[Bug tree-optimization/38645] [4.4 Regression] ICE with volatile

[jamborm at gcc dot gnu dot org]

------- Comment #3 from jamborm at gcc dot gnu dot org  2008-12-30 16:58 -------
I discussed  this bug with  Richi on IRC  and was told that  we should
avoid  having the  statement marked  as volatile  since it  is  not an
access  through  a volatile  variable  in  the  original source  code.
Instead, we should try preventing the propagation of &a to uses of p.

The patch  is below, it  prevents ccp and  dom1 from carrying  out the
propagation.  However,  fwprop does  this propagation, albeit  using a
VIEW_CONVERT_EXP and, moreover, fold_stmt_inplace() it calls marks the
statement  volatile  anyway.  This  propagation  is performed  because
fwprop checks types (not decls) for volatility and apparently the type
of a is not volatile (as opposed to a[0] which probably is).

The declaration of  a is volatile though.  However,  when I changed it
to  check the  declaration,  nrv pass  did  the propagation,  creating
invalid gimple along the way.   Because the such change would possibly
be incorrect  it is  not included below.   Anyway, the test  case with
this patch compiles  ans it is possibly slightly  more correct, though
the final result is not all that different.

2008-12-30  Martin Jambor  <mjambor@suse.cz>

        * tree-ssa-dom.c (cprop_operand): Always check the return value of
        may_propagate_copy.
        * tree-ssa-copy.c (may_propagate_copy): Do not allow propagation
        of addresses of volatile declarations.

Index: gcc/tree-ssa-dom.c
===================================================================
--- gcc/tree-ssa-dom.c  (revision 142962)
+++ gcc/tree-ssa-dom.c  (working copy)
@@ -2092,7 +2092,7 @@ cprop_operand (gimple stmt, use_operand_
       /* Certain operands are not allowed to be copy propagated due
         to their interaction with exception handling and some GCC
         extensions.  */
-      else if (!may_propagate_copy (op, val))
+      if (!may_propagate_copy (op, val))
        return false;

       /* Do not propagate copies if the propagated value is at a deeper loop
Index: gcc/tree-ssa-copy.c
===================================================================
--- gcc/tree-ssa-copy.c (revision 142962)
+++ gcc/tree-ssa-copy.c (working copy)
@@ -73,6 +73,15 @@ may_propagate_copy (tree dest, tree orig
       && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (dest))
     return false;

+  if (TREE_CODE (orig) == ADDR_EXPR)
+  {
+    tree base = TREE_OPERAND (orig, 0);
+    while (handled_component_p (base))
+      base = TREE_OPERAND (base, 0);
+    if (TREE_THIS_VOLATILE (base))
+       return false;
+  }
+
   /* For memory partitions, copies are OK as long as the memory symbol
      belongs to the partition.  */
   if (TREE_CODE (dest) == SSA_NAME


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38645