1 /* --------------------------------------------------------------  */
13 /* - Redistributions of source code must retain the above copyright*/
16 /* - Redistributions in binary form must reproduce the above       */
21 /* - Neither the name of IBM Corporation nor the names of its      */
29 /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE        */
39 /* --------------------------------------------------------------  */
56 static __inline vector float _recipf4(vector float a)  in _recipf4()  argument
58   /* This function has been designed to provide a  in _recipf4()
61    * single precision input <a>. This includes:  in _recipf4()
66    *    mode of the processor - truncated toward zero.  in _recipf4()
71    *    the input <a> has a zero exponent. A reciprocal  in _recipf4()
74    * 6) If a non-compliant IEEE result is produced, the  in _recipf4()
75    *    a DIFF exception is generated.  in _recipf4()
84   /* If a has a zero exponent, then set the divide by zero  in _recipf4()
87   (void)si_frest((qword)(a));  in _recipf4()
92   mant_a = spu_sel(a, one, exp_mask);  in _recipf4()
95    * followed by one iteration of the Newton-Raphson.   in _recipf4()
106   x1 = spu_sel(x0, x1, spu_cmpgt((vector signed int)(err), -1));  in _recipf4()
111    * values by computing a multiplier (mult) that will force the   in _recipf4()
115    * inoput <a> is a denorm or zero.  in _recipf4()
117   exp_a = spu_and((vector unsigned int)a, exp_mask);  in _recipf4()
121    * either the dividend <a> is a denorm/zero, or the computed exponent is   in _recipf4()
122    * less than or equal to a biased 0), force the multiplier to 0.0.  in _recipf4()
130   mult = spu_or(mult, (vector float)spu_rlmask(spu_cmpeq(exp_a, 0), -1));  in _recipf4()