[Patch] Fix libstdc++/9563

[Paolo Carlini]

Jerry Quinn wrote:

>Seems like this should reduce the impact of the extra required checking.
>
I agree. However, I have double checked a few other widespread
implementations: no-one seems to implement the optimization and
its additional complexity. Is it really worty?

But... I have got an idea... Do you believe would be ok rewriting
the code simply like this:

	: _M_os(__os)
	{
	  // XXX MT
	  if (__os.tie() && __os.good())
              ^^^^^^^^^^^^^^^^^^^^^^^^^
            __os.tie()->flush();

	  if (__os.good())
            _M_ok = true;
	  else
	    {
	      _M_ok = false;
	      __os.setstate(ios_base::failbit);
	    }
        }

First blush, there is nothing wrong with checking that __os.tie() is not
a null pointer even if we are still not sure that __os.good() is true!

And in this way if __os.tie() is a null pointer we call __os.good() just
one time as in your solution.

What do you think?
Paolo.