xref: /freebsd/lib/msun/src/s_fma.c (revision c697fb7f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2005-2011 David Schultz <das@FreeBSD.ORG>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 #include <fenv.h>
33 #include <float.h>
34 #include <math.h>
35 
36 #include "math_private.h"
37 
38 /*
39  * A struct dd represents a floating-point number with twice the precision
40  * of a double.  We maintain the invariant that "hi" stores the 53 high-order
41  * bits of the result.
42  */
43 struct dd {
44 	double hi;
45 	double lo;
46 };
47 
48 /*
49  * Compute a+b exactly, returning the exact result in a struct dd.  We assume
50  * that both a and b are finite, but make no assumptions about their relative
51  * magnitudes.
52  */
53 static inline struct dd
54 dd_add(double a, double b)
55 {
56 	struct dd ret;
57 	double s;
58 
59 	ret.hi = a + b;
60 	s = ret.hi - a;
61 	ret.lo = (a - (ret.hi - s)) + (b - s);
62 	return (ret);
63 }
64 
65 /*
66  * Compute a+b, with a small tweak:  The least significant bit of the
67  * result is adjusted into a sticky bit summarizing all the bits that
68  * were lost to rounding.  This adjustment negates the effects of double
69  * rounding when the result is added to another number with a higher
70  * exponent.  For an explanation of round and sticky bits, see any reference
71  * on FPU design, e.g.,
72  *
73  *     J. Coonen.  An Implementation Guide to a Proposed Standard for
74  *     Floating-Point Arithmetic.  Computer, vol. 13, no. 1, Jan 1980.
75  */
76 static inline double
77 add_adjusted(double a, double b)
78 {
79 	struct dd sum;
80 	uint64_t hibits, lobits;
81 
82 	sum = dd_add(a, b);
83 	if (sum.lo != 0) {
84 		EXTRACT_WORD64(hibits, sum.hi);
85 		if ((hibits & 1) == 0) {
86 			/* hibits += (int)copysign(1.0, sum.hi * sum.lo) */
87 			EXTRACT_WORD64(lobits, sum.lo);
88 			hibits += 1 - ((hibits ^ lobits) >> 62);
89 			INSERT_WORD64(sum.hi, hibits);
90 		}
91 	}
92 	return (sum.hi);
93 }
94 
95 /*
96  * Compute ldexp(a+b, scale) with a single rounding error. It is assumed
97  * that the result will be subnormal, and care is taken to ensure that
98  * double rounding does not occur.
99  */
100 static inline double
101 add_and_denormalize(double a, double b, int scale)
102 {
103 	struct dd sum;
104 	uint64_t hibits, lobits;
105 	int bits_lost;
106 
107 	sum = dd_add(a, b);
108 
109 	/*
110 	 * If we are losing at least two bits of accuracy to denormalization,
111 	 * then the first lost bit becomes a round bit, and we adjust the
112 	 * lowest bit of sum.hi to make it a sticky bit summarizing all the
113 	 * bits in sum.lo. With the sticky bit adjusted, the hardware will
114 	 * break any ties in the correct direction.
115 	 *
116 	 * If we are losing only one bit to denormalization, however, we must
117 	 * break the ties manually.
118 	 */
119 	if (sum.lo != 0) {
120 		EXTRACT_WORD64(hibits, sum.hi);
121 		bits_lost = -((int)(hibits >> 52) & 0x7ff) - scale + 1;
122 		if ((bits_lost != 1) ^ (int)(hibits & 1)) {
123 			/* hibits += (int)copysign(1.0, sum.hi * sum.lo) */
124 			EXTRACT_WORD64(lobits, sum.lo);
125 			hibits += 1 - (((hibits ^ lobits) >> 62) & 2);
126 			INSERT_WORD64(sum.hi, hibits);
127 		}
128 	}
129 	return (ldexp(sum.hi, scale));
130 }
131 
132 /*
133  * Compute a*b exactly, returning the exact result in a struct dd.  We assume
134  * that both a and b are normalized, so no underflow or overflow will occur.
135  * The current rounding mode must be round-to-nearest.
136  */
137 static inline struct dd
138 dd_mul(double a, double b)
139 {
140 	static const double split = 0x1p27 + 1.0;
141 	struct dd ret;
142 	double ha, hb, la, lb, p, q;
143 
144 	p = a * split;
145 	ha = a - p;
146 	ha += p;
147 	la = a - ha;
148 
149 	p = b * split;
150 	hb = b - p;
151 	hb += p;
152 	lb = b - hb;
153 
154 	p = ha * hb;
155 	q = ha * lb + la * hb;
156 
157 	ret.hi = p + q;
158 	ret.lo = p - ret.hi + q + la * lb;
159 	return (ret);
160 }
161 
162 /*
163  * Fused multiply-add: Compute x * y + z with a single rounding error.
164  *
165  * We use scaling to avoid overflow/underflow, along with the
166  * canonical precision-doubling technique adapted from:
167  *
168  *	Dekker, T.  A Floating-Point Technique for Extending the
169  *	Available Precision.  Numer. Math. 18, 224-242 (1971).
170  *
171  * This algorithm is sensitive to the rounding precision.  FPUs such
172  * as the i387 must be set in double-precision mode if variables are
173  * to be stored in FP registers in order to avoid incorrect results.
174  * This is the default on FreeBSD, but not on many other systems.
175  *
176  * Hardware instructions should be used on architectures that support it,
177  * since this implementation will likely be several times slower.
178  */
179 double
180 fma(double x, double y, double z)
181 {
182 	double xs, ys, zs, adj;
183 	struct dd xy, r;
184 	int oround;
185 	int ex, ey, ez;
186 	int spread;
187 
188 	/*
189 	 * Handle special cases. The order of operations and the particular
190 	 * return values here are crucial in handling special cases involving
191 	 * infinities, NaNs, overflows, and signed zeroes correctly.
192 	 */
193 	if (x == 0.0 || y == 0.0)
194 		return (x * y + z);
195 	if (z == 0.0)
196 		return (x * y);
197 	if (!isfinite(x) || !isfinite(y))
198 		return (x * y + z);
199 	if (!isfinite(z))
200 		return (z);
201 
202 	xs = frexp(x, &ex);
203 	ys = frexp(y, &ey);
204 	zs = frexp(z, &ez);
205 	oround = fegetround();
206 	spread = ex + ey - ez;
207 
208 	/*
209 	 * If x * y and z are many orders of magnitude apart, the scaling
210 	 * will overflow, so we handle these cases specially.  Rounding
211 	 * modes other than FE_TONEAREST are painful.
212 	 */
213 	if (spread < -DBL_MANT_DIG) {
214 		feraiseexcept(FE_INEXACT);
215 		if (!isnormal(z))
216 			feraiseexcept(FE_UNDERFLOW);
217 		switch (oround) {
218 		case FE_TONEAREST:
219 			return (z);
220 		case FE_TOWARDZERO:
221 			if (x > 0.0 ^ y < 0.0 ^ z < 0.0)
222 				return (z);
223 			else
224 				return (nextafter(z, 0));
225 		case FE_DOWNWARD:
226 			if (x > 0.0 ^ y < 0.0)
227 				return (z);
228 			else
229 				return (nextafter(z, -INFINITY));
230 		default:	/* FE_UPWARD */
231 			if (x > 0.0 ^ y < 0.0)
232 				return (nextafter(z, INFINITY));
233 			else
234 				return (z);
235 		}
236 	}
237 	if (spread <= DBL_MANT_DIG * 2)
238 		zs = ldexp(zs, -spread);
239 	else
240 		zs = copysign(DBL_MIN, zs);
241 
242 	fesetround(FE_TONEAREST);
243 	/* work around clang bug 8100 */
244 	volatile double vxs = xs;
245 
246 	/*
247 	 * Basic approach for round-to-nearest:
248 	 *
249 	 *     (xy.hi, xy.lo) = x * y		(exact)
250 	 *     (r.hi, r.lo)   = xy.hi + z	(exact)
251 	 *     adj = xy.lo + r.lo		(inexact; low bit is sticky)
252 	 *     result = r.hi + adj		(correctly rounded)
253 	 */
254 	xy = dd_mul(vxs, ys);
255 	r = dd_add(xy.hi, zs);
256 
257 	spread = ex + ey;
258 
259 	if (r.hi == 0.0) {
260 		/*
261 		 * When the addends cancel to 0, ensure that the result has
262 		 * the correct sign.
263 		 */
264 		fesetround(oround);
265 		volatile double vzs = zs; /* XXX gcc CSE bug workaround */
266 		return (xy.hi + vzs + ldexp(xy.lo, spread));
267 	}
268 
269 	if (oround != FE_TONEAREST) {
270 		/*
271 		 * There is no need to worry about double rounding in directed
272 		 * rounding modes.
273 		 */
274 		fesetround(oround);
275 		/* work around clang bug 8100 */
276 		volatile double vrlo = r.lo;
277 		adj = vrlo + xy.lo;
278 		return (ldexp(r.hi + adj, spread));
279 	}
280 
281 	adj = add_adjusted(r.lo, xy.lo);
282 	if (spread + ilogb(r.hi) > -1023)
283 		return (ldexp(r.hi + adj, spread));
284 	else
285 		return (add_and_denormalize(r.hi, adj, spread));
286 }
287 
288 #if (LDBL_MANT_DIG == 53)
289 __weak_reference(fma, fmal);
290 #endif
291