Fix performance regression in Qthreads 1.21 by adjusting the exact ty…

…pe of fencing we do for spinlocks.
sandialabs · Oct 24, 2024 · ad176e5 · ad176e5
1 parent f2b88c5
commit ad176e5
Showing 1 changed file with 12 additions and 1 deletion.
diff --git a/include/qt_atomics.h b/include/qt_atomics.h
@@ -21,8 +21,19 @@
 #define THREAD_FENCE_MEM_RELEASE THREAD_FENCE_MEM_RELEASE_IMPL
 #endif
 
+#if QTHREAD_ASSEMBLY_ARCH == QTHREAD_IA32 ||                                   \
+  QTHREAD_ASSEMBLY_ARCH == QTHREAD_AMD64
 #define SPINLOCK_BODY()                                                        \
-  do { MACHINE_FENCE; } while (0)
+  do { __asm__ __volatile__("pause" ::: "memory"); } while (0)
+#elif QTHREAD_ASSEMBLY_ARCH == QTHREAD_ARM ||                                  \
+  QTHREAD_ASSEMBLY_ARCH == QTHREAD_ARMV8_A64
+#define SPINLOCK_BODY()                                                        \
+  do { __asm__ __volatile__("yield" ::: "memory"); } while (0)
+#else
+#define SPINLOCK_BODY()                                                        \
+  do { atomic_thread_fence(memory_order_acq_rel;                               \
+  } while (0)
+#endif
 
 #define QTHREAD_FASTLOCK_SETUP()                                               \
   do {                                                                         \