| 1 | #include "duckdb/function/scalar/string_functions.hpp" |
| 2 | |
| 3 | #include "duckdb/common/exception.hpp" |
| 4 | #include "duckdb/common/vector_operations/vector_operations.hpp" |
| 5 | #include "duckdb/planner/expression/bound_function_expression.hpp" |
| 6 | |
| 7 | namespace duckdb { |
| 8 | |
| 9 | template <class UNSIGNED, int NEEDLE_SIZE> |
| 10 | static idx_t ContainsUnaligned(const unsigned char *haystack, idx_t haystack_size, const unsigned char *needle, |
| 11 | idx_t base_offset) { |
| 12 | if (NEEDLE_SIZE > haystack_size) { |
| 13 | // needle is bigger than haystack: haystack cannot contain needle |
| 14 | return DConstants::INVALID_INDEX; |
| 15 | } |
| 16 | // contains for a small unaligned needle (3/5/6/7 bytes) |
| 17 | // we perform unsigned integer comparisons to check for equality of the entire needle in a single comparison |
| 18 | // this implementation is inspired by the memmem implementation of freebsd |
| 19 | |
| 20 | // first we set up the needle and the first NEEDLE_SIZE characters of the haystack as UNSIGNED integers |
| 21 | UNSIGNED needle_entry = 0; |
| 22 | UNSIGNED haystack_entry = 0; |
| 23 | const UNSIGNED start = (sizeof(UNSIGNED) * 8) - 8; |
| 24 | const UNSIGNED shift = (sizeof(UNSIGNED) - NEEDLE_SIZE) * 8; |
| 25 | for (int i = 0; i < NEEDLE_SIZE; i++) { |
| 26 | needle_entry |= UNSIGNED(needle[i]) << UNSIGNED(start - i * 8); |
| 27 | haystack_entry |= UNSIGNED(haystack[i]) << UNSIGNED(start - i * 8); |
| 28 | } |
| 29 | // now we perform the actual search |
| 30 | for (idx_t offset = NEEDLE_SIZE; offset < haystack_size; offset++) { |
| 31 | // for this position we first compare the haystack with the needle |
| 32 | if (haystack_entry == needle_entry) { |
| 33 | return base_offset + offset - NEEDLE_SIZE; |
| 34 | } |
| 35 | // now we adjust the haystack entry by |
| 36 | // (1) removing the left-most character (shift by 8) |
| 37 | // (2) adding the next character (bitwise or, with potential shift) |
| 38 | // this shift is only necessary if the needle size is not aligned with the unsigned integer size |
| 39 | // (e.g. needle size 3, unsigned integer size 4, we need to shift by 1) |
| 40 | haystack_entry = (haystack_entry << 8) | ((UNSIGNED(haystack[offset])) << shift); |
| 41 | } |
| 42 | if (haystack_entry == needle_entry) { |
| 43 | return base_offset + haystack_size - NEEDLE_SIZE; |
| 44 | } |
| 45 | return DConstants::INVALID_INDEX; |
| 46 | } |
| 47 | |
| 48 | template <class UNSIGNED> |
| 49 | static idx_t ContainsAligned(const unsigned char *haystack, idx_t haystack_size, const unsigned char *needle, |
| 50 | idx_t base_offset) { |
| 51 | if (sizeof(UNSIGNED) > haystack_size) { |
| 52 | // needle is bigger than haystack: haystack cannot contain needle |
| 53 | return DConstants::INVALID_INDEX; |
| 54 | } |
| 55 | // contains for a small needle aligned with unsigned integer (2/4/8) |
| 56 | // similar to ContainsUnaligned, but simpler because we only need to do a reinterpret cast |
| 57 | auto needle_entry = Load<UNSIGNED>(needle); |
| 58 | for (idx_t offset = 0; offset <= haystack_size - sizeof(UNSIGNED); offset++) { |
| 59 | // for this position we first compare the haystack with the needle |
| 60 | auto haystack_entry = Load<UNSIGNED>(haystack + offset); |
| 61 | if (needle_entry == haystack_entry) { |
| 62 | return base_offset + offset; |
| 63 | } |
| 64 | } |
| 65 | return DConstants::INVALID_INDEX; |
| 66 | } |
| 67 | |
| 68 | idx_t ContainsGeneric(const unsigned char *haystack, idx_t haystack_size, const unsigned char *needle, |
| 69 | idx_t needle_size, idx_t base_offset) { |
| 70 | if (needle_size > haystack_size) { |
| 71 | // needle is bigger than haystack: haystack cannot contain needle |
| 72 | return DConstants::INVALID_INDEX; |
| 73 | } |
| 74 | // this implementation is inspired by Raphael Javaux's faststrstr (https://github.com/RaphaelJ/fast_strstr) |
| 75 | // generic contains; note that we can't use strstr because we don't have null-terminated strings anymore |
| 76 | // we keep track of a shifting window sum of all characters with window size equal to needle_size |
| 77 | // this shifting sum is used to avoid calling into memcmp; |
| 78 | // we only need to call into memcmp when the window sum is equal to the needle sum |
| 79 | // when that happens, the characters are potentially the same and we call into memcmp to check if they are |
| 80 | uint32_t sums_diff = 0; |
| 81 | for (idx_t i = 0; i < needle_size; i++) { |
| 82 | sums_diff += haystack[i]; |
| 83 | sums_diff -= needle[i]; |
| 84 | } |
| 85 | idx_t offset = 0; |
| 86 | while (true) { |
| 87 | if (sums_diff == 0 && haystack[offset] == needle[0]) { |
| 88 | if (memcmp(s1: haystack + offset, s2: needle, n: needle_size) == 0) { |
| 89 | return base_offset + offset; |
| 90 | } |
| 91 | } |
| 92 | if (offset >= haystack_size - needle_size) { |
| 93 | return DConstants::INVALID_INDEX; |
| 94 | } |
| 95 | sums_diff -= haystack[offset]; |
| 96 | sums_diff += haystack[offset + needle_size]; |
| 97 | offset++; |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | idx_t ContainsFun::Find(const unsigned char *haystack, idx_t haystack_size, const unsigned char *needle, |
| 102 | idx_t needle_size) { |
| 103 | D_ASSERT(needle_size > 0); |
| 104 | // start off by performing a memchr to find the first character of the |
| 105 | auto location = memchr(s: haystack, c: needle[0], n: haystack_size); |
| 106 | if (location == nullptr) { |
| 107 | return DConstants::INVALID_INDEX; |
| 108 | } |
| 109 | idx_t base_offset = const_uchar_ptr_cast(src: location) - haystack; |
| 110 | haystack_size -= base_offset; |
| 111 | haystack = const_uchar_ptr_cast(src: location); |
| 112 | // switch algorithm depending on needle size |
| 113 | switch (needle_size) { |
| 114 | case 1: |
| 115 | return base_offset; |
| 116 | case 2: |
| 117 | return ContainsAligned<uint16_t>(haystack, haystack_size, needle, base_offset); |
| 118 | case 3: |
| 119 | return ContainsUnaligned<uint32_t, 3>(haystack, haystack_size, needle, base_offset); |
| 120 | case 4: |
| 121 | return ContainsAligned<uint32_t>(haystack, haystack_size, needle, base_offset); |
| 122 | case 5: |
| 123 | return ContainsUnaligned<uint64_t, 5>(haystack, haystack_size, needle, base_offset); |
| 124 | case 6: |
| 125 | return ContainsUnaligned<uint64_t, 6>(haystack, haystack_size, needle, base_offset); |
| 126 | case 7: |
| 127 | return ContainsUnaligned<uint64_t, 7>(haystack, haystack_size, needle, base_offset); |
| 128 | case 8: |
| 129 | return ContainsAligned<uint64_t>(haystack, haystack_size, needle, base_offset); |
| 130 | default: |
| 131 | return ContainsGeneric(haystack, haystack_size, needle, needle_size, base_offset); |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | idx_t ContainsFun::Find(const string_t &haystack_s, const string_t &needle_s) { |
| 136 | auto haystack = const_uchar_ptr_cast(src: haystack_s.GetData()); |
| 137 | auto haystack_size = haystack_s.GetSize(); |
| 138 | auto needle = const_uchar_ptr_cast(src: needle_s.GetData()); |
| 139 | auto needle_size = needle_s.GetSize(); |
| 140 | if (needle_size == 0) { |
| 141 | // empty needle: always true |
| 142 | return 0; |
| 143 | } |
| 144 | return ContainsFun::Find(haystack, haystack_size, needle, needle_size); |
| 145 | } |
| 146 | |
| 147 | struct ContainsOperator { |
| 148 | template <class TA, class TB, class TR> |
| 149 | static inline TR Operation(TA left, TB right) { |
| 150 | return ContainsFun::Find(left, right) != DConstants::INVALID_INDEX; |
| 151 | } |
| 152 | }; |
| 153 | |
| 154 | ScalarFunction ContainsFun::GetFunction() { |
| 155 | return ScalarFunction("contains" , // name of the function |
| 156 | {LogicalType::VARCHAR, LogicalType::VARCHAR}, // argument list |
| 157 | LogicalType::BOOLEAN, // return type |
| 158 | ScalarFunction::BinaryFunction<string_t, string_t, bool, ContainsOperator>); |
| 159 | } |
| 160 | |
| 161 | void ContainsFun::RegisterFunction(BuiltinFunctions &set) { |
| 162 | set.AddFunction(function: GetFunction()); |
| 163 | } |
| 164 | |
| 165 | } // namespace duckdb |
| 166 | |