[D3D12] DXBC: Don't use indexable temps unless needed

src/xenia/gpu/dxbc_shader_translator.cc

@@ -42,81 +42,7 @@ void DxbcShaderTranslator::Reset() {
 
 void DxbcShaderTranslator::CompleteVertexShaderCode() {}
 
-void DxbcShaderTranslator::CompletePixelShaderCode() {
+void DxbcShaderTranslator::CompletePixelShaderCode() {}
-  // Remap color outputs from Xbox 360 to Direct3D 12 indices.
-  // temp = xe_color_output_map;
-  // [unroll] for (uint i = 0; i < 4; ++i) {
-  //   o[i] = x1[temp[i]];
-  // }
-  shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOV) |
-                         ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
-  shader_code_.push_back(
-      ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
-      ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
-          D3D10_SB_OPERAND_4_COMPONENT_MASK_MODE) |
-      D3D10_SB_OPERAND_4_COMPONENT_MASK_ALL |
-      ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_TEMP) |
-      ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_1D) |
-      ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
-          0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32));
-  shader_code_.push_back(uint32_t(TempRegister::kColorOutputMap));
-  shader_code_.push_back(
-      ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
-      ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
-          D3D10_SB_OPERAND_4_COMPONENT_SWIZZLE_MODE) |
-      D3D10_SB_OPERAND_4_COMPONENT_NOSWIZZLE |
-      ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_CONSTANT_BUFFER) |
-      ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_3D) |
-      ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
-          0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32) |
-      ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
-          1, D3D10_SB_OPERAND_INDEX_IMMEDIATE32) |
-      ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
-          2, D3D10_SB_OPERAND_INDEX_IMMEDIATE32));
-  shader_code_.push_back(uint32_t(RdefConstantBufferIndex::kSystemConstants));
-  shader_code_.push_back(uint32_t(CbufferRegister::kSystemConstants));
-  shader_code_.push_back(kSysConst_ColorOutputMap_Vec);
-  ++stat_.instruction_count;
-  for (uint32_t i = 0; i < 4; ++i) {
-    shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOV) |
-                           ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(7));
-    shader_code_.push_back(
-        ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
-        ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
-            D3D10_SB_OPERAND_4_COMPONENT_MASK_MODE) |
-        D3D10_SB_OPERAND_4_COMPONENT_MASK_ALL |
-        ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_OUTPUT) |
-        ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_1D) |
-        ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
-            0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32));
-    shader_code_.push_back(i);
-    shader_code_.push_back(
-        ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
-        ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
-            D3D10_SB_OPERAND_4_COMPONENT_SWIZZLE_MODE) |
-        D3D10_SB_OPERAND_4_COMPONENT_NOSWIZZLE |
-        ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_INDEXABLE_TEMP) |
-        ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_2D) |
-        ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
-            0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32) |
-        ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
-            1, D3D10_SB_OPERAND_INDEX_RELATIVE));
-    shader_code_.push_back(1);
-    shader_code_.push_back(
-        ENCODE_D3D10_SB_OPERAND_NUM_COMPONENTS(D3D10_SB_OPERAND_4_COMPONENT) |
-        ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECTION_MODE(
-            D3D10_SB_OPERAND_4_COMPONENT_SELECT_1_MODE) |
-        ENCODE_D3D10_SB_OPERAND_4_COMPONENT_SELECT_1(i) |
-        ENCODE_D3D10_SB_OPERAND_TYPE(D3D10_SB_OPERAND_TYPE_TEMP) |
-        ENCODE_D3D10_SB_OPERAND_INDEX_DIMENSION(D3D10_SB_OPERAND_INDEX_1D) |
-        ENCODE_D3D10_SB_OPERAND_INDEX_REPRESENTATION(
-            0, D3D10_SB_OPERAND_INDEX_IMMEDIATE32));
-    shader_code_.push_back(uint32_t(TempRegister::kColorOutputMap));
-    ++stat_.instruction_count;
-    ++stat_.array_instruction_count;
-    ++stat_.mov_instruction_count;
-  }
-}
 
 void DxbcShaderTranslator::CompleteShaderCode() {
   // Write stage-specific epilogue.
@@ -943,7 +869,18 @@ void DxbcShaderTranslator::WriteShaderCode() {
     }
   }
 
-  // General-purpose registers (x0).
+  // Temporary registers - system registers, and also Xbox 360 general-purpose
+  // registers if not using dynamic indexing.
+  stat_.temp_register_count =
+      uint32_t(TempRegister::kCount) +
+      (uses_register_relative_addressing() ? 0 : register_count());
+  shader_object_.push_back(
+      ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_DCL_TEMPS) |
+      ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(2));
+  shader_object_.push_back(stat_.temp_register_count);
+
+  // General-purpose registers if using dynamic indexing (x0).
+  if (uses_register_relative_addressing()) {
     shader_object_.push_back(
         ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_DCL_INDEXABLE_TEMP) |
         ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(4));
@@ -953,16 +890,6 @@ void DxbcShaderTranslator::WriteShaderCode() {
     // 4 components in each.
     shader_object_.push_back(4);
     stat_.temp_array_count += register_count();
-
-  // Color outputs on the Xbox 360 side (x1), for remapping to D3D12 bindings.
-  if (is_pixel_shader()) {
-    shader_object_.push_back(
-        ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_DCL_INDEXABLE_TEMP) |
-        ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(4));
-    shader_object_.push_back(1);
-    shader_object_.push_back(4);
-    shader_object_.push_back(4);
-    stat_.temp_array_count += 4;
   }
 
   // Initialize the depth output if used, which must be initialized on every

src/xenia/gpu/dxbc_shader_translator.h

@@ -130,6 +130,7 @@ class DxbcShaderTranslator : public ShaderTranslator {
       kPSInPointParametersRegister + 1;
 
   enum class TempRegister {
+    // Color output map in pixel shader epilogue.
     kColorOutputMap,
 
     kCount,

src/xenia/gpu/shader_translator.h

@@ -45,6 +45,11 @@ class ShaderTranslator {
   bool is_vertex_shader() const { return shader_type_ == ShaderType::kVertex; }
   // True if the current shader is a pixel shader.
   bool is_pixel_shader() const { return shader_type_ == ShaderType::kPixel; }
+  // True if the current shader addresses general-purpose registers with dynamic
+  // indices.
+  bool uses_register_relative_addressing() const {
+    return uses_register_relative_addressing_;
+  }
   // A list of all vertex bindings, populated before translation occurs.
   const std::vector<Shader::VertexBinding>& vertex_bindings() const {
     return vertex_bindings_;