fix opt/12441

[Jan Hubicka]

> The block comment immediately above Jan's change explicitly says not to
> accept SUBREGs, and why.  This PR is exactly that case.  Shame on you.
> 
> The problem is not in reload, per se.  The problem is that with the
> way we treat multi-word data types such as DImode, we really will run
> out of registers.
> 
> This would be safe to pull back to 3.3.3 if someone would like to do
> the requisite sanity check build.

This patch has been originaly made to fix regression on x86-64 where we
produced SUBREGed registers (this is why i didn't want to revert it
completely as mentioned earlier on the thread).
Perhaps we can allow just SUBREGs that does not increase number of words
in the memory addresses?
I will try to dig out the original testcase if it still fails.

Honza