diff --git a/CMakeLists.txt b/CMakeLists.txt
index 76544a01..4a57a3ae 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -15,6 +15,7 @@ option(MI_USE_CXX           "Use the C++ compiler to compile the library (instea
 option(MI_SEE_ASM           "Generate assembly files" OFF)
 option(MI_OSX_INTERPOSE     "Use interpose to override standard malloc on macOS" ON)
 option(MI_OSX_ZONE          "Use malloc zone to override standard malloc on macOS" ON) 
+option(MI_WIN_REDIRECT      "Use redirection module ('mimalloc-redirect') on Windows if compiling mimalloc as a DLL" ON)
 option(MI_LOCAL_DYNAMIC_TLS "Use slightly slower, dlopen-compatible TLS mechanism (Unix)" OFF)
 option(MI_BUILD_SHARED      "Build shared library" ON)
 option(MI_BUILD_STATIC      "Build static library" ON)
@@ -105,6 +106,19 @@ if(MI_OVERRIDE)
   endif()
 endif()
 
+if(WIN32)
+  if (MI_WIN_REDIRECT)
+    if (MSVC_C_ARCHITECTURE_ID MATCHES "ARM")
+      message(STATUS "Cannot use redirection on Windows ARM (MI_WIN_REDIRECT=OFF)")
+      set(MI_WIN_REDIRECT OFF)
+    endif()
+  endif()
+  if (NOT MI_WIN_REDIRECT)
+    # use a negative define for backward compatibility
+    list(APPEND mi_defines MI_WIN_NOREDIRECT=1)
+  endif()
+endif()
+
 if(MI_SECURE)
   message(STATUS "Set full secure build (MI_SECURE=ON)")
   list(APPEND mi_defines MI_SECURE=4)
@@ -323,8 +337,8 @@ if(MI_BUILD_SHARED)
       $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
       $<INSTALL_INTERFACE:${mi_install_incdir}>
   )
-  if(WIN32)
-    # On windows copy the mimalloc redirection dll too.
+  if(WIN32 AND MI_WIN_REDIRECT)
+    # On windows, link and copy the mimalloc redirection dll too.
     if(CMAKE_SIZEOF_VOID_P EQUAL 4)
       set(MIMALLOC_REDIRECT_SUFFIX "32")
     else()
diff --git a/src/init.c b/src/init.c
index 19124afe..66e71ef2 100644
--- a/src/init.c
+++ b/src/init.c
@@ -475,7 +475,7 @@ void _mi_heap_set_default_direct(mi_heap_t* heap)  {
 // --------------------------------------------------------
 // Run functions on process init/done, and thread init/done
 // --------------------------------------------------------
-static void mi_process_done(void);
+static void mi_cdecl mi_process_done(void);
 
 static bool os_preloading = true;    // true until this module is initialized
 static bool mi_redirected = false;   // true if malloc redirects to mi_malloc
@@ -490,7 +490,7 @@ mi_decl_nodiscard bool mi_is_redirected(void) mi_attr_noexcept {
 }
 
 // Communicate with the redirection module on Windows
-#if defined(_WIN32) && defined(MI_SHARED_LIB)
+#if defined(_WIN32) && defined(MI_SHARED_LIB) && !defined(MI_WIN_NOREDIRECT)
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -506,8 +506,8 @@ mi_decl_export void _mi_redirect_entry(DWORD reason) {
     mi_thread_done();
   }
 }
-__declspec(dllimport) bool mi_allocator_init(const char** message);
-__declspec(dllimport) void mi_allocator_done(void);
+__declspec(dllimport) bool mi_cdecl mi_allocator_init(const char** message);
+__declspec(dllimport) void mi_cdecl mi_allocator_done(void);
 #ifdef __cplusplus
 }
 #endif
@@ -606,7 +606,7 @@ void mi_process_init(void) mi_attr_noexcept {
 }
 
 // Called when the process is done (through `at_exit`)
-static void mi_process_done(void) {
+static void mi_cdecl mi_process_done(void) {
   // only shutdown if we were initialized
   if (!_mi_process_is_initialized) return;
   // ensure we are called once
diff --git a/src/options.c b/src/options.c
index 56a08cff..367bc0d2 100644
--- a/src/options.c
+++ b/src/options.c
@@ -170,7 +170,7 @@ void mi_option_disable(mi_option_t option) {
 }
 
 
-static void mi_out_stderr(const char* msg, void* arg) {
+static void mi_cdecl mi_out_stderr(const char* msg, void* arg) {
   MI_UNUSED(arg);
   if (msg == NULL) return;
   #ifdef _WIN32
@@ -203,7 +203,7 @@ static void mi_out_stderr(const char* msg, void* arg) {
 static char out_buf[MI_MAX_DELAY_OUTPUT+1];
 static _Atomic(size_t) out_len;
 
-static void mi_out_buf(const char* msg, void* arg) {
+static void mi_cdecl mi_out_buf(const char* msg, void* arg) {
   MI_UNUSED(arg);
   if (msg==NULL) return;
   if (mi_atomic_load_relaxed(&out_len)>=MI_MAX_DELAY_OUTPUT) return;
@@ -235,7 +235,7 @@ static void mi_out_buf_flush(mi_output_fun* out, bool no_more_buf, void* arg) {
 
 // Once this module is loaded, switch to this routine
 // which outputs to stderr and the delayed output buffer.
-static void mi_out_buf_stderr(const char* msg, void* arg) {
+static void mi_cdecl mi_out_buf_stderr(const char* msg, void* arg) {
   mi_out_stderr(msg,arg);
   mi_out_buf(msg,arg);
 }
diff --git a/src/stats.c b/src/stats.c
index 134a7bcb..f82c7c67 100644
--- a/src/stats.c
+++ b/src/stats.c
@@ -267,7 +267,7 @@ static void mi_buffered_flush(buffered_t* buf) {
   buf->used = 0;
 }
 
-static void mi_buffered_out(const char* msg, void* arg) {
+static void mi_cdecl mi_buffered_out(const char* msg, void* arg) {
   buffered_t* buf = (buffered_t*)arg;
   if (msg==NULL || buf==NULL) return;
   for (const char* src = msg; *src != 0; src++) {