fp_extend_impl.inc 4.53 KB
//=-lib/fp_extend_impl.inc - low precision -> high precision conversion -*-- -//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file implements a fairly generic conversion from a narrower to a wider
// IEEE-754 floating-point type.  The constants and types defined following the
// includes below parameterize the conversion.
//
// It does not support types that don't use the usual IEEE-754 interchange
// formats; specifically, some work would be needed to adapt it to
// (for example) the Intel 80-bit format or PowerPC double-double format.
//
// Note please, however, that this implementation is only intended to support
// *widening* operations; if you need to convert to a *narrower* floating-point
// type (e.g. double -> float), then this routine will not do what you want it
// to.
//
// It also requires that integer types at least as large as both formats
// are available on the target platform; this may pose a problem when trying
// to add support for quad on some 32-bit systems, for example.  You also may
// run into trouble finding an appropriate CLZ function for wide source types;
// you will likely need to roll your own on some platforms.
//
// Finally, the following assumptions are made:
//
// 1. Floating-point types and integer types have the same endianness on the
//    target platform.
//
// 2. Quiet NaNs, if supported, are indicated by the leading bit of the
//    significand field being set.
//
//===----------------------------------------------------------------------===//

#include "fp_extend.h"

static __inline dst_t __extendXfYf2__(src_t a) {
  // Various constants whose values follow from the type parameters.
  // Any reasonable optimizer will fold and propagate all of these.
  const int srcBits = sizeof(src_t) * CHAR_BIT;
  const int srcExpBits = srcBits - srcSigBits - 1;
  const int srcInfExp = (1 << srcExpBits) - 1;
  const int srcExpBias = srcInfExp >> 1;

  const src_rep_t srcMinNormal = SRC_REP_C(1) << srcSigBits;
  const src_rep_t srcInfinity = (src_rep_t)srcInfExp << srcSigBits;
  const src_rep_t srcSignMask = SRC_REP_C(1) << (srcSigBits + srcExpBits);
  const src_rep_t srcAbsMask = srcSignMask - 1;
  const src_rep_t srcQNaN = SRC_REP_C(1) << (srcSigBits - 1);
  const src_rep_t srcNaNCode = srcQNaN - 1;

  const int dstBits = sizeof(dst_t) * CHAR_BIT;
  const int dstExpBits = dstBits - dstSigBits - 1;
  const int dstInfExp = (1 << dstExpBits) - 1;
  const int dstExpBias = dstInfExp >> 1;

  const dst_rep_t dstMinNormal = DST_REP_C(1) << dstSigBits;

  // Break a into a sign and representation of the absolute value.
  const src_rep_t aRep = srcToRep(a);
  const src_rep_t aAbs = aRep & srcAbsMask;
  const src_rep_t sign = aRep & srcSignMask;
  dst_rep_t absResult;

  // If sizeof(src_rep_t) < sizeof(int), the subtraction result is promoted
  // to (signed) int.  To avoid that, explicitly cast to src_rep_t.
  if ((src_rep_t)(aAbs - srcMinNormal) < srcInfinity - srcMinNormal) {
    // a is a normal number.
    // Extend to the destination type by shifting the significand and
    // exponent into the proper position and rebiasing the exponent.
    absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits);
    absResult += (dst_rep_t)(dstExpBias - srcExpBias) << dstSigBits;
  }

  else if (aAbs >= srcInfinity) {
    // a is NaN or infinity.
    // Conjure the result by beginning with infinity, then setting the qNaN
    // bit (if needed) and right-aligning the rest of the trailing NaN
    // payload field.
    absResult = (dst_rep_t)dstInfExp << dstSigBits;
    absResult |= (dst_rep_t)(aAbs & srcQNaN) << (dstSigBits - srcSigBits);
    absResult |= (dst_rep_t)(aAbs & srcNaNCode) << (dstSigBits - srcSigBits);
  }

  else if (aAbs) {
    // a is denormal.
    // renormalize the significand and clear the leading bit, then insert
    // the correct adjusted exponent in the destination type.
    const int scale = src_rep_t_clz(aAbs) - src_rep_t_clz(srcMinNormal);
    absResult = (dst_rep_t)aAbs << (dstSigBits - srcSigBits + scale);
    absResult ^= dstMinNormal;
    const int resultExponent = dstExpBias - srcExpBias - scale + 1;
    absResult |= (dst_rep_t)resultExponent << dstSigBits;
  }

  else {
    // a is zero.
    absResult = 0;
  }

  // Apply the signbit to the absolute value.
  const dst_rep_t result = absResult | (dst_rep_t)sign << (dstBits - srcBits);
  return dstFromRep(result);
}