2 files changed, 30 insertions, 15 deletions

diff --git a/mesalib/src/mesa/program/ir_to_mesa.cpp b/mesalib/src/mesa/program/ir_to_mesa.cpp
index debadb9a3..1ef609fe1 100644
--- a/mesalib/src/mesa/program/ir_to_mesa.cpp
+++ b/mesalib/src/mesa/program/ir_to_mesa.cpp
@@ -641,8 +641,6 @@ src_reg
 ir_to_mesa_visitor::get_temp(const glsl_type *type)
 {
    src_reg src;
-   int swizzle[4];
-   int i;
 
    src.file = PROGRAM_TEMPORARY;
    src.index = next_temp;
@@ -652,12 +650,7 @@ ir_to_mesa_visitor::get_temp(const glsl_type *type)
    if (type->is_array() || type->is_record()) {
       src.swizzle = SWIZZLE_NOOP;
    } else {
-      for (i = 0; i < type->vector_elements; i++)
-	 swizzle[i] = i;
-      for (; i < 4; i++)
-	 swizzle[i] = type->vector_elements - 1;
-      src.swizzle = MAKE_SWIZZLE4(swizzle[0], swizzle[1],
-				  swizzle[2], swizzle[3]);
+      src.swizzle = swizzle_for_size(type->vector_elements);
    }
    src.negate = 0;
 
@@ -915,10 +908,30 @@ ir_to_mesa_visitor::try_emit_sat(ir_expression *ir)
    sat_src->accept(this);
    src_reg src = this->result;
 
-   this->result = get_temp(ir->type);
-   ir_to_mesa_instruction *inst;
-   inst = emit(ir, OPCODE_MOV, dst_reg(this->result), src);
-   inst->saturate = true;
+   /* If we generated an expression instruction into a temporary in
+    * processing the saturate's operand, apply the saturate to that
+    * instruction.  Otherwise, generate a MOV to do the saturate.
+    *
+    * Note that we have to be careful to only do this optimization if
+    * the instruction in question was what generated src->result.  For
+    * example, ir_dereference_array might generate a MUL instruction
+    * to create the reladdr, and return us a src reg using that
+    * reladdr.  That MUL result is not the value we're trying to
+    * saturate.
+    */
+   ir_expression *sat_src_expr = sat_src->as_expression();
+   ir_to_mesa_instruction *new_inst;
+   new_inst = (ir_to_mesa_instruction *)this->instructions.get_tail();
+   if (sat_src_expr && (sat_src_expr->operation == ir_binop_mul ||
+			sat_src_expr->operation == ir_binop_add ||
+			sat_src_expr->operation == ir_binop_dot)) {
+      new_inst->saturate = true;
+   } else {
+      this->result = get_temp(ir->type);
+      ir_to_mesa_instruction *inst;
+      inst = emit(ir, OPCODE_MOV, dst_reg(this->result), src);
+      inst->saturate = true;
+   }
 
    return true;
 }
diff --git a/mesalib/src/mesa/program/prog_optimize.c b/mesalib/src/mesa/program/prog_optimize.c
index f4a7a638d..3340ce049 100644
--- a/mesalib/src/mesa/program/prog_optimize.c
+++ b/mesalib/src/mesa/program/prog_optimize.c
@@ -472,8 +472,7 @@ can_downward_mov_be_modifed(const struct prog_instruction *mov)
       mov->SrcReg[0].HasIndex2 == 0 &&
       mov->SrcReg[0].RelAddr2 == 0 &&
       mov->DstReg.RelAddr == 0 &&
-      mov->DstReg.CondMask == COND_TR &&
-      mov->SaturateMode == SATURATE_OFF;
+      mov->DstReg.CondMask == COND_TR;
 }
 
 
@@ -482,7 +481,8 @@ can_upward_mov_be_modifed(const struct prog_instruction *mov)
 {
    return
       can_downward_mov_be_modifed(mov) &&
-      mov->DstReg.File == PROGRAM_TEMPORARY;
+      mov->DstReg.File == PROGRAM_TEMPORARY &&
+      mov->SaturateMode == SATURATE_OFF;
 }
 
 
@@ -657,6 +657,8 @@ _mesa_merge_mov_into_inst(struct prog_instruction *inst,
    if (mask != (inst->DstReg.WriteMask & mask))
       return GL_FALSE;
 
+   inst->SaturateMode |= mov->SaturateMode;
+
    /* Depending on the instruction, we may need to recompute the swizzles.
     * Also, some other instructions (like TEX) are not linear. We will only
     * consider completely active sources and destinations