09/04/2013 23:12, pdv:
I've adapted rowBreakPoint() and also removed the following call to rowWidth(), where the same calculations are done again. The on screen layout is now OK, except for a small remaining issue with RTL-languages.
Yes, we should not have several pieces of code that do the same calculations. What is the data structure that you use to remember the row characteristics?
I also notice that the inlineCompletion code is polluting the TextMetrics code, I wonder why the completion string is not just inserted in the paragraph and removed by the DEPM mechanism. I suspect this would be cleaner.
Unfortunately the extra time taken to calculate the correct widths in rowBreakPoint() eats too much of the time gained gained by avoiding single_character_painting and although still somewhat faster than before I find the scrolling speed again no longer acceptable. It looks indeed that this can only be solved by caching the widths.
Yes, probably. Remember also that a bit of profiling can help. Just reconfigure with --enable-build-type=prof and use Shark.app on LyX.
Finally, there is probably some parallel effort from Lin Wei (look for the "Word wrapping problem" thread) for breaking rows properly in chinese/english documents. I suspect that this work will become easier to get right after you have put all the word breaking code in one place.
JMarc
