slice valarrays with std::string -> segfault

[Paolo Carlini]

Gabriel Dos Reis wrote:
> I just saw this message.  As said in previous mails, your analysis is
> correct that the use of _copy instead of _copy_construct was
> thoughtless -- other constructors carefully use _copy_construct.
>   
Ok, thanks.
> [...]
>
> | +    __valarray_construct(_Array<_Tp> __a, size_t __n, size_t __s, _Array<_Tp> __b)
>
> Can we do without this additional layer?
> Otherwise, the patch is correct with that change.
>   
Really, I have to spend some serious time on the issue. Right now I'm 
missing a lot of important details. Maybe you can guide me a bit?

For example, I see the other conversion constructors (both 
standard-mandated and not, i.e., from gslice_array, mask_array, 
indirect_array, _Expr<>) also calling __valarray_copy... I'm afraid 
something can go wrong there too, but I don't know sufficiently well 
<valarray> to tell for sure, at the moment...

As regards avoiding the additional layer, in principle seems certainly 
feasible, I was quickly mimicking what currently happens when 
__valarray_copy is used (i.e., those intermediate templates around 2/3 
of valarray_array.h).

Paolo.