[x86] Failure to optimize vector shuffle in conversion

[GabrielRavier]

Bugzilla Link	49081
Version	trunk
OS	Linux
Depends On	#35080
CC	@alexey-bataev,@anton-afanasyev,@topperc,@RKSimon,@phoebewang,@rotateright

Extended Description

typedef int v4si __attribute__((vector_size(16)));
typedef float v4sf __attribute__((vector_size(16)));

v4sf f(v4si f)
{
    return (v4sf){(float)f[1], (float)f[1], (float)f[2], (float)f[3]};
}

With -O3, GCC outputs this:

f(int __vector(4)):
  pshufd xmm0, xmm0, 229
  cvtdq2ps xmm0, xmm0
  ret

LLVM outputs this:

f(int __vector(4)):
  pshufd xmm1, xmm0, 85 # xmm1 = xmm0[1,1,1,1]
  cvtdq2ps xmm1, xmm1
  pshufd xmm2, xmm0, 238 # xmm2 = xmm0[2,3,2,3]
  cvtdq2ps xmm2, xmm2
  pshufd xmm0, xmm0, 255 # xmm0 = xmm0[3,3,3,3]
  cvtdq2ps xmm0, xmm0
  unpcklps xmm2, xmm0 # xmm2 = xmm2[0],xmm0[0],xmm2[1],xmm0[1]
  shufps xmm1, xmm2, 64 # xmm1 = xmm1[0,0],xmm2[0,1]
  movaps xmm0, xmm1
  ret

[RKSimon]

Current Codegen: https://godbolt.org/z/oYn1df

Probably an SLP issue? [Bug #​35732] looks very similar

[anton-afanasyev]

This issue is to be fixed by http://reviews.llvm.org/D57059 after committing (in process of review now). We have conversion for f[1], f[2] and f[3] here, for three fps, so I have checked patch "Initial support for the vectorization of the non-power-of-2 vectors" fits here:

./opt -slp-vectorizer -instcombine -S pr49081.ll
...
define dso_local <4 x float> @​foo(<4 x i32> %0) {
%shuffle = shufflevector <4 x i32> %0, <4 x i32> poison, <4 x i32> <i32 1, i32 2, i32 3, i32 undef>
%2 = sitofp <4 x i32> %shuffle to <4 x float>
%3 = shufflevector <4 x float> %2, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 1, i32 2>
ret <4 x float> %3
}

[anton-afanasyev]

Commited llvm/test/Transforms/SLPVectorizer/X86/pr49081.ll to track fixing.

[rotateright]

This issue is to be fixed by http://reviews.llvm.org/D57059 after committing
(in process of review now). We have conversion for f[1], f[2] and f[3] here,
for three fps, so I have checked patch "Initial support for the
vectorization of the non-power-of-2 vectors" fits here:

./opt -slp-vectorizer -instcombine -S pr49081.ll
...
define dso_local <4 x float> @​foo(<4 x i32> %0) {
%shuffle = shufflevector <4 x i32> %0, <4 x i32> poison, <4 x i32> <i32 1,
i32 2, i32 3, i32 undef>
%2 = sitofp <4 x i32> %shuffle to <4 x float>
%3 = shufflevector <4 x float> %2, <4 x float> undef, <4 x i32> <i32 0,
i32 0, i32 1, i32 2>
ret <4 x float> %3
}

For reference, that sequence was not optimizing in IR or backend, so added an instcombine transform to make it easier for SDAG:
https://reviews.llvm.org/rG0bab0f616119

[RKSimon]

resolving - this was fixed by #65476