JIT: Set naive likelihoods when initializing preds #98192
Merged
Conversation
This file contains 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
dotnet-issue-labeler
bot
added
the
area-CodeGen-coreclr
ghost
assigned 
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsPart of #93020. Per discussion on #98054, set "naive" edge likelihoods (in other words, assume every successor edge is equally likely to be taken) in cc @dotnet/jit-contrib
|
amanasifkhalid
commented
Feb 8, 2024
src/coreclr/jit/fgflow.cpp
Outdated
|
|
||
| const unsigned numSucc = blockPred->NumSucc(); | ||
| assert(numSucc > 0); | ||
| flow->setLikelihood(1.0 / numSucc); |
There was a problem hiding this comment.
@AndyAyersMS you mentioned the likelihood calculation should consider the edge's duplicate count. I ran into issues with BBJ_COND blocks with true/false branches to the same target, as NumSucc handles this case by returning 1 for such blocks. In such cases, if we calculate the likelihood as (1.0 / numSucc) / dupCount, the edge likelihood overflows to (1.0 / 1) * 2 = 2.0. So leaving the dup count out of the calculation seems to work for that case, but I think this will give us wonky likelihoods for BBJ_SWITCH blocks with duplicate successors.
Is it possible to encounter BBJ_SWITCH blocks with duplicate successors this early on in the JIT? If so, we can use BasicBlock::NumSucc(Compiler *) to get the number of unique switch successors to get the edge likelihoods right, though it'll probably cost a bit more TP.
There was a problem hiding this comment.
For switches having dups is pretty common.
Maybe you just need to special-case BBJ_COND and check for the odd cases where both targets are the same block.
Diff results for #98192Throughput diffsThroughput diffs for windows/arm64 ran on linux/x64Overall (+0.03% to +0.05%)
MinOpts (+0.03% to +0.11%)
FullOpts (+0.03% to +0.05%)
Details here |
Diff results for #98192Assembly diffsAssembly diffs for linux/arm64 ran on windows/x64Diffs are based on 1,610,272 contexts (368,644 MinOpts, 1,241,628 FullOpts). MISSED contexts: 3,428 (0.21%) Overall (-16 bytes)
FullOpts (-16 bytes)
Assembly diffs for linux/x64 ran on windows/x64Diffs are based on 1,620,764 contexts (360,162 MinOpts, 1,260,602 FullOpts). MISSED contexts: 3,086 (0.19%) Overall (-15 bytes)
FullOpts (-15 bytes)
Assembly diffs for osx/arm64 ran on windows/x64Diffs are based on 1,733,061 contexts (561,303 MinOpts, 1,171,758 FullOpts). MISSED contexts: 3,460 (0.20%) Overall (-16 bytes)
FullOpts (-16 bytes)
Assembly diffs for windows/arm64 ran on windows/x64Diffs are based on 1,477,297 contexts (263,527 MinOpts, 1,213,770 FullOpts). MISSED contexts: 3,464 (0.23%) Overall (-16 bytes)
FullOpts (-16 bytes)
Assembly diffs for windows/x64 ran on windows/x64Diffs are based on 1,999,231 contexts (587,594 MinOpts, 1,411,637 FullOpts). MISSED contexts: 3,657 (0.18%) Overall (-2 bytes)
FullOpts (-2 bytes)
Details here Assembly diffs for linux/arm ran on windows/x86Diffs are based on 1,449,330 contexts (345,734 MinOpts, 1,103,596 FullOpts). MISSED contexts: 55,656 (3.70%) Overall (+0 bytes)
FullOpts (+0 bytes)
Assembly diffs for windows/x86 ran on windows/x86Diffs are based on 1,618,717 contexts (327,626 MinOpts, 1,291,091 FullOpts). MISSED contexts: base: 11,019 (0.68%), diff: 11,022 (0.68%) Overall (+38 bytes)
FullOpts (+38 bytes)
Details here Throughput diffsThroughput diffs for linux/arm64 ran on windows/x64Overall (+0.03% to +0.05%)
MinOpts (+0.03% to +0.11%)
FullOpts (+0.03% to +0.05%)
Throughput diffs for linux/x64 ran on windows/x64Overall (+0.03% to +0.05%)
MinOpts (+0.03% to +0.15%)
FullOpts (+0.03% to +0.06%)
Throughput diffs for osx/arm64 ran on windows/x64Overall (+0.03% to +0.05%)
MinOpts (+0.03% to +0.11%)
FullOpts (+0.03% to +0.05%)
Throughput diffs for windows/x64 ran on windows/x64Overall (+0.03% to +0.06%)
MinOpts (+0.03% to +0.15%)
FullOpts (+0.03% to +0.06%)
Details here Throughput diffs for linux/arm ran on windows/x86Overall (+0.03% to +0.04%)
MinOpts (+0.03% to +0.10%)
FullOpts (+0.03% to +0.04%)
Throughput diffs for windows/x86 ran on windows/x86Overall (+0.03% to +0.05%)
MinOpts (+0.02% to +0.16%)
FullOpts (+0.03% to +0.05%)
Details here Throughput diffs for linux/arm64 ran on linux/x64Overall (+0.03% to +0.05%)
MinOpts (+0.03% to +0.12%)
FullOpts (+0.03% to +0.06%)
Throughput diffs for linux/x64 ran on linux/x64Overall (+0.04% to +0.06%)
MinOpts (+0.03% to +0.16%)
FullOpts (+0.04% to +0.06%)
Details here |
Diff results for #98192Throughput diffsThroughput diffs for linux/arm64 ran on linux/x64Overall (+0.02% to +0.04%)
MinOpts (+0.02% to +0.09%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for linux/x64 ran on linux/x64Overall (+0.03% to +0.04%)
MinOpts (+0.03% to +0.14%)
FullOpts (+0.03% to +0.04%)
Throughput diffs for osx/arm64 ran on linux/x64Overall (+0.02% to +0.04%)
MinOpts (+0.02% to +0.11%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for windows/arm64 ran on linux/x64Overall (+0.02% to +0.04%)
MinOpts (+0.02% to +0.10%)
FullOpts (+0.02% to +0.04%)
Details here |
Diff results for #98192Assembly diffsAssembly diffs for linux/arm64 ran on windows/x64Diffs are based on 1,610,272 contexts (368,644 MinOpts, 1,241,628 FullOpts). MISSED contexts: 3,428 (0.21%) Overall (-16 bytes)
FullOpts (-16 bytes)
Assembly diffs for linux/x64 ran on windows/x64Diffs are based on 1,620,764 contexts (360,162 MinOpts, 1,260,602 FullOpts). MISSED contexts: 3,086 (0.19%) Overall (-15 bytes)
FullOpts (-15 bytes)
Assembly diffs for osx/arm64 ran on windows/x64Diffs are based on 1,733,061 contexts (561,303 MinOpts, 1,171,758 FullOpts). MISSED contexts: 3,460 (0.20%) Overall (-16 bytes)
FullOpts (-16 bytes)
Assembly diffs for windows/arm64 ran on windows/x64Diffs are based on 1,477,297 contexts (263,527 MinOpts, 1,213,770 FullOpts). MISSED contexts: 3,464 (0.23%) Overall (-16 bytes)
FullOpts (-16 bytes)
Assembly diffs for windows/x64 ran on windows/x64Diffs are based on 1,999,231 contexts (587,594 MinOpts, 1,411,637 FullOpts). MISSED contexts: 3,657 (0.18%) Overall (-2 bytes)
FullOpts (-2 bytes)
Details here Assembly diffs for linux/arm ran on windows/x86Diffs are based on 1,449,330 contexts (345,734 MinOpts, 1,103,596 FullOpts). MISSED contexts: 55,656 (3.70%) Overall (+0 bytes)
FullOpts (+0 bytes)
Assembly diffs for windows/x86 ran on windows/x86Diffs are based on 1,618,717 contexts (327,626 MinOpts, 1,291,091 FullOpts). MISSED contexts: base: 11,019 (0.68%), diff: 11,022 (0.68%) Overall (+38 bytes)
FullOpts (+38 bytes)
Details here Throughput diffsThroughput diffs for linux/arm64 ran on windows/x64Overall (+0.02% to +0.04%)
MinOpts (+0.02% to +0.10%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for linux/x64 ran on windows/x64Overall (+0.03% to +0.04%)
MinOpts (+0.03% to +0.14%)
FullOpts (+0.03% to +0.04%)
Throughput diffs for osx/arm64 ran on windows/x64Overall (+0.02% to +0.04%)
MinOpts (+0.02% to +0.11%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for windows/arm64 ran on windows/x64Overall (+0.02% to +0.04%)
MinOpts (+0.02% to +0.10%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for windows/x64 ran on windows/x64Overall (+0.03% to +0.05%)
MinOpts (+0.03% to +0.13%)
FullOpts (+0.03% to +0.05%)
Details here Throughput diffs for linux/arm64 ran on linux/x64Overall (+0.02% to +0.04%)
MinOpts (+0.02% to +0.09%)
FullOpts (+0.02% to +0.04%)
Throughput diffs for linux/x64 ran on linux/x64Overall (+0.02% to +0.04%)
MinOpts (+0.02% to +0.12%)
FullOpts (+0.02% to +0.04%)
Details here Throughput diffs for linux/arm ran on windows/x86Overall (+0.02% to +0.03%)
MinOpts (+0.02% to +0.09%)
FullOpts (+0.02% to +0.03%)
Throughput diffs for windows/x86 ran on windows/x86Overall (+0.02% to +0.03%)
MinOpts (+0.01% to +0.14%)
FullOpts (+0.02% to +0.03%)
Details here |
|
@AndyAyersMS PTAL. TP is slightly more reasonable now. I took a look at |
AndyAyersMS
approved these changes
Feb 9, 2024
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
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.
Part of #93020. Per discussion on #98054, set "naive" edge likelihoods (in other words, assume every successor edge is equally likely to be taken) in
fgAddRefPredwhen initializing preds.cc @dotnet/jit-contrib