From adc0eb87f62f966968c4daeb03714d997013ab23 Mon Sep 17 00:00:00 2001
From: Triang3l <triang3l@yandex.ru>
Date: Thu, 20 Dec 2018 10:14:18 +0300
Subject: [PATCH] [GPU] Gather used memexport constants in shader translator

---
 src/xenia/gpu/shader.h             | 10 ++++++++--
 src/xenia/gpu/shader_translator.cc | 17 +++++++++++++++++
 src/xenia/gpu/shader_translator.h  |  7 +++++++
 3 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/src/xenia/gpu/shader.h b/src/xenia/gpu/shader.h
index f0f31ef73..5653bf49e 100644
--- a/src/xenia/gpu/shader.h
+++ b/src/xenia/gpu/shader.h
@@ -31,8 +31,8 @@ enum class InstructionStorageTarget {
   kPosition,
   // Result is stored to the point size export (gl_PointSize).
   kPointSize,
-  // Result is stored as memexport destination address.
-  // [physical >> 2, ??, ??, ??]
+  // Result is stored as memexport destination address
+  // (see xenos::xe_gpu_memexport_stream_t).
   kExportAddress,
   // Result is stored to memexport destination data.
   kExportData,
@@ -583,6 +583,11 @@ class Shader {
     return constant_register_map_;
   }
 
+  // All c# registers used as the addend in MAD operations to eA.
+  const std::vector<uint32_t>& memexport_stream_constants() const {
+    return memexport_stream_constants_;
+  }
+
   // Returns true if the given color target index [0-3].
   bool writes_color_target(int i) const { return writes_color_targets_[i]; }
 
@@ -634,6 +639,7 @@ class Shader {
   std::vector<TextureBinding> texture_bindings_;
   ConstantRegisterMap constant_register_map_ = {0};
   bool writes_color_targets_[4] = {false, false, false, false};
+  std::vector<uint32_t> memexport_stream_constants_;
 
   bool is_valid_ = false;
   bool is_translated_ = false;
diff --git a/src/xenia/gpu/shader_translator.cc b/src/xenia/gpu/shader_translator.cc
index 2f9609b51..4d65a8d37 100644
--- a/src/xenia/gpu/shader_translator.cc
+++ b/src/xenia/gpu/shader_translator.cc
@@ -63,6 +63,7 @@ void ShaderTranslator::Reset() {
     writes_color_targets_[i] = false;
   }
   writes_depth_ = false;
+  memexport_stream_constants_.clear();
 }
 
 bool ShaderTranslator::GatherAllBindingInformation(Shader* shader) {
@@ -174,6 +175,10 @@ bool ShaderTranslator::TranslateInternal(Shader* shader) {
   for (size_t i = 0; i < xe::countof(writes_color_targets_); ++i) {
     shader->writes_color_targets_[i] = writes_color_targets_[i];
   }
+  shader->memexport_stream_constants_.clear();
+  for (uint32_t memexport_stream_constant : memexport_stream_constants_) {
+    shader->memexport_stream_constants_.push_back(memexport_stream_constant);
+  }
 
   shader->is_valid_ = true;
   shader->is_translated_ = true;
@@ -282,6 +287,18 @@ void ShaderTranslator::GatherInstructionInformation(
                   writes_depth_ = true;
                 }
               }
+              // Store used memexport constants because CPU code needs addresses
+              // and sizes. eA is (hopefully) always written to using:
+              // mad eA, r#, const0100, c#
+              // (though there are some exceptions, shaders in Halo 3 for some
+              // reason set eA to zeros, but the swizzle of the constant is not
+              // .xyzw in this case, and they don't write to eM#).
+              if (op.vector_dest() == 32 &&
+                  op.vector_opcode() == AluVectorOpcode::kMad &&
+                  op.vector_write_mask() == 0b1111 && !op.src_is_temp(3) &&
+                  op.src_swizzle(3) == 0) {
+                memexport_stream_constants_.insert(op.src_reg(3));
+              }
             } else {
               if (op.is_vector_dest_relative()) {
                 uses_register_dynamic_addressing_ = true;
diff --git a/src/xenia/gpu/shader_translator.h b/src/xenia/gpu/shader_translator.h
index 238daea7d..b79bf2fac 100644
--- a/src/xenia/gpu/shader_translator.h
+++ b/src/xenia/gpu/shader_translator.h
@@ -11,6 +11,7 @@
 #define XENIA_GPU_SHADER_TRANSLATOR_H_
 
 #include <memory>
+#include <set>
 #include <string>
 #include <vector>
 
@@ -65,6 +66,11 @@ class ShaderTranslator {
   const std::vector<Shader::TextureBinding>& texture_bindings() const {
     return texture_bindings_;
   }
+  // All c# registers used as the addend in MAD operations to eA, populated
+  // before translation occurs.
+  const std::set<uint32_t>& memexport_stream_constants() const {
+    return memexport_stream_constants_;
+  }
 
   // Current line number in the ucode disassembly.
   size_t ucode_disasm_line_number() const { return ucode_disasm_line_number_; }
@@ -230,6 +236,7 @@ class ShaderTranslator {
   bool uses_register_dynamic_addressing_ = false;
   bool writes_color_targets_[4] = {false, false, false, false};
   bool writes_depth_ = false;
+  std::set<uint32_t> memexport_stream_constants_;
 
   static const AluOpcodeInfo alu_vector_opcode_infos_[0x20];
   static const AluOpcodeInfo alu_scalar_opcode_infos_[0x40];