//===-- Square root of IEEE 754 floating point numbers ----------*- 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_SQRT_H #define LLVM_LIBC_UTILS_FPUTIL_SQRT_H #include "FPBits.h" #include "utils/CPP/TypeTraits.h" namespace __llvm_libc { namespace fputil { namespace internal { template static inline void normalize(int &exponent, typename FPBits::UIntType &mantissa); template <> inline void normalize(int &exponent, uint32_t &mantissa) { // Use binary search to shift the leading 1 bit. // With MantissaWidth = 23, it will take // ceil(log2(23)) = 5 steps checking the mantissa bits as followed: // Step 1: 0000 0000 0000 XXXX XXXX XXXX // Step 2: 0000 00XX XXXX XXXX XXXX XXXX // Step 3: 000X XXXX XXXX XXXX XXXX XXXX // Step 4: 00XX XXXX XXXX XXXX XXXX XXXX // Step 5: 0XXX XXXX XXXX XXXX XXXX XXXX constexpr int nsteps = 5; // = ceil(log2(MantissaWidth)) constexpr uint32_t bounds[nsteps] = {1 << 12, 1 << 18, 1 << 21, 1 << 22, 1 << 23}; constexpr int shifts[nsteps] = {12, 6, 3, 2, 1}; for (int i = 0; i < nsteps; ++i) { if (mantissa < bounds[i]) { exponent -= shifts[i]; mantissa <<= shifts[i]; } } } template <> inline void normalize(int &exponent, uint64_t &mantissa) { // Use binary search to shift the leading 1 bit similar to float. // With MantissaWidth = 52, it will take // ceil(log2(52)) = 6 steps checking the mantissa bits. constexpr int nsteps = 6; // = ceil(log2(MantissaWidth)) constexpr uint64_t bounds[nsteps] = {1ULL << 26, 1ULL << 39, 1ULL << 46, 1ULL << 49, 1ULL << 51, 1ULL << 52}; constexpr int shifts[nsteps] = {27, 14, 7, 4, 2, 1}; for (int i = 0; i < nsteps; ++i) { if (mantissa < bounds[i]) { exponent -= shifts[i]; mantissa <<= shifts[i]; } } } #if !(defined(__x86_64__) || defined(__i386__)) template <> inline void normalize(int &exponent, __uint128_t &mantissa) { // Use binary search to shift the leading 1 bit similar to float. // With MantissaWidth = 112, it will take // ceil(log2(112)) = 7 steps checking the mantissa bits. constexpr int nsteps = 7; // = ceil(log2(MantissaWidth)) constexpr __uint128_t bounds[nsteps] = { __uint128_t(1) << 56, __uint128_t(1) << 84, __uint128_t(1) << 98, __uint128_t(1) << 105, __uint128_t(1) << 109, __uint128_t(1) << 111, __uint128_t(1) << 112}; constexpr int shifts[nsteps] = {57, 29, 15, 8, 4, 2, 1}; for (int i = 0; i < nsteps; ++i) { if (mantissa < bounds[i]) { exponent -= shifts[i]; mantissa <<= shifts[i]; } } } #endif } // namespace internal // Correctly rounded IEEE 754 SQRT with round to nearest, ties to even. // Shift-and-add algorithm. template ::Value, int> = 0> static inline T sqrt(T x) { using UIntType = typename FPBits::UIntType; constexpr UIntType One = UIntType(1) << MantissaWidth::value; FPBits bits(x); if (bits.isInfOrNaN()) { if (bits.sign && (bits.mantissa == 0)) { // sqrt(-Inf) = NaN return FPBits::buildNaN(One >> 1); } else { // sqrt(NaN) = NaN // sqrt(+Inf) = +Inf return x; } } else if (bits.isZero()) { // sqrt(+0) = +0 // sqrt(-0) = -0 return x; } else if (bits.sign) { // sqrt( negative numbers ) = NaN return FPBits::buildNaN(One >> 1); } else { int xExp = bits.getExponent(); UIntType xMant = bits.mantissa; // Step 1a: Normalize denormal input and append hiddent bit to the mantissa if (bits.exponent == 0) { ++xExp; // let xExp be the correct exponent of One bit. internal::normalize(xExp, xMant); } else { xMant |= One; } // Step 1b: Make sure the exponent is even. if (xExp & 1) { --xExp; xMant <<= 1; } // After step 1b, x = 2^(xExp) * xMant, where xExp is even, and // 1 <= xMant < 4. So sqrt(x) = 2^(xExp / 2) * y, with 1 <= y < 2. // Notice that the output of sqrt is always in the normal range. // To perform shift-and-add algorithm to find y, let denote: // y(n) = 1.y_1 y_2 ... y_n, we can define the nth residue to be: // r(n) = 2^n ( xMant - y(n)^2 ). // That leads to the following recurrence formula: // r(n) = 2*r(n-1) - y_n*[ 2*y(n-1) + 2^(-n-1) ] // with the initial conditions: y(0) = 1, and r(0) = x - 1. // So the nth digit y_n of the mantissa of sqrt(x) can be found by: // y_n = 1 if 2*r(n-1) >= 2*y(n - 1) + 2^(-n-1) // 0 otherwise. UIntType y = One; UIntType r = xMant - One; for (UIntType current_bit = One >> 1; current_bit; current_bit >>= 1) { r <<= 1; UIntType tmp = (y << 1) + current_bit; // 2*y(n - 1) + 2^(-n-1) if (r >= tmp) { r -= tmp; y += current_bit; } } // We compute one more iteration in order to round correctly. bool lsb = y & 1; // Least significant bit bool rb = false; // Round bit r <<= 2; UIntType tmp = (y << 2) + 1; if (r >= tmp) { r -= tmp; rb = true; } // Remove hidden bit and append the exponent field. xExp = ((xExp >> 1) + FPBits::exponentBias); y = (y - One) | (static_cast(xExp) << MantissaWidth::value); // Round to nearest, ties to even if (rb && (lsb || (r != 0))) { ++y; } return *reinterpret_cast(&y); } } } // namespace fputil } // namespace __llvm_libc #if (defined(__x86_64__) || defined(__i386__)) #include "SqrtLongDoubleX86.h" #endif // defined(__x86_64__) || defined(__i386__) #endif // LLVM_LIBC_UTILS_FPUTIL_SQRT_H