Move stackref buffer to per-eval loop to reduce interp stack usage

Feature or enhancement

Proposal:

The interpreter main loop's stack usage is huge. We should try to reduce it a little. Currently, the stackref buffer takes up 10 words on 64-bit machines. We could lessen that by moving it to the heap (thread state).

This might mean slightly less perf due to worse locality and one memory indirection. So let's benchmark this to be sure.

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

Linked PRs

gh-138115: Move stackref buffer to thread state to reduce interp stack usage #138116

@colesbury

Thanks to talking to @colesbury , there's a cheaper approach here.

Allocate per-thread/per-interpreter buffer for stackref scratch, similar to the datastack chunk.
At each _PyThreadState_PushFrame, do the bounds check to see if we have enough space left on the stackref buffer fro co->co_stacksize. If we don't, allocate a new chunk.
At each callsite, no bounds check is needed thanks to 2., just bump the pointer in the stackref chunk and use the scratch in STACKREFS_TO_PYOBJECTS.
After each call, shrink the pointer in the stackref chunk.

I think it needs to be per-thread (even with the GIL) so that the pointer bumps/shrinks in (3) and (4) match up. I think with a single per-interpreter stack you could get incorrect interleavings.

Thanks, yes that seems right. I think we need this now for correctness in 3.14 and 3.15 though, to un-crash Clang builds. So it's not just about perf anymore (though the perf should be around the same in most cases).

See #148284 for an actual segfault in the wild on Clang 21 builds.

This seems like a bit of a complex change to backport to 3.14.

I'm not sure if this would work, but a simpler 3.14 change might be:

In 3.14, when building without the tail call interpreter, declare the temporary storage for STACKREFS_TO_PYOBJECTS once inside PyEval_FrameDefault
Use that common PyObject** array in each STACKREFS_TO_PYOBJECTS when not building with the tail call interpreter

@colesbury

Just checking, I built CPython with CFLAGS="-g" LDFLAGS="-fuse-ld=lld-22" RANLIB=llvm-ranlib-22 CC=clang-22 ./configure --enable-optimizations --with-lto --enable-shared && make clean && make -j18

LD_LIBRARY_PATH=/home/ken/Documents/GitHub/cpython llvm-objdump-22 --disassemble-symbols=_PyEval_EvalFrameDefault --source ./libpython3.14.so > 1.txt

gives me the following dissassembly:

1.txt

with @colesbury suggested fix, I get

2.txt

Stack usage went from subq $0x6f8, %rsp to subq $0x268, %rsp, or roughly 1/3rd now. Wow!

Is this still relevant, or can we close it?

Navigation Menu

Repository navigation

Move stackref buffer to per-eval loop to reduce interp stack usage #138115

Feature or enhancement

Proposal:

Has this already been discussed elsewhere?

Links to previous discussion of this feature:

Linked PRs

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Move stackref buffer to per-eval loop to reduce interp stack usage #138115

Description

Feature or enhancement

Proposal:

Has this already been discussed elsewhere?

Links to previous discussion of this feature:

Linked PRs

Activity

Fidget-Spinner commented on Apr 9, 2026

colesbury commented on Apr 9, 2026

Fidget-Spinner commented on Apr 9, 2026

colesbury commented on Apr 9, 2026

Fidget-Spinner commented on Apr 9, 2026

sergey-miryanov commented on May 4, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions