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
| //= lib/fp_trunc_impl.inc - high precision -> low precision conversion *-*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements a fairly generic conversion from a wider to a narrower
// IEEE-754 floating-point type in the default (round to nearest, ties to even)
// rounding mode. The constants and types defined following the includes below
// parameterize the conversion.
//
// This routine can be trivially adapted to support conversions to
// half-precision or from quad-precision. It does not support types that don't
// use the usual IEEE-754 interchange formats; specifically, some work would be
// needed to adapt it to (for example) the Intel 80-bit format or PowerPC
// double-double format.
//
// Note please, however, that this implementation is only intended to support
// *narrowing* operations; if you need to convert to a *wider* floating-point
// type (e.g. float -> double), then this routine will not do what you want it
// to.
//
// It also requires that integer types at least as large as both formats
// are available on the target platform; this may pose a problem when trying
// to add support for quad on some 32-bit systems, for example.
//
// Finally, the following assumptions are made:
//
// 1. Floating-point types and integer types have the same endianness on the
// target platform.
//
// 2. Quiet NaNs, if supported, are indicated by the leading bit of the
// significand field being set.
//
//===----------------------------------------------------------------------===//
#include "fp_trunc.h"
static __inline dst_t __truncXfYf2__(src_t a) {
// Various constants whose values follow from the type parameters.
// Any reasonable optimizer will fold and propagate all of these.
const int srcBits = sizeof(src_t) * CHAR_BIT;
const int srcExpBits = srcBits - srcSigBits - 1;
const int srcInfExp = (1 << srcExpBits) - 1;
const int srcExpBias = srcInfExp >> 1;
const src_rep_t srcMinNormal = SRC_REP_C(1) << srcSigBits;
const src_rep_t srcSignificandMask = srcMinNormal - 1;
const src_rep_t srcInfinity = (src_rep_t)srcInfExp << srcSigBits;
const src_rep_t srcSignMask = SRC_REP_C(1) << (srcSigBits + srcExpBits);
const src_rep_t srcAbsMask = srcSignMask - 1;
const src_rep_t roundMask = (SRC_REP_C(1) << (srcSigBits - dstSigBits)) - 1;
const src_rep_t halfway = SRC_REP_C(1) << (srcSigBits - dstSigBits - 1);
const src_rep_t srcQNaN = SRC_REP_C(1) << (srcSigBits - 1);
const src_rep_t srcNaNCode = srcQNaN - 1;
const int dstBits = sizeof(dst_t) * CHAR_BIT;
const int dstExpBits = dstBits - dstSigBits - 1;
const int dstInfExp = (1 << dstExpBits) - 1;
const int dstExpBias = dstInfExp >> 1;
const int underflowExponent = srcExpBias + 1 - dstExpBias;
const int overflowExponent = srcExpBias + dstInfExp - dstExpBias;
const src_rep_t underflow = (src_rep_t)underflowExponent << srcSigBits;
const src_rep_t overflow = (src_rep_t)overflowExponent << srcSigBits;
const dst_rep_t dstQNaN = DST_REP_C(1) << (dstSigBits - 1);
const dst_rep_t dstNaNCode = dstQNaN - 1;
// Break a into a sign and representation of the absolute value.
const src_rep_t aRep = srcToRep(a);
const src_rep_t aAbs = aRep & srcAbsMask;
const src_rep_t sign = aRep & srcSignMask;
dst_rep_t absResult;
if (aAbs - underflow < aAbs - overflow) {
// The exponent of a is within the range of normal numbers in the
// destination format. We can convert by simply right-shifting with
// rounding and adjusting the exponent.
absResult = aAbs >> (srcSigBits - dstSigBits);
absResult -= (dst_rep_t)(srcExpBias - dstExpBias) << dstSigBits;
const src_rep_t roundBits = aAbs & roundMask;
// Round to nearest.
if (roundBits > halfway)
absResult++;
// Tie to even.
else if (roundBits == halfway)
absResult += absResult & 1;
} else if (aAbs > srcInfinity) {
// a is NaN.
// Conjure the result by beginning with infinity, setting the qNaN
// bit and inserting the (truncated) trailing NaN field.
absResult = (dst_rep_t)dstInfExp << dstSigBits;
absResult |= dstQNaN;
absResult |=
((aAbs & srcNaNCode) >> (srcSigBits - dstSigBits)) & dstNaNCode;
} else if (aAbs >= overflow) {
// a overflows to infinity.
absResult = (dst_rep_t)dstInfExp << dstSigBits;
} else {
// a underflows on conversion to the destination type or is an exact
// zero. The result may be a denormal or zero. Extract the exponent
// to get the shift amount for the denormalization.
const int aExp = aAbs >> srcSigBits;
const int shift = srcExpBias - dstExpBias - aExp + 1;
const src_rep_t significand = (aRep & srcSignificandMask) | srcMinNormal;
// Right shift by the denormalization amount with sticky.
if (shift > srcSigBits) {
absResult = 0;
} else {
const bool sticky = (significand << (srcBits - shift)) != 0;
src_rep_t denormalizedSignificand = significand >> shift | sticky;
absResult = denormalizedSignificand >> (srcSigBits - dstSigBits);
const src_rep_t roundBits = denormalizedSignificand & roundMask;
// Round to nearest
if (roundBits > halfway)
absResult++;
// Ties to even
else if (roundBits == halfway)
absResult += absResult & 1;
}
}
// Apply the signbit to the absolute value.
const dst_rep_t result = absResult | sign >> (srcBits - dstBits);
return dstFromRep(result);
}
|