[llvm] [ValueTracking] Don't use CondContext in dataflow analysis of phi nodes (PR #100316)

[via llvm-commits]

goldsteinn wrote:

Okay, took me a second, do I understand correctly that the issue is that since `%mask` is in-out its only the output that has the `%cmp2=false` context?

If thats correct, I think this works to fix it, although I wonder if we can do this a bit more precisely? I don't think this could be an issue outside of the context of a loop and it would be nice be able to keep this context for say analyzing the result of a switch/diamond branch.

https://github.com/llvm/llvm-project/pull/100316