From 78c70642952f49a4362de0abae35a4e539b8d130 Mon Sep 17 00:00:00 2001
From: Eyck Jentzsch <eyck@minres.com>
Date: Fri, 26 Mar 2021 08:23:47 +0000
Subject: [PATCH] update groovy template to extract used registers

---
 gen_input/templates/interp/CORENAME.cpp.gtl |   4 +-
 src/vm/interp/vm_tgf_c.cpp                  | 329 ++++++++++++++------
 2 files changed, 242 insertions(+), 91 deletions(-)

diff --git a/gen_input/templates/interp/CORENAME.cpp.gtl b/gen_input/templates/interp/CORENAME.cpp.gtl
index f305cbb..ad43e15 100644
--- a/gen_input/templates/interp/CORENAME.cpp.gtl
+++ b/gen_input/templates/interp/CORENAME.cpp.gtl
@@ -201,9 +201,11 @@ private:
             <%}%>
         }
         // prepare execution
-        uint${addrDataWidth}_t* X = reinterpret_cast<uint${addrDataWidth}_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + ${instr.length/8};
         // execute instruction
         <%instr.behavior.eachLine{%>${it}
diff --git a/src/vm/interp/vm_tgf_c.cpp b/src/vm/interp/vm_tgf_c.cpp
index 11e811a..ef6ce60 100644
--- a/src/vm/interp/vm_tgf_c.cpp
+++ b/src/vm/interp/vm_tgf_c.cpp
@@ -379,9 +379,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd != 0) *(X+rd) = (int32_t)imm; 
@@ -411,9 +413,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd != 0) *(X+rd) = *PC + (int32_t)imm; 
@@ -443,9 +447,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -479,9 +485,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -516,9 +524,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(*(X+rs1) == *(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -549,9 +559,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(*(X+rs1) != *(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -582,9 +594,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *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); 
@@ -615,9 +629,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *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); 
@@ -648,9 +664,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(*(X+rs1) < *(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -681,9 +699,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(*(X+rs1) >= *(X+rs2)) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<13>(imm); 
@@ -714,9 +734,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -750,9 +772,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -790,9 +814,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -830,9 +856,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -866,9 +894,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -906,9 +936,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         writeSpace1(traits::MEM, *(X+rs1) + (int16_t)sext<12>(imm), (int8_t)*(X+rs2));
@@ -939,9 +971,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -976,9 +1010,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -1013,9 +1049,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) + (int16_t)sext<12>(imm); 
@@ -1046,9 +1084,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = (int32_t)*(X+rs1) < (int16_t)sext<12>(imm)?  1 :  0; 
@@ -1079,9 +1119,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) < (int16_t)sext<12>(imm)?  1 :  0; 
@@ -1112,9 +1154,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) ^ (int16_t)sext<12>(imm); 
@@ -1145,9 +1189,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) | (int16_t)sext<12>(imm); 
@@ -1178,9 +1224,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) & (int16_t)sext<12>(imm); 
@@ -1211,9 +1259,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(shamt >  31) {
@@ -1249,9 +1299,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(shamt >  31) {
@@ -1287,9 +1339,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(shamt >  31) {
@@ -1325,9 +1379,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) + *(X+rs2); 
@@ -1358,9 +1414,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) - *(X+rs2); 
@@ -1391,9 +1449,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) << (*(X+rs2) & (traits::XLEN - 1)); 
@@ -1424,9 +1484,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = (int32_t)*(X+rs1) < (int32_t)*(X+rs2)?  1 :  0; 
@@ -1457,9 +1519,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = (uint32_t)*(X+rs1) < (uint32_t)*(X+rs2)?  1 :  0; 
@@ -1490,9 +1554,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) ^ *(X+rs2); 
@@ -1523,9 +1589,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) >> (*(X+rs2) & (traits::XLEN - 1)); 
@@ -1556,9 +1624,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = (int32_t)*(X+rs1) >> (*(X+rs2) & (traits::XLEN - 1)); 
@@ -1589,9 +1659,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) | *(X+rs2); 
@@ -1622,9 +1694,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         if(rd !=  0) *(X+rd) = *(X+rs1) & *(X+rs2); 
@@ -1656,9 +1730,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         writeSpace1(traits::FENCE, traits::fence, pred << 4 | succ);
@@ -1689,9 +1763,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         writeSpace2(traits::FENCE, traits::fencei, imm);
@@ -1716,9 +1790,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         raise(0,  11);
@@ -1743,9 +1817,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         raise(0,  3);
@@ -1770,9 +1844,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         leave(0);
@@ -1797,9 +1871,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         leave(1);
@@ -1824,9 +1898,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         leave(3);
@@ -1851,9 +1925,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         wait(1);
@@ -1883,9 +1957,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -1919,9 +1993,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -1962,9 +2038,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2000,9 +2078,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2038,9 +2118,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2074,9 +2156,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2111,9 +2195,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2148,9 +2234,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2186,9 +2274,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2224,9 +2314,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2262,9 +2354,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2300,9 +2394,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2342,9 +2438,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2380,9 +2478,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2422,9 +2522,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 4;
         // execute instruction
         {
@@ -2459,9 +2561,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2495,9 +2599,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2532,9 +2638,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2568,9 +2676,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         *(X+rs1) = *(X+rs1) + (int8_t)sext<6>(imm);
@@ -2596,9 +2706,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2628,9 +2738,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2663,9 +2775,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2697,9 +2811,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2732,9 +2848,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(nzimm) *(X+2) = *(X+2) + (int16_t)sext<10>(nzimm); 
@@ -2765,9 +2883,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2800,9 +2920,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(shamt) {
@@ -2839,9 +2961,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2874,9 +2998,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2909,9 +3035,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2944,9 +3072,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -2979,9 +3109,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3013,9 +3145,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         pc_assign(*NEXT_PC) = *PC + (int16_t)sext<12>(imm);
@@ -3045,9 +3177,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(*(X+(rs1 + 8)) == 0) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<9>(imm); 
@@ -3077,9 +3211,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(*(X+(rs1 + 8)) != 0) pc_assign(*NEXT_PC) = *PC + (int16_t)sext<9>(imm); 
@@ -3109,9 +3245,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(nzuimm) *(X+rs1) = *(X+rs1) << nzuimm; 
@@ -3141,9 +3279,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(rd) {
@@ -3177,9 +3317,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         *(X+rd) = *(X+rs2);
@@ -3208,9 +3350,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         if(rs1) pc_assign(*NEXT_PC) = *(X+rs1); 
@@ -3241,9 +3385,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         *(X+rd) = *(X+rd) + *(X+rs2);
@@ -3272,9 +3418,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3302,9 +3450,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         raise(0,  3);
@@ -3334,9 +3482,11 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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
         *NEXT_PC = *PC + 2;
         // execute instruction
         {
@@ -3364,9 +3514,9 @@ private:
             
         }
         // prepare execution
-        uint32_t* X = reinterpret_cast<uint32_t*>(this->regs_base_ptr+arch::traits<ARCH>::reg_byte_offsets[arch::traits<ARCH>::X0]);
         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// calculate next pc value
         *NEXT_PC = *PC + 2;
         // execute instruction
         raise(0,  2);
@@ -3453,8 +3603,7 @@ 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) == 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) && (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)