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
|
/*
* The LLVM Compiler Infrastructure
*
* This file is distributed under the University of Illinois Open Source
* License. See LICENSE.TXT for details.
*/
#define DOUBLE_PRECISION
#include "fp_lib.h"
// This file implements double-precision soft-float multiplication with the
// IEEE-754 default rounding (to nearest, ties to even).
#define loWord(a) (a & 0xffffffffU)
#define hiWord(a) (a >> 32)
// 64x64 -> 128 wide multiply for platforms that don't have such an operation;
// some 64-bit platforms have this operation, but they tend to have hardware
// floating-point, so we don't bother with a special case for them here.
static inline void wideMultiply(rep_t a, rep_t b, rep_t *hi, rep_t *lo) {
// Each of the component 32x32 -> 64 products
const uint64_t plolo = loWord(a) * loWord(b);
const uint64_t plohi = loWord(a) * hiWord(b);
const uint64_t philo = hiWord(a) * loWord(b);
const uint64_t phihi = hiWord(a) * hiWord(b);
// Sum terms that compute to lo in a way that allows us to get the carry
const uint64_t r0 = loWord(plolo);
const uint64_t r1 = hiWord(plolo) + loWord(plohi) + loWord(philo);
*lo = r0 + (r1 << 32);
// Sum terms contributing to hi with the carry from lo
*hi = hiWord(plohi) + hiWord(philo) + hiWord(r1) + phihi;
}
fp_t __muldf3(fp_t a, fp_t b) {
const unsigned int aExponent = toRep(a) >> significandBits & maxExponent;
const unsigned int bExponent = toRep(b) >> significandBits & maxExponent;
const rep_t productSign = (toRep(a) ^ toRep(b)) & signBit;
rep_t aSignificand = toRep(a) & significandMask;
rep_t bSignificand = toRep(b) & significandMask;
int scale = 0;
// Detect if a or b is zero, denormal, infinity, or NaN.
if (aExponent-1U >= maxExponent-1U || bExponent-1U >= maxExponent-1U) {
const rep_t aAbs = toRep(a) & absMask;
const rep_t bAbs = toRep(b) & absMask;
// NaN * anything = qNaN
if (aAbs > infRep) return fromRep(toRep(a) | quietBit);
// anything * NaN = qNaN
if (bAbs > infRep) return fromRep(toRep(b) | quietBit);
if (aAbs == infRep) {
// infinity * non-zero = +/- infinity
if (bAbs) return fromRep(aAbs | productSign);
// infinity * zero = NaN
else return fromRep(qnanRep);
}
if (bAbs == infRep) {
// non-zero * infinity = +/- infinity
if (aAbs) return fromRep(bAbs | productSign);
// zero * infinity = NaN
else return fromRep(qnanRep);
}
// zero * anything = +/- zero
if (!aAbs) return fromRep(productSign);
// anything * zero = +/- zero
if (!bAbs) return fromRep(productSign);
// one or both of a or b is denormal, the other (if applicable) is a
// normal number. Renormalize one or both of a and b, and set scale to
// include the necessary exponent adjustment.
if (aAbs < implicitBit) scale += normalize(&aSignificand);
if (bAbs < implicitBit) scale += normalize(&bSignificand);
}
// Or in the implicit significand bit. (If we fell through from the
// denormal path it was already set by normalize( ), but setting it twice
// won't hurt anything.)
aSignificand |= implicitBit;
bSignificand |= implicitBit;
// Get the significand of a*b. Before multiplying the significands, shift
// one of them left to left-align it in the field. Thus, the product will
// have (exponentBits + 2) integral digits, all but two of which must be
// zero. Normalizing this result is just a conditional left-shift by one
// and bumping the exponent accordingly.
rep_t productHi, productLo;
wideMultiply(aSignificand, bSignificand << exponentBits,
&productHi, &productLo);
int productExponent = aExponent + bExponent - exponentBias + scale;
// Normalize the significand, adjust exponent if needed.
if (productHi & implicitBit) productExponent++;
else wideLeftShift(&productHi, &productLo, 1);
// If we have overflowed the type, return +/- infinity.
if (productExponent >= maxExponent) return fromRep(infRep | productSign);
if (productExponent <= 0) {
// Result is denormal before rounding
//
// If the result is so small that it just underflows to zero, return
// a zero of the appropriate sign. Mathematically there is no need to
// handle this case separately, but we make it a special case to
// simplify the shift logic.
const int shift = 1 - productExponent;
if (shift >= typeWidth) return fromRep(productSign);
// Otherwise, shift the significand of the result so that the round
// bit is the high bit of productLo.
wideRightShiftWithSticky(&productHi, &productLo, shift);
}
else {
// Result is normal before rounding; insert the exponent.
productHi &= significandMask;
productHi |= (rep_t)productExponent << significandBits;
}
// Insert the sign of the result:
productHi |= productSign;
// Final rounding. The final result may overflow to infinity, or underflow
// to zero, but those are the correct results in those cases. We use the
// default IEEE-754 round-to-nearest, ties-to-even rounding mode.
if (productLo > signBit) productHi++;
if (productLo == signBit) productHi += productHi & 1;
return fromRep(productHi);
}
|
