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

emit_x64_floating_point: Optimize 64-bit EmitFPRSqrtEstimate

Browse source
This commit is contained in:
MerryMage 2021-04-25 22:44:12 +01:00
parent 7bc9e36ed7
commit 3f74a839b9
1 changed files with 57 additions and 26 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

83
src/backend/x64/emit_x64_floating_point.cpp
View file

@ -1005,47 +1005,32 @@ static void EmitFPRSqrtEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* i
const Xbyak::Xmm value = ctx.reg_alloc.ScratchXmm(); const Xbyak::Xmm value = ctx.reg_alloc.ScratchXmm();
[[maybe_unused]] const Xbyak::Reg32 tmp = ctx.reg_alloc.ScratchGpr().cvt32(); [[maybe_unused]] const Xbyak::Reg32 tmp = ctx.reg_alloc.ScratchGpr().cvt32();
Xbyak::Label fallback, bad_values, end; Xbyak::Label fallback, bad_values, end, default_nan;
bool needs_fallback = false;
if constexpr (fsize == 64) { code.movaps(value, operand);
code.cvtsd2ss(value, operand);
if (ctx.FPCR().RMode() == FP::RoundingMode::TowardsMinusInfinity || ctx.FPCR().RMode() == FP::RoundingMode::TowardsZero) { code.movaps(xmm0, code.MConst(xword, fsize == 32 ? 0xFFFF8000 : 0xFFFF'F000'0000'0000));
code.ucomiss(value, code.MConst(xword, FP::FPInfo<u32>::MaxNormal(false)));
code.je(bad_values, code.T_NEAR);
}
} else {
code.movaps(value, operand);
}
code.movaps(xmm0, code.MConst(xword, 0xFFFF8000));
code.pand(value, xmm0); code.pand(value, xmm0);
code.por(value, code.MConst(xword, 0x00008000)); code.por(value, code.MConst(xword, fsize == 32 ? 0x00008000 : 0x0000'1000'0000'0000));
// Detect NaNs, negatives, zeros, denormals and infinities // Detect NaNs, negatives, zeros, denormals and infinities
code.ucomiss(value, code.MConst(xword, 0x00800000)); FCODE(ucomis)(value, code.MConst(xword, FPT(1) << FP::FPInfo<FPT>::explicit_mantissa_width));
code.jna(bad_values, code.T_NEAR); code.jna(bad_values, code.T_NEAR);
code.sqrtss(value, value); FCODE(sqrts)(value, value);
ICODE(mov)(result, code.MConst(xword, FP::FPValue<FPT, false, 0, 1>()));
FCODE(divs)(result, value);
code.movd(result, code.MConst(xword, 0x3F800000)); ICODE(padd)(result, code.MConst(xword, fsize == 32 ? 0x00004000 : 0x0000'0800'0000'0000));
code.divss(result, value);
code.paddd(result, code.MConst(xword, 0x00004000));
code.pand(result, xmm0); code.pand(result, xmm0);
if constexpr (fsize == 64) {
code.cvtss2sd(result, result);
}
code.L(end); code.L(end);
code.SwitchToFarCode(); code.SwitchToFarCode();
code.L(bad_values); code.L(bad_values);
bool needs_fallback = false;
if constexpr (fsize == 32) { if constexpr (fsize == 32) {
Xbyak::Label default_nan;
code.movd(tmp, operand); code.movd(tmp, operand);
if (!ctx.FPCR().FZ()) { if (!ctx.FPCR().FZ()) {
@ -1079,7 +1064,53 @@ static void EmitFPRSqrtEstimate(BlockOfCode& code, EmitContext& ctx, IR::Inst* i
code.movd(result, code.MConst(xword, 0x7FC00000)); code.movd(result, code.MConst(xword, 0x7FC00000));
code.jmp(end, code.T_NEAR); code.jmp(end, code.T_NEAR);
} else { } else {
needs_fallback = true; Xbyak::Label nan, zero;
code.movaps(value, operand);
DenormalsAreZero<fsize>(code, ctx, {value});
code.pxor(result, result);
code.ucomisd(value, result);
if (ctx.FPCR().DN()) {
code.jc(default_nan);
code.je(zero);
} else {
code.jp(nan);
code.je(zero);
code.jc(default_nan);
}
if (!ctx.FPCR().FZ()) {
needs_fallback = true;
code.jmp(fallback);
} else {
// result = 0
code.jmp(end, code.T_NEAR);
}
code.L(zero);
if (code.HasAVX()) {
code.vpor(result, value, code.MConst(xword, 0x7FF0'0000'0000'0000));
} else {
code.movaps(result, value);
code.por(result, code.MConst(xword, 0x7FF0'0000'0000'0000));
}
code.jmp(end, code.T_NEAR);
code.L(nan);
if (!ctx.FPCR().DN()) {
if (code.HasAVX()) {
code.vpor(result, operand, code.MConst(xword, 0x0008'0000'0000'0000));
} else {
code.movaps(result, operand);
code.por(result, code.MConst(xword, 0x0008'0000'0000'0000));
}
code.jmp(end, code.T_NEAR);
}
code.L(default_nan);
code.movq(result, code.MConst(xword, 0x7FF8'0000'0000'0000));
code.jmp(end, code.T_NEAR);
} }
code.L(fallback); code.L(fallback);