[llvm] [InstCombine] Support gep nuw in icmp folds (PR #118472)

[Hans Wennborg via llvm-commits]

zmodem wrote:

I see @bgra8 already raised this, but maybe it's worth discussing further.

There is a fair amount of test fallout from this (the profile merging issue in LLVM mentioned above, at least one issue in Chromium, and many more internally). We can fix those, but what concerns me are the issues not caught by tests. It doesn't seem unlikely that this optimization could open up security holes by removing these kinds of bounds checks.

While such code was already technically broken, it would have been hard for developers to notice. Could -Wtautological-compare catch some of these (like the `Check` example above)? Could UBSan flag pointer arithmetic overflows? Could ASan flag comparisons of invalid pointers?

https://github.com/llvm/llvm-project/pull/118472