Re: PR java/17689 [4.0 Regression] New testsuite failures indicatesignal unwinding is broken

[Andrew Haley <aph at redhat dot com>]

Diego Novillo writes:
 > On Mon, 2004-09-27 at 10:38, Andrew Haley wrote:
 > > It seems like a write to a volatile mem is being deleted.  This may
 > > not be a real bug, because libgcj is obtaining the mem by doing some
 > > blatantly illegal addressing into the stack.
 > > 
 > > I should rework i386-signal.h to access the kernel struct in a legal
 > > way.  Until then, this patch makes gcj usable again.
 > > 
 > > However, I think the regression is down to this patch:
 > > 
 > Yes, and I think I know why.  If we are processing a mix of volatile and
 > non-volatile pointers, we'll mess things up because the code just copies
 > TREE_THIS_VOLATILE from the type.
 > 
 > I'm sorry for the mess.  Odd that it didn't trigger here.  *blush*
 > If you can't test now, let me know and I'll have it tested later today.

I couldn't get your patch to work, but this did: 

Index: tree-ssa-alias.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/tree-ssa-alias.c,v
retrieving revision 2.43
diff -c -2 -p -r2.43 tree-ssa-alias.c
*** tree-ssa-alias.c	23 Sep 2004 14:34:19 -0000	2.43
--- tree-ssa-alias.c	27 Sep 2004 16:20:48 -0000
*************** create_memory_tag (tree type, bool is_ty
*** 2029,2032 ****
--- 2029,2035 ----
    DECL_CONTEXT (tag) = current_function_decl;
  
+   /* If the pointed-to type is volatile, so is the tag.  */
+   TREE_THIS_VOLATILE (tag) |= TREE_THIS_VOLATILE (type);
+ 
    /* Memory tags are by definition addressable.  This also prevents
       is_gimple_ref frome confusing memory tags with optimizable
*************** get_tmt_for (tree ptr, struct alias_info
*** 2128,2132 ****
  
    /* If the pointed-to type is volatile, so is the tag.  */
!   TREE_THIS_VOLATILE (tag) = TREE_THIS_VOLATILE (tag_type);
  
    /* Make sure that the type tag has the same alias set as the
--- 2131,2136 ----
  
    /* If the pointed-to type is volatile, so is the tag.  */
!   if (TREE_THIS_VOLATILE (tag_type))
!     TREE_THIS_VOLATILE (tag) = 1;
  
    /* Make sure that the type tag has the same alias set as the

???

Andrew.