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
The addon does support some configuration options. These are passed
as properties of the `matchBrackets` CodeMirror option. Just passing
the boolean there hides that fact.
Bug: T270317
Bug: T270237
Change-Id: Iaa4b5ed8ef538e76cd1c96a09485e143112f1ae0
This patch also removes all remaining FIXMEs. This code is not
a bottleneck. There is nothing to do.
Bug: T270317
Change-Id: Ie034440c98d8064a22811a1b569237dddb7b7436
We "forgot" to update this addon when we did the update in
I6f0f030 (T258999).
Bug: T258999
Bug: T270317
Change-Id: Iab29e9e36f34b76551ddac497e40dc76669ba7c7
Basic tests to show that the highlighting classes have been attached
to the expected elements.
TODO in later patches:
* tests for the wikitext 2017 editor
Bug: T270240
Change-Id: I01ebd9881d38dd877f19ee3bb4fdcbb74d43afaf
Optimizations for the code introduced in Ic403e0a:
* Skip this entirely when something is selected (as discussed
in Ic403e0a).
* Use a combination of existing methods. I benchmarked these
again. This approach is "significantly" slower compared to
the custom code from before. However, "significantly" here
means something like 1 nanosecond vs. 4 nanoseconds. Both
is effectively nothing.
* Use the same approach in another place. This one is triggered
every time a change is made, e.g. a character typed. I
benchmarked this as well. The new code is about 500x faster
(yes, seriously).
Bug: T269094
Change-Id: I00fe595a89be7a257e27ed28d38568c81483338b
This is a convenient way to propagate alternate configuration during
browser testing.
Bug: T270240
Change-Id: Ica6399c53499be7f930e8d13b838ad265b66cdf4
The fallback technique makes the whole edit surface semi-transparent,
so reset native selections to full opacity.
Change-Id: If6cd585b1a09c549781fe82a3bdf18d64ac597b5
* using CodeMirror addon matchBrackets
* highlights the matching bracket of a pair
* highlights brackets when cursor is inside a pair
* feature usable in source code editor
Bug: T261857
Change-Id: Ib01d9919a47bb29684b54501644b01936b57972a
I had to make some CSS selectors more specific, because the
library changed
.CodeMirror pre
to
.CodeMirror pre.CodeMirror-line,
.CodeMirror pre.CodeMirror-line-like
This is only relevant for entire lines (implemented as <pre>
elements). Most of the custom CSS is for characters, not lines.
In my tests in the Wikitext editor as well as VisualEditor I
could not spot any difference between the old and new version.
Bug: T258999
Change-Id: I6f0f030f972838727f3ef220feb105264f122798
