We had a mix of 0.6.2 and 0.6.5 (which is the current release),
this unifies to the latest version. It also enables the union
feature which removes the discriminant, reducing memory usage.
Before this change, if we needed to create a Font which we've already
created, but at a new size, then we'd fetch the FontTemplateInfo again.
If the bytes of the font are held in memory, then this could be
expensive as we need to pass those bytes over IPC.
FontTemplateData gets passed over IPC during the communication between
FontContext and FontCacheThread. Serializing and deserializing these
bytes is expensive, so this change ensures that we only do that when the
bytes can't be read from disk. A similar strategy is already used on
macos and windows.
The performance problem was particularly noticeable after implenting
font fallback, where the content process would potentially work through
a list of fonts, trying to find one which contains a certain glyph. That
could result in lots of font bytes going over IPC.
Prior to this change, if none of the fonts specified in CSS contained a
glyph for a codepoint, we tried only one fallback font. If that font
didn't contain the glyph, we'd give up.
With this change, we try multiple fonts in turn. The font names we try
differ across each platform, and based on the codepoint we're trying to
match. The current implementation is heavily inspired by the analogous
code in Gecko, but I've used to ucd lib to make it more readable,
whereas Gecko matches raw unicode ranges.
This fixes some of the issues reported in #17267, although colour emoji
support is not implemented.
== Notes on changes to WPT metadata ==
=== css/css-text/i18n/css3-text-line-break-opclns-* ===
A bunch of these have started failing on macos when they previously
passed.
These tests check that the browser automatically inserts line breaks
near certain characters that are classified as "opening and closing
punctuation". The idea is that if we have e.g. an opening parenthesis,
it does not make sense for it to appear at the end of a line box; it
should "stick" to the next character and go into the next line box.
Before this change, a lot of these codepoints rendered as a missing
glyph on Mac and Linux. In some cases, that meant that the test was
passing.
After this change, a bunch of these codepoints are now rendering glyphs
on Mac (but not Linux). In some cases, the test should continue to pass
where it previously did when rendering with the missing glyph.
However, it seems this has also exposed a layout bug. The "ref" div in
these tests contains a <br> element, and it seems that this, combined
with these punctuation characters, makes the spacing between glyphs ever
so slightly different to the "test" div. (Speculation: might be
something to do with shaping?)
Therefore I've had to mark a bunch of these tests failing on mac.
=== css/css-text/i18n/css3-text-line-break-baspglwj-* ===
Some of these previously passed on Mac due to a missing glyph. Now that
we're rendering the correct glyph, they are failing.
=== css/css-text/word-break/word-break-normal-bo-000.html ===
The characters now render correctly on Mac, and the test is passing. But
we do not find a suitable fallback font on Linux, so it is still failing
on that platform.
=== css/css-text/word-break/word-break-break-all-007.html ===
This was previously passing on Mac, but only because missing character
glyphs were rendered. Now that a fallback font is able to be found, it
(correctly) fails.
=== mozilla/tests/css/font_fallback_* ===
These are new tests added in this commit. 01 and 02 are marked failing
on Linux because the builders don't have the appropriate fonts installed
(that will be a follow-up).
Fix build errors from rebase
FontTemplateDescriptor can no longer just derive(Hash). We need to
implement it on each component part, because the components now
generally wrap floats, which do not impl Hash because of NaN. However in
this case we know that we won't have a NaN, so it is safe to manually
impl Hash.
Move DL items from gfx to layout and implement corner clipping
Implement corner clipping.
Remove PixelFormat from WebrenderImageInfo.
Use WebRender text shadow.
Remove MallocSizeOf and Deserialize for DL items.
Closes#19649, closes#19680, closes#19802
<!-- Please describe your changes on the following line: -->
---
<!-- Thank you for contributing to Servo! Please replace each `[ ]` by `[X]` when the step is complete, and replace `__` with appropriate data: -->
- [x] `./mach build -d` does not report any errors
- [x] `./mach test-tidy` does not report any errors
- [x] These changes fix #__ (github issue number if applicable).
<!-- Either: -->
- [x] There are tests for these changes OR
- [ ] These changes do not require tests because _____
<!-- Also, please make sure that "Allow edits from maintainers" checkbox is checked, so that we can help you if you get stuck somewhere along the way.-->
<!-- Pull requests that do not address these steps are welcome, but they will require additional verification as part of the review process. -->
<!-- Reviewable:start -->
---
This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/servo/servo/20420)
<!-- Reviewable:end -->
Implement corner clipping.
Remove PixelFormat from WebrenderImageInfo.
Use WebRender text shadow.
Remove MallocSizeOf and Deserialize for DL items.
Closes#19649, #19680, #19802
Add an IndexableText structure for text queries.
Instead of linear search for a node this now uses a HashMap.
Remove the now irrelevant fields from TextDisplayItem.
Lazy load fonts in a FontGroup
The first commit message explains this so I'll just copy it here:
---
This is a step towards fixing #17267. To fix that, we need to be able to
try various different fallback fonts in turn, which would become
unweildy with the prior eager-loading strategy.
Prior to this change, FontGroup loaded up all Font instances, including
the fallback font, before any of them were checked for the presence of
the glyphs we're trying to render.
So for the following CSS:
font-family: Helvetica, Arial;
The FontGroup would contain a Font instance for Helvetica, and a Font
instance for Arial, and a Font instance for the fallback font.
It may be that Helvetica contains glyphs for every character in the
document, and therefore Arial and the fallback font are not needed at
all.
This change makes the strategy lazy, so that we'll only create a Font
for Arial if we cannot find a glyph within Helvetica. I've also
substantially refactored the existing code in the process and added
some documentation along the way.
---
I've added some tests in the second commit, but it required quite a bit of gymnastics to make it possible to write such a test. I'm not sure if the added complexity to the production code is worth it?
On the other hand, having this infrastructure in place may be useful for testing future changes in this area, and also possibly brings us a step closer to extracting a library as discussed in #4901. (What I mean by that is: it reduces coupling between `FontCacheThread` and `FontContext` -- the latter would have a place in such a library, the former wouldn't.)
<!-- Reviewable:start -->
---
This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/servo/servo/20021)
<!-- Reviewable:end -->
Unfortunately, this required quite a bit of changes to the non-test
code. That's because FontContext depends on a FontCacheThread, which in
turn depends on a CoreResourceThread and therefore lots of other data
structures.
It seemed like it would be very difficult to instantiate a FontContext
as it was, and even if we could it seems like overkill to have all these
data structures present for a relatively focused test.
Therefore, I created a FontSource trait which represents the interface
which FontContext uses to talk to FontCacheThread. FontCacheThread then
implements FontSource. Then, in the test, we can create a dummy
implementation of FontSource rather than using FontCacheThread.
This actually has the advantage that we can make our dummy
implementation behave in certain specific way which are useful for
testing, for example it can count the number of times
find_font_template() is called, which helps us verify that
caching/lazy-loading is working as intended.
This is a step towards fixing #17267. To fix that, we need to be able to
try various different fallback fonts in turn, which would become
unweildy with the prior eager-loading strategy.
Prior to this change, FontGroup loaded up all Font instances, including
the fallback font, before any of them were checked for the presence of
the glyphs we're trying to render.
So for the following CSS:
font-family: Helvetica, Arial;
The FontGroup would contain a Font instance for Helvetica, and a Font
instance for Arial, and a Font instance for the fallback font.
It may be that Helvetica contains glyphs for every character in the
document, and therefore Arial and the fallback font are not needed at
all.
This change makes the strategy lazy, so that we'll only create a Font
for Arial if we cannot find a glyph within Helvetica. I've also
substantially refactored the existing code in the process and added
some documentation along the way.
Fix FontTemplateDescriptor under FreeType
Issue #17321. Under Linux, using "font-family: sans-serif" previously
caused Servo to select the "UltraLight" face (of DejaVu Sans). There were
two reasons for this:
1. Font weight was only retrieved from the OS/2 table for bold faces.
This neglected to retrieve the weight information for "lighter than
normal" weight faces. This meant that the UltraLight face appeared as
normal weight, and was selected.
2. Retrieval of font stretch information from the OS/2 table was not
implemented at all.
<!-- Reviewable:start -->
---
This change is [<img src="https://reviewable.io/review_button.svg" height="34" align="absmiddle" alt="Reviewable"/>](https://reviewable.io/reviews/servo/servo/19928)
<!-- Reviewable:end -->
Issue #17321. Under Linux, using "font-family: sans-serif" previously
caused Servo to select the "UltraLight" face (of DejaVu Sans). There
were two reasons for this:
1. Font weight was only retrieved from the OS/2 table for bold faces.
This neglected to retrieve the weight information for "lighter than
normal" weight faces. This meant that the UltraLight face appeared as
normal weight, and was selected.
2. Retrieval of font stretch information from the OS/2 table was not
implemented at all.
Previously anything with the wrong stretch/italicness was considered
equally bad. Now we consider the wrong weight to be least bad, wrong
boldness of intermediate badness, and wrong itallicness to be most
bad.
The current implementation for boldness assumes that if Freetype
claims a font is not bold then it must have normal weight. However
fonts with lighter-than-normal weight are not claimed to be bold.
Therefore we get rid of this check.
This allows servo to use the ExternalScrollId API from WebRender fixing
some issues related to duplicate scroll root ids.
Fixes#17176.
Fixes#19287.
Fixes#19648.
