[Mesa-dev] [PATCH 10/16] i965/fs: Eliminate null-dst instructions without side-effects.

[Kenneth Graunke]

On Tuesday, January 20, 2015 10:13:09 AM Matt Turner wrote:
> On Tue, Jan 20, 2015 at 12:53 AM, Kenneth Graunke <kenneth at whitecape.org> wrote:
> > On Monday, January 19, 2015 03:31:09 PM Matt Turner wrote:
> >> ---
> >>  src/mesa/drivers/dri/i965/brw_fs_dead_code_eliminate.cpp | 11 +++++++++++
> >>  1 file changed, 11 insertions(+)
> >>
> >> diff --git a/src/mesa/drivers/dri/i965/brw_fs_dead_code_eliminate.cpp b/src/mesa/drivers/dri/i965/brw_fs_dead_code_eliminate.cpp
> >> index 81be4de..d66808b 100644
> >> --- a/src/mesa/drivers/dri/i965/brw_fs_dead_code_eliminate.cpp
> >> +++ b/src/mesa/drivers/dri/i965/brw_fs_dead_code_eliminate.cpp
> >> @@ -85,6 +85,17 @@ fs_visitor::dead_code_eliminate()
> >>              }
> >>           }
> >>
> >> +         if ((inst->opcode != BRW_OPCODE_IF &&
> >> +              inst->opcode != BRW_OPCODE_WHILE) &&
> >> +             inst->dst.is_null() &&
> >> +             !inst->has_side_effects() &&
> >> +             !inst->writes_flag() &&
> >> +             !inst->writes_accumulator) {
> >> +            inst->opcode = BRW_OPCODE_NOP;
> >> +            progress = true;
> >> +            continue;
> >> +         }
> >> +
> >>           if (inst->dst.file == GRF) {
> >>              if (!inst->is_partial_write()) {
> >>                 int var = live_intervals->var_from_reg(inst->dst);
> >>
> >
> > Seems like these should be handled too...
> >
> > - BRW_OPCODE_ELSE
> > - FS_OPCODE_DISCARD_JUMP
> > - FS_OPCODE_PLACEHOLDER_HALT
> > - SHADER_OPCODE_SHADER_TIME_ADD
> 
> These have BAD_FILE destinations.

Oh, my mistake - I confused BAD_FILE and ARF NULL for a moment.
You're right, of course.

> > - SHADER_OPCODE_GEN4_SCRATCH_READ
> > - SHADER_OPCODE_GEN4_SCRATCH_WRITE
> > - SHADER_OPCODE_GEN7_SCRATCH_READ
> 
> The READs have non-null destinations (they have to return the data somewhere).

Right.  I was thinking about scratch, and just pasted all the opcodes...oops.

> And we only emit SCRATCH_* from spilling registers as part of register
> allocation. We can't ever call dead code elimination after we've
> assigned registers. (Not only do we not, but it couldn't work)

That makes sense.  However, in the vec4 backend, we do emit scratch reads
and writes prior to optimization - in move_grf_array_access_to_scratch().
I've been meaning to port that over to the scalar backend.  At which
point, SHADER_OPCODE_GEN4_SCRATCH_WRITE would be a problem.

But it's not yet, so,
Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20150120/7b37e6a3/attachment.sig>