[Perf] Windows/x64: 5 Regressions on 7/17/2026 12:15:56 AM +00:00

[performanceautofiler]

Run Information

Name	Value
Architecture	x64
OS	Windows 10.0.22631
Queue	ViperWindows
Baseline	b78a8f8b1d15898ba39eb4fd63e2578edcf16c37
Compare	09f41d9272e6dcff30bfa15653e124d013167e05
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Runtime.Intrinsics.Tests.Perf_Vector128Of<SByte>

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
SquareRootBenchmark - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	26.78 ns	29.27 ns	1.09	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
py .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Runtime.Intrinsics.Tests.Perf_Vector128Of<SByte>*'

Details

System.Runtime.Intrinsics.Tests.Perf_Vector128Of<SByte>.SquareRootBenchmark

ETL Files

Histogram

JIT Disasms

Docs

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

---

Run Information

Name	Value
Architecture	x64
OS	Windows 10.0.22631
Queue	ViperWindows
Baseline	b78a8f8b1d15898ba39eb4fd63e2578edcf16c37
Compare	09f41d9272e6dcff30bfa15653e124d013167e05
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Runtime.Intrinsics.Tests.Perf_Vector128Of<Byte>

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
SquareRootBenchmark - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	26.85 ns	29.64 ns	1.10	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
py .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Runtime.Intrinsics.Tests.Perf_Vector128Of<Byte>*'

Details

System.Runtime.Intrinsics.Tests.Perf_Vector128Of<Byte>.SquareRootBenchmark

ETL Files

Histogram

JIT Disasms

Docs

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

---

Run Information

Name	Value
Architecture	x64
OS	Windows 10.0.22631
Queue	ViperWindows
Baseline	b78a8f8b1d15898ba39eb4fd63e2578edcf16c37
Compare	09f41d9272e6dcff30bfa15653e124d013167e05
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Runtime.Intrinsics.Tests.Perf_Vector128Of<Int16>

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
SquareRootBenchmark - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	20.87 ns	22.48 ns	1.08	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
py .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Runtime.Intrinsics.Tests.Perf_Vector128Of<Int16>*'

Details

System.Runtime.Intrinsics.Tests.Perf_Vector128Of<Int16>.SquareRootBenchmark

ETL Files

Histogram

JIT Disasms

Docs

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

---

Run Information

Name	Value
Architecture	x64
OS	Windows 10.0.22631
Queue	ViperWindows
Baseline	b78a8f8b1d15898ba39eb4fd63e2578edcf16c37
Compare	09f41d9272e6dcff30bfa15653e124d013167e05
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Numerics.Tests.Perf_VectorOf<UInt16>

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
SquareRootBenchmark - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	22.67 ns	25.16 ns	1.11	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
py .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Numerics.Tests.Perf_VectorOf<UInt16>*'

Details

System.Numerics.Tests.Perf_VectorOf<UInt16>.SquareRootBenchmark

ETL Files

Histogram

JIT Disasms

Docs

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

---

Run Information

Name	Value
Architecture	x64
OS	Windows 10.0.22631
Queue	ViperWindows
Baseline	b78a8f8b1d15898ba39eb4fd63e2578edcf16c37
Compare	09f41d9272e6dcff30bfa15653e124d013167e05
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Numerics.Tests.Perf_VectorOf<Int16>

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
SquareRootBenchmark - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	22.48 ns	25.34 ns	1.13	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
py .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Numerics.Tests.Perf_VectorOf<Int16>*'

Details

System.Numerics.Tests.Perf_VectorOf<Int16>.SquareRootBenchmark

ETL Files

Histogram

JIT Disasms

Docs

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

[LoopedBard3]

🔍 Automated Triage Analysis

Summary: Verified regression in SquareRootBenchmark — caused by commit d61e485 ("JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integr"). 5 tests bisected to related culprit commits.

Finding Confidence: 5/5 (analysis accuracy)
Regression Confidence: 5/5 (likelihood of true regression)

Likely Cause: d61e4857b5b9822ade487780da34141cb98792ef

🤖 Proposed Next Actions: Known-cause regression — transfer & notify — 0 done, 3 dry-run (rule R4_known_regression_new — see Full Analysis for the table). · 🦆 Re-analysis: ✅ concurs

📊 Full Analysis (click to expand)

Summary: All 5 SquareRoot vector benchmarks on small integral element types (SByte, Byte, Int16, UInt16) regressed +7.7–12.7% as a clean step change, attributed with high confidence to commit d61e4857 (#128604), which added saturating float→small-integral conversion clamping in the CoreCLR JIT (morph.cpp). Finding confidence 5/5 (300 prior runs, z=13–90, tight post-change cluster); regression confidence 5/5 (surgical change directly on the affected code path, and only the small integral types the commit touches regressed).

Issue Overview

• Issue: [Perf] Windows/x64: 5 Regressions on 7/17/2026 12:15:56 AM +00:00 #131158 - [Perf] Windows/x64: 5 Regressions on 7/17/2026 12:15:56 AM +00:00
• Date: 2026-07-17 | Queue: Windows.11.Amd64.Viper.Perf | OS/Arch: Windows 10.0.22631/x64
• Commits: b78a8f8b → 09f41d92 (40 commits)
• Labels: untriaged, perf-regression, os-windows, arch-x64, branch-refs/heads/main, runtime-coreclr, kind-micro, compilationmode-tiered, runkind-micro
• Runtime stack: Desktop + CoreCLR (x64) — confirmed by runtime-coreclr label and CoreClr:true run-config.
• Historical Data: Available (300 prior runs per test, 20 after; excellent density)

Test Analysis

Test Name	Baseline	Compare	Δ%	μ	σ	n	z	Thresh%	Noise?	Assessment
Perf_Vector128Of<SByte>.SquareRootBenchmark	26.778	29.270	+9.31%	26.983	0.176	300	13.0	5%	N	Step change, stable
Perf_Vector128Of<Byte>.SquareRootBenchmark	26.854	29.637	+10.36%	26.954	0.095	300	28.2	5%	N	Step change, stable
Perf_Vector128Of<Int16>.SquareRootBenchmark	20.868	22.484	+7.74%	20.854	0.018	300	90.6	5%	N	Step change, rock-solid history
Perf_VectorOf<UInt16>.SquareRootBenchmark	22.670	25.160	+10.98%	22.398	0.121	300	22.8	5%	N	Step change, stable
Perf_VectorOf<Int16>.SquareRootBenchmark	22.480	25.339	+12.72%	22.757	2.275	300	1.14	19.99%	N (Borderline)	Step change; z/threshold distorted by 2 outliers (~50) in history

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

Finding Confidence: 5/5 - 300 prior runs per test, chronologically ordered with no gaps; four tests show extreme z-scores (13–90) with tight post-change clusters. The fifth (Vector<Int16>) is statistically "Borderline" only because two infrastructure outliers (~50ns, 2.2× the ~22.5 cluster) inflate σ to 2.275; excluding them, it is a clean +12% step matching its siblings.

Candidate Stack Relevance

Commit	Touched Paths (summary)	Relevance	Notes
d61e4857	src/coreclr/jit/morph.cpp, gentree.cpp, valuenum.cpp, importercalls.cpp, assertionprop.cpp	Relevant	CoreCLR JIT change to float→small-integral cast expansion (adds saturating min/max clamp on HW-intrinsics targets incl. x64). Directly on the codegen path for integer-element vector Sqrt.
56946bbc	src/coreclr/jit/* (Vector<T> size mismatch)	Relevant	CoreCLR JIT, but scoped to cross-targeting Vector<T> size mismatch handling; not on the Sqrt/conversion path.
475838a7	src/coreclr/jit/* (GenTreeCall::Equals)	Relevant	JIT call-node equality/CSE change; no plausible effect on vector Sqrt lowering.
34d8d669	src/mono/, WASM SIMD intrinsics	LessLikely	WASM/Mono SIMD light-up; not exercised on x64 CoreCLR.
6c68540143	src/coreclr/* WASM genCodeForConstant	LessLikely	WASM-specific codegen; not exercised on x64.

Confirmed Regressions

The five tests share one root cause, so they are grouped. Representative test: Perf_VectorOf<Int16>.SquareRootBenchmark (+12.72%).

System.Runtime.Intrinsics.Tests.Perf_Vector128Of<SByte / Byte / Int16>.SquareRootBenchmark and System.Numerics.Tests.Perf_VectorOf<UInt16 / Int16>.SquareRootBenchmark

• Delta: +7.74% to +12.72% across the five small-integral element types (see Test Analysis table)
• Historical Context: μ within ~1% of baseline for all; z=13–90 for four tests; adaptive threshold 5% (19.99% for the outlier-contaminated Vector<Int16>, whose true cluster CV is <0.5%)
• Assessment: Step change — every test jumps to a new, tightly clustered performance tier entirely above its historical max, at the same build.
• Finding Confidence: 5/5 - comprehensive, stable history; consistent magnitude and timing across all five independent tests.
• Regression Confidence: 5/5 - surgical change on the exact code path; only the small integral element types the commit touches (sbyte/byte/short/ushort) regressed.
• Likely Cause: d61e4857 - "JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types (JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types #128604)"
  • Affected Areas: JIT / SIMD codegen (fgMorphExpandCast, float→small-int conversion)
  • Stack Relevance: Relevant
  • Rationale: Vector128.Sqrt/Vector.Sqrt on integer element types is expanded by the JIT as int→float, sqrt, then float→small-integral conversion. This commit changed fgMorphExpandCast in src/coreclr/jit/morph.cpp so that, on FEATURE_HW_INTRINSICS targets (which includes x64), the R → smallT conversion now performs float-domain min/max clamping to saturate at the destination bounds before the cast. That inserts extra min/max + compare instructions into precisely the float→smallint step these benchmarks exercise, explaining a uniform ~8–13% slowdown. The change is scoped to sbyte/byte/short/ushort — exactly the five regressed element types — while float/double vector Sqrt (no small-int conversion) is unaffected. The commit was authored 2026-07-16 23:38 UTC, ~37 minutes before the regression run, and lies within the range.
  • Candidates Considered:
    • 56946bbc - JIT Vector<T> size mismatch; unrelated to Sqrt/conversion lowering.
    • 475838a7 - JIT GenTreeCall::Equals/CSE; no vector-Sqrt effect.
    • 34d8d669 - WASM/Mono SIMD light-up; not exercised on x64 CoreCLR.
• Verification Status: Not verified (mechanism-based attribution). Given the surgical fit, verification is optional; a confirmatory RunWithBuildAtHash at d61e485 vs. its parent 0bc487c would raise finding confidence to a bisected 5/5.

Open Questions

• Historical Data: Available and abundant (300 runs/test). One caveat: two ~50ns outliers in the Vector<Int16> history inflate its σ and mask the z-score; the underlying step is clearly real.
• Confidence Limitations: None material. Attribution is single-commit and mechanism-backed; the 40-commit range contains only one change touching the relevant float→small-int conversion path.
• Verification Needed: Optional confirmatory rerun/bisection at d61e485 vs parent 0bc487c to formally close the loop.
• Bisection Status: Not performed (single high-confidence candidate identified).

Recommended Actions

1. Confirm root cause - Run RunWithBuildAtHash on d61e4857 vs parent 0bc487c7 for the five SquareRoot benchmarks to verify the delta reproduces.
2. Source review - Review the FEATURE_HW_INTRINSICS float-domain min/max clamping path added in fgMorphExpandCast (src/coreclr/jit/morph.cpp) to assess whether the saturation clamp for integer-element vector Sqrt can be emitted more cheaply (e.g., avoid redundant clamps when the value range is provably in-bounds).
3. Correctness vs. perf tradeoff - This is a correctness fix (saturating conversion semantics); coordinate with the JIT/JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types #128604 authors on whether the added cost is acceptable or optimizable, rather than reverting.

Bisection Results

Test	Status	Culprit Commit	Relevance	Reported Diff (issue)	Bisected Diff (at culprit)	Duration
SquareRootBenchmark	✅ Culprit found	d61e4857b5b9	✅ RELATED	26.8→29.3 ns (+9%)	26.7→29.3 ns (+9%)	-
SquareRootBenchmark	✅ Culprit found	d61e4857b5b9	✅ RELATED	26.9→29.6 ns (+10%)	26.8→29.6 ns (+10%)	-
SquareRootBenchmark	✅ Culprit found	d61e4857b5b9	✅ RELATED	20.9→22.5 ns (+8%)	20.8→22.5 ns (+8%)	-
SquareRootBenchmark	✅ Culprit found	d61e4857b5b9	✅ RELATED	22.7→25.2 ns (+11%)	22.4→25.2 ns (+13%)	-
SquareRootBenchmark	✅ Culprit found	d61e4857b5b9	✅ RELATED	22.5→25.3 ns (+13%)	22.5→25.0 ns (+11%)	2176s

Total bisection time: 2176s
Success rate: 5/5 tests

System.Runtime.Intrinsics.Tests.Perf_Vector128Of.SquareRootBenchmark

• Culprit: d61e4857b5b9
• Commit Message: JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types (JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types #128604)
• Relevance: RELATED
• Explanation: The commit adds saturating clamp logic to the JIT's float/double → small-integral-type cast expansion (fgMorphExpandCast in morph.cpp), using float-domain min/max clamping before cast expansion on HW_INTRINSICS targets (x64). The benchmark operates on Vector128 (an sbyte small-integral type) and its SquareRoot path performs exactly these float→small-int conversions, so the added clamp instructions land directly in the hot path. The +9.3% regression and its precise magnitude match are fully consistent with the extra per-element saturation work.

Helix jobs

Job ID	Submitted by bisection	Last status
de55ccac-569c-4848-93cc-2f4c559f330d	✅	submitted
a8481141-b6cc-49a0-9d3e-601a6b666f7e	✅	submitted
f37b3d0b-fb00-4068-8905-764bcff6df2d	✅	completed

Per-commit measurements

Commit	Position	Measurements
b78a8f8b1d15	baseline / range start	26.781 ns
9221c0834e45	idx 10	26.775 ns
15c8879b23d3	idx 20	26.775 ns
bf44912e0d7b	idx 22	26.791 ns
0bc487c79352	idx 26, last good	26.748 ns
d61e4857b5b9	idx 27, culprit	29.251 ns
475838a780ea	idx 28	29.316 ns
3a7bba7af4ac	idx 30	29.279 ns
09f41d9272e6	head	29.232 ns

Bisection details

…[earlier output omitted]

- Baseline Value: 26.748 ns
- Culprit Value: 29.251 ns
- Summary: Double-bisected the 40-commit range on the default tiered-JIT x64/Windows config; good measurements clustered at ~26.75 ns and regressed ones at ~29.25 ns. Narrowed to a single transition: last-good 0bc487c7 (26.748 ns) → culprit d61e4857 (29.251 ns).

HELIX JOBS:
- de55ccac-569c-4848-93cc-2f4c559f330d
- a8481141-b6cc-49a0-9d3e-601a6b666f7e
- f37b3d0b-fb00-4068-8905-764bcff6df2d

MEASUREMENTS:
- b78a8f8b1d15 (baseline / range start): 26.781 ns
- 9221c0834e45 (idx 10): 26.775 ns
- 15c8879b23d3 (idx 20): 26.775 ns
- bf44912e0d7b (idx 22): 26.791 ns
- 0bc487c79352 (idx 26, last good): 26.748 ns
- d61e4857b5b9 (idx 27, culprit): 29.251 ns
- 475838a780ea (idx 28): 29.316 ns
- 3a7bba7af4ac (idx 30): 29.279 ns
- 09f41d9272e6 (head): 29.232 ns

Notes: default-JIT config matched the reported baseline/compare exactly (no run-config mismatch); the ~9.3% step is ~7σ above the 2σ noise floor, so single-sample measurements were decisive. Cleaning up temp files.

● Remove temp output files (shell)
│ rm -f /tmp/1784630570000-copilot-tool-output-456c11.txt
│ /tmp/1784631219813-copilot-tool-output-4f5667.txt; echo cleaned
└ 2 lines…

Bisection succeeded. Culprit: d61e4857 — "JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types (#128604)", RELATED with Confidence 5. Last-good 0bc487c7 = 26.748 ns → culprit = 29.251 ns (+9.3%), matching the reported regression exactly. The added JIT saturation clamping for float→small-integral casts lands directly in the Vector128<SByte> SquareRoot hot path.

System.Runtime.Intrinsics.Tests.Perf_Vector128Of.SquareRootBenchmark

• Culprit: d61e4857b5b9
• Commit Message: JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types (JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types #128604)
• Relevance: RELATED
• Explanation: The benchmark is Vector128.Sqrt(Value1) with T=Byte; the software fallback converts each byte to float/double, computes the square root, and converts the float result back to byte. This PR modifies the JIT's fgMorphExpandCast to add float-domain min/max clamping for float/double → small-integral (byte) casts on FEATURE_HW_INTRINSICS (x64) targets, injecting extra clamp instructions into exactly the hot per-element float→byte conversion path this test exercises — a clear ~10% mechanism.

Helix jobs

Job ID	Submitted by bisection	Last status
bbc27737-def2-4e72-906b-e9dc38632899	✅	submitted
3d0c11a7-085d-4d05-b60b-a8840085e0e7	✅	submitted
c45b7868-c6e2-4a8c-b788-7b64c71ebcf0	✅	submitted
86f34b2e-23a9-4bc9-8a26-6d8a4c5c5609	✅	submitted

Per-commit measurements

Commit	Position	Measurements
(baseline)	baseline	26.844 ns
b78a8f8b1d15	baseline / parent of range	26.854 ns
4c9e771cbd5e	idx 7, good	26.829 ns
14a226f36172	idx 14, good	26.847 ns
7c1cb8af886a	idx 21, good	26.806 ns
0bc487c79352	idx 26, good / last-good parent	26.822 ns
d61e4857b5b9	idx 27, culprit	29.619 ns
475838a780ea	idx 28, regressed	29.622 ns
34d8d6692375	idx 29, regressed	29.562 ns
09f41d9272e6	head / compare, regressed	29.617 ns

Bisection details

…[earlier output omitted]

: JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types (#128604)
- Commits Tested: 11
- Confidence: 5
- Commit Relevance: RELATED
- Relevance Explanation: The benchmark is Vector128.Sqrt(Value1) with T=Byte; the software fallback converts each byte to float/double, computes the square root, and converts the float result back to byte. This PR modifies the JIT's fgMorphExpandCast to add float-domain min/max clamping for float/double → small-integral (byte) casts on FEATURE_HW_INTRINSICS (x64) targets, injecting extra clamp instructions into exactly the hot per-element float→byte conversion path this test exercises — a clear ~10% mechanism.
- Baseline Value: 26.844 ns (measured at last-good parent commit 0bc487c7)
- Culprit Value: 29.619 ns
- Summary: Bisected the 40-commit range with 4 multi-commit Helix jobs on the default tiered-JIT config. Endpoints reproduced the reported baseline (26.85) and compare (29.64) exactly; narrowing (idx 29 bad → idx 7/14/21 good → idx 24/26 good, 28 bad → idx 27 bad) isolated the step to d61e4857, whose parent 0bc487c7 measured 26.84 while d61e4857 measured 29.62. The ~2.8 ns step is ~14× the 2σ noise floor (0.19 ns) and was reproduced at adjacent commits, so single-sample bisection is unambiguous.

HELIX JOBS:
- bbc27737-def2-4e72-906b-e9dc38632899
- 3d0c11a7-085d-4d05-b60b-a8840085e0e7
- c45b7868-c6e2-4a8c-b788-7b64c71ebcf0
- 86f34b2e-23a9-4bc9-8a26-6d8a4c5c5609

MEASUREMENTS:
- b78a8f8b1d15 (baseline / parent of range): 26.854 ns
- 4c9e771cbd5e (idx 7, good): 26.829 ns
- 14a226f36172 (idx 14, good): 26.847 ns
- 7c1cb8af886a (idx 21, good): 26.806 ns
- 2f7ae92e105c (idx 24): no valid measurement (result dropped)
- 0bc487c79352 (idx 26, good / last-good parent): 26.822 ns; re-run 26.844 ns
- d61e4857b5b9 (idx 27, CULPRIT): 29.619 ns
- 475838a780ea (idx 28, regressed): 29.622 ns; re-run 29.570 ns
- 34d8d6692375 (idx 29, regressed): 29.562 ns
- 09f41d9272e6 (head / compare, regressed): 29.617 ns

System.Runtime.Intrinsics.Tests.Perf_Vector128Of.SquareRootBenchmark

• Culprit: d61e4857b5b9
• Commit Message: JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types (JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types #128604)
• Relevance: RELATED
• Explanation: The commit changes the JIT (morph.cpp) expansion of float/double → small integral (short) conversions, adding float-domain min/max saturation clamping before cast expansion on FEATURE_HW_INTRINSICS (x64) targets. The Perf_Vector128Of.SquareRootBenchmark exercises exactly this path (square root produces floating values narrowed back to Int16), so the added clamp instructions land directly in the hot loop, a clear and plausible mechanism for the +7.7% slowdown.

Helix jobs

Job ID	Submitted by bisection	Last status
c63a68e5-97d9-4720-8308-5dc28208a180	✅	submitted
0213918c-7254-4d70-b76b-78f5f73f6dc4	✅	submitted
d5bb72aa-be97-4770-8d9b-f1ce353e60dd	✅	submitted

Per-commit measurements

Commit	Position	Measurements
b78a8f8b1d15	baseline	20.860 ns
6bd973a54170	idx10, good	20.823 ns
d0eee82ec66a	idx20, good	20.825 ns
34d8d6692375	idx30, bad	22.475 ns
09f41d9272e6	compare/head, bad	22.548 ns
0bc487c79352	last-good before culprit	20.843 ns
2f7ae92e105c	good	20.843 ns
7c1cb8af886a	good	20.863 ns
d61e4857b5b9	culprit	22.488 ns
e0e5f8304a80	good	20.870 ns

Bisection details

…[earlier output omitted]

57…
└ {"index":null,"sha":"d61e4857b5b9822ade487780da34141cb98792ef","abbrevSha":"d...

Bisection complete. The culprit is d61e48, a JIT change to float/double→small-integral saturating conversions that directly affects the Vector128 SquareRoot code path (float→short casts). Clean separation, far above the noise floor.

BISECTION RESULT:
- Culprit Commit: d61e4857b5b9822ade487780da34141cb98792ef
- Commit Message: JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types (#128604)
- Commits Tested: 10
- Confidence: 5
- Commit Relevance: RELATED
- Relevance Explanation: The commit changes the JIT (morph.cpp) expansion of float/double → small integral (short) conversions, adding float-domain min/max saturation clamping before cast expansion on FEATURE_HW_INTRINSICS (x64) targets. The Perf_Vector128Of<Int16>.SquareRootBenchmark exercises exactly this path (square root produces floating values narrowed back to Int16), so the added clamp instructions land directly in the hot loop, a clear and plausible mechanism for the +7.7% slowdown.
- Baseline Value: 20.843 ns
- Culprit Value: 22.488 ns
- Summary: Double-bisection over the 40-commit range located a single clean, monotonic step from ~20.84 ns (good) to ~22.49 ns (bad). The last-good commit 0bc487 measured 20.843 ns and its immediate child d61e48 measured 22.488 ns; the ~1.6 ns jump is ~45× the 2σ noise floor (0.036 ns), so single samples suffice.

HELIX JOBS:
- c63a68e5-97d9-4720-8308-5dc28208a180
- 0213918c-7254-4d70-b76b-78f5f73f6dc4
- d5bb72aa-be97-4770-8d9b-f1ce353e60dd

MEASUREMENTS:
- b78a8f8b1d15 (baseline): 20.860 ns
- 6bd973a54170 (idx10, good): 20.823 ns
- d0eee82ec66a (idx20, good): 20.825 ns
- 7c1cb8af886a (good): 20.863 ns
- 2f7ae92e105c (good): 20.843 ns
- e0e5f8304a80 (good): 20.870 ns
- 0bc487c79352 (last-good before culprit): 20.843 ns
- d61e4857b5b9 (culprit): 22.488 ns
- 34d8d6692375 (idx30, bad): 22.475 ns
- 09f41d9272e6 (compare/head, bad): 22.548 ns

System.Numerics.Tests.Perf_VectorOf.SquareRootBenchmark

• Culprit: d61e4857b5b9
• Commit Message: JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types (JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types #128604)
• Relevance: RELATED
• Explanation: The commit modifies the JIT's fgMorphExpandCast (morph.cpp) to add float-domain min/max saturation clamping for float/double → small integral type conversions (byte, ushort, etc.) on FEATURE_HW_INTRINSICS targets like x64. Perf_VectorOf.SquareRootBenchmark computes Vector square roots, whose implementation converts float sqrt results back to ushort — exactly the float→small-integral cast path this change makes more expensive by inserting extra clamp instructions in the hot loop. The mechanism (added clamping in a vectorized conversion hot path) plausibly and directly explains the +11% regression.

Helix jobs

Job ID	Submitted by bisection	Last status
b938668f-b167-472b-a949-d2781c869f7e	✅	submitted
6045935e-ba3a-45bc-b87d-93dfb69f2282	✅	submitted
83fb8d29-b3c3-4454-9b4c-23011dc9a452	✅	submitted

Per-commit measurements

Commit	Position	Measurements
b78a8f8b	baseline	22.743 ns
9221c083	idx10	22.338 ns
15c8879b	idx20	22.928 ns
bf44912e	idx23	22.317 ns
e0e5f830	idx25	22.297, 22.352 ns
0bc487c7	idx26, last good	22.373 ns
d61e4857b5b9	idx27 re-run	25.241, 25.178 ns
3a7bba7a	idx30	25.156 ns
09f41d92	head	25.165 ns

Bisection details

…[earlier output omitted]

- Commits Tested: 9
- Confidence: 5
- Commit Relevance: RELATED
- Relevance Explanation: The commit modifies the JIT's fgMorphExpandCast (morph.cpp) to add float-domain min/max saturation clamping for float/double → small integral type conversions (byte, ushort, etc.) on FEATURE_HW_INTRINSICS targets like x64. Perf_VectorOf<UInt16>.SquareRootBenchmark computes Vector<ushort> square roots, whose implementation converts float sqrt results back to ushort — exactly the float→small-integral cast path this change makes more expensive by inserting extra clamp instructions in the hot loop. The mechanism (added clamping in a vectorized conversion hot path) plausibly and directly explains the +11% regression.
- Baseline Value: 22.373 ns (last-good commit 0bc487c7)
- Culprit Value: 25.241 ns (culprit commit d61e4857; also 25.178 ns in a prior job)
- Summary: A double-bisection over the 40-commit range localized a single clean step change (~22.4 ns → ~25.2 ns, +12%, far above the 0.242 ns 2σ noise floor). Baseline/mid (idx20) measured good, idx30/head measured bad; successive Helix jobs narrowed to idx26 (0bc487c7, 22.37 ns ✓) good and idx27 (d61e4857, 25.18/25.24 ns ✗) bad, with the boundary re-confirmed in a second job. Absolute values match the reported baseline (22.67) and compare (25.16) under the default JIT config, so no run-config mismatch applies.

HELIX JOBS:
- b938668f-b167-472b-a949-d2781c869f7e
- 6045935e-ba3a-45bc-b87d-93dfb69f2282
- 83fb8d29-b3c3-4454-9b4c-23011dc9a452

MEASUREMENTS:
- b78a8f8b (baseline): 22.743 ns
- 9221c083 (idx10): 22.338 ns
- 15c8879b (idx20): 22.928 ns
- bf44912e (idx23): 22.317 ns
- e0e5f830 (idx25): 22.297 ns
- e0e5f830 (idx25 re-run): 22.352 ns
- 0bc487c7 (idx26, last good): 22.373 ns
- d61e4857 (idx27, culprit): 25.178 ns
- d61e4857 (idx27 re-run): 25.241 ns
- 3a7bba7a (idx30): 25.156 ns
- 09f41d92 (head): 25.165 ns

Bisection succeeded via three successful Helix jobs; no failures encountered.

System.Numerics.Tests.Perf_VectorOf.SquareRootBenchmark

• Culprit: d61e4857b5b9
• Commit Message: JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types (JIT, AOT preinit, and CoreLib: Saturate float/double conversions to small integral types #128604)
• Relevance: RELATED
• Explanation: The commit modifies JIT morph.cpp (fgMorphExpandCast) to add float-domain min/max clamping before every unchecked float/double → small integral (short/Int16) cast on FEATURE_HW_INTRINSICS (x64) targets. The Perf_VectorOf.SquareRootBenchmark computes a square root over a Vector, which internally converts to floating point, takes sqrt, and converts back to Int16 — landing squarely on the newly added saturating-conversion code path. The extra clamp instructions injected into this hot loop are a plausible, direct mechanism for the ~+11-12% slowdown.

Helix jobs

Job ID	Submitted by bisection	Last status
8a48efe6-d28e-4153-b0ce-e5f979dcddb4	✅	submitted
6cfccc75-fa92-4c78-af7b-ba21538d504b	✅	submitted
ecf6c6fa-c526-41ff-9006-fdf8203f4b45	✅	submitted

Per-commit measurements

Commit	Position	Measurements
b78a8f8b1d15	baseline, good	22.640 ns
9221c0834e45	idx10, good	22.880 ns
15c8879b23d3	idx20, good	22.570 ns
e0e5f8304a80	idx25, good	22.920 ns
0bc487c79352	idx26, last-good	22.480 ns
475838a780ea	idx28, bad	25.040 ns
34d8d6692375	idx29, bad	25.280 ns
3a7bba7af4ac	idx30, bad	25.000 ns
874b9e9eea4a	idx35, bad	25.320 ns
09f41d9272e6	head, bad	25.080 ns
d61e4857b5b9	idx27, culprit / first-bad	25.000 ns

Bisection details

…[earlier output omitted]

Lib: Saturate float/double conversions to small integral types (#128604)
- Commits Tested: 11
- Confidence: 3
- Commit Relevance: RELATED
- Relevance Explanation: The commit modifies JIT `morph.cpp` (fgMorphExpandCast) to add float-domain min/max clamping before every unchecked float/double → small integral (short/Int16) cast on FEATURE_HW_INTRINSICS (x64) targets. The Perf_VectorOf<Int16>.SquareRootBenchmark computes a square root over a Vector<Int16>, which internally converts to floating point, takes sqrt, and converts back to Int16 — landing squarely on the newly added saturating-conversion code path. The extra clamp instructions injected into this hot loop are a plausible, direct mechanism for the ~+11-12% slowdown.
- Baseline Value: 22.48 ns
- Culprit Value: 25.00 ns
- Summary: Ran double-bisection over the 40-commit range via Helix (default tiered-JIT x64 config). Good commits (baseline, idx10, idx20, idx25, idx26) clustered at 22.5–22.9 ns; every commit from idx27 (d61e485) onward clustered at 25.0–25.3 ns, isolating a clean, reproducible step at d61e485.

HELIX JOBS:
- 8a48efe6-d28e-4153-b0ce-e5f979dcddb4
- 6cfccc75-fa92-4c78-af7b-ba21538d504b
- ecf6c6fa-c526-41ff-9006-fdf8203f4b45

MEASUREMENTS:
- b78a8f8b1d15 (baseline, good): 22.64 ns
- 9221c0834e45 (idx10, good): 22.88 ns
- 15c8879b23d3 (idx20, good): 22.57 ns
- e0e5f8304a80 (idx25, good): 22.92 ns
- 0bc487c79352 (idx26, last-good): 22.48 ns
- d61e4857b5b9 (idx27, culprit / first-bad): 25.00 ns
- 475838a780ea (idx28, bad): 25.04 ns
- 34d8d6692375 (idx29, bad): 25.28 ns
- 3a7bba7af4ac (idx30, bad): 25.00 ns
- 874b9e9eea4a (idx35, bad): 25.32 ns
- 09f41d9272e6 (head, bad): 25.08 ns

Note on confidence: the reported regression delta (2.86 ns, +12.7%) is smaller than this benchmark's 2σ noise floor (≈4.55 ns), so per the noise-floor guidance confidence is capped at 3/5 even though the step is highly consistent across 6 good vs. 6 bad single-sample measurements and backed by a clear functional mechanism.

🤖 Proposed Next Actions

Rule: R4_known_regression_new — Known-cause regression — transfer & notify.

Verified regression attributed to commit d61e485 (finding 5/5, regression 5/5, bisection RELATED). No prior linked autofile issue found.

#	Action	Status	Details
1	Back-reference cause PR	⏭️ Skipped (dry-run)	on culprit PR for d61e4857b5b9
2	Transfer issue	⏭️ Skipped (dry-run)	→ dotnet/runtime
3	Ping owners	⏭️ Skipped (dry-run)	no owners resolved

🦆 Pre-bisection Re-analysis (advisory)

✅ Agrees with the pre-bisection plan — advisory only, not applied

Rationale: All five findings show noise=False with change confidence 5/5, relevant stacks, and a single consistent likely culprit, so the noise call and the candidate set are well-supported. No noisy test is wasting a bisection slot and no clearly-regressed test is missing from the queue.

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

🦆 Re-analysis (advisory)

✅ Agrees with the analysis — advisory only, not applied

Rationale: All five findings converged on the same culprit via successful bisection with RELATED relevance and no alignment warnings, and the noise=False call is well supported, so the high finding and change confidence are justified.

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

---

Generated by PerfTriageAgent on 2026-07-21 14:56 UTC

[DrewScoggins]

BIsected to #128604

Related regressions:

• [Perf] Linux/x64: 8 Regressions on 7/17/2026 12:15:56 AM +00:00 perf-autofiling-issues#76506

[dotnet-policy-service]

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

[tannergooding]

This is expected/by-design, was required for correctness.