I checked how_Py_rotateright_uintptr() is compiled on x86-64:

• clang -O3 with__builtin_rotateright64():ror rax, cl
• gcc -O3 with(x >> bits) | (x << (8 * SIZEOF_UINTPTR_T - bits)):ror rbx,cl

As expected, theROR instruction is used in both cases.

I usedgdb: in Python, I runimport _testinternalcapi, and then in gdb I typedisassemble rotateright_uintptr.

For that, I made the following change. Thecheck_rotateright_uintptr() cannot be used to check the machine code: sincecheck_rotateright_uintptr() arguments are constant, GCC and clang just pre-compute the function result at build. The funciton is no longer called at runtime.

Change:

diff --git a/Modules/_testinternalcapi.c b/Modules/_testinternalcapi.cindex 2d888e0ae2c..a08b3dfb168 100644--- a/Modules/_testinternalcapi.c+++ b/Modules/_testinternalcapi.c@@ -290,6 +290,23 @@ test_rotateright_uintptr(PyObject *self, PyObject *Py_UNUSED(args)) }+static PyObject*+rotateright_uintptr(PyObject *self, PyObject *args)+{+    PyObject *obj;+    int bits;+    if (!PyArg_ParseTuple(args, "O!i", &PyLong_Type, &obj, &bits)) {+        return NULL;+    }+    unsigned long long x = PyLong_AsUnsignedLongLong(obj);+    if (x == (unsigned long long)-1 && PyErr_Occurred()) {+        return NULL;+    }+    uintptr_t y = _Py_rotateright_uintptr((uintptr_t)x, bits);+    return PyLong_FromUnsignedLongLong(y);+}++ #define TO_PTR(ch) ((void*)(uintptr_t)ch) #define FROM_PTR(ptr) ((uintptr_t)ptr) #define VALUE(key) (1 + ((int)(key) - 'a'))@@ -1689,6 +1706,7 @@ static PyMethodDef module_functions[] = {     {"test_popcount", test_popcount, METH_NOARGS},     {"test_bit_length", test_bit_length, METH_NOARGS},     {"test_rotateright_uintptr", test_rotateright_uintptr, METH_NOARGS},+    {"rotateright_uintptr", rotateright_uintptr, METH_VARARGS},     {"test_hashtable", test_hashtable, METH_NOARGS},     {"get_config", test_get_config, METH_NOARGS},     {"set_config", test_set_config, METH_O},

I checked how _Py_rotateright_uintptr() is compiled on x86-64

Adding_Py_rotateright_uintptr() static inline function added too much code, and it's not needed: current code is already compiled to the expected most efficient x86-64ROR instruction. I removed the_Py_rotateright_uintptr() function.

Changes:

• PR rebased on main.
• Rename the function to Py_HashPointer() followinggh-111545: Add PyHash_Double() function #112095 (comment) discussion.
• Keep _Py_HashPointerRaw() as a static inline function to avoid code duplication.
• No longer use SIZEOF_UINTPTR_T, but sizeof().

@encukou: Please review the updated PR.

@serhiy-storchaka: Do you want to review this change?

I rebased the PR on main to get a NoGIL fix.

@encukou:

(And that the rest of the C-API WG won't have major objections, but Idon't think that's likely.)

PEP 731 says:

The working group may operate primarily through reviews of GitHub issues and PRs.

Are you suggesting that the C API cannot be modified anymore unless whole C API Working Group approve a change? Or are you suggesting that the C API Working Group must review all C API additions? I don't think that was the plan. Well, the plan is not well defined.

Are you suggesting that the C API cannot be modified anymore unless whole C API Working Group approve a change?

No, I'm saying that the others might disagree -- and that I don't think it's very likely that they will disagree.
I'd optimistically merge this (edit: after checking with Serhiy), but be prepared to fix it up or revert.

Adding new C API should include the WG up until we've developed and agreed upon the guidelines for adding new C API.

@encukou createdcapi-workgroup/decisions#1

@iritkatriel:

I'm missing some mention of the use case for this (not necessarily here). The issue talks about _Py_HashDouble, there is no mention of Py_HashPointer.

ProposedPy_HashDouble() API requires the caller to callPy_HashPointer(obj) for NaN: see#112449

Moreover, I removed the private function_Py_HashPointer() in Python 3.13 alpha1 and it broke a few projects. A code search on PyPI top 5,000 projects found 4 projects:

• cffi (1.16.0)
• gmpy2 (2.1.5)
• numba (0.58.1)
• tree_sitter (0.20.4)

Finally, igraph-0.11.2 contains a copy of the private_Py_HashPointer() function insrc/_igraph/pyhelpers.c.

I merged my PR. The API was approved by the C API Working Group:capi-workgroup/decisions#1