This commit works around an issue in std::initializer_list design.
By using a detail::json_ref proxy with a mutable value inside,
rvalue-ness of an input to list initializer is remembered and
used later to move from the proxy instead of copying.
Travis found an error with Clang 3.8's sanitizers, see https://travis-ci.org/nlohmann/json/jobs/256366699. Unfortunately, I cannot reproduce this error with clang version 6.0.0 (trunk 308825) locally. However, this seems to be an issue, because so far, we did not reset a value after moving from it.
A complete rewrite of the string escape function. It now provides codepoint-to-\uxxxx escaping. Invalid UTF-8 byte sequences are not escaped, but copied as-is. I haven’t spent much time optimizing the code - but the library now agrees with Python on every single Unicode character’s escaping (see file test/data/json_nlohmann_tests/all_unicode_ascii.json).
Other minor changes: replaced "size_t" by "std::size_t"
It makes no sense to have this special exception. Instead of throwing when an input adapter is created, it is better to detect a parse error in later usage when an EOF is "read" unexpectedly.
These breaks were just added to silence a GCC warning - the GCC is right about warning as it cannot know that the expect function will not return in these two scenarios.
It seems these functions are not required any more. The code was added in commit 7e32457 to fix issue #171. There are still regression tests for #171, so when this commit passes the CI, the functions may be removed for good.
When an empty vector was passed to the parse function, an empty iterator range was used to construct an input iterator. Unfortunately, we then cannot use the start iterator to derive a pointer from.
Found with Xcode's undefined behavior sanitizer.
