Right now frame visibility doesn't take -moz-element into account, but it's not that hard. We can do so by having nsSVGIDRenderingObserver mark the subtree of the primary frame of the element it's observer visible. (Until it stops observing that element, in which case it can decrement the visible counts of the frames in the subtree.) Right now the frame walking is extremely simple and doesn't cross into subdocuments or try to do anything clever about scrollframes. We can extend this further in followups; I just want to get the basic, common cases working first.

Here's the patch.

OK, this needed to be reworked, because the frames associated with a content node are not necessarily available at the time that nsSVGIDRenderingObserver starts observing it. To make this work, I've instead added a list of content nodes to the pres shell. If a content node is in the list, every frame in its subtree is considered visible. We can then rely on the normal frame visibility update mechanism to the keep the frame visibility list correct. What's really nice is that this eliminates the limitations of the previous approach, because we're reusing the existing frame walker. For example, we now traverse subdocuments correctly.

This patch actually implements visibility tracking that "sees through" nsSVGIDRenderingObserver, and therefore through -moz-element. The idea is the same as in the previous version of the patch, but now we make use of Increment/DecrementVisibleCountOfContentSubtree().

OK, so investigation of the failing test above made me realize that there is a simpler and better way to implement this. Rather than forcing all frames in a subtree to be visible, just call MarkFramesInSubtreeVisible() and pass the visual overflow rect of the primary frame of the content we care about. The old version of the patch was buggy, and this should fix that, but even better, it means that for e.g. scrollframes we will only mark visible the frames that are actually visible through the scrollframe, and not every single frame that may be scrolled way out of view. It also makes the patch even shorter and simpler, which is nice. =)

Alright, moving the list of content nodes that we're forcing to be visible to the document turned out to be the trick. The pres shell doesn't have the right lifetime; the same document can have many pres shells over time. The try job above is green. I'll upload the final version of the patches now.

This updated version of part one, as implied by the comment above, has one major change: the list of content nodes that we're forcing to be considered visible has moved to nsIDocument/nsDocument. That gives the list the same lifetime as the content nodes themselves (unless the content is removed from the document, of course, but we handle that in part 2 in the ElementChanged() callback). With this change, these patches are now green on try. I've made a couple of cosmetic changes as well, adding more comments and renaming the methods to (hopefully) clarify what they do a bit more.

This version of part 2 is basically just rebased against the updated version of part 1; the idea remains the same.