get_filetype_from_line runs modeline_re.search() on the first and last lines of the buffer, and because the pattern is a modeline start followed by .* and an ft=/syn= assignment, a line with many starts and no assignment restarts the scan at every start, which is quadratic in the line length. guess_lexer() calls this before anything else, so guessing the language of a single 96 KB line of 'vi:' takes about 51 s (48 KB: 10 s); with this change it takes 5 ms, and 1 MB takes 50 ms. The .* can skip over any later start, so only the leftmost one can decide the result: find it with the prefix alone and run the existing pattern once with match() there, which returns the same filetype as before (checked against the old regex on 300k random lines).