IR: Implement VectorSaturated{Signed,Unsigned}{Add,Sub}
This commit is contained in:
MerryMage
parent
3a50d444dc
commit
4e90754873
5 changed files with 471 additions and 0 deletions
|
|
@ -257,6 +257,7 @@ if (ARCHITECTURE_x86_64)
|
||||||
backend/x64/emit_x64_sm4.cpp
|
backend/x64/emit_x64_sm4.cpp
|
||||||
backend/x64/emit_x64_vector.cpp
|
backend/x64/emit_x64_vector.cpp
|
||||||
backend/x64/emit_x64_vector_floating_point.cpp
|
backend/x64/emit_x64_vector_floating_point.cpp
|
||||||
|
backend/x64/emit_x64_vector_saturation.cpp
|
||||||
backend/x64/exception_handler.h
|
backend/x64/exception_handler.h
|
||||||
backend/x64/hostloc.cpp
|
backend/x64/hostloc.cpp
|
||||||
backend/x64/hostloc.h
|
backend/x64/hostloc.h
|
||||||
|
|
|
||||||
370
src/backend/x64/emit_x64_vector_saturation.cpp
Normal file
370
src/backend/x64/emit_x64_vector_saturation.cpp
Normal file
|
|
@ -0,0 +1,370 @@
|
||||||
|
/* This file is part of the dynarmic project.
|
||||||
|
* Copyright (c) 2016 MerryMage
|
||||||
|
* SPDX-License-Identifier: 0BSD
|
||||||
|
*/
|
||||||
|
|
||||||
|
#include "backend/x64/block_of_code.h"
|
||||||
|
#include "backend/x64/emit_x64.h"
|
||||||
|
#include "common/common_types.h"
|
||||||
|
#include "frontend/ir/microinstruction.h"
|
||||||
|
#include "frontend/ir/opcodes.h"
|
||||||
|
|
||||||
|
namespace Dynarmic::Backend::X64 {
|
||||||
|
|
||||||
|
using namespace Xbyak::util;
|
||||||
|
|
||||||
|
namespace {
|
||||||
|
|
||||||
|
void EmitVectorSaturatedNative(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, void (Xbyak::CodeGenerator::*saturated_fn)(const Xbyak::Mmx& mmx, const Xbyak::Operand&), void (Xbyak::CodeGenerator::*unsaturated_fn)(const Xbyak::Mmx& mmx, const Xbyak::Operand&), void (Xbyak::CodeGenerator::*sub_fn)(const Xbyak::Mmx& mmx, const Xbyak::Operand&)) {
|
||||||
|
const auto overflow_inst = inst->GetAssociatedPseudoOperation(IR::Opcode::GetOverflowFromOp);
|
||||||
|
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
|
||||||
|
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
|
||||||
|
const Xbyak::Xmm addend = ctx.reg_alloc.UseXmm(args[1]);
|
||||||
|
|
||||||
|
if (overflow_inst) {
|
||||||
|
code.movaps(xmm0, result);
|
||||||
|
}
|
||||||
|
|
||||||
|
(code.*saturated_fn)(result, addend);
|
||||||
|
|
||||||
|
if (overflow_inst) {
|
||||||
|
const Xbyak::Reg8 overflow = ctx.reg_alloc.ScratchGpr().cvt8();
|
||||||
|
|
||||||
|
(code.*unsaturated_fn)(xmm0, addend);
|
||||||
|
(code.*sub_fn)(xmm0, result);
|
||||||
|
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
|
||||||
|
code.ptest(xmm0, xmm0);
|
||||||
|
} else {
|
||||||
|
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
|
||||||
|
code.pxor(tmp, tmp);
|
||||||
|
code.pcmpeqw(xmm0, tmp);
|
||||||
|
code.pmovmskb(overflow.cvt32(), xmm0);
|
||||||
|
code.xor_(overflow.cvt32(), 0xFFFF);
|
||||||
|
code.test(overflow.cvt32(), overflow.cvt32());
|
||||||
|
}
|
||||||
|
code.setnz(overflow);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(overflow_inst, overflow);
|
||||||
|
ctx.EraseInstruction(overflow_inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(inst, result);
|
||||||
|
}
|
||||||
|
|
||||||
|
enum class Op {
|
||||||
|
Add,
|
||||||
|
Sub,
|
||||||
|
};
|
||||||
|
|
||||||
|
template<Op op, size_t esize>
|
||||||
|
void EmitVectorSignedSaturated(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
static_assert(esize == 32 || esize == 64);
|
||||||
|
constexpr u64 msb_mask = esize == 32 ? 0x8000000080000000 : 0x8000000000000000;
|
||||||
|
|
||||||
|
const auto overflow_inst = inst->GetAssociatedPseudoOperation(IR::Opcode::GetOverflowFromOp);
|
||||||
|
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
|
||||||
|
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
|
||||||
|
const Xbyak::Xmm arg = ctx.reg_alloc.UseXmm(args[1]);
|
||||||
|
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
|
||||||
|
|
||||||
|
// TODO AVX-512: vpternlog, vpsraq
|
||||||
|
// TODO AVX2 implementation
|
||||||
|
|
||||||
|
code.movaps(xmm0, result);
|
||||||
|
code.movaps(tmp, result);
|
||||||
|
|
||||||
|
if constexpr (op == Op::Add) {
|
||||||
|
if constexpr (esize == 32) {
|
||||||
|
code.paddd(result, arg);
|
||||||
|
} else {
|
||||||
|
code.paddq(result, arg);
|
||||||
|
}
|
||||||
|
} else {
|
||||||
|
if constexpr (esize == 32) {
|
||||||
|
code.psubd(result, arg);
|
||||||
|
} else {
|
||||||
|
code.psubq(result, arg);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
code.pxor(tmp, result);
|
||||||
|
code.pxor(xmm0, arg);
|
||||||
|
if constexpr (op == Op::Add) {
|
||||||
|
code.pandn(xmm0, tmp);
|
||||||
|
} else {
|
||||||
|
code.pand(xmm0, tmp);
|
||||||
|
}
|
||||||
|
|
||||||
|
code.movaps(tmp, result);
|
||||||
|
code.psrad(tmp, 31);
|
||||||
|
if constexpr (esize == 64) {
|
||||||
|
code.pshufd(tmp, tmp, 0b11110101);
|
||||||
|
}
|
||||||
|
code.pxor(tmp, code.MConst(xword, msb_mask, msb_mask));
|
||||||
|
|
||||||
|
if (overflow_inst) {
|
||||||
|
const Xbyak::Reg8 overflow = ctx.reg_alloc.ScratchGpr().cvt8();
|
||||||
|
|
||||||
|
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
|
||||||
|
code.ptest(xmm0, code.MConst(xword, msb_mask, msb_mask));
|
||||||
|
} else {
|
||||||
|
if constexpr (esize == 32) {
|
||||||
|
code.movmskps(overflow.cvt32(), xmm0);
|
||||||
|
} else {
|
||||||
|
code.movmskpd(overflow.cvt32(), xmm0);
|
||||||
|
}
|
||||||
|
code.test(overflow.cvt32(), overflow.cvt32());
|
||||||
|
}
|
||||||
|
code.setnz(overflow);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(overflow_inst, overflow);
|
||||||
|
ctx.EraseInstruction(overflow_inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
|
||||||
|
if constexpr (esize == 32) {
|
||||||
|
code.blendvps(result, tmp);
|
||||||
|
} else {
|
||||||
|
code.blendvpd(result, tmp);
|
||||||
|
}
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(inst, result);
|
||||||
|
} else {
|
||||||
|
code.psrad(xmm0, 31);
|
||||||
|
if constexpr (esize == 64) {
|
||||||
|
code.pshufd(xmm0, xmm0, 0b11110101);
|
||||||
|
}
|
||||||
|
|
||||||
|
code.pand(tmp, xmm0);
|
||||||
|
code.pandn(xmm0, result);
|
||||||
|
code.por(tmp, xmm0);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(inst, tmp);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
} // anonymous namespace
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorSignedSaturatedAdd8(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSaturatedNative(code, ctx, inst, &Xbyak::CodeGenerator::paddsb, &Xbyak::CodeGenerator::paddb, &Xbyak::CodeGenerator::psubb);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorSignedSaturatedAdd16(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSaturatedNative(code, ctx, inst, &Xbyak::CodeGenerator::paddsw, &Xbyak::CodeGenerator::paddw, &Xbyak::CodeGenerator::psubw);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorSignedSaturatedAdd32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSignedSaturated<Op::Add, 32>(code, ctx, inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorSignedSaturatedAdd64(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSignedSaturated<Op::Add, 64>(code, ctx, inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorSignedSaturatedSub8(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSaturatedNative(code, ctx, inst, &Xbyak::CodeGenerator::psubsb, &Xbyak::CodeGenerator::psubb, &Xbyak::CodeGenerator::psubb);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorSignedSaturatedSub16(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSaturatedNative(code, ctx, inst, &Xbyak::CodeGenerator::psubsw, &Xbyak::CodeGenerator::psubw, &Xbyak::CodeGenerator::psubw);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorSignedSaturatedSub32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSignedSaturated<Op::Sub, 32>(code, ctx, inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorSignedSaturatedSub64(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSignedSaturated<Op::Sub, 64>(code, ctx, inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorUnsignedSaturatedAdd8(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSaturatedNative(code, ctx, inst, &Xbyak::CodeGenerator::paddusb, &Xbyak::CodeGenerator::paddb, &Xbyak::CodeGenerator::psubb);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorUnsignedSaturatedAdd16(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSaturatedNative(code, ctx, inst, &Xbyak::CodeGenerator::paddusw, &Xbyak::CodeGenerator::paddw, &Xbyak::CodeGenerator::psubw);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorUnsignedSaturatedAdd32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
const auto overflow_inst = inst->GetAssociatedPseudoOperation(IR::Opcode::GetOverflowFromOp);
|
||||||
|
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
|
||||||
|
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
|
||||||
|
const Xbyak::Xmm addend = ctx.reg_alloc.UseXmm(args[1]);
|
||||||
|
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
|
||||||
|
|
||||||
|
// TODO AVX2, AVX-512: vpternlog
|
||||||
|
|
||||||
|
code.movaps(tmp, result);
|
||||||
|
code.movaps(xmm0, result);
|
||||||
|
|
||||||
|
code.pxor(xmm0, addend);
|
||||||
|
code.pand(tmp, addend);
|
||||||
|
code.paddd(result, addend);
|
||||||
|
|
||||||
|
code.psrld(xmm0, 1);
|
||||||
|
code.paddd(tmp, xmm0);
|
||||||
|
code.psrad(tmp, 31);
|
||||||
|
|
||||||
|
code.por(result, tmp);
|
||||||
|
|
||||||
|
if (overflow_inst) {
|
||||||
|
const Xbyak::Reg8 overflow = ctx.reg_alloc.ScratchGpr().cvt8();
|
||||||
|
|
||||||
|
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
|
||||||
|
code.ptest(tmp, tmp);
|
||||||
|
} else {
|
||||||
|
code.movmskps(overflow.cvt32(), tmp);
|
||||||
|
code.test(overflow.cvt32(), overflow.cvt32());
|
||||||
|
}
|
||||||
|
code.setnz(overflow);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(overflow_inst, overflow);
|
||||||
|
ctx.EraseInstruction(overflow_inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(inst, result);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorUnsignedSaturatedAdd64(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
const auto overflow_inst = inst->GetAssociatedPseudoOperation(IR::Opcode::GetOverflowFromOp);
|
||||||
|
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
|
||||||
|
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
|
||||||
|
const Xbyak::Xmm addend = ctx.reg_alloc.UseXmm(args[1]);
|
||||||
|
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
|
||||||
|
|
||||||
|
// TODO AVX2, AVX-512: vpternlog
|
||||||
|
|
||||||
|
code.movaps(tmp, result);
|
||||||
|
code.movaps(xmm0, result);
|
||||||
|
|
||||||
|
code.pxor(xmm0, addend);
|
||||||
|
code.pand(tmp, addend);
|
||||||
|
code.paddq(result, addend);
|
||||||
|
|
||||||
|
code.psrlq(xmm0, 1);
|
||||||
|
code.paddq(tmp, xmm0);
|
||||||
|
code.psrad(tmp, 31);
|
||||||
|
code.pshufd(tmp, tmp, 0b11110101);
|
||||||
|
|
||||||
|
code.por(result, tmp);
|
||||||
|
|
||||||
|
if (overflow_inst) {
|
||||||
|
const Xbyak::Reg8 overflow = ctx.reg_alloc.ScratchGpr().cvt8();
|
||||||
|
|
||||||
|
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
|
||||||
|
code.ptest(tmp, tmp);
|
||||||
|
} else {
|
||||||
|
code.movmskpd(overflow.cvt32(), tmp);
|
||||||
|
code.test(overflow.cvt32(), overflow.cvt32());
|
||||||
|
}
|
||||||
|
code.setnz(overflow);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(overflow_inst, overflow);
|
||||||
|
ctx.EraseInstruction(overflow_inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(inst, result);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorUnsignedSaturatedSub8(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSaturatedNative(code, ctx, inst, &Xbyak::CodeGenerator::psubusb, &Xbyak::CodeGenerator::psubb, &Xbyak::CodeGenerator::psubb);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorUnsignedSaturatedSub16(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
EmitVectorSaturatedNative(code, ctx, inst, &Xbyak::CodeGenerator::psubusw, &Xbyak::CodeGenerator::psubw, &Xbyak::CodeGenerator::psubw);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorUnsignedSaturatedSub32(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
const auto overflow_inst = inst->GetAssociatedPseudoOperation(IR::Opcode::GetOverflowFromOp);
|
||||||
|
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
|
||||||
|
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
|
||||||
|
const Xbyak::Xmm subtrahend = ctx.reg_alloc.UseXmm(args[1]);
|
||||||
|
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
|
||||||
|
|
||||||
|
// TODO AVX2, AVX-512: vpternlog
|
||||||
|
|
||||||
|
code.movaps(tmp, result);
|
||||||
|
code.movaps(xmm0, subtrahend);
|
||||||
|
|
||||||
|
code.pxor(tmp, subtrahend);
|
||||||
|
code.psubd(result, subtrahend);
|
||||||
|
code.pand(xmm0, tmp);
|
||||||
|
|
||||||
|
code.psrld(tmp, 1);
|
||||||
|
code.psubd(tmp, xmm0);
|
||||||
|
code.psrad(tmp, 31);
|
||||||
|
|
||||||
|
if (overflow_inst) {
|
||||||
|
const Xbyak::Reg8 overflow = ctx.reg_alloc.ScratchGpr().cvt8();
|
||||||
|
|
||||||
|
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
|
||||||
|
code.ptest(tmp, tmp);
|
||||||
|
} else {
|
||||||
|
code.movmskps(overflow.cvt32(), tmp);
|
||||||
|
code.test(overflow.cvt32(), overflow.cvt32());
|
||||||
|
}
|
||||||
|
code.setnz(overflow);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(overflow_inst, overflow);
|
||||||
|
ctx.EraseInstruction(overflow_inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
code.pandn(tmp, result);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(inst, tmp);
|
||||||
|
}
|
||||||
|
|
||||||
|
void EmitX64::EmitVectorUnsignedSaturatedSub64(EmitContext& ctx, IR::Inst* inst) {
|
||||||
|
const auto overflow_inst = inst->GetAssociatedPseudoOperation(IR::Opcode::GetOverflowFromOp);
|
||||||
|
|
||||||
|
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
|
||||||
|
|
||||||
|
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
|
||||||
|
const Xbyak::Xmm subtrahend = ctx.reg_alloc.UseXmm(args[1]);
|
||||||
|
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
|
||||||
|
|
||||||
|
// TODO AVX2, AVX-512: vpternlog
|
||||||
|
|
||||||
|
code.movaps(tmp, result);
|
||||||
|
code.movaps(xmm0, subtrahend);
|
||||||
|
|
||||||
|
code.pxor(tmp, subtrahend);
|
||||||
|
code.psubq(result, subtrahend);
|
||||||
|
code.pand(xmm0, tmp);
|
||||||
|
|
||||||
|
code.psrlq(tmp, 1);
|
||||||
|
code.psubq(tmp, xmm0);
|
||||||
|
code.psrad(tmp, 31);
|
||||||
|
code.pshufd(tmp, tmp, 0b11110101);
|
||||||
|
|
||||||
|
if (overflow_inst) {
|
||||||
|
const Xbyak::Reg8 overflow = ctx.reg_alloc.ScratchGpr().cvt8();
|
||||||
|
|
||||||
|
if (code.DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
|
||||||
|
code.ptest(tmp, tmp);
|
||||||
|
} else {
|
||||||
|
code.movmskpd(overflow.cvt32(), tmp);
|
||||||
|
code.test(overflow.cvt32(), overflow.cvt32());
|
||||||
|
}
|
||||||
|
code.setnz(overflow);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(overflow_inst, overflow);
|
||||||
|
ctx.EraseInstruction(overflow_inst);
|
||||||
|
}
|
||||||
|
|
||||||
|
code.pandn(tmp, result);
|
||||||
|
|
||||||
|
ctx.reg_alloc.DefineValue(inst, tmp);
|
||||||
|
}
|
||||||
|
|
||||||
|
} // namespace Dynarmic::Backend::X64
|
||||||
|
|
@ -635,6 +635,83 @@ ResultAndOverflow<U32> IREmitter::UnsignedSaturation(const U32& a, size_t bit_si
|
||||||
return {result, overflow};
|
return {result, overflow};
|
||||||
}
|
}
|
||||||
|
|
||||||
|
ResultAndOverflow<U128> IREmitter::VectorSignedSaturatedAdd(size_t esize, const U128& a, const U128& b) {
|
||||||
|
const auto result = [&]{
|
||||||
|
switch (esize) {
|
||||||
|
case 8:
|
||||||
|
return Inst<U128>(Opcode::VectorSignedSaturatedAdd8, a, b);
|
||||||
|
case 16:
|
||||||
|
return Inst<U128>(Opcode::VectorSignedSaturatedAdd16, a, b);
|
||||||
|
case 32:
|
||||||
|
return Inst<U128>(Opcode::VectorSignedSaturatedAdd32, a, b);
|
||||||
|
case 64:
|
||||||
|
return Inst<U128>(Opcode::VectorSignedSaturatedAdd64, a, b);
|
||||||
|
default:
|
||||||
|
UNREACHABLE();
|
||||||
|
}
|
||||||
|
}();
|
||||||
|
const auto overflow = Inst<U1>(Opcode::GetOverflowFromOp, result);
|
||||||
|
return {result, overflow};
|
||||||
|
}
|
||||||
|
|
||||||
|
ResultAndOverflow<U128> IREmitter::VectorSignedSaturatedSub(size_t esize, const U128& a, const U128& b) {
|
||||||
|
const auto result = [&]{
|
||||||
|
switch (esize) {
|
||||||
|
case 8:
|
||||||
|
return Inst<U128>(Opcode::VectorSignedSaturatedSub8, a, b);
|
||||||
|
case 16:
|
||||||
|
return Inst<U128>(Opcode::VectorSignedSaturatedSub16, a, b);
|
||||||
|
case 32:
|
||||||
|
return Inst<U128>(Opcode::VectorSignedSaturatedSub32, a, b);
|
||||||
|
case 64:
|
||||||
|
return Inst<U128>(Opcode::VectorSignedSaturatedSub64, a, b);
|
||||||
|
default:
|
||||||
|
UNREACHABLE();
|
||||||
|
}
|
||||||
|
}();
|
||||||
|
const auto overflow = Inst<U1>(Opcode::GetOverflowFromOp, result);
|
||||||
|
return {result, overflow};
|
||||||
|
}
|
||||||
|
|
||||||
|
ResultAndOverflow<U128> IREmitter::VectorUnsignedSaturatedAdd(size_t esize, const U128& a, const U128& b) {
|
||||||
|
const auto result = [&]{
|
||||||
|
switch (esize) {
|
||||||
|
case 8:
|
||||||
|
return Inst<U128>(Opcode::VectorUnsignedSaturatedAdd8, a, b);
|
||||||
|
case 16:
|
||||||
|
return Inst<U128>(Opcode::VectorUnsignedSaturatedAdd16, a, b);
|
||||||
|
case 32:
|
||||||
|
return Inst<U128>(Opcode::VectorUnsignedSaturatedAdd32, a, b);
|
||||||
|
case 64:
|
||||||
|
return Inst<U128>(Opcode::VectorUnsignedSaturatedAdd64, a, b);
|
||||||
|
default:
|
||||||
|
UNREACHABLE();
|
||||||
|
}
|
||||||
|
}();
|
||||||
|
const auto overflow = Inst<U1>(Opcode::GetOverflowFromOp, result);
|
||||||
|
return {result, overflow};
|
||||||
|
}
|
||||||
|
|
||||||
|
ResultAndOverflow<U128> IREmitter::VectorUnsignedSaturatedSub(size_t esize, const U128& a, const U128& b) {
|
||||||
|
const auto result = [&]{
|
||||||
|
switch (esize) {
|
||||||
|
case 8:
|
||||||
|
return Inst<U128>(Opcode::VectorUnsignedSaturatedSub8, a, b);
|
||||||
|
case 16:
|
||||||
|
return Inst<U128>(Opcode::VectorUnsignedSaturatedSub16, a, b);
|
||||||
|
case 32:
|
||||||
|
return Inst<U128>(Opcode::VectorUnsignedSaturatedSub32, a, b);
|
||||||
|
case 64:
|
||||||
|
return Inst<U128>(Opcode::VectorUnsignedSaturatedSub64, a, b);
|
||||||
|
default:
|
||||||
|
UNREACHABLE();
|
||||||
|
}
|
||||||
|
}();
|
||||||
|
const auto overflow = Inst<U1>(Opcode::GetOverflowFromOp, result);
|
||||||
|
return {result, overflow};
|
||||||
|
}
|
||||||
|
|
||||||
|
|
||||||
ResultAndGE<U32> IREmitter::PackedAddU8(const U32& a, const U32& b) {
|
ResultAndGE<U32> IREmitter::PackedAddU8(const U32& a, const U32& b) {
|
||||||
const auto result = Inst<U32>(Opcode::PackedAddU8, a, b);
|
const auto result = Inst<U32>(Opcode::PackedAddU8, a, b);
|
||||||
const auto ge = Inst<U32>(Opcode::GetGEFromOp, result);
|
const auto ge = Inst<U32>(Opcode::GetGEFromOp, result);
|
||||||
|
|
|
||||||
|
|
@ -166,6 +166,11 @@ public:
|
||||||
ResultAndOverflow<UAny> UnsignedSaturatedSub(const UAny& a, const UAny& b);
|
ResultAndOverflow<UAny> UnsignedSaturatedSub(const UAny& a, const UAny& b);
|
||||||
ResultAndOverflow<U32> UnsignedSaturation(const U32& a, size_t bit_size_to_saturate_to);
|
ResultAndOverflow<U32> UnsignedSaturation(const U32& a, size_t bit_size_to_saturate_to);
|
||||||
|
|
||||||
|
ResultAndOverflow<U128> VectorSignedSaturatedAdd(size_t esize, const U128& a, const U128& b);
|
||||||
|
ResultAndOverflow<U128> VectorSignedSaturatedSub(size_t esize, const U128& a, const U128& b);
|
||||||
|
ResultAndOverflow<U128> VectorUnsignedSaturatedAdd(size_t esize, const U128& a, const U128& b);
|
||||||
|
ResultAndOverflow<U128> VectorUnsignedSaturatedSub(size_t esize, const U128& a, const U128& b);
|
||||||
|
|
||||||
ResultAndGE<U32> PackedAddU8(const U32& a, const U32& b);
|
ResultAndGE<U32> PackedAddU8(const U32& a, const U32& b);
|
||||||
ResultAndGE<U32> PackedAddS8(const U32& a, const U32& b);
|
ResultAndGE<U32> PackedAddS8(const U32& a, const U32& b);
|
||||||
ResultAndGE<U32> PackedAddU16(const U32& a, const U32& b);
|
ResultAndGE<U32> PackedAddU16(const U32& a, const U32& b);
|
||||||
|
|
|
||||||
|
|
@ -194,6 +194,24 @@ OPCODE(UnsignedSaturatedSub32, U32, U32,
|
||||||
OPCODE(UnsignedSaturatedSub64, U64, U64, U64 )
|
OPCODE(UnsignedSaturatedSub64, U64, U64, U64 )
|
||||||
OPCODE(UnsignedSaturation, U32, U32, U8 )
|
OPCODE(UnsignedSaturation, U32, U32, U8 )
|
||||||
|
|
||||||
|
// Vector saturated instructions
|
||||||
|
OPCODE(VectorSignedSaturatedAdd8, U128, U128, U128 )
|
||||||
|
OPCODE(VectorSignedSaturatedAdd16, U128, U128, U128 )
|
||||||
|
OPCODE(VectorSignedSaturatedAdd32, U128, U128, U128 )
|
||||||
|
OPCODE(VectorSignedSaturatedAdd64, U128, U128, U128 )
|
||||||
|
OPCODE(VectorSignedSaturatedSub8, U128, U128, U128 )
|
||||||
|
OPCODE(VectorSignedSaturatedSub16, U128, U128, U128 )
|
||||||
|
OPCODE(VectorSignedSaturatedSub32, U128, U128, U128 )
|
||||||
|
OPCODE(VectorSignedSaturatedSub64, U128, U128, U128 )
|
||||||
|
OPCODE(VectorUnsignedSaturatedAdd8, U128, U128, U128 )
|
||||||
|
OPCODE(VectorUnsignedSaturatedAdd16, U128, U128, U128 )
|
||||||
|
OPCODE(VectorUnsignedSaturatedAdd32, U128, U128, U128 )
|
||||||
|
OPCODE(VectorUnsignedSaturatedAdd64, U128, U128, U128 )
|
||||||
|
OPCODE(VectorUnsignedSaturatedSub8, U128, U128, U128 )
|
||||||
|
OPCODE(VectorUnsignedSaturatedSub16, U128, U128, U128 )
|
||||||
|
OPCODE(VectorUnsignedSaturatedSub32, U128, U128, U128 )
|
||||||
|
OPCODE(VectorUnsignedSaturatedSub64, U128, U128, U128 )
|
||||||
|
|
||||||
// Packed instructions
|
// Packed instructions
|
||||||
OPCODE(PackedAddU8, U32, U32, U32 )
|
OPCODE(PackedAddU8, U32, U32, U32 )
|
||||||
OPCODE(PackedAddS8, U32, U32, U32 )
|
OPCODE(PackedAddS8, U32, U32, U32 )
|
||||||
|
|
|
||||||
Loading…
Reference in a new issue