On Fri, Oct 10, 2008 at 8:54 PM, Tobias Grosser <[EMAIL PROTECTED]> wrote: > On Fri, 2008-10-10 at 20:35 +0200, Manuel López-Ibáñez wrote: >> 2008/10/10 Tobias Grosser <[EMAIL PROTECTED]>: >> > >> > Now: >> > ---- >> > >> > -fgraphite: Do nothing. >> > -floop-block, -floop-interchange, -floop-stripmine: Try these >> > transformations. >> >> This only applies to the branch, does it? There is no do-nothing >> command-line option in trunk, is there? > > No, that's the same for trunk. But it seems my description was not that > clear. In fact "-fgraphite" does a lot. > > If you enable "-fgraphite" the complete CFG graph is scanned and we > generate the GRAPHITE representation for every SCoP. This is great to > make compile tests for the transformation from GIMPLE to GRAPHITE. So > e.g. the polyhedron test suite can be transformed to GRAPHITE, without > any failure. And most of the SPEC test suite also. > > But we only generate GIMPLE out of GRAPHITE, if we applied a > transformation. This was thought to reduce compile time, as we do not > regenerate code for SCoPs, that never changed. > > If you want to test a graphite transformation, use "-floop-block" (The > one implemented and used for testing). > > What I would like to discuss (and later implement) is how we could allow > to specify a identity transformation, that enables code generation for > all SCoPs. This transformation would probably require higher compile > time and generate slower code, but can be used to test the GRAPHITE code > generation better. > > But as you see in my description, I would also like to change the > meaning of -fgraphite. > > Solution one would make it to generate the identity transformation. > Solution two would make it an automatic optimizer (yet unimplemented), > that could at the moment default to all available graphite > transformations.
For 4.4 I would like us to not expose -fgraphite to the user, as it will only generate bugreports and never be profitable for them. If you think it makes sense to have the -fgraphite debugging only then you can add a new option for debugging as you suggested - but I think -fgraphite can as well always generate code to be a more useful debugging option? So, can we for 4.4 remove the -fgraphite option (or at least its user-level documentation) as well as -floop-stripmine? Thanks, Richard.