[PATCH] D40873: [PGO] handle infinite loop properly in pgo instrumention

[Vedant Kumar via Phabricator via llvm-commits]

vsk added a comment.

I don't have my head wrapped around the cause of the leak yet. Was the issue that getAnalysis<LoopInfo...>() recomputed BFI, thereby leaking the existing analysis? If so, would it be sensible for the pass manager to provide a cached BFI instead of allowing the leak? (I'm not suggesting changing the PM, just curious about how these things fit together.)

At a more basic level: Say CFGMST depends on BFI, which in turn depends on {BPI, LI}. Is it possible for CFGMST to be scheduled after a competing pass "DestroysBPI", which preserves BFI but invalidates BPI? If so, how do we avoid that?


https://reviews.llvm.org/D40873