Skip to content

Windows arm64 zend and standard extension supports #9115

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 8 commits into from
Closed

Windows arm64 zend and standard extension supports #9115

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
a2a53cb
Port zend_cpuid for windows arm64
dixyes Dec 1, 2021
ba53063
Fix zend_atomic windows arm64 build
dixyes Jul 20, 2022
496266d
Fix windows arm64 multiply
dixyes Jul 23, 2022
100311c
Enable arm64 neon for windows in standard extension
dixyes Jul 23, 2022
e82edbd
Enable arm64 neon for windows in zend_hash.c
dixyes Jul 23, 2022
a58e5ce
Use ZEND_ENABLE_ZVAL_LONG64
dixyes Jul 23, 2022
18b5b17
Workaround for msvc arm64 optimization bug
dixyes Jul 23, 2022
00c340b
Add clang checks in __zend_cpuid for windows
dixyes Jul 24, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions Zend/zend_atomic.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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) {
Expand Down
18 changes: 9 additions & 9 deletions Zend/zend_cpuinfo.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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"
Expand All @@ -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];
Expand All @@ -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 /* fail back to zero */
static void __zend_cpuid(uint32_t func, uint32_t subfunc, zend_cpu_info *cpuinfo) {
cpuinfo->eax = 0;
}
Expand Down
4 changes: 2 additions & 2 deletions Zend/zend_hash.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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

Expand Down Expand Up @@ -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);
Expand Down
13 changes: 13 additions & 0 deletions Zend/zend_multiply.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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); \
Expand Down
9 changes: 8 additions & 1 deletion ext/json/json_encoder.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nice to have a comment referring to the bug report here! :)

Copy link
Contributor Author

@dixyes dixyes Aug 3, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@cmb69 this bug also affects x86_64 at VS 2022 17.2, should we remove && defined(_M_ARM64) ?

see ruby issue ruby workaround vs community issue

generated asm code from vs 17.2 on my environment for x86_64:

$LN14:
  000000000000027C: F2 0F 10 07        movsd       xmm0,mmword ptr [rdi]
  0000000000000280: E8 00 00 00 00     call        _dclass
  0000000000000285: F2 0F 10 0F        movsd       xmm1,mmword ptr [rdi]
  0000000000000289: 48 8D 0D 00 00 00  lea         rcx,[$SG72158]
                    00
  0000000000000290: 45 33 C0           xor         r8d,r8d
  0000000000000293: 66 48 0F 7E CA     movq        rdx,xmm1
  0000000000000298: 66 83 F8 01        cmp         ax,1
  000000000000029C: 41 0F 95 C0        setne       r8b
  00000000000002A0: E8 00 00 00 00     call        printf ----- debug printf
  00000000000002A5: F2 0F 10 07        movsd       xmm0,mmword ptr [rdi]
  00000000000002A9: E8 00 00 00 00     call        _dclass
  00000000000002AE: 66 83 F8 01        cmp         ax,1 <- only checked infinite
  00000000000002B2: 74 16              je          00000000000002CA
  00000000000002B4: F2 0F 10 0F        movsd       xmm1,mmword ptr [rdi]
  00000000000002B8: 44 8B C5           mov         r8d,ebp
  00000000000002BB: 48 8B CB           mov         rcx,rbx
  00000000000002BE: E8 00 00 00 00     call        php_json_encode_double
  00000000000002C3: 33 C0              xor         eax,eax
  00000000000002C5: E9 E7 FD FF FF     jmp         00000000000000B1
  00000000000002CA: C7 46 08 07 00 00  mov         dword ptr [rsi+8],7
                    00
  00000000000002D1: BF 01 00 00 00     mov         edi,1
  00000000000002D6: 48 8B 03           mov         rax,qword ptr [rbx]
  00000000000002D9: 48 85 C0           test        rax,rax
  00000000000002DC: 74 0D              je          00000000000002EB
  00000000000002DE: 48 8B 78 10        mov         rdi,qword ptr [rax+10h]
  00000000000002E2: 48 FF C7           inc         rdi
  00000000000002E5: 48 3B 7B 08        cmp         rdi,qword ptr [rbx+8]
  00000000000002E9: 72 0B              jb          00000000000002F6
  00000000000002EB: 48 8B D7           mov         rdx,rdi
  00000000000002EE: 48 8B CB           mov         rcx,rbx
  00000000000002F1: E8 00 00 00 00     call        smart_str_erealloc@@16
  00000000000002F6: 48 8B 03           mov         rax,qword ptr [rbx]
  00000000000002F9: C6 44 38 17 30     mov         byte ptr [rax+rdi+17h],30h
  00000000000002FE: 48 8B 03           mov         rax,qword ptr [rbx]
  0000000000000301: 48 89 78 10        mov         qword ptr [rax+10h],rdi
  0000000000000305: 33 C0              xor         eax,eax
  0000000000000307: E9 A5 FD FF FF     jmp         00000000000000B1

correct code generated by vs 17.3 preview:

$LN14:
  000000000000027C: 0F 29 74 24 60     movaps      xmmword ptr [rsp+60h],xmm6
  0000000000000281: F2 0F 10 36        movsd       xmm6,mmword ptr [rsi]
  0000000000000285: 0F 28 C6           movaps      xmm0,xmm6
  0000000000000288: E8 00 00 00 00     call        _dclass
  000000000000028D: BF 01 00 00 00     mov         edi,1
  0000000000000292: 66 3B C7           cmp         ax,di
  0000000000000295: 74 11              je          00000000000002A8
  0000000000000297: 66 0F 2E F6        ucomisd     xmm6,xmm6
  000000000000029B: 0F 9A C0           setp        al
  000000000000029E: 40 3A C7           cmp         al,dil
  00000000000002A1: 74 05              je          00000000000002A8
  00000000000002A3: 44 8B C7           mov         r8d,edi
  00000000000002A6: EB 03              jmp         00000000000002AB
  00000000000002A8: 45 33 C0           xor         r8d,r8d
  00000000000002AB: 0F 28 CE           movaps      xmm1,xmm6
  00000000000002AE: 48 8D 0D 00 00 00  lea         rcx,[$SG72170]
                    00
  00000000000002B5: 66 48 0F 7E F2     movq        rdx,xmm6
  00000000000002BA: E8 00 00 00 00     call        printf ----- debug printf
  00000000000002BF: F2 0F 10 36        movsd       xmm6,mmword ptr [rsi]
  00000000000002C3: 0F 28 C6           movaps      xmm0,xmm6
  00000000000002C6: E8 00 00 00 00     call        _dclass
  00000000000002CB: 66 3B C7           cmp         ax,di <-- check INF first
  00000000000002CE: 74 26              je          00000000000002F6
  00000000000002D0: 66 0F 2E F6        ucomisd     xmm6,xmm6
  00000000000002D4: 0F 9A C0           setp        al <-- checked NaN also
  00000000000002D7: 40 3A C7           cmp         al,dil
  00000000000002DA: 74 1A              je          00000000000002F6
  00000000000002DC: 45 8B C6           mov         r8d,r14d
  00000000000002DF: 0F 28 CE           movaps      xmm1,xmm6
  00000000000002E2: 48 8B CB           mov         rcx,rbx
  00000000000002E5: E8 00 00 00 00     call        php_json_encode_double
  00000000000002EA: 0F 28 74 24 60     movaps      xmm6,xmmword ptr [rsp+60h]
  00000000000002EF: 33 C0              xor         eax,eax
  00000000000002F1: E9 BC FD FF FF     jmp         00000000000000B2
  00000000000002F6: C7 45 08 07 00 00  mov         dword ptr [rbp+8],7
                    00
  00000000000002FD: 48 8B 03           mov         rax,qword ptr [rbx]
  0000000000000300: 48 85 C0           test        rax,rax
  0000000000000303: 74 0D              je          0000000000000312
  0000000000000305: 48 8B 78 10        mov         rdi,qword ptr [rax+10h]
  0000000000000309: 48 FF C7           inc         rdi
  000000000000030C: 48 3B 7B 08        cmp         rdi,qword ptr [rbx+8]
  0000000000000310: 72 0B              jb          000000000000031D
  0000000000000312: 48 8B D7           mov         rdx,rdi
  0000000000000315: 48 8B CB           mov         rcx,rbx
  0000000000000318: E8 00 00 00 00     call        smart_str_erealloc@@16
  000000000000031D: 48 8B 03           mov         rax,qword ptr [rbx]
  0000000000000320: 0F 28 74 24 60     movaps      xmm6,xmmword ptr [rsp+60h]
  0000000000000325: C6 44 38 17 30     mov         byte ptr [rax+rdi+17h],30h
  000000000000032A: 48 8B 03           mov         rax,qword ptr [rbx]
  000000000000032D: 48 89 78 10        mov         qword ptr [rax+10h],rdi
  0000000000000331: 33 C0              xor         eax,eax
  0000000000000333: E9 7A FD FF FF     jmp         00000000000000B2

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Well, given that the official builds still use VS 2019, and that issue is fixed with 17.3, I think we can leave that as is.

zend_never_inline
#else
inline
#endif
bool php_json_is_valid_double(double d) /* {{{ */
{
return !zend_isinf(d) && !zend_isnan(d);
}
Expand Down
2 changes: 1 addition & 1 deletion ext/opcache/ZendAccelerator.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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));
Expand Down
12 changes: 6 additions & 6 deletions ext/standard/base64.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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)
Expand Down Expand Up @@ -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);
Expand Down Expand Up @@ -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));
Expand Down Expand Up @@ -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);
Expand Down
31 changes: 22 additions & 9 deletions ext/standard/string.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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)
Expand Down Expand Up @@ -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--;
Expand Down Expand Up @@ -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];
Expand Down Expand Up @@ -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
Expand All @@ -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 {
Expand All @@ -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) {
Expand All @@ -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;
Expand All @@ -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) {
Expand Down Expand Up @@ -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;
Expand Down Expand Up @@ -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 */
Expand Down