[Bug tree-optimization/55723] SLP vectorization vs loop: SLP more efficient: loop vectorization inefficient in presence of multiple "blends"

[vincenzo.innocente at cern dot ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55723

vincenzo Innocente <vincenzo.innocente at cern dot ch> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|SLP vectorization vs loop:  |SLP vectorization vs loop:
                   |SLP more efficient!         |SLP more efficient: loop
                   |                            |vectorization inefficient
                   |                            |in presence of multiple
                   |                            |"blends"

--- Comment #1 from vincenzo Innocente <vincenzo.innocente at cern dot ch> 2012-12-17 19:25:37 UTC ---
moving the second blending  before the polynomial makes the two loops to
produce almost identical code
This is not always possible though.
Bug in the loop optimizer?

template<typename Float>
inline
Float atan(Float t) {
  constexpr float PIO4F = 0.7853981633974483096f;
  constexpr Float zero = {0};
  Float z= (t > 0.4142135623730950f) ? (t-1.0f)/(t+1.0f) : t;
  Float ret = ( t > 0.4142135623730950f ) ? zero+PIO4F : zero;

  Float z2 = z * z;
  ret +=
    ((( 8.05374449538e-2f * z2
    - 1.38776856032E-1f) * z2
      + 1.99777106478E-1f) * z2
     - 3.33329491539E-1f) * z2 * z
    + z;

  return ret;
}