[Bug middle-end/46675] [4.6 Regression] profiledbootstrap failed

[ubizjak at gmail dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=46675

--- Comment #3 from Uros Bizjak <ubizjak at gmail dot com> 2010-11-27 10:21:20 
UTC ---
Created attachment 22545
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22545
preprocessed file

The problem is in -fprofile-use. Corresponding .gcda file is not needed to
trigger the problem, so:

~/gcc-build/prev-gcc/cc1 -O2 verify-impl.i

produces correct object file, while

~/gcc-build/prev-gcc/cc1 -O2 -fprofile-use verify-impl.i

produces wrong object file (please note that execution counts are estimated).

You can conveniently grep for "printf" in the dumps, since I have left a
debugging printf in the loop:

        for (i = 0; i < npairs; ++i)
          {
        jint key = get_int ();
        if (i > 0 && key <= lastkey)
          verify_fail_pc ("lookupswitch pairs unsorted", vfr->start_PC);
        lastkey = key;
        push_jump (get_int ());
        printf ("i = %i\n", i);
          }

While the _.optimized dump without -fprofile-use has similar structure to the
original source:

--cut here--
<bb 164>:
  key_215 = get_int ();
  D.21282_216 = i_224 > 0;
  D.21283_217 = key_1 >= key_215;
  D.21284_218 = D.21283_217 & D.21282_216;
  if (D.21284_218 != 0)
    goto <bb 165>;
  else
    goto <bb 166>;

<bb 165>:
  vfr.14_219 = vfr;
  D.21287_220 = vfr.14_219->start_PC;
  verify_fail_pc ("lookupswitch pairs unsorted", D.21287_220);

<bb 166>:

<bb 167>:
  # key_1 = PHI <key_215(166), key_1193(291)>
  # i_1121 = PHI <i_224(166), 0(291)>
  D.21288_222 = get_int ();
  push_jump (D.21288_222);
  printf (&"i = %i\n"[0], i_1121);
  i_224 = i_1121 + 1;
  if (i_224 != npairs_212)
    goto <bb 164>;
  else
    goto <bb 168>;

<bb 168>:
  prephitmp.551_882 = vfr;
  prephitmp.551_882->PC = -1;
  goto <bb 21>;

<L228>:
  D.21290 = vfy_pop_type (10); [return slot optimization]
  check_return_type (D.21290);
  prephitmp.551_883 = vfr;
  prephitmp.551_883->PC = -1;
  goto <bb 21>;
--cut here--

And using -fprofile-use, it looks somehow strange:

--cut here--
<bb 309>:
  # prephitmp.1170_3787 = PHI <prephitmp.1170_3781(307),
prephitmp.1170_3786(308)>
  D.25872_2132 = prephitmp.1170_3787->current_state;
  merge_into (npc_2135, D.25872_2132);
  printf (&"i = %i\n"[0], 0);
  if (npairs_1969 > 1)
    goto <bb 290>;
  else
    goto <bb 310>;

<bb 310>:
  prephitmp.1288_882 = vfr;
  prephitmp.1288_882->PC = -1;
  goto <bb 12>;
--cut here--

And sure enough, npairs_1969 is nowhere incremented.