[issue4580] slicing of memoryviews when itemsize != 1 is wrong

[Nick Coghlan]

Nick Coghlan <ncoghlan at gmail.com> added the comment:

Trying another way of getting the point across... if Py_buffer wasn't
defined in PEP 3118, there would have to be some *other* API there
whereby the memoryview implementation could ask an object for a
description of the data layout of the object's buffer.

It is Py_buffer which allows the implementation of memoryview to even be
contemplated.

The one point where the PEP is silent (and I think this might be what
you've been attempting to get at) is as to whether or not it is OK for a
client to *modify* the contents of the Py_buffer structure before
passing it back in to PyBuffer_Release.

And to that I would say that a robust type implementation shouldn't be
assuming *anything* about the contents of the Py_buffer passed to
PyBuffer_Release. If it needs to do more than decrement a reference
count when the buffer is released (e.g. releasing memory allocated for
shapes and strides information), then it should be doing its own
internal bookkeeping based on the address of the Py_buffer object
originally passed to PyObject_GetBuffer.

_______________________________________
Python tracker <report at bugs.python.org>
<http://bugs.python.org/issue4580>
_______________________________________