[D3D12] DXBC internal r# stack allocation

src/xenia/gpu/dxbc_shader_translator.cc

@@ -35,11 +35,29 @@ void DxbcShaderTranslator::Reset() {
   shader_code_.clear();
 
   rdef_constants_used_ = 0;
+  system_temp_count_current_ = 0;
+  system_temp_count_max_ = 0;
   writes_depth_ = false;
 
   std::memset(&stat_, 0, sizeof(stat_));
 }
 
+uint32_t DxbcShaderTranslator::PushSystemTemp() {
+  uint32_t register_index = system_temp_count_current_;
+  if (!uses_register_relative_addressing()) {
+    // Guest shader registers first if they're not in x0.
+    register_index += register_count();
+  }
+  ++system_temp_count_current_;
+  system_temp_count_max_ =
+      std::max(system_temp_count_max_, system_temp_count_current_);
+}
+
+void DxbcShaderTranslator::PopSystemTemp(uint32_t count) {
+  assert_true(count <= system_temp_count_current_);
+  system_temp_count_current_ -= std::min(count, system_temp_count_current_);
+}
+
 void DxbcShaderTranslator::CompleteVertexShaderCode() {}
 
 void DxbcShaderTranslator::CompletePixelShaderCode() {}
@@ -869,15 +887,17 @@ void DxbcShaderTranslator::WriteShaderCode() {
     }
   }
 
-  // Temporary registers - system registers, and also Xbox 360 general-purpose
+  // Temporary registers - guest general-purpose registers if not using dynamic
-  // registers if not using dynamic indexing.
+  // indexing and Xenia internal registers.
   stat_.temp_register_count =
-      uint32_t(TempRegister::kCount) +
+      (uses_register_relative_addressing() ? 0 : register_count()) +
-      (uses_register_relative_addressing() ? 0 : register_count());
+      system_temp_count_max_;
-  shader_object_.push_back(
+  if (stat_.temp_register_count != 0) {
-      ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_DCL_TEMPS) |
+    shader_object_.push_back(
-      ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(2));
+        ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_DCL_TEMPS) |
-  shader_object_.push_back(stat_.temp_register_count);
+        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()) {

src/xenia/gpu/dxbc_shader_translator.h

@@ -129,12 +129,10 @@ class DxbcShaderTranslator : public ShaderTranslator {
   static constexpr uint32_t kPSInPositionRegister =
       kPSInPointParametersRegister + 1;
 
-  enum class TempRegister {
+  // Allocates a new r# register for internal use and returns its index.
-    // Color output map in pixel shader epilogue.
+  uint32_t PushSystemTemp();
-    kColorOutputMap,
+  // Frees the last allocated internal r# registers for later reuse.
-
+  void PopSystemTemp(uint32_t count = 1);
-    kCount,
-  };
 
   // Writing the epilogue.
   void CompleteVertexShaderCode();
@@ -268,6 +266,12 @@ class DxbcShaderTranslator : public ShaderTranslator {
   static const RdefConstantBufferIndex
       constant_buffer_dcl_order_[size_t(RdefConstantBufferIndex::kCount)];
 
+  // Number of currently allocated Xenia internal r# registers.
+  uint32_t system_temp_count_current_;
+  // Total maximum number of temporary registers ever used during this
+  // translation (for the declaration).
+  uint32_t system_temp_count_max_;
+
   bool writes_depth_;
 
   // The STAT chunk (based on Wine d3dcompiler_parse_stat).
@@ -294,7 +298,7 @@ class DxbcShaderTranslator : public ShaderTranslator {
     uint32_t texture_comp_instructions;
     uint32_t texture_bias_instructions;
     uint32_t texture_gradient_instructions;
-    // Moving to outputs only.
+    // Not including indexable temp load/store.
     uint32_t mov_instruction_count;
     // Unknown in Wine.
     uint32_t movc_instruction_count;