frontend/ir_emitter: Add half-precision opcode for FPRecipStepFused
This commit is contained in:
parent
68d8cd2b13
commit
6da0411111
5 changed files with 48 additions and 31 deletions
|
@ -787,40 +787,42 @@ template<size_t fsize>
|
||||||
static void EmitFPRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
|
static void EmitFPRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
|
||||||
using FPT = mp::unsigned_integer_of_size<fsize>;
|
using FPT = mp::unsigned_integer_of_size<fsize>;
|
||||||
|
|
||||||
if (code.DoesCpuSupport(Xbyak::util::Cpu::tFMA)) {
|
if constexpr (fsize != 16) {
|
||||||
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
if (code.DoesCpuSupport(Xbyak::util::Cpu::tFMA)) {
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
|
||||||
Xbyak::Label end, fallback;
|
Xbyak::Label end, fallback;
|
||||||
|
|
||||||
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
|
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
|
||||||
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
|
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
|
||||||
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
|
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
|
||||||
|
|
||||||
code.movaps(result, code.MConst(xword, FP::FPValue<FPT, false, 0, 2>()));
|
code.movaps(result, code.MConst(xword, FP::FPValue<FPT, false, 0, 2>()));
|
||||||
FCODE(vfnmadd231s)(result, operand1, operand2);
|
FCODE(vfnmadd231s)(result, operand1, operand2);
|
||||||
FCODE(ucomis)(result, result);
|
FCODE(ucomis)(result, result);
|
||||||
code.jp(fallback, code.T_NEAR);
|
code.jp(fallback, code.T_NEAR);
|
||||||
code.L(end);
|
code.L(end);
|
||||||
|
|
||||||
code.SwitchToFarCode();
|
code.SwitchToFarCode();
|
||||||
code.L(fallback);
|
code.L(fallback);
|
||||||
|
|
||||||
code.sub(rsp, 8);
|
code.sub(rsp, 8);
|
||||||
ABI_PushCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
ABI_PushCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
||||||
code.movq(code.ABI_PARAM1, operand1);
|
code.movq(code.ABI_PARAM1, operand1);
|
||||||
code.movq(code.ABI_PARAM2, operand2);
|
code.movq(code.ABI_PARAM2, operand2);
|
||||||
code.mov(code.ABI_PARAM3.cvt32(), ctx.FPCR().Value());
|
code.mov(code.ABI_PARAM3.cvt32(), ctx.FPCR().Value());
|
||||||
code.lea(code.ABI_PARAM4, code.ptr[code.r15 + code.GetJitStateInfo().offsetof_fpsr_exc]);
|
code.lea(code.ABI_PARAM4, code.ptr[code.r15 + code.GetJitStateInfo().offsetof_fpsr_exc]);
|
||||||
code.CallFunction(&FP::FPRecipStepFused<FPT>);
|
code.CallFunction(&FP::FPRecipStepFused<FPT>);
|
||||||
code.movq(result, code.ABI_RETURN);
|
code.movq(result, code.ABI_RETURN);
|
||||||
ABI_PopCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
ABI_PopCallerSaveRegistersAndAdjustStackExcept(code, HostLocXmmIdx(result.getIdx()));
|
||||||
code.add(rsp, 8);
|
code.add(rsp, 8);
|
||||||
|
|
||||||
code.jmp(end, code.T_NEAR);
|
code.jmp(end, code.T_NEAR);
|
||||||
code.SwitchToNearCode();
|
code.SwitchToNearCode();
|
||||||
|
|
||||||
ctx.reg_alloc.DefineValue(inst, result);
|
ctx.reg_alloc.DefineValue(inst, result);
|
||||||
return;
|
return;
|
||||||
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
@ -830,6 +832,10 @@ static void EmitFPRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst*
|
||||||
code.CallFunction(&FP::FPRecipStepFused<FPT>);
|
code.CallFunction(&FP::FPRecipStepFused<FPT>);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitFPRecipStepFused16(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitFPRecipStepFused<16>(code, ctx, inst);
|
||||||
|
}
|
||||||
|
|
||||||
void EmitX64::EmitFPRecipStepFused32(EmitContext& ctx, IR::Inst* inst) {
|
void EmitX64::EmitFPRecipStepFused32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
EmitFPRecipStepFused<32>(code, ctx, inst);
|
EmitFPRecipStepFused<32>(code, ctx, inst);
|
||||||
}
|
}
|
||||||
|
|
|
@ -1943,11 +1943,20 @@ U16U32U64 IREmitter::FPRecipExponent(const U16U32U64& a) {
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
U32U64 IREmitter::FPRecipStepFused(const U32U64& a, const U32U64& b) {
|
U16U32U64 IREmitter::FPRecipStepFused(const U16U32U64& a, const U16U32U64& b) {
|
||||||
if (a.GetType() == Type::U32) {
|
ASSERT(a.GetType() == b.GetType());
|
||||||
|
|
||||||
|
switch (a.GetType()) {
|
||||||
|
case Type::U16:
|
||||||
|
return Inst<U16>(Opcode::FPRecipStepFused16, a, b);
|
||||||
|
case Type::U32:
|
||||||
return Inst<U32>(Opcode::FPRecipStepFused32, a, b);
|
return Inst<U32>(Opcode::FPRecipStepFused32, a, b);
|
||||||
|
case Type::U64:
|
||||||
|
return Inst<U64>(Opcode::FPRecipStepFused64, a, b);
|
||||||
|
default:
|
||||||
|
UNREACHABLE();
|
||||||
|
return U16U32U64{};
|
||||||
}
|
}
|
||||||
return Inst<U64>(Opcode::FPRecipStepFused64, a, b);
|
|
||||||
}
|
}
|
||||||
|
|
||||||
U16U32U64 IREmitter::FPRoundInt(const U16U32U64& a, FP::RoundingMode rounding, bool exact) {
|
U16U32U64 IREmitter::FPRoundInt(const U16U32U64& a, FP::RoundingMode rounding, bool exact) {
|
||||||
|
|
|
@ -307,7 +307,7 @@ public:
|
||||||
U16U32U64 FPNeg(const U16U32U64& a);
|
U16U32U64 FPNeg(const U16U32U64& a);
|
||||||
U32U64 FPRecipEstimate(const U32U64& a);
|
U32U64 FPRecipEstimate(const U32U64& a);
|
||||||
U16U32U64 FPRecipExponent(const U16U32U64& a);
|
U16U32U64 FPRecipExponent(const U16U32U64& a);
|
||||||
U32U64 FPRecipStepFused(const U32U64& a, const U32U64& b);
|
U16U32U64 FPRecipStepFused(const U16U32U64& a, const U16U32U64& b);
|
||||||
U16U32U64 FPRoundInt(const U16U32U64& a, FP::RoundingMode rounding, bool exact);
|
U16U32U64 FPRoundInt(const U16U32U64& a, FP::RoundingMode rounding, bool exact);
|
||||||
U32U64 FPRSqrtEstimate(const U32U64& a);
|
U32U64 FPRSqrtEstimate(const U32U64& a);
|
||||||
U32U64 FPRSqrtStepFused(const U32U64& a, const U32U64& b);
|
U32U64 FPRSqrtStepFused(const U32U64& a, const U32U64& b);
|
||||||
|
|
|
@ -277,6 +277,7 @@ bool Inst::ReadsFromAndWritesToFPSRCumulativeExceptionBits() const {
|
||||||
case Opcode::FPRecipExponent16:
|
case Opcode::FPRecipExponent16:
|
||||||
case Opcode::FPRecipExponent32:
|
case Opcode::FPRecipExponent32:
|
||||||
case Opcode::FPRecipExponent64:
|
case Opcode::FPRecipExponent64:
|
||||||
|
case Opcode::FPRecipStepFused16:
|
||||||
case Opcode::FPRecipStepFused32:
|
case Opcode::FPRecipStepFused32:
|
||||||
case Opcode::FPRecipStepFused64:
|
case Opcode::FPRecipStepFused64:
|
||||||
case Opcode::FPRoundInt16:
|
case Opcode::FPRoundInt16:
|
||||||
|
|
|
@ -496,6 +496,7 @@ OPCODE(FPRecipEstimate64, U64, U64
|
||||||
OPCODE(FPRecipExponent16, U16, U16 )
|
OPCODE(FPRecipExponent16, U16, U16 )
|
||||||
OPCODE(FPRecipExponent32, U32, U32 )
|
OPCODE(FPRecipExponent32, U32, U32 )
|
||||||
OPCODE(FPRecipExponent64, U64, U64 )
|
OPCODE(FPRecipExponent64, U64, U64 )
|
||||||
|
OPCODE(FPRecipStepFused16, U16, U16, U16 )
|
||||||
OPCODE(FPRecipStepFused32, U32, U32, U32 )
|
OPCODE(FPRecipStepFused32, U32, U32, U32 )
|
||||||
OPCODE(FPRecipStepFused64, U64, U64, U64 )
|
OPCODE(FPRecipStepFused64, U64, U64, U64 )
|
||||||
OPCODE(FPRoundInt16, U16, U16, U8, U1 )
|
OPCODE(FPRoundInt16, U16, U16, U8, U1 )
|
||||||
|
|
Loading…
Reference in a new issue