Make sure that tp_as_xxx methods are never NULL in PyType_Ready()

[vstinner]

PyType_Ready() now sets type tp_as_xxx members to a structure filled of NULL if a member is NULL. It avoids checking if tp_as_xxx is NULL in Objects/abstract.c functions.

[vstinner]

I wrote a patch to run a microbenchmark on PyNumber_Add() with two integers (1+2):

Details

diff --git a/Modules/_testcapimodule.c b/Modules/_testcapimodule.c
index eb769294fd2..c388711fdf1 100644
--- a/Modules/_testcapimodule.c
+++ b/Modules/_testcapimodule.c
@@ -2879,6 +2879,47 @@ uptime_bsd(PyObject *Py_UNUSED(self), PyObject *Py_UNUSED(args))
 #endif
 
 
+static PyObject*
+bench_add(PyObject *Py_UNUSED(self), PyObject *args)
+{
+    Py_ssize_t loops;
+    if (!PyArg_ParseTuple(args, "n", &loops)) {
+        return NULL;
+    }
+
+    PyObject *one = PyLong_FromLong(1);
+    PyObject *two = PyLong_FromLong(1);
+    assert(one != NULL && two != NULL);
+
+    PyTime_t t1, t2;
+    (void)PyTime_PerfCounterRaw(&t1);
+    PyObject *sum;
+
+#define ADD() \
+        sum = PyNumber_Add(one, two); \
+        if (sum == NULL) { \
+            return NULL; \
+        } \
+        Py_DECREF(sum);
+
+    for (Py_ssize_t i=0; i < loops; i++) {
+        ADD();
+        ADD();
+        ADD();
+        ADD();
+        ADD();
+        ADD();
+        ADD();
+        ADD();
+        ADD();
+        ADD();
+    }
+
+    (void)PyTime_PerfCounterRaw(&t2);
+    return PyFloat_FromDouble(PyTime_AsSecondsDouble(t2 - t1));
+}
+
+
 static PyMethodDef TestMethods[] = {
     {"set_errno",               set_errno,                       METH_VARARGS},
     {"test_config",             test_config,                     METH_NOARGS},
@@ -2978,6 +3019,7 @@ static PyMethodDef TestMethods[] = {
 #ifdef HAVE_SYSCTLBYNAME
     {"uptime_bsd", uptime_bsd, METH_NOARGS},
 #endif
+    {"bench_add", bench_add, METH_VARARGS},
     {NULL, NULL} /* sentinel */
 };
 

Benchmark:

import pyperf
import _testcapi
runner = pyperf.Runner()
runner.bench_time_func('int + int', _testcapi.bench_add)

Result: Mean +- std dev: [isol_ref] 88.1 ns +- 1.4 ns -> [isol_change] 87.4 ns +- 0.0 ns: 1.01x faster.

Well... the difference is quite minor. So I'm not sure that the change is worth it.

[vstinner]

I also tried adding PyNumberMethods _as_number; to PyTypeObject and modify PyNumber_Add() to use it to avoid one pointer indirection. Again, there is no significant performance difference on PyNumber_Add(1, 1), so this change is not worth it. I prefer to close my PR.

This work would be easier to do if PyTypeObject structure was opaque, but it's part of the public C API, so it cannot be easily changed :-( We have to keep tp_as_number pointer (and check if it's not-NULL before using it, for now).

[ZeroIntensity]

The performance impact might be better than you think! See #149180.