From c80b7cb3d01705badadc367031229fd413ece1e2 Mon Sep 17 00:00:00 2001
From: Luke Street <luke@street.dev>
Date: Fri, 15 Aug 2025 10:03:13 -0600
Subject: [PATCH] Integrate mimalloc (#88)

---
 CMakeLists.txt   | 19 ++++++++++++++++---
 Dockerfile       |  2 +-
 dll/kernel32.cpp | 27 ++++++++++++++++++---------
 3 files changed, 35 insertions(+), 13 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 6c89538..72b0c2c 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -5,14 +5,27 @@ if(NOT CMAKE_BUILD_TYPE)
         "Build type options: Debug Release RelWithDebInfo MinSizeRel" FORCE)
 endif()
 
+set(CMAKE_C_FLAGS_INIT "-m32")
+set(CMAKE_CXX_FLAGS_INIT "-m32")
+set(CMAKE_EXE_LINKER_FLAGS_INIT "-m32")
+set(CMAKE_SHARED_LINKER_FLAGS_INIT "-m32")
+
 project(wibo LANGUAGES CXX)
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -fno-pie -no-pie -D_LARGEFILE64_SOURCE")
+
 find_package(Filesystem REQUIRED)
 
-set(CMAKE_CXX_STANDARD 17)
-set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -m32 -Wall -fno-pie -no-pie -D_LARGEFILE64_SOURCE")
+include(FetchContent)
+FetchContent_Declare(
+    mimalloc
+    GIT_REPOSITORY https://github.com/microsoft/mimalloc.git
+    GIT_TAG        dfa50c37d951128b1e77167dd9291081aa88eea4  # v3.1.5
+)
+FetchContent_MakeAvailable(mimalloc)
 
 include_directories(.)
 add_executable(wibo
@@ -35,5 +48,5 @@ add_executable(wibo
     processes.cpp
     strutil.cpp
 )
-target_link_libraries(wibo PRIVATE std::filesystem)
+target_link_libraries(wibo PRIVATE std::filesystem mimalloc-static)
 install(TARGETS wibo DESTINATION bin)
diff --git a/Dockerfile b/Dockerfile
index 2fa409c..d618896 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -2,7 +2,7 @@
 FROM --platform=linux/i386 alpine:latest AS build
 
 # Install dependencies
-RUN apk add --no-cache cmake ninja g++ linux-headers binutils
+RUN apk add --no-cache cmake ninja g++ linux-headers binutils git
 
 # Copy source files
 COPY . /wibo
diff --git a/dll/kernel32.cpp b/dll/kernel32.cpp
index e030bdd..e7068c9 100644
--- a/dll/kernel32.cpp
+++ b/dll/kernel32.cpp
@@ -10,7 +10,7 @@
 #include <fnmatch.h>
 #include <string>
 #include "strutil.h"
-#include <malloc.h>
+#include <mimalloc.h>
 #include <random>
 #include <stdarg.h>
 #include <system_error>
@@ -46,9 +46,9 @@ namespace kernel32 {
 	static void *doAlloc(unsigned int dwBytes, bool zero) {
 		if (dwBytes == 0)
 			dwBytes = 1;
-		void *ret = malloc(dwBytes);
+		void *ret = mi_malloc_aligned(dwBytes, 8);
 		if (ret && zero) {
-			memset(ret, 0, malloc_usable_size(ret));
+			memset(ret, 0, mi_usable_size(ret));
 		}
 		return ret;
 	}
@@ -56,9 +56,9 @@ namespace kernel32 {
 	static void *doRealloc(void *mem, unsigned int dwBytes, bool zero) {
 		if (dwBytes == 0)
 			dwBytes = 1;
-		size_t oldSize = malloc_usable_size(mem);
-		void *ret = realloc(mem, dwBytes);
-		size_t newSize = malloc_usable_size(ret);
+		size_t oldSize = mi_usable_size(mem);
+		void *ret = mi_realloc_aligned(mem, dwBytes, 8);
+		size_t newSize = mi_usable_size(ret);
 		if (ret && zero && newSize > oldSize) {
 			memset((char*)ret + oldSize, 0, newSize - oldSize);
 		}
@@ -353,9 +353,17 @@ namespace kernel32 {
 
 	int WIN_FUNC InitOnceBeginInitialize(LPINIT_ONCE lpInitOnce, DWORD dwFlags, PBOOL fPending, LPVOID* lpContext) {
 		DEBUG_LOG("STUB: InitOnceBeginInitialize\n");
+		if (fPending != nullptr) {
+			*fPending = TRUE;
+		}
 		return 1;
 	}
 
+	BOOL WIN_FUNC InitOnceComplete(LPINIT_ONCE lpInitOnce, DWORD dwFlags, LPVOID lpContext) {
+		DEBUG_LOG("STUB: InitOnceComplete\n");
+		return TRUE;
+	}
+
 	void WIN_FUNC AcquireSRWLockShared(void *SRWLock) { DEBUG_LOG("STUB: AcquireSRWLockShared(%p)\n", SRWLock); }
 
 	void WIN_FUNC ReleaseSRWLockShared(void *SRWLock) { DEBUG_LOG("STUB: ReleaseSRWLockShared(%p)\n", SRWLock); }
@@ -485,7 +493,7 @@ namespace kernel32 {
 		bufSize++;
 
 		// Step 2, actually build that buffer
-		char *buffer = (char *) malloc(bufSize);
+		char *buffer = (char *) mi_malloc(bufSize);
 		char *ptr = buffer;
 		work = environ;
 
@@ -515,7 +523,7 @@ namespace kernel32 {
 		bufSizeW++;
 
 		// Step 2, actually build that buffer
-		uint16_t *buffer = (uint16_t *) malloc(bufSizeW * 2);
+		uint16_t *buffer = (uint16_t *) mi_malloc(bufSizeW * 2);
 		uint16_t *ptr = buffer;
 		work = environ;
 
@@ -1896,7 +1904,7 @@ namespace kernel32 {
 
 	unsigned int WIN_FUNC HeapSize(void *hHeap, unsigned int dwFlags, void *lpMem) {
 		DEBUG_LOG("HeapSize(heap=%p, flags=%x, mem=%p)\n", hHeap, dwFlags, lpMem);
-		return malloc_usable_size(lpMem);
+		return mi_usable_size(lpMem);
 	}
 
 	void *WIN_FUNC GetProcessHeap() {
@@ -2348,6 +2356,7 @@ static void *resolveByName(const char *name) {
 	if (strcmp(name, "EnterCriticalSection") == 0) return (void *) kernel32::EnterCriticalSection;
 	if (strcmp(name, "LeaveCriticalSection") == 0) return (void *) kernel32::LeaveCriticalSection;
 	if (strcmp(name, "InitOnceBeginInitialize") == 0) return (void *) kernel32::InitOnceBeginInitialize;
+	if (strcmp(name, "InitOnceComplete") == 0) return (void *) kernel32::InitOnceComplete;
 	if (strcmp(name, "AcquireSRWLockShared") == 0) return (void *) kernel32::AcquireSRWLockShared;
 	if (strcmp(name, "ReleaseSRWLockShared") == 0) return (void *) kernel32::ReleaseSRWLockShared;
 	if (strcmp(name, "AcquireSRWLockExclusive") == 0) return (void *) kernel32::AcquireSRWLockExclusive;