[llvm] [ValueTracking] Use dominating conditions in computeConstantRange (PR #210412)

[Joel Walker via llvm-commits]

Joel-Wwalker wrote:

Thanks — verified: plain `-passes=sccp` folds all three positive tests, so those shapes are indeed within SCCP's existing coverage (and IPSCCP sees them before flattening at -O2/-O3).

One data point from the patch's A/B sweep: the four LoopVectorize/PhaseOrdering test changes (`sub nsw` -> `sub nuw nsw` on the vector trip count) come from InstCombine running after the vectorizer, where SCCP/CVP don't run again — so that particular gain isn't reachable from the SCCP placement. Whether it justifies the walk is exactly the compile-time question, so the numbers and your call on the design boundary decide it; no objection from me either way.

Agreed on the root cause — the flattened single-block form is what actually blocks #207131 at -O3, and this patch doesn't claim to address that.

https://github.com/llvm/llvm-project/pull/210412