Avoid overhead of cast() calls when not type checking

[correctmost]

Type of Changes

	Type
✓	🔨 Refactoring

Description

While profiling Pylint with the yt-dlp codebase, I noticed that cast was called 1.277 million times:

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
  1277992    0.130    0.000    0.130    0.000 /usr/lib/python3.12/typing.py:2173(cast)

This PR avoids 1.1 million of these calls.

Background

python/cpython#116290 (comment) has this to say about cast:

My own vision for call inlining is that it's mostly going to make "trivial" functions disappear -- a good example would be typing.cast(), which returns its second argument unchanged. I currently sometimes hesitate about putting that in because of the perceived overhead of even the most trivial function call [...]

Stats

Before

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
  1277992    0.130    0.000    0.130    0.000 /usr/lib/python3.12/typing.py:2173(cast)

Command	Mean [s]	Min [s]	Max [s]	Relative
pylint --recursive=y .	36.214 ± 0.381	35.590	36.608	1.00

After

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
   176834    0.022    0.000    0.022    0.000 /usr/lib/python3.12/typing.py:2173(cast)

Command	Mean [s]	Min [s]	Max [s]	Relative
pylint --recursive=y .	35.195 ± 0.165	34.974	35.326	1.00

Reproduction notes

• The tables were generated by hyperfine
  • hyperfine --ignore-failure --warmup 2 --runs 5 --export-markdown=baseline.md 'pylint --recursive=y .'

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 92.98%. Comparing base (15207a7) to head (26c6840).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2500      +/-   ##
==========================================
- Coverage   92.98%   92.98%   -0.01%     
==========================================
  Files          93       93              
  Lines       11040    11039       -1     
==========================================
- Hits        10266    10265       -1     
  Misses        774      774              

Flag	Coverage Δ
linux	92.86% <ø> (-0.01%)	⬇️
pypy	92.98% <ø> (-0.01%)	⬇️
windows	92.97% <ø> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files	Coverage Δ
astroid/transforms.py	100.00% <ø> (ø)

[Pierre-Sassoulas]

Thank you for the merge request and the benchmark. This avoid a lot of cast call but at the cost of the same number of if check. I'm not convinced the net result is positive (as the given benchmark do not include the information regarding the cost of the additionnal 'if' afaiu) reread the summary and the diff is surprisingly large.

[correctmost]

Thanks for the review :).

The numbers from the hyperfine runs will have some noise because my computer is not perfectly tuned to eliminate variance, the Pylint runs involve disk I/O, etc.

But the cProfile stats and these additional timeit stats make me believe that there is a small speed-up for large codebases:

# Without the if check
python -m timeit -s 'from typing import cast, TYPE_CHECKING; x=1' -n 1000000 'x=cast(int, x)'
1000000 loops, best of 5: 40.1 nsec per loop

# With the if check
python -m timeit -s 'from typing import cast, TYPE_CHECKING; x=1' -n 1000000 'if TYPE_CHECKING: x=cast(int, x)'
1000000 loops, best of 5: 9.63 nsec per loop