[Bug]: tuple() does not validate `__len__` / `__length_hint__` before iterating

[jseop-lim]

Description

tuple(obj) starts iterating before validating __len__ / __length_hint__, so any exception raised by __getitem__ propagates instead of the TypeError / ValueError that CPython raises when the length hint is invalid. This is a CPython compatibility issue.

Root Cause

ConstructTupleNode.generic() obtains an iterator via PyObjectGetIter and immediately hands it to CreateStorageFromIteratorNode without passing a length hint. __len__ / __length_hint__ are never validated, so the Exception raised by __getitem__ during iteration propagates unchanged.

CPython's PySequence_Tuple instead calls PyObject_LengthHint(v, 10) before iteration, which validates the return value of __len__ / __length_hint__ and raises TypeError / ValueError first — before __getitem__ is ever invoked.

CPython reference: Objects/abstract.c#L91-L146, Objects/abstract.c#L2053

// PyObject_LengthHint validates __len__ / __length_hint__ before any iteration.
// __len__ path — a negative return value raises ValueError("__len__() should return >= 0").
if (_PyObject_HasLen(o)) {
    res = PyObject_Length(o);
    if (res < 0) {
        if (!_PyErr_ExceptionMatches(tstate, PyExc_TypeError)) {
            return -1;   // propagate non-TypeError (e.g. ValueError) as-is
        }
        _PyErr_Clear(tstate);
    } else {
        return res;
    }
}
// __length_hint__ path — a non-int return value raises TypeError.
if (!PyLong_Check(result)) {
    PyErr_Format(PyExc_TypeError, "__length_hint__ must be an integer, not %.100s",
        Py_TYPE(result)->tp_name);
    return -1;
}

Reproduction

class BadLengthHint:
    def __getitem__(self, index):
        raise Exception

    def __length_hint__(self):
        return None

tuple(BadLengthHint())

Output

GraalPy:

Exception

CPython:

TypeError: __length_hint__ must be an integer, not NoneType

Environment

• GraalPy 25.0.2 (Python 3.12.8)
• CPython v3.12.13
• OS: Debian 12

Additional context

The same GraalPy already does this correctly on the list(iterable) path (ListBuiltins.listIterable), which uses IteratorNodes.GetLength. Only the tuple() path is missing the pre-validation.

graalpython/graalpython/com.oracle.graal.python/src/com/oracle/graal/python/builtins/objects/list/ListBuiltins.java

Lines 267 to 279 in 4f02c3d

	static PNone listIterable(VirtualFrame frame, PList list, Object iterable,
	@Bind Node inliningTarget,
	// exclusive for truffle-interpreted-performance
	@Exclusive @Cached ClearListStorageNode clearStorageNode,
	@Cached IteratorNodes.GetLength lenNode,
	@Cached PyObjectGetIter getIter,
	@Cached CreateStorageFromIteratorNode storageNode) {
	clearStorageNode.execute(inliningTarget, list);
	int len = lenNode.execute(frame, inliningTarget, iterable);
	Object iterObj = getIter.execute(frame, inliningTarget, iterable);
	list.setSequenceStorage(storageNode.execute(frame, iterObj, len));
	return PNone.NONE;
	}

Applying the list()-style fix directly did not work because of the @Fallback(excludeForUncached = true) annotation on ConstructTupleNode.generic(), so this is being filed as an issue only.