Table reflow can be made faster in at least 3 ways. The 1st is mostly
implemented. Although the 2nd seems doable, after some preliminary
testing on the block code, it may bring out additional bugs in the block
code.
1) Minimize the number of reflows which require rebalancing the columns
(and then usually reflowing every cell to fit). The pres shell will
notify frames which are ancestors of an incremental reflow command
target that a reflow command will be coming. Tables can keep track of
the number of pending reflows of descendants and not rebalance (and
reflow every cell) when there are still reflows coming. This is tables'
answer to reflow coalescing.
Acutally, it is a little more complicated than this to account for the
pres shell's ability to interrupt the reflow process. So the pres shell
will have an additional reflow command queue which frames can append
reflow commands to (indirectly) and this queue is guaranteed to be
processed before the pres shell gives up control. Tables request such
reflow commands when they decide to delay their rebalancing as above.
They can also cancel the commands if they are no longer needed.
2) When new row or cell frames get appended, reflow the new cells
initially at the current width of the appropriate table column(s) rather
than 1st with an unconstrained width and then 2nd with a real one. If
the table doesn't need to be rebalanced or if optimizaion #1 holds, then
the number of reflows on cells has been reduced. This makes the
assumption that during an initial reflow a block can set
max-element-size and maximum width even if it isn't being reflowed with
an unconstrained width. The assumption appears to be correct.
3) Remove all of the cases where each cell in the table is reflowed
again (after its initial reflows) to get max-element-size and maximum
width. Once this information is obtained it isn't necessary to get it
again unless an incremental reflow affects a particulay cell.
In addition, some optimizations have recently gone into the tree which
have improved table reflow performance including: (a) stopped
rebalancing nested tables just to get their maximum widths, (b) removed
an excessive number of colspan calculations when a colspan spans a
percent column, (b) cached the last available width on the row group to
avoid doing anything in degenerate yet common cases.
