win32: use native threading APIs instead of pthreads

* use fiber local storage if NO_PTHREADS is defined
* use critical sections instead of mutexes

Contributing-author: Ben Viglietta <benvi@microsoft.com>

Author: DHowett

arc.m

@@ -13,9 +13,36 @@
 
 #ifndef NO_PTHREADS
 #include <pthread.h>
-pthread_key_t ARCThreadKey;
+typedef pthread_key_t arc_tls_key_t;
+#define arc_tls_store pthread_setspecific
+#define arc_tls_load pthread_getspecific
+#define TLS_CALLBACK(name) void name
+
+static inline pthread_key_t arc_tls_key_create(void(*cleanupFunction)(void*))
+{
+	pthread_key_t key;
+	pthread_key_create(&key, cleanupFunction);
+	return key;
+}
+#else // if defined(NO_PTHREADS)
+// We're using the Fiber-Local Storage APIs on Windows
+// because the TLS APIs won't pass app certification.
+// Additionally, the FLS API surface is 1:1 mapped to
+// the TLS API surface when fibers are not in use.
+#include <Windows.h>
+typedef DWORD arc_tls_key_t;
+#define arc_tls_store FlsSetValue
+#define arc_tls_load FlsGetValue
+#define TLS_CALLBACK(name) void WINAPI name
+
+static inline DWORD arc_tls_key_create(void WINAPI(*cleanupFunction)(void*))
+{
+	return FlsAlloc(cleanupFunction);
+}
 #endif
 
+arc_tls_key_t ARCThreadKey;
+
 extern void _NSConcreteMallocBlock;
 extern void _NSConcreteStackBlock;
 extern void _NSConcreteGlobalBlock;
@@ -59,17 +86,13 @@ - (void)release;
 
 static inline struct arc_tls* getARCThreadData(void)
 {
-#ifdef NO_PTHREADS
-	return NULL;
-#else
-	struct arc_tls *tls = pthread_getspecific(ARCThreadKey);
+	struct arc_tls *tls = arc_tls_load(ARCThreadKey);
 	if (NULL == tls)
 	{
 		tls = calloc(sizeof(struct arc_tls), 1);
-		pthread_setspecific(ARCThreadKey, tls);
+		arc_tls_store(ARCThreadKey, tls);
 	}
 	return tls;
-#endif
 }
 int count = 0;
 int poolCount = 0;
@@ -132,7 +155,7 @@ static void emptyPool(struct arc_tls *tls, id *stop)
 	//fprintf(stderr, "New insert: %p.  Stop: %p\n", tls->pool->insert, stop);
 }
 
-static void cleanupPools(struct arc_tls* tls)
+static TLS_CALLBACK(cleanupPools)(struct arc_tls* tls)
 {
 	if (tls->returnRetained)
 	{
@@ -515,9 +538,7 @@ PRIVATE void init_arc(void)
 {
 	weak_ref_initialize(&weakRefs, 128);
 	INIT_LOCK(weakRefLock);
-#ifndef NO_PTHREADS
-	pthread_key_create(&ARCThreadKey, (void(*)(void*))cleanupPools);
-#endif
+	ARCThreadKey = arc_tls_key_create((void(*)(void*))cleanupPools);
 }
 
 void* block_load_weak(void *block);

lock.h

@@ -10,11 +10,11 @@
 #define BOOL _WINBOOL
 #	include <windows.h>
 #undef BOOL
-typedef HANDLE mutex_t;
-#	define INIT_LOCK(x) x = CreateMutex(NULL, FALSE, NULL)
-#	define LOCK(x) WaitForSingleObject(*x, INFINITE)
-#	define UNLOCK(x) ReleaseMutex(*x)
-#	define DESTROY_LOCK(x) CloseHandle(*x)
+typedef CRITICAL_SECTION mutex_t;
+#	define INIT_LOCK(x) InitializeCriticalSection(&(x))
+#	define LOCK(x) EnterCriticalSection(x)
+#	define UNLOCK(x) LeaveCriticalSection(x)
+#	define DESTROY_LOCK(x) DeleteCriticalSection(x)
 #else
 
 #	include <pthread.h>