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>
Changes promised in the replies to the review of #5620:
- read() reuses the owned copy: a copied input (an rvalue byte container,
a stream, FILE*, wide characters, an iterator range) is appended to the
cleared owned buffer instead of a fresh std::string being move-assigned
over it. A moved-in std::string still replaces it.
- The error replay runs the library parser with a SAX handler that stores
nothing and throws the reported exception, instead of
BasicJsonType::parse(), which built a DOM up to the error.
- Pointers to and arrays of wide characters, and FILE*, take the input
adapter path (converted to UTF-8 into an owned buffer) instead of
failing a static_assert. Wide pointers are NUL-terminated.
- Streams are read in blocks through rdbuf() into the buffer, instead of
character by character through the input adapter; eofbit is set at the
end, as by the library's stream adapter.
- tape/tape_size/tape_cap are now nodes/node_count/node_capacity.
- The builder's members get descriptive, m_-prefixed names.
- The shallow frame stack is a std::array (in a wrapper with a
user-provided constructor: -Weffc++ wants the member initialized, and
value-initializing the 64 frames costs up to 30 % on small documents).
- The value dispatch macro has no NEXT parameter; the switch is left with
break, and the macro is undefined after its last use.
- literal() and digits_exceed() deduce the length from the string literal;
each literal has one helper; the two `if (!is_float)` blocks are merged.
- arena_ensure() sits right before the writes; the decode arena is a
std::vector<char>.
- The builder records the deepest nesting, and materialize() sizes its
stack from it once (growing it only if a subtree is deeper).
- Comment fixes (hex4, slow_string, the NUL rule, put_utf8, 0 - m, the
arena resize, grow()) and formatting (macro backslashes, digit cases,
one power of ten per line, value_t assertion next to is_container()).
- build_with's template parameters follow the builder's order.
Greg's idea of keeping the source offset in `off` and the arena offset in
`next` for escaped strings was measured and not adopted: string access got
7-18 % slower with a select, and 3.5-18 % slower with a branch-free field
pick (Apple M-series, clang -O3), parsing was unchanged.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
emit() moves the node array when it grows, and the subtraction read base
in the same expression, so the order was unspecified. MSVC Release builds
without forced inlining read the old base; the container index then
pointed outside the array and close() wrote out of bounds.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
reserve(n) and the growth of the index computed n * sizeof(node) without a
check, which wraps around on 32-bit targets for inputs of about 1 GiB and
allocates a too small array. Throw std::bad_alloc for a count beyond the
address space and clamp the growth step to it.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The non-little-endian path of the node writer wrote the integer's native
word over len and next, so len got the high half there. Compose and split
the value explicitly (len is the low half, next the high half); the
little-endian path stays a plain memcpy.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Add json_document and json_view, a read-only, zero-copy index of a
JSON text, as the first public slice of the zero-copy view (#5295).
A parse produces a flat array of 16-byte nodes in document order,
one per value and one per object key. Strings stay in the source
text; escaped strings are decoded into an arena. Integers are
converted while their digits are in the cache; floats keep only
their digit layout and are converted on read. Containers store the
size of their subtree, so a reader can step over one in constant
time. A document makes a handful of allocations, however many
values it has.
The parser accepts exactly what json::parse accepts, with every
combination of ignore_comments and ignore_trailing_commas, with and
without a trailing NUL, and under JSON_STRICT_NUL_HANDLING. It is
portable C++11 and does not depend on byte order.
basic_json_document adds parse, parse_copy, accept, read (reuses a
document's memory), root, is_discarded, source, owns_source,
node_count, memory_usage, and shrink_to_fit. basic_json_view adds
type, the is_* queries, operator bool, size, empty, materialize,
and source_offset. A parse error throws the same exception
basic_json::parse would throw for the same input, message and
position included.
detail::abi_config keeps JSON_STRICT_NUL_HANDLING and
JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON readable after json.hpp
undefines them, in the ABI namespace so they always match the
basic_json in use.
A NUL byte that ends a // comment is the end of the input, as in
parse() since #5696.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Niels Lohmann