On Thu, 12 Nov 2026 16:00:22 GMT, Phil Race <[email protected]> wrote:
> Whilst this fix is probably fine, I'm not at all sure it relates to the bug > that it is being fixed under. Both cases should be allocating native memory, > not Java heap memory, and the memory usage would be similar. Perhaps FFM has > reclamation delays. And do you actually know that the reporters of this were > using massively long strings ? I mean the whole of War and Peace is only > about 3 million letters, and no one lays that out at once. I hope. We don't know the reporters text, but you're right that a single shaping run should be of a gigantic size to lead to a minute long freeze. There's a regression in JBR 25 that exacerbated this issue by causing newlines to split text into separate runs, so there's now another multiplier - number of runs and that causes a much faster growth. I think quite a normal (but long) log file may cause this. That regression is to be fixed within JBR, but the problem on OpenJDK side while rare to occur is still a real one, I think > src/java.desktop/share/classes/sun/font/SunLayoutEngine.java line 72: > >> 70: // See hb_buffer_t::CONTEXT_LENGTH in hb-buffer.hh >> 71: // test/jdk/java/awt/font/TextLayout/ShapingContextTest.java checks >> the value in case of a future change. >> 72: private static final int HB_CONTEXT_LENGTH = 10; > > FWIW this works as long as you are testing your bundled harfbuzz, but doesn't > guarantee what will be right on some end-user system with a different > harfbuzz. > So could we not be more conservative, rather than being exact ? > There seems to be minimal (no) benefit until the text gets very long anyway .. True, I've measured a bit different numbers here, 1024 seems to be big enough but just as fast on realistic input (up to 500k) ------------- PR Comment: https://git.openjdk.org/jdk/pull/33853#issuecomment-6917742247 PR Review Comment: https://git.openjdk.org/jdk/pull/33853#discussion_r4586322295
