CPU/Recompiler: Write exception exits to far code buffer

Keeps the hot path nice and clean.

src/common/jit_code_buffer.cpp

@@ -7,10 +7,12 @@
 #include <sys/mman.h>
 #endif
 
-JitCodeBuffer::JitCodeBuffer(size_t size)
+JitCodeBuffer::JitCodeBuffer(size_t size /* = 64 * 1024 * 1024 */, size_t far_code_size /* = 0 */)
 {
+  m_total_size = size + far_code_size;
+
 #if defined(Y_PLATFORM_WINDOWS)
-  m_code_ptr = VirtualAlloc(nullptr, size, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
+  m_code_ptr = VirtualAlloc(nullptr, m_total_size, MEM_COMMIT, PAGE_EXECUTE_READWRITE);
 #elif defined(Y_PLATFORM_LINUX) || defined(Y_PLATFORM_ANDROID)
   m_code_ptr = mmap(nullptr, size, PROT_READ | PROT_WRITE | PROT_EXEC, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
 #else
@@ -20,6 +22,10 @@ JitCodeBuffer::JitCodeBuffer(size_t size)
   m_code_size = size;
   m_code_used = 0;
 
+  m_far_code_ptr = static_cast<u8*>(m_code_ptr) + size;
+  m_far_code_size = far_code_size;
+  m_far_code_used = 0;
+
   if (!m_code_ptr)
     Panic("Failed to allocate code space.");
 }
@@ -27,34 +33,39 @@ JitCodeBuffer::JitCodeBuffer(size_t size)
 JitCodeBuffer::~JitCodeBuffer()
 {
 #if defined(Y_PLATFORM_WINDOWS)
-  VirtualFree(m_code_ptr, m_code_size, MEM_RELEASE);
+  VirtualFree(m_code_ptr, m_total_size, MEM_RELEASE);
 #elif defined(Y_PLATFORM_LINUX) || defined(Y_PLATFORM_ANDROID)
-  munmap(m_code_ptr, m_code_size);
+  munmap(m_code_ptr, m_total_size);
 #endif
 }
 
 void JitCodeBuffer::CommitCode(size_t length)
 {
-  //     // Function alignment?
-  //     size_t extra_bytes = ((length % 16) != 0) ? (16 - (length % 16)) : 0;
-  //     for (size_t i = 0; i < extra_bytes; i++)
-  //         reinterpret_cast<char*>(m_free_code_ptr)[i] = 0xCC;
-
   Assert(length <= (m_code_size - m_code_used));
-  m_free_code_ptr = reinterpret_cast<char*>(m_free_code_ptr) + length;
+  m_free_code_ptr = reinterpret_cast<u8*>(m_free_code_ptr) + length;
   m_code_used += length;
 }
 
+void JitCodeBuffer::CommitFarCode(size_t length)
+{
+  Assert(length <= (m_far_code_size - m_far_code_used));
+  m_free_far_code_ptr = reinterpret_cast<u8*>(m_free_far_code_ptr) + length;
+  m_far_code_used += length;
+}
+
 void JitCodeBuffer::Reset()
 {
 #if defined(Y_PLATFORM_WINDOWS)
-  FlushInstructionCache(GetCurrentProcess(), m_code_ptr, m_code_size);
+  FlushInstructionCache(GetCurrentProcess(), m_code_ptr, m_total_size);
 #elif defined(Y_PLATFORM_LINUX) || defined(Y_PLATFORM_ANDROID)
 // TODO
 #endif
 
   m_free_code_ptr = m_code_ptr;
   m_code_used = 0;
+
+  m_free_far_code_ptr = m_far_code_ptr;
+  m_far_code_used = 0;
 }
 
 void JitCodeBuffer::Align(u32 alignment, u8 padding_value)

src/common/jit_code_buffer.h

@@ -4,13 +4,18 @@
 class JitCodeBuffer
 {
 public:
-  JitCodeBuffer(size_t size = 64 * 1024 * 1024);
+  JitCodeBuffer(size_t size = 64 * 1024 * 1024, size_t far_code_size = 0);
   ~JitCodeBuffer();
 
+  void Reset();
+
   void* GetFreeCodePointer() const { return m_free_code_ptr; }
   size_t GetFreeCodeSpace() const { return (m_code_size - m_code_used); }
   void CommitCode(size_t length);
-  void Reset();
+
+  void* GetFreeFarCodePointer() const { return m_free_far_code_ptr; }
+  size_t GetFreeFarCodeSpace() const { return (m_far_code_size - m_far_code_used); }
+  void CommitFarCode(size_t length);
 
   /// Adjusts the free code pointer to the specified alignment, padding with bytes.
   /// Assumes alignment is a power-of-two.
@@ -21,5 +26,12 @@ private:
   void* m_free_code_ptr;
   size_t m_code_size;
   size_t m_code_used;
+
+  void* m_far_code_ptr;
+  void* m_free_far_code_ptr;
+  size_t m_far_code_size;
+  size_t m_far_code_used;
+
+  size_t m_total_size;
 };
 

src/core/cpu_code_cache.cpp

@@ -9,8 +9,11 @@ Log_SetChannel(CPU::CodeCache);
 
 namespace CPU {
 
-bool USE_CODE_CACHE = true;
+bool USE_CODE_CACHE = false;
-bool USE_RECOMPILER = true;
+bool USE_RECOMPILER = false;
+
+static constexpr size_t RECOMPILER_CODE_CACHE_SIZE = 32 * 1024 * 1024;
+static constexpr size_t RECOMPILER_FAR_CODE_CACHE_SIZE = 32 * 1024 * 1024;
 
 CodeCache::CodeCache() = default;
 
@@ -22,7 +25,7 @@ void CodeCache::Initialize(System* system, Core* core, Bus* bus)
   m_core = core;
   m_bus = bus;
 
-  m_code_buffer = std::make_unique<JitCodeBuffer>();
+  m_code_buffer = std::make_unique<JitCodeBuffer>(RECOMPILER_CODE_CACHE_SIZE, RECOMPILER_FAR_CODE_CACHE_SIZE);
   m_asm_functions = std::make_unique<Recompiler::ASMFunctions>();
   m_asm_functions->Generate(m_code_buffer.get());
 }
@@ -50,7 +53,7 @@ void CodeCache::Execute()
 
 #if 0
     const u32 tick = m_system->GetGlobalTickCounter() + m_core->GetPendingTicks();
-    if (tick == 8950812)
+    if (tick == 58672386)
       __debugbreak();
 #endif
 
@@ -184,6 +187,11 @@ bool CodeCache::CompileBlock(CodeBlock* block)
   bool is_branch_delay_slot = false;
   bool is_load_delay_slot = false;
 
+#if 0
+  if (pc == 0x0005aa90)
+    __debugbreak();
+#endif
+
   for (;;)
   {
     CodeBlockInstruction cbi = {};
@@ -247,7 +255,10 @@ bool CodeCache::CompileBlock(CodeBlock* block)
   if (USE_RECOMPILER)
   {
     // Ensure we're not going to run out of space while compiling this block.
-    if (m_code_buffer->GetFreeCodeSpace() < (block->instructions.size() * Recompiler::MAX_HOST_BYTES_PER_INSTRUCTION))
+    if (m_code_buffer->GetFreeCodeSpace() <
+          (block->instructions.size() * Recompiler::MAX_NEAR_HOST_BYTES_PER_INSTRUCTION) ||
+        m_code_buffer->GetFreeFarCodeSpace() <
+          (block->instructions.size() * Recompiler::MAX_FAR_HOST_BYTES_PER_INSTRUCTION))
     {
       Log_WarningPrintf("Out of code space, flushing all blocks.");
       Reset();

src/core/cpu_recompiler_code_generator.cpp

@@ -7,7 +7,8 @@ namespace CPU::Recompiler {
 
 CodeGenerator::CodeGenerator(Core* cpu, JitCodeBuffer* code_buffer, const ASMFunctions& asm_functions)
   : m_cpu(cpu), m_code_buffer(code_buffer), m_asm_functions(asm_functions), m_register_cache(*this),
-    m_emit(code_buffer->GetFreeCodeSpace(), code_buffer->GetFreeCodePointer())
+    m_near_emitter(code_buffer->GetFreeCodeSpace(), code_buffer->GetFreeCodePointer()),
+    m_far_emitter(code_buffer->GetFreeFarCodeSpace(), code_buffer->GetFreeFarCodePointer()), m_emit(&m_near_emitter)
 {
   InitHostRegs();
 }
@@ -608,7 +609,7 @@ void CodeGenerator::BlockPrologue()
 void CodeGenerator::BlockEpilogue()
 {
 #if defined(_DEBUG) && defined(Y_CPU_X64)
-  m_emit.nop();
+  m_emit->nop();
 #endif
 
   m_register_cache.FlushAllGuestRegisters(true, true);
@@ -632,7 +633,7 @@ void CodeGenerator::InstructionPrologue(const CodeBlockInstruction& cbi, TickCou
                                         bool force_sync /* = false */)
 {
 #if defined(_DEBUG) && defined(Y_CPU_X64)
-  m_emit.nop();
+  m_emit->nop();
 #endif
 
   // reset dirty flags

src/core/cpu_recompiler_code_generator.h

@@ -33,9 +33,6 @@ public:
   static const char* GetHostRegName(HostReg reg, RegSize size = HostPointerSize);
   static void AlignCodeBuffer(JitCodeBuffer* code_buffer);
 
-  RegisterCache& GetRegisterCache() { return m_register_cache; }
-  CodeEmitter& GetCodeEmitter() { return m_emit; }
-
   bool CompileBlock(const CodeBlock* block, CodeBlock::HostCodePointer* out_host_code, u32* out_host_code_size);
 
   //////////////////////////////////////////////////////////////////////////
@@ -146,14 +143,18 @@ private:
   Value ConvertValueSize(const Value& value, RegSize size, bool sign_extend);
   void ConvertValueSizeInPlace(Value* value, RegSize size, bool sign_extend);
 
+  void SwitchToFarCode();
+  void SwitchToNearCode();
+  void* GetCurrentNearCodePointer() const;
+  void* GetCurrentFarCodePointer() const;
+
   //////////////////////////////////////////////////////////////////////////
   // Code Generation Helpers
   //////////////////////////////////////////////////////////////////////////
   // branch target, memory address, etc
   void BlockPrologue();
   void BlockEpilogue();
-  void InstructionPrologue(const CodeBlockInstruction& cbi, TickCount cycles,
+  void InstructionPrologue(const CodeBlockInstruction& cbi, TickCount cycles, bool force_sync = false);
-                           bool force_sync = false);
   void InstructionEpilogue(const CodeBlockInstruction& cbi);
   void SyncCurrentInstructionPC();
   void SyncPC();
@@ -182,7 +183,9 @@ private:
   const CodeBlockInstruction* m_block_start = nullptr;
   const CodeBlockInstruction* m_block_end = nullptr;
   RegisterCache m_register_cache;
-  CodeEmitter m_emit;
+  CodeEmitter m_near_emitter;
+  CodeEmitter m_far_emitter;
+  CodeEmitter* m_emit;
 
   u32 m_delayed_pc_add = 0;
   TickCount m_delayed_cycles_add = 0;
@@ -197,4 +200,4 @@ private:
   bool m_next_load_delay_dirty = false;
 };
 
-} // namespace CPU_X86::Recompiler
+} // namespace CPU::Recompiler

src/core/cpu_recompiler_types.h

@@ -35,7 +35,8 @@ constexpr HostReg HostReg_Invalid = static_cast<HostReg>(HostReg_Count);
 constexpr RegSize HostPointerSize = RegSize_64;
 
 // A reasonable "maximum" number of bytes per instruction.
-constexpr u32 MAX_HOST_BYTES_PER_INSTRUCTION = 128;
+constexpr u32 MAX_NEAR_HOST_BYTES_PER_INSTRUCTION = 64;
+constexpr u32 MAX_FAR_HOST_BYTES_PER_INSTRUCTION = 128;
 
 // Are shifts implicitly masked to 0..31?
 constexpr bool SHIFTS_ARE_IMPLICITLY_MASKED = true;