JIT: use blend rather then repair for profile inconsistencies #85171
Conversation
If we have a partial profile then the current count reconstruction will adjust the exit likelihood of some loop exit when it hits a capped loop. But for multiple exit loops we might wish to see some profile flow out of all exits, not just one. In `ludcmp` we choose to send all the profile weights down an early return path, leaving the bulk of the method with zero counts. Instead of trying increasingly elaborate repair schemes, we will now use blend mode for these sorts of problems; this gives a more balanced count redistribution. I also updated blend to use the same logic as repair if a block has zero weights, since presumably whatever likelihood was assigned there during reconstruction is not well supported. Fixes the `ludcmp` regression with PGO over no PGO, noted in dotnet#84264 (comment)
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsIf we have a partial profile then the current count reconstruction will adjust the exit likelihood of some loop exit when it hits a capped loop. But for multiple exit loops we might wish to see some profile flow out of all exits, not just one. In Instead of trying increasingly elaborate repair schemes, we will now use blend mode for these sorts of problems; this gives a more balanced count redistribution. I also updated blend to use the same logic as repair if a block has zero weights, since presumably whatever likelihood was assigned there during reconstruction is not well supported. Fixes the
|
|
@EgorBo PTAL |
|
/azp run runtime-coreclr pgo, runtime-coreclr libraries-pgo |
|
Azure Pipelines successfully started running 2 pipeline(s). |
|
Diffs are fairly surgical: |
Yes. It is mainly just this one method where the old repair process left most of the method cold and so we didn't clone loops like we need to for good perf. |
|
Failure is known issue, |
ghost
locked as resolved and limited conversation to collaborators
If we have a partial profile then the current count reconstruction will adjust the exit likelihood of some loop exit when it hits a capped loop. But for multiple exit loops we might wish to see some profile flow out of all exits, not just one.
In
ludcmpwe choose to send all the profile weights down an early return path, leaving the bulk of the method with zero counts.Instead of trying increasingly elaborate repair schemes, we will now use blend mode for these sorts of problems; this gives a more balanced count redistribution.
I also updated blend to use the same logic as repair if a block has zero weights, since presumably whatever likelihood was assigned there during reconstruction is not well supported.
Fixes the
ludcmpregression with PGO over no PGO, noted in #84264 (comment)