Repository navigation
Optimize _PyObject_IS_GC in free-threaded build #114733
Description
Feature or enhancement
@nascheme writes the following suggestion (from #114564 (comment)):
_PyObject_IS_GC() could be optimized in the nogil build. That function is a bit expensive since it has to check for tp_is_gc() and possibly call it. Now that all objects have ob_gc_bits, we could store the "IS GC" status as another bit there, when the object is allocated.
I think this might require a bit of refactoring, but worth doing eventually. At a glance:
_PyObject_GC_New()(and related functions) could initializeob_gc_bitswith the "IS GC" setPyType_GenericAlloc()would also need to initializeob_gc_bitswith the "IS GC" set if the type is a GC type._Py_NewReference()requires some care. It. currently initializesob_gc_bitsto zero and it's used after "allocating" objects from freelists.
Activity
corona10 commentedon Oct 19, 2026
I will re-evaluate this optimization, my last year's experiment was not worth doing that.
Let's rewrite the patch and if it does not change signifacnt changes, lets' close the issue.
cc @mpage @nascheme
corona10 commentedon Oct 19, 2026
Let me get metrics first based on pyperformance.
corona10 commentedon Oct 19, 2026
sample_gc_collect.txt
sample_gc_traversal.txt
Benchmarks based on gc_collect and gc_traversal show this is too small a portion for optimization; I would like to close this issue and revisit later if we have to.