PERFORMANCE · EVERY DECLARATION TIMED
Declaration benchmark
The target is to check every concrete declaration within 100 ms. Select a longer deadline to explore slower dependencies, and the area to measure: the rebuilt library, the archive, or the Cubist tests.
Loading the latest benchmark…
What is measured?
Each declaration is elaborated and admitted by the native C kernel's instruction interface, compiled to WebAssembly: an untrusted driver derives the elaborated body one kernel instruction per typing rule, each checked and answered with a judgement, finding the equality steps where types must agree, and the kernel registers the closed judgement as the definition. Dependencies are checked once, in source import order. The time includes elaboration and admission; it excludes reading source files. Kernel time is the instruction kernel's: every derivation the declaration needed, the checks of its parts that the elaborator asks for as it goes and the admission of its body, which added the judgements shown to the kernel's judgement graph. Elaborator time is the rest: elaboration itself, its search for implicit arguments and the term checker's normal forms it consults included. Inspection and full normalization are not requested.
The selected deadline is polled cooperatively in JavaScript and C. Cleanup or a long operation between polls can exceed the deadline; anything over the selected deadline fails the speed target. Timings vary by machine and cache state. A timeout is a performance result, not a mathematical refutation.
Blocked entries record the first unavailable dependency reached, with its root blocker. Failed entries may indicate an elaborator or migration defect, unsupported syntax, or a faulty proof. An intended failure is a refusal or a block that its source states, as the Cubist tests state theirs, in a comment above the declaration: // Error: E606: Type mismatch: …. It needs nothing, and is listed last; one that fails otherwise than its comment states needs fixing.
The area is the modules measured: the rebuilt library (library/), the archive (archive/first-library/), or the Cubist tests (cubist-tests/), whose stated errors are listed as intended failures. Each area is checked as its modules are, importing as they do, and has its own saved benchmark. A saved benchmark names the revision it measured, and a row links to its source there, on the repository, when its module was measured exactly as committed; a row measured from working changes has no link. The rows of a run in this browser open in the workspace.
The Run button checks the selected area's current sources locally in a worker, with live progress. Cancel stops the worker. A browser run does not upload results or change the saved benchmark. Keep this tab active for comparable timings. The command-line equivalent is npm run benchmark:cubical -- --area=library, archive (the default) or tests.
| Declaration | Category | Time | Elaborator time | Kernel time | Judgements | Details / root blocker |
|---|