fast_copy.diff

diff --git a/ext/opcache/ZendAccelerator.c b/ext/opcache/ZendAccelerator.c
index 1c870e9..1c2c41f 100644
--- a/ext/opcache/ZendAccelerator.c
+++ b/ext/opcache/ZendAccelerator.c
@@ -1358,10 +1358,10 @@ static zend_persistent_script *store_script_in_file_cache(zend_persistent_script
 	memory_used = zend_accel_script_persist_calc(new_persistent_script, NULL, 0, 0);

 	/* Allocate memory block */
-#ifdef __SSE2__
+#if defined(__AVX__) || defined(__SSE2__)
 	/* Align to 64-byte boundary */
-	ZCG(mem) = zend_arena_alloc(&CG(arena), memory_used + 64);
-	ZCG(mem) = (void*)(((zend_uintptr_t)ZCG(mem) + 63L) & ~63L);
+	ZCG(mem) = zend_arena_alloc(&CG(arena), memory_used + FAST_COPY_PAD);
+	ZCG(mem) = (void*)(((zend_uintptr_t)ZCG(mem) + FAST_COPY_MASK) & ~FAST_COPY_MASK);
 #else
 	ZCG(mem) = zend_arena_alloc(&CG(arena), memory_used);
 #endif
@@ -1465,12 +1465,12 @@ static zend_persistent_script *cache_script_in_shared_memory(zend_persistent_scr
 	memory_used = zend_accel_script_persist_calc(new_persistent_script, key, key_length, 1);

 	/* Allocate shared memory */
-#ifdef __SSE2__
+#if defined(__AVX__) || defined(__SSE2__)
 	/* Align to 64-byte boundary */
-	ZCG(mem) = zend_shared_alloc(memory_used + 64);
+	ZCG(mem) = zend_shared_alloc(memory_used + FAST_COPY_PAD);
 	if (ZCG(mem)) {
 		memset(ZCG(mem), 0, memory_used + 64);
-		ZCG(mem) = (void*)(((zend_uintptr_t)ZCG(mem) + 63L) & ~63L);
+		ZCG(mem) = (void*)(((zend_uintptr_t)ZCG(mem) + FAST_COPY_MASK) & ~FAST_COPY_MASK);
 	}
 #else
 	ZCG(mem) = zend_shared_alloc(memory_used);
diff --git a/ext/opcache/zend_accelerator_util_funcs.c b/ext/opcache/zend_accelerator_util_funcs.c
index bd46a86..5acbc88 100644
--- a/ext/opcache/zend_accelerator_util_funcs.c
+++ b/ext/opcache/zend_accelerator_util_funcs.c
@@ -596,8 +596,86 @@ static void zend_accel_class_hash_copy(HashTable *target, HashTable *source, uni
 	return;
 }

-#ifdef __SSE2__
-# include <mmintrin.h>
+#if defined(__AVX__)
+# include <nmmintrin.h>
+# if defined(__GNUC__) && defined(__i386__)
+static zend_always_inline void fast_memcpy(void *dest, const void *src, size_t size)
+{
+	size_t delta = (char*)dest - (char*)src;
+
+	__asm__ volatile (
+		".align 16\n\t"
+		".LL0%=:\n\t"
+		"prefetchnta 0x40(%1)\n\t"
+		"vmovaps (%1), %%ymm0\n\t"
+		"vmovaps 0x20(%1), %%ymm1\n\t"
+		"vmovaps 0x40(%1), %%ymm2\n\t"
+		"vmovaps 0x60(%1), %%ymm3\n\t"
+		"vmovaps %%ymm0, (%1,%2)\n\t"
+		"vmovaps %%ymm1, 0x20(%1,%2)\n\t"
+		"vmovaps %%ymm2, 0x40(%1,%2)\n\t"
+		"vmovaps %%ymm3, 0x60(%1,%2)\n\t"
+		"addl $0x80, %1\n\t"
+		"subl $0x80, %0\n\t"
+		"ja .LL0%="
+		: "+r"(size),
+		"+r"(src)
+		: "r"(delta)
+		: "cc", "memory", "%ymm0", "%ymm1", "%ymm2", "%ymm3");
+}
+# elif defined(__GNUC__) && defined(__x86_64__)
+static zend_always_inline void fast_memcpy(void *dest, const void *src, size_t size)
+{
+	size_t delta = (char*)dest - (char*)src;
+
+	__asm__ volatile (
+		".align 16\n\t"
+		".LL0%=:\n\t"
+		"prefetchnta 0x40(%1)\n\t"
+		"vmovaps (%1), %%ymm0\n\t"
+		"vmovaps 0x20(%1), %%ymm1\n\t"
+		"vmovaps 0x40(%1), %%ymm2\n\t"
+		"vmovaps 0x60(%1), %%ymm3\n\t"
+		"vmovaps %%ymm0, (%1,%2)\n\t"
+		"vmovaps %%ymm1, 0x20(%1,%2)\n\t"
+		"vmovaps %%ymm2, 0x40(%1,%2)\n\t"
+		"vmovaps %%ymm3, 0x60(%1,%2)\n\t"
+		"addq $0x80, %1\n\t"
+		"subq $0x80, %0\n\t"
+		"ja .LL0%="
+		: "+r"(size),
+		"+r"(src)
+		: "r"(delta)
+		: "cc", "memory", "%ymm0", "%ymm1", "%ymm2", "%ymm3");
+}
+# else
+static zend_always_inline void fast_memcpy(void *dest, const void *src, size_t size)
+{
+	__m256i *dqdest = (__m256i*)dest;
+	const __m256i *dqsrc  = (const __m256i*)src;
+	const __m256i *end  = (const __m256i*)((const char*)src + size);
+
+	do {
+#ifdef PHP_WIN32
+		_mm_prefetch((const char *)(dqsrc + 4), _MM_HINT_NTA);
+#else
+		_mm_prefetch(dqsrc + 4, _MM_HINT_NTA);
+#endif
+
+		__m256i ymm0 = _mm256_load_ps(dqsrc + 0);
+		__m256i ymm1 = _mm256_load_ps(dqsrc + 1);
+		__m256i ymm2 = _mm256_load_ps(dqsrc + 2);
+		__m256i ymm3 = _mm256_load_ps(dqsrc + 3);
+		dqsrc  += 4;
+		_mm256_store_ps(dqdest + 0, ymm0);
+		_mm256_store_ps(dqdest + 1, ymm1);
+		_mm256_store_ps(dqdest + 2, ymm2);
+		_mm256_store_ps(dqdest + 3, ymm3);
+		dqdest += 4;
+	} while (dqsrc != end);
+}
+# endif
+#elif defined(__SSE2__)
 # include <emmintrin.h>
 # if defined(__GNUC__) && defined(__i386__)
 static zend_always_inline void fast_memcpy(void *dest, const void *src, size_t size)
@@ -691,11 +769,11 @@ zend_op_array* zend_accel_load_script(zend_persistent_script *persistent_script,
 		ZCG(current_persistent_script) = persistent_script;
 		ZCG(arena_mem) = NULL;
 		if (EXPECTED(persistent_script->arena_size)) {
-#ifdef __SSE2__
+#if defined(__AVX__) || defined(__SSE2__)
 			/* Target address must be aligned to 64-byte boundary */
 			_mm_prefetch(persistent_script->arena_mem, _MM_HINT_NTA);
-			ZCG(arena_mem) = zend_arena_alloc(&CG(arena), persistent_script->arena_size + 64);
-			ZCG(arena_mem) = (void*)(((zend_uintptr_t)ZCG(arena_mem) + 63L) & ~63L);
+			ZCG(arena_mem) = zend_arena_alloc(&CG(arena), persistent_script->arena_size + FAST_COPY_PAD);
+			ZCG(arena_mem) = (void*)(((zend_uintptr_t)ZCG(arena_mem) + FAST_COPY_MASK) & ~FAST_COPY_MASK);
 			fast_memcpy(ZCG(arena_mem), persistent_script->arena_mem, persistent_script->arena_size);
 #else
 			ZCG(arena_mem) = zend_arena_alloc(&CG(arena), persistent_script->arena_size);
diff --git a/ext/opcache/zend_accelerator_util_funcs.h b/ext/opcache/zend_accelerator_util_funcs.h
index 5d7834a..8fb7acd 100644
--- a/ext/opcache/zend_accelerator_util_funcs.h
+++ b/ext/opcache/zend_accelerator_util_funcs.h
@@ -41,6 +41,14 @@ unsigned int zend_adler32(unsigned int checksum, signed char *buf, uint32_t len)

 unsigned int zend_accel_script_checksum(zend_persistent_script *persistent_script);

+#if defined(__AVX__)
+#define FAST_COPY_PAD  128L
+#define FAST_COPY_MASK 127L
+#elif defined(__SSE2__)
+#define FAST_COPY_PAD  64L
+#define FAST_COPY_MASK 63L
+#endif
+
 #endif /* ZEND_ACCELERATOR_UTIL_FUNCS_H */

 /*
diff --git a/ext/opcache/zend_file_cache.c b/ext/opcache/zend_file_cache.c
index de54949..a2c985c 100644
--- a/ext/opcache/zend_file_cache.c
+++ b/ext/opcache/zend_file_cache.c
@@ -809,10 +809,9 @@ int zend_file_cache_script_store(zend_persistent_script *script, int in_shm)
 		return FAILURE;
 	}

-#ifdef __SSE2__
-	/* Align to 64-byte boundary */
-	mem = emalloc(script->size + 64);
-	buf = (void*)(((zend_uintptr_t)mem + 63L) & ~63L);
+#if defined(__AVX__) || defined(__SSE2__)
+	mem = emalloc(script->size + FAST_COPY_PAD);
+	buf = (void*)(((zend_uintptr_t)mem + FAST_COPY_MASK) & ~FAST_COPY_MASK);
 #else
 	mem = buf = emalloc(script->size);
 #endif
@@ -1391,10 +1390,9 @@ zend_persistent_script *zend_file_cache_script_load(zend_file_handle *file_handl
 	}

 	checkpoint = zend_arena_checkpoint(CG(arena));
-#ifdef __SSE2__
-	/* Align to 64-byte boundary */
-	mem = zend_arena_alloc(&CG(arena), info.mem_size + info.str_size + 64);
-	mem = (void*)(((zend_uintptr_t)mem + 63L) & ~63L);
+#if defined(__AVX__) || defined(__SSE2__)
+	mem = zend_arena_alloc(&CG(arena), info.mem_size + info.str_size + FAST_COPY_PAD);
+	mem = (void*)(((zend_uintptr_t)mem + FAST_COPY_MASK) & ~FAST_COPY_MASK);
 #else
 	mem = zend_arena_alloc(&CG(arena), info.mem_size + info.str_size);
 #endif
@@ -1452,10 +1450,9 @@ zend_persistent_script *zend_file_cache_script_load(zend_file_handle *file_handl
 			goto use_process_mem;
 		}

-#ifdef __SSE2__
-		/* Align to 64-byte boundary */
-		buf = zend_shared_alloc(info.mem_size + 64);
-		buf = (void*)(((zend_uintptr_t)buf + 63L) & ~63L);
+#if defined(__AVX__) || defined(__SSE2__)
+		buf = zend_shared_alloc(info.mem_size + FAST_COPY_PAD);
+		buf = (void*)(((zend_uintptr_t)buf + FAST_COPY_MASK) & ~FAST_COPY_MASK);
 #else
 		buf = zend_shared_alloc(info.mem_size);
 #endif
diff --git a/ext/opcache/zend_persist.c b/ext/opcache/zend_persist.c
index 8c16538..7476a7c 100644
--- a/ext/opcache/zend_persist.c
+++ b/ext/opcache/zend_persist.c
@@ -27,6 +27,7 @@
 #include "zend_vm.h"
 #include "zend_constants.h"
 #include "zend_operators.h"
+#include "zend_accelerator_util_funcs.h"

 #define zend_accel_store(p, size) \
 	    (p = _zend_shared_memdup((void*)p, size, 1))
@@ -870,9 +871,9 @@ zend_persistent_script *zend_accel_script_persist(zend_persistent_script *script
 	}
 	zend_accel_store_interned_string(script->script.filename);

-#ifdef __SSE2__
+#if defined(__AVX__) || defined(__SSE2__)
 	/* Align to 64-byte boundary */
-	ZCG(mem) = (void*)(((zend_uintptr_t)ZCG(mem) + 63L) & ~63L);
+	ZCG(mem) = (void*)(((zend_uintptr_t)ZCG(mem) + FAST_COPY_MASK) & ~FAST_COPY_MASK);
 #else
 	ZEND_ASSERT(((zend_uintptr_t)ZCG(mem) & 0x7) == 0); /* should be 8 byte aligned */
 #endif
diff --git a/ext/opcache/zend_persist_calc.c b/ext/opcache/zend_persist_calc.c
index eb80232..c925d43 100644
--- a/ext/opcache/zend_persist_calc.c
+++ b/ext/opcache/zend_persist_calc.c
@@ -25,6 +25,7 @@
 #include "zend_extensions.h"
 #include "zend_shared_alloc.h"
 #include "zend_operators.h"
+#include "zend_accelerator_util_funcs.h"

 #define ADD_DUP_SIZE(m,s)  ZCG(current_persistent_script)->size += zend_shared_memdup_size((void*)m, s)
 #define ADD_SIZE(m)        ZCG(current_persistent_script)->size += ZEND_ALIGNED_SIZE(m)
@@ -415,9 +416,9 @@ uint32_t zend_accel_script_persist_calc(zend_persistent_script *new_persistent_s
 	}
 	ADD_STRING(new_persistent_script->script.filename);

-#ifdef __SSE2__
+#if defined(__AVX__) || defined(__SSE2__)
 	/* Align size to 64-byte boundary */
-	new_persistent_script->size = (new_persistent_script->size + 63) & ~63;
+	new_persistent_script->size = (new_persistent_script->size + FAST_COPY_MASK) & ~FAST_COPY_MASK;
 #endif

 	if (new_persistent_script->script.class_table.nNumUsed != new_persistent_script->script.class_table.nNumOfElements) {
@@ -430,9 +431,9 @@ uint32_t zend_accel_script_persist_calc(zend_persistent_script *new_persistent_s
 	zend_hash_persist_calc(&new_persistent_script->script.function_table, zend_persist_op_array_calc);
 	zend_persist_op_array_calc_ex(&new_persistent_script->script.main_op_array);

-#ifdef __SSE2__
+#if defined(__AVX__) || defined(__SSE2__)
 	/* Align size to 64-byte boundary */
-	new_persistent_script->arena_size = (new_persistent_script->arena_size + 63) & ~63;
+	new_persistent_script->arena_size = (new_persistent_script->arena_size + FAST_COPY_MASK) & ~FAST_COPY_MASK;
 #endif

 	new_persistent_script->size += new_persistent_script->arena_size;