diff --git a/rpcs3/Emu/CPU/CPUThread.cpp b/rpcs3/Emu/CPU/CPUThread.cpp
index f055e72cf8..bb952393aa 100644
--- a/rpcs3/Emu/CPU/CPUThread.cpp
+++ b/rpcs3/Emu/CPU/CPUThread.cpp
@@ -145,7 +145,7 @@ void CPUThread::NextPc(u8 instr_size)
 	}
 }
 
-void CPUThread::SetBranch(const u64 pc)
+void CPUThread::SetBranch(const u64 pc, bool record_branch)
 {
 	if(!Memory.IsGoodAddr(m_offset + pc))
 	{
@@ -155,6 +155,9 @@ void CPUThread::SetBranch(const u64 pc)
 
 	m_is_branch = true;
 	nPC = pc;
+
+	if(record_branch)
+		CallStackBranch(pc);
 }
 
 void CPUThread::SetPc(const u64 pc)
diff --git a/rpcs3/Emu/CPU/CPUThread.h b/rpcs3/Emu/CPU/CPUThread.h
index bc8d4a0cad..bde7140c8b 100644
--- a/rpcs3/Emu/CPU/CPUThread.h
+++ b/rpcs3/Emu/CPU/CPUThread.h
@@ -132,7 +132,7 @@ public:
 	int ThreadStatus();
 
 	void NextPc(u8 instr_size);
-	void SetBranch(const u64 pc);
+	void SetBranch(const u64 pc, bool record_branch = false);
 	void SetPc(const u64 pc);
 	void SetEntry(const u64 entry);
 
@@ -171,15 +171,50 @@ public:
 	virtual void Exec();
 	void ExecOnce();
 
-	Stack<u64> m_call_stack;
+	struct CallStackItem
+	{
+		u64 pc;
+		u64 branch_pc;
+	};
+
+	Stack<CallStackItem> m_call_stack;
+
 	wxString CallStackToString()
 	{
 		wxString ret = "Call Stack:\n==========\n";
+
 		for(uint i=0; i<m_call_stack.GetCount(); ++i)
-			ret += wxString::Format("0x%llx\n", m_call_stack.Get(i));
+		{
+			ret += wxString::Format("0x%llx -> 0x%llx\n", m_call_stack[i].pc, m_call_stack[i].branch_pc);
+		}
+
 		return ret;
 	}
 
+	void CallStackBranch(u64 pc)
+	{
+		for(int i=m_call_stack.GetCount() - 1; i >= 0; --i)
+		{
+			if(CallStackGetNextPC(m_call_stack[i].pc) == pc)
+			{
+				m_call_stack.RemoveAt(i, m_call_stack.GetCount() - i);
+				return;
+			}
+		}
+
+		CallStackItem new_item;
+
+		new_item.branch_pc = pc;
+		new_item.pc = PC;
+
+		m_call_stack.AddCpy(new_item);
+	}
+
+	virtual u64 CallStackGetNextPC(u64 pc)
+	{
+		return pc + 4;
+	}
+
 protected:
 	virtual void DoReset()=0;
 	virtual void DoRun()=0;
diff --git a/rpcs3/Emu/Cell/PPUInterpreter.h b/rpcs3/Emu/Cell/PPUInterpreter.h
index 96c0856b61..44c5eb67fd 100644
--- a/rpcs3/Emu/Cell/PPUInterpreter.h
+++ b/rpcs3/Emu/Cell/PPUInterpreter.h
@@ -2059,7 +2059,7 @@ private:
 	void BC(u32 bo, u32 bi, s32 bd, u32 aa, u32 lk)
 	{
 		if(!CheckCondition(bo, bi)) return;
-		CPU.SetBranch(branchTarget((aa ? 0 : CPU.PC), bd));
+		CPU.SetBranch(branchTarget((aa ? 0 : CPU.PC), bd), lk);
 		if(lk) CPU.LR = CPU.PC + 4;
 	}
 	void SC(s32 sc_code)
@@ -2074,9 +2074,8 @@ private:
 	}
 	void B(s32 ll, u32 aa, u32 lk)
 	{
-		CPU.SetBranch(branchTarget(aa ? 0 : CPU.PC, ll));
+		CPU.SetBranch(branchTarget(aa ? 0 : CPU.PC, ll), lk);
 		if(lk) CPU.LR = CPU.PC + 4;
-		CPU.m_call_stack.Push(branchTarget(aa ? 0 : CPU.PC, ll)); //This does not affect emulation
 	}
 	void MCRF(u32 crfd, u32 crfs)
 	{
@@ -2085,9 +2084,8 @@ private:
 	void BCLR(u32 bo, u32 bi, u32 bh, u32 lk)
 	{
 		if(!CheckCondition(bo, bi)) return;
-		CPU.SetBranch(branchTarget(0, CPU.LR));
+		CPU.SetBranch(branchTarget(0, CPU.LR), true);
 		if(lk) CPU.LR = CPU.PC + 4;
-		CPU.m_call_stack.Pop(); //This does not affect emulation
 	}
 	void CRNOR(u32 crbd, u32 crba, u32 crbb)
 	{
@@ -2136,7 +2134,7 @@ private:
 	{
 		if(bo & 0x10 || CPU.IsCR(bi) == (bo & 0x8))
 		{
-			CPU.SetBranch(branchTarget(0, CPU.CTR));
+			CPU.SetBranch(branchTarget(0, CPU.CTR), true);
 			if(lk) CPU.LR = CPU.PC + 4;
 		}
 	}