This fixes Savannah bug #26867.
Previously, only integers have been used which can lead to serious
rounding errors.
However, fractional values are only used internally; after the
charstrings (of either Type 1 or 2) have been processed, the
resulting coordinates get rounded to integers currently -- before
applying scaling. This should be fixed; at the same time a new load
flag should be introduced, to be used in combination with
FT_LOAD_NO_SCALE, which indicates that font units are returned in
16.16 format. Similarly, the incremental interface should be
extended to allow fractional values for metrics.
* include/freetype/internal/psaux.h (T1_BuilderRec): Remove `shift'
field.
* include/freetype/internal/pshints.h (T1_Hints_SetStemFunc,
T1_Hints_SetStem3Func): Use FT_Fixed for coordinates.
* src/psaux/psobjs.c: Include FT_INTERNAL_CALC_H.
(t1_build_add_point): Always convert fixed to integer.
* src/psaux/t1decode.c (t1_decoder_parse_charstrings):
Use 16.16 format everywhere (except for large integers followed by a
`div').
[CAN_HANDLE_NON_INTEGRAL_T1_OPERANDS]: Remove #ifdef and activate
code uncoditionally.
Add support for random numbers and update remaining code
accordingly; this should work now.
(t1_operator_seac): Updated.
* src/psaux/pshrec.c: Include FT_INTERNAL_CALC_H.
(ps_hints_t1stem3, t1_hints_stem): Updated.
* src/cid/cidgload.c: Include FT_INTERNAL_CALC_H.
(cid_load_glyph) [FT_CONFIG_OPTION_INCREMENTAL],
(cid_face_compute_max_advance, cid_slot_load_glyph): Updated.
* src/type1/t1gload.c (T1_Parse_Glyph_And_Get_Char_String)
[FT_CONFIG_OPTION_INCREMENTAL], (T1_Get_Advances, T1_Load_Glyph):
Updated.
* src/type1/t1load.c: Include FT_INTERNAL_CALC_H.
* src/type1/t1objs.c (T1_Face_Init): Updated.
Normally, the SCANMODE instruction (if present) to set the drop-out
mode in a TrueType font is located in the `prep' table only and thus
valid for all glyphs. However, there are fonts like `pala.ttf'
which additionally contain this instruction in the hinting code of
some glyphs (but not all). As a result it can happen that a
composite glyph needs multiple drop-out modes for its subglyphs
since the rendering state gets reset for each subglyph.
FreeType collects the hinted outlines from all subglyphs, then it
sends the data to the rasterizer. It also sends the drop-out mode
-- after hinting has been applied -- and here is the error: It sends
the drop-out mode of the last subglyph only; drop-out modes of all
other subglyphs are lost.
This patch fixes the problem; it adds a second, alternative
mechanism to pass the drop-out mode: For each contour, the
rasterizer now checks the first `tags' array element. If bit 2 is
set, bits 5-7 contain the contour's drop-out mode, overriding the
global drop-out mode.
* include/freetype/ftimage.h (FT_CURVE_TAG_HAS_SCANMODE): New macro.
* src/truetype/ttgload.c (TT_Hint_Glyph): Store drop-out mode in
`tags[0]'.
* src/raster/ftraster.c (Flow_Up, Overshoot_Top, Overshoot_Bottom):
Use bits 3-5 instead of 0-2.
(New_Profile): Set the drop-out mode in the profile's `flags' field.
(Decompose_Curve): Check `tags[0]' and set `dropOutControl' if
necessary.
(Vertical_Sweep_Drop, Horizontal_Sweep_Drop,
Horizontal_Gray_Sweep_Drop, Draw_Sweep): Use the profile's drop-out
mode.
Two new constraints are introduced to better identify a `stub' -- a
concept which is only vaguely described in the OpenType
specification. The old code was too rigorous and suppressed more
pixel than it should.
. The intersection of the two profiles with the scanline is less
than a half pixel. Code related to this was already present in
the sources but has been commented out.
. The endpoint of the original contour forming a profile has a
distance (`overshoot') less than half a pixel to the scanline.
Note that the two additional conditions fix almost all differences
to the Windows rasterizer, but some problematic cases remain.
* src/raster/ftraster.c (Overshoot_Top, Overshoot_Bottom): New
macros for the `flags' field in the `TProfile' structure.
(IS_BOTTOM_OVERSHOOT, IS_TOP_OVERSHOOT): New macros.
(New_Profile, End_Profile): Pass overshoot flag as an argument and
set it accordingly.
Update callers.
(Vertical_Sweep_Drop, Horizontal_Sweep_Drop): Implement the two new
constraints.
* src/raster/ftraster.c (Set_High_Precision): Add two more bits to
the precision. This corrects rendering of some small glyphs, for
example, glyph `xi' in verdana.ttf at 13 ppem. Testing with ftbench
on my GNU/Linux box I don't see a performance degradation.
