Reviewed by Ryosuke Niwa.
[BiDi] [CSS3] MASTER: Add support for the unicode-bidi:isolate CSS property
https://bugs.webkit.org/show_bug.cgi?id=50912
Source/WebCore:
This patch adds support for CSS3 unicode-bidi: isolate property, under the -webkit- vendor prefix.
Parsing support was added in a previous patch, this wires up the RenderStyle values
to code changes in the BidiResolver.
The effect of this patch is that it makes it possible to "isolate" runs of text
so that their RTL-ness or LTR-ness does not bleed out into the rest of your text
and effect layout. This is important because many unicode characters (like parenthesis, ':', '-', etc.)
do not have intrinsic directionality and are affected by whatever characters come before/after.
If you have usernames which include RTL text, if you inject those usernames in your page
you might end up with nearby characters moving!
(like 'RTL USERNAME - my awesome site' as a title, could end up as
'my awesome site - USERNAME RTL' when correct would be 'USERNAME RTL - my awesome site'.)
This patch makes it possible to wrap sections of text in isolated spans, so that
they correctly order all their RTL/LTR contents, but also correctly participate in the
larger RTL/LTR ordering without affecting nearby characters.
Because much of this code is old and rarely touched, I've included extra background
information in hopes of expanding my set of potential reviewers:
WebKit uses the standard "Unicode Bidi Algorithm" henceforth known as the UBA.
The UBA is defined at http://unicode.org/reports/tr9/ for those not faint of heart.
Text layout is done per-block (<div>, <p>, etc), and begins with a string of text
(which in our case comes from the rendering tree) and a specified width.
First: Text is measured and wrapped into lines.
Second: The UBA is run over the lines of text.
Third: WebKit builds InlineBoxes (its linebox tree) and eventually render the text.
This patch modifies our UBA to ignore all text content inside "isolated" inlines (treating them as neutral characters)
and then adds another step after running the UBA, where we run the UBA recursively on any
previously identified "isolated" content.
The result of the UBA is an ordered list of "runs" of text with the RTL runs
correctly RTL and the LTR runs LTR.
The UBA does three things:
1. It assigns a "class" to each character in a text stream (like neutral, strongly-RTL, strongly-LTR, etc.)
2. Divides the text stream up into "runs" of characters of the same directionality (all RTL, all LTR).
3. Re-orders those runs.
The UBA in WebKit is implemented by BidiResolver<T> in BidiResolver.h
The InlineBidiResolver (BidiResolver specialization which knows about the rendering tree)
walks along its InlineIterators, looking at each character and running the
Unicode Bidi Algorithm (UBA). It walks through the rendering tree subtree under
a block, using a (poorly named) bidiNext function which returns the next inline object.
Each inline object (or text character there-in) has a corresponding meaning in the UBA
such as a "strong RTL" character or a "neutral" character. The UBA reads these sequence
of characters, and figures out what direction (RTL or LTR) to assign to any neutral
characters it encounters, based on surrounding characters.
As the InlineBidiResolver is walking the rendering tree, the InlineIterator::advance()
function calls bidiNext(), which in turn can call notifyObserverEnteredObject/notifyObserverWillExitObject
notifying InlineBidiResolver that it is entering or exiting an "isolated"
span, at which point it will either start or stop ignoring the stream of characters
from the InlineIterator. When the InlineBidiResolver is ignoring the stream of
characters, instead of creating separate BidiRuns at each RTL/LTR boundary
as it normally would, it instead creates one "fake" run for the entire
isolated span. These fake runs participate in the normal UBA run ordering process,
but after the main UBA, a second pass is made where we examine
the list of isolatedRuns() and run the UBA on each of them, replacing the fake
run we previously inserted, with the resulting list of runs from that inner UBA run.
The way it "ignores" characters is by treating them all as neutral when inside an isolate.
Thus all the characters end up grouped in a single run, but their directionality (as a group)
is correctly affected by any surrounding strong characters.
If you understood that last paragraph, than the rest of the change is just plumbing.
I added a huge number of FIXMEs to this code, because this code has a variety of
design choices (or lack there of) which make some of this very difficult.
For example the bidiNext iterator function has two sets of mutually exclusive
parameters and can be used optionally with or without an observer. Prior to this
change there was only ever one object which cared about observing a walk over inlines
and that was InlineBidiResolver. This patch (regretfully) templatizes bidiNext
to support a new Observer type. The correct fix would be to rip bidiNext into
multiple functions and rip need for observation out of InlineBidiResolver.
Unfortunately I've tried both in separate bugs and failed. This code is very very
old and very poorly understood. We're slowly moving forward, this is another tiny step.
This is my fourth iteration of this patch (I'm happy to do more!), but I believe
it's a good compromise between fixing all of the design gotcha's of our bidi
system and doing the minimum amount to add this killer CSS feature.
I ran the PLT. (It averaged 0.2% faster with this change, but I attribute that to noise).
Test: css3/unicode-bidi-isolate-basic.html and css3/unicode-bidi-isolate-aharon.html
* platform/text/BidiResolver.h:
(WebCore::BidiCharacterRun::setNext):
- Needed by the new replaceRunWithRuns function.
(WebCore::BidiResolver::BidiResolver):
(WebCore::BidiResolver::~BidiResolver):
(WebCore::BidiResolver::enterIsolate):
(WebCore::BidiResolver::exitIsolate):
(WebCore::BidiResolver::inIsolate):
(WebCore::BidiResolver::isolatedRuns):
- Used to track isolated spans of text as they're encoutered.
They're stuffed away here to be processed recursively
after the main UBA has done its thang.
(WebCore::::appendRun):
(WebCore::::embed):
(WebCore::::commitExplicitEmbedding):
(WebCore::::createBidiRunsForLine):
* platform/text/BidiRunList.h:
(WebCore::::replaceRunWithRuns):
- This effectively takes all the runs from one runlist and adds them to
this one, replacing the fake run we inserted during a previous pass of the UBA.
- This RunList now owns the runs, so we call clear() on the other RunList
so that we don't end up double-freeing the runs.
(WebCore::::clear):
- This allows us to "take" runs from another run list and then clear it.
* rendering/BidiRun.h:
(WebCore::BidiRun::object):
* rendering/InlineIterator.h:
(WebCore::InlineIterator::object):
(WebCore::InlineIterator::offset):
(WebCore::notifyObserverEnteredObject): Mostly just renaming and adding a FIXME about plaintext.
(WebCore::notifyObserverWillExitObject): Mostly just renaming.
(WebCore::addPlaceholderRunForIsolatedInline):
(WebCore::isIsolatedInline):
(WebCore::InlineBidiResolver::appendRun):
* rendering/RenderBlockLineLayout.cpp:
(WebCore::statusWithDirection):
(WebCore::constructBidiRuns):
- This is the heavy-lifting of this change. This function
runs the UBA recursively on all the previously identified isolated spans.
- If we encounter more isolated spans in our run, we just add them to the
main list an keep going. Because the runs are linked lists and we have
direct pointers to our placeholder objects, we don't care what order
we process the placeholders in, so long as when we're done, they're all processed.
(WebCore::RenderBlock::layoutInlineChildren):
LayoutTests:
Two new tests for testing unicode-bidi: isolate behavior.
Note that the test from Aharon Lanin has one failing subtest
I've asked him if the test might have a typo in:
https://bugs.webkit.org/show_bug.cgi?id=50912#c30
* css3/unicode-bidi-isolate-aharon.html: Added.
- Some various unicode-bidi: isolate tests from Aharon.
* css3/unicode-bidi-isolate-basic.html: Added.
- This test tries all possible orderings of strong-LTR, strong-RTL and neutral characters
across unicode-bidi: isolate spans to make sure that we match expected rendering.
- A little red bleeds through the test, but that appears to be from anti-aliasing
and possible automatic font kerning, not layout failures.
* platform/mac/css3/unicode-bidi-isolate-aharon-expected.png: Added.
* platform/mac/css3/unicode-bidi-isolate-aharon-expected.txt: Added.
* platform/mac/css3/unicode-bidi-isolate-basic-expected.png: Added.
* platform/mac/css3/unicode-bidi-isolate-basic-expected.txt: Added.
git-svn-id: http://svn.webkit.org/repository/webkit/trunk@94775 268f45cc-cd09-0410-ab3c-d52691b4dbfc
13 files changed
