Unroll collect_call_path to speed up common cases

[charliermarsh]

Summary

This PR just naively unrolls collect_call_path to handle attribute resolutions of up to eight segments. In profiling via Instruments, it seems to be about 4x faster for a very hot code path (4% of total execution time on main, 1% here).

Profiling by running RAYON_NUM_THREADS=1 cargo instruments -t time --profile release-debug --time-limit 10000 -p ruff_cli -o FromSlice.trace -- check crates/ruff/resources/test/cpython --silent -e --no-cache --select ALL, and modifying the linter to loop infinitely up to the specified time (10 seconds) to increase sample size.

Before:

After:

[charliermarsh]

Empty else, not related.

[github-actions]

PR Check Results

Ecosystem

✅ ecosystem check detected no changes.

Benchmark

Linux

group                                      main                                   pr
-----                                      ----                                   --
formatter/large/dataset.py                 1.00      9.8±0.02ms     4.2 MB/sec    1.00      9.8±0.03ms     4.2 MB/sec
formatter/numpy/ctypeslib.py               1.00   1886.4±1.80µs     8.8 MB/sec    1.00  1888.0±12.67µs     8.8 MB/sec
formatter/numpy/globals.py                 1.00    205.5±1.79µs    14.4 MB/sec    1.00    205.3±0.37µs    14.4 MB/sec
formatter/pydantic/types.py                1.01      4.2±0.01ms     6.0 MB/sec    1.00      4.2±0.02ms     6.1 MB/sec
linter/all-rules/large/dataset.py          1.03     13.8±0.07ms     3.0 MB/sec    1.00     13.3±0.11ms     3.0 MB/sec
linter/all-rules/numpy/ctypeslib.py        1.02      3.5±0.03ms     4.8 MB/sec    1.00      3.4±0.02ms     4.9 MB/sec
linter/all-rules/numpy/globals.py          1.01    373.7±1.20µs     7.9 MB/sec    1.00    369.4±2.89µs     8.0 MB/sec
linter/all-rules/pydantic/types.py         1.02      6.1±0.01ms     4.1 MB/sec    1.00      6.0±0.01ms     4.2 MB/sec
linter/default-rules/large/dataset.py      1.00      7.0±0.01ms     5.8 MB/sec    1.00      6.9±0.01ms     5.9 MB/sec
linter/default-rules/numpy/ctypeslib.py    1.00   1413.3±2.96µs    11.8 MB/sec    1.00   1415.5±1.69µs    11.8 MB/sec
linter/default-rules/numpy/globals.py      1.00    147.4±0.29µs    20.0 MB/sec    1.00    147.6±0.67µs    20.0 MB/sec
linter/default-rules/pydantic/types.py     1.00      3.1±0.01ms     8.3 MB/sec    1.00      3.1±0.01ms     8.2 MB/sec

Windows

group                                      main                                   pr
-----                                      ----                                   --
formatter/large/dataset.py                 1.00     14.4±0.61ms     2.8 MB/sec    1.27     18.3±0.58ms     2.2 MB/sec
formatter/numpy/ctypeslib.py               1.00      2.9±0.17ms     5.8 MB/sec    1.17      3.3±0.13ms     5.0 MB/sec
formatter/numpy/globals.py                 1.00   311.8±14.48µs     9.5 MB/sec    1.14   354.2±17.15µs     8.3 MB/sec
formatter/pydantic/types.py                1.00      6.5±0.44ms     3.9 MB/sec    1.13      7.3±0.26ms     3.5 MB/sec
linter/all-rules/large/dataset.py          1.00     20.3±0.72ms     2.0 MB/sec    1.00     20.4±0.62ms  2044.0 KB/sec
linter/all-rules/numpy/ctypeslib.py        1.00      5.3±0.21ms     3.2 MB/sec    1.02      5.3±0.24ms     3.1 MB/sec
linter/all-rules/numpy/globals.py          1.00   625.4±27.38µs     4.7 MB/sec    1.02   640.1±32.57µs     4.6 MB/sec
linter/all-rules/pydantic/types.py         1.00      9.0±0.35ms     2.8 MB/sec    1.03      9.3±0.47ms     2.7 MB/sec
linter/default-rules/large/dataset.py      1.00     10.4±0.44ms     3.9 MB/sec    1.07     11.1±0.33ms     3.7 MB/sec
linter/default-rules/numpy/ctypeslib.py    1.00      2.2±0.11ms     7.7 MB/sec    1.02      2.2±0.08ms     7.5 MB/sec
linter/default-rules/numpy/globals.py      1.00   263.3±11.99µs    11.2 MB/sec    1.01   266.7±15.57µs    11.1 MB/sec
linter/default-rules/pydantic/types.py     1.00      4.7±0.14ms     5.5 MB/sec    1.00      4.7±0.21ms     5.4 MB/sec

[sbrugman]

and modifying the linter to loop infinitely

@charliermarsh Could you please share the exact diff? Was inspired by this PR to experiment with some of the benchmarking in #5811, but couldn't reproduce this setup (it ended up giving irrelevant profiles) so ended up with running over CPython

[konstin]

@sbrugman You can get the same result with cargo run --bin ruff_dev -- repeat --no-cache --repeat 10000 crates/ruff/resources/test/cpython (it's ruff check with an additional --repeat option)

[charliermarsh]

Oh that seems a lot better than what I did. I will post the diff regardless in a bit.

[konstin]

The return type of collect_call_path also seems like an opportunity for optimization: We generally either want to check against a specific length (e.g. 2) and don't need to save anything otherwise or iterate over the results (e.g. the widely used resolve_call_path), where we only want to iterate over elements from left to write (we could do this with an Iterator that is implemented recursively, but i'm not sure if the extra function calls aren't making this slower)

[konstin]

That's a great speedup

[charliermarsh]

The return type of collect_call_path also seems like an opportunity for optimization...

Agreed.