[Bug tree-optimization/53044] New: completely peel loops that do not run a constant time

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53044

             Bug #: 53044
           Summary: completely peel loops that do not run a constant time
    Classification: Unclassified
           Product: gcc
           Version: 4.8.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassig...@gcc.gnu.org
        ReportedBy: rgue...@gcc.gnu.org


There exist loops where we do not know how many times they run.  Consider

int objects[2];

void foo (int n)
{
  int i;
  for (i = 0; i < n; ++i)
    objects[i] = i;
}

we do know the loop will run at most twice (due to undefined behavior
analysis, the info is accessible via max_loop_iterations (loop)).

It could be beneficial to completely peel such loops (retaining the
exit checks, of course).  At least for loops that do not run many
times (like at most once or twice, as in the above example).

int objects[1];

void foo (int n)
{
  int i;
  for (i = 0; i < n; ++i)
    objects[i] = i;
}

jump threading should handle the above case from VRP, but currently
VRP transforms the above to

<bb 2>:
  if (n_2(D) > 0)
    goto <bb 4>;
  else
    goto <bb 3>;

<bb 3>:
  return;

<bb 4>:

<bb 5>:
  ivtmp.5_1 = 0;
  i_9 = 0;
  MEM[(int[1] *)&objects] = 0;
  ivtmp.5_8 = 1;
  i_7 = 1;
  if (n_2(D) > 1)
    goto <bb 7>;
  else
    goto <bb 6>;

<bb 6>:
  goto <bb 3>;

<bb 7>:
  goto <bb 5>;

thus, a conditionally endless loop (ugh, a side-effect that we even preserve).