This is the mail archive of the
binutils@sourceware.cygnus.com
mailing list for the binutils project.
Re: MIPS opcodes patch: change move from andu to or
Andrew Haley writes:
> > OK to commit?
> >
> > For opcodes:
> > * mips-opc.c (move): Use "or" rather than "addu" or "daddu".
> >
> > We deliberately chose ``addu'' and ``daddu'' rather than ``or''
> > because several MIPS chips have two pipelined addition units but only
> > one pipelined logical unit. Those chips can move two pairs of
> > registers around at once if we use ``addu'', but they can only move
> > one pair if we use ``or''.
>
> Aha! That's a *very* good reason. I'll add a comment, if that's OK
> with you.
>
> > Is there any reason we should use ``or''? Your comment mentions
> > something about ``gcc -mips3 -gp32''; can't we fix gcc?
>
> Yeah. The problem is that -mgp32 is supposed to generate 32-bit code,
> but any move insns get converted into "dadd". I don't really think
> that this is important, since the insns will have the same effect, but
> it looks rather odd.
The overall problem is that the -mgp32 flag was never fully implemented
(I don't know why), and no-one's ever had time to get it working. One
of the problems with it is that gas neither recieves nor handles the
-mgp32 flag. It only sees the -mips3 flag (or gets -mips3 it by default),
and produces a "daddu" for "move" (and perhaps other related problems).
I thought that the "move" -> "or" patch would be a low cost fix to part
of the -mgp32 problems, but I didn't think of the pipeline issue.
> I suppose that we could change gcc to ouput "or" insns for 32-bit
> moves, but that would degrade performance in the case that you
> describe.
Getting -mgp32 working in mips gas is on my list of things to do; I'll
see about moving it up.
-gavin...