diff --git a/.vscode/settings.json b/.vscode/settings.json
new file mode 100644
index 0000000..bff7192
--- /dev/null
+++ b/.vscode/settings.json
@@ -0,0 +1,75 @@
+{
+    "files.associations": {
+        "array": "cpp",
+        "__bit_reference": "cpp",
+        "__config": "cpp",
+        "__debug": "cpp",
+        "__errc": "cpp",
+        "__functional_base": "cpp",
+        "__hash_table": "cpp",
+        "__locale": "cpp",
+        "__mutex_base": "cpp",
+        "__node_handle": "cpp",
+        "__nullptr": "cpp",
+        "__split_buffer": "cpp",
+        "__string": "cpp",
+        "__threading_support": "cpp",
+        "__tree": "cpp",
+        "__tuple": "cpp",
+        "algorithm": "cpp",
+        "atomic": "cpp",
+        "bit": "cpp",
+        "bitset": "cpp",
+        "cctype": "cpp",
+        "chrono": "cpp",
+        "cmath": "cpp",
+        "complex": "cpp",
+        "csignal": "cpp",
+        "cstdarg": "cpp",
+        "cstddef": "cpp",
+        "cstdint": "cpp",
+        "cstdio": "cpp",
+        "cstdlib": "cpp",
+        "cstring": "cpp",
+        "ctime": "cpp",
+        "cwchar": "cpp",
+        "cwctype": "cpp",
+        "deque": "cpp",
+        "exception": "cpp",
+        "fstream": "cpp",
+        "functional": "cpp",
+        "initializer_list": "cpp",
+        "iomanip": "cpp",
+        "ios": "cpp",
+        "iosfwd": "cpp",
+        "iostream": "cpp",
+        "istream": "cpp",
+        "iterator": "cpp",
+        "limits": "cpp",
+        "locale": "cpp",
+        "map": "cpp",
+        "memory": "cpp",
+        "mutex": "cpp",
+        "new": "cpp",
+        "numeric": "cpp",
+        "optional": "cpp",
+        "ostream": "cpp",
+        "random": "cpp",
+        "ratio": "cpp",
+        "set": "cpp",
+        "sstream": "cpp",
+        "stack": "cpp",
+        "stdexcept": "cpp",
+        "streambuf": "cpp",
+        "string": "cpp",
+        "string_view": "cpp",
+        "system_error": "cpp",
+        "tuple": "cpp",
+        "type_traits": "cpp",
+        "typeinfo": "cpp",
+        "unordered_map": "cpp",
+        "utility": "cpp",
+        "vector": "cpp"
+    },
+    "C_Cpp.errorSquiggles": "Disabled"
+}
\ No newline at end of file
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 75c5a1a..94f4e3d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -21,6 +21,7 @@ else ()
     message(STATUS "Compiler is recent enough to support C++17.")
 endif ()
 
+# set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++1z -pthread -msse4.2 -mbmi2 -Wall")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++1z -pthread -Wall")
 set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -DNDEBUG -march=native -O3")
 set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -fsanitize=address -fno-omit-frame-pointer -O0 -g -DDEBUG")
diff --git a/include/xcdat/bit_tools.hpp b/include/xcdat/bit_tools.hpp
index 559993f..b814b40 100644
--- a/include/xcdat/bit_tools.hpp
+++ b/include/xcdat/bit_tools.hpp
@@ -3,8 +3,13 @@
 #include <cstdint>
 #include <cstdlib>
 
-#include <immintrin.h>
+#ifdef __SSE4_2__
 #include <nmmintrin.h>
+#endif
+
+#ifdef __BMI2__
+#include <immintrin.h>
+#endif
 
 // From https://github.com/ot/succinct
 namespace xcdat::bit_tools {
@@ -28,8 +33,37 @@ inline std::uint64_t popcount(std::uint64_t x) {
 #endif
 }
 
+static constexpr std::uint8_t debruijn64_mapping[64] = {
+    63, 0,  58, 1,  59, 47, 53, 2,  60, 39, 48, 27, 54, 33, 42, 3,  61, 51, 37, 40, 49, 18,
+    28, 20, 55, 30, 34, 11, 43, 14, 22, 4,  62, 57, 46, 52, 38, 26, 32, 41, 50, 36, 17, 19,
+    29, 10, 13, 21, 56, 45, 25, 31, 35, 16, 9,  12, 44, 24, 15, 8,  23, 7,  6,  5,
+};
+
+static constexpr std::uint64_t debruijn64 = 0x07EDD5E59A4E28C2ULL;
+
+// return the position of the single bit set in the word x
+inline std::uint8_t bit_position(std::uint64_t x) {
+    return debruijn64_mapping[(x * debruijn64) >> 58];
+}
+
 inline std::uint64_t msb(std::uint64_t x) {
+#ifdef __SSE4_2__
     return x == 0 ? 0 : 63 - __builtin_clzll(x);
+#else
+    if (x == 0) {
+        return 0;
+    }
+    // right-saturate the word
+    x |= x >> 1;
+    x |= x >> 2;
+    x |= x >> 4;
+    x |= x >> 8;
+    x |= x >> 16;
+    x |= x >> 32;
+    // isolate the MSB
+    x ^= x >> 1;
+    return bit_position(x);
+#endif
 }
 
 inline std::uint64_t uleq_step_9(std::uint64_t x, std::uint64_t y) {
diff --git a/include/xcdat/code_table.hpp b/include/xcdat/code_table.hpp
new file mode 100644
index 0000000..4604c2b
--- /dev/null
+++ b/include/xcdat/code_table.hpp
@@ -0,0 +1,85 @@
+#pragma once
+
+#include <array>
+#include <string_view>
+
+#include "mm_vector.hpp"
+
+namespace xcdat {
+
+class code_table {
+  private:
+    std::uint64_t m_max_length = 0;
+    std::array<std::uint8_t, 512> m_table;
+    mm_vector<std::uint8_t> m_alphabet;
+
+    struct cf_type {
+        std::uint8_t ch;
+        std::uint64_t freq;
+    };
+
+  public:
+    code_table() = default;
+
+    virtual ~code_table() = default;
+
+    void build(const std::vector<std::string_view>& keys) {
+        std::array<cf_type, 256> counter;
+        for (std::uint32_t ch = 0; ch < 256; ++ch) {
+            counter[ch] = {static_cast<std::uint8_t>(ch), 0};
+        }
+
+        m_max_length = 0;
+        for (const auto& key : keys) {
+            for (std::uint8_t ch : key) {
+                counter[ch].freq += 1;
+            }
+            m_max_length = std::max<std::uint64_t>(m_max_length, key.length());
+        }
+
+        {
+            std::vector<std::uint8_t> alphabet;
+            for (const auto& cf : counter) {
+                if (cf.freq != 0) {
+                    alphabet.push_back(cf.ch);
+                }
+            }
+            m_alphabet.steal(alphabet);
+        }
+
+        std::sort(counter.begin(), counter.end(), [](const cf_type& a, const cf_type& b) { return a.freq > b.freq; });
+
+        for (std::uint32_t ch = 0; ch < 256; ++ch) {
+            m_table[counter[ch].ch] = static_cast<std::uint8_t>(ch);
+        }
+        for (std::uint32_t ch = 0; ch < 256; ++ch) {
+            m_table[m_table[ch] + 256] = static_cast<std::uint8_t>(ch);
+        }
+    }
+
+    inline std::uint64_t alphabet_size() const {
+        return m_alphabet.size();
+    }
+
+    inline std::uint64_t max_length() const {
+        return m_max_length;
+    }
+
+    inline std::uint8_t get_code(char ch) const {
+        return m_table[static_cast<std::uint8_t>(ch)];
+    }
+
+    inline char get_char(std::uint8_t cd) const {
+        return static_cast<char>(m_table[cd + 256]);
+    }
+
+    inline auto begin() const {
+        return m_alphabet.begin();
+    }
+
+    inline auto end() const {
+        return m_alphabet.end();
+    }
+};
+
+}  // namespace xcdat
diff --git a/include/xcdat/compact_vector.hpp b/include/xcdat/compact_vector.hpp
new file mode 100644
index 0000000..fac1d1b
--- /dev/null
+++ b/include/xcdat/compact_vector.hpp
@@ -0,0 +1,71 @@
+#pragma once
+
+#include "mm_vector.hpp"
+#include "utils.hpp"
+
+namespace xcdat {
+
+class compact_vector {
+  private:
+    std::uint64_t m_size = 0;
+    std::uint64_t m_bits = 0;
+    std::uint64_t m_mask = 0;
+    mm_vector<std::uint64_t> m_chunks;
+
+  public:
+    compact_vector() = default;
+
+    template <class Vec>
+    compact_vector(const Vec& vec) {
+        build(vec);
+    }
+
+    virtual ~compact_vector() = default;
+
+    template <class Vec>
+    void build(const Vec& vec) {
+        const std::uint64_t maxv = *std::max_element(vec.begin(), vec.end());
+
+        m_size = vec.size();
+        m_bits = utils::bits_for_int(maxv);
+        m_mask = (1ULL << m_bits) - 1;
+
+        std::vector<std::uint64_t> chunks(utils::words_for_bits(m_size * m_bits));
+
+        for (std::uint64_t i = 0; i < m_size; i++) {
+            const auto [quo, mod] = utils::decompose<64>(i * m_bits);
+            chunks[quo] &= ~(m_mask << mod);
+            chunks[quo] |= (vec[i] & m_mask) << mod;
+            if (64 < mod + m_bits) {
+                const std::uint64_t diff = 64ULL - mod;
+                chunks[quo + 1] &= ~(m_mask >> diff);
+                chunks[quo + 1] |= (vec[i] & m_mask) >> diff;
+            }
+        }
+        m_chunks.steal(chunks);
+    }
+
+    inline std::uint64_t operator[](std::uint64_t i) const {
+        assert(i < m_size);
+        const auto [quo, mod] = utils::decompose<64>(i * m_bits);
+        if (mod + m_bits <= 64) {
+            return (m_chunks[quo] >> mod) & m_mask;
+        } else {
+            return ((m_chunks[quo] >> mod) | (m_chunks[quo + 1] << (64 - mod))) & m_mask;
+        }
+    }
+
+    inline std::uint64_t size() const {
+        return m_size;
+    }
+
+    inline std::uint64_t bits() const {
+        return m_bits;
+    }
+
+    inline std::uint64_t memory_in_bytes() const {
+        return m_chunks.size() * sizeof(std::uint64_t);
+    }
+};
+
+}  // namespace xcdat
\ No newline at end of file
diff --git a/include/xcdat/dac_bc.hpp b/include/xcdat/dac_bc.hpp
new file mode 100644
index 0000000..b30da58
--- /dev/null
+++ b/include/xcdat/dac_bc.hpp
@@ -0,0 +1,123 @@
+#pragma once
+
+#include <array>
+
+#include "bit_vector.hpp"
+#include "compact_vector.hpp"
+
+namespace xcdat {
+
+class dac_bc {
+  public:
+    static constexpr std::uint32_t l1_bits = 8;
+    static constexpr std::uint32_t max_levels = sizeof(std::uint64_t);
+
+  private:
+    std::uint32_t m_num_levels = 0;
+    std::uint64_t m_num_frees = 0;
+    std::array<mm_vector<std::uint8_t>, max_levels> m_bytes;
+    std::array<bit_vector, max_levels - 1> m_next_flags;
+    compact_vector m_links;
+    bit_vector m_leaf_flags;
+
+  public:
+    dac_bc() = default;
+
+    template <class BcUnit>
+    dac_bc(const std::vector<BcUnit>& bc_units, bit_vector::builder&& leaf_flags) {
+        std::array<std::vector<std::uint8_t>, max_levels> bytes;
+        std::array<bit_vector::builder, max_levels - 1> next_flags;
+        std::vector<std::uint64_t> links;
+
+        bytes[0].reserve(bc_units.size() * 2);
+        next_flags[0].reserve(bc_units.size() * 2);
+        links.reserve(bc_units.size());
+
+        m_num_levels = 0;
+
+        auto append_unit = [&](std::uint64_t x) {
+            std::uint32_t j = 0;
+            bytes[j].push_back(static_cast<std::uint8_t>(x & 0xFF));
+            next_flags[j].push_back(true);
+            x >>= 8;
+            while (x) {
+                ++j;
+                bytes[j].push_back(static_cast<std::uint8_t>(x & 0xFF));
+                next_flags[j].push_back(true);
+                x >>= 8;
+            }
+            next_flags[j].set_bit(next_flags[j].size() - 1, false);
+            m_num_levels = std::max(m_num_levels, j);
+        };
+
+        auto append_leaf = [&](std::uint64_t x) {
+            bytes[0].push_back(static_cast<std::uint8_t>(x & 0xFF));
+            next_flags[0].push_back(false);
+            links.push_back(x >> 8);
+        };
+
+        for (std::uint64_t i = 0; i < bc_units.size(); ++i) {
+            if (leaf_flags[i]) {
+                append_leaf(bc_units[i].base);
+            } else {
+                append_unit(bc_units[i].base ^ i);
+            }
+            append_unit(bc_units[i].check ^ i);
+            if (bc_units[i].check == i) {
+                m_num_frees += 1;
+            }
+        }
+
+        // release
+        for (uint8_t i = 0; i < m_num_levels; ++i) {
+            m_bytes[i].steal(bytes[i]);
+            m_next_flags[i].build(next_flags[i], true, false);
+        }
+        m_bytes[m_num_levels].steal(bytes[m_num_levels]);
+        m_links.build(links);
+        m_leaf_flags.build(leaf_flags, true, false);
+    }
+
+    virtual ~dac_bc() = default;
+
+    inline std::uint64_t base(std::uint64_t i) const {
+        return access(i * 2) ^ i;
+    }
+
+    inline std::uint64_t check(std::uint64_t i) const {
+        return access(i * 2 + 1) ^ i;
+    }
+
+    inline std::uint64_t link(std::uint64_t i) const {
+        return m_bytes[0][i * 2] | (m_links[m_leaf_flags.rank(i)] << 8);
+    }
+
+    inline bool is_leaf(std::uint64_t i) const {
+        return m_leaf_flags[i];
+    }
+
+    inline bool is_used(std::uint64_t i) const {
+        return check(i) != i;
+    }
+
+    inline std::uint64_t num_units() const {
+        return m_bytes[0].size() / 2;
+    }
+
+    inline std::uint64_t num_leaves() const {
+        return m_leaf_flags.num_ones();
+    }
+
+  private:
+    inline std::uint64_t access(std::uint64_t i) const {
+        std::uint32_t j = 0;
+        std::uint64_t x = m_bytes[j][i];
+        while (j < m_num_levels and m_next_flags[j][i]) {
+            i = m_next_flags[j++].rank(i);
+            x |= static_cast<std::uint64_t>(m_bytes[j][i]) << (j * 8);
+        }
+        return x;
+    }
+};
+
+}  // namespace xcdat
\ No newline at end of file
diff --git a/include/xcdat/shared_tail.hpp b/include/xcdat/shared_tail.hpp
new file mode 100644
index 0000000..6c7b7c0
--- /dev/null
+++ b/include/xcdat/shared_tail.hpp
@@ -0,0 +1,147 @@
+#pragma once
+
+#include <algorithm>
+#include <functional>
+#include <string>
+#include <string_view>
+#include <vector>
+
+#include "bit_vector.hpp"
+#include "mm_vector.hpp"
+
+namespace xcdat {
+
+//! TAIL implementation with suffix merge
+class shared_tail {
+  public:
+    struct suffix_type {
+        std::string_view str;
+        std::uint64_t npos;
+
+        inline char operator[](std::uint64_t i) const {
+            return str[length() - i - 1];
+        }
+        inline std::uint64_t length() const {
+            return str.length();
+        }
+
+        inline const char* begin() const {
+            return str.data();
+        }
+        inline const char* end() const {
+            return str.data() + str.length();
+        }
+
+        inline std::reverse_iterator<const char*> rbegin() const {
+            return std::make_reverse_iterator(str.data() + str.length());
+        }
+        inline std::reverse_iterator<const char*> rend() const {
+            return std::make_reverse_iterator(str.data());
+        }
+    };
+
+    class builder {
+      private:
+        // Buffer
+        std::vector<suffix_type> m_suffixes;
+
+        // Released
+        std::vector<char> m_chars;
+        bit_vector::builder m_term_flags;
+
+      public:
+        builder() = default;
+
+        virtual ~builder() = default;
+
+        void set_suffix(std::string_view str, std::uint64_t npos) {
+            m_suffixes.push_back({str, npos});
+        }
+
+        // setter(npos, tpos): Set units[npos].base = tpos.
+        void complete(bool bin_mode, const std::function<void(std::uint64_t, std::uint64_t)>& setter) {
+            std::sort(m_suffixes.begin(), m_suffixes.end(), [](const suffix_type& a, const suffix_type& b) {
+                return std::lexicographical_compare(std::rbegin(a), std::rend(a), std::rbegin(b), std::rend(b));
+            });
+
+            // Sentinel for an empty suffix
+            m_chars.emplace_back('\0');
+            if (bin_mode) {
+                m_term_flags.push_back(false);
+            }
+
+            const suffix_type dmmy_suffix = {{nullptr, 0}, 0};
+            const suffix_type* prev_suffix = &dmmy_suffix;
+
+            std::uint64_t prev_tpos = 0;
+
+            for (std::uint64_t i = m_suffixes.size(); i > 0; --i) {
+                const suffix_type& curr_suffix = m_suffixes[i - 1];
+                if (curr_suffix.length() == 0) {
+                    // throw TrieBuilder::Exception("A suffix is empty.");
+                }
+
+                std::uint64_t match = 0;
+                while ((match < curr_suffix.length()) && (match < prev_suffix->length()) &&
+                       ((*prev_suffix)[match] == curr_suffix[match])) {
+                    ++match;
+                }
+
+                if ((match == curr_suffix.length()) && (prev_suffix->length() != 0)) {  // sharable
+                    setter(curr_suffix.npos, prev_tpos + (prev_suffix->length() - match));
+                    prev_tpos += prev_suffix->length() - match;
+                } else {  // append
+                    setter(curr_suffix.npos, m_chars.size());
+                    prev_tpos = m_chars.size();
+                    std::copy(curr_suffix.begin(), curr_suffix.end(), std::back_inserter(m_chars));
+                    if (bin_mode) {
+                        for (std::uint64_t j = 1; j < curr_suffix.length(); ++j) {
+                            m_term_flags.push_back(false);
+                        }
+                        m_term_flags.push_back(true);
+                    } else {
+                        m_chars.emplace_back('\0');
+                    }
+                }
+
+                prev_suffix = &curr_suffix;
+            }
+        }
+
+        friend class shared_tail;
+    };
+
+  private:
+    mm_vector<char> m_chars;
+    bit_vector m_term_flags;
+
+  public:
+    shared_tail() = default;
+
+    explicit shared_tail(builder& b) {
+        m_chars.steal(b.m_chars);
+        m_term_flags.build(b.m_term_flags);
+    }
+
+    ~shared_tail() = default;
+
+    inline bool bin_mode() const {
+        return m_term_flags.size() == 0;
+    }
+
+    inline bool match(std::string_view key, size_t tpos) const {}
+
+    inline void decode(std::string& decoded, size_t tpos) const {
+        if (bin_mode()) {
+            do {
+                decoded.push_back(m_chars[tpos]);
+            } while (!m_term_flags[tpos++]);
+        } else {
+            do {
+                decoded.push_back(m_chars[tpos++]);
+            } while (m_chars[tpos]);
+        }
+    }
+};
+
+}  // namespace xcdat
\ No newline at end of file
diff --git a/include/xcdat/trie.hpp b/include/xcdat/trie.hpp
new file mode 100644
index 0000000..58adafe
--- /dev/null
+++ b/include/xcdat/trie.hpp
@@ -0,0 +1,107 @@
+#pragma once
+
+#include <optional>
+#include <string>
+
+#include "dac_bc.hpp"
+#include "trie_builder.hpp"
+
+namespace xcdat {
+
+class trie {
+  public:
+  private:
+    std::uint64_t m_num_keys = 0;
+    code_table m_table;
+    dac_bc m_bc;
+    bit_vector m_term_flags;
+    shared_tail m_tail;
+
+  public:
+    trie() = default;
+
+    virtual ~trie() = default;
+
+    static trie build(const std::vector<std::string_view>& keys, bool bin_mode = false) {
+        trie_builder b(keys, 8, bin_mode);
+        return trie(b);
+    }
+
+    inline std::uint64_t num_keys() const {
+        return m_num_keys;
+    }
+
+    inline bool bin_mode() const {
+        return m_tail.bin_mode();
+    }
+
+    inline std::uint64_t alphabet_size() const {
+        return m_table.alphabet_size();
+    }
+
+    inline std::uint64_t max_length() const {
+        return m_table.max_length();
+    }
+
+    inline std::optional<std::uint64_t> lookup(std::string_view key) const {
+        std::uint64_t kpos = 0, npos = 0;
+        while (!m_bc.is_leaf(npos)) {
+            if (kpos == key.length()) {
+                if (m_term_flags[npos]) {
+                    return npos_to_id(npos);
+                }
+                return std::nullopt;
+            }
+            const auto cpos = m_bc.base(npos) ^ m_table.get_code(key[kpos++]);
+            if (m_bc.check(cpos) != npos) {
+                return std::nullopt;
+            }
+            npos = cpos;
+        }
+        const std::uint64_t tpos = m_bc.link(npos);
+        if (!m_tail.match(key.substr(kpos, key.length() - kpos), tpos)) {
+            return std::nullopt;
+        }
+        return npos_to_id(npos);
+    }
+
+    inline std::string access(std::uint64_t id) const {
+        if (num_keys() <= id) {
+            return {};
+        }
+
+        std::string decoded;
+        decoded.reserve(max_length());
+
+        auto npos = id_to_npos(id);
+        auto tpos = m_bc.is_leaf(npos) ? m_bc.link(npos) : UINT64_MAX;
+
+        while (npos != 0) {
+            const auto ppos = m_bc.check(npos);
+            decoded.push_back(m_table.get_char(m_bc.base(ppos) ^ npos));
+            npos = ppos;
+        }
+
+        std::reverse(decoded.begin(), decoded.end());
+
+        if (tpos != 0 && tpos != UINT64_MAX) {
+            m_tail.decode(decoded, tpos);
+        }
+        return decoded;
+    }
+
+  private:
+    trie(trie_builder& b)
+        : m_num_keys(b.m_keys.size()), m_table(b.m_table), m_bc(b.m_units, b.m_leaf_flags),
+          m_term_flags(b.m_term_flags, true, true), m_tail(b.m_suffixes) {}
+
+    inline std::uint64_t npos_to_id(std::uint64_t npos) const {
+        return m_term_flags.rank(npos);
+    };
+
+    inline std::uint64_t id_to_npos(std::uint64_t id) const {
+        return m_term_flags.select(id);
+    };
+};
+
+}  // namespace xcdat
\ No newline at end of file
diff --git a/include/xcdat/trie_builder.hpp b/include/xcdat/trie_builder.hpp
new file mode 100644
index 0000000..b712b11
--- /dev/null
+++ b/include/xcdat/trie_builder.hpp
@@ -0,0 +1,246 @@
+#pragma once
+
+#include <algorithm>
+#include <string_view>
+
+#include "code_table.hpp"
+#include "dac_bc.hpp"
+#include "shared_tail.hpp"
+
+namespace xcdat {
+
+class trie_builder {
+  public:
+    struct unit_type {
+        std::uint64_t base;
+        std::uint64_t check;
+    };
+
+  private:
+    static constexpr std::uint64_t taboo_npos = 1;
+    static constexpr std::uint64_t free_blocks = 16;
+
+    const std::vector<std::string_view>& m_keys;
+    const std::uint32_t m_l1_bits;  // # of bits for L1 layer of DACs
+    const std::uint64_t m_l1_size;
+
+    bool m_bin_mode = false;
+
+    code_table m_table;
+    std::vector<unit_type> m_units;
+    bit_vector::builder m_leaf_flags;
+    bit_vector::builder m_term_flags;
+    bit_vector::builder m_used_flags;
+    std::vector<std::uint64_t> m_heads;  // for L1 blocks
+    std::vector<std::uint8_t> m_edges;
+    shared_tail::builder m_suffixes;
+
+  public:
+    trie_builder(const std::vector<std::string_view>& keys, std::uint32_t l1_bits, bool bin_mode)
+        : m_keys(keys), m_l1_bits(l1_bits), m_l1_size(1ULL << l1_bits), m_bin_mode(bin_mode) {
+        if (m_keys.empty()) {
+            // throw TrieBuilder::Exception("The input data is empty.");
+        }
+
+        // Reserve
+        {
+            std::uint64_t init_capa = 1;
+            while (init_capa < m_keys.size()) {
+                init_capa <<= 1;
+            }
+            m_units.reserve(init_capa);
+            m_leaf_flags.reserve(init_capa);
+            m_term_flags.reserve(init_capa);
+            m_used_flags.reserve(init_capa);
+            m_heads.reserve(init_capa >> m_l1_bits);
+            m_edges.reserve(256);
+        }
+
+        // Initialize an empty list.
+        for (std::uint64_t npos = 0; npos < 256; ++npos) {
+            m_units.push_back(unit_type{npos + 1, npos - 1});
+            m_leaf_flags.push_back(false);
+            m_term_flags.push_back(false);
+            m_used_flags.push_back(false);
+        }
+        m_units[255].base = 0;
+        m_units[0].check = 255;
+
+        for (std::uint64_t npos = 0; npos < 256; npos += m_l1_size) {
+            m_heads.push_back(npos);
+        }
+
+        // Fix the root
+        use_unit(0);
+        m_units[0].check = taboo_npos;
+        m_used_flags.set_bit(taboo_npos, true);
+        m_heads[taboo_npos >> m_l1_bits] = m_units[taboo_npos].base;
+
+        // Build the code table
+        m_table.build(keys);
+        m_bin_mode |= (*m_table.begin() == '\0');
+
+        // Build the BC unites
+        arrange(0, m_keys.size(), 0, 0);
+
+        m_suffixes.complete(m_bin_mode, [&](std::uint64_t npos, std::uint64_t tpos) { m_units[npos].base = tpos; });
+    }
+
+    virtual ~trie_builder() = default;
+
+  private:
+    inline void use_unit(std::uint64_t npos) {
+        m_used_flags.set_bit(npos);
+
+        const auto next = m_units[npos].base;
+        const auto prev = m_units[npos].check;
+        m_units[prev].base = next;
+        m_units[next].check = prev;
+
+        const auto lpos = npos >> m_l1_bits;
+        if (m_heads[lpos] == npos) {
+            m_heads[lpos] = (lpos != next >> m_l1_bits) ? taboo_npos : next;
+        }
+    }
+
+    inline void close_block(std::uint64_t bpos) {
+        const auto beg_npos = bpos * 256;
+        const auto end_npos = beg_npos + 256;
+
+        for (auto npos = beg_npos; npos < end_npos; ++npos) {
+            if (!m_used_flags[npos]) {
+                use_unit(npos);
+                m_used_flags.set_bit(npos, false);
+                m_units[npos].base = npos;
+                m_units[npos].check = npos;
+            }
+        }
+
+        for (auto npos = beg_npos; npos < end_npos; npos += m_l1_size) {
+            m_heads[npos >> m_l1_bits] = taboo_npos;
+        }
+    }
+
+    void expand() {
+        const auto old_size = static_cast<std::uint64_t>(m_units.size());
+        const auto new_size = old_size + 256;
+
+        for (auto npos = old_size; npos < new_size; ++npos) {
+            m_units.push_back({npos + 1, npos - 1});
+            m_leaf_flags.push_back(false);
+            m_term_flags.push_back(false);
+            m_used_flags.push_back(false);
+        }
+
+        {
+            const auto last_npos = m_units[taboo_npos].check;
+            m_units[old_size].check = last_npos;
+            m_units[last_npos].base = old_size;
+            m_units[new_size - 1].base = taboo_npos;
+            m_units[taboo_npos].check = new_size - 1;
+        }
+
+        for (auto npos = old_size; npos < new_size; npos += m_l1_size) {
+            m_heads.push_back(npos);
+        }
+
+        const auto bpos = old_size / 256;
+        if (free_blocks <= bpos) {
+            close_block(bpos - free_blocks);
+        }
+    }
+
+    void arrange(std::uint64_t beg, std::uint64_t end, std::uint64_t depth, std::uint64_t npos) {
+        if (m_keys[beg].length() == depth) {
+            m_term_flags.set_bit(npos, true);
+            if (++beg == end) {  // without link?
+                m_units[npos].base = 0;  // with an empty suffix
+                m_leaf_flags.set_bit(npos, true);
+                return;
+            }
+        } else if (beg + 1 == end) {  // leaf?
+            m_term_flags.set_bit(npos, true);
+            m_leaf_flags.set_bit(npos, true);
+            m_suffixes.set_suffix({m_keys[beg].data() + depth, m_keys[beg].length() - depth}, npos);
+            return;
+        }
+
+        // fetching edges
+        {
+            m_edges.clear();
+            auto ch = static_cast<std::uint8_t>(m_keys[beg][depth]);
+            for (auto i = beg + 1; i < end; ++i) {
+                const auto next_ch = static_cast<std::uint8_t>(m_keys[i][depth]);
+                if (ch != next_ch) {
+                    if (next_ch < ch) {
+                        // throw TrieBuilder::Exception("The input data is not in lexicographical order.");
+                    }
+                    m_edges.push_back(ch);
+                    ch = next_ch;
+                }
+            }
+            m_edges.push_back(ch);
+        }
+
+        const auto base = xcheck(npos >> m_l1_bits);
+        if (m_units.size() <= base) {
+            expand();
+        }
+
+        // defining new edges
+        m_units[npos].base = base;
+        for (const auto ch : m_edges) {
+            const auto child_id = base ^ m_table.get_code(ch);
+            use_unit(child_id);
+            m_units[child_id].check = npos;
+        }
+
+        // following the children
+        auto i = beg;
+        auto ch = static_cast<uint8_t>(m_keys[beg][depth]);
+        for (auto j = beg + 1; j < end; ++j) {
+            const auto next_ch = static_cast<uint8_t>(m_keys[j][depth]);
+            if (ch != next_ch) {
+                arrange(i, j, depth + 1, base ^ m_table.get_code(ch));
+                ch = next_ch;
+                i = j;
+            }
+        }
+        arrange(i, end, depth + 1, base ^ m_table.get_code(ch));
+    }
+
+    inline std::uint64_t xcheck(std::uint64_t lpos) const {
+        if (m_units[taboo_npos].base == taboo_npos) {  // Full?
+            return m_units.size() ^ m_table.get_code(m_edges[0]);
+        }
+
+        // search in the same L1 block
+        for (auto i = m_heads[lpos]; i != taboo_npos && i >> m_l1_bits == lpos; i = m_units[i].base) {
+            const auto base = i ^ m_table.get_code(m_edges[0]);
+            if (is_target(base)) {
+                return base;  // base / block_size_ == lpos
+            }
+        }
+
+        for (auto i = m_units[taboo_npos].base; i != taboo_npos; i = m_units[i].base) {
+            const auto base = i ^ m_table.get_code(m_edges[0]);
+            if (is_target(base)) {
+                return base;  // base / block_size_ != lpos
+            }
+        }
+        return m_units.size() ^ m_table.get_code(m_edges[0]);
+    }
+
+    inline bool is_target(std::uint64_t base) const {
+        for (const auto ch : m_edges) {
+            if (m_used_flags[base ^ m_table.get_code(ch)]) {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    friend class trie;
+};
+
+}  // namespace xcdat
\ No newline at end of file
diff --git a/include/xcdat/utils.hpp b/include/xcdat/utils.hpp
index d3e4d5e..8febf65 100644
--- a/include/xcdat/utils.hpp
+++ b/include/xcdat/utils.hpp
@@ -8,7 +8,7 @@ namespace xcdat::utils {
 
 template <std::uint64_t N>
 constexpr std::tuple<std::uint64_t, std::uint64_t> decompose(std::uint64_t x) {
-    return std::make_tuple(x / N, x % N);
+    return {x / N, x % N};
 }
 
 template <class T = std::uint64_t>
diff --git a/test/test_bc.cpp b/test/test_bc.cpp
new file mode 100644
index 0000000..522ce16
--- /dev/null
+++ b/test/test_bc.cpp
@@ -0,0 +1,72 @@
+#define DOCTEST_CONFIG_IMPLEMENT_WITH_MAIN
+
+#include <algorithm>
+#include <random>
+
+#include <xcdat/dac_bc.hpp>
+
+#include "doctest/doctest.h"
+
+struct bc_unit {
+    std::uint64_t base;
+    std::uint64_t check;
+};
+
+std::vector<bc_unit> make_random_units(std::uint64_t n) {
+    static constexpr std::uint64_t seed = 13;
+
+    std::mt19937_64 engine(seed);
+    std::uniform_int_distribution<std::uint64_t> dist(0, n - 1);
+
+    std::vector<bc_unit> bc_units(n);
+    for (std::uint64_t i = 0; i < n; i++) {
+        bc_units[i].base = dist(engine);
+        bc_units[i].check = dist(engine);
+    }
+    return bc_units;
+}
+
+std::vector<bool> make_random_bits(std::uint64_t n, double dens) {
+    static constexpr std::uint64_t seed = 17;
+
+    std::mt19937_64 engine(seed);
+    std::uniform_real_distribution<double> dist(0.0, 1.0);
+
+    std::vector<bool> bits(n);
+    for (std::uint64_t i = 0; i < n; i++) {
+        bits[i] = dist(engine) < dens;
+    }
+    return bits;
+}
+
+xcdat::bit_vector::builder to_bit_vector_builder(const std::vector<bool>& bits) {
+    xcdat::bit_vector::builder bvb(bits.size());
+    for (std::uint64_t i = 0; i < bits.size(); i++) {
+        bvb.set_bit(i, bits[i]);
+    }
+    return bvb;
+}
+
+std::uint64_t get_num_ones(const std::vector<bool>& bits) {
+    return std::accumulate(bits.begin(), bits.end(), 0ULL);
+}
+
+TEST_CASE("Test xcdat::dac_bc") {
+    auto bc_units = make_random_units(10000);
+    auto leaf_flags = make_random_bits(10000, 0.2);
+
+    xcdat::dac_bc bc(bc_units, to_bit_vector_builder(leaf_flags));
+
+    REQUIRE_EQ(bc.num_units(), bc_units.size());
+    REQUIRE_EQ(bc.num_leaves(), get_num_ones(leaf_flags));
+
+    for (std::uint64_t i = 0; i < bc.num_units(); i++) {
+        REQUIRE_EQ(bc.is_leaf(i), leaf_flags[i]);
+        if (leaf_flags[i]) {
+            REQUIRE_EQ(bc.link(i), bc_units[i].base);
+        } else {
+            REQUIRE_EQ(bc.base(i), bc_units[i].base);
+        }
+        REQUIRE_EQ(bc.check(i), bc_units[i].check);
+    }
+}
\ No newline at end of file
diff --git a/test/test_bit_vector.cpp b/test/test_bit_vector.cpp
index 4ff425b..8046855 100644
--- a/test/test_bit_vector.cpp
+++ b/test/test_bit_vector.cpp
@@ -7,10 +7,12 @@
 
 #include "doctest/doctest.h"
 
-std::vector<bool> generate_random_bits(std::uint64_t n) {
+std::vector<bool> make_random_bits(std::uint64_t n) {
     static constexpr std::uint64_t seed = 13;
+
     std::vector<bool> bits;
-    std::mt19937 engine(seed);
+    std::mt19937_64 engine(seed);
+
     for (std::uint64_t i = 0; i < n; i++) {
         bits.push_back(engine() & 1);
     }
@@ -31,16 +33,15 @@ std::uint64_t select_naive(const std::vector<bool>& bits, std::uint64_t n) {
         if (bits[i]) {
             if (n == 0) {
                 break;
-            } else {
-                n -= 1;
             }
+            n -= 1;
         }
     }
     return i;
 }
 
 TEST_CASE("Test bit_vector::builder with resize") {
-    const auto bits = generate_random_bits(10000);
+    const auto bits = make_random_bits(10000);
 
     xcdat::bit_vector::builder b;
     b.resize(bits.size());
@@ -56,7 +57,7 @@ TEST_CASE("Test bit_vector::builder with resize") {
 }
 
 TEST_CASE("Test bit_vector::builder with push_back") {
-    const auto bits = generate_random_bits(10000);
+    const auto bits = make_random_bits(10000);
 
     xcdat::bit_vector::builder b;
     b.reserve(bits.size());
@@ -73,7 +74,7 @@ TEST_CASE("Test bit_vector::builder with push_back") {
 }
 
 TEST_CASE("Test bit_vector") {
-    const auto bits = generate_random_bits(10000);
+    const auto bits = make_random_bits(10000);
 
     xcdat::bit_vector bv;
     {
@@ -93,6 +94,7 @@ TEST_CASE("Test bit_vector") {
 
     static constexpr std::uint64_t seed = 17;
     std::mt19937_64 engine(seed);
+
     {
         std::uniform_int_distribution<std::uint64_t> dist(0, bv.size());
         for (std::uint64_t r = 0; r < 100; r++) {
diff --git a/test/test_compact_vector.cpp b/test/test_compact_vector.cpp
new file mode 100644
index 0000000..55f0daa
--- /dev/null
+++ b/test/test_compact_vector.cpp
@@ -0,0 +1,43 @@
+#define DOCTEST_CONFIG_IMPLEMENT_WITH_MAIN
+
+#include <algorithm>
+#include <random>
+
+#include <xcdat/compact_vector.hpp>
+
+#include "doctest/doctest.h"
+
+std::vector<std::uint64_t> make_random_ints(std::uint64_t n) {
+    static constexpr std::uint64_t seed = 13;
+
+    std::mt19937_64 engine(seed);
+    std::uniform_int_distribution<std::uint64_t> dist(0, UINT16_MAX);
+
+    std::vector<std::uint64_t> ints(n);
+    for (std::uint64_t i = 0; i < n; i++) {
+        ints[i] = dist(engine);
+    }
+    return ints;
+}
+
+TEST_CASE("Test xcdat::compact_vector (tiny)") {
+    std::vector<std::uint64_t> ints = {2, 0, 14, 456, 32, 5544, 23};
+    xcdat::compact_vector cv(ints);
+
+    REQUIRE_EQ(cv.size(), ints.size());
+
+    for (std::uint64_t i = 0; i < ints.size(); i++) {
+        REQUIRE_EQ(cv[i], ints[i]);
+    }
+}
+
+TEST_CASE("Test xcdat::compact_vector (random)") {
+    std::vector<std::uint64_t> ints = make_random_ints(10000);
+    xcdat::compact_vector cv(ints);
+
+    REQUIRE_EQ(cv.size(), ints.size());
+
+    for (std::uint64_t i = 0; i < ints.size(); i++) {
+        REQUIRE_EQ(cv[i], ints[i]);
+    }
+}