GeezerVison

Move bool.__new__ to Argument Clinic with @vectorcall, dropping the
hand-written bool_new and bool_vectorcall. The generated vectorcall name
matches the hand-written one so the tp_vectorcall slot is unchanged, and
argument error messages are unchanged on both paths.

pyperf microbenchmark (PGO+LTO, run serially), no significant change on
the vectorcall path:

| Benchmark      | base    | bool                  |
|----------------|:-------:|:---------------------:|
| bool.__new__   | 23.0 ns | 21.7 ns: 1.06x faster |
| bool()         | 10.1 ns | 10.4 ns: 1.03x slower |
| bool(int)      | 12.2 ns | 12.6 ns: 1.03x slower |
| bool(True)     | 11.1 ns | 11.4 ns: 1.03x slower |
| Geometric mean | (ref)   | 1.00x slower          |

Benchmark hidden because not significant (2): bool(__bool__), bool(list)

Instructions per call (callgrind, inclusive of bool_vectorcall or
bool_new, PGO+LTO):

| Case                  | base | bool | delta |
|-----------------------|-----:|-----:|------:|
| bool()                |   17 |   18 |    +1 |
| bool(True)            |   20 |   23 |    +3 |
| bool(12345)           |   53 |   49 |    -4 |
| bool([1])             |   50 |   47 |    -3 |
| bool.__new__(bool, 1) |  127 |   69 |   -58 |

The compiler folds PyObject_IsTrue(Py_False) for bool(), so the default
costs nothing; the small deltas are register allocation. bool.__new__
drops PyArg_UnpackTuple for inline clinic parsing.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

msullivan left a comment