i965/vec4: check swizzle before discarding a uniform on a 3src operand
authorAlejandro Piñeiro <apinheiro@igalia.com>
Wed, 23 Sep 2015 17:22:17 +0000 (19:22 +0200)
committerAlejandro Piñeiro <apinheiro@igalia.com>
Thu, 24 Sep 2015 19:12:53 +0000 (21:12 +0200)
Without this commit, copy propagation is discarded if it involves
a uniform with an instruction that has 3 sources. But 3 sourced
instructions can access scalar values.

For example, this is what vec4_visitor::fix_3src_operand() is already
doing:

   if (src.file == UNIFORM && brw_is_single_value_swizzle(src.swizzle))
      return src;

Shader-db results (unfiltered) on NIR:
total instructions in shared programs: 6259650 -> 6241985 (-0.28%)
instructions in affected programs:     812755 -> 795090 (-2.17%)
helped:                                7930
HURT:                                  0

Shader-db results (unfiltered) on IR:
total instructions in shared programs: 6445822 -> 6441788 (-0.06%)
instructions in affected programs:     296630 -> 292596 (-1.36%)
helped:                                2533
HURT:                                  0

v2:
- Updated commit message, using Matt Turner suggestions
- Move the check after we've created the final value, as Jason
  Ekstrand suggested
- Clean up the condition

v3:
- Move the check back to the original place, to keep things
  tidy, as suggested by Jason Ekstrand

v4:
- Fixed missing is_single_value_swizzle() as pointed by Jason Ekstrand

Reviewed-by: Matt Turner <mattst88@gmail.com>
src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp

index d3f0ddde258f861fec240588df3b706c00bcce6b..5b6444e3210ee8777ca378a8a8184c39523536d1 100644 (file)
@@ -325,7 +325,11 @@ try_copy_propagate(const struct brw_device_info *devinfo,
        inst->opcode == SHADER_OPCODE_GEN4_SCRATCH_WRITE)
       return false;
 
-   if (inst->is_3src() && value.file == UNIFORM)
+   unsigned composed_swizzle = brw_compose_swizzle(inst->src[arg].swizzle,
+                                                   value.swizzle);
+   if (inst->is_3src() &&
+       value.file == UNIFORM &&
+       !brw_is_single_value_swizzle(composed_swizzle))
       return false;
 
    if (inst->is_send_from_grf())
@@ -380,8 +384,7 @@ try_copy_propagate(const struct brw_device_info *devinfo,
    if (inst->src[arg].negate)
       value.negate = !value.negate;
 
-   value.swizzle = brw_compose_swizzle(inst->src[arg].swizzle,
-                                       value.swizzle);
+   value.swizzle = composed_swizzle;
    if (has_source_modifiers &&
        value.type != inst->src[arg].type) {
       assert(can_change_source_types(inst));