[Perf] Linux/arm64: 49 Improvements on 8/5/2026 1:23:13 AM +00:00 #77882

Activity

1. performanceautofiler

   added

   ampere

   arch-arm64

   bdn-update

   branch-refs/heads/main

   compilationmode-tiered

   kind-micro

   os-linux

   perf-improvement

   runkind-micro

   runtime-coreclr

   untriaged

   on Aug 13, 2026
2. performanceautofiler commented on Aug 13, 2026

   

   performanceautofilerbot

   on Aug 13, 2026 – with PerformanceAutoFiler

   Author

   More actions

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in System.Numerics.Tensors.Tests.Perf_FloatingPointTensorPrimitives<Double>

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   FusedMultiplyAdd_Vectors - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	1.58 μs	1.37 μs	0.87	0.08	False
   FusedMultiplyAdd_Vectors - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	61.65 ns	53.78 ns	0.87	0.09	False
   FusedMultiplyAdd_ScalarAddend - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	833.19 ns	732.76 ns	0.88	0.26	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.Numerics.Tensors.Tests.Perf_FloatingPointTensorPrimitives<Double>*'

   Details

   System.Numerics.Tensors.Tests.Perf_FloatingPointTensorPrimitives<Double>.FusedMultiplyAdd_Vectors(BufferLength: 3079)

   ETL Files

   Histogram

   JIT Disasms

   System.Numerics.Tensors.Tests.Perf_FloatingPointTensorPrimitives<Double>.FusedMultiplyAdd_Vectors(BufferLength: 128)

   ETL Files

   Histogram

   JIT Disasms

   System.Numerics.Tensors.Tests.Perf_FloatingPointTensorPrimitives<Double>.FusedMultiplyAdd_ScalarAddend(BufferLength: 3079)

   ETL Files

   Histogram

   JIT Disasms

   Docs

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

   ---

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in Microsoft.Extensions.Logging.FormattingOverhead

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   FourArguments_EnumerableArgument - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	695.64 ns	492.64 ns	0.71	0.57	False
   FourArguments_DefineMessage - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	277.21 ns	209.56 ns	0.76	0.31	False
   TwoArguments - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	184.13 ns	140.51 ns	0.76	0.19	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 'Microsoft.Extensions.Logging.FormattingOverhead*'

   Details

   Microsoft.Extensions.Logging.FormattingOverhead.FourArguments_EnumerableArgument

   ETL Files

   Histogram

   JIT Disasms

   Microsoft.Extensions.Logging.FormattingOverhead.FourArguments_DefineMessage

   ETL Files

   Histogram

   JIT Disasms

   Microsoft.Extensions.Logging.FormattingOverhead.TwoArguments

   ETL Files

   Histogram

   JIT Disasms

   Docs

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

   ---

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in System.Text.Json.Tests.Utf8JsonReaderCommentsTests

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   Utf8JsonReaderCommentParsing - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	4.46 μs	2.27 μs	0.51	0.00	False
   Utf8JsonReaderCommentParsing - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	4.41 μs	2.26 μs	0.51	0.01	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.Text.Json.Tests.Utf8JsonReaderCommentsTests*'

   Details

   System.Text.Json.Tests.Utf8JsonReaderCommentsTests.Utf8JsonReaderCommentParsing(CommentHandling: Skip, SegmentSize: 0, TestCase: LongMultiLine)

   ETL Files

   Histogram

   JIT Disasms

   System.Text.Json.Tests.Utf8JsonReaderCommentsTests.Utf8JsonReaderCommentParsing(CommentHandling: Allow, SegmentSize: 0, TestCase: LongMultiLine)

   ETL Files

   Histogram

   JIT Disasms

   Docs

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

   ---

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in System.Numerics.Tensors.Tests.Perf_NumberTensorPrimitives<Double>

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   Max_Vector - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	851.09 ns	668.32 ns	0.79	0.30	False
   AddMultiply_Vectors - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	1.62 μs	1.41 μs	0.87	0.06	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.Numerics.Tensors.Tests.Perf_NumberTensorPrimitives<Double>*'

   Details

   System.Numerics.Tensors.Tests.Perf_NumberTensorPrimitives<Double>.Max_Vector(BufferLength: 3079)

   ETL Files

   Histogram

   JIT Disasms

   System.Numerics.Tensors.Tests.Perf_NumberTensorPrimitives<Double>.AddMultiply_Vectors(BufferLength: 3079)

   ETL Files

   Histogram

   JIT Disasms

   Docs

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

   ---

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in BenchmarksGame.KNucleotide_1

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   RunBench - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	281.27 ms	247.89 ms	0.88	0.00	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 'BenchmarksGame.KNucleotide_1*'

   Details

   BenchmarksGame.KNucleotide_1.RunBench

   ETL Files

   Histogram

   JIT Disasms

   Docs

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

   ---

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in System.Collections.TryGetValueTrue<String, String>

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   OrderedDictionary - Duration of single invocation ADX - Test Multi Config Graph	10.57 μs	9.48 μs	0.90	0.11	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.TryGetValueTrue<String, String>*'

   Details

   System.Collections.TryGetValueTrue<String, String>.OrderedDictionary(Size: 512)

   ETL Files

   Histogram

   JIT Disasms

   Docs

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

   ---

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in System.Numerics.Tests.Perf_BigInteger

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   ToStringX - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	78.38 ns	54.03 ns	0.69	0.18	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.Numerics.Tests.Perf_BigInteger*'

   Details

   System.Numerics.Tests.Perf_BigInteger.ToStringX(numberString: 123)

   ETL Files

   Histogram

   JIT Disasms

   Docs

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

   ---

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in Benchstone.BenchI.BubbleSort2

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   Test - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	33.84 ms	31.70 ms	0.94	0.10	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 'Benchstone.BenchI.BubbleSort2*'

   Details

   Benchstone.BenchI.BubbleSort2.Test

   ETL Files

   Histogram

   JIT Disasms

   Docs

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

   ---

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in System.Text.Json.Serialization.Tests.ReadJson<HashSet<String>>

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   DeserializeFromReader - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	12.91 μs	11.48 μs	0.89	0.20	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.Text.Json.Serialization.Tests.ReadJson<HashSet<String>>*'

   Details

   System.Text.Json.Serialization.Tests.ReadJson<HashSet<String>>.DeserializeFromReader(Mode: SourceGen)

   ETL Files

   Histogram

   JIT Disasms

   Docs

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

   ---

   Run Information

   Name	Value
   Architecture	arm64
   OS	ubuntu 22.04
   Queue	AmpereUbuntu
   Baseline	2dd5330804f9d4f49599d82c90a67892d4dd8af0
   Compare	c8f97975ab0d8ff84aebf176829e17f7c7052f03
   Diff	Diff
   Configs	CompilationMode:tiered, RunKind:micro

   Improvements in System.Collections.ContainsKeyTrue<String, String>

   Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
   OrderedDictionary - Duration of single invocation ADX - Test Multi Config Graph	10.51 μs	9.48 μs	0.90	0.09	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.ContainsKeyTrue<String, String>*'

   Details

   System.Collections.ContainsKeyTrue<String, String>.OrderedDictionary(Size: 512)

   ETL Files

   Histogram

   JIT Disasms

   Docs

   Profiling workflow for dotnet/runtime repository
   Benchmarking workflow for dotnet/runtime repository
3. EgorBo commented on Aug 13, 2026

   

   EgorBo

   on Aug 13, 2026

   Member

   More actions

   Should be dotnet/runtime#129688 I think (we were expecting improvements from it)

   👍React with 👍1Reacted by Jonathan Davies
4. LoopedBard3 commented on Aug 13, 2026

   

   LoopedBard3

   on Aug 13, 2026

   Last edited by LoopedBard3

   Member

   More actions

   🔍 Automated Triage Analysis

   Summary: Verified improvement in Utf8JsonReaderCommentParsing(CommentHandling: Skip, SegmentSize: 0, TestCase: LongMultiLine) — caused by commit 23ec3d982ce2.

   Finding Confidence: 5/5 (analysis accuracy)
   Improvement Confidence: 5/5 (likelihood of true improvement)

   Likely Cause: 23ec3d982ce25e7b3037296eeed2dec0d0837795

   🤖 Proposed Next Actions: Known-cause improvement — link & close — 0 done, 3 dry-run (rule R2_known_improvement — see Full Analysis for the table).

   📊 Full Analysis (click to expand)

   Summary: All 49 reported movements are genuine performance improvements — the 12 statistically sampled tests all show clean step-change gains (−6% to −62%) with strong signals (|z| = 2.5–26.1) over 297–300 historical runs, so finding confidence is very high (5/5). The gains span unrelated areas (Reflection TypeName parsing, ConcurrentQueue.IsEmpty, Utf8JsonReader comment parsing, Logging formatting, Tensors, BigInteger), which points at one or more broad JIT codegen changes in this 35-commit window rather than a single component fix; the leading candidate is f107b23 ("JIT: Avoid unnecessary argument spills in indirect call transformer"), but with multiple plausible JIT commits the attribution is moderate (improvement confidence 3/5) and bisection is recommended.

   Issue Overview

   • Issue: [Perf] Linux/arm64: 49 Improvements on 8/5/2026 1:23:13 AM +00:00 #77882 - [Perf] Linux/arm64: 49 Improvements on 8/5/2026 1:23:13 AM +00:00
   • Type: Performance Improvement
   • Date: 2026-08-05 | Queue: Ubuntu.2204.Arm64.Perf | OS/Arch: linux/arm64
   • Commits: 2dd53308 → c8f97975 (35 commits)
   • Labels: perf-improvement, arch-arm64, os-linux, runtime-coreclr, ampere, kind-micro, compilationmode-tiered, runkind-micro, bdn-update
   • Runtime stack: Desktop + CoreCLR (arm64, tiered) — arm64 CoreCLR JIT is the exercised codegen path.
   • Historical Data: Available (297–300 prior runs per sampled test; 1,197 before-runs, ~24–25 after-runs)

   Test Analysis

   12 representative tests sampled from the 49 improvements (units: ns; KNucleotide/BubbleSort in ns of the reported harness scale). All classified Improvement / StepChange by AnalyzeHistoricalData.

   Test Name	Baseline	Compare	Δ%	μ	σ	n	z	Thresh%	Noise?	Assessment
   Concurrent.IsEmpty.Queue(512)	11.547	4.377	-62.1%	10.333	0.904	299	-6.59	17.5%	N	Step change, new low plateau
   Concurrent.IsEmpty.Queue(0)	12.623	5.959	-52.8%	11.672	0.749	300	-7.63	12.8%	N	Step change, ~54% floor
   Utf8JsonReaderCommentParsing(Skip,LongMultiLine)	4464	2272	-49.1%	4398.9	81.6	300	-26.07	5%	N	Massive step, CV 1.9%
   BigInteger.ToStringX(123)	78.38	54.03	-31.1%	68.71	5.82	298	-2.52	16.9%	N	Step, landed mid-after-window
   Logging.FourArguments_EnumerableArgument	695.64	492.64	-29.2%	690.18	51.93	297	-3.80	15.1%	N	Step, no before/after overlap
   Logging.TwoArguments	184.13	140.51	-23.7%	175.03	5.73	298	-6.02	6.6%	N	Step, stable baseline
   Tensors.Max_Vector(3079)	851.09	668.32	-21.5%	805.81	50.33	298	-2.73	12.5%	N	Step over prior upward drift
   TypeName.Parse_AssemblyQualifiedName(Int32[])	1158.21	899.78	-22.3%	1128.21	25.72	300	-8.88	5%	N	Step, disjoint ranges
   TypeName.Parse_AssemblyQualifiedName(int)	1073.18	842.59	-21.5%	1075.28	23.46	300	-9.92	5%	N	Step, stable baseline
   BenchmarksGame.KNucleotide_1.RunBench	281.27M	247.89M	-11.9%	288.99M	4.76M	300	-8.64	5%	N	Step on top of prior drift
   Benchstone.BubbleSort2.Test	33.84M	31.70M	-6.3%	33.80M	0.71M	300	-2.95	5%	N	Modest but genuine step
   Collections.ContainsKeyTrue.OrderedDictionary(512)	10512	9481	-9.8%	10464.5	227.5	300	-4.32	5%	N	Step below historical floor

   Legend: μ=mean, σ=stddev, n=runs, z=z-score, Thresh%=adaptive threshold

   Finding Confidence: 5/5 - Comprehensive historical data (≈300 before-runs per test, HasSufficientData=true), tight after-clusters disjoint from before-ranges, no noise classifications. The full set of 49 tests was not each run through the statistical tool, but the sampled spread (largest, mid, and smallest deltas across every affected namespace) is uniformly clean.

   Candidate Stack Relevance

   Detected stack: Desktop + CoreCLR (arm64). Relevant = src/coreclr/jit (arm64 codegen exercised). Shared = BCL/eng/native. LessLikely = mono/wasm-only trees.

   Commit	Touched Paths (summary)	Relevance	Notes
   f107b23	src/coreclr/jit (indirect call transformer)	Relevant	Reduces arg spills on indirect/virtual/GDV calls — broad codegen impact across dispatch-heavy benchmarks.
   eb3dce1	src/coreclr/jit/objectalloc.cpp	Relevant	CEA (enumerator) cloning fix — affects IEnumerable/GDV allocation paths (e.g. Logging EnumerableArgument, span enumeration).
   ca4ed7d	src/coreclr, src/native	Relevant	Restores symbol-free decoding for CORINFO_HELP_NEW* allocation helpers — touches the allocation-helper fast path used everywhere.
   23ec3d9	src/coreclr (arm64 SIMD)	Relevant	arm64 ExtractMostSignificantBits consumer optimization — plausibly explains the Tensors vector gains but not the scalar/parse benchmarks.
   18f89a4	src/coreclr (x64 barrier)	LessLikely	VOLATILE_MEMORY_BARRIER change is x64-Unix only; arm64 not affected.
   e275aeb	src/coreclr (xarch bit idioms)	LessLikely	bzhi/bts/btc emit is x86/x64 only; arm64 not exercised.
   ae97cad	src/coreclr (wasm Virtual IP)	LessLikely	WASM-specific JIT path; not exercised by CoreCLR arm64.

   Confirmed Improvements

   The 12 sampled tests below are all confirmed genuine improvements (representative of the full 49). Attribution is shared across the whole group because the movements co-occur on the same date/commit and no single commit obviously accounts for every area.

   System.Text.Json Utf8JsonReaderCommentParsing (Skip/Allow, LongMultiLine) (-49.1% / -48.7%)

   • Delta: 4464 → 2272 ns (-49.1%)
   • Historical Context: μ=4398.9 (σ=81.6, n=300); z=-26.07; adaptive threshold=5%
   • Assessment: Massive step change, historically ultra-stable (CV=1.9%). Highest-confidence signal in the set.
   • Finding Confidence: 5/5 - excellent, stable history; compare cluster 2207–2272 cleanly disjoint.
   • Improvement Confidence: 3/5 - unambiguous real gain, but single-commit attribution unverified within a 35-commit range.
   • Likely Cause: f107b23 - "JIT: Avoid unnecessary argument spills in indirect call transformer"
     • Affected Areas: JIT / codegen (indirect & virtual dispatch, GDV)
     • Stack Relevance: Relevant
     • Rationale: Utf8JsonReader comment parsing is virtual-dispatch and helper-call heavy; reducing argument spills in the indirect-call transformer improves such hot loops broadly. Timing aligns with the compare commit. Attribution is not verified — a ~50% halving is large for this change alone, so bisection is warranted.
     • Candidates Considered:
       • eb3dce1 - CEA/enumerator allocation cloning fix
       • ca4ed7d - allocation-helper decoding restore
   • Verification Status: Not verified (bisection recommended)

   System.Collections.Concurrent.IsEmpty.Queue (Size 512 / 0) (-62.1% / -52.8%)

   • Delta: 11.547 → 4.377 ns (-62.1%); 12.623 → 5.959 ns (-52.8%)
   • Historical Context: μ=10.333/11.672 (σ=0.90/0.75, n=299/300); z=-6.59/-7.63; adaptive threshold=17.5%/12.8%
   • Assessment: Both step to a new low plateau; identical timing suggests a shared root cause.
   • Finding Confidence: 5/5 - large before-history; compare clusters (4.37–4.40, 5.67–6.02) well below all prior regimes.
   • Improvement Confidence: 3/5 - clearly real; attribution shared with the group, unverified.
   • Likely Cause: f107b23 (group attribution); alternately allocation/JIT commits above.
     • Stack Relevance: Relevant
     • Rationale: IsEmpty on ConcurrentQueue is a tiny hot method dominated by field reads and branch codegen; a broad JIT codegen/inlining improvement matches the >50% drop. The two sizes moving together on the same commit corroborate a shared JIT cause.
   • Verification Status: Not verified

   System.Reflection.Metadata.Perf_TypeName.* (Parse/ParseAndGet family) (-20% to -22%)

   • Delta: e.g. Parse_AssemblyQualifiedName(Int32[]): 1158.21 → 899.78 ns (-22.3%)
   • Historical Context: μ≈1128/1075 (σ≈26/23, n=300); z=-8.88/-9.92; adaptive threshold=5%
   • Assessment: Consistent ~20% step across the entire TypeName parsing family (30+ variants), very stable history.
   • Finding Confidence: 5/5 - tight, disjoint before/after ranges.
   • Improvement Confidence: 3/5 - real and highly consistent across the family; commit-level attribution unverified.
   • Likely Cause: f107b23 (group attribution)
     • Stack Relevance: Relevant
     • Rationale: TypeName parsing is span- and dispatch-heavy; the uniform ~20% shift across every parameterization is characteristic of a codegen-level improvement rather than an algorithmic change to a single method.
   • Verification Status: Not verified

   Microsoft.Extensions.Logging FormattingOverhead (FourArguments_EnumerableArgument -29.2%, TwoArguments -23.7%, FourArguments_DefineMessage -24.4%)

   • Historical Context: e.g. TwoArguments μ=175.0 (σ=5.7, n=298); z=-6.02; threshold=6.6%
   • Assessment: Step change, after-values entirely below before-range.
   • Finding Confidence: 5/5
   • Improvement Confidence: 3/5
   • Likely Cause: eb3dce1 (CEA/enumerator cloning) is especially plausible for EnumerableArgument; group JIT attribution otherwise.
     • Stack Relevance: Relevant
     • Rationale: The EnumerableArgument path materializes an enumerator; the CEA cloning fix directly affects such allocation/enumeration codegen. Unverified.
   • Verification Status: Not verified

   Tensors, BigInteger, KNucleotide, BubbleSort, OrderedDictionary (-6% to -31%)

   • Tensors.Max_Vector(3079): 851.09 → 668.32 (-21.5%, z=-2.73) — plausibly aided by arm64 SIMD commit 23ec3d9.
   • BigInteger.ToStringX(123): 78.38 → 54.03 (-31.1%, z=-2.52) — improvement landed mid-after-window; confirm plateau.
   • KNucleotide.RunBench: -11.9% (z=-8.64); BubbleSort2: -6.3% (z=-2.95); OrderedDictionary ContainsKey/TryGetValue(512): ~-9.8% (z=-4.32).
   • Finding Confidence: 4–5/5 (BigInteger/Tensors slightly lower due to multi-regime/drift history).
   • Improvement Confidence: 3/5 - genuine; distributed across areas, group JIT attribution.
   • Verification Status: Not verified

   Open Questions

   • Historical Data: Available and strong (≈300 runs/test); BigInteger and ConcurrentQueue show multi-regime history (CV 8.5%/8.7%) that widens thresholds but does not overturn the step-change signal.
   • Confidence Limitations: Large commit range (35 commits) with multiple plausible JIT candidates; no single commit obviously explains the full breadth (parse + JSON + collections + tensors), so the group attribution to [f107b23] is a hypothesis, not a confirmed cause.
   • Verification Needed: Bisection across [2dd53308]…[c8f97975] on a high-signal, cheap test (e.g. TypeName.Parse_AssemblyQualifiedName or ConcurrentQueue.IsEmpty) to isolate the responsible commit(s).
   • Bisection Status: Not performed (to be handled separately).
   • Correlation Signal: 12 co-occurring sibling arm64 improvement issues in the same window strongly reinforce that these are real, shared-cause improvements rather than noise.

   Recommended Actions

   1. Bisection - Run bisection on commits 2dd53308…c8f97975 using RunWithBuildAtHash on System.Reflection.Metadata.Perf_TypeName.Parse_AssemblyQualifiedName(input: typeof(int)) and System.Collections.Concurrent.IsEmpty<String>.Queue(Size: 512) (high z, cheap, distinct areas) to isolate the improving commit(s).
   2. Verification - Confirm the leading JIT candidate f107b23 with RunWithBuildAtHash vs its parent; if it does not reproduce the breadth, test [eb3dce1] and [ca4ed7d].
   3. Follow-up - Confirm BigInteger.ToStringX(123) plateau (~54) holds across additional after-runs, since the gain landed mid-window.

   Bisection Results

   Test	Status	Culprit Commit	Relevance	Reported Diff (issue)	Bisected Diff (at culprit)
   Int32[]))	🔁 Did not reproduce	-	-	1.2→0.9 μs (-22%)	≈ (baseline ≈ head)
   Queue(Size: 512)	🔁 Did not reproduce	-	-	11.5→4.4 ns (-62%)	≈ (baseline ≈ head)
   FusedMultiplyAdd_Vectors(BufferLength...	🔁 Did not reproduce	-	-	1.6→1.4 μs (-13%)	≈ (baseline ≈ head)
   FourArguments_EnumerableArgument	🔁 Did not reproduce	-	-	695.6→492.6 ns (-29%)	≈ (baseline ≈ head)
   Utf8JsonReaderCommentParsing(CommentH...	✅ Culprit found	23ec3d982ce2	✅ RELATED	4.5→2.3 μs (-49%)	4.4→2.3 μs (-48%)
   RunBench	🔁 Did not reproduce	-	-	281.3→247.9 ms (-12%)	≈ (baseline ≈ head)
   Test	🔁 Did not reproduce	-	-	33.8→31.7 ms (-6%)	≈ (baseline ≈ head)

   Total bisection time: 70m 28s
   Tests measured: 7/7
   Culprit found: 1/7

   System.Reflection.Metadata.Perf_TypeName.Parse_AssemblyQualifiedName(input: typeof(System.Int32[]))

   • Status: inconclusive

   Bisection details

   • Baseline: 1158.205 ns

   Per-commit measurements (oldest → newest) (ns)

   #	Commit	Result	Measurement
   0	2dd5330804f9	baseline	958.976 ns
   12	24630fa3a58e	-	922.866 ns
   23	f7d2d521667a	-	901.942 ns
   35	c8f97975ab0d	head	932.826 ns

   Helix jobs

   Job ID
   4eeb69ba-5669-43bb-8479-13e2ced997b4

   System.Collections.Concurrent.IsEmpty<String>.Queue(Size: 512)

   • Status: inconclusive

   Bisection details

   • Baseline: 11.547 ns

   Per-commit measurements (oldest → newest) (ns)

   #	Commit	Result	Measurement
   0	2dd5330804f9	baseline	4.396 ns
   12	24630fa3a58e	-	4.378 ns
   23	f7d2d521667a	-	4.383 ns
   35	c8f97975ab0d	head	4.393 ns

   Helix jobs

   Job ID
   e2193ea7-246a-48b9-b5da-0b48d719c52b

   System.Numerics.Tensors.Tests.Perf_FloatingPointTensorPrimitives<Double>.FusedMultiplyAdd_Vectors(BufferLength: 3079)

   • Status: inconclusive

   Bisection details

   • Baseline: 1.582 μs

   Per-commit measurements (oldest → newest) (μs)

   #	Commit	Result	Measurement
   0	2dd5330804f9	baseline	1.388 μs
   12	24630fa3a58e	-	1.384 μs
   23	f7d2d521667a	-	1.388 μs
   35	c8f97975ab0d	head	1.367 μs

   Helix jobs

   Job ID
   937ec104-3f5f-48ae-a822-f1327e79f7d0

   Microsoft.Extensions.Logging.FormattingOverhead.FourArguments_EnumerableArgument

   • Status: inconclusive

   Bisection details

   • Baseline: 695.637 ns

   Per-commit measurements (oldest → newest) (ns)

   #	Commit	Result	Measurement
   0	2dd5330804f9	baseline	530.421 ns
   12	24630fa3a58e	-	513.159 ns
   23	f7d2d521667a	-	514.868 ns
   35	c8f97975ab0d	head	603.236 ns

   Helix jobs

   Job ID
   e6361b2a-e037-43dd-addd-3107397fdcff

   System.Text.Json.Tests.Utf8JsonReaderCommentsTests.Utf8JsonReaderCommentParsing(CommentHandling: Skip, SegmentSize: 0, TestCase: LongMultiLine)

   • Culprit: 23ec3d982ce2
   • Relevance: RELATED
   • Explanation: Commit 23ec3d98 'arm64: Optimize ARM64 compare mask ExtractMostSignificantBits consumers (#129688)' is a JIT optimization targeting Arm64 vectorized ExtractMostSignificantBits code. The benchmark runs on Arm64 (Neoverse-N1) and Utf8JsonReader comment scanning uses vectorized most-significant-bit extraction, so the ~2x speedup (4378 ns -> 2266 ns) at exactly this commit is directly explained.

   Bisection details

   • Culprit: 23ec3d982ce2
   • Relevance: RELATED
   • Baseline: 4.464 μs
   • Culprit value: 2.263 μs

   Per-commit measurements (oldest → newest) (μs)

   #	Commit	Result	Measurement
   0	2dd5330804f9	baseline/good	4.511 μs
   12	24630fa3a58e	good	4.419 μs
   23	f7d2d521667a	good	4.379 μs
   24	23ec3d982ce2	culprit/bad	2.263 μs
   25	18f89a452f2a	bad	2.266 μs
   27	ff5613f81da6	bad	2.207 μs
   31	173e6b40e02c	bad	2.244 μs
   35	c8f97975ab0d	head/bad	2.239 μs

   Helix jobs

   Job ID
   34d5c276-628d-4556-bdbe-a2d62e61c546
   8f3b08ee-ebae-441d-8d69-9892a5d2a9fa
   591db1c3-3d71-48d3-8884-1e609fd7604e

   BenchmarksGame.KNucleotide_1.RunBench

   • Status: inconclusive

   Bisection details

   • Baseline: 281.270 ms

   Per-commit measurements (oldest → newest) (ms)

   #	Commit	Result	Measurement
   0	2dd5330804f9	baseline	247.971 ms
   12	24630fa3a58e	-	248.464 ms
   23	f7d2d521667a	-	247.943 ms
   35	c8f97975ab0d	head	248.445 ms

   Helix jobs

   Job ID
   88c1a361-554e-4627-ada5-a50c2855114a

   Benchstone.BenchI.BubbleSort2.Test

   • Status: inconclusive

   Bisection details

   • Baseline: 33.838 ms

   Per-commit measurements (oldest → newest) (ms)

   #	Commit	Result	Measurement
   0	2dd5330804f9	baseline	31.729 ms
   12	24630fa3a58e	-	31.671 ms
   23	f7d2d521667a	-	31.600 ms
   35	c8f97975ab0d	head	31.701 ms

   Helix jobs

   Job ID
   6db00153-bd46-4ced-8d41-060eae0197ce

   🤖 Proposed Next Actions

   Rule: R2_known_improvement — Known-cause improvement — link & close.

   Verified improvement attributed to commit 23ec3d982ce2 (finding 5/5, improvement 5/5, bisection RELATED).

   #	Action	Status	Details
   1	Back-reference cause PR	⏭️ Skipped (dry-run)	on culprit PR for 23ec3d982ce2
   2	Remove label	⏭️ Skipped (dry-run)	untriaged
   3	Close issue	⏭️ Skipped (dry-run)	reason completed

   🦆 Pre-bisection Re-analysis (advisory)

   ⚠️ Disputes the pre-bisection plan — advisory only, not applied

   Flags: wasted-bisection-slot, possible-noise

   Rationale: The Tensors, KNucleotide, and BubbleSort candidates all derive from the single group with finding confidence 0/5, unknown stack relevance, and no likely commit, so queuing three expensive bisection slots for an unverified finding is wasteful; the JSON, Concurrent Queue, TypeName, and Logging candidates are well-supported (findingConf 5/5, Relevant) and should stay.

   Suggested drop from bisection: System.Numerics.Tensors.Tests.Perf_FloatingPointTensorPrimitives<Double>.FusedMultiplyAdd_Vectors(BufferLength: 3079), BenchmarksGame.KNucleotide_1.RunBench, Benchstone.BenchI.BubbleSort2.Test (advisory — not applied)

   This critique is advisory in the current configuration and did not change which tests were bisected.

   🦆 Re-analysis (advisory)

   ⚠️ Disputes the analysis — advisory only, not applied

   Flags: overconfident, weak-bisection-evidence

   Rationale: The overall Improvement Confidence of 5/5 is not justified because every per-finding change confidence is only 3/5 and six of seven bisections were inconclusive; only the Utf8JsonReader benchmark converged. The culprit 23ec3d9 is acceptable since it is the sole SHA the bisection actually surfaced (relevance RELATED), so it is not disputed, but the change confidence should be lowered to match the single-benchmark evidence.

   Suggested confidence: Change Confidence → 3/5 (advisory — not applied)

   This critique is advisory in the current configuration and did not change the confidences, culprit, or proposed actions above.

   ---

   Generated by PerfTriageAgent on 2026-08-13 12:02 UTC