[llvm] [GISel][AArch64][AMDGPU][RISCV] Canonicalize (sub X, C) -> (add X, -C) (PR #114309)

[Craig Topper via llvm-commits]

topperc wrote:

> Hi @topperc We have CI internally which tracks number of fallbacks in SPEC 2017 benchmark and we see a massive uptick in the number of fallbacks in `gcc_r` benchmark between 4th Nov and 5th Nov. We compile SPEC 2017 for AArch64 with and without SVE. In NOSVE case, the number of fallbacks went from 8 to 28.
> 
> Unfortunately, our CI doesn't have the capability to bisect to a change but I took a quick look on `git log` and I see this is the only relevant patch which could have caused the uptick.
> 
> Is this expected? If you have access to SPEC, can you please try compiling before and after your patch by passing "+nosve"?

I built clang and spec2017 on my M2 Macbook. I'm seeing 32 fallbacks on 502.gcc_r before and after my change. I rolled back to the last patch from Nov 2 and I see 32 there too.

https://github.com/llvm/llvm-project/pull/114309