boost::core::launder bug in BOOST_MSVC

[alandefreitas]

boostorg/optional@8375df7 included a workaround for a problem related to boost::core::launder. The problem was reported at boostorg/optional#122.

The problem happens due to an MSVC bug and it would be best to bypass MSVC directly in Boost.Core than to have this logic replicated in other libraries.

[pdimov]

To clarify, the problem is an internal compiler error with toolset=msvc-14.1 cxxstd=17 variant=release address-model=32, caused by the use of boost::core::launder (which calls std::launder in this case).

[Lastique]

Perhaps, it could be emulated with __asm { nop ptr }; or something?

[pdimov]

I don't think this is what it does.

[Lastique]

I think, something like this should have the effect of std::launder:

template< typename T >
T* launder(T* ptr)
{
    T* res;
    __asm
    {
        mov eax, ptr
        mov res, eax
    };
    return res;
}

But that's two instructions. I suggested a nop because it may be enough for the compiler to stop assuming anything about ptr, and that's just one instruction. Sure, ideally, there should be zero instructions, though.

[pdimov]

MSVC doesn't perform the optimizations against which std::launder guards, so the correct emulation of it today is a no-op.

If it starts doing them some day, and for some reason std::launder doesn't work on that hypothetical future version only in 32 bit mode, maybe the above will work.