Again, if the item does not exist, an assert will fail.

If the named item does not exist then an assert will fail, which
does not match this description. I don't think the actual behavour
particularly needs documenting.

Lyndon Brown authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

The previous commit made use of compound literals. MSVC was identified
as not yet having support for compound literals in C++ though, which is
considered a blocking problem to their use. This work around was
suggested by Alexandre Janniaux.

Lyndon Brown authored 6 years ago and Rémi Denis-Courmont committed 3 years ago

I.e. check the type of the values provided, rather than just blindly
casting, and thus allow mistakes to get caught with compilation errors.

I once lost some hours tracking down the cause of a crash that this type
strictness would have prevented.

Note, now that `change_string_list()` and `change_integer_list()` use type
strictness for the value and label arrays to help make sure they're the
right types, we can more safely make use of `ARRAY_SIZE()` to more cleanly
get the length.

`set_callback[s]()` is excluded from this for obvious reasons.

I've taken the opportunity to ditch remaining use of Hungarian notation
and to fix a param name that should be singular rather than plural
(psz_caps on `add_module[_list]()`.

Re-implemented using compound literals upon excellent suggestion by
Alexandre Janniaux.

vlc_stream_NewURL() can spawn stream filters which may alter the content
or even the nature of the stream, which is not suitable here.

vlc_access_NewMRL() would avoid loading stream filters at all. But it
would also skip testing the cache filter, thus missing much of the point
of the test case.

So use a reproducible pseudorandom input using a constant seed, which
is known not to trigger any stream filter.

Fixes #26569.

libvlc_audio_output_set() had no practical effects as the audio output
was consistently created before the application had the opportunity to
select a non-default module. This fixes it.

An alternative would be to remove the function entirely. Exposing
module names to LibVLC applications is bad design.

Historically this function would set the device for a specified audio
output module. This required the application to somehow "know" what
audio output module it wanted to use, which is very suboptimal for
forward compatibility and portability.

Worse yet, this design failed outright in VLC 2.2 with the addition of
MMDevice and PulseAudio. This lead to the convention of leaving the
module parameter NULL to change the device of the currently active
audio output.

As with the removal of libvlc_audio_output_device_list_get(), it seems
saner to simply remove the old broken semantics.

This function does not work with MMDevice and PulseAudio, and does not
support hot-plug. To enumerate devices, the newer function
libvlc_audio_output_device_enum() ought to be used, which also works
with MMDevice and PulseAudio, and track audio device events.

No amount of documentational warnings will save developers from using
this broken function. It is better to remove it now that there is the
opportunity to do so in the form of the LibVLC 4.0 binary compatibility
break.

Brad Smith authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

Fixes #26573

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

Brad Smith authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

Fixes #26572

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

This applies the user-requested transform to the video format, rather
than the pictures. Doing this serves multiple purposes:
 - the filter chain format can be computed correctly (notably the SAR),
 - the transform filter can process any format, including opaque ones.

Instead of creating a sub-chain with a singular hard-coded transform
filter instance, the chain video converter treats mismatched video
format orientations normally as other mismatched parameters.

In practice, this will trigger the new orientation video converter
instead of the transform video filter, and the result will be
essentially the same... when using non-opaque picture formats.

Unlike transform, this checks the video format, not some configuration
parameters. Thus a video format can undergo reorientation directly,
without having the chain video converter create a dummy chain for it.

The code is heavily based on the transform filter.

...rather than a very ad-hoc ersatz. This works because transform and
orientation are idendity-mapped.

No functional changes.

Steve Lhomme authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

Rather than the number of samples which depends directly on it.

Fixes #26565.

Disabling the window is sufficient here.

Steve Lhomme authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

The remaining module usage was removed with 7b8c6467.

Down the line, we probably should not have one directory per ISA
extension. This replicates the model from the hw/*/ directory.

isa/ seems better than arch/ because:
 - "Instruction Set Architecture" is more specific than "Architecture",
 - it saves one character,
 - no modules/ subdirectory starts with I yet (so this will not break
   auto-completion habits).

This also moves non-NEON-specific files to isa/arm/.

See gettext documentation.

This follows the established cargo cult of not assuming that the
toolchain supports SIMD.

This detects support for the RISC-V "V" Vector Extension at run-time
from the device tree via /proc/cpuinfo.

The number of tabulation character can vary with versions. Case in
point, RISC-V started with one tabluration, but has two in more recent
kernel versions, to maintain visual alignment.

This detects compiler support for the RISC-V "V" Vector Extension if it
enabled in the compiler flags (or compiler machine definition).

Lyndon Brown authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

Only two such defines were already in use, the rest were pointlessly
duplicating the text, which should be identical in both interfaces and so
it makes sense to have shared defines.

This happens to fix a few unwanted inconsistencies:
 - MacOS was using lowercase "settings" in two places.
 - Qt was using completely different language for the hotkey and ML tag
   tooltips to everything else.
 - A trivial difference between input tab tooltips.

This provides for audio input/output on embedded Linux.

Alexandre Janniaux authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

It works much better with dark mode theme.

Alexandre Janniaux authored 3 years ago and Jean-Baptiste Kempf committed 3 years ago

Not generating the tree view is the default setting, but it's actually
pretty useful and pretty common (rustdoc, msdn, mdn, awesome doxygen).