Volker, 0. I’d like to see what the crashing stack frame is when NOT on harfbuzz, because there should be no change.
1. Phil addressed this one 2. I will take a look on AIX. I’ll see if I can build Harfbuzz itself on AIX at first. 3. Yes, I’ll work on putting some docs together for that test. Actually, other tests such as the Font2DTest would be a good test. Or, any text in a complex script such as Hindi. 4. I’m not sure. A threaded stress test would be good. Perhaps Harfbuzz-in-JDK could leverage the hotspot/src/os_cpu/../vm/atomic_os_cp.inline.hpp you mentioned. Behdad - file is here http://hg.openjdk.java.net/jdk9/client/hotspot/file/c8e212fb27d0/src/os_cpu/linux_x86/vm/atomic_linux_x86.inline.hpp - any comments on this or other items? -s > > -------- Original Message -------- > Subject: Re: [OpenJDK 2D-Dev] Question regarding the integration of > harfbuzz (JEP 258) > Date: Wed, 09 Dec 2015 09:02:33 -0800 > From: Philip Race <philip.r...@oracle.com> <mailto:philip.r...@oracle.com> > Organization: Oracle Corporation > To: 2d-dev@openjdk.java.net <mailto:2d-dev@openjdk.java.net> > > Hi Volker, > > Running with ICU should stop harfbuzz use completely so if > you are still seeing a crash, perhaps it is due to something else entirely. > Especially since a "simple AWT program" should not load layout anyway. > There was a fair amount of new and changed code pushed recently ... > > Yes, we need to run harfbuzz in multiple concurrent threads in the JDK. > A single threaded lock on harfbuzz would throttle text operations. > I suppose it is possible that we do not strictly need it as we currently > create a > new harfbuzz "instance" each time we go to invoke layout but that > is something that at least theoretically could change. > > The approach to synchronization is all from up-stream harfbuzz. > The harfbuzz mailing list might be the place to discuss that. > I'd expect it does work on AIX as a principal driver here is that > the IBM sponsored ICU project has deprecated its layout engine > and IBM/ICU are moving to harfbuzz so I'd be somewhat astonished > if they haven't yet tried it on AIX. Did you try just getting a copy > of the full hotspot library and running its configure on AIX to see if > harfbuzz can work out which options it wants there ? Obviously > I did not try that since I don't have AIX. Steven Loomis of IBM > who submitted the JEP should be able to help with a lot of this. > > -phil. > > On 12/8/15, 10:50 AM, Volker Simonis wrote: > > Hi, > > > > the integration of harfbuzz broke our AIX build because there's no > > implementation available for the hb_atomic macros in > > hb-atomic-private.hh. I've fixed that locally but still get a crash > > when running a simple AWT program (even with > > -Dsun.font.layoutengine=icu). > > > > I'm curretnly debugging the problem, but I have some general questions: > > > > 1. Do we really need the multi-threaded features of harfbuzz in OpenJDK? > > > > 2. Why does hurbuzz require these synchronization macros? From a first > > glance it doesn't seem to use multiple threads by itself. Does it have > > some global data structures which it needs to protect if harfbuzz is > > called from various threads? > > > > 3. Is there a way to test the hurfbuzz engine on a new platform? I saw > > the new test TestLayoutVsICU.java but it is marked as "manual" and it > > seems to require external fonts (at least some of which seem to be > > freely available). Could you please provide at least a minimal > > description on how this test can be run? > > > > 4. Is there a way to test that the implementation of the hb_atomic > > macros in hb-atomic-private.hh is correct (i.e. a way to stress > > harfbuzz from multiple threads)? I think this is important as I > > couldn't find a good description of the requirements for the atomic > > macros and on weak memory model architectures we should really get > > this right. For example there's no description of which fences are > > required around the various atomic operations. Getting this right > > across all platforms in the HotSpot (see > > hotspot/src/os_cpu/../vm/atomic_os_cp.inline.hpp) is already hard > > enough so I'm a little concerned that we are now introducing similar > > primitives in the class library. > > > > Thank you and best regards, > > Volker
