A32: Implement ASIMD VRECPS
This commit is contained in:
parent
9eef4f7471
commit
8f506c80c3
7 changed files with 23 additions and 13 deletions
|
@ -1176,6 +1176,7 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
|
|||
if constexpr (fsize != 16) {
|
||||
if (code.HasFMA() && code.HasAVX()) {
|
||||
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||
const bool fpcr_controlled = args[2].GetImmediateU1();
|
||||
|
||||
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
|
||||
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
|
||||
|
@ -1184,8 +1185,10 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
|
|||
|
||||
Xbyak::Label end, fallback;
|
||||
|
||||
MaybeStandardFPSCRValue(code, ctx, fpcr_controlled, [&]{
|
||||
code.movaps(result, GetVectorOf<fsize, false, 0, 2>(code));
|
||||
FCODE(vfnmadd231p)(result, operand1, operand2);
|
||||
});
|
||||
|
||||
FCODE(vcmpunordp)(tmp, result, result);
|
||||
code.vptest(tmp, tmp);
|
||||
|
@ -1196,7 +1199,7 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
|
|||
code.L(fallback);
|
||||
code.sub(rsp, 8);
|
||||
ABI_PushCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
||||
EmitThreeOpFallbackWithoutRegAlloc(code, ctx, result, operand1, operand2, fallback_fn);
|
||||
EmitThreeOpFallbackWithoutRegAlloc(code, ctx, result, operand1, operand2, fallback_fn, fpcr_controlled);
|
||||
ABI_PopCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
||||
code.add(rsp, 8);
|
||||
code.jmp(end, code.T_NEAR);
|
||||
|
@ -1207,7 +1210,7 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
|
|||
}
|
||||
}
|
||||
|
||||
EmitThreeOpFallback(code, ctx, inst, fallback_fn);
|
||||
EmitThreeOpFallback<FpcrControlledArgument::Present>(code, ctx, inst, fallback_fn);
|
||||
}
|
||||
|
||||
void EmitX64::EmitFPVectorRecipStepFused16(EmitContext& ctx, IR::Inst* inst) {
|
||||
|
|
|
@ -46,7 +46,7 @@ INST(asimd_VMUL_float, "VMUL (floating-point)", "111100110D0znnnndddd110
|
|||
//INST(asimd_VACGE, "VACGE", "111100110-CC--------1110---1----") // ASIMD
|
||||
INST(asimd_VMAX_float, "VMAX (floating-point)", "111100100D0znnnndddd1111NQM0mmmm") // ASIMD
|
||||
INST(asimd_VMIN_float, "VMIN (floating-point)", "111100100D1znnnndddd1111NQM0mmmm") // ASIMD
|
||||
//INST(asimd_VRECPS, "VRECPS", "111100100-0C--------1111---1----") // ASIMD
|
||||
INST(asimd_VRECPS, "VRECPS", "111100100D0znnnndddd1111NQM1mmmm") // ASIMD
|
||||
//INST(asimd_VRSQRTS, "VRSQRTS", "111100100-1C--------1111---1----") // ASIMD
|
||||
|
||||
// Two registers and a scalar
|
||||
|
|
|
@ -413,4 +413,10 @@ bool ArmTranslatorVisitor::asimd_VMIN_float(bool D, bool sz, size_t Vn, size_t V
|
|||
});
|
||||
}
|
||||
|
||||
bool ArmTranslatorVisitor::asimd_VRECPS(bool D, bool sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm) {
|
||||
return FloatingPointInstruction(*this, D, sz, Vn, Vd, N, Q, M, Vm, [this](const auto&, const auto& reg_n, const auto& reg_m) {
|
||||
return ir.FPVectorRecipStepFused(32, reg_n, reg_m, false);
|
||||
});
|
||||
}
|
||||
|
||||
} // namespace Dynarmic::A32
|
||||
|
|
|
@ -471,6 +471,7 @@ struct ArmTranslatorVisitor final {
|
|||
bool asimd_VMUL_float(bool D, bool sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
|
||||
bool asimd_VMAX_float(bool D, bool sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
|
||||
bool asimd_VMIN_float(bool D, bool sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
|
||||
bool asimd_VRECPS(bool D, bool sz, size_t Vn, size_t Vd, bool N, bool Q, bool M, size_t Vm);
|
||||
|
||||
// Two registers and a shift amount
|
||||
bool asimd_SHR(bool U, bool D, size_t imm6, size_t Vd, bool L, bool Q, bool M, size_t Vm);
|
||||
|
|
|
@ -2452,14 +2452,14 @@ U128 IREmitter::FPVectorRecipEstimate(size_t esize, const U128& a) {
|
|||
UNREACHABLE();
|
||||
}
|
||||
|
||||
U128 IREmitter::FPVectorRecipStepFused(size_t esize, const U128& a, const U128& b) {
|
||||
U128 IREmitter::FPVectorRecipStepFused(size_t esize, const U128& a, const U128& b, bool fpcr_controlled) {
|
||||
switch (esize) {
|
||||
case 16:
|
||||
return Inst<U128>(Opcode::FPVectorRecipStepFused16, a, b);
|
||||
return Inst<U128>(Opcode::FPVectorRecipStepFused16, a, b, Imm1(fpcr_controlled));
|
||||
case 32:
|
||||
return Inst<U128>(Opcode::FPVectorRecipStepFused32, a, b);
|
||||
return Inst<U128>(Opcode::FPVectorRecipStepFused32, a, b, Imm1(fpcr_controlled));
|
||||
case 64:
|
||||
return Inst<U128>(Opcode::FPVectorRecipStepFused64, a, b);
|
||||
return Inst<U128>(Opcode::FPVectorRecipStepFused64, a, b, Imm1(fpcr_controlled));
|
||||
}
|
||||
UNREACHABLE();
|
||||
}
|
||||
|
|
|
@ -361,7 +361,7 @@ public:
|
|||
U128 FPVectorPairedAdd(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
||||
U128 FPVectorPairedAddLower(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
||||
U128 FPVectorRecipEstimate(size_t esize, const U128& a);
|
||||
U128 FPVectorRecipStepFused(size_t esize, const U128& a, const U128& b);
|
||||
U128 FPVectorRecipStepFused(size_t esize, const U128& a, const U128& b, bool fpcr_controlled = true);
|
||||
U128 FPVectorRoundInt(size_t esize, const U128& operand, FP::RoundingMode rounding, bool exact);
|
||||
U128 FPVectorRSqrtEstimate(size_t esize, const U128& a);
|
||||
U128 FPVectorRSqrtStepFused(size_t esize, const U128& a, const U128& b);
|
||||
|
|
|
@ -616,9 +616,9 @@ OPCODE(FPVectorPairedAddLower64, U128, U128
|
|||
OPCODE(FPVectorRecipEstimate16, U128, U128 )
|
||||
OPCODE(FPVectorRecipEstimate32, U128, U128 )
|
||||
OPCODE(FPVectorRecipEstimate64, U128, U128 )
|
||||
OPCODE(FPVectorRecipStepFused16, U128, U128, U128 )
|
||||
OPCODE(FPVectorRecipStepFused32, U128, U128, U128 )
|
||||
OPCODE(FPVectorRecipStepFused64, U128, U128, U128 )
|
||||
OPCODE(FPVectorRecipStepFused16, U128, U128, U128, U1 )
|
||||
OPCODE(FPVectorRecipStepFused32, U128, U128, U128, U1 )
|
||||
OPCODE(FPVectorRecipStepFused64, U128, U128, U128, U1 )
|
||||
OPCODE(FPVectorRoundInt16, U128, U128, U8, U1 )
|
||||
OPCODE(FPVectorRoundInt32, U128, U128, U8, U1 )
|
||||
OPCODE(FPVectorRoundInt64, U128, U128, U8, U1 )
|
||||
|
|
Loading…
Reference in a new issue