[llvm] [DFAJumpThreading] Avoid exploring the paths that never come back (PR #85505)

[via llvm-commits]

XChy wrote:

> My general concern with not preserving LoopInfo is that we still continue using LoopInfo even after it has been invalidated. Let's say we have two candidates for DFA jump threading in a function. When we process the first one, we modify the CFG without (fully) updating LoopInfo. Then for the second one, we query LoopInfo -- but the LoopInfo is broken now. Can this cause any issues? I find it hard to tell...

Yes, if we process more than one candidate, that would be a problem (though there are already other underlying problems if we thread more). These problems should be solved in the future.

https://github.com/llvm/llvm-project/pull/85505