[Bug]: Assorted CPython compatibility divergences (15 items) #1291

Describe the bug

Found by differential testing against CPython 3.12. All 15 items reproduce on GraalPy 25.4.4.1.1, and CPython 3.12.13 and 3.13.16 give the same expected output for each. Every item under "Steps to reproduce" has a reproduction, both outputs and the cause, with a suggested fix folded underneath.

Operating system

Linux

CPU architecture

ARM64

GraalPy version

25.4.4.1.1 (Python 3.13.14); code read on master c1c5cef

JDK version

No response

Context configuration

No response

Steps to reproduce

1. classmethod objects are callable

cm = classmethod(lambda: 1)

ClassmethodCommonBuiltins.CallNode registers tp_call in the slots that classmethod shares with classmethod_descriptor (ClassmethodCommonBuiltins.java#L126-L135). In CPython PyClassMethod_Type has no tp_call (funcobject.c#L1139); only classmethod_descriptor is callable.

2. type() with a non-type base crashes when the error message is read

try:
    type('X', (3,), {})
except TypeError as e:
    print('TypeError:', e)

GraalPy:

com.oracle.truffle.api.dsl.UnsupportedSpecializationException: Unexpected values provided for TypeNodesFactory.GetTpNameNodeGen.Uncached@5009421e: [null, 3], [null,Integer]

CPython:

TypeError: metaclass conflict: the metaclass of a derived class must be a (non-strict) subclass of the metaclasses of all its bases

calculateMetaclass (TypeBuiltins.java#L401-L423) checks IsAcceptableBaseNode on each base before comparing metaclasses, and raises TYPE_IS_NOT_ACCEPTABLE_BASE_TYPE, whose %N needs a type, for the int 3. Formatting the message calls GetTpNameNode on a non-type, which has no specialization for it. CPython's _PyType_CalculateMetaclass (typeobject.c#L3057-L3089) compares Py_TYPE(base) first, so a non-type base always gives the metaclass conflict error. Bases such as len, lambda: 0 or functools.partial(len) crash the same way.

3. type() puts __module__ after the __dict__ and __weakref__ descriptors

CreateTypeNode.makeType (TypeNodes.java#L1865-L1876) creates the class, which installs the descriptors, and only then sets __module__. CPython sets __module__ (typeobject.c#L3674) before PyType_Ready adds the descriptors. A class statement is not affected because its namespace already holds __module__.

4. NotImplementedType has no __bool__

import warnings
warnings.simplefilter('error')

NotImplementedBuiltins (NotImplementedBuiltins.java#L67-L107) defines no nb_bool. CPython's notimplemented_bool (object.c#L1962-L1975) warns and returns true.

5. frozenset() returns a frozenset subclass instance unchanged

class F(frozenset):
    pass

frozensetIdentity (FrozenSetBuiltins.java#L104-L109) returns the argument for any PFrozenSet. CPython takes that shortcut only for an exact frozenset (setobject.c#L995).

6. repr() of the smallest subnormals is not the shortest string

reprDigits (FloatFormatter.java#L775-L778) takes its digits from Double.toString, which picks the closest decimal of length 1 or 2 when one digit would round-trip. CPython's float_repr uses the shortest string that round-trips (floatobject.c#L391-L394).

7. compile() raises TypeError for an out-of-range optimize

checkOptimize (BuiltinFunctions.java#L1137-L1142) raises TypeError. CPython raises ValueError (bltinmodule.c#L766-L770).

8. Three-argument pow() raises ValueError for operands complex cannot convert

ComplexBuiltins.PowerNode.error (ComplexBuiltins.java#L1192-L1198) raises as soon as a modulus is given. CPython's complex_pow (complexobject.c#L507-L517) converts both operands first and returns NotImplemented when one is not a number.

9. list subclass + list returns the subclass

import operator

class L(list):
    pass

ListBuiltins.ConcatNode allocates the result with the class from GetClassForNewListNode (ListNodes.java#L196-L204), which is the receiver's class. CPython's list_concat always builds an exact list (listobject.c#L531). The result skips L.__init__.

10. bytearray item assignment checks the index before the value

b = bytearray(b'x')

SetSubscriptNode.set and SetItemNode.set (ByteArrayBuiltins.java#L221-L263) bounds-check the index before the value is converted. CPython converts the value first (bytearrayobject.c#L606-L616, GH-91153).

11. math.gcd() and math.lcm() with one argument call abs() before __index__

import math

class Idx:
    def __index__(self):
        return 6

The one-argument specializations GcdNode.gcdOne and LcmNode.gcdOne (MathModuleBuiltins.java#L969-L975, #L1125-L1131) both compute index(abs(x)). CPython computes abs(index(x)) (mathmodule.c#L767-L773, #L842-L848).

12. itertools.combinations() and combinations_with_replacement() check r before the iterable

Both constructors (CombinationsBuiltins.java#L134-L139, CombinationsWithReplacementBuiltins.java#L109-L113) raise for r < 0 before converting the iterable. CPython builds the pool first (itertoolsmodule.c#L2577-L2583, #L2893-L2899).

13. itertools.count() keeps an int subclass start

CountNode.construct (CountBuiltins.java#L114-L133) stores start as passed. CPython's fast mode (itertoolsmodule.c#L4028-L4067) keeps an int start with step 1 as a C Py_ssize_t, so it produces exact ints.

14. functools.lru_cache() requires an integer typed

import functools

LruCacheWrapperBuiltins converts typed with ClinicConversion.Int (LruCacheWrapperBuiltins.java#L123). CPython parses it with p (_functoolsmodule.c#L1168).

15. bisect functions treat an explicit hi=-1 as an empty range

import bisect

GraalPy has no _bisect, so bisect keeps the pure-Python functions from the bundled bisect.py (bisect.py#L111-L114), which replace only hi is None with len(a) (#L36, #L89). CPython runs the C _bisect, where -1 is the default of hi and becomes len(a) (_bisectmodule.c#L60-L64, #L244-L248). CPython declined to change its pure-Python version for this (python/cpython#125889).

Expected behavior

The CPython column of each item. CPython 3.12.13 and 3.13.16 give the same output for all of them.

Stack trace

Additional context

No response