Repository navigation
Perf: cheaper request cycle and cold path #573
Description
This was generated by AI during a pre-4.0 audit of main at 9cd88e7.
Warm resolve is at its floor. Measured on CPython 3.14.7 (cross-checked 3.11-3.13), full suite green against the patched copy:
| Change | Where | Effect |
|---|---|---|
CacheRegistry and ContextRegistry as plain __slots__ classes; build ContextRegistry positionally |
cache_registry.py:72, context_registry.py:7, container.py:88-89 |
kw-only dataclass call 138 ns vs 63 ns positional; G6 -19%, G7 -6% |
Memoize compiled code on a tuple key (arity, names, static, cached) instead of the frozen _Shape dataclass |
resolver_compiler.py:190-196 |
G8 cold 24.4 -> 20.8 us (-15%) |
Inline find_context into the ContextProvider resolver (keep in + []) |
resolver_compiler.py:280 |
direct context resolve 160 -> 134 ns, G9 -6%; removes a second frame per context node |
| Skip the cache close when nothing was created | container.py:213,221 |
request cycle with nothing cached -6 to -11% |
Dropping abc.ABC from AbstractProvider (another G8 -4%, Marker.resolve -30%) is in the provider-surface issue.
Rejected: building the child via object.__new__ (another -15-20% on G6 but duplicates construction logic) and is instead of == for scope checks (3.14-only gain).
Acceptance
Before/after numbers for G6, G7, G8, G9 in the PR; frame-budget test still green.
Activity
perf: cheaper request cycle and cold path (#573)