//===-- A class to store a normalized floating point number -----*- C++ -*-===// // // 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 // //===----------------------------------------------------------------------===// #ifndef LLVM_LIBC_UTILS_FPUTIL_NORMAL_FLOAT_H #define LLVM_LIBC_UTILS_FPUTIL_NORMAL_FLOAT_H #include "FPBits.h" #include "utils/CPP/TypeTraits.h" #include namespace __llvm_libc { namespace fputil { // A class which stores the normalized form of a floating point value. // The special IEEE-754 bits patterns of Zero, infinity and NaNs are // are not handled by this class. // // A normalized floating point number is of this form: // (-1)*sign * 2^exponent * // where is of the form 1.<...>. template struct NormalFloat { static_assert( cpp::IsFloatingPointType::Value, "NormalFloat template parameter has to be a floating point type."); using UIntType = typename FPBits::UIntType; static constexpr UIntType one = (UIntType(1) << MantissaWidth::value); // Unbiased exponent value. int32_t exponent; UIntType mantissa; // We want |UIntType| to have atleast one bit more than the actual mantissa // bit width to accommodate the implicit 1 value. static_assert(sizeof(UIntType) * 8 >= MantissaWidth::value + 1, "Bad type for mantissa in NormalFloat."); bool sign; NormalFloat(int32_t e, UIntType m, bool s) : exponent(e), mantissa(m), sign(s) { if (mantissa >= one) return; unsigned normalizationShift = evaluateNormalizationShift(mantissa); mantissa = mantissa << normalizationShift; exponent -= normalizationShift; } explicit NormalFloat(T x) { initFromBits(FPBits(x)); } explicit NormalFloat(FPBits bits) { initFromBits(bits); } // Compares this normalized number with another normalized number. // Returns -1 is this number is less than |other|, 0 if this number is equal // to |other|, and 1 if this number is greater than |other|. int cmp(const NormalFloat &other) const { if (sign != other.sign) return sign ? -1 : 1; if (exponent > other.exponent) { return sign ? -1 : 1; } else if (exponent == other.exponent) { if (mantissa > other.mantissa) return sign ? -1 : 1; else if (mantissa == other.mantissa) return 0; else return sign ? 1 : -1; } else { return sign ? 1 : -1; } } // Returns a new normalized floating point number which is equal in value // to this number multiplied by 2^e. That is: // new = this * 2^e NormalFloat mul2(int e) const { NormalFloat result = *this; result.exponent += e; return result; } operator T() const { int biasedExponent = exponent + FPBits::exponentBias; // Max exponent is of the form 0xFF...E. That is why -2 and not -1. constexpr int maxExponentValue = (1 << ExponentWidth::value) - 2; if (biasedExponent > maxExponentValue) { return sign ? T(FPBits::negInf()) : T(FPBits::inf()); } FPBits result(T(0.0)); result.setSign(sign); constexpr int subnormalExponent = -FPBits::exponentBias + 1; if (exponent < subnormalExponent) { unsigned shift = subnormalExponent - exponent; // Since exponent > subnormalExponent, shift is strictly greater than // zero. if (shift <= MantissaWidth::value + 1) { // Generate a subnormal number. Might lead to loss of precision. // We round to nearest and round halfway cases to even. const UIntType shiftOutMask = (UIntType(1) << shift) - 1; const UIntType shiftOutValue = mantissa & shiftOutMask; const UIntType halfwayValue = UIntType(1) << (shift - 1); result.setUnbiasedExponent(0); result.setMantissa(mantissa >> shift); UIntType newMantissa = result.getMantissa(); if (shiftOutValue > halfwayValue) { newMantissa += 1; } else if (shiftOutValue == halfwayValue) { // Round to even. if (result.getMantissa() & 0x1) newMantissa += 1; } result.setMantissa(newMantissa); // Adding 1 to mantissa can lead to overflow. This can only happen if // mantissa was all ones (0b111..11). For such a case, we will carry // the overflow into the exponent. if (newMantissa == one) result.setUnbiasedExponent(1); return T(result); } else { return T(result); } } result.setUnbiasedExponent(exponent + FPBits::exponentBias); result.setMantissa(mantissa); return T(result); } private: void initFromBits(FPBits bits) { sign = bits.getSign(); if (bits.isInfOrNaN() || bits.isZero()) { // Ignore special bit patterns. Implementations deal with them separately // anyway so this should not be a problem. exponent = 0; mantissa = 0; return; } // Normalize subnormal numbers. if (bits.getUnbiasedExponent() == 0) { unsigned shift = evaluateNormalizationShift(bits.getMantissa()); mantissa = UIntType(bits.getMantissa()) << shift; exponent = 1 - FPBits::exponentBias - shift; } else { exponent = bits.getUnbiasedExponent() - FPBits::exponentBias; mantissa = one | bits.getMantissa(); } } unsigned evaluateNormalizationShift(UIntType m) { unsigned shift = 0; for (; (one & m) == 0 && (shift < MantissaWidth::value); m <<= 1, ++shift) ; return shift; } }; #ifdef SPECIAL_X86_LONG_DOUBLE template <> inline void NormalFloat::initFromBits(FPBits bits) { sign = bits.getSign(); if (bits.isInfOrNaN() || bits.isZero()) { // Ignore special bit patterns. Implementations deal with them separately // anyway so this should not be a problem. exponent = 0; mantissa = 0; return; } if (bits.getUnbiasedExponent() == 0) { if (bits.getImplicitBit() == 0) { // Since we ignore zero value, the mantissa in this case is non-zero. int normalizationShift = evaluateNormalizationShift(bits.getMantissa()); exponent = -16382 - normalizationShift; mantissa = (bits.getMantissa() << normalizationShift); } else { exponent = -16382; mantissa = one | bits.getMantissa(); } } else { if (bits.getImplicitBit() == 0) { // Invalid number so just store 0 similar to a NaN. exponent = 0; mantissa = 0; } else { exponent = bits.getUnbiasedExponent() - 16383; mantissa = one | bits.getMantissa(); } } } template <> inline NormalFloat::operator long double() const { int biasedExponent = exponent + FPBits::exponentBias; // Max exponent is of the form 0xFF...E. That is why -2 and not -1. constexpr int maxExponentValue = (1 << ExponentWidth::value) - 2; if (biasedExponent > maxExponentValue) { return sign ? FPBits::negInf() : FPBits::inf(); } FPBits result(0.0l); result.setSign(sign); constexpr int subnormalExponent = -FPBits::exponentBias + 1; if (exponent < subnormalExponent) { unsigned shift = subnormalExponent - exponent; if (shift <= MantissaWidth::value + 1) { // Generate a subnormal number. Might lead to loss of precision. // We round to nearest and round halfway cases to even. const UIntType shiftOutMask = (UIntType(1) << shift) - 1; const UIntType shiftOutValue = mantissa & shiftOutMask; const UIntType halfwayValue = UIntType(1) << (shift - 1); result.setUnbiasedExponent(0); result.setMantissa(mantissa >> shift); UIntType newMantissa = result.getMantissa(); if (shiftOutValue > halfwayValue) { newMantissa += 1; } else if (shiftOutValue == halfwayValue) { // Round to even. if (result.getMantissa() & 0x1) newMantissa += 1; } result.setMantissa(newMantissa); // Adding 1 to mantissa can lead to overflow. This can only happen if // mantissa was all ones (0b111..11). For such a case, we will carry // the overflow into the exponent and set the implicit bit to 1. if (newMantissa == one) { result.setUnbiasedExponent(1); result.setImplicitBit(1); } else { result.setImplicitBit(0); } return static_cast(result); } else { return static_cast(result); } } result.setUnbiasedExponent(biasedExponent); result.setMantissa(mantissa); result.setImplicitBit(1); return static_cast(result); } #endif // SPECIAL_X86_LONG_DOUBLE } // namespace fputil } // namespace __llvm_libc #endif // LLVM_LIBC_UTILS_FPUTIL_NORMAL_FLOAT_H