suyu/dynarmic
26
1
Fork 1
Code Issues Pull requests Projects Releases Packages Wiki Activity

frontend/ir_emitter: Add half-precision variant of FPRoundInt

Browse source
This commit is contained in:
Lioncash 2019-04-13 17:40:56 -04:00 committed by MerryMage
parent 61cec94a19
commit ad0c698f89
5 changed files with 25 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
src/backend/x64/emit_x64_floating_point.cpp
View file

@ -843,7 +843,7 @@ static void EmitFPRound(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, siz
const bool exact = inst->GetArg(2).GetU1();
const auto round_imm = ConvertRoundingModeToX64Immediate(rounding_mode);
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41) && round_imm && !exact) {
if (fsize != 16 && code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41) && round_imm && !exact) {
if (fsize == 64) {
FPTwoOp<64>(code, ctx, inst, [&](Xbyak::Xmm result) {
code.roundsd(result, result, *round_imm);
@ -857,7 +857,9 @@ static void EmitFPRound(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, siz
return;
}
using fsize_list = mp::list<mp::vlift<size_t(32)>, mp::vlift<size_t(64)>>;
using fsize_list = mp::list<mp::vlift<size_t(16)>,
mp::vlift<size_t(32)>,
mp::vlift<size_t(64)>>;
using rounding_list = mp::list<
std::integral_constant<FP::RoundingMode, FP::RoundingMode::ToNearest_TieEven>,
std::integral_constant<FP::RoundingMode, FP::RoundingMode::TowardsPlusInfinity>,
@ -897,6 +899,10 @@ static void EmitFPRound(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, siz
code.CallFunction(lut.at(std::make_tuple(fsize, rounding_mode, exact)));
}
void EmitX64::EmitFPRoundInt16(EmitContext& ctx, IR::Inst* inst) {
EmitFPRound(code, ctx, inst, 16);
}
void EmitX64::EmitFPRoundInt32(EmitContext& ctx, IR::Inst* inst) {
EmitFPRound(code, ctx, inst, 32);
}

18
src/frontend/ir/ir_emitter.cpp
View file

@ -1950,11 +1950,21 @@ U32U64 IREmitter::FPRecipStepFused(const U32U64& a, const U32U64& b) {
return Inst<U64>(Opcode::FPRecipStepFused64, a, b);
}
U32U64 IREmitter::FPRoundInt(const U32U64& a, FP::RoundingMode rounding, bool exact) {
if (a.GetType() == Type::U32) {
return Inst<U32>(Opcode::FPRoundInt32, a, static_cast<u8>(rounding), Imm1(exact));
U16U32U64 IREmitter::FPRoundInt(const U16U32U64& a, FP::RoundingMode rounding, bool exact) {
const u8 rounding_value = static_cast<u8>(rounding);
const IR::U1 exact_imm = Imm1(exact);
switch (a.GetType()) {
case Type::U16:
return Inst<U16>(Opcode::FPRoundInt16, a, rounding_value, exact_imm);
case Type::U32:
return Inst<U32>(Opcode::FPRoundInt32, a, rounding_value, exact_imm);
case Type::U64:
return Inst<U64>(Opcode::FPRoundInt64, a, rounding_value, exact_imm);
default:
UNREACHABLE();
return U16U32U64{};
}
return Inst<U64>(Opcode::FPRoundInt64, a, static_cast<u8>(rounding), Imm1(exact));
}
U32U64 IREmitter::FPRSqrtEstimate(const U32U64& a) {

2
src/frontend/ir/ir_emitter.h
View file

@ -308,7 +308,7 @@ public:
U32U64 FPRecipEstimate(const U32U64& a);
U16U32U64 FPRecipExponent(const U16U32U64& a);
U32U64 FPRecipStepFused(const U32U64& a, const U32U64& b);
U32U64 FPRoundInt(const U32U64& a, FP::RoundingMode rounding, bool exact);
U16U32U64 FPRoundInt(const U16U32U64& a, FP::RoundingMode rounding, bool exact);
U32U64 FPRSqrtEstimate(const U32U64& a);
U32U64 FPRSqrtStepFused(const U32U64& a, const U32U64& b);
U32U64 FPSqrt(const U32U64& a);

1
src/frontend/ir/microinstruction.cpp
View file

@ -279,6 +279,7 @@ bool Inst::ReadsFromAndWritesToFPSRCumulativeExceptionBits() const {
case Opcode::FPRecipExponent64:
case Opcode::FPRecipStepFused32:
case Opcode::FPRecipStepFused64:
case Opcode::FPRoundInt16:
case Opcode::FPRoundInt32:
case Opcode::FPRoundInt64:
case Opcode::FPRSqrtEstimate32:

1
src/frontend/ir/opcodes.inc
View file

@ -498,6 +498,7 @@ OPCODE(FPRecipExponent32, U32, U32
OPCODE(FPRecipExponent64, U64, U64 )
OPCODE(FPRecipStepFused32, U32, U32, U32 )
OPCODE(FPRecipStepFused64, U64, U64, U64 )
OPCODE(FPRoundInt16, U16, U16, U8, U1 )
OPCODE(FPRoundInt32, U32, U32, U8, U1 )
OPCODE(FPRoundInt64, U64, U64, U8, U1 )
OPCODE(FPRSqrtEstimate32, U32, U32 )