[hauntsaninja]

Authored by Codex, here's its description below.

Note that its baseline doesn't include #22270 so it probably overstates the win a little

TypeInfo.get_method scanned every name in each MRO entry when a method was absent. On colour, native sampling attributed 21% of the baseline profile to this lookup.

The semantic analyzer inserts the unnumbered "name-redefinition" before any numbered redefinitions. Check for that key first. When it exists, select the same lexicographically last definition as before. This keeps the fallback that dataclass InitVars need.

Both binaries checked the same source trees cold. The baseline was b286c11c49ef.

    input     median before -> after     paired wall reduction
    self         6.227 s ->  6.169 s             0.84%
    torch        3.251 s ->  3.278 s            -0.11%
    pandas      24.372 s -> 23.622 s             2.82%
    colour      38.643 s -> 29.050 s            24.69%
    jax         29.032 s -> 27.883 s             3.39%
    zulip       14.505 s -> 14.154 s             2.19%

The torch interval spans zero (-0.75% to +0.47%). Each pair ran both Clang 18 release binaries on the same CPU set, with an empty mypy cache. The order alternated over 12 pairs for self, 14 for torch 2.1.0, and eight per project from mypy_primer 9889850. The last column averages within-pair wall time ratios. All pairs produced byte-identical diagnostics. The checker, fine-grained incremental, and dependency suites passed (9,134 tests).

An independent multifile mypyc build reduced colour wall time by 26.01% across six pairs (95% interval 25.66% to 26.67%). Its self result was inconclusive (0.20%, interval -0.64% to 1.15%). A common import path and common typeshed files gave a 0.60% self improvement across 14 pairs with the single-unit build.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[JukkaL]

Thanks for the fix, this is a nice win.