Call Protocol¶
CPython supports two different calling protocols: tp_call and vectorcall.
The tp_call Protocol¶
Instances of classes that set tp_call are callable.
The signature of the slot is:
PyObject *tp_call(PyObject *callable, PyObject *args, PyObject *kwargs);
A call is made using a tuple for the positional arguments
and a dict for the keyword arguments, similarly to
callable(*args, **kwargs) in Python code.
args must be non-NULL (use an empty tuple if there are no arguments)
but kwargs may be NULL if there are no keyword arguments.
This convention is not only used by tp_call:
tp_new and tp_init
also pass arguments this way.
To call an object, use PyObject_Call() or another
call API.
The Vectorcall Protocol¶
Added in version 3.9.
The vectorcall protocol was introduced in PEP 590 as an additional protocol for making calls more efficient.
As rule of thumb, CPython will prefer the vectorcall for internal calls
if the callable supports it. However, this is not a hard rule.
Additionally, some third-party extensions use tp_call directly
(rather than using PyObject_Call()).
Therefore, a class supporting vectorcall must also implement
tp_call.
Moreover, the callable must behave the same
regardless of which protocol is used.
The recommended way to achieve this is by setting
tp_call to PyVectorcall_Call().
This bears repeating:
Warning
A class supporting vectorcall must also implement
tp_call with the same semantics.
Changed in version 3.12: The Py_TPFLAGS_HAVE_VECTORCALL flag is now removed from a class
when the class’s __call__() method is reassigned.
(This internally sets tp_call only, and thus
may make it behave differently than the vectorcall function.)
In earlier Python versions, vectorcall should only be used with
immutable or static types.
A class should not implement vectorcall if that would be slower than tp_call. For example, if the callee needs to convert the arguments to an args tuple and kwargs dict anyway, then there is no point in implementing vectorcall.
Classes can implement the vectorcall protocol by enabling the
Py_TPFLAGS_HAVE_VECTORCALL flag and setting
tp_vectorcall_offset to the offset inside the
object structure where a vectorcallfunc appears.
This is a pointer to a function with the following signature:
-
typedef PyObject *(*vectorcallfunc)(PyObject *callable, PyObject *const *args, size_t nargsf, PyObject *kwnames)¶
- Part of the Stable ABI since version 3.12.
callable is the object being called.
- args is a C array consisting of the positional arguments followed by the
values of the keyword arguments. This can be NULL if there are no arguments.
- nargsf is the number of positional arguments plus possibly the
PY_VECTORCALL_ARGUMENTS_OFFSETflag. To get the actual number of positional arguments from nargsf, usePyVectorcall_NARGS().
- kwnames is a tuple containing the names of the keyword arguments;
in other words, the keys of the kwargs dict. These names must be strings (instances of
stror a subclass) and they must be unique. If there are no keyword arguments, then kwnames can instead be NULL.
-
PY_VECTORCALL_ARGUMENTS_OFFSET¶
- Part of the Stable ABI since version 3.12.
If this flag is set in a vectorcall nargsf argument, the callee is allowed to temporarily change
args[-1]. In other words, args points to argument 1 (not 0) in the allocated vector. The callee must restore the value ofargs[-1]before returning.For
PyObject_VectorcallMethod(), this flag means instead thatargs[0]may be changed.Whenever they can do so cheaply (without additional allocation), callers are encouraged to use
PY_VECTORCALL_ARGUMENTS_OFFSET. Doing so will allow callables such as bound methods to make their onward calls (which include a prepended self argument) very efficiently.Added in version 3.8.
To call an object that implements vectorcall, use a call API
function as with any other callable.
PyObject_Vectorcall() will usually be most efficient.
Recursion Control¶
When using tp_call, callees do not need to worry about
recursion: CPython uses
Py_EnterRecursiveCall() and Py_LeaveRecursiveCall()
for calls made using tp_call.
For efficiency, this is not the case for calls done using vectorcall: the callee should use Py_EnterRecursiveCall and Py_LeaveRecursiveCall if needed.
Vectorcall Support API¶
-
Py_ssize_t PyVectorcall_NARGS(size_t nargsf)¶
- Part of the Stable ABI since version 3.12.
Given a vectorcall nargsf argument, return the actual number of arguments. Currently equivalent to:
(Py_ssize_t)(nargsf & ~PY_VECTORCALL_ARGUMENTS_OFFSET)
However, the function
PyVectorcall_NARGSshould be used to allow for future extensions.Added in version 3.8.
-
vectorcallfunc PyVectorcall_Function(PyObject *op)¶
If op does not support the vectorcall protocol (either because the type does not or because the specific instance does not), return NULL. Otherwise, return the vectorcall function pointer stored in op. This function never raises an exception.
This is mostly useful to check whether or not op supports vectorcall, which can be done by checking
PyVectorcall_Function(op) != NULL.Added in version 3.9.
-
PyObject *PyVectorcall_Call(PyObject *callable, PyObject *tuple, PyObject *dict)¶
- Part of the Stable ABI since version 3.12.
Call callable’s
vectorcallfuncwith positional and keyword arguments given in a tuple and dict, respectively.This is a specialized function, intended to be put in the
tp_callslot or be used in an implementation oftp_call. It does not check thePy_TPFLAGS_HAVE_VECTORCALLflag and it does not fall back totp_call.Added in version 3.8.
Object Calling API¶
Various functions are available for calling a Python object. Each converts its arguments to a convention supported by the called object – either tp_call or vectorcall. In order to do as little conversion as possible, pick one that best fits the format of data you have available.
The following table summarizes the available functions; please see individual documentation for details.
Function |
callable |
args |
kwargs |
|---|---|---|---|
|
tuple |
dict/ |
|
|
— |
— |
|
|
1 object |
— |
|
|
tuple/ |
— |
|
|
format |
— |
|
obj + |
format |
— |
|
|
variadic |
— |
|
obj + name |
variadic |
— |
|
obj + name |
— |
— |
|
obj + name |
1 object |
— |
|
|
vectorcall |
vectorcall |
|
|
vectorcall |
dict/ |
|
arg + name |
vectorcall |