Shader: Initial implementation of x86_x64 JIT compiler for Pica vertex shaders.

- Config: Add an option for selecting to use shader JIT or interpreter.
- Qt: Add a menu option for enabling/disabling the shader JIT.

externals/nihstro

@@ -1 +1 @@
-Subproject commit 676254f71e0a7ef0aca8acce078d3c3dc80ccf70
+Subproject commit 445cba0b2ff8d348368e32698e4760a670260bfc

src/citra/citra.cpp

@@ -71,6 +71,7 @@ int main(int argc, char **argv) {
     EmuWindow_GLFW* emu_window = new EmuWindow_GLFW;
 
     VideoCore::g_hw_renderer_enabled = Settings::values.use_hw_renderer;
+    VideoCore::g_shader_jit_enabled = Settings::values.use_shader_jit;
 
     System::Init(emu_window);
 

src/citra/config.cpp

@@ -61,6 +61,7 @@ void Config::ReadValues() {
 
     // Renderer
     Settings::values.use_hw_renderer = glfw_config->GetBoolean("Renderer", "use_hw_renderer", false);
+    Settings::values.use_shader_jit = glfw_config->GetBoolean("Renderer", "use_shader_jit", true);
 
     Settings::values.bg_red   = (float)glfw_config->GetReal("Renderer", "bg_red",   1.0);
     Settings::values.bg_green = (float)glfw_config->GetReal("Renderer", "bg_green", 1.0);

src/citra/default_ini.h

@@ -42,6 +42,10 @@ frame_skip =
 # 0 (default): Software, 1: Hardware
 use_hw_renderer =
 
+# Whether to use the Just-In-Time (JIT) compiler for shader emulation
+# 0 : Interpreter (slow), 1 (default): JIT (fast)
+use_shader_jit =
+
 # The clear color for the renderer. What shows up on the sides of the bottom screen.
 # Must be in range of 0.0-1.0. Defaults to 1.0 for all.
 bg_red =

src/citra_qt/config.cpp

@@ -44,6 +44,7 @@ void Config::ReadValues() {
 
     qt_config->beginGroup("Renderer");
     Settings::values.use_hw_renderer = qt_config->value("use_hw_renderer", false).toBool();
+    Settings::values.use_shader_jit = qt_config->value("use_shader_jit", true).toBool();
 
     Settings::values.bg_red   = qt_config->value("bg_red",   1.0).toFloat();
     Settings::values.bg_green = qt_config->value("bg_green", 1.0).toFloat();
@@ -77,6 +78,7 @@ void Config::SaveValues() {
 
     qt_config->beginGroup("Renderer");
     qt_config->setValue("use_hw_renderer", Settings::values.use_hw_renderer);
+    qt_config->setValue("use_shader_jit", Settings::values.use_shader_jit);
 
     // Cast to double because Qt's written float values are not human-readable
     qt_config->setValue("bg_red",   (double)Settings::values.bg_red);

src/citra_qt/main.cpp

@@ -131,6 +131,9 @@ GMainWindow::GMainWindow() : emu_thread(nullptr)
     ui.action_Use_Hardware_Renderer->setChecked(Settings::values.use_hw_renderer);
     SetHardwareRendererEnabled(ui.action_Use_Hardware_Renderer->isChecked());
 
+    ui.action_Use_Shader_JIT->setChecked(Settings::values.use_shader_jit);
+    SetShaderJITEnabled(ui.action_Use_Shader_JIT->isChecked());
+
     ui.action_Single_Window_Mode->setChecked(settings.value("singleWindowMode", true).toBool());
     ToggleWindowMode();
 
@@ -144,6 +147,7 @@ GMainWindow::GMainWindow() : emu_thread(nullptr)
     connect(ui.action_Pause, SIGNAL(triggered()), this, SLOT(OnPauseGame()));
     connect(ui.action_Stop, SIGNAL(triggered()), this, SLOT(OnStopGame()));
     connect(ui.action_Use_Hardware_Renderer, SIGNAL(triggered(bool)), this, SLOT(SetHardwareRendererEnabled(bool)));
+    connect(ui.action_Use_Shader_JIT, SIGNAL(triggered(bool)), this, SLOT(SetShaderJITEnabled(bool)));
     connect(ui.action_Single_Window_Mode, SIGNAL(triggered(bool)), this, SLOT(ToggleWindowMode()));
     connect(ui.action_Hotkeys, SIGNAL(triggered()), this, SLOT(OnOpenHotkeysDialog()));
 
@@ -331,6 +335,10 @@ void GMainWindow::SetHardwareRendererEnabled(bool enabled) {
     VideoCore::g_hw_renderer_enabled = enabled;
 }
 
+void GMainWindow::SetShaderJITEnabled(bool enabled) {
+    VideoCore::g_shader_jit_enabled = enabled;
+}
+
 void GMainWindow::ToggleWindowMode() {
     if (ui.action_Single_Window_Mode->isChecked()) {
         // Render in the main window...

src/citra_qt/main.h

@@ -70,6 +70,7 @@ private slots:
     void OnConfigure();
     void OnDisplayTitleBars(bool);
     void SetHardwareRendererEnabled(bool);
+    void SetShaderJITEnabled(bool);
     void ToggleWindowMode();
 
 private:

src/citra_qt/main.ui

@@ -66,6 +66,7 @@
     <addaction name="action_Stop"/>
     <addaction name="separator"/>
     <addaction name="action_Use_Hardware_Renderer"/>
+    <addaction name="action_Use_Shader_JIT"/>
     <addaction name="action_Configure"/>
    </widget>
    <widget class="QMenu" name="menu_View">
@@ -153,6 +154,14 @@
     <string>Use Hardware Renderer</string>
    </property>
   </action>
+  <action name="action_Use_Shader_JIT">
+   <property name="checkable">
+    <bool>true</bool>
+   </property>
+   <property name="text">
+    <string>Use Shader JIT</string>
+   </property>
+  </action>
   <action name="action_Configure">
    <property name="text">
     <string>Configure ...</string>

src/core/settings.h

@@ -53,6 +53,7 @@ struct Values {
 
     // Renderer
     bool use_hw_renderer;
+    bool use_shader_jit;
 
     float bg_red;
     float bg_green;

src/video_core/CMakeLists.txt

@@ -13,6 +13,7 @@ set(SRCS
             rasterizer.cpp
             shader/shader.cpp
             shader/shader_interpreter.cpp
+            shader/shader_jit.cpp
             utils.cpp
             video_core.cpp
             )
@@ -38,10 +39,19 @@ set(HEADERS
             renderer_base.h
             shader/shader.h
             shader/shader_interpreter.h
+            shader/shader_jit.h
             utils.h
             video_core.h
             )
 
+if(_M_X86_64)
+    set(SRCS ${SRCS}
+            shader/shader_jit_x64.cpp)
+else()
+    set(SRCS ${SRCS}
+            shader/shader_jit_fake.cpp)
+endif()
+
 create_directory_groups(${SRCS} ${HEADERS})
 
 add_library(video_core STATIC ${SRCS} ${HEADERS})

src/video_core/pica.cpp

@@ -6,6 +6,7 @@
 #include <unordered_map>
 
 #include "pica.h"
+#include "shader/shader.h"
 
 namespace Pica {
 
@@ -84,6 +85,8 @@ void Init() {
 }
 
 void Shutdown() {
+    Shader::Shutdown();
+
     memset(&g_state, 0, sizeof(State));
 }
 

src/video_core/shader/shader.cpp

@@ -2,21 +2,52 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#include "common/logging/log.h"
+#include <memory>
+#include <unordered_map>
+
+#include "common/hash.h"
+#include "common/make_unique.h"
 #include "common/profiler.h"
 
 #include "video_core/debug_utils/debug_utils.h"
 #include "video_core/pica.h"
+#include "video_core/video_core.h"
 
 #include "shader.h"
 #include "shader_interpreter.h"
+#include "shader_jit.h"
 
 namespace Pica {
 
 namespace Shader {
 
+#ifdef ARCHITECTURE_x86_64
+
+static std::unordered_map<u64, CompiledShader*> shader_map;
+static JitCompiler jit;
+static CompiledShader* jit_shader;
+
+#endif // ARCHITECTURE_x86_64
+
 void Setup(UnitState& state) {
-    // TODO(bunnei): This will be used by the JIT in a subsequent patch
+#ifdef ARCHITECTURE_x86_64
+    if (VideoCore::g_shader_jit_enabled) {
+        u64 cache_key = (Common::ComputeHash64(&g_state.vs.program_code, sizeof(g_state.vs.program_code)) ^
+            Common::ComputeHash64(&g_state.vs.swizzle_data, sizeof(g_state.vs.swizzle_data)) ^
+            g_state.regs.vs.main_offset);
+
+        auto iter = shader_map.find(cache_key);
+        if (iter != shader_map.end()) {
+            jit_shader = iter->second;
+        } else {
+            jit_shader = jit.Compile();
+            shader_map.emplace(cache_key, jit_shader);
+        }
+    }
+}
+
+void Shutdown() {
+    shader_map.clear();
 }
 
 static Common::Profiling::TimingCategory shader_category("Vertex Shader");
@@ -54,7 +85,14 @@ OutputVertex Run(UnitState& state, const InputVertex& input, int num_attributes)
     state.conditional_code[0] = false;
     state.conditional_code[1] = false;
 
+#ifdef ARCHITECTURE_x86_64
+    if (VideoCore::g_shader_jit_enabled)
+        jit_shader(&state);
+    else
+        RunInterpreter(state);
+#else
     RunInterpreter(state);
+#endif
 
 #if PICA_DUMP_SHADERS
     DebugUtils::DumpShader(setup.program_code.data(), state.debug.max_offset, setup.swizzle_data.data(),

src/video_core/shader/shader.h

@@ -149,6 +149,9 @@ struct UnitState {
  */
 void Setup(UnitState& state);
 
+/// Performs any cleanup when the emulator is shutdown
+void Shutdown();
+
 /**
  * Runs the currently setup shader
  * @param state Shader unit state, must be setup per shader and per shader unit

src/video_core/shader/shader_jit.cpp

@@ -0,0 +1,36 @@
+// Copyright 2015 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "video_core/pica.h"
+
+#include "shader.h"
+#include "shader_jit.h"
+
+namespace Pica {
+
+namespace Shader {
+
+JitShader::JitShader() : jitted(nullptr) {
+}
+
+void JitShader::DoJit(JitCompiler& jit) {
+    jitted = jit.Compile();
+}
+
+void JitShader::Run(UnitState& state) {
+    if (jitted)
+        jitted(&state);
+}
+
+JitCompiler::JitCompiler() {
+    AllocCodeSpace(1024 * 1024 * 4);
+}
+
+void JitCompiler::Clear() {
+    ClearCodeSpace();
+}
+
+} // namespace Shader
+
+} // namespace Pica

src/video_core/shader/shader_jit.h

@@ -0,0 +1,85 @@
+// Copyright 2015 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <nihstro/shader_bytecode.h>
+
+#if defined(_M_X86_64)
+#include "common/x64_emitter.h"
+#else
+#include "common/fake_emitter.h"
+#endif
+
+#include "video_core/pica.h"
+
+#include "shader.h"
+
+using nihstro::Instruction;
+using nihstro::OpCode;
+using nihstro::SwizzlePattern;
+
+namespace Pica {
+
+namespace Shader {
+
+using CompiledShader = void(void* state);
+
+/**
+ * This class implements the shader JIT compiler. It recompiles a Pica shader program into x86_64
+ * code that can be executed on the host machine directly.
+ */
+class JitCompiler : public Gen::XCodeBlock {
+public:
+    JitCompiler();
+
+    CompiledShader* Compile();
+
+    void Clear();
+
+    void Compile_ADD(Instruction instr);
+    void Compile_DP3(Instruction instr);
+    void Compile_DP4(Instruction instr);
+    void Compile_MUL(Instruction instr);
+    void Compile_FLR(Instruction instr);
+    void Compile_MAX(Instruction instr);
+    void Compile_MIN(Instruction instr);
+    void Compile_RCP(Instruction instr);
+    void Compile_RSQ(Instruction instr);
+    void Compile_MOVA(Instruction instr);
+    void Compile_MOV(Instruction instr);
+    void Compile_SLTI(Instruction instr);
+    void Compile_NOP(Instruction instr);
+    void Compile_END(Instruction instr);
+    void Compile_CALL(Instruction instr);
+    void Compile_CALLC(Instruction instr);
+    void Compile_CALLU(Instruction instr);
+    void Compile_IF(Instruction instr);
+    void Compile_LOOP(Instruction instr);
+    void Compile_JMP(Instruction instr);
+    void Compile_CMP(Instruction instr);
+    void Compile_MAD(Instruction instr);
+
+private:
+    void Compile_Block(unsigned stop);
+    void Compile_NextInstr(unsigned* offset);
+
+#if defined(_M_X86_64)
+    void Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg, Gen::X64Reg dest);
+    void Compile_DestEnable(Instruction instr, Gen::X64Reg dest);
+
+    void Compile_EvaluateCondition(Instruction instr);
+    void Compile_UniformCondition(Instruction instr);
+#endif
+
+    /// Pointer to the variable that stores the current Pica code offset. Used to handle nested code blocks.
+    unsigned* offset_ptr = nullptr;
+
+    bool done = false;
+    bool looping = false;
+};
+
+} // Shader
+
+} // Pica

src/video_core/shader/shader_jit_fake.cpp

@@ -0,0 +1,91 @@
+// Copyright 2015 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "common/fake_emitter.h"
+
+#include "video_core/shader/shader.h"
+#include "video_core/shader/shader_jit.h"
+
+namespace Pica {
+
+namespace Shader {
+
+using namespace FakeGen;
+
+void Jit::Comp_ADD(Instruction instr) {
+}
+
+void Jit::Comp_DP3(Instruction instr) {
+}
+
+void Jit::Comp_DP4(Instruction instr) {
+}
+
+void Jit::Comp_MUL(Instruction instr) {
+}
+
+void Jit::Comp_FLR(Instruction instr) {
+}
+
+void Jit::Comp_MAX(Instruction instr) {
+}
+
+void Jit::Comp_MIN(Instruction instr) {
+}
+
+void Jit::Comp_MOVA(Instruction instr) {
+}
+
+void Jit::Comp_MOV(Instruction instr) {
+}
+
+void Jit::Comp_SLTI(Instruction instr) {
+}
+
+void Jit::Comp_RCP(Instruction instr) {
+}
+
+void Jit::Comp_RSQ(Instruction instr) {
+}
+
+void Jit::Comp_NOP(Instruction instr) {
+}
+
+void Jit::Comp_END(Instruction instr) {
+}
+
+void Jit::Comp_CALL(Instruction instr) {
+}
+
+void Jit::Comp_CALLC(Instruction instr) {
+}
+
+void Jit::Comp_CALLU(Instruction instr) {
+}
+
+void Jit::Comp_CMP(Instruction instr) {
+}
+
+void Jit::Comp_MAD(Instruction instr) {
+}
+
+void Jit::Comp_IF(Instruction instr) {
+}
+
+void Jit::Comp_LOOP(Instruction instr) {
+}
+
+void Jit::Comp_JMP(Instruction instr) {
+}
+
+void Jit::Comp_NextInstr(unsigned* offset) {
+}
+
+CompiledShader Jit::Compile() {
+    return nullptr;
+}
+
+} // namespace Shader
+
+} // namespace Pica

src/video_core/shader/shader_jit_x64.cpp

@@ -0,0 +1,669 @@
+// Copyright 2015 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <smmintrin.h>
+
+#include "common/abi.h"
+#include "common/cpu_detect.h"
+#include "common/x64_emitter.h"
+
+#include "shader.h"
+#include "shader_jit.h"
+
+namespace Pica {
+
+namespace Shader {
+
+using namespace Gen;
+
+typedef void (JitCompiler::*JitFunction)(Instruction instr);
+
+const JitFunction instr_table[64] = {
+    &JitCompiler::Compile_ADD,      // add
+    &JitCompiler::Compile_DP3,      // dp3
+    &JitCompiler::Compile_DP4,      // dp4
+    nullptr,                        // dph
+    nullptr,                        // unknown
+    nullptr,                        // ex2
+    nullptr,                        // lg2
+    nullptr,                        // unknown
+    &JitCompiler::Compile_MUL,      // mul
+    nullptr,                        // lge
+    nullptr,                        // slt
+    &JitCompiler::Compile_FLR,      // flr
+    &JitCompiler::Compile_MAX,      // max
+    &JitCompiler::Compile_MIN,      // min
+    &JitCompiler::Compile_RCP,      // rcp
+    &JitCompiler::Compile_RSQ,      // rsq
+    nullptr,                        // unknown
+    nullptr,                        // unknown
+    &JitCompiler::Compile_MOVA,     // mova
+    &JitCompiler::Compile_MOV,      // mov
+    nullptr,                        // unknown
+    nullptr,                        // unknown
+    nullptr,                        // unknown
+    nullptr,                        // unknown
+    nullptr,                        // dphi
+    nullptr,                        // unknown
+    nullptr,                        // sgei
+    &JitCompiler::Compile_SLTI,     // slti
+    nullptr,                        // unknown
+    nullptr,                        // unknown
+    nullptr,                        // unknown
+    nullptr,                        // unknown
+    nullptr,                        // unknown
+    &JitCompiler::Compile_NOP,      // nop
+    &JitCompiler::Compile_END,      // end
+    nullptr,                        // break
+    &JitCompiler::Compile_CALL,     // call
+    &JitCompiler::Compile_CALLC,    // callc
+    &JitCompiler::Compile_CALLU,    // callu
+    &JitCompiler::Compile_IF,       // ifu
+    &JitCompiler::Compile_IF,       // ifc
+    &JitCompiler::Compile_LOOP,     // loop
+    nullptr,                        // emit
+    nullptr,                        // sete
+    &JitCompiler::Compile_JMP,      // jmpc
+    &JitCompiler::Compile_JMP,      // jmpu
+    &JitCompiler::Compile_CMP,      // cmp
+    &JitCompiler::Compile_CMP,      // cmp
+    &JitCompiler::Compile_MAD,      // madi
+    &JitCompiler::Compile_MAD,      // madi
+    &JitCompiler::Compile_MAD,      // madi
+    &JitCompiler::Compile_MAD,      // madi
+    &JitCompiler::Compile_MAD,      // madi
+    &JitCompiler::Compile_MAD,      // madi
+    &JitCompiler::Compile_MAD,      // madi
+    &JitCompiler::Compile_MAD,      // madi
+    &JitCompiler::Compile_MAD,      // mad
+    &JitCompiler::Compile_MAD,      // mad
+    &JitCompiler::Compile_MAD,      // mad
+    &JitCompiler::Compile_MAD,      // mad
+    &JitCompiler::Compile_MAD,      // mad
+    &JitCompiler::Compile_MAD,      // mad
+    &JitCompiler::Compile_MAD,      // mad
+    &JitCompiler::Compile_MAD,      // mad
+};
+
+// The following is used to alias some commonly used registers. Generally, RAX-RDX and XMM0-XMM3 can
+// be used as scratch registers within a compiler function. The other registers have designated
+// purposes, as documented below:
+
+/// Pointer to the uniform memory
+static const X64Reg UNIFORMS = R10;
+/// The two 32-bit VS address offset registers set by the MOVA instruction
+static const X64Reg ADDROFFS_REG = R11;
+/// VS loop count register
+static const X64Reg LOOPCOUNT_REG = R12;
+/// Current VS loop iteration number (we could probably use LOOPCOUNT_REG, but this quicker)
+static const X64Reg LOOPCOUNT = RSI;
+/// Number to increment LOOPCOUNT_REG by on each loop iteration
+static const X64Reg LOOPINC = RDI;
+/// Result of the previous CMP instruction for the X-component comparison
+static const X64Reg COND0 = R13;
+/// Result of the previous CMP instruction for the Y-component comparison
+static const X64Reg COND1 = R14;
+/// Pointer to the UnitState instance for the current VS unit
+static const X64Reg STATE = R15;
+/// SIMD scratch register
+static const X64Reg SCRATCH = XMM0;
+/// Loaded with the first swizzled source register, otherwise can be used as a scratch register
+static const X64Reg SRC1 = XMM1;
+/// Loaded with the second swizzled source register, otherwise can be used as a scratch register
+static const X64Reg SRC2 = XMM2;
+/// Loaded with the third swizzled source register, otherwise can be used as a scratch register
+static const X64Reg SRC3 = XMM3;
+/// Constant vector of [1.0f, 1.0f, 1.0f, 1.0f], used to efficiently set a vector to one
+static const X64Reg ONE = XMM14;
+/// Constant vector of [-0.f, -0.f, -0.f, -0.f], used to efficiently negate a vector with XOR
+static const X64Reg NEGBIT = XMM15;
+
+/// Raw constant for the source register selector that indicates no swizzling is performed
+static const u8 NO_SRC_REG_SWIZZLE = 0x1b;
+/// Raw constant for the destination register enable mask that indicates all components are enabled
+static const u8 NO_DEST_REG_MASK = 0xf;
+
+/**
+ * Loads and swizzles a source register into the specified XMM register.
+ * @param instr VS instruction, used for determining how to load the source register
+ * @param src_num Number indicating which source register to load (1 = src1, 2 = src2, 3 = src3)
+ * @param src_reg SourceRegister object corresponding to the source register to load
+ * @param dest Destination XMM register to store the loaded, swizzled source register
+ */
+void JitCompiler::Compile_SwizzleSrc(Instruction instr, unsigned src_num, SourceRegister src_reg, X64Reg dest) {
+    X64Reg src_ptr;
+    std::size_t src_offset;
+
+    if (src_reg.GetRegisterType() == RegisterType::FloatUniform) {
+        src_ptr = UNIFORMS;
+        src_offset = src_reg.GetIndex() * sizeof(float24) * 4;
+    } else {
+        src_ptr = STATE;
+        src_offset = UnitState::InputOffset(src_reg);
+    }
+
+    unsigned operand_desc_id;
+    if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD ||
+        instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
+        // The MAD and MADI instructions do not use the address offset registers, so loading the
+        // source is a bit simpler here
+
+        operand_desc_id = instr.mad.operand_desc_id;
+
+        // Load the source
+        MOVAPS(dest, MDisp(src_ptr, src_offset));
+    } else {
+        operand_desc_id = instr.common.operand_desc_id;
+
+        const bool is_inverted = (0 != (instr.opcode.Value().GetInfo().subtype & OpCode::Info::SrcInversed));
+        unsigned offset_src = is_inverted ? 2 : 1;
+
+        if (src_num == offset_src && instr.common.address_register_index != 0) {
+            switch (instr.common.address_register_index) {
+            case 1: // address offset 1
+                MOV(32, R(RBX), R(ADDROFFS_REG));
+                break;
+            case 2: // address offset 2
+                MOV(64, R(RBX), R(ADDROFFS_REG));
+                SHR(64, R(RBX), Imm8(32));
+                break;
+            case 3: // adddress offet 3
+                MOV(64, R(RBX), R(LOOPCOUNT_REG));
+                break;
+            default:
+                UNREACHABLE();
+                break;
+            }
+
+            MOVAPS(dest, MComplex(src_ptr, RBX, 1, src_offset));
+        } else {
+            // Load the source
+            MOVAPS(dest, MDisp(src_ptr, src_offset));
+        }
+    }
+
+    SwizzlePattern swiz = { g_state.vs.swizzle_data[operand_desc_id] };
+
+    // Generate instructions for source register swizzling as needed
+    u8 sel = swiz.GetRawSelector(src_num);
+    if (sel != NO_SRC_REG_SWIZZLE) {
+        // Selector component order needs to be reversed for the SHUFPS instruction
+        sel = ((sel & 0xc0) >> 6) | ((sel & 3) << 6) | ((sel & 0xc) << 2) | ((sel & 0x30) >> 2);
+
+        // Shuffle inputs for swizzle
+        SHUFPS(dest, R(dest), sel);
+    }
+
+    // If the source register should be negated, flip the negative bit using XOR
+    const bool negate[] = { swiz.negate_src1, swiz.negate_src2, swiz.negate_src3 };
+    if (negate[src_num - 1]) {
+        XORPS(dest, R(NEGBIT));
+    }
+}
+
+void JitCompiler::Compile_DestEnable(Instruction instr,X64Reg src) {
+    DestRegister dest;
+    unsigned operand_desc_id;
+    if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MAD ||
+        instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
+        operand_desc_id = instr.mad.operand_desc_id;
+        dest = instr.mad.dest.Value();
+    } else {
+        operand_desc_id = instr.common.operand_desc_id;
+        dest = instr.common.dest.Value();
+    }
+
+    SwizzlePattern swiz = { g_state.vs.swizzle_data[operand_desc_id] };
+
+    // If all components are enabled, write the result to the destination register
+    if (swiz.dest_mask == NO_DEST_REG_MASK) {
+        // Store dest back to memory
+        MOVAPS(MDisp(STATE, UnitState::OutputOffset(dest)), src);
+
+    } else {
+        // Not all components are enabled, so mask the result when storing to the destination register...
+        MOVAPS(SCRATCH, MDisp(STATE, UnitState::OutputOffset(dest)));
+
+        if (Common::cpu_info.bSSE4_1) {
+            u8 mask = ((swiz.dest_mask & 1) << 3) | ((swiz.dest_mask & 8) >> 3) | ((swiz.dest_mask & 2) << 1) | ((swiz.dest_mask & 4) >> 1);
+            BLENDPS(SCRATCH, R(src), mask);
+        } else {
+            MOVAPS(XMM4, R(src));
+            UNPCKHPS(XMM4, R(SCRATCH)); // Unpack X/Y components of source and destination
+            UNPCKLPS(SCRATCH, R(src)); // Unpack Z/W components of source and destination
+
+            // Compute selector to selectively copy source components to destination for SHUFPS instruction
+            u8 sel = ((swiz.DestComponentEnabled(0) ? 1 : 0) << 0) |
+                     ((swiz.DestComponentEnabled(1) ? 3 : 2) << 2) |
+                     ((swiz.DestComponentEnabled(2) ? 0 : 1) << 4) |
+                     ((swiz.DestComponentEnabled(3) ? 2 : 3) << 6);
+            SHUFPS(SCRATCH, R(XMM4), sel);
+        }
+
+        // Store dest back to memory
+        MOVAPS(MDisp(STATE, UnitState::OutputOffset(dest)), SCRATCH);
+    }
+}
+
+void JitCompiler::Compile_EvaluateCondition(Instruction instr) {
+    // Note: NXOR is used below to check for equality
+    switch (instr.flow_control.op) {
+    case Instruction::FlowControlType::Or:
+        MOV(32, R(RAX), R(COND0));
+        MOV(32, R(RBX), R(COND1));
+        XOR(32, R(RAX), Imm32(instr.flow_control.refx.Value() ^ 1));
+        XOR(32, R(RBX), Imm32(instr.flow_control.refy.Value() ^ 1));
+        OR(32, R(RAX), R(RBX));
+        break;
+
+    case Instruction::FlowControlType::And:
+        MOV(32, R(RAX), R(COND0));
+        MOV(32, R(RBX), R(COND1));
+        XOR(32, R(RAX), Imm32(instr.flow_control.refx.Value() ^ 1));
+        XOR(32, R(RBX), Imm32(instr.flow_control.refy.Value() ^ 1));
+        AND(32, R(RAX), R(RBX));
+        break;
+
+    case Instruction::FlowControlType::JustX:
+        MOV(32, R(RAX), R(COND0));
+        XOR(32, R(RAX), Imm32(instr.flow_control.refx.Value() ^ 1));
+        break;
+
+    case Instruction::FlowControlType::JustY:
+        MOV(32, R(RAX), R(COND1));
+        XOR(32, R(RAX), Imm32(instr.flow_control.refy.Value() ^ 1));
+        break;
+    }
+}
+
+void JitCompiler::Compile_UniformCondition(Instruction instr) {
+    int offset = offsetof(decltype(g_state.vs.uniforms), b) + (instr.flow_control.bool_uniform_id * sizeof(bool));
+    CMP(sizeof(bool) * 8, MDisp(UNIFORMS, offset), Imm8(0));
+}
+
+void JitCompiler::Compile_ADD(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+    ADDPS(SRC1, R(SRC2));
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_DP3(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+
+    if (Common::cpu_info.bSSE4_1) {
+        DPPS(SRC1, R(SRC2), 0x7f);
+    } else {
+        MULPS(SRC1, R(SRC2));
+
+        MOVAPS(SRC2, R(SRC1));
+        SHUFPS(SRC2, R(SRC2), _MM_SHUFFLE(1, 1, 1, 1));
+
+        MOVAPS(SRC3, R(SRC1));
+        SHUFPS(SRC3, R(SRC3), _MM_SHUFFLE(2, 2, 2, 2));
+
+        SHUFPS(SRC1, R(SRC1), _MM_SHUFFLE(0, 0, 0, 0));
+        ADDPS(SRC1, R(SRC2));
+        ADDPS(SRC1, R(SRC3));
+    }
+
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_DP4(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+
+    if (Common::cpu_info.bSSE4_1) {
+        DPPS(SRC1, R(SRC2), 0xff);
+    } else {
+        MULPS(SRC1, R(SRC2));
+
+        MOVAPS(SRC2, R(SRC1));
+        SHUFPS(SRC1, R(SRC1), _MM_SHUFFLE(2, 3, 0, 1)); // XYZW -> ZWXY
+        ADDPS(SRC1, R(SRC2));
+
+        MOVAPS(SRC2, R(SRC1));
+        SHUFPS(SRC1, R(SRC1), _MM_SHUFFLE(0, 1, 2, 3)); // XYZW -> WZYX
+        ADDPS(SRC1, R(SRC2));
+    }
+
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_MUL(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+    MULPS(SRC1, R(SRC2));
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_FLR(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+
+    if (Common::cpu_info.bSSE4_1) {
+        ROUNDFLOORPS(SRC1, R(SRC1));
+    } else {
+        CVTPS2DQ(SRC1, R(SRC1));
+        CVTDQ2PS(SRC1, R(SRC1));
+    }
+
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_MAX(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+    MAXPS(SRC1, R(SRC2));
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_MIN(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+    MINPS(SRC1, R(SRC2));
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_MOVA(Instruction instr) {
+    SwizzlePattern swiz = { g_state.vs.swizzle_data[instr.common.operand_desc_id] };
+
+    if (!swiz.DestComponentEnabled(0) && !swiz.DestComponentEnabled(1)) {
+        return; // NoOp
+    }
+
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+
+    // Convert floats to integers (only care about X and Y components)
+    CVTPS2DQ(SRC1, R(SRC1));
+
+    // Get result
+    MOVQ_xmm(R(RAX), SRC1);
+    SHL(64, R(RAX), Imm8(4)); // Multiply by 16 to be used as an offset later
+
+    // Handle destination enable
+    if (swiz.DestComponentEnabled(0) && swiz.DestComponentEnabled(1)) {
+        MOV(64, R(ADDROFFS_REG), R(RAX)); // Overwrite both
+    } else {
+        if (swiz.DestComponentEnabled(0)) {
+            // Preserve Y-component
+
+            // Clear low 32 bits of previous address register
+            MOV(32, R(RBX), R(ADDROFFS_REG));
+            XOR(64, R(ADDROFFS_REG), R(RBX));
+
+            // Clear high 32-bits of new address register
+            MOV(32, R(RAX), R(RAX));
+        } else if (swiz.DestComponentEnabled(1)) {
+            // Preserve X-component
+
+            // Clear high 32-bits of previous address register
+            MOV(32, R(ADDROFFS_REG), R(ADDROFFS_REG));
+
+            // Clear low 32 bits of new address register
+            MOV(32, R(RBX), R(RAX));
+            XOR(64, R(RAX), R(RBX));
+        }
+
+        OR(64, R(ADDROFFS_REG), R(RAX)); // Combine result
+    }
+}
+
+void JitCompiler::Compile_MOV(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_SLTI(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1i, SRC1);
+    Compile_SwizzleSrc(instr, 1, instr.common.src2i, SRC2);
+
+    CMPSS(SRC1, R(SRC2), CMP_LT);
+    ANDPS(SRC1, R(ONE));
+
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_RCP(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+
+    // TODO(bunnei): RCPPS is a pretty rough approximation, this might cause problems if Pica
+    // performs this operation more accurately. This should be checked on hardware.
+    RCPPS(SRC1, R(SRC1));
+
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_RSQ(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+
+    // TODO(bunnei): RSQRTPS is a pretty rough approximation, this might cause problems if Pica
+    // performs this operation more accurately. This should be checked on hardware.
+    RSQRTPS(SRC1, R(SRC1));
+
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_NOP(Instruction instr) {
+}
+
+void JitCompiler::Compile_END(Instruction instr) {
+    ABI_PopAllCalleeSavedRegsAndAdjustStack();
+    RET();
+    done = true;
+}
+
+void JitCompiler::Compile_CALL(Instruction instr) {
+    unsigned offset = instr.flow_control.dest_offset;
+    while (offset < (instr.flow_control.dest_offset + instr.flow_control.num_instructions)) {
+        Compile_NextInstr(&offset);
+    }
+}
+
+void JitCompiler::Compile_CALLC(Instruction instr) {
+    Compile_EvaluateCondition(instr);
+    FixupBranch b = J_CC(CC_Z, true);
+    Compile_CALL(instr);
+    SetJumpTarget(b);
+}
+
+void JitCompiler::Compile_CALLU(Instruction instr) {
+    Compile_UniformCondition(instr);
+    FixupBranch b = J_CC(CC_Z, true);
+    Compile_CALL(instr);
+    SetJumpTarget(b);
+}
+
+void JitCompiler::Compile_CMP(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.common.src1, SRC1);
+    Compile_SwizzleSrc(instr, 2, instr.common.src2, SRC2);
+
+    static const u8 cmp[] = { CMP_EQ, CMP_NEQ, CMP_LT, CMP_LE, CMP_NLE, CMP_NLT };
+
+    if (instr.common.compare_op.x == instr.common.compare_op.y) {
+        // Compare X-component and Y-component together
+        CMPPS(SRC1, R(SRC2), cmp[instr.common.compare_op.x]);
+
+        MOVQ_xmm(R(COND0), SRC1);
+        MOV(64, R(COND1), R(COND0));
+    } else {
+        // Compare X-component
+        MOVAPS(SCRATCH, R(SRC1));
+        CMPSS(SCRATCH, R(SRC2), cmp[instr.common.compare_op.x]);
+
+        // Compare Y-component
+        CMPPS(SRC1, R(SRC2), cmp[instr.common.compare_op.y]);
+
+        MOVQ_xmm(R(COND0), SCRATCH);
+        MOVQ_xmm(R(COND1), SRC1);
+    }
+
+    SHR(32, R(COND0), Imm8(31));
+    SHR(64, R(COND1), Imm8(63));
+}
+
+void JitCompiler::Compile_MAD(Instruction instr) {
+    Compile_SwizzleSrc(instr, 1, instr.mad.src1, SRC1);
+
+    if (instr.opcode.Value().EffectiveOpCode() == OpCode::Id::MADI) {
+        Compile_SwizzleSrc(instr, 2, instr.mad.src2i, SRC2);
+        Compile_SwizzleSrc(instr, 3, instr.mad.src3i, SRC3);
+    } else {
+        Compile_SwizzleSrc(instr, 2, instr.mad.src2, SRC2);
+        Compile_SwizzleSrc(instr, 3, instr.mad.src3, SRC3);
+    }
+
+    if (Common::cpu_info.bFMA) {
+        VFMADD213PS(SRC1, SRC2, R(SRC3));
+    } else {
+        MULPS(SRC1, R(SRC2));
+        ADDPS(SRC1, R(SRC3));
+    }
+
+    Compile_DestEnable(instr, SRC1);
+}
+
+void JitCompiler::Compile_IF(Instruction instr) {
+    ASSERT_MSG(instr.flow_control.dest_offset > *offset_ptr, "Backwards if-statements not supported");
+
+    // Evaluate the "IF" condition
+    if (instr.opcode.Value() == OpCode::Id::IFU) {
+        Compile_UniformCondition(instr);
+    } else if (instr.opcode.Value() == OpCode::Id::IFC) {
+        Compile_EvaluateCondition(instr);
+    }
+    FixupBranch b = J_CC(CC_Z, true);
+
+    // Compile the code that corresponds to the condition evaluating as true
+    Compile_Block(instr.flow_control.dest_offset - 1);
+
+    // If there isn't an "ELSE" condition, we are done here
+    if (instr.flow_control.num_instructions == 0) {
+        SetJumpTarget(b);
+        return;
+    }
+
+    FixupBranch b2 = J(true);
+
+    SetJumpTarget(b);
+
+    // This code corresponds to the "ELSE" condition
+    // Comple the code that corresponds to the condition evaluating as false
+    Compile_Block(instr.flow_control.dest_offset + instr.flow_control.num_instructions - 1);
+
+    SetJumpTarget(b2);
+}
+
+void JitCompiler::Compile_LOOP(Instruction instr) {
+    ASSERT_MSG(instr.flow_control.dest_offset > *offset_ptr, "Backwards loops not supported");
+    ASSERT_MSG(!looping, "Nested loops not supported");
+
+    looping = true;
+
+    int offset = offsetof(decltype(g_state.vs.uniforms), i) + (instr.flow_control.int_uniform_id * sizeof(Math::Vec4<u8>));
+    MOV(32, R(LOOPCOUNT), MDisp(UNIFORMS, offset));
+    MOV(32, R(LOOPCOUNT_REG), R(LOOPCOUNT));
+    SHR(32, R(LOOPCOUNT_REG), Imm8(8));
+    AND(32, R(LOOPCOUNT_REG), Imm32(0xff)); // Y-component is the start
+    MOV(32, R(LOOPINC), R(LOOPCOUNT));
+    SHR(32, R(LOOPINC), Imm8(16));
+    MOVZX(32, 8, LOOPINC, R(LOOPINC)); // Z-component is the incrementer
+    MOVZX(32, 8, LOOPCOUNT, R(LOOPCOUNT)); // X-component is iteration count
+    ADD(32, R(LOOPCOUNT), Imm8(1)); // Iteration count is X-component + 1
+
+    auto loop_start = GetCodePtr();
+
+    Compile_Block(instr.flow_control.dest_offset);
+
+    ADD(32, R(LOOPCOUNT_REG), R(LOOPINC)); // Increment LOOPCOUNT_REG by Z-component
+    SUB(32, R(LOOPCOUNT), Imm8(1)); // Increment loop count by 1
+    J_CC(CC_NZ, loop_start); // Loop if not equal
+
+    looping = false;
+}
+
+void JitCompiler::Compile_JMP(Instruction instr) {
+    ASSERT_MSG(instr.flow_control.dest_offset > *offset_ptr, "Backwards jumps not supported");
+
+    if (instr.opcode.Value() == OpCode::Id::JMPC)
+        Compile_EvaluateCondition(instr);
+    else if (instr.opcode.Value() == OpCode::Id::JMPU)
+        Compile_UniformCondition(instr);
+    else
+        UNREACHABLE();
+
+    FixupBranch b = J_CC(CC_NZ, true);
+
+    Compile_Block(instr.flow_control.dest_offset);
+
+    SetJumpTarget(b);
+}
+
+void JitCompiler::Compile_Block(unsigned stop) {
+    // Save current offset pointer
+    unsigned* prev_offset_ptr = offset_ptr;
+    unsigned offset = *prev_offset_ptr;
+
+    while (offset <= stop)
+        Compile_NextInstr(&offset);
+
+    // Restore current offset pointer
+    offset_ptr = prev_offset_ptr;
+    *offset_ptr = offset;
+}
+
+void JitCompiler::Compile_NextInstr(unsigned* offset) {
+    offset_ptr = offset;
+
+    Instruction instr = *(Instruction*)&g_state.vs.program_code[(*offset_ptr)++];
+    OpCode::Id opcode = instr.opcode.Value();
+    auto instr_func = instr_table[static_cast<unsigned>(opcode)];
+
+    if (instr_func) {
+        // JIT the instruction!
+        ((*this).*instr_func)(instr);
+    } else {
+        // Unhandled instruction
+        LOG_CRITICAL(HW_GPU, "Unhandled instruction: 0x%02x (0x%08x)", instr.opcode.Value(), instr.hex);
+    }
+}
+
+CompiledShader* JitCompiler::Compile() {
+    const u8* start = GetCodePtr();
+    const auto& code = g_state.vs.program_code;
+    unsigned offset = g_state.regs.vs.main_offset;
+
+    ABI_PushAllCalleeSavedRegsAndAdjustStack();
+
+    MOV(PTRBITS, R(STATE), R(ABI_PARAM1));
+    MOV(PTRBITS, R(UNIFORMS), ImmPtr(&g_state.vs.uniforms));
+
+    // Zero address/loop  registers
+    XOR(64, R(ADDROFFS_REG_0), R(ADDROFFS_REG_0));
+    XOR(64, R(ADDROFFS_REG_1), R(ADDROFFS_REG_1));
+    XOR(64, R(LOOPCOUNT_REG), R(LOOPCOUNT_REG));
+
+    // Used to set a register to one
+    static const __m128 one = { 1.f, 1.f, 1.f, 1.f };
+    MOV(PTRBITS, R(RAX), ImmPtr(&one));
+    MOVAPS(ONE, MDisp(RAX, 0));
+
+    // Used to negate registers
+    static const __m128 neg = { -0.f, -0.f, -0.f, -0.f };
+    MOV(PTRBITS, R(RAX), ImmPtr(&neg));
+    MOVAPS(NEGBIT, MDisp(RAX, 0));
+
+    looping = false;
+    done = false;
+    while (offset < g_state.vs.program_code.size()) {
+        Compile_NextInstr(&offset);
+    }
+
+    return (CompiledShader*)start;
+}
+
+} // namespace Shader
+
+} // namespace Pica

src/video_core/video_core.cpp

@@ -23,6 +23,7 @@ EmuWindow*      g_emu_window    = nullptr;     ///< Frontend emulator window
 RendererBase*   g_renderer      = nullptr;     ///< Renderer plugin
 
 std::atomic<bool> g_hw_renderer_enabled;
+std::atomic<bool> g_shader_jit_enabled;
 
 /// Initialize the video core
 void Init(EmuWindow* emu_window) {

src/video_core/video_core.h

@@ -32,8 +32,9 @@ static const int kScreenBottomHeight    = 240;  ///< 3DS bottom screen height
 extern RendererBase*   g_renderer;              ///< Renderer plugin
 extern EmuWindow*      g_emu_window;            ///< Emu window
 
-// TODO: Wrap this in a user settings struct along with any other graphics settings (often set from qt ui)
+// TODO: Wrap these in a user settings struct along with any other graphics settings (often set from qt ui)
 extern std::atomic<bool> g_hw_renderer_enabled;
+extern std::atomic<bool> g_shader_jit_enabled;
 
 /// Start the video core
 void Start();