use simd_saturating_{add, sub} on neon

[folkertdev]

The idea behind this is that e.g. miri and cranelift don't know all of the llvm intrinsics (and adding them is possible but a lot of extra work). Using the generic simd function means logic only needs to be implemented once for all targets.

The simd_saturating_* functions generate the same instructions for the saturating add/sub operations https://godbolt.org/z/zc35doTc6

[rustbot]

Thanks for the pull request, and welcome! The Rust team is excited to review your changes, and you should hear from @Amanieu (or someone else) some time within the next two weeks.

[folkertdev]

I don't think this idea works because of this LLVM issue I opened llvm/llvm-project#94463. So unless we can get that fixed (or somehow accept much worse codegen) this PR does not work

[Amanieu]

OK let's defer this for now until the LLVM issue is sorted.

[folkertdev]

@Jamesbarford given that you work at ARM, is there any chance you could pass this on to the right person (specifically the LLVM issue llvm/llvm-project#94463). Not only does it clean up stdarch a bit, but it's especially valuable for std::simd.

[Jamesbarford]

Sure, I've mentioned it 👍

[sayantn]

We can still use simd_saturating_add for arm, right? There are actually quite a few such occurrences in ARM (e.g. using llvm.fma instead of just using simd_fma). Also there are a few which use wrong LLVM intrinsics, I will send a PR to correct those soon

[folkertdev]

I actually fixed llvm/llvm-project#94463 in llvm/llvm-project#140454. I just re-ran CI here to check if there are any other issues, and there were, resulting in llvm/llvm-project#142323 and a fix in llvm/llvm-project#142342.

So, things are looking good for being able to use these operations for aarch64 with the next llvm release.

For arm: we can use the platform-independent versions if they don't cause regressions. If that's the case (looks like it for these operations at least) then yes, we can (and should!) use the simd_ intrinsics.

[sayantn]

The ARM intrinsics are actually using the same LLVM intrinsics that simd_saturating_{add,sub} use, so it shouldn't cause any regressions