[llvm] [AArch64][GlobalISel] Take abs scalar codegen closer to SDAG (PR #84886)

[Madhur Amilkanthwar via llvm-commits]

madhur13490 wrote:

> Do you have an mca proof that smax is better than xor?

I understand the question but my intention of the patch is to take the codegen closer to what SDAG is doing. However, if you see the i8 case, it is not quite close yet and I am also not sure if the final assembly is right. Ideally, it should just be cmp+cneg (which is what SDAG) is doing.

If we get it closer to SDAG's codegen, then the benefit would be visible as Global ISel's current lowering leads to 3 instructions whereas SDAG is using just cmp+cneg.  
I am looking for suggestions around this.

https://github.com/llvm/llvm-project/pull/84886