[Bug middle-end/122493] New: is_eye in leela is not split because of empty basic block

[hubicka at gcc dot gnu.org via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122493

            Bug ID: 122493
           Summary: is_eye in leela is not split because of empty basic
                    block
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hubicka at gcc dot gnu.org
  Target Milestone: ---

is_eye in leela would benefit from partial inlining that does not happen.  We
have:

  <bb 2> [local count: 1073741822]:
  _1 = (long unsigned int) i_62(D);
  # VUSE <.MEM_64(D)>
  _4 = MEM <struct array> [(short unsigned int &)this_63(D) + 7076].elems[_1];
  _5 = (int) _4;
  _6 = (long unsigned int) color_66(D);
  # VUSE <.MEM_64(D)>
  _8 = MEM <const struct array> [(const int &)&s_eyemask].elems[_6];
  ownsurrounded_68 = _5 & _8;
  if (ownsurrounded_68 == 0)
    goto <bb 3>; [34.00%]
  else
    goto <bb 5>; [66.00%]

  <bb 3> [local count: 365072224]:
  // predicted unlikely by early return (on trees) predictor.

  <bb 4> [local count: 636386382]:
  # .MEM_21 = PHI <.MEM_64(D)(3), .MEM_79(7), .MEM_79(9)>
  goto <bb 10>; [100.00%]

  <bb 5> [local count: 708669599]:
  # .MEM_69 = VDEF <.MEM_64(D)>
  colorcount[0] = 0;
  # .MEM_70 = VDEF <.MEM_69>
  colorcount[1] = 0;
  # .MEM_71 = VDEF <.MEM_70>
  colorcount[3] = 0;
  _9 = i_62(D) + -1;
  # VUSE <.MEM_71>
  _10 = this_63(D)->m_boardsize;

....

  <bb 10> [local count: 1073741824]:
  # _60 = PHI <0(4), 1(6), 1(8)>
  # .MEM_61 = PHI <.MEM_21(4), .MEM_79(6), .MEM_79(8)>
  # .MEM_84 = VDEF <.MEM_61>
  colorcount ={v} {CLOBBER(eos)};
  # VUSE <.MEM_84> 
  return _60;

}


and we would like to split at BB5. However the tail also does:
  <bb 8> [local count: 354334800]:
  _57 = color_66(D) == 0;
  _58 = (int) _57;
  # VUSE <.MEM_79>
  _59 = colorcount[_58];
  if (_59 != 0)
    goto <bb 9>; [34.00%]
  else 
    goto <bb 10>; [66.00%]



  <bb 7> [local count: 150840325]:
  // predicted unlikely by early return (on trees) predictor.
  goto <bb 4>; [100.00%]

Whch means that it sometimes terminated by jumping to return block but
sometimes to bb 4 which is forwarder to return block.  This makes the split
point not recognized.  Sceduling extra merge_phi:

diff --git a/gcc/passes.def b/gcc/passes.def
index fac04cd86c7..a8a89b980a5 100644
--- a/gcc/passes.def
+++ b/gcc/passes.def
@@ -97,6 +97,7 @@ along with GCC; see the file COPYING3.  If not see
          NEXT_PASS (pass_merge_phi);
           NEXT_PASS (pass_dse);
          NEXT_PASS (pass_cd_dce, false /* update_address_taken_p */, true /*
remove_unused_locals */);
+         NEXT_PASS (pass_merge_phi);
          NEXT_PASS (pass_phiopt, true /* early_p */);
          /* Cleanup eh is done before tail recusision to remove empty (only
clobbers)
             finally blocks which would block tail recursion.  */

solves the problem. However I think this can be done by cfgcleanup itself.