[llvm] [InstCombine] Canonicalize xor with disjoint ops to or disjoint (PR #133139)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> I don't think this transformation is profitable. Generally cmov is cheaper than mul. InstCombiner performs similar transformations in foldSelectICmpAndBinOp for add/xor/shl/lshr/ashr/..., but not for mul.

It's fewer IR instructions, so it's better canonical form. Undoing for profitability is a backend issue 


https://github.com/llvm/llvm-project/pull/133139