CachedInterpreter: Callback Disassembler

Source/Core/Core/CMakeLists.txt

@@ -479,6 +479,7 @@ add_library(core
   PatchEngine.h
   PowerPC/BreakPoints.cpp
   PowerPC/BreakPoints.h
+  PowerPC/CachedInterpreter/CachedInterpreter_Disassembler.cpp
   PowerPC/CachedInterpreter/CachedInterpreter.cpp
   PowerPC/CachedInterpreter/CachedInterpreter.h
   PowerPC/CachedInterpreter/CachedInterpreterBlockCache.cpp

Source/Core/Core/HLE/HLE.cpp

@@ -195,6 +195,11 @@ u32 GetHookByFunctionAddress(PPCSymbolDB& ppc_symbol_db, u32 address)
   return (symbol && symbol->address == address) ? index : 0;
 }
 
+const char* GetHookNameByIndex(u32 index)
+{
+  return os_patches[index].name;
+}
+
 HookType GetHookTypeByIndex(u32 index)
 {
   return os_patches[index].type;

Source/Core/Core/HLE/HLE.h

@@ -69,6 +69,7 @@ void ExecuteFromJIT(u32 current_pc, u32 hook_index, Core::System& system);
 u32 GetHookByAddress(u32 address);
 // Returns the HLE hook index if the address matches the function start
 u32 GetHookByFunctionAddress(PPCSymbolDB& ppc_symbol_db, u32 address);
+const char* GetHookNameByIndex(u32 index);
 HookType GetHookTypeByIndex(u32 index);
 HookFlag GetHookFlagsByIndex(u32 index);
 

Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.cpp

@@ -370,15 +370,16 @@ bool CachedInterpreter::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
             {interpreter, Interpreter::GetInterpreterOp(op.inst), js.compilerPC, op.inst},
             power_pc,
             js.downcountAmount};
-        Write(op.canEndBlock ? InterpretAndCheckExceptions<true> :
-                               InterpretAndCheckExceptions<false>,
+        Write(op.canEndBlock ? CallbackCast(InterpretAndCheckExceptions<true>) :
+                               CallbackCast(InterpretAndCheckExceptions<false>),
               operands);
       }
       else
       {
         const InterpretOperands operands = {interpreter, Interpreter::GetInterpreterOp(op.inst),
                                             js.compilerPC, op.inst};
-        Write(op.canEndBlock ? Interpret<true> : Interpret<false>, operands);
+        Write(op.canEndBlock ? CallbackCast(Interpret<true>) : CallbackCast(Interpret<false>),
+              operands);
       }
 
       if (op.branchIsIdleLoop)

Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.h

@@ -46,6 +46,8 @@ public:
   void Jit(u32 address, bool clear_cache_and_retry_on_failure);
   bool DoJit(u32 address, JitBlock* b, u32 nextPC);
 
+  static std::size_t Disassemble(const JitBlock& block, std::ostream& stream);
+
   JitBaseBlockCache* GetBlockCache() override { return &m_block_cache; }
   const char* GetName() const override { return "Cached Interpreter"; }
   const CommonAsmRoutinesBase* GetAsmRoutines() override { return nullptr; }
@@ -74,20 +76,33 @@ private:
   struct CheckIdleOperands;
 
   static s32 StartProfiledBlock(PowerPC::PowerPCState& ppc_state,
-                                const StartProfiledBlockOperands& profile_data);
+                                const StartProfiledBlockOperands& operands);
+  static s32 StartProfiledBlock(std::ostream& stream, const StartProfiledBlockOperands& operands);
   template <bool profiled>
   static s32 EndBlock(PowerPC::PowerPCState& ppc_state, const EndBlockOperands<profiled>& operands);
+  template <bool profiled>
+  static s32 EndBlock(std::ostream& stream, const EndBlockOperands<profiled>& operands);
   template <bool write_pc>
   static s32 Interpret(PowerPC::PowerPCState& ppc_state, const InterpretOperands& operands);
   template <bool write_pc>
+  static s32 Interpret(std::ostream& stream, const InterpretOperands& operands);
+  template <bool write_pc>
   static s32 InterpretAndCheckExceptions(PowerPC::PowerPCState& ppc_state,
                                          const InterpretAndCheckExceptionsOperands& operands);
+  template <bool write_pc>
+  static s32 InterpretAndCheckExceptions(std::ostream& stream,
+                                         const InterpretAndCheckExceptionsOperands& operands);
   static s32 HLEFunction(PowerPC::PowerPCState& ppc_state, const HLEFunctionOperands& operands);
+  static s32 HLEFunction(std::ostream& stream, const HLEFunctionOperands& operands);
   static s32 WriteBrokenBlockNPC(PowerPC::PowerPCState& ppc_state,
                                  const WriteBrokenBlockNPCOperands& operands);
+  static s32 WriteBrokenBlockNPC(std::ostream& stream, const WriteBrokenBlockNPCOperands& operands);
   static s32 CheckFPU(PowerPC::PowerPCState& ppc_state, const CheckHaltOperands& operands);
+  static s32 CheckFPU(std::ostream& stream, const CheckHaltOperands& operands);
   static s32 CheckBreakpoint(PowerPC::PowerPCState& ppc_state, const CheckHaltOperands& operands);
+  static s32 CheckBreakpoint(std::ostream& stream, const CheckHaltOperands& operands);
   static s32 CheckIdle(PowerPC::PowerPCState& ppc_state, const CheckIdleOperands& operands);
+  static s32 CheckIdle(std::ostream& stream, const CheckIdleOperands& operands);
 
   HyoutaUtilities::RangeSizeSet<u8*> m_free_ranges;
   CachedInterpreterBlockCache m_block_cache;

Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreterEmitter.h

@@ -4,6 +4,7 @@
 #pragma once
 
 #include <cstddef>
+#include <iosfwd>
 #include <type_traits>
 
 #include "Common/CodeBlock.h"
@@ -24,6 +25,11 @@ protected:
   using Callback = s32 (*)(PowerPC::PowerPCState& ppc_state, const Operands& operands);
   using AnyCallback = s32 (*)(PowerPC::PowerPCState& ppc_state, const void* operands);
 
+  template <class Operands>
+  static consteval Callback<Operands> CallbackCast(Callback<Operands> callback)
+  {
+    return callback;
+  }
   template <class Operands>
   static AnyCallback AnyCallbackCast(Callback<Operands> callback)
   {
@@ -31,6 +37,21 @@ protected:
   }
   static consteval AnyCallback AnyCallbackCast(AnyCallback callback) { return callback; }
 
+  // Disassemble callbacks will always return the distance to the next callback.
+  template <class Operands>
+  using Disassemble = s32 (*)(std::ostream& stream, const Operands& operands);
+  using AnyDisassemble = s32 (*)(std::ostream& stream, const void* operands);
+
+  template <class Operands>
+  static AnyDisassemble AnyDisassembleCast(Disassemble<Operands> disassemble)
+  {
+    return reinterpret_cast<AnyDisassemble>(disassemble);
+  }
+  static consteval AnyDisassemble AnyDisassembleCast(AnyDisassemble disassemble)
+  {
+    return disassemble;
+  }
+
 public:
   CachedInterpreterEmitter() = default;
   explicit CachedInterpreterEmitter(u8* begin, u8* end) : m_code(begin), m_code_end(end) {}
@@ -63,6 +84,7 @@ public:
   }
 
   static s32 PoisonCallback(PowerPC::PowerPCState& ppc_state, const void* operands);
+  static s32 PoisonCallback(std::ostream& stream, const void* operands);
 
 private:
   void Write(AnyCallback callback, const void* operands, std::size_t size);

Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter_Disassembler.cpp

@@ -0,0 +1,143 @@
+// Copyright 2024 Dolphin Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "Core/PowerPC/CachedInterpreter/CachedInterpreter.h"
+
+#include <algorithm>
+#include <array>
+#include <mutex>
+#include <utility>
+
+#include <fmt/format.h>
+#include <fmt/ostream.h>
+
+#include "Core/HLE/HLE.h"
+
+s32 CachedInterpreterEmitter::PoisonCallback(std::ostream& stream, const void* operands)
+{
+  stream << "PoisonCallback()\n";
+  return sizeof(AnyCallback);
+}
+
+s32 CachedInterpreter::StartProfiledBlock(std::ostream& stream,
+                                          const StartProfiledBlockOperands& operands)
+{
+  stream << "StartProfiledBlock()\n";
+  return sizeof(AnyCallback) + sizeof(operands);
+}
+
+template <bool profiled>
+s32 CachedInterpreter::EndBlock(std::ostream& stream, const EndBlockOperands<profiled>& operands)
+{
+  fmt::println(stream, "EndBlock<profiled={}>(downcount={}, num_load_stores={}, num_fp_inst={})",
+               profiled, operands.downcount, operands.num_load_stores, operands.num_fp_inst);
+  return sizeof(AnyCallback) + sizeof(operands);
+}
+
+template <bool write_pc>
+s32 CachedInterpreter::Interpret(std::ostream& stream, const InterpretOperands& operands)
+{
+  fmt::println(stream, "Interpret<write_pc={:5}>(current_pc=0x{:08x}, inst=0x{:08x})", write_pc,
+               operands.current_pc, operands.inst.hex);
+  return sizeof(AnyCallback) + sizeof(operands);
+}
+
+template <bool write_pc>
+s32 CachedInterpreter::InterpretAndCheckExceptions(
+    std::ostream& stream, const InterpretAndCheckExceptionsOperands& operands)
+{
+  fmt::println(stream,
+               "InterpretAndCheckExceptions<write_pc={:5}>(current_pc=0x{:08x}, inst=0x{:08x}, "
+               "downcount={})",
+               write_pc, operands.current_pc, operands.inst.hex, operands.downcount);
+  return sizeof(AnyCallback) + sizeof(operands);
+}
+
+s32 CachedInterpreter::HLEFunction(std::ostream& stream, const HLEFunctionOperands& operands)
+{
+  const auto& [system, current_pc, hook_index] = operands;
+  fmt::println(stream, "HLEFunction(current_pc=0x{:08x}, hook_index={}) [\"{}\"]", current_pc,
+               hook_index, HLE::GetHookNameByIndex(hook_index));
+  return sizeof(AnyCallback) + sizeof(operands);
+}
+
+s32 CachedInterpreter::WriteBrokenBlockNPC(std::ostream& stream,
+                                           const WriteBrokenBlockNPCOperands& operands)
+{
+  const auto& [current_pc] = operands;
+  fmt::println(stream, "WriteBrokenBlockNPC(current_pc=0x{:08x})", current_pc);
+  return sizeof(AnyCallback) + sizeof(operands);
+}
+
+s32 CachedInterpreter::CheckFPU(std::ostream& stream, const CheckHaltOperands& operands)
+{
+  const auto& [power_pc, current_pc, downcount] = operands;
+  fmt::println(stream, "CheckFPU(current_pc=0x{:08x}, downcount={})", current_pc, downcount);
+  return sizeof(AnyCallback) + sizeof(operands);
+}
+
+s32 CachedInterpreter::CheckBreakpoint(std::ostream& stream, const CheckHaltOperands& operands)
+{
+  const auto& [power_pc, current_pc, downcount] = operands;
+  fmt::println(stream, "CheckBreakpoint(current_pc=0x{:08x}, downcount={})", current_pc, downcount);
+  return sizeof(AnyCallback) + sizeof(operands);
+}
+
+s32 CachedInterpreter::CheckIdle(std::ostream& stream, const CheckIdleOperands& operands)
+{
+  const auto& [core_timing, idle_pc] = operands;
+  fmt::println(stream, "CheckIdle(idle_pc=0x{:08x})", idle_pc);
+  return sizeof(AnyCallback) + sizeof(operands);
+}
+
+static std::once_flag s_sorted_lookup_flag;
+
+std::size_t CachedInterpreter::Disassemble(const JitBlock& block, std::ostream& stream)
+{
+  using LookupKV = std::pair<AnyCallback, AnyDisassemble>;
+
+  // clang-format off
+#define LOOKUP_KV(...) {AnyCallbackCast(__VA_ARGS__), AnyDisassembleCast(__VA_ARGS__)}
+  // clang-format on
+
+  // Function addresses aren't known at compile-time, so this array is sorted at run-time.
+  static auto sorted_lookup = std::to_array<LookupKV>({
+      LOOKUP_KV(CachedInterpreter::PoisonCallback),
+      LOOKUP_KV(CachedInterpreter::StartProfiledBlock),
+      LOOKUP_KV(CachedInterpreter::EndBlock<false>),
+      LOOKUP_KV(CachedInterpreter::EndBlock<true>),
+      LOOKUP_KV(CachedInterpreter::Interpret<false>),
+      LOOKUP_KV(CachedInterpreter::Interpret<true>),
+      LOOKUP_KV(CachedInterpreter::InterpretAndCheckExceptions<false>),
+      LOOKUP_KV(CachedInterpreter::InterpretAndCheckExceptions<true>),
+      LOOKUP_KV(CachedInterpreter::HLEFunction),
+      LOOKUP_KV(CachedInterpreter::WriteBrokenBlockNPC),
+      LOOKUP_KV(CachedInterpreter::CheckFPU),
+      LOOKUP_KV(CachedInterpreter::CheckBreakpoint),
+      LOOKUP_KV(CachedInterpreter::CheckIdle),
+  });
+
+#undef LOOKUP_KV
+
+  std::call_once(s_sorted_lookup_flag, []() {
+    const auto end = std::ranges::sort(sorted_lookup, {}, &LookupKV::first);
+    ASSERT_MSG(DYNA_REC, std::ranges::adjacent_find(sorted_lookup, {}, &LookupKV::first) == end,
+               "Sorted lookup should not contain duplicate keys.");
+  });
+
+  std::size_t instruction_count = 0;
+  for (const u8* normal_entry = block.normalEntry; normal_entry != block.near_end;
+       ++instruction_count)
+  {
+    const auto callback = *reinterpret_cast<const AnyCallback*>(normal_entry);
+    const auto kv = std::ranges::lower_bound(sorted_lookup, callback, {}, &LookupKV::first);
+    if (kv != sorted_lookup.end() && kv->first == callback)
+    {
+      normal_entry += kv->second(stream, normal_entry + sizeof(AnyCallback));
+      continue;
+    }
+    stream << "UNKNOWN OR ILLEGAL CALLBACK\n";
+    break;
+  }
+  return instruction_count;
+}

Source/Core/DolphinLib.props

@@ -1089,6 +1089,7 @@
     <ClCompile Include="Core\NetworkCaptureLogger.cpp" />
     <ClCompile Include="Core\PatchEngine.cpp" />
     <ClCompile Include="Core\PowerPC\BreakPoints.cpp" />
+    <ClCompile Include="Core\PowerPC\CachedInterpreter\CachedInterpreter_Disassembler.cpp" />
     <ClCompile Include="Core\PowerPC\CachedInterpreter\CachedInterpreter.cpp" />
     <ClCompile Include="Core\PowerPC\CachedInterpreter\CachedInterpreterBlockCache.cpp" />
     <ClCompile Include="Core\PowerPC\CachedInterpreter\CachedInterpreterEmitter.cpp" />