ir_to_mesa: Do not emit OPCODE_SEQ or OPCODE_SNE

Nothing that consumes the output of this backend consumes them
navtively.  This is *not* the way i915 has implemented these
instructions, but, as far as I am able to tell, this is the way both the
Cg compiler and the HLSL compiler implement these operations.

Signed-off-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>

diff --git a/src/mesa/program/ir_to_mesa.cpp b/src/mesa/program/ir_to_mesa.cpp
index c5b67ad706a931369b2f42ac0a24e46884fed7ff..d5b3693661c23fd84936d8f4b0e5974b50acf214 100644 (file)
--- a/src/mesa/program/ir_to_mesa.cpp
+++ b/src/mesa/program/ir_to_mesa.cpp
@@ -304,6 +304,22 @@ public:
 
    void emit_swz(ir_expression *ir);
 
+   void emit_equality_comparison(ir_expression *ir, enum prog_opcode op,
+                                 dst_reg dst,
+                                 const src_reg &src0, const src_reg &src1);
+
+   inline void emit_sne(ir_expression *ir, dst_reg dst,
+                        const src_reg &src0, const src_reg &src1)
+   {
+      emit_equality_comparison(ir, OPCODE_SLT, dst, src0, src1);
+   }
+
+   inline void emit_seq(ir_expression *ir, dst_reg dst,
+                        const src_reg &src0, const src_reg &src1)
+   {
+      emit_equality_comparison(ir, OPCODE_SGE, dst, src0, src1);
+   }
+
    bool process_move_condition(ir_rvalue *ir);
 
    void copy_propagate(void);
@@ -912,6 +928,46 @@ ir_to_mesa_visitor::emit_swz(ir_expression *ir)
    this->result = result_src;
 }
 
+void
+ir_to_mesa_visitor::emit_equality_comparison(ir_expression *ir,
+                                             enum prog_opcode op,
+                                             dst_reg dst,
+                                             const src_reg &src0,
+                                             const src_reg &src1)
+{
+   src_reg difference;
+   src_reg abs_difference = get_temp(glsl_type::vec4_type);
+   const src_reg zero = src_reg_for_float(0.0);
+
+   /* x == y is equivalent to -abs(x-y) >= 0.  Since all of the code that
+    * consumes the generated IR is pretty dumb, take special care when one
+    * of the operands is zero.
+    *
+    * Similarly, x != y is equivalent to -abs(x-y) < 0.
+    */
+   if (src0.file == zero.file &&
+       src0.index == zero.index &&
+       src0.swizzle == zero.swizzle) {
+      difference = src1;
+   } else if (src1.file == zero.file &&
+              src1.index == zero.index &&
+              src1.swizzle == zero.swizzle) {
+      difference = src0;
+   } else {
+      difference = get_temp(glsl_type::vec4_type);
+
+      src_reg tmp_src = src0;
+      tmp_src.negate = ~tmp_src.negate;
+
+      emit(ir, OPCODE_ADD, dst_reg(difference), tmp_src, src1);
+   }
+
+   emit(ir, OPCODE_ABS, dst_reg(abs_difference), difference);
+
+   abs_difference.negate = ~abs_difference.negate;
+   emit(ir, op, dst, abs_difference, zero);
+}
+
 void
 ir_to_mesa_visitor::visit(ir_expression *ir)
 {
@@ -1086,17 +1142,17 @@ ir_to_mesa_visitor::visit(ir_expression *ir)
       emit(ir, OPCODE_SGE, result_dst, op[0], op[1]);
       break;
    case ir_binop_equal:
-      emit(ir, OPCODE_SEQ, result_dst, op[0], op[1]);
+      emit_seq(ir, result_dst, op[0], op[1]);
       break;
    case ir_binop_nequal:
-      emit(ir, OPCODE_SNE, result_dst, op[0], op[1]);
+      emit_sne(ir, result_dst, op[0], op[1]);
       break;
    case ir_binop_all_equal:
       /* "==" operator producing a scalar boolean. */
       if (ir->operands[0]->type->is_vector() ||
          ir->operands[1]->type->is_vector()) {
         src_reg temp = get_temp(glsl_type::vec4_type);
-        emit(ir, OPCODE_SNE, dst_reg(temp), op[0], op[1]);
+         emit_sne(ir, dst_reg(temp), op[0], op[1]);
 
         /* After the dot-product, the value will be an integer on the
          * range [0,4].  Zero becomes 1.0, and positive values become zero.
@@ -1111,7 +1167,7 @@ ir_to_mesa_visitor::visit(ir_expression *ir)
         sge_src.negate = ~sge_src.negate;
         emit(ir, OPCODE_SGE, result_dst, sge_src, src_reg_for_float(0.0));
       } else {
-        emit(ir, OPCODE_SEQ, result_dst, op[0], op[1]);
+         emit_seq(ir, result_dst, op[0], op[1]);
       }
       break;
    case ir_binop_any_nequal:
@@ -1124,7 +1180,7 @@ ir_to_mesa_visitor::visit(ir_expression *ir)
              ir->operands[1]->as_constant()->is_zero()) {
             temp = op[0];
          } else {
-            emit(ir, OPCODE_SNE, dst_reg(temp), op[0], op[1]);
+            emit_sne(ir, dst_reg(temp), op[0], op[1]);
          }
 
         /* After the dot-product, the value will be an integer on the
@@ -1147,12 +1203,12 @@ ir_to_mesa_visitor::visit(ir_expression *ir)
            emit(ir, OPCODE_SLT, result_dst, slt_src, src_reg_for_float(0.0));
         }
       } else {
-        emit(ir, OPCODE_SNE, result_dst, op[0], op[1]);
+         emit_sne(ir, result_dst, op[0], op[1]);
       }
       break;
 
    case ir_binop_logic_xor:
-      emit(ir, OPCODE_SNE, result_dst, op[0], op[1]);
+      emit_sne(ir, result_dst, op[0], op[1]);
       break;
 
    case ir_binop_logic_or: {
@@ -1217,8 +1273,7 @@ ir_to_mesa_visitor::visit(ir_expression *ir)
       break;
    case ir_unop_f2b:
    case ir_unop_i2b:
-      emit(ir, OPCODE_SNE, result_dst,
-                         op[0], src_reg_for_float(0.0));
+      emit_sne(ir, result_dst, op[0], src_reg_for_float(0.0));
       break;
    case ir_unop_bitcast_f2i: // Ignore these 4, they can't happen here anyway
    case ir_unop_bitcast_f2u: