Skip to content

[Perf] Windows/x64: BenchStone Regressions on 1/29/2025 7:05:56 PM +00:00 #112136

Closed
Closed
[Perf] Windows/x64: BenchStone Regressions on 1/29/2025 7:05:56 PM +00:00#112136
Assignees
amanasifkhalid
Labels
arch-x64area-CodeGen-coreclrCLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMIos-windowsruntime-coreclrspecific to the CoreCLR runtimetenet-performancePerformance related issuetenet-performance-benchmarksIssue from performance benchmark
Milestone
10.0.0

Description

@performanceautofiler
performanceautofiler
bot
opened on Feb 4, 2025

Run Information

Name Value
Architecture x64
OS Windows 10.0.22621
Queue TigerWindows
Baseline bfa02766bdde4657ddd1f734e274d43e6ba3c036
Compare f3d90b2f622691253225ddcd61b8ae8a087955ae
Diff Diff
Configs CompilationMode:tiered, RunKind:micro

Regressions in Benchstone.BenchI.Array2

Benchmark Baseline Test Test/Base Test Quality Edge Detector Baseline IR Compare IR IR Ratio
918.94 ms 1.28 secs 1.40 0.01 False

graph
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 'Benchstone.BenchI.Array2*'
Details

Benchstone.BenchI.Array2.Test

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 10.0.22621
Queue TigerWindows
Baseline bfa02766bdde4657ddd1f734e274d43e6ba3c036
Compare f3d90b2f622691253225ddcd61b8ae8a087955ae
Diff Diff
Configs CompilationMode:tiered, RunKind:micro

Regressions in Benchstone.BenchI.NDhrystone

Benchmark Baseline Test Test/Base Test Quality Edge Detector Baseline IR Compare IR IR Ratio
326.78 ms 361.55 ms 1.11 0.02 False

graph
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 'Benchstone.BenchI.NDhrystone*'
Details

Benchstone.BenchI.NDhrystone.Test

ETL Files

Histogram

JIT Disasms

Docs

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

Activity

  1. performanceautofiler
    added
    arch-x64
    os-windows
    runtime-coreclrspecific to the CoreCLR runtime
    untriagedNew issue has not been triaged by the area owner
    on Feb 4, 2025
  2. LoopedBard3
    transferred this issue fromdotnet/perf-autofiling-issueson Feb 4, 2025
  3. ghost added
    needs-area-labelAn area label is needed to ensure this gets routed to the appropriate area owners
    on Feb 4, 2025
  4. LoopedBard3
    changed the title [-][Perf] Windows/x64: 2 Regressions on 1/29/2025 7:05:56 PM +00:00[/-] [+][Perf] Windows/x64: BenchStone Regressions on 1/29/2025 7:05:56 PM +00:00[/+] on Feb 4, 2025
  5. LoopedBard3
    added
    tenet-performancePerformance related issue
    tenet-performance-benchmarksIssue from performance benchmark
    on Feb 4, 2025

  6. LoopedBard3 commented on Feb 4, 2025

    @LoopedBard3
    LoopedBard3
    on Feb 4, 2025
    Member

    Likely #111873, though only 2 regressions on this configuration. FYI @amanasifkhalid

    Related Regressions:

  7. amanasifkhalid
    added
    area-CodeGen-coreclrCLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI
    and removed
    untriagedNew issue has not been triaged by the area owner
    needs-area-labelAn area label is needed to ensure this gets routed to the appropriate area owners
    on Feb 4, 2025
  8. amanasifkhalid
    added this to the 10.0.0 milestone on Feb 4, 2025
  9. amanasifkhalid
    self-assigned this
    on Feb 4, 2025

  10. dotnet-policy-service commented on Feb 4, 2025

    @dotnet-policy-service
    dotnet-policy-serviceBot
    on Feb 4, 2025 – with Dotnet Policy Service
    Contributor

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

  11. amanasifkhalid commented on Mar 24, 2025

    @amanasifkhalid
    amanasifkhalid
    on Mar 24, 2025
    Contributor

    I took a look at the Benchstone regressions as part of #113108, and the key change in #111873 is that we're calculating fgCalledCount (the JIT's estimate of how many times a method was called before it tiered up) differently. If the profile data is already consistent (which is the case for the Benchstone benchmarks), the new method shouldn't produce different results, but small floating-point rounding errors can subtly change fgCalledCount, and thus change relative block weight computations. CSE is pretty sensitive to this, and it's a known issue that CSE mixes relative and absolute block weights in its cost computation, so trivial changes to how relative weights are computed can churn CSE quite a bit. I ran into this while working on #112151, and tried to quirk my way around it. I'll keep this issue open for now, and revisit it once #112151 is merged to see if those quirks improve any CSE-related regressions.

  12. amanasifkhalid commented on Jun 25, 2025

    @amanasifkhalid
    amanasifkhalid
    on Jun 25, 2025
    Contributor
    Notes Recent Score Orig Score Ampere Ubuntu Tiger Ubuntu Tiger Windows Viper Ubuntu Viper Windows Benchmark
    Fixed, regressed by #115119 1.73 2.03 1.73
    2.03    		 System.Text.Json.Document.Tests.Perf_DocumentParse.Parse(IsDataIndented: True, TestRandomAccess: True, TestCase: Json400KB)
    maybe #111023 1.71 1.12 1.71
    1.12    		 System.Text.Json.Node.Tests.Perf_ParseThenWrite.ParseThenWrite(IsDataIndented: True, TestCase: Json400KB)
    comment 1.12 1.15 1.11
    1.13    		 1.13
    1.18    		 Benchstone.BenchI.IniArray.Test
    comment 1.09 1.32 1.09
    1.32    		 System.Collections.ContainsFalse(Int32).ImmutableSortedSet(Size: 512)
    comment 1.09 1.09 1.09
    1.09    		 LinqBenchmarks.Where00LinqQueryX
    comment 1.07 1.09 1.07
    1.09    		 LinqBenchmarks.Where00LinqMethodX
    comment 1.06 1.08 1.06
    1.08    		 LinqBenchmarks.Where00ForX
    comment 1.06 1.38 1.05
    1.71    		 1.02
    1.15    		 1.12
    1.35    		 System.Collections.IterateFor(String).ImmutableList(Size: 512)
    comment 1.04 1.19 1.11
    1.10    		 0.84
    1.10    		 1.21
    1.40    		 Benchstone.BenchI.Array2.Test
    comment 1.03 1.37 0.98
    1.63    		 1.12
    1.32    		 1.00
    1.19    		 System.Collections.IterateFor(Int32).ImmutableList(Size: 512)
    1.02 1.11 1.02
    1.11    		 Benchstone.BenchI.NDhrystone.Test
    1.00 1.18 1.00
    1.18    		 System.Collections.IterateFor(Int32).IList(Size: 512)
    0.97 1.29 1.05
    1.60    		 0.90
    1.18    		 0.97
    1.14    		 System.Collections.IterateFor(Int32).ImmutableSortedSet(Size: 512)
    0.96 1.26 1.02
    1.66    		 0.94
    1.15    		 0.93
    1.17    		 0.96
    1.14    		 System.Collections.IterateFor(String).ImmutableSortedSet(Size: 512)

  13. amanasifkhalid commented on Jun 25, 2025

    @amanasifkhalid
    amanasifkhalid
    on Jun 25, 2025
    Contributor

    I took another look at the Benchstone regressions, and I'm no longer seeing meaningful diffs in Tier1. However, #111936 changed OSR codegen such that fgOptimizeUncondBranchToSimpleCond kicks in more often, making some loops unnatural and thus pessimizing downstream phases. Here are the metric diffs for Benchstone.BenchI.Array2.Test:

    - LoopsFoundDuringOpts                      : 4
- LoopAlignmentCandidates                   : 1
- HoistedExpressions                        : 3
- CseCount                                  : 4
+ LoopsFoundDuringOpts                      : 2
+ LoopAlignmentCandidates                   : 0
+ HoistedExpressions                        : 0
+ CseCount                                  : 3

    In local runs of these benchmarks, I don't always make it out of OSR, so this affects the actual benchmark runs. It might make sense to delay tail duplication (and the rest of the non-simplifying transformations in fgUpdateFlowGraph), but I'm not keen on opening that can of worms in .NET 10, considering the amount of additional loop cloning it may unlock.

  14. amanasifkhalid commented on Jun 25, 2025

    @amanasifkhalid
    amanasifkhalid
    on Jun 25, 2025
    Contributor

    For System.Collections.ContainsFalse(Int32).ImmutableSortedSet(Size: 512), we were previously underestimating fgCalledCount by about tenfold. The new calculation means the loop body's relative weight is much lower, such that 3-opt doesn't align it as well as it previously did. We previously had the following layout:

    ---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight     IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0029]  1                             1     3498 [???..???)-> BB02(1)                 (always)                     LIR IBC internal
BB08 [0024]  1       BB07                  4.18 14609 [015..016)-> BB12(1)                 (always)                     i LIR IBC bwd
BB12 [0028]  3       BB08,BB10,BB11        9.00 31468 [015..016)-> BB14(1)                 (always)                     i LIR IBC nullcheck bwd
BB14 [0007]  2       BB13,BB12             9.00 31468 [015..026)-> BB05(0),BB15(1)         ( cond )                     i LIR IBC internal bwd
BB15 [0004]  1       BB14                  9.00 31468 [028..02C)-> BB17(0.517),BB16(0.483) ( cond )                     i LIR IBC bwd
BB17 [0006]  1       BB15                  4.65 16282 [03A..048)-> BB02(1)                 (always)                     i LIR IBC hascall bwd-src r-tail
BB02 [0000]  3       BB16,BB17,BB01       18.99 66434 [000..00B)-> BB04(1),BB03(0)         ( cond )                     i LIR IBC bwd bwd-target
BB04 [0012]  1       BB02                 10.00 34966 [000..013)-> BB06(0.9),BB05(0.1)     ( cond )                     i LIR IBC bwd
BB06 [0002]  1       BB04                  9.00 31468 [015..015)-> BB13(0),BB07(1)         ( cond )                     i LIR IBC bwd
BB07 [0008]  1       BB06                  6.75 23614 [015..016)-> BB09(0.381),BB08(0.619) ( cond )                     i LIR IBC nullcheck bwd
BB09 [0025]  1       BB07                  2.57  9005 [015..016)-> BB11(0.885),BB10(0.115) ( cond )                     i LIR IBC bwd
BB11 [0027]  1       BB09                  2.28  7971 [015..016)-> BB12(1)                 (always)                     i LIR IBC bwd
BB10 [0026]  1       BB09                  0.30  1034 [015..016)-> BB12(1)                 (always)                     i LIR IBC bwd
BB16 [0005]  1       BB15                  4.34 15186 [02C..03A)-> BB02(1)                 (always)                     i LIR IBC hascall bwd-src r-tail
BB05 [0001]  2       BB14,BB04             1     3498 [013..015)                           (return)                     i LIR IBC
BB13 [0009]  1       BB06                  0        0 [???..???)-> BB14(1)                 (always)                     i LIR IBC rare internal hascall gcsafe bwd
BB03 [0011]  1       BB02                  0        0 [000..001)                           (throw )                     i LIR IBC rare hascall gcsafe bwd
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

    We now have this:

    ---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight     IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0029]  1                             1    48269 [???..???)-> BB02(1)                 (always)                     LIR IBC internal
BB02 [0000]  3       BB16,BB17,BB01        1.90 91721 [000..00B)-> BB04(1),BB03(0)         ( cond )                     i LIR IBC bwd bwd-target
BB04 [0012]  1       BB02                  1    48269 [000..013)-> BB06(0.9),BB05(0.0998)  ( cond )                     i LIR IBC bwd
BB06 [0002]  1       BB04                  0.90 43452 [015..015)-> BB13(0),BB07(1)         ( cond )                     i LIR IBC bwd
BB07 [0008]  1       BB06                  0.68 32606 [015..016)-> BB09(0.381),BB08(0.619) ( cond )                     i LIR IBC nullcheck bwd
BB08 [0024]  1       BB07                  0.42 20172 [015..016)-> BB12(1)                 (always)                     i LIR IBC bwd
BB12 [0028]  3       BB08,BB10,BB11        0.90 43452 [015..016)-> BB14(1)                 (always)                     i LIR IBC nullcheck bwd
BB14 [0007]  2       BB13,BB12             0.90 43452 [015..026)-> BB05(0),BB15(1)         ( cond )                     i LIR IBC internal bwd
BB15 [0004]  1       BB14                  0.90 43452 [028..02C)-> BB17(0.515),BB16(0.485) ( cond )                     i LIR IBC bwd
BB17 [0006]  1       BB15                  0.46 22384 [03A..048)-> BB02(1)                 (always)                     i LIR IBC hascall bwd-src r-tail
BB09 [0025]  1       BB07                  0.26 12434 [015..016)-> BB11(0.885),BB10(0.115) ( cond )                     i LIR IBC bwd
BB11 [0027]  1       BB09                  0.23 11006 [015..016)-> BB12(1)                 (always)                     i LIR IBC bwd
BB10 [0026]  1       BB09                  0.03  1428 [015..016)-> BB12(1)                 (always)                     i LIR IBC bwd
BB16 [0005]  1       BB15                  0.44 21068 [02C..03A)-> BB02(1)                 (always)                     i LIR IBC hascall bwd-src r-tail
BB05 [0001]  2       BB14,BB04             0.10  4817 [013..015)                           (return)                     i LIR IBC
BB13 [0009]  1       BB06                  0        0 [???..???)-> BB14(1)                 (always)                     i LIR IBC rare internal hascall gcsafe bwd
BB03 [0011]  1       BB02                  0        0 [000..001)                           (throw )                     i LIR IBC rare hascall gcsafe bwd
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

    Considering the regression is only on Tiger machines, I'm inclined to believe the JCC erratum is partially to blame for this.

  15. amanasifkhalid commented on Jun 25, 2025

    @amanasifkhalid
    amanasifkhalid
    on Jun 25, 2025
    Contributor

    The Linq benchmarks never make it out of OSR on my machine, and the more aggressive tail duplication after #111936 changes (pessimizes?) a few downstream opts:

    - RedundantBranchesEliminated               : 3
- JumpThreadingsPerformed                   : 7
- CseCount                                  : 3
+ RedundantBranchesEliminated               : 2
+ JumpThreadingsPerformed                   : 2
+ CseCount                                  : 4

    More tail duplication might mean fewer branches for RBO to consider eliminating, so this decrease doesn't look bad on its own. It's possible the additional CSE wasn't worth it, and is responsible for the (relatively small) regression in these benchmarks.

  16. amanasifkhalid commented on Jun 25, 2025

    @amanasifkhalid
    amanasifkhalid
    on Jun 25, 2025
    Contributor

    The diffs for System.Collections.IterateFor(String).ImmutableList(Size: 512) stem from profile incorporation computing slightly different edge likelihoods for a few conditions in an inlinee:

    ---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight       IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0015]  1                            3682k 3682304 [000..00E)-> BB03(0.559),BB02(0.441) ( cond )                     IBC
BB02 [0016]  1       BB01                 1623k 1623159 [00E..01B)                           (return)                     IBC
BB03 [0017]  1       BB01                 2059k 2059145 [01B..029)-> BB05(0.223),BB04(0.777) ( cond )                     IBC
BB04 [0018]  1       BB03                 1601k 1600916 [029..044)                           (return)                     IBC
BB05 [0019]  1       BB03                  458k  458229 [044..04B)                           (return)                     IBC
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

    vs

    ---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight       IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0015]  1                            3814k 3813888 [000..00E)-> BB03(0.562),BB02(0.438) ( cond )                     IBC
BB02 [0016]  1       BB01                 1670k 1669990 [00E..01B)                           (return)                     IBC
BB03 [0017]  1       BB01                 2144k 2143898 [01B..029)-> BB05(0.221),BB04(0.779) ( cond )                     IBC
BB04 [0018]  1       BB03                 1669k 1669387 [029..044)                           (return)                     IBC
BB05 [0019]  1       BB03                  475k  474511 [044..04B)                           (return)                     IBC
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

    This change is just enough to move a jump within a hot loop:

    ---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight       IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0000]  1                             1       7194 [000..013)-> BB17(0.00195),BB16(0.998)   ( cond )                     i LIR IBC
BB17 [0025]  2       BB01,BB12             1.00    7194 [028..02A)                           (return)                     i LIR IBC
BB08 [0016]  1       BB07                225.63 1623159 [013..014)-> BB12(1)                 (always)                     i LIR IBC hascall gcsafe bwd
BB12 [0020]  3       BB08,BB10,BB11      511.86 3682304 [013..028)-> BB16(0.998),BB17(0.00195)   ( cond )                     i LIR IBC nullcheck bwd
BB16 [0024]  2       BB01,BB12           511.86 3682304 [013..014)-> BB04(0),BB03(1)         ( cond )                     i LIR IBC nullcheck bwd
BB03 [0006]  1       BB16                511.86 3682304 [013..014)-> BB07(1),BB04(0)         ( cond )                     i LIR IBC bwd
BB07 [0013]  1       BB03                511.86 3682304 [013..014)-> BB09(0.559),BB08(0.441) ( cond )                     i LIR IBC bwd
BB09 [0017]  1       BB07                286.23 2059145 [013..014)-> BB11(0.223),BB10(0.777) ( cond )                     i LIR IBC bwd
BB10 [0018]  1       BB09                222.53 1600916 [013..014)-> BB12(1)                 (always)                     i LIR IBC hascall gcsafe bwd
BB11 [0019]  1       BB09                 63.70  458229 [013..014)-> BB12(1)                 (always)                     i LIR IBC nullcheck bwd
BB04 [0007]  2       BB03,BB16             0          0 [013..014)                           (throw )                     i LIR IBC rare gcsafe bwd
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

    vs

    ---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight       IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0000]  1                             1       7459 [000..013)-> BB17(0.00195),BB16(0.998)   ( cond )                     i LIR IBC
BB17 [0025]  2       BB01,BB12             1.00    7459 [028..02A)                           (return)                     i LIR IBC
BB11 [0019]  1       BB09                 63.62  474511 [013..014)-> BB12(1)                 (always)                     i LIR IBC nullcheck bwd
BB08 [0016]  1       BB07                223.89 1669990 [013..014)-> BB12(1)                 (always)                     i LIR IBC hascall gcsafe bwd
BB12 [0020]  3       BB08,BB10,BB11      511.31 3813888 [013..028)-> BB16(0.998),BB17(0.00195)   ( cond )                     i LIR IBC nullcheck bwd
BB16 [0024]  2       BB01,BB12           511.31 3813888 [013..014)-> BB04(0),BB03(1)         ( cond )                     i LIR IBC nullcheck bwd
BB03 [0006]  1       BB16                511.31 3813888 [013..014)-> BB07(1),BB04(0)         ( cond )                     i LIR IBC bwd
BB07 [0013]  1       BB03                511.31 3813888 [013..014)-> BB09(0.562),BB08(0.438) ( cond )                     i LIR IBC bwd
BB09 [0017]  1       BB07                287.42 2143898 [013..014)-> BB11(0.221),BB10(0.779) ( cond )                     i LIR IBC bwd
BB10 [0018]  1       BB09                223.81 1669387 [013..014)-> BB12(1)                 (always)                     i LIR IBC hascall gcsafe bwd
BB04 [0007]  2       BB03,BB16             0          0 [013..014)                           (throw )                     i LIR IBC rare gcsafe bwd
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

    I suspect the profile churn is just from execution count imprecision from run to run. As for the block layout changes, it's possible the conditional jump BB09 -> BB11 is incorrectly predicted to be taken if it is a backward jump.

    With a mainline JIT, the layout returns to the baseline:

    ---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BBnum BBid ref try hnd preds           weight        IBC [IL range]   [jump]                            [EH region]        [flags]
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
BB01 [0000]  1                             1       58704 [000..013)-> BB04(0.00198),BB20(0.998)   ( cond )                   i LIR IBC
BB04 [0003]  2       BB01,BB14             1.00    58704 [028..02A)                           (return)                   i LIR IBC
BB16 [0016]  1       BB13                222.58 13066437 [013..014)-> BB14(1)                 (always)                   i LIR IBC hascall gcsafe bwd
BB14 [0020]  3       BB16,BB18,BB19      504.81 29634560 [013..028)-> BB20(0.998),BB04(0.00198)   ( cond )                   i LIR IBC nullcheck bwd
BB20 [0023]  2       BB01,BB14           504.81 29634560 [013..014)-> BB08(0),BB07(1)         ( cond )                   i LIR IBC nullcheck bwd
BB07 [0006]  1       BB20                504.81 29634560 [013..014)-> BB13(1),BB08(0)         ( cond )                   i LIR IBC bwd
BB13 [0013]  1       BB07                504.81 29634560 [013..014)-> BB17(0.559),BB16(0.441) ( cond )                   i LIR IBC bwd
BB17 [0017]  1       BB13                282.23 16568123 [013..014)-> BB19(0.223),BB18(0.777) ( cond )                   i LIR IBC bwd
BB18 [0018]  1       BB17                219.37 12877676 [013..014)-> BB14(1)                 (always)                   i LIR IBC hascall gcsafe bwd
BB19 [0019]  1       BB17                 62.87  3690447 [013..014)-> BB14(1)                 (always)                   i LIR IBC nullcheck bwd
BB08 [0007]  2       BB07,BB20             0           0 [013..014)                           (throw )                   i LIR IBC rare gcsafe bwd
---------------------------------------------------------------------------------------------------------------------------------------------------------------------

    I suspect we're getting unlucky from run to run with the edge likelihoods; this would explain the spikiness of this benchmark's graph.

  17. amanasifkhalid commented on Jun 25, 2025

    @amanasifkhalid
    amanasifkhalid
    on Jun 25, 2025
    Contributor

    I couldn't find any meaningful diffs for the System.Text.Json.Document.Tests benchmarks. I suspect #111023 could be to blame.

  18. amanasifkhalid commented on Jun 25, 2025

    @amanasifkhalid
    amanasifkhalid
    on Jun 25, 2025
    Contributor

    I might try moving tail duplication later in the JIT to reduce its interference with loop finding, but I'm hesitant to introduce more loop-related churn into .NET 10, especially since I have inversion-related regressions to look at still. I think these regressions are small and limited enough in frequency to accept for now.

  19. amanasifkhalid
    closed this as completedon Jun 25, 2025
  20. github-actions
    locked and limited conversation to collaborators on Jul 26, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

arch-x64area-CodeGen-coreclrCLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMIos-windowsruntime-coreclrspecific to the CoreCLR runtimetenet-performancePerformance related issuetenet-performance-benchmarksIssue from performance benchmark

Type

No type

Projects

No projects

    Milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions