[Bug middle-end/47653] [4.6 Regression] gcc.c-torture/execute/930614-2.c FAILs with -fstack-check=generic -fgraphite-identity

["law at redhat dot com" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47653

Jeffrey A. Law <law at redhat dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2
                 CC|                            |law at redhat dot com
         AssignedTo|unassigned at gcc dot       |spop at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #2 from Jeffrey A. Law <law at redhat dot com> 2011-02-18 15:36:46 UTC ---
This appears to be a bug in the graphite transformations. Prior to graphite
transformations we have the following function with two nested loops:

  # BLOCK 2 freq:139
  # PRED: ENTRY [100.0%]  (fallthru,exec)
  # .MEM_13 = VDEF <.MEM_12(D)>
  saved_stack.1_3 = __builtin_stack_save ();
  # .MEM_14 = VDEF <.MEM_13>
  x.0_4 = __builtin_alloca (256);
  goto <bb 8>;
  # SUCC: 8 [100.0%]  (fallthru,exec)

  # BLOCK 3 freq:7901
  # PRED: 4 [88.9%]  (true,exec)
  # SUCC: 4 [100.0%]  (fallthru,exec)

  # BLOCK 4 freq:8889
  # PRED: 3 [100.0%]  (fallthru,exec) 9 [100.0%]  (fallthru,exec)
  # j_23 = PHI <j_7(3), j_5(9)>
  # .MEM_24 = PHI <.MEM_15(3), .MEM_6(9)>
  # .MEM_15 = VDEF <.MEM_24>
  *x.0_4[j_5][j_23] = 0;
  j_7 = j_23 + 1;
  if (j_7 <= 7)
    goto <bb 3>;
  else
    goto <bb 5>;
  # SUCC: 3 [88.9%]  (true,exec) 5 [11.1%]  (false,exec)

  # BLOCK 5 freq:988
  # PRED: 4 [11.1%]  (false,exec)
  # .MEM_19 = PHI <.MEM_15(4)>
  # SUCC: 6 [100.0%]  (fallthru,exec)

  # BLOCK 6 freq:1111
  # PRED: 5 [100.0%]  (fallthru,exec) 8 [11.1%]  (false,exec)
  # .MEM_25 = PHI <.MEM_19(5), .MEM_6(8)>
  i_8 = j_5 + 1;
  if (i_8 <= 7)
    goto <bb 7>;
  else
    goto <bb 10>;
  # SUCC: 7 [88.9%]  (true,exec) 10 [11.1%]  (false,exec)

  # BLOCK 7 freq:988
  # PRED: 6 [88.9%]  (true,exec)
  # SUCC: 8 [100.0%]  (fallthru,exec)

  # BLOCK 8 freq:1111
  # PRED: 7 [100.0%]  (fallthru,exec) 2 [100.0%]  (fallthru,exec)
  # j_5 = PHI <i_8(7), 0(2)>
  # .MEM_6 = PHI <.MEM_25(7), .MEM_14(2)>
  if (j_5 <= 7)
    goto <bb 9>;
  else
    goto <bb 6>;
  # SUCC: 9 [88.9%]  (true,exec) 6 [11.1%]  (false,exec)

  # BLOCK 9 freq:988
  # PRED: 8 [88.9%]  (true,exec)
  goto <bb 4>;
  # SUCC: 4 [100.0%]  (fallthru,exec)

  # BLOCK 10 freq:139
  # PRED: 6 [11.1%]  (false,exec)
  # .MEM_16 = VDEF <.MEM_25>
  __builtin_stack_restore (saved_stack.1_3);
  return 0;
  # SUCC: EXIT [100.0%]

Pretty simple stuff.

Graphite transforms it into:

  # BLOCK 2 freq:313
  # PRED: ENTRY [100.0%]  (fallthru,exec)
  # .MEM_13 = VDEF <.MEM_12(D)>
  saved_stack.1_3 = __builtin_stack_save ();
  # .MEM_14 = VDEF <.MEM_13>
  x.0_4 = __builtin_alloca (256);
  # SUCC: 3 [100.0%]  (fallthru)

  # BLOCK 3 freq:2500
  # PRED: 2 [100.0%]  (fallthru) 8 [100.0%]  (fallthru,dfs_back)
  # graphite_IV.5_11 = PHI <0(2), graphite_IV.5_1(8)>
  # .MEM_33 = PHI <.MEM_14(2), .MEM_34(8)>
  D.2701_22 = 8 - graphite_IV.5_11;
  D.2702_21 = D.2701_22 > 0;
  if (D.2702_21 != 0)
    goto <bb 4>;
  else
    goto <bb 7>;
  # SUCC: 4 [50.0%]  (true) 7 [50.0%]  (false)

  # BLOCK 4 freq:1250
  # PRED: 3 [50.0%]  (true)
  D.2704_10 = (<unnamed-signed:64>) graphite_IV.5_11;
  D.2705_2 = D.2704_10 * 4294967295;
  D.2706_18 = D.2705_2 + 7;
  # SUCC: 5 [100.0%]  (fallthru)

  # BLOCK 5 freq:10000
  # PRED: 4 [100.0%]  (fallthru) 6 [100.0%]  (fallthru,dfs_back)
  # graphite_IV.6_17 = PHI <0(4), graphite_IV.6_26(6)>
  # .MEM_35 = PHI <.MEM_33(4), .MEM_27(6)>
  D.2707_28 = (int) graphite_IV.5_11;
  D.2708_29 = (int) graphite_IV.5_11;
  D.2709_30 = (int) graphite_IV.6_17;
  D.2710_31 = D.2708_29 + D.2709_30;
  # .MEM_27 = VDEF <.MEM_35>
  *x.0_4[D.2707_28][D.2710_31] = 0;
  graphite_IV.6_26 = graphite_IV.6_17 + 1;
  if (graphite_IV.6_17 < D.2706_18)
    goto <bb 6>;
  else
    goto <bb 7>;
  # SUCC: 6 [87.5%]  (true) 7 [12.5%]  (loop_exit,false)

  # BLOCK 6 freq:8750
  # PRED: 5 [87.5%]  (true)
  goto <bb 5>;
  # SUCC: 5 [100.0%]  (fallthru,dfs_back)

  # BLOCK 7 freq:2500
  # PRED: 5 [12.5%]  (loop_exit,false) 3 [50.0%]  (false)
  # .MEM_34 = PHI <.MEM_27(5), .MEM_33(3)>
  graphite_IV.5_1 = graphite_IV.5_11 + 1;
  if (graphite_IV.5_11 < 7)
    goto <bb 8>;
  else
    goto <bb 9>;
  # SUCC: 8 [87.5%]  (true) 9 [12.5%]  (loop_exit,false)

  # BLOCK 8 freq:2188
  # PRED: 7 [87.5%]  (true)
  goto <bb 3>;
  # SUCC: 3 [100.0%]  (fallthru,dfs_back)

  # BLOCK 9 freq:313
  # PRED: 7 [12.5%]  (loop_exit,false)
  # .MEM_32 = PHI <.MEM_34(7)>
  # .MEM_16 = VDEF <.MEM_32>
  __builtin_stack_restore (saved_stack.1_3);
  return 0;
  # SUCC: EXIT [100.0%]


Of particular interest is the assignment to D.2705_2 in block #4:

  D.2705_2 = D.2704_10 * 4294967295;
  D.2706_18 = D.2705_2 + 7;

That makes absolutely no sense.  Particularly since D.2706_18 is later used to
control loop termination in BB5.    Note D.2706 is a 64bit type, so we really
are multiplying by 4294967295.  Needless to say this causes the loop
termination condition to do something different than was originally intended.