Hi,

I was encountering wacky behavior in my hacked copy of Kaffe when
running lots of threads in some simple tests.  (I couldn't get the
release version to duplicate the problem, but I think its still a
problem in 0.9.2.)  Basically, a bunch of stack variables were getting
trashed in intrp/machine.c::virtualMachine() (the interpreter loop.)
I ran 'gcc -O2 -Wall' on it, and gcc tells me that a bunch of
variables might get trashed by longjmp()---exactly the variables that
were getting trashed.  There is a single setjmp() in virtualMachine().
I got rid of the warnings (and the problem!) by putting the setjmp in
a separate function (I read that the problems with setjmp/longjmp are
limited to the active stack frame).  Attached is a patch for
intrp/machine.c.  Anyone with more experience with setjmp/longjmp care
to explain what was going on?  ;)

-Pat

--- /home/grad/tullmann/java/kaffe-0.9.2/kaffe/kaffevm/intrp/machine.c Sat Oct  4 04:26:57 1997
+++ /home/grad/tullmann/tmp/machine.c   Wed Oct 22 11:59:46 1997
@@ -68,6 +68,12 @@
 /* Misc stuff */
 Hjava_lang_Object* exceptionObject;
 
+static int
+exceptionSetjmp(vmException *mjbuf)
+{
+       return setjmp(mjbuf->jbuf);
+}
+
 void
 virtualMachine(methods* meth, slots* arg, slots* retval)
 {
@@ -140,7 +146,7 @@
                TCTX(currentThread)->exceptPtr = &mjbuf;
        }
        if (meth->exception_table != 0 || (methaccflags & ACC_SYNCHRONISED)) {
-               if (setjmp(mjbuf.jbuf) != 0) {
+               if (exceptionSetjmp(&mjbuf) != 0) {
                        assert(TCTX(currentThread)->exceptPtr == &mjbuf);
                        npc = mjbuf.pc;
                        sp = &lcl[meth->localsz + meth->stacksz - 1];