New SPU interpreter (SPU fast)

Use LLVM to build SPU interpreter.
Simplify interpreter loop.

rpcs3/Emu/CPU/CPUTranslator.h

@@ -923,6 +923,12 @@ public:
 		return llvm_value_t<T>::get_type(m_context);
 	}
 
+	template <typename R, typename... Args>
+	llvm::FunctionType* get_ftype()
+	{
+		return llvm::FunctionType::get(get_type<R>(), {get_type<Args>()...}, false);
+	}
+
 	template <typename T>
 	using value_t = llvm_value_t<T>;
 
@@ -1083,6 +1089,15 @@ public:
 		return result;
 	}
 
+	template <typename T, typename V>
+	auto vsplat(V v)
+	{
+		value_t<T> result;
+		static_assert(result.is_vector);
+		result.value = m_ir->CreateVectorSplat(result.is_vector, v.eval(m_ir));
+		return result;
+	}
+
 	// Min
 	template <typename T>
 	auto min(T a, T b)
@@ -1257,6 +1272,19 @@ public:
 		return result;
 	}
 
+	llvm::Value* load_const(llvm::GlobalVariable* g, llvm::Value* i)
+	{
+		return m_ir->CreateLoad(m_ir->CreateGEP(g, {m_ir->getInt64(0), m_ir->CreateZExtOrTrunc(i, get_type<u64>())}));
+	}
+
+	template <typename T, typename I>
+	value_t<T> load_const(llvm::GlobalVariable* g, I i)
+	{
+		value_t<T> result;
+		result.value = load_const(g, i.eval(m_ir));
+		return result;
+	}
+
 	template <typename R = v128>
 	R get_const_vector(llvm::Constant*, u32 a, u32 b);
 

rpcs3/Emu/Cell/SPURecompiler.h

@@ -108,6 +108,9 @@ public:
 	// All dispatchers (array allocated in jit memory)
 	static atomic_t<spu_function_t>* const g_dispatcher;
 
+	// Interpreter entry point
+	static spu_function_t g_interpreter;
+
 	struct passive_lock
 	{
 		spu_runtime& _this;
@@ -253,7 +256,7 @@ public:
 	static std::unique_ptr<spu_recompiler_base> make_asmjit_recompiler();
 
 	// Create recompiler instance (LLVM)
-	static std::unique_ptr<spu_recompiler_base> make_llvm_recompiler();
+	static std::unique_ptr<spu_recompiler_base> make_llvm_recompiler(u8 magn = 0);
 
 	enum : u8
 	{

rpcs3/Emu/Cell/SPUThread.cpp

@@ -23,16 +23,6 @@
 #include <atomic>
 #include <thread>
 
-const bool s_use_ssse3 =
-#ifdef _MSC_VER
-	utils::has_ssse3();
-#elif __SSSE3__
-	true;
-#else
-	false;
-#define _mm_shuffle_epi8
-#endif
-
 #ifdef _MSC_VER
 bool operator ==(const u128& lhs, const u128& rhs)
 {
@@ -597,6 +587,23 @@ void spu_thread::cpu_task()
 		return;
 	}
 
+	if (spu_runtime::g_interpreter)
+	{
+		while (true)
+		{
+			if (UNLIKELY(state))
+			{
+				if (check_state())
+					break;
+			}
+
+			spu_runtime::g_interpreter(*this, vm::_ptr<u8>(offset), nullptr);
+		}
+
+		cpu_stop();
+		return;
+	}
+
 	// Select opcode table
 	const auto& table = *(
 		g_cfg.core.spu_decoder == spu_decoder_type::precise ? &g_spu_interpreter_precise.get_table() :
@@ -605,11 +612,6 @@ void spu_thread::cpu_task()
 
 	// LS pointer
 	const auto base = vm::_ptr<const u8>(offset);
-	const auto bswap4 = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3);
-
-	v128 _op;
-	using func_t = decltype(&spu_interpreter::UNK);
-	func_t func0, func1, func2, func3, func4, func5;
 
 	while (true)
 	{
@@ -617,62 +619,11 @@ void spu_thread::cpu_task()
 		{
 			if (check_state())
 				break;
-
-			// Decode single instruction (may be step)
-			const u32 op = *reinterpret_cast<const be_t<u32>*>(base + pc);
-			if (table[spu_decode(op)](*this, {op})) { pc += 4; }
-			continue;
 		}
 
-		if (pc % 16 || !s_use_ssse3)
+		const u32 op = *reinterpret_cast<const be_t<u32>*>(base + pc);
-		{
+		if (table[spu_decode(op)](*this, {op}))
-			// Unaligned
-			const u32 op = *reinterpret_cast<const be_t<u32>*>(base + pc);
-			if (table[spu_decode(op)](*this, {op})) { pc += 4; }
-			continue;
-		}
-
-		// Reinitialize
-		_op.vi =  _mm_shuffle_epi8(_mm_load_si128(reinterpret_cast<const __m128i*>(base + pc)), bswap4);
-		func0 = table[spu_decode(_op._u32[0])];
-		func1 = table[spu_decode(_op._u32[1])];
-		func2 = table[spu_decode(_op._u32[2])];
-		func3 = table[spu_decode(_op._u32[3])];
-
-		while (LIKELY(func0(*this, {_op._u32[0]})))
-		{
 			pc += 4;
-			if (LIKELY(func1(*this, {_op._u32[1]})))
-			{
-				pc += 4;
-				u32 op2 = _op._u32[2];
-				u32 op3 = _op._u32[3];
-				_op.vi =  _mm_shuffle_epi8(_mm_load_si128(reinterpret_cast<const __m128i*>(base + pc + 8)), bswap4);
-				func0 = table[spu_decode(_op._u32[0])];
-				func1 = table[spu_decode(_op._u32[1])];
-				func4 = table[spu_decode(_op._u32[2])];
-				func5 = table[spu_decode(_op._u32[3])];
-				if (LIKELY(func2(*this, {op2})))
-				{
-					pc += 4;
-					if (LIKELY(func3(*this, {op3})))
-					{
-						pc += 4;
-						func2 = func4;
-						func3 = func5;
-
-						if (UNLIKELY(state))
-						{
-							break;
-						}
-						continue;
-					}
-					break;
-				}
-				break;
-			}
-			break;
-		}
 	}
 
 	cpu_stop();

rpcs3/rpcs3qt/settings_dialog.cpp

@@ -291,7 +291,12 @@ settings_dialog::settings_dialog(std::shared_ptr<gui_settings> guiSettings, std:
 		ui->accurateXFloat->setEnabled(checked);
 	});
 
-	ui->accurateXFloat->setEnabled(ui->spu_llvm->isChecked());
+	connect(ui->spu_fast, &QAbstractButton::toggled, [this](bool checked)
+	{
+		ui->accurateXFloat->setEnabled(checked);
+	});
+
+	ui->accurateXFloat->setEnabled(ui->spu_llvm->isChecked() || ui->spu_fast->isChecked());
 
 #ifndef LLVM_AVAILABLE
 	ui->ppu_llvm->setEnabled(false);