using frame pointer + constant in reg_equiv_constant

[Joern Rennecke]

> I disagree strongly because of how reg_equiv_constant entries can be
> substituted into insns without any checks as to the validity of the
> substitution.

But it is really an extension of register elimination: the constants are of
the form frame_pointer + CONST or arg_pointer + CONST, and this gets
transformed into stack_pointer + CONST or hard_frame_pointer + CONST
just like an eliminated frame or arg pointer.  The usual reloading
is done on the result, so any invalidity introduced by the substitution
gets fixed up.