Reduce TP for targets with more than 64 Registers Part 1 #112704

DeepakRajendrakumaran · 2025-02-19T18:32:47Z

Theory:

My profiling shows among others 3 methods whose tpdiff spikes when we have more than 64 registers – processKills(), freeRegisters(), processBlockStartLocations(). These 3 iterate over a regMaskTP and modifies it. The idea here is to work on 2regMaskSmalls separately instead of regMaskTP

Expectation:

Targets with less than 64 registers will not see any impact. Targets with more than 64 registers(currently only arm64) will see tpdiff improvement.

Relevant issue:

#113670

Result

Using arm64 as proxy for target with more than 64 registers, from CI

linux arm64

Overall (-0.87% to -0.61%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	-0.71%
benchmarks.run_pgo.linux.arm64.checked.mch	-0.75%
benchmarks.run_tiered.linux.arm64.checked.mch	-0.87%
coreclr_tests.run.linux.arm64.checked.mch	-0.78%
libraries.crossgen2.linux.arm64.checked.mch	-0.68%
libraries.pmi.linux.arm64.checked.mch	-0.67%
libraries_tests.run.linux.arm64.Release.mch	-0.80%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	-0.61%
realworld.run.linux.arm64.checked.mch	-0.64%
smoke_tests.nativeaot.linux.arm64.checked.mch	-0.72%

MinOpts (-1.39% to -0.88%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	-0.99%
benchmarks.run_pgo.linux.arm64.checked.mch	-1.13%
benchmarks.run_tiered.linux.arm64.checked.mch	-1.12%
coreclr_tests.run.linux.arm64.checked.mch	-0.88%
libraries.crossgen2.linux.arm64.checked.mch	-0.89%
libraries.pmi.linux.arm64.checked.mch	-0.98%
libraries_tests.run.linux.arm64.Release.mch	-1.12%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	-0.89%
realworld.run.linux.arm64.checked.mch	-1.25%
smoke_tests.nativeaot.linux.arm64.checked.mch	-1.39%

FullOpts (-0.72% to -0.60%)

Collection	PDIFF
benchmarks.run.linux.arm64.checked.mch	-0.70%
benchmarks.run_pgo.linux.arm64.checked.mch	-0.71%
benchmarks.run_tiered.linux.arm64.checked.mch	-0.63%
coreclr_tests.run.linux.arm64.checked.mch	-0.71%
libraries.crossgen2.linux.arm64.checked.mch	-0.68%
libraries.pmi.linux.arm64.checked.mch	-0.67%
libraries_tests.run.linux.arm64.Release.mch	-0.70%
libraries_tests_no_tiered_compilation.run.linux.arm64.Release.mch	-0.60%
realworld.run.linux.arm64.checked.mch	-0.63%
smoke_tests.nativeaot.linux.arm64.checked.mch	-0.72%

Local tests done:

I added some temporary changes locally so that x64 has more than 64 registers and ran superpmi tpdiff with/without the optimization to simulate impact with more than 64 registers

The results from local testing is below

Overall (-0.66% to -0.47%)

Collection	PDIFF
aspnet.run.windows.x64.checked.mch	-0.66%
benchmarks.run.windows.x64.checked.mch	-0.51%
benchmarks.run_pgo.windows.x64.checked.mch	-0.64%
benchmarks.run_pgo_optrepeat.windows.x64.checked.mch	-0.51%
benchmarks.run_tiered.windows.x64.checked.mch	-0.59%
coreclr_tests.run.windows.x64.checked.mch	-0.63%
libraries.crossgen2.windows.x64.checked.mch	-0.49%
libraries.pmi.windows.x64.checked.mch	-0.57%
libraries_tests.run.windows.x64.Release.mch	-0.66%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	-0.54%
realworld.run.windows.x64.checked.mch	-0.54%
smoke_tests.nativeaot.windows.x64.checked.mch	-0.47%

MinOpts (-1.14% to -0.65%)

Collection	PDIFF
aspnet.run.windows.x64.checked.mch	-0.87%
benchmarks.run.windows.x64.checked.mch	-0.75%
benchmarks.run_pgo.windows.x64.checked.mch	-0.86%
benchmarks.run_pgo_optrepeat.windows.x64.checked.mch	-0.75%
benchmarks.run_tiered.windows.x64.checked.mch	-0.82%
coreclr_tests.run.windows.x64.checked.mch	-0.69%
libraries.crossgen2.windows.x64.checked.mch	-0.65%
libraries.pmi.windows.x64.checked.mch	-0.71%
libraries_tests.run.windows.x64.Release.mch	-0.92%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	-0.75%
realworld.run.windows.x64.checked.mch	-1.14%
smoke_tests.nativeaot.windows.x64.checked.mch	-0.93%

FullOpts (-0.62% to -0.43%)

Collection	PDIFF
aspnet.run.windows.x64.checked.mch	-0.62%
benchmarks.run.windows.x64.checked.mch	-0.51%
benchmarks.run_pgo.windows.x64.checked.mch	-0.60%
benchmarks.run_pgo_optrepeat.windows.x64.checked.mch	-0.51%
benchmarks.run_tiered.windows.x64.checked.mch	-0.43%
coreclr_tests.run.windows.x64.checked.mch	-0.59%
libraries.crossgen2.windows.x64.checked.mch	-0.49%
libraries.pmi.windows.x64.checked.mch	-0.57%
libraries_tests.run.windows.x64.Release.mch	-0.57%
libraries_tests_no_tiered_compilation.run.windows.x64.Release.mch	-0.53%
realworld.run.windows.x64.checked.mch	-0.53%
smoke_tests.nativeaot.windows.x64.checked.mch	-0.47%

dotnet-policy-service · 2025-02-19T18:33:29Z

Tagging subscribers to this area: @hoyosjs
See info in area-owners.md if you want to be subscribed.

… SingleTypeRegSet instead.

DeepakRajendrakumaran · 2025-03-18T22:56:07Z

@BruceForstall @kunalspathak @jakobbotsch

kunalspathak · 2025-03-19T16:54:18Z

/azp run runtime-coreclr superpmi-replay

azure-pipelines · 2025-03-19T16:54:35Z

Azure Pipelines successfully started running 1 pipeline(s).

kunalspathak · 2025-03-19T17:28:09Z

src/coreclr/jit/lsra.cpp

    {
-        regNumber  killedReg        = genFirstRegNumFromMaskAndToggle(killedRegs);
+        regNumber killedReg = (regNumber)(BitOperations::BitScanForward(killedRegs) + regBase);


consider having an equivalent method of genFirstRegNumFromMaskAndToggle() that takes SingleTypeRegSet as parameter and operate on it. You can also pass lowBase and highBase, if required.

You are right. I should not have added regBase where I did either. The following xor is slightly incorrect.

Should be

regNumber killedReg = (regNumber)(genFirstRegNumFromMaskAndToggle(killedRegs) + regBase); RegRecord* regRecord = getRegisterRecord(killedReg);

Made this change as well. I'm done making changes now

kunalspathak · 2025-03-19T17:28:23Z

src/coreclr/jit/lsra.cpp

+//   regBase - `0` or `64` based on the `killedRegs` being processed
+//
+void LinearScan::freeKilledRegs(RefPosition* killRefPosition,
+                                regMaskSmall killedRegs,


Why using regMaskSmall instead of SingleTypeRegSet?

This reverts commit 65acca3.

kunalspathak · 2025-03-19T22:31:03Z

/azp run runtime-coreclr superpmi-replay

azure-pipelines · 2025-03-19T22:31:19Z

Azure Pipelines successfully started running 1 pipeline(s).

kunalspathak

LGTM

BruceForstall · 2025-03-19T22:56:43Z

cc @dotnet/jit-contrib

kunalspathak · 2025-03-20T15:23:20Z

/ba-g infra related changes

dotnet-issue-labeler bot added the area-Infrastructure-coreclr label Feb 19, 2025

dotnet-policy-service bot added the community-contribution Indicates that the PR has been added by a community member label Feb 19, 2025

DeepakRajendrakumaran changed the title ~~Testing out r2r tests~~ TP diff check Mar 4, 2025

DeepakRajendrakumaran force-pushed the 112329 branch 3 times, most recently from ae81451 to 58b7c1a Compare March 11, 2025 00:16

DeepakRajendrakumaran force-pushed the 112329 branch 2 times, most recently from 4acef83 to f726c2e Compare March 17, 2025 23:39

DeepakRajendrakumaran changed the title ~~TP diff check~~ Reduce TP for targets with more than 64 Registers Mar 17, 2025

DeepakRajendrakumaran force-pushed the 112329 branch 2 times, most recently from 32047d1 to 7ec2042 Compare March 18, 2025 00:50

DeepakRajendrakumaran changed the title ~~Reduce TP for targets with more than 64 Registers~~ Draft : https://github.com/dotnet/runtime/pull/112704/filesReduce TP for targets with more than 64 Registers Mar 18, 2025

DeepakRajendrakumaran changed the title ~~Draft : https://github.com/dotnet/runtime/pull/112704/filesReduce TP for targets with more than 64 Registers~~ Draft : Reduce TP for targets with more than 64 Registers Mar 18, 2025

build-analysis bot mentioned this pull request Mar 18, 2025

The Operation will be canceled. The next steps may not contain expected logs. dotnet/dnceng#3008

Open

3 tasks

DeepakRajendrakumaran force-pushed the 112329 branch 2 times, most recently from ac9e948 to a32c3cd Compare March 18, 2025 21:38

Reduce tpdiff when regMaskTP has more than 64 registers by working on…

a32c3cd

… SingleTypeRegSet instead.

DeepakRajendrakumaran marked this pull request as ready for review March 18, 2025 22:56

DeepakRajendrakumaran changed the title ~~Draft : Reduce TP for targets with more than 64 Registers~~ Reduce TP for targets with more than 64 Registers Mar 18, 2025

DeepakRajendrakumaran changed the title ~~Reduce TP for targets with more than 64 Registers~~ Reduce TP for targets with more than 64 Registers Part 1 Mar 18, 2025

kunalspathak reviewed Mar 19, 2025

View reviewed changes

DeepakRajendrakumaran added 4 commits March 19, 2025 11:26

Handle liveRegs locally as SingleTypeRegSet.

65acca3

Addressing review comments.

8aff345

Revert "Handle liveRegs locally as SingleTypeRegSet."

1f92535

This reverts commit 65acca3.

Fix formatting.

ab32d2f

build-analysis bot mentioned this pull request Mar 19, 2025

/root/helix/work/correlation/scripts/<hash>/execute.sh: Permission denied dotnet/dnceng#3412

Open

3 tasks

kunalspathak approved these changes Mar 19, 2025

View reviewed changes

kunalspathak merged commit bb5c6c3 into dotnet:main Mar 20, 2025
110 of 113 checks passed

DeepakRajendrakumaran mentioned this pull request Mar 20, 2025

Reduce iteration over regmaskTP in relevant methods #113670

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Reduce TP for targets with more than 64 Registers Part 1 #112704

Reduce TP for targets with more than 64 Registers Part 1 #112704

DeepakRajendrakumaran commented Feb 19, 2025 •

edited

Loading

dotnet-policy-service bot commented Feb 19, 2025

DeepakRajendrakumaran commented Mar 18, 2025

kunalspathak commented Mar 19, 2025

azure-pipelines bot commented Mar 19, 2025

kunalspathak Mar 19, 2025

DeepakRajendrakumaran Mar 19, 2025

DeepakRajendrakumaran Mar 19, 2025

kunalspathak Mar 19, 2025

DeepakRajendrakumaran Mar 19, 2025

kunalspathak commented Mar 19, 2025

azure-pipelines bot commented Mar 19, 2025

kunalspathak left a comment

BruceForstall commented Mar 19, 2025

kunalspathak commented Mar 20, 2025

Reduce TP for targets with more than 64 Registers Part 1 #112704

Reduce TP for targets with more than 64 Registers Part 1 #112704

Conversation

DeepakRajendrakumaran commented Feb 19, 2025 • edited Loading

Theory:

Expectation:

Relevant issue:

Result

linux arm64

Local tests done:

dotnet-policy-service bot commented Feb 19, 2025

DeepakRajendrakumaran commented Mar 18, 2025

kunalspathak commented Mar 19, 2025

azure-pipelines bot commented Mar 19, 2025

kunalspathak Mar 19, 2025

Choose a reason for hiding this comment

DeepakRajendrakumaran Mar 19, 2025

Choose a reason for hiding this comment

DeepakRajendrakumaran Mar 19, 2025

Choose a reason for hiding this comment

kunalspathak Mar 19, 2025

Choose a reason for hiding this comment

DeepakRajendrakumaran Mar 19, 2025

Choose a reason for hiding this comment

kunalspathak commented Mar 19, 2025

azure-pipelines bot commented Mar 19, 2025

kunalspathak left a comment

Choose a reason for hiding this comment

BruceForstall commented Mar 19, 2025

kunalspathak commented Mar 20, 2025

DeepakRajendrakumaran commented Feb 19, 2025 •

edited

Loading