[llvm] [InstCombine] Make backedge check in op of phi transform more precise (PR #106075)

[Nikita Popov via llvm-commits]

nikic wrote:

@mikaelholmen I've reduced this to `opt -S -passes="function(constraint-elimination,loop-mssa(licm,simple-loop-unswitch))" -verify-scev` with https://gist.github.com/nikic/ab65b31e3f95ff66385e27498c8edf60, so this is a pre-existing issue (either in SCEV or one of the loop optimizations).

https://github.com/llvm/llvm-project/pull/106075