Format f16 values correctly in intrinsic-test

[sayantn]

Fix some x86 intrinsics too

• _mm_{u}comineq_sh (these should be unordered, i.e. should return true if either operand is nan)
• _mm_mask_cvt{epi16_epi8, pd_ps, pd_epi32} (top 64 bits should be 0)

[sayantn]

r? @folkertdev

Just lmk before merging, will need to remove the last commit

[rustbot]

Failed to set assignee to folk: invalid assignee

Note: Only org members with at least the repository "read" role, users with write permissions, or people who have commented on the PR may be assigned.

[folkertdev]

cc @tgross35 was there some caveat to this approach?

[tgross35]

Not sure about the rest but just regarding formatting, these were added at a time that we didn't have Debug for f16

[folkertdev]

Right just double-checking that there is not some edge case where we'd get incorrect outputs by upcasting to f32.

[folkertdev]

This should be good to go then with the most recent commit removed

[folkertdev]

Do we document what clang bug(s) these are?

[sayantn]

yeah I was thinking about putting some clang PRs to correct these. in summary,

• _mm{256}_extract_epi{8,16} does (int) (unsigned {char,short}), but afaik C does sign-extension whenever the target type is signed, so this is getting sign-extended, whereas Intel specifies that it should zero-extend.
• _mm512_mask_reduce_{max,min}_{ps,pd} uses {f32,f64}::{MIN,MAX} in place of masked-out values, but Intel specifies it should use {f32,f64}::{NEG_INFINITY,INFINITY}