and therefore freeing the memory only when the i_refcount successfully
went to 0.

The problem was that i_refcount is an unsigned variable, and was being
decremented twice, once erroneously by mosaic_bridge, and then again in
the picture's original pf_release function. If i_refcount started at 1,
it wrapped to the maximum unsigned value rather than -1, failing the
refcount tests in the pf_release function.

Patch Authors: Phil Roffe and David Grellscheid
Signed-off-by: Antoine Cellerier <dionoea@videolan.org>

Sample can be found at ftp://streams.videolan.org/streams-videolan/matroska/MLP_5.1.mka
It seems to work, but I have difficulties to find a reference output.
Support for TrueHD is not yet tested because of lack of correct sample.

First step of fixing linking errors against libx264

contribs: don't compile ffmpeg with mdynamic-no-pic on PPC-Macs, as this is a really bad idea. You need PIC in shared libs on this platform.

This slows down the compilation speed by 5%, but that should be fine. Pointed by nobody.loopback on the forum.

It fixes #2648.

This removes a lot of false true positives.

It was based on xxcv investigation and preliminary patch.
It should fix #2644.

Always increase i_bytes_per_frame to avoid cycling between 2 values with
some files.

YES, imagine, including stdint.h is apparently a good idea from time to time!

macosx: get rid of the infamous 'get' prefixes everywhere and finally respect Cocoa's naming scheme here

quartztext, controlling playback with the Apple Remote and embedded AGL vouts (aka safari/mozilla plugin vouts) are not supported by now. Quartztext and AGL vout will need a complete rewrite.

The OS X interface no longer depends on the QuickTime framework.

Check for broken ld on Darwin 9 only

 It happens when RLE offsets overlap (en general, because both fields
share the same data).
 Its is probably present in 0.9.

Note that the x86_64 package doesn't include libsdl, libsdl_image and goom2k4, as these aren't 64bit compatible.