Handle numbers that do not fit narrow number types in the binary readers (#5607)

Browse Source

* Handle numbers that do not fit narrow number types in the binary readers

With custom number types narrower than the values in a binary document,
for example basic_json<..., std::int32_t, std::uint32_t, float>, every
binary reader (CBOR, MessagePack, UBJSON, BJData, BSON, BON8) passed the
decoded number to the SAX interface with an implicit conversion: the
integer 5000000000 silently became 705032704, and a finite double such as
1e300 became infinity. The lexer handles the same values in JSON text: an
integer that fits neither integer type is stored as number_float_t, and a
finite number that overflows number_float_t is rejected with
out_of_range.406.

Pass every number read from binary input through three helpers that
apply the lexer's rules:
- emit_signed(): number_integer_t, else number_unsigned_t for a
  non-negative value, else number_float_t
- emit_unsigned(): number_unsigned_t, else number_float_t
- emit_float(): out_of_range.406 if a finite value overflows
  number_float_t; infinity and NaN are passed on

For consistency, a CBOR negative integer below the range of
number_integer_t is now stored as number_float_t, like a too small
integer in JSON text, instead of being rejected with parse_error.112.
With the default number types, this is the only change in behavior.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Fix MSVC and clang 3.5 in the narrow number type test

MSVC types 3000000000 and 5000000000 as unsigned long, so
json(-3000000000) triggered C4146 (unary minus on an unsigned type),
which /WX turns into an error. Use LL literals, as elsewhere in the
tests.

clang 3.5 cannot convert the lambdas in the braced initializer of the
format table to function pointers. Use named functions instead.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Check integer-to-float fallbacks for overflow in the binary readers

emit_signed, emit_unsigned, and the CBOR negative integer fallback now
pass their number_float_t fallback through emit_float, so a value that
overflows number_float_t is rejected with out_of_range.406 like a
floating-point value, instead of silently becoming infinity. This only
matters for a number_float_t that cannot represent 2^64, such as a
half-precision type. The CBOR value -1 - n is computed as long double so
that emit_float sees a finite value.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

* Use the input_format member instead of passing the format to binary_reader helpers

The helpers (get_number, get_to, get_string, get_binary, get_bytes,
emit_signed, emit_unsigned, emit_float, unexpect_eof, exception_message)
are members of binary_reader, which already stores the format it was
constructed with, so the parameter was redundant.

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

---------

Signed-off-by: Niels Lohmann <mail@nlohmann.me>

...

This commit is contained in:

Niels Lohmann

2026-10-04 17:53:29 +02:00

committed by GitHub

parent d8d47be4a5

commit 6b4b825af2

9 changed files with 743 additions and 448 deletions

Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL

Download Patch File Download Diff File Expand all files Collapse all files

single_include/nlohmann/json.hpp

+283 -211

File diff suppressed because it is too large Load Diff