This doesn't switch all of the users yet, but is a step in the right
direction.
--HG--
extra : rebase_source : 91b4fef3f67586179c119208d000cf7629e04963
This patch enables sharing of an nsAttrValue's MiscContainer between nodes for style rules. MiscContainers of type eCSSStyleRule are now refcounted (with some clever struct packing to ensure that the amount of memory allocated for MiscContainer remains unchanged on 32 and 64 bit). This infrastructure can be used to share most MiscContainer types in the future if we find advantages to sharing other types than just eCSSStyleRuley. A cache mapping strings to MiscContainers has been added to nsHTMLCSSStyleSheet. MiscContainers can be shared between nsAttrValues when one nsAttrValue is SetTo another nsAttrValue or when there is a cache hit in this cache. This patch also adds the ability to tell a style rule that it belongs to an nsHTMLCSSStyleSheet, with appropriate accessor functions to separate that from the existing case of belonging to an nsCSSStyleSheet.
The primary use case is to reduce memory use for pages that have lots of inline style attributes with the same value. This can happen easily with large pages that are automatically generated. An (admittedly pathological) testcase in Bug 686975 sees over 250 MB of memory savings with this change. Reusing the same MiscContainer for multiple nodes saves the overhead of maintaining separate copies of the string containing the serialized value of the style attribute and of creating separate style rules for each node. Eliminating duplicate style rules enables further savings in layout through style context sharing. The testcase sees the amount of memory used by style contexts go from over 250 MB to 10 KB.
Because the cache is based on the text value of the style attribute, it will not handle attributes that have different text values but are parsed into identical style rules. We also do not attempt to share MiscContainers when the node's base URI differs from the document URI. The effect of these limitations is expected to be low.
Previously we snapped the results of nsDisplayItem::GetBounds and
nsDisplayItem::GetOpaqueRegion internally. By tracking which display items were
inside transforms, we disabled snapping quite conservatively whenever an ancestor
had a transform, which is undesirable.
With this patch, we don't snap inside GetBounds or GetOpaqueRegion, but just return
a boolean flag indicating whether the item will draw with snapping or not. This flag
is conservative so that "true" means we will snap (if the graphics context has a transform
that allows snapping), but "false" means we might or might not snap (so it's always safe
to return false).
FrameLayerBuilder takes over responsibility for snapping item bounds. When it converts
display item bounds to layer pixel coordinates, it checks the snap flag returned from
the display item and checks whether the transform when we draw into the layer will be
a known scale (the ContainerParameters scale factors) plus integer translation. If both
are true, we snap the item bounds when converting to layer pixel coordinates. With
this approach, we can snap item bounds even when the items have ancestors with active
transforms.
This is the third of three patches to rework the way we handle getting
the font inflation container and width data during reflow, which are
needed so that we can sometimes honor inflation during intrinsic width
calculation (which we need to do to make some form controls inflate
correctly).
