-Wunused-variable triggers for temporaries with side effects when bound to a reference

[cor3ntin]

The following code warns on guard2 and guard3 (but not guard) despite all 3 variable having side effects on destruction.
This feels inconsistent and surprising.
GCC emits no warning in any of these cases.

struct RAIIWrapper {
    RAIIWrapper();
    ~RAIIWrapper();
};

void foo() {
    auto const  guard = RAIIWrapper();
    auto const& guard2 = RAIIWrapper(); //unused variable 'guard2'
    auto && guard3 = RAIIWrapper(); //unused variable 'guard3'
}

https://godbolt.org/z/zr7dsevoc

[dwblaikie]

My personal hope/take on this is that the warning should be aggressive like this (because it lets us warn on things like unused std::string), but we should have an attribute for types with scoped side effects (and then code like RAIIWrapper(), func(); is valid - since func() happens between the wrapper's ctor and dtor, but RAIIWrapper(); is still "unused" - but there might be a third type with arbitrary (not even scoped) side effects, where type_name(); would be valid/not-warned)

[dwblaikie]

Oh, yeah, I guess this is just about reference lifetime extended variables - this doesn't lead us to being able to warn on std::string currently.

Yeah, we shouldn't warn on the reference extended cases anymore than we should on the direct variable cases.