1 |
|
|
//= lib/fp_trunc_impl.inc - high precision -> low precision conversion *-*-===// |
2 |
|
|
// |
3 |
|
|
// The LLVM Compiler Infrastructure |
4 |
|
|
// |
5 |
|
|
// This file is dual licensed under the MIT and the University of Illinois Open |
6 |
|
|
// Source Licenses. See LICENSE.TXT for details. |
7 |
|
|
// |
8 |
|
|
//===----------------------------------------------------------------------===// |
9 |
|
|
// |
10 |
|
|
// This file implements a fairly generic conversion from a wider to a narrower |
11 |
|
|
// IEEE-754 floating-point type in the default (round to nearest, ties to even) |
12 |
|
|
// rounding mode. The constants and types defined following the includes below |
13 |
|
|
// parameterize the conversion. |
14 |
|
|
// |
15 |
|
|
// This routine can be trivially adapted to support conversions to |
16 |
|
|
// half-precision or from quad-precision. It does not support types that don't |
17 |
|
|
// use the usual IEEE-754 interchange formats; specifically, some work would be |
18 |
|
|
// needed to adapt it to (for example) the Intel 80-bit format or PowerPC |
19 |
|
|
// double-double format. |
20 |
|
|
// |
21 |
|
|
// Note please, however, that this implementation is only intended to support |
22 |
|
|
// *narrowing* operations; if you need to convert to a *wider* floating-point |
23 |
|
|
// type (e.g. float -> double), then this routine will not do what you want it |
24 |
|
|
// to. |
25 |
|
|
// |
26 |
|
|
// It also requires that integer types at least as large as both formats |
27 |
|
|
// are available on the target platform; this may pose a problem when trying |
28 |
|
|
// to add support for quad on some 32-bit systems, for example. |
29 |
|
|
// |
30 |
|
|
// Finally, the following assumptions are made: |
31 |
|
|
// |
32 |
|
|
// 1. floating-point types and integer types have the same endianness on the |
33 |
|
|
// target platform |
34 |
|
|
// |
35 |
|
|
// 2. quiet NaNs, if supported, are indicated by the leading bit of the |
36 |
|
|
// significand field being set |
37 |
|
|
// |
38 |
|
|
//===----------------------------------------------------------------------===// |
39 |
|
|
|
40 |
|
|
#include "fp_trunc.h" |
41 |
|
|
|
42 |
|
|
static __inline dst_t __truncXfYf2__(src_t a) { |
43 |
|
|
// Various constants whose values follow from the type parameters. |
44 |
|
|
// Any reasonable optimizer will fold and propagate all of these. |
45 |
|
|
const int srcBits = sizeof(src_t)*CHAR_BIT; |
46 |
|
|
const int srcExpBits = srcBits - srcSigBits - 1; |
47 |
|
|
const int srcInfExp = (1 << srcExpBits) - 1; |
48 |
|
|
const int srcExpBias = srcInfExp >> 1; |
49 |
|
|
|
50 |
|
|
const src_rep_t srcMinNormal = SRC_REP_C(1) << srcSigBits; |
51 |
|
|
const src_rep_t srcSignificandMask = srcMinNormal - 1; |
52 |
|
|
const src_rep_t srcInfinity = (src_rep_t)srcInfExp << srcSigBits; |
53 |
|
|
const src_rep_t srcSignMask = SRC_REP_C(1) << (srcSigBits + srcExpBits); |
54 |
|
|
const src_rep_t srcAbsMask = srcSignMask - 1; |
55 |
|
|
const src_rep_t roundMask = (SRC_REP_C(1) << (srcSigBits - dstSigBits)) - 1; |
56 |
|
|
const src_rep_t halfway = SRC_REP_C(1) << (srcSigBits - dstSigBits - 1); |
57 |
|
|
const src_rep_t srcQNaN = SRC_REP_C(1) << (srcSigBits - 1); |
58 |
|
|
const src_rep_t srcNaNCode = srcQNaN - 1; |
59 |
|
|
|
60 |
|
|
const int dstBits = sizeof(dst_t)*CHAR_BIT; |
61 |
|
|
const int dstExpBits = dstBits - dstSigBits - 1; |
62 |
|
|
const int dstInfExp = (1 << dstExpBits) - 1; |
63 |
|
|
const int dstExpBias = dstInfExp >> 1; |
64 |
|
|
|
65 |
|
|
const int underflowExponent = srcExpBias + 1 - dstExpBias; |
66 |
|
|
const int overflowExponent = srcExpBias + dstInfExp - dstExpBias; |
67 |
|
|
const src_rep_t underflow = (src_rep_t)underflowExponent << srcSigBits; |
68 |
|
|
const src_rep_t overflow = (src_rep_t)overflowExponent << srcSigBits; |
69 |
|
|
|
70 |
|
|
const dst_rep_t dstQNaN = DST_REP_C(1) << (dstSigBits - 1); |
71 |
|
|
const dst_rep_t dstNaNCode = dstQNaN - 1; |
72 |
|
|
|
73 |
|
|
// Break a into a sign and representation of the absolute value |
74 |
|
|
const src_rep_t aRep = srcToRep(a); |
75 |
|
|
const src_rep_t aAbs = aRep & srcAbsMask; |
76 |
|
|
const src_rep_t sign = aRep & srcSignMask; |
77 |
|
|
dst_rep_t absResult; |
78 |
|
|
|
79 |
|
|
if (aAbs - underflow < aAbs - overflow) { |
80 |
|
|
// The exponent of a is within the range of normal numbers in the |
81 |
|
|
// destination format. We can convert by simply right-shifting with |
82 |
|
|
// rounding and adjusting the exponent. |
83 |
|
|
absResult = aAbs >> (srcSigBits - dstSigBits); |
84 |
|
|
absResult -= (dst_rep_t)(srcExpBias - dstExpBias) << dstSigBits; |
85 |
|
|
|
86 |
|
|
const src_rep_t roundBits = aAbs & roundMask; |
87 |
|
|
// Round to nearest |
88 |
|
|
if (roundBits > halfway) |
89 |
|
|
absResult++; |
90 |
|
|
// Ties to even |
91 |
|
|
else if (roundBits == halfway) |
92 |
|
|
absResult += absResult & 1; |
93 |
|
|
} |
94 |
|
|
else if (aAbs > srcInfinity) { |
95 |
|
|
// a is NaN. |
96 |
|
|
// Conjure the result by beginning with infinity, setting the qNaN |
97 |
|
|
// bit and inserting the (truncated) trailing NaN field. |
98 |
|
|
absResult = (dst_rep_t)dstInfExp << dstSigBits; |
99 |
|
|
absResult |= dstQNaN; |
100 |
|
|
absResult |= ((aAbs & srcNaNCode) >> (srcSigBits - dstSigBits)) & dstNaNCode; |
101 |
|
|
} |
102 |
|
|
else if (aAbs >= overflow) { |
103 |
|
|
// a overflows to infinity. |
104 |
|
|
absResult = (dst_rep_t)dstInfExp << dstSigBits; |
105 |
|
|
} |
106 |
|
|
else { |
107 |
|
|
// a underflows on conversion to the destination type or is an exact |
108 |
|
|
// zero. The result may be a denormal or zero. Extract the exponent |
109 |
|
|
// to get the shift amount for the denormalization. |
110 |
|
|
const int aExp = aAbs >> srcSigBits; |
111 |
|
|
const int shift = srcExpBias - dstExpBias - aExp + 1; |
112 |
|
|
|
113 |
|
|
const src_rep_t significand = (aRep & srcSignificandMask) | srcMinNormal; |
114 |
|
|
|
115 |
|
|
// Right shift by the denormalization amount with sticky. |
116 |
|
|
if (shift > srcSigBits) { |
117 |
|
|
absResult = 0; |
118 |
|
|
} else { |
119 |
|
|
const bool sticky = significand << (srcBits - shift); |
120 |
|
|
src_rep_t denormalizedSignificand = significand >> shift | sticky; |
121 |
|
|
absResult = denormalizedSignificand >> (srcSigBits - dstSigBits); |
122 |
|
|
const src_rep_t roundBits = denormalizedSignificand & roundMask; |
123 |
|
|
// Round to nearest |
124 |
|
|
if (roundBits > halfway) |
125 |
|
|
absResult++; |
126 |
|
|
// Ties to even |
127 |
|
|
else if (roundBits == halfway) |
128 |
|
|
absResult += absResult & 1; |
129 |
|
|
} |
130 |
|
|
} |
131 |
|
|
|
132 |
|
|
// Apply the signbit to (dst_t)abs(a). |
133 |
|
|
const dst_rep_t result = absResult | sign >> (srcBits - dstBits); |
134 |
|
|
return dstFromRep(result); |
135 |
|
|
} |