Merge pull request #12691 from mitaclaw/jit-profiling-restoration

JitCache: Software Profiling Restoration

Source/Android/app/src/main/java/org/dolphinemu/dolphinemu/NativeLibrary.java

@@ -385,16 +385,9 @@ public final class NativeLibrary
   public static native boolean IsRunningAndUnpaused();
 
   /**
-   * Enables or disables CPU block profiling
-   *
-   * @param enable
+   * Writes out the JitBlock Cache log dump
    */
-  public static native void SetProfiling(boolean enable);
-
-  /**
-   * Writes out the block profile results
-   */
-  public static native void WriteProfileResults();
+  public static native void WriteJitBlockLogDump();
 
   /**
    * Native EGL functions not exposed by Java bindings
@@ -454,6 +447,14 @@ public final class NativeLibrary
 
   private static native String GetCurrentTitleDescriptionUnchecked();
 
+  @Keep
+  public static void displayToastMsg(final String text, final boolean long_length)
+  {
+    final int length = long_length ? Toast.LENGTH_LONG : Toast.LENGTH_SHORT;
+    new Handler(Looper.getMainLooper())
+            .post(() -> Toast.makeText(DolphinApplication.getAppContext(), text, length).show());
+  }
+
   @Keep
   public static boolean displayAlertMsg(final String caption, final String text,
           final boolean yesNo, final boolean isWarning, final boolean nonBlocking)
@@ -466,9 +467,7 @@ public final class NativeLibrary
     // and are allowed to block. As a fallback, we can use toasts.
     if (emulationActivity == null || nonBlocking)
     {
-      new Handler(Looper.getMainLooper()).post(
-              () -> Toast.makeText(DolphinApplication.getAppContext(), text, Toast.LENGTH_LONG)
-                      .show());
+      displayToastMsg(text, true);
     }
     else
     {

Source/Android/app/src/main/java/org/dolphinemu/dolphinemu/features/settings/model/BooleanSetting.kt

@@ -214,6 +214,12 @@ enum class BooleanSetting(
         "JitRegisterCacheOff",
         false
     ),
+    MAIN_DEBUG_JIT_ENABLE_PROFILING(
+        Settings.FILE_DOLPHIN,
+        Settings.SECTION_DEBUG,
+        "JitEnableProfiling",
+        false
+    ),
     MAIN_EMULATE_SKYLANDER_PORTAL(
         Settings.FILE_DOLPHIN,
         Settings.SECTION_EMULATED_USB_DEVICES,

Source/Android/app/src/main/java/org/dolphinemu/dolphinemu/features/settings/ui/SettingsFragmentPresenter.kt

@@ -1978,6 +1978,26 @@ class SettingsFragmentPresenter(
             )
         )
 
+        sl.add(HeaderSetting(context, R.string.debug_jit_profiling_header, 0))
+        sl.add(
+            SwitchSetting(
+                context,
+                BooleanSetting.MAIN_DEBUG_JIT_ENABLE_PROFILING,
+                R.string.debug_jit_enable_block_profiling,
+                0
+           )
+        )
+        sl.add(
+            RunRunnable(
+                context,
+                R.string.debug_jit_write_block_log_dump,
+                0,
+                0,
+                0,
+                true
+            ) { NativeLibrary.WriteJitBlockLogDump() }
+        )
+
         sl.add(HeaderSetting(context, R.string.debug_jit_header, 0))
         sl.add(
             SwitchSetting(

Source/Android/app/src/main/res/values/strings.xml

@@ -406,6 +406,9 @@
     <string name="debug_fastmem">Disable Fastmem</string>
     <string name="debug_fastmem_arena">Disable Fastmem Arena</string>
     <string name="debug_large_entry_points_map">Disable Large Entry Points Map</string>
+    <string name="debug_jit_profiling_header">Jit Profiling</string>
+    <string name="debug_jit_enable_block_profiling">Enable Jit Block Profiling</string>
+    <string name="debug_jit_write_block_log_dump">Write Jit Block Log Dump</string>
     <string name="debug_jit_header">Jit</string>
     <string name="debug_jitoff">Jit Disabled</string>
     <string name="debug_jitloadstoreoff">Jit Load Store Disabled</string>

Source/Android/jni/AndroidCommon/IDCache.cpp

@@ -12,6 +12,7 @@ static JavaVM* s_java_vm;
 static jclass s_string_class;
 
 static jclass s_native_library_class;
+static jmethodID s_display_toast_msg;
 static jmethodID s_display_alert_msg;
 static jmethodID s_update_touch_pointer;
 static jmethodID s_on_title_changed;
@@ -146,6 +147,11 @@ jclass GetNativeLibraryClass()
   return s_native_library_class;
 }
 
+jmethodID GetDisplayToastMsg()
+{
+  return s_display_toast_msg;
+}
+
 jmethodID GetDisplayAlertMsg()
 {
   return s_display_alert_msg;
@@ -528,6 +534,8 @@ JNIEXPORT jint JNI_OnLoad(JavaVM* vm, void* reserved)
 
   const jclass native_library_class = env->FindClass("org/dolphinemu/dolphinemu/NativeLibrary");
   s_native_library_class = reinterpret_cast<jclass>(env->NewGlobalRef(native_library_class));
+  s_display_toast_msg =
+      env->GetStaticMethodID(s_native_library_class, "displayToastMsg", "(Ljava/lang/String;Z)V");
   s_display_alert_msg = env->GetStaticMethodID(s_native_library_class, "displayAlertMsg",
                                                "(Ljava/lang/String;Ljava/lang/String;ZZZ)Z");
   s_update_touch_pointer =

Source/Android/jni/AndroidCommon/IDCache.h

@@ -12,6 +12,7 @@ JNIEnv* GetEnvForThread();
 jclass GetStringClass();
 
 jclass GetNativeLibraryClass();
+jmethodID GetDisplayToastMsg();
 jmethodID GetDisplayAlertMsg();
 jmethodID GetUpdateTouchPointer();
 jmethodID GetOnTitleChanged();

Source/Android/jni/MainAndroid.cpp

@@ -6,6 +6,7 @@
 #include <android/native_window_jni.h>
 #include <cstdio>
 #include <cstdlib>
+#include <fmt/format.h>
 #include <jni.h>
 #include <memory>
 #include <mutex>
@@ -22,6 +23,7 @@
 #include "Common/Event.h"
 #include "Common/FileUtil.h"
 #include "Common/Flag.h"
+#include "Common/IOFile.h"
 #include "Common/IniFile.h"
 #include "Common/Logging/LogManager.h"
 #include "Common/MsgHandler.h"
@@ -42,7 +44,6 @@
 #include "Core/Host.h"
 #include "Core/PowerPC/JitInterface.h"
 #include "Core/PowerPC/PowerPC.h"
-#include "Core/PowerPC/Profiler.h"
 #include "Core/State.h"
 #include "Core/System.h"
 
@@ -404,26 +405,34 @@ JNIEXPORT jint JNICALL Java_org_dolphinemu_dolphinemu_NativeLibrary_GetMaxLogLev
   return static_cast<jint>(Common::Log::MAX_LOGLEVEL);
 }
 
-JNIEXPORT void JNICALL Java_org_dolphinemu_dolphinemu_NativeLibrary_SetProfiling(JNIEnv*, jclass,
-                                                                                 jboolean enable)
+JNIEXPORT void JNICALL Java_org_dolphinemu_dolphinemu_NativeLibrary_WriteJitBlockLogDump(
+    JNIEnv* env, jclass native_library_class)
 {
   HostThreadLock guard;
   auto& system = Core::System::GetInstance();
   auto& jit_interface = system.GetJitInterface();
-  const Core::CPUThreadGuard cpu_guard(system);
-  jit_interface.ClearCache(cpu_guard);
-  jit_interface.SetProfilingState(enable ? JitInterface::ProfilingState::Enabled :
-                                           JitInterface::ProfilingState::Disabled);
-}
-
-JNIEXPORT void JNICALL Java_org_dolphinemu_dolphinemu_NativeLibrary_WriteProfileResults(JNIEnv*,
-                                                                                        jclass)
-{
-  HostThreadLock guard;
-  std::string filename = File::GetUserPath(D_DUMP_IDX) + "Debug/profiler.txt";
-  File::CreateFullPath(filename);
-  auto& jit_interface = Core::System::GetInstance().GetJitInterface();
-  jit_interface.WriteProfileResults(filename);
+  if (jit_interface.GetCore() == nullptr)
+  {
+    env->CallStaticVoidMethod(native_library_class, IDCache::GetDisplayToastMsg(),
+                              ToJString(env, Common::GetStringT("JIT is not active")),
+                              static_cast<jboolean>(false));
+    return;
+  }
+  const std::string filename = fmt::format("{}{}.txt", File::GetUserPath(D_DUMPDEBUG_JITBLOCKS_IDX),
+                                           SConfig::GetInstance().GetGameID());
+  File::IOFile f(filename, "w");
+  if (!f)
+  {
+    env->CallStaticVoidMethod(
+        native_library_class, IDCache::GetDisplayToastMsg(),
+        ToJString(env, Common::FmtFormatT("Failed to open \"{0}\" for writing.", filename)),
+        static_cast<jboolean>(false));
+    return;
+  }
+  jit_interface.JitBlockLogDump(Core::CPUThreadGuard{system}, f.GetHandle());
+  env->CallStaticVoidMethod(native_library_class, IDCache::GetDisplayToastMsg(),
+                            ToJString(env, Common::FmtFormatT("Wrote to \"{0}\".", filename)),
+                            static_cast<jboolean>(false));
 }
 
 // Surface Handling

Source/Core/Common/CMakeLists.txt

@@ -108,8 +108,6 @@ add_library(common
   Network.h
   PcapFile.cpp
   PcapFile.h
-  PerformanceCounter.cpp
-  PerformanceCounter.h
   Profiler.cpp
   Profiler.h
   QoSSession.cpp

Source/Core/Common/CommonPaths.h

@@ -77,6 +77,7 @@
 #define DUMP_SSL_DIR "SSL"
 #define DUMP_DEBUG_DIR "Debug"
 #define DUMP_DEBUG_BRANCHWATCH_DIR "BranchWatch"
+#define DUMP_DEBUG_JITBLOCKS_DIR "JitBlocks"
 #define LOGS_DIR "Logs"
 #define MAIL_LOGS_DIR "Mail"
 #define SHADERS_DIR "Shaders"

Source/Core/Common/FileUtil.cpp

@@ -859,6 +859,8 @@ static void RebuildUserDirectories(unsigned int dir_index)
     s_user_paths[D_DUMPDEBUG_IDX] = s_user_paths[D_DUMP_IDX] + DUMP_DEBUG_DIR DIR_SEP;
     s_user_paths[D_DUMPDEBUG_BRANCHWATCH_IDX] =
         s_user_paths[D_DUMPDEBUG_IDX] + DUMP_DEBUG_BRANCHWATCH_DIR DIR_SEP;
+    s_user_paths[D_DUMPDEBUG_JITBLOCKS_IDX] =
+        s_user_paths[D_DUMPDEBUG_IDX] + DUMP_DEBUG_JITBLOCKS_DIR DIR_SEP;
     s_user_paths[D_LOGS_IDX] = s_user_paths[D_USER_IDX] + LOGS_DIR DIR_SEP;
     s_user_paths[D_MAILLOGS_IDX] = s_user_paths[D_LOGS_IDX] + MAIL_LOGS_DIR DIR_SEP;
     s_user_paths[D_THEMES_IDX] = s_user_paths[D_USER_IDX] + THEMES_DIR DIR_SEP;
@@ -938,6 +940,8 @@ static void RebuildUserDirectories(unsigned int dir_index)
     s_user_paths[D_DUMPDEBUG_IDX] = s_user_paths[D_DUMP_IDX] + DUMP_DEBUG_DIR DIR_SEP;
     s_user_paths[D_DUMPDEBUG_BRANCHWATCH_IDX] =
         s_user_paths[D_DUMPDEBUG_IDX] + DUMP_DEBUG_BRANCHWATCH_DIR DIR_SEP;
+    s_user_paths[D_DUMPDEBUG_JITBLOCKS_IDX] =
+        s_user_paths[D_DUMPDEBUG_IDX] + DUMP_DEBUG_JITBLOCKS_DIR DIR_SEP;
     s_user_paths[F_MEM1DUMP_IDX] = s_user_paths[D_DUMP_IDX] + MEM1_DUMP;
     s_user_paths[F_MEM2DUMP_IDX] = s_user_paths[D_DUMP_IDX] + MEM2_DUMP;
     s_user_paths[F_ARAMDUMP_IDX] = s_user_paths[D_DUMP_IDX] + ARAM_DUMP;

Source/Core/Common/FileUtil.h

@@ -54,6 +54,7 @@ enum
   D_DUMPSSL_IDX,
   D_DUMPDEBUG_IDX,
   D_DUMPDEBUG_BRANCHWATCH_IDX,
+  D_DUMPDEBUG_JITBLOCKS_IDX,
   D_LOAD_IDX,
   D_LOGS_IDX,
   D_MAILLOGS_IDX,

Source/Core/Common/PerformanceCounter.cpp

@@ -1,47 +0,0 @@
-// Copyright 2014 Dolphin Emulator Project
-// SPDX-License-Identifier: GPL-2.0-or-later
-
-#if !defined(_WIN32)
-#include "Common/PerformanceCounter.h"
-
-#include <cstdint>
-#include <ctime>
-
-#include <unistd.h>
-
-#include "Common/CommonTypes.h"
-
-#if defined(_POSIX_TIMERS) && _POSIX_TIMERS > 0
-#if defined(_POSIX_MONOTONIC_CLOCK) && _POSIX_MONOTONIC_CLOCK > 0
-#define DOLPHIN_CLOCK CLOCK_MONOTONIC
-#else
-#define DOLPHIN_CLOCK CLOCK_REALTIME
-#endif
-#endif
-
-bool QueryPerformanceCounter(u64* out)
-{
-#if defined(_POSIX_TIMERS) && _POSIX_TIMERS > 0
-  timespec tp;
-  if (clock_gettime(DOLPHIN_CLOCK, &tp))
-    return false;
-  *out = (u64)tp.tv_nsec + (u64)1000000000 * (u64)tp.tv_sec;
-  return true;
-#else
-  *out = 0;
-  return false;
-#endif
-}
-
-bool QueryPerformanceFrequency(u64* out)
-{
-#if defined(_POSIX_TIMERS) && _POSIX_TIMERS > 0
-  *out = 1000000000;
-  return true;
-#else
-  *out = 1;
-  return false;
-#endif
-}
-
-#endif

Source/Core/Common/PerformanceCounter.h

@@ -1,16 +0,0 @@
-// Copyright 2014 Dolphin Emulator Project
-// SPDX-License-Identifier: GPL-2.0-or-later
-
-#pragma once
-
-#if !defined(_WIN32)
-
-#include <cstdint>
-
-#include "Common/CommonTypes.h"
-
-typedef u64 LARGE_INTEGER;
-bool QueryPerformanceCounter(u64* out);
-bool QueryPerformanceFrequency(u64* lpFrequency);
-
-#endif

Source/Core/Core/CMakeLists.txt

@@ -523,7 +523,6 @@ add_library(core
   PowerPC/PPCSymbolDB.h
   PowerPC/PPCTables.cpp
   PowerPC/PPCTables.h
-  PowerPC/Profiler.h
   PowerPC/SignatureDB/CSVSignatureDB.cpp
   PowerPC/SignatureDB/CSVSignatureDB.h
   PowerPC/SignatureDB/DSYSignatureDB.cpp

Source/Core/Core/Config/MainSettings.cpp

@@ -509,6 +509,8 @@ const Info<bool> MAIN_DEBUG_JIT_SYSTEM_REGISTERS_OFF{
 const Info<bool> MAIN_DEBUG_JIT_BRANCH_OFF{{System::Main, "Debug", "JitBranchOff"}, false};
 const Info<bool> MAIN_DEBUG_JIT_REGISTER_CACHE_OFF{{System::Main, "Debug", "JitRegisterCacheOff"},
                                                    false};
+const Info<bool> MAIN_DEBUG_JIT_ENABLE_PROFILING{{System::Main, "Debug", "JitEnableProfiling"},
+                                                 false};
 
 // Main.BluetoothPassthrough
 

Source/Core/Core/Config/MainSettings.h

@@ -335,6 +335,7 @@ extern const Info<bool> MAIN_DEBUG_JIT_PAIRED_OFF;
 extern const Info<bool> MAIN_DEBUG_JIT_SYSTEM_REGISTERS_OFF;
 extern const Info<bool> MAIN_DEBUG_JIT_BRANCH_OFF;
 extern const Info<bool> MAIN_DEBUG_JIT_REGISTER_CACHE_OFF;
+extern const Info<bool> MAIN_DEBUG_JIT_ENABLE_PROFILING;
 
 // Main.BluetoothPassthrough
 

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

@@ -316,7 +316,7 @@ void CachedInterpreter::Jit(u32 address)
   js.numFloatingPointInst = 0;
   js.curBlock = b;
 
-  b->normalEntry = GetCodePtr();
+  b->normalEntry = b->near_begin = GetCodePtr();
 
   for (u32 i = 0; i < code_block.m_num_instructions; i++)
   {
@@ -382,6 +382,10 @@ void CachedInterpreter::Jit(u32 address)
   }
   m_code.emplace_back();
 
+  b->near_end = GetCodePtr();
+  b->far_begin = nullptr;
+  b->far_end = nullptr;
+
   b->codeSize = static_cast<u32>(GetCodePtr() - b->normalEntry);
   b->originalSize = code_block.m_num_instructions;
 

Source/Core/Core/PowerPC/Gekko.h

@@ -931,6 +931,7 @@ enum CPUEmuFeatureFlags : u32
   FEATURE_FLAG_MSR_DR = 1 << 0,
   FEATURE_FLAG_MSR_IR = 1 << 1,
   FEATURE_FLAG_PERFMON = 1 << 2,
+  FEATURE_FLAG_END_OF_ENUMERATION,
 };
 
 constexpr s32 SignExt16(s16 x)

Source/Core/Core/PowerPC/Jit64/Jit.cpp

@@ -19,7 +19,6 @@
 #include "Common/GekkoDisassembler.h"
 #include "Common/IOFile.h"
 #include "Common/Logging/Log.h"
-#include "Common/PerformanceCounter.h"
 #include "Common/StringUtil.h"
 #include "Common/Swap.h"
 #include "Common/x64ABI.h"
@@ -43,7 +42,6 @@
 #include "Core/PowerPC/MMU.h"
 #include "Core/PowerPC/PPCAnalyst.h"
 #include "Core/PowerPC/PowerPC.h"
-#include "Core/PowerPC/Profiler.h"
 #include "Core/System.h"
 
 using namespace Gen;
@@ -454,20 +452,11 @@ bool Jit64::Cleanup()
     did_something = true;
   }
 
-  if (jo.profile_blocks)
+  if (IsProfilingEnabled())
   {
     ABI_PushRegistersAndAdjustStack({}, 0);
-    // get end tic
-    MOV(64, R(ABI_PARAM1), ImmPtr(&js.curBlock->profile_data.ticStop));
-    ABI_CallFunction(QueryPerformanceCounter);
-    // tic counter += (end tic - start tic)
-    MOV(64, R(RSCRATCH2), ImmPtr(&js.curBlock->profile_data));
-    MOV(64, R(RSCRATCH), MDisp(RSCRATCH2, offsetof(JitBlock::ProfileData, ticStop)));
-    SUB(64, R(RSCRATCH), MDisp(RSCRATCH2, offsetof(JitBlock::ProfileData, ticStart)));
-    ADD(64, R(RSCRATCH), MDisp(RSCRATCH2, offsetof(JitBlock::ProfileData, ticCounter)));
-    ADD(64, MDisp(RSCRATCH2, offsetof(JitBlock::ProfileData, downcountCounter)),
-        Imm32(js.downcountAmount));
-    MOV(64, MDisp(RSCRATCH2, offsetof(JitBlock::ProfileData, ticCounter)), R(RSCRATCH));
+    ABI_CallFunctionPC(&JitBlock::ProfileData::EndProfiling, js.curBlock->profile_data.get(),
+                       js.downcountAmount);
     ABI_PopRegistersAndAdjustStack({}, 0);
     did_something = true;
   }
@@ -773,7 +762,7 @@ void Jit64::Jit(u32 em_address, bool clear_cache_and_retry_on_failure)
     EnableBlockLink();
     EnableOptimization();
 
-    if (!jo.profile_blocks)
+    if (!IsProfilingEnabled())
     {
       if (m_system.GetCPU().IsStepping())
       {
@@ -899,15 +888,9 @@ bool Jit64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
   }
 
   // Conditionally add profiling code.
-  if (jo.profile_blocks)
-  {
-    // get start tic
-    MOV(64, R(ABI_PARAM1), ImmPtr(&b->profile_data.ticStart));
-    int offset = static_cast<int>(offsetof(JitBlock::ProfileData, runCount)) -
-                 static_cast<int>(offsetof(JitBlock::ProfileData, ticStart));
-    ADD(64, MDisp(ABI_PARAM1, offset), Imm8(1));
-    ABI_CallFunction(QueryPerformanceCounter);
-  }
+  if (IsProfilingEnabled())
+    ABI_CallFunctionP(&JitBlock::ProfileData::BeginProfiling, b->profile_data.get());
+
 #if defined(_DEBUG) || defined(DEBUGFAST) || defined(NAN_CHECK)
   // should help logged stack-traces become more accurate
   MOV(32, PPCSTATE(pc), Imm32(js.blockStart));

Source/Core/Core/PowerPC/JitArm64/Jit.cpp

@@ -10,7 +10,6 @@
 #include "Common/Logging/Log.h"
 #include "Common/MathUtil.h"
 #include "Common/MsgHandler.h"
-#include "Common/PerformanceCounter.h"
 #include "Common/StringUtil.h"
 
 #include "Core/ConfigManager.h"
@@ -26,7 +25,6 @@
 #include "Core/PowerPC/JitArm64/JitArm64_RegCache.h"
 #include "Core/PowerPC/JitInterface.h"
 #include "Core/PowerPC/PowerPC.h"
-#include "Core/PowerPC/Profiler.h"
 #include "Core/System.h"
 
 using namespace Arm64Gen;
@@ -408,7 +406,11 @@ void JitArm64::WriteExit(u32 destination, bool LK, u32 exit_address_after_return
                          ARM64Reg exit_address_after_return_reg)
 {
   Cleanup();
-  EndTimeProfile(js.curBlock);
+  if (IsProfilingEnabled())
+  {
+    ABI_CallFunction(&JitBlock::ProfileData::EndProfiling, js.curBlock->profile_data.get(),
+                     js.downcountAmount);
+  }
   DoDownCount();
 
   LK &= m_enable_blr_optimization;
@@ -509,7 +511,11 @@ void JitArm64::WriteExit(Arm64Gen::ARM64Reg dest, bool LK, u32 exit_address_afte
     MOV(DISPATCHER_PC, dest);
 
   Cleanup();
-  EndTimeProfile(js.curBlock);
+  if (IsProfilingEnabled())
+  {
+    ABI_CallFunction(&JitBlock::ProfileData::EndProfiling, js.curBlock->profile_data.get(),
+                     js.downcountAmount);
+  }
   DoDownCount();
 
   LK &= m_enable_blr_optimization;
@@ -672,7 +678,11 @@ void JitArm64::WriteBLRExit(Arm64Gen::ARM64Reg dest)
     MOV(DISPATCHER_PC, dest);
 
   Cleanup();
-  EndTimeProfile(js.curBlock);
+  if (IsProfilingEnabled())
+  {
+    ABI_CallFunction(&JitBlock::ProfileData::EndProfiling, js.curBlock->profile_data.get(),
+                     js.downcountAmount);
+  }
 
   // Check if {PPC_PC, feature_flags} matches the current state, then RET to ARM_PC.
   LDP(IndexType::Post, ARM64Reg::X2, ARM64Reg::X1, ARM64Reg::SP, 16);
@@ -736,7 +746,11 @@ void JitArm64::WriteExceptionExit(ARM64Reg dest, bool only_external, bool always
   if (!always_exception)
     SetJumpTarget(no_exceptions);
 
-  EndTimeProfile(js.curBlock);
+  if (IsProfilingEnabled())
+  {
+    ABI_CallFunction(&JitBlock::ProfileData::EndProfiling, js.curBlock->profile_data.get(),
+                     js.downcountAmount);
+  }
   DoDownCount();
 
   B(dispatcher);
@@ -804,44 +818,6 @@ void JitArm64::DumpCode(const u8* start, const u8* end)
   WARN_LOG_FMT(DYNA_REC, "Code dump from {} to {}:\n{}", fmt::ptr(start), fmt::ptr(end), output);
 }
 
-void JitArm64::BeginTimeProfile(JitBlock* b)
-{
-  MOVP2R(ARM64Reg::X0, &b->profile_data);
-  LDR(IndexType::Unsigned, ARM64Reg::X1, ARM64Reg::X0, offsetof(JitBlock::ProfileData, runCount));
-  ADD(ARM64Reg::X1, ARM64Reg::X1, 1);
-
-  // Fetch the current counter register
-  CNTVCT(ARM64Reg::X2);
-
-  // stores runCount and ticStart
-  STP(IndexType::Signed, ARM64Reg::X1, ARM64Reg::X2, ARM64Reg::X0,
-      offsetof(JitBlock::ProfileData, runCount));
-}
-
-void JitArm64::EndTimeProfile(JitBlock* b)
-{
-  if (!jo.profile_blocks)
-    return;
-
-  // Fetch the current counter register
-  CNTVCT(ARM64Reg::X1);
-
-  MOVP2R(ARM64Reg::X0, &b->profile_data);
-
-  LDR(IndexType::Unsigned, ARM64Reg::X2, ARM64Reg::X0, offsetof(JitBlock::ProfileData, ticStart));
-  SUB(ARM64Reg::X1, ARM64Reg::X1, ARM64Reg::X2);
-
-  // loads ticCounter and downcountCounter
-  LDP(IndexType::Signed, ARM64Reg::X2, ARM64Reg::X3, ARM64Reg::X0,
-      offsetof(JitBlock::ProfileData, ticCounter));
-  ADD(ARM64Reg::X2, ARM64Reg::X2, ARM64Reg::X1);
-  ADDI2R(ARM64Reg::X3, ARM64Reg::X3, js.downcountAmount, ARM64Reg::X1);
-
-  // stores ticCounter and downcountCounter
-  STP(IndexType::Signed, ARM64Reg::X2, ARM64Reg::X3, ARM64Reg::X0,
-      offsetof(JitBlock::ProfileData, ticCounter));
-}
-
 void JitArm64::Run()
 {
   ProtectStack();
@@ -933,7 +909,7 @@ void JitArm64::Jit(u32 em_address, bool clear_cache_and_retry_on_failure)
     SetBlockLinkingEnabled(true);
     SetOptimizationEnabled(true);
 
-    if (!jo.profile_blocks)
+    if (!IsProfilingEnabled())
     {
       if (cpu.IsStepping())
       {
@@ -1052,11 +1028,8 @@ bool JitArm64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
   b->normalEntry = GetWritableCodePtr();
 
   // Conditionally add profiling code.
-  if (jo.profile_blocks)
-  {
-    // get start tic
-    BeginTimeProfile(b);
-  }
+  if (IsProfilingEnabled())
+    ABI_CallFunction(&JitBlock::ProfileData::BeginProfiling, b->profile_data.get());
 
   if (code_block.m_gqr_used.Count() == 1 &&
       js.pairedQuantizeAddresses.find(js.blockStart) == js.pairedQuantizeAddresses.end())
@@ -1246,7 +1219,11 @@ bool JitArm64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
         FixupBranch no_breakpoint = CBZ(ARM64Reg::W0);
 
         Cleanup();
-        EndTimeProfile(js.curBlock);
+        if (IsProfilingEnabled())
+        {
+          ABI_CallFunction(&JitBlock::ProfileData::EndProfiling, b->profile_data.get(),
+                           js.downcountAmount);
+        }
         DoDownCount();
         B(dispatcher_exit);