PCSX2/pcsx2
1
0
Fork 0
mirror of https://github.com/PCSX2/pcsx2.git synced 2024-10-07 07:23:28 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

3rdparty: Update xbyak to 7.05.1

Browse Source
This commit is contained in:
JordanTheToaster 2024-02-23 21:39:04 +00:00 committed by refractionpcsx2
parent f00f0cc846
commit 9c9d2b7c9c
3 changed files with 1940 additions and 1371 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

822
3rdparty/xbyak/xbyak/xbyak.h vendored
View File

File diff suppressed because it is too large Load Diff

2439
3rdparty/xbyak/xbyak/xbyak_mnemonic.h vendored
View File

File diff suppressed because it is too large Load Diff

50
3rdparty/xbyak/xbyak/xbyak_util.h vendored
View File

@ -27,7 +27,7 @@
#include "xbyak.h"
#endif // XBYAK_ONLY_CLASS_CPU
#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
#if defined(__i386__) || (defined(__x86_64__) && !defined(__arm64ec__)) || defined(_M_IX86) || (defined(_M_X64) && !defined(_M_ARM64EC))
#define XBYAK_INTEL_CPU_SPECIFIC
#endif
@ -144,6 +144,7 @@ private:
uint32_t dataCacheSize_[maxNumberCacheLevels];
uint32_t coresSharignDataCache_[maxNumberCacheLevels];
uint32_t dataCacheLevels_;
uint32_t avx10version_;
uint32_t get32bitAsBE(const char *x) const
{
@ -176,7 +177,7 @@ private:
}
uint32_t extractBit(uint32_t val, uint32_t base, uint32_t end)
{
return (val >> base) & ((1u << (end - base)) - 1);
return (val >> base) & ((1u << (end + 1 - base)) - 1);
}
void setNumCores()
{
@ -470,6 +471,14 @@ public:
XBYAK_DEFINE_TYPE(79, tSM3);
XBYAK_DEFINE_TYPE(80, tSM4);
XBYAK_DEFINE_TYPE(81, tAVX_VNNI_INT16);
XBYAK_DEFINE_TYPE(82, tAPX_F);
XBYAK_DEFINE_TYPE(83, tAVX10);
XBYAK_DEFINE_TYPE(84, tAESKLE);
XBYAK_DEFINE_TYPE(85, tWIDE_KL);
XBYAK_DEFINE_TYPE(86, tKEYLOCKER);
XBYAK_DEFINE_TYPE(87, tKEYLOCKER_WIDE);
XBYAK_DEFINE_TYPE(88, tSSE4a);
XBYAK_DEFINE_TYPE(89, tCLWB);
#undef XBYAK_SPLIT_ID
#undef XBYAK_DEFINE_TYPE
@ -481,6 +490,7 @@ public:
, dataCacheSize_()
, coresSharignDataCache_()
, dataCacheLevels_(0)
, avx10version_(0)
{
uint32_t data[4] = {};
const uint32_t& EAX = data[0];
@ -515,13 +525,14 @@ public:
if (maxExtendedNum >= 0x80000001) {
getCpuid(0x80000001, data);
if (EDX & (1U << 31)) type_ |= t3DN;
if (EDX & (1U << 30)) type_ |= tE3DN;
if (EDX & (1U << 27)) type_ |= tRDTSCP;
if (EDX & (1U << 22)) type_ |= tMMX2;
if (EDX & (1U << 15)) type_ |= tCMOV;
if (ECX & (1U << 5)) type_ |= tLZCNT;
if (ECX & (1U << 6)) type_ |= tSSE4a;
if (ECX & (1U << 8)) type_ |= tPREFETCHW;
if (EDX & (1U << 15)) type_ |= tCMOV;
if (EDX & (1U << 22)) type_ |= tMMX2;
if (EDX & (1U << 27)) type_ |= tRDTSCP;
if (EDX & (1U << 30)) type_ |= tE3DN;
if (EDX & (1U << 31)) type_ |= t3DN;
}
if (maxExtendedNum >= 0x80000008) {
@ -540,8 +551,8 @@ public:
if (ECX & (1U << 25)) type_ |= tAESNI;
if (ECX & (1U << 26)) type_ |= tXSAVE;
if (ECX & (1U << 27)) type_ |= tOSXSAVE;
if (ECX & (1U << 30)) type_ |= tRDRAND;
if (ECX & (1U << 29)) type_ |= tF16C;
if (ECX & (1U << 30)) type_ |= tRDRAND;
if (EDX & (1U << 15)) type_ |= tCMOV;
if (EDX & (1U << 23)) type_ |= tMMX;
@ -552,8 +563,8 @@ public:
// check XFEATURE_ENABLED_MASK[2:1] = '11b'
uint64_t bv = getXfeature();
if ((bv & 6) == 6) {
if (ECX & (1U << 28)) type_ |= tAVX;
if (ECX & (1U << 12)) type_ |= tFMA;
if (ECX & (1U << 28)) type_ |= tAVX;
// do *not* check AVX-512 state on macOS because it has on-demand AVX-512 support
#if !defined(__APPLE__)
if (((bv >> 5) & 7) == 7)
@ -587,21 +598,23 @@ public:
const uint32_t maxNumSubLeaves = EAX;
if (type_ & tAVX && (EBX & (1U << 5))) type_ |= tAVX2;
if (EBX & (1U << 3)) type_ |= tBMI1;
if (EBX & (1U << 4)) type_ |= tHLE;
if (EBX & (1U << 8)) type_ |= tBMI2;
if (EBX & (1U << 9)) type_ |= tENHANCED_REP;
if (EBX & (1U << 11)) type_ |= tRTM;
if (EBX & (1U << 14)) type_ |= tMPX;
if (EBX & (1U << 18)) type_ |= tRDSEED;
if (EBX & (1U << 19)) type_ |= tADX;
if (EBX & (1U << 20)) type_ |= tSMAP;
if (EBX & (1U << 23)) type_ |= tCLFLUSHOPT;
if (EBX & (1U << 4)) type_ |= tHLE;
if (EBX & (1U << 11)) type_ |= tRTM;
if (EBX & (1U << 14)) type_ |= tMPX;
if (EBX & (1U << 24)) type_ |= tCLWB;
if (EBX & (1U << 29)) type_ |= tSHA;
if (ECX & (1U << 0)) type_ |= tPREFETCHWT1;
if (ECX & (1U << 5)) type_ |= tWAITPKG;
if (ECX & (1U << 8)) type_ |= tGFNI;
if (ECX & (1U << 9)) type_ |= tVAES;
if (ECX & (1U << 10)) type_ |= tVPCLMULQDQ;
if (ECX & (1U << 23)) type_ |= tKEYLOCKER;
if (ECX & (1U << 25)) type_ |= tCLDEMOTE;
if (ECX & (1U << 27)) type_ |= tMOVDIRI;
if (ECX & (1U << 28)) type_ |= tMOVDIR64B;
@ -627,8 +640,20 @@ public:
if (EDX & (1U << 5)) type_ |= tAVX_NE_CONVERT;
if (EDX & (1U << 10)) type_ |= tAVX_VNNI_INT16;
if (EDX & (1U << 14)) type_ |= tPREFETCHITI;
if (EDX & (1U << 19)) type_ |= tAVX10;
if (EDX & (1U << 21)) type_ |= tAPX_F;
}
}
if (maxNum >= 0x19) {
getCpuidEx(0x19, 0, data);
if (EBX & (1U << 0)) type_ |= tAESKLE;
if (EBX & (1U << 2)) type_ |= tWIDE_KL;
if (type_ & (tKEYLOCKER|tAESKLE|tWIDE_KL)) type_ |= tKEYLOCKER_WIDE;
}
if (has(tAVX10) && maxNum >= 0x24) {
getCpuidEx(0x24, 0, data);
avx10version_ = EBX & mask(7);
}
setFamily();
setNumCores();
setCacheHierarchy();
@ -645,6 +670,7 @@ public:
{
return (type & type_) == type;
}
int getAVX10version() const { return avx10version_; }
};
#ifndef XBYAK_ONLY_CLASS_CPU