call_function() and do_call_core() functions of Python/ceval.c use C_TRACE() macro to call the profiler, but PyObject_Call() and similar functions like _PyObject_FastCallKeywords() don't.

Is it a bug or a deliberate choice for performance?

Since PyObject_Call() and similar functions have now fast paths, I don't think that it's still needed to have these fast paths in ceval.c too. But I kept fast paths in ceval.c because of C_TRACE().

Output of attached profiler_c_code.py example:
---
haypo@selma$ ./python profiler_c_code.py 
(<frame object at 0xf9d788>, 'call', None)
(<frame object at 0xf9d788>, 'return', 'h')
(<frame object at 0xf9d788>, 'call', None)
(<frame object at 0xf9d788>, 'return', 'e')
(<frame object at 0xf9d788>, 'call', None)
(<frame object at 0xf9d788>, 'return', 'l')
(<frame object at 0xf9d788>, 'call', None)
(<frame object at 0xf9d788>, 'return', 'l')
(<frame object at 0xf9d788>, 'call', None)
(<frame object at 0xf9d788>, 'return', 'o')
(<frame object at 0xf9d498>, 'c_call', <built-in function setprofile>)
---

Except of sys.setprofile(), the profiler doesn't see any C calls! list() and map() are simply "hidden".

So tools like coverage miss a lot of C call? Again, is it a compromise between performance and correctness, or just a regular bug?

I always assumed that enabling profiling only from the Python bytecode interpreter was a deliberate choice.

issue30990 is related

FYI, _lsprof uses PyObject_Call()
https://github.com/python/cpython/blob/ea68d83933e6de6cabfb115ec1b8888301947369/Modules/_lsprof.c#L120-L123

If PyObject_Call() trigger profiler, lsprof should avoid recursion.

I would prefer to wait with this issue until PEP 580 has been decided. If it's accepted, it will change function calling code a lot. As I wrote in PEP 580, I was planning to have a PEP on profiling anyway.

This issue has no activity for 3 years and was created 4 years ago. I close it.

stage: resolved

messages: + msg287402