[llvm] [X86] Try to scale i8/i16 VPERMV3 to i32 (PR #97212)

[Phoebe Wang via llvm-commits]

phoebewang wrote:

> Do you have real world examples of this?

No. Just to address a comment from @goldsteinn in https://github.com/llvm/llvm-project/pull/96414#discussion_r1650296310

I didn't create an approximate example, and guess we may have scaled most cases already.

> I mention this as I'm intending to patch X86InstCombineIntrinsic.cpp so that VPERMV3 style intrinsics with constant masks get converted to regular shuffles like most other x86 shuffle intrinsics. I think this should address most cases unless you know it can occur in the backend.

That would be great! I'll close this.

https://github.com/llvm/llvm-project/pull/97212