In T270880 an example with a slash in <ref name="a/b"> is
described. The same issue happens with several other characters
including the closing bracket, e.g. <ref name="a>b">. This patch
fixes all of this by accepting _all_ characters between double
and single quotes.
Bug: T270880
Change-Id: I03a1e1a25af692dc703b44a57b2d23d6fc15c8c9
Introduces a new config variable `CodeMirrorLineNumberingNamespaces`
that can restrict line numbering to only appear for specified
namespaces. Setting to null enables everywhere.
This takes some liberties with the `lib` module, turning it into a
container for shared functionality. This can be pursued in later
work, by cleaning up duplicated code in this repo.
FIXME: failed to deduplicate the code for now.
Bug: T267911
Change-Id: Ida2b33eef38edc57d29756ec472c6f2c83bd7b11
The issue can be reproduced as described in T278840. What
happens is that an (auto)clear is triggered and removes all
marks, but the cached values in `currentMarks` remain. The next
time the same marks are found, they are discarded and don't
show up, because the cache says they are already there, when
they are not.
Bug: T278840
Change-Id: If83bd99e924f579854cfe4b01fab4ef86892933b
Adds a custom class for matched brackets to allow better integration
with custom bracket styles. The brackets won't be bold in the 2017WTE.
Bold font might lead to misalignment there. See ticket.
Note: box-shadow seems to be supported for quite some time by all
relevant browsers
Bug: T270926
Change-Id: Ica1e301f63a106a96db3bfaba4b2f322af64b009
These changes to the color scheme are hidden behind a feature
flag for the time being.
Bug: T271895
Change-Id: I0a4b03e0f3bc8239f31edbbd5ae55661607b76f6
This does not have an effect any more with all the other
optimizations in place.
This reverts commit 094f20902c.
Bug: T274369
Change-Id: I288039a35270093bd22b5a073e70f6b769088c13
I was wondering why the performance when editing wikitext is
still so bad, and profiled it again. Turns out
StringStream.match() is still the bottleneck (even if already
100 times better than before Icbb1122).
The method is called with many different patterns from
mode/mediawiki/mediawiki.js. I profiled them individually and
found a single outlier. The idea is the same as in Icbb1122.
A pattern that is able to find something *in* a string is
doing nothing but wasting time, as StringStream.match() ignores
every result that is not at the start of the string.
The change adds the missing ^ anchor and wraps the regex pattern
from mw.config.get( 'wgUrlProtocols' ) (that is something like
"ftp:\/\/|http:\/\/|https:\/\/|…") in (?:…), which is a
non-matching group. This is necessary because of the | in the
pattern. The result is a pattern that looks like /^(?:…|…|…)/i.
I remember looking at this code while working on Icbb1122, but
didn't include it in the patch, and then forgot about it.
Bug: T270237
Bug: T270317
Change-Id: Iea2fd116b68704c3186b0edf965006cc7c6eda82
My previous patch Icbb1122 focused on the behavior of the
matchbrackets addon when the text is *edited*. This patch here
is about moving the cursor without changing the text. I
realized the addon re-draws everything every time the cursor
moves, even if the highlighted pair of brackets is still the
same. This triggers very expensive code in the CodeMirror lib.
I had a look at this expensive code, but did not found an easy
win. It just is what it is: an expensive re-draw. Instead I
introduced a caching layer that remembers the positions of the
previously highlighted brackets and bails out as early as
possible when nothing changed.
The biggest chunk of code is that "did something change?"
comparison. It looks expensive, but typically isn't. There are
typically only 2 elements in the array for a single
opening–closing pair. (Possibly more when there are multiple
text selections.) The elements in the two arrays are typically
in the same order. (Except the cursor is on the closing
bracket.) Which means the nested `every` → `for` loop will
typically be executed 2 times only – one time for each of the
2 elements.
I won't upload this change upstream because it is only relevant
together with our custom "in the middle" bracket highlighting.
With our customization we have many, many situations where the
highlighted brackets don't change. This (almost) doesn't happen
upstream.
Bug: T270317
Change-Id: I789b45362388f0818e797f789f6af427a35e3e06
While working on T270317, I realized the performance of the
matchbrackets addon is not as good as described in
T270237#6739993. The issue with my original benchmark was that
I did it with a single pair of brackets with thousands of
characters between. A paragraph with thousands of brackets
behaves much worse. So bad that I feels painful when moving
the cursor.
Lowering the limit to something in the middle (between the
original 1000 and my 10000) makes it behave much, much better
on my machine.
Bug: T270237
Bug: T270317
Change-Id: I31f850f4c7778d6b5ff1d0eb17fdaf0edf7ae019
My upstream patch was accepted within 9 minutes:
https://github.com/codemirror/CodeMirror/pull/6565
Note: This backport includes another upstream commit that fixed
some typos.
Bug: T269096
Change-Id: Ib5b64214d7536bc952886f45290d537eab2f9bbb
The addon does have 3 settings:
- maxHighlightLineLength is for the current line where the
cursor is. Bracket matching is simply not done when the
current line is longer. The default is 1000, which is rather
low.
- maxScanLineLength is for every other line that is scanned in
the process. I don't understand why, but this limit is 10x
higher.
- maxScanLines is the number of lines that can be scanned.
Simply raising the first to be 10000 as well fixes our issue.
Note that CodeMirror does have a limit of 10000 anyway. It's
called maxHighlightLength there. Lines that are longer get
syntax highlighting only for the first 10000 characters. The
rest of the line is black. Using the same limit in the addon
makes it's behavior consistent. Means: The user can see when
the syntax highlighting stops, and bracket matching stops
working the same time.
I benchmarked with both settings. It doesn't have a measurable
effect. Bracket matching is done in <1ms in both cases.
Bug: T270237
Change-Id: Ia56bf4c2fb023c9f117376242221d39f51196173
Less ambiguous variable names. Less duplication. The minor
issues have been introduced in T270317.
Bug: T270317
Change-Id: I366396a61b76e19293ce8d14c2f346b97498fe40
I continued profiling the matchbrackets addon for T270237 and run into
performance issues that turned out to be unrelated to the addon. The
flame graph highlighted a "match" function. Note this is not the
String.match() from JavaScript, but something in the CodeMirror lib:
StringStream.prototype.match = function (pattern) {
var match = this.string.match(pattern);
if (match && match.index > 0) {
return null;
}
return match;
}
(Note: I simplified this code so we can focus on the bug.)
When the pattern is a regular expression, it's executed via
JavaScript's String.match(). The function then checks if there was a
match and if it's at the start of the string. If not, it's not a
match and doesn't return one.
In other words: Even if there is a match somewhere in the string, the
function acts as if there was no match.
When we change all patterns to be anchored via ^, they don't scan the
entire string any more but return much ealier when there is no match
at the start of the string. We are effectively replacing nested loops
(hidden in the patterns) with single calls.
This bug exists since 2014.
The disabled line in the matchbrackets addon is just dead code. I
don't remove it to document the fact that we disabled it.
Bug: T270237
Bug: T270317
Change-Id: Icbb1122e6a3b26c0606726ff405e108931d185be
