[Bug middle-end/55481] [4.8 regression] -O2 generates a wrong-code infinite loop in C++Benchmark's simple_types_constant_folding int8 xor test

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55481



--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> 2012-11-27 
16:10:29 UTC ---

The following even more reduced testcase is not fixed by the patch pending

for PR35634.



typedef signed char int8_t;



#define SIZE 13



static inline int8_t do_shift() { return 31; }



int main(int argc, char** argv)

{

  int i;

  for(i = 0; i < 1000; ++i)

    {

      int8_t result = 0;

      int n;

      for (n = 0; n < SIZE; ++n)

        result += do_shift();

      int8_t temp = (int8_t)((int8_t)SIZE * ((int8_t)(23) | (int8_t)(10)));

      if (result != temp)

        __builtin_abort ();

    }

  return 0;

}





With the C++ frontend we correctly do



        D.2194 = do_shift ();

        D.2207 = (int) result;

        D.2208 = (int) D.2194;

        D.2209 = D.2207 + D.2208;

        result = (int8_t) D.2209;



anyway, but then IVOPTs comes around and changes it:



   <bb 4>:

   # result_17 = PHI <result_8(3), 0(7)>

-  # n_18 = PHI <n_9(3), 0(7)>

-  # ivtmp_12 = PHI <ivtmp_13(3), 13(7)>

   _6 = (int) result_17;

   _7 = _6 + 31;

-  result_8 = (int8_t) _7;

-  n_9 = n_18 + 1;

-  ivtmp_13 = ivtmp_12 - 1;

-  if (ivtmp_13 != 0)

+  _1 = result_17 + 31;

+  result_8 = _1;

+  if (result_8 != -109)

     goto <bb 3>;



performing the computation in int8_t again which causes VRP to "miscompile"

the test.