Suboptimal code gen in Clang for _mm_setzero_ps with #pragma STDC FENV_ACCESS ON on x64

[pavlejosipovic]

_mm_setzero_ps() is supposed to generate xorps xmm0, xmm0 but if STDC FENV_ACCESS ON is set (it get set by -ffp-model=strict)
than it generates
xor eax, eax cvtsi2ss xmm0, eax shufps xmm0, xmm0, 0

This happens only on Clang, with GCC or MSVC this doesn't happen.
Below are code snippets for STDC FENV_ACCESS OFF/ON
https://godbolt.org/z/nqx9ejKre
https://godbolt.org/z/7hKzh95E4

[topperc]

I can make dumb fix for this in the headers. Probably need some better constant folding somewhere to do this generically.

cc @kpneal

[llvmbot]

@llvm/issue-subscribers-backend-x86

[topperc]

Hopefully fixed after c9320bc

[pavlejosipovic]

@topperc would it help if make a separate issue to complain about constant folding? I've noticed issue in the same environment if somebody does something like this float(0) instead of writing 0.0f