[Bug]: PEP 688 `__buffer__` missing on builtin types

[jseop-lim]

Describe the bug

GraalPy registers no __buffer__ or __release_buffer__ on its managed types. So dir(b"") returns 77 names where CPython 3.12 returns 78, bytearray().__buffer__(0) raises AttributeError, and isinstance(b"", collections.abc.Buffer) is False. The buffer protocol itself works; only the PEP 688 Python-level dunders are missing.

Operating system

Linux

CPU architecture

ARM64

GraalPy version

25.2.4 (Python 3.12.8); also reproduced on 25.0.2

JDK version

No response

Context configuration

No response

Steps to reproduce

import collections.abc

print(len(dir(b"")), "__buffer__" in dir(b""))
print(isinstance(b"", collections.abc.Buffer))
bytearray(b"hello").__buffer__(0)

GraalPy 25.2.4:

77 False
False
Traceback (most recent call last):
  File "/tmp/buf.py", line 5, in <module>
    bytearray(b"hello").__buffer__(0)
AttributeError: 'bytearray' object has no attribute '__buffer__'

Expected behavior

CPython 3.12.13 raises nothing, and __buffer__(0) returns a memoryview:

78 True
True

Stack trace

Additional context

Every builtin buffer type is affected, not just bytes. "__buffer__" in dir(t) / "__release_buffer__" in dir(t):

type	GraalPy 25.2.4	CPython 3.12.13
bytes	False / False	True / False
bytearray	False / False	True / True
memoryview	False / False	True / True
array.array	False / False	True / True
mmap.mmap	False / False	True / True

CPython's test_buffer.TestPythonBufferProtocol.test_call_builtins covers this case, calling both dunders on a bytearray (test_buffer.py#L4589-L4595). It is missing from unittest_tags/test_buffer.txt, which tags only the four cases built on user-defined classes, so CI never sees the gap.

Root Cause

Nothing in the Java core registers the dunder: git grep __buffer__ -- graalpython/com.oracle.graal.python/src returns no hits at 730597a0. bytes.__dict__ has no entry for it, and dir() only merges the MRO type dicts (ObjectBuiltins.java#L869 into TypeBuiltins.java#L1240), so it reports 77.

CPython synthesizes the wrapper in add_operators, which PyType_Ready calls to walk slotdefs and add a descriptor to tp_dict for every slot a type defines. The two slots are declared at Objects/typeobject.c#L9501-L9506, and bytes gets bf_getbuffer from bytes_as_buffer, installed as tp_as_buffer at #L2969:

static PyBufferProcs bytes_as_buffer = {
    (getbufferproc)bytes_buffer_getbuffer,
    NULL,
};

bf_releasebuffer is NULL there, so __release_buffer__ is not added. That is why the dir() difference is exactly one name.

The protocol underneath is already there: PBytesLike exports PythonBufferAccessLibrary (PBytesLike.java#L62-L64) and memoryview(b"ab") works. The C extension layer carries the same BUFSLOT definitions (cext/src/typeobject.c#L10461-L10466), but only for native types. Managed builtin types are the gap.

See PEP 688 and object.__buffer__.

Fix Suggestion

Each managed buffer type needs __buffer__, and __release_buffer__ only where the CPython counterpart defines bf_releasebuffer. bytes takes the first alone, the other four take both. Registering the dunders as builtins per type and synthesizing them at type initialization for every type exporting PythonBufferAccessLibrary, the way add_operators walks slotdefs, both seem workable; exporting that library does not by itself separate the two groups above, so which approach fits GraalPy's type setup is better judged by the maintainers.