| 1 | #ifndef BOOST_UNORDERED_DETAIL_MULX_HPP |
| 2 | #define BOOST_UNORDERED_DETAIL_MULX_HPP |
| 3 | |
| 4 | // Copyright 2022 Peter Dimov. |
| 5 | // Copyright 2022 Joaquin M Lopez Munoz. |
| 6 | // Distributed under the Boost Software License, Version 1.0. |
| 7 | // https://www.boost.org/LICENSE_1_0.txt) |
| 8 | |
| 9 | #include <boost/cstdint.hpp> |
| 10 | #include <climits> |
| 11 | #include <cstddef> |
| 12 | |
| 13 | #if defined(_MSC_VER) && !defined(__clang__) |
| 14 | # include <intrin.h> |
| 15 | #endif |
| 16 | |
| 17 | namespace boost { |
| 18 | namespace unordered { |
| 19 | namespace detail { |
| 20 | |
| 21 | // Bit mixer based on the mulx primitive |
| 22 | |
| 23 | #if defined(_MSC_VER) && defined(_M_X64) && !defined(__clang__) |
| 24 | |
| 25 | __forceinline boost::uint64_t mulx64( boost::uint64_t x, boost::uint64_t y ) |
| 26 | { |
| 27 | boost::uint64_t r2; |
| 28 | boost::uint64_t r = _umul128( x, y, &r2 ); |
| 29 | return r ^ r2; |
| 30 | } |
| 31 | |
| 32 | #elif defined(_MSC_VER) && defined(_M_ARM64) && !defined(__clang__) |
| 33 | |
| 34 | __forceinline boost::uint64_t mulx64( boost::uint64_t x, boost::uint64_t y ) |
| 35 | { |
| 36 | boost::uint64_t r = x * y; |
| 37 | boost::uint64_t r2 = __umulh( x, y ); |
| 38 | return r ^ r2; |
| 39 | } |
| 40 | |
| 41 | #elif defined(__SIZEOF_INT128__) |
| 42 | |
| 43 | inline boost::uint64_t mulx64( boost::uint64_t x, boost::uint64_t y ) |
| 44 | { |
| 45 | __uint128_t r = (__uint128_t)x * y; |
| 46 | return (boost::uint64_t)r ^ (boost::uint64_t)( r >> 64 ); |
| 47 | } |
| 48 | |
| 49 | #else |
| 50 | |
| 51 | inline boost::uint64_t mulx64( boost::uint64_t x, boost::uint64_t y ) |
| 52 | { |
| 53 | boost::uint64_t x1 = (boost::uint32_t)x; |
| 54 | boost::uint64_t x2 = x >> 32; |
| 55 | |
| 56 | boost::uint64_t y1 = (boost::uint32_t)y; |
| 57 | boost::uint64_t y2 = y >> 32; |
| 58 | |
| 59 | boost::uint64_t r3 = x2 * y2; |
| 60 | |
| 61 | boost::uint64_t r2a = x1 * y2; |
| 62 | |
| 63 | r3 += r2a >> 32; |
| 64 | |
| 65 | boost::uint64_t r2b = x2 * y1; |
| 66 | |
| 67 | r3 += r2b >> 32; |
| 68 | |
| 69 | boost::uint64_t r1 = x1 * y1; |
| 70 | |
| 71 | boost::uint64_t r2 = (r1 >> 32) + (boost::uint32_t)r2a + (boost::uint32_t)r2b; |
| 72 | |
| 73 | r1 = (r2 << 32) + (boost::uint32_t)r1; |
| 74 | r3 += r2 >> 32; |
| 75 | |
| 76 | return r1 ^ r3; |
| 77 | } |
| 78 | |
| 79 | #endif |
| 80 | |
| 81 | inline boost::uint32_t mulx32( boost::uint32_t x, boost::uint32_t y ) |
| 82 | { |
| 83 | boost::uint64_t r = (boost::uint64_t)x * y; |
| 84 | |
| 85 | #if defined(__MSVC_RUNTIME_CHECKS) |
| 86 | |
| 87 | return (boost::uint32_t)(r & UINT32_MAX) ^ (boost::uint32_t)(r >> 32); |
| 88 | |
| 89 | #else |
| 90 | |
| 91 | return (boost::uint32_t)r ^ (boost::uint32_t)(r >> 32); |
| 92 | |
| 93 | #endif |
| 94 | } |
| 95 | |
| 96 | #if defined(SIZE_MAX) |
| 97 | #if ((((SIZE_MAX >> 16) >> 16) >> 16) >> 15) != 0 |
| 98 | #define BOOST_UNORDERED_64B_ARCHITECTURE /* >64 bits assumed as 64 bits */ |
| 99 | #endif |
| 100 | #elif defined(UINTPTR_MAX) /* used as proxy for std::size_t */ |
| 101 | #if ((((UINTPTR_MAX >> 16) >> 16) >> 16) >> 15) != 0 |
| 102 | #define BOOST_UNORDERED_64B_ARCHITECTURE |
| 103 | #endif |
| 104 | #endif |
| 105 | |
| 106 | inline std::size_t mulx( std::size_t x ) noexcept |
| 107 | { |
| 108 | #if defined(BOOST_UNORDERED_64B_ARCHITECTURE) |
| 109 | |
| 110 | // multiplier is phi |
| 111 | return (std::size_t)mulx64( x: (boost::uint64_t)x, y: 0x9E3779B97F4A7C15ull ); |
| 112 | |
| 113 | #else /* 32 bits assumed */ |
| 114 | |
| 115 | // multiplier from https://arxiv.org/abs/2001.05304 |
| 116 | return mulx32( x, 0xE817FB2Du ); |
| 117 | |
| 118 | #endif |
| 119 | } |
| 120 | |
| 121 | #ifdef BOOST_UNORDERED_64B_ARCHITECTURE |
| 122 | #undef BOOST_UNORDERED_64B_ARCHITECTURE |
| 123 | #endif |
| 124 | |
| 125 | } // namespace detail |
| 126 | } // namespace unordered |
| 127 | } // namespace boost |
| 128 | |
| 129 | #endif // #ifndef BOOST_UNORDERED_DETAIL_MULX_HPP |
| 130 | |