Don't inline slow path functions in the interpreter loop

[mpage]

Feature or enhancement

Proposal:

I was looking at the IPA inlining dump as part of investigating the root cause of gh-132295 and noticed that we're inlining a number of slow path functions related to specialization, instrumentation, and error handling in the interpreter loop when performing LTO (full list here):

_Py_Specialize_CallKw
_Py_Specialize_CompareOp
_Py_Specialize_ContainsOp
_Py_Specialize_ForIter
_Py_Specialize_LoadGlobal
_Py_Specialize_LoadSuperAttr
_Py_Specialize_Send
_Py_Specialize_StoreAttr
_Py_Specialize_StoreSubscr
_Py_Specialize_ToBool
_Py_Specialize_UnpackSequence
_Py_call_instrumentation
_Py_call_instrumentation_instruction
_Py_call_instrumentation_line
get_exception_handler

I suspect these are being inlined because we use unit tests for collecting PGO data. These should be executed relatively infrequently for "normal" Python code. We should mark them as noinline.

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

Linked PRs

• gh-132336: Mark a few "slow path" functions used by the interpreter loop as noinline #132337
• gh-132336: Only mark slow path functions used by the interpreter loop as noinline in free-threaded builds #132424

[thesamesam]

The cold attribute may be a fit here.

[Yhg1s]

As I understand it, cold is ignored when using PGO, at least in GCC, so I don't think that would have the desired effect.

[thesamesam]

Ah, you're right:

When profile feedback is available, via -fprofile-use, cold functions are automatically detected and this attribute is ignored.

[markshannon]

What we do or do not want to inline depends on whether it is a tail-calling build or not.

For the tailcalling, many of the above function are on the hot path of the instruction they are part of.
For example, get_exception_handler is only called in one place, it is on the hot/only path of exception_unwind. The instrumentation functions in the INSTRUMENTED instruction, likewise.

For the normal build, the size of PyEval_EvalDefault is the main concern and maybe we should be choosing functions based on size, as much as whether they are cold.

Maybe using better profiles is the way to do this?

[mpage]

What we do or do not want to inline depends on whether it is a tail-calling build or not.

This is a good point.

Maybe using better profiles is the way to do this?

I think we should do both. For the non-tail-calling interpreter we should mark the specialization functions as noinline, since that's more reliable than trying to coax a black box into doing what we want. I also think we should continue marking the instrumentation and exception handling functions as noinline since we know that those should be rare / exceptional paths.