JIT: stopping preference for ML CSE heuristic#98063
Merged
AndyAyersMS merged 3 commits intodotnet:mainfrom Feb 8, 2024
Merged
Conversation
Some initial attempts to improve the modelling of "stopping early" when doing CSEs. This adds a simplistic register pressure estimate modelled on the one we have now, where we find the weight of the Nth tracked local and use that as a reference weight for deciding when a CSE might start becoming costly. Contributes to dotnet#92915.
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsSome initial attempts to improve the modelling of "stopping early" when doing CSEs. This adds a simplistic register pressure estimate modelled on the one we have now, where we find the weight of the Nth tracked local and use that as a reference weight for deciding when a CSE might start becoming costly. Contributes to #92915.
|
Member
Author
|
@EgorBo PTAL Not as effective as I'd hoped. I think I may actually need multiple "stopping" choices to represent different strategies for stopping early and to let ML tune them (analogous perhaps to the current "aggressive, moderate, conservative" thresholds), but that will take more work as there is a strong bias in the current code that there's just one way to stop. |
Diff results for #98063Throughput diffsThroughput diffs for linux/arm64 ran on windows/x64Overall (+0.01% to +0.04%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for linux/x64 ran on windows/x64Overall (+0.01% to +0.03%)
FullOpts (+0.02% to +0.03%)
Throughput diffs for osx/arm64 ran on windows/x64Overall (+0.01% to +0.03%)
FullOpts (+0.02% to +0.03%)
Throughput diffs for windows/arm64 ran on windows/x64Overall (+0.01% to +0.03%)
MinOpts (-0.00% to +0.01%)
FullOpts (+0.02% to +0.03%)
Throughput diffs for windows/x64 ran on windows/x64Overall (+0.01% to +0.03%)
FullOpts (+0.02% to +0.03%)
Details here Throughput diffs for windows/x86 ran on linux/x86Overall (+0.02% to +0.03%)
FullOpts (+0.02% to +0.03%)
Details here |
Diff results for #98063Throughput diffsThroughput diffs for linux/arm64 ran on linux/x64Overall (+0.02% to +0.04%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for linux/x64 ran on linux/x64Overall (+0.01% to +0.03%)
FullOpts (+0.02% to +0.03%)
Details here Throughput diffs for linux/arm ran on windows/x86Overall (+0.02% to +0.04%)
FullOpts (+0.02% to +0.04%)
Details here |
EgorBo
reviewed
Feb 7, 2024
|
|
||
| // Stopping "parameter" | ||
| // | ||
| m_registerPressure = CNT_CALLEE_TRASH + CNT_CALLEE_SAVED; |
Member
There was a problem hiding this comment.
does it need to take floating points into account here? (e.g. CNT_CALLEE_SAVED_FLOAT)
Member
Author
There was a problem hiding this comment.
Historically we haven't, and there are very few FP cses (maybe 1% of all candidates).
EgorBo
approved these changes
Feb 7, 2024
Diff results for #98063Throughput diffsThroughput diffs for linux/arm64 ran on linux/x64Overall (+0.02% to +0.04%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for linux/x64 ran on linux/x64Overall (+0.01% to +0.03%)
FullOpts (+0.02% to +0.03%)
Throughput diffs for osx/arm64 ran on linux/x64Overall (+0.01% to +0.03%)
FullOpts (+0.02% to +0.03%)
Throughput diffs for windows/arm64 ran on linux/x64Overall (+0.01% to +0.04%)
MinOpts (-0.01% to +0.00%)
FullOpts (+0.02% to +0.04%)
Details here |
Diff results for #98063Throughput diffsThroughput diffs for linux/arm64 ran on windows/x64Overall (+0.01% to +0.04%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for windows/x64 ran on windows/x64Overall (+0.01% to +0.03%)
FullOpts (+0.02% to +0.03%)
Details here Throughput diffs for linux/arm ran on windows/x86Overall (+0.02% to +0.04%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for windows/x86 ran on windows/x86Overall (+0.02% to +0.03%)
FullOpts (+0.02% to +0.03%)
Details here |
8 tasks
This was referenced Feb 8, 2024
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Some initial attempts to improve the modelling of "stopping early" when doing CSEs. This adds a simplistic register pressure estimate modelled on the one we have now, where we find the weight of the Nth tracked local and use that as a reference weight for deciding when a CSE might start becoming costly.
Contributes to #92915.