After removing :before and :after pseudo selectors, a class name may be empty, causing an invalid selector string when joined. Remove empty elements before calling querySelectorAll.
Previously, pseudo elements would be processed as they were found in the DOM tree, which was
an expensive operation as each element's computed :before and :after style was checked for
'content' styles.
This commit traverses the user's stylesheets for :before and :after selectors, gathers the classes
affected, selects all elements that likely have a pseudo element present, then checks computed style.
If there is actually an element present, it is created but *not* appended to the DOM until after
all elements have been processed.
After all elements have been found and created, they are added to the DOM in a single batch, and the original
pseudo elements are hidden in a single batch. This prevents the layout invalidation / relayout loop that was
occuring previously, and in my tests speeds parsing by as much as 50% or more, depending on how many
pseudo elements your page uses.
Additionally, this commit contains a bugfix to the handling of ":before" pseudo elements; the browser effectively
inserts them as the first child of the element, not before the element. This fixes a few rendering inconsistencies
and complicated pages look almost perfect in my tests.
In my testing, the major time sink is parsing. This commit adds a setTimeout() around parsing
of each item so control can return to the browser. This increases the total time it takes to finish
a screenshot but will not freeze the browser when it does. This is a good option when e.g. doing
error reporting, where you might not want to freeze the browser while sending debugging information
back to your server.
1. when stacking without z-index, positioned > floated > normal flow
2. supports negative z-index
3. new stacking context formed when opacity < 1 (standard)
4. new stacking context formed for position:fixed (no standard yet, done in mobile webkit and chrome 22+)
