`std::from_chars` is locale-independent and thus avoids the TOCTOU
issue #5198 when it's available.
This commit introduces a `from_chars` utility in the `detail` namespace
which merely wraps `std::from_chars` if it is available and emulates its
API and behavior using the `strto*` family of functions when it is not.
As of this commit, the `strto*`-based fallback is still locale-dependent
and only matches `std::from_chars` when the "C" locale is used.
Availability of `std::from_chars` for floating point numbers is varied.
For instance, as of libc++ 22, `std::from_chars` for `long double` isn't
supported and, consequently, the feature-test macro `__cpp_lib_to_chars`
is not set, while the overloads for `float` and `double` are available.
To avoid complicating matters further, the feature-test macro is taken
as the authoritative indicator for (full) `std::from_chars` support and
the fallback is used otherwise.
If `std::from_chars` is available, our wrapper additionally tweaks its
output in the case of floating-point out-of-range results. The behavior
in this case is poorly (under-)specified and existing implementations do
not agree. The wrapper's tweaks are intended to ensure consistent
behavior across all toolchains in this edge case. In practice, this
should only take effect when parsing out-of-range floats on libstdc++.
The unit tests for the `from_chars` helper are intentionally compiled in
both C++11 and C++17 modes. This ensures that the test expectations
align with the actual behavior we're trying to model. The tests are not
aiming to cover all of floating-point parsing comprehensively, as
`std::from_chars` is specified to behave mostly the same as `strto*`
and instead focuses on the edge cases where the two differ:
- leading whitespace is not tolerated
- `+` sign is not allowed (outside of the exponent)
- out-parameter is left unchanged in case of `invalid_argument` and
integral `result_out_of_range` errors
Signed-off-by: Jonas Greitemann <jgreitemann@gmail.com>
When an integer JSON token exceeds the range of the underlying 64-bit
unsigned/integer types, the lexer instead identifies the token type as
`value_float`. When the value exceeds the range of the underlying
floating-point type, the token type remains as `value_float`.
This was previously not tested explicitly for the edge case values and
subtly breaking this behavior right around the edge case was not caught
by any other test.
Signed-off-by: Jonas Greitemann <jgreitemann@gmail.com>
Jonas Greitemann
Niels Lohmann
