reload bug

[Joern Rennecke]

> A note, if I move your new code out one nesting level, then it appears to
> work.  ie, I put your new code before this statement:
> 
>       /* I is nonneg if this reload used a register.
>          If reload_reg_rtx[r] is 0, this is an optional reload
>          that we opted to ignore.  */
> 
>       if (i >= 0 && reload_reg_rtx[r] != 0)
> [ ... ]
> 
> I have no clue if that's the right thing to do or not.

Yes, it is the right thing to do.

Basically, there are two ways we manage references to pseudos:
If we can make a reload, even if it's optional, the reload itself
carries the information.  emit_reload_insns has to use that information to
update its tables.
If for some reason we can't make a reload at all, a USE or CLOBBER of the
pseudo is created.  CLOBBERs are directly evaluated by emit_reload_insns,
while USEs are evaluated by delete_output_reload.