1 /* Compute x * y + z as ternary operation.
2 Copyright (C) 2010-2013 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4 Contributed by Jakub Jelinek <jakub@redhat.com>, 2010.
6 The GNU C Library is free software; you can redistribute it and/or
7 modify it under the terms of the GNU Lesser General Public
8 License as published by the Free Software Foundation; either
9 version 2.1 of the License, or (at your option) any later version.
11 The GNU C Library is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 Lesser General Public License for more details.
16 You should have received a copy of the GNU Lesser General Public
17 License along with the GNU C Library; if not, see
18 <http://www.gnu.org/licenses/>. */
24 #include <math_private.h>
27 /* This implementation uses rounding to odd to avoid problems with
28 double rounding. See a paper by Boldo and Melquiond:
29 http://www.lri.fr/~melquion/doc/08-tc.pdf */
32 __fma (double x
, double y
, double z
)
34 union ieee754_double u
, v
, w
;
39 if (__builtin_expect (u
.ieee
.exponent
+ v
.ieee
.exponent
40 >= 0x7ff + IEEE754_DOUBLE_BIAS
- DBL_MANT_DIG
, 0)
41 || __builtin_expect (u
.ieee
.exponent
>= 0x7ff - DBL_MANT_DIG
, 0)
42 || __builtin_expect (v
.ieee
.exponent
>= 0x7ff - DBL_MANT_DIG
, 0)
43 || __builtin_expect (w
.ieee
.exponent
>= 0x7ff - DBL_MANT_DIG
, 0)
44 || __builtin_expect (u
.ieee
.exponent
+ v
.ieee
.exponent
45 <= IEEE754_DOUBLE_BIAS
+ DBL_MANT_DIG
, 0))
47 /* If z is Inf, but x and y are finite, the result should be
49 if (w
.ieee
.exponent
== 0x7ff
50 && u
.ieee
.exponent
!= 0x7ff
51 && v
.ieee
.exponent
!= 0x7ff)
53 /* If z is zero and x are y are nonzero, compute the result
54 as x * y to avoid the wrong sign of a zero result if x * y
56 if (z
== 0 && x
!= 0 && y
!= 0)
58 /* If x or y or z is Inf/NaN, or if x * y is zero, compute as
60 if (u
.ieee
.exponent
== 0x7ff
61 || v
.ieee
.exponent
== 0x7ff
62 || w
.ieee
.exponent
== 0x7ff
66 /* If fma will certainly overflow, compute as x * y. */
67 if (u
.ieee
.exponent
+ v
.ieee
.exponent
> 0x7ff + IEEE754_DOUBLE_BIAS
)
69 /* If x * y is less than 1/4 of DBL_DENORM_MIN, neither the
70 result nor whether there is underflow depends on its exact
71 value, only on its sign. */
72 if (u
.ieee
.exponent
+ v
.ieee
.exponent
73 < IEEE754_DOUBLE_BIAS
- DBL_MANT_DIG
- 2)
75 int neg
= u
.ieee
.negative
^ v
.ieee
.negative
;
76 double tiny
= neg
? -0x1p
-1074 : 0x1p
-1074;
77 if (w
.ieee
.exponent
>= 3)
79 /* Scaling up, adding TINY and scaling down produces the
80 correct result, because in round-to-nearest mode adding
81 TINY has no effect and in other modes double rounding is
82 harmless. But it may not produce required underflow
84 v
.d
= z
* 0x1p
54 + tiny
;
85 if (TININESS_AFTER_ROUNDING
86 ? v
.ieee
.exponent
< 55
87 : (w
.ieee
.exponent
== 0
88 || (w
.ieee
.exponent
== 1
89 && w
.ieee
.negative
!= neg
90 && w
.ieee
.mantissa1
== 0
91 && w
.ieee
.mantissa0
== 0)))
93 volatile double force_underflow
= x
* y
;
94 (void) force_underflow
;
98 if (u
.ieee
.exponent
+ v
.ieee
.exponent
99 >= 0x7ff + IEEE754_DOUBLE_BIAS
- DBL_MANT_DIG
)
101 /* Compute 1p-53 times smaller result and multiply
103 if (u
.ieee
.exponent
> v
.ieee
.exponent
)
104 u
.ieee
.exponent
-= DBL_MANT_DIG
;
106 v
.ieee
.exponent
-= DBL_MANT_DIG
;
107 /* If x + y exponent is very large and z exponent is very small,
108 it doesn't matter if we don't adjust it. */
109 if (w
.ieee
.exponent
> DBL_MANT_DIG
)
110 w
.ieee
.exponent
-= DBL_MANT_DIG
;
113 else if (w
.ieee
.exponent
>= 0x7ff - DBL_MANT_DIG
)
116 If z exponent is very large and x and y exponents are
117 very small, adjust them up to avoid spurious underflows,
119 if (u
.ieee
.exponent
+ v
.ieee
.exponent
120 <= IEEE754_DOUBLE_BIAS
+ DBL_MANT_DIG
)
122 if (u
.ieee
.exponent
> v
.ieee
.exponent
)
123 u
.ieee
.exponent
+= 2 * DBL_MANT_DIG
+ 2;
125 v
.ieee
.exponent
+= 2 * DBL_MANT_DIG
+ 2;
127 else if (u
.ieee
.exponent
> v
.ieee
.exponent
)
129 if (u
.ieee
.exponent
> DBL_MANT_DIG
)
130 u
.ieee
.exponent
-= DBL_MANT_DIG
;
132 else if (v
.ieee
.exponent
> DBL_MANT_DIG
)
133 v
.ieee
.exponent
-= DBL_MANT_DIG
;
134 w
.ieee
.exponent
-= DBL_MANT_DIG
;
137 else if (u
.ieee
.exponent
>= 0x7ff - DBL_MANT_DIG
)
139 u
.ieee
.exponent
-= DBL_MANT_DIG
;
141 v
.ieee
.exponent
+= DBL_MANT_DIG
;
145 else if (v
.ieee
.exponent
>= 0x7ff - DBL_MANT_DIG
)
147 v
.ieee
.exponent
-= DBL_MANT_DIG
;
149 u
.ieee
.exponent
+= DBL_MANT_DIG
;
153 else /* if (u.ieee.exponent + v.ieee.exponent
154 <= IEEE754_DOUBLE_BIAS + DBL_MANT_DIG) */
156 if (u
.ieee
.exponent
> v
.ieee
.exponent
)
157 u
.ieee
.exponent
+= 2 * DBL_MANT_DIG
+ 2;
159 v
.ieee
.exponent
+= 2 * DBL_MANT_DIG
+ 2;
160 if (w
.ieee
.exponent
<= 4 * DBL_MANT_DIG
+ 6)
163 w
.ieee
.exponent
+= 2 * DBL_MANT_DIG
+ 2;
168 /* Otherwise x * y should just affect inexact
176 /* Ensure correct sign of exact 0 + 0. */
177 if (__builtin_expect ((x
== 0 || y
== 0) && z
== 0, 0))
181 libc_feholdexcept_setround (&env
, FE_TONEAREST
);
183 /* Multiplication m1 + m2 = x * y using Dekker's algorithm. */
184 #define C ((1 << (DBL_MANT_DIG + 1) / 2) + 1)
192 double m2
= (((x1
* y1
- m1
) + x1
* y2
) + x2
* y1
) + x2
* y2
;
194 /* Addition a1 + a2 = z + m1 using Knuth's algorithm. */
201 feclearexcept (FE_INEXACT
);
203 /* If the result is an exact zero, ensure it has the correct
205 if (a1
== 0 && m2
== 0)
207 libc_feupdateenv (&env
);
208 /* Ensure that round-to-nearest value of z + m1 is not
210 asm volatile ("" : "=m" (z
) : "m" (z
));
214 libc_fesetround (FE_TOWARDZERO
);
216 /* Perform m2 + a2 addition with round to odd. */
219 if (__builtin_expect (adjust
< 0, 0))
221 if ((u
.ieee
.mantissa1
& 1) == 0)
222 u
.ieee
.mantissa1
|= libc_fetestexcept (FE_INEXACT
) != 0;
224 /* Ensure the addition is not scheduled after fetestexcept call. */
225 math_force_eval (v
.d
);
228 /* Reset rounding mode and test for inexact simultaneously. */
229 int j
= libc_feupdateenv_test (&env
, FE_INEXACT
) != 0;
231 if (__builtin_expect (adjust
== 0, 1))
233 if ((u
.ieee
.mantissa1
& 1) == 0 && u
.ieee
.exponent
!= 0x7ff)
234 u
.ieee
.mantissa1
|= j
;
235 /* Result is a1 + u.d. */
238 else if (__builtin_expect (adjust
> 0, 1))
240 if ((u
.ieee
.mantissa1
& 1) == 0 && u
.ieee
.exponent
!= 0x7ff)
241 u
.ieee
.mantissa1
|= j
;
242 /* Result is a1 + u.d, scaled up. */
243 return (a1
+ u
.d
) * 0x1p
53;
247 /* If a1 + u.d is exact, the only rounding happens during
250 return v
.d
* 0x1p
-108;
251 /* If result rounded to zero is not subnormal, no double
252 rounding will occur. */
253 if (v
.ieee
.exponent
> 108)
254 return (a1
+ u
.d
) * 0x1p
-108;
255 /* If v.d * 0x1p-108 with round to zero is a subnormal above
256 or equal to DBL_MIN / 2, then v.d * 0x1p-108 shifts mantissa
257 down just by 1 bit, which means v.ieee.mantissa1 |= j would
258 change the round bit, not sticky or guard bit.
259 v.d * 0x1p-108 never normalizes by shifting up,
260 so round bit plus sticky bit should be already enough
261 for proper rounding. */
262 if (v
.ieee
.exponent
== 108)
264 /* If the exponent would be in the normal range when
265 rounding to normal precision with unbounded exponent
266 range, the exact result is known and spurious underflows
267 must be avoided on systems detecting tininess after
269 if (TININESS_AFTER_ROUNDING
)
272 if (w
.ieee
.exponent
== 109)
273 return w
.d
* 0x1p
-108;
275 /* v.ieee.mantissa1 & 2 is LSB bit of the result before rounding,
276 v.ieee.mantissa1 & 1 is the round bit and j is our sticky
279 w
.ieee
.mantissa1
= ((v
.ieee
.mantissa1
& 3) << 1) | j
;
280 w
.ieee
.negative
= v
.ieee
.negative
;
281 v
.ieee
.mantissa1
&= ~3U;
286 v
.ieee
.mantissa1
|= j
;
287 return v
.d
* 0x1p
-108;
291 weak_alias (__fma
, fma
)
294 #ifdef NO_LONG_DOUBLE
295 strong_alias (__fma
, __fmal
)
296 weak_alias (__fmal
, fmal
)