| Issue |
240763
|
| Summary |
[X86][AVX512] failed to fold vfixupimm pattern
|
| Labels |
backend:X86,
missed-optimization,
floating-point
|
| Assignees |
|
| Reporter |
Andarwinux
|
https://wunkolo.github.io/post/2022/10/vfixupimm-signum/
```c++
#include <cstdint>
#include <immintrin.h>
float signum(float Value)
{
if( Value < 0.0 )
return -1.0;
if( Value > 0.0 )
return 1.0;
return 0.0;
}
// Substitution codes for use with vfixupimm
enum class FpFixup : std::uint8_t
{
Dest = 0b0000, // Preserve destination
Norm_Src = 0b0001, // Source operand (Denormal as positive-zero)
QNaN_Src = 0b0010, // QNaN with sign of source (Denormal as positive-zero)
IndefNaN = 0b0011, // Indefinite QNaN (Negative QNaN with no payload)
NegInf = 0b0100, // -Infinity
PosInf = 0b0101, // +Infinity
Inf_Src = 0b0110, // Infinity with sign of source (Denormal as positive-zero)
NegZero = 0b0111, // -0.0
PosZero = 0b1000, // +0.0
NegOne = 0b1001, // -1.0
PosOne = 0b1010, // +1.0
Half = 0b1011, // 0.5
Ninety = 0b1100, // 90.0
HalfPi = 0b1101, // PI/2
PosMax = 0b1110, // +{FLT_MAX,DBL_MAX}
NegMax = 0b1111, // -{FLT_MAX,DBL_MAX}
};
// Generates a 32-bit LUT element for the vfixupimm instruction
constexpr std::uint32_t FixupLUT(
FpFixup src_qnan = FpFixup::Dest, FpFixup src_snan = FpFixup::Dest,
FpFixup src_zero = FpFixup::Dest, FpFixup src_posone = FpFixup::Dest,
FpFixup src_neginf = FpFixup::Dest, FpFixup src_posinf = FpFixup::Dest,
FpFixup src_neg = FpFixup::Dest, FpFixup src_pos = FpFixup::Dest)
{
std::uint32_t fixup_lut = 0;
fixup_lut |= ((static_cast<std::uint32_t>(src_qnan ) & 0b1111) << 0);
fixup_lut |= ((static_cast<std::uint32_t>(src_snan ) & 0b1111) << 4);
fixup_lut |= ((static_cast<std::uint32_t>(src_zero ) & 0b1111) << 8);
fixup_lut |= ((static_cast<std::uint32_t>(src_posone) & 0b1111) << 12);
fixup_lut |= ((static_cast<std::uint32_t>(src_neginf) & 0b1111) << 16);
fixup_lut |= ((static_cast<std::uint32_t>(src_posinf) & 0b1111) << 20);
fixup_lut |= ((static_cast<std::uint32_t>(src_neg ) & 0b1111) << 24);
fixup_lut |= ((static_cast<std::uint32_t>(src_pos ) & 0b1111) << 28);
return fixup_lut;
}
float signum_avx512(__m128 Value)
{
constexpr std::uint32_t FixupOp = FixupLUT(
FpFixup::Dest, FpFixup::Dest, FpFixup::PosZero, FpFixup::PosOne,
FpFixup::NegOne, FpFixup::PosOne, FpFixup::NegOne, FpFixup::PosOne);
const __m128 FixUp
= _mm_fixupimm_ss(Value, Value, _mm_cvtsi32_si128(FixupOp), 0);
return _mm_cvtss_f32(FixUp);
}
```
```asm
.LCPI0_0:
.long 0x3f800000
.LCPI0_1:
.long 0xbf800000
signum(float):
vxorps xmm2, xmm2, xmm2
vcmpltss k1, xmm2, xmm0
vmovss xmm1 {k1} {z}, dword ptr [rip + .LCPI0_0]
vcmpltss k1, xmm0, xmm2
vmovss xmm1 {k1}, dword ptr [rip + .LCPI0_1]
vmovaps xmm0, xmm1
ret
.LCPI1_0:
.long 2846468096
.long 0
.long 0
.long 0
signum_avx512(float vector[4]):
vfixupimmss xmm0, xmm0, dword ptr [rip + .LCPI1_0], 0
ret
```
https://godbolt.org/z/nPj4Tbsa8
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs