[llvm] [InstCombine] Fold icmp of gep chains (PR #146714)

[Yingwei Zheng via llvm-commits]

dtcxzyw wrote:

We hit a tricky regression when the code contains both pointer comparison and difference sharing the same operands.
See https://github.com/facebook/zstd/blob/1dbc2e09084e843f6c0dcc2d0791610015c50979/lib/decompress/zstd_decompress.c#L1526-L1528 and https://github.com/dtcxzyw/llvm-opt-benchmark/pull/2530/commits/1b6d543c62c13b43afedded56bd972fc228a0d96#r2183037608

It is hard to reassociate operands and reuse the offset...



https://github.com/llvm/llvm-project/pull/146714