Auto merge of #136823 - matthiaskrgr:rollup-vp20mk1, r=matthiaskrgr

Rollup of 6 pull requests

Successful merges:

 - #136419 (adding autodiff tests)
 - #136628 (ci: upgrade to crosstool-ng 1.27.0)
 - #136681 (resolve `llvm-config` path properly on cross builds)
 - #136714 (Update `compiler-builtins` to 0.1.146)
 - #136731 (rustc_middle: parallel: TyCtxt: remove "unsafe impl DynSend/DynSync")
 - #136791 (Disable DWARF in linker options for i686-unknown-uefi)

Failed merges:

 - #136767 (improve host/cross target checking)

r? `@ghost`
`@rustbot` modify labels: rollup

compiler/rustc_ast/src/expand/autodiff_attrs.rs

@@ -30,14 +30,6 @@ pub enum DiffMode {
     Forward,
     /// The target function, to be created using reverse mode AD.
     Reverse,
-    /// The target function, to be created using forward mode AD.
-    /// This target function will also be used as a source for higher order derivatives,
-    /// so compute it before all Forward/Reverse targets and optimize it through llvm.
-    ForwardFirst,
-    /// The target function, to be created using reverse mode AD.
-    /// This target function will also be used as a source for higher order derivatives,
-    /// so compute it before all Forward/Reverse targets and optimize it through llvm.
-    ReverseFirst,
 }
 
 /// Dual and Duplicated (and their Only variants) are getting lowered to the same Enzyme Activity.
@@ -92,10 +84,10 @@ pub struct AutoDiffAttrs {
 
 impl DiffMode {
     pub fn is_rev(&self) -> bool {
-        matches!(self, DiffMode::Reverse | DiffMode::ReverseFirst)
+        matches!(self, DiffMode::Reverse)
     }
     pub fn is_fwd(&self) -> bool {
-        matches!(self, DiffMode::Forward | DiffMode::ForwardFirst)
+        matches!(self, DiffMode::Forward)
     }
 }
 
@@ -106,8 +98,6 @@ impl Display for DiffMode {
             DiffMode::Source => write!(f, "Source"),
             DiffMode::Forward => write!(f, "Forward"),
             DiffMode::Reverse => write!(f, "Reverse"),
-            DiffMode::ForwardFirst => write!(f, "ForwardFirst"),
-            DiffMode::ReverseFirst => write!(f, "ReverseFirst"),
         }
     }
 }
@@ -125,12 +115,12 @@ pub fn valid_ret_activity(mode: DiffMode, activity: DiffActivity) -> bool {
     match mode {
         DiffMode::Error => false,
         DiffMode::Source => false,
-        DiffMode::Forward | DiffMode::ForwardFirst => {
+        DiffMode::Forward => {
             activity == DiffActivity::Dual
                 || activity == DiffActivity::DualOnly
                 || activity == DiffActivity::Const
         }
-        DiffMode::Reverse | DiffMode::ReverseFirst => {
+        DiffMode::Reverse => {
             activity == DiffActivity::Const
                 || activity == DiffActivity::Active
                 || activity == DiffActivity::ActiveOnly
@@ -166,10 +156,10 @@ pub fn valid_input_activity(mode: DiffMode, activity: DiffActivity) -> bool {
     return match mode {
         DiffMode::Error => false,
         DiffMode::Source => false,
-        DiffMode::Forward | DiffMode::ForwardFirst => {
+        DiffMode::Forward => {
             matches!(activity, Dual | DualOnly | Const)
         }
-        DiffMode::Reverse | DiffMode::ReverseFirst => {
+        DiffMode::Reverse => {
             matches!(activity, Active | ActiveOnly | Duplicated | DuplicatedOnly | Const)
         }
     };
@@ -200,8 +190,6 @@ impl FromStr for DiffMode {
             "Source" => Ok(DiffMode::Source),
             "Forward" => Ok(DiffMode::Forward),
             "Reverse" => Ok(DiffMode::Reverse),
-            "ForwardFirst" => Ok(DiffMode::ForwardFirst),
-            "ReverseFirst" => Ok(DiffMode::ReverseFirst),
             _ => Err(()),
         }
     }

compiler/rustc_codegen_cranelift/patches/0029-stdlib-Disable-f16-and-f128-in-compiler-builtins.patch

@@ -16,8 +16,8 @@ index 7165c3e48af..968552ad435 100644
 
  [dependencies]
  core = { path = "../core" }
--compiler_builtins = { version = "=0.1.145", features = ['rustc-dep-of-std'] }
-+compiler_builtins = { version = "=0.1.145", features = ['rustc-dep-of-std', 'no-f16-f128'] }
+-compiler_builtins = { version = "=0.1.146", features = ['rustc-dep-of-std'] }
++compiler_builtins = { version = "=0.1.146", features = ['rustc-dep-of-std', 'no-f16-f128'] }
 
  [dev-dependencies]
  rand = { version = "0.8.5", default-features = false, features = ["alloc"] }

compiler/rustc_codegen_llvm/src/back/lto.rs

@@ -606,10 +606,31 @@ pub(crate) fn run_pass_manager(
 
     // If this rustc version was build with enzyme/autodiff enabled, and if users applied the
     // `#[autodiff]` macro at least once, then we will later call llvm_optimize a second time.
-    let first_run = true;
     debug!("running llvm pm opt pipeline");
     unsafe {
-        write::llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, first_run)?;
+        write::llvm_optimize(
+            cgcx,
+            dcx,
+            module,
+            config,
+            opt_level,
+            opt_stage,
+            write::AutodiffStage::DuringAD,
+        )?;
+    }
+    // FIXME(ZuseZ4): Make this more granular
+    if cfg!(llvm_enzyme) && !thin {
+        unsafe {
+            write::llvm_optimize(
+                cgcx,
+                dcx,
+                module,
+                config,
+                opt_level,
+                llvm::OptStage::FatLTO,
+                write::AutodiffStage::PostAD,
+            )?;
+        }
     }
     debug!("lto done");
     Ok(())

compiler/rustc_codegen_llvm/src/back/write.rs

@@ -530,14 +530,24 @@ fn get_instr_profile_output_path(config: &ModuleConfig) -> Option<CString> {
     config.instrument_coverage.then(|| c"default_%m_%p.profraw".to_owned())
 }
 
+// PreAD will run llvm opts but disable size increasing opts (vectorization, loop unrolling)
+// DuringAD is the same as above, but also runs the enzyme opt and autodiff passes.
+// PostAD will run all opts, including size increasing opts.
+#[derive(Debug, Eq, PartialEq)]
+pub(crate) enum AutodiffStage {
+    PreAD,
+    DuringAD,
+    PostAD,
+}
+
 pub(crate) unsafe fn llvm_optimize(
     cgcx: &CodegenContext<LlvmCodegenBackend>,
     dcx: DiagCtxtHandle<'_>,
     module: &ModuleCodegen<ModuleLlvm>,
     config: &ModuleConfig,
     opt_level: config::OptLevel,
     opt_stage: llvm::OptStage,
-    skip_size_increasing_opts: bool,
+    autodiff_stage: AutodiffStage,
 ) -> Result<(), FatalError> {
     // Enzyme:
     // The whole point of compiler based AD is to differentiate optimized IR instead of unoptimized
@@ -550,12 +560,16 @@ pub(crate) unsafe fn llvm_optimize(
     let unroll_loops;
     let vectorize_slp;
     let vectorize_loop;
+    let run_enzyme = cfg!(llvm_enzyme) && autodiff_stage == AutodiffStage::DuringAD;
 
     // When we build rustc with enzyme/autodiff support, we want to postpone size-increasing
-    // optimizations until after differentiation. FIXME(ZuseZ4): Before shipping on nightly,
+    // optimizations until after differentiation. Our pipeline is thus: (opt + enzyme), (full opt).
+    // We therefore have two calls to llvm_optimize, if autodiff is used.
+    //
+    // FIXME(ZuseZ4): Before shipping on nightly,
     // we should make this more granular, or at least check that the user has at least one autodiff
     // call in their code, to justify altering the compilation pipeline.
-    if skip_size_increasing_opts && cfg!(llvm_enzyme) {
+    if cfg!(llvm_enzyme) && autodiff_stage != AutodiffStage::PostAD {
         unroll_loops = false;
         vectorize_slp = false;
         vectorize_loop = false;
@@ -565,7 +579,7 @@ pub(crate) unsafe fn llvm_optimize(
         vectorize_slp = config.vectorize_slp;
         vectorize_loop = config.vectorize_loop;
     }
-    trace!(?unroll_loops, ?vectorize_slp, ?vectorize_loop);
+    trace!(?unroll_loops, ?vectorize_slp, ?vectorize_loop, ?run_enzyme);
     let using_thin_buffers = opt_stage == llvm::OptStage::PreLinkThinLTO || config.bitcode_needed();
     let pgo_gen_path = get_pgo_gen_path(config);
     let pgo_use_path = get_pgo_use_path(config);
@@ -633,6 +647,7 @@ pub(crate) unsafe fn llvm_optimize(
             vectorize_loop,
             config.no_builtins,
             config.emit_lifetime_markers,
+            run_enzyme,
             sanitizer_options.as_ref(),
             pgo_gen_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
             pgo_use_path.as_ref().map_or(std::ptr::null(), |s| s.as_ptr()),
@@ -684,18 +699,14 @@ pub(crate) unsafe fn optimize(
             _ => llvm::OptStage::PreLinkNoLTO,
         };
 
-        // If we know that we will later run AD, then we disable vectorization and loop unrolling
-        let skip_size_increasing_opts = cfg!(llvm_enzyme);
+        // If we know that we will later run AD, then we disable vectorization and loop unrolling.
+        // Otherwise we pretend AD is already done and run the normal opt pipeline (=PostAD).
+        // FIXME(ZuseZ4): Make this more granular, only set PreAD if we actually have autodiff
+        // usages, not just if we build rustc with autodiff support.
+        let autodiff_stage =
+            if cfg!(llvm_enzyme) { AutodiffStage::PreAD } else { AutodiffStage::PostAD };
         return unsafe {
-            llvm_optimize(
-                cgcx,
-                dcx,
-                module,
-                config,
-                opt_level,
-                opt_stage,
-                skip_size_increasing_opts,
-            )
+            llvm_optimize(cgcx, dcx, module, config, opt_level, opt_stage, autodiff_stage)
         };
     }
     Ok(())

compiler/rustc_codegen_llvm/src/builder/autodiff.rs

@@ -4,10 +4,9 @@ use rustc_ast::expand::autodiff_attrs::{AutoDiffAttrs, AutoDiffItem, DiffActivit
 use rustc_codegen_ssa::ModuleCodegen;
 use rustc_codegen_ssa::back::write::ModuleConfig;
 use rustc_errors::FatalError;
-use rustc_session::config::Lto;
 use tracing::{debug, trace};
 
-use crate::back::write::{llvm_err, llvm_optimize};
+use crate::back::write::llvm_err;
 use crate::builder::SBuilder;
 use crate::context::SimpleCx;
 use crate::declare::declare_simple_fn;
@@ -53,8 +52,6 @@ fn generate_enzyme_call<'ll>(
     let mut ad_name: String = match attrs.mode {
         DiffMode::Forward => "__enzyme_fwddiff",
         DiffMode::Reverse => "__enzyme_autodiff",
-        DiffMode::ForwardFirst => "__enzyme_fwddiff",
-        DiffMode::ReverseFirst => "__enzyme_autodiff",
         _ => panic!("logic bug in autodiff, unrecognized mode"),
     }
     .to_string();
@@ -153,7 +150,7 @@ fn generate_enzyme_call<'ll>(
             _ => {}
         }
 
-        trace!("matching autodiff arguments");
+        debug!("matching autodiff arguments");
         // We now handle the issue that Rust level arguments not always match the llvm-ir level
         // arguments. A slice, `&[f32]`, for example, is represented as a pointer and a length on
         // llvm-ir level. The number of activities matches the number of Rust level arguments, so we
@@ -164,10 +161,10 @@ fn generate_enzyme_call<'ll>(
         let mut activity_pos = 0;
         let outer_args: Vec<&llvm::Value> = get_params(outer_fn);
         while activity_pos < inputs.len() {
-            let activity = inputs[activity_pos as usize];
+            let diff_activity = inputs[activity_pos as usize];
             // Duplicated arguments received a shadow argument, into which enzyme will write the
             // gradient.
-            let (activity, duplicated): (&Metadata, bool) = match activity {
+            let (activity, duplicated): (&Metadata, bool) = match diff_activity {
                 DiffActivity::None => panic!("not a valid input activity"),
                 DiffActivity::Const => (enzyme_const, false),
                 DiffActivity::Active => (enzyme_out, false),
@@ -222,7 +219,15 @@ fn generate_enzyme_call<'ll>(
                     // A duplicated pointer will have the following two outer_fn arguments:
                     // (..., ptr, ptr, ...). We add the following llvm-ir to our __enzyme call:
                     // (..., metadata! enzyme_dup, ptr, ptr, ...).
-                    assert!(llvm::LLVMRustGetTypeKind(next_outer_ty) == llvm::TypeKind::Pointer);
+                    if matches!(
+                        diff_activity,
+                        DiffActivity::Duplicated | DiffActivity::DuplicatedOnly
+                    ) {
+                        assert!(
+                            llvm::LLVMRustGetTypeKind(next_outer_ty) == llvm::TypeKind::Pointer
+                        );
+                    }
+                    // In the case of Dual we don't have assumptions, e.g. f32 would be valid.
                     args.push(next_outer_arg);
                     outer_pos += 2;
                     activity_pos += 1;
@@ -277,7 +282,7 @@ pub(crate) fn differentiate<'ll>(
     module: &'ll ModuleCodegen<ModuleLlvm>,
     cgcx: &CodegenContext<LlvmCodegenBackend>,
     diff_items: Vec<AutoDiffItem>,
-    config: &ModuleConfig,
+    _config: &ModuleConfig,
 ) -> Result<(), FatalError> {
     for item in &diff_items {
         trace!("{}", item);
@@ -318,29 +323,6 @@ pub(crate) fn differentiate<'ll>(
 
     // FIXME(ZuseZ4): support SanitizeHWAddress and prevent illegal/unsupported opts
 
-    if let Some(opt_level) = config.opt_level {
-        let opt_stage = match cgcx.lto {
-            Lto::Fat => llvm::OptStage::PreLinkFatLTO,
-            Lto::Thin | Lto::ThinLocal => llvm::OptStage::PreLinkThinLTO,
-            _ if cgcx.opts.cg.linker_plugin_lto.enabled() => llvm::OptStage::PreLinkThinLTO,
-            _ => llvm::OptStage::PreLinkNoLTO,
-        };
-        // This is our second opt call, so now we run all opts,
-        // to make sure we get the best performance.
-        let skip_size_increasing_opts = false;
-        trace!("running Module Optimization after differentiation");
-        unsafe {
-            llvm_optimize(
-                cgcx,
-                diag_handler.handle(),
-                module,
-                config,
-                opt_level,
-                opt_stage,
-                skip_size_increasing_opts,
-            )?
-        };
-    }
     trace!("done with differentiate()");
 
     Ok(())

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -2382,6 +2382,7 @@ unsafe extern "C" {
         LoopVectorize: bool,
         DisableSimplifyLibCalls: bool,
         EmitLifetimeMarkers: bool,
+        RunEnzyme: bool,
         SanitizerOptions: Option<&SanitizerOptions>,
         PGOGenPath: *const c_char,
         PGOUsePath: *const c_char,

compiler/rustc_codegen_ssa/src/codegen_attrs.rs

@@ -914,8 +914,6 @@ fn autodiff_attrs(tcx: TyCtxt<'_>, id: DefId) -> Option<AutoDiffAttrs> {
     let mode = match mode.as_str() {
         "Forward" => DiffMode::Forward,
         "Reverse" => DiffMode::Reverse,
-        "ForwardFirst" => DiffMode::ForwardFirst,
-        "ReverseFirst" => DiffMode::ReverseFirst,
         _ => {
             span_bug!(mode.span, "rustc_autodiff attribute contains invalid mode");
         }

compiler/rustc_llvm/build.rs

@@ -193,6 +193,10 @@ fn main() {
         cfg.define(&flag, None);
     }
 
+    if tracked_env_var_os("LLVM_ENZYME").is_some() {
+        cfg.define("ENZYME", None);
+    }
+
     if tracked_env_var_os("LLVM_RUSTLLVM").is_some() {
         cfg.define("LLVM_RUSTLLVM", None);
     }

compiler/rustc_llvm/llvm-wrapper/PassWrapper.cpp

@@ -688,14 +688,20 @@ struct LLVMRustSanitizerOptions {
   bool SanitizeKernelAddressRecover;
 };
 
+// This symbol won't be available or used when Enzyme is not enabled
+#ifdef ENZYME
+extern "C" void registerEnzyme(llvm::PassBuilder &PB);
+#endif
+
 extern "C" LLVMRustResult LLVMRustOptimize(
     LLVMModuleRef ModuleRef, LLVMTargetMachineRef TMRef,
     LLVMRustPassBuilderOptLevel OptLevelRust, LLVMRustOptStage OptStage,
     bool IsLinkerPluginLTO, bool NoPrepopulatePasses, bool VerifyIR,
     bool LintIR, bool UseThinLTOBuffers, bool MergeFunctions, bool UnrollLoops,
     bool SLPVectorize, bool LoopVectorize, bool DisableSimplifyLibCalls,
-    bool EmitLifetimeMarkers, LLVMRustSanitizerOptions *SanitizerOptions,
-    const char *PGOGenPath, const char *PGOUsePath, bool InstrumentCoverage,
+    bool EmitLifetimeMarkers, bool RunEnzyme,
+    LLVMRustSanitizerOptions *SanitizerOptions, const char *PGOGenPath,
+    const char *PGOUsePath, bool InstrumentCoverage,
     const char *InstrProfileOutput, const char *PGOSampleUsePath,
     bool DebugInfoForProfiling, void *LlvmSelfProfiler,
     LLVMRustSelfProfileBeforePassCallback BeforePassCallback,
@@ -1010,6 +1016,18 @@ extern "C" LLVMRustResult LLVMRustOptimize(
     MPM.addPass(NameAnonGlobalPass());
   }
 
+  // now load "-enzyme" pass:
+#ifdef ENZYME
+  if (RunEnzyme) {
+    registerEnzyme(PB);
+    if (auto Err = PB.parsePassPipeline(MPM, "enzyme")) {
+      std::string ErrMsg = toString(std::move(Err));
+      LLVMRustSetLastError(ErrMsg.c_str());
+      return LLVMRustResult::Failure;
+    }
+  }
+#endif
+
   // Upgrade all calls to old intrinsics first.
   for (Module::iterator I = TheModule->begin(), E = TheModule->end(); I != E;)
     UpgradeCallsToIntrinsic(&*I++); // must be post-increment, as we remove

compiler/rustc_middle/src/ty/context.rs

@@ -1301,9 +1301,6 @@ pub struct TyCtxt<'tcx> {
     gcx: &'tcx GlobalCtxt<'tcx>,
 }
 
-// Explicitly implement `DynSync` and `DynSend` for `TyCtxt` to short circuit trait resolution.
-unsafe impl DynSend for TyCtxt<'_> {}
-unsafe impl DynSync for TyCtxt<'_> {}
 fn _assert_tcx_fields() {
     sync::assert_dyn_sync::<&'_ GlobalCtxt<'_>>();
     sync::assert_dyn_send::<&'_ GlobalCtxt<'_>>();