The hb_font_t object holds on to tables of font data, acquired through
the MinikinFont::GetTable interface, which is based on copying data
into caller-owned buffers. Now that we're caching lots of hb_font_t's,
the cost of these buffers is significant.
This patch moves to a different interface, inspired by HarfBuzz's
hb_reference_table API, where the font can provide a pointer to the
actual font data (which will often be mmap'ed, so it doesn't even
consume physical RAM).
Bug: 27860101
Change-Id: Id766ab16a8d342bf7322a90e076e801271d527d4
am: 1ea4165
* commit '1ea4165cef7651770fe28a0eada3da593bb149ad':
Purge hb font on Minikin font destruction
Change-Id: Idb2034353407e8b390cc07862803f846d5adea0c
This patch eagerly purges the corresponding hb_font_t object from
the HbFontCache when the underlying MinikinFont is destroyed. After
that, the key will no longer be accessed, so having the entry is
wastes memory.
Bug: 27251075
Bug: 27860101
Change-Id: I1b98016133fe3baf6525ac37d970a65ddccadb4f
am: dbcbe1f
* commit 'dbcbe1f426b17242f2c548fb3df5e2b6a659ac50':
Support multiple locales for font language settings.
Change-Id: Iad60bf1c01f309e70c3b1ad22e28b22afe44b33f
Some fonts support multiple scripts, for example, some fonts for
Korean supports not only "Kore" but also "Jamo".
To select fonts based on their multiple languages, this CL introduces
the following changes:
- Compares all languages of the font family and use the maximum score
for font selection.
- Even if each language of the font family doesn't support the requested
language, the font get score of 2 if the requested font is covered by
all of the languages of the font family. For example, the font for
"ko-Hang,ko-Hani" gets score of 2 for the requested language "ko-Kore".
Bug: 26687969
Change-Id: I7f13b51464c9b01982bb573251d77052b9ddbd70
am: c9c0359
* commit 'c9c0359b1d0711c0dceeef7c853c6377c6a2af08':
Try putting combining marks in the same font run
Change-Id: I6f76b791acde2dc8209e8d33da02d148a9f790d8
Indic combining marks, when combined with a common character such as
a hyphen or a dotted circle, used to get rendered in a different font
due to the greedy algorithm used in determining runs, which resulted
in the base character and the combining mark getting rendered in
separate font runs, resulting in a dotted circle appearing in phrases
such as "100-ാം" (0031 0030 0030 002D 0D3E 0D02).
This change makes combining marks change the font run of the base
character if the base character is supported in the same font as the
combining mark, similar to the support for emoji modifiers and the
combining keycap.
Bug: 25036888
Bug: 24535344
Change-Id: I8e2798e8ecb8efaf723a0fd02c05c6fbdef8b365
Implement the change proposed in UTC document L2/16-043R
(http://www.unicode.org/L2/L2016/16043r-line-break-pr-po.txt) to make
sure we do not break between letters and currency symbols.
Bug: 24959657
Change-Id: Ia29d0e5625f84870bd910d0c6e19036d17206704
An emoji base with an emoji modifier renders as a single glyph and
thus should not be a line break. Current (Unicode 8) logic does
indicate a line break, so we override the results of the ICU line
break iterator. The code references a proposal to improve Unicode
behavior; when that is adopted and we upgrade ICU accordingly, the
special-case code should be deleted, but the tests can remain.
Bug: 27343378
Change-Id: I5de9c53e9a34c503816f9131e3d894e6f7a57d13
If skin tone is specified, the base emoji should be emoji style even
if it is text presentation default emoji.
This patch also removes wrong test case which expects default emoji
presentation but it is controlled by family order in /etc/fonts.xml
and there is no special logic for default presentation in minikin.
Thus the default presentation unit test should not be in minikin.
Bug: 27342346
Change-Id: I74a2b2feab4d559535049e368cfd833063cce81c
An emoji with a modifier should be treated as a single grapheme, i.e.
it should not be possible to place the cursor between the base and
modifier. This patch implements the proposed Rule GB9c from Mark
Davis's proposal entitled "Fixing breaking properties for emoji",
L2/16-011R3.
The patch also skips over variation sequences attached the to the
preceding character, for computing grapheme cluster boundaries.
Bug: 26829153
Change-Id: Iff5bc2bb8e5246223a017c7cf33acfbf63817f16
Due to the way emoji ZWJ sequences are defined, the ICU line breaking
algorithm determines that there are valid line breaks inside the
sequence. This patch suppresses these line breaks.
This is an adaptation of I225ebebc0f4186e4b8f48fee399c4a62b3f0218a
into the nyc-dev branch.
Bug: 25433289
Change-Id: I84b50b1e6ef13d436965eab389659d02a30d100f
With this CL, measureText is used for getRunAdvance,
getOffsetForAdvance and line breaking.
Bug: 24505153
Change-Id: Ib699f6b1391b46537736fc274cdb41686586b550
In cases when a word (as defined by the ICU break iterator) overlaps a
style boundary, the returned wordStart can be extend before the range
currently being measured for layout. When we try to hyphenate the
resulting substrings, we get a negative range, which crashes. This
patch disables hyphenation in this case.
Bug: 27237112
Change-Id: I76d04b39dd3b4d6d267aaaf4bebc9ab361891646
