[Bug tree-optimization/48316] New: missed CSE / reassoc with array offsets

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48316

           Summary: missed CSE / reassoc with array offsets
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: rguenth@gcc.gnu.org


int
foo (int *p, int i)
{
  int i1 = i + 1;
  int i2 = i + 2;
  return p[i1] + p[i2];
}

int
bar (int *p, unsigned long i)
{
  unsigned long i1 = i + 1;
  unsigned long i2 = i + 2;
  return p[i1] + p[i2];
}

For both testcases (the latter being the more "optimal" input due to
pointer-plus-expr constraints) we miss to CSE the multiplication
of i by 4 which makes the memory references not trivially independent
(based on the same pointer, offsetted by different constants).  Such
a case causes vectorization for alias checks being inserted for
gfortran.dg/reassoc_4.f with --param max-completely-peeled-insns=4000

IL on x86_64 is

<bb 2>:
  i1_2 = i_1(D) + 1;
  i2_3 = i_1(D) + 2;
  D.2702_4 = (long unsigned int) i1_2;
  D.2703_5 = D.2702_4 * 4;
  D.2704_7 = p_6(D) + D.2703_5;
  D.2705_8 = MEM[(int *)D.2704_7];
  D.2706_9 = (long unsigned int) i2_3;
  D.2707_10 = D.2706_9 * 4;
  D.2708_11 = p_6(D) + D.2707_10;
  D.2709_12 = MEM[(int *)D.2708_11];
  D.2701_13 = D.2705_8 + D.2709_12;
  return D.2701_13;

vs.

<bb 2>:
  i1_2 = i_1(D) + 1;
  i2_3 = i_1(D) + 2;
  D.2694_4 = i1_2 * 4;
  D.2695_6 = p_5(D) + D.2694_4;
  D.2696_7 = MEM[(int *)D.2695_6];
  D.2697_8 = i2_3 * 4;
  D.2698_9 = p_5(D) + D.2697_8;
  D.2699_10 = MEM[(int *)D.2698_9];
  D.2693_11 = D.2696_7 + D.2699_10;
  return D.2693_11;

For the reassoc_4.f testcase the question is whether either SCEV or
data-dependence can be enhanced to handle the cases (the multiplications
are in BB2, outside of any loop).