extendsfdf2.c revision 214152
1//===-- lib/extendsfdf2.c - single -> double conversion -----------*- C -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements a fairly generic conversion from a narrower to a wider 11// IEEE-754 floating-point type. The constants and types defined following the 12// includes below parameterize the conversion. 13// 14// This routine can be trivially adapted to support conversions from 15// half-precision or to quad-precision. It does not support types that don't 16// use the usual IEEE-754 interchange formats; specifically, some work would be 17// needed to adapt it to (for example) the Intel 80-bit format or PowerPC 18// double-double format. 19// 20// Note please, however, that this implementation is only intended to support 21// *widening* operations; if you need to convert to a *narrower* floating-point 22// type (e.g. double -> float), then this routine will not do what you want it 23// to. 24// 25// It also requires that integer types at least as large as both formats 26// are available on the target platform; this may pose a problem when trying 27// to add support for quad on some 32-bit systems, for example. You also may 28// run into trouble finding an appropriate CLZ function for wide source types; 29// you will likely need to roll your own on some platforms. 30// 31// Finally, the following assumptions are made: 32// 33// 1. floating-point types and integer types have the same endianness on the 34// target platform 35// 36// 2. quiet NaNs, if supported, are indicated by the leading bit of the 37// significand field being set 38// 39//===----------------------------------------------------------------------===// 40 41#include <stdint.h> 42#include <limits.h> 43 44typedef float src_t; 45typedef uint32_t src_rep_t; 46#define SRC_REP_C UINT32_C 47static const int srcSigBits = 23; 48#define src_rep_t_clz __builtin_clz 49 50typedef double dst_t; 51typedef uint64_t dst_rep_t; 52#define DST_REP_C UINT64_C 53static const int dstSigBits = 52; 54 55// End of specialization parameters. Two helper routines for conversion to and 56// from the representation of floating-point data as integer values follow. 57 58static inline src_rep_t srcToRep(src_t x) { 59 const union { src_t f; src_rep_t i; } rep = {.f = x}; 60 return rep.i; 61} 62 63static inline dst_t dstFromRep(dst_rep_t x) { 64 const union { dst_t f; dst_rep_t i; } rep = {.i = x}; 65 return rep.f; 66} 67 68// End helper routines. Conversion implementation follows. 69 70dst_t __extendsfdf2(src_t a) { 71 72 // Various constants whose values follow from the type parameters. 73 // Any reasonable optimizer will fold and propagate all of these. 74 const int srcBits = sizeof(src_t)*CHAR_BIT; 75 const int srcExpBits = srcBits - srcSigBits - 1; 76 const int srcInfExp = (1 << srcExpBits) - 1; 77 const int srcExpBias = srcInfExp >> 1; 78 79 const src_rep_t srcMinNormal = SRC_REP_C(1) << srcSigBits; 80 const src_rep_t srcInfinity = (src_rep_t)srcInfExp << srcSigBits; 81 const src_rep_t srcSignMask = SRC_REP_C(1) << (srcSigBits + srcExpBits); 82 const src_rep_t srcAbsMask = srcSignMask - 1; 83 const src_rep_t srcQNaN = SRC_REP_C(1) << (srcSigBits - 1); 84 const src_rep_t srcNaNCode = srcQNaN - 1; 85 86 const int dstBits = sizeof(dst_t)*CHAR_BIT; 87 const int dstExpBits = dstBits - dstSigBits - 1; 88 const int dstInfExp = (1 << dstExpBits) - 1; 89 const int dstExpBias = dstInfExp >> 1; 90 91 const dst_rep_t dstMinNormal = DST_REP_C(1) << dstSigBits; 92 93 // Break a into a sign and representation of the absolute value 94 const src_rep_t aRep = srcToRep(a); 95 const src_rep_t aAbs = aRep & srcAbsMask; 96 const src_rep_t sign = aRep & srcSignMask; 97 dst_rep_t absResult; 98 99 if (aAbs - srcMinNormal < srcInfinity - srcMinNormal) { 100 // a is a normal number. 101 // Extend to the destination type by shifting the significand and 102 // exponent into the proper position and rebiasing the exponent. 103 absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits); 104 absResult += (dst_rep_t)(dstExpBias - srcExpBias) << dstSigBits; 105 } 106 107 else if (aAbs >= srcInfinity) { 108 // a is NaN or infinity. 109 // Conjure the result by beginning with infinity, then setting the qNaN 110 // bit (if needed) and right-aligning the rest of the trailing NaN 111 // payload field. 112 absResult = (dst_rep_t)dstInfExp << dstSigBits; 113 absResult |= (dst_rep_t)(aAbs & srcQNaN) << (dstSigBits - srcSigBits); 114 absResult |= aAbs & srcNaNCode; 115 } 116 117 else if (aAbs) { 118 // a is denormal. 119 // renormalize the significand and clear the leading bit, then insert 120 // the correct adjusted exponent in the destination type. 121 const int scale = src_rep_t_clz(aAbs) - src_rep_t_clz(srcMinNormal); 122 absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits + scale); 123 absResult ^= dstMinNormal; 124 const int resultExponent = dstExpBias - srcExpBias - scale + 1; 125 absResult |= (dst_rep_t)resultExponent << dstSigBits; 126 } 127 128 else { 129 // a is zero. 130 absResult = 0; 131 } 132 133 // Apply the signbit to (dst_t)abs(a). 134 const dst_rep_t result = absResult | (dst_rep_t)sign << (dstBits - srcBits); 135 return dstFromRep(result); 136} 137