PA question

[Richard Kenner]

    We don't consider it valid, but instead have special patterns to
    detect the cases where we can figure out what the base register is.

Ah, I see.  Like movqi+2.  I'm seeing an unrecognized insn with such a
pattern coming from a large block of Ada code.  I suspect that what's
happening is that combine is making it due to the setting of
REGNO_POINTER_FLAG and then somebody is changing it to use a different pseudo
without that flag.  I'll have to look into it more closely.