On Wednesday, September 7th, 2022 at 8:50 AM, Edward Catmur <ecatmur_at_[hidden]> wrote:


> > > [...] the point of
> > > this proposal was probably to allow
> > > variant to return a pointer to its
> > > own buffer without breaking the
> > > type system.
> >

> >

> > My suggested use of `std::visit` doesn't "break the type system" as far as I can tell.
>

>

> I conjecture that what ZY meant by this was that the implementation can bypass the overhead of switching on the index, forming a pointer to the derived object, and casting that to the base type, instead just forming a pointer to the base type directly from the buffer (assuming of course that the base type is indeed the first base and so has the same address in each alternative), while not resulting in UB from an invalid pointer cast.
>

> But a modern implementation can eliminate the overhead anyway by looking through the call to std::visit; https://godbolt.org/z/a6o46TTbE compiles to an unimpeachable:
>

> f(std::variant<std::monostate, D1, D2>&):
> xor eax, eax
> cmp BYTE PTR [rdi+1], 0
> cmovne rax, rdi
> ret


Yes, nice codegen. However, discussions
around design vs. QoI always come in two
folds:

// on top of the std::variant abstraction
a) it exposes a sound property (in this case,
    the equivalency between a sum type and
    a closed set of subtypes), and
    it is a useful (in this case, adapt code that
    relies on subtype polymorphism. You can
    call v.common_base()->virtual_function()
    as a transit).

or

// write it yourself
b) non-portable codegen is a QoI issue;
    file a bug to your vendor (in this case,
    clang & libc++ and msvc
    https://godbolt.org/z/3a1Gc8P6n), and
    quality of debug codegen is still QoI;
    at worst, add some pragma.


I recommend writing a paper if you are
motivated.

--
Zhihao Yuan, ID lichray
The best way to predict the future is to invent it.
_______________________________________________