ACP: Add unchecked arithmetic methods for integer simd

[CrazyboyQCD]

Proposal

Add unchecked arithmetic methods for integer simd

Problem statement

uncheked_(add, sub, mul, div, shl, shr, neg, exact_div) methods are added for integers but not for integer simd.

Motivating examples or use cases

These methods will be helpful in performant code.

Solution sketch

1. Add unchecked intrinsics
2. Add inherent unchecked methods

Alternatives

Links and related work

rust-lang/portable-simd#14
rust-lang/portable-simd#481

[kennytm]

"integer simd" means e.g. the std::simd::Simd<i32; 4> type rather than e.g. std::arch::x86::__m128i right?

[CrazyboyQCD]

Yes.

[programmerjake]

note we're missing unchecked_rem (it's listed in the intrinsics, but not yet exposed as an integer method).

this requires adding simd_*_unchecked intrinsics to the compiler, and then creating a PR in portable-simd to add those methods.

That said, other than div/rem, the existing Simd<T, N> operations are generally equivalent to int::wrapping_op so are already about as efficient as you can get

[pitaj]

This needs an actual motivating example.

[CrazyboyQCD]

This needs an actual motivating example.

Here is a code snippet rewrote from https://github.com/dmtrKovalenko/odiff/blob/c6018838899469195c64782db431ddea3ca61b8a/src/color_delta.zig#L106C1-L153C2

Rust version:
https://godbolt.org/z/jWv1oY34G

Zig version:
https://godbolt.org/z/G4f7jnzWP

IR comparision(source 1 is the rust version, source 2 is the zig version, source 3 is the rust version with no-overflowing signs manually added):
https://godbolt.org/z/ExMWEM8re

Thought there are gaps after adding the signs due to usages of instruction , no-overflowing signs does produces more verctorization instructions.

[dtolnay]

We discussed this ACP in today's standard library API team meeting. From our end we would be happy adding this functionality on core::simd::Simd. @calebzulawski could you take a look as well and confirm that adding all these would make sense? Or just a subset of them?

[calebzulawski]

This seems reasonable to me, as @programmerjake said I think the main effort is getting the intrinsics into the compiler. Note that we don't currently have a design for the matching checked functions, because we can't practically return None.

I probably won't have time to implement this right now and I don't think it's critical to implement before stabilizing std::simd but I'd be happy to review PRs.

Edit: I think we might as well implement all of them to be consistent. It's not clear to me what e.g. add nsw accomplishes for SIMD but if we can encode the extra metadata we probably should.