Use SIMD intrinsics for vector shifts

[sayantn]

Retrying #1928, this time actually checking for overflow. LLVM sees through this, and eliminates the branch/selects https://godbolt.org/z/WsYvfjPas

cc @RalfJung this also removes uses of a few intrinsics from avx2

[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

[eduardosm]

simd_shl and simd_shr are documented to be UB on overflow, but the implementations of this PR will unconditionally call these intrinsics.

[sayantn]

Ah, so I need to add a simd_and to make it safe. Thanks for pointing it out

[sayantn]

It looks like adding that one simd_and messes up the whole LLVM optimization. Strictly speaking, this is fine as long as we are only in LLVM backend, LLVM generates poison values if the shift value overflows and we are filtering them out in the next step. But obviously that's not a good argument, as the main motive for this change is to make it more portable across backends. I will look into what needs to be done

edit: https://godbolt.org/z/99zrnb5Gj

edit2: somehow using simd_select(good, count, T::splat(dummy_value)) works. It lets LLVM collapse the 2 selects, while also keeping the simd_shl call safe from Rust (https://godbolt.org/z/cc4EKfeb7)

[sayantn]

https://godbolt.org/z/cafKWsf6x I think these should do the job, @eduardosm could you check please?

edit: seems to pass miri on overflow playground

[eduardosm]

https://godbolt.org/z/cafKWsf6x I think these should do the job, @eduardosm could you check please?

That looks good to me 👍🏻