//===-- 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_SRC_SUPPORT_FPUTIL_NORMAL_FLOAT_H #define LLVM_LIBC_SRC_SUPPORT_FPUTIL_NORMAL_FLOAT_H #include "FPBits.h" #include "src/__support/CPP/type_traits.h" #include "src/__support/common.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::is_floating_point_v, "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; LIBC_INLINE NormalFloat(int32_t e, UIntType m, bool s) : exponent(e), mantissa(m), sign(s) { if (mantissa >= ONE) return; unsigned normalization_shift = evaluate_normalization_shift(mantissa); mantissa = mantissa << normalization_shift; exponent -= normalization_shift; } LIBC_INLINE explicit NormalFloat(T x) { init_from_bits(FPBits(x)); } LIBC_INLINE explicit NormalFloat(FPBits bits) { init_from_bits(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|. LIBC_INLINE 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 LIBC_INLINE NormalFloat mul2(int e) const { NormalFloat result = *this; result.exponent += e; return result; } LIBC_INLINE operator T() const { int biased_exponent = exponent + FPBits::EXPONENT_BIAS; // Max exponent is of the form 0xFF...E. That is why -2 and not -1. constexpr int MAX_EXPONENT_VALUE = (1 << ExponentWidth::VALUE) - 2; if (biased_exponent > MAX_EXPONENT_VALUE) { return sign ? T(FPBits::neg_inf()) : T(FPBits::inf()); } FPBits result(T(0.0)); result.set_sign(sign); constexpr int SUBNORMAL_EXPONENT = -FPBits::EXPONENT_BIAS + 1; if (exponent < SUBNORMAL_EXPONENT) { unsigned shift = SUBNORMAL_EXPONENT - 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 shift_out_mask = (UIntType(1) << shift) - 1; const UIntType shift_out_value = mantissa & shift_out_mask; const UIntType halfway_value = UIntType(1) << (shift - 1); result.set_unbiased_exponent(0); result.set_mantissa(mantissa >> shift); UIntType new_mantissa = result.get_mantissa(); if (shift_out_value > halfway_value) { new_mantissa += 1; } else if (shift_out_value == halfway_value) { // Round to even. if (result.get_mantissa() & 0x1) new_mantissa += 1; } result.set_mantissa(new_mantissa); // 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 (new_mantissa == ONE) result.set_unbiased_exponent(1); return T(result); } else { return T(result); } } result.set_unbiased_exponent(exponent + FPBits::EXPONENT_BIAS); result.set_mantissa(mantissa); return T(result); } private: LIBC_INLINE void init_from_bits(FPBits bits) { sign = bits.get_sign(); if (bits.is_inf_or_nan() || bits.is_zero()) { // 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.get_unbiased_exponent() == 0) { unsigned shift = evaluate_normalization_shift(bits.get_mantissa()); mantissa = UIntType(bits.get_mantissa()) << shift; exponent = 1 - FPBits::EXPONENT_BIAS - shift; } else { exponent = bits.get_unbiased_exponent() - FPBits::EXPONENT_BIAS; mantissa = ONE | bits.get_mantissa(); } } LIBC_INLINE unsigned evaluate_normalization_shift(UIntType m) { unsigned shift = 0; for (; (ONE & m) == 0 && (shift < MantissaWidth::VALUE); m <<= 1, ++shift) ; return shift; } }; #ifdef SPECIAL_X86_LONG_DOUBLE template <> LIBC_INLINE void NormalFloat::init_from_bits(FPBits bits) { sign = bits.get_sign(); if (bits.is_inf_or_nan() || bits.is_zero()) { // Ignore special bit patterns. Implementations deal with them separately // anyway so this should not be a problem. exponent = 0; mantissa = 0; return; } if (bits.get_unbiased_exponent() == 0) { if (bits.get_implicit_bit() == 0) { // Since we ignore zero value, the mantissa in this case is non-zero. int normalization_shift = evaluate_normalization_shift(bits.get_mantissa()); exponent = -16382 - normalization_shift; mantissa = (bits.get_mantissa() << normalization_shift); } else { exponent = -16382; mantissa = ONE | bits.get_mantissa(); } } else { if (bits.get_implicit_bit() == 0) { // Invalid number so just store 0 similar to a NaN. exponent = 0; mantissa = 0; } else { exponent = bits.get_unbiased_exponent() - 16383; mantissa = ONE | bits.get_mantissa(); } } } template <> LIBC_INLINE NormalFloat::operator long double() const { int biased_exponent = exponent + FPBits::EXPONENT_BIAS; // Max exponent is of the form 0xFF...E. That is why -2 and not -1. constexpr int MAX_EXPONENT_VALUE = (1 << ExponentWidth::VALUE) - 2; if (biased_exponent > MAX_EXPONENT_VALUE) { return sign ? FPBits::neg_inf() : FPBits::inf(); } FPBits result(0.0l); result.set_sign(sign); constexpr int SUBNORMAL_EXPONENT = -FPBits::EXPONENT_BIAS + 1; if (exponent < SUBNORMAL_EXPONENT) { unsigned shift = SUBNORMAL_EXPONENT - 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 shift_out_mask = (UIntType(1) << shift) - 1; const UIntType shift_out_value = mantissa & shift_out_mask; const UIntType halfway_value = UIntType(1) << (shift - 1); result.set_unbiased_exponent(0); result.set_mantissa(mantissa >> shift); UIntType new_mantissa = result.get_mantissa(); if (shift_out_value > halfway_value) { new_mantissa += 1; } else if (shift_out_value == halfway_value) { // Round to even. if (result.get_mantissa() & 0x1) new_mantissa += 1; } result.set_mantissa(new_mantissa); // 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 (new_mantissa == ONE) { result.set_unbiased_exponent(1); result.set_implicit_bit(1); } else { result.set_implicit_bit(0); } return static_cast(result); } else { return static_cast(result); } } result.set_unbiased_exponent(biased_exponent); result.set_mantissa(mantissa); result.set_implicit_bit(1); return static_cast(result); } #endif // SPECIAL_X86_LONG_DOUBLE } // namespace fputil } // namespace __llvm_libc #endif // LLVM_LIBC_SRC_SUPPORT_FPUTIL_NORMAL_FLOAT_H