Re: sign/zero extension patterns

[Richard Earnshaw <Richard dot Earnshaw at arm dot com>]

On Wed, 2004-04-28 at 08:22, Nathan Sidwell wrote:
> Richard Earnshaw wrote:
> 
> > 
> > Try making the constraint for operand 1 "memory_operand".
> Yup, that worked. Thanks!
> 
> I didn't realise that one needn't always have a register_operand pattern.

The general rule is that everything you accept in a define_expand you
must either match in a define_insn or translate into a series of insns
that will match.  In this case an alternative would have been to
translate non-memory operands into independent shift insns.  But that's
probably somewhat redundant in this case because the compiler knows how
to do that if your machine descriptions can't.

R.