This way we'll be able to take different paths for the sequential and parallel
traversals in some concrete cases.
This is a preliminar patch to fix bug 1332525.
It turns out that it's problematic to embed ThreadLocalStyleContext within
LayoutContext, because parameterizing the former on TElement (which we do
in the next patch) infects all the traversal stuff with the trait parameters,
which we don't really want.
In general, it probably makes sense to use separate scoped TLS types for
the separate DOM and Flow tree passes, so we can add a different ScopedTLS
type for the Flow pass if we ever need it.
We also reorder the |scope| and |shared| parameters in parallel.rs, because
it aligns more with the order in style/parallel.rs. I did this when I was
adding a TLS parameter to all these functions, which I realized we don't need
for now.
This allows us to get rid of a bunch of lifetimes and simplify a lot of code. It
also lets us get rid of that nasty lifetime transmute, which is awesome.
The situation with thread-local contexts is still suboptimal, but we fix that in
subsequent patches.
I noticed that our current behavior in ContentRangeInserted is incorrect. Unlike
ContentInserted (where this code lived originally), ContentRangeInserted takes a
start and end element. I'm not sure if we ever take that path for new content that
needs style, but it seemed sketchy. And generally, it seems nice to just always
style new content the same way (though we still need to style NAC by the subtree
root, since it hasn't been attached to the parent yet).
For situations where there is indeed only one unstyled child, the traversal
overhead should be neglible, since we special-case the single-element in
parallel.rs to avoid calling into rayon.
Being more explicit about what we want here also makes us more robust against
the other handful of callpaths that can take us into
nsCSSFrameConstructor::{ContentRangeInserted,ContentAppended}. Currently we
can call StyleNewSubtree on an already-styled element via RecreateFramesForContent,
which triggers an assertion in the servo traversal.
MozReview-Commit-ID: DqCGh90deHH
The primary idea of this patch is to ditch the rigid enum of Previous/Current
styles, and replace it with a series of indicators for the various types of
work that needs to be performed (expanding snapshots, rematching, recascading,
and damage processing). This loses us a little bit of sanity checking (since
the up-to-date-ness of our style is no longer baked into the type system), but
gives us a lot more flexibility that we'll need going forward (especially when
we separate matching from cascading). We also eliminate get_styling_mode in
favor of a method on the traversal.
This patch does a few other things as ridealongs:
* Temporarily eliminates the handling for transfering ownership of styles to the
frame. We'll need this again at some point, but for now it's causing too much
complexity for a half-implemented feature.
* Ditches TRestyleDamage, which is no longer necessary post-crate-merge, and is
a constant source of compilation failures from either needing to be imported
or being unnecessarily imported (which varies between gecko and servo).
* Expands Snapshots for the traversal root, which was missing before.
* Fixes up the skip_root stuff to avoid visiting the skipped root.
* Unifies parallel traversal and avoids spawning for a single work item.
* Adds an explicit pre_traverse step do any pre-processing and determine whether
we need to traverse at all.
MozReview-Commit-ID: IKhLAkAigXE
This patch introduces infrastructure for the rule tree, and constructs it.
We don't use it yet, nor have good heuristics for GC'ing it, but this should not
happen anymore once we store the rule node reference in the node.
I haven't messed up with memory orders because I want to do a try run with it,
then mess with them.
Take down the ApplicableDeclarationsCache, use the rule tree for doing the cascade.
We don't need this for Gecko, and it's hard to implement in that case because
there's nowhere obvious to put it (we don't plan to create TSDs for non-dirty
nodes, and non-dirty nodes can have dirty children which require the
children_to_process atomic). There are various solutions here, but punting is
the easiest.
We'll need to rethink this if/when we need to do a bottom-up traversal for
Gecko.
The style candidate cache had regressed a few times (see #12534), and my
intuition is that being able to disable all style sharing with a single rule in
the page is really unfortunate.
This commit redesigns the style sharing cache in order to be a optimistic cache,
but then reject candidates if they match different sibling-affecting selectors
in the page, for example.
So far the numbers have improved, but not so much as I'd wanted (~10%/20% of
non-incremental restyling time in general). The current implementation is really
dumb though (we recompute and re-match a lot of stuff), so we should be able to
optimise it quite a bit.
I have different ideas for improving it (that may or may not work), apart of the
low-hanging fruit like don't re-matching candidates all the time but I have to
measure the real impact.
Also, I need to verify it against try.
This commit adds hooks to the Servo style traversal to avoid traversing all the
DOM for every restyle. Additionally it changes the behavior of the dirty flag to
be propagated top down, to prevent extra overhead when an element is dirtied.
This commit doesn't aim to change the behavior on Servo just yet, since Servo
might rely on a full bottom up reconstruction of the flows. I'll need to double
check and implement that separately.
As a follow-up, we could move all the data living under a mutex in the
SharedLayoutContext only in order to create the local context to the same place.
This should increase animation performance when there are multiple animations in
one page that happen to be on different threads.
There's a bit of flickering when unpausing where the node has the original
state, but I'm not totally sure where it comes from, posibly from
PropertyAnimation returning None due to no styles changing?
