ShuriZma
/
dolphin
mirror of https://github.com/dolphin-emu/dolphin.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

JitArm64: Skip checking last input for NaN for non-SIMD operations 

AArch64's handling of NaNs in arithmetic instructions matches PowerPC's
as long as no more than one of the operands is NaN. If we know that all
inputs except the last input are non-NaN, we can therefore skip checking
the last input. This is an optimization that in principle only works for
non-SIMD operations, but ps_sumX effectively is non-SIMD as far as the
arithmetic part of it is concerned, so we can use it there too.
This commit is contained in:
JosJuice 2023-08-09 19:46:27 +02:00
parent 95f06ef231
commit 8274dcbfe4
2 changed files with 43 additions and 66 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

47
Source/Core/Core/PowerPC/JitArm64/JitArm64_FloatingPoint.cpp
View File

@ -80,9 +80,6 @@ void JitArm64::fp_arith(UGeckoInstruction inst)
const bool fma = use_b && use_c;
const bool negate_result = (op5 & ~0x1) == 30;
// Addition and subtraction can't generate new NaNs, they can only take NaNs from inputs
const bool can_generate_nan = (op5 & ~0x1) != 20;
const bool output_is_single = inst.OPCD == 59;
const bool inaccurate_fma = op5 > 25 && !Config::Get(Config::SESSION_USE_FMA);
const bool round_c = use_c && output_is_single && !js.op->fprIsSingle[inst.FC];
@ -203,45 +200,35 @@ void JitArm64::fp_arith(UGeckoInstruction inst)
if (use_c && VA != VC && (!use_b || VB != VC))
inputs.push_back(VC);
// If any inputs are NaNs, pick the first NaN of them and set its quiet bit
for (size_t i = 0; i < inputs.size(); ++i)
// If any inputs are NaNs, pick the first NaN of them and set its quiet bit.
// However, we can skip checking the last input, because if exactly one input is NaN, AArch64
// arithmetic instructions automatically pick that NaN and make it quiet, just like we want.
for (size_t i = 0; i < inputs.size() - 1; ++i)
{
// Skip checking if the input is a NaN if it's the last input and we're guaranteed to have at
// least one NaN input
const bool check_input = can_generate_nan || i != inputs.size() - 1;
const ARM64Reg input = inputs[i];
FixupBranch skip;
if (check_input)
{
m_float_emit.FCMP(input);
skip = B(CCFlags::CC_VC);
}
m_float_emit.FCMP(input);
FixupBranch skip = B(CCFlags::CC_VC);
// Make the NaN quiet
m_float_emit.FADD(VD, input, input);
nan_fixups.push_back(B());
if (check_input)
SetJumpTarget(skip);
SetJumpTarget(skip);
}
std::optional<FixupBranch> nan_early_fixup;
if (can_generate_nan)
if (negate_result)
{
// There was no NaN in any of the inputs, so the NaN must have been generated by the
// arithmetic instruction. In this case, the result is already correct.
if (negate_result)
{
if (result_reg != VD)
m_float_emit.MOV(EncodeRegToDouble(VD), EncodeRegToDouble(result_reg));
nan_fixups.push_back(B());
}
else
{
nan_early_fixup = B();
}
// If we have a NaN, we must not execute FNEG.
if (result_reg != VD)
m_float_emit.MOV(EncodeRegToDouble(VD), EncodeRegToDouble(result_reg));
nan_fixups.push_back(B());
}
else
{
nan_early_fixup = B();
}
SwitchToNearCode();

62
Source/Core/Core/PowerPC/JitArm64/JitArm64_Paired.cpp
View File

@ -384,45 +384,38 @@ void JitArm64::ps_sumX(UGeckoInstruction inst)
m_float_emit.DUP(size, reg_encoder(V0), reg_encoder(VB), 1);
FixupBranch a_nan_done, b_nan_done;
FixupBranch a_nan_done;
if (m_accurate_nans)
{
const auto check_nan = [&](ARM64Reg input) {
m_float_emit.FCMP(scalar_reg_encoder(input));
FixupBranch not_nan = B(CCFlags::CC_VC);
FixupBranch nan = B();
SetJumpTarget(not_nan);
m_float_emit.FCMP(scalar_reg_encoder(VA));
FixupBranch a_not_nan = B(CCFlags::CC_VC);
FixupBranch a_nan = B();
SetJumpTarget(a_not_nan);
SwitchToFarCode();
SetJumpTarget(nan);
SwitchToFarCode();
SetJumpTarget(a_nan);
if (upper)
{
m_float_emit.FADD(scalar_reg_encoder(V0), scalar_reg_encoder(input),
scalar_reg_encoder(input));
m_float_emit.TRN1(size, reg_encoder(VD), reg_encoder(VC), reg_encoder(V0));
}
else if (d != c)
{
m_float_emit.FADD(scalar_reg_encoder(VD), scalar_reg_encoder(input),
scalar_reg_encoder(input));
m_float_emit.INS(size, VD, 1, VC, 1);
}
else
{
m_float_emit.FADD(scalar_reg_encoder(V0), scalar_reg_encoder(input),
scalar_reg_encoder(input));
m_float_emit.INS(size, VD, 0, V0, 0);
}
if (upper)
{
m_float_emit.FADD(scalar_reg_encoder(V0), scalar_reg_encoder(VA), scalar_reg_encoder(VA));
m_float_emit.TRN1(size, reg_encoder(VD), reg_encoder(VC), reg_encoder(V0));
}
else if (d != c)
{
m_float_emit.FADD(scalar_reg_encoder(VD), scalar_reg_encoder(VA), scalar_reg_encoder(VA));
m_float_emit.INS(size, VD, 1, VC, 1);
}
else
{
m_float_emit.FADD(scalar_reg_encoder(V0), scalar_reg_encoder(VA), scalar_reg_encoder(VA));
m_float_emit.INS(size, VD, 0, V0, 0);
}
FixupBranch nan_done = B();
SwitchToNearCode();
FixupBranch a_nan_done = B();
SwitchToNearCode();
return nan_done;
};
a_nan_done = check_nan(VA);
b_nan_done = check_nan(V0);
// If exactly one input is NaN, AArch64 arithmetic instructions automatically pick that NaN
// and make it quiet, just like we want. So if rA isn't NaN, we can skip checking rB.
}
if (upper)
@ -442,10 +435,7 @@ void JitArm64::ps_sumX(UGeckoInstruction inst)
}
if (m_accurate_nans)
{
SetJumpTarget(a_nan_done);
SetJumpTarget(b_nan_done);
}
fpr.Unlock(V0);
if (temp_gpr != ARM64Reg::INVALID_REG)