Hi, On Fri, 2018-05-11 at 12:55 +0900, Yasumasa Suenaga wrote: > Hi, > > I found 1 GCC optimization issue, but it is not enough. > > http://cr.openjdk.java.net/~ysuenaga/openjdk-fc28/gcc8-opt/gdb-before.txt > > I traced class loading from `modules`. Above log is copy of GDB > console in ImageFileReader::verify_location(). > The code expects `*next++` increments after referring the value, but > it do not so. > > Thus I fixed the code as following changeset: > http://cr.openjdk.java.net/~ysuenaga/openjdk-fc28/gcc8-opt/wip.00/ > > It passes module name check, but it fails parent (package name) check. > GDB log is here: > http://cr.openjdk.java.net/~ysuenaga/openjdk-fc28/gcc8-opt/gdb-after.txt > > I guess the value of parent should be `java/lang`, but it sets to > `sun/refrect/generics/factory`. > Do you have anything to think about?
The problematic library was indeed libjimage.so. More below. > 2018-05-09 17:22 GMT+09:00 Severin Gehwolf <sgehw...@redhat.com>: > > Hi, > > > > Note that slowdebug builds work: > > > > $ ./build/linux-x86_64-normal-server-slowdebug/images/jdk/bin/java -version > > openjdk version "11-internal" 2018-09-25 > > OpenJDK Runtime Environment (slowdebug build > > 11-internal+0-adhoc.sgehwolf.openjdk-hs) > > OpenJDK 64-Bit Server VM (slowdebug build > > 11-internal+0-adhoc.sgehwolf.openjdk-hs, mixed mode) > > > > In summary we have: > > > > * Build fails with fastdebug/release debug config (-O3) > > * Build succeeds with slowdebug (-O0) > > * F28 has GCC 8 > > * Older GCC-based builds continue to work for fastdebug/release config > > > > * JDK 10.0.1 builds fine with GCC 8. > > > > So far showing all symptoms of either a GCC bug or some UB in recent > > OpenJDK code which breaks with new optimizations done in GCC 8. > > > > I'll continue to investigate what it is... Looks like another case of UB in the JDK. This time libjimage.so with a signed integer overflow: https://bugs.openjdk.java.net/browse/JDK-8203223 Thanks, Severin
