10 files changed, 175 insertions, 66 deletions

diff --git a/externals/CMakeLists.txt b/externals/CMakeLists.txt
index aa01acda8..dcf6c460f 100644
--- a/externals/CMakeLists.txt
+++ b/externals/CMakeLists.txt
@@ -10,6 +10,14 @@ target_include_directories(catch-single-include INTERFACE catch/single_include)
 # Crypto++
 add_subdirectory(cryptopp)
 
+# Dynarmic
+if (ARCHITECTURE_x86_64)
+    add_library(xbyak INTERFACE)
+    set(DYNARMIC_TESTS OFF)
+    set(DYNARMIC_NO_BUNDLED_FMT ON)
+    add_subdirectory(dynarmic)
+endif()
+
 # libfmt
 add_subdirectory(fmt)
 
@@ -49,7 +57,7 @@ target_include_directories(unicorn-headers INTERFACE ./unicorn/include)
 # Xbyak
 if (ARCHITECTURE_x86_64)
     # Defined before "dynarmic" above
-    add_library(xbyak INTERFACE)
+    # add_library(xbyak INTERFACE)
     target_include_directories(xbyak INTERFACE ./xbyak/xbyak)
     target_compile_definitions(xbyak INTERFACE XBYAK_NO_OP_NAMES)
 endif()
diff --git a/externals/dynarmic b/externals/dynarmic
-Subproject 305fba50babf736d77b71c5a44dd6b0ccb7f9d1
+Subproject a5caa7cd8d5741d34dcf0b3447b5f1c3f7333d5
diff --git a/src/citra_qt/debugger/wait_tree.cpp b/src/citra_qt/debugger/wait_tree.cpp
index c066a3e17..eefbcb9f1 100644
--- a/src/citra_qt/debugger/wait_tree.cpp
+++ b/src/citra_qt/debugger/wait_tree.cpp
@@ -168,7 +168,7 @@ QString WaitTreeThread::GetText() const {
     }
     QString pc_info = tr(" PC = 0x%1 LR = 0x%2")
                           .arg(thread.context.pc, 8, 16, QLatin1Char('0'))
-                          .arg(thread.context.lr, 8, 16, QLatin1Char('0'));
+                          .arg(thread.context.cpu_registers[31], 8, 16, QLatin1Char('0'));
     return WaitTreeWaitObject::GetText() + pc_info + " (" + status + ") ";
 }
 
diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
index 025177b89..e8e98a095 100644
--- a/src/core/CMakeLists.txt
+++ b/src/core/CMakeLists.txt
@@ -171,7 +171,7 @@ set(HEADERS
 
 create_directory_groups(${SRCS} ${HEADERS})
 add_library(core STATIC ${SRCS} ${HEADERS})
-target_link_libraries(core PUBLIC common PRIVATE audio_core network video_core)
+target_link_libraries(core PUBLIC common PRIVATE audio_core dynarmic network video_core)
 target_link_libraries(core PUBLIC Boost::boost PRIVATE cryptopp fmt lz4_static unicorn)
 if (ENABLE_WEB_SERVICE)
     target_link_libraries(core PUBLIC json-headers web_service)
diff --git a/src/core/arm/arm_interface.h b/src/core/arm/arm_interface.h
index c0d6e2604..531875006 100644
--- a/src/core/arm/arm_interface.h
+++ b/src/core/arm/arm_interface.h
@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <array>
 #include "common/common_types.h"
 #include "core/hle/kernel/vm_manager.h"
 
@@ -13,15 +14,12 @@ public:
     virtual ~ARM_Interface() {}
 
     struct ThreadContext {
-        u64 cpu_registers[30];
-        u64 lr;
+        std::array<u64, 31> cpu_registers;
         u64 sp;
         u64 pc;
         u64 cpsr;
-        u128 fpu_registers[32];
+        std::array<u128, 32> fpu_registers;
         u64 fpscr;
-        u64 fpexc;
-
 
         // TODO(bunnei): Fix once we have proper support for tpidrro_el0, etc. in the JIT
         VAddr tls_address;
@@ -75,9 +73,9 @@ public:
      */
     virtual void SetReg(int index, u64 value) = 0;
 
-    virtual const u128& GetExtReg(int index) const = 0;
+    virtual u128 GetExtReg(int index) const = 0;
 
-    virtual void SetExtReg(int index, u128& value) = 0;
+    virtual void SetExtReg(int index, u128 value) = 0;
 
     /**
      * Gets the value of a VFP register
diff --git a/src/core/arm/dynarmic/arm_dynarmic.cpp b/src/core/arm/dynarmic/arm_dynarmic.cpp
index 12b7fc926..c1a63e064 100644
--- a/src/core/arm/dynarmic/arm_dynarmic.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic.cpp
@@ -2,43 +2,114 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <memory>
+#include <dynarmic/A64/a64.h>
+#include <dynarmic/A64/config.h>
 #include "core/arm/dynarmic/arm_dynarmic.h"
-
-ARM_Dynarmic::ARM_Dynarmic() {
-    UNIMPLEMENTED();
-}
-
-void ARM_Dynarmic::MapBackingMemory(VAddr /*address*/, size_t /*size*/, u8* /*memory*/,
-                                    Kernel::VMAPermission /*perms*/) {
-    UNIMPLEMENTED();
-}
-
-void ARM_Dynarmic::SetPC(u64 /*pc*/) {
-    UNIMPLEMENTED();
+#include "core/core_timing.h"
+#include "core/hle/kernel/svc.h"
+#include "core/memory.h"
+
+class ARM_Dynarmic_Callbacks : public Dynarmic::A64::UserCallbacks {
+public:
+    explicit ARM_Dynarmic_Callbacks(ARM_Dynarmic& parent) : parent(parent) {}
+    ~ARM_Dynarmic_Callbacks() = default;
+
+    virtual u8 MemoryRead8(u64 vaddr) override {
+        return Memory::Read8(vaddr);
+    }
+    virtual u16 MemoryRead16(u64 vaddr) override {
+        return Memory::Read16(vaddr);
+    }
+    virtual u32 MemoryRead32(u64 vaddr) override {
+        return Memory::Read32(vaddr);
+    }
+    virtual u64 MemoryRead64(u64 vaddr) override {
+        return Memory::Read64(vaddr);
+    }
+
+    virtual void MemoryWrite8(u64 vaddr, u8 value) override {
+        Memory::Write8(vaddr, value);
+    }
+    virtual void MemoryWrite16(u64 vaddr, u16 value) override {
+        Memory::Write16(vaddr, value);
+    }
+    virtual void MemoryWrite32(u64 vaddr, u32 value) override {
+        Memory::Write32(vaddr, value);
+    }
+    virtual void MemoryWrite64(u64 vaddr, u64 value) override {
+        Memory::Write64(vaddr, value);
+    }
+
+    virtual void InterpreterFallback(u64 pc, size_t num_instructions) override {
+        ARM_Interface::ThreadContext ctx;
+        parent.SaveContext(ctx);
+        parent.inner_unicorn.LoadContext(ctx);
+        parent.inner_unicorn.ExecuteInstructions(num_instructions);
+        parent.inner_unicorn.SaveContext(ctx);
+        parent.LoadContext(ctx);
+        num_interpreted_instructions += num_instructions;
+    }
+
+    virtual void CallSVC(u32 swi) override {
+        printf("svc %x\n", swi);
+        Kernel::CallSVC(swi);
+    }
+
+    virtual void AddTicks(u64 ticks) override {
+        if (ticks > ticks_remaining) {
+            ticks_remaining = 0;
+            return;
+        }
+        ticks -= ticks_remaining;
+    }
+    virtual u64 GetTicksRemaining() override {
+        return ticks_remaining;
+    }
+
+    ARM_Dynarmic& parent;
+    size_t ticks_remaining = 0;
+    size_t num_interpreted_instructions = 0;
+    u64 tpidrr0_el0 = 0;
+};
+
+ARM_Dynarmic::ARM_Dynarmic()
+    : cb(std::make_unique<ARM_Dynarmic_Callbacks>(*this)),
+      jit(Dynarmic::A64::UserConfig{cb.get()}) {
+    ARM_Interface::ThreadContext ctx;
+    inner_unicorn.SaveContext(ctx);
+    LoadContext(ctx);
+}
+
+ARM_Dynarmic::~ARM_Dynarmic() = default;
+
+void ARM_Dynarmic::MapBackingMemory(u64 address, size_t size, u8* memory,
+                                    Kernel::VMAPermission perms) {
+    inner_unicorn.MapBackingMemory(address, size, memory, perms);
+}
+
+void ARM_Dynarmic::SetPC(u64 pc) {
+    jit.SetPC(pc);
 }
 
 u64 ARM_Dynarmic::GetPC() const {
-    UNIMPLEMENTED();
-    return {};
+    return jit.GetPC();
 }
 
-u64 ARM_Dynarmic::GetReg(int /*index*/) const {
-    UNIMPLEMENTED();
-    return {};
+u64 ARM_Dynarmic::GetReg(int index) const {
+    return jit.GetRegister(index);
 }
 
-void ARM_Dynarmic::SetReg(int /*index*/, u64 /*value*/) {
-    UNIMPLEMENTED();
+void ARM_Dynarmic::SetReg(int index, u64 value) {
+    jit.SetRegister(index, value);
 }
 
-const u128& ARM_Dynarmic::GetExtReg(int /*index*/) const {
-    UNIMPLEMENTED();
-    static constexpr u128 res{};
-    return res;
+u128 ARM_Dynarmic::GetExtReg(int index) const {
+    return jit.GetVector(index);
 }
 
-void ARM_Dynarmic::SetExtReg(int /*index*/, u128& /*value*/) {
-    UNIMPLEMENTED();
+void ARM_Dynarmic::SetExtReg(int index, u128 value) {
+    jit.SetVector(index, value);
 }
 
 u32 ARM_Dynarmic::GetVFPReg(int /*index*/) const {
@@ -51,41 +122,56 @@ void ARM_Dynarmic::SetVFPReg(int /*index*/, u32 /*value*/) {
 }
 
 u32 ARM_Dynarmic::GetCPSR() const {
-    UNIMPLEMENTED();
-    return {};
+    return jit.GetPstate();
 }
 
-void ARM_Dynarmic::SetCPSR(u32 /*cpsr*/) {
-    UNIMPLEMENTED();
+void ARM_Dynarmic::SetCPSR(u32 cpsr) {
+    jit.SetPstate(cpsr);
 }
 
-VAddr ARM_Dynarmic::GetTlsAddress() const {
-    UNIMPLEMENTED();
-    return {};
+u64 ARM_Dynarmic::GetTlsAddress() const {
+    return cb->tpidrr0_el0;
 }
 
-void ARM_Dynarmic::SetTlsAddress(VAddr /*address*/) {
-    UNIMPLEMENTED();
+void ARM_Dynarmic::SetTlsAddress(u64 address) {
+    cb->tpidrr0_el0 = address;
 }
 
-void ARM_Dynarmic::ExecuteInstructions(int /*num_instructions*/) {
-    UNIMPLEMENTED();
+void ARM_Dynarmic::ExecuteInstructions(int num_instructions) {
+    cb->ticks_remaining = num_instructions;
+    jit.Run();
+    CoreTiming::AddTicks(num_instructions - cb->num_interpreted_instructions);
+    cb->num_interpreted_instructions = 0;
 }
 
-void ARM_Dynarmic::SaveContext(ARM_Interface::ThreadContext& /*ctx*/) {
-    UNIMPLEMENTED();
+void ARM_Dynarmic::SaveContext(ARM_Interface::ThreadContext& ctx) {
+    ctx.cpu_registers = jit.GetRegisters();
+    ctx.sp = jit.GetSP();
+    ctx.pc = jit.GetPC();
+    ctx.cpsr = jit.GetPstate();
+    ctx.fpu_registers = jit.GetVectors();
+    ctx.fpscr = jit.GetFpcr();
+    ctx.tls_address = cb->tpidrr0_el0;
 }
 
-void ARM_Dynarmic::LoadContext(const ARM_Interface::ThreadContext& /*ctx*/) {
-    UNIMPLEMENTED();
+void ARM_Dynarmic::LoadContext(const ARM_Interface::ThreadContext& ctx) {
+    jit.SetRegisters(ctx.cpu_registers);
+    jit.SetSP(ctx.sp);
+    jit.SetPC(ctx.pc);
+    jit.SetPstate(ctx.cpsr);
+    jit.SetVectors(ctx.fpu_registers);
+    jit.SetFpcr(ctx.fpscr);
+    cb->tpidrr0_el0 = ctx.tls_address;
 }
 
 void ARM_Dynarmic::PrepareReschedule() {
-    UNIMPLEMENTED();
+    if (jit.IsExecuting()) {
+        jit.HaltExecution();
+    }
 }
 
 void ARM_Dynarmic::ClearInstructionCache() {
-    UNIMPLEMENTED();
+    jit.ClearCache();
 }
 
 void ARM_Dynarmic::PageTableChanged() {
diff --git a/src/core/arm/dynarmic/arm_dynarmic.h b/src/core/arm/dynarmic/arm_dynarmic.h
index c62d02ec9..947fab3f4 100644
--- a/src/core/arm/dynarmic/arm_dynarmic.h
+++ b/src/core/arm/dynarmic/arm_dynarmic.h
@@ -4,12 +4,18 @@
 
 #pragma once
 
+#include <memory>
+#include <dynarmic/A64/a64.h>
 #include "common/common_types.h"
 #include "core/arm/arm_interface.h"
+#include "core/arm/unicorn/arm_unicorn.h"
+
+class ARM_Dynarmic_Callbacks;
 
 class ARM_Dynarmic final : public ARM_Interface {
 public:
     ARM_Dynarmic();
+    ~ARM_Dynarmic();
 
     void MapBackingMemory(VAddr address, size_t size, u8* memory,
                           Kernel::VMAPermission perms) override;
@@ -18,8 +24,8 @@ public:
     u64 GetPC() const override;
     u64 GetReg(int index) const override;
     void SetReg(int index, u64 value) override;
-    const u128& GetExtReg(int index) const override;
-    void SetExtReg(int index, u128& value) override;
+    u128 GetExtReg(int index) const override;
+    void SetExtReg(int index, u128 value) override;
     u32 GetVFPReg(int index) const override;
     void SetVFPReg(int index, u32 value) override;
     u32 GetCPSR() const override;
@@ -35,4 +41,10 @@ public:
 
     void ClearInstructionCache() override;
     void PageTableChanged() override;
+
+private:
+    friend class ARM_Dynarmic_Callbacks;
+    std::unique_ptr<ARM_Dynarmic_Callbacks> cb;
+    Dynarmic::A64::Jit jit;
+    ARM_Unicorn inner_unicorn;
 };
diff --git a/src/core/arm/unicorn/arm_unicorn.cpp b/src/core/arm/unicorn/arm_unicorn.cpp
index 1d15152e5..dc886a4f7 100644
--- a/src/core/arm/unicorn/arm_unicorn.cpp
+++ b/src/core/arm/unicorn/arm_unicorn.cpp
@@ -108,13 +108,13 @@ void ARM_Unicorn::SetReg(int regn, u64 val) {
     CHECKED(uc_reg_write(uc, treg, &val));
 }
 
-const u128& ARM_Unicorn::GetExtReg(int /*index*/) const {
+u128 ARM_Unicorn::GetExtReg(int /*index*/) const {
     UNIMPLEMENTED();
     static constexpr u128 res{};
     return res;
 }
 
-void ARM_Unicorn::SetExtReg(int /*index*/, u128& /*value*/) {
+void ARM_Unicorn::SetExtReg(int /*index*/, u128 /*value*/) {
     UNIMPLEMENTED();
 }
 
@@ -168,10 +168,12 @@ void ARM_Unicorn::SaveContext(ARM_Interface::ThreadContext& ctx) {
         uregs[i] = UC_ARM64_REG_X0 + i;
         tregs[i] = &ctx.cpu_registers[i];
     }
+    uregs[29] = UC_ARM64_REG_X29;
+    tregs[29] = (void*)&ctx.cpu_registers[29];
+    uregs[30] = UC_ARM64_REG_X30;
+    tregs[30] = (void*)&ctx.cpu_registers[30];
 
-    CHECKED(uc_reg_read_batch(uc, uregs, tregs, 29));
-    CHECKED(uc_reg_read(uc, UC_ARM64_REG_X29, &ctx.cpu_registers[29]));
-    CHECKED(uc_reg_read(uc, UC_ARM64_REG_X30, &ctx.lr));
+    CHECKED(uc_reg_read_batch(uc, uregs, tregs, 31));
 
     ctx.tls_address = GetTlsAddress();
 
@@ -195,10 +197,12 @@ void ARM_Unicorn::LoadContext(const ARM_Interface::ThreadContext& ctx) {
         uregs[i] = UC_ARM64_REG_X0 + i;
         tregs[i] = (void*)&ctx.cpu_registers[i];
     }
+    uregs[29] = UC_ARM64_REG_X29;
+    tregs[29] = (void*)&ctx.cpu_registers[29];
+    uregs[30] = UC_ARM64_REG_X30;
+    tregs[30] = (void*)&ctx.cpu_registers[30];
 
-    CHECKED(uc_reg_write_batch(uc, uregs, tregs, 29));
-    CHECKED(uc_reg_write(uc, UC_ARM64_REG_X29, &ctx.cpu_registers[29]));
-    CHECKED(uc_reg_write(uc, UC_ARM64_REG_X30, &ctx.lr));
+    CHECKED(uc_reg_write_batch(uc, uregs, tregs, 31));
 
     SetTlsAddress(ctx.tls_address);
 
diff --git a/src/core/arm/unicorn/arm_unicorn.h b/src/core/arm/unicorn/arm_unicorn.h
index 90db3eb7f..81cce0721 100644
--- a/src/core/arm/unicorn/arm_unicorn.h
+++ b/src/core/arm/unicorn/arm_unicorn.h
@@ -9,7 +9,6 @@
 #include "core/arm/arm_interface.h"
 
 class ARM_Unicorn final : public ARM_Interface {
-
 public:
     ARM_Unicorn();
     ~ARM_Unicorn();
@@ -19,8 +18,8 @@ public:
     u64 GetPC() const override;
     u64 GetReg(int index) const override;
     void SetReg(int index, u64 value) override;
-    const u128& GetExtReg(int index) const override;
-    void SetExtReg(int index, u128& value) override;
+    u128 GetExtReg(int index) const override;
+    void SetExtReg(int index, u128 value) override;
     u32 GetVFPReg(int index) const override;
     void SetVFPReg(int index, u32 value) override;
     u32 GetCPSR() const override;
diff --git a/src/core/core.cpp b/src/core/core.cpp
index 40ef58f59..0240f946b 100644
--- a/src/core/core.cpp
+++ b/src/core/core.cpp
@@ -6,6 +6,7 @@
 #include <utility>
 #include "audio_core/audio_core.h"
 #include "common/logging/log.h"
+#include "core/arm/dynarmic/arm_dynarmic.h"
 #include "core/arm/unicorn/arm_unicorn.h"
 #include "core/core.h"
 #include "core/core_timing.h"
@@ -139,7 +140,8 @@ void System::Reschedule() {
 System::ResultStatus System::Init(EmuWindow* emu_window, u32 system_mode) {
     LOG_DEBUG(HW_Memory, "initialized OK");
 
-    cpu_core = std::make_unique<ARM_Unicorn>();
+    // TODO: Configuration option
+    cpu_core = std::make_unique<ARM_Dynarmic>();
     telemetry_session = std::make_unique<Core::TelemetrySession>();
 
     CoreTiming::Init();