This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: [new-regalloc] What is the status on current sources
- To: Geert Bosch <bosch at gnat dot com>
- Subject: Re: [new-regalloc] What is the status on current sources
- From: Michael Matz <matzmich at cs dot tu-berlin dot de>
- Date: Sat, 10 Feb 2001 19:50:14 +0100 (MET)
- cc: Daniel Berlin <dberlin at redhat dot com>, "gcc at gcc dot gnu dot org" <gcc at gcc dot gnu dot org>
Hi,
On Sat, 10 Feb 2001, Geert Bosch wrote:
>
> I couldn't agree more that it is too early to worry about absolute
> performance, but I do think we need to be clear that the various algorithms
> can be implemented without regressing to algorithms that inherently
> require O(n**2) time.
Of course.
> An essential part of the implementation of the dataflow algorithms is
> to use DFS post-order traversal.
If iterative analysis is used, yes.
> Since it seems that flow.c is the place where we do our dataflow analysis,
> I wonder why you decided to implement a new df.c unit.
I didn't decide anything ;) Daniel and me were in need for def/use chains
(I only reg-def reg-use lists), flow.c didn't provide that, the existing
df.c did. You can guess the implication ;)
> It seems to me that it would be preferable to enhance the existing
> flow.c instead of writing a different piece of code that mostly does
> the same.
It does in some parts mostly the same as some parts of flow.c ;) In
flow.c are many things: BB detection/manipulation, edge manipulation, insn
insertion, dead code removal, live register analysis, dealing with
conditional execution, converting to autoinc/dec, loop tree analysis and
more. df.c only deals with global data flow. And at least I am a fan of
moderately sized files ;) df.c also has a nice interface.
Ciao,
Michael.