This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Fw: Scheduling an early complete loop unrolling pass?
- From: Dorit Nuzman <DORIT at il dot ibm dot com>
- To: Ayal Zaks <ZAKS at il dot ibm dot com>
- Cc: bergner at vnet dot ibm dot com, gcc at gcc dot gnu dot org, Ira Rosen <IRAR at il dot ibm dot com>, Paolo Bonzini <paolo dot bonzini at lu dot unisi dot ch>, Zdenek Dvorak <rakdver at atrey dot karlin dot mff dot cuni dot cz>, Richard Guenther <rguenther at suse dot de>
- Date: Fri, 9 Feb 2007 20:09:51 +0200
- Subject: Re: Fw: Scheduling an early complete loop unrolling pass?
Ayal Zaks/Haifa/IBM wrote on 07/02/2007 23:17:54:
> ...
>
> >Ah, right... I wonder if we can keep the loop structure in place, even
> >after completely unrolling the loop - I mean the 'struct loop' in
> >'current_loops' (not the actual CFG), so that the "SLP in loops" would
have
> >a chance to at least consider vectorizing this "loop".
> Having a "loop" structure for a piece of CFG that is not a loop, was
> used in some other compiler we worked with - the notion of 'region'
> was such that it corresponded to loops, and in addition the entire
> function belonged to a "universal" region (Peter - please correct if
> I'm wrong). But I think you were looking for some marking of a basic
> block saying "this used to be a loop but got completely unrolled". I
> wonder how much such a "dummy" loop structure can really help the
> vectorizer, except for (convenience of) keeping intact the driver
> that traverses all such structures or the hanging of additional data
> off of them.
>
yes, the idea was just to save time in looking for such occurrences, and
leverage our loop-based-slp. Indeed, there may be user-unrolled code and
other occurrences for which this trick won't help, and as stated earlier,
we probably shouldn't care too much about these cases anyhow. We can play
with this idea at some later point when loop-based-slp is in place.
thanks,
dorit
> Ayal.