xref: /freebsd/lib/msun/src/s_fma.c (revision 2008043f386721d58158e37e0d7e50df8095942d)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include <fenv.h>
31 #include <float.h>
32 #include <math.h>
33 
34 #include "math_private.h"
35 
36 #ifdef USE_BUILTIN_FMA
37 double
38 fma(double x, double y, double z)
39 {
40 	return (__builtin_fma(x, y, z));
41 }
42 #else
43 /*
44  * A struct dd represents a floating-point number with twice the precision
45  * of a double.  We maintain the invariant that "hi" stores the 53 high-order
46  * bits of the result.
47  */
48 struct dd {
49 	double hi;
50 	double lo;
51 };
52 
53 /*
54  * Compute a+b exactly, returning the exact result in a struct dd.  We assume
55  * that both a and b are finite, but make no assumptions about their relative
56  * magnitudes.
57  */
58 static inline struct dd
59 dd_add(double a, double b)
60 {
61 	struct dd ret;
62 	double s;
63 
64 	ret.hi = a + b;
65 	s = ret.hi - a;
66 	ret.lo = (a - (ret.hi - s)) + (b - s);
67 	return (ret);
68 }
69 
70 /*
71  * Compute a+b, with a small tweak:  The least significant bit of the
72  * result is adjusted into a sticky bit summarizing all the bits that
73  * were lost to rounding.  This adjustment negates the effects of double
74  * rounding when the result is added to another number with a higher
75  * exponent.  For an explanation of round and sticky bits, see any reference
76  * on FPU design, e.g.,
77  *
78  *     J. Coonen.  An Implementation Guide to a Proposed Standard for
79  *     Floating-Point Arithmetic.  Computer, vol. 13, no. 1, Jan 1980.
80  */
81 static inline double
82 add_adjusted(double a, double b)
83 {
84 	struct dd sum;
85 	uint64_t hibits, lobits;
86 
87 	sum = dd_add(a, b);
88 	if (sum.lo != 0) {
89 		EXTRACT_WORD64(hibits, sum.hi);
90 		if ((hibits & 1) == 0) {
91 			/* hibits += (int)copysign(1.0, sum.hi * sum.lo) */
92 			EXTRACT_WORD64(lobits, sum.lo);
93 			hibits += 1 - ((hibits ^ lobits) >> 62);
94 			INSERT_WORD64(sum.hi, hibits);
95 		}
96 	}
97 	return (sum.hi);
98 }
99 
100 /*
101  * Compute ldexp(a+b, scale) with a single rounding error. It is assumed
102  * that the result will be subnormal, and care is taken to ensure that
103  * double rounding does not occur.
104  */
105 static inline double
106 add_and_denormalize(double a, double b, int scale)
107 {
108 	struct dd sum;
109 	uint64_t hibits, lobits;
110 	int bits_lost;
111 
112 	sum = dd_add(a, b);
113 
114 	/*
115 	 * If we are losing at least two bits of accuracy to denormalization,
116 	 * then the first lost bit becomes a round bit, and we adjust the
117 	 * lowest bit of sum.hi to make it a sticky bit summarizing all the
118 	 * bits in sum.lo. With the sticky bit adjusted, the hardware will
119 	 * break any ties in the correct direction.
120 	 *
121 	 * If we are losing only one bit to denormalization, however, we must
122 	 * break the ties manually.
123 	 */
124 	if (sum.lo != 0) {
125 		EXTRACT_WORD64(hibits, sum.hi);
126 		bits_lost = -((int)(hibits >> 52) & 0x7ff) - scale + 1;
127 		if ((bits_lost != 1) ^ (int)(hibits & 1)) {
128 			/* hibits += (int)copysign(1.0, sum.hi * sum.lo) */
129 			EXTRACT_WORD64(lobits, sum.lo);
130 			hibits += 1 - (((hibits ^ lobits) >> 62) & 2);
131 			INSERT_WORD64(sum.hi, hibits);
132 		}
133 	}
134 	return (ldexp(sum.hi, scale));
135 }
136 
137 /*
138  * Compute a*b exactly, returning the exact result in a struct dd.  We assume
139  * that both a and b are normalized, so no underflow or overflow will occur.
140  * The current rounding mode must be round-to-nearest.
141  */
142 static inline struct dd
143 dd_mul(double a, double b)
144 {
145 	static const double split = 0x1p27 + 1.0;
146 	struct dd ret;
147 	double ha, hb, la, lb, p, q;
148 
149 	p = a * split;
150 	ha = a - p;
151 	ha += p;
152 	la = a - ha;
153 
154 	p = b * split;
155 	hb = b - p;
156 	hb += p;
157 	lb = b - hb;
158 
159 	p = ha * hb;
160 	q = ha * lb + la * hb;
161 
162 	ret.hi = p + q;
163 	ret.lo = p - ret.hi + q + la * lb;
164 	return (ret);
165 }
166 
167 /*
168  * Fused multiply-add: Compute x * y + z with a single rounding error.
169  *
170  * We use scaling to avoid overflow/underflow, along with the
171  * canonical precision-doubling technique adapted from:
172  *
173  *	Dekker, T.  A Floating-Point Technique for Extending the
174  *	Available Precision.  Numer. Math. 18, 224-242 (1971).
175  *
176  * This algorithm is sensitive to the rounding precision.  FPUs such
177  * as the i387 must be set in double-precision mode if variables are
178  * to be stored in FP registers in order to avoid incorrect results.
179  * This is the default on FreeBSD, but not on many other systems.
180  *
181  * Hardware instructions should be used on architectures that support it,
182  * since this implementation will likely be several times slower.
183  */
184 double
185 fma(double x, double y, double z)
186 {
187 	double xs, ys, zs, adj;
188 	struct dd xy, r;
189 	int oround;
190 	int ex, ey, ez;
191 	int spread;
192 
193 	/*
194 	 * Handle special cases. The order of operations and the particular
195 	 * return values here are crucial in handling special cases involving
196 	 * infinities, NaNs, overflows, and signed zeroes correctly.
197 	 */
198 	if (x == 0.0 || y == 0.0)
199 		return (x * y + z);
200 	if (z == 0.0)
201 		return (x * y);
202 	if (!isfinite(x) || !isfinite(y))
203 		return (x * y + z);
204 	if (!isfinite(z))
205 		return (z);
206 
207 	xs = frexp(x, &ex);
208 	ys = frexp(y, &ey);
209 	zs = frexp(z, &ez);
210 	oround = fegetround();
211 	spread = ex + ey - ez;
212 
213 	/*
214 	 * If x * y and z are many orders of magnitude apart, the scaling
215 	 * will overflow, so we handle these cases specially.  Rounding
216 	 * modes other than FE_TONEAREST are painful.
217 	 */
218 	if (spread < -DBL_MANT_DIG) {
219 		feraiseexcept(FE_INEXACT);
220 		if (!isnormal(z))
221 			feraiseexcept(FE_UNDERFLOW);
222 		switch (oround) {
223 		case FE_TONEAREST:
224 			return (z);
225 		case FE_TOWARDZERO:
226 			if (x > 0.0 ^ y < 0.0 ^ z < 0.0)
227 				return (z);
228 			else
229 				return (nextafter(z, 0));
230 		case FE_DOWNWARD:
231 			if (x > 0.0 ^ y < 0.0)
232 				return (z);
233 			else
234 				return (nextafter(z, -INFINITY));
235 		default:	/* FE_UPWARD */
236 			if (x > 0.0 ^ y < 0.0)
237 				return (nextafter(z, INFINITY));
238 			else
239 				return (z);
240 		}
241 	}
242 	if (spread <= DBL_MANT_DIG * 2)
243 		zs = ldexp(zs, -spread);
244 	else
245 		zs = copysign(DBL_MIN, zs);
246 
247 	fesetround(FE_TONEAREST);
248 	/* work around clang bug 8100 */
249 	volatile double vxs = xs;
250 
251 	/*
252 	 * Basic approach for round-to-nearest:
253 	 *
254 	 *     (xy.hi, xy.lo) = x * y		(exact)
255 	 *     (r.hi, r.lo)   = xy.hi + z	(exact)
256 	 *     adj = xy.lo + r.lo		(inexact; low bit is sticky)
257 	 *     result = r.hi + adj		(correctly rounded)
258 	 */
259 	xy = dd_mul(vxs, ys);
260 	r = dd_add(xy.hi, zs);
261 
262 	spread = ex + ey;
263 
264 	if (r.hi == 0.0) {
265 		/*
266 		 * When the addends cancel to 0, ensure that the result has
267 		 * the correct sign.
268 		 */
269 		fesetround(oround);
270 		volatile double vzs = zs; /* XXX gcc CSE bug workaround */
271 		return (xy.hi + vzs + ldexp(xy.lo, spread));
272 	}
273 
274 	if (oround != FE_TONEAREST) {
275 		/*
276 		 * There is no need to worry about double rounding in directed
277 		 * rounding modes.
278 		 */
279 		fesetround(oround);
280 		/* work around clang bug 8100 */
281 		volatile double vrlo = r.lo;
282 		adj = vrlo + xy.lo;
283 		return (ldexp(r.hi + adj, spread));
284 	}
285 
286 	adj = add_adjusted(r.lo, xy.lo);
287 	if (spread + ilogb(r.hi) > -1023)
288 		return (ldexp(r.hi + adj, spread));
289 	else
290 		return (add_and_denormalize(r.hi, adj, spread));
291 }
292 #endif /* !USE_BUILTIN_FMA */
293 
294 #if (LDBL_MANT_DIG == 53)
295 __weak_reference(fma, fmal);
296 #endif
297