[Bug bootstrap/49786] [4.7 Regression] bootstrap failed with bootstrap-profiled

[ian at airs dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49786

Ian Lance Taylor <ian at airs dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2011.07.20 22:03:04
         AssignedTo|unassigned at gcc dot       |jamborm at gcc dot gnu.org
                   |gnu.org                     |
     Ever Confirmed|0                           |1

--- Comment #12 from Ian Lance Taylor <ian at airs dot com> 2011-07-20 22:03:04 UTC ---
I was able to recreate the problem.  The count is going negative at line 1926
of ipa-cp.c, the first line setting cs->count here:

  for (cs = new_node->callees; cs ; cs = cs->next_callee)
    if (cs->frequency)
      cs->count = cs->count * new_sum / orig_node_count;
    else
      cs->count = 0;

In the test case, cs->count, new_sum, and orig_node_count all == 3870557758. 
Computing 3870557758 * 3870557758 overflows to a negative number.  Dividing by
3870557758 leaves the number still negative.  This then leads to the failure.

For reference, profile_info->sum_max == 2619109064700.

I don't know if these numbers are plausible.

I think that this code probably needs to be written along the lines of
scale_bbs_frequencies_gcov_type.  I will leave this for Martin.