Merge pull request #99168 from RandomShaper/even_better_spinlock

SpinLock: Overhaul false sharing prevention

core/os/spin_lock.h

@@ -31,8 +31,15 @@
 #ifndef SPIN_LOCK_H
 #define SPIN_LOCK_H
 
+#include "core/os/thread.h"
 #include "core/typedefs.h"
 
+#ifdef THREADS_ENABLED
+
+// Note the implementations below avoid false sharing by ensuring their
+// sizes match the assumed cache line. We can't use align attributes
+// because these objects may end up unaligned in semi-tightly packed arrays.
+
 #ifdef _MSC_VER
 #include <intrin.h>
 #endif
@@ -42,7 +49,10 @@
 #include <os/lock.h>
 
 class SpinLock {
-	mutable os_unfair_lock _lock = OS_UNFAIR_LOCK_INIT;
+	union {
+		mutable os_unfair_lock _lock = OS_UNFAIR_LOCK_INIT;
+		char aligner[Thread::CACHE_LINE_BYTES];
+	};
 
 public:
 	_ALWAYS_INLINE_ void lock() const {
@@ -54,9 +64,7 @@ public:
 	}
 };
 
-#else
-
-#include "core/os/thread.h"
+#else // __APPLE__
 
 #include <atomic>
 
@@ -84,8 +92,11 @@ _ALWAYS_INLINE_ static void _cpu_pause() {
 
 static_assert(std::atomic_bool::is_always_lock_free);
 
-class alignas(Thread::CACHE_LINE_BYTES) SpinLock {
-	mutable std::atomic<bool> locked = ATOMIC_VAR_INIT(false);
+class SpinLock {
+	union {
+		mutable std::atomic<bool> locked = ATOMIC_VAR_INIT(false);
+		char aligner[Thread::CACHE_LINE_BYTES];
+	};
 
 public:
 	_ALWAYS_INLINE_ void lock() const {
@@ -107,4 +118,14 @@ public:
 
 #endif // __APPLE__
 
+#else // THREADS_ENABLED
+
+class SpinLock {
+public:
+	void lock() const {}
+	void unlock() const {}
+};
+
+#endif // THREADS_ENABLED
+
 #endif // SPIN_LOCK_H