* Fix develop CI: GCC 4.8, Clang 3.4 and clang-tidy errors
- unit-json_patch.cpp: move multi-line raw strings out of CHECK(), which
GCC 4.8 cannot parse inside a macro argument
- unit-recursion_depth_limit.cpp: value-initialize the const std::hash
(Clang 3.4 rejects default initialization of a const object)
- json_view: silence modernize-use-transparent-functors (std::less<> is
C++14), bugprone-suspicious-stringview-data-usage (the length comes from
the node) and the deliberate use-after-clear() / owning-memory hits in
unit-json_view_builder.cpp; use a raw string literal there
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix clang-tidy errors in the std::u8string tests
Test the lvalue path with the const original instead of a mutable copy
and the rvalue path with a temporary (misc-const-correctness), and mark
the intentional char8_t array with NOLINT (avoid-c-arrays).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Read a std::u8string from a JSON string and store it as one in every configuration
In C++20, json(std::u8string) was stored as a JSON string only when
std::filesystem::path support was available (the to_json overload sat
inside the JSON_HAS_FILESYSTEM guard), and there was no matching
from_json, so get<std::u8string>() threw type_error.302 ("type must be
array, but is string"). Without std::filesystem, the same expression
still produced an array of code units (the 3.12.0 behavior).
Move the to_json overload out of the filesystem guard so that a
std::u8string is a string in every configuration, build the value
directly from the code units (this also works for a custom string_t
such as the tests' alt_string), and add a from_json overload that reads
a JSON string. The array of numbers written by 3.12.0 is still accepted;
other types throw type_error.302 ("type must be string or array").
std::u8string_view and char8_t arrays are unchanged.
The tests now run in C++20/23 (the file mentions JSON_HAS_CPP_20/23),
also without std::filesystem (test-conversions2_no_filesystem), and with
ordered_json and a custom string type. The new behavior is documented.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fill the std::u8string with resize and indexing instead of push_back
The push_back loop is not vectorized: reading a 1 kB string took about
2.2 us against 0.39 us with resize and direct element access (clang -O2).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Write the std::u8string from_json template header in the compact form
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Ignore libstdc++ char to char8_t sign change in the clang sanitizer job
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Two bugs of 3.12.0 are fixed on develop without a note on the pages.
With JSON_DIAGNOSTIC_POSITIONS, start_pos() of a string that contains
escape sequences was too large; the position is now that of the opening
quote. A parse error for an unexpected number that is followed by a line
feed reported column 0 (for "1 2" and a line feed, line 1, column 0); for
"1", line feed, "2", line feed it now reports line 2, column 1.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* docs: state that NaN ordering inside containers depends on the standard
When the element-wise comparison of arrays or objects reaches a NaN and
a number (or another NaN) before any other difference, the result
depends on the language standard. Before C++20 the pair counts as
equivalent and the following elements (or the sizes) decide, so
[1] <= [NaN] is true. Since C++20 the comparison is unordered and
<, <=, >, >= all yield false. Document this on operator<, <=, >, >=
and <=>.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* docs: fix the FILE* notes of parse() and the locking promise
The performance note of parse() said a FILE* is read one character at a
time. With glibc and the Apple C library the characters in the FILE's
buffer are scanned in bulk; only elsewhere is it one character at a
time. The file is locked once only where the C library has unlocked
stdio functions (not on musl, MinGW, newlib, with
JSON_DISABLE_UNLOCKED_FILE_IO, or under MSan). Qualify the promise on
parse.md, sax_parse.md, the parsing page and the performance page.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* docs: fix integer ranges in the MessagePack and BJData tables
number_integer_t is std::int64_t, so number_integer cannot hold values
above 9223372036854775807; those are number_unsigned. The MessagePack
row for uint 64 listed the range up to 18446744073709551615, and the
BJData table had a uint64 row for number_integer that can never be
selected. The CBOR table was already corrected.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* docs: name the macros that turn off FILE* locking and bulk scans
parse() and sax_parse() now say that the FILE* is locked unless
JSON_DISABLE_UNLOCKED_FILE_IO is defined to 1, and the input-source
table and the performance page name JSON_DISABLE_STREAM_BULK_SCAN and
JSON_DISABLE_UNLOCKED_FILE_IO with what each turns off.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The test for json_view documents over an alternative string type called alt_editable_document::load() and save(). Both exist only on little-endian targets, so the test suite did not compile on big-endian targets such as s390x. Guard the two image lines with the condition the other image tests use (unit-json_view_image.cpp, unit-json_view_positions.cpp).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Add JSON_UBJSON_VALUELESS_ELEMENTS_PER_BYTE to set the UBJSON whole-input budget
The whole-input budget for the elements of optimized UBJSON arrays of
type Z, T and F (1048576 plus 4 per byte read) rejects files written by
3.12.0 that have long arrays of these types, such as 1100 rows of 1000
nulls with a `$Z` annotation, and there was no way to read them. The
per-byte rate is now the macro JSON_UBJSON_VALUELESS_ELEMENTS_PER_BYTE.
Semantics:
- default 4, so the default build is unchanged (the generated code of
from_ubjson is identical at -O2 up to source line numbers);
- the value must be a positive integer: 0, negative, non-integral, bool,
character and wider-than-uintmax_t values fail to compile with a
static_assert that names the macro;
- a value that does not fit into std::size_t is the largest size_t, so
one spelling works for 32- and 64-bit targets;
- the limit base + rate * bytes_read saturates at the largest size_t
instead of wrapping around, for any rate and input size; a rate of
131072 (2^20 / 8) or more cannot be exceeded by any input (the
shortest header of an array of more than 32767 elements is 8 bytes),
which switches the whole-input budget off;
- the limit of one array (1048576) and the base (1048576) stay fixed
constants, so the writer fallback still follows the base only and a
reader with any rate reads everything the writer produces;
- the macro is not part of the ABI tag and must have the same value in
every translation unit; it is documented that way.
The memory the arrays can take in a DOM is about 20 bytes per rate and
byte of input (measured, macOS, containers of 9 * rate elements:
25 bytes per input byte at rate 1, 81 at 4, 312 at 16, 1226 at 64, 5493
at 256) plus about 33 MB for the base. A 3.12.0 row of N values has a
header of b bytes (5 up to 255, 6 up to 32767, 8 above), so a table is
read completely if N <= rate * b, and otherwise up to
(1048576 + h * rate, h = outer header: 6, 7 or 9 bytes) / (N - rate * b) rows; 1100 rows of 1000 nulls need
a rate of 8. The new page of the macro documents the formulas, the
measurements, and two examples that are run by the documentation checks.
The tests of the budget move to tests/src/unit-ubjson_valueless.cpp, which
derives every boundary from the macro and is built with the rates 1,
1000000 and 0xFFFFFFFFFFFFFFFF in addition to the default (the last two
check that nothing is refused and that the limit does not wrap around);
they fail with the unchanged header for rate 1, 1000000 and the maximum.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Refuse char8_t for JSON_UBJSON_VALUELESS_ELEMENTS_PER_BYTE and clarify its documentation
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Report an empty array for UBJSON `[$N#<n>` instead of n elements
A no-op (N) is not a value, so an optimized array of type N holds no
element whatever count it declares. The reader skipped the elements but
still passed the declared count to the SAX parser: a handler was told
start_array(n) and no element followed, and the DOM parser reserved
min(n, 16384) values (256 KB) for an array that stays empty. Nine bytes
(`[$N#l 0x7FFFFFFF`) therefore committed 256 KB: 2000 of them in one
input, 18 KB, committed 512 MB of heap (capacity 163,840,000 values for
the 10000 arrays of the original probe: 2.6 GB for 70 KB of input).
Untouched pages keep the resident size small, but the commit charge and
a 32-bit address space are consumed.
The reader now announces 0 elements for an optimized array of type N.
The loop that skips the elements is unchanged (it never ran). A count
that does not fit into std::size_t is still rejected by the count
reader; a count that is too large for the array type of the DOM (before:
out_of_range.408 "excessive array size" from the DOM handler) is
accepted now, because the array is empty.
Behaviour change: sax_parse reports start_array(0) for `[$N#n` (before:
n); from_ubjson no longer throws out_of_range.408 for an N-typed array
whose count exceeds the maximum size of the array type.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Clarify the reservation note for announced array counts
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Two views compared equal when the tokens of two floats were equal, and
the same node was equal to itself. An edit writes a NaN as the token
"nan", so a view of it (or of an array holding it) was equal to itself
and to another NaN, while json(NaN) == json(NaN) and view == json are
false. The token shortcut now excludes the edited "nan" token (found by
the missing digit layout), dispatched on the document mode by a tag: a
read-only document cannot hold a NaN, and its walk is unchanged
(instruction count and cycles on canada, citm and twitter within
noise). For editable documents the check costs about 2 % instructions
on canada (cycles within noise). The same-node shortcut is kept for
read-only documents and for editable ones that were never edited; a
document with edits is walked, so comparing it with itself is linear
instead of constant.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
image_check::full accepted a float node whose token is an integer literal
that fits 64 bits (e.g. 10446744073709551616 changed from 18446744073709551616).
The parser stores an integer node for such a token, so the loaded
document differed from what parsing the same text gives. check_float_ranges
now requires an integer token of a float node not to fit uint64_t (or
int64_t when negative).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
dump() said a view never throws type_error.316 because the text is always
validated. A document loaded with image_check::bounds or none can hold
invalid UTF-8: dump() writes it unchanged, and dump(ensure_ascii) throws
type_error.316. (The matching correction of number_token() is part of the
read() documentation change below which this commit is stacked.)
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Document that read() must not be given a view of the document itself
basic_json_document::read() releases the previous text, the decoded
strings and the edits before it parses its input. An input that points
into them (source(), get_string(), number_token(), or a string of an
edited document) is read after its memory was cleared or freed.
A guard that copies such input costs about 1 ns per read() (5 % of a
read of a 7-byte input), which this API exists to avoid, so state the
precondition in read.md, lifetime.md and the pages of the views that
hand out such strings instead, and show how to copy the string first.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Assert in debug builds that read() is not given its own text
read() releases the document's text, decoded strings, owned image and
edits before it parses, so an input that points into them is read after
release. Without NDEBUG, read() now asserts that a borrowed input, a
copied contiguous input and (before C++17) a string_ref input are not
held by the document: its owned text, its decoded strings, an owned
image, and the current buffer of edited strings. Release builds keep
the documented precondition without a check, which would cost 2-4.5 %
per read() of a small message.
JSON_ASSERT is undefined at the end of json.hpp, so the view uses
assert() through a new internal NLOHMANN_VIEW_ASSERT.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Cause: #5476 made the DOM parsers reserve storage (up to 16384 elements)
for a counted array of CBOR, MessagePack, UBJSON and BJData when it
opens, before any input backs the count. The arrays that are open at the
same time keep their storage until the parse ends, so N nested headers
that announce 16384 elements and are never completed hold
N * 16384 * sizeof(json): 100,000 nested CBOR headers `99 40 00`
(300,000 bytes) peaked at 1.65 GiB, 3.12.0 at 12 MiB; the same for
MessagePack `DC 40 00`, UBJSON `[#I 16384` and BJData.
Fix: json_sax_dom_error_state, the base of the three DOM parsers, keeps a
credit of elements that may still be reserved: 65,536 at the start plus
2 for every byte the reader has read. An array that announces no more
than the credit left is reserved as before (one comparison and one
subtraction per counted array); a larger one is reserved what is left
and grows as its elements arrive. binary_reader tells a DOM parser (a
handler whose pointer converts to json_sax_dom_error_state*; a handler
that derives from one privately or protectedly, or any other SAX handler,
is not told anything and reserves as in the base of this change) where to
find its count of characters read for the duration of sax_parse(). An
RAII guard puts back the previous binding on every exit, also by
exception, so the parser never keeps a pointer to a reader that is gone
(unbound, only the per-array cap of 16384 applies); the public SAX
interface is unchanged, and start_array() still reports the announced number. A
bound per array by the remaining input would not help, since the same
bytes would back every open level; the bound has to be over all arrays.
The elements of an array take at least one byte of input each, so the
credit is more than honest input needs: none of the 19 inputs measured
(canada, citm_catalog, twitter, mesh, update-center, random, numbers,
animation, ik_rig, instruments, geo_response, jeopardy and 7 synthetic
ones, each as CBOR, MessagePack, UBJSON and BJData, with and without
size/type optimization) was given less than it announced. A value that
nests five or more arrays of 16384 elements, each with an array as its
first element, is the exception: the later ones grow as before.
Measurements (clang -O2, macOS arm64, peak memory footprint of
`/usr/bin/time -l`, reproducer: n nested headers):
n = 20,000 n = 100,000
develop fix 3.12.0 develop fix 3.12.0
CBOR 338.7 5.9 8.2 1687.7 23.3 11.8 MiB
MessagePack 338.9 6.0 7.3 1687.9 23.8 11.9 MiB
UBJSON 339.0 7.6 8.0 1687.9 30.1 11.6 MiB
BJData 339.0 7.3 7.9 1688.0 30.8 11.6 MiB
That is at most (65,536 + 2 per input byte) values: 1 MiB plus 32 bytes
per input byte with 16-byte values (a value is 24 bytes with
JSON_DIAGNOSTICS and up to 40 with JSON_DIAGNOSTIC_POSITIONS), as
documented in untrusted_input.md. Lower fixed caps do not get there: a
cap of 64 elements still peaks at 113 MiB (n = 100,000) and costs 4.7 %
on arrays of 300 elements (mat300.cbor) and 5.0 % on arrays of 10,000
(arrays_of_big.cbor).
Reader instructions retired (min of 7, 108 rows: 18 inputs in the 6 forms
CBOR, MessagePack, UBJSON and BJData plain and optimized; jeopardy, 46 MB,
is within 0.1 %), compared with the base of this change: -0.19 % to +0.37 % in the formats that announce
sizes (canada.cbor 81.9 M -> 82.1 M, many_small.cbor 333.5 M -> 334.2 M);
the benefit of #5476 is kept (without any reserve canada.cbor takes
116.8 M, +42.6 %, many_small.cbor +71.5 %, mat300.cbor +81.6 %). Rows of
plain UBJSON and BJData, whose arrays announce no size, move by -2.7 % to
+2.6 % without any code of theirs changing: the base alternates
between two instruction counts per process (452 M and 463 M for
many_pairs.ubjson, equal in cycles), and a struct with three unused
members added to the same class reproduces the 474 M mode.
Tests (CBOR, MessagePack, UBJSON, BJData): 200 nested headers that
announce 16384 elements peak at 52,433,600 bytes of DOM allocations on
the base and below 128 Ki values with the fix (scaled by sizeof(value),
so that it holds with the diagnostic macros) (counting allocator, no
process memory measured); an array that completes is still reserved in
one allocation; CBOR also checks that a DOM parser is not bound to a finished
parse (ASan: stack-use-after-scope without the guard), that wrappers
derived publicly (also with a parse inside the parse), privately and
protectedly compile and are bound or not as described, that an array that got less than it
announced decodes and that a SAX handler is told the announced size.
The new tests take 4.6 s under Valgrind at -O0 (gcc:12, all four files).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
A no-op (N) is not a value, so an optimized array of type N holds no
element whatever count it declares. The reader skipped the elements but
still passed the declared count to the SAX parser: a handler was told
start_array(n) and no element followed, and the DOM parser reserved
min(n, 16384) values (256 KB) for an array that stays empty. Nine bytes
(`[$N#l 0x7FFFFFFF`) therefore committed 256 KB: 2000 of them in one
input, 18 KB, committed 512 MB of heap (capacity 163,840,000 values for
the 10000 arrays of the original probe: 2.6 GB for 70 KB of input).
Untouched pages keep the resident size small, but the commit charge and
a 32-bit address space are consumed.
The reader now announces 0 elements for an optimized array of type N.
The loop that skips the elements is unchanged (it never ran). A count
that does not fit into std::size_t is still rejected by the count
reader; a count that is too large for the array type of the DOM (before:
out_of_range.408 "excessive array size" from the DOM handler) is
accepted now, because the array is empty.
Behaviour change: sax_parse reports start_array(0) for `[$N#n` (before:
n); from_ubjson no longer throws out_of_range.408 for an N-typed array
whose count exceeds the maximum size of the array type.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
An optimized UBJSON array of type Z, T or F has no payload, so its
declared count is not bounded by the input. The limit of 1048576 applied
to each array on its own, so 9 bytes (`[$Z#l` and a count) produced
1048576 values and every further 9 bytes another 1048576: 362 bytes
allocated 661 MB in a DOM, and a SAX handler received 1M events per 9
bytes.
The reader now adds the counts of all such arrays of one input and
rejects the input with out_of_range.408 when the sum exceeds 1048576
plus 4 for every byte read so far. The base is the per-array limit, so
an array that was accepted on its own before is still accepted at the
start of an input. These arrays add about 80 bytes of memory per input
byte (measured 77; one `Z` per byte needs 32), plus the base; the memory
of a document as a whole is not bounded by this. The check runs once per
valueless array, before its start event, with plain size_t arithmetic
that cannot overflow for a 32-bit size_t. Containers with a payload type
and objects (every member needs a key) are bounded by the input already
and are unchanged. BJData rejects these types.
to_ubjson counts the elements of the typed arrays it writes in one call
and writes an array in the unoptimized form (`[#n Z Z ...`) once it would
bring the total above 1048576, which a reader accepts at any input size.
The check is made only for a typed array of a valueless type, so BJData
and other arrays do not pay for it. Everything to_ubjson writes in
version 3.13.0 can therefore be read back.
Two consequences:
- to_ubjson output changes for values with more than 1048576 elements in
arrays of type Z, T or F: such an array is written at one byte per
element, e.g. 270000 rows of 100 nulls take 26.8 MB instead of 1.35 MB.
- Output of version 3.12.0 and earlier (typed arrays without such a
total) is rejected by from_ubjson when the arrays hold more than
1048576 elements plus 4 per byte of input. 3.12.0 writes a row of up to 255 nulls in 5 bytes (6 above
255), so a table is rejected from 1075 rows of 1000 nulls or 209723
rows of 25 (accepted: 1000 rows of 1000, 209000 rows of 25); rows of
20 or fewer are never rejected.
Measured on the test data: no file of json_test_data (converted with
to_ubjson in all three modes, 846 round trips) holds an array of this
kind with more than 0.17 values per byte. All of them are read back.
Reproducer (K arrays of 1048576 nulls), Apple clang -O2:
K=40 (362 bytes): before 41943040 values, 664 MB; after rejected, 33 MB
K=2 (20 bytes): before 2097152 values, 53 MB; after rejected, 33 MB
The new tests count events with a SAX handler where they can: building
and comparing million-element values made the first version take 176 s
under Valgrind, the current one about 22 s.
Behaviour change: an input that holds more than 1048576 elements in
arrays of type Z, T or F at more than 4 per byte is now rejected.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The Dependabot bump to 10.1.0 (#6059) conflicts with the pinned
opentelemetry-api 1.37.0 (importlib-metadata<8.8.0), so the Semgrep job
failed with ResolutionImpossible. Ignore importlib-metadata in Dependabot
since semgrep pins it indirectly.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Read iterator pairs of std::string and std::vector as pointer ranges
Before C++20, a pair of iterators of a contiguous container cannot be told
from any other random-access iterator, so json::parse(s.begin(), s.end())
on a std::string or std::vector<char> was read one character at a time,
without the bulk scans of strings, numbers and indentation that the
container itself gets (the container overload of input_adapter() already
maps it to a pointer range). The iterators of std::string and of a
std::vector of 1-byte integral elements are now matched exactly and mapped
to the pointer range they cover; an empty range is not dereferenced. The
character type stays the same, so the results and error positions do not
change.
parse(first, last) of a std::string gets 1.7-2.0x as fast in C++17 on an
Apple M1 Max, accept(first, last) 2.7-3.1x, and from_cbor(first, last) of a
std::vector 8-23% faster.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix CI: clang-tidy, unused result, Infer false positive in tests
- unit-deserialization.cpp: raw string literals, and assign the result
of parse() in CHECK_THROWS_AS (-Wunused-result)
- unit-bson.cpp: mark Infer's NULLPTR_DEREFERENCE on from_bson() with an
iterator pair as a false positive
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix implicit char conversion in iterator-pair test
The test built std::vector<unsigned char> and <signed char> from a char
range, which -fsanitize=integer reports as an implicit conversion.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Tests: sweep_allocator::construct() accepts pointers to const
Recent libc++ (emsdk latest) probes construct(const T*, const T&) in an
unevaluated context while checking for trivial copies; the static_cast to
void* in the noexcept specification then fails to compile.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Tests: guard sweep_allocator::construct() against const types instead of const_cast
The const_cast added in the previous commit is rejected by
cppcoreguidelines-pro-type-const-cast (clang-tidy). Make construct() not
viable for const U instead, which is what libc++ probes for.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix CI: astyle and clang-tidy for sweep_allocator::construct()
astyle 3.4.13 reformats the enable_if<!...> default template argument
(Check amalgamation). Delete the const overload instead, which makes the
libc++ probe for const objects fail in the SFINAE context just the same
and needs no <type_traits>.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Iterator ranges: astyle-safe trait members; reserve in the test helper
The byte check and the container match are static constexpr members
instead of integral_constant arguments and a default template argument
with comparisons, which astyle spaced out.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* SWAR: test for special bytes in the reduced form
The usual test for a zero byte, (x - ones) & ~x & high, needs `& ~x` only
to keep bytes from 0x80 out. The tests for quotes, backslashes and control
characters in string_scan.hpp and the serializer also flag every byte from
0x80 with `v & high`, so the `& ~x` terms can be dropped. Clang finds this
reduction itself; GCC does not and emits 24 instead of 16 instructions.
The masks are the same for 1.3 billion random words and for every pair of
adjacent byte values. With GCC 15 on x86-64: accept() -2 to -4 %,
sax_parse() -2 %, dump() -3 %. Clang is unchanged.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Vector scans: one compare for the quote and the control characters
The SSE2/NEON tests for the bytes a string run stops at compared with the
quote, with the backslash and with 0x20 separately. Flipping bit 1 maps
0x00-0x1F onto itself and the quote onto 0x20, and every other ASCII byte
to 33 or more, so (c ^ 2) < 33 finds the quote and the control characters
with one compare: signed where the bytes from 0x80 stop the run too,
unsigned (or with a saturating subtraction / minimum on SSE2) where they
do not. This saves one vector operation per 16 bytes in all twelve vector
tests: the lexer's vector_plain_run() and vector_utf8_run(), dump()'s
copy_plain_halves() and copy_plain_vectors(), and json_view's scanner and
serializer. The DEL tests of the ensure_ascii paths stay.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Tests: sweep_allocator::construct() accepts pointers to const
Recent libc++ (emsdk latest) probes construct(const T*, const T&) in an
unevaluated context while checking for trivial copies; the static_cast to
void* in the noexcept specification then fails to compile.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
(cherry picked from commit 794bf70823c12a2a1cb2a8a95f34dfb6357746a8)
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Tests: guard sweep_allocator::construct() against const types instead of const_cast
The const_cast added in the previous commit is rejected by
cppcoreguidelines-pro-type-const-cast (clang-tidy). Make construct() not
viable for const U instead, which is what libc++ probes for.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
(cherry picked from commit de099fa607001b710862fd1bbecef3609c54cbc4)
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix CI: astyle and clang-tidy for sweep_allocator::construct()
astyle 3.4.13 reformats the enable_if<!...> default template argument
(Check amalgamation). Delete the const overload instead, which makes the
libc++ probe for const objects fail in the SFINAE context just the same
and needs no <type_traits>.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
semgrep requires exact or compatible-release versions of mcp, boltons and
the opentelemetry packages. Bumping them alone makes the Semgrep job fail
at pip install (see #6054, #6055, #6057, #6058, #6060, #6061), so leave
them to move with the semgrep bump.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Regenerate the hash-pinned requirements file. This raises the transitive
pins pyjwt 2.13.0 -> 2.15.1 and mcp 1.23.3 -> 1.29.0 (also click 8.1.8 ->
8.4.2, new structlog 26.1.0), closing 18 of the 23 open Dependabot alerts.
The remaining alerts cannot be fixed until Semgrep relaxes its pins:
mcp needs 1.30.0 (Semgrep pins mcp==1.29.0) and opentelemetry-api needs
1.45.1 (Semgrep pins ~=1.37.0).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* docs: update the README for the release
- Show the animated SVG feature slideshow (light and dark variant) instead
of json.gif.
- Thank 20 further contributors and extend two existing entries.
- Fix the bad_serializer example (invalid template header, second function
named to_json instead of from_json, missing template keyword).
- Add "Zero-copy views" to the table of contents, link ordered_json to its
API page, and point the speed design goal to the performance page.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* chore: remove the stale ChangeLog.md and docs/json.gif
ChangeLog.md was last generated in January 2025 and still ends at
"unreleased (2024-12-22)"; 3.12.0 shipped without an entry. The release
notes live on the GitHub releases page and in docs/mkdocs/docs/home/releases/.
Remove the file, the Makefile target that called github_changelog_generator,
and the README credit for that tool.
docs/json.gif is an old copy of the feature slideshow that nothing
references; the home page uses the SVGs built by docs/mkdocs/scripts/slideshow/.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* chore: remove the unreferenced docs/mkdocs/docs/images/json.gif
The README header and the home page use the SVG slideshow built by
docs/mkdocs/scripts/slideshow/.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
dump() batches float tokens of 16 or 17 digits with at least 8 fraction
digits. A damaged image loaded with image_check::bounds can record a
token with no digit before the point and 17 after it, which no JSON
number has. Its q = -17 is outside the range of locate_token(), which
then read fives[17] past its 17 entries; on x86-64 the resulting digits
had 18 digits and failed the assertion of write_decimal() (OSS-Fuzz
issue 583898810, json_view_image_fuzzer).
The batch now takes 8 to 16 fraction digits, which every valid token of
16 or 17 digits has, so it batches the same tokens of valid documents.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Install Semgrep from a new hash-pinned requirements file in
cmake/requirements (covered by Dependabot) instead of a version-only
"pip install semgrep==1.168.0", and pass --require-hashes to the astyle
and natvis installs in check_amalgamation.yml, whose requirements files
got their hashes in #6046. Addresses the Scorecard Pinned-Dependencies
code-scanning alerts for semgrep.yml and check_amalgamation.yml.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* docs: blog post: describe the performance work of #6020, #6028-#6031 and #6033
Theme 1 gets the iterator pairs of std::string and std::vector, a new
subsection on the fused contiguous parser after yyjson's reader, and the
FILE buffer window after gnulib's freadptr(). Theme 2 shows the reduced
SWAR test and the (c ^ 2) < 33 vector compare. Themes 4 and 6 get the
binary strings without a zero fill and the 64 KiB window of the vector
sink. Appendices H and J list the new pull requests and their checks.
The numbers of the suite, the timeline, the comparison and the compile
times are not remeasured yet.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* docs: blog post: remeasure on develop with #6029, #6031, #6033 and #6039
Rerun on develop 7fc8312c64 with the open pull requests #6029, #6031,
#6033 and #6039 merged, measured on November 7: the suite (median of 12
runs, now 3.3x), the timeline (358 commits, #6031 shown as a pending
step on the candidate, #6028 labeled, a tick for November), the compile
times, the stream gap, the allocator check and the comparison on the
Mac and on x86-64, including the held-out files.
The comparison now leaves out the "copy untimed" variants, as the post
describes, and the held-out chart no longer counts json::accept() as
another library. Appendix D, the Intel column of Appendix B and the
GCC 16/Clang 23 check on the Mac are still from November 3.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Parse JSON in one block of memory with a fused pointer-based parser
The parser and the lexer exchange one token at a time, and the lexer keeps
its state (the current character, a pending unget, where the token starts)
in members, which every character pays for: json::accept() of twitter.json
took 2.8x as long as json_document::accept(). detail::contiguous_parser
reads text in one block of memory through a local pointer, with the
grammar, the whitespace and the tokens in one loop (after yyjson's reader),
and drives a SAX handler: parse()'s DOM handler, accept()'s handler, or the
user's handler of sax_parse(). It is used with the default options (no
callback, no comments, no trailing commas; sax_parse() with strict = true).
Errors are reported by the library parser, so the exception, its message and
its position stay the same. contiguous_parser stops at the token that the
library parser would reject, before reporting it, so the events up to the
error are the same; then:
- parse() runs the library parser with a handler that only reports the
error (json_sax_error_reporter), with the lexer of parse(), which is
instantiated anyway;
- sax_parse() runs it with a handler that forwards only parse_error() to
the user's handler (json_sax_error_forwarder), so no event arrives twice;
- accept() returns the verdict.
A handler that returns false stops the parse as before. An error at the end
of a large input costs parse() about 1.2x of what it cost before.
On an Apple M1 Max: accept() 1.6-2.0x as fast, parse() 10-16% faster,
ordered_json 11-16%, sax_parse() with a user's handler 1.36-1.46x, small
messages 12-22% (parse()) and 2.1-2.4x (accept()). On x86-64 (GCC 15,
Clang 21): accept() 1.6-1.9x, parse() 5-17%, ordered_json 12-19%,
sax_parse() 1.32-1.57x. A translation unit that parses a string compiles 3%
(-O0) and 9% (-O2) slower.
unit-contiguous_parser.cpp compares the contiguous path with the byte path
(std::list iterators) on corner cases, generated documents, their
truncations and mutations, deep nesting, and narrow number types: the same
values with the same number types and float bits, exception messages,
accept() verdicts and sax_parse() events.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix CI: unused parameter without exceptions, cpplint, BUILD.bazel
- json_sax_error_reporter::parse_error: mark the exception unused, as
JSON_THROW does not use it with JSON_NOEXCEPTION
- unit-contiguous_parser.cpp: compare the exceptions of both paths only
when exceptions are enabled (invalid input aborted the test otherwise)
- silence cpplint's readability/braces false positives on the if
statements inside the parser's macros
- add contiguous_parser.hpp to BUILD.bazel
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* accept(): skip digits without computing their value
accept() needs the value of a number only for a float whose exponent may
reach the largest value. skip_digits() finds the end of the digits without
the multiply-adds of accumulate_digits(), and once an 8-byte word fails the
digit test, the remaining loop needs no bounds check: a non-digit is in
that word. The value is computed afterwards for the rare floats that need
it.
accept() of canada.json: -16 % on Apple M1, -22 % with GCC 15 and -21 %
with Clang 21 on x86-64; files with many numbers -9 to -15 %.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Tests: mark parser callback in unit-contiguous_parser noexcept
Fixes the -Werror=noexcept failure with GCC 26 (ci_test_standards_gcc (26)).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Tests: sweep_allocator::construct() accepts pointers to const
Recent libc++ (emsdk latest) probes construct(const T*, const T&) in an
unevaluated context while checking for trivial copies; the static_cast to
void* in the noexcept specification then fails to compile.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
(cherry picked from commit 794bf70823c12a2a1cb2a8a95f34dfb6357746a8)
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Tests: guard sweep_allocator::construct() against const types instead of const_cast
The const_cast added in the previous commit is rejected by
cppcoreguidelines-pro-type-const-cast (clang-tidy). Make construct() not
viable for const U instead, which is what libc++ probes for.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
(cherry picked from commit de099fa607001b710862fd1bbecef3609c54cbc4)
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix CI: astyle and clang-tidy for sweep_allocator::construct()
astyle 3.4.13 reformats the enable_if<!...> default template argument
(Check amalgamation). Delete the const overload instead, which makes the
libc++ probe for const objects fail in the SFINAE context just the same
and needs no <type_traits>.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N5NasEsrzHhSNouxSoEwqz
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Contiguous parser: review fixes (wording, macros, diagnostic positions)
- "default parser" instead of "library parser" in comments and tests
- the three comments in json.hpp say the same thing
- current() replaces the repeated `p != end ? *p : 0`
- the backslashes of each macro are aligned
- the docs say that JSON_DIAGNOSTIC_POSITIONS turns the fast path off
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
Add sha256 hashes for every release file to the requirements of the
astyle and natvis tools, and install Meson in the ubuntu workflow from a
new hashed requirements file with --require-hashes instead of an
unpinned "pip install meson". pip verifies all installs from these files
once they contain hashes, including the ones of the Makefile and
cmake/ci.cmake. Addresses the Scorecard Pinned-Dependencies alerts for
ubuntu.yml; the check_amalgamation.yml flags follow once develop has the
hashes.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Autobuild compiled the whole test suite. The resulting database was so
large that the hosted runner repeatedly lost communication with the
server while importing the TRAP files (about 80 minutes in), which left
failed CodeQL analyses on develop and on pull requests. Analyze the
sources without building them instead.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
#6018 split features/json_view.md into features/json_view/ sub-pages while
#6022 added a blog post linking to the old page, so `mkdocs build --strict`
failed on develop (Publish documentation run 42577793091). Point the links
to json_view/index.md and json_view/editing.md.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Split the json_view feature page into sub-pages
The single features/json_view.md page grew to ~600 lines covering eight
topics. Split it into features/json_view/ with an overview page and
pages for ownership and lifetime, differences from basic_json, reading
values, generic code, editing, and images.
The new differences page collects what was spread over several sections
("What is the same", "What is different", "Duplicate keys", and notes in
the conversion and serialization sections) under one summary table, and
adds the parts of the basic_json interface a view does not have. The
other pages now link to it instead of restating the rules.
Internal links across the documentation point to the new pages.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Docs: fix stale comments and trim repeated text in json_view pages
- fig-jsonview.js/.css: point the header comments at the new pages
- reading.md: views' strings are valid as long as the document (and a
borrowed source text), as lifetime.md says
- reading.md, generic_code.md: link to differences.md instead of
restating the member-order and repeated-key rules
- differences.md: drop the third mention of ordering comparisons, name
rbegin()/rend() in the iterator row
- exceptions.md: link editable documents to editing.md
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
#5835 replaced the switch on the value type in operator*, operator->,
operator++, operator-- and operator== of iter_impl with an if chain that
tests for an array first. Every step over an object then pays for a
failed test and a taken branch: with the iterator inlined, walking
twitter.json took 14% more cycles than with 3.12.0, and citm_catalog.json
3.5% more.
Test for objects first. Walking every value of the three files now
takes fewer cycles than with 3.12.0 (Apple M1 Max, Apple clang 21):
citm_catalog.json 1.21x, twitter.json 1.04x, canada.json 1.17x as fast.
canada.json, which holds only arrays, gives back part of the gain of
#5835 (develop: 1.62x); the other orders tried (arrays first without
hints, objects first without hints, a switch) each left one of the
files slower than 3.12.0.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Niels Lohmann
dependabot[bot]
