This works more or less like SNDVOL (and like PulseAudio with flat
volumes not disabled): the VLC volume represents the product of the
endpoint volume by the simple volume.

cross-pkg-config tools have a correct default search path. The contrib
path should complement it, not replace it. If the caller intends to
(ab)use the native pkg-config for cross-compilation, the caller is
responsible for selecting an adequate default path, not VLC configure,
just as with every other build tools.

fixes #10125, #10199

This is superseded by a proper fix.
This reverts commit fa396d19.

This reverts commit c3bb23cd.

If Start() failed, IDirectSound_Release() was called twice.

Signed-off-by: Rémi Denis-Courmont <remi@remlab.net> with changes

Fix a crash occurring when CreateDSBufferPCM() fails in Start() and
afterwards Stop() uses uninitialized pointer values when releasing
objects.
Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>

NSFullScreenWindowMask is not defined here.

Defining the category to a nonexisting class is not allowed, as well as providing
an interface for this class without implementation.

libavutil does what we do already

Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>

No functional change.

close #10716

Signed-off-by: Rémi Denis-Courmont <remi@remlab.net>

libavutil will always check for all CPU features that it understands,
even those VLC masked. Thus masking features provides no savings.

libavutil runs more tests than VLC. Depending on the platform and
operating system combination, libavutil is either as conservative or
more conservative than VLC. As such masking features provides no extra
safety.

So overall, GetVlcDspMask() is useless, at least nowadays.

Android ARM NEON seems controversial so it is kept for now.