[Bug tree-optimization/56125] [4.7/4.8 Regression] -O2 -ffast-math generates bad code when dividing a double by the square of another double.

["jakub at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56125

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |jakub at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> 2013-01-28 10:27:19 UTC ---
Created attachment 29292
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=29292
gcc48-pr56125.patch

The bug is that the last two optimizations of pow (where 2c resp. 3c is a
non-zero integer) silently assume that the earlier optimizations already
handled the cases where already c (or 2c for the last optimization) is an
integer.
But that doesn't have to be the case, as shown by the testcase, the integer
optimization is guarded by c in [-1,2] or optimization for speed.

The && optimize_function_for_speed_p () (or should that be
optimize_insn_for_speed_p?, the pow folding is inconsistent in that, and I
don't see e.g. rtl_profile_for_bb being called during this pass to make it
accurate)
is up for discussions, say on x86_64
__attribute__((cold)) double
foo (double x, double n)
{
  double u = __builtin_pow (x, -1.5);
  return u;
}
with it we get smaller code:
        movsd   .LC0(%rip), %xmm1
        jmp     pow
compared to:
        sqrtsd  %xmm0, %xmm1
        mulsd   %xmm0, %xmm1
        movsd   .LC0(%rip), %xmm0
        divsd   %xmm1, %xmm0
without it, 7 bytes shorter.