This change was made in 359fe90 due to recommendations from tooling.
On its face this change makes sense since it removes a copy that
isn't always necessary. In practice it caused ergonomic issues due
to now being forced to make a copy of the render states when needed.
The performance gains of eliding this copy are unsubstantiated. We
have not done any profiling to measure its impact. For lack of such
measurements I'd rather err on the side of improved user experience.
If future benchmarks prove this copy is rather expensive then we
can reconsider removing it with that evidence in mind.
While I don't suspect there are any uninitialize variable bugs
present, it's still good to err on the side of safety and provide
an initial value nonetheless.
While you can't easily run these unit tests, it's easy enough to
still compile them so we don't need to go out of way to prohibit
this. CMake has some support for running tests on the target OS
so perhaps in the future we find a compelling way to run the test
suite on a mobile OS.
Character traits are only standardized for character types of which
std::uint8_t is not. All major C++ implementations happen to define
this specialization but because it is not standard C++ they are
allowed to remove it as LLVM has done by deprecating this specialization
in LLVM 18. It is slated for removal in LLVM 19. To avoid compilation
errors and to get ahead of any deprecation warnings when LLVM 18 ships
we need to define our own std::uint8_t character traits.
SFML 4 will have access to C++20's std::u8string which should let us
remove this code.
sf::RenderWindow still inherits a virtual destructor from a base
class so there's no need to explicitly declare a virtual destructor.
I added a test to ensure this property was not broken.
