[Bug tree-optimization/93042] New: bit-field optimizations get in the way of interchange

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93042

            Bug ID: 93042
           Summary: bit-field optimizations get in the way of interchange
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---

Take:
#define N 100
#define M 1111
struct S { int a : 3; int b : 17; int c : 12; };
struct S A[N][M];

int __attribute__((noinline))
foo (void)
{
  int i, j;

  for( i = 0; i < M; i++)
    for( j = 0; j < N; j++)
      {
        int t = 0;
        struct S s1 = A[j][i];
        if (s1.b)
          t++;
        A[j][i].a = t;
      }

  return A[0][0].a + A[N-1][M-1].a;
}
----- CUT ----
If we change the loop slightly, to not to use a temp struct or even add a temp
var to hold s1.b, the loop can be interchanged.
That is:
int __attribute__((noinline))
foo1 (void)
{
  int i, j;

  for( i = 0; i < M; i++)
    for( j = 0; j < N; j++)
      {
        int t = 0;
        if (A[j][i].b)
          t++;
        A[j][i].a = t;
      }

  return A[0][0].a + A[N-1][M-1].a;
}
int __attribute__((noinline))
foo2 (void)
{
  int i, j;

  for( i = 0; i < M; i++)
    for( j = 0; j < N; j++)
      {
        int t = 0;
        struct S s1 = A[j][i];
        int t1 = s1.b;
        if (t1)
          t++;
        A[j][i].a = t;
      }

  return A[0][0].a + A[N-1][M-1].a;
}
----- CUT ----
This shows up slightly different when I add lowering of bit-field accesses but
in a different way.