[VFABI] sincos is no longer vectorizable on trunk

[Andarwinux]

https://godbolt.org/z/cKh4MY9f7

[MacDue]

On AArch64 we have tests to guard this: llvm/test/Transforms/LoopVectorize/AArch64/multiple-result-intrinsics.ll + llvm/test/CodeGen/AArch64/veclib-llvm.sincos.ll, and I think they are still functioning. It's probably worth adding a similar test for x86/AMDLIBM.

[Andarwinux]

On AArch64 we have tests to guard this: llvm/test/Transforms/LoopVectorize/AArch64/multiple-result-intrinsics.ll + llvm/test/CodeGen/AArch64/veclib-llvm.sincos.ll, and I think they are still functioning. It's probably worth adding a similar test for x86/AMDLIBM.

https://godbolt.org/z/ffsonMx5e

But it doesn't seem to work on aarch64 either.

[MacDue]

I see, it looks like it's now doing a generic expansion rather than lowering to vector libcalls.

[fhahn]

Ideally this kind of thing would be caught end-to-end, possibly via loop-vectorization micro benchmarks. It would also require setting up a job to run those automatically on platforms that have support in their math libraries

[MacDue]

The AArch64 side of this should be fixed by #167996, which adds a Clang test.