Multithreaded scaling regression in 3.13.1t

[colesbury]

Bug report

Bug description:

#127114 changed the logic to immortalize certain objects when the first thread is created instead of the first non-main thread.

3.13.0:

sgross@devfair0453:/raid/sgross/clean-cpython/Tools/ftscalingbench$ uv run -p 3.13.0t ftscalingbench.py
Running benchmarks with 20 threads
object_cfunction 14.3x faster
cmodule_function 14.3x faster
object_lookup_special 14.3x faster
mult_constant 13.8x faster
generator 13.8x faster
pymethod 14.3x faster
pyfunction 14.0x faster
module_function 14.2x faster
load_string_const 2.5x slower
load_tuple_const 13.5x faster
create_pyobject 14.0x faster
create_closure 14.0x faster
create_dict 13.0x faster
thread_local_read 3.0x slower

3.13.1:

sgross@devfair0453:/raid/sgross/clean-cpython/Tools/ftscalingbench$ uv run -p 3.13.1t ftscalingbench.py
Running benchmarks with 20 threads
object_cfunction 1.6x slower
cmodule_function 1.9x slower
object_lookup_special 1.7x slower
mult_constant 14.3x faster
generator 2.7x faster
pymethod 2.3x slower
pyfunction 3.1x slower
module_function 2.1x slower
load_string_const 2.1x slower
load_tuple_const 13.1x faster
create_pyobject 3.6x slower
create_closure 4.3x slower
create_dict 15.0x faster
thread_local_read 3.0x slower

CPython versions tested on:

3.13

Operating systems tested on:

No response

Linked PRs

• [3.13] gh-131988: Fix a multithreaded scaling regression #131989

[vstinner]

I added "release-blocker" since it would be nice to get this important performance fix in the next 3.13 release.

[hugovk]

The next 3.13 release is tomorrow, please review #131989.