softjit: Fix PixelFuncID arg on non-Windows x64.

Oops, this is of course not put on the stack, it's in R8.
This commit is contained in:
Unknown W. Brackets 2021-11-28 08:54:36 -08:00
parent 2d8fdd8cf4
commit 9653c33d9c
2 changed files with 25 additions and 7 deletions

View File

@ -89,7 +89,6 @@ SingleFunc PixelJitCache::CompileSingle(const PixelFuncID &id) {
#else #else
// Must save: RBX, RSP, RBP, R12-R15 // Must save: RBX, RSP, RBP, R12-R15
regCache_.Add(R8, RegCache::GEN_INVALID);
regCache_.Add(R9, RegCache::GEN_INVALID); regCache_.Add(R9, RegCache::GEN_INVALID);
regCache_.Add(XMM4, RegCache::VEC_INVALID); regCache_.Add(XMM4, RegCache::VEC_INVALID);
@ -98,9 +97,9 @@ SingleFunc PixelJitCache::CompileSingle(const PixelFuncID &id) {
regCache_.Add(RDX, RegCache::GEN_ARG_Z); regCache_.Add(RDX, RegCache::GEN_ARG_Z);
regCache_.Add(RCX, RegCache::GEN_ARG_FOG); regCache_.Add(RCX, RegCache::GEN_ARG_FOG);
regCache_.Add(XMM0, RegCache::VEC_ARG_COLOR); regCache_.Add(XMM0, RegCache::VEC_ARG_COLOR);
regCache_.Add(R8, RegCache::GEN_ARG_ID);
// Here we just have the return and padding to align RPB. stackIDOffset_ = -1;
stackIDOffset_ = 16;
#endif #endif
// Initially, disallow spill for args (they get unlocked when no longer needed.) // Initially, disallow spill for args (they get unlocked when no longer needed.)
@ -109,6 +108,8 @@ SingleFunc PixelJitCache::CompileSingle(const PixelFuncID &id) {
regCache_.ForceRetain(RegCache::GEN_ARG_Z); regCache_.ForceRetain(RegCache::GEN_ARG_Z);
regCache_.ForceRetain(RegCache::GEN_ARG_FOG); regCache_.ForceRetain(RegCache::GEN_ARG_FOG);
regCache_.ForceRetain(RegCache::VEC_ARG_COLOR); regCache_.ForceRetain(RegCache::VEC_ARG_COLOR);
if (regCache_.Has(RegCache::GEN_ARG_ID))
regCache_.ForceRetain(RegCache::GEN_ARG_ID);
BeginWrite(); BeginWrite();
const u8 *start = AlignCode16(); const u8 *start = AlignCode16();
@ -145,6 +146,8 @@ SingleFunc PixelJitCache::CompileSingle(const PixelFuncID &id) {
} }
discards_.clear(); discards_.clear();
if (regCache_.Has(RegCache::GEN_ARG_ID))
regCache_.ForceRelease(RegCache::GEN_ARG_ID);
regCache_.Reset(success); regCache_.Reset(success);
if (!success) { if (!success) {
@ -1412,8 +1415,15 @@ bool PixelJitCache::Jit_Dither(const PixelFuncID &id) {
LEA(32, valueReg, MComplex(argXReg, valueReg, 8, offsetof(PixelFuncID, cached.ditherMatrix))); LEA(32, valueReg, MComplex(argXReg, valueReg, 8, offsetof(PixelFuncID, cached.ditherMatrix)));
// Okay, now abuse argXReg to read the PixelFuncID pointer on the stack. // Okay, now abuse argXReg to read the PixelFuncID pointer on the stack.
MOV(PTRBITS, R(argXReg), MDisp(RSP, stackIDOffset_)); if (regCache_.Has(RegCache::GEN_ARG_ID)) {
MOVSX(32, 16, valueReg, MRegSum(argXReg, valueReg)); X64Reg idReg = regCache_.Find(RegCache::GEN_ARG_ID);
MOVSX(32, 16, valueReg, MRegSum(idReg, valueReg));
regCache_.Unlock(idReg, RegCache::GEN_ARG_ID);
} else {
_assert_(stackIDOffset_ != -1);
MOV(PTRBITS, R(argXReg), MDisp(RSP, stackIDOffset_));
MOVSX(32, 16, valueReg, MRegSum(argXReg, valueReg));
}
#endif #endif
if (argXReg != INVALID_REG) { if (argXReg != INVALID_REG) {
regCache_.Unlock(argXReg, RegCache::GEN_ARG_X); regCache_.Unlock(argXReg, RegCache::GEN_ARG_X);
@ -1597,8 +1607,15 @@ bool PixelJitCache::Jit_WriteColor(const PixelFuncID &id) {
#else #else
maskReg = regCache_.Alloc(RegCache::GEN_TEMP3); maskReg = regCache_.Alloc(RegCache::GEN_TEMP3);
// Load the pre-converted and combined write mask. // Load the pre-converted and combined write mask.
MOV(PTRBITS, R(maskReg), MDisp(RSP, stackIDOffset_)); if (regCache_.Has(RegCache::GEN_ARG_ID)) {
MOV(32, R(maskReg), MDisp(maskReg, offsetof(PixelFuncID, cached.colorWriteMask))); X64Reg idReg = regCache_.Find(RegCache::GEN_ARG_ID);
MOV(32, R(maskReg), MDisp(idReg, offsetof(PixelFuncID, cached.colorWriteMask)));
regCache_.Unlock(idReg, RegCache::GEN_ARG_ID);
} else {
_assert_(stackIDOffset_ != -1);
MOV(PTRBITS, R(maskReg), MDisp(RSP, stackIDOffset_));
MOV(32, R(maskReg), MDisp(maskReg, offsetof(PixelFuncID, cached.colorWriteMask)));
}
#endif #endif
} }

View File

@ -97,6 +97,7 @@ struct RegCache {
GEN_ARG_Y = 0x0181, GEN_ARG_Y = 0x0181,
GEN_ARG_Z = 0x0182, GEN_ARG_Z = 0x0182,
GEN_ARG_FOG = 0x0183, GEN_ARG_FOG = 0x0183,
GEN_ARG_ID = 0x0184,
VEC_ARG_COLOR = 0x0080, VEC_ARG_COLOR = 0x0080,
VEC_ARG_MASK = 0x0081, VEC_ARG_MASK = 0x0081,