void Jit64::ps_sign(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITPairedOff);
FALLBACK_IF(inst.Rc);
int d = inst.FD;
int b = inst.FB;
fpr.Lock(d, b);
if (d != b)
{
fpr.BindToRegister(d, false);
MOVAPD(fpr.RX(d), fpr.R(b));
}
else
{
fpr.BindToRegister(d, true);
}
switch (inst.SUBOP10)
{
case 40: //neg
PXOR(fpr.RX(d), M((void*)&psSignBits));
break;
case 136: //nabs
POR(fpr.RX(d), M((void*)&psSignBits));
break;
case 264: //abs
PAND(fpr.RX(d), M((void*)&psAbsMask));
break;
}
fpr.UnlockAll();
}
void Jit64::ps_sel(UGeckoInstruction inst)
{
// we can't use (V)BLENDVPD here because it just looks at the sign bit
// but we need -0 = +0
INSTRUCTION_START
JITDISABLE(bJITPairedOff);
FALLBACK_IF(inst.Rc);
int d = inst.FD;
int a = inst.FA;
int b = inst.FB;
int c = inst.FC;
fpr.Lock(a, b, c, d);
MOVAPD(XMM0, fpr.R(a));
PXOR(XMM1, R(XMM1));
// XMM0 = XMM0 < 0 ? all 1s : all 0s
CMPPD(XMM0, R(XMM1), LT);
MOVAPD(XMM1, R(XMM0));
PAND(XMM0, fpr.R(b));
PANDN(XMM1, fpr.R(c));
POR(XMM0, R(XMM1));
fpr.BindToRegister(d, false);
MOVAPD(fpr.RX(d), R(XMM0));
fpr.UnlockAll();
}
void Jit64::fsign(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITFloatingPointOff);
FALLBACK_IF(inst.Rc);
int d = inst.FD;
int b = inst.FB;
fpr.Lock(b, d);
fpr.BindToRegister(d, true, true);
MOVSD(XMM0, fpr.R(b));
switch (inst.SUBOP10) {
case 40: // fnegx
PXOR(XMM0, M((void*)&psSignBits2));
break;
case 264: // fabsx
PAND(XMM0, M((void*)&psAbsMask2));
break;
case 136: // fnabs
POR(XMM0, M((void*)&psSignBits2));
break;
default:
PanicAlert("fsign bleh");
break;
}
MOVSD(fpr.R(d), XMM0);
fpr.UnlockAll();
}
bool SamplerJitCache::Jit_Decode4444() {
MOVD_xmm(fpScratchReg1, R(resultReg));
PUNPCKLBW(fpScratchReg1, R(fpScratchReg1));
if (RipAccessible(color4444mask)) {
PAND(fpScratchReg1, M(color4444mask)); // rip accessible
} else {
Crash();
}
MOVSS(fpScratchReg2, R(fpScratchReg1));
MOVSS(fpScratchReg3, R(fpScratchReg1));
PSRLW(fpScratchReg2, 4);
PSLLW(fpScratchReg3, 4);
POR(fpScratchReg1, R(fpScratchReg2));
POR(fpScratchReg1, R(fpScratchReg3));
MOVD_xmm(R(resultReg), fpScratchReg1);
return true;
}
void Jit::Comp_FPU2op(MIPSOpcode op) {
CONDITIONAL_DISABLE;
int fs = _FS;
int fd = _FD;
switch (op & 0x3f)
{
case 5: //F(fd) = fabsf(F(fs)); break; //abs
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
MOVSS(fpr.RX(fd), fpr.R(fs));
PAND(fpr.RX(fd), M(ssNoSignMask));
break;
case 6: //F(fd) = F(fs); break; //mov
if (fd != fs) {
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
MOVSS(fpr.RX(fd), fpr.R(fs));
}
break;
case 7: //F(fd) = -F(fs); break; //neg
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
MOVSS(fpr.RX(fd), fpr.R(fs));
PXOR(fpr.RX(fd), M(ssSignBits2));
break;
case 4: //F(fd) = sqrtf(F(fs)); break; //sqrt
fpr.SpillLock(fd, fs); // this probably works, just badly tested
fpr.MapReg(fd, fd == fs, true);
SQRTSS(fpr.RX(fd), fpr.R(fs));
break;
case 13: //FsI(fd) = F(fs)>=0 ? (int)floorf(F(fs)) : (int)ceilf(F(fs)); break;//trunc.w.s
{
fpr.SpillLock(fs, fd);
fpr.StoreFromRegister(fd);
CVTTSS2SI(EAX, fpr.R(fs));
// Did we get an indefinite integer value?
CMP(32, R(EAX), Imm32(0x80000000));
FixupBranch skip = J_CC(CC_NE);
MOVSS(XMM0, fpr.R(fs));
XORPS(XMM1, R(XMM1));
CMPSS(XMM0, R(XMM1), CMP_LT);
// At this point, -inf = 0xffffffff, inf/nan = 0x00000000.
// We want -inf to be 0x80000000 inf/nan to be 0x7fffffff, so we flip those bits.
MOVD_xmm(R(EAX), XMM0);
XOR(32, R(EAX), Imm32(0x7fffffff));
SetJumpTarget(skip);
MOV(32, fpr.R(fd), R(EAX));
}
break;
case 32: //F(fd) = (float)FsI(fs); break; //cvt.s.w
// Store to memory so we can read it as an integer value.
fpr.StoreFromRegister(fs);
CVTSI2SS(XMM0, fpr.R(fs));
MOVSS(fpr.R(fd), XMM0);
break;
case 12: //FsI(fd) = (int)floorf(F(fs)+0.5f); break; //round.w.s
case 14: //FsI(fd) = (int)ceilf (F(fs)); break; //ceil.w.s
case 15: //FsI(fd) = (int)floorf(F(fs)); break; //floor.w.s
case 36: //FsI(fd) = (int) F(fs); break; //cvt.w.s
default:
DISABLE;
return;
}
fpr.ReleaseSpillLocks();
}
void Jit::Comp_FPU2op(u32 op)
{
CONDITIONAL_DISABLE;
int fs = _FS;
int fd = _FD;
switch (op & 0x3f)
{
case 5: //F(fd) = fabsf(F(fs)); break; //abs
fpr.Lock(fd, fs);
fpr.BindToRegister(fd, fd == fs, true);
MOVSS(fpr.RX(fd), fpr.R(fs));
PAND(fpr.RX(fd), M((void *)ssNoSignMask));
fpr.UnlockAll();
break;
case 6: //F(fd) = F(fs); break; //mov
if (fd != fs) {
fpr.Lock(fd, fs);
fpr.BindToRegister(fd, fd == fs, true);
MOVSS(fpr.RX(fd), fpr.R(fs));
fpr.UnlockAll();
}
break;
case 7: //F(fd) = -F(fs); break; //neg
fpr.Lock(fd, fs);
fpr.BindToRegister(fd, fd == fs, true);
MOVSS(fpr.RX(fd), fpr.R(fs));
PXOR(fpr.RX(fd), M((void *)ssSignBits2));
fpr.UnlockAll();
break;
case 12: //FsI(fd) = (int)floorf(F(fs)+0.5f); break; //round.w.s
case 4: //F(fd) = sqrtf(F(fs)); break; //sqrt
/* fpr.Lock(fd, fs); // this probably works, just badly tested
fpr.BindToRegister(fd, fd == fs, true);
SQRTSS(fpr.RX(fd), fpr.R(fs));
fpr.UnlockAll();
break;*/
Comp_Generic(op);
return;
case 13: //FsI(fd) = F(fs)>=0 ? (int)floorf(F(fs)) : (int)ceilf(F(fs)); break;//trunc.w.s
fpr.Lock(fs, fd);
fpr.StoreFromRegister(fd);
CVTTSS2SI(EAX, fpr.R(fs));
MOV(32, fpr.R(fd), R(EAX));
fpr.UnlockAll();
break;
case 14: //FsI(fd) = (int)ceilf (F(fs)); break; //ceil.w.s
case 15: //FsI(fd) = (int)floorf(F(fs)); break; //floor.w.s
case 32: //F(fd) = (float)FsI(fs); break; //cvt.s.w
case 36: //FsI(fd) = (int) F(fs); break; //cvt.w.s
default:
Comp_Generic(op);
return;
}
}
void Jit::Comp_FPU2op(MIPSOpcode op) {
CONDITIONAL_DISABLE;
int fs = _FS;
int fd = _FD;
auto execRounding = [&](void (XEmitter::*conv)(X64Reg, OpArg), int setMXCSR) {
fpr.SpillLock(fs);
// Small optimization: 0 is our default mode anyway.
if (setMXCSR == 0 && !js.hasSetRounding) {
setMXCSR = -1;
}
if (setMXCSR != -1) {
STMXCSR(M(&mxcsrTemp));
MOV(32, R(TEMPREG), M(&mxcsrTemp));
AND(32, R(TEMPREG), Imm32(~(3 << 13)));
OR(32, R(TEMPREG), Imm32(setMXCSR << 13));
MOV(32, M(&mips_->temp), R(TEMPREG));
LDMXCSR(M(&mips_->temp));
}
(this->*conv)(TEMPREG, fpr.R(fs));
// Did we get an indefinite integer value?
CMP(32, R(TEMPREG), Imm32(0x80000000));
FixupBranch skip = J_CC(CC_NE);
MOVSS(XMM0, fpr.R(fs));
XORPS(XMM1, R(XMM1));
CMPSS(XMM0, R(XMM1), CMP_LT);
// At this point, -inf = 0xffffffff, inf/nan = 0x00000000.
// We want -inf to be 0x80000000 inf/nan to be 0x7fffffff, so we flip those bits.
MOVD_xmm(R(TEMPREG), XMM0);
XOR(32, R(TEMPREG), Imm32(0x7fffffff));
SetJumpTarget(skip);
fpr.DiscardR(fd);
MOV(32, fpr.R(fd), R(TEMPREG));
if (setMXCSR != -1) {
LDMXCSR(M(&mxcsrTemp));
}
};
switch (op & 0x3f) {
case 5: //F(fd) = fabsf(F(fs)); break; //abs
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
if (fd != fs) {
MOVSS(fpr.RX(fd), fpr.R(fs));
}
PAND(fpr.RX(fd), M(ssNoSignMask));
break;
case 6: //F(fd) = F(fs); break; //mov
if (fd != fs) {
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
MOVSS(fpr.RX(fd), fpr.R(fs));
}
break;
case 7: //F(fd) = -F(fs); break; //neg
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fd == fs, true);
if (fd != fs) {
MOVSS(fpr.RX(fd), fpr.R(fs));
}
XORPS(fpr.RX(fd), M(ssSignBits2));
break;
case 4: //F(fd) = sqrtf(F(fs)); break; //sqrt
fpr.SpillLock(fd, fs); // this probably works, just badly tested
fpr.MapReg(fd, fd == fs, true);
SQRTSS(fpr.RX(fd), fpr.R(fs));
break;
case 13: //FsI(fd) = F(fs)>=0 ? (int)floorf(F(fs)) : (int)ceilf(F(fs)); break;//trunc.w.s
execRounding(&XEmitter::CVTTSS2SI, -1);
break;
case 32: //F(fd) = (float)FsI(fs); break; //cvt.s.w
fpr.SpillLock(fd, fs);
fpr.MapReg(fd, fs == fd, true);
if (fpr.R(fs).IsSimpleReg()) {
CVTDQ2PS(fpr.RX(fd), fpr.R(fs));
} else {
// If fs was fd, we'd be in the case above since we mapped fd.
MOVSS(fpr.RX(fd), fpr.R(fs));
CVTDQ2PS(fpr.RX(fd), fpr.R(fd));
}
break;
case 36: //FsI(fd) = (int) F(fs); break; //cvt.w.s
// Uses the current rounding mode.
execRounding(&XEmitter::CVTSS2SI, -1);
break;
case 12: //FsI(fd) = (int)floorf(F(fs)+0.5f); break; //round.w.s
execRounding(&XEmitter::CVTSS2SI, 0);
break;
case 14: //FsI(fd) = (int)ceilf (F(fs)); break; //ceil.w.s
execRounding(&XEmitter::CVTSS2SI, 2);
break;
case 15: //FsI(fd) = (int)floorf(F(fs)); break; //floor.w.s
execRounding(&XEmitter::CVTSS2SI, 1);
break;
default:
DISABLE;
return;
}
fpr.ReleaseSpillLocks();
}