add test for codegen of SIMD vector from array repeat

[folkertdev]

fixes #97804

It appears that this issue was fixed silently in LLVM 19. The original codegen was terrible, but starting at LLVM 19 opt is able to generate good code.

https://llvm.godbolt.org/z/5vq8scP6q

cc @programmerjake

[rustbot]

r? @Mark-Simulacrum

rustbot has assigned @Mark-Simulacrum.
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

[folkertdev]

Hmm, it is fixed on x86_64 and aarch64, but not for other targets where vector types require a target feature.

• for s390x it only optimizes when the vector target feature is available
• for wasm32 even with +simd128 it does not optimize (should just be a splat I think?)

Do we consider that fixed?

[folkertdev]

I did file llvm/llvm-project#178940 for the wasm issue, cc @nikic

[workingjubilee]

For s390x without +vector, it probably comes from type legalization: LLVM, if it can't fit a type or operation with that type to a series of machine operations, will, during... lowering to SelectionDAG, I think?... "scalarize" the operations into operations on a valid type.

[folkertdev]

I'm a bit out of my depth here, but I would expect the transformation to a vector splat to happen much earlier than that. You're right that it does appear use information about what types are legal for the current target though.

[workingjubilee]

I think this is fine to merge if the test gets qualified with the appropriate architectures or whatever.

[Mark-Simulacrum]

@rustbot author

Yeah, with appropriate ignore-X added seems OK to approve, though we should probably leave the issue open in that case?

[rustbot]

Reminder, once the PR becomes ready for a review, use @rustbot ready.

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

[folkertdev]

I personally think the issue can be closed because we have and use simd_splat now instead of an array repeat expression. So the missed optimization caused worse code in practice before, while now for wasm that should not be the case unless you manually experiment with repr(simd) and don't use simd_splat.

@rustbot ready