diff --git a/Source/Core/Common/Arm64Emitter.cpp b/Source/Core/Common/Arm64Emitter.cpp
index bd72df2e94..ed1fa2b626 100644
--- a/Source/Core/Common/Arm64Emitter.cpp
+++ b/Source/Core/Common/Arm64Emitter.cpp
@@ -1218,6 +1218,14 @@ void ARM64XEmitter::MRS(ARM64Reg Rt, PStateField field)
   EncodeSystemInst(o0 | 4, op1, CRn, CRm, op2, DecodeReg(Rt));
 }
 
+void ARM64XEmitter::CNTVCT(Arm64Gen::ARM64Reg Rt)
+{
+  _assert_msg_(DYNA_REC, Is64Bit(Rt), "CNTVCT: Rt must be 64-bit");
+
+  // MRS <Xt>, CNTVCT_EL0 ; Read CNTVCT_EL0 into Xt
+  EncodeSystemInst(3 | 4, 3, 0xe, 0, 2, DecodeReg(Rt));
+}
+
 void ARM64XEmitter::HINT(SystemHint op)
 {
   EncodeSystemInst(0, 3, 2, 0, op, WSP);
diff --git a/Source/Core/Common/Arm64Emitter.h b/Source/Core/Common/Arm64Emitter.h
index 4b5bc9f137..2f49795519 100644
--- a/Source/Core/Common/Arm64Emitter.h
+++ b/Source/Core/Common/Arm64Emitter.h
@@ -603,9 +603,9 @@ public:
 
   // System
   void _MSR(PStateField field, u8 imm);
-
   void _MSR(PStateField field, ARM64Reg Rt);
   void MRS(ARM64Reg Rt, PStateField field);
+  void CNTVCT(ARM64Reg Rt);
 
   void HINT(SystemHint op);
   void CLREX();
diff --git a/Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.cpp b/Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.cpp
index 0067a7f3f4..ff7739905c 100644
--- a/Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.cpp
+++ b/Source/Core/Core/PowerPC/CachedInterpreter/CachedInterpreter.cpp
@@ -200,7 +200,6 @@ void CachedInterpreter::Jit(u32 address)
 
   b->checkedEntry = GetCodePtr();
   b->normalEntry = GetCodePtr();
-  b->runCount = 0;
 
   for (u32 i = 0; i < code_block.m_num_instructions; i++)
   {
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit.cpp b/Source/Core/Core/PowerPC/Jit64/Jit.cpp
index 9f45d150f6..1a8cd83972 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit.cpp
@@ -16,6 +16,7 @@
 #include "Common/File.h"
 #include "Common/Logging/Log.h"
 #include "Common/MemoryUtil.h"
+#include "Common/PerformanceCounter.h"
 #include "Common/StringUtil.h"
 #include "Common/x64ABI.h"
 #include "Core/Core.h"
@@ -370,6 +371,23 @@ bool Jit64::Cleanup()
     did_something = true;
   }
 
+  if (Profiler::g_ProfileBlocks)
+  {
+    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_PopRegistersAndAdjustStack({}, 0);
+  }
+
   return did_something;
 }
 
@@ -627,7 +645,6 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer* code_buf, JitBloc
   const u8* start =
       AlignCode4();  // TODO: Test if this or AlignCode16 make a difference from GetCodePtr
   b->checkedEntry = start;
-  b->runCount = 0;
 
   // Downcount flag check. The last block decremented downcounter, and the flag should still be
   // available.
@@ -650,13 +667,12 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer* code_buf, JitBloc
   // Conditionally add profiling code.
   if (Profiler::g_ProfileBlocks)
   {
-    MOV(64, R(RSCRATCH), ImmPtr(&b->runCount));
-    ADD(32, MatR(RSCRATCH), Imm8(1));
-    b->ticCounter = 0;
-    b->ticStart = 0;
-    b->ticStop = 0;
     // get start tic
-    PROFILER_QUERY_PERFORMANCE_COUNTER(&b->ticStart);
+    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 defined(_DEBUG) || defined(DEBUGFAST) || defined(NAN_CHECK)
   // should help logged stack-traces become more accurate
@@ -731,16 +747,6 @@ const u8* Jit64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer* code_buf, JitBloc
 
     if (i == (code_block.m_num_instructions - 1))
     {
-      if (Profiler::g_ProfileBlocks)
-      {
-        // WARNING - cmp->branch merging will screw this up.
-        PROFILER_VPUSH;
-        // get end tic
-        PROFILER_QUERY_PERFORMANCE_COUNTER(&b->ticStop);
-        // tic counter += (end tic - start tic)
-        PROFILER_UPDATE_TIME(b);
-        PROFILER_VPOP;
-      }
       js.isLastInstruction = true;
     }
 
diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit.cpp b/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
index a507485043..4b27c13f00 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit.cpp
@@ -36,15 +36,6 @@ constexpr size_t SAFE_STACK_SIZE = 512 * 1024;
 constexpr size_t GUARD_SIZE = 0x10000;  // two guards - bottom (permanent) and middle (see above)
 constexpr size_t GUARD_OFFSET = STACK_SIZE - SAFE_STACK_SIZE - GUARD_SIZE;
 
-static bool HasCycleCounters()
-{
-  // Bit needs to be set to support cycle counters
-  const u32 PMUSERENR_CR = 0x4;
-  u32 reg;
-  asm("mrs %[val], PMUSERENR_EL0" : [val] "=r"(reg));
-  return !!(reg & PMUSERENR_CR);
-}
-
 void JitArm64::Init()
 {
   InitializeInstructionTables();
@@ -72,8 +63,6 @@ void JitArm64::Init()
 
   AllocStack();
   GenerateAsm();
-
-  m_supports_cycle_counter = HasCycleCounters();
 }
 
 bool JitArm64::HandleFault(uintptr_t access_address, SContext* ctx)
@@ -172,6 +161,7 @@ void JitArm64::FallBackToInterpreter(UGeckoInstruction inst)
       ARM64Reg WA = gpr.GetReg();
       LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(npc));
       WriteExceptionExit(WA);
+      gpr.Unlock(WA);
     }
     else
     {
@@ -185,6 +175,7 @@ void JitArm64::FallBackToInterpreter(UGeckoInstruction inst)
       FixupBranch c = B(CC_EQ);
       WriteExceptionExit(WA);
       SetJumpTarget(c);
+      gpr.Unlock(WA);
     }
   }
 
@@ -222,6 +213,7 @@ void JitArm64::HLEFunction(UGeckoInstruction inst)
   ARM64Reg WA = gpr.GetReg();
   LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(npc));
   WriteExit(WA);
+  gpr.Unlock(WA);
 }
 
 void JitArm64::DoNothing(UGeckoInstruction inst)
@@ -239,21 +231,16 @@ void JitArm64::Cleanup()
 {
   if (jo.optimizeGatherPipe && js.fifoBytesSinceCheck > 0)
   {
-    gpr.Lock(W0);
     MOVP2R(X0, &GPFifo::FastCheckGatherPipe);
     BLR(X0);
-    gpr.Unlock(W0);
   }
 }
 
 void JitArm64::DoDownCount()
 {
-  ARM64Reg WA = gpr.GetReg();
-  LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(downcount));
-  ARM64Reg WB = gpr.GetReg();
-  SUBSI2R(WA, WA, js.downcountAmount, WB);
-  STR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(downcount));
-  gpr.Unlock(WA, WB);
+  LDR(INDEX_UNSIGNED, W0, PPC_REG, PPCSTATE_OFF(downcount));
+  SUBSI2R(W0, W0, js.downcountAmount, W1);
+  STR(INDEX_UNSIGNED, W0, PPC_REG, PPCSTATE_OFF(downcount));
 }
 
 void JitArm64::ResetStack()
@@ -303,9 +290,7 @@ void JitArm64::WriteExit(u32 destination, bool LK, u32 exit_address_after_return
 {
   Cleanup();
   DoDownCount();
-
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
+  EndTimeProfile(js.curBlock);
 
   LK &= m_enable_blr_optimization;
 
@@ -342,17 +327,14 @@ void JitArm64::WriteExit(u32 destination, bool LK, u32 exit_address_after_return
 
 void JitArm64::WriteExit(Arm64Gen::ARM64Reg dest, bool LK, u32 exit_address_after_return)
 {
-  Cleanup();
-  DoDownCount();
-
-  LK &= m_enable_blr_optimization;
-
   if (dest != DISPATCHER_PC)
     MOV(DISPATCHER_PC, dest);
-  gpr.Unlock(dest);
 
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
+  Cleanup();
+  DoDownCount();
+  EndTimeProfile(js.curBlock);
+
+  LK &= m_enable_blr_optimization;
 
   if (!LK)
   {
@@ -418,35 +400,28 @@ void JitArm64::WriteBLRExit(Arm64Gen::ARM64Reg dest)
     return;
   }
 
+  if (dest != DISPATCHER_PC)
+    MOV(DISPATCHER_PC, dest);
+
   Cleanup();
-
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
-
-  ARM64Reg code = gpr.GetReg();
-  ARM64Reg pc = gpr.GetReg();
+  EndTimeProfile(js.curBlock);
 
   // Check if {ARM_PC, PPC_PC} matches the current state.
-  LDP(INDEX_POST, EncodeRegTo64(code), EncodeRegTo64(pc), SP, 16);
-  CMP(pc, dest);
+  LDP(INDEX_POST, X2, X1, SP, 16);
+  CMP(W1, DISPATCHER_PC);
   FixupBranch no_match = B(CC_NEQ);
 
-  DoDownCount();
+  DoDownCount();  // overwrites X0 + X1
 
-  RET(EncodeRegTo64(code));
+  RET(X2);
 
   SetJumpTarget(no_match);
 
   DoDownCount();
 
-  if (dest != DISPATCHER_PC)
-    MOV(DISPATCHER_PC, dest);
-
   ResetStack();
 
   B(dispatcher);
-
-  gpr.Unlock(dest, pc, code);
 }
 
 void JitArm64::WriteExceptionExit(u32 destination, bool only_external)
@@ -469,39 +444,34 @@ void JitArm64::WriteExceptionExit(u32 destination, bool only_external)
 
   SetJumpTarget(no_exceptions);
 
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
+  EndTimeProfile(js.curBlock);
 
   B(dispatcher);
 }
 
 void JitArm64::WriteExceptionExit(ARM64Reg dest, bool only_external)
 {
+  if (dest != DISPATCHER_PC)
+    MOV(DISPATCHER_PC, dest);
+
   Cleanup();
   DoDownCount();
 
-  ARM64Reg WA = gpr.GetReg();
-  LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(Exceptions));
-  FixupBranch no_exceptions = CBZ(WA);
-  gpr.Unlock(WA);
+  LDR(INDEX_UNSIGNED, W30, PPC_REG, PPCSTATE_OFF(Exceptions));
+  FixupBranch no_exceptions = CBZ(W30);
 
-  STR(INDEX_UNSIGNED, dest, PPC_REG, PPCSTATE_OFF(pc));
-  STR(INDEX_UNSIGNED, dest, PPC_REG, PPCSTATE_OFF(npc));
+  STR(INDEX_UNSIGNED, DISPATCHER_PC, PPC_REG, PPCSTATE_OFF(pc));
+  STR(INDEX_UNSIGNED, DISPATCHER_PC, PPC_REG, PPCSTATE_OFF(npc));
   if (only_external)
-    MOVP2R(EncodeRegTo64(dest), &PowerPC::CheckExternalExceptions);
+    MOVP2R(EncodeRegTo64(DISPATCHER_PC), &PowerPC::CheckExternalExceptions);
   else
-    MOVP2R(EncodeRegTo64(dest), &PowerPC::CheckExceptions);
-  BLR(EncodeRegTo64(dest));
-  LDR(INDEX_UNSIGNED, dest, PPC_REG, PPCSTATE_OFF(npc));
+    MOVP2R(EncodeRegTo64(DISPATCHER_PC), &PowerPC::CheckExceptions);
+  BLR(EncodeRegTo64(DISPATCHER_PC));
+  LDR(INDEX_UNSIGNED, DISPATCHER_PC, PPC_REG, PPCSTATE_OFF(npc));
 
   SetJumpTarget(no_exceptions);
 
-  if (dest != DISPATCHER_PC)
-    MOV(DISPATCHER_PC, dest);
-  gpr.Unlock(dest);
-
-  if (Profiler::g_ProfileBlocks)
-    EndTimeProfile(js.curBlock);
+  EndTimeProfile(js.curBlock);
 
   B(dispatcher);
 }
@@ -514,66 +484,39 @@ void JitArm64::DumpCode(const u8* start, const u8* end)
   WARN_LOG(DYNA_REC, "Code dump from %p to %p:\n%s", start, end, output.c_str());
 }
 
-void JitArm64::EmitResetCycleCounters()
-{
-  const u32 PMCR_EL0_E = 1;
-  const u32 PMCR_EL0_P = 2;
-  const u32 PMCR_EL0_C = 4;
-  const u32 PMCR_EL0_LC = 0x40;
-  _MSR(FIELD_PMCR_EL0, X0);
-  MOVI2R(X1, PMCR_EL0_E | PMCR_EL0_P | PMCR_EL0_C | PMCR_EL0_LC);
-  ORR(X0, X0, X1);
-  MRS(X0, FIELD_PMCR_EL0);
-}
-
-void JitArm64::EmitGetCycles(Arm64Gen::ARM64Reg reg)
-{
-  _MSR(FIELD_PMCCNTR_EL0, reg);
-}
-
 void JitArm64::BeginTimeProfile(JitBlock* b)
 {
-  b->ticCounter = 0;
-  b->ticStart = 0;
-  b->ticStop = 0;
+  MOVP2R(X0, &b->profile_data);
+  LDR(INDEX_UNSIGNED, X1, X0, offsetof(JitBlock::ProfileData, runCount));
+  ADD(X1, X1, 1);
 
-  if (m_supports_cycle_counter)
-  {
-    EmitResetCycleCounters();
-    EmitGetCycles(X1);
-    MOVP2R(X0, &b->ticStart);
-    STR(INDEX_UNSIGNED, X1, X0, 0);
-  }
-  else
-  {
-    MOVP2R(X1, &QueryPerformanceCounter);
-    MOVP2R(X0, &b->ticStart);
-    BLR(X1);
-  }
+  // Fetch the current counter register
+  CNTVCT(X2);
+
+  // stores runCount and ticStart
+  STP(INDEX_SIGNED, X1, X2, X0, offsetof(JitBlock::ProfileData, runCount));
 }
 
 void JitArm64::EndTimeProfile(JitBlock* b)
 {
-  if (m_supports_cycle_counter)
-  {
-    EmitGetCycles(X2);
-    MOVP2R(X0, &b->ticStart);
-  }
-  else
-  {
-    MOVP2R(X1, &QueryPerformanceCounter);
-    MOVP2R(X0, &b->ticStop);
-    BLR(X1);
+  if (!Profiler::g_ProfileBlocks)
+    return;
 
-    MOVP2R(X0, &b->ticStart);
-    LDR(INDEX_UNSIGNED, X2, X0, 8);  // Stop
-  }
+  // Fetch the current counter register
+  CNTVCT(X1);
 
-  LDR(INDEX_UNSIGNED, X1, X0, 0);   // Start
-  LDR(INDEX_UNSIGNED, X3, X0, 16);  // Counter
-  SUB(X2, X2, X1);
-  ADD(X3, X3, X2);
-  STR(INDEX_UNSIGNED, X3, X0, 16);
+  MOVP2R(X0, &b->profile_data);
+
+  LDR(INDEX_UNSIGNED, X2, X0, offsetof(JitBlock::ProfileData, ticStart));
+  SUB(X1, X1, X2);
+
+  // loads ticCounter and downcountCounter
+  LDP(INDEX_SIGNED, X2, X3, X0, offsetof(JitBlock::ProfileData, ticCounter));
+  ADD(X2, X2, X1);
+  ADDI2R(X3, X3, js.downcountAmount, X1);
+
+  // stores ticCounter and downcountCounter
+  STP(INDEX_SIGNED, X2, X3, X0, offsetof(JitBlock::ProfileData, ticCounter));
 }
 
 void JitArm64::Run()
@@ -657,7 +600,6 @@ void JitArm64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer* code_buf, JitBlock*
 
   const u8* start = GetCodePtr();
   b->checkedEntry = start;
-  b->runCount = 0;
 
   // Downcount flag check, Only valid for linked blocks
   {
@@ -673,15 +615,6 @@ void JitArm64::DoJit(u32 em_address, PPCAnalyst::CodeBuffer* code_buf, JitBlock*
   // Conditionally add profiling code.
   if (Profiler::g_ProfileBlocks)
   {
-    ARM64Reg WA = gpr.GetReg();
-    ARM64Reg WB = gpr.GetReg();
-    ARM64Reg XA = EncodeRegTo64(WA);
-    ARM64Reg XB = EncodeRegTo64(WB);
-    MOVP2R(XA, &b->runCount);
-    LDR(INDEX_UNSIGNED, XB, XA, 0);
-    ADD(XB, XB, 1);
-    STR(INDEX_UNSIGNED, XB, XA, 0);
-    gpr.Unlock(WA, WB);
     // get start tic
     BeginTimeProfile(b);
   }
diff --git a/Source/Core/Core/PowerPC/JitArm64/Jit.h b/Source/Core/Core/PowerPC/JitArm64/Jit.h
index 5685440ca1..918ceb678c 100644
--- a/Source/Core/Core/PowerPC/JitArm64/Jit.h
+++ b/Source/Core/Core/PowerPC/JitArm64/Jit.h
@@ -174,9 +174,6 @@ private:
   static void InitializeInstructionTables();
   void CompileInstruction(PPCAnalyst::CodeOp& op);
 
-  void EmitResetCycleCounters();
-  void EmitGetCycles(Arm64Gen::ARM64Reg reg);
-
   // Simple functions to switch between near and far code emitting
   void SwitchToFarCode()
   {
@@ -253,9 +250,6 @@ private:
   Arm64Gen::ARM64CodeBlock farcode;
   u8* nearcode;  // Backed up when we switch to far code.
 
-  // Do we support cycle counter profiling?
-  bool m_supports_cycle_counter;
-
   bool m_enable_blr_optimization;
   bool m_cleanup_after_stackfault = false;
   u8* m_stack_base = nullptr;
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_Branch.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_Branch.cpp
index 82325b0f49..da49e0a3bf 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_Branch.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_Branch.cpp
@@ -67,8 +67,8 @@ void JitArm64::rfi(UGeckoInstruction inst)
   LDR(INDEX_UNSIGNED, WA, PPC_REG, PPCSTATE_OFF(spr[SPR_SRR0]));
   gpr.Unlock(WB, WC);
 
-  // WA is unlocked in this function
   WriteExceptionExit(WA);
+  gpr.Unlock(WA);
 }
 
 void JitArm64::bx(UGeckoInstruction inst)
@@ -220,6 +220,8 @@ void JitArm64::bcctrx(UGeckoInstruction inst)
   AND(WA, WA, 30, 29);  // Wipe the bottom 2 bits.
 
   WriteExit(WA, inst.LK_3, js.compilerPC + 4);
+
+  gpr.Unlock(WA);
 }
 
 void JitArm64::bclrx(UGeckoInstruction inst)
@@ -275,6 +277,8 @@ void JitArm64::bclrx(UGeckoInstruction inst)
 
   WriteBLRExit(WA);
 
+  gpr.Unlock(WA);
+
   if (conditional)
     SwitchToNearCode();
 
diff --git a/Source/Core/Core/PowerPC/JitCommon/JitCache.h b/Source/Core/Core/PowerPC/JitCommon/JitCache.h
index 76e2d4dbf9..7b8e44b1b2 100644
--- a/Source/Core/Core/PowerPC/JitCommon/JitCache.h
+++ b/Source/Core/Core/PowerPC/JitCommon/JitCache.h
@@ -49,7 +49,6 @@ struct JitBlock
   // The number of PPC instructions represented by this block. Mostly
   // useful for logging.
   u32 originalSize;
-  int runCount;  // for profiling.
 
   // Information about exits to a known address from this block.
   // This is used to implement block linking.
@@ -65,11 +64,15 @@ struct JitBlock
   // This set stores all physical addresses of all occupied instructions.
   std::set<u32> physical_addresses;
 
-  // we don't really need to save start and stop
-  // TODO (mb2): ticStart and ticStop -> "local var" mean "in block" ... low priority ;)
-  u64 ticStart;    // for profiling - time.
-  u64 ticStop;     // for profiling - time.
-  u64 ticCounter;  // for profiling - time.
+  // Block profiling data, structure is inlined in Jit.cpp
+  struct ProfileData
+  {
+    u64 ticCounter;
+    u64 downcountCounter;
+    u64 runCount;
+    u64 ticStart;
+    u64 ticStop;
+  } profile_data = {};
 
   // This tracks the position if this block within the fast block cache.
   // We allow each block to have only one map entry.
diff --git a/Source/Core/Core/PowerPC/JitInterface.cpp b/Source/Core/Core/PowerPC/JitInterface.cpp
index 2369eb75d6..284b7ea366 100644
--- a/Source/Core/Core/PowerPC/JitInterface.cpp
+++ b/Source/Core/Core/PowerPC/JitInterface.cpp
@@ -119,12 +119,12 @@ void GetProfileResults(ProfileStats* prof_stats)
 
   QueryPerformanceFrequency((LARGE_INTEGER*)&prof_stats->countsPerSec);
   g_jit->GetBlockCache()->RunOnBlocks([&prof_stats](const JitBlock& block) {
-    // Rough heuristic.  Mem instructions should cost more.
-    u64 cost = block.originalSize * (block.runCount / 4);
-    u64 timecost = block.ticCounter;
+    const auto& data = block.profile_data;
+    u64 cost = data.downcountCounter;
+    u64 timecost = data.ticCounter;
     // Todo: tweak.
-    if (block.runCount >= 1)
-      prof_stats->block_stats.emplace_back(block.effectiveAddress, cost, timecost, block.runCount,
+    if (data.runCount >= 1)
+      prof_stats->block_stats.emplace_back(block.effectiveAddress, cost, timecost, data.runCount,
                                            block.codeSize);
     prof_stats->cost_sum += cost;
     prof_stats->timecost_sum += timecost;
diff --git a/Source/Core/Core/PowerPC/Profiler.cpp b/Source/Core/Core/PowerPC/Profiler.cpp
index 25a31cf740..0378c4a5b5 100644
--- a/Source/Core/Core/PowerPC/Profiler.cpp
+++ b/Source/Core/Core/PowerPC/Profiler.cpp
@@ -5,11 +5,12 @@
 #include "Core/PowerPC/Profiler.h"
 
 #include <string>
+#include "Common/PerformanceCounter.h"
 #include "Core/PowerPC/JitInterface.h"
 
 namespace Profiler
 {
-bool g_ProfileBlocks;
+bool g_ProfileBlocks = false;
 
 void WriteProfileResults(const std::string& filename)
 {
diff --git a/Source/Core/Core/PowerPC/Profiler.h b/Source/Core/Core/PowerPC/Profiler.h
index 6146bee4d1..33a8b225bb 100644
--- a/Source/Core/Core/PowerPC/Profiler.h
+++ b/Source/Core/Core/PowerPC/Profiler.h
@@ -10,37 +10,6 @@
 
 #include "Common/CommonTypes.h"
 
-#include "Common/PerformanceCounter.h"
-
-#if defined(_M_X86_64)
-
-#define PROFILER_QUERY_PERFORMANCE_COUNTER(pt)                                                     \
-  MOV(64, R(ABI_PARAM1), Imm64(reinterpret_cast<u64>(pt)));                                        \
-  ABI_CallFunction(QueryPerformanceCounter)
-
-// block->ticCounter += block->ticStop - block->ticStart
-#define PROFILER_UPDATE_TIME(block)                                                                \
-  MOV(64, R(RSCRATCH2), Imm64((u64)block));                                                        \
-  MOV(64, R(RSCRATCH), MDisp(RSCRATCH2, offsetof(struct JitBlock, ticStop)));                      \
-  SUB(64, R(RSCRATCH), MDisp(RSCRATCH2, offsetof(struct JitBlock, ticStart)));                     \
-  ADD(64, R(RSCRATCH), MDisp(RSCRATCH2, offsetof(struct JitBlock, ticCounter)));                   \
-  MOV(64, MDisp(RSCRATCH2, offsetof(struct JitBlock, ticCounter)), R(RSCRATCH));
-
-#define PROFILER_VPUSH                                                                             \
-  BitSet32 registersInUse = CallerSavedRegistersInUse();                                           \
-  ABI_PushRegistersAndAdjustStack(registersInUse, 0);
-
-#define PROFILER_VPOP ABI_PopRegistersAndAdjustStack(registersInUse, 0);
-
-#else
-
-#define PROFILER_QUERY_PERFORMANCE_COUNTER(pt)
-#define PROFILER_UPDATE_TIME(b)
-#define PROFILER_VPUSH
-#define PROFILER_VPOP
-
-#endif
-
 struct BlockStat
 {
   BlockStat(u32 _addr, u64 c, u64 ticks, u64 run, u32 size)
diff --git a/Source/Core/DolphinWX/Debugger/JitWindow.cpp b/Source/Core/DolphinWX/Debugger/JitWindow.cpp
index a62d1ba2f2..bf2eab365c 100644
--- a/Source/Core/DolphinWX/Debugger/JitWindow.cpp
+++ b/Source/Core/DolphinWX/Debugger/JitWindow.cpp
@@ -87,6 +87,7 @@ void CJitWindow::Compare(u32 em_address)
   PPCAnalyst::CodeBlock code_block;
   PPCAnalyst::PPCAnalyzer analyzer;
   analyzer.SetOption(PPCAnalyst::PPCAnalyzer::OPTION_CONDITIONAL_CONTINUE);
+  analyzer.SetOption(PPCAnalyst::PPCAnalyzer::OPTION_BRANCH_FOLLOW);
 
   code_block.m_stats = &st;
   code_block.m_gpa = &gpa;