Semantic Diff

#9270 Benchmark project builds and output sizes in CI

3 commits

bf35764db1628906bf68c424735c68da4d1711fb..pr-9270

7b1226611 · cknitt + Claude · Nov 08, 2026, 21:23 UTC

Benchmark project builds and output sizes in CI

The benchmarks only replayed single bsc calls, so the build system's own work never showed. They now also build tests/tests on one worker domain (clean, no-op, leaf edit, shared edit) and measure the .cmi/.cmti/.cmt/.cmj sizes and the playground bundle; rewatch reports its workers' allocations, which the runtime's exit report leaves out, under REWATCH_REPORT_WORKER_ALLOCATIONS.

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.
78a68c5e9 · cknitt + Claude · Nov 08, 2026, 21:33 UTC

Count instructions for the small builds only

Under cachegrind the shared-edit build took about 40 s a side and its instructions differed by 9.5% between base and head of this pull request: with two domains collecting, they depend on how the domains interleave.

ded1cf367 · cknitt + Claude · Nov 09, 2026, 06:41 UTC

Don't alert on builds that count the main domain only

A base whose build system doesn't report its workers' allocations counts only the main domain, which the PR adding the report showed as regressions of up to +7900%; those rows are now marked as having other inputs.