[llvm] [InstCombine] Fold -X / -Y -> X / Y (PR #88422)

[Yingwei Zheng via llvm-commits]

dtcxzyw wrote:

> > I will accept this patch when we meet this pattern in some real-world applications.
> 
> If we see it in an actual application, it means that some customer has encountered an optimization opportunity which we missed. Handling potential issues before they happen is a reasonable thing to to.

I disagree. Adding unused folds makes InstCombine slow and buggy. It is also a maintenance burden for our LLVM developers and researchers who develop fuzzers and formal-method based tools.

> What concerns do you have with this patch as-is?

I am not willing to pay for the expensive call to `computeKnownBits(Op1, 0, &I)`.


https://github.com/llvm/llvm-project/pull/88422