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

emit_x64{,_vector}_floating_point: Fix non-FMA execution

Browse source 
Avoid repeated calls to GetArgumentInfo
This commit is contained in:
MerryMage 2021-01-02 20:40:32 +00:00
parent 6023bcd8ad
commit 3806284cbe
2 changed files with 33 additions and 41 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

39
src/backend/x64/emit_x64_floating_point.cpp
View file

@ -283,21 +283,7 @@ void FPThreeOp(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Function fn)
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
if (ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) { if (ctx.FPCR().DN() || ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand = ctx.reg_alloc.UseScratchXmm(args[1]);
if constexpr (std::is_member_function_pointer_v<Function>) {
(code.*fn)(result, operand);
} else {
fn(result, operand);
}
ctx.reg_alloc.DefineValue(inst, result);
return;
}
if (ctx.FPCR().DN()) {
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand = ctx.reg_alloc.UseScratchXmm(args[1]); const Xbyak::Xmm operand = ctx.reg_alloc.UseScratchXmm(args[1]);
@ -307,7 +293,9 @@ void FPThreeOp(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, Function fn)
fn(result, operand); fn(result, operand);
} }
if (!ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
ForceToDefaultNaN<fsize>(code, result); ForceToDefaultNaN<fsize>(code, result);
}
ctx.reg_alloc.DefineValue(inst, result); ctx.reg_alloc.DefineValue(inst, result);
return; return;
@ -605,9 +593,9 @@ template<size_t fsize>
static void EmitFPMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) { static void EmitFPMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
using FPT = mp::unsigned_integer_of_size<fsize>; using FPT = mp::unsigned_integer_of_size<fsize>;
if constexpr (fsize != 16) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
if constexpr (fsize != 16) {
if (code.HasFMA() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) { if (code.HasFMA() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
@ -680,7 +668,6 @@ static void EmitFPMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
} }
} }
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
ctx.reg_alloc.HostCall(inst, args[0], args[1], args[2]); ctx.reg_alloc.HostCall(inst, args[0], args[1], args[2]);
code.mov(code.ABI_PARAM4.cvt32(), ctx.FPCR().Value()); code.mov(code.ABI_PARAM4.cvt32(), ctx.FPCR().Value());
#ifdef _WIN32 #ifdef _WIN32
@ -834,10 +821,10 @@ 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 constexpr (fsize != 16) {
if (code.HasFMA() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
if constexpr (fsize != 16) {
if (code.HasFMA() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
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]);
@ -852,8 +839,6 @@ static void EmitFPRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst*
} }
if (code.HasFMA()) { if (code.HasFMA()) {
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]);
@ -888,8 +873,6 @@ static void EmitFPRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst*
} }
if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) { if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(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();
@ -903,7 +886,6 @@ static void EmitFPRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst*
} }
} }
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
ctx.reg_alloc.HostCall(inst, args[0], args[1]); ctx.reg_alloc.HostCall(inst, args[0], args[1]);
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]);
@ -1038,10 +1020,10 @@ template<size_t fsize>
static void EmitFPRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) { static void EmitFPRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
using FPT = mp::unsigned_integer_of_size<fsize>; using FPT = mp::unsigned_integer_of_size<fsize>;
if constexpr (fsize != 16) {
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
if constexpr (fsize != 16) {
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
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();
@ -1055,8 +1037,6 @@ static void EmitFPRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst*
} }
if (code.HasFMA() && code.HasAVX()) { if (code.HasFMA() && code.HasAVX()) {
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]);
@ -1103,8 +1083,6 @@ static void EmitFPRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst*
} }
if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) { if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(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();
@ -1119,7 +1097,6 @@ static void EmitFPRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst*
} }
} }
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
ctx.reg_alloc.HostCall(inst, args[0], args[1]); ctx.reg_alloc.HostCall(inst, args[0], args[1]);
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]);

21
src/backend/x64/emit_x64_vector_floating_point.cpp
View file

@ -989,10 +989,10 @@ void EmitFPVectorMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
}; };
if constexpr (fsize != 16) { if constexpr (fsize != 16) {
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const bool fpcr_controlled = args[3].GetImmediateU1(); const bool fpcr_controlled = args[3].GetImmediateU1();
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm xmm_c = ctx.reg_alloc.UseXmm(args[2]); const Xbyak::Xmm xmm_c = ctx.reg_alloc.UseXmm(args[2]);
@ -1006,6 +1006,9 @@ void EmitFPVectorMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
} }
if (code.HasFMA() && code.HasAVX()) { if (code.HasFMA() && code.HasAVX()) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const bool fpcr_controlled = args[3].GetImmediateU1();
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseXmm(args[0]); const Xbyak::Xmm xmm_a = ctx.reg_alloc.UseXmm(args[0]);
const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]); const Xbyak::Xmm xmm_b = ctx.reg_alloc.UseXmm(args[1]);
@ -1041,6 +1044,8 @@ void EmitFPVectorMulAdd(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
} }
if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) { if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseScratchXmm(args[1]); const Xbyak::Xmm operand2 = ctx.reg_alloc.UseScratchXmm(args[1]);
const Xbyak::Xmm operand3 = ctx.reg_alloc.UseXmm(args[2]); const Xbyak::Xmm operand3 = ctx.reg_alloc.UseXmm(args[2]);
@ -1247,10 +1252,10 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
}; };
if constexpr (fsize != 16) { if constexpr (fsize != 16) {
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const bool fpcr_controlled = args[2].GetImmediateU1(); const bool fpcr_controlled = args[2].GetImmediateU1();
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
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]);
@ -1265,6 +1270,9 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
} }
if (code.HasFMA() && code.HasAVX()) { 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 result = ctx.reg_alloc.ScratchXmm();
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]);
@ -1297,6 +1305,8 @@ static void EmitRecipStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
} }
if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) { if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(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();
@ -1454,10 +1464,10 @@ static void EmitRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
}; };
if constexpr (fsize != 16) { if constexpr (fsize != 16) {
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const bool fpcr_controlled = args[2].GetImmediateU1(); const bool fpcr_controlled = args[2].GetImmediateU1();
if (code.HasFMA() && code.HasAVX() && ctx.HasOptimization(OptimizationFlag::Unsafe_InaccurateNaN)) {
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
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]);
@ -1473,6 +1483,9 @@ static void EmitRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
} }
if (code.HasFMA() && code.HasAVX()) { 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 result = ctx.reg_alloc.ScratchXmm();
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]);
@ -1511,6 +1524,8 @@ static void EmitRSqrtStepFused(BlockOfCode& code, EmitContext& ctx, IR::Inst* in
} }
if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) { if (ctx.HasOptimization(OptimizationFlag::Unsafe_UnfuseFMA)) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm operand1 = ctx.reg_alloc.UseScratchXmm(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();