3 Commits

Include renames

Author	SHA1	Message	Date
Niels Lohmann	dd7686a269	json_view: faster reading, indexing and C-stream parsing (#5888) ... * Read C streams into a json_document in blocks json_document::parse(FILE*) (and read(), accept(), parse_copy()) went through the library's file input adapter and appended the text character by character with fgetc. Read it with fread in blocks instead, like the std::istream overload already does. Parsing citm_catalog.json (1.7 MB) from a FILE*: Apple M1: 80.6 M -> 8.4 M instructions, 4.8 ms -> 0.55 ms (parse of the same text from a std::string: 8.7 M) Linux x86-64 Ir (whole run), GCC 13 / Clang 18: 47.0 M -> 13.7 M / 45.8 M -> 14.2 M The parsed documents and the error cases are identical (checked over the test corpus and 30 mutants of each file). A null FILE* is handled as before (by the library's adapter). Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Index arrays without subtrees directly operator[](idx), at(idx) and back() skipped the elements of an array one after the other, whatever they were. In an array none of whose elements is an array or object (numbers, strings, booleans, null) the elements are adjacent nodes, which the array node's own sizes tell: its subtree has exactly one node more than it has elements. Such an element is now the node idx after the array node; any other array is walked as before. Reading r[i] for every i of an array of 3000 integers: Apple M1: 28.8 M instructions, 3.1 ms -> 0.1 M instructions, 4 us Linux x86-64 Ir (whole run), GCC 13 / Clang 18: 84.1 M -> 3.1 M / 75.0 M -> 3.0 M Arrays that contain arrays are unchanged. Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Skip the bounds checks of number and string tokens in NUL-terminated input For a NUL-terminated input (std::string, C string, string literal), the parser reads the terminator instead of comparing p with the end, via cursor::cur(). Number and string scanning still tested `p != e` before each look at the next byte (after the integer digits for '.', for 'e', after the exponent sign and digits, and after a string for its closing quote), although the terminator can never be one of those characters. They now use cur(), which is the plain load for NUL-terminated input and the bounds check for ptr+len input, as before. Parsing, Apple M1 instructions per parse: twitter 3.126 M -> 3.089 M (-1.2 %), citm_catalog 8.70 M -> 8.58 M (-1.3 %), jeopardy 306.1 M -> 302.5 M (-1.2 %), signed integers 281 M -> 266 M (-5.4 %), canada and floats -0.1 % / -0.9 %. Linux x86-64 Ir, GCC 13 / Clang 18: twitter -0.5 % / -0.0 %, signed integers +0.2 % / -0.7 %. Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Make the index text of items() on arrays lazy basic_json_view::items() on an array formatted every element's index with std::to_string when the item was made, whether or not the loop asks for the key. The item now keeps the position and makes the text on the first key() call. Objects are not affected. Iterating over the items of an array of 3000 numbers and reading only the values: Apple M1: 25.5 us -> 4.6 us (0.45 M -> ~0 net instructions) Linux x86-64 Ir (whole run), GCC 13 / Clang 18: 14.9 M -> 4.1 M / 15.3 M -> 4.6 M Reading key() costs what it did (24.5 us). Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Make the default of value() with a C string only if the member is missing basic_json_view::value(key, "text") and value(pointer, "text") built a std::string from the default first and passed it on, so a call allocated for the default even when the member exists, and copied it again when it does not. They now look the member up and make the result directly: from the member if there is one, else from the C string. Behavior, including the exceptions, is the same. Two calls (member found, member missing) with a 40 character default, Apple M1: 1.6 K -> 0.7 K instructions Linux x86-64 Ir (500 rounds), GCC 13 / Clang 18: 2.46 M -> 2.22 M / 2.40 M -> 2.15 M Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Materialize scalars directly and keep the stack of small subtrees local basic_json_view::materialize() (and get<BasicJsonType>()) always set up the SAX handler that parse() uses and a std::vector for the open containers, sized from the document's deepest nesting. For a scalar, which is most of what is materialized one value at a time, neither is needed: the value is made directly (the same constructors the handler ends up calling). For arrays and objects the stack of open containers is now a local array for the first 16 levels, and the vector is only used for deeper documents. Materializing five scalars of a small document, Apple M1: 3.6 K -> 0.9 K instructions Linux x86-64 Ir (500 rounds), GCC 13 / Clang 18: 2.96 M -> 2.37 M / 2.90 M -> 2.29 M Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Copy the text between escapes while scanning it A string that contains an escape is decoded by builder::slow_string(). After each escape it scanned the next run of plain bytes (8 or 16 byte-by-byte steps, then a vector loop), and then copied that run to the decode arena in a second pass. The run is now copied 16 bytes at a time while it is scanned (one vector load, a store of the same 16 bytes, and the mask of the bytes that end the run), and the write position lives in a local variable, not in the builder, since a store of a character may alias the builder's members and made the compiler reload it after every escape. A run that ends at a quote needs no scan, and non-ASCII text goes through the UTF-8 check as before; the portable code is unchanged. Parsing, Apple M1 instructions: jeopardy.json 304.0 M -> 293.8 M (-3.4 %), 3000 strings with escapes 1.72 M -> 1.54 M (-10 %), twitter -0.6 %. Linux x86-64 Ir, GCC 13 / Clang 18: escaped strings -7.6 % / -11.4 %, twitter 0.0 % / 0.0 %. Documents and error messages are identical on 900 k random strings of escapes, UTF-8 and control characters (also under ASan), and on mutants of the test corpus. Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Hash the last bytes of a key of a large object without a loop key_hash() mixed a key eight bytes at a time and then packed the remaining 0-7 bytes one by one, in a loop. The last word of a key of eight or more bytes is now read as the last eight bytes of the key (overlapping the word before), and a key of 2-7 bytes as two overlapping words of four or two bytes. A product only moves bits upwards, and the slot is taken from the low bits: the high bits of the last word are folded down before the multiplication, since keys that differ only in their last characters ("user_17", "user_18") would otherwise share slots. The hash is not stored, so its values may change. Looking up every key of an object of 1000 random keys, Apple M1 instructions per 1000 lookups: keys of 5 bytes 175 K -> 93 K, hex strings of 8 bytes 177 K -> 168 K, keys of 12 bytes 196 K -> 186 K Linux x86-64 Ir (whole run), GCC 13 / Clang 18: 5-byte keys -4.7 % / -8.8 %, 12-byte keys -3.3 % / -7.4 %. Building the tables while parsing such objects is cheaper as well (5-byte keys: 13.3 us -> 10.1 us). Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Compare short keys of a large object without a call of memcmp find_indexed() compared the key of a member the table points to with memcmp, a library call for any length. Up to 16 bytes, the bytes are now compared as two overlapping words (the first and the last 8, 4, or 2 bytes), as find_member() already does for the objects without a table; longer keys still use memcmp. Looking up every key of an object of 1000 random keys, Apple M1 instructions per 1000 lookups (with the previous commit): keys of 5 bytes 93 K -> 72 K, hex strings of 8 bytes 168 K -> 148 K, keys of 12 bytes 186 K -> 170 K, keys of 24 bytes 198 K -> 193 K (unchanged, memcmp). Linux x86-64 Ir (whole run), GCC 13 / Clang 18: 5-byte keys -7.2 % / -7.3 %, 12-byte keys -5.9 % / -6.6 %. Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Read an unedited editable document like a read-only one Views of a json_editable_document followed edits at every step: the lookup, the element access, back(), dump() and materialize() were built for nodes that can be moved or linked, though no edit changes the index until the first one (every edit creates the edit state before it touches a node). Until then these now take the read-only code: one check of the document's edit state per call, not per member or per node. After the first edit nothing changes. Views that were made earlier keep working, as the check is made when they are used. Unedited editable document, Apple M1: dump() of twitter.json 1.40 M -> 1.22 M instructions (-12.6 %); lookups unchanged. Linux x86-64 Ir, GCC 13 / Clang 18: dump -1.7 % / -1.0 %, lookups -0.1 % / 0.0 %. Walking with iterators is unchanged, and so is dump() after an edit. Belongs to the PR "json_view: read-side speedups". Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Format the numbered test keys of json_view's index with a stream The keys were made with snprintf() and a format from a list, which Flawfinder reports as an error (CWE-134, a format string that is not a literal). A stream now formats them; the 30000 keys of the test are the same as before. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Put the key number of the json_view index test in parentheses astyle formatted i * step after << as a pointer declaration (i* step), which failed the amalgamation check. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Fix the compiler and lint findings of json_view's read side and tests - find() and contains() dispatch on Editable for the member lookup; for documents that cannot be edited both branches were the same call (GCC -Wduplicated-branches) - json_view.hpp includes <cstdio> for the FILE overload (cpplint) - the arena check macro of the builder is on one line (cpplint read it as an if body without braces) - tests: no nested conditional (also MSVC C4018), no useless cast, parentheses around mixed operators, raw string literals, += instead of temporary concatenations, and NOLINT for the fixed seeds and a find() != npos, as elsewhere in the tests Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Grow the decode arena through a function and store its position back on every return slow_string() kept its write position in locals and grew the arena through the macro NLOHMANN_VIEW_ENSURE. The macro is now the always-inline member arena_ensure() (a lambda is not inlined at every call site by clang, which then keeps the position in memory: +11 % instructions on escape-heavy strings), and a guard stores the position back into m_arena_write on every return instead of by hand on the successful one. Instruction counts are unchanged (clang) or 1 % lower (GCC). Also word the complexity of back() as that of at(). Signed-off-by: Niels Lohmann <mail@nlohmann.me> --------- Signed-off-by: Niels Lohmann <mail@nlohmann.me>	2026-10-24 07:41:06 +02:00
Niels Lohmann	e9970066c5	Scan strings with SSE2, SSSE3, or NEON (#5853) ... * Scan strings with SSE2, SSSE3, or NEON The string fast path found the end of a run of plain characters 8 bytes at a time (SWAR) and validated non-ASCII text one sequence at a time, calling validate_one_utf8() out of line for each. The vector code of json_view's scanner, in the new shared header detail/input/simd_scan.hpp, now does both 16 bytes at a time: SSE2 finds the end of a run on x86-64, and the UTF-8 check of Keiser and Lemire runs with NEON on AArch64 and with SSSE3 on x86-64. SSSE3 is not part of x86-64, so the check is compiled for it with a function attribute and used when the CPU has it (cached in a constant-initialized atomic), else the scalar check; code compiled with -mssse3 skips the CPU check. The first word of a run is still checked with SWAR (the first four with NEON, where a vector mask costs more), as most keys and short values end there, and the scalar helpers are always inlined. On anything else, the scalar code is unchanged. The new macro JSON_NO_SIMD selects the portable code everywhere, including the SSE2/NEON digit formatting of the floating-point writer. The lexer tests now also run with JSON_NO_SIMD and, on x86-64, with -mssse3, and a new test places every byte that ends a run at each offset across several 16-byte blocks, after ASCII and after CJK text. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Keep the string scan from slowing down short strings and dump() on Arm Three benchmarks were 7-9% slower than develop on Apple M1 (Apple Clang, -O3 -flto): Accept/jeopardy, Dump/citm_catalog, and Dump/jeopardy. - dump(): the serializer passes a string without its closing quote, so plain_run_end() ran out of input and called the out-of-line NEON loop for every string, with fewer than 16 bytes left. The loop is now only called for 16 bytes or more. - accept(): no string of jeopardy.json reaches the vector UTF-8 check, yet compiling it into the lexer's scan function made it 5% slower: the function runs once per token and now saved more registers and kept a stack slot. lexer_string_bulk_run() checks the first four words inline and continues out of line, and the serializer keeps string_bulk_run() inline, which is faster there. - The byte that ends a run is read from the scanned buffer instead of through get(), which Apple Clang merged with the other get() calls that load the read position again right after it was stored (3%). Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Define string_bulk_run_from() only where lexer_string_bulk_run() calls it With NEON and simdutf in C++17, lexer_string_bulk_run() takes the string_bulk_run() path, so the out-of-line helper was unused. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Do not dispatch the SSSE3 string scan with GCC's C++20 modules GCC does not keep the target("ssse3") attribute on functions it reads back from a module interface, so the module build (ci_module_cpp20, gcc:latest) rejected the intrinsics in vector_utf8_run() with "needs isa option -mssse3". With __cpp_modules, GCC now uses the portable UTF-8 check like other compilers without the target attribute; the SSE2 scan stays, and -mssse3 or more still compiles the SSSE3 check directly. Clang is not affected. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Include <cpuid.h> only once when json.hpp and json_view.hpp are both used The SSSE3 dispatch of the string scan and the one of json_view each include <cpuid.h>. Before GCC 11 the header has no include guard, so a translation unit that includes both headers fails with GCC 10: "redefinition of 'unsigned int __get_cpuid_max(unsigned int, unsigned int*)'". Guard both includes with NLOHMANN_JSON_CPUID_INCLUDED. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Add an ABI tag for JSON_NO_SIMD JSON_NO_SIMD selects between two definitions of the internal scanning and formatting functions. Like JSON_VIEW_NO_SIMD (_vns), it now appends _ns to the inline namespace, so translation units built with and without it can be linked into one program without ODR violations. Signed-off-by: Niels Lohmann <mail@nlohmann.me> --------- Signed-off-by: Niels Lohmann <mail@nlohmann.me>	2026-10-21 20:00:50 +02:00
Niels Lohmann	3ff37ead75	Scan json_view strings with SIMD and index large objects (#5628) ... * Scan json_view strings with SIMD and index large objects Speed up json_view's parser with SIMD scanning and a hash table for large objects. Long runs of string bytes are scanned 16 bytes at a time with NEON (AArch64, GCC and Clang) and SSE2 (x86-64), both baseline instruction sets. Keys keep 16 table checks before the vector loop, because their lengths repeat from record to record; string values get 8, because their lengths vary more. Non-ASCII text is validated 16 bytes at a time with simdjson's "lookup4" check (Keiser and Lemire, 2021), with NEON on AArch64 and, on x86-64, with SSSE3. SSSE3 is not part of baseline x86-64, so the check is compiled for SSSE3 with a function attribute and used only where CPUID reports it, which all x86-64 CPUs since about 2011 do; the answer is cached in a statically initialized atomic, so there is no guard of a local static and no global constructor. The same input is accepted either way. JSON_VIEW_NO_SIMD selects the portable code. On x86-64, string runs are now checked vector-first: one SSE2 compare from the first byte finds the end of most keys and short values, instead of a branch per byte for the first 8-16 bytes. AArch64 keeps the byte-wise steps, where a NEON mask costs more and the branches predict well. Entering an object or array no longer stalls: open() stores the parent's frame field by field instead of building it on the stack and reading it back with wider loads, which waited for the narrower stores to retire. Objects with 128 members or more get an open-addressing hash table built when the object closes, so operator[], at(), find(), contains(), count(), value(), and JSON pointers take constant time on average in such objects; of duplicate keys, the first is kept, as for the linear search. The idea comes from Boost.JSON. simdjson is credited in simd.hpp's SPDX block, the README, and license.md. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Fix GCC module build and MSVC C4127 in the json_view SIMD code GCC ignores the target attribute in modules, so the SSSE3 dispatch is disabled for the module interface (the check stays portable, SSE2 is kept). Make the 8-vs-16 byte unrolling condition in scan_string_run a preprocessor/template split to avoid a constant condition (C4127). Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Remove an accidentally committed GCC module cache Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Bound the probe length of the large-object hash index The key hash is not seeded, so keys chosen to collide made building the table quadratic (20,000 colliding keys took 470 ms to parse). A key may now sit at most 64 slots from its home slot; if a key would sit further away, the table is dropped and the object is searched linearly. Lookups stop after the same distance. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Release the spare capacity of the hash index in shrink_to_fit shrink_to_fit() now trims the tables of large objects like the node array and the decoded strings, and the list of large objects is released as soon as the tables are built. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Use the portable string scan on non-x86 targets that define __SSE2__ WebAssembly with -msse2 defines __SSE2__, but has no <cpuid.h> or x86 intrinsics headers. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Document the hash index number in node::extra Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Keep the last of duplicate keys in the object hash index Lookups in objects with 128 members or more now return the last member of a repeated key, like the linear search of smaller objects and like materialize() and basic_json::parse(). build_object_index let the first occurrence win, so the same text gave different results depending on the size of the object. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Keep the first of duplicate keys in the object hash index again Lookups in objects with 128 members or more return the first member of a repeated key again, like the linear search of smaller objects. This undoes the code change of a69542046; its test now expects the first member from lookups (with and without a table) and the last value from materialize() and basic_json::parse(). Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Fix clang-tidy 22 findings in the json_view headers Silence bugprone-casting-through-void for the SSE loads (a reinterpret_cast would trip -Wcast-align=strict), use auto for a cast initialiser, add parentheses to a mixed expression, and name bugprone-std-namespace-modification in the NOLINTs of the tuple_size/tuple_element specialisations. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Fix clang-tidy 22 findings in unit-json_view.cpp and unit-json_view_builder.cpp Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Avoid GCC useless-cast and strict-overflow warnings in unit-json_view.cpp Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Index strings with size_t in the colliding-keys test (MSVC C4244 on Win32) Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Make json_view.hpp pass include-what-you-use ci_single_binaries runs IWYU with --error on every header. For json_view.hpp it suggested adding <array> (std::array is used), the headers that json.hpp already provides (abi_config, abi_macros, input_adapters, json_pointer, cpp_future, string_concat, value_t, json_fwd), and <version> for std::nullptr_t, and removing <cstddef>. - include <array> - keep <cstddef> (nullptr_t, size_t; IWYU attributes them to <version> and <cstring>) - tell IWYU not to suggest the headers that json.hpp provides: the amalgamated json_view.hpp only includes json.hpp, so including them here would duplicate their definitions Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Move the dump, comparison, and large-object tests to unit-json_view_dump.cpp The dump and comparison tests moved to their own file in the dump pull request; the hash index tests of this branch join them there. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Regenerate the amalgamated headers Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Follow the renames from the review of the view builder The review of #5620 renamed document_data's tape/tape_size/tape_cap to nodes/node_count/node_capacity and made the decode arena a std::vector<char>; adapt the code of this change to it. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Rename slots in the object index for Qt compatibility Qt defines slots as a macro, so a Qt header included before json_view.hpp broke the declarations of the local variable (see #5865). The test variable is renamed as well. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Address the review of the json_view object index and SIMD scan From gregmarr's review of #5628: - Docs and comments say that the hash index is one of the object's keys. - object_index.hpp: conditional_static_cast for the hash, "slot" for the table position, no flag for a degenerate table, and a static_assert that the smallest indexed object is larger than the probe window. - scan.hpp unrolls with NLOHMANN_VIEW_REPEAT_FIRST_8/SECOND_8. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Regenerate the amalgamated headers Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Encode JSON_VIEW_NO_SIMD and JSON_VIEW_USE_SSSE3 in the ABI namespace Both macros select between definitions of the same inline functions, so translation units that disagree on them violated the ODR. Like the other ABI-affecting macros, they now add a tag to the inline namespace (_vns and _vssse3), and the docs replace the "Define consistently" warning with the usual ABI compatibility note. From gregmarr's review of #5628. nlohmann_json.natvis covers twice as many tags now, so it lists four times as many namespaces (14.7 MB instead of 3.4 MB). Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Regenerate the amalgamated headers Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Do not include type_traits.hpp directly in the object index json.hpp, which object_index.hpp already includes, provides it; included directly, the json_view amalgamation (which only treats nlohmann/json.hpp as external) copied the json internals into json_view.hpp a second time, so the fuzzers, which include both single headers, failed with redefinitions (make_void, integer_sequence, priority_tag, ...). Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Compare keys with key_equals() in the object hash index As the linear lookups do since the second review round of #5622. Signed-off-by: Niels Lohmann <mail@nlohmann.me> * Regenerate the amalgamated headers Signed-off-by: Niels Lohmann <mail@nlohmann.me> --------- Signed-off-by: Niels Lohmann <mail@nlohmann.me>	2026-10-21 10:21:28 +02:00