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

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

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
}

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

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