[Bug tree-optimization/56273] [4.8 regression] Bogus -Warray-bounds warning

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56273

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> 2013-02-11 09:32:31 UTC ---
(In reply to comment #2)
> The problem is a missing VRP.  Basically ivopts changes:
>   if (j_10 != 9)
> 
> Which works into:
> 
>   j_10 = ivtmp.12_56;
>   if (ivtmp.12_56 != 9)
>     goto <bb 7>;
>   else
>     goto <bb 6>;
> 
> 
> Which does not work as VRP cannot figure out j_10 will never be 9.
> I have some patches to VRP which improves this but I don't remember if it fixes
> this case where there is an assignment which is used later on.

A similar case was PR55079 which was

   j_10 = ivtmp.12_56 + 1;
   if (ivtmp.12_56 == 5)
     ...

I handled already.  Extending this to copies should be easy.  If I do that
it still doesn't work.

That's because we warn for stuff[j_15] and j_15 is [9, 16] (we completely
peel the loop).  The access is guarded with

  <bb 30>:
  j_55 = ivtmp.12_56 + 8;
  j_72 = ivtmp.12_56 + 8;
  if (j_72 != 9)
    goto <bb 31>;
  else
    goto <bb 3>;

  <bb 31>:
  # j_15 = PHI <j_55(30)>
  _12 = stuff[j_15].a;
  if (_12 != 0)

but j_72 is [9, 16] as well and ivtmp.12_56: [1, 9].  Improving VRP to
see that j_55 is != 9 in bb31 will only lead to [10, 16] which is also
out-of-bounds.

It might very well be that folding the < test to != in VRP1 causes most
of the issues.  Indeed -fdisable-tree-vrp1 "fixes" the testcase.  We
should probably delay that optimization to VRP2 only, thus don't do

Folding statement: if (i_1 <= 8)
Simplified relational if (i_1 <= 8)
 into if (i_1 != 9)

(simplify_cond_using_ranges).