Update external c++ unordered maps for benchmarking.

4 files changed, 578 insertions, 685 deletions

diff --git a/benchmarks/external/ankerl/unordered_dense.h b/benchmarks/external/ankerl/unordered_dense.h
index 304ef3bc..220c65b8 100644
--- a/benchmarks/external/ankerl/unordered_dense.h
+++ b/benchmarks/external/ankerl/unordered_dense.h
@@ -1,7 +1,7 @@
 ///////////////////////// ankerl::unordered_dense::{map, set} /////////////////////////
 
 // A fast & densely stored hashmap and hashset based on robin-hood backward shift deletion.
-// Version 1.0.2
+// Version 1.3.1
 // https://github.com/martinus/unordered_dense
 //
 // Licensed under the MIT License <http://opensource.org/licenses/MIT>.
@@ -30,35 +30,56 @@
 #define ANKERL_UNORDERED_DENSE_H
 
 // see https://semver.org/spec/v2.0.0.html
-#define ANKERL_UNORDERED_DENSE_VERSION_MAJOR 1 // incompatible API changes
-#define ANKERL_UNORDERED_DENSE_VERSION_MINOR 0 // add functionality in a backwards compatible manner
-#define ANKERL_UNORDERED_DENSE_VERSION_PATCH 2 // backwards compatible bug fixes
+#define ANKERL_UNORDERED_DENSE_VERSION_MAJOR 1 // NOLINT(cppcoreguidelines-macro-usage) incompatible API changes
+#define ANKERL_UNORDERED_DENSE_VERSION_MINOR 3 // NOLINT(cppcoreguidelines-macro-usage) backwards compatible functionality
+#define ANKERL_UNORDERED_DENSE_VERSION_PATCH 1 // NOLINT(cppcoreguidelines-macro-usage) backwards compatible bug fixes
+
+// API versioning with inline namespace, see https://www.foonathan.net/2018/11/inline-namespaces/
+#define ANKERL_UNORDERED_DENSE_VERSION_CONCAT1(major, minor, patch) v##major##_##minor##_##patch
+#define ANKERL_UNORDERED_DENSE_VERSION_CONCAT(major, minor, patch) ANKERL_UNORDERED_DENSE_VERSION_CONCAT1(major, minor, patch)
+#define ANKERL_UNORDERED_DENSE_NAMESPACE   \
+    ANKERL_UNORDERED_DENSE_VERSION_CONCAT( \
+        ANKERL_UNORDERED_DENSE_VERSION_MAJOR, ANKERL_UNORDERED_DENSE_VERSION_MINOR, ANKERL_UNORDERED_DENSE_VERSION_PATCH)
+
+#if defined(_MSVC_LANG)
+#    define ANKERL_UNORDERED_DENSE_CPP_VERSION _MSVC_LANG
+#else
+#    define ANKERL_UNORDERED_DENSE_CPP_VERSION __cplusplus
+#endif
+
+#if defined(__GNUC__)
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#    define ANKERL_UNORDERED_DENSE_PACK(decl) decl __attribute__((__packed__))
+#elif defined(_MSC_VER)
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#    define ANKERL_UNORDERED_DENSE_PACK(decl) __pragma(pack(push, 1)) decl __pragma(pack(pop))
+#endif
 
-#if 0 // __cplusplus < 201703L
+#if ANKERL_UNORDERED_DENSE_CPP_VERSION < 201703L
 #    error ankerl::unordered_dense requires C++17 or higher
 #else
-
-#    include <algorithm>
-#    include <array>
-#    include <cstdint>
-#    include <cstring>
-#    include <functional>
-#    include <initializer_list>
-#    include <limits>
-#    include <memory>
-#    include <stdexcept>
-#    include <string>
-#    include <string_view>
-#    include <type_traits>
-#    include <utility>
-#    include <vector>
-
-#    define ANKERL_UNORDERED_DENSE_PMR 0
+#    include <array>            // for array
+#    include <cstdint>          // for uint64_t, uint32_t, uint8_t, UINT64_C
+#    include <cstring>          // for size_t, memcpy, memset
+#    include <functional>       // for equal_to, hash
+#    include <initializer_list> // for initializer_list
+#    include <iterator>         // for pair, distance
+#    include <limits>           // for numeric_limits
+#    include <memory>           // for allocator, allocator_traits, shared_ptr
+#    include <stdexcept>        // for out_of_range
+#    include <string>           // for basic_string
+#    include <string_view>      // for basic_string_view, hash
+#    include <tuple>            // for forward_as_tuple
+#    include <type_traits>      // for enable_if_t, declval, conditional_t, ena...
+#    include <utility>          // for forward, exchange, pair, as_const, piece...
+#    include <vector>           // for vector
+
+#    define ANKERL_UNORDERED_DENSE_PMR 0 // NOLINT(cppcoreguidelines-macro-usage)
 #    if defined(__has_include)
 #        if __has_include(<memory_resource>)
 #            undef ANKERL_UNORDERED_DENSE_PMR
-#            define ANKERL_UNORDERED_DENSE_PMR 1
-#            include <memory_resource>
+#            define ANKERL_UNORDERED_DENSE_PMR 1 // NOLINT(cppcoreguidelines-macro-usage)
+#            include <memory_resource>           // for polymorphic_allocator
 #        endif
 #    endif
 
@@ -68,14 +89,15 @@
 #    endif
 
 #    if defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__clang__)
-#        define ANKERL_UNORDERED_DENSE_LIKELY(x) __builtin_expect(x, 1)
-#        define ANKERL_UNORDERED_DENSE_UNLIKELY(x) __builtin_expect(x, 0)
+#        define ANKERL_UNORDERED_DENSE_LIKELY(x) __builtin_expect(x, 1)   // NOLINT(cppcoreguidelines-macro-usage)
+#        define ANKERL_UNORDERED_DENSE_UNLIKELY(x) __builtin_expect(x, 0) // NOLINT(cppcoreguidelines-macro-usage)
 #    else
-#        define ANKERL_UNORDERED_DENSE_LIKELY(x) (x)
-#        define ANKERL_UNORDERED_DENSE_UNLIKELY(x) (x)
+#        define ANKERL_UNORDERED_DENSE_LIKELY(x) (x)   // NOLINT(cppcoreguidelines-macro-usage)
+#        define ANKERL_UNORDERED_DENSE_UNLIKELY(x) (x) // NOLINT(cppcoreguidelines-macro-usage)
 #    endif
 
 namespace ankerl::unordered_dense {
+inline namespace ANKERL_UNORDERED_DENSE_NAMESPACE {
 
 // hash ///////////////////////////////////////////////////////////////////////
 
@@ -122,7 +144,7 @@ static inline void mum(uint64_t* a, uint64_t* b) {
 // read functions. WARNING: we don't care about endianness, so results are different on big endian!
 [[nodiscard]] static inline auto r8(const uint8_t* p) -> uint64_t {
     uint64_t v{};
-    std::memcpy(&v, p, 8);
+    std::memcpy(&v, p, 8U);
     return v;
 }
 
@@ -137,7 +159,7 @@ static inline void mum(uint64_t* a, uint64_t* b) {
     return (static_cast<uint64_t>(p[0]) << 16U) | (static_cast<uint64_t>(p[k >> 1U]) << 8U) | p[k - 1];
 }
 
-[[nodiscard]] static inline auto hash(void const* key, size_t len) -> uint64_t {
+[[maybe_unused]] [[nodiscard]] static inline auto hash(void const* key, size_t len) -> uint64_t {
     static constexpr auto secret = std::array{UINT64_C(0xa0761d6478bd642f),
                                               UINT64_C(0xe7037ed1a0b428db),
                                               UINT64_C(0x8ebc6af09c88c6e3),
@@ -191,70 +213,74 @@ static inline void mum(uint64_t* a, uint64_t* b) {
 } // namespace detail::wyhash
 
 template <typename T, typename Enable = void>
-struct hash : public std::hash<T> {
+struct hash {
     using is_avalanching = void;
     auto operator()(T const& obj) const noexcept(noexcept(std::declval<std::hash<T>>().operator()(std::declval<T const&>())))
-        -> size_t {
-        return static_cast<size_t>(detail::wyhash::hash(std::hash<T>::operator()(obj)));
+        -> uint64_t {
+        return detail::wyhash::hash(std::hash<T>{}(obj));
     }
 };
 
 template <typename CharT>
 struct hash<std::basic_string<CharT>> {
     using is_avalanching = void;
-    auto operator()(std::basic_string<CharT> const& str) const noexcept -> size_t {
-        return static_cast<size_t>(detail::wyhash::hash(str.data(), sizeof(CharT) * str.size()));
+    auto operator()(std::basic_string<CharT> const& str) const noexcept -> uint64_t {
+        return detail::wyhash::hash(str.data(), sizeof(CharT) * str.size());
     }
 };
 
 template <typename CharT>
 struct hash<std::basic_string_view<CharT>> {
     using is_avalanching = void;
-    auto operator()(std::basic_string_view<CharT> const& sv) const noexcept -> size_t {
-        return static_cast<size_t>(detail::wyhash::hash(sv.data(), sizeof(CharT) * sv.size()));
+    auto operator()(std::basic_string_view<CharT> const& sv) const noexcept -> uint64_t {
+        return detail::wyhash::hash(sv.data(), sizeof(CharT) * sv.size());
     }
 };
 
 template <class T>
 struct hash<T*> {
     using is_avalanching = void;
-    auto operator()(T* ptr) const noexcept -> size_t {
-        return static_cast<size_t>(detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr)));
+    auto operator()(T* ptr) const noexcept -> uint64_t {
+        // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
+        return detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr));
     }
 };
 
 template <class T>
 struct hash<std::unique_ptr<T>> {
     using is_avalanching = void;
-    auto operator()(std::unique_ptr<T> const& ptr) const noexcept -> size_t {
-        return static_cast<size_t>(detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr.get())));
+    auto operator()(std::unique_ptr<T> const& ptr) const noexcept -> uint64_t {
+        // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
+        return detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr.get()));
     }
 };
 
 template <class T>
 struct hash<std::shared_ptr<T>> {
     using is_avalanching = void;
-    auto operator()(std::shared_ptr<T> const& ptr) const noexcept -> size_t {
-        return static_cast<size_t>(detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr.get())));
+    auto operator()(std::shared_ptr<T> const& ptr) const noexcept -> uint64_t {
+        // NOLINTNEXTLINE(cppcoreguidelines-pro-type-reinterpret-cast)
+        return detail::wyhash::hash(reinterpret_cast<uintptr_t>(ptr.get()));
     }
 };
 
 template <typename Enum>
 struct hash<Enum, typename std::enable_if<std::is_enum<Enum>::value>::type> {
     using is_avalanching = void;
-    auto operator()(Enum e) const noexcept -> size_t {
-        using Underlying = typename std::underlying_type_t<Enum>;
-        return static_cast<size_t>(detail::wyhash::hash(static_cast<Underlying>(e)));
+    auto operator()(Enum e) const noexcept -> uint64_t {
+        using underlying = typename std::underlying_type_t<Enum>;
+        return detail::wyhash::hash(static_cast<underlying>(e));
     }
 };
 
-#    define ANKERL_UNORDERED_DENSE_HASH_STATICCAST(T)                                         \
-        template <>                                                                           \
-        struct hash<T> {                                                                      \
-            using is_avalanching = void;                                                      \
-            auto operator()(T const& obj) const noexcept -> size_t {                          \
-                return static_cast<size_t>(detail::wyhash::hash(static_cast<uint64_t>(obj))); \
-            }                                                                                 \
+// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
+#    define ANKERL_UNORDERED_DENSE_HASH_STATICCAST(T)                    \
+        template <>                                                      \
+        struct hash<T> {                                                 \
+            using is_avalanching = void;                                 \
+            auto operator()(T const& obj) const noexcept -> uint64_t {   \
+                return detail::wyhash::hash(static_cast<uint64_t>(obj)); \
+            }                                                            \
         }
 
 #    if defined(__GNUC__) && !defined(__clang__)
@@ -266,7 +292,7 @@ ANKERL_UNORDERED_DENSE_HASH_STATICCAST(bool);
 ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char);
 ANKERL_UNORDERED_DENSE_HASH_STATICCAST(signed char);
 ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned char);
-#    if __cplusplus >= 202002L
+#    if ANKERL_UNORDERED_DENSE_CPP_VERSION >= 202002L
 ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char8_t);
 #    endif
 ANKERL_UNORDERED_DENSE_HASH_STATICCAST(char16_t);
@@ -285,6 +311,28 @@ ANKERL_UNORDERED_DENSE_HASH_STATICCAST(unsigned long long);
 #        pragma GCC diagnostic pop
 #    endif
 
+// bucket_type //////////////////////////////////////////////////////////
+
+namespace bucket_type {
+
+struct standard {
+    static constexpr uint32_t dist_inc = 1U << 8U;             // skip 1 byte fingerprint
+    static constexpr uint32_t fingerprint_mask = dist_inc - 1; // mask for 1 byte of fingerprint
+
+    uint32_t m_dist_and_fingerprint; // upper 3 byte: distance to original bucket. lower byte: fingerprint from hash
+    uint32_t m_value_idx;            // index into the m_values vector.
+};
+
+ANKERL_UNORDERED_DENSE_PACK(struct big {
+    static constexpr uint32_t dist_inc = 1U << 8U;             // skip 1 byte fingerprint
+    static constexpr uint32_t fingerprint_mask = dist_inc - 1; // mask for 1 byte of fingerprint
+
+    uint32_t m_dist_and_fingerprint; // upper 3 byte: distance to original bucket. lower byte: fingerprint from hash
+    size_t m_value_idx;              // index into the m_values vector.
+});
+
+} // namespace bucket_type
+
 namespace detail {
 
 struct nonesuch {};
@@ -313,6 +361,9 @@ using detect_avalanching = typename T::is_avalanching;
 template <typename T>
 using detect_is_transparent = typename T::is_transparent;
 
+template <typename T>
+using detect_iterator = typename T::iterator;
+
 template <typename H, typename KE>
 using is_transparent =
     std::enable_if_t<is_detected_v<detect_is_transparent, H> && is_detected_v<detect_is_transparent, KE>, bool>;
@@ -322,79 +373,101 @@ template <class Key,
           class T, // when void, treat it as a set.
           class Hash,
           class KeyEqual,
-          class Allocator>
+          class AllocatorOrContainer,
+          class Bucket>
 class table {
-    struct Bucket;
-    using ValueContainer =
-        typename std::vector<typename std::conditional_t<std::is_void_v<T>, Key, std::pair<Key, T>>, Allocator>;
-    using BucketAlloc = typename std::allocator_traits<Allocator>::template rebind_alloc<Bucket>;
-    using BucketAllocTraits = std::allocator_traits<BucketAlloc>;
+public:
+    using value_container_type = std::conditional_t<
+        is_detected_v<detect_iterator, AllocatorOrContainer>,
+        AllocatorOrContainer,
+        typename std::vector<typename std::conditional_t<std::is_void_v<T>, Key, std::pair<Key, T>>, AllocatorOrContainer>>;
 
-    static constexpr uint32_t BUCKET_DIST_INC = 1U << 8U;                    // skip 1 byte fingerprint
-    static constexpr uint32_t BUCKET_FINGERPRINT_MASK = BUCKET_DIST_INC - 1; // mask for 1 byte of fingerprint
-    static constexpr uint8_t INITIAL_SHIFTS = 64 - 3;                        // 2^(64-m_shift) number of buckets
-    static constexpr float DEFAULT_MAX_LOAD_FACTOR = 0.8F;
+private:
+    using bucket_alloc =
+        typename std::allocator_traits<typename value_container_type::allocator_type>::template rebind_alloc<Bucket>;
+    using bucket_alloc_traits = std::allocator_traits<bucket_alloc>;
+
+    static constexpr uint8_t initial_shifts = 64 - 3; // 2^(64-m_shift) number of buckets
+    static constexpr float default_max_load_factor = 0.8F;
 
 public:
     using key_type = Key;
     using mapped_type = T;
-    using value_type = typename ValueContainer::value_type;
-    using size_type = typename ValueContainer::size_type;
-    using difference_type = typename ValueContainer::difference_type;
+    using value_type = typename value_container_type::value_type;
+    using size_type = typename value_container_type::size_type;
+    using difference_type = typename value_container_type::difference_type;
     using hasher = Hash;
     using key_equal = KeyEqual;
-    using allocator_type = typename ValueContainer::allocator_type;
-    using reference = typename ValueContainer::reference;
-    using const_reference = typename ValueContainer::const_reference;
-    using pointer = typename ValueContainer::pointer;
-    using const_pointer = typename ValueContainer::const_pointer;
-    using iterator = typename ValueContainer::iterator;
-    using const_iterator = typename ValueContainer::const_iterator;
+    using allocator_type = typename value_container_type::allocator_type;
+    using reference = typename value_container_type::reference;
+    using const_reference = typename value_container_type::const_reference;
+    using pointer = typename value_container_type::pointer;
+    using const_pointer = typename value_container_type::const_pointer;
+    using iterator = typename value_container_type::iterator;
+    using const_iterator = typename value_container_type::const_iterator;
+    using bucket_type = Bucket;
 
 private:
-    struct Bucket {
-        uint32_t dist_and_fingerprint; // upper 3 byte: distance to original bucket. lower byte: fingerprint from hash
-        uint32_t value_idx;            // index into the m_values vector.
-    };
+    using value_idx_type = decltype(Bucket::m_value_idx);
+    using dist_and_fingerprint_type = decltype(Bucket::m_dist_and_fingerprint);
+
     static_assert(std::is_trivially_destructible_v<Bucket>, "assert there's no need to call destructor / std::destroy");
     static_assert(std::is_trivially_copyable_v<Bucket>, "assert we can just memset / memcpy");
 
-    ValueContainer m_values{}; // Contains all the key-value pairs in one densely stored container. No holes.
-    Bucket* m_buckets_start = nullptr;
-    Bucket* m_buckets_end = nullptr;
-    uint32_t m_max_bucket_capacity = 0;
-    float m_max_load_factor = DEFAULT_MAX_LOAD_FACTOR;
+    value_container_type m_values{}; // Contains all the key-value pairs in one densely stored container. No holes.
+    typename std::allocator_traits<bucket_alloc>::pointer m_buckets{};
+    size_t m_num_buckets = 0;
+    size_t m_max_bucket_capacity = 0;
+    float m_max_load_factor = default_max_load_factor;
     Hash m_hash{};
     KeyEqual m_equal{};
-    uint8_t m_shifts = INITIAL_SHIFTS;
+    uint8_t m_shifts = initial_shifts;
 
-    [[nodiscard]] auto next(Bucket const* bucket) const -> Bucket const* {
-        return ANKERL_UNORDERED_DENSE_UNLIKELY(bucket + 1 == m_buckets_end) ? m_buckets_start : bucket + 1;
+    [[nodiscard]] auto next(value_idx_type bucket_idx) const -> value_idx_type {
+        return ANKERL_UNORDERED_DENSE_UNLIKELY(bucket_idx + 1U == m_num_buckets)
+                   ? 0
+                   : static_cast<value_idx_type>(bucket_idx + 1U);
     }
 
-    [[nodiscard]] auto next(Bucket* bucket) -> Bucket* {
-        return ANKERL_UNORDERED_DENSE_UNLIKELY(bucket + 1 == m_buckets_end) ? m_buckets_start : bucket + 1;
+    // Helper to access bucket through pointer types
+    [[nodiscard]] static constexpr auto at(typename std::allocator_traits<bucket_alloc>::pointer bucket_ptr, size_t offset)
+        -> Bucket& {
+        return *(bucket_ptr + static_cast<typename std::allocator_traits<bucket_alloc>::difference_type>(offset));
     }
 
+    // use the dist_inc and dist_dec functions so that uint16_t types work without warning
+    [[nodiscard]] static constexpr auto dist_inc(dist_and_fingerprint_type x) -> dist_and_fingerprint_type {
+        return static_cast<dist_and_fingerprint_type>(x + Bucket::dist_inc);
+    }
+
+    [[nodiscard]] static constexpr auto dist_dec(dist_and_fingerprint_type x) -> dist_and_fingerprint_type {
+        return static_cast<dist_and_fingerprint_type>(x - Bucket::dist_inc);
+    }
+
+    // The goal of mixed_hash is to always produce a high quality 64bit hash.
     template <typename K>
     [[nodiscard]] constexpr auto mixed_hash(K const& key) const -> uint64_t {
         if constexpr (is_detected_v<detect_avalanching, Hash>) {
-            return m_hash(key);
+            // we know that the hash is good because is_avalanching.
+            if constexpr (sizeof(decltype(m_hash(key))) < sizeof(uint64_t)) {
+                // 32bit hash and is_avalanching => multiply with a constant to avalanche bits upwards
+                return m_hash(key) * UINT64_C(0x9ddfea08eb382d69);
+            } else {
+                // 64bit and is_avalanching => only use the hash itself.
+                return m_hash(key);
+            }
         } else {
+            // not is_avalanching => apply wyhash
             return wyhash::hash(m_hash(key));
         }
     }
 
-    [[nodiscard]] constexpr auto dist_and_fingerprint_from_hash(uint64_t hash) const -> uint32_t {
-        return BUCKET_DIST_INC | (hash & BUCKET_FINGERPRINT_MASK);
+    [[nodiscard]] constexpr auto dist_and_fingerprint_from_hash(uint64_t hash) const -> dist_and_fingerprint_type {
+        return Bucket::dist_inc | (static_cast<dist_and_fingerprint_type>(hash) & Bucket::fingerprint_mask);
     }
 
-    [[nodiscard]] constexpr auto bucket_from_hash(uint64_t hash) const -> Bucket const* {
-        return m_buckets_start + (hash >> m_shifts);
-    }
-
-    [[nodiscard]] constexpr auto bucket_from_hash(uint64_t hash) -> Bucket* {
-        return m_buckets_start + (hash >> m_shifts);
+    [[nodiscard]] constexpr auto bucket_idx_from_hash(uint64_t hash) const -> value_idx_type {
+        return static_cast<value_idx_type>(hash >> m_shifts);
     }
 
     [[nodiscard]] static constexpr auto get_key(value_type const& vt) -> key_type const& {
@@ -406,51 +479,45 @@ private:
     }
 
     template <typename K>
-    [[nodiscard]] auto next_while_less(K const& key) -> std::pair<uint32_t, Bucket*> {
-        auto const& pair = std::as_const(*this).next_while_less(key);
-        return {pair.first, const_cast<Bucket*>(pair.second)}; // NOLINT(cppcoreguidelines-pro-type-const-cast)
-    }
-
-    template <typename K>
-    [[nodiscard]] auto next_while_less(K const& key) const -> std::pair<uint32_t, Bucket const*> {
+    [[nodiscard]] auto next_while_less(K const& key) const -> Bucket {
         auto hash = mixed_hash(key);
         auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
-        auto const* bucket = bucket_from_hash(hash);
+        auto bucket_idx = bucket_idx_from_hash(hash);
 
-        while (dist_and_fingerprint < bucket->dist_and_fingerprint) {
-            dist_and_fingerprint += BUCKET_DIST_INC;
-            bucket = next(bucket);
+        while (dist_and_fingerprint < at(m_buckets, bucket_idx).m_dist_and_fingerprint) {
+            dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+            bucket_idx = next(bucket_idx);
         }
-        return {dist_and_fingerprint, bucket};
+        return {dist_and_fingerprint, bucket_idx};
     }
 
-    void place_and_shift_up(Bucket bucket, Bucket* place) {
-        while (0 != place->dist_and_fingerprint) {
-            bucket = std::exchange(*place, bucket);
-            bucket.dist_and_fingerprint += BUCKET_DIST_INC;
+    void place_and_shift_up(Bucket bucket, value_idx_type place) {
+        while (0 != at(m_buckets, place).m_dist_and_fingerprint) {
+            bucket = std::exchange(at(m_buckets, place), bucket);
+            bucket.m_dist_and_fingerprint = dist_inc(bucket.m_dist_and_fingerprint);
             place = next(place);
         }
-        *place = bucket;
+        at(m_buckets, place) = bucket;
     }
 
-    [[nodiscard]] static constexpr auto calc_num_buckets(uint8_t shifts) -> uint64_t {
-        return UINT64_C(1) << (64U - shifts);
+    [[nodiscard]] static constexpr auto calc_num_buckets(uint8_t shifts) -> size_t {
+        return std::min(max_bucket_count(), size_t{1} << (64U - shifts));
     }
 
     [[nodiscard]] constexpr auto calc_shifts_for_size(size_t s) const -> uint8_t {
-        auto shifts = INITIAL_SHIFTS;
-        while (shifts > 0 && static_cast<uint64_t>(calc_num_buckets(shifts) * max_load_factor()) < s) {
+        auto shifts = initial_shifts;
+        while (shifts > 0 && static_cast<size_t>(static_cast<float>(calc_num_buckets(shifts)) * max_load_factor()) < s) {
             --shifts;
         }
         return shifts;
     }
 
-    // assumes m_values has data, m_buckets_start=m_buckets_end=nullptr, m_shifts is INITIAL_SHIFTS
+    // assumes m_values has data, m_buckets=m_buckets_end=nullptr, m_shifts is INITIAL_SHIFTS
     void copy_buckets(table const& other) {
         if (!empty()) {
             m_shifts = other.m_shifts;
             allocate_buckets_from_shift();
-            std::memcpy(m_buckets_start, other.m_buckets_start, sizeof(Bucket) * bucket_count());
+            std::memcpy(m_buckets, other.m_buckets, sizeof(Bucket) * bucket_count());
         }
     }
 
@@ -462,30 +529,37 @@ private:
     }
 
     void deallocate_buckets() {
-        auto bucket_alloc = BucketAlloc(m_values.get_allocator());
-        BucketAllocTraits::deallocate(bucket_alloc, m_buckets_start, bucket_count());
-        m_buckets_start = nullptr;
-        m_buckets_end = nullptr;
+        auto ba = bucket_alloc(m_values.get_allocator());
+        if (nullptr != m_buckets) {
+            bucket_alloc_traits::deallocate(ba, m_buckets, bucket_count());
+        }
+        m_buckets = nullptr;
+        m_num_buckets = 0;
         m_max_bucket_capacity = 0;
     }
 
     void allocate_buckets_from_shift() {
-        auto bucket_alloc = BucketAlloc(m_values.get_allocator());
-        auto num_buckets = calc_num_buckets(m_shifts);
-        m_buckets_start = BucketAllocTraits::allocate(bucket_alloc, num_buckets);
-        m_buckets_end = m_buckets_start + num_buckets;
-        m_max_bucket_capacity = static_cast<uint64_t>(num_buckets * max_load_factor());
+        auto ba = bucket_alloc(m_values.get_allocator());
+        m_num_buckets = calc_num_buckets(m_shifts);
+        m_buckets = bucket_alloc_traits::allocate(ba, m_num_buckets);
+        if (m_num_buckets == max_bucket_count()) {
+            // reached the maximum, make sure we can use each bucket
+            m_max_bucket_capacity = max_bucket_count();
+        } else {
+            m_max_bucket_capacity = static_cast<value_idx_type>(static_cast<float>(m_num_buckets) * max_load_factor());
+        }
     }
 
     void clear_buckets() {
-        if (m_buckets_start != nullptr) {
-            std::memset(m_buckets_start, 0, sizeof(Bucket) * bucket_count());
+        if (m_buckets != nullptr) {
+            std::memset(&*m_buckets, 0, sizeof(Bucket) * bucket_count());
         }
     }
 
     void clear_and_fill_buckets_from_values() {
         clear_buckets();
-        for (uint32_t value_idx = 0, end_idx = static_cast<uint32_t>(m_values.size()); value_idx < end_idx; ++value_idx) {
+        for (value_idx_type value_idx = 0, end_idx = static_cast<value_idx_type>(m_values.size()); value_idx < end_idx;
+             ++value_idx) {
             auto const& key = get_key(m_values[value_idx]);
             auto [dist_and_fingerprint, bucket] = next_while_less(key);
 
@@ -495,22 +569,26 @@ private:
     }
 
     void increase_size() {
+        if (ANKERL_UNORDERED_DENSE_UNLIKELY(m_max_bucket_capacity == max_bucket_count())) {
+            throw std::overflow_error("ankerl::unordered_dense: reached max bucket size, cannot increase size");
+        }
         --m_shifts;
         deallocate_buckets();
         allocate_buckets_from_shift();
         clear_and_fill_buckets_from_values();
     }
 
-    void do_erase(Bucket* bucket) {
-        auto const value_idx_to_remove = bucket->value_idx;
+    void do_erase(value_idx_type bucket_idx) {
+        auto const value_idx_to_remove = at(m_buckets, bucket_idx).m_value_idx;
 
         // shift down until either empty or an element with correct spot is found
-        auto* next_bucket = next(bucket);
-        while (next_bucket->dist_and_fingerprint >= BUCKET_DIST_INC * 2) {
-            *bucket = {next_bucket->dist_and_fingerprint - BUCKET_DIST_INC, next_bucket->value_idx};
-            bucket = std::exchange(next_bucket, next(next_bucket));
+        auto next_bucket_idx = next(bucket_idx);
+        while (at(m_buckets, next_bucket_idx).m_dist_and_fingerprint >= Bucket::dist_inc * 2) {
+            at(m_buckets, bucket_idx) = {dist_dec(at(m_buckets, next_bucket_idx).m_dist_and_fingerprint),
+                                         at(m_buckets, next_bucket_idx).m_value_idx};
+            bucket_idx = std::exchange(next_bucket_idx, next(next_bucket_idx));
         }
-        *bucket = {};
+        at(m_buckets, bucket_idx) = {};
 
         // update m_values
         if (value_idx_to_remove != m_values.size() - 1) {
@@ -520,13 +598,13 @@ private:
 
             // update the values_idx of the moved entry. No need to play the info game, just look until we find the values_idx
             auto mh = mixed_hash(get_key(val));
-            bucket = bucket_from_hash(mh);
+            bucket_idx = bucket_idx_from_hash(mh);
 
-            auto const values_idx_back = static_cast<uint32_t>(m_values.size() - 1);
-            while (values_idx_back != bucket->value_idx) {
-                bucket = next(bucket);
+            auto const values_idx_back = static_cast<value_idx_type>(m_values.size() - 1);
+            while (values_idx_back != at(m_buckets, bucket_idx).m_value_idx) {
+                bucket_idx = next(bucket_idx);
             }
-            bucket->value_idx = value_idx_to_remove;
+            at(m_buckets, bucket_idx).m_value_idx = value_idx_to_remove;
         }
         m_values.pop_back();
     }
@@ -537,17 +615,18 @@ private:
             return 0;
         }
 
-        auto [dist_and_fingerprint, bucket] = next_while_less(key);
+        auto [dist_and_fingerprint, bucket_idx] = next_while_less(key);
 
-        while (dist_and_fingerprint == bucket->dist_and_fingerprint && !m_equal(key, get_key(m_values[bucket->value_idx]))) {
-            dist_and_fingerprint += BUCKET_DIST_INC;
-            bucket = next(bucket);
+        while (dist_and_fingerprint == at(m_buckets, bucket_idx).m_dist_and_fingerprint &&
+               !m_equal(key, get_key(m_values[at(m_buckets, bucket_idx).m_value_idx]))) {
+            dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+            bucket_idx = next(bucket_idx);
         }
 
-        if (dist_and_fingerprint != bucket->dist_and_fingerprint) {
+        if (dist_and_fingerprint != at(m_buckets, bucket_idx).m_dist_and_fingerprint) {
             return 0;
         }
-        do_erase(bucket);
+        do_erase(bucket_idx);
         return 1;
     }
 
@@ -561,65 +640,82 @@ private:
     }
 
     template <typename K, typename... Args>
+    auto do_place_element(dist_and_fingerprint_type dist_and_fingerprint, value_idx_type bucket_idx, K&& key, Args&&... args)
+        -> std::pair<iterator, bool> {
+
+        // emplace the new value. If that throws an exception, no harm done; index is still in a valid state
+        m_values.emplace_back(std::piecewise_construct,
+                              std::forward_as_tuple(std::forward<K>(key)),
+                              std::forward_as_tuple(std::forward<Args>(args)...));
+
+        // place element and shift up until we find an empty spot
+        auto value_idx = static_cast<value_idx_type>(m_values.size() - 1);
+        place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
+        return {begin() + static_cast<difference_type>(value_idx), true};
+    }
+
+    template <typename K, typename... Args>
     auto do_try_emplace(K&& key, Args&&... args) -> std::pair<iterator, bool> {
-        if (is_full()) {
+        if (ANKERL_UNORDERED_DENSE_UNLIKELY(is_full())) {
             increase_size();
         }
 
         auto hash = mixed_hash(key);
         auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
-        auto* bucket = bucket_from_hash(hash);
+        auto bucket_idx = bucket_idx_from_hash(hash);
 
-        while (dist_and_fingerprint <= bucket->dist_and_fingerprint) {
-            if (dist_and_fingerprint == bucket->dist_and_fingerprint && m_equal(key, m_values[bucket->value_idx].first)) {
-                return {begin() + bucket->value_idx, false};
+        while (true) {
+            auto* bucket = &at(m_buckets, bucket_idx);
+            if (dist_and_fingerprint == bucket->m_dist_and_fingerprint) {
+                if (m_equal(key, m_values[bucket->m_value_idx].first)) {
+                    return {begin() + static_cast<difference_type>(bucket->m_value_idx), false};
+                }
+            } else if (dist_and_fingerprint > bucket->m_dist_and_fingerprint) {
+                return do_place_element(dist_and_fingerprint, bucket_idx, std::forward<K>(key), std::forward<Args>(args)...);
             }
-            dist_and_fingerprint += BUCKET_DIST_INC;
-            bucket = next(bucket);
+            dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+            bucket_idx = next(bucket_idx);
         }
-
-        // emplace the new value. If that throws an exception, no harm done; index is still in a valid state
-        m_values.emplace_back(std::piecewise_construct,
-                              std::forward_as_tuple(std::forward<K>(key)),
-                              std::forward_as_tuple(std::forward<Args>(args)...));
-
-        // place element and shift up until we find an empty spot
-        uint32_t value_idx = static_cast<uint32_t>(m_values.size()) - 1;
-        place_and_shift_up({dist_and_fingerprint, value_idx}, bucket);
-        return {begin() + value_idx, true};
     }
 
     template <typename K>
     auto do_find(K const& key) -> iterator {
-        if (empty()) {
+        if (ANKERL_UNORDERED_DENSE_UNLIKELY(empty())) {
             return end();
         }
 
         auto mh = mixed_hash(key);
         auto dist_and_fingerprint = dist_and_fingerprint_from_hash(mh);
-        auto const* bucket = bucket_from_hash(mh);
+        auto bucket_idx = bucket_idx_from_hash(mh);
+        auto* bucket = &at(m_buckets, bucket_idx);
 
         // unrolled loop. *Always* check a few directly, then enter the loop. This is faster.
-        if (dist_and_fingerprint == bucket->dist_and_fingerprint && m_equal(key, get_key(m_values[bucket->value_idx]))) {
-            return begin() + bucket->value_idx;
+        if (dist_and_fingerprint == bucket->m_dist_and_fingerprint && m_equal(key, get_key(m_values[bucket->m_value_idx]))) {
+            return begin() + static_cast<difference_type>(bucket->m_value_idx);
         }
-        dist_and_fingerprint += BUCKET_DIST_INC;
-        bucket = next(bucket);
+        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+        bucket_idx = next(bucket_idx);
+        bucket = &at(m_buckets, bucket_idx);
 
-        if (dist_and_fingerprint == bucket->dist_and_fingerprint && m_equal(key, get_key(m_values[bucket->value_idx]))) {
-            return begin() + bucket->value_idx;
+        if (dist_and_fingerprint == bucket->m_dist_and_fingerprint && m_equal(key, get_key(m_values[bucket->m_value_idx]))) {
+            return begin() + static_cast<difference_type>(bucket->m_value_idx);
         }
-        dist_and_fingerprint += BUCKET_DIST_INC;
-        bucket = next(bucket);
-
-        do {
-            if (dist_and_fingerprint == bucket->dist_and_fingerprint && m_equal(key, get_key(m_values[bucket->value_idx]))) {
-                return begin() + bucket->value_idx;
+        dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+        bucket_idx = next(bucket_idx);
+        bucket = &at(m_buckets, bucket_idx);
+
+        while (true) {
+            if (dist_and_fingerprint == bucket->m_dist_and_fingerprint) {
+                if (m_equal(key, get_key(m_values[bucket->m_value_idx]))) {
+                    return begin() + static_cast<difference_type>(bucket->m_value_idx);
+                }
+            } else if (dist_and_fingerprint > bucket->m_dist_and_fingerprint) {
+                return end();
             }
-            dist_and_fingerprint += BUCKET_DIST_INC;
-            bucket = next(bucket);
-        } while (dist_and_fingerprint <= bucket->dist_and_fingerprint);
-        return end();
+            dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+            bucket_idx = next(bucket_idx);
+            bucket = &at(m_buckets, bucket_idx);
+        }
     }
 
     template <typename K>
@@ -634,19 +730,22 @@ public:
     explicit table(size_t /*bucket_count*/,
                    Hash const& hash = Hash(),
                    KeyEqual const& equal = KeyEqual(),
-                   Allocator const& alloc = Allocator())
-        : m_values(alloc)
+                   AllocatorOrContainer const& alloc_or_container = AllocatorOrContainer())
+        : m_values(alloc_or_container)
         , m_hash(hash)
-        , m_equal(equal) {}
+        , m_equal(equal) {
+        // If alloc_or_container is a container with elements, we don't want the data that was in it
+        m_values.clear();
+    }
 
-    table(size_t bucket_count, Allocator const& alloc)
-        : table(bucket_count, Hash(), KeyEqual(), alloc) {}
+    table(size_t bucket_count, AllocatorOrContainer const& alloc_or_container)
+        : table(bucket_count, Hash(), KeyEqual(), alloc_or_container) {}
 
-    table(size_t bucket_count, Hash const& hash, Allocator const& alloc)
-        : table(bucket_count, hash, KeyEqual(), alloc) {}
+    table(size_t bucket_count, Hash const& hash, AllocatorOrContainer const& alloc_or_container)
+        : table(bucket_count, hash, KeyEqual(), alloc_or_container) {}
 
-    explicit table(Allocator const& alloc)
-        : table(0, Hash(), KeyEqual(), alloc) {}
+    explicit table(AllocatorOrContainer const& alloc_or_container)
+        : table(0, Hash(), KeyEqual(), alloc_or_container) {}
 
     template <class InputIt>
     table(InputIt first,
@@ -654,24 +753,25 @@ public:
           size_type bucket_count = 0,
           Hash const& hash = Hash(),
           KeyEqual const& equal = KeyEqual(),
-          Allocator const& alloc = Allocator())
-        : table(bucket_count, hash, equal, alloc) {
+          AllocatorOrContainer const& alloc_or_container = AllocatorOrContainer())
+        : table(bucket_count, hash, equal, alloc_or_container) {
         insert(first, last);
     }
 
     template <class InputIt>
-    table(InputIt first, InputIt last, size_type bucket_count, Allocator const& alloc)
-        : table(first, last, bucket_count, Hash(), KeyEqual(), alloc) {}
+    table(InputIt first, InputIt last, size_type bucket_count, AllocatorOrContainer const& alloc_or_container)
+        : table(first, last, bucket_count, Hash(), KeyEqual(), alloc_or_container) {}
 
     template <class InputIt>
-    table(InputIt first, InputIt last, size_type bucket_count, Hash const& hash, Allocator const& alloc)
-        : table(first, last, bucket_count, hash, KeyEqual(), alloc) {}
+    table(
+        InputIt first, InputIt last, size_type bucket_count, Hash const& hash, AllocatorOrContainer const& alloc_or_container)
+        : table(first, last, bucket_count, hash, KeyEqual(), alloc_or_container) {}
 
     table(table const& other)
         : table(other, other.m_values.get_allocator()) {}
 
-    table(table const& other, Allocator const& alloc)
-        : m_values(other.m_values, alloc)
+    table(table const& other, AllocatorOrContainer const& alloc_or_container)
+        : m_values(other.m_values, alloc_or_container)
         , m_max_load_factor(other.m_max_load_factor)
         , m_hash(other.m_hash)
         , m_equal(other.m_equal) {
@@ -681,15 +781,15 @@ public:
     table(table&& other) noexcept
         : table(std::move(other), other.m_values.get_allocator()) {}
 
-    table(table&& other, Allocator const& alloc) noexcept
-        : m_values(std::move(other.m_values), alloc)
-        , m_buckets_start(std::exchange(other.m_buckets_start, nullptr))
-        , m_buckets_end(std::exchange(other.m_buckets_end, nullptr))
+    table(table&& other, AllocatorOrContainer const& alloc_or_container) noexcept
+        : m_values(std::move(other.m_values), alloc_or_container)
+        , m_buckets(std::exchange(other.m_buckets, nullptr))
+        , m_num_buckets(std::exchange(other.m_num_buckets, 0))
         , m_max_bucket_capacity(std::exchange(other.m_max_bucket_capacity, 0))
-        , m_max_load_factor(std::exchange(other.m_max_load_factor, DEFAULT_MAX_LOAD_FACTOR))
+        , m_max_load_factor(std::exchange(other.m_max_load_factor, default_max_load_factor))
         , m_hash(std::exchange(other.m_hash, {}))
         , m_equal(std::exchange(other.m_equal, {}))
-        , m_shifts(std::exchange(other.m_shifts, INITIAL_SHIFTS)) {
+        , m_shifts(std::exchange(other.m_shifts, initial_shifts)) {
         other.m_values.clear();
     }
 
@@ -697,20 +797,23 @@ public:
           size_t bucket_count = 0,
           Hash const& hash = Hash(),
           KeyEqual const& equal = KeyEqual(),
-          Allocator const& alloc = Allocator())
-        : table(bucket_count, hash, equal, alloc) {
+          AllocatorOrContainer const& alloc_or_container = AllocatorOrContainer())
+        : table(bucket_count, hash, equal, alloc_or_container) {
         insert(ilist);
     }
 
-    table(std::initializer_list<value_type> ilist, size_type bucket_count, const Allocator& alloc)
-        : table(ilist, bucket_count, Hash(), KeyEqual(), alloc) {}
+    table(std::initializer_list<value_type> ilist, size_type bucket_count, AllocatorOrContainer const& alloc_or_container)
+        : table(ilist, bucket_count, Hash(), KeyEqual(), alloc_or_container) {}
 
-    table(std::initializer_list<value_type> init, size_type bucket_count, Hash const& hash, Allocator const& alloc)
-        : table(init, bucket_count, hash, KeyEqual(), alloc) {}
+    table(std::initializer_list<value_type> init,
+          size_type bucket_count,
+          Hash const& hash,
+          AllocatorOrContainer const& alloc_or_container)
+        : table(init, bucket_count, hash, KeyEqual(), alloc_or_container) {}
 
     ~table() {
-        auto bucket_alloc = BucketAlloc(m_values.get_allocator());
-        BucketAllocTraits::deallocate(bucket_alloc, m_buckets_start, bucket_count());
+        auto ba = bucket_alloc(m_values.get_allocator());
+        bucket_alloc_traits::deallocate(ba, m_buckets, bucket_count());
     }
 
     auto operator=(table const& other) -> table& {
@@ -720,25 +823,25 @@ public:
             m_max_load_factor = other.m_max_load_factor;
             m_hash = other.m_hash;
             m_equal = other.m_equal;
-            m_shifts = INITIAL_SHIFTS;
+            m_shifts = initial_shifts;
             copy_buckets(other);
         }
         return *this;
     }
 
     auto operator=(table&& other) noexcept(
-        noexcept(std::is_nothrow_move_assignable_v<ValueContainer>&& std::is_nothrow_move_assignable_v<Hash>&&
+        noexcept(std::is_nothrow_move_assignable_v<value_container_type>&& std::is_nothrow_move_assignable_v<Hash>&&
                      std::is_nothrow_move_assignable_v<KeyEqual>)) -> table& {
         if (&other != this) {
             deallocate_buckets(); // deallocate before m_values is set (might have another allocator)
             m_values = std::move(other.m_values);
-            m_buckets_start = std::exchange(other.m_buckets_start, nullptr);
-            m_buckets_end = std::exchange(other.m_buckets_end, nullptr);
+            m_buckets = std::exchange(other.m_buckets, nullptr);
+            m_num_buckets = std::exchange(other.m_num_buckets, 0);
             m_max_bucket_capacity = std::exchange(other.m_max_bucket_capacity, 0);
-            m_max_load_factor = std::exchange(other.m_max_load_factor, DEFAULT_MAX_LOAD_FACTOR);
+            m_max_load_factor = std::exchange(other.m_max_load_factor, default_max_load_factor);
             m_hash = std::exchange(other.m_hash, {});
             m_equal = std::exchange(other.m_equal, {});
-            m_shifts = std::exchange(other.m_shifts, INITIAL_SHIFTS);
+            m_shifts = std::exchange(other.m_shifts, initial_shifts);
             other.m_values.clear();
         }
         return *this;
@@ -790,8 +893,12 @@ public:
         return m_values.size();
     }
 
-    [[nodiscard]] auto max_size() const noexcept -> size_t {
-        return std::numeric_limits<uint32_t>::max();
+    [[nodiscard]] static constexpr auto max_size() noexcept -> size_t {
+        if constexpr (std::numeric_limits<value_idx_type>::max() == std::numeric_limits<size_t>::max()) {
+            return size_t{1} << (sizeof(value_idx_type) * 8 - 1);
+        } else {
+            return size_t{1} << (sizeof(value_idx_type) * 8);
+        }
     }
 
     // modifiers //////////////////////////////////////////////////////////////
@@ -839,6 +946,67 @@ public:
         insert(ilist.begin(), ilist.end());
     }
 
+    // nonstandard API: *this is emptied.
+    // Also see "A Standard flat_map" https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2022/p0429r9.pdf
+    auto extract() && -> value_container_type {
+        return std::move(m_values);
+    }
+
+    // nonstandard API:
+    // Discards the internally held container and replaces it with the one passed. Erases non-unique elements.
+    auto replace(value_container_type&& container) {
+        if (container.size() > max_size()) {
+            throw std::out_of_range("ankerl::unordered_dense::map::replace(): too many elements");
+        }
+
+        auto shifts = calc_shifts_for_size(container.size());
+        if (0 == m_num_buckets || shifts < m_shifts || container.get_allocator() != m_values.get_allocator()) {
+            m_shifts = shifts;
+            deallocate_buckets();
+            allocate_buckets_from_shift();
+        }
+        clear_buckets();
+
+        m_values = std::move(container);
+
+        // can't use clear_and_fill_buckets_from_values() because container elements might not be unique
+        auto value_idx = value_idx_type{};
+
+        // loop until we reach the end of the container. duplicated entries will be replaced with back().
+        while (value_idx != static_cast<value_idx_type>(m_values.size())) {
+            auto const& key = get_key(m_values[value_idx]);
+
+            auto hash = mixed_hash(key);
+            auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
+            auto bucket_idx = bucket_idx_from_hash(hash);
+
+            bool key_found = false;
+            while (true) {
+                auto const& bucket = at(m_buckets, bucket_idx);
+                if (dist_and_fingerprint > bucket.m_dist_and_fingerprint) {
+                    break;
+                }
+                if (dist_and_fingerprint == bucket.m_dist_and_fingerprint &&
+                    m_equal(key, m_values[bucket.m_value_idx].first)) {
+                    key_found = true;
+                    break;
+                }
+                dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+                bucket_idx = next(bucket_idx);
+            }
+
+            if (key_found) {
+                if (value_idx != static_cast<value_idx_type>(m_values.size() - 1)) {
+                    m_values[value_idx] = std::move(m_values.back());
+                }
+                m_values.pop_back();
+            } else {
+                place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
+                ++value_idx;
+            }
+        }
+    }
+
     template <class M, typename Q = T, std::enable_if_t<!std::is_void_v<Q>, bool> = true>
     auto insert_or_assign(Key const& key, M&& mapped) -> std::pair<iterator, bool> {
         return do_insert_or_assign(key, std::forward<M>(mapped));
@@ -869,23 +1037,23 @@ public:
         auto& val = m_values.emplace_back(std::forward<Args>(args)...);
         auto hash = mixed_hash(get_key(val));
         auto dist_and_fingerprint = dist_and_fingerprint_from_hash(hash);
-        auto* bucket = bucket_from_hash(hash);
+        auto bucket_idx = bucket_idx_from_hash(hash);
 
-        while (dist_and_fingerprint <= bucket->dist_and_fingerprint) {
-            if (dist_and_fingerprint == bucket->dist_and_fingerprint &&
-                m_equal(get_key(val), get_key(m_values[bucket->value_idx]))) {
+        while (dist_and_fingerprint <= at(m_buckets, bucket_idx).m_dist_and_fingerprint) {
+            if (dist_and_fingerprint == at(m_buckets, bucket_idx).m_dist_and_fingerprint &&
+                m_equal(get_key(val), get_key(m_values[at(m_buckets, bucket_idx).m_value_idx]))) {
                 m_values.pop_back(); // value was already there, so get rid of it
-                return {begin() + bucket->value_idx, false};
+                return {begin() + static_cast<difference_type>(at(m_buckets, bucket_idx).m_value_idx), false};
             }
-            dist_and_fingerprint += BUCKET_DIST_INC;
-            bucket = next(bucket);
+            dist_and_fingerprint = dist_inc(dist_and_fingerprint);
+            bucket_idx = next(bucket_idx);
         }
 
         // value is new, place the bucket and shift up until we find an empty spot
-        uint32_t value_idx = static_cast<uint32_t>(m_values.size()) - 1;
-        place_and_shift_up({dist_and_fingerprint, value_idx}, bucket);
+        auto value_idx = static_cast<value_idx_type>(m_values.size() - 1);
+        place_and_shift_up({dist_and_fingerprint, value_idx}, bucket_idx);
 
-        return {begin() + value_idx, true};
+        return {begin() + static_cast<difference_type>(value_idx), true};
     }
 
     template <class... Args>
@@ -915,15 +1083,15 @@ public:
 
     auto erase(iterator it) -> iterator {
         auto hash = mixed_hash(get_key(*it));
-        auto* bucket = bucket_from_hash(hash);
+        auto bucket_idx = bucket_idx_from_hash(hash);
 
-        auto const value_idx_to_remove = static_cast<uint32_t>(it - cbegin());
-        while (bucket->value_idx != value_idx_to_remove) {
-            bucket = next(bucket);
+        auto const value_idx_to_remove = static_cast<value_idx_type>(it - cbegin());
+        while (at(m_buckets, bucket_idx).m_value_idx != value_idx_to_remove) {
+            bucket_idx = next(bucket_idx);
         }
 
-        do_erase(bucket);
-        return begin() + value_idx_to_remove;
+        do_erase(bucket_idx);
+        return begin() + static_cast<difference_type>(value_idx_to_remove);
     }
 
     auto erase(const_iterator it) -> iterator {
@@ -963,8 +1131,8 @@ public:
         return do_erase_key(std::forward<K>(key));
     }
 
-    void swap(table& other) noexcept(noexcept(std::is_nothrow_swappable_v<ValueContainer>&& std::is_nothrow_swappable_v<Hash>&&
-                                                  std::is_nothrow_swappable_v<KeyEqual>)) {
+    void swap(table& other) noexcept(noexcept(std::is_nothrow_swappable_v<value_container_type>&&
+                                                  std::is_nothrow_swappable_v<Hash>&& std::is_nothrow_swappable_v<KeyEqual>)) {
         using std::swap;
         swap(other, *this);
     }
@@ -1055,17 +1223,17 @@ public:
     // bucket interface ///////////////////////////////////////////////////////
 
     auto bucket_count() const noexcept -> size_t { // NOLINT(modernize-use-nodiscard)
-        return m_buckets_end - m_buckets_start;
+        return m_num_buckets;
     }
 
-    auto max_bucket_count() const noexcept -> size_t { // NOLINT(modernize-use-nodiscard)
-        return std::numeric_limits<uint32_t>::max();
+    static constexpr auto max_bucket_count() noexcept -> size_t { // NOLINT(modernize-use-nodiscard)
+        return max_size();
     }
 
     // hash policy ////////////////////////////////////////////////////////////
 
     [[nodiscard]] auto load_factor() const -> float {
-        return bucket_count() ? static_cast<float>(size()) / bucket_count() : 0.0F;
+        return bucket_count() ? static_cast<float>(size()) / static_cast<float>(bucket_count()) : 0.0F;
     }
 
     [[nodiscard]] auto max_load_factor() const -> float {
@@ -1074,10 +1242,13 @@ public:
 
     void max_load_factor(float ml) {
         m_max_load_factor = ml;
-        m_max_bucket_capacity = static_cast<uint32_t>(bucket_count() * max_load_factor());
+        if (m_num_buckets != max_bucket_count()) {
+            m_max_bucket_capacity = static_cast<value_idx_type>(static_cast<float>(bucket_count()) * max_load_factor());
+        }
     }
 
     void rehash(size_t count) {
+        count = std::min(count, max_size());
         auto shifts = calc_shifts_for_size(std::max(count, size()));
         if (shifts != m_shifts) {
             m_shifts = shifts;
@@ -1089,8 +1260,10 @@ public:
     }
 
     void reserve(size_t capa) {
+        capa = std::min(capa, max_size());
+        m_values.reserve(capa);
         auto shifts = calc_shifts_for_size(std::max(capa, size()));
-        if (shifts < m_shifts) {
+        if (0 == m_num_buckets || shifts < m_shifts) {
             m_shifts = shifts;
             deallocate_buckets();
             allocate_buckets_from_shift();
@@ -1108,6 +1281,11 @@ public:
         return m_equal;
     }
 
+    // nonstandard API: expose the underlying values container
+    [[nodiscard]] auto values() const noexcept -> value_container_type const& {
+        return m_values;
+    }
+
     // non-member functions ///////////////////////////////////////////////////
 
     friend auto operator==(table const& a, table const& b) -> bool {
@@ -1145,21 +1323,30 @@ template <class Key,
           class T,
           class Hash = hash<Key>,
           class KeyEqual = std::equal_to<Key>,
-          class Allocator = std::allocator<std::pair<Key, T>>>
-using map = detail::table<Key, T, Hash, KeyEqual, Allocator>;
+          class AllocatorOrContainer = std::allocator<std::pair<Key, T>>,
+          class Bucket = bucket_type::standard>
+using map = detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket>;
 
-template <class Key, class Hash = hash<Key>, class KeyEqual = std::equal_to<Key>, class Allocator = std::allocator<Key>>
-using set = detail::table<Key, void, Hash, KeyEqual, Allocator>;
+template <class Key,
+          class Hash = hash<Key>,
+          class KeyEqual = std::equal_to<Key>,
+          class AllocatorOrContainer = std::allocator<Key>,
+          class Bucket = bucket_type::standard>
+using set = detail::table<Key, void, Hash, KeyEqual, AllocatorOrContainer, Bucket>;
 
 #    if ANKERL_UNORDERED_DENSE_PMR
 
 namespace pmr {
 
-template <class Key, class T, class Hash = hash<Key>, class KeyEqual = std::equal_to<Key>>
-using map = detail::table<Key, T, Hash, KeyEqual, std::pmr::polymorphic_allocator<std::pair<Key, T>>>;
+template <class Key,
+          class T,
+          class Hash = hash<Key>,
+          class KeyEqual = std::equal_to<Key>,
+          class Bucket = bucket_type::standard>
+using map = detail::table<Key, T, Hash, KeyEqual, std::pmr::polymorphic_allocator<std::pair<Key, T>>, Bucket>;
 
-template <class Key, class Hash = hash<Key>, class KeyEqual = std::equal_to<Key>>
-using set = detail::table<Key, void, Hash, KeyEqual, std::pmr::polymorphic_allocator<Key>>;
+template <class Key, class Hash = hash<Key>, class KeyEqual = std::equal_to<Key>, class Bucket = bucket_type::standard>
+using set = detail::table<Key, void, Hash, KeyEqual, std::pmr::polymorphic_allocator<Key>, Bucket>;
 
 } // namespace pmr
 
@@ -1167,22 +1354,27 @@ using set = detail::table<Key, void, Hash, KeyEqual, std::pmr::polymorphic_alloc
 
 // deduction guides ///////////////////////////////////////////////////////////
 
-// TODO not yet implemented
+// deduction guides for alias templates are only possible since C++20
+// see https://en.cppreference.com/w/cpp/language/class_template_argument_deduction
 
+} // namespace ANKERL_UNORDERED_DENSE_NAMESPACE
 } // namespace ankerl::unordered_dense
 
 // std extensions /////////////////////////////////////////////////////////////
 
 namespace std { // NOLINT(cert-dcl58-cpp)
 
-template <class Key, class T, class Hash, class KeyEqual, class Allocator, class Pred>
-auto erase_if(ankerl::unordered_dense::detail::table<Key, T, Hash, KeyEqual, Allocator>& map, Pred pred) -> size_t {
+template <class Key, class T, class Hash, class KeyEqual, class AllocatorOrContainer, class Bucket, class Pred>
+auto erase_if(ankerl::unordered_dense::detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket>& map, Pred pred)
+    -> size_t {
+    using map_t = ankerl::unordered_dense::detail::table<Key, T, Hash, KeyEqual, AllocatorOrContainer, Bucket>;
+
     // going back to front because erase() invalidates the end iterator
     auto const old_size = map.size();
     auto idx = old_size;
     while (idx) {
         --idx;
-        auto it = map.begin() + idx;
+        auto it = map.begin() + static_cast<typename map_t::difference_type>(idx);
         if (pred(*it)) {
             map.erase(it);
         }
diff --git a/benchmarks/external/emhash/hash_table7.hpp b/benchmarks/external/emhash/hash_table7.hpp
index 7daedff2..c4761a0a 100644
--- a/benchmarks/external/emhash/hash_table7.hpp
+++ b/benchmarks/external/emhash/hash_table7.hpp
@@ -1,5 +1,5 @@
 // emhash7::HashMap for C++11/14/17
-// version 2.1.2
+// version 2.2.3
 // https://github.com/ktprime/ktprime/blob/master/hash_table7.hpp
 //
 // Licensed under the MIT License <http://opensource.org/licenses/MIT>.
@@ -159,10 +159,9 @@ of resizing granularity. Ignoring variance, the expected occurrences of list siz
 
 namespace emhash7 {
 
-    constexpr static float EMH_DEFAULT_LOAD_FACTOR =  0.80f;
 #ifdef EMH_SIZE_TYPE_16BIT
     typedef uint16_t size_type;
-    static constexpr size_type INACTIVE = 0xFFFE;
+    static constexpr size_type INACTIVE = 0xFFFF;
 #elif EMH_SIZE_TYPE_64BIT
     typedef uint64_t size_type;
     static constexpr size_type INACTIVE = 0 - 0x1ull;
@@ -298,21 +297,25 @@ struct entry {
         std::swap(first, o.first);
     }
 
-#ifndef EMH_ORDER_KV
-    Second second;//int
+#if EMH_ORDER_KV || EMH_SIZE_TYPE_64BIT
+    Second second;
     size_type bucket;
-    First first; //long
+    First first;
 #else
-    First first; //long
+    First first;
     size_type bucket;
-    Second second;//int
+    Second second;
 #endif
-};// __attribute__ ((packed));
+};
 
 /// A cache-friendly hash table with open addressing, linear/qua probing and power-of-two capacity
 template <typename KeyT, typename ValueT, typename HashT = std::hash<KeyT>, typename EqT = std::equal_to<KeyT>>
 class HashMap
 {
+#ifndef EMH_DEFAULT_LOAD_FACTOR
+    constexpr static float EMH_DEFAULT_LOAD_FACTOR = 0.80f;
+#endif
+
 public:
     typedef HashMap<KeyT, ValueT, HashT, EqT> htype;
     typedef std::pair<KeyT,ValueT>            value_type;
@@ -350,7 +353,11 @@ public:
         iterator() : _map(nullptr) { }
         iterator(const const_iterator& it) : _map(it._map), _bucket(it._bucket), _from(it._from), _bmask(it._bmask) { }
         iterator(const htype* hash_map, size_type bucket, bool) : _map(hash_map), _bucket(bucket) { init(); }
+#if EMH_ITER_SAFE
+        iterator(const htype* hash_map, size_type bucket) : _map(hash_map), _bucket(bucket) { init(); }
+#else
         iterator(const htype* hash_map, size_type bucket) : _map(hash_map), _bucket(bucket) { _bmask = _from = 0; }
+#endif
 
         void init()
         {
@@ -445,7 +452,11 @@ public:
 
         const_iterator(const iterator& it) : _map(it._map), _bucket(it._bucket), _from(it._from), _bmask(it._bmask) { }
         const_iterator(const htype* hash_map, size_type bucket, bool) : _map(hash_map), _bucket(bucket) { init(); }
+#if EMH_ITER_SAFE
+        const_iterator(const htype* hash_map, size_type bucket) : _map(hash_map), _bucket(bucket) { init(); }
+#else
         const_iterator(const htype* hash_map, size_type bucket) : _map(hash_map), _bucket(bucket) { _bmask = _from = 0; }
+#endif
 
         void init()
         {
@@ -522,23 +533,23 @@ public:
         rehash(bucket);
     }
 
-    HashMap(size_type bucket = 2, float mlf = EMH_DEFAULT_LOAD_FACTOR)
+    HashMap(size_type bucket = 2, float mlf = EMH_DEFAULT_LOAD_FACTOR) noexcept
     {
         init(bucket, mlf);
     }
 
-    size_type AllocSize(size_type num_buckets) const
+    size_t AllocSize(uint64_t num_buckets) const
     {
-        return num_buckets * sizeof(PairT) + EPACK_SIZE * sizeof(PairT) + (num_buckets + 7) / 8 + BIT_PACK;
+        return (num_buckets + EPACK_SIZE) * sizeof(PairT) + (num_buckets + 7) / 8 + BIT_PACK;
     }
 
-    HashMap(const HashMap& rhs)
+    HashMap(const HashMap& rhs) noexcept
     {
         _pairs = (PairT*)malloc(AllocSize(rhs._num_buckets));
         clone(rhs);
     }
 
-    HashMap(HashMap&& rhs)
+    HashMap(HashMap&& rhs) noexcept
     {
 #ifndef EMH_ZERO_MOVE
         init(4);
@@ -564,7 +575,7 @@ public:
             emplace(*first);
     }
 
-    HashMap& operator=(const HashMap& rhs)
+    HashMap& operator= (const HashMap& rhs) noexcept
     {
         if (this == &rhs)
             return *this;
@@ -581,7 +592,7 @@ public:
         return *this;
     }
 
-    HashMap& operator=(HashMap&& rhs)
+    HashMap& operator= (HashMap&& rhs) noexcept
     {
         if (this != &rhs) {
             swap(rhs);
@@ -621,7 +632,7 @@ public:
     void clone(const HashMap& rhs)
     {
         _hasher      = rhs._hasher;
-//        _eq          = rhs._eq;
+        //_eq          = rhs._eq;
 
         _num_filled  = rhs._num_filled;
         _mask        = rhs._mask;
@@ -646,18 +657,17 @@ public:
     void swap(HashMap& rhs)
     {
         std::swap(_hasher, rhs._hasher);
-        //      std::swap(_eq, rhs._eq);
+        //std::swap(_eq, rhs._eq);
         std::swap(_pairs, rhs._pairs);
         std::swap(_num_buckets, rhs._num_buckets);
         std::swap(_num_filled, rhs._num_filled);
         std::swap(_mask, rhs._mask);
         std::swap(_mlf, rhs._mlf);
         std::swap(_bitmask, rhs._bitmask);
-        //std::swap(EMH_BUCKET(_pairs, _num_buckets), EMH_BUCKET(rhs._pairs, rhs._num_buckets));
     }
 
     // -------------------------------------------------------------
-    iterator begin()
+    iterator begin() noexcept
     {
 #ifdef EMH_ZERO_MOVE
         if (0 == _num_filled)
@@ -672,7 +682,7 @@ public:
         return it.next();
     }
 
-    const_iterator cbegin() const
+    const_iterator cbegin() const noexcept
     {
 #ifdef EMH_ZERO_MOVE
         if (0 == _num_filled)
@@ -697,9 +707,9 @@ public:
         return {this, bucket, true};
     }
 
-    const_iterator begin() const { return cbegin(); }
+    const_iterator begin() const noexcept { return cbegin(); }
 
-    iterator end() { return {this, _num_buckets}; }
+    iterator end() noexcept { return {this, _num_buckets}; }
     const_iterator cend() const { return {this, _num_buckets}; }
     const_iterator end() const { return cend(); }
 
@@ -719,10 +729,9 @@ public:
     }
 
     constexpr float max_load_factor() const { return (1 << 27) / (float)_mlf; }
-    constexpr size_type max_size() const { return (1ull << (sizeof(size_type) * 8 - 2)); }
+    constexpr size_type max_size() const { return 1ull << (sizeof(size_type) * 8 - 1); }
     constexpr size_type max_bucket_count() const { return max_size(); }
 
-#if EMH_STATIS
     size_type bucket_main() const
     {
         auto main_size = 0;
@@ -733,6 +742,7 @@ public:
         return main_size;
     }
 
+#if EMH_STATIS
     //Returns the bucket number where the element with key k is located.
     size_type bucket(const KeyT& key) const
     {
@@ -981,7 +991,7 @@ public:
         }
     }
 
-    /// Returns false if key isn't found.
+#ifdef EMH_EXT
     bool try_get(const KeyT& key, ValueT& val) const noexcept
     {
         const auto bucket = find_filled_bucket(key);
@@ -1012,6 +1022,7 @@ public:
         const auto bucket = find_filled_bucket(key);
         return bucket == _num_buckets ? ValueT() : EMH_VAL(_pairs, bucket);
     }
+#endif
 
     // -----------------------------------------------------
     template<typename K = KeyT, typename V = ValueT>
@@ -1066,7 +1077,7 @@ public:
         return do_insert(value);
     }
 
-    std::pair<iterator, bool> insert(value_type && value)
+    std::pair<iterator, bool> insert(value_type&& value)
     {
         check_expand_need();
         return do_insert(std::move(value));
@@ -1087,34 +1098,6 @@ public:
             do_insert(it->first, it->second);
     }
 
-#if 0
-    template <typename Iter>
-    void insert2(Iter begin, Iter end)
-    {
-        Iter citbeg = begin;
-        Iter citend = begin;
-        reserve(std::distance(begin, end) + _num_filled);
-        for (; begin != end; ++begin) {
-            if (try_insert_mainbucket(begin->first, begin->second) == INACTIVE) {
-                std::swap(*begin, *citend++);
-            }
-        }
-
-        for (; citbeg != citend; ++citbeg)
-            insert(*citbeg);
-    }
-
-    size_type try_insert_mainbucket(const KeyT& key, const ValueT& val)
-    {
-        const auto bucket = hash_key(key) & _mask;
-        if (!EMH_EMPTY(_pairs, bucket))
-            return INACTIVE;
-
-        EMH_NEW(key, val, bucket);
-        return bucket;
-    }
-#endif
-
     template <typename Iter>
     void insert_unique(Iter begin, Iter end)
     {
@@ -1123,15 +1106,10 @@ public:
             do_insert_unqiue(*begin);
     }
 
-    /// Same as above, but contains(key) MUST be false
-    size_type insert_unique(KeyT&& key, ValueT&& val)
-    {
-        return do_insert_unqiue(std::move(key), std::forward<ValueT>(val));
-    }
-
-    size_type insert_unique(const KeyT& key, ValueT&& val)
+    template<typename K, typename V>
+    size_type insert_unique(K&& key, V&& val)
     {
-        return do_insert_unqiue(key, std::forward<ValueT>(val));
+        return do_insert_unqiue(std::forward<K>(key), std::forward<V>(val));
     }
 
     size_type insert_unique(value_type&& value)
@@ -1157,7 +1135,7 @@ public:
     std::pair<iterator, bool> insert_or_assign(KeyT&& key, ValueT&& val) { return do_assign(std::move(key), std::forward<ValueT>(val)); }
 
     template <typename... Args>
-    inline std::pair<iterator, bool> emplace(Args&&... args)
+    inline std::pair<iterator, bool> emplace(Args&&... args) noexcept
     {
         check_expand_need();
         return do_insert(std::forward<Args>(args)...);
@@ -1186,13 +1164,13 @@ public:
     }
 
     template <class... Args>
-    inline size_type emplace_unique(Args&&... args)
+    inline size_type emplace_unique(Args&&... args) noexcept
     {
         return insert_unique(std::forward<Args>(args)...);
     }
 
     /* Check if inserting a new value rather than overwriting an old entry */
-    ValueT& operator[](const KeyT& key)
+    ValueT& operator[](const KeyT& key) noexcept
     {
         check_expand_need();
 
@@ -1205,7 +1183,7 @@ public:
         return EMH_VAL(_pairs, bucket);
     }
 
-    ValueT& operator[](KeyT&& key)
+    ValueT& operator[](KeyT&& key) noexcept
     {
         check_expand_need();
 
@@ -1261,16 +1239,6 @@ public:
         clear_bucket(bucket);
     }
 
-    iterator erase(const_iterator first, const_iterator last)
-    {
-        auto iend = cend();
-        auto next = first;
-        for (; next.bucket() < last.bucket() && next != iend; )
-            next = erase(next);
-
-        return {this, next.bucket()};
-    }
-
     template<typename Pred>
     size_type erase_if(Pred pred)
     {
@@ -1313,8 +1281,10 @@ public:
     /// Remove all elements, keeping full capacity.
     void clear()
     {
-        if (!is_triviall_destructable() && _num_filled)
+        if (!is_triviall_destructable() && _num_filled) {
             memset(_bitmask, 0xFFFFFFFF, (_num_buckets + 7) / 8);
+            if (_num_buckets < 8) _bitmask[0] = (1 << _num_buckets) - 1;
+        }
         else if (_num_filled)
             clearkv();
 
@@ -1351,10 +1321,12 @@ public:
         if (required_buckets < _num_filled)
             return;
 
-        auto num_buckets = _num_filled > (1u << 16) ? (1u << 16) : 2u;
-        while (num_buckets < required_buckets) { num_buckets *= 2; }
-
+        uint64_t buckets = _num_filled > (1u << 16) ? (1u << 16) : 2u;
+        while (buckets < required_buckets) { buckets *= 2; }
+        assert(buckets < max_size() && buckets > _num_filled);
         //TODO: throwOverflowError
+
+        auto num_buckets = (size_type)buckets;
         auto old_num_filled = _num_filled;
         auto old_mask  = _num_buckets - 1;
         auto old_pairs = _pairs;
@@ -1391,8 +1363,9 @@ public:
         if (_num_filled > EMH_REHASH_LOG) {
             auto mbucket = bucket_main();
             char buff[255] = {0};
-            sprintf(buff, "    _num_filled/aver_size/K.V/pack/ = %u/%2.lf/%s.%s/%zd",
-                    _num_filled, double (_num_filled) / mbucket, typeid(KeyT).name(), typeid(ValueT).name(), sizeof(_pairs[0]));
+            sprintf(buff, "    _num_filled/collision/main/K.V/pack/ = %u/%.2lf%%(%.2lf%%)/%s.%s/%zd",
+                    _num_filled, 200.0f * (_num_filled - mbucket) / _mask,  100.0f * mbucket / _mask,
+                    typeid(KeyT).name(), typeid(ValueT).name(), sizeof(_pairs[0]));
 #ifdef EMH_LOG
             static size_t ihashs = 0;
             EMH_LOG << "rhash_nums = " << ihashs ++ << "|" <<__FUNCTION__ << "|" << buff << endl;
@@ -1538,12 +1511,7 @@ private:
         if (EMH_EMPTY(_pairs, bucket))
             return _num_buckets;
 
-        auto next_bucket = EMH_BUCKET(_pairs, bucket);
-        if (_eq(key, EMH_KEY(_pairs, bucket)))
-            return bucket;
-        else if (next_bucket == bucket)
-            return _num_buckets;
-
+        auto next_bucket = bucket;
         while (true) {
             if (_eq(key, EMH_KEY(_pairs, next_bucket)))
                 return next_bucket;
@@ -1565,36 +1533,7 @@ private:
         if (EMH_EMPTY(_pairs, bucket))
             return _num_buckets;
 
-#if 1
-        if (_eq(key, EMH_KEY(_pairs, bucket)))
-            return bucket;
-
-        auto next_bucket = EMH_BUCKET(_pairs, bucket);
-        if (next_bucket == bucket)
-            return _num_buckets;
-#elif 0
-        else if (_eq(key, EMH_KEY(_pairs, bucket)))
-            return bucket;
-        else if (next_bucket == bucket)
-            return _num_buckets;
-
-        else if (_eq(key, EMH_KEY(_pairs, next_bucket)))
-            return next_bucket;
-        const auto nbucket = EMH_BUCKET(_pairs, next_bucket);
-        if (nbucket == next_bucket)
-            return _num_buckets;
-        next_bucket = nbucket;
-#elif 0
-        const auto bucket = hash_key(key) & _mask;
-        if (EMH_EMPTY(_pairs, bucket))
-            return _num_buckets;
-        else if (_eq(key, EMH_KEY(_pairs, bucket)))
-            return bucket;
-
-        auto next_bucket = EMH_BUCKET(_pairs, bucket);
-        if (next_bucket == bucket)
-            return _num_buckets;
-#endif
+        auto next_bucket = bucket;
 //        else if (bucket != (hash_key(bucket_key) & _mask))
 //            return _num_buckets;
 
@@ -1628,11 +1567,7 @@ private:
             EMH_BUCKET(_pairs, new_bucket) = new_bucket;
         EMH_BUCKET(_pairs, prev_bucket) = new_bucket;
 
-        //set new bucket bit
         EMH_SET(new_bucket);
-
-        //clear kickout bit
-        //EMH_CLS(kbucket);
         return kbucket;
     }
 
@@ -1643,8 +1578,8 @@ private:
 ** put new key in its main position; otherwise (colliding bucket is in its main
 ** position), new key goes to an empty position. ***/
 
-    template<typename Key=KeyT>
-    size_type find_or_allocate(const Key& key, bool& isempty)
+//    template<typename Key=KeyT>
+    size_type find_or_allocate(const KeyT& key, bool& isempty)
     {
         const auto bucket = hash_key(key) & _mask;
         const auto& bucket_key = EMH_KEY(_pairs, bucket);
@@ -1699,28 +1634,36 @@ private:
     // key is not in this map. Find a place to put it.
     size_type find_empty_bucket(const size_type bucket_from, const size_t main_bucket)
     {
-#if 1 || __arm64__ || __aarch64__
+#ifdef EMH_ALIGN64 // only works 64bit
         const auto boset  = bucket_from % MASK_BIT;
         auto* const align = _bitmask + bucket_from / MASK_BIT;
         const auto bmask  = ((size_t)align[1] << (MASK_BIT - boset)) | (align[0] >> boset);
-#else
-        const auto boset  = bucket_from % 8;
-        auto* const align = (uint8_t*)_bitmask + bucket_from / 8;
-        const auto bmask  = *(size_t*)(align) >> boset;
-#endif
         if (EMH_LIKELY(bmask != 0))
             return bucket_from + CTZ(bmask);
+#elif EMH_ITER_SAFE
+        const auto boset = bucket_from % 8;
+        auto* const start = (uint8_t*)_bitmask + bucket_from / 8;
+        size_t bmask; memcpy(&bmask, start + 0, sizeof(bmask)); bmask >>= boset;// bmask |= ((size_t)start[8] << (SIZE_BIT - boset));
+        if (EMH_LIKELY(bmask != 0))
+            return bucket_from + CTZ(bmask);
+#else
+        const auto boset  = main_bucket % 8;
+        auto* const align = (uint8_t*)_bitmask + main_bucket / 8;
+        const size_t bmask  = (*(size_t*)(align) >> boset);// & 0xF0F0F0F0FF0FF0FFull;//
+        if (EMH_LIKELY(bmask != 0))
+            return main_bucket + CTZ(bmask);
+#endif
 
         const auto qmask = _mask / SIZE_BIT;
         if (1) {
-            const auto step = (bucket_from - SIZE_BIT / 2) & qmask;
+            const auto step = (main_bucket - SIZE_BIT / 4) & qmask;
             const auto bmask3 = *((size_t*)_bitmask + step);
             if (bmask3 != 0)
                 return step * SIZE_BIT + CTZ(bmask3);
         }
 
         auto& _last = EMH_BUCKET(_pairs, _num_buckets);
-        for (; ; ) { //2.4.7
+        for (; ;) {
             const auto bmask2 = *((size_t*)_bitmask + _last);
             if (bmask2 != 0)
                 return _last * SIZE_BIT + CTZ(bmask2);
@@ -1733,6 +1676,38 @@ private:
             }
             _last = (_last + 1) & qmask;
         }
+
+        return 0;
+    }
+
+    // key is not in this map. Find a place to put it.
+    size_type find_unique_empty(const size_type bucket_from, const size_t main_bucket)
+    {
+#ifdef EMH_ALIGN64
+        const auto boset  = bucket_from % MASK_BIT;
+        auto* const align = _bitmask + bucket_from / MASK_BIT;
+        const auto bmask  = ((size_t)align[1] << (MASK_BIT - boset)) | (align[0] >> boset);
+        static_assert(sizeof(size_t) > 4);
+#elif EMH_ITER_SAFE
+        const auto boset = bucket_from % 8;
+        auto* const start = (uint8_t*)_bitmask + bucket_from / 8;
+        size_t bmask; memcpy(&bmask, start + 0, sizeof(bmask)); bmask >>= boset;
+#else
+        const auto boset  = bucket_from % 8;
+        auto* const align = (uint8_t*)_bitmask + bucket_from / 8;
+        const auto bmask  = (*(size_t*)(align) >> boset); //maybe not aligned and warning
+#endif
+        if (EMH_LIKELY(bmask != 0))
+            return bucket_from + CTZ(bmask);
+
+        const auto qmask = _mask / SIZE_BIT;
+        for (auto last = (bucket_from + _mask) & qmask; ;) {
+            const auto bmask2 = *((size_t*)_bitmask + last);// & 0xF0F0F0F0FF0FF0FFull;
+            if (EMH_LIKELY(bmask2 != 0))
+                return last * SIZE_BIT + CTZ(bmask2);
+            last = (last + 1) & qmask;
+        }
+
         return 0;
     }
 
@@ -1780,16 +1755,17 @@ private:
             next_bucket = find_last_bucket(next_bucket);
 
         //find a new empty and link it to tail
-        return EMH_BUCKET(_pairs, next_bucket) = find_empty_bucket(next_bucket, bucket);
+        return EMH_BUCKET(_pairs, next_bucket) = find_unique_empty(next_bucket, bucket);
     }
 
+#if EMH_INT_HASH
     static constexpr uint64_t KC = UINT64_C(11400714819323198485);
     static inline uint64_t hash64(uint64_t key)
     {
-#if __SIZEOF_INT128__ && EMH_FIBONACCI_HASH == 1
+#if __SIZEOF_INT128__ && EMH_INT_HASH == 1
         __uint128_t r = key; r *= KC;
         return (uint64_t)(r >> 64) + (uint64_t)r;
-#elif EMH_FIBONACCI_HASH == 2
+#elif EMH_INT_HASH == 2
         //MurmurHash3Mixer
         uint64_t h = key;
         h ^= h >> 33;
@@ -1798,18 +1774,20 @@ private:
         h *= 0xc4ceb9fe1a85ec53;
         h ^= h >> 33;
         return h;
-#elif _WIN64 && EMH_FIBONACCI_HASH == 1
+#elif _WIN64 && EMH_INT_HASH == 1
         uint64_t high;
         return _umul128(key, KC, &high) + high;
-#elif EMH_FIBONACCI_HASH == 3
+#elif EMH_INT_HASH == 3
         auto ror  = (key >> 32) | (key << 32);
         auto low  = key * 0xA24BAED4963EE407ull;
         auto high = ror * 0x9FB21C651E98DF25ull;
         auto mix  = low + high;
         return mix;
-#elif EMH_FIBONACCI_HASH == 1
+#elif EMH_INT_HASH == 1
         uint64_t r = key * UINT64_C(0xca4bcaa75ec3f625);
         return (r >> 32) + r;
+#elif EMH_WYHASH64
+        return wyhash64(key, KC);
 #else
         uint64_t x = key;
         x = (x ^ (x >> 30)) * UINT64_C(0xbf58476d1ce4e5b9);
@@ -1818,16 +1796,15 @@ private:
         return x;
 #endif
     }
+#endif
 
     template<typename UType, typename std::enable_if<std::is_integral<UType>::value, size_type>::type = 0>
     inline size_type hash_key(const UType key) const
     {
-#ifdef EMH_FIBONACCI_HASH
+#if EMH_INT_HASH
         return hash64(key);
 #elif EMH_IDENTITY_HASH
         return key + (key >> (sizeof(UType) * 4));
-#elif EMH_WYHASH64
-        return wyhash64(key, KC);
 #else
         return (size_type)_hasher(key);
 #endif
@@ -1836,7 +1813,7 @@ private:
     template<typename UType, typename std::enable_if<std::is_same<UType, std::string>::value, size_type>::type = 0>
     inline size_type hash_key(const UType& key) const
     {
-#ifdef WYHASH_LITTLE_ENDIAN
+#if WYHASH_LITTLE_ENDIAN
         return wyhash(key.data(), key.size(), key.size());
 #else
         return (size_type)_hasher(key);
@@ -1846,11 +1823,7 @@ private:
     template<typename UType, typename std::enable_if<!std::is_integral<UType>::value && !std::is_same<UType, std::string>::value, size_type>::type = 0>
     inline size_type hash_key(const UType& key) const
     {
-#ifdef EMH_FIBONACCI_HASH
-        return _hasher(key) * KC;
-#else
         return (size_type)_hasher(key);
-#endif
     }
 
 private:
@@ -1870,15 +1843,15 @@ private:
     static constexpr uint32_t SIZE_BIT = sizeof(size_t) * 8;
     static constexpr uint32_t EPACK_SIZE = sizeof(PairT) >= sizeof(size_t) == 0 ? 1 : 2; // > 1
 };
-} // namespace emhash
+}
+// namespace emhash7
 #if __cplusplus >= 201103L
-//template <class Key, class Val> using emhash7 = emhash7::HashMap<Key, Val, std::hash<Key>, std::equal_to<Key>>;
+//template <class Key, class Val> using ehmap7 = emhash7::HashMap<Key, Val, std::hash<Key>, std::equal_to<Key>>;
 #endif
 
 //TODO
 //2. improve rehash and find miss performance(reduce peak memory)
 //3. dump or Serialization interface
 //4. node hash map support
-//5. support load_factor > 1.0
-//6. add grow ration
-//8. ... https://godbolt.org/
+//5. load_factor > 1.0 && add grow ration
+//... https://godbolt.org/
diff --git a/benchmarks/external/emhash/wyhash.h b/benchmarks/external/emhash/wyhash.h
deleted file mode 100644
index b89352a7..00000000
--- a/benchmarks/external/emhash/wyhash.h
+++ /dev/null
@@ -1,272 +0,0 @@
-// This is free and unencumbered software released into the public domain under The Unlicense (http://unlicense.org/)
-// main repo: https://github.com/wangyi-fudan/wyhash
-// author: 王一 Wang Yi <[email protected]>
-// contributors: Reini Urban, Dietrich Epp, Joshua Haberman, Tommy Ettinger, Daniel Lemire, Otmar Ertl, cocowalla, leo-yuriev, Diego Barrios Romero, paulie-g, dumblob, Yann Collet, ivte-ms, hyb, James Z.M. Gao, easyaspi314 (Devin), TheOneric
-
-/* quick example:
-   uint64_t _wyp[4];
-   make_secret(time(NULL),_wyp);
-   string s="fjsakfdsjkf";
-   uint64_t hash=wyhash(s.c_str(), s.size(), 0, _wyp);
-*/
-
-#ifndef wyhash_final_version_3
-#define wyhash_final_version_3
-
-#ifndef WYHASH_CONDOM
-//protections that produce different results:
-//1: normal valid behavior
-//2: extra protection against entropy loss (probability=2^-63), aka. "blind multiplication"
-#define WYHASH_CONDOM 1
-#endif
-
-#ifndef WYHASH_32BIT_MUM
-//0: normal version, slow on 32 bit systems
-//1: faster on 32 bit systems but produces different results, incompatible with wy2u0k function
-#define WYHASH_32BIT_MUM 0
-#endif
-
-//includes
-#include <stdint.h>
-#include <string.h>
-#if defined(_MSC_VER) && defined(_M_X64)
-  #include <intrin.h>
-  #pragma intrinsic(_umul128)
-#endif
-
-//likely and unlikely macros
-#if defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__clang__)
-  #define _likely_(x)  __builtin_expect(x,1)
-  #define _unlikely_(x)  __builtin_expect(x,0)
-#else
-  #define _likely_(x) (x)
-  #define _unlikely_(x) (x)
-#endif
-
-//128bit multiply function
-static inline uint64_t _wyrot(uint64_t x) { return (x>>32)|(x<<32); }
-static inline void _wymum(uint64_t *A, uint64_t *B){
-#if(WYHASH_32BIT_MUM)
-  uint64_t hh=(*A>>32)*(*B>>32), hl=(*A>>32)*(uint32_t)*B, lh=(uint32_t)*A*(*B>>32), ll=(uint64_t)(uint32_t)*A*(uint32_t)*B;
-  #if(WYHASH_CONDOM>1)
-  *A^=_wyrot(hl)^hh; *B^=_wyrot(lh)^ll;
-  #else
-  *A=_wyrot(hl)^hh; *B=_wyrot(lh)^ll;
-  #endif
-#elif defined(__SIZEOF_INT128__)
-  __uint128_t r=*A; r*=*B;
-  #if(WYHASH_CONDOM>1)
-  *A^=(uint64_t)r; *B^=(uint64_t)(r>>64);
-  #else
-  *A=(uint64_t)r; *B=(uint64_t)(r>>64);
-  #endif
-#elif defined(_MSC_VER) && defined(_M_X64)
-  #if(WYHASH_CONDOM>1)
-  uint64_t  a,  b;
-  a=_umul128(*A,*B,&b);
-  *A^=a;  *B^=b;
-  #else
-  *A=_umul128(*A,*B,B);
-  #endif
-#else
-  uint64_t ha=*A>>32, hb=*B>>32, la=(uint32_t)*A, lb=(uint32_t)*B, hi, lo;
-  uint64_t rh=ha*hb, rm0=ha*lb, rm1=hb*la, rl=la*lb, t=rl+(rm0<<32), c=t<rl;
-  lo=t+(rm1<<32); c+=lo<t; hi=rh+(rm0>>32)+(rm1>>32)+c;
-  #if(WYHASH_CONDOM>1)
-  *A^=lo;  *B^=hi;
-  #else
-  *A=lo;  *B=hi;
-  #endif
-#endif
-}
-
-//multiply and xor mix function, aka MUM
-static inline uint64_t _wymix(uint64_t A, uint64_t B){ _wymum(&A,&B); return A^B; }
-
-//endian macros
-#ifndef WYHASH_LITTLE_ENDIAN
-  #if defined(_WIN32) || defined(__LITTLE_ENDIAN__) || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
-    #define WYHASH_LITTLE_ENDIAN 1
-  #elif defined(__BIG_ENDIAN__) || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
-    #define WYHASH_LITTLE_ENDIAN 0
-  #else
-    #warning could not determine endianness! Falling back to little endian.
-    #define WYHASH_LITTLE_ENDIAN 1
-  #endif
-#endif
-
-//read functions
-#if (WYHASH_LITTLE_ENDIAN)
-static inline uint64_t _wyr8(const uint8_t *p) { uint64_t v; memcpy(&v, p, 8); return v;}
-static inline uint64_t _wyr4(const uint8_t *p) { uint32_t v; memcpy(&v, p, 4); return v;}
-#elif defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__clang__)
-static inline uint64_t _wyr8(const uint8_t *p) { uint64_t v; memcpy(&v, p, 8); return __builtin_bswap64(v);}
-static inline uint64_t _wyr4(const uint8_t *p) { uint32_t v; memcpy(&v, p, 4); return __builtin_bswap32(v);}
-#elif defined(_MSC_VER)
-static inline uint64_t _wyr8(const uint8_t *p) { uint64_t v; memcpy(&v, p, 8); return _byteswap_uint64(v);}
-static inline uint64_t _wyr4(const uint8_t *p) { uint32_t v; memcpy(&v, p, 4); return _byteswap_ulong(v);}
-#else
-static inline uint64_t _wyr8(const uint8_t *p) {
-  uint64_t v; memcpy(&v, p, 8);
-  return (((v >> 56) & 0xff)| ((v >> 40) & 0xff00)| ((v >> 24) & 0xff0000)| ((v >>  8) & 0xff000000)| ((v <<  8) & 0xff00000000)| ((v << 24) & 0xff0000000000)| ((v << 40) & 0xff000000000000)| ((v << 56) & 0xff00000000000000));
-}
-static inline uint64_t _wyr4(const uint8_t *p) {
-  uint32_t v; memcpy(&v, p, 4);
-  return (((v >> 24) & 0xff)| ((v >>  8) & 0xff00)| ((v <<  8) & 0xff0000)| ((v << 24) & 0xff000000));
-}
-#endif
-static inline uint64_t _wyr3(const uint8_t *p, size_t k) { return (((uint64_t)p[0])<<16)|(((uint64_t)p[k>>1])<<8)|p[k-1];}
-//wyhash main function
-static inline uint64_t wyhash(const void *key, size_t len, uint64_t seed, const uint64_t *secret){
-  const uint8_t *p=(const uint8_t *)key; seed^=*secret; uint64_t a, b;
-  if(_likely_(len<=16)){
-    if(_likely_(len>=4)){ a=(_wyr4(p)<<32)|_wyr4(p+((len>>3)<<2)); b=(_wyr4(p+len-4)<<32)|_wyr4(p+len-4-((len>>3)<<2)); }
-    else if(_likely_(len>0)){ a=_wyr3(p,len); b=0;}
-    else a=b=0;
-  }
-  else{
-    size_t i=len;
-    if(_unlikely_(i>48)){
-      uint64_t see1=seed, see2=seed;
-      do{
-        seed=_wymix(_wyr8(p)^secret[1],_wyr8(p+8)^seed);
-        see1=_wymix(_wyr8(p+16)^secret[2],_wyr8(p+24)^see1);
-        see2=_wymix(_wyr8(p+32)^secret[3],_wyr8(p+40)^see2);
-        p+=48; i-=48;
-      }while(_likely_(i>48));
-      seed^=see1^see2;
-    }
-    while(_unlikely_(i>16)){  seed=_wymix(_wyr8(p)^secret[1],_wyr8(p+8)^seed);  i-=16; p+=16;  }
-    a=_wyr8(p+i-16);  b=_wyr8(p+i-8);
-  }
-  return _wymix(secret[1]^len,_wymix(a^secret[1],b^seed));
-}
-
-//the default secret parameters
-static const uint64_t _wyp[4] = {0xa0761d6478bd642full, 0xe7037ed1a0b428dbull, 0x8ebc6af09c88c6e3ull, 0x589965cc75374cc3ull};
-
-static inline uint64_t wyhash(const void *key, size_t len, uint64_t seed) { return wyhash(key, len, seed, _wyp); }
-
-//a useful 64bit-64bit mix function to produce deterministic pseudo random numbers that can pass BigCrush and PractRand
-static inline uint64_t wyhash64(uint64_t A, uint64_t B){ A^=0xa0761d6478bd642full; B^=0xe7037ed1a0b428dbull; _wymum(&A,&B); return _wymix(A^0xa0761d6478bd642full,B^0xe7037ed1a0b428dbull);}
-
-//The wyrand PRNG that pass BigCrush and PractRand
-static inline uint64_t wyrand(uint64_t *seed){ *seed+=0xa0761d6478bd642full; return _wymix(*seed,*seed^0xe7037ed1a0b428dbull);}
-
-//convert any 64 bit pseudo random numbers to uniform distribution [0,1). It can be combined with wyrand, wyhash64 or wyhash.
-static inline double wy2u01(uint64_t r){ const double _wynorm=1.0/(1ull<<52); return (r>>12)*_wynorm;}
-
-//convert any 64 bit pseudo random numbers to APPROXIMATE Gaussian distribution. It can be combined with wyrand, wyhash64 or wyhash.
-static inline double wy2gau(uint64_t r){ const double _wynorm=1.0/(1ull<<20); return ((r&0x1fffff)+((r>>21)&0x1fffff)+((r>>42)&0x1fffff))*_wynorm-3.0;}
-
-#if(!WYHASH_32BIT_MUM)
-//fast range integer random number generation on [0,k) credit to Daniel Lemire. May not work when WYHASH_32BIT_MUM=1. It can be combined with wyrand, wyhash64 or wyhash.
-static inline uint64_t wy2u0k(uint64_t r, uint64_t k){ _wymum(&r,&k); return k; }
-#endif
-
-//make your own secret
-static inline void make_secret(uint64_t seed, uint64_t *secret){
-  uint8_t c[] = {15, 23, 27, 29, 30, 39, 43, 45, 46, 51, 53, 54, 57, 58, 60, 71, 75, 77, 78, 83, 85, 86, 89, 90, 92, 99, 101, 102, 105, 106, 108, 113, 114, 116, 120, 135, 139, 141, 142, 147, 149, 150, 153, 154, 156, 163, 165, 166, 169, 170, 172, 177, 178, 180, 184, 195, 197, 198, 201, 202, 204, 209, 210, 212, 216, 225, 226, 228, 232, 240 };
-  for(size_t i=0;i<4;i++){
-    uint8_t ok;
-    do{
-      ok=1; secret[i]=0;
-      for(size_t j=0;j<64;j+=8) secret[i]|=((uint64_t)c[wyrand(&seed)%sizeof(c)])<<j;
-      if(secret[i]%2==0){ ok=0; continue; }
-      for(size_t j=0;j<i;j++) {
-#if defined(__GNUC__) || defined(__INTEL_COMPILER) || defined(__clang__)
-        if(__builtin_popcountll(secret[j]^secret[i])!=32){ ok=0; break; }
-#elif defined(_MSC_VER) && defined(_M_X64)
-        if(_mm_popcnt_u64(secret[j]^secret[i])!=32){ ok=0; break; }
-#else
-        //manual popcount
-        uint64_t x = secret[j]^secret[i];
-        x -= (x >> 1) & 0x5555555555555555;
-        x = (x & 0x3333333333333333) + ((x >> 2) & 0x3333333333333333);
-        x = (x + (x >> 4)) & 0x0f0f0f0f0f0f0f0f;
-        x = (x * 0x0101010101010101) >> 56;
-        if(x!=32){ ok=0; break; }
-#endif
-      }
-    }while(!ok);
-  }
-}
-
-/*  This is world's fastest hash map: 2x faster than bytell_hash_map.
-    It does not store the keys, but only the hash/signature of keys.
-    First we use pos=hash1(key) to approximately locate the bucket.
-    Then we search signature=hash2(key) from pos linearly.
-    If we find a bucket with matched signature we report the bucket
-    Or if we meet a bucket whose signature=0, we report a new position to insert
-    The signature collision probability is very low as we usually searched N~10 buckets.
-    By combining hash1 and hash2, we acturally have 128 bit anti-collision strength.
-    hash1 and hash2 can be the same function, resulting lower collision resistance but faster.
-    The signature is 64 bit, but can be modified to 32 bit if necessary for save space.
-    The above two can be activated by define WYHASHMAP_WEAK_SMALL_FAST
-    simple examples:
-    const	size_t	size=213432;
-    vector<wyhashmap_t>	idx(size);	//	allocate the index of fixed size. idx MUST be zeroed.
-    vector<value_class>	value(size);	//	we only care about the index, user should maintain his own value vectors.
-    string  key="dhskfhdsj"	//	the object to be inserted into idx
-    size_t	pos=wyhashmap(idx.data(), idx.size(), key.c_str(), key.size(), 1);	//	get the position and insert
-    if(pos<size)	value[pos]++;	//	we process the vallue
-    else	cerr<<"map is full\n";
-    pos=wyhashmap(idx.data(), idx.size(), key.c_str(), key.size(), 0);	// just lookup by setting insert=0
-    if(pos<size)	value[pos]++;	//	we process the vallue
-    else	cerr<<"the key does not exist\n";
-*/
-/*
-#ifdef	WYHASHMAP_WEAK_SMALL_FAST	// for small hashmaps whose size < 2^24 and acceptable collision
-typedef	uint32_t	wyhashmap_t;
-#else
-typedef	uint64_t	wyhashmap_t;
-#endif
-
-static	inline	size_t	wyhashmap(wyhashmap_t	*idx,	size_t	idx_size,	const	void *key, size_t	key_size,	uint8_t	insert, uint64_t *secret){
-	size_t	i=1;	uint64_t	h2;	wyhashmap_t	sig;
-	do{	sig=h2=wyhash(key,key_size,i,secret);	i++;	}while(_unlikely_(!sig));
-#ifdef	WYHASHMAP_WEAK_SMALL_FAST
-	size_t	i0=wy2u0k(h2,idx_size);
-#else
-	size_t	i0=wy2u0k(wyhash(key,key_size,0,secret),idx_size);
-#endif
-	for(i=i0;	i<idx_size&&idx[i]&&idx[i]!=sig;	i++);
-	if(_unlikely_(i==idx_size)){
-		for(i=0;	i<i0&&idx[i]&&idx[i]!=sig;  i++);
-		if(i==i0)	return	idx_size;
-	}
-	if(!idx[i]){
-		if(insert)	idx[i]=sig;
-		else	return	idx_size;
-	}
-	return	i;
-}
-*/
-#endif
-
-/* The Unlicense
-This is free and unencumbered software released into the public domain.
-
-Anyone is free to copy, modify, publish, use, compile, sell, or
-distribute this software, either in source code form or as a compiled
-binary, for any purpose, commercial or non-commercial, and by any
-means.
-
-In jurisdictions that recognize copyright laws, the author or authors
-of this software dedicate any and all copyright interest in the
-software to the public domain. We make this dedication for the benefit
-of the public at large and to the detriment of our heirs and
-successors. We intend this dedication to be an overt act of
-relinquishment in perpetuity of all present and future rights to this
-software under copyright law.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
-EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
-MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
-IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
-OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
-ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
-OTHER DEALINGS IN THE SOFTWARE.
-
-For more information, please refer to <http://unlicense.org/>
-*/
diff --git a/benchmarks/external/update.sh b/benchmarks/external/update.sh
index c9681485..45b472b1 100644
--- a/benchmarks/external/update.sh
+++ b/benchmarks/external/update.sh
@@ -18,7 +18,7 @@ wget "$tsl_r/robin_growth_policy.h" -O "tsl/robin_growth_policy.h"
 wget "$tsl_r/robin_hash.h"          -O "tsl/robin_hash.h"
 wget "$tsl_r/robin_map.h"           -O "tsl/robin_map.h"
 
-wget "$ktprime/thirdparty/wyhash.h" -O "emhash/wyhash.h"
+#wget "$ktprime/thirdparty/wyhash.h" -O "emhash/wyhash.h"
 wget "$ktprime/hash_table7.hpp" -O "emhash/hash_table7.hpp"
 
 #wget "$tsl_h/hopscotch_growth_policy.h" -O "tsl/hopscotch_growth_policy.h"