extendsfdf2.c revision 37b97d1cf4501b94347e0b4e880f4b25825a289f
1//===-- lib/extendsfdf2.c - single -> double conversion -----------*- C -*-===// 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 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 44#include "int_lib.h" 45 46typedef float src_t; 47typedef uint32_t src_rep_t; 48#define SRC_REP_C UINT32_C 49static const int srcSigBits = 23; 50#define src_rep_t_clz __builtin_clz 51 52typedef double dst_t; 53typedef uint64_t dst_rep_t; 54#define DST_REP_C UINT64_C 55static const int dstSigBits = 52; 56 57// End of specialization parameters. Two helper routines for conversion to and 58// from the representation of floating-point data as integer values follow. 59 60static inline src_rep_t srcToRep(src_t x) { 61 const union { src_t f; src_rep_t i; } rep = {.f = x}; 62 return rep.i; 63} 64 65static inline dst_t dstFromRep(dst_rep_t x) { 66 const union { dst_t f; dst_rep_t i; } rep = {.i = x}; 67 return rep.f; 68} 69 70// End helper routines. Conversion implementation follows. 71 72ARM_EABI_FNALIAS(f2d, extendsfdf2); 73 74dst_t __extendsfdf2(src_t a) { 75 76 // Various constants whose values follow from the type parameters. 77 // Any reasonable optimizer will fold and propagate all of these. 78 const int srcBits = sizeof(src_t)*CHAR_BIT; 79 const int srcExpBits = srcBits - srcSigBits - 1; 80 const int srcInfExp = (1 << srcExpBits) - 1; 81 const int srcExpBias = srcInfExp >> 1; 82 83 const src_rep_t srcMinNormal = SRC_REP_C(1) << srcSigBits; 84 const src_rep_t srcInfinity = (src_rep_t)srcInfExp << srcSigBits; 85 const src_rep_t srcSignMask = SRC_REP_C(1) << (srcSigBits + srcExpBits); 86 const src_rep_t srcAbsMask = srcSignMask - 1; 87 const src_rep_t srcQNaN = SRC_REP_C(1) << (srcSigBits - 1); 88 const src_rep_t srcNaNCode = srcQNaN - 1; 89 90 const int dstBits = sizeof(dst_t)*CHAR_BIT; 91 const int dstExpBits = dstBits - dstSigBits - 1; 92 const int dstInfExp = (1 << dstExpBits) - 1; 93 const int dstExpBias = dstInfExp >> 1; 94 95 const dst_rep_t dstMinNormal = DST_REP_C(1) << dstSigBits; 96 97 // Break a into a sign and representation of the absolute value 98 const src_rep_t aRep = srcToRep(a); 99 const src_rep_t aAbs = aRep & srcAbsMask; 100 const src_rep_t sign = aRep & srcSignMask; 101 dst_rep_t absResult; 102 103 if (aAbs - srcMinNormal < srcInfinity - srcMinNormal) { 104 // a is a normal number. 105 // Extend to the destination type by shifting the significand and 106 // exponent into the proper position and rebiasing the exponent. 107 absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits); 108 absResult += (dst_rep_t)(dstExpBias - srcExpBias) << dstSigBits; 109 } 110 111 else if (aAbs >= srcInfinity) { 112 // a is NaN or infinity. 113 // Conjure the result by beginning with infinity, then setting the qNaN 114 // bit (if needed) and right-aligning the rest of the trailing NaN 115 // payload field. 116 absResult = (dst_rep_t)dstInfExp << dstSigBits; 117 absResult |= (dst_rep_t)(aAbs & srcQNaN) << (dstSigBits - srcSigBits); 118 absResult |= aAbs & srcNaNCode; 119 } 120 121 else if (aAbs) { 122 // a is denormal. 123 // renormalize the significand and clear the leading bit, then insert 124 // the correct adjusted exponent in the destination type. 125 const int scale = src_rep_t_clz(aAbs) - src_rep_t_clz(srcMinNormal); 126 absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits + scale); 127 absResult ^= dstMinNormal; 128 const int resultExponent = dstExpBias - srcExpBias - scale + 1; 129 absResult |= (dst_rep_t)resultExponent << dstSigBits; 130 } 131 132 else { 133 // a is zero. 134 absResult = 0; 135 } 136 137 // Apply the signbit to (dst_t)abs(a). 138 const dst_rep_t result = absResult | (dst_rep_t)sign << (dstBits - srcBits); 139 return dstFromRep(result); 140} 141