Fixed the trivial import bug that prevented the Python 3 version of MPO
from running. On the way fixed the trivial C bug that prevented
Convert.c from compiling properly in a Mac OS X environment for a
Python 3 target.
1) Renamed USE_INLINE to PIL_USE_INLINE to avoid conflicts with
other headers/libraries.
2) Replace __WIN32__ and WIN32 with _WIN32
3) Don't define WIN32 when the compiler is MSVC but not on Windows
Why would you even...
4) Don't define strcasecmp if you're not even going to use it.
5) Don't include Windows.h with undefs for compilers newer than
1998 everywhere.
6) Don't surpress warnings for MSVC++ 4.0. People still using
MSVC++ 4.0 deserve it.
7) Don't include things that are already included in Windows.h
The previous version of the function didn't generate correct wide lines
of even width.
The most notable changes are:
* Make variable names far more descriptive about the process.
* Corrected the width calculation; we should deduct one pixel from the
width because the pixels at the center of the line doesn't count for
the triangles.
* Now we calculate *two* ratios, one for left/top displacement (dxmin)
and one for right/bottom (dxmax), this fix the behavior with lines
of even width.
It can probably be optimized.
The (previously refactored) polygon_generic function didn't draw
consistent polygons (equilateral polygons were not equilateral nor
symmetrical).
The most notable changes are:
* The horizontal edges are searched for when finding the polygon
boundaries, drawn and discarded from the edge list used to detect
intersections.
* The intersections are now checked and calculated from the current value
of the scan line (ymin) instead of in the middle (ymin + 0.5).
* Because the change in the scan line behavior, we should duplicate
the intersections in the maximum Y value of an edge or there will be
draw errors with concave and complex polygons.
* The rounds of the X coordinates in the hline function calls are
switched to draw the inner pixels.
* Removed the ugly micro-optimization of qsort at the end.
This implementation of the scan line algorithm may not be technically
correct, it's not optimized and it have problems with some edge cases,
like a wide line from (x0, y) to (x1, y + 1), therefore it should be
reviewed in the future.
The rounds used in the library are towards positive/negative infinity.
Since we're in a cartesian plane, rounds around zero are much more
convenient, so we can use positive and negative values with consistent
results.
The functions `polygon8`, `polygon32` and `polygon32rgba` all have exactly
the same logic in code, only changes the hline function called inside.
Now all the logic is contained in `polygon_generic` which gets a
function pointer to the hline handler, so there is no duplicated code
anymore.
