Re: Fix build on x86-64

[Wu Zhou <woodzltc at cn dot ibm dot com>]

On Wed, 18 Jan 2006, Andrew Cagney wrote:

> Wu Zhou wrote:
> 
> > On Tue, 17 Jan 2006, Andrew Cagney wrote:
> > 
> >  
> > > Wu Zhou wrote:
> > > 
> > > Overall the results are pretty good, two notes though:
> > >    
> > 
> > Hi Anrew, what about "FAIL: textiter/a.out"?  Is it a serious problem?  I
> > remember that it ever pass for an earlier cvs.  Now it failed on both x86
> > and ppc64.
> >  
> That's interesting; I'll re-check my local results.  It shouldn't be a serious
> problem.
> The bug first hit some java-gnome JNI bindings code, fortunatly it was
> possible to rewrite the code to avoid the problem.  There's obviously an
> ongoing risk that fresh code will trip over the problem, the gui developers
> feel that the risk is relatively low.
> 

That is okay.  This test still fail on x86 and ppc64 though.

> > > > FAIL: cniinner/a.out
> > > >      
> > > not so good; I'm guessing that on ppc-64 it tickled some other bug.
> > > Perhaphs
> > > try that test on gcc-4.1?
> > >    
> > 
> > Is gcc-4.1 out already?  I tested on gcc-4.0.1, 4,0.2, Redhat 4.0.1 and your
> > 4.0.2 compilers.  All of them give the same error message:
> > 
> > [root@plinuxt2 tests]# make -f Makefile.401 cniinner/a.out
> > mkdir -p cniinner
> > /opt/gcc401/bin/gcj -I../../../frysk/frysk-imports/tests -d . -C
> > ../../../frysk/frysk-imports/tests/cniinner/Parent.java
> > /opt/gcc401/bin/gcj -I../../../frysk/frysk-imports/tests -d . -C
> > ../../../frysk/frysk-imports/tests/cniinner/Child.java
> > /opt/gcc401/bin/gcjh -I. cniinner.Parent
> > /opt/gcc401/bin/gcjh -I. cniinner.Child
> > /opt/gcc401/bin/gcjh -I. 'cniinner.Child$Nested'
> > /opt/gcc401/bin/gcj -g -I. -c -o cniinner/native.o
> > ../../../frysk/frysk-imports/tests/cniinner/native.cxx
> > /opt/gcc401/bin/gcj -g -I. --main=cniinner.Child -o cniinner/a.out \
> >        ../../../frysk/frysk-imports/tests/cniinner/Parent.java \
> >        ../../../frysk/frysk-imports/tests/cniinner/Child.java \
> >        cniinner/native.o
> > ../../../frysk/frysk-imports/tests/cniinner/Parent.java:3: internal compiler
> > error: in mark_reference_fields, at java/boehm.c:105
> > Please submit a full bug report,
> > with preprocessed source if appropriate.
> > See <URL:http://gcc.gnu.org/bugs.html> for instructions.
> > make: *** [cniinner/a.out] Error 1
> > 
> > Maybe I can submit a bug report for GNU java community?
> >  
> The intent is for Fedora Core 5 to ship with 4.1, the test releases (FC-5 Test
> 2 is just out) include preliminary versions of the compiler.

I made a test with the 20060106 snapshot of gcc-4.1, it still failed.  
Janis Johnson also confirmed that it failed in the same way on GCC trunk, 
gcc-3.4.2 building in 32-bits mode, with "-m64" option however, it will 
report only an undefined reference:

/tmp/ccggTln9.o(.toc+0x10): undefined reference to `cniinner::child::class$'
collect2: ld returned 1 exit status

> 
> I think the java group would be very interested  The original i386 bug
> (botched object method offsets leaving a virtual address null) resulted in a
> segmentation fault.  You've got something new different.
> 

Before submitting a problem reprot, I want to ask if I can make these code 
public (Parent.java, Child.java and native.cxx)?  I guess it is, but I 
still want to make sure.  :-)

> > > > FAIL: rh177240/strace-clone-exec.sh
> > > > 
> > > >      
> > > Now there is likely a kernel bug.  It would explain this failure:
> > > 
> > >    
> > 
> > Can you point me to any link about this kernel bug?  Thanks a lot!
> >  
> The above was testing for a regression in both the FC-5 and correspoinding
> upstream kernel.  The FC-5 bug report is here:
> https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=177240 sounds like it has
> tickled another bug.

I am now believing that it is the same bug.  Because I run the original 
test on a 2.6.15 kernel.  I switch the kenel back to RHEL4 U2 shipped one, 
it report PASS this time.  So the imports check on PPC64 only reports one 
FAILURE now,  Great!

BTW, I want to ask a question, to make Frysk work on PPC64, what extra 
work needed to be done?  I means these code different than the common 
ones.  If you can share your thought, that is much appreciated. 

Regards
- Wu Zhou