Current performance

Loading the last daily measurement…

Metric

Compare Native with

By language area

Area	Assessment	Compared	Weakest workload	Language coverage

Area assessments use the geometric mean of Native relative to the selected comparison. Language coverage links to the ordinary-language probes of the same family; a fast area with differing probes is not complete support.

By workload

Workload	Native	Pure Go	Assessment

Compiler self-compilation

The current Native compiler builds its own source at the measured revision. These diagnostic values track build cost and emitted QBE size; they are separate from workload comparisons.

Recent compiler measurements

Measured	Revision	Self build	QBE size	QBE emission

Other languages: weekly small-input comparison ↗

Measurement details and how to read the results

Changes you can act on

Current, previous and frozen Native compilers run the same workload on the same host. TypeRB Go compiles the identical source. Pure Go uses hand-written programs: the three from the detailed comparison and one for each language-area workload. Language-area workloads use features newer than the frozen Native baseline, so they are compared with the previous day, TypeRB Go and Pure Go. Implementation differences contribute to the result. Lower time, memory and size are better. Differences below 5% are labelled “About the same”; this is a screening aid, not proof of equivalence or a statistical acceptance threshold.

Runtime ranges show the retained minimum and maximum. Trends retain up to 60 snapshots and show only the current measurement series. Input or measurement-policy changes start a new series.

Freshness without blocking development

Performance changes are measured once a day, around 06:23 JST, or on demand. GitHub scheduling may be delayed. An unchanged day needs no benchmark run. The daily update is independent of PR acceptance.

Numerical regressions are published. A failed measurement is shown explicitly; infrastructure failures preserve the last measured snapshot.

Read the measurements in scope

These are small workloads, not a general language score. Builds use clean output directories with a warm Go cache. Runtime includes fresh-process startup, with warm filesystem caches. Startup measurements include the launch harness; very short CPU and RSS observations may be below measurement resolution.

Peak RSS uses GNU time, not a summed process-tree sample. The allocation probes are separate from timing. Long-lived memory stability and full distribution size remain separate checks. Method and coverage ↗