PyPycLoader can't read cached .pyc files

[brett.cannon]

http://bugs.python.org/review/15030/diff/5254/Lib/importlib/abc.py
File Lib/importlib/abc.py (right):

http://bugs.python.org/review/15030/diff/5254/Lib/importlib/abc.py#newcode308
Lib/importlib/abc.py:308: bytecode = data[12:]  # Python 3.3 added file size
field to header
Line is too long. Move the comment to the line above.

Also, change comment to "Python 3.3 added a file size field, so we might need to
ignore 4 bytes for that."

http://bugs.python.org/review/15030/diff/5254/Lib/importlib/abc.py#newcode310
Lib/importlib/abc.py:310: return marshal.loads(bytecode)
Don't bother with slicing the bytecode off of data again. Instead, to make it
more obvious what has occurred, do bytecode[4:].

[brett.cannon]

Almost there! Very minor tweaks.

http://bugs.python.org/review/15030/diff/5258/Lib/importlib/abc.py
File Lib/importlib/abc.py (right):

http://bugs.python.org/review/15030/diff/5258/Lib/importlib/abc.py#newcode312
Lib/importlib/abc.py:312: return marshal.loads(bytecode)
Drop the assignment above and make this:

return marshal.loads(bytecode[4:])

http://bugs.python.org/review/15030/diff/5258/Lib/importlib/abc.py#newcode313
Lib/importlib/abc.py:313: except EOFError:  # If we get EOF, raise original
ValueError exception
Move the comment down a line (inside the 'except' clause), make sure it wraps at
80 characters, and change it to "EOF suggests that there was bad bytecode to
begin with, so re-raise original exception."

[brett.cannon]

There's a bug in the calculation of the file size field and the test is doing
more than it needs to (and I do realize, Marc, that the test is originally from
Ronan and not you).

http://bugs.python.org/review/15030/diff/5266/Lib/importlib/abc.py
File Lib/importlib/abc.py (right):

http://bugs.python.org/review/15030/diff/5266/Lib/importlib/abc.py#newcode326
Lib/importlib/abc.py:326: data.extend(_bootstrap._w_long(len(bytecode)))
It is the length of the source, not the bytecode; see
Lib/importlib/_bootstrap.py:SourceLoader.get_code() for how importlib already
does it.

http://bugs.python.org/review/15030/diff/5266/Lib/importlib/test/source/test_...
File Lib/importlib/test/source/test_abc_loader.py (right):

http://bugs.python.org/review/15030/diff/5266/Lib/importlib/test/source/test_...
Lib/importlib/test/source/test_abc_loader.py:462: name = 'foo'
Module uses 'mod' for the module name in the tests.

http://bugs.python.org/review/15030/diff/5266/Lib/importlib/test/source/test_...
Lib/importlib/test/source/test_abc_loader.py:463: with
source_util.create_modules(name) as fpath:
I think this is the wrong approach. PyPycLoaderMock provides everything
necessary to avoid writing bytecode.  You can make it return bytecode for
get_data() that has the size field to verify it parses it properly, and you can
introspect what write_bytecode() was called with to make sure the field was set.
That way you avoid any writing of files and thus slowing the test down (it still
requires the writes_bytecode_files flag, though, else get_code() won't write the
fake bytecode out).