Feature or enhancement

Proposal:

At the moment, thePy_buffer view onto the emptybytearray() orarray.array(typecode) has arbitrary alignment. This is fine in C, since the buffers are length 0. I found on Linux x86-64 with at least CPython 3.14 and building 3.15a1 from source, the buffer backingbytearray is reliably aligned on an odd pointer. E.g., given the script code

importarrayfromctypesimport*classpy_buffer(Structure):_fields_= [        ("buf",c_void_p),        ("obj",py_object),        ("len",c_ssize_t),        ("itemsize",c_ssize_t),        ("readonly",c_int),        ("ndim",c_int),        ("format",c_char_p),        ("shape",POINTER(c_ssize_t)),        ("strides",POINTER(c_ssize_t)),        ("suboffsets",POINTER(c_ssize_t)),        ("internal",c_void_p),    ]defbuf_ptr(buffer):buf=py_buffer()assertnotpythonapi.PyObject_GetBuffer(py_object(buffer),byref(buf),0)ptr=buf.bufpythonapi.PyBuffer_Release(byref(buf))returnptrptr=buf_ptr(bytearray())print(f"bytearray()     :{ptr:#x}")ptr=buf_ptr(bytes())print(f"bytes()         :{ptr:#x}")fortin"bBwhHiIlLqQfd":ptr=buf_ptr(array.array(t))print(f"array.array('{t}'):{ptr:#x}")

I got, for example

bytearray()     : 0x586d06546869bytes()         : 0x586d064fe218array.array('b'): 0x7ba46a224b99array.array('B'): 0x7ba46a224b99array.array('w'): 0x7ba46a224b99array.array('h'): 0x7ba46a224b99array.array('H'): 0x7ba46a224b99array.array('i'): 0x7ba46a224b99array.array('I'): 0x7ba46a224b99array.array('l'): 0x7ba46a224b99array.array('L'): 0x7ba46a224b99array.array('q'): 0x7ba46a224b99array.array('Q'): 0x7ba46a224b99array.array('f'): 0x7ba46a224b99array.array('d'): 0x7ba46a224b99

Some other languages (Rust, for example) have stricter alignment requirements to produce slice-like object from a raw pointer, even if the number of elements you're allowed to read from it is 0. Concretely in Rust, any&[T] has to be backed by a non-null pointer that's aligned for typeT, even if the slice is empty.

It would be convenient for interop if the default empty buffers ofbytearray andarray.array are aligned for any type; so this fast-path default case doesn't need special handling when taking out direct views onto the data buffer.

I came across this innumpy/numpy#30062, where the Pickle protocol-5 implementation for an empty array causes an emptybytearray to be loaded from the pickle stream (assuming inline buffers) and directly loaded to back the Numpy array, which was detectable in a Rust extension module I own.

Has this already been discussed elsewhere?

This is a minor feature, which does not need previous discussion elsewhere

Links to previous discussion of this feature:

No response

Linked PRs

• gh-140557: Force alignment of emptybytearray andarray.array buffers #140559