frv: asm reloads for pseudos assigned to memory fail

[Richard Henderson]

On Sat, Jul 17, 2004 at 01:58:03AM -0300, Alexandre Oliva wrote:
> Even if I relaxed the check to be non-strict (since the next round of
> reloading would presumably fix things up, I thought), it failed
> because the pseudo had become a mem(fp+offset), but the insn had
> already been recognized as one that doesn't accept memory inputs, so
> it just fell apart.

Yes.  Reload *relies* on having a single move pattern.  FRV has
split its move patterns into at least two pieces, which is 
pretty much always going to fail.

Your patch merely papers over the problem.  Merge all the move
patterns back into one.


r~