Hi Thibault,
thanks the detailed analysis. :)
I do not know if it's worth to create a jdk bug, if we can not reproduce
the bug reliably and can not reduce the set.
-Pascal
Am 06.09.2015 um 16:29 schrieb Thibault Kruse:
I believe this kind of error normally only happens when java in
invoking native C/C++ code.
I could reproduce locally on java-9-oracle 9.0-b78 with either:
./gradlew :groovy-groovysh:test -Pindy=true.
Also I could sometimes reproduce like this:
./gradlew :groovy-console:test -Pindy=true
Running individual test classes only does not seem to reproduce the
error. Reducing the number of tested classes reduces the likelyhood of
the filure occuring (see https://en.wikipedia.org/wiki/Heisenbug).
Even worse, re-running the command does not guarantee to reproduce the
failure, in particular the console tests were fickle in that respect.
I tried to remove some test classes to get a smaller set, but at some
point the failure would not reoccur, even when restoring all test
classes and running a clean build.
That points to something like a race condition, or GC-related, or some
other Voodoo (like http://openjdk.java.net/jeps/197).
I could reproduce going back to
9148f794c168b531d0 2015-03-15 09:25:12 Merge branch 'GROOVY_2_4_X'
(by backporting a jdk9 fix from 83d680877c440)
So this is not related to any recent commit. Also this means I have
checked many gradle versions, and it's not related to any recent
change of gradle versions.
So I believe this looks like either a bug in java9 or an
incompatibility with Java9 of groovy or any library used during
testing.
I think it would be worth reporting this at http://bugreport.java.com/.
On Sun, Sep 6, 2015 at 10:14 AM, Thibault Kruse
<tibokr...@googlemail.com> wrote:
It fails for both a console and a groovysh test. There is some output
for these tests earlier in the report saying:
[:groovy-groovysh:test] *** Error in
`/opt/jdk9-build/j2sdk-image/bin/java': double free or corruption
(fasttop): 0x00007fe038b0d300 ***
It might be good to bisect this to the change causing this if possible.
On Sun, Sep 6, 2015 at 9:16 AM, Pascal Schumacher
<pascalschumac...@gmx.net> wrote:
GroovyShell Test all work now :), but now gradle exits with an non-zero exit
code when running groovy-console tests with indy:
http://ci.groovy-lang.org/viewLog.html?buildId=26577&buildTypeId=Groovy_Jdk9Build&tab=buildLog#_focus=85490&state=85490
Anybody has an idea why?
-Pascal
Am 01.09.2015 um 19:19 schrieb Pascal Schumacher:
Merged. Thanks a lot. :)
- Pascal
Am 01.09.2015 um 12:36 schrieb Thibault Kruse:
See https://github.com/apache/incubator-groovy/pull/107 for a possible
fix
On Tue, Sep 1, 2015 at 11:13 AM, Thibault Kruse
<tibokr...@googlemail.com> wrote:
The tests makes sure that completing java.util. includes 'Set', as in
java.util.Set. I'll assume java9 does not move the Set class. I guess
there is a change affecting method
PackageHelperImpl.getClassnames(), which was originally copied from
jline1.
I see it has adaptations for jigsaw made by Cedric:
https://github.com/apache/incubator-groovy/commit/0e384ec3
(not pointing fingers, just analyzing the code)
And the breaking test follows that new codepath.
Debugging a bit, I notice that the files contained in 'jrt:/' does not
contains only "/modules/java.base/java/util/Set.class". At a glance,
it seems the assumptions in
PackageHelperImpl.getPackagesAndClassesFromJigsaw() do not hold with
the current Java9 implementation, since it produces a Class
java.base.java.util.Set, but not java.util.Set.
A quick fix is to change:
-if (elems) {
- elems = elems[3..<elems.length]
+if (elems && elems.length > 2) {
+ elems = elems[3..<elems.length]
in *two* places in that method.
I have no idea whether there is a standard for the folder layout
FileSystems.newFileSystem(URI.create("jrt:/")) should return, or
whether and why this has changed since Cedric made his additions.
However this might be related:
http://mail.openjdk.java.net/pipermail/jigsaw-dev/2014-November/004044.html
