Re: Is there any generator generator, not parser generator?

  David Chase <chase@world.std.com> writes:
> There are other ways to approach this; on a
> machine with many registers (i.e., not x86) you could pessimistically
> pre-allocate registers; during pattern- matching you'd discover that
> you weren't using those registers after all, but you might do well
> enough anyway. You could perform some amount of "semantic"
> scheduling/pipelining, in which you simply assumed that each abstract
> operation had a latency associated with it, and work from that;


Why do these phases in the wrong order when you can do them in the
right order? Just do instruction selection before instruction
schedling and register allocation, and the problems mentioned above go
away.


There's still a phase ordering problem between scheduling and register
allocation, but if there are enough registers, doing register
allocation last will work well (there are lots of more complex methods
to deal with this).


In Section 4 of [ertl&pirker97] you can find a description of a basic
block code generator that uses tree parsing for instruction selection,
and performs instruction scheduling and local register allocation.


@InProceedings{ertl&pirker97,
    author = {M. Anton Ertl and Christian Pirker},
    title = {The Structure of a {Forth} Native Code Compiler},
    booktitle = {EuroForth '97 Conference Proceedings},
    year = {1997},
    address = {Oxford},
    pages = {107--116},
    url = {http://www.complang.tuwien.ac.at/papers/ertl&pirker97.ps.gz},
    abstract = {Writing a sophisticated Forth native code compiler
                                    poses some tasks that are not discussed in compiler
                                    text books. Some of these tasks arise from specific
                                    language features, others from the requirement for
                                    very fast compilation to maintain interactivity. In
                                    this paper we describe some of the more interesting
                                    data structures and algorithms used in the RAFTS
                                    prototype.}
}


- anton
--
M. Anton Ertl Some things have to be seen to be believed
anton@mips.complang.tuwien.ac.at Most things have to be believed to be seen
http://www.complang.tuwien.ac.at/anton/home.html

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