Doug Evans wrote:
> I _think_ the problem centers around the implicit widening
> in the original code.  The subtraction is done in QImode
> and a QImode value is subsequently added to a DImode value.

Well, it's more due to a missed explicit widening.

> [Digression:
> In a world in which Umodes don't exist, implicit widening operations
> should probably be disallowed, though part of me will loathe the extra
> amount of typing required.]

It would be unlogical to have to write zero-extensions different
from sign-extensions.  So, since the mode doesn't have signedness
information, implicit widening would be missing that information.

> I don't know the architecture, so I don't know what the desired effect
> of (abs DI (some-QI-value)) is.  Making a wild guess, what about
> something like:
> 
> -                (set acc (abs DI (sub (subword QI rm 0) (subword QI rn 0))))
> +                (set acc (zext DI (abs QI (sub (subword QI rm 0) (subword QI rn 0)))))
> 
> Or maybe it's
> -                (set acc (abs DI (sub (subword QI rm 0) (subword QI rn 0))))
> +                (set acc (zext DI (sub (subword QI rm 0) (subword QI rn 0))))
> 
> [and so on for the rest of the bytes]

No, the zero-extension should happen before the subtraction.
		
-- 
--------------------------
SuperH (UK) Ltd.
2410 Aztec West / Almondsbury / BRISTOL / BS32 4QX
T:+44 1454 465658