On Oct 28, 2015 9:12 PM, "Kenneth Graunke" <<a href="mailto:kenneth@whitecape.org">kenneth@whitecape.org</a>> wrote: > > On Wednesday, October 28, 2015 02:58:07 PM Kristian Høgsberg wrote: > > On Wed, Oct 28, 2015 at 2:34 PM, Jason Ekstrand <<a href="mailto:jason@jlekstrand.net">jason@jlekstrand.net</a>> wrote: > > > On Wed, Oct 28, 2015 at 2:32 PM, Jason Ekstrand <<a href="mailto:jason@jlekstrand.net">jason@jlekstrand.net</a>> wrote: > > >> This series adds a nir_pass datastructure and some helpers for managing > > >> optimization and lowering passes. I've been meaning to get around to this > > >> for some time. There are a couple of primary benifits to this: > > >> > > >> First, this gives us a central place to put things such as validating the > > >> shader, printing it if it changes, etc. Right now, the i965 backend calls > > >> nir_validate_shader after each pass. We would also like to add something > > >> like we have in the i965 backend where it can be set to dump the IR to a > > >> file after every pass that changess it. > > >> > > >> Mor importantly, though, it moves metadata out of the passes them selves > > >> and into the runner. In the process of putting this series together, I > > >> found at least 3 or 4 optimization passes that don't properly invalidate > > >> metadata. By putting a metadata_preserved field in nir_pass and handling > > >> metadata in the pass runner, we make it much less likely that a pass will > > >> get this wrong. LLVM has a similar optimization pass architecture for > > >> precicely this reason. > > >> > > >> As a nice little side-benifit, we no longer have to iterate over all of the > > >> overloads with non-NULL impl pointers in each pass. > > > > > > Once again, git-send-email failed to send the last patch for whatever > > > reason. The entire series can be found here: > > > > > > <a href="http://cgit.freedesktop.org/~jekstrand/mesa/log/?h=wip/nir-pass">http://cgit.freedesktop.org/~jekstrand/mesa/log/?h=wip/nir-pass</a> > > > > Nice. Series, > > > > Reviewed-by: Kristian Høgsberg <<a href="mailto:krh@bitplanet.net">krh@bitplanet.net</a>> > > I plan to review this as well, so please hold off on pushing it for a > little while. Thanks! By all means, go ahead. It'd also be nice, while you're at it to weigh in on how to handle passing arguments to passes. There are a number of ideas thrown back-and-forth between Connor and myself on how to do it (if you can sort through through the bike-shedding). --Jason