Jan Vroonhof <vroonhof(a)math.ethz.ch&gt; writes: The highlighting is there, but isearch itself always happens in one window. If you try to change the window (or frame or buffer or anything), isearch is done. That's how it's been for ages. For some time I considered "correctly" handling this case by highlighting all visible regions in all windows, but I rejected the idea because it would be too much work for practically no gain. When the isearch is in progress, we tend to look at the window where it happens. When the isearch ends, so does the highlighting. It is; that's why window-end suits me and that's why I reported a bug. Yes, it would be nice, but it's hard to do right. Ideally, you'd want the code to be called after the redisplay has calculated what needs to be done, but before it is actually done. But your code could change the screen layout and thus invalidate the previous calculation. This would require a recalculation, leading to your code being reinvoked, etc. It's very easy to infloop there. You can either do what I do and change things on your own risk, or do what lazy shot does, and enjoy the flicker. For isearch, the two approaches amount to (almost) the same thing, and I believe mine is simpler. Oh, and note that even in the case of isearch the `do-on-demand' code might prove deficient. For the highlighting to be meaningful in all cases, it requires an anchoring point, normally provided by the location of point in isearch. If you just try to find matches beginning with a random point, your code will suffer from the "banana phenomenon"; see my article on comp.editors for details. But none of it is really relevant. Philosophical discussions are nice, but I only really wanted to draw attention to a bug in window-end. I do not understand the question. Huh? It is quite obvious that lack of redisplay causes window-start to return the wrong value. If my understanding is correct, the GUARANTEE argument was invented *exactly* for that purpose -- to get the correct value without forcing a redisplay. Bugs notwithstanding, it is very strange that window-end is provided a GUARANTEE argument, while window-start is not. That code is buried deep in redisplay. I don't think you can just "force it to be run" without also triggering redisplay. Or, do you mean that the redisplay code should be run, while the redisplay *output* should be suppressed?