[llvm] [InstCombine][CVP] Fold length-one memset with variable fill (PR #213240)

[via llvm-commits]

hsnaveen2u wrote:

Thanks for the suggestion. I reworked the change to remove the CVP/LVI dependency.

The series is now split as follows:-
1. InstCombine folds a length-one memset with a variable fill value into an i8 store.
2. AggressiveInstCombine handles nonconstant lengths known to be in [0, 1].
It uses computeKnownBits() through a SimplifyQuery containing the assumption cache, dominator tree and DomConditionCache so that both the directly masked case and the dominating-condition case are recognized.  

https://github.com/llvm/llvm-project/pull/213240