diff --git a/src/AssemblyGeneratorX86.cpp b/src/AssemblyGeneratorX86.cpp
index 7bbb658..1922b0e 100644
--- a/src/AssemblyGeneratorX86.cpp
+++ b/src/AssemblyGeneratorX86.cpp
@@ -499,9 +499,6 @@ namespace RandomX {
 		gena(instr);
 		asmCode << "\tcmp rsp, rbp" << std::endl;
 		asmCode << "\tje short not_taken_ret_" << i << std::endl;
-		asmCode << "\tcmp " << regR32[instr.regb % RegistersCount] << ", " << instr.imm32 << std::endl;
-		asmCode << "\t" << jumpCondition(instr, true);
-		asmCode << " short not_taken_ret_" << i << std::endl;
 		asmCode << "\txor rax, qword ptr [rsp + 8]" << std::endl;
 		gencr(instr);
 		asmCode << "\tret 8" << std::endl;
diff --git a/src/InterpretedVirtualMachine.cpp b/src/InterpretedVirtualMachine.cpp
index 84b4b11..2bb4d75 100644
--- a/src/InterpretedVirtualMachine.cpp
+++ b/src/InterpretedVirtualMachine.cpp
@@ -26,6 +26,9 @@ along with RandomX.  If not, see<http://www.gnu.org/licenses/>.
 #include <stdexcept>
 #include <sstream>
 #include <cmath>
+#ifdef STATS
+#include <algorithm>
+#endif
 
 #ifdef FPUCHECK
 constexpr bool fpuCheck = true;
@@ -54,6 +57,9 @@ namespace RandomX {
 
 	void InterpretedVirtualMachine::execute() {
 		while (ic > 0) {
+#ifdef STATS
+			count_instructions[pc]++;
+#endif
 			auto& inst = p(pc);
 			if(trace) std::cout << inst.getName() << " (" << std::dec << pc << ")" << std::endl;
 			pc = (pc + 1) % ProgramLength;
@@ -61,6 +67,9 @@ namespace RandomX {
 			(this->*handler)(inst);
 			ic--;
 		}
+#ifdef STATS
+		count_endstack += stack.size();
+#endif
 	}
 
 	convertible_t InterpretedVirtualMachine::loada(Instruction& inst) {
@@ -284,9 +293,13 @@ namespace RandomX {
 #ifdef STATS
 			count_CALL_taken++;
 			count_jump_taken[inst.locb & 7]++;
+			count_retdepth = std::max(0, count_retdepth - 1);
 #endif
 			stackPush(a);
 			stackPush(pc);
+#ifdef STATS
+			count_max_stack = std::max(count_max_stack, (int)stack.size());
+#endif
 			pc += (inst.imm8 & 127) + 1;
 			pc = pc % ProgramLength;
 			if (trace) std::cout << std::hex << a.u64 << std::endl;
@@ -306,10 +319,11 @@ namespace RandomX {
 		convertible_t a = loada(inst);
 		convertible_t b = loadbr1(inst);
 		convertible_t& c = getcr(inst);
-		if (stack.size() > 0 && JMP_COND(inst.locb, reg.r[inst.regb % RegistersCount], inst.imm32)) {
+		if (stack.size() > 0) {
 #ifdef STATS
 			count_RET_taken++;
-			count_jump_taken[inst.locb & 7]++;
+			count_retdepth++;
+			count_retdepth_max = std::max(count_retdepth_max, count_retdepth);
 #endif
 			auto raddr = stackPopAddress();
 			auto retval = stackPopValue();
diff --git a/src/InterpretedVirtualMachine.hpp b/src/InterpretedVirtualMachine.hpp
index 1427c1b..5a6f49a 100644
--- a/src/InterpretedVirtualMachine.hpp
+++ b/src/InterpretedVirtualMachine.hpp
@@ -18,7 +18,7 @@ along with RandomX.  If not, see<http://www.gnu.org/licenses/>.
 */
 
 #pragma once
-//#define STATS
+#define STATS
 #include "VirtualMachine.hpp"
 #include "Program.hpp"
 #include <vector>
@@ -43,41 +43,46 @@ namespace RandomX {
 		std::vector<convertible_t> stack;
 		uint64_t pc, ic;
 #ifdef STATS
-		int count_ADD_64;
-		int count_ADD_32;
-		int count_SUB_64;
-		int count_SUB_32;
-		int count_MUL_64;
-		int count_MULH_64;
-		int count_MUL_32;
-		int count_IMUL_32;
-		int count_IMULH_64;
-		int count_DIV_64;
-		int count_IDIV_64;
-		int count_AND_64;
-		int count_AND_32;
-		int count_OR_64;
-		int count_OR_32;
-		int count_XOR_64;
-		int count_XOR_32;
-		int count_SHL_64;
-		int count_SHR_64;
-		int count_SAR_64;
-		int count_ROL_64;
-		int count_ROR_64;
-		int count_FPADD;
-		int count_FPSUB;
-		int count_FPMUL;
-		int count_FPDIV;
-		int count_FPSQRT;
-		int count_FPROUND;
-		int count_CALL_taken;
-		int count_CALL_not_taken;
-		int count_RET_stack_empty;
-		int count_RET_taken;
-		int count_RET_not_taken;
+		int count_ADD_64 = 0;
+		int count_ADD_32 = 0;
+		int count_SUB_64 = 0;
+		int count_SUB_32 = 0;
+		int count_MUL_64 = 0;
+		int count_MULH_64 = 0;
+		int count_MUL_32 = 0;
+		int count_IMUL_32 = 0;
+		int count_IMULH_64 = 0;
+		int count_DIV_64 = 0;
+		int count_IDIV_64 = 0;
+		int count_AND_64 = 0;
+		int count_AND_32 = 0;
+		int count_OR_64 = 0;
+		int count_OR_32 = 0;
+		int count_XOR_64 = 0;
+		int count_XOR_32 = 0;
+		int count_SHL_64 = 0;
+		int count_SHR_64 = 0;
+		int count_SAR_64 = 0;
+		int count_ROL_64 = 0;
+		int count_ROR_64 = 0;
+		int count_FPADD = 0;
+		int count_FPSUB = 0;
+		int count_FPMUL = 0;
+		int count_FPDIV = 0;
+		int count_FPSQRT = 0;
+		int count_FPROUND = 0;
+		int count_CALL_taken = 0;
+		int count_CALL_not_taken = 0;
+		int count_RET_stack_empty = 0;
+		int count_RET_taken = 0;
+		int count_RET_not_taken = 0;
 		int count_jump_taken[8] = { 0 };
 		int count_jump_not_taken[8] = { 0 };
+		int count_max_stack = 0;
+		int count_retdepth = 0;
+		int count_retdepth_max = 0;
+		int count_endstack = 0;
+		int count_instructions[ProgramLength] = { 0 };
 #endif
 
 		convertible_t loada(Instruction&);
diff --git a/src/JitCompilerX86.cpp b/src/JitCompilerX86.cpp
index 6f83e40..644fd9b 100644
--- a/src/JitCompilerX86.cpp
+++ b/src/JitCompilerX86.cpp
@@ -712,11 +712,6 @@ namespace RandomX {
 			crlen = 17;
 		}
 		emit(0x74e53b48); //cmp rsp, rbp; je
-		emitByte(20 + crlen);
-		emit(uint16_t(0x8141)); //cmp regb, imm32
-		emitByte(0xf8 + (instr.regb % RegistersCount));
-		emit(instr.imm32);
-		emitByte(jumpCondition(instr, true));
 		emitByte(11 + crlen);
 		emitByte(0x48);
 		emit(0x08244433); //xor rax,QWORD PTR [rsp+0x8]