lib/extendsfdf2.c

   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 "int_lib.h"
  42
  43 typedef float src_t;
  44 typedef uint32_t src_rep_t;
  45 #define SRC_REP_C UINT32_C
  46 static const int srcSigBits = 23;
  47 #define src_rep_t_clz __builtin_clz
  48
  49 typedef double dst_t;
  50 typedef uint64_t dst_rep_t;
  51 #define DST_REP_C UINT64_C
  52 static const int dstSigBits = 52;
  53
  54 // End of specialization parameters.  Two helper routines for conversion to and
  55 // from the representation of floating-point data as integer values follow.
  56
  57 static inline src_rep_t srcToRep(src_t x) {
  58     const union { src_t f; src_rep_t i; } rep = {.f = x};
  59     return rep.i;
  60 }
  61
  62 static inline dst_t dstFromRep(dst_rep_t x) {
  63     const union { dst_t f; dst_rep_t i; } rep = {.i = x};
  64     return rep.f;
  65 }
  66
  67 // End helper routines.  Conversion implementation follows.
  68
  69 ARM_EABI_FNALIAS(f2d, extendsfdf2)
  70
  71 dst_t __extendsfdf2(src_t a) {
  72
  73     // Various constants whose values follow from the type parameters.
  74     // Any reasonable optimizer will fold and propagate all of these.
  75     const int srcBits = sizeof(src_t)*CHAR_BIT;
  76     const int srcExpBits = srcBits - srcSigBits - 1;
  77     const int srcInfExp = (1 << srcExpBits) - 1;
  78     const int srcExpBias = srcInfExp >> 1;
  79
  80     const src_rep_t srcMinNormal = SRC_REP_C(1) << srcSigBits;
  81     const src_rep_t srcInfinity = (src_rep_t)srcInfExp << srcSigBits;
  82     const src_rep_t srcSignMask = SRC_REP_C(1) << (srcSigBits + srcExpBits);
  83     const src_rep_t srcAbsMask = srcSignMask - 1;
  84     const src_rep_t srcQNaN = SRC_REP_C(1) << (srcSigBits - 1);
  85     const src_rep_t srcNaNCode = srcQNaN - 1;
  86
  87     const int dstBits = sizeof(dst_t)*CHAR_BIT;
  88     const int dstExpBits = dstBits - dstSigBits - 1;
  89     const int dstInfExp = (1 << dstExpBits) - 1;
  90     const int dstExpBias = dstInfExp >> 1;
  91
  92     const dst_rep_t dstMinNormal = DST_REP_C(1) << dstSigBits;
  93
  94     // Break a into a sign and representation of the absolute value
  95     const src_rep_t aRep = srcToRep(a);
  96     const src_rep_t aAbs = aRep & srcAbsMask;
  97     const src_rep_t sign = aRep & srcSignMask;
  98     dst_rep_t absResult;
  99
 100     if (aAbs - srcMinNormal < srcInfinity - srcMinNormal) {
 101         // a is a normal number.
 102         // Extend to the destination type by shifting the significand and
 103         // exponent into the proper position and rebiasing the exponent.
 104         absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits);
 105         absResult += (dst_rep_t)(dstExpBias - srcExpBias) << dstSigBits;
 106     }
 107
 108     else if (aAbs >= srcInfinity) {
 109         // a is NaN or infinity.
 110         // Conjure the result by beginning with infinity, then setting the qNaN
 111         // bit (if needed) and right-aligning the rest of the trailing NaN
 112         // payload field.
 113         absResult = (dst_rep_t)dstInfExp << dstSigBits;
 114         absResult |= (dst_rep_t)(aAbs & srcQNaN) << (dstSigBits - srcSigBits);
 115         absResult |= aAbs & srcNaNCode;
 116     }
 117
 118     else if (aAbs) {
 119         // a is denormal.
 120         // renormalize the significand and clear the leading bit, then insert
 121         // the correct adjusted exponent in the destination type.
 122         const int scale = src_rep_t_clz(aAbs) - src_rep_t_clz(srcMinNormal);
 123         absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits + scale);
 124         absResult ^= dstMinNormal;
 125         const int resultExponent = dstExpBias - srcExpBias - scale + 1;
 126         absResult |= (dst_rep_t)resultExponent << dstSigBits;
 127     }
 128
 129     else {
 130         // a is zero.
 131         absResult = 0;
 132     }
 133
 134     // Apply the signbit to (dst_t)abs(a).
 135     const dst_rep_t result = absResult | (dst_rep_t)sign << (dstBits - srcBits);
 136     return dstFromRep(result);
 137 }