On Wed, Oct 9, 2013 at 3:35 AM, Eric Anholt <e...@anholt.net> wrote: > Chia-I Wu <olva...@gmail.com> writes: > >> From: Chia-I Wu <o...@lunarg.com> >> >> FS_OPCODE_OVERWRITE_DST is used to indicate that the destination register is >> (completely) overwritten. No code is emitted, but the liveness analysis can >> use it as a hint to add the destination register to DEF bitset. This is >> needed because it is hard to figure out if some partial writes combined >> constitute a complete write during liveness analysis, while it is easier for >> the FS visitor to know if that is the case. > > I'm not a fan of this one (particularly the adding of scheduling > barriers). Any other ways you looked at to deal with this? The goal is for liveness analysis to know that a register is not a live-in. It should be possible to make the liveness analysis smarter, knowing that these partial writes or SENDs in the basic block make a full write. But that's quite some code to figure that out, while we already know whether it is the case or not when visiting an ir_texture.
An alternative would be to add a flag, force_def, to fs_inst. The liveness analysis can use that to help decide whether the dst should be in the DEF or not. > > Patch 1, 2, 4 are: > > Reviewed-by: Eric Anholt <e...@anholt.net> > > and I wouldn't mind seeing them land early. Thanks. As this series could not be applied cleanly since texture gather support, I will commit patch 1, 2, 4 first and send v2 for the rest. -- o...@lunarg.com _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
