CPU: Refactor execution mode switching

Fixes single step breaking in branch delay slots with recompiler.
Simplifies initialization.
Removes multiple sources of truth for fastmem.

src/core/bus.cpp

@@ -143,8 +143,6 @@ static RAM_SIZE_REG s_RAM_SIZE = {};
 
 static std::string s_tty_line_buffer;
 
-static CPUFastmemMode s_fastmem_mode = CPUFastmemMode::Disabled;
-
 #ifdef ENABLE_MMAP_FASTMEM
 static SharedMemoryMappingArea s_fastmem_arena;
 static std::vector<std::pair<u8*, size_t>> s_fastmem_ram_views;
@@ -161,6 +159,8 @@ static void SetRAMSize(bool enable_8mb_ram);
 static std::tuple<TickCount, TickCount, TickCount> CalculateMemoryTiming(MEMDELAY mem_delay, COMDELAY common_delay);
 static void RecalculateMemoryTimings();
 
+static void MapFastmemViews();
+static void UnmapFastmemViews();
 static u8* GetLUTFastmemPointer(u32 address, u8* ram_ptr);
 
 static void SetRAMPageWritable(u32 page_index, bool writable);
@@ -252,6 +252,7 @@ bool Bus::AllocateMemoryMap(bool export_shared_memory, Error* error)
 
   VERBOSE_LOG("LUTs are mapped at {}.", static_cast<void*>(g_memory_handlers));
   g_memory_handlers_isc = g_memory_handlers + MEMORY_LUT_SLOTS;
+  g_ram_mapped_size = RAM_8MB_SIZE;
   SetHandlers();
 
 #ifndef __ANDROID__
@@ -348,7 +349,7 @@ bool Bus::ReallocateMemoryMap(bool export_shared_memory, Error* error)
   if (System::IsValid())
   {
     CPU::CodeCache::InvalidateAllRAMBlocks();
-    UpdateFastmemViews(CPUFastmemMode::Disabled);
+    UnmapFastmemViews();
 
     ram_backup.resize(RAM_8MB_SIZE);
     std::memcpy(ram_backup.data(), g_unprotected_ram, RAM_8MB_SIZE);
@@ -365,7 +366,7 @@ bool Bus::ReallocateMemoryMap(bool export_shared_memory, Error* error)
     UpdateMappedRAMSize();
     std::memcpy(g_unprotected_ram, ram_backup.data(), RAM_8MB_SIZE);
     std::memcpy(g_bios, bios_backup.data(), BIOS_SIZE);
-    UpdateFastmemViews(g_settings.cpu_fastmem_mode);
+    MapFastmemViews();
   }
 
   return true;
@@ -380,11 +381,10 @@ void Bus::CleanupMemoryMap()
 #endif
 }
 
-bool Bus::Initialize()
+void Bus::Initialize()
 {
   SetRAMSize(g_settings.enable_8mb_ram);
-  Reset();
-  return true;
+  MapFastmemViews();
 }
 
 void Bus::SetRAMSize(bool enable_8mb_ram)
@@ -400,7 +400,7 @@ void Bus::SetRAMSize(bool enable_8mb_ram)
 
 void Bus::Shutdown()
 {
-  UpdateFastmemViews(CPUFastmemMode::Disabled);
+  UnmapFastmemViews();
   CPU::g_state.fastmem_base = nullptr;
 
   g_ram_mask = 0;
@@ -439,8 +439,7 @@ bool Bus::DoState(StateWrapper& sw)
   {
     const bool using_8mb_ram = (ram_size == RAM_8MB_SIZE);
     SetRAMSize(using_8mb_ram);
-    UpdateFastmemViews(s_fastmem_mode);
-    CPU::UpdateMemoryPointers();
+    RemapFastmemViews();
   }
 
   sw.Do(&g_exp1_access_time);
@@ -526,18 +525,13 @@ void Bus::RecalculateMemoryTimings()
             g_spu_access_time[2] + 1);
 }
 
-CPUFastmemMode Bus::GetFastmemMode()
-{
-  return s_fastmem_mode;
-}
-
 void* Bus::GetFastmemBase(bool isc)
 {
 #ifdef ENABLE_MMAP_FASTMEM
-  if (s_fastmem_mode == CPUFastmemMode::MMap)
+  if (g_settings.cpu_fastmem_mode == CPUFastmemMode::MMap)
     return isc ? nullptr : s_fastmem_arena.BasePointer();
 #endif
-  if (s_fastmem_mode == CPUFastmemMode::LUT)
+  if (g_settings.cpu_fastmem_mode == CPUFastmemMode::LUT)
     return reinterpret_cast<u8*>(s_fastmem_lut + (isc ? (FASTMEM_LUT_SIZE * sizeof(void*)) : 0));
 
   return nullptr;
@@ -548,24 +542,16 @@ u8* Bus::GetLUTFastmemPointer(u32 address, u8* ram_ptr)
   return ram_ptr - address;
 }
 
-void Bus::UpdateFastmemViews(CPUFastmemMode mode)
+void Bus::MapFastmemViews()
 {
-#ifndef ENABLE_MMAP_FASTMEM
-  Assert(mode != CPUFastmemMode::MMap);
-#else
-  for (const auto& it : s_fastmem_ram_views)
-    s_fastmem_arena.Unmap(it.first, it.second);
-  s_fastmem_ram_views.clear();
-#endif
+  Assert(s_fastmem_ram_views.empty());
 
-  s_fastmem_mode = mode;
-  if (mode == CPUFastmemMode::Disabled)
-    return;
-
-#ifdef ENABLE_MMAP_FASTMEM
+  const CPUFastmemMode mode = g_settings.cpu_fastmem_mode;
   if (mode == CPUFastmemMode::MMap)
   {
+#ifdef ENABLE_MMAP_FASTMEM
     auto MapRAM = [](u32 base_address) {
+      // No need to check mapped RAM range here, we only ever fastmem map the first 2MB.
       u8* map_address = s_fastmem_arena.BasePointer() + base_address;
       if (!s_fastmem_arena.Map(s_shmem_handle, 0, map_address, g_ram_size, PageProtect::ReadWrite)) [[unlikely]]
       {
@@ -600,57 +586,80 @@ void Bus::UpdateFastmemViews(CPUFastmemMode mode)
 
     // KSEG1 - uncached
     MapRAM(0xA0000000);
-
-    return;
-  }
+#else
+    Panic("MMap fastmem should not be selected on this platform.");
 #endif
-
-  if (!s_fastmem_lut)
+  }
+  else if (mode == CPUFastmemMode::LUT)
   {
-    s_fastmem_lut = static_cast<u8**>(std::malloc(sizeof(u8*) * FASTMEM_LUT_SLOTS));
-    Assert(s_fastmem_lut);
+    if (!s_fastmem_lut)
+    {
+      s_fastmem_lut = static_cast<u8**>(std::malloc(sizeof(u8*) * FASTMEM_LUT_SLOTS));
+      Assert(s_fastmem_lut);
 
-    INFO_LOG("Fastmem base (software): {}", static_cast<void*>(s_fastmem_lut));
-  }
+      INFO_LOG("Fastmem base (software): {}", static_cast<void*>(s_fastmem_lut));
+    }
 
-  // This assumes the top 4KB of address space is not mapped. It shouldn't be on any sane OSes.
-  for (u32 i = 0; i < FASTMEM_LUT_SLOTS; i++)
-    s_fastmem_lut[i] = GetLUTFastmemPointer(i << FASTMEM_LUT_PAGE_SHIFT, nullptr);
+    // This assumes the top 4KB of address space is not mapped. It shouldn't be on any sane OSes.
+    for (u32 i = 0; i < FASTMEM_LUT_SLOTS; i++)
+      s_fastmem_lut[i] = GetLUTFastmemPointer(i << FASTMEM_LUT_PAGE_SHIFT, nullptr);
 
-  auto MapRAM = [](u32 base_address) {
-    u8* ram_ptr = g_ram + (base_address & g_ram_mask);
-    for (u32 address = 0; address < g_ram_size; address += FASTMEM_LUT_PAGE_SIZE)
-    {
-      const u32 lut_index = (base_address + address) >> FASTMEM_LUT_PAGE_SHIFT;
-      s_fastmem_lut[lut_index] = GetLUTFastmemPointer(base_address + address, ram_ptr);
-      ram_ptr += FASTMEM_LUT_PAGE_SIZE;
-    }
-  };
+    auto MapRAM = [](u32 base_address) {
+      // Don't map RAM that isn't accessible.
+      if ((base_address & CPU::PHYSICAL_MEMORY_ADDRESS_MASK) >= g_ram_mapped_size)
+        return;
+
+      u8* ram_ptr = g_ram + (base_address & g_ram_mask);
+      for (u32 address = 0; address < g_ram_size; address += FASTMEM_LUT_PAGE_SIZE)
+      {
+        const u32 lut_index = (base_address + address) >> FASTMEM_LUT_PAGE_SHIFT;
+        s_fastmem_lut[lut_index] = GetLUTFastmemPointer(base_address + address, ram_ptr);
+        ram_ptr += FASTMEM_LUT_PAGE_SIZE;
+      }
+    };
+
+    // KUSEG - cached
+    MapRAM(0x00000000);
+    MapRAM(0x00200000);
+    MapRAM(0x00400000);
+    MapRAM(0x00600000);
+
+    // KSEG0 - cached
+    MapRAM(0x80000000);
+    MapRAM(0x80200000);
+    MapRAM(0x80400000);
+    MapRAM(0x80600000);
+
+    // KSEG1 - uncached
+    MapRAM(0xA0000000);
+    MapRAM(0xA0200000);
+    MapRAM(0xA0400000);
+    MapRAM(0xA0600000);
+  }
 
-  // KUSEG - cached
-  MapRAM(0x00000000);
-  MapRAM(0x00200000);
-  MapRAM(0x00400000);
-  MapRAM(0x00600000);
+  CPU::UpdateMemoryPointers();
+}
 
-  // KSEG0 - cached
-  MapRAM(0x80000000);
-  MapRAM(0x80200000);
-  MapRAM(0x80400000);
-  MapRAM(0x80600000);
+void Bus::UnmapFastmemViews()
+{
+#ifdef ENABLE_MMAP_FASTMEM
+  for (const auto& it : s_fastmem_ram_views)
+    s_fastmem_arena.Unmap(it.first, it.second);
+  s_fastmem_ram_views.clear();
+#endif
+}
 
-  // KSEG1 - uncached
-  MapRAM(0xA0000000);
-  MapRAM(0xA0200000);
-  MapRAM(0xA0400000);
-  MapRAM(0xA0600000);
+void Bus::RemapFastmemViews()
+{
+  UnmapFastmemViews();
+  MapFastmemViews();
 }
 
 bool Bus::CanUseFastmemForAddress(VirtualMemoryAddress address)
 {
   const PhysicalMemoryAddress paddr = address & CPU::PHYSICAL_MEMORY_ADDRESS_MASK;
 
-  switch (s_fastmem_mode)
+  switch (g_settings.cpu_fastmem_mode)
   {
 #ifdef ENABLE_MMAP_FASTMEM
     case CPUFastmemMode::MMap:
@@ -706,7 +715,7 @@ void Bus::SetRAMPageWritable(u32 page_index, bool writable)
   }
 
 #ifdef ENABLE_MMAP_FASTMEM
-  if (s_fastmem_mode == CPUFastmemMode::MMap)
+  if (g_settings.cpu_fastmem_mode == CPUFastmemMode::MMap)
   {
     const PageProtect protect = writable ? PageProtect::ReadWrite : PageProtect::ReadOnly;
 
@@ -734,7 +743,7 @@ void Bus::ClearRAMCodePageFlags()
     ERROR_LOG("Failed to restore RAM protection to read-write.");
 
 #ifdef ENABLE_MMAP_FASTMEM
-  if (s_fastmem_mode == CPUFastmemMode::MMap)
+  if (g_settings.cpu_fastmem_mode == CPUFastmemMode::MMap)
   {
     // unprotect fastmem pages
     for (const auto& it : s_fastmem_ram_views)
@@ -1041,7 +1050,8 @@ bool Bus::SideloadEXE(const std::string& path, Error* error)
   }
   if (okay)
   {
-    const std::optional<DynamicHeapArray<u8>> exe_data = FileSystem::ReadBinaryFile(System::GetExeOverride().c_str(), error);
+    const std::optional<DynamicHeapArray<u8>> exe_data =
+      FileSystem::ReadBinaryFile(System::GetExeOverride().c_str(), error);
     okay = (exe_data.has_value() && InjectExecutable(exe_data->cspan(), true, error));
     if (!okay)
       Error::AddPrefixFmt(error, "Failed to load {}: ", Path::GetFileName(path));
@@ -1961,6 +1971,8 @@ void Bus::SetHandlers()
 
 void Bus::UpdateMappedRAMSize()
 {
+  const u32 prev_mapped_size = g_ram_mapped_size;
+
   switch (s_RAM_SIZE.memory_window)
   {
     case 4: // 2MB memory + 6MB unmapped
@@ -2003,6 +2015,10 @@ void Bus::UpdateMappedRAMSize()
     }
     break;
   }
+
+  // Fastmem needs to be remapped.
+  if (prev_mapped_size != g_ram_mapped_size)
+    RemapFastmemViews();
 }
 
 #undef SET

src/core/bus.h

@@ -127,7 +127,7 @@ bool ReallocateMemoryMap(bool export_shared_memory, Error* error);
 /// Should be called when the process crashes, to avoid leaking.
 void CleanupMemoryMap();
 
-bool Initialize();
+void Initialize();
 void Shutdown();
 void Reset();
 bool DoState(StateWrapper& sw);
@@ -144,9 +144,8 @@ ALWAYS_INLINE_RELEASE static FP* OffsetHandlerArray(void** handlers, MemoryAcces
                                (((static_cast<size_t>(size) * 2) + static_cast<size_t>(type)) * MEMORY_LUT_SIZE));
 }
 
-CPUFastmemMode GetFastmemMode();
 void* GetFastmemBase(bool isc);
-void UpdateFastmemViews(CPUFastmemMode mode);
+void RemapFastmemViews();
 bool CanUseFastmemForAddress(VirtualMemoryAddress address);
 
 void SetExpansionROM(std::vector<u8> data);

src/core/cpu_code_cache.cpp

@@ -102,7 +102,6 @@ static void BacklinkBlocks(u32 pc, const void* dst);
 static void UnlinkBlockExits(Block* block);
 static void ResetCodeBuffer();
 
-static void ClearASMFunctions();
 static void CompileASMFunctions();
 static bool CompileBlock(Block* block);
 static PageFaultHandler::HandlerResult HandleFastmemException(void* exception_pc, void* fault_address, bool is_write);
@@ -174,7 +173,7 @@ bool CPU::CodeCache::IsUsingAnyRecompiler()
 
 bool CPU::CodeCache::IsUsingFastmem()
 {
-  return IsUsingAnyRecompiler() && g_settings.cpu_fastmem_mode != CPUFastmemMode::Disabled;
+  return g_settings.cpu_fastmem_mode != CPUFastmemMode::Disabled;
 }
 
 bool CPU::CodeCache::ProcessStartup(Error* error)
@@ -214,37 +213,12 @@ void CPU::CodeCache::ProcessShutdown()
 #endif
 }
 
-void CPU::CodeCache::Initialize()
-{
-  Assert(s_blocks.empty());
-
-  if (IsUsingAnyRecompiler())
-  {
-    ResetCodeBuffer();
-    CompileASMFunctions();
-    ResetCodeLUT();
-  }
-
-  Bus::UpdateFastmemViews(IsUsingAnyRecompiler() ? g_settings.cpu_fastmem_mode : CPUFastmemMode::Disabled);
-  CPU::UpdateMemoryPointers();
-}
-
-void CPU::CodeCache::Shutdown()
-{
-  ClearBlocks();
-  ClearASMFunctions();
-
-  Bus::UpdateFastmemViews(CPUFastmemMode::Disabled);
-  CPU::UpdateMemoryPointers();
-}
-
 void CPU::CodeCache::Reset()
 {
   ClearBlocks();
 
   if (IsUsingAnyRecompiler())
   {
-    ClearASMFunctions();
     ResetCodeBuffer();
     CompileASMFunctions();
     ResetCodeLUT();
@@ -1560,25 +1534,14 @@ const void* CPU::CodeCache::GetInterpretUncachedBlockFunction()
   }
 }
 
-void CPU::CodeCache::ClearASMFunctions()
+void CPU::CodeCache::CompileASMFunctions()
 {
-  g_enter_recompiler = nullptr;
-  g_compile_or_revalidate_block = nullptr;
-  g_check_events_and_dispatch = nullptr;
-  g_run_events_and_dispatch = nullptr;
-  g_dispatcher = nullptr;
-  g_interpret_block = nullptr;
-  g_discard_and_recompile_block = nullptr;
+  MemMap::BeginCodeWrite();
 
 #ifdef _DEBUG
   s_total_instructions_compiled = 0;
   s_total_host_instructions_emitted = 0;
 #endif
-}
-
-void CPU::CodeCache::CompileASMFunctions()
-{
-  MemMap::BeginCodeWrite();
 
   const u32 asm_size = EmitASMFunctions(GetFreeCodePointer(), GetFreeCodeSpace());
 

src/core/cpu_code_cache.h

@@ -22,12 +22,6 @@ bool ProcessStartup(Error* error);
 /// Frees resources, call once at shutdown.
 void ProcessShutdown();
 
-/// Initializes resources for the system.
-void Initialize();
-
-/// Frees resources used by the system.
-void Shutdown();
-
 /// Runs the system.
 [[noreturn]] void Execute();
 

src/core/cpu_core.cpp

@@ -29,7 +29,14 @@
 Log_SetChannel(CPU::Core);
 
 namespace CPU {
-static bool ShouldUseInterpreter();
+enum class ExecutionBreakType
+{
+  None,
+  ExecuteOneInstruction,
+  SingleStep,
+  Breakpoint,
+};
+
 static void UpdateLoadDelay();
 static void Branch(u32 target);
 static void FlushLoadDelay();
@@ -68,6 +75,8 @@ static void LogInstruction(u32 bits, u32 pc, bool regs);
 static void HandleWriteSyscall();
 static void HandlePutcSyscall();
 static void HandlePutsSyscall();
+
+static void CheckForExecutionModeChange();
 [[noreturn]] static void ExecuteInterpreter();
 
 template<PGXPMode pgxp_mode, bool debug>
@@ -104,8 +113,8 @@ static constexpr u32 INVALID_BREAKPOINT_PC = UINT32_C(0xFFFFFFFF);
 static std::array<std::vector<Breakpoint>, static_cast<u32>(BreakpointType::Count)> s_breakpoints;
 static u32 s_breakpoint_counter = 1;
 static u32 s_last_breakpoint_check_pc = INVALID_BREAKPOINT_PC;
-static bool s_single_step = false;
-static bool s_break_after_instruction = false;
+static CPUExecutionMode s_current_execution_mode = CPUExecutionMode::Interpreter;
+static ExecutionBreakType s_break_type = ExecutionBreakType::None;
 } // namespace CPU
 
 bool CPU::IsTraceEnabled()
@@ -163,14 +172,14 @@ void CPU::Initialize()
   // From nocash spec.
   g_state.cop0_regs.PRID = UINT32_C(0x00000002);
 
+  s_current_execution_mode = g_settings.cpu_execution_mode;
   g_state.using_debug_dispatcher = false;
-  g_state.using_interpreter = ShouldUseInterpreter();
+  g_state.using_interpreter = (s_current_execution_mode == CPUExecutionMode::Interpreter);
   for (BreakpointList& bps : s_breakpoints)
     bps.clear();
   s_breakpoint_counter = 1;
   s_last_breakpoint_check_pc = INVALID_BREAKPOINT_PC;
-  s_single_step = false;
-  s_break_after_instruction = false;
+  s_break_type = ExecutionBreakType::None;
 
   UpdateMemoryPointers();
   UpdateDebugDispatcherFlag();
@@ -280,33 +289,23 @@ bool CPU::DoState(StateWrapper& sw)
     sw.Do(&g_state.icache_data);
   }
 
-  bool using_interpreter = g_state.using_interpreter;
-  sw.DoEx(&using_interpreter, 67, g_state.using_interpreter);
+  sw.DoEx(&g_state.using_interpreter, 67, g_state.using_interpreter);
 
   if (sw.IsReading())
   {
-    // Since the recompilers do not use npc/next_instruction, and the icache emulation doesn't actually fill the data,
-    // only the tags, if we save state with the recompiler, then load state with the interpreter, we're most likely
-    // going to crash. Clear both in the case that we are switching.
-    if (using_interpreter != g_state.using_interpreter)
-    {
-      WARNING_LOG("Current execution mode does not match save state. Resetting icache state.");
-      ExecutionModeChanged();
-    }
-
-    UpdateMemoryPointers();
+    // Trigger an execution mode change if the state was/wasn't using the interpreter.
+    s_current_execution_mode =
+      g_state.using_interpreter ?
+        CPUExecutionMode::Interpreter :
+        ((g_settings.cpu_execution_mode == CPUExecutionMode::Interpreter) ? CPUExecutionMode::CachedInterpreter :
+                                                                            g_settings.cpu_execution_mode);
     g_state.gte_completion_tick = 0;
+    UpdateMemoryPointers();
   }
 
   return !sw.HasError();
 }
 
-ALWAYS_INLINE_RELEASE bool CPU::ShouldUseInterpreter()
-{
-  // Currently, any breakpoints require the interpreter.
-  return (g_settings.cpu_execution_mode == CPUExecutionMode::Interpreter || g_state.using_debug_dispatcher);
-}
-
 void CPU::SetPC(u32 new_pc)
 {
   DebugAssert(Common::IsAlignedPow2(new_pc, 4));
@@ -1994,9 +1993,14 @@ void CPU::DispatchInterrupt()
   TimingEvents::UpdateCPUDowncount();
 }
 
+CPUExecutionMode CPU::GetCurrentExecutionMode()
+{
+  return s_current_execution_mode;
+}
+
 bool CPU::UpdateDebugDispatcherFlag()
 {
-  const bool has_any_breakpoints = HasAnyBreakpoints() || s_single_step;
+  const bool has_any_breakpoints = (HasAnyBreakpoints() || s_break_type == ExecutionBreakType::SingleStep);
 
   const auto& dcic = g_state.cop0_regs.dcic;
   const bool has_cop0_breakpoints = dcic.super_master_enable_1 && dcic.super_master_enable_2 &&
@@ -2010,12 +2014,73 @@ bool CPU::UpdateDebugDispatcherFlag()
 
   DEV_LOG("{} debug dispatcher", use_debug_dispatcher ? "Now using" : "No longer using");
   g_state.using_debug_dispatcher = use_debug_dispatcher;
+  return true;
+}
 
-  // Switching to interpreter?
-  if (g_state.using_interpreter != ShouldUseInterpreter())
-    ExecutionModeChanged();
+void CPU::CheckForExecutionModeChange()
+{
+  // Currently, any breakpoints require the interpreter.
+  const CPUExecutionMode new_execution_mode =
+    (g_state.using_debug_dispatcher ? CPUExecutionMode::Interpreter : g_settings.cpu_execution_mode);
+  if (s_current_execution_mode == new_execution_mode) [[likely]]
+  {
+    DebugAssert(g_state.using_interpreter == (s_current_execution_mode == CPUExecutionMode::Interpreter));
+    return;
+  }
 
-  return true;
+  WARNING_LOG("Execution mode changed from {} to {}", Settings::GetCPUExecutionModeName(s_current_execution_mode),
+              Settings::GetCPUExecutionModeName(new_execution_mode));
+
+  const bool new_interpreter = (new_execution_mode == CPUExecutionMode::Interpreter);
+  if (g_state.using_interpreter != new_interpreter)
+  {
+    // Have to clear out the icache too, only the tags are valid in the recs.
+    ClearICache();
+    g_state.bus_error = false;
+
+    if (new_interpreter)
+    {
+      // Switching to interpreter. Set up the pipeline.
+      // We'll also need to fetch the next instruction to execute.
+      if (!SafeReadInstruction(g_state.pc, &g_state.next_instruction.bits)) [[unlikely]]
+      {
+        g_state.next_instruction.bits = 0;
+        ERROR_LOG("Failed to read current instruction from 0x{:08X}", g_state.pc);
+      }
+
+      g_state.npc = g_state.pc + sizeof(Instruction);
+    }
+    else
+    {
+      // Switching to recompiler. We can't start a rec block in a branch delay slot, so we need to execute the
+      // instruction if we're currently in one.
+      if (g_state.next_instruction_is_branch_delay_slot) [[unlikely]]
+      {
+        while (g_state.next_instruction_is_branch_delay_slot)
+        {
+          WARNING_LOG("EXECMODE: Executing instruction at 0x{:08X} because it is in a branch delay slot.", g_state.pc);
+          if (fastjmp_set(&s_jmp_buf) == 0)
+          {
+            s_break_type = ExecutionBreakType::ExecuteOneInstruction;
+            g_state.using_debug_dispatcher = true;
+            ExecuteInterpreter();
+          }
+        }
+
+        // Need to restart the whole process again, because the branch slot could change the debug flag.
+        UpdateDebugDispatcherFlag();
+        CheckForExecutionModeChange();
+        return;
+      }
+    }
+  }
+
+  s_current_execution_mode = new_execution_mode;
+  g_state.using_interpreter = new_interpreter;
+
+  // Wipe out code cache when switching modes.
+  if (!new_interpreter)
+    CPU::CodeCache::Reset();
 }
 
 [[noreturn]] void CPU::ExitExecution()
@@ -2269,20 +2334,11 @@ ALWAYS_INLINE_RELEASE bool CPU::CheckBreakpointList(BreakpointType type, Virtual
 
 ALWAYS_INLINE_RELEASE void CPU::ExecutionBreakpointCheck()
 {
-  if (s_single_step) [[unlikely]]
-  {
-    // single step ignores breakpoints, since it stops anyway
-    s_single_step = false;
-    s_break_after_instruction = true;
-    Host::ReportDebuggerMessage(fmt::format("Stepped to 0x{:08X}.", g_state.npc));
-    return;
-  }
-
   if (s_breakpoints[static_cast<u32>(BreakpointType::Execute)].empty()) [[likely]]
     return;
 
   const u32 pc = g_state.pc;
-  if (pc == s_last_breakpoint_check_pc) [[unlikely]]
+  if (pc == s_last_breakpoint_check_pc || s_break_type == ExecutionBreakType::ExecuteOneInstruction) [[unlikely]]
   {
     // we don't want to trigger the same breakpoint which just paused us repeatedly.
     return;
@@ -2292,7 +2348,7 @@ ALWAYS_INLINE_RELEASE void CPU::ExecutionBreakpointCheck()
 
   if (CheckBreakpointList(BreakpointType::Execute, pc)) [[unlikely]]
   {
-    s_single_step = false;
+    s_break_type = ExecutionBreakType::None;
     ExitExecution();
   }
 }
@@ -2301,8 +2357,8 @@ template<MemoryAccessType type>
 ALWAYS_INLINE_RELEASE void CPU::MemoryBreakpointCheck(VirtualMemoryAddress address)
 {
   const BreakpointType bptype = (type == MemoryAccessType::Read) ? BreakpointType::Read : BreakpointType::Write;
-  if (CheckBreakpointList(bptype, address))
-    s_break_after_instruction = true;
+  if (CheckBreakpointList(bptype, address)) [[unlikely]]
+    s_break_type = ExecutionBreakType::Breakpoint;
 }
 
 template<PGXPMode pgxp_mode, bool debug>
@@ -2364,10 +2420,12 @@ template<PGXPMode pgxp_mode, bool debug>
 
       if constexpr (debug)
       {
-        if (s_break_after_instruction)
+        if (s_break_type != ExecutionBreakType::None) [[unlikely]]
         {
-          s_break_after_instruction = false;
-          System::PauseSystem(true);
+          const ExecutionBreakType break_type = std::exchange(s_break_type, ExecutionBreakType::None);
+          if (break_type >= ExecutionBreakType::SingleStep)
+            System::PauseSystem(true);
+
           UpdateDebugDispatcherFlag();
           ExitExecution();
         }
@@ -2412,6 +2470,8 @@ void CPU::ExecuteInterpreter()
 
 void CPU::Execute()
 {
+  CheckForExecutionModeChange();
+
   if (fastjmp_set(&s_jmp_buf) != 0)
     return;
 
@@ -2423,7 +2483,7 @@ void CPU::Execute()
 
 void CPU::SetSingleStepFlag()
 {
-  s_single_step = true;
+  s_break_type = ExecutionBreakType::SingleStep;
   if (UpdateDebugDispatcherFlag())
     System::InterruptExecution();
 }
@@ -2571,41 +2631,6 @@ void CPU::UpdateMemoryPointers()
   g_state.fastmem_base = Bus::GetFastmemBase(g_state.cop0_regs.sr.Isc);
 }
 
-void CPU::ExecutionModeChanged()
-{
-  const bool prev_interpreter = g_state.using_interpreter;
-
-  UpdateDebugDispatcherFlag();
-
-  // Clear out bus errors in case only memory exceptions are toggled on.
-  g_state.bus_error = false;
-
-  // Have to clear out the icache too, only the tags are valid in the recs.
-  ClearICache();
-
-  // Switching to interpreter?
-  g_state.using_interpreter = ShouldUseInterpreter();
-  if (g_state.using_interpreter != prev_interpreter && !prev_interpreter)
-  {
-    // Before we return, set npc to pc so that we can switch from recs to int.
-    // We'll also need to fetch the next instruction to execute.
-    if (!SafeReadInstruction(g_state.pc, &g_state.next_instruction.bits)) [[unlikely]]
-    {
-      g_state.next_instruction.bits = 0;
-      ERROR_LOG("Failed to read current instruction from 0x{:08X}", g_state.pc);
-    }
-
-    g_state.npc = g_state.pc + sizeof(Instruction);
-  }
-
-  // Wipe out code cache when switching back to recompiler.
-  if (!g_state.using_interpreter && prev_interpreter)
-    CPU::CodeCache::Reset();
-
-  UpdateDebugDispatcherFlag();
-  System::InterruptExecution();
-}
-
 template<bool add_ticks, bool icache_read, u32 word_count, bool raise_exceptions>
 ALWAYS_INLINE_RELEASE bool CPU::DoInstructionRead(PhysicalMemoryAddress address, void* data)
 {

src/core/cpu_core.h

@@ -132,9 +132,9 @@ void Shutdown();
 void Reset();
 bool DoState(StateWrapper& sw);
 void ClearICache();
+CPUExecutionMode GetCurrentExecutionMode();
 bool UpdateDebugDispatcherFlag();
 void UpdateMemoryPointers();
-void ExecutionModeChanged();
 
 /// Executes interpreter loop.
 void Execute();

src/core/system.cpp

@@ -1832,20 +1832,13 @@ bool System::Initialize(bool force_software_renderer, Error* error)
 
   TimingEvents::Initialize();
 
+  Bus::Initialize();
   CPU::Initialize();
 
-  if (!Bus::Initialize())
-  {
-    CPU::Shutdown();
-    return false;
-  }
-
-  CPU::CodeCache::Initialize();
-
   if (!CreateGPU(force_software_renderer ? GPURenderer::Software : g_settings.gpu_renderer, false, error))
   {
-    Bus::Shutdown();
     CPU::Shutdown();
+    Bus::Shutdown();
     return false;
   }
 
@@ -1933,9 +1926,8 @@ void System::DestroySystem()
   g_gpu.reset();
   DMA::Shutdown();
   CPU::PGXP::Shutdown();
-  CPU::CodeCache::Shutdown();
-  Bus::Shutdown();
   CPU::Shutdown();
+  Bus::Shutdown();
   TimingEvents::Shutdown();
   ClearRunningGame();
 
@@ -3383,6 +3375,9 @@ void System::UpdateDisplayVSync()
   // Avoid flipping vsync on and off by manually throttling when vsync is on.
   const GPUVSyncMode vsync_mode = GetEffectiveVSyncMode();
   const bool allow_present_throttle = ShouldAllowPresentThrottle();
+  if (g_gpu_device->GetVSyncMode() == vsync_mode && g_gpu_device->IsPresentThrottleAllowed() == allow_present_throttle)
+    return;
+
   VERBOSE_LOG("VSync: {}{}{}", vsync_modes[static_cast<size_t>(vsync_mode)],
               s_syncing_to_host_with_vsync ? " (for throttling)" : "",
               allow_present_throttle ? " (present throttle allowed)" : "");
@@ -4246,22 +4241,18 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
     if (g_settings.emulation_speed != old_settings.emulation_speed)
       UpdateThrottlePeriod();
 
-    if (g_settings.cpu_execution_mode != old_settings.cpu_execution_mode ||
-        g_settings.cpu_fastmem_mode != old_settings.cpu_fastmem_mode)
+    if (g_settings.cpu_execution_mode != old_settings.cpu_execution_mode)
     {
       Host::AddIconOSDMessage("CPUExecutionModeSwitch", ICON_FA_MICROCHIP,
                               fmt::format(TRANSLATE_FS("OSDMessage", "Switching to {} CPU execution mode."),
                                           TRANSLATE_SV("CPUExecutionMode", Settings::GetCPUExecutionModeDisplayName(
                                                                              g_settings.cpu_execution_mode))),
                               Host::OSD_INFO_DURATION);
-      CPU::ExecutionModeChanged();
-      if (old_settings.cpu_execution_mode != CPUExecutionMode::Interpreter)
-        CPU::CodeCache::Shutdown();
-      if (g_settings.cpu_execution_mode != CPUExecutionMode::Interpreter)
-        CPU::CodeCache::Initialize();
+      CPU::UpdateDebugDispatcherFlag();
+      InterruptExecution();
     }
 
-    if (CPU::CodeCache::IsUsingAnyRecompiler() &&
+    if (CPU::GetCurrentExecutionMode() != CPUExecutionMode::Interpreter &&
         (g_settings.cpu_recompiler_memory_exceptions != old_settings.cpu_recompiler_memory_exceptions ||
          g_settings.cpu_recompiler_block_linking != old_settings.cpu_recompiler_block_linking ||
          g_settings.cpu_recompiler_icache != old_settings.cpu_recompiler_icache ||
@@ -4270,12 +4261,21 @@ void System::CheckForSettingsChanges(const Settings& old_settings)
       Host::AddIconOSDMessage("CPUFlushAllBlocks", ICON_FA_MICROCHIP,
                               TRANSLATE_STR("OSDMessage", "Recompiler options changed, flushing all blocks."),
                               Host::OSD_INFO_DURATION);
-      CPU::ExecutionModeChanged();
+      CPU::CodeCache::Reset();
+      CPU::g_state.bus_error = false;
     }
     else if (g_settings.cpu_execution_mode == CPUExecutionMode::Interpreter &&
              g_settings.bios_tty_logging != old_settings.bios_tty_logging)
     {
-      CPU::UpdateDebugDispatcherFlag();
+      // TTY interception requires debug dispatcher.
+      if (CPU::UpdateDebugDispatcherFlag())
+        InterruptExecution();
+    }
+
+    if (g_settings.cpu_fastmem_mode != old_settings.cpu_fastmem_mode)
+    {
+      // Reallocate fastmem area, even if it's not being used.
+      Bus::RemapFastmemViews();
     }
 
     if (g_settings.enable_cheats != old_settings.enable_cheats)

src/util/gpu_device.h

@@ -711,6 +711,7 @@ public:
 
   ALWAYS_INLINE GPUVSyncMode GetVSyncMode() const { return m_vsync_mode; }
   ALWAYS_INLINE bool IsVSyncModeBlocking() const { return (m_vsync_mode == GPUVSyncMode::FIFO); }
+  ALWAYS_INLINE bool IsPresentThrottleAllowed() const { return m_allow_present_throttle; }
   virtual void SetVSyncMode(GPUVSyncMode mode, bool allow_present_throttle) = 0;
 
   ALWAYS_INLINE bool IsDebugDevice() const { return m_debug_device; }