Teaching GCC about fmadd variant

Teaching GCC about fmadd variant

[Daniel Glastonbury]

Hi,
  I'm looking for some help in teaching gcc how to use the fmadd variants
that my mips has.  Unlike the mips IV madd instruction, this mips has a
variant that goes something like this:

mula  fs, ft    : ACC <- fs*ft
madda fs, ft    : ACC <- ACC + fs*ft
madd  fd, fs, ft: fd <- ACC + fs*ft

The problem is there is no general way to move to/from the ACC register.

I tried to teach gcc about a new register, the fp acc, and then specify insn
patterns to try match on a new register contraint, which I called F.

here are the example insn patterns I'm using:
(define_insn ""
  [(set (match_operand:SF 0 "register_operand" "=f")
  (plus:SF (mult:SF (match_operand:SF 1 "register_operand" "f")
        (match_operand:SF 2 "register_operand" "f"))
     (match_operand:SF 3 "register_operand" "F")))]
     "TARGET_MIPS5900"
     "madd.s\\t%0,%1,%2"
     [(set_attr "type" "fmadd")
     (set_attr "mode" "SF")
     (set_attr "length" "1")])

(define_insn "R5900_maddas"
  [(set (match_operand:SF 0 "register_operand" "=F")
 (plus:SF (mult:SF (match_operand:SF 1 "register_operand" "f")
     (match_operand:SF 2 "register_operand" "f"))
   (match_operand:SF 3 "register_operand" "F")))]
  "TARGET_MIPS5900"
  "madda.s\\t%1,%2"
  [(set_attr "type" "fmadd")
   (set_attr "mode" "SF")
   (set_attr "length" "1")])

(define_insn ""
  [(set (match_operand:SF 0 "register_operand" "=F")
  (mult:SF (match_operand:SF 1 "register_operand" "f")
     (match_operand:SF 2 "register_operand" "f")))]
     "TARGET_MIPS5900"
     "mula.s\\t%1,%2"
     [(set_attr "type" "fmadd")
     (set_attr "mode" "SF")
     (set_attr "length" "1")])

The problem is that I get an ICE because the compiler can't generate a
reload.  I believe it's trying to load the computation from a regular
mult:SF into the ACC register.

Can anyone point me in the right direction.  I've tried to glean as much
information as I could from the gcc internals manual, and from
mips.{c,h,md}, but I can't work out how to tell gcc to use a different
register.

I'm sorry if this is the wrong forum to ask this question.

thanks
Dan
--
Daniel Glastonbury, Senior Programmer, Ratbag Pty Ltd
Level 8, 63 Pirie Street, Adelaide, SA, Australia. 5000
T: +618 8223 5830 F: +618 8223 5746 W: http://www.ratbaggames.com

Re: Teaching GCC about fmadd variant

[Daniel Jacobowitz]

On Mon, Apr 22, 2002 at 05:47:10PM +0930, Daniel Glastonbury wrote:
> Hi,
>   I'm looking for some help in teaching gcc how to use the fmadd variants
> that my mips has.  Unlike the mips IV madd instruction, this mips has a
> variant that goes something like this:
> 
> mula  fs, ft    : ACC <- fs*ft
> madda fs, ft    : ACC <- ACC + fs*ft
> madd  fd, fs, ft: fd <- ACC + fs*ft
> 
> The problem is there is no general way to move to/from the ACC register.

(Over my head, but) do you still get reasonable benefit from this if
you define simple-minded patterns to move from/to the accumulator? 
i.e.:
   load fs into ACC: load 1.0 into ft; mula fs, ft
   load ACC into fd: load 0.0 into ft; madd fd, ft, ft

On the other hand, this won't help if you were actually seeing it use
the ACC unnecessarily.  I believe that would need to be worked around
with costs.  You're going to have problems with combine (discussed on
one GCC list, probably this one, a few weeks ago) getting fmadd to
work well, also.

-- 
Daniel Jacobowitz                           Carnegie Mellon University
MontaVista Software                         Debian GNU/Linux Developer

Re: Teaching GCC about fmadd variant

[Daniel Glastonbury]

> (Over my head, but) do you still get reasonable benefit from this if
> you define simple-minded patterns to move from/to the accumulator?
> i.e.:
>    load fs into ACC: load 1.0 into ft; mula fs, ft
>    load ACC into fd: load 0.0 into ft; madd fd, ft, ft
>
> On the other hand, this won't help if you were actually seeing it use
> the ACC unnecessarily.  I believe that would need to be worked around
> with costs.  You're going to have problems with combine (discussed on
> one GCC list, probably this one, a few weeks ago) getting fmadd to
> work well, also.

Thanks for your reply.  It seems everyone is busy with GCC 3.1.

How I was thinking of achieving this was by looking for patterns and trying
to `rewrite' the registers from $f1 to ACC, for example.  Unfortunately I
don't really know where to do this.  I'll keep nosing about the code and see
if I can understand how it all works.  I'll look up the combine discussion
too.

thanks
Dan