Fixed cpuinfo in LLVM GCC & Added AVX2 detection

Seems it only defines __builtin_cpu_supports but no __builtin_cpu_init
(Apple LLVM version 9.0.0 (clang-900.0.38))

3 files changed, 39 insertions, 15 deletions

diff --git a/Zend/zend.c b/Zend/zend.c
index dc0cab8cc0..e62f9e1a48 100644
--- a/Zend/zend.c
+++ b/Zend/zend.c
@@ -743,6 +743,7 @@ int zend_startup(zend_utility_functions *utility_functions, char **extensions) /
 #endif
 
 	zend_cpu_startup();
+	zend_cpu_supports(ZEND_CPU_FEATURE_AVX2);
 
 #ifdef ZEND_WIN32
 	php_win32_cp_set_by_id(65001);
diff --git a/Zend/zend_cpuinfo.c b/Zend/zend_cpuinfo.c
index 988ceb2521..43252a7cca 100644
--- a/Zend/zend_cpuinfo.c
+++ b/Zend/zend_cpuinfo.c
@@ -29,46 +29,53 @@ typedef struct _zend_cpu_info {
 static zend_cpu_info cpuinfo = {0};
 
 #if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
-static void __zend_cpuid(uint32_t func, uint32_t subfunc) {
+static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo) {
 	__asm__ __volatile__ (
 		"cpuid"
-		: "=a"(cpuinfo.eax), "=b"(cpuinfo.ebx), "=c"(cpuinfo.ecx), "=d"(cpuinfo.edx)
+		: "=a"(cpuinfo->eax), "=b"(cpuinfo->ebx), "=c"(cpuinfo->ecx), "=d"(cpuinfo->edx)
 		: "a"(func), "c"(subfunc)
 	);
 }
 #elif defined(ZEND_WIN32)
 # include <intrin.h>
-static void __zend_cpuid(uint32_t func, uint32_t subfunc) {
+static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo) {
 	int regs[4];
 
 	__cpuidex(regs, func, subfunc);
 
-	cpuinfo.eax = regs[0];
-	cpuinfo.ebx = regs[1];
-	cpuinfo.ecx = regs[2];
-	cpuinfo.edx = regs[3];
+	cpuinfo->eax = regs[0];
+	cpuinfo->ebx = regs[1];
+	cpuinfo->ecx = regs[2];
+	cpuinfo->edx = regs[3];
 }
 #else
-static void __zend_cpuid(uint32_t func, uint32_t subfunc) {
-	cpuinfo.eax = 0;
+static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo) {
+	cpuinfo->eax = 0;
 }
 #endif
 
 void zend_cpu_startup(void)
 {
 	if (!cpuinfo.initialized) {
+		zend_cpu_info ebx;
+
 		cpuinfo.initialized = 1;
-		__zend_cpuid(0, 0);
+		__zend_cpuid(0, 0, &cpuinfo);
 		if (cpuinfo.eax == 0) {
 			return;
 		}
-		__zend_cpuid(1, 0);
+		__zend_cpuid(1, 0, &cpuinfo);
+		/* for avx2 */
+		__zend_cpuid(7, 0, &ebx);
+		cpuinfo.ebx = ebx.ebx;
 	}
 }
 
 ZEND_API int zend_cpu_supports(zend_cpu_feature feature) {
 	if (feature & ZEND_CPU_EDX_MASK) {
 		return (cpuinfo.edx & (feature & ~ZEND_CPU_EDX_MASK));
+	} else if (feature & ZEND_CPU_EBX_MASK) {
+		return (cpuinfo.ebx & (feature & ~ZEND_CPU_EBX_MASK));
 	} else {
 		return (cpuinfo.ecx & feature);
 	}
diff --git a/Zend/zend_cpuinfo.h b/Zend/zend_cpuinfo.h
index c077366bd8..d0d3a936e1 100644
--- a/Zend/zend_cpuinfo.h
+++ b/Zend/zend_cpuinfo.h
@@ -21,6 +21,7 @@
 
 #include "zend.h"
 
+#define ZEND_CPU_EBX_MASK     (1<<30)
 #define ZEND_CPU_EDX_MASK     (1<<31)
 
 typedef enum _zend_cpu_feature {
@@ -55,9 +56,12 @@ typedef enum _zend_cpu_feature {
 	ZEND_CPU_FEATURE_OSXSAVE		= (1<<27) ,
 	ZEND_CPU_FEATURE_AVX			= (1<<28),
 	ZEND_CPU_FEATURE_F16C			= (1<<29),
-	ZEND_CPU_FEATURE_RDRAND			= (1<<30),
+	/* intentionally don't support	= (1<<30) */
 	/* intentionally don't support	= (1<<31) */
 
+	/* EBX */
+	ZEND_CPU_FEATURE_AVX2			= (1<<5 | ZEND_CPU_EBX_MASK),
+
 	/* EDX */
 	ZEND_CPU_FEATURE_FPU			= (1<<0 | ZEND_CPU_EDX_MASK),
 	ZEND_CPU_FEATURE_VME			= (1<<1 | ZEND_CPU_EDX_MASK),
@@ -88,47 +92,59 @@ typedef enum _zend_cpu_feature {
 	ZEND_CPU_FEATURE_SSE2			= (1<<26 | ZEND_CPU_EDX_MASK),
 	ZEND_CPU_FEATURE_SS				= (1<<27 | ZEND_CPU_EDX_MASK),
 	ZEND_CPU_FEATURE_HT				= (1<<28 | ZEND_CPU_EDX_MASK),
-	ZEND_CPU_FEATURE_TM				= (1<<29 | ZEND_CPU_EDX_MASK),
-	ZEND_CPU_FEATURE_IA64			= (1<<30 | ZEND_CPU_EDX_MASK)
+	ZEND_CPU_FEATURE_TM				= (1<<29 | ZEND_CPU_EDX_MASK)
+	/*intentionally don't support   = (1<<30 | ZEND_CPU_EDX_MASK)*/
 	/*intentionally don't support   = (1<<31 | ZEND_CPU_EDX_MASK)*/
 } zend_cpu_feature;
 
 void zend_cpu_startup();
 ZEND_API int zend_cpu_supports(zend_cpu_feature feature);
 
-#ifdef PHP_HAVE_BUILTIN_CPU_SUPPORTS
+#if PHP_HAVE_BUILTIN_CPU_SUPPORTS
 /* NOTE: you should use following inline function in
  * resolver functions (ifunc), as it could be called
  * before all PLT symbols are resloved. in other words,
  * resolver functions should not depends any external
  * functions */
 static zend_always_inline int zend_cpu_support_sse2() {
+#if PHP_HAVE_BUILTIN_CPU_INIT
 	__builtin_cpu_init();
+#endif
 	return __builtin_cpu_supports("sse2");
 }
 
 static zend_always_inline int zend_cpu_support_sse3() {
+#if PHP_HAVE_BUILTIN_CPU_INIT
 	__builtin_cpu_init();
+#endif
 	return __builtin_cpu_supports("sse3");
 }
 
 static zend_always_inline int zend_cpu_support_sse41() {
+#if PHP_HAVE_BUILTIN_CPU_INIT
 	__builtin_cpu_init();
+#endif
 	return __builtin_cpu_supports("sse4.1");
 }
 
 static zend_always_inline int zend_cpu_support_sse42() {
+#if PHP_HAVE_BUILTIN_CPU_INIT
 	__builtin_cpu_init();
+#endif
 	return __builtin_cpu_supports("sse4.2");
 }
 
 static zend_always_inline int zend_cpu_support_avx() {
+#if PHP_HAVE_BUILTIN_CPU_INIT
 	__builtin_cpu_init();
+#endif
 	return __builtin_cpu_supports("avx");
 }
 
 static zend_always_inline int zend_cpu_support_avx2() {
+#if PHP_HAVE_BUILTIN_CPU_INIT
 	__builtin_cpu_init();
+#endif
 	return __builtin_cpu_supports("avx2");
 }
 #else