Use platform-specific yield/pause instructions

4 files changed, 38 insertions, 20 deletions

diff --git a/src/pthreads.c b/src/pthreads.c
index 2d945a0..430ca79 100644
--- a/src/pthreads.c
+++ b/src/pthreads.c
@@ -108,8 +108,7 @@ static void wait_worker_threads(struct pthreadpool* threadpool) {
 
 	/* Spin-wait */
 	for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) {
-		/* This fence serves as a sleep instruction */
-		pthreadpool_fence_acquire();
+		pthreadpool_yield();
 
 		#if PTHREADPOOL_USE_FUTEX
 			has_active_threads = pthreadpool_load_acquire_uint32_t(&threadpool->has_active_threads);
@@ -151,8 +150,7 @@ static uint32_t wait_for_new_command(
 	if ((last_flags & PTHREADPOOL_FLAG_YIELD_WORKERS) == 0) {
 		/* Spin-wait loop */
 		for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) {
-			/* This fence serves as a sleep instruction */
-			pthreadpool_fence_acquire();
+			pthreadpool_yield();
 
 			command = pthreadpool_load_acquire_uint32_t(&threadpool->command);
 			if (command != last_command) {
diff --git a/src/threadpool-atomics.h b/src/threadpool-atomics.h
index 135166f..a8491e4 100644
--- a/src/threadpool-atomics.h
+++ b/src/threadpool-atomics.h
@@ -4,12 +4,19 @@
 #include <stddef.h>
 #include <stdint.h>
 
+/* SSE-specific headers */
+#if defined(__i386__) || defined(__i686__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
+	#include <xmmintrin.h>
+#endif
+
+/* ARM-specific headers */
+#if defined(__ARM_ACLE)
+	#include <arm_acle.h>
+#endif
+
 /* MSVC-specific headers */
 #ifdef _MSC_VER
 	#include <intrin.h>
-	#if defined(_M_IX86) || defined(_M_X64) || defined(_M_AMD64)
-		#include <immintrin.h>
-	#endif
 #endif
 
 
@@ -123,7 +130,7 @@
 	static inline void pthreadpool_fence_release() {
 		__c11_atomic_thread_fence(__ATOMIC_RELEASE);
 	}
-#elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_AMD64))
+#elif defined(_MSC_VER) && defined(_M_X64)
 	typedef volatile uint32_t pthreadpool_atomic_uint32_t;
 	typedef volatile size_t   pthreadpool_atomic_size_t;
 	typedef void *volatile    pthreadpool_atomic_void_p;
@@ -701,3 +708,21 @@
 #else
 	#error "Platform-specific implementation of threadpool-atomics.h required"
 #endif
+
+#if defined(__i386__) || defined(__i686__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
+	static inline void pthreadpool_yield() {
+		_mm_pause();
+	}
+#elif defined(__ARM_ACLE) || defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
+	static inline void pthreadpool_yield() {
+		__yield();
+	}
+#elif defined(__GNUC__) && (defined(__arm__) || defined(__aarch64__))
+	static inline void pthreadpool_yield() {
+		__asm__ __volatile__("yield");
+	}
+#else
+	static inline void pthreadpool_yield() {
+		pthreadpool_fence_acquire();
+	}
+#endif
diff --git a/src/threadpool-utils.h b/src/threadpool-utils.h
index d95e3a5..91e2445 100644
--- a/src/threadpool-utils.h
+++ b/src/threadpool-utils.h
@@ -4,21 +4,18 @@
 #include <stddef.h>
 
 /* SSE-specific headers */
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	#include <xmmintrin.h>
 #endif
 
 /* MSVC-specific headers */
 #if defined(_MSC_VER)
 	#include <intrin.h>
-	#if defined(_M_IX86) || defined(_M_X64) || defined(_M_AMD64)
-		#include <immintrin.h>
-	#endif
 #endif
 
 
 struct fpu_state {
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	uint32_t mxcsr;
 #elif defined(__GNUC__) && defined(__arm__) && defined(__ARM_FP) && (__ARM_FP != 0) || defined(_MSC_VER) && defined(_M_ARM)
 	uint32_t fpscr;
@@ -31,7 +28,7 @@ struct fpu_state {
 
 static inline struct fpu_state get_fpu_state() {
 	struct fpu_state state = { 0 };
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	state.mxcsr = (uint32_t) _mm_getcsr();
 #elif defined(_MSC_VER) && defined(_M_ARM)
 	state.fpscr = (uint32_t) _MoveFromCoprocessor(10, 7, 1, 0, 0);
@@ -46,7 +43,7 @@ static inline struct fpu_state get_fpu_state() {
 }
 
 static inline void set_fpu_state(const struct fpu_state state) {
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	_mm_setcsr((unsigned int) state.mxcsr);
 #elif defined(_MSC_VER) && defined(_M_ARM)
 	_MoveToCoprocessor((int) state.fpscr, 10, 7, 1, 0, 0);
@@ -60,7 +57,7 @@ static inline void set_fpu_state(const struct fpu_state state) {
 }
 
 static inline void disable_fpu_denormals() {
-#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
+#if defined(__SSE__) || defined(__x86_64__) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1)
 	_mm_setcsr(_mm_getcsr() | 0x8040);
 #elif defined(_MSC_VER) && defined(_M_ARM)
 	int fpscr = _MoveFromCoprocessor(10, 7, 1, 0, 0);
diff --git a/src/windows.c b/src/windows.c
index 19e534f..c9b88f7 100644
--- a/src/windows.c
+++ b/src/windows.c
@@ -35,8 +35,7 @@ static void wait_worker_threads(struct pthreadpool* threadpool, uint32_t event_i
 
 	/* Spin-wait */
 	for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) {
-		/* This fence serves as a sleep instruction */
-		pthreadpool_fence_acquire();
+		pthreadpool_yield();
 
 		active_threads = pthreadpool_load_acquire_size_t(&threadpool->active_threads);
 		if (active_threads == 0) {
@@ -63,8 +62,7 @@ static uint32_t wait_for_new_command(
 	if ((last_flags & PTHREADPOOL_FLAG_YIELD_WORKERS) == 0) {
 		/* Spin-wait loop */
 		for (uint32_t i = PTHREADPOOL_SPIN_WAIT_ITERATIONS; i != 0; i--) {
-			/* This fence serves as a sleep instruction */
-			pthreadpool_fence_acquire();
+			pthreadpool_yield();
 
 			command = pthreadpool_load_acquire_uint32_t(&threadpool->command);
 			if (command != last_command) {