diff --git a/npc/cmake/ChiselBuild.cmake b/npc/cmake/ChiselBuild.cmake
index 789b476..f446790 100644
--- a/npc/cmake/ChiselBuild.cmake
+++ b/npc/cmake/ChiselBuild.cmake
@@ -1,20 +1,19 @@
-# -- Add an always run target to generate verilog files with sbt/bloop,
-#    as we don't know if the result files will be different from cmake
-# NOTE: Must reconfigure if we add new files in SCALA_CORE directory
+# -- Add an always run target to generate verilog files with sbt/bloop, as we
+# don't know if the result files will be different from cmake NOTE: Must
+# reconfigure if we add new files in SCALA_CORE directory
 file(GLOB_RECURSE SCALA_CORE_SOURCES "${SCALA_CORE}/src/main/scala/*.scala")
 file(GLOB_RECURSE SCALA_CORE_RESOURCES "${SCALA_CORE}/src/main/resources/*")
 message(STATUS "Found scala source file: ${SCALA_CORE_SOURCES}")
-set(CHISEL_DEPENDENCY ${SCALA_CORE_SOURCES} ${SCALA_CORE_RESOURCES} ${SCALA_CORE}/build.sbt)
+set(CHISEL_DEPENDENCY ${SCALA_CORE_SOURCES} ${SCALA_CORE_RESOURCES}
+                      ${SCALA_CORE}/build.sbt)
 
 if(BUILD_USE_BLOOP)
+  message(STATUS "Building core using bloop")
   set(CHISEL_TARGET bloop_${TOPMODULE})
   set(CHISEL_TEST_TARGET bloop_${TOPMODULE}_test)
   # Export sbt build config to bloop
   if(NOT EXISTS ${SCALA_CORE}/.bloop)
-    execute_process(
-      COMMAND sbt bloopInstall
-      WORKING_DIRECTORY ${SCALA_CORE}
-    )
+    execute_process(COMMAND sbt bloopInstall WORKING_DIRECTORY ${SCALA_CORE})
   endif()
   string(REPLACE " " ";" CHISEL_EMIT_ARGS_LIST ${CHISEL_EMIT_ARGS})
   list(TRANSFORM CHISEL_EMIT_ARGS_LIST PREPEND "--args;")
@@ -24,14 +23,11 @@ if(BUILD_USE_BLOOP)
     WORKING_DIRECTORY ${SCALA_CORE}
     DEPENDS ${CHISEL_DEPENDENCY}
     COMMAND_EXPAND_LISTS
-    COMMENT "Run bloop from CMake"
-  )
-#   add_test(
-#     NAME bloop_${TOPMODULE}_test
-#     COMMAND bloop test
-#     WORKING_DIRECTORY ${SCALA_CORE}
-#   )
+    COMMENT "Run bloop from CMake")
+  # add_test( NAME bloop_${TOPMODULE}_test COMMAND bloop test WORKING_DIRECTORY
+  # ${SCALA_CORE} )
 else()
+  message(STATUS "Building core using sbt")
   set(CHISEL_TARGET sbt_${TOPMODULE})
   set(CHISEL_TEST_TARGET sbt_${TOPMODULE}_test)
   add_custom_command(
@@ -40,19 +36,16 @@ else()
     WORKING_DIRECTORY ${SCALA_CORE}
     DEPENDS ${CHISEL_DEPENDENCY}
     VERBATIM
-    COMMENT "Run sbt from CMake"
-  )
+    COMMENT "Run sbt from CMake")
   add_test(
     NAME sbt_${TOPMODULE}_test
     COMMAND sbt test
-    WORKING_DIRECTORY ${SCALA_CORE}
-  )
+    WORKING_DIRECTORY ${SCALA_CORE})
 endif()
 
 if(NOT EXISTS ${CHISEL_OUTPUT_TOPMODULE})
-  # Probably cold build, generate verilog at configure time to produce top module file
-  execute_process(
-    COMMAND sbt "run ${CHISEL_EMIT_ARGS}"
-    WORKING_DIRECTORY ${SCALA_CORE}
-  )
+  # Probably cold build, generate verilog at configure time to produce top
+  # module file
+  execute_process(COMMAND sbt "run ${CHISEL_EMIT_ARGS}"
+                  WORKING_DIRECTORY ${SCALA_CORE})
 endif()
diff --git a/npc/core/src/main/scala/components/ALU.scala b/npc/core/src/main/scala/components/ALU.scala
index 44e9c14..0c1f0ae 100644
--- a/npc/core/src/main/scala/components/ALU.scala
+++ b/npc/core/src/main/scala/components/ALU.scala
@@ -6,7 +6,8 @@ import shapeless.{HNil, ::}
 
 class ALUControlInterface extends Bundle {
   object OpSelect extends ChiselEnum {
-    val aOpAdd, aOpSub, aOpNot, aOpAnd, aOpOr, aOpXor, aOpSlt, aOpSltu, aOpSll, aOpSrl, aOpSra = Value
+    val aOpAdd, aOpSub, aOpNot, aOpAnd, aOpOr, aOpXor, aOpSlt, aOpSltu, aOpSll,
+        aOpSrl, aOpSra = Value
   }
   object SrcASelect extends ChiselEnum {
     val aSrcARs1, aSrcAPc, aSrcAZero = Value
@@ -54,19 +55,21 @@ class ALU[T <: UInt](tpe: T) extends Module {
 
   import control.OpSelect._
 
-  out.result := MuxLookup(control.op, 0.U)(Seq(
-    aOpAdd -> add,
-    aOpSub -> sub,
-    aOpNot -> not,
-    aOpAnd -> and,
-    aOpOr  -> or,
-    aOpXor -> xor,
-    aOpSlt -> slt,
-    aOpSltu -> sltu,
-    aOpSll -> sll,
-    aOpSrl -> srl,
-    aOpSra -> sra.asUInt
-  ))
+  out.result := MuxLookup(control.op, 0.U)(
+    Seq(
+      aOpAdd -> add,
+      aOpSub -> sub,
+      aOpNot -> not,
+      aOpAnd -> and,
+      aOpOr -> or,
+      aOpXor -> xor,
+      aOpSlt -> slt,
+      aOpSltu -> sltu,
+      aOpSll -> sll,
+      aOpSrl -> srl,
+      aOpSra -> sra.asUInt
+    )
+  )
 }
 
 object ALU {
diff --git a/npc/core/src/main/scala/components/Mem.scala b/npc/core/src/main/scala/components/Mem.scala
index 24f9927..a9d3e22 100644
--- a/npc/core/src/main/scala/components/Mem.scala
+++ b/npc/core/src/main/scala/components/Mem.scala
@@ -1,13 +1,15 @@
 package flow.components
 
 import chisel3._
+import chisel3.experimental.noPrefix
 import chisel3.util.HasBlackBoxPath
 import chisel3.util.HasBlackBoxResource
 import chisel3.util.log2Ceil
-import chisel3.experimental.noPrefix
-import scala.collection.SeqMap
 import flow.components
-import shapeless.{HNil, ::}
+import shapeless.::
+import shapeless.HNil
+
+import scala.collection.SeqMap
 
 class RamControlInterface(addrWidth: Int) extends Bundle {
   val valid = Input(Bool())
diff --git a/npc/csrc/Flow/config.cpp b/npc/csrc/Flow/config.cpp
index 035fe24..e8e224b 100644
--- a/npc/csrc/Flow/config.cpp
+++ b/npc/csrc/Flow/config.cpp
@@ -11,9 +11,6 @@ void Config::cli_parse(int argc, char **argv) {
                "Memory file is in text format");
   app.add_option("--wav", wavefile, "output .vcd file path");
   app.add_flag("-i", interactive, "Launch sdb for interactive session");
-  app.add_option("--diff-lib", lib_ref,
-                 "Dynamic library file of difftest reference")
-      ->check(CLI::ExistingFile);
 
   try {
     app.parse(argc, argv);
diff --git a/npc/include/components.hpp b/npc/include/components.hpp
index cfd05f1..887ded0 100644
--- a/npc/include/components.hpp
+++ b/npc/include/components.hpp
@@ -14,32 +14,48 @@
 #include <string>
 
 template <typename T, std::size_t nr> class _RegistersBase {
-  std::array<T, nr> regs;
-  T pc;
   std::shared_ptr<spdlog::logger> logger = spdlog::stdout_color_mt("registers");
 
-  virtual T fetch_pc() const;
-  virtual T fetch_reg(std::size_t id) const;
+  virtual T fetch_pc() const = 0;
+  virtual T fetch_reg(std::size_t id) const = 0;
 
 public:
   T operator[](size_t id) const { return fetch_reg(id); }
   T get_pc() const { return fetch_pc(); }
-  void update() {
-    for (int i = 0; i < regs.size(); i++) {
-      regs[i] = fetch_reg(i);
-    }
-  }
 };
 
 template <std::size_t n> class Memory {
-  paddr_t pmem_start, pmem_end;
-
 public:
   std::array<uint8_t, n> mem;
-  // TODO: Read memory file before init and use memcpy to initialize memory.
+
   Memory(std::filesystem::path filepath, bool is_binary, paddr_t pmem_start,
          paddr_t pmem_end)
       : pmem_start(pmem_start), pmem_end(pmem_end) {
+    read_memory(filepath, is_binary);
+  }
+
+  const word_t &operator[](std::size_t addr) { return this->read(addr); }
+
+  void transfer(paddr_t addr, uint8_t data[], size_t len, bool is_write) {
+    if (is_write) {
+      // memcpy(guest_to_host(addr), data, len);
+      size_t offset = (addr - pmem_start);
+      std::copy(data, data + len, &mem[offset]);
+    } else {
+      // memcpy(data, guest_to_host(addr), len);
+      size_t offset = (addr - pmem_start);
+      std::copy(&mem[offset], &mem[offset + len], data);
+    }
+  }
+
+  bool in_pmem(paddr_t addr) const {
+    return addr >= pmem_start && addr <= pmem_end;
+  }
+
+private:
+  paddr_t pmem_start, pmem_end;
+
+  void read_memory(std::filesystem::path filepath, bool is_binary) {
     if (!std::filesystem::exists(filepath))
       throw std::runtime_error("Memory file not found");
     if (is_binary) {
@@ -55,30 +71,12 @@ public:
       }
     }
   }
-  const word_t &operator[](std::size_t addr) { return this->read(addr); }
-  void transfer(paddr_t addr, uint8_t data[], size_t len, bool is_write) {
-    if (is_write) {
-      // memcpy(guest_to_host(addr), data, len);
-      size_t offset = (addr - pmem_start);
-      std::copy(data, data + len, &mem[offset]);
-    } else {
-      // memcpy(data, guest_to_host(addr), len);
-      size_t offset = (addr - pmem_start);
-      std::copy(&mem[offset], &mem[offset + len], data);
-    }
-  }
-  bool in_pmem(paddr_t addr) const {
-    return addr >= pmem_start && addr <= pmem_end;
-  }
 };
 
 template <typename Mem, typename DevMap> class MemoryMap {
-  std::unique_ptr<Mem> ram;
-  std::unique_ptr<DevMap> devices;
-  std::shared_ptr<spdlog::logger> logger = spdlog::stdout_color_mt("mmap");
-
 public:
-  MemoryMap(std::unique_ptr<Mem> &&ram, std::unique_ptr<DevMap> &&devices)
+  MemoryMap(std::unique_ptr<Mem> &&ram,
+            std::unique_ptr<DevMap> &&devices) noexcept
       : ram(std::move(ram)), devices(std::move(devices)) {}
 
   void write(paddr_t waddr, word_t wdata, char wmask) {
@@ -122,5 +120,10 @@ public:
   void trace(paddr_t addr, bool is_read, word_t pc = 0, word_t value = 0) {
     logger->trace("[{}] 0x{:x}", is_read ? 'R' : 'W', this->read(addr));
   }
+
+private:
+  std::unique_ptr<Mem> ram;
+  std::unique_ptr<DevMap> devices;
+  std::shared_ptr<spdlog::logger> logger = spdlog::stdout_color_mt("mmap");
 };
 #endif
diff --git a/npc/include/config.hpp b/npc/include/config.hpp
index 8c12f31..8d6bf33 100644
--- a/npc/include/config.hpp
+++ b/npc/include/config.hpp
@@ -12,7 +12,6 @@ struct Config {
   bool interactive{false};
   bool memory_file_binary = {true};
   std::filesystem::path wavefile;
-  std::filesystem::path lib_ref;
   void cli_parse(int argc, char **argv);
 };
 
diff --git a/npc/include/trm_difftest.hpp b/npc/include/trm_difftest.hpp
deleted file mode 100644
index ef8bde0..0000000
--- a/npc/include/trm_difftest.hpp
+++ /dev/null
@@ -1,102 +0,0 @@
-#ifndef _DIFFTEST_DIFFTEST_H_
-#define _DIFFTEST_DIFFTEST_H_
-#include "disasm.hpp"
-#include "types.h"
-#include <cassert>
-#include <components.hpp>
-#include <cstdint>
-#include <cstdlib>
-#include <dlfcn.h>
-#include <filesystem>
-#include <functional>
-#include <iostream>
-#include <map>
-#include <memory>
-#include <ostream>
-#include <stdexcept>
-#include <trm_interface.hpp>
-
-using paddr_t = uint32_t;
-struct DifftestTrmInterface : public TrmInterface {
-  TrmInterface &dut;
-  TrmInterface &ref;
-
-  DifftestTrmInterface(TrmInterface &dut, TrmInterface &ref, void *mem,
-                       size_t mem_size)
-      : dut(dut), ref(ref) {
-    init = [this, mem, mem_size](int n) {
-      this->ref.init(n);
-      this->dut.init(n);
-      paddr_t reset_vector = 0x80000000;
-      this->ref.memcpy(reset_vector, mem, mem_size, TRM_TO_MACHINE);
-      this->dut.memcpy(reset_vector, mem, mem_size, TRM_TO_MACHINE);
-      fetch_state();
-    };
-    exec = [this](uint64_t n) {
-      bool enable_disasm = true;
-      if (n > 30) {
-        enable_disasm = false;
-      }
-
-      while (n--) {
-        word_t pc = this->ref.at("pc");
-        word_t inst = this->ref.at(pc);
-        if (enable_disasm)
-          std::cout << d.disassemble(pc, (uint8_t *)&inst, WORD_BYTES)
-                    << std::endl;
-        if (inst == 1048691) {
-          // ebreak
-          throw TrmRuntimeException(TrmRuntimeException::EBREAK, "ebreak");
-        }
-        this->ref.exec(1);
-        this->dut.exec(1);
-        this->ref.fetch_state();
-        this->dut.fetch_state();
-        if (*(CPUState *)this->ref.cpu_state !=
-            *(CPUState *)this->dut.cpu_state) {
-          throw TrmRuntimeException(TrmRuntimeException::DIFFTEST_FAILED,
-                                    "Difftest failed");
-        }
-      }
-    };
-
-    // NOTE: Different from normal Trm, we copy 2 * sizeof(CPUState) to/from p,
-    // which represents ref_state and dut state
-    regcpy = [this](void *p, bool direction) {
-      // this->ref.regcpy(p, direction);
-      // this->dut.regcpy(p, direction);
-    };
-
-    memcpy = [this](paddr_t paddr, void *p, size_t n, bool direction) {
-      this->dut.memcpy(paddr, p, n, direction);
-      this->ref.memcpy(paddr, (uint8_t *)p + n, n, direction);
-    };
-  }
-
-  word_t at(std::string name) const override {
-    if (name.empty()) {
-      throw std::runtime_error("Empty register name");
-    } else if (name[0] == 'r') {
-      std::cout << name.substr(1) << std::endl;
-      this->ref.at(name.substr(1));
-    } else if (name[0] == 'd') {
-      this->dut.at(name.substr(1));
-    } else {
-      throw std::runtime_error("Register name provided to difftest interface "
-                               "must start with r or d.");
-    }
-    return 0;
-  }
-
-  word_t at(paddr_t addr) const override {
-    std::cout << ref.at(addr) << "\t" << dut.at(addr) << std::endl;
-    return dut.at(addr);
-  }
-
-  void print(std::ostream &os) const override {
-    os << "REF state:\n"
-       << *(CPUState *)ref.cpu_state << "DUT state:\n"
-       << *(CPUState *)dut.cpu_state << std::endl;
-  }
-};
-#endif
\ No newline at end of file
diff --git a/npc/include/trm_interface.hpp b/npc/include/trm_interface.hpp
deleted file mode 100644
index 5410820..0000000
--- a/npc/include/trm_interface.hpp
+++ /dev/null
@@ -1,158 +0,0 @@
-#ifndef _NPC_TRM_INTERFACE_HEADER_FILE_
-#define _NPC_TRM_INTERFACE_HEADER_FILE_
-#include <disasm.hpp>
-#include <dlfcn.h>
-#include <filesystem>
-#include <functional>
-#include <stdexcept>
-#include <string>
-#include <types.h>
-
-extern Disassembler d;
-
-template <typename R, size_t nr_reg> struct CPUStateBase {
-  R reg[nr_reg] = {0};
-  word_t pc = 0x80000000;
-
-  static const std::map<std::string, int> inline regs_by_name =
-      riscv32_regs_by_name;
-  CPUStateBase() {
-    for (int i = 0; i < nr_reg; i++)
-      reg[i] = 0;
-  }
-
-  bool operator==(const CPUStateBase &other) const {
-    if (pc != other.pc)
-      return false;
-    for (int i = 0; i < nr_reg; ++i) {
-      if (reg[i] != other.reg[i])
-        return false;
-    }
-    return true;
-  }
-
-  bool operator!=(const CPUStateBase &other) const {
-    return !(*this == other); // Reuse the == operator for != implementation
-  }
-
-  /* This does not update the register!!! */
-  R at(std::string name) {
-    // FIXME: Using this to get pc seems broken
-    return name == "pc" ? pc : reg[regs_by_name.at(name)];
-  }
-
-  uint32_t reg_str2val(const char *name, bool *success) {
-    try {
-      *success = true;
-      return this->at(name);
-    } catch (std::runtime_error) {
-      *success = false;
-      return 0;
-    }
-  }
-};
-
-template <typename R, size_t nr_reg>
-std::ostream &operator<<(std::ostream &os, const CPUStateBase<R, nr_reg> &cpu) {
-  os << "PC: " << std::hex << cpu.pc << std::endl;
-  for (int i = 0; i < nr_reg; i++) {
-    os << "reg " << std::dec << std::setw(2) << i << ":" << std::hex
-       << std::setw(10) << cpu.reg[i];
-    if (i % 4 == 3) {
-      os << std::endl;
-    } else {
-      os << " | ";
-    }
-  }
-  return os;
-}
-
-using CPUState = CPUStateBase<word_t, REG_COUNT>;
-
-enum { TRM_FROM_MACHINE, TRM_TO_MACHINE };
-
-class TrmInterface {
-protected:
-  using memcpy_t = void (*)(paddr_t, void *, size_t, bool);
-  using regcpy_t = void (*)(void *, bool);
-  using exec_t = void (*)(uint64_t);
-  using init_t = void (*)(int);
-  std::function<void(void *, bool)> regcpy;
-
-public:
-  std::function<void(uint64_t)> exec;
-  std::function<void(int)> init;
-  // TODO: paddr_t can probably changed to (void *)?
-  std::function<void(paddr_t, void *, size_t, bool)> memcpy;
-  // Managed by callee
-  void *cpu_state;
-
-  TrmInterface() {}
-  TrmInterface(memcpy_t f_memcpy, regcpy_t f_regcpy, exec_t f_exec,
-               init_t f_init, void *cpu_state)
-      : memcpy(f_memcpy), regcpy(f_regcpy), exec(f_exec), init(f_init),
-        cpu_state(cpu_state) {}
-
-  void fetch_state() { this->regcpy(cpu_state, TRM_FROM_MACHINE); }
-  void push_state() { this->regcpy(cpu_state, TRM_TO_MACHINE); }
-  virtual word_t at(std::string) const = 0;
-  virtual word_t at(word_t addr) const = 0;
-  virtual void print(std::ostream &os) const = 0;
-};
-
-class TrmRuntimeException : public std::exception {
-private:
-  const char *msg_;
-  int code_;
-
-public:
-  enum { EBREAK, DIFFTEST_FAILED };
-  TrmRuntimeException(int code, const char *message)
-      : code_(code), msg_(message) {}
-
-  virtual const char *what() const throw() { return msg_; }
-
-  int error_code() const { return code_; }
-};
-
-struct RefTrmInterface : TrmInterface {
-  RefTrmInterface(std::filesystem::path lib_file) {
-    void *handle = dlopen(lib_file.c_str(), RTLD_LAZY);
-    if (handle == nullptr) {
-      throw std::runtime_error("Failed to open diff library file");
-    };
-    memcpy = (memcpy_t)dlsym(handle, "difftest_memcpy");
-    if (handle == nullptr) {
-      throw std::runtime_error("Failed to find `difftest_memcpy`");
-    };
-    regcpy = (regcpy_t)dlsym(handle, "difftest_regcpy");
-    if (handle == nullptr) {
-      throw std::runtime_error("Failed to find `difftest_regcpy`");
-    };
-    exec = (exec_t)dlsym(handle, "difftest_exec");
-    if (handle == nullptr) {
-      throw std::runtime_error("Failed to find `difftest_exec`");
-    };
-    init = (init_t)dlsym(handle, "difftest_init");
-    if (handle == nullptr) {
-      throw std::runtime_error("Failed to find `difftest_init`");
-    };
-    cpu_state = new CPUState{};
-  }
-
-  ~RefTrmInterface() { delete (CPUState *)cpu_state; }
-
-  word_t at(std::string name) const override {
-    return ((CPUState *)cpu_state)->at(name);
-  }
-
-  word_t at(paddr_t addr) const override {
-    word_t buf;
-    this->memcpy(addr, &buf, sizeof(word_t), TRM_FROM_MACHINE);
-    return buf;
-  }
-
-  void print(std::ostream &os) const override { os << *(CPUState *)cpu_state; }
-};
-
-#endif
diff --git a/npc/include/types.h b/npc/include/types.h
index fcc14e4..f34f168 100644
--- a/npc/include/types.h
+++ b/npc/include/types.h
@@ -4,8 +4,7 @@
 extern "C" {
 #endif
 #include <gdbstub.h>
-#include <inttypes.h>
-#include <stddef.h>
+#include <stdint.h>
 
 typedef uint32_t word_t;
 typedef int32_t sword_t;
@@ -15,10 +14,8 @@ static const sword_t SWORD_T_MIN = INT32_MIN;
 #define WORD_BYTES 4
 #define REG_COUNT 32
 
-#define FMT_WORD "0x%08x"
 typedef uint32_t vaddr_t;
 typedef uint32_t paddr_t;
-#define FMT_ADDR "0x%08x"
 typedef uint16_t ioaddr_t;
 
 struct Breakpoint {
@@ -31,18 +28,8 @@ struct Breakpoint {
 #endif
 
 #ifdef __cplusplus
-#include <map>
-#include <string>
 #include <vector>
 
-const std::map<std::string, int> riscv32_regs_by_name{
-    {"$0", 0},  {"ra", 1},  {"sp", 2},   {"gp", 3},   {"tp", 4},  {"t0", 5},
-    {"t1", 6},  {"t2", 7},  {"s0", 8},   {"s1", 9},   {"a0", 10}, {"a1", 11},
-    {"a2", 12}, {"a3", 13}, {"a4", 14},  {"a5", 15},  {"a6", 16}, {"a7", 17},
-    {"s2", 18}, {"s3", 19}, {"s4", 20},  {"s5", 21},  {"s6", 22}, {"s7", 23},
-    {"s8", 24}, {"s9", 25}, {"s10", 26}, {"s11", 27}, {"t3", 28}, {"t4", 29},
-    {"t5", 30}, {"t6", 31}};
-
 struct DbgState {
   std::vector<Breakpoint> *bp;
 };
diff --git a/npc/include/vpi_wrapper.hpp b/npc/include/vpi_wrapper.hpp
index 25f26fb..22b1876 100644
--- a/npc/include/vpi_wrapper.hpp
+++ b/npc/include/vpi_wrapper.hpp
@@ -23,6 +23,11 @@ class _RegistersVPI : public _RegistersBase<T, nr> {
 public:
   _RegistersVPI<T, nr>(const std::string regs_prefix,
                        const std::string pcname) {
+    init_handlers(regs_prefix, pcname);
+  }
+
+private:
+  void init_handlers(const std::string regs_prefix, const std::string pcname) {
     for (int i = 0; i < nr; i++) {
       std::string regname = regs_prefix + std::to_string(i);
       vpiHandle vh = vpi_handle_by_name((PLI_BYTE8 *)regname.c_str(), nullptr);