[llvm] [InstCombine] Fold mul (lshr exact (X, N)), 2^N + 1 -> add (X , lshr exact (X, N)) (PR #95042)

[via llvm-commits]

AtariDreams wrote:

> The code doesn't match the proofs (the proofs have freeze instructions, the code does not).

Is it okay to use freeze even if that counts as an extra instruction in the IR? Because that's why did a last second replacement with a check for NUW/NSW for the optimization to apply at all.

https://github.com/llvm/llvm-project/pull/95042