Gutter Crash [was Re: 21.2-b19 "Sneaky Bastard 19991102" crashes when using toolbar for Ediff control buffer]

[No it is not week-end here but I have 21.2 on our new linux box] ALTMARK(a)de.ibm.com writes: Ok, here is what happens. The ediff window is very small. Thus if you run with the buffer tabs, the buffer tabs take all the place of the window, so the buffers window isn't actually displayed. Only the modeline is. (Side issue: This window really shouldn't have tabs) (Side issue: Why is the size of the gutters not taken into account to set the frame size?) Then you move the mouse outside. The mouse code tries to find the nearest point in the buffer (pixel_to_glyph_translation), because the displayed parts of the buffer are not near, it thinks the last displayed line is closest and thus calls end_of_last_line (which is really start_end_of_last_line). These kinds of computation are really done in redisplay and thus start_end_of_last_line calls point_in_line_start_cache to update the cache that says where the display lines begin. As the window really isn't displayed the display structures are empty and thus point_in_line_start_cache does a quick minimal redisplay (without actually updating the screen), that is the call to update_line_start_cache on line 7254. As the cache is empty it hits this code. while (startp < old_lb || low_bound == -1) { int ic_elt; Bufpos new_startp; regenerate_window (w, startp, point, CMOTION_DISP); update_internal_cache_list (w, CMOTION_DISP); regenerate_window does that redisplaying.. However only the modeline is displayed. update_internal_cache_list then copies all info form the redisplay structures to a temporary working copy of the line start cache. However as there aren't any real buffer lines displayed there is no info to copy and the cache stays empty. That turns to be a known problem /* If this assert is triggered then regenerate_window failed to layout a single line. That is not supposed to be possible because we impose a minimum height on the buffer and override vertical clip when we are in here. */ /* #### Ah, but it is because the window may temporarily exist but not have any lines at all if the minibuffer is real big. Look into that situation better. */ if (!Dynarr_length (internal_cache)) { if (old_lb == -1 && low_bound == -1) { updating_line_start_cache = 0; w->line_cache_validation_override--; return; } and thus update_line_start_cache returns without doing anything and leaving the cache empty. Then point_in_line_start_cache returns "-1" because the value of window-start couldn't be find in the buffer. This then triggers the abort. The question is now how to solve it. Should we enforce the conditions that part of a window should always be displayed? Should we remove the abort and return a sensible guess? Should generate_window be fixed to always layout a line regardless of whether it fixed? Additionally, I see no gutter code in pixel_to_glyph_translation, is that correct? Andy? Jan