This change comes along with 2a7bb4596f56 ans is only meant to reduce
pointer casting in the code.
* include/freetype/ftimage.h (FT_Outline): Do it.
* src/*: Update `FT_Outline` users.
In C it is undefined behavior to do arithmetic on a null pointer, including
adding zero. When using NotoSansKhmer[wdth,wght].ttf UBSAN produces a report
like
ttgxvar.c:1052:31: runtime error: applying zero offset to null pointer
when adding zero to `varData->deltaSet` (which is null) to produce `bytes`.
Protect against all the potential issues of this kind by returning early if
`varData->regionIdxCount == 0`.
* src/truetype/ttgxvar.c (tt_var_get_item_delta): early return on no regions
This doubles the number or allowed points, see
https://github.com/harfbuzz/harfbuzz/issues/4752
Although it is hardly practical to use more than 32767 points,
other font engines seem to support it.
* docs/CHANGES: Announce it.
* include/freetype/ftimage.h (FT_Outline): Do it and update limits.
* src/*: Update `FT_Outline` users.
Instead of validating variation axes in every access, OpenType specs
suggest that peak = 0 be used to tag invalid ranges. This implements
just that once during loading.
* src/cff/cffload.c (cff_blend_build_vector): Move the range checks...
(cff_vstore_load): ... here.
* src/truetype/ttgxvar.c (tt_var_get_item_delta): Ditto...
(tt_var_load_item_variation_store): ... ditto.
This should achieve quicker results for common cases.
* src/cff/cffload.c (cff_blend_build_vector): Rearrange conditionals.
* src/truetype/ttgxvar.c (tt_var_get_item_delta): Ditto.
* include/freetype/internal/t1types.h: Host PS_DesignMap and PS_Blend.
* include/freetype/ftmm.h: Host and document TT_MAX_MM_XXX.
* include/freetype/t1tables.h: Remove them from here.
`T1_FIELD_ZERO` is used to zero initialize a `T1_FieldRec`.
`T1_FIELD_ZERO` is currently initilizing `T1_FieldRec::location` with a
`T1_FieldType` and `T1_FieldRec::type` with a `T1_FieldLocation`. This
was detected with `-Wenum-conversion`.
* include/freetype/internal/psaux.h (T1_FIELD_ZERO): correct order of
initalizers
In 32 bit builds `FT_ULong` is 32 bits and can silently overflow when a
large number is read into one and then it is summed or multiplied with
another number. Checks for range overflow must be written so that they
themselves do not overflow. Also ensure that the table_size is always the
first part of the range check and consistently use `<` or `<=`.
* src/sfnt/ttcolr.c (tt_face_load_colr): Avoid overflow.
(find_base_glyph_v1_record): Remove old work-around.
Bug: https://issues.chromium.org/issues/41495455
Bug: https://issues.chromium.org/issues/40945818
FT_SqrtFixed (95b0fe2a6dff) is faster and does not overflow.
* src/sdf/ftsdfcommin.h (square_root): Replace with a macro.
* src/sdf/ftsdfcommin.c (square_root): Remove function.
The upper limit of 4095 is implied by the SFNT header format
where the multiplication by 16 would overflow without it.
* src/sfnt/sfwoff.c (woff_open_font): Updated.
* src/sfnt/sfwoff2.c (woff2_open_font): Ditto.
The existing code already disallows zero table woff2 overall, but still
allows for individual CollectionFontEntry to create font instances with
zero tables. Such fonts are not useful so error early.
This also fixes an MSAN discovered issue where if a CollectionFontEntry
numTables is zero then the sfnt_header was not fully initialized.
* src/sfnt/sfwoff2.c (woff2_open_font): error on zero tables, always
initalize sfnt_header
Fixes: https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=68384
Resetting of the size nodes can crash, if a parent face purge happens
simultaneously and destroys the node. It is safer to create a new node.
Fixes#1270, reopened with a separate issue.
* src/cache/ftcbasic.c (FTC_BasicAttrRec): Match type of `load_flags`
to its main purpose in `FT_Load_Glyph`.
(FTC_ImageCache_Lookup{,Scaler},FTC_SBitCache_Lookup{,Scaler}): Updated.
Manipulate the cache after a face is requested or a size is looked up
successfully. Fixes#1270.
* src/cache/ftcmanag.c (ftc_size_node_init, ftc_size_node_reset,
ftc_face_node_init): Check for errors before accepting a change.
* src/cache/ftcmru.c (FTC_MruList_New): Do nothing if reset fails.
With horizontal bisections, the smallest section is a whole single
scanline. Almost horizontal lines or other complex scanlines can
easily overflow the rendering pool. Switching to vertical bisections
splits the scanlines and should rule out the overflows. Fixes#1269.
* src/smooth/ftgrays.c (gray_convert_glyph): Bisect vertically.
When creating a CID parser the location of the 'StartData' or '/sfnts'
tokens needs to be known. However, the token parser requires that the
entire document be in memory and flattening the entire stream into memory is
to be avoided.
To avoid forcing the entire stream into memory, previously this code would
scan through the stream looking for 'StartData' or '/sfnts' as strings.
However, these strings could have been in a comment or string token, so the
stream would be read into memory up to that point and the parser run to
check that these strings were actually tokens. This forced a parser restart
from the beginning each time; as a result, data with many 'StartData'
non-tokens would take n^2 time to check.
* src/cid/cidparse.c (cid_parser_new): Change algorithm to make the initial
scan look for the last possible 'StartData' or '/sfnts' string in the
stream. The stream is read forward instead of backward as a typical normal
CID font will have one 'StartData' toward the beginning of the data and it
it much faster to read the data from beginning to end instead of end to
beginning. For memory-based fonts the limit is set to the end of the stream
since the stream is already in memory. Then the parser is run once to look
for 'StartData' or '/sfnts' tokens. If they are found the parser is re-set
to reflect this new information.
Reported as
https://issues.chromium.org/issues/40201695
For default variable instances `cff_face_init` did not set the blend. This
mostly worked as later use of the unset blend produced the default
variation. However, if a user called `TT_Get_MM_Var` the blend would be
partially set up, but not fully. In particular the number of axes, the axis
definitions, and the instance locations would be set up, but not the current
instance location (`coords` and `normalizedcoords`). This could lead to the
default instances of CFF2 fonts erroring on any use of `blend`.
Ensure the default variable instance is fully set up by always calling
`FT_Set_Named_Instance` on a variable face.
* src/cff/cffobjs.c (cff_face_init): Call `FT_Set_Named_Instance` on
default instances.
* src/truetype/ttobjs.c (tt_face_init): Ditto.
Fixes#1268.
This commit adds support for kerning from 'GPOS' tables, while maintaining
support for basic 'kern' tables. `FT_HAS_KERNING` will be true for a font
with either available and `FT_Get_Kerning` will still use the basic 'kern'
table data if avilable, otherwise check the GPOS 'kern' feature.
This feature is disabled by default; it can be enabled with the
`TT_CONFIG_OPTION_GPOS_KERNING` flag.
Only basic kerning (pair positioning with just an x advance) is supported
from the GPOS layout features; support for that was added to make the
existing `FT_Get_Kerning` API more consistently functional. FreeType does
not intend to extend itself to further GPOS functionality though; a
higher-level library like HarfBuzz can be used instead for that.
* include/freetype/config/ftoption.h, include/devel/ftoption.h
(TT_CONFIG_OPTION_GPOS_KERNING): New configuration option.
* include/freetype/internal/fttrace.h: Add `ttgpos` trace handler.
* include/freetype/internal/sfnt.h (SFNT_Interface): Add `load_gpos` and
`get_gpos_kerning` fields.
(FT_DEFINE_SFNT_INTERFACE): Updated.
* include/freetype/internal/tttypes.h: Include `fttypes.h`.
(TT_FaceRec) [TT_CONFIG_OPTION_GPOS_KERNING]: Add `gpos_table` and
`gpos_kerning_available` fields.
* src/sfnt/ttgpos.c, src/sfnt/ttgpos.h: New files.
* src/sfnt/sfdriver.c [TT_CONFIG_OPTION_GPOS_KERNING]: Include `ttgpos.h`.
(sfnt_interface): Updated.
* src/sfnt/sfnt.c: Include `ttgpos.c`.
* src/sfnt/sfobjs.c [TT_CONFIG_OPTION_GPOS_KERNING]: Include `ttgpos.h`.
(sfnt_load_face) [TT_CONFIG_OPTION_GPOS_KERNING]: Load and free GPOS kerning
data; check GPOS kerning availability.
* src/truetype/ttdriver.c (tt_get_kerning): Use GPOS kerning if there's no
'kern' table.
As a result of 7b308a29dd10, the regular 64-bit execution is now faster
than SSE2. The rendering speed of script fonts at 64 ppem or larger is
improved by about 3% without SSE2. See !314 for the testing results.
* src/smooth/ftgrays.c (gray_render_conic)[FT_INT64]: Remove SSE2 code.
A font has surfaced with `post` version 1.0 and fewer than 258 glyphs.
Its glyphs did not correspond to their names. We now reject such `post`
strictly following specifications.
* src/sfnt/ttpost.c (tt_face_get_ps_name): Check the number of glyphs
for version 1.0.
Reported as
https://bugs.chromium.org/p/chromium/issues/detail?id=1505216
* src/sfnt/ttcolr.c (find_base_glyph_v1_record): Guard access of the search
pointer during binary search. The pointer needs to be checked as we go as
the test that compares number of v1 glyphs with table size at the time of
loading the table is not sufficient on its own.
A scenario is possible in which the `BaseGlyphRecord` list extends into
non-`BaseGlyphRecord` parts of the 'COLR' v1 table (but passed the size
comparison check). Then, at those locations, invalid glyph ID values are
read and may provoke an invalid read due to reassigning min and max values
during the binary search.
This is a follow-up to commit 26a7f047,
[cff] Make blend operator work with floats in private dicts.
which addressed the 'party baseline' bug. However, the reporting user
indicated that the default location and some other points in design space
rendered OK, but other points in design space still had problems. The most
obvious issue being that the x-heights of lower-case letters did not align;
see
https://github.com/adobe-fonts/source-serif/issues/121#issuecomment-1773794136
After some analysis we determined that this was due to an interaction
between `BlueValue` rounding and the zone-based algorithm. In short, for a
point to be considered in a zone it must fall within the bounds of the zone.
(There is a slop factor in some cases, but only a very small one.) In the
Adobe-contributed side of the code, point values are not integer-rounded,
instead they're kept as (some form of) fixed. Rounding just the `BlueValues`
means that points that need to be considered within a zone will fall outside
of it at some points in design space.
The majority of this patch changes the storage and parsing of `BlueValues`
to keep them as `FT_Fixed`. No significant code changes were needed because
the values are converted to `Fixed` anyway when stored in `CF_BlueRec`. No
attempt was made to address problems in the older pshinter code beyond
converting the values from `FT_Fixed` to `FT_Short` when copying the private
dictionary. (However, as the point values are also rounded in that code,
the problem is much less likely to occur, although inconsistency between
rounding and truncation could cause an analogous problem.)
* include/freetype/internal/cfftypes.h (CFF_PrivateRec): Use `FT_Fixed` for
`blue_values`, `other_blues`, `family_blues`, and `family_other_blues`.
* src/cff/cffload.c (cff_blend_doBlend): Updated.
* src/cff/cffobjs.c (CFF_fixedToInt): New macro.
(cff_make_private_dict): Use it.
* src/cff/cffparse.h (cff_kind_delta_fixed): New enum value.
* src/cff/cffparse.c (do_fixed): Updated.
(CFF_FIELD_DELTA, CFF_FIELD_DELTA_FIXED, CFF_DELTA_KIND): New set of macros,
replacing `CFF_FIELD_DELTA`.
(cff_parser_run): Updated to handle fixed-float deltas.
* src/cff/cfftoken.h: Updated to use `CFF_FIELD_DELTA_FIXED` for blue
values.
* src/psaux/psblues.c (cf2_blueToFixed): Removed, no longer needed.
(cf2_blues_init): Updated.
* src/pxaux/psft.c, src/pxaux/psft.h (cf2_getBlueValues, cf2_getOtherBlues,
cf2_getFamilyBlues, cf2_getFamilyOtherBlues): Updated signatures.
* src/psaux/psobjs.c (t1_make_subfont): Updated.
MSBuild.exe can now be fired from the root folder without options.
* builds/windows/vc2010/freetype.sln: Relocated to...
* MSBuild.sln: ... here with minor changes.
* MSBuild.rsp: New file with command line options.
* docs/INSTALL: Reference `MSBuild.sln`.
While it is recommended to have an explicit point at each curve
extrema, they might be missing or outline could be rotated. This
leads to excessive bisections in raster to find them. This change
helps to decrease the number of bisections. The scanline
intersections remain monotonous, of course.
* src/raster/ftraster.c (Conic_To, Cubic_To): Check that control
points cross the scanlines to bisect.
This reduces the code duplication.
* src/raster/ftraster.c (Function_Sweep_Span): Change signature.
(Vertical_Sweep_Drop, Horizontal_Sweep_Drop): Focus on pixel setting
and move duplicated the dropout control logic to...
(Draw_Sweep): ... this function and refactor.
(Vertical_Sweep_Span, Horizontal_Sweep_Span): Minor.
This results in noticeable performance improvement.
* src/raster/ftraster.c (Insert_Y_Turns): All bottom y-turns are still
sorted and stored; only the maximum top y-turn is recorded now. The
function is renamed.
(End_Profile): Updated accordingly.
(Convert_Glyph): Reserve space for the top value.
* src/raster/ftraster.c (black_TWorker): Remove the fresh flag.
(New_Profile): Set the starting scanline here based on the current
coordinate and initialize the joint crossing if necessary.
(Line_Up, Bezier_Up): Do not deal with fresh and joint starts at all.
(Line_Down, Bezier_Down): Simplify.
(Conic_To, Cubic_To): Update the current coordinate after each
subsection.
* src/raster/ftraster.c (Line_Up, Bezier_Up): Deal with the scanline
joints directly based on the initial y-coordinate.
(New_Profile, black_TWorker): Remove the boolean flag.
* src/raster/ftraster.c (black_TWorker): Store signed maximum indexes
in both directions instead of unsigned dimensions of the original bitmap.
(*_Sweep_*, Render_Glyph, ft_black_render): Updated all users.
* src/raster/ftraster.c (TProfile): Get rid of `countL`.
(Draw_Sweep): Use `start` for countdown to activation.
(Horizontal_Sweep_Drop, Vertical_Sweep_Drop): Rely on `height` and
`offset` to verify profile ends for the stub detection.
* src/raster/ftraster.c (s/Sort/Increment): Rename this function to
reflect its true purpose, delete exhausted profiles here...
(Draw_Sweep): ... instead of here.
While curving close to a pixel center, vertical and horizontal pass
might split the curve differently and cause a rare dropout. This
makes the split condition invariant of the sweep direction and more
robust.
* src/raster/ftraster.c (Bezier_Up): Modify the split condition.
The jitter exception used to be applied when two neighboring pixels
were barely inside the outline. One the left one was turned on then,
which contradicts the OpenType specifications. Intended to remove
glitches, it caused disappearing lines and was softened by adding an
exception to the exception (#54589).
* src/raster/ftraster.c (Vertical_Sweep_Span): Drop the jitter exception.
* src/raster/ftraster.c (Draw_Sweep): Use y-turns to set the range,
correctly set the initial position, directly loop through y_turns,
and remove a 5-gray remnant at exit.
This is mostly cosmetic and removes a few casts, plus Short is
promoted to Int in calculations anyway.
* src/raster/ftraster.c (Vertical_Sweep_Init, Vertical_Sweep_Span,
Vertical_Sweep_Drop, Horizontal_Sweep_Init, Horizontal_Sweep_Span,
Horizontal_Sweep_Drop, Draw_Sweep): Mostly s/Short/Int/ and remove
casting.
This only helps to delay the pool overflow and bisections to larger
sizes and benefits only very intricate glyphs at reasonable sizes.
* src/raster/ftraster.c (TProfile): Use Int instead of Long or PLong
when it is sufficient.
(New_Profile, End_Profuile, Bezier_Up, Sort): Updated accordingly.
* src/raster/ftraster.c (TProfile): Include the variable array member
and repackage with pointers first.
(New_Profile): Advance the top using the variable array pointer.
These ancient builtins have been supported by clang since 2013. We
condition it somewhat stricter but still around 2017. This is more
portable than `__has_builtin`. Fixes#1260.
* src/raster/ftraster.c (New_Profile): Set important fields only and
delay setting `gProfile` until...
(End_Profile): ... it is checked to be valid here.
(Convert_Glyph): Updated.
This fixes a subtle bug when the last profile in a contour was not
properly short-circuited if it was still empty at `End_Profile`.
We finalize all linking in `Finalize_Profile_List` now and do nothing
else there. The turns are added in `End_Profile`.
* src/raster/ftraster.c (Insert_Y_Turn): Moved up unchanged.
(End_Profile): Take care of turns but set only preliminary linking.
(Finalize_Profile_Table): Take care of linking and null-termination.
(Convert_Glyph): Adjusted accordingly.
* src/raster/ftraster.c (End_Profile): Do not initiate next profile.
(New_Profile): Fully initiate new profile.
(Convert_Glyph): Clean up variables, initialize `fProfile` here.
The general square root calculations are not necessary in FreeType.
For vector normalization or length, FreeType uses special functions.
It is, however, required in the legacy CFF specifications.
* src/base/ftcalc.c (FT_SqrtFixed): New function that uses either
Babylonian or bit-wise algorithm, whichever is faster for the given
situation.
* include/freetype/internal/ftcalc.h (FT_SqrtFixed): Declare it.
When matching the keywords, we avoid calculating their lengths by
checking the stored values. This itself is a sufficient pre-check
before diving into `memcmp`. Therefore, we remove explicit check of
the first characters.
* include/freetype/internal/psaux.h (T1_FieldRec): Store length.
* src/cid/cidload.c (cid_parse_dict): Use `memcmp` and stored length.
* src/type1/t1load.c (parse_dict): Ditto.
* src/type42/t42parse.c (t42_parse_dict): Ditto.
* src/cff/cffobjs.c (remove_style): Rewrite using pointers.
(remove_subset_prefix): Unwrap loop and use `memmove`.
* src/truetype/ttobjs.c (tt_skip_pdffont_random_tag): Unwrap loop
and avoid `strlen`.
This also reduces the used heap size by a large factor.
From Behdad.
* src/autofit/afcjk.h (AF_CJKAxisRec): Use `AF_BLUE_STRINGSET_MAX_LEN`.
* src/autofit/aflatin.h (AF_LatinAxisRec): Ditto.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
