refactor: make replaceBulk truly bulk-processing, try to optimise a bit

miiverse extra load times gone from 15 seconds to 2 seconds, though that's mostly because I changed the search range 😎

Makefile

@@ -33,7 +33,7 @@ CFLAGS	:=	-g -Wall -O2 -ffunction-sections \
 
 CFLAGS	+=	$(INCLUDE) -D__WIIU__ -D__WUT__ -D__WUPS__
 
-CXXFLAGS	:= $(CFLAGS)
+CXXFLAGS	:= $(CFLAGS) -std=c++20
 
 ASFLAGS	:=	-g $(ARCH)
 LDFLAGS	=	-g $(ARCH) $(RPXSPECS) -Wl,-Map,$(notdir $*.map)

src/patches/olv_applet.cpp

@@ -21,7 +21,6 @@
 #include <optional>
 #include <coreinit/debug.h>
 #include <coreinit/filesystem.h>
-#include <coreinit/memory.h>
 #include <nsysnet/nssl.h>
 
 #include "ca_pem.h" // generated at buildtime
@@ -36,32 +35,32 @@ const char wave_new[] = {
         0x00, 0x00, 0x00, 0x00, 0x2E, 0x70, 0x72, 0x65, 0x74, 0x65, 0x6E, 0x64,
         0x6F, 0x2E, 0x63, 0x63, 0x00
 };
-
-const char miiverse_green_highlight[] = {
+const unsigned char miiverse_green_highlight[] = {
         0x82, 0xff, 0x05, 0xff, 0x82, 0xff, 0x05, 0xff, 0x1d, 0xff, 0x04, 0xff, 0x1d, 0xff, 0x04, 0xff
 };
-
-const char juxt_purple_highlight[] = {
+const unsigned char juxt_purple_highlight[] = {
         0x5d, 0x4a, 0x9a, 0xff, 0x5d, 0x4a, 0x9a, 0xff, 0x5d, 0x4a, 0x9a, 0xff, 0x5d, 0x4a, 0x9a, 0xff
 };
-
-const char miiverse_green_touch1[] = {
+const unsigned char miiverse_green_touch1[] = {
         0x94, 0xd9, 0x2a, 0x00, 0x57, 0xbd, 0x12, 0xff
 };
-
-const char juxt_purple_touch1[] = {
+const unsigned char juxt_purple_touch1[] = {
         0x5d, 0x4a, 0x9a, 0x00, 0x5d, 0x4a, 0x9a, 0xff
 };
-
-const char miiverse_green_touch2[] = {
+const unsigned char miiverse_green_touch2[] = {
         0x57, 0xbd, 0x12, 0x00, 0x94, 0xd9, 0x2a, 0xff
 };
-
-const char juxt_purple_touch2[] = {
+const unsigned char juxt_purple_touch2[] = {
         0x5d, 0x4a, 0x9a, 0x00, 0x5d, 0x4a, 0x9a, 0xff
 };
 
-static std::optional <FSFileHandle> rootca_pem_handle{};
+const replacement replacements[] = {
+        {miiverse_green_highlight, juxt_purple_highlight},
+        {miiverse_green_touch1,    juxt_purple_touch1},
+        {miiverse_green_touch2,    juxt_purple_touch2},
+};
+
+static std::optional<FSFileHandle> rootca_pem_handle{};
 
 DECL_FUNCTION(int, FSOpenFile, FSClient *client, FSCmdBlock *block, char *path, const char *mode, uint32_t *handle,
               int error) {
@@ -101,17 +100,11 @@ DECL_FUNCTION(FSStatus, FSReadFile, FSClient *client, FSCmdBlock *block, uint8_t
     }
 
     if (rootca_pem_handle && *rootca_pem_handle == handle) {
-        memset(buffer, 0, size);
-        strcpy((char *) buffer, (const char *) ca_pem);
+        strlcpy((char *) buffer, (const char *) ca_pem, size * count);
 
         //this can't be done above (in the FSOpenFile hook) since it's not loaded yet.
-        replaceBulk(0x10000000, 0x10000000, miiverse_green_highlight, sizeof(miiverse_green_highlight),
-                    juxt_purple_highlight, sizeof(juxt_purple_highlight));
-        replaceBulk(0x10000000, 0x10000000, miiverse_green_touch1, sizeof(miiverse_green_touch1), juxt_purple_touch1,
-                    sizeof(juxt_purple_touch1));
-        replaceBulk(0x10000000, 0x10000000, miiverse_green_touch2, sizeof(miiverse_green_touch2), juxt_purple_touch2,
-                    sizeof(juxt_purple_touch2));
-
+        //the hardcoded offsets suck but they really are at Random Places In The Heap
+        replaceBulk(0x11000000, 0x02000000, replacements);
         return (FSStatus) count;
     }
 

src/utils/replace_mem.cpp

@@ -17,6 +17,7 @@
 
 #include <kernel/kernel.h>
 #include <coreinit/memorymap.h>
+#include <algorithm>
 
 bool replace(uint32_t start, uint32_t size, const char* original_val, size_t original_val_sz, const char* new_val, size_t new_val_sz) {
     for (uint32_t addr = start; addr < start + size - original_val_sz; addr++) {
@@ -32,13 +33,34 @@ bool replace(uint32_t start, uint32_t size, const char* original_val, size_t ori
     return false;
 }
 
-void replaceBulk(uint32_t start, uint32_t size, const char* original_val, size_t original_val_sz, const char* new_val, size_t new_val_sz) {
-    for (uint32_t addr = start; addr < start + size - original_val_sz; addr++) {
-        int ret = memcmp(original_val, (void*)addr, original_val_sz);
-        if (ret == 0) {
-            DEBUG_FUNCTION_LINE("found bulk @%08x", addr);
-            KernelCopyData(OSEffectiveToPhysical(addr), OSEffectiveToPhysical((uint32_t)new_val), new_val_sz);
-            DEBUG_FUNCTION_LINE("new bulk   @%08x", addr);
+void replaceBulk(uint32_t start, uint32_t size, std::span<const replacement> replacements) {
+    // work out the biggest input replacement
+    auto max_sz = std::max_element(replacements.begin(), replacements.end(), [](auto& a, auto& b) {
+        return a.orig.size_bytes() < b.orig.size_bytes();
+    })->orig.size_bytes();
+
+    int counts[replacements.size()];
+    for (auto& c : counts) {
+        c = 0;
+    }
+
+    for (uint32_t addr = start; addr < start + size - max_sz; addr++) {
+        for (int i = 0; i < (int)replacements.size(); i++) {
+            const auto& replacement = replacements[i];
+
+            int ret = memcmp((void*)addr, replacement.orig.data(), replacement.orig.size_bytes());
+            if (ret == 0) {
+                KernelCopyData(
+                        OSEffectiveToPhysical(addr),
+                        OSEffectiveToPhysical((uint32_t)replacement.repl.data()),
+                        replacement.repl.size_bytes()
+                );
+                counts[i]++;
+                break; // don't check the other replacements
+            }
         }
     }
+    for (auto c : counts) {
+        DEBUG_FUNCTION_LINE("replaced %d times", c);
+    }
 }

src/utils/replace_mem.h

@@ -16,6 +16,13 @@
 
 #include <cstdint>
 #include <cstddef>
+#include <span>
 
 bool replace(uint32_t start, uint32_t size, const char* original_val, size_t original_val_sz, const char* new_val, size_t new_val_sz);
-void replaceBulk(uint32_t start, uint32_t size, const char* original_val, size_t original_val_sz, const char* new_val, size_t new_val_sz);
+
+struct replacement {
+    std::span<const uint8_t> orig;
+    std::span<const uint8_t> repl;
+};
+
+void replaceBulk(uint32_t start, uint32_t size, std::span<const replacement> replacements);