[Bug tree-optimization/46675] [4.6 Regression] profiledbootstrap failed

[rguenth at gcc dot gnu.org]

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

--- Comment #9 from Richard Guenther <rguenth at gcc dot gnu.org> 2010-11-27 
20:12:44 UTC ---
The bug seems to be:

Visiting PHI node: i_1950 = PHI <i_1028(291), i_224(510)>

    Argument #0 (291 -> 503 executable)
        i_1028
        Value: [0, 0]

    Argument #1 (510 -> 503 executable)
        i_224
        Value: [1, 2]
(analyze_scalar_evolution
  (loop_nb = 16)
  (scalar = i_1950)
(get_scalar_evolution
  (scalar = i_1950)
  (scalar_evolution = {0, +, 1}_16))
(set_scalar_evolution
  instantiated_below = 291
  (scalar = i_1950)
  (scalar_evolution = {0, +, 1}_16))
)
(instantiate_scev
  (instantiate_below = 291)
  (evolution_loop = 16)
  (chrec = {0, +, 1}_16)
  (res = {0, +, 1}_16))
i_1950: loop information indicates does not overflow
(analyze_scalar_evolution
  (loop_nb = 16)
  (scalar = i_1950)
(get_scalar_evolution
  (scalar = i_1950)
  (scalar_evolution = {0, +, 1}_16))
(set_scalar_evolution
  instantiated_below = 291
  (scalar = i_1950)
  (scalar_evolution = {0, +, 1}_16))
)
(instantiate_scev
  (instantiate_below = 291)
  (evolution_loop = 16)
  (chrec = {0, +, 1}_16)
  (res = {0, +, 1}_16))
Found new range for i_1950: [0, 0]

while the previous range was i_1950: [0, 1]

This means that number of iteration analysis has gone wrong.

Analyzing # of iterations of loop 16
  exit condition [1, + , 1](no_overflow) < n_1968 - -2147483648
  bounds on difference of bases: 0 ... -2
  result:
    # of iterations (unsigned int) n_1968 + 2147483647, bounded by 0
Statement (exit)if (i_224 < npairs_1969)
 is executed at most (unsigned int) n_1968 + 2147483647 (bounded by 0) + 1
times in loop 16.

the "bounded by 0" is why things go wrong.  The original statement defining
the exit condition bound is

  npairs_1969 = n_1968 - -2147483648;


Testcase:

volatile int j;
int __attribute__((noinline))
foo(int n)
{
  int i = 0, npairs;
  npairs = n - (- __INT_MAX__ - 1);
  if (npairs > 0)
    {
      do
        {
          ++j;
          i = i + 1;
        }
      while (i < npairs);
    }
  return 0;
}
extern void abort (void);
int main()
{
  j = 0;
  foo (- __INT_MAX__ - 1 + 5);
  if (j != 5)
    abort ();
  return 0;
}

is optimized to an endless loop by the first VRP pass.