[llvm] [LV] Add vplan folds for (X - (X & Y)) -> (X & ~Y) (PR #226466)

[David Sherwood via llvm-commits]

david-arm wrote:

Hi @lukel97, that change might be useful by itself regardless of this PR? I think what I'm scared of here is playing a game of whack-a-mole, which I think is inevitable without duplicating the entire behaviour of instcombine (along with all the associated analysis) in vplan. I personally believe we're probably going to have to accept some limitations and carefully assess the risk/reward with such transformations. In a way, we've made life a lot more difficult for ourselves by implementing a new form of IR (vplan) that nothing else in LLVM understands, severely restricting our ability to make use of existing code.

https://github.com/llvm/llvm-project/pull/226466