[Mesa-dev] [PATCH 2/4] i965/vs: Add support for copy propagation of the UNIFORM and ATTR files.

[Eric Anholt]

Removes 2.0% of the instructions from 35.7% of vertex shaders in shader-db.
---
 src/mesa/drivers/dri/i965/brw_vec4.cpp             |   13 +++++
 src/mesa/drivers/dri/i965/brw_vec4.h               |    2 +
 .../drivers/dri/i965/brw_vec4_copy_propagation.cpp |   58 +++++++++++++++++++-
 3 files changed, 72 insertions(+), 1 deletions(-)

diff --git a/src/mesa/drivers/dri/i965/brw_vec4.cpp b/src/mesa/drivers/dri/i965/brw_vec4.cpp
index 492e579..5d3d672 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4.cpp
@@ -32,6 +32,19 @@ extern "C" {
 namespace brw {
 
 bool
+vec4_instruction::is_math()
+{
+   return (opcode == SHADER_OPCODE_RCP ||
+	   opcode == SHADER_OPCODE_RSQ ||
+	   opcode == SHADER_OPCODE_SQRT ||
+	   opcode == SHADER_OPCODE_EXP2 ||
+	   opcode == SHADER_OPCODE_LOG2 ||
+	   opcode == SHADER_OPCODE_SIN ||
+	   opcode == SHADER_OPCODE_COS ||
+	   opcode == SHADER_OPCODE_POW);
+}
+
+bool
 src_reg::equals(src_reg *r)
 {
    return (file == r->file &&
diff --git a/src/mesa/drivers/dri/i965/brw_vec4.h b/src/mesa/drivers/dri/i965/brw_vec4.h
index 865f7d1..04ac81f9 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4.h
+++ b/src/mesa/drivers/dri/i965/brw_vec4.h
@@ -275,6 +275,8 @@ public:
     */
    ir_instruction *ir;
    const char *annotation;
+
+   bool is_math();
 };
 
 class vec4_visitor : public ir_visitor
diff --git a/src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp b/src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp
index 1e24e2e..4b33df1 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_copy_propagation.cpp
@@ -153,6 +153,61 @@ try_constant_propagation(vec4_instruction *inst, int arg, src_reg *values[4])
    return false;
 }
 
+static bool
+try_copy_propagation(struct intel_context *intel,
+		     vec4_instruction *inst, int arg, src_reg *values[4])
+{
+   /* For constant propagation, we only handle the same constant
+    * across all 4 channels.  Some day, we should handle the 8-bit
+    * float vector format, which would let us constant propagate
+    * vectors better.
+    */
+   src_reg value = *values[0];
+   for (int i = 1; i < 4; i++) {
+      /* This is equals() except we don't care about the swizzle. */
+      if (value.file != values[i]->file ||
+	  value.reg != values[i]->reg ||
+	  value.reg_offset != values[i]->reg_offset ||
+	  value.type != values[i]->type ||
+	  value.negate != values[i]->negate ||
+	  value.abs != values[i]->abs) {
+	 return false;
+      }
+   }
+
+   /* Compute the swizzle of the original register by swizzling the
+    * component loaded from each value according to the swizzle of
+    * operand we're going to change.
+    */
+   int s[4];
+   for (int i = 0; i < 4; i++) {
+      s[i] = BRW_GET_SWZ(values[i]->swizzle,
+			 BRW_GET_SWZ(inst->src[arg].swizzle, i));
+   }
+   value.swizzle = BRW_SWIZZLE4(s[0], s[1], s[2], s[3]);
+
+   if (value.file != UNIFORM &&
+       value.file != ATTR)
+      return false;
+
+   if (inst->src[arg].abs) {
+      value.negate = false;
+      value.abs = true;
+   }
+   if (inst->src[arg].negate)
+      value.negate = true;
+
+   /* FINISHME: We can't copy-propagate things that aren't normal
+    * vec8s into gen6 math instructions, because of the weird src
+    * handling for those instructions.  Just ignore them for now.
+    */
+   if (intel->gen >= 6 && inst->is_math())
+      return false;
+
+   inst->src[arg] = value;
+   return true;
+}
+
 bool
 vec4_visitor::opt_copy_propagation()
 {
@@ -216,7 +271,8 @@ vec4_visitor::opt_copy_propagation()
 	 if (c != 4)
 	    continue;
 
-	 if (try_constant_propagation(inst, i, values))
+	 if (try_constant_propagation(inst, i, values) ||
+	     try_copy_propagation(intel, inst, i, values))
 	    progress = true;
       }
 
-- 
1.7.5.4