Skip to content

Capture child-only JIT disassembly in frozen native-parent replay #692

Closed

Description

@thomhurst

Summary

The frozen BenchmarkDotNet native-parent replay cannot currently retain JIT disassembly from only its benchmark child processes. Scoped dispatch timing controls have not identified a production defect; inspecting the compiled dispatch methods across the original executable images is the next bounded diagnostic, with all instrumented timings excluded.

Reproduction

Run .github/scripts/frozen_bdn_control.py against retained Performance run 43409168876 with --case scoped-dispatch --native-executor tools/FrozenNativeHost/bin/Release/net10.0/FrozenNativeHost.dll. There is no child-only JIT capture option. The existing frozen_scope_control.py --jit-disassembly selects ScopeOnly/ScopeAndHandlerResolution and uses a console host, so it cannot capture this original DirectScopedDispatch native-parent runner.

Proposal

Reuse .NET's native DOTNET_JitDisasm, DOTNET_JitDisasmDiffable and DOTNET_JitStdOutFile settings and the existing child-wrapper environment mechanism. Add an opt-in method filter to the frozen native-parent launcher and trusted-main workflow, retain one distinct raw disassembly file per launch, fail visibly on missing capture while preserving all outcomes, and label all capture timings/allocations as excluded from performance evidence. Keep default uninstrumented replay unchanged and reject mixed instrumentation. Add focused launcher coverage for child/parent isolation, shell quoting, missing capture and all four outcomes.

This does not establish which code ran historically, serialize concurrent JIT output, prove equivalence or justify a production patch. Missing/precompiled/inlined/interleaved methods remain explicit limitations. Use the resulting capture only to compare observed compiled dispatch code and tiers within/across images before selecting further experiments.

Tracks the next diagnostic for #613. No shipped API or user documentation changes.

Activity

  1. thomhurst commented on Nov 13, 2026

    @thomhurst
    OwnerAuthor

    Implemented in draft PR #694, commit 6179915c70bd98996612e37d0c4a01c55c5d562f (remote head verified). All 25 focused launcher tests pass; the real .NET 10.0.14 ARM64 smoke validates native JIT output only. Validation source/binaries/logs are archived and read-back hash-verified under /Users/tom/git/Relayer-evidence/pr-694/6179915c70bd98996612e37d0c4a01c55c5d562f/; complete manifest/location in the PR. Codex review requested for this exact head, Claude automatic review in progress; no findings were present when review surfaces were fetched.

    After review/merge, #613 can perform its bounded code-capture diagnostic on trusted main. Do not compare instrumented timings/allocations, claim historical code identity, infer equivalence, rerun the Performance gate or patch production speculatively. No production/per-message/public docs changes. No services or background local processes started; releasing issue-692 and pr-694.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    P1High priorityarea: performanceBenchmarks and hot-path work

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions