1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
|
/*-
* Copyright (c) 2005-2011 David Schultz <das@FreeBSD.ORG>
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include <fenv.h>
#include <float.h>
#include <math.h>
#include "fpmath.h"
/*
* A struct dd represents a floating-point number with twice the precision
* of a long double. We maintain the invariant that "hi" stores the high-order
* bits of the result.
*/
struct dd {
long double hi;
long double lo;
};
/*
* Compute a+b exactly, returning the exact result in a struct dd. We assume
* that both a and b are finite, but make no assumptions about their relative
* magnitudes.
*/
static inline struct dd
dd_add(long double a, long double b)
{
struct dd ret;
long double s;
ret.hi = a + b;
s = ret.hi - a;
ret.lo = (a - (ret.hi - s)) + (b - s);
return (ret);
}
/*
* Compute a+b, with a small tweak: The least significant bit of the
* result is adjusted into a sticky bit summarizing all the bits that
* were lost to rounding. This adjustment negates the effects of double
* rounding when the result is added to another number with a higher
* exponent. For an explanation of round and sticky bits, see any reference
* on FPU design, e.g.,
*
* J. Coonen. An Implementation Guide to a Proposed Standard for
* Floating-Point Arithmetic. Computer, vol. 13, no. 1, Jan 1980.
*/
static inline long double
add_adjusted(long double a, long double b)
{
struct dd sum;
union IEEEl2bits u;
sum = dd_add(a, b);
if (sum.lo != 0) {
u.e = sum.hi;
if ((u.bits.manl & 1) == 0)
sum.hi = nextafterl(sum.hi, INFINITY * sum.lo);
}
return (sum.hi);
}
/*
* Compute ldexp(a+b, scale) with a single rounding error. It is assumed
* that the result will be subnormal, and care is taken to ensure that
* double rounding does not occur.
*/
static inline long double
add_and_denormalize(long double a, long double b, int scale)
{
struct dd sum;
int bits_lost;
union IEEEl2bits u;
sum = dd_add(a, b);
/*
* If we are losing at least two bits of accuracy to denormalization,
* then the first lost bit becomes a round bit, and we adjust the
* lowest bit of sum.hi to make it a sticky bit summarizing all the
* bits in sum.lo. With the sticky bit adjusted, the hardware will
* break any ties in the correct direction.
*
* If we are losing only one bit to denormalization, however, we must
* break the ties manually.
*/
if (sum.lo != 0) {
u.e = sum.hi;
bits_lost = -u.bits.exp - scale + 1;
if (bits_lost != 1 ^ (int)(u.bits.manl & 1))
sum.hi = nextafterl(sum.hi, INFINITY * sum.lo);
}
return (ldexp(sum.hi, scale));
}
/*
* Compute a*b exactly, returning the exact result in a struct dd. We assume
* that both a and b are normalized, so no underflow or overflow will occur.
* The current rounding mode must be round-to-nearest.
*/
static inline struct dd
dd_mul(long double a, long double b)
{
#if LDBL_MANT_DIG == 64
static const long double split = 0x1p32L + 1.0;
#elif LDBL_MANT_DIG == 113
static const long double split = 0x1p57L + 1.0;
#endif
struct dd ret;
long double ha, hb, la, lb, p, q;
p = a * split;
ha = a - p;
ha += p;
la = a - ha;
p = b * split;
hb = b - p;
hb += p;
lb = b - hb;
p = ha * hb;
q = ha * lb + la * hb;
ret.hi = p + q;
ret.lo = p - ret.hi + q + la * lb;
return (ret);
}
/*
* Fused multiply-add: Compute x * y + z with a single rounding error.
*
* We use scaling to avoid overflow/underflow, along with the
* canonical precision-doubling technique adapted from:
*
* Dekker, T. A Floating-Point Technique for Extending the
* Available Precision. Numer. Math. 18, 224-242 (1971).
*/
long double
fmal(long double x, long double y, long double z)
{
long double xs, ys, zs, adj;
struct dd xy, r;
int oround;
int ex, ey, ez;
int spread;
/*
* Handle special cases. The order of operations and the particular
* return values here are crucial in handling special cases involving
* infinities, NaNs, overflows, and signed zeroes correctly.
*/
if (x == 0.0 || y == 0.0)
return (x * y + z);
if (z == 0.0)
return (x * y);
if (!isfinite(x) || !isfinite(y))
return (x * y + z);
if (!isfinite(z))
return (z);
xs = frexpl(x, &ex);
ys = frexpl(y, &ey);
zs = frexpl(z, &ez);
oround = fegetround();
spread = ex + ey - ez;
/*
* If x * y and z are many orders of magnitude apart, the scaling
* will overflow, so we handle these cases specially. Rounding
* modes other than FE_TONEAREST are painful.
*/
if (spread < -LDBL_MANT_DIG) {
feraiseexcept(FE_INEXACT);
if (!isnormal(z))
feraiseexcept(FE_UNDERFLOW);
switch (oround) {
case FE_TONEAREST:
return (z);
case FE_TOWARDZERO:
if (x > 0.0 ^ y < 0.0 ^ z < 0.0)
return (z);
else
return (nextafterl(z, 0));
case FE_DOWNWARD:
if (x > 0.0 ^ y < 0.0)
return (z);
else
return (nextafterl(z, -INFINITY));
default: /* FE_UPWARD */
if (x > 0.0 ^ y < 0.0)
return (nextafterl(z, INFINITY));
else
return (z);
}
}
if (spread <= LDBL_MANT_DIG * 2)
zs = ldexpl(zs, -spread);
else
zs = copysignl(LDBL_MIN, zs);
fesetround(FE_TONEAREST);
/*
* Basic approach for round-to-nearest:
*
* (xy.hi, xy.lo) = x * y (exact)
* (r.hi, r.lo) = xy.hi + z (exact)
* adj = xy.lo + r.lo (inexact; low bit is sticky)
* result = r.hi + adj (correctly rounded)
*/
xy = dd_mul(xs, ys);
r = dd_add(xy.hi, zs);
spread = ex + ey;
if (r.hi == 0.0) {
/*
* When the addends cancel to 0, ensure that the result has
* the correct sign.
*/
fesetround(oround);
volatile long double vzs = zs; /* XXX gcc CSE bug workaround */
return (xy.hi + vzs + ldexpl(xy.lo, spread));
}
if (oround != FE_TONEAREST) {
/*
* There is no need to worry about double rounding in directed
* rounding modes.
*/
fesetround(oround);
adj = r.lo + xy.lo;
return (ldexpl(r.hi + adj, spread));
}
adj = add_adjusted(r.lo, xy.lo);
if (spread + ilogbl(r.hi) > -16383)
return (ldexpl(r.hi + adj, spread));
else
return (add_and_denormalize(r.hi, adj, spread));
}
|