When slicing or cloning a memoryview object that has been previously cast to change its format string, the new memoryview's format shares a pointer to the parent's format string, which may be deleted at any time.

This manifests as a corrupt format when using the new memoryview object, causing crashes or unexpected behavior.

Tested on:
Python 3.4.0 (default, Apr 11 2014, 13:05:11) [GCC 4.8.2] on linux
Python 3.4.1 (v3.4.1:c0e311e010fc, May 18 2014, 10:45:13) [MSC v.1600 64 bit (AMD64)] on win32
Python 3.5.0a0 (default:cb8606fc84df, Oct 18 2014, 14:55:44) [GCC 4.8.2] on linux

Fix memoryview object allocations to copy format string

Thanks for the detailed report!  Making a private copy of 'format' for
each memoryview generally sounds like the best solution.  However, format
strings can be arbitrarily large, so we'd need to store the copy in the
ob_array after shape, strides and suboffsets.

This of course would slow down memoryview creation in the general case.

Given that the disappearing format strings are only created during casting,
I think we can get away with a local solution (see patch).

New changeset e9c1fca50b46 by Stefan Krah in branch '3.4':
Issue #22668: Ensure that format strings survive slicing after casting.
https://hg.python.org/cpython/rev/e9c1fca50b46

New changeset 37112bd3dfb3 by Stefan Krah in branch 'default':
Closes #22668: Merge from 3.4.
https://hg.python.org/cpython/rev/37112bd3dfb3

The buildbots aren't happy.  This one errors in the added test:

http://buildbot.python.org/all/builders/System%20Z%20Linux%203.x/builds/2595/steps/test/logs/stdio

Another one had a MemoryError in lib2to3 tests.  Not sure what is going on, but the one above at least is related to this issue.

(That's not a very descriptive name for a test, by the way, though it did allow me to find the issue easily ;)

New changeset 9a4af12dcc9d by Stefan Krah in branch '3.4':
Issue #22668: Remove endianness assumption in test.
https://hg.python.org/cpython/rev/9a4af12dcc9d

New changeset da0ca7b1351f by Stefan Krah in branch 'default':
Issue #22668: Merge from 3.4.
https://hg.python.org/cpython/rev/da0ca7b1351f

Thanks, David.  The tests assumed little-endian, which is now fixed.

The MemoryError is sporadic and unrelated -- the OpenIndiana bot
often has system load issues.

I don't see any more related failures, please reopen if you do. :)