[tree-ssa] Another aliasing speedup

[law at redhat dot com]

While investigating elimination of the double-simplification of trees,
I stumbled on the solution for a compile-time performance issue I tabled
in the aliasing code.

Specifically we're walking an excessive number of nodes when searching
for variables.  If you debug find_vars_r and its calls from walk_tree
you'd see that we walk into the children of type nodes, children of
_DECL nodes and such.  None of those walks are useful.

This patch stops the walker from descending into children unnecessarily
by clearing walk_subtrees.  It cuts another 35-40% off the alias analysis
time.  In real terms it's about 1.5 seconds of my components of cc1 test.

Anyway, now that's done and I'll return to trying to avoid double-
gimplification of our trees.


	* tree-dfa.c (find_vars_r): Clear *walk_subtrees appropriately
	to avoid useless walking of subtrees.

Index: tree-dfa.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/Attic/tree-dfa.c,v
retrieving revision 1.1.4.82
diff -c -3 -p -r1.1.4.82 tree-dfa.c
*** tree-dfa.c	26 Feb 2003 16:26:03 -0000	1.1.4.82
--- tree-dfa.c	27 Feb 2003 22:48:59 -0000
*************** find_vars_r (tp, walk_subtrees, data)
*** 2103,2108 ****
--- 2103,2126 ----
    struct walk_state *walk_state = (struct walk_state *)data;
    int saved_is_store = walk_state->is_store;
  
+   /* Type and constant nodes have no interesting children.  Ignore them.  */
+   if (TYPE_P (t) || TREE_CODE_CLASS (TREE_CODE (t)) == 'c')
+     {
+       *walk_subtrees = 0;
+       return NULL_TREE;
+     }
+ 
+   /* DECL nodes have no interesting children.  */
+   if (DECL_P (t))
+     {
+       *walk_subtrees = 0;
+ 
+       /* If this _DECL node is not interesting to the SSA builder,
+          then we can just return now.  */
+       if (! SSA_VAR_P (t)) 
+ 	return NULL_TREE;
+     }
+ 
    if (TREE_CODE (t) == MODIFY_EXPR || TREE_CODE (t) == INIT_EXPR)
      {
        walk_state->is_store = 1;