On Wed, May 20, 2015 at 12:02 PM, Matt Turner <matts...@gmail.com> wrote: > On Wed, May 20, 2015 at 11:09 AM, Jason Ekstrand <ja...@jlekstrand.net> wrote: >> Now that we're using NIR by default and everything seems stable there, we >> have about 3000 lines of extra code lying around that we don't need. We >> might as well delete it. :-) >> >> This series together with one of the patches I sent yesterday can be found >> in the review/fs-nir-only branch on my freedesktop: >> >> http://cgit.freedesktop.org/~jekstrand/mesa/log/?h=review/fs-nir-only > > Thanks Jason. > > text data bss dec hex filename > 5040325 192800 26512 5259637 504175 before/i965_dri.so > 4932875 193112 26448 5152435 4e9eb3 after/i965_dri.so > > Looks like this cuts about 2% of the .text size. Unexpectedly, it > increases the .data size? > > I compiled this branch with gcc and clang and didn't see any > additional warnings. > > The areas I'm aware of that NIR produces more instructions than the > visitor code are: > > - the texture in then/else cases of an if > - Missing try_emit_line() for gen <= 5 > - Emits an extra CMP for discards (I've got a WIP patch) > - Missing emit_gen6_if(). Probably want to do this differently (after > optimizations) anyway > > I'm okay with not blocking on any of those. > > Do you know of any others?
Not off-hand. When we throw in GCM, we start having issues with CMP instructions being too far away from their use and not cmod propagation not working. However, that's not really a problem at the moment. --Jason _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
