[clang] Revert "[DebugInfo] Ignore undefined constexpr constructors in constructor homing." (PR #221566)

[Avi Kivity via cfe-commits]

avikivity wrote:

> > The original case where I got hit with this was some abseil map, the type without debugging information was std::pair.
> > libstdc++ has these constructors for pair:
> > ```c++
> >       constexpr pair(const pair&) = default;    ///< Copy constructor
> > 
> >       constexpr pair(pair&&) = default;         ///< Move constructor
> > 
> > 
> > 
> >       template<typename... _Args1, typename... _Args2>
> > 
> >         _GLIBCXX20_CONSTEXPR
> > 
> >         pair(piecewise_construct_t, tuple<_Args1...>, tuple<_Args2...>);
> > 
> > 
> > 
> >       template<typename... _Args1, size_t... _Indexes1,
> > 
> >                typename... _Args2, size_t... _Indexes2>
> > 
> >         _GLIBCXX20_CONSTEXPR
> > 
> >         pair(tuple<_Args1...>&, tuple<_Args2...>&,
> > 
> >              _Index_tuple<_Indexes1...>, _Index_tuple<_Indexes2...>);
> > 
> > 
> > 
> >       constexpr
> > 
> >       explicit(__not_<__and_<__is_implicitly_default_constructible<_T1>,
> > 
> >                              __is_implicitly_default_constructible<_T2>>>())
> > 
> >       pair()
> > 
> >       noexcept(is_nothrow_default_constructible_v<_T1>
> > 
> >                 && is_nothrow_default_constructible_v<_T2>)
> > 
> >       requires is_default_constructible_v<_T1>
> > 
> >                && is_default_constructible_v<_T2>
> > 
> >       : first(), second()
> > 
> >       { }
> > ```
> > 
> > 
> >     
> >       
> >     
> > 
> >       
> >     
> > 
> >     
> >   
> > I see that abseil uses the piecewise_construct constructor, which at the point of the template definition, is declared but not defined. It is defined later.
> > Could it be that at the point the code checks whether the constexpr constructor is defined or not, the compiler hasn't instantiated the constructor definition?
> 
> Could you write a minimal repro that compiles but lacks debug info? I'm happy to try to fix-forward, and if we can't we can roll back the original change.

See https://github.com/llvm/llvm-project/pull/221566#discussion_r3944740708. Apparently there's a loophole that allows using a struct that was never constructed.

https://github.com/llvm/llvm-project/pull/221566