Benchmarking shows that this provides a very slighty performance
boost when rendering fonts with lots of quadratic bezier arcs,
compared to the recursive arc splitting, but only when SSE2 is
available, or on 64-bit CPUs.
On a 2017 Core i5-7300U CPU on Linux/x86_64:
./ftbench -p -s10 -t5 -cb .../DroidSansFallbackFull.ttf
Before: 4.033 us/op (best of 5 runs for all numbers)
After: 3.876 us/op
./ftbench -p -s60 -t5 -cb .../DroidSansFallbackFull.ttf
Before: 13.467 us/op
After: 13.385 us/op
This speeds up the smooth rasterizer by avoiding a
conditional branches in the hot path. Namely:
- Define a fixed "null cell" which will be pointed
to whenever the current cell is outside of the current
target region. This avoids a "ras.cell != NULL"
check in the FT_INTEGRATE() macro.
- Also use the null cell as a sentinel at the end of
all ycells[] linked-lists, by setting its x coordinate
to INT_MAX. This avoids a 'if (!cell)' check in
gray_set_cell() as well.
- Slightly change the worker struct fields to perform
a little less operations during rendering.
Example results (on a 2013 Corei5-3337U CPU)
out/ftbench -p -s10 -t5 -bc /usr/share/fonts/truetype/droid/DroidSansFallbackFull.ttf
Before: 5.472 us/op
After: 5.275 us/op
out/ftbench -p -s60 -t5 -bc /usr/share/fonts/truetype/droid/DroidSansFallbackFull.ttf
Before: 17.988 us/op
After: 17.389 us/op
Replaces download-test-fonts.sh with download-test-fonts.py
which does the same work, and also avoids downloading anything
if the files are already installed with the right content.
Now uses the first 8 byte of each file's sha256 hash for the
digest.
I am currently trying to compile FreeType for CHERI-extended ISAs
(CHERI-RISC-V and Arm's Morello), but I am getting compiler warnings
from the `FT_UINT_TO_POINTER` macro. When compiling with the CHERI
Clang compiler, not using `uinptr_t` for casts between integers an
pointers results in the following `-Werror` build failures:
```
In file included from .../src/truetype/truetype.c:22:
.../src/truetype/ttgload.c:1925:22: error:
cast from provenance-free integer type to pointer type will
give pointer that can not be dereferenced
[-Werror,-Wcheri-capability-misuse]
node->data = FT_UINT_TO_POINTER( glyph_index );
^
.../include/freetype/internal/compiler-macros.h:79:34: note:
expanded from macro 'FT_UINT_TO_POINTER'
```
* include/freetype/internal/compiler-macros.h (FT_UINT_TO_POINTER):
The ISO C standard compliant fix for this would be to use
`uintptr_t` from `stdint.h`, but I am not sure if this is supported
by the minimum compiler version. Therefore, use the
compiler-defined `__UINTPTR_TYPE__` macro (supported in GCC 4.6+ and
Clang since about 3.0) before checking for `_WIN64` and falling back
to `unsigned long`.
This was already true (though undocumented) most of the time, but
not if `FT_NEW` inside `FT_Stream_New` failed or if the
`FT_OPEN_XXX` flags were bad.
Normally, `FT_Open_Face` calls `FT_Stream_New`, which returns the
user-supplied stream unchanged, and in case of any subsequent error
in `FT_Open_Face`, the stream is closed via `FT_Stream_Free`.
Up to now, however, `FT_Stream_New` allocates a new stream even if
it is already given one by the user. If this allocation fails, the
user-supplied stream is not returned to `FT_Open_Face` and never
closed. Moreover, the user cannot detect this situation: all they
see is that `FT_Open_Face` returns `FT_Err_Out_Of_Memory`, but that
can also happen after a different allocation fails within the main
body of `FT_Open_Face`, when the user's stream has already been
closed by `FT_Open_Face`. It is plausible that the user stream's
`close` method frees memory allocated for the stream object itself,
so the user cannot defensively free it upon `FT_Open_Face` failure
lest it ends up doubly freed. All in all, this ends up leaking the
memory/resources used by user's stream.
Furthermore, `FT_Stream_New` simply returns an error if the
`FT_OPEN_XXX` flags are unsupported, which can mean either an
invalid combination of flags or a perfectly innocent
`FT_OPEN_STREAM` on a FreeType build that lacks stream support.
With this patch, the user-supplied stream is closed even in these
cases, so the user can be sure that if `FT_Open_Face` failed, the
stream is definitely closed.
* src/base/ftobjs.c (FT_Stream_New): Don't allocate a buffer
unnecessarily.
Move error-handling code to make the control flow more obvious.
Close user-supplied stream if the flags are unsupported.
`FT_Stream_Open` always sets `pathname.pointer`, so remove the
redundant (re)assignment. None of the `FT_Stream_Open...` functions
uses `stream->memory`, so keep just one assignment at the end,
shared among all possible control flow paths.
('Unsupported flags' that may need a stream closure can be either an
invalid combination of multiple `FT_OPEN_XXX` mode flags or a clean
`FT_OPEN_STREAM` flag on a FreeType build that lacks stream
support.)
The three modes are mutually exclusive, and the documentation of the
`FT_OPEN_XXX` constants notes this. However, there was no check to
validate this in the code, and the documentation on `FT_Open_Args`
claimed that the corresponding bits were checked in a well-defined
order, implying it was valid (if useless) to specify more than one.
Ironically, this documented order did not agree with the actual
code, so it could not be relied upon; hopefully, nobody did this and
nobody will be hurt by the new validation.
Even if multiple mode bits were allowed, they could cause memory
leaks: if both `FT_OPEN_STREAM` and `stream` are set along with
either `FT_OPEN_MEMORY` or `FT_OPEN_PATHNAME`, then `FT_Stream_New`
allocated a new stream but `FT_Open_Face` marked it as an 'external'
stream, so the stream object was never released.
* src/base/ftobjs.c (FT_Stream_New): Reject incompatible
`FT_OPEN_XXX` flags.
Without this change the build of `unix/ftsystem.c` fails because the
`ftconfig.h` header that defines macros such as `HAVE_UNISTD_H` and
`HAVE_FCNTL_H` is only being generated for Linux, macOS, and Cygwin
systems:
```
.../builds/unix/ftsystem.c:258:32: error:
use of undeclared identifier 'O_RDONLY'
file = open( filepathname, O_RDONLY );
```
Instead of hardcoding a list of operating systems for this check,
update the logic that decides whether to build the file and set a
boolean flag that can be checked instead.
Move the af_sort_xxx() functions from afangles.c to afhints.c
in order to get rid of the obsolete angle-related types,
macros and function definitions.
This feature was always experimental, and probably nevery worked
properly. This patch completely removes it from the source code,
except for a documentation block describing it for historical
purpose.
This code has always been experimental and was never compiled
anyway (FT_OPTION_AUTOFIT2 does not appear in ftoption.h or even
any of our build files).
It is undefined behavior to pass `NULL` to `memcpy`. `coords' is
passed to `memcpy` but `TT_Get_MM_Blend` and `TT_Get_Var_Design`
explictly call `tt_set_mm_blend` with `coords` as `NULL`. In
addition, `TT_Set_MM_Blend` has a similar possible issue.
'COLR' v1 fonts do not necessarily need to have a layer list; for
this reason, 'fontTools' recently started generating fonts in a way
that drops the layer list if there are no layers in it. This
results in the layer list offset becoming zero, which FreeType
treated as an invalid table. Fix that and handle the case for layer
list offset being 0. This slightly changes how we need to calculate
the starting offset for paints.
* src/sfnt/ttcolr.c (tt_face_load_colr): Handle case of layer list
offset being zero without outright rejecting table.
The second pass also fixes horizontal lines through the pixel centers.
* src/raster/ftraster.c (black_TWorker): Do not use `second_pass'.
(Render_Glyph): Skip the second pass only with the appropriate flag.
* src/raster/ftraster.c (black_TWorker): Move `band_stack' from here..
(Render_Single_Pass): ... to here and accept limit arguments.
(Render_Glyph): Updated.
* src/raster/ftraster.c (black_TWorker): Move `arcs' from here...
(Conic_To, Cubic_To): ... to here to tighten their scope.
(Bezier_Up, Bezier_Down): ... Take the current `arc' argument.
Paint tables can appear before the `base_glyphs_v1` offset if the
font is produced with the layer list before the base glyph list. In
this case paint tables can occur after the layer list but before the
base glyph list. Checks in the 'COLR' v1 code were rejecting fonts
with this layout. Improve these checks by calculating a minimum
offset after which paint tables can occur and use that in safety
checks.
* src/sfnt/ttcolr.c (Colr, tt_face_load_colr): Declare
`paint_start_v1` and calculate that as the minimum of the end of
layer list and base glyph list.
(get_child_table_pointer, read_paint, tt_face_get_paint_layers):
Use that in safety checks.
* src/raster/ftraster.c (black_TWorker): Replace the current line
offset with the pointer and drop the increment.
(Function_Sweep_Init): Take values as arguments instead of pointers.
(Vertical_Sweep_*, Horizontal_Sweep_Init, Draw_Sweep): Updated.
We can bisect a band until it is just a single scan line. This might
be slow and cause time-outs but if we need to impose limits it should
be elsewhere.
* src/raster/ftraster.c (Render_Single_Pass): Tweak sub-banding.
Some practical debugging work has shown that displaying level X of
an `FT_TRACEX` macro in the output of `FT2_DEBUG="... -v"` would be
very helpful to find out which trace level should be selected. As
an example, we now get output like
```
[ttobjs:2] TTF driver
[ttobjs:2] SFNT driver
[sfobjs:2] not a font using the SFNT container format
[t1objs:2] Type 1 driver
[stream:7] FT_Stream_EnterFrame: 14 bytes
```
* include/freetype/internal/ftdebug.h (FT_LOGGING_TAGX): New macro.
(FT_LOG): Use it to add the trace level to the logging tag.
* include/freetype/internal/fttrace.h (FT_MAX_TRACE_LEVEL_LENGTH):
Adjust.
* docs/DEBUG: Updated.
FT_Render_Glyph picked up FAILURE or 1 returned from the raster
function, which became a confusing error code. Instead, return
Raster_Overflow in the unlikely event that banding does not help or
another meaningful error.
* src/smooth/ftgrays.c (gray_convert_glyph_inner, gray_convert_glyph):
Use Raster_Overflow when the rendering pool is exhausted and return it
if banding does not help.
(gray_raster_render): Use Smooth_Err_Ok.
* src/raster/ftraster.c (Render_Single_Pass): Return Raster_Overflow
if banding does not help or another error code.
* include/freetype/ftcolor.h (FT_PaintTransformed, FT_PaintFormat,
FT_COLR_Paint): Do it to make it harmonize with other names such as
'PaintTranslate'.
* src/sfnt/ttcolr.c (read_paint, tt_face_get_paint): Ditto.
* include/freetype/freetype.h: Cut section layer managament
containing 'COLR' v0 and v1 API and move it to `ftcolor.h` as
requested by Werner on freetype-devel.
* include/freetype/ftcolor.h: Paste that section.
* tests/README.md: documentation explaining how to
build and run the tests with the Meson build.
* tests/scripts/download-test-fonts.sh: Simple script
to download test font files to the 'tests/data/'
folder.
* meson.build, meson_options.txt, tests/meson.build:
Add 'tests' option to enable building and running the
test programs, disabled by default.
* tests/issue-1063/main.c: Simple regression test to
exhibit issue 1063.
* .gitignore: Ignore the content of the tests/data/
folder for now.
We need to inverse inheritance of FT_GlyphSlot_Internal so that we
have a chance to free the rendered bitmap from the parent slot.
* src/type42/t42objs.c (T42_GlyphSlot_Init): Remove the internal parts
of the child `ttslot' and replace it with the parent structure.
(T42_GlyphSlot_Done): Updated accordingly.
David Turner