Resolve the json_sax.hpp conflicts (keep the shared error state next to the
object key index; the binary parser uses the shared error state as well) and
drop the else after return in the diff stack-frame code.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Reserve the parser's nesting stack once
The stack of open arrays and objects was a std::vector<std::uint8_t> that
allocated at the first push_back and then reallocated at 1, 2, 4, 8 bytes,
so a document nested four levels deep made four allocations before it
parsed a value. A document that starts with a container now reserves 32
levels in one allocation; deeper documents grow as before. The loop that
reads the tokens is the code it was, so large documents are not affected
(a stack object with its own storage, and a reserve at every container,
both cost the token loop instructions or cycles).
Fixed cost of parsing a 124-byte object (macOS instructions): accept
-7.6 %, DOM -2.5 %; no change for a single number.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Build arrays of the DOM parser in one exact-size allocation
Reading an array of unknown length appended its elements to a std::vector
one at a time, so a small array reallocated (and moved its elements) at
sizes 1, 2, 4, ... The DOM SAX parser now collects the elements of arrays
whose length is unknown on one stack shared by all open arrays and moves
them into an array of exactly the right size when the closing bracket
arrives. An array that grows past 16 elements switches to appending to its
own vector, so large arrays pay no second copy. Binary formats, which
announce the length, and builds with JSON_DIAGNOSTICS or
JSON_DIAGNOSTIC_POSITIONS (whose values need their final place from the
start) keep storing into the array directly.
The stacks of open containers are reserved on first use instead of growing
from one element.
The binary readers get a DOM parser of their own, json_sax_dom_binary_parser:
it is the plain parser that stores every value directly in its container
(the code the readers used before). Their containers announce their size, so
the bookkeeping of the collecting parser (frame entries, the destination of a
finished container) was pure overhead for them: with the collecting parser
FromBinaryShape nested was 3-12 % and containers up to 5 % more instructions
(GCC) than before. With the plain parser Linux Ir of FromBinaryShape and
FromBinaryBuffer is within 0.8 % of develop (Clang: identical).
canada.json has 56 000 two-element arrays that took three allocations each
and now take two. DOM parse Ir on Linux against develop (GCC / Clang): canada
-7.0 % / -7.7 %, citm -0.8 % / -1.6 %, 124-byte object -5.4 % / -5.5 %;
accept and SAX are untouched.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Reuse one value for the key events of the callback parser
parse() with a callback built a new string value for every key it passed to
the callback (a heap allocation and release per key), and copied the key a
second time into the object. The callback parser now keeps one string value
for key events, which it assigns, and moves the key into the object.
Callback parse, Linux Ir (GCC / Clang): twitter -13.3 % / -11.7 %, citm
-9.5 % / -7.6 %. (Earlier macOS measurement: twitter -19.5 % instructions.)
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Remember where the DOM parser stores the next value
The DOM SAX parser looked up the innermost open container for every value:
the end of its frame stack, then the object or array pointer in the frame.
Compilers have to reload all of that after each value is stored, since a
store of a value may alias the frame. The parser now keeps the innermost
container's kind and, for an array that stores directly, the array itself
in members, updated when a container opens or closes.
Flat arrays of numbers are the case that mattered: parsing floats.json
(1 000 000 doubles) into a DOM took 2 % more cycles than before the
previous commit on macOS and now takes 0.4 % fewer than the code before it
(signed_ints.json -0.8 %). Linux Ir is unchanged (canada -6.1 % GCC /
-6.9 % Clang, citm -0.3 % / -0.8 %, twitter +0.2 % / 0.0 %).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Look up the keys of big ordered_json objects in a hash table
ordered_map keeps its elements in a vector and searches it linearly for
every key it adds, so parsing an object with n keys into ordered_json took
O(n^2) key comparisons: 33 billion instructions for one object with 30 000
keys, where json needs 245 million. From 64 keys on, the DOM parser now
keeps a hash table of the keys of the open object (hashes and positions
only; the elements stay in the vector) and finds duplicate keys there. The
result is the same, including the position of a key that occurs twice.
Smaller objects, std::map based objects and key types without std::hash
are untouched.
ordered_json parse, macOS instructions: object with 1 000 keys 64.8 M ->
14.6 M, 10 000 keys 5 104 M -> 38.7 M, 30 000 keys 33 358 M -> 89 M.
Linux Ir of an object with 2 000 keys: 43.0 M -> 1.9 M (GCC), 37.9 M ->
1.8 M (Clang).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Trim the work the DOM parser does per container and per member
Opening and closing an array or object cost the DOM parser more than before
the exact-size array change, which showed on documents of many small
objects (jeopardy.json: +1.4 % Ir with GCC). This takes it back:
- the stack of open containers keeps its first eight entries in the parser
object instead of in a vector, so it needs no capacity check beyond the
push itself and no allocation for typical documents;
- a container is created in the place it ends up in, instead of as a
temporary that is then moved there;
- a new object member, which is null, is constructed over the null instead
of being swapped with it (only with the default base class of basic_json);
- the key index that big ordered_json objects use is only a part of the
stack entries of objects that can have one;
- an array that outgrows the stack reserves twice what it has so far, as a
vector that grew by itself would have.
DOM parse Ir of the series up to here against develop, Linux GCC / Clang:
canada -6.6 % / -7.3 %, citm +1.2 % / +0.2 %, jeopardy +0.4 % / -0.1 %,
124-byte object -6.5 % / -6.2 %. (twitter measures +6.8 % / +6.6 % at this
point in steady state, almost all of it in glibc's malloc/consolidate for the
new allocation sizes; with the last commit of the series it is -0.1 % / 0.0 %.)
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Test parsing containers of the sizes where the DOM parser changes storage
The DOM parser collects the elements of an array on a stack (up to 16) and
the members of an ordered_json object (up to 32) before it moves them into
the container, and uses a hash table for keys from 64 members on. Parse
arrays and objects of the sizes around these limits, nested into each other,
with repeated keys, and as random trees that are built next to their text,
and compare json and ordered_json against the value built through the API.
The tests pass on the code from before these changes as well. The random
trees are what finds the mistakes that are easy to make here: while the
collecting change was being written, a collected array that outgrew the
stack at the moment its parent had just reached the limit left its elements
in the parent, and only these tests (142 failed assertions) showed it.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Do not convert floats that cannot overflow in accept()
accept() only has to find out whether a float token is finite: json_sax_acceptor
ignores the value, and the parser rejects only a float that overflows. A
float whose digits fit into 19 digits and whose decimal exponent is at most
max_exponent10 - 19 is below 10^max_exponent10 and cannot overflow, so the
lexer, told by accept() that the value is not needed, skips the conversion
for those. A token that may overflow, with more digits, or in another float
format is converted as before, so 1e309 and its neighbours are rejected as
they were (the new test compares accept() with parse() on tokens around the
limit).
json::accept(): canada.json Ir -8.6 % (GCC) / -10.9 % (Clang), 124-byte
object -3.2 %; cycles on the M1 (release benchmarks, LTO): canada -7.8 %,
floats.json -9.0 %. The check tests the accept() flag first, so the parser
that needs the value pays one load and branch: SaxParse canada +0.35 % Ir
(GCC and Clang), DOM canada +0.4 %.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Collect the members of ordered_json objects on a stack until they are known
An ordered_json object keeps its members in a vector and appends every new
member to it, so reading an object of unknown size reallocated the vector at
1, 2, 4, 8, ... members, and ordered_map copies all keys on every
reallocation. The DOM SAX parser now collects the members of such objects on
one stack shared by all open objects, as it does the elements of arrays, and
moves them into a vector of exactly the right size when the closing brace
arrives. An object that gets past 32 members switches to appending to its
own vector (and, from 64, to the key index of the earlier commit). The
members are the same, in the same order, and a key that occurs twice keeps
the place of its first occurrence. std::map based objects, objects whose
size the input announces (binary formats), and builds with JSON_DIAGNOSTICS
or JSON_DIAGNOSTIC_POSITIONS are untouched.
What only ordered objects need (the collected members, the key indexes) is
a member of the parser only for them, and the code that stores a value tries
the array that takes its elements directly first. An array that outgrows
the stack now reserves 32 elements, the capacity a vector that grew by
itself has at that point; with the larger reserve before, parsing
floats.json (1 000 000 numbers) was 2 % slower on Apple M1 than before this
series.
ordered_json parse, Linux Ir against develop (GCC / Clang): citm -11.6 % /
-11.7 %, twitter -8.4 % / -8.4 %, jeopardy (12 000 objects of 7 keys)
-10.5 % / -8.0 %, canada -3.9 % / -5.2 %, 124-byte object -20.9 % / -16.5 %.
macOS time (M1, earlier base): citm -19 %, twitter -13 %, jeopardy -13 %,
canada -14 %, 124-byte object -26 %. json parse: steady-state twitter -0.1 %
(see the previous commits), citm -1.2 %.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Rename the key index's match callback so Flawfinder doesn't read it as std::equal
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Bind the parser test's strings by reference in the range-for loop
GCC and Clang 14 (-Wrange-loop-construct) reported the copy.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Collect container values only in basic_json's own parse functions
A json_sax_dom_parser that a user constructs (as in the sax_no_exception
example) must hold the values read before a parse error, as on develop:
its value is used after the error. The values of containers of unknown
size are now collected on a stack only with the new template parameter
Collect (default false), which json::parse() and the json_view
materialization set; without it, every value is stored in place as it
arrives. A test checks the values kept after an error for json and
ordered_json.
Also fixes the clang-tidy, cpplint and Infer findings: the frame stack
keeps its first frames in a std::array (left uninitialized) and the
others in a std::unique_ptr, auto for create<>(), NOLINT for member
initializers -Weffc++ needs, and an @infer-ignore for the nullptr that
place_container() does not use.
json::parse() and ordered_json::parse() instructions are unchanged
(Apple clang; twitter, citm, canada, jeopardy: -0.03 % to +0.25 %).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Clarify the DOM parser's collection stacks and rename reenter_enclosing_container()
Address review comments: explain why a closed container erases only the top
of the shared stack, set the payload before the type, rename the helper that
makes the enclosing container current again, and declare the parser's state
stack right before reserving it.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Export json_sax_dom_parser from the C++20 module for MSVC
MSVC needs the detail classes that basic_json befriends to be visible in
the module (#3970, as json_sax_dom_callback_parser); with its new
template parameter, json_sax_dom_parser failed in ci_module_cpp20
(C2039, not a member of nlohmann::detail).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Keep the overflow-limit numbers of the parser test in a std::vector
clang-tidy (cppcoreguidelines-avoid-c-arrays) rejected the C array.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix the clang-tidy findings of the ordered_json parser tests
The container sizes are a std::array, the expected text uses raw string
literals, the generator's arithmetic is in parentheses, auto for the
result of make<>(), and single characters are appended as characters.
clang-tidy now reports nothing on the lines the branch adds to
unit-ordered_json.cpp and unit-class_parser.cpp (apart from checks CI's
clang-tidy does not have).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* 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>
* Write the indented output of json_view's dump() with the fast writer
The compact writer of dump() (cursor in a register, fixed-size copies of
tokens) served dump(-1) only; every dump(indent), and everything with
ensure_ascii that was indented, went through a generic walk that calls put()
for each token and put_repeated() (a memset call) for each line break. The
fast writer now writes indented output, too, with or without ensure_ascii, and
the generic walk is gone. A line break is a copy of a 64-byte ruler (a line
break and 63 indentation characters) when the line is at most 64 bytes, and a
store plus memset beyond that. The depth is counted only in the indented
instantiations (counting it in the compact one cost an editable document 5 %
with Clang). The indented walks are out of line (dump_pretty()): inlined
into dump() next to the compact ones, they made compact dumps of integer
arrays 7 % slower.
A decoded string (one with escape sequences) of a damaged image loaded with
image_check::bounds can contain invalid UTF-8. As before, the compact dump
writes such bytes unchanged, and the indented dump throws type_error.316; its
runs are found by the lexer's validating scan (a scalar validation of every
sequence made dump(2) of decoded CJK text twice as slow as before).
dump(2) instructions per call, macOS arm64 (before -> after):
twitter citm canada jeopardy(4k)
json 3.75M -> 1.85M 8.28M -> 3.93M 67.1M -> 56.0M 9.56M -> 4.81M
editable 3.82M -> 1.84M 8.48M -> 4.02M 68.1M -> 56.2M 9.66M -> 4.75M
Linux instructions (GCC 13 / Clang 18), dump(2) and dump(2, ' ', true):
twitter -25 % / -30 % -11 % / -11 %
citm -29 % / -28 % -24 % / -24 %
jeopardy(4k) -25 % / -32 % -20 % / -24 %
CJK text with \u escapes -6 % / -11 % -2 % / -2 %
The compact dump is unchanged (within +0.4 % on Linux, within noise on the Mac).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Scan the strings of json_view's dump() with ensure_ascii 16 bytes at a time
With ensure_ascii, a string of the source that has a byte >= 0x7F was written
by a generic path: put() per piece and a validate-and-decode step for every
character beyond ASCII. It now goes through a writer with the cursor in a
local variable. The part of the string that has nothing to escape is found
16 bytes at a time (SSE2 or NEON; the existing 8-byte scan without vector
support) and copied; strings that are ASCII throughout are still decided by
the 8-byte check of dump() with ensure_ascii (a 16-byte scan for them was
2-5 % slower for short strings). The two-byte sequences (lead C2..DF) and
the three-byte sequences whose continuation bytes may have any value (lead
E1..EF except ED) are decoded with a few compares; E0, ED, four-byte
sequences, and damaged input go through validate_one_utf8() as before, so
damaged images still throw type_error.316.
dump(-1, ' ', true) instructions per call, macOS arm64 (develop -> now);
the compact dump without ensure_ascii and dumps of ASCII text are unchanged:
twitter 5.42M -> 3.57M citm 3.28M -> 3.24M jeopardy(4k) 5.78M -> 5.65M
CJK text, raw 66.5M -> 35.3M escaped \u 32.7M -> 17.0M
Latin text with accents 11.5M -> 9.4M ASCII text 4.40M -> 4.40M
Linux instructions (GCC 13 / Clang 18), dump(-1, ' ', true):
twitter -9 % / -33 % CJK text -14 % / -47 % Latin text -8 % / -18 %
citm -0.2 % / -0.4 % ASCII text -0.1 % / -0.3 %
The compact dump without ensure_ascii is unchanged (canada 0.0 %).
The unit test covers every character of the Basic Multilingual Plane, and a
character that needs an escape at each position of strings of 0..40 bytes.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Find the end of the plain run of a decoded string 16 bytes at a time with SSE2
The strings of a document that had escape sequences are written in runs
without a quote, backslash, or control character; the compact dump found the
end of a run 8 bytes at a time with a SWAR test. With SSE2 (x86-64) it now
compares 16 bytes per step. NEON keeps the SWAR code: there the mask of a
step costs more than it saves on the short runs that are typical here.
Instructions of dump(), Linux (GCC 13 / Clang 18):
twitter.json 1.99M -> 1.88M (-5.2 %) / 1.97M -> 1.92M (-3.0 %)
CJK text with \u escapes 5.98M -> 4.52M (-24.5 %) / 5.30M -> 4.25M (-19.9 %)
ASCII text with a few escapes 4.28M -> 4.09M (-4.4 %) / 4.24M -> 4.11M (-3.2 %)
jeopardy (4,000 questions) 5.08M -> 4.78M (-5.8 %) / 4.97M -> 4.80M (-3.4 %)
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Rename depth_counter's open() and close() to enter() and leave()
Flawfinder took them for the file functions open() and close().
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix the clang-tidy findings in json_view's dump and its tests
- the SSE2 loop of the string scan is within NOLINTBEGIN/END
(portability-simd-intrinsics), as in to_chars.hpp
- depth_counter<false>'s leave() and get() keep the interface of
depth_counter<true> instead of being static
- write_ascii() compares the room left as a size_t variable
- the test's texts are a std::array, and a JSON fragment with escapes
is a raw string literal
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Move dump_ascii() and dump_pretty() next to dump()
Address a review comment: the two dispatchers that dump() calls now follow
it directly.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
- throw_on_discarded() and throw_integer_out_of_range() take an
input_format_t and call input_format_name() themselves
- has_no_children(), set_parents() and max_size() check array, then
object, instead of calling is_primitive() first
- parse_impl() chains like accept_impl(); sax_parse_impl() uses if/else
- std::array for the control character messages; negation<> in
write_object_key(); clarify the is_primitive() comment
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
- json_sax_dom_error_state::is_errored() cannot be constexpr: the class
is not a literal type (GCC 4.8, GCC -Wpedantic, Clang 3.5).
- to_binary_vector: destroy the vector writer before returning, as the
output_vector_sink sets the size of the vector in its destructor.
Without NRVO (MSVC Debug in C++14 mode), the vector was moved out
before it was resized.
- parser::sax_parse: return false instead of result after the early exit
(cppcheck identicalConditionAfterEarlyExit).
- json.hpp: export detail/exceptions.hpp for IWYU, so json_view.hpp can
name exception_id without including a detail header.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Taking the second-byte range from utf8_shape() made BON8 decoding 2-3% slower
(FromBinaryShape/scalars, FromBinaryBuffer, alignment-pinned builds), so the byte
path keeps its hand-written ranges. The exhaustive validator test still checks
that it accepts the same sequences as the other validators.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Moving parse_array_index out of json_pointer stopped clang from inlining it
into the pointer resolution (JsonPointerAt/citm_catalog +17%,
JsonPointerAt/twitter +8%); JSON_HEDLEY_ALWAYS_INLINE restores the baseline.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Calling get_number_value() from from_json(ArithmeticType&) made
GetArray<std::array<int, 1000>> 2.4x slower (it changes the inlining of the
std::array conversion), so that overload keeps its own switch, with a comment.
get_arithmetic_value() still uses the shared helper.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Move json_pointer::parse_array_index() and its status enum to namespace detail
(json_pointer keeps a forwarding member), and let the json_view pointer
resolution use it instead of its own copy of the RFC 6901 index grammar. Both
accepted the same tokens with the same outcome. Also update the comment on
detail::decode(), which listed four of the five UTF-8 validators and now points
to the test that keeps them in sync.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
sax_parse keeps its if (result) guard and return value. Also adds the
clang-tidy NOLINT for the 2.3.6 message table.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Holds errored, allow_exceptions, parse_error() and is_errored(). The scalar
handlers stay in each class (parse hot path).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
get_bon8_string takes the valid range of the second byte from utf8_shape()
(0xF5..0xF7 stay errors, unget behavior unchanged). validate_one_utf8 keeps its
if-chain (hot bulk path); the new test checks decode, validate_one_utf8,
utf8_shape, json::accept and the BON8 reader against each other for every lead
and second byte.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
- detail::exception_id at every numeric id; throw_* helpers take it
- hex_byte, hex4_lower, utf8_shape replace local hex/table code
- throw_array_index_out_of_range/_past_end, throw_key_not_found
- has_native_float_format replaces the hand-written IEEE check
No change of messages, ids or output.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The three number cases of get_arithmetic_value and the arithmetic
from_json share one helper (boolean stays accepted only by the latter);
the array and object reserve overload pairs become one.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Object and array stay the first branches; everything else is the
primitive case. operator== and operator< share the different-container
check.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The leading-slash and invalid '~' escape checks are shared helpers; the
reference-token walks test object and array first and treat everything
else as primitive.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The four create() bodies build their message through one protected
static helper; the diagnostics parent walk handles array/object and
ignores all other types.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
1.1: build ascii_copyable() at compile time from is_ascii_copyable and
test find_ascii_copyable_run against it for every byte. 1.2/2.5.4: use
detail::hex4_lower in both \uXXXX writers. 2.1.6: factor the repeated
flush/emit blocks in dump_escaped_impl into always-inline helpers.
2.1.7: comment why the escape copy does not use put_buffer.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The CBOR/MessagePack/UBJSON/BSON/BJData/BON8 literals in throw_on_discarded and
throw_integer_out_of_range calls now come from detail::input_format_name; the
texts are identical, and UBJSON vs BJData still follows use_bjdata.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
write_cbor_key and write_msgpack_key were the same SFINAE pair apart from the
string/scalar writers; both now delegate to write_object_key<Walk>. UBJSON key
and to_format_length stay as they are.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
The 13 oa.write_characters(reinterpret_cast<const CharType*>(x.data()), n)
sites now call write_bytes(const void*, size_t); copy_characters sites and the
buffer-based writes are untouched.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
write_ubjson_start_array/_object are now write_ubjson_start_container<IsArray>,
keeping the array-only begin()+1 scan and excessive_valueless rule.
ubjson_frame derives from binary_container_frame instead of duplicating it.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
write_cbor_iterative and write_msgpack_iterative (and their *_value_or_push
helpers) were identical but for the head, scalar, and key writers. They are now
one template over a small policy struct, so everything inlines. BON8 keeps its
own walker (open string run, closing marker), noted in a comment.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Array and object prefixes differed only in the fix marker and the 16/32-bit
markers; both now delegate to write_msgpack_container_prefix. Output is
byte-identical.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
write_bson_entry_header and write_bson_string re-derived the error-handler
policy differently. Both now use sanitize_utf8_for_bson_write, which treats
strict like keep (the size pass already validated) and sanitizes replace/ignore
into the bytes the size pass measured. No context plumbing is needed.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Add detail::is_primitive(value_t), detail::hex4_lower(), detail::utf8_shape(),
detail::input_format_name(), detail::empty_input_message(), the exception ids
116, 319, 320 and 416 to detail::exception_id, and make
detail::is_ascii_copyable() constexpr. Later commits switch the duplicated
code over to them.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
#5876 added JSON_LITTLE_ENDIAN_TARGET and JSON_BIG_ENDIAN_TARGET to
macro_scope.hpp. They must stay macros because they select code in #if,
so suppress the clang-tidy check as simd_scan.hpp already does.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Flatten into a std::map by inserting next to the previous entry
flatten() added every path with operator[], so each insertion into the
std::map of the result searched the tree from the root and compared long
shared path prefixes, which took half the time on canada.json. Paths come
out in nearly sorted order, so they are now inserted with emplace_hint
next to the entry added last, for object types that have it. ordered_map
keeps its append without lookup, and other object types keep
operator[].
Path segments are also appended in place, without a temporary string
for each escaped key or array index.
flatten() of canada.json, citm_catalog.json and twitter.json is 1.4x to
1.9x as fast; ordered_json is unchanged (1.05x). The output is
unchanged.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Unflatten without splitting keys or walking each path from the root
unflatten() split every key into a vector of reference tokens, stored the
pointer prefixes in a tree to decide which values become arrays, and then
looked up every token of every key from the root again.
Keys are now checked by a scan that raises the same parse_error.107/108 as
split(), skipping the part a key shares with the previous one, which was
checked already. The prefixes that denote arrays are kept in a small hash
set of key slices; the array/object decision is unchanged. The walk keeps
the values along the previous key's path, so only the part of a key that
differs from the previous one is looked up. Keys of a flattened object are
usually sorted, so that part is short.
unflatten() of canada.json, citm_catalog.json and twitter.json is 1.9x to
2.3x as fast; the output and the exceptions are unchanged.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Apply plain JSON Patch operations without building a json_pointer
patch() built a json_pointer for every operation, which splits the path
into a vector of strings, and for "add" and "remove" a second one for
the parent. That took 20% to 45% of the time not spent copying the
document.
"add", "remove", "replace" and "test" with a path that has no escapes and
denotes existing values (the parent, for "add"), as diff() creates them,
are now applied by resolving the path in place, reusing one string for
the reference tokens. Every other operation, and every operation that
would fail, takes the existing code, so the results, the exceptions and
the state left after a failing operation are unchanged.
patch() of the diffs of canada.json, citm_catalog.json and twitter.json
is 1.02x to 1.10x as fast; the rest is the copy of the document.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Build diff operations member by member
diff() created every operation from an initializer list, so each
{"op", "replace"} pair became a temporary array with two heap strings
before it was turned into an object member; an operation took about 25
allocations. The operations are now built as objects whose members are
assigned one by one. Paths of members and elements are also built in
place, without a temporary string for each escaped key or array index.
diff() of canada.json, citm_catalog.json and twitter.json, with every
eighth scalar changed, is 1.3x to 1.6x as fast; the output is unchanged.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Diff objects with the same keys in lockstep and keep one path buffer
diff() looked up every key of the source object in the target object,
and built a new string for the path of every value it descended into.
Objects of the same size usually have the same keys in the same order,
so they are now walked together as long as that holds, without a
lookup; objects that differ fall back to the lookup from the first key
that does not match. The path is one buffer that is extended before
descending and shortened again afterwards. Objects whose members are
equal but in a different order (possible with ordered_json) still get
no operations.
Together with the previous commit, diff() of canada.json,
citm_catalog.json and twitter.json is 1.5x to 2.0x as fast; ordered_json
1.65x to 1.85x. The output is unchanged.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Use std::array for the index buffer in append_index()
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Declare the partner iterator in the loop initializer in diff_object_keys()
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix GCC and clang-tidy warnings in the JSON pointer speedups
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Use standard library helpers where available, with C++11 fallbacks
Add detail helpers that call the standard library function when the
library provides it (checked with the __cpp_lib_* feature-test macros, so
no new ABI-tag macros) and a C++11 implementation otherwise:
- detail::clamp (std::clamp): by value, so that static constexpr members
need no out-of-line definition before C++17; used in json_view's dump()
buffer sizing, where step() can now be a static constexpr member.
- detail::bit_cast (std::bit_cast): replaces dtoa_impl::reinterpret_bits
and the memcpy bit copies in the number parser, to_chars, the json_view
serializer, and binary_writer::to_char_type.
- detail::byteswap (std::byteswap): one implementation instead of the
binary writer's own byte_swap() overloads and the bswap chain in
to_chars' store_msb_first().
- count_leading_zeros/count_trailing_zeros: two 32-bit _BitScanReverse/
_BitScanForward calls for MSVC x86/ARM32 instead of the loop, and
std::countl_zero/countr_zero for compilers without builtins.
- conjunction, disjunction, negation, bool_constant, and uncvref_t map to
their std counterparts in C++17/C++20.
Builtins and intrinsics stay ahead of the <bit> functions, as MSVC's
std::countl_zero checks for LZCNT at run time unless AVX2 is enabled.
The generated code is unchanged with Clang (C++11/20/23, arm64 and
x86-64) and GCC 16 for parsing, dumping, and the binary formats. The
binary reader keeps its own byte_swap() and the writer its memcpy calls:
the smaller code from detail::byteswap/std::bit_cast there made GCC 16
inline get_number()/write_number() differently, which slowed reading by
10-18 % and writing by about 1 %.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Name the target byte order once (JSON_LITTLE_ENDIAN_TARGET, JSON_BIG_ENDIAN_TARGET)
to_chars.hpp and bit_ops.hpp repeated the long __BYTE_ORDER__/_MSC_VER
conditions; macro_scope.hpp now defines them once (undefined again in
macro_unscope.hpp).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Explain why the serializer growth bound is not a clamp
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
With ensure_ascii, the view's dump wrote every string through the escaping
path, which decodes UTF-8 byte by byte, even strings of the source that are
plain ASCII. A string of the source without escape sequences has no quote,
backslash, or control character, so ensure_ascii leaves it unchanged
exactly when it has no byte >= 0x7F. The dump now checks that eight bytes
at a time (one add and one or per word) and copies such strings like the
dump without ensure_ascii.
The compact walk (dump_compact) now also serves ensure_ascii, through a
separate out-of-line instantiation, so the walk without ensure_ascii keeps
its speed.
Parsing is unchanged. A flag set by the parser would let the dump skip the
check and was about twice as fast for compact output, but it costs every
parse of string-heavy input 1-2% with GCC.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
dump() now checks strings of up to 32 bytes against a table of the bytes
that can be written verbatim, four bytes per step, and copies them in one
go if all of them can. That is cheaper than setting up the vector scanner,
whose setup costs more than such a string takes to scan; SQLite's
jsonAppendString() writes strings with the same kind of table loop. Most
object keys and many values are this short (citm_catalog averages 8 bytes
per string, jeopardy 14).
GCC 16 and 13 stopped inlining put_indent() and put_literal() into the
pretty-printing loops once dump_escaped() had grown, which made indented
output up to 13% slower; both are now always inlined.
The new test puts every ASCII character and a 2-byte character at every
position of strings of 1 to 40 bytes and compares the output with the
reference implementation of the existing UTF-8 tests.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
json_sax_dom_parser::key() copied the key from the lexer's buffer into the
object. The lexer starts every token with an empty buffer (reset()), so the
key can be moved instead, which saves an allocation and a copy for every key
that does not fit into the string's inline buffer. json_sax_dom_parser
already moved string values the same way. ordered_map has no rvalue
operator[], so ordered_json still copies keys.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Match literals without a stack array and keep parser states in bytes
scan_after_whitespace() built the expected text of true, false, and null in
local std::array objects. With -fstack-protector-strong (the default of
Ubuntu's GCC, among others), a local array gives the function a stack canary
check, which then ran on every token. scan_literal() now takes the text
after the first character as a string literal, with its length as a template
parameter; for contiguous input it compares the rest at once and falls back
to the byte loop on a mismatch, so errors are reported at the same
character. It is out of line, which keeps the token dispatch small.
The parser kept the nesting of arrays and objects in a std::vector<bool>,
whose push_back() GCC calls out of line; it is a std::vector<std::uint8_t>
now.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Skip the literal fast path test without exceptions
The test case compares the error messages of both lexer paths, so it does
not compile with JSON_NOEXCEPTION (the caught exception does not exist).
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Allow the string literal parameter of scan_literal() for clang-tidy
The parameter takes the literal by array reference so that its length is
a template parameter; this is the same exemption the C array overloads of
to_json() and from_json() have.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* 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>
After the digits are converted while scanning, copying the token text into
token_buffer (an out-of-line std::string::append) was about 10% of parsing
canada.json. The only reader is the string argument of
json_sax::number_float(), and the library's own handlers
(json_sax_dom_parser, json_sax_dom_callback_parser, json_sax_acceptor)
ignore it.
For contiguous input, the fast path now keeps a pointer to the token in the
input, and get_string() copies it on first use. The parser passes an empty
string to the handlers marked by the new internal trait
sax_ignores_number_text instead of calling get_string(), so json::parse()
and json::accept() never copy float text. User handlers, and
JSON_DIAGNOSTIC_POSITIONS (which reads get_string().size()), still see the
exact text. Keys and strings, which never have pending text, use the
unchecked get_string_token(). Input read through a window into a stream
(std::istream, FILE*) copies the text right away, because the next read may
refill the window.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Accumulate number digits while scanning them
The contiguous number fast path validated a token byte by byte, copied it
into token_buffer, and then converted it in out-of-line functions that
scanned the digits a second time: parse_float_native() called
parse_float_significand(), which returned its result through memory and
called std::any_of (out of line with GCC) twice per number, and
parse_integer_signed() divided by ten on every digit for its overflow check.
scan_number_bulk_contiguous() now accumulates the significand while it
validates the digits, eight at a time where at least eight bytes remain
(is_eight_digits() and parse_eight_digits(), after fast_float). Floats with
at most 19 digits go straight to decimal_to_float(); integers with at most
18 digits are stored without an overflow check (a narrower number type
still rejects values that do not round-trip). Longer tokens take the
existing path. has_nonzero_digit() is a plain loop and
parse_float_significand() is always inlined, which helps that path.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Skip the digit accumulation checks without exceptions
The section compares error messages from both lexer paths and parses
out-of-range numbers, so it does not compile (the caught exception does not
exist) and would abort with JSON_NOEXCEPTION.
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Share the significant-digit loop of the integer part and the fraction
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Fix silent wrong results for custom number types
Squashed onto develop cd6b428ef (CI fixes from #5860). Original commits:
- Fix silent wrong results for custom number types
- Fix CI: build the out_of_range.407 message from string_t and fix a link
- Fix CI: diagnostics message, clang-tidy, and MinGW section limit
- Format the out_of_range.407 number without dump()
- Move the std::deque array test out of unit-regression3.cpp
- Fix clang-tidy and clang 3.4/3.5 failures in the custom number type changes
- Fix CI: cast the 128-bit unsigned value explicitly in the integer fast path
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Name the range checks of integer_fits_max/min
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Speed up writing CBOR and MessagePack, above all integer arrays
Squashed onto develop cd6b428ef (CI fixes from #5860). Original commits:
- Append to the output vector through a position-tracking sink
- Write CBOR integers without branching on their sign or size
- Write MessagePack integers without branching on their sign or size
- Use std::array for the integer size class table
- Fix clang-tidy and GCC warnings in the integer array test
- Regenerate the amalgamation
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
* Convert the streamed bytes explicitly in the binary writer sink tests
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
---------
Signed-off-by: Niels Lohmann <mail@nlohmann.me>
Niels Lohmann