Re: [boost] [optional] get() misses optional r-value overload in contrast to operator* and value()

Subject: Re: [boost] [optional] get() misses optional r-value overload in contrast to operator* and value()
From: Peter Dimov (lists_at_[hidden])
Date: 2018-03-23 22:20:23

Viktor Sehr wrote:

> Clang (6.0) emits "ud2" instructions for both auto&& and const auto&. If 
> the const auto& can bind directly to an r-value, this must mean Clang is 
> wrong, or am I missing something?

"ud2" is Clang's way of encoding undefined behavior. This is an invalid 
instruction that crashes. Both auto&& and auto const& do the same thing - 
bind to the rvalue reference returned by value() &&.

> GCC on the other hand emits code for both functions, although with slight 
> difference (I'm not skilled enough in assembler to analyze the GCC output 
> in further detail)
>
> Code example: https://godbolt.org/g/7fjszE <https://godbolt.org/g/uNJoZB>

GCC crashes too, in another creative way. It returns *p, where the pointer p 
is [rsp+8], which is NULL. If you change the functions to

int * func1() {
    const auto& x = func0().value();
    return &*x;
}

you'll see that Clang returns NULL and GCC returns the value in [rsp+8] 
which is also NULL.

So both compilers deliberately dereference a NULL pointer because they see 
that the reference `x` is no longer valid.