[PATCH] D81981: [PGO] Supplement PGO profile with Sample profile

[Wei Mi via Phabricator via llvm-commits]

wmi added a comment.

In D81981#2099702 <https://reviews.llvm.org/D81981#2099702>, @wenlei wrote:

> If how we select BB to instrument depends on a switch, we would need instrumentation build and optimizing build to have consistent switch, otherwise counters could mismatch even if CFG checksum matches? I guess that's one reason why it'd be good to avoid different ways of selecting BBs.


PGO has already had a common practice to ensure profile-gen and profile-use to have the same flags. So the flag to enable inserting counter in entry block won't cause too much trouble.

> Would it be possible to tweak/cheat the edge weights just for MST so entry BB is pinned to be non-MST node hence guaranteed to be instrumented directly?

I consider that but I don't know how to do that. From what I currently understand, MST is to select some edges with highest frequencies and pruning those edges won't affect the inference of the profile of all the edges. We can prune edge when selecting MST but we cannot guarantee a node is selected as an instrumented BB during that phase. Deciding which BB to instrument is done in getInstrBB -- by choosing whether to instrument src node or dst node for each edge.

I am new to this part so if you know there is way to do that, please let me know. That is very appreciated.


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D81981/new/

https://reviews.llvm.org/D81981