Implement arithmetic operation traits for x86 SIMD types

[sayantn]

ACP: rust-lang/libs-team#628
Tracking Issue: rust-lang/rust#145066

Unresolved questions: Should we have Not and Rem implementations? At least Not seems quite useful.

I will send the PRs of other archs after this is merged, as they also use the common macro

[rustbot]

r? @folkertdev

rustbot has assigned @folkertdev.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[sayantn]

CI is stuck on #1897 🤦🏽

[folkertdev]

yeah that will likely take a couple of days at least

[bjorn3]

These are missing a #[target_feature(enable = "...")] matching the _mm* caller, so the MIR inliner will not inline them and LLVM may not inline them either.

[sayantn]

simd_* intrinsics call the platform-independent LLVM intrinsics/instructions. The conversion to ASM checks the available feature flags, and uses them as available. See rust-lang/libs-team#628 (comment)

[folkertdev]

Do you want to decide on those unresolved questions here, or later, or bounce it back to T-libs?

• simd_rem is not really used (I find 2 occurrences), suggesting that there is no instruction support for this operation and we should probably steer users away from it
• we don't have a simd_not, but simd_xor(v, splat(all_one_bits)) should do something efficient

[folkertdev]

Also @Amanieu merging this PR insta-stabilizes these instances, so just checking: that is intentional, right?

I feel like in that case we should at least implement the same operations for all stable simd types in the same rust release, i.e. at least make sure the aarch64 ones make it into the same release as the x86 ones.