void JitArm::ps_rsqrte(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITPairedOff);
FALLBACK_IF(inst.Rc);
u32 b = inst.FB, d = inst.FD;
ARMReg vB0 = fpr.R0(b);
ARMReg vB1 = fpr.R1(b);
ARMReg vD0 = fpr.R0(d, false);
ARMReg vD1 = fpr.R1(d, false);
ARMReg fpscrReg = gpr.GetReg();
ARMReg V0 = D1;
ARMReg rA = gpr.GetReg();
MOVI2R(fpscrReg, (u32)&PPC_NAN);
VLDR(V0, fpscrReg, 0);
LDR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
VCMP(vB0);
VMRS(_PC);
FixupBranch Less0 = B_CC(CC_LT);
VMOV(vD0, V0);
SetFPException(fpscrReg, FPSCR_VXSQRT);
FixupBranch SkipOrr0 = B();
SetJumpTarget(Less0);
SetCC(CC_EQ);
ORR(rA, rA, 1);
SetCC();
SetJumpTarget(SkipOrr0);
VCMP(vB1);
VMRS(_PC);
FixupBranch Less1 = B_CC(CC_LT);
VMOV(vD1, V0);
SetFPException(fpscrReg, FPSCR_VXSQRT);
FixupBranch SkipOrr1 = B();
SetJumpTarget(Less1);
SetCC(CC_EQ);
ORR(rA, rA, 2);
SetCC();
SetJumpTarget(SkipOrr1);
CMP(rA, 0);
FixupBranch noException = B_CC(CC_EQ);
SetFPException(fpscrReg, FPSCR_ZX);
SetJumpTarget(noException);
VCVT(S0, vB0, 0);
VCVT(S1, vB1, 0);
NEONXEmitter nemit(this);
nemit.VRSQRTE(F_32, D0, D0);
VCVT(vD0, S0, 0);
VCVT(vD1, S1, 0);
STR(fpscrReg, R9, PPCSTATE_OFF(fpscr));
gpr.Unlock(fpscrReg, rA);
}
void JitArm::ps_sel(UGeckoInstruction inst)
{
INSTRUCTION_START
JITDISABLE(bJITPairedOff);
FALLBACK_IF(inst.Rc);
u32 a = inst.FA, b = inst.FB, c = inst.FC, d = inst.FD;
ARMReg vA0 = fpr.R0(a);
ARMReg vA1 = fpr.R1(a);
ARMReg vB0 = fpr.R0(b);
ARMReg vB1 = fpr.R1(b);
ARMReg vC0 = fpr.R0(c);
ARMReg vC1 = fpr.R1(c);
ARMReg vD0 = fpr.R0(d, false);
ARMReg vD1 = fpr.R1(d, false);
VCMP(vA0);
VMRS(_PC);
FixupBranch GT0 = B_CC(CC_GE);
VMOV(vD0, vB0);
FixupBranch EQ0 = B();
SetJumpTarget(GT0);
VMOV(vD0, vC0);
SetJumpTarget(EQ0);
VCMP(vA1);
VMRS(_PC);
FixupBranch GT1 = B_CC(CC_GE);
VMOV(vD1, vB1);
FixupBranch EQ1 = B();
SetJumpTarget(GT1);
VMOV(vD1, vC1);
SetJumpTarget(EQ1);
}
void Jit::Comp_FPUComp(u32 op) {
CONDITIONAL_DISABLE;
int opc = op & 0xF;
if (opc >= 8) opc -= 8; // alias
if (opc == 0)//f, sf (signalling false)
{
MOVI2R(R0, 0);
STR(R0, CTXREG, offsetof(MIPSState, fpcond));
return;
}
int fs = _FS;
int ft = _FT;
fpr.MapInIn(fs, ft);
VCMP(fpr.R(fs), fpr.R(ft));
VMRS_APSR(); // Move FP flags from FPSCR to APSR (regular flags).
switch(opc)
{
case 1: // un, ngle (unordered)
SetCC(CC_VS);
MOVI2R(R0, 1);
SetCC(CC_VC);
break;
case 2: // eq, seq (equal, ordered)
SetCC(CC_EQ);
MOVI2R(R0, 1);
SetCC(CC_NEQ);
break;
case 3: // ueq, ngl (equal, unordered)
SetCC(CC_EQ);
MOVI2R(R0, 1);
SetCC(CC_NEQ);
MOVI2R(R0, 0);
SetCC(CC_VS);
MOVI2R(R0, 1);
SetCC(CC_AL);
STR(R0, CTXREG, offsetof(MIPSState, fpcond));
return;
case 4: // olt, lt (less than, ordered)
SetCC(CC_LO);
MOVI2R(R0, 1);
SetCC(CC_HS);
break;
case 5: // ult, nge (less than, unordered)
SetCC(CC_LT);
MOVI2R(R0, 1);
SetCC(CC_GE);
break;
case 6: // ole, le (less equal, ordered)
SetCC(CC_LS);
MOVI2R(R0, 1);
SetCC(CC_HI);
break;
case 7: // ule, ngt (less equal, unordered)
SetCC(CC_LE);
MOVI2R(R0, 1);
SetCC(CC_GT);
break;
default:
Comp_Generic(op);
return;
}
MOVI2R(R0, 0);
SetCC(CC_AL);
STR(R0, CTXREG, offsetof(MIPSState, fpcond));
}
void Jit::ApplyPrefixD(const u8 *vregs, VectorSize sz) {
_assert_(js.prefixDFlag & ArmJitState::PREFIX_KNOWN);
if (!js.prefixD) return;
int n = GetNumVectorElements(sz);
for (int i = 0; i < n; i++) {
if (js.VfpuWriteMask(i))
continue;
// TODO: These clampers are wrong - put this into google
// and look at the plot: abs(x) - abs(x-0.5) + 0.5
// It's too steep.
// Also, they mishandle NaN and Inf.
int sat = (js.prefixD >> (i * 2)) & 3;
if (sat == 1) {
// clamped = fabs(x) - fabs(x-0.5f) + 0.5f; // [ 0, 1]
fpr.MapRegV(vregs[i], MAP_DIRTY);
MOVI2F(S0, 0.0f, R0);
MOVI2F(S1, 1.0f, R0);
VCMP(fpr.V(vregs[i]), S0);
VMRS_APSR(); // Move FP flags from FPSCR to APSR (regular flags).
SetCC(CC_LE);
VMOV(fpr.V(vregs[i]), S0);
SetCC(CC_AL);
VCMP(fpr.V(vregs[i]), S1);
VMRS_APSR(); // Move FP flags from FPSCR to APSR (regular flags).
SetCC(CC_GT);
VMOV(fpr.V(vregs[i]), S1);
SetCC(CC_AL);
/*
VABS(S1, fpr.V(vregs[i])); // S1 = fabs(x)
VSUB(fpr.V(vregs[i]), fpr.V(vregs[i]), S0); // S2 = fabs(x-0.5f) {VABD}
VABS(fpr.V(vregs[i]), fpr.V(vregs[i]));
VSUB(fpr.V(vregs[i]), S1, fpr.V(vregs[i])); // v[i] = S1 - S2 + 0.5f
VADD(fpr.V(vregs[i]), fpr.V(vregs[i]), S0);*/
} else if (sat == 3) {
fpr.MapRegV(vregs[i], MAP_DIRTY);
MOVI2F(S0, -1.0f, R0);
MOVI2F(S1, 1.0f, R0);
VCMP(fpr.V(vregs[i]), S0);
VMRS_APSR(); // Move FP flags from FPSCR to APSR (regular flags).
SetCC(CC_LT);
VMOV(fpr.V(vregs[i]), S0);
SetCC(CC_AL);
VCMP(fpr.V(vregs[i]), S1);
VMRS_APSR(); // Move FP flags from FPSCR to APSR (regular flags).
SetCC(CC_GT);
VMOV(fpr.V(vregs[i]), S1);
SetCC(CC_AL);
// clamped = fabs(x) - fabs(x-1.0f); // [-1, 1]
/*
fpr.MapRegV(vregs[i], MAP_DIRTY);
MOVI2F(S0, 1.0f, R0);
VABS(S1, fpr.V(vregs[i])); // S1 = fabs(x)
VSUB(fpr.V(vregs[i]), fpr.V(vregs[i]), S0); // S2 = fabs(x-1.0f) {VABD}
VABS(fpr.V(vregs[i]), fpr.V(vregs[i]));
VSUB(fpr.V(vregs[i]), S1, fpr.V(vregs[i])); // v[i] = S1 - S2
*/
}
}
}
