json_view: accept() checks the text without writing an index

json_document::accept() ran the builder that writes the node index and
discarded the index. For input without comments and trailing commas it now
runs a second instantiation of the builder, with a template parameter Store
that is false: the grammar, number, string, UTF-8 and nesting checks are the
same code, but no node is written, the container nodes are not patched when
they close, the node array is not reserved, and no list of large objects is
kept. What only builds the arguments of a node (integer conversion, the digit
counts of floats) is not run either. accept() with comments or trailing commas
and parse() are not changed; the builder that stores is the one parse() always
used. A flag of the document passes the request from accept() to build().

accept() on twitter, citm_catalog, canada and the two minified files, A/B in
one binary, 6 code layouts, time after / before (geometric mean of the five;
the other 14 files of the set are faster too, none is slower):

- Apple M1 Max, Apple clang 21: 0.89
- x86-64 Haswell, GCC 16.2: 0.75
- x86-64 Haswell, Clang 23: 0.78

(semanticscholar-corpus.json: 0.52 on both x86 compilers.) parse() runs the same
instructions as before; its timings moved by up to 3 % on a few small files with
unchanged instruction counts (code placement).

Tests: accept() against json::accept() for the same inputs with all four
option combinations and with and without a terminating NUL: 2,000 generated
documents and damaged copies of them, nesting from 63 to 1000 levels (closed and
not), objects with 300 members, a null byte in the text, and float overflow for
a float document. Checked against a deliberately broken validator.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>
This commit is contained in:
Niels Lohmann committed 2026-10-30 22:18:14 +01:00
1 parent 755bb16d7a
commit dea2a4434a
6 files changed
+179 -27

No files matched your search

@@ -13,9 +13,11 @@ function never throws an exception for invalid input, and the returned `#!cpp bo
is returned.
The result is the same as that of `BasicJsonType::accept()` by design; the difference is how the check is made. This
function runs the document's index builder (and discards the index), while `BasicJsonType::accept()` runs the SAX
parser with a handler that ignores every event. Use this one if you go on to parse the same input with
[`parse()`](parse.md) anyway, so that both use the same code; either is fine for a plain yes/no answer.
function runs the document's builder, which checks the text as [`parse()`](parse.md) does but writes no index (for
input without comments and trailing commas; with either option it builds the index and discards it), while
`BasicJsonType::accept()` runs the SAX parser with a handler that ignores every event. Use this one if you go on to
parse the same input with [`parse()`](parse.md) anyway, so that both use the same code; either is fine for a plain
yes/no answer.
## Template parameters