[v3] Fix / clean-up config vs crosses (5/n): avoid a TRY_RUN

[Paolo Carlini]

Mark Mitchell wrote:
> We need something different.  The current mechanism (test -r) means that
> a cross to (say) x86 GNU/Linux may yield different results than a native
> build, since in the cross case the answer (whatever it is) will be
> determined in some different way.  That's exactly the situation we need
> to avoid.  To a first approximation, as long as there are IS_NATIVE
> checks, we've not yet solved the cross-compilation problem.
>   
Yes, I understand the usual issue, in general terms. Yesterday just 
wanted to remove a TRY_RUN, the only RUN in the entire configury.
> I would suggest either:
>
> A1. Hard-code the answer based on the target OS (e.g., GNU/Linux has
> /dev/urandom).
> A2. If you like, provide a configure option to override the default.
>   
Ok, thanks. I think that over the last days we covered quite a bit of 
ground about this native vs crosses issue. Let's give the other changes 
a bit of time to settle, then I will work on this improvement. By the 
way, I would appreciate some feedback about
> or:
>
> B. Test at run-time, the first time you want to use the special file.
>   
In this specific case, I think this is unfortunately a no-no, because 
the layout of a class changes depending on that test and it's too late 
to change it. For C++0x I'll make sure to design things differently...

Paolo.