I've committed this patch for amdgcn.

This changes the procedure calling ABI such that vector arguments are 
passed in vector registers, rather than on the stack as before.

The ABI for scalar functions is the same for arguments, but the return 
value has now moved to a vector register; keeping it the same for all 
types simplifies the compiler implementation. If a significant down-side 
is found then we can move to having multiple return locations, and worry 
about how to fix the "untyped" calls then.

There's no "standard ABI" for this target, and there are no third party 
binaries with which to retain compatibility, so we're free to make 
whatever changes we wish.

Andrew