| Commit message (Collapse) | Author | Age | Files | Lines |
| |
|
|
|
|
| |
This will be needed for the next commit.
|
|
|
|
| |
...instead of passing them as arguments everywhere.
|
| |
|
| |
|
| |
|
|
|
|
| |
This allowed writes to one past the end of `families` and `fullnames`.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Normally, we delay loading Dwrite.dll until runtime to allow building
and running DirectWrite-enabled libass binaries on old Windows versions
that lack DirectWrite. However, this is forbidden in WinRT/UWP.
DirectWrite is present in all versions of Windows that support WinRT/UWP,
so we lose nothing by requiring it.
Older Windows SDKs (Microsoft and MinGW alike) lack <winapifamily.h>,
so include it only if we really need it.
Based on VLC patch for libass:
https://github.com/videolan/vlc/commit/eedb57aa96d2bc0046a6da2e081c75ae9edf8fd5
and on this autoconf code:
https://github.com/lu-zero/mfx_dispatch/commit/c51a54c15f51579804030c70592c0a26065f1242
Note: the VLC patch retained an unconditional call to FreeLibrary
in destroy_provider. However, FreeLibrary does not actually expect
NULL as argument, so this commit adds a guard to that call.
Perhaps FreeLibrary(NULL) simply fails cleanly and that's why
this has not caused VLC problems, but we should not rely on this.
|
| |
|
| |
|
| |
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently implemented only for coretext,
but other providers may/should add this later.
This improves fallback font choice by considering the whole extended family.
This also fixes remaining cases of Core Text fallback failure caused by
differences among the names recognized by Core Text, the name we choose
to return from get_fallback, and the names we seek in find_font.
This partly reverts commit 152d0484e98f118d01987138695cf40579c9e297.
This fixes https://github.com/libass/libass/issues/512.
|
|
|
|
|
|
|
|
|
|
|
| |
We ask Core Text to match multiple name fields, and a single font may
match several at once. In this case, by default, Core Text duly returns
the font multiple times, and we add it to fontselect multiple times.
This does not break anything, but it is a waste, so ask Core Text
to deduplicate the returned list.
This is supposedly available since Mac OS X 10.5.
We support Core Text on 10.6+.
|
|
|
|
| |
Allows compilers to print warnings when the args we pass don't match the string
|
|
|
|
|
| |
- ass_drawing was logging size_t as %d
- ass_font was logging ASS_StringView as %s (could crash!)
|
|
|
|
|
|
|
| |
match_fonts is supposed to add all the fonts with the same family name
for fontselect. Commit d325c633f4ac32fcb4c76a3fb5926737a45c38d0 adds
only one font, which leads to inability to select (using family name)
correct fonts from families with multiple styles.
|
| |
|
|
|
|
|
|
|
|
| |
get_bitmap_glyph copies this struct via assignment. But to allow
safe assignment-copy, it seems the whole struct must be initialized,
rather than merely each of its members.
Fixes CID 314186 found by Coverity Scan.
|
|
|
|
|
|
|
|
|
|
|
|
| |
scan_fonts can be very slow when many fonts were installed in
the system. Use IDWriteGdiInterop to create a font by its name
when needed instead of scanning all installed fonts during the
initializing stage to get better performance. In case of a
non-existent font, the fallback mechanism should do its work.
Fixes https://github.com/libass/libass/issues/334.
Signed-off-by: Oleg Oshmyan <chortos@inbox.lv>
|
|
|
|
|
| |
GCC and MSVC (at least) warn about this cast as the function types
appear to be incompatible. This is correct with GetProcAddress.
|
|
|
|
|
|
|
|
| |
Second argument of cglobal macro for fill_halfplane in 64-bit mode
should always be 6 to preload all function arguments into registers.
It doesn't matter under *nix as the first 6 arguments are already
in registers according to calling conventions, but under WIN64
it's only the first 4.
|
|
|
|
|
|
| |
All Create and Copy functions create new objects, so all of them may fail.
Keep using SAFE_CFRelease only in shared cleanup clauses.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Since commit be0d1613f79a95073d18d96a60e1394abf9316a2,
get_fallback can return a name from Core Text that is different
from all names that match_fonts registers for the same font.
This causes font fallback to fail as find_fonts fails to
find any font having the name received from get_fallback.
(Moreover, libass will keep retrying fallback for other glyphs,
and the coretext match_fonts will keep readding the same fonts
over and over again.)
Commit 152d0484e98f118d01987138695cf40579c9e297 attempted
to fix font fallback by getting a full name from Core Text
instead of a family name when it was noticed that Core Text's
family name can come from TT_NAME_ID_TYPOGRAPHIC_FAMILY
in addition to TT_NAME_ID_FONT_FAMILY, which we fetch in
get_font_info, but the problem goes deeper: Core Text can return
Macintosh-platform names that don't match Microsoft-platform
names, or the font might have no Microsoft-platform names at all.
Furthermore, returning a full name breaks style matching:
get_fallback does not know about weight, slant etc. and is
expected to return a whole family of fonts that will be
lazy-loaded in its entirety (if applicable) and searched
through for the best stylistic match by find_fonts.
To fix fallback while preserving maximum name portability
in the spirit of be0d1613f79a95073d18d96a60e1394abf9316a2,
use the same API in get_fallback as is used in match_fonts
to look up a family name.
Note: this could be more efficient if ass_get_font_info
could be told to return just an arbitrary family name.
This fixes https://github.com/libass/libass/issues/457.
|
|
|
|
|
|
|
|
| |
If a font contains no recognized Microsoft-platform family names,
use the family name reported by Core Text.
Skip FreeType's face->family_name because it is an ASCII
bastardization of a family name and may not exist at all.
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
These glyphs are effectively invisible.
VSFilter draws no border around them, either (because for BorderStyle 1
it finds no pixels to draw a border around, while BorderStyle 3 gets
scaled a second time by the same \fscx or \fscy), except current MPC-HC
(which doesn't double-scale BorderStyle 3).
Since nothing is to be seen on screen, skip the rest of the pipeline
altogether. They still affect the layout of other glyphs that may
have nonzero scales, so we can only do this after text layout is done.
This avoids subsequent divisions by zero when sheared pre-scale coordinates
are computed from post-scale coordinates at scale zero. (We eagerly scale
all coordinates before shearing because it makes text layout easier.)
In apply_baseline_shear itself, the shear displacement for an \fscy0
glyph is always zero, so skipping its calculation is safe. For an
\fscx0 glyph, we currently divide 0 by 0, which we want to avoid.
Logically, this is supposed to take the cluster's pre-\fscx advance
and compute shear from that, but this commit does not do this. When
in VSFilter-compatible mode, we're actually supposed to reset the
baseline when \fscx becomes nonzero, so this won't matter. Meanwhile,
when we're trying to do a "sensible" thing, it's not clear whether
shear-before-scale makes sense in the first place. At any rate, this
improves upon the current behavior by avoiding the zero division,
and at least on x86 this actually happens to match our current layouts
in situations where the current code produces visible output at all.
Fixes https://github.com/libass/libass/issues/185.
Fixes https://github.com/libass/libass/issues/413.
Fixes https://github.com/libass/libass/issues/428.
Fixes https://github.com/libass/libass/issues/443.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The current code calculates the effect of \fay on advances
early in retrieve_glyph, only to ignore it until reorder_text,
which then partially undoes these calculations.
Instead, just calculate it once, when all the necessary
properties have been fully determined.
The earliest point to do this would be within reorder_text,
but to simplify future logic changes, delay this until all
other text layout steps are done. This matches VSFilter, which
applies shear after alignment. This would also have neutralized
the bug that the previous commit fixed, because glyphs[0].pos.y
really is always 0 until shear is applied. This also opens
the door for another commit to skip glyphs that are invisible
but affect layout.
In addition to simplifying code and calculations, this commit fixes
a bug: the current code for undoing shear assumes constant scale_x/y
from the last \fay change or line break point to the next such point,
producing strange, wrong positions when scale_x/y do change.
This fixes https://github.com/libass/libass/issues/465.
|
|
|
|
|
| |
Put continue on its own line.
Don't mutate & reset info.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
In most cases, text_info->glyphs[0].pos.y is 0. This makes the bbox
in compute_string_bbox have bbox.y_min = -text_info->lines[0].asc.
Thus, during event positioning, we refer to the text's top and
bottom edge coordinates in two equivalent ways: \an4-6, \pos, \move
and Scroll effects use bbox.y_min and bbox.y_max, while \an1-3
and \an7-9 use -text_info->lines[0].asc and text_info->height -
text_info->lines[0].asc.
However, text->glyphs[0].pos.y is not *actually always* 0.
Specifically, it is nonzero for events that begin with right-to-left
text with nonzero \fay (at least two visible glyphs with the same \fay).
In this case, the anchoring point we use for vertical positioning
is different between \an1-3,7-9 and the other cases. For \an1-3,7-9,
it is the middle point of the text's left edge, which makes sense.
But for \an4-6, \pos, \move and Scroll effects, it is shifted up.
It is easiest to describe for a single-line event: the anchor is
the middle point of the *left edge of the rightmost glyph*. That is,
the anchor point is not the middle of *either* edge of the full text,
and it is hard to understand intuitively, more so in a multiline event.
VSFilter in the same situation places the anchor in the middle of the
whole event's left edge, just as we do for \an1-3 and \an7-9.
This also affects the position of rotation origin,
which is likewise unintuitive and VSFilter-incompatible.
Fix this by making compute_string_bbox ignore text->glyphs[0].pos.y
even in the rare case that it is nonzero. This makes our behavior
obvious and VSFilter-compatible, and our code easier to understand
(as the reader no longer has to wonder whether there is a difference
between the various alignments, wonder *why* there seems to be no
difference after testing common cases, and make the wrong conclusion
that glyphs[0].pos.y is always 0).
|
|
|
|
|
|
|
|
|
|
|
| |
Previously only both only worked when ass_set_fonts was called after all
embedded and memory fonts were already added. Especially for embedded
fonts this meant it won't work for most users, except mpv, including our
own utilities, even if extract_fonts was set.
Now that it works, enable extract_fonts in our utilities.
GitHub: fixes #266
|
|
|
|
|
|
|
|
| |
As the function only takes a library handle it is imposssible to adjust
it to be safe to call otherwise. FT_Done_Face needs to be called before
the font data can be freed.
Luckily it seems barely anyone even used this (mpv, ffmpeg, VLC and
Aegisub don't)
|
|
|
|
|
|
| |
The latter has overflow protections for size_t indexation and element
size > 1, the former none whatsoever. Change all related vars to size_t.
Additionally, switch to exponential buffer growth.
|
| |
|
| |
|
|
|
|
| |
Fixes https://github.com/libass/libass/issues/475
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Change list:
- Fixed differences from C version introduced
in f23b9ed64bd4ccf249c686616dd3f51a69d285dc.
- Common macro for SSE2 and AVX2 versions.
- Reduced register usage and efficient 32-bit version.
- Full width memory operations instead of half-register.
- Vectorized handling of width tails instead of byte/word loops.
- Vectorized initial population of temporary buffer and final line fill.
- Interleaved layout of temporary buffer.
- Great speedup overall.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Change list:
- No special handling of unaligned case.
- Common macro for SSE2 and AVX2 versions,
AVX2 has got significantly faster.
- Reduced register usage and efficient 32-bit version.
- Full width memory operations instead of half-register.
- Vectorized handling of width tails instead of byte loops.
|
|
|
|
|
|
|
|
|
|
| |
Change list:
- No special handling of unaligned case (removed in previous commit).
- Common macro for both add_bitmaps() and sub_bitmaps().
- Reduced register usage and efficient 32-bit version.
- add_bitmaps() no longer rely on zero padding.
- Vectorized handling of width tails (instead of byte loop
in sub_bitmaps(), great speedup for non-empty tails).
|
| |
|
| |
|
|
|
|
|
|
|
| |
Those particular instances were harmless, but in general trigraphs can
cause unintended behaviour changes. Not having any trigraphs in the code
makes it easier to catch unintended ones via a simple regex.
(In GNU-C trigraphs are disabled by default, but enabled in ISO-C)
|
|
|
|
|
|
| |
M_PI is not part of ISO C and also not part of base POSIX, but an XSI
extension. For simplicity, just always use our own macro for PI.
The value is identical to the one used in glibc and musl.
|
|
|
|
|
|
|
|
| |
And move fallback declarations to ass_compat.h
As ass_compat.h is already included in every source file we no longer
need to include _both_ string.h and ass_utils.h to use str(n)dup.
Definitions are still in ass_utils.c since a separate source file just
for two functions seemed overkill.
|
|
|
|
|
|
|
|
|
| |
This could help us in future bug reports and we already do so for
shapers and freetype. Also include a guess about the source code
version, which may not be entirely reliable. Eg if custom tarballs are
involved or if patches are added after configure was called, the value
will be wrong.
But hopefully it will be "good enough" for most cases.
|
|
|
|
|
|
| |
This matches VSFilter.
Fixes https://github.com/libass/libass/issues/487.
|
| |
|
|
|
|
|
|
| |
Our hash function doesn't handle zero-length buffers,
but it was not a problem until 82b225b3d6653091d028b39d561d185ed76a7be5.
Found by OSS-Fuzz.
|
|
|
|
| |
This should fix the warnings introduced with nasm 2.15
|
| |
|
|
|
|
| |
That eliminates most uses of strdup() in the rendering process.
|
| |
|
| |
|
| |
|
|
|
|
|
|
| |
There is a small typo in libass/ass_outline.h.
Should read `outline` rather than `ouline`.
|
|
|
|
|
| |
Both parameters are copied and are immutable. Make it clear through the
function prototype
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
VSFilter does collision by calculating the intersection of two events
rectangles and treating events as colliding iff their intersection
has a non-zero area.
As a consequence zero-area events cannot collide with anything.
https://github.com/Cyberbeing/xy-VSFilter/blob/xy_sub_filter_rc4@%7B2020-05-17%7D/src/subtitles/RTS.cpp#L1627
libass does not do collision by intersection area, but because we
ignore edge-edge intersections results are identical to VSFilter
for non-zero-area events.
To reconcile the results for zero-area events, treat them as explicitly
positioned as they effectively are in VSFilter. This allows us to keep
our (probably) faster collision code and saves us from unnecessary
collsion checks for zero-area events.
Fixes https://github.com/libass/libass/issues/444
|
|
|
|
|
|
| |
Since text_info is not re-initialised before every event, border_x could
carry over to following events with a smaller \xbord value.
This was broken in commmit 236bc18d39c080853d1145c7b978fb8f47e31deb
|
| |
|
| |
|
|
|
|
|
|
|
|
| |
Omissions in commit 910211f1c0078e37546f73e95306724358b89be2.
Microsoft's C library does not support %zu until Universal CRT
(Visual Studio 2015). At worst, this verbose-level message will
look wrong and be useless.
|
|