[Bug rtl-optimization/55838] [4.6/4.7/4.8 Regression] ICE in extract_insn (unrecognizable insn) with -O -funroll-loops

[steven at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55838



Steven Bosscher <steven at gcc dot gnu.org> changed:



           What    |Removed                     |Added

----------------------------------------------------------------------------

             Status|ASSIGNED                    |NEW

         AssignedTo|steven at gcc dot gnu.org   |unassigned at gcc dot

                   |                            |gnu.org



--- Comment #9 from Steven Bosscher <steven at gcc dot gnu.org> 2013-01-01 
21:02:55 UTC ---

In general_operand, the following check fails:



951      if (CONST_INT_P (op)

952          && mode != VOIDmode

953          && trunc_int_for_mode (INTVAL (op), mode) != INTVAL (op))

954        return 0;



with op=(const_int 129) and trunc_int_for_mode (129, QImode)=-127.



This happens because trunc_int_for_mode sign-extends 129 (0b10000001):



65      if (width < HOST_BITS_PER_WIDE_INT)

66        {

67          HOST_WIDE_INT sign = 1;

68          sign <<= width - 1;

69          c &= (sign << 1) - 1;

70          c ^= sign;

71          c -= sign;

72        }



with width=8 for QImode and HOST_BITS_PER_WIDE_INT is 64 in my case. So:

sign=128 (0b10000000), c=129 (0b10000001)

c &= (sign << 1) - 1 = 129

c ^= sign = 1

c -= sign = 1 - 128 = -127 (0b11111111)



???



For someone with better RTL-fu than me...