aio-libs / multidict / 36020212647

87%

push

github

Compare the hash before taking a reference in the lock-free probes (#1570)

<!-- Thank you for your contribution! -->

## What do these changes do?

On the free-threaded build `d[key]`, `d.get()`, `d.getone()` and
`key in d` probe the table without the critical section, and for every
entry on the probe chain took a reference to the entry's identity with
`try_get_ref()` (an atomic load, a `TryIncRef`, a re-load and a
`Py_DECREF`) before even comparing hashes. The two lock-free probes now
compare the hash first, so a mismatch costs no reference traffic, and
treat an identity that is the very same object as the one being looked
up as a match without taking the reference: the caller's own reference
keeps that object alive, so its address cannot have been reused for
anything else. Only an identity that is a different object still goes
through `try_get_ref()` for the string compare, and the `NULL` and
`_MD_NEED_LOCK` fallbacks are unchanged.

The GIL build has no lock-free probe and is untouched; the pure-Python
implementation has nothing corresponding.

Instruction counts per operation, `benchmarks/callgrind_driver.py`,
CPython 3.14.7t, `PYTHONHASHSEED=0`:

| operation, free-threaded | `MultiDict` before | after | |
`CIMultiDict` before | after | |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| `d[key]` | 291 | 249 | -14% | 378 | 336 | -11% |
| `key in d` | 253 | 209 | -18% | 340 | 295 | -13% |
| `d.get(key)`, miss | 409 | 388 | -5% | 660 | 639 | -3% |
| `d[istr]` | | | | 292 | 250 | -14% |
| `d.update(other)`, 100 existing keys | 48,915 | 48,283 | -1% | 57,863
| 57,251 | -1% |

`dict` on the same build: `d[key]` 224, `key in d` 235. Every other
row is unchanged to the instruction.

Wall clock, interleaved before/after runs pinned to a P-core on an idle
machine, median ns per operation:

| operation, free-threaded | `MultiDict` before | after | |
`CIMultiDict` before | after | | `dict` |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| `d[ke... (continued)

777 of 1554 branches covered (50.0%)

Branch coverage included in aggregate %.

7693 of 8107 relevant lines covered (94.89%)

1.9 hits per line