Up to now, we simply took a snapshot of the Graphics State after the
`prep' table has been executed, and right before a glyph's bytecode
was run it got reloaded. However, as Greg Hitchcock has told us in
private communication, reference points get reset to zero in the MS
rasterizer and we follow in due course. While reasonable, this is
undocumented behaviour.
Most notably, this fixes the rendering of Arial's `x' glyph in
subpixel hinting mode.
Many compilers such as gcc and clang optimize away pointer overflow
checks `p + n < p', because pointer overflow is undefined behavior.
Use a safe form `n > p_limit - p' instead.
Also avoid possible integer overflow issues, for example, using
`num_glyphs > ( p_limit - p ) / 2' rather than `num_glyphs * 2'
given a large `num_glyphs'.
* src/sfnt/ttsbit0.c (tt_sbit_decoder_load_image): Implement it.
* src/truetype/ttsubpix.h: Updated.
(SPH_X_SCALING_RULES_SIZE): Moved and renamed to...
* src/truetype/ttsubpix.c (X_SCALING_RULES_SIZE): This.
(sph_X_SCALING_Rules): Removed.
(scale_test_tweak): Make function static.
(sph_test_tweak_x_scaling): New function.
* src/truetype/ttgload.c (TT_Process_Simple_Glyph): Updated.
* include/freetype/fttrigon.h (FT_Hypot): Declare it.
* src/base/fttrigon.c (FT_Hypot): Define it.
* src/truetype/ttgload.c (TT_Process_Composite_Component): Use it
instead of explicit expressions.
* src/truetype/ttinterp.c (Current_Ratio, Normalize): Use it instead
of TT_VecLen.
(TT_VecLen): Removed.
* src/truetype/ttobjs.c (tt_size_reset): The Windows rendering
engine uses rounded values of the ascender and descender to compute
the TrueType font height.
It was not taking bit_depth into consideration when blitting!
* src/sfnt/ttsbit0.c (tt_sbit_decoder_load_byte_aligned,
* tt_sbit_decoder_load_bit_aligned): Handle bit
depth.
This patches fixes many issues with the ttsubpix implementation.
1. Data tables are defined, instead of declared, in the header, and
thus copied into each source file that includes it.
2. These tables were defined as global, mutable, visible variables,
and thus costing private RAM to every process that loads the
library (> 50 KB / process, this is huge!).
Additionally, this also made the library export the symbols
completely needlessly.
3. Missing `sph_' and `SPH_' prefixes to some of the definitions.
Note that this doesn't try to fix the incredibly inefficient storage
format for the data tables used by the code. This one will require
another pass in the future.
* src/truetype/ttinterp.h (MAX_NAME_SIZE, MAX_CLASS_MEMBERS):
Renamed to...
(SPH_MAX_NAME_SIZE, SPH_MAX_CLASS_MEMBERS): This.
Update all users.
(SPH_TweakRule, SPH_ScaleRule): Decorate with `const' where
appropriate.
(Font_Class): Rename to...
(SPH_Font_Class): This. Decorate with `const' where appropriate.
* src/truetype/ttsubpix.h (scale_test_tweak, sph_test_tweak):
Decorate arguments with `const' where appropriate.
Move font tweaking tables to...
* src/truetype/ttsubpic.c: This file and decorate them with `static'
and `const' where appropriate.
(X_SCALING_Rules, X_SCALING_RULES_SIZE): Renamed to...
(spu_X_SCALING_Rules, SPH_X_SCALING_RULES_SIZE): This.
Update all users.
Unit vector components are stored as 2.14 fixed-point numbers. In
order to calculate all 14 bits accurately, a short vector to be
normalized has to be upscaled to at least 14 bits before its length
is calculated. This has been safe since accurate CORDIC algorithms
were adopted.
* src/truetype/ttinterp.c (Normalize): Scale short vectors by 0x4000.
Back in the days, vector length calculations were not very accurate
and the vector normalization function, Normalize, had to meticulously
correct the errors for long vectors [commit b7ef2b0968]. It was no
longer necessary after accurate CORDIC algorithms were adopted, but
the code remained. It is time to kill it.
* src/truetype/ttinterp.c (Normalize): Remove error compensation.
(TT_VecLen): Remove any mention of old less accurate implementation.
After the next release we are going to remove the code completely.
* devel/ftoption.h, include/freetype/config/ftoption.h
(FT_CONFIG_OPTION_OLD_INTERNALS): Comment out.
* docs/CHANGES: Document it.
The recent optimizations of CORDIC iterations drastically reduce
the expansion factor. The vector components with MSB of 29 are now
safe from overflow.
* src/base/fttrigon.c (FT_TRIG_SAFE_MSB): New macro.
(ft_trig_prenorm): Use it and remove dead code.
* src/base/fttrigon.c: Document the algorithm in a large comment.
(FT_TRIG_COSCALE): Remove macro.
(FT_Tan: Use `FT_TRIG_SCALE' instead.
(FT_Cos, FT_Vector_Unit): Ditto and round the return values.
After we get within 45 degrees by means of true 90-degree rotations,
we can remove initial 45-degree CORDIC iteration and start from
atan(1/2) pseudorotation, reducing expansion factor thereby.
* src/base/fttrigon.c (FT_TRIG_SCALE, FT_TRIG_COSCALE): Update macros.
(ft_trig_pseudo_rotate, ft_trig_pseudo_polarize): Update.
* src/tools/cordic.py: Bring up to date with trigonometric core.
* docs/CHANGES: Old typo.
Problem reported by Robin Watts <robin.watts@artifex.com>.
* src/base/ftoutln.c (FT_Outline_New_Internal): Ensure that
`numContours' and `numPoints' fit into FT_Outline's `n_points' and
`n_contours', respectively.
FreeType used to rely on a 24-step iteration CORDIC algorithm to
calculate trigonometric functions and rotate vectors. It turns out
that once the vector is in the right half-plane, the initial rotation
by 63 degrees is not necessary. The algorithm is perfectly capable
to converge to any angle starting from the second 45 degree rotation.
This patch removes the first rotation and makes it a 23-step CORDIC
algorithm.
* src/base/fttrigon.c (FT_TRIG_SCALE, FT_TRIG_COSCALE): Update macro
values.
(ft_trig_pseudo_rotate, ft_trig_pseudo_polarize): Remove initial
rotation.
