[Perf] Linux/arm64: Regressions in System.Collections.TryAddGiventSize

[performanceautofiler]

Run Information

Name	Value
Architecture	arm64
OS	ubuntu 20.04
Queue	AmpereUbuntu
Baseline	60be20ccdea212f1fa841f01114cc72324e7d56f
Compare	276aa2dc8f7cf5350df4c1bf8cdf056f207b1518
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Collections.TryAddGiventSize<Int32>

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
Dictionary - Duration of single invocation 📝 - Benchmark Source 📈 - ADX Test Multi Config Graph	5.68 μs	6.21 μs	1.09	0.03	False

Test Report

Repro

General Docs link: https://github.com/dotnet/performance/blob/main/docs/benchmarking-workflow-dotnet-runtime.md

git clone https://github.com/dotnet/performance.git
python3 .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Collections.TryAddGiventSize<Int32>*'

Payloads

Baseline
Compare

System.Collections.TryAddGiventSize<Int32>.Dictionary(Count: 512)

ETL Files

Histogram

JIT Disasms

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository

[EgorBo]

Could be #95340

[EgorBo]

cc @amanasifkhalid

[ghost]

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

Issue Details

---

Run Information

Name	Value
Architecture	arm64
OS	ubuntu 20.04
Queue	AmpereUbuntu
Baseline	60be20ccdea212f1fa841f01114cc72324e7d56f
Compare	276aa2dc8f7cf5350df4c1bf8cdf056f207b1518
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Collections.TryAddGiventSize<Int32>

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
Dictionary - Duration of single invocation 📝 - Benchmark Source 📈 - ADX Test Multi Config Graph	5.68 μs	6.21 μs	1.09	0.03	False

Test Report

Repro

General Docs link: https://github.com/dotnet/performance/blob/main/docs/benchmarking-workflow-dotnet-runtime.md

git clone https://github.com/dotnet/performance.git
python3 .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Collections.TryAddGiventSize<Int32>*'

Payloads

Baseline
Compare

System.Collections.TryAddGiventSize<Int32>.Dictionary(Count: 512)

ETL Files

Histogram

JIT Disasms

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository

Author:	performanceautofiler[bot]
Assignees:	-
Labels:	arch-arm64, os-linux, area-CodeGen-coreclr, untriaged, runtime-coreclr, PGO
Milestone:	-

[amanasifkhalid]

I took a look at the disassemblies of this benchmark with the base/diff builds, and I didn't see any differences in codegen on Linux ARM64. I couldn't find any jumps to the next block in Tier0, so I wouldn't expect the optimization touched by #95340 to do anything different for this benchmark. I guess this regression could be caused by lower TP in the JIT, though I'm skeptical of that hypothesis, too. I don't have a Linux ARM64 box to try this on, but when targeting Unix ARM64 on Windows x64 with SPMI and compiling the benchmark method System.Collections.TryAddGiventSize'1[int]:Dictionary():System.Collections.Generic.Dictionary'2[int,int]:this, the TP diff is a small improvement (0.02%) over the baseline JIT (though this may not be the case when executing on Linux ARM64).

Looking at my changes in #95340, I would expect TP to only improve: For unoptimized methods with jumps that can be removed, we now skip emitting those jumps, which easily pays for the checks in BasicBlock::CanRemoveJumpToNext. And for unoptimized methods with no jumps to the next block (such as the above benchmark), we now skip checking if optimizations are enabled (or if BBF_NONE_QUIRK is set), and immediately short-circuit on the JumpsToNext check, so I'd expect this path to be slightly faster, too. I might be too focused on the benchmark method itself, but I don't think #95340 is to blame, here.

@EgorBo do you know what the "Test Quality" score means, in this case? Is that a measure of its flakiness? Is a larger score better? Thank you!

[amanasifkhalid]

This seems to have since resolved itself: Test report.