Ruta graveolens  ·  notes from a language experiment  ·  cultivated since 2025
Field study № 2

Performance

This dashboard publishes synthetic compiler phase probes and deterministic scaling diagnostics. They help locate compiler regressions; they are not representative applications and do not measure compiled-program runtime performance. Representative RUE-901 scenarios will remain a separate metric family and headline.

Measured commits

Primary observationloading measurements

Compilation time by measured commit

Absolute milliseconds. Lower is better. Select a point or choose two commits below.

compilation time
new measurement segment

Compiler phases

Aligned to the same measured commits. The aggregate compile timer is excluded.

Field notesevents in the measured history

What changed around these measurements

Measurement and runner changes
Explore the full performance evolution
Secondary observationnormalized, segmented history

Long-term performance evolution

This view shows normalized compile speed, not milliseconds. The first measurement in each connected machine segment is assigned 100, and higher is faster: roughly, 105 means compile speed is 5% above that segment’s baseline. Faint points are measurements; solid daily or weekly trends include observed variation. Compare direction only within a connected segment. Do not rank machines by their indexes or interpret a change across a disconnected measurement boundary.

Calendar range
Raw measurements, gaps, and coverage
Supporting observationsopen as needed

Other diagnostic views

Workload measurements

Absolute time for each benchmark makes a localized regression visible.

Memory and output size
Peak memory
Combined output binary size
Latest workload metrics
Synthetic compiler scaling diagnostics

Generated phase probes show whether latency and peak-memory growth outpace input growth. They are not representative applications and are excluded from the aggregate headline.

Representative compiler build/query scenarios

Cold root builds and fixed CompilerSession edits use one tracked multi-module project. They measure compiler work, not generated-program runtime, and remain separate from phase probes and scaling diagnostics.

Measurement records, coverage, provenance, and raw samples

Technical evidence for reproducing or auditing a point. Fields are labeled; raw samples remain available without exposing serialized objects.

Methodology

Benchmarks run automatically against trunk on Linux x86-64, Linux ARM64, and macOS ARM64 runners. Each point represents a measured commit. When commits land faster than a run completes, an in-flight run may be cancelled in favor of the newest commit.

The primary charts use absolute time on one machine. Comparisons only cross points that the benchmark model marks comparable, and their classification accounts for observed variation. A disconnected line begins a new baseline; it does not imply that compiler performance jumped. Normalized indexes are reserved for comparing the direction of change within each machine because absolute timings from unlike machines should not be overlaid.