Re: SPARC compiler optimisation

vanroy@prl.dec.com
Mon, 9 Mar 1992 13:45:07 GMT

          From comp.compilers

Related articles
[4 earlier articles]
Re: SPARC compiler optimisation grunwald@foobar.cs.colorado.edu (1992-02-22)
Re: SPARC compiler optimisation andrew@highland.oz.au (1992-02-26)
Re: SPARC compiler optimisation dmk@craycos.com (1992-02-27)
Re: SPARC compiler optimisation nickh@CS.CMU.EDU (1992-02-28)
Re: SPARC compiler optimisation nickh@CS.CMU.EDU (1992-03-02)
Re: SPARC compiler optimisation preston@dawn.cs.rice.edu (1992-03-02)
Re: SPARC compiler optimisation vanroy@prl.dec.com (1992-03-09)
| List of all articles for this month |
Newsgroups: comp.compilers
From: vanroy@prl.dec.com
Keywords: optimize, sparc
Organization: Compilers Central
References: 92-03-011
Date: Mon, 9 Mar 1992 13:45:07 GMT

In article 92-03-011 Preston Briggs writes:
> I think of aligned LDD and STD as kind of a mistake for RISC machines,
> since they can't be generated by compilers. I expect they're usually
> justified for their usefulness in hand-coded library routines, especially
> block copies and such, where alignment can be tested.


I don't agree with this. With proper design of both compiler and
architecture, these instructions can be quite useful. Here's an example.
At Berkeley, we ported the Aquarius Prolog compiler to both the VLSI-BAM
chip (a RISC-like processor with extensions for Prolog) and the SPARC.


Both of these processors have LDD and STD instructions, but they are
slightly different. It's this difference that makes the difference.


On the VLSI-BAM the LDD and STD instructions can load or store _any_ two
registers. The only condition for a load or store from
register+displacement R+D is that R+D be aligned on a double-word
boundary. This condition is easily satisfied by our compiler; we measure
the performance improvement for the double-word memory port to be about
17% (see article in ISCA 1990; this figure is a lump sum for LDD, STD,
STDC, PUSHD, and PUSHDC).


There are two additional conditions on the SPARC: the source or
destination registers must be _consecutive_ and the first register number
must be _even_. These conditions are too strong for our compiler; only a
small number of LDD & STD instructions are generated. (However, we did
not show that a smarter compiler could not accommodate these additional
conditions.)


I conclude that properly designed double-word loads and stores can indeed
be fruitfully used by a RISC machine, as shown by the VLSI-BAM. However,
I do not know whether this is true for the SPARC as well.


BTW, the VLSI-BAM was built together with a cache board, and the
combination achieves its rated speed.


Regards,
Peter Van Roy
--
Peter Van Roy
Digital Equipment Corporation Net: vanroy@prl.dec.com
Paris Research Laboratory Tel: [33] (1) 47 14 28 65
85, avenue Victor Hugo Fax: [33] (1) 47 14 28 99
92563 RUEIL MALMAISON CEDEX
FRANCE
--


Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.