This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: DDG - Implementing Swing Modulo Scheduling in GCC (cont.)
- From: Ayal Zaks <ZAKS at il dot ibm dot com>
- To: Richard Henderson <rth at redhat dot com>
- Cc: Mostafa Hagog <MUSTAFA at il dot ibm dot com>, gcc at gcc dot gnu dot org, David Edelsohn <dje at watson dot ibm dot com>, Vladimir Makarov <vmakarov at redhat dot com>, canqun at yahoo dot com dot cn
- Date: Thu, 11 Dec 2003 18:48:49 +0200
- Subject: Re: DDG - Implementing Swing Modulo Scheduling in GCC (cont.)
>> Our solution is to build a new and different representation of the data
>> dependence information.
>
>Did you look at using the existing def-use graph (du.c)?
The existing def-use graph (in df.c that is) deals with explicit defs, uses
and register information which we don't need.
We need to know all dependences that exist between insns (flow, anti,
output,
or other scheduling fences), and their latency. The original Swing (and
most
other) Modulo-Scheduling algorithm does not care if the dependence is
register
or memory, flow/anti or output. We model that because we have some ideas
how
to improve on the original algorithm. The basic dependence information in
the
schedulers LOG_LINKS is suitable, after enhancing it with edge structures
and
loop-carried edges.
Ayal.