diff --git a/include/common.h b/include/common.h
index 52b03ef20115e9bf4956cb81514b182c97d6eb74..a5d8474c3b6a43b8a085c465955e73a0bbdc306d 100644
--- a/include/common.h
+++ b/include/common.h
@@ -1,7 +1,7 @@
 #ifndef __COMMON_H__
 #define __COMMON_H__
 
-#define DEBUG
+//#define DEBUG
 //#define DIFF_TEST
 
 #define ISA64
diff --git a/include/monitor/diff-test.h b/include/monitor/diff-test.h
index 8c9c43d10966b5285c636a307fff6aedfb889d72..8aef569037382adbc7ee241267be6a2388178605 100644
--- a/include/monitor/diff-test.h
+++ b/include/monitor/diff-test.h
@@ -22,7 +22,7 @@ extern void (*ref_difftest_exec)(uint64_t n);
 static inline bool difftest_check_reg(const char *name, vaddr_t pc, rtlreg_t ref, rtlreg_t dut) {
   if (ref != dut) {
 #ifdef ISA64
-    Log("%s is different after executing instruction at pc = 0x%16lx, right = 0x%16lx, wrong = 0x%16lx",
+    Log("%s is different after executing instruction at pc = 0x%016lx, right = 0x%016lx, wrong = 0x%016lx",
         name, pc, ref, dut);
 #else
     Log("%s is different after executing instruction at pc = 0x%08x, right = 0x%08x, wrong = 0x%08x",
diff --git a/include/rtl/rtl.h b/include/rtl/rtl.h
index 77477d03eb44738bf29feb278a80e330ad1cf8eb..3d9ecd9610eaa6e2769645da27f2752559f098dc 100644
--- a/include/rtl/rtl.h
+++ b/include/rtl/rtl.h
@@ -13,7 +13,7 @@ bool interpret_relop(uint32_t relop, const rtlreg_t src1, const rtlreg_t src2);
 
 /* RTL basic instructions */
 
-static inline void interpret_rtl_li(rtlreg_t* dest, uint64_t imm) {
+static inline void interpret_rtl_li(rtlreg_t* dest, rtlreg_t imm) {
   *dest = imm;
 }
 
@@ -158,17 +158,24 @@ static inline void rtl_not(rtlreg_t *dest, const rtlreg_t* src1) {
 static inline void rtl_sext(rtlreg_t* dest, const rtlreg_t* src1, int width) {
   // dest <- signext(src1[(width * 8 - 1) .. 0])
 //  TODO();
+
+#ifdef ISA64
   if (width == 8) {
     rtl_mv(dest, src1);
   } else {
     assert(width == 1 || width == 2 || width == 4);
     rtl_shli(dest, src1, (8 - width) * 8);
-#ifdef ISA64
     rtl_sar64i(dest, dest, (8 - width) * 8);
+  }
 #else
-    rtl_sari(dest, dest, (8 - width) * 8);
-#endif
+  if (width == 4) {
+    rtl_mv(dest, src1);
+  } else {
+    assert(width == 1 || width == 2);
+    rtl_shli(dest, src1, (4 - width) * 8);
+    rtl_sari(dest, dest, (4 - width) * 8);
   }
+#endif
 }
 
 static inline void rtl_setrelopi(uint32_t relop, rtlreg_t *dest,
diff --git a/runall.sh b/runall.sh
index d0dad91a6e80e1ce7f52e2ed55d64172f2ffdce6..c2b8e8b52d3c59082cf43bb34b34b662d7b42908 100644
--- a/runall.sh
+++ b/runall.sh
@@ -30,7 +30,7 @@ for file in $files; do
 
   if (grep 'nemu: .*HIT GOOD TRAP' $logfile > /dev/null) then
     echo -e "\033[1;32mPASS!\033[0m"
-    #rm $logfile
+    rm $logfile
   else
     echo -e "\033[1;31mFAIL!\033[0m see $logfile for more information"
     if (test -e $ori_log) then
diff --git a/src/cpu/inv.c b/src/cpu/inv.c
index c255f51f538726a0a7b5392c2c0e6b8ddbdff8ee..4afab7e907729ad957b3abafdfc601892f3e2c77 100644
--- a/src/cpu/inv.c
+++ b/src/cpu/inv.c
@@ -5,9 +5,9 @@ void display_inv_msg(vaddr_t pc) {
 
 #ifdef ISA64
   printf("There are two cases which will trigger this unexpected exception:\n"
-      "1. The instruction at PC = 0x%16lx is not implemented.\n"
+      "1. The instruction at PC = 0x%016lx is not implemented.\n"
       "2. Something is implemented incorrectly.\n", pc);
-  printf("Find this PC(0x%16lx) in the disassembling result to distinguish which case it is.\n\n", pc);
+  printf("Find this PC(0x%016lx) in the disassembling result to distinguish which case it is.\n\n", pc);
 #else
   printf("There are two cases which will trigger this unexpected exception:\n"
       "1. The instruction at PC = 0x%08x is not implemented.\n"
diff --git a/src/device/io/map.c b/src/device/io/map.c
index 25668bcdb37f028e83db71acbbc03e3589ccf2d3..8f969fd78eb4544a2081bc71bced95d4b777032b 100644
--- a/src/device/io/map.c
+++ b/src/device/io/map.c
@@ -19,7 +19,7 @@ uint8_t* new_space(int size) {
 static inline void check_bound(IOMap *map, paddr_t addr) {
 #ifdef ISA64
   Assert(map != NULL && addr <= map->high && addr >= map->low,
-      "address (0x%08x) is out of bound {%s} [0x%08x, 0x%08x] at pc = 0x%16lx",
+      "address (0x%08x) is out of bound {%s} [0x%08x, 0x%08x] at pc = 0x%016lx",
       addr, (map ? map->name : "???"), (map ? map->low : 0), (map ? map->high : 0), cpu.pc);
 #else
   Assert(map != NULL && addr <= map->high && addr >= map->low,
diff --git a/src/isa/riscv32/decode.c b/src/isa/riscv32/decode.c
index 336ee24de6db84f6461db181256750d60509f284..2f5614ab667ecd99940d93cc9c3bf4ddd64c829e 100644
--- a/src/isa/riscv32/decode.c
+++ b/src/isa/riscv32/decode.c
@@ -1,7 +1,7 @@
 #include "cpu/decode.h"
 #include "rtl/rtl.h"
 
-// decode operand helper (i & r)
+// decode operand helper
 #define make_DopHelper(name) void concat(decode_op_, name) (Operand *op, uint32_t val, bool load_val)
 
 static inline make_DopHelper(i) {
@@ -16,13 +16,12 @@ static inline make_DopHelper(r) {
   op->type = OP_TYPE_REG;
   op->reg = val;
   if (load_val) {
-    rtl_lr(&op->val, op->reg, 4);   // here reg is index of reg, width is ignored
+    rtl_lr(&op->val, op->reg, 4);
   }
 
   print_Dop(op->str, OP_STR_SIZE, "%s", reg_name(op->reg, 4));
 }
 
-// decode helper (I, R, U, J, B, ld, st, csr)
 make_DHelper(I) {
   decode_op_r(id_src, decinfo.isa.instr.rs1, true);
   decode_op_i(id_src2, decinfo.isa.instr.simm11_0, true);
diff --git a/src/isa/riscv64/exec/special.c b/src/isa/riscv64/exec/special.c
index 7cfb8bb708069b3dfe2473cc2d8cb44c019e051a..f89cc6e25042063a7b04111cd7c8bca834984492 100644
--- a/src/isa/riscv64/exec/special.c
+++ b/src/isa/riscv64/exec/special.c
@@ -9,7 +9,7 @@ make_EHelper(inv) {
   instr[0] = instr_fetch(pc, 4);
   instr[1] = instr_fetch(pc, 4);
 
-  printf("invalid opcode(PC = 0x%16lx): %08x %08x ...\n\n",
+  printf("invalid opcode(PC = 0x%016lx): %08x %08x ...\n\n",
       cpu.pc, instr[0], instr[1]);
 
   display_inv_msg(cpu.pc);
diff --git a/src/isa/riscv64/reg.c b/src/isa/riscv64/reg.c
index 4023f4fc118e8376287bae3ec871f4d63f87fc65..b98abbe0a52e629fb7d29db097f248c1f6dd91da 100644
--- a/src/isa/riscv64/reg.c
+++ b/src/isa/riscv64/reg.c
@@ -10,12 +10,12 @@ const char *regsl[] = {
 void isa_reg_display() {
   int i;
   for (i = 0; i < 32; i ++) {
-    printf("%s: 0x%16lx ", regsl[i], cpu.gpr[i]._64);
+    printf("%s: 0x%016lx ", regsl[i], cpu.gpr[i]._64);
     if (i % 4 == 3) {
       printf("\n");
     }
   }
-  printf("pc: 0x%16lx\n", cpu.pc);
+  printf("pc: 0x%016lx\n", cpu.pc);
 }
 
 rtlreg_t isa_reg_str2val(const char *s, bool *success) {
diff --git a/src/monitor/cpu-exec.c b/src/monitor/cpu-exec.c
index ad25a3eb282a4b7272b2cb069117d6f401b50f1c..54e7fb96211f8e3852d49b6195c2551486544b9f 100644
--- a/src/monitor/cpu-exec.c
+++ b/src/monitor/cpu-exec.c
@@ -62,7 +62,7 @@ void cpu_exec(uint64_t n) {
     WP *wp = scan_watchpoint();
     if(wp != NULL) {
 #ifdef ISA64
-      printf("\n\nHint watchpoint %d at address 0x%16lx, expr = %s\n", wp->NO, ori_pc, wp->expr);
+      printf("\n\nHint watchpoint %d at address 0x%016lx, expr = %s\n", wp->NO, ori_pc, wp->expr);
 #else
       printf("\n\nHint watchpoint %d at address 0x%08x, expr = %s\n", wp->NO, ori_pc, wp->expr);
 #endif
@@ -88,7 +88,7 @@ void cpu_exec(uint64_t n) {
 
     case NEMU_END: case NEMU_ABORT:
 #ifdef ISA64
-      _Log("nemu: %s\33[0m at pc = 0x%16lx\n\n",
+      _Log("nemu: %s\33[0m at pc = 0x%016lx\n\n",
           (nemu_state.state == NEMU_ABORT ? "\33[1;31mABORT" :
            (nemu_state.halt_ret == 0 ? "\33[1;32mHIT GOOD TRAP" : "\33[1;31mHIT BAD TRAP")),
           nemu_state.halt_pc);
diff --git a/src/monitor/debug/expr.c b/src/monitor/debug/expr.c
index 23a0ab1b9699d601b3b405803e4e8feaccf46ed6..f870f01fb04c30413f952082234c23f73f15eede 100644
--- a/src/monitor/debug/expr.c
+++ b/src/monitor/debug/expr.c
@@ -175,7 +175,7 @@ static int find_dominated_op(int s, int e, bool *success) {
 
 uint32_t isa_reg_str2val(const char*, bool *);
 
-static uint64_t eval(int s, int e, bool *success) {
+static rtlreg_t eval(int s, int e, bool *success) {
   if (s > e) {
     // bad expression
     *success = false;
@@ -183,7 +183,7 @@ static uint64_t eval(int s, int e, bool *success) {
   }
   else if (s == e) {
     // single token
-    uint64_t val;
+    rtlreg_t val;
     switch (tokens[s].type) {
       case TK_REG: val = isa_reg_str2val(tokens[s].str + 1, success); // +1 to skip '$'
                 if (!*success) { return 0; }
@@ -205,7 +205,7 @@ static uint64_t eval(int s, int e, bool *success) {
 
     int op_type = tokens[dominated_op].type;
     if (op_type == '!' || op_type == TK_NEG || op_type == TK_REF) {
-      uint64_t val = eval(dominated_op + 1, e, success);
+      rtlreg_t val = eval(dominated_op + 1, e, success);
       if (!*success) { return 0; }
 
       switch (op_type) {
@@ -216,9 +216,9 @@ static uint64_t eval(int s, int e, bool *success) {
       }
     }
 
-    uint64_t val1 = eval(s, dominated_op - 1, success);
+    rtlreg_t val1 = eval(s, dominated_op - 1, success);
     if (!*success) { return 0; }
-    uint64_t val2 = eval(dominated_op + 1, e, success);
+    rtlreg_t val2 = eval(dominated_op + 1, e, success);
     if (!*success) { return 0; }
 
     switch (op_type) {
@@ -236,7 +236,7 @@ static uint64_t eval(int s, int e, bool *success) {
   }
 }
 
-uint64_t expr(char *e, bool *success) {
+rtlreg_t expr(char *e, bool *success) {
   if (!make_token(e)) {
     *success = false;
     return 0;
diff --git a/src/monitor/debug/log.c b/src/monitor/debug/log.c
index b027509fae04e8e87d8e588edb34591c3e457133..d917a858471ed642fbfe105d0099f24194fa62a1 100644
--- a/src/monitor/debug/log.c
+++ b/src/monitor/debug/log.c
@@ -23,10 +23,10 @@ void strcatf(char *buf, const char *fmt, ...) {
 
 void asm_print(vaddr_t ori_pc, int instr_len, bool print_flag) {
 #ifdef ISA64
-  snprintf(tempbuf, sizeof(tempbuf), "%16lx:   %s%*.s%s", ori_pc, log_bytebuf,
+  snprintf(tempbuf, sizeof(tempbuf), "%016lx:   %s%*.s%s", ori_pc, log_bytebuf,
       50 - (12 + 3 * instr_len), "", log_asmbuf);
 #else
-  snprintf(tempbuf, sizeof(tempbuf), "%08x:   %s%*.s%s", ori_pc, log_bytebuf,
+  snprintf(tempbuf, sizeof(tempbuf), "%8x:   %s%*.s%s", ori_pc, log_bytebuf,
       50 - (12 + 3 * instr_len), "", log_asmbuf);
 #endif
   log_write("%s\n", tempbuf);
diff --git a/src/monitor/debug/ui.c b/src/monitor/debug/ui.c
index 3362a65f61f3f9c51be920d0b6aa985b8e51777f..11fd05063455a80133d5f4bd53f9af6051cd9792 100644
--- a/src/monitor/debug/ui.c
+++ b/src/monitor/debug/ui.c
@@ -93,7 +93,7 @@ static int cmd_p(char *args) {
     if(success) { printf("0x%lx\n", r); }
 #else
     uint32_t r = expr(args, &success);
-    if(success) { printf("0x%x\n", r); }
+    if(success) { printf("%d\n", r); }
 #endif
     else { printf("Bad expression\n"); }
   }
@@ -128,7 +128,7 @@ static int cmd_x(char *args) {
       for (i = 0; i < n; i ++) {
         if (i % 4 == 0) {
 #ifdef ISA64
-          printf("0x%16lx: ", addr);
+          printf("0x%016lx: ", addr);
 #else
           printf("0x%08x: ", addr);
 #endif
diff --git a/tools/qemu-diff/src/diff-test.c b/tools/qemu-diff/src/diff-test.c
index 1c214ee2c59b2ebf31e0cc7604026160c361e55b..2589fb56ced56eac77be2f089320d82a72862634 100644
--- a/tools/qemu-diff/src/diff-test.c
+++ b/tools/qemu-diff/src/diff-test.c
@@ -68,6 +68,7 @@ void difftest_init(void) {
     printf("Connect to QEMU successfully\n");
 
     atexit(gdb_exit);
+    
     init_isa();
   }
 }
diff --git a/tools/qemu-diff/src/gdb-host.c b/tools/qemu-diff/src/gdb-host.c
index 7e30617716582a076fd51a3f9ee2f79753ac5821..341ab851ec5455c3b75fee1f1984f5e74b82f5a9 100644
--- a/tools/qemu-diff/src/gdb-host.c
+++ b/tools/qemu-diff/src/gdb-host.c
@@ -90,7 +90,6 @@ bool gdb_setregs(union isa_gdb_regs *r) {
 }
 
 bool gdb_si(void) {
-
   char buf[] = "vCont;s:1";
   gdb_send(conn, (const uint8_t *)buf, strlen(buf));
   size_t size;