diff --git a/.github/workflows/main.yml b/.github/workflows/main.yml
index 900be9957a23fc..b1ed3794e2b2c8 100644
--- a/.github/workflows/main.yml
+++ b/.github/workflows/main.yml
@@ -169,8 +169,11 @@ jobs:
       NO_PERL: 1
       GIT_CONFIG_PARAMETERS: "'user.name=CI' 'user.email=ci@git'"
     runs-on: windows-latest
+    strategy:
+      matrix:
+        arch: [x64, arm64]
     concurrency:
-      group: vs-build-${{ github.ref }}
+      group: vs-build-${{ github.ref }}-${{ matrix.arch }}
       cancel-in-progress: ${{ needs.ci-config.outputs.skip_concurrent == 'yes' }}
     steps:
     - uses: actions/checkout@v4
@@ -189,14 +192,14 @@ jobs:
       uses: microsoft/setup-msbuild@v2
     - name: copy dlls to root
       shell: cmd
-      run: compat\vcbuild\vcpkg_copy_dlls.bat release
+      run: compat\vcbuild\vcpkg_copy_dlls.bat release ${{ matrix.arch }}-windows
     - name: generate Visual Studio solution
       shell: bash
       run: |
-        cmake `pwd`/contrib/buildsystems/ -DCMAKE_PREFIX_PATH=`pwd`/compat/vcbuild/vcpkg/installed/x64-windows \
-        -DNO_GETTEXT=YesPlease -DPERL_TESTS=OFF -DPYTHON_TESTS=OFF -DCURL_NO_CURL_CMAKE=ON
+        cmake `pwd`/contrib/buildsystems/ -DCMAKE_PREFIX_PATH=`pwd`/compat/vcbuild/vcpkg/installed/${{ matrix.arch }}-windows \
+        -DNO_GETTEXT=YesPlease -DPERL_TESTS=OFF -DPYTHON_TESTS=OFF -DCURL_NO_CURL_CMAKE=ON -DCMAKE_GENERATOR_PLATFORM=${{ matrix.arch }} -DVCPKG_ARCH=${{ matrix.arch }}-windows -DHOST_CPU=${{ matrix.arch }}
     - name: MSBuild
-      run: msbuild git.sln -property:Configuration=Release -property:Platform=x64 -maxCpuCount:4 -property:PlatformToolset=v142
+      run: msbuild git.sln -property:Configuration=Release -property:Platform=${{ matrix.arch }} -maxCpuCount:4 -property:PlatformToolset=v142
     - name: bundle artifact tar
       shell: bash
       env:
@@ -210,7 +213,7 @@ jobs:
     - name: upload tracked files and build artifacts
       uses: actions/upload-artifact@v4
       with:
-        name: vs-artifacts
+        name: vs-artifacts-${{ matrix.arch }}
         path: artifacts
   vs-test:
     name: win+VS test
@@ -228,7 +231,7 @@ jobs:
     - name: download tracked files and build artifacts
       uses: actions/download-artifact@v4
       with:
-        name: vs-artifacts
+        name: vs-artifacts-x64
         path: ${{github.workspace}}
     - name: extract tracked files and build artifacts
       shell: bash
diff --git a/.github/workflows/nano-server.yml b/.github/workflows/nano-server.yml
new file mode 100644
index 00000000000000..3d943c23d2616d
--- /dev/null
+++ b/.github/workflows/nano-server.yml
@@ -0,0 +1,76 @@
+name: Windows Nano Server tests
+
+on:
+  workflow_dispatch:
+
+env:
+  DEVELOPER: 1
+
+jobs:
+  test-nano-server:
+    runs-on: windows-2022
+    env:
+      WINDBG_DIR: "C:/Program Files (x86)/Windows Kits/10/Debuggers/x64"
+      IMAGE: mcr.microsoft.com/powershell:nanoserver-ltsc2022
+
+    steps:
+      - uses: actions/checkout@v4
+      - uses: git-for-windows/setup-git-for-windows-sdk@v1
+      - name: build Git
+        shell: bash
+        run: make -j15
+      - name: pull nanoserver image
+        shell: bash
+        run: docker pull $IMAGE
+      - name: run nano-server test
+        shell: bash
+        run: |
+          docker run \
+            --user "ContainerAdministrator" \
+            -v "$WINDBG_DIR:C:/dbg" \
+            -v "$(cygpath -aw /mingw64/bin):C:/mingw64-bin" \
+            -v "$(cygpath -aw .):C:/test" \
+            $IMAGE pwsh.exe -Command '
+              # Extend the PATH to include the `.dll` files in /mingw64/bin/
+              $env:PATH += ";C:\mingw64-bin"
+
+              # For each executable to test pick some no-operation set of
+              # flags/subcommands or something that should quickly result in an
+              # error with known exit code that is not a negative 32-bit
+              # number, and set the expected return code appropriately.
+              #
+              # Only test executables that could be expected to run in a UI
+              # less environment.
+              #
+              # ( Executable path, arguments, expected return code )
+              # also note space is required before close parenthesis (a
+              # powershell quirk when defining nested arrays like this)
+
+              $executables_to_test = @(
+                  ("C:\test\git.exe", "", 1 ),
+                  ("C:\test\scalar.exe", "version", 0 )
+              )
+
+              foreach ($executable in $executables_to_test)
+              {
+                  Write-Output "Now testing $($executable[0])"
+                  &$executable[0] $executable[1]
+                  if ($LASTEXITCODE -ne $executable[2]) {
+                      # if we failed, run the debugger to find out what function
+                      # or DLL could not be found and then exit the script with
+                      # failure The missing DLL or EXE will be referenced near
+                      # the end of the output
+
+                      # Set a flag to have the debugger show loader stub
+                      # diagnostics. This requires running as administrator,
+                      # otherwise the flag will be ignored.
+                      C:\dbg\gflags -i $executable[0] +SLS
+
+                      C:\dbg\cdb.exe -c "g" -c "q" $executable[0] $executable[1]
+
+                      exit 1
+                  }
+              }
+
+              exit 0
+            '
diff --git a/.gitignore b/.gitignore
index e82aa19df03fc8..9dd1fbc61e6c82 100644
--- a/.gitignore
+++ b/.gitignore
@@ -19,6 +19,7 @@
 /git-apply
 /git-archimport
 /git-archive
+/git-backfill
 /git-bisect
 /git-blame
 /git-branch
@@ -164,6 +165,7 @@
 /git-submodule
 /git-submodule--helper
 /git-subtree
+/git-survey
 /git-svn
 /git-switch
 /git-symbolic-ref
@@ -250,3 +252,4 @@ Release/
 /git.VC.db
 *.dSYM
 /contrib/buildsystems/out
+CMakeSettings.json
diff --git a/Documentation/config.txt b/Documentation/config.txt
index 8c0b3ed8075214..939cc1387992f8 100644
--- a/Documentation/config.txt
+++ b/Documentation/config.txt
@@ -518,6 +518,8 @@ include::config/safe.txt[]
 
 include::config/sendemail.txt[]
 
+include::config/sendpack.txt[]
+
 include::config/sequencer.txt[]
 
 include::config/showbranch.txt[]
@@ -534,6 +536,8 @@ include::config/status.txt[]
 
 include::config/submodule.txt[]
 
+include::config/survey.txt[]
+
 include::config/tag.txt[]
 
 include::config/tar.txt[]
@@ -554,4 +558,6 @@ include::config/versionsort.txt[]
 
 include::config/web.txt[]
 
+include::config/windows.txt[]
+
 include::config/worktree.txt[]
diff --git a/Documentation/config/feature.txt b/Documentation/config/feature.txt
index f061b64b748449..cb49ff2604a632 100644
--- a/Documentation/config/feature.txt
+++ b/Documentation/config/feature.txt
@@ -20,6 +20,10 @@ walking fewer objects.
 +
 * `pack.allowPackReuse=multi` may improve the time it takes to create a pack by
 reusing objects from multiple packs instead of just one.
++
+* `pack.usePathWalk` may speed up packfile creation and make the packfiles be
+significantly smaller in the presence of certain filename collisions with Git's
+default name-hash.
 
 feature.manyFiles::
 	Enable config options that optimize for repos with many files in the
diff --git a/Documentation/config/http.txt b/Documentation/config/http.txt
index a14371b5c96a51..ddda264dbed9b7 100644
--- a/Documentation/config/http.txt
+++ b/Documentation/config/http.txt
@@ -218,11 +218,13 @@ http.sslBackend::
 
 http.schannelCheckRevoke::
 	Used to enforce or disable certificate revocation checks in cURL
-	when http.sslBackend is set to "schannel". Defaults to `true` if
-	unset. Only necessary to disable this if Git consistently errors
-	and the message is about checking the revocation status of a
-	certificate. This option is ignored if cURL lacks support for
-	setting the relevant SSL option at runtime.
+	when http.sslBackend is set to "schannel" via "true" and "false",
+	respectively. Another accepted value is "best-effort" (the default)
+	in which case revocation checks are performed, but errors due to
+	revocation list distribution points that are offline are silently
+	ignored, as well as errors due to certificates missing revocation
+	list distribution points. This option is ignored if cURL lacks
+	support for setting the relevant SSL option at runtime.
 
 http.schannelUseSSLCAInfo::
 	As of cURL v7.60.0, the Secure Channel backend can use the
@@ -232,6 +234,11 @@ http.schannelUseSSLCAInfo::
 	when the `schannel` backend was configured via `http.sslBackend`,
 	unless `http.schannelUseSSLCAInfo` overrides this behavior.
 
+http.sslAutoClientCert::
+	As of cURL v7.77.0, the Secure Channel backend won't automatically
+	send client certificates from the Windows Certificate Store anymore.
+	To opt in to the old behavior, http.sslAutoClientCert can be set.
+
 http.pinnedPubkey::
 	Public key of the https service. It may either be the filename of
 	a PEM or DER encoded public key file or a string starting with
diff --git a/Documentation/config/pack.txt b/Documentation/config/pack.txt
index da527377fafcb6..08d06271177006 100644
--- a/Documentation/config/pack.txt
+++ b/Documentation/config/pack.txt
@@ -155,6 +155,14 @@ pack.useSparse::
 	commits contain certain types of direct renames. Default is
 	`true`.
 
+pack.usePathWalk::
+	When true, git will default to using the '--path-walk' option in
+	'git pack-objects' when the '--revs' option is present. This
+	algorithm groups objects by path to maximize the ability to
+	compute delta chains across historical versions of the same
+	object. This may disable other options, such as using bitmaps to
+	enumerate objects.
+
 pack.preferBitmapTips::
 	When selecting which commits will receive bitmaps, prefer a
 	commit at the tip of any reference that is a suffix of any value
diff --git a/Documentation/config/sendpack.txt b/Documentation/config/sendpack.txt
new file mode 100644
index 00000000000000..e306f657fba7dd
--- /dev/null
+++ b/Documentation/config/sendpack.txt
@@ -0,0 +1,5 @@
+sendpack.sideband::
+	Allows to disable the side-band-64k capability for send-pack even
+	when it is advertised by the server. Makes it possible to work
+	around a limitation in the git for windows implementation together
+	with the dump git protocol. Defaults to true.
diff --git a/Documentation/config/survey.txt b/Documentation/config/survey.txt
new file mode 100644
index 00000000000000..9e594a2092f225
--- /dev/null
+++ b/Documentation/config/survey.txt
@@ -0,0 +1,14 @@
+survey.*::
+	These variables adjust the default behavior of the `git survey`
+	command. The intention is that this command could be run in the
+	background with these options.
++
+--
+	verbose::
+		This boolean value implies the `--[no-]verbose` option.
+	progress::
+		This boolean value implies the `--[no-]progress` option.
+	top::
+		This integer value implies `--top=<N>`, specifying the
+		number of entries in the detail tables.
+--
diff --git a/Documentation/config/windows.txt b/Documentation/config/windows.txt
new file mode 100644
index 00000000000000..fdaaf1c65504f3
--- /dev/null
+++ b/Documentation/config/windows.txt
@@ -0,0 +1,4 @@
+windows.appendAtomically::
+	By default, append atomic API is used on windows. But it works only with
+	local disk files, if you're working on a network file system, you should
+	set it false to turn it off.
diff --git a/Documentation/git-backfill.txt b/Documentation/git-backfill.txt
new file mode 100644
index 00000000000000..2e1f8505209cbb
--- /dev/null
+++ b/Documentation/git-backfill.txt
@@ -0,0 +1,60 @@
+git-backfill(1)
+===============
+
+NAME
+----
+git-backfill - Download missing objects in a partial clone
+
+
+SYNOPSIS
+--------
+[verse]
+(EXPERIMENTAL) 'git backfill' [--batch-size=<n>] [--[no-]sparse]
+
+DESCRIPTION
+-----------
+
+Blobless partial clones are created using `git clone --filter=blob:none`
+and then configure the local repository such that the Git client avoids
+downloading blob objects unless they are required for a local operation.
+This initially means that the clone and later fetches download reachable
+commits and trees but no blobs. Later operations that change the `HEAD`
+pointer, such as `git checkout` or `git merge`, may need to download
+missing blobs in order to complete their operation.
+
+In the worst cases, commands that compute blob diffs, such as `git blame`,
+become very slow as they download the missing blobs in single-blob
+requests to satisfy the missing object as the Git command needs it. This
+leads to multiple download requests and no ability for the Git server to
+provide delta compression across those objects.
+
+The `git backfill` command provides a way for the user to request that
+Git downloads the missing blobs (with optional filters) such that the
+missing blobs representing historical versions of files can be downloaded
+in batches. The `backfill` command attempts to optimize the request by
+grouping blobs that appear at the same path, hopefully leading to good
+delta compression in the packfile sent by the server.
+
+By default, `git backfill` downloads all blobs reachable from the `HEAD`
+commit. This set can be restricted or expanded using various options.
+
+OPTIONS
+-------
+
+--batch-size=<n>::
+	Specify a minimum size for a batch of missing objects to request
+	from the server. This size may be exceeded by the last set of
+	blobs seen at a given path. Default batch size is 16,000.
+
+--[no-]sparse::
+	Only download objects if they appear at a path that matches the
+	current sparse-checkout. If the sparse-checkout feature is enabled,
+	then `--sparse` is assumed and can be disabled with `--no-sparse`.
+
+SEE ALSO
+--------
+linkgit:git-clone[1].
+
+GIT
+---
+Part of the linkgit:git[1] suite
diff --git a/Documentation/git-pack-objects.txt b/Documentation/git-pack-objects.txt
index e32404c6aaee30..1a3023e2067f28 100644
--- a/Documentation/git-pack-objects.txt
+++ b/Documentation/git-pack-objects.txt
@@ -15,7 +15,8 @@ SYNOPSIS
 	[--revs [--unpacked | --all]] [--keep-pack=<pack-name>]
 	[--cruft] [--cruft-expiration=<time>]
 	[--stdout [--filter=<filter-spec>] | <base-name>]
-	[--shallow] [--keep-true-parents] [--[no-]sparse] < <object-list>
+	[--shallow] [--keep-true-parents] [--[no-]sparse]
+	[--full-name-hash] [--path-walk] < <object-list>
 
 
 DESCRIPTION
@@ -345,6 +346,16 @@ raise an error.
 	Restrict delta matches based on "islands". See DELTA ISLANDS
 	below.
 
+--path-walk::
+	By default, `git pack-objects` walks objects in an order that
+	presents trees and blobs in an order unrelated to the path they
+	appear relative to a commit's root tree. The `--path-walk` option
+	enables a different walking algorithm that organizes trees and
+	blobs by path. This has the potential to improve delta compression
+	especially in the presence of filenames that cause collisions in
+	Git's default name-hash algorithm. Due to changing how the objects
+	are walked, this option is not compatible with `--delta-islands`,
+	`--shallow`, or `--filter`.
 
 DELTA ISLANDS
 -------------
diff --git a/Documentation/git-repack.txt b/Documentation/git-repack.txt
index c902512a9e89b0..717e0a8e41a03b 100644
--- a/Documentation/git-repack.txt
+++ b/Documentation/git-repack.txt
@@ -9,7 +9,9 @@ git-repack - Pack unpacked objects in a repository
 SYNOPSIS
 --------
 [verse]
-'git repack' [-a] [-A] [-d] [-f] [-F] [-l] [-n] [-q] [-b] [-m] [--window=<n>] [--depth=<n>] [--threads=<n>] [--keep-pack=<pack-name>] [--write-midx]
+'git repack' [-a] [-A] [-d] [-f] [-F] [-l] [-n] [-q] [-b] [-m]
+	[--window=<n>] [--depth=<n>] [--threads=<n>] [--keep-pack=<pack-name>]
+	[--write-midx] [--full-name-hash] [--path-walk]
 
 DESCRIPTION
 -----------
@@ -249,6 +251,19 @@ linkgit:git-multi-pack-index[1]).
 	Write a multi-pack index (see linkgit:git-multi-pack-index[1])
 	containing the non-redundant packs.
 
+--path-walk::
+	This option passes the `--path-walk` option to the underlying
+	`git pack-options` process (see linkgit:git-pack-objects[1]).
+	By default, `git pack-objects` walks objects in an order that
+	presents trees and blobs in an order unrelated to the path they
+	appear relative to a commit's root tree. The `--path-walk` option
+	enables a different walking algorithm that organizes trees and
+	blobs by path. This has the potential to improve delta compression
+	especially in the presence of filenames that cause collisions in
+	Git's default name-hash algorithm. Due to changing how the objects
+	are walked, this option is not compatible with `--delta-islands`
+	or `--filter`.
+
 CONFIGURATION
 -------------
 
diff --git a/Documentation/git-survey.txt b/Documentation/git-survey.txt
new file mode 100644
index 00000000000000..44f3a0568b7697
--- /dev/null
+++ b/Documentation/git-survey.txt
@@ -0,0 +1,83 @@
+git-survey(1)
+=============
+
+NAME
+----
+git-survey - EXPERIMENTAL: Measure various repository dimensions of scale
+
+SYNOPSIS
+--------
+[verse]
+(EXPERIMENTAL!) 'git survey' <options>
+
+DESCRIPTION
+-----------
+
+Survey the repository and measure various dimensions of scale.
+
+As repositories grow to "monorepo" size, certain data shapes can cause
+performance problems.  `git-survey` attempts to measure and report on
+known problem areas.
+
+Ref Selection and Reachable Objects
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+In this first analysis phase, `git survey` will iterate over the set of
+requested branches, tags, and other refs and treewalk over all of the
+reachable commits, trees, and blobs and generate various statistics.
+
+OPTIONS
+-------
+
+--progress::
+	Show progress.  This is automatically enabled when interactive.
+
+Ref Selection
+~~~~~~~~~~~~~
+
+The following options control the set of refs that `git survey` will examine.
+By default, `git survey` will look at tags, local branches, and remote refs.
+If any of the following options are given, the default set is cleared and
+only refs for the given options are added.
+
+--all-refs::
+	Use all refs.  This includes local branches, tags, remote refs,
+	notes, and stashes.  This option overrides all of the following.
+
+--branches::
+	Add local branches (`refs/heads/`) to the set.
+
+--tags::
+	Add tags (`refs/tags/`) to the set.
+
+--remotes::
+	Add remote branches (`refs/remote/`) to the set.
+
+--detached::
+	Add HEAD to the set.
+
+--other::
+	Add notes (`refs/notes/`) and stashes (`refs/stash/`) to the set.
+
+OUTPUT
+------
+
+By default, `git survey` will print information about the repository in a
+human-readable format that includes overviews and tables.
+
+References Summary
+~~~~~~~~~~~~~~~~~~
+
+The references summary includes a count of each kind of reference,
+including branches, remote refs, and tags (split by "all" and
+"annotated").
+
+Reachable Object Summary
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+The reachable object summary shows the total number of each kind of Git
+object, including tags, commits, trees, and blobs.
+
+GIT
+---
+Part of the linkgit:git[1] suite
diff --git a/Documentation/technical/api-path-walk.txt b/Documentation/technical/api-path-walk.txt
new file mode 100644
index 00000000000000..83bfe3d665e9fb
--- /dev/null
+++ b/Documentation/technical/api-path-walk.txt
@@ -0,0 +1,82 @@
+Path-Walk API
+=============
+
+The path-walk API is used to walk reachable objects, but to visit objects
+in batches based on a common path they appear in, or by type.
+
+For example, all reachable commits are visited in a group. All tags are
+visited in a group. Then, all root trees are visited. At some point, all
+blobs reachable via a path `my/dir/to/A` are visited. When there are
+multiple paths possible to reach the same object, then only one of those
+paths is used to visit the object.
+
+When walking a range of commits with some `UNINTERESTING` objects, the
+objects with the `UNINTERESTING` flag are included in these batches. In
+order to walk `UNINTERESTING` objects, the `--boundary` option must be
+used in the commit walk in order to visit `UNINTERESTING` commits.
+
+Basics
+------
+
+To use the path-walk API, include `path-walk.h` and call
+`walk_objects_by_path()` with a customized `path_walk_info` struct. The
+struct is used to set all of the options for how the walk should proceed.
+Let's dig into the different options and their use.
+
+`path_fn` and `path_fn_data`::
+	The most important option is the `path_fn` option, which is a
+	function pointer to the callback that can execute logic on the
+	object IDs for objects grouped by type and path. This function
+	also receives a `data` value that corresponds to the
+	`path_fn_data` member, for providing custom data structures to
+	this callback function.
+
+`revs`::
+	To configure the exact details of the reachable set of objects,
+	use the `revs` member and initialize it using the revision
+	machinery in `revision.h`. Initialize `revs` using calls such as
+	`setup_revisions()` or `parse_revision_opt()`. Do not call
+	`prepare_revision_walk()`, as that will be called within
+	`walk_objects_by_path()`.
++
+It is also important that you do not specify the `--objects` flag for the
+`revs` struct. The revision walk should only be used to walk commits, and
+the objects will be walked in a separate way based on those starting
+commits.
++
+If you want the path-walk API to emit `UNINTERESTING` objects based on the
+commit walk's boundary, be sure to set `revs.boundary` so the boundary
+commits are emitted.
+
+`commits`, `blobs`, `trees`, `tags`::
+	By default, these members are enabled and signal that the path-walk
+	API should call the `path_fn` on objects of these types. Specialized
+	applications could disable some options to make it simpler to walk
+	the objects or to have fewer calls to `path_fn`.
++
+While it is possible to walk only commits in this way, consumers would be
+better off using the revision walk API instead.
+
+`prune_all_uninteresting`::
+	By default, all reachable paths are emitted by the path-walk API.
+	This option allows consumers to declare that they are not
+	interested in paths where all included objects are marked with the
+	`UNINTERESTING` flag. This requires using the `boundary` option in
+	the revision walk so that the walk emits commits marked with the
+	`UNINTERESTING` flag.
+
+`pl`::
+	This pattern list pointer allows focusing the path-walk search to
+	a set of patterns, only emitting paths that match the given
+	patterns. See linkgit:gitignore[5] or
+	linkgit:git-sparse-checkout[1] for details about pattern lists.
+	When the pattern list uses cone-mode patterns, then the path-walk
+	API can prune the set of paths it walks to improve performance.
+
+Examples
+--------
+
+See example usages in:
+	`t/helper/test-path-walk.c`,
+	`builtin/backfill.c`,
+	`builtin/pack-objects.c`
diff --git a/Makefile b/Makefile
index 2c6dad8a7513be..334eff4f75c029 100644
--- a/Makefile
+++ b/Makefile
@@ -475,6 +475,11 @@ include shared.mak
 #
 #     CURL_LDFLAGS=-lcurl
 #
+# Define LAZYLOAD_LIBCURL to dynamically load the libcurl; This can be useful
+# if Multiple libcurl versions exist (with different file names) that link to
+# various SSL/TLS backends, to support the `http.sslBackend` runtime switch in
+# such a scenario.
+#
 # === Optional library: libpcre2 ===
 #
 # Define USE_LIBPCRE if you have and want to use libpcre. Various
@@ -816,12 +821,14 @@ TEST_BUILTINS_OBJS += test-lazy-init-name-hash.o
 TEST_BUILTINS_OBJS += test-match-trees.o
 TEST_BUILTINS_OBJS += test-mergesort.o
 TEST_BUILTINS_OBJS += test-mktemp.o
+TEST_BUILTINS_OBJS += test-name-hash.o
 TEST_BUILTINS_OBJS += test-online-cpus.o
 TEST_BUILTINS_OBJS += test-pack-mtimes.o
 TEST_BUILTINS_OBJS += test-parse-options.o
 TEST_BUILTINS_OBJS += test-parse-pathspec-file.o
 TEST_BUILTINS_OBJS += test-partial-clone.o
 TEST_BUILTINS_OBJS += test-path-utils.o
+TEST_BUILTINS_OBJS += test-path-walk.o
 TEST_BUILTINS_OBJS += test-pcre2-config.o
 TEST_BUILTINS_OBJS += test-pkt-line.o
 TEST_BUILTINS_OBJS += test-proc-receive.o
@@ -1098,6 +1105,7 @@ LIB_OBJS += parse-options.o
 LIB_OBJS += patch-delta.o
 LIB_OBJS += patch-ids.o
 LIB_OBJS += path.o
+LIB_OBJS += path-walk.o
 LIB_OBJS += pathspec.o
 LIB_OBJS += pkt-line.o
 LIB_OBJS += preload-index.o
@@ -1205,6 +1213,7 @@ BUILTIN_OBJS += builtin/am.o
 BUILTIN_OBJS += builtin/annotate.o
 BUILTIN_OBJS += builtin/apply.o
 BUILTIN_OBJS += builtin/archive.o
+BUILTIN_OBJS += builtin/backfill.o
 BUILTIN_OBJS += builtin/bisect.o
 BUILTIN_OBJS += builtin/blame.o
 BUILTIN_OBJS += builtin/branch.o
@@ -1309,6 +1318,7 @@ BUILTIN_OBJS += builtin/sparse-checkout.o
 BUILTIN_OBJS += builtin/stash.o
 BUILTIN_OBJS += builtin/stripspace.o
 BUILTIN_OBJS += builtin/submodule--helper.o
+BUILTIN_OBJS += builtin/survey.o
 BUILTIN_OBJS += builtin/symbolic-ref.o
 BUILTIN_OBJS += builtin/tag.o
 BUILTIN_OBJS += builtin/unpack-file.o
@@ -1332,6 +1342,7 @@ BUILTIN_OBJS += builtin/write-tree.o
 # upstream unnecessarily (making merging in future changes easier).
 THIRD_PARTY_SOURCES += compat/inet_ntop.c
 THIRD_PARTY_SOURCES += compat/inet_pton.c
+THIRD_PARTY_SOURCES += compat/mimalloc/%
 THIRD_PARTY_SOURCES += compat/nedmalloc/%
 THIRD_PARTY_SOURCES += compat/obstack.%
 THIRD_PARTY_SOURCES += compat/poll/%
@@ -1651,10 +1662,23 @@ else
 		CURL_LIBCURL =
         endif
 
-        ifndef CURL_LDFLAGS
-		CURL_LDFLAGS = $(eval CURL_LDFLAGS := $$(shell $$(CURL_CONFIG) --libs))$(CURL_LDFLAGS)
+        ifdef LAZYLOAD_LIBCURL
+		LAZYLOAD_LIBCURL_OBJ = compat/lazyload-curl.o
+		OBJECTS += $(LAZYLOAD_LIBCURL_OBJ)
+		# The `CURL_STATICLIB` constant must be defined to avoid seeing the functions
+		# declared as DLL imports
+		CURL_CFLAGS = -DCURL_STATICLIB
+ifneq ($(uname_S),MINGW)
+ifneq ($(uname_S),Windows)
+		CURL_LIBCURL = -ldl
+endif
+endif
+        else
+                ifndef CURL_LDFLAGS
+			CURL_LDFLAGS = $(eval CURL_LDFLAGS := $$(shell $$(CURL_CONFIG) --libs))$(CURL_LDFLAGS)
+                endif
+		CURL_LIBCURL += $(CURL_LDFLAGS)
         endif
-	CURL_LIBCURL += $(CURL_LDFLAGS)
 
         ifndef CURL_CFLAGS
 		CURL_CFLAGS = $(eval CURL_CFLAGS := $$(shell $$(CURL_CONFIG) --cflags))$(CURL_CFLAGS)
@@ -1675,7 +1699,7 @@ else
         endif
         ifdef USE_CURL_FOR_IMAP_SEND
 		BASIC_CFLAGS += -DUSE_CURL_FOR_IMAP_SEND
-		IMAP_SEND_BUILDDEPS = http.o
+		IMAP_SEND_BUILDDEPS = http.o $(LAZYLOAD_LIBCURL_OBJ)
 		IMAP_SEND_LDFLAGS += $(CURL_LIBCURL)
         endif
         ifndef NO_EXPAT
@@ -2130,6 +2154,43 @@ ifdef USE_NED_ALLOCATOR
 	OVERRIDE_STRDUP = YesPlease
 endif
 
+ifdef USE_MIMALLOC
+	MIMALLOC_OBJS = \
+		compat/mimalloc/alloc-aligned.o \
+		compat/mimalloc/alloc.o \
+		compat/mimalloc/arena.o \
+		compat/mimalloc/bitmap.o \
+		compat/mimalloc/heap.o \
+		compat/mimalloc/init.o \
+		compat/mimalloc/options.o \
+		compat/mimalloc/os.o \
+		compat/mimalloc/page.o \
+		compat/mimalloc/random.o \
+		compat/mimalloc/prim/windows/prim.o \
+		compat/mimalloc/segment.o \
+		compat/mimalloc/segment-cache.o \
+		compat/mimalloc/segment-map.o \
+		compat/mimalloc/stats.o
+
+	COMPAT_CFLAGS += -Icompat/mimalloc -DMI_DEBUG=0 -DUSE_MIMALLOC --std=gnu11
+	COMPAT_OBJS += $(MIMALLOC_OBJS)
+
+$(MIMALLOC_OBJS): COMPAT_CFLAGS += -DBANNED_H
+
+$(MIMALLOC_OBJS): COMPAT_CFLAGS += \
+	-Wno-attributes \
+	-Wno-unknown-pragmas \
+	-Wno-array-bounds
+
+ifdef DEVELOPER
+$(MIMALLOC_OBJS): COMPAT_CFLAGS += \
+	-Wno-pedantic \
+	-Wno-declaration-after-statement \
+	-Wno-old-style-definition \
+	-Wno-missing-prototypes
+endif
+endif
+
 ifdef OVERRIDE_STRDUP
 	COMPAT_CFLAGS += -DOVERRIDE_STRDUP
 	COMPAT_OBJS += compat/strdup.o
@@ -2856,10 +2917,10 @@ git-imap-send$X: imap-send.o $(IMAP_SEND_BUILDDEPS) GIT-LDFLAGS $(GITLIBS)
 	$(QUIET_LINK)$(CC) $(ALL_CFLAGS) -o $@ $(ALL_LDFLAGS) $(filter %.o,$^) \
 		$(IMAP_SEND_LDFLAGS) $(LIBS)
 
-git-http-fetch$X: http.o http-walker.o http-fetch.o GIT-LDFLAGS $(GITLIBS)
+git-http-fetch$X: http.o http-walker.o http-fetch.o $(LAZYLOAD_LIBCURL_OBJ) GIT-LDFLAGS $(GITLIBS)
 	$(QUIET_LINK)$(CC) $(ALL_CFLAGS) -o $@ $(ALL_LDFLAGS) $(filter %.o,$^) \
 		$(CURL_LIBCURL) $(LIBS)
-git-http-push$X: http.o http-push.o GIT-LDFLAGS $(GITLIBS)
+git-http-push$X: http.o http-push.o $(LAZYLOAD_LIBCURL_OBJ) GIT-LDFLAGS $(GITLIBS)
 	$(QUIET_LINK)$(CC) $(ALL_CFLAGS) -o $@ $(ALL_LDFLAGS) $(filter %.o,$^) \
 		$(CURL_LIBCURL) $(EXPAT_LIBEXPAT) $(LIBS)
 
@@ -2869,7 +2930,7 @@ $(REMOTE_CURL_ALIASES): $(REMOTE_CURL_PRIMARY)
 	ln -s $< $@ 2>/dev/null || \
 	cp $< $@
 
-$(REMOTE_CURL_PRIMARY): remote-curl.o http.o http-walker.o GIT-LDFLAGS $(GITLIBS)
+$(REMOTE_CURL_PRIMARY): remote-curl.o http.o http-walker.o $(LAZYLOAD_LIBCURL_OBJ) GIT-LDFLAGS $(GITLIBS)
 	$(QUIET_LINK)$(CC) $(ALL_CFLAGS) -o $@ $(ALL_LDFLAGS) $(filter %.o,$^) \
 		$(CURL_LIBCURL) $(EXPAT_LIBEXPAT) $(LIBS)
 
@@ -3747,12 +3808,15 @@ ifdef MSVC
 	$(RM) $(patsubst %.o,%.o.pdb,$(OBJECTS))
 	$(RM) headless-git.o.pdb
 	$(RM) $(patsubst %.exe,%.pdb,$(OTHER_PROGRAMS))
+	$(RM) $(patsubst %.exe,%.ilk,$(OTHER_PROGRAMS))
 	$(RM) $(patsubst %.exe,%.iobj,$(OTHER_PROGRAMS))
 	$(RM) $(patsubst %.exe,%.ipdb,$(OTHER_PROGRAMS))
 	$(RM) $(patsubst %.exe,%.pdb,$(PROGRAMS))
+	$(RM) $(patsubst %.exe,%.ilk,$(PROGRAMS))
 	$(RM) $(patsubst %.exe,%.iobj,$(PROGRAMS))
 	$(RM) $(patsubst %.exe,%.ipdb,$(PROGRAMS))
 	$(RM) $(patsubst %.exe,%.pdb,$(TEST_PROGRAMS))
+	$(RM) $(patsubst %.exe,%.ilk,$(TEST_PROGRAMS))
 	$(RM) $(patsubst %.exe,%.iobj,$(TEST_PROGRAMS))
 	$(RM) $(patsubst %.exe,%.ipdb,$(TEST_PROGRAMS))
 	$(RM) compat/vcbuild/MSVC-DEFS-GEN
diff --git a/abspath.c b/abspath.c
index 1202cde23dbc9b..0c17e98654e4b0 100644
--- a/abspath.c
+++ b/abspath.c
@@ -93,6 +93,9 @@ static char *strbuf_realpath_1(struct strbuf *resolved, const char *path,
 			goto error_out;
 	}
 
+	if (platform_strbuf_realpath(resolved, path))
+		return resolved->buf;
+
 	strbuf_addstr(&remaining, path);
 	get_root_part(resolved, &remaining);
 
diff --git a/azure-pipelines.yml b/azure-pipelines.yml
new file mode 100644
index 00000000000000..fed26341d49a4a
--- /dev/null
+++ b/azure-pipelines.yml
@@ -0,0 +1,393 @@
+variables:
+  Agent.Source.Git.ShallowFetchDepth: 1
+  GIT_CONFIG_PARAMETERS: "'checkout.workers=56' 'user.name=CI' 'user.email=ci@git'"
+
+jobs:
+- job: windows_build
+  displayName: Windows Build
+  condition: succeeded()
+  pool:
+    vmImage: windows-latest
+  timeoutInMinutes: 240
+  steps:
+  - bash: git clone --bare --depth=1 --filter=blob:none --single-branch -b main https://github.com/git-for-windows/git-sdk-64
+    displayName: 'clone git-sdk-64'
+  - bash: git clone --depth=1 --single-branch -b main https://github.com/git-for-windows/build-extra
+    displayName: 'clone build-extra'
+  - bash: sh -x ./build-extra/please.sh create-sdk-artifact --sdk=git-sdk-64.git --out=git-sdk-64-minimal minimal-sdk
+    displayName: 'build git-sdk-64-minimal-sdk'
+  - bash: |
+      # Let Git ignore the SDK and the test-cache
+      printf "%s\n" /git-sdk-64.git/ /build-extra/ /git-sdk-64-minimal/ /test-cache/ >>'.git/info/exclude'
+    displayName: 'Ignore untracked directories'
+  - bash: ci/make-test-artifacts.sh artifacts
+    displayName: Build
+    env:
+      HOME: $(Build.SourcesDirectory)
+      MSYSTEM: MINGW64
+      DEVELOPER: 1
+      NO_PERL: 1
+      PATH: "$(Build.SourcesDirectory)\\git-sdk-64-minimal\\mingw64\\bin;$(Build.SourcesDirectory)\\git-sdk-64-minimal\\usr\\bin;C:\\Windows\\system32;C:\\Windows;C:\\Windows\\system32\\wbem"
+  - task: PublishPipelineArtifact@0
+    displayName: 'Publish Pipeline Artifact: test artifacts'
+    inputs:
+      artifactName: 'windows-artifacts'
+      targetPath: '$(Build.SourcesDirectory)\artifacts'
+  - task: PublishPipelineArtifact@0
+    displayName: 'Publish Pipeline Artifact: git-sdk-64-minimal'
+    inputs:
+      artifactName: 'git-sdk-64-minimal'
+      targetPath: '$(Build.SourcesDirectory)\git-sdk-64-minimal'
+
+- job: windows_test
+  displayName: Windows Test
+  dependsOn: windows_build
+  condition: succeeded()
+  pool:
+    vmImage: windows-latest
+  timeoutInMinutes: 240
+  strategy:
+    parallel: 10
+  steps:
+  - task: DownloadPipelineArtifact@0
+    displayName: 'Download Pipeline Artifact: test artifacts'
+    inputs:
+      artifactName: 'windows-artifacts'
+      targetPath: '$(Build.SourcesDirectory)'
+  - task: DownloadPipelineArtifact@0
+    displayName: 'Download Pipeline Artifact: git-sdk-64-minimal'
+    inputs:
+      artifactName: 'git-sdk-64-minimal'
+      targetPath: '$(Build.SourcesDirectory)\git-sdk-64-minimal'
+  - bash: |
+      test -f artifacts.tar.gz || {
+        echo No test artifacts found\; skipping >&2
+        exit 0
+      }
+      tar xf artifacts.tar.gz || exit 1
+
+      # Let Git ignore the SDK and the test-cache
+      printf '%s\n' /git-sdk-64.git/ /build-extra/ /git-sdk-64-minimal/ /test-cache/ >>.git/info/exclude
+
+      ci/run-test-slice.sh $SYSTEM_JOBPOSITIONINPHASE $SYSTEM_TOTALJOBSINPHASE || {
+        ci/print-test-failures.sh
+        exit 1
+      }
+    displayName: 'Test (parallel)'
+    env:
+      HOME: $(Build.SourcesDirectory)
+      MSYSTEM: MINGW64
+      NO_SVN_TESTS: 1
+      GIT_TEST_SKIP_REBASE_P: 1
+      PATH: "$(Build.SourcesDirectory)\\git-sdk-64-minimal\\mingw64\\bin;$(Build.SourcesDirectory)\\git-sdk-64-minimal\\usr\\bin\\core_perl;$(Build.SourcesDirectory)\\git-sdk-64-minimal\\usr\\bin;C:\\Windows\\system32;C:\\Windows;C:\\Windows\\system32\\wbem"
+  - task: PublishTestResults@2
+    displayName: 'Publish Test Results **/TEST-*.xml'
+    inputs:
+      mergeTestResults: true
+      testRunTitle: 'windows'
+      platform: Windows
+      publishRunAttachments: false
+    condition: succeededOrFailed()
+  - task: PublishBuildArtifacts@1
+    displayName: 'Publish trash directories of failed tests'
+    condition: failed()
+    inputs:
+      PathtoPublish: t/failed-test-artifacts
+      ArtifactName: failed-test-artifacts
+
+- job: vs_build
+  displayName: Visual Studio Build
+  condition: succeeded()
+  pool:
+    vmImage: windows-latest
+  timeoutInMinutes: 240
+  steps:
+  - bash: git clone --bare --depth=1 --filter=blob:none --single-branch -b main https://github.com/git-for-windows/git-sdk-64
+    displayName: 'clone git-sdk-64'
+  - bash: git clone --depth=1 --single-branch -b main https://github.com/git-for-windows/build-extra
+    displayName: 'clone build-extra'
+  - bash: sh -x ./build-extra/please.sh create-sdk-artifact --sdk=git-sdk-64.git --out=git-sdk-64-minimal minimal-sdk
+    displayName: 'build git-sdk-64-minimal-sdk'
+  - bash: |
+      # Let Git ignore the SDK and the test-cache
+      printf "%s\n" /git-sdk-64-minimal/ /test-cache/ >>'.git/info/exclude'
+    displayName: 'Ignore untracked directories'
+  - bash: make NDEBUG=1 DEVELOPER=1 vcxproj
+    displayName: Generate Visual Studio Solution
+    env:
+      HOME: $(Build.SourcesDirectory)
+      MSYSTEM: MINGW64
+      DEVELOPER: 1
+      NO_PERL: 1
+      PATH: "$(Build.SourcesDirectory)\\git-sdk-64-minimal\\mingw64\\bin;$(Build.SourcesDirectory)\\git-sdk-64-minimal\\usr\\bin;C:\\Windows\\system32;C:\\Windows;C:\\Windows\\system32\\wbem"
+  - powershell: |
+      $urlbase = "https://dev.azure.com/git/git/_apis/build/builds"
+      $id = ((Invoke-WebRequest -UseBasicParsing "${urlbase}?definitions=9&statusFilter=completed&resultFilter=succeeded&`$top=1").content | ConvertFrom-JSON).value[0].id
+      $downloadUrl = ((Invoke-WebRequest -UseBasicParsing "${urlbase}/$id/artifacts").content | ConvertFrom-JSON).value[0].resource.downloadUrl
+      (New-Object Net.WebClient).DownloadFile($downloadUrl, "compat.zip")
+      Expand-Archive compat.zip -DestinationPath . -Force
+      Remove-Item compat.zip
+    displayName: 'Download vcpkg artifacts'
+  - task: MSBuild@1
+    inputs:
+      solution: git.sln
+      platform: x64
+      configuration: Release
+      maximumCpuCount: 4
+      msbuildArguments: /p:PlatformToolset=v142
+  - bash: |
+      ./compat/vcbuild/vcpkg_copy_dlls.bat release &&
+      mkdir -p artifacts &&
+      eval "$(make -n artifacts-tar INCLUDE_DLLS_IN_ARTIFACTS=YesPlease ARTIFACTS_DIRECTORY=artifacts | grep ^tar)"
+    displayName: Bundle artifact tar
+    env:
+      HOME: $(Build.SourcesDirectory)
+      MSYSTEM: MINGW64
+      DEVELOPER: 1
+      NO_PERL: 1
+      MSVC: 1
+      VCPKG_ROOT: $(Build.SourcesDirectory)\compat\vcbuild\vcpkg
+      PATH: "$(Build.SourcesDirectory)\\git-sdk-64-minimal\\mingw64\\bin;$(Build.SourcesDirectory)\\git-sdk-64-minimal\\usr\\bin;C:\\Windows\\system32;C:\\Windows;C:\\Windows\\system32\\wbem"
+  - powershell: |
+      $tag = (Invoke-WebRequest -UseBasicParsing "https://gitforwindows.org/latest-tag.txt").content
+      $version = (Invoke-WebRequest -UseBasicParsing "https://gitforwindows.org/latest-version.txt").content
+      $url = "https://github.com/git-for-windows/git/releases/download/${tag}/PortableGit-${version}-64-bit.7z.exe"
+      (New-Object Net.WebClient).DownloadFile($url,"PortableGit.exe")
+      & .\PortableGit.exe -y -oartifacts\PortableGit
+      # Wait until it is unpacked
+      while (-not @(Remove-Item -ErrorAction SilentlyContinue PortableGit.exe; $?)) { sleep 1 }
+    displayName: Download & extract portable Git
+  - task: PublishPipelineArtifact@0
+    displayName: 'Publish Pipeline Artifact: MSVC test artifacts'
+    inputs:
+      artifactName: 'vs-artifacts'
+      targetPath: '$(Build.SourcesDirectory)\artifacts'
+
+- job: vs_test
+  displayName: Visual Studio Test
+  dependsOn: vs_build
+  condition: succeeded()
+  pool:
+    vmImage: windows-latest
+  timeoutInMinutes: 240
+  strategy:
+    parallel: 10
+  steps:
+  - task: DownloadPipelineArtifact@0
+    displayName: 'Download Pipeline Artifact: VS test artifacts'
+    inputs:
+      artifactName: 'vs-artifacts'
+      targetPath: '$(Build.SourcesDirectory)'
+  - bash: |
+      test -f artifacts.tar.gz || {
+        echo No test artifacts found\; skipping >&2
+        exit 0
+      }
+      tar xf artifacts.tar.gz || exit 1
+
+      # Let Git ignore the SDK and the test-cache
+      printf '%s\n' /PortableGit/ /test-cache/ >>.git/info/exclude
+
+      cd t &&
+      PATH="$PWD/helper:$PATH" &&
+      test-tool.exe run-command testsuite --jobs=10 -V -x --write-junit-xml \
+              $(test-tool.exe path-utils slice-tests \
+                      $SYSTEM_JOBPOSITIONINPHASE $SYSTEM_TOTALJOBSINPHASE t[0-9]*.sh)
+    displayName: 'Test (parallel)'
+    env:
+      HOME: $(Build.SourcesDirectory)
+      MSYSTEM: MINGW64
+      NO_SVN_TESTS: 1
+      GIT_TEST_SKIP_REBASE_P: 1
+      PATH: "$(Build.SourcesDirectory)\\PortableGit\\mingw64\\bin;$(Build.SourcesDirectory)\\PortableGit\\usr\\bin;C:\\Windows\\system32;C:\\Windows;C:\\Windows\\system32\\wbem"
+  - task: PublishTestResults@2
+    displayName: 'Publish Test Results **/TEST-*.xml'
+    inputs:
+      mergeTestResults: true
+      testRunTitle: 'vs'
+      platform: Windows
+      publishRunAttachments: false
+    condition: succeededOrFailed()
+  - task: PublishBuildArtifacts@1
+    displayName: 'Publish trash directories of failed tests'
+    condition: failed()
+    inputs:
+      PathtoPublish: t/failed-test-artifacts
+      ArtifactName: failed-vs-test-artifacts
+
+- job: linux_clang
+  displayName: linux-clang
+  condition: succeeded()
+  pool:
+    vmImage: ubuntu-latest
+  steps:
+  - bash: |
+       export CC=clang || exit 1
+
+       ci/install-dependencies.sh || exit 1
+       ci/run-build-and-tests.sh || {
+           ci/print-test-failures.sh
+           exit 1
+       }
+    displayName: 'ci/run-build-and-tests.sh'
+  - task: PublishTestResults@2
+    displayName: 'Publish Test Results **/TEST-*.xml'
+    inputs:
+      mergeTestResults: true
+      testRunTitle: 'linux-clang'
+      platform: Linux
+      publishRunAttachments: false
+    condition: succeededOrFailed()
+  - task: PublishBuildArtifacts@1
+    displayName: 'Publish trash directories of failed tests'
+    condition: failed()
+    inputs:
+      PathtoPublish: t/failed-test-artifacts
+      ArtifactName: failed-test-artifacts
+
+- job: linux_gcc
+  displayName: linux-gcc
+  condition: succeeded()
+  pool:
+    vmImage: ubuntu-latest
+  steps:
+  - bash: |
+       ci/install-dependencies.sh || exit 1
+       ci/run-build-and-tests.sh || {
+           ci/print-test-failures.sh
+           exit 1
+       }
+    displayName: 'ci/run-build-and-tests.sh'
+  - task: PublishTestResults@2
+    displayName: 'Publish Test Results **/TEST-*.xml'
+    inputs:
+      mergeTestResults: true
+      testRunTitle: 'linux-gcc'
+      platform: Linux
+      publishRunAttachments: false
+    condition: succeededOrFailed()
+  - task: PublishBuildArtifacts@1
+    displayName: 'Publish trash directories of failed tests'
+    condition: failed()
+    inputs:
+      PathtoPublish: t/failed-test-artifacts
+      ArtifactName: failed-test-artifacts
+
+- job: osx_clang
+  displayName: osx-clang
+  condition: succeeded()
+  pool:
+    vmImage: macOS-latest
+  steps:
+  - bash: |
+       export CC=clang
+
+       ci/install-dependencies.sh || exit 1
+       ci/run-build-and-tests.sh || {
+           ci/print-test-failures.sh
+           exit 1
+       }
+    displayName: 'ci/run-build-and-tests.sh'
+  - task: PublishTestResults@2
+    displayName: 'Publish Test Results **/TEST-*.xml'
+    inputs:
+      mergeTestResults: true
+      testRunTitle: 'osx-clang'
+      platform: macOS
+      publishRunAttachments: false
+    condition: succeededOrFailed()
+  - task: PublishBuildArtifacts@1
+    displayName: 'Publish trash directories of failed tests'
+    condition: failed()
+    inputs:
+      PathtoPublish: t/failed-test-artifacts
+      ArtifactName: failed-test-artifacts
+
+- job: osx_gcc
+  displayName: osx-gcc
+  condition: succeeded()
+  pool:
+    vmImage: macOS-latest
+  steps:
+  - bash: |
+       ci/install-dependencies.sh || exit 1
+       ci/run-build-and-tests.sh || {
+           ci/print-test-failures.sh
+           exit 1
+       }
+    displayName: 'ci/run-build-and-tests.sh'
+  - task: PublishTestResults@2
+    displayName: 'Publish Test Results **/TEST-*.xml'
+    inputs:
+      mergeTestResults: true
+      testRunTitle: 'osx-gcc'
+      platform: macOS
+      publishRunAttachments: false
+    condition: succeededOrFailed()
+  - task: PublishBuildArtifacts@1
+    displayName: 'Publish trash directories of failed tests'
+    condition: failed()
+    inputs:
+      PathtoPublish: t/failed-test-artifacts
+      ArtifactName: failed-test-artifacts
+
+- job: linux32
+  displayName: Linux32
+  condition: succeeded()
+  pool:
+    vmImage: ubuntu-latest
+  steps:
+  - bash: |
+       res=0
+       sudo AGENT_OS="$AGENT_OS" BUILD_BUILDNUMBER="$BUILD_BUILDNUMBER" BUILD_REPOSITORY_URI="$BUILD_REPOSITORY_URI" BUILD_SOURCEBRANCH="$BUILD_SOURCEBRANCH" BUILD_SOURCEVERSION="$BUILD_SOURCEVERSION" SYSTEM_PHASENAME="$SYSTEM_PHASENAME" SYSTEM_TASKDEFINITIONSURI="$SYSTEM_TASKDEFINITIONSURI" SYSTEM_TEAMPROJECT="$SYSTEM_TEAMPROJECT" CC=$CC MAKEFLAGS="$MAKEFLAGS" jobname=linux32 bash -lxc ci/run-docker.sh || res=1
+
+       sudo chmod a+r t/out/TEST-*.xml
+       test ! -d t/failed-test-artifacts || sudo chmod a+r t/failed-test-artifacts
+
+       exit $res
+    displayName: 'jobname=linux32 ci/run-docker.sh'
+  - task: PublishTestResults@2
+    displayName: 'Publish Test Results **/TEST-*.xml'
+    inputs:
+      mergeTestResults: true
+      testRunTitle: 'linux32'
+      platform: Linux
+      publishRunAttachments: false
+    condition: succeededOrFailed()
+  - task: PublishBuildArtifacts@1
+    displayName: 'Publish trash directories of failed tests'
+    condition: failed()
+    inputs:
+      PathtoPublish: t/failed-test-artifacts
+      ArtifactName: failed-test-artifacts
+
+- job: static_analysis
+  displayName: StaticAnalysis
+  condition: succeeded()
+  pool:
+    vmImage: ubuntu-22.04
+  steps:
+  - bash: |
+       sudo apt-get update &&
+       sudo apt-get install -y coccinelle libcurl4-openssl-dev libssl-dev libexpat-dev gettext &&
+
+       export jobname=StaticAnalysis &&
+
+       ci/run-static-analysis.sh || exit 1
+    displayName: 'ci/run-static-analysis.sh'
+
+- job: documentation
+  displayName: Documentation
+  condition: succeeded()
+  pool:
+    vmImage: ubuntu-latest
+  steps:
+  - bash: |
+       sudo apt-get update &&
+       sudo apt-get install -y asciidoc xmlto asciidoctor docbook-xsl-ns &&
+
+       export ALREADY_HAVE_ASCIIDOCTOR=yes. &&
+       export jobname=Documentation &&
+
+       ci/test-documentation.sh || exit 1
+    displayName: 'ci/test-documentation.sh'
diff --git a/builtin.h b/builtin.h
index f7b166b33484d3..5f64730cf0273d 100644
--- a/builtin.h
+++ b/builtin.h
@@ -120,6 +120,7 @@ int cmd_am(int argc, const char **argv, const char *prefix, struct repository *r
 int cmd_annotate(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_apply(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_archive(int argc, const char **argv, const char *prefix, struct repository *repo);
+int cmd_backfill(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_bisect(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_blame(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_branch(int argc, const char **argv, const char *prefix, struct repository *repo);
@@ -230,6 +231,7 @@ int cmd_sparse_checkout(int argc, const char **argv, const char *prefix, struct
 int cmd_status(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_stash(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_stripspace(int argc, const char **argv, const char *prefix, struct repository *repo);
+int cmd_survey(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_submodule__helper(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_switch(int argc, const char **argv, const char *prefix, struct repository *repo);
 int cmd_symbolic_ref(int argc, const char **argv, const char *prefix, struct repository *repo);
diff --git a/builtin/backfill.c b/builtin/backfill.c
new file mode 100644
index 00000000000000..6fab33fccf4f7e
--- /dev/null
+++ b/builtin/backfill.c
@@ -0,0 +1,154 @@
+#define USE_THE_REPOSITORY_VARIABLE /* for core_apply_sparse_checkout */
+
+#include "builtin.h"
+#include "git-compat-util.h"
+#include "config.h"
+#include "parse-options.h"
+#include "repository.h"
+#include "commit.h"
+#include "dir.h"
+#include "environment.h"
+#include "hex.h"
+#include "tree.h"
+#include "tree-walk.h"
+#include "object.h"
+#include "object-store-ll.h"
+#include "oid-array.h"
+#include "oidset.h"
+#include "promisor-remote.h"
+#include "strmap.h"
+#include "string-list.h"
+#include "revision.h"
+#include "trace2.h"
+#include "progress.h"
+#include "packfile.h"
+#include "path-walk.h"
+
+static const char * const builtin_backfill_usage[] = {
+	N_("(EXPERIMENTAL) git backfill [--batch-size=<n>] [--[no-]sparse]"),
+	NULL
+};
+
+struct backfill_context {
+	struct repository *repo;
+	struct oid_array current_batch;
+	size_t batch_size;
+	int sparse;
+};
+
+static void clear_backfill_context(struct backfill_context *ctx)
+{
+	oid_array_clear(&ctx->current_batch);
+}
+
+static void download_batch(struct backfill_context *ctx)
+{
+	promisor_remote_get_direct(ctx->repo,
+				   ctx->current_batch.oid,
+				   ctx->current_batch.nr);
+	oid_array_clear(&ctx->current_batch);
+
+	/*
+	 * We likely have a new packfile. Add it to the packed list to
+	 * avoid possible duplicate downloads of the same objects.
+	 */
+	reprepare_packed_git(ctx->repo);
+}
+
+static int fill_missing_blobs(const char *path UNUSED,
+			      struct oid_array *list,
+			      enum object_type type,
+			      void *data)
+{
+	struct backfill_context *ctx = data;
+
+	if (type != OBJ_BLOB)
+		return 0;
+
+	for (size_t i = 0; i < list->nr; i++) {
+		off_t size = 0;
+		struct object_info info = OBJECT_INFO_INIT;
+		info.disk_sizep = &size;
+		if (oid_object_info_extended(ctx->repo,
+					     &list->oid[i],
+					     &info,
+					     OBJECT_INFO_FOR_PREFETCH) ||
+		    !size)
+			oid_array_append(&ctx->current_batch, &list->oid[i]);
+	}
+
+	if (ctx->current_batch.nr >= ctx->batch_size)
+		download_batch(ctx);
+
+	return 0;
+}
+
+static int do_backfill(struct backfill_context *ctx)
+{
+	struct rev_info revs;
+	struct path_walk_info info = PATH_WALK_INFO_INIT;
+	int ret;
+
+	if (ctx->sparse) {
+		CALLOC_ARRAY(info.pl, 1);
+		if (get_sparse_checkout_patterns(info.pl)) {
+			clear_pattern_list(info.pl);
+			free(info.pl);
+			return error(_("problem loading sparse-checkout"));
+		}
+	}
+
+	repo_init_revisions(ctx->repo, &revs, "");
+	handle_revision_arg("HEAD", &revs, 0, 0);
+
+	info.blobs = 1;
+	info.tags = info.commits = info.trees = 0;
+
+	info.revs = &revs;
+	info.path_fn = fill_missing_blobs;
+	info.path_fn_data = ctx;
+
+	ret = walk_objects_by_path(&info);
+
+	/* Download the objects that did not fill a batch. */
+	if (!ret)
+		download_batch(ctx);
+
+	clear_backfill_context(ctx);
+	release_revisions(&revs);
+	if (info.pl) {
+		clear_pattern_list(info.pl);
+		free(info.pl);
+	}
+	return ret;
+}
+
+int cmd_backfill(int argc, const char **argv, const char *prefix, struct repository *repo)
+{
+	struct backfill_context ctx = {
+		.repo = repo,
+		.current_batch = OID_ARRAY_INIT,
+		.batch_size = 50000,
+		.sparse = 0,
+	};
+	struct option options[] = {
+		OPT_INTEGER(0, "batch-size", &ctx.batch_size,
+			    N_("Minimun number of objects to request at a time")),
+		OPT_BOOL(0, "sparse", &ctx.sparse,
+			 N_("Restrict the missing objects to the current sparse-checkout")),
+		OPT_END(),
+	};
+
+	if (argc == 2 && !strcmp(argv[1], "-h"))
+		usage_with_options(builtin_backfill_usage, options);
+
+	argc = parse_options(argc, argv, prefix, options, builtin_backfill_usage,
+			     0);
+
+	repo_config(repo, git_default_config, NULL);
+
+	if (ctx.sparse < 0)
+		ctx.sparse = core_apply_sparse_checkout;
+
+	return do_backfill(&ctx);
+}
diff --git a/builtin/clean.c b/builtin/clean.c
index 053c94fc6bd12a..df39efcc50fbc4 100644
--- a/builtin/clean.c
+++ b/builtin/clean.c
@@ -40,6 +40,10 @@ static const char *msg_remove = N_("Removing %s\n");
 static const char *msg_would_remove = N_("Would remove %s\n");
 static const char *msg_skip_git_dir = N_("Skipping repository %s\n");
 static const char *msg_would_skip_git_dir = N_("Would skip repository %s\n");
+#ifndef CAN_UNLINK_MOUNT_POINTS
+static const char *msg_skip_mount_point = N_("Skipping mount point %s\n");
+static const char *msg_would_skip_mount_point = N_("Would skip mount point %s\n");
+#endif
 static const char *msg_warn_remove_failed = N_("failed to remove %s");
 static const char *msg_warn_lstat_failed = N_("could not lstat %s\n");
 static const char *msg_skip_cwd = N_("Refusing to remove current working directory\n");
@@ -184,6 +188,29 @@ static int remove_dirs(struct strbuf *path, const char *prefix, int force_flag,
 		goto out;
 	}
 
+	if (is_mount_point(path)) {
+#ifndef CAN_UNLINK_MOUNT_POINTS
+		if (!quiet) {
+			quote_path(path->buf, prefix, &quoted, 0);
+			printf(dry_run ?
+			       _(msg_would_skip_mount_point) :
+			       _(msg_skip_mount_point), quoted.buf);
+		}
+		*dir_gone = 0;
+#else
+		if (!dry_run && unlink(path->buf)) {
+			int saved_errno = errno;
+			quote_path(path->buf, prefix, &quoted, 0);
+			errno = saved_errno;
+			warning_errno(_(msg_warn_remove_failed), quoted.buf);
+			*dir_gone = 0;
+			ret = -1;
+		}
+#endif
+
+		goto out;
+	}
+
 	dir = opendir(path->buf);
 	if (!dir) {
 		/* an empty dir could be removed even if it is unreadble */
diff --git a/builtin/pack-objects.c b/builtin/pack-objects.c
index 1c3b8426515c42..4eeaf788b117f8 100644
--- a/builtin/pack-objects.c
+++ b/builtin/pack-objects.c
@@ -41,6 +41,9 @@
 #include "promisor-remote.h"
 #include "pack-mtimes.h"
 #include "parse-options.h"
+#include "blob.h"
+#include "tree.h"
+#include "path-walk.h"
 
 /*
  * Objects we are going to pack are collected in the `to_pack` structure.
@@ -217,6 +220,7 @@ static int delta_search_threads;
 static int pack_to_stdout;
 static int sparse;
 static int thin;
+static int path_walk = -1;
 static int num_preferred_base;
 static struct progress *progress_state;
 
@@ -269,6 +273,14 @@ struct configured_exclusion {
 static struct oidmap configured_exclusions;
 
 static struct oidset excluded_by_config;
+static int use_full_name_hash = -1;
+
+static inline uint32_t pack_name_hash_fn(const char *name)
+{
+	if (use_full_name_hash)
+		return pack_full_name_hash(name);
+	return pack_name_hash(name);
+}
 
 /*
  * stats
@@ -1686,7 +1698,7 @@ static int add_object_entry(const struct object_id *oid, enum object_type type,
 		return 0;
 	}
 
-	create_object_entry(oid, type, pack_name_hash(name),
+	create_object_entry(oid, type, pack_name_hash_fn(name),
 			    exclude, name && no_try_delta(name),
 			    found_pack, found_offset);
 	return 1;
@@ -1900,7 +1912,7 @@ static void add_preferred_base_object(const char *name)
 {
 	struct pbase_tree *it;
 	size_t cmplen;
-	unsigned hash = pack_name_hash(name);
+	unsigned hash = pack_name_hash_fn(name);
 
 	if (!num_preferred_base || check_pbase_path(hash))
 		return;
@@ -2913,6 +2925,7 @@ static void find_deltas(struct object_entry **list, unsigned *list_size,
 struct thread_params {
 	pthread_t thread;
 	struct object_entry **list;
+	struct packing_region *regions;
 	unsigned list_size;
 	unsigned remaining;
 	int window;
@@ -3155,6 +3168,234 @@ static int add_ref_tag(const char *tag UNUSED, const char *referent UNUSED, cons
 	return 0;
 }
 
+static int should_attempt_deltas(struct object_entry *entry)
+{
+	if (DELTA(entry))
+		return 0;
+
+	if (!entry->type_valid ||
+	    oe_size_less_than(&to_pack, entry, 50))
+		return 0;
+
+	if (entry->no_try_delta)
+		return 0;
+
+	if (!entry->preferred_base) {
+		if (oe_type(entry) < 0)
+			die(_("unable to get type of object %s"),
+				oid_to_hex(&entry->idx.oid));
+	} else if (oe_type(entry) < 0) {
+		/*
+		 * This object is not found, but we
+		 * don't have to include it anyway.
+		 */
+		return 0;
+	}
+
+	return 1;
+}
+
+static void find_deltas_for_region(struct object_entry *list UNUSED,
+				   struct packing_region *region,
+				   unsigned int *processed)
+{
+	struct object_entry **delta_list;
+	uint32_t delta_list_nr = 0;
+
+	ALLOC_ARRAY(delta_list, region->nr);
+	for (uint32_t i = 0; i < region->nr; i++) {
+		struct object_entry *entry = to_pack.objects + region->start + i;
+		if (should_attempt_deltas(entry))
+			delta_list[delta_list_nr++] = entry;
+	}
+
+	QSORT(delta_list, delta_list_nr, type_size_sort);
+	find_deltas(delta_list, &delta_list_nr, window, depth, processed);
+	free(delta_list);
+}
+
+static void find_deltas_by_region(struct object_entry *list,
+				  struct packing_region *regions,
+				  uint32_t start, uint32_t nr)
+{
+	unsigned int processed = 0;
+	uint32_t progress_nr;
+
+	if (!nr)
+		return;
+
+	progress_nr = regions[nr - 1].start + regions[nr - 1].nr;
+
+	if (progress)
+		progress_state = start_progress(_("Compressing objects by path"),
+						progress_nr);
+
+	while (nr--)
+		find_deltas_for_region(list,
+				       &regions[start++],
+				       &processed);
+
+	display_progress(progress_state, progress_nr);
+	stop_progress(&progress_state);
+}
+
+static void *threaded_find_deltas_by_path(void *arg)
+{
+	struct thread_params *me = arg;
+
+	progress_lock();
+	while (me->remaining) {
+		while (me->remaining) {
+			progress_unlock();
+			find_deltas_for_region(to_pack.objects,
+					       me->regions,
+					       me->processed);
+			progress_lock();
+			me->remaining--;
+			me->regions++;
+		}
+
+		me->working = 0;
+		pthread_cond_signal(&progress_cond);
+		progress_unlock();
+
+		/*
+		 * We must not set ->data_ready before we wait on the
+		 * condition because the main thread may have set it to 1
+		 * before we get here. In order to be sure that new
+		 * work is available if we see 1 in ->data_ready, it
+		 * was initialized to 0 before this thread was spawned
+		 * and we reset it to 0 right away.
+		 */
+		pthread_mutex_lock(&me->mutex);
+		while (!me->data_ready)
+			pthread_cond_wait(&me->cond, &me->mutex);
+		me->data_ready = 0;
+		pthread_mutex_unlock(&me->mutex);
+
+		progress_lock();
+	}
+	progress_unlock();
+	/* leave ->working 1 so that this doesn't get more work assigned */
+	return NULL;
+}
+
+static void ll_find_deltas_by_region(struct object_entry *list,
+				     struct packing_region *regions,
+				     uint32_t start, uint32_t nr)
+{
+	struct thread_params *p;
+	int i, ret, active_threads = 0;
+	unsigned int processed = 0;
+	uint32_t progress_nr;
+	init_threaded_search();
+
+	if (!nr)
+		return;
+
+	progress_nr =  regions[nr - 1].start + regions[nr - 1].nr;
+	if (delta_search_threads <= 1) {
+		find_deltas_by_region(list, regions, start, nr);
+		cleanup_threaded_search();
+		return;
+	}
+
+	if (progress > pack_to_stdout)
+		fprintf_ln(stderr, _("Path-based delta compression using up to %d threads"),
+			   delta_search_threads);
+	CALLOC_ARRAY(p, delta_search_threads);
+
+	if (progress)
+		progress_state = start_progress(_("Compressing objects by path"),
+						progress_nr);
+	/* Partition the work amongst work threads. */
+	for (i = 0; i < delta_search_threads; i++) {
+		unsigned sub_size = nr / (delta_search_threads - i);
+
+		p[i].window = window;
+		p[i].depth = depth;
+		p[i].processed = &processed;
+		p[i].working = 1;
+		p[i].data_ready = 0;
+
+		p[i].regions = regions;
+		p[i].list_size = sub_size;
+		p[i].remaining = sub_size;
+
+		regions += sub_size;
+		nr -= sub_size;
+	}
+
+	/* Start work threads. */
+	for (i = 0; i < delta_search_threads; i++) {
+		if (!p[i].list_size)
+			continue;
+		pthread_mutex_init(&p[i].mutex, NULL);
+		pthread_cond_init(&p[i].cond, NULL);
+		ret = pthread_create(&p[i].thread, NULL,
+				     threaded_find_deltas_by_path, &p[i]);
+		if (ret)
+			die(_("unable to create thread: %s"), strerror(ret));
+		active_threads++;
+	}
+
+	/*
+	 * Now let's wait for work completion.  Each time a thread is done
+	 * with its work, we steal half of the remaining work from the
+	 * thread with the largest number of unprocessed objects and give
+	 * it to that newly idle thread.  This ensure good load balancing
+	 * until the remaining object list segments are simply too short
+	 * to be worth splitting anymore.
+	 */
+	while (active_threads) {
+		struct thread_params *target = NULL;
+		struct thread_params *victim = NULL;
+		unsigned sub_size = 0;
+
+		progress_lock();
+		for (;;) {
+			for (i = 0; !target && i < delta_search_threads; i++)
+				if (!p[i].working)
+					target = &p[i];
+			if (target)
+				break;
+			pthread_cond_wait(&progress_cond, &progress_mutex);
+		}
+
+		for (i = 0; i < delta_search_threads; i++)
+			if (p[i].remaining > 2*window &&
+			    (!victim || victim->remaining < p[i].remaining))
+				victim = &p[i];
+		if (victim) {
+			sub_size = victim->remaining / 2;
+			target->regions = victim->regions + victim->remaining - sub_size;
+			victim->list_size -= sub_size;
+			victim->remaining -= sub_size;
+		}
+		target->list_size = sub_size;
+		target->remaining = sub_size;
+		target->working = 1;
+		progress_unlock();
+
+		pthread_mutex_lock(&target->mutex);
+		target->data_ready = 1;
+		pthread_cond_signal(&target->cond);
+		pthread_mutex_unlock(&target->mutex);
+
+		if (!sub_size) {
+			pthread_join(target->thread, NULL);
+			pthread_cond_destroy(&target->cond);
+			pthread_mutex_destroy(&target->mutex);
+			active_threads--;
+		}
+	}
+	cleanup_threaded_search();
+	free(p);
+
+	display_progress(progress_state, progress_nr);
+	stop_progress(&progress_state);
+}
+
 static void prepare_pack(int window, int depth)
 {
 	struct object_entry **delta_list;
@@ -3179,39 +3420,21 @@ static void prepare_pack(int window, int depth)
 	if (!to_pack.nr_objects || !window || !depth)
 		return;
 
+	if (path_walk)
+		ll_find_deltas_by_region(to_pack.objects, to_pack.regions,
+					 0, to_pack.nr_regions);
+
 	ALLOC_ARRAY(delta_list, to_pack.nr_objects);
 	nr_deltas = n = 0;
 
 	for (i = 0; i < to_pack.nr_objects; i++) {
 		struct object_entry *entry = to_pack.objects + i;
 
-		if (DELTA(entry))
-			/* This happens if we decided to reuse existing
-			 * delta from a pack.  "reuse_delta &&" is implied.
-			 */
-			continue;
-
-		if (!entry->type_valid ||
-		    oe_size_less_than(&to_pack, entry, 50))
+		if (!should_attempt_deltas(entry))
 			continue;
 
-		if (entry->no_try_delta)
-			continue;
-
-		if (!entry->preferred_base) {
+		if (!entry->preferred_base)
 			nr_deltas++;
-			if (oe_type(entry) < 0)
-				die(_("unable to get type of object %s"),
-				    oid_to_hex(&entry->idx.oid));
-		} else {
-			if (oe_type(entry) < 0) {
-				/*
-				 * This object is not found, but we
-				 * don't have to include it anyway.
-				 */
-				continue;
-			}
-		}
 
 		delta_list[n++] = entry;
 	}
@@ -3410,7 +3633,7 @@ static void show_object_pack_hint(struct object *object, const char *name,
 	 * here using a now in order to perhaps improve the delta selection
 	 * process.
 	 */
-	oe->hash = pack_name_hash(name);
+	oe->hash = pack_name_hash_fn(name);
 	oe->no_try_delta = name && no_try_delta(name);
 
 	stdin_packs_hints_nr++;
@@ -3560,7 +3783,7 @@ static void add_cruft_object_entry(const struct object_id *oid, enum object_type
 	entry = packlist_find(&to_pack, oid);
 	if (entry) {
 		if (name) {
-			entry->hash = pack_name_hash(name);
+			entry->hash = pack_name_hash_fn(name);
 			entry->no_try_delta = no_try_delta(name);
 		}
 	} else {
@@ -3583,7 +3806,7 @@ static void add_cruft_object_entry(const struct object_id *oid, enum object_type
 			return;
 		}
 
-		entry = create_object_entry(oid, type, pack_name_hash(name),
+		entry = create_object_entry(oid, type, pack_name_hash_fn(name),
 					    0, name && no_try_delta(name),
 					    pack, offset);
 	}
@@ -4129,6 +4352,88 @@ static void mark_bitmap_preferred_tips(void)
 	}
 }
 
+static inline int is_oid_interesting(struct repository *repo,
+				     struct object_id *oid)
+{
+	struct object *o = lookup_object(repo, oid);
+	return o && !(o->flags & UNINTERESTING);
+}
+
+static int add_objects_by_path(const char *path,
+			       struct oid_array *oids,
+			       enum object_type type,
+			       void *data)
+{
+	size_t oe_start = to_pack.nr_objects;
+	size_t oe_end;
+	unsigned int *processed = data;
+
+	/*
+	 * First, add all objects to the packing data, including the ones
+	 * marked UNINTERESTING (translated to 'exclude') as they can be
+	 * used as delta bases.
+	 */
+	for (size_t i = 0; i < oids->nr; i++) {
+		int exclude;
+		struct object_info oi = OBJECT_INFO_INIT;
+		struct object_id *oid = &oids->oid[i];
+
+		/* Skip objects that do not exist locally. */
+		if ((exclude_promisor_objects || arg_missing_action != MA_ERROR) &&
+		    oid_object_info_extended(the_repository, oid, &oi,
+					     OBJECT_INFO_FOR_PREFETCH) < 0)
+			continue;
+
+		exclude = !is_oid_interesting(the_repository, oid);
+
+		if (exclude && !thin)
+			continue;
+
+		add_object_entry(oid, type, path, exclude);
+	}
+
+	oe_end = to_pack.nr_objects;
+
+	/* We can skip delta calculations if it is a no-op. */
+	if (oe_end == oe_start || !window)
+		return 0;
+
+	ALLOC_GROW(to_pack.regions,
+		   to_pack.nr_regions + 1,
+		   to_pack.nr_regions_alloc);
+
+	to_pack.regions[to_pack.nr_regions].start = oe_start;
+	to_pack.regions[to_pack.nr_regions].nr = oe_end - oe_start;
+	to_pack.nr_regions++;
+
+	*processed += oids->nr;
+	display_progress(progress_state, *processed);
+
+	return 0;
+}
+
+static void get_object_list_path_walk(struct rev_info *revs)
+{
+	struct path_walk_info info = PATH_WALK_INFO_INIT;
+	unsigned int processed = 0;
+
+	info.revs = revs;
+	info.path_fn = add_objects_by_path;
+	info.path_fn_data = &processed;
+	revs->tag_objects = 1;
+
+	/*
+	 * Allow the --[no-]sparse option to be interesting here, if only
+	 * for testing purposes. Paths with no interesting objects will not
+	 * contribute to the resulting pack, but only create noisy preferred
+	 * base objects.
+	 */
+	info.prune_all_uninteresting = sparse;
+
+	if (walk_objects_by_path(&info))
+		die(_("failed to pack objects via path-walk"));
+}
+
 static void get_object_list(struct rev_info *revs, int ac, const char **av)
 {
 	struct setup_revision_opt s_r_opt = {
@@ -4175,7 +4480,7 @@ static void get_object_list(struct rev_info *revs, int ac, const char **av)
 
 	warn_on_object_refname_ambiguity = save_warning;
 
-	if (use_bitmap_index && !get_object_list_from_bitmap(revs))
+	if (use_bitmap_index && !path_walk && !get_object_list_from_bitmap(revs))
 		return;
 
 	if (use_delta_islands)
@@ -4184,15 +4489,19 @@ static void get_object_list(struct rev_info *revs, int ac, const char **av)
 	if (write_bitmap_index)
 		mark_bitmap_preferred_tips();
 
-	if (prepare_revision_walk(revs))
-		die(_("revision walk setup failed"));
-	mark_edges_uninteresting(revs, show_edge, sparse);
-
 	if (!fn_show_object)
 		fn_show_object = show_object;
-	traverse_commit_list(revs,
-			     show_commit, fn_show_object,
-			     NULL);
+
+	if (path_walk) {
+		get_object_list_path_walk(revs);
+	} else {
+		if (prepare_revision_walk(revs))
+			die(_("revision walk setup failed"));
+		mark_edges_uninteresting(revs, show_edge, sparse);
+		traverse_commit_list(revs,
+				show_commit, fn_show_object,
+				NULL);
+	}
 
 	if (unpack_unreachable_expiration) {
 		revs->ignore_missing_links = 1;
@@ -4402,6 +4711,8 @@ int cmd_pack_objects(int argc,
 			 N_("use the sparse reachability algorithm")),
 		OPT_BOOL(0, "thin", &thin,
 			 N_("create thin packs")),
+		OPT_BOOL(0, "path-walk", &path_walk,
+			 N_("use the path-walk API to walk objects when possible")),
 		OPT_BOOL(0, "shallow", &shallow,
 			 N_("create packs suitable for shallow fetches")),
 		OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep_on_disk,
@@ -4435,6 +4746,8 @@ int cmd_pack_objects(int argc,
 		OPT_STRING_LIST(0, "uri-protocol", &uri_protocols,
 				N_("protocol"),
 				N_("exclude any configured uploadpack.blobpackfileuri with this protocol")),
+		OPT_BOOL(0, "full-name-hash", &use_full_name_hash,
+			 N_("(EXPERIMENTAL!) optimize delta compression across identical path names over time")),
 		OPT_END(),
 	};
 
@@ -4469,6 +4782,17 @@ int cmd_pack_objects(int argc,
 	if (pack_to_stdout != !base_name || argc)
 		usage_with_options(pack_usage, pack_objects_options);
 
+	if (path_walk < 0) {
+		if (use_bitmap_index > 0 ||
+		    !use_internal_rev_list)
+			path_walk = 0;
+		else if (the_repository->gitdir &&
+			 the_repository->settings.pack_use_path_walk)
+			path_walk = 1;
+		else
+			path_walk = git_env_bool("GIT_TEST_PACK_PATH_WALK", 0);
+	}
+
 	if (depth < 0)
 		depth = 0;
 	if (depth >= (1 << OE_DEPTH_BITS)) {
@@ -4485,7 +4809,27 @@ int cmd_pack_objects(int argc,
 		window = 0;
 
 	strvec_push(&rp, "pack-objects");
-	if (thin) {
+
+	if (path_walk && filter_options.choice) {
+		warning(_("cannot use --filter with --path-walk"));
+		path_walk = 0;
+	}
+	if (path_walk && use_delta_islands) {
+		warning(_("cannot use delta islands with --path-walk"));
+		path_walk = 0;
+	}
+	if (path_walk && shallow) {
+		warning(_("cannot use --shallow with --path-walk"));
+		path_walk = 0;
+	}
+	if (path_walk) {
+		strvec_push(&rp, "--boundary");
+		 /*
+		  * We must disable the bitmaps because we are removing
+		  * the --objects / --objects-edge[-aggressive] options.
+		  */
+		use_bitmap_index = 0;
+	} else if (thin) {
 		use_internal_rev_list = 1;
 		strvec_push(&rp, shallow
 				? "--objects-edge-aggressive"
@@ -4590,6 +4934,11 @@ int cmd_pack_objects(int argc,
 	if (pack_to_stdout || !rev_list_all)
 		write_bitmap_index = 0;
 
+	if (write_bitmap_index && use_full_name_hash > 0)
+		die(_("currently, the --full-name-hash option is incompatible with --write-bitmap-index"));
+	if (use_full_name_hash < 0)
+		use_full_name_hash = git_env_bool("GIT_TEST_FULL_NAME_HASH", 0);
+
 	if (use_delta_islands)
 		strvec_push(&rp, "--topo-order");
 
diff --git a/builtin/repack.c b/builtin/repack.c
index 0c6dad7df47a16..01db1b8f52913a 100644
--- a/builtin/repack.c
+++ b/builtin/repack.c
@@ -41,7 +41,9 @@ static int run_update_server_info = 1;
 static char *packdir, *packtmp_name, *packtmp;
 
 static const char *const git_repack_usage[] = {
-	N_("git repack [<options>]"),
+	N_("git repack [-a] [-A] [-d] [-f] [-F] [-l] [-n] [-q] [-b] [-m]\n"
+	   "[--window=<n>] [--depth=<n>] [--threads=<n>] [--keep-pack=<pack-name>]\n"
+	   "[--write-midx] [--full-name-hash] [--path-walk]"),
 	NULL
 };
 
@@ -60,6 +62,8 @@ struct pack_objects_args {
 	int no_reuse_object;
 	int quiet;
 	int local;
+	int full_name_hash;
+	int path_walk;
 	struct list_objects_filter_options filter_options;
 };
 
@@ -308,6 +312,10 @@ static void prepare_pack_objects(struct child_process *cmd,
 		strvec_pushf(&cmd->args, "--no-reuse-delta");
 	if (args->no_reuse_object)
 		strvec_pushf(&cmd->args, "--no-reuse-object");
+	if (args->full_name_hash)
+		strvec_pushf(&cmd->args, "--full-name-hash");
+	if (args->path_walk)
+		strvec_pushf(&cmd->args, "--path-walk");
 	if (args->local)
 		strvec_push(&cmd->args,  "--local");
 	if (args->quiet)
@@ -1205,6 +1213,10 @@ int cmd_repack(int argc,
 				N_("pass --no-reuse-delta to git-pack-objects")),
 		OPT_BOOL('F', NULL, &po_args.no_reuse_object,
 				N_("pass --no-reuse-object to git-pack-objects")),
+		OPT_BOOL(0, "full-name-hash", &po_args.full_name_hash,
+				N_("(EXPERIMENTAL!) pass --full-name-hash to git-pack-objects")),
+		OPT_BOOL(0, "path-walk", &po_args.path_walk,
+				N_("(EXPERIMENTAL!) pass --path-walk to git-pack-objects")),
 		OPT_NEGBIT('n', NULL, &run_update_server_info,
 				N_("do not run git-update-server-info"), 1),
 		OPT__QUIET(&po_args.quiet, N_("be quiet")),
diff --git a/builtin/survey.c b/builtin/survey.c
new file mode 100644
index 00000000000000..a86b728d6a2671
--- /dev/null
+++ b/builtin/survey.c
@@ -0,0 +1,930 @@
+#define USE_THE_REPOSITORY_VARIABLE
+
+#include "builtin.h"
+#include "config.h"
+#include "environment.h"
+#include "hex.h"
+#include "object.h"
+#include "object-name.h"
+#include "object-store-ll.h"
+#include "parse-options.h"
+#include "path-walk.h"
+#include "progress.h"
+#include "ref-filter.h"
+#include "refs.h"
+#include "revision.h"
+#include "strbuf.h"
+#include "strvec.h"
+#include "tag.h"
+#include "trace2.h"
+#include "color.h"
+
+static const char * const survey_usage[] = {
+	N_("(EXPERIMENTAL!) git survey <options>"),
+	NULL,
+};
+
+struct survey_refs_wanted {
+	int want_all_refs; /* special override */
+
+	int want_branches;
+	int want_tags;
+	int want_remotes;
+	int want_detached;
+	int want_other; /* see FILTER_REFS_OTHERS -- refs/notes/, refs/stash/ */
+};
+
+static struct survey_refs_wanted default_ref_options = {
+	.want_all_refs = 1,
+};
+
+struct survey_opts {
+	int verbose;
+	int show_progress;
+	int top_nr;
+	struct survey_refs_wanted refs;
+};
+
+struct survey_report_ref_summary {
+	size_t refs_nr;
+	size_t branches_nr;
+	size_t remote_refs_nr;
+	size_t tags_nr;
+	size_t tags_annotated_nr;
+	size_t others_nr;
+	size_t unknown_nr;
+};
+
+struct survey_report_object_summary {
+	size_t commits_nr;
+	size_t tags_nr;
+	size_t trees_nr;
+	size_t blobs_nr;
+};
+
+/**
+ * For some category given by 'label', count the number of objects
+ * that match that label along with the on-disk size and the size
+ * after decompressing (both with delta bases and zlib).
+ */
+struct survey_report_object_size_summary {
+	char *label;
+	size_t nr;
+	size_t disk_size;
+	size_t inflated_size;
+	size_t num_missing;
+};
+
+typedef int (*survey_top_cmp)(void *v1, void *v2);
+
+static int cmp_by_nr(void *v1, void *v2)
+{
+	struct survey_report_object_size_summary *s1 = v1;
+	struct survey_report_object_size_summary *s2 = v2;
+
+	if (s1->nr < s2->nr)
+		return -1;
+	if (s1->nr > s2->nr)
+		return 1;
+	return 0;
+}
+
+static int cmp_by_disk_size(void *v1, void *v2)
+{
+	struct survey_report_object_size_summary *s1 = v1;
+	struct survey_report_object_size_summary *s2 = v2;
+
+	if (s1->disk_size < s2->disk_size)
+		return -1;
+	if (s1->disk_size > s2->disk_size)
+		return 1;
+	return 0;
+}
+
+static int cmp_by_inflated_size(void *v1, void *v2)
+{
+	struct survey_report_object_size_summary *s1 = v1;
+	struct survey_report_object_size_summary *s2 = v2;
+
+	if (s1->inflated_size < s2->inflated_size)
+		return -1;
+	if (s1->inflated_size > s2->inflated_size)
+		return 1;
+	return 0;
+}
+
+/**
+ * Store a list of "top" categories by some sorting function. When
+ * inserting a new category, reorder the list and free the one that
+ * got ejected (if any).
+ */
+struct survey_report_top_table {
+	const char *name;
+	survey_top_cmp cmp_fn;
+	size_t nr;
+	size_t alloc;
+
+	/**
+	 * 'data' stores an array of structs and must be cast into
+	 * the proper array type before evaluating an index.
+	 */
+	void *data;
+};
+
+static void init_top_sizes(struct survey_report_top_table *top,
+			   size_t limit, const char *name,
+			   survey_top_cmp cmp)
+{
+	struct survey_report_object_size_summary *sz_array;
+
+	top->name = name;
+	top->cmp_fn = cmp;
+	top->alloc = limit;
+	top->nr = 0;
+
+	CALLOC_ARRAY(sz_array, limit);
+	top->data = sz_array;
+}
+
+MAYBE_UNUSED
+static void clear_top_sizes(struct survey_report_top_table *top)
+{
+	struct survey_report_object_size_summary *sz_array = top->data;
+
+	for (size_t i = 0; i < top->nr; i++)
+		free(sz_array[i].label);
+	free(sz_array);
+}
+
+static void maybe_insert_into_top_size(struct survey_report_top_table *top,
+				       struct survey_report_object_size_summary *summary)
+{
+	struct survey_report_object_size_summary *sz_array = top->data;
+	size_t pos = top->nr;
+
+	/* Compare against list from the bottom. */
+	while (pos > 0 && top->cmp_fn(&sz_array[pos - 1], summary) < 0)
+		pos--;
+
+	/* Not big enough! */
+	if (pos >= top->alloc)
+		return;
+
+	/* We need to shift the data. */
+	if (top->nr == top->alloc)
+		free(sz_array[top->nr - 1].label);
+	else
+		top->nr++;
+
+	for (size_t i = top->nr - 1; i > pos; i--)
+		memcpy(&sz_array[i], &sz_array[i - 1], sizeof(*sz_array));
+
+	memcpy(&sz_array[pos], summary, sizeof(*summary));
+	sz_array[pos].label = xstrdup(summary->label);
+}
+
+/**
+ * This struct contains all of the information that needs to be printed
+ * at the end of the exploration of the repository and its references.
+ */
+struct survey_report {
+	struct survey_report_ref_summary refs;
+	struct survey_report_object_summary reachable_objects;
+
+	struct survey_report_object_size_summary *by_type;
+
+	struct survey_report_top_table *top_paths_by_count;
+	struct survey_report_top_table *top_paths_by_disk;
+	struct survey_report_top_table *top_paths_by_inflate;
+};
+
+#define REPORT_TYPE_COMMIT 0
+#define REPORT_TYPE_TREE 1
+#define REPORT_TYPE_BLOB 2
+#define REPORT_TYPE_TAG 3
+#define REPORT_TYPE_COUNT 4
+
+struct survey_context {
+	struct repository *repo;
+
+	/* Options that control what is done. */
+	struct survey_opts opts;
+
+	/* Info for output only. */
+	struct survey_report report;
+
+	/*
+	 * The rest of the members are about enabling the activity
+	 * of the 'git survey' command, including ref listings, object
+	 * pointers, and progress.
+	 */
+
+	struct progress *progress;
+	size_t progress_nr;
+	size_t progress_total;
+
+	struct strvec refs;
+	struct ref_array ref_array;
+};
+
+static void clear_survey_context(struct survey_context *ctx)
+{
+	ref_array_clear(&ctx->ref_array);
+	strvec_clear(&ctx->refs);
+}
+
+struct survey_table {
+	const char *table_name;
+	struct strvec header;
+	struct strvec *rows;
+	size_t rows_nr;
+	size_t rows_alloc;
+};
+
+#define SURVEY_TABLE_INIT {	\
+	.header = STRVEC_INIT,	\
+}
+
+static void clear_table(struct survey_table *table)
+{
+	strvec_clear(&table->header);
+	for (size_t i = 0; i < table->rows_nr; i++)
+		strvec_clear(&table->rows[i]);
+	free(table->rows);
+}
+
+static void insert_table_rowv(struct survey_table *table, ...)
+{
+	va_list ap;
+	char *arg;
+	ALLOC_GROW(table->rows, table->rows_nr + 1, table->rows_alloc);
+
+	memset(&table->rows[table->rows_nr], 0, sizeof(struct strvec));
+
+	va_start(ap, table);
+	while ((arg = va_arg(ap, char *)))
+		strvec_push(&table->rows[table->rows_nr], arg);
+	va_end(ap);
+
+	table->rows_nr++;
+}
+
+#define SECTION_SEGMENT "========================================"
+#define SECTION_SEGMENT_LEN 40
+static const char *section_line = SECTION_SEGMENT
+				  SECTION_SEGMENT
+				  SECTION_SEGMENT
+				  SECTION_SEGMENT;
+static const size_t section_len = 4 * SECTION_SEGMENT_LEN;
+
+static void print_table_title(const char *name, size_t *widths, size_t nr)
+{
+	size_t width = 3 * (nr - 1);
+	size_t min_width = strlen(name);
+
+	for (size_t i = 0; i < nr; i++)
+		width += widths[i];
+
+	if (width < min_width)
+		width = min_width;
+
+	if (width > section_len)
+		width = section_len;
+
+	printf("\n%s\n%.*s\n", name, (int)width, section_line);
+}
+
+static void print_row_plaintext(struct strvec *row, size_t *widths)
+{
+	static struct strbuf line = STRBUF_INIT;
+	strbuf_setlen(&line, 0);
+
+	for (size_t i = 0; i < row->nr; i++) {
+		const char *str = row->v[i];
+		size_t len = strlen(str);
+		if (i)
+			strbuf_add(&line, " | ", 3);
+		strbuf_addchars(&line, ' ', widths[i] - len);
+		strbuf_add(&line, str, len);
+	}
+	printf("%s\n", line.buf);
+}
+
+static void print_divider_plaintext(size_t *widths, size_t nr)
+{
+	static struct strbuf line = STRBUF_INIT;
+	strbuf_setlen(&line, 0);
+
+	for (size_t i = 0; i < nr; i++) {
+		if (i)
+			strbuf_add(&line, "-+-", 3);
+		strbuf_addchars(&line, '-', widths[i]);
+	}
+	printf("%s\n", line.buf);
+}
+
+static void print_table_plaintext(struct survey_table *table)
+{
+	size_t *column_widths;
+	size_t columns_nr = table->header.nr;
+	CALLOC_ARRAY(column_widths, columns_nr);
+
+	for (size_t i = 0; i < columns_nr; i++) {
+		column_widths[i] = strlen(table->header.v[i]);
+
+		for (size_t j = 0; j < table->rows_nr; j++) {
+			size_t rowlen = strlen(table->rows[j].v[i]);
+			if (column_widths[i] < rowlen)
+				column_widths[i] = rowlen;
+		}
+	}
+
+	print_table_title(table->table_name, column_widths, columns_nr);
+	print_row_plaintext(&table->header, column_widths);
+	print_divider_plaintext(column_widths, columns_nr);
+
+	for (size_t j = 0; j < table->rows_nr; j++)
+		print_row_plaintext(&table->rows[j], column_widths);
+
+	free(column_widths);
+}
+
+static void survey_report_plaintext_refs(struct survey_context *ctx)
+{
+	struct survey_report_ref_summary *refs = &ctx->report.refs;
+	struct survey_table table = SURVEY_TABLE_INIT;
+
+	table.table_name = _("REFERENCES SUMMARY");
+
+	strvec_push(&table.header, _("Ref Type"));
+	strvec_push(&table.header, _("Count"));
+
+	if (ctx->opts.refs.want_all_refs || ctx->opts.refs.want_branches) {
+		char *fmt = xstrfmt("%"PRIuMAX"", (uintmax_t)refs->branches_nr);
+		insert_table_rowv(&table, _("Branches"), fmt, NULL);
+		free(fmt);
+	}
+
+	if (ctx->opts.refs.want_all_refs || ctx->opts.refs.want_remotes) {
+		char *fmt = xstrfmt("%"PRIuMAX"", (uintmax_t)refs->remote_refs_nr);
+		insert_table_rowv(&table, _("Remote refs"), fmt, NULL);
+		free(fmt);
+	}
+
+	if (ctx->opts.refs.want_all_refs || ctx->opts.refs.want_tags) {
+		char *fmt = xstrfmt("%"PRIuMAX"", (uintmax_t)refs->tags_nr);
+		insert_table_rowv(&table, _("Tags (all)"), fmt, NULL);
+		free(fmt);
+		fmt = xstrfmt("%"PRIuMAX"", (uintmax_t)refs->tags_annotated_nr);
+		insert_table_rowv(&table, _("Tags (annotated)"), fmt, NULL);
+		free(fmt);
+	}
+
+	print_table_plaintext(&table);
+	clear_table(&table);
+}
+
+static void survey_report_plaintext_reachable_object_summary(struct survey_context *ctx)
+{
+	struct survey_report_object_summary *objs = &ctx->report.reachable_objects;
+	struct survey_table table = SURVEY_TABLE_INIT;
+	char *fmt;
+
+	table.table_name = _("REACHABLE OBJECT SUMMARY");
+
+	strvec_push(&table.header, _("Object Type"));
+	strvec_push(&table.header, _("Count"));
+
+	fmt = xstrfmt("%"PRIuMAX"", (uintmax_t)objs->tags_nr);
+	insert_table_rowv(&table, _("Tags"), fmt, NULL);
+	free(fmt);
+
+	fmt = xstrfmt("%"PRIuMAX"", (uintmax_t)objs->commits_nr);
+	insert_table_rowv(&table, _("Commits"), fmt, NULL);
+	free(fmt);
+
+	fmt = xstrfmt("%"PRIuMAX"", (uintmax_t)objs->trees_nr);
+	insert_table_rowv(&table, _("Trees"), fmt, NULL);
+	free(fmt);
+
+	fmt = xstrfmt("%"PRIuMAX"", (uintmax_t)objs->blobs_nr);
+	insert_table_rowv(&table, _("Blobs"), fmt, NULL);
+	free(fmt);
+
+	print_table_plaintext(&table);
+	clear_table(&table);
+}
+
+static void survey_report_object_sizes(const char *title,
+				       const char *categories,
+				       struct survey_report_object_size_summary *summary,
+				       size_t summary_nr)
+{
+	struct survey_table table = SURVEY_TABLE_INIT;
+	table.table_name = title;
+
+	strvec_push(&table.header, categories);
+	strvec_push(&table.header, _("Count"));
+	strvec_push(&table.header, _("Disk Size"));
+	strvec_push(&table.header, _("Inflated Size"));
+
+	for (size_t i = 0; i < summary_nr; i++) {
+		char *label_str =  xstrdup(summary[i].label);
+		char *nr_str = xstrfmt("%"PRIuMAX, (uintmax_t)summary[i].nr);
+		char *disk_str = xstrfmt("%"PRIuMAX, (uintmax_t)summary[i].disk_size);
+		char *inflate_str = xstrfmt("%"PRIuMAX, (uintmax_t)summary[i].inflated_size);
+
+		insert_table_rowv(&table, label_str, nr_str,
+				  disk_str, inflate_str, NULL);
+
+		free(label_str);
+		free(nr_str);
+		free(disk_str);
+		free(inflate_str);
+	}
+
+	print_table_plaintext(&table);
+	clear_table(&table);
+}
+
+static void survey_report_plaintext_sorted_size(
+		struct survey_report_top_table *top)
+{
+	survey_report_object_sizes(top->name,  _("Path"),
+				   top->data, top->nr);
+}
+
+static void survey_report_plaintext(struct survey_context *ctx)
+{
+	printf("GIT SURVEY for \"%s\"\n", ctx->repo->worktree);
+	printf("-----------------------------------------------------\n");
+	survey_report_plaintext_refs(ctx);
+	survey_report_plaintext_reachable_object_summary(ctx);
+	survey_report_object_sizes(_("TOTAL OBJECT SIZES BY TYPE"),
+				   _("Object Type"),
+				   ctx->report.by_type,
+				   REPORT_TYPE_COUNT);
+
+	survey_report_plaintext_sorted_size(
+		&ctx->report.top_paths_by_count[REPORT_TYPE_TREE]);
+	survey_report_plaintext_sorted_size(
+		&ctx->report.top_paths_by_count[REPORT_TYPE_BLOB]);
+
+	survey_report_plaintext_sorted_size(
+		&ctx->report.top_paths_by_disk[REPORT_TYPE_TREE]);
+	survey_report_plaintext_sorted_size(
+		&ctx->report.top_paths_by_disk[REPORT_TYPE_BLOB]);
+
+	survey_report_plaintext_sorted_size(
+		&ctx->report.top_paths_by_inflate[REPORT_TYPE_TREE]);
+	survey_report_plaintext_sorted_size(
+		&ctx->report.top_paths_by_inflate[REPORT_TYPE_BLOB]);
+}
+
+/*
+ * After parsing the command line arguments, figure out which refs we
+ * should scan.
+ *
+ * If ANY were given in positive sense, then we ONLY include them and
+ * do not use the builtin values.
+ */
+static void fixup_refs_wanted(struct survey_context *ctx)
+{
+	struct survey_refs_wanted *rw = &ctx->opts.refs;
+
+	/*
+	 * `--all-refs` overrides and enables everything.
+	 */
+	if (rw->want_all_refs == 1) {
+		rw->want_branches = 1;
+		rw->want_tags = 1;
+		rw->want_remotes = 1;
+		rw->want_detached = 1;
+		rw->want_other = 1;
+		return;
+	}
+
+	/*
+	 * If none of the `--<ref-type>` were given, we assume all
+	 * of the builtin unspecified values.
+	 */
+	if (rw->want_branches == -1 &&
+	    rw->want_tags == -1 &&
+	    rw->want_remotes == -1 &&
+	    rw->want_detached == -1 &&
+	    rw->want_other == -1) {
+		*rw = default_ref_options;
+		return;
+	}
+
+	/*
+	 * Since we only allow positive boolean values on the command
+	 * line, we will only have true values where they specified
+	 * a `--<ref-type>`.
+	 *
+	 * So anything that still has an unspecified value should be
+	 * set to false.
+	 */
+	if (rw->want_branches == -1)
+		rw->want_branches = 0;
+	if (rw->want_tags == -1)
+		rw->want_tags = 0;
+	if (rw->want_remotes == -1)
+		rw->want_remotes = 0;
+	if (rw->want_detached == -1)
+		rw->want_detached = 0;
+	if (rw->want_other == -1)
+		rw->want_other = 0;
+}
+
+static int survey_load_config_cb(const char *var, const char *value,
+				 const struct config_context *cctx, void *pvoid)
+{
+	struct survey_context *ctx = pvoid;
+
+	if (!strcmp(var, "survey.verbose")) {
+		ctx->opts.verbose = git_config_bool(var, value);
+		return 0;
+	}
+	if (!strcmp(var, "survey.progress")) {
+		ctx->opts.show_progress = git_config_bool(var, value);
+		return 0;
+	}
+	if (!strcmp(var, "survey.top")) {
+		ctx->opts.top_nr = git_config_bool(var, value);
+		return 0;
+	}
+
+	return git_default_config(var, value, cctx, pvoid);
+}
+
+static void survey_load_config(struct survey_context *ctx)
+{
+	git_config(survey_load_config_cb, ctx);
+}
+
+static void do_load_refs(struct survey_context *ctx,
+			 struct ref_array *ref_array)
+{
+	struct ref_filter filter = REF_FILTER_INIT;
+	struct ref_sorting *sorting;
+	struct string_list sorting_options = STRING_LIST_INIT_DUP;
+
+	string_list_append(&sorting_options, "objectname");
+	sorting = ref_sorting_options(&sorting_options);
+
+	if (ctx->opts.refs.want_detached)
+		strvec_push(&ctx->refs, "HEAD");
+
+	if (ctx->opts.refs.want_all_refs) {
+		strvec_push(&ctx->refs, "refs/");
+	} else {
+		if (ctx->opts.refs.want_branches)
+			strvec_push(&ctx->refs, "refs/heads/");
+		if (ctx->opts.refs.want_tags)
+			strvec_push(&ctx->refs, "refs/tags/");
+		if (ctx->opts.refs.want_remotes)
+			strvec_push(&ctx->refs, "refs/remotes/");
+		if (ctx->opts.refs.want_other) {
+			strvec_push(&ctx->refs, "refs/notes/");
+			strvec_push(&ctx->refs, "refs/stash/");
+		}
+	}
+
+	filter.name_patterns = ctx->refs.v;
+	filter.ignore_case = 0;
+	filter.match_as_path = 1;
+
+	if (ctx->opts.show_progress) {
+		ctx->progress_total = 0;
+		ctx->progress = start_progress(_("Scanning refs..."), 0);
+	}
+
+	filter_refs(ref_array, &filter, FILTER_REFS_KIND_MASK);
+
+	if (ctx->opts.show_progress) {
+		ctx->progress_total = ref_array->nr;
+		display_progress(ctx->progress, ctx->progress_total);
+	}
+
+	ref_array_sort(sorting, ref_array);
+
+	stop_progress(&ctx->progress);
+	ref_filter_clear(&filter);
+	ref_sorting_release(sorting);
+}
+
+/*
+ * The REFS phase:
+ *
+ * Load the set of requested refs and assess them for scalablity problems.
+ * Use that set to start a treewalk to all reachable objects and assess
+ * them.
+ *
+ * This data will give us insights into the repository itself (the number
+ * of refs, the size and shape of the DAG, the number and size of the
+ * objects).
+ *
+ * Theoretically, this data is independent of the on-disk representation
+ * (e.g. independent of packing concerns).
+ */
+static void survey_phase_refs(struct survey_context *ctx)
+{
+	trace2_region_enter("survey", "phase/refs", ctx->repo);
+	do_load_refs(ctx, &ctx->ref_array);
+
+	ctx->report.refs.refs_nr = ctx->ref_array.nr;
+	for (int i = 0; i < ctx->ref_array.nr; i++) {
+		unsigned long size;
+		struct ref_array_item *item = ctx->ref_array.items[i];
+
+		switch (item->kind) {
+		case FILTER_REFS_TAGS:
+			ctx->report.refs.tags_nr++;
+			if (oid_object_info(ctx->repo,
+					    &item->objectname,
+					    &size) == OBJ_TAG)
+				ctx->report.refs.tags_annotated_nr++;
+			break;
+
+		case FILTER_REFS_BRANCHES:
+			ctx->report.refs.branches_nr++;
+			break;
+
+		case FILTER_REFS_REMOTES:
+			ctx->report.refs.remote_refs_nr++;
+			break;
+
+		case FILTER_REFS_OTHERS:
+			ctx->report.refs.others_nr++;
+			break;
+
+		default:
+			ctx->report.refs.unknown_nr++;
+			break;
+		}
+	}
+
+	trace2_region_leave("survey", "phase/refs", ctx->repo);
+}
+
+static void increment_object_counts(
+		struct survey_report_object_summary *summary,
+		enum object_type type,
+		size_t nr)
+{
+	switch (type) {
+	case OBJ_COMMIT:
+		summary->commits_nr += nr;
+		break;
+
+	case OBJ_TREE:
+		summary->trees_nr += nr;
+		break;
+
+	case OBJ_BLOB:
+		summary->blobs_nr += nr;
+		break;
+
+	case OBJ_TAG:
+		summary->tags_nr += nr;
+		break;
+
+	default:
+		break;
+	}
+}
+
+static void increment_totals(struct survey_context *ctx,
+			     struct oid_array *oids,
+			     struct survey_report_object_size_summary *summary)
+{
+	for (size_t i = 0; i < oids->nr; i++) {
+		struct object_info oi = OBJECT_INFO_INIT;
+		unsigned oi_flags = OBJECT_INFO_FOR_PREFETCH;
+		unsigned long object_length = 0;
+		off_t disk_sizep = 0;
+		enum object_type type;
+
+		oi.typep = &type;
+		oi.sizep = &object_length;
+		oi.disk_sizep = &disk_sizep;
+
+		if (oid_object_info_extended(ctx->repo, &oids->oid[i],
+					     &oi, oi_flags) < 0) {
+			summary->num_missing++;
+		} else {
+			summary->nr++;
+			summary->disk_size += disk_sizep;
+			summary->inflated_size += object_length;
+		}
+	}
+}
+
+static void increment_object_totals(struct survey_context *ctx,
+				    struct oid_array *oids,
+				    enum object_type type,
+				    const char *path)
+{
+	struct survey_report_object_size_summary *total;
+	struct survey_report_object_size_summary summary = { 0 };
+
+	increment_totals(ctx, oids, &summary);
+
+	switch (type) {
+	case OBJ_COMMIT:
+		total = &ctx->report.by_type[REPORT_TYPE_COMMIT];
+		break;
+
+	case OBJ_TREE:
+		total = &ctx->report.by_type[REPORT_TYPE_TREE];
+		break;
+
+	case OBJ_BLOB:
+		total = &ctx->report.by_type[REPORT_TYPE_BLOB];
+		break;
+
+	case OBJ_TAG:
+		total = &ctx->report.by_type[REPORT_TYPE_TAG];
+		break;
+
+	default:
+		BUG("No other type allowed");
+	}
+
+	total->nr += summary.nr;
+	total->disk_size += summary.disk_size;
+	total->inflated_size += summary.inflated_size;
+	total->num_missing += summary.num_missing;
+
+	if (type == OBJ_TREE || type == OBJ_BLOB) {
+		int index = type == OBJ_TREE ?
+			    REPORT_TYPE_TREE : REPORT_TYPE_BLOB;
+		struct survey_report_top_table *top;
+
+		/*
+		 * Temporarily store (const char *) here, but it will
+		 * be duped if inserted and will not be freed.
+		 */
+		summary.label = (char *)path;
+
+		top = ctx->report.top_paths_by_count;
+		maybe_insert_into_top_size(&top[index], &summary);
+
+		top = ctx->report.top_paths_by_disk;
+		maybe_insert_into_top_size(&top[index], &summary);
+
+		top = ctx->report.top_paths_by_inflate;
+		maybe_insert_into_top_size(&top[index], &summary);
+	}
+}
+
+static int survey_objects_path_walk_fn(const char *path,
+				       struct oid_array *oids,
+				       enum object_type type,
+				       void *data)
+{
+	struct survey_context *ctx = data;
+
+	increment_object_counts(&ctx->report.reachable_objects,
+				type, oids->nr);
+	increment_object_totals(ctx, oids, type, path);
+
+	ctx->progress_nr += oids->nr;
+	display_progress(ctx->progress, ctx->progress_nr);
+
+	return 0;
+}
+
+static void initialize_report(struct survey_context *ctx)
+{
+	CALLOC_ARRAY(ctx->report.by_type, REPORT_TYPE_COUNT);
+	ctx->report.by_type[REPORT_TYPE_COMMIT].label = xstrdup(_("Commits"));
+	ctx->report.by_type[REPORT_TYPE_TREE].label = xstrdup(_("Trees"));
+	ctx->report.by_type[REPORT_TYPE_BLOB].label = xstrdup(_("Blobs"));
+	ctx->report.by_type[REPORT_TYPE_TAG].label = xstrdup(_("Tags"));
+
+	CALLOC_ARRAY(ctx->report.top_paths_by_count, REPORT_TYPE_COUNT);
+	init_top_sizes(&ctx->report.top_paths_by_count[REPORT_TYPE_TREE],
+		       ctx->opts.top_nr, _("TOP DIRECTORIES BY COUNT"), cmp_by_nr);
+	init_top_sizes(&ctx->report.top_paths_by_count[REPORT_TYPE_BLOB],
+		       ctx->opts.top_nr, _("TOP FILES BY COUNT"), cmp_by_nr);
+
+	CALLOC_ARRAY(ctx->report.top_paths_by_disk, REPORT_TYPE_COUNT);
+	init_top_sizes(&ctx->report.top_paths_by_disk[REPORT_TYPE_TREE],
+		       ctx->opts.top_nr, _("TOP DIRECTORIES BY DISK SIZE"), cmp_by_disk_size);
+	init_top_sizes(&ctx->report.top_paths_by_disk[REPORT_TYPE_BLOB],
+		       ctx->opts.top_nr, _("TOP FILES BY DISK SIZE"), cmp_by_disk_size);
+
+	CALLOC_ARRAY(ctx->report.top_paths_by_inflate, REPORT_TYPE_COUNT);
+	init_top_sizes(&ctx->report.top_paths_by_inflate[REPORT_TYPE_TREE],
+		       ctx->opts.top_nr, _("TOP DIRECTORIES BY INFLATED SIZE"), cmp_by_inflated_size);
+	init_top_sizes(&ctx->report.top_paths_by_inflate[REPORT_TYPE_BLOB],
+		       ctx->opts.top_nr, _("TOP FILES BY INFLATED SIZE"), cmp_by_inflated_size);
+}
+
+static void survey_phase_objects(struct survey_context *ctx)
+{
+	struct rev_info revs = REV_INFO_INIT;
+	struct path_walk_info info = PATH_WALK_INFO_INIT;
+	unsigned int add_flags = 0;
+
+	trace2_region_enter("survey", "phase/objects", ctx->repo);
+
+	info.revs = &revs;
+	info.path_fn = survey_objects_path_walk_fn;
+	info.path_fn_data = ctx;
+
+	initialize_report(ctx);
+
+	repo_init_revisions(ctx->repo, &revs, "");
+	revs.tag_objects = 1;
+
+	ctx->progress_nr = 0;
+	ctx->progress_total = ctx->ref_array.nr;
+	if (ctx->opts.show_progress)
+		ctx->progress = start_progress(_("Preparing object walk"),
+					       ctx->progress_total);
+	for (int i = 0; i < ctx->ref_array.nr; i++) {
+		struct ref_array_item *item = ctx->ref_array.items[i];
+		add_pending_oid(&revs, NULL, &item->objectname, add_flags);
+		display_progress(ctx->progress, ++(ctx->progress_nr));
+	}
+	stop_progress(&ctx->progress);
+
+	ctx->progress_nr = 0;
+	ctx->progress_total = 0;
+	if (ctx->opts.show_progress)
+		ctx->progress = start_progress(_("Walking objects"), 0);
+	walk_objects_by_path(&info);
+	stop_progress(&ctx->progress);
+
+	release_revisions(&revs);
+	trace2_region_leave("survey", "phase/objects", ctx->repo);
+}
+
+int cmd_survey(int argc, const char **argv, const char *prefix, struct repository *repo)
+{
+	static struct survey_context ctx = {
+		.opts = {
+			.verbose = 0,
+			.show_progress = -1, /* defaults to isatty(2) */
+			.top_nr = 10,
+
+			.refs.want_all_refs = -1,
+
+			.refs.want_branches = -1, /* default these to undefined */
+			.refs.want_tags = -1,
+			.refs.want_remotes = -1,
+			.refs.want_detached = -1,
+			.refs.want_other = -1,
+		},
+		.refs = STRVEC_INIT,
+	};
+
+	static struct option survey_options[] = {
+		OPT__VERBOSE(&ctx.opts.verbose, N_("verbose output")),
+		OPT_BOOL(0, "progress", &ctx.opts.show_progress, N_("show progress")),
+		OPT_INTEGER('n', "top", &ctx.opts.top_nr,
+			    N_("number of entries to include in detail tables")),
+
+		OPT_BOOL_F(0, "all-refs", &ctx.opts.refs.want_all_refs, N_("include all refs"),          PARSE_OPT_NONEG),
+
+		OPT_BOOL_F(0, "branches", &ctx.opts.refs.want_branches, N_("include branches"),          PARSE_OPT_NONEG),
+		OPT_BOOL_F(0, "tags",     &ctx.opts.refs.want_tags,     N_("include tags"),              PARSE_OPT_NONEG),
+		OPT_BOOL_F(0, "remotes",  &ctx.opts.refs.want_remotes,  N_("include all remotes refs"),  PARSE_OPT_NONEG),
+		OPT_BOOL_F(0, "detached", &ctx.opts.refs.want_detached, N_("include detached HEAD"),     PARSE_OPT_NONEG),
+		OPT_BOOL_F(0, "other",    &ctx.opts.refs.want_other,    N_("include notes and stashes"), PARSE_OPT_NONEG),
+
+		OPT_END(),
+	};
+
+	if (argc == 2 && !strcmp(argv[1], "-h"))
+		usage_with_options(survey_usage, survey_options);
+
+	if (isatty(2))
+		color_fprintf_ln(stderr,
+				 want_color_fd(2, GIT_COLOR_AUTO) ? GIT_COLOR_YELLOW : "",
+				 "(THIS IS EXPERIMENTAL, EXPECT THE OUTPUT FORMAT TO CHANGE!)");
+
+	ctx.repo = repo;
+
+	prepare_repo_settings(ctx.repo);
+	survey_load_config(&ctx);
+
+	argc = parse_options(argc, argv, prefix, survey_options, survey_usage, 0);
+
+	if (ctx.opts.show_progress < 0)
+		ctx.opts.show_progress = isatty(2);
+
+	fixup_refs_wanted(&ctx);
+
+	survey_phase_refs(&ctx);
+
+	survey_phase_objects(&ctx);
+
+	survey_report_plaintext(&ctx);
+
+	clear_survey_context(&ctx);
+	return 0;
+}
diff --git a/ci/lib.sh b/ci/lib.sh
index 6a1267fbcb3c8b..38d5e641ca3ffc 100755
--- a/ci/lib.sh
+++ b/ci/lib.sh
@@ -225,6 +225,12 @@ then
 
 	GIT_TEST_OPTS="--write-junit-xml"
 	JOBS=10
+	case "$CI_OS_NAME" in
+	linux) runs_on_pool=ubuntu-latest;;
+	macos|osx) runs_on_pool=macos-latest;;
+	windows_nt) runs_on_pool=windows-latest;;
+	*) echo "Unhandled OS: $CI_OS_NAME" >&2; exit 1;;
+	esac
 elif test true = "$GITHUB_ACTIONS"
 then
 	CI_TYPE=github-actions
diff --git a/ci/run-build-and-tests.sh b/ci/run-build-and-tests.sh
index 76667a1277720d..8e35528bf05cd7 100755
--- a/ci/run-build-and-tests.sh
+++ b/ci/run-build-and-tests.sh
@@ -5,11 +5,6 @@
 
 . ${0%/*}/lib.sh
 
-case "$CI_OS_NAME" in
-windows*) cmd //c mklink //j t\\.prove "$(cygpath -aw "$cache_dir/.prove")";;
-*) ln -s "$cache_dir/.prove" t/.prove;;
-esac
-
 run_tests=t
 
 case "$jobname" in
@@ -31,6 +26,8 @@ linux-TEST-vars)
 	export GIT_TEST_NO_WRITE_REV_INDEX=1
 	export GIT_TEST_CHECKOUT_WORKERS=2
 	export GIT_TEST_PACK_USE_BITMAP_BOUNDARY_TRAVERSAL=1
+	export GIT_TEST_FULL_NAME_HASH=1
+	export GIT_TEST_PACK_PATH_WALK=1
 	;;
 linux-clang)
 	export GIT_TEST_DEFAULT_HASH=sha1
@@ -73,5 +70,9 @@ case "$jobname" in
 	;;
 esac
 
+case " $MAKE_TARGETS " in
+*" all "*) make -C contrib/subtree test;;
+esac
+
 check_unignored_build_artifacts
 save_good_tree
diff --git a/ci/run-test-slice.sh b/ci/run-test-slice.sh
index e167e646f79e3d..6e21260e17543b 100755
--- a/ci/run-test-slice.sh
+++ b/ci/run-test-slice.sh
@@ -5,11 +5,6 @@
 
 . ${0%/*}/lib.sh
 
-case "$CI_OS_NAME" in
-windows*) cmd //c mklink //j t\\.prove "$(cygpath -aw "$cache_dir/.prove")";;
-*) ln -s "$cache_dir/.prove" t/.prove;;
-esac
-
 group "Run tests" make --quiet -C t T="$(cd t &&
 	./helper/test-tool path-utils slice-tests "$1" "$2" t[0-9]*.sh |
 	tr '\n' ' ')" ||
@@ -20,4 +15,7 @@ if [ "$1" == "0" ] ; then
 	group "Run unit tests" make --quiet -C t unit-tests-test-tool
 fi
 
+# Run the git subtree tests only if main tests succeeded
+test 0 != "$1" || make -C contrib/subtree test
+
 check_unignored_build_artifacts
diff --git a/command-list.txt b/command-list.txt
index e0bb87b3b5c278..ecc9d2281a0909 100644
--- a/command-list.txt
+++ b/command-list.txt
@@ -60,6 +60,7 @@ git-annotate                            ancillaryinterrogators
 git-apply                               plumbingmanipulators            complete
 git-archimport                          foreignscminterface
 git-archive                             mainporcelain
+git-backfill                            mainporcelain           history
 git-bisect                              mainporcelain           info
 git-blame                               ancillaryinterrogators          complete
 git-branch                              mainporcelain           history
@@ -186,6 +187,7 @@ git-stash                               mainporcelain
 git-status                              mainporcelain           info
 git-stripspace                          purehelpers
 git-submodule                           mainporcelain
+git-survey                              mainporcelain
 git-svn                                 foreignscminterface
 git-switch                              mainporcelain           history
 git-symbolic-ref                        plumbingmanipulators
diff --git a/common-main.c b/common-main.c
index 8e68ac9e42993d..75cf58abf645f0 100644
--- a/common-main.c
+++ b/common-main.c
@@ -77,6 +77,13 @@ static void check_bug_if_BUG(void)
 /* We wrap exit() to call common_exit() in git-compat-util.h */
 int common_exit(const char *file, int line, int code)
 {
+	/*
+	 *  Windows Filtering Platform driver provided by the security software
+	 * may change buffer type of stdout from _IONBF to _IOFBF.
+	 * It will no output without fflush manually.
+	 */
+	fflush(stdout);
+
 	/*
 	 * For non-POSIX systems: Take the lowest 8 bits of the "code"
 	 * to e.g. turn -1 into 255. On a POSIX system this is
diff --git a/compat/bswap.h b/compat/bswap.h
index 512f6f4b9937c8..a443e99eef2f1c 100644
--- a/compat/bswap.h
+++ b/compat/bswap.h
@@ -35,7 +35,19 @@ static inline uint64_t default_bswap64(uint64_t val)
 #undef bswap32
 #undef bswap64
 
-#if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
+/**
+ * __has_builtin is available since Clang 10 and GCC 10.
+ * Below is a fallback for older compilers.
+ */
+#ifndef __has_builtin
+	#define __has_builtin(x) 0
+#endif
+
+#if __has_builtin(__builtin_bswap32) && __has_builtin(__builtin_bswap64)
+#define bswap32(x) __builtin_bswap32((x))
+#define bswap64(x) __builtin_bswap64((x))
+
+#elif defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
 
 #define bswap32 git_bswap32
 static inline uint32_t git_bswap32(uint32_t x)
diff --git a/compat/lazyload-curl.c b/compat/lazyload-curl.c
new file mode 100644
index 00000000000000..2f2b2ebc27993e
--- /dev/null
+++ b/compat/lazyload-curl.c
@@ -0,0 +1,418 @@
+#include "../git-compat-util.h"
+#include "../git-curl-compat.h"
+#ifndef WIN32
+#include <dlfcn.h>
+#endif
+
+/*
+ * The ABI version of libcurl is encoded in its shared libraries' file names.
+ * This ABI version has not changed since October 2006 and is unlikely to be
+ * changed in the future. See https://curl.se/libcurl/abi.html for details.
+ */
+#define LIBCURL_ABI_VERSION "4"
+
+typedef void (*func_t)(void);
+
+#ifndef WIN32
+#ifdef __APPLE__
+#define LIBCURL_FILE_NAME(base) base "." LIBCURL_ABI_VERSION ".dylib"
+#else
+#define LIBCURL_FILE_NAME(base) base ".so." LIBCURL_ABI_VERSION
+#endif
+
+static void *load_library(const char *name)
+{
+	return dlopen(name, RTLD_LAZY);
+}
+
+static func_t load_function(void *handle, const char *name)
+{
+	/*
+	 * Casting the return value of `dlsym()` to a function pointer is
+	 * explicitly allowed in recent POSIX standards, but GCC complains
+	 * about this in pedantic mode nevertheless. For more about this issue,
+	 * see https://stackoverflow.com/q/31526876/1860823 and
+	 * http://stackoverflow.com/a/36385690/1905491.
+	 */
+	func_t f;
+	*(void **)&f = dlsym(handle, name);
+	return f;
+}
+#else
+#define LIBCURL_FILE_NAME(base) base "-" LIBCURL_ABI_VERSION ".dll"
+
+static void *load_library(const char *name)
+{
+	size_t name_size = strlen(name) + 1;
+	const char *path = getenv("PATH");
+	char dll_path[MAX_PATH];
+
+	while (path && *path) {
+		const char *sep = strchrnul(path, ';');
+		size_t len = sep - path;
+
+		if (len && len + name_size < sizeof(dll_path)) {
+			memcpy(dll_path, path, len);
+			dll_path[len] = '/';
+			memcpy(dll_path + len + 1, name, name_size);
+
+			if (!access(dll_path, R_OK)) {
+				wchar_t wpath[MAX_PATH];
+				int wlen = MultiByteToWideChar(CP_UTF8, 0, dll_path, -1, wpath, ARRAY_SIZE(wpath));
+				void *res = wlen ? (void *)LoadLibraryExW(wpath, NULL, 0) : NULL;
+				if (!res) {
+					DWORD err = GetLastError();
+					char buf[1024];
+
+					if (!FormatMessageA(FORMAT_MESSAGE_FROM_SYSTEM |
+							    FORMAT_MESSAGE_ARGUMENT_ARRAY |
+							    FORMAT_MESSAGE_IGNORE_INSERTS,
+							    NULL, err, LANG_NEUTRAL,
+							    buf, sizeof(buf) - 1, NULL))
+						xsnprintf(buf, sizeof(buf), "last error: %ld", err);
+					error("LoadLibraryExW() failed with: %s", buf);
+				}
+				return res;
+			}
+		}
+
+		path = *sep ? sep + 1 : NULL;
+	}
+
+	return NULL;
+}
+
+static func_t load_function(void *handle, const char *name)
+{
+	return (func_t)GetProcAddress((HANDLE)handle, name);
+}
+#endif
+
+typedef struct curl_version_info_data *(*curl_version_info_type)(CURLversion version);
+static curl_version_info_type curl_version_info_func;
+
+typedef char *(*curl_easy_escape_type)(CURL *handle, const char *string, int length);
+static curl_easy_escape_type curl_easy_escape_func;
+
+typedef void (*curl_free_type)(void *p);
+static curl_free_type curl_free_func;
+
+typedef CURLcode (*curl_global_init_type)(long flags);
+static curl_global_init_type curl_global_init_func;
+
+typedef CURLsslset (*curl_global_sslset_type)(curl_sslbackend id, const char *name, const curl_ssl_backend ***avail);
+static curl_global_sslset_type curl_global_sslset_func;
+
+typedef void (*curl_global_cleanup_type)(void);
+static curl_global_cleanup_type curl_global_cleanup_func;
+
+typedef struct curl_slist *(*curl_slist_append_type)(struct curl_slist *list, const char *data);
+static curl_slist_append_type curl_slist_append_func;
+
+typedef void (*curl_slist_free_all_type)(struct curl_slist *list);
+static curl_slist_free_all_type curl_slist_free_all_func;
+
+typedef const char *(*curl_easy_strerror_type)(CURLcode error);
+static curl_easy_strerror_type curl_easy_strerror_func;
+
+typedef CURLM *(*curl_multi_init_type)(void);
+static curl_multi_init_type curl_multi_init_func;
+
+typedef CURLMcode (*curl_multi_add_handle_type)(CURLM *multi_handle, CURL *curl_handle);
+static curl_multi_add_handle_type curl_multi_add_handle_func;
+
+typedef CURLMcode (*curl_multi_remove_handle_type)(CURLM *multi_handle, CURL *curl_handle);
+static curl_multi_remove_handle_type curl_multi_remove_handle_func;
+
+typedef CURLMcode (*curl_multi_fdset_type)(CURLM *multi_handle, fd_set *read_fd_set, fd_set *write_fd_set, fd_set *exc_fd_set, int *max_fd);
+static curl_multi_fdset_type curl_multi_fdset_func;
+
+typedef CURLMcode (*curl_multi_perform_type)(CURLM *multi_handle, int *running_handles);
+static curl_multi_perform_type curl_multi_perform_func;
+
+typedef CURLMcode (*curl_multi_cleanup_type)(CURLM *multi_handle);
+static curl_multi_cleanup_type curl_multi_cleanup_func;
+
+typedef CURLMsg *(*curl_multi_info_read_type)(CURLM *multi_handle, int *msgs_in_queue);
+static curl_multi_info_read_type curl_multi_info_read_func;
+
+typedef const char *(*curl_multi_strerror_type)(CURLMcode error);
+static curl_multi_strerror_type curl_multi_strerror_func;
+
+typedef CURLMcode (*curl_multi_timeout_type)(CURLM *multi_handle, long *milliseconds);
+static curl_multi_timeout_type curl_multi_timeout_func;
+
+typedef CURL *(*curl_easy_init_type)(void);
+static curl_easy_init_type curl_easy_init_func;
+
+typedef CURLcode (*curl_easy_perform_type)(CURL *curl);
+static curl_easy_perform_type curl_easy_perform_func;
+
+typedef void (*curl_easy_cleanup_type)(CURL *curl);
+static curl_easy_cleanup_type curl_easy_cleanup_func;
+
+typedef CURL *(*curl_easy_duphandle_type)(CURL *curl);
+static curl_easy_duphandle_type curl_easy_duphandle_func;
+
+typedef CURLcode (*curl_easy_getinfo_long_type)(CURL *curl, CURLINFO info, long *value);
+static curl_easy_getinfo_long_type curl_easy_getinfo_long_func;
+
+typedef CURLcode (*curl_easy_getinfo_pointer_type)(CURL *curl, CURLINFO info, void **value);
+static curl_easy_getinfo_pointer_type curl_easy_getinfo_pointer_func;
+
+typedef CURLcode (*curl_easy_getinfo_off_t_type)(CURL *curl, CURLINFO info, curl_off_t *value);
+static curl_easy_getinfo_off_t_type curl_easy_getinfo_off_t_func;
+
+typedef CURLcode (*curl_easy_setopt_long_type)(CURL *curl, CURLoption opt, long value);
+static curl_easy_setopt_long_type curl_easy_setopt_long_func;
+
+typedef CURLcode (*curl_easy_setopt_pointer_type)(CURL *curl, CURLoption opt, void *value);
+static curl_easy_setopt_pointer_type curl_easy_setopt_pointer_func;
+
+typedef CURLcode (*curl_easy_setopt_off_t_type)(CURL *curl, CURLoption opt, curl_off_t value);
+static curl_easy_setopt_off_t_type curl_easy_setopt_off_t_func;
+
+static char ssl_backend[64];
+
+static void lazy_load_curl(void)
+{
+	static int initialized;
+	void *libcurl = NULL;
+	func_t curl_easy_getinfo_func, curl_easy_setopt_func;
+
+	if (initialized)
+		return;
+
+	initialized = 1;
+	if (ssl_backend[0]) {
+		char dll_name[64 + 16];
+		snprintf(dll_name, sizeof(dll_name) - 1,
+			 LIBCURL_FILE_NAME("libcurl-%s"), ssl_backend);
+		libcurl = load_library(dll_name);
+	}
+	if (!libcurl)
+		libcurl = load_library(LIBCURL_FILE_NAME("libcurl"));
+	if (!libcurl)
+		die("failed to load library '%s'", LIBCURL_FILE_NAME("libcurl"));
+
+	curl_version_info_func = (curl_version_info_type)load_function(libcurl, "curl_version_info");
+	curl_easy_escape_func = (curl_easy_escape_type)load_function(libcurl, "curl_easy_escape");
+	curl_free_func = (curl_free_type)load_function(libcurl, "curl_free");
+	curl_global_init_func = (curl_global_init_type)load_function(libcurl, "curl_global_init");
+	curl_global_sslset_func = (curl_global_sslset_type)load_function(libcurl, "curl_global_sslset");
+	curl_global_cleanup_func = (curl_global_cleanup_type)load_function(libcurl, "curl_global_cleanup");
+	curl_slist_append_func = (curl_slist_append_type)load_function(libcurl, "curl_slist_append");
+	curl_slist_free_all_func = (curl_slist_free_all_type)load_function(libcurl, "curl_slist_free_all");
+	curl_easy_strerror_func = (curl_easy_strerror_type)load_function(libcurl, "curl_easy_strerror");
+	curl_multi_init_func = (curl_multi_init_type)load_function(libcurl, "curl_multi_init");
+	curl_multi_add_handle_func = (curl_multi_add_handle_type)load_function(libcurl, "curl_multi_add_handle");
+	curl_multi_remove_handle_func = (curl_multi_remove_handle_type)load_function(libcurl, "curl_multi_remove_handle");
+	curl_multi_fdset_func = (curl_multi_fdset_type)load_function(libcurl, "curl_multi_fdset");
+	curl_multi_perform_func = (curl_multi_perform_type)load_function(libcurl, "curl_multi_perform");
+	curl_multi_cleanup_func = (curl_multi_cleanup_type)load_function(libcurl, "curl_multi_cleanup");
+	curl_multi_info_read_func = (curl_multi_info_read_type)load_function(libcurl, "curl_multi_info_read");
+	curl_multi_strerror_func = (curl_multi_strerror_type)load_function(libcurl, "curl_multi_strerror");
+	curl_multi_timeout_func = (curl_multi_timeout_type)load_function(libcurl, "curl_multi_timeout");
+	curl_easy_init_func = (curl_easy_init_type)load_function(libcurl, "curl_easy_init");
+	curl_easy_perform_func = (curl_easy_perform_type)load_function(libcurl, "curl_easy_perform");
+	curl_easy_cleanup_func = (curl_easy_cleanup_type)load_function(libcurl, "curl_easy_cleanup");
+	curl_easy_duphandle_func = (curl_easy_duphandle_type)load_function(libcurl, "curl_easy_duphandle");
+
+	curl_easy_getinfo_func = load_function(libcurl, "curl_easy_getinfo");
+	curl_easy_getinfo_long_func = (curl_easy_getinfo_long_type)curl_easy_getinfo_func;
+	curl_easy_getinfo_pointer_func = (curl_easy_getinfo_pointer_type)curl_easy_getinfo_func;
+	curl_easy_getinfo_off_t_func = (curl_easy_getinfo_off_t_type)curl_easy_getinfo_func;
+
+	curl_easy_setopt_func = load_function(libcurl, "curl_easy_setopt");
+	curl_easy_setopt_long_func = (curl_easy_setopt_long_type)curl_easy_setopt_func;
+	curl_easy_setopt_pointer_func = (curl_easy_setopt_pointer_type)curl_easy_setopt_func;
+	curl_easy_setopt_off_t_func = (curl_easy_setopt_off_t_type)curl_easy_setopt_func;
+}
+
+struct curl_version_info_data *curl_version_info(CURLversion version)
+{
+	lazy_load_curl();
+	return curl_version_info_func(version);
+}
+
+char *curl_easy_escape(CURL *handle, const char *string, int length)
+{
+	lazy_load_curl();
+	return curl_easy_escape_func(handle, string, length);
+}
+
+void curl_free(void *p)
+{
+	lazy_load_curl();
+	curl_free_func(p);
+}
+
+CURLcode curl_global_init(long flags)
+{
+	lazy_load_curl();
+	return curl_global_init_func(flags);
+}
+
+CURLsslset curl_global_sslset(curl_sslbackend id, const char *name, const curl_ssl_backend ***avail)
+{
+	if (name && strlen(name) < sizeof(ssl_backend))
+		strlcpy(ssl_backend, name, sizeof(ssl_backend));
+
+	lazy_load_curl();
+	return curl_global_sslset_func(id, name, avail);
+}
+
+void curl_global_cleanup(void)
+{
+	lazy_load_curl();
+	curl_global_cleanup_func();
+}
+
+struct curl_slist *curl_slist_append(struct curl_slist *list, const char *data)
+{
+	lazy_load_curl();
+	return curl_slist_append_func(list, data);
+}
+
+void curl_slist_free_all(struct curl_slist *list)
+{
+	lazy_load_curl();
+	curl_slist_free_all_func(list);
+}
+
+const char *curl_easy_strerror(CURLcode error)
+{
+	lazy_load_curl();
+	return curl_easy_strerror_func(error);
+}
+
+CURLM *curl_multi_init(void)
+{
+	lazy_load_curl();
+	return curl_multi_init_func();
+}
+
+CURLMcode curl_multi_add_handle(CURLM *multi_handle, CURL *curl_handle)
+{
+	lazy_load_curl();
+	return curl_multi_add_handle_func(multi_handle, curl_handle);
+}
+
+CURLMcode curl_multi_remove_handle(CURLM *multi_handle, CURL *curl_handle)
+{
+	lazy_load_curl();
+	return curl_multi_remove_handle_func(multi_handle, curl_handle);
+}
+
+CURLMcode curl_multi_fdset(CURLM *multi_handle, fd_set *read_fd_set, fd_set *write_fd_set, fd_set *exc_fd_set, int *max_fd)
+{
+	lazy_load_curl();
+	return curl_multi_fdset_func(multi_handle, read_fd_set, write_fd_set, exc_fd_set, max_fd);
+}
+
+CURLMcode curl_multi_perform(CURLM *multi_handle, int *running_handles)
+{
+	lazy_load_curl();
+	return curl_multi_perform_func(multi_handle, running_handles);
+}
+
+CURLMcode curl_multi_cleanup(CURLM *multi_handle)
+{
+	lazy_load_curl();
+	return curl_multi_cleanup_func(multi_handle);
+}
+
+CURLMsg *curl_multi_info_read(CURLM *multi_handle, int *msgs_in_queue)
+{
+	lazy_load_curl();
+	return curl_multi_info_read_func(multi_handle, msgs_in_queue);
+}
+
+const char *curl_multi_strerror(CURLMcode error)
+{
+	lazy_load_curl();
+	return curl_multi_strerror_func(error);
+}
+
+CURLMcode curl_multi_timeout(CURLM *multi_handle, long *milliseconds)
+{
+	lazy_load_curl();
+	return curl_multi_timeout_func(multi_handle, milliseconds);
+}
+
+CURL *curl_easy_init(void)
+{
+	lazy_load_curl();
+	return curl_easy_init_func();
+}
+
+CURLcode curl_easy_perform(CURL *curl)
+{
+	lazy_load_curl();
+	return curl_easy_perform_func(curl);
+}
+
+void curl_easy_cleanup(CURL *curl)
+{
+	lazy_load_curl();
+	curl_easy_cleanup_func(curl);
+}
+
+CURL *curl_easy_duphandle(CURL *curl)
+{
+	lazy_load_curl();
+	return curl_easy_duphandle_func(curl);
+}
+
+#ifndef CURL_IGNORE_DEPRECATION
+#define CURL_IGNORE_DEPRECATION(x) x
+#endif
+
+#ifndef CURLOPTTYPE_BLOB
+#define CURLOPTTYPE_BLOB 40000
+#endif
+
+#undef curl_easy_getinfo
+CURLcode curl_easy_getinfo(CURL *curl, CURLINFO info, ...)
+{
+	va_list ap;
+	CURLcode res;
+
+	va_start(ap, info);
+	lazy_load_curl();
+	CURL_IGNORE_DEPRECATION(
+		if (info >= CURLINFO_LONG && info < CURLINFO_DOUBLE)
+			res = curl_easy_getinfo_long_func(curl, info, va_arg(ap, long *));
+		else if ((info >= CURLINFO_STRING && info < CURLINFO_LONG) ||
+			 (info >= CURLINFO_SLIST && info < CURLINFO_SOCKET))
+			res = curl_easy_getinfo_pointer_func(curl, info, va_arg(ap, void **));
+		else if (info >= CURLINFO_OFF_T)
+			res = curl_easy_getinfo_off_t_func(curl, info, va_arg(ap, curl_off_t *));
+		else
+			die("%s:%d: TODO (info: %d)!", __FILE__, __LINE__, info);
+	)
+	va_end(ap);
+	return res;
+}
+
+#undef curl_easy_setopt
+CURLcode curl_easy_setopt(CURL *curl, CURLoption opt, ...)
+{
+	va_list ap;
+	CURLcode res;
+
+	va_start(ap, opt);
+	lazy_load_curl();
+	CURL_IGNORE_DEPRECATION(
+		if (opt >= CURLOPTTYPE_LONG && opt < CURLOPTTYPE_OBJECTPOINT)
+			res = curl_easy_setopt_long_func(curl, opt, va_arg(ap, long));
+		else if (opt >= CURLOPTTYPE_OBJECTPOINT && opt < CURLOPTTYPE_OFF_T)
+			res = curl_easy_setopt_pointer_func(curl, opt, va_arg(ap, void *));
+		else if (opt >= CURLOPTTYPE_OFF_T && opt < CURLOPTTYPE_BLOB)
+			res = curl_easy_setopt_off_t_func(curl, opt, va_arg(ap, curl_off_t));
+		else
+			die("%s:%d: TODO (opt: %d)!", __FILE__, __LINE__, opt);
+	)
+	va_end(ap);
+	return res;
+}
diff --git a/compat/mimalloc/LICENSE b/compat/mimalloc/LICENSE
new file mode 100644
index 00000000000000..670b668a0c928e
--- /dev/null
+++ b/compat/mimalloc/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2018-2021 Microsoft Corporation, Daan Leijen
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/compat/mimalloc/alloc-aligned.c b/compat/mimalloc/alloc-aligned.c
new file mode 100644
index 00000000000000..e975af5f7c2ad4
--- /dev/null
+++ b/compat/mimalloc/alloc-aligned.c
@@ -0,0 +1,298 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/prim.h"  // mi_prim_get_default_heap
+
+#include <string.h>     // memset
+
+// ------------------------------------------------------
+// Aligned Allocation
+// ------------------------------------------------------
+
+// Fallback primitive aligned allocation -- split out for better codegen
+static mi_decl_noinline void* mi_heap_malloc_zero_aligned_at_fallback(mi_heap_t* const heap, const size_t size, const size_t alignment, const size_t offset, const bool zero) mi_attr_noexcept
+{
+  mi_assert_internal(size <= PTRDIFF_MAX);
+  mi_assert_internal(alignment != 0 && _mi_is_power_of_two(alignment));
+
+  const uintptr_t align_mask = alignment - 1;  // for any x, `(x & align_mask) == (x % alignment)`
+  const size_t padsize = size + MI_PADDING_SIZE;
+
+  // use regular allocation if it is guaranteed to fit the alignment constraints
+  if (offset==0 && alignment<=padsize && padsize<=MI_MAX_ALIGN_GUARANTEE && (padsize&align_mask)==0) {
+    void* p = _mi_heap_malloc_zero(heap, size, zero);
+    mi_assert_internal(p == NULL || ((uintptr_t)p % alignment) == 0);
+    return p;
+  }
+
+  void* p;
+  size_t oversize;
+  if mi_unlikely(alignment > MI_ALIGNMENT_MAX) {
+    // use OS allocation for very large alignment and allocate inside a huge page (dedicated segment with 1 page)
+    // This can support alignments >= MI_SEGMENT_SIZE by ensuring the object can be aligned at a point in the
+    // first (and single) page such that the segment info is `MI_SEGMENT_SIZE` bytes before it (so it can be found by aligning the pointer down)
+    if mi_unlikely(offset != 0) {
+      // todo: cannot support offset alignment for very large alignments yet
+      #if MI_DEBUG > 0
+      _mi_error_message(EOVERFLOW, "aligned allocation with a very large alignment cannot be used with an alignment offset (size %zu, alignment %zu, offset %zu)\n", size, alignment, offset);
+      #endif
+      return NULL;
+    }
+    oversize = (size <= MI_SMALL_SIZE_MAX ? MI_SMALL_SIZE_MAX + 1 /* ensure we use generic malloc path */ : size);
+    p = _mi_heap_malloc_zero_ex(heap, oversize, false, alignment); // the page block size should be large enough to align in the single huge page block
+    // zero afterwards as only the area from the aligned_p may be committed!
+    if (p == NULL) return NULL;
+  }
+  else {
+    // otherwise over-allocate
+    oversize = size + alignment - 1;
+    p = _mi_heap_malloc_zero(heap, oversize, zero);
+    if (p == NULL) return NULL;
+  }
+
+  // .. and align within the allocation
+  const uintptr_t poffset = ((uintptr_t)p + offset) & align_mask;
+  const uintptr_t adjust  = (poffset == 0 ? 0 : alignment - poffset);
+  mi_assert_internal(adjust < alignment);
+  void* aligned_p = (void*)((uintptr_t)p + adjust);
+  if (aligned_p != p) {
+    mi_page_t* page = _mi_ptr_page(p);
+    mi_page_set_has_aligned(page, true);
+    _mi_padding_shrink(page, (mi_block_t*)p, adjust + size);
+  }
+  // todo: expand padding if overallocated ?
+
+  mi_assert_internal(mi_page_usable_block_size(_mi_ptr_page(p)) >= adjust + size);
+  mi_assert_internal(p == _mi_page_ptr_unalign(_mi_ptr_segment(aligned_p), _mi_ptr_page(aligned_p), aligned_p));
+  mi_assert_internal(((uintptr_t)aligned_p + offset) % alignment == 0);
+  mi_assert_internal(mi_usable_size(aligned_p)>=size);
+  mi_assert_internal(mi_usable_size(p) == mi_usable_size(aligned_p)+adjust);
+
+  // now zero the block if needed
+  if (alignment > MI_ALIGNMENT_MAX) {
+    // for the tracker, on huge aligned allocations only from the start of the large block is defined
+    mi_track_mem_undefined(aligned_p, size);
+    if (zero) {
+      _mi_memzero_aligned(aligned_p, mi_usable_size(aligned_p));
+    }
+  }
+
+  if (p != aligned_p) {
+    mi_track_align(p,aligned_p,adjust,mi_usable_size(aligned_p));
+  }
+  return aligned_p;
+}
+
+// Primitive aligned allocation
+static void* mi_heap_malloc_zero_aligned_at(mi_heap_t* const heap, const size_t size, const size_t alignment, const size_t offset, const bool zero) mi_attr_noexcept
+{
+  // note: we don't require `size > offset`, we just guarantee that the address at offset is aligned regardless of the allocated size.
+  if mi_unlikely(alignment == 0 || !_mi_is_power_of_two(alignment)) { // require power-of-two (see <https://en.cppreference.com/w/c/memory/aligned_alloc>)
+    #if MI_DEBUG > 0
+    _mi_error_message(EOVERFLOW, "aligned allocation requires the alignment to be a power-of-two (size %zu, alignment %zu)\n", size, alignment);
+    #endif
+    return NULL;
+  }
+
+  if mi_unlikely(size > PTRDIFF_MAX) {          // we don't allocate more than PTRDIFF_MAX (see <https://sourceware.org/ml/libc-announce/2019/msg00001.html>)
+    #if MI_DEBUG > 0
+    _mi_error_message(EOVERFLOW, "aligned allocation request is too large (size %zu, alignment %zu)\n", size, alignment);
+    #endif
+    return NULL;
+  }
+  const uintptr_t align_mask = alignment-1;       // for any x, `(x & align_mask) == (x % alignment)`
+  const size_t padsize = size + MI_PADDING_SIZE;  // note: cannot overflow due to earlier size > PTRDIFF_MAX check
+
+  // try first if there happens to be a small block available with just the right alignment
+  if mi_likely(padsize <= MI_SMALL_SIZE_MAX && alignment <= padsize) {
+    mi_page_t* page = _mi_heap_get_free_small_page(heap, padsize);
+    const bool is_aligned = (((uintptr_t)page->free+offset) & align_mask)==0;
+    if mi_likely(page->free != NULL && is_aligned)
+    {
+      #if MI_STAT>1
+      mi_heap_stat_increase(heap, malloc, size);
+      #endif
+      void* p = _mi_page_malloc(heap, page, padsize, zero); // TODO: inline _mi_page_malloc
+      mi_assert_internal(p != NULL);
+      mi_assert_internal(((uintptr_t)p + offset) % alignment == 0);
+      mi_track_malloc(p,size,zero);
+      return p;
+    }
+  }
+  // fallback
+  return mi_heap_malloc_zero_aligned_at_fallback(heap, size, alignment, offset, zero);
+}
+
+
+// ------------------------------------------------------
+// Optimized mi_heap_malloc_aligned / mi_malloc_aligned
+// ------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_malloc_aligned_at(mi_heap_t* heap, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_malloc_zero_aligned_at(heap, size, alignment, offset, false);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_malloc_aligned(mi_heap_t* heap, size_t size, size_t alignment) mi_attr_noexcept {
+  if mi_unlikely(alignment == 0 || !_mi_is_power_of_two(alignment)) return NULL;
+  #if !MI_PADDING
+  // without padding, any small sized allocation is naturally aligned (see also `_mi_segment_page_start`)
+  if mi_likely(_mi_is_power_of_two(size) && size >= alignment && size <= MI_SMALL_SIZE_MAX)
+  #else
+  // with padding, we can only guarantee this for fixed alignments
+  if mi_likely((alignment == sizeof(void*) || (alignment == MI_MAX_ALIGN_SIZE && size > (MI_MAX_ALIGN_SIZE/2)))
+		&& size <= MI_SMALL_SIZE_MAX)
+  #endif
+  {
+    // fast path for common alignment and size
+    return mi_heap_malloc_small(heap, size);
+  }
+  else {
+    return mi_heap_malloc_aligned_at(heap, size, alignment, 0);
+  }
+}
+
+// ensure a definition is emitted
+#if defined(__cplusplus)
+static void* _mi_heap_malloc_aligned = (void*)&mi_heap_malloc_aligned;
+#endif
+
+// ------------------------------------------------------
+// Aligned Allocation
+// ------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_zalloc_aligned_at(mi_heap_t* heap, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_malloc_zero_aligned_at(heap, size, alignment, offset, true);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_zalloc_aligned(mi_heap_t* heap, size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_zalloc_aligned_at(heap, size, alignment, 0);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_calloc_aligned_at(mi_heap_t* heap, size_t count, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(count, size, &total)) return NULL;
+  return mi_heap_zalloc_aligned_at(heap, total, alignment, offset);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_calloc_aligned(mi_heap_t* heap, size_t count, size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_calloc_aligned_at(heap,count,size,alignment,0);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_malloc_aligned_at(size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_malloc_aligned_at(mi_prim_get_default_heap(), size, alignment, offset);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_malloc_aligned(size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_malloc_aligned(mi_prim_get_default_heap(), size, alignment);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_zalloc_aligned_at(size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_zalloc_aligned_at(mi_prim_get_default_heap(), size, alignment, offset);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_zalloc_aligned(size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_zalloc_aligned(mi_prim_get_default_heap(), size, alignment);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_calloc_aligned_at(size_t count, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_calloc_aligned_at(mi_prim_get_default_heap(), count, size, alignment, offset);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_calloc_aligned(size_t count, size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_calloc_aligned(mi_prim_get_default_heap(), count, size, alignment);
+}
+
+
+// ------------------------------------------------------
+// Aligned re-allocation
+// ------------------------------------------------------
+
+static void* mi_heap_realloc_zero_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset, bool zero) mi_attr_noexcept {
+  mi_assert(alignment > 0);
+  if (alignment <= sizeof(uintptr_t)) return _mi_heap_realloc_zero(heap,p,newsize,zero);
+  if (p == NULL) return mi_heap_malloc_zero_aligned_at(heap,newsize,alignment,offset,zero);
+  size_t size = mi_usable_size(p);
+  if (newsize <= size && newsize >= (size - (size / 2))
+      && (((uintptr_t)p + offset) % alignment) == 0) {
+    return p;  // reallocation still fits, is aligned and not more than 50% waste
+  }
+  else {
+    // note: we don't zero allocate upfront so we only zero initialize the expanded part
+    void* newp = mi_heap_malloc_aligned_at(heap,newsize,alignment,offset);
+    if (newp != NULL) {
+      if (zero && newsize > size) {
+	// also set last word in the previous allocation to zero to ensure any padding is zero-initialized
+	size_t start = (size >= sizeof(intptr_t) ? size - sizeof(intptr_t) : 0);
+	_mi_memzero((uint8_t*)newp + start, newsize - start);
+      }
+      _mi_memcpy_aligned(newp, p, (newsize > size ? size : newsize));
+      mi_free(p); // only free if successful
+    }
+    return newp;
+  }
+}
+
+static void* mi_heap_realloc_zero_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, bool zero) mi_attr_noexcept {
+  mi_assert(alignment > 0);
+  if (alignment <= sizeof(uintptr_t)) return _mi_heap_realloc_zero(heap,p,newsize,zero);
+  size_t offset = ((uintptr_t)p % alignment); // use offset of previous allocation (p can be NULL)
+  return mi_heap_realloc_zero_aligned_at(heap,p,newsize,alignment,offset,zero);
+}
+
+mi_decl_nodiscard void* mi_heap_realloc_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_realloc_zero_aligned_at(heap,p,newsize,alignment,offset,false);
+}
+
+mi_decl_nodiscard void* mi_heap_realloc_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment) mi_attr_noexcept {
+  return mi_heap_realloc_zero_aligned(heap,p,newsize,alignment,false);
+}
+
+mi_decl_nodiscard void* mi_heap_rezalloc_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_realloc_zero_aligned_at(heap, p, newsize, alignment, offset, true);
+}
+
+mi_decl_nodiscard void* mi_heap_rezalloc_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment) mi_attr_noexcept {
+  return mi_heap_realloc_zero_aligned(heap, p, newsize, alignment, true);
+}
+
+mi_decl_nodiscard void* mi_heap_recalloc_aligned_at(mi_heap_t* heap, void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(newcount, size, &total)) return NULL;
+  return mi_heap_rezalloc_aligned_at(heap, p, total, alignment, offset);
+}
+
+mi_decl_nodiscard void* mi_heap_recalloc_aligned(mi_heap_t* heap, void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(newcount, size, &total)) return NULL;
+  return mi_heap_rezalloc_aligned(heap, p, total, alignment);
+}
+
+mi_decl_nodiscard void* mi_realloc_aligned_at(void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_realloc_aligned_at(mi_prim_get_default_heap(), p, newsize, alignment, offset);
+}
+
+mi_decl_nodiscard void* mi_realloc_aligned(void* p, size_t newsize, size_t alignment) mi_attr_noexcept {
+  return mi_heap_realloc_aligned(mi_prim_get_default_heap(), p, newsize, alignment);
+}
+
+mi_decl_nodiscard void* mi_rezalloc_aligned_at(void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_rezalloc_aligned_at(mi_prim_get_default_heap(), p, newsize, alignment, offset);
+}
+
+mi_decl_nodiscard void* mi_rezalloc_aligned(void* p, size_t newsize, size_t alignment) mi_attr_noexcept {
+  return mi_heap_rezalloc_aligned(mi_prim_get_default_heap(), p, newsize, alignment);
+}
+
+mi_decl_nodiscard void* mi_recalloc_aligned_at(void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_recalloc_aligned_at(mi_prim_get_default_heap(), p, newcount, size, alignment, offset);
+}
+
+mi_decl_nodiscard void* mi_recalloc_aligned(void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_recalloc_aligned(mi_prim_get_default_heap(), p, newcount, size, alignment);
+}
diff --git a/compat/mimalloc/alloc.c b/compat/mimalloc/alloc.c
new file mode 100644
index 00000000000000..ae272c1fb54504
--- /dev/null
+++ b/compat/mimalloc/alloc.c
@@ -0,0 +1,1056 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2022, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#ifndef _DEFAULT_SOURCE
+#define _DEFAULT_SOURCE   // for realpath() on Linux
+#endif
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"   // _mi_prim_thread_id()
+
+#include <string.h>      // memset, strlen (for mi_strdup)
+#include <stdlib.h>      // malloc, abort
+
+// ------------------------------------------------------
+// Allocation
+// ------------------------------------------------------
+
+// Fast allocation in a page: just pop from the free list.
+// Fall back to generic allocation only if the list is empty.
+extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size, bool zero) mi_attr_noexcept {
+  mi_assert_internal(page->xblock_size==0||mi_page_block_size(page) >= size);
+  mi_block_t* const block = page->free;
+  if mi_unlikely(block == NULL) {
+    return _mi_malloc_generic(heap, size, zero, 0);
+  }
+  mi_assert_internal(block != NULL && _mi_ptr_page(block) == page);
+  // pop from the free list
+  page->used++;
+  page->free = mi_block_next(page, block);
+  mi_assert_internal(page->free == NULL || _mi_ptr_page(page->free) == page);
+  #if MI_DEBUG>3
+  if (page->free_is_zero) {
+    mi_assert_expensive(mi_mem_is_zero(block+1,size - sizeof(*block)));
+  }
+  #endif
+
+  // allow use of the block internally
+  // note: when tracking we need to avoid ever touching the MI_PADDING since
+  // that is tracked by valgrind etc. as non-accessible (through the red-zone, see `mimalloc/track.h`)
+  mi_track_mem_undefined(block, mi_page_usable_block_size(page));
+
+  // zero the block? note: we need to zero the full block size (issue #63)
+  if mi_unlikely(zero) {
+    mi_assert_internal(page->xblock_size != 0); // do not call with zero'ing for huge blocks (see _mi_malloc_generic)
+    mi_assert_internal(page->xblock_size >= MI_PADDING_SIZE);
+    if (page->free_is_zero) {
+      block->next = 0;
+      mi_track_mem_defined(block, page->xblock_size - MI_PADDING_SIZE);
+    }
+    else {
+      _mi_memzero_aligned(block, page->xblock_size - MI_PADDING_SIZE);
+    }
+  }
+
+#if (MI_DEBUG>0) && !MI_TRACK_ENABLED && !MI_TSAN
+  if (!zero && !mi_page_is_huge(page)) {
+    memset(block, MI_DEBUG_UNINIT, mi_page_usable_block_size(page));
+  }
+#elif (MI_SECURE!=0)
+  if (!zero) { block->next = 0; } // don't leak internal data
+#endif
+
+#if (MI_STAT>0)
+  const size_t bsize = mi_page_usable_block_size(page);
+  if (bsize <= MI_MEDIUM_OBJ_SIZE_MAX) {
+    mi_heap_stat_increase(heap, normal, bsize);
+    mi_heap_stat_counter_increase(heap, normal_count, 1);
+#if (MI_STAT>1)
+    const size_t bin = _mi_bin(bsize);
+    mi_heap_stat_increase(heap, normal_bins[bin], 1);
+#endif
+  }
+#endif
+
+#if MI_PADDING // && !MI_TRACK_ENABLED
+  mi_padding_t* const padding = (mi_padding_t*)((uint8_t*)block + mi_page_usable_block_size(page));
+  ptrdiff_t delta = ((uint8_t*)padding - (uint8_t*)block - (size - MI_PADDING_SIZE));
+  #if (MI_DEBUG>=2)
+  mi_assert_internal(delta >= 0 && mi_page_usable_block_size(page) >= (size - MI_PADDING_SIZE + delta));
+  #endif
+  mi_track_mem_defined(padding,sizeof(mi_padding_t));  // note: re-enable since mi_page_usable_block_size may set noaccess
+  padding->canary = (uint32_t)(mi_ptr_encode(page,block,page->keys));
+  padding->delta  = (uint32_t)(delta);
+  #if MI_PADDING_CHECK
+  if (!mi_page_is_huge(page)) {
+    uint8_t* fill = (uint8_t*)padding - delta;
+    const size_t maxpad = (delta > MI_MAX_ALIGN_SIZE ? MI_MAX_ALIGN_SIZE : delta); // set at most N initial padding bytes
+    for (size_t i = 0; i < maxpad; i++) { fill[i] = MI_DEBUG_PADDING; }
+  }
+  #endif
+#endif
+
+  return block;
+}
+
+static inline mi_decl_restrict void* mi_heap_malloc_small_zero(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept {
+  mi_assert(heap != NULL);
+  #if MI_DEBUG
+  const uintptr_t tid = _mi_thread_id();
+  mi_assert(heap->thread_id == 0 || heap->thread_id == tid); // heaps are thread local
+  #endif
+  mi_assert(size <= MI_SMALL_SIZE_MAX);
+  #if (MI_PADDING)
+  if (size == 0) { size = sizeof(void*); }
+  #endif
+  mi_page_t* page = _mi_heap_get_free_small_page(heap, size + MI_PADDING_SIZE);
+  void* const p = _mi_page_malloc(heap, page, size + MI_PADDING_SIZE, zero);
+  mi_track_malloc(p,size,zero);
+  #if MI_STAT>1
+  if (p != NULL) {
+    if (!mi_heap_is_initialized(heap)) { heap = mi_prim_get_default_heap(); }
+    mi_heap_stat_increase(heap, malloc, mi_usable_size(p));
+  }
+  #endif
+  #if MI_DEBUG>3
+  if (p != NULL && zero) {
+    mi_assert_expensive(mi_mem_is_zero(p, size));
+  }
+  #endif
+  return p;
+}
+
+// allocate a small block
+mi_decl_nodiscard extern inline mi_decl_restrict void* mi_heap_malloc_small(mi_heap_t* heap, size_t size) mi_attr_noexcept {
+  return mi_heap_malloc_small_zero(heap, size, false);
+}
+
+mi_decl_nodiscard extern inline mi_decl_restrict void* mi_malloc_small(size_t size) mi_attr_noexcept {
+  return mi_heap_malloc_small(mi_prim_get_default_heap(), size);
+}
+
+// The main allocation function
+extern inline void* _mi_heap_malloc_zero_ex(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept {
+  if mi_likely(size <= MI_SMALL_SIZE_MAX) {
+    mi_assert_internal(huge_alignment == 0);
+    return mi_heap_malloc_small_zero(heap, size, zero);
+  }
+  else {
+    mi_assert(heap!=NULL);
+    mi_assert(heap->thread_id == 0 || heap->thread_id == _mi_thread_id());   // heaps are thread local
+    void* const p = _mi_malloc_generic(heap, size + MI_PADDING_SIZE, zero, huge_alignment);  // note: size can overflow but it is detected in malloc_generic
+    mi_track_malloc(p,size,zero);
+    #if MI_STAT>1
+    if (p != NULL) {
+      if (!mi_heap_is_initialized(heap)) { heap = mi_prim_get_default_heap(); }
+      mi_heap_stat_increase(heap, malloc, mi_usable_size(p));
+    }
+    #endif
+    #if MI_DEBUG>3
+    if (p != NULL && zero) {
+      mi_assert_expensive(mi_mem_is_zero(p, size));
+    }
+    #endif
+    return p;
+  }
+}
+
+extern inline void* _mi_heap_malloc_zero(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept {
+  return _mi_heap_malloc_zero_ex(heap, size, zero, 0);
+}
+
+mi_decl_nodiscard extern inline mi_decl_restrict void* mi_heap_malloc(mi_heap_t* heap, size_t size) mi_attr_noexcept {
+  return _mi_heap_malloc_zero(heap, size, false);
+}
+
+mi_decl_nodiscard extern inline mi_decl_restrict void* mi_malloc(size_t size) mi_attr_noexcept {
+  return mi_heap_malloc(mi_prim_get_default_heap(), size);
+}
+
+// zero initialized small block
+mi_decl_nodiscard mi_decl_restrict void* mi_zalloc_small(size_t size) mi_attr_noexcept {
+  return mi_heap_malloc_small_zero(mi_prim_get_default_heap(), size, true);
+}
+
+mi_decl_nodiscard extern inline mi_decl_restrict void* mi_heap_zalloc(mi_heap_t* heap, size_t size) mi_attr_noexcept {
+  return _mi_heap_malloc_zero(heap, size, true);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_zalloc(size_t size) mi_attr_noexcept {
+  return mi_heap_zalloc(mi_prim_get_default_heap(),size);
+}
+
+
+// ------------------------------------------------------
+// Check for double free in secure and debug mode
+// This is somewhat expensive so only enabled for secure mode 4
+// ------------------------------------------------------
+
+#if (MI_ENCODE_FREELIST && (MI_SECURE>=4 || MI_DEBUG!=0))
+// linear check if the free list contains a specific element
+static bool mi_list_contains(const mi_page_t* page, const mi_block_t* list, const mi_block_t* elem) {
+  while (list != NULL) {
+    if (elem==list) return true;
+    list = mi_block_next(page, list);
+  }
+  return false;
+}
+
+static mi_decl_noinline bool mi_check_is_double_freex(const mi_page_t* page, const mi_block_t* block) {
+  // The decoded value is in the same page (or NULL).
+  // Walk the free lists to verify positively if it is already freed
+  if (mi_list_contains(page, page->free, block) ||
+      mi_list_contains(page, page->local_free, block) ||
+      mi_list_contains(page, mi_page_thread_free(page), block))
+  {
+    _mi_error_message(EAGAIN, "double free detected of block %p with size %zu\n", block, mi_page_block_size(page));
+    return true;
+  }
+  return false;
+}
+
+#define mi_track_page(page,access)  { size_t psize; void* pstart = _mi_page_start(_mi_page_segment(page),page,&psize); mi_track_mem_##access( pstart, psize); }
+
+static inline bool mi_check_is_double_free(const mi_page_t* page, const mi_block_t* block) {
+  bool is_double_free = false;
+  mi_block_t* n = mi_block_nextx(page, block, page->keys); // pretend it is freed, and get the decoded first field
+  if (((uintptr_t)n & (MI_INTPTR_SIZE-1))==0 &&  // quick check: aligned pointer?
+      (n==NULL || mi_is_in_same_page(block, n))) // quick check: in same page or NULL?
+  {
+    // Suspicous: decoded value a in block is in the same page (or NULL) -- maybe a double free?
+    // (continue in separate function to improve code generation)
+    is_double_free = mi_check_is_double_freex(page, block);
+  }
+  return is_double_free;
+}
+#else
+static inline bool mi_check_is_double_free(const mi_page_t* page, const mi_block_t* block) {
+  MI_UNUSED(page);
+  MI_UNUSED(block);
+  return false;
+}
+#endif
+
+// ---------------------------------------------------------------------------
+// Check for heap block overflow by setting up padding at the end of the block
+// ---------------------------------------------------------------------------
+
+#if MI_PADDING // && !MI_TRACK_ENABLED
+static bool mi_page_decode_padding(const mi_page_t* page, const mi_block_t* block, size_t* delta, size_t* bsize) {
+  *bsize = mi_page_usable_block_size(page);
+  const mi_padding_t* const padding = (mi_padding_t*)((uint8_t*)block + *bsize);
+  mi_track_mem_defined(padding,sizeof(mi_padding_t));
+  *delta = padding->delta;
+  uint32_t canary = padding->canary;
+  uintptr_t keys[2];
+  keys[0] = page->keys[0];
+  keys[1] = page->keys[1];
+  bool ok = ((uint32_t)mi_ptr_encode(page,block,keys) == canary && *delta <= *bsize);
+  mi_track_mem_noaccess(padding,sizeof(mi_padding_t));
+  return ok;
+}
+
+// Return the exact usable size of a block.
+static size_t mi_page_usable_size_of(const mi_page_t* page, const mi_block_t* block) {
+  size_t bsize;
+  size_t delta;
+  bool ok = mi_page_decode_padding(page, block, &delta, &bsize);
+  mi_assert_internal(ok); mi_assert_internal(delta <= bsize);
+  return (ok ? bsize - delta : 0);
+}
+
+// When a non-thread-local block is freed, it becomes part of the thread delayed free
+// list that is freed later by the owning heap. If the exact usable size is too small to
+// contain the pointer for the delayed list, then shrink the padding (by decreasing delta)
+// so it will later not trigger an overflow error in `mi_free_block`.
+void _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size) {
+  size_t bsize;
+  size_t delta;
+  bool ok = mi_page_decode_padding(page, block, &delta, &bsize);
+  mi_assert_internal(ok);
+  if (!ok || (bsize - delta) >= min_size) return;  // usually already enough space
+  mi_assert_internal(bsize >= min_size);
+  if (bsize < min_size) return;  // should never happen
+  size_t new_delta = (bsize - min_size);
+  mi_assert_internal(new_delta < bsize);
+  mi_padding_t* padding = (mi_padding_t*)((uint8_t*)block + bsize);
+  mi_track_mem_defined(padding,sizeof(mi_padding_t));
+  padding->delta = (uint32_t)new_delta;
+  mi_track_mem_noaccess(padding,sizeof(mi_padding_t));
+}
+#else
+static size_t mi_page_usable_size_of(const mi_page_t* page, const mi_block_t* block) {
+  MI_UNUSED(block);
+  return mi_page_usable_block_size(page);
+}
+
+void _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size) {
+  MI_UNUSED(page);
+  MI_UNUSED(block);
+  MI_UNUSED(min_size);
+}
+#endif
+
+#if MI_PADDING && MI_PADDING_CHECK
+
+static bool mi_verify_padding(const mi_page_t* page, const mi_block_t* block, size_t* size, size_t* wrong) {
+  size_t bsize;
+  size_t delta;
+  bool ok = mi_page_decode_padding(page, block, &delta, &bsize);
+  *size = *wrong = bsize;
+  if (!ok) return false;
+  mi_assert_internal(bsize >= delta);
+  *size = bsize - delta;
+  if (!mi_page_is_huge(page)) {
+    uint8_t* fill = (uint8_t*)block + bsize - delta;
+    const size_t maxpad = (delta > MI_MAX_ALIGN_SIZE ? MI_MAX_ALIGN_SIZE : delta); // check at most the first N padding bytes
+    mi_track_mem_defined(fill, maxpad);
+    for (size_t i = 0; i < maxpad; i++) {
+      if (fill[i] != MI_DEBUG_PADDING) {
+	*wrong = bsize - delta + i;
+	ok = false;
+	break;
+      }
+    }
+    mi_track_mem_noaccess(fill, maxpad);
+  }
+  return ok;
+}
+
+static void mi_check_padding(const mi_page_t* page, const mi_block_t* block) {
+  size_t size;
+  size_t wrong;
+  if (!mi_verify_padding(page,block,&size,&wrong)) {
+    _mi_error_message(EFAULT, "buffer overflow in heap block %p of size %zu: write after %zu bytes\n", block, size, wrong );
+  }
+}
+
+#else
+
+static void mi_check_padding(const mi_page_t* page, const mi_block_t* block) {
+  MI_UNUSED(page);
+  MI_UNUSED(block);
+}
+
+#endif
+
+// only maintain stats for smaller objects if requested
+#if (MI_STAT>0)
+static void mi_stat_free(const mi_page_t* page, const mi_block_t* block) {
+  #if (MI_STAT < 2)
+  MI_UNUSED(block);
+  #endif
+  mi_heap_t* const heap = mi_heap_get_default();
+  const size_t bsize = mi_page_usable_block_size(page);
+  #if (MI_STAT>1)
+  const size_t usize = mi_page_usable_size_of(page, block);
+  mi_heap_stat_decrease(heap, malloc, usize);
+  #endif
+  if (bsize <= MI_MEDIUM_OBJ_SIZE_MAX) {
+    mi_heap_stat_decrease(heap, normal, bsize);
+    #if (MI_STAT > 1)
+    mi_heap_stat_decrease(heap, normal_bins[_mi_bin(bsize)], 1);
+    #endif
+  }
+  else if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
+    mi_heap_stat_decrease(heap, large, bsize);
+  }
+  else {
+    mi_heap_stat_decrease(heap, huge, bsize);
+  }
+}
+#else
+static void mi_stat_free(const mi_page_t* page, const mi_block_t* block) {
+  MI_UNUSED(page); MI_UNUSED(block);
+}
+#endif
+
+#if MI_HUGE_PAGE_ABANDON
+#if (MI_STAT>0)
+// maintain stats for huge objects
+static void mi_stat_huge_free(const mi_page_t* page) {
+  mi_heap_t* const heap = mi_heap_get_default();
+  const size_t bsize = mi_page_block_size(page); // to match stats in `page.c:mi_page_huge_alloc`
+  if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
+    mi_heap_stat_decrease(heap, large, bsize);
+  }
+  else {
+    mi_heap_stat_decrease(heap, huge, bsize);
+  }
+}
+#else
+static void mi_stat_huge_free(const mi_page_t* page) {
+  MI_UNUSED(page);
+}
+#endif
+#endif
+
+// ------------------------------------------------------
+// Free
+// ------------------------------------------------------
+
+// multi-threaded free (or free in huge block if compiled with MI_HUGE_PAGE_ABANDON)
+static mi_decl_noinline void _mi_free_block_mt(mi_page_t* page, mi_block_t* block)
+{
+  // The padding check may access the non-thread-owned page for the key values.
+  // that is safe as these are constant and the page won't be freed (as the block is not freed yet).
+  mi_check_padding(page, block);
+  _mi_padding_shrink(page, block, sizeof(mi_block_t));       // for small size, ensure we can fit the delayed thread pointers without triggering overflow detection
+
+  // huge page segments are always abandoned and can be freed immediately
+  mi_segment_t* segment = _mi_page_segment(page);
+  if (segment->kind == MI_SEGMENT_HUGE) {
+    #if MI_HUGE_PAGE_ABANDON
+    // huge page segments are always abandoned and can be freed immediately
+    mi_stat_huge_free(page);
+    _mi_segment_huge_page_free(segment, page, block);
+    return;
+    #else
+    // huge pages are special as they occupy the entire segment
+    // as these are large we reset the memory occupied by the page so it is available to other threads
+    // (as the owning thread needs to actually free the memory later).
+    _mi_segment_huge_page_reset(segment, page, block);
+    #endif
+  }
+
+  #if (MI_DEBUG>0) && !MI_TRACK_ENABLED && !MI_TSAN        // note: when tracking, cannot use mi_usable_size with multi-threading
+  if (segment->kind != MI_SEGMENT_HUGE) {                  // not for huge segments as we just reset the content
+    memset(block, MI_DEBUG_FREED, mi_usable_size(block));
+  }
+  #endif
+
+  // Try to put the block on either the page-local thread free list, or the heap delayed free list.
+  mi_thread_free_t tfreex;
+  bool use_delayed;
+  mi_thread_free_t tfree = mi_atomic_load_relaxed(&page->xthread_free);
+  do {
+    use_delayed = (mi_tf_delayed(tfree) == MI_USE_DELAYED_FREE);
+    if mi_unlikely(use_delayed) {
+      // unlikely: this only happens on the first concurrent free in a page that is in the full list
+      tfreex = mi_tf_set_delayed(tfree,MI_DELAYED_FREEING);
+    }
+    else {
+      // usual: directly add to page thread_free list
+      mi_block_set_next(page, block, mi_tf_block(tfree));
+      tfreex = mi_tf_set_block(tfree,block);
+    }
+  } while (!mi_atomic_cas_weak_release(&page->xthread_free, &tfree, tfreex));
+
+  if mi_unlikely(use_delayed) {
+    // racy read on `heap`, but ok because MI_DELAYED_FREEING is set (see `mi_heap_delete` and `mi_heap_collect_abandon`)
+    mi_heap_t* const heap = (mi_heap_t*)(mi_atomic_load_acquire(&page->xheap)); //mi_page_heap(page);
+    mi_assert_internal(heap != NULL);
+    if (heap != NULL) {
+      // add to the delayed free list of this heap. (do this atomically as the lock only protects heap memory validity)
+      mi_block_t* dfree = mi_atomic_load_ptr_relaxed(mi_block_t, &heap->thread_delayed_free);
+      do {
+	mi_block_set_nextx(heap,block,dfree, heap->keys);
+      } while (!mi_atomic_cas_ptr_weak_release(mi_block_t,&heap->thread_delayed_free, &dfree, block));
+    }
+
+    // and reset the MI_DELAYED_FREEING flag
+    tfree = mi_atomic_load_relaxed(&page->xthread_free);
+    do {
+      tfreex = tfree;
+      mi_assert_internal(mi_tf_delayed(tfree) == MI_DELAYED_FREEING);
+      tfreex = mi_tf_set_delayed(tfree,MI_NO_DELAYED_FREE);
+    } while (!mi_atomic_cas_weak_release(&page->xthread_free, &tfree, tfreex));
+  }
+}
+
+// regular free
+static inline void _mi_free_block(mi_page_t* page, bool local, mi_block_t* block)
+{
+  // and push it on the free list
+  //const size_t bsize = mi_page_block_size(page);
+  if mi_likely(local) {
+    // owning thread can free a block directly
+    if mi_unlikely(mi_check_is_double_free(page, block)) return;
+    mi_check_padding(page, block);
+    #if (MI_DEBUG>0) && !MI_TRACK_ENABLED && !MI_TSAN
+    if (!mi_page_is_huge(page)) {   // huge page content may be already decommitted
+      memset(block, MI_DEBUG_FREED, mi_page_block_size(page));
+    }
+    #endif
+    mi_block_set_next(page, block, page->local_free);
+    page->local_free = block;
+    page->used--;
+    if mi_unlikely(mi_page_all_free(page)) {
+      _mi_page_retire(page);
+    }
+    else if mi_unlikely(mi_page_is_in_full(page)) {
+      _mi_page_unfull(page);
+    }
+  }
+  else {
+    _mi_free_block_mt(page,block);
+  }
+}
+
+
+// Adjust a block that was allocated aligned, to the actual start of the block in the page.
+mi_block_t* _mi_page_ptr_unalign(const mi_segment_t* segment, const mi_page_t* page, const void* p) {
+  mi_assert_internal(page!=NULL && p!=NULL);
+  const size_t diff   = (uint8_t*)p - _mi_page_start(segment, page, NULL);
+  const size_t adjust = (diff % mi_page_block_size(page));
+  return (mi_block_t*)((uintptr_t)p - adjust);
+}
+
+
+void mi_decl_noinline _mi_free_generic(const mi_segment_t* segment, mi_page_t* page, bool is_local, void* p) mi_attr_noexcept {
+  mi_block_t* const block = (mi_page_has_aligned(page) ? _mi_page_ptr_unalign(segment, page, p) : (mi_block_t*)p);
+  mi_stat_free(page, block);    // stat_free may access the padding
+  mi_track_free_size(block, mi_page_usable_size_of(page,block));
+  _mi_free_block(page, is_local, block);
+}
+
+// Get the segment data belonging to a pointer
+// This is just a single `and` in assembly but does further checks in debug mode
+// (and secure mode) if this was a valid pointer.
+static inline mi_segment_t* mi_checked_ptr_segment(const void* p, const char* msg)
+{
+  MI_UNUSED(msg);
+  mi_assert(p != NULL);
+
+#if (MI_DEBUG>0)
+  if mi_unlikely(((uintptr_t)p & (MI_INTPTR_SIZE - 1)) != 0) {
+    _mi_error_message(EINVAL, "%s: invalid (unaligned) pointer: %p\n", msg, p);
+    return NULL;
+  }
+#endif
+
+  mi_segment_t* const segment = _mi_ptr_segment(p);
+  mi_assert_internal(segment != NULL);
+
+#if (MI_DEBUG>0)
+  if mi_unlikely(!mi_is_in_heap_region(p)) {
+  #if (MI_INTPTR_SIZE == 8 && defined(__linux__))
+    if (((uintptr_t)p >> 40) != 0x7F) { // linux tends to align large blocks above 0x7F000000000 (issue #640)
+  #else
+    {
+  #endif
+      _mi_warning_message("%s: pointer might not point to a valid heap region: %p\n"
+	"(this may still be a valid very large allocation (over 64MiB))\n", msg, p);
+      if mi_likely(_mi_ptr_cookie(segment) == segment->cookie) {
+	_mi_warning_message("(yes, the previous pointer %p was valid after all)\n", p);
+      }
+    }
+  }
+#endif
+#if (MI_DEBUG>0 || MI_SECURE>=4)
+  if mi_unlikely(_mi_ptr_cookie(segment) != segment->cookie) {
+    _mi_error_message(EINVAL, "%s: pointer does not point to a valid heap space: %p\n", msg, p);
+    return NULL;
+  }
+#endif
+
+  return segment;
+}
+
+// Free a block
+// fast path written carefully to prevent spilling on the stack
+void mi_free(void* p) mi_attr_noexcept
+{
+  if mi_unlikely(p == NULL) return;
+  mi_segment_t* const segment = mi_checked_ptr_segment(p,"mi_free");
+  const bool          is_local= (_mi_prim_thread_id() == mi_atomic_load_relaxed(&segment->thread_id));
+  mi_page_t* const    page    = _mi_segment_page_of(segment, p);
+
+  if mi_likely(is_local) {                       // thread-local free?
+    if mi_likely(page->flags.full_aligned == 0)  // and it is not a full page (full pages need to move from the full bin), nor has aligned blocks (aligned blocks need to be unaligned)
+    {
+      mi_block_t* const block = (mi_block_t*)p;
+      if mi_unlikely(mi_check_is_double_free(page, block)) return;
+      mi_check_padding(page, block);
+      mi_stat_free(page, block);
+      #if (MI_DEBUG>0) && !MI_TRACK_ENABLED  && !MI_TSAN
+      memset(block, MI_DEBUG_FREED, mi_page_block_size(page));
+      #endif
+      mi_track_free_size(p, mi_page_usable_size_of(page,block)); // faster then mi_usable_size as we already know the page and that p is unaligned
+      mi_block_set_next(page, block, page->local_free);
+      page->local_free = block;
+      if mi_unlikely(--page->used == 0) {   // using this expression generates better code than: page->used--; if (mi_page_all_free(page))
+	_mi_page_retire(page);
+      }
+    }
+    else {
+      // page is full or contains (inner) aligned blocks; use generic path
+      _mi_free_generic(segment, page, true, p);
+    }
+  }
+  else {
+    // not thread-local; use generic path
+    _mi_free_generic(segment, page, false, p);
+  }
+}
+
+// return true if successful
+bool _mi_free_delayed_block(mi_block_t* block) {
+  // get segment and page
+  const mi_segment_t* const segment = _mi_ptr_segment(block);
+  mi_assert_internal(_mi_ptr_cookie(segment) == segment->cookie);
+  mi_assert_internal(_mi_thread_id() == segment->thread_id);
+  mi_page_t* const page = _mi_segment_page_of(segment, block);
+
+  // Clear the no-delayed flag so delayed freeing is used again for this page.
+  // This must be done before collecting the free lists on this page -- otherwise
+  // some blocks may end up in the page `thread_free` list with no blocks in the
+  // heap `thread_delayed_free` list which may cause the page to be never freed!
+  // (it would only be freed if we happen to scan it in `mi_page_queue_find_free_ex`)
+  if (!_mi_page_try_use_delayed_free(page, MI_USE_DELAYED_FREE, false /* dont overwrite never delayed */)) {
+    return false;
+  }
+
+  // collect all other non-local frees to ensure up-to-date `used` count
+  _mi_page_free_collect(page, false);
+
+  // and free the block (possibly freeing the page as well since used is updated)
+  _mi_free_block(page, true, block);
+  return true;
+}
+
+// Bytes available in a block
+mi_decl_noinline static size_t mi_page_usable_aligned_size_of(const mi_segment_t* segment, const mi_page_t* page, const void* p) mi_attr_noexcept {
+  const mi_block_t* block = _mi_page_ptr_unalign(segment, page, p);
+  const size_t size = mi_page_usable_size_of(page, block);
+  const ptrdiff_t adjust = (uint8_t*)p - (uint8_t*)block;
+  mi_assert_internal(adjust >= 0 && (size_t)adjust <= size);
+  return (size - adjust);
+}
+
+static inline size_t _mi_usable_size(const void* p, const char* msg) mi_attr_noexcept {
+  if (p == NULL) return 0;
+  const mi_segment_t* const segment = mi_checked_ptr_segment(p, msg);
+  const mi_page_t* const page = _mi_segment_page_of(segment, p);
+  if mi_likely(!mi_page_has_aligned(page)) {
+    const mi_block_t* block = (const mi_block_t*)p;
+    return mi_page_usable_size_of(page, block);
+  }
+  else {
+    // split out to separate routine for improved code generation
+    return mi_page_usable_aligned_size_of(segment, page, p);
+  }
+}
+
+mi_decl_nodiscard size_t mi_usable_size(const void* p) mi_attr_noexcept {
+  return _mi_usable_size(p, "mi_usable_size");
+}
+
+
+// ------------------------------------------------------
+// Allocation extensions
+// ------------------------------------------------------
+
+void mi_free_size(void* p, size_t size) mi_attr_noexcept {
+  MI_UNUSED_RELEASE(size);
+  mi_assert(p == NULL || size <= _mi_usable_size(p,"mi_free_size"));
+  mi_free(p);
+}
+
+void mi_free_size_aligned(void* p, size_t size, size_t alignment) mi_attr_noexcept {
+  MI_UNUSED_RELEASE(alignment);
+  mi_assert(((uintptr_t)p % alignment) == 0);
+  mi_free_size(p,size);
+}
+
+void mi_free_aligned(void* p, size_t alignment) mi_attr_noexcept {
+  MI_UNUSED_RELEASE(alignment);
+  mi_assert(((uintptr_t)p % alignment) == 0);
+  mi_free(p);
+}
+
+mi_decl_nodiscard extern inline mi_decl_restrict void* mi_heap_calloc(mi_heap_t* heap, size_t count, size_t size) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(count,size,&total)) return NULL;
+  return mi_heap_zalloc(heap,total);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_calloc(size_t count, size_t size) mi_attr_noexcept {
+  return mi_heap_calloc(mi_prim_get_default_heap(),count,size);
+}
+
+// Uninitialized `calloc`
+mi_decl_nodiscard extern mi_decl_restrict void* mi_heap_mallocn(mi_heap_t* heap, size_t count, size_t size) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(count, size, &total)) return NULL;
+  return mi_heap_malloc(heap, total);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_mallocn(size_t count, size_t size) mi_attr_noexcept {
+  return mi_heap_mallocn(mi_prim_get_default_heap(),count,size);
+}
+
+// Expand (or shrink) in place (or fail)
+void* mi_expand(void* p, size_t newsize) mi_attr_noexcept {
+  #if MI_PADDING
+  // we do not shrink/expand with padding enabled
+  MI_UNUSED(p); MI_UNUSED(newsize);
+  return NULL;
+  #else
+  if (p == NULL) return NULL;
+  const size_t size = _mi_usable_size(p,"mi_expand");
+  if (newsize > size) return NULL;
+  return p; // it fits
+  #endif
+}
+
+void* _mi_heap_realloc_zero(mi_heap_t* heap, void* p, size_t newsize, bool zero) mi_attr_noexcept {
+  // if p == NULL then behave as malloc.
+  // else if size == 0 then reallocate to a zero-sized block (and don't return NULL, just as mi_malloc(0)).
+  // (this means that returning NULL always indicates an error, and `p` will not have been freed in that case.)
+  const size_t size = _mi_usable_size(p,"mi_realloc"); // also works if p == NULL (with size 0)
+  if mi_unlikely(newsize <= size && newsize >= (size / 2) && newsize > 0) {  // note: newsize must be > 0 or otherwise we return NULL for realloc(NULL,0)
+    mi_assert_internal(p!=NULL);
+    // todo: do not track as the usable size is still the same in the free; adjust potential padding?
+    // mi_track_resize(p,size,newsize)
+    // if (newsize < size) { mi_track_mem_noaccess((uint8_t*)p + newsize, size - newsize); }
+    return p;  // reallocation still fits and not more than 50% waste
+  }
+  void* newp = mi_heap_malloc(heap,newsize);
+  if mi_likely(newp != NULL) {
+    if (zero && newsize > size) {
+      // also set last word in the previous allocation to zero to ensure any padding is zero-initialized
+      const size_t start = (size >= sizeof(intptr_t) ? size - sizeof(intptr_t) : 0);
+      _mi_memzero((uint8_t*)newp + start, newsize - start);
+    }
+    else if (newsize == 0) {
+      ((uint8_t*)newp)[0] = 0; // work around for applications that expect zero-reallocation to be zero initialized (issue #725)
+    }
+    if mi_likely(p != NULL) {
+      const size_t copysize = (newsize > size ? size : newsize);
+      mi_track_mem_defined(p,copysize);  // _mi_useable_size may be too large for byte precise memory tracking..
+      _mi_memcpy(newp, p, copysize);
+      mi_free(p); // only free the original pointer if successful
+    }
+  }
+  return newp;
+}
+
+mi_decl_nodiscard void* mi_heap_realloc(mi_heap_t* heap, void* p, size_t newsize) mi_attr_noexcept {
+  return _mi_heap_realloc_zero(heap, p, newsize, false);
+}
+
+mi_decl_nodiscard void* mi_heap_reallocn(mi_heap_t* heap, void* p, size_t count, size_t size) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(count, size, &total)) return NULL;
+  return mi_heap_realloc(heap, p, total);
+}
+
+
+// Reallocate but free `p` on errors
+mi_decl_nodiscard void* mi_heap_reallocf(mi_heap_t* heap, void* p, size_t newsize) mi_attr_noexcept {
+  void* newp = mi_heap_realloc(heap, p, newsize);
+  if (newp==NULL && p!=NULL) mi_free(p);
+  return newp;
+}
+
+mi_decl_nodiscard void* mi_heap_rezalloc(mi_heap_t* heap, void* p, size_t newsize) mi_attr_noexcept {
+  return _mi_heap_realloc_zero(heap, p, newsize, true);
+}
+
+mi_decl_nodiscard void* mi_heap_recalloc(mi_heap_t* heap, void* p, size_t count, size_t size) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(count, size, &total)) return NULL;
+  return mi_heap_rezalloc(heap, p, total);
+}
+
+
+mi_decl_nodiscard void* mi_realloc(void* p, size_t newsize) mi_attr_noexcept {
+  return mi_heap_realloc(mi_prim_get_default_heap(),p,newsize);
+}
+
+mi_decl_nodiscard void* mi_reallocn(void* p, size_t count, size_t size) mi_attr_noexcept {
+  return mi_heap_reallocn(mi_prim_get_default_heap(),p,count,size);
+}
+
+// Reallocate but free `p` on errors
+mi_decl_nodiscard void* mi_reallocf(void* p, size_t newsize) mi_attr_noexcept {
+  return mi_heap_reallocf(mi_prim_get_default_heap(),p,newsize);
+}
+
+mi_decl_nodiscard void* mi_rezalloc(void* p, size_t newsize) mi_attr_noexcept {
+  return mi_heap_rezalloc(mi_prim_get_default_heap(), p, newsize);
+}
+
+mi_decl_nodiscard void* mi_recalloc(void* p, size_t count, size_t size) mi_attr_noexcept {
+  return mi_heap_recalloc(mi_prim_get_default_heap(), p, count, size);
+}
+
+
+
+// ------------------------------------------------------
+// strdup, strndup, and realpath
+// ------------------------------------------------------
+
+// `strdup` using mi_malloc
+mi_decl_nodiscard mi_decl_restrict char* mi_heap_strdup(mi_heap_t* heap, const char* s) mi_attr_noexcept {
+  if (s == NULL) return NULL;
+  size_t n = strlen(s);
+  char* t = (char*)mi_heap_malloc(heap,n+1);
+  if (t == NULL) return NULL;
+  _mi_memcpy(t, s, n);
+  t[n] = 0;
+  return t;
+}
+
+mi_decl_nodiscard mi_decl_restrict char* mi_strdup(const char* s) mi_attr_noexcept {
+  return mi_heap_strdup(mi_prim_get_default_heap(), s);
+}
+
+// `strndup` using mi_malloc
+mi_decl_nodiscard mi_decl_restrict char* mi_heap_strndup(mi_heap_t* heap, const char* s, size_t n) mi_attr_noexcept {
+  if (s == NULL) return NULL;
+  const char* end = (const char*)memchr(s, 0, n);  // find end of string in the first `n` characters (returns NULL if not found)
+  const size_t m = (end != NULL ? (size_t)(end - s) : n);  // `m` is the minimum of `n` or the end-of-string
+  mi_assert_internal(m <= n);
+  char* t = (char*)mi_heap_malloc(heap, m+1);
+  if (t == NULL) return NULL;
+  _mi_memcpy(t, s, m);
+  t[m] = 0;
+  return t;
+}
+
+mi_decl_nodiscard mi_decl_restrict char* mi_strndup(const char* s, size_t n) mi_attr_noexcept {
+  return mi_heap_strndup(mi_prim_get_default_heap(),s,n);
+}
+
+#ifndef __wasi__
+// `realpath` using mi_malloc
+#ifdef _WIN32
+#ifndef PATH_MAX
+#define PATH_MAX MAX_PATH
+#endif
+#include <windows.h>
+mi_decl_nodiscard mi_decl_restrict char* mi_heap_realpath(mi_heap_t* heap, const char* fname, char* resolved_name) mi_attr_noexcept {
+  // todo: use GetFullPathNameW to allow longer file names
+  char buf[PATH_MAX];
+  DWORD res = GetFullPathNameA(fname, PATH_MAX, (resolved_name == NULL ? buf : resolved_name), NULL);
+  if (res == 0) {
+    errno = GetLastError(); return NULL;
+  }
+  else if (res > PATH_MAX) {
+    errno = EINVAL; return NULL;
+  }
+  else if (resolved_name != NULL) {
+    return resolved_name;
+  }
+  else {
+    return mi_heap_strndup(heap, buf, PATH_MAX);
+  }
+}
+#else
+/*
+#include <unistd.h>  // pathconf
+static size_t mi_path_max(void) {
+  static size_t path_max = 0;
+  if (path_max <= 0) {
+    long m = pathconf("/",_PC_PATH_MAX);
+    if (m <= 0) path_max = 4096;      // guess
+    else if (m < 256) path_max = 256; // at least 256
+    else path_max = m;
+  }
+  return path_max;
+}
+*/
+char* mi_heap_realpath(mi_heap_t* heap, const char* fname, char* resolved_name) mi_attr_noexcept {
+  if (resolved_name != NULL) {
+    return realpath(fname,resolved_name);
+  }
+  else {
+    char* rname = realpath(fname, NULL);
+    if (rname == NULL) return NULL;
+    char* result = mi_heap_strdup(heap, rname);
+    free(rname);  // use regular free! (which may be redirected to our free but that's ok)
+    return result;
+  }
+  /*
+    const size_t n  = mi_path_max();
+    char* buf = (char*)mi_malloc(n+1);
+    if (buf == NULL) {
+      errno = ENOMEM;
+      return NULL;
+    }
+    char* rname  = realpath(fname,buf);
+    char* result = mi_heap_strndup(heap,rname,n); // ok if `rname==NULL`
+    mi_free(buf);
+    return result;
+  }
+  */
+}
+#endif
+
+mi_decl_nodiscard mi_decl_restrict char* mi_realpath(const char* fname, char* resolved_name) mi_attr_noexcept {
+  return mi_heap_realpath(mi_prim_get_default_heap(),fname,resolved_name);
+}
+#endif
+
+/*-------------------------------------------------------
+C++ new and new_aligned
+The standard requires calling into `get_new_handler` and
+throwing the bad_alloc exception on failure. If we compile
+with a C++ compiler we can implement this precisely. If we
+use a C compiler we cannot throw a `bad_alloc` exception
+but we call `exit` instead (i.e. not returning).
+-------------------------------------------------------*/
+
+#ifdef __cplusplus
+#include <new>
+static bool mi_try_new_handler(bool nothrow) {
+  #if defined(_MSC_VER) || (__cplusplus >= 201103L)
+    std::new_handler h = std::get_new_handler();
+  #else
+    std::new_handler h = std::set_new_handler();
+    std::set_new_handler(h);
+  #endif
+  if (h==NULL) {
+    _mi_error_message(ENOMEM, "out of memory in 'new'");
+    if (!nothrow) {
+      throw std::bad_alloc();
+    }
+    return false;
+  }
+  else {
+    h();
+    return true;
+  }
+}
+#else
+typedef void (*std_new_handler_t)(void);
+
+#if (defined(__GNUC__) || (defined(__clang__) && !defined(_MSC_VER)))  // exclude clang-cl, see issue #631
+std_new_handler_t __attribute__((weak)) _ZSt15get_new_handlerv(void) {
+  return NULL;
+}
+static std_new_handler_t mi_get_new_handler(void) {
+  return _ZSt15get_new_handlerv();
+}
+#else
+// note: on windows we could dynamically link to `?get_new_handler@std@@YAP6AXXZXZ`.
+static std_new_handler_t mi_get_new_handler() {
+  return NULL;
+}
+#endif
+
+static bool mi_try_new_handler(bool nothrow) {
+  std_new_handler_t h = mi_get_new_handler();
+  if (h==NULL) {
+    _mi_error_message(ENOMEM, "out of memory in 'new'");
+    if (!nothrow) {
+      abort();  // cannot throw in plain C, use abort
+    }
+    return false;
+  }
+  else {
+    h();
+    return true;
+  }
+}
+#endif
+
+mi_decl_export mi_decl_noinline void* mi_heap_try_new(mi_heap_t* heap, size_t size, bool nothrow ) {
+  void* p = NULL;
+  while(p == NULL && mi_try_new_handler(nothrow)) {
+    p = mi_heap_malloc(heap,size);
+  }
+  return p;
+}
+
+static mi_decl_noinline void* mi_try_new(size_t size, bool nothrow) {
+  return mi_heap_try_new(mi_prim_get_default_heap(), size, nothrow);
+}
+
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_alloc_new(mi_heap_t* heap, size_t size) {
+  void* p = mi_heap_malloc(heap,size);
+  if mi_unlikely(p == NULL) return mi_heap_try_new(heap, size, false);
+  return p;
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_new(size_t size) {
+  return mi_heap_alloc_new(mi_prim_get_default_heap(), size);
+}
+
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_alloc_new_n(mi_heap_t* heap, size_t count, size_t size) {
+  size_t total;
+  if mi_unlikely(mi_count_size_overflow(count, size, &total)) {
+    mi_try_new_handler(false);  // on overflow we invoke the try_new_handler once to potentially throw std::bad_alloc
+    return NULL;
+  }
+  else {
+    return mi_heap_alloc_new(heap,total);
+  }
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_new_n(size_t count, size_t size) {
+  return mi_heap_alloc_new_n(mi_prim_get_default_heap(), size, count);
+}
+
+
+mi_decl_nodiscard mi_decl_restrict void* mi_new_nothrow(size_t size) mi_attr_noexcept {
+  void* p = mi_malloc(size);
+  if mi_unlikely(p == NULL) return mi_try_new(size, true);
+  return p;
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_new_aligned(size_t size, size_t alignment) {
+  void* p;
+  do {
+    p = mi_malloc_aligned(size, alignment);
+  }
+  while(p == NULL && mi_try_new_handler(false));
+  return p;
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_new_aligned_nothrow(size_t size, size_t alignment) mi_attr_noexcept {
+  void* p;
+  do {
+    p = mi_malloc_aligned(size, alignment);
+  }
+  while(p == NULL && mi_try_new_handler(true));
+  return p;
+}
+
+mi_decl_nodiscard void* mi_new_realloc(void* p, size_t newsize) {
+  void* q;
+  do {
+    q = mi_realloc(p, newsize);
+  } while (q == NULL && mi_try_new_handler(false));
+  return q;
+}
+
+mi_decl_nodiscard void* mi_new_reallocn(void* p, size_t newcount, size_t size) {
+  size_t total;
+  if mi_unlikely(mi_count_size_overflow(newcount, size, &total)) {
+    mi_try_new_handler(false);  // on overflow we invoke the try_new_handler once to potentially throw std::bad_alloc
+    return NULL;
+  }
+  else {
+    return mi_new_realloc(p, total);
+  }
+}
+
+// ------------------------------------------------------
+// ensure explicit external inline definitions are emitted!
+// ------------------------------------------------------
+
+#ifdef __cplusplus
+void* _mi_externs[] = {
+  (void*)&_mi_page_malloc,
+  (void*)&_mi_heap_malloc_zero,
+  (void*)&_mi_heap_malloc_zero_ex,
+  (void*)&mi_malloc,
+  (void*)&mi_malloc_small,
+  (void*)&mi_zalloc_small,
+  (void*)&mi_heap_malloc,
+  (void*)&mi_heap_zalloc,
+  (void*)&mi_heap_malloc_small,
+  // (void*)&mi_heap_alloc_new,
+  // (void*)&mi_heap_alloc_new_n
+};
+#endif
diff --git a/compat/mimalloc/arena.c b/compat/mimalloc/arena.c
new file mode 100644
index 00000000000000..879ee9e7e773d4
--- /dev/null
+++ b/compat/mimalloc/arena.c
@@ -0,0 +1,935 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* ----------------------------------------------------------------------------
+"Arenas" are fixed area's of OS memory from which we can allocate
+large blocks (>= MI_ARENA_MIN_BLOCK_SIZE, 4MiB).
+In contrast to the rest of mimalloc, the arenas are shared between
+threads and need to be accessed using atomic operations.
+
+Arenas are used to for huge OS page (1GiB) reservations or for reserving
+OS memory upfront which can be improve performance or is sometimes needed
+on embedded devices. We can also employ this with WASI or `sbrk` systems
+to reserve large arenas upfront and be able to reuse the memory more effectively.
+
+The arena allocation needs to be thread safe and we use an atomic bitmap to allocate.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+
+#include <string.h>  // memset
+#include <errno.h>   // ENOMEM
+
+#include "bitmap.h"  // atomic bitmap
+
+/* -----------------------------------------------------------
+  Arena allocation
+----------------------------------------------------------- */
+
+// Block info: bit 0 contains the `in_use` bit, the upper bits the
+// size in count of arena blocks.
+typedef uintptr_t mi_block_info_t;
+#define MI_ARENA_BLOCK_SIZE   (MI_SEGMENT_SIZE)        // 64MiB  (must be at least MI_SEGMENT_ALIGN)
+#define MI_ARENA_MIN_OBJ_SIZE (MI_ARENA_BLOCK_SIZE/2)  // 32MiB
+#define MI_MAX_ARENAS         (112)                    // not more than 126 (since we use 7 bits in the memid and an arena index + 1)
+
+// A memory arena descriptor
+typedef struct mi_arena_s {
+  mi_arena_id_t id;                       // arena id; 0 for non-specific
+  mi_memid_t memid;                       // memid of the memory area
+  _Atomic(uint8_t*) start;                // the start of the memory area
+  size_t   block_count;                   // size of the area in arena blocks (of `MI_ARENA_BLOCK_SIZE`)
+  size_t   field_count;                   // number of bitmap fields (where `field_count * MI_BITMAP_FIELD_BITS >= block_count`)
+  size_t   meta_size;                     // size of the arena structure itself (including its bitmaps)
+  mi_memid_t meta_memid;                  // memid of the arena structure itself (OS or static allocation)
+  int      numa_node;                     // associated NUMA node
+  bool     exclusive;                     // only allow allocations if specifically for this arena
+  bool     is_large;                      // memory area consists of large- or huge OS pages (always committed)
+  _Atomic(size_t) search_idx;             // optimization to start the search for free blocks
+  _Atomic(mi_msecs_t) purge_expire;       // expiration time when blocks should be decommitted from `blocks_decommit`.
+  mi_bitmap_field_t* blocks_dirty;        // are the blocks potentially non-zero?
+  mi_bitmap_field_t* blocks_committed;    // are the blocks committed? (can be NULL for memory that cannot be decommitted)
+  mi_bitmap_field_t* blocks_purge;        // blocks that can be (reset) decommitted. (can be NULL for memory that cannot be (reset) decommitted)
+  mi_bitmap_field_t  blocks_inuse[1];     // in-place bitmap of in-use blocks (of size `field_count`)
+} mi_arena_t;
+
+
+// The available arenas
+static mi_decl_cache_align _Atomic(mi_arena_t*) mi_arenas[MI_MAX_ARENAS];
+static mi_decl_cache_align _Atomic(size_t)      mi_arena_count; // = 0
+
+
+//static bool mi_manage_os_memory_ex2(void* start, size_t size, bool is_large, int numa_node, bool exclusive, mi_memid_t memid, mi_arena_id_t* arena_id) mi_attr_noexcept;
+
+/* -----------------------------------------------------------
+  Arena id's
+  id = arena_index + 1
+----------------------------------------------------------- */
+
+static size_t mi_arena_id_index(mi_arena_id_t id) {
+  return (size_t)(id <= 0 ? MI_MAX_ARENAS : id - 1);
+}
+
+static mi_arena_id_t mi_arena_id_create(size_t arena_index) {
+  mi_assert_internal(arena_index < MI_MAX_ARENAS);
+  return (int)arena_index + 1;
+}
+
+mi_arena_id_t _mi_arena_id_none(void) {
+  return 0;
+}
+
+static bool mi_arena_id_is_suitable(mi_arena_id_t arena_id, bool arena_is_exclusive, mi_arena_id_t req_arena_id) {
+  return ((!arena_is_exclusive && req_arena_id == _mi_arena_id_none()) ||
+	  (arena_id == req_arena_id));
+}
+
+bool _mi_arena_memid_is_suitable(mi_memid_t memid, mi_arena_id_t request_arena_id) {
+  if (memid.memkind == MI_MEM_ARENA) {
+    return mi_arena_id_is_suitable(memid.mem.arena.id, memid.mem.arena.is_exclusive, request_arena_id);
+  }
+  else {
+    return mi_arena_id_is_suitable(0, false, request_arena_id);
+  }
+}
+
+bool _mi_arena_memid_is_os_allocated(mi_memid_t memid) {
+  return (memid.memkind == MI_MEM_OS);
+}
+
+/* -----------------------------------------------------------
+  Arena allocations get a (currently) 16-bit memory id where the
+  lower 8 bits are the arena id, and the upper bits the block index.
+----------------------------------------------------------- */
+
+static size_t mi_block_count_of_size(size_t size) {
+  return _mi_divide_up(size, MI_ARENA_BLOCK_SIZE);
+}
+
+static size_t mi_arena_block_size(size_t bcount) {
+  return (bcount * MI_ARENA_BLOCK_SIZE);
+}
+
+static size_t mi_arena_size(mi_arena_t* arena) {
+  return mi_arena_block_size(arena->block_count);
+}
+
+static mi_memid_t mi_memid_create_arena(mi_arena_id_t id, bool is_exclusive, mi_bitmap_index_t bitmap_index) {
+  mi_memid_t memid = _mi_memid_create(MI_MEM_ARENA);
+  memid.mem.arena.id = id;
+  memid.mem.arena.block_index = bitmap_index;
+  memid.mem.arena.is_exclusive = is_exclusive;
+  return memid;
+}
+
+static bool mi_arena_memid_indices(mi_memid_t memid, size_t* arena_index, mi_bitmap_index_t* bitmap_index) {
+  mi_assert_internal(memid.memkind == MI_MEM_ARENA);
+  *arena_index = mi_arena_id_index(memid.mem.arena.id);
+  *bitmap_index = memid.mem.arena.block_index;
+  return memid.mem.arena.is_exclusive;
+}
+
+
+
+/* -----------------------------------------------------------
+  Special static area for mimalloc internal structures
+  to avoid OS calls (for example, for the arena metadata)
+----------------------------------------------------------- */
+
+#define MI_ARENA_STATIC_MAX  (MI_INTPTR_SIZE*MI_KiB)  // 8 KiB on 64-bit
+
+static uint8_t mi_arena_static[MI_ARENA_STATIC_MAX];
+static _Atomic(size_t) mi_arena_static_top;
+
+static void* mi_arena_static_zalloc(size_t size, size_t alignment, mi_memid_t* memid) {
+  *memid = _mi_memid_none();
+  if (size == 0 || size > MI_ARENA_STATIC_MAX) return NULL;
+  if ((mi_atomic_load_relaxed(&mi_arena_static_top) + size) > MI_ARENA_STATIC_MAX) return NULL;
+
+  // try to claim space
+  if (alignment == 0) { alignment = 1; }
+  const size_t oversize = size + alignment - 1;
+  if (oversize > MI_ARENA_STATIC_MAX) return NULL;
+  const size_t oldtop = mi_atomic_add_acq_rel(&mi_arena_static_top, oversize);
+  size_t top = oldtop + oversize;
+  if (top > MI_ARENA_STATIC_MAX) {
+    // try to roll back, ok if this fails
+    mi_atomic_cas_strong_acq_rel(&mi_arena_static_top, &top, oldtop);
+    return NULL;
+  }
+
+  // success
+  *memid = _mi_memid_create(MI_MEM_STATIC);
+  const size_t start = _mi_align_up(oldtop, alignment);
+  uint8_t* const p = &mi_arena_static[start];
+  _mi_memzero(p, size);
+  return p;
+}
+
+static void* mi_arena_meta_zalloc(size_t size, mi_memid_t* memid, mi_stats_t* stats) {
+  *memid = _mi_memid_none();
+
+  // try static
+  void* p = mi_arena_static_zalloc(size, MI_ALIGNMENT_MAX, memid);
+  if (p != NULL) return p;
+
+  // or fall back to the OS
+  return _mi_os_alloc(size, memid, stats);
+}
+
+static void mi_arena_meta_free(void* p, mi_memid_t memid, size_t size, mi_stats_t* stats) {
+  if (mi_memkind_is_os(memid.memkind)) {
+    _mi_os_free(p, size, memid, stats);
+  }
+  else {
+    mi_assert(memid.memkind == MI_MEM_STATIC);
+  }
+}
+
+static void* mi_arena_block_start(mi_arena_t* arena, mi_bitmap_index_t bindex) {
+  return (arena->start + mi_arena_block_size(mi_bitmap_index_bit(bindex)));
+}
+
+
+/* -----------------------------------------------------------
+  Thread safe allocation in an arena
+----------------------------------------------------------- */
+
+// claim the `blocks_inuse` bits
+static bool mi_arena_try_claim(mi_arena_t* arena, size_t blocks, mi_bitmap_index_t* bitmap_idx)
+{
+  size_t idx = 0; // mi_atomic_load_relaxed(&arena->search_idx);  // start from last search; ok to be relaxed as the exact start does not matter
+  if (_mi_bitmap_try_find_from_claim_across(arena->blocks_inuse, arena->field_count, idx, blocks, bitmap_idx)) {
+    mi_atomic_store_relaxed(&arena->search_idx, mi_bitmap_index_field(*bitmap_idx));  // start search from found location next time around
+    return true;
+  };
+  return false;
+}
+
+
+/* -----------------------------------------------------------
+  Arena Allocation
+----------------------------------------------------------- */
+
+static mi_decl_noinline void* mi_arena_try_alloc_at(mi_arena_t* arena, size_t arena_index, size_t needed_bcount,
+						    bool commit, mi_memid_t* memid, mi_os_tld_t* tld)
+{
+  MI_UNUSED(arena_index);
+  mi_assert_internal(mi_arena_id_index(arena->id) == arena_index);
+
+  mi_bitmap_index_t bitmap_index;
+  if (!mi_arena_try_claim(arena, needed_bcount, &bitmap_index)) return NULL;
+
+  // claimed it!
+  void* p = mi_arena_block_start(arena, bitmap_index);
+  *memid = mi_memid_create_arena(arena->id, arena->exclusive, bitmap_index);
+  memid->is_pinned = arena->memid.is_pinned;
+
+  // none of the claimed blocks should be scheduled for a decommit
+  if (arena->blocks_purge != NULL) {
+    // this is thread safe as a potential purge only decommits parts that are not yet claimed as used (in `blocks_inuse`).
+    _mi_bitmap_unclaim_across(arena->blocks_purge, arena->field_count, needed_bcount, bitmap_index);
+  }
+
+  // set the dirty bits (todo: no need for an atomic op here?)
+  if (arena->memid.initially_zero && arena->blocks_dirty != NULL) {
+    memid->initially_zero = _mi_bitmap_claim_across(arena->blocks_dirty, arena->field_count, needed_bcount, bitmap_index, NULL);
+  }
+
+  // set commit state
+  if (arena->blocks_committed == NULL) {
+    // always committed
+    memid->initially_committed = true;
+  }
+  else if (commit) {
+    // commit requested, but the range may not be committed as a whole: ensure it is committed now
+    memid->initially_committed = true;
+    bool any_uncommitted;
+    _mi_bitmap_claim_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index, &any_uncommitted);
+    if (any_uncommitted) {
+      bool commit_zero = false;
+      if (!_mi_os_commit(p, mi_arena_block_size(needed_bcount), &commit_zero, tld->stats)) {
+	memid->initially_committed = false;
+      }
+      else {
+	if (commit_zero) { memid->initially_zero = true; }
+      }
+    }
+  }
+  else {
+    // no need to commit, but check if already fully committed
+    memid->initially_committed = _mi_bitmap_is_claimed_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index);
+  }
+
+  return p;
+}
+
+// allocate in a speficic arena
+static void* mi_arena_try_alloc_at_id(mi_arena_id_t arena_id, bool match_numa_node, int numa_node, size_t size, size_t alignment,
+				       bool commit, bool allow_large, mi_arena_id_t req_arena_id, mi_memid_t* memid, mi_os_tld_t* tld )
+{
+  MI_UNUSED_RELEASE(alignment);
+  mi_assert_internal(alignment <= MI_SEGMENT_ALIGN);
+  const size_t bcount = mi_block_count_of_size(size);
+  const size_t arena_index = mi_arena_id_index(arena_id);
+  mi_assert_internal(arena_index < mi_atomic_load_relaxed(&mi_arena_count));
+  mi_assert_internal(size <= mi_arena_block_size(bcount));
+
+  // Check arena suitability
+  mi_arena_t* arena = mi_atomic_load_ptr_acquire(mi_arena_t, &mi_arenas[arena_index]);
+  if (arena == NULL) return NULL;
+  if (!allow_large && arena->is_large) return NULL;
+  if (!mi_arena_id_is_suitable(arena->id, arena->exclusive, req_arena_id)) return NULL;
+  if (req_arena_id == _mi_arena_id_none()) { // in not specific, check numa affinity
+    const bool numa_suitable = (numa_node < 0 || arena->numa_node < 0 || arena->numa_node == numa_node);
+    if (match_numa_node) { if (!numa_suitable) return NULL; }
+		    else { if (numa_suitable) return NULL; }
+  }
+
+  // try to allocate
+  void* p = mi_arena_try_alloc_at(arena, arena_index, bcount, commit, memid, tld);
+  mi_assert_internal(p == NULL || _mi_is_aligned(p, alignment));
+  return p;
+}
+
+
+// allocate from an arena with fallback to the OS
+static mi_decl_noinline void* mi_arena_try_alloc(int numa_node, size_t size, size_t alignment,
+						  bool commit, bool allow_large,
+						  mi_arena_id_t req_arena_id, mi_memid_t* memid, mi_os_tld_t* tld )
+{
+  MI_UNUSED(alignment);
+  mi_assert_internal(alignment <= MI_SEGMENT_ALIGN);
+  const size_t max_arena = mi_atomic_load_relaxed(&mi_arena_count);
+  if mi_likely(max_arena == 0) return NULL;
+
+  if (req_arena_id != _mi_arena_id_none()) {
+    // try a specific arena if requested
+    if (mi_arena_id_index(req_arena_id) < max_arena) {
+      void* p = mi_arena_try_alloc_at_id(req_arena_id, true, numa_node, size, alignment, commit, allow_large, req_arena_id, memid, tld);
+      if (p != NULL) return p;
+    }
+  }
+  else {
+    // try numa affine allocation
+    for (size_t i = 0; i < max_arena; i++) {
+      void* p = mi_arena_try_alloc_at_id(mi_arena_id_create(i), true, numa_node, size, alignment, commit, allow_large, req_arena_id, memid, tld);
+      if (p != NULL) return p;
+    }
+
+    // try from another numa node instead..
+    if (numa_node >= 0) {  // if numa_node was < 0 (no specific affinity requested), all arena's have been tried already
+      for (size_t i = 0; i < max_arena; i++) {
+	void* p = mi_arena_try_alloc_at_id(mi_arena_id_create(i), false /* only proceed if not numa local */, numa_node, size, alignment, commit, allow_large, req_arena_id, memid, tld);
+	if (p != NULL) return p;
+      }
+    }
+  }
+  return NULL;
+}
+
+// try to reserve a fresh arena space
+static bool mi_arena_reserve(size_t req_size, bool allow_large, mi_arena_id_t req_arena_id, mi_arena_id_t *arena_id)
+{
+  if (_mi_preloading()) return false;  // use OS only while pre loading
+  if (req_arena_id != _mi_arena_id_none()) return false;
+
+  const size_t arena_count = mi_atomic_load_acquire(&mi_arena_count);
+  if (arena_count > (MI_MAX_ARENAS - 4)) return false;
+
+  size_t arena_reserve = mi_option_get_size(mi_option_arena_reserve);
+  if (arena_reserve == 0) return false;
+
+  if (!_mi_os_has_virtual_reserve()) {
+    arena_reserve = arena_reserve/4;  // be conservative if virtual reserve is not supported (for some embedded systems for example)
+  }
+  arena_reserve = _mi_align_up(arena_reserve, MI_ARENA_BLOCK_SIZE);
+  if (arena_count >= 8 && arena_count <= 128) {
+    arena_reserve = ((size_t)1<<(arena_count/8)) * arena_reserve;  // scale up the arena sizes exponentially
+  }
+  if (arena_reserve < req_size) return false;  // should be able to at least handle the current allocation size
+
+  // commit eagerly?
+  bool arena_commit = false;
+  if (mi_option_get(mi_option_arena_eager_commit) == 2)      { arena_commit = _mi_os_has_overcommit(); }
+  else if (mi_option_get(mi_option_arena_eager_commit) == 1) { arena_commit = true; }
+
+  return (mi_reserve_os_memory_ex(arena_reserve, arena_commit, allow_large, false /* exclusive */, arena_id) == 0);
+}
+
+
+void* _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool commit, bool allow_large,
+			      mi_arena_id_t req_arena_id, mi_memid_t* memid, mi_os_tld_t* tld)
+{
+  mi_assert_internal(memid != NULL && tld != NULL);
+  mi_assert_internal(size > 0);
+  *memid = _mi_memid_none();
+
+  const int numa_node = _mi_os_numa_node(tld); // current numa node
+
+  // try to allocate in an arena if the alignment is small enough and the object is not too small (as for heap meta data)
+  if (size >= MI_ARENA_MIN_OBJ_SIZE && alignment <= MI_SEGMENT_ALIGN && align_offset == 0) {
+    void* p = mi_arena_try_alloc(numa_node, size, alignment, commit, allow_large, req_arena_id, memid, tld);
+    if (p != NULL) return p;
+
+    // otherwise, try to first eagerly reserve a new arena
+    if (req_arena_id == _mi_arena_id_none()) {
+      mi_arena_id_t arena_id = 0;
+      if (mi_arena_reserve(size, allow_large, req_arena_id, &arena_id)) {
+	// and try allocate in there
+	mi_assert_internal(req_arena_id == _mi_arena_id_none());
+	p = mi_arena_try_alloc_at_id(arena_id, true, numa_node, size, alignment, commit, allow_large, req_arena_id, memid, tld);
+	if (p != NULL) return p;
+      }
+    }
+  }
+
+  // if we cannot use OS allocation, return NULL
+  if (mi_option_is_enabled(mi_option_limit_os_alloc) || req_arena_id != _mi_arena_id_none()) {
+    errno = ENOMEM;
+    return NULL;
+  }
+
+  // finally, fall back to the OS
+  if (align_offset > 0) {
+    return _mi_os_alloc_aligned_at_offset(size, alignment, align_offset, commit, allow_large, memid, tld->stats);
+  }
+  else {
+    return _mi_os_alloc_aligned(size, alignment, commit, allow_large, memid, tld->stats);
+  }
+}
+
+void* _mi_arena_alloc(size_t size, bool commit, bool allow_large, mi_arena_id_t req_arena_id, mi_memid_t* memid, mi_os_tld_t* tld)
+{
+  return _mi_arena_alloc_aligned(size, MI_ARENA_BLOCK_SIZE, 0, commit, allow_large, req_arena_id, memid, tld);
+}
+
+
+void* mi_arena_area(mi_arena_id_t arena_id, size_t* size) {
+  if (size != NULL) *size = 0;
+  size_t arena_index = mi_arena_id_index(arena_id);
+  if (arena_index >= MI_MAX_ARENAS) return NULL;
+  mi_arena_t* arena = mi_atomic_load_ptr_acquire(mi_arena_t, &mi_arenas[arena_index]);
+  if (arena == NULL) return NULL;
+  if (size != NULL) { *size = mi_arena_block_size(arena->block_count); }
+  return arena->start;
+}
+
+
+/* -----------------------------------------------------------
+  Arena purge
+----------------------------------------------------------- */
+
+static long mi_arena_purge_delay(void) {
+  // <0 = no purging allowed, 0=immediate purging, >0=milli-second delay
+  return (mi_option_get(mi_option_purge_delay) * mi_option_get(mi_option_arena_purge_mult));
+}
+
+// reset or decommit in an arena and update the committed/decommit bitmaps
+// assumes we own the area (i.e. blocks_in_use is claimed by us)
+static void mi_arena_purge(mi_arena_t* arena, size_t bitmap_idx, size_t blocks, mi_stats_t* stats) {
+  mi_assert_internal(arena->blocks_committed != NULL);
+  mi_assert_internal(arena->blocks_purge != NULL);
+  mi_assert_internal(!arena->memid.is_pinned);
+  const size_t size = mi_arena_block_size(blocks);
+  void* const p = mi_arena_block_start(arena, bitmap_idx);
+  bool needs_recommit;
+  if (_mi_bitmap_is_claimed_across(arena->blocks_committed, arena->field_count, blocks, bitmap_idx)) {
+    // all blocks are committed, we can purge freely
+    needs_recommit = _mi_os_purge(p, size, stats);
+  }
+  else {
+    // some blocks are not committed -- this can happen when a partially committed block is freed
+    // in `_mi_arena_free` and it is conservatively marked as uncommitted but still scheduled for a purge
+    // we need to ensure we do not try to reset (as that may be invalid for uncommitted memory),
+    // and also undo the decommit stats (as it was already adjusted)
+    mi_assert_internal(mi_option_is_enabled(mi_option_purge_decommits));
+    needs_recommit = _mi_os_purge_ex(p, size, false /* allow reset? */, stats);
+    _mi_stat_increase(&stats->committed, size);
+  }
+
+  // clear the purged blocks
+  _mi_bitmap_unclaim_across(arena->blocks_purge, arena->field_count, blocks, bitmap_idx);
+  // update committed bitmap
+  if (needs_recommit) {
+    _mi_bitmap_unclaim_across(arena->blocks_committed, arena->field_count, blocks, bitmap_idx);
+  }
+}
+
+// Schedule a purge. This is usually delayed to avoid repeated decommit/commit calls.
+// Note: assumes we (still) own the area as we may purge immediately
+static void mi_arena_schedule_purge(mi_arena_t* arena, size_t bitmap_idx, size_t blocks, mi_stats_t* stats) {
+  mi_assert_internal(arena->blocks_purge != NULL);
+  const long delay = mi_arena_purge_delay();
+  if (delay < 0) return;  // is purging allowed at all?
+
+  if (_mi_preloading() || delay == 0) {
+    // decommit directly
+    mi_arena_purge(arena, bitmap_idx, blocks, stats);
+  }
+  else {
+    // schedule decommit
+    mi_msecs_t expire = mi_atomic_loadi64_relaxed(&arena->purge_expire);
+    if (expire != 0) {
+      mi_atomic_addi64_acq_rel(&arena->purge_expire, delay/10);  // add smallish extra delay
+    }
+    else {
+      mi_atomic_storei64_release(&arena->purge_expire, _mi_clock_now() + delay);
+    }
+    _mi_bitmap_claim_across(arena->blocks_purge, arena->field_count, blocks, bitmap_idx, NULL);
+  }
+}
+
+// purge a range of blocks
+// return true if the full range was purged.
+// assumes we own the area (i.e. blocks_in_use is claimed by us)
+static bool mi_arena_purge_range(mi_arena_t* arena, size_t idx, size_t startidx, size_t bitlen, size_t purge, mi_stats_t* stats) {
+  const size_t endidx = startidx + bitlen;
+  size_t bitidx = startidx;
+  bool all_purged = false;
+  while (bitidx < endidx) {
+    // count consequetive ones in the purge mask
+    size_t count = 0;
+    while (bitidx + count < endidx && (purge & ((size_t)1 << (bitidx + count))) != 0) {
+      count++;
+    }
+    if (count > 0) {
+      // found range to be purged
+      const mi_bitmap_index_t range_idx = mi_bitmap_index_create(idx, bitidx);
+      mi_arena_purge(arena, range_idx, count, stats);
+      if (count == bitlen) {
+	all_purged = true;
+      }
+    }
+    bitidx += (count+1); // +1 to skip the zero bit (or end)
+  }
+  return all_purged;
+}
+
+// returns true if anything was purged
+static bool mi_arena_try_purge(mi_arena_t* arena, mi_msecs_t now, bool force, mi_stats_t* stats)
+{
+  if (arena->memid.is_pinned || arena->blocks_purge == NULL) return false;
+  mi_msecs_t expire = mi_atomic_loadi64_relaxed(&arena->purge_expire);
+  if (expire == 0) return false;
+  if (!force && expire > now) return false;
+
+  // reset expire (if not already set concurrently)
+  mi_atomic_casi64_strong_acq_rel(&arena->purge_expire, &expire, 0);
+
+  // potential purges scheduled, walk through the bitmap
+  bool any_purged = false;
+  bool full_purge = true;
+  for (size_t i = 0; i < arena->field_count; i++) {
+    size_t purge = mi_atomic_load_relaxed(&arena->blocks_purge[i]);
+    if (purge != 0) {
+      size_t bitidx = 0;
+      while (bitidx < MI_BITMAP_FIELD_BITS) {
+	// find consequetive range of ones in the purge mask
+	size_t bitlen = 0;
+	while (bitidx + bitlen < MI_BITMAP_FIELD_BITS && (purge & ((size_t)1 << (bitidx + bitlen))) != 0) {
+	  bitlen++;
+	}
+	// try to claim the longest range of corresponding in_use bits
+	const mi_bitmap_index_t bitmap_index = mi_bitmap_index_create(i, bitidx);
+	while( bitlen > 0 ) {
+	  if (_mi_bitmap_try_claim(arena->blocks_inuse, arena->field_count, bitlen, bitmap_index)) {
+	    break;
+	  }
+	  bitlen--;
+	}
+	// actual claimed bits at `in_use`
+	if (bitlen > 0) {
+	  // read purge again now that we have the in_use bits
+	  purge = mi_atomic_load_acquire(&arena->blocks_purge[i]);
+	  if (!mi_arena_purge_range(arena, i, bitidx, bitlen, purge, stats)) {
+	    full_purge = false;
+	  }
+	  any_purged = true;
+	  // release the claimed `in_use` bits again
+	  _mi_bitmap_unclaim(arena->blocks_inuse, arena->field_count, bitlen, bitmap_index);
+	}
+	bitidx += (bitlen+1);  // +1 to skip the zero (or end)
+      } // while bitidx
+    } // purge != 0
+  }
+  // if not fully purged, make sure to purge again in the future
+  if (!full_purge) {
+    const long delay = mi_arena_purge_delay();
+    mi_msecs_t expected = 0;
+    mi_atomic_casi64_strong_acq_rel(&arena->purge_expire,&expected,_mi_clock_now() + delay);
+  }
+  return any_purged;
+}
+
+static void mi_arenas_try_purge( bool force, bool visit_all, mi_stats_t* stats ) {
+  if (_mi_preloading() || mi_arena_purge_delay() <= 0) return;  // nothing will be scheduled
+
+  const size_t max_arena = mi_atomic_load_acquire(&mi_arena_count);
+  if (max_arena == 0) return;
+
+  // allow only one thread to purge at a time
+  static mi_atomic_guard_t purge_guard;
+  mi_atomic_guard(&purge_guard)
+  {
+    mi_msecs_t now = _mi_clock_now();
+    size_t max_purge_count = (visit_all ? max_arena : 1);
+    for (size_t i = 0; i < max_arena; i++) {
+      mi_arena_t* arena = mi_atomic_load_ptr_acquire(mi_arena_t, &mi_arenas[i]);
+      if (arena != NULL) {
+	if (mi_arena_try_purge(arena, now, force, stats)) {
+	  if (max_purge_count <= 1) break;
+	  max_purge_count--;
+	}
+      }
+    }
+  }
+}
+
+
+/* -----------------------------------------------------------
+  Arena free
+----------------------------------------------------------- */
+
+void _mi_arena_free(void* p, size_t size, size_t committed_size, mi_memid_t memid, mi_stats_t* stats) {
+  mi_assert_internal(size > 0 && stats != NULL);
+  mi_assert_internal(committed_size <= size);
+  if (p==NULL) return;
+  if (size==0) return;
+  const bool all_committed = (committed_size == size);
+
+  if (mi_memkind_is_os(memid.memkind)) {
+    // was a direct OS allocation, pass through
+    if (!all_committed && committed_size > 0) {
+      // if partially committed, adjust the committed stats (as `_mi_os_free` will increase decommit by the full size)
+      _mi_stat_decrease(&stats->committed, committed_size);
+    }
+    _mi_os_free(p, size, memid, stats);
+  }
+  else if (memid.memkind == MI_MEM_ARENA) {
+    // allocated in an arena
+    size_t arena_idx;
+    size_t bitmap_idx;
+    mi_arena_memid_indices(memid, &arena_idx, &bitmap_idx);
+    mi_assert_internal(arena_idx < MI_MAX_ARENAS);
+    mi_arena_t* arena = mi_atomic_load_ptr_acquire(mi_arena_t,&mi_arenas[arena_idx]);
+    mi_assert_internal(arena != NULL);
+    const size_t blocks = mi_block_count_of_size(size);
+
+    // checks
+    if (arena == NULL) {
+      _mi_error_message(EINVAL, "trying to free from non-existent arena: %p, size %zu, memid: 0x%zx\n", p, size, memid);
+      return;
+    }
+    mi_assert_internal(arena->field_count > mi_bitmap_index_field(bitmap_idx));
+    if (arena->field_count <= mi_bitmap_index_field(bitmap_idx)) {
+      _mi_error_message(EINVAL, "trying to free from non-existent arena block: %p, size %zu, memid: 0x%zx\n", p, size, memid);
+      return;
+    }
+
+    // need to set all memory to undefined as some parts may still be marked as no_access (like padding etc.)
+    mi_track_mem_undefined(p,size);
+
+    // potentially decommit
+    if (arena->memid.is_pinned || arena->blocks_committed == NULL) {
+      mi_assert_internal(all_committed);
+    }
+    else {
+      mi_assert_internal(arena->blocks_committed != NULL);
+      mi_assert_internal(arena->blocks_purge != NULL);
+
+      if (!all_committed) {
+	// mark the entire range as no longer committed (so we recommit the full range when re-using)
+	_mi_bitmap_unclaim_across(arena->blocks_committed, arena->field_count, blocks, bitmap_idx);
+	mi_track_mem_noaccess(p,size);
+	if (committed_size > 0) {
+	  // if partially committed, adjust the committed stats (is it will be recommitted when re-using)
+	  // in the delayed purge, we now need to not count a decommit if the range is not marked as committed.
+	  _mi_stat_decrease(&stats->committed, committed_size);
+	}
+	// note: if not all committed, it may be that the purge will reset/decommit the entire range
+	// that contains already decommitted parts. Since purge consistently uses reset or decommit that
+	// works (as we should never reset decommitted parts).
+      }
+      // (delay) purge the entire range
+      mi_arena_schedule_purge(arena, bitmap_idx, blocks, stats);
+    }
+
+    // and make it available to others again
+    bool all_inuse = _mi_bitmap_unclaim_across(arena->blocks_inuse, arena->field_count, blocks, bitmap_idx);
+    if (!all_inuse) {
+      _mi_error_message(EAGAIN, "trying to free an already freed arena block: %p, size %zu\n", p, size);
+      return;
+    };
+  }
+  else {
+    // arena was none, external, or static; nothing to do
+    mi_assert_internal(memid.memkind < MI_MEM_OS);
+  }
+
+  // purge expired decommits
+  mi_arenas_try_purge(false, false, stats);
+}
+
+// destroy owned arenas; this is unsafe and should only be done using `mi_option_destroy_on_exit`
+// for dynamic libraries that are unloaded and need to release all their allocated memory.
+static void mi_arenas_unsafe_destroy(void) {
+  const size_t max_arena = mi_atomic_load_relaxed(&mi_arena_count);
+  size_t new_max_arena = 0;
+  for (size_t i = 0; i < max_arena; i++) {
+    mi_arena_t* arena = mi_atomic_load_ptr_acquire(mi_arena_t, &mi_arenas[i]);
+    if (arena != NULL) {
+      if (arena->start != NULL && mi_memkind_is_os(arena->memid.memkind)) {
+	mi_atomic_store_ptr_release(mi_arena_t, &mi_arenas[i], NULL);
+	_mi_os_free(arena->start, mi_arena_size(arena), arena->memid, &_mi_stats_main);
+      }
+      else {
+	new_max_arena = i;
+      }
+      mi_arena_meta_free(arena, arena->meta_memid, arena->meta_size, &_mi_stats_main);
+    }
+  }
+
+  // try to lower the max arena.
+  size_t expected = max_arena;
+  mi_atomic_cas_strong_acq_rel(&mi_arena_count, &expected, new_max_arena);
+}
+
+// Purge the arenas; if `force_purge` is true, amenable parts are purged even if not yet expired
+void _mi_arena_collect(bool force_purge, mi_stats_t* stats) {
+  mi_arenas_try_purge(force_purge, true /* visit all */, stats);
+}
+
+// destroy owned arenas; this is unsafe and should only be done using `mi_option_destroy_on_exit`
+// for dynamic libraries that are unloaded and need to release all their allocated memory.
+void _mi_arena_unsafe_destroy_all(mi_stats_t* stats) {
+  mi_arenas_unsafe_destroy();
+  _mi_arena_collect(true /* force purge */, stats);  // purge non-owned arenas
+}
+
+// Is a pointer inside any of our arenas?
+bool _mi_arena_contains(const void* p) {
+  const size_t max_arena = mi_atomic_load_relaxed(&mi_arena_count);
+  for (size_t i = 0; i < max_arena; i++) {
+    mi_arena_t* arena = mi_atomic_load_ptr_acquire(mi_arena_t, &mi_arenas[i]);
+    if (arena != NULL && arena->start <= (const uint8_t*)p && arena->start + mi_arena_block_size(arena->block_count) > (const uint8_t*)p) {
+      return true;
+    }
+  }
+  return false;
+}
+
+
+/* -----------------------------------------------------------
+  Add an arena.
+----------------------------------------------------------- */
+
+static bool mi_arena_add(mi_arena_t* arena, mi_arena_id_t* arena_id) {
+  mi_assert_internal(arena != NULL);
+  mi_assert_internal((uintptr_t)mi_atomic_load_ptr_relaxed(uint8_t,&arena->start) % MI_SEGMENT_ALIGN == 0);
+  mi_assert_internal(arena->block_count > 0);
+  if (arena_id != NULL) { *arena_id = -1; }
+
+  size_t i = mi_atomic_increment_acq_rel(&mi_arena_count);
+  if (i >= MI_MAX_ARENAS) {
+    mi_atomic_decrement_acq_rel(&mi_arena_count);
+    return false;
+  }
+  arena->id = mi_arena_id_create(i);
+  mi_atomic_store_ptr_release(mi_arena_t,&mi_arenas[i], arena);
+  if (arena_id != NULL) { *arena_id = arena->id; }
+  return true;
+}
+
+static bool mi_manage_os_memory_ex2(void* start, size_t size, bool is_large, int numa_node, bool exclusive, mi_memid_t memid, mi_arena_id_t* arena_id) mi_attr_noexcept
+{
+  if (arena_id != NULL) *arena_id = _mi_arena_id_none();
+  if (size < MI_ARENA_BLOCK_SIZE) return false;
+
+  if (is_large) {
+    mi_assert_internal(memid.initially_committed && memid.is_pinned);
+  }
+
+  const size_t bcount = size / MI_ARENA_BLOCK_SIZE;
+  const size_t fields = _mi_divide_up(bcount, MI_BITMAP_FIELD_BITS);
+  const size_t bitmaps = (memid.is_pinned ? 2 : 4);
+  const size_t asize  = sizeof(mi_arena_t) + (bitmaps*fields*sizeof(mi_bitmap_field_t));
+  mi_memid_t meta_memid;
+  mi_arena_t* arena   = (mi_arena_t*)mi_arena_meta_zalloc(asize, &meta_memid, &_mi_stats_main); // TODO: can we avoid allocating from the OS?
+  if (arena == NULL) return false;
+
+  // already zero'd due to os_alloc
+  // _mi_memzero(arena, asize);
+  arena->id = _mi_arena_id_none();
+  arena->memid = memid;
+  arena->exclusive = exclusive;
+  arena->meta_size = asize;
+  arena->meta_memid = meta_memid;
+  arena->block_count = bcount;
+  arena->field_count = fields;
+  arena->start = (uint8_t*)start;
+  arena->numa_node    = numa_node; // TODO: or get the current numa node if -1? (now it allows anyone to allocate on -1)
+  arena->is_large     = is_large;
+  arena->purge_expire = 0;
+  arena->search_idx   = 0;
+  arena->blocks_dirty = &arena->blocks_inuse[fields]; // just after inuse bitmap
+  arena->blocks_committed = (arena->memid.is_pinned ? NULL : &arena->blocks_inuse[2*fields]); // just after dirty bitmap
+  arena->blocks_purge  = (arena->memid.is_pinned ? NULL : &arena->blocks_inuse[3*fields]); // just after committed bitmap
+  // initialize committed bitmap?
+  if (arena->blocks_committed != NULL && arena->memid.initially_committed) {
+    memset((void*)arena->blocks_committed, 0xFF, fields*sizeof(mi_bitmap_field_t)); // cast to void* to avoid atomic warning
+  }
+
+  // and claim leftover blocks if needed (so we never allocate there)
+  ptrdiff_t post = (fields * MI_BITMAP_FIELD_BITS) - bcount;
+  mi_assert_internal(post >= 0);
+  if (post > 0) {
+    // don't use leftover bits at the end
+    mi_bitmap_index_t postidx = mi_bitmap_index_create(fields - 1, MI_BITMAP_FIELD_BITS - post);
+    _mi_bitmap_claim(arena->blocks_inuse, fields, post, postidx, NULL);
+  }
+  return mi_arena_add(arena, arena_id);
+
+}
+
+bool mi_manage_os_memory_ex(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept {
+  mi_memid_t memid = _mi_memid_create(MI_MEM_EXTERNAL);
+  memid.initially_committed = is_committed;
+  memid.initially_zero = is_zero;
+  memid.is_pinned = is_large;
+  return mi_manage_os_memory_ex2(start,size,is_large,numa_node,exclusive,memid, arena_id);
+}
+
+// Reserve a range of regular OS memory
+int mi_reserve_os_memory_ex(size_t size, bool commit, bool allow_large, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept {
+  if (arena_id != NULL) *arena_id = _mi_arena_id_none();
+  size = _mi_align_up(size, MI_ARENA_BLOCK_SIZE); // at least one block
+  mi_memid_t memid;
+  void* start = _mi_os_alloc_aligned(size, MI_SEGMENT_ALIGN, commit, allow_large, &memid, &_mi_stats_main);
+  if (start == NULL) return ENOMEM;
+  const bool is_large = memid.is_pinned; // todo: use separate is_large field?
+  if (!mi_manage_os_memory_ex2(start, size, is_large, -1 /* numa node */, exclusive, memid, arena_id)) {
+    _mi_os_free_ex(start, size, commit, memid, &_mi_stats_main);
+    _mi_verbose_message("failed to reserve %zu k memory\n", _mi_divide_up(size, 1024));
+    return ENOMEM;
+  }
+  _mi_verbose_message("reserved %zu KiB memory%s\n", _mi_divide_up(size, 1024), is_large ? " (in large os pages)" : "");
+  return 0;
+}
+
+
+// Manage a range of regular OS memory
+bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node) mi_attr_noexcept {
+  return mi_manage_os_memory_ex(start, size, is_committed, is_large, is_zero, numa_node, false /* exclusive? */, NULL);
+}
+
+// Reserve a range of regular OS memory
+int mi_reserve_os_memory(size_t size, bool commit, bool allow_large) mi_attr_noexcept {
+  return mi_reserve_os_memory_ex(size, commit, allow_large, false, NULL);
+}
+
+
+/* -----------------------------------------------------------
+  Debugging
+----------------------------------------------------------- */
+
+static size_t mi_debug_show_bitmap(const char* prefix, mi_bitmap_field_t* fields, size_t field_count ) {
+  size_t inuse_count = 0;
+  for (size_t i = 0; i < field_count; i++) {
+    char buf[MI_BITMAP_FIELD_BITS + 1];
+    uintptr_t field = mi_atomic_load_relaxed(&fields[i]);
+    for (size_t bit = 0; bit < MI_BITMAP_FIELD_BITS; bit++) {
+      bool inuse = ((((uintptr_t)1 << bit) & field) != 0);
+      if (inuse) inuse_count++;
+      buf[MI_BITMAP_FIELD_BITS - 1 - bit] = (inuse ? 'x' : '.');
+    }
+    buf[MI_BITMAP_FIELD_BITS] = 0;
+    _mi_verbose_message("%s%s\n", prefix, buf);
+  }
+  return inuse_count;
+}
+
+void mi_debug_show_arenas(void) mi_attr_noexcept {
+  size_t max_arenas = mi_atomic_load_relaxed(&mi_arena_count);
+  for (size_t i = 0; i < max_arenas; i++) {
+    mi_arena_t* arena = mi_atomic_load_ptr_relaxed(mi_arena_t, &mi_arenas[i]);
+    if (arena == NULL) break;
+    size_t inuse_count = 0;
+    _mi_verbose_message("arena %zu: %zu blocks with %zu fields\n", i, arena->block_count, arena->field_count);
+    inuse_count += mi_debug_show_bitmap("  ", arena->blocks_inuse, arena->field_count);
+    _mi_verbose_message("  blocks in use ('x'): %zu\n", inuse_count);
+  }
+}
+
+
+/* -----------------------------------------------------------
+  Reserve a huge page arena.
+----------------------------------------------------------- */
+// reserve at a specific numa node
+int mi_reserve_huge_os_pages_at_ex(size_t pages, int numa_node, size_t timeout_msecs, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept {
+  if (arena_id != NULL) *arena_id = -1;
+  if (pages==0) return 0;
+  if (numa_node < -1) numa_node = -1;
+  if (numa_node >= 0) numa_node = numa_node % _mi_os_numa_node_count();
+  size_t hsize = 0;
+  size_t pages_reserved = 0;
+  mi_memid_t memid;
+  void* p = _mi_os_alloc_huge_os_pages(pages, numa_node, timeout_msecs, &pages_reserved, &hsize, &memid);
+  if (p==NULL || pages_reserved==0) {
+    _mi_warning_message("failed to reserve %zu GiB huge pages\n", pages);
+    return ENOMEM;
+  }
+  _mi_verbose_message("numa node %i: reserved %zu GiB huge pages (of the %zu GiB requested)\n", numa_node, pages_reserved, pages);
+
+  if (!mi_manage_os_memory_ex2(p, hsize, true, numa_node, exclusive, memid, arena_id)) {
+    _mi_os_free(p, hsize, memid, &_mi_stats_main);
+    return ENOMEM;
+  }
+  return 0;
+}
+
+int mi_reserve_huge_os_pages_at(size_t pages, int numa_node, size_t timeout_msecs) mi_attr_noexcept {
+  return mi_reserve_huge_os_pages_at_ex(pages, numa_node, timeout_msecs, false, NULL);
+}
+
+// reserve huge pages evenly among the given number of numa nodes (or use the available ones as detected)
+int mi_reserve_huge_os_pages_interleave(size_t pages, size_t numa_nodes, size_t timeout_msecs) mi_attr_noexcept {
+  if (pages == 0) return 0;
+
+  // pages per numa node
+  size_t numa_count = (numa_nodes > 0 ? numa_nodes : _mi_os_numa_node_count());
+  if (numa_count <= 0) numa_count = 1;
+  const size_t pages_per = pages / numa_count;
+  const size_t pages_mod = pages % numa_count;
+  const size_t timeout_per = (timeout_msecs==0 ? 0 : (timeout_msecs / numa_count) + 50);
+
+  // reserve evenly among numa nodes
+  for (size_t numa_node = 0; numa_node < numa_count && pages > 0; numa_node++) {
+    size_t node_pages = pages_per;  // can be 0
+    if (numa_node < pages_mod) node_pages++;
+    int err = mi_reserve_huge_os_pages_at(node_pages, (int)numa_node, timeout_per);
+    if (err) return err;
+    if (pages < node_pages) {
+      pages = 0;
+    }
+    else {
+      pages -= node_pages;
+    }
+  }
+
+  return 0;
+}
+
+int mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept {
+  MI_UNUSED(max_secs);
+  _mi_warning_message("mi_reserve_huge_os_pages is deprecated: use mi_reserve_huge_os_pages_interleave/at instead\n");
+  if (pages_reserved != NULL) *pages_reserved = 0;
+  int err = mi_reserve_huge_os_pages_interleave(pages, 0, (size_t)(max_secs * 1000.0));
+  if (err==0 && pages_reserved!=NULL) *pages_reserved = pages;
+  return err;
+}
diff --git a/compat/mimalloc/bitmap.c b/compat/mimalloc/bitmap.c
new file mode 100644
index 00000000000000..878f0ab3250a47
--- /dev/null
+++ b/compat/mimalloc/bitmap.c
@@ -0,0 +1,432 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2023 Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* ----------------------------------------------------------------------------
+Concurrent bitmap that can set/reset sequences of bits atomically,
+represeted as an array of fields where each field is a machine word (`size_t`)
+
+There are two api's; the standard one cannot have sequences that cross
+between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
+
+The `_across` postfixed functions do allow sequences that can cross over
+between the fields. (This is used in arena allocation)
+---------------------------------------------------------------------------- */
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "bitmap.h"
+
+/* -----------------------------------------------------------
+  Bitmap definition
+----------------------------------------------------------- */
+
+// The bit mask for a given number of blocks at a specified bit index.
+static inline size_t mi_bitmap_mask_(size_t count, size_t bitidx) {
+  mi_assert_internal(count + bitidx <= MI_BITMAP_FIELD_BITS);
+  mi_assert_internal(count > 0);
+  if (count >= MI_BITMAP_FIELD_BITS) return MI_BITMAP_FIELD_FULL;
+  if (count == 0) return 0;
+  return ((((size_t)1 << count) - 1) << bitidx);
+}
+
+
+/* -----------------------------------------------------------
+  Claim a bit sequence atomically
+----------------------------------------------------------- */
+
+// Try to atomically claim a sequence of `count` bits in a single
+// field at `idx` in `bitmap`. Returns `true` on success.
+inline bool _mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx, const size_t count, mi_bitmap_index_t* bitmap_idx)
+{
+  mi_assert_internal(bitmap_idx != NULL);
+  mi_assert_internal(count <= MI_BITMAP_FIELD_BITS);
+  mi_assert_internal(count > 0);
+  mi_bitmap_field_t* field = &bitmap[idx];
+  size_t map  = mi_atomic_load_relaxed(field);
+  if (map==MI_BITMAP_FIELD_FULL) return false; // short cut
+
+  // search for 0-bit sequence of length count
+  const size_t mask = mi_bitmap_mask_(count, 0);
+  const size_t bitidx_max = MI_BITMAP_FIELD_BITS - count;
+
+#ifdef MI_HAVE_FAST_BITSCAN
+  size_t bitidx = mi_ctz(~map);    // quickly find the first zero bit if possible
+#else
+  size_t bitidx = 0;               // otherwise start at 0
+#endif
+  size_t m = (mask << bitidx);     // invariant: m == mask shifted by bitidx
+
+  // scan linearly for a free range of zero bits
+  while (bitidx <= bitidx_max) {
+    const size_t mapm = (map & m);
+    if (mapm == 0) {  // are the mask bits free at bitidx?
+      mi_assert_internal((m >> bitidx) == mask); // no overflow?
+      const size_t newmap = (map | m);
+      mi_assert_internal((newmap^map) >> bitidx == mask);
+      if (!mi_atomic_cas_strong_acq_rel(field, &map, newmap)) {  // TODO: use weak cas here?
+	// no success, another thread claimed concurrently.. keep going (with updated `map`)
+	continue;
+      }
+      else {
+	// success, we claimed the bits!
+	*bitmap_idx = mi_bitmap_index_create(idx, bitidx);
+	return true;
+      }
+    }
+    else {
+      // on to the next bit range
+#ifdef MI_HAVE_FAST_BITSCAN
+      mi_assert_internal(mapm != 0);
+      const size_t shift = (count == 1 ? 1 : (MI_INTPTR_BITS - mi_clz(mapm) - bitidx));
+      mi_assert_internal(shift > 0 && shift <= count);
+#else
+      const size_t shift = 1;
+#endif
+      bitidx += shift;
+      m <<= shift;
+    }
+  }
+  // no bits found
+  return false;
+}
+
+// Find `count` bits of 0 and set them to 1 atomically; returns `true` on success.
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+// `count` can be at most MI_BITMAP_FIELD_BITS and will never cross fields.
+bool _mi_bitmap_try_find_from_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx) {
+  size_t idx = start_field_idx;
+  for (size_t visited = 0; visited < bitmap_fields; visited++, idx++) {
+    if (idx >= bitmap_fields) { idx = 0; } // wrap
+    if (_mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
+      return true;
+    }
+  }
+  return false;
+}
+
+// Like _mi_bitmap_try_find_from_claim but with an extra predicate that must be fullfilled
+bool _mi_bitmap_try_find_from_claim_pred(mi_bitmap_t bitmap, const size_t bitmap_fields,
+	    const size_t start_field_idx, const size_t count,
+	    mi_bitmap_pred_fun_t pred_fun, void* pred_arg,
+	    mi_bitmap_index_t* bitmap_idx) {
+  size_t idx = start_field_idx;
+  for (size_t visited = 0; visited < bitmap_fields; visited++, idx++) {
+    if (idx >= bitmap_fields) idx = 0; // wrap
+    if (_mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
+      if (pred_fun == NULL || pred_fun(*bitmap_idx, pred_arg)) {
+	return true;
+      }
+      // predicate returned false, unclaim and look further
+      _mi_bitmap_unclaim(bitmap, bitmap_fields, count, *bitmap_idx);
+    }
+  }
+  return false;
+}
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  const size_t idx = mi_bitmap_index_field(bitmap_idx);
+  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
+  const size_t mask = mi_bitmap_mask_(count, bitidx);
+  mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
+  // mi_assert_internal((bitmap[idx] & mask) == mask);
+  const size_t prev = mi_atomic_and_acq_rel(&bitmap[idx], ~mask);
+  return ((prev & mask) == mask);
+}
+
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_zero) {
+  const size_t idx = mi_bitmap_index_field(bitmap_idx);
+  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
+  const size_t mask = mi_bitmap_mask_(count, bitidx);
+  mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
+  //mi_assert_internal(any_zero != NULL || (bitmap[idx] & mask) == 0);
+  size_t prev = mi_atomic_or_acq_rel(&bitmap[idx], mask);
+  if (any_zero != NULL) { *any_zero = ((prev & mask) != mask); }
+  return ((prev & mask) == 0);
+}
+
+// Returns `true` if all `count` bits were 1. `any_ones` is `true` if there was at least one bit set to one.
+static bool mi_bitmap_is_claimedx(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_ones) {
+  const size_t idx = mi_bitmap_index_field(bitmap_idx);
+  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
+  const size_t mask = mi_bitmap_mask_(count, bitidx);
+  mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
+  const size_t field = mi_atomic_load_relaxed(&bitmap[idx]);
+  if (any_ones != NULL) { *any_ones = ((field & mask) != 0); }
+  return ((field & mask) == mask);
+}
+
+// Try to set `count` bits at `bitmap_idx` from 0 to 1 atomically.
+// Returns `true` if successful when all previous `count` bits were 0.
+bool _mi_bitmap_try_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  const size_t idx = mi_bitmap_index_field(bitmap_idx);
+  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
+  const size_t mask = mi_bitmap_mask_(count, bitidx);
+  mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
+  size_t expected = mi_atomic_load_relaxed(&bitmap[idx]);
+  do  {
+    if ((expected & mask) != 0) return false;
+  }
+  while (!mi_atomic_cas_strong_acq_rel(&bitmap[idx], &expected, expected | mask));
+  mi_assert_internal((expected & mask) == 0);
+  return true;
+}
+
+
+bool _mi_bitmap_is_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  return mi_bitmap_is_claimedx(bitmap, bitmap_fields, count, bitmap_idx, NULL);
+}
+
+bool _mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  bool any_ones;
+  mi_bitmap_is_claimedx(bitmap, bitmap_fields, count, bitmap_idx, &any_ones);
+  return any_ones;
+}
+
+
+//--------------------------------------------------------------------------
+// the `_across` functions work on bitmaps where sequences can cross over
+// between the fields. This is used in arena allocation
+//--------------------------------------------------------------------------
+
+// Try to atomically claim a sequence of `count` bits starting from the field
+// at `idx` in `bitmap` and crossing into subsequent fields. Returns `true` on success.
+// Only needs to consider crossing into the next fields (see `mi_bitmap_try_find_from_claim_across`)
+static bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t idx, const size_t count, const size_t retries, mi_bitmap_index_t* bitmap_idx)
+{
+  mi_assert_internal(bitmap_idx != NULL);
+
+  // check initial trailing zeros
+  mi_bitmap_field_t* field = &bitmap[idx];
+  size_t map = mi_atomic_load_relaxed(field);
+  const size_t initial = mi_clz(map);  // count of initial zeros starting at idx
+  mi_assert_internal(initial <= MI_BITMAP_FIELD_BITS);
+  if (initial == 0)     return false;
+  if (initial >= count) return _mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx);    // no need to cross fields (this case won't happen for us)
+  if (_mi_divide_up(count - initial, MI_BITMAP_FIELD_BITS) >= (bitmap_fields - idx)) return false; // not enough entries
+
+  // scan ahead
+  size_t found = initial;
+  size_t mask = 0;     // mask bits for the final field
+  while(found < count) {
+    field++;
+    map = mi_atomic_load_relaxed(field);
+    const size_t mask_bits = (found + MI_BITMAP_FIELD_BITS <= count ? MI_BITMAP_FIELD_BITS : (count - found));
+    mi_assert_internal(mask_bits > 0 && mask_bits <= MI_BITMAP_FIELD_BITS);
+    mask = mi_bitmap_mask_(mask_bits, 0);
+    if ((map & mask) != 0) return false;  // some part is already claimed
+    found += mask_bits;
+  }
+  mi_assert_internal(field < &bitmap[bitmap_fields]);
+
+  // we found a range of contiguous zeros up to the final field; mask contains mask in the final field
+  // now try to claim the range atomically
+  mi_bitmap_field_t* const final_field = field;
+  const size_t final_mask = mask;
+  mi_bitmap_field_t* const initial_field = &bitmap[idx];
+  const size_t initial_idx = MI_BITMAP_FIELD_BITS - initial;
+  const size_t initial_mask = mi_bitmap_mask_(initial, initial_idx);
+
+  // initial field
+  size_t newmap;
+  field = initial_field;
+  map = mi_atomic_load_relaxed(field);
+  do {
+    newmap = (map | initial_mask);
+    if ((map & initial_mask) != 0) { goto rollback; };
+  } while (!mi_atomic_cas_strong_acq_rel(field, &map, newmap));
+
+  // intermediate fields
+  while (++field < final_field) {
+    newmap = MI_BITMAP_FIELD_FULL;
+    map = 0;
+    if (!mi_atomic_cas_strong_acq_rel(field, &map, newmap)) { goto rollback; }
+  }
+
+  // final field
+  mi_assert_internal(field == final_field);
+  map = mi_atomic_load_relaxed(field);
+  do {
+    newmap = (map | final_mask);
+    if ((map & final_mask) != 0) { goto rollback; }
+  } while (!mi_atomic_cas_strong_acq_rel(field, &map, newmap));
+
+  // claimed!
+  *bitmap_idx = mi_bitmap_index_create(idx, initial_idx);
+  return true;
+
+rollback:
+  // roll back intermediate fields
+  // (we just failed to claim `field` so decrement first)
+  while (--field > initial_field) {
+    newmap = 0;
+    map = MI_BITMAP_FIELD_FULL;
+    mi_assert_internal(mi_atomic_load_relaxed(field) == map);
+    mi_atomic_store_release(field, newmap);
+  }
+  if (field == initial_field) {               // (if we failed on the initial field, `field + 1 == initial_field`)
+    map = mi_atomic_load_relaxed(field);
+    do {
+      mi_assert_internal((map & initial_mask) == initial_mask);
+      newmap = (map & ~initial_mask);
+    } while (!mi_atomic_cas_strong_acq_rel(field, &map, newmap));
+  }
+  // retry? (we make a recursive call instead of goto to be able to use const declarations)
+  if (retries <= 2) {
+    return mi_bitmap_try_find_claim_field_across(bitmap, bitmap_fields, idx, count, retries+1, bitmap_idx);
+  }
+  else {
+    return false;
+  }
+}
+
+
+// Find `count` bits of zeros and set them to 1 atomically; returns `true` on success.
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx) {
+  mi_assert_internal(count > 0);
+  if (count <= 2) {
+    // we don't bother with crossover fields for small counts
+    return _mi_bitmap_try_find_from_claim(bitmap, bitmap_fields, start_field_idx, count, bitmap_idx);
+  }
+
+  // visit the fields
+  size_t idx = start_field_idx;
+  for (size_t visited = 0; visited < bitmap_fields; visited++, idx++) {
+    if (idx >= bitmap_fields) { idx = 0; } // wrap
+    // first try to claim inside a field
+    if (count <= MI_BITMAP_FIELD_BITS) {
+      if (_mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
+	return true;
+      }
+    }
+    // if that fails, then try to claim across fields
+    if (mi_bitmap_try_find_claim_field_across(bitmap, bitmap_fields, idx, count, 0, bitmap_idx)) {
+      return true;
+    }
+  }
+  return false;
+}
+
+// Helper for masks across fields; returns the mid count, post_mask may be 0
+static size_t mi_bitmap_mask_across(mi_bitmap_index_t bitmap_idx, size_t bitmap_fields, size_t count, size_t* pre_mask, size_t* mid_mask, size_t* post_mask) {
+  MI_UNUSED(bitmap_fields);
+  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
+  if mi_likely(bitidx + count <= MI_BITMAP_FIELD_BITS) {
+    *pre_mask = mi_bitmap_mask_(count, bitidx);
+    *mid_mask = 0;
+    *post_mask = 0;
+    mi_assert_internal(mi_bitmap_index_field(bitmap_idx) < bitmap_fields);
+    return 0;
+  }
+  else {
+    const size_t pre_bits = MI_BITMAP_FIELD_BITS - bitidx;
+    mi_assert_internal(pre_bits < count);
+    *pre_mask = mi_bitmap_mask_(pre_bits, bitidx);
+    count -= pre_bits;
+    const size_t mid_count = (count / MI_BITMAP_FIELD_BITS);
+    *mid_mask = MI_BITMAP_FIELD_FULL;
+    count %= MI_BITMAP_FIELD_BITS;
+    *post_mask = (count==0 ? 0 : mi_bitmap_mask_(count, 0));
+    mi_assert_internal(mi_bitmap_index_field(bitmap_idx) + mid_count + (count==0 ? 0 : 1) < bitmap_fields);
+    return mid_count;
+  }
+}
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  size_t idx = mi_bitmap_index_field(bitmap_idx);
+  size_t pre_mask;
+  size_t mid_mask;
+  size_t post_mask;
+  size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
+  bool all_one = true;
+  mi_bitmap_field_t* field = &bitmap[idx];
+  size_t prev = mi_atomic_and_acq_rel(field++, ~pre_mask);   // clear first part
+  if ((prev & pre_mask) != pre_mask) all_one = false;
+  while(mid_count-- > 0) {
+    prev = mi_atomic_and_acq_rel(field++, ~mid_mask);        // clear mid part
+    if ((prev & mid_mask) != mid_mask) all_one = false;
+  }
+  if (post_mask!=0) {
+    prev = mi_atomic_and_acq_rel(field, ~post_mask);         // clear end part
+    if ((prev & post_mask) != post_mask) all_one = false;
+  }
+  return all_one;
+}
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero) {
+  size_t idx = mi_bitmap_index_field(bitmap_idx);
+  size_t pre_mask;
+  size_t mid_mask;
+  size_t post_mask;
+  size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
+  bool all_zero = true;
+  bool any_zero = false;
+  _Atomic(size_t)*field = &bitmap[idx];
+  size_t prev = mi_atomic_or_acq_rel(field++, pre_mask);
+  if ((prev & pre_mask) != 0) all_zero = false;
+  if ((prev & pre_mask) != pre_mask) any_zero = true;
+  while (mid_count-- > 0) {
+    prev = mi_atomic_or_acq_rel(field++, mid_mask);
+    if ((prev & mid_mask) != 0) all_zero = false;
+    if ((prev & mid_mask) != mid_mask) any_zero = true;
+  }
+  if (post_mask!=0) {
+    prev = mi_atomic_or_acq_rel(field, post_mask);
+    if ((prev & post_mask) != 0) all_zero = false;
+    if ((prev & post_mask) != post_mask) any_zero = true;
+  }
+  if (pany_zero != NULL) { *pany_zero = any_zero; }
+  return all_zero;
+}
+
+
+// Returns `true` if all `count` bits were 1.
+// `any_ones` is `true` if there was at least one bit set to one.
+static bool mi_bitmap_is_claimedx_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_ones) {
+  size_t idx = mi_bitmap_index_field(bitmap_idx);
+  size_t pre_mask;
+  size_t mid_mask;
+  size_t post_mask;
+  size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
+  bool all_ones = true;
+  bool any_ones = false;
+  mi_bitmap_field_t* field = &bitmap[idx];
+  size_t prev = mi_atomic_load_relaxed(field++);
+  if ((prev & pre_mask) != pre_mask) all_ones = false;
+  if ((prev & pre_mask) != 0) any_ones = true;
+  while (mid_count-- > 0) {
+    prev = mi_atomic_load_relaxed(field++);
+    if ((prev & mid_mask) != mid_mask) all_ones = false;
+    if ((prev & mid_mask) != 0) any_ones = true;
+  }
+  if (post_mask!=0) {
+    prev = mi_atomic_load_relaxed(field);
+    if ((prev & post_mask) != post_mask) all_ones = false;
+    if ((prev & post_mask) != 0) any_ones = true;
+  }
+  if (pany_ones != NULL) { *pany_ones = any_ones; }
+  return all_ones;
+}
+
+bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  return mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, NULL);
+}
+
+bool _mi_bitmap_is_any_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  bool any_ones;
+  mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, &any_ones);
+  return any_ones;
+}
diff --git a/compat/mimalloc/bitmap.h b/compat/mimalloc/bitmap.h
new file mode 100644
index 00000000000000..9ba15d5d6f09ea
--- /dev/null
+++ b/compat/mimalloc/bitmap.h
@@ -0,0 +1,115 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2023 Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* ----------------------------------------------------------------------------
+Concurrent bitmap that can set/reset sequences of bits atomically,
+represeted as an array of fields where each field is a machine word (`size_t`)
+
+There are two api's; the standard one cannot have sequences that cross
+between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
+(this is used in region allocation)
+
+The `_across` postfixed functions do allow sequences that can cross over
+between the fields. (This is used in arena allocation)
+---------------------------------------------------------------------------- */
+#pragma once
+#ifndef MI_BITMAP_H
+#define MI_BITMAP_H
+
+/* -----------------------------------------------------------
+  Bitmap definition
+----------------------------------------------------------- */
+
+#define MI_BITMAP_FIELD_BITS   (8*MI_SIZE_SIZE)
+#define MI_BITMAP_FIELD_FULL   (~((size_t)0))   // all bits set
+
+// An atomic bitmap of `size_t` fields
+typedef _Atomic(size_t)  mi_bitmap_field_t;
+typedef mi_bitmap_field_t*  mi_bitmap_t;
+
+// A bitmap index is the index of the bit in a bitmap.
+typedef size_t mi_bitmap_index_t;
+
+// Create a bit index.
+static inline mi_bitmap_index_t mi_bitmap_index_create(size_t idx, size_t bitidx) {
+  mi_assert_internal(bitidx < MI_BITMAP_FIELD_BITS);
+  return (idx*MI_BITMAP_FIELD_BITS) + bitidx;
+}
+
+// Create a bit index.
+static inline mi_bitmap_index_t mi_bitmap_index_create_from_bit(size_t full_bitidx) {
+  return mi_bitmap_index_create(full_bitidx / MI_BITMAP_FIELD_BITS, full_bitidx % MI_BITMAP_FIELD_BITS);
+}
+
+// Get the field index from a bit index.
+static inline size_t mi_bitmap_index_field(mi_bitmap_index_t bitmap_idx) {
+  return (bitmap_idx / MI_BITMAP_FIELD_BITS);
+}
+
+// Get the bit index in a bitmap field
+static inline size_t mi_bitmap_index_bit_in_field(mi_bitmap_index_t bitmap_idx) {
+  return (bitmap_idx % MI_BITMAP_FIELD_BITS);
+}
+
+// Get the full bit index
+static inline size_t mi_bitmap_index_bit(mi_bitmap_index_t bitmap_idx) {
+  return bitmap_idx;
+}
+
+/* -----------------------------------------------------------
+  Claim a bit sequence atomically
+----------------------------------------------------------- */
+
+// Try to atomically claim a sequence of `count` bits in a single
+// field at `idx` in `bitmap`. Returns `true` on success.
+bool _mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
+
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+// For now, `count` can be at most MI_BITMAP_FIELD_BITS and will never cross fields.
+bool _mi_bitmap_try_find_from_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
+
+// Like _mi_bitmap_try_find_from_claim but with an extra predicate that must be fullfilled
+typedef bool (mi_cdecl *mi_bitmap_pred_fun_t)(mi_bitmap_index_t bitmap_idx, void* pred_arg);
+bool _mi_bitmap_try_find_from_claim_pred(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_pred_fun_t pred_fun, void* pred_arg, mi_bitmap_index_t* bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+// Try to set `count` bits at `bitmap_idx` from 0 to 1 atomically.
+// Returns `true` if successful when all previous `count` bits were 0.
+bool _mi_bitmap_try_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_zero);
+
+bool _mi_bitmap_is_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+bool _mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+
+//--------------------------------------------------------------------------
+// the `_across` functions work on bitmaps where sequences can cross over
+// between the fields. This is used in arena allocation
+//--------------------------------------------------------------------------
+
+// Find `count` bits of zeros and set them to 1 atomically; returns `true` on success.
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero);
+
+bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+bool _mi_bitmap_is_any_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+#endif
diff --git a/compat/mimalloc/heap.c b/compat/mimalloc/heap.c
new file mode 100644
index 00000000000000..dab8c4bf8ae388
--- /dev/null
+++ b/compat/mimalloc/heap.c
@@ -0,0 +1,626 @@
+/*----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"  // mi_prim_get_default_heap
+
+#include <string.h>  // memset, memcpy
+
+#if defined(_MSC_VER) && (_MSC_VER < 1920)
+#pragma warning(disable:4204)  // non-constant aggregate initializer
+#endif
+
+/* -----------------------------------------------------------
+  Helpers
+----------------------------------------------------------- */
+
+// return `true` if ok, `false` to break
+typedef bool (heap_page_visitor_fun)(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2);
+
+// Visit all pages in a heap; returns `false` if break was called.
+static bool mi_heap_visit_pages(mi_heap_t* heap, heap_page_visitor_fun* fn, void* arg1, void* arg2)
+{
+  if (heap==NULL || heap->page_count==0) return 0;
+
+  // visit all pages
+  #if MI_DEBUG>1
+  size_t total = heap->page_count;
+  size_t count = 0;
+  #endif
+
+  for (size_t i = 0; i <= MI_BIN_FULL; i++) {
+    mi_page_queue_t* pq = &heap->pages[i];
+    mi_page_t* page = pq->first;
+    while(page != NULL) {
+      mi_page_t* next = page->next; // save next in case the page gets removed from the queue
+      mi_assert_internal(mi_page_heap(page) == heap);
+      #if MI_DEBUG>1
+      count++;
+      #endif
+      if (!fn(heap, pq, page, arg1, arg2)) return false;
+      page = next; // and continue
+    }
+  }
+  mi_assert_internal(count == total);
+  return true;
+}
+
+
+#if MI_DEBUG>=2
+static bool mi_heap_page_is_valid(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2) {
+  MI_UNUSED(arg1);
+  MI_UNUSED(arg2);
+  MI_UNUSED(pq);
+  mi_assert_internal(mi_page_heap(page) == heap);
+  mi_segment_t* segment = _mi_page_segment(page);
+  mi_assert_internal(segment->thread_id == heap->thread_id);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  return true;
+}
+#endif
+#if MI_DEBUG>=3
+static bool mi_heap_is_valid(mi_heap_t* heap) {
+  mi_assert_internal(heap!=NULL);
+  mi_heap_visit_pages(heap, &mi_heap_page_is_valid, NULL, NULL);
+  return true;
+}
+#endif
+
+
+
+
+/* -----------------------------------------------------------
+  "Collect" pages by migrating `local_free` and `thread_free`
+  lists and freeing empty pages. This is done when a thread
+  stops (and in that case abandons pages if there are still
+  blocks alive)
+----------------------------------------------------------- */
+
+typedef enum mi_collect_e {
+  MI_NORMAL,
+  MI_FORCE,
+  MI_ABANDON
+} mi_collect_t;
+
+
+static bool mi_heap_page_collect(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg_collect, void* arg2 ) {
+  MI_UNUSED(arg2);
+  MI_UNUSED(heap);
+  mi_assert_internal(mi_heap_page_is_valid(heap, pq, page, NULL, NULL));
+  mi_collect_t collect = *((mi_collect_t*)arg_collect);
+  _mi_page_free_collect(page, collect >= MI_FORCE);
+  if (mi_page_all_free(page)) {
+    // no more used blocks, free the page.
+    // note: this will free retired pages as well.
+    _mi_page_free(page, pq, collect >= MI_FORCE);
+  }
+  else if (collect == MI_ABANDON) {
+    // still used blocks but the thread is done; abandon the page
+    _mi_page_abandon(page, pq);
+  }
+  return true; // don't break
+}
+
+static bool mi_heap_page_never_delayed_free(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2) {
+  MI_UNUSED(arg1);
+  MI_UNUSED(arg2);
+  MI_UNUSED(heap);
+  MI_UNUSED(pq);
+  _mi_page_use_delayed_free(page, MI_NEVER_DELAYED_FREE, false);
+  return true; // don't break
+}
+
+static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect)
+{
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return;
+
+  const bool force = collect >= MI_FORCE;
+  _mi_deferred_free(heap, force);
+
+  // note: never reclaim on collect but leave it to threads that need storage to reclaim
+  const bool force_main =
+    #ifdef NDEBUG
+      collect == MI_FORCE
+    #else
+      collect >= MI_FORCE
+    #endif
+      && _mi_is_main_thread() && mi_heap_is_backing(heap) && !heap->no_reclaim;
+
+  if (force_main) {
+    // the main thread is abandoned (end-of-program), try to reclaim all abandoned segments.
+    // if all memory is freed by now, all segments should be freed.
+    _mi_abandoned_reclaim_all(heap, &heap->tld->segments);
+  }
+
+  // if abandoning, mark all pages to no longer add to delayed_free
+  if (collect == MI_ABANDON) {
+    mi_heap_visit_pages(heap, &mi_heap_page_never_delayed_free, NULL, NULL);
+  }
+
+  // free all current thread delayed blocks.
+  // (if abandoning, after this there are no more thread-delayed references into the pages.)
+  _mi_heap_delayed_free_all(heap);
+
+  // collect retired pages
+  _mi_heap_collect_retired(heap, force);
+
+  // collect all pages owned by this thread
+  mi_heap_visit_pages(heap, &mi_heap_page_collect, &collect, NULL);
+  mi_assert_internal( collect != MI_ABANDON || mi_atomic_load_ptr_acquire(mi_block_t,&heap->thread_delayed_free) == NULL );
+
+  // collect abandoned segments (in particular, purge expired parts of segments in the abandoned segment list)
+  // note: forced purge can be quite expensive if many threads are created/destroyed so we do not force on abandonment
+  _mi_abandoned_collect(heap, collect == MI_FORCE /* force? */, &heap->tld->segments);
+
+  // collect segment local caches
+  if (force) {
+    _mi_segment_thread_collect(&heap->tld->segments);
+  }
+
+  // collect regions on program-exit (or shared library unload)
+  if (force && _mi_is_main_thread() && mi_heap_is_backing(heap)) {
+    _mi_thread_data_collect();  // collect thread data cache
+    _mi_arena_collect(true /* force purge */, &heap->tld->stats);
+  }
+}
+
+void _mi_heap_collect_abandon(mi_heap_t* heap) {
+  mi_heap_collect_ex(heap, MI_ABANDON);
+}
+
+void mi_heap_collect(mi_heap_t* heap, bool force) mi_attr_noexcept {
+  mi_heap_collect_ex(heap, (force ? MI_FORCE : MI_NORMAL));
+}
+
+void mi_collect(bool force) mi_attr_noexcept {
+  mi_heap_collect(mi_prim_get_default_heap(), force);
+}
+
+
+/* -----------------------------------------------------------
+  Heap new
+----------------------------------------------------------- */
+
+mi_heap_t* mi_heap_get_default(void) {
+  mi_thread_init();
+  return mi_prim_get_default_heap();
+}
+
+static bool mi_heap_is_default(const mi_heap_t* heap) {
+  return (heap == mi_prim_get_default_heap());
+}
+
+
+mi_heap_t* mi_heap_get_backing(void) {
+  mi_heap_t* heap = mi_heap_get_default();
+  mi_assert_internal(heap!=NULL);
+  mi_heap_t* bheap = heap->tld->heap_backing;
+  mi_assert_internal(bheap!=NULL);
+  mi_assert_internal(bheap->thread_id == _mi_thread_id());
+  return bheap;
+}
+
+mi_decl_nodiscard mi_heap_t* mi_heap_new_in_arena(mi_arena_id_t arena_id) {
+  mi_heap_t* bheap = mi_heap_get_backing();
+  mi_heap_t* heap = mi_heap_malloc_tp(bheap, mi_heap_t);  // todo: OS allocate in secure mode?
+  if (heap == NULL) return NULL;
+  _mi_memcpy_aligned(heap, &_mi_heap_empty, sizeof(mi_heap_t));
+  heap->tld = bheap->tld;
+  heap->thread_id = _mi_thread_id();
+  heap->arena_id = arena_id;
+  _mi_random_split(&bheap->random, &heap->random);
+  heap->cookie = _mi_heap_random_next(heap) | 1;
+  heap->keys[0] = _mi_heap_random_next(heap);
+  heap->keys[1] = _mi_heap_random_next(heap);
+  heap->no_reclaim = true;  // don't reclaim abandoned pages or otherwise destroy is unsafe
+  // push on the thread local heaps list
+  heap->next = heap->tld->heaps;
+  heap->tld->heaps = heap;
+  return heap;
+}
+
+mi_decl_nodiscard mi_heap_t* mi_heap_new(void) {
+  return mi_heap_new_in_arena(_mi_arena_id_none());
+}
+
+bool _mi_heap_memid_is_suitable(mi_heap_t* heap, mi_memid_t memid) {
+  return _mi_arena_memid_is_suitable(memid, heap->arena_id);
+}
+
+uintptr_t _mi_heap_random_next(mi_heap_t* heap) {
+  return _mi_random_next(&heap->random);
+}
+
+// zero out the page queues
+static void mi_heap_reset_pages(mi_heap_t* heap) {
+  mi_assert_internal(heap != NULL);
+  mi_assert_internal(mi_heap_is_initialized(heap));
+  // TODO: copy full empty heap instead?
+  memset(&heap->pages_free_direct, 0, sizeof(heap->pages_free_direct));
+  _mi_memcpy_aligned(&heap->pages, &_mi_heap_empty.pages, sizeof(heap->pages));
+  heap->thread_delayed_free = NULL;
+  heap->page_count = 0;
+}
+
+// called from `mi_heap_destroy` and `mi_heap_delete` to free the internal heap resources.
+static void mi_heap_free(mi_heap_t* heap) {
+  mi_assert(heap != NULL);
+  mi_assert_internal(mi_heap_is_initialized(heap));
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return;
+  if (mi_heap_is_backing(heap)) return; // dont free the backing heap
+
+  // reset default
+  if (mi_heap_is_default(heap)) {
+    _mi_heap_set_default_direct(heap->tld->heap_backing);
+  }
+
+  // remove ourselves from the thread local heaps list
+  // linear search but we expect the number of heaps to be relatively small
+  mi_heap_t* prev = NULL;
+  mi_heap_t* curr = heap->tld->heaps;
+  while (curr != heap && curr != NULL) {
+    prev = curr;
+    curr = curr->next;
+  }
+  mi_assert_internal(curr == heap);
+  if (curr == heap) {
+    if (prev != NULL) { prev->next = heap->next; }
+		 else { heap->tld->heaps = heap->next; }
+  }
+  mi_assert_internal(heap->tld->heaps != NULL);
+
+  // and free the used memory
+  mi_free(heap);
+}
+
+
+/* -----------------------------------------------------------
+  Heap destroy
+----------------------------------------------------------- */
+
+static bool _mi_heap_page_destroy(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2) {
+  MI_UNUSED(arg1);
+  MI_UNUSED(arg2);
+  MI_UNUSED(heap);
+  MI_UNUSED(pq);
+
+  // ensure no more thread_delayed_free will be added
+  _mi_page_use_delayed_free(page, MI_NEVER_DELAYED_FREE, false);
+
+  // stats
+  const size_t bsize = mi_page_block_size(page);
+  if (bsize > MI_MEDIUM_OBJ_SIZE_MAX) {
+    if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
+      mi_heap_stat_decrease(heap, large, bsize);
+    }
+    else {
+      mi_heap_stat_decrease(heap, huge, bsize);
+    }
+  }
+#if (MI_STAT)
+  _mi_page_free_collect(page, false);  // update used count
+  const size_t inuse = page->used;
+  if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
+    mi_heap_stat_decrease(heap, normal, bsize * inuse);
+#if (MI_STAT>1)
+    mi_heap_stat_decrease(heap, normal_bins[_mi_bin(bsize)], inuse);
+#endif
+  }
+  mi_heap_stat_decrease(heap, malloc, bsize * inuse);  // todo: off for aligned blocks...
+#endif
+
+  /// pretend it is all free now
+  mi_assert_internal(mi_page_thread_free(page) == NULL);
+  page->used = 0;
+
+  // and free the page
+  // mi_page_free(page,false);
+  page->next = NULL;
+  page->prev = NULL;
+  _mi_segment_page_free(page,false /* no force? */, &heap->tld->segments);
+
+  return true; // keep going
+}
+
+void _mi_heap_destroy_pages(mi_heap_t* heap) {
+  mi_heap_visit_pages(heap, &_mi_heap_page_destroy, NULL, NULL);
+  mi_heap_reset_pages(heap);
+}
+
+#if MI_TRACK_HEAP_DESTROY
+static bool mi_cdecl mi_heap_track_block_free(const mi_heap_t* heap, const mi_heap_area_t* area, void* block, size_t block_size, void* arg) {
+  MI_UNUSED(heap); MI_UNUSED(area);  MI_UNUSED(arg); MI_UNUSED(block_size);
+  mi_track_free_size(block,mi_usable_size(block));
+  return true;
+}
+#endif
+
+void mi_heap_destroy(mi_heap_t* heap) {
+  mi_assert(heap != NULL);
+  mi_assert(mi_heap_is_initialized(heap));
+  mi_assert(heap->no_reclaim);
+  mi_assert_expensive(mi_heap_is_valid(heap));
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return;
+  if (!heap->no_reclaim) {
+    // don't free in case it may contain reclaimed pages
+    mi_heap_delete(heap);
+  }
+  else {
+    // track all blocks as freed
+    #if MI_TRACK_HEAP_DESTROY
+    mi_heap_visit_blocks(heap, true, mi_heap_track_block_free, NULL);
+    #endif
+    // free all pages
+    _mi_heap_destroy_pages(heap);
+    mi_heap_free(heap);
+  }
+}
+
+// forcefully destroy all heaps in the current thread
+void _mi_heap_unsafe_destroy_all(void) {
+  mi_heap_t* bheap = mi_heap_get_backing();
+  mi_heap_t* curr = bheap->tld->heaps;
+  while (curr != NULL) {
+    mi_heap_t* next = curr->next;
+    if (curr->no_reclaim) {
+      mi_heap_destroy(curr);
+    }
+    else {
+      _mi_heap_destroy_pages(curr);
+    }
+    curr = next;
+  }
+}
+
+/* -----------------------------------------------------------
+  Safe Heap delete
+----------------------------------------------------------- */
+
+// Transfer the pages from one heap to the other
+static void mi_heap_absorb(mi_heap_t* heap, mi_heap_t* from) {
+  mi_assert_internal(heap!=NULL);
+  if (from==NULL || from->page_count == 0) return;
+
+  // reduce the size of the delayed frees
+  _mi_heap_delayed_free_partial(from);
+
+  // transfer all pages by appending the queues; this will set a new heap field
+  // so threads may do delayed frees in either heap for a while.
+  // note: appending waits for each page to not be in the `MI_DELAYED_FREEING` state
+  // so after this only the new heap will get delayed frees
+  for (size_t i = 0; i <= MI_BIN_FULL; i++) {
+    mi_page_queue_t* pq = &heap->pages[i];
+    mi_page_queue_t* append = &from->pages[i];
+    size_t pcount = _mi_page_queue_append(heap, pq, append);
+    heap->page_count += pcount;
+    from->page_count -= pcount;
+  }
+  mi_assert_internal(from->page_count == 0);
+
+  // and do outstanding delayed frees in the `from` heap
+  // note: be careful here as the `heap` field in all those pages no longer point to `from`,
+  // turns out to be ok as `_mi_heap_delayed_free` only visits the list and calls a
+  // the regular `_mi_free_delayed_block` which is safe.
+  _mi_heap_delayed_free_all(from);
+  #if !defined(_MSC_VER) || (_MSC_VER > 1900) // somehow the following line gives an error in VS2015, issue #353
+  mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_block_t,&from->thread_delayed_free) == NULL);
+  #endif
+
+  // and reset the `from` heap
+  mi_heap_reset_pages(from);
+}
+
+// Safe delete a heap without freeing any still allocated blocks in that heap.
+void mi_heap_delete(mi_heap_t* heap)
+{
+  mi_assert(heap != NULL);
+  mi_assert(mi_heap_is_initialized(heap));
+  mi_assert_expensive(mi_heap_is_valid(heap));
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return;
+
+  if (!mi_heap_is_backing(heap)) {
+    // tranfer still used pages to the backing heap
+    mi_heap_absorb(heap->tld->heap_backing, heap);
+  }
+  else {
+    // the backing heap abandons its pages
+    _mi_heap_collect_abandon(heap);
+  }
+  mi_assert_internal(heap->page_count==0);
+  mi_heap_free(heap);
+}
+
+mi_heap_t* mi_heap_set_default(mi_heap_t* heap) {
+  mi_assert(heap != NULL);
+  mi_assert(mi_heap_is_initialized(heap));
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return NULL;
+  mi_assert_expensive(mi_heap_is_valid(heap));
+  mi_heap_t* old = mi_prim_get_default_heap();
+  _mi_heap_set_default_direct(heap);
+  return old;
+}
+
+
+
+
+/* -----------------------------------------------------------
+  Analysis
+----------------------------------------------------------- */
+
+// static since it is not thread safe to access heaps from other threads.
+static mi_heap_t* mi_heap_of_block(const void* p) {
+  if (p == NULL) return NULL;
+  mi_segment_t* segment = _mi_ptr_segment(p);
+  bool valid = (_mi_ptr_cookie(segment) == segment->cookie);
+  mi_assert_internal(valid);
+  if mi_unlikely(!valid) return NULL;
+  return mi_page_heap(_mi_segment_page_of(segment,p));
+}
+
+bool mi_heap_contains_block(mi_heap_t* heap, const void* p) {
+  mi_assert(heap != NULL);
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return false;
+  return (heap == mi_heap_of_block(p));
+}
+
+
+static bool mi_heap_page_check_owned(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* p, void* vfound) {
+  MI_UNUSED(heap);
+  MI_UNUSED(pq);
+  bool* found = (bool*)vfound;
+  mi_segment_t* segment = _mi_page_segment(page);
+  void* start = _mi_page_start(segment, page, NULL);
+  void* end   = (uint8_t*)start + (page->capacity * mi_page_block_size(page));
+  *found = (p >= start && p < end);
+  return (!*found); // continue if not found
+}
+
+bool mi_heap_check_owned(mi_heap_t* heap, const void* p) {
+  mi_assert(heap != NULL);
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return false;
+  if (((uintptr_t)p & (MI_INTPTR_SIZE - 1)) != 0) return false;  // only aligned pointers
+  bool found = false;
+  mi_heap_visit_pages(heap, &mi_heap_page_check_owned, (void*)p, &found);
+  return found;
+}
+
+bool mi_check_owned(const void* p) {
+  return mi_heap_check_owned(mi_prim_get_default_heap(), p);
+}
+
+/* -----------------------------------------------------------
+  Visit all heap blocks and areas
+  Todo: enable visiting abandoned pages, and
+	enable visiting all blocks of all heaps across threads
+----------------------------------------------------------- */
+
+// Separate struct to keep `mi_page_t` out of the public interface
+typedef struct mi_heap_area_ex_s {
+  mi_heap_area_t area;
+  mi_page_t*     page;
+} mi_heap_area_ex_t;
+
+static bool mi_heap_area_visit_blocks(const mi_heap_area_ex_t* xarea, mi_block_visit_fun* visitor, void* arg) {
+  mi_assert(xarea != NULL);
+  if (xarea==NULL) return true;
+  const mi_heap_area_t* area = &xarea->area;
+  mi_page_t* page = xarea->page;
+  mi_assert(page != NULL);
+  if (page == NULL) return true;
+
+  _mi_page_free_collect(page,true);
+  mi_assert_internal(page->local_free == NULL);
+  if (page->used == 0) return true;
+
+  const size_t bsize = mi_page_block_size(page);
+  const size_t ubsize = mi_page_usable_block_size(page); // without padding
+  size_t   psize;
+  uint8_t* pstart = _mi_page_start(_mi_page_segment(page), page, &psize);
+
+  if (page->capacity == 1) {
+    // optimize page with one block
+    mi_assert_internal(page->used == 1 && page->free == NULL);
+    return visitor(mi_page_heap(page), area, pstart, ubsize, arg);
+  }
+
+  // create a bitmap of free blocks.
+  #define MI_MAX_BLOCKS   (MI_SMALL_PAGE_SIZE / sizeof(void*))
+  uintptr_t free_map[MI_MAX_BLOCKS / sizeof(uintptr_t)];
+  memset(free_map, 0, sizeof(free_map));
+
+  #if MI_DEBUG>1
+  size_t free_count = 0;
+  #endif
+  for (mi_block_t* block = page->free; block != NULL; block = mi_block_next(page,block)) {
+    #if MI_DEBUG>1
+    free_count++;
+    #endif
+    mi_assert_internal((uint8_t*)block >= pstart && (uint8_t*)block < (pstart + psize));
+    size_t offset = (uint8_t*)block - pstart;
+    mi_assert_internal(offset % bsize == 0);
+    size_t blockidx = offset / bsize;  // Todo: avoid division?
+    mi_assert_internal( blockidx < MI_MAX_BLOCKS);
+    size_t bitidx = (blockidx / sizeof(uintptr_t));
+    size_t bit = blockidx - (bitidx * sizeof(uintptr_t));
+    free_map[bitidx] |= ((uintptr_t)1 << bit);
+  }
+  mi_assert_internal(page->capacity == (free_count + page->used));
+
+  // walk through all blocks skipping the free ones
+  #if MI_DEBUG>1
+  size_t used_count = 0;
+  #endif
+  for (size_t i = 0; i < page->capacity; i++) {
+    size_t bitidx = (i / sizeof(uintptr_t));
+    size_t bit = i - (bitidx * sizeof(uintptr_t));
+    uintptr_t m = free_map[bitidx];
+    if (bit == 0 && m == UINTPTR_MAX) {
+      i += (sizeof(uintptr_t) - 1); // skip a run of free blocks
+    }
+    else if ((m & ((uintptr_t)1 << bit)) == 0) {
+      #if MI_DEBUG>1
+      used_count++;
+      #endif
+      uint8_t* block = pstart + (i * bsize);
+      if (!visitor(mi_page_heap(page), area, block, ubsize, arg)) return false;
+    }
+  }
+  mi_assert_internal(page->used == used_count);
+  return true;
+}
+
+typedef bool (mi_heap_area_visit_fun)(const mi_heap_t* heap, const mi_heap_area_ex_t* area, void* arg);
+
+
+static bool mi_heap_visit_areas_page(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* vfun, void* arg) {
+  MI_UNUSED(heap);
+  MI_UNUSED(pq);
+  mi_heap_area_visit_fun* fun = (mi_heap_area_visit_fun*)vfun;
+  mi_heap_area_ex_t xarea;
+  const size_t bsize = mi_page_block_size(page);
+  const size_t ubsize = mi_page_usable_block_size(page);
+  xarea.page = page;
+  xarea.area.reserved = page->reserved * bsize;
+  xarea.area.committed = page->capacity * bsize;
+  xarea.area.blocks = _mi_page_start(_mi_page_segment(page), page, NULL);
+  xarea.area.used = page->used;   // number of blocks in use (#553)
+  xarea.area.block_size = ubsize;
+  xarea.area.full_block_size = bsize;
+  return fun(heap, &xarea, arg);
+}
+
+// Visit all heap pages as areas
+static bool mi_heap_visit_areas(const mi_heap_t* heap, mi_heap_area_visit_fun* visitor, void* arg) {
+  if (visitor == NULL) return false;
+  return mi_heap_visit_pages((mi_heap_t*)heap, &mi_heap_visit_areas_page, (void*)(visitor), arg); // note: function pointer to void* :-{
+}
+
+// Just to pass arguments
+typedef struct mi_visit_blocks_args_s {
+  bool  visit_blocks;
+  mi_block_visit_fun* visitor;
+  void* arg;
+} mi_visit_blocks_args_t;
+
+static bool mi_heap_area_visitor(const mi_heap_t* heap, const mi_heap_area_ex_t* xarea, void* arg) {
+  mi_visit_blocks_args_t* args = (mi_visit_blocks_args_t*)arg;
+  if (!args->visitor(heap, &xarea->area, NULL, xarea->area.block_size, args->arg)) return false;
+  if (args->visit_blocks) {
+    return mi_heap_area_visit_blocks(xarea, args->visitor, args->arg);
+  }
+  else {
+    return true;
+  }
+}
+
+// Visit all blocks in a heap
+bool mi_heap_visit_blocks(const mi_heap_t* heap, bool visit_blocks, mi_block_visit_fun* visitor, void* arg) {
+  mi_visit_blocks_args_t args = { visit_blocks, visitor, arg };
+  return mi_heap_visit_areas(heap, &mi_heap_area_visitor, &args);
+}
diff --git a/compat/mimalloc/init.c b/compat/mimalloc/init.c
new file mode 100644
index 00000000000000..4ec5812e3ce1d0
--- /dev/null
+++ b/compat/mimalloc/init.c
@@ -0,0 +1,713 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2022, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/prim.h"
+
+#include <string.h>  // memcpy, memset
+#include <stdlib.h>  // atexit
+
+
+// Empty page used to initialize the small free pages array
+const mi_page_t _mi_page_empty = {
+  0, false, false, false,
+  0,       // capacity
+  0,       // reserved capacity
+  { 0 },   // flags
+  false,   // is_zero
+  0,       // retire_expire
+  NULL,    // free
+  0,       // used
+  0,       // xblock_size
+  NULL,    // local_free
+  #if (MI_PADDING || MI_ENCODE_FREELIST)
+  { 0, 0 },
+  #endif
+  MI_ATOMIC_VAR_INIT(0), // xthread_free
+  MI_ATOMIC_VAR_INIT(0), // xheap
+  NULL, NULL
+  #if MI_INTPTR_SIZE==8
+  , { 0 }  // padding
+  #endif
+};
+
+#define MI_PAGE_EMPTY() ((mi_page_t*)&_mi_page_empty)
+
+#if (MI_SMALL_WSIZE_MAX==128)
+#if (MI_PADDING>0) && (MI_INTPTR_SIZE >= 8)
+#define MI_SMALL_PAGES_EMPTY  { MI_INIT128(MI_PAGE_EMPTY), MI_PAGE_EMPTY(), MI_PAGE_EMPTY() }
+#elif (MI_PADDING>0)
+#define MI_SMALL_PAGES_EMPTY  { MI_INIT128(MI_PAGE_EMPTY), MI_PAGE_EMPTY(), MI_PAGE_EMPTY(), MI_PAGE_EMPTY() }
+#else
+#define MI_SMALL_PAGES_EMPTY  { MI_INIT128(MI_PAGE_EMPTY), MI_PAGE_EMPTY() }
+#endif
+#else
+#error "define right initialization sizes corresponding to MI_SMALL_WSIZE_MAX"
+#endif
+
+// Empty page queues for every bin
+#define QNULL(sz)  { NULL, NULL, (sz)*sizeof(uintptr_t) }
+#define MI_PAGE_QUEUES_EMPTY \
+  { QNULL(1), \
+    QNULL(     1), QNULL(     2), QNULL(     3), QNULL(     4), QNULL(     5), QNULL(     6), QNULL(     7), QNULL(     8), /* 8 */ \
+    QNULL(    10), QNULL(    12), QNULL(    14), QNULL(    16), QNULL(    20), QNULL(    24), QNULL(    28), QNULL(    32), /* 16 */ \
+    QNULL(    40), QNULL(    48), QNULL(    56), QNULL(    64), QNULL(    80), QNULL(    96), QNULL(   112), QNULL(   128), /* 24 */ \
+    QNULL(   160), QNULL(   192), QNULL(   224), QNULL(   256), QNULL(   320), QNULL(   384), QNULL(   448), QNULL(   512), /* 32 */ \
+    QNULL(   640), QNULL(   768), QNULL(   896), QNULL(  1024), QNULL(  1280), QNULL(  1536), QNULL(  1792), QNULL(  2048), /* 40 */ \
+    QNULL(  2560), QNULL(  3072), QNULL(  3584), QNULL(  4096), QNULL(  5120), QNULL(  6144), QNULL(  7168), QNULL(  8192), /* 48 */ \
+    QNULL( 10240), QNULL( 12288), QNULL( 14336), QNULL( 16384), QNULL( 20480), QNULL( 24576), QNULL( 28672), QNULL( 32768), /* 56 */ \
+    QNULL( 40960), QNULL( 49152), QNULL( 57344), QNULL( 65536), QNULL( 81920), QNULL( 98304), QNULL(114688), QNULL(131072), /* 64 */ \
+    QNULL(163840), QNULL(196608), QNULL(229376), QNULL(262144), QNULL(327680), QNULL(393216), QNULL(458752), QNULL(524288), /* 72 */ \
+    QNULL(MI_MEDIUM_OBJ_WSIZE_MAX + 1  /* 655360, Huge queue */), \
+    QNULL(MI_MEDIUM_OBJ_WSIZE_MAX + 2) /* Full queue */ }
+
+#define MI_STAT_COUNT_NULL()  {0,0,0,0}
+
+// Empty statistics
+#if MI_STAT>1
+#define MI_STAT_COUNT_END_NULL()  , { MI_STAT_COUNT_NULL(), MI_INIT32(MI_STAT_COUNT_NULL) }
+#else
+#define MI_STAT_COUNT_END_NULL()
+#endif
+
+#define MI_STATS_NULL  \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), \
+  { 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 }, \
+  { 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 } \
+  MI_STAT_COUNT_END_NULL()
+
+
+// Empty slice span queues for every bin
+#define SQNULL(sz)  { NULL, NULL, sz }
+#define MI_SEGMENT_SPAN_QUEUES_EMPTY \
+  { SQNULL(1), \
+    SQNULL(     1), SQNULL(     2), SQNULL(     3), SQNULL(     4), SQNULL(     5), SQNULL(     6), SQNULL(     7), SQNULL(    10), /*  8 */ \
+    SQNULL(    12), SQNULL(    14), SQNULL(    16), SQNULL(    20), SQNULL(    24), SQNULL(    28), SQNULL(    32), SQNULL(    40), /* 16 */ \
+    SQNULL(    48), SQNULL(    56), SQNULL(    64), SQNULL(    80), SQNULL(    96), SQNULL(   112), SQNULL(   128), SQNULL(   160), /* 24 */ \
+    SQNULL(   192), SQNULL(   224), SQNULL(   256), SQNULL(   320), SQNULL(   384), SQNULL(   448), SQNULL(   512), SQNULL(   640), /* 32 */ \
+    SQNULL(   768), SQNULL(   896), SQNULL(  1024) /* 35 */ }
+
+
+// --------------------------------------------------------
+// Statically allocate an empty heap as the initial
+// thread local value for the default heap,
+// and statically allocate the backing heap for the main
+// thread so it can function without doing any allocation
+// itself (as accessing a thread local for the first time
+// may lead to allocation itself on some platforms)
+// --------------------------------------------------------
+
+mi_decl_cache_align const mi_heap_t _mi_heap_empty = {
+  NULL,
+  MI_SMALL_PAGES_EMPTY,
+  MI_PAGE_QUEUES_EMPTY,
+  MI_ATOMIC_VAR_INIT(NULL),
+  0,                // tid
+  0,                // cookie
+  0,                // arena id
+  { 0, 0 },         // keys
+  { {0}, {0}, 0, true }, // random
+  0,                // page count
+  MI_BIN_FULL, 0,   // page retired min/max
+  NULL,             // next
+  false
+};
+
+#define tld_empty_stats  ((mi_stats_t*)((uint8_t*)&tld_empty + offsetof(mi_tld_t,stats)))
+#define tld_empty_os     ((mi_os_tld_t*)((uint8_t*)&tld_empty + offsetof(mi_tld_t,os)))
+
+mi_decl_cache_align static const mi_tld_t tld_empty = {
+  0,
+  false,
+  NULL, NULL,
+  { MI_SEGMENT_SPAN_QUEUES_EMPTY, 0, 0, 0, 0, tld_empty_stats, tld_empty_os }, // segments
+  { 0, tld_empty_stats }, // os
+  { MI_STATS_NULL }       // stats
+};
+
+mi_threadid_t _mi_thread_id(void) mi_attr_noexcept {
+  return _mi_prim_thread_id();
+}
+
+// the thread-local default heap for allocation
+mi_decl_thread mi_heap_t* _mi_heap_default = (mi_heap_t*)&_mi_heap_empty;
+
+extern mi_heap_t _mi_heap_main;
+
+static mi_tld_t tld_main = {
+  0, false,
+  &_mi_heap_main, & _mi_heap_main,
+  { MI_SEGMENT_SPAN_QUEUES_EMPTY, 0, 0, 0, 0, &tld_main.stats, &tld_main.os }, // segments
+  { 0, &tld_main.stats },  // os
+  { MI_STATS_NULL }       // stats
+};
+
+mi_heap_t _mi_heap_main = {
+  &tld_main,
+  MI_SMALL_PAGES_EMPTY,
+  MI_PAGE_QUEUES_EMPTY,
+  MI_ATOMIC_VAR_INIT(NULL),
+  0,                // thread id
+  0,                // initial cookie
+  0,                // arena id
+  { 0, 0 },         // the key of the main heap can be fixed (unlike page keys that need to be secure!)
+  { {0x846ca68b}, {0}, 0, true },  // random
+  0,                // page count
+  MI_BIN_FULL, 0,   // page retired min/max
+  NULL,             // next heap
+  false             // can reclaim
+};
+
+bool _mi_process_is_initialized = false;  // set to `true` in `mi_process_init`.
+
+mi_stats_t _mi_stats_main = { MI_STATS_NULL };
+
+
+static void mi_heap_main_init(void) {
+  if (_mi_heap_main.cookie == 0) {
+    _mi_heap_main.thread_id = _mi_thread_id();
+    _mi_heap_main.cookie = 1;
+    #if defined(_WIN32) && !defined(MI_SHARED_LIB)
+      _mi_random_init_weak(&_mi_heap_main.random);    // prevent allocation failure during bcrypt dll initialization with static linking
+    #else
+      _mi_random_init(&_mi_heap_main.random);
+    #endif
+    _mi_heap_main.cookie  = _mi_heap_random_next(&_mi_heap_main);
+    _mi_heap_main.keys[0] = _mi_heap_random_next(&_mi_heap_main);
+    _mi_heap_main.keys[1] = _mi_heap_random_next(&_mi_heap_main);
+  }
+}
+
+mi_heap_t* _mi_heap_main_get(void) {
+  mi_heap_main_init();
+  return &_mi_heap_main;
+}
+
+
+/* -----------------------------------------------------------
+  Initialization and freeing of the thread local heaps
+----------------------------------------------------------- */
+
+// note: in x64 in release build `sizeof(mi_thread_data_t)` is under 4KiB (= OS page size).
+typedef struct mi_thread_data_s {
+  mi_heap_t  heap;  // must come first due to cast in `_mi_heap_done`
+  mi_tld_t   tld;
+  mi_memid_t memid;
+} mi_thread_data_t;
+
+
+// Thread meta-data is allocated directly from the OS. For
+// some programs that do not use thread pools and allocate and
+// destroy many OS threads, this may causes too much overhead
+// per thread so we maintain a small cache of recently freed metadata.
+
+#define TD_CACHE_SIZE (16)
+static _Atomic(mi_thread_data_t*) td_cache[TD_CACHE_SIZE];
+
+static mi_thread_data_t* mi_thread_data_zalloc(void) {
+  // try to find thread metadata in the cache
+  bool is_zero = false;
+  mi_thread_data_t* td = NULL;
+  for (int i = 0; i < TD_CACHE_SIZE; i++) {
+    td = mi_atomic_load_ptr_relaxed(mi_thread_data_t, &td_cache[i]);
+    if (td != NULL) {
+      // found cached allocation, try use it
+      td = mi_atomic_exchange_ptr_acq_rel(mi_thread_data_t, &td_cache[i], NULL);
+      if (td != NULL) {
+	break;
+      }
+    }
+  }
+
+  // if that fails, allocate as meta data
+  if (td == NULL) {
+    mi_memid_t memid;
+    td = (mi_thread_data_t*)_mi_os_alloc(sizeof(mi_thread_data_t), &memid, &_mi_stats_main);
+    if (td == NULL) {
+      // if this fails, try once more. (issue #257)
+      td = (mi_thread_data_t*)_mi_os_alloc(sizeof(mi_thread_data_t), &memid, &_mi_stats_main);
+      if (td == NULL) {
+	// really out of memory
+	_mi_error_message(ENOMEM, "unable to allocate thread local heap metadata (%zu bytes)\n", sizeof(mi_thread_data_t));
+      }
+    }
+    if (td != NULL) {
+      td->memid = memid;
+      is_zero = memid.initially_zero;
+    }
+  }
+
+  if (td != NULL && !is_zero) {
+    _mi_memzero_aligned(td, sizeof(*td));
+  }
+  return td;
+}
+
+static void mi_thread_data_free( mi_thread_data_t* tdfree ) {
+  // try to add the thread metadata to the cache
+  for (int i = 0; i < TD_CACHE_SIZE; i++) {
+    mi_thread_data_t* td = mi_atomic_load_ptr_relaxed(mi_thread_data_t, &td_cache[i]);
+    if (td == NULL) {
+      mi_thread_data_t* expected = NULL;
+      if (mi_atomic_cas_ptr_weak_acq_rel(mi_thread_data_t, &td_cache[i], &expected, tdfree)) {
+	return;
+      }
+    }
+  }
+  // if that fails, just free it directly
+  _mi_os_free(tdfree, sizeof(mi_thread_data_t), tdfree->memid, &_mi_stats_main);
+}
+
+void _mi_thread_data_collect(void) {
+  // free all thread metadata from the cache
+  for (int i = 0; i < TD_CACHE_SIZE; i++) {
+    mi_thread_data_t* td = mi_atomic_load_ptr_relaxed(mi_thread_data_t, &td_cache[i]);
+    if (td != NULL) {
+      td = mi_atomic_exchange_ptr_acq_rel(mi_thread_data_t, &td_cache[i], NULL);
+      if (td != NULL) {
+	_mi_os_free(td, sizeof(mi_thread_data_t), td->memid, &_mi_stats_main);
+      }
+    }
+  }
+}
+
+// Initialize the thread local default heap, called from `mi_thread_init`
+static bool _mi_heap_init(void) {
+  if (mi_heap_is_initialized(mi_prim_get_default_heap())) return true;
+  if (_mi_is_main_thread()) {
+    // mi_assert_internal(_mi_heap_main.thread_id != 0);  // can happen on freeBSD where alloc is called before any initialization
+    // the main heap is statically allocated
+    mi_heap_main_init();
+    _mi_heap_set_default_direct(&_mi_heap_main);
+    //mi_assert_internal(_mi_heap_default->tld->heap_backing == mi_prim_get_default_heap());
+  }
+  else {
+    // use `_mi_os_alloc` to allocate directly from the OS
+    mi_thread_data_t* td = mi_thread_data_zalloc();
+    if (td == NULL) return false;
+
+    mi_tld_t*  tld = &td->tld;
+    mi_heap_t* heap = &td->heap;
+    _mi_memcpy_aligned(tld, &tld_empty, sizeof(*tld));
+    _mi_memcpy_aligned(heap, &_mi_heap_empty, sizeof(*heap));
+    heap->thread_id = _mi_thread_id();
+    #if defined(_WIN32) && !defined(MI_SHARED_LIB)
+    _mi_random_init_weak(&heap->random); // match mi_heap_main_init()
+    #else
+    _mi_random_init(&heap->random);
+    #endif
+    heap->cookie  = _mi_heap_random_next(heap) | 1;
+    heap->keys[0] = _mi_heap_random_next(heap);
+    heap->keys[1] = _mi_heap_random_next(heap);
+    heap->tld = tld;
+    tld->heap_backing = heap;
+    tld->heaps = heap;
+    tld->segments.stats = &tld->stats;
+    tld->segments.os = &tld->os;
+    tld->os.stats = &tld->stats;
+    _mi_heap_set_default_direct(heap);
+  }
+  return false;
+}
+
+// Free the thread local default heap (called from `mi_thread_done`)
+static bool _mi_heap_done(mi_heap_t* heap) {
+  if (!mi_heap_is_initialized(heap)) return true;
+
+  // reset default heap
+  _mi_heap_set_default_direct(_mi_is_main_thread() ? &_mi_heap_main : (mi_heap_t*)&_mi_heap_empty);
+
+  // switch to backing heap
+  heap = heap->tld->heap_backing;
+  if (!mi_heap_is_initialized(heap)) return false;
+
+  // delete all non-backing heaps in this thread
+  mi_heap_t* curr = heap->tld->heaps;
+  while (curr != NULL) {
+    mi_heap_t* next = curr->next; // save `next` as `curr` will be freed
+    if (curr != heap) {
+      mi_assert_internal(!mi_heap_is_backing(curr));
+      mi_heap_delete(curr);
+    }
+    curr = next;
+  }
+  mi_assert_internal(heap->tld->heaps == heap && heap->next == NULL);
+  mi_assert_internal(mi_heap_is_backing(heap));
+
+  // collect if not the main thread
+  if (heap != &_mi_heap_main) {
+    _mi_heap_collect_abandon(heap);
+  }
+
+  // merge stats
+  _mi_stats_done(&heap->tld->stats);
+
+  // free if not the main thread
+  if (heap != &_mi_heap_main) {
+    // the following assertion does not always hold for huge segments as those are always treated
+    // as abondened: one may allocate it in one thread, but deallocate in another in which case
+    // the count can be too large or negative. todo: perhaps not count huge segments? see issue #363
+    // mi_assert_internal(heap->tld->segments.count == 0 || heap->thread_id != _mi_thread_id());
+    mi_thread_data_free((mi_thread_data_t*)heap);
+  }
+  else {
+    #if 0
+    // never free the main thread even in debug mode; if a dll is linked statically with mimalloc,
+    // there may still be delete/free calls after the mi_fls_done is called. Issue #207
+    _mi_heap_destroy_pages(heap);
+    mi_assert_internal(heap->tld->heap_backing == &_mi_heap_main);
+    #endif
+  }
+  return false;
+}
+
+
+
+// --------------------------------------------------------
+// Try to run `mi_thread_done()` automatically so any memory
+// owned by the thread but not yet released can be abandoned
+// and re-owned by another thread.
+//
+// 1. windows dynamic library:
+//     call from DllMain on DLL_THREAD_DETACH
+// 2. windows static library:
+//     use `FlsAlloc` to call a destructor when the thread is done
+// 3. unix, pthreads:
+//     use a pthread key to call a destructor when a pthread is done
+//
+// In the last two cases we also need to call `mi_process_init`
+// to set up the thread local keys.
+// --------------------------------------------------------
+
+// Set up handlers so `mi_thread_done` is called automatically
+static void mi_process_setup_auto_thread_done(void) {
+  static bool tls_initialized = false; // fine if it races
+  if (tls_initialized) return;
+  tls_initialized = true;
+  _mi_prim_thread_init_auto_done();
+  _mi_heap_set_default_direct(&_mi_heap_main);
+}
+
+
+bool _mi_is_main_thread(void) {
+  return (_mi_heap_main.thread_id==0 || _mi_heap_main.thread_id == _mi_thread_id());
+}
+
+static _Atomic(size_t) thread_count = MI_ATOMIC_VAR_INIT(1);
+
+size_t  _mi_current_thread_count(void) {
+  return mi_atomic_load_relaxed(&thread_count);
+}
+
+// This is called from the `mi_malloc_generic`
+void mi_thread_init(void) mi_attr_noexcept
+{
+  // ensure our process has started already
+  mi_process_init();
+
+  // initialize the thread local default heap
+  // (this will call `_mi_heap_set_default_direct` and thus set the
+  //  fiber/pthread key to a non-zero value, ensuring `_mi_thread_done` is called)
+  if (_mi_heap_init()) return;  // returns true if already initialized
+
+  _mi_stat_increase(&_mi_stats_main.threads, 1);
+  mi_atomic_increment_relaxed(&thread_count);
+  //_mi_verbose_message("thread init: 0x%zx\n", _mi_thread_id());
+}
+
+void mi_thread_done(void) mi_attr_noexcept {
+  _mi_thread_done(NULL);
+}
+
+void _mi_thread_done(mi_heap_t* heap)
+{
+  // calling with NULL implies using the default heap
+  if (heap == NULL) {
+    heap = mi_prim_get_default_heap();
+    if (heap == NULL) return;
+  }
+
+  // prevent re-entrancy through heap_done/heap_set_default_direct (issue #699)
+  if (!mi_heap_is_initialized(heap)) {
+    return;
+  }
+
+  // adjust stats
+  mi_atomic_decrement_relaxed(&thread_count);
+  _mi_stat_decrease(&_mi_stats_main.threads, 1);
+
+  // check thread-id as on Windows shutdown with FLS the main (exit) thread may call this on thread-local heaps...
+  if (heap->thread_id != _mi_thread_id()) return;
+
+  // abandon the thread local heap
+  if (_mi_heap_done(heap)) return;  // returns true if already ran
+}
+
+void _mi_heap_set_default_direct(mi_heap_t* heap)  {
+  mi_assert_internal(heap != NULL);
+  #if defined(MI_TLS_SLOT)
+  mi_prim_tls_slot_set(MI_TLS_SLOT,heap);
+  #elif defined(MI_TLS_PTHREAD_SLOT_OFS)
+  *mi_tls_pthread_heap_slot() = heap;
+  #elif defined(MI_TLS_PTHREAD)
+  // we use _mi_heap_default_key
+  #else
+  _mi_heap_default = heap;
+  #endif
+
+  // ensure the default heap is passed to `_mi_thread_done`
+  // setting to a non-NULL value also ensures `mi_thread_done` is called.
+  _mi_prim_thread_associate_default_heap(heap);
+}
+
+
+// --------------------------------------------------------
+// Run functions on process init/done, and thread init/done
+// --------------------------------------------------------
+static void mi_cdecl mi_process_done(void);
+
+static bool os_preloading = true;    // true until this module is initialized
+static bool mi_redirected = false;   // true if malloc redirects to mi_malloc
+
+// Returns true if this module has not been initialized; Don't use C runtime routines until it returns false.
+bool mi_decl_noinline _mi_preloading(void) {
+  return os_preloading;
+}
+
+mi_decl_nodiscard bool mi_is_redirected(void) mi_attr_noexcept {
+  return mi_redirected;
+}
+
+// Communicate with the redirection module on Windows
+#if defined(_WIN32) && defined(MI_SHARED_LIB) && !defined(MI_WIN_NOREDIRECT)
+#ifdef __cplusplus
+extern "C" {
+#endif
+mi_decl_export void _mi_redirect_entry(DWORD reason) {
+  // called on redirection; careful as this may be called before DllMain
+  if (reason == DLL_PROCESS_ATTACH) {
+    mi_redirected = true;
+  }
+  else if (reason == DLL_PROCESS_DETACH) {
+    mi_redirected = false;
+  }
+  else if (reason == DLL_THREAD_DETACH) {
+    mi_thread_done();
+  }
+}
+__declspec(dllimport) bool mi_cdecl mi_allocator_init(const char** message);
+__declspec(dllimport) void mi_cdecl mi_allocator_done(void);
+#ifdef __cplusplus
+}
+#endif
+#else
+static bool mi_allocator_init(const char** message) {
+  if (message != NULL) *message = NULL;
+  return true;
+}
+static void mi_allocator_done(void) {
+  // nothing to do
+}
+#endif
+
+// Called once by the process loader
+static void mi_process_load(void) {
+  mi_heap_main_init();
+  #if defined(__APPLE__) || defined(MI_TLS_RECURSE_GUARD)
+  volatile mi_heap_t* dummy = _mi_heap_default; // access TLS to allocate it before setting tls_initialized to true;
+  if (dummy == NULL) return;                    // use dummy or otherwise the access may get optimized away (issue #697)
+  #endif
+  os_preloading = false;
+  mi_assert_internal(_mi_is_main_thread());
+  #if !(defined(_WIN32) && defined(MI_SHARED_LIB))  // use Dll process detach (see below) instead of atexit (issue #521)
+  atexit(&mi_process_done);
+  #endif
+  _mi_options_init();
+  mi_process_setup_auto_thread_done();
+  mi_process_init();
+  if (mi_redirected) _mi_verbose_message("malloc is redirected.\n");
+
+  // show message from the redirector (if present)
+  const char* msg = NULL;
+  mi_allocator_init(&msg);
+  if (msg != NULL && (mi_option_is_enabled(mi_option_verbose) || mi_option_is_enabled(mi_option_show_errors))) {
+    _mi_fputs(NULL,NULL,NULL,msg);
+  }
+
+  // reseed random
+  _mi_random_reinit_if_weak(&_mi_heap_main.random);
+}
+
+#if defined(_WIN32) && (defined(_M_IX86) || defined(_M_X64))
+#include <intrin.h>
+mi_decl_cache_align bool _mi_cpu_has_fsrm = false;
+
+static void mi_detect_cpu_features(void) {
+  // FSRM for fast rep movsb support (AMD Zen3+ (~2020) or Intel Ice Lake+ (~2017))
+  int32_t cpu_info[4];
+  __cpuid(cpu_info, 7);
+  _mi_cpu_has_fsrm = ((cpu_info[3] & (1 << 4)) != 0); // bit 4 of EDX : see <https://en.wikipedia.org/wiki/CPUID#EAX=7,_ECX=0:_Extended_Features>
+}
+#else
+static void mi_detect_cpu_features(void) {
+  // nothing
+}
+#endif
+
+// Initialize the process; called by thread_init or the process loader
+void mi_process_init(void) mi_attr_noexcept {
+  // ensure we are called once
+  static mi_atomic_once_t process_init;
+	#if _MSC_VER < 1920
+	mi_heap_main_init(); // vs2017 can dynamically re-initialize _mi_heap_main
+	#endif
+  if (!mi_atomic_once(&process_init)) return;
+  _mi_process_is_initialized = true;
+  _mi_verbose_message("process init: 0x%zx\n", _mi_thread_id());
+  mi_process_setup_auto_thread_done();
+
+  mi_detect_cpu_features();
+  _mi_os_init();
+  mi_heap_main_init();
+  #if MI_DEBUG
+  _mi_verbose_message("debug level : %d\n", MI_DEBUG);
+  #endif
+  _mi_verbose_message("secure level: %d\n", MI_SECURE);
+  _mi_verbose_message("mem tracking: %s\n", MI_TRACK_TOOL);
+  #if MI_TSAN
+  _mi_verbose_message("thread santizer enabled\n");
+  #endif
+  mi_thread_init();
+
+  #if defined(_WIN32)
+  // On windows, when building as a static lib the FLS cleanup happens to early for the main thread.
+  // To avoid this, set the FLS value for the main thread to NULL so the fls cleanup
+  // will not call _mi_thread_done on the (still executing) main thread. See issue #508.
+  _mi_prim_thread_associate_default_heap(NULL);
+  #endif
+
+  mi_stats_reset();  // only call stat reset *after* thread init (or the heap tld == NULL)
+  mi_track_init();
+
+  if (mi_option_is_enabled(mi_option_reserve_huge_os_pages)) {
+    size_t pages = mi_option_get_clamp(mi_option_reserve_huge_os_pages, 0, 128*1024);
+    long reserve_at = mi_option_get(mi_option_reserve_huge_os_pages_at);
+    if (reserve_at != -1) {
+      mi_reserve_huge_os_pages_at(pages, reserve_at, pages*500);
+    } else {
+      mi_reserve_huge_os_pages_interleave(pages, 0, pages*500);
+    }
+  }
+  if (mi_option_is_enabled(mi_option_reserve_os_memory)) {
+    long ksize = mi_option_get(mi_option_reserve_os_memory);
+    if (ksize > 0) {
+      mi_reserve_os_memory((size_t)ksize*MI_KiB, true /* commit? */, true /* allow large pages? */);
+    }
+  }
+}
+
+// Called when the process is done (through `at_exit`)
+static void mi_cdecl mi_process_done(void) {
+  // only shutdown if we were initialized
+  if (!_mi_process_is_initialized) return;
+  // ensure we are called once
+  static bool process_done = false;
+  if (process_done) return;
+  process_done = true;
+
+  // release any thread specific resources and ensure _mi_thread_done is called on all but the main thread
+  _mi_prim_thread_done_auto_done();
+
+  #ifndef MI_SKIP_COLLECT_ON_EXIT
+    #if (MI_DEBUG || !defined(MI_SHARED_LIB))
+    // free all memory if possible on process exit. This is not needed for a stand-alone process
+    // but should be done if mimalloc is statically linked into another shared library which
+    // is repeatedly loaded/unloaded, see issue #281.
+    mi_collect(true /* force */ );
+    #endif
+  #endif
+
+  // Forcefully release all retained memory; this can be dangerous in general if overriding regular malloc/free
+  // since after process_done there might still be other code running that calls `free` (like at_exit routines,
+  // or C-runtime termination code.
+  if (mi_option_is_enabled(mi_option_destroy_on_exit)) {
+    mi_collect(true /* force */);
+    _mi_heap_unsafe_destroy_all();     // forcefully release all memory held by all heaps (of this thread only!)
+    _mi_arena_unsafe_destroy_all(& _mi_heap_main_get()->tld->stats);
+  }
+
+  if (mi_option_is_enabled(mi_option_show_stats) || mi_option_is_enabled(mi_option_verbose)) {
+    mi_stats_print(NULL);
+  }
+  mi_allocator_done();
+  _mi_verbose_message("process done: 0x%zx\n", _mi_heap_main.thread_id);
+  os_preloading = true; // don't call the C runtime anymore
+}
+
+
+
+#if defined(_WIN32) && defined(MI_SHARED_LIB)
+  // Windows DLL: easy to hook into process_init and thread_done
+  __declspec(dllexport) BOOL WINAPI DllMain(HINSTANCE inst, DWORD reason, LPVOID reserved) {
+    MI_UNUSED(reserved);
+    MI_UNUSED(inst);
+    if (reason==DLL_PROCESS_ATTACH) {
+      mi_process_load();
+    }
+    else if (reason==DLL_PROCESS_DETACH) {
+      mi_process_done();
+    }
+    else if (reason==DLL_THREAD_DETACH) {
+      if (!mi_is_redirected()) {
+	mi_thread_done();
+      }
+    }
+    return TRUE;
+  }
+
+#elif defined(_MSC_VER)
+  // MSVC: use data section magic for static libraries
+  // See <https://www.codeguru.com/cpp/misc/misc/applicationcontrol/article.php/c6945/Running-Code-Before-and-After-Main.htm>
+  static int _mi_process_init(void) {
+    mi_process_load();
+    return 0;
+  }
+  typedef int(*_mi_crt_callback_t)(void);
+  #if defined(_M_X64) || defined(_M_ARM64)
+    __pragma(comment(linker, "/include:" "_mi_msvc_initu"))
+    #pragma section(".CRT$XIU", long, read)
+  #else
+    __pragma(comment(linker, "/include:" "__mi_msvc_initu"))
+  #endif
+  #pragma data_seg(".CRT$XIU")
+  mi_decl_externc _mi_crt_callback_t _mi_msvc_initu[] = { &_mi_process_init };
+  #pragma data_seg()
+
+#elif defined(__cplusplus)
+  // C++: use static initialization to detect process start
+  static bool _mi_process_init(void) {
+    mi_process_load();
+    return (_mi_heap_main.thread_id != 0);
+  }
+  static bool mi_initialized = _mi_process_init();
+
+#elif defined(__GNUC__) || defined(__clang__)
+  // GCC,Clang: use the constructor attribute
+  static void __attribute__((constructor)) _mi_process_init(void) {
+    mi_process_load();
+  }
+
+#else
+#pragma message("define a way to call mi_process_load on your platform")
+#endif
diff --git a/compat/mimalloc/mimalloc.h b/compat/mimalloc/mimalloc.h
new file mode 100644
index 00000000000000..7e3b5dd66e91a0
--- /dev/null
+++ b/compat/mimalloc/mimalloc.h
@@ -0,0 +1,566 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_H
+#define MIMALLOC_H
+
+#define MI_MALLOC_VERSION 212   // major + 2 digits minor
+
+// ------------------------------------------------------
+// Compiler specific attributes
+// ------------------------------------------------------
+
+#ifdef __cplusplus
+  #if (__cplusplus >= 201103L) || (_MSC_VER > 1900)  // C++11
+    #define mi_attr_noexcept   noexcept
+  #else
+    #define mi_attr_noexcept   throw()
+  #endif
+#else
+  #define mi_attr_noexcept
+#endif
+
+#if defined(__cplusplus) && (__cplusplus >= 201703)
+  #define mi_decl_nodiscard    [[nodiscard]]
+#elif (defined(__GNUC__) && (__GNUC__ >= 4)) || defined(__clang__)  // includes clang, icc, and clang-cl
+  #define mi_decl_nodiscard    __attribute__((warn_unused_result))
+#elif defined(_HAS_NODISCARD)
+  #define mi_decl_nodiscard    _NODISCARD
+#elif (_MSC_VER >= 1700)
+  #define mi_decl_nodiscard    _Check_return_
+#else
+  #define mi_decl_nodiscard
+#endif
+
+#if defined(_MSC_VER) || defined(__MINGW32__)
+  #if !defined(MI_SHARED_LIB)
+    #define mi_decl_export
+  #elif defined(MI_SHARED_LIB_EXPORT)
+    #define mi_decl_export              __declspec(dllexport)
+  #else
+    #define mi_decl_export              __declspec(dllimport)
+  #endif
+  #if defined(__MINGW32__)
+    #define mi_decl_restrict
+    #define mi_attr_malloc              __attribute__((malloc))
+  #else
+    #if (_MSC_VER >= 1900) && !defined(__EDG__)
+      #define mi_decl_restrict          __declspec(allocator) __declspec(restrict)
+    #else
+      #define mi_decl_restrict          __declspec(restrict)
+    #endif
+    #define mi_attr_malloc
+  #endif
+  #define mi_cdecl                      __cdecl
+  #define mi_attr_alloc_size(s)
+  #define mi_attr_alloc_size2(s1,s2)
+  #define mi_attr_alloc_align(p)
+#elif defined(__GNUC__)                 // includes clang and icc
+  #if defined(MI_SHARED_LIB) && defined(MI_SHARED_LIB_EXPORT)
+    #define mi_decl_export              __attribute__((visibility("default")))
+  #else
+    #define mi_decl_export
+  #endif
+  #define mi_cdecl                      // leads to warnings... __attribute__((cdecl))
+  #define mi_decl_restrict
+  #define mi_attr_malloc                __attribute__((malloc))
+  #if (defined(__clang_major__) && (__clang_major__ < 4)) || (__GNUC__ < 5)
+    #define mi_attr_alloc_size(s)
+    #define mi_attr_alloc_size2(s1,s2)
+    #define mi_attr_alloc_align(p)
+  #elif defined(__INTEL_COMPILER)
+    #define mi_attr_alloc_size(s)       __attribute__((alloc_size(s)))
+    #define mi_attr_alloc_size2(s1,s2)  __attribute__((alloc_size(s1,s2)))
+    #define mi_attr_alloc_align(p)
+  #else
+    #define mi_attr_alloc_size(s)       __attribute__((alloc_size(s)))
+    #define mi_attr_alloc_size2(s1,s2)  __attribute__((alloc_size(s1,s2)))
+    #define mi_attr_alloc_align(p)      __attribute__((alloc_align(p)))
+  #endif
+#else
+  #define mi_cdecl
+  #define mi_decl_export
+  #define mi_decl_restrict
+  #define mi_attr_malloc
+  #define mi_attr_alloc_size(s)
+  #define mi_attr_alloc_size2(s1,s2)
+  #define mi_attr_alloc_align(p)
+#endif
+
+// ------------------------------------------------------
+// Includes
+// ------------------------------------------------------
+
+#include "git-compat-util.h"
+
+#include <stdbool.h>    // bool
+#include <stdint.h>     // INTPTR_MAX
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// ------------------------------------------------------
+// Standard malloc interface
+// ------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_malloc(size_t size)  mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_calloc(size_t count, size_t size)  mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(1,2);
+mi_decl_nodiscard mi_decl_export void* mi_realloc(void* p, size_t newsize)      mi_attr_noexcept mi_attr_alloc_size(2);
+mi_decl_export void* mi_expand(void* p, size_t newsize)                         mi_attr_noexcept mi_attr_alloc_size(2);
+
+mi_decl_export void mi_free(void* p) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_strdup(const char* s) mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_strndup(const char* s, size_t n) mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_realpath(const char* fname, char* resolved_name) mi_attr_noexcept mi_attr_malloc;
+
+// ------------------------------------------------------
+// Extended functionality
+// ------------------------------------------------------
+#define MI_SMALL_WSIZE_MAX  (128)
+#define MI_SMALL_SIZE_MAX   (MI_SMALL_WSIZE_MAX*sizeof(void*))
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_malloc_small(size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_zalloc_small(size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_zalloc(size_t size)       mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_mallocn(size_t count, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(1,2);
+mi_decl_nodiscard mi_decl_export void* mi_reallocn(void* p, size_t count, size_t size)        mi_attr_noexcept mi_attr_alloc_size2(2,3);
+mi_decl_nodiscard mi_decl_export void* mi_reallocf(void* p, size_t newsize)                   mi_attr_noexcept mi_attr_alloc_size(2);
+
+mi_decl_nodiscard mi_decl_export size_t mi_usable_size(const void* p) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export size_t mi_good_size(size_t size)     mi_attr_noexcept;
+
+
+// ------------------------------------------------------
+// Internals
+// ------------------------------------------------------
+
+typedef void (mi_cdecl mi_deferred_free_fun)(bool force, unsigned long long heartbeat, void* arg);
+mi_decl_export void mi_register_deferred_free(mi_deferred_free_fun* deferred_free, void* arg) mi_attr_noexcept;
+
+typedef void (mi_cdecl mi_output_fun)(const char* msg, void* arg);
+mi_decl_export void mi_register_output(mi_output_fun* out, void* arg) mi_attr_noexcept;
+
+typedef void (mi_cdecl mi_error_fun)(int err, void* arg);
+mi_decl_export void mi_register_error(mi_error_fun* fun, void* arg);
+
+mi_decl_export void mi_collect(bool force)    mi_attr_noexcept;
+mi_decl_export int  mi_version(void)          mi_attr_noexcept;
+mi_decl_export void mi_stats_reset(void)      mi_attr_noexcept;
+mi_decl_export void mi_stats_merge(void)      mi_attr_noexcept;
+mi_decl_export void mi_stats_print(void* out) mi_attr_noexcept;  // backward compatibility: `out` is ignored and should be NULL
+mi_decl_export void mi_stats_print_out(mi_output_fun* out, void* arg) mi_attr_noexcept;
+
+mi_decl_export void mi_process_init(void)     mi_attr_noexcept;
+mi_decl_export void mi_thread_init(void)      mi_attr_noexcept;
+mi_decl_export void mi_thread_done(void)      mi_attr_noexcept;
+mi_decl_export void mi_thread_stats_print_out(mi_output_fun* out, void* arg) mi_attr_noexcept;
+
+mi_decl_export void mi_process_info(size_t* elapsed_msecs, size_t* user_msecs, size_t* system_msecs,
+				    size_t* current_rss, size_t* peak_rss,
+				    size_t* current_commit, size_t* peak_commit, size_t* page_faults) mi_attr_noexcept;
+
+// -------------------------------------------------------------------------------------
+// Aligned allocation
+// Note that `alignment` always follows `size` for consistency with unaligned
+// allocation, but unfortunately this differs from `posix_memalign` and `aligned_alloc`.
+// -------------------------------------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_malloc_aligned(size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1) mi_attr_alloc_align(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_malloc_aligned_at(size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_zalloc_aligned(size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1) mi_attr_alloc_align(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_zalloc_aligned_at(size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_calloc_aligned(size_t count, size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(1,2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_calloc_aligned_at(size_t count, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(1,2);
+mi_decl_nodiscard mi_decl_export void* mi_realloc_aligned(void* p, size_t newsize, size_t alignment) mi_attr_noexcept mi_attr_alloc_size(2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export void* mi_realloc_aligned_at(void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size(2);
+
+
+// -------------------------------------------------------------------------------------
+// Heaps: first-class, but can only allocate from the same thread that created it.
+// -------------------------------------------------------------------------------------
+
+struct mi_heap_s;
+typedef struct mi_heap_s mi_heap_t;
+
+mi_decl_nodiscard mi_decl_export mi_heap_t* mi_heap_new(void);
+mi_decl_export void       mi_heap_delete(mi_heap_t* heap);
+mi_decl_export void       mi_heap_destroy(mi_heap_t* heap);
+mi_decl_export mi_heap_t* mi_heap_set_default(mi_heap_t* heap);
+mi_decl_export mi_heap_t* mi_heap_get_default(void);
+mi_decl_export mi_heap_t* mi_heap_get_backing(void);
+mi_decl_export void       mi_heap_collect(mi_heap_t* heap, bool force) mi_attr_noexcept;
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_malloc(mi_heap_t* heap, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_zalloc(mi_heap_t* heap, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_calloc(mi_heap_t* heap, size_t count, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(2, 3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_mallocn(mi_heap_t* heap, size_t count, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(2, 3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_malloc_small(mi_heap_t* heap, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+
+mi_decl_nodiscard mi_decl_export void* mi_heap_realloc(mi_heap_t* heap, void* p, size_t newsize)              mi_attr_noexcept mi_attr_alloc_size(3);
+mi_decl_nodiscard mi_decl_export void* mi_heap_reallocn(mi_heap_t* heap, void* p, size_t count, size_t size)  mi_attr_noexcept mi_attr_alloc_size2(3,4);
+mi_decl_nodiscard mi_decl_export void* mi_heap_reallocf(mi_heap_t* heap, void* p, size_t newsize)             mi_attr_noexcept mi_attr_alloc_size(3);
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_heap_strdup(mi_heap_t* heap, const char* s)            mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_heap_strndup(mi_heap_t* heap, const char* s, size_t n) mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_heap_realpath(mi_heap_t* heap, const char* fname, char* resolved_name) mi_attr_noexcept mi_attr_malloc;
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_malloc_aligned(mi_heap_t* heap, size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_malloc_aligned_at(mi_heap_t* heap, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_zalloc_aligned(mi_heap_t* heap, size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_zalloc_aligned_at(mi_heap_t* heap, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_calloc_aligned(mi_heap_t* heap, size_t count, size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(2, 3) mi_attr_alloc_align(4);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_calloc_aligned_at(mi_heap_t* heap, size_t count, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(2, 3);
+mi_decl_nodiscard mi_decl_export void* mi_heap_realloc_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment) mi_attr_noexcept mi_attr_alloc_size(3) mi_attr_alloc_align(4);
+mi_decl_nodiscard mi_decl_export void* mi_heap_realloc_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size(3);
+
+
+// --------------------------------------------------------------------------------
+// Zero initialized re-allocation.
+// Only valid on memory that was originally allocated with zero initialization too.
+// e.g. `mi_calloc`, `mi_zalloc`, `mi_zalloc_aligned` etc.
+// see <https://github.com/microsoft/mimalloc/issues/63#issuecomment-508272992>
+// --------------------------------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_export void* mi_rezalloc(void* p, size_t newsize)                mi_attr_noexcept mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export void* mi_recalloc(void* p, size_t newcount, size_t size)  mi_attr_noexcept mi_attr_alloc_size2(2,3);
+
+mi_decl_nodiscard mi_decl_export void* mi_rezalloc_aligned(void* p, size_t newsize, size_t alignment) mi_attr_noexcept mi_attr_alloc_size(2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export void* mi_rezalloc_aligned_at(void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export void* mi_recalloc_aligned(void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept mi_attr_alloc_size2(2,3) mi_attr_alloc_align(4);
+mi_decl_nodiscard mi_decl_export void* mi_recalloc_aligned_at(void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size2(2,3);
+
+mi_decl_nodiscard mi_decl_export void* mi_heap_rezalloc(mi_heap_t* heap, void* p, size_t newsize)                mi_attr_noexcept mi_attr_alloc_size(3);
+mi_decl_nodiscard mi_decl_export void* mi_heap_recalloc(mi_heap_t* heap, void* p, size_t newcount, size_t size)  mi_attr_noexcept mi_attr_alloc_size2(3,4);
+
+mi_decl_nodiscard mi_decl_export void* mi_heap_rezalloc_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment) mi_attr_noexcept mi_attr_alloc_size(3) mi_attr_alloc_align(4);
+mi_decl_nodiscard mi_decl_export void* mi_heap_rezalloc_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size(3);
+mi_decl_nodiscard mi_decl_export void* mi_heap_recalloc_aligned(mi_heap_t* heap, void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept mi_attr_alloc_size2(3,4) mi_attr_alloc_align(5);
+mi_decl_nodiscard mi_decl_export void* mi_heap_recalloc_aligned_at(mi_heap_t* heap, void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size2(3,4);
+
+
+// ------------------------------------------------------
+// Analysis
+// ------------------------------------------------------
+
+mi_decl_export bool mi_heap_contains_block(mi_heap_t* heap, const void* p);
+mi_decl_export bool mi_heap_check_owned(mi_heap_t* heap, const void* p);
+mi_decl_export bool mi_check_owned(const void* p);
+
+// An area of heap space contains blocks of a single size.
+typedef struct mi_heap_area_s {
+  void*  blocks;      // start of the area containing heap blocks
+  size_t reserved;    // bytes reserved for this area (virtual)
+  size_t committed;   // current available bytes for this area
+  size_t used;        // number of allocated blocks
+  size_t block_size;  // size in bytes of each block
+  size_t full_block_size; // size in bytes of a full block including padding and metadata.
+} mi_heap_area_t;
+
+typedef bool (mi_cdecl mi_block_visit_fun)(const mi_heap_t* heap, const mi_heap_area_t* area, void* block, size_t block_size, void* arg);
+
+mi_decl_export bool mi_heap_visit_blocks(const mi_heap_t* heap, bool visit_all_blocks, mi_block_visit_fun* visitor, void* arg);
+
+// Experimental
+mi_decl_nodiscard mi_decl_export bool mi_is_in_heap_region(const void* p) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export bool mi_is_redirected(void) mi_attr_noexcept;
+
+mi_decl_export int mi_reserve_huge_os_pages_interleave(size_t pages, size_t numa_nodes, size_t timeout_msecs) mi_attr_noexcept;
+mi_decl_export int mi_reserve_huge_os_pages_at(size_t pages, int numa_node, size_t timeout_msecs) mi_attr_noexcept;
+
+mi_decl_export int  mi_reserve_os_memory(size_t size, bool commit, bool allow_large) mi_attr_noexcept;
+mi_decl_export bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node) mi_attr_noexcept;
+
+mi_decl_export void mi_debug_show_arenas(void) mi_attr_noexcept;
+
+// Experimental: heaps associated with specific memory arena's
+typedef int mi_arena_id_t;
+mi_decl_export void* mi_arena_area(mi_arena_id_t arena_id, size_t* size);
+mi_decl_export int   mi_reserve_huge_os_pages_at_ex(size_t pages, int numa_node, size_t timeout_msecs, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept;
+mi_decl_export int   mi_reserve_os_memory_ex(size_t size, bool commit, bool allow_large, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept;
+mi_decl_export bool  mi_manage_os_memory_ex(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept;
+
+#if MI_MALLOC_VERSION >= 182
+// Create a heap that only allocates in the specified arena
+mi_decl_nodiscard mi_decl_export mi_heap_t* mi_heap_new_in_arena(mi_arena_id_t arena_id);
+#endif
+
+// deprecated
+mi_decl_export int  mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept;
+
+
+// ------------------------------------------------------
+// Convenience
+// ------------------------------------------------------
+
+#define mi_malloc_tp(tp)                ((tp*)mi_malloc(sizeof(tp)))
+#define mi_zalloc_tp(tp)                ((tp*)mi_zalloc(sizeof(tp)))
+#define mi_calloc_tp(tp,n)              ((tp*)mi_calloc(n,sizeof(tp)))
+#define mi_mallocn_tp(tp,n)             ((tp*)mi_mallocn(n,sizeof(tp)))
+#define mi_reallocn_tp(p,tp,n)          ((tp*)mi_reallocn(p,n,sizeof(tp)))
+#define mi_recalloc_tp(p,tp,n)          ((tp*)mi_recalloc(p,n,sizeof(tp)))
+
+#define mi_heap_malloc_tp(hp,tp)        ((tp*)mi_heap_malloc(hp,sizeof(tp)))
+#define mi_heap_zalloc_tp(hp,tp)        ((tp*)mi_heap_zalloc(hp,sizeof(tp)))
+#define mi_heap_calloc_tp(hp,tp,n)      ((tp*)mi_heap_calloc(hp,n,sizeof(tp)))
+#define mi_heap_mallocn_tp(hp,tp,n)     ((tp*)mi_heap_mallocn(hp,n,sizeof(tp)))
+#define mi_heap_reallocn_tp(hp,p,tp,n)  ((tp*)mi_heap_reallocn(hp,p,n,sizeof(tp)))
+#define mi_heap_recalloc_tp(hp,p,tp,n)  ((tp*)mi_heap_recalloc(hp,p,n,sizeof(tp)))
+
+
+// ------------------------------------------------------
+// Options
+// ------------------------------------------------------
+
+typedef enum mi_option_e {
+  // stable options
+  mi_option_show_errors,              // print error messages
+  mi_option_show_stats,               // print statistics on termination
+  mi_option_verbose,                  // print verbose messages
+  // the following options are experimental (see src/options.h)
+  mi_option_eager_commit,             // eager commit segments? (after `eager_commit_delay` segments) (=1)
+  mi_option_arena_eager_commit,       // eager commit arenas? Use 2 to enable just on overcommit systems (=2)
+  mi_option_purge_decommits,          // should a memory purge decommit (or only reset) (=1)
+  mi_option_allow_large_os_pages,     // allow large (2MiB) OS pages, implies eager commit
+  mi_option_reserve_huge_os_pages,    // reserve N huge OS pages (1GiB/page) at startup
+  mi_option_reserve_huge_os_pages_at, // reserve huge OS pages at a specific NUMA node
+  mi_option_reserve_os_memory,        // reserve specified amount of OS memory in an arena at startup
+  mi_option_deprecated_segment_cache,
+  mi_option_deprecated_page_reset,
+  mi_option_abandoned_page_purge,     // immediately purge delayed purges on thread termination
+  mi_option_deprecated_segment_reset,
+  mi_option_eager_commit_delay,
+  mi_option_purge_delay,              // memory purging is delayed by N milli seconds; use 0 for immediate purging or -1 for no purging at all.
+  mi_option_use_numa_nodes,           // 0 = use all available numa nodes, otherwise use at most N nodes.
+  mi_option_limit_os_alloc,           // 1 = do not use OS memory for allocation (but only programmatically reserved arenas)
+  mi_option_os_tag,                   // tag used for OS logging (macOS only for now)
+  mi_option_max_errors,               // issue at most N error messages
+  mi_option_max_warnings,             // issue at most N warning messages
+  mi_option_max_segment_reclaim,
+  mi_option_destroy_on_exit,          // if set, release all memory on exit; sometimes used for dynamic unloading but can be unsafe.
+  mi_option_arena_reserve,            // initial memory size in KiB for arena reservation (1GiB on 64-bit)
+  mi_option_arena_purge_mult,
+  mi_option_purge_extend_delay,
+  _mi_option_last,
+  // legacy option names
+  mi_option_large_os_pages = mi_option_allow_large_os_pages,
+  mi_option_eager_region_commit = mi_option_arena_eager_commit,
+  mi_option_reset_decommits = mi_option_purge_decommits,
+  mi_option_reset_delay = mi_option_purge_delay,
+  mi_option_abandoned_page_reset = mi_option_abandoned_page_purge
+} mi_option_t;
+
+
+mi_decl_nodiscard mi_decl_export bool mi_option_is_enabled(mi_option_t option);
+mi_decl_export void mi_option_enable(mi_option_t option);
+mi_decl_export void mi_option_disable(mi_option_t option);
+mi_decl_export void mi_option_set_enabled(mi_option_t option, bool enable);
+mi_decl_export void mi_option_set_enabled_default(mi_option_t option, bool enable);
+
+mi_decl_nodiscard mi_decl_export long   mi_option_get(mi_option_t option);
+mi_decl_nodiscard mi_decl_export long   mi_option_get_clamp(mi_option_t option, long min, long max);
+mi_decl_nodiscard mi_decl_export size_t mi_option_get_size(mi_option_t option);
+mi_decl_export void mi_option_set(mi_option_t option, long value);
+mi_decl_export void mi_option_set_default(mi_option_t option, long value);
+
+
+// -------------------------------------------------------------------------------------------------------
+// "mi" prefixed implementations of various posix, Unix, Windows, and C++ allocation functions.
+// (This can be convenient when providing overrides of these functions as done in `mimalloc-override.h`.)
+// note: we use `mi_cfree` as "checked free" and it checks if the pointer is in our heap before free-ing.
+// -------------------------------------------------------------------------------------------------------
+
+mi_decl_export void  mi_cfree(void* p) mi_attr_noexcept;
+mi_decl_export void* mi__expand(void* p, size_t newsize) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export size_t mi_malloc_size(const void* p)        mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export size_t mi_malloc_good_size(size_t size)     mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export size_t mi_malloc_usable_size(const void *p) mi_attr_noexcept;
+
+mi_decl_export int mi_posix_memalign(void** p, size_t alignment, size_t size)   mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_memalign(size_t alignment, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2) mi_attr_alloc_align(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_valloc(size_t size)  mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_pvalloc(size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_aligned_alloc(size_t alignment, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2) mi_attr_alloc_align(1);
+
+mi_decl_nodiscard mi_decl_export void* mi_reallocarray(void* p, size_t count, size_t size) mi_attr_noexcept mi_attr_alloc_size2(2,3);
+mi_decl_nodiscard mi_decl_export int   mi_reallocarr(void* p, size_t count, size_t size) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export void* mi_aligned_recalloc(void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export void* mi_aligned_offset_recalloc(void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept;
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict unsigned short* mi_wcsdup(const unsigned short* s) mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict unsigned char*  mi_mbsdup(const unsigned char* s)  mi_attr_noexcept mi_attr_malloc;
+mi_decl_export int mi_dupenv_s(char** buf, size_t* size, const char* name)                      mi_attr_noexcept;
+mi_decl_export int mi_wdupenv_s(unsigned short** buf, size_t* size, const unsigned short* name) mi_attr_noexcept;
+
+mi_decl_export void mi_free_size(void* p, size_t size)                           mi_attr_noexcept;
+mi_decl_export void mi_free_size_aligned(void* p, size_t size, size_t alignment) mi_attr_noexcept;
+mi_decl_export void mi_free_aligned(void* p, size_t alignment)                   mi_attr_noexcept;
+
+// The `mi_new` wrappers implement C++ semantics on out-of-memory instead of directly returning `NULL`.
+// (and call `std::get_new_handler` and potentially raise a `std::bad_alloc` exception).
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new(size_t size)                   mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new_aligned(size_t size, size_t alignment) mi_attr_malloc mi_attr_alloc_size(1) mi_attr_alloc_align(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new_nothrow(size_t size)           mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new_aligned_nothrow(size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1) mi_attr_alloc_align(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new_n(size_t count, size_t size)   mi_attr_malloc mi_attr_alloc_size2(1, 2);
+mi_decl_nodiscard mi_decl_export void* mi_new_realloc(void* p, size_t newsize)                mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export void* mi_new_reallocn(void* p, size_t newcount, size_t size) mi_attr_alloc_size2(2, 3);
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_alloc_new(mi_heap_t* heap, size_t size)                mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_alloc_new_n(mi_heap_t* heap, size_t count, size_t size) mi_attr_malloc mi_attr_alloc_size2(2, 3);
+
+#ifdef __cplusplus
+}
+#endif
+
+// ---------------------------------------------------------------------------------------------
+// Implement the C++ std::allocator interface for use in STL containers.
+// (note: see `mimalloc-new-delete.h` for overriding the new/delete operators globally)
+// ---------------------------------------------------------------------------------------------
+#ifdef __cplusplus
+
+#include <cstddef>     // std::size_t
+#include <cstdint>     // PTRDIFF_MAX
+#if (__cplusplus >= 201103L) || (_MSC_VER > 1900)  // C++11
+#include <type_traits> // std::true_type
+#include <utility>     // std::forward
+#endif
+
+template<class T> struct _mi_stl_allocator_common {
+  typedef T                 value_type;
+  typedef std::size_t       size_type;
+  typedef std::ptrdiff_t    difference_type;
+  typedef value_type&       reference;
+  typedef value_type const& const_reference;
+  typedef value_type*       pointer;
+  typedef value_type const* const_pointer;
+
+  #if ((__cplusplus >= 201103L) || (_MSC_VER > 1900))  // C++11
+  using propagate_on_container_copy_assignment = std::true_type;
+  using propagate_on_container_move_assignment = std::true_type;
+  using propagate_on_container_swap            = std::true_type;
+  template <class U, class ...Args> void construct(U* p, Args&& ...args) { ::new(p) U(std::forward<Args>(args)...); }
+  template <class U> void destroy(U* p) mi_attr_noexcept { p->~U(); }
+  #else
+  void construct(pointer p, value_type const& val) { ::new(p) value_type(val); }
+  void destroy(pointer p) { p->~value_type(); }
+  #endif
+
+  size_type     max_size() const mi_attr_noexcept { return (PTRDIFF_MAX/sizeof(value_type)); }
+  pointer       address(reference x) const        { return &x; }
+  const_pointer address(const_reference x) const  { return &x; }
+};
+
+template<class T> struct mi_stl_allocator : public _mi_stl_allocator_common<T> {
+  using typename _mi_stl_allocator_common<T>::size_type;
+  using typename _mi_stl_allocator_common<T>::value_type;
+  using typename _mi_stl_allocator_common<T>::pointer;
+  template <class U> struct rebind { typedef mi_stl_allocator<U> other; };
+
+  mi_stl_allocator()                                             mi_attr_noexcept = default;
+  mi_stl_allocator(const mi_stl_allocator&)                      mi_attr_noexcept = default;
+  template<class U> mi_stl_allocator(const mi_stl_allocator<U>&) mi_attr_noexcept { }
+  mi_stl_allocator  select_on_container_copy_construction() const { return *this; }
+  void              deallocate(T* p, size_type) { mi_free(p); }
+
+  #if (__cplusplus >= 201703L)  // C++17
+  mi_decl_nodiscard T* allocate(size_type count) { return static_cast<T*>(mi_new_n(count, sizeof(T))); }
+  mi_decl_nodiscard T* allocate(size_type count, const void*) { return allocate(count); }
+  #else
+  mi_decl_nodiscard pointer allocate(size_type count, const void* = 0) { return static_cast<pointer>(mi_new_n(count, sizeof(value_type))); }
+  #endif
+
+  #if ((__cplusplus >= 201103L) || (_MSC_VER > 1900))  // C++11
+  using is_always_equal = std::true_type;
+  #endif
+};
+
+template<class T1,class T2> bool operator==(const mi_stl_allocator<T1>& , const mi_stl_allocator<T2>& ) mi_attr_noexcept { return true; }
+template<class T1,class T2> bool operator!=(const mi_stl_allocator<T1>& , const mi_stl_allocator<T2>& ) mi_attr_noexcept { return false; }
+
+
+#if (__cplusplus >= 201103L) || (_MSC_VER >= 1900)  // C++11
+#define MI_HAS_HEAP_STL_ALLOCATOR 1
+
+#include <memory>      // std::shared_ptr
+
+// Common base class for STL allocators in a specific heap
+template<class T, bool _mi_destroy> struct _mi_heap_stl_allocator_common : public _mi_stl_allocator_common<T> {
+  using typename _mi_stl_allocator_common<T>::size_type;
+  using typename _mi_stl_allocator_common<T>::value_type;
+  using typename _mi_stl_allocator_common<T>::pointer;
+
+  _mi_heap_stl_allocator_common(mi_heap_t* hp) : heap(hp) { }    /* will not delete nor destroy the passed in heap */
+
+  #if (__cplusplus >= 201703L)  // C++17
+  mi_decl_nodiscard T* allocate(size_type count) { return static_cast<T*>(mi_heap_alloc_new_n(this->heap.get(), count, sizeof(T))); }
+  mi_decl_nodiscard T* allocate(size_type count, const void*) { return allocate(count); }
+  #else
+  mi_decl_nodiscard pointer allocate(size_type count, const void* = 0) { return static_cast<pointer>(mi_heap_alloc_new_n(this->heap.get(), count, sizeof(value_type))); }
+  #endif
+
+  #if ((__cplusplus >= 201103L) || (_MSC_VER > 1900))  // C++11
+  using is_always_equal = std::false_type;
+  #endif
+
+  void collect(bool force) { mi_heap_collect(this->heap.get(), force); }
+  template<class U> bool is_equal(const _mi_heap_stl_allocator_common<U, _mi_destroy>& x) const { return (this->heap == x.heap); }
+
+protected:
+  std::shared_ptr<mi_heap_t> heap;
+  template<class U, bool D> friend struct _mi_heap_stl_allocator_common;
+
+  _mi_heap_stl_allocator_common() {
+    mi_heap_t* hp = mi_heap_new();
+    this->heap.reset(hp, (_mi_destroy ? &heap_destroy : &heap_delete));  /* calls heap_delete/destroy when the refcount drops to zero */
+  }
+  _mi_heap_stl_allocator_common(const _mi_heap_stl_allocator_common& x) mi_attr_noexcept : heap(x.heap) { }
+  template<class U> _mi_heap_stl_allocator_common(const _mi_heap_stl_allocator_common<U, _mi_destroy>& x) mi_attr_noexcept : heap(x.heap) { }
+
+private:
+  static void heap_delete(mi_heap_t* hp)  { if (hp != NULL) { mi_heap_delete(hp); } }
+  static void heap_destroy(mi_heap_t* hp) { if (hp != NULL) { mi_heap_destroy(hp); } }
+};
+
+// STL allocator allocation in a specific heap
+template<class T> struct mi_heap_stl_allocator : public _mi_heap_stl_allocator_common<T, false> {
+  using typename _mi_heap_stl_allocator_common<T, false>::size_type;
+  mi_heap_stl_allocator() : _mi_heap_stl_allocator_common<T, false>() { } // creates fresh heap that is deleted when the destructor is called
+  mi_heap_stl_allocator(mi_heap_t* hp) : _mi_heap_stl_allocator_common<T, false>(hp) { }  // no delete nor destroy on the passed in heap
+  template<class U> mi_heap_stl_allocator(const mi_heap_stl_allocator<U>& x) mi_attr_noexcept : _mi_heap_stl_allocator_common<T, false>(x) { }
+
+  mi_heap_stl_allocator select_on_container_copy_construction() const { return *this; }
+  void deallocate(T* p, size_type) { mi_free(p); }
+  template<class U> struct rebind { typedef mi_heap_stl_allocator<U> other; };
+};
+
+template<class T1, class T2> bool operator==(const mi_heap_stl_allocator<T1>& x, const mi_heap_stl_allocator<T2>& y) mi_attr_noexcept { return (x.is_equal(y)); }
+template<class T1, class T2> bool operator!=(const mi_heap_stl_allocator<T1>& x, const mi_heap_stl_allocator<T2>& y) mi_attr_noexcept { return (!x.is_equal(y)); }
+
+
+// STL allocator allocation in a specific heap, where `free` does nothing and
+// the heap is destroyed in one go on destruction -- use with care!
+template<class T> struct mi_heap_destroy_stl_allocator : public _mi_heap_stl_allocator_common<T, true> {
+  using typename _mi_heap_stl_allocator_common<T, true>::size_type;
+  mi_heap_destroy_stl_allocator() : _mi_heap_stl_allocator_common<T, true>() { } // creates fresh heap that is destroyed when the destructor is called
+  mi_heap_destroy_stl_allocator(mi_heap_t* hp) : _mi_heap_stl_allocator_common<T, true>(hp) { }  // no delete nor destroy on the passed in heap
+  template<class U> mi_heap_destroy_stl_allocator(const mi_heap_destroy_stl_allocator<U>& x) mi_attr_noexcept : _mi_heap_stl_allocator_common<T, true>(x) { }
+
+  mi_heap_destroy_stl_allocator select_on_container_copy_construction() const { return *this; }
+  void deallocate(T*, size_type) { /* do nothing as we destroy the heap on destruct. */ }
+  template<class U> struct rebind { typedef mi_heap_destroy_stl_allocator<U> other; };
+};
+
+template<class T1, class T2> bool operator==(const mi_heap_destroy_stl_allocator<T1>& x, const mi_heap_destroy_stl_allocator<T2>& y) mi_attr_noexcept { return (x.is_equal(y)); }
+template<class T1, class T2> bool operator!=(const mi_heap_destroy_stl_allocator<T1>& x, const mi_heap_destroy_stl_allocator<T2>& y) mi_attr_noexcept { return (!x.is_equal(y)); }
+
+#endif // C++11
+
+#endif // __cplusplus
+
+#endif
diff --git a/compat/mimalloc/mimalloc/atomic.h b/compat/mimalloc/mimalloc/atomic.h
new file mode 100644
index 00000000000000..c6b8146ffdb049
--- /dev/null
+++ b/compat/mimalloc/mimalloc/atomic.h
@@ -0,0 +1,385 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023 Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_ATOMIC_H
+#define MIMALLOC_ATOMIC_H
+
+// --------------------------------------------------------------------------------------------
+// Atomics
+// We need to be portable between C, C++, and MSVC.
+// We base the primitives on the C/C++ atomics and create a mimimal wrapper for MSVC in C compilation mode.
+// This is why we try to use only `uintptr_t` and `<type>*` as atomic types.
+// To gain better insight in the range of used atomics, we use explicitly named memory order operations
+// instead of passing the memory order as a parameter.
+// -----------------------------------------------------------------------------------------------
+
+#if defined(__cplusplus)
+// Use C++ atomics
+#include <atomic>
+#define  _Atomic(tp)            std::atomic<tp>
+#define  mi_atomic(name)        std::atomic_##name
+#define  mi_memory_order(name)  std::memory_order_##name
+#if !defined(ATOMIC_VAR_INIT) || (__cplusplus >= 202002L) // c++20, see issue #571
+ #define MI_ATOMIC_VAR_INIT(x)  x
+#else
+ #define MI_ATOMIC_VAR_INIT(x)  ATOMIC_VAR_INIT(x)
+#endif
+#elif defined(_MSC_VER)
+// Use MSVC C wrapper for C11 atomics
+#define  _Atomic(tp)            tp
+#define  MI_ATOMIC_VAR_INIT(x)  x
+#define  mi_atomic(name)        mi_atomic_##name
+#define  mi_memory_order(name)  mi_memory_order_##name
+#else
+// Use C11 atomics
+#include <stdatomic.h>
+#define  mi_atomic(name)        atomic_##name
+#define  mi_memory_order(name)  memory_order_##name
+#if !defined(ATOMIC_VAR_INIT) || (__STDC_VERSION__ >= 201710L) // c17, see issue #735
+ #define MI_ATOMIC_VAR_INIT(x) x
+#else
+ #define MI_ATOMIC_VAR_INIT(x) ATOMIC_VAR_INIT(x)
+#endif
+#endif
+
+// Various defines for all used memory orders in mimalloc
+#define mi_atomic_cas_weak(p,expected,desired,mem_success,mem_fail)  \
+  mi_atomic(compare_exchange_weak_explicit)(p,expected,desired,mem_success,mem_fail)
+
+#define mi_atomic_cas_strong(p,expected,desired,mem_success,mem_fail)  \
+  mi_atomic(compare_exchange_strong_explicit)(p,expected,desired,mem_success,mem_fail)
+
+#define mi_atomic_load_acquire(p)                mi_atomic(load_explicit)(p,mi_memory_order(acquire))
+#define mi_atomic_load_relaxed(p)                mi_atomic(load_explicit)(p,mi_memory_order(relaxed))
+#define mi_atomic_store_release(p,x)             mi_atomic(store_explicit)(p,x,mi_memory_order(release))
+#define mi_atomic_store_relaxed(p,x)             mi_atomic(store_explicit)(p,x,mi_memory_order(relaxed))
+#define mi_atomic_exchange_release(p,x)          mi_atomic(exchange_explicit)(p,x,mi_memory_order(release))
+#define mi_atomic_exchange_acq_rel(p,x)          mi_atomic(exchange_explicit)(p,x,mi_memory_order(acq_rel))
+#define mi_atomic_cas_weak_release(p,exp,des)    mi_atomic_cas_weak(p,exp,des,mi_memory_order(release),mi_memory_order(relaxed))
+#define mi_atomic_cas_weak_acq_rel(p,exp,des)    mi_atomic_cas_weak(p,exp,des,mi_memory_order(acq_rel),mi_memory_order(acquire))
+#define mi_atomic_cas_strong_release(p,exp,des)  mi_atomic_cas_strong(p,exp,des,mi_memory_order(release),mi_memory_order(relaxed))
+#define mi_atomic_cas_strong_acq_rel(p,exp,des)  mi_atomic_cas_strong(p,exp,des,mi_memory_order(acq_rel),mi_memory_order(acquire))
+
+#define mi_atomic_add_relaxed(p,x)               mi_atomic(fetch_add_explicit)(p,x,mi_memory_order(relaxed))
+#define mi_atomic_sub_relaxed(p,x)               mi_atomic(fetch_sub_explicit)(p,x,mi_memory_order(relaxed))
+#define mi_atomic_add_acq_rel(p,x)               mi_atomic(fetch_add_explicit)(p,x,mi_memory_order(acq_rel))
+#define mi_atomic_sub_acq_rel(p,x)               mi_atomic(fetch_sub_explicit)(p,x,mi_memory_order(acq_rel))
+#define mi_atomic_and_acq_rel(p,x)               mi_atomic(fetch_and_explicit)(p,x,mi_memory_order(acq_rel))
+#define mi_atomic_or_acq_rel(p,x)                mi_atomic(fetch_or_explicit)(p,x,mi_memory_order(acq_rel))
+
+#define mi_atomic_increment_relaxed(p)           mi_atomic_add_relaxed(p,(uintptr_t)1)
+#define mi_atomic_decrement_relaxed(p)           mi_atomic_sub_relaxed(p,(uintptr_t)1)
+#define mi_atomic_increment_acq_rel(p)           mi_atomic_add_acq_rel(p,(uintptr_t)1)
+#define mi_atomic_decrement_acq_rel(p)           mi_atomic_sub_acq_rel(p,(uintptr_t)1)
+
+static inline void mi_atomic_yield(void);
+static inline intptr_t mi_atomic_addi(_Atomic(intptr_t)*p, intptr_t add);
+static inline intptr_t mi_atomic_subi(_Atomic(intptr_t)*p, intptr_t sub);
+
+
+#if defined(__cplusplus) || !defined(_MSC_VER)
+
+// In C++/C11 atomics we have polymorphic atomics so can use the typed `ptr` variants (where `tp` is the type of atomic value)
+// We use these macros so we can provide a typed wrapper in MSVC in C compilation mode as well
+#define mi_atomic_load_ptr_acquire(tp,p)                mi_atomic_load_acquire(p)
+#define mi_atomic_load_ptr_relaxed(tp,p)                mi_atomic_load_relaxed(p)
+
+// In C++ we need to add casts to help resolve templates if NULL is passed
+#if defined(__cplusplus)
+#define mi_atomic_store_ptr_release(tp,p,x)             mi_atomic_store_release(p,(tp*)x)
+#define mi_atomic_store_ptr_relaxed(tp,p,x)             mi_atomic_store_relaxed(p,(tp*)x)
+#define mi_atomic_cas_ptr_weak_release(tp,p,exp,des)    mi_atomic_cas_weak_release(p,exp,(tp*)des)
+#define mi_atomic_cas_ptr_weak_acq_rel(tp,p,exp,des)    mi_atomic_cas_weak_acq_rel(p,exp,(tp*)des)
+#define mi_atomic_cas_ptr_strong_release(tp,p,exp,des)  mi_atomic_cas_strong_release(p,exp,(tp*)des)
+#define mi_atomic_exchange_ptr_release(tp,p,x)          mi_atomic_exchange_release(p,(tp*)x)
+#define mi_atomic_exchange_ptr_acq_rel(tp,p,x)          mi_atomic_exchange_acq_rel(p,(tp*)x)
+#else
+#define mi_atomic_store_ptr_release(tp,p,x)             mi_atomic_store_release(p,x)
+#define mi_atomic_store_ptr_relaxed(tp,p,x)             mi_atomic_store_relaxed(p,x)
+#define mi_atomic_cas_ptr_weak_release(tp,p,exp,des)    mi_atomic_cas_weak_release(p,exp,des)
+#define mi_atomic_cas_ptr_weak_acq_rel(tp,p,exp,des)    mi_atomic_cas_weak_acq_rel(p,exp,des)
+#define mi_atomic_cas_ptr_strong_release(tp,p,exp,des)  mi_atomic_cas_strong_release(p,exp,des)
+#define mi_atomic_exchange_ptr_release(tp,p,x)          mi_atomic_exchange_release(p,x)
+#define mi_atomic_exchange_ptr_acq_rel(tp,p,x)          mi_atomic_exchange_acq_rel(p,x)
+#endif
+
+// These are used by the statistics
+static inline int64_t mi_atomic_addi64_relaxed(volatile int64_t* p, int64_t add) {
+  return mi_atomic(fetch_add_explicit)((_Atomic(int64_t)*)p, add, mi_memory_order(relaxed));
+}
+static inline void mi_atomic_maxi64_relaxed(volatile int64_t* p, int64_t x) {
+  int64_t current = mi_atomic_load_relaxed((_Atomic(int64_t)*)p);
+  while (current < x && !mi_atomic_cas_weak_release((_Atomic(int64_t)*)p, &current, x)) { /* nothing */ };
+}
+
+// Used by timers
+#define mi_atomic_loadi64_acquire(p)            mi_atomic(load_explicit)(p,mi_memory_order(acquire))
+#define mi_atomic_loadi64_relaxed(p)            mi_atomic(load_explicit)(p,mi_memory_order(relaxed))
+#define mi_atomic_storei64_release(p,x)         mi_atomic(store_explicit)(p,x,mi_memory_order(release))
+#define mi_atomic_storei64_relaxed(p,x)         mi_atomic(store_explicit)(p,x,mi_memory_order(relaxed))
+
+#define mi_atomic_casi64_strong_acq_rel(p,e,d)  mi_atomic_cas_strong_acq_rel(p,e,d)
+#define mi_atomic_addi64_acq_rel(p,i)           mi_atomic_add_acq_rel(p,i)
+
+
+#elif defined(_MSC_VER)
+
+// MSVC C compilation wrapper that uses Interlocked operations to model C11 atomics.
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include <intrin.h>
+#ifdef _WIN64
+typedef LONG64   msc_intptr_t;
+#define MI_64(f) f##64
+#else
+typedef LONG     msc_intptr_t;
+#define MI_64(f) f
+#endif
+
+typedef enum mi_memory_order_e {
+  mi_memory_order_relaxed,
+  mi_memory_order_consume,
+  mi_memory_order_acquire,
+  mi_memory_order_release,
+  mi_memory_order_acq_rel,
+  mi_memory_order_seq_cst
+} mi_memory_order;
+
+static inline uintptr_t mi_atomic_fetch_add_explicit(_Atomic(uintptr_t)*p, uintptr_t add, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedExchangeAdd)((volatile msc_intptr_t*)p, (msc_intptr_t)add);
+}
+static inline uintptr_t mi_atomic_fetch_sub_explicit(_Atomic(uintptr_t)*p, uintptr_t sub, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedExchangeAdd)((volatile msc_intptr_t*)p, -((msc_intptr_t)sub));
+}
+static inline uintptr_t mi_atomic_fetch_and_explicit(_Atomic(uintptr_t)*p, uintptr_t x, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedAnd)((volatile msc_intptr_t*)p, (msc_intptr_t)x);
+}
+static inline uintptr_t mi_atomic_fetch_or_explicit(_Atomic(uintptr_t)*p, uintptr_t x, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedOr)((volatile msc_intptr_t*)p, (msc_intptr_t)x);
+}
+static inline bool mi_atomic_compare_exchange_strong_explicit(_Atomic(uintptr_t)*p, uintptr_t* expected, uintptr_t desired, mi_memory_order mo1, mi_memory_order mo2) {
+  (void)(mo1); (void)(mo2);
+  uintptr_t read = (uintptr_t)MI_64(_InterlockedCompareExchange)((volatile msc_intptr_t*)p, (msc_intptr_t)desired, (msc_intptr_t)(*expected));
+  if (read == *expected) {
+    return true;
+  }
+  else {
+    *expected = read;
+    return false;
+  }
+}
+static inline bool mi_atomic_compare_exchange_weak_explicit(_Atomic(uintptr_t)*p, uintptr_t* expected, uintptr_t desired, mi_memory_order mo1, mi_memory_order mo2) {
+  return mi_atomic_compare_exchange_strong_explicit(p, expected, desired, mo1, mo2);
+}
+static inline uintptr_t mi_atomic_exchange_explicit(_Atomic(uintptr_t)*p, uintptr_t exchange, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedExchange)((volatile msc_intptr_t*)p, (msc_intptr_t)exchange);
+}
+static inline void mi_atomic_thread_fence(mi_memory_order mo) {
+  (void)(mo);
+  _Atomic(uintptr_t) x = 0;
+  mi_atomic_exchange_explicit(&x, 1, mo);
+}
+static inline uintptr_t mi_atomic_load_explicit(_Atomic(uintptr_t) const* p, mi_memory_order mo) {
+  (void)(mo);
+#if defined(_M_IX86) || defined(_M_X64)
+  return *p;
+#else
+  uintptr_t x = *p;
+  if (mo > mi_memory_order_relaxed) {
+    while (!mi_atomic_compare_exchange_weak_explicit(p, &x, x, mo, mi_memory_order_relaxed)) { /* nothing */ };
+  }
+  return x;
+#endif
+}
+static inline void mi_atomic_store_explicit(_Atomic(uintptr_t)*p, uintptr_t x, mi_memory_order mo) {
+  (void)(mo);
+#if defined(_M_IX86) || defined(_M_X64)
+  *p = x;
+#else
+  mi_atomic_exchange_explicit(p, x, mo);
+#endif
+}
+static inline int64_t mi_atomic_loadi64_explicit(_Atomic(int64_t)*p, mi_memory_order mo) {
+  (void)(mo);
+#if defined(_M_X64)
+  return *p;
+#else
+  int64_t old = *p;
+  int64_t x = old;
+  while ((old = InterlockedCompareExchange64(p, x, old)) != x) {
+    x = old;
+  }
+  return x;
+#endif
+}
+static inline void mi_atomic_storei64_explicit(_Atomic(int64_t)*p, int64_t x, mi_memory_order mo) {
+  (void)(mo);
+#if defined(x_M_IX86) || defined(_M_X64)
+  *p = x;
+#else
+  InterlockedExchange64(p, x);
+#endif
+}
+
+// These are used by the statistics
+static inline int64_t mi_atomic_addi64_relaxed(volatile _Atomic(int64_t)*p, int64_t add) {
+#ifdef _WIN64
+  return (int64_t)mi_atomic_addi((int64_t*)p, add);
+#else
+  int64_t current;
+  int64_t sum;
+  do {
+    current = *p;
+    sum = current + add;
+  } while (_InterlockedCompareExchange64(p, sum, current) != current);
+  return current;
+#endif
+}
+static inline void mi_atomic_maxi64_relaxed(volatile _Atomic(int64_t)*p, int64_t x) {
+  int64_t current;
+  do {
+    current = *p;
+  } while (current < x && _InterlockedCompareExchange64(p, x, current) != current);
+}
+
+static inline void mi_atomic_addi64_acq_rel(volatile _Atomic(int64_t*)p, int64_t i) {
+  mi_atomic_addi64_relaxed(p, i);
+}
+
+static inline bool mi_atomic_casi64_strong_acq_rel(volatile _Atomic(int64_t*)p, int64_t* exp, int64_t des) {
+  int64_t read = _InterlockedCompareExchange64(p, des, *exp);
+  if (read == *exp) {
+    return true;
+  }
+  else {
+    *exp = read;
+    return false;
+  }
+}
+
+// The pointer macros cast to `uintptr_t`.
+#define mi_atomic_load_ptr_acquire(tp,p)                (tp*)mi_atomic_load_acquire((_Atomic(uintptr_t)*)(p))
+#define mi_atomic_load_ptr_relaxed(tp,p)                (tp*)mi_atomic_load_relaxed((_Atomic(uintptr_t)*)(p))
+#define mi_atomic_store_ptr_release(tp,p,x)             mi_atomic_store_release((_Atomic(uintptr_t)*)(p),(uintptr_t)(x))
+#define mi_atomic_store_ptr_relaxed(tp,p,x)             mi_atomic_store_relaxed((_Atomic(uintptr_t)*)(p),(uintptr_t)(x))
+#define mi_atomic_cas_ptr_weak_release(tp,p,exp,des)    mi_atomic_cas_weak_release((_Atomic(uintptr_t)*)(p),(uintptr_t*)exp,(uintptr_t)des)
+#define mi_atomic_cas_ptr_weak_acq_rel(tp,p,exp,des)    mi_atomic_cas_weak_acq_rel((_Atomic(uintptr_t)*)(p),(uintptr_t*)exp,(uintptr_t)des)
+#define mi_atomic_cas_ptr_strong_release(tp,p,exp,des)  mi_atomic_cas_strong_release((_Atomic(uintptr_t)*)(p),(uintptr_t*)exp,(uintptr_t)des)
+#define mi_atomic_exchange_ptr_release(tp,p,x)          (tp*)mi_atomic_exchange_release((_Atomic(uintptr_t)*)(p),(uintptr_t)x)
+#define mi_atomic_exchange_ptr_acq_rel(tp,p,x)          (tp*)mi_atomic_exchange_acq_rel((_Atomic(uintptr_t)*)(p),(uintptr_t)x)
+
+#define mi_atomic_loadi64_acquire(p)    mi_atomic(loadi64_explicit)(p,mi_memory_order(acquire))
+#define mi_atomic_loadi64_relaxed(p)    mi_atomic(loadi64_explicit)(p,mi_memory_order(relaxed))
+#define mi_atomic_storei64_release(p,x) mi_atomic(storei64_explicit)(p,x,mi_memory_order(release))
+#define mi_atomic_storei64_relaxed(p,x) mi_atomic(storei64_explicit)(p,x,mi_memory_order(relaxed))
+
+
+#endif
+
+
+// Atomically add a signed value; returns the previous value.
+static inline intptr_t mi_atomic_addi(_Atomic(intptr_t)*p, intptr_t add) {
+  return (intptr_t)mi_atomic_add_acq_rel((_Atomic(uintptr_t)*)p, (uintptr_t)add);
+}
+
+// Atomically subtract a signed value; returns the previous value.
+static inline intptr_t mi_atomic_subi(_Atomic(intptr_t)*p, intptr_t sub) {
+  return (intptr_t)mi_atomic_addi(p, -sub);
+}
+
+typedef _Atomic(uintptr_t) mi_atomic_once_t;
+
+// Returns true only on the first invocation
+static inline bool mi_atomic_once( mi_atomic_once_t* once ) {
+  if (mi_atomic_load_relaxed(once) != 0) return false;     // quick test
+  uintptr_t expected = 0;
+  return mi_atomic_cas_strong_acq_rel(once, &expected, (uintptr_t)1); // try to set to 1
+}
+
+typedef _Atomic(uintptr_t) mi_atomic_guard_t;
+
+// Allows only one thread to execute at a time
+#define mi_atomic_guard(guard) \
+  uintptr_t _mi_guard_expected = 0; \
+  for(bool _mi_guard_once = true; \
+      _mi_guard_once && mi_atomic_cas_strong_acq_rel(guard,&_mi_guard_expected,(uintptr_t)1); \
+      (mi_atomic_store_release(guard,(uintptr_t)0), _mi_guard_once = false) )
+
+
+
+// Yield
+#if defined(__cplusplus)
+#include <thread>
+static inline void mi_atomic_yield(void) {
+  std::this_thread::yield();
+}
+#elif defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+static inline void mi_atomic_yield(void) {
+  YieldProcessor();
+}
+#elif defined(__SSE2__)
+#include <emmintrin.h>
+static inline void mi_atomic_yield(void) {
+  _mm_pause();
+}
+#elif (defined(__GNUC__) || defined(__clang__)) && \
+      (defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__armel__) || defined(__ARMEL__) || \
+       defined(__aarch64__) || defined(__powerpc__) || defined(__ppc__) || defined(__PPC__)) || defined(__POWERPC__)
+#if defined(__x86_64__) || defined(__i386__)
+static inline void mi_atomic_yield(void) {
+  __asm__ volatile ("pause" ::: "memory");
+}
+#elif defined(__aarch64__)
+static inline void mi_atomic_yield(void) {
+  __asm__ volatile("wfe");
+}
+#elif (defined(__arm__) && __ARM_ARCH__ >= 7)
+static inline void mi_atomic_yield(void) {
+  __asm__ volatile("yield" ::: "memory");
+}
+#elif defined(__powerpc__) || defined(__ppc__) || defined(__PPC__) || defined(__POWERPC__)
+#ifdef __APPLE__
+static inline void mi_atomic_yield(void) {
+  __asm__ volatile ("or r27,r27,r27" ::: "memory");
+}
+#else
+static inline void mi_atomic_yield(void) {
+  __asm__ __volatile__ ("or 27,27,27" ::: "memory");
+}
+#endif
+#elif defined(__armel__) || defined(__ARMEL__)
+static inline void mi_atomic_yield(void) {
+  __asm__ volatile ("nop" ::: "memory");
+}
+#endif
+#elif defined(__sun)
+// Fallback for other archs
+#include <synch.h>
+static inline void mi_atomic_yield(void) {
+  smt_pause();
+}
+#elif defined(__wasi__)
+#include <sched.h>
+static inline void mi_atomic_yield(void) {
+  sched_yield();
+}
+#else
+#include <unistd.h>
+static inline void mi_atomic_yield(void) {
+  sleep(0);
+}
+#endif
+
+
+#endif // __MIMALLOC_ATOMIC_H
diff --git a/compat/mimalloc/mimalloc/internal.h b/compat/mimalloc/mimalloc/internal.h
new file mode 100644
index 00000000000000..f076bc6a40f977
--- /dev/null
+++ b/compat/mimalloc/mimalloc/internal.h
@@ -0,0 +1,979 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_INTERNAL_H
+#define MIMALLOC_INTERNAL_H
+
+
+// --------------------------------------------------------------------------
+// This file contains the interal API's of mimalloc and various utility
+// functions and macros.
+// --------------------------------------------------------------------------
+
+#include "mimalloc/types.h"
+#include "mimalloc/track.h"
+
+#if (MI_DEBUG>0)
+#define mi_trace_message(...)  _mi_trace_message(__VA_ARGS__)
+#else
+#define mi_trace_message(...)
+#endif
+
+#define MI_CACHE_LINE          64
+#if defined(_MSC_VER)
+#pragma warning(disable:4127)   // suppress constant conditional warning (due to MI_SECURE paths)
+#pragma warning(disable:26812)  // unscoped enum warning
+#define mi_decl_noinline        __declspec(noinline)
+#define mi_decl_thread          __declspec(thread)
+#define mi_decl_cache_align     __declspec(align(MI_CACHE_LINE))
+#elif (defined(__GNUC__) && (__GNUC__ >= 3)) || defined(__clang__) // includes clang and icc
+#define mi_decl_noinline        __attribute__((noinline))
+#define mi_decl_thread          __thread
+#define mi_decl_cache_align     __attribute__((aligned(MI_CACHE_LINE)))
+#else
+#define mi_decl_noinline
+#define mi_decl_thread          __thread        // hope for the best :-)
+#define mi_decl_cache_align
+#endif
+
+#if defined(__EMSCRIPTEN__) && !defined(__wasi__)
+#define __wasi__
+#endif
+
+#if defined(__cplusplus)
+#define mi_decl_externc       extern "C"
+#else
+#define mi_decl_externc
+#endif
+
+// pthreads
+#if !defined(_WIN32) && !defined(__wasi__)
+#define  MI_USE_PTHREADS
+#include <pthread.h>
+#endif
+
+// "options.c"
+void       _mi_fputs(mi_output_fun* out, void* arg, const char* prefix, const char* message);
+void       _mi_fprintf(mi_output_fun* out, void* arg, const char* fmt, ...);
+void       _mi_warning_message(const char* fmt, ...);
+void       _mi_verbose_message(const char* fmt, ...);
+void       _mi_trace_message(const char* fmt, ...);
+void       _mi_options_init(void);
+void       _mi_error_message(int err, const char* fmt, ...);
+
+// random.c
+void       _mi_random_init(mi_random_ctx_t* ctx);
+void       _mi_random_init_weak(mi_random_ctx_t* ctx);
+void       _mi_random_reinit_if_weak(mi_random_ctx_t * ctx);
+void       _mi_random_split(mi_random_ctx_t* ctx, mi_random_ctx_t* new_ctx);
+uintptr_t  _mi_random_next(mi_random_ctx_t* ctx);
+uintptr_t  _mi_heap_random_next(mi_heap_t* heap);
+uintptr_t  _mi_os_random_weak(uintptr_t extra_seed);
+static inline uintptr_t _mi_random_shuffle(uintptr_t x);
+
+// init.c
+extern mi_decl_cache_align mi_stats_t       _mi_stats_main;
+extern mi_decl_cache_align const mi_page_t  _mi_page_empty;
+bool       _mi_is_main_thread(void);
+size_t     _mi_current_thread_count(void);
+bool       _mi_preloading(void);           // true while the C runtime is not initialized yet
+mi_threadid_t _mi_thread_id(void) mi_attr_noexcept;
+mi_heap_t*    _mi_heap_main_get(void);     // statically allocated main backing heap
+void       _mi_thread_done(mi_heap_t* heap);
+void       _mi_thread_data_collect(void);
+
+// os.c
+void       _mi_os_init(void);                                            // called from process init
+void*      _mi_os_alloc(size_t size, mi_memid_t* memid, mi_stats_t* stats);
+void       _mi_os_free(void* p, size_t size, mi_memid_t memid, mi_stats_t* stats);
+void       _mi_os_free_ex(void* p, size_t size, bool still_committed, mi_memid_t memid, mi_stats_t* stats);
+
+size_t     _mi_os_page_size(void);
+size_t     _mi_os_good_alloc_size(size_t size);
+bool       _mi_os_has_overcommit(void);
+bool       _mi_os_has_virtual_reserve(void);
+
+bool       _mi_os_purge(void* p, size_t size, mi_stats_t* stats);
+bool       _mi_os_reset(void* addr, size_t size, mi_stats_t* tld_stats);
+bool       _mi_os_commit(void* p, size_t size, bool* is_zero, mi_stats_t* stats);
+bool       _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats);
+bool       _mi_os_protect(void* addr, size_t size);
+bool       _mi_os_unprotect(void* addr, size_t size);
+bool       _mi_os_purge(void* p, size_t size, mi_stats_t* stats);
+bool       _mi_os_purge_ex(void* p, size_t size, bool allow_reset, mi_stats_t* stats);
+
+void*      _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* stats);
+void*      _mi_os_alloc_aligned_at_offset(size_t size, size_t alignment, size_t align_offset, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* tld_stats);
+
+void*      _mi_os_get_aligned_hint(size_t try_alignment, size_t size);
+bool       _mi_os_use_large_page(size_t size, size_t alignment);
+size_t     _mi_os_large_page_size(void);
+
+void*      _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, mi_msecs_t max_secs, size_t* pages_reserved, size_t* psize, mi_memid_t* memid);
+
+// arena.c
+mi_arena_id_t _mi_arena_id_none(void);
+void       _mi_arena_free(void* p, size_t size, size_t still_committed_size, mi_memid_t memid, mi_stats_t* stats);
+void*      _mi_arena_alloc(size_t size, bool commit, bool allow_large, mi_arena_id_t req_arena_id, mi_memid_t* memid, mi_os_tld_t* tld);
+void*      _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool commit, bool allow_large, mi_arena_id_t req_arena_id, mi_memid_t* memid, mi_os_tld_t* tld);
+bool       _mi_arena_memid_is_suitable(mi_memid_t memid, mi_arena_id_t request_arena_id);
+bool       _mi_arena_contains(const void* p);
+void       _mi_arena_collect(bool force_purge, mi_stats_t* stats);
+void       _mi_arena_unsafe_destroy_all(mi_stats_t* stats);
+
+// "segment-map.c"
+void       _mi_segment_map_allocated_at(const mi_segment_t* segment);
+void       _mi_segment_map_freed_at(const mi_segment_t* segment);
+
+// "segment.c"
+mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, size_t page_alignment, mi_segments_tld_t* tld, mi_os_tld_t* os_tld);
+void       _mi_segment_page_free(mi_page_t* page, bool force, mi_segments_tld_t* tld);
+void       _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld);
+bool       _mi_segment_try_reclaim_abandoned( mi_heap_t* heap, bool try_all, mi_segments_tld_t* tld);
+void       _mi_segment_thread_collect(mi_segments_tld_t* tld);
+
+#if MI_HUGE_PAGE_ABANDON
+void       _mi_segment_huge_page_free(mi_segment_t* segment, mi_page_t* page, mi_block_t* block);
+#else
+void       _mi_segment_huge_page_reset(mi_segment_t* segment, mi_page_t* page, mi_block_t* block);
+#endif
+
+uint8_t*   _mi_segment_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size); // page start for any page
+void       _mi_abandoned_reclaim_all(mi_heap_t* heap, mi_segments_tld_t* tld);
+void       _mi_abandoned_await_readers(void);
+void       _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t* tld);
+
+// "page.c"
+void*      _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment)  mi_attr_noexcept mi_attr_malloc;
+
+void       _mi_page_retire(mi_page_t* page) mi_attr_noexcept;                  // free the page if there are no other pages with many free blocks
+void       _mi_page_unfull(mi_page_t* page);
+void       _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force);   // free the page
+void       _mi_page_abandon(mi_page_t* page, mi_page_queue_t* pq);            // abandon the page, to be picked up by another thread...
+void       _mi_heap_delayed_free_all(mi_heap_t* heap);
+bool       _mi_heap_delayed_free_partial(mi_heap_t* heap);
+void       _mi_heap_collect_retired(mi_heap_t* heap, bool force);
+
+void       _mi_page_use_delayed_free(mi_page_t* page, mi_delayed_t delay, bool override_never);
+bool       _mi_page_try_use_delayed_free(mi_page_t* page, mi_delayed_t delay, bool override_never);
+size_t     _mi_page_queue_append(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_queue_t* append);
+void       _mi_deferred_free(mi_heap_t* heap, bool force);
+
+void       _mi_page_free_collect(mi_page_t* page,bool force);
+void       _mi_page_reclaim(mi_heap_t* heap, mi_page_t* page);   // callback from segments
+
+size_t     _mi_bin_size(uint8_t bin);           // for stats
+uint8_t    _mi_bin(size_t size);                // for stats
+
+// "heap.c"
+void       _mi_heap_destroy_pages(mi_heap_t* heap);
+void       _mi_heap_collect_abandon(mi_heap_t* heap);
+void       _mi_heap_set_default_direct(mi_heap_t* heap);
+bool       _mi_heap_memid_is_suitable(mi_heap_t* heap, mi_memid_t memid);
+void       _mi_heap_unsafe_destroy_all(void);
+
+// "stats.c"
+void       _mi_stats_done(mi_stats_t* stats);
+mi_msecs_t  _mi_clock_now(void);
+mi_msecs_t  _mi_clock_end(mi_msecs_t start);
+mi_msecs_t  _mi_clock_start(void);
+
+// "alloc.c"
+void*       _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size, bool zero) mi_attr_noexcept;  // called from `_mi_malloc_generic`
+void*       _mi_heap_malloc_zero(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept;
+void*       _mi_heap_malloc_zero_ex(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept;     // called from `_mi_heap_malloc_aligned`
+void*       _mi_heap_realloc_zero(mi_heap_t* heap, void* p, size_t newsize, bool zero) mi_attr_noexcept;
+mi_block_t* _mi_page_ptr_unalign(const mi_segment_t* segment, const mi_page_t* page, const void* p);
+bool        _mi_free_delayed_block(mi_block_t* block);
+void        _mi_free_generic(const mi_segment_t* segment, mi_page_t* page, bool is_local, void* p) mi_attr_noexcept;  // for runtime integration
+void        _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size);
+
+// option.c, c primitives
+char        _mi_toupper(char c);
+int         _mi_strnicmp(const char* s, const char* t, size_t n);
+void        _mi_strlcpy(char* dest, const char* src, size_t dest_size);
+void        _mi_strlcat(char* dest, const char* src, size_t dest_size);
+size_t      _mi_strlen(const char* s);
+size_t      _mi_strnlen(const char* s, size_t max_len);
+
+
+#if MI_DEBUG>1
+bool        _mi_page_is_valid(mi_page_t* page);
+#endif
+
+
+// ------------------------------------------------------
+// Branches
+// ------------------------------------------------------
+
+#if defined(__GNUC__) || defined(__clang__)
+#define mi_unlikely(x)     (__builtin_expect(!!(x),false))
+#define mi_likely(x)       (__builtin_expect(!!(x),true))
+#elif (defined(__cplusplus) && (__cplusplus >= 202002L)) || (defined(_MSVC_LANG) && _MSVC_LANG >= 202002L)
+#define mi_unlikely(x)     (x) [[unlikely]]
+#define mi_likely(x)       (x) [[likely]]
+#else
+#define mi_unlikely(x)     (x)
+#define mi_likely(x)       (x)
+#endif
+
+#ifndef __has_builtin
+#define __has_builtin(x)  0
+#endif
+
+
+/* -----------------------------------------------------------
+  Error codes passed to `_mi_fatal_error`
+  All are recoverable but EFAULT is a serious error and aborts by default in secure mode.
+  For portability define undefined error codes using common Unix codes:
+  <https://www-numi.fnal.gov/offline_software/srt_public_context/WebDocs/Errors/unix_system_errors.html>
+----------------------------------------------------------- */
+#include <errno.h>
+#ifndef EAGAIN         // double free
+#define EAGAIN (11)
+#endif
+#ifndef ENOMEM         // out of memory
+#define ENOMEM (12)
+#endif
+#ifndef EFAULT         // corrupted free-list or meta-data
+#define EFAULT (14)
+#endif
+#ifndef EINVAL         // trying to free an invalid pointer
+#define EINVAL (22)
+#endif
+#ifndef EOVERFLOW      // count*size overflow
+#define EOVERFLOW (75)
+#endif
+
+
+/* -----------------------------------------------------------
+  Inlined definitions
+----------------------------------------------------------- */
+#define MI_UNUSED(x)     (void)(x)
+#if (MI_DEBUG>0)
+#define MI_UNUSED_RELEASE(x)
+#else
+#define MI_UNUSED_RELEASE(x)  MI_UNUSED(x)
+#endif
+
+#define MI_INIT4(x)   x(),x(),x(),x()
+#define MI_INIT8(x)   MI_INIT4(x),MI_INIT4(x)
+#define MI_INIT16(x)  MI_INIT8(x),MI_INIT8(x)
+#define MI_INIT32(x)  MI_INIT16(x),MI_INIT16(x)
+#define MI_INIT64(x)  MI_INIT32(x),MI_INIT32(x)
+#define MI_INIT128(x) MI_INIT64(x),MI_INIT64(x)
+#define MI_INIT256(x) MI_INIT128(x),MI_INIT128(x)
+
+
+#include <string.h>
+// initialize a local variable to zero; use memset as compilers optimize constant sized memset's
+#define _mi_memzero_var(x)  memset(&x,0,sizeof(x))
+
+// Is `x` a power of two? (0 is considered a power of two)
+static inline bool _mi_is_power_of_two(uintptr_t x) {
+  return ((x & (x - 1)) == 0);
+}
+
+// Is a pointer aligned?
+static inline bool _mi_is_aligned(void* p, size_t alignment) {
+  mi_assert_internal(alignment != 0);
+  return (((uintptr_t)p % alignment) == 0);
+}
+
+// Align upwards
+static inline uintptr_t _mi_align_up(uintptr_t sz, size_t alignment) {
+  mi_assert_internal(alignment != 0);
+  uintptr_t mask = alignment - 1;
+  if ((alignment & mask) == 0) {  // power of two?
+    return ((sz + mask) & ~mask);
+  }
+  else {
+    return (((sz + mask)/alignment)*alignment);
+  }
+}
+
+// Align downwards
+static inline uintptr_t _mi_align_down(uintptr_t sz, size_t alignment) {
+  mi_assert_internal(alignment != 0);
+  uintptr_t mask = alignment - 1;
+  if ((alignment & mask) == 0) { // power of two?
+    return (sz & ~mask);
+  }
+  else {
+    return ((sz / alignment) * alignment);
+  }
+}
+
+// Divide upwards: `s <= _mi_divide_up(s,d)*d < s+d`.
+static inline uintptr_t _mi_divide_up(uintptr_t size, size_t divider) {
+  mi_assert_internal(divider != 0);
+  return (divider == 0 ? size : ((size + divider - 1) / divider));
+}
+
+// Is memory zero initialized?
+static inline bool mi_mem_is_zero(const void* p, size_t size) {
+  for (size_t i = 0; i < size; i++) {
+    if (((uint8_t*)p)[i] != 0) return false;
+  }
+  return true;
+}
+
+
+// Align a byte size to a size in _machine words_,
+// i.e. byte size == `wsize*sizeof(void*)`.
+static inline size_t _mi_wsize_from_size(size_t size) {
+  mi_assert_internal(size <= SIZE_MAX - sizeof(uintptr_t));
+  return (size + sizeof(uintptr_t) - 1) / sizeof(uintptr_t);
+}
+
+// Overflow detecting multiply
+#if __has_builtin(__builtin_umul_overflow) || (defined(__GNUC__) && (__GNUC__ >= 5))
+#include <limits.h>      // UINT_MAX, ULONG_MAX
+#if defined(_CLOCK_T)    // for Illumos
+#undef _CLOCK_T
+#endif
+static inline bool mi_mul_overflow(size_t count, size_t size, size_t* total) {
+  #if (SIZE_MAX == ULONG_MAX)
+    return __builtin_umull_overflow(count, size, (unsigned long *)total);
+  #elif (SIZE_MAX == UINT_MAX)
+    return __builtin_umul_overflow(count, size, (unsigned int *)total);
+  #else
+    return __builtin_umulll_overflow(count, size, (unsigned long long *)total);
+  #endif
+}
+#else /* __builtin_umul_overflow is unavailable */
+static inline bool mi_mul_overflow(size_t count, size_t size, size_t* total) {
+  #define MI_MUL_NO_OVERFLOW ((size_t)1 << (4*sizeof(size_t)))  // sqrt(SIZE_MAX)
+  *total = count * size;
+  // note: gcc/clang optimize this to directly check the overflow flag
+  return ((size >= MI_MUL_NO_OVERFLOW || count >= MI_MUL_NO_OVERFLOW) && size > 0 && (SIZE_MAX / size) < count);
+}
+#endif
+
+// Safe multiply `count*size` into `total`; return `true` on overflow.
+static inline bool mi_count_size_overflow(size_t count, size_t size, size_t* total) {
+  if (count==1) {  // quick check for the case where count is one (common for C++ allocators)
+    *total = size;
+    return false;
+  }
+  else if mi_unlikely(mi_mul_overflow(count, size, total)) {
+    #if MI_DEBUG > 0
+    _mi_error_message(EOVERFLOW, "allocation request is too large (%zu * %zu bytes)\n", count, size);
+    #endif
+    *total = SIZE_MAX;
+    return true;
+  }
+  else return false;
+}
+
+
+/*----------------------------------------------------------------------------------------
+  Heap functions
+------------------------------------------------------------------------------------------- */
+
+extern const mi_heap_t _mi_heap_empty;  // read-only empty heap, initial value of the thread local default heap
+
+static inline bool mi_heap_is_backing(const mi_heap_t* heap) {
+  return (heap->tld->heap_backing == heap);
+}
+
+static inline bool mi_heap_is_initialized(mi_heap_t* heap) {
+  mi_assert_internal(heap != NULL);
+  return (heap != &_mi_heap_empty);
+}
+
+static inline uintptr_t _mi_ptr_cookie(const void* p) {
+  extern mi_heap_t _mi_heap_main;
+  mi_assert_internal(_mi_heap_main.cookie != 0);
+  return ((uintptr_t)p ^ _mi_heap_main.cookie);
+}
+
+/* -----------------------------------------------------------
+  Pages
+----------------------------------------------------------- */
+
+static inline mi_page_t* _mi_heap_get_free_small_page(mi_heap_t* heap, size_t size) {
+  mi_assert_internal(size <= (MI_SMALL_SIZE_MAX + MI_PADDING_SIZE));
+  const size_t idx = _mi_wsize_from_size(size);
+  mi_assert_internal(idx < MI_PAGES_DIRECT);
+  return heap->pages_free_direct[idx];
+}
+
+// Segment that contains the pointer
+// Large aligned blocks may be aligned at N*MI_SEGMENT_SIZE (inside a huge segment > MI_SEGMENT_SIZE),
+// and we need align "down" to the segment info which is `MI_SEGMENT_SIZE` bytes before it;
+// therefore we align one byte before `p`.
+static inline mi_segment_t* _mi_ptr_segment(const void* p) {
+  mi_assert_internal(p != NULL);
+  return (mi_segment_t*)(((uintptr_t)p - 1) & ~MI_SEGMENT_MASK);
+}
+
+static inline mi_page_t* mi_slice_to_page(mi_slice_t* s) {
+  mi_assert_internal(s->slice_offset== 0 && s->slice_count > 0);
+  return (mi_page_t*)(s);
+}
+
+static inline mi_slice_t* mi_page_to_slice(mi_page_t* p) {
+  mi_assert_internal(p->slice_offset== 0 && p->slice_count > 0);
+  return (mi_slice_t*)(p);
+}
+
+// Segment belonging to a page
+static inline mi_segment_t* _mi_page_segment(const mi_page_t* page) {
+  mi_segment_t* segment = _mi_ptr_segment(page);
+  mi_assert_internal(segment == NULL || ((mi_slice_t*)page >= segment->slices && (mi_slice_t*)page < segment->slices + segment->slice_entries));
+  return segment;
+}
+
+static inline mi_slice_t* mi_slice_first(const mi_slice_t* slice) {
+  mi_slice_t* start = (mi_slice_t*)((uint8_t*)slice - slice->slice_offset);
+  mi_assert_internal(start >= _mi_ptr_segment(slice)->slices);
+  mi_assert_internal(start->slice_offset == 0);
+  mi_assert_internal(start + start->slice_count > slice);
+  return start;
+}
+
+// Get the page containing the pointer (performance critical as it is called in mi_free)
+static inline mi_page_t* _mi_segment_page_of(const mi_segment_t* segment, const void* p) {
+  mi_assert_internal(p > (void*)segment);
+  ptrdiff_t diff = (uint8_t*)p - (uint8_t*)segment;
+  mi_assert_internal(diff > 0 && diff <= (ptrdiff_t)MI_SEGMENT_SIZE);
+  size_t idx = (size_t)diff >> MI_SEGMENT_SLICE_SHIFT;
+  mi_assert_internal(idx <= segment->slice_entries);
+  mi_slice_t* slice0 = (mi_slice_t*)&segment->slices[idx];
+  mi_slice_t* slice = mi_slice_first(slice0);  // adjust to the block that holds the page data
+  mi_assert_internal(slice->slice_offset == 0);
+  mi_assert_internal(slice >= segment->slices && slice < segment->slices + segment->slice_entries);
+  return mi_slice_to_page(slice);
+}
+
+// Quick page start for initialized pages
+static inline uint8_t* _mi_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size) {
+  return _mi_segment_page_start(segment, page, page_size);
+}
+
+// Get the page containing the pointer
+static inline mi_page_t* _mi_ptr_page(void* p) {
+  return _mi_segment_page_of(_mi_ptr_segment(p), p);
+}
+
+// Get the block size of a page (special case for huge objects)
+static inline size_t mi_page_block_size(const mi_page_t* page) {
+  const size_t bsize = page->xblock_size;
+  mi_assert_internal(bsize > 0);
+  if mi_likely(bsize < MI_HUGE_BLOCK_SIZE) {
+    return bsize;
+  }
+  else {
+    size_t psize;
+    _mi_segment_page_start(_mi_page_segment(page), page, &psize);
+    return psize;
+  }
+}
+
+static inline bool mi_page_is_huge(const mi_page_t* page) {
+  return (_mi_page_segment(page)->kind == MI_SEGMENT_HUGE);
+}
+
+// Get the usable block size of a page without fixed padding.
+// This may still include internal padding due to alignment and rounding up size classes.
+static inline size_t mi_page_usable_block_size(const mi_page_t* page) {
+  return mi_page_block_size(page) - MI_PADDING_SIZE;
+}
+
+// size of a segment
+static inline size_t mi_segment_size(mi_segment_t* segment) {
+  return segment->segment_slices * MI_SEGMENT_SLICE_SIZE;
+}
+
+static inline uint8_t* mi_segment_end(mi_segment_t* segment) {
+  return (uint8_t*)segment + mi_segment_size(segment);
+}
+
+// Thread free access
+static inline mi_block_t* mi_page_thread_free(const mi_page_t* page) {
+  return (mi_block_t*)(mi_atomic_load_relaxed(&((mi_page_t*)page)->xthread_free) & ~3);
+}
+
+static inline mi_delayed_t mi_page_thread_free_flag(const mi_page_t* page) {
+  return (mi_delayed_t)(mi_atomic_load_relaxed(&((mi_page_t*)page)->xthread_free) & 3);
+}
+
+// Heap access
+static inline mi_heap_t* mi_page_heap(const mi_page_t* page) {
+  return (mi_heap_t*)(mi_atomic_load_relaxed(&((mi_page_t*)page)->xheap));
+}
+
+static inline void mi_page_set_heap(mi_page_t* page, mi_heap_t* heap) {
+  mi_assert_internal(mi_page_thread_free_flag(page) != MI_DELAYED_FREEING);
+  mi_atomic_store_release(&page->xheap,(uintptr_t)heap);
+}
+
+// Thread free flag helpers
+static inline mi_block_t* mi_tf_block(mi_thread_free_t tf) {
+  return (mi_block_t*)(tf & ~0x03);
+}
+static inline mi_delayed_t mi_tf_delayed(mi_thread_free_t tf) {
+  return (mi_delayed_t)(tf & 0x03);
+}
+static inline mi_thread_free_t mi_tf_make(mi_block_t* block, mi_delayed_t delayed) {
+  return (mi_thread_free_t)((uintptr_t)block | (uintptr_t)delayed);
+}
+static inline mi_thread_free_t mi_tf_set_delayed(mi_thread_free_t tf, mi_delayed_t delayed) {
+  return mi_tf_make(mi_tf_block(tf),delayed);
+}
+static inline mi_thread_free_t mi_tf_set_block(mi_thread_free_t tf, mi_block_t* block) {
+  return mi_tf_make(block, mi_tf_delayed(tf));
+}
+
+// are all blocks in a page freed?
+// note: needs up-to-date used count, (as the `xthread_free` list may not be empty). see `_mi_page_collect_free`.
+static inline bool mi_page_all_free(const mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  return (page->used == 0);
+}
+
+// are there any available blocks?
+static inline bool mi_page_has_any_available(const mi_page_t* page) {
+  mi_assert_internal(page != NULL && page->reserved > 0);
+  return (page->used < page->reserved || (mi_page_thread_free(page) != NULL));
+}
+
+// are there immediately available blocks, i.e. blocks available on the free list.
+static inline bool mi_page_immediate_available(const mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  return (page->free != NULL);
+}
+
+// is more than 7/8th of a page in use?
+static inline bool mi_page_mostly_used(const mi_page_t* page) {
+  if (page==NULL) return true;
+  uint16_t frac = page->reserved / 8U;
+  return (page->reserved - page->used <= frac);
+}
+
+static inline mi_page_queue_t* mi_page_queue(const mi_heap_t* heap, size_t size) {
+  return &((mi_heap_t*)heap)->pages[_mi_bin(size)];
+}
+
+
+
+//-----------------------------------------------------------
+// Page flags
+//-----------------------------------------------------------
+static inline bool mi_page_is_in_full(const mi_page_t* page) {
+  return page->flags.x.in_full;
+}
+
+static inline void mi_page_set_in_full(mi_page_t* page, bool in_full) {
+  page->flags.x.in_full = in_full;
+}
+
+static inline bool mi_page_has_aligned(const mi_page_t* page) {
+  return page->flags.x.has_aligned;
+}
+
+static inline void mi_page_set_has_aligned(mi_page_t* page, bool has_aligned) {
+  page->flags.x.has_aligned = has_aligned;
+}
+
+
+/* -------------------------------------------------------------------
+Encoding/Decoding the free list next pointers
+
+This is to protect against buffer overflow exploits where the
+free list is mutated. Many hardened allocators xor the next pointer `p`
+with a secret key `k1`, as `p^k1`. This prevents overwriting with known
+values but might be still too weak: if the attacker can guess
+the pointer `p` this  can reveal `k1` (since `p^k1^p == k1`).
+Moreover, if multiple blocks can be read as well, the attacker can
+xor both as `(p1^k1) ^ (p2^k1) == p1^p2` which may reveal a lot
+about the pointers (and subsequently `k1`).
+
+Instead mimalloc uses an extra key `k2` and encodes as `((p^k2)<<<k1)+k1`.
+Since these operations are not associative, the above approaches do not
+work so well any more even if the `p` can be guesstimated. For example,
+for the read case we can subtract two entries to discard the `+k1` term,
+but that leads to `((p1^k2)<<<k1) - ((p2^k2)<<<k1)` at best.
+We include the left-rotation since xor and addition are otherwise linear
+in the lowest bit. Finally, both keys are unique per page which reduces
+the re-use of keys by a large factor.
+
+We also pass a separate `null` value to be used as `NULL` or otherwise
+`(k2<<<k1)+k1` would appear (too) often as a sentinel value.
+------------------------------------------------------------------- */
+
+static inline bool mi_is_in_same_segment(const void* p, const void* q) {
+  return (_mi_ptr_segment(p) == _mi_ptr_segment(q));
+}
+
+static inline bool mi_is_in_same_page(const void* p, const void* q) {
+  mi_segment_t* segment = _mi_ptr_segment(p);
+  if (_mi_ptr_segment(q) != segment) return false;
+  // assume q may be invalid // return (_mi_segment_page_of(segment, p) == _mi_segment_page_of(segment, q));
+  mi_page_t* page = _mi_segment_page_of(segment, p);
+  size_t psize;
+  uint8_t* start = _mi_segment_page_start(segment, page, &psize);
+  return (start <= (uint8_t*)q && (uint8_t*)q < start + psize);
+}
+
+static inline uintptr_t mi_rotl(uintptr_t x, uintptr_t shift) {
+  shift %= MI_INTPTR_BITS;
+  return (shift==0 ? x : ((x << shift) | (x >> (MI_INTPTR_BITS - shift))));
+}
+static inline uintptr_t mi_rotr(uintptr_t x, uintptr_t shift) {
+  shift %= MI_INTPTR_BITS;
+  return (shift==0 ? x : ((x >> shift) | (x << (MI_INTPTR_BITS - shift))));
+}
+
+static inline void* mi_ptr_decode(const void* null, const mi_encoded_t x, const uintptr_t* keys) {
+  void* p = (void*)(mi_rotr(x - keys[0], keys[0]) ^ keys[1]);
+  return (p==null ? NULL : p);
+}
+
+static inline mi_encoded_t mi_ptr_encode(const void* null, const void* p, const uintptr_t* keys) {
+  uintptr_t x = (uintptr_t)(p==NULL ? null : p);
+  return mi_rotl(x ^ keys[1], keys[0]) + keys[0];
+}
+
+static inline mi_block_t* mi_block_nextx( const void* null, const mi_block_t* block, const uintptr_t* keys ) {
+  mi_track_mem_defined(block,sizeof(mi_block_t));
+  mi_block_t* next;
+  #ifdef MI_ENCODE_FREELIST
+  next = (mi_block_t*)mi_ptr_decode(null, block->next, keys);
+  #else
+  MI_UNUSED(keys); MI_UNUSED(null);
+  next = (mi_block_t*)block->next;
+  #endif
+  mi_track_mem_noaccess(block,sizeof(mi_block_t));
+  return next;
+}
+
+static inline void mi_block_set_nextx(const void* null, mi_block_t* block, const mi_block_t* next, const uintptr_t* keys) {
+  mi_track_mem_undefined(block,sizeof(mi_block_t));
+  #ifdef MI_ENCODE_FREELIST
+  block->next = mi_ptr_encode(null, next, keys);
+  #else
+  MI_UNUSED(keys); MI_UNUSED(null);
+  block->next = (mi_encoded_t)next;
+  #endif
+  mi_track_mem_noaccess(block,sizeof(mi_block_t));
+}
+
+static inline mi_block_t* mi_block_next(const mi_page_t* page, const mi_block_t* block) {
+  #ifdef MI_ENCODE_FREELIST
+  mi_block_t* next = mi_block_nextx(page,block,page->keys);
+  // check for free list corruption: is `next` at least in the same page?
+  // TODO: check if `next` is `page->block_size` aligned?
+  if mi_unlikely(next!=NULL && !mi_is_in_same_page(block, next)) {
+    _mi_error_message(EFAULT, "corrupted free list entry of size %zub at %p: value 0x%zx\n", mi_page_block_size(page), block, (uintptr_t)next);
+    next = NULL;
+  }
+  return next;
+  #else
+  MI_UNUSED(page);
+  return mi_block_nextx(page,block,NULL);
+  #endif
+}
+
+static inline void mi_block_set_next(const mi_page_t* page, mi_block_t* block, const mi_block_t* next) {
+  #ifdef MI_ENCODE_FREELIST
+  mi_block_set_nextx(page,block,next, page->keys);
+  #else
+  MI_UNUSED(page);
+  mi_block_set_nextx(page,block,next,NULL);
+  #endif
+}
+
+
+// -------------------------------------------------------------------
+// commit mask
+// -------------------------------------------------------------------
+
+static inline void mi_commit_mask_create_empty(mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    cm->mask[i] = 0;
+  }
+}
+
+static inline void mi_commit_mask_create_full(mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    cm->mask[i] = ~((size_t)0);
+  }
+}
+
+static inline bool mi_commit_mask_is_empty(const mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    if (cm->mask[i] != 0) return false;
+  }
+  return true;
+}
+
+static inline bool mi_commit_mask_is_full(const mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    if (cm->mask[i] != ~((size_t)0)) return false;
+  }
+  return true;
+}
+
+// defined in `segment.c`:
+size_t _mi_commit_mask_committed_size(const mi_commit_mask_t* cm, size_t total);
+size_t _mi_commit_mask_next_run(const mi_commit_mask_t* cm, size_t* idx);
+
+#define mi_commit_mask_foreach(cm,idx,count) \
+  idx = 0; \
+  while ((count = _mi_commit_mask_next_run(cm,&idx)) > 0) {
+
+#define mi_commit_mask_foreach_end() \
+    idx += count; \
+  }
+
+
+
+/* -----------------------------------------------------------
+  memory id's
+----------------------------------------------------------- */
+
+static inline mi_memid_t _mi_memid_create(mi_memkind_t memkind) {
+  mi_memid_t memid;
+  _mi_memzero_var(memid);
+  memid.memkind = memkind;
+  return memid;
+}
+
+static inline mi_memid_t _mi_memid_none(void) {
+  return _mi_memid_create(MI_MEM_NONE);
+}
+
+static inline mi_memid_t _mi_memid_create_os(bool committed, bool is_zero, bool is_large) {
+  mi_memid_t memid = _mi_memid_create(MI_MEM_OS);
+  memid.initially_committed = committed;
+  memid.initially_zero = is_zero;
+  memid.is_pinned = is_large;
+  return memid;
+}
+
+
+// -------------------------------------------------------------------
+// Fast "random" shuffle
+// -------------------------------------------------------------------
+
+static inline uintptr_t _mi_random_shuffle(uintptr_t x) {
+  if (x==0) { x = 17; }   // ensure we don't get stuck in generating zeros
+#if (MI_INTPTR_SIZE==8)
+  // by Sebastiano Vigna, see: <http://xoshiro.di.unimi.it/splitmix64.c>
+  x ^= x >> 30;
+  x *= 0xbf58476d1ce4e5b9UL;
+  x ^= x >> 27;
+  x *= 0x94d049bb133111ebUL;
+  x ^= x >> 31;
+#elif (MI_INTPTR_SIZE==4)
+  // by Chris Wellons, see: <https://nullprogram.com/blog/2018/07/31/>
+  x ^= x >> 16;
+  x *= 0x7feb352dUL;
+  x ^= x >> 15;
+  x *= 0x846ca68bUL;
+  x ^= x >> 16;
+#endif
+  return x;
+}
+
+// -------------------------------------------------------------------
+// Optimize numa node access for the common case (= one node)
+// -------------------------------------------------------------------
+
+int    _mi_os_numa_node_get(mi_os_tld_t* tld);
+size_t _mi_os_numa_node_count_get(void);
+
+extern _Atomic(size_t) _mi_numa_node_count;
+static inline int _mi_os_numa_node(mi_os_tld_t* tld) {
+  if mi_likely(mi_atomic_load_relaxed(&_mi_numa_node_count) == 1) { return 0; }
+  else return _mi_os_numa_node_get(tld);
+}
+static inline size_t _mi_os_numa_node_count(void) {
+  const size_t count = mi_atomic_load_relaxed(&_mi_numa_node_count);
+  if mi_likely(count > 0) { return count; }
+  else return _mi_os_numa_node_count_get();
+}
+
+
+
+// -----------------------------------------------------------------------
+// Count bits: trailing or leading zeros (with MI_INTPTR_BITS on all zero)
+// -----------------------------------------------------------------------
+
+#if defined(__GNUC__)
+
+#include <limits.h>       // LONG_MAX
+#define MI_HAVE_FAST_BITSCAN
+static inline size_t mi_clz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (INTPTR_MAX == LONG_MAX)
+  return __builtin_clzl(x);
+#else
+  return __builtin_clzll(x);
+#endif
+}
+static inline size_t mi_ctz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (INTPTR_MAX == LONG_MAX)
+  return __builtin_ctzl(x);
+#else
+  return __builtin_ctzll(x);
+#endif
+}
+
+#elif defined(_MSC_VER)
+
+#include <limits.h>       // LONG_MAX
+#include <intrin.h>       // BitScanReverse64
+#define MI_HAVE_FAST_BITSCAN
+static inline size_t mi_clz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+  unsigned long idx;
+#if (INTPTR_MAX == LONG_MAX)
+  _BitScanReverse(&idx, x);
+#else
+  _BitScanReverse64(&idx, x);
+#endif
+  return ((MI_INTPTR_BITS - 1) - idx);
+}
+static inline size_t mi_ctz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+  unsigned long idx;
+#if (INTPTR_MAX == LONG_MAX)
+  _BitScanForward(&idx, x);
+#else
+  _BitScanForward64(&idx, x);
+#endif
+  return idx;
+}
+
+#else
+static inline size_t mi_ctz32(uint32_t x) {
+  // de Bruijn multiplication, see <http://supertech.csail.mit.edu/papers/debruijn.pdf>
+  static const unsigned char debruijn[32] = {
+    0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
+    31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
+  };
+  if (x==0) return 32;
+  return debruijn[((x & -(int32_t)x) * 0x077CB531UL) >> 27];
+}
+static inline size_t mi_clz32(uint32_t x) {
+  // de Bruijn multiplication, see <http://supertech.csail.mit.edu/papers/debruijn.pdf>
+  static const uint8_t debruijn[32] = {
+    31, 22, 30, 21, 18, 10, 29, 2, 20, 17, 15, 13, 9, 6, 28, 1,
+    23, 19, 11, 3, 16, 14, 7, 24, 12, 4, 8, 25, 5, 26, 27, 0
+  };
+  if (x==0) return 32;
+  x |= x >> 1;
+  x |= x >> 2;
+  x |= x >> 4;
+  x |= x >> 8;
+  x |= x >> 16;
+  return debruijn[(uint32_t)(x * 0x07C4ACDDUL) >> 27];
+}
+
+static inline size_t mi_clz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (MI_INTPTR_BITS <= 32)
+  return mi_clz32((uint32_t)x);
+#else
+  size_t count = mi_clz32((uint32_t)(x >> 32));
+  if (count < 32) return count;
+  return (32 + mi_clz32((uint32_t)x));
+#endif
+}
+static inline size_t mi_ctz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (MI_INTPTR_BITS <= 32)
+  return mi_ctz32((uint32_t)x);
+#else
+  size_t count = mi_ctz32((uint32_t)x);
+  if (count < 32) return count;
+  return (32 + mi_ctz32((uint32_t)(x>>32)));
+#endif
+}
+
+#endif
+
+// "bit scan reverse": Return index of the highest bit (or MI_INTPTR_BITS if `x` is zero)
+static inline size_t mi_bsr(uintptr_t x) {
+  return (x==0 ? MI_INTPTR_BITS : MI_INTPTR_BITS - 1 - mi_clz(x));
+}
+
+
+// ---------------------------------------------------------------------------------
+// Provide our own `_mi_memcpy` for potential performance optimizations.
+//
+// For now, only on Windows with msvc/clang-cl we optimize to `rep movsb` if
+// we happen to run on x86/x64 cpu's that have "fast short rep movsb" (FSRM) support
+// (AMD Zen3+ (~2020) or Intel Ice Lake+ (~2017). See also issue #201 and pr #253.
+// ---------------------------------------------------------------------------------
+
+#if !MI_TRACK_ENABLED && defined(_WIN32) && (defined(_M_IX86) || defined(_M_X64))
+#include <intrin.h>
+extern bool _mi_cpu_has_fsrm;
+static inline void _mi_memcpy(void* dst, const void* src, size_t n) {
+  if (_mi_cpu_has_fsrm) {
+    __movsb((unsigned char*)dst, (const unsigned char*)src, n);
+  }
+  else {
+    memcpy(dst, src, n);
+  }
+}
+static inline void _mi_memzero(void* dst, size_t n) {
+  if (_mi_cpu_has_fsrm) {
+    __stosb((unsigned char*)dst, 0, n);
+  }
+  else {
+    memset(dst, 0, n);
+  }
+}
+#else
+static inline void _mi_memcpy(void* dst, const void* src, size_t n) {
+  memcpy(dst, src, n);
+}
+static inline void _mi_memzero(void* dst, size_t n) {
+  memset(dst, 0, n);
+}
+#endif
+
+// -------------------------------------------------------------------------------
+// The `_mi_memcpy_aligned` can be used if the pointers are machine-word aligned
+// This is used for example in `mi_realloc`.
+// -------------------------------------------------------------------------------
+
+#if (defined(__GNUC__) && (__GNUC__ >= 4)) || defined(__clang__)
+// On GCC/CLang we provide a hint that the pointers are word aligned.
+static inline void _mi_memcpy_aligned(void* dst, const void* src, size_t n) {
+  mi_assert_internal(((uintptr_t)dst % MI_INTPTR_SIZE == 0) && ((uintptr_t)src % MI_INTPTR_SIZE == 0));
+  void* adst = __builtin_assume_aligned(dst, MI_INTPTR_SIZE);
+  const void* asrc = __builtin_assume_aligned(src, MI_INTPTR_SIZE);
+  _mi_memcpy(adst, asrc, n);
+}
+
+static inline void _mi_memzero_aligned(void* dst, size_t n) {
+  mi_assert_internal((uintptr_t)dst % MI_INTPTR_SIZE == 0);
+  void* adst = __builtin_assume_aligned(dst, MI_INTPTR_SIZE);
+  _mi_memzero(adst, n);
+}
+#else
+// Default fallback on `_mi_memcpy`
+static inline void _mi_memcpy_aligned(void* dst, const void* src, size_t n) {
+  mi_assert_internal(((uintptr_t)dst % MI_INTPTR_SIZE == 0) && ((uintptr_t)src % MI_INTPTR_SIZE == 0));
+  _mi_memcpy(dst, src, n);
+}
+
+static inline void _mi_memzero_aligned(void* dst, size_t n) {
+  mi_assert_internal((uintptr_t)dst % MI_INTPTR_SIZE == 0);
+  _mi_memzero(dst, n);
+}
+#endif
+
+
+#endif
diff --git a/compat/mimalloc/mimalloc/prim.h b/compat/mimalloc/mimalloc/prim.h
new file mode 100644
index 00000000000000..1e55cb5f8802d7
--- /dev/null
+++ b/compat/mimalloc/mimalloc/prim.h
@@ -0,0 +1,323 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_PRIM_H
+#define MIMALLOC_PRIM_H
+
+
+// --------------------------------------------------------------------------
+// This file specifies the primitive portability API.
+// Each OS/host needs to implement these primitives, see `src/prim`
+// for implementations on Window, macOS, WASI, and Linux/Unix.
+//
+// note: on all primitive functions, we always have result parameters != NUL, and:
+//  addr != NULL and page aligned
+//  size > 0     and page aligned
+//  return value is an error code an int where 0 is success.
+// --------------------------------------------------------------------------
+
+// OS memory configuration
+typedef struct mi_os_mem_config_s {
+  size_t  page_size;            // 4KiB
+  size_t  large_page_size;      // 2MiB
+  size_t  alloc_granularity;    // smallest allocation size (on Windows 64KiB)
+  bool    has_overcommit;       // can we reserve more memory than can be actually committed?
+  bool    must_free_whole;      // must allocated blocks be freed as a whole (false for mmap, true for VirtualAlloc)
+  bool    has_virtual_reserve;  // supports virtual address space reservation? (if true we can reserve virtual address space without using commit or physical memory)
+} mi_os_mem_config_t;
+
+// Initialize
+void _mi_prim_mem_init( mi_os_mem_config_t* config );
+
+// Free OS memory
+int _mi_prim_free(void* addr, size_t size );
+
+// Allocate OS memory. Return NULL on error.
+// The `try_alignment` is just a hint and the returned pointer does not have to be aligned.
+// If `commit` is false, the virtual memory range only needs to be reserved (with no access)
+// which will later be committed explicitly using `_mi_prim_commit`.
+// `is_zero` is set to true if the memory was zero initialized (as on most OS's)
+// pre: !commit => !allow_large
+//      try_alignment >= _mi_os_page_size() and a power of 2
+int _mi_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, void** addr);
+
+// Commit memory. Returns error code or 0 on success.
+// For example, on Linux this would make the memory PROT_READ|PROT_WRITE.
+// `is_zero` is set to true if the memory was zero initialized (e.g. on Windows)
+int _mi_prim_commit(void* addr, size_t size, bool* is_zero);
+
+// Decommit memory. Returns error code or 0 on success. The `needs_recommit` result is true
+// if the memory would need to be re-committed. For example, on Windows this is always true,
+// but on Linux we could use MADV_DONTNEED to decommit which does not need a recommit.
+// pre: needs_recommit != NULL
+int _mi_prim_decommit(void* addr, size_t size, bool* needs_recommit);
+
+// Reset memory. The range keeps being accessible but the content might be reset.
+// Returns error code or 0 on success.
+int _mi_prim_reset(void* addr, size_t size);
+
+// Protect memory. Returns error code or 0 on success.
+int _mi_prim_protect(void* addr, size_t size, bool protect);
+
+// Allocate huge (1GiB) pages possibly associated with a NUMA node.
+// `is_zero` is set to true if the memory was zero initialized (as on most OS's)
+// pre: size > 0  and a multiple of 1GiB.
+//      numa_node is either negative (don't care), or a numa node number.
+int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, bool* is_zero, void** addr);
+
+// Return the current NUMA node
+size_t _mi_prim_numa_node(void);
+
+// Return the number of logical NUMA nodes
+size_t _mi_prim_numa_node_count(void);
+
+// Clock ticks
+mi_msecs_t _mi_prim_clock_now(void);
+
+// Return process information (only for statistics)
+typedef struct mi_process_info_s {
+  mi_msecs_t  elapsed;
+  mi_msecs_t  utime;
+  mi_msecs_t  stime;
+  size_t      current_rss;
+  size_t      peak_rss;
+  size_t      current_commit;
+  size_t      peak_commit;
+  size_t      page_faults;
+} mi_process_info_t;
+
+void _mi_prim_process_info(mi_process_info_t* pinfo);
+
+// Default stderr output. (only for warnings etc. with verbose enabled)
+// msg != NULL && _mi_strlen(msg) > 0
+void _mi_prim_out_stderr( const char* msg );
+
+// Get an environment variable. (only for options)
+// name != NULL, result != NULL, result_size >= 64
+bool _mi_prim_getenv(const char* name, char* result, size_t result_size);
+
+
+// Fill a buffer with strong randomness; return `false` on error or if
+// there is no strong randomization available.
+bool _mi_prim_random_buf(void* buf, size_t buf_len);
+
+// Called on the first thread start, and should ensure `_mi_thread_done` is called on thread termination.
+void _mi_prim_thread_init_auto_done(void);
+
+// Called on process exit and may take action to clean up resources associated with the thread auto done.
+void _mi_prim_thread_done_auto_done(void);
+
+// Called when the default heap for a thread changes
+void _mi_prim_thread_associate_default_heap(mi_heap_t* heap);
+
+
+//-------------------------------------------------------------------
+// Thread id: `_mi_prim_thread_id()`
+//
+// Getting the thread id should be performant as it is called in the
+// fast path of `_mi_free` and we specialize for various platforms as
+// inlined definitions. Regular code should call `init.c:_mi_thread_id()`.
+// We only require _mi_prim_thread_id() to return a unique id
+// for each thread (unequal to zero).
+//-------------------------------------------------------------------
+
+// defined in `init.c`; do not use these directly
+extern mi_decl_thread mi_heap_t* _mi_heap_default;  // default heap to allocate from
+extern bool _mi_process_is_initialized;             // has mi_process_init been called?
+
+static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept;
+
+#if defined(_WIN32)
+
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
+  // Windows: works on Intel and ARM in both 32- and 64-bit
+  return (uintptr_t)NtCurrentTeb();
+}
+
+// We use assembly for a fast thread id on the main platforms. The TLS layout depends on
+// both the OS and libc implementation so we use specific tests for each main platform.
+// If you test on another platform and it works please send a PR :-)
+// see also https://akkadia.org/drepper/tls.pdf for more info on the TLS register.
+#elif defined(__GNUC__) && ( \
+	   (defined(__GLIBC__)   && (defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__aarch64__))) \
+	|| (defined(__APPLE__)   && (defined(__x86_64__) || defined(__aarch64__))) \
+	|| (defined(__BIONIC__)  && (defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__aarch64__))) \
+	|| (defined(__FreeBSD__) && (defined(__x86_64__) || defined(__i386__) || defined(__aarch64__))) \
+	|| (defined(__OpenBSD__) && (defined(__x86_64__) || defined(__i386__) || defined(__aarch64__))) \
+      )
+
+static inline void* mi_prim_tls_slot(size_t slot) mi_attr_noexcept {
+  void* res;
+  const size_t ofs = (slot*sizeof(void*));
+  #if defined(__i386__)
+    __asm__("movl %%gs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : );  // x86 32-bit always uses GS
+  #elif defined(__APPLE__) && defined(__x86_64__)
+    __asm__("movq %%gs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : );  // x86_64 macOSX uses GS
+  #elif defined(__x86_64__) && (MI_INTPTR_SIZE==4)
+    __asm__("movl %%fs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : );  // x32 ABI
+  #elif defined(__x86_64__)
+    __asm__("movq %%fs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : );  // x86_64 Linux, BSD uses FS
+  #elif defined(__arm__)
+    void** tcb; MI_UNUSED(ofs);
+    __asm__ volatile ("mrc p15, 0, %0, c13, c0, 3\nbic %0, %0, #3" : "=r" (tcb));
+    res = tcb[slot];
+  #elif defined(__aarch64__)
+    void** tcb; MI_UNUSED(ofs);
+    #if defined(__APPLE__) // M1, issue #343
+    __asm__ volatile ("mrs %0, tpidrro_el0\nbic %0, %0, #7" : "=r" (tcb));
+    #else
+    __asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
+    #endif
+    res = tcb[slot];
+  #endif
+  return res;
+}
+
+// setting a tls slot is only used on macOS for now
+static inline void mi_prim_tls_slot_set(size_t slot, void* value) mi_attr_noexcept {
+  const size_t ofs = (slot*sizeof(void*));
+  #if defined(__i386__)
+    __asm__("movl %1,%%gs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : );  // 32-bit always uses GS
+  #elif defined(__APPLE__) && defined(__x86_64__)
+    __asm__("movq %1,%%gs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : );  // x86_64 macOS uses GS
+  #elif defined(__x86_64__) && (MI_INTPTR_SIZE==4)
+    __asm__("movl %1,%%fs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : );  // x32 ABI
+  #elif defined(__x86_64__)
+    __asm__("movq %1,%%fs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : );  // x86_64 Linux, BSD uses FS
+  #elif defined(__arm__)
+    void** tcb; MI_UNUSED(ofs);
+    __asm__ volatile ("mrc p15, 0, %0, c13, c0, 3\nbic %0, %0, #3" : "=r" (tcb));
+    tcb[slot] = value;
+  #elif defined(__aarch64__)
+    void** tcb; MI_UNUSED(ofs);
+    #if defined(__APPLE__) // M1, issue #343
+    __asm__ volatile ("mrs %0, tpidrro_el0\nbic %0, %0, #7" : "=r" (tcb));
+    #else
+    __asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
+    #endif
+    tcb[slot] = value;
+  #endif
+}
+
+static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
+  #if defined(__BIONIC__)
+    // issue #384, #495: on the Bionic libc (Android), slot 1 is the thread id
+    // see: https://github.com/aosp-mirror/platform_bionic/blob/c44b1d0676ded732df4b3b21c5f798eacae93228/libc/platform/bionic/tls_defines.h#L86
+    return (uintptr_t)mi_prim_tls_slot(1);
+  #else
+    // in all our other targets, slot 0 is the thread id
+    // glibc: https://sourceware.org/git/?p=glibc.git;a=blob_plain;f=sysdeps/x86_64/nptl/tls.h
+    // apple: https://github.com/apple/darwin-xnu/blob/main/libsyscall/os/tsd.h#L36
+    return (uintptr_t)mi_prim_tls_slot(0);
+  #endif
+}
+
+#else
+
+// otherwise use portable C, taking the address of a thread local variable (this is still very fast on most platforms).
+static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
+  return (uintptr_t)&_mi_heap_default;
+}
+
+#endif
+
+
+
+/* ----------------------------------------------------------------------------------------
+The thread local default heap: `_mi_prim_get_default_heap()`
+This is inlined here as it is on the fast path for allocation functions.
+
+On most platforms (Windows, Linux, FreeBSD, NetBSD, etc), this just returns a
+__thread local variable (`_mi_heap_default`). With the initial-exec TLS model this ensures
+that the storage will always be available (allocated on the thread stacks).
+
+On some platforms though we cannot use that when overriding `malloc` since the underlying
+TLS implementation (or the loader) will call itself `malloc` on a first access and recurse.
+We try to circumvent this in an efficient way:
+- macOSX : we use an unused TLS slot from the OS allocated slots (MI_TLS_SLOT). On OSX, the
+	   loader itself calls `malloc` even before the modules are initialized.
+- OpenBSD: we use an unused slot from the pthread block (MI_TLS_PTHREAD_SLOT_OFS).
+- DragonFly: defaults are working but seem slow compared to freeBSD (see PR #323)
+------------------------------------------------------------------------------------------- */
+
+static inline mi_heap_t* mi_prim_get_default_heap(void);
+
+#if defined(MI_MALLOC_OVERRIDE)
+#if defined(__APPLE__) // macOS
+  #define MI_TLS_SLOT               89  // seems unused?
+  // #define MI_TLS_RECURSE_GUARD 1
+  // other possible unused ones are 9, 29, __PTK_FRAMEWORK_JAVASCRIPTCORE_KEY4 (94), __PTK_FRAMEWORK_GC_KEY9 (112) and __PTK_FRAMEWORK_OLDGC_KEY9 (89)
+  // see <https://github.com/rweichler/substrate/blob/master/include/pthread_machdep.h>
+#elif defined(__OpenBSD__)
+  // use end bytes of a name; goes wrong if anyone uses names > 23 characters (ptrhread specifies 16)
+  // see <https://github.com/openbsd/src/blob/master/lib/libc/include/thread_private.h#L371>
+  #define MI_TLS_PTHREAD_SLOT_OFS   (6*sizeof(int) + 4*sizeof(void*) + 24)
+  // #elif defined(__DragonFly__)
+  // #warning "mimalloc is not working correctly on DragonFly yet."
+  // #define MI_TLS_PTHREAD_SLOT_OFS   (4 + 1*sizeof(void*))  // offset `uniqueid` (also used by gdb?) <https://github.com/DragonFlyBSD/DragonFlyBSD/blob/master/lib/libthread_xu/thread/thr_private.h#L458>
+#elif defined(__ANDROID__)
+  // See issue #381
+  #define MI_TLS_PTHREAD
+#endif
+#endif
+
+
+#if defined(MI_TLS_SLOT)
+
+static inline mi_heap_t* mi_prim_get_default_heap(void) {
+  mi_heap_t* heap = (mi_heap_t*)mi_prim_tls_slot(MI_TLS_SLOT);
+  if mi_unlikely(heap == NULL) {
+    #ifdef __GNUC__
+    __asm(""); // prevent conditional load of the address of _mi_heap_empty
+    #endif
+    heap = (mi_heap_t*)&_mi_heap_empty;
+  }
+  return heap;
+}
+
+#elif defined(MI_TLS_PTHREAD_SLOT_OFS)
+
+static inline mi_heap_t** mi_prim_tls_pthread_heap_slot(void) {
+  pthread_t self = pthread_self();
+  #if defined(__DragonFly__)
+  if (self==NULL) return NULL;
+  #endif
+  return (mi_heap_t**)((uint8_t*)self + MI_TLS_PTHREAD_SLOT_OFS);
+}
+
+static inline mi_heap_t* mi_prim_get_default_heap(void) {
+  mi_heap_t** pheap = mi_prim_tls_pthread_heap_slot();
+  if mi_unlikely(pheap == NULL) return _mi_heap_main_get();
+  mi_heap_t* heap = *pheap;
+  if mi_unlikely(heap == NULL) return (mi_heap_t*)&_mi_heap_empty;
+  return heap;
+}
+
+#elif defined(MI_TLS_PTHREAD)
+
+extern pthread_key_t _mi_heap_default_key;
+static inline mi_heap_t* mi_prim_get_default_heap(void) {
+  mi_heap_t* heap = (mi_unlikely(_mi_heap_default_key == (pthread_key_t)(-1)) ? _mi_heap_main_get() : (mi_heap_t*)pthread_getspecific(_mi_heap_default_key));
+  return (mi_unlikely(heap == NULL) ? (mi_heap_t*)&_mi_heap_empty : heap);
+}
+
+#else // default using a thread local variable; used on most platforms.
+
+static inline mi_heap_t* mi_prim_get_default_heap(void) {
+  #if defined(MI_TLS_RECURSE_GUARD)
+  if (mi_unlikely(!_mi_process_is_initialized)) return _mi_heap_main_get();
+  #endif
+  return _mi_heap_default;
+}
+
+#endif  // mi_prim_get_default_heap()
+
+
+
+#endif  // MIMALLOC_PRIM_H
diff --git a/compat/mimalloc/mimalloc/track.h b/compat/mimalloc/mimalloc/track.h
new file mode 100644
index 00000000000000..fa1a048d846a9c
--- /dev/null
+++ b/compat/mimalloc/mimalloc/track.h
@@ -0,0 +1,147 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_TRACK_H
+#define MIMALLOC_TRACK_H
+
+/* ------------------------------------------------------------------------------------------------------
+Track memory ranges with macros for tools like Valgrind address sanitizer, or other memory checkers.
+These can be defined for tracking allocation:
+
+  #define mi_track_malloc_size(p,reqsize,size,zero)
+  #define mi_track_free_size(p,_size)
+
+The macros are set up such that the size passed to `mi_track_free_size`
+always matches the size of `mi_track_malloc_size`. (currently, `size == mi_usable_size(p)`).
+The `reqsize` is what the user requested, and `size >= reqsize`.
+The `size` is either byte precise (and `size==reqsize`) if `MI_PADDING` is enabled,
+or otherwise it is the usable block size which may be larger than the original request.
+Use `_mi_block_size_of(void* p)` to get the full block size that was allocated (including padding etc).
+The `zero` parameter is `true` if the allocated block is zero initialized.
+
+Optional:
+
+  #define mi_track_align(p,alignedp,offset,size)
+  #define mi_track_resize(p,oldsize,newsize)
+  #define mi_track_init()
+
+The `mi_track_align` is called right after a `mi_track_malloc` for aligned pointers in a block.
+The corresponding `mi_track_free` still uses the block start pointer and original size (corresponding to the `mi_track_malloc`).
+The `mi_track_resize` is currently unused but could be called on reallocations within a block.
+`mi_track_init` is called at program start.
+
+The following macros are for tools like asan and valgrind to track whether memory is
+defined, undefined, or not accessible at all:
+
+  #define mi_track_mem_defined(p,size)
+  #define mi_track_mem_undefined(p,size)
+  #define mi_track_mem_noaccess(p,size)
+
+-------------------------------------------------------------------------------------------------------*/
+
+#if MI_TRACK_VALGRIND
+// valgrind tool
+
+#define MI_TRACK_ENABLED      1
+#define MI_TRACK_HEAP_DESTROY 1           // track free of individual blocks on heap_destroy
+#define MI_TRACK_TOOL         "valgrind"
+
+#include <valgrind/valgrind.h>
+#include <valgrind/memcheck.h>
+
+#define mi_track_malloc_size(p,reqsize,size,zero) VALGRIND_MALLOCLIKE_BLOCK(p,size,MI_PADDING_SIZE /*red zone*/,zero)
+#define mi_track_free_size(p,_size)               VALGRIND_FREELIKE_BLOCK(p,MI_PADDING_SIZE /*red zone*/)
+#define mi_track_resize(p,oldsize,newsize)        VALGRIND_RESIZEINPLACE_BLOCK(p,oldsize,newsize,MI_PADDING_SIZE /*red zone*/)
+#define mi_track_mem_defined(p,size)              VALGRIND_MAKE_MEM_DEFINED(p,size)
+#define mi_track_mem_undefined(p,size)            VALGRIND_MAKE_MEM_UNDEFINED(p,size)
+#define mi_track_mem_noaccess(p,size)             VALGRIND_MAKE_MEM_NOACCESS(p,size)
+
+#elif MI_TRACK_ASAN
+// address sanitizer
+
+#define MI_TRACK_ENABLED      1
+#define MI_TRACK_HEAP_DESTROY 0
+#define MI_TRACK_TOOL         "asan"
+
+#include <sanitizer/asan_interface.h>
+
+#define mi_track_malloc_size(p,reqsize,size,zero) ASAN_UNPOISON_MEMORY_REGION(p,size)
+#define mi_track_free_size(p,size)                ASAN_POISON_MEMORY_REGION(p,size)
+#define mi_track_mem_defined(p,size)              ASAN_UNPOISON_MEMORY_REGION(p,size)
+#define mi_track_mem_undefined(p,size)            ASAN_UNPOISON_MEMORY_REGION(p,size)
+#define mi_track_mem_noaccess(p,size)             ASAN_POISON_MEMORY_REGION(p,size)
+
+#elif MI_TRACK_ETW
+// windows event tracing
+
+#define MI_TRACK_ENABLED      1
+#define MI_TRACK_HEAP_DESTROY 1
+#define MI_TRACK_TOOL         "ETW"
+
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include "../src/prim/windows/etw.h"
+
+#define mi_track_init()                           EventRegistermicrosoft_windows_mimalloc();
+#define mi_track_malloc_size(p,reqsize,size,zero) EventWriteETW_MI_ALLOC((UINT64)(p), size)
+#define mi_track_free_size(p,size)                EventWriteETW_MI_FREE((UINT64)(p), size)
+
+#else
+// no tracking
+
+#define MI_TRACK_ENABLED      0
+#define MI_TRACK_HEAP_DESTROY 0
+#define MI_TRACK_TOOL         "none"
+
+#define mi_track_malloc_size(p,reqsize,size,zero)
+#define mi_track_free_size(p,_size)
+
+#endif
+
+// -------------------
+// Utility definitions
+
+#ifndef mi_track_resize
+#define mi_track_resize(p,oldsize,newsize)      mi_track_free_size(p,oldsize); mi_track_malloc(p,newsize,false)
+#endif
+
+#ifndef mi_track_align
+#define mi_track_align(p,alignedp,offset,size)  mi_track_mem_noaccess(p,offset)
+#endif
+
+#ifndef mi_track_init
+#define mi_track_init()
+#endif
+
+#ifndef mi_track_mem_defined
+#define mi_track_mem_defined(p,size)
+#endif
+
+#ifndef mi_track_mem_undefined
+#define mi_track_mem_undefined(p,size)
+#endif
+
+#ifndef mi_track_mem_noaccess
+#define mi_track_mem_noaccess(p,size)
+#endif
+
+
+#if MI_PADDING
+#define mi_track_malloc(p,reqsize,zero) \
+  if ((p)!=NULL) { \
+    mi_assert_internal(mi_usable_size(p)==(reqsize)); \
+    mi_track_malloc_size(p,reqsize,reqsize,zero); \
+  }
+#else
+#define mi_track_malloc(p,reqsize,zero) \
+  if ((p)!=NULL) { \
+    mi_assert_internal(mi_usable_size(p)>=(reqsize)); \
+    mi_track_malloc_size(p,reqsize,mi_usable_size(p),zero); \
+  }
+#endif
+
+#endif
diff --git a/compat/mimalloc/mimalloc/types.h b/compat/mimalloc/mimalloc/types.h
new file mode 100644
index 00000000000000..7616f37e4b978f
--- /dev/null
+++ b/compat/mimalloc/mimalloc/types.h
@@ -0,0 +1,670 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_TYPES_H
+#define MIMALLOC_TYPES_H
+
+// --------------------------------------------------------------------------
+// This file contains the main type definitions for mimalloc:
+// mi_heap_t      : all data for a thread-local heap, contains
+//                  lists of all managed heap pages.
+// mi_segment_t   : a larger chunk of memory (32GiB) from where pages
+//                  are allocated.
+// mi_page_t      : a mimalloc page (usually 64KiB or 512KiB) from
+//                  where objects are allocated.
+// --------------------------------------------------------------------------
+
+
+#include <stddef.h>   // ptrdiff_t
+#include <stdint.h>   // uintptr_t, uint16_t, etc
+#include "mimalloc/atomic.h"  // _Atomic
+
+#ifdef _MSC_VER
+#pragma warning(disable:4214) // bitfield is not int
+#endif
+
+// Minimal alignment necessary. On most platforms 16 bytes are needed
+// due to SSE registers for example. This must be at least `sizeof(void*)`
+#ifndef MI_MAX_ALIGN_SIZE
+#define MI_MAX_ALIGN_SIZE  16   // sizeof(max_align_t)
+#endif
+
+// ------------------------------------------------------
+// Variants
+// ------------------------------------------------------
+
+// Define NDEBUG in the release version to disable assertions.
+// #define NDEBUG
+
+// Define MI_TRACK_<tool> to enable tracking support
+// #define MI_TRACK_VALGRIND 1
+// #define MI_TRACK_ASAN     1
+// #define MI_TRACK_ETW      1
+
+// Define MI_STAT as 1 to maintain statistics; set it to 2 to have detailed statistics (but costs some performance).
+// #define MI_STAT 1
+
+// Define MI_SECURE to enable security mitigations
+// #define MI_SECURE 1  // guard page around metadata
+// #define MI_SECURE 2  // guard page around each mimalloc page
+// #define MI_SECURE 3  // encode free lists (detect corrupted free list (buffer overflow), and invalid pointer free)
+// #define MI_SECURE 4  // checks for double free. (may be more expensive)
+
+#if !defined(MI_SECURE)
+#define MI_SECURE 0
+#endif
+
+// Define MI_DEBUG for debug mode
+// #define MI_DEBUG 1  // basic assertion checks and statistics, check double free, corrupted free list, and invalid pointer free.
+// #define MI_DEBUG 2  // + internal assertion checks
+// #define MI_DEBUG 3  // + extensive internal invariant checking (cmake -DMI_DEBUG_FULL=ON)
+#if !defined(MI_DEBUG)
+#if !defined(NDEBUG) || defined(_DEBUG)
+#define MI_DEBUG 2
+#else
+#define MI_DEBUG 0
+#endif
+#endif
+
+// Reserve extra padding at the end of each block to be more resilient against heap block overflows.
+// The padding can detect buffer overflow on free.
+#if !defined(MI_PADDING) && (MI_SECURE>=3 || MI_DEBUG>=1 || (MI_TRACK_VALGRIND || MI_TRACK_ASAN || MI_TRACK_ETW))
+#define MI_PADDING  1
+#endif
+
+// Check padding bytes; allows byte-precise buffer overflow detection
+#if !defined(MI_PADDING_CHECK) && MI_PADDING && (MI_SECURE>=3 || MI_DEBUG>=1)
+#define MI_PADDING_CHECK 1
+#endif
+
+
+// Encoded free lists allow detection of corrupted free lists
+// and can detect buffer overflows, modify after free, and double `free`s.
+#if (MI_SECURE>=3 || MI_DEBUG>=1)
+#define MI_ENCODE_FREELIST  1
+#endif
+
+
+// We used to abandon huge pages but to eagerly deallocate if freed from another thread,
+// but that makes it not possible to visit them during a heap walk or include them in a
+// `mi_heap_destroy`. We therefore instead reset/decommit the huge blocks if freed from
+// another thread so most memory is available until it gets properly freed by the owning thread.
+// #define MI_HUGE_PAGE_ABANDON 1
+
+
+// ------------------------------------------------------
+// Platform specific values
+// ------------------------------------------------------
+
+// ------------------------------------------------------
+// Size of a pointer.
+// We assume that `sizeof(void*)==sizeof(intptr_t)`
+// and it holds for all platforms we know of.
+//
+// However, the C standard only requires that:
+//  p == (void*)((intptr_t)p))
+// but we also need:
+//  i == (intptr_t)((void*)i)
+// or otherwise one might define an intptr_t type that is larger than a pointer...
+// ------------------------------------------------------
+
+#if INTPTR_MAX > INT64_MAX
+# define MI_INTPTR_SHIFT (4)  // assume 128-bit  (as on arm CHERI for example)
+#elif INTPTR_MAX == INT64_MAX
+# define MI_INTPTR_SHIFT (3)
+#elif INTPTR_MAX == INT32_MAX
+# define MI_INTPTR_SHIFT (2)
+#else
+#error platform pointers must be 32, 64, or 128 bits
+#endif
+
+#if SIZE_MAX == UINT64_MAX
+# define MI_SIZE_SHIFT (3)
+typedef int64_t  mi_ssize_t;
+#elif SIZE_MAX == UINT32_MAX
+# define MI_SIZE_SHIFT (2)
+typedef int32_t  mi_ssize_t;
+#else
+#error platform objects must be 32 or 64 bits
+#endif
+
+#if (SIZE_MAX/2) > LONG_MAX
+# define MI_ZU(x)  x##ULL
+# define MI_ZI(x)  x##LL
+#else
+# define MI_ZU(x)  x##UL
+# define MI_ZI(x)  x##L
+#endif
+
+#define MI_INTPTR_SIZE  (1<<MI_INTPTR_SHIFT)
+#define MI_INTPTR_BITS  (MI_INTPTR_SIZE*8)
+
+#define MI_SIZE_SIZE  (1<<MI_SIZE_SHIFT)
+#define MI_SIZE_BITS  (MI_SIZE_SIZE*8)
+
+#define MI_KiB     (MI_ZU(1024))
+#define MI_MiB     (MI_KiB*MI_KiB)
+#define MI_GiB     (MI_MiB*MI_KiB)
+
+
+// ------------------------------------------------------
+// Main internal data-structures
+// ------------------------------------------------------
+
+// Main tuning parameters for segment and page sizes
+// Sizes for 64-bit (usually divide by two for 32-bit)
+#define MI_SEGMENT_SLICE_SHIFT            (13 + MI_INTPTR_SHIFT)         // 64KiB  (32KiB on 32-bit)
+
+#if MI_INTPTR_SIZE > 4
+#define MI_SEGMENT_SHIFT                  ( 9 + MI_SEGMENT_SLICE_SHIFT)  // 32MiB
+#else
+#define MI_SEGMENT_SHIFT                  ( 7 + MI_SEGMENT_SLICE_SHIFT)  // 4MiB on 32-bit
+#endif
+
+#define MI_SMALL_PAGE_SHIFT               (MI_SEGMENT_SLICE_SHIFT)       // 64KiB
+#define MI_MEDIUM_PAGE_SHIFT              ( 3 + MI_SMALL_PAGE_SHIFT)     // 512KiB
+
+
+// Derived constants
+#define MI_SEGMENT_SIZE                   (MI_ZU(1)<<MI_SEGMENT_SHIFT)
+#define MI_SEGMENT_ALIGN                  MI_SEGMENT_SIZE
+#define MI_SEGMENT_MASK                   ((uintptr_t)(MI_SEGMENT_ALIGN - 1))
+#define MI_SEGMENT_SLICE_SIZE             (MI_ZU(1)<< MI_SEGMENT_SLICE_SHIFT)
+#define MI_SLICES_PER_SEGMENT             (MI_SEGMENT_SIZE / MI_SEGMENT_SLICE_SIZE) // 1024
+
+#define MI_SMALL_PAGE_SIZE                (MI_ZU(1)<<MI_SMALL_PAGE_SHIFT)
+#define MI_MEDIUM_PAGE_SIZE               (MI_ZU(1)<<MI_MEDIUM_PAGE_SHIFT)
+
+#define MI_SMALL_OBJ_SIZE_MAX             (MI_SMALL_PAGE_SIZE/4)   // 8KiB on 64-bit
+#define MI_MEDIUM_OBJ_SIZE_MAX            (MI_MEDIUM_PAGE_SIZE/4)  // 128KiB on 64-bit
+#define MI_MEDIUM_OBJ_WSIZE_MAX           (MI_MEDIUM_OBJ_SIZE_MAX/MI_INTPTR_SIZE)
+#define MI_LARGE_OBJ_SIZE_MAX             (MI_SEGMENT_SIZE/2)      // 32MiB on 64-bit
+#define MI_LARGE_OBJ_WSIZE_MAX            (MI_LARGE_OBJ_SIZE_MAX/MI_INTPTR_SIZE)
+
+// Maximum number of size classes. (spaced exponentially in 12.5% increments)
+#define MI_BIN_HUGE  (73U)
+
+#if (MI_MEDIUM_OBJ_WSIZE_MAX >= 655360)
+#error "mimalloc internal: define more bins"
+#endif
+
+// Maximum slice offset (15)
+#define MI_MAX_SLICE_OFFSET               ((MI_ALIGNMENT_MAX / MI_SEGMENT_SLICE_SIZE) - 1)
+
+// Used as a special value to encode block sizes in 32 bits.
+#define MI_HUGE_BLOCK_SIZE                ((uint32_t)(2*MI_GiB))
+
+// blocks up to this size are always allocated aligned
+#define MI_MAX_ALIGN_GUARANTEE            (8*MI_MAX_ALIGN_SIZE)
+
+// Alignments over MI_ALIGNMENT_MAX are allocated in dedicated huge page segments
+#define MI_ALIGNMENT_MAX                  (MI_SEGMENT_SIZE >> 1)
+
+
+// ------------------------------------------------------
+// Mimalloc pages contain allocated blocks
+// ------------------------------------------------------
+
+// The free lists use encoded next fields
+// (Only actually encodes when MI_ENCODED_FREELIST is defined.)
+typedef uintptr_t  mi_encoded_t;
+
+// thread id's
+typedef size_t     mi_threadid_t;
+
+// free lists contain blocks
+typedef struct mi_block_s {
+  mi_encoded_t next;
+} mi_block_t;
+
+
+// The delayed flags are used for efficient multi-threaded free-ing
+typedef enum mi_delayed_e {
+  MI_USE_DELAYED_FREE   = 0, // push on the owning heap thread delayed list
+  MI_DELAYED_FREEING    = 1, // temporary: another thread is accessing the owning heap
+  MI_NO_DELAYED_FREE    = 2, // optimize: push on page local thread free queue if another block is already in the heap thread delayed free list
+  MI_NEVER_DELAYED_FREE = 3  // sticky, only resets on page reclaim
+} mi_delayed_t;
+
+
+// The `in_full` and `has_aligned` page flags are put in a union to efficiently
+// test if both are false (`full_aligned == 0`) in the `mi_free` routine.
+#if !MI_TSAN
+typedef union mi_page_flags_s {
+  uint8_t full_aligned;
+  struct {
+    uint8_t in_full : 1;
+    uint8_t has_aligned : 1;
+  } x;
+} mi_page_flags_t;
+#else
+// under thread sanitizer, use a byte for each flag to suppress warning, issue #130
+typedef union mi_page_flags_s {
+  uint16_t full_aligned;
+  struct {
+    uint8_t in_full;
+    uint8_t has_aligned;
+  } x;
+} mi_page_flags_t;
+#endif
+
+// Thread free list.
+// We use the bottom 2 bits of the pointer for mi_delayed_t flags
+typedef uintptr_t mi_thread_free_t;
+
+// A page contains blocks of one specific size (`block_size`).
+// Each page has three list of free blocks:
+// `free` for blocks that can be allocated,
+// `local_free` for freed blocks that are not yet available to `mi_malloc`
+// `thread_free` for freed blocks by other threads
+// The `local_free` and `thread_free` lists are migrated to the `free` list
+// when it is exhausted. The separate `local_free` list is necessary to
+// implement a monotonic heartbeat. The `thread_free` list is needed for
+// avoiding atomic operations in the common case.
+//
+//
+// `used - |thread_free|` == actual blocks that are in use (alive)
+// `used - |thread_free| + |free| + |local_free| == capacity`
+//
+// We don't count `freed` (as |free|) but use `used` to reduce
+// the number of memory accesses in the `mi_page_all_free` function(s).
+//
+// Notes:
+// - Access is optimized for `mi_free` and `mi_page_alloc` (in `alloc.c`)
+// - Using `uint16_t` does not seem to slow things down
+// - The size is 8 words on 64-bit which helps the page index calculations
+//   (and 10 words on 32-bit, and encoded free lists add 2 words. Sizes 10
+//    and 12 are still good for address calculation)
+// - To limit the structure size, the `xblock_size` is 32-bits only; for
+//   blocks > MI_HUGE_BLOCK_SIZE the size is determined from the segment page size
+// - `thread_free` uses the bottom bits as a delayed-free flags to optimize
+//   concurrent frees where only the first concurrent free adds to the owning
+//   heap `thread_delayed_free` list (see `alloc.c:mi_free_block_mt`).
+//   The invariant is that no-delayed-free is only set if there is
+//   at least one block that will be added, or as already been added, to
+//   the owning heap `thread_delayed_free` list. This guarantees that pages
+//   will be freed correctly even if only other threads free blocks.
+typedef struct mi_page_s {
+  // "owned" by the segment
+  uint32_t              slice_count;       // slices in this page (0 if not a page)
+  uint32_t              slice_offset;      // distance from the actual page data slice (0 if a page)
+  uint8_t               is_committed : 1;  // `true` if the page virtual memory is committed
+  uint8_t               is_zero_init : 1;  // `true` if the page was initially zero initialized
+
+  // layout like this to optimize access in `mi_malloc` and `mi_free`
+  uint16_t              capacity;          // number of blocks committed, must be the first field, see `segment.c:page_clear`
+  uint16_t              reserved;          // number of blocks reserved in memory
+  mi_page_flags_t       flags;             // `in_full` and `has_aligned` flags (8 bits)
+  uint8_t               free_is_zero : 1;  // `true` if the blocks in the free list are zero initialized
+  uint8_t               retire_expire : 7; // expiration count for retired blocks
+
+  mi_block_t*           free;              // list of available free blocks (`malloc` allocates from this list)
+  uint32_t              used;              // number of blocks in use (including blocks in `local_free` and `thread_free`)
+  uint32_t              xblock_size;       // size available in each block (always `>0`)
+  mi_block_t*           local_free;        // list of deferred free blocks by this thread (migrates to `free`)
+
+  #if (MI_ENCODE_FREELIST || MI_PADDING)
+  uintptr_t             keys[2];           // two random keys to encode the free lists (see `_mi_block_next`) or padding canary
+  #endif
+
+  _Atomic(mi_thread_free_t) xthread_free;  // list of deferred free blocks freed by other threads
+  _Atomic(uintptr_t)        xheap;
+
+  struct mi_page_s*     next;              // next page owned by this thread with the same `block_size`
+  struct mi_page_s*     prev;              // previous page owned by this thread with the same `block_size`
+
+  // 64-bit 9 words, 32-bit 12 words, (+2 for secure)
+  #if MI_INTPTR_SIZE==8
+  uintptr_t padding[1];
+  #endif
+} mi_page_t;
+
+
+
+// ------------------------------------------------------
+// Mimalloc segments contain mimalloc pages
+// ------------------------------------------------------
+
+typedef enum mi_page_kind_e {
+  MI_PAGE_SMALL,    // small blocks go into 64KiB pages inside a segment
+  MI_PAGE_MEDIUM,   // medium blocks go into medium pages inside a segment
+  MI_PAGE_LARGE,    // larger blocks go into a page of just one block
+  MI_PAGE_HUGE,     // huge blocks (> 16 MiB) are put into a single page in a single segment.
+} mi_page_kind_t;
+
+typedef enum mi_segment_kind_e {
+  MI_SEGMENT_NORMAL, // MI_SEGMENT_SIZE size with pages inside.
+  MI_SEGMENT_HUGE,   // > MI_LARGE_SIZE_MAX segment with just one huge page inside.
+} mi_segment_kind_t;
+
+// ------------------------------------------------------
+// A segment holds a commit mask where a bit is set if
+// the corresponding MI_COMMIT_SIZE area is committed.
+// The MI_COMMIT_SIZE must be a multiple of the slice
+// size. If it is equal we have the most fine grained
+// decommit (but setting it higher can be more efficient).
+// The MI_MINIMAL_COMMIT_SIZE is the minimal amount that will
+// be committed in one go which can be set higher than
+// MI_COMMIT_SIZE for efficiency (while the decommit mask
+// is still tracked in fine-grained MI_COMMIT_SIZE chunks)
+// ------------------------------------------------------
+
+#define MI_MINIMAL_COMMIT_SIZE      (1*MI_SEGMENT_SLICE_SIZE)
+#define MI_COMMIT_SIZE              (MI_SEGMENT_SLICE_SIZE)              // 64KiB
+#define MI_COMMIT_MASK_BITS         (MI_SEGMENT_SIZE / MI_COMMIT_SIZE)
+#define MI_COMMIT_MASK_FIELD_BITS    MI_SIZE_BITS
+#define MI_COMMIT_MASK_FIELD_COUNT  (MI_COMMIT_MASK_BITS / MI_COMMIT_MASK_FIELD_BITS)
+
+#if (MI_COMMIT_MASK_BITS != (MI_COMMIT_MASK_FIELD_COUNT * MI_COMMIT_MASK_FIELD_BITS))
+#error "the segment size must be exactly divisible by the (commit size * size_t bits)"
+#endif
+
+typedef struct mi_commit_mask_s {
+  size_t mask[MI_COMMIT_MASK_FIELD_COUNT];
+} mi_commit_mask_t;
+
+typedef mi_page_t  mi_slice_t;
+typedef int64_t    mi_msecs_t;
+
+
+// Memory can reside in arena's, direct OS allocated, or statically allocated. The memid keeps track of this.
+typedef enum mi_memkind_e {
+  MI_MEM_NONE,      // not allocated
+  MI_MEM_EXTERNAL,  // not owned by mimalloc but provided externally (via `mi_manage_os_memory` for example)
+  MI_MEM_STATIC,    // allocated in a static area and should not be freed (for arena meta data for example)
+  MI_MEM_OS,        // allocated from the OS
+  MI_MEM_OS_HUGE,   // allocated as huge os pages
+  MI_MEM_OS_REMAP,  // allocated in a remapable area (i.e. using `mremap`)
+  MI_MEM_ARENA      // allocated from an arena (the usual case)
+} mi_memkind_t;
+
+static inline bool mi_memkind_is_os(mi_memkind_t memkind) {
+  return (memkind >= MI_MEM_OS && memkind <= MI_MEM_OS_REMAP);
+}
+
+typedef struct mi_memid_os_info {
+  void*         base;               // actual base address of the block (used for offset aligned allocations)
+  size_t        alignment;          // alignment at allocation
+} mi_memid_os_info_t;
+
+typedef struct mi_memid_arena_info {
+  size_t        block_index;        // index in the arena
+  mi_arena_id_t id;                 // arena id (>= 1)
+  bool          is_exclusive;       // the arena can only be used for specific arena allocations
+} mi_memid_arena_info_t;
+
+typedef struct mi_memid_s {
+  union {
+    mi_memid_os_info_t    os;       // only used for MI_MEM_OS
+    mi_memid_arena_info_t arena;    // only used for MI_MEM_ARENA
+  } mem;
+  bool          is_pinned;          // `true` if we cannot decommit/reset/protect in this memory (e.g. when allocated using large OS pages)
+  bool          initially_committed;// `true` if the memory was originally allocated as committed
+  bool          initially_zero;     // `true` if the memory was originally zero initialized
+  mi_memkind_t  memkind;
+} mi_memid_t;
+
+
+// Segments are large allocated memory blocks (8mb on 64 bit) from
+// the OS. Inside segments we allocated fixed size _pages_ that
+// contain blocks.
+typedef struct mi_segment_s {
+  // constant fields
+  mi_memid_t        memid;              // memory id for arena allocation
+  bool              allow_decommit;
+  bool              allow_purge;
+  size_t            segment_size;
+
+  // segment fields
+  mi_msecs_t        purge_expire;
+  mi_commit_mask_t  purge_mask;
+  mi_commit_mask_t  commit_mask;
+
+  _Atomic(struct mi_segment_s*) abandoned_next;
+
+  // from here is zero initialized
+  struct mi_segment_s* next;            // the list of freed segments in the cache (must be first field, see `segment.c:mi_segment_init`)
+
+  size_t            abandoned;          // abandoned pages (i.e. the original owning thread stopped) (`abandoned <= used`)
+  size_t            abandoned_visits;   // count how often this segment is visited in the abandoned list (to force reclaim it it is too long)
+  size_t            used;               // count of pages in use
+  uintptr_t         cookie;             // verify addresses in debug mode: `mi_ptr_cookie(segment) == segment->cookie`
+
+  size_t            segment_slices;      // for huge segments this may be different from `MI_SLICES_PER_SEGMENT`
+  size_t            segment_info_slices; // initial slices we are using segment info and possible guard pages.
+
+  // layout like this to optimize access in `mi_free`
+  mi_segment_kind_t kind;
+  size_t            slice_entries;       // entries in the `slices` array, at most `MI_SLICES_PER_SEGMENT`
+  _Atomic(mi_threadid_t) thread_id;      // unique id of the thread owning this segment
+
+  mi_slice_t        slices[MI_SLICES_PER_SEGMENT+1];  // one more for huge blocks with large alignment
+} mi_segment_t;
+
+
+// ------------------------------------------------------
+// Heaps
+// Provide first-class heaps to allocate from.
+// A heap just owns a set of pages for allocation and
+// can only be allocate/reallocate from the thread that created it.
+// Freeing blocks can be done from any thread though.
+// Per thread, the segments are shared among its heaps.
+// Per thread, there is always a default heap that is
+// used for allocation; it is initialized to statically
+// point to an empty heap to avoid initialization checks
+// in the fast path.
+// ------------------------------------------------------
+
+// Thread local data
+typedef struct mi_tld_s mi_tld_t;
+
+// Pages of a certain block size are held in a queue.
+typedef struct mi_page_queue_s {
+  mi_page_t* first;
+  mi_page_t* last;
+  size_t     block_size;
+} mi_page_queue_t;
+
+#define MI_BIN_FULL  (MI_BIN_HUGE+1)
+
+// Random context
+typedef struct mi_random_cxt_s {
+  uint32_t input[16];
+  uint32_t output[16];
+  int      output_available;
+  bool     weak;
+} mi_random_ctx_t;
+
+
+// In debug mode there is a padding structure at the end of the blocks to check for buffer overflows
+#if (MI_PADDING)
+typedef struct mi_padding_s {
+  uint32_t canary; // encoded block value to check validity of the padding (in case of overflow)
+  uint32_t delta;  // padding bytes before the block. (mi_usable_size(p) - delta == exact allocated bytes)
+} mi_padding_t;
+#define MI_PADDING_SIZE   (sizeof(mi_padding_t))
+#define MI_PADDING_WSIZE  ((MI_PADDING_SIZE + MI_INTPTR_SIZE - 1) / MI_INTPTR_SIZE)
+#else
+#define MI_PADDING_SIZE   0
+#define MI_PADDING_WSIZE  0
+#endif
+
+#define MI_PAGES_DIRECT   (MI_SMALL_WSIZE_MAX + MI_PADDING_WSIZE + 1)
+
+
+// A heap owns a set of pages.
+struct mi_heap_s {
+  mi_tld_t*             tld;
+  mi_page_t*            pages_free_direct[MI_PAGES_DIRECT];  // optimize: array where every entry points a page with possibly free blocks in the corresponding queue for that size.
+  mi_page_queue_t       pages[MI_BIN_FULL + 1];              // queue of pages for each size class (or "bin")
+  _Atomic(mi_block_t*)  thread_delayed_free;
+  mi_threadid_t         thread_id;                           // thread this heap belongs too
+  mi_arena_id_t         arena_id;                            // arena id if the heap belongs to a specific arena (or 0)
+  uintptr_t             cookie;                              // random cookie to verify pointers (see `_mi_ptr_cookie`)
+  uintptr_t             keys[2];                             // two random keys used to encode the `thread_delayed_free` list
+  mi_random_ctx_t       random;                              // random number context used for secure allocation
+  size_t                page_count;                          // total number of pages in the `pages` queues.
+  size_t                page_retired_min;                    // smallest retired index (retired pages are fully free, but still in the page queues)
+  size_t                page_retired_max;                    // largest retired index into the `pages` array.
+  mi_heap_t*            next;                                // list of heaps per thread
+  bool                  no_reclaim;                          // `true` if this heap should not reclaim abandoned pages
+};
+
+
+
+// ------------------------------------------------------
+// Debug
+// ------------------------------------------------------
+
+#if !defined(MI_DEBUG_UNINIT)
+#define MI_DEBUG_UNINIT     (0xD0)
+#endif
+#if !defined(MI_DEBUG_FREED)
+#define MI_DEBUG_FREED      (0xDF)
+#endif
+#if !defined(MI_DEBUG_PADDING)
+#define MI_DEBUG_PADDING    (0xDE)
+#endif
+
+#if (MI_DEBUG)
+// use our own assertion to print without memory allocation
+void _mi_assert_fail(const char* assertion, const char* fname, unsigned int line, const char* func );
+#define mi_assert(expr)     ((expr) ? (void)0 : _mi_assert_fail(#expr,__FILE__,__LINE__,__func__))
+#else
+#define mi_assert(x)
+#endif
+
+#if (MI_DEBUG>1)
+#define mi_assert_internal    mi_assert
+#else
+#define mi_assert_internal(x)
+#endif
+
+#if (MI_DEBUG>2)
+#define mi_assert_expensive   mi_assert
+#else
+#define mi_assert_expensive(x)
+#endif
+
+// ------------------------------------------------------
+// Statistics
+// ------------------------------------------------------
+
+#ifndef MI_STAT
+#if (MI_DEBUG>0)
+#define MI_STAT 2
+#else
+#define MI_STAT 0
+#endif
+#endif
+
+typedef struct mi_stat_count_s {
+  int64_t allocated;
+  int64_t freed;
+  int64_t peak;
+  int64_t current;
+} mi_stat_count_t;
+
+typedef struct mi_stat_counter_s {
+  int64_t total;
+  int64_t count;
+} mi_stat_counter_t;
+
+typedef struct mi_stats_s {
+  mi_stat_count_t segments;
+  mi_stat_count_t pages;
+  mi_stat_count_t reserved;
+  mi_stat_count_t committed;
+  mi_stat_count_t reset;
+  mi_stat_count_t purged;
+  mi_stat_count_t page_committed;
+  mi_stat_count_t segments_abandoned;
+  mi_stat_count_t pages_abandoned;
+  mi_stat_count_t threads;
+  mi_stat_count_t normal;
+  mi_stat_count_t huge;
+  mi_stat_count_t large;
+  mi_stat_count_t malloc;
+  mi_stat_count_t segments_cache;
+  mi_stat_counter_t pages_extended;
+  mi_stat_counter_t mmap_calls;
+  mi_stat_counter_t commit_calls;
+  mi_stat_counter_t reset_calls;
+  mi_stat_counter_t purge_calls;
+  mi_stat_counter_t page_no_retire;
+  mi_stat_counter_t searches;
+  mi_stat_counter_t normal_count;
+  mi_stat_counter_t huge_count;
+  mi_stat_counter_t large_count;
+#if MI_STAT>1
+  mi_stat_count_t normal_bins[MI_BIN_HUGE+1];
+#endif
+} mi_stats_t;
+
+
+void _mi_stat_increase(mi_stat_count_t* stat, size_t amount);
+void _mi_stat_decrease(mi_stat_count_t* stat, size_t amount);
+void _mi_stat_counter_increase(mi_stat_counter_t* stat, size_t amount);
+
+#if (MI_STAT)
+#define mi_stat_increase(stat,amount)         _mi_stat_increase( &(stat), amount)
+#define mi_stat_decrease(stat,amount)         _mi_stat_decrease( &(stat), amount)
+#define mi_stat_counter_increase(stat,amount) _mi_stat_counter_increase( &(stat), amount)
+#else
+#define mi_stat_increase(stat,amount)         (void)0
+#define mi_stat_decrease(stat,amount)         (void)0
+#define mi_stat_counter_increase(stat,amount) (void)0
+#endif
+
+#define mi_heap_stat_counter_increase(heap,stat,amount)  mi_stat_counter_increase( (heap)->tld->stats.stat, amount)
+#define mi_heap_stat_increase(heap,stat,amount)  mi_stat_increase( (heap)->tld->stats.stat, amount)
+#define mi_heap_stat_decrease(heap,stat,amount)  mi_stat_decrease( (heap)->tld->stats.stat, amount)
+
+// ------------------------------------------------------
+// Thread Local data
+// ------------------------------------------------------
+
+// A "span" is is an available range of slices. The span queues keep
+// track of slice spans of at most the given `slice_count` (but more than the previous size class).
+typedef struct mi_span_queue_s {
+  mi_slice_t* first;
+  mi_slice_t* last;
+  size_t      slice_count;
+} mi_span_queue_t;
+
+#define MI_SEGMENT_BIN_MAX (35)     // 35 == mi_segment_bin(MI_SLICES_PER_SEGMENT)
+
+// OS thread local data
+typedef struct mi_os_tld_s {
+  size_t                region_idx;   // start point for next allocation
+  mi_stats_t*           stats;        // points to tld stats
+} mi_os_tld_t;
+
+
+// Segments thread local data
+typedef struct mi_segments_tld_s {
+  mi_span_queue_t     spans[MI_SEGMENT_BIN_MAX+1];  // free slice spans inside segments
+  size_t              count;        // current number of segments;
+  size_t              peak_count;   // peak number of segments
+  size_t              current_size; // current size of all segments
+  size_t              peak_size;    // peak size of all segments
+  mi_stats_t*         stats;        // points to tld stats
+  mi_os_tld_t*        os;           // points to os stats
+} mi_segments_tld_t;
+
+// Thread local data
+struct mi_tld_s {
+  unsigned long long  heartbeat;     // monotonic heartbeat count
+  bool                recurse;       // true if deferred was called; used to prevent infinite recursion.
+  mi_heap_t*          heap_backing;  // backing heap of this thread (cannot be deleted)
+  mi_heap_t*          heaps;         // list of heaps in this thread (so we can abandon all when the thread terminates)
+  mi_segments_tld_t   segments;      // segment tld
+  mi_os_tld_t         os;            // os tld
+  mi_stats_t          stats;         // statistics
+};
+
+#endif
diff --git a/compat/mimalloc/options.c b/compat/mimalloc/options.c
new file mode 100644
index 00000000000000..3a3090d9acfc94
--- /dev/null
+++ b/compat/mimalloc/options.c
@@ -0,0 +1,571 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"  // mi_prim_out_stderr
+
+#include <stdio.h>      // FILE
+#include <stdlib.h>     // abort
+#include <stdarg.h>
+
+
+static long mi_max_error_count   = 16; // stop outputting errors after this (use < 0 for no limit)
+static long mi_max_warning_count = 16; // stop outputting warnings after this (use < 0 for no limit)
+
+static void mi_add_stderr_output(void);
+
+int mi_version(void) mi_attr_noexcept {
+  return MI_MALLOC_VERSION;
+}
+
+
+// --------------------------------------------------------
+// Options
+// These can be accessed by multiple threads and may be
+// concurrently initialized, but an initializing data race
+// is ok since they resolve to the same value.
+// --------------------------------------------------------
+typedef enum mi_init_e {
+  UNINIT,       // not yet initialized
+  DEFAULTED,    // not found in the environment, use default value
+  INITIALIZED   // found in environment or set explicitly
+} mi_init_t;
+
+typedef struct mi_option_desc_s {
+  long        value;  // the value
+  mi_init_t   init;   // is it initialized yet? (from the environment)
+  mi_option_t option; // for debugging: the option index should match the option
+  const char* name;   // option name without `mimalloc_` prefix
+  const char* legacy_name; // potential legacy option name
+} mi_option_desc_t;
+
+#define MI_OPTION(opt)                  mi_option_##opt, #opt, NULL
+#define MI_OPTION_LEGACY(opt,legacy)    mi_option_##opt, #opt, #legacy
+
+static mi_option_desc_t options[_mi_option_last] =
+{
+  // stable options
+  #if MI_DEBUG || defined(MI_SHOW_ERRORS)
+  { 1, UNINIT, MI_OPTION(show_errors) },
+  #else
+  { 0, UNINIT, MI_OPTION(show_errors) },
+  #endif
+  { 0, UNINIT, MI_OPTION(show_stats) },
+  { 0, UNINIT, MI_OPTION(verbose) },
+
+  // the following options are experimental and not all combinations make sense.
+  { 1, UNINIT, MI_OPTION(eager_commit) },               // commit per segment directly (4MiB)  (but see also `eager_commit_delay`)
+  { 2, UNINIT, MI_OPTION_LEGACY(arena_eager_commit,eager_region_commit) }, // eager commit arena's? 2 is used to enable this only on an OS that has overcommit (i.e. linux)
+  { 1, UNINIT, MI_OPTION_LEGACY(purge_decommits,reset_decommits) },        // purge decommits memory (instead of reset) (note: on linux this uses MADV_DONTNEED for decommit)
+  { 0, UNINIT, MI_OPTION_LEGACY(allow_large_os_pages,large_os_pages) },    // use large OS pages, use only with eager commit to prevent fragmentation of VMA's
+  { 0, UNINIT, MI_OPTION(reserve_huge_os_pages) },      // per 1GiB huge pages
+  {-1, UNINIT, MI_OPTION(reserve_huge_os_pages_at) },   // reserve huge pages at node N
+  { 0, UNINIT, MI_OPTION(reserve_os_memory)     },
+  { 0, UNINIT, MI_OPTION(deprecated_segment_cache) },   // cache N segments per thread
+  { 0, UNINIT, MI_OPTION(deprecated_page_reset) },      // reset page memory on free
+  { 0, UNINIT, MI_OPTION_LEGACY(abandoned_page_purge,abandoned_page_reset) },       // reset free page memory when a thread terminates
+  { 0, UNINIT, MI_OPTION(deprecated_segment_reset) },   // reset segment memory on free (needs eager commit)
+#if defined(__NetBSD__)
+  { 0, UNINIT, MI_OPTION(eager_commit_delay) },         // the first N segments per thread are not eagerly committed
+#else
+  { 1, UNINIT, MI_OPTION(eager_commit_delay) },         // the first N segments per thread are not eagerly committed (but per page in the segment on demand)
+#endif
+  { 10,  UNINIT, MI_OPTION_LEGACY(purge_delay,reset_delay) },  // purge delay in milli-seconds
+  { 0,   UNINIT, MI_OPTION(use_numa_nodes) },           // 0 = use available numa nodes, otherwise use at most N nodes.
+  { 0,   UNINIT, MI_OPTION(limit_os_alloc) },           // 1 = do not use OS memory for allocation (but only reserved arenas)
+  { 100, UNINIT, MI_OPTION(os_tag) },                   // only apple specific for now but might serve more or less related purpose
+  { 16,  UNINIT, MI_OPTION(max_errors) },               // maximum errors that are output
+  { 16,  UNINIT, MI_OPTION(max_warnings) },             // maximum warnings that are output
+  { 8,   UNINIT, MI_OPTION(max_segment_reclaim)},       // max. number of segment reclaims from the abandoned segments per try.
+  { 0,   UNINIT, MI_OPTION(destroy_on_exit)},           // release all OS memory on process exit; careful with dangling pointer or after-exit frees!
+  #if (MI_INTPTR_SIZE>4)
+  { 1024L * 1024L, UNINIT, MI_OPTION(arena_reserve) },  // reserve memory N KiB at a time
+  #else
+  {  128L * 1024L, UNINIT, MI_OPTION(arena_reserve) },
+  #endif
+  { 10,  UNINIT, MI_OPTION(arena_purge_mult) },        // purge delay multiplier for arena's
+  { 1,   UNINIT, MI_OPTION_LEGACY(purge_extend_delay, decommit_extend_delay) },
+};
+
+static void mi_option_init(mi_option_desc_t* desc);
+
+void _mi_options_init(void) {
+  // called on process load; should not be called before the CRT is initialized!
+  // (e.g. do not call this from process_init as that may run before CRT initialization)
+  mi_add_stderr_output(); // now it safe to use stderr for output
+  for(int i = 0; i < _mi_option_last; i++ ) {
+    mi_option_t option = (mi_option_t)i;
+    long l = mi_option_get(option); MI_UNUSED(l); // initialize
+    // if (option != mi_option_verbose)
+    {
+      mi_option_desc_t* desc = &options[option];
+      _mi_verbose_message("option '%s': %ld\n", desc->name, desc->value);
+    }
+  }
+  mi_max_error_count = mi_option_get(mi_option_max_errors);
+  mi_max_warning_count = mi_option_get(mi_option_max_warnings);
+}
+
+mi_decl_nodiscard long mi_option_get(mi_option_t option) {
+  mi_assert(option >= 0 && option < _mi_option_last);
+  if (option < 0 || option >= _mi_option_last) return 0;
+  mi_option_desc_t* desc = &options[option];
+  mi_assert(desc->option == option);  // index should match the option
+  if mi_unlikely(desc->init == UNINIT) {
+    mi_option_init(desc);
+  }
+  return desc->value;
+}
+
+mi_decl_nodiscard long mi_option_get_clamp(mi_option_t option, long min, long max) {
+  long x = mi_option_get(option);
+  return (x < min ? min : (x > max ? max : x));
+}
+
+mi_decl_nodiscard size_t mi_option_get_size(mi_option_t option) {
+  mi_assert_internal(option == mi_option_reserve_os_memory || option == mi_option_arena_reserve);
+  long x = mi_option_get(option);
+  return (x < 0 ? 0 : (size_t)x * MI_KiB);
+}
+
+void mi_option_set(mi_option_t option, long value) {
+  mi_assert(option >= 0 && option < _mi_option_last);
+  if (option < 0 || option >= _mi_option_last) return;
+  mi_option_desc_t* desc = &options[option];
+  mi_assert(desc->option == option);  // index should match the option
+  desc->value = value;
+  desc->init = INITIALIZED;
+}
+
+void mi_option_set_default(mi_option_t option, long value) {
+  mi_assert(option >= 0 && option < _mi_option_last);
+  if (option < 0 || option >= _mi_option_last) return;
+  mi_option_desc_t* desc = &options[option];
+  if (desc->init != INITIALIZED) {
+    desc->value = value;
+  }
+}
+
+mi_decl_nodiscard bool mi_option_is_enabled(mi_option_t option) {
+  return (mi_option_get(option) != 0);
+}
+
+void mi_option_set_enabled(mi_option_t option, bool enable) {
+  mi_option_set(option, (enable ? 1 : 0));
+}
+
+void mi_option_set_enabled_default(mi_option_t option, bool enable) {
+  mi_option_set_default(option, (enable ? 1 : 0));
+}
+
+void mi_option_enable(mi_option_t option) {
+  mi_option_set_enabled(option,true);
+}
+
+void mi_option_disable(mi_option_t option) {
+  mi_option_set_enabled(option,false);
+}
+
+static void mi_cdecl mi_out_stderr(const char* msg, void* arg) {
+  MI_UNUSED(arg);
+  if (msg != NULL && msg[0] != 0) {
+    _mi_prim_out_stderr(msg);
+  }
+}
+
+// Since an output function can be registered earliest in the `main`
+// function we also buffer output that happens earlier. When
+// an output function is registered it is called immediately with
+// the output up to that point.
+#ifndef MI_MAX_DELAY_OUTPUT
+#define MI_MAX_DELAY_OUTPUT ((size_t)(32*1024))
+#endif
+static char out_buf[MI_MAX_DELAY_OUTPUT+1];
+static _Atomic(size_t) out_len;
+
+static void mi_cdecl mi_out_buf(const char* msg, void* arg) {
+  MI_UNUSED(arg);
+  if (msg==NULL) return;
+  if (mi_atomic_load_relaxed(&out_len)>=MI_MAX_DELAY_OUTPUT) return;
+  size_t n = _mi_strlen(msg);
+  if (n==0) return;
+  // claim space
+  size_t start = mi_atomic_add_acq_rel(&out_len, n);
+  if (start >= MI_MAX_DELAY_OUTPUT) return;
+  // check bound
+  if (start+n >= MI_MAX_DELAY_OUTPUT) {
+    n = MI_MAX_DELAY_OUTPUT-start-1;
+  }
+  _mi_memcpy(&out_buf[start], msg, n);
+}
+
+static void mi_out_buf_flush(mi_output_fun* out, bool no_more_buf, void* arg) {
+  if (out==NULL) return;
+  // claim (if `no_more_buf == true`, no more output will be added after this point)
+  size_t count = mi_atomic_add_acq_rel(&out_len, (no_more_buf ? MI_MAX_DELAY_OUTPUT : 1));
+  // and output the current contents
+  if (count>MI_MAX_DELAY_OUTPUT) count = MI_MAX_DELAY_OUTPUT;
+  out_buf[count] = 0;
+  out(out_buf,arg);
+  if (!no_more_buf) {
+    out_buf[count] = '\n'; // if continue with the buffer, insert a newline
+  }
+}
+
+
+// Once this module is loaded, switch to this routine
+// which outputs to stderr and the delayed output buffer.
+static void mi_cdecl mi_out_buf_stderr(const char* msg, void* arg) {
+  mi_out_stderr(msg,arg);
+  mi_out_buf(msg,arg);
+}
+
+
+
+// --------------------------------------------------------
+// Default output handler
+// --------------------------------------------------------
+
+// Should be atomic but gives errors on many platforms as generally we cannot cast a function pointer to a uintptr_t.
+// For now, don't register output from multiple threads.
+static mi_output_fun* volatile mi_out_default; // = NULL
+static _Atomic(void*) mi_out_arg; // = NULL
+
+static mi_output_fun* mi_out_get_default(void** parg) {
+  if (parg != NULL) { *parg = mi_atomic_load_ptr_acquire(void,&mi_out_arg); }
+  mi_output_fun* out = mi_out_default;
+  return (out == NULL ? &mi_out_buf : out);
+}
+
+void mi_register_output(mi_output_fun* out, void* arg) mi_attr_noexcept {
+  mi_out_default = (out == NULL ? &mi_out_stderr : out); // stop using the delayed output buffer
+  mi_atomic_store_ptr_release(void,&mi_out_arg, arg);
+  if (out!=NULL) mi_out_buf_flush(out,true,arg);         // output all the delayed output now
+}
+
+// add stderr to the delayed output after the module is loaded
+static void mi_add_stderr_output(void) {
+  mi_assert_internal(mi_out_default == NULL);
+  mi_out_buf_flush(&mi_out_stderr, false, NULL); // flush current contents to stderr
+  mi_out_default = &mi_out_buf_stderr;           // and add stderr to the delayed output
+}
+
+// --------------------------------------------------------
+// Messages, all end up calling `_mi_fputs`.
+// --------------------------------------------------------
+static _Atomic(size_t) error_count;   // = 0;  // when >= max_error_count stop emitting errors
+static _Atomic(size_t) warning_count; // = 0;  // when >= max_warning_count stop emitting warnings
+
+// When overriding malloc, we may recurse into mi_vfprintf if an allocation
+// inside the C runtime causes another message.
+// In some cases (like on macOS) the loader already allocates which
+// calls into mimalloc; if we then access thread locals (like `recurse`)
+// this may crash as the access may call _tlv_bootstrap that tries to
+// (recursively) invoke malloc again to allocate space for the thread local
+// variables on demand. This is why we use a _mi_preloading test on such
+// platforms. However, C code generator may move the initial thread local address
+// load before the `if` and we therefore split it out in a separate funcion.
+static mi_decl_thread bool recurse = false;
+
+static mi_decl_noinline bool mi_recurse_enter_prim(void) {
+  if (recurse) return false;
+  recurse = true;
+  return true;
+}
+
+static mi_decl_noinline void mi_recurse_exit_prim(void) {
+  recurse = false;
+}
+
+static bool mi_recurse_enter(void) {
+  #if defined(__APPLE__) || defined(MI_TLS_RECURSE_GUARD)
+  if (_mi_preloading()) return false;
+  #endif
+  return mi_recurse_enter_prim();
+}
+
+static void mi_recurse_exit(void) {
+  #if defined(__APPLE__) || defined(MI_TLS_RECURSE_GUARD)
+  if (_mi_preloading()) return;
+  #endif
+  mi_recurse_exit_prim();
+}
+
+void _mi_fputs(mi_output_fun* out, void* arg, const char* prefix, const char* message) {
+  if (out==NULL || (void*)out==(void*)stdout || (void*)out==(void*)stderr) { // TODO: use mi_out_stderr for stderr?
+    if (!mi_recurse_enter()) return;
+    out = mi_out_get_default(&arg);
+    if (prefix != NULL) out(prefix, arg);
+    out(message, arg);
+    mi_recurse_exit();
+  }
+  else {
+    if (prefix != NULL) out(prefix, arg);
+    out(message, arg);
+  }
+}
+
+// Define our own limited `fprintf` that avoids memory allocation.
+// We do this using `snprintf` with a limited buffer.
+static void mi_vfprintf( mi_output_fun* out, void* arg, const char* prefix, const char* fmt, va_list args ) {
+  char buf[512];
+  if (fmt==NULL) return;
+  if (!mi_recurse_enter()) return;
+  vsnprintf(buf,sizeof(buf)-1,fmt,args);
+  mi_recurse_exit();
+  _mi_fputs(out,arg,prefix,buf);
+}
+
+void _mi_fprintf( mi_output_fun* out, void* arg, const char* fmt, ... ) {
+  va_list args;
+  va_start(args,fmt);
+  mi_vfprintf(out,arg,NULL,fmt,args);
+  va_end(args);
+}
+
+static void mi_vfprintf_thread(mi_output_fun* out, void* arg, const char* prefix, const char* fmt, va_list args) {
+  if (prefix != NULL && _mi_strnlen(prefix,33) <= 32 && !_mi_is_main_thread()) {
+    char tprefix[64];
+    snprintf(tprefix, sizeof(tprefix), "%sthread 0x%llx: ", prefix, (unsigned long long)_mi_thread_id());
+    mi_vfprintf(out, arg, tprefix, fmt, args);
+  }
+  else {
+    mi_vfprintf(out, arg, prefix, fmt, args);
+  }
+}
+
+void _mi_trace_message(const char* fmt, ...) {
+  if (mi_option_get(mi_option_verbose) <= 1) return;  // only with verbose level 2 or higher
+  va_list args;
+  va_start(args, fmt);
+  mi_vfprintf_thread(NULL, NULL, "mimalloc: ", fmt, args);
+  va_end(args);
+}
+
+void _mi_verbose_message(const char* fmt, ...) {
+  if (!mi_option_is_enabled(mi_option_verbose)) return;
+  va_list args;
+  va_start(args,fmt);
+  mi_vfprintf(NULL, NULL, "mimalloc: ", fmt, args);
+  va_end(args);
+}
+
+static void mi_show_error_message(const char* fmt, va_list args) {
+  if (!mi_option_is_enabled(mi_option_verbose)) {
+    if (!mi_option_is_enabled(mi_option_show_errors)) return;
+    if (mi_max_error_count >= 0 && (long)mi_atomic_increment_acq_rel(&error_count) > mi_max_error_count) return;
+  }
+  mi_vfprintf_thread(NULL, NULL, "mimalloc: error: ", fmt, args);
+}
+
+void _mi_warning_message(const char* fmt, ...) {
+  if (!mi_option_is_enabled(mi_option_verbose)) {
+    if (!mi_option_is_enabled(mi_option_show_errors)) return;
+    if (mi_max_warning_count >= 0 && (long)mi_atomic_increment_acq_rel(&warning_count) > mi_max_warning_count) return;
+  }
+  va_list args;
+  va_start(args,fmt);
+  mi_vfprintf_thread(NULL, NULL, "mimalloc: warning: ", fmt, args);
+  va_end(args);
+}
+
+
+#if MI_DEBUG
+void _mi_assert_fail(const char* assertion, const char* fname, unsigned line, const char* func ) {
+  _mi_fprintf(NULL, NULL, "mimalloc: assertion failed: at \"%s\":%u, %s\n  assertion: \"%s\"\n", fname, line, (func==NULL?"":func), assertion);
+  abort();
+}
+#endif
+
+// --------------------------------------------------------
+// Errors
+// --------------------------------------------------------
+
+static mi_error_fun* volatile  mi_error_handler; // = NULL
+static _Atomic(void*) mi_error_arg;     // = NULL
+
+static void mi_error_default(int err) {
+  MI_UNUSED(err);
+#if (MI_DEBUG>0)
+  if (err==EFAULT) {
+    #ifdef _MSC_VER
+    __debugbreak();
+    #endif
+    abort();
+  }
+#endif
+#if (MI_SECURE>0)
+  if (err==EFAULT) {  // abort on serious errors in secure mode (corrupted meta-data)
+    abort();
+  }
+#endif
+#if defined(MI_XMALLOC)
+  if (err==ENOMEM || err==EOVERFLOW) { // abort on memory allocation fails in xmalloc mode
+    abort();
+  }
+#endif
+}
+
+void mi_register_error(mi_error_fun* fun, void* arg) {
+  mi_error_handler = fun;  // can be NULL
+  mi_atomic_store_ptr_release(void,&mi_error_arg, arg);
+}
+
+void _mi_error_message(int err, const char* fmt, ...) {
+  // show detailed error message
+  va_list args;
+  va_start(args, fmt);
+  mi_show_error_message(fmt, args);
+  va_end(args);
+  // and call the error handler which may abort (or return normally)
+  if (mi_error_handler != NULL) {
+    mi_error_handler(err, mi_atomic_load_ptr_acquire(void,&mi_error_arg));
+  }
+  else {
+    mi_error_default(err);
+  }
+}
+
+// --------------------------------------------------------
+// Initialize options by checking the environment
+// --------------------------------------------------------
+char _mi_toupper(char c) {
+  if (c >= 'a' && c <= 'z') return (c - 'a' + 'A');
+		       else return c;
+}
+
+int _mi_strnicmp(const char* s, const char* t, size_t n) {
+  if (n == 0) return 0;
+  for (; *s != 0 && *t != 0 && n > 0; s++, t++, n--) {
+    if (_mi_toupper(*s) != _mi_toupper(*t)) break;
+  }
+  return (n == 0 ? 0 : *s - *t);
+}
+
+void _mi_strlcpy(char* dest, const char* src, size_t dest_size) {
+  if (dest==NULL || src==NULL || dest_size == 0) return;
+  // copy until end of src, or when dest is (almost) full
+  while (*src != 0 && dest_size > 1) {
+    *dest++ = *src++;
+    dest_size--;
+  }
+  // always zero terminate
+  *dest = 0;
+}
+
+void _mi_strlcat(char* dest, const char* src, size_t dest_size) {
+  if (dest==NULL || src==NULL || dest_size == 0) return;
+  // find end of string in the dest buffer
+  while (*dest != 0 && dest_size > 1) {
+    dest++;
+    dest_size--;
+  }
+  // and catenate
+  _mi_strlcpy(dest, src, dest_size);
+}
+
+size_t _mi_strlen(const char* s) {
+  if (s==NULL) return 0;
+  size_t len = 0;
+  while(s[len] != 0) { len++; }
+  return len;
+}
+
+size_t _mi_strnlen(const char* s, size_t max_len) {
+  if (s==NULL) return 0;
+  size_t len = 0;
+  while(s[len] != 0 && len < max_len) { len++; }
+  return len;
+}
+
+#ifdef MI_NO_GETENV
+static bool mi_getenv(const char* name, char* result, size_t result_size) {
+  MI_UNUSED(name);
+  MI_UNUSED(result);
+  MI_UNUSED(result_size);
+  return false;
+}
+#else
+static bool mi_getenv(const char* name, char* result, size_t result_size) {
+  if (name==NULL || result == NULL || result_size < 64) return false;
+  return _mi_prim_getenv(name,result,result_size);
+}
+#endif
+
+// TODO: implement ourselves to reduce dependencies on the C runtime
+#include <stdlib.h> // strtol
+#include <string.h> // strstr
+
+
+static void mi_option_init(mi_option_desc_t* desc) {
+  // Read option value from the environment
+  char s[64 + 1];
+  char buf[64+1];
+  _mi_strlcpy(buf, "mimalloc_", sizeof(buf));
+  _mi_strlcat(buf, desc->name, sizeof(buf));
+  bool found = mi_getenv(buf, s, sizeof(s));
+  if (!found && desc->legacy_name != NULL) {
+    _mi_strlcpy(buf, "mimalloc_", sizeof(buf));
+    _mi_strlcat(buf, desc->legacy_name, sizeof(buf));
+    found = mi_getenv(buf, s, sizeof(s));
+    if (found) {
+      _mi_warning_message("environment option \"mimalloc_%s\" is deprecated -- use \"mimalloc_%s\" instead.\n", desc->legacy_name, desc->name);
+    }
+  }
+
+  if (found) {
+    size_t len = _mi_strnlen(s, sizeof(buf) - 1);
+    for (size_t i = 0; i < len; i++) {
+      buf[i] = _mi_toupper(s[i]);
+    }
+    buf[len] = 0;
+    if (buf[0] == 0 || strstr("1;TRUE;YES;ON", buf) != NULL) {
+      desc->value = 1;
+      desc->init = INITIALIZED;
+    }
+    else if (strstr("0;FALSE;NO;OFF", buf) != NULL) {
+      desc->value = 0;
+      desc->init = INITIALIZED;
+    }
+    else {
+      char* end = buf;
+      long value = strtol(buf, &end, 10);
+      if (desc->option == mi_option_reserve_os_memory || desc->option == mi_option_arena_reserve) {
+	// this option is interpreted in KiB to prevent overflow of `long`
+	if (*end == 'K') { end++; }
+	else if (*end == 'M') { value *= MI_KiB; end++; }
+	else if (*end == 'G') { value *= MI_MiB; end++; }
+	else { value = (value + MI_KiB - 1) / MI_KiB; }
+	if (end[0] == 'I' && end[1] == 'B') { end += 2; }
+	else if (*end == 'B') { end++; }
+      }
+      if (*end == 0) {
+	desc->value = value;
+	desc->init = INITIALIZED;
+      }
+      else {
+	// set `init` first to avoid recursion through _mi_warning_message on mimalloc_verbose.
+	desc->init = DEFAULTED;
+	if (desc->option == mi_option_verbose && desc->value == 0) {
+	  // if the 'mimalloc_verbose' env var has a bogus value we'd never know
+	  // (since the value defaults to 'off') so in that case briefly enable verbose
+	  desc->value = 1;
+	  _mi_warning_message("environment option mimalloc_%s has an invalid value.\n", desc->name);
+	  desc->value = 0;
+	}
+	else {
+	  _mi_warning_message("environment option mimalloc_%s has an invalid value.\n", desc->name);
+	}
+      }
+    }
+    mi_assert_internal(desc->init != UNINIT);
+  }
+  else if (!_mi_preloading()) {
+    desc->init = DEFAULTED;
+  }
+}
diff --git a/compat/mimalloc/os.c b/compat/mimalloc/os.c
new file mode 100644
index 00000000000000..bf9de1be0fdb49
--- /dev/null
+++ b/compat/mimalloc/os.c
@@ -0,0 +1,689 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"
+
+
+/* -----------------------------------------------------------
+  Initialization.
+  On windows initializes support for aligned allocation and
+  large OS pages (if MIMALLOC_LARGE_OS_PAGES is true).
+----------------------------------------------------------- */
+
+static mi_os_mem_config_t mi_os_mem_config = {
+  4096,   // page size
+  0,      // large page size (usually 2MiB)
+  4096,   // allocation granularity
+  true,   // has overcommit?  (if true we use MAP_NORESERVE on mmap systems)
+  false,  // must free whole? (on mmap systems we can free anywhere in a mapped range, but on Windows we must free the entire span)
+  true    // has virtual reserve? (if true we can reserve virtual address space without using commit or physical memory)
+};
+
+bool _mi_os_has_overcommit(void) {
+  return mi_os_mem_config.has_overcommit;
+}
+
+bool _mi_os_has_virtual_reserve(void) {
+  return mi_os_mem_config.has_virtual_reserve;
+}
+
+
+// OS (small) page size
+size_t _mi_os_page_size(void) {
+  return mi_os_mem_config.page_size;
+}
+
+// if large OS pages are supported (2 or 4MiB), then return the size, otherwise return the small page size (4KiB)
+size_t _mi_os_large_page_size(void) {
+  return (mi_os_mem_config.large_page_size != 0 ? mi_os_mem_config.large_page_size : _mi_os_page_size());
+}
+
+bool _mi_os_use_large_page(size_t size, size_t alignment) {
+  // if we have access, check the size and alignment requirements
+  if (mi_os_mem_config.large_page_size == 0 || !mi_option_is_enabled(mi_option_allow_large_os_pages)) return false;
+  return ((size % mi_os_mem_config.large_page_size) == 0 && (alignment % mi_os_mem_config.large_page_size) == 0);
+}
+
+// round to a good OS allocation size (bounded by max 12.5% waste)
+size_t _mi_os_good_alloc_size(size_t size) {
+  size_t align_size;
+  if (size < 512*MI_KiB) align_size = _mi_os_page_size();
+  else if (size < 2*MI_MiB) align_size = 64*MI_KiB;
+  else if (size < 8*MI_MiB) align_size = 256*MI_KiB;
+  else if (size < 32*MI_MiB) align_size = 1*MI_MiB;
+  else align_size = 4*MI_MiB;
+  if mi_unlikely(size >= (SIZE_MAX - align_size)) return size; // possible overflow?
+  return _mi_align_up(size, align_size);
+}
+
+void _mi_os_init(void) {
+  _mi_prim_mem_init(&mi_os_mem_config);
+}
+
+
+/* -----------------------------------------------------------
+  Util
+-------------------------------------------------------------- */
+bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats);
+bool _mi_os_commit(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats);
+
+static void* mi_align_up_ptr(void* p, size_t alignment) {
+  return (void*)_mi_align_up((uintptr_t)p, alignment);
+}
+
+static void* mi_align_down_ptr(void* p, size_t alignment) {
+  return (void*)_mi_align_down((uintptr_t)p, alignment);
+}
+
+
+/* -----------------------------------------------------------
+  aligned hinting
+-------------------------------------------------------------- */
+
+// On 64-bit systems, we can do efficient aligned allocation by using
+// the 2TiB to 30TiB area to allocate those.
+#if (MI_INTPTR_SIZE >= 8)
+static mi_decl_cache_align _Atomic(uintptr_t)aligned_base;
+
+// Return a MI_SEGMENT_SIZE aligned address that is probably available.
+// If this returns NULL, the OS will determine the address but on some OS's that may not be
+// properly aligned which can be more costly as it needs to be adjusted afterwards.
+// For a size > 1GiB this always returns NULL in order to guarantee good ASLR randomization;
+// (otherwise an initial large allocation of say 2TiB has a 50% chance to include (known) addresses
+//  in the middle of the 2TiB - 6TiB address range (see issue #372))
+
+#define MI_HINT_BASE ((uintptr_t)2 << 40)  // 2TiB start
+#define MI_HINT_AREA ((uintptr_t)4 << 40)  // upto 6TiB   (since before win8 there is "only" 8TiB available to processes)
+#define MI_HINT_MAX  ((uintptr_t)30 << 40) // wrap after 30TiB (area after 32TiB is used for huge OS pages)
+
+void* _mi_os_get_aligned_hint(size_t try_alignment, size_t size)
+{
+  if (try_alignment <= 1 || try_alignment > MI_SEGMENT_SIZE) return NULL;
+  size = _mi_align_up(size, MI_SEGMENT_SIZE);
+  if (size > 1*MI_GiB) return NULL;  // guarantee the chance of fixed valid address is at most 1/(MI_HINT_AREA / 1<<30) = 1/4096.
+  #if (MI_SECURE>0)
+  size += MI_SEGMENT_SIZE;        // put in `MI_SEGMENT_SIZE` virtual gaps between hinted blocks; this splits VLA's but increases guarded areas.
+  #endif
+
+  uintptr_t hint = mi_atomic_add_acq_rel(&aligned_base, size);
+  if (hint == 0 || hint > MI_HINT_MAX) {   // wrap or initialize
+    uintptr_t init = MI_HINT_BASE;
+    #if (MI_SECURE>0 || MI_DEBUG==0)       // security: randomize start of aligned allocations unless in debug mode
+    uintptr_t r = _mi_heap_random_next(mi_prim_get_default_heap());
+    init = init + ((MI_SEGMENT_SIZE * ((r>>17) & 0xFFFFF)) % MI_HINT_AREA);  // (randomly 20 bits)*4MiB == 0 to 4TiB
+    #endif
+    uintptr_t expected = hint + size;
+    mi_atomic_cas_strong_acq_rel(&aligned_base, &expected, init);
+    hint = mi_atomic_add_acq_rel(&aligned_base, size); // this may still give 0 or > MI_HINT_MAX but that is ok, it is a hint after all
+  }
+  if (hint%try_alignment != 0) return NULL;
+  return (void*)hint;
+}
+#else
+void* _mi_os_get_aligned_hint(size_t try_alignment, size_t size) {
+  MI_UNUSED(try_alignment); MI_UNUSED(size);
+  return NULL;
+}
+#endif
+
+
+/* -----------------------------------------------------------
+  Free memory
+-------------------------------------------------------------- */
+
+static void mi_os_free_huge_os_pages(void* p, size_t size, mi_stats_t* stats);
+
+static void mi_os_prim_free(void* addr, size_t size, bool still_committed, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  mi_assert_internal((size % _mi_os_page_size()) == 0);
+  if (addr == NULL || size == 0) return; // || _mi_os_is_huge_reserved(addr)
+  int err = _mi_prim_free(addr, size);
+  if (err != 0) {
+    _mi_warning_message("unable to free OS memory (error: %d (0x%x), size: 0x%zx bytes, address: %p)\n", err, err, size, addr);
+  }
+  mi_stats_t* stats = &_mi_stats_main;
+  if (still_committed) { _mi_stat_decrease(&stats->committed, size); }
+  _mi_stat_decrease(&stats->reserved, size);
+}
+
+void _mi_os_free_ex(void* addr, size_t size, bool still_committed, mi_memid_t memid, mi_stats_t* tld_stats) {
+  if (mi_memkind_is_os(memid.memkind)) {
+    size_t csize = _mi_os_good_alloc_size(size);
+    void* base = addr;
+    // different base? (due to alignment)
+    if (memid.mem.os.base != NULL) {
+      mi_assert(memid.mem.os.base <= addr);
+      mi_assert((uint8_t*)memid.mem.os.base + memid.mem.os.alignment >= (uint8_t*)addr);
+      base = memid.mem.os.base;
+      csize += ((uint8_t*)addr - (uint8_t*)memid.mem.os.base);
+    }
+    // free it
+    if (memid.memkind == MI_MEM_OS_HUGE) {
+      mi_assert(memid.is_pinned);
+      mi_os_free_huge_os_pages(base, csize, tld_stats);
+    }
+    else {
+      mi_os_prim_free(base, csize, still_committed, tld_stats);
+    }
+  }
+  else {
+    // nothing to do
+    mi_assert(memid.memkind < MI_MEM_OS);
+  }
+}
+
+void  _mi_os_free(void* p, size_t size, mi_memid_t memid, mi_stats_t* tld_stats) {
+  _mi_os_free_ex(p, size, true, memid, tld_stats);
+}
+
+
+/* -----------------------------------------------------------
+   Primitive allocation from the OS.
+-------------------------------------------------------------- */
+
+// Note: the `try_alignment` is just a hint and the returned pointer is not guaranteed to be aligned.
+static void* mi_os_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, mi_stats_t* stats) {
+  mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
+  mi_assert_internal(is_zero != NULL);
+  mi_assert_internal(is_large != NULL);
+  if (size == 0) return NULL;
+  if (!commit) { allow_large = false; }
+  if (try_alignment == 0) { try_alignment = 1; } // avoid 0 to ensure there will be no divide by zero when aligning
+
+  *is_zero = false;
+  void* p = NULL;
+  int err = _mi_prim_alloc(size, try_alignment, commit, allow_large, is_large, is_zero, &p);
+  if (err != 0) {
+    _mi_warning_message("unable to allocate OS memory (error: %d (0x%x), size: 0x%zx bytes, align: 0x%zx, commit: %d, allow large: %d)\n", err, err, size, try_alignment, commit, allow_large);
+  }
+  mi_stat_counter_increase(stats->mmap_calls, 1);
+  if (p != NULL) {
+    _mi_stat_increase(&stats->reserved, size);
+    if (commit) {
+      _mi_stat_increase(&stats->committed, size);
+      // seems needed for asan (or `mimalloc-test-api` fails)
+      #ifdef MI_TRACK_ASAN
+      if (*is_zero) { mi_track_mem_defined(p,size); }
+	       else { mi_track_mem_undefined(p,size); }
+      #endif
+    }
+  }
+  return p;
+}
+
+
+// Primitive aligned allocation from the OS.
+// This function guarantees the allocated memory is aligned.
+static void* mi_os_prim_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, void** base, mi_stats_t* stats) {
+  mi_assert_internal(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0));
+  mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
+  mi_assert_internal(is_large != NULL);
+  mi_assert_internal(is_zero != NULL);
+  mi_assert_internal(base != NULL);
+  if (!commit) allow_large = false;
+  if (!(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0))) return NULL;
+  size = _mi_align_up(size, _mi_os_page_size());
+
+  // try first with a hint (this will be aligned directly on Win 10+ or BSD)
+  void* p = mi_os_prim_alloc(size, alignment, commit, allow_large, is_large, is_zero, stats);
+  if (p == NULL) return NULL;
+
+  // aligned already?
+  if (((uintptr_t)p % alignment) == 0) {
+    *base = p;
+  }
+  else {
+    // if not aligned, free it, overallocate, and unmap around it
+    _mi_warning_message("unable to allocate aligned OS memory directly, fall back to over-allocation (size: 0x%zx bytes, address: %p, alignment: 0x%zx, commit: %d)\n", size, p, alignment, commit);
+    mi_os_prim_free(p, size, commit, stats);
+    if (size >= (SIZE_MAX - alignment)) return NULL; // overflow
+    const size_t over_size = size + alignment;
+
+    if (mi_os_mem_config.must_free_whole) {  // win32 virtualAlloc cannot free parts of an allocate block
+      // over-allocate uncommitted (virtual) memory
+      p = mi_os_prim_alloc(over_size, 1 /*alignment*/, false /* commit? */, false /* allow_large */, is_large, is_zero, stats);
+      if (p == NULL) return NULL;
+
+      // set p to the aligned part in the full region
+      // note: this is dangerous on Windows as VirtualFree needs the actual base pointer
+      // this is handled though by having the `base` field in the memid's
+      *base = p; // remember the base
+      p = mi_align_up_ptr(p, alignment);
+
+      // explicitly commit only the aligned part
+      if (commit) {
+	_mi_os_commit(p, size, NULL, stats);
+      }
+    }
+    else  { // mmap can free inside an allocation
+      // overallocate...
+      p = mi_os_prim_alloc(over_size, 1, commit, false, is_large, is_zero, stats);
+      if (p == NULL) return NULL;
+
+      // and selectively unmap parts around the over-allocated area. (noop on sbrk)
+      void* aligned_p = mi_align_up_ptr(p, alignment);
+      size_t pre_size = (uint8_t*)aligned_p - (uint8_t*)p;
+      size_t mid_size = _mi_align_up(size, _mi_os_page_size());
+      size_t post_size = over_size - pre_size - mid_size;
+      mi_assert_internal(pre_size < over_size&& post_size < over_size&& mid_size >= size);
+      if (pre_size > 0)  { mi_os_prim_free(p, pre_size, commit, stats); }
+      if (post_size > 0) { mi_os_prim_free((uint8_t*)aligned_p + mid_size, post_size, commit, stats); }
+      // we can return the aligned pointer on `mmap` (and sbrk) systems
+      p = aligned_p;
+      *base = aligned_p; // since we freed the pre part, `*base == p`.
+    }
+  }
+
+  mi_assert_internal(p == NULL || (p != NULL && *base != NULL && ((uintptr_t)p % alignment) == 0));
+  return p;
+}
+
+
+/* -----------------------------------------------------------
+  OS API: alloc and alloc_aligned
+----------------------------------------------------------- */
+
+void* _mi_os_alloc(size_t size, mi_memid_t* memid, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  *memid = _mi_memid_none();
+  mi_stats_t* stats = &_mi_stats_main;
+  if (size == 0) return NULL;
+  size = _mi_os_good_alloc_size(size);
+  bool os_is_large = false;
+  bool os_is_zero  = false;
+  void* p = mi_os_prim_alloc(size, 0, true, false, &os_is_large, &os_is_zero, stats);
+  if (p != NULL) {
+    *memid = _mi_memid_create_os(true, os_is_zero, os_is_large);
+  }
+  return p;
+}
+
+void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* tld_stats)
+{
+  MI_UNUSED(&_mi_os_get_aligned_hint); // suppress unused warnings
+  MI_UNUSED(tld_stats);
+  *memid = _mi_memid_none();
+  if (size == 0) return NULL;
+  size = _mi_os_good_alloc_size(size);
+  alignment = _mi_align_up(alignment, _mi_os_page_size());
+
+  bool os_is_large = false;
+  bool os_is_zero  = false;
+  void* os_base = NULL;
+  void* p = mi_os_prim_alloc_aligned(size, alignment, commit, allow_large, &os_is_large, &os_is_zero, &os_base, &_mi_stats_main /*tld->stats*/ );
+  if (p != NULL) {
+    *memid = _mi_memid_create_os(commit, os_is_zero, os_is_large);
+    memid->mem.os.base = os_base;
+    memid->mem.os.alignment = alignment;
+  }
+  return p;
+}
+
+/* -----------------------------------------------------------
+  OS aligned allocation with an offset. This is used
+  for large alignments > MI_ALIGNMENT_MAX. We use a large mimalloc
+  page where the object can be aligned at an offset from the start of the segment.
+  As we may need to overallocate, we need to free such pointers using `mi_free_aligned`
+  to use the actual start of the memory region.
+----------------------------------------------------------- */
+
+void* _mi_os_alloc_aligned_at_offset(size_t size, size_t alignment, size_t offset, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* tld_stats) {
+  mi_assert(offset <= MI_SEGMENT_SIZE);
+  mi_assert(offset <= size);
+  mi_assert((alignment % _mi_os_page_size()) == 0);
+  *memid = _mi_memid_none();
+  if (offset > MI_SEGMENT_SIZE) return NULL;
+  if (offset == 0) {
+    // regular aligned allocation
+    return _mi_os_alloc_aligned(size, alignment, commit, allow_large, memid, tld_stats);
+  }
+  else {
+    // overallocate to align at an offset
+    const size_t extra = _mi_align_up(offset, alignment) - offset;
+    const size_t oversize = size + extra;
+    void* const start = _mi_os_alloc_aligned(oversize, alignment, commit, allow_large, memid, tld_stats);
+    if (start == NULL) return NULL;
+
+    void* const p = (uint8_t*)start + extra;
+    mi_assert(_mi_is_aligned((uint8_t*)p + offset, alignment));
+    // decommit the overallocation at the start
+    if (commit && extra > _mi_os_page_size()) {
+      _mi_os_decommit(start, extra, tld_stats);
+    }
+    return p;
+  }
+}
+
+/* -----------------------------------------------------------
+  OS memory API: reset, commit, decommit, protect, unprotect.
+----------------------------------------------------------- */
+
+// OS page align within a given area, either conservative (pages inside the area only),
+// or not (straddling pages outside the area is possible)
+static void* mi_os_page_align_areax(bool conservative, void* addr, size_t size, size_t* newsize) {
+  mi_assert(addr != NULL && size > 0);
+  if (newsize != NULL) *newsize = 0;
+  if (size == 0 || addr == NULL) return NULL;
+
+  // page align conservatively within the range
+  void* start = (conservative ? mi_align_up_ptr(addr, _mi_os_page_size())
+    : mi_align_down_ptr(addr, _mi_os_page_size()));
+  void* end = (conservative ? mi_align_down_ptr((uint8_t*)addr + size, _mi_os_page_size())
+    : mi_align_up_ptr((uint8_t*)addr + size, _mi_os_page_size()));
+  ptrdiff_t diff = (uint8_t*)end - (uint8_t*)start;
+  if (diff <= 0) return NULL;
+
+  mi_assert_internal((conservative && (size_t)diff <= size) || (!conservative && (size_t)diff >= size));
+  if (newsize != NULL) *newsize = (size_t)diff;
+  return start;
+}
+
+static void* mi_os_page_align_area_conservative(void* addr, size_t size, size_t* newsize) {
+  return mi_os_page_align_areax(true, addr, size, newsize);
+}
+
+bool _mi_os_commit(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  mi_stats_t* stats = &_mi_stats_main;
+  if (is_zero != NULL) { *is_zero = false; }
+  _mi_stat_increase(&stats->committed, size);  // use size for precise commit vs. decommit
+  _mi_stat_counter_increase(&stats->commit_calls, 1);
+
+  // page align range
+  size_t csize;
+  void* start = mi_os_page_align_areax(false /* conservative? */, addr, size, &csize);
+  if (csize == 0) return true;
+
+  // commit
+  bool os_is_zero = false;
+  int err = _mi_prim_commit(start, csize, &os_is_zero);
+  if (err != 0) {
+    _mi_warning_message("cannot commit OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", err, err, start, csize);
+    return false;
+  }
+  if (os_is_zero && is_zero != NULL) {
+    *is_zero = true;
+    mi_assert_expensive(mi_mem_is_zero(start, csize));
+  }
+  // note: the following seems required for asan (otherwise `mimalloc-test-stress` fails)
+  #ifdef MI_TRACK_ASAN
+  if (os_is_zero) { mi_track_mem_defined(start,csize); }
+	     else { mi_track_mem_undefined(start,csize); }
+  #endif
+  return true;
+}
+
+static bool mi_os_decommit_ex(void* addr, size_t size, bool* needs_recommit, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  mi_stats_t* stats = &_mi_stats_main;
+  mi_assert_internal(needs_recommit!=NULL);
+  _mi_stat_decrease(&stats->committed, size);
+
+  // page align
+  size_t csize;
+  void* start = mi_os_page_align_area_conservative(addr, size, &csize);
+  if (csize == 0) return true;
+
+  // decommit
+  *needs_recommit = true;
+  int err = _mi_prim_decommit(start,csize,needs_recommit);
+  if (err != 0) {
+    _mi_warning_message("cannot decommit OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", err, err, start, csize);
+  }
+  mi_assert_internal(err == 0);
+  return (err == 0);
+}
+
+bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* tld_stats) {
+  bool needs_recommit;
+  return mi_os_decommit_ex(addr, size, &needs_recommit, tld_stats);
+}
+
+
+// Signal to the OS that the address range is no longer in use
+// but may be used later again. This will release physical memory
+// pages and reduce swapping while keeping the memory committed.
+// We page align to a conservative area inside the range to reset.
+bool _mi_os_reset(void* addr, size_t size, mi_stats_t* stats) {
+  // page align conservatively within the range
+  size_t csize;
+  void* start = mi_os_page_align_area_conservative(addr, size, &csize);
+  if (csize == 0) return true;  // || _mi_os_is_huge_reserved(addr)
+  _mi_stat_increase(&stats->reset, csize);
+  _mi_stat_counter_increase(&stats->reset_calls, 1);
+
+  #if (MI_DEBUG>1) && !MI_SECURE && !MI_TRACK_ENABLED // && !MI_TSAN
+  memset(start, 0, csize); // pretend it is eagerly reset
+  #endif
+
+  int err = _mi_prim_reset(start, csize);
+  if (err != 0) {
+    _mi_warning_message("cannot reset OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", err, err, start, csize);
+  }
+  return (err == 0);
+}
+
+
+// either resets or decommits memory, returns true if the memory needs
+// to be recommitted if it is to be re-used later on.
+bool _mi_os_purge_ex(void* p, size_t size, bool allow_reset, mi_stats_t* stats)
+{
+  if (mi_option_get(mi_option_purge_delay) < 0) return false;  // is purging allowed?
+  _mi_stat_counter_increase(&stats->purge_calls, 1);
+  _mi_stat_increase(&stats->purged, size);
+
+  if (mi_option_is_enabled(mi_option_purge_decommits) &&   // should decommit?
+      !_mi_preloading())                                   // don't decommit during preloading (unsafe)
+  {
+    bool needs_recommit = true;
+    mi_os_decommit_ex(p, size, &needs_recommit, stats);
+    return needs_recommit;
+  }
+  else {
+    if (allow_reset) {  // this can sometimes be not allowed if the range is not fully committed
+      _mi_os_reset(p, size, stats);
+    }
+    return false;  // needs no recommit
+  }
+}
+
+// either resets or decommits memory, returns true if the memory needs
+// to be recommitted if it is to be re-used later on.
+bool _mi_os_purge(void* p, size_t size, mi_stats_t * stats) {
+  return _mi_os_purge_ex(p, size, true, stats);
+}
+
+// Protect a region in memory to be not accessible.
+static  bool mi_os_protectx(void* addr, size_t size, bool protect) {
+  // page align conservatively within the range
+  size_t csize = 0;
+  void* start = mi_os_page_align_area_conservative(addr, size, &csize);
+  if (csize == 0) return false;
+  /*
+  if (_mi_os_is_huge_reserved(addr)) {
+	  _mi_warning_message("cannot mprotect memory allocated in huge OS pages\n");
+  }
+  */
+  int err = _mi_prim_protect(start,csize,protect);
+  if (err != 0) {
+    _mi_warning_message("cannot %s OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", (protect ? "protect" : "unprotect"), err, err, start, csize);
+  }
+  return (err == 0);
+}
+
+bool _mi_os_protect(void* addr, size_t size) {
+  return mi_os_protectx(addr, size, true);
+}
+
+bool _mi_os_unprotect(void* addr, size_t size) {
+  return mi_os_protectx(addr, size, false);
+}
+
+
+
+/* ----------------------------------------------------------------------------
+Support for allocating huge OS pages (1Gib) that are reserved up-front
+and possibly associated with a specific NUMA node. (use `numa_node>=0`)
+-----------------------------------------------------------------------------*/
+#define MI_HUGE_OS_PAGE_SIZE  (MI_GiB)
+
+
+#if (MI_INTPTR_SIZE >= 8)
+// To ensure proper alignment, use our own area for huge OS pages
+static mi_decl_cache_align _Atomic(uintptr_t)  mi_huge_start; // = 0
+
+// Claim an aligned address range for huge pages
+static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) {
+  if (total_size != NULL) *total_size = 0;
+  const size_t size = pages * MI_HUGE_OS_PAGE_SIZE;
+
+  uintptr_t start = 0;
+  uintptr_t end = 0;
+  uintptr_t huge_start = mi_atomic_load_relaxed(&mi_huge_start);
+  do {
+    start = huge_start;
+    if (start == 0) {
+      // Initialize the start address after the 32TiB area
+      start = ((uintptr_t)32 << 40);  // 32TiB virtual start address
+    #if (MI_SECURE>0 || MI_DEBUG==0)      // security: randomize start of huge pages unless in debug mode
+      uintptr_t r = _mi_heap_random_next(mi_prim_get_default_heap());
+      start = start + ((uintptr_t)MI_HUGE_OS_PAGE_SIZE * ((r>>17) & 0x0FFF));  // (randomly 12bits)*1GiB == between 0 to 4TiB
+    #endif
+    }
+    end = start + size;
+    mi_assert_internal(end % MI_SEGMENT_SIZE == 0);
+  } while (!mi_atomic_cas_strong_acq_rel(&mi_huge_start, &huge_start, end));
+
+  if (total_size != NULL) *total_size = size;
+  return (uint8_t*)start;
+}
+#else
+static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) {
+  MI_UNUSED(pages);
+  if (total_size != NULL) *total_size = 0;
+  return NULL;
+}
+#endif
+
+// Allocate MI_SEGMENT_SIZE aligned huge pages
+void* _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, mi_msecs_t max_msecs, size_t* pages_reserved, size_t* psize, mi_memid_t* memid) {
+  *memid = _mi_memid_none();
+  if (psize != NULL) *psize = 0;
+  if (pages_reserved != NULL) *pages_reserved = 0;
+  size_t size = 0;
+  uint8_t* start = mi_os_claim_huge_pages(pages, &size);
+  if (start == NULL) return NULL; // or 32-bit systems
+
+  // Allocate one page at the time but try to place them contiguously
+  // We allocate one page at the time to be able to abort if it takes too long
+  // or to at least allocate as many as available on the system.
+  mi_msecs_t start_t = _mi_clock_start();
+  size_t page = 0;
+  bool all_zero = true;
+  while (page < pages) {
+    // allocate a page
+    bool is_zero = false;
+    void* addr = start + (page * MI_HUGE_OS_PAGE_SIZE);
+    void* p = NULL;
+    int err = _mi_prim_alloc_huge_os_pages(addr, MI_HUGE_OS_PAGE_SIZE, numa_node, &is_zero, &p);
+    if (!is_zero) { all_zero = false;  }
+    if (err != 0) {
+      _mi_warning_message("unable to allocate huge OS page (error: %d (0x%x), address: %p, size: %zx bytes)\n", err, err, addr, MI_HUGE_OS_PAGE_SIZE);
+      break;
+    }
+
+    // Did we succeed at a contiguous address?
+    if (p != addr) {
+      // no success, issue a warning and break
+      if (p != NULL) {
+	_mi_warning_message("could not allocate contiguous huge OS page %zu at %p\n", page, addr);
+	mi_os_prim_free(p, MI_HUGE_OS_PAGE_SIZE, true, &_mi_stats_main);
+      }
+      break;
+    }
+
+    // success, record it
+    page++;  // increase before timeout check (see issue #711)
+    _mi_stat_increase(&_mi_stats_main.committed, MI_HUGE_OS_PAGE_SIZE);
+    _mi_stat_increase(&_mi_stats_main.reserved, MI_HUGE_OS_PAGE_SIZE);
+
+    // check for timeout
+    if (max_msecs > 0) {
+      mi_msecs_t elapsed = _mi_clock_end(start_t);
+      if (page >= 1) {
+	mi_msecs_t estimate = ((elapsed / (page+1)) * pages);
+	if (estimate > 2*max_msecs) { // seems like we are going to timeout, break
+	  elapsed = max_msecs + 1;
+	}
+      }
+      if (elapsed > max_msecs) {
+	_mi_warning_message("huge OS page allocation timed out (after allocating %zu page(s))\n", page);
+	break;
+      }
+    }
+  }
+  mi_assert_internal(page*MI_HUGE_OS_PAGE_SIZE <= size);
+  if (pages_reserved != NULL) { *pages_reserved = page; }
+  if (psize != NULL) { *psize = page * MI_HUGE_OS_PAGE_SIZE; }
+  if (page != 0) {
+    mi_assert(start != NULL);
+    *memid = _mi_memid_create_os(true /* is committed */, all_zero, true /* is_large */);
+    memid->memkind = MI_MEM_OS_HUGE;
+    mi_assert(memid->is_pinned);
+    #ifdef MI_TRACK_ASAN
+    if (all_zero) { mi_track_mem_defined(start,size); }
+    #endif
+  }
+  return (page == 0 ? NULL : start);
+}
+
+// free every huge page in a range individually (as we allocated per page)
+// note: needed with VirtualAlloc but could potentially be done in one go on mmap'd systems.
+static void mi_os_free_huge_os_pages(void* p, size_t size, mi_stats_t* stats) {
+  if (p==NULL || size==0) return;
+  uint8_t* base = (uint8_t*)p;
+  while (size >= MI_HUGE_OS_PAGE_SIZE) {
+    mi_os_prim_free(base, MI_HUGE_OS_PAGE_SIZE, true, stats);
+    size -= MI_HUGE_OS_PAGE_SIZE;
+    base += MI_HUGE_OS_PAGE_SIZE;
+  }
+}
+
+/* ----------------------------------------------------------------------------
+Support NUMA aware allocation
+-----------------------------------------------------------------------------*/
+
+_Atomic(size_t)  _mi_numa_node_count; // = 0   // cache the node count
+
+size_t _mi_os_numa_node_count_get(void) {
+  size_t count = mi_atomic_load_acquire(&_mi_numa_node_count);
+  if (count <= 0) {
+    long ncount = mi_option_get(mi_option_use_numa_nodes); // given explicitly?
+    if (ncount > 0) {
+      count = (size_t)ncount;
+    }
+    else {
+      count = _mi_prim_numa_node_count(); // or detect dynamically
+      if (count == 0) count = 1;
+    }
+    mi_atomic_store_release(&_mi_numa_node_count, count); // save it
+    _mi_verbose_message("using %zd numa regions\n", count);
+  }
+  return count;
+}
+
+int _mi_os_numa_node_get(mi_os_tld_t* tld) {
+  MI_UNUSED(tld);
+  size_t numa_count = _mi_os_numa_node_count();
+  if (numa_count<=1) return 0; // optimize on single numa node systems: always node 0
+  // never more than the node count and >= 0
+  size_t numa_node = _mi_prim_numa_node();
+  if (numa_node >= numa_count) { numa_node = numa_node % numa_count; }
+  return (int)numa_node;
+}
diff --git a/compat/mimalloc/page-queue.c b/compat/mimalloc/page-queue.c
new file mode 100644
index 00000000000000..5619a81f9917fe
--- /dev/null
+++ b/compat/mimalloc/page-queue.c
@@ -0,0 +1,332 @@
+/*----------------------------------------------------------------------------
+Copyright (c) 2018-2020, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* -----------------------------------------------------------
+  Definition of page queues for each block size
+----------------------------------------------------------- */
+
+#ifndef MI_IN_PAGE_C
+#error "this file should be included from 'page.c'"
+#endif
+
+/* -----------------------------------------------------------
+  Minimal alignment in machine words (i.e. `sizeof(void*)`)
+----------------------------------------------------------- */
+
+#if (MI_MAX_ALIGN_SIZE > 4*MI_INTPTR_SIZE)
+  #error "define alignment for more than 4x word size for this platform"
+#elif (MI_MAX_ALIGN_SIZE > 2*MI_INTPTR_SIZE)
+  #define MI_ALIGN4W   // 4 machine words minimal alignment
+#elif (MI_MAX_ALIGN_SIZE > MI_INTPTR_SIZE)
+  #define MI_ALIGN2W   // 2 machine words minimal alignment
+#else
+  // ok, default alignment is 1 word
+#endif
+
+
+/* -----------------------------------------------------------
+  Queue query
+----------------------------------------------------------- */
+
+
+static inline bool mi_page_queue_is_huge(const mi_page_queue_t* pq) {
+  return (pq->block_size == (MI_MEDIUM_OBJ_SIZE_MAX+sizeof(uintptr_t)));
+}
+
+static inline bool mi_page_queue_is_full(const mi_page_queue_t* pq) {
+  return (pq->block_size == (MI_MEDIUM_OBJ_SIZE_MAX+(2*sizeof(uintptr_t))));
+}
+
+static inline bool mi_page_queue_is_special(const mi_page_queue_t* pq) {
+  return (pq->block_size > MI_MEDIUM_OBJ_SIZE_MAX);
+}
+
+/* -----------------------------------------------------------
+  Bins
+----------------------------------------------------------- */
+
+// Return the bin for a given field size.
+// Returns MI_BIN_HUGE if the size is too large.
+// We use `wsize` for the size in "machine word sizes",
+// i.e. byte size == `wsize*sizeof(void*)`.
+static inline uint8_t mi_bin(size_t size) {
+  size_t wsize = _mi_wsize_from_size(size);
+  uint8_t bin;
+  if (wsize <= 1) {
+    bin = 1;
+  }
+  #if defined(MI_ALIGN4W)
+  else if (wsize <= 4) {
+    bin = (uint8_t)((wsize+1)&~1); // round to double word sizes
+  }
+  #elif defined(MI_ALIGN2W)
+  else if (wsize <= 8) {
+    bin = (uint8_t)((wsize+1)&~1); // round to double word sizes
+  }
+  #else
+  else if (wsize <= 8) {
+    bin = (uint8_t)wsize;
+  }
+  #endif
+  else if (wsize > MI_MEDIUM_OBJ_WSIZE_MAX) {
+    bin = MI_BIN_HUGE;
+  }
+  else {
+    #if defined(MI_ALIGN4W)
+    if (wsize <= 16) { wsize = (wsize+3)&~3; } // round to 4x word sizes
+    #endif
+    wsize--;
+    // find the highest bit
+    uint8_t b = (uint8_t)mi_bsr(wsize);  // note: wsize != 0
+    // and use the top 3 bits to determine the bin (~12.5% worst internal fragmentation).
+    // - adjust with 3 because we use do not round the first 8 sizes
+    //   which each get an exact bin
+    bin = ((b << 2) + (uint8_t)((wsize >> (b - 2)) & 0x03)) - 3;
+    mi_assert_internal(bin < MI_BIN_HUGE);
+  }
+  mi_assert_internal(bin > 0 && bin <= MI_BIN_HUGE);
+  return bin;
+}
+
+
+
+/* -----------------------------------------------------------
+  Queue of pages with free blocks
+----------------------------------------------------------- */
+
+uint8_t _mi_bin(size_t size) {
+  return mi_bin(size);
+}
+
+size_t _mi_bin_size(uint8_t bin) {
+  return _mi_heap_empty.pages[bin].block_size;
+}
+
+// Good size for allocation
+size_t mi_good_size(size_t size) mi_attr_noexcept {
+  if (size <= MI_MEDIUM_OBJ_SIZE_MAX) {
+    return _mi_bin_size(mi_bin(size));
+  }
+  else {
+    return _mi_align_up(size,_mi_os_page_size());
+  }
+}
+
+#if (MI_DEBUG>1)
+static bool mi_page_queue_contains(mi_page_queue_t* queue, const mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  mi_page_t* list = queue->first;
+  while (list != NULL) {
+    mi_assert_internal(list->next == NULL || list->next->prev == list);
+    mi_assert_internal(list->prev == NULL || list->prev->next == list);
+    if (list == page) break;
+    list = list->next;
+  }
+  return (list == page);
+}
+
+#endif
+
+#if (MI_DEBUG>1)
+static bool mi_heap_contains_queue(const mi_heap_t* heap, const mi_page_queue_t* pq) {
+  return (pq >= &heap->pages[0] && pq <= &heap->pages[MI_BIN_FULL]);
+}
+#endif
+
+static mi_page_queue_t* mi_page_queue_of(const mi_page_t* page) {
+  uint8_t bin = (mi_page_is_in_full(page) ? MI_BIN_FULL : mi_bin(page->xblock_size));
+  mi_heap_t* heap = mi_page_heap(page);
+  mi_assert_internal(heap != NULL && bin <= MI_BIN_FULL);
+  mi_page_queue_t* pq = &heap->pages[bin];
+  mi_assert_internal(bin >= MI_BIN_HUGE || page->xblock_size == pq->block_size);
+  mi_assert_expensive(mi_page_queue_contains(pq, page));
+  return pq;
+}
+
+static mi_page_queue_t* mi_heap_page_queue_of(mi_heap_t* heap, const mi_page_t* page) {
+  uint8_t bin = (mi_page_is_in_full(page) ? MI_BIN_FULL : mi_bin(page->xblock_size));
+  mi_assert_internal(bin <= MI_BIN_FULL);
+  mi_page_queue_t* pq = &heap->pages[bin];
+  mi_assert_internal(mi_page_is_in_full(page) || page->xblock_size == pq->block_size);
+  return pq;
+}
+
+// The current small page array is for efficiency and for each
+// small size (up to 256) it points directly to the page for that
+// size without having to compute the bin. This means when the
+// current free page queue is updated for a small bin, we need to update a
+// range of entries in `_mi_page_small_free`.
+static inline void mi_heap_queue_first_update(mi_heap_t* heap, const mi_page_queue_t* pq) {
+  mi_assert_internal(mi_heap_contains_queue(heap,pq));
+  size_t size = pq->block_size;
+  if (size > MI_SMALL_SIZE_MAX) return;
+
+  mi_page_t* page = pq->first;
+  if (pq->first == NULL) page = (mi_page_t*)&_mi_page_empty;
+
+  // find index in the right direct page array
+  size_t start;
+  size_t idx = _mi_wsize_from_size(size);
+  mi_page_t** pages_free = heap->pages_free_direct;
+
+  if (pages_free[idx] == page) return;  // already set
+
+  // find start slot
+  if (idx<=1) {
+    start = 0;
+  }
+  else {
+    // find previous size; due to minimal alignment upto 3 previous bins may need to be skipped
+    uint8_t bin = mi_bin(size);
+    const mi_page_queue_t* prev = pq - 1;
+    while( bin == mi_bin(prev->block_size) && prev > &heap->pages[0]) {
+      prev--;
+    }
+    start = 1 + _mi_wsize_from_size(prev->block_size);
+    if (start > idx) start = idx;
+  }
+
+  // set size range to the right page
+  mi_assert(start <= idx);
+  for (size_t sz = start; sz <= idx; sz++) {
+    pages_free[sz] = page;
+  }
+}
+
+/*
+static bool mi_page_queue_is_empty(mi_page_queue_t* queue) {
+  return (queue->first == NULL);
+}
+*/
+
+static void mi_page_queue_remove(mi_page_queue_t* queue, mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(mi_page_queue_contains(queue, page));
+  mi_assert_internal(page->xblock_size == queue->block_size || (page->xblock_size > MI_MEDIUM_OBJ_SIZE_MAX && mi_page_queue_is_huge(queue))  || (mi_page_is_in_full(page) && mi_page_queue_is_full(queue)));
+  mi_heap_t* heap = mi_page_heap(page);
+
+  if (page->prev != NULL) page->prev->next = page->next;
+  if (page->next != NULL) page->next->prev = page->prev;
+  if (page == queue->last)  queue->last = page->prev;
+  if (page == queue->first) {
+    queue->first = page->next;
+    // update first
+    mi_assert_internal(mi_heap_contains_queue(heap, queue));
+    mi_heap_queue_first_update(heap,queue);
+  }
+  heap->page_count--;
+  page->next = NULL;
+  page->prev = NULL;
+  // mi_atomic_store_ptr_release(mi_atomic_cast(void*, &page->heap), NULL);
+  mi_page_set_in_full(page,false);
+}
+
+
+static void mi_page_queue_push(mi_heap_t* heap, mi_page_queue_t* queue, mi_page_t* page) {
+  mi_assert_internal(mi_page_heap(page) == heap);
+  mi_assert_internal(!mi_page_queue_contains(queue, page));
+  #if MI_HUGE_PAGE_ABANDON
+  mi_assert_internal(_mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
+  #endif
+  mi_assert_internal(page->xblock_size == queue->block_size ||
+		      (page->xblock_size > MI_MEDIUM_OBJ_SIZE_MAX) ||
+			(mi_page_is_in_full(page) && mi_page_queue_is_full(queue)));
+
+  mi_page_set_in_full(page, mi_page_queue_is_full(queue));
+  // mi_atomic_store_ptr_release(mi_atomic_cast(void*, &page->heap), heap);
+  page->next = queue->first;
+  page->prev = NULL;
+  if (queue->first != NULL) {
+    mi_assert_internal(queue->first->prev == NULL);
+    queue->first->prev = page;
+    queue->first = page;
+  }
+  else {
+    queue->first = queue->last = page;
+  }
+
+  // update direct
+  mi_heap_queue_first_update(heap, queue);
+  heap->page_count++;
+}
+
+
+static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* from, mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(mi_page_queue_contains(from, page));
+  mi_assert_expensive(!mi_page_queue_contains(to, page));
+
+  mi_assert_internal((page->xblock_size == to->block_size && page->xblock_size == from->block_size) ||
+		     (page->xblock_size == to->block_size && mi_page_queue_is_full(from)) ||
+		     (page->xblock_size == from->block_size && mi_page_queue_is_full(to)) ||
+		     (page->xblock_size > MI_LARGE_OBJ_SIZE_MAX && mi_page_queue_is_huge(to)) ||
+		     (page->xblock_size > MI_LARGE_OBJ_SIZE_MAX && mi_page_queue_is_full(to)));
+
+  mi_heap_t* heap = mi_page_heap(page);
+  if (page->prev != NULL) page->prev->next = page->next;
+  if (page->next != NULL) page->next->prev = page->prev;
+  if (page == from->last)  from->last = page->prev;
+  if (page == from->first) {
+    from->first = page->next;
+    // update first
+    mi_assert_internal(mi_heap_contains_queue(heap, from));
+    mi_heap_queue_first_update(heap, from);
+  }
+
+  page->prev = to->last;
+  page->next = NULL;
+  if (to->last != NULL) {
+    mi_assert_internal(heap == mi_page_heap(to->last));
+    to->last->next = page;
+    to->last = page;
+  }
+  else {
+    to->first = page;
+    to->last = page;
+    mi_heap_queue_first_update(heap, to);
+  }
+
+  mi_page_set_in_full(page, mi_page_queue_is_full(to));
+}
+
+// Only called from `mi_heap_absorb`.
+size_t _mi_page_queue_append(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_queue_t* append) {
+  mi_assert_internal(mi_heap_contains_queue(heap,pq));
+  mi_assert_internal(pq->block_size == append->block_size);
+
+  if (append->first==NULL) return 0;
+
+  // set append pages to new heap and count
+  size_t count = 0;
+  for (mi_page_t* page = append->first; page != NULL; page = page->next) {
+    // inline `mi_page_set_heap` to avoid wrong assertion during absorption;
+    // in this case it is ok to be delayed freeing since both "to" and "from" heap are still alive.
+    mi_atomic_store_release(&page->xheap, (uintptr_t)heap);
+    // set the flag to delayed free (not overriding NEVER_DELAYED_FREE) which has as a
+    // side effect that it spins until any DELAYED_FREEING is finished. This ensures
+    // that after appending only the new heap will be used for delayed free operations.
+    _mi_page_use_delayed_free(page, MI_USE_DELAYED_FREE, false);
+    count++;
+  }
+
+  if (pq->last==NULL) {
+    // take over afresh
+    mi_assert_internal(pq->first==NULL);
+    pq->first = append->first;
+    pq->last = append->last;
+    mi_heap_queue_first_update(heap, pq);
+  }
+  else {
+    // append to end
+    mi_assert_internal(pq->last!=NULL);
+    mi_assert_internal(append->first!=NULL);
+    pq->last->next = append->first;
+    append->first->prev = pq->last;
+    pq->last = append->last;
+  }
+  return count;
+}
diff --git a/compat/mimalloc/page.c b/compat/mimalloc/page.c
new file mode 100644
index 00000000000000..211204aa79e59d
--- /dev/null
+++ b/compat/mimalloc/page.c
@@ -0,0 +1,939 @@
+/*----------------------------------------------------------------------------
+Copyright (c) 2018-2020, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* -----------------------------------------------------------
+  The core of the allocator. Every segment contains
+  pages of a certain block size. The main function
+  exported is `mi_malloc_generic`.
+----------------------------------------------------------- */
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+
+/* -----------------------------------------------------------
+  Definition of page queues for each block size
+----------------------------------------------------------- */
+
+#define MI_IN_PAGE_C
+#include "page-queue.c"
+#undef MI_IN_PAGE_C
+
+
+/* -----------------------------------------------------------
+  Page helpers
+----------------------------------------------------------- */
+
+// Index a block in a page
+static inline mi_block_t* mi_page_block_at(const mi_page_t* page, void* page_start, size_t block_size, size_t i) {
+  MI_UNUSED(page);
+  mi_assert_internal(page != NULL);
+  mi_assert_internal(i <= page->reserved);
+  return (mi_block_t*)((uint8_t*)page_start + (i * block_size));
+}
+
+static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t size, mi_tld_t* tld);
+static void mi_page_extend_free(mi_heap_t* heap, mi_page_t* page, mi_tld_t* tld);
+
+#if (MI_DEBUG>=3)
+static size_t mi_page_list_count(mi_page_t* page, mi_block_t* head) {
+  size_t count = 0;
+  while (head != NULL) {
+    mi_assert_internal(page == _mi_ptr_page(head));
+    count++;
+    head = mi_block_next(page, head);
+  }
+  return count;
+}
+
+/*
+// Start of the page available memory
+static inline uint8_t* mi_page_area(const mi_page_t* page) {
+  return _mi_page_start(_mi_page_segment(page), page, NULL);
+}
+*/
+
+static bool mi_page_list_is_valid(mi_page_t* page, mi_block_t* p) {
+  size_t psize;
+  uint8_t* page_area = _mi_page_start(_mi_page_segment(page), page, &psize);
+  mi_block_t* start = (mi_block_t*)page_area;
+  mi_block_t* end   = (mi_block_t*)(page_area + psize);
+  while(p != NULL) {
+    if (p < start || p >= end) return false;
+    p = mi_block_next(page, p);
+  }
+#if MI_DEBUG>3 // generally too expensive to check this
+  if (page->free_is_zero) {
+    const size_t ubsize = mi_page_usable_block_size(page);
+    for (mi_block_t* block = page->free; block != NULL; block = mi_block_next(page, block)) {
+      mi_assert_expensive(mi_mem_is_zero(block + 1, ubsize - sizeof(mi_block_t)));
+    }
+  }
+#endif
+  return true;
+}
+
+static bool mi_page_is_valid_init(mi_page_t* page) {
+  mi_assert_internal(page->xblock_size > 0);
+  mi_assert_internal(page->used <= page->capacity);
+  mi_assert_internal(page->capacity <= page->reserved);
+
+  mi_segment_t* segment = _mi_page_segment(page);
+  uint8_t* start = _mi_page_start(segment,page,NULL);
+  mi_assert_internal(start == _mi_segment_page_start(segment,page,NULL));
+  //const size_t bsize = mi_page_block_size(page);
+  //mi_assert_internal(start + page->capacity*page->block_size == page->top);
+
+  mi_assert_internal(mi_page_list_is_valid(page,page->free));
+  mi_assert_internal(mi_page_list_is_valid(page,page->local_free));
+
+  #if MI_DEBUG>3 // generally too expensive to check this
+  if (page->free_is_zero) {
+    const size_t ubsize = mi_page_usable_block_size(page);
+    for(mi_block_t* block = page->free; block != NULL; block = mi_block_next(page,block)) {
+      mi_assert_expensive(mi_mem_is_zero(block + 1, ubsize - sizeof(mi_block_t)));
+    }
+  }
+  #endif
+
+  #if !MI_TRACK_ENABLED && !MI_TSAN
+  mi_block_t* tfree = mi_page_thread_free(page);
+  mi_assert_internal(mi_page_list_is_valid(page, tfree));
+  //size_t tfree_count = mi_page_list_count(page, tfree);
+  //mi_assert_internal(tfree_count <= page->thread_freed + 1);
+  #endif
+
+  size_t free_count = mi_page_list_count(page, page->free) + mi_page_list_count(page, page->local_free);
+  mi_assert_internal(page->used + free_count == page->capacity);
+
+  return true;
+}
+
+extern bool _mi_process_is_initialized;             // has mi_process_init been called?
+
+bool _mi_page_is_valid(mi_page_t* page) {
+  mi_assert_internal(mi_page_is_valid_init(page));
+  #if MI_SECURE
+  mi_assert_internal(page->keys[0] != 0);
+  #endif
+  if (mi_page_heap(page)!=NULL) {
+    mi_segment_t* segment = _mi_page_segment(page);
+
+    mi_assert_internal(!_mi_process_is_initialized || segment->thread_id==0 || segment->thread_id == mi_page_heap(page)->thread_id);
+    #if MI_HUGE_PAGE_ABANDON
+    if (segment->kind != MI_SEGMENT_HUGE)
+    #endif
+    {
+      mi_page_queue_t* pq = mi_page_queue_of(page);
+      mi_assert_internal(mi_page_queue_contains(pq, page));
+      mi_assert_internal(pq->block_size==mi_page_block_size(page) || mi_page_block_size(page) > MI_MEDIUM_OBJ_SIZE_MAX || mi_page_is_in_full(page));
+      mi_assert_internal(mi_heap_contains_queue(mi_page_heap(page),pq));
+    }
+  }
+  return true;
+}
+#endif
+
+void _mi_page_use_delayed_free(mi_page_t* page, mi_delayed_t delay, bool override_never) {
+  while (!_mi_page_try_use_delayed_free(page, delay, override_never)) {
+    mi_atomic_yield();
+  }
+}
+
+bool _mi_page_try_use_delayed_free(mi_page_t* page, mi_delayed_t delay, bool override_never) {
+  mi_thread_free_t tfreex;
+  mi_delayed_t     old_delay;
+  mi_thread_free_t tfree;
+  size_t yield_count = 0;
+  do {
+    tfree = mi_atomic_load_acquire(&page->xthread_free); // note: must acquire as we can break/repeat this loop and not do a CAS;
+    tfreex = mi_tf_set_delayed(tfree, delay);
+    old_delay = mi_tf_delayed(tfree);
+    if mi_unlikely(old_delay == MI_DELAYED_FREEING) {
+      if (yield_count >= 4) return false;  // give up after 4 tries
+      yield_count++;
+      mi_atomic_yield(); // delay until outstanding MI_DELAYED_FREEING are done.
+      // tfree = mi_tf_set_delayed(tfree, MI_NO_DELAYED_FREE); // will cause CAS to busy fail
+    }
+    else if (delay == old_delay) {
+      break; // avoid atomic operation if already equal
+    }
+    else if (!override_never && old_delay == MI_NEVER_DELAYED_FREE) {
+      break; // leave never-delayed flag set
+    }
+  } while ((old_delay == MI_DELAYED_FREEING) ||
+	   !mi_atomic_cas_weak_release(&page->xthread_free, &tfree, tfreex));
+
+  return true; // success
+}
+
+/* -----------------------------------------------------------
+  Page collect the `local_free` and `thread_free` lists
+----------------------------------------------------------- */
+
+// Collect the local `thread_free` list using an atomic exchange.
+// Note: The exchange must be done atomically as this is used right after
+// moving to the full list in `mi_page_collect_ex` and we need to
+// ensure that there was no race where the page became unfull just before the move.
+static void _mi_page_thread_free_collect(mi_page_t* page)
+{
+  mi_block_t* head;
+  mi_thread_free_t tfreex;
+  mi_thread_free_t tfree = mi_atomic_load_relaxed(&page->xthread_free);
+  do {
+    head = mi_tf_block(tfree);
+    tfreex = mi_tf_set_block(tfree,NULL);
+  } while (!mi_atomic_cas_weak_acq_rel(&page->xthread_free, &tfree, tfreex));
+
+  // return if the list is empty
+  if (head == NULL) return;
+
+  // find the tail -- also to get a proper count (without data races)
+  uint32_t max_count = page->capacity; // cannot collect more than capacity
+  uint32_t count = 1;
+  mi_block_t* tail = head;
+  mi_block_t* next;
+  while ((next = mi_block_next(page,tail)) != NULL && count <= max_count) {
+    count++;
+    tail = next;
+  }
+  // if `count > max_count` there was a memory corruption (possibly infinite list due to double multi-threaded free)
+  if (count > max_count) {
+    _mi_error_message(EFAULT, "corrupted thread-free list\n");
+    return; // the thread-free items cannot be freed
+  }
+
+  // and append the current local free list
+  mi_block_set_next(page,tail, page->local_free);
+  page->local_free = head;
+
+  // update counts now
+  page->used -= count;
+}
+
+void _mi_page_free_collect(mi_page_t* page, bool force) {
+  mi_assert_internal(page!=NULL);
+
+  // collect the thread free list
+  if (force || mi_page_thread_free(page) != NULL) {  // quick test to avoid an atomic operation
+    _mi_page_thread_free_collect(page);
+  }
+
+  // and the local free list
+  if (page->local_free != NULL) {
+    if mi_likely(page->free == NULL) {
+      // usual case
+      page->free = page->local_free;
+      page->local_free = NULL;
+      page->free_is_zero = false;
+    }
+    else if (force) {
+      // append -- only on shutdown (force) as this is a linear operation
+      mi_block_t* tail = page->local_free;
+      mi_block_t* next;
+      while ((next = mi_block_next(page, tail)) != NULL) {
+	tail = next;
+      }
+      mi_block_set_next(page, tail, page->free);
+      page->free = page->local_free;
+      page->local_free = NULL;
+      page->free_is_zero = false;
+    }
+  }
+
+  mi_assert_internal(!force || page->local_free == NULL);
+}
+
+
+
+/* -----------------------------------------------------------
+  Page fresh and retire
+----------------------------------------------------------- */
+
+// called from segments when reclaiming abandoned pages
+void _mi_page_reclaim(mi_heap_t* heap, mi_page_t* page) {
+  mi_assert_expensive(mi_page_is_valid_init(page));
+
+  mi_assert_internal(mi_page_heap(page) == heap);
+  mi_assert_internal(mi_page_thread_free_flag(page) != MI_NEVER_DELAYED_FREE);
+  #if MI_HUGE_PAGE_ABANDON
+  mi_assert_internal(_mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
+  #endif
+
+  // TODO: push on full queue immediately if it is full?
+  mi_page_queue_t* pq = mi_page_queue(heap, mi_page_block_size(page));
+  mi_page_queue_push(heap, pq, page);
+  mi_assert_expensive(_mi_page_is_valid(page));
+}
+
+// allocate a fresh page from a segment
+static mi_page_t* mi_page_fresh_alloc(mi_heap_t* heap, mi_page_queue_t* pq, size_t block_size, size_t page_alignment) {
+  #if !MI_HUGE_PAGE_ABANDON
+  mi_assert_internal(pq != NULL);
+  mi_assert_internal(mi_heap_contains_queue(heap, pq));
+  mi_assert_internal(page_alignment > 0 || block_size > MI_MEDIUM_OBJ_SIZE_MAX || block_size == pq->block_size);
+  #endif
+  mi_page_t* page = _mi_segment_page_alloc(heap, block_size, page_alignment, &heap->tld->segments, &heap->tld->os);
+  if (page == NULL) {
+    // this may be out-of-memory, or an abandoned page was reclaimed (and in our queue)
+    return NULL;
+  }
+  mi_assert_internal(page_alignment >0 || block_size > MI_MEDIUM_OBJ_SIZE_MAX || _mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
+  mi_assert_internal(pq!=NULL || page->xblock_size != 0);
+  mi_assert_internal(pq!=NULL || mi_page_block_size(page) >= block_size);
+  // a fresh page was found, initialize it
+  const size_t full_block_size = ((pq == NULL || mi_page_queue_is_huge(pq)) ? mi_page_block_size(page) : block_size); // see also: mi_segment_huge_page_alloc
+  mi_assert_internal(full_block_size >= block_size);
+  mi_page_init(heap, page, full_block_size, heap->tld);
+  mi_heap_stat_increase(heap, pages, 1);
+  if (pq != NULL) { mi_page_queue_push(heap, pq, page); }
+  mi_assert_expensive(_mi_page_is_valid(page));
+  return page;
+}
+
+// Get a fresh page to use
+static mi_page_t* mi_page_fresh(mi_heap_t* heap, mi_page_queue_t* pq) {
+  mi_assert_internal(mi_heap_contains_queue(heap, pq));
+  mi_page_t* page = mi_page_fresh_alloc(heap, pq, pq->block_size, 0);
+  if (page==NULL) return NULL;
+  mi_assert_internal(pq->block_size==mi_page_block_size(page));
+  mi_assert_internal(pq==mi_page_queue(heap, mi_page_block_size(page)));
+  return page;
+}
+
+/* -----------------------------------------------------------
+   Do any delayed frees
+   (put there by other threads if they deallocated in a full page)
+----------------------------------------------------------- */
+void _mi_heap_delayed_free_all(mi_heap_t* heap) {
+  while (!_mi_heap_delayed_free_partial(heap)) {
+    mi_atomic_yield();
+  }
+}
+
+// returns true if all delayed frees were processed
+bool _mi_heap_delayed_free_partial(mi_heap_t* heap) {
+  // take over the list (note: no atomic exchange since it is often NULL)
+  mi_block_t* block = mi_atomic_load_ptr_relaxed(mi_block_t, &heap->thread_delayed_free);
+  while (block != NULL && !mi_atomic_cas_ptr_weak_acq_rel(mi_block_t, &heap->thread_delayed_free, &block, NULL)) { /* nothing */ };
+  bool all_freed = true;
+
+  // and free them all
+  while(block != NULL) {
+    mi_block_t* next = mi_block_nextx(heap,block, heap->keys);
+    // use internal free instead of regular one to keep stats etc correct
+    if (!_mi_free_delayed_block(block)) {
+      // we might already start delayed freeing while another thread has not yet
+      // reset the delayed_freeing flag; in that case delay it further by reinserting the current block
+      // into the delayed free list
+      all_freed = false;
+      mi_block_t* dfree = mi_atomic_load_ptr_relaxed(mi_block_t, &heap->thread_delayed_free);
+      do {
+	mi_block_set_nextx(heap, block, dfree, heap->keys);
+      } while (!mi_atomic_cas_ptr_weak_release(mi_block_t,&heap->thread_delayed_free, &dfree, block));
+    }
+    block = next;
+  }
+  return all_freed;
+}
+
+/* -----------------------------------------------------------
+  Unfull, abandon, free and retire
+----------------------------------------------------------- */
+
+// Move a page from the full list back to a regular list
+void _mi_page_unfull(mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  mi_assert_internal(mi_page_is_in_full(page));
+  if (!mi_page_is_in_full(page)) return;
+
+  mi_heap_t* heap = mi_page_heap(page);
+  mi_page_queue_t* pqfull = &heap->pages[MI_BIN_FULL];
+  mi_page_set_in_full(page, false); // to get the right queue
+  mi_page_queue_t* pq = mi_heap_page_queue_of(heap, page);
+  mi_page_set_in_full(page, true);
+  mi_page_queue_enqueue_from(pq, pqfull, page);
+}
+
+static void mi_page_to_full(mi_page_t* page, mi_page_queue_t* pq) {
+  mi_assert_internal(pq == mi_page_queue_of(page));
+  mi_assert_internal(!mi_page_immediate_available(page));
+  mi_assert_internal(!mi_page_is_in_full(page));
+
+  if (mi_page_is_in_full(page)) return;
+  mi_page_queue_enqueue_from(&mi_page_heap(page)->pages[MI_BIN_FULL], pq, page);
+  _mi_page_free_collect(page,false);  // try to collect right away in case another thread freed just before MI_USE_DELAYED_FREE was set
+}
+
+
+// Abandon a page with used blocks at the end of a thread.
+// Note: only call if it is ensured that no references exist from
+// the `page->heap->thread_delayed_free` into this page.
+// Currently only called through `mi_heap_collect_ex` which ensures this.
+void _mi_page_abandon(mi_page_t* page, mi_page_queue_t* pq) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  mi_assert_internal(pq == mi_page_queue_of(page));
+  mi_assert_internal(mi_page_heap(page) != NULL);
+
+  mi_heap_t* pheap = mi_page_heap(page);
+
+  // remove from our page list
+  mi_segments_tld_t* segments_tld = &pheap->tld->segments;
+  mi_page_queue_remove(pq, page);
+
+  // page is no longer associated with our heap
+  mi_assert_internal(mi_page_thread_free_flag(page)==MI_NEVER_DELAYED_FREE);
+  mi_page_set_heap(page, NULL);
+
+#if (MI_DEBUG>1) && !MI_TRACK_ENABLED
+  // check there are no references left..
+  for (mi_block_t* block = (mi_block_t*)pheap->thread_delayed_free; block != NULL; block = mi_block_nextx(pheap, block, pheap->keys)) {
+    mi_assert_internal(_mi_ptr_page(block) != page);
+  }
+#endif
+
+  // and abandon it
+  mi_assert_internal(mi_page_heap(page) == NULL);
+  _mi_segment_page_abandon(page,segments_tld);
+}
+
+
+// Free a page with no more free blocks
+void _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  mi_assert_internal(pq == mi_page_queue_of(page));
+  mi_assert_internal(mi_page_all_free(page));
+  mi_assert_internal(mi_page_thread_free_flag(page)!=MI_DELAYED_FREEING);
+
+  // no more aligned blocks in here
+  mi_page_set_has_aligned(page, false);
+
+  mi_heap_t* heap = mi_page_heap(page);
+
+  // remove from the page list
+  // (no need to do _mi_heap_delayed_free first as all blocks are already free)
+  mi_segments_tld_t* segments_tld = &heap->tld->segments;
+  mi_page_queue_remove(pq, page);
+
+  // and free it
+  mi_page_set_heap(page,NULL);
+  _mi_segment_page_free(page, force, segments_tld);
+}
+
+// Retire parameters
+#define MI_MAX_RETIRE_SIZE    (MI_MEDIUM_OBJ_SIZE_MAX)
+#define MI_RETIRE_CYCLES      (16)
+
+// Retire a page with no more used blocks
+// Important to not retire too quickly though as new
+// allocations might coming.
+// Note: called from `mi_free` and benchmarks often
+// trigger this due to freeing everything and then
+// allocating again so careful when changing this.
+void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  mi_assert_internal(mi_page_all_free(page));
+
+  mi_page_set_has_aligned(page, false);
+
+  // don't retire too often..
+  // (or we end up retiring and re-allocating most of the time)
+  // NOTE: refine this more: we should not retire if this
+  // is the only page left with free blocks. It is not clear
+  // how to check this efficiently though...
+  // for now, we don't retire if it is the only page left of this size class.
+  mi_page_queue_t* pq = mi_page_queue_of(page);
+  if mi_likely(page->xblock_size <= MI_MAX_RETIRE_SIZE && !mi_page_queue_is_special(pq)) {  // not too large && not full or huge queue?
+    if (pq->last==page && pq->first==page) { // the only page in the queue?
+      mi_stat_counter_increase(_mi_stats_main.page_no_retire,1);
+      page->retire_expire = 1 + (page->xblock_size <= MI_SMALL_OBJ_SIZE_MAX ? MI_RETIRE_CYCLES : MI_RETIRE_CYCLES/4);
+      mi_heap_t* heap = mi_page_heap(page);
+      mi_assert_internal(pq >= heap->pages);
+      const size_t index = pq - heap->pages;
+      mi_assert_internal(index < MI_BIN_FULL && index < MI_BIN_HUGE);
+      if (index < heap->page_retired_min) heap->page_retired_min = index;
+      if (index > heap->page_retired_max) heap->page_retired_max = index;
+      mi_assert_internal(mi_page_all_free(page));
+      return; // dont't free after all
+    }
+  }
+  _mi_page_free(page, pq, false);
+}
+
+// free retired pages: we don't need to look at the entire queues
+// since we only retire pages that are at the head position in a queue.
+void _mi_heap_collect_retired(mi_heap_t* heap, bool force) {
+  size_t min = MI_BIN_FULL;
+  size_t max = 0;
+  for(size_t bin = heap->page_retired_min; bin <= heap->page_retired_max; bin++) {
+    mi_page_queue_t* pq   = &heap->pages[bin];
+    mi_page_t*       page = pq->first;
+    if (page != NULL && page->retire_expire != 0) {
+      if (mi_page_all_free(page)) {
+	page->retire_expire--;
+	if (force || page->retire_expire == 0) {
+	  _mi_page_free(pq->first, pq, force);
+	}
+	else {
+	  // keep retired, update min/max
+	  if (bin < min) min = bin;
+	  if (bin > max) max = bin;
+	}
+      }
+      else {
+	page->retire_expire = 0;
+      }
+    }
+  }
+  heap->page_retired_min = min;
+  heap->page_retired_max = max;
+}
+
+
+/* -----------------------------------------------------------
+  Initialize the initial free list in a page.
+  In secure mode we initialize a randomized list by
+  alternating between slices.
+----------------------------------------------------------- */
+
+#define MI_MAX_SLICE_SHIFT  (6)   // at most 64 slices
+#define MI_MAX_SLICES       (1UL << MI_MAX_SLICE_SHIFT)
+#define MI_MIN_SLICES       (2)
+
+static void mi_page_free_list_extend_secure(mi_heap_t* const heap, mi_page_t* const page, const size_t bsize, const size_t extend, mi_stats_t* const stats) {
+  MI_UNUSED(stats);
+  #if (MI_SECURE<=2)
+  mi_assert_internal(page->free == NULL);
+  mi_assert_internal(page->local_free == NULL);
+  #endif
+  mi_assert_internal(page->capacity + extend <= page->reserved);
+  mi_assert_internal(bsize == mi_page_block_size(page));
+  void* const page_area = _mi_page_start(_mi_page_segment(page), page, NULL);
+
+  // initialize a randomized free list
+  // set up `slice_count` slices to alternate between
+  size_t shift = MI_MAX_SLICE_SHIFT;
+  while ((extend >> shift) == 0) {
+    shift--;
+  }
+  const size_t slice_count = (size_t)1U << shift;
+  const size_t slice_extend = extend / slice_count;
+  mi_assert_internal(slice_extend >= 1);
+  mi_block_t* blocks[MI_MAX_SLICES];   // current start of the slice
+  size_t      counts[MI_MAX_SLICES];   // available objects in the slice
+  for (size_t i = 0; i < slice_count; i++) {
+    blocks[i] = mi_page_block_at(page, page_area, bsize, page->capacity + i*slice_extend);
+    counts[i] = slice_extend;
+  }
+  counts[slice_count-1] += (extend % slice_count);  // final slice holds the modulus too (todo: distribute evenly?)
+
+  // and initialize the free list by randomly threading through them
+  // set up first element
+  const uintptr_t r = _mi_heap_random_next(heap);
+  size_t current = r % slice_count;
+  counts[current]--;
+  mi_block_t* const free_start = blocks[current];
+  // and iterate through the rest; use `random_shuffle` for performance
+  uintptr_t rnd = _mi_random_shuffle(r|1); // ensure not 0
+  for (size_t i = 1; i < extend; i++) {
+    // call random_shuffle only every INTPTR_SIZE rounds
+    const size_t round = i%MI_INTPTR_SIZE;
+    if (round == 0) rnd = _mi_random_shuffle(rnd);
+    // select a random next slice index
+    size_t next = ((rnd >> 8*round) & (slice_count-1));
+    while (counts[next]==0) {                            // ensure it still has space
+      next++;
+      if (next==slice_count) next = 0;
+    }
+    // and link the current block to it
+    counts[next]--;
+    mi_block_t* const block = blocks[current];
+    blocks[current] = (mi_block_t*)((uint8_t*)block + bsize);  // bump to the following block
+    mi_block_set_next(page, block, blocks[next]);   // and set next; note: we may have `current == next`
+    current = next;
+  }
+  // prepend to the free list (usually NULL)
+  mi_block_set_next(page, blocks[current], page->free);  // end of the list
+  page->free = free_start;
+}
+
+static mi_decl_noinline void mi_page_free_list_extend( mi_page_t* const page, const size_t bsize, const size_t extend, mi_stats_t* const stats)
+{
+  MI_UNUSED(stats);
+  #if (MI_SECURE <= 2)
+  mi_assert_internal(page->free == NULL);
+  mi_assert_internal(page->local_free == NULL);
+  #endif
+  mi_assert_internal(page->capacity + extend <= page->reserved);
+  mi_assert_internal(bsize == mi_page_block_size(page));
+  void* const page_area = _mi_page_start(_mi_page_segment(page), page, NULL );
+
+  mi_block_t* const start = mi_page_block_at(page, page_area, bsize, page->capacity);
+
+  // initialize a sequential free list
+  mi_block_t* const last = mi_page_block_at(page, page_area, bsize, page->capacity + extend - 1);
+  mi_block_t* block = start;
+  while(block <= last) {
+    mi_block_t* next = (mi_block_t*)((uint8_t*)block + bsize);
+    mi_block_set_next(page,block,next);
+    block = next;
+  }
+  // prepend to free list (usually `NULL`)
+  mi_block_set_next(page, last, page->free);
+  page->free = start;
+}
+
+/* -----------------------------------------------------------
+  Page initialize and extend the capacity
+----------------------------------------------------------- */
+
+#define MI_MAX_EXTEND_SIZE    (4*1024)      // heuristic, one OS page seems to work well.
+#if (MI_SECURE>0)
+#define MI_MIN_EXTEND         (8*MI_SECURE) // extend at least by this many
+#else
+#define MI_MIN_EXTEND         (4)
+#endif
+
+// Extend the capacity (up to reserved) by initializing a free list
+// We do at most `MI_MAX_EXTEND` to avoid touching too much memory
+// Note: we also experimented with "bump" allocation on the first
+// allocations but this did not speed up any benchmark (due to an
+// extra test in malloc? or cache effects?)
+static void mi_page_extend_free(mi_heap_t* heap, mi_page_t* page, mi_tld_t* tld) {
+  MI_UNUSED(tld);
+  mi_assert_expensive(mi_page_is_valid_init(page));
+  #if (MI_SECURE<=2)
+  mi_assert(page->free == NULL);
+  mi_assert(page->local_free == NULL);
+  if (page->free != NULL) return;
+  #endif
+  if (page->capacity >= page->reserved) return;
+
+  size_t page_size;
+  _mi_page_start(_mi_page_segment(page), page, &page_size);
+  mi_stat_counter_increase(tld->stats.pages_extended, 1);
+
+  // calculate the extend count
+  const size_t bsize = (page->xblock_size < MI_HUGE_BLOCK_SIZE ? page->xblock_size : page_size);
+  size_t extend = page->reserved - page->capacity;
+  mi_assert_internal(extend > 0);
+
+  size_t max_extend = (bsize >= MI_MAX_EXTEND_SIZE ? MI_MIN_EXTEND : MI_MAX_EXTEND_SIZE/(uint32_t)bsize);
+  if (max_extend < MI_MIN_EXTEND) { max_extend = MI_MIN_EXTEND; }
+  mi_assert_internal(max_extend > 0);
+
+  if (extend > max_extend) {
+    // ensure we don't touch memory beyond the page to reduce page commit.
+    // the `lean` benchmark tests this. Going from 1 to 8 increases rss by 50%.
+    extend = max_extend;
+  }
+
+  mi_assert_internal(extend > 0 && extend + page->capacity <= page->reserved);
+  mi_assert_internal(extend < (1UL<<16));
+
+  // and append the extend the free list
+  if (extend < MI_MIN_SLICES || MI_SECURE==0) { //!mi_option_is_enabled(mi_option_secure)) {
+    mi_page_free_list_extend(page, bsize, extend, &tld->stats );
+  }
+  else {
+    mi_page_free_list_extend_secure(heap, page, bsize, extend, &tld->stats);
+  }
+  // enable the new free list
+  page->capacity += (uint16_t)extend;
+  mi_stat_increase(tld->stats.page_committed, extend * bsize);
+  mi_assert_expensive(mi_page_is_valid_init(page));
+}
+
+// Initialize a fresh page
+static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi_tld_t* tld) {
+  mi_assert(page != NULL);
+  mi_segment_t* segment = _mi_page_segment(page);
+  mi_assert(segment != NULL);
+  mi_assert_internal(block_size > 0);
+  // set fields
+  mi_page_set_heap(page, heap);
+  page->xblock_size = (block_size < MI_HUGE_BLOCK_SIZE ? (uint32_t)block_size : MI_HUGE_BLOCK_SIZE); // initialize before _mi_segment_page_start
+  size_t page_size;
+  const void* page_start = _mi_segment_page_start(segment, page, &page_size);
+  MI_UNUSED(page_start);
+  mi_track_mem_noaccess(page_start,page_size);
+  mi_assert_internal(mi_page_block_size(page) <= page_size);
+  mi_assert_internal(page_size <= page->slice_count*MI_SEGMENT_SLICE_SIZE);
+  mi_assert_internal(page_size / block_size < (1L<<16));
+  page->reserved = (uint16_t)(page_size / block_size);
+  mi_assert_internal(page->reserved > 0);
+  #if (MI_PADDING || MI_ENCODE_FREELIST)
+  page->keys[0] = _mi_heap_random_next(heap);
+  page->keys[1] = _mi_heap_random_next(heap);
+  #endif
+  page->free_is_zero = page->is_zero_init;
+  #if MI_DEBUG>2
+  if (page->is_zero_init) {
+    mi_track_mem_defined(page_start, page_size);
+    mi_assert_expensive(mi_mem_is_zero(page_start, page_size));
+  }
+  #endif
+
+  mi_assert_internal(page->is_committed);
+  mi_assert_internal(page->capacity == 0);
+  mi_assert_internal(page->free == NULL);
+  mi_assert_internal(page->used == 0);
+  mi_assert_internal(page->xthread_free == 0);
+  mi_assert_internal(page->next == NULL);
+  mi_assert_internal(page->prev == NULL);
+  mi_assert_internal(page->retire_expire == 0);
+  mi_assert_internal(!mi_page_has_aligned(page));
+  #if (MI_PADDING || MI_ENCODE_FREELIST)
+  mi_assert_internal(page->keys[0] != 0);
+  mi_assert_internal(page->keys[1] != 0);
+  #endif
+  mi_assert_expensive(mi_page_is_valid_init(page));
+
+  // initialize an initial free list
+  mi_page_extend_free(heap,page,tld);
+  mi_assert(mi_page_immediate_available(page));
+}
+
+
+/* -----------------------------------------------------------
+  Find pages with free blocks
+-------------------------------------------------------------*/
+
+// Find a page with free blocks of `page->block_size`.
+static mi_page_t* mi_page_queue_find_free_ex(mi_heap_t* heap, mi_page_queue_t* pq, bool first_try)
+{
+  // search through the pages in "next fit" order
+  #if MI_STAT
+  size_t count = 0;
+  #endif
+  mi_page_t* page = pq->first;
+  while (page != NULL)
+  {
+    mi_page_t* next = page->next; // remember next
+    #if MI_STAT
+    count++;
+    #endif
+
+    // 0. collect freed blocks by us and other threads
+    _mi_page_free_collect(page, false);
+
+    // 1. if the page contains free blocks, we are done
+    if (mi_page_immediate_available(page)) {
+      break;  // pick this one
+    }
+
+    // 2. Try to extend
+    if (page->capacity < page->reserved) {
+      mi_page_extend_free(heap, page, heap->tld);
+      mi_assert_internal(mi_page_immediate_available(page));
+      break;
+    }
+
+    // 3. If the page is completely full, move it to the `mi_pages_full`
+    // queue so we don't visit long-lived pages too often.
+    mi_assert_internal(!mi_page_is_in_full(page) && !mi_page_immediate_available(page));
+    mi_page_to_full(page, pq);
+
+    page = next;
+  } // for each page
+
+  mi_heap_stat_counter_increase(heap, searches, count);
+
+  if (page == NULL) {
+    _mi_heap_collect_retired(heap, false); // perhaps make a page available?
+    page = mi_page_fresh(heap, pq);
+    if (page == NULL && first_try) {
+      // out-of-memory _or_ an abandoned page with free blocks was reclaimed, try once again
+      page = mi_page_queue_find_free_ex(heap, pq, false);
+    }
+  }
+  else {
+    mi_assert(pq->first == page);
+    page->retire_expire = 0;
+  }
+  mi_assert_internal(page == NULL || mi_page_immediate_available(page));
+  return page;
+}
+
+
+
+// Find a page with free blocks of `size`.
+static inline mi_page_t* mi_find_free_page(mi_heap_t* heap, size_t size) {
+  mi_page_queue_t* pq = mi_page_queue(heap,size);
+  mi_page_t* page = pq->first;
+  if (page != NULL) {
+   #if (MI_SECURE>=3) // in secure mode, we extend half the time to increase randomness
+    if (page->capacity < page->reserved && ((_mi_heap_random_next(heap) & 1) == 1)) {
+      mi_page_extend_free(heap, page, heap->tld);
+      mi_assert_internal(mi_page_immediate_available(page));
+    }
+    else
+   #endif
+    {
+      _mi_page_free_collect(page,false);
+    }
+
+    if (mi_page_immediate_available(page)) {
+      page->retire_expire = 0;
+      return page; // fast path
+    }
+  }
+  return mi_page_queue_find_free_ex(heap, pq, true);
+}
+
+
+/* -----------------------------------------------------------
+  Users can register a deferred free function called
+  when the `free` list is empty. Since the `local_free`
+  is separate this is deterministically called after
+  a certain number of allocations.
+----------------------------------------------------------- */
+
+static mi_deferred_free_fun* volatile deferred_free = NULL;
+static _Atomic(void*) deferred_arg; // = NULL
+
+void _mi_deferred_free(mi_heap_t* heap, bool force) {
+  heap->tld->heartbeat++;
+  if (deferred_free != NULL && !heap->tld->recurse) {
+    heap->tld->recurse = true;
+    deferred_free(force, heap->tld->heartbeat, mi_atomic_load_ptr_relaxed(void,&deferred_arg));
+    heap->tld->recurse = false;
+  }
+}
+
+void mi_register_deferred_free(mi_deferred_free_fun* fn, void* arg) mi_attr_noexcept {
+  deferred_free = fn;
+  mi_atomic_store_ptr_release(void,&deferred_arg, arg);
+}
+
+
+/* -----------------------------------------------------------
+  General allocation
+----------------------------------------------------------- */
+
+// Large and huge page allocation.
+// Huge pages are allocated directly without being in a queue.
+// Because huge pages contain just one block, and the segment contains
+// just that page, we always treat them as abandoned and any thread
+// that frees the block can free the whole page and segment directly.
+// Huge pages are also use if the requested alignment is very large (> MI_ALIGNMENT_MAX).
+static mi_page_t* mi_large_huge_page_alloc(mi_heap_t* heap, size_t size, size_t page_alignment) {
+  size_t block_size = _mi_os_good_alloc_size(size);
+  mi_assert_internal(mi_bin(block_size) == MI_BIN_HUGE || page_alignment > 0);
+  bool is_huge = (block_size > MI_LARGE_OBJ_SIZE_MAX || page_alignment > 0);
+  #if MI_HUGE_PAGE_ABANDON
+  mi_page_queue_t* pq = (is_huge ? NULL : mi_page_queue(heap, block_size));
+  #else
+  mi_page_queue_t* pq = mi_page_queue(heap, is_huge ? MI_HUGE_BLOCK_SIZE : block_size); // not block_size as that can be low if the page_alignment > 0
+  mi_assert_internal(!is_huge || mi_page_queue_is_huge(pq));
+  #endif
+  mi_page_t* page = mi_page_fresh_alloc(heap, pq, block_size, page_alignment);
+  if (page != NULL) {
+    mi_assert_internal(mi_page_immediate_available(page));
+
+    if (is_huge) {
+      mi_assert_internal(_mi_page_segment(page)->kind == MI_SEGMENT_HUGE);
+      mi_assert_internal(_mi_page_segment(page)->used==1);
+      #if MI_HUGE_PAGE_ABANDON
+      mi_assert_internal(_mi_page_segment(page)->thread_id==0); // abandoned, not in the huge queue
+      mi_page_set_heap(page, NULL);
+      #endif
+    }
+    else {
+      mi_assert_internal(_mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
+    }
+
+    const size_t bsize = mi_page_usable_block_size(page);  // note: not `mi_page_block_size` to account for padding
+    if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
+      mi_heap_stat_increase(heap, large, bsize);
+      mi_heap_stat_counter_increase(heap, large_count, 1);
+    }
+    else {
+      mi_heap_stat_increase(heap, huge, bsize);
+      mi_heap_stat_counter_increase(heap, huge_count, 1);
+    }
+  }
+  return page;
+}
+
+
+// Allocate a page
+// Note: in debug mode the size includes MI_PADDING_SIZE and might have overflowed.
+static mi_page_t* mi_find_page(mi_heap_t* heap, size_t size, size_t huge_alignment) mi_attr_noexcept {
+  // huge allocation?
+  const size_t req_size = size - MI_PADDING_SIZE;  // correct for padding_size in case of an overflow on `size`
+  if mi_unlikely(req_size > (MI_MEDIUM_OBJ_SIZE_MAX - MI_PADDING_SIZE) || huge_alignment > 0) {
+    if mi_unlikely(req_size > PTRDIFF_MAX) {  // we don't allocate more than PTRDIFF_MAX (see <https://sourceware.org/ml/libc-announce/2019/msg00001.html>)
+      _mi_error_message(EOVERFLOW, "allocation request is too large (%zu bytes)\n", req_size);
+      return NULL;
+    }
+    else {
+      return mi_large_huge_page_alloc(heap,size,huge_alignment);
+    }
+  }
+  else {
+    // otherwise find a page with free blocks in our size segregated queues
+    #if MI_PADDING
+    mi_assert_internal(size >= MI_PADDING_SIZE);
+    #endif
+    return mi_find_free_page(heap, size);
+  }
+}
+
+// Generic allocation routine if the fast path (`alloc.c:mi_page_malloc`) does not succeed.
+// Note: in debug mode the size includes MI_PADDING_SIZE and might have overflowed.
+// The `huge_alignment` is normally 0 but is set to a multiple of MI_SEGMENT_SIZE for
+// very large requested alignments in which case we use a huge segment.
+void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept
+{
+  mi_assert_internal(heap != NULL);
+
+  // initialize if necessary
+  if mi_unlikely(!mi_heap_is_initialized(heap)) {
+    heap = mi_heap_get_default(); // calls mi_thread_init
+    if mi_unlikely(!mi_heap_is_initialized(heap)) { return NULL; }
+  }
+  mi_assert_internal(mi_heap_is_initialized(heap));
+
+  // call potential deferred free routines
+  _mi_deferred_free(heap, false);
+
+  // free delayed frees from other threads (but skip contended ones)
+  _mi_heap_delayed_free_partial(heap);
+
+  // find (or allocate) a page of the right size
+  mi_page_t* page = mi_find_page(heap, size, huge_alignment);
+  if mi_unlikely(page == NULL) { // first time out of memory, try to collect and retry the allocation once more
+    mi_heap_collect(heap, true /* force */);
+    page = mi_find_page(heap, size, huge_alignment);
+  }
+
+  if mi_unlikely(page == NULL) { // out of memory
+    const size_t req_size = size - MI_PADDING_SIZE;  // correct for padding_size in case of an overflow on `size`
+    _mi_error_message(ENOMEM, "unable to allocate memory (%zu bytes)\n", req_size);
+    return NULL;
+  }
+
+  mi_assert_internal(mi_page_immediate_available(page));
+  mi_assert_internal(mi_page_block_size(page) >= size);
+
+  // and try again, this time succeeding! (i.e. this should never recurse through _mi_page_malloc)
+  if mi_unlikely(zero && page->xblock_size == 0) {
+    // note: we cannot call _mi_page_malloc with zeroing for huge blocks; we zero it afterwards in that case.
+    void* p = _mi_page_malloc(heap, page, size, false);
+    mi_assert_internal(p != NULL);
+    _mi_memzero_aligned(p, mi_page_usable_block_size(page));
+    return p;
+  }
+  else {
+    return _mi_page_malloc(heap, page, size, zero);
+  }
+}
diff --git a/compat/mimalloc/prim/windows/prim.c b/compat/mimalloc/prim/windows/prim.c
new file mode 100644
index 00000000000000..d060833c5b644d
--- /dev/null
+++ b/compat/mimalloc/prim/windows/prim.c
@@ -0,0 +1,622 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+// This file is included in `src/prim/prim.c`
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"
+#include <stdio.h>   // fputs, stderr
+
+
+//---------------------------------------------
+// Dynamically bind Windows API points for portability
+//---------------------------------------------
+
+// We use VirtualAlloc2 for aligned allocation, but it is only supported on Windows 10 and Windows Server 2016.
+// So, we need to look it up dynamically to run on older systems. (use __stdcall for 32-bit compatibility)
+// NtAllocateVirtualAllocEx is used for huge OS page allocation (1GiB)
+// We define a minimal MEM_EXTENDED_PARAMETER ourselves in order to be able to compile with older SDK's.
+typedef enum MI_MEM_EXTENDED_PARAMETER_TYPE_E {
+  MiMemExtendedParameterInvalidType = 0,
+  MiMemExtendedParameterAddressRequirements,
+  MiMemExtendedParameterNumaNode,
+  MiMemExtendedParameterPartitionHandle,
+  MiMemExtendedParameterUserPhysicalHandle,
+  MiMemExtendedParameterAttributeFlags,
+  MiMemExtendedParameterMax
+} MI_MEM_EXTENDED_PARAMETER_TYPE;
+
+typedef struct DECLSPEC_ALIGN(8) MI_MEM_EXTENDED_PARAMETER_S {
+  struct { DWORD64 Type : 8; DWORD64 Reserved : 56; } Type;
+  union  { DWORD64 ULong64; PVOID Pointer; SIZE_T Size; HANDLE Handle; DWORD ULong; } Arg;
+} MI_MEM_EXTENDED_PARAMETER;
+
+typedef struct MI_MEM_ADDRESS_REQUIREMENTS_S {
+  PVOID  LowestStartingAddress;
+  PVOID  HighestEndingAddress;
+  SIZE_T Alignment;
+} MI_MEM_ADDRESS_REQUIREMENTS;
+
+#define MI_MEM_EXTENDED_PARAMETER_NONPAGED_HUGE   0x00000010
+
+#include <winternl.h>
+typedef PVOID    (__stdcall *PVirtualAlloc2)(HANDLE, PVOID, SIZE_T, ULONG, ULONG, MI_MEM_EXTENDED_PARAMETER*, ULONG);
+typedef NTSTATUS (__stdcall *PNtAllocateVirtualMemoryEx)(HANDLE, PVOID*, SIZE_T*, ULONG, ULONG, MI_MEM_EXTENDED_PARAMETER*, ULONG);
+static PVirtualAlloc2 pVirtualAlloc2 = NULL;
+static PNtAllocateVirtualMemoryEx pNtAllocateVirtualMemoryEx = NULL;
+
+// Similarly, GetNumaProcesorNodeEx is only supported since Windows 7
+typedef struct MI_PROCESSOR_NUMBER_S { WORD Group; BYTE Number; BYTE Reserved; } MI_PROCESSOR_NUMBER;
+
+typedef VOID (__stdcall *PGetCurrentProcessorNumberEx)(MI_PROCESSOR_NUMBER* ProcNumber);
+typedef BOOL (__stdcall *PGetNumaProcessorNodeEx)(MI_PROCESSOR_NUMBER* Processor, PUSHORT NodeNumber);
+typedef BOOL (__stdcall* PGetNumaNodeProcessorMaskEx)(USHORT Node, PGROUP_AFFINITY ProcessorMask);
+typedef BOOL (__stdcall *PGetNumaProcessorNode)(UCHAR Processor, PUCHAR NodeNumber);
+static PGetCurrentProcessorNumberEx pGetCurrentProcessorNumberEx = NULL;
+static PGetNumaProcessorNodeEx      pGetNumaProcessorNodeEx = NULL;
+static PGetNumaNodeProcessorMaskEx  pGetNumaNodeProcessorMaskEx = NULL;
+static PGetNumaProcessorNode        pGetNumaProcessorNode = NULL;
+
+//---------------------------------------------
+// Enable large page support dynamically (if possible)
+//---------------------------------------------
+
+static bool win_enable_large_os_pages(size_t* large_page_size)
+{
+  static bool large_initialized = false;
+  if (large_initialized) return (_mi_os_large_page_size() > 0);
+  large_initialized = true;
+
+  // Try to see if large OS pages are supported
+  // To use large pages on Windows, we first need access permission
+  // Set "Lock pages in memory" permission in the group policy editor
+  // <https://devblogs.microsoft.com/oldnewthing/20110128-00/?p=11643>
+  unsigned long err = 0;
+  HANDLE token = NULL;
+  BOOL ok = OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &token);
+  if (ok) {
+    TOKEN_PRIVILEGES tp;
+    ok = LookupPrivilegeValue(NULL, TEXT("SeLockMemoryPrivilege"), &tp.Privileges[0].Luid);
+    if (ok) {
+      tp.PrivilegeCount = 1;
+      tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
+      ok = AdjustTokenPrivileges(token, FALSE, &tp, 0, (PTOKEN_PRIVILEGES)NULL, 0);
+      if (ok) {
+	err = GetLastError();
+	ok = (err == ERROR_SUCCESS);
+	if (ok && large_page_size != NULL) {
+	  *large_page_size = GetLargePageMinimum();
+	}
+      }
+    }
+    CloseHandle(token);
+  }
+  if (!ok) {
+    if (err == 0) err = GetLastError();
+    _mi_warning_message("cannot enable large OS page support, error %lu\n", err);
+  }
+  return (ok!=0);
+}
+
+
+//---------------------------------------------
+// Initialize
+//---------------------------------------------
+
+void _mi_prim_mem_init( mi_os_mem_config_t* config )
+{
+  config->has_overcommit = false;
+  config->must_free_whole = true;
+  config->has_virtual_reserve = true;
+  // get the page size
+  SYSTEM_INFO si;
+  GetSystemInfo(&si);
+  if (si.dwPageSize > 0) { config->page_size = si.dwPageSize; }
+  if (si.dwAllocationGranularity > 0) { config->alloc_granularity = si.dwAllocationGranularity; }
+  // get the VirtualAlloc2 function
+  HINSTANCE  hDll;
+  hDll = LoadLibrary(TEXT("kernelbase.dll"));
+  if (hDll != NULL) {
+    // use VirtualAlloc2FromApp if possible as it is available to Windows store apps
+    pVirtualAlloc2 = (PVirtualAlloc2)(void (*)(void))GetProcAddress(hDll, "VirtualAlloc2FromApp");
+    if (pVirtualAlloc2==NULL) pVirtualAlloc2 = (PVirtualAlloc2)(void (*)(void))GetProcAddress(hDll, "VirtualAlloc2");
+    FreeLibrary(hDll);
+  }
+  // NtAllocateVirtualMemoryEx is used for huge page allocation
+  hDll = LoadLibrary(TEXT("ntdll.dll"));
+  if (hDll != NULL) {
+    pNtAllocateVirtualMemoryEx = (PNtAllocateVirtualMemoryEx)(void (*)(void))GetProcAddress(hDll, "NtAllocateVirtualMemoryEx");
+    FreeLibrary(hDll);
+  }
+  // Try to use Win7+ numa API
+  hDll = LoadLibrary(TEXT("kernel32.dll"));
+  if (hDll != NULL) {
+    pGetCurrentProcessorNumberEx = (PGetCurrentProcessorNumberEx)(void (*)(void))GetProcAddress(hDll, "GetCurrentProcessorNumberEx");
+    pGetNumaProcessorNodeEx = (PGetNumaProcessorNodeEx)(void (*)(void))GetProcAddress(hDll, "GetNumaProcessorNodeEx");
+    pGetNumaNodeProcessorMaskEx = (PGetNumaNodeProcessorMaskEx)(void (*)(void))GetProcAddress(hDll, "GetNumaNodeProcessorMaskEx");
+    pGetNumaProcessorNode = (PGetNumaProcessorNode)(void (*)(void))GetProcAddress(hDll, "GetNumaProcessorNode");
+    FreeLibrary(hDll);
+  }
+  if (mi_option_is_enabled(mi_option_allow_large_os_pages) || mi_option_is_enabled(mi_option_reserve_huge_os_pages)) {
+    win_enable_large_os_pages(&config->large_page_size);
+  }
+}
+
+
+//---------------------------------------------
+// Free
+//---------------------------------------------
+
+int _mi_prim_free(void* addr, size_t size ) {
+  MI_UNUSED(size);
+  DWORD errcode = 0;
+  bool err = (VirtualFree(addr, 0, MEM_RELEASE) == 0);
+  if (err) { errcode = GetLastError(); }
+  if (errcode == ERROR_INVALID_ADDRESS) {
+    // In mi_os_mem_alloc_aligned the fallback path may have returned a pointer inside
+    // the memory region returned by VirtualAlloc; in that case we need to free using
+    // the start of the region.
+    MEMORY_BASIC_INFORMATION info = { 0 };
+    VirtualQuery(addr, &info, sizeof(info));
+    if (info.AllocationBase < addr && ((uint8_t*)addr - (uint8_t*)info.AllocationBase) < (ptrdiff_t)MI_SEGMENT_SIZE) {
+      errcode = 0;
+      err = (VirtualFree(info.AllocationBase, 0, MEM_RELEASE) == 0);
+      if (err) { errcode = GetLastError(); }
+    }
+  }
+  return (int)errcode;
+}
+
+
+//---------------------------------------------
+// VirtualAlloc
+//---------------------------------------------
+
+static void* win_virtual_alloc_prim(void* addr, size_t size, size_t try_alignment, DWORD flags) {
+  #if (MI_INTPTR_SIZE >= 8)
+  // on 64-bit systems, try to use the virtual address area after 2TiB for 4MiB aligned allocations
+  if (addr == NULL) {
+    void* hint = _mi_os_get_aligned_hint(try_alignment,size);
+    if (hint != NULL) {
+      void* p = VirtualAlloc(hint, size, flags, PAGE_READWRITE);
+      if (p != NULL) return p;
+      _mi_verbose_message("warning: unable to allocate hinted aligned OS memory (%zu bytes, error code: 0x%x, address: %p, alignment: %zu, flags: 0x%x)\n", size, GetLastError(), hint, try_alignment, flags);
+      // fall through on error
+    }
+  }
+  #endif
+  // on modern Windows try use VirtualAlloc2 for aligned allocation
+  if (try_alignment > 1 && (try_alignment % _mi_os_page_size()) == 0 && pVirtualAlloc2 != NULL) {
+    MI_MEM_ADDRESS_REQUIREMENTS reqs = { 0, 0, 0 };
+    reqs.Alignment = try_alignment;
+    MI_MEM_EXTENDED_PARAMETER param = { {0, 0}, {0} };
+    param.Type.Type = MiMemExtendedParameterAddressRequirements;
+    param.Arg.Pointer = &reqs;
+    void* p = (*pVirtualAlloc2)(GetCurrentProcess(), addr, size, flags, PAGE_READWRITE, &param, 1);
+    if (p != NULL) return p;
+    _mi_warning_message("unable to allocate aligned OS memory (%zu bytes, error code: 0x%x, address: %p, alignment: %zu, flags: 0x%x)\n", size, GetLastError(), addr, try_alignment, flags);
+    // fall through on error
+  }
+  // last resort
+  return VirtualAlloc(addr, size, flags, PAGE_READWRITE);
+}
+
+static void* win_virtual_alloc(void* addr, size_t size, size_t try_alignment, DWORD flags, bool large_only, bool allow_large, bool* is_large) {
+  mi_assert_internal(!(large_only && !allow_large));
+  static _Atomic(size_t) large_page_try_ok; // = 0;
+  void* p = NULL;
+  // Try to allocate large OS pages (2MiB) if allowed or required.
+  if ((large_only || _mi_os_use_large_page(size, try_alignment))
+      && allow_large && (flags&MEM_COMMIT)!=0 && (flags&MEM_RESERVE)!=0) {
+    size_t try_ok = mi_atomic_load_acquire(&large_page_try_ok);
+    if (!large_only && try_ok > 0) {
+      // if a large page allocation fails, it seems the calls to VirtualAlloc get very expensive.
+      // therefore, once a large page allocation failed, we don't try again for `large_page_try_ok` times.
+      mi_atomic_cas_strong_acq_rel(&large_page_try_ok, &try_ok, try_ok - 1);
+    }
+    else {
+      // large OS pages must always reserve and commit.
+      *is_large = true;
+      p = win_virtual_alloc_prim(addr, size, try_alignment, flags | MEM_LARGE_PAGES);
+      if (large_only) return p;
+      // fall back to non-large page allocation on error (`p == NULL`).
+      if (p == NULL) {
+	mi_atomic_store_release(&large_page_try_ok,10UL);  // on error, don't try again for the next N allocations
+      }
+    }
+  }
+  // Fall back to regular page allocation
+  if (p == NULL) {
+    *is_large = ((flags&MEM_LARGE_PAGES) != 0);
+    p = win_virtual_alloc_prim(addr, size, try_alignment, flags);
+  }
+  //if (p == NULL) { _mi_warning_message("unable to allocate OS memory (%zu bytes, error code: 0x%x, address: %p, alignment: %zu, flags: 0x%x, large only: %d, allow large: %d)\n", size, GetLastError(), addr, try_alignment, flags, large_only, allow_large); }
+  return p;
+}
+
+int _mi_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, void** addr) {
+  mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
+  mi_assert_internal(commit || !allow_large);
+  mi_assert_internal(try_alignment > 0);
+  *is_zero = true;
+  int flags = MEM_RESERVE;
+  if (commit) { flags |= MEM_COMMIT; }
+  *addr = win_virtual_alloc(NULL, size, try_alignment, flags, false, allow_large, is_large);
+  return (*addr != NULL ? 0 : (int)GetLastError());
+}
+
+
+//---------------------------------------------
+// Commit/Reset/Protect
+//---------------------------------------------
+#ifdef _MSC_VER
+#pragma warning(disable:6250)   // suppress warning calling VirtualFree without MEM_RELEASE (for decommit)
+#endif
+
+int _mi_prim_commit(void* addr, size_t size, bool* is_zero) {
+  *is_zero = false;
+  /*
+  // zero'ing only happens on an initial commit... but checking upfront seems expensive..
+  _MEMORY_BASIC_INFORMATION meminfo; _mi_memzero_var(meminfo);
+  if (VirtualQuery(addr, &meminfo, size) > 0) {
+    if ((meminfo.State & MEM_COMMIT) == 0) {
+      *is_zero = true;
+    }
+  }
+  */
+  // commit
+  void* p = VirtualAlloc(addr, size, MEM_COMMIT, PAGE_READWRITE);
+  if (p == NULL) return (int)GetLastError();
+  return 0;
+}
+
+int _mi_prim_decommit(void* addr, size_t size, bool* needs_recommit) {
+  BOOL ok = VirtualFree(addr, size, MEM_DECOMMIT);
+  *needs_recommit = true;  // for safety, assume always decommitted even in the case of an error.
+  return (ok ? 0 : (int)GetLastError());
+}
+
+int _mi_prim_reset(void* addr, size_t size) {
+  void* p = VirtualAlloc(addr, size, MEM_RESET, PAGE_READWRITE);
+  mi_assert_internal(p == addr);
+  #if 0
+  if (p != NULL) {
+    VirtualUnlock(addr,size); // VirtualUnlock after MEM_RESET removes the memory directly from the working set
+  }
+  #endif
+  return (p != NULL ? 0 : (int)GetLastError());
+}
+
+int _mi_prim_protect(void* addr, size_t size, bool protect) {
+  DWORD oldprotect = 0;
+  BOOL ok = VirtualProtect(addr, size, protect ? PAGE_NOACCESS : PAGE_READWRITE, &oldprotect);
+  return (ok ? 0 : (int)GetLastError());
+}
+
+
+//---------------------------------------------
+// Huge page allocation
+//---------------------------------------------
+
+static void* _mi_prim_alloc_huge_os_pagesx(void* hint_addr, size_t size, int numa_node)
+{
+  const DWORD flags = MEM_LARGE_PAGES | MEM_COMMIT | MEM_RESERVE;
+
+  win_enable_large_os_pages(NULL);
+
+  MI_MEM_EXTENDED_PARAMETER params[3] = { {{0,0},{0}},{{0,0},{0}},{{0,0},{0}} };
+  // on modern Windows try use NtAllocateVirtualMemoryEx for 1GiB huge pages
+  static bool mi_huge_pages_available = true;
+  if (pNtAllocateVirtualMemoryEx != NULL && mi_huge_pages_available) {
+    params[0].Type.Type = MiMemExtendedParameterAttributeFlags;
+    params[0].Arg.ULong64 = MI_MEM_EXTENDED_PARAMETER_NONPAGED_HUGE;
+    ULONG param_count = 1;
+    if (numa_node >= 0) {
+      param_count++;
+      params[1].Type.Type = MiMemExtendedParameterNumaNode;
+      params[1].Arg.ULong = (unsigned)numa_node;
+    }
+    SIZE_T psize = size;
+    void* base = hint_addr;
+    NTSTATUS err = (*pNtAllocateVirtualMemoryEx)(GetCurrentProcess(), &base, &psize, flags, PAGE_READWRITE, params, param_count);
+    if (err == 0 && base != NULL) {
+      return base;
+    }
+    else {
+      // fall back to regular large pages
+      mi_huge_pages_available = false; // don't try further huge pages
+      _mi_warning_message("unable to allocate using huge (1GiB) pages, trying large (2MiB) pages instead (status 0x%lx)\n", err);
+    }
+  }
+  // on modern Windows try use VirtualAlloc2 for numa aware large OS page allocation
+  if (pVirtualAlloc2 != NULL && numa_node >= 0) {
+    params[0].Type.Type = MiMemExtendedParameterNumaNode;
+    params[0].Arg.ULong = (unsigned)numa_node;
+    return (*pVirtualAlloc2)(GetCurrentProcess(), hint_addr, size, flags, PAGE_READWRITE, params, 1);
+  }
+
+  // otherwise use regular virtual alloc on older windows
+  return VirtualAlloc(hint_addr, size, flags, PAGE_READWRITE);
+}
+
+int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, bool* is_zero, void** addr) {
+  *is_zero = true;
+  *addr = _mi_prim_alloc_huge_os_pagesx(hint_addr,size,numa_node);
+  return (*addr != NULL ? 0 : (int)GetLastError());
+}
+
+
+//---------------------------------------------
+// Numa nodes
+//---------------------------------------------
+
+size_t _mi_prim_numa_node(void) {
+  USHORT numa_node = 0;
+  if (pGetCurrentProcessorNumberEx != NULL && pGetNumaProcessorNodeEx != NULL) {
+    // Extended API is supported
+    MI_PROCESSOR_NUMBER pnum;
+    (*pGetCurrentProcessorNumberEx)(&pnum);
+    USHORT nnode = 0;
+    BOOL ok = (*pGetNumaProcessorNodeEx)(&pnum, &nnode);
+    if (ok) { numa_node = nnode; }
+  }
+  else if (pGetNumaProcessorNode != NULL) {
+    // Vista or earlier, use older API that is limited to 64 processors. Issue #277
+    DWORD pnum = GetCurrentProcessorNumber();
+    UCHAR nnode = 0;
+    BOOL ok = pGetNumaProcessorNode((UCHAR)pnum, &nnode);
+    if (ok) { numa_node = nnode; }
+  }
+  return numa_node;
+}
+
+size_t _mi_prim_numa_node_count(void) {
+  ULONG numa_max = 0;
+  GetNumaHighestNodeNumber(&numa_max);
+  // find the highest node number that has actual processors assigned to it. Issue #282
+  while(numa_max > 0) {
+    if (pGetNumaNodeProcessorMaskEx != NULL) {
+      // Extended API is supported
+      GROUP_AFFINITY affinity;
+      if ((*pGetNumaNodeProcessorMaskEx)((USHORT)numa_max, &affinity)) {
+	if (affinity.Mask != 0) break;  // found the maximum non-empty node
+      }
+    }
+    else {
+      // Vista or earlier, use older API that is limited to 64 processors.
+      ULONGLONG mask;
+      if (GetNumaNodeProcessorMask((UCHAR)numa_max, &mask)) {
+	if (mask != 0) break; // found the maximum non-empty node
+      };
+    }
+    // max node was invalid or had no processor assigned, try again
+    numa_max--;
+  }
+  return ((size_t)numa_max + 1);
+}
+
+
+//----------------------------------------------------------------
+// Clock
+//----------------------------------------------------------------
+
+static mi_msecs_t mi_to_msecs(LARGE_INTEGER t) {
+  static LARGE_INTEGER mfreq; // = 0
+  if (mfreq.QuadPart == 0LL) {
+    LARGE_INTEGER f;
+    QueryPerformanceFrequency(&f);
+    mfreq.QuadPart = f.QuadPart/1000LL;
+    if (mfreq.QuadPart == 0) mfreq.QuadPart = 1;
+  }
+  return (mi_msecs_t)(t.QuadPart / mfreq.QuadPart);
+}
+
+mi_msecs_t _mi_prim_clock_now(void) {
+  LARGE_INTEGER t;
+  QueryPerformanceCounter(&t);
+  return mi_to_msecs(t);
+}
+
+
+//----------------------------------------------------------------
+// Process Info
+//----------------------------------------------------------------
+
+#include <windows.h>
+#include <psapi.h>
+
+static mi_msecs_t filetime_msecs(const FILETIME* ftime) {
+  ULARGE_INTEGER i;
+  i.LowPart = ftime->dwLowDateTime;
+  i.HighPart = ftime->dwHighDateTime;
+  mi_msecs_t msecs = (i.QuadPart / 10000); // FILETIME is in 100 nano seconds
+  return msecs;
+}
+
+typedef BOOL (WINAPI *PGetProcessMemoryInfo)(HANDLE, PPROCESS_MEMORY_COUNTERS, DWORD);
+static PGetProcessMemoryInfo pGetProcessMemoryInfo = NULL;
+
+void _mi_prim_process_info(mi_process_info_t* pinfo)
+{
+  FILETIME ct;
+  FILETIME ut;
+  FILETIME st;
+  FILETIME et;
+  GetProcessTimes(GetCurrentProcess(), &ct, &et, &st, &ut);
+  pinfo->utime = filetime_msecs(&ut);
+  pinfo->stime = filetime_msecs(&st);
+
+  // load psapi on demand
+  if (pGetProcessMemoryInfo == NULL) {
+    HINSTANCE hDll = LoadLibrary(TEXT("psapi.dll"));
+    if (hDll != NULL) {
+      pGetProcessMemoryInfo = (PGetProcessMemoryInfo)(void (*)(void))GetProcAddress(hDll, "GetProcessMemoryInfo");
+    }
+  }
+
+  // get process info
+  PROCESS_MEMORY_COUNTERS info;
+  memset(&info, 0, sizeof(info));
+  if (pGetProcessMemoryInfo != NULL) {
+    pGetProcessMemoryInfo(GetCurrentProcess(), &info, sizeof(info));
+  }
+  pinfo->current_rss    = (size_t)info.WorkingSetSize;
+  pinfo->peak_rss       = (size_t)info.PeakWorkingSetSize;
+  pinfo->current_commit = (size_t)info.PagefileUsage;
+  pinfo->peak_commit    = (size_t)info.PeakPagefileUsage;
+  pinfo->page_faults    = (size_t)info.PageFaultCount;
+}
+
+//----------------------------------------------------------------
+// Output
+//----------------------------------------------------------------
+
+void _mi_prim_out_stderr( const char* msg )
+{
+  // on windows with redirection, the C runtime cannot handle locale dependent output
+  // after the main thread closes so we use direct console output.
+  if (!_mi_preloading()) {
+    // _cputs(msg);  // _cputs cannot be used at is aborts if it fails to lock the console
+    static HANDLE hcon = INVALID_HANDLE_VALUE;
+    static bool hconIsConsole;
+    if (hcon == INVALID_HANDLE_VALUE) {
+      CONSOLE_SCREEN_BUFFER_INFO sbi;
+      hcon = GetStdHandle(STD_ERROR_HANDLE);
+      hconIsConsole = ((hcon != INVALID_HANDLE_VALUE) && GetConsoleScreenBufferInfo(hcon, &sbi));
+    }
+    const size_t len = _mi_strlen(msg);
+    if (len > 0 && len < UINT32_MAX) {
+      DWORD written = 0;
+      if (hconIsConsole) {
+	WriteConsoleA(hcon, msg, (DWORD)len, &written, NULL);
+      }
+      else if (hcon != INVALID_HANDLE_VALUE) {
+	// use direct write if stderr was redirected
+	WriteFile(hcon, msg, (DWORD)len, &written, NULL);
+      }
+      else {
+	// finally fall back to fputs after all
+	fputs(msg, stderr);
+      }
+    }
+  }
+}
+
+
+//----------------------------------------------------------------
+// Environment
+//----------------------------------------------------------------
+
+// On Windows use GetEnvironmentVariable instead of getenv to work
+// reliably even when this is invoked before the C runtime is initialized.
+// i.e. when `_mi_preloading() == true`.
+// Note: on windows, environment names are not case sensitive.
+bool _mi_prim_getenv(const char* name, char* result, size_t result_size) {
+  result[0] = 0;
+  size_t len = GetEnvironmentVariableA(name, result, (DWORD)result_size);
+  return (len > 0 && len < result_size);
+}
+
+
+
+//----------------------------------------------------------------
+// Random
+//----------------------------------------------------------------
+
+#if defined(MI_USE_RTLGENRANDOM) // || defined(__cplusplus)
+// We prefer to use BCryptGenRandom instead of (the unofficial) RtlGenRandom but when using
+// dynamic overriding, we observed it can raise an exception when compiled with C++, and
+// sometimes deadlocks when also running under the VS debugger.
+// In contrast, issue #623 implies that on Windows Server 2019 we need to use BCryptGenRandom.
+// To be continued..
+#pragma comment (lib,"advapi32.lib")
+#define RtlGenRandom  SystemFunction036
+mi_decl_externc BOOLEAN NTAPI RtlGenRandom(PVOID RandomBuffer, ULONG RandomBufferLength);
+
+bool _mi_prim_random_buf(void* buf, size_t buf_len) {
+  return (RtlGenRandom(buf, (ULONG)buf_len) != 0);
+}
+
+#else
+
+#ifndef BCRYPT_USE_SYSTEM_PREFERRED_RNG
+#define BCRYPT_USE_SYSTEM_PREFERRED_RNG 0x00000002
+#endif
+
+typedef LONG (NTAPI *PBCryptGenRandom)(HANDLE, PUCHAR, ULONG, ULONG);
+static  PBCryptGenRandom pBCryptGenRandom = NULL;
+
+bool _mi_prim_random_buf(void* buf, size_t buf_len) {
+  if (pBCryptGenRandom == NULL) {
+    HINSTANCE hDll = LoadLibrary(TEXT("bcrypt.dll"));
+    if (hDll != NULL) {
+      pBCryptGenRandom = (PBCryptGenRandom)(void (*)(void))GetProcAddress(hDll, "BCryptGenRandom");
+    }
+    if (pBCryptGenRandom == NULL) return false;
+  }
+  return (pBCryptGenRandom(NULL, (PUCHAR)buf, (ULONG)buf_len, BCRYPT_USE_SYSTEM_PREFERRED_RNG) >= 0);
+}
+
+#endif  // MI_USE_RTLGENRANDOM
+
+//----------------------------------------------------------------
+// Thread init/done
+//----------------------------------------------------------------
+
+#if !defined(MI_SHARED_LIB)
+
+// use thread local storage keys to detect thread ending
+#include <fibersapi.h>
+#if (_WIN32_WINNT < 0x600)  // before Windows Vista
+WINBASEAPI DWORD WINAPI FlsAlloc( _In_opt_ PFLS_CALLBACK_FUNCTION lpCallback );
+WINBASEAPI PVOID WINAPI FlsGetValue( _In_ DWORD dwFlsIndex );
+WINBASEAPI BOOL  WINAPI FlsSetValue( _In_ DWORD dwFlsIndex, _In_opt_ PVOID lpFlsData );
+WINBASEAPI BOOL  WINAPI FlsFree(_In_ DWORD dwFlsIndex);
+#endif
+
+static DWORD mi_fls_key = (DWORD)(-1);
+
+static void NTAPI mi_fls_done(PVOID value) {
+  mi_heap_t* heap = (mi_heap_t*)value;
+  if (heap != NULL) {
+    _mi_thread_done(heap);
+    FlsSetValue(mi_fls_key, NULL);  // prevent recursion as _mi_thread_done may set it back to the main heap, issue #672
+  }
+}
+
+void _mi_prim_thread_init_auto_done(void) {
+  mi_fls_key = FlsAlloc(&mi_fls_done);
+}
+
+void _mi_prim_thread_done_auto_done(void) {
+  // call thread-done on all threads (except the main thread) to prevent
+  // dangling callback pointer if statically linked with a DLL; Issue #208
+  FlsFree(mi_fls_key);
+}
+
+void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {
+  mi_assert_internal(mi_fls_key != (DWORD)(-1));
+  FlsSetValue(mi_fls_key, heap);
+}
+
+#else
+
+// Dll; nothing to do as in that case thread_done is handled through the DLL_THREAD_DETACH event.
+
+void _mi_prim_thread_init_auto_done(void) {
+}
+
+void _mi_prim_thread_done_auto_done(void) {
+}
+
+void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {
+  MI_UNUSED(heap);
+}
+
+#endif
diff --git a/compat/mimalloc/random.c b/compat/mimalloc/random.c
new file mode 100644
index 00000000000000..2a18b5aa992dad
--- /dev/null
+++ b/compat/mimalloc/random.c
@@ -0,0 +1,254 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/prim.h"    // _mi_prim_random_buf
+#include <string.h>       // memset
+
+/* ----------------------------------------------------------------------------
+We use our own PRNG to keep predictable performance of random number generation
+and to avoid implementations that use a lock. We only use the OS provided
+random source to initialize the initial seeds. Since we do not need ultimate
+performance but we do rely on the security (for secret cookies in secure mode)
+we use a cryptographically secure generator (chacha20).
+-----------------------------------------------------------------------------*/
+
+#define MI_CHACHA_ROUNDS (20)   // perhaps use 12 for better performance?
+
+
+/* ----------------------------------------------------------------------------
+Chacha20 implementation as the original algorithm with a 64-bit nonce
+and counter: https://en.wikipedia.org/wiki/Salsa20
+The input matrix has sixteen 32-bit values:
+Position  0 to  3: constant key
+Position  4 to 11: the key
+Position 12 to 13: the counter.
+Position 14 to 15: the nonce.
+
+The implementation uses regular C code which compiles very well on modern compilers.
+(gcc x64 has no register spills, and clang 6+ uses SSE instructions)
+-----------------------------------------------------------------------------*/
+
+static inline uint32_t rotl(uint32_t x, uint32_t shift) {
+  return (x << shift) | (x >> (32 - shift));
+}
+
+static inline void qround(uint32_t x[16], size_t a, size_t b, size_t c, size_t d) {
+  x[a] += x[b]; x[d] = rotl(x[d] ^ x[a], 16);
+  x[c] += x[d]; x[b] = rotl(x[b] ^ x[c], 12);
+  x[a] += x[b]; x[d] = rotl(x[d] ^ x[a], 8);
+  x[c] += x[d]; x[b] = rotl(x[b] ^ x[c], 7);
+}
+
+static void chacha_block(mi_random_ctx_t* ctx)
+{
+  // scramble into `x`
+  uint32_t x[16];
+  for (size_t i = 0; i < 16; i++) {
+    x[i] = ctx->input[i];
+  }
+  for (size_t i = 0; i < MI_CHACHA_ROUNDS; i += 2) {
+    qround(x, 0, 4,  8, 12);
+    qround(x, 1, 5,  9, 13);
+    qround(x, 2, 6, 10, 14);
+    qround(x, 3, 7, 11, 15);
+    qround(x, 0, 5, 10, 15);
+    qround(x, 1, 6, 11, 12);
+    qround(x, 2, 7,  8, 13);
+    qround(x, 3, 4,  9, 14);
+  }
+
+  // add scrambled data to the initial state
+  for (size_t i = 0; i < 16; i++) {
+    ctx->output[i] = x[i] + ctx->input[i];
+  }
+  ctx->output_available = 16;
+
+  // increment the counter for the next round
+  ctx->input[12] += 1;
+  if (ctx->input[12] == 0) {
+    ctx->input[13] += 1;
+    if (ctx->input[13] == 0) {  // and keep increasing into the nonce
+      ctx->input[14] += 1;
+    }
+  }
+}
+
+static uint32_t chacha_next32(mi_random_ctx_t* ctx) {
+  if (ctx->output_available <= 0) {
+    chacha_block(ctx);
+    ctx->output_available = 16; // (assign again to suppress static analysis warning)
+  }
+  const uint32_t x = ctx->output[16 - ctx->output_available];
+  ctx->output[16 - ctx->output_available] = 0; // reset once the data is handed out
+  ctx->output_available--;
+  return x;
+}
+
+static inline uint32_t read32(const uint8_t* p, size_t idx32) {
+  const size_t i = 4*idx32;
+  return ((uint32_t)p[i+0] | (uint32_t)p[i+1] << 8 | (uint32_t)p[i+2] << 16 | (uint32_t)p[i+3] << 24);
+}
+
+static void chacha_init(mi_random_ctx_t* ctx, const uint8_t key[32], uint64_t nonce)
+{
+  // since we only use chacha for randomness (and not encryption) we
+  // do not _need_ to read 32-bit values as little endian but we do anyways
+  // just for being compatible :-)
+  memset(ctx, 0, sizeof(*ctx));
+  for (size_t i = 0; i < 4; i++) {
+    const uint8_t* sigma = (uint8_t*)"expand 32-byte k";
+    ctx->input[i] = read32(sigma,i);
+  }
+  for (size_t i = 0; i < 8; i++) {
+    ctx->input[i + 4] = read32(key,i);
+  }
+  ctx->input[12] = 0;
+  ctx->input[13] = 0;
+  ctx->input[14] = (uint32_t)nonce;
+  ctx->input[15] = (uint32_t)(nonce >> 32);
+}
+
+static void chacha_split(mi_random_ctx_t* ctx, uint64_t nonce, mi_random_ctx_t* ctx_new) {
+  memset(ctx_new, 0, sizeof(*ctx_new));
+  _mi_memcpy(ctx_new->input, ctx->input, sizeof(ctx_new->input));
+  ctx_new->input[12] = 0;
+  ctx_new->input[13] = 0;
+  ctx_new->input[14] = (uint32_t)nonce;
+  ctx_new->input[15] = (uint32_t)(nonce >> 32);
+  mi_assert_internal(ctx->input[14] != ctx_new->input[14] || ctx->input[15] != ctx_new->input[15]); // do not reuse nonces!
+  chacha_block(ctx_new);
+}
+
+
+/* ----------------------------------------------------------------------------
+Random interface
+-----------------------------------------------------------------------------*/
+
+#if MI_DEBUG>1
+static bool mi_random_is_initialized(mi_random_ctx_t* ctx) {
+  return (ctx != NULL && ctx->input[0] != 0);
+}
+#endif
+
+void _mi_random_split(mi_random_ctx_t* ctx, mi_random_ctx_t* ctx_new) {
+  mi_assert_internal(mi_random_is_initialized(ctx));
+  mi_assert_internal(ctx != ctx_new);
+  chacha_split(ctx, (uintptr_t)ctx_new /*nonce*/, ctx_new);
+}
+
+uintptr_t _mi_random_next(mi_random_ctx_t* ctx) {
+  mi_assert_internal(mi_random_is_initialized(ctx));
+  #if MI_INTPTR_SIZE <= 4
+    return chacha_next32(ctx);
+  #elif MI_INTPTR_SIZE == 8
+    return (((uintptr_t)chacha_next32(ctx) << 32) | chacha_next32(ctx));
+  #else
+  # error "define mi_random_next for this platform"
+  #endif
+}
+
+
+/* ----------------------------------------------------------------------------
+To initialize a fresh random context.
+If we cannot get good randomness, we fall back to weak randomness based on a timer and ASLR.
+-----------------------------------------------------------------------------*/
+
+uintptr_t _mi_os_random_weak(uintptr_t extra_seed) {
+  uintptr_t x = (uintptr_t)&_mi_os_random_weak ^ extra_seed; // ASLR makes the address random
+  x ^= _mi_prim_clock_now();
+  // and do a few randomization steps
+  uintptr_t max = ((x ^ (x >> 17)) & 0x0F) + 1;
+  for (uintptr_t i = 0; i < max; i++) {
+    x = _mi_random_shuffle(x);
+  }
+  mi_assert_internal(x != 0);
+  return x;
+}
+
+static void mi_random_init_ex(mi_random_ctx_t* ctx, bool use_weak) {
+  uint8_t key[32];
+  if (use_weak || !_mi_prim_random_buf(key, sizeof(key))) {
+    // if we fail to get random data from the OS, we fall back to a
+    // weak random source based on the current time
+    #if !defined(__wasi__)
+    if (!use_weak) { _mi_warning_message("unable to use secure randomness\n"); }
+    #endif
+    uintptr_t x = _mi_os_random_weak(0);
+    for (size_t i = 0; i < 8; i++) {  // key is eight 32-bit words.
+      x = _mi_random_shuffle(x);
+      ((uint32_t*)key)[i] = (uint32_t)x;
+    }
+    ctx->weak = true;
+  }
+  else {
+    ctx->weak = false;
+  }
+  chacha_init(ctx, key, (uintptr_t)ctx /*nonce*/ );
+}
+
+void _mi_random_init(mi_random_ctx_t* ctx) {
+  mi_random_init_ex(ctx, false);
+}
+
+void _mi_random_init_weak(mi_random_ctx_t * ctx) {
+  mi_random_init_ex(ctx, true);
+}
+
+void _mi_random_reinit_if_weak(mi_random_ctx_t * ctx) {
+  if (ctx->weak) {
+    _mi_random_init(ctx);
+  }
+}
+
+/* --------------------------------------------------------
+test vectors from <https://tools.ietf.org/html/rfc8439>
+----------------------------------------------------------- */
+/*
+static bool array_equals(uint32_t* x, uint32_t* y, size_t n) {
+  for (size_t i = 0; i < n; i++) {
+    if (x[i] != y[i]) return false;
+  }
+  return true;
+}
+static void chacha_test(void)
+{
+  uint32_t x[4] = { 0x11111111, 0x01020304, 0x9b8d6f43, 0x01234567 };
+  uint32_t x_out[4] = { 0xea2a92f4, 0xcb1cf8ce, 0x4581472e, 0x5881c4bb };
+  qround(x, 0, 1, 2, 3);
+  mi_assert_internal(array_equals(x, x_out, 4));
+
+  uint32_t y[16] = {
+       0x879531e0,  0xc5ecf37d,  0x516461b1,  0xc9a62f8a,
+       0x44c20ef3,  0x3390af7f,  0xd9fc690b,  0x2a5f714c,
+       0x53372767,  0xb00a5631,  0x974c541a,  0x359e9963,
+       0x5c971061,  0x3d631689,  0x2098d9d6,  0x91dbd320 };
+  uint32_t y_out[16] = {
+       0x879531e0,  0xc5ecf37d,  0xbdb886dc,  0xc9a62f8a,
+       0x44c20ef3,  0x3390af7f,  0xd9fc690b,  0xcfacafd2,
+       0xe46bea80,  0xb00a5631,  0x974c541a,  0x359e9963,
+       0x5c971061,  0xccc07c79,  0x2098d9d6,  0x91dbd320 };
+  qround(y, 2, 7, 8, 13);
+  mi_assert_internal(array_equals(y, y_out, 16));
+
+  mi_random_ctx_t r = {
+    { 0x61707865, 0x3320646e, 0x79622d32, 0x6b206574,
+      0x03020100, 0x07060504, 0x0b0a0908, 0x0f0e0d0c,
+      0x13121110, 0x17161514, 0x1b1a1918, 0x1f1e1d1c,
+      0x00000001, 0x09000000, 0x4a000000, 0x00000000 },
+    {0},
+    0
+  };
+  uint32_t r_out[16] = {
+       0xe4e7f110, 0x15593bd1, 0x1fdd0f50, 0xc47120a3,
+       0xc7f4d1c7, 0x0368c033, 0x9aaa2204, 0x4e6cd4c3,
+       0x466482d2, 0x09aa9f07, 0x05d7c214, 0xa2028bd9,
+       0xd19c12b5, 0xb94e16de, 0xe883d0cb, 0x4e3c50a2 };
+  chacha_block(&r);
+  mi_assert_internal(array_equals(r.output, r_out, 16));
+}
+*/
diff --git a/compat/mimalloc/segment-cache.c b/compat/mimalloc/segment-cache.c
new file mode 100644
index 00000000000000..e69de29bb2d1d6
diff --git a/compat/mimalloc/segment-map.c b/compat/mimalloc/segment-map.c
new file mode 100644
index 00000000000000..3cd2127e56c1a7
--- /dev/null
+++ b/compat/mimalloc/segment-map.c
@@ -0,0 +1,153 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* -----------------------------------------------------------
+  The following functions are to reliably find the segment or
+  block that encompasses any pointer p (or NULL if it is not
+  in any of our segments).
+  We maintain a bitmap of all memory with 1 bit per MI_SEGMENT_SIZE (64MiB)
+  set to 1 if it contains the segment meta data.
+----------------------------------------------------------- */
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+
+#if (MI_INTPTR_SIZE==8)
+#define MI_MAX_ADDRESS    ((size_t)40 << 40)  // 40TB (to include huge page areas)
+#else
+#define MI_MAX_ADDRESS    ((size_t)2 << 30)   // 2Gb
+#endif
+
+#define MI_SEGMENT_MAP_BITS  (MI_MAX_ADDRESS / MI_SEGMENT_SIZE)
+#define MI_SEGMENT_MAP_SIZE  (MI_SEGMENT_MAP_BITS / 8)
+#define MI_SEGMENT_MAP_WSIZE (MI_SEGMENT_MAP_SIZE / MI_INTPTR_SIZE)
+
+static _Atomic(uintptr_t) mi_segment_map[MI_SEGMENT_MAP_WSIZE + 1];  // 2KiB per TB with 64MiB segments
+
+static size_t mi_segment_map_index_of(const mi_segment_t* segment, size_t* bitidx) {
+  mi_assert_internal(_mi_ptr_segment(segment + 1) == segment); // is it aligned on MI_SEGMENT_SIZE?
+  if ((uintptr_t)segment >= MI_MAX_ADDRESS) {
+    *bitidx = 0;
+    return MI_SEGMENT_MAP_WSIZE;
+  }
+  else {
+    const uintptr_t segindex = ((uintptr_t)segment) / MI_SEGMENT_SIZE;
+    *bitidx = segindex % MI_INTPTR_BITS;
+    const size_t mapindex = segindex / MI_INTPTR_BITS;
+    mi_assert_internal(mapindex < MI_SEGMENT_MAP_WSIZE);
+    return mapindex;
+  }
+}
+
+void _mi_segment_map_allocated_at(const mi_segment_t* segment) {
+  size_t bitidx;
+  size_t index = mi_segment_map_index_of(segment, &bitidx);
+  mi_assert_internal(index <= MI_SEGMENT_MAP_WSIZE);
+  if (index==MI_SEGMENT_MAP_WSIZE) return;
+  uintptr_t mask = mi_atomic_load_relaxed(&mi_segment_map[index]);
+  uintptr_t newmask;
+  do {
+    newmask = (mask | ((uintptr_t)1 << bitidx));
+  } while (!mi_atomic_cas_weak_release(&mi_segment_map[index], &mask, newmask));
+}
+
+void _mi_segment_map_freed_at(const mi_segment_t* segment) {
+  size_t bitidx;
+  size_t index = mi_segment_map_index_of(segment, &bitidx);
+  mi_assert_internal(index <= MI_SEGMENT_MAP_WSIZE);
+  if (index == MI_SEGMENT_MAP_WSIZE) return;
+  uintptr_t mask = mi_atomic_load_relaxed(&mi_segment_map[index]);
+  uintptr_t newmask;
+  do {
+    newmask = (mask & ~((uintptr_t)1 << bitidx));
+  } while (!mi_atomic_cas_weak_release(&mi_segment_map[index], &mask, newmask));
+}
+
+// Determine the segment belonging to a pointer or NULL if it is not in a valid segment.
+static mi_segment_t* _mi_segment_of(const void* p) {
+  if (p == NULL) return NULL;
+  mi_segment_t* segment = _mi_ptr_segment(p);
+  mi_assert_internal(segment != NULL);
+  size_t bitidx;
+  size_t index = mi_segment_map_index_of(segment, &bitidx);
+  // fast path: for any pointer to valid small/medium/large object or first MI_SEGMENT_SIZE in huge
+  const uintptr_t mask = mi_atomic_load_relaxed(&mi_segment_map[index]);
+  if mi_likely((mask & ((uintptr_t)1 << bitidx)) != 0) {
+    return segment; // yes, allocated by us
+  }
+  if (index==MI_SEGMENT_MAP_WSIZE) return NULL;
+
+  // TODO: maintain max/min allocated range for efficiency for more efficient rejection of invalid pointers?
+
+  // search downwards for the first segment in case it is an interior pointer
+  // could be slow but searches in MI_INTPTR_SIZE * MI_SEGMENT_SIZE (512MiB) steps trough
+  // valid huge objects
+  // note: we could maintain a lowest index to speed up the path for invalid pointers?
+  size_t lobitidx;
+  size_t loindex;
+  uintptr_t lobits = mask & (((uintptr_t)1 << bitidx) - 1);
+  if (lobits != 0) {
+    loindex = index;
+    lobitidx = mi_bsr(lobits);    // lobits != 0
+  }
+  else if (index == 0) {
+    return NULL;
+  }
+  else {
+    mi_assert_internal(index > 0);
+    uintptr_t lomask = mask;
+    loindex = index;
+    do {
+      loindex--;
+      lomask = mi_atomic_load_relaxed(&mi_segment_map[loindex]);
+    } while (lomask != 0 && loindex > 0);
+    if (lomask == 0) return NULL;
+    lobitidx = mi_bsr(lomask);    // lomask != 0
+  }
+  mi_assert_internal(loindex < MI_SEGMENT_MAP_WSIZE);
+  // take difference as the addresses could be larger than the MAX_ADDRESS space.
+  size_t diff = (((index - loindex) * (8*MI_INTPTR_SIZE)) + bitidx - lobitidx) * MI_SEGMENT_SIZE;
+  segment = (mi_segment_t*)((uint8_t*)segment - diff);
+
+  if (segment == NULL) return NULL;
+  mi_assert_internal((void*)segment < p);
+  bool cookie_ok = (_mi_ptr_cookie(segment) == segment->cookie);
+  mi_assert_internal(cookie_ok);
+  if mi_unlikely(!cookie_ok) return NULL;
+  if (((uint8_t*)segment + mi_segment_size(segment)) <= (uint8_t*)p) return NULL; // outside the range
+  mi_assert_internal(p >= (void*)segment && (uint8_t*)p < (uint8_t*)segment + mi_segment_size(segment));
+  return segment;
+}
+
+// Is this a valid pointer in our heap?
+static bool  mi_is_valid_pointer(const void* p) {
+  return ((_mi_segment_of(p) != NULL) || (_mi_arena_contains(p)));
+}
+
+mi_decl_nodiscard mi_decl_export bool mi_is_in_heap_region(const void* p) mi_attr_noexcept {
+  return mi_is_valid_pointer(p);
+}
+
+/*
+// Return the full segment range belonging to a pointer
+static void* mi_segment_range_of(const void* p, size_t* size) {
+  mi_segment_t* segment = _mi_segment_of(p);
+  if (segment == NULL) {
+    if (size != NULL) *size = 0;
+    return NULL;
+  }
+  else {
+    if (size != NULL) *size = segment->segment_size;
+    return segment;
+  }
+  mi_assert_expensive(page == NULL || mi_segment_is_valid(_mi_page_segment(page),tld));
+  mi_assert_internal(page == NULL || (mi_segment_page_size(_mi_page_segment(page)) - (MI_SECURE == 0 ? 0 : _mi_os_page_size())) >= block_size);
+  mi_reset_delayed(tld);
+  mi_assert_internal(page == NULL || mi_page_not_in_queue(page, tld));
+  return page;
+}
+*/
diff --git a/compat/mimalloc/segment.c b/compat/mimalloc/segment.c
new file mode 100644
index 00000000000000..6b901f6cc80f13
--- /dev/null
+++ b/compat/mimalloc/segment.c
@@ -0,0 +1,1617 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2020, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+
+#include <string.h>  // memset
+#include <stdio.h>
+
+#define MI_PAGE_HUGE_ALIGN   (256*1024)
+
+static void mi_segment_try_purge(mi_segment_t* segment, bool force, mi_stats_t* stats);
+
+
+// -------------------------------------------------------------------
+// commit mask
+// -------------------------------------------------------------------
+
+static bool mi_commit_mask_all_set(const mi_commit_mask_t* commit, const mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    if ((commit->mask[i] & cm->mask[i]) != cm->mask[i]) return false;
+  }
+  return true;
+}
+
+static bool mi_commit_mask_any_set(const mi_commit_mask_t* commit, const mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    if ((commit->mask[i] & cm->mask[i]) != 0) return true;
+  }
+  return false;
+}
+
+static void mi_commit_mask_create_intersect(const mi_commit_mask_t* commit, const mi_commit_mask_t* cm, mi_commit_mask_t* res) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    res->mask[i] = (commit->mask[i] & cm->mask[i]);
+  }
+}
+
+static void mi_commit_mask_clear(mi_commit_mask_t* res, const mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    res->mask[i] &= ~(cm->mask[i]);
+  }
+}
+
+static void mi_commit_mask_set(mi_commit_mask_t* res, const mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    res->mask[i] |= cm->mask[i];
+  }
+}
+
+static void mi_commit_mask_create(size_t bitidx, size_t bitcount, mi_commit_mask_t* cm) {
+  mi_assert_internal(bitidx < MI_COMMIT_MASK_BITS);
+  mi_assert_internal((bitidx + bitcount) <= MI_COMMIT_MASK_BITS);
+  if (bitcount == MI_COMMIT_MASK_BITS) {
+    mi_assert_internal(bitidx==0);
+    mi_commit_mask_create_full(cm);
+  }
+  else if (bitcount == 0) {
+    mi_commit_mask_create_empty(cm);
+  }
+  else {
+    mi_commit_mask_create_empty(cm);
+    size_t i = bitidx / MI_COMMIT_MASK_FIELD_BITS;
+    size_t ofs = bitidx % MI_COMMIT_MASK_FIELD_BITS;
+    while (bitcount > 0) {
+      mi_assert_internal(i < MI_COMMIT_MASK_FIELD_COUNT);
+      size_t avail = MI_COMMIT_MASK_FIELD_BITS - ofs;
+      size_t count = (bitcount > avail ? avail : bitcount);
+      size_t mask = (count >= MI_COMMIT_MASK_FIELD_BITS ? ~((size_t)0) : (((size_t)1 << count) - 1) << ofs);
+      cm->mask[i] = mask;
+      bitcount -= count;
+      ofs = 0;
+      i++;
+    }
+  }
+}
+
+size_t _mi_commit_mask_committed_size(const mi_commit_mask_t* cm, size_t total) {
+  mi_assert_internal((total%MI_COMMIT_MASK_BITS)==0);
+  size_t count = 0;
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    size_t mask = cm->mask[i];
+    if (~mask == 0) {
+      count += MI_COMMIT_MASK_FIELD_BITS;
+    }
+    else {
+      for (; mask != 0; mask >>= 1) {  // todo: use popcount
+	if ((mask&1)!=0) count++;
+      }
+    }
+  }
+  // we use total since for huge segments each commit bit may represent a larger size
+  return ((total / MI_COMMIT_MASK_BITS) * count);
+}
+
+
+size_t _mi_commit_mask_next_run(const mi_commit_mask_t* cm, size_t* idx) {
+  size_t i = (*idx) / MI_COMMIT_MASK_FIELD_BITS;
+  size_t ofs = (*idx) % MI_COMMIT_MASK_FIELD_BITS;
+  size_t mask = 0;
+  // find first ones
+  while (i < MI_COMMIT_MASK_FIELD_COUNT) {
+    mask = cm->mask[i];
+    mask >>= ofs;
+    if (mask != 0) {
+      while ((mask&1) == 0) {
+	mask >>= 1;
+	ofs++;
+      }
+      break;
+    }
+    i++;
+    ofs = 0;
+  }
+  if (i >= MI_COMMIT_MASK_FIELD_COUNT) {
+    // not found
+    *idx = MI_COMMIT_MASK_BITS;
+    return 0;
+  }
+  else {
+    // found, count ones
+    size_t count = 0;
+    *idx = (i*MI_COMMIT_MASK_FIELD_BITS) + ofs;
+    do {
+      mi_assert_internal(ofs < MI_COMMIT_MASK_FIELD_BITS && (mask&1) == 1);
+      do {
+	count++;
+	mask >>= 1;
+      } while ((mask&1) == 1);
+      if ((((*idx + count) % MI_COMMIT_MASK_FIELD_BITS) == 0)) {
+	i++;
+	if (i >= MI_COMMIT_MASK_FIELD_COUNT) break;
+	mask = cm->mask[i];
+	ofs = 0;
+      }
+    } while ((mask&1) == 1);
+    mi_assert_internal(count > 0);
+    return count;
+  }
+}
+
+
+/* --------------------------------------------------------------------------------
+  Segment allocation
+
+  If a  thread ends, it "abandons" pages with used blocks
+  and there is an abandoned segment list whose segments can
+  be reclaimed by still running threads, much like work-stealing.
+-------------------------------------------------------------------------------- */
+
+
+/* -----------------------------------------------------------
+   Slices
+----------------------------------------------------------- */
+
+
+static const mi_slice_t* mi_segment_slices_end(const mi_segment_t* segment) {
+  return &segment->slices[segment->slice_entries];
+}
+
+static uint8_t* mi_slice_start(const mi_slice_t* slice) {
+  mi_segment_t* segment = _mi_ptr_segment(slice);
+  mi_assert_internal(slice >= segment->slices && slice < mi_segment_slices_end(segment));
+  return ((uint8_t*)segment + ((slice - segment->slices)*MI_SEGMENT_SLICE_SIZE));
+}
+
+
+/* -----------------------------------------------------------
+   Bins
+----------------------------------------------------------- */
+// Use bit scan forward to quickly find the first zero bit if it is available
+
+static inline size_t mi_slice_bin8(size_t slice_count) {
+  if (slice_count<=1) return slice_count;
+  mi_assert_internal(slice_count <= MI_SLICES_PER_SEGMENT);
+  slice_count--;
+  size_t s = mi_bsr(slice_count);  // slice_count > 1
+  if (s <= 2) return slice_count + 1;
+  size_t bin = ((s << 2) | ((slice_count >> (s - 2))&0x03)) - 4;
+  return bin;
+}
+
+static inline size_t mi_slice_bin(size_t slice_count) {
+  mi_assert_internal(slice_count*MI_SEGMENT_SLICE_SIZE <= MI_SEGMENT_SIZE);
+  mi_assert_internal(mi_slice_bin8(MI_SLICES_PER_SEGMENT) <= MI_SEGMENT_BIN_MAX);
+  size_t bin = mi_slice_bin8(slice_count);
+  mi_assert_internal(bin <= MI_SEGMENT_BIN_MAX);
+  return bin;
+}
+
+static inline size_t mi_slice_index(const mi_slice_t* slice) {
+  mi_segment_t* segment = _mi_ptr_segment(slice);
+  ptrdiff_t index = slice - segment->slices;
+  mi_assert_internal(index >= 0 && index < (ptrdiff_t)segment->slice_entries);
+  return index;
+}
+
+
+/* -----------------------------------------------------------
+   Slice span queues
+----------------------------------------------------------- */
+
+static void mi_span_queue_push(mi_span_queue_t* sq, mi_slice_t* slice) {
+  // todo: or push to the end?
+  mi_assert_internal(slice->prev == NULL && slice->next==NULL);
+  slice->prev = NULL; // paranoia
+  slice->next = sq->first;
+  sq->first = slice;
+  if (slice->next != NULL) slice->next->prev = slice;
+		     else sq->last = slice;
+  slice->xblock_size = 0; // free
+}
+
+static mi_span_queue_t* mi_span_queue_for(size_t slice_count, mi_segments_tld_t* tld) {
+  size_t bin = mi_slice_bin(slice_count);
+  mi_span_queue_t* sq = &tld->spans[bin];
+  mi_assert_internal(sq->slice_count >= slice_count);
+  return sq;
+}
+
+static void mi_span_queue_delete(mi_span_queue_t* sq, mi_slice_t* slice) {
+  mi_assert_internal(slice->xblock_size==0 && slice->slice_count>0 && slice->slice_offset==0);
+  // should work too if the queue does not contain slice (which can happen during reclaim)
+  if (slice->prev != NULL) slice->prev->next = slice->next;
+  if (slice == sq->first) sq->first = slice->next;
+  if (slice->next != NULL) slice->next->prev = slice->prev;
+  if (slice == sq->last) sq->last = slice->prev;
+  slice->prev = NULL;
+  slice->next = NULL;
+  slice->xblock_size = 1; // no more free
+}
+
+
+/* -----------------------------------------------------------
+ Invariant checking
+----------------------------------------------------------- */
+
+static bool mi_slice_is_used(const mi_slice_t* slice) {
+  return (slice->xblock_size > 0);
+}
+
+
+#if (MI_DEBUG>=3)
+static bool mi_span_queue_contains(mi_span_queue_t* sq, mi_slice_t* slice) {
+  for (mi_slice_t* s = sq->first; s != NULL; s = s->next) {
+    if (s==slice) return true;
+  }
+  return false;
+}
+
+static bool mi_segment_is_valid(mi_segment_t* segment, mi_segments_tld_t* tld) {
+  mi_assert_internal(segment != NULL);
+  mi_assert_internal(_mi_ptr_cookie(segment) == segment->cookie);
+  mi_assert_internal(segment->abandoned <= segment->used);
+  mi_assert_internal(segment->thread_id == 0 || segment->thread_id == _mi_thread_id());
+  mi_assert_internal(mi_commit_mask_all_set(&segment->commit_mask, &segment->purge_mask)); // can only decommit committed blocks
+  //mi_assert_internal(segment->segment_info_size % MI_SEGMENT_SLICE_SIZE == 0);
+  mi_slice_t* slice = &segment->slices[0];
+  const mi_slice_t* end = mi_segment_slices_end(segment);
+  size_t used_count = 0;
+  mi_span_queue_t* sq;
+  while(slice < end) {
+    mi_assert_internal(slice->slice_count > 0);
+    mi_assert_internal(slice->slice_offset == 0);
+    size_t index = mi_slice_index(slice);
+    size_t maxindex = (index + slice->slice_count >= segment->slice_entries ? segment->slice_entries : index + slice->slice_count) - 1;
+    if (mi_slice_is_used(slice)) { // a page in use, we need at least MAX_SLICE_OFFSET valid back offsets
+      used_count++;
+      for (size_t i = 0; i <= MI_MAX_SLICE_OFFSET && index + i <= maxindex; i++) {
+	mi_assert_internal(segment->slices[index + i].slice_offset == i*sizeof(mi_slice_t));
+	mi_assert_internal(i==0 || segment->slices[index + i].slice_count == 0);
+	mi_assert_internal(i==0 || segment->slices[index + i].xblock_size == 1);
+      }
+      // and the last entry as well (for coalescing)
+      const mi_slice_t* last = slice + slice->slice_count - 1;
+      if (last > slice && last < mi_segment_slices_end(segment)) {
+	mi_assert_internal(last->slice_offset == (slice->slice_count-1)*sizeof(mi_slice_t));
+	mi_assert_internal(last->slice_count == 0);
+	mi_assert_internal(last->xblock_size == 1);
+      }
+    }
+    else {  // free range of slices; only last slice needs a valid back offset
+      mi_slice_t* last = &segment->slices[maxindex];
+      if (segment->kind != MI_SEGMENT_HUGE || slice->slice_count <= (segment->slice_entries - segment->segment_info_slices)) {
+	mi_assert_internal((uint8_t*)slice == (uint8_t*)last - last->slice_offset);
+      }
+      mi_assert_internal(slice == last || last->slice_count == 0 );
+      mi_assert_internal(last->xblock_size == 0 || (segment->kind==MI_SEGMENT_HUGE && last->xblock_size==1));
+      if (segment->kind != MI_SEGMENT_HUGE && segment->thread_id != 0) { // segment is not huge or abandoned
+	sq = mi_span_queue_for(slice->slice_count,tld);
+	mi_assert_internal(mi_span_queue_contains(sq,slice));
+      }
+    }
+    slice = &segment->slices[maxindex+1];
+  }
+  mi_assert_internal(slice == end);
+  mi_assert_internal(used_count == segment->used + 1);
+  return true;
+}
+#endif
+
+/* -----------------------------------------------------------
+ Segment size calculations
+----------------------------------------------------------- */
+
+static size_t mi_segment_info_size(mi_segment_t* segment) {
+  return segment->segment_info_slices * MI_SEGMENT_SLICE_SIZE;
+}
+
+static uint8_t* _mi_segment_page_start_from_slice(const mi_segment_t* segment, const mi_slice_t* slice, size_t xblock_size, size_t* page_size)
+{
+  ptrdiff_t idx = slice - segment->slices;
+  size_t psize = (size_t)slice->slice_count * MI_SEGMENT_SLICE_SIZE;
+  // make the start not OS page aligned for smaller blocks to avoid page/cache effects
+  // note: the offset must always be an xblock_size multiple since we assume small allocations
+  // are aligned (see `mi_heap_malloc_aligned`).
+  size_t start_offset = 0;
+  if (xblock_size >= MI_INTPTR_SIZE) {
+    if (xblock_size <= 64) { start_offset = 3*xblock_size; }
+    else if (xblock_size <= 512) { start_offset = xblock_size; }
+  }
+  if (page_size != NULL) { *page_size = psize - start_offset; }
+  return (uint8_t*)segment + ((idx*MI_SEGMENT_SLICE_SIZE) + start_offset);
+}
+
+// Start of the page available memory; can be used on uninitialized pages
+uint8_t* _mi_segment_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size)
+{
+  const mi_slice_t* slice = mi_page_to_slice((mi_page_t*)page);
+  uint8_t* p = _mi_segment_page_start_from_slice(segment, slice, page->xblock_size, page_size);
+  mi_assert_internal(page->xblock_size > 0 || _mi_ptr_page(p) == page);
+  mi_assert_internal(_mi_ptr_segment(p) == segment);
+  return p;
+}
+
+
+static size_t mi_segment_calculate_slices(size_t required, size_t* pre_size, size_t* info_slices) {
+  size_t page_size = _mi_os_page_size();
+  size_t isize     = _mi_align_up(sizeof(mi_segment_t), page_size);
+  size_t guardsize = 0;
+
+  if (MI_SECURE>0) {
+    // in secure mode, we set up a protected page in between the segment info
+    // and the page data (and one at the end of the segment)
+    guardsize = page_size;
+    if (required > 0) {
+      required = _mi_align_up(required, MI_SEGMENT_SLICE_SIZE) + page_size;
+    }
+  }
+
+  if (pre_size != NULL) *pre_size = isize;
+  isize = _mi_align_up(isize + guardsize, MI_SEGMENT_SLICE_SIZE);
+  if (info_slices != NULL) *info_slices = isize / MI_SEGMENT_SLICE_SIZE;
+  size_t segment_size = (required==0 ? MI_SEGMENT_SIZE : _mi_align_up( required + isize + guardsize, MI_SEGMENT_SLICE_SIZE) );
+  mi_assert_internal(segment_size % MI_SEGMENT_SLICE_SIZE == 0);
+  return (segment_size / MI_SEGMENT_SLICE_SIZE);
+}
+
+
+/* ----------------------------------------------------------------------------
+Segment caches
+We keep a small segment cache per thread to increase local
+reuse and avoid setting/clearing guard pages in secure mode.
+------------------------------------------------------------------------------- */
+
+static void mi_segments_track_size(long segment_size, mi_segments_tld_t* tld) {
+  if (segment_size>=0) _mi_stat_increase(&tld->stats->segments,1);
+		  else _mi_stat_decrease(&tld->stats->segments,1);
+  tld->count += (segment_size >= 0 ? 1 : -1);
+  if (tld->count > tld->peak_count) tld->peak_count = tld->count;
+  tld->current_size += segment_size;
+  if (tld->current_size > tld->peak_size) tld->peak_size = tld->current_size;
+}
+
+static void mi_segment_os_free(mi_segment_t* segment, mi_segments_tld_t* tld) {
+  segment->thread_id = 0;
+  _mi_segment_map_freed_at(segment);
+  mi_segments_track_size(-((long)mi_segment_size(segment)),tld);
+  if (MI_SECURE>0) {
+    // _mi_os_unprotect(segment, mi_segment_size(segment)); // ensure no more guard pages are set
+    // unprotect the guard pages; we cannot just unprotect the whole segment size as part may be decommitted
+    size_t os_pagesize = _mi_os_page_size();
+    _mi_os_unprotect((uint8_t*)segment + mi_segment_info_size(segment) - os_pagesize, os_pagesize);
+    uint8_t* end = (uint8_t*)segment + mi_segment_size(segment) - os_pagesize;
+    _mi_os_unprotect(end, os_pagesize);
+  }
+
+  // purge delayed decommits now? (no, leave it to the arena)
+  // mi_segment_try_purge(segment,true,tld->stats);
+
+  const size_t size = mi_segment_size(segment);
+  const size_t csize = _mi_commit_mask_committed_size(&segment->commit_mask, size);
+
+  _mi_abandoned_await_readers();  // wait until safe to free
+  _mi_arena_free(segment, mi_segment_size(segment), csize, segment->memid, tld->stats);
+}
+
+// called by threads that are terminating
+void _mi_segment_thread_collect(mi_segments_tld_t* tld) {
+  MI_UNUSED(tld);
+  // nothing to do
+}
+
+
+/* -----------------------------------------------------------
+   Commit/Decommit ranges
+----------------------------------------------------------- */
+
+static void mi_segment_commit_mask(mi_segment_t* segment, bool conservative, uint8_t* p, size_t size, uint8_t** start_p, size_t* full_size, mi_commit_mask_t* cm) {
+  mi_assert_internal(_mi_ptr_segment(p + 1) == segment);
+  mi_assert_internal(segment->kind != MI_SEGMENT_HUGE);
+  mi_commit_mask_create_empty(cm);
+  if (size == 0 || size > MI_SEGMENT_SIZE || segment->kind == MI_SEGMENT_HUGE) return;
+  const size_t segstart = mi_segment_info_size(segment);
+  const size_t segsize = mi_segment_size(segment);
+  if (p >= (uint8_t*)segment + segsize) return;
+
+  size_t pstart = (p - (uint8_t*)segment);
+  mi_assert_internal(pstart + size <= segsize);
+
+  size_t start;
+  size_t end;
+  if (conservative) {
+    // decommit conservative
+    start = _mi_align_up(pstart, MI_COMMIT_SIZE);
+    end   = _mi_align_down(pstart + size, MI_COMMIT_SIZE);
+    mi_assert_internal(start >= segstart);
+    mi_assert_internal(end <= segsize);
+  }
+  else {
+    // commit liberal
+    start = _mi_align_down(pstart, MI_MINIMAL_COMMIT_SIZE);
+    end   = _mi_align_up(pstart + size, MI_MINIMAL_COMMIT_SIZE);
+  }
+  if (pstart >= segstart && start < segstart) {  // note: the mask is also calculated for an initial commit of the info area
+    start = segstart;
+  }
+  if (end > segsize) {
+    end = segsize;
+  }
+
+  mi_assert_internal(start <= pstart && (pstart + size) <= end);
+  mi_assert_internal(start % MI_COMMIT_SIZE==0 && end % MI_COMMIT_SIZE == 0);
+  *start_p   = (uint8_t*)segment + start;
+  *full_size = (end > start ? end - start : 0);
+  if (*full_size == 0) return;
+
+  size_t bitidx = start / MI_COMMIT_SIZE;
+  mi_assert_internal(bitidx < MI_COMMIT_MASK_BITS);
+
+  size_t bitcount = *full_size / MI_COMMIT_SIZE; // can be 0
+  if (bitidx + bitcount > MI_COMMIT_MASK_BITS) {
+    _mi_warning_message("commit mask overflow: idx=%zu count=%zu start=%zx end=%zx p=0x%p size=%zu fullsize=%zu\n", bitidx, bitcount, start, end, p, size, *full_size);
+  }
+  mi_assert_internal((bitidx + bitcount) <= MI_COMMIT_MASK_BITS);
+  mi_commit_mask_create(bitidx, bitcount, cm);
+}
+
+static bool mi_segment_commit(mi_segment_t* segment, uint8_t* p, size_t size, mi_stats_t* stats) {
+  mi_assert_internal(mi_commit_mask_all_set(&segment->commit_mask, &segment->purge_mask));
+
+  // commit liberal
+  uint8_t* start = NULL;
+  size_t   full_size = 0;
+  mi_commit_mask_t mask;
+  mi_segment_commit_mask(segment, false /* conservative? */, p, size, &start, &full_size, &mask);
+  if (mi_commit_mask_is_empty(&mask) || full_size == 0) return true;
+
+  if (!mi_commit_mask_all_set(&segment->commit_mask, &mask)) {
+    // committing
+    bool is_zero = false;
+    mi_commit_mask_t cmask;
+    mi_commit_mask_create_intersect(&segment->commit_mask, &mask, &cmask);
+    _mi_stat_decrease(&_mi_stats_main.committed, _mi_commit_mask_committed_size(&cmask, MI_SEGMENT_SIZE)); // adjust for overlap
+    if (!_mi_os_commit(start, full_size, &is_zero, stats)) return false;
+    mi_commit_mask_set(&segment->commit_mask, &mask);
+  }
+
+  // increase purge expiration when using part of delayed purges -- we assume more allocations are coming soon.
+  if (mi_commit_mask_any_set(&segment->purge_mask, &mask)) {
+    segment->purge_expire = _mi_clock_now() + mi_option_get(mi_option_purge_delay);
+  }
+
+  // always clear any delayed purges in our range (as they are either committed now)
+  mi_commit_mask_clear(&segment->purge_mask, &mask);
+  return true;
+}
+
+static bool mi_segment_ensure_committed(mi_segment_t* segment, uint8_t* p, size_t size, mi_stats_t* stats) {
+  mi_assert_internal(mi_commit_mask_all_set(&segment->commit_mask, &segment->purge_mask));
+  // note: assumes commit_mask is always full for huge segments as otherwise the commit mask bits can overflow
+  if (mi_commit_mask_is_full(&segment->commit_mask) && mi_commit_mask_is_empty(&segment->purge_mask)) return true; // fully committed
+  mi_assert_internal(segment->kind != MI_SEGMENT_HUGE);
+  return mi_segment_commit(segment, p, size, stats);
+}
+
+static bool mi_segment_purge(mi_segment_t* segment, uint8_t* p, size_t size, mi_stats_t* stats) {
+  mi_assert_internal(mi_commit_mask_all_set(&segment->commit_mask, &segment->purge_mask));
+  if (!segment->allow_purge) return true;
+
+  // purge conservative
+  uint8_t* start = NULL;
+  size_t   full_size = 0;
+  mi_commit_mask_t mask;
+  mi_segment_commit_mask(segment, true /* conservative? */, p, size, &start, &full_size, &mask);
+  if (mi_commit_mask_is_empty(&mask) || full_size==0) return true;
+
+  if (mi_commit_mask_any_set(&segment->commit_mask, &mask)) {
+    // purging
+    mi_assert_internal((void*)start != (void*)segment);
+    mi_assert_internal(segment->allow_decommit);
+    const bool decommitted = _mi_os_purge(start, full_size, stats);  // reset or decommit
+    if (decommitted) {
+      mi_commit_mask_t cmask;
+      mi_commit_mask_create_intersect(&segment->commit_mask, &mask, &cmask);
+      _mi_stat_increase(&_mi_stats_main.committed, full_size - _mi_commit_mask_committed_size(&cmask, MI_SEGMENT_SIZE)); // adjust for double counting
+      mi_commit_mask_clear(&segment->commit_mask, &mask);
+    }
+  }
+
+  // always clear any scheduled purges in our range
+  mi_commit_mask_clear(&segment->purge_mask, &mask);
+  return true;
+}
+
+static void mi_segment_schedule_purge(mi_segment_t* segment, uint8_t* p, size_t size, mi_stats_t* stats) {
+  if (!segment->allow_purge) return;
+
+  if (mi_option_get(mi_option_purge_delay) == 0) {
+    mi_segment_purge(segment, p, size, stats);
+  }
+  else {
+    // register for future purge in the purge mask
+    uint8_t* start = NULL;
+    size_t   full_size = 0;
+    mi_commit_mask_t mask;
+    mi_segment_commit_mask(segment, true /*conservative*/, p, size, &start, &full_size, &mask);
+    if (mi_commit_mask_is_empty(&mask) || full_size==0) return;
+
+    // update delayed commit
+    mi_assert_internal(segment->purge_expire > 0 || mi_commit_mask_is_empty(&segment->purge_mask));
+    mi_commit_mask_t cmask;
+    mi_commit_mask_create_intersect(&segment->commit_mask, &mask, &cmask);  // only purge what is committed; span_free may try to decommit more
+    mi_commit_mask_set(&segment->purge_mask, &cmask);
+    mi_msecs_t now = _mi_clock_now();
+    if (segment->purge_expire == 0) {
+      // no previous purgess, initialize now
+      segment->purge_expire = now + mi_option_get(mi_option_purge_delay);
+    }
+    else if (segment->purge_expire <= now) {
+      // previous purge mask already expired
+      if (segment->purge_expire + mi_option_get(mi_option_purge_extend_delay) <= now) {
+	mi_segment_try_purge(segment, true, stats);
+      }
+      else {
+	segment->purge_expire = now + mi_option_get(mi_option_purge_extend_delay); // (mi_option_get(mi_option_purge_delay) / 8); // wait a tiny bit longer in case there is a series of free's
+      }
+    }
+    else {
+      // previous purge mask is not yet expired, increase the expiration by a bit.
+      segment->purge_expire += mi_option_get(mi_option_purge_extend_delay);
+    }
+  }
+}
+
+static void mi_segment_try_purge(mi_segment_t* segment, bool force, mi_stats_t* stats) {
+  if (!segment->allow_purge || mi_commit_mask_is_empty(&segment->purge_mask)) return;
+  mi_msecs_t now = _mi_clock_now();
+  if (!force && now < segment->purge_expire) return;
+
+  mi_commit_mask_t mask = segment->purge_mask;
+  segment->purge_expire = 0;
+  mi_commit_mask_create_empty(&segment->purge_mask);
+
+  size_t idx;
+  size_t count;
+  mi_commit_mask_foreach(&mask, idx, count) {
+    // if found, decommit that sequence
+    if (count > 0) {
+      uint8_t* p = (uint8_t*)segment + (idx*MI_COMMIT_SIZE);
+      size_t size = count * MI_COMMIT_SIZE;
+      mi_segment_purge(segment, p, size, stats);
+    }
+  }
+  mi_commit_mask_foreach_end()
+  mi_assert_internal(mi_commit_mask_is_empty(&segment->purge_mask));
+}
+
+
+/* -----------------------------------------------------------
+   Span free
+----------------------------------------------------------- */
+
+static bool mi_segment_is_abandoned(mi_segment_t* segment) {
+  return (segment->thread_id == 0);
+}
+
+// note: can be called on abandoned segments
+static void mi_segment_span_free(mi_segment_t* segment, size_t slice_index, size_t slice_count, bool allow_purge, mi_segments_tld_t* tld) {
+  mi_assert_internal(slice_index < segment->slice_entries);
+  mi_span_queue_t* sq = (segment->kind == MI_SEGMENT_HUGE || mi_segment_is_abandoned(segment)
+			  ? NULL : mi_span_queue_for(slice_count,tld));
+  if (slice_count==0) slice_count = 1;
+  mi_assert_internal(slice_index + slice_count - 1 < segment->slice_entries);
+
+  // set first and last slice (the intermediates can be undetermined)
+  mi_slice_t* slice = &segment->slices[slice_index];
+  slice->slice_count = (uint32_t)slice_count;
+  mi_assert_internal(slice->slice_count == slice_count); // no overflow?
+  slice->slice_offset = 0;
+  if (slice_count > 1) {
+    mi_slice_t* last = &segment->slices[slice_index + slice_count - 1];
+    last->slice_count = 0;
+    last->slice_offset = (uint32_t)(sizeof(mi_page_t)*(slice_count - 1));
+    last->xblock_size = 0;
+  }
+
+  // perhaps decommit
+  if (allow_purge) {
+    mi_segment_schedule_purge(segment, mi_slice_start(slice), slice_count * MI_SEGMENT_SLICE_SIZE, tld->stats);
+  }
+
+  // and push it on the free page queue (if it was not a huge page)
+  if (sq != NULL) mi_span_queue_push( sq, slice );
+	     else slice->xblock_size = 0; // mark huge page as free anyways
+}
+
+/*
+// called from reclaim to add existing free spans
+static void mi_segment_span_add_free(mi_slice_t* slice, mi_segments_tld_t* tld) {
+  mi_segment_t* segment = _mi_ptr_segment(slice);
+  mi_assert_internal(slice->xblock_size==0 && slice->slice_count>0 && slice->slice_offset==0);
+  size_t slice_index = mi_slice_index(slice);
+  mi_segment_span_free(segment,slice_index,slice->slice_count,tld);
+}
+*/
+
+static void mi_segment_span_remove_from_queue(mi_slice_t* slice, mi_segments_tld_t* tld) {
+  mi_assert_internal(slice->slice_count > 0 && slice->slice_offset==0 && slice->xblock_size==0);
+  mi_assert_internal(_mi_ptr_segment(slice)->kind != MI_SEGMENT_HUGE);
+  mi_span_queue_t* sq = mi_span_queue_for(slice->slice_count, tld);
+  mi_span_queue_delete(sq, slice);
+}
+
+// note: can be called on abandoned segments
+static mi_slice_t* mi_segment_span_free_coalesce(mi_slice_t* slice, mi_segments_tld_t* tld) {
+  mi_assert_internal(slice != NULL && slice->slice_count > 0 && slice->slice_offset == 0);
+  mi_segment_t* segment = _mi_ptr_segment(slice);
+  bool is_abandoned = mi_segment_is_abandoned(segment);
+
+  // for huge pages, just mark as free but don't add to the queues
+  if (segment->kind == MI_SEGMENT_HUGE) {
+    // issue #691: segment->used can be 0 if the huge page block was freed while abandoned (reclaim will get here in that case)
+    mi_assert_internal((segment->used==0 && slice->xblock_size==0) || segment->used == 1);  // decreased right after this call in `mi_segment_page_clear`
+    slice->xblock_size = 0;  // mark as free anyways
+    // we should mark the last slice `xblock_size=0` now to maintain invariants but we skip it to
+    // avoid a possible cache miss (and the segment is about to be freed)
+    return slice;
+  }
+
+  // otherwise coalesce the span and add to the free span queues
+  size_t slice_count = slice->slice_count;
+  mi_slice_t* next = slice + slice->slice_count;
+  mi_assert_internal(next <= mi_segment_slices_end(segment));
+  if (next < mi_segment_slices_end(segment) && next->xblock_size==0) {
+    // free next block -- remove it from free and merge
+    mi_assert_internal(next->slice_count > 0 && next->slice_offset==0);
+    slice_count += next->slice_count; // extend
+    if (!is_abandoned) { mi_segment_span_remove_from_queue(next, tld); }
+  }
+  if (slice > segment->slices) {
+    mi_slice_t* prev = mi_slice_first(slice - 1);
+    mi_assert_internal(prev >= segment->slices);
+    if (prev->xblock_size==0) {
+      // free previous slice -- remove it from free and merge
+      mi_assert_internal(prev->slice_count > 0 && prev->slice_offset==0);
+      slice_count += prev->slice_count;
+      if (!is_abandoned) { mi_segment_span_remove_from_queue(prev, tld); }
+      slice = prev;
+    }
+  }
+
+  // and add the new free page
+  mi_segment_span_free(segment, mi_slice_index(slice), slice_count, true, tld);
+  return slice;
+}
+
+
+
+/* -----------------------------------------------------------
+   Page allocation
+----------------------------------------------------------- */
+
+// Note: may still return NULL if committing the memory failed
+static mi_page_t* mi_segment_span_allocate(mi_segment_t* segment, size_t slice_index, size_t slice_count, mi_segments_tld_t* tld) {
+  mi_assert_internal(slice_index < segment->slice_entries);
+  mi_slice_t* const slice = &segment->slices[slice_index];
+  mi_assert_internal(slice->xblock_size==0 || slice->xblock_size==1);
+
+  // commit before changing the slice data
+  if (!mi_segment_ensure_committed(segment, _mi_segment_page_start_from_slice(segment, slice, 0, NULL), slice_count * MI_SEGMENT_SLICE_SIZE, tld->stats)) {
+    return NULL;  // commit failed!
+  }
+
+  // convert the slices to a page
+  slice->slice_offset = 0;
+  slice->slice_count = (uint32_t)slice_count;
+  mi_assert_internal(slice->slice_count == slice_count);
+  const size_t bsize = slice_count * MI_SEGMENT_SLICE_SIZE;
+  slice->xblock_size = (uint32_t)(bsize >= MI_HUGE_BLOCK_SIZE ? MI_HUGE_BLOCK_SIZE : bsize);
+  mi_page_t*  page = mi_slice_to_page(slice);
+  mi_assert_internal(mi_page_block_size(page) == bsize);
+
+  // set slice back pointers for the first MI_MAX_SLICE_OFFSET entries
+  size_t extra = slice_count-1;
+  if (extra > MI_MAX_SLICE_OFFSET) extra = MI_MAX_SLICE_OFFSET;
+  if (slice_index + extra >= segment->slice_entries) extra = segment->slice_entries - slice_index - 1;  // huge objects may have more slices than avaiable entries in the segment->slices
+
+  mi_slice_t* slice_next = slice + 1;
+  for (size_t i = 1; i <= extra; i++, slice_next++) {
+    slice_next->slice_offset = (uint32_t)(sizeof(mi_slice_t)*i);
+    slice_next->slice_count = 0;
+    slice_next->xblock_size = 1;
+  }
+
+  // and also for the last one (if not set already) (the last one is needed for coalescing and for large alignments)
+  // note: the cast is needed for ubsan since the index can be larger than MI_SLICES_PER_SEGMENT for huge allocations (see #543)
+  mi_slice_t* last = slice + slice_count - 1;
+  mi_slice_t* end = (mi_slice_t*)mi_segment_slices_end(segment);
+  if (last > end) last = end;
+  if (last > slice) {
+    last->slice_offset = (uint32_t)(sizeof(mi_slice_t) * (last - slice));
+    last->slice_count = 0;
+    last->xblock_size = 1;
+  }
+
+  // and initialize the page
+  page->is_committed = true;
+  segment->used++;
+  return page;
+}
+
+static void mi_segment_slice_split(mi_segment_t* segment, mi_slice_t* slice, size_t slice_count, mi_segments_tld_t* tld) {
+  mi_assert_internal(_mi_ptr_segment(slice) == segment);
+  mi_assert_internal(slice->slice_count >= slice_count);
+  mi_assert_internal(slice->xblock_size > 0); // no more in free queue
+  if (slice->slice_count <= slice_count) return;
+  mi_assert_internal(segment->kind != MI_SEGMENT_HUGE);
+  size_t next_index = mi_slice_index(slice) + slice_count;
+  size_t next_count = slice->slice_count - slice_count;
+  mi_segment_span_free(segment, next_index, next_count, false /* don't purge left-over part */, tld);
+  slice->slice_count = (uint32_t)slice_count;
+}
+
+static mi_page_t* mi_segments_page_find_and_allocate(size_t slice_count, mi_arena_id_t req_arena_id, mi_segments_tld_t* tld) {
+  mi_assert_internal(slice_count*MI_SEGMENT_SLICE_SIZE <= MI_LARGE_OBJ_SIZE_MAX);
+  // search from best fit up
+  mi_span_queue_t* sq = mi_span_queue_for(slice_count, tld);
+  if (slice_count == 0) slice_count = 1;
+  while (sq <= &tld->spans[MI_SEGMENT_BIN_MAX]) {
+    for (mi_slice_t* slice = sq->first; slice != NULL; slice = slice->next) {
+      if (slice->slice_count >= slice_count) {
+	// found one
+	mi_segment_t* segment = _mi_ptr_segment(slice);
+	if (_mi_arena_memid_is_suitable(segment->memid, req_arena_id)) {
+	  // found a suitable page span
+	  mi_span_queue_delete(sq, slice);
+
+	  if (slice->slice_count > slice_count) {
+	    mi_segment_slice_split(segment, slice, slice_count, tld);
+	  }
+	  mi_assert_internal(slice != NULL && slice->slice_count == slice_count && slice->xblock_size > 0);
+	  mi_page_t* page = mi_segment_span_allocate(segment, mi_slice_index(slice), slice->slice_count, tld);
+	  if (page == NULL) {
+	    // commit failed; return NULL but first restore the slice
+	    mi_segment_span_free_coalesce(slice, tld);
+	    return NULL;
+	  }
+	  return page;
+	}
+      }
+    }
+    sq++;
+  }
+  // could not find a page..
+  return NULL;
+}
+
+
+/* -----------------------------------------------------------
+   Segment allocation
+----------------------------------------------------------- */
+
+static mi_segment_t* mi_segment_os_alloc( size_t required, size_t page_alignment, bool eager_delayed, mi_arena_id_t req_arena_id,
+					  size_t* psegment_slices, size_t* ppre_size, size_t* pinfo_slices,
+					  bool commit, mi_segments_tld_t* tld, mi_os_tld_t* os_tld)
+
+{
+  mi_memid_t memid;
+  bool   allow_large = (!eager_delayed && (MI_SECURE == 0)); // only allow large OS pages once we are no longer lazy
+  size_t align_offset = 0;
+  size_t alignment = MI_SEGMENT_ALIGN;
+
+  if (page_alignment > 0) {
+    // mi_assert_internal(huge_page != NULL);
+    mi_assert_internal(page_alignment >= MI_SEGMENT_ALIGN);
+    alignment = page_alignment;
+    const size_t info_size = (*pinfo_slices) * MI_SEGMENT_SLICE_SIZE;
+    align_offset = _mi_align_up( info_size, MI_SEGMENT_ALIGN );
+    const size_t extra = align_offset - info_size;
+    // recalculate due to potential guard pages
+    *psegment_slices = mi_segment_calculate_slices(required + extra, ppre_size, pinfo_slices);
+  }
+
+  const size_t segment_size = (*psegment_slices) * MI_SEGMENT_SLICE_SIZE;
+  mi_segment_t* segment = (mi_segment_t*)_mi_arena_alloc_aligned(segment_size, alignment, align_offset, commit, allow_large, req_arena_id, &memid, os_tld);
+  if (segment == NULL) {
+    return NULL;  // failed to allocate
+  }
+
+  // ensure metadata part of the segment is committed
+  mi_commit_mask_t commit_mask;
+  if (memid.initially_committed) {
+    mi_commit_mask_create_full(&commit_mask);
+  }
+  else {
+    // at least commit the info slices
+    const size_t commit_needed = _mi_divide_up((*pinfo_slices)*MI_SEGMENT_SLICE_SIZE, MI_COMMIT_SIZE);
+    mi_assert_internal(commit_needed>0);
+    mi_commit_mask_create(0, commit_needed, &commit_mask);
+    mi_assert_internal(commit_needed*MI_COMMIT_SIZE >= (*pinfo_slices)*MI_SEGMENT_SLICE_SIZE);
+    if (!_mi_os_commit(segment, commit_needed*MI_COMMIT_SIZE, NULL, tld->stats)) {
+      _mi_arena_free(segment,segment_size,0,memid,tld->stats);
+      return NULL;
+    }
+  }
+  mi_assert_internal(segment != NULL && (uintptr_t)segment % MI_SEGMENT_SIZE == 0);
+
+  segment->memid = memid;
+  segment->allow_decommit = !memid.is_pinned;
+  segment->allow_purge = segment->allow_decommit && (mi_option_get(mi_option_purge_delay) >= 0);
+  segment->segment_size = segment_size;
+  segment->commit_mask = commit_mask;
+  segment->purge_expire = 0;
+  mi_commit_mask_create_empty(&segment->purge_mask);
+  mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, NULL);  // tsan
+
+  mi_segments_track_size((long)(segment_size), tld);
+  _mi_segment_map_allocated_at(segment);
+  return segment;
+}
+
+
+// Allocate a segment from the OS aligned to `MI_SEGMENT_SIZE` .
+static mi_segment_t* mi_segment_alloc(size_t required, size_t page_alignment, mi_arena_id_t req_arena_id, mi_segments_tld_t* tld, mi_os_tld_t* os_tld, mi_page_t** huge_page)
+{
+  mi_assert_internal((required==0 && huge_page==NULL) || (required>0 && huge_page != NULL));
+
+  // calculate needed sizes first
+  size_t info_slices;
+  size_t pre_size;
+  size_t segment_slices = mi_segment_calculate_slices(required, &pre_size, &info_slices);
+
+  // Commit eagerly only if not the first N lazy segments (to reduce impact of many threads that allocate just a little)
+  const bool eager_delay = (// !_mi_os_has_overcommit() &&             // never delay on overcommit systems
+			    _mi_current_thread_count() > 1 &&       // do not delay for the first N threads
+			    tld->count < (size_t)mi_option_get(mi_option_eager_commit_delay));
+  const bool eager = !eager_delay && mi_option_is_enabled(mi_option_eager_commit);
+  bool commit = eager || (required > 0);
+
+  // Allocate the segment from the OS
+  mi_segment_t* segment = mi_segment_os_alloc(required, page_alignment, eager_delay, req_arena_id,
+					      &segment_slices, &pre_size, &info_slices, commit, tld, os_tld);
+  if (segment == NULL) return NULL;
+
+  // zero the segment info? -- not always needed as it may be zero initialized from the OS
+  if (!segment->memid.initially_zero) {
+    ptrdiff_t ofs    = offsetof(mi_segment_t, next);
+    size_t    prefix = offsetof(mi_segment_t, slices) - ofs;
+    size_t    zsize  = prefix + (sizeof(mi_slice_t) * (segment_slices + 1)); // one more
+    _mi_memzero((uint8_t*)segment + ofs, zsize);
+  }
+
+  // initialize the rest of the segment info
+  const size_t slice_entries = (segment_slices > MI_SLICES_PER_SEGMENT ? MI_SLICES_PER_SEGMENT : segment_slices);
+  segment->segment_slices = segment_slices;
+  segment->segment_info_slices = info_slices;
+  segment->thread_id = _mi_thread_id();
+  segment->cookie = _mi_ptr_cookie(segment);
+  segment->slice_entries = slice_entries;
+  segment->kind = (required == 0 ? MI_SEGMENT_NORMAL : MI_SEGMENT_HUGE);
+
+  // _mi_memzero(segment->slices, sizeof(mi_slice_t)*(info_slices+1));
+  _mi_stat_increase(&tld->stats->page_committed, mi_segment_info_size(segment));
+
+  // set up guard pages
+  size_t guard_slices = 0;
+  if (MI_SECURE>0) {
+    // in secure mode, we set up a protected page in between the segment info
+    // and the page data, and at the end of the segment.
+    size_t os_pagesize = _mi_os_page_size();
+    mi_assert_internal(mi_segment_info_size(segment) - os_pagesize >= pre_size);
+    _mi_os_protect((uint8_t*)segment + mi_segment_info_size(segment) - os_pagesize, os_pagesize);
+    uint8_t* end = (uint8_t*)segment + mi_segment_size(segment) - os_pagesize;
+    mi_segment_ensure_committed(segment, end, os_pagesize, tld->stats);
+    _mi_os_protect(end, os_pagesize);
+    if (slice_entries == segment_slices) segment->slice_entries--; // don't use the last slice :-(
+    guard_slices = 1;
+  }
+
+  // reserve first slices for segment info
+  mi_page_t* page0 = mi_segment_span_allocate(segment, 0, info_slices, tld);
+  mi_assert_internal(page0!=NULL); if (page0==NULL) return NULL; // cannot fail as we always commit in advance
+  mi_assert_internal(segment->used == 1);
+  segment->used = 0; // don't count our internal slices towards usage
+
+  // initialize initial free pages
+  if (segment->kind == MI_SEGMENT_NORMAL) { // not a huge page
+    mi_assert_internal(huge_page==NULL);
+    mi_segment_span_free(segment, info_slices, segment->slice_entries - info_slices, false /* don't purge */, tld);
+  }
+  else {
+    mi_assert_internal(huge_page!=NULL);
+    mi_assert_internal(mi_commit_mask_is_empty(&segment->purge_mask));
+    mi_assert_internal(mi_commit_mask_is_full(&segment->commit_mask));
+    *huge_page = mi_segment_span_allocate(segment, info_slices, segment_slices - info_slices - guard_slices, tld);
+    mi_assert_internal(*huge_page != NULL); // cannot fail as we commit in advance
+  }
+
+  mi_assert_expensive(mi_segment_is_valid(segment,tld));
+  return segment;
+}
+
+
+static void mi_segment_free(mi_segment_t* segment, bool force, mi_segments_tld_t* tld) {
+  MI_UNUSED(force);
+  mi_assert_internal(segment != NULL);
+  mi_assert_internal(segment->next == NULL);
+  mi_assert_internal(segment->used == 0);
+
+  // Remove the free pages
+  mi_slice_t* slice = &segment->slices[0];
+  const mi_slice_t* end = mi_segment_slices_end(segment);
+  #if MI_DEBUG>1
+  size_t page_count = 0;
+  #endif
+  while (slice < end) {
+    mi_assert_internal(slice->slice_count > 0);
+    mi_assert_internal(slice->slice_offset == 0);
+    mi_assert_internal(mi_slice_index(slice)==0 || slice->xblock_size == 0); // no more used pages ..
+    if (slice->xblock_size == 0 && segment->kind != MI_SEGMENT_HUGE) {
+      mi_segment_span_remove_from_queue(slice, tld);
+    }
+    #if MI_DEBUG>1
+    page_count++;
+    #endif
+    slice = slice + slice->slice_count;
+  }
+  mi_assert_internal(page_count == 2); // first page is allocated by the segment itself
+
+  // stats
+  _mi_stat_decrease(&tld->stats->page_committed, mi_segment_info_size(segment));
+
+  // return it to the OS
+  mi_segment_os_free(segment, tld);
+}
+
+
+/* -----------------------------------------------------------
+   Page Free
+----------------------------------------------------------- */
+
+static void mi_segment_abandon(mi_segment_t* segment, mi_segments_tld_t* tld);
+
+// note: can be called on abandoned pages
+static mi_slice_t* mi_segment_page_clear(mi_page_t* page, mi_segments_tld_t* tld) {
+  mi_assert_internal(page->xblock_size > 0);
+  mi_assert_internal(mi_page_all_free(page));
+  mi_segment_t* segment = _mi_ptr_segment(page);
+  mi_assert_internal(segment->used > 0);
+
+  size_t inuse = page->capacity * mi_page_block_size(page);
+  _mi_stat_decrease(&tld->stats->page_committed, inuse);
+  _mi_stat_decrease(&tld->stats->pages, 1);
+
+  // reset the page memory to reduce memory pressure?
+  if (segment->allow_decommit && mi_option_is_enabled(mi_option_deprecated_page_reset)) {
+    size_t psize;
+    uint8_t* start = _mi_page_start(segment, page, &psize);
+    _mi_os_reset(start, psize, tld->stats);
+  }
+
+  // zero the page data, but not the segment fields
+  page->is_zero_init = false;
+  ptrdiff_t ofs = offsetof(mi_page_t, capacity);
+  _mi_memzero((uint8_t*)page + ofs, sizeof(*page) - ofs);
+  page->xblock_size = 1;
+
+  // and free it
+  mi_slice_t* slice = mi_segment_span_free_coalesce(mi_page_to_slice(page), tld);
+  segment->used--;
+  // cannot assert segment valid as it is called during reclaim
+  // mi_assert_expensive(mi_segment_is_valid(segment, tld));
+  return slice;
+}
+
+void _mi_segment_page_free(mi_page_t* page, bool force, mi_segments_tld_t* tld)
+{
+  mi_assert(page != NULL);
+
+  mi_segment_t* segment = _mi_page_segment(page);
+  mi_assert_expensive(mi_segment_is_valid(segment,tld));
+
+  // mark it as free now
+  mi_segment_page_clear(page, tld);
+  mi_assert_expensive(mi_segment_is_valid(segment, tld));
+
+  if (segment->used == 0) {
+    // no more used pages; remove from the free list and free the segment
+    mi_segment_free(segment, force, tld);
+  }
+  else if (segment->used == segment->abandoned) {
+    // only abandoned pages; remove from free list and abandon
+    mi_segment_abandon(segment,tld);
+  }
+}
+
+
+/* -----------------------------------------------------------
+Abandonment
+
+When threads terminate, they can leave segments with
+live blocks (reachable through other threads). Such segments
+are "abandoned" and will be reclaimed by other threads to
+reuse their pages and/or free them eventually
+
+We maintain a global list of abandoned segments that are
+reclaimed on demand. Since this is shared among threads
+the implementation needs to avoid the A-B-A problem on
+popping abandoned segments: <https://en.wikipedia.org/wiki/ABA_problem>
+We use tagged pointers to avoid accidentally identifying
+reused segments, much like stamped references in Java.
+Secondly, we maintain a reader counter to avoid resetting
+or decommitting segments that have a pending read operation.
+
+Note: the current implementation is one possible design;
+another way might be to keep track of abandoned segments
+in the arenas/segment_cache's. This would have the advantage of keeping
+all concurrent code in one place and not needing to deal
+with ABA issues. The drawback is that it is unclear how to
+scan abandoned segments efficiently in that case as they
+would be spread among all other segments in the arenas.
+----------------------------------------------------------- */
+
+// Use the bottom 20-bits (on 64-bit) of the aligned segment pointers
+// to put in a tag that increments on update to avoid the A-B-A problem.
+#define MI_TAGGED_MASK   MI_SEGMENT_MASK
+typedef uintptr_t        mi_tagged_segment_t;
+
+static mi_segment_t* mi_tagged_segment_ptr(mi_tagged_segment_t ts) {
+  return (mi_segment_t*)(ts & ~MI_TAGGED_MASK);
+}
+
+static mi_tagged_segment_t mi_tagged_segment(mi_segment_t* segment, mi_tagged_segment_t ts) {
+  mi_assert_internal(((uintptr_t)segment & MI_TAGGED_MASK) == 0);
+  uintptr_t tag = ((ts & MI_TAGGED_MASK) + 1) & MI_TAGGED_MASK;
+  return ((uintptr_t)segment | tag);
+}
+
+// This is a list of visited abandoned pages that were full at the time.
+// this list migrates to `abandoned` when that becomes NULL. The use of
+// this list reduces contention and the rate at which segments are visited.
+static mi_decl_cache_align _Atomic(mi_segment_t*)       abandoned_visited; // = NULL
+
+// The abandoned page list (tagged as it supports pop)
+static mi_decl_cache_align _Atomic(mi_tagged_segment_t) abandoned;         // = NULL
+
+// Maintain these for debug purposes (these counts may be a bit off)
+static mi_decl_cache_align _Atomic(size_t)           abandoned_count;
+static mi_decl_cache_align _Atomic(size_t)           abandoned_visited_count;
+
+// We also maintain a count of current readers of the abandoned list
+// in order to prevent resetting/decommitting segment memory if it might
+// still be read.
+static mi_decl_cache_align _Atomic(size_t)           abandoned_readers; // = 0
+
+// Push on the visited list
+static void mi_abandoned_visited_push(mi_segment_t* segment) {
+  mi_assert_internal(segment->thread_id == 0);
+  mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_segment_t,&segment->abandoned_next) == NULL);
+  mi_assert_internal(segment->next == NULL);
+  mi_assert_internal(segment->used > 0);
+  mi_segment_t* anext = mi_atomic_load_ptr_relaxed(mi_segment_t, &abandoned_visited);
+  do {
+    mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, anext);
+  } while (!mi_atomic_cas_ptr_weak_release(mi_segment_t, &abandoned_visited, &anext, segment));
+  mi_atomic_increment_relaxed(&abandoned_visited_count);
+}
+
+// Move the visited list to the abandoned list.
+static bool mi_abandoned_visited_revisit(void)
+{
+  // quick check if the visited list is empty
+  if (mi_atomic_load_ptr_relaxed(mi_segment_t, &abandoned_visited) == NULL) return false;
+
+  // grab the whole visited list
+  mi_segment_t* first = mi_atomic_exchange_ptr_acq_rel(mi_segment_t, &abandoned_visited, NULL);
+  if (first == NULL) return false;
+
+  // first try to swap directly if the abandoned list happens to be NULL
+  mi_tagged_segment_t afirst;
+  mi_tagged_segment_t ts = mi_atomic_load_relaxed(&abandoned);
+  if (mi_tagged_segment_ptr(ts)==NULL) {
+    size_t count = mi_atomic_load_relaxed(&abandoned_visited_count);
+    afirst = mi_tagged_segment(first, ts);
+    if (mi_atomic_cas_strong_acq_rel(&abandoned, &ts, afirst)) {
+      mi_atomic_add_relaxed(&abandoned_count, count);
+      mi_atomic_sub_relaxed(&abandoned_visited_count, count);
+      return true;
+    }
+  }
+
+  // find the last element of the visited list: O(n)
+  mi_segment_t* last = first;
+  mi_segment_t* next;
+  while ((next = mi_atomic_load_ptr_relaxed(mi_segment_t, &last->abandoned_next)) != NULL) {
+    last = next;
+  }
+
+  // and atomically prepend to the abandoned list
+  // (no need to increase the readers as we don't access the abandoned segments)
+  mi_tagged_segment_t anext = mi_atomic_load_relaxed(&abandoned);
+  size_t count;
+  do {
+    count = mi_atomic_load_relaxed(&abandoned_visited_count);
+    mi_atomic_store_ptr_release(mi_segment_t, &last->abandoned_next, mi_tagged_segment_ptr(anext));
+    afirst = mi_tagged_segment(first, anext);
+  } while (!mi_atomic_cas_weak_release(&abandoned, &anext, afirst));
+  mi_atomic_add_relaxed(&abandoned_count, count);
+  mi_atomic_sub_relaxed(&abandoned_visited_count, count);
+  return true;
+}
+
+// Push on the abandoned list.
+static void mi_abandoned_push(mi_segment_t* segment) {
+  mi_assert_internal(segment->thread_id == 0);
+  mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_segment_t, &segment->abandoned_next) == NULL);
+  mi_assert_internal(segment->next == NULL);
+  mi_assert_internal(segment->used > 0);
+  mi_tagged_segment_t next;
+  mi_tagged_segment_t ts = mi_atomic_load_relaxed(&abandoned);
+  do {
+    mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, mi_tagged_segment_ptr(ts));
+    next = mi_tagged_segment(segment, ts);
+  } while (!mi_atomic_cas_weak_release(&abandoned, &ts, next));
+  mi_atomic_increment_relaxed(&abandoned_count);
+}
+
+// Wait until there are no more pending reads on segments that used to be in the abandoned list
+// called for example from `arena.c` before decommitting
+void _mi_abandoned_await_readers(void) {
+  size_t n;
+  do {
+    n = mi_atomic_load_acquire(&abandoned_readers);
+    if (n != 0) mi_atomic_yield();
+  } while (n != 0);
+}
+
+// Pop from the abandoned list
+static mi_segment_t* mi_abandoned_pop(void) {
+  mi_segment_t* segment;
+  // Check efficiently if it is empty (or if the visited list needs to be moved)
+  mi_tagged_segment_t ts = mi_atomic_load_relaxed(&abandoned);
+  segment = mi_tagged_segment_ptr(ts);
+  if mi_likely(segment == NULL) {
+    if mi_likely(!mi_abandoned_visited_revisit()) { // try to swap in the visited list on NULL
+      return NULL;
+    }
+  }
+
+  // Do a pop. We use a reader count to prevent
+  // a segment to be decommitted while a read is still pending,
+  // and a tagged pointer to prevent A-B-A link corruption.
+  // (this is called from `region.c:_mi_mem_free` for example)
+  mi_atomic_increment_relaxed(&abandoned_readers);  // ensure no segment gets decommitted
+  mi_tagged_segment_t next = 0;
+  ts = mi_atomic_load_acquire(&abandoned);
+  do {
+    segment = mi_tagged_segment_ptr(ts);
+    if (segment != NULL) {
+      mi_segment_t* anext = mi_atomic_load_ptr_relaxed(mi_segment_t, &segment->abandoned_next);
+      next = mi_tagged_segment(anext, ts); // note: reads the segment's `abandoned_next` field so should not be decommitted
+    }
+  } while (segment != NULL && !mi_atomic_cas_weak_acq_rel(&abandoned, &ts, next));
+  mi_atomic_decrement_relaxed(&abandoned_readers);  // release reader lock
+  if (segment != NULL) {
+    mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, NULL);
+    mi_atomic_decrement_relaxed(&abandoned_count);
+  }
+  return segment;
+}
+
+/* -----------------------------------------------------------
+   Abandon segment/page
+----------------------------------------------------------- */
+
+static void mi_segment_abandon(mi_segment_t* segment, mi_segments_tld_t* tld) {
+  mi_assert_internal(segment->used == segment->abandoned);
+  mi_assert_internal(segment->used > 0);
+  mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_segment_t, &segment->abandoned_next) == NULL);
+  mi_assert_internal(segment->abandoned_visits == 0);
+  mi_assert_expensive(mi_segment_is_valid(segment,tld));
+
+  // remove the free pages from the free page queues
+  mi_slice_t* slice = &segment->slices[0];
+  const mi_slice_t* end = mi_segment_slices_end(segment);
+  while (slice < end) {
+    mi_assert_internal(slice->slice_count > 0);
+    mi_assert_internal(slice->slice_offset == 0);
+    if (slice->xblock_size == 0) { // a free page
+      mi_segment_span_remove_from_queue(slice,tld);
+      slice->xblock_size = 0; // but keep it free
+    }
+    slice = slice + slice->slice_count;
+  }
+
+  // perform delayed decommits (forcing is much slower on mstress)
+  mi_segment_try_purge(segment, mi_option_is_enabled(mi_option_abandoned_page_purge) /* force? */, tld->stats);
+
+  // all pages in the segment are abandoned; add it to the abandoned list
+  _mi_stat_increase(&tld->stats->segments_abandoned, 1);
+  mi_segments_track_size(-((long)mi_segment_size(segment)), tld);
+  segment->thread_id = 0;
+  mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, NULL);
+  segment->abandoned_visits = 1;   // from 0 to 1 to signify it is abandoned
+  mi_abandoned_push(segment);
+}
+
+void _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld) {
+  mi_assert(page != NULL);
+  mi_assert_internal(mi_page_thread_free_flag(page)==MI_NEVER_DELAYED_FREE);
+  mi_assert_internal(mi_page_heap(page) == NULL);
+  mi_segment_t* segment = _mi_page_segment(page);
+
+  mi_assert_expensive(mi_segment_is_valid(segment,tld));
+  segment->abandoned++;
+
+  _mi_stat_increase(&tld->stats->pages_abandoned, 1);
+  mi_assert_internal(segment->abandoned <= segment->used);
+  if (segment->used == segment->abandoned) {
+    // all pages are abandoned, abandon the entire segment
+    mi_segment_abandon(segment, tld);
+  }
+}
+
+/* -----------------------------------------------------------
+  Reclaim abandoned pages
+----------------------------------------------------------- */
+
+static mi_slice_t* mi_slices_start_iterate(mi_segment_t* segment, const mi_slice_t** end) {
+  mi_slice_t* slice = &segment->slices[0];
+  *end = mi_segment_slices_end(segment);
+  mi_assert_internal(slice->slice_count>0 && slice->xblock_size>0); // segment allocated page
+  slice = slice + slice->slice_count; // skip the first segment allocated page
+  return slice;
+}
+
+// Possibly free pages and check if free space is available
+static bool mi_segment_check_free(mi_segment_t* segment, size_t slices_needed, size_t block_size, mi_segments_tld_t* tld)
+{
+  mi_assert_internal(block_size < MI_HUGE_BLOCK_SIZE);
+  mi_assert_internal(mi_segment_is_abandoned(segment));
+  bool has_page = false;
+
+  // for all slices
+  const mi_slice_t* end;
+  mi_slice_t* slice = mi_slices_start_iterate(segment, &end);
+  while (slice < end) {
+    mi_assert_internal(slice->slice_count > 0);
+    mi_assert_internal(slice->slice_offset == 0);
+    if (mi_slice_is_used(slice)) { // used page
+      // ensure used count is up to date and collect potential concurrent frees
+      mi_page_t* const page = mi_slice_to_page(slice);
+      _mi_page_free_collect(page, false);
+      if (mi_page_all_free(page)) {
+	// if this page is all free now, free it without adding to any queues (yet)
+	mi_assert_internal(page->next == NULL && page->prev==NULL);
+	_mi_stat_decrease(&tld->stats->pages_abandoned, 1);
+	segment->abandoned--;
+	slice = mi_segment_page_clear(page, tld); // re-assign slice due to coalesce!
+	mi_assert_internal(!mi_slice_is_used(slice));
+	if (slice->slice_count >= slices_needed) {
+	  has_page = true;
+	}
+      }
+      else {
+	if (page->xblock_size == block_size && mi_page_has_any_available(page)) {
+	  // a page has available free blocks of the right size
+	  has_page = true;
+	}
+      }
+    }
+    else {
+      // empty span
+      if (slice->slice_count >= slices_needed) {
+	has_page = true;
+      }
+    }
+    slice = slice + slice->slice_count;
+  }
+  return has_page;
+}
+
+// Reclaim an abandoned segment; returns NULL if the segment was freed
+// set `right_page_reclaimed` to `true` if it reclaimed a page of the right `block_size` that was not full.
+static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap, size_t requested_block_size, bool* right_page_reclaimed, mi_segments_tld_t* tld) {
+  mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_segment_t, &segment->abandoned_next) == NULL);
+  mi_assert_expensive(mi_segment_is_valid(segment, tld));
+  if (right_page_reclaimed != NULL) { *right_page_reclaimed = false; }
+
+  segment->thread_id = _mi_thread_id();
+  segment->abandoned_visits = 0;
+  mi_segments_track_size((long)mi_segment_size(segment), tld);
+  mi_assert_internal(segment->next == NULL);
+  _mi_stat_decrease(&tld->stats->segments_abandoned, 1);
+
+  // for all slices
+  const mi_slice_t* end;
+  mi_slice_t* slice = mi_slices_start_iterate(segment, &end);
+  while (slice < end) {
+    mi_assert_internal(slice->slice_count > 0);
+    mi_assert_internal(slice->slice_offset == 0);
+    if (mi_slice_is_used(slice)) {
+      // in use: reclaim the page in our heap
+      mi_page_t* page = mi_slice_to_page(slice);
+      mi_assert_internal(page->is_committed);
+      mi_assert_internal(mi_page_thread_free_flag(page)==MI_NEVER_DELAYED_FREE);
+      mi_assert_internal(mi_page_heap(page) == NULL);
+      mi_assert_internal(page->next == NULL && page->prev==NULL);
+      _mi_stat_decrease(&tld->stats->pages_abandoned, 1);
+      segment->abandoned--;
+      // set the heap again and allow delayed free again
+      mi_page_set_heap(page, heap);
+      _mi_page_use_delayed_free(page, MI_USE_DELAYED_FREE, true); // override never (after heap is set)
+      _mi_page_free_collect(page, false); // ensure used count is up to date
+      if (mi_page_all_free(page)) {
+	// if everything free by now, free the page
+	slice = mi_segment_page_clear(page, tld);   // set slice again due to coalesceing
+      }
+      else {
+	// otherwise reclaim it into the heap
+	_mi_page_reclaim(heap, page);
+	if (requested_block_size == page->xblock_size && mi_page_has_any_available(page)) {
+	  if (right_page_reclaimed != NULL) { *right_page_reclaimed = true; }
+	}
+      }
+    }
+    else {
+      // the span is free, add it to our page queues
+      slice = mi_segment_span_free_coalesce(slice, tld); // set slice again due to coalesceing
+    }
+    mi_assert_internal(slice->slice_count>0 && slice->slice_offset==0);
+    slice = slice + slice->slice_count;
+  }
+
+  mi_assert(segment->abandoned == 0);
+  if (segment->used == 0) {  // due to page_clear
+    mi_assert_internal(right_page_reclaimed == NULL || !(*right_page_reclaimed));
+    mi_segment_free(segment, false, tld);
+    return NULL;
+  }
+  else {
+    return segment;
+  }
+}
+
+
+void _mi_abandoned_reclaim_all(mi_heap_t* heap, mi_segments_tld_t* tld) {
+  mi_segment_t* segment;
+  while ((segment = mi_abandoned_pop()) != NULL) {
+    mi_segment_reclaim(segment, heap, 0, NULL, tld);
+  }
+}
+
+static mi_segment_t* mi_segment_try_reclaim(mi_heap_t* heap, size_t needed_slices, size_t block_size, bool* reclaimed, mi_segments_tld_t* tld)
+{
+  *reclaimed = false;
+  mi_segment_t* segment;
+  long max_tries = mi_option_get_clamp(mi_option_max_segment_reclaim, 8, 1024);     // limit the work to bound allocation times
+  while ((max_tries-- > 0) && ((segment = mi_abandoned_pop()) != NULL)) {
+    segment->abandoned_visits++;
+    // todo: an arena exclusive heap will potentially visit many abandoned unsuitable segments
+    // and push them into the visited list and use many tries. Perhaps we can skip non-suitable ones in a better way?
+    bool is_suitable = _mi_heap_memid_is_suitable(heap, segment->memid);
+    bool has_page = mi_segment_check_free(segment,needed_slices,block_size,tld); // try to free up pages (due to concurrent frees)
+    if (segment->used == 0) {
+      // free the segment (by forced reclaim) to make it available to other threads.
+      // note1: we prefer to free a segment as that might lead to reclaiming another
+      // segment that is still partially used.
+      // note2: we could in principle optimize this by skipping reclaim and directly
+      // freeing but that would violate some invariants temporarily)
+      mi_segment_reclaim(segment, heap, 0, NULL, tld);
+    }
+    else if (has_page && is_suitable) {
+      // found a large enough free span, or a page of the right block_size with free space
+      // we return the result of reclaim (which is usually `segment`) as it might free
+      // the segment due to concurrent frees (in which case `NULL` is returned).
+      return mi_segment_reclaim(segment, heap, block_size, reclaimed, tld);
+    }
+    else if (segment->abandoned_visits > 3 && is_suitable) {
+      // always reclaim on 3rd visit to limit the abandoned queue length.
+      mi_segment_reclaim(segment, heap, 0, NULL, tld);
+    }
+    else {
+      // otherwise, push on the visited list so it gets not looked at too quickly again
+      mi_segment_try_purge(segment, true /* force? */, tld->stats); // force purge if needed as we may not visit soon again
+      mi_abandoned_visited_push(segment);
+    }
+  }
+  return NULL;
+}
+
+
+void _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t* tld)
+{
+  mi_segment_t* segment;
+  int max_tries = (force ? 16*1024 : 1024); // limit latency
+  if (force) {
+    mi_abandoned_visited_revisit();
+  }
+  while ((max_tries-- > 0) && ((segment = mi_abandoned_pop()) != NULL)) {
+    mi_segment_check_free(segment,0,0,tld); // try to free up pages (due to concurrent frees)
+    if (segment->used == 0) {
+      // free the segment (by forced reclaim) to make it available to other threads.
+      // note: we could in principle optimize this by skipping reclaim and directly
+      // freeing but that would violate some invariants temporarily)
+      mi_segment_reclaim(segment, heap, 0, NULL, tld);
+    }
+    else {
+      // otherwise, purge if needed and push on the visited list
+      // note: forced purge can be expensive if many threads are destroyed/created as in mstress.
+      mi_segment_try_purge(segment, force, tld->stats);
+      mi_abandoned_visited_push(segment);
+    }
+  }
+}
+
+/* -----------------------------------------------------------
+   Reclaim or allocate
+----------------------------------------------------------- */
+
+static mi_segment_t* mi_segment_reclaim_or_alloc(mi_heap_t* heap, size_t needed_slices, size_t block_size, mi_segments_tld_t* tld, mi_os_tld_t* os_tld)
+{
+  mi_assert_internal(block_size < MI_HUGE_BLOCK_SIZE);
+  mi_assert_internal(block_size <= MI_LARGE_OBJ_SIZE_MAX);
+
+  // 1. try to reclaim an abandoned segment
+  bool reclaimed;
+  mi_segment_t* segment = mi_segment_try_reclaim(heap, needed_slices, block_size, &reclaimed, tld);
+  if (reclaimed) {
+    // reclaimed the right page right into the heap
+    mi_assert_internal(segment != NULL);
+    return NULL; // pretend out-of-memory as the page will be in the page queue of the heap with available blocks
+  }
+  else if (segment != NULL) {
+    // reclaimed a segment with a large enough empty span in it
+    return segment;
+  }
+  // 2. otherwise allocate a fresh segment
+  return mi_segment_alloc(0, 0, heap->arena_id, tld, os_tld, NULL);
+}
+
+
+/* -----------------------------------------------------------
+   Page allocation
+----------------------------------------------------------- */
+
+static mi_page_t* mi_segments_page_alloc(mi_heap_t* heap, mi_page_kind_t page_kind, size_t required, size_t block_size, mi_segments_tld_t* tld, mi_os_tld_t* os_tld)
+{
+  mi_assert_internal(required <= MI_LARGE_OBJ_SIZE_MAX && page_kind <= MI_PAGE_LARGE);
+
+  // find a free page
+  size_t page_size = _mi_align_up(required, (required > MI_MEDIUM_PAGE_SIZE ? MI_MEDIUM_PAGE_SIZE : MI_SEGMENT_SLICE_SIZE));
+  size_t slices_needed = page_size / MI_SEGMENT_SLICE_SIZE;
+  mi_assert_internal(slices_needed * MI_SEGMENT_SLICE_SIZE == page_size);
+  mi_page_t* page = mi_segments_page_find_and_allocate(slices_needed, heap->arena_id, tld); //(required <= MI_SMALL_SIZE_MAX ? 0 : slices_needed), tld);
+  if (page==NULL) {
+    // no free page, allocate a new segment and try again
+    if (mi_segment_reclaim_or_alloc(heap, slices_needed, block_size, tld, os_tld) == NULL) {
+      // OOM or reclaimed a good page in the heap
+      return NULL;
+    }
+    else {
+      // otherwise try again
+      return mi_segments_page_alloc(heap, page_kind, required, block_size, tld, os_tld);
+    }
+  }
+  mi_assert_internal(page != NULL && page->slice_count*MI_SEGMENT_SLICE_SIZE == page_size);
+  mi_assert_internal(_mi_ptr_segment(page)->thread_id == _mi_thread_id());
+  mi_segment_try_purge(_mi_ptr_segment(page), false, tld->stats);
+  return page;
+}
+
+
+
+/* -----------------------------------------------------------
+   Huge page allocation
+----------------------------------------------------------- */
+
+static mi_page_t* mi_segment_huge_page_alloc(size_t size, size_t page_alignment, mi_arena_id_t req_arena_id, mi_segments_tld_t* tld, mi_os_tld_t* os_tld)
+{
+  mi_page_t* page = NULL;
+  mi_segment_t* segment = mi_segment_alloc(size,page_alignment,req_arena_id,tld,os_tld,&page);
+  if (segment == NULL || page==NULL) return NULL;
+  mi_assert_internal(segment->used==1);
+  mi_assert_internal(mi_page_block_size(page) >= size);
+  #if MI_HUGE_PAGE_ABANDON
+  segment->thread_id = 0; // huge segments are immediately abandoned
+  #endif
+
+  // for huge pages we initialize the xblock_size as we may
+  // overallocate to accommodate large alignments.
+  size_t psize;
+  uint8_t* start = _mi_segment_page_start(segment, page, &psize);
+  page->xblock_size = (psize > MI_HUGE_BLOCK_SIZE ? MI_HUGE_BLOCK_SIZE : (uint32_t)psize);
+
+  // decommit the part of the prefix of a page that will not be used; this can be quite large (close to MI_SEGMENT_SIZE)
+  if (page_alignment > 0 && segment->allow_decommit) {
+    uint8_t* aligned_p = (uint8_t*)_mi_align_up((uintptr_t)start, page_alignment);
+    mi_assert_internal(_mi_is_aligned(aligned_p, page_alignment));
+    mi_assert_internal(psize - (aligned_p - start) >= size);
+    uint8_t* decommit_start = start + sizeof(mi_block_t);              // for the free list
+    ptrdiff_t decommit_size = aligned_p - decommit_start;
+    _mi_os_reset(decommit_start, decommit_size, &_mi_stats_main);   // note: cannot use segment_decommit on huge segments
+  }
+
+  return page;
+}
+
+#if MI_HUGE_PAGE_ABANDON
+// free huge block from another thread
+void _mi_segment_huge_page_free(mi_segment_t* segment, mi_page_t* page, mi_block_t* block) {
+  // huge page segments are always abandoned and can be freed immediately by any thread
+  mi_assert_internal(segment->kind==MI_SEGMENT_HUGE);
+  mi_assert_internal(segment == _mi_page_segment(page));
+  mi_assert_internal(mi_atomic_load_relaxed(&segment->thread_id)==0);
+
+  // claim it and free
+  mi_heap_t* heap = mi_heap_get_default(); // issue #221; don't use the internal get_default_heap as we need to ensure the thread is initialized.
+  // paranoia: if this it the last reference, the cas should always succeed
+  size_t expected_tid = 0;
+  if (mi_atomic_cas_strong_acq_rel(&segment->thread_id, &expected_tid, heap->thread_id)) {
+    mi_block_set_next(page, block, page->free);
+    page->free = block;
+    page->used--;
+    page->is_zero = false;
+    mi_assert(page->used == 0);
+    mi_tld_t* tld = heap->tld;
+    _mi_segment_page_free(page, true, &tld->segments);
+  }
+#if (MI_DEBUG!=0)
+  else {
+    mi_assert_internal(false);
+  }
+#endif
+}
+
+#else
+// reset memory of a huge block from another thread
+void _mi_segment_huge_page_reset(mi_segment_t* segment, mi_page_t* page, mi_block_t* block) {
+  MI_UNUSED(page);
+  mi_assert_internal(segment->kind == MI_SEGMENT_HUGE);
+  mi_assert_internal(segment == _mi_page_segment(page));
+  mi_assert_internal(page->used == 1); // this is called just before the free
+  mi_assert_internal(page->free == NULL);
+  if (segment->allow_decommit) {
+    size_t csize = mi_usable_size(block);
+    if (csize > sizeof(mi_block_t)) {
+      csize = csize - sizeof(mi_block_t);
+      uint8_t* p = (uint8_t*)block + sizeof(mi_block_t);
+      _mi_os_reset(p, csize, &_mi_stats_main);  // note: cannot use segment_decommit on huge segments
+    }
+  }
+}
+#endif
+
+/* -----------------------------------------------------------
+   Page allocation and free
+----------------------------------------------------------- */
+mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, size_t page_alignment, mi_segments_tld_t* tld, mi_os_tld_t* os_tld) {
+  mi_page_t* page;
+  if mi_unlikely(page_alignment > MI_ALIGNMENT_MAX) {
+    mi_assert_internal(_mi_is_power_of_two(page_alignment));
+    mi_assert_internal(page_alignment >= MI_SEGMENT_SIZE);
+    if (page_alignment < MI_SEGMENT_SIZE) { page_alignment = MI_SEGMENT_SIZE; }
+    page = mi_segment_huge_page_alloc(block_size,page_alignment,heap->arena_id,tld,os_tld);
+  }
+  else if (block_size <= MI_SMALL_OBJ_SIZE_MAX) {
+    page = mi_segments_page_alloc(heap,MI_PAGE_SMALL,block_size,block_size,tld,os_tld);
+  }
+  else if (block_size <= MI_MEDIUM_OBJ_SIZE_MAX) {
+    page = mi_segments_page_alloc(heap,MI_PAGE_MEDIUM,MI_MEDIUM_PAGE_SIZE,block_size,tld, os_tld);
+  }
+  else if (block_size <= MI_LARGE_OBJ_SIZE_MAX) {
+    page = mi_segments_page_alloc(heap,MI_PAGE_LARGE,block_size,block_size,tld, os_tld);
+  }
+  else {
+    page = mi_segment_huge_page_alloc(block_size,page_alignment,heap->arena_id,tld,os_tld);
+  }
+  mi_assert_internal(page == NULL || _mi_heap_memid_is_suitable(heap, _mi_page_segment(page)->memid));
+  mi_assert_expensive(page == NULL || mi_segment_is_valid(_mi_page_segment(page),tld));
+  return page;
+}
diff --git a/compat/mimalloc/stats.c b/compat/mimalloc/stats.c
new file mode 100644
index 00000000000000..6817e07aa1ee9f
--- /dev/null
+++ b/compat/mimalloc/stats.c
@@ -0,0 +1,467 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"
+
+#include <stdio.h>  // snprintf
+#include <string.h> // memset
+
+#if defined(_MSC_VER) && (_MSC_VER < 1920)
+#pragma warning(disable:4204)  // non-constant aggregate initializer
+#endif
+
+/* -----------------------------------------------------------
+  Statistics operations
+----------------------------------------------------------- */
+
+static bool mi_is_in_main(void* stat) {
+  return ((uint8_t*)stat >= (uint8_t*)&_mi_stats_main
+	 && (uint8_t*)stat < ((uint8_t*)&_mi_stats_main + sizeof(mi_stats_t)));
+}
+
+static void mi_stat_update(mi_stat_count_t* stat, int64_t amount) {
+  if (amount == 0) return;
+  if (mi_is_in_main(stat))
+  {
+    // add atomically (for abandoned pages)
+    int64_t current = mi_atomic_addi64_relaxed(&stat->current, amount);
+    mi_atomic_maxi64_relaxed(&stat->peak, current + amount);
+    if (amount > 0) {
+      mi_atomic_addi64_relaxed(&stat->allocated,amount);
+    }
+    else {
+      mi_atomic_addi64_relaxed(&stat->freed, -amount);
+    }
+  }
+  else {
+    // add thread local
+    stat->current += amount;
+    if (stat->current > stat->peak) stat->peak = stat->current;
+    if (amount > 0) {
+      stat->allocated += amount;
+    }
+    else {
+      stat->freed += -amount;
+    }
+  }
+}
+
+void _mi_stat_counter_increase(mi_stat_counter_t* stat, size_t amount) {
+  if (mi_is_in_main(stat)) {
+    mi_atomic_addi64_relaxed( &stat->count, 1 );
+    mi_atomic_addi64_relaxed( &stat->total, (int64_t)amount );
+  }
+  else {
+    stat->count++;
+    stat->total += amount;
+  }
+}
+
+void _mi_stat_increase(mi_stat_count_t* stat, size_t amount) {
+  mi_stat_update(stat, (int64_t)amount);
+}
+
+void _mi_stat_decrease(mi_stat_count_t* stat, size_t amount) {
+  mi_stat_update(stat, -((int64_t)amount));
+}
+
+// must be thread safe as it is called from stats_merge
+static void mi_stat_add(mi_stat_count_t* stat, const mi_stat_count_t* src, int64_t unit) {
+  if (stat==src) return;
+  if (src->allocated==0 && src->freed==0) return;
+  mi_atomic_addi64_relaxed( &stat->allocated, src->allocated * unit);
+  mi_atomic_addi64_relaxed( &stat->current, src->current * unit);
+  mi_atomic_addi64_relaxed( &stat->freed, src->freed * unit);
+  // peak scores do not work across threads..
+  mi_atomic_addi64_relaxed( &stat->peak, src->peak * unit);
+}
+
+static void mi_stat_counter_add(mi_stat_counter_t* stat, const mi_stat_counter_t* src, int64_t unit) {
+  if (stat==src) return;
+  mi_atomic_addi64_relaxed( &stat->total, src->total * unit);
+  mi_atomic_addi64_relaxed( &stat->count, src->count * unit);
+}
+
+// must be thread safe as it is called from stats_merge
+static void mi_stats_add(mi_stats_t* stats, const mi_stats_t* src) {
+  if (stats==src) return;
+  mi_stat_add(&stats->segments, &src->segments,1);
+  mi_stat_add(&stats->pages, &src->pages,1);
+  mi_stat_add(&stats->reserved, &src->reserved, 1);
+  mi_stat_add(&stats->committed, &src->committed, 1);
+  mi_stat_add(&stats->reset, &src->reset, 1);
+  mi_stat_add(&stats->purged, &src->purged, 1);
+  mi_stat_add(&stats->page_committed, &src->page_committed, 1);
+
+  mi_stat_add(&stats->pages_abandoned, &src->pages_abandoned, 1);
+  mi_stat_add(&stats->segments_abandoned, &src->segments_abandoned, 1);
+  mi_stat_add(&stats->threads, &src->threads, 1);
+
+  mi_stat_add(&stats->malloc, &src->malloc, 1);
+  mi_stat_add(&stats->segments_cache, &src->segments_cache, 1);
+  mi_stat_add(&stats->normal, &src->normal, 1);
+  mi_stat_add(&stats->huge, &src->huge, 1);
+  mi_stat_add(&stats->large, &src->large, 1);
+
+  mi_stat_counter_add(&stats->pages_extended, &src->pages_extended, 1);
+  mi_stat_counter_add(&stats->mmap_calls, &src->mmap_calls, 1);
+  mi_stat_counter_add(&stats->commit_calls, &src->commit_calls, 1);
+  mi_stat_counter_add(&stats->reset_calls, &src->reset_calls, 1);
+  mi_stat_counter_add(&stats->purge_calls, &src->purge_calls, 1);
+
+  mi_stat_counter_add(&stats->page_no_retire, &src->page_no_retire, 1);
+  mi_stat_counter_add(&stats->searches, &src->searches, 1);
+  mi_stat_counter_add(&stats->normal_count, &src->normal_count, 1);
+  mi_stat_counter_add(&stats->huge_count, &src->huge_count, 1);
+  mi_stat_counter_add(&stats->large_count, &src->large_count, 1);
+#if MI_STAT>1
+  for (size_t i = 0; i <= MI_BIN_HUGE; i++) {
+    if (src->normal_bins[i].allocated > 0 || src->normal_bins[i].freed > 0) {
+      mi_stat_add(&stats->normal_bins[i], &src->normal_bins[i], 1);
+    }
+  }
+#endif
+}
+
+/* -----------------------------------------------------------
+  Display statistics
+----------------------------------------------------------- */
+
+// unit > 0 : size in binary bytes
+// unit == 0: count as decimal
+// unit < 0 : count in binary
+static void mi_printf_amount(int64_t n, int64_t unit, mi_output_fun* out, void* arg, const char* fmt) {
+  char buf[32]; buf[0] = 0;
+  int  len = 32;
+  const char* suffix = (unit <= 0 ? " " : "B");
+  const int64_t base = (unit == 0 ? 1000 : 1024);
+  if (unit>0) n *= unit;
+
+  const int64_t pos = (n < 0 ? -n : n);
+  if (pos < base) {
+    if (n!=1 || suffix[0] != 'B') {  // skip printing 1 B for the unit column
+      snprintf(buf, len, "%d   %-3s", (int)n, (n==0 ? "" : suffix));
+    }
+  }
+  else {
+    int64_t divider = base;
+    const char* magnitude = "K";
+    if (pos >= divider*base) { divider *= base; magnitude = "M"; }
+    if (pos >= divider*base) { divider *= base; magnitude = "G"; }
+    const int64_t tens = (n / (divider/10));
+    const long whole = (long)(tens/10);
+    const long frac1 = (long)(tens%10);
+    char unitdesc[8];
+    snprintf(unitdesc, 8, "%s%s%s", magnitude, (base==1024 ? "i" : ""), suffix);
+    snprintf(buf, len, "%ld.%ld %-3s", whole, (frac1 < 0 ? -frac1 : frac1), unitdesc);
+  }
+  _mi_fprintf(out, arg, (fmt==NULL ? "%12s" : fmt), buf);
+}
+
+
+static void mi_print_amount(int64_t n, int64_t unit, mi_output_fun* out, void* arg) {
+  mi_printf_amount(n,unit,out,arg,NULL);
+}
+
+static void mi_print_count(int64_t n, int64_t unit, mi_output_fun* out, void* arg) {
+  if (unit==1) _mi_fprintf(out, arg, "%12s"," ");
+	  else mi_print_amount(n,0,out,arg);
+}
+
+static void mi_stat_print_ex(const mi_stat_count_t* stat, const char* msg, int64_t unit, mi_output_fun* out, void* arg, const char* notok ) {
+  _mi_fprintf(out, arg,"%10s:", msg);
+  if (unit > 0) {
+    mi_print_amount(stat->peak, unit, out, arg);
+    mi_print_amount(stat->allocated, unit, out, arg);
+    mi_print_amount(stat->freed, unit, out, arg);
+    mi_print_amount(stat->current, unit, out, arg);
+    mi_print_amount(unit, 1, out, arg);
+    mi_print_count(stat->allocated, unit, out, arg);
+    if (stat->allocated > stat->freed) {
+      _mi_fprintf(out, arg, "  ");
+      _mi_fprintf(out, arg, (notok == NULL ? "not all freed" : notok));
+      _mi_fprintf(out, arg, "\n");
+    }
+    else {
+      _mi_fprintf(out, arg, "  ok\n");
+    }
+  }
+  else if (unit<0) {
+    mi_print_amount(stat->peak, -1, out, arg);
+    mi_print_amount(stat->allocated, -1, out, arg);
+    mi_print_amount(stat->freed, -1, out, arg);
+    mi_print_amount(stat->current, -1, out, arg);
+    if (unit==-1) {
+      _mi_fprintf(out, arg, "%24s", "");
+    }
+    else {
+      mi_print_amount(-unit, 1, out, arg);
+      mi_print_count((stat->allocated / -unit), 0, out, arg);
+    }
+    if (stat->allocated > stat->freed)
+      _mi_fprintf(out, arg, "  not all freed!\n");
+    else
+      _mi_fprintf(out, arg, "  ok\n");
+  }
+  else {
+    mi_print_amount(stat->peak, 1, out, arg);
+    mi_print_amount(stat->allocated, 1, out, arg);
+    _mi_fprintf(out, arg, "%11s", " ");  // no freed
+    mi_print_amount(stat->current, 1, out, arg);
+    _mi_fprintf(out, arg, "\n");
+  }
+}
+
+static void mi_stat_print(const mi_stat_count_t* stat, const char* msg, int64_t unit, mi_output_fun* out, void* arg) {
+  mi_stat_print_ex(stat, msg, unit, out, arg, NULL);
+}
+
+static void mi_stat_peak_print(const mi_stat_count_t* stat, const char* msg, int64_t unit, mi_output_fun* out, void* arg) {
+  _mi_fprintf(out, arg, "%10s:", msg);
+  mi_print_amount(stat->peak, unit, out, arg);
+  _mi_fprintf(out, arg, "\n");
+}
+
+static void mi_stat_counter_print(const mi_stat_counter_t* stat, const char* msg, mi_output_fun* out, void* arg ) {
+  _mi_fprintf(out, arg, "%10s:", msg);
+  mi_print_amount(stat->total, -1, out, arg);
+  _mi_fprintf(out, arg, "\n");
+}
+
+
+static void mi_stat_counter_print_avg(const mi_stat_counter_t* stat, const char* msg, mi_output_fun* out, void* arg) {
+  const int64_t avg_tens = (stat->count == 0 ? 0 : (stat->total*10 / stat->count));
+  const long avg_whole = (long)(avg_tens/10);
+  const long avg_frac1 = (long)(avg_tens%10);
+  _mi_fprintf(out, arg, "%10s: %5ld.%ld avg\n", msg, avg_whole, avg_frac1);
+}
+
+
+static void mi_print_header(mi_output_fun* out, void* arg ) {
+  _mi_fprintf(out, arg, "%10s: %11s %11s %11s %11s %11s %11s\n", "heap stats", "peak   ", "total   ", "freed   ", "current   ", "unit   ", "count   ");
+}
+
+#if MI_STAT>1
+static void mi_stats_print_bins(const mi_stat_count_t* bins, size_t max, const char* fmt, mi_output_fun* out, void* arg) {
+  bool found = false;
+  char buf[64];
+  for (size_t i = 0; i <= max; i++) {
+    if (bins[i].allocated > 0) {
+      found = true;
+      int64_t unit = _mi_bin_size((uint8_t)i);
+      snprintf(buf, 64, "%s %3lu", fmt, (long)i);
+      mi_stat_print(&bins[i], buf, unit, out, arg);
+    }
+  }
+  if (found) {
+    _mi_fprintf(out, arg, "\n");
+    mi_print_header(out, arg);
+  }
+}
+#endif
+
+
+
+//------------------------------------------------------------
+// Use an output wrapper for line-buffered output
+// (which is nice when using loggers etc.)
+//------------------------------------------------------------
+typedef struct buffered_s {
+  mi_output_fun* out;   // original output function
+  void*          arg;   // and state
+  char*          buf;   // local buffer of at least size `count+1`
+  size_t         used;  // currently used chars `used <= count`
+  size_t         count; // total chars available for output
+} buffered_t;
+
+static void mi_buffered_flush(buffered_t* buf) {
+  buf->buf[buf->used] = 0;
+  _mi_fputs(buf->out, buf->arg, NULL, buf->buf);
+  buf->used = 0;
+}
+
+static void mi_cdecl mi_buffered_out(const char* msg, void* arg) {
+  buffered_t* buf = (buffered_t*)arg;
+  if (msg==NULL || buf==NULL) return;
+  for (const char* src = msg; *src != 0; src++) {
+    char c = *src;
+    if (buf->used >= buf->count) mi_buffered_flush(buf);
+    mi_assert_internal(buf->used < buf->count);
+    buf->buf[buf->used++] = c;
+    if (c == '\n') mi_buffered_flush(buf);
+  }
+}
+
+//------------------------------------------------------------
+// Print statistics
+//------------------------------------------------------------
+
+static void _mi_stats_print(mi_stats_t* stats, mi_output_fun* out0, void* arg0) mi_attr_noexcept {
+  // wrap the output function to be line buffered
+  char buf[256];
+  buffered_t buffer = { out0, arg0, NULL, 0, 255 };
+  buffer.buf = buf;
+  mi_output_fun* out = &mi_buffered_out;
+  void* arg = &buffer;
+
+  // and print using that
+  mi_print_header(out,arg);
+  #if MI_STAT>1
+  mi_stats_print_bins(stats->normal_bins, MI_BIN_HUGE, "normal",out,arg);
+  #endif
+  #if MI_STAT
+  mi_stat_print(&stats->normal, "normal", (stats->normal_count.count == 0 ? 1 : -(stats->normal.allocated / stats->normal_count.count)), out, arg);
+  mi_stat_print(&stats->large, "large", (stats->large_count.count == 0 ? 1 : -(stats->large.allocated / stats->large_count.count)), out, arg);
+  mi_stat_print(&stats->huge, "huge", (stats->huge_count.count == 0 ? 1 : -(stats->huge.allocated / stats->huge_count.count)), out, arg);
+  mi_stat_count_t total = { 0,0,0,0 };
+  mi_stat_add(&total, &stats->normal, 1);
+  mi_stat_add(&total, &stats->large, 1);
+  mi_stat_add(&total, &stats->huge, 1);
+  mi_stat_print(&total, "total", 1, out, arg);
+  #endif
+  #if MI_STAT>1
+  mi_stat_print(&stats->malloc, "malloc req", 1, out, arg);
+  _mi_fprintf(out, arg, "\n");
+  #endif
+  mi_stat_print_ex(&stats->reserved, "reserved", 1, out, arg, "");
+  mi_stat_print_ex(&stats->committed, "committed", 1, out, arg, "");
+  mi_stat_peak_print(&stats->reset, "reset", 1, out, arg );
+  mi_stat_peak_print(&stats->purged, "purged", 1, out, arg );
+  mi_stat_print(&stats->page_committed, "touched", 1, out, arg);
+  mi_stat_print(&stats->segments, "segments", -1, out, arg);
+  mi_stat_print(&stats->segments_abandoned, "-abandoned", -1, out, arg);
+  mi_stat_print(&stats->segments_cache, "-cached", -1, out, arg);
+  mi_stat_print(&stats->pages, "pages", -1, out, arg);
+  mi_stat_print(&stats->pages_abandoned, "-abandoned", -1, out, arg);
+  mi_stat_counter_print(&stats->pages_extended, "-extended", out, arg);
+  mi_stat_counter_print(&stats->page_no_retire, "-noretire", out, arg);
+  mi_stat_counter_print(&stats->mmap_calls, "mmaps", out, arg);
+  mi_stat_counter_print(&stats->commit_calls, "commits", out, arg);
+  mi_stat_counter_print(&stats->reset_calls, "resets", out, arg);
+  mi_stat_counter_print(&stats->purge_calls, "purges", out, arg);
+  mi_stat_print(&stats->threads, "threads", -1, out, arg);
+  mi_stat_counter_print_avg(&stats->searches, "searches", out, arg);
+  _mi_fprintf(out, arg, "%10s: %5zu\n", "numa nodes", _mi_os_numa_node_count());
+
+  size_t elapsed;
+  size_t user_time;
+  size_t sys_time;
+  size_t current_rss;
+  size_t peak_rss;
+  size_t current_commit;
+  size_t peak_commit;
+  size_t page_faults;
+  mi_process_info(&elapsed, &user_time, &sys_time, &current_rss, &peak_rss, &current_commit, &peak_commit, &page_faults);
+  _mi_fprintf(out, arg, "%10s: %5ld.%03ld s\n", "elapsed", elapsed/1000, elapsed%1000);
+  _mi_fprintf(out, arg, "%10s: user: %ld.%03ld s, system: %ld.%03ld s, faults: %lu, rss: ", "process",
+	      user_time/1000, user_time%1000, sys_time/1000, sys_time%1000, (unsigned long)page_faults );
+  mi_printf_amount((int64_t)peak_rss, 1, out, arg, "%s");
+  if (peak_commit > 0) {
+    _mi_fprintf(out, arg, ", commit: ");
+    mi_printf_amount((int64_t)peak_commit, 1, out, arg, "%s");
+  }
+  _mi_fprintf(out, arg, "\n");
+}
+
+static mi_msecs_t mi_process_start; // = 0
+
+static mi_stats_t* mi_stats_get_default(void) {
+  mi_heap_t* heap = mi_heap_get_default();
+  return &heap->tld->stats;
+}
+
+static void mi_stats_merge_from(mi_stats_t* stats) {
+  if (stats != &_mi_stats_main) {
+    mi_stats_add(&_mi_stats_main, stats);
+    memset(stats, 0, sizeof(mi_stats_t));
+  }
+}
+
+void mi_stats_reset(void) mi_attr_noexcept {
+  mi_stats_t* stats = mi_stats_get_default();
+  if (stats != &_mi_stats_main) { memset(stats, 0, sizeof(mi_stats_t)); }
+  memset(&_mi_stats_main, 0, sizeof(mi_stats_t));
+  if (mi_process_start == 0) { mi_process_start = _mi_clock_start(); };
+}
+
+void mi_stats_merge(void) mi_attr_noexcept {
+  mi_stats_merge_from( mi_stats_get_default() );
+}
+
+void _mi_stats_done(mi_stats_t* stats) {  // called from `mi_thread_done`
+  mi_stats_merge_from(stats);
+}
+
+void mi_stats_print_out(mi_output_fun* out, void* arg) mi_attr_noexcept {
+  mi_stats_merge_from(mi_stats_get_default());
+  _mi_stats_print(&_mi_stats_main, out, arg);
+}
+
+void mi_stats_print(void* out) mi_attr_noexcept {
+  // for compatibility there is an `out` parameter (which can be `stdout` or `stderr`)
+  mi_stats_print_out((mi_output_fun*)out, NULL);
+}
+
+void mi_thread_stats_print_out(mi_output_fun* out, void* arg) mi_attr_noexcept {
+  _mi_stats_print(mi_stats_get_default(), out, arg);
+}
+
+
+// ----------------------------------------------------------------
+// Basic timer for convenience; use milli-seconds to avoid doubles
+// ----------------------------------------------------------------
+
+static mi_msecs_t mi_clock_diff;
+
+mi_msecs_t _mi_clock_now(void) {
+  return _mi_prim_clock_now();
+}
+
+mi_msecs_t _mi_clock_start(void) {
+  if (mi_clock_diff == 0.0) {
+    mi_msecs_t t0 = _mi_clock_now();
+    mi_clock_diff = _mi_clock_now() - t0;
+  }
+  return _mi_clock_now();
+}
+
+mi_msecs_t _mi_clock_end(mi_msecs_t start) {
+  mi_msecs_t end = _mi_clock_now();
+  return (end - start - mi_clock_diff);
+}
+
+
+// --------------------------------------------------------
+// Basic process statistics
+// --------------------------------------------------------
+
+mi_decl_export void mi_process_info(size_t* elapsed_msecs, size_t* user_msecs, size_t* system_msecs, size_t* current_rss, size_t* peak_rss, size_t* current_commit, size_t* peak_commit, size_t* page_faults) mi_attr_noexcept
+{
+  mi_process_info_t pinfo;
+  _mi_memzero_var(pinfo);
+  pinfo.elapsed        = _mi_clock_end(mi_process_start);
+  pinfo.current_commit = (size_t)(mi_atomic_loadi64_relaxed((_Atomic(int64_t)*)&_mi_stats_main.committed.current));
+  pinfo.peak_commit    = (size_t)(mi_atomic_loadi64_relaxed((_Atomic(int64_t)*)&_mi_stats_main.committed.peak));
+  pinfo.current_rss    = pinfo.current_commit;
+  pinfo.peak_rss       = pinfo.peak_commit;
+  pinfo.utime          = 0;
+  pinfo.stime          = 0;
+  pinfo.page_faults    = 0;
+
+  _mi_prim_process_info(&pinfo);
+
+  if (elapsed_msecs!=NULL)  *elapsed_msecs  = (pinfo.elapsed < 0 ? 0 : (pinfo.elapsed < (mi_msecs_t)PTRDIFF_MAX ? (size_t)pinfo.elapsed : PTRDIFF_MAX));
+  if (user_msecs!=NULL)     *user_msecs     = (pinfo.utime < 0 ? 0 : (pinfo.utime < (mi_msecs_t)PTRDIFF_MAX ? (size_t)pinfo.utime : PTRDIFF_MAX));
+  if (system_msecs!=NULL)   *system_msecs   = (pinfo.stime < 0 ? 0 : (pinfo.stime < (mi_msecs_t)PTRDIFF_MAX ? (size_t)pinfo.stime : PTRDIFF_MAX));
+  if (current_rss!=NULL)    *current_rss    = pinfo.current_rss;
+  if (peak_rss!=NULL)       *peak_rss       = pinfo.peak_rss;
+  if (current_commit!=NULL) *current_commit = pinfo.current_commit;
+  if (peak_commit!=NULL)    *peak_commit    = pinfo.peak_commit;
+  if (page_faults!=NULL)    *page_faults    = pinfo.page_faults;
+}
diff --git a/compat/mingw.c b/compat/mingw.c
index 1d5b211b548dab..80ece6083c1f68 100644
--- a/compat/mingw.c
+++ b/compat/mingw.c
@@ -21,6 +21,8 @@
 #include "gettext.h"
 #define SECURITY_WIN32
 #include <sspi.h>
+#include "../write-or-die.h"
+#include "../repository.h"
 
 #define HCAST(type, handle) ((type)(intptr_t)handle)
 
@@ -614,6 +616,7 @@ static int is_local_named_pipe_path(const char *filename)
 
 int mingw_open (const char *filename, int oflags, ...)
 {
+	static int append_atomically = -1;
 	typedef int (*open_fn_t)(wchar_t const *wfilename, int oflags, ...);
 	va_list args;
 	unsigned mode;
@@ -630,7 +633,16 @@ int mingw_open (const char *filename, int oflags, ...)
 		return -1;
 	}
 
-	if ((oflags & O_APPEND) && !is_local_named_pipe_path(filename))
+	/*
+	 * Only set append_atomically to default value(1) when repo is initialized
+	 * and fail to get config value
+	 */
+	if (append_atomically < 0 && the_repository && the_repository->commondir &&
+		git_config_get_bool("windows.appendatomically", &append_atomically))
+		append_atomically = 1;
+
+	if (append_atomically && (oflags & O_APPEND) &&
+		!is_local_named_pipe_path(filename))
 		open_fn = mingw_open_append;
 	else if (!(oflags & ~(O_ACCMODE | O_NOINHERIT)))
 		open_fn = mingw_open_existing;
@@ -776,14 +788,33 @@ ssize_t mingw_write(int fd, const void *buf, size_t len)
 {
 	ssize_t result = write(fd, buf, len);
 
-	if (result < 0 && (errno == EINVAL || errno == ENOSPC) && buf) {
+	if (result < 0 && (errno == EINVAL || errno == EBADF || errno == ENOSPC) && buf) {
 		int orig = errno;
 
 		/* check if fd is a pipe */
 		HANDLE h = (HANDLE) _get_osfhandle(fd);
-		if (GetFileType(h) != FILE_TYPE_PIPE)
+		if (GetFileType(h) != FILE_TYPE_PIPE) {
+			if (orig == EINVAL) {
+				wchar_t path[MAX_PATH];
+				DWORD ret = GetFinalPathNameByHandleW(h, path,
+								ARRAY_SIZE(path), 0);
+				UINT drive_type = ret > 0 && ret < ARRAY_SIZE(path) ?
+					GetDriveTypeW(path) : DRIVE_UNKNOWN;
+
+				/*
+				 * The default atomic append causes such an error on
+				 * network file systems, in such a case, it should be
+				 * turned off via config.
+				 *
+				 * `drive_type` of UNC path: DRIVE_NO_ROOT_DIR
+				 */
+				if (DRIVE_NO_ROOT_DIR == drive_type || DRIVE_REMOTE == drive_type)
+					warning("invalid write operation detected; you may try:\n"
+						"\n\tgit config windows.appendAtomically false");
+			}
+
 			errno = orig;
-		else if (orig == EINVAL)
+		} else if (orig == EINVAL || errno == EBADF)
 			errno = EPIPE;
 		else {
 			DWORD buf_size;
@@ -1141,11 +1172,19 @@ unsigned int sleep (unsigned int seconds)
 char *mingw_mktemp(char *template)
 {
 	wchar_t wtemplate[MAX_PATH];
+	int offset = 0;
+
 	if (xutftowcs_path(wtemplate, template) < 0)
 		return NULL;
+
+	if (is_dir_sep(template[0]) && !is_dir_sep(template[1]) &&
+	    iswalpha(wtemplate[0]) && wtemplate[1] == L':') {
+		/* We have an absolute path missing the drive prefix */
+		offset = 2;
+	}
 	if (!_wmktemp(wtemplate))
 		return NULL;
-	if (xwcstoutf(template, wtemplate, strlen(template) + 1) < 0)
+	if (xwcstoutf(template, wtemplate + offset, strlen(template) + 1) < 0)
 		return NULL;
 	return template;
 }
@@ -1207,37 +1246,114 @@ struct tm *localtime_r(const time_t *timep, struct tm *result)
 }
 #endif
 
+char *mingw_strbuf_realpath(struct strbuf *resolved, const char *path)
+{
+	wchar_t wpath[MAX_PATH];
+	HANDLE h;
+	DWORD ret;
+	int len;
+	const char *last_component = NULL;
+	char *append = NULL;
+
+	if (xutftowcs_path(wpath, path) < 0)
+		return NULL;
+
+	h = CreateFileW(wpath, 0,
+			FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, NULL,
+			OPEN_EXISTING, FILE_FLAG_BACKUP_SEMANTICS, NULL);
+
+	/*
+	 * strbuf_realpath() allows the last path component to not exist. If
+	 * that is the case, now it's time to try without last component.
+	 */
+	if (h == INVALID_HANDLE_VALUE &&
+	    GetLastError() == ERROR_FILE_NOT_FOUND) {
+		/* cut last component off of `wpath` */
+		wchar_t *p = wpath + wcslen(wpath);
+
+		while (p != wpath)
+			if (*(--p) == L'/' || *p == L'\\')
+				break; /* found start of last component */
+
+		if (p != wpath && (last_component = find_last_dir_sep(path))) {
+			append = xstrdup(last_component + 1); /* skip directory separator */
+			/*
+			 * Do not strip the trailing slash at the drive root, otherwise
+			 * the path would be e.g. `C:` (which resolves to the
+			 * _current_ directory on that drive).
+			 */
+			if (p[-1] == L':')
+				p[1] = L'\0';
+			else
+				*p = L'\0';
+			h = CreateFileW(wpath, 0, FILE_SHARE_READ |
+					FILE_SHARE_WRITE | FILE_SHARE_DELETE,
+					NULL, OPEN_EXISTING,
+					FILE_FLAG_BACKUP_SEMANTICS, NULL);
+		}
+	}
+
+	if (h == INVALID_HANDLE_VALUE) {
+realpath_failed:
+		FREE_AND_NULL(append);
+		return NULL;
+	}
+
+	ret = GetFinalPathNameByHandleW(h, wpath, ARRAY_SIZE(wpath), 0);
+	CloseHandle(h);
+	if (!ret || ret >= ARRAY_SIZE(wpath))
+		goto realpath_failed;
+
+	len = wcslen(wpath) * 3;
+	strbuf_grow(resolved, len);
+	len = xwcstoutf(resolved->buf, normalize_ntpath(wpath), len);
+	if (len < 0)
+		goto realpath_failed;
+	resolved->len = len;
+
+	if (append) {
+		/* Use forward-slash, like `normalize_ntpath()` */
+		strbuf_complete(resolved, '/');
+		strbuf_addstr(resolved, append);
+		FREE_AND_NULL(append);
+	}
+
+	return resolved->buf;
+
+}
+
 char *mingw_getcwd(char *pointer, int len)
 {
 	wchar_t cwd[MAX_PATH], wpointer[MAX_PATH];
 	DWORD ret = GetCurrentDirectoryW(ARRAY_SIZE(cwd), cwd);
+	HANDLE hnd;
 
 	if (!ret || ret >= ARRAY_SIZE(cwd)) {
 		errno = ret ? ENAMETOOLONG : err_win_to_posix(GetLastError());
 		return NULL;
 	}
-	ret = GetLongPathNameW(cwd, wpointer, ARRAY_SIZE(wpointer));
-	if (!ret && GetLastError() == ERROR_ACCESS_DENIED) {
-		HANDLE hnd = CreateFileW(cwd, 0,
-			FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, NULL,
-			OPEN_EXISTING, FILE_FLAG_BACKUP_SEMANTICS, NULL);
-		if (hnd == INVALID_HANDLE_VALUE)
-			return NULL;
+	hnd = CreateFileW(cwd, 0,
+			  FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, NULL,
+			  OPEN_EXISTING, FILE_FLAG_BACKUP_SEMANTICS, NULL);
+	if (hnd != INVALID_HANDLE_VALUE) {
 		ret = GetFinalPathNameByHandleW(hnd, wpointer, ARRAY_SIZE(wpointer), 0);
 		CloseHandle(hnd);
-		if (!ret || ret >= ARRAY_SIZE(wpointer))
-			return NULL;
+		if (!ret || ret >= ARRAY_SIZE(wpointer)) {
+			ret = GetLongPathNameW(cwd, wpointer, ARRAY_SIZE(wpointer));
+			if (!ret || ret >= ARRAY_SIZE(wpointer)) {
+				errno = ret ? ENAMETOOLONG : err_win_to_posix(GetLastError());
+				return NULL;
+			}
+		}
 		if (xwcstoutf(pointer, normalize_ntpath(wpointer), len) < 0)
 			return NULL;
 		return pointer;
 	}
-	if (!ret || ret >= ARRAY_SIZE(wpointer))
-		return NULL;
-	if (GetFileAttributesW(wpointer) == INVALID_FILE_ATTRIBUTES) {
+	if (GetFileAttributesW(cwd) == INVALID_FILE_ATTRIBUTES) {
 		errno = ENOENT;
 		return NULL;
 	}
-	if (xwcstoutf(pointer, wpointer, len) < 0)
+	if (xwcstoutf(pointer, cwd, len) < 0)
 		return NULL;
 	convert_slashes(pointer);
 	return pointer;
@@ -1339,7 +1455,7 @@ static const char *quote_arg_msys2(const char *arg)
 
 static const char *parse_interpreter(const char *cmd)
 {
-	static char buf[100];
+	static char buf[MAX_PATH];
 	char *p, *opt;
 	ssize_t n; /* read() can return negative values */
 	int fd;
@@ -2115,18 +2231,150 @@ static void ensure_socket_initialization(void)
 	initialized = 1;
 }
 
+static int winsock_error_to_errno(DWORD err)
+{
+	switch (err) {
+	case WSAEINTR: return EINTR;
+	case WSAEBADF: return EBADF;
+	case WSAEACCES: return EACCES;
+	case WSAEFAULT: return EFAULT;
+	case WSAEINVAL: return EINVAL;
+	case WSAEMFILE: return EMFILE;
+	case WSAEWOULDBLOCK: return EWOULDBLOCK;
+	case WSAEINPROGRESS: return EINPROGRESS;
+	case WSAEALREADY: return EALREADY;
+	case WSAENOTSOCK: return ENOTSOCK;
+	case WSAEDESTADDRREQ: return EDESTADDRREQ;
+	case WSAEMSGSIZE: return EMSGSIZE;
+	case WSAEPROTOTYPE: return EPROTOTYPE;
+	case WSAENOPROTOOPT: return ENOPROTOOPT;
+	case WSAEPROTONOSUPPORT: return EPROTONOSUPPORT;
+	case WSAEOPNOTSUPP: return EOPNOTSUPP;
+	case WSAEAFNOSUPPORT: return EAFNOSUPPORT;
+	case WSAEADDRINUSE: return EADDRINUSE;
+	case WSAEADDRNOTAVAIL: return EADDRNOTAVAIL;
+	case WSAENETDOWN: return ENETDOWN;
+	case WSAENETUNREACH: return ENETUNREACH;
+	case WSAENETRESET: return ENETRESET;
+	case WSAECONNABORTED: return ECONNABORTED;
+	case WSAECONNRESET: return ECONNRESET;
+	case WSAENOBUFS: return ENOBUFS;
+	case WSAEISCONN: return EISCONN;
+	case WSAENOTCONN: return ENOTCONN;
+	case WSAETIMEDOUT: return ETIMEDOUT;
+	case WSAECONNREFUSED: return ECONNREFUSED;
+	case WSAELOOP: return ELOOP;
+	case WSAENAMETOOLONG: return ENAMETOOLONG;
+	case WSAEHOSTUNREACH: return EHOSTUNREACH;
+	case WSAENOTEMPTY: return ENOTEMPTY;
+	/* No errno equivalent; default to EIO */
+	case WSAESOCKTNOSUPPORT:
+	case WSAEPFNOSUPPORT:
+	case WSAESHUTDOWN:
+	case WSAETOOMANYREFS:
+	case WSAEHOSTDOWN:
+	case WSAEPROCLIM:
+	case WSAEUSERS:
+	case WSAEDQUOT:
+	case WSAESTALE:
+	case WSAEREMOTE:
+	case WSASYSNOTREADY:
+	case WSAVERNOTSUPPORTED:
+	case WSANOTINITIALISED:
+	case WSAEDISCON:
+	case WSAENOMORE:
+	case WSAECANCELLED:
+	case WSAEINVALIDPROCTABLE:
+	case WSAEINVALIDPROVIDER:
+	case WSAEPROVIDERFAILEDINIT:
+	case WSASYSCALLFAILURE:
+	case WSASERVICE_NOT_FOUND:
+	case WSATYPE_NOT_FOUND:
+	case WSA_E_NO_MORE:
+	case WSA_E_CANCELLED:
+	case WSAEREFUSED:
+	case WSAHOST_NOT_FOUND:
+	case WSATRY_AGAIN:
+	case WSANO_RECOVERY:
+	case WSANO_DATA:
+	case WSA_QOS_RECEIVERS:
+	case WSA_QOS_SENDERS:
+	case WSA_QOS_NO_SENDERS:
+	case WSA_QOS_NO_RECEIVERS:
+	case WSA_QOS_REQUEST_CONFIRMED:
+	case WSA_QOS_ADMISSION_FAILURE:
+	case WSA_QOS_POLICY_FAILURE:
+	case WSA_QOS_BAD_STYLE:
+	case WSA_QOS_BAD_OBJECT:
+	case WSA_QOS_TRAFFIC_CTRL_ERROR:
+	case WSA_QOS_GENERIC_ERROR:
+	case WSA_QOS_ESERVICETYPE:
+	case WSA_QOS_EFLOWSPEC:
+	case WSA_QOS_EPROVSPECBUF:
+	case WSA_QOS_EFILTERSTYLE:
+	case WSA_QOS_EFILTERTYPE:
+	case WSA_QOS_EFILTERCOUNT:
+	case WSA_QOS_EOBJLENGTH:
+	case WSA_QOS_EFLOWCOUNT:
+#ifndef _MSC_VER
+	case WSA_QOS_EUNKNOWNPSOBJ:
+#endif
+	case WSA_QOS_EPOLICYOBJ:
+	case WSA_QOS_EFLOWDESC:
+	case WSA_QOS_EPSFLOWSPEC:
+	case WSA_QOS_EPSFILTERSPEC:
+	case WSA_QOS_ESDMODEOBJ:
+	case WSA_QOS_ESHAPERATEOBJ:
+	case WSA_QOS_RESERVED_PETYPE:
+	default: return EIO;
+	}
+}
+
+/*
+ * On Windows, `errno` is a global macro to a function call.
+ * This makes it difficult to debug and single-step our mappings.
+ */
+static inline void set_wsa_errno(void)
+{
+	DWORD wsa = WSAGetLastError();
+	int e = winsock_error_to_errno(wsa);
+	errno = e;
+
+#ifdef DEBUG_WSA_ERRNO
+	fprintf(stderr, "winsock error: %d -> %d\n", wsa, e);
+	fflush(stderr);
+#endif
+}
+
+static inline int winsock_return(int ret)
+{
+	if (ret < 0)
+		set_wsa_errno();
+
+	return ret;
+}
+
+#define WINSOCK_RETURN(x) do { return winsock_return(x); } while (0)
+
 #undef gethostname
 int mingw_gethostname(char *name, int namelen)
 {
-    ensure_socket_initialization();
-    return gethostname(name, namelen);
+	ensure_socket_initialization();
+	WINSOCK_RETURN(gethostname(name, namelen));
 }
 
 #undef gethostbyname
 struct hostent *mingw_gethostbyname(const char *host)
 {
+	struct hostent *ret;
+
 	ensure_socket_initialization();
-	return gethostbyname(host);
+
+	ret = gethostbyname(host);
+	if (!ret)
+		set_wsa_errno();
+
+	return ret;
 }
 
 #undef getaddrinfo
@@ -2134,7 +2382,7 @@ int mingw_getaddrinfo(const char *node, const char *service,
 		      const struct addrinfo *hints, struct addrinfo **res)
 {
 	ensure_socket_initialization();
-	return getaddrinfo(node, service, hints, res);
+	WINSOCK_RETURN(getaddrinfo(node, service, hints, res));
 }
 
 int mingw_socket(int domain, int type, int protocol)
@@ -2154,7 +2402,7 @@ int mingw_socket(int domain, int type, int protocol)
 		 * in errno so that _if_ someone looks up the code somewhere,
 		 * then it is at least the number that are usually listed.
 		 */
-		errno = WSAGetLastError();
+		set_wsa_errno();
 		return -1;
 	}
 	/* convert into a file descriptor */
@@ -2170,35 +2418,35 @@ int mingw_socket(int domain, int type, int protocol)
 int mingw_connect(int sockfd, struct sockaddr *sa, size_t sz)
 {
 	SOCKET s = (SOCKET)_get_osfhandle(sockfd);
-	return connect(s, sa, sz);
+	WINSOCK_RETURN(connect(s, sa, sz));
 }
 
 #undef bind
 int mingw_bind(int sockfd, struct sockaddr *sa, size_t sz)
 {
 	SOCKET s = (SOCKET)_get_osfhandle(sockfd);
-	return bind(s, sa, sz);
+	WINSOCK_RETURN(bind(s, sa, sz));
 }
 
 #undef setsockopt
 int mingw_setsockopt(int sockfd, int lvl, int optname, void *optval, int optlen)
 {
 	SOCKET s = (SOCKET)_get_osfhandle(sockfd);
-	return setsockopt(s, lvl, optname, (const char*)optval, optlen);
+	WINSOCK_RETURN(setsockopt(s, lvl, optname, (const char*)optval, optlen));
 }
 
 #undef shutdown
 int mingw_shutdown(int sockfd, int how)
 {
 	SOCKET s = (SOCKET)_get_osfhandle(sockfd);
-	return shutdown(s, how);
+	WINSOCK_RETURN(shutdown(s, how));
 }
 
 #undef listen
 int mingw_listen(int sockfd, int backlog)
 {
 	SOCKET s = (SOCKET)_get_osfhandle(sockfd);
-	return listen(s, backlog);
+	WINSOCK_RETURN(listen(s, backlog));
 }
 
 #undef accept
@@ -2209,6 +2457,11 @@ int mingw_accept(int sockfd1, struct sockaddr *sa, socklen_t *sz)
 	SOCKET s1 = (SOCKET)_get_osfhandle(sockfd1);
 	SOCKET s2 = accept(s1, sa, sz);
 
+	if (s2 == INVALID_SOCKET) {
+		set_wsa_errno();
+		return -1;
+	}
+
 	/* convert into a file descriptor */
 	if ((sockfd2 = _open_osfhandle(s2, O_RDWR|O_BINARY)) < 0) {
 		int err = errno;
@@ -2686,6 +2939,28 @@ pid_t waitpid(pid_t pid, int *status, int options)
 	return -1;
 }
 
+int mingw_is_mount_point(struct strbuf *path)
+{
+	WIN32_FIND_DATAW findbuf = { 0 };
+	HANDLE handle;
+	wchar_t wfilename[MAX_PATH];
+	int wlen = xutftowcs_path(wfilename, path->buf);
+	if (wlen < 0)
+		die(_("could not get long path for '%s'"), path->buf);
+
+	/* remove trailing slash, if any */
+	if (wlen > 0 && wfilename[wlen - 1] == L'/')
+		wfilename[--wlen] = L'\0';
+
+	handle = FindFirstFileW(wfilename, &findbuf);
+	if (handle == INVALID_HANDLE_VALUE)
+		return 0;
+	FindClose(handle);
+
+	return (findbuf.dwFileAttributes & FILE_ATTRIBUTE_REPARSE_POINT) &&
+		(findbuf.dwReserved0 == IO_REPARSE_TAG_MOUNT_POINT);
+}
+
 int xutftowcsn(wchar_t *wcs, const char *utfs, size_t wcslen, int utflen)
 {
 	int upos = 0, wpos = 0;
@@ -2771,6 +3046,59 @@ int xwcstoutf(char *utf, const wchar_t *wcs, size_t utflen)
 	return -1;
 }
 
+#ifdef ENSURE_MSYSTEM_IS_SET
+static size_t append_system_bin_dirs(char *path, size_t size)
+{
+#if !defined(RUNTIME_PREFIX) || !defined(HAVE_WPGMPTR)
+	return 0;
+#else
+	char prefix[32768];
+	const char *slash;
+	size_t len = xwcstoutf(prefix, _wpgmptr, sizeof(prefix)), off = 0;
+
+	if (len == 0 || len >= sizeof(prefix) ||
+	    !(slash = find_last_dir_sep(prefix)))
+		return 0;
+	/* strip trailing `git.exe` */
+	len = slash - prefix;
+
+	/* strip trailing `cmd` or `mingw64\bin` or `mingw32\bin` or `bin` or `libexec\git-core` */
+	if (strip_suffix_mem(prefix, &len, "\\mingw64\\libexec\\git-core") ||
+	    strip_suffix_mem(prefix, &len, "\\mingw64\\bin"))
+		off += xsnprintf(path + off, size - off,
+				 "%.*s\\mingw64\\bin;", (int)len, prefix);
+	else if (strip_suffix_mem(prefix, &len, "\\mingw32\\libexec\\git-core") ||
+		 strip_suffix_mem(prefix, &len, "\\mingw32\\bin"))
+		off += xsnprintf(path + off, size - off,
+				 "%.*s\\mingw32\\bin;", (int)len, prefix);
+	else if (strip_suffix_mem(prefix, &len, "\\cmd") ||
+		 strip_suffix_mem(prefix, &len, "\\bin") ||
+		 strip_suffix_mem(prefix, &len, "\\libexec\\git-core"))
+		off += xsnprintf(path + off, size - off,
+				 "%.*s\\mingw%d\\bin;", (int)len, prefix,
+				 (int)(sizeof(void *) * 8));
+	else
+		return 0;
+
+	off += xsnprintf(path + off, size - off,
+			 "%.*s\\usr\\bin;", (int)len, prefix);
+	return off;
+#endif
+}
+#endif
+
+static int is_system32_path(const char *path)
+{
+	WCHAR system32[MAX_PATH], wpath[MAX_PATH];
+
+	if (xutftowcs_path(wpath, path) < 0 ||
+	    !GetSystemDirectoryW(system32, ARRAY_SIZE(system32)) ||
+	    _wcsicmp(system32, wpath))
+		return 0;
+
+	return 1;
+}
+
 static void setup_windows_environment(void)
 {
 	char *tmp = getenv("TMPDIR");
@@ -2795,9 +3123,20 @@ static void setup_windows_environment(void)
 		convert_slashes(tmp);
 	}
 
-	/* simulate TERM to enable auto-color (see color.c) */
-	if (!getenv("TERM"))
-		setenv("TERM", "cygwin", 1);
+
+	/*
+	 * Make sure TERM is set up correctly to enable auto-color
+	 * (see color.c .) Use "cygwin" for older OS releases which
+	 * works correctly with MSYS2 utilities on older consoles.
+	 */
+	if (!getenv("TERM")) {
+		if ((GetVersion() >> 16) < 15063)
+			setenv("TERM", "cygwin", 0);
+		else {
+			setenv("TERM", "xterm-256color", 0);
+			setenv("COLORTERM", "truecolor", 0);
+		}
+	}
 
 	/* calculate HOME if not set */
 	if (!getenv("HOME")) {
@@ -2811,7 +3150,8 @@ static void setup_windows_environment(void)
 			strbuf_addstr(&buf, tmp);
 			if ((tmp = getenv("HOMEPATH"))) {
 				strbuf_addstr(&buf, tmp);
-				if (is_directory(buf.buf))
+				if (!is_system32_path(buf.buf) &&
+				    is_directory(buf.buf))
 					setenv("HOME", buf.buf, 1);
 				else
 					tmp = NULL; /* use $USERPROFILE */
@@ -2822,6 +3162,37 @@ static void setup_windows_environment(void)
 		if (!tmp && (tmp = getenv("USERPROFILE")))
 			setenv("HOME", tmp, 1);
 	}
+
+	if (!getenv("PLINK_PROTOCOL"))
+		setenv("PLINK_PROTOCOL", "ssh", 0);
+
+#ifdef ENSURE_MSYSTEM_IS_SET
+	if (!(tmp = getenv("MSYSTEM")) || !tmp[0]) {
+		const char *home = getenv("HOME"), *path = getenv("PATH");
+		char buf[32768];
+		size_t off = 0;
+
+		xsnprintf(buf, sizeof(buf),
+			  "MINGW%d", (int)(sizeof(void *) * 8));
+		setenv("MSYSTEM", buf, 1);
+
+		if (home)
+			off += xsnprintf(buf + off, sizeof(buf) - off,
+					 "%s\\bin;", home);
+		off += append_system_bin_dirs(buf + off, sizeof(buf) - off);
+		if (path)
+			off += xsnprintf(buf + off, sizeof(buf) - off,
+					 "%s", path);
+		else if (off > 0)
+			buf[off - 1] = '\0';
+		else
+			buf[0] = '\0';
+		setenv("PATH", buf, 1);
+	}
+#endif
+
+	if (!getenv("LC_ALL") && !getenv("LC_CTYPE") && !getenv("LANG"))
+		setenv("LC_CTYPE", "C.UTF-8", 1);
 }
 
 static PSID get_current_user_sid(void)
@@ -2924,9 +3295,7 @@ int is_path_owned_by_current_sid(const char *path, struct strbuf *report)
 				    DACL_SECURITY_INFORMATION,
 				    &sid, NULL, NULL, NULL, &descriptor);
 
-	if (err != ERROR_SUCCESS)
-		error(_("failed to get owner for '%s' (%ld)"), path, err);
-	else if (sid && IsValidSid(sid)) {
+	if (err == ERROR_SUCCESS && sid && IsValidSid(sid)) {
 		/* Now, verify that the SID matches the current user's */
 		static PSID current_user_sid;
 		BOOL is_member;
@@ -3253,6 +3622,7 @@ int wmain(int argc, const wchar_t **wargv)
 #endif
 
 	maybe_redirect_std_handles();
+	fsync_object_files = 1;
 
 	/* determine size of argv and environ conversion buffer */
 	maxlen = wcslen(wargv[0]);
diff --git a/compat/mingw.h b/compat/mingw.h
index ebfb8ba423b754..4967a649894142 100644
--- a/compat/mingw.h
+++ b/compat/mingw.h
@@ -454,9 +454,16 @@ static inline void convert_slashes(char *path)
 		if (*path == '\\')
 			*path = '/';
 }
+struct strbuf;
+int mingw_is_mount_point(struct strbuf *path);
+#define is_mount_point mingw_is_mount_point
+#define CAN_UNLINK_MOUNT_POINTS 1
 #define PATH_SEP ';'
 char *mingw_query_user_email(void);
 #define query_user_email mingw_query_user_email
+struct strbuf;
+char *mingw_strbuf_realpath(struct strbuf *resolved, const char *path);
+#define platform_strbuf_realpath mingw_strbuf_realpath
 #if !defined(__MINGW64_VERSION_MAJOR) && (!defined(_MSC_VER) || _MSC_VER < 1800)
 #define PRIuMAX "I64u"
 #define PRId64 "I64d"
diff --git a/compat/vcbuild/README b/compat/vcbuild/README
index 29ec1d0f104b80..9ac9760397f479 100644
--- a/compat/vcbuild/README
+++ b/compat/vcbuild/README
@@ -6,7 +6,11 @@ The Steps to Build Git with VS2015 or VS2017 from the command line.
    Prompt or from an SDK bash window:
 
    $ cd <repo_root>
-   $ ./compat/vcbuild/vcpkg_install.bat
+   $ ./compat/vcbuild/vcpkg_install.bat x64-windows
+
+   or
+
+   $ ./compat/vcbuild/vcpkg_install.bat arm64-windows
 
    The vcpkg tools and all of the third-party sources will be installed
    in this folder:
@@ -37,27 +41,17 @@ The Steps to Build Git with VS2015 or VS2017 from the command line.
 
 ================================================================
 
-Alternatively, run `make vcxproj` and then load the generated `git.sln` in
-Visual Studio. The initial build will install the vcpkg system and build the
+Alternatively, just open Git's top-level directory in Visual Studio, via
+`File>Open>Folder...`. This will use CMake internally to generate the
+project definitions. It will also install the vcpkg system and build the
 dependencies automatically. This will take a while.
 
-Instead of generating the `git.sln` file yourself (which requires a full Git
-for Windows SDK), you may want to consider fetching the `vs/master` branch of
-https://github.com/git-for-windows/git instead (which is updated automatically
-via CI running `make vcxproj`). The `vs/master` branch does not require a Git
-for Windows to build, but you can run the test scripts in a regular Git Bash.
-
-Note that `make vcxproj` will automatically add and commit the generated `.sln`
-and `.vcxproj` files to the repo. This is necessary to allow building a
-fully-testable Git in Visual Studio, where a regular Git Bash can be used to
-run the test scripts (as opposed to a full Git for Windows SDK): a number of
-build targets, such as Git commands implemented as Unix shell scripts (where
-`@@SHELL_PATH@@` and other placeholders are interpolated) require a full-blown
-Git for Windows SDK (which is about 10x the size of a regular Git for Windows
-installation).
-
-If your plan is to open a Pull Request with Git for Windows, it is a good idea
-to drop this commit before submitting.
+You can also generate the Visual Studio solution manually by downloading
+and running CMake explicitly rather than letting Visual Studio doing
+that implicitly.
+
+Another, deprecated option is to run `make vcxproj`. This option is
+superseded by the CMake-based build, and will be removed at some point.
 
 ================================================================
 The Steps of Build Git with VS2008
diff --git a/compat/vcbuild/find_vs_env.bat b/compat/vcbuild/find_vs_env.bat
index b35d264c0e6bed..379b16296e09c2 100644
--- a/compat/vcbuild/find_vs_env.bat
+++ b/compat/vcbuild/find_vs_env.bat
@@ -99,6 +99,7 @@ REM ================================================================
 
    SET sdk_dir=%WindowsSdkDir%
    SET sdk_ver=%WindowsSDKVersion%
+   SET sdk_ver_bin_dir=%WindowsSdkVerBinPath%%tgt%
    SET si=%sdk_dir%Include\%sdk_ver%
    SET sdk_includes=-I"%si%ucrt" -I"%si%um" -I"%si%shared"
    SET sl=%sdk_dir%lib\%sdk_ver%
@@ -130,6 +131,7 @@ REM ================================================================
 
    SET sdk_dir=%WindowsSdkDir%
    SET sdk_ver=%WindowsSDKVersion%
+   SET sdk_ver_bin_dir=%WindowsSdkVerBinPath%bin\amd64
    SET si=%sdk_dir%Include\%sdk_ver%
    SET sdk_includes=-I"%si%ucrt" -I"%si%um" -I"%si%shared" -I"%si%winrt"
    SET sl=%sdk_dir%lib\%sdk_ver%
@@ -160,6 +162,11 @@ REM ================================================================
    echo msvc_includes=%msvc_includes%
    echo msvc_libs=%msvc_libs%
 
+   echo sdk_ver_bin_dir=%sdk_ver_bin_dir%
+   SET X1=%sdk_ver_bin_dir:C:=/C%
+   SET X2=%X1:\=/%
+   echo sdk_ver_bin_dir_msys=%X2%
+
    echo sdk_includes=%sdk_includes%
    echo sdk_libs=%sdk_libs%
 
diff --git a/compat/vcbuild/scripts/clink.pl b/compat/vcbuild/scripts/clink.pl
index 3bd824154be381..677d44e46f98d6 100755
--- a/compat/vcbuild/scripts/clink.pl
+++ b/compat/vcbuild/scripts/clink.pl
@@ -15,6 +15,7 @@
 my @lflags = ();
 my $is_linking = 0;
 my $is_debug = 0;
+my $is_gui = 0;
 while (@ARGV) {
 	my $arg = shift @ARGV;
 	if ("$arg" eq "-DDEBUG") {
@@ -56,7 +57,8 @@
 		# need to use that instead?
 		foreach my $flag (@lflags) {
 			if ($flag =~ /^-LIBPATH:(.*)/) {
-				foreach my $l ("libcurl_imp.lib", "libcurl.lib") {
+				my $libcurl = $is_debug ? "libcurl-d.lib" : "libcurl.lib";
+				foreach my $l ("libcurl_imp.lib", $libcurl) {
 					if (-f "$1/$l") {
 						$lib = $l;
 						last;
@@ -66,7 +68,11 @@
 		}
 		push(@args, $lib);
 	} elsif ("$arg" eq "-lexpat") {
+	    if ($is_debug) {
+		push(@args, "libexpatd.lib");
+	    } else {
 		push(@args, "libexpat.lib");
+	    }
 	} elsif ("$arg" =~ /^-L/ && "$arg" ne "-LTCG") {
 		$arg =~ s/^-L/-LIBPATH:/;
 		push(@lflags, $arg);
@@ -118,11 +124,23 @@
 		push(@cflags, "-wd4996");
 	} elsif ("$arg" =~ /^-W[a-z]/) {
 		# let's ignore those
+	} elsif ("$arg" eq "-fno-stack-protector") {
+		# eat this
+	} elsif ("$arg" eq "-mwindows") {
+		$is_gui = 1;
 	} else {
 		push(@args, $arg);
 	}
 }
 if ($is_linking) {
+	if ($is_gui) {
+		push(@args, "-ENTRY:wWinMainCRTStartup");
+		push(@args, "-SUBSYSTEM:WINDOWS");
+	} else {
+		push(@args, "-ENTRY:wmainCRTStartup");
+		push(@args, "-SUBSYSTEM:CONSOLE");
+	}
+
 	push(@args, @lflags);
 	unshift(@args, "link.exe");
 } else {
diff --git a/compat/vcbuild/scripts/rc.pl b/compat/vcbuild/scripts/rc.pl
new file mode 100644
index 00000000000000..7bca4cd81c6c63
--- /dev/null
+++ b/compat/vcbuild/scripts/rc.pl
@@ -0,0 +1,46 @@
+#!/usr/bin/perl -w
+######################################################################
+# Compile Resources on Windows
+#
+# This is a wrapper to facilitate the compilation of Git with MSVC
+# using GNU Make as the build system. So, instead of manipulating the
+# Makefile into something nasty, just to support non-space arguments
+# etc, we use this wrapper to fix the command line options
+#
+######################################################################
+use strict;
+my @args = ();
+my @input = ();
+
+while (@ARGV) {
+	my $arg = shift @ARGV;
+	if ("$arg" =~ /^-[dD]/) {
+		# GIT_VERSION gets passed with too many
+		# layers of dquote escaping.
+		$arg =~ s/\\"/"/g;
+
+		push(@args, $arg);
+
+	} elsif ("$arg" eq "-i") {
+		my $arg = shift @ARGV;
+		# TODO complain if NULL or is dashed ??
+		push(@input, $arg);
+
+	} elsif ("$arg" eq "-o") {
+		my $arg = shift @ARGV;
+		# TODO complain if NULL or is dashed ??
+		push(@args, "-fo$arg");
+
+	} else {
+		push(@args, $arg);
+	}
+}
+
+push(@args, "-nologo");
+push(@args, "-v");
+push(@args, @input);
+
+unshift(@args, "rc.exe");
+printf("**** @args\n");
+
+exit (system(@args) != 0);
diff --git a/compat/vcbuild/vcpkg_copy_dlls.bat b/compat/vcbuild/vcpkg_copy_dlls.bat
index 13661c14f8705c..8bea0cbf83b6cf 100644
--- a/compat/vcbuild/vcpkg_copy_dlls.bat
+++ b/compat/vcbuild/vcpkg_copy_dlls.bat
@@ -15,7 +15,12 @@ REM ================================================================
 	@FOR /F "delims=" %%D IN ("%~dp0") DO @SET cwd=%%~fD
 	cd %cwd%
 
-	SET arch=x64-windows
+	SET arch=%2
+	IF NOT DEFINED arch (
+		echo defaulting to 'x64-windows`. Invoke %0 with 'x86-windows', 'x64-windows', or 'arm64-windows'
+		set arch=x64-windows
+	)
+
 	SET inst=%cwd%vcpkg\installed\%arch%
 
 	IF [%1]==[release] (
diff --git a/compat/vcbuild/vcpkg_install.bat b/compat/vcbuild/vcpkg_install.bat
index ebd0bad242a8ca..575c65c20ba307 100644
--- a/compat/vcbuild/vcpkg_install.bat
+++ b/compat/vcbuild/vcpkg_install.bat
@@ -31,11 +31,24 @@ REM ================================================================
 
 	SETLOCAL EnableDelayedExpansion
 
+	SET arch=%1
+	IF NOT DEFINED arch (
+		echo defaulting to 'x64-windows`. Invoke %0 with 'x86-windows', 'x64-windows', or 'arm64-windows'
+		set arch=x64-windows
+	)
+
 	@FOR /F "delims=" %%D IN ("%~dp0") DO @SET cwd=%%~fD
 	cd %cwd%
 
 	dir vcpkg\vcpkg.exe >nul 2>nul && GOTO :install_libraries
 
+	git.exe version 2>nul
+	IF ERRORLEVEL 1 (
+	echo "***"
+	echo "Git not found. Please adjust your CMD path or Git install option."
+	echo "***"
+	EXIT /B 1 )
+
 	echo Fetching vcpkg in %cwd%vcpkg
 	git.exe clone https://github.com/Microsoft/vcpkg vcpkg
 	IF ERRORLEVEL 1 ( EXIT /B 1 )
@@ -48,9 +61,8 @@ REM ================================================================
 	echo Successfully installed %cwd%vcpkg\vcpkg.exe
 
 :install_libraries
-	SET arch=x64-windows
 
-	echo Installing third-party libraries...
+	echo Installing third-party libraries(%arch%)...
 	FOR %%i IN (zlib expat libiconv openssl libssh2 curl) DO (
 	    cd %cwd%vcpkg
 	    IF NOT EXIST "packages\%%i_%arch%" CALL :sub__install_one %%i
@@ -73,8 +85,47 @@ REM ================================================================
 :sub__install_one
 	echo     Installing package %1...
 
-	.\vcpkg.exe install %1:%arch%
+	call :%1_features
+
+	REM vcpkg may not be reliable on slow, intermittent or proxy
+	REM connections, see e.g.
+	REM https://social.msdn.microsoft.com/Forums/windowsdesktop/en-US/4a8f7be5-5e15-4213-a7bb-ddf424a954e6/winhttpsendrequest-ends-with-12002-errorhttptimeout-after-21-seconds-no-matter-what-timeout?forum=windowssdk
+	REM which explains the hidden 21 second timeout
+	REM (last post by Dave : Microsoft - Windows Networking team)
+
+	.\vcpkg.exe install %1%features%:%arch%
 	IF ERRORLEVEL 1 ( EXIT /B 1 )
 
 	echo     Finished %1
 	goto :EOF
+
+::
+:: features for each vcpkg to install
+:: there should be an entry here for each package to install
+:: 'set features=' means use the default otherwise
+:: 'set features=[comma-delimited-feature-set]' is the syntax
+::
+
+:zlib_features
+set features=
+goto :EOF
+
+:expat_features
+set features=
+goto :EOF
+
+:libiconv_features
+set features=
+goto :EOF
+
+:openssl_features
+set features=
+goto :EOF
+
+:libssh2_features
+set features=
+goto :EOF
+
+:curl_features
+set features=[core,openssl,schannel]
+goto :EOF
diff --git a/compat/win32/path-utils.c b/compat/win32/path-utils.c
index 966ef779b9ca9b..c4fea0301b5ecc 100644
--- a/compat/win32/path-utils.c
+++ b/compat/win32/path-utils.c
@@ -2,6 +2,9 @@
 
 #include "../../git-compat-util.h"
 #include "../../environment.h"
+#include "../../wrapper.h"
+#include "../../strbuf.h"
+#include "../../versioncmp.h"
 
 int win32_has_dos_drive_prefix(const char *path)
 {
@@ -89,3 +92,199 @@ int win32_fspathcmp(const char *a, const char *b)
 {
 	return win32_fspathncmp(a, b, (size_t)-1);
 }
+
+static int read_at(int fd, char *buffer, size_t offset, size_t size)
+{
+	if (lseek(fd, offset, SEEK_SET) < 0) {
+		fprintf(stderr, "could not seek to 0x%x\n", (unsigned int)offset);
+		return -1;
+	}
+
+	return read_in_full(fd, buffer, size);
+}
+
+static size_t le16(const char *buffer)
+{
+	unsigned char *u = (unsigned char *)buffer;
+	return u[0] | (u[1] << 8);
+}
+
+static size_t le32(const char *buffer)
+{
+	return le16(buffer) | (le16(buffer + 2) << 16);
+}
+
+/*
+ * Determine the Go version of a given executable, if it was built with Go.
+ *
+ * This recapitulates the logic from
+ * https://github.com/golang/go/blob/master/src/cmd/go/internal/version/version.go
+ * (without requiring the user to install `go.exe` to find out).
+ */
+static ssize_t get_go_version(const char *path, char *go_version, size_t go_version_size)
+{
+	int fd = open(path, O_RDONLY);
+	char buffer[1024];
+	off_t offset;
+	size_t num_sections, opt_header_size, i;
+	char *p = NULL, *q;
+	ssize_t res = -1;
+
+	if (fd < 0)
+		return -1;
+
+	if (read_in_full(fd, buffer, 2) < 0)
+		goto fail;
+
+	/*
+	 * Parse the PE file format, for more details, see
+	 * https://en.wikipedia.org/wiki/Portable_Executable#Layout and
+	 * https://learn.microsoft.com/en-us/windows/win32/debug/pe-format
+	 */
+	if (buffer[0] != 'M' || buffer[1] != 'Z')
+		goto fail;
+
+	if (read_at(fd, buffer, 0x3c, 4) < 0)
+		goto fail;
+
+	/* Read the `PE\0\0` signature and the COFF file header */
+	offset = le32(buffer);
+	if (read_at(fd, buffer, offset, 24) < 0)
+		goto fail;
+
+	if (buffer[0] != 'P' || buffer[1] != 'E' || buffer[2] != '\0' || buffer[3] != '\0')
+		goto fail;
+
+	num_sections = le16(buffer + 6);
+	opt_header_size = le16(buffer + 20);
+	offset += 24; /* skip file header */
+
+	/*
+	 * Validate magic number 0x10b or 0x20b, for full details see
+	 * https://learn.microsoft.com/en-us/windows/win32/debug/pe-format#optional-header-standard-fields-image-only
+	 */
+	if (read_at(fd, buffer, offset, 2) < 0 ||
+	    ((i = le16(buffer)) != 0x10b && i != 0x20b))
+		goto fail;
+
+	offset += opt_header_size;
+
+	for (i = 0; i < num_sections; i++) {
+		if (read_at(fd, buffer, offset + i * 40, 40) < 0)
+			goto fail;
+
+		/*
+		 * For full details about the section headers, see
+		 * https://learn.microsoft.com/en-us/windows/win32/debug/pe-format#section-table-section-headers
+		 */
+		if ((le32(buffer + 36) /* characteristics */ & ~0x600000) /* IMAGE_SCN_ALIGN_32BYTES */ ==
+		    (/* IMAGE_SCN_CNT_INITIALIZED_DATA */ 0x00000040 |
+		     /* IMAGE_SCN_MEM_READ */ 0x40000000 |
+		     /* IMAGE_SCN_MEM_WRITE */ 0x80000000)) {
+			size_t size = le32(buffer + 16); /* "SizeOfRawData " */
+			size_t pointer = le32(buffer + 20); /* "PointerToRawData " */
+
+			/*
+			 * Skip the section if either size or pointer is 0, see
+			 * https://github.com/golang/go/blob/go1.21.0/src/debug/buildinfo/buildinfo.go#L333
+			 * for full details.
+			 *
+			 * Merely seeing a non-zero size will not actually do,
+			 * though: he size must be at least `buildInfoSize`,
+			 * i.e. 32, and we expect a UVarint (at least another
+			 * byte) _and_ the bytes representing the string,
+			 * which we expect to start with the letters "go" and
+			 * continue with the Go version number.
+			 */
+			if (size < 32 + 1 + 2 + 1 || !pointer)
+				continue;
+
+			p = malloc(size);
+
+			if (!p || read_at(fd, p, pointer, size) < 0)
+				goto fail;
+
+			/*
+			 * Look for the build information embedded by Go, see
+			 * https://github.com/golang/go/blob/go1.21.0/src/debug/buildinfo/buildinfo.go#L165-L175
+			 * for full details.
+			 *
+			 * Note: Go contains code to enforce alignment along a
+			 * 16-byte boundary. In practice, no `.exe` has been
+			 * observed that required any adjustment, therefore
+			 * this here code skips that logic for simplicity.
+			 */
+			q = memmem(p, size - 18, "\xff Go buildinf:", 14);
+			if (!q)
+				goto fail;
+			/*
+			 * Decode the build blob. For full details, see
+			 * https://github.com/golang/go/blob/go1.21.0/src/debug/buildinfo/buildinfo.go#L177-L191
+			 *
+			 * Note: The `endianness` values observed in practice
+			 * were always 2, therefore the complex logic to handle
+			 * any other value is skipped for simplicty.
+			 */
+			if ((q[14] == 8 || q[14] == 4) && q[15] == 2) {
+				/*
+				 * Only handle a Go version string with fewer
+				 * than 128 characters, so the Go UVarint at
+				 * q[32] that indicates the string's length must
+				 * be only one byte (without the high bit set).
+				 */
+				if ((q[32] & 0x80) ||
+				    !q[32] ||
+				    (q + 33 + q[32] - p) > (ssize_t)size ||
+				    q[32] + 1 > (ssize_t)go_version_size)
+					goto fail;
+				res = q[32];
+				memcpy(go_version, q + 33, res);
+				go_version[res] = '\0';
+				break;
+			}
+		}
+	}
+
+fail:
+	free(p);
+	close(fd);
+	return res;
+}
+
+void win32_warn_about_git_lfs_on_windows7(int exit_code, const char *argv0)
+{
+	char buffer[128], *git_lfs = NULL;
+	const char *p;
+
+	/*
+	 * Git LFS v3.5.1 fails with an Access Violation on Windows 7; That
+	 * would usually show up as an exit code 0xc0000005. For some reason
+	 * (probably because at this point, we no longer have the _original_
+	 * HANDLE that was returned by `CreateProcess()`) we observe other
+	 * values like 0xb00 and 0x2 instead. Since the exact exit code
+	 * seems to be inconsistent, we check for a non-zero exit status.
+	 */
+	if (exit_code == 0)
+		return;
+	if (GetVersion() >> 16 > 7601)
+		return; /* Warn only on Windows 7 or older */
+	if (!istarts_with(argv0, "git-lfs ") &&
+	    strcasecmp(argv0, "git-lfs"))
+		return;
+	if (!(git_lfs = locate_in_PATH("git-lfs")))
+		return;
+	if (get_go_version(git_lfs, buffer, sizeof(buffer)) > 0 &&
+	    skip_prefix(buffer, "go", &p) &&
+	    versioncmp("1.21.0", p) <= 0)
+		warning("This program was built with Go v%s\n"
+			"i.e. without support for this Windows version:\n"
+			"\n\t%s\n"
+			"\n"
+			"To work around this, you can download and install a "
+			"working version from\n"
+			"\n"
+			"\thttps://github.com/git-lfs/git-lfs/releases/tag/"
+			"v3.4.1\n",
+			p, git_lfs);
+	free(git_lfs);
+}
diff --git a/compat/win32/path-utils.h b/compat/win32/path-utils.h
index a561c700e75713..a69483c332c1a7 100644
--- a/compat/win32/path-utils.h
+++ b/compat/win32/path-utils.h
@@ -34,4 +34,7 @@ int win32_fspathcmp(const char *a, const char *b);
 int win32_fspathncmp(const char *a, const char *b, size_t count);
 #define fspathncmp win32_fspathncmp
 
+void win32_warn_about_git_lfs_on_windows7(int exit_code, const char *argv0);
+#define warn_about_git_lfs_on_windows7 win32_warn_about_git_lfs_on_windows7
+
 #endif
diff --git a/compat/win32/pthread.c b/compat/win32/pthread.c
index 58980a529c3eb9..4cc6356f757d80 100644
--- a/compat/win32/pthread.c
+++ b/compat/win32/pthread.c
@@ -21,8 +21,8 @@ static unsigned __stdcall win32_start_routine(void *arg)
 	return 0;
 }
 
-int pthread_create(pthread_t *thread, const void *attr UNUSED,
-		   void *(*start_routine)(void *), void *arg)
+int win32_pthread_create(pthread_t *thread, const void *attr UNUSED,
+			 void *(*start_routine)(void *), void *arg)
 {
 	thread->arg = arg;
 	thread->start_routine = start_routine;
@@ -53,7 +53,7 @@ int win32_pthread_join(pthread_t *thread, void **value_ptr)
 	}
 }
 
-pthread_t pthread_self(void)
+pthread_t win32_pthread_self(void)
 {
 	pthread_t t = { NULL };
 	t.tid = GetCurrentThreadId();
diff --git a/compat/win32/pthread.h b/compat/win32/pthread.h
index e2b5c4f64c9b91..476e78b911b857 100644
--- a/compat/win32/pthread.h
+++ b/compat/win32/pthread.h
@@ -50,8 +50,9 @@ typedef struct {
 	DWORD tid;
 } pthread_t;
 
-int pthread_create(pthread_t *thread, const void *unused,
-		   void *(*start_routine)(void*), void *arg);
+int win32_pthread_create(pthread_t *thread, const void *unused,
+			 void *(*start_routine)(void*), void *arg);
+#define pthread_create win32_pthread_create
 
 /*
  * To avoid the need of copying a struct, we use small macro wrapper to pass
@@ -62,7 +63,8 @@ int pthread_create(pthread_t *thread, const void *unused,
 int win32_pthread_join(pthread_t *thread, void **value_ptr);
 
 #define pthread_equal(t1, t2) ((t1).tid == (t2).tid)
-pthread_t pthread_self(void);
+pthread_t win32_pthread_self(void);
+#define pthread_self win32_pthread_self
 
 static inline void NORETURN pthread_exit(void *ret)
 {
diff --git a/compat/winansi.c b/compat/winansi.c
index ac2ffb78691a7d..0e0baf36e7ec7d 100644
--- a/compat/winansi.c
+++ b/compat/winansi.c
@@ -575,6 +575,9 @@ static void detect_msys_tty(int fd)
 	if (!NT_SUCCESS(NtQueryObject(h, ObjectNameInformation,
 			buffer, sizeof(buffer) - 2, &result)))
 		return;
+	if (result < sizeof(*nameinfo) || !nameinfo->Name.Buffer ||
+		!nameinfo->Name.Length)
+		return;
 	name = nameinfo->Name.Buffer;
 	name[nameinfo->Name.Length / sizeof(*name)] = 0;
 
@@ -593,6 +596,49 @@ static void detect_msys_tty(int fd)
 
 #endif
 
+static HANDLE std_console_handle;
+static DWORD std_console_mode = ENABLE_VIRTUAL_TERMINAL_PROCESSING;
+static UINT std_console_code_page = CP_UTF8;
+
+static void reset_std_console(void)
+{
+	if (std_console_mode != ENABLE_VIRTUAL_TERMINAL_PROCESSING)
+		SetConsoleMode(std_console_handle, std_console_mode);
+	if (std_console_code_page != CP_UTF8)
+		SetConsoleOutputCP(std_console_code_page);
+}
+
+static int enable_virtual_processing(void)
+{
+	std_console_handle = GetStdHandle(STD_OUTPUT_HANDLE);
+	if (std_console_handle == INVALID_HANDLE_VALUE ||
+	    !GetConsoleMode(std_console_handle, &std_console_mode)) {
+		std_console_handle = GetStdHandle(STD_ERROR_HANDLE);
+		if (std_console_handle == INVALID_HANDLE_VALUE ||
+		    !GetConsoleMode(std_console_handle, &std_console_mode))
+		return 0;
+	}
+
+	std_console_code_page = GetConsoleOutputCP();
+	if (std_console_code_page != CP_UTF8)
+		SetConsoleOutputCP(CP_UTF8);
+	if (!std_console_code_page)
+		std_console_code_page = CP_UTF8;
+
+	atexit(reset_std_console);
+
+	if (std_console_mode & ENABLE_VIRTUAL_TERMINAL_PROCESSING)
+		return 1;
+
+	if (!SetConsoleMode(std_console_handle,
+			    std_console_mode |
+			    ENABLE_PROCESSED_OUTPUT |
+			    ENABLE_VIRTUAL_TERMINAL_PROCESSING))
+		return 0;
+
+	return 1;
+}
+
 /*
  * Wrapper for isatty().  Most calls in the main git code
  * call isatty(1 or 2) to see if the instance is interactive
@@ -631,6 +677,9 @@ void winansi_init(void)
 		return;
 	}
 
+	if (enable_virtual_processing())
+		return;
+
 	/* create a named pipe to communicate with the console thread */
 	if (swprintf(name, ARRAY_SIZE(name) - 1, L"\\\\.\\pipe\\winansi%lu",
 		     GetCurrentProcessId()) < 0)
diff --git a/config.c b/config.c
index 50f2d17b399446..95b7de0522d180 100644
--- a/config.c
+++ b/config.c
@@ -1392,8 +1392,8 @@ int git_config_color(char *dest, const char *var, const char *value)
 	return 0;
 }
 
-static int git_default_core_config(const char *var, const char *value,
-				   const struct config_context *ctx, void *cb)
+int git_default_core_config(const char *var, const char *value,
+			    const struct config_context *ctx, void *cb)
 {
 	/* This needs a better name */
 	if (!strcmp(var, "core.filemode")) {
diff --git a/config.h b/config.h
index 5c730c4f899160..e4199bbdc07685 100644
--- a/config.h
+++ b/config.h
@@ -167,6 +167,8 @@ typedef int (*config_fn_t)(const char *, const char *,
 
 int git_default_config(const char *, const char *,
 		       const struct config_context *, void *);
+int git_default_core_config(const char *var, const char *value,
+			    const struct config_context *ctx, void *cb);
 
 /**
  * Read a specific file in git-config format.
diff --git a/config.mak.dev b/config.mak.dev
index 0fd8cc4d355ebb..a5d6dd67dfd77e 100644
--- a/config.mak.dev
+++ b/config.mak.dev
@@ -22,8 +22,10 @@ endif
 
 ifneq ($(uname_S),FreeBSD)
 ifneq ($(or $(filter gcc6,$(COMPILER_FEATURES)),$(filter clang7,$(COMPILER_FEATURES))),)
+ifndef USE_MIMALLOC
 DEVELOPER_CFLAGS += -std=gnu99
 endif
+endif
 else
 # FreeBSD cannot limit to C99 because its system headers unconditionally
 # rely on C11 features.
diff --git a/config.mak.uname b/config.mak.uname
index b12d4e168ae119..9734af587d56da 100644
--- a/config.mak.uname
+++ b/config.mak.uname
@@ -440,7 +440,7 @@ ifeq ($(uname_S),Windows)
 	# link.exe next to, and required by, cl.exe, we have to prepend this
 	# onto the existing $PATH.
 	#
-	SANE_TOOL_PATH ?= $(msvc_bin_dir_msys)
+	SANE_TOOL_PATH ?= $(msvc_bin_dir_msys):$(sdk_ver_bin_dir_msys)
 	HAVE_ALLOCA_H = YesPlease
 	NO_PREAD = YesPlease
 	NEEDS_CRYPTO_WITH_SSL = YesPlease
@@ -485,6 +485,7 @@ ifeq ($(uname_S),Windows)
 	NO_POSIX_GOODIES = UnfortunatelyYes
 	NATIVE_CRLF = YesPlease
 	DEFAULT_HELP_FORMAT = html
+	SKIP_DASHED_BUILT_INS = YabbaDabbaDoo
 ifeq (/mingw64,$(subst 32,64,$(prefix)))
 	# Move system config into top-level /etc/
 	ETC_GITCONFIG = ../etc/gitconfig
@@ -494,20 +495,22 @@ endif
 	CC = compat/vcbuild/scripts/clink.pl
 	AR = compat/vcbuild/scripts/lib.pl
 	CFLAGS =
-	BASIC_CFLAGS = -nologo -I. -Icompat/vcbuild/include -DWIN32 -D_CONSOLE -DHAVE_STRING_H -D_CRT_SECURE_NO_WARNINGS -D_CRT_NONSTDC_NO_DEPRECATE
+	BASIC_CFLAGS = -nologo -I. -Icompat/vcbuild/include -DWIN32 -D_CONSOLE -DHAVE_STRING_H -D_CRT_SECURE_NO_WARNINGS -D_CRT_NONSTDC_NO_DEPRECATE -MP -std:c11
 	COMPAT_OBJS = compat/msvc.o compat/winansi.o \
 		compat/win32/flush.o \
 		compat/win32/path-utils.o \
 		compat/win32/pthread.o compat/win32/syslog.o \
 		compat/win32/trace2_win32_process_info.o \
 		compat/win32/dirent.o
-	COMPAT_CFLAGS = -D__USE_MINGW_ACCESS -DDETECT_MSYS_TTY -DNOGDI -DHAVE_STRING_H -Icompat -Icompat/regex -Icompat/win32 -DSTRIP_EXTENSION=\".exe\"
-	BASIC_LDFLAGS = -IGNORE:4217 -IGNORE:4049 -NOLOGO -ENTRY:wmainCRTStartup -SUBSYSTEM:CONSOLE
+	COMPAT_CFLAGS = -D__USE_MINGW_ACCESS -DDETECT_MSYS_TTY -DENSURE_MSYSTEM_IS_SET -DNOGDI -DHAVE_STRING_H -Icompat -Icompat/regex -Icompat/win32 -DSTRIP_EXTENSION=\".exe\"
+	BASIC_LDFLAGS = -IGNORE:4217 -IGNORE:4049 -NOLOGO
 	# invalidcontinue.obj allows Git's source code to close the same file
 	# handle twice, or to access the osfhandle of an already-closed stdout
 	# See https://msdn.microsoft.com/en-us/library/ms235330.aspx
 	EXTLIBS = user32.lib advapi32.lib shell32.lib wininet.lib ws2_32.lib invalidcontinue.obj kernel32.lib ntdll.lib
+	GITLIBS += git.res
 	PTHREAD_LIBS =
+	RC = compat/vcbuild/scripts/rc.pl
 	lib =
 	BASIC_CFLAGS += $(vcpkg_inc) $(sdk_includes) $(msvc_includes)
 ifndef DEBUG
@@ -677,6 +680,7 @@ ifeq ($(uname_S),MINGW)
 	FSMONITOR_DAEMON_BACKEND = win32
 	FSMONITOR_OS_SETTINGS = win32
 
+	SKIP_DASHED_BUILT_INS = YabbaDabbaDoo
 	RUNTIME_PREFIX = YesPlease
 	HAVE_WPGMPTR = YesWeDo
 	NO_ST_BLOCKS_IN_STRUCT_STAT = YesPlease
@@ -691,7 +695,8 @@ ifeq ($(uname_S),MINGW)
 	DEFAULT_HELP_FORMAT = html
 	HAVE_PLATFORM_PROCINFO = YesPlease
 	CSPRNG_METHOD = rtlgenrandom
-	BASIC_LDFLAGS += -municode
+	BASIC_LDFLAGS += -municode -Wl,--tsaware
+	LAZYLOAD_LIBCURL = YesDoThatPlease
 	COMPAT_CFLAGS += -DNOGDI -Icompat -Icompat/win32
 	COMPAT_CFLAGS += -DSTRIP_EXTENSION=\".exe\"
 	COMPAT_OBJS += compat/mingw.o compat/winansi.o \
@@ -724,13 +729,17 @@ ifeq ($(uname_S),MINGW)
 		prefix = /mingw64
 		HOST_CPU = x86_64
 		BASIC_LDFLAGS += -Wl,--pic-executable,-e,mainCRTStartup
+        else ifeq (CLANGARM64,$(MSYSTEM))
+		prefix = /clangarm64
+		HOST_CPU = aarch64
+		BASIC_LDFLAGS += -Wl,--pic-executable,-e,mainCRTStartup
         else
 		COMPAT_CFLAGS += -D_USE_32BIT_TIME_T
 		BASIC_LDFLAGS += -Wl,--large-address-aware
         endif
 	CC = gcc
 	COMPAT_CFLAGS += -D__USE_MINGW_ANSI_STDIO=0 -DDETECT_MSYS_TTY \
-		-fstack-protector-strong
+		-DENSURE_MSYSTEM_IS_SET -fstack-protector-strong
 	EXTLIBS += -lntdll
 	EXTRA_PROGRAMS += headless-git$X
 	INSTALL = /bin/install
@@ -738,7 +747,8 @@ ifeq ($(uname_S),MINGW)
 	HAVE_LIBCHARSET_H = YesPlease
 	USE_GETTEXT_SCHEME = fallthrough
 	USE_LIBPCRE = YesPlease
-	USE_NED_ALLOCATOR = YesPlease
+	USE_MIMALLOC = YesPlease
+	NO_PYTHON =
         ifeq (/mingw64,$(subst 32,64,$(prefix)))
 		# Move system config into top-level /etc/
 		ETC_GITCONFIG = ../etc/gitconfig
@@ -768,7 +778,7 @@ vcxproj:
 
 	# Make .vcxproj files and add them
 	perl contrib/buildsystems/generate -g Vcxproj
-	git add -f git.sln {*,*/lib,t/helper/*}/*.vcxproj
+	git add -f git.sln {*,*/lib.proj,t/helper/*,reftable/libreftable{,_test}.proj}/*.vcxproj
 
 	# Generate the LinkOrCopyBuiltins.targets and LinkOrCopyRemoteHttp.targets file
 	(echo '<Project xmlns="http://schemas.microsoft.com/developer/msbuild/2003">' && \
@@ -778,7 +788,7 @@ vcxproj:
 	   echo '    <Copy SourceFiles="$$(OutDir)\git.exe" DestinationFiles="$$(OutDir)\'"$$name"'" SkipUnchangedFiles="true" UseHardlinksIfPossible="true" />'; \
 	 done && \
 	 echo '  </Target>' && \
-	 echo '</Project>') >git/LinkOrCopyBuiltins.targets
+	 echo '</Project>') >git.proj/LinkOrCopyBuiltins.targets
 	(echo '<Project xmlns="http://schemas.microsoft.com/developer/msbuild/2003">' && \
 	 echo '  <Target Name="CopyBuiltins_AfterBuild" AfterTargets="AfterBuild">' && \
 	 for name in $(REMOTE_CURL_ALIASES); \
@@ -786,8 +796,8 @@ vcxproj:
 	   echo '    <Copy SourceFiles="$$(OutDir)\'"$(REMOTE_CURL_PRIMARY)"'" DestinationFiles="$$(OutDir)\'"$$name"'" SkipUnchangedFiles="true" UseHardlinksIfPossible="true" />'; \
 	 done && \
 	 echo '  </Target>' && \
-	 echo '</Project>') >git-remote-http/LinkOrCopyRemoteHttp.targets
-	git add -f git/LinkOrCopyBuiltins.targets git-remote-http/LinkOrCopyRemoteHttp.targets
+	 echo '</Project>') >git-remote-http.proj/LinkOrCopyRemoteHttp.targets
+	git add -f git.proj/LinkOrCopyBuiltins.targets git-remote-http.proj/LinkOrCopyRemoteHttp.targets
 
 	# Add generated headers
 	$(MAKE) MSVC=1 SKIP_VCPKG=1 prefix=/mingw64 $(GENERATED_H)
@@ -800,9 +810,11 @@ vcxproj:
 	sed -i '/^git_broken_path_fix ".*/d' git-sh-setup
 	git add -f $(SCRIPT_LIB) $(SCRIPTS)
 
+ifndef NO_PERL
 	# Add Perl module
 	$(MAKE) $(LIB_PERL_GEN)
 	git add -f perl/build
+endif
 
 	# Add bin-wrappers, for testing
 	rm -rf bin-wrappers/
diff --git a/contrib/buildsystems/CMakeLists.txt b/contrib/buildsystems/CMakeLists.txt
index 10dc54fdcb647b..c21eeabd088f1d 100644
--- a/contrib/buildsystems/CMakeLists.txt
+++ b/contrib/buildsystems/CMakeLists.txt
@@ -14,6 +14,11 @@ Note: Visual Studio also has the option of opening `CMakeLists.txt`
 directly; Using this option, Visual Studio will not find the source code,
 though, therefore the `File>Open>Folder...` option is preferred.
 
+Visual Studio does not produce a .sln solution file nor the .vcxproj files
+that may be required by VS extension tools.
+
+To generate the .sln/.vcxproj files run CMake manually, as described below.
+
 Instructions to run CMake manually:
 
     mkdir -p contrib/buildsystems/out
@@ -22,7 +27,7 @@ Instructions to run CMake manually:
 
 This will build the git binaries in contrib/buildsystems/out
 directory (our top-level .gitignore file knows to ignore contents of
-this directory).
+this directory). The project .sln and .vcxproj files are also generated.
 
 Possible build configurations(-DCMAKE_BUILD_TYPE) with corresponding
 compiler flags
@@ -35,17 +40,16 @@ empty(default) :
 NOTE: -DCMAKE_BUILD_TYPE is optional. For multi-config generators like Visual Studio
 this option is ignored
 
-This process generates a Makefile(Linux/*BSD/MacOS) , Visual Studio solution(Windows) by default.
+This process generates a Makefile(Linux/*BSD/MacOS), Visual Studio solution(Windows) by default.
 Run `make` to build Git on Linux/*BSD/MacOS.
 Open git.sln on Windows and build Git.
 
-NOTE: By default CMake uses Makefile as the build tool on Linux and Visual Studio in Windows,
-to use another tool say `ninja` add this to the command line when configuring.
-`-G Ninja`
-
 NOTE: By default CMake will install vcpkg locally to your source tree on configuration,
 to avoid this, add `-DNO_VCPKG=TRUE` to the command line when configuring.
 
+The Visual Studio default generator changed in v16.6 from its Visual Studio
+implemenation to `Ninja` This required changes to many CMake scripts.
+
 ]]
 cmake_minimum_required(VERSION 3.14)
 
@@ -59,15 +63,29 @@ endif()
 
 if(NOT DEFINED CMAKE_EXPORT_COMPILE_COMMANDS)
 	set(CMAKE_EXPORT_COMPILE_COMMANDS TRUE)
+	message("settting CMAKE_EXPORT_COMPILE_COMMANDS: ${CMAKE_EXPORT_COMPILE_COMMANDS}")
 endif()
 
 if(USE_VCPKG)
 	set(VCPKG_DIR "${CMAKE_SOURCE_DIR}/compat/vcbuild/vcpkg")
+	message("WIN32: ${WIN32}") # show its underlying text values
+	message("VCPKG_DIR: ${VCPKG_DIR}")
+	message("VCPKG_ARCH: ${VCPKG_ARCH}") # maybe unset
+	message("MSVC: ${MSVC}")
+	message("CMAKE_GENERATOR: ${CMAKE_GENERATOR}")
+	message("CMAKE_CXX_COMPILER_ID: ${CMAKE_CXX_COMPILER_ID}")
+	message("CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+	message("CMAKE_EXPORT_COMPILE_COMMANDS: ${CMAKE_EXPORT_COMPILE_COMMANDS}")
+	message("ENV(CMAKE_EXPORT_COMPILE_COMMANDS): $ENV{CMAKE_EXPORT_COMPILE_COMMANDS}")
 	if(NOT EXISTS ${VCPKG_DIR})
 		message("Initializing vcpkg and building the Git's dependencies (this will take a while...)")
-		execute_process(COMMAND ${CMAKE_SOURCE_DIR}/compat/vcbuild/vcpkg_install.bat)
+		execute_process(COMMAND ${CMAKE_SOURCE_DIR}/compat/vcbuild/vcpkg_install.bat ${VCPKG_ARCH})
 	endif()
-	list(APPEND CMAKE_PREFIX_PATH "${VCPKG_DIR}/installed/x64-windows")
+	if(NOT EXISTS ${VCPKG_ARCH})
+		message("VCPKG_ARCH: unset, using 'x64-windows'")
+		set(VCPKG_ARCH "x64-windows") # default from vcpkg_install.bat
+	endif()
+	list(APPEND CMAKE_PREFIX_PATH "${VCPKG_DIR}/installed/${VCPKG_ARCH}")
 
 	# In the vcpkg edition, we need this to be able to link to libcurl
 	set(CURL_NO_CURL_CMAKE ON)
@@ -208,11 +226,19 @@ if(CMAKE_C_COMPILER_ID STREQUAL "MSVC")
 	set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG ${CMAKE_BINARY_DIR})
 	set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_RELEASE ${CMAKE_BINARY_DIR})
 	add_compile_options(/MP /std:c11)
+	add_link_options(/MANIFEST:NO)
 endif()
 
 #default behaviour
 include_directories(${CMAKE_SOURCE_DIR})
-add_compile_definitions(GIT_HOST_CPU="${CMAKE_SYSTEM_PROCESSOR}")
+
+# When cross-compiling, define HOST_CPU as the canonical name of the CPU on
+# which the built Git will run (for instance "x86_64").
+if(NOT HOST_CPU)
+	add_compile_definitions(GIT_HOST_CPU="${CMAKE_SYSTEM_PROCESSOR}")
+else()
+	add_compile_definitions(GIT_HOST_CPU="${HOST_CPU}")
+endif()
 add_compile_definitions(SHA256_BLK INTERNAL_QSORT RUNTIME_PREFIX)
 add_compile_definitions(NO_OPENSSL SHA1_DC SHA1DC_NO_STANDARD_INCLUDES
 			SHA1DC_INIT_SAFE_HASH_DEFAULT=0
@@ -745,6 +771,7 @@ if(WIN32)
 	endif()
 
 	add_executable(headless-git ${CMAKE_SOURCE_DIR}/compat/win32/headless.c)
+	list(APPEND PROGRAMS_BUILT headless-git)
 	if(CMAKE_C_COMPILER_ID STREQUAL "GNU" OR CMAKE_C_COMPILER_ID STREQUAL "Clang")
 		target_link_options(headless-git PUBLIC -municode -Wl,-subsystem,windows)
 	elseif(CMAKE_C_COMPILER_ID STREQUAL "MSVC")
@@ -945,7 +972,7 @@ list(TRANSFORM git_perl_scripts PREPEND "${CMAKE_BINARY_DIR}/")
 
 #install
 foreach(program ${PROGRAMS_BUILT})
-if(program MATCHES "^(git|git-shell|scalar)$")
+if(program MATCHES "^(git|git-shell|headless-git|scalar)$")
 install(TARGETS ${program}
 	RUNTIME DESTINATION bin)
 else()
@@ -1207,7 +1234,7 @@ string(REPLACE "@USE_GETTEXT_SCHEME@" "" git_build_options "${git_build_options}
 string(REPLACE "@USE_LIBPCRE2@" "" git_build_options "${git_build_options}")
 string(REPLACE "@X@" "${EXE_EXTENSION}" git_build_options "${git_build_options}")
 if(USE_VCPKG)
-	string(APPEND git_build_options "PATH=\"$PATH:$TEST_DIRECTORY/../compat/vcbuild/vcpkg/installed/x64-windows/bin\"\n")
+	string(APPEND git_build_options "PATH=\"$PATH:$TEST_DIRECTORY/../compat/vcbuild/vcpkg/installed/${VCPKG_ARCH}/bin\"\n")
 endif()
 file(WRITE ${CMAKE_BINARY_DIR}/GIT-BUILD-OPTIONS ${git_build_options})
 
diff --git a/contrib/buildsystems/Generators/Vcxproj.pm b/contrib/buildsystems/Generators/Vcxproj.pm
index b2e68a16715e39..d62249bf7f1010 100644
--- a/contrib/buildsystems/Generators/Vcxproj.pm
+++ b/contrib/buildsystems/Generators/Vcxproj.pm
@@ -58,8 +58,8 @@ sub createProject {
     my $uuid = generate_guid($name);
     $$build_structure{"$prefix${target}_GUID"} = $uuid;
     my $vcxproj = $target;
-    $vcxproj =~ s/(.*\/)?(.*)/$&\/$2.vcxproj/;
-    $vcxproj =~ s/([^\/]*)(\/lib)\/(lib.vcxproj)/$1$2\/$1_$3/;
+    $vcxproj =~ s/(.*\/)?(.*)/$&.proj\/$2.vcxproj/;
+    $vcxproj =~ s/([^\/]*)(\/lib\.proj)\/(lib.vcxproj)/$1$2\/$1_$3/;
     $$build_structure{"$prefix${target}_VCXPROJ"} = $vcxproj;
 
     my @srcs = sort(map("$rel_dir\\$_", @{$$build_structure{"$prefix${name}_SOURCES"}}));
@@ -77,7 +77,7 @@ sub createProject {
     my $libs_release = "\n    ";
     my $libs_debug = "\n    ";
     if (!$static_library && $name ne 'headless-git') {
-      $libs_release = join(";", sort(grep /^(?!libgit\.lib|xdiff\/lib\.lib|vcs-svn\/lib\.lib|reftable\/libreftable\.lib)/, @{$$build_structure{"$prefix${name}_LIBS"}}));
+      $libs_release = join(";", sort(grep /^(?!libgit\.lib|xdiff\/lib\.lib|vcs-svn\/lib\.lib|reftable\/libreftable(_test)?\.lib)/, @{$$build_structure{"$prefix${name}_LIBS"}}));
       $libs_debug = $libs_release;
       $libs_debug =~ s/zlib\.lib/zlibd\.lib/g;
       $libs_debug =~ s/libexpat\.lib/libexpatd\.lib/g;
@@ -88,8 +88,21 @@ sub createProject {
     $defines =~ s/</&lt;/g;
     $defines =~ s/>/&gt;/g;
     $defines =~ s/\'//g;
+    $defines =~ s/\\"/"/g;
 
-    die "Could not create the directory $target for $label project!\n" unless (-d "$target" || mkdir "$target");
+    my $rcdefines = $defines;
+    $rcdefines =~ s/(?<!\\)"/\\$&/g;
+
+    my $entrypoint = 'wmainCRTStartup';
+    my $subsystem = 'Console';
+    if (grep /^-mwindows$/, @{$$build_structure{"$prefix${name}_LFLAGS"}}) {
+        $entrypoint = 'wWinMainCRTStartup';
+        $subsystem = 'Windows';
+    }
+
+    my $dir = $vcxproj;
+    $dir =~ s/\/[^\/]*$//;
+    die "Could not create the directory $dir for $label project!\n" unless (-d "$dir" || mkdir "$dir");
 
     open F, ">$vcxproj" or die "Could not open $vcxproj for writing!\n";
     binmode F, ":crlf :utf8";
@@ -114,12 +127,21 @@ sub createProject {
       <Configuration>Release</Configuration>
       <Platform>x64</Platform>
     </ProjectConfiguration>
+    <ProjectConfiguration Include="Debug|ARM64">
+      <Configuration>Debug</Configuration>
+      <Platform>ARM64</Platform>
+    </ProjectConfiguration>
+    <ProjectConfiguration Include="Release|ARM64">
+      <Configuration>Release</Configuration>
+      <Platform>ARM64</Platform>
+    </ProjectConfiguration>
   </ItemGroup>
   <PropertyGroup Label="Globals">
     <ProjectGuid>$uuid</ProjectGuid>
     <Keyword>Win32Proj</Keyword>
     <VCPKGArch Condition="'\$(Platform)'=='Win32'">x86-windows</VCPKGArch>
-    <VCPKGArch Condition="'\$(Platform)'!='Win32'">x64-windows</VCPKGArch>
+    <VCPKGArch Condition="'\$(Platform)'=='x64'">x64-windows</VCPKGArch>
+    <VCPKGArch Condition="'\$(Platform)'=='ARM64'">arm64-windows</VCPKGArch>
     <VCPKGArchDirectory>$cdup\\compat\\vcbuild\\vcpkg\\installed\\\$(VCPKGArch)</VCPKGArchDirectory>
     <VCPKGBinDirectory Condition="'\$(Configuration)'=='Debug'">\$(VCPKGArchDirectory)\\debug\\bin</VCPKGBinDirectory>
     <VCPKGLibDirectory Condition="'\$(Configuration)'=='Debug'">\$(VCPKGArchDirectory)\\debug\\lib</VCPKGLibDirectory>
@@ -140,7 +162,7 @@ sub createProject {
   </PropertyGroup>
   <PropertyGroup>
     <ConfigurationType>$config_type</ConfigurationType>
-    <PlatformToolset>v140</PlatformToolset>
+    <PlatformToolset>v142</PlatformToolset>
     <!-- <CharacterSet>UTF-8</CharacterSet> -->
     <OutDir>..\\</OutDir>
     <!-- <IntDir>\$(ProjectDir)\$(Configuration)\\</IntDir> -->
@@ -166,6 +188,7 @@ sub createProject {
       <InlineFunctionExpansion>OnlyExplicitInline</InlineFunctionExpansion>
       <PrecompiledHeader />
       <DebugInformationFormat>ProgramDatabase</DebugInformationFormat>
+      <LanguageStandard_C>stdc11</LanguageStandard_C>
     </ClCompile>
     <Lib>
       <SuppressStartupBanner>true</SuppressStartupBanner>
@@ -174,9 +197,8 @@ sub createProject {
       <AdditionalLibraryDirectories>\$(VCPKGLibDirectory);%(AdditionalLibraryDirectories)</AdditionalLibraryDirectories>
       <AdditionalDependencies>\$(VCPKGLibs);\$(AdditionalDependencies)</AdditionalDependencies>
       <AdditionalOptions>invalidcontinue.obj %(AdditionalOptions)</AdditionalOptions>
-      <EntryPointSymbol>wmainCRTStartup</EntryPointSymbol>
-      <ManifestFile>$cdup\\compat\\win32\\git.manifest</ManifestFile>
-      <SubSystem>Console</SubSystem>
+      <EntryPointSymbol>$entrypoint</EntryPointSymbol>
+      <SubSystem>$subsystem</SubSystem>
     </Link>
 EOM
     if ($target eq 'libgit') {
@@ -184,7 +206,7 @@ EOM
     <PreBuildEvent Condition="!Exists('$cdup\\compat\\vcbuild\\vcpkg\\installed\\\$(VCPKGArch)\\include\\openssl\\ssl.h')">
       <Message>Initialize VCPKG</Message>
       <Command>del "$cdup\\compat\\vcbuild\\vcpkg"</Command>
-      <Command>call "$cdup\\compat\\vcbuild\\vcpkg_install.bat"</Command>
+      <Command>call "$cdup\\compat\\vcbuild\\vcpkg_install.bat" \$(VCPKGArch)</Command>
     </PreBuildEvent>
 EOM
     }
@@ -201,6 +223,9 @@ EOM
       <PreprocessorDefinitions>WIN32;_DEBUG;$defines;%(PreprocessorDefinitions)</PreprocessorDefinitions>
       <RuntimeLibrary>MultiThreadedDebugDLL</RuntimeLibrary>
     </ClCompile>
+    <ResourceCompile>
+      <PreprocessorDefinitions>WIN32;_DEBUG;$rcdefines;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+    </ResourceCompile>
     <Link>
       <GenerateDebugInformation>true</GenerateDebugInformation>
     </Link>
@@ -214,6 +239,9 @@ EOM
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <FavorSizeOrSpeed>Speed</FavorSizeOrSpeed>
     </ClCompile>
+    <ResourceCompile>
+      <PreprocessorDefinitions>WIN32;NDEBUG;$rcdefines;%(PreprocessorDefinitions)</PreprocessorDefinitions>
+    </ResourceCompile>
     <Link>
       <GenerateDebugInformation>true</GenerateDebugInformation>
       <EnableCOMDATFolding>true</EnableCOMDATFolding>
@@ -223,9 +251,15 @@ EOM
   <ItemGroup>
 EOM
     foreach(@sources) {
-        print F << "EOM";
+        if (/\.rc$/) {
+            print F << "EOM";
+    <ResourceCompile Include="$_" />
+EOM
+        } else {
+            print F << "EOM";
     <ClCompile Include="$_" />
 EOM
+        }
     }
     print F << "EOM";
   </ItemGroup>
@@ -233,26 +267,31 @@ EOM
     if ((!$static_library || $target =~ 'vcs-svn' || $target =~ 'xdiff') && !($name =~ /headless-git/)) {
       my $uuid_libgit = $$build_structure{"LIBS_libgit_GUID"};
       my $uuid_libreftable = $$build_structure{"LIBS_reftable/libreftable_GUID"};
+      my $uuid_libreftable_test = $$build_structure{"LIBS_reftable/libreftable_test_GUID"};
       my $uuid_xdiff_lib = $$build_structure{"LIBS_xdiff/lib_GUID"};
 
       print F << "EOM";
   <ItemGroup>
-    <ProjectReference Include="$cdup\\libgit\\libgit.vcxproj">
+    <ProjectReference Include="$cdup\\libgit.proj\\libgit.vcxproj">
       <Project>$uuid_libgit</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>
     </ProjectReference>
 EOM
       if (!($name =~ /xdiff|libreftable/)) {
         print F << "EOM";
-    <ProjectReference Include="$cdup\\reftable\\libreftable\\libreftable.vcxproj">
+    <ProjectReference Include="$cdup\\reftable\\libreftable.proj\\libreftable.vcxproj">
       <Project>$uuid_libreftable</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>
     </ProjectReference>
+    <ProjectReference Include="$cdup\\reftable\\libreftable_test.proj\\libreftable_test.vcxproj">
+      <Project>$uuid_libreftable_test</Project>
+      <ReferenceOutputAssembly>false</ReferenceOutputAssembly>
+    </ProjectReference>
 EOM
       }
       if (!($name =~ 'xdiff')) {
         print F << "EOM";
-    <ProjectReference Include="$cdup\\xdiff\\lib\\xdiff_lib.vcxproj">
+    <ProjectReference Include="$cdup\\xdiff\\lib.proj\\xdiff_lib.vcxproj">
       <Project>$uuid_xdiff_lib</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>
     </ProjectReference>
@@ -261,7 +300,7 @@ EOM
       if ($name =~ /(test-(line-buffer|svn-fe)|^git-remote-testsvn)\.exe$/) {
         my $uuid_vcs_svn_lib = $$build_structure{"LIBS_vcs-svn/lib_GUID"};
         print F << "EOM";
-    <ProjectReference Include="$cdup\\vcs-svn\\lib\\vcs-svn_lib.vcxproj">
+    <ProjectReference Include="$cdup\\vcs-svn\\lib.proj\\vcs-svn_lib.vcxproj">
       <Project>$uuid_vcs_svn_lib</Project>
       <ReferenceOutputAssembly>false</ReferenceOutputAssembly>
     </ProjectReference>
@@ -338,7 +377,7 @@ sub createGlueProject {
 	my $vcxproj = $build_structure{"APPS_${appname}_VCXPROJ"};
 	$vcxproj =~ s/\//\\/g;
         $appname =~ s/.*\///;
-        print F "\"${appname}\", \"${vcxproj}\", \"${uuid}\"";
+        print F "\"${appname}.proj\", \"${vcxproj}\", \"${uuid}\"";
         print F "$SLN_POST";
     }
     foreach (@libs) {
@@ -348,15 +387,17 @@ sub createGlueProject {
         my $vcxproj = $build_structure{"LIBS_${libname}_VCXPROJ"};
 	$vcxproj =~ s/\//\\/g;
         $libname =~ s/\//_/g;
-        print F "\"${libname}\", \"${vcxproj}\", \"${uuid}\"";
+        print F "\"${libname}.proj\", \"${vcxproj}\", \"${uuid}\"";
         print F "$SLN_POST";
     }
 
     print F << "EOM";
 Global
 	GlobalSection(SolutionConfigurationPlatforms) = preSolution
+		Debug|ARM64 = Debug|ARM64
 		Debug|x64 = Debug|x64
 		Debug|x86 = Debug|x86
+		Release|ARM64 = Release|ARM64
 		Release|x64 = Release|x64
 		Release|x86 = Release|x86
 	EndGlobalSection
@@ -367,10 +408,14 @@ EOM
     foreach (@apps) {
         my $appname = $_;
         my $uuid = $build_structure{"APPS_${appname}_GUID"};
+        print F "\t\t${uuid}.Debug|ARM64.ActiveCfg = Debug|ARM64\n";
+        print F "\t\t${uuid}.Debug|ARM64.Build.0 = Debug|ARM64\n";
         print F "\t\t${uuid}.Debug|x64.ActiveCfg = Debug|x64\n";
         print F "\t\t${uuid}.Debug|x64.Build.0 = Debug|x64\n";
         print F "\t\t${uuid}.Debug|x86.ActiveCfg = Debug|Win32\n";
         print F "\t\t${uuid}.Debug|x86.Build.0 = Debug|Win32\n";
+        print F "\t\t${uuid}.Release|ARM64.ActiveCfg = Release|ARM64\n";
+        print F "\t\t${uuid}.Release|ARM64.Build.0 = Release|ARM64\n";
         print F "\t\t${uuid}.Release|x64.ActiveCfg = Release|x64\n";
         print F "\t\t${uuid}.Release|x64.Build.0 = Release|x64\n";
         print F "\t\t${uuid}.Release|x86.ActiveCfg = Release|Win32\n";
@@ -379,10 +424,14 @@ EOM
     foreach (@libs) {
         my $libname = $_;
         my $uuid = $build_structure{"LIBS_${libname}_GUID"};
+        print F "\t\t${uuid}.Debug|ARM64.ActiveCfg = Debug|ARM64\n";
+        print F "\t\t${uuid}.Debug|ARM64.Build.0 = Debug|ARM64\n";
         print F "\t\t${uuid}.Debug|x64.ActiveCfg = Debug|x64\n";
         print F "\t\t${uuid}.Debug|x64.Build.0 = Debug|x64\n";
         print F "\t\t${uuid}.Debug|x86.ActiveCfg = Debug|Win32\n";
         print F "\t\t${uuid}.Debug|x86.Build.0 = Debug|Win32\n";
+        print F "\t\t${uuid}.Release|ARM64.ActiveCfg = Release|ARM64\n";
+        print F "\t\t${uuid}.Release|ARM64.Build.0 = Release|ARM64\n";
         print F "\t\t${uuid}.Release|x64.ActiveCfg = Release|x64\n";
         print F "\t\t${uuid}.Release|x64.Build.0 = Release|x64\n";
         print F "\t\t${uuid}.Release|x86.ActiveCfg = Release|Win32\n";
diff --git a/contrib/buildsystems/engine.pl b/contrib/buildsystems/engine.pl
index 069be7e4befcd7..ee4fca200cc506 100755
--- a/contrib/buildsystems/engine.pl
+++ b/contrib/buildsystems/engine.pl
@@ -165,7 +165,7 @@ sub parseMakeOutput
             next;
         }
 
-        if($text =~ / -c /) {
+        if($text =~ / -c / || $text =~ / -i \S+\.rc /) {
             # compilation
             handleCompileLine($text, $line);
 
@@ -263,16 +263,15 @@ sub handleCompileLine
         if ("$part" eq "-o") {
             # ignore object file
             shift @parts;
-        } elsif ("$part" eq "-c") {
+        } elsif ("$part" eq "-c" || "$part" eq "-i" || "$part" =~ /^-fno-/ || "$part" eq '-pedantic') {
             # ignore compile flag
-        } elsif ("$part" eq "-c") {
         } elsif ($part =~ /^.?-I/) {
             push(@incpaths, $part);
         } elsif ($part =~ /^.?-D/) {
             push(@defines, $part);
         } elsif ($part =~ /^-/) {
             push(@cflags, $part);
-        } elsif ($part =~ /\.(c|cc|cpp)$/) {
+        } elsif ($part =~ /\.(c|cc|cpp|rc)$/) {
             $sourcefile = $part;
         } else {
             die "Unhandled compiler option @ line $lineno: $part";
@@ -359,7 +358,7 @@ sub handleLinkLine
             push(@libs, $part);
         } elsif ($part eq 'invalidcontinue.obj') {
             # ignore - known to MSVC
-        } elsif ($part =~ /\.o$/) {
+        } elsif ($part =~ /\.(o|res)$/) {
             push(@objfiles, $part);
         } elsif ($part =~ /\.obj$/) {
             # do nothing, 'make' should not be producing .obj, only .o files
@@ -373,6 +372,7 @@ sub handleLinkLine
         my $sourcefile = $_;
         $sourcefile =~ s/^headless-git\.o$/compat\/win32\/headless.c/;
         $sourcefile =~ s/\.o$/.c/;
+        $sourcefile =~ s/\.res$/.rc/;
         push(@sources, $sourcefile);
         push(@cflags, @{$compile_options{"${sourcefile}_CFLAGS"}});
         push(@defines, @{$compile_options{"${sourcefile}_DEFINES"}});
diff --git a/contrib/subtree/Makefile b/contrib/subtree/Makefile
index 6fa7496bfdb3fd..6f6e90c4cb49b6 100644
--- a/contrib/subtree/Makefile
+++ b/contrib/subtree/Makefile
@@ -94,7 +94,7 @@ $(GIT_SUBTREE_TEST): $(GIT_SUBTREE)
 	cp $< $@
 
 test: $(GIT_SUBTREE_TEST)
-	$(MAKE) -C t/ test
+	$(MAKE) -C t/ all
 
 clean:
 	$(RM) $(GIT_SUBTREE)
diff --git a/dir.c b/dir.c
index 5b2181e5899ce9..16ccfe7e4e868d 100644
--- a/dir.c
+++ b/dir.c
@@ -1093,10 +1093,6 @@ static void invalidate_directory(struct untracked_cache *uc,
 		dir->dirs[i]->recurse = 0;
 }
 
-static int add_patterns_from_buffer(char *buf, size_t size,
-				    const char *base, int baselen,
-				    struct pattern_list *pl);
-
 /* Flags for add_patterns() */
 #define PATTERN_NOFOLLOW (1<<0)
 
@@ -1186,9 +1182,9 @@ static int add_patterns(const char *fname, const char *base, int baselen,
 	return 0;
 }
 
-static int add_patterns_from_buffer(char *buf, size_t size,
-				    const char *base, int baselen,
-				    struct pattern_list *pl)
+int add_patterns_from_buffer(char *buf, size_t size,
+			     const char *base, int baselen,
+			     struct pattern_list *pl)
 {
 	char *orig = buf;
 	int i, lineno = 1;
diff --git a/dir.h b/dir.h
index a3a2f00f5d9273..6cfef5df66091b 100644
--- a/dir.h
+++ b/dir.h
@@ -467,6 +467,9 @@ void add_patterns_from_file(struct dir_struct *, const char *fname);
 int add_patterns_from_blob_to_list(struct object_id *oid,
 				   const char *base, int baselen,
 				   struct pattern_list *pl);
+int add_patterns_from_buffer(char *buf, size_t size,
+			     const char *base, int baselen,
+			     struct pattern_list *pl);
 void parse_path_pattern(const char **string, int *patternlen, unsigned *flags, int *nowildcardlen);
 void add_pattern(const char *string, const char *base,
 		 int baselen, struct pattern_list *pl, int srcpos);
diff --git a/git-compat-util.h b/git-compat-util.h
index e283c46c6fa06e..e42c61193f582a 100644
--- a/git-compat-util.h
+++ b/git-compat-util.h
@@ -176,7 +176,7 @@ DISABLE_WARNING(-Wsign-compare)
 /* Approximation of the length of the decimal representation of this type. */
 #define decimal_length(x)	((int)(sizeof(x) * 2.56 + 0.5) + 1)
 
-#ifdef __MINGW64__
+#if defined(__MINGW32__) || defined(__MINGW64__)
 #define _POSIX_C_SOURCE 1
 #elif defined(__sun__)
  /*
@@ -201,7 +201,9 @@ DISABLE_WARNING(-Wsign-compare)
 #define _ALL_SOURCE 1
 #define _GNU_SOURCE 1
 #define _BSD_SOURCE 1
+#ifndef _DEFAULT_SOURCE
 #define _DEFAULT_SOURCE 1
+#endif
 #define _NETBSD_SOURCE 1
 #define _SGI_SOURCE 1
 
@@ -441,6 +443,16 @@ char *gitdirname(char *);
 # include <sys/sysinfo.h>
 #endif
 
+#ifdef USE_MIMALLOC
+#include "mimalloc.h"
+#define malloc mi_malloc
+#define calloc mi_calloc
+#define realloc mi_realloc
+#define free mi_free
+#define strdup mi_strdup
+#define strndup mi_strndup
+#endif
+
 /* On most systems <netdb.h> would have given us this, but
  * not on some systems (e.g. z/OS).
  */
@@ -537,6 +549,13 @@ static inline int git_offset_1st_component(const char *path)
 #define fspathncmp git_fspathncmp
 #endif
 
+#ifndef warn_about_git_lfs_on_windows7
+static inline void warn_about_git_lfs_on_windows7(int exit_code UNUSED,
+						  const char *argv0 UNUSED)
+{
+}
+#endif
+
 #ifndef is_valid_path
 #define is_valid_path(path) 1
 #endif
@@ -626,10 +645,18 @@ static inline int git_has_dir_sep(const char *path)
 #define has_dir_sep(path) git_has_dir_sep(path)
 #endif
 
+#ifndef is_mount_point
+#define is_mount_point is_mount_point_via_stat
+#endif
+
 #ifndef query_user_email
 #define query_user_email() NULL
 #endif
 
+#ifndef platform_strbuf_realpath
+#define platform_strbuf_realpath(resolved, path) NULL
+#endif
+
 #ifdef __TANDEM
 #include <floss.h(floss_execl,floss_execlp,floss_execv,floss_execvp)>
 #include <floss.h(floss_getpwuid)>
diff --git a/git-curl-compat.h b/git-curl-compat.h
index 703756ba851d5b..8291755f50454f 100644
--- a/git-curl-compat.h
+++ b/git-curl-compat.h
@@ -45,4 +45,12 @@
 #define GIT_CURL_HAVE_CURLOPT_PROTOCOLS_STR 1
 #endif
 
+/**
+ * CURLSSLOPT_AUTO_CLIENT_CERT was added in 7.77.0, released in May
+ * 2021.
+ */
+#if LIBCURL_VERSION_NUM >= 0x074d00
+#define GIT_CURL_HAVE_CURLSSLOPT_AUTO_CLIENT_CERT
+#endif
+
 #endif
diff --git a/git-gui/git-gui.sh b/git-gui/git-gui.sh
index 887d6d596c16fc..fc8b761666c9ee 100755
--- a/git-gui/git-gui.sh
+++ b/git-gui/git-gui.sh
@@ -101,6 +101,9 @@ proc _which {what args} {
 
 	if {[is_Windows] && [lsearch -exact $args -script] >= 0} {
 		set suffix {}
+	} elseif {[string match *$_search_exe $what]} {
+		# The search string already has the file extension
+		set suffix {}
 	} else {
 		set suffix $_search_exe
 	}
@@ -2078,6 +2081,7 @@ set all_icons(U$ui_index)   file_merge
 set all_icons(T$ui_index)   file_statechange
 
 set all_icons(_$ui_workdir) file_plain
+set all_icons(A$ui_workdir) file_plain
 set all_icons(M$ui_workdir) file_mod
 set all_icons(D$ui_workdir) file_question
 set all_icons(U$ui_workdir) file_merge
@@ -2104,6 +2108,7 @@ foreach i {
 		{A_ {mc "Staged for commit"}}
 		{AM {mc "Portions staged for commit"}}
 		{AD {mc "Staged for commit, missing"}}
+		{AA {mc "Intended to be added"}}
 
 		{_D {mc "Missing"}}
 		{D_ {mc "Staged for removal"}}
diff --git a/git-gui/lib/diff.tcl b/git-gui/lib/diff.tcl
index d657bfec05b498..9a593a3f6e1547 100644
--- a/git-gui/lib/diff.tcl
+++ b/git-gui/lib/diff.tcl
@@ -568,7 +568,8 @@ proc apply_or_revert_hunk {x y revert} {
 	if {$current_diff_side eq $ui_index} {
 		set failed_msg [mc "Failed to unstage selected hunk."]
 		lappend apply_cmd --reverse --cached
-		if {[string index $mi 0] ne {M}} {
+		set file_state [string index $mi 0]
+		if {$file_state ne {M} && $file_state ne {A}} {
 			unlock_index
 			return
 		}
@@ -581,7 +582,8 @@ proc apply_or_revert_hunk {x y revert} {
 			lappend apply_cmd --cached
 		}
 
-		if {[string index $mi 1] ne {M}} {
+		set file_state [string index $mi 1]
+		if {$file_state ne {M} && $file_state ne {A}} {
 			unlock_index
 			return
 		}
@@ -673,7 +675,8 @@ proc apply_or_revert_range_or_line {x y revert} {
 		set failed_msg [mc "Failed to unstage selected line."]
 		set to_context {+}
 		lappend apply_cmd --reverse --cached
-		if {[string index $mi 0] ne {M}} {
+		set file_state [string index $mi 0]
+		if {$file_state ne {M} && $file_state ne {A}} {
 			unlock_index
 			return
 		}
@@ -688,7 +691,8 @@ proc apply_or_revert_range_or_line {x y revert} {
 			lappend apply_cmd --cached
 		}
 
-		if {[string index $mi 1] ne {M}} {
+		set file_state [string index $mi 1]
+		if {$file_state ne {M} && $file_state ne {A}} {
 			unlock_index
 			return
 		}
diff --git a/git.c b/git.c
index 71d644dc1c5990..71f4a9c37236ab 100644
--- a/git.c
+++ b/git.c
@@ -506,6 +506,7 @@ static struct cmd_struct commands[] = {
 	{ "annotate", cmd_annotate, RUN_SETUP },
 	{ "apply", cmd_apply, RUN_SETUP_GENTLY },
 	{ "archive", cmd_archive, RUN_SETUP_GENTLY },
+	{ "backfill", cmd_backfill, RUN_SETUP },
 	{ "bisect", cmd_bisect, RUN_SETUP },
 	{ "blame", cmd_blame, RUN_SETUP },
 	{ "branch", cmd_branch, RUN_SETUP | DELAY_PAGER_CONFIG },
@@ -626,6 +627,7 @@ static struct cmd_struct commands[] = {
 	{ "status", cmd_status, RUN_SETUP | NEED_WORK_TREE },
 	{ "stripspace", cmd_stripspace },
 	{ "submodule--helper", cmd_submodule__helper, RUN_SETUP },
+	{ "survey", cmd_survey, RUN_SETUP },
 	{ "switch", cmd_switch, RUN_SETUP | NEED_WORK_TREE },
 	{ "symbolic-ref", cmd_symbolic_ref, RUN_SETUP },
 	{ "tag", cmd_tag, RUN_SETUP | DELAY_PAGER_CONFIG },
diff --git a/git.rc.in b/git.rc.in
index e69444eef3f0c5..cd671dc2f0d134 100644
--- a/git.rc.in
+++ b/git.rc.in
@@ -1,3 +1,4 @@
+#include<winuser.h>
 1 VERSIONINFO
 FILEVERSION     @GIT_MAJOR_VERSION@,@GIT_MINOR_VERSION@,@GIT_MICRO_VERSION@,@GIT_PATCH_LEVEL@
 PRODUCTVERSION  @GIT_MAJOR_VERSION@,@GIT_MINOR_VERSION@,@GIT_MICRO_VERSION@,@GIT_PATCH_LEVEL@
@@ -12,6 +13,7 @@ BEGIN
       VALUE "OriginalFilename", "git.exe\0"
       VALUE "ProductName", "Git\0"
       VALUE "ProductVersion", "@GIT_VERSION@\0"
+      VALUE "FileVersion", "@GIT_VERSION@\0"
     END
   END
 
diff --git a/gitk-git/gitk b/gitk-git/gitk
index 47a7c1d29c430f..5c8790dccd2683 100755
--- a/gitk-git/gitk
+++ b/gitk-git/gitk
@@ -9,6 +9,141 @@ exec wish "$0" -- "$@"
 
 package require Tk
 
+######################################################################
+##
+## Enabling platform-specific code paths
+
+proc is_MacOSX {} {
+	if {[tk windowingsystem] eq {aqua}} {
+		return 1
+	}
+	return 0
+}
+
+proc is_Windows {} {
+	if {$::tcl_platform(platform) eq {windows}} {
+		return 1
+	}
+	return 0
+}
+
+set _iscygwin {}
+proc is_Cygwin {} {
+	global _iscygwin
+	if {$_iscygwin eq {}} {
+		if {[string match "CYGWIN_*" $::tcl_platform(os)]} {
+			set _iscygwin 1
+		} else {
+			set _iscygwin 0
+		}
+	}
+	return $_iscygwin
+}
+
+######################################################################
+##
+## PATH lookup
+
+set _search_path {}
+proc _which {what args} {
+	global env _search_exe _search_path
+
+	if {$_search_path eq {}} {
+		if {[is_Cygwin] && [regexp {^(/|\.:)} $env(PATH)]} {
+			set _search_path [split [exec cygpath \
+				--windows \
+				--path \
+				--absolute \
+				$env(PATH)] {;}]
+			set _search_exe .exe
+		} elseif {[is_Windows]} {
+			set gitguidir [file dirname [info script]]
+			regsub -all ";" $gitguidir "\\;" gitguidir
+			set env(PATH) "$gitguidir;$env(PATH)"
+			set _search_path [split $env(PATH) {;}]
+			# Skip empty `PATH` elements
+			set _search_path [lsearch -all -inline -not -exact \
+				$_search_path ""]
+			set _search_exe .exe
+		} else {
+			set _search_path [split $env(PATH) :]
+			set _search_exe {}
+		}
+	}
+
+	if {[is_Windows] && [lsearch -exact $args -script] >= 0} {
+		set suffix {}
+	} else {
+		set suffix $_search_exe
+	}
+
+	foreach p $_search_path {
+		set p [file join $p $what$suffix]
+		if {[file exists $p]} {
+			return [file normalize $p]
+		}
+	}
+	return {}
+}
+
+proc sanitize_command_line {command_line from_index} {
+	set i $from_index
+	while {$i < [llength $command_line]} {
+		set cmd [lindex $command_line $i]
+		if {[file pathtype $cmd] ne "absolute"} {
+			set fullpath [_which $cmd]
+			if {$fullpath eq ""} {
+				throw {NOT-FOUND} "$cmd not found in PATH"
+			}
+			lset command_line $i $fullpath
+		}
+
+		# handle piped commands, e.g. `exec A | B`
+		for {incr i} {$i < [llength $command_line]} {incr i} {
+			if {[lindex $command_line $i] eq "|"} {
+				incr i
+				break
+			}
+		}
+	}
+	return $command_line
+}
+
+# Override `exec` to avoid unsafe PATH lookup
+
+rename exec real_exec
+
+proc exec {args} {
+	# skip options
+	for {set i 0} {$i < [llength $args]} {incr i} {
+		set arg [lindex $args $i]
+		if {$arg eq "--"} {
+			incr i
+			break
+		}
+		if {[string range $arg 0 0] ne "-"} {
+			break
+		}
+	}
+	set args [sanitize_command_line $args $i]
+	uplevel 1 real_exec $args
+}
+
+# Override `open` to avoid unsafe PATH lookup
+
+rename open real_open
+
+proc open {args} {
+	set arg0 [lindex $args 0]
+	if {[string range $arg0 0 0] eq "|"} {
+		set command_line [string trim [string range $arg0 1 end]]
+		lset args 0 "| [sanitize_command_line $command_line 0]"
+	}
+	uplevel 1 real_open $args
+}
+
+# End of safe PATH lookup stuff
+
 proc hasworktree {} {
     return [expr {[exec git rev-parse --is-bare-repository] == "false" &&
                   [exec git rev-parse --is-inside-git-dir] == "false"}]
diff --git a/http.c b/http.c
index c8fc15aa118d3b..b6a5b87a4c3d63 100644
--- a/http.c
+++ b/http.c
@@ -143,7 +143,13 @@ static char *cached_accept_language;
 
 static char *http_ssl_backend;
 
-static int http_schannel_check_revoke = 1;
+static int http_schannel_check_revoke_mode =
+#ifdef CURLSSLOPT_REVOKE_BEST_EFFORT
+	CURLSSLOPT_REVOKE_BEST_EFFORT;
+#else
+	CURLSSLOPT_NO_REVOKE;
+#endif
+
 /*
  * With the backend being set to `schannel`, setting sslCAinfo would override
  * the Certificate Store in cURL v7.60.0 and later, which is not what we want
@@ -151,6 +157,8 @@ static int http_schannel_check_revoke = 1;
  */
 static int http_schannel_use_ssl_cainfo;
 
+static int http_auto_client_cert;
+
 static int always_auth_proactively(void)
 {
 	return http_proactive_auth != PROACTIVE_AUTH_NONE &&
@@ -418,7 +426,19 @@ static int http_options(const char *var, const char *value,
 	}
 
 	if (!strcmp("http.schannelcheckrevoke", var)) {
-		http_schannel_check_revoke = git_config_bool(var, value);
+		if (value && !strcmp(value, "best-effort")) {
+			http_schannel_check_revoke_mode =
+#ifdef CURLSSLOPT_REVOKE_BEST_EFFORT
+				CURLSSLOPT_REVOKE_BEST_EFFORT;
+#else
+				CURLSSLOPT_NO_REVOKE;
+			warning(_("%s=%s unsupported by current cURL"),
+				var, value);
+#endif
+		} else
+			http_schannel_check_revoke_mode =
+				(git_config_bool(var, value) ?
+				 0 : CURLSSLOPT_NO_REVOKE);
 		return 0;
 	}
 
@@ -427,6 +447,11 @@ static int http_options(const char *var, const char *value,
 		return 0;
 	}
 
+	if (!strcmp("http.sslautoclientcert", var)) {
+		http_auto_client_cert = git_config_bool(var, value);
+		return 0;
+	}
+
 	if (!strcmp("http.minsessions", var)) {
 		min_curl_sessions = git_config_int(var, value, ctx->kvi);
 		if (min_curl_sessions > 1)
@@ -1044,9 +1069,20 @@ static CURL *get_curl_handle(void)
 	}
 #endif
 
-	if (http_ssl_backend && !strcmp("schannel", http_ssl_backend) &&
-	    !http_schannel_check_revoke) {
-		curl_easy_setopt(result, CURLOPT_SSL_OPTIONS, CURLSSLOPT_NO_REVOKE);
+	if (http_ssl_backend && !strcmp("schannel", http_ssl_backend)) {
+		long ssl_options = 0;
+		if (http_schannel_check_revoke_mode) {
+			ssl_options |= http_schannel_check_revoke_mode;
+		}
+
+		if (http_auto_client_cert) {
+#ifdef GIT_CURL_HAVE_CURLSSLOPT_AUTO_CLIENT_CERT
+			ssl_options |= CURLSSLOPT_AUTO_CLIENT_CERT;
+#endif
+		}
+
+		if (ssl_options)
+			curl_easy_setopt(result, CURLOPT_SSL_OPTIONS, ssl_options);
 	}
 
 	if (http_proactive_auth != PROACTIVE_AUTH_NONE)
diff --git a/meson.build b/meson.build
index 0064eb64f546a6..38cdec55728f6a 100644
--- a/meson.build
+++ b/meson.build
@@ -358,6 +358,7 @@ libgit_sources = [
   'patch-delta.c',
   'patch-ids.c',
   'path.c',
+  'path-walk.c',
   'pathspec.c',
   'pkt-line.c',
   'preload-index.c',
@@ -486,6 +487,7 @@ builtin_sources = [
   'builtin/annotate.c',
   'builtin/apply.c',
   'builtin/archive.c',
+  'builtin/backfill.c',
   'builtin/bisect.c',
   'builtin/blame.c',
   'builtin/branch.c',
@@ -590,6 +592,7 @@ builtin_sources = [
   'builtin/stash.c',
   'builtin/stripspace.c',
   'builtin/submodule--helper.c',
+  'builtin/survey.c',
   'builtin/symbolic-ref.c',
   'builtin/tag.c',
   'builtin/unpack-file.c',
diff --git a/object-file.c b/object-file.c
index f84dcd2f2a7b88..82c04083a0d665 100644
--- a/object-file.c
+++ b/object-file.c
@@ -1935,9 +1935,9 @@ void *read_object_with_reference(struct repository *r,
 }
 
 static void hash_object_body(const struct git_hash_algo *algo, git_hash_ctx *c,
-			     const void *buf, unsigned long len,
+			     const void *buf, size_t len,
 			     struct object_id *oid,
-			     char *hdr, int *hdrlen)
+			     char *hdr, size_t *hdrlen)
 {
 	algo->init_fn(c);
 	algo->update_fn(c, hdr, *hdrlen);
@@ -1946,23 +1946,23 @@ static void hash_object_body(const struct git_hash_algo *algo, git_hash_ctx *c,
 }
 
 static void write_object_file_prepare(const struct git_hash_algo *algo,
-				      const void *buf, unsigned long len,
+				      const void *buf, size_t len,
 				      enum object_type type, struct object_id *oid,
-				      char *hdr, int *hdrlen)
+				      char *hdr, size_t *hdrlen)
 {
 	git_hash_ctx c;
 
 	/* Generate the header */
 	*hdrlen = format_object_header(hdr, *hdrlen, type, len);
 
-	/* Sha1.. */
+	/* Hash (function pointers) computation */
 	hash_object_body(algo, &c, buf, len, oid, hdr, hdrlen);
 }
 
 static void write_object_file_prepare_literally(const struct git_hash_algo *algo,
-				      const void *buf, unsigned long len,
+				      const void *buf, size_t len,
 				      const char *type, struct object_id *oid,
-				      char *hdr, int *hdrlen)
+				      char *hdr, size_t *hdrlen)
 {
 	git_hash_ctx c;
 
@@ -2084,17 +2084,17 @@ int finalize_object_file_flags(const char *tmpfile, const char *filename,
 }
 
 static void hash_object_file_literally(const struct git_hash_algo *algo,
-				       const void *buf, unsigned long len,
+				       const void *buf, size_t len,
 				       const char *type, struct object_id *oid)
 {
 	char hdr[MAX_HEADER_LEN];
-	int hdrlen = sizeof(hdr);
+	size_t hdrlen = sizeof(hdr);
 
 	write_object_file_prepare_literally(algo, buf, len, type, oid, hdr, &hdrlen);
 }
 
 void hash_object_file(const struct git_hash_algo *algo, const void *buf,
-		      unsigned long len, enum object_type type,
+		      size_t len, enum object_type type,
 		      struct object_id *oid)
 {
 	hash_object_file_literally(algo, buf, len, type_name(type), oid);
@@ -2460,7 +2460,7 @@ int stream_loose_object(struct input_stream *in_stream, size_t len,
 	return err;
 }
 
-int write_object_file_flags(const void *buf, unsigned long len,
+int write_object_file_flags(const void *buf, size_t len,
 			    enum object_type type, struct object_id *oid,
 			    struct object_id *compat_oid_in, unsigned flags)
 {
@@ -2469,7 +2469,7 @@ int write_object_file_flags(const void *buf, unsigned long len,
 	const struct git_hash_algo *compat = repo->compat_hash_algo;
 	struct object_id compat_oid;
 	char hdr[MAX_HEADER_LEN];
-	int hdrlen = sizeof(hdr);
+	size_t hdrlen = sizeof(hdr);
 
 	/* Generate compat_oid */
 	if (compat) {
@@ -2500,7 +2500,7 @@ int write_object_file_flags(const void *buf, unsigned long len,
 	return 0;
 }
 
-int write_object_file_literally(const void *buf, unsigned long len,
+int write_object_file_literally(const void *buf, size_t len,
 				const char *type, struct object_id *oid,
 				unsigned flags)
 {
@@ -2509,8 +2509,8 @@ int write_object_file_literally(const void *buf, unsigned long len,
 	const struct git_hash_algo *algo = repo->hash_algo;
 	const struct git_hash_algo *compat = repo->compat_hash_algo;
 	struct object_id compat_oid;
-	int hdrlen, status = 0;
-	int compat_type = -1;
+	size_t hdrlen;
+	int status = 0, compat_type = -1;
 
 	if (compat) {
 		compat_type = type_from_string_gently(type, -1, 1);
diff --git a/object-store-ll.h b/object-store-ll.h
index cd3bd5bd99f78c..f9eee3433a112c 100644
--- a/object-store-ll.h
+++ b/object-store-ll.h
@@ -270,10 +270,10 @@ void *repo_read_object_file(struct repository *r,
 int oid_object_info(struct repository *r, const struct object_id *, unsigned long *);
 
 void hash_object_file(const struct git_hash_algo *algo, const void *buf,
-		      unsigned long len, enum object_type type,
+		      size_t len, enum object_type type,
 		      struct object_id *oid);
 
-int write_object_file_flags(const void *buf, unsigned long len,
+int write_object_file_flags(const void *buf, size_t len,
 			    enum object_type type, struct object_id *oid,
 			    struct object_id *comapt_oid_in, unsigned flags);
 static inline int write_object_file(const void *buf, unsigned long len,
@@ -282,7 +282,7 @@ static inline int write_object_file(const void *buf, unsigned long len,
 	return write_object_file_flags(buf, len, type, oid, NULL, 0);
 }
 
-int write_object_file_literally(const void *buf, unsigned long len,
+int write_object_file_literally(const void *buf, size_t len,
 				const char *type, struct object_id *oid,
 				unsigned flags);
 int stream_loose_object(struct input_stream *in_stream, size_t len,
diff --git a/pack-objects.h b/pack-objects.h
index 3f6f5042030041..5ec9efdf4af210 100644
--- a/pack-objects.h
+++ b/pack-objects.h
@@ -119,11 +119,23 @@ struct object_entry {
 	unsigned ext_base:1; /* delta_idx points outside packlist */
 };
 
+/**
+ * A packing region is a section of the packing_data.objects array
+ * as given by a starting index and a number of elements.
+ */
+struct packing_region {
+	uint32_t start;
+	uint32_t nr;
+};
+
 struct packing_data {
 	struct repository *repo;
 	struct object_entry *objects;
 	uint32_t nr_objects, nr_alloc;
 
+	struct packing_region *regions;
+	uint32_t nr_regions, nr_regions_alloc;
+
 	int32_t *index;
 	uint32_t index_size;
 
@@ -208,6 +220,27 @@ static inline uint32_t pack_name_hash(const char *name)
 	return hash;
 }
 
+static inline uint32_t pack_full_name_hash(const char *name)
+{
+	const uint32_t bigp = 1234572167U;
+	uint32_t c, hash = bigp;
+
+	if (!name)
+		return 0;
+
+	/*
+	 * Do the simplest thing that will resemble pseudo-randomness: add
+	 * random multiples of a large prime number with a binary shift.
+	 * The goal is not to be cryptographic, but to be generally
+	 * uniformly distributed.
+	 */
+	while ((c = *name++) != 0) {
+		hash += c * bigp;
+		hash = (hash >> 5) | (hash << 27);
+	}
+	return hash;
+}
+
 static inline enum object_type oe_type(const struct object_entry *e)
 {
 	return e->type_valid ? e->type_ : OBJ_BAD;
diff --git a/path-walk.c b/path-walk.c
new file mode 100644
index 00000000000000..c0be95922c1615
--- /dev/null
+++ b/path-walk.c
@@ -0,0 +1,425 @@
+/*
+ * path-walk.c: implementation for path-based walks of the object graph.
+ */
+#include "git-compat-util.h"
+#include "path-walk.h"
+#include "blob.h"
+#include "commit.h"
+#include "dir.h"
+#include "hashmap.h"
+#include "hex.h"
+#include "object.h"
+#include "oid-array.h"
+#include "repository.h"
+#include "revision.h"
+#include "string-list.h"
+#include "strmap.h"
+#include "tag.h"
+#include "trace2.h"
+#include "tree.h"
+#include "tree-walk.h"
+
+struct type_and_oid_list
+{
+	enum object_type type;
+	struct oid_array oids;
+	int maybe_interesting;
+};
+
+#define TYPE_AND_OID_LIST_INIT { \
+	.type = OBJ_NONE, 	 \
+	.oids = OID_ARRAY_INIT	 \
+}
+
+struct path_walk_context {
+	/**
+	 * Repeats of data in 'struct path_walk_info' for
+	 * access with fewer characters.
+	 */
+	struct repository *repo;
+	struct rev_info *revs;
+	struct path_walk_info *info;
+
+	/**
+	 * Map a path to a 'struct type_and_oid_list'
+	 * containing the objects discovered at that
+	 * path.
+	 */
+	struct strmap paths_to_lists;
+
+	/**
+	 * Store the current list of paths in a stack, to
+	 * facilitate depth-first-search without recursion.
+	 */
+	struct string_list path_stack;
+};
+
+static int add_children(struct path_walk_context *ctx,
+			const char *base_path,
+			struct object_id *oid)
+{
+	struct tree_desc desc;
+	struct name_entry entry;
+	struct strbuf path = STRBUF_INIT;
+	size_t base_len;
+	struct tree *tree = lookup_tree(ctx->repo, oid);
+
+	if (!tree) {
+		error(_("failed to walk children of tree %s: not found"),
+		      oid_to_hex(oid));
+		return -1;
+	} else if (parse_tree_gently(tree, 1)) {
+		die("bad tree object %s", oid_to_hex(oid));
+	}
+
+	strbuf_addstr(&path, base_path);
+	base_len = path.len;
+
+	parse_tree(tree);
+	init_tree_desc(&desc, &tree->object.oid, tree->buffer, tree->size);
+	while (tree_entry(&desc, &entry)) {
+		struct type_and_oid_list *list;
+		struct object *o;
+		/* Not actually true, but we will ignore submodules later. */
+		enum object_type type = S_ISDIR(entry.mode) ? OBJ_TREE : OBJ_BLOB;
+
+		/* Skip submodules. */
+		if (S_ISGITLINK(entry.mode))
+			continue;
+
+		/* If the caller doesn't want blobs, then don't bother. */
+		if (!ctx->info->blobs && type == OBJ_BLOB)
+			continue;
+
+		if (type == OBJ_TREE) {
+			struct tree *child = lookup_tree(ctx->repo, &entry.oid);
+			o = child ? &child->object : NULL;
+		} else if (type == OBJ_BLOB) {
+			struct blob *child = lookup_blob(ctx->repo, &entry.oid);
+			o = child ? &child->object : NULL;
+		} else {
+			/* Wrong type? */
+			continue;
+		}
+
+		if (!o) /* report error?*/
+			continue;
+
+		/* Skip this object if already seen. */
+		if (o->flags & SEEN)
+			continue;
+		o->flags |= SEEN;
+
+		strbuf_setlen(&path, base_len);
+		strbuf_add(&path, entry.path, entry.pathlen);
+
+		/*
+		 * Trees will end with "/" for concatenation and distinction
+		 * from blobs at the same path.
+		 */
+		if (type == OBJ_TREE)
+			strbuf_addch(&path, '/');
+
+		if (ctx->info->pl) {
+			int dtype;
+			enum pattern_match_result match;
+			match = path_matches_pattern_list(path.buf, path.len,
+							  path.buf + base_len, &dtype,
+							  ctx->info->pl,
+							  ctx->repo->index);
+
+			if (ctx->info->pl->use_cone_patterns &&
+			    match == NOT_MATCHED)
+				continue;
+			else if (!ctx->info->pl->use_cone_patterns &&
+				 type == OBJ_BLOB &&
+				 match != MATCHED)
+				continue;
+		}
+
+		if (!(list = strmap_get(&ctx->paths_to_lists, path.buf))) {
+			CALLOC_ARRAY(list, 1);
+			list->type = type;
+			strmap_put(&ctx->paths_to_lists, path.buf, list);
+			string_list_append(&ctx->path_stack, path.buf);
+		}
+		if (!(o->flags & UNINTERESTING))
+			list->maybe_interesting = 1;
+		oid_array_append(&list->oids, &entry.oid);
+	}
+
+	free_tree_buffer(tree);
+	strbuf_release(&path);
+	return 0;
+}
+
+/*
+ * For each path in paths_to_explore, walk the trees another level
+ * and add any found blobs to the batch (but only if they exist and
+ * haven't been added yet).
+ */
+static int walk_path(struct path_walk_context *ctx,
+		     const char *path)
+{
+	struct type_and_oid_list *list;
+	int ret = 0;
+
+	list = strmap_get(&ctx->paths_to_lists, path);
+
+	if (ctx->info->prune_all_uninteresting) {
+		/*
+		 * This is true if all objects were UNINTERESTING
+		 * when added to the list.
+		 */
+		if (!list->maybe_interesting)
+			return 0;
+
+		/*
+		 * But it's still possible that the objects were set
+		 * as UNINTERESTING after being added. Do a quick check.
+		 */
+		list->maybe_interesting = 0;
+		for (size_t i = 0;
+		     !list->maybe_interesting && i < list->oids.nr;
+		     i++) {
+			if (list->type == OBJ_TREE) {
+				struct tree *t = lookup_tree(ctx->repo,
+							     &list->oids.oid[i]);
+				if (t && !(t->object.flags & UNINTERESTING))
+					list->maybe_interesting = 1;
+			} else {
+				struct blob *b = lookup_blob(ctx->repo,
+							     &list->oids.oid[i]);
+				if (b && !(b->object.flags & UNINTERESTING))
+					list->maybe_interesting = 1;
+			}
+		}
+
+		/* We have confirmed that all objects are UNINTERESTING. */
+		if (!list->maybe_interesting)
+			return 0;
+	}
+
+	/* Evaluate function pointer on this data, if requested. */
+	if ((list->type == OBJ_TREE && ctx->info->trees) ||
+	    (list->type == OBJ_BLOB && ctx->info->blobs))
+		ret = ctx->info->path_fn(path, &list->oids, list->type,
+					ctx->info->path_fn_data);
+
+	/* Expand data for children. */
+	if (list->type == OBJ_TREE) {
+		for (size_t i = 0; i < list->oids.nr; i++) {
+			ret |= add_children(ctx,
+					    path,
+					    &list->oids.oid[i]);
+		}
+	}
+
+	oid_array_clear(&list->oids);
+	strmap_remove(&ctx->paths_to_lists, path, 1);
+	return ret;
+}
+
+static void clear_strmap(struct strmap *map)
+{
+	struct hashmap_iter iter;
+	struct strmap_entry *e;
+
+	hashmap_for_each_entry(&map->map, &iter, e, ent) {
+		struct type_and_oid_list *list = e->value;
+		oid_array_clear(&list->oids);
+	}
+	strmap_clear(map, 1);
+	strmap_init(map);
+}
+
+/**
+ * Given the configuration of 'info', walk the commits based on 'info->revs' and
+ * call 'info->path_fn' on each discovered path.
+ *
+ * Returns nonzero on an error.
+ */
+int walk_objects_by_path(struct path_walk_info *info)
+{
+	const char *root_path = "";
+	int ret = 0, has_uninteresting = 0;
+	size_t commits_nr = 0, paths_nr = 0;
+	struct commit *c;
+	struct type_and_oid_list *root_tree_list;
+	struct type_and_oid_list *commit_list;
+	struct path_walk_context ctx = {
+		.repo = info->revs->repo,
+		.revs = info->revs,
+		.info = info,
+		.path_stack = STRING_LIST_INIT_DUP,
+		.paths_to_lists = STRMAP_INIT
+	};
+	struct oidset root_tree_set = OIDSET_INIT;
+
+	trace2_region_enter("path-walk", "commit-walk", info->revs->repo);
+
+	CALLOC_ARRAY(commit_list, 1);
+	commit_list->type = OBJ_COMMIT;
+
+	if (info->tags)
+		info->revs->tag_objects = 1;
+
+	/* Insert a single list for the root tree into the paths. */
+	CALLOC_ARRAY(root_tree_list, 1);
+	root_tree_list->type = OBJ_TREE;
+	root_tree_list->maybe_interesting = 1;
+	strmap_put(&ctx.paths_to_lists, root_path, root_tree_list);
+
+	/*
+	 * Set these values before preparing the walk to catch
+	 * lightweight tags pointing to non-commits.
+	 */
+	info->revs->blob_objects = info->blobs;
+	info->revs->tree_objects = info->trees;
+
+	if (prepare_revision_walk(info->revs))
+		die(_("failed to setup revision walk"));
+
+	info->revs->blob_objects = info->revs->tree_objects = 0;
+
+	if (info->tags) {
+		struct oid_array tagged_blob_list = OID_ARRAY_INIT;
+		struct oid_array tags = OID_ARRAY_INIT;
+
+		trace2_region_enter("path-walk", "tag-walk", info->revs->repo);
+
+		/*
+		 * Walk any pending objects at this point, but they should only
+		 * be tags.
+		 */
+		for (size_t i = 0; i < info->revs->pending.nr; i++) {
+			struct object_array_entry *pending = info->revs->pending.objects + i;
+			struct object *obj = pending->item;
+
+			if (obj->type == OBJ_COMMIT || obj->flags & SEEN)
+				continue;
+
+			while (obj->type == OBJ_TAG) {
+				struct tag *tag = lookup_tag(info->revs->repo,
+							     &obj->oid);
+				if (!(obj->flags & SEEN)) {
+					obj->flags |= SEEN;
+					oid_array_append(&tags, &obj->oid);
+				}
+				obj = tag->tagged;
+			}
+
+			if ((obj->flags & SEEN))
+				continue;
+			obj->flags |= SEEN;
+
+			switch (obj->type) {
+			case OBJ_TREE:
+				if (info->trees)
+					oid_array_append(&root_tree_list->oids, &obj->oid);
+				break;
+
+			case OBJ_BLOB:
+				if (info->blobs)
+					oid_array_append(&tagged_blob_list, &obj->oid);
+				break;
+
+			case OBJ_COMMIT:
+				/* Make sure it is in the object walk */
+				add_pending_object(info->revs, obj, "");
+				break;
+
+			default:
+				BUG("should not see any other type here");
+			}
+		}
+
+		info->path_fn("", &tags, OBJ_TAG, info->path_fn_data);
+
+		if (tagged_blob_list.nr && info->blobs)
+			info->path_fn("/tagged-blobs", &tagged_blob_list, OBJ_BLOB,
+				      info->path_fn_data);
+
+		trace2_data_intmax("path-walk", ctx.repo, "tags", tags.nr);
+		trace2_region_leave("path-walk", "tag-walk", info->revs->repo);
+		oid_array_clear(&tags);
+		oid_array_clear(&tagged_blob_list);
+	}
+
+	while ((c = get_revision(info->revs))) {
+		struct object_id *oid;
+		struct tree *t;
+		commits_nr++;
+
+		if (info->commits)
+			oid_array_append(&commit_list->oids,
+					 &c->object.oid);
+
+		/* If we only care about commits, then skip trees. */
+		if (!info->trees && !info->blobs)
+			continue;
+
+		oid = get_commit_tree_oid(c);
+		t = lookup_tree(info->revs->repo, oid);
+
+		if (t) {
+			if (t->object.flags & SEEN)
+				continue;
+			t->object.flags |= SEEN;
+
+			if (!oidset_insert(&root_tree_set, oid))
+				oid_array_append(&root_tree_list->oids, oid);
+		} else {
+			warning("could not find tree %s", oid_to_hex(oid));
+		}
+
+		if (t && (c->object.flags & UNINTERESTING)) {
+			t->object.flags |= UNINTERESTING;
+			has_uninteresting = 1;
+		}
+	}
+
+	trace2_data_intmax("path-walk", ctx.repo, "commits", commits_nr);
+	trace2_region_leave("path-walk", "commit-walk", info->revs->repo);
+
+	/* Track all commits. */
+	if (info->commits)
+		ret = info->path_fn("", &commit_list->oids, OBJ_COMMIT,
+				    info->path_fn_data);
+	oid_array_clear(&commit_list->oids);
+	free(commit_list);
+
+	/*
+	 * Before performing a DFS of our paths and emitting them as interesting,
+	 * do a full walk of the trees to distribute the UNINTERESTING bit. Use
+	 * the sparse algorithm if prune_all_uninteresting was set.
+	 */
+	if (has_uninteresting) {
+		trace2_region_enter("path-walk", "uninteresting-walk", info->revs->repo);
+		if (info->prune_all_uninteresting)
+			mark_trees_uninteresting_sparse(ctx.repo, &root_tree_set);
+		else
+			mark_trees_uninteresting_dense(ctx.repo, &root_tree_set);
+		trace2_region_leave("path-walk", "uninteresting-walk", info->revs->repo);
+	}
+	oidset_clear(&root_tree_set);
+
+	string_list_append(&ctx.path_stack, root_path);
+
+	trace2_region_enter("path-walk", "path-walk", info->revs->repo);
+	while (!ret && ctx.path_stack.nr) {
+		char *path = ctx.path_stack.items[ctx.path_stack.nr - 1].string;
+		ctx.path_stack.nr--;
+		paths_nr++;
+
+		ret = walk_path(&ctx, path);
+
+		free(path);
+	}
+	trace2_data_intmax("path-walk", ctx.repo, "paths", paths_nr);
+	trace2_region_leave("path-walk", "path-walk", info->revs->repo);
+
+	clear_strmap(&ctx.paths_to_lists);
+	string_list_clear(&ctx.path_stack, 0);
+	return ret;
+}
diff --git a/path-walk.h b/path-walk.h
new file mode 100644
index 00000000000000..090cda3b5cf8f4
--- /dev/null
+++ b/path-walk.h
@@ -0,0 +1,75 @@
+/*
+ * path-walk.h : Methods and structures for walking the object graph in batches
+ * by the paths that can reach those objects.
+ */
+#include "object.h" /* Required for 'enum object_type'. */
+
+struct rev_info;
+struct oid_array;
+struct pattern_list;
+
+/**
+ * The type of a function pointer for the method that is called on a list of
+ * objects reachable at a given path.
+ */
+typedef int (*path_fn)(const char *path,
+		       struct oid_array *oids,
+		       enum object_type type,
+		       void *data);
+
+struct path_walk_info {
+	/**
+	 * revs provides the definitions for the commit walk, including
+	 * which commits are UNINTERESTING or not.
+	 */
+	struct rev_info *revs;
+
+	/**
+	 * The caller wishes to execute custom logic on objects reachable at a
+	 * given path. Every reachable object will be visited exactly once, and
+	 * the first path to see an object wins. This may not be a stable choice.
+	 */
+	path_fn path_fn;
+	void *path_fn_data;
+	/**
+	 * Initialize which object types the path_fn should be called on. This
+	 * could also limit the walk to skip blobs if not set.
+	 */
+	int commits;
+	int trees;
+	int blobs;
+	int tags;
+
+	/**
+	 * When 'prune_all_uninteresting' is set and a path has all objects
+	 * marked as UNINTERESTING, then the path-walk will not visit those
+	 * objects. It will not call path_fn on those objects and will not
+	 * walk the children of such trees.
+	 */
+	int prune_all_uninteresting;
+
+	/**
+	 * Specify a sparse-checkout definition to match our paths to. Do not
+	 * walk outside of this sparse definition. If the patterns are in
+	 * cone mode, then the search may prune directories that are outside
+	 * of the cone. If not in cone mode, then all tree paths will be
+	 * explored but the path_fn will only be called when the path matches
+	 * the sparse-checkout patterns.
+	 */
+	struct pattern_list *pl;
+};
+
+#define PATH_WALK_INFO_INIT {   \
+	.blobs = 1,		\
+	.trees = 1,		\
+	.commits = 1,		\
+	.tags = 1,		\
+}
+
+/**
+ * Given the configuration of 'info', walk the commits based on 'info->revs' and
+ * call 'info->path_fn' on each discovered path.
+ *
+ * Returns nonzero on an error.
+ */
+int walk_objects_by_path(struct path_walk_info *info);
diff --git a/path.c b/path.c
index 07964f5d32c977..6256f53b556050 100644
--- a/path.c
+++ b/path.c
@@ -1226,6 +1226,45 @@ char *strip_path_suffix(const char *path, const char *suffix)
 	return offset == -1 ? NULL : xstrndup(path, offset);
 }
 
+int is_mount_point_via_stat(struct strbuf *path)
+{
+	size_t len = path->len;
+	dev_t current_dev;
+	struct stat st;
+
+	if (!strcmp("/", path->buf))
+		return 1;
+
+	strbuf_addstr(path, "/.");
+	if (lstat(path->buf, &st)) {
+		/*
+		 * If we cannot access the current directory, we cannot say
+		 * that it is a bind mount.
+		 */
+		strbuf_setlen(path, len);
+		return 0;
+	}
+	current_dev = st.st_dev;
+
+	/* Now look at the parent directory */
+	strbuf_addch(path, '.');
+	if (lstat(path->buf, &st)) {
+		/*
+		 * If we cannot access the parent directory, we cannot say
+		 * that it is a bind mount.
+		 */
+		strbuf_setlen(path, len);
+		return 0;
+	}
+	strbuf_setlen(path, len);
+
+	/*
+	 * If the device ID differs between current and parent directory,
+	 * then it is a bind mount.
+	 */
+	return current_dev != st.st_dev;
+}
+
 int daemon_avoid_alias(const char *p)
 {
 	int sl, ndot;
@@ -1443,6 +1482,7 @@ int looks_like_command_line_option(const char *str)
 char *xdg_config_home_for(const char *subdir, const char *filename)
 {
 	const char *home, *config_home;
+	char *home_config = NULL;
 
 	assert(subdir);
 	assert(filename);
@@ -1451,10 +1491,26 @@ char *xdg_config_home_for(const char *subdir, const char *filename)
 		return mkpathdup("%s/%s/%s", config_home, subdir, filename);
 
 	home = getenv("HOME");
-	if (home)
-		return mkpathdup("%s/.config/%s/%s", home, subdir, filename);
+	if (home && *home)
+		home_config = mkpathdup("%s/.config/%s/%s", home, subdir, filename);
+
+	#ifdef WIN32
+	{
+		const char *appdata = getenv("APPDATA");
+		if (appdata && *appdata) {
+			char *appdata_config = mkpathdup("%s/Git/%s", appdata, filename);
+			if (file_exists(appdata_config)) {
+				if (home_config && file_exists(home_config))
+					warning("'%s' was ignored because '%s' exists.", home_config, appdata_config);
+				free(home_config);
+				return appdata_config;
+			}
+			free(appdata_config);
+		}
+	}
+	#endif
 
-	return NULL;
+	return home_config;
 }
 
 char *xdg_config_home(const char *filename)
diff --git a/path.h b/path.h
index 5f6c85e5f8d795..9b0b5d2718e5f8 100644
--- a/path.h
+++ b/path.h
@@ -183,6 +183,7 @@ int normalize_path_copy(char *dst, const char *src);
 int strbuf_normalize_path(struct strbuf *src);
 int longest_ancestor_length(const char *path, struct string_list *prefixes);
 char *strip_path_suffix(const char *path, const char *suffix);
+int is_mount_point_via_stat(struct strbuf *path);
 int daemon_avoid_alias(const char *path);
 
 /*
diff --git a/prompt.c b/prompt.c
index f21c5bf1c7e42f..a07b135050b9c2 100644
--- a/prompt.c
+++ b/prompt.c
@@ -80,7 +80,7 @@ int git_read_line_interactively(struct strbuf *line)
 	int ret;
 
 	fflush(stdout);
-	ret = strbuf_getline_lf(line, stdin);
+	ret = strbuf_getline(line, stdin);
 	if (ret != EOF)
 		strbuf_trim_trailing_newline(line);
 
diff --git a/repo-settings.c b/repo-settings.c
index 9d16d5399e3751..633b154c20ab57 100644
--- a/repo-settings.c
+++ b/repo-settings.c
@@ -47,11 +47,13 @@ void prepare_repo_settings(struct repository *r)
 		r->settings.fetch_negotiation_algorithm = FETCH_NEGOTIATION_SKIPPING;
 		r->settings.pack_use_bitmap_boundary_traversal = 1;
 		r->settings.pack_use_multi_pack_reuse = 1;
+		r->settings.pack_use_path_walk = 1;
 	}
 	if (manyfiles) {
 		r->settings.index_version = 4;
 		r->settings.index_skip_hash = 1;
 		r->settings.core_untracked_cache = UNTRACKED_CACHE_WRITE;
+		r->settings.pack_use_path_walk = 1;
 	}
 
 	/* Commit graph config or default, does not cascade (simple) */
@@ -66,6 +68,7 @@ void prepare_repo_settings(struct repository *r)
 
 	/* Boolean config or default, does not cascade (simple)  */
 	repo_cfg_bool(r, "pack.usesparse", &r->settings.pack_use_sparse, 1);
+	repo_cfg_bool(r, "pack.usepathwalk", &r->settings.pack_use_path_walk, 0);
 	repo_cfg_bool(r, "core.multipackindex", &r->settings.core_multi_pack_index, 1);
 	repo_cfg_bool(r, "index.sparse", &r->settings.sparse_index, 0);
 	repo_cfg_bool(r, "index.skiphash", &r->settings.index_skip_hash, r->settings.index_skip_hash);
diff --git a/repo-settings.h b/repo-settings.h
index 93ea0c32741125..7e8869ac4efcb8 100644
--- a/repo-settings.h
+++ b/repo-settings.h
@@ -53,6 +53,7 @@ struct repo_settings {
 	enum untracked_cache_setting core_untracked_cache;
 
 	int pack_use_sparse;
+	int pack_use_path_walk;
 	enum fetch_negotiation_setting fetch_negotiation_algorithm;
 
 	int core_multi_pack_index;
diff --git a/repository.c b/repository.c
index 1a6a62bbd03a5d..468fe580a5d6b6 100644
--- a/repository.c
+++ b/repository.c
@@ -136,7 +136,7 @@ static void repo_set_commondir(struct repository *repo,
 {
 	struct strbuf sb = STRBUF_INIT;
 
-	free(repo->commondir);
+	FREE_AND_NULL(repo->commondir);
 
 	if (commondir) {
 		repo->different_commondir = 1;
diff --git a/revision.c b/revision.c
index 474fa1e767c8d8..32d949b36feb80 100644
--- a/revision.c
+++ b/revision.c
@@ -220,6 +220,21 @@ static void add_children_by_path(struct repository *r,
 	free_tree_buffer(tree);
 }
 
+void mark_trees_uninteresting_dense(struct repository *r,
+				    struct oidset *trees)
+{
+	struct object_id *oid;
+	struct oidset_iter iter;
+
+	oidset_iter_init(trees, &iter);
+	while ((oid = oidset_iter_next(&iter))) {
+		struct tree *tree = lookup_tree(r, oid);
+
+		if (tree->object.flags & UNINTERESTING)
+			mark_tree_contents_uninteresting(r, tree);
+	}
+}
+
 void mark_trees_uninteresting_sparse(struct repository *r,
 				     struct oidset *trees)
 {
diff --git a/revision.h b/revision.h
index 71e984c452b8d7..8938b2db112e2a 100644
--- a/revision.h
+++ b/revision.h
@@ -487,6 +487,7 @@ void put_revision_mark(const struct rev_info *revs,
 
 void mark_parents_uninteresting(struct rev_info *revs, struct commit *commit);
 void mark_tree_uninteresting(struct repository *r, struct tree *tree);
+void mark_trees_uninteresting_dense(struct repository *r, struct oidset *trees);
 void mark_trees_uninteresting_sparse(struct repository *r, struct oidset *trees);
 
 void show_object_with_name(FILE *, struct object *, const char *);
diff --git a/run-command.c b/run-command.c
index 402138b8b53ae8..ca546563f985c3 100644
--- a/run-command.c
+++ b/run-command.c
@@ -576,6 +576,7 @@ static int wait_or_whine(pid_t pid, const char *argv0, int in_signal)
 		 */
 		code += 128;
 	} else if (WIFEXITED(status)) {
+		warn_about_git_lfs_on_windows7(status, argv0);
 		code = WEXITSTATUS(status);
 	} else {
 		if (!in_signal)
diff --git a/scalar.c b/scalar.c
index f24bcd0169cee4..22932ad9d6da4e 100644
--- a/scalar.c
+++ b/scalar.c
@@ -170,6 +170,7 @@ static int set_recommended_config(int reconfigure)
 		{ "core.autoCRLF", "false" },
 		{ "core.safeCRLF", "false" },
 		{ "fetch.showForcedUpdates", "false" },
+		{ "pack.usePathWalk", "true" },
 		{ NULL, NULL },
 	};
 	int i;
diff --git a/send-pack.c b/send-pack.c
index 7e8321368379ef..0d39ee7859d067 100644
--- a/send-pack.c
+++ b/send-pack.c
@@ -494,7 +494,7 @@ int send_pack(struct send_pack_args *args,
 	int need_pack_data = 0;
 	int allow_deleting_refs = 0;
 	int status_report = 0;
-	int use_sideband = 0;
+	int use_sideband = 1;
 	int quiet_supported = 0;
 	int agent_supported = 0;
 	int advertise_sid = 0;
@@ -518,6 +518,7 @@ int send_pack(struct send_pack_args *args,
 		goto out;
 	}
 
+	git_config_get_bool("sendpack.sideband", &use_sideband);
 	git_config_get_bool("push.negotiate", &push_negotiate);
 	if (push_negotiate) {
 		trace2_region_enter("send_pack", "push_negotiate", the_repository);
@@ -539,8 +540,7 @@ int send_pack(struct send_pack_args *args,
 		allow_deleting_refs = 1;
 	if (server_supports("ofs-delta"))
 		args->use_ofs_delta = 1;
-	if (server_supports("side-band-64k"))
-		use_sideband = 1;
+	use_sideband = use_sideband && server_supports("side-band-64k");
 	if (server_supports("quiet"))
 		quiet_supported = 1;
 	if (server_supports("agent"))
diff --git a/setup.c b/setup.c
index 8a488f3e7c74b1..b33b15d20d62b7 100644
--- a/setup.c
+++ b/setup.c
@@ -1781,10 +1781,19 @@ const char *setup_git_directory_gently(int *nongit_ok)
 		break;
 	case GIT_DIR_INVALID_OWNERSHIP:
 		if (!nongit_ok) {
+			struct strbuf prequoted = STRBUF_INIT;
 			struct strbuf quoted = STRBUF_INIT;
 
 			strbuf_complete(&report, '\n');
-			sq_quote_buf_pretty(&quoted, dir.buf);
+
+#ifdef __MINGW32__
+			if (dir.buf[0] == '/')
+				strbuf_addstr(&prequoted, "%(prefix)/");
+#endif
+
+			strbuf_add(&prequoted, dir.buf, dir.len);
+			sq_quote_buf_pretty(&quoted, prequoted.buf);
+
 			die(_("detected dubious ownership in repository at '%s'\n"
 			      "%s"
 			      "To add an exception for this directory, call:\n"
@@ -2586,7 +2595,7 @@ int init_db(const char *git_dir, const char *real_git_dir,
 	 * have set up the repository format such that we can evaluate
 	 * includeIf conditions correctly in the case of re-initialization.
 	 */
-	git_config(platform_core_config, NULL);
+	git_config(git_default_core_config, NULL);
 
 	safe_create_dir(git_dir, 0);
 
diff --git a/sha1dc_git.c b/sha1dc_git.c
index 9b675a046ee699..fe58d7962a30c9 100644
--- a/sha1dc_git.c
+++ b/sha1dc_git.c
@@ -27,10 +27,9 @@ void git_SHA1DCFinal(unsigned char hash[20], SHA1_CTX *ctx)
 /*
  * Same as SHA1DCUpdate, but adjust types to match git's usual interface.
  */
-void git_SHA1DCUpdate(SHA1_CTX *ctx, const void *vdata, unsigned long len)
+void git_SHA1DCUpdate(SHA1_CTX *ctx, const void *vdata, size_t len)
 {
 	const char *data = vdata;
-	/* We expect an unsigned long, but sha1dc only takes an int */
 	while (len > INT_MAX) {
 		SHA1DCUpdate(ctx, data, INT_MAX);
 		data += INT_MAX;
diff --git a/sha1dc_git.h b/sha1dc_git.h
index f6f880cabea382..0bcf1aa84b7241 100644
--- a/sha1dc_git.h
+++ b/sha1dc_git.h
@@ -15,7 +15,7 @@ void git_SHA1DCInit(SHA1_CTX *);
 #endif
 
 void git_SHA1DCFinal(unsigned char [20], SHA1_CTX *);
-void git_SHA1DCUpdate(SHA1_CTX *ctx, const void *data, unsigned long len);
+void git_SHA1DCUpdate(SHA1_CTX *ctx, const void *data, size_t len);
 
 #define platform_SHA_IS_SHA1DC /* used by "test-tool sha1-is-sha1dc" */
 
diff --git a/t/README b/t/README
index e84824dc002932..17f16c2af409e2 100644
--- a/t/README
+++ b/t/README
@@ -415,6 +415,10 @@ GIT_TEST_PACK_SPARSE=<boolean> if disabled will default the pack-objects
 builtin to use the non-sparse object walk. This can still be overridden by
 the --sparse command-line argument.
 
+GIT_TEST_PACK_PATH_WALK=<boolean> if enabled will default the pack-objects
+builtin to use the path-walk API for the object walk. This can still be
+overridden by the --no-path-walk command-line argument.
+
 GIT_TEST_PRELOAD_INDEX=<boolean> exercises the preload-index code path
 by overriding the minimum number of cache entries required per thread.
 
@@ -471,6 +475,10 @@ a test and then fails then the whole test run will abort. This can help to make
 sure the expected tests are executed and not silently skipped when their
 dependency breaks or is simply not present in a new environment.
 
+GIT_TEST_FULL_NAME_HASH=<boolean>, when true, sets the default name-hash
+function in 'git pack-objects' to be the one used by the --full-name-hash
+option.
+
 Naming Tests
 ------------
 
diff --git a/t/helper/meson.build b/t/helper/meson.build
index 5e83884246edc7..1d6154ce9756db 100644
--- a/t/helper/meson.build
+++ b/t/helper/meson.build
@@ -34,12 +34,14 @@ test_tool_sources = [
   'test-match-trees.c',
   'test-mergesort.c',
   'test-mktemp.c',
+  'test-name-hash.c',
   'test-online-cpus.c',
   'test-pack-mtimes.c',
   'test-parse-options.c',
   'test-parse-pathspec-file.c',
   'test-partial-clone.c',
   'test-path-utils.c',
+  'test-path-walk.c',
   'test-pcre2-config.c',
   'test-pkt-line.c',
   'test-proc-receive.c',
diff --git a/t/helper/test-name-hash.c b/t/helper/test-name-hash.c
new file mode 100644
index 00000000000000..e4ecd159b76400
--- /dev/null
+++ b/t/helper/test-name-hash.c
@@ -0,0 +1,24 @@
+/*
+ * test-name-hash.c: Read a list of paths over stdin and report on their
+ * name-hash and full name-hash.
+ */
+
+#include "test-tool.h"
+#include "git-compat-util.h"
+#include "pack-objects.h"
+#include "strbuf.h"
+
+int cmd__name_hash(int argc UNUSED, const char **argv UNUSED)
+{
+	struct strbuf line = STRBUF_INIT;
+
+	while (!strbuf_getline(&line, stdin)) {
+		uint32_t name_hash = pack_name_hash(line.buf);
+		uint32_t full_hash = pack_full_name_hash(line.buf);
+
+		printf("%10"PRIu32"\t%10"PRIu32"\t%s\n", name_hash, full_hash, line.buf);
+	}
+
+	strbuf_release(&line);
+	return 0;
+}
diff --git a/t/helper/test-path-walk.c b/t/helper/test-path-walk.c
new file mode 100644
index 00000000000000..405c0f43be1694
--- /dev/null
+++ b/t/helper/test-path-walk.c
@@ -0,0 +1,133 @@
+#define USE_THE_REPOSITORY_VARIABLE
+
+#include "test-tool.h"
+#include "dir.h"
+#include "environment.h"
+#include "hex.h"
+#include "object-name.h"
+#include "object.h"
+#include "pretty.h"
+#include "revision.h"
+#include "setup.h"
+#include "parse-options.h"
+#include "strbuf.h"
+#include "path-walk.h"
+#include "oid-array.h"
+
+static const char * const path_walk_usage[] = {
+	N_("test-tool path-walk <options> -- <revision-options>"),
+	NULL
+};
+
+struct path_walk_test_data {
+	uintmax_t commit_nr;
+	uintmax_t tree_nr;
+	uintmax_t blob_nr;
+	uintmax_t tag_nr;
+};
+
+static int emit_block(const char *path, struct oid_array *oids,
+		      enum object_type type, void *data)
+{
+	struct path_walk_test_data *tdata = data;
+	const char *typestr;
+
+	switch (type) {
+	case OBJ_COMMIT:
+		typestr = "COMMIT";
+		tdata->commit_nr += oids->nr;
+		break;
+
+	case OBJ_TREE:
+		typestr = "TREE";
+		tdata->tree_nr += oids->nr;
+		break;
+
+	case OBJ_BLOB:
+		typestr = "BLOB";
+		tdata->blob_nr += oids->nr;
+		break;
+
+	case OBJ_TAG:
+		typestr = "TAG";
+		tdata->tag_nr += oids->nr;
+		break;
+
+	default:
+		BUG("we do not understand this type");
+	}
+
+	for (size_t i = 0; i < oids->nr; i++) {
+		struct object *o = lookup_unknown_object(the_repository,
+							 &oids->oid[i]);
+		printf("%s:%s:%s%s\n", typestr, path, oid_to_hex(&oids->oid[i]),
+		       o->flags & UNINTERESTING ? ":UNINTERESTING" : "");
+	}
+
+	return 0;
+}
+
+int cmd__path_walk(int argc, const char **argv)
+{
+	int res, stdin_pl = 0;
+	struct rev_info revs = REV_INFO_INIT;
+	struct path_walk_info info = PATH_WALK_INFO_INIT;
+	struct path_walk_test_data data = { 0 };
+	struct option options[] = {
+		OPT_BOOL(0, "blobs", &info.blobs,
+			 N_("toggle inclusion of blob objects")),
+		OPT_BOOL(0, "commits", &info.commits,
+			 N_("toggle inclusion of commit objects")),
+		OPT_BOOL(0, "tags", &info.tags,
+			 N_("toggle inclusion of tag objects")),
+		OPT_BOOL(0, "trees", &info.trees,
+			 N_("toggle inclusion of tree objects")),
+		OPT_BOOL(0, "prune", &info.prune_all_uninteresting,
+			 N_("toggle pruning of uninteresting paths")),
+		OPT_BOOL(0, "stdin-pl", &stdin_pl,
+			 N_("read a pattern list over stdin")),
+		OPT_END(),
+	};
+
+	setup_git_directory();
+	revs.repo = the_repository;
+
+	argc = parse_options(argc, argv, NULL,
+			     options, path_walk_usage,
+			     PARSE_OPT_KEEP_UNKNOWN_OPT | PARSE_OPT_KEEP_ARGV0);
+
+	if (argc > 1)
+		setup_revisions(argc, argv, &revs, NULL);
+	else
+		usage(path_walk_usage[0]);
+
+	info.revs = &revs;
+	info.path_fn = emit_block;
+	info.path_fn_data = &data;
+
+	if (stdin_pl) {
+		struct strbuf in = STRBUF_INIT;
+		CALLOC_ARRAY(info.pl, 1);
+
+		info.pl->use_cone_patterns = 1;
+
+		strbuf_fread(&in, 2048, stdin);
+		add_patterns_from_buffer(in.buf, in.len, "", 0, info.pl);
+		strbuf_release(&in);
+	}
+
+	res = walk_objects_by_path(&info);
+
+	printf("commits:%" PRIuMAX "\n"
+	       "trees:%" PRIuMAX "\n"
+	       "blobs:%" PRIuMAX "\n"
+	       "tags:%" PRIuMAX "\n",
+	       data.commit_nr, data.tree_nr, data.blob_nr, data.tag_nr);
+
+	if (info.pl) {
+		clear_pattern_list(info.pl);
+		free(info.pl);
+	}
+	release_revisions(&revs);
+	return res;
+}
diff --git a/t/helper/test-tool.c b/t/helper/test-tool.c
index 4a7aa993ba99e2..dd9a6d277318c1 100644
--- a/t/helper/test-tool.c
+++ b/t/helper/test-tool.c
@@ -44,6 +44,7 @@ static struct test_cmd cmds[] = {
 	{ "match-trees", cmd__match_trees },
 	{ "mergesort", cmd__mergesort },
 	{ "mktemp", cmd__mktemp },
+	{ "name-hash", cmd__name_hash },
 	{ "online-cpus", cmd__online_cpus },
 	{ "pack-mtimes", cmd__pack_mtimes },
 	{ "parse-options", cmd__parse_options },
@@ -52,6 +53,7 @@ static struct test_cmd cmds[] = {
 	{ "parse-subcommand", cmd__parse_subcommand },
 	{ "partial-clone", cmd__partial_clone },
 	{ "path-utils", cmd__path_utils },
+	{ "path-walk", cmd__path_walk },
 	{ "pcre2-config", cmd__pcre2_config },
 	{ "pkt-line", cmd__pkt_line },
 	{ "proc-receive", cmd__proc_receive },
diff --git a/t/helper/test-tool.h b/t/helper/test-tool.h
index 21802ac27da37f..2dc5ec515feea9 100644
--- a/t/helper/test-tool.h
+++ b/t/helper/test-tool.h
@@ -37,6 +37,7 @@ int cmd__lazy_init_name_hash(int argc, const char **argv);
 int cmd__match_trees(int argc, const char **argv);
 int cmd__mergesort(int argc, const char **argv);
 int cmd__mktemp(int argc, const char **argv);
+int cmd__name_hash(int argc, const char **argv);
 int cmd__online_cpus(int argc, const char **argv);
 int cmd__pack_mtimes(int argc, const char **argv);
 int cmd__parse_options(int argc, const char **argv);
@@ -45,6 +46,7 @@ int cmd__parse_pathspec_file(int argc, const char** argv);
 int cmd__parse_subcommand(int argc, const char **argv);
 int cmd__partial_clone(int argc, const char **argv);
 int cmd__path_utils(int argc, const char **argv);
+int cmd__path_walk(int argc, const char **argv);
 int cmd__pcre2_config(int argc, const char **argv);
 int cmd__pkt_line(int argc, const char **argv);
 int cmd__proc_receive(int argc, const char **argv);
diff --git a/t/meson.build b/t/meson.build
index 602ebfe6a26021..95282e2d894c53 100644
--- a/t/meson.build
+++ b/t/meson.build
@@ -721,6 +721,7 @@ integration_tests = [
   't5617-clone-submodules-remote.sh',
   't5618-alternate-refs.sh',
   't5619-clone-local-ambiguous-transport.sh',
+  't5620-backfill.sh',
   't5700-protocol-v1.sh',
   't5701-git-serve.sh',
   't5702-protocol-v2.sh',
@@ -829,6 +830,7 @@ integration_tests = [
   't6500-gc.sh',
   't6501-freshen-objects.sh',
   't6600-test-reach.sh',
+  't6601-path-walk.sh',
   't6700-tree-depth.sh',
   't7001-mv.sh',
   't7002-mv-sparse-checkout.sh',
@@ -955,6 +957,7 @@ integration_tests = [
   't8012-blame-colors.sh',
   't8013-blame-ignore-revs.sh',
   't8014-blame-ignore-fuzzy.sh',
+  't8100-git-survey.sh',
   't9001-send-email.sh',
   't9002-column.sh',
   't9003-help-autocorrect.sh',
diff --git a/t/perf/p5313-pack-objects.sh b/t/perf/p5313-pack-objects.sh
new file mode 100755
index 00000000000000..2e10895621ce75
--- /dev/null
+++ b/t/perf/p5313-pack-objects.sh
@@ -0,0 +1,98 @@
+#!/bin/sh
+
+test_description='Tests pack performance using bitmaps'
+. ./perf-lib.sh
+
+GIT_TEST_PASSING_SANITIZE_LEAK=0
+export GIT_TEST_PASSING_SANITIZE_LEAK
+
+test_perf_large_repo
+
+test_expect_success 'create rev input' '
+	cat >in-thin <<-EOF &&
+	$(git rev-parse HEAD)
+	^$(git rev-parse HEAD~1)
+	EOF
+
+	cat >in-big <<-EOF
+	$(git rev-parse HEAD)
+	^$(git rev-parse HEAD~1000)
+	EOF
+'
+
+test_perf 'thin pack' '
+	git pack-objects --thin --stdout --revs --sparse  <in-thin >out
+'
+
+test_size 'thin pack size' '
+	test_file_size out
+'
+
+test_perf 'thin pack with --full-name-hash' '
+	git pack-objects --thin --stdout --revs --sparse --full-name-hash <in-thin >out
+'
+
+test_size 'thin pack size with --full-name-hash' '
+	test_file_size out
+'
+
+test_perf 'thin pack with --path-walk' '
+	git pack-objects --thin --stdout --revs --sparse --path-walk <in-thin >out
+'
+
+test_size 'thin pack size with --path-walk' '
+	test_file_size out
+'
+
+test_perf 'big pack' '
+	git pack-objects --stdout --revs --sparse  <in-big >out
+'
+
+test_size 'big pack size' '
+	test_file_size out
+'
+
+test_perf 'big pack with --full-name-hash' '
+	git pack-objects --stdout --revs --sparse --full-name-hash <in-big >out
+'
+
+test_size 'big pack size with --full-name-hash' '
+	test_file_size out
+'
+
+test_perf 'big pack with --path-walk' '
+	git pack-objects --stdout --revs --sparse --path-walk <in-big >out
+'
+
+test_size 'big pack size with --path-walk' '
+	test_file_size out
+'
+
+test_perf 'repack' '
+	git repack -adf
+'
+
+test_size 'repack size' '
+	pack=$(ls .git/objects/pack/pack-*.pack) &&
+	test_file_size "$pack"
+'
+
+test_perf 'repack with --full-name-hash' '
+	git repack -adf --full-name-hash
+'
+
+test_size 'repack size with --full-name-hash' '
+	pack=$(ls .git/objects/pack/pack-*.pack) &&
+	test_file_size "$pack"
+'
+
+test_perf 'repack with --path-walk' '
+	git repack -adf --path-walk
+'
+
+test_size 'repack size with --path-walk' '
+	pack=$(ls .git/objects/pack/pack-*.pack) &&
+	test_file_size "$pack"
+'
+
+test_done
diff --git a/t/perf/p5314-name-hash.sh b/t/perf/p5314-name-hash.sh
new file mode 100755
index 00000000000000..9fe26612facc1d
--- /dev/null
+++ b/t/perf/p5314-name-hash.sh
@@ -0,0 +1,41 @@
+#!/bin/sh
+
+test_description='Tests pack performance using bitmaps'
+. ./perf-lib.sh
+
+GIT_TEST_PASSING_SANITIZE_LEAK=0
+export GIT_TEST_PASSING_SANITIZE_LEAK
+
+test_perf_large_repo
+
+test_size 'paths at head' '
+	git ls-tree -r --name-only HEAD >path-list &&
+	wc -l <path-list
+'
+
+test_size 'number of distinct name-hashes' '
+	cat path-list | test-tool name-hash >name-hashes &&
+	cat name-hashes | awk "{ print \$1; }" | sort -n | uniq -c >name-hash-count &&
+	wc -l <name-hash-count
+'
+
+test_size 'number of distinct full-name-hashes' '
+	cat name-hashes | awk "{ print \$2; }" | sort -n | uniq -c >full-name-hash-count &&
+	wc -l <full-name-hash-count
+'
+
+test_size 'maximum multiplicity of name-hashes' '
+	cat name-hash-count | \
+		sort -nr | \
+		head -n 1 | \
+		awk "{ print \$1; }"
+'
+
+test_size 'maximum multiplicity of fullname-hashes' '
+	cat full-name-hash-count | \
+		sort -nr | \
+		head -n 1 | \
+		awk "{ print \$1; }"
+'
+
+test_done
diff --git a/t/t0014-alias.sh b/t/t0014-alias.sh
index 854d59ec58c25a..30708146887d19 100755
--- a/t/t0014-alias.sh
+++ b/t/t0014-alias.sh
@@ -38,10 +38,10 @@ test_expect_success 'looping aliases - internal execution' '
 #'
 
 test_expect_success 'run-command formats empty args properly' '
-    test_must_fail env GIT_TRACE=1 git frotz a "" b " " c 2>actual.raw &&
-    sed -ne "/run_command:/s/.*trace: run_command: //p" actual.raw >actual &&
-    echo "git-frotz a '\'''\'' b '\'' '\'' c" >expect &&
-    test_cmp expect actual
+	test_must_fail env GIT_TRACE=1 git frotz a "" b " " c 2>actual.raw &&
+	sed -ne "/run_command:/s/.*trace: run_command: //p" actual.raw >actual &&
+	echo "git-frotz a '\'''\'' b '\'' '\'' c" >expect &&
+	test_cmp expect actual
 '
 
 test_expect_success 'tracing a shell alias with arguments shows trace of prepared command' '
diff --git a/t/t0060-path-utils.sh b/t/t0060-path-utils.sh
index dbb2e73bcd912a..1fc51c5b46736e 100755
--- a/t/t0060-path-utils.sh
+++ b/t/t0060-path-utils.sh
@@ -281,6 +281,14 @@ test_expect_success SYMLINKS 'real path works on symlinks' '
 	test_cmp expect actual
 '
 
+test_expect_success MINGW 'real path works near drive root' '
+	# we need a non-existing path at the drive root; simply skip if C:/xyz exists
+	if test ! -e C:/xyz
+	then
+		test C:/xyz = $(test-tool path-utils real_path C:/xyz)
+	fi
+'
+
 test_expect_success SYMLINKS 'prefix_path works with absolute paths to work tree symlinks' '
 	ln -s target symlink &&
 	echo "symlink" >expect &&
@@ -598,7 +606,8 @@ test_expect_success !VALGRIND,RUNTIME_PREFIX,CAN_EXEC_IN_PWD 'RUNTIME_PREFIX wor
 	cp "$GIT_EXEC_PATH"/git$X pretend/bin/ &&
 	GIT_EXEC_PATH= ./pretend/bin/git here >actual &&
 	echo HERE >expect &&
-	test_cmp expect actual'
+	test_cmp expect actual
+'
 
 test_expect_success !VALGRIND,RUNTIME_PREFIX,CAN_EXEC_IN_PWD '%(prefix)/ works' '
 	mkdir -p pretend/bin &&
@@ -609,4 +618,32 @@ test_expect_success !VALGRIND,RUNTIME_PREFIX,CAN_EXEC_IN_PWD '%(prefix)/ works'
 	test_cmp expect actual
 '
 
+test_expect_success MINGW 'MSYSTEM/PATH is adjusted if necessary' '
+	mkdir -p "$HOME"/bin pretend/mingw64/bin \
+		pretend/mingw64/libexec/git-core pretend/usr/bin &&
+	cp "$GIT_EXEC_PATH"/git.exe pretend/mingw64/bin/ &&
+	cp "$GIT_EXEC_PATH"/git.exe pretend/mingw64/libexec/git-core/ &&
+	# copy the .dll files, if any (happens when building via CMake)
+	case "$GIT_EXEC_PATH"/*.dll in
+	*/"*.dll") ;; # no `.dll` files to be copied
+	*)
+		cp "$GIT_EXEC_PATH"/*.dll pretend/mingw64/bin/ &&
+		cp "$GIT_EXEC_PATH"/*.dll pretend/mingw64/libexec/git-core/
+		;;
+	esac &&
+	echo "env | grep MSYSTEM=" | write_script "$HOME"/bin/git-test-home &&
+	echo "echo mingw64" | write_script pretend/mingw64/bin/git-test-bin &&
+	echo "echo usr" | write_script pretend/usr/bin/git-test-bin2 &&
+
+	(
+		MSYSTEM= &&
+		GIT_EXEC_PATH= &&
+		pretend/mingw64/libexec/git-core/git.exe test-home >actual &&
+		pretend/mingw64/libexec/git-core/git.exe test-bin >>actual &&
+		pretend/mingw64/bin/git.exe test-bin2 >>actual
+	) &&
+	test_write_lines MSYSTEM=$MSYSTEM mingw64 usr >expect &&
+	test_cmp expect actual
+'
+
 test_done
diff --git a/t/t0411-clone-from-partial.sh b/t/t0411-clone-from-partial.sh
index 196fc617843cb9..9e6bca56255be9 100755
--- a/t/t0411-clone-from-partial.sh
+++ b/t/t0411-clone-from-partial.sh
@@ -59,6 +59,12 @@ test_expect_success 'pack-objects should fetch from promisor remote and execute
 
 test_expect_success 'clone from promisor remote does not lazy-fetch by default' '
 	rm -f script-executed &&
+
+	# The --path-walk feature of "git pack-objects" is not
+	# compatible with this kind of fetch from an incomplete repo.
+	GIT_TEST_PACK_PATH_WALK=0 &&
+	export GIT_TEST_PACK_PATH_WALK &&
+
 	test_must_fail git clone evil no-lazy 2>err &&
 	test_grep "lazy fetching disabled" err &&
 	test_path_is_missing script-executed
diff --git a/t/t0450/txt-help-mismatches b/t/t0450/txt-help-mismatches
index 28003f18c924bb..c4a15fd0cb885a 100644
--- a/t/t0450/txt-help-mismatches
+++ b/t/t0450/txt-help-mismatches
@@ -45,7 +45,6 @@ rebase
 remote
 remote-ext
 remote-fd
-repack
 reset
 restore
 rev-parse
diff --git a/t/t1007-hash-object.sh b/t/t1007-hash-object.sh
index a0481139de5519..5850c27a9cc315 100755
--- a/t/t1007-hash-object.sh
+++ b/t/t1007-hash-object.sh
@@ -49,6 +49,9 @@ test_expect_success 'setup' '
 
 	example sha1:ddd3f836d3e3fbb7ae289aa9ae83536f76956399
 	example sha256:b44fe1fe65589848253737db859bd490453510719d7424daab03daf0767b85ae
+
+	large5GB sha1:0be2be10a4c8764f32c4bf372a98edc731a4b204
+	large5GB sha256:dc18ca621300c8d3cfa505a275641ebab00de189859e022a975056882d313e64
 	EOF
 '
 
@@ -265,4 +268,40 @@ test_expect_success '--stdin outside of repository (uses SHA-1)' '
 	test_cmp expect actual
 '
 
+test_expect_success EXPENSIVE,SIZE_T_IS_64BIT,!LONG_IS_64BIT \
+		'files over 4GB hash literally' '
+	test-tool genzeros $((5*1024*1024*1024)) >big &&
+	test_oid large5GB >expect &&
+	git hash-object --stdin --literally <big >actual &&
+	test_cmp expect actual
+'
+
+test_expect_success EXPENSIVE,SIZE_T_IS_64BIT,!LONG_IS_64BIT \
+		'files over 4GB hash correctly via --stdin' '
+	{ test -f big || test-tool genzeros $((5*1024*1024*1024)) >big; } &&
+	test_oid large5GB >expect &&
+	git hash-object --stdin <big >actual &&
+	test_cmp expect actual
+'
+
+test_expect_success EXPENSIVE,SIZE_T_IS_64BIT,!LONG_IS_64BIT \
+		'files over 4GB hash correctly' '
+	{ test -f big || test-tool genzeros $((5*1024*1024*1024)) >big; } &&
+	test_oid large5GB >expect &&
+	git hash-object -- big >actual &&
+	test_cmp expect actual
+'
+
+# This clean filter does nothing, other than excercising the interface.
+# We ensure that cleaning doesn't mangle large files on 64-bit Windows.
+test_expect_success EXPENSIVE,SIZE_T_IS_64BIT,!LONG_IS_64BIT \
+		'hash filtered files over 4GB correctly' '
+	{ test -f big || test-tool genzeros $((5*1024*1024*1024)) >big; } &&
+	test_oid large5GB >expect &&
+	test_config filter.null-filter.clean "cat" &&
+	echo "big filter=null-filter" >.gitattributes &&
+	git hash-object -- big >actual &&
+	test_cmp expect actual
+'
+
 test_done
diff --git a/t/t3700-add.sh b/t/t3700-add.sh
index df580a5806b4f1..10185f31f019a9 100755
--- a/t/t3700-add.sh
+++ b/t/t3700-add.sh
@@ -548,4 +548,15 @@ test_expect_success CASE_INSENSITIVE_FS 'path is case-insensitive' '
 	git add "$downcased"
 '
 
+test_expect_success MINGW 'can add files via NTFS junctions' '
+	test_when_finished "cmd //c rmdir junction && rm -rf target" &&
+	test_create_repo target &&
+	cmd //c "mklink /j junction target" &&
+	>target/via-junction &&
+	git -C junction add "$(pwd)/junction/via-junction" &&
+	echo via-junction >expect &&
+	git -C target diff --cached --name-only >actual &&
+	test_cmp expect actual
+'
+
 test_done
diff --git a/t/t3701-add-interactive.sh b/t/t3701-add-interactive.sh
index b8a05d95f3f130..81d0487b3b90eb 100755
--- a/t/t3701-add-interactive.sh
+++ b/t/t3701-add-interactive.sh
@@ -1166,6 +1166,27 @@ test_expect_success 'checkout -p patch editing of added file' '
 	)
 '
 
+test_expect_success EXPENSIVE 'add -i with a lot of files' '
+	git reset --hard &&
+	x160=0123456789012345678901234567890123456789 &&
+	x160=$x160$x160$x160$x160 &&
+	y= &&
+	i=0 &&
+	while test $i -le 200
+	do
+		name=$(printf "%s%03d" $x160 $i) &&
+		echo $name >$name &&
+		git add -N $name &&
+		y="${y}y$LF" &&
+		i=$(($i+1)) ||
+		exit 1
+	done &&
+	echo "$y" | git add -p -- . &&
+	git diff --cached >staged &&
+	test_line_count = 1407 staged &&
+	git reset --hard
+'
+
 test_expect_success 'show help from add--helper' '
 	git reset --hard &&
 	cat >expect <<-EOF &&
diff --git a/t/t5300-pack-object.sh b/t/t5300-pack-object.sh
index d1d6248558238c..1fffa7209b785b 100755
--- a/t/t5300-pack-object.sh
+++ b/t/t5300-pack-object.sh
@@ -689,4 +689,40 @@ do
 	'
 done
 
+# The following test is not necessarily a permanent choice, but since we do not
+# have a "name hash version" bit in the .bitmap file format, we cannot write the
+# full-name hash values into the .bitmap file without risking breakage later.
+#
+# TODO: Make these compatible in the future and replace this test with the
+# expected behavior when both are specified.
+test_expect_success '--full-name-hash and --write-bitmap-index are incompatible' '
+	test_must_fail git pack-objects base --all \
+		--full-name-hash --write-bitmap-index 2>err &&
+	grep incompatible err &&
+
+	# --stdout option silently removes --write-bitmap-index
+	git pack-objects --stdout --all --full-name-hash --write-bitmap-index >out
+'
+
+# Basic "repack everything" test
+test_expect_success '--path-walk pack everything' '
+	git -C server rev-parse HEAD >in &&
+	GIT_PROGRESS_DELAY=0 git -C server pack-objects \
+		--stdout --revs --path-walk --progress <in >out.pack 2>err &&
+	grep "Compressing objects by path" err &&
+	git -C server index-pack --stdin <out.pack
+'
+
+# Basic "thin pack" test
+test_expect_success '--path-walk thin pack' '
+	cat >in <<-EOF &&
+	$(git -C server rev-parse HEAD)
+	^$(git -C server rev-parse HEAD~2)
+	EOF
+	GIT_PROGRESS_DELAY=0 git -C server pack-objects \
+		--thin --stdout --revs --path-walk --progress <in >out.pack 2>err &&
+	grep "Compressing objects by path" err &&
+	git -C server index-pack --fix-thin --stdin <out.pack
+'
+
 test_done
diff --git a/t/t5306-pack-nobase.sh b/t/t5306-pack-nobase.sh
index 805d60ff3179ce..609399d54fbb5b 100755
--- a/t/t5306-pack-nobase.sh
+++ b/t/t5306-pack-nobase.sh
@@ -59,6 +59,11 @@ test_expect_success 'indirectly clone patch_clone' '
 	 git pull ../.git &&
 	 test $(git rev-parse HEAD) = $B &&
 
+	# The --path-walk feature of "git pack-objects" is not
+	# compatible with this kind of fetch from an incomplete repo.
+	GIT_TEST_PACK_PATH_WALK=0 &&
+	export GIT_TEST_PACK_PATH_WALK &&
+
 	 git pull ../patch_clone/.git &&
 	 test $(git rev-parse HEAD) = $C
 	)
diff --git a/t/t5310-pack-bitmaps.sh b/t/t5310-pack-bitmaps.sh
index eabfcd7ff67d2f..fac653c120efa9 100755
--- a/t/t5310-pack-bitmaps.sh
+++ b/t/t5310-pack-bitmaps.sh
@@ -26,6 +26,32 @@ has_any () {
 	grep -Ff "$1" "$2"
 }
 
+# Since name-hash values are stored in the .bitmap files, add a test
+# that checks that the name-hash calculations are stable across versions.
+# Not exhaustive, but these hashing algorithms would be hard to change
+# without causing deviations here.
+test_expect_success 'name-hash value stability' '
+	cat >names <<-\EOF &&
+	first
+	second
+	third
+	one-long-enough-for-collisions
+	two-long-enough-for-collisions
+	EOF
+
+	test-tool name-hash <names >out &&
+
+	cat >expect <<-\EOF &&
+	2582249472	3109209818	first
+	2289942528	3781118409	second
+	2300837888	3028707182	third
+	2544516325	3241327563	one-long-enough-for-collisions
+	2544516325	4207880830	two-long-enough-for-collisions
+	EOF
+
+	test_cmp expect out
+'
+
 test_bitmap_cases () {
 	writeLookupTable=false
 	for i in "$@"
@@ -128,8 +154,9 @@ test_bitmap_cases () {
 		ls .git/objects/pack/ | grep bitmap >output &&
 		test_line_count = 1 output &&
 		# verify equivalent packs are generated with/without using bitmap index
-		packasha1=$(git pack-objects --no-use-bitmap-index --all packa </dev/null) &&
-		packbsha1=$(git pack-objects --use-bitmap-index --all packb </dev/null) &&
+		# Be careful to not use the path-walk option in either case.
+		packasha1=$(git pack-objects --no-use-bitmap-index --no-path-walk --all packa </dev/null) &&
+		packbsha1=$(git pack-objects --use-bitmap-index --no-path-walk --all packb </dev/null) &&
 		list_packed_objects packa-$packasha1.idx >packa.objects &&
 		list_packed_objects packb-$packbsha1.idx >packb.objects &&
 		test_cmp packa.objects packb.objects
@@ -358,6 +385,14 @@ test_bitmap_cases () {
 		git init --bare client.git &&
 		(
 			cd client.git &&
+
+			# This test relies on reusing a delta, but if the
+			# path-walk machinery is engaged, the base object
+			# is considered too small to use during the
+			# dynamic computation, so is not used.
+			GIT_TEST_PACK_PATH_WALK=0 &&
+			export GIT_TEST_PACK_PATH_WALK &&
+
 			git config transfer.unpackLimit 1 &&
 			git fetch .. delta-reuse-old:delta-reuse-old &&
 			git fetch .. delta-reuse-new:delta-reuse-new &&
diff --git a/t/t5316-pack-delta-depth.sh b/t/t5316-pack-delta-depth.sh
index 32cf4227451ff7..167c3a35234576 100755
--- a/t/t5316-pack-delta-depth.sh
+++ b/t/t5316-pack-delta-depth.sh
@@ -89,15 +89,18 @@ max_chain() {
 # adjusted (or scrapped if the heuristics have become too unreliable)
 test_expect_success 'packing produces a long delta' '
 	# Use --window=0 to make sure we are seeing reused deltas,
-	# not computing a new long chain.
-	pack=$(git pack-objects --all --window=0 </dev/null pack) &&
+	# not computing a new long chain. (Also avoid the --path-walk
+	# option as it may break delta chains.)
+	pack=$(git pack-objects --all --window=0 --no-path-walk </dev/null pack) &&
 	echo 9 >expect &&
 	max_chain pack-$pack.pack >actual &&
 	test_cmp expect actual
 '
 
 test_expect_success '--depth limits depth' '
-	pack=$(git pack-objects --all --depth=5 </dev/null pack) &&
+	# Avoid --path-walk to avoid breaking delta chains across path
+	# boundaries.
+	pack=$(git pack-objects --all --depth=5 --no-path-walk </dev/null pack) &&
 	echo 5 >expect &&
 	max_chain pack-$pack.pack >actual &&
 	test_cmp expect actual
diff --git a/t/t5332-multi-pack-reuse.sh b/t/t5332-multi-pack-reuse.sh
index 57cad7708f80f9..395d09444ced72 100755
--- a/t/t5332-multi-pack-reuse.sh
+++ b/t/t5332-multi-pack-reuse.sh
@@ -7,6 +7,13 @@ test_description='pack-objects multi-pack reuse'
 
 GIT_TEST_MULTI_PACK_INDEX=0
 GIT_TEST_MULTI_PACK_INDEX_WRITE_INCREMENTAL=0
+
+# The --path-walk option does not consider the preferred pack
+# at all for reusing deltas, so this variable changes the
+# behavior of this test, if enabled.
+GIT_TEST_PACK_PATH_WALK=0
+export GIT_TEST_PACK_PATH_WALK
+
 objdir=.git/objects
 packdir=$objdir/pack
 
diff --git a/t/t5505-remote.sh b/t/t5505-remote.sh
index 519f7973e31287..86498c62b29273 100755
--- a/t/t5505-remote.sh
+++ b/t/t5505-remote.sh
@@ -941,8 +941,8 @@ test_expect_success '"remote show" does not show symbolic refs' '
 	(
 		cd three &&
 		git remote show origin >output &&
-		! grep "^ *HEAD$" < output &&
-		! grep -i stale < output
+		! grep "^ *HEAD$" <output &&
+		! grep -i stale <output
 	)
 '
 
@@ -1145,7 +1145,7 @@ test_expect_success 'migrate a remote from named file in $GIT_DIR/branches' '
 	(
 		cd six &&
 		git remote rm origin &&
-		mkdir .git/branches &&
+		mkdir -p .git/branches &&
 		echo "$origin_url#main" >.git/branches/origin &&
 		git remote rename origin origin &&
 		test_path_is_missing .git/branches/origin &&
@@ -1160,8 +1160,8 @@ test_expect_success 'migrate a remote from named file in $GIT_DIR/branches (2)'
 	(
 		cd seven &&
 		git remote rm origin &&
-		mkdir .git/branches &&
-		echo "quux#foom" > .git/branches/origin &&
+		mkdir -p .git/branches &&
+		echo "quux#foom" >.git/branches/origin &&
 		git remote rename origin origin &&
 		test_path_is_missing .git/branches/origin &&
 		test "$(git config remote.origin.url)" = "quux" &&
diff --git a/t/t5510-fetch.sh b/t/t5510-fetch.sh
index 2d9587059f5083..3ef509cd058d26 100755
--- a/t/t5510-fetch.sh
+++ b/t/t5510-fetch.sh
@@ -1240,7 +1240,12 @@ test_expect_success 'all boundary commits are excluded' '
 	test_tick &&
 	git merge otherside &&
 	ad=$(git log --no-walk --format=%ad HEAD) &&
-	git bundle create twoside-boundary.bdl main --since="$ad" &&
+
+	# If the --full-name-hash function is used here, then no delta
+	# pair is found and the bundle does not expand to three objects
+	# when fixing the thin object.
+	GIT_TEST_FULL_NAME_HASH=0 \
+		git bundle create twoside-boundary.bdl main --since="$ad" &&
 	test_bundle_object_count --thin twoside-boundary.bdl 3
 '
 
diff --git a/t/t5516-fetch-push.sh b/t/t5516-fetch-push.sh
index 041d7d806fea57..cc0fad65bfce9f 100755
--- a/t/t5516-fetch-push.sh
+++ b/t/t5516-fetch-push.sh
@@ -979,8 +979,8 @@ test_expect_success 'fetch with branches' '
 	mk_empty testrepo &&
 	git branch second $the_first_commit &&
 	git checkout second &&
-	mkdir testrepo/.git/branches &&
-	echo ".." > testrepo/.git/branches/branch1 &&
+	mkdir -p testrepo/.git/branches &&
+	echo ".." >testrepo/.git/branches/branch1 &&
 	(
 		cd testrepo &&
 		git fetch branch1 &&
@@ -993,8 +993,8 @@ test_expect_success 'fetch with branches' '
 
 test_expect_success 'fetch with branches containing #' '
 	mk_empty testrepo &&
-	mkdir testrepo/.git/branches &&
-	echo "..#second" > testrepo/.git/branches/branch2 &&
+	mkdir -p testrepo/.git/branches &&
+	echo "..#second" >testrepo/.git/branches/branch2 &&
 	(
 		cd testrepo &&
 		git fetch branch2 &&
@@ -1010,8 +1010,8 @@ test_expect_success 'push with branches' '
 	git checkout second &&
 
 	test_when_finished "rm -rf .git/branches" &&
-	mkdir .git/branches &&
-	echo "testrepo" > .git/branches/branch1 &&
+	mkdir -p .git/branches &&
+	echo "testrepo" >.git/branches/branch1 &&
 
 	git push branch1 &&
 	(
@@ -1026,8 +1026,8 @@ test_expect_success 'push with branches containing #' '
 	mk_empty testrepo &&
 
 	test_when_finished "rm -rf .git/branches" &&
-	mkdir .git/branches &&
-	echo "testrepo#branch3" > .git/branches/branch2 &&
+	mkdir -p .git/branches &&
+	echo "testrepo#branch3" >.git/branches/branch2 &&
 
 	git push branch2 &&
 	(
@@ -1557,7 +1557,7 @@ EOF
 	git init no-thin &&
 	git --git-dir=no-thin/.git config receive.unpacklimit 0 &&
 	git push no-thin/.git refs/heads/main:refs/heads/foo &&
-	echo modified >> path1 &&
+	echo modified >>path1 &&
 	git commit -am modified &&
 	git repack -adf &&
 	rcvpck="git receive-pack --reject-thin-pack-for-testing" &&
diff --git a/t/t5580-unc-paths.sh b/t/t5580-unc-paths.sh
index 65ef1a3628ee94..e9df367d5777fd 100755
--- a/t/t5580-unc-paths.sh
+++ b/t/t5580-unc-paths.sh
@@ -20,14 +20,11 @@ fi
 UNCPATH="$(winpwd)"
 case "$UNCPATH" in
 [A-Z]:*)
+	WITHOUTDRIVE="${UNCPATH#?:}"
 	# Use administrative share e.g. \\localhost\C$\git-sdk-64\usr\src\git
 	# (we use forward slashes here because MSYS2 and Git accept them, and
 	# they are easier on the eyes)
-	UNCPATH="//localhost/${UNCPATH%%:*}\$/${UNCPATH#?:}"
-	test -d "$UNCPATH" || {
-		skip_all='could not access administrative share; skipping'
-		test_done
-	}
+	UNCPATH="//localhost/${UNCPATH%%:*}\$$WITHOUTDRIVE"
 	;;
 *)
 	skip_all='skipping UNC path tests, cannot determine current path as UNC'
@@ -35,6 +32,18 @@ case "$UNCPATH" in
 	;;
 esac
 
+test_expect_success 'clone into absolute path lacking a drive prefix' '
+	USINGBACKSLASHES="$(echo "$WITHOUTDRIVE"/without-drive-prefix |
+		tr / \\\\)" &&
+	git clone . "$USINGBACKSLASHES" &&
+	test -f without-drive-prefix/.git/HEAD
+'
+
+test -d "$UNCPATH" || {
+	skip_all='could not access administrative share; skipping'
+	test_done
+}
+
 test_expect_success setup '
 	test_commit initial
 '
diff --git a/t/t5601-clone.sh b/t/t5601-clone.sh
index d0c18660e33113..079dda1367958a 100755
--- a/t/t5601-clone.sh
+++ b/t/t5601-clone.sh
@@ -7,6 +7,16 @@ export GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME
 
 . ./test-lib.sh
 
+# This test script contains test cases that need to create symbolic links. To
+# make sure that these test cases are exercised in Git for Windows, where (for
+# historical reasons) `ln -s` creates copies by default, let's specifically ask
+# for `ln -s` to create symbolic links whenever possible.
+if test_have_prereq MINGW
+then
+	MSYS=${MSYS+$MSYS }winsymlinks:nativestrict
+	export MSYS
+fi
+
 X=
 test_have_prereq !MINGW || X=.exe
 
@@ -78,6 +88,13 @@ test_expect_success 'clone respects GIT_WORK_TREE' '
 
 '
 
+test_expect_success CASE_INSENSITIVE_FS 'core.worktree is not added due to path case' '
+
+	mkdir UPPERCASE &&
+	git clone src "$(pwd)/uppercase" &&
+	test "unset" = "$(git -C UPPERCASE config --default unset core.worktree)"
+'
+
 test_expect_success 'clone from hooks' '
 
 	test_create_repo r0 &&
diff --git a/t/t5616-partial-clone.sh b/t/t5616-partial-clone.sh
index 4650451964339d..20d946cc16ac22 100755
--- a/t/t5616-partial-clone.sh
+++ b/t/t5616-partial-clone.sh
@@ -515,7 +515,18 @@ test_expect_success 'fetch lazy-fetches only to resolve deltas' '
 	# Exercise to make sure it works. Git will not fetch anything from the
 	# promisor remote other than for the big tree (because it needs to
 	# resolve the delta).
-	GIT_TRACE_PACKET="$(pwd)/trace" git -C client \
+	#
+	# TODO: the --full-name-hash option is disabled here, since this test
+	# is fundamentally broken! When GIT_TEST_FULL_NAME_HASH=1, the server
+	# recognizes delta bases in a different way and then sends a _blob_ to
+	# the client with a delta base that the client does not have! This is
+	# because the client is cloned from "promisor-server" with tree:0 but
+	# is now fetching from "server" withot any filter. This is violating the
+	# promise to the server that all reachable objects exist and could be
+	# used as delta bases!
+	GIT_TRACE_PACKET="$(pwd)/trace" \
+	GIT_TEST_FULL_NAME_HASH=0 \
+		git -C client \
 		fetch "file://$(pwd)/server" main &&
 
 	# Verify the assumption that the client needed to fetch the delta base
@@ -534,7 +545,18 @@ test_expect_success 'fetch lazy-fetches only to resolve deltas, protocol v2' '
 	# Exercise to make sure it works. Git will not fetch anything from the
 	# promisor remote other than for the big blob (because it needs to
 	# resolve the delta).
-	GIT_TRACE_PACKET="$(pwd)/trace" git -C client \
+	#
+	# TODO: the --full-name-hash option is disabled here, since this test
+	# is fundamentally broken! When GIT_TEST_FULL_NAME_HASH=1, the server
+	# recognizes delta bases in a different way and then sends a _blob_ to
+	# the client with a delta base that the client does not have! This is
+	# because the client is cloned from "promisor-server" with tree:0 but
+	# is now fetching from "server" withot any filter. This is violating the
+	# promise to the server that all reachable objects exist and could be
+	# used as delta bases!
+	GIT_TRACE_PACKET="$(pwd)/trace" \
+	GIT_TEST_FULL_NAME_HASH=0 \
+		git -C client \
 		fetch "file://$(pwd)/server" main &&
 
 	# Verify that protocol version 2 was used.
diff --git a/t/t5620-backfill.sh b/t/t5620-backfill.sh
new file mode 100755
index 00000000000000..eecf03d5199ea9
--- /dev/null
+++ b/t/t5620-backfill.sh
@@ -0,0 +1,178 @@
+#!/bin/sh
+
+test_description='git backfill on partial clones'
+
+GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME=main
+export GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME
+
+. ./test-lib.sh
+
+# We create objects in the 'src' repo.
+test_expect_success 'setup repo for object creation' '
+	echo "{print \$1}" >print_1.awk &&
+	echo "{print \$2}" >print_2.awk &&
+
+	git init src &&
+
+	mkdir -p src/a/b/c &&
+	mkdir -p src/d/e &&
+
+	for i in 1 2
+	do
+		for n in 1 2 3 4
+		do
+			echo "Version $i of file $n" > src/file.$n.txt &&
+			echo "Version $i of file a/$n" > src/a/file.$n.txt &&
+			echo "Version $i of file a/b/$n" > src/a/b/file.$n.txt &&
+			echo "Version $i of file a/b/c/$n" > src/a/b/c/file.$n.txt &&
+			echo "Version $i of file d/$n" > src/d/file.$n.txt &&
+			echo "Version $i of file d/e/$n" > src/d/e/file.$n.txt &&
+			git -C src add . &&
+			git -C src commit -m "Iteration $n" || return 1
+		done
+	done
+'
+
+# Clone 'src' into 'srv.bare' so we have a bare repo to be our origin
+# server for the partial clone.
+test_expect_success 'setup bare clone for server' '
+	git clone --bare "file://$(pwd)/src" srv.bare &&
+	git -C srv.bare config --local uploadpack.allowfilter 1 &&
+	git -C srv.bare config --local uploadpack.allowanysha1inwant 1
+'
+
+# do basic partial clone from "srv.bare"
+test_expect_success 'do partial clone 1, backfill gets all objects' '
+	git clone --no-checkout --filter=blob:none	\
+		--single-branch --branch=main 		\
+		"file://$(pwd)/srv.bare" backfill1 &&
+
+	# Backfill with no options gets everything reachable from HEAD.
+	GIT_TRACE2_EVENT="$(pwd)/backfill-file-trace" git \
+		-C backfill1 backfill &&
+
+	# We should have engaged the partial clone machinery
+	test_trace2_data promisor fetch_count 48 <backfill-file-trace &&
+
+	# No more missing objects!
+	git -C backfill1 rev-list --quiet --objects --missing=print HEAD >revs2 &&
+	test_line_count = 0 revs2
+'
+
+test_expect_success 'do partial clone 2, backfill batch size' '
+	git clone --no-checkout --filter=blob:none	\
+		--single-branch --branch=main 		\
+		"file://$(pwd)/srv.bare" backfill2 &&
+
+	GIT_TRACE2_EVENT="$(pwd)/batch-trace" git \
+		-C backfill2 backfill --batch-size=20 &&
+
+	# Batches were used
+	test_trace2_data promisor fetch_count 20 <batch-trace >matches &&
+	test_line_count = 2 matches &&
+	test_trace2_data promisor fetch_count 8 <batch-trace &&
+
+	# No more missing objects!
+	git -C backfill2 rev-list --quiet --objects --missing=print HEAD >revs2 &&
+	test_line_count = 0 revs2
+'
+
+test_expect_success 'backfill --sparse without sparse-checkout fails' '
+	git init not-sparse &&
+	test_must_fail git -C not-sparse backfill --sparse 2>err &&
+	grep "problem loading sparse-checkout" err
+'
+
+test_expect_success 'backfill --sparse' '
+	git clone --sparse --filter=blob:none		\
+		--single-branch --branch=main 		\
+		"file://$(pwd)/srv.bare" backfill3 &&
+
+	# Initial checkout includes four files at root.
+	git -C backfill3 rev-list --quiet --objects --missing=print HEAD >missing &&
+	test_line_count = 44 missing &&
+
+	# Initial sparse-checkout is just the files at root, so we get the
+	# older versions of the four files at tip.
+	GIT_TRACE2_EVENT="$(pwd)/sparse-trace1" git \
+		-C backfill3 backfill --sparse &&
+	test_trace2_data promisor fetch_count 4 <sparse-trace1 &&
+	test_trace2_data path-walk paths 5 <sparse-trace1 &&
+	git -C backfill3 rev-list --quiet --objects --missing=print HEAD >missing &&
+	test_line_count = 40 missing &&
+
+	# Expand the sparse-checkout to include 'd' recursively. This
+	# engages the algorithm to skip the trees for 'a'. Note that
+	# the "sparse-checkout set" command downloads the objects at tip
+	# to satisfy the current checkout.
+	git -C backfill3 sparse-checkout set d &&
+	GIT_TRACE2_EVENT="$(pwd)/sparse-trace2" git \
+		-C backfill3 backfill --sparse &&
+	test_trace2_data promisor fetch_count 8 <sparse-trace2 &&
+	test_trace2_data path-walk paths 15 <sparse-trace2 &&
+	git -C backfill3 rev-list --quiet --objects --missing=print HEAD >missing &&
+	test_line_count = 24 missing &&
+
+	# Disabling the --sparse option (on by default) will download everything
+	git -C backfill3 backfill --no-sparse &&
+	git -C backfill3 rev-list --quiet --objects --missing=print HEAD >missing &&
+	test_line_count = 0 missing
+'
+
+test_expect_success 'backfill --sparse without cone mode' '
+	git clone --no-checkout --filter=blob:none		\
+		--single-branch --branch=main 		\
+		"file://$(pwd)/srv.bare" backfill4 &&
+
+	# No blobs yet
+	git -C backfill4 rev-list --quiet --objects --missing=print HEAD >missing &&
+	test_line_count = 48 missing &&
+
+	# Define sparse-checkout by filename regardless of parent directory.
+	# This downloads 6 blobs to satisfy the checkout.
+	git -C backfill4 sparse-checkout set --no-cone "**/file.1.txt" &&
+	git -C backfill4 checkout main &&
+
+	GIT_TRACE2_EVENT="$(pwd)/no-cone-trace1" git \
+		-C backfill4 backfill --sparse &&
+	test_trace2_data promisor fetch_count 6 <no-cone-trace1 &&
+
+	# This walk needed to visit all directories to search for these paths.
+	test_trace2_data path-walk paths 12 <no-cone-trace1 &&
+	git -C backfill4 rev-list --quiet --objects --missing=print HEAD >missing &&
+	test_line_count = 36 missing
+'
+
+. "$TEST_DIRECTORY"/lib-httpd.sh
+start_httpd
+
+test_expect_success 'create a partial clone over HTTP' '
+	SERVER="$HTTPD_DOCUMENT_ROOT_PATH/server" &&
+	rm -rf "$SERVER" repo &&
+	git clone --bare "file://$(pwd)/src" "$SERVER" &&
+	test_config -C "$SERVER" uploadpack.allowfilter 1 &&
+	test_config -C "$SERVER" uploadpack.allowanysha1inwant 1 &&
+
+	git clone --no-checkout --filter=blob:none \
+		"$HTTPD_URL/smart/server" backfill-http
+'
+
+test_expect_success 'backfilling over HTTP succeeds' '
+	GIT_TRACE2_EVENT="$(pwd)/backfill-http-trace" git \
+		-C backfill-http backfill &&
+
+	# We should have engaged the partial clone machinery
+	test_trace2_data promisor fetch_count 48 <backfill-http-trace &&
+
+	# Confirm all objects are present, none missing.
+	git -C backfill-http rev-list --objects --all >rev-list-out &&
+	awk "{print \$1;}" <rev-list-out >oids &&
+	GIT_TRACE2_EVENT="$(pwd)/walk-trace" git -C backfill-http \
+		cat-file --batch-check <oids >batch-out &&
+	! grep missing batch-out
+'
+
+# DO NOT add non-httpd-specific tests here, because the last part of this
+# test script is only executed when httpd is available and enabled.
+
+test_done
diff --git a/t/t5801-remote-helpers.sh b/t/t5801-remote-helpers.sh
index d21877150ed82e..3917da47276825 100755
--- a/t/t5801-remote-helpers.sh
+++ b/t/t5801-remote-helpers.sh
@@ -262,7 +262,7 @@ test_expect_success 'push update refs failure' '
 	echo "update fail" >>file &&
 	git commit -a -m "update fail" &&
 	git rev-parse --verify testgit/origin/heads/update >expect &&
-	test_expect_code 1 env GIT_REMOTE_TESTGIT_FAILURE="non-fast forward" \
+	test_must_fail env GIT_REMOTE_TESTGIT_FAILURE="non-fast forward" \
 		git push origin update &&
 	git rev-parse --verify testgit/origin/heads/update >actual &&
 	test_cmp expect actual
diff --git a/t/t6020-bundle-misc.sh b/t/t6020-bundle-misc.sh
index 4ce62feaa26fde..5317649b7223b8 100755
--- a/t/t6020-bundle-misc.sh
+++ b/t/t6020-bundle-misc.sh
@@ -246,7 +246,11 @@ test_expect_success 'create bundle with --since option' '
 	EOF
 	test_cmp expect actual &&
 
-	git bundle create since.bdl \
+	# If the --full-name-hash option is used, then one fewer
+	# delta base is found and this counts a different number
+	# of objects after performing --fix-thin.
+	GIT_TEST_FULL_NAME_HASH=0 \
+		git bundle create since.bdl \
 		--since "Thu Apr 7 15:27:00 2005 -0700" \
 		--all &&
 
diff --git a/t/t6601-path-walk.sh b/t/t6601-path-walk.sh
new file mode 100755
index 00000000000000..312bf3c19c176a
--- /dev/null
+++ b/t/t6601-path-walk.sh
@@ -0,0 +1,336 @@
+#!/bin/sh
+
+test_description='direct path-walk API tests'
+
+. ./test-lib.sh
+
+test_expect_success 'setup test repository' '
+	git checkout -b base &&
+
+	# Make some objects that will only be reachable
+	# via non-commit tags.
+	mkdir child &&
+	echo file >child/file &&
+	git add child &&
+	git commit -m "will abandon" &&
+	git tag -a -m "tree" tree-tag HEAD^{tree} &&
+	echo file2 >file2 &&
+	git add file2 &&
+	git commit --amend -m "will abandon" &&
+	git tag tree-tag2 HEAD^{tree} &&
+
+	echo blob >file &&
+	blob_oid=$(git hash-object -t blob -w --stdin <file) &&
+	git tag -a -m "blob" blob-tag "$blob_oid" &&
+	echo blob2 >file2 &&
+	blob2_oid=$(git hash-object -t blob -w --stdin <file2) &&
+	git tag blob-tag2 "$blob2_oid" &&
+
+	rm -fr child file file2 &&
+
+	mkdir left &&
+	mkdir right &&
+	echo a >a &&
+	echo b >left/b &&
+	echo c >right/c &&
+	git add . &&
+	git commit --amend -m "first" &&
+	git tag -m "first" first HEAD &&
+
+	echo d >right/d &&
+	git add right &&
+	git commit -m "second" &&
+	git tag -a -m "second (under)" second.1 HEAD &&
+	git tag -a -m "second (top)" second.2 second.1 &&
+
+	# Set up file/dir collision in history.
+	rm a &&
+	mkdir a &&
+	echo a >a/a &&
+	echo bb >left/b &&
+	git add a left &&
+	git commit -m "third" &&
+	git tag -a -m "third" third &&
+
+	git checkout -b topic HEAD~1 &&
+	echo cc >right/c &&
+	git commit -a -m "topic" &&
+	git tag -a -m "fourth" fourth
+'
+
+test_expect_success 'all' '
+	test-tool path-walk -- --all >out &&
+
+	cat >expect <<-EOF &&
+	COMMIT::$(git rev-parse topic)
+	COMMIT::$(git rev-parse base)
+	COMMIT::$(git rev-parse base~1)
+	COMMIT::$(git rev-parse base~2)
+	commits:4
+	TREE::$(git rev-parse topic^{tree})
+	TREE::$(git rev-parse base^{tree})
+	TREE::$(git rev-parse base~1^{tree})
+	TREE::$(git rev-parse base~2^{tree})
+	TREE::$(git rev-parse refs/tags/tree-tag^{})
+	TREE::$(git rev-parse refs/tags/tree-tag2^{})
+	TREE:a/:$(git rev-parse base:a)
+	TREE:left/:$(git rev-parse base:left)
+	TREE:left/:$(git rev-parse base~2:left)
+	TREE:right/:$(git rev-parse topic:right)
+	TREE:right/:$(git rev-parse base~1:right)
+	TREE:right/:$(git rev-parse base~2:right)
+	TREE:child/:$(git rev-parse refs/tags/tree-tag^{}:child)
+	trees:13
+	BLOB:a:$(git rev-parse base~2:a)
+	BLOB:file2:$(git rev-parse refs/tags/tree-tag2^{}:file2)
+	BLOB:left/b:$(git rev-parse base~2:left/b)
+	BLOB:left/b:$(git rev-parse base:left/b)
+	BLOB:right/c:$(git rev-parse base~2:right/c)
+	BLOB:right/c:$(git rev-parse topic:right/c)
+	BLOB:right/d:$(git rev-parse base~1:right/d)
+	BLOB:/tagged-blobs:$(git rev-parse refs/tags/blob-tag^{})
+	BLOB:/tagged-blobs:$(git rev-parse refs/tags/blob-tag2^{})
+	BLOB:child/file:$(git rev-parse refs/tags/tree-tag^{}:child/file)
+	blobs:10
+	TAG::$(git rev-parse refs/tags/first)
+	TAG::$(git rev-parse refs/tags/second.1)
+	TAG::$(git rev-parse refs/tags/second.2)
+	TAG::$(git rev-parse refs/tags/third)
+	TAG::$(git rev-parse refs/tags/fourth)
+	TAG::$(git rev-parse refs/tags/tree-tag)
+	TAG::$(git rev-parse refs/tags/blob-tag)
+	tags:7
+	EOF
+
+	sort expect >expect.sorted &&
+	sort out >out.sorted &&
+
+	test_cmp expect.sorted out.sorted
+'
+
+test_expect_success 'base & topic, sparse' '
+	cat >patterns <<-EOF &&
+	/*
+	!/*/
+	/left/
+	EOF
+
+	test-tool path-walk --stdin-pl -- base topic <patterns >out &&
+
+	cat >expect <<-EOF &&
+	COMMIT::$(git rev-parse topic)
+	COMMIT::$(git rev-parse base)
+	COMMIT::$(git rev-parse base~1)
+	COMMIT::$(git rev-parse base~2)
+	commits:4
+	TREE::$(git rev-parse topic^{tree})
+	TREE::$(git rev-parse base^{tree})
+	TREE::$(git rev-parse base~1^{tree})
+	TREE::$(git rev-parse base~2^{tree})
+	TREE:left/:$(git rev-parse base:left)
+	TREE:left/:$(git rev-parse base~2:left)
+	trees:6
+	BLOB:a:$(git rev-parse base~2:a)
+	BLOB:left/b:$(git rev-parse base~2:left/b)
+	BLOB:left/b:$(git rev-parse base:left/b)
+	blobs:3
+	tags:0
+	EOF
+
+	sort expect >expect.sorted &&
+	sort out >out.sorted &&
+
+	test_cmp expect.sorted out.sorted
+'
+
+test_expect_success 'topic only' '
+	test-tool path-walk -- topic >out &&
+
+	cat >expect <<-EOF &&
+	COMMIT::$(git rev-parse topic)
+	COMMIT::$(git rev-parse base~1)
+	COMMIT::$(git rev-parse base~2)
+	commits:3
+	TREE::$(git rev-parse topic^{tree})
+	TREE::$(git rev-parse base~1^{tree})
+	TREE::$(git rev-parse base~2^{tree})
+	TREE:left/:$(git rev-parse base~2:left)
+	TREE:right/:$(git rev-parse topic:right)
+	TREE:right/:$(git rev-parse base~1:right)
+	TREE:right/:$(git rev-parse base~2:right)
+	trees:7
+	BLOB:a:$(git rev-parse base~2:a)
+	BLOB:left/b:$(git rev-parse base~2:left/b)
+	BLOB:right/c:$(git rev-parse base~2:right/c)
+	BLOB:right/c:$(git rev-parse topic:right/c)
+	BLOB:right/d:$(git rev-parse base~1:right/d)
+	blobs:5
+	tags:0
+	EOF
+
+	sort expect >expect.sorted &&
+	sort out >out.sorted &&
+
+	test_cmp expect.sorted out.sorted
+'
+
+test_expect_success 'topic, not base' '
+	test-tool path-walk -- topic --not base >out &&
+
+	cat >expect <<-EOF &&
+	COMMIT::$(git rev-parse topic)
+	commits:1
+	TREE::$(git rev-parse topic^{tree})
+	TREE:left/:$(git rev-parse topic:left)
+	TREE:right/:$(git rev-parse topic:right)
+	trees:3
+	BLOB:a:$(git rev-parse topic:a)
+	BLOB:left/b:$(git rev-parse topic:left/b)
+	BLOB:right/c:$(git rev-parse topic:right/c)
+	BLOB:right/d:$(git rev-parse topic:right/d)
+	blobs:4
+	tags:0
+	EOF
+
+	sort expect >expect.sorted &&
+	sort out >out.sorted &&
+
+	test_cmp expect.sorted out.sorted
+'
+
+test_expect_success 'topic, not base, only blobs' '
+	test-tool path-walk --no-trees --no-commits \
+		-- topic --not base >out &&
+
+	cat >expect <<-EOF &&
+	commits:0
+	trees:0
+	BLOB:a:$(git rev-parse topic:a)
+	BLOB:left/b:$(git rev-parse topic:left/b)
+	BLOB:right/c:$(git rev-parse topic:right/c)
+	BLOB:right/d:$(git rev-parse topic:right/d)
+	blobs:4
+	tags:0
+	EOF
+
+	sort expect >expect.sorted &&
+	sort out >out.sorted &&
+
+	test_cmp expect.sorted out.sorted
+'
+
+# No, this doesn't make a lot of sense for the path-walk API,
+# but it is possible to do.
+test_expect_success 'topic, not base, only commits' '
+	test-tool path-walk --no-blobs --no-trees \
+		-- topic --not base >out &&
+
+	cat >expect <<-EOF &&
+	COMMIT::$(git rev-parse topic)
+	commits:1
+	trees:0
+	blobs:0
+	tags:0
+	EOF
+
+	sort expect >expect.sorted &&
+	sort out >out.sorted &&
+
+	test_cmp expect.sorted out.sorted
+'
+
+test_expect_success 'topic, not base, only trees' '
+	test-tool path-walk --no-blobs --no-commits \
+		-- topic --not base >out &&
+
+	cat >expect <<-EOF &&
+	commits:0
+	TREE::$(git rev-parse topic^{tree})
+	TREE:left/:$(git rev-parse topic:left)
+	TREE:right/:$(git rev-parse topic:right)
+	trees:3
+	blobs:0
+	tags:0
+	EOF
+
+	sort expect >expect.sorted &&
+	sort out >out.sorted &&
+
+	test_cmp expect.sorted out.sorted
+'
+
+test_expect_success 'topic, not base, boundary' '
+	test-tool path-walk -- --boundary topic --not base >out &&
+
+	cat >expect <<-EOF &&
+	COMMIT::$(git rev-parse topic)
+	COMMIT::$(git rev-parse base~1):UNINTERESTING
+	commits:2
+	TREE::$(git rev-parse topic^{tree})
+	TREE::$(git rev-parse base~1^{tree}):UNINTERESTING
+	TREE:left/:$(git rev-parse base~1:left):UNINTERESTING
+	TREE:right/:$(git rev-parse topic:right)
+	TREE:right/:$(git rev-parse base~1:right):UNINTERESTING
+	trees:5
+	BLOB:a:$(git rev-parse base~1:a):UNINTERESTING
+	BLOB:left/b:$(git rev-parse base~1:left/b):UNINTERESTING
+	BLOB:right/c:$(git rev-parse base~1:right/c):UNINTERESTING
+	BLOB:right/c:$(git rev-parse topic:right/c)
+	BLOB:right/d:$(git rev-parse base~1:right/d):UNINTERESTING
+	blobs:5
+	tags:0
+	EOF
+
+	sort expect >expect.sorted &&
+	sort out >out.sorted &&
+
+	test_cmp expect.sorted out.sorted
+'
+
+test_expect_success 'topic, not base, boundary with pruning' '
+	test-tool path-walk --prune -- --boundary topic --not base >out &&
+
+	cat >expect <<-EOF &&
+	COMMIT::$(git rev-parse topic)
+	COMMIT::$(git rev-parse base~1):UNINTERESTING
+	commits:2
+	TREE::$(git rev-parse topic^{tree})
+	TREE::$(git rev-parse base~1^{tree}):UNINTERESTING
+	TREE:right/:$(git rev-parse topic:right)
+	TREE:right/:$(git rev-parse base~1:right):UNINTERESTING
+	trees:4
+	BLOB:right/c:$(git rev-parse base~1:right/c):UNINTERESTING
+	BLOB:right/c:$(git rev-parse topic:right/c)
+	blobs:2
+	tags:0
+	EOF
+
+	sort expect >expect.sorted &&
+	sort out >out.sorted &&
+
+	test_cmp expect.sorted out.sorted
+'
+
+test_expect_success 'trees are reported exactly once' '
+	test_when_finished "rm -rf unique-trees" &&
+	test_create_repo unique-trees &&
+	(
+		cd unique-trees &&
+		mkdir initial &&
+		test_commit initial/file &&
+
+		git switch -c move-to-top &&
+		git mv initial/file.t ./ &&
+		test_tick &&
+		git commit -m moved &&
+
+		git update-ref refs/heads/other HEAD
+	) &&
+
+	test-tool -C unique-trees path-walk -- --all >out &&
+	tree=$(git -C unique-trees rev-parse HEAD:) &&
+	grep "$tree" out >out-filtered &&
+	test_line_count = 1 out-filtered
+'
+
+test_done
diff --git a/t/t7300-clean.sh b/t/t7300-clean.sh
index 00d4070156243b..6f16f3893191e7 100755
--- a/t/t7300-clean.sh
+++ b/t/t7300-clean.sh
@@ -800,4 +800,14 @@ test_expect_success 'traverse into directories that may have ignored entries' '
 	)
 '
 
+test_expect_success MINGW 'clean does not traverse mount points' '
+	mkdir target &&
+	>target/dont-clean-me &&
+	git init with-mountpoint &&
+	cmd //c "mklink /j with-mountpoint\\mountpoint target" &&
+	git -C with-mountpoint clean -dfx &&
+	test_path_is_missing with-mountpoint/mountpoint &&
+	test_path_is_file target/dont-clean-me
+'
+
 test_done
diff --git a/t/t7406-submodule-update.sh b/t/t7406-submodule-update.sh
index 297c6c3b5cc4b8..15e29f538cacc1 100755
--- a/t/t7406-submodule-update.sh
+++ b/t/t7406-submodule-update.sh
@@ -14,6 +14,15 @@ export GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME
 
 . ./test-lib.sh
 
+# This test script contains test cases that need to create symbolic links. To
+# make sure that these test cases are exercised in Git for Windows, where (for
+# historical reasons) `ln -s` creates copies by default, let's specifically ask
+# for `ln -s` to create symbolic links whenever possible.
+if test_have_prereq MINGW
+then
+	MSYS=${MSYS+$MSYS }winsymlinks:nativestrict
+	export MSYS
+fi
 
 compare_head()
 {
@@ -1093,12 +1102,16 @@ test_expect_success 'submodule update --quiet passes quietness to fetch with a s
 	) &&
 	git clone super4 super5 &&
 	(cd super5 &&
+	 # This test variable will create a "warning" message to stderr
+	 GIT_TEST_PACK_PATH_WALK=0 \
 	 git submodule update --quiet --init --depth=1 submodule3 >out 2>err &&
 	 test_must_be_empty out &&
 	 test_must_be_empty err
 	) &&
 	git clone super4 super6 &&
 	(cd super6 &&
+	 # This test variable will create a "warning" message to stderr
+	 GIT_TEST_PACK_PATH_WALK=0 \
 	 git submodule update --init --depth=1 submodule3 >out 2>err &&
 	 test_file_not_empty out &&
 	 test_file_not_empty err
diff --git a/t/t7502-commit-porcelain.sh b/t/t7502-commit-porcelain.sh
index b37e2018a74a7b..c38b96b66cd20a 100755
--- a/t/t7502-commit-porcelain.sh
+++ b/t/t7502-commit-porcelain.sh
@@ -623,6 +623,48 @@ test_expect_success 'cleanup commit messages (scissors option,-F,-e, scissors on
 	test_must_be_empty actual
 '
 
+test_expect_success 'helper-editor' '
+
+	write_script lf-to-crlf.sh <<-\EOF
+	sed "s/\$/Q/" <"$1" | tr Q "\\015" >"$1".new &&
+	mv -f "$1".new "$1"
+	EOF
+'
+
+test_expect_success 'cleanup commit messages (scissors option,-F,-e, CR/LF line endings)' '
+
+	test_config core.editor "\"$PWD/lf-to-crlf.sh\"" &&
+	scissors="# ------------------------ >8 ------------------------" &&
+
+	test_write_lines >text \
+	"# Keep this comment" "" " $scissors" \
+	"# Keep this comment, too" "$scissors" \
+	"# Remove this comment" "$scissors" \
+	"Remove this comment, too" &&
+
+	test_write_lines >expect \
+	"# Keep this comment" "" " $scissors" \
+	"# Keep this comment, too" &&
+
+	git commit --cleanup=scissors -e -F text --allow-empty &&
+	git cat-file -p HEAD >raw &&
+	sed -e "1,/^\$/d" raw >actual &&
+	test_cmp expect actual
+'
+
+test_expect_success 'cleanup commit messages (scissors option,-F,-e, scissors on first line, CR/LF line endings)' '
+
+	scissors="# ------------------------ >8 ------------------------" &&
+	test_write_lines >text \
+	"$scissors" \
+	"# Remove this comment and any following lines" &&
+	cp text /tmp/test2-text &&
+	git commit --cleanup=scissors -e -F text --allow-empty --allow-empty-message &&
+	git cat-file -p HEAD >raw &&
+	sed -e "1,/^\$/d" raw >actual &&
+	test_must_be_empty actual
+'
+
 test_expect_success 'cleanup commit messages (strip option,-F)' '
 
 	echo >>negative &&
diff --git a/t/t7700-repack.sh b/t/t7700-repack.sh
index be1188e7365b19..1feb6643e69979 100755
--- a/t/t7700-repack.sh
+++ b/t/t7700-repack.sh
@@ -776,6 +776,13 @@ test_expect_success 'repack -ad cleans up old .tmp-* packs' '
 	test_must_be_empty tmpfiles
 '
 
+test_expect_success '--full-name-hash option passes through to pack-objects' '
+	GIT_TRACE2_EVENT="$(pwd)/full-trace.txt" \
+		git repack -a --full-name-hash &&
+	test_subcommand_flex git pack-objects --full-name-hash <full-trace.txt
+'
+
+
 test_expect_success 'setup for update-server-info' '
 	git init update-server-info &&
 	test_commit -C update-server-info message
diff --git a/t/t8100-git-survey.sh b/t/t8100-git-survey.sh
new file mode 100755
index 00000000000000..8c6edfcae0c6c2
--- /dev/null
+++ b/t/t8100-git-survey.sh
@@ -0,0 +1,102 @@
+#!/bin/sh
+
+test_description='git survey'
+
+GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME=main
+export GIT_TEST_DEFAULT_INITIAL_BRANCH_NAME
+
+TEST_PASSES_SANITIZE_LEAK=0
+export TEST_PASSES_SANITIZE_LEAK
+
+. ./test-lib.sh
+
+test_expect_success 'git survey -h shows experimental warning' '
+	test_expect_code 129 git survey -h 2>usage &&
+	grep "EXPERIMENTAL!" usage
+'
+
+test_expect_success 'create a semi-interesting repo' '
+	test_commit_bulk 10 &&
+	git tag -a -m one one HEAD~5 &&
+	git tag -a -m two two HEAD~3 &&
+	git tag -a -m three three two &&
+	git tag -a -m four four three &&
+	git update-ref -d refs/tags/three &&
+	git update-ref -d refs/tags/two
+'
+
+test_expect_success 'git survey --progress' '
+	GIT_PROGRESS_DELAY=0 git survey --all-refs --progress >out 2>err &&
+	grep "Preparing object walk" err
+'
+
+test_expect_success 'git survey (default)' '
+	git survey --all-refs >out 2>err &&
+	test_line_count = 0 err &&
+
+	test_oid_cache <<-EOF &&
+	commits_size_on_disk sha1:     1523
+	commits_size_on_disk sha256:     1811
+
+	commits_size sha1:         2153
+	commits_size sha256:         2609
+
+	trees_size_on_disk sha1:      495
+	trees_size_on_disk sha256:      635
+
+	trees_size sha1:         1706
+	trees_size sha256:         2366
+
+	tags_size sha1:          528
+	tags_size sha256:          624
+
+	tags_size_on_disk sha1:      510
+	tags_size_on_disk sha256:      569
+	EOF
+
+	tr , " " >expect <<-EOF &&
+	GIT SURVEY for "$(pwd)"
+	-----------------------------------------------------
+
+	REFERENCES SUMMARY
+	========================
+	,       Ref Type | Count
+	-----------------+------
+	,       Branches |     1
+	     Remote refs |     0
+	      Tags (all) |     2
+	Tags (annotated) |     2
+
+	REACHABLE OBJECT SUMMARY
+	========================
+	Object Type | Count
+	------------+------
+	       Tags |     4
+	    Commits |    10
+	      Trees |    10
+	      Blobs |    10
+
+	TOTAL OBJECT SIZES BY TYPE
+	===============================================
+	Object Type | Count | Disk Size | Inflated Size
+	------------+-------+-----------+--------------
+	    Commits |    10 | $(test_oid commits_size_on_disk) | $(test_oid commits_size)
+	      Trees |    10 | $(test_oid trees_size_on_disk) | $(test_oid trees_size)
+	      Blobs |    10 |       191 |           101
+	       Tags |     4 | $(test_oid tags_size_on_disk) | $(test_oid tags_size)
+	EOF
+
+	lines=$(wc -l <expect) &&
+	head -n $lines out >out-trimmed &&
+	test_cmp expect out-trimmed &&
+
+	for type in "DIRECTORIES" "FILES"
+	do
+		for metric in "COUNT" "DISK SIZE" "INFLATED SIZE"
+		do
+			grep "TOP $type BY $metric" out || return 1
+		done || return 1
+	done
+'
+
+test_done
diff --git a/t/t9350-fast-export.sh b/t/t9350-fast-export.sh
index 40427883ec6dbf..abffd267e259c4 100755
--- a/t/t9350-fast-export.sh
+++ b/t/t9350-fast-export.sh
@@ -801,4 +801,15 @@ test_expect_success 'fast-export handles --end-of-options' '
 	test_cmp expect actual
 '
 
+cat > expected << EOF
+reset refs/heads/master
+from $(git rev-parse master)
+
+EOF
+
+test_expect_failure 'refs are updated even if no commits need to be exported' '
+	git fast-export master..master > actual &&
+	test_cmp expected actual
+'
+
 test_done
diff --git a/t/test-lib-functions.sh b/t/test-lib-functions.sh
index 78e054ab503a65..860ce3f260ab71 100644
--- a/t/test-lib-functions.sh
+++ b/t/test-lib-functions.sh
@@ -1886,6 +1886,33 @@ test_subcommand () {
 	fi
 }
 
+
+# Check that the given subcommand was run with the given set of
+# arguments in order (but with possible extra arguments).
+#
+#	test_subcommand_flex [!] <command> <args>... < <trace>
+#
+# If the first parameter passed is !, this instead checks that
+# the given command was not called.
+#
+test_subcommand_flex () {
+	local negate=
+	if test "$1" = "!"
+	then
+		negate=t
+		shift
+	fi
+
+	local expr="$(printf '"%s".*' "$@")"
+
+	if test -n "$negate"
+	then
+		! grep "\[$expr\]"
+	else
+		grep "\[$expr\]"
+	fi
+}
+
 # Check that the given command was invoked as part of the
 # trace2-format trace on stdin.
 #
diff --git a/transport-helper.c b/transport-helper.c
index d457b425501a74..7513fd7eea05e4 100644
--- a/transport-helper.c
+++ b/transport-helper.c
@@ -22,6 +22,8 @@
 #include "packfile.h"
 
 static int debug;
+/* TODO: put somewhere sensible, e.g. git_transport_options? */
+static int auto_gc = 1;
 
 struct helper_data {
 	char *name;
@@ -499,10 +501,25 @@ static int get_exporter(struct transport *transport,
 	for (size_t i = 0; i < revlist_args->nr; i++)
 		strvec_push(&fastexport->args, revlist_args->items[i].string);
 
+	strvec_push(&fastexport->args, "--");
+
 	fastexport->git_cmd = 1;
 	return start_command(fastexport);
 }
 
+static void check_helper_status(struct helper_data *data)
+{
+	int pid, status;
+
+	pid = waitpid(data->helper->pid, &status, WNOHANG);
+	if (pid < 0)
+		die("Could not retrieve status of remote helper '%s'",
+		    data->name);
+	if (pid > 0 && WIFEXITED(status))
+		die("Remote helper '%s' died with %d",
+		    data->name, WEXITSTATUS(status));
+}
+
 static int fetch_with_import(struct transport *transport,
 			     int nr_heads, struct ref **to_fetch)
 {
@@ -539,6 +556,7 @@ static int fetch_with_import(struct transport *transport,
 
 	if (finish_command(&fastimport))
 		die(_("error while running fast-import"));
+	check_helper_status(data);
 
 	/*
 	 * The fast-import stream of a remote helper that advertises
@@ -572,6 +590,13 @@ static int fetch_with_import(struct transport *transport,
 		}
 	}
 	strbuf_release(&buf);
+	if (auto_gc) {
+		struct child_process cmd = CHILD_PROCESS_INIT;
+
+		cmd.git_cmd = 1;
+		strvec_pushl(&cmd.args, "gc", "--auto", "--quiet", NULL);
+		run_command(&cmd);
+	}
 	return 0;
 }
 
@@ -1158,6 +1183,7 @@ static int push_refs_with_export(struct transport *transport,
 
 	if (finish_command(&exporter))
 		die(_("error while running fast-export"));
+	check_helper_status(data);
 	if (push_update_refs_status(data, remote_refs, flags))
 		return 1;
 
diff --git a/wt-status.c b/wt-status.c
index 3ee918176422c7..9609ba260fe804 100644
--- a/wt-status.c
+++ b/wt-status.c
@@ -40,7 +40,7 @@
 #define UF_DELAY_WARNING_IN_MS (2 * 1000)
 
 static const char cut_line[] =
-"------------------------ >8 ------------------------\n";
+"------------------------ >8 ------------------------";
 
 static char default_wt_status_colors[][COLOR_MAXLEN] = {
 	GIT_COLOR_NORMAL, /* WT_STATUS_HEADER */
@@ -1096,15 +1096,22 @@ static void wt_longstatus_print_other(struct wt_status *s,
 	status_printf_ln(s, GIT_COLOR_NORMAL, "%s", "");
 }
 
+static inline int starts_with_newline(const char *p)
+{
+    return *p == '\n' || (*p == '\r' && p[1] == '\n');
+}
+
 size_t wt_status_locate_end(const char *s, size_t len)
 {
 	const char *p;
 	struct strbuf pattern = STRBUF_INIT;
 
 	strbuf_addf(&pattern, "\n%s %s", comment_line_str, cut_line);
-	if (starts_with(s, pattern.buf + 1))
+	if (starts_with(s, pattern.buf + 1) &&
+	    starts_with_newline(s + pattern.len - 1))
 		len = 0;
-	else if ((p = strstr(s, pattern.buf))) {
+	else if ((p = strstr(s, pattern.buf)) &&
+		 starts_with_newline(p + pattern.len)) {
 		size_t newlen = p - s + 1;
 		if (newlen < len)
 			len = newlen;