Avx512

[minybot]

fmadd: ps,pd
fnmadd: ps,pd
fmsub: ps,pd
fnmsub: ps,pd
fmaddsub: ps,pd
fmsubadd: ps,pd

add_round: ps,pd
sub_round: ps,pd
mul_round: ps,pd
div_round: ps,pd
sqrt_round: ps,pd

cvt_roundps_epi32
cvt_roundps_epu32

[rust-highfive]

r? @Amanieu

(rust_highfive has picked a reviewer for you, use r? to override)

[minybot]

_mm512_fmsub_ps can be vfmadd132ps or vfmadd213ps or vfmadd231ps.
In x86_64-unknown-linux-gnu, it uses vfmadd213ps.
In x86_64-pc-windows-msvc, it uses vfmadd231ps.
Does it support something like assert_instr(vfmadd213ps|vfmadd231ps)?

[Amanieu]

assert_instr can match a prefix, so you can use assert_instr(vfmadd). Please add a comment next to it to explain that the compiler can generate different instructions for this intrinsic.

[Amanieu]

For rounding modes can you do the same thing as the constify_imm5_sae macro: we want to panic in case of an invalid rounding mode since that could crash the compiler if we let it through to the LLVM intrinsic.

[jrmuizel]

It looks like this broke the check-style tests. It wasn't caught before merge because check-style passes without checking anything if there's no rustfmt available for the particular nightly.