[PATCH] D39346: [LV] [ScalarEvolution] Fix PR34965 - Cache pointer stride information before LV code gen

[Diego Caballero via Phabricator via llvm-commits]

dcaballe added a comment.

Thanks for the feedback.

Before posting this review, I got some early feedback regarding fixing/extending SCEV for this case and it didn't look good. Maybe, we could special-case it but a generic solution, if I understood correctly, could require representing non-inductive phis in SCEV. This is not my area of expertise so, hopefully, we'll get more feedback about it. I'll ping llvm-dev if this doesn't happen in the short term.

In any case, I agree with Hideki and Hal. We should cache any analysis information before we start transforming the code. We shouldn't expect the analysis information from the original loop to be preserved after transforming the code. VPlan will address this issue in a better way but, in the meantime, this should be a reasonable fix.

Thanks,
Diego


https://reviews.llvm.org/D39346