[llvm] [AArch64] Check for negative numbers when adjusting icmps (PR #141151)

[via llvm-commits]

AZero13 wrote:

There are SOME regressions here, but this isn't the patch's fault.

Overall, the improvements outweigh the regressions

Or rather, the cause of the regressions is not something the original code was ever doing for the sake of optimization, but rather luck.

You see, the change of CMN 1 to cmp was in some places allowing better CSE to happen, but with this patch, that's undone, but this also means cse is happening in more places.

The fix is not to reverse this patch, but it is exposing a huge problem with the design of this code: we should be taking CSE into account too.

But how? Espcially when CSE is probably not looking at this from the DAG level and is a different pass? 

https://github.com/llvm/llvm-project/pull/141151