1 /* @(#)k_rem_pio2.c 5.1 93/09/24 */
2 /*
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  *
6  * Developed at SunPro, a Sun Microsystems, Inc. business.
7  * Permission to use, copy, modify, and distribute this
8  * software is freely granted, provided that this notice
9  * is preserved.
10  * ====================================================
11  */
12 
13 #if defined(LIBM_SCCS) && !defined(lint)
14 static char rcsid[] = "$NetBSD: k_rem_pio2.c,v 1.7 1995/05/10 20:46:25 jtc Exp $";
15 #endif
16 
17 /*
18  * __kernel_rem_pio2(x,y,e0,nx,prec,ipio2)
19  * double x[],y[]; int e0,nx,prec; int ipio2[];
20  *
21  * __kernel_rem_pio2 return the last three digits of N with
22  *		y = x - N*pi/2
23  * so that |y| < pi/2.
24  *
25  * The method is to compute the integer (mod 8) and fraction parts of
26  * (2/pi)*x without doing the full multiplication. In general we
27  * skip the part of the product that are known to be a huge integer (
28  * more accurately, = 0 mod 8 ). Thus the number of operations are
29  * independent of the exponent of the input.
30  *
31  * (2/pi) is represented by an array of 24-bit integers in ipio2[].
32  *
33  * Input parameters:
34  * 	x[]	The input value (must be positive) is broken into nx
35  *		pieces of 24-bit integers in double precision format.
36  *		x[i] will be the i-th 24 bit of x. The scaled exponent
37  *		of x[0] is given in input parameter e0 (i.e., x[0]*2^e0
38  *		match x's up to 24 bits.
39  *
40  *		Example of breaking a double positive z into x[0]+x[1]+x[2]:
41  *			e0 = ilogb(z)-23
42  *			z  = scalbn(z,-e0)
43  *		for i = 0,1,2
44  *			x[i] = floor(z)
45  *			z    = (z-x[i])*2**24
46  *
47  *
48  *	y[]	output result in an array of double precision numbers.
49  *		The dimension of y[] is:
50  *			24-bit  precision	1
51  *			53-bit  precision	2
52  *			64-bit  precision	2
53  *			113-bit precision	3
54  *		The actual value is the sum of them. Thus for 113-bit
55  *		precision, one may have to do something like:
56  *
57  *		long double t,w,r_head, r_tail;
58  *		t = (long double)y[2] + (long double)y[1];
59  *		w = (long double)y[0];
60  *		r_head = t+w;
61  *		r_tail = w - (r_head - t);
62  *
63  *	e0	The exponent of x[0]
64  *
65  *	nx	dimension of x[]
66  *
67  *  	prec	an integer indicating the precision:
68  *			0	24  bits (single)
69  *			1	53  bits (double)
70  *			2	64  bits (extended)
71  *			3	113 bits (quad)
72  *
73  *	ipio2[]
74  *		integer array, contains the (24*i)-th to (24*i+23)-th
75  *		bit of 2/pi after binary point. The corresponding
76  *		floating value is
77  *
78  *			ipio2[i] * 2^(-24(i+1)).
79  *
80  * External function:
81  *	double scalbn(), floor();
82  *
83  *
84  * Here is the description of some local variables:
85  *
86  * 	jk	jk+1 is the initial number of terms of ipio2[] needed
87  *		in the computation. The recommended value is 2,3,4,
88  *		6 for single, double, extended,and quad.
89  *
90  * 	jz	local integer variable indicating the number of
91  *		terms of ipio2[] used.
92  *
93  *	jx	nx - 1
94  *
95  *	jv	index for pointing to the suitable ipio2[] for the
96  *		computation. In general, we want
97  *			( 2^e0*x[0] * ipio2[jv-1]*2^(-24jv) )/8
98  *		is an integer. Thus
99  *			e0-3-24*jv >= 0 or (e0-3)/24 >= jv
100  *		Hence jv = max(0,(e0-3)/24).
101  *
102  *	jp	jp+1 is the number of terms in PIo2[] needed, jp = jk.
103  *
104  * 	q[]	double array with integral value, representing the
105  *		24-bits chunk of the product of x and 2/pi.
106  *
107  *	q0	the corresponding exponent of q[0]. Note that the
108  *		exponent for q[i] would be q0-24*i.
109  *
110  *	PIo2[]	double precision array, obtained by cutting pi/2
111  *		into 24 bits chunks.
112  *
113  *	f[]	ipio2[] in floating point
114  *
115  *	iq[]	integer array by breaking up q[] in 24-bits chunk.
116  *
117  *	fq[]	final product of x*(2/pi) in fq[0],..,fq[jk]
118  *
119  *	ih	integer. If >0 it indicates q[] is >= 0.5, hence
120  *		it also indicates the *sign* of the result.
121  *
122  */
123 
124 
125 /*
126  * Constants:
127  * The hexadecimal values are the intended ones for the following
128  * constants. The decimal values may be used, provided that the
129  * compiler will convert from decimal to binary accurately enough
130  * to produce the hexadecimal values shown.
131  */
132 
133 #include <math.h>
134 #include <math-narrow-eval.h>
135 #include <math_private.h>
136 #include <libc-diag.h>
137 
138 static const int init_jk[] = {2,3,4,6}; /* initial value for jk */
139 
140 static const double PIo2[] = {
141   1.57079625129699707031e+00, /* 0x3FF921FB, 0x40000000 */
142   7.54978941586159635335e-08, /* 0x3E74442D, 0x00000000 */
143   5.39030252995776476554e-15, /* 0x3CF84698, 0x80000000 */
144   3.28200341580791294123e-22, /* 0x3B78CC51, 0x60000000 */
145   1.27065575308067607349e-29, /* 0x39F01B83, 0x80000000 */
146   1.22933308981111328932e-36, /* 0x387A2520, 0x40000000 */
147   2.73370053816464559624e-44, /* 0x36E38222, 0x80000000 */
148   2.16741683877804819444e-51, /* 0x3569F31D, 0x00000000 */
149 };
150 
151 static const double
152   zero   = 0.0,
153   one    = 1.0,
154   two24  = 1.67772160000000000000e+07, /* 0x41700000, 0x00000000 */
155   twon24 = 5.96046447753906250000e-08; /* 0x3E700000, 0x00000000 */
156 
157 int
__kernel_rem_pio2(double * x,double * y,int e0,int nx,int prec,const int32_t * ipio2)158 __kernel_rem_pio2 (double *x, double *y, int e0, int nx, int prec,
159                    const int32_t *ipio2)
160 {
161   int32_t jz, jx, jv, jp, jk, carry, n, iq[20], i, j, k, m, q0, ih;
162   double z, fw, f[20], fq[20], q[20];
163 
164   /* initialize jk*/
165   jk = init_jk[prec];
166   jp = jk;
167 
168   /* determine jx,jv,q0, note that 3>q0 */
169   jx = nx - 1;
170   jv = (e0 - 3) / 24; if (jv < 0)
171     jv = 0;
172   q0 = e0 - 24 * (jv + 1);
173 
174   /* set up f[0] to f[jx+jk] where f[jx+jk] = ipio2[jv+jk] */
175   j = jv - jx; m = jx + jk;
176   for (i = 0; i <= m; i++, j++)
177     f[i] = (j < 0) ? zero : (double) ipio2[j];
178 
179   /* compute q[0],q[1],...q[jk] */
180   for (i = 0; i <= jk; i++)
181     {
182       for (j = 0, fw = 0.0; j <= jx; j++)
183 	fw += x[j] * f[jx + i - j];
184       q[i] = fw;
185     }
186 
187   jz = jk;
188 recompute:
189   /* distill q[] into iq[] reversingly */
190   for (i = 0, j = jz, z = q[jz]; j > 0; i++, j--)
191     {
192       fw = (double) ((int32_t) (twon24 * z));
193       iq[i] = (int32_t) (z - two24 * fw);
194       z = q[j - 1] + fw;
195     }
196 
197   /* compute n */
198   z = __scalbn (z, q0);                 /* actual value of z */
199   z -= 8.0 * floor (z * 0.125);               /* trim off integer >= 8 */
200   n = (int32_t) z;
201   z -= (double) n;
202   ih = 0;
203   if (q0 > 0)           /* need iq[jz-1] to determine n */
204     {
205       i = (iq[jz - 1] >> (24 - q0)); n += i;
206       iq[jz - 1] -= i << (24 - q0);
207       ih = iq[jz - 1] >> (23 - q0);
208     }
209   else if (q0 == 0)
210     ih = iq[jz - 1] >> 23;
211   else if (z >= 0.5)
212     ih = 2;
213 
214   if (ih > 0)           /* q > 0.5 */
215     {
216       n += 1; carry = 0;
217       for (i = 0; i < jz; i++)          /* compute 1-q */
218 	{
219 	  j = iq[i];
220 	  if (carry == 0)
221 	    {
222 	      if (j != 0)
223 		{
224 		  carry = 1; iq[i] = 0x1000000 - j;
225 		}
226 	    }
227 	  else
228 	    iq[i] = 0xffffff - j;
229 	}
230       if (q0 > 0)               /* rare case: chance is 1 in 12 */
231 	{
232 	  switch (q0)
233 	    {
234 	    case 1:
235 	      iq[jz - 1] &= 0x7fffff; break;
236 	    case 2:
237 	      iq[jz - 1] &= 0x3fffff; break;
238 	    }
239 	}
240       if (ih == 2)
241 	{
242 	  z = one - z;
243 	  if (carry != 0)
244 	    z -= __scalbn (one, q0);
245 	}
246     }
247 
248   /* check if recomputation is needed */
249   if (z == zero)
250     {
251       j = 0;
252       for (i = jz - 1; i >= jk; i--)
253 	j |= iq[i];
254       if (j == 0)      /* need recomputation */
255 	{
256 	  /* On s390x gcc 6.1 -O3 produces the warning "array subscript is below
257 	     array bounds [-Werror=array-bounds]".  Only __ieee754_rem_pio2l
258 	     calls __kernel_rem_pio2 for normal numbers and |x| > pi/4 in case
259 	     of ldbl-96 and |x| > 3pi/4 in case of ldbl-128[ibm].
260 	     Thus x can't be zero and ipio2 is not zero, too.  Thus not all iq[]
261 	     values can't be zero.  */
262 	  DIAG_PUSH_NEEDS_COMMENT;
263 	  DIAG_IGNORE_NEEDS_COMMENT (6.1, "-Warray-bounds");
264 	  for (k = 1; iq[jk - k] == 0; k++)
265 	    ;                               /* k = no. of terms needed */
266 	  DIAG_POP_NEEDS_COMMENT;
267 
268 	  for (i = jz + 1; i <= jz + k; i++) /* add q[jz+1] to q[jz+k] */
269 	    {
270 	      f[jx + i] = (double) ipio2[jv + i];
271 	      for (j = 0, fw = 0.0; j <= jx; j++)
272 		fw += x[j] * f[jx + i - j];
273 	      q[i] = fw;
274 	    }
275 	  jz += k;
276 	  goto recompute;
277 	}
278     }
279 
280   /* chop off zero terms */
281   if (z == 0.0)
282     {
283       jz -= 1; q0 -= 24;
284       while (iq[jz] == 0)
285 	{
286 	  jz--; q0 -= 24;
287 	}
288     }
289   else           /* break z into 24-bit if necessary */
290     {
291       z = __scalbn (z, -q0);
292       if (z >= two24)
293 	{
294 	  fw = (double) ((int32_t) (twon24 * z));
295 	  iq[jz] = (int32_t) (z - two24 * fw);
296 	  jz += 1; q0 += 24;
297 	  iq[jz] = (int32_t) fw;
298 	}
299       else
300 	iq[jz] = (int32_t) z;
301     }
302 
303   /* convert integer "bit" chunk to floating-point value */
304   fw = __scalbn (one, q0);
305   for (i = jz; i >= 0; i--)
306     {
307       q[i] = fw * (double) iq[i]; fw *= twon24;
308     }
309 
310   /* compute PIo2[0,...,jp]*q[jz,...,0] */
311   for (i = jz; i >= 0; i--)
312     {
313       for (fw = 0.0, k = 0; k <= jp && k <= jz - i; k++)
314 	fw += PIo2[k] * q[i + k];
315       fq[jz - i] = fw;
316     }
317 
318   /* compress fq[] into y[] */
319   switch (prec)
320     {
321     case 0:
322       fw = 0.0;
323       for (i = jz; i >= 0; i--)
324 	fw += fq[i];
325       y[0] = (ih == 0) ? fw : -fw;
326       break;
327     case 1:
328     case 2:;
329       double fv = 0.0;
330       for (i = jz; i >= 0; i--)
331 	fv = math_narrow_eval (fv + fq[i]);
332       y[0] = (ih == 0) ? fv : -fv;
333       /* GCC mainline (to be GCC 9), as of 2018-05-22 on i686, warns
334 	 that fq[0] may be used uninitialized.  This is not possible
335 	 because jz is always nonnegative when the above loop
336 	 initializing fq is executed, because the result is never zero
337 	 to full precision (this function is not called for zero
338 	 arguments).  */
339       DIAG_PUSH_NEEDS_COMMENT;
340       DIAG_IGNORE_NEEDS_COMMENT (9, "-Wmaybe-uninitialized");
341       fv = math_narrow_eval (fq[0] - fv);
342       DIAG_POP_NEEDS_COMMENT;
343       for (i = 1; i <= jz; i++)
344 	fv = math_narrow_eval (fv + fq[i]);
345       y[1] = (ih == 0) ? fv : -fv;
346       break;
347     case 3:             /* painful */
348       for (i = jz; i > 0; i--)
349 	{
350 	  double fv = math_narrow_eval (fq[i - 1] + fq[i]);
351 	  fq[i] += fq[i - 1] - fv;
352 	  fq[i - 1] = fv;
353 	}
354       for (i = jz; i > 1; i--)
355 	{
356 	  double fv = math_narrow_eval (fq[i - 1] + fq[i]);
357 	  fq[i] += fq[i - 1] - fv;
358 	  fq[i - 1] = fv;
359 	}
360       for (fw = 0.0, i = jz; i >= 2; i--)
361 	fw += fq[i];
362       if (ih == 0)
363 	{
364 	  y[0] = fq[0]; y[1] = fq[1]; y[2] = fw;
365 	}
366       else
367 	{
368 	  y[0] = -fq[0]; y[1] = -fq[1]; y[2] = -fw;
369 	}
370     }
371   return n & 7;
372 }
373