* getText() could be called multiple times on a ParserCache object
which would fire the ParserOutputPostCacheTransform handler
multiple times.
But, I could not track down how this could happen right now.
* As a separate issue, while conceptually there are no restrictions
against calling getText() multiple times, there is a semantics and
performance issue if that did actually happen. getText() does a
bunch of transformations and makes no effort to avoid duplicate
work. It will accumulate more transformations over time via the
OutputTranform pipeline and it is preferable for getText() and/or
the OutputTransform pipeline to guarantee semantics where the
pipeline won't be run multiple times on the same content. That will
free both hook handlers (like this) and the transforms themselves
to avoid checks as in this patch.
This patch should be reverted once such a change is made to core.
Bug: T351461
Change-Id: If5dfa0954e3fd2b7dbea1ed29b475be07f0f3986
This was being overridden by a later rule that only affected reply
buttons on mobile.
Bug: T351542
Change-Id: Iad2f6b703827cccf46ad5909d11fe3912d7023d5
This was causing a misalignment on mobile pages that loaded the DT
styles but didn't have visual enhancements enabled.
Bug: T351044
Change-Id: Ib5e392943331467c3d29ee2b10d3bbe4995137d0
* This patch enables DT to work with Parsoid HTML without changing
the functionality for legacy HTML.
* The code comments document some of the decisions being made here.
Some of these decisions are temporary and need better solutions
but this patch will let us run visual diff tests and expose any
other latent bugs.
TODO
----
* We need to add new tests to verify CommentFormatter expectations
for Parsoid HTML. I'll tackle this in a followup patch.
Known issues:
-------------
* Performance: Since the getText() transformed output in ParserOutput
is not cached, if DiscussionTools is to switch over to Parsoid HTML,
we have to add some form of caching of the transformed output because
transformHtml can take a couple seconds in the p99 case which is too
long to render uncached!
* Longer-term: Since this hook is called when getText() is
called, all calls to getText() will now invoke this handler
(which will return but still has to do a bunch of checks to
determine this won't apply). Presumably, transformHtml() is
idempotent because when some other code (other extensions, for ex)
calls getText(), we will run the transfromHtml() on previously
transformed content.
My understanding is that getText() is going the way of the dodo
and that getText() callers will have to explicit call the output
transform pipeline code (and presumably this issue of repeatedly
calling the same transforms on previously transformed content will
be addressed there).
* Some CSS doesn't apply to Parsoid HTML because intervening <section>
tags interfere with existing query selectors -- will be addressed
separately.
Bug: T341010
Change-Id: I9846193656cdc658f5237df0a133d9d4dcc20d00
This requires something like an invalid timestamp, to produce a reply
link whose associated comment ID is going to not be findable.
Bug: T350633
Change-Id: Ib50c11096b9af9961b74309b60524a4b986e04aa
According to EventLogging 'editattemptstep' data the error hasn't
happened in the last 90 days, and according to Logstash data the
warning hasn't happened in the last 90 days either.
The problem this guards against shouldn't be possible in a world
without RESTBase, but keep the checks as assertions just in case.
Change-Id: Id7eaf14294f8a7bb877f6a0e00a90976e560fc54
While this method is not a huge bottleneck in this codebase it still
sticks out because it calls end() and array_pop() literally millions
of times. (Tested by running the unit test suite, which currently
takes about 30 seconds on my machine.)
Because of the way the method is used in this codebase (see especially
ImmutableRange::computePosition) $a is almost always a sub-element of
$b, or the other way around. It's almost never necessary to go all the
way up to the root element. We can use this additional knowledge and
stop much, much earlier. The extra code is worth it because we know it
will succeed very, very often.
When I measure the runtime of this method alone it goes down to less
than 1% of the previous runtime. The final loop at the end of the
method is almost never executed now (about 30 times in 15,000).
I also micro-benchmarked the final loop and optimized it to work with
passive array-indexes instead of actively manipulating the array with
array_pop().
Change-Id: Iffcaa8848780a85fde38e322649050c687567f29
It does the same as before.
I think performance is not a concern here, and wasn't my motivation
either. But I hope this makes the code easier to read and to reason
with.
I added a pure unit test case (without involving an actual Language
object) to cover the previously uncovered digits feature.
Change-Id: I6a0fc86035817eabb42b55e58183ae094c052aa6
I was curious why running the CommentParserTest takes so long. I
found this is one of the bottlenecks because it's called so often,
but many link titles that are parsed as user names turn out to be
something else. This little hack speeds up the test by 15% and has
probably a similar impact in production scenarios.
Change-Id: I5a0b3a49ba5793c8a345baaa7118fed500c082b6
I was curious why running some of the PHPUnit tests in this code base
takes so long. While I could not spot an obvious bottleneck I found
a lot of code that is extremely hot, e.g. called a hundred thousand
times. A few obvious optimizations are possible in this code, e.g.
not calling the surprisingly expensive DOMCompat::getClassList
multiple times.
Change-Id: If22bbc1aedd2c36db1ab2343de5737009050b7bb
