[llvm] [AArch64][GlobalISel] Expand abs.v4i8 to v4i16 and abs.v2s16 to v2s32 (PR #81231)

[Dhruv Chawla via llvm-commits]

dc03-work wrote:

> > Hello. I had seen the same Extract problems. There is a patch for improving Inserts in #81453, and I was looking to do something similar for Extracts too once that goes in, but it may need to fix ANYEXT too. @chuongg3 is looking into stores (and maybe loads) of v4i8 types at the moment too.
> > There is a chance that the extracts in your case might have come from #75662. It looks like it should just be using %29 = %10 though, and if it could learn to remove those intermediate instructions then the problem might go simplified away? I will keep taking a look - let me know if you come up with anything further. I'm not against this patch if we need it, but I currently (weakly) think that widening will be better in general than promoting if we can get there and be consistent about it.
> 
> This patch is a pretty safe bet for now, I think it would be best to follow Madhur's suggestion and implement the widening support in a follow-up patch. For now, this solution works well enough and I feel it should be landed because it matches SDAG and stops the fallback.

Ping! @davemgreen 

https://github.com/llvm/llvm-project/pull/81231