[PATCH] D113939: [runtime-unroll] Remove restriction about unrolling multiple exit loops

[Philip Reames via Phabricator via llvm-commits]

reames added a comment.

Just so folks know what to expect here, let me briefly summarize.

At this point, we have enough cases reported that regress than in order to drive this forward, I'd likely need to do some cost modeling changes.  Heuristic changes are famously delicate and expensive time wise.  Its really hard for me to justify that time as things stand.

I'm currently waiting on a couple of examples from my client where this causes regressions on internal workloads.  I'm going to take a look at them to see if anything glaring falls out, but we're expecting the answer to simply be that we unrolled an unprofitable case.  Given these couple of regressions seem to be exceptions, I'm probably going to end up recommending my client enable this downstream and close the upstream review.

In the meantime, if anyone wants to send me IR for an "interesting" regression, I'm happy to take a look at other sources to see if ideas fall out.  I don't plan to proactively investigate test-suite results just because I've found in the past that the investigation cost is high, and the expected value in this case is quite low.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D113939/new/

https://reviews.llvm.org/D113939