ARM: Fix occasional safepoint crash

[Andrew Haley]

We have been seeing occasional crashes, sometimes in the garbage
collector, on ARM.  These tended to happen at times of high memory
allocation and therefore lots of GC activity.

With help from Andrew Dinn I found the problem.  The frame walking
code that finds roots during garbage collection wals the stack, so you
have to make sure that the saved SP is correct.  It wasn't: the C++
interpreter assumes that the TOS item is saved at sp[1], not sp[0].
That is to say, the code used to get the address of the top item is:

inline intptr_t* frame::interpreter_frame_tos_address() const {
  return get_interpreterState()->_stack + 1;
}

It's easy enough to fix this: adjust the saved SP when we enter a
safepoint.

Andrew.


diff -r f293db049783 src/cpu/zero/vm/thumb2.cpp
--- a/src/cpu/zero/vm/thumb2.cpp        Mon May 28 08:01:18 2012 -0400
+++ b/src/cpu/zero/vm/thumb2.cpp        Mon May 28 08:39:18 2012 -0400
@@ -4244,6 +4244,15 @@
   //if (offset != 0) {
     Thumb2_save_locals(jinfo, stackdepth);
     //}
+
+  // The frame walking code used by the garbage collector
+  // (frame::interpreter_frame_tos_address()) assumes that the stack
+  // pointer points one word below the top item on the stack, so we
+  // have to adjust the SP saved in istate accordingly.  If we don't,
+  // the value on TOS won't be seen by the GC and we will crash later.
+  sub_imm(jinfo->codebuf, ARM_R0, Rstack, 4);
+  store_istate(jinfo, ARM_R0, ISTATE_STACK, stackdepth);
+
   mov_imm(jinfo->codebuf, ARM_R1, bci+CONSTMETHOD_CODEOFFSET);
   add_imm(jinfo->codebuf, ARM_R2, ISTATE_REG(jinfo),
          ISTATE_OFFSET(jinfo, stackdepth, 0));