SSA: Improve performance of finding relevant phi input nodes.

[aschackmull]

Figuring out whether a phi-input node is necessary involves checking whether it's guard-equivalent to the prior reference. The code used to do this naively by simply checking whether there's a guard that controls one but not the other. This led to some extremely poor performance in large repos (30+ mins, billions of tuples, non-termination). Instead we can walk the dominator tree to see if we can get from the one to the other without passing a branch edge of a guard. That's equivalent and avoids the join with valueControls entirely. I've checked one of the non-terminating cases, and there the predicate can now be evaluated in mere seconds.

[Copilot]

Pull request overview

This PR improves the performance of finding relevant phi input nodes in SSA analysis by replacing a naive guard-equivalence check with a more efficient dominator tree walk approach. The change addresses severe performance issues in large repositories where the previous implementation could take 30+ minutes or fail to terminate.

Key changes:

• Introduced a new dominator tree-based algorithm to determine guard-equivalence between phi inputs and their predecessors
• Added helper predicates to support the new implementation
• Replaced the previous join-heavy approach that used valueControls with a recursive walk up the dominator tree

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[hvitved]

LGTM, thanks for addressing. I have started DCA runs.

[aschackmull]

There was a performance problem in some huge pieces of JS code where the reachability recursion got very large. Switching to fastTC fixed the problem.