diff --git a/CMakeLists.txt b/CMakeLists.txt
index 8449c41..3977695 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -34,7 +34,7 @@ set(LIB_SOURCES
     src/plugin/instruction_count.cpp
     src/plugin/cycle_estimate.cpp
 )
-if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/src/iss/tgf_b.cpp)
+if(EXISTS src/iss/tgf_b.cpp)
     set(LIB_SOURCES ${LIB_SOURCES}  src/iss/tgf_b.cpp src/vm/interp/vm_tgf_b.cpp)
 endif()
 if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/src/iss/tgf_c.cpp)
@@ -43,8 +43,8 @@ endif()
 if(WITH_LLVM)
 set(LIB_SOURCES ${LIB_SOURCES}
 	src/vm/llvm/fp_impl.cpp
-	src/vm/llvm/vm_tgf_b.cpp
-	src/vm/llvm/vm_tgf_c.cpp
+#	src/vm/llvm/vm_tgf_b.cpp
+#	src/vm/llvm/vm_tgf_c.cpp
 )
 endif()
 
@@ -52,10 +52,10 @@ endif()
 add_library(${PROJECT_NAME} SHARED ${LIB_SOURCES})
 # list code gen dependencies
 if(TARGET TGF_B_src)
-    add_dependencies(tgfs TGF_B_src)
+    add_dependencies(${PROJECT_NAME} TGF_B_src)
 endif()
 if(TARGET TGF_C_src)
-    add_dependencies(tgfs TGF_C_src)
+    add_dependencies(${PROJECT_NAME} TGF_C_src)
 endif()
 
 target_compile_options(${PROJECT_NAME} PRIVATE -Wno-shift-count-overflow)
diff --git a/gen_input/templates/interp/CORENAME.cpp.gtl b/gen_input/templates/interp/CORENAME.cpp.gtl
index ad43e15..4d1e12c 100644
--- a/gen_input/templates/interp/CORENAME.cpp.gtl
+++ b/gen_input/templates/interp/CORENAME.cpp.gtl
@@ -203,9 +203,11 @@ private:
         // prepare execution
         uint${addrDataWidth}_t* PC = reinterpret_cast<uint${addrDataWidth}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint${addrDataWidth}_t* NEXT_PC = reinterpret_cast<uint${addrDataWidth}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
-        // used registers<%instr.usedVariables.each{%>
-        uint${it.value.size}_t* ${it.key} = reinterpret_cast<uint${it.value.size}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::${it.key}]);
-        <%}%>// calculate next pc value
+        // used registers<%instr.usedVariables.each{ k,v->
+            if(v.isArray) {%>
+        uint${v.type.size}_t* ${k} = reinterpret_cast<uint${v.type.size}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::${k}0]);<% }else{ %> 
+        uint${v.type.size}_t* ${k} = reinterpret_cast<uint${v.type.size}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::${k}]);
+        <%}}%>// calculate next pc value
         *NEXT_PC = *PC + ${instr.length/8};
         // execute instruction
         <%instr.behavior.eachLine{%>${it}
@@ -292,7 +294,8 @@ typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e co
             res = fetch_ins(virt_addr_t{access_type::FETCH, new_pc}, data);
             if(res!=iss::Ok) throw simulation_stopped(0);
         }
-        if ((cond & finish_cond_e::JUMP_TO_SELF) && (insn == 0x0000006f || (insn&0xffff)==0xa001)) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
+        if ((cond & finish_cond_e::JUMP_TO_SELF) == finish_cond_e::JUMP_TO_SELF &&
+                (insn == 0x0000006f || (insn&0xffff)==0xa001)) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
         auto lut_val = extract_fields(insn);
         auto f = qlut[insn & 0x3][lut_val];
         if (!f)
diff --git a/src/vm/interp/vm_tgf_c.cpp b/src/vm/interp/vm_tgf_c.cpp
index ef6ce60..8123d5e 100644
--- a/src/vm/interp/vm_tgf_c.cpp
+++ b/src/vm/interp/vm_tgf_c.cpp
@@ -382,8 +382,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd != 0) *(X+rd) = (int32_t)imm; 
@@ -416,8 +415,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd != 0) *(X+rd) = *PC + (int32_t)imm; 
@@ -450,8 +448,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -488,8 +485,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -527,8 +523,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(*(X+rs1) == *(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -562,8 +557,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(*(X+rs1) != *(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -597,8 +591,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if((int32_t)*(X+rs1) < (int32_t)*(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -632,8 +625,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if((int32_t)*(X+rs1) >= (int32_t)*(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -667,8 +659,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(*(X+rs1) < *(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -702,8 +693,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(*(X+rs1) >= *(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -737,8 +727,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -775,8 +764,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -817,8 +805,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -859,8 +846,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -897,8 +883,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -939,8 +924,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         writeSpace1(traits::MEM, *(X+rs1) + (int16_t)sext<12>(imm), (int8_t)*(X+rs2));
@@ -974,8 +958,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -1013,8 +996,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -1052,8 +1034,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) + (int16_t)sext<12>(imm); 
@@ -1087,8 +1068,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = (int32_t)*(X+rs1) < (int16_t)sext<12>(imm)?  1 :  0; 
@@ -1122,8 +1102,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) < (int16_t)sext<12>(imm)?  1 :  0; 
@@ -1157,8 +1136,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) ^ (int16_t)sext<12>(imm); 
@@ -1192,8 +1170,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) | (int16_t)sext<12>(imm); 
@@ -1227,8 +1204,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) & (int16_t)sext<12>(imm); 
@@ -1262,8 +1238,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(shamt >  31) {
@@ -1302,8 +1277,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(shamt >  31) {
@@ -1342,8 +1316,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(shamt >  31) {
@@ -1382,8 +1355,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) + *(X+rs2); 
@@ -1417,8 +1389,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) - *(X+rs2); 
@@ -1452,8 +1423,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) << (*(X+rs2) & (traits::XLEN - 1)); 
@@ -1487,8 +1457,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = (int32_t)*(X+rs1) < (int32_t)*(X+rs2)?  1 :  0; 
@@ -1522,8 +1491,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = (uint32_t)*(X+rs1) < (uint32_t)*(X+rs2)?  1 :  0; 
@@ -1557,8 +1525,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) ^ *(X+rs2); 
@@ -1592,8 +1559,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) >> (*(X+rs2) & (traits::XLEN - 1)); 
@@ -1627,8 +1593,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = (int32_t)*(X+rs1) >> (*(X+rs2) & (traits::XLEN - 1)); 
@@ -1662,8 +1627,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) | *(X+rs2); 
@@ -1697,8 +1661,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) & *(X+rs2); 
@@ -1996,8 +1959,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2041,8 +2003,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2081,8 +2042,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2121,8 +2081,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2159,8 +2118,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2198,8 +2156,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2237,8 +2194,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2277,8 +2233,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2317,8 +2272,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2357,8 +2311,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2397,8 +2350,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2441,8 +2393,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2481,8 +2432,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2525,8 +2475,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2564,8 +2513,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2602,8 +2550,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2641,8 +2588,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2679,8 +2625,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         *(X+rs1) = *(X+rs1) + (int8_t)sext<6>(imm);
@@ -2741,8 +2686,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2778,8 +2722,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2814,8 +2757,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2851,8 +2793,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(nzimm) *(X+2) = *(X+2) + (int16_t)sext<10>(nzimm); 
@@ -2886,8 +2827,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2923,8 +2863,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(shamt) {
@@ -2964,8 +2903,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3001,8 +2939,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3038,8 +2975,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3075,8 +3011,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3112,8 +3047,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3180,8 +3114,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(*(X+(rs1 + 8)) == 0) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<9>(imm); 
@@ -3214,8 +3147,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(*(X+(rs1 + 8)) != 0) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<9>(imm); 
@@ -3248,8 +3180,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(nzuimm) *(X+rs1) = *(X+rs1) << nzuimm; 
@@ -3282,8 +3213,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(rd) {
@@ -3320,8 +3250,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         *(X+rd) = *(X+rs2);
@@ -3353,8 +3282,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(rs1) pc_assign(*NEXT_PC) = *(X+rs1); 
@@ -3388,8 +3316,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         *(X+rd) = *(X+rd) + *(X+rs2);
@@ -3421,8 +3348,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3485,8 +3411,7 @@ private:
         uint32_t* PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::PC]);
         uint32_t* NEXT_PC = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::NEXT_PC]);
         // used registers
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X]);
-        // calculate next pc value
+        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3603,7 +3528,8 @@ typename vm_base<ARCH>::virt_addr_t vm_impl<ARCH>::execute_inst(finish_cond_e co
             res = fetch_ins(virt_addr_t{access_type::FETCH, new_pc}, data);
             if(res!=iss::Ok) throw simulation_stopped(0);
         }
-        if ((cond & finish_cond_e::JUMP_TO_SELF) && (insn == 0x0000006f || (insn&0xffff)==0xa001)) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
+        if ((cond & finish_cond_e::JUMP_TO_SELF) == finish_cond_e::JUMP_TO_SELF &&
+                (insn == 0x0000006f || (insn&0xffff)==0xa001)) throw simulation_stopped(0); // 'J 0' or 'C.J 0'
         auto lut_val = extract_fields(insn);
         auto f = qlut[insn & 0x3][lut_val];
         if (!f)