[Bug tree-optimization/63168] loop header copying fails - not vectorized: latch block not empty

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63168

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
It looks like do_while_loop_p is fooled by the CFG structure and hits

static bool
do_while_loop_p (struct loop *loop)
{
  gimple stmt = last_stmt (loop->latch);

  /* If the latch of the loop is not empty, it is not a do-while loop.  */
  if (stmt
      && gimple_code (stmt) != GIMPLE_LABEL)
    return false;

the loop at CH time looks like
      |
      v
/-> <bb 5>
|      exit test  --->
|      |
|   <bb 3>
|      ...
|      |
\   <bb 4>
 --   (empty but with a PHI)

phiopt introduces this singleton PHI:

   <bb 3>:
   max_8 = *it_6;
-  if (max_8 > max_2)
-    goto <bb 5>;
-  else
-    goto <bb 4>;
+  max_10 = MAX_EXPR <max_2, max_8>;

   <bb 4>:
+  # max_9 = PHI <max_10(3)>

   <bb 5>:
-  # max_9 = PHI <max_2(4), max_8(3)>
-
-  <bb 6>:
-  # it_1 = PHI <it_6(5), it_4(D)(2)>
-  # max_2 = PHI <max_9(5), max_5(2)>
+  # it_1 = PHI <it_6(4), it_4(D)(2)>
+  # max_2 = PHI <max_9(4), max_5(2)>

it looks like the CH predicate relies on copy-propagated form
(a latch block with an SSA name copy should also be considered empty IMHO).

I think that cfgcleanup should merge

  <bb A>
   ....
    | single-succ/pred
  <bb B>
    # var = PHI <arg>

and rewrite the PHIs to copies.  Its comment even says so:

  /* Merging the blocks may create new opportunities for folding
     conditional branches (due to the elimination of single-valued PHI
     nodes).  */
  if (single_succ_p (bb)
      && can_merge_blocks_p (bb, single_succ (bb)))
    {
      merge_blocks (bb, single_succ (bb));
      return true;
    }

Ah, but

/* Checks whether we can merge block B into block A.  */

static bool
gimple_can_merge_blocks_p (basic_block a, basic_block b)
{
...
  /* Protect the loop latches.  */
  if (current_loops && b->loop_father->latch == b)
    return false;

triggers here (merge_blocks doesn't handle merging with a latch).

Mine.