[Bug tree-optimization/88405] New: Missed DSE opportunity

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88405

            Bug ID: 88405
           Summary: Missed DSE opportunity
           Product: gcc
           Version: 9.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ktkachov at gcc dot gnu.org
  Target Milestone: ---

For the following code:

#define MATRIX_SIZE 512

static double a[MATRIX_SIZE][MATRIX_SIZE];
static double b[MATRIX_SIZE][MATRIX_SIZE];
static double c[MATRIX_SIZE][MATRIX_SIZE];

double
foo (void) {
  double s;
  int i, j, k;
  /* Section A */
  for (i = 0; i < MATRIX_SIZE; i++) {
    for (j = 0; j < MATRIX_SIZE; j++) {
      a[i][j] = (double)i * (double)j;
      b[i][j] = (double)i / (double)(j+5);
    }
  }
  /* Section B */
  for (j = 0; j < MATRIX_SIZE; j++) {
    for (i = 0; i < MATRIX_SIZE; i++) {
      s = 0;
      for (k = 0; k < MATRIX_SIZE; k++) {
        s += a[i][k] * b[k][j];
      }
      c[i][j] = s;
    }
  }
  s = 0.0; // (1)
#if 0
  /* Section C */
  for (i = 0; i < MATRIX_SIZE; i++) {
    for (j = 0; j < MATRIX_SIZE; j++) {
      s += c[i][j];
    }
  }
#endif
  return s;
}

GCC does not manage to eliminate the code up to (1) and retains the expensive
Section A.

Clang manages to eliminate much more and produces:
foo:                                    // @foo
// %bb.0:                               // %entry
        orr     w8, wzr, #0x200
.LBB0_1:                                // %vector.ph
                                        // =>This Inner Loop Header: Depth=1
        subs    x8, x8, #1              // =1
        b.ne    .LBB0_1
// %bb.2:                               // %for.cond20.preheader.preheader
        fmov    d0, xzr
        ret


on aarch64. This happens at -O3 as well as -O2 as well as other targets (occurs
also on x86)