New -Wall issues

[Paolo Carlini]

Jonathan Wakely wrote:
> Detecting uninitialised data requires the value-propagation pass of
> the optimiser, at -O0 it doesn't happen.
>   
Of course.

I think there are 2 real, and separate, issues here:

1- Whether the standard actually mandates that __r must remain
uninitialized in this case, or, in other terms, whether a random __r
doesn't cause problem for the functioning of the class after
construction, whether __r uninitialized is really intended, in yet other
terms.

2- Once more, it seems that the pragma system_headers doesn't really
suppress all the warnings, this apparently makes for a compiler PR, if
we don't have already one...

Paolo.