1 //=-lib/fp_extend_impl.inc - low precision -> high precision conversion -*-- -//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file implements a fairly generic conversion from a narrower to a wider
10 // IEEE-754 floating-point type. The constants and types defined following the
11 // includes below parameterize the conversion.
13 // It does not support types that don't use the usual IEEE-754 interchange
14 // formats; specifically, some work would be needed to adapt it to
15 // (for example) the Intel 80-bit format or PowerPC double-double format.
17 // Note please, however, that this implementation is only intended to support
18 // *widening* operations; if you need to convert to a *narrower* floating-point
19 // type (e.g. double -> float), then this routine will not do what you want it
22 // It also requires that integer types at least as large as both formats
23 // are available on the target platform; this may pose a problem when trying
24 // to add support for quad on some 32-bit systems, for example. You also may
25 // run into trouble finding an appropriate CLZ function for wide source types;
26 // you will likely need to roll your own on some platforms.
28 // Finally, the following assumptions are made:
30 // 1. Floating-point types and integer types have the same endianness on the
33 // 2. Quiet NaNs, if supported, are indicated by the leading bit of the
34 // significand field being set.
36 //===----------------------------------------------------------------------===//
38 #include "fp_extend.h"
40 static __inline dst_t __extendXfYf2__(src_t a) {
41 // Various constants whose values follow from the type parameters.
42 // Any reasonable optimizer will fold and propagate all of these.
43 const int srcBits = sizeof(src_t) * CHAR_BIT;
44 const int srcExpBits = srcBits - srcSigBits - 1;
45 const int srcInfExp = (1 << srcExpBits) - 1;
46 const int srcExpBias = srcInfExp >> 1;
48 const src_rep_t srcMinNormal = SRC_REP_C(1) << srcSigBits;
49 const src_rep_t srcInfinity = (src_rep_t)srcInfExp << srcSigBits;
50 const src_rep_t srcSignMask = SRC_REP_C(1) << (srcSigBits + srcExpBits);
51 const src_rep_t srcAbsMask = srcSignMask - 1;
52 const src_rep_t srcQNaN = SRC_REP_C(1) << (srcSigBits - 1);
53 const src_rep_t srcNaNCode = srcQNaN - 1;
55 const int dstBits = sizeof(dst_t) * CHAR_BIT;
56 const int dstExpBits = dstBits - dstSigBits - 1;
57 const int dstInfExp = (1 << dstExpBits) - 1;
58 const int dstExpBias = dstInfExp >> 1;
60 const dst_rep_t dstMinNormal = DST_REP_C(1) << dstSigBits;
62 // Break a into a sign and representation of the absolute value.
63 const src_rep_t aRep = srcToRep(a);
64 const src_rep_t aAbs = aRep & srcAbsMask;
65 const src_rep_t sign = aRep & srcSignMask;
68 // If sizeof(src_rep_t) < sizeof(int), the subtraction result is promoted
69 // to (signed) int. To avoid that, explicitly cast to src_rep_t.
70 if ((src_rep_t)(aAbs - srcMinNormal) < srcInfinity - srcMinNormal) {
71 // a is a normal number.
72 // Extend to the destination type by shifting the significand and
73 // exponent into the proper position and rebiasing the exponent.
74 absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits);
75 absResult += (dst_rep_t)(dstExpBias - srcExpBias) << dstSigBits;
78 else if (aAbs >= srcInfinity) {
79 // a is NaN or infinity.
80 // Conjure the result by beginning with infinity, then setting the qNaN
81 // bit (if needed) and right-aligning the rest of the trailing NaN
83 absResult = (dst_rep_t)dstInfExp << dstSigBits;
84 absResult |= (dst_rep_t)(aAbs & srcQNaN) << (dstSigBits - srcSigBits);
85 absResult |= (dst_rep_t)(aAbs & srcNaNCode) << (dstSigBits - srcSigBits);
90 // renormalize the significand and clear the leading bit, then insert
91 // the correct adjusted exponent in the destination type.
92 const int scale = src_rep_t_clz(aAbs) - src_rep_t_clz(srcMinNormal);
93 absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits + scale);
94 absResult ^= dstMinNormal;
95 const int resultExponent = dstExpBias - srcExpBias - scale + 1;
96 absResult |= (dst_rep_t)resultExponent << dstSigBits;
104 // Apply the signbit to the absolute value.
105 const dst_rep_t result = absResult | (dst_rep_t)sign << (dstBits - srcBits);
106 return dstFromRep(result);