Re: PDP-10 backend for gcc

[lars brinkhoff <lars at nocrew dot org>]

Alan Lehotsky <lehotsky@tiac.net> writes:
> At 8:05 AM +0200 8/25/00, lars brinkhoff wrote:
> > > On Thu, Aug 24, 2000 at 07:50:09AM +0200, lars brinkhoff wrote:
> > > > I may be writing a PDP-10 backend for gcc during the next year.  It's
> > > > a word-adressed 36-bit machine with a 30-bit virtual address space.
> > Does it seem like it would be possible to write a back end using these
> > pointer formats?
> 
> I've worked on a compiler that supported TWO different sized pointers
> (24 and 32 bits) and there are a lot of problems. [I also worked on
> a compiler for the DEC-10 (Bliss36), so I know a lot about your
> target machine. [If you need someone to bounce ideas off, feel free
> to mail me....]

Thanks!

> The biggest problem you will encounter is the fact that the way that
> RTL is currently generated, you don't really get the opportunity to
> provide any type information about a pointer.  There's basically
> a Pmode that's supposed to be correct for ALL pointers.

So, a hack to make gcc support two pointer formats (global one-word
byte pointers and ordinary global pointers) could be a quick-and-
dirty solution?

> Certainly, if you wanted to sacrifice performance, you could make
> ALL pointers be byte-pointer format (with 36 bit fields) and
> pretty much everything would work.

The problem with this is that one-word global byte pointers can't
point to 36-bit bytes.  (Or so I'm told.)

> Alternatively, you may need to propose an extension to the RTL and 
> the code generator generator so that you can maintain the necessary 
> type information.
> 
> IMHO, such an extension would be a major contribution to the GCC
> community, because these problems keep cropping up in DSP ports
> and the current solutions are less than satisfactory.

I'm quite willing to try this, but I'm not sure that my employer want
to pay for it.