Fixing livelock caused by lf stuff (or at least making it harder to hit).

src/alloy/frontend/ppc/ppc_context.h

@@ -172,6 +172,10 @@ typedef struct alignas(64) PPCContext_s {
 
   // Reserve address for load acquire/store release. Shared.
   uint64_t* reserve_address;
+  uint64_t* reserve_value;
+
+  // Used to shuttle data into externs. Contents volatile.
+  uint64_t scratch;
 
   // Runtime-specific data pointer. Used on callbacks to get access to the
   // current runtime and its data.

src/alloy/frontend/ppc/ppc_emit_control.cc

@@ -607,22 +607,38 @@ XEEMITTER(mtspr, 0x7C0003A6, XFX)(PPCHIRBuilder& f, InstrData& i) {
   return 0;
 }
 
-// TODO(benvanik): MSR is used for toggling interrupts, and it'd be nice to
-//                 obey that setting. It's usually guarding atomic stores.
+// MSR is used for toggling interrupts (among other things).
+// We track it here for taking a global processor lock, as lots of lockfree
+// code requires it. Sequences of mtmsr/lwar/stcw/mtmsr come up a lot, and
+// without the lock here threads can livelock.
 
 XEEMITTER(mfmsr, 0x7C0000A6, X)(PPCHIRBuilder& f, InstrData& i) {
-  f.Nop();
+  f.StoreGPR(i.X.RT, f.LoadMSR());
   return 0;
 }
 
 XEEMITTER(mtmsr, 0x7C000124, X)(PPCHIRBuilder& f, InstrData& i) {
-  f.Nop();
+  if (i.X.RA & 0x01) {
+    // L = 1
+    f.StoreMSR(f.ZeroExtend(f.LoadGPR(i.X.RT), INT64_TYPE));
     return 0;
+  } else {
+    // L = 0
+    XEINSTRNOTIMPLEMENTED();
+    return 1;
+  }
 }
 
 XEEMITTER(mtmsrd, 0x7C000164, X)(PPCHIRBuilder& f, InstrData& i) {
-  f.Nop();
+  if (i.X.RA & 0x01) {
+    // L = 1
+    f.StoreMSR(f.LoadGPR(i.X.RT));
     return 0;
+  } else {
+    // L = 0
+    XEINSTRNOTIMPLEMENTED();
+    return 1;
+  }
 }
 
 void RegisterEmitCategoryControl() {

src/alloy/frontend/ppc/ppc_frontend.cc

@@ -13,6 +13,7 @@
 #include <alloy/frontend/ppc/ppc_disasm.h>
 #include <alloy/frontend/ppc/ppc_emit.h>
 #include <alloy/frontend/ppc/ppc_translator.h>
+#include <alloy/runtime/runtime.h>
 
 namespace alloy {
 namespace frontend {
@@ -58,12 +59,37 @@ PPCFrontend::~PPCFrontend() {
   translator_pool_.Reset();
 }
 
+void CheckGlobalLock(PPCContext* ppc_state, void* arg0, void* arg1) {
+  ppc_state->scratch = 0x8000;
+}
+void HandleGlobalLock(PPCContext* ppc_state, void* arg0, void* arg1) {
+  std::mutex* global_lock = reinterpret_cast<std::mutex*>(arg0);
+  volatile bool* global_lock_taken = reinterpret_cast<bool*>(arg1);
+  uint64_t value = ppc_state->scratch;
+  if (value == 0x8000) {
+    global_lock->unlock();
+    *global_lock_taken = false;
+  } else if (value == ppc_state->r[13]) {
+    global_lock->lock();
+    *global_lock_taken = true;
+  }
+}
+
 int PPCFrontend::Initialize() {
   int result = Frontend::Initialize();
   if (result) {
     return result;
   }
 
+  void* arg0 = reinterpret_cast<void*>(&builtins_.global_lock);
+  void* arg1 = reinterpret_cast<void*>(&builtins_.global_lock_taken);
+  builtins_.check_global_lock = runtime_->DefineBuiltin(
+      "CheckGlobalLock", (FunctionInfo::ExternHandler)CheckGlobalLock, arg0,
+      arg1);
+  builtins_.handle_global_lock = runtime_->DefineBuiltin(
+      "HandleGlobalLock", (FunctionInfo::ExternHandler)HandleGlobalLock, arg0,
+      arg1);
+
   return result;
 }
 

src/alloy/frontend/ppc/ppc_frontend.h

@@ -10,6 +10,8 @@
 #ifndef ALLOY_FRONTEND_PPC_PPC_FRONTEND_H_
 #define ALLOY_FRONTEND_PPC_PPC_FRONTEND_H_
 
+#include <mutex>
+
 #include <alloy/frontend/frontend.h>
 #include <alloy/type_pool.h>
 
@@ -19,6 +21,13 @@ namespace ppc {
 
 class PPCTranslator;
 
+struct PPCBuiltins {
+  std::mutex global_lock;
+  bool global_lock_taken;
+  runtime::FunctionInfo* check_global_lock;
+  runtime::FunctionInfo* handle_global_lock;
+};
+
 class PPCFrontend : public Frontend {
  public:
   PPCFrontend(runtime::Runtime* runtime);
@@ -26,6 +35,8 @@ class PPCFrontend : public Frontend {
 
   int Initialize() override;
 
+  PPCBuiltins* builtins() { return &builtins_; }
+
   int DeclareFunction(runtime::FunctionInfo* symbol_info) override;
   int DefineFunction(runtime::FunctionInfo* symbol_info,
                      uint32_t debug_info_flags, uint32_t trace_flags,
@@ -33,6 +44,7 @@ class PPCFrontend : public Frontend {
 
  private:
   TypePool<PPCTranslator, PPCFrontend*> translator_pool_;
+  PPCBuiltins builtins_;
 };
 
 }  // namespace ppc

src/alloy/frontend/ppc/ppc_hir_builder.cc

@@ -308,6 +308,20 @@ void PPCHIRBuilder::UpdateCR6(Value* src_value) {
   // TOOD(benvanik): trace CR.
 }
 
+Value* PPCHIRBuilder::LoadMSR() {
+  // bit 48 = EE; interrupt enabled
+  // bit 62 = RI; recoverable interrupt
+  // return 8000h if unlocked, else 0
+  CallExtern(frontend_->builtins()->check_global_lock);
+  return LoadContext(offsetof(PPCContext, scratch), INT64_TYPE);
+}
+
+void PPCHIRBuilder::StoreMSR(Value* value) {
+  // if & 0x8000 == 0, lock, else unlock
+  StoreContext(offsetof(PPCContext, scratch), ZeroExtend(value, INT64_TYPE));
+  CallExtern(frontend_->builtins()->handle_global_lock);
+}
+
 Value* PPCHIRBuilder::LoadFPSCR() {
   return LoadContext(offsetof(PPCContext, fpscr), INT64_TYPE);
 }
@@ -397,7 +411,12 @@ Value* PPCHIRBuilder::LoadAcquire(Value* address, TypeName type,
                                   uint32_t load_flags) {
   AtomicExchange(LoadContext(offsetof(PPCContext, reserve_address), INT64_TYPE),
                  Truncate(address, INT32_TYPE));
-  return Load(address, type, load_flags);
+  Value* value = Load(address, type, load_flags);
+  // Save the value so that we can compare it later in StoreRelease.
+  AtomicExchange(
+      LoadContext(offsetof(PPCContext, reserve_value), INT64_TYPE),
+      value);
+  return value;
 }
 
 Value* PPCHIRBuilder::StoreRelease(Value* address, Value* value,
@@ -405,7 +424,13 @@ Value* PPCHIRBuilder::StoreRelease(Value* address, Value* value,
   Value* old_address = AtomicExchange(
       LoadContext(offsetof(PPCContext, reserve_address), INT64_TYPE),
       LoadZero(INT32_TYPE));
-  Value* eq = CompareEQ(Truncate(address, INT32_TYPE), old_address);
+  // HACK: ensure the reservation addresses match AND the value hasn't changed.
+  Value* old_value = AtomicExchange(
+      LoadContext(offsetof(PPCContext, reserve_value), INT64_TYPE),
+      LoadZero(value->type));
+  Value* current_value = Load(address, value->type);
+  Value* eq = And(CompareEQ(Truncate(address, INT32_TYPE), old_address),
+                  CompareEQ(current_value, old_value));
   StoreContext(offsetof(PPCContext, cr0.cr0_eq), eq);
   StoreContext(offsetof(PPCContext, cr0.cr0_lt), LoadZero(INT8_TYPE));
   StoreContext(offsetof(PPCContext, cr0.cr0_gt), LoadZero(INT8_TYPE));

src/alloy/frontend/ppc/ppc_hir_builder.h

@@ -57,6 +57,8 @@ class PPCHIRBuilder : public hir::HIRBuilder {
   void UpdateCR(uint32_t n, Value* lhs, bool is_signed = true);
   void UpdateCR(uint32_t n, Value* lhs, Value* rhs, bool is_signed = true);
   void UpdateCR6(Value* src_value);
+  Value* LoadMSR();
+  void StoreMSR(Value* value);
   Value* LoadFPSCR();
   void StoreFPSCR(Value* value);
   Value* LoadXER();

src/alloy/frontend/ppc/test/alloy-ppc-test.cc

@@ -52,6 +52,7 @@ class ThreadState : public alloy::runtime::ThreadState {
 
     // Stash pointers to common structures that callbacks may need.
     context_->reserve_address = memory_->reserve_address();
+    context_->reserve_value = memory_->reserve_value();
     context_->membase = memory_->membase();
     context_->runtime = runtime;
     context_->thread_state = this;

src/alloy/memory.cc

@@ -13,7 +13,11 @@
 
 namespace alloy {
 
-Memory::Memory() : membase_(nullptr), reserve_address_(0), trace_base_(0) {
+Memory::Memory()
+    : membase_(nullptr),
+      reserve_address_(0),
+      reserve_value_(0),
+      trace_base_(0) {
   system_page_size_ = poly::page_size();
 }
 
@@ -64,6 +68,7 @@ uint64_t Memory::SearchAligned(uint64_t start, uint64_t end,
 SimpleMemory::SimpleMemory(size_t capacity) : memory_(capacity) {
   membase_ = reinterpret_cast<uint8_t*>(memory_.data());
   reserve_address_ = capacity - 8;
+  reserve_value_ = capacity - 16;
 }
 
 SimpleMemory::~SimpleMemory() = default;

src/alloy/memory.h

@@ -25,6 +25,7 @@ class Memory {
     return membase_ + guest_address;
   };
   inline uint64_t* reserve_address() { return &reserve_address_; }
+  inline uint64_t* reserve_value() { return &reserve_value_; }
 
   // TODO(benvanik): remove with GPU refactor.
   virtual uint64_t page_table() const = 0;
@@ -56,6 +57,7 @@ class Memory {
   size_t system_page_size_;
   uint8_t* membase_;
   uint64_t reserve_address_;
+  uint64_t reserve_value_;
   uint64_t trace_base_;
 };
 

src/alloy/runtime/runtime.cc

@@ -28,11 +28,25 @@ namespace runtime {
 using alloy::backend::Backend;
 using alloy::frontend::Frontend;
 
+class BuiltinModule : public Module {
+ public:
+  BuiltinModule(Runtime* runtime) : Module(runtime), name_("builtin") {}
+  const std::string& name() const override { return name_; }
+  bool ContainsAddress(uint64_t address) override {
+    return (address & 0x1FFFFFFF0) == 0x100000000;
+  }
+
+ private:
+  std::string name_;
+};
+
 Runtime::Runtime(Memory* memory, uint32_t debug_info_flags,
                  uint32_t trace_flags)
     : memory_(memory),
       debug_info_flags_(debug_info_flags),
-      trace_flags_(trace_flags) {}
+      trace_flags_(trace_flags),
+      builtin_module_(nullptr),
+      next_builtin_address_(0x100000000ull) {}
 
 Runtime::~Runtime() {
   {
@@ -53,6 +67,10 @@ int Runtime::Initialize(std::unique_ptr<Frontend> frontend,
   // Create debugger first. Other types hook up to it.
   debugger_.reset(new Debugger(this));
 
+  std::unique_ptr<Module> builtin_module(new BuiltinModule(this));
+  builtin_module_ = builtin_module.get();
+  modules_.push_back(std::move(builtin_module));
+
   if (frontend_ || backend_) {
     return 1;
   }
@@ -127,6 +145,22 @@ std::vector<Module*> Runtime::GetModules() {
   return clone;
 }
 
+FunctionInfo* Runtime::DefineBuiltin(const std::string& name,
+                                     FunctionInfo::ExternHandler handler,
+                                     void* arg0, void* arg1) {
+  uint64_t address = next_builtin_address_;
+  next_builtin_address_ += 4;
+
+  FunctionInfo* fn_info;
+  builtin_module_->DeclareFunction(address, &fn_info);
+  fn_info->set_end_address(address + 4);
+  fn_info->set_name(name);
+  fn_info->SetupExtern(handler, arg0, arg1);
+  fn_info->set_status(runtime::SymbolInfo::STATUS_DECLARED);
+
+  return fn_info;
+}
+
 std::vector<Function*> Runtime::FindFunctionsWithAddress(uint64_t address) {
   return entry_table_.FindWithAddress(address);
 }

src/alloy/runtime/runtime.h

@@ -45,6 +45,11 @@ class Runtime {
   Module* GetModule(const std::string& name) { return GetModule(name.c_str()); }
   std::vector<Module*> GetModules();
 
+  Module* builtin_module() const { return builtin_module_; }
+  FunctionInfo* DefineBuiltin(const std::string& name,
+                              FunctionInfo::ExternHandler handler, void* arg0,
+                              void* arg1);
+
   std::vector<Function*> FindFunctionsWithAddress(uint64_t address);
 
   int LookupFunctionInfo(uint64_t address, FunctionInfo** out_symbol_info);
@@ -71,6 +76,8 @@ class Runtime {
   EntryTable entry_table_;
   std::mutex modules_lock_;
   std::vector<std::unique_ptr<Module>> modules_;
+  Module* builtin_module_;
+  uint64_t next_builtin_address_;
 };
 
 }  // namespace runtime

src/alloy/test/util.h

@@ -47,6 +47,7 @@ class ThreadState : public alloy::runtime::ThreadState {
 
     // Stash pointers to common structures that callbacks may need.
     context_->reserve_address = memory_->reserve_address();
+    context_->reserve_value = memory_->reserve_value();
     context_->membase = memory_->membase();
     context_->runtime = runtime;
     context_->thread_state = this;

src/xenia/cpu/xenon_thread_state.cc

@@ -33,6 +33,7 @@ XenonThreadState::XenonThreadState(XenonRuntime* runtime, uint32_t thread_id,
 
   // Stash pointers to common structures that callbacks may need.
   context_->reserve_address = memory_->reserve_address();
+  context_->reserve_value = memory_->reserve_value();
   context_->membase = memory_->membase();
   context_->runtime = runtime;
   context_->thread_state = this;