paradoxical subreg problem [ patch attached ]

[Richard Earnshaw]

>   3. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is defined. 

And doesn't return NIL for the mode of op0.  The ARM port can do this for 
HImode, since there are variants of the ARM (older ones) that don't extend 
half-word memory fetches.

>  In that
>   case we know those bits are zeros or ones; however we do not know if
>   they are the same as the upper bits of op1.  It seems to me we should
>   not make the simplification in this case.

R.