[PATCH] Fix target/15286 (reload of paradoxical subregs)

[David Edelsohn]

> It wouldn't bother me if, for 4.1, someone did the work to
> disallow paradoxical subregs.  It ought to be just as easy for
> us to generate
> 
> 	(set (reg:DI temp) (foo:DI ...))
> 	(set (reg:SI dest) (subreg:SI (reg:DI temp) 0))
> 
> instead of
> 
> 	(set (subreg:DI (reg:SI dest) 0) (foo:DI ...))

	Would the common part of the compiler generate the alternate
sequence or the target-specific backend?

	For the pattern that caused the problem, the machine description
was generating reg:DI, not subreg.

David