[Bug tree-optimization/48702] [4.6/4.7 Regression] optimization regression with gcc-4.6 on x86_64-unknown-linux-gnu

[xinliangli at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48702

--- Comment #14 from davidxl <xinliangli at gmail dot com> 2011-05-17 17:17:11 UTC ---
(In reply to comment #13)
> (In reply to comment #12)
> > (In reply to comment #3)
> > > I have a patch that makes it fail on trunk as well.  IVOPTs generates
> > > 
> > >   for (p = &a; p != &a - 3; --p)
> > >     *(p + 3) = ...
> > > 
> > > and alias analysis doesn't like this invalid pointer.
> > 
> > 
> > I wonder why ivopt does not select the iv candidate whose base is &a+3.
> 
> I think that one is not in the list of initial candidates.
> 
> I think for that sake we would want to add stripped &base + object size (+ 1?)
> (if we know it) as IV candidate iff iv->step is negative, iff iv->step is
> positive continue to add &base.
> 
> I don't like the alias oracle fixups too much, but I guess we have to
> benchmark both alternatives.

The candidate is actually there:

candidate 8
  var_before ivtmp.13
  var_after ivtmp.13
  incremented before exit test
  type long unsigned int
  base (long unsigned int) &MEM[(int *)&array][3]{lb: 0 sz: 4}
  step 0x0fffffffffffffffc
  base object (void *) &array

The following patch fixes the problem. Is it ok?

David


--- tree-ssa-loop-ivopts.c      (revision 173278)
+++ tree-ssa-loop-ivopts.c      (working copy)
@@ -3968,7 +3968,7 @@ get_computation_cost_at (struct ivopts_d
                          int *inv_expr_id)
 {
   tree ubase = use->iv->base, ustep = use->iv->step;
-  tree cbase, cstep;
+  tree cbase, cstep, cbase_strip;
   tree utype = TREE_TYPE (ubase), ctype;
   unsigned HOST_WIDE_INT cstepi, offset = 0;
   HOST_WIDE_INT ratio, aratio;
@@ -4026,6 +4026,13 @@ get_computation_cost_at (struct ivopts_d
   if (!constant_multiple_of (ustep, cstep, &rat))
     return infinite_cost;

+  cbase_strip = STRIP_NOPS (cbase);
+  /* Avoid confusing aliaser.  */
+  if (TREE_CODE (cbase_strip) == ADDR_EXPR
+      && TREE_CODE (TREE_OPERAND (cbase_strip, 0)) == VAR_DECL
+      && (HOST_WIDE_INT) cstepi < 0)
+    return infinite_cost;
+
   if (double_int_fits_in_shwi_p (rat))
     ratio = double_int_to_shwi (rat);
   else