89.87% Lines (213/237)
100.00% Functions (36/36)
| TLA | Baseline | Branch | ||||||
|---|---|---|---|---|---|---|---|---|
| Line | Hits | Code | Line | Hits | Code | |||
| 1 | // Copyright 2020-2023 Daniel Lemire | 1 | // Copyright 2020-2023 Daniel Lemire | |||||
| 2 | // Copyright 2023 Matt Borland | 2 | // Copyright 2023 Matt Borland | |||||
| 3 | // Distributed under the Boost Software License, Version 1.0. | 3 | // Distributed under the Boost Software License, Version 1.0. | |||||
| 4 | // https://www.boost.org/LICENSE_1_0.txt | 4 | // https://www.boost.org/LICENSE_1_0.txt | |||||
| 5 | // | 5 | // | |||||
| 6 | // Derivative of: https://github.com/fastfloat/fast_float | 6 | // Derivative of: https://github.com/fastfloat/fast_float | |||||
| 7 | 7 | |||||||
| 8 | #ifndef BOOST_JSON_DETAIL_CHARCONV_DETAIL_FASTFLOAT_BIGINT_HPP | 8 | #ifndef BOOST_JSON_DETAIL_CHARCONV_DETAIL_FASTFLOAT_BIGINT_HPP | |||||
| 9 | #define BOOST_JSON_DETAIL_CHARCONV_DETAIL_FASTFLOAT_BIGINT_HPP | 9 | #define BOOST_JSON_DETAIL_CHARCONV_DETAIL_FASTFLOAT_BIGINT_HPP | |||||
| 10 | 10 | |||||||
| 11 | #include <boost/json/detail/charconv/detail/fast_float/float_common.hpp> | 11 | #include <boost/json/detail/charconv/detail/fast_float/float_common.hpp> | |||||
| 12 | #include <algorithm> | 12 | #include <algorithm> | |||||
| 13 | #include <cstdint> | 13 | #include <cstdint> | |||||
| 14 | #include <climits> | 14 | #include <climits> | |||||
| 15 | #include <cstring> | 15 | #include <cstring> | |||||
| 16 | 16 | |||||||
| 17 | namespace boost { namespace json { namespace detail { namespace charconv { namespace detail { namespace fast_float { | 17 | namespace boost { namespace json { namespace detail { namespace charconv { namespace detail { namespace fast_float { | |||||
| 18 | 18 | |||||||
| 19 | // the limb width: we want efficient multiplication of double the bits in | 19 | // the limb width: we want efficient multiplication of double the bits in | |||||
| 20 | // limb, or for 64-bit limbs, at least 64-bit multiplication where we can | 20 | // limb, or for 64-bit limbs, at least 64-bit multiplication where we can | |||||
| 21 | // extract the high and low parts efficiently. this is every 64-bit | 21 | // extract the high and low parts efficiently. this is every 64-bit | |||||
| 22 | // architecture except for sparc, which emulates 128-bit multiplication. | 22 | // architecture except for sparc, which emulates 128-bit multiplication. | |||||
| 23 | // we might have platforms where `CHAR_BIT` is not 8, so let's avoid | 23 | // we might have platforms where `CHAR_BIT` is not 8, so let's avoid | |||||
| 24 | // doing `8 * sizeof(limb)`. | 24 | // doing `8 * sizeof(limb)`. | |||||
| 25 | #if defined(BOOST_JSON_FASTFLOAT_64BIT) && !defined(__sparc) | 25 | #if defined(BOOST_JSON_FASTFLOAT_64BIT) && !defined(__sparc) | |||||
| 26 | #define BOOST_JSON_FASTFLOAT_64BIT_LIMB 1 | 26 | #define BOOST_JSON_FASTFLOAT_64BIT_LIMB 1 | |||||
| 27 | typedef uint64_t limb; | 27 | typedef uint64_t limb; | |||||
| 28 | constexpr size_t limb_bits = 64; | 28 | constexpr size_t limb_bits = 64; | |||||
| 29 | #else | 29 | #else | |||||
| 30 | #define BOOST_JSON_FASTFLOAT_32BIT_LIMB | 30 | #define BOOST_JSON_FASTFLOAT_32BIT_LIMB | |||||
| 31 | typedef uint32_t limb; | 31 | typedef uint32_t limb; | |||||
| 32 | constexpr size_t limb_bits = 32; | 32 | constexpr size_t limb_bits = 32; | |||||
| 33 | #endif | 33 | #endif | |||||
| 34 | 34 | |||||||
| 35 | typedef span<limb> limb_span; | 35 | typedef span<limb> limb_span; | |||||
| 36 | 36 | |||||||
| 37 | // number of bits in a bigint. this needs to be at least the number | 37 | // number of bits in a bigint. this needs to be at least the number | |||||
| 38 | // of bits required to store the largest bigint, which is | 38 | // of bits required to store the largest bigint, which is | |||||
| 39 | // `log2(10**(digits + max_exp))`, or `log2(10**(767 + 342))`, or | 39 | // `log2(10**(digits + max_exp))`, or `log2(10**(767 + 342))`, or | |||||
| 40 | // ~3600 bits, so we round to 4000. | 40 | // ~3600 bits, so we round to 4000. | |||||
| 41 | constexpr size_t bigint_bits = 4000; | 41 | constexpr size_t bigint_bits = 4000; | |||||
| 42 | constexpr size_t bigint_limbs = bigint_bits / limb_bits; | 42 | constexpr size_t bigint_limbs = bigint_bits / limb_bits; | |||||
| 43 | 43 | |||||||
| 44 | // vector-like type that is allocated on the stack. the entire | 44 | // vector-like type that is allocated on the stack. the entire | |||||
| 45 | // buffer is pre-allocated, and only the length changes. | 45 | // buffer is pre-allocated, and only the length changes. | |||||
| 46 | template <uint16_t size> | 46 | template <uint16_t size> | |||||
| 47 | struct stackvec { | 47 | struct stackvec { | |||||
| 48 | limb data[size]; | 48 | limb data[size]; | |||||
| 49 | // we never need more than 150 limbs | 49 | // we never need more than 150 limbs | |||||
| 50 | uint16_t length{0}; | 50 | uint16_t length{0}; | |||||
| 51 | 51 | |||||||
| HITCBC | 52 | 12693 | stackvec() = default; | 52 | 12693 | stackvec() = default; | ||
| 53 | stackvec(const stackvec &) = delete; | 53 | stackvec(const stackvec &) = delete; | |||||
| 54 | stackvec &operator=(const stackvec &) = delete; | 54 | stackvec &operator=(const stackvec &) = delete; | |||||
| 55 | stackvec(stackvec &&) = delete; | 55 | stackvec(stackvec &&) = delete; | |||||
| 56 | stackvec &operator=(stackvec &&other) = delete; | 56 | stackvec &operator=(stackvec &&other) = delete; | |||||
| 57 | 57 | |||||||
| 58 | // create stack vector from existing limb span. | 58 | // create stack vector from existing limb span. | |||||
| HITCBC | 59 | 2024 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 stackvec(limb_span s) { | 59 | 2024 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 stackvec(limb_span s) { | ||
| HITCBC | 60 | 2024 | try_extend(s); | 60 | 2024 | try_extend(s); | ||
| HITCBC | 61 | 2024 | } | 61 | 2024 | } | ||
| 62 | 62 | |||||||
| HITCBC | 63 | 255428 | BOOST_JSON_CXX14_CONSTEXPR limb& operator[](size_t index) noexcept { | 63 | 255428 | BOOST_JSON_CXX14_CONSTEXPR limb& operator[](size_t index) noexcept { | ||
| HITCBC | 64 | 255428 | BOOST_ASSERT(index < length); | 64 | 255428 | BOOST_ASSERT(index < length); | ||
| HITCBC | 65 | 255428 | return data[index]; | 65 | 255428 | return data[index]; | ||
| 66 | } | 66 | } | |||||
| HITCBC | 67 | 6198 | BOOST_JSON_CXX14_CONSTEXPR const limb& operator[](size_t index) const noexcept { | 67 | 6198 | BOOST_JSON_CXX14_CONSTEXPR const limb& operator[](size_t index) const noexcept { | ||
| HITCBC | 68 | 6198 | BOOST_ASSERT(index < length); | 68 | 6198 | BOOST_ASSERT(index < length); | ||
| HITCBC | 69 | 6198 | return data[index]; | 69 | 6198 | return data[index]; | ||
| 70 | } | 70 | } | |||||
| 71 | // index from the end of the container | 71 | // index from the end of the container | |||||
| HITCBC | 72 | 9129 | BOOST_JSON_CXX14_CONSTEXPR const limb& rindex(size_t index) const noexcept { | 72 | 9129 | BOOST_JSON_CXX14_CONSTEXPR const limb& rindex(size_t index) const noexcept { | ||
| HITCBC | 73 | 9129 | BOOST_ASSERT(index < length); | 73 | 9129 | BOOST_ASSERT(index < length); | ||
| HITCBC | 74 | 9129 | size_t rindex = length - index - 1; | 74 | 9129 | size_t rindex = length - index - 1; | ||
| HITCBC | 75 | 9129 | return data[rindex]; | 75 | 9129 | return data[rindex]; | ||
| 76 | } | 76 | } | |||||
| 77 | 77 | |||||||
| 78 | // set the length, without bounds checking. | 78 | // set the length, without bounds checking. | |||||
| HITCBC | 79 | 28303 | BOOST_JSON_CXX14_CONSTEXPR void set_len(size_t len) noexcept { | 79 | 28303 | BOOST_JSON_CXX14_CONSTEXPR void set_len(size_t len) noexcept { | ||
| HITCBC | 80 | 28303 | length = uint16_t(len); | 80 | 28303 | length = uint16_t(len); | ||
| HITCBC | 81 | 28303 | } | 81 | 28303 | } | ||
| HITCBC | 82 | 259478 | constexpr size_t len() const noexcept { | 82 | 259478 | constexpr size_t len() const noexcept { | ||
| HITCBC | 83 | 259478 | return length; | 83 | 259478 | return length; | ||
| 84 | } | 84 | } | |||||
| HITCBC | 85 | 3789 | constexpr bool is_empty() const noexcept { | 85 | 3789 | constexpr bool is_empty() const noexcept { | ||
| HITCBC | 86 | 3789 | return length == 0; | 86 | 3789 | return length == 0; | ||
| 87 | } | 87 | } | |||||
| HITCBC | 88 | 45829 | constexpr size_t capacity() const noexcept { | 88 | 45829 | constexpr size_t capacity() const noexcept { | ||
| HITCBC | 89 | 45829 | return size; | 89 | 45829 | return size; | ||
| 90 | } | 90 | } | |||||
| 91 | // append item to vector, without bounds checking | 91 | // append item to vector, without bounds checking | |||||
| HITCBC | 92 | 27233 | BOOST_JSON_CXX14_CONSTEXPR void push_unchecked(limb value) noexcept { | 92 | 27233 | BOOST_JSON_CXX14_CONSTEXPR void push_unchecked(limb value) noexcept { | ||
| HITCBC | 93 | 27233 | data[length] = value; | 93 | 27233 | data[length] = value; | ||
| HITCBC | 94 | 27233 | length++; | 94 | 27233 | length++; | ||
| HITCBC | 95 | 27233 | } | 95 | 27233 | } | ||
| 96 | // append item to vector, returning if item was added | 96 | // append item to vector, returning if item was added | |||||
| HITCBC | 97 | 25622 | BOOST_JSON_CXX14_CONSTEXPR bool try_push(limb value) noexcept { | 97 | 25622 | BOOST_JSON_CXX14_CONSTEXPR bool try_push(limb value) noexcept { | ||
| HITCBC | 98 | 25622 | if (len() < capacity()) { | 98 | 25622 | if (len() < capacity()) { | ||
| HITCBC | 99 | 25622 | push_unchecked(value); | 99 | 25622 | push_unchecked(value); | ||
| HITCBC | 100 | 25622 | return true; | 100 | 25622 | return true; | ||
| 101 | } else { | 101 | } else { | |||||
| MISUBC | 102 | ✗ | return false; | 102 | ✗ | return false; | ||
| 103 | } | 103 | } | |||||
| 104 | } | 104 | } | |||||
| 105 | // add items to the vector, from a span, without bounds checking | 105 | // add items to the vector, from a span, without bounds checking | |||||
| HITCBC | 106 | 10120 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 void extend_unchecked(limb_span s) noexcept { | 106 | 10120 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 void extend_unchecked(limb_span s) noexcept { | ||
| HITCBC | 107 | 10120 | limb* ptr = data + length; | 107 | 10120 | limb* ptr = data + length; | ||
| HITCBC | 108 | 10120 | std::copy_n(s.ptr, s.len(), ptr); | 108 | 10120 | std::copy_n(s.ptr, s.len(), ptr); | ||
| HITCBC | 109 | 10120 | set_len(len() + s.len()); | 109 | 10120 | set_len(len() + s.len()); | ||
| HITCBC | 110 | 10120 | } | 110 | 10120 | } | ||
| 111 | // try to add items to the vector, returning if items were added | 111 | // try to add items to the vector, returning if items were added | |||||
| HITCBC | 112 | 10120 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool try_extend(limb_span s) noexcept { | 112 | 10120 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool try_extend(limb_span s) noexcept { | ||
| HITCBC | 113 | 10120 | if (len() + s.len() <= capacity()) { | 113 | 10120 | if (len() + s.len() <= capacity()) { | ||
| HITCBC | 114 | 10120 | extend_unchecked(s); | 114 | 10120 | extend_unchecked(s); | ||
| HITCBC | 115 | 10120 | return true; | 115 | 10120 | return true; | ||
| 116 | } else { | 116 | } else { | |||||
| MISUBC | 117 | ✗ | return false; | 117 | ✗ | return false; | ||
| 118 | } | 118 | } | |||||
| 119 | } | 119 | } | |||||
| 120 | // resize the vector, without bounds checking | 120 | // resize the vector, without bounds checking | |||||
| 121 | // if the new size is longer than the vector, assign value to each | 121 | // if the new size is longer than the vector, assign value to each | |||||
| 122 | // appended item. | 122 | // appended item. | |||||
| 123 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 123 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| HITCBC | 124 | 7673 | void resize_unchecked(size_t new_len, limb value) noexcept { | 124 | 7673 | void resize_unchecked(size_t new_len, limb value) noexcept { | ||
| HITCBC | 125 | 7673 | if (new_len > len()) { | 125 | 7673 | if (new_len > len()) { | ||
| HITCBC | 126 | 7673 | size_t count = new_len - len(); | 126 | 7673 | size_t count = new_len - len(); | ||
| HITCBC | 127 | 7673 | limb* first = data + len(); | 127 | 7673 | limb* first = data + len(); | ||
| HITCBC | 128 | 7673 | limb* last = first + count; | 128 | 7673 | limb* last = first + count; | ||
| HITCBC | 129 | 7673 | ::std::fill(first, last, value); | 129 | 7673 | ::std::fill(first, last, value); | ||
| HITCBC | 130 | 7673 | set_len(new_len); | 130 | 7673 | set_len(new_len); | ||
| 131 | } else { | 131 | } else { | |||||
| MISUBC | 132 | ✗ | set_len(new_len); | 132 | ✗ | set_len(new_len); | ||
| 133 | } | 133 | } | |||||
| HITCBC | 134 | 7673 | } | 134 | 7673 | } | ||
| 135 | // try to resize the vector, returning if the vector was resized. | 135 | // try to resize the vector, returning if the vector was resized. | |||||
| HITCBC | 136 | 7673 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool try_resize(size_t new_len, limb value) noexcept { | 136 | 7673 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool try_resize(size_t new_len, limb value) noexcept { | ||
| HITCBC | 137 | 7673 | if (new_len > capacity()) { | 137 | 7673 | if (new_len > capacity()) { | ||
| MISUBC | 138 | ✗ | return false; | 138 | ✗ | return false; | ||
| 139 | } else { | 139 | } else { | |||||
| HITCBC | 140 | 7673 | resize_unchecked(new_len, value); | 140 | 7673 | resize_unchecked(new_len, value); | ||
| HITCBC | 141 | 7673 | return true; | 141 | 7673 | return true; | ||
| 142 | } | 142 | } | |||||
| 143 | } | 143 | } | |||||
| 144 | // check if any limbs are non-zero after the given index. | 144 | // check if any limbs are non-zero after the given index. | |||||
| 145 | // this needs to be done in reverse order, since the index | 145 | // this needs to be done in reverse order, since the index | |||||
| 146 | // is relative to the most significant limbs. | 146 | // is relative to the most significant limbs. | |||||
| HITCBC | 147 | 1375 | BOOST_JSON_CXX14_CONSTEXPR bool nonzero(size_t index) const noexcept { | 147 | 1375 | BOOST_JSON_CXX14_CONSTEXPR bool nonzero(size_t index) const noexcept { | ||
| HITCBC | 148 | 1375 | while (index < len()) { | 148 | 1375 | while (index < len()) { | ||
| HITCBC | 149 | 1369 | if (rindex(index) != 0) { | 149 | 1369 | if (rindex(index) != 0) { | ||
| HITCBC | 150 | 1369 | return true; | 150 | 1369 | return true; | ||
| 151 | } | 151 | } | |||||
| MISUBC | 152 | ✗ | index++; | 152 | ✗ | index++; | ||
| 153 | } | 153 | } | |||||
| HITCBC | 154 | 6 | return false; | 154 | 6 | return false; | ||
| 155 | } | 155 | } | |||||
| 156 | // normalize the big integer, so most-significant zero limbs are removed. | 156 | // normalize the big integer, so most-significant zero limbs are removed. | |||||
| HITCBC | 157 | 3635 | BOOST_JSON_CXX14_CONSTEXPR void normalize() noexcept { | 157 | 3635 | BOOST_JSON_CXX14_CONSTEXPR void normalize() noexcept { | ||
| HITCBC | 158 | 3635 | while (len() > 0 && rindex(0) == 0) { | 158 | 3635 | while (len() > 0 && rindex(0) == 0) { | ||
| MISUBC | 159 | ✗ | length--; | 159 | ✗ | length--; | ||
| 160 | } | 160 | } | |||||
| HITCBC | 161 | 3635 | } | 161 | 3635 | } | ||
| 162 | }; | 162 | }; | |||||
| 163 | 163 | |||||||
| 164 | BOOST_FORCEINLINE BOOST_JSON_CXX14_CONSTEXPR_NO_INLINE | 164 | BOOST_FORCEINLINE BOOST_JSON_CXX14_CONSTEXPR_NO_INLINE | |||||
| 165 | uint64_t empty_hi64(bool& truncated) noexcept { | 165 | uint64_t empty_hi64(bool& truncated) noexcept { | |||||
| MISUBC | 166 | ✗ | truncated = false; | 166 | ✗ | truncated = false; | ||
| MISUBC | 167 | ✗ | return 0; | 167 | ✗ | return 0; | ||
| 168 | } | 168 | } | |||||
| 169 | 169 | |||||||
| 170 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 170 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| 171 | uint64_t uint64_hi64(uint64_t r0, bool& truncated) noexcept { | 171 | uint64_t uint64_hi64(uint64_t r0, bool& truncated) noexcept { | |||||
| MISUBC | 172 | ✗ | truncated = false; | 172 | ✗ | truncated = false; | ||
| MISUBC | 173 | ✗ | int shl = leading_zeroes(r0); | 173 | ✗ | int shl = leading_zeroes(r0); | ||
| MISUBC | 174 | ✗ | return r0 << shl; | 174 | ✗ | return r0 << shl; | ||
| 175 | } | 175 | } | |||||
| 176 | 176 | |||||||
| 177 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 177 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| 178 | uint64_t uint64_hi64(uint64_t r0, uint64_t r1, bool& truncated) noexcept { | 178 | uint64_t uint64_hi64(uint64_t r0, uint64_t r1, bool& truncated) noexcept { | |||||
| HITCBC | 179 | 1375 | int shl = leading_zeroes(r0); | 179 | 1375 | int shl = leading_zeroes(r0); | ||
| HITCBC | 180 | 1375 | if (shl == 0) { | 180 | 1375 | if (shl == 0) { | ||
| HITCBC | 181 | 27 | truncated = r1 != 0; | 181 | 27 | truncated = r1 != 0; | ||
| HITCBC | 182 | 27 | return r0; | 182 | 27 | return r0; | ||
| 183 | } else { | 183 | } else { | |||||
| HITCBC | 184 | 1348 | int shr = 64 - shl; | 184 | 1348 | int shr = 64 - shl; | ||
| HITCBC | 185 | 1348 | truncated = (r1 << shl) != 0; | 185 | 1348 | truncated = (r1 << shl) != 0; | ||
| HITCBC | 186 | 1348 | return (r0 << shl) | (r1 >> shr); | 186 | 1348 | return (r0 << shl) | (r1 >> shr); | ||
| 187 | } | 187 | } | |||||
| 188 | } | 188 | } | |||||
| 189 | 189 | |||||||
| 190 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 190 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| 191 | uint64_t uint32_hi64(uint32_t r0, bool& truncated) noexcept { | 191 | uint64_t uint32_hi64(uint32_t r0, bool& truncated) noexcept { | |||||
| 192 | return uint64_hi64(r0, truncated); | 192 | return uint64_hi64(r0, truncated); | |||||
| 193 | } | 193 | } | |||||
| 194 | 194 | |||||||
| 195 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 195 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| 196 | uint64_t uint32_hi64(uint32_t r0, uint32_t r1, bool& truncated) noexcept { | 196 | uint64_t uint32_hi64(uint32_t r0, uint32_t r1, bool& truncated) noexcept { | |||||
| 197 | uint64_t x0 = r0; | 197 | uint64_t x0 = r0; | |||||
| 198 | uint64_t x1 = r1; | 198 | uint64_t x1 = r1; | |||||
| 199 | return uint64_hi64((x0 << 32) | x1, truncated); | 199 | return uint64_hi64((x0 << 32) | x1, truncated); | |||||
| 200 | } | 200 | } | |||||
| 201 | 201 | |||||||
| 202 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 202 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| 203 | uint64_t uint32_hi64(uint32_t r0, uint32_t r1, uint32_t r2, bool& truncated) noexcept { | 203 | uint64_t uint32_hi64(uint32_t r0, uint32_t r1, uint32_t r2, bool& truncated) noexcept { | |||||
| 204 | uint64_t x0 = r0; | 204 | uint64_t x0 = r0; | |||||
| 205 | uint64_t x1 = r1; | 205 | uint64_t x1 = r1; | |||||
| 206 | uint64_t x2 = r2; | 206 | uint64_t x2 = r2; | |||||
| 207 | return uint64_hi64(x0, (x1 << 32) | x2, truncated); | 207 | return uint64_hi64(x0, (x1 << 32) | x2, truncated); | |||||
| 208 | } | 208 | } | |||||
| 209 | 209 | |||||||
| 210 | // add two small integers, checking for overflow. | 210 | // add two small integers, checking for overflow. | |||||
| 211 | // we want an efficient operation. for msvc, where | 211 | // we want an efficient operation. for msvc, where | |||||
| 212 | // we don't have built-in intrinsics, this is still | 212 | // we don't have built-in intrinsics, this is still | |||||
| 213 | // pretty fast. | 213 | // pretty fast. | |||||
| 214 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 214 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| 215 | limb scalar_add(limb x, limb y, bool& overflow) noexcept { | 215 | limb scalar_add(limb x, limb y, bool& overflow) noexcept { | |||||
| 216 | limb z; | 216 | limb z; | |||||
| 217 | // gcc and clang | 217 | // gcc and clang | |||||
| 218 | #if defined(__has_builtin) | 218 | #if defined(__has_builtin) | |||||
| 219 | #if __has_builtin(__builtin_add_overflow) | 219 | #if __has_builtin(__builtin_add_overflow) | |||||
| HITCBC | 220 | 38490 | if (!cpp20_and_in_constexpr()) { | 220 | 38490 | if (!cpp20_and_in_constexpr()) { | ||
| HITCBC | 221 | 38490 | overflow = __builtin_add_overflow(x, y, &z); | 221 | 38490 | overflow = __builtin_add_overflow(x, y, &z); | ||
| HITCBC | 222 | 38490 | return z; | 222 | 38490 | return z; | ||
| 223 | } | 223 | } | |||||
| 224 | #endif | 224 | #endif | |||||
| 225 | #endif | 225 | #endif | |||||
| 226 | 226 | |||||||
| 227 | // generic, this still optimizes correctly on MSVC. | 227 | // generic, this still optimizes correctly on MSVC. | |||||
| MISUBC | 228 | ✗ | z = x + y; | 228 | ✗ | z = x + y; | ||
| MISUBC | 229 | ✗ | overflow = z < x; | 229 | ✗ | overflow = z < x; | ||
| MISUBC | 230 | ✗ | return z; | 230 | ✗ | return z; | ||
| 231 | } | 231 | } | |||||
| 232 | 232 | |||||||
| 233 | // multiply two small integers, getting both the high and low bits. | 233 | // multiply two small integers, getting both the high and low bits. | |||||
| 234 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 234 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| 235 | limb scalar_mul(limb x, limb y, limb& carry) noexcept { | 235 | limb scalar_mul(limb x, limb y, limb& carry) noexcept { | |||||
| 236 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | 236 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | |||||
| 237 | #if defined(__SIZEOF_INT128__) | 237 | #if defined(__SIZEOF_INT128__) | |||||
| 238 | // GCC and clang both define it as an extension. | 238 | // GCC and clang both define it as an extension. | |||||
| HITCBC | 239 | 80707 | __uint128_t z = __uint128_t(x) * __uint128_t(y) + __uint128_t(carry); | 239 | 80707 | __uint128_t z = __uint128_t(x) * __uint128_t(y) + __uint128_t(carry); | ||
| HITCBC | 240 | 80707 | carry = limb(z >> limb_bits); | 240 | 80707 | carry = limb(z >> limb_bits); | ||
| HITCBC | 241 | 80707 | return limb(z); | 241 | 80707 | return limb(z); | ||
| 242 | #else | 242 | #else | |||||
| 243 | // fallback, no native 128-bit integer multiplication with carry. | 243 | // fallback, no native 128-bit integer multiplication with carry. | |||||
| 244 | // on msvc, this optimizes identically, somehow. | 244 | // on msvc, this optimizes identically, somehow. | |||||
| 245 | value128 z = full_multiplication(x, y); | 245 | value128 z = full_multiplication(x, y); | |||||
| 246 | bool overflow; | 246 | bool overflow; | |||||
| 247 | z.low = scalar_add(z.low, carry, overflow); | 247 | z.low = scalar_add(z.low, carry, overflow); | |||||
| 248 | z.high += uint64_t(overflow); // cannot overflow | 248 | z.high += uint64_t(overflow); // cannot overflow | |||||
| 249 | carry = z.high; | 249 | carry = z.high; | |||||
| 250 | return z.low; | 250 | return z.low; | |||||
| 251 | #endif | 251 | #endif | |||||
| 252 | #else | 252 | #else | |||||
| 253 | uint64_t z = uint64_t(x) * uint64_t(y) + uint64_t(carry); | 253 | uint64_t z = uint64_t(x) * uint64_t(y) + uint64_t(carry); | |||||
| 254 | carry = limb(z >> limb_bits); | 254 | carry = limb(z >> limb_bits); | |||||
| 255 | return limb(z); | 255 | return limb(z); | |||||
| 256 | #endif | 256 | #endif | |||||
| 257 | } | 257 | } | |||||
| 258 | 258 | |||||||
| 259 | // add scalar value to bigint starting from offset. | 259 | // add scalar value to bigint starting from offset. | |||||
| 260 | // used in grade school multiplication | 260 | // used in grade school multiplication | |||||
| 261 | template <uint16_t size> | 261 | template <uint16_t size> | |||||
| 262 | inline BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 262 | inline BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| HITCBC | 263 | 9437 | bool small_add_from(stackvec<size>& vec, limb y, size_t start) noexcept { | 263 | 9437 | bool small_add_from(stackvec<size>& vec, limb y, size_t start) noexcept { | ||
| HITCBC | 264 | 9437 | size_t index = start; | 264 | 9437 | size_t index = start; | ||
| HITCBC | 265 | 9437 | limb carry = y; | 265 | 9437 | limb carry = y; | ||
| 266 | bool overflow; | 266 | bool overflow; | |||||
| HITCBC | 267 | 16116 | while (carry != 0 && index < vec.len()) { | 267 | 16116 | while (carry != 0 && index < vec.len()) { | ||
| HITCBC | 268 | 6679 | vec[index] = scalar_add(vec[index], carry, overflow); | 268 | 6679 | vec[index] = scalar_add(vec[index], carry, overflow); | ||
| HITCBC | 269 | 6679 | carry = limb(overflow); | 269 | 6679 | carry = limb(overflow); | ||
| HITCBC | 270 | 6679 | index += 1; | 270 | 6679 | index += 1; | ||
| 271 | } | 271 | } | |||||
| HITCBC | 272 | 9437 | if (carry != 0) { | 272 | 9437 | if (carry != 0) { | ||
| HITCBC | 273 | 2986 | BOOST_JSON_FASTFLOAT_TRY(vec.try_push(carry)); | 273 | 2986 | BOOST_JSON_FASTFLOAT_TRY(vec.try_push(carry)); | ||
| 274 | } | 274 | } | |||||
| HITCBC | 275 | 9437 | return true; | 275 | 9437 | return true; | ||
| 276 | } | 276 | } | |||||
| 277 | 277 | |||||||
| 278 | // add scalar value to bigint. | 278 | // add scalar value to bigint. | |||||
| 279 | template <uint16_t size> | 279 | template <uint16_t size> | |||||
| 280 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 280 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| 281 | bool small_add(stackvec<size>& vec, limb y) noexcept { | 281 | bool small_add(stackvec<size>& vec, limb y) noexcept { | |||||
| HITCBC | 282 | 9437 | return small_add_from(vec, y, 0); | 282 | 9437 | return small_add_from(vec, y, 0); | ||
| 283 | } | 283 | } | |||||
| 284 | 284 | |||||||
| 285 | // multiply bigint by scalar value. | 285 | // multiply bigint by scalar value. | |||||
| 286 | template <uint16_t size> | 286 | template <uint16_t size> | |||||
| 287 | inline BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 287 | inline BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| HITCBC | 288 | 28029 | bool small_mul(stackvec<size>& vec, limb y) noexcept { | 288 | 28029 | bool small_mul(stackvec<size>& vec, limb y) noexcept { | ||
| HITCBC | 289 | 28029 | limb carry = 0; | 289 | 28029 | limb carry = 0; | ||
| HITCBC | 290 | 108736 | for (size_t index = 0; index < vec.len(); index++) { | 290 | 108736 | for (size_t index = 0; index < vec.len(); index++) { | ||
| HITCBC | 291 | 161414 | vec[index] = scalar_mul(vec[index], y, carry); | 291 | 161414 | vec[index] = scalar_mul(vec[index], y, carry); | ||
| 292 | } | 292 | } | |||||
| HITCBC | 293 | 28029 | if (carry != 0) { | 293 | 28029 | if (carry != 0) { | ||
| HITCBC | 294 | 21102 | BOOST_JSON_FASTFLOAT_TRY(vec.try_push(carry)); | 294 | 21102 | BOOST_JSON_FASTFLOAT_TRY(vec.try_push(carry)); | ||
| 295 | } | 295 | } | |||||
| HITCBC | 296 | 28029 | return true; | 296 | 28029 | return true; | ||
| 297 | } | 297 | } | |||||
| 298 | 298 | |||||||
| 299 | // add bigint to bigint starting from index. | 299 | // add bigint to bigint starting from index. | |||||
| 300 | // used in grade school multiplication | 300 | // used in grade school multiplication | |||||
| 301 | template <uint16_t size> | 301 | template <uint16_t size> | |||||
| 302 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 302 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| HITCBC | 303 | 8096 | bool large_add_from(stackvec<size>& x, limb_span y, size_t start) noexcept { | 303 | 8096 | bool large_add_from(stackvec<size>& x, limb_span y, size_t start) noexcept { | ||
| 304 | // the effective x buffer is from `xstart..x.len()`, so exit early | 304 | // the effective x buffer is from `xstart..x.len()`, so exit early | |||||
| 305 | // if we can't get that current range. | 305 | // if we can't get that current range. | |||||
| HITCBC | 306 | 8096 | if (x.len() < start || y.len() > x.len() - start) { | 306 | 8096 | if (x.len() < start || y.len() > x.len() - start) { | ||
| HITCBC | 307 | 7673 | BOOST_JSON_FASTFLOAT_TRY(x.try_resize(y.len() + start, 0)); | 307 | 7673 | BOOST_JSON_FASTFLOAT_TRY(x.try_resize(y.len() + start, 0)); | ||
| 308 | } | 308 | } | |||||
| 309 | 309 | |||||||
| HITCBC | 310 | 8096 | bool carry = false; | 310 | 8096 | bool carry = false; | ||
| HITCBC | 311 | 34341 | for (size_t index = 0; index < y.len(); index++) { | 311 | 34341 | for (size_t index = 0; index < y.len(); index++) { | ||
| HITCBC | 312 | 26245 | limb xi = x[index + start]; | 312 | 26245 | limb xi = x[index + start]; | ||
| HITCBC | 313 | 26245 | limb yi = y[index]; | 313 | 26245 | limb yi = y[index]; | ||
| HITCBC | 314 | 26245 | bool c1 = false; | 314 | 26245 | bool c1 = false; | ||
| HITCBC | 315 | 26245 | bool c2 = false; | 315 | 26245 | bool c2 = false; | ||
| HITCBC | 316 | 26245 | xi = scalar_add(xi, yi, c1); | 316 | 26245 | xi = scalar_add(xi, yi, c1); | ||
| HITCBC | 317 | 26245 | if (carry) { | 317 | 26245 | if (carry) { | ||
| HITCBC | 318 | 5566 | xi = scalar_add(xi, 1, c2); | 318 | 5566 | xi = scalar_add(xi, 1, c2); | ||
| 319 | } | 319 | } | |||||
| HITCBC | 320 | 26245 | x[index + start] = xi; | 320 | 26245 | x[index + start] = xi; | ||
| HITCBC | 321 | 26245 | carry = c1 | c2; | 321 | 26245 | carry = c1 | c2; | ||
| 322 | } | 322 | } | |||||
| 323 | 323 | |||||||
| 324 | // handle overflow | 324 | // handle overflow | |||||
| HITCBC | 325 | 8096 | if (carry) { | 325 | 8096 | if (carry) { | ||
| MISUBC | 326 | ✗ | BOOST_JSON_FASTFLOAT_TRY(small_add_from(x, 1, y.len() + start)); | 326 | ✗ | BOOST_JSON_FASTFLOAT_TRY(small_add_from(x, 1, y.len() + start)); | ||
| 327 | } | 327 | } | |||||
| HITCBC | 328 | 8096 | return true; | 328 | 8096 | return true; | ||
| 329 | } | 329 | } | |||||
| 330 | 330 | |||||||
| 331 | // add bigint to bigint. | 331 | // add bigint to bigint. | |||||
| 332 | template <uint16_t size> | 332 | template <uint16_t size> | |||||
| 333 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 333 | BOOST_FORCEINLINE BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| 334 | bool large_add_from(stackvec<size>& x, limb_span y) noexcept { | 334 | bool large_add_from(stackvec<size>& x, limb_span y) noexcept { | |||||
| 335 | return large_add_from(x, y, 0); | 335 | return large_add_from(x, y, 0); | |||||
| 336 | } | 336 | } | |||||
| 337 | 337 | |||||||
| 338 | // grade-school multiplication algorithm | 338 | // grade-school multiplication algorithm | |||||
| 339 | template <uint16_t size> | 339 | template <uint16_t size> | |||||
| 340 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 340 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| HITCBC | 341 | 2024 | bool long_mul(stackvec<size>& x, limb_span y) noexcept { | 341 | 2024 | bool long_mul(stackvec<size>& x, limb_span y) noexcept { | ||
| HITCBC | 342 | 2024 | limb_span xs = limb_span(x.data, x.len()); | 342 | 2024 | limb_span xs = limb_span(x.data, x.len()); | ||
| HITCBC | 343 | 2024 | stackvec<size> z(xs); | 343 | 2024 | stackvec<size> z(xs); | ||
| HITCBC | 344 | 2024 | limb_span zs = limb_span(z.data, z.len()); | 344 | 2024 | limb_span zs = limb_span(z.data, z.len()); | ||
| 345 | 345 | |||||||
| HITCBC | 346 | 2024 | if (y.len() != 0) { | 346 | 2024 | if (y.len() != 0) { | ||
| HITCBC | 347 | 2024 | limb y0 = y[0]; | 347 | 2024 | limb y0 = y[0]; | ||
| HITCBC | 348 | 2024 | BOOST_JSON_FASTFLOAT_TRY(small_mul(x, y0)); | 348 | 2024 | BOOST_JSON_FASTFLOAT_TRY(small_mul(x, y0)); | ||
| HITCBC | 349 | 10120 | for (size_t index = 1; index < y.len(); index++) { | 349 | 10120 | for (size_t index = 1; index < y.len(); index++) { | ||
| HITCBC | 350 | 8096 | limb yi = y[index]; | 350 | 8096 | limb yi = y[index]; | ||
| HITCBC | 351 | 8096 | stackvec<size> zi; | 351 | 8096 | stackvec<size> zi; | ||
| HITCBC | 352 | 8096 | if (yi != 0) { | 352 | 8096 | if (yi != 0) { | ||
| 353 | // re-use the same buffer throughout | 353 | // re-use the same buffer throughout | |||||
| HITCBC | 354 | 8096 | zi.set_len(0); | 354 | 8096 | zi.set_len(0); | ||
| HITCBC | 355 | 8096 | BOOST_JSON_FASTFLOAT_TRY(zi.try_extend(zs)); | 355 | 8096 | BOOST_JSON_FASTFLOAT_TRY(zi.try_extend(zs)); | ||
| HITCBC | 356 | 8096 | BOOST_JSON_FASTFLOAT_TRY(small_mul(zi, yi)); | 356 | 8096 | BOOST_JSON_FASTFLOAT_TRY(small_mul(zi, yi)); | ||
| HITCBC | 357 | 8096 | limb_span zis = limb_span(zi.data, zi.len()); | 357 | 8096 | limb_span zis = limb_span(zi.data, zi.len()); | ||
| HITCBC | 358 | 8096 | BOOST_JSON_FASTFLOAT_TRY(large_add_from(x, zis, index)); | 358 | 8096 | BOOST_JSON_FASTFLOAT_TRY(large_add_from(x, zis, index)); | ||
| 359 | } | 359 | } | |||||
| 360 | } | 360 | } | |||||
| 361 | } | 361 | } | |||||
| 362 | 362 | |||||||
| HITCBC | 363 | 2024 | x.normalize(); | 363 | 2024 | x.normalize(); | ||
| HITCBC | 364 | 2024 | return true; | 364 | 2024 | return true; | ||
| 365 | } | 365 | } | |||||
| 366 | 366 | |||||||
| 367 | // grade-school multiplication algorithm | 367 | // grade-school multiplication algorithm | |||||
| 368 | template <uint16_t size> | 368 | template <uint16_t size> | |||||
| 369 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 | 369 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 | |||||
| HITCBC | 370 | 2024 | bool large_mul(stackvec<size>& x, limb_span y) noexcept { | 370 | 2024 | bool large_mul(stackvec<size>& x, limb_span y) noexcept { | ||
| HITCBC | 371 | 2024 | if (y.len() == 1) { | 371 | 2024 | if (y.len() == 1) { | ||
| MISUBC | 372 | ✗ | BOOST_JSON_FASTFLOAT_TRY(small_mul(x, y[0])); | 372 | ✗ | BOOST_JSON_FASTFLOAT_TRY(small_mul(x, y[0])); | ||
| 373 | } else { | 373 | } else { | |||||
| HITCBC | 374 | 2024 | BOOST_JSON_FASTFLOAT_TRY(long_mul(x, y)); | 374 | 2024 | BOOST_JSON_FASTFLOAT_TRY(long_mul(x, y)); | ||
| 375 | } | 375 | } | |||||
| HITCBC | 376 | 2024 | return true; | 376 | 2024 | return true; | ||
| 377 | } | 377 | } | |||||
| 378 | 378 | |||||||
| 379 | template <typename = void> | 379 | template <typename = void> | |||||
| 380 | struct pow5_tables { | 380 | struct pow5_tables { | |||||
| 381 | static constexpr uint32_t large_step = 135; | 381 | static constexpr uint32_t large_step = 135; | |||||
| 382 | static constexpr uint64_t small_power_of_5[] = { | 382 | static constexpr uint64_t small_power_of_5[] = { | |||||
| 383 | 1UL, 5UL, 25UL, 125UL, 625UL, 3125UL, 15625UL, 78125UL, 390625UL, | 383 | 1UL, 5UL, 25UL, 125UL, 625UL, 3125UL, 15625UL, 78125UL, 390625UL, | |||||
| 384 | 1953125UL, 9765625UL, 48828125UL, 244140625UL, 1220703125UL, | 384 | 1953125UL, 9765625UL, 48828125UL, 244140625UL, 1220703125UL, | |||||
| 385 | 6103515625UL, 30517578125UL, 152587890625UL, 762939453125UL, | 385 | 6103515625UL, 30517578125UL, 152587890625UL, 762939453125UL, | |||||
| 386 | 3814697265625UL, 19073486328125UL, 95367431640625UL, 476837158203125UL, | 386 | 3814697265625UL, 19073486328125UL, 95367431640625UL, 476837158203125UL, | |||||
| 387 | 2384185791015625UL, 11920928955078125UL, 59604644775390625UL, | 387 | 2384185791015625UL, 11920928955078125UL, 59604644775390625UL, | |||||
| 388 | 298023223876953125UL, 1490116119384765625UL, 7450580596923828125UL, | 388 | 298023223876953125UL, 1490116119384765625UL, 7450580596923828125UL, | |||||
| 389 | }; | 389 | }; | |||||
| 390 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | 390 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | |||||
| 391 | constexpr static limb large_power_of_5[] = { | 391 | constexpr static limb large_power_of_5[] = { | |||||
| 392 | 1414648277510068013UL, 9180637584431281687UL, 4539964771860779200UL, | 392 | 1414648277510068013UL, 9180637584431281687UL, 4539964771860779200UL, | |||||
| 393 | 10482974169319127550UL, 198276706040285095UL}; | 393 | 10482974169319127550UL, 198276706040285095UL}; | |||||
| 394 | #else | 394 | #else | |||||
| 395 | constexpr static limb large_power_of_5[] = { | 395 | constexpr static limb large_power_of_5[] = { | |||||
| 396 | 4279965485U, 329373468U, 4020270615U, 2137533757U, 4287402176U, | 396 | 4279965485U, 329373468U, 4020270615U, 2137533757U, 4287402176U, | |||||
| 397 | 1057042919U, 1071430142U, 2440757623U, 381945767U, 46164893U}; | 397 | 1057042919U, 1071430142U, 2440757623U, 381945767U, 46164893U}; | |||||
| 398 | #endif | 398 | #endif | |||||
| 399 | }; | 399 | }; | |||||
| 400 | 400 | |||||||
| 401 | template <typename T> | 401 | template <typename T> | |||||
| 402 | constexpr uint32_t pow5_tables<T>::large_step; | 402 | constexpr uint32_t pow5_tables<T>::large_step; | |||||
| 403 | 403 | |||||||
| 404 | template <typename T> | 404 | template <typename T> | |||||
| 405 | constexpr uint64_t pow5_tables<T>::small_power_of_5[]; | 405 | constexpr uint64_t pow5_tables<T>::small_power_of_5[]; | |||||
| 406 | 406 | |||||||
| 407 | template <typename T> | 407 | template <typename T> | |||||
| 408 | constexpr limb pow5_tables<T>::large_power_of_5[]; | 408 | constexpr limb pow5_tables<T>::large_power_of_5[]; | |||||
| 409 | 409 | |||||||
| 410 | // big integer type. implements a small subset of big integer | 410 | // big integer type. implements a small subset of big integer | |||||
| 411 | // arithmetic, using simple algorithms since asymptotically | 411 | // arithmetic, using simple algorithms since asymptotically | |||||
| 412 | // faster algorithms are slower for a small number of limbs. | 412 | // faster algorithms are slower for a small number of limbs. | |||||
| 413 | // all operations assume the big-integer is normalized. | 413 | // all operations assume the big-integer is normalized. | |||||
| 414 | struct bigint : pow5_tables<> { | 414 | struct bigint : pow5_tables<> { | |||||
| 415 | // storage of the limbs, in little-endian order. | 415 | // storage of the limbs, in little-endian order. | |||||
| 416 | stackvec<bigint_limbs> vec; | 416 | stackvec<bigint_limbs> vec; | |||||
| 417 | 417 | |||||||
| HITCBC | 418 | 2986 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bigint(): vec() {} | 418 | 2986 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bigint(): vec() {} | ||
| 419 | bigint(const bigint &) = delete; | 419 | bigint(const bigint &) = delete; | |||||
| 420 | bigint &operator=(const bigint &) = delete; | 420 | bigint &operator=(const bigint &) = delete; | |||||
| 421 | bigint(bigint &&) = delete; | 421 | bigint(bigint &&) = delete; | |||||
| 422 | bigint &operator=(bigint &&other) = delete; | 422 | bigint &operator=(bigint &&other) = delete; | |||||
| 423 | 423 | |||||||
| HITCBC | 424 | 1611 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bigint(uint64_t value): vec() { | 424 | 1611 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bigint(uint64_t value): vec() { | ||
| 425 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | 425 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | |||||
| HITCBC | 426 | 1611 | vec.push_unchecked(value); | 426 | 1611 | vec.push_unchecked(value); | ||
| 427 | #else | 427 | #else | |||||
| 428 | vec.push_unchecked(uint32_t(value)); | 428 | vec.push_unchecked(uint32_t(value)); | |||||
| 429 | vec.push_unchecked(uint32_t(value >> 32)); | 429 | vec.push_unchecked(uint32_t(value >> 32)); | |||||
| 430 | #endif | 430 | #endif | |||||
| HITCBC | 431 | 1611 | vec.normalize(); | 431 | 1611 | vec.normalize(); | ||
| HITCBC | 432 | 1611 | } | 432 | 1611 | } | ||
| 433 | 433 | |||||||
| 434 | // get the high 64 bits from the vector, and if bits were truncated. | 434 | // get the high 64 bits from the vector, and if bits were truncated. | |||||
| 435 | // this is to get the significant digits for the float. | 435 | // this is to get the significant digits for the float. | |||||
| HITCBC | 436 | 1375 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 uint64_t hi64(bool& truncated) const noexcept { | 436 | 1375 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 uint64_t hi64(bool& truncated) const noexcept { | ||
| 437 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | 437 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | |||||
| HITCBC | 438 | 1375 | if (vec.len() == 0) { | 438 | 1375 | if (vec.len() == 0) { | ||
| MISUBC | 439 | ✗ | return empty_hi64(truncated); | 439 | ✗ | return empty_hi64(truncated); | ||
| HITCBC | 440 | 1375 | } else if (vec.len() == 1) { | 440 | 1375 | } else if (vec.len() == 1) { | ||
| MISUBC | 441 | ✗ | return uint64_hi64(vec.rindex(0), truncated); | 441 | ✗ | return uint64_hi64(vec.rindex(0), truncated); | ||
| 442 | } else { | 442 | } else { | |||||
| HITCBC | 443 | 1375 | uint64_t result = uint64_hi64(vec.rindex(0), vec.rindex(1), truncated); | 443 | 1375 | uint64_t result = uint64_hi64(vec.rindex(0), vec.rindex(1), truncated); | ||
| HITCBC | 444 | 1375 | truncated |= vec.nonzero(2); | 444 | 1375 | truncated |= vec.nonzero(2); | ||
| HITCBC | 445 | 1375 | return result; | 445 | 1375 | return result; | ||
| 446 | } | 446 | } | |||||
| 447 | #else | 447 | #else | |||||
| 448 | if (vec.len() == 0) { | 448 | if (vec.len() == 0) { | |||||
| 449 | return empty_hi64(truncated); | 449 | return empty_hi64(truncated); | |||||
| 450 | } else if (vec.len() == 1) { | 450 | } else if (vec.len() == 1) { | |||||
| 451 | return uint32_hi64(vec.rindex(0), truncated); | 451 | return uint32_hi64(vec.rindex(0), truncated); | |||||
| 452 | } else if (vec.len() == 2) { | 452 | } else if (vec.len() == 2) { | |||||
| 453 | return uint32_hi64(vec.rindex(0), vec.rindex(1), truncated); | 453 | return uint32_hi64(vec.rindex(0), vec.rindex(1), truncated); | |||||
| 454 | } else { | 454 | } else { | |||||
| 455 | uint64_t result = uint32_hi64(vec.rindex(0), vec.rindex(1), vec.rindex(2), truncated); | 455 | uint64_t result = uint32_hi64(vec.rindex(0), vec.rindex(1), vec.rindex(2), truncated); | |||||
| 456 | truncated |= vec.nonzero(3); | 456 | truncated |= vec.nonzero(3); | |||||
| 457 | return result; | 457 | return result; | |||||
| 458 | } | 458 | } | |||||
| 459 | #endif | 459 | #endif | |||||
| 460 | } | 460 | } | |||||
| 461 | 461 | |||||||
| 462 | // compare two big integers, returning the large value. | 462 | // compare two big integers, returning the large value. | |||||
| 463 | // assumes both are normalized. if the return value is | 463 | // assumes both are normalized. if the return value is | |||||
| 464 | // negative, other is larger, if the return value is | 464 | // negative, other is larger, if the return value is | |||||
| 465 | // positive, this is larger, otherwise they are equal. | 465 | // positive, this is larger, otherwise they are equal. | |||||
| 466 | // the limbs are stored in little-endian order, so we | 466 | // the limbs are stored in little-endian order, so we | |||||
| 467 | // must compare the limbs in ever order. | 467 | // must compare the limbs in ever order. | |||||
| HITCBC | 468 | 1611 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 int compare(const bigint& other) const noexcept { | 468 | 1611 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 int compare(const bigint& other) const noexcept { | ||
| HITCBC | 469 | 1611 | if (vec.len() > other.vec.len()) { | 469 | 1611 | if (vec.len() > other.vec.len()) { | ||
| MISUBC | 470 | ✗ | return 1; | 470 | ✗ | return 1; | ||
| HITCBC | 471 | 1611 | } else if (vec.len() < other.vec.len()) { | 471 | 1611 | } else if (vec.len() < other.vec.len()) { | ||
| MISUBC | 472 | ✗ | return -1; | 472 | ✗ | return -1; | ||
| 473 | } else { | 473 | } else { | |||||
| HITCBC | 474 | 3099 | for (size_t index = vec.len(); index > 0; index--) { | 474 | 3099 | for (size_t index = vec.len(); index > 0; index--) { | ||
| HITCBC | 475 | 3099 | limb xi = vec[index - 1]; | 475 | 3099 | limb xi = vec[index - 1]; | ||
| HITCBC | 476 | 3099 | limb yi = other.vec[index - 1]; | 476 | 3099 | limb yi = other.vec[index - 1]; | ||
| HITCBC | 477 | 3099 | if (xi > yi) { | 477 | 3099 | if (xi > yi) { | ||
| HITCBC | 478 | 767 | return 1; | 478 | 767 | return 1; | ||
| HITCBC | 479 | 2332 | } else if (xi < yi) { | 479 | 2332 | } else if (xi < yi) { | ||
| HITCBC | 480 | 844 | return -1; | 480 | 844 | return -1; | ||
| 481 | } | 481 | } | |||||
| 482 | } | 482 | } | |||||
| MISUBC | 483 | ✗ | return 0; | 483 | ✗ | return 0; | ||
| 484 | } | 484 | } | |||||
| 485 | } | 485 | } | |||||
| 486 | 486 | |||||||
| 487 | // shift left each limb n bits, carrying over to the new limb | 487 | // shift left each limb n bits, carrying over to the new limb | |||||
| 488 | // returns true if we were able to shift all the digits. | 488 | // returns true if we were able to shift all the digits. | |||||
| HITCBC | 489 | 2931 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool shl_bits(size_t n) noexcept { | 489 | 2931 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool shl_bits(size_t n) noexcept { | ||
| 490 | // Internally, for each item, we shift left by n, and add the previous | 490 | // Internally, for each item, we shift left by n, and add the previous | |||||
| 491 | // right shifted limb-bits. | 491 | // right shifted limb-bits. | |||||
| 492 | // For example, we transform (for u8) shifted left 2, to: | 492 | // For example, we transform (for u8) shifted left 2, to: | |||||
| 493 | // b10100100 b01000010 | 493 | // b10100100 b01000010 | |||||
| 494 | // b10 b10010001 b00001000 | 494 | // b10 b10010001 b00001000 | |||||
| HITCBC | 495 | 2931 | BOOST_ASSERT(n != 0); | 495 | 2931 | BOOST_ASSERT(n != 0); | ||
| HITCBC | 496 | 2931 | BOOST_ASSERT(n < sizeof(limb) * 8); | 496 | 2931 | BOOST_ASSERT(n < sizeof(limb) * 8); | ||
| 497 | 497 | |||||||
| HITCBC | 498 | 2931 | size_t shl = n; | 498 | 2931 | size_t shl = n; | ||
| HITCBC | 499 | 2931 | size_t shr = limb_bits - shl; | 499 | 2931 | size_t shr = limb_bits - shl; | ||
| HITCBC | 500 | 2931 | limb prev = 0; | 500 | 2931 | limb prev = 0; | ||
| HITCBC | 501 | 17014 | for (size_t index = 0; index < vec.len(); index++) { | 501 | 17014 | for (size_t index = 0; index < vec.len(); index++) { | ||
| HITCBC | 502 | 14083 | limb xi = vec[index]; | 502 | 14083 | limb xi = vec[index]; | ||
| HITCBC | 503 | 14083 | vec[index] = (xi << shl) | (prev >> shr); | 503 | 14083 | vec[index] = (xi << shl) | (prev >> shr); | ||
| HITCBC | 504 | 14083 | prev = xi; | 504 | 14083 | prev = xi; | ||
| 505 | } | 505 | } | |||||
| 506 | 506 | |||||||
| HITCBC | 507 | 2931 | limb carry = prev >> shr; | 507 | 2931 | limb carry = prev >> shr; | ||
| HITCBC | 508 | 2931 | if (carry != 0) { | 508 | 2931 | if (carry != 0) { | ||
| HITCBC | 509 | 1534 | return vec.try_push(carry); | 509 | 1534 | return vec.try_push(carry); | ||
| 510 | } | 510 | } | |||||
| HITCBC | 511 | 1397 | return true; | 511 | 1397 | return true; | ||
| 512 | } | 512 | } | |||||
| 513 | 513 | |||||||
| 514 | // move the limbs left by `n` limbs. | 514 | // move the limbs left by `n` limbs. | |||||
| HITCBC | 515 | 2414 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool shl_limbs(size_t n) noexcept { | 515 | 2414 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool shl_limbs(size_t n) noexcept { | ||
| HITCBC | 516 | 2414 | BOOST_ASSERT(n != 0); | 516 | 2414 | BOOST_ASSERT(n != 0); | ||
| HITCBC | 517 | 2414 | if (n + vec.len() > vec.capacity()) { | 517 | 2414 | if (n + vec.len() > vec.capacity()) { | ||
| MISUBC | 518 | ✗ | return false; | 518 | ✗ | return false; | ||
| HITCBC | 519 | 2414 | } else if (!vec.is_empty()) { | 519 | 2414 | } else if (!vec.is_empty()) { | ||
| 520 | // move limbs | 520 | // move limbs | |||||
| HITCBC | 521 | 2414 | limb* dst = vec.data + n; | 521 | 2414 | limb* dst = vec.data + n; | ||
| HITCBC | 522 | 2414 | const limb* src = vec.data; | 522 | 2414 | const limb* src = vec.data; | ||
| HITCBC | 523 | 2414 | std::copy_backward(src, src + vec.len(), dst + vec.len()); | 523 | 2414 | std::copy_backward(src, src + vec.len(), dst + vec.len()); | ||
| 524 | // fill in empty limbs | 524 | // fill in empty limbs | |||||
| HITCBC | 525 | 2414 | limb* first = vec.data; | 525 | 2414 | limb* first = vec.data; | ||
| HITCBC | 526 | 2414 | limb* last = first + n; | 526 | 2414 | limb* last = first + n; | ||
| HITCBC | 527 | 2414 | ::std::fill(first, last, 0); | 527 | 2414 | ::std::fill(first, last, 0); | ||
| HITCBC | 528 | 2414 | vec.set_len(n + vec.len()); | 528 | 2414 | vec.set_len(n + vec.len()); | ||
| HITCBC | 529 | 2414 | return true; | 529 | 2414 | return true; | ||
| 530 | } else { | 530 | } else { | |||||
| MISUBC | 531 | ✗ | return true; | 531 | ✗ | return true; | ||
| 532 | } | 532 | } | |||||
| 533 | } | 533 | } | |||||
| 534 | 534 | |||||||
| 535 | // move the limbs left by `n` bits. | 535 | // move the limbs left by `n` bits. | |||||
| HITCBC | 536 | 2984 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool shl(size_t n) noexcept { | 536 | 2984 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool shl(size_t n) noexcept { | ||
| HITCBC | 537 | 2984 | size_t rem = n % limb_bits; | 537 | 2984 | size_t rem = n % limb_bits; | ||
| HITCBC | 538 | 2984 | size_t div = n / limb_bits; | 538 | 2984 | size_t div = n / limb_bits; | ||
| HITCBC | 539 | 2984 | if (rem != 0) { | 539 | 2984 | if (rem != 0) { | ||
| HITCBC | 540 | 2931 | BOOST_JSON_FASTFLOAT_TRY(shl_bits(rem)); | 540 | 2931 | BOOST_JSON_FASTFLOAT_TRY(shl_bits(rem)); | ||
| 541 | } | 541 | } | |||||
| HITCBC | 542 | 2984 | if (div != 0) { | 542 | 2984 | if (div != 0) { | ||
| HITCBC | 543 | 2414 | BOOST_JSON_FASTFLOAT_TRY(shl_limbs(div)); | 543 | 2414 | BOOST_JSON_FASTFLOAT_TRY(shl_limbs(div)); | ||
| 544 | } | 544 | } | |||||
| HITCBC | 545 | 2984 | return true; | 545 | 2984 | return true; | ||
| 546 | } | 546 | } | |||||
| 547 | 547 | |||||||
| 548 | // get the number of leading zeros in the bigint. | 548 | // get the number of leading zeros in the bigint. | |||||
| HITCBC | 549 | 1375 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 int ctlz() const noexcept { | 549 | 1375 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 int ctlz() const noexcept { | ||
| HITCBC | 550 | 1375 | if (vec.is_empty()) { | 550 | 1375 | if (vec.is_empty()) { | ||
| MISUBC | 551 | ✗ | return 0; | 551 | ✗ | return 0; | ||
| 552 | } else { | 552 | } else { | |||||
| 553 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | 553 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | |||||
| HITCBC | 554 | 2750 | return leading_zeroes(vec.rindex(0)); | 554 | 2750 | return leading_zeroes(vec.rindex(0)); | ||
| 555 | #else | 555 | #else | |||||
| 556 | // no use defining a specialized leading_zeroes for a 32-bit type. | 556 | // no use defining a specialized leading_zeroes for a 32-bit type. | |||||
| 557 | uint64_t r0 = vec.rindex(0); | 557 | uint64_t r0 = vec.rindex(0); | |||||
| 558 | return leading_zeroes(r0 << 32); | 558 | return leading_zeroes(r0 << 32); | |||||
| 559 | #endif | 559 | #endif | |||||
| 560 | } | 560 | } | |||||
| 561 | } | 561 | } | |||||
| 562 | 562 | |||||||
| 563 | // get the number of bits in the bigint. | 563 | // get the number of bits in the bigint. | |||||
| HITCBC | 564 | 1375 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 int bit_length() const noexcept { | 564 | 1375 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 int bit_length() const noexcept { | ||
| HITCBC | 565 | 1375 | int lz = ctlz(); | 565 | 1375 | int lz = ctlz(); | ||
| HITCBC | 566 | 1375 | return int(limb_bits * vec.len()) - lz; | 566 | 1375 | return int(limb_bits * vec.len()) - lz; | ||
| 567 | } | 567 | } | |||||
| 568 | 568 | |||||||
| HITCBC | 569 | 9437 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool mul(limb y) noexcept { | 569 | 9437 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool mul(limb y) noexcept { | ||
| HITCBC | 570 | 9437 | return small_mul(vec, y); | 570 | 9437 | return small_mul(vec, y); | ||
| 571 | } | 571 | } | |||||
| 572 | 572 | |||||||
| HITCBC | 573 | 9437 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool add(limb y) noexcept { | 573 | 9437 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool add(limb y) noexcept { | ||
| HITCBC | 574 | 18874 | return small_add(vec, y); | 574 | 18874 | return small_add(vec, y); | ||
| 575 | } | 575 | } | |||||
| 576 | 576 | |||||||
| 577 | // multiply as if by 2 raised to a power. | 577 | // multiply as if by 2 raised to a power. | |||||
| HITCBC | 578 | 2984 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool pow2(uint32_t exp) noexcept { | 578 | 2984 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool pow2(uint32_t exp) noexcept { | ||
| HITCBC | 579 | 2984 | return shl(exp); | 579 | 2984 | return shl(exp); | ||
| 580 | } | 580 | } | |||||
| 581 | 581 | |||||||
| 582 | // multiply as if by 5 raised to a power. | 582 | // multiply as if by 5 raised to a power. | |||||
| HITCBC | 583 | 2986 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool pow5(uint32_t exp) noexcept { | 583 | 2986 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool pow5(uint32_t exp) noexcept { | ||
| 584 | // multiply by a power of 5 | 584 | // multiply by a power of 5 | |||||
| HITCBC | 585 | 2986 | constexpr size_t large_length = sizeof(large_power_of_5) / sizeof(limb); | 585 | 2986 | constexpr size_t large_length = sizeof(large_power_of_5) / sizeof(limb); | ||
| HITCBC | 586 | 2986 | limb_span large = limb_span(large_power_of_5, large_length); | 586 | 2986 | limb_span large = limb_span(large_power_of_5, large_length); | ||
| HITCBC | 587 | 5010 | while (exp >= large_step) { | 587 | 5010 | while (exp >= large_step) { | ||
| HITCBC | 588 | 2024 | BOOST_JSON_FASTFLOAT_TRY(large_mul(vec, large)); | 588 | 2024 | BOOST_JSON_FASTFLOAT_TRY(large_mul(vec, large)); | ||
| HITCBC | 589 | 2024 | exp -= large_step; | 589 | 2024 | exp -= large_step; | ||
| 590 | } | 590 | } | |||||
| 591 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | 591 | #ifdef BOOST_JSON_FASTFLOAT_64BIT_LIMB | |||||
| HITCBC | 592 | 2986 | constexpr uint32_t small_step = 27; | 592 | 2986 | constexpr uint32_t small_step = 27; | ||
| HITCBC | 593 | 2986 | constexpr limb max_native = 7450580596923828125UL; | 593 | 2986 | constexpr limb max_native = 7450580596923828125UL; | ||
| 594 | #else | 594 | #else | |||||
| 595 | constexpr uint32_t small_step = 13; | 595 | constexpr uint32_t small_step = 13; | |||||
| 596 | constexpr limb max_native = 1220703125U; | 596 | constexpr limb max_native = 1220703125U; | |||||
| 597 | #endif | 597 | #endif | |||||
| HITCBC | 598 | 8590 | while (exp >= small_step) { | 598 | 8590 | while (exp >= small_step) { | ||
| HITCBC | 599 | 5604 | BOOST_JSON_FASTFLOAT_TRY(small_mul(vec, max_native)); | 599 | 5604 | BOOST_JSON_FASTFLOAT_TRY(small_mul(vec, max_native)); | ||
| HITCBC | 600 | 5604 | exp -= small_step; | 600 | 5604 | exp -= small_step; | ||
| 601 | } | 601 | } | |||||
| HITCBC | 602 | 2986 | if (exp != 0) { | 602 | 2986 | if (exp != 0) { | ||
| 603 | // Work around clang bug https://godbolt.org/z/zedh7rrhc | 603 | // Work around clang bug https://godbolt.org/z/zedh7rrhc | |||||
| 604 | // This is similar to https://github.com/llvm/llvm-project/issues/47746, | 604 | // This is similar to https://github.com/llvm/llvm-project/issues/47746, | |||||
| 605 | // except the workaround described there don't work here | 605 | // except the workaround described there don't work here | |||||
| HITCBC | 606 | 2868 | BOOST_JSON_FASTFLOAT_TRY( | 606 | 2868 | BOOST_JSON_FASTFLOAT_TRY( | ||
| 607 | small_mul(vec, limb(((void)small_power_of_5[0], small_power_of_5[exp]))) | 607 | small_mul(vec, limb(((void)small_power_of_5[0], small_power_of_5[exp]))) | |||||
| 608 | ); | 608 | ); | |||||
| 609 | } | 609 | } | |||||
| 610 | 610 | |||||||
| HITCBC | 611 | 2986 | return true; | 611 | 2986 | return true; | ||
| 612 | } | 612 | } | |||||
| 613 | 613 | |||||||
| 614 | // multiply as if by 10 raised to a power. | 614 | // multiply as if by 10 raised to a power. | |||||
| HITCBC | 615 | 1375 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool pow10(uint32_t exp) noexcept { | 615 | 1375 | BOOST_JSON_FASTFLOAT_CONSTEXPR20 bool pow10(uint32_t exp) noexcept { | ||
| HITCBC | 616 | 1375 | BOOST_JSON_FASTFLOAT_TRY(pow5(exp)); | 616 | 1375 | BOOST_JSON_FASTFLOAT_TRY(pow5(exp)); | ||
| HITCBC | 617 | 1375 | return pow2(exp); | 617 | 1375 | return pow2(exp); | ||
| 618 | } | 618 | } | |||||
| 619 | }; | 619 | }; | |||||
| 620 | 620 | |||||||
| 621 | }}}}}} // namespace fast_float | 621 | }}}}}} // namespace fast_float | |||||
| 622 | 622 | |||||||
| 623 | #endif | 623 | #endif | |||||