Re: JNI on ARM/XScale

[Andrew Haley <aph at redhat dot com>]

Bryce McKinlay writes:
 > Jari Korva wrote:
 > 
 > >On Thu, 30 Sep 2004, Bryce McKinlay wrote:
 > >  
 > >
 > >>Hmm, so if you printf() lib_name immediately before and after the
 > >>JvSynchronize(), it prints the correct value before the call and
 > >>corrupted afterwards? If so, that sounds like a C++ compiler bug, as a
 > >>value on the stack is somehow being overwritten.
 > >>
 > >>You could try building natRuntime.cc with -O0 to confirm this theory.
 > >>    
 > >>
 > >
 > >Thanks again! I removed -O2 from libjava/Makefile, touched natRuntime.cc
 > >and ran make again. Then I rebuild my JNI-HelloWorld - and it really
 > >worked!
 > >  
 > >
 > 
 > It'd be interesting to know if you can still reproduce this with the 
 > latest code from CVS. If so, then it would be worth trying to isolate a 
 > small test case and reporting a compiler bug.

Indeed.

 > Andrew Haley wrote:
 > 
 > > > I also tried printfs instead of debugger on -O2 optimized binary. That
 > > > way, it seems that the value is corrupted already before JvSynchronize.
 > > > The trace and modified source are below (the most bizarre thing is that it
 > > > jumps directly from trace 3-1 to 4).
 > >
 > >There's nothing weird about that: all it means is that you're getting
 > >an ArrayIndexOutOfBoundsException.  This is Bad.
 > >  
 > >
 > 
 > ... but probably "normal" if stack traces are not working correctly on 
 > this version/target?

Well, that depends on what " not working correctly" means.  It might
mean something very bad like wandering off into never-never land, or
something quite harmeless.

Andrew.