gh-126618: Fix representation of itertools.count(sys.maxsize)

[picnixz]

Backports for 3.12 will be manual since the free-threaded code will not be needed.

• Issue: Incorrect representation of itertools.count(sys.maxsize) #126618

[picnixz]

Converting to draft until the CI failures are fixed by #126617.

[skirpichev]

Here is a more simple workaround:

diff --git a/Modules/itertoolsmodule.c b/Modules/itertoolsmodule.c
index 1201fa0949..a00b867d73 100644
--- a/Modules/itertoolsmodule.c
+++ b/Modules/itertoolsmodule.c
@@ -3415,7 +3418,7 @@ count_next(countobject *lz)
 static PyObject *
 count_repr(countobject *lz)
 {
-    if (lz->cnt != PY_SSIZE_T_MAX)
+    if (lz->long_cnt == NULL)
         return PyUnicode_FromFormat("%s(%zd)",
                                     _PyType_Name(Py_TYPE(lz)), lz->cnt);
 

But, perhaps, your solution is better, as it ensure that either:

assert(cnt != PY_SSIZE_T_MAX && long_cnt == NULL && long_step==1)

or

assert(cnt == PY_SSIZE_T_MAX && long_cnt != NULL && long_step != NULL)

kept.

[picnixz]

Yea, I wanted to keep those invariants but if performances matter more, we may use your solution (it would reduce the number of operations to perform in general).

[rhettinger]

Is there a simpler fix? I was expecting a much more minor edit to just the repr code.

[picnixz]

Is there a simpler fix? I was expecting a much more minor edit to just the repr code.

Sergey's patch could be the solution if you don't mind having the invariants invalid.

[skirpichev]

Perhaps, here is slightly another alternative to this fix and #126617:

diff --git a/Modules/itertoolsmodule.c b/Modules/itertoolsmodule.c
index 78fbdcdf77..6ca93294b6 100644
--- a/Modules/itertoolsmodule.c
+++ b/Modules/itertoolsmodule.c
@@ -3291,9 +3291,6 @@ itertools_count_impl(PyTypeObject *type, PyObject *long_cnt,
                 PyErr_Clear();
                 fast_mode = 0;
             }
-            else if (cnt == PY_SSIZE_T_MAX) {
-                fast_mode = 0;
-            }
         }
     } else {
         cnt = 0;
@@ -3325,7 +3322,7 @@ itertools_count_impl(PyTypeObject *type, PyObject *long_cnt,
     else
         cnt = PY_SSIZE_T_MAX;
 
-    assert((cnt != PY_SSIZE_T_MAX && long_cnt == NULL && fast_mode) ||
+    assert((long_cnt == NULL && fast_mode) ||
            (cnt == PY_SSIZE_T_MAX && long_cnt != NULL && !fast_mode));
     assert(!fast_mode ||
            (PyLong_Check(long_step) && PyLong_AS_LONG(long_step) == 1));
@@ -3418,7 +3415,7 @@ count_next(countobject *lz)
 static PyObject *
 count_repr(countobject *lz)
 {
-    if (lz->cnt != PY_SSIZE_T_MAX)
+    if (lz->long_cnt == NULL)
         return PyUnicode_FromFormat("%s(%zd)",
                                     _PyType_Name(Py_TYPE(lz)), lz->cnt);
 

This reverts fix from #126617 and adjusts assert instead. This will require also documentation change ("Counting logic and invariants" comment).

[skirpichev]

It seems, CI failure is related to the patch.

Alternative solution (from the thread) is presented in #127048.

[picnixz]

The CI failure is in importlib so I don't really know but I'll have a look at your patch when I am back.

[skirpichev]

The CI failure is in importlib so I don't really know

Yes, it's not obvious how this failure is triggered. But your code alters free-threading code - and that failure clearly belongs to it, I did several updates.

[picnixz]

Closing in favor of #127048.