Windows arm64 zend and standard extension support

* Port zend_cpuid for windows arm64
* Fix zend_atomic windows arm64 build
* Fix windows arm64 multiply
* Enable arm64 neon for windows in standard extension
* Enable arm64 neon for windows in zend_hash.c
* Workaround for msvc arm64 optimization bug

Closes GH-9115.

Author: dixyes

Zend/zend_atomic.h

@@ -57,6 +57,13 @@ BEGIN_EXTERN_C()
 
 #if ZEND_WIN32
 
+#ifndef InterlockedExchange8
+#define InterlockedExchange8 _InterlockedExchange8
+#endif
+#ifndef InterlockedOr8
+#define InterlockedOr8 _InterlockedOr8
+#endif
+
 #define ZEND_ATOMIC_BOOL_INIT(obj, desired) ((obj)->value = (desired))
 
 static zend_always_inline bool zend_atomic_bool_exchange_ex(zend_atomic_bool *obj, bool desired) {

Zend/zend_cpuinfo.c

@@ -28,15 +28,15 @@ typedef struct _zend_cpu_info {
 
 static zend_cpu_info cpuinfo = {0};
 
-#if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
-# if defined(HAVE_CPUID_H) && defined(HAVE_CPUID_COUNT)
-# include <cpuid.h>
+#if (defined(__GNUC__) || defined(__clang__)) && (defined(__i386__) || defined(__x86_64__))
+# if defined(HAVE_CPUID_H) && defined(HAVE_CPUID_COUNT) /* use cpuid.h functions */
+#  include <cpuid.h>
 static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo) {
 	__cpuid_count(func, subfunc, cpuinfo->eax, cpuinfo->ebx, cpuinfo->ecx, cpuinfo->edx);
 }
-# else
+# else /* use inline asm */
 static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo) {
-#if defined(__i386__) && (defined(__pic__) || defined(__PIC__))
+#  if defined(__i386__) && (defined(__pic__) || defined(__PIC__))
 	/* PIC on i386 uses %ebx, so preserve it. */
 	__asm__ __volatile__ (
 		"pushl  %%ebx\n"
@@ -46,16 +46,16 @@ static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo
 		: "=a"(cpuinfo->eax), "=r"(cpuinfo->ebx), "=c"(cpuinfo->ecx), "=d"(cpuinfo->edx)
 		: "a"(func), "c"(subfunc)
 	);
-#else
+#  else
 	__asm__ __volatile__ (
 		"cpuid"
 		: "=a"(cpuinfo->eax), "=b"(cpuinfo->ebx), "=c"(cpuinfo->ecx), "=d"(cpuinfo->edx)
 		: "a"(func), "c"(subfunc)
 	);
-#endif
+#  endif
 }
 # endif
-#elif defined(ZEND_WIN32) && !defined(__clang__)
+#elif defined(_MSC_VER) && !defined(__clang__) && (defined(_M_X64) || defined(_M_IX86)) /* use MSVC __cpuidex intrin */
 # include <intrin.h>
 static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo) {
 	int regs[4];
@@ -67,7 +67,7 @@ static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo
 	cpuinfo->ecx = regs[2];
 	cpuinfo->edx = regs[3];
 }
-#else
+#else /* fall back to zero */
 static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo) {
 	cpuinfo->eax = 0;
 }

Zend/zend_hash.c

@@ -22,7 +22,7 @@
 #include "zend_globals.h"
 #include "zend_variables.h"
 
-#if defined(__aarch64__)
+#if defined(__aarch64__) || defined(_M_ARM64)
 # include <arm_neon.h>
 #endif
 
@@ -183,7 +183,7 @@ static zend_always_inline void zend_hash_real_init_mixed_ex(HashTable *ht)
 			_mm_storeu_si128((__m128i*)&HT_HASH_EX(data,  8), xmm0);
 			_mm_storeu_si128((__m128i*)&HT_HASH_EX(data, 12), xmm0);
 		} while (0);
-#elif defined(__aarch64__)
+#elif defined(__aarch64__) || defined(_M_ARM64)
 		do {
 			int32x4_t t = vdupq_n_s32(-1);
 			vst1q_s32((int32_t*)&HT_HASH_EX(data,  0), t);

Zend/zend_multiply.h

@@ -94,6 +94,19 @@
 		(dval) = (double)(a) * (double)(b); \
 	} \
 } while (0)
+# elif defined(_M_ARM64)
+#  pragma intrinsic(__mulh)
+#  define ZEND_SIGNED_MULTIPLY_LONG(a, b, lval, dval, usedval) do {       \
+	__int64 __high = __mulh((a), (b)); \
+	__int64 __low  = (a) * (b); \
+	if ((__low >> 63I64) == __high) { \
+		(usedval) = 0; \
+		(lval) = __low; \
+	} else { \
+		(usedval) = 1; \
+		(dval) = (double)(a) * (double)(b); \
+	} \
+} while (0)
 # else
 #  define ZEND_SIGNED_MULTIPLY_LONG(a, b, lval, dval, usedval) do {	\
 	zend_long   __lres  = (a) * (b);										\

ext/json/json_encoder.c

@@ -71,7 +71,14 @@ static inline void php_json_pretty_print_indent(smart_str *buf, int options, php
 
 /* }}} */
 
-static inline bool php_json_is_valid_double(double d) /* {{{ */
+static
+#if defined(_MSC_VER) && defined(_M_ARM64)
+// MSVC bug: https://developercommunity.visualstudio.com/t/corrupt-optimization-on-arm64-with-Ox-/10102551
+zend_never_inline
+#else
+inline
+#endif
+bool php_json_is_valid_double(double d) /* {{{ */
 {
 	return !zend_isinf(d) && !zend_isnan(d);
 }

ext/opcache/ZendAccelerator.c

@@ -2190,7 +2190,7 @@ zend_op_array *persistent_compile_file(zend_file_handle *file_handle, int type)
 		ZCSG(hits)++; /* TBFixed: may lose one hit */
 		persistent_script->dynamic_members.hits++; /* see above */
 #else
-#ifdef _M_X64
+#if ZEND_ENABLE_ZVAL_LONG64
 		InterlockedIncrement64(&ZCSG(hits));
 #else
 		InterlockedIncrement(&ZCSG(hits));

ext/standard/base64.c

@@ -51,7 +51,7 @@ static const short base64_reverse_table[256] = {
 };
 /* }}} */
 
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(_M_ARM64)
 #include <arm_neon.h>
 
 static zend_always_inline uint8x16_t encode_toascii(const uint8x16_t input, const uint8x16x2_t shift_LUT)
@@ -118,11 +118,11 @@ static zend_always_inline unsigned char *neon_base64_encode(const unsigned char
 	*left = inl;
 	return out;
 }
-#endif /* __aarch64__ */
+#endif /* defined(__aarch64__) || defined(_M_ARM64) */
 
 static zend_always_inline unsigned char *php_base64_encode_impl(const unsigned char *in, size_t inl, unsigned char *out) /* {{{ */
 {
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(_M_ARM64)
 	if (inl >= 16 * 3) {
 		size_t left = 0;
 		out = neon_base64_encode(in, inl, out, &left);
@@ -161,7 +161,7 @@ static zend_always_inline unsigned char *php_base64_encode_impl(const unsigned c
 }
 /* }}} */
 
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(_M_ARM64)
 static zend_always_inline uint8x16_t decode_fromascii(const uint8x16_t input, uint8x16_t *error, const uint8x16x2_t shiftLUT, const uint8x16x2_t maskLUT, const uint8x16x2_t bitposLUT) {
 	const uint8x16_t higher_nibble = vshrq_n_u8(input, 4);
 	const uint8x16_t lower_nibble = vandq_u8(input, vdupq_n_u8(0x0f));
@@ -241,14 +241,14 @@ static zend_always_inline size_t neon_base64_decode(const unsigned char *in, siz
 	*left = inl;
 	return out - out_orig;
 }
-#endif /* __aarch64__ */
+#endif /* defined(__aarch64__) || defined(_M_ARM64) */
 
 static zend_always_inline int php_base64_decode_impl(const unsigned char *in, size_t inl, unsigned char *out, size_t *outl, bool strict) /* {{{ */
 {
 	int ch;
 	size_t i = 0, padding = 0, j = *outl;
 
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(_M_ARM64)
 	if (inl >= 16 * 4) {
 		size_t left = 0;
 		j += neon_base64_decode(in, inl, out, &left);

ext/standard/string.c

@@ -3338,7 +3338,7 @@ PHP_FUNCTION(strtr)
 /* {{{ Reverse a string */
 #ifdef ZEND_INTRIN_SSSE3_NATIVE
 #include <tmmintrin.h>
-#elif defined(__aarch64__)
+#elif defined(__aarch64__) || defined(_M_ARM64)
 #include <arm_neon.h>
 #endif
 PHP_FUNCTION(strrev)
@@ -3385,6 +3385,19 @@ PHP_FUNCTION(strrev)
 			e -= 16;
 		} while (e - s > 15);
 	}
+#elif defined(_M_ARM64)
+	if (e - s > 15) {
+		do {
+			const __n128 str = vld1q_u8((uint8_t *)(e - 15));
+			/* Synthesize rev128 with a rev64 + ext. */
+			/* strange force cast limit on windows: you cannot convert anything */
+			const __n128 rev = vrev64q_u8(str);
+			const __n128 ext = vextq_u64(rev, rev, 1);
+			vst1q_u8((uint8_t *)p, ext);
+			p += 16;
+			e -= 16;
+		} while (e - s > 15);
+	}
 #endif
 	while (e >= s) {
 		*p++ = *e--;
@@ -3864,7 +3877,7 @@ zend_string *php_addslashes_sse42(zend_string *str)
 /* }}} */
 #endif
 
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(_M_ARM64)
 typedef union {
 	uint8_t mem[16];
 	uint64_t dw[2];
@@ -3899,7 +3912,7 @@ static zend_always_inline char *aarch64_add_slashes(quad_word res, const char *s
 	}
 	return target;
 }
-#endif /* __aarch64__ */
+#endif /* defined(__aarch64__) || defined(_M_ARM64) */
 
 #ifndef ZEND_INTRIN_SSE4_2_NATIVE
 # ifdef ZEND_INTRIN_SSE4_2_RESOLVER
@@ -3921,7 +3934,7 @@ PHPAPI zend_string *php_addslashes(zend_string *str)
 	source = ZSTR_VAL(str);
 	end = source + ZSTR_LEN(str);
 
-# ifdef __aarch64__
+# if defined(__aarch64__) || defined(_M_ARM64)
 	quad_word res = {0};
 	if (ZSTR_LEN(str) > 15) {
 		do {
@@ -3932,7 +3945,7 @@ PHPAPI zend_string *php_addslashes(zend_string *str)
 		} while ((end - source) > 15);
 	}
 	/* Finish the last 15 bytes or less with the scalar loop. */
-# endif /* __aarch64__ */
+# endif /* defined(__aarch64__) || defined(_M_ARM64) */
 
 	while (source < end) {
 		switch (*source) {
@@ -3955,7 +3968,7 @@ PHPAPI zend_string *php_addslashes(zend_string *str)
 	memcpy(ZSTR_VAL(new_str), ZSTR_VAL(str), offset);
 	target = ZSTR_VAL(new_str) + offset;
 
-# ifdef __aarch64__
+# if defined(__aarch64__) || defined(_M_ARM64)
 	if (res.dw[0] | res.dw[1]) {
 		target = aarch64_add_slashes(res, source, target);
 		source += 16;
@@ -3971,7 +3984,7 @@ PHPAPI zend_string *php_addslashes(zend_string *str)
 		}
 	}
 	/* Finish the last 15 bytes or less with the scalar loop. */
-# endif /* __aarch64__ */
+# endif /* defined(__aarch64__) || defined(_M_ARM64) */
 
 	while (source < end) {
 		switch (*source) {
@@ -4010,7 +4023,7 @@ PHPAPI zend_string *php_addslashes(zend_string *str)
  * be careful, this edits the string in-place */
 static zend_always_inline char *php_stripslashes_impl(const char *str, char *out, size_t len)
 {
-#ifdef __aarch64__
+#if defined(__aarch64__) || defined(_M_ARM64)
 	while (len > 15) {
 		uint8x16_t x = vld1q_u8((uint8_t *)str);
 		quad_word q;
@@ -4040,7 +4053,7 @@ static zend_always_inline char *php_stripslashes_impl(const char *str, char *out
 		}
 	}
 	/* Finish the last 15 bytes or less with the scalar loop. */
-#endif /* __aarch64__ */
+#endif /* defined(__aarch64__) || defined(_M_ARM64) */
 	while (len > 0) {
 		if (*str == '\\') {
 			str++;				/* skip the slash */