The EPS encoder wasn't part of Gohlke's test suite, so the previous "fixes"
there were only expected syntactic ones. This gives a cleaner fix to the
encoder.
The decoder doesn't work in round-trip due to a missing eps_decoder method
on the core module, but it's clear it worked at some point.
This is Gohlke's fix for two issues: negative indexes on paths were not
resolved to the correct index, and path slicing didn't work.
His fix for slicing is related to the one found at:
http://renesd.blogspot.com/2009/07/python3-c-api-simple-slicing-sqslice.html
With this commit, all 79 tests (82 minus the three skipped ones) run
successfully on Python 2.6.8, Python 2.7.3rc2, and Python 3.2.3.
There are two main issues fixed with this commit:
* bytes vs. str: All file, image, and palette data are now handled as
bytes. A new _binary module consolidates the hacks needed to do this
across Python versions. tostring/fromstring methods have been renamed to
tobytes/frombytes, but the Python 2.6/2.7 versions alias them to the old
names for compatibility. Users should move to tobytes/frombytes.
One other potentially-breaking change is that text data in image files
(such as tags, comments) are now explicitly handled with a specific
character encoding in mind. This works well with the Unicode str in
Python 3, but may trip up old code expecting a straight byte-for-byte
translation to a Python string. This also required a change to Gohlke's
tags tests (in Tests/test_file_png.py) to expect Unicode strings from
the code.
* True div vs. floor div: Many division operations used the "/" operator
to do floor division, which is now the "//" operator in Python 3. These
were fixed.
As of this commit, on the first pass, I have one failing test (improper
handling of a slice object in a C module, test_imagepath.py) in Python 3,
and three that that I haven't tried running yet (test_imagegl,
test_imagegrab, and test_imageqt). I also haven't tested anything on
Windows. All but the three skipped tests run flawlessly against Pythons
2.6 and 2.7.
int() is really now long() in py3k, but to avoid breaking 2.6/2.7, we leave
the integer types where they are and just map long to int in py3k.
Also, pretty proud of myself for finding an easy way of detecting py3k.
This is, I guess, a few things the Python devs were just fed up with.
* "while 1" is now "while True"
* Types are compared with isinstance instead of ==
* Sort a list in one go with sorted()
My own twist is to also replace type('') with str, type(()) with tuple,
type([]) with list, type(1) with int, and type(5000.0) with float.
In py3k, imports are absolute unless using the "from . import" syntax.
This commit also solves a recursive import between Image, ImageColor, and
ImagePalette by delay-importing ImagePalette in Image.
I'm not too keen on this commit because the syntax is ugly. I might go back
and prefer the prettier "from PIL import".
What's really going on is that map() and filter() return iterators in py3k.
I've just gone ahead and turned them all into list comprehensions, because
I find them much easier to read.
y.has_key(x) is gone (use x in y), and keys(), values(), items(), and
range() all return views.
Some iterables needed to be packed into lists, either because the code
expected a list (such as "range(256) * 3") or because the original
collection was being modified (automatic global declarations).
The Tiff ImageFileDictionary is a special case and will be dealt with in
another commit.
This gets the putdata test case to run correctly under 2.6/2.7. It fixes an
issue where the value 0xFFFFFFFF (which is long in old Python) isn't
recognized and putdata tries to parse it as a tuple.
The original fix comes from Christoph Gohlke. It was adapted to work in
both 2.* and 3.*.
Most of the differences are in tobytes/tostring naming and expected
behavior of the bytes() constructor. The latter was usually easy to fix
with the right bytes literal.
This is a good preview of what will have to happen in the Python 3 code.
This is Christoph Gohlke's test suite from his personal PIL package found
at http://www.lfd.uci.edu/~gohlke/pythonlibs/.
This is just to bring it in as a separate commit. Future commits will align
it with Pillow.
This commit also renames some functions from "fromstring" and the like to
"frombytes". I'll probably need to come back later and update any
references to "string," here or in the docs.
I also noticed that encode allocates some data for certain codecs, but
never frees them. That would be a good bug to fix. I fixed the one where it
outright stole a pointer from Python.
This commit:
* Adds Python 3 module initialization functions. I split out the main init
of each module into a static setup_module function.
* Adds a py3.h which unifies int/long in Python 3 and unicode/bytes in
Python 2. _imagingft.c unfortunately looks a little kludgy after this
because it was already using PyUnicode functions, and I had to mix and
match there manually.
With this commit, the modules all build successfully under Python 3.
What this commit does NOT do is patch all of the uses of PyArg_ParseTuple
and Py_BuildValue, which all need to be checked for proper use of bytes
and unicode codes. It also does not let selftest.py run yet, because there
are probably hundreds of issues to fix in the Python code itself.
Python 3 enables C's strict aliasing rules for the first time, which means
you need to be careful about the ways you reference pointers. Here, we're
using a char[4] as an INT32, so we cast between them using a union.
I'm pretty sure this preserves the intent of the code. HAVE_UNICODE is now
assumed, and PyString is only used if we're not in Py3k.
Since this is the only file that uses PyUnicode, I'm going to go ahead and
#define PyUnicode and PyBytes back to PyString for 2.6, and explicitly
change out every call so I have to check them all.
Other ports have taken advantage of the fact that Python 3 has wrappers for
the old buffer protocol, but there's a significant disadvantage: you can't
let the buffered object know when you're done with it.
Since Python 2.6 supports the new protocol, we just go ahead and move to
it.
This updates several Python type definitions and uses to bring us closer
to Python 3 compatibility. This includes:
* Replacing staticforward and statichere with static. These were a hack for
old compilers and are not supported/needed anymore.
* Using Py_TYPE() instead of ob_type; ob_type is hidden in Py3.
* Replacing getattr with getters/setters. getattr is sort-of supported in
Py3, but Py_FindMethod is not. So we just use the newer
methods/getsetters mechanisms and start using PyType_Ready everywhere.
* Use PyVarObject_HEAD_INIT for types, since types are PyVarObject.
* Use PyMODINIT_FUNC for module initialization functions.
There are some tab/space issues in this commit. I'm set for spaces; the
source is a little schizo.
