Re: [RFC] Implementing Swing Modulo Scheduling in GCC

[Sebastian Pop <sebastian dot pop at cri dot ensmp dot fr>]

On Tue, Mar 30, 2004 at 05:14:08PM +0200, Ayal Zaks wrote:
> (from Feb. 3rd ...)
> >> [snip]
> >> 3. Pass dependence info from tree-ssa (monev analyzer?), and maintain
> >> this info until reaching the modulo-scheduler (primary concern: loop
> >> unroller), following Dan's suggestion:
> >
> >If you want you can include your work in the lno-branch.  Then, I will
> >try to fill in the data dependence structures that you need for SMS.
> 
> We'd like to better understand this option:
> we have an innermost single-basic block loop at the RTL level (after
> loop-unrolling/peeling), and we need to figure out if there is an
> inter-loop (flow, anti or output) dependence between a pair of stores
> (or store and load) inside the loop. We are only interested in knowing if
> such a dependence exists of a bounded (say 10) distance, and if so what
> that distance is.

Yes, the classic per loop dependence distances are computed in the LNO
branch.

> Can LNO's tree-based analysis possibly fill in our structures at a post-
> unrolling/peeling RTL stage?
> 

Yes, this will be a little technical, but not infeasible.  I can
retrieve the files comming with dependence.c that send the information
to the back-end, or maybe Dan Berlin already knows the tricks used in
these old patches...