From 5dcff19efce8cd114972b221199d591d0c4d53ce Mon Sep 17 00:00:00 2001
From: Triang3l <triang3l@yandex.ru>
Date: Sun, 2 Sep 2018 16:50:00 +0300
Subject: [PATCH] [D3D12] DXBC internal registers

---
 src/xenia/gpu/dxbc_shader_translator.cc | 30 ++++++++++++++++++++++++-
 src/xenia/gpu/dxbc_shader_translator.h  | 11 ++++++++-
 2 files changed, 39 insertions(+), 2 deletions(-)

diff --git a/src/xenia/gpu/dxbc_shader_translator.cc b/src/xenia/gpu/dxbc_shader_translator.cc
index 1c0f5e69d..efb4a0aac 100644
--- a/src/xenia/gpu/dxbc_shader_translator.cc
+++ b/src/xenia/gpu/dxbc_shader_translator.cc
@@ -65,7 +65,7 @@ void DxbcShaderTranslator::Reset() {
   std::memset(&stat_, 0, sizeof(stat_));
 }
 
-uint32_t DxbcShaderTranslator::PushSystemTemp() {
+uint32_t DxbcShaderTranslator::PushSystemTemp(bool zero) {
   uint32_t register_index = system_temp_count_current_;
   if (!uses_register_dynamic_addressing()) {
     // Guest shader registers first if they're not in x0.
@@ -74,6 +74,23 @@ uint32_t DxbcShaderTranslator::PushSystemTemp() {
   ++system_temp_count_current_;
   system_temp_count_max_ =
       std::max(system_temp_count_max_, system_temp_count_current_);
+
+  if (zero) {
+    shader_code_.push_back(ENCODE_D3D10_SB_OPCODE_TYPE(D3D10_SB_OPCODE_MOV) |
+                           ENCODE_D3D10_SB_TOKENIZED_INSTRUCTION_LENGTH(8));
+    shader_code_.push_back(
+        EncodeVectorMaskedOperand(D3D10_SB_OPERAND_TYPE_TEMP, 0b1111, 1));
+    shader_code_.push_back(register_index);
+    shader_code_.push_back(EncodeVectorSwizzledOperand(
+        D3D10_SB_OPERAND_TYPE_IMMEDIATE32, kSwizzleXYZW, 0));
+    shader_code_.push_back(0);
+    shader_code_.push_back(0);
+    shader_code_.push_back(0);
+    shader_code_.push_back(0);
+    ++stat_.instruction_count;
+    ++stat_.mov_instruction_count;
+  }
+
   return register_index;
 }
 
@@ -371,6 +388,11 @@ void DxbcShaderTranslator::StartTranslation() {
   } else if (is_pixel_shader()) {
     StartPixelShader();
   }
+
+  // Request global system temporary variables.
+  system_temp_ps_pc_p0_a0_ = PushSystemTemp(true);
+  system_temp_aL_ = PushSystemTemp(true);
+  system_temp_loop_count_ = PushSystemTemp(true);
 }
 
 void DxbcShaderTranslator::CompleteVertexShader() {}
@@ -378,6 +400,12 @@ void DxbcShaderTranslator::CompleteVertexShader() {}
 void DxbcShaderTranslator::CompletePixelShader() {}
 
 void DxbcShaderTranslator::CompleteShaderCode() {
+  // Release the following system temporary values so epilogue can reuse them:
+  // - system_temp_ps_pc_p0_a0_.
+  // - system_temp_aL_.
+  // - system_temp_loop_count_.
+  PopSystemTemp(3);
+
   // Write stage-specific epilogue.
   if (is_vertex_shader()) {
     CompleteVertexShader();
diff --git a/src/xenia/gpu/dxbc_shader_translator.h b/src/xenia/gpu/dxbc_shader_translator.h
index 2d42b5649..bcfc70d9a 100644
--- a/src/xenia/gpu/dxbc_shader_translator.h
+++ b/src/xenia/gpu/dxbc_shader_translator.h
@@ -180,7 +180,7 @@ class DxbcShaderTranslator : public ShaderTranslator {
   }
 
   // Allocates a new r# register for internal use and returns its index.
-  uint32_t PushSystemTemp();
+  uint32_t PushSystemTemp(bool zero = false);
   // Frees the last allocated internal r# registers for later reuse.
   void PopSystemTemp(uint32_t count = 1);
 
@@ -327,6 +327,15 @@ class DxbcShaderTranslator : public ShaderTranslator {
   // translation (for the declaration).
   uint32_t system_temp_count_max_;
 
+  // Temporary register ID for previous scalar result, program counter,
+  // predicate and absolute address register.
+  uint32_t system_temp_ps_pc_p0_a0_;
+  // Loop index stack - .x is the active loop, shifted right to .yzw on push.
+  uint32_t system_temp_aL_;
+  // Loop counter stack, .x is the active loop. Represents number of times
+  // remaining to loop.
+  uint32_t system_temp_loop_count_;
+
   bool writes_depth_;
 
   // The STAT chunk (based on Wine d3dcompiler_parse_stat).