SH failures - analyzis

[Jan Hubicka]

Hi,
I am looking at the SH port (and trying to get it running properly so I can experiment more with the .align code)
and it appears to have serious problem in reload.

The case is that register is subregged to lowpart (subreg:HI (reg:SI x) 2).  This is correct for little endian
machines (SH does both).

Reload then puts register X to stack and generate memory for it - it is valid for SH, as it allows
constant indexes in memory that are divisible by 4.  During subreg elimination, the compilation crashes,
as memory if offsetted by 2 and thats no longer valid address.

What appears to be the proper solution?

Honza