[cknitt]

The benchmarks job replays single bsc calls only, so the build system's own work (scanning, dependency graph, freshness checks, in-process .cmi loading, scheduling) never showed up. This adds:

• Project builds: tests/tests (~680 modules) built with rescript build on one worker domain (REWATCH_COMPILER_DOMAINS=1): clean, no-op, after editing a leaf test, and after editing test_utils.res (~200 dependents). Allocated words for all four, including the workers' allocations that the build reports with REWATCH_REPORT_WORKER_ALLOCATIONS (Report the build's worker allocations when asked #9271, below this PR). Instructions under cachegrind for the no-op and leaf-edit builds only: the bigger builds would take minutes, the replayed builds cover their compiler work, and with two domains collecting their instruction counts depend on interleaving (the shared edit differed by 9.5% between base and head of an earlier run). Both sides build the base's tests/tests from a fixed directory, with -warn-error dropped, so new tests or warnings in a PR can't break either build.
• Output sizes: total .cmi / .cmti / .cmt / .cmj bytes for runtime, Belt and tests/tests, plus the playground's compiler.js. A new "Output size" section, alerting like the others at 1%.

Local runs (linux-arm64): clean build 1.4 s, the other three 0.1–0.6 s; repeated runs agree within 0.0004% (allocations), sizes are identical.

Job time: 266 s in an earlier run of this change (was 234 s on master; runner speed varies by more than that). The two playground builds are 75 s of it; the critical path (Windows build) stays ~550 s.

Replaces #9270, split so that this PR's base already reports the workers' allocations.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Kbf3gGPsF3fHZn9rBxA7Hm

[github-actions]

Changelog

No changelog entry.

Semantic Diff

1 commit · Full page

Benchmark project builds and output sizes in CI

_{1673f76 · cknitt + Claude · Nov 09, 2026, 06:46 UTC}

The benchmarks only replayed single bsc calls, so the build system's own work never showed. They now also build the base's tests/tests on one worker domain (clean, no-op, leaf edit, shared edit), adding the workers' allocations the build reports, and measure the .cmi/.cmti/.cmt/.cmj sizes and the playground bundle. Only the no-op and leaf-edit builds run under cachegrind: the larger ones take minutes, and with two domains collecting their instructions depend on how the domains interleave.

Review

• Why build the base's tests/tests on both sides? The base compiler builds them already, so a pull request adding tests with new syntax can't break the base measurement.

[github-actions]

✅ Compiler benchmarks

Allocated words

In total: 806,109,534 → 806,109,309 (±0.00%).

None of the 261 benchmarks changed by 0.1% or more.

Instructions

In total: 11,321,687,111 → 11,321,641,809 (±0.00%).

None of the 35 benchmarks changed by 0.1% or more.

Output size

In total: 29,806,856 → 29,806,856 (±0.00%).

None of the 13 benchmarks changed by 0.1% or more.

Pull request 1673f76 compared with base a5e4a65 in this run.

[pkg-pr-new]

Open in StackBlitz

rescript

npm i https://pkg.pr.new/rescript@9272

@rescript/belt

npm i https://pkg.pr.new/@rescript/belt@9272

@rescript/darwin-arm64

npm i https://pkg.pr.new/@rescript/darwin-arm64@9272

@rescript/linux-arm64

npm i https://pkg.pr.new/@rescript/linux-arm64@9272

@rescript/linux-x64

npm i https://pkg.pr.new/@rescript/linux-x64@9272

@rescript/runtime

npm i https://pkg.pr.new/@rescript/runtime@9272

@rescript/win32-x64

npm i https://pkg.pr.new/@rescript/win32-x64@9272

commit: 1673f76