Repository navigation
bench: commit class_attrs_methods jitstats baselines - #952
youknowone merged 2 commits into
Conversation
class_attrs_methods regained its method-call inline (loops_compiled 2 -> 1, dynasm 0.07s user vs the 1.7s residual round trip it gated at max-pypy-ratio=824), but nothing pins the trace shape: the perf gate is too loose to notice the callee residualizing again. Record the dynasm/cranelift structural baselines (loops_compiled=1, loops_aborted=0) so check.py's regression floor reddens a bare run the moment the method call falls back to a residual with its own func-entry trace. Recorded via check.py --synthetic-only --snapshot on macOS arm64; no wasm toolchain locally, so the wasm baseline is left for a runner that has one. Co-Authored-By: Claude Fable 5 <[email protected]>
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: ASSERTIVE Plan: Pro Plus Run ID: 📒 Files selected for processing (1)
WalkthroughThe change adds a ChangesClass attributes JIT statistics
Estimated code review effort: 1 (Trivial) | ~2 minutes Possibly related PRs
Suggested reviewers: Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8be024a66d
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| guard_failures=1 | ||
| internal_compile_panics=0 | ||
| loops_aborted=0 | ||
| loops_compiled=1 |
There was a problem hiding this comment.
In the default pyre/check.py path this value is never compared: _jit_stats_regression_floor only considers the badness fields (loops_aborted, internal_compile_panics, and descr-set counters), and the CI --snapshot-diff invocation skips synthetic benches with --no-synthetic. If this benchmark regresses back to a separate method-entry trace (1 loop -> 2 loops) without aborting, the new baseline still lets a bare check pass; add an exact/default gate for this synthetic stat (or tighten the perf gate) rather than just recording the value.
Useful? React with 👍 / 👎.
| # these on every run, which is the only gate wasm has — run_synthetic_bench | ||
| # passes vs_pypy=None for wasm, so no perf ratio can catch a wasm-only trace | ||
| # abort storm there. | ||
| !/bench/synth/class_attrs_methods.*.jitstats |
There was a problem hiding this comment.
On Linux CI, installing wasm32-unknown-unknown makes wasm part of DEFAULT_BACKENDS, and _apply_snapshot_gate only applies the regression floor when pyre/bench/synth/<name>.wasm.jitstats exists. This new unignore pattern covers class_attrs_methods.wasm.jitstats, but only the native baseline files were committed, so wasm-only abort/panic/descr-counter regressions for this synthetic bench still have no committed floor (unlike the other exceptions). Add the 5-field wasm baseline when enabling this benchmark.
Useful? React with 👍 / 👎.
What
Commit
class_attrs_methods.{dynasm,cranelift}.jitstatsstructural baselines(plus their
pyre/.gitignoreexception line, following the existingcomprehension_object_append_hot/const_arg_call_resume/nested_list_comprehension_hotpattern).Why
class_attrs_methodsis themax-pypy-ratio=824fixture — the loosest perfgate in the tree, calibrated (#853) when its hot
c.add(i & 7)method call wasdeclined by the method-form inline gate and ran as a full
CallMayForceRresidual round trip per iteration (fresh callee
PyFrame, warmstate dispatchinto a separately compiled function-entry trace via
execute_token,re-boxing — ~1.7s user on an M-series dynasm build vs 0.09s CPython).
Current main inlines the call again:
loops_compiled2 → 1, dynasm 0.07s /cranelift 0.09s user. But nothing pins that trace shape — with an 824x
allowance, the callee residualizing again (and recompiling its own func-entry
loop) would sail through a bare
pyre/check.py. The committed baselines makecheck.py's jit-stats regression floor catch exactly that structural
regression, timing noise aside.
Baselines were recorded with
pyre/check.py --synthetic-only --synthetic-pattern class_attrs_methods.py --snapshotand verified green on a bare re-run (dynasm 1/1, cranelift 1/1, 4.2x/6.2x vs
pypy). No wasm baseline: no wasm toolchain on this machine — can follow up
from a runner that has one, as the other baselined benches carry a wasm file.
🤖 Generated with Claude Code
Summary by CodeRabbit