homepage

Finally, here is my C implementation of BytesIO. The code is well tested
and include the proper unit tests. The only remaining issues are:
 - The behavior of the close() method.
 - The failure of test_profile and test_cProfile.
Currently, I have no idea how to fix the tests for the profilers. The
weird thing is both pass when run with:
 % ./python Lib/test/regrtest.py -R: test_profile

Any chance of uploading a single patch that contains all these changes?
The profile tests often fail when io.py changes because they happen to
depend on "golden output" which includes line numbers of code in io.py
that happens to be traced during the test. I don't know why -R: makes
it pass but it could be that it doesn't compare the golden output.

So, here's one big patch. I have updated the behavior of close(), so that 
> The profile tests often fail when io.py changes because they happen to
> depend on "golden output" which includes line numbers of code in io.py
> that happens to be traced during the test. 
Yes, I knew that. But, how can I fix the test so that it passes even if
_bytesio is not available?
Oh, one more thing. In the misc fixes for io.py, I added a checkClosed
in IOBase.readline(). As a side-effect, this make __next__ raises a
ValueError, instead of StopIteration. Is that correct? 
>>> f = open("load.py")
[45681 refs]
>>> next(f)
'import sys\n'
[45700 refs]
>>> f.close()
[45703 refs]
>>> next(f)
Traceback (most recent call last):
 File "<stdin>", line 1, in <module>
 File "/home/alex/src/python.org/py3k/Lib/io.py", line 1440, in __next__
 line = self.readline()
 File "/home/alex/src/python.org/py3k/Lib/io.py", line 1449, in readline
 raise ValueError("read from closed file")
ValueError: read from closed file
[45703 refs]

[grrr, I eat my words]
> So, here's one big patch. I have updated the behavior of close(), so that 
... it matches the behavior of 2.x.
> As a side-effect, this make __next__ raises a
ValueError, instead of StopIteration.
... when the file is closed.

I got a patch that also fixes the profiler tests. That was easy after
all. :-)

Here is a patch against the latest trunk (r61578) that includes the
accelerator module of io.BytesIO with its test suite. The patch also
changes the behavior of the truncate method to imply a seek(). Please
review!

Bumping priority so that this gets reviewed before next release.

From a quick glance: I may be wrong, but it looks like in
bytesio_writelines(), you'll have some reference leaks if the `while`
loop exits early because bytesio_write() returns NULL; `item` and `it`
should be decred'ed before returning.

Yes, that was a leak. It should be fixed now.
Merci Antoine!

Alexandre, is it normal that the unit tests look much less complete in
your latest patch than they were in the previous one?
Also, they don't test the Python fallback implementation anymore.

Oops, I forgot to include the unit tests, with "svn add", when I made
the diff.

There is a small bug in the last patch I posted. The unit tests assumed
(wrongly) that accelerator module for io.StringIO (i.e., _stringio) was
present.

Ok I took a detailed look at _bytesio.c.
In write_bytes() there is the following resizing logic:
 if (self->pos + len > self->string_size) {
 if (resize_buffer(self, self->pos + len) < 0)
 return -1;
 }
Replacing `self->string_size` with `self->buf_size` should avoid some
spurious reallocations.
For some reason, using the help() function on a BytesIO instance does
not display the class help.
Overall, the new module looks fine. I can't say anything about the io.py
or _fileio.c fixes.

One last thing: you probably noticed it, but there is one test which
needs correcting in test_StringIO. It's due to the fact that calling
next() on a closed BytesIO object raises ValueError now rather than
StopIteration.

I think that check in write_bytes() is a guard to avoid resize_buffer()
from truncating the string stored in the buffer. However, I am not sure
if it is still necessary.
I don't know why help() doesn't work on BytesIO instances. But, the
problem doesn't seems to be in my code, since `help(sys.stdin)` doesn't
work either.
Finally regarding test_StringIO, see msg59460. Anyway, test_StringIO is
superseded by test_memoryio included my patch and should be removed from
the stdlib.

Well, by construction self->buf_size should always be greater than
self->string_size, so I don't think there's any risk of truncating
anything here.
I tried the change, and the tests still ran fine.

I'm going to review the patch later. How are we going to back port the
stuff to 2.6?

Hey Alexandre!
The latest patch doesn't apply cleanly. Please provide a new patch.

Here's an updated patch.

Oops, I forgot again to "svn add" the new files.

Do I need to look at this, or is the review going well without my
interference?

Hi Guido,
The patch changes a minor things to io.py to allow io.BytesIO to pass my
test suite, so you may want to check it out. Other than that, I think
the review is going fine.

Since nothing happened for almost one month here, perhaps it would be
better to merge Alexandre's changes as they are? We'll see quickly if
unexpected things happen :-)
It would also allow to progress on other issues, e.g. #2523.
Also, I stand by the suggestion I made about the resizing logic, but it
shouldn't be a showstopper either. We can improve that later.

Alexandre, you have until the end of the week to commit it or someone else 
will. =) I don't want this going in so close to the alphas, but it should 
go in almost immediately afterwards.

By the way, is someone already working on a C-accelerated TextIO?

Patch committed in r62778.
Antoine wrote:
> Also, I stand by the suggestion I made about the resizing logic, but it
> shouldn't be a showstopper either. We can improve that later.
I made your suggested change to the resizing logic. Thanks again for the
review! 
> By the way, is someone already working on a C-accelerated TextIO?
Yes. I was waiting to commit the C optimization for io.BytesIO, before
continuing the one for io.StringIO. The implementation is mostly
done---it just needs to be updated with respect to the recent changes to
TextIO.