Merge pull request #5184 from MerryMage/dyn-update-2020-04

Update dynarmic (Apr 2020)

.travis.yml

@@ -32,7 +32,7 @@ matrix:
     - os: osx
       env: NAME="macos build"
       sudo: false
-      osx_image: xcode10
+      osx_image: xcode10.2
       install: "./.travis/macos/deps.sh"
       script: "./.travis/macos/build.sh"
       after_success: "./.travis/macos/upload.sh"

externals/dynarmic

@@ -1 +1 @@
-Subproject commit 4e6848d1c9e8dadc70595c15b5589f8b14aad478
+Subproject commit b58048a5a88ad6184d64f16cfd2c5d63a1952e77

src/citra_qt/debugger/registers.cpp

@@ -103,8 +103,6 @@ void RegistersWidget::OnEmulationStopping() {
 
     vfp_system_registers->child(0)->setText(1, QString{});
     vfp_system_registers->child(1)->setText(1, QString{});
-    vfp_system_registers->child(2)->setText(1, QString{});
-    vfp_system_registers->child(3)->setText(1, QString{});
 
     setEnabled(false);
 }
@@ -188,16 +186,12 @@ void RegistersWidget::CreateVFPSystemRegisterChildren() {
 
     vfp_system_registers->addChild(fpscr);
     vfp_system_registers->addChild(fpexc);
-    vfp_system_registers->addChild(new QTreeWidgetItem(QStringList(QStringLiteral("FPINST"))));
-    vfp_system_registers->addChild(new QTreeWidgetItem(QStringList(QStringLiteral("FPINST2"))));
 }
 
 void RegistersWidget::UpdateVFPSystemRegisterValues() {
     // Todo: handle all cores
     const u32 fpscr_val = Core::GetCore(0).GetVFPSystemReg(VFP_FPSCR);
     const u32 fpexc_val = Core::GetCore(0).GetVFPSystemReg(VFP_FPEXC);
-    const u32 fpinst_val = Core::GetCore(0).GetVFPSystemReg(VFP_FPINST);
-    const u32 fpinst2_val = Core::GetCore(0).GetVFPSystemReg(VFP_FPINST2);
 
     QTreeWidgetItem* const fpscr = vfp_system_registers->child(0);
     fpscr->setText(1, QStringLiteral("0x%1").arg(fpscr_val, 8, 16, QLatin1Char('0')));
@@ -237,9 +231,4 @@ void RegistersWidget::UpdateVFPSystemRegisterValues() {
     fpexc->child(5)->setText(1, QString::number((fpexc_val >> 28) & 1));
     fpexc->child(6)->setText(1, QString::number((fpexc_val >> 30) & 1));
     fpexc->child(7)->setText(1, QString::number((fpexc_val >> 31) & 1));
-
-    vfp_system_registers->child(2)->setText(
-        1, QStringLiteral("0x%1").arg(fpinst_val, 8, 16, QLatin1Char('0')));
-    vfp_system_registers->child(3)->setText(
-        1, QStringLiteral("0x%1").arg(fpinst2_val, 8, 16, QLatin1Char('0')));
 }

src/common/ring_buffer.h

@@ -9,6 +9,7 @@
 #include <atomic>
 #include <cstddef>
 #include <cstring>
+#include <new>
 #include <type_traits>
 #include <vector>
 #include "common/common_types.h"
@@ -29,7 +30,7 @@ class RingBuffer {
     static_assert(capacity < std::numeric_limits<std::size_t>::max() / 2 / granularity);
     static_assert((capacity & (capacity - 1)) == 0, "capacity must be a power of two");
     // Ensure lock-free.
-    static_assert(std::atomic<std::size_t>::is_always_lock_free);
+    static_assert(std::atomic_size_t::is_always_lock_free);
 
 public:
     /// Pushes slots into the ring buffer
@@ -100,10 +101,22 @@ public:
     }
 
 private:
-    // It is important to align the below variables for performance reasons:
+    // It is important to separate the below atomics for performance reasons:
     // Having them on the same cache-line would result in false-sharing between them.
-    alignas(128) std::atomic<std::size_t> m_read_index{0};
+    // TODO: Remove this ifdef whenever clang and GCC support
-    alignas(128) std::atomic<std::size_t> m_write_index{0};
+    //       std::hardware_destructive_interference_size.
+#if defined(_MSC_VER) && _MSC_VER >= 1911
+    static constexpr std::size_t padding_size =
+        std::hardware_destructive_interference_size - sizeof(std::atomic_size_t);
+#else
+    static constexpr std::size_t padding_size = 128 - sizeof(std::atomic_size_t);
+#endif
+
+    std::atomic_size_t m_read_index{0};
+    char padding1[padding_size];
+
+    std::atomic_size_t m_write_index{0};
+    char padding2[padding_size];
 
     std::array<T, granularity * capacity> m_data;
 };

src/core/arm/dynarmic/arm_dynarmic.cpp

@@ -9,7 +9,6 @@
 #include "common/microprofile.h"
 #include "core/arm/dynarmic/arm_dynarmic.h"
 #include "core/arm/dynarmic/arm_dynarmic_cp15.h"
-#include "core/arm/dyncom/arm_dyncom_interpreter.h"
 #include "core/core.h"
 #include "core/core_timing.h"
 #include "core/gdbstub/gdbstub.h"
@@ -102,24 +101,9 @@ public:
     }
 
     void InterpreterFallback(VAddr pc, std::size_t num_instructions) override {
-        parent.interpreter_state->Reg = parent.jit->Regs();
+        // Should never happen.
-        parent.interpreter_state->Cpsr = parent.jit->Cpsr();
+        UNREACHABLE_MSG("InterpeterFallback reached with pc = 0x{:08x}, code = 0x{:08x}, num = {}",
-        parent.interpreter_state->Reg[15] = pc;
+                        pc, MemoryReadCode(pc), num_instructions);
-        parent.interpreter_state->ExtReg = parent.jit->ExtRegs();
-        parent.interpreter_state->VFP[VFP_FPSCR] = parent.jit->Fpscr();
-        parent.interpreter_state->NumInstrsToExecute = num_instructions;
-
-        InterpreterMainLoop(parent.interpreter_state.get());
-
-        bool is_thumb = (parent.interpreter_state->Cpsr & (1 << 5)) != 0;
-        parent.interpreter_state->Reg[15] &= (is_thumb ? 0xFFFFFFFE : 0xFFFFFFFC);
-
-        parent.jit->Regs() = parent.interpreter_state->Reg;
-        parent.jit->SetCpsr(parent.interpreter_state->Cpsr);
-        parent.jit->ExtRegs() = parent.interpreter_state->ExtReg;
-        parent.jit->SetFpscr(parent.interpreter_state->VFP[VFP_FPSCR]);
-
-        parent.interpreter_state->ServeBreak();
     }
 
     void CallSVC(std::uint32_t swi) override {
@@ -135,14 +119,18 @@ public:
             if (GDBStub::IsConnected()) {
                 parent.jit->HaltExecution();
                 parent.SetPC(pc);
-                Kernel::Thread* thread =
+                parent.ServeBreak();
-                    parent.system.Kernel().GetCurrentThreadManager().GetCurrentThread();
-                parent.SaveContext(thread->context);
-                GDBStub::Break();
-                GDBStub::SendTrap(thread, 5);
                 return;
             }
             break;
+        case Dynarmic::A32::Exception::SendEvent:
+        case Dynarmic::A32::Exception::SendEventLocal:
+        case Dynarmic::A32::Exception::WaitForInterrupt:
+        case Dynarmic::A32::Exception::WaitForEvent:
+        case Dynarmic::A32::Exception::Yield:
+        case Dynarmic::A32::Exception::PreloadData:
+        case Dynarmic::A32::Exception::PreloadDataWithIntentToWrite:
+            return;
         }
         ASSERT_MSG(false, "ExceptionRaised(exception = {}, pc = {:08X}, code = {:08X})",
                    static_cast<std::size_t>(exception), pc, MemoryReadCode(pc));
@@ -161,12 +149,10 @@ public:
     Memory::MemorySystem& memory;
 };
 
-ARM_Dynarmic::ARM_Dynarmic(Core::System* system, Memory::MemorySystem& memory,
+ARM_Dynarmic::ARM_Dynarmic(Core::System* system, Memory::MemorySystem& memory, u32 id,
-                           PrivilegeMode initial_mode, u32 id,
                            std::shared_ptr<Core::Timing::Timer> timer)
     : ARM_Interface(id, timer), system(*system), memory(memory),
       cb(std::make_unique<DynarmicUserCallbacks>(*this)) {
-    interpreter_state = std::make_shared<ARMul_State>(system, memory, initial_mode);
     PageTableChanged();
 }
 
@@ -182,7 +168,11 @@ void ARM_Dynarmic::Run() {
 }
 
 void ARM_Dynarmic::Step() {
-    cb->InterpreterFallback(jit->Regs()[15], 1);
+    jit->Step();
+
+    if (GDBStub::IsConnected()) {
+        ServeBreak();
+    }
 }
 
 void ARM_Dynarmic::SetPC(u32 pc) {
@@ -210,21 +200,25 @@ void ARM_Dynarmic::SetVFPReg(int index, u32 value) {
 }
 
 u32 ARM_Dynarmic::GetVFPSystemReg(VFPSystemRegister reg) const {
-    if (reg == VFP_FPSCR) {
+    switch (reg) {
+    case VFP_FPSCR:
         return jit->Fpscr();
+    case VFP_FPEXC:
+        return fpexc;
     }
-
+    UNREACHABLE_MSG("Unknown VFP system register: {}", static_cast<size_t>(reg));
-    // Dynarmic does not implement and/or expose other VFP registers, fallback to interpreter state
-    return interpreter_state->VFP[reg];
 }
 
 void ARM_Dynarmic::SetVFPSystemReg(VFPSystemRegister reg, u32 value) {
-    if (reg == VFP_FPSCR) {
+    switch (reg) {
+    case VFP_FPSCR:
         jit->SetFpscr(value);
+        return;
+    case VFP_FPEXC:
+        fpexc = value;
+        return;
     }
-
+    UNREACHABLE_MSG("Unknown VFP system register: {}", static_cast<size_t>(reg));
-    // Dynarmic does not implement and/or expose other VFP registers, fallback to interpreter state
-    interpreter_state->VFP[reg] = value;
 }
 
 u32 ARM_Dynarmic::GetCPSR() const {
@@ -236,11 +230,25 @@ void ARM_Dynarmic::SetCPSR(u32 cpsr) {
 }
 
 u32 ARM_Dynarmic::GetCP15Register(CP15Register reg) {
-    return interpreter_state->CP15[reg];
+    switch (reg) {
+    case CP15_THREAD_UPRW:
+        return cp15_state.cp15_thread_uprw;
+    case CP15_THREAD_URO:
+        return cp15_state.cp15_thread_uro;
+    }
+    UNREACHABLE_MSG("Unknown CP15 register: {}", static_cast<size_t>(reg));
 }
 
 void ARM_Dynarmic::SetCP15Register(CP15Register reg, u32 value) {
-    interpreter_state->CP15[reg] = value;
+    switch (reg) {
+    case CP15_THREAD_UPRW:
+        cp15_state.cp15_thread_uprw = value;
+        return;
+    case CP15_THREAD_URO:
+        cp15_state.cp15_thread_uro = value;
+        return;
+    }
+    UNREACHABLE_MSG("Unknown CP15 register: {}", static_cast<size_t>(reg));
 }
 
 std::unique_ptr<ARM_Interface::ThreadContext> ARM_Dynarmic::NewContext() const {
@@ -252,7 +260,7 @@ void ARM_Dynarmic::SaveContext(const std::unique_ptr<ThreadContext>& arg) {
     ASSERT(ctx);
 
     jit->SaveContext(ctx->ctx);
-    ctx->fpexc = interpreter_state->VFP[VFP_FPEXC];
+    ctx->fpexc = fpexc;
 }
 
 void ARM_Dynarmic::LoadContext(const std::unique_ptr<ThreadContext>& arg) {
@@ -260,7 +268,7 @@ void ARM_Dynarmic::LoadContext(const std::unique_ptr<ThreadContext>& arg) {
     ASSERT(ctx);
 
     jit->LoadContext(ctx->ctx);
-    interpreter_state->VFP[VFP_FPEXC] = ctx->fpexc;
+    fpexc = ctx->fpexc;
 }
 
 void ARM_Dynarmic::PrepareReschedule() {
@@ -270,11 +278,9 @@ void ARM_Dynarmic::PrepareReschedule() {
 }
 
 void ARM_Dynarmic::ClearInstructionCache() {
-    // TODO: Clear interpreter cache when appropriate.
     for (const auto& j : jits) {
         j.second->ClearCache();
     }
-    interpreter_state->instruction_cache.clear();
 }
 
 void ARM_Dynarmic::InvalidateCacheRange(u32 start_address, std::size_t length) {
@@ -295,11 +301,18 @@ void ARM_Dynarmic::PageTableChanged() {
     jits.emplace(current_page_table, std::move(new_jit));
 }
 
+void ARM_Dynarmic::ServeBreak() {
+    Kernel::Thread* thread = system.Kernel().GetCurrentThreadManager().GetCurrentThread();
+    SaveContext(thread->context);
+    GDBStub::Break();
+    GDBStub::SendTrap(thread, 5);
+}
+
 std::unique_ptr<Dynarmic::A32::Jit> ARM_Dynarmic::MakeJit() {
     Dynarmic::A32::UserConfig config;
     config.callbacks = cb.get();
     config.page_table = &current_page_table->pointers;
-    config.coprocessors[15] = std::make_shared<DynarmicCP15>(interpreter_state);
+    config.coprocessors[15] = std::make_shared<DynarmicCP15>(cp15_state);
     config.define_unpredictable_behaviour = true;
     return std::make_unique<Dynarmic::A32::Jit>(config);
 }

src/core/arm/dynarmic/arm_dynarmic.h

@@ -9,7 +9,7 @@
 #include <dynarmic/A32/a32.h>
 #include "common/common_types.h"
 #include "core/arm/arm_interface.h"
-#include "core/arm/skyeye_common/armstate.h"
+#include "core/arm/dynarmic/arm_dynarmic_cp15.h"
 
 namespace Memory {
 struct PageTable;
@@ -24,8 +24,8 @@ class DynarmicUserCallbacks;
 
 class ARM_Dynarmic final : public ARM_Interface {
 public:
-    ARM_Dynarmic(Core::System* system, Memory::MemorySystem& memory, PrivilegeMode initial_mode,
+    ARM_Dynarmic(Core::System* system, Memory::MemorySystem& memory, u32 id,
-                 u32 id, std::shared_ptr<Core::Timing::Timer> timer);
+                 std::shared_ptr<Core::Timing::Timer> timer);
     ~ARM_Dynarmic() override;
 
     void Run() override;
@@ -55,14 +55,18 @@ public:
     void PageTableChanged() override;
 
 private:
+    void ServeBreak();
+
     friend class DynarmicUserCallbacks;
     Core::System& system;
     Memory::MemorySystem& memory;
     std::unique_ptr<DynarmicUserCallbacks> cb;
     std::unique_ptr<Dynarmic::A32::Jit> MakeJit();
 
+    u32 fpexc = 0;
+    CP15State cp15_state;
+
     Dynarmic::A32::Jit* jit = nullptr;
     Memory::PageTable* current_page_table = nullptr;
     std::map<Memory::PageTable*, std::unique_ptr<Dynarmic::A32::Jit>> jits;
-    std::shared_ptr<ARMul_State> interpreter_state;
 };

src/core/arm/dynarmic/arm_dynarmic_cp15.cpp

@@ -10,14 +10,14 @@ using Callback = Dynarmic::A32::Coprocessor::Callback;
 using CallbackOrAccessOneWord = Dynarmic::A32::Coprocessor::CallbackOrAccessOneWord;
 using CallbackOrAccessTwoWords = Dynarmic::A32::Coprocessor::CallbackOrAccessTwoWords;
 
-DynarmicCP15::DynarmicCP15(const std::shared_ptr<ARMul_State>& state) : interpreter_state(state) {}
+DynarmicCP15::DynarmicCP15(CP15State& state) : state(state) {}
 
 DynarmicCP15::~DynarmicCP15() = default;
 
-boost::optional<Callback> DynarmicCP15::CompileInternalOperation(bool two, unsigned opc1,
+std::optional<Callback> DynarmicCP15::CompileInternalOperation(bool two, unsigned opc1,
-                                                                 CoprocReg CRd, CoprocReg CRn,
+                                                               CoprocReg CRd, CoprocReg CRn,
-                                                                 CoprocReg CRm, unsigned opc2) {
+                                                               CoprocReg CRm, unsigned opc2) {
-    return boost::none;
+    return std::nullopt;
 }
 
 CallbackOrAccessOneWord DynarmicCP15::CompileSendOneWord(bool two, unsigned opc1, CoprocReg CRn,
@@ -26,31 +26,31 @@ CallbackOrAccessOneWord DynarmicCP15::CompileSendOneWord(bool two, unsigned opc1
 
     if (!two && CRn == CoprocReg::C7 && opc1 == 0 && CRm == CoprocReg::C5 && opc2 == 4) {
         // This is a dummy write, we ignore the value written here.
-        return &interpreter_state->CP15[CP15_FLUSH_PREFETCH_BUFFER];
+        return &state.cp15_flush_prefetch_buffer;
     }
 
     if (!two && CRn == CoprocReg::C7 && opc1 == 0 && CRm == CoprocReg::C10) {
         switch (opc2) {
         case 4:
             // This is a dummy write, we ignore the value written here.
-            return &interpreter_state->CP15[CP15_DATA_SYNC_BARRIER];
+            return &state.cp15_data_sync_barrier;
         case 5:
             // This is a dummy write, we ignore the value written here.
-            return &interpreter_state->CP15[CP15_DATA_MEMORY_BARRIER];
+            return &state.cp15_data_memory_barrier;
         default:
-            return boost::blank{};
+            return std::monostate{};
         }
     }
 
     if (!two && CRn == CoprocReg::C13 && opc1 == 0 && CRm == CoprocReg::C0 && opc2 == 2) {
-        return &interpreter_state->CP15[CP15_THREAD_UPRW];
+        return &state.cp15_thread_uprw;
     }
 
-    return boost::blank{};
+    return std::monostate{};
 }
 
 CallbackOrAccessTwoWords DynarmicCP15::CompileSendTwoWords(bool two, unsigned opc, CoprocReg CRm) {
-    return boost::blank{};
+    return std::monostate{};
 }
 
 CallbackOrAccessOneWord DynarmicCP15::CompileGetOneWord(bool two, unsigned opc1, CoprocReg CRn,
@@ -60,29 +60,27 @@ CallbackOrAccessOneWord DynarmicCP15::CompileGetOneWord(bool two, unsigned opc1,
     if (!two && CRn == CoprocReg::C13 && opc1 == 0 && CRm == CoprocReg::C0) {
         switch (opc2) {
         case 2:
-            return &interpreter_state->CP15[CP15_THREAD_UPRW];
+            return &state.cp15_thread_uprw;
         case 3:
-            return &interpreter_state->CP15[CP15_THREAD_URO];
+            return &state.cp15_thread_uro;
         default:
-            return boost::blank{};
+            return std::monostate{};
         }
     }
 
-    return boost::blank{};
+    return std::monostate{};
 }
 
 CallbackOrAccessTwoWords DynarmicCP15::CompileGetTwoWords(bool two, unsigned opc, CoprocReg CRm) {
-    return boost::blank{};
+    return std::monostate{};
 }
 
-boost::optional<Callback> DynarmicCP15::CompileLoadWords(bool two, bool long_transfer,
+std::optional<Callback> DynarmicCP15::CompileLoadWords(bool two, bool long_transfer, CoprocReg CRd,
-                                                         CoprocReg CRd,
+                                                       std::optional<u8> option) {
-                                                         boost::optional<u8> option) {
+    return std::nullopt;
-    return boost::none;
 }
 
-boost::optional<Callback> DynarmicCP15::CompileStoreWords(bool two, bool long_transfer,
+std::optional<Callback> DynarmicCP15::CompileStoreWords(bool two, bool long_transfer, CoprocReg CRd,
-                                                          CoprocReg CRd,
+                                                        std::optional<u8> option) {
-                                                          boost::optional<u8> option) {
+    return std::nullopt;
-    return boost::none;
 }

src/core/arm/dynarmic/arm_dynarmic_cp15.h

@@ -8,29 +8,35 @@
 #include <dynarmic/A32/coprocessor.h>
 #include "common/common_types.h"
 
-struct ARMul_State;
+struct CP15State {
+    u32 cp15_thread_uprw = 0;
+    u32 cp15_thread_uro = 0;
+    u32 cp15_flush_prefetch_buffer = 0; ///< dummy value
+    u32 cp15_data_sync_barrier = 0;     ///< dummy value
+    u32 cp15_data_memory_barrier = 0;   ///< dummy value
+};
 
 class DynarmicCP15 final : public Dynarmic::A32::Coprocessor {
 public:
     using CoprocReg = Dynarmic::A32::CoprocReg;
 
-    explicit DynarmicCP15(const std::shared_ptr<ARMul_State>&);
+    explicit DynarmicCP15(CP15State&);
     ~DynarmicCP15() override;
 
-    boost::optional<Callback> CompileInternalOperation(bool two, unsigned opc1, CoprocReg CRd,
+    std::optional<Callback> CompileInternalOperation(bool two, unsigned opc1, CoprocReg CRd,
-                                                       CoprocReg CRn, CoprocReg CRm,
+                                                     CoprocReg CRn, CoprocReg CRm,
-                                                       unsigned opc2) override;
+                                                     unsigned opc2) override;
     CallbackOrAccessOneWord CompileSendOneWord(bool two, unsigned opc1, CoprocReg CRn,
                                                CoprocReg CRm, unsigned opc2) override;
     CallbackOrAccessTwoWords CompileSendTwoWords(bool two, unsigned opc, CoprocReg CRm) override;
     CallbackOrAccessOneWord CompileGetOneWord(bool two, unsigned opc1, CoprocReg CRn, CoprocReg CRm,
                                               unsigned opc2) override;
     CallbackOrAccessTwoWords CompileGetTwoWords(bool two, unsigned opc, CoprocReg CRm) override;
-    boost::optional<Callback> CompileLoadWords(bool two, bool long_transfer, CoprocReg CRd,
+    std::optional<Callback> CompileLoadWords(bool two, bool long_transfer, CoprocReg CRd,
-                                               boost::optional<u8> option) override;
+                                             std::optional<u8> option) override;
-    boost::optional<Callback> CompileStoreWords(bool two, bool long_transfer, CoprocReg CRd,
+    std::optional<Callback> CompileStoreWords(bool two, bool long_transfer, CoprocReg CRd,
-                                                boost::optional<u8> option) override;
+                                              std::optional<u8> option) override;
 
 private:
-    std::shared_ptr<ARMul_State> interpreter_state;
+    CP15State& state;
 };

src/core/core.cpp

@@ -267,7 +267,7 @@ System::ResultStatus System::Init(Frontend::EmuWindow& emu_window, u32 system_mo
 #ifdef ARCHITECTURE_x86_64
         for (std::size_t i = 0; i < num_cores; ++i) {
             cpu_cores.push_back(
-                std::make_shared<ARM_Dynarmic>(this, *memory, USER32MODE, i, timing->GetTimer(i)));
+                std::make_shared<ARM_Dynarmic>(this, *memory, i, timing->GetTimer(i)));
         }
 #else
         for (std::size_t i = 0; i < num_cores; ++i) {