mirrors/dynarmic
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

IR: Implement IR instructions VectorEqual{8,16,32,64,128}

Browse source
This commit is contained in:
MerryMage 2018-01-26 01:52:06 +00:00
parent 4ce9c65cfb
commit d5283e46e8
4 changed files with 190 additions and 101 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

261
src/backend_x64/emit_x64_vector.cpp
View file

@ -17,6 +17,18 @@ namespace BackendX64 {
using namespace Xbyak::util;
template <typename Function>
static void EmitVectorOperation(BlockOfCode* code, EmitContext& ctx, IR::Inst* inst, Function fn) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm xmm_a = ctx.reg_alloc.UseScratchXmm(args[0]);
Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
(code->*fn)(xmm_a, xmm_b);
ctx.reg_alloc.DefineValue(inst, xmm_a);
}
void EmitX64::EmitVectorGetElement8(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
ASSERT(args[1].IsImmediate());
@ -90,17 +102,6 @@ void EmitX64::EmitVectorGetElement64(EmitContext& ctx, IR::Inst* inst) {
ctx.reg_alloc.DefineValue(inst, dest);
}
static void EmitVectorOperation(BlockOfCode* code, EmitContext& ctx, IR::Inst* inst, void (Xbyak::CodeGenerator::*fn)(const Xbyak::Mmx& mmx, const Xbyak::Operand&)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm xmm_a = ctx.reg_alloc.UseScratchXmm(args[0]);
Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
(code->*fn)(xmm_a, xmm_b);
ctx.reg_alloc.DefineValue(inst, xmm_a);
}
void EmitX64::EmitVectorAdd8(EmitContext& ctx, IR::Inst* inst) {
EmitVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::paddb);
}
@ -121,6 +122,95 @@ void EmitX64::EmitVectorAnd(EmitContext& ctx, IR::Inst* inst) {
EmitVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::pand);
}
void EmitX64::EmitVectorLowerBroadcast8(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
if (code->DoesCpuSupport(Xbyak::util::Cpu::tSSSE3)) {
Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code->pxor(tmp, tmp);
code->pshufb(a, tmp);
code->movq(a, a);
} else {
code->punpcklbw(a, a);
code->pshuflw(a, a, 0);
}
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorLowerBroadcast16(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->pshuflw(a, a, 0);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorLowerBroadcast32(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->pshuflw(a, a, 0b01000100);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorBroadcast8(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
if (code->DoesCpuSupport(Xbyak::util::Cpu::tSSSE3)) {
Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code->pxor(tmp, tmp);
code->pshufb(a, tmp);
} else {
code->punpcklbw(a, a);
code->pshuflw(a, a, 0);
code->punpcklqdq(a, a);
}
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorBroadcast16(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->pshuflw(a, a, 0);
code->punpcklqdq(a, a);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorBroadcast32(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->pshufd(a, a, 0);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorBroadcast64(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->punpcklqdq(a, a);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorOr(EmitContext& ctx, IR::Inst* inst) {
EmitVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::por);
}
@ -141,6 +231,65 @@ void EmitX64::EmitVectorNot(EmitContext& ctx, IR::Inst* inst) {
ctx.reg_alloc.DefineValue(inst, xmm_a);
}
void EmitX64::EmitVectorEqual8(EmitContext& ctx, IR::Inst* inst) {
EmitVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::pcmpeqb);
}
void EmitX64::EmitVectorEqual16(EmitContext& ctx, IR::Inst* inst) {
EmitVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::pcmpeqw);
}
void EmitX64::EmitVectorEqual32(EmitContext& ctx, IR::Inst* inst) {
EmitVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::pcmpeqd);
}
void EmitX64::EmitVectorEqual64(EmitContext& ctx, IR::Inst* inst) {
if (code->DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
EmitVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::pcmpeqq);
return;
}
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm xmm_a = ctx.reg_alloc.UseScratchXmm(args[0]);
Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code->pcmpeqd(xmm_a, xmm_b);
code->pshufd(tmp, xmm_a, 0b10110001);
code->pand(xmm_a, tmp);
ctx.reg_alloc.DefineValue(inst, xmm_a);
}
void EmitX64::EmitVectorEqual128(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
if (code->DoesCpuSupport(Xbyak::util::Cpu::tSSE41)) {
Xbyak::Xmm xmm_a = ctx.reg_alloc.UseScratchXmm(args[0]);
Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code->pcmpeqq(xmm_a, xmm_b);
code->pshufd(tmp, xmm_a, 0b01001110);
code->pand(xmm_a, tmp);
ctx.reg_alloc.DefineValue(inst, xmm_a);
} else {
Xbyak::Xmm xmm_a = ctx.reg_alloc.UseScratchXmm(args[0]);
Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code->pcmpeqd(xmm_a, xmm_b);
code->pshufd(tmp, xmm_a, 0b10110001);
code->pand(xmm_a, tmp);
code->pshufd(tmp, xmm_a, 0b01001110);
code->pand(xmm_a, tmp);
ctx.reg_alloc.DefineValue(inst, xmm_a);
}
}
void EmitX64::EmitVectorLowerPairedAdd8(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
@ -298,96 +447,6 @@ void EmitX64::EmitVectorPairedAdd64(EmitContext& ctx, IR::Inst* inst) {
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorLowerBroadcast8(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
if (code->DoesCpuSupport(Xbyak::util::Cpu::tSSSE3)) {
Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code->pxor(tmp, tmp);
code->pshufb(a, tmp);
code->movq(a, a);
} else {
code->punpcklbw(a, a);
code->pshuflw(a, a, 0);
}
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorLowerBroadcast16(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->pshuflw(a, a, 0);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorLowerBroadcast32(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->pshuflw(a, a, 0b01000100);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorBroadcast8(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
if (code->DoesCpuSupport(Xbyak::util::Cpu::tSSSE3)) {
Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
code->pxor(tmp, tmp);
code->pshufb(a, tmp);
} else {
code->punpcklbw(a, a);
code->pshuflw(a, a, 0);
code->punpcklqdq(a, a);
}
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorBroadcast16(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->pshuflw(a, a, 0);
code->punpcklqdq(a, a);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorBroadcast32(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->pshufd(a, a, 0);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorBroadcast64(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
code->punpcklqdq(a, a);
ctx.reg_alloc.DefineValue(inst, a);
}
void EmitX64::EmitVectorZeroUpper(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);

20
src/frontend/ir/ir_emitter.cpp
View file

@ -745,6 +745,26 @@ U128 IREmitter::VectorBroadcast64(const U64& a) {
return Inst<U128>(Opcode::VectorBroadcast64, a);
}
U128 IREmitter::VectorEqual8(const U128& a, const U128& b) {
return Inst<U128>(Opcode::VectorEqual8, a, b);
}
U128 IREmitter::VectorEqual16(const U128& a, const U128& b) {
return Inst<U128>(Opcode::VectorEqual16, a, b);
}
U128 IREmitter::VectorEqual32(const U128& a, const U128& b) {
return Inst<U128>(Opcode::VectorEqual32, a, b);
}
U128 IREmitter::VectorEqual64(const U128& a, const U128& b) {
return Inst<U128>(Opcode::VectorEqual64, a, b);
}
U128 IREmitter::VectorEqual128(const U128& a, const U128& b) {
return Inst<U128>(Opcode::VectorEqual128, a, b);
}
U128 IREmitter::VectorLowerPairedAdd8(const U128& a, const U128& b) {
return Inst<U128>(Opcode::VectorLowerPairedAdd8, a, b);
}

5
src/frontend/ir/ir_emitter.h
View file

@ -200,6 +200,11 @@ public:
U128 VectorBroadcast16(const U16& a);
U128 VectorBroadcast32(const U32& a);
U128 VectorBroadcast64(const U64& a);
U128 VectorEqual8(const U128& a, const U128& b);
U128 VectorEqual16(const U128& a, const U128& b);
U128 VectorEqual32(const U128& a, const U128& b);
U128 VectorEqual64(const U128& a, const U128& b);
U128 VectorEqual128(const U128& a, const U128& b);
U128 VectorLowerPairedAdd8(const U128& a, const U128& b);
U128 VectorLowerPairedAdd16(const U128& a, const U128& b);
U128 VectorLowerPairedAdd32(const U128& a, const U128& b);

5
src/frontend/ir/opcodes.inc
View file

@ -185,6 +185,11 @@ OPCODE(VectorBroadcast8, T::U128, T::U8
OPCODE(VectorBroadcast16, T::U128, T::U16 )
OPCODE(VectorBroadcast32, T::U128, T::U32 )
OPCODE(VectorBroadcast64, T::U128, T::U64 )
OPCODE(VectorEqual8, T::U128, T::U128, T::U128 )
OPCODE(VectorEqual16, T::U128, T::U128, T::U128 )
OPCODE(VectorEqual32, T::U128, T::U128, T::U128 )
OPCODE(VectorEqual64, T::U128, T::U128, T::U128 )
OPCODE(VectorEqual128, T::U128, T::U128, T::U128 )
OPCODE(VectorLowerPairedAdd8, T::U128, T::U128, T::U128 )
OPCODE(VectorLowerPairedAdd16, T::U128, T::U128, T::U128 )
OPCODE(VectorLowerPairedAdd32, T::U128, T::U128, T::U128 )