[cmaloney]

decimal.Decimal is a heap type that would like to use AC Vectorcall. Heap types have no static type object so switch from Py_Is to an assert that the shadowed tp_new / tp_init function is the expected one to validate the vectorcall is ony used with the exact function it is providing a "vector" option for.

Heap types can have zero-argument __new__ and __init__ so add support for that producing the same error message via a new helper _PyArg_NoPositionalStack().

--
Minified from GH-157968. This does not use Py_tp_token as that only exists on Heap objects and want both static and heap objects to work with the same code.

• Issue: Generate vectorcall code to parse arguments using Argument Clinic #87613

[skirpichev]

Has merge conflicts.

[skirpichev]

Maybe you include Decimal case in this pr?

[cmaloney]

@skirpichev uipdated to fix conflicts and include Decimal

[skirpichev]

With same benchmark setup as in #158207:

decimal_factorial: Mean +- std dev: [ref] 1.06 sec +- 0.02 sec -> [patch] 1.06 sec +- 0.01 sec: 1.00x faster
decimal_pi: Mean +- std dev: [ref] 1.90 sec +- 0.02 sec -> [patch] 1.87 sec +- 0.01 sec: 1.02x faster

Geometric mean: 1.01x faster

Decimal(0): Mean +- std dev: [ref-micro] 584 ns +- 7 ns -> [patch-micro] 513 ns +- 8 ns: 1.14x faster
Decimal(123): Mean +- std dev: [ref-micro] 584 ns +- 10 ns -> [patch-micro] 515 ns +- 6 ns: 1.13x faster
Decimal('123'): Mean +- std dev: [ref-micro] 743 ns +- 14 ns -> [patch-micro] 704 ns +- 11 ns: 1.06x faster

Geometric mean: 1.11x faster

CC @vstinner

[cmaloney]

cpython/Modules/_decimal/_decimal.c

Lines 3232 to 3263 in 5b55b50

	static PyObject *
	PyDecType_FromObjectExact(PyTypeObject *type, PyObject *v, PyObject *context)
	{
	decimal_state *state = get_module_state_from_ctx(context);
	if (v == NULL) {
	return PyDecType_FromSsizeExact(type, 0, context);
	}
	else if (PyDec_Check(state, v)) {
	return PyDecType_FromDecimalExact(type, v, context);
	}
	else if (PyUnicode_Check(v)) {
	return PyDecType_FromUnicodeExactWS(type, v, context);
	}
	else if (PyLong_Check(v)) {
	return PyDecType_FromLongExact(type, v, context);
	}
	else if (PyTuple_Check(v) || PyList_Check(v)) {
	return PyDecType_FromSequenceExact(type, v, context);
	}
	else if (PyFloat_Check(v)) {
	if (dec_addstatus(context, MPD_Float_operation)) {
	return NULL;
	}
	return PyDecType_FromFloatExact(type, v, context);
	}
	else {
	PyErr_Format(PyExc_TypeError,
	"conversion from %s to Decimal is not supported",
	Py_TYPE(v)->tp_name);
	return NULL;
	}
	}

There is quite a bit more performance that should be unlockable for decimal construction if we get so Argument Clinic can inline the type check tree and/or provide to the interpreter dispatch loop a set of typed overloads (reduces instruction count a lot...). For bytes inlining that makes VC go from 1.1x -> 1.6x over fastcall.

[vstinner]

LGTM if you revert the _decimal change.

[vstinner]

Please revert the _decimal change. I would prefer a dedicated change for that, with benchmarks.

[skirpichev]

BTW, there are benchmarks. IMO, it would be better to leave this suggestion for a separate pr: #160115 (comment)

[vstinner]

Well, I would still prefer to change AC first, and then change decimal in a separated PR.

Thanks for the benchmarks, I missed them.

[vstinner]

LGTM