nintendo-emulators/dolphin
1
0
Fork 0
mirror of https://github.com/dolphin-emu/dolphin synced 2025-10-05 16:03:02 +02:00
Code Issues Packages Projects Releases Wiki Activity

Jit64: dcbz, use VMOVAPS with YMM registers on AVX CPUs

Browse Source 
LLVM does this, so let's do it as well.
This commit is contained in:
Martino Fontana
2025-09-02 15:12:37 +02:00
parent 5e69da7eba
commit d1ba849876
3 changed files with 38 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
Source/Core/Common/x64Emitter.cpp
View File

@@ -1858,19 +1858,23 @@ void XEmitter::WriteVEXOp4(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, co
Write8((u8)regOp3 << 4);
}
void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg,
int W, int extrabytes)
void CheckAVXSupport()
{
if (!cpu_info.bAVX)
PanicAlertFmt("Trying to use AVX on a system that doesn't support it. Bad programmer.");
}
void XEmitter::WriteAVXOp(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg,
int W, int extrabytes)
{
CheckAVXSupport();
WriteVEXOp(opPrefix, op, regOp1, regOp2, arg, W, extrabytes);
}
void XEmitter::WriteAVXOp4(u8 opPrefix, u16 op, X64Reg regOp1, X64Reg regOp2, const OpArg& arg,
X64Reg regOp3, int W)
{
if (!cpu_info.bAVX)
PanicAlertFmt("Trying to use AVX on a system that doesn't support it. Bad programmer.");
CheckAVXSupport();
WriteVEXOp4(opPrefix, op, regOp1, regOp2, arg, regOp3, W);
}
@@ -3030,6 +3034,19 @@ void XEmitter::VPXOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg)
WriteAVXOp(0x66, 0xEF, regOp1, regOp2, arg);
}
void XEmitter::VMOVAPS(const OpArg& arg, X64Reg regOp)
{
WriteAVXOp(0x00, 0x29, X64Reg::INVALID_REG, regOp, arg);
}
void XEmitter::VZEROUPPER()
{
CheckAVXSupport();
Write8(0xC5);
Write8(0xF8);
Write8(0x77);
}
void XEmitter::VFMADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg)
{
WriteFMA3Op(0x98, regOp1, regOp2, arg);

4
Source/Core/Common/x64Emitter.h
View File

@@ -876,6 +876,10 @@ public:
void VPOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VPXOR(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VMOVAPS(const OpArg& arg, X64Reg regOp);
void VZEROUPPER();
// FMA3
void VFMADD132PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);
void VFMADD213PS(X64Reg regOp1, X64Reg regOp2, const OpArg& arg);

16
Source/Core/Core/PowerPC/Jit64/Jit_LoadStore.cpp
View File

@@ -8,6 +8,7 @@
#include "Common/Assert.h"
#include "Common/BitSet.h"
#include "Common/CPUDetect.h"
#include "Common/CommonTypes.h"
#include "Common/MsgHandler.h"
#include "Common/x64ABI.h"
@@ -469,9 +470,18 @@ void Jit64::dcbz(UGeckoInstruction inst)
FixupBranch slow = J_CC(CC_Z, Jump::Near);
// Fast path: compute full address, then zero out 32 bytes of memory.
XORPS(XMM0, R(XMM0));
MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 0), XMM0);
MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 16), XMM0);
if (cpu_info.bAVX)
{
VXORPS(XMM0, XMM0, R(XMM0));
VMOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 0), YMM0);
VZEROUPPER();
}
else
{
XORPS(XMM0, R(XMM0));
MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 16), XMM0);
MOVAPS(MComplex(RMEM, RSCRATCH, SCALE_1, 0), XMM0);
}
// Slow path: call the general-case code.
SwitchToFarCode();