diff --git a/Source/Core/Core/CoreTiming.cpp b/Source/Core/Core/CoreTiming.cpp
index 0a03329096..d3e766450f 100644
--- a/Source/Core/Core/CoreTiming.cpp
+++ b/Source/Core/Core/CoreTiming.cpp
@@ -380,6 +380,7 @@ void Idle()
     Fifo::FlushGpu();
   }
 
+  PowerPC::UpdatePerformanceMonitor(PowerPC::ppcState.downcount, 0, 0);
   s_idled_cycles += DowncountToCycles(PowerPC::ppcState.downcount);
   PowerPC::ppcState.downcount = 0;
 }
diff --git a/Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.cpp b/Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.cpp
index eea5b63988..17b91a6627 100644
--- a/Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.cpp
+++ b/Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.cpp
@@ -134,6 +134,17 @@ static void EndBlock(UGeckoInstruction data)
 {
   PC = NPC;
   PowerPC::ppcState.downcount -= data.hex;
+  PowerPC::UpdatePerformanceMonitor(data.hex, 0, 0);
+}
+
+static void UpdateNumLoadStoreInstructions(UGeckoInstruction data)
+{
+  PowerPC::UpdatePerformanceMonitor(0, data.hex, 0);
+}
+
+static void UpdateNumFloatingPointInstructions(UGeckoInstruction data)
+{
+  PowerPC::UpdatePerformanceMonitor(0, 0, data.hex);
 }
 
 static void WritePC(UGeckoInstruction data)
@@ -230,6 +241,8 @@ void CachedInterpreter::Jit(u32 address)
   js.firstFPInstructionFound = false;
   js.fifoBytesSinceCheck = 0;
   js.downcountAmount = 0;
+  js.numLoadStoreInst = 0;
+  js.numFloatingPointInst = 0;
   js.curBlock = b;
 
   b->checkedEntry = GetCodePtr();
@@ -240,6 +253,10 @@ void CachedInterpreter::Jit(u32 address)
     PPCAnalyst::CodeOp& op = m_code_buffer[i];
 
     js.downcountAmount += op.opinfo->numCycles;
+    if (op.opinfo->flags & FL_LOADSTORE)
+      ++js.numLoadStoreInst;
+    if (op.opinfo->flags & FL_USE_FPU)
+      ++js.numFloatingPointInst;
 
     if (HandleFunctionHooking(op.address))
       break;
@@ -274,13 +291,19 @@ void CachedInterpreter::Jit(u32 address)
       if (idle_loop)
         m_code.emplace_back(CheckIdle, js.blockStart);
       if (endblock)
+      {
         m_code.emplace_back(EndBlock, js.downcountAmount);
+        m_code.emplace_back(UpdateNumLoadStoreInstructions, js.numLoadStoreInst);
+        m_code.emplace_back(UpdateNumFloatingPointInstructions, js.numFloatingPointInst);
+      }
     }
   }
   if (code_block.m_broken)
   {
     m_code.emplace_back(WriteBrokenBlockNPC, nextPC);
     m_code.emplace_back(EndBlock, js.downcountAmount);
+    m_code.emplace_back(UpdateNumLoadStoreInstructions, js.numLoadStoreInst);
+    m_code.emplace_back(UpdateNumFloatingPointInstructions, js.numFloatingPointInst);
   }
   m_code.emplace_back();
 
diff --git a/Source/Core/Core/PowerPC/Interpreter/Interpreter.cpp b/Source/Core/Core/PowerPC/Interpreter/Interpreter.cpp
index 6702fb798c..8835945416 100644
--- a/Source/Core/Core/PowerPC/Interpreter/Interpreter.cpp
+++ b/Source/Core/Core/PowerPC/Interpreter/Interpreter.cpp
@@ -219,7 +219,11 @@ int Interpreter::SingleStepInner()
   }
 
   UpdatePC();
-  return PPCTables::GetOpInfo(m_prev_inst)->numCycles;
+
+  const GekkoOPInfo* opinfo = PPCTables::GetOpInfo(m_prev_inst);
+  PowerPC::UpdatePerformanceMonitor(opinfo->numCycles, (opinfo->flags & FL_LOADSTORE) != 0,
+                                    (opinfo->flags & FL_USE_FPU) != 0);
+  return opinfo->numCycles;
 }
 
 void Interpreter::SingleStep()
diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit.cpp b/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
index 43b9462600..38e6ae8ed3 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
@@ -244,6 +244,16 @@ void JitArm64::Cleanup()
     BLR(X0);
     SetJumpTarget(exit);
   }
+
+  // SPEED HACK: MMCR0/MMCR1 should be checked at run-time, not at compile time.
+  if (MMCR0.Hex || MMCR1.Hex)
+  {
+    MOVP2R(X30, &PowerPC::UpdatePerformanceMonitor);
+    MOVI2R(X0, js.downcountAmount);
+    MOVI2R(X1, js.numLoadStoreInst);
+    MOVI2R(X2, js.numFloatingPointInst);
+    BLR(X30);
+  }
 }
 
 void JitArm64::DoDownCount()
@@ -620,6 +630,8 @@ void JitArm64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
   js.skipInstructions = 0;
   js.curBlock = b;
   js.carryFlagSet = false;
+  js.numLoadStoreInst = 0;
+  js.numFloatingPointInst = 0;
 
   u8* const start = GetWritableCodePtr();
   b->checkedEntry = start;
@@ -807,6 +819,12 @@ void JitArm64::DoJit(u32 em_address, JitBlock* b, u32 nextPC)
       // If we have a register that will never be used again, flush it.
       gpr.StoreRegisters(~op.gprInUse);
       fpr.StoreRegisters(~op.fprInUse);
+
+      if (opinfo->flags & FL_LOADSTORE)
+        ++js.numLoadStoreInst;
+
+      if (opinfo->flags & FL_USE_FPU)
+        ++js.numFloatingPointInst;
     }
 
     i += js.skipInstructions;