Skip to content

gh-162350: Use AC for array.array constructor - #162419

Open
skirpichev wants to merge 5 commits into
python:mainfrom
skirpichev:array.array-vc/162350
Open

skirpichev wants to merge 5 commits into
python:mainfrom
skirpichev:array.array-vc/162350

Conversation

@skirpichev

@skirpichev skirpichev commented Nov 22, 2026 •

Copy link
Copy Markdown
Member

This fixes it's signature and adds also support for Vectorcall protocol.

This fixes it's signature and adds also support for Vectorcall protocol.
@skirpichev

skirpichev commented Nov 22, 2026 •

Copy link
Copy Markdown
Member Author

Microbenchmark (default options, gcc 14.2):

Benchmark ref patch
array of 0 elts 5.08 us 3.86 us: 1.32x faster
array of 1 elts 6.23 us 5.02 us: 1.24x faster
array of 2 elts 6.51 us 5.29 us: 1.23x faster
array of 3 elts 6.79 us 5.57 us: 1.22x faster
array of 100 elts 30.8 us 30.9 us: 1.00x slower
Geometric mean (ref) 1.20x faster
Details
from array import array
import pyperf
runner = pyperf.Runner()
tp = 'i'
for xs in [[], [1], [1, 2], [1, 2, 3], list(range(100))]:
    s = f"array of {len(xs)} elts"
    runner.bench_func(s, array, tp, xs)
array of 0 elts: Mean +- std dev: [ref] 5.08 us +- 0.13 us -> [patch] 3.86 us +- 0.02 us: 1.32x faster
array of 1 elts: Mean +- std dev: [ref] 6.23 us +- 0.04 us -> [patch] 5.02 us +- 0.23 us: 1.24x faster
array of 2 elts: Mean +- std dev: [ref] 6.51 us +- 0.40 us -> [patch] 5.29 us +- 0.03 us: 1.23x faster
array of 3 elts: Mean +- std dev: [ref] 6.79 us +- 0.27 us -> [patch] 5.57 us +- 0.04 us: 1.22x faster
array of 100 elts: Mean +- std dev: [ref] 30.8 us +- 0.2 us -> [patch] 30.9 us +- 0.1 us: 1.00x slower

Geometric mean: 1.20x faster

No significant difference for scimark benchmark.

@cmaloney cmaloney left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Slightly outside these lines make_array currently exists and uses array_new which means an unnecessary tuple construction on pickle related code paths. Be nice to have the two make_array callers now just directly call array_new_impl (both have the right variables as locals already) which means make_array can go away entirely.

cpython/Modules/arraymodule.c

Lines 2170 to 2209 in 02fe78f

/* Forward declaration. */
static PyObject *array_new(PyTypeObject *type, PyObject *args, PyObject *kwds);
/*
* Internal: This function wraps the array constructor--i.e., array_new()--to
* allow the creation of array objects from C code without having to deal
* directly the tuple argument of array_new(). The typecode argument is a
* string, like "i" or "f" for example, representing an array
* type code. The items argument is a bytes or a list object from which
* contains the initial value of the array.
*
* On success, this functions returns the array object created. Otherwise,
* NULL is returned to indicate a failure.
*/
static PyObject *
make_array(PyTypeObject *arraytype, const char *typecode, PyObject *items)
{
PyObject *new_args;
PyObject *array_obj;
PyObject *typecode_obj;
assert(arraytype != NULL);
assert(items != NULL);
typecode_obj = PyUnicode_FromString(typecode);
if (typecode_obj == NULL)
return NULL;
new_args = _PyTuple_FromPairSteal(typecode_obj, Py_NewRef(items));
if (new_args == NULL) {
return NULL;
}
array_obj = array_new(arraytype, new_args, NULL);
Py_DECREF(new_args);
if (array_obj == NULL)
return NULL;
return array_obj;
}

Comment thread Modules/arraymodule.c Resolved
Comment thread Modules/arraymodule.c Resolved
Comment thread Modules/arraymodule.c Resolved
@cmaloney

cmaloney commented Nov 22, 2026 •

Copy link
Copy Markdown
Contributor

A couple small changes with this but I think they're all reasonable:

Call Before After
array() array() takes at least 1 argument (0 given) array expected at least 1 argument, got 0
array('b',1,2) array() takes at most 2 arguments (3 given) array expected at most 2 arguments, got 3
array(typecode='b') array.array() takes no keyword arguments array() takes no keyword arguments
inspect.signature(array) ValueError (typecode, initializer=(), /)
__doc__ first line array(typecode [, initializer]) -> array Create array object.

@skirpichev

Copy link
Copy Markdown
Member Author

I'll run scimark benchmark on this, but not expect too much, maybe few percents.

Using AC for constructor seems to be a good idea anyway, due to removal of the internal helper and fixed signature.

@cmaloney

Copy link
Copy Markdown
Contributor

👍, I tried making the base rules for "what makes a good refactor" that: cases where it's just adding @vectorcall and setting the slot have a higher performance requirement than cases where it's refactoring towards Argument Clinic. As long as we're doing the rewrite/refactor also adding vectorcall at the same time is a small extra cost with a reasonable micro-benchmark performance improvement.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants