diff --git a/.gitattributes b/.gitattributes
index 1534e778..0332e031 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -9,3 +9,4 @@
 *.patch binary
 *.dll binary
 *.lib binary
+*.exe binary
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 29243b4f..353d171b 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -337,6 +337,7 @@ if (MI_BUILD_STATIC)
   endif()
 
   install(TARGETS mimalloc-static EXPORT mimalloc DESTINATION ${mi_install_objdir} LIBRARY)
+  install(EXPORT mimalloc DESTINATION ${mi_install_cmakedir})
 endif()
 
 # install include files
diff --git a/bin/mimalloc-redirect.dll b/bin/mimalloc-redirect.dll
index b7bf1d09..83b6bd4f 100644
Binary files a/bin/mimalloc-redirect.dll and b/bin/mimalloc-redirect.dll differ
diff --git a/bin/mimalloc-redirect.lib b/bin/mimalloc-redirect.lib
index 550db8ec..059fb870 100644
Binary files a/bin/mimalloc-redirect.lib and b/bin/mimalloc-redirect.lib differ
diff --git a/bin/mimalloc-redirect32.dll b/bin/mimalloc-redirect32.dll
index 7ba303af..2892d459 100644
Binary files a/bin/mimalloc-redirect32.dll and b/bin/mimalloc-redirect32.dll differ
diff --git a/bin/mimalloc-redirect32.lib b/bin/mimalloc-redirect32.lib
index 66173060..7dadab3d 100644
Binary files a/bin/mimalloc-redirect32.lib and b/bin/mimalloc-redirect32.lib differ
diff --git a/bin/minject.exe b/bin/minject.exe
new file mode 100644
index 00000000..e576c71f
Binary files /dev/null and b/bin/minject.exe differ
diff --git a/bin/minject32.exe b/bin/minject32.exe
new file mode 100644
index 00000000..1eb8a75d
Binary files /dev/null and b/bin/minject32.exe differ
diff --git a/include/mimalloc-types.h b/include/mimalloc-types.h
index 957115c8..41364286 100644
--- a/include/mimalloc-types.h
+++ b/include/mimalloc-types.h
@@ -447,9 +447,15 @@ struct mi_heap_s {
 // 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
diff --git a/src/init.c b/src/init.c
index 80cc5c5b..e691df7f 100644
--- a/src/init.c
+++ b/src/init.c
@@ -520,14 +520,14 @@ void mi_process_init(void) mi_attr_noexcept {
   #endif
   _mi_verbose_message("secure level: %d\n", MI_SECURE);
   mi_thread_init();
-#if defined(_WIN32) && !defined(MI_SHARED_LIB)
-  /* When building as a static lib the FLS cleanup happens to early for the main thread.
-   * To avoid that set the FLS value for the main thread to NULL; the eventual
-   * mi_fls_done() execution won't call _mi_thread_done().
-   * The latter function is later called explicitly from mi_process_done(). 
-   * See GitHub issue #508 for more background and explanation. */
+
+  #if defined(_WIN32) && !defined(MI_SHARED_LIB)
+  // 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.
   FlsSetValue(mi_fls_key, NULL);
-#endif
+  #endif
+
   mi_stats_reset();  // only call stat reset *after* thread init (or the heap tld == NULL)
 
   if (mi_option_is_enabled(mi_option_reserve_huge_os_pages)) {
@@ -557,9 +557,7 @@ static void mi_process_done(void) {
   process_done = true;
 
   #if defined(_WIN32) && !defined(MI_SHARED_LIB)
-  // Explicitly clean up main thread. See comment in mi_process_init() for reason
-  _mi_thread_done(_mi_heap_default);
-  FlsFree(mi_fls_key);            // call thread-done on all threads to prevent dangling callback pointer if statically linked with a DLL; Issue #208
+  FlsFree(mi_fls_key);  // call thread-done on all threads (except the main thread) to prevent dangling callback pointer if statically linked with a DLL; Issue #208
   #endif
   
   #if (MI_DEBUG != 0) || !defined(MI_SHARED_LIB)  
@@ -600,7 +598,7 @@ static void mi_process_done(void) {
     mi_process_load();
     return 0;
   }
-  typedef int(*_crt_cb)(void);
+  typedef int(*_mi_crt_callback_t)(void);
   #if defined(_M_X64) || defined(_M_ARM64)
     __pragma(comment(linker, "/include:" "_mi_msvc_initu"))
     #pragma section(".CRT$XIU", long, read)
@@ -608,7 +606,7 @@ static void mi_process_done(void) {
     __pragma(comment(linker, "/include:" "__mi_msvc_initu"))
   #endif
   #pragma data_seg(".CRT$XIU")
-  extern "C" _crt_cb _mi_msvc_initu[] = { &_mi_process_init };
+  extern "C" _mi_crt_callback_t _mi_msvc_initu[] = { &_mi_process_init };
   #pragma data_seg()
 
 #elif defined(__cplusplus)
diff --git a/src/region.c b/src/region.c
index f864f73b..2d73025e 100644
--- a/src/region.c
+++ b/src/region.c
@@ -94,7 +94,7 @@ typedef struct mem_region_s {
   mi_bitmap_field_t         commit;      // track if committed per block
   mi_bitmap_field_t         reset;       // track if reset per block
   _Atomic(size_t)           arena_memid; // if allocated from a (huge page) arena
-  size_t                    padding;     // round to 8 fields
+  _Atomic(size_t)           padding;     // round to 8 fields (needs to be atomic for msvc, see issue #508)
 } mem_region_t;
 
 // The region map
diff --git a/test/main-override.cpp b/test/main-override.cpp
index 8834f2c7..e0dba5a3 100644
--- a/test/main-override.cpp
+++ b/test/main-override.cpp
@@ -50,7 +50,7 @@ int main() {
    tsan_numa_test();
    strdup_test();
 
-  //test_mt_shutdown();
+  test_mt_shutdown();
   //fail_aslr();
   bench_alloc_large();
   mi_stats_print(NULL);
@@ -76,7 +76,7 @@ public:
 static void various_tests() {
   atexit(free_p);
   void* p1 = malloc(78);
-  void* p2 = mi_malloc_aligned(16, 24);
+  void* p2 = mi_malloc_aligned(24, 16);
   free(p1);
   p1 = malloc(8);
   char* s = mi_strdup("hello\n");