1 files changed, 131 insertions, 76 deletions

diff --git a/src/core/arm/dynarmic/arm_dynarmic_64.cpp b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
index d96226c41..d8d3a38db 100644
--- a/src/core/arm/dynarmic/arm_dynarmic_64.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic_64.cpp
@@ -1,6 +1,5 @@
-// Copyright 2018 yuzu emulator team
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
+// SPDX-FileCopyrightText: Copyright 2018 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
 
 #include <cinttypes>
 #include <memory>
@@ -26,6 +25,9 @@ namespace Core {
 using Vector = Dynarmic::A64::Vector;
 using namespace Common::Literals;
 
+constexpr Dynarmic::HaltReason break_loop = Dynarmic::HaltReason::UserDefined2;
+constexpr Dynarmic::HaltReason svc_call = Dynarmic::HaltReason::UserDefined3;
+
 class DynarmicCallbacks64 : public Dynarmic::A64::UserCallbacks {
 public:
     explicit DynarmicCallbacks64(ARM_Dynarmic_64& parent_)
@@ -81,6 +83,7 @@ public:
     }
 
     void InterpreterFallback(u64 pc, std::size_t num_instructions) override {
+        parent.LogBacktrace();
         LOG_ERROR(Core_ARM,
                   "Unimplemented instruction @ 0x{:X} for {} instructions (instr = {:08X})", pc,
                   num_instructions, MemoryReadCode(pc));
@@ -93,17 +96,19 @@ public:
             static constexpr u64 ICACHE_LINE_SIZE = 64;
 
             const u64 cache_line_start = value & ~(ICACHE_LINE_SIZE - 1);
-            parent.InvalidateCacheRange(cache_line_start, ICACHE_LINE_SIZE);
+            parent.system.InvalidateCpuInstructionCacheRange(cache_line_start, ICACHE_LINE_SIZE);
             break;
         }
         case Dynarmic::A64::InstructionCacheOperation::InvalidateAllToPoU:
-            parent.ClearInstructionCache();
+            parent.system.InvalidateCpuInstructionCaches();
             break;
         case Dynarmic::A64::InstructionCacheOperation::InvalidateAllToPoUInnerSharable:
         default:
             LOG_DEBUG(Core_ARM, "Unprocesseed instruction cache operation: {}", op);
             break;
         }
+
+        parent.jit.load()->HaltExecution(Dynarmic::HaltReason::CacheInvalidation);
     }
 
     void ExceptionRaised(u64 pc, Dynarmic::A64::Exception exception) override {
@@ -116,21 +121,19 @@ public:
             return;
         case Dynarmic::A64::Exception::Breakpoint:
         default:
+            parent.LogBacktrace();
             ASSERT_MSG(false, "ExceptionRaised(exception = {}, pc = {:08X}, code = {:08X})",
                        static_cast<std::size_t>(exception), pc, MemoryReadCode(pc));
         }
     }
 
     void CallSVC(u32 swi) override {
-        parent.svc_called = true;
         parent.svc_swi = swi;
-        parent.jit->HaltExecution();
+        parent.jit.load()->HaltExecution(svc_call);
     }
 
     void AddTicks(u64 ticks) override {
-        if (parent.uses_wall_clock) {
-            return;
-        }
+        ASSERT_MSG(!parent.uses_wall_clock, "This should never happen - dynarmic ticking disabled");
 
         // Divide the number of ticks by the amount of CPU cores. TODO(Subv): This yields only a
         // rough approximation of the amount of executed ticks in the system, it may be thrown off
@@ -145,12 +148,8 @@ public:
     }
 
     u64 GetTicksRemaining() override {
-        if (parent.uses_wall_clock) {
-            if (!parent.interrupt_handlers[parent.core_index].IsInterrupted()) {
-                return minimum_run_cycles;
-            }
-            return 0U;
-        }
+        ASSERT_MSG(!parent.uses_wall_clock, "This should never happen - dynarmic ticking disabled");
+
         return std::max<s64>(parent.system.CoreTiming().GetDowncount(), 0);
     }
 
@@ -206,11 +205,19 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
 
     // Timing
     config.wall_clock_cntpct = uses_wall_clock;
+    config.enable_cycle_counting = !uses_wall_clock;
 
     // Code cache size
     config.code_cache_size = 512_MiB;
     config.far_code_offset = 400_MiB;
 
+    // null_jit
+    if (!page_table) {
+        // Don't waste too much memory on null_jit
+        config.code_cache_size = 8_MiB;
+        config.far_code_offset = 4_MiB;
+    }
+
     // Safe optimizations
     if (Settings::values.cpu_debug_mode) {
         if (!Settings::values.cpuopt_page_tables) {
@@ -246,35 +253,41 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
         if (!Settings::values.cpuopt_recompile_exclusives) {
             config.recompile_on_exclusive_fastmem_failure = false;
         }
-    }
+    } else {
+        // Unsafe optimizations
+        if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::Unsafe) {
+            config.unsafe_optimizations = true;
+            if (Settings::values.cpuopt_unsafe_unfuse_fma) {
+                config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_UnfuseFMA;
+            }
+            if (Settings::values.cpuopt_unsafe_reduce_fp_error) {
+                config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_ReducedErrorFP;
+            }
+            if (Settings::values.cpuopt_unsafe_inaccurate_nan) {
+                config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_InaccurateNaN;
+            }
+            if (Settings::values.cpuopt_unsafe_fastmem_check) {
+                config.fastmem_address_space_bits = 64;
+            }
+            if (Settings::values.cpuopt_unsafe_ignore_global_monitor) {
+                config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_IgnoreGlobalMonitor;
+            }
+        }
 
-    // Unsafe optimizations
-    if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::Unsafe) {
-        config.unsafe_optimizations = true;
-        if (Settings::values.cpuopt_unsafe_unfuse_fma) {
+        // Curated optimizations
+        if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::Auto) {
+            config.unsafe_optimizations = true;
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_UnfuseFMA;
-        }
-        if (Settings::values.cpuopt_unsafe_reduce_fp_error) {
-            config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_ReducedErrorFP;
-        }
-        if (Settings::values.cpuopt_unsafe_inaccurate_nan) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_InaccurateNaN;
-        }
-        if (Settings::values.cpuopt_unsafe_fastmem_check) {
             config.fastmem_address_space_bits = 64;
-        }
-        if (Settings::values.cpuopt_unsafe_ignore_global_monitor) {
             config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_IgnoreGlobalMonitor;
         }
-    }
 
-    // Curated optimizations
-    if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::Auto) {
-        config.unsafe_optimizations = true;
-        config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_UnfuseFMA;
-        config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_InaccurateNaN;
-        config.fastmem_address_space_bits = 64;
-        config.optimizations |= Dynarmic::OptimizationFlag::Unsafe_IgnoreGlobalMonitor;
+        // Paranoia mode for debugging optimizations
+        if (Settings::values.cpu_accuracy.GetValue() == Settings::CPUAccuracy::Paranoid) {
+            config.unsafe_optimizations = false;
+            config.optimizations = Dynarmic::no_optimizations;
+        }
     }
 
     return std::make_shared<Dynarmic::A64::Jit>(config);
@@ -282,20 +295,18 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
 
 void ARM_Dynarmic_64::Run() {
     while (true) {
-        jit->Run();
-        if (!svc_called) {
-            break;
+        const auto hr = jit.load()->Run();
+        if (Has(hr, svc_call)) {
+            Kernel::Svc::Call(system, svc_swi);
         }
-        svc_called = false;
-        Kernel::Svc::Call(system, svc_swi);
-        if (shutdown) {
+        if (Has(hr, break_loop) || !uses_wall_clock) {
             break;
         }
     }
 }
 
 void ARM_Dynarmic_64::Step() {
-    jit->Step();
+    jit.load()->Step();
 }
 
 ARM_Dynarmic_64::ARM_Dynarmic_64(System& system_, CPUInterrupts& interrupt_handlers_,
@@ -304,40 +315,44 @@ ARM_Dynarmic_64::ARM_Dynarmic_64(System& system_, CPUInterrupts& interrupt_handl
     : ARM_Interface{system_, interrupt_handlers_, uses_wall_clock_},
       cb(std::make_unique<DynarmicCallbacks64>(*this)), core_index{core_index_},
       exclusive_monitor{dynamic_cast<DynarmicExclusiveMonitor&>(exclusive_monitor_)},
-      jit(MakeJit(nullptr, 48)) {}
+      null_jit{MakeJit(nullptr, 48)}, jit{null_jit.get()} {}
 
 ARM_Dynarmic_64::~ARM_Dynarmic_64() = default;
 
 void ARM_Dynarmic_64::SetPC(u64 pc) {
-    jit->SetPC(pc);
+    jit.load()->SetPC(pc);
 }
 
 u64 ARM_Dynarmic_64::GetPC() const {
-    return jit->GetPC();
+    return jit.load()->GetPC();
+}
+
+u64 ARM_Dynarmic_64::GetSP() const {
+    return jit.load()->GetSP();
 }
 
 u64 ARM_Dynarmic_64::GetReg(int index) const {
-    return jit->GetRegister(index);
+    return jit.load()->GetRegister(index);
 }
 
 void ARM_Dynarmic_64::SetReg(int index, u64 value) {
-    jit->SetRegister(index, value);
+    jit.load()->SetRegister(index, value);
 }
 
 u128 ARM_Dynarmic_64::GetVectorReg(int index) const {
-    return jit->GetVector(index);
+    return jit.load()->GetVector(index);
 }
 
 void ARM_Dynarmic_64::SetVectorReg(int index, u128 value) {
-    jit->SetVector(index, value);
+    jit.load()->SetVector(index, value);
 }
 
 u32 ARM_Dynarmic_64::GetPSTATE() const {
-    return jit->GetPstate();
+    return jit.load()->GetPstate();
 }
 
 void ARM_Dynarmic_64::SetPSTATE(u32 pstate) {
-    jit->SetPstate(pstate);
+    jit.load()->SetPstate(pstate);
 }
 
 u64 ARM_Dynarmic_64::GetTlsAddress() const {
@@ -357,42 +372,47 @@ void ARM_Dynarmic_64::SetTPIDR_EL0(u64 value) {
 }
 
 void ARM_Dynarmic_64::SaveContext(ThreadContext64& ctx) {
-    ctx.cpu_registers = jit->GetRegisters();
-    ctx.sp = jit->GetSP();
-    ctx.pc = jit->GetPC();
-    ctx.pstate = jit->GetPstate();
-    ctx.vector_registers = jit->GetVectors();
-    ctx.fpcr = jit->GetFpcr();
-    ctx.fpsr = jit->GetFpsr();
+    Dynarmic::A64::Jit* j = jit.load();
+    ctx.cpu_registers = j->GetRegisters();
+    ctx.sp = j->GetSP();
+    ctx.pc = j->GetPC();
+    ctx.pstate = j->GetPstate();
+    ctx.vector_registers = j->GetVectors();
+    ctx.fpcr = j->GetFpcr();
+    ctx.fpsr = j->GetFpsr();
     ctx.tpidr = cb->tpidr_el0;
 }
 
 void ARM_Dynarmic_64::LoadContext(const ThreadContext64& ctx) {
-    jit->SetRegisters(ctx.cpu_registers);
-    jit->SetSP(ctx.sp);
-    jit->SetPC(ctx.pc);
-    jit->SetPstate(ctx.pstate);
-    jit->SetVectors(ctx.vector_registers);
-    jit->SetFpcr(ctx.fpcr);
-    jit->SetFpsr(ctx.fpsr);
+    Dynarmic::A64::Jit* j = jit.load();
+    j->SetRegisters(ctx.cpu_registers);
+    j->SetSP(ctx.sp);
+    j->SetPC(ctx.pc);
+    j->SetPstate(ctx.pstate);
+    j->SetVectors(ctx.vector_registers);
+    j->SetFpcr(ctx.fpcr);
+    j->SetFpsr(ctx.fpsr);
     SetTPIDR_EL0(ctx.tpidr);
 }
 
 void ARM_Dynarmic_64::PrepareReschedule() {
-    jit->HaltExecution();
-    shutdown = true;
+    jit.load()->HaltExecution(break_loop);
+}
+
+void ARM_Dynarmic_64::SignalInterrupt() {
+    jit.load()->HaltExecution(break_loop);
 }
 
 void ARM_Dynarmic_64::ClearInstructionCache() {
-    jit->ClearCache();
+    jit.load()->ClearCache();
 }
 
 void ARM_Dynarmic_64::InvalidateCacheRange(VAddr addr, std::size_t size) {
-    jit->InvalidateCacheRange(addr, size);
+    jit.load()->InvalidateCacheRange(addr, size);
 }
 
 void ARM_Dynarmic_64::ClearExclusiveState() {
-    jit->ClearExclusiveState();
+    jit.load()->ClearExclusiveState();
 }
 
 void ARM_Dynarmic_64::PageTableChanged(Common::PageTable& page_table,
@@ -403,13 +423,48 @@ void ARM_Dynarmic_64::PageTableChanged(Common::PageTable& page_table,
     auto key = std::make_pair(&page_table, new_address_space_size_in_bits);
     auto iter = jit_cache.find(key);
     if (iter != jit_cache.end()) {
-        jit = iter->second;
+        jit.store(iter->second.get());
         LoadContext(ctx);
         return;
     }
-    jit = MakeJit(&page_table, new_address_space_size_in_bits);
+    std::shared_ptr new_jit = MakeJit(&page_table, new_address_space_size_in_bits);
+    jit.store(new_jit.get());
     LoadContext(ctx);
-    jit_cache.emplace(key, jit);
+    jit_cache.emplace(key, std::move(new_jit));
+}
+
+std::vector<ARM_Interface::BacktraceEntry> ARM_Dynarmic_64::GetBacktrace(Core::System& system,
+                                                                         u64 fp, u64 lr) {
+    std::vector<BacktraceEntry> out;
+    auto& memory = system.Memory();
+
+    // fp (= r29) points to the last frame record.
+    // Note that this is the frame record for the *previous* frame, not the current one.
+    // Note we need to subtract 4 from our last read to get the proper address
+    // Frame records are two words long:
+    // fp+0 : pointer to previous frame record
+    // fp+8 : value of lr for frame
+    while (true) {
+        out.push_back({"", 0, lr, 0, ""});
+        if (!fp) {
+            break;
+        }
+        lr = memory.Read64(fp + 8) - 4;
+        fp = memory.Read64(fp);
+    }
+
+    SymbolicateBacktrace(system, out);
+
+    return out;
+}
+
+std::vector<ARM_Interface::BacktraceEntry> ARM_Dynarmic_64::GetBacktraceFromContext(
+    System& system, const ThreadContext64& ctx) {
+    return GetBacktrace(system, ctx.cpu_registers[29], ctx.cpu_registers[30]);
+}
+
+std::vector<ARM_Interface::BacktraceEntry> ARM_Dynarmic_64::GetBacktrace() const {
+    return GetBacktrace(system, GetReg(29), GetReg(30));
 }
 
 } // namespace Core