about summary refs log tree commit diff
path: root/sysdeps/ieee754/dbl-64/s_fma.c
blob: 55ccf46c9a239541a42b02ede64f550ccc8c3df4 (plain) (blame)
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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
/* Compute x * y + z as ternary operation.
   Copyright (C) 2010-2021 Free Software Foundation, Inc.
   This file is part of the GNU C Library.

   The GNU C Library is free software; you can redistribute it and/or
   modify it under the terms of the GNU Lesser General Public
   License as published by the Free Software Foundation; either
   version 2.1 of the License, or (at your option) any later version.

   The GNU C Library is distributed in the hope that it will be useful,
   but WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
   Lesser General Public License for more details.

   You should have received a copy of the GNU Lesser General Public
   License along with the GNU C Library; if not, see
   <https://www.gnu.org/licenses/>.  */

#include <float.h>
#include <math.h>
#include <fenv.h>
#include <ieee754.h>
#include <math-barriers.h>
#include <fenv_private.h>
#include <libm-alias-double.h>
#include <tininess.h>
#include <math-use-builtins.h>

/* This implementation uses rounding to odd to avoid problems with
   double rounding.  See a paper by Boldo and Melquiond:
   http://www.lri.fr/~melquion/doc/08-tc.pdf  */

double
__fma (double x, double y, double z)
{
#if USE_FMA_BUILTIN
  return __builtin_fma (x, y, z);
#else
  /* Use generic implementation.  */
  union ieee754_double u, v, w;
  int adjust = 0;
  u.d = x;
  v.d = y;
  w.d = z;
  if (__builtin_expect (u.ieee.exponent + v.ieee.exponent
			>= 0x7ff + IEEE754_DOUBLE_BIAS - DBL_MANT_DIG, 0)
      || __builtin_expect (u.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0)
      || __builtin_expect (v.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0)
      || __builtin_expect (w.ieee.exponent >= 0x7ff - DBL_MANT_DIG, 0)
      || __builtin_expect (u.ieee.exponent + v.ieee.exponent
			   <= IEEE754_DOUBLE_BIAS + DBL_MANT_DIG, 0))
    {
      /* If z is Inf, but x and y are finite, the result should be
	 z rather than NaN.  */
      if (w.ieee.exponent == 0x7ff
	  && u.ieee.exponent != 0x7ff
	  && v.ieee.exponent != 0x7ff)
	return (z + x) + y;
      /* If z is zero and x are y are nonzero, compute the result
	 as x * y to avoid the wrong sign of a zero result if x * y
	 underflows to 0.  */
      if (z == 0 && x != 0 && y != 0)
	return x * y;
      /* If x or y or z is Inf/NaN, or if x * y is zero, compute as
	 x * y + z.  */
      if (u.ieee.exponent == 0x7ff
	  || v.ieee.exponent == 0x7ff
	  || w.ieee.exponent == 0x7ff
	  || x == 0
	  || y == 0)
	return x * y + z;
      /* If fma will certainly overflow, compute as x * y.  */
      if (u.ieee.exponent + v.ieee.exponent > 0x7ff + IEEE754_DOUBLE_BIAS)
	return x * y;
      /* If x * y is less than 1/4 of DBL_TRUE_MIN, neither the
	 result nor whether there is underflow depends on its exact
	 value, only on its sign.  */
      if (u.ieee.exponent + v.ieee.exponent
	  < IEEE754_DOUBLE_BIAS - DBL_MANT_DIG - 2)
	{
	  int neg = u.ieee.negative ^ v.ieee.negative;
	  double tiny = neg ? -0x1p-1074 : 0x1p-1074;
	  if (w.ieee.exponent >= 3)
	    return tiny + z;
	  /* Scaling up, adding TINY and scaling down produces the
	     correct result, because in round-to-nearest mode adding
	     TINY has no effect and in other modes double rounding is
	     harmless.  But it may not produce required underflow
	     exceptions.  */
	  v.d = z * 0x1p54 + tiny;
	  if (TININESS_AFTER_ROUNDING
	      ? v.ieee.exponent < 55
	      : (w.ieee.exponent == 0
		 || (w.ieee.exponent == 1
		     && w.ieee.negative != neg
		     && w.ieee.mantissa1 == 0
		     && w.ieee.mantissa0 == 0)))
	    {
	      double force_underflow = x * y;
	      math_force_eval (force_underflow);
	    }
	  return v.d * 0x1p-54;
	}
      if (u.ieee.exponent + v.ieee.exponent
	  >= 0x7ff + IEEE754_DOUBLE_BIAS - DBL_MANT_DIG)
	{
	  /* Compute 1p-53 times smaller result and multiply
	     at the end.  */
	  if (u.ieee.exponent > v.ieee.exponent)
	    u.ieee.exponent -= DBL_MANT_DIG;
	  else
	    v.ieee.exponent -= DBL_MANT_DIG;
	  /* If x + y exponent is very large and z exponent is very small,
	     it doesn't matter if we don't adjust it.  */
	  if (w.ieee.exponent > DBL_MANT_DIG)
	    w.ieee.exponent -= DBL_MANT_DIG;
	  adjust = 1;
	}
      else if (w.ieee.exponent >= 0x7ff - DBL_MANT_DIG)
	{
	  /* Similarly.
	     If z exponent is very large and x and y exponents are
	     very small, adjust them up to avoid spurious underflows,
	     rather than down.  */
	  if (u.ieee.exponent + v.ieee.exponent
	      <= IEEE754_DOUBLE_BIAS + 2 * DBL_MANT_DIG)
	    {
	      if (u.ieee.exponent > v.ieee.exponent)
		u.ieee.exponent += 2 * DBL_MANT_DIG + 2;
	      else
		v.ieee.exponent += 2 * DBL_MANT_DIG + 2;
	    }
	  else if (u.ieee.exponent > v.ieee.exponent)
	    {
	      if (u.ieee.exponent > DBL_MANT_DIG)
		u.ieee.exponent -= DBL_MANT_DIG;
	    }
	  else if (v.ieee.exponent > DBL_MANT_DIG)
	    v.ieee.exponent -= DBL_MANT_DIG;
	  w.ieee.exponent -= DBL_MANT_DIG;
	  adjust = 1;
	}
      else if (u.ieee.exponent >= 0x7ff - DBL_MANT_DIG)
	{
	  u.ieee.exponent -= DBL_MANT_DIG;
	  if (v.ieee.exponent)
	    v.ieee.exponent += DBL_MANT_DIG;
	  else
	    v.d *= 0x1p53;
	}
      else if (v.ieee.exponent >= 0x7ff - DBL_MANT_DIG)
	{
	  v.ieee.exponent -= DBL_MANT_DIG;
	  if (u.ieee.exponent)
	    u.ieee.exponent += DBL_MANT_DIG;
	  else
	    u.d *= 0x1p53;
	}
      else /* if (u.ieee.exponent + v.ieee.exponent
		  <= IEEE754_DOUBLE_BIAS + DBL_MANT_DIG) */
	{
	  if (u.ieee.exponent > v.ieee.exponent)
	    u.ieee.exponent += 2 * DBL_MANT_DIG + 2;
	  else
	    v.ieee.exponent += 2 * DBL_MANT_DIG + 2;
	  if (w.ieee.exponent <= 4 * DBL_MANT_DIG + 6)
	    {
	      if (w.ieee.exponent)
		w.ieee.exponent += 2 * DBL_MANT_DIG + 2;
	      else
		w.d *= 0x1p108;
	      adjust = -1;
	    }
	  /* Otherwise x * y should just affect inexact
	     and nothing else.  */
	}
      x = u.d;
      y = v.d;
      z = w.d;
    }

  /* Ensure correct sign of exact 0 + 0.  */
  if (__glibc_unlikely ((x == 0 || y == 0) && z == 0))
    {
      x = math_opt_barrier (x);
      return x * y + z;
    }

  fenv_t env;
  libc_feholdexcept_setround (&env, FE_TONEAREST);

  /* Multiplication m1 + m2 = x * y using Dekker's algorithm.  */
#define C ((1 << (DBL_MANT_DIG + 1) / 2) + 1)
  double x1 = x * C;
  double y1 = y * C;
  double m1 = x * y;
  x1 = (x - x1) + x1;
  y1 = (y - y1) + y1;
  double x2 = x - x1;
  double y2 = y - y1;
  double m2 = (((x1 * y1 - m1) + x1 * y2) + x2 * y1) + x2 * y2;

  /* Addition a1 + a2 = z + m1 using Knuth's algorithm.  */
  double a1 = z + m1;
  double t1 = a1 - z;
  double t2 = a1 - t1;
  t1 = m1 - t1;
  t2 = z - t2;
  double a2 = t1 + t2;
  /* Ensure the arithmetic is not scheduled after feclearexcept call.  */
  math_force_eval (m2);
  math_force_eval (a2);
  feclearexcept (FE_INEXACT);

  /* If the result is an exact zero, ensure it has the correct sign.  */
  if (a1 == 0 && m2 == 0)
    {
      libc_feupdateenv (&env);
      /* Ensure that round-to-nearest value of z + m1 is not reused.  */
      z = math_opt_barrier (z);
      return z + m1;
    }

  libc_fesetround (FE_TOWARDZERO);

  /* Perform m2 + a2 addition with round to odd.  */
  u.d = a2 + m2;

  if (__glibc_unlikely (adjust < 0))
    {
      if ((u.ieee.mantissa1 & 1) == 0)
	u.ieee.mantissa1 |= libc_fetestexcept (FE_INEXACT) != 0;
      v.d = a1 + u.d;
      /* Ensure the addition is not scheduled after fetestexcept call.  */
      math_force_eval (v.d);
    }

  /* Reset rounding mode and test for inexact simultaneously.  */
  int j = libc_feupdateenv_test (&env, FE_INEXACT) != 0;

  if (__glibc_likely (adjust == 0))
    {
      if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7ff)
	u.ieee.mantissa1 |= j;
      /* Result is a1 + u.d.  */
      return a1 + u.d;
    }
  else if (__glibc_likely (adjust > 0))
    {
      if ((u.ieee.mantissa1 & 1) == 0 && u.ieee.exponent != 0x7ff)
	u.ieee.mantissa1 |= j;
      /* Result is a1 + u.d, scaled up.  */
      return (a1 + u.d) * 0x1p53;
    }
  else
    {
      /* If a1 + u.d is exact, the only rounding happens during
	 scaling down.  */
      if (j == 0)
	return v.d * 0x1p-108;
      /* If result rounded to zero is not subnormal, no double
	 rounding will occur.  */
      if (v.ieee.exponent > 108)
	return (a1 + u.d) * 0x1p-108;
      /* If v.d * 0x1p-108 with round to zero is a subnormal above
	 or equal to DBL_MIN / 2, then v.d * 0x1p-108 shifts mantissa
	 down just by 1 bit, which means v.ieee.mantissa1 |= j would
	 change the round bit, not sticky or guard bit.
	 v.d * 0x1p-108 never normalizes by shifting up,
	 so round bit plus sticky bit should be already enough
	 for proper rounding.  */
      if (v.ieee.exponent == 108)
	{
	  /* If the exponent would be in the normal range when
	     rounding to normal precision with unbounded exponent
	     range, the exact result is known and spurious underflows
	     must be avoided on systems detecting tininess after
	     rounding.  */
	  if (TININESS_AFTER_ROUNDING)
	    {
	      w.d = a1 + u.d;
	      if (w.ieee.exponent == 109)
		return w.d * 0x1p-108;
	    }
	  /* v.ieee.mantissa1 & 2 is LSB bit of the result before rounding,
	     v.ieee.mantissa1 & 1 is the round bit and j is our sticky
	     bit.  */
	  w.d = 0.0;
	  w.ieee.mantissa1 = ((v.ieee.mantissa1 & 3) << 1) | j;
	  w.ieee.negative = v.ieee.negative;
	  v.ieee.mantissa1 &= ~3U;
	  v.d *= 0x1p-108;
	  w.d *= 0x1p-2;
	  return v.d + w.d;
	}
      v.ieee.mantissa1 |= j;
      return v.d * 0x1p-108;
    }
#endif /* ! USE_FMA_BUILTIN  */
}
#ifndef __fma
libm_alias_double (__fma, fma)
#endif