[llvm] [JumpThreading] Refractor redirectValuesFromPredecessorsToPhi: avoid unnecessary map resizing in gatherIncomingValuesToPhi (PR #173596)

[via llvm-commits]

int-zjt wrote:

Just for context: I tried implementing a fast path for certain patterns to skip the map population initially. I dropped it because it didn't fundamentally solve the resizing overhead for other cases. Also, handling the merging of `undef` and `poison` values for the same block correctly in that fast path turned out to be tricky without making the code much more complicated.

https://github.com/llvm/llvm-project/pull/173596