gh-139716: Make PyStackRef_FromPyObjectSteal function very lightweight for GIL build as on FT build

[efimov-mikhail]

This is another attempt to provide the same StackRef flagging scheme for all builds (GIL, FT, STACKREF_DEBUG).
Instead of adding _Py_IsImmortal check to PyStackRef_FromPyObjectSteal for FT builds as in #141675, we make no checks at all in this function.
As a result, immortal objects will be DECREF'ed with no effect at reference closing.

• Issue: Add tests on basic PyStackRef functions #139716

[efimov-mikhail]

CC @Fidget-Spinner@colesbury@sergey-miryanov

[efimov-mikhail]

If this PR will be merged we can remove PyStackRef_FromPyObjectStealMortal at all.
OTOH, I'd like to provide this change in a separate PR.

[Fidget-Spinner]

This is another attempt to provide the same StackRef flagging scheme for all builds (GIL, FT, STACKREF_DEBUG). Instead of adding _Py_IsImmortal check to PyStackRef_FromPyObjectSteal for FR builds as in #141675, we make no checks at all in this function. As a result, immortal objects will be DECREF'ed with no effect at reference closing.

• Issue: Add tests on basic PyStackRef functions #139716

We can't rely on Py_DECREF to do the immortal check, as that is more expensive than simply checking a bit on a pointer. You have to do an atomic read of memory now.

[efimov-mikhail]

We can't rely on Py_DECREF to do the immortal check, as that is more expensive than simply checking a bit on a pointer. You have to do an atomic read of memory now.

It seems that I'm not quite understand, when we will do additional atomic reads.
Do you mean FT builds? For those nothing changes since Py_DECREF_MORTAL is Py_DECREF on them.
And there's no atomic reads at Py_DECREF on GIL builds.

[Fidget-Spinner]

Sorry I think I'm misunderstanding. This PR moves the immortal check from the stackref to the Py_DECREF right?

[efimov-mikhail]

Sorry I think I'm misunderstanding. This PR moves the immortal check from the stackref to the Py_DECREF right?

Yes. If we steal stack reference from the immortal object then there will be no actual decrefing at Py_DECREF.

[kumaraditya303]

We can't rely on Py_DECREF to do the immortal check, as that is more expensive than simply checking a bit on a pointer. You have to do an atomic read of memory now.

Actually for immortal check, only ob_ref_local is read with relaxed order so it will get compiled to plain load. That will still be more instructions executed that checking a single bit though.

[colesbury]

1. Please benchmark this
2. I think @markshannon preferred the current scheme

[markshannon]

I like the idea of merging the two implementations, but why choose the slower scheme, not the faster one?
AFAICT, the FT variant has an extra memory access for the immortality check when doing a CLOSE or DUP.

[github-actions]

This PR is stale because it has been open for 30 days with no activity.