Add --profile to record a CPU profile of every sample

[NullVoxPopuli-ai-agent]

pnpm bench --profile=<dir> records a CPU profile of every sample, cut to the time between :start and :done. pnpm profile:report shows where that time went, or what changed against another profiled run.

pnpm bench --framework=ember --bench=all --profile=profiles/ember-main
pnpm profile:report profiles/ember-main
pnpm profile:report profiles/ember-pr --base=profiles/ember-main

How it records

• Each sample runs inside a Chrome trace with only blink.user_timing and the V8 CPU profiler on. The trace has the marks and the profiler samples on one clock, so the cut at :start and :done is exact. The Profiler domain alone has no clock in common with performance.mark.
• Only the page's main thread counts. Worker threads and other renderers are left out.
• Per bench, the directory gets <bench>.jsonl (one line per sample: self time per frame, and time per stack) and <bench>.trace.json.gz (the full trace of the first sample, for DevTools or Perfetto).
• The app's source maps are copied into <dir>/<framework>/maps, because the next build replaces dist.
• The result file records PROFILE in its args. Profiling slows the page down, so a profiled run is only comparable with other profiled runs.
• Profiling also changes the difference between two builds, not only the times. On the ember apps at 8x, it added 10 to 20 points to one build's gain on the sync benches. Use unprofiled runs to compare times, and profiled runs to find where the time goes.

What the report shows

Per bench: self time per package, self time per function, and inclusive time per function. With --base, it shows the largest changes for each instead.

Frames go through the source maps. ember-source ships no maps of its own, but its dist keeps real names, so the report reads the function name from the original source line. It also drops the content hash from chunk names (tracked-JcbMdnul.js becomes tracked.js), so two different ember-source builds line up by function and file.

Example, ember on 1 item, 100k updates at 8x, 3 samples:

self time by function
      27.0 ms  21.5%  set ember-source/shared-chunks/tracked.js
      26.3 ms  21.0%  dirtyTagFor ember-source/shared-chunks/meta.js
      19.0 ms  15.1%  MS common/src/tests/one-item.js
      15.4 ms  12.2%  (garbage collector)
      7.12 ms   5.7%  setter ember-source/shared-chunks/tracked-value.js

Size

At 8x, a trace has a sample about every 180 µs. A 100 ms sample is about 300 KB of trace and 12 KB of JSON lines. Only the first trace per bench is kept.

🤖 Generated with Claude Code

[vercel]

@NullVoxPopuli is attempting to deploy a commit to the NullVoxPopuli's projects Team on Vercel.

A member of the Team first needs to authorize it.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
rere-benchmark	Ready	Preview	Oct 3, 2026 11:03pm UTC