Added instruction counting

src/binary_translation/InstructionBlock.cpp

@@ -28,6 +28,8 @@ void InstructionBlock::GenerateCode()
     auto ir_builder = Module()->IrBuilder();
     ir_builder->SetInsertPoint(entry_basic_block);
 
+    module->GenerateIncInstructionCount();
+
     instruction->GenerateCode(this);
 }
 

src/binary_translation/ModuleGen.cpp

@@ -46,6 +46,15 @@ void ModuleGen::Run()
     GenerateBlockAddressArray();
 }
 
+void ModuleGen::GenerateIncInstructionCount()
+{
+    auto load = ir_builder->CreateLoad(instruction_count);
+    auto inc = ir_builder->CreateAdd(load, ir_builder->getInt32(1));
+    auto store = ir_builder->CreateStore(inc, instruction_count);
+    tbaa->TagInstructionCount(load);
+    tbaa->TagInstructionCount(store);
+}
+
 void ModuleGen::BranchReadPC()
 {
     if (verify)
@@ -111,6 +120,9 @@ void ModuleGen::GenerateGlobals()
 
     // bool Verify - contains the value of verify for citra usage
     new GlobalVariable(*module, ir_builder->getInt1Ty(), true, GlobalValue::ExternalLinkage, ir_builder->getInt1(verify), "Verify");
+
+    instruction_count = new GlobalVariable(*Module(), ir_builder->getInt32Ty(), false, GlobalValue::ExternalLinkage,
+        ir_builder->getInt32(0), "InstructionCount");
 }
 
 void ModuleGen::GenerateBlockAddressArray()

src/binary_translation/ModuleGen.h

@@ -27,6 +27,7 @@ public:
 
     void Run();
 
+    void GenerateIncInstructionCount();
     // Generate code to read pc and run all following instructions, used in cases of indirect branch
     void BranchReadPC();
     // Generate code to write to pc and run all following instructions, used in cases of direct branch
@@ -80,6 +81,11 @@ private:
      */
     llvm::ArrayType *block_address_array_type;
     llvm::GlobalVariable *block_address_array;
+    /*
+     * i32 InstructionCount;
+     *  The count of instructions executed
+     */
+    llvm::GlobalVariable *instruction_count;
     /*
      * i8 *GetBlockAddress(u32 pc)
      *  Returns the address of the block for the instruction at pc

src/binary_translation/TBAA.cpp

@@ -20,6 +20,7 @@ void TBAA::GenerateTags()
         register_nodes[i] = md_builder.createTBAAScalarTypeNode(ss.str(), tbaa_root);
     }
     const_node = md_builder.createTBAAScalarTypeNode("Readonly", tbaa_root);
+    instruction_count_node = md_builder.createTBAAScalarTypeNode("InstructionCount", tbaa_root);
 }
 
 void TBAA::TagRegister(Instruction* instruction, Register reg)
@@ -30,4 +31,9 @@ void TBAA::TagRegister(Instruction* instruction, Register reg)
 void TBAA::TagConst(Instruction* instruction)
 {
     instruction->setMetadata(LLVMContext::MD_tbaa, const_node);
+}
+
+void TBAA::TagInstructionCount(llvm::Instruction* instruction)
+{
+    instruction->setMetadata(LLVMContext::MD_tbaa, instruction_count_node);
 }

src/binary_translation/TBAA.h

@@ -21,9 +21,11 @@ public:
 
     void TagRegister(llvm::Instruction *instruction, Register reg);
     void TagConst(llvm::Instruction *instruction);
+    void TagInstructionCount(llvm::Instruction *instruction);
 private:
     llvm::MDNode *register_nodes[RegisterCount];
     // Tag for everything that is never written.
     // Since it is never written, one tag works
     llvm::MDNode *const_node;
+    llvm::MDNode *instruction_count_node;
 };

src/core/arm/dyncom/arm_dyncom_interpreter.cpp

@@ -3985,7 +3985,7 @@ unsigned InterpreterMainLoop(ARMul_State* cpu) {
 
 
 #if ENABLE_BINARY_TRANSLATION
-        BinaryTranslationLoader::Run();
+        num_instrs = BinaryTranslationLoader::Run(num_instrs);
 #endif
 
         if (cpu->TFlag)

src/core/binary_translation/BinaryTranslationLoader.cpp

@@ -19,6 +19,7 @@ std::unique_ptr<RuntimeDyld::LoadedObjectInfo> g_loaded_object_info;
 
 void(*g_run_function)();
 bool(*g_can_run_function)();
+uint32_t *g_instruction_count;
 
 // Used by the verifier
 struct SavedState
@@ -117,7 +118,16 @@ void BinaryTranslationLoader::Load(FileUtil::IOFile& file)
 
     g_run_function = static_cast<decltype(g_run_function)>(g_dyld->getSymbolAddress("Run"));
     g_can_run_function = static_cast<decltype(g_can_run_function)>(g_dyld->getSymbolAddress("CanRun"));
-    g_verify = *static_cast<bool*>(g_dyld->getSymbolAddress("Verify"));
+    auto verify_ptr = static_cast<bool*>(g_dyld->getSymbolAddress("Verify"));
+    g_instruction_count = static_cast<uint32_t *>(g_dyld->getSymbolAddress("InstructionCount"));
+
+    if (!g_run_function || !g_can_run_function || !verify_ptr || !g_instruction_count)
+    {
+        LOG_WARNING(Loader, "Cannot load optimized file, missing critical function");
+        return;
+    }
+
+    g_verify = *verify_ptr;
 
     g_enabled = true;
 }
@@ -145,18 +155,19 @@ bool BinaryTranslationLoader::CanRun(bool specific_address)
     return true;
 }
 
-void BinaryTranslationLoader::Run()
+uint32_t BinaryTranslationLoader::Run(uint32_t instruction_count)
 {
     // No need to check the PC, Run does it anyway
-    if (!CanRun(false)) return;
+    if (!CanRun(false)) return instruction_count;
     // If verify is enabled, it will run opcodes
-    if (g_verify) return;
+    if (g_verify) return instruction_count;
 
-    RunInternal();
+    return RunInternal(instruction_count);
 }
 
-void BinaryTranslationLoader::RunInternal()
+uint32_t BinaryTranslationLoader::RunInternal(uint32_t instruction_count)
 {
+    *g_instruction_count = instruction_count;
     g_run_function();
 
     g_state->TFlag = g_state->Reg[15] & 1;
@@ -164,6 +175,8 @@ void BinaryTranslationLoader::RunInternal()
         g_state->Reg[15] &= 0xfffffffe;
     else
         g_state->Reg[15] &= 0xfffffffc;
+
+    return *g_instruction_count;
 }
 
 void Swap(void *a, void *b, size_t size)
@@ -198,7 +211,7 @@ void BinaryTranslationLoader::VerifyCallback()
         g_state_copy.SwapWith(*g_state);
 
         // Run the opcode
-        RunInternal();
+        RunInternal(0);
 
         // Test
         auto current_as_saved_state = SavedState(*g_state);

src/core/binary_translation/BinaryTranslationLoader.h

@@ -13,8 +13,9 @@ public:
     // If specific_address, checks the specific PC too
     static bool CanRun(bool specific_address);
     // Runs the state provided at SetCpuState.
-    static void Run();
+    // Returns instruction_count + number of instructions executed
+    static uint32_t Run(uint32_t instruction_count);
     // Link between Run and VerifyCallback
-    static void RunInternal();
+    static uint32_t RunInternal(uint32_t instruction_count);
     static void VerifyCallback();
 };