beartype / plum / 33876858524

99%

push

github

perf(type): memoise the wrapper used by signature comparison (#301)

`Signature.__eq__` and `__le__` compare hints through `_type_hint_eq` /
`_type_hint_le`, each of which rebuilds `TypeHintWrapper(_substitute_any(h))`
from scratch. For a function no cache can serve -- one dispatching on
`Literal`, `list[int]`, `Callable[...]` -- full resolution runs on every call,
so those two helpers run about 14 times per call, and each time pay the
hashability probe, the `lru_cache` lookup for the rewrite, and `beartype`'s own
wrapper lookup.

The hints are fixed per signature, so memoise the finished wrapper instead of
the rewrite alone. Unhashable hints (`Annotated[int, {"a": 1}]`) fall back to
building it, as `_substitute_any` already does.

Uncached dispatch on a `Literal`, median of `tests/benchmarks`:

| path | before | after |
|---|---:|---:|
| `Literal` dispatch | 34,959 ns | 15,750 ns (2.22x) |
| registration (5 shapes) | 46.7-60.8 us | 43.5-57.3 us (~1.07x) |

Every cached path is unchanged: faithful dispatch 388 ns, `type[X]` 3,709 ns,
`invoke` 119 ns, annotated return 481 ns, all within noise of before.

This also repairs a regression. `Literal` dispatch cost 27.7 us before #296 and
35.7 us after it, flat across the sixteen commits in between -- the added
`_substitute_any` call in each comparison is the difference. The memo takes it
below the pre-#296 figure rather than merely back to it, because it also skips
the wrapper construction that was always there.

Like `_substitute_any`'s own cache, this one is not cleared by
`clear_all_cache`: both are keyed on an already-resolved hint, so
`type_mapping` and late-delivered `ModuleType`s are handled before this point.

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

10 of 10 new or added lines in 1 file covered. (100.0%)

1046 of 1053 relevant lines covered (99.34%)

6.73 hits per line