xref: /freebsd/contrib/arm-optimized-routines/math/sincosf.c (revision 31914882fca502069810b9e9ddea4bcd8136a4f4)
1*31914882SAlex Richardson /*
2*31914882SAlex Richardson  * Single-precision sin/cos function.
3*31914882SAlex Richardson  *
4*31914882SAlex Richardson  * Copyright (c) 2018-2019, Arm Limited.
5*31914882SAlex Richardson  * SPDX-License-Identifier: MIT
6*31914882SAlex Richardson  */
7*31914882SAlex Richardson 
8*31914882SAlex Richardson #include <stdint.h>
9*31914882SAlex Richardson #include <math.h>
10*31914882SAlex Richardson #include "math_config.h"
11*31914882SAlex Richardson #include "sincosf.h"
12*31914882SAlex Richardson 
13*31914882SAlex Richardson /* Fast sincosf implementation.  Worst-case ULP is 0.5607, maximum relative
14*31914882SAlex Richardson    error is 0.5303 * 2^-23.  A single-step range reduction is used for
15*31914882SAlex Richardson    small values.  Large inputs have their range reduced using fast integer
16*31914882SAlex Richardson    arithmetic.  */
17*31914882SAlex Richardson void
18*31914882SAlex Richardson sincosf (float y, float *sinp, float *cosp)
19*31914882SAlex Richardson {
20*31914882SAlex Richardson   double x = y;
21*31914882SAlex Richardson   double s;
22*31914882SAlex Richardson   int n;
23*31914882SAlex Richardson   const sincos_t *p = &__sincosf_table[0];
24*31914882SAlex Richardson 
25*31914882SAlex Richardson   if (abstop12 (y) < abstop12 (pio4))
26*31914882SAlex Richardson     {
27*31914882SAlex Richardson       double x2 = x * x;
28*31914882SAlex Richardson 
29*31914882SAlex Richardson       if (unlikely (abstop12 (y) < abstop12 (0x1p-12f)))
30*31914882SAlex Richardson 	{
31*31914882SAlex Richardson 	  if (unlikely (abstop12 (y) < abstop12 (0x1p-126f)))
32*31914882SAlex Richardson 	    /* Force underflow for tiny y.  */
33*31914882SAlex Richardson 	    force_eval_float (x2);
34*31914882SAlex Richardson 	  *sinp = y;
35*31914882SAlex Richardson 	  *cosp = 1.0f;
36*31914882SAlex Richardson 	  return;
37*31914882SAlex Richardson 	}
38*31914882SAlex Richardson 
39*31914882SAlex Richardson       sincosf_poly (x, x2, p, 0, sinp, cosp);
40*31914882SAlex Richardson     }
41*31914882SAlex Richardson   else if (abstop12 (y) < abstop12 (120.0f))
42*31914882SAlex Richardson     {
43*31914882SAlex Richardson       x = reduce_fast (x, p, &n);
44*31914882SAlex Richardson 
45*31914882SAlex Richardson       /* Setup the signs for sin and cos.  */
46*31914882SAlex Richardson       s = p->sign[n & 3];
47*31914882SAlex Richardson 
48*31914882SAlex Richardson       if (n & 2)
49*31914882SAlex Richardson 	p = &__sincosf_table[1];
50*31914882SAlex Richardson 
51*31914882SAlex Richardson       sincosf_poly (x * s, x * x, p, n, sinp, cosp);
52*31914882SAlex Richardson     }
53*31914882SAlex Richardson   else if (likely (abstop12 (y) < abstop12 (INFINITY)))
54*31914882SAlex Richardson     {
55*31914882SAlex Richardson       uint32_t xi = asuint (y);
56*31914882SAlex Richardson       int sign = xi >> 31;
57*31914882SAlex Richardson 
58*31914882SAlex Richardson       x = reduce_large (xi, &n);
59*31914882SAlex Richardson 
60*31914882SAlex Richardson       /* Setup signs for sin and cos - include original sign.  */
61*31914882SAlex Richardson       s = p->sign[(n + sign) & 3];
62*31914882SAlex Richardson 
63*31914882SAlex Richardson       if ((n + sign) & 2)
64*31914882SAlex Richardson 	p = &__sincosf_table[1];
65*31914882SAlex Richardson 
66*31914882SAlex Richardson       sincosf_poly (x * s, x * x, p, n, sinp, cosp);
67*31914882SAlex Richardson     }
68*31914882SAlex Richardson   else
69*31914882SAlex Richardson     {
70*31914882SAlex Richardson       /* Return NaN if Inf or NaN for both sin and cos.  */
71*31914882SAlex Richardson       *sinp = *cosp = y - y;
72*31914882SAlex Richardson #if WANT_ERRNO
73*31914882SAlex Richardson       /* Needed to set errno for +-Inf, the add is a hack to work
74*31914882SAlex Richardson 	 around a gcc register allocation issue: just passing y
75*31914882SAlex Richardson 	 affects code generation in the fast path.  */
76*31914882SAlex Richardson       __math_invalidf (y + y);
77*31914882SAlex Richardson #endif
78*31914882SAlex Richardson     }
79*31914882SAlex Richardson }
80