[PATCH] D147456: [AutoFDO] Stale profile matching(part 1)

[Lei Wang via Phabricator via llvm-commits]

wlei added a comment.

In D147456#4247167 <https://reviews.llvm.org/D147456#4247167>, @mingmingl wrote:

> In D147456#4247075 <https://reviews.llvm.org/D147456#4247075>, @wlei wrote:
>
>> Thank you for all the feedback, I will try to address it later. I saw you all are curious the algorithm part, here it is: https://reviews.llvm.org/D147545.
>>
>>> Just browsed the patch, I don't seem to find how FUncToMatchingsMap is populated which is the core of the stale profile matching algorithm.  Does algorithm use some anchor points to do the remapping?
>>
>> Yes, here is the algorithm part : https://reviews.llvm.org/D147545, it leverages the callee name/ calltarget name as anchor which doesn't need any other additional info.
>
> Just curious, I wonder if there //happens// to be a measurement of the percentage of samples recovered (that would have been dropped) for csspgo. I guess I'm trying see how much headroom remains with D147545 <https://reviews.llvm.org/D147545> :)

Good point, so far no, it's doable, we have the https://reviews.llvm.org/D136627 to compute how many samples are dropped due to the CSSPGO's checksum mismatch, which is what this stale profile matching want to reuse. Similarly, after the matching in D147545 <https://reviews.llvm.org/D147545>, we can compute how many samples/callsites in profile still remains mismatched.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D147456/new/

https://reviews.llvm.org/D147456