From 93906428798ff6f4e1b71bf2afc087f6920dddf6 Mon Sep 17 00:00:00 2001
From: daan <daan@microsoft.com>
Date: Wed, 3 Jul 2019 13:00:43 -0700
Subject: [PATCH] fix windows dynamic malloc overried when both ucrtbase and
 msvcrt are loaded; also fix virtualalloc2 on 32-bit

---
 src/alloc-override-win.c | 12 +++++++-----
 src/os.c                 |  2 +-
 test/main-override.cpp   |  2 +-
 3 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/src/alloc-override-win.c b/src/alloc-override-win.c
index 7158e920..9c091389 100644
--- a/src/alloc-override-win.c
+++ b/src/alloc-override-win.c
@@ -526,7 +526,7 @@ static void mi_module_resolve(HMODULE mod) {
   // see if any patches apply
   for (size_t i = 0; patches[i].name != NULL; i++) {
     mi_patch_t* patch = &patches[i];
-    if (!patch->applied && patch->original==NULL) {
+    if (!patch->applied && patch->original==NULL) { // we apply only the first found dll
       void* addr = GetProcAddress(mod, patch->name);
       if (addr != NULL) {
         // found it! set the address
@@ -554,9 +554,11 @@ static bool mi_patches_resolve(void) {
   size_t count = needed / sizeof(HMODULE);
   size_t ucrtbase_index = 0;
   size_t mimalloc_index = 0;
-  // iterate through the loaded modules
-  for (size_t i = 0; i < count; i++) {
-    HMODULE mod = modules[i];
+  // iterate through the loaded modules; do this from the end so we prefer the
+  // first loaded DLL as sometimes both "msvcr" and "ucrt" are both loaded and we should
+  // override "ucrt" in that situation.
+  for (size_t i = count; i > 0; i--) {
+    HMODULE mod = modules[i-1];
     char filename[MAX_PATH] = { 0 };
     DWORD slen = GetModuleFileName(mod, filename, MAX_PATH);
     if (slen > 0 && slen < MAX_PATH) {
@@ -566,7 +568,7 @@ static bool mi_patches_resolve(void) {
       const char* basename = (lastsep==NULL ? filename : lastsep+1);
       if (i==0                                    // main module to allow static crt linking
         || _strnicmp(basename, "ucrt", 4) == 0    // new ucrtbase.dll in windows 10
-        || _strnicmp(basename, "msvcr", 5) == 0)  // older runtimes
+        || _strnicmp(basename, "msvcr", 5) == 0)  // older runtimes      
       {
         // remember indices so we can check load order (in debug mode)
         if (_stricmp(basename, MIMALLOC_NAME) == 0) mimalloc_index = i;
diff --git a/src/os.c b/src/os.c
index d952726a..2957e0f9 100644
--- a/src/os.c
+++ b/src/os.c
@@ -62,7 +62,7 @@ static size_t mi_os_good_alloc_size(size_t size, size_t alignment) {
 #if defined(_WIN32)
 // 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.
-typedef PVOID (*VirtualAlloc2Ptr)(HANDLE, PVOID, SIZE_T, ULONG, ULONG, MEM_EXTENDED_PARAMETER*, ULONG );
+typedef PVOID (__stdcall *VirtualAlloc2Ptr)(HANDLE, PVOID, SIZE_T, ULONG, ULONG, MEM_EXTENDED_PARAMETER*, ULONG );
 static VirtualAlloc2Ptr pVirtualAlloc2 = NULL;
 
 void _mi_os_init(void) {
diff --git a/test/main-override.cpp b/test/main-override.cpp
index 406d3bea..e24ed6d3 100644
--- a/test/main-override.cpp
+++ b/test/main-override.cpp
@@ -13,7 +13,7 @@ void free_p() {
 }
 
 int main() {
-  mi_stats_reset();
+  mi_stats_reset();  
   atexit(free_p);
   void* p1 = malloc(78);
   void* p2 = malloc(24);