Compiler performance
Headline index
The geometric mean of each workload's median against this epoch's baseline. 1.00 is the baseline; lower is faster. It is dimensionless — it answers whether anything is moving, not how long a build takes. A move in one of n workloads shifts it by roughly 1/n of its true size, so the per-workload panels carry the full-size signal.
Workloads
What “flagged” means
A workload is flagged when its median differs from the median of the trailing window of runs by more than k times the two dispersions combined — the rule in ADR-0067 §5, with k and the window length pinned by the epoch. Three things follow that the word alone does not say:
- It is measured against noise, not against a percentage. A 1% move on a quiet workload can qualify where 5% on a noisy one does not.
- It has no direction. The rule takes an absolute difference, so a large speedup is flagged exactly as a regression is.
- It compares against the trailing window, not the previous commit, so it can disagree with the commit-to-commit delta beside it.
Not enough history means the trailing window is not full yet. That is not the same as stable, and is deliberately not rendered as though it were. Hover a workload, or the phase chart, for the figures behind the verdict.
Phase composition —
Absolute milliseconds by published phase. These bands are mutually exclusive and sum exactly to compiler-root time, which is what makes stacking them honest. mixed_parallel is time when more than one phase was active; unattributed is compiler time no phase describes. Both are findings, not rounding.
Field notes
Disclosures
Driver overhead
Process wall time outside compiler root: startup, output publication, and other driver cost. Real time you wait for, deliberately outside the phase stack.
Memory and binary size
Collection health
Download the derived data — rebuilt from the raw run objects on the performance-data-v1 branch, which stores observations and nothing derived.
No measurements have been collected yet. The measurement system is in place; this page fills in once collection has run on trunk and each epoch has a baseline. A baseline must be the first complete, valid run at a declared revision, so it cannot exist before one has been measured.