[llvm] [ValueTracking] Support dominating known bits condition in and/or (PR #74728)

[Nikita Popov via llvm-commits]

nikic wrote:

@dtcxzyw It's been a while, but IIRC I addressed most of the "interesting" regressions. The main one left is some sub-optimality in IV widening (we don't make good use of the fact that top bits are zero for the widened LCSSA phi node), which would ideally be solved by introducing trunc nuw/nsw, though I guess we could also add special support for it during widening.

https://github.com/llvm/llvm-project/pull/74728