- With SSE2, string runs are checked 16 bytes at a time from their first
byte: one compare finds the end of most keys and short values, faster on
x86-64 than a branch per byte. AArch64 keeps the byte steps (there, a
NEON mask costs more and the branches predict well; vector-first was 20%
slower on Apple M1).
- open() stores the parent's frame field by field. Built on the stack and
copied, it was read back by loads wider than its stores, which waited for
them (store forwarding fails): 18% of the time on citm_catalog.json.
Parse on x86-64 (GCC 13 / Clang 18, us): twitter 352 -> 306 / 316 -> 288,
citm_catalog 1000 -> 770 / 843 -> 692, canada 1941 -> 1687 / 1978 -> 1722.
Unchanged on Apple M1.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The vector UTF-8 check of json_view needed SSSE3 at compile time
(JSON_VIEW_USE_SSSE3 with -mssse3), so default x86-64 builds validated
non-ASCII text one sequence at a time. The check is now compiled for
SSSE3 with a function attribute (GCC 4.9 and later, Clang; MSVC compiles
the intrinsics anyway) and used where CPUID reports SSSE3. The answer is
kept in an atomic that is initialized at compile time, so neither a
guard nor a global constructor is needed. The definitions do not depend
on compiler flags, so there is no ODR issue. JSON_VIEW_USE_SSSE3 now only
skips the CPU check.
On x86-64 Linux (Haswell), twitter.json parses 23% faster with Clang 18
and 25-34% faster with GCC 13, now ahead of yyjson.
Also always inline read_eight_bytes() and parse_eight_digits(): GCC
called both in the number loops of the lexer and of json_view (52 call
sites), which cost about 10% on citm_catalog.json at -O2.
Document that reusing a document with read() avoids the page faults of
a fresh node index (about 40% of a 55 MB parse on x86-64 Linux).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Long runs of string bytes are scanned 16 at a time with NEON (AArch64, with
GCC and Clang) and SSE2 (x86-64): both belong to the baseline instruction
sets. A signed compare with 0x20 finds control characters and non-ASCII
bytes at once. Keys keep 16 table checks before the vector loop (their
lengths repeat from record to record, so the branches predict well);
string values have 8, as their lengths vary more.
Non-ASCII text is validated 16 bytes at a time with the "lookup4" check of
simdjson (J. Keiser and D. Lemire, "Validating UTF-8 In Less Than One
Instruction Per Byte", 2021): with NEON, and on x86-64 with SSSE3 if
JSON_VIEW_USE_SSSE3 is defined (SSSE3 is not part of x86-64, and the code
must not depend on the flags of a translation unit). JSON_VIEW_NO_SIMD
selects the portable code. The vector code sits in
detail/view/simd.hpp; the same input is accepted either way.
json_document::parse, best of 7 runs in separate processes (M1 Max):
poet.json (CJK text) -72%, random.json -25%, twitter.json -22%,
gsoc-2018.json -20%, semanticscholar -19%, github_events -11%,
apache_builds -9.5%, canada/citm -5/-6%; lottie +4%, tree-pretty +2.5%.
Tests: every two-byte sequence and three- and four-byte sequences with
continuation bytes at the edges of their ranges, at every offset around
the vector blocks of keys and values, cut short, and long runs of text
with a damaged byte, against json::accept and json::parse. CMake builds
the parser tests again with JSON_VIEW_NO_SIMD, and on x86-64 with
JSON_VIEW_USE_SSSE3 and -mssse3; the macros are documented.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Since whole blocks of eight digits are read directly, parse_upto8() only
gets fewer than eight digits; its eight-digit case was dead code.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
A block of eight digits of a number token lies inside the input (the
digits were counted while scanning, or are recorded in the digit
layout), so parse_upto19() reads it without the bounds check of the
last, partial block. Traversing canada.json: -11% instructions, -6%
cycles.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
- tables as std::array; the frames of the first 64 levels stay a C array
(not initialized on purpose, NOLINT)
- \u escapes are decoded with the library's hex_codepoint() instead of a
second table
- the parse failure is private, with an accessor; the special member
functions of the builder are all declared
- no nested conditional operators; explicit parentheses; a repeated
branch body merged; auto for casts
- the test's C arrays, fixed seed, and escaped literals are marked, as in
the other tests
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Internal parts of the zero-copy view (#5295), under detail/view and not
included by json.hpp, so that users of json.hpp compile nothing of it:
- macro_scope.hpp/macro_unscope.hpp: the few macros the view needs, under
its own prefix (json.hpp undefines its own at its end); the throw macro
honors JSON_NOEXCEPTION and JSON_THROW_USER like JSON_THROW
- string_ref.hpp: std::string_view from C++17 on, else a small stand-in
- node.hpp: the 16-byte node of the index; its kinds are value_t values
(checked by a static_assert)
- document_data.hpp: the storage of a parsed document (node array, decode
arena, owned input)
- scan.hpp: string and digit scanning with unrolled checks at fixed offsets
(after yyjson) and the library's SWAR and UTF-8 checks, independent of the
byte order
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Niels Lohmann