Performance loss for `str.rstrip()` for 3.13+

[Alaeddine22]

Description
Starting from Python3.12.0, I noticed that the rstrip function is slower than Python3.11.11
The following is a small table that compares the results of my test with different Python versions:

Python3.11.10 VS Python3.11.11: I had almost the same value.
Python3.11.11 VS Python3.12.0: I had 25.91% loss.
Python3.11.11 VS Python3.12.9: I had 26.91% loss.
Python3.11.11 VS Python3.13.2: I had 15.94% loss.

Any explanation of this performance decrease ?
Thank you in advance.

Reproduction

• pyenv install {version}
• pyenv shell {version}
• python3 -m timeit -n 1000 "for _ in range(10000): 'toto '.rstrip()"

Python versions tested on:
3.11.10
3.11.11
3.12.0
3.12.9
3.13.2

Operating systems tested on:
WSL

[picnixz]

It could come from the fact that instantiating the for loop becomes slower. Instead, I suggest using

pip install pyperf
python -m pyperf timeit -s "s = 'toto    '" "s.rstrip()"

And then report the different timings depending on the version.

---

Also, there are noticeable differences in the interpreter between 3.11 and 3.12+, so this may also be something entirely different. I don't think we changed much str.rstrip but I may be wrong.

And the reason could also be because of WSL (I'll benchmark more tomorrow on my Linux)

[eendebakpt]

Is the same difference there for strip? I recently found that sometimes strip is faster than rstrip, even though it does more work. This might be related to the PGO.

[tanmayadmuthe]

Python 3.12's PEP 683 introduced immortal objects, adding minor overhead to string operations like rstrip().

Key Fixes:

• Use removesuffix() for suffix removal (2.4x faster than rstrip in 3.12)
• Rebuild Python with GCC ≥11 (CC=gcc-11 ./configure --enable-optimizations)
• Regex pre-compilation: re.compile(r'\s+$').sub('', s)

Regression reduced to 15.94% in 3.13.2 (vs 26% in 3.12.x), with ongoing optimizations.

[picnixz]

I can validate the following performance drops on a PGO+LTO build, but only on 3.13 and 3.14:

+-----------+---------+-----------------+-----------------------+-----------------------+
| Benchmark | 3.11    | 3.12            | 3.13                  | main                  |
+===========+=========+=================+=======================+=======================+
| timeit    | 15.4 ns | not significant | 16.5 ns: 1.07x slower | 17.9 ns: 1.16x slower |
+-----------+---------+-----------------+-----------------------+-----------------------+

For strip() it's a bit different but we still lose something:

+-----------+------------+-----------------------+-----------------------+-----------------------+
| Benchmark | strip-3.11 | strip-3.12            | strip-3.13            | strip-main            |
+===========+============+=======================+=======================+=======================+
| timeit    | 15.3 ns    | 14.5 ns: 1.06x faster | 16.6 ns: 1.09x slower | 17.5 ns: 1.14x slower |
+-----------+------------+-----------------------+-----------------------+-----------------------+

[picnixz]

The diassembly of 3.11 vs 3.12+ is a bit different as well:

  0           0 RESUME                   0

  1           2 LOAD_CONST               0 ('toto ')
              4 LOAD_METHOD              0 (rstrip)
             26 PRECALL                  0
             30 CALL                     0
             40 POP_TOP
             42 LOAD_CONST               1 (None)
             44 RETURN_VALUE

For 3.12:

  0           0 RESUME                   0

  1           2 LOAD_CONST               0 ('toto ')
              4 LOAD_ATTR                1 (NULL|self + rstrip)
             24 CALL                     0
             32 POP_TOP
             34 RETURN_CONST             1 (None)

However, the diassembly for 3.13 is almost the same the one for 3.12:

  0           RESUME                   0

  1           LOAD_CONST               0 ('toto ')
              LOAD_ATTR                1 (rstrip + NULL|self)
              CALL                     0
              POP_TOP
              RETURN_CONST             1 (None)

so I don't understand why 3.11 and 3.12 have almost no differences but 3.13 and main have one. So I think we need to accept this loss =/

[Alaeddine22]

Thank you @picnixz fot the results.
Can you share the steps you've done to get them ?