[Bug tree-optimization/58326] [4.9 Regression] ICE in check_loop_closed_ssa_use, at tree-ssa-loop-manip.c:551

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58326

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
Ok, so for one, fix_bb_placements for the case of moved subloop (headers):

      if (from->loop_father->header == from)
        {
          /* Subloop header, maybe move the loop upward.  */
          if (!fix_loop_placement (from->loop_father, irred_invalidated))
            continue;
          target_loop = loop_outer (from->loop_father);
        }

doesn't include the header (or any of its body BBs in this case) in the
set of loop_closed_ssa_invalidated blocks.  At least the header can
use a name defined in the previous outer loop (and I fail to see why
a random loop body block couldn't).

But then, rewrite-into-LC-ssa does

static void
find_uses_to_rename_bb (basic_block bb, bitmap *use_blocks, bitmap need_phis)
{
  gimple_stmt_iterator bsi;
  edge e;
  edge_iterator ei;

  FOR_EACH_EDGE (e, ei, bb->succs)
    for (bsi = gsi_start_phis (e->dest); !gsi_end_p (bsi); gsi_next (&bsi))
      {
        gimple phi = gsi_stmt (bsi);
        if (! virtual_operand_p (gimple_phi_result (phi)))
          find_uses_to_rename_use (bb, PHI_ARG_DEF_FROM_EDGE (phi, e),
                                   use_blocks, need_phis);
      }

so for uses in PHIs it considers all destinations of edges out of
"changed" blocks (but not uses in PHIs of the block itself).  For
some weird reason.  Well - in this case one predecessor is in the
set of changed blocks and one (the one from the backedge) isn't.

Fixing fix_bb_placements to include all blocks of moved sub-loops
in loop_closed_ssa_invalidated fixes the testcases, but still the
above looks odd.