[Bug tree-optimization/61607] DOM missed jump threading and destroyed loops

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61607

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
With the propagation limitation removed we get

  Registering jump thread: (2, 4) incoming edge;  (4, 5) joiner;  (5, 7)
normal;
  Cancelling jump thread: (2, 4) incoming edge;  (4, 5) joiner;  (5, 7) normal;
Jump threading proved probability of edge 7->9 too small (it is 3900, should be
5000).
  Threaded jump 5 --> 7 to 10
fix_loop_structure: removing loop 1
flow_loops_find: discovered new loop 2 with header 4

  <bb 5>:
  # inter0p_13 = PHI <inter0p_2(4)>
  # inter1p_14 = PHI <inter0p_2(4)>
  if (inter0p_2 != 0)
    goto <bb 6>;
  else
    goto <bb 10>;

  <bb 6>:
  inter[0] = 1;

  <bb 7>:
  if (inter0p_2 != 0)
    goto <bb 8>;
  else
    goto <bb 9>;

  <bb 8>:
  inter[1] = 1;

  <bb 9>:
  foo (&inter);
  inter ={v} {CLOBBER};
  return;

  <bb 10>:
  goto <bb 9>;

that still misses the threading of the true path.  And it still destroys
loop info from:

static bool
thread_block_1 (basic_block bb, bool noloop_only, bool joiners)
{
...
  /* If we thread the latch of the loop to its exit, the loop ceases to
     exist.  Make sure we do not restrict ourselves in order to preserve
     this loop.  */
  if (loop->header == bb)
    {
      e = loop_latch_edge (loop);
      vec<jump_thread_edge *> *path = THREAD_PATH (e);

      if (path
          && (((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK && joiners)
              || ((*path)[1]->type == EDGE_COPY_SRC_BLOCK && !joiners)))
        {
          for (unsigned int i = 1; i < path->length (); i++)
            {
              edge e2 = (*path)[i]->e;

              if (loop_exit_edge_p (loop, e2))
                {
                  loop->header = NULL;
                  loop->latch = NULL;
                  loops_state_set (LOOPS_NEED_FIXUP);

but it seems we cancel the threading later...

(gdb) 
913                   delete_jump_thread_path (path);

so it would be better if we destroyed the loop only if necessary
(I would have expected the actual edge redirection code takes care of that).
Cancelling loops unnecessarily is very bad.