Upgrade to fmt 7.1.3

diff --git a/LICENSE.adoc b/LICENSE.adoc
index fa1f32095c6533341294e1743dd4416649cb469f..9ccdb0717ee3b27cb01b0d27a12f9b87eae866b1 100644 (file)
--- a/LICENSE.adoc
+++ b/LICENSE.adoc
@@ -451,7 +451,7 @@ SOFTWARE.
 src/third_party/fmt/*.h and src/third_party/format.cpp
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-This is a subset of https://fmt.dev[fmt] 7.0.3 with the following license:
+This is a subset of https://fmt.dev[fmt] 7.1.3 with the following license:
 
 -------------------------------------------------------------------------------
 Formatting library for C++

diff --git a/src/third_party/fmt/core.h b/src/third_party/fmt/core.h
index 031bf869e13798523d6928810d4450434c742d61..0a81e0ccd953aa4775289ff8ac73b6cbc6e3f16f 100644 (file)
--- a/src/third_party/fmt/core.h
+++ b/src/third_party/fmt/core.h
@@ -18,7 +18,7 @@
 #include <vector>
 
 // The fmt library version in the form major * 10000 + minor * 100 + patch.
-#define FMT_VERSION 70003
+#define FMT_VERSION 70103
 
 #ifdef __clang__
 #  define FMT_CLANG_VERSION (__clang_major__ * 100 + __clang_minor__)
@@ -57,6 +57,7 @@
 #  define FMT_MSC_VER 0
 #  define FMT_SUPPRESS_MSC_WARNING(n)
 #endif
+
 #ifdef __has_feature
 #  define FMT_HAS_FEATURE(x) __has_feature(x)
 #else
@@ -64,7 +65,7 @@
 #endif
 
 #if defined(__has_include) && !defined(__INTELLISENSE__) && \
-    !(FMT_ICC_VERSION && FMT_ICC_VERSION < 1600)
+    (!FMT_ICC_VERSION || FMT_ICC_VERSION >= 1600)
 #  define FMT_HAS_INCLUDE(x) __has_include(x)
 #else
 #  define FMT_HAS_INCLUDE(x) 0
@@ -152,7 +153,7 @@
 #  if FMT_HAS_CPP14_ATTRIBUTE(deprecated) || FMT_MSC_VER >= 1900
 #    define FMT_DEPRECATED [[deprecated]]
 #  else
-#    if defined(__GNUC__) || defined(__clang__)
+#    if (defined(__GNUC__) && !defined(__LCC__)) || defined(__clang__)
 #      define FMT_DEPRECATED __attribute__((deprecated))
 #    elif FMT_MSC_VER
 #      define FMT_DEPRECATED __declspec(deprecated)
@@ -177,9 +178,17 @@
 #  endif
 #endif
 
-#ifndef FMT_BEGIN_NAMESPACE
+#ifndef FMT_USE_INLINE_NAMESPACES
 #  if FMT_HAS_FEATURE(cxx_inline_namespaces) || FMT_GCC_VERSION >= 404 || \
-      FMT_MSC_VER >= 1900
+      (FMT_MSC_VER >= 1900 && !_MANAGED)
+#    define FMT_USE_INLINE_NAMESPACES 1
+#  else
+#    define FMT_USE_INLINE_NAMESPACES 0
+#  endif
+#endif
+
+#ifndef FMT_BEGIN_NAMESPACE
+#  if FMT_USE_INLINE_NAMESPACES
 #    define FMT_INLINE_NAMESPACE inline namespace
 #    define FMT_END_NAMESPACE \
       }                       \
@@ -269,8 +278,7 @@ struct monostate {};
 
 namespace detail {
 
-// A helper function to suppress bogus "conditional expression is constant"
-// warnings.
+// A helper function to suppress "conditional expression is constant" warnings.
 template <typename T> constexpr T const_check(T value) { return value; }
 
 FMT_NORETURN FMT_API void assert_fail(const char* file, int line,
@@ -299,7 +307,8 @@ template <typename T> struct std_string_view {};
 
 #ifdef FMT_USE_INT128
 // Do nothing.
-#elif defined(__SIZEOF_INT128__) && !FMT_NVCC && !(FMT_CLANG_VERSION && FMT_MSC_VER)
+#elif defined(__SIZEOF_INT128__) && !FMT_NVCC && \
+    !(FMT_CLANG_VERSION && FMT_MSC_VER)
 #  define FMT_USE_INT128 1
 using int128_t = __int128_t;
 using uint128_t = __uint128_t;
@@ -506,6 +515,18 @@ template <typename S> struct char_t_impl<S, enable_if_t<is_string<S>::value>> {
   using type = typename result::value_type;
 };
 
+// Reports a compile-time error if S is not a valid format string.
+template <typename..., typename S, FMT_ENABLE_IF(!is_compile_string<S>::value)>
+FMT_INLINE void check_format_string(const S&) {
+#ifdef FMT_ENFORCE_COMPILE_STRING
+  static_assert(is_compile_string<S>::value,
+                "FMT_ENFORCE_COMPILE_STRING requires all format strings to use "
+                "FMT_STRING.");
+#endif
+}
+template <typename..., typename S, FMT_ENABLE_IF(is_compile_string<S>::value)>
+void check_format_string(S);
+
 struct error_handler {
   constexpr error_handler() = default;
   constexpr error_handler(const error_handler&) = default;
@@ -545,8 +566,9 @@ class basic_format_parse_context : private ErrorHandler {
   using iterator = typename basic_string_view<Char>::iterator;
 
   explicit constexpr basic_format_parse_context(
-      basic_string_view<Char> format_str, ErrorHandler eh = {})
-      : ErrorHandler(eh), format_str_(format_str), next_arg_id_(0) {}
+      basic_string_view<Char> format_str, ErrorHandler eh = {},
+      int next_arg_id = 0)
+      : ErrorHandler(eh), format_str_(format_str), next_arg_id_(next_arg_id) {}
 
   /**
     Returns an iterator to the beginning of the format string range being
@@ -616,8 +638,24 @@ template <typename T, typename Context>
 using has_formatter =
     std::is_constructible<typename Context::template formatter_type<T>>;
 
+// Checks whether T is a container with contiguous storage.
+template <typename T> struct is_contiguous : std::false_type {};
+template <typename Char>
+struct is_contiguous<std::basic_string<Char>> : std::true_type {};
+
 namespace detail {
 
+// Extracts a reference to the container from back_insert_iterator.
+template <typename Container>
+inline Container& get_container(std::back_insert_iterator<Container> it) {
+  using bi_iterator = std::back_insert_iterator<Container>;
+  struct accessor : bi_iterator {
+    accessor(bi_iterator iter) : bi_iterator(iter) {}
+    using bi_iterator::container;
+  };
+  return *accessor(it).container;
+}
+
 /**
   \rst
   A contiguous memory buffer with an optional growing ability. It is an internal
@@ -640,6 +678,8 @@ template <typename T> class buffer {
         size_(sz),
         capacity_(cap) {}
 
+  ~buffer() = default;
+
   /** Sets the buffer data and capacity. */
   void set(T* buf_data, size_t buf_capacity) FMT_NOEXCEPT {
     ptr_ = buf_data;
@@ -655,7 +695,6 @@ template <typename T> class buffer {
 
   buffer(const buffer&) = delete;
   void operator=(const buffer&) = delete;
-  virtual ~buffer() = default;
 
   T* begin() FMT_NOEXCEPT { return ptr_; }
   T* end() FMT_NOEXCEPT { return ptr_ + size_; }
@@ -675,24 +714,26 @@ template <typename T> class buffer {
   /** Returns a pointer to the buffer data. */
   const T* data() const FMT_NOEXCEPT { return ptr_; }
 
-  /**
-    Resizes the buffer. If T is a POD type new elements may not be initialized.
-   */
-  void resize(size_t new_size) {
-    reserve(new_size);
-    size_ = new_size;
-  }
-
   /** Clears this buffer. */
   void clear() { size_ = 0; }
 
-  /** Reserves space to store at least *capacity* elements. */
-  void reserve(size_t new_capacity) {
+  // Tries resizing the buffer to contain *count* elements. If T is a POD type
+  // the new elements may not be initialized.
+  void try_resize(size_t count) {
+    try_reserve(count);
+    size_ = count <= capacity_ ? count : capacity_;
+  }
+
+  // Tries increasing the buffer capacity to *new_capacity*. It can increase the
+  // capacity by a smaller amount than requested but guarantees there is space
+  // for at least one additional element either by increasing the capacity or by
+  // flushing the buffer if it is full.
+  void try_reserve(size_t new_capacity) {
     if (new_capacity > capacity_) grow(new_capacity);
   }
 
   void push_back(const T& value) {
-    reserve(size_ + 1);
+    try_reserve(size_ + 1);
     ptr_[size_++] = value;
   }
 
@@ -705,32 +746,150 @@ template <typename T> class buffer {
   }
 };
 
-// A container-backed buffer.
+struct buffer_traits {
+  explicit buffer_traits(size_t) {}
+  size_t count() const { return 0; }
+  size_t limit(size_t size) { return size; }
+};
+
+class fixed_buffer_traits {
+ private:
+  size_t count_ = 0;
+  size_t limit_;
+
+ public:
+  explicit fixed_buffer_traits(size_t limit) : limit_(limit) {}
+  size_t count() const { return count_; }
+  size_t limit(size_t size) {
+    size_t n = limit_ > count_ ? limit_ - count_ : 0;
+    count_ += size;
+    return size < n ? size : n;
+  }
+};
+
+// A buffer that writes to an output iterator when flushed.
+template <typename OutputIt, typename T, typename Traits = buffer_traits>
+class iterator_buffer final : public Traits, public buffer<T> {
+ private:
+  OutputIt out_;
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+
+ protected:
+  void grow(size_t) final FMT_OVERRIDE {
+    if (this->size() == buffer_size) flush();
+  }
+  void flush();
+
+ public:
+  explicit iterator_buffer(OutputIt out, size_t n = buffer_size)
+      : Traits(n),
+        buffer<T>(data_, 0, buffer_size),
+        out_(out) {}
+  ~iterator_buffer() { flush(); }
+
+  OutputIt out() {
+    flush();
+    return out_;
+  }
+  size_t count() const { return Traits::count() + this->size(); }
+};
+
+template <typename T> class iterator_buffer<T*, T> final : public buffer<T> {
+ protected:
+  void grow(size_t) final FMT_OVERRIDE {}
+
+ public:
+  explicit iterator_buffer(T* out, size_t = 0) : buffer<T>(out, 0, ~size_t()) {}
+
+  T* out() { return &*this->end(); }
+};
+
+// A buffer that writes to a container with the contiguous storage.
 template <typename Container>
-class container_buffer : public buffer<typename Container::value_type> {
+class iterator_buffer<std::back_insert_iterator<Container>,
+                      enable_if_t<is_contiguous<Container>::value,
+                                  typename Container::value_type>>
+    final : public buffer<typename Container::value_type> {
  private:
   Container& container_;
 
  protected:
-  void grow(size_t capacity) FMT_OVERRIDE {
+  void grow(size_t capacity) final FMT_OVERRIDE {
     container_.resize(capacity);
     this->set(&container_[0], capacity);
   }
 
  public:
-  explicit container_buffer(Container& c)
+  explicit iterator_buffer(Container& c)
       : buffer<typename Container::value_type>(c.size()), container_(c) {}
+  explicit iterator_buffer(std::back_insert_iterator<Container> out, size_t = 0)
+      : iterator_buffer(get_container(out)) {}
+  std::back_insert_iterator<Container> out() {
+    return std::back_inserter(container_);
+  }
 };
 
-// Extracts a reference to the container from back_insert_iterator.
-template <typename Container>
-inline Container& get_container(std::back_insert_iterator<Container> it) {
-  using bi_iterator = std::back_insert_iterator<Container>;
-  struct accessor : bi_iterator {
-    accessor(bi_iterator iter) : bi_iterator(iter) {}
-    using bi_iterator::container;
-  };
-  return *accessor(it).container;
+// A buffer that counts the number of code units written discarding the output.
+template <typename T = char> class counting_buffer final : public buffer<T> {
+ private:
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+  size_t count_ = 0;
+
+ protected:
+  void grow(size_t) final FMT_OVERRIDE {
+    if (this->size() != buffer_size) return;
+    count_ += this->size();
+    this->clear();
+  }
+
+ public:
+  counting_buffer() : buffer<T>(data_, 0, buffer_size) {}
+
+  size_t count() { return count_ + this->size(); }
+};
+
+// An output iterator that appends to the buffer.
+// It is used to reduce symbol sizes for the common case.
+template <typename T>
+class buffer_appender : public std::back_insert_iterator<buffer<T>> {
+  using base = std::back_insert_iterator<buffer<T>>;
+
+ public:
+  explicit buffer_appender(buffer<T>& buf) : base(buf) {}
+  buffer_appender(base it) : base(it) {}
+
+  buffer_appender& operator++() {
+    base::operator++();
+    return *this;
+  }
+
+  buffer_appender operator++(int) {
+    buffer_appender tmp = *this;
+    ++*this;
+    return tmp;
+  }
+};
+
+// Maps an output iterator into a buffer.
+template <typename T, typename OutputIt>
+iterator_buffer<OutputIt, T> get_buffer(OutputIt);
+template <typename T> buffer<T>& get_buffer(buffer_appender<T>);
+
+template <typename OutputIt> OutputIt get_buffer_init(OutputIt out) {
+  return out;
+}
+template <typename T> buffer<T>& get_buffer_init(buffer_appender<T> out) {
+  return get_container(out);
+}
+
+template <typename Buffer>
+auto get_iterator(Buffer& buf) -> decltype(buf.out()) {
+  return buf.out();
+}
+template <typename T> buffer_appender<T> get_iterator(buffer<T>& buf) {
+  return buffer_appender<T>(buf);
 }
 
 template <typename T, typename Char = char, typename Enable = void>
@@ -759,7 +918,8 @@ template <typename Char> struct named_arg_info {
 template <typename T, typename Char, size_t NUM_ARGS, size_t NUM_NAMED_ARGS>
 struct arg_data {
   // args_[0].named_args points to named_args_ to avoid bloating format_args.
-  T args_[1 + (NUM_ARGS != 0 ? NUM_ARGS : 1)];
+  // +1 to workaround a bug in gcc 7.5 that causes duplicated-branches warning.
+  T args_[1 + (NUM_ARGS != 0 ? NUM_ARGS : +1)];
   named_arg_info<Char> named_args_[NUM_NAMED_ARGS];
 
   template <typename... U>
@@ -771,7 +931,8 @@ struct arg_data {
 
 template <typename T, typename Char, size_t NUM_ARGS>
 struct arg_data<T, Char, NUM_ARGS, 0> {
-  T args_[NUM_ARGS != 0 ? NUM_ARGS : 1];
+  // +1 to workaround a bug in gcc 7.5 that causes duplicated-branches warning.
+  T args_[NUM_ARGS != 0 ? NUM_ARGS : +1];
 
   template <typename... U>
   FMT_INLINE arg_data(const U&... init) : args_{init...} {}
@@ -959,6 +1120,8 @@ enum { long_short = sizeof(long) == sizeof(int) };
 using long_type = conditional_t<long_short, int, long long>;
 using ulong_type = conditional_t<long_short, unsigned, unsigned long long>;
 
+struct unformattable {};
+
 // Maps formatting arguments to core types.
 template <typename Context> struct arg_mapper {
   using char_type = typename Context::char_type;
@@ -1067,15 +1230,7 @@ template <typename Context> struct arg_mapper {
     return map(val.value);
   }
 
-  int map(...) {
-    constexpr bool formattable = sizeof(Context) == 0;
-    static_assert(
-        formattable,
-        "Cannot format argument. To make type T formattable provide a "
-        "formatter<T> specialization: "
-        "https://fmt.dev/latest/api.html#formatting-user-defined-types");
-    return 0;
-  }
+  unformattable map(...) { return {}; }
 };
 
 // A type constant after applying arg_mapper<Context>.
@@ -1199,15 +1354,25 @@ FMT_CONSTEXPR_DECL FMT_INLINE auto visit_format_arg(
   return vis(monostate());
 }
 
-// Checks whether T is a container with contiguous storage.
-template <typename T> struct is_contiguous : std::false_type {};
-template <typename Char>
-struct is_contiguous<std::basic_string<Char>> : std::true_type {};
-template <typename Char>
-struct is_contiguous<detail::buffer<Char>> : std::true_type {};
+template <typename T> struct formattable : std::false_type {};
 
 namespace detail {
 
+// A workaround for gcc 4.8 to make void_t work in a SFINAE context.
+template <typename... Ts> struct void_t_impl { using type = void; };
+template <typename... Ts>
+using void_t = typename detail::void_t_impl<Ts...>::type;
+
+template <typename It, typename T, typename Enable = void>
+struct is_output_iterator : std::false_type {};
+
+template <typename It, typename T>
+struct is_output_iterator<
+    It, T,
+    void_t<typename std::iterator_traits<It>::iterator_category,
+           decltype(*std::declval<It>() = std::declval<T>())>>
+    : std::true_type {};
+
 template <typename OutputIt>
 struct is_back_insert_iterator : std::false_type {};
 template <typename Container>
@@ -1219,6 +1384,9 @@ struct is_contiguous_back_insert_iterator : std::false_type {};
 template <typename Container>
 struct is_contiguous_back_insert_iterator<std::back_insert_iterator<Container>>
     : is_contiguous<Container> {};
+template <typename Char>
+struct is_contiguous_back_insert_iterator<buffer_appender<Char>>
+    : std::true_type {};
 
 // A type-erased reference to an std::locale to avoid heavy <locale> include.
 class locale_ref {
@@ -1250,13 +1418,24 @@ FMT_CONSTEXPR basic_format_arg<Context> make_arg(const T& value) {
   return arg;
 }
 
+template <typename T> int check(unformattable) {
+  static_assert(
+      formattable<T>(),
+      "Cannot format an argument. To make type T formattable provide a "
+      "formatter<T> specialization: https://fmt.dev/latest/api.html#udt");
+  return 0;
+}
+template <typename T, typename U> inline const U& check(const U& val) {
+  return val;
+}
+
 // The type template parameter is there to avoid an ODR violation when using
 // a fallback formatter in one translation unit and an implicit conversion in
 // another (not recommended).
 template <bool IS_PACKED, typename Context, type, typename T,
           FMT_ENABLE_IF(IS_PACKED)>
 inline value<Context> make_arg(const T& val) {
-  return arg_mapper<Context>().map(val);
+  return check<T>(arg_mapper<Context>().map(val));
 }
 
 template <bool IS_PACKED, typename Context, type, typename T,
@@ -1356,13 +1535,13 @@ template <typename OutputIt, typename Char> class basic_format_context {
 
 template <typename Char>
 using buffer_context =
-    basic_format_context<std::back_insert_iterator<detail::buffer<Char>>, Char>;
+    basic_format_context<detail::buffer_appender<Char>, Char>;
 using format_context = buffer_context<char>;
 using wformat_context = buffer_context<wchar_t>;
 
-// Workaround a bug in gcc: https://stackoverflow.com/q/62767544/471164.
+// Workaround an alias issue: https://stackoverflow.com/q/62767544/471164.
 #define FMT_BUFFER_CONTEXT(Char) \
-  basic_format_context<std::back_insert_iterator<detail::buffer<Char>>, Char>
+  basic_format_context<detail::buffer_appender<Char>, Char>
 
 /**
   \rst
@@ -1414,7 +1593,7 @@ class format_arg_store
 
 /**
   \rst
-  Constructs an `~fmt::format_arg_store` object that contains references to
+  Constructs a `~fmt::format_arg_store` object that contains references to
   arguments and can be implicitly converted to `~fmt::format_args`. `Context`
   can be omitted in which case it defaults to `~fmt::context`.
   See `~fmt::arg` for lifetime considerations.
@@ -1426,6 +1605,27 @@ inline format_arg_store<Context, Args...> make_format_args(
   return {args...};
 }
 
+/**
+  \rst
+  Constructs a `~fmt::format_arg_store` object that contains references
+  to arguments and can be implicitly converted to `~fmt::format_args`.
+  If ``format_str`` is a compile-time string then `make_args_checked` checks
+  its validity at compile time.
+  \endrst
+ */
+template <typename... Args, typename S, typename Char = char_t<S>>
+inline auto make_args_checked(const S& format_str,
+                              const remove_reference_t<Args>&... args)
+    -> format_arg_store<buffer_context<Char>, remove_reference_t<Args>...> {
+  static_assert(
+      detail::count<(
+              std::is_base_of<detail::view, remove_reference_t<Args>>::value &&
+              std::is_reference<Args>::value)...>() == 0,
+      "passing views as lvalues is disallowed");
+  detail::check_format_string<Args...>(format_str);
+  return {args...};
+}
+
 /**
   \rst
   Returns a named argument to be used in a formatting function. It should only
@@ -1729,7 +1929,14 @@ template <typename Context> class basic_format_args {
   }
 };
 
-/** An alias to ``basic_format_args<context>``. */
+#ifdef FMT_ARM_ABI_COMPATIBILITY
+/** An alias to ``basic_format_args<format_context>``. */
+// Separate types would result in shorter symbols but break ABI compatibility
+// between clang and gcc on ARM (#1919).
+using format_args = basic_format_args<format_context>;
+using wformat_args = basic_format_args<wformat_context>;
+#else
+// DEPRECATED! These are kept for ABI compatibility.
 // It is a separate type rather than an alias to make symbols readable.
 struct format_args : basic_format_args<format_context> {
   template <typename... Args>
@@ -1738,31 +1945,9 @@ struct format_args : basic_format_args<format_context> {
 struct wformat_args : basic_format_args<wformat_context> {
   using basic_format_args::basic_format_args;
 };
-
-namespace detail {
-
-// Reports a compile-time error if S is not a valid format string.
-template <typename..., typename S, FMT_ENABLE_IF(!is_compile_string<S>::value)>
-FMT_INLINE void check_format_string(const S&) {
-#ifdef FMT_ENFORCE_COMPILE_STRING
-  static_assert(is_compile_string<S>::value,
-                "FMT_ENFORCE_COMPILE_STRING requires all format strings to use "
-                "FMT_STRING.");
 #endif
-}
-template <typename..., typename S, FMT_ENABLE_IF(is_compile_string<S>::value)>
-void check_format_string(S);
 
-template <typename... Args, typename S, typename Char = char_t<S>>
-inline format_arg_store<buffer_context<Char>, remove_reference_t<Args>...>
-make_args_checked(const S& format_str,
-                  const remove_reference_t<Args>&... args) {
-  static_assert(count<(std::is_base_of<view, remove_reference_t<Args>>::value &&
-                       std::is_reference<Args>::value)...>() == 0,
-                "passing views as lvalues is disallowed");
-  check_format_string<Args...>(format_str);
-  return {args...};
-}
+namespace detail {
 
 template <typename Char, FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
 std::basic_string<Char> vformat(
@@ -1772,9 +1957,10 @@ std::basic_string<Char> vformat(
 FMT_API std::string vformat(string_view format_str, format_args args);
 
 template <typename Char>
-typename FMT_BUFFER_CONTEXT(Char)::iterator vformat_to(
+void vformat_to(
     buffer<Char>& buf, basic_string_view<Char> format_str,
-    basic_format_args<FMT_BUFFER_CONTEXT(type_identity_t<Char>)> args);
+    basic_format_args<FMT_BUFFER_CONTEXT(type_identity_t<Char>)> args,
+    detail::locale_ref loc = {});
 
 template <typename Char, typename Args,
           FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
@@ -1789,26 +1975,80 @@ inline void vprint_mojibake(std::FILE*, string_view, format_args) {}
 /** Formats a string and writes the output to ``out``. */
 // GCC 8 and earlier cannot handle std::back_insert_iterator<Container> with
 // vformat_to<ArgFormatter>(...) overload, so SFINAE on iterator type instead.
-template <
-    typename OutputIt, typename S, typename Char = char_t<S>,
-    FMT_ENABLE_IF(detail::is_contiguous_back_insert_iterator<OutputIt>::value)>
-OutputIt vformat_to(
-    OutputIt out, const S& format_str,
-    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
-  auto& c = detail::get_container(out);
-  detail::container_buffer<remove_reference_t<decltype(c)>> buf(c);
+template <typename OutputIt, typename S, typename Char = char_t<S>,
+          bool enable = detail::is_output_iterator<OutputIt, Char>::value>
+auto vformat_to(OutputIt out, const S& format_str,
+                basic_format_args<buffer_context<type_identity_t<Char>>> args)
+    -> typename std::enable_if<enable, OutputIt>::type {
+  decltype(detail::get_buffer<Char>(out)) buf(detail::get_buffer_init(out));
   detail::vformat_to(buf, to_string_view(format_str), args);
-  return out;
+  return detail::get_iterator(buf);
+}
+
+/**
+ \rst
+ Formats arguments, writes the result to the output iterator ``out`` and returns
+ the iterator past the end of the output range.
+
+ **Example**::
+
+   std::vector<char> out;
+   fmt::format_to(std::back_inserter(out), "{}", 42);
+ \endrst
+ */
+// We cannot use FMT_ENABLE_IF because of a bug in gcc 8.3.
+template <typename OutputIt, typename S, typename... Args,
+          bool enable = detail::is_output_iterator<OutputIt, char_t<S>>::value>
+inline auto format_to(OutputIt out, const S& format_str, Args&&... args) ->
+    typename std::enable_if<enable, OutputIt>::type {
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
+  return vformat_to(out, to_string_view(format_str), vargs);
+}
+
+template <typename OutputIt> struct format_to_n_result {
+  /** Iterator past the end of the output range. */
+  OutputIt out;
+  /** Total (not truncated) output size. */
+  size_t size;
+};
+
+template <typename OutputIt, typename Char, typename... Args,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value)>
+inline format_to_n_result<OutputIt> vformat_to_n(
+    OutputIt out, size_t n, basic_string_view<Char> format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  detail::iterator_buffer<OutputIt, Char, detail::fixed_buffer_traits> buf(out,
+                                                                           n);
+  detail::vformat_to(buf, format_str, args);
+  return {buf.out(), buf.count()};
+}
+
+/**
+ \rst
+ Formats arguments, writes up to ``n`` characters of the result to the output
+ iterator ``out`` and returns the total output size and the iterator past the
+ end of the output range.
+ \endrst
+ */
+template <typename OutputIt, typename S, typename... Args,
+          bool enable = detail::is_output_iterator<OutputIt, char_t<S>>::value>
+inline auto format_to_n(OutputIt out, size_t n, const S& format_str,
+                        const Args&... args) ->
+    typename std::enable_if<enable, format_to_n_result<OutputIt>>::type {
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
+  return vformat_to_n(out, n, to_string_view(format_str), vargs);
 }
 
-template <typename Container, typename S, typename... Args,
-          FMT_ENABLE_IF(
-              is_contiguous<Container>::value&& detail::is_string<S>::value)>
-inline std::back_insert_iterator<Container> format_to(
-    std::back_insert_iterator<Container> out, const S& format_str,
-    Args&&... args) {
-  return vformat_to(out, to_string_view(format_str),
-                    detail::make_args_checked<Args...>(format_str, args...));
+/**
+  Returns the number of characters in the output of
+  ``format(format_str, args...)``.
+ */
+template <typename... Args>
+inline size_t formatted_size(string_view format_str, Args&&... args) {
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
+  detail::counting_buffer<> buf;
+  detail::vformat_to(buf, format_str, vargs);
+  return buf.count();
 }
 
 template <typename S, typename Char = char_t<S>>
@@ -1832,7 +2072,7 @@ FMT_INLINE std::basic_string<Char> vformat(
 // std::basic_string<char_t<S>> to reduce the symbol size.
 template <typename S, typename... Args, typename Char = char_t<S>>
 FMT_INLINE std::basic_string<Char> format(const S& format_str, Args&&... args) {
-  const auto& vargs = detail::make_args_checked<Args...>(format_str, args...);
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
   return detail::vformat(to_string_view(format_str), vargs);
 }
 
@@ -1852,7 +2092,7 @@ FMT_API void vprint(std::FILE*, string_view, format_args);
  */
 template <typename S, typename... Args, typename Char = char_t<S>>
 inline void print(std::FILE* f, const S& format_str, Args&&... args) {
-  const auto& vargs = detail::make_args_checked<Args...>(format_str, args...);
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
   return detail::is_unicode<Char>()
              ? vprint(f, to_string_view(format_str), vargs)
              : detail::vprint_mojibake(f, to_string_view(format_str), vargs);
@@ -1871,7 +2111,7 @@ inline void print(std::FILE* f, const S& format_str, Args&&... args) {
  */
 template <typename S, typename... Args, typename Char = char_t<S>>
 inline void print(const S& format_str, Args&&... args) {
-  const auto& vargs = detail::make_args_checked<Args...>(format_str, args...);
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
   return detail::is_unicode<Char>()
              ? vprint(to_string_view(format_str), vargs)
              : detail::vprint_mojibake(stdout, to_string_view(format_str),

diff --git a/src/third_party/fmt/format-inl.h b/src/third_party/fmt/format-inl.h
index d8c9c8a5ee58b4f1fd3073c5aea25ae370208ed5..8f2fe7354a9cef2ee3cf3cbdd10a5678c250a973 100644 (file)
--- a/src/third_party/fmt/format-inl.h
+++ b/src/third_party/fmt/format-inl.h
@@ -13,32 +13,19 @@
 #include <climits>
 #include <cmath>
 #include <cstdarg>
-#include <cstring>  // for std::memmove
+#include <cstring>  // std::memmove
 #include <cwchar>
 #include <exception>
 
-#include "format.h"
-#if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
 #  include <locale>
 #endif
 
 #ifdef _WIN32
-#  if !defined(NOMINMAX) && !defined(WIN32_LEAN_AND_MEAN)
-#    define NOMINMAX
-#    define WIN32_LEAN_AND_MEAN
-#    include <windows.h>
-#    undef WIN32_LEAN_AND_MEAN
-#    undef NOMINMAX
-#  else
-#    include <windows.h>
-#  endif
-#  include <io.h>
+#  include <io.h>  // _isatty
 #endif
 
-#ifdef _MSC_VER
-#  pragma warning(push)
-#  pragma warning(disable : 4702)  // unreachable code
-#endif
+#include "format.h"
 
 // Dummy implementations of strerror_r and strerror_s called if corresponding
 // system functions are not available.
@@ -79,8 +66,8 @@ inline int fmt_snprintf(char* buffer, size_t size, const char* format, ...) {
 //   ERANGE - buffer is not large enough to store the error message
 //   other  - failure
 // Buffer should be at least of size 1.
-FMT_FUNC int safe_strerror(int error_code, char*& buffer,
-                           size_t buffer_size) FMT_NOEXCEPT {
+inline int safe_strerror(int error_code, char*& buffer,
+                         size_t buffer_size) FMT_NOEXCEPT {
   FMT_ASSERT(buffer != nullptr && buffer_size != 0, "invalid buffer");
 
   class dispatcher {
@@ -145,7 +132,7 @@ FMT_FUNC void format_error_code(detail::buffer<char>& out, int error_code,
   // Report error code making sure that the output fits into
   // inline_buffer_size to avoid dynamic memory allocation and potential
   // bad_alloc.
-  out.resize(0);
+  out.try_resize(0);
   static const char SEP[] = ": ";
   static const char ERROR_STR[] = "error ";
   // Subtract 2 to account for terminating null characters in SEP and ERROR_STR.
@@ -156,7 +143,7 @@ FMT_FUNC void format_error_code(detail::buffer<char>& out, int error_code,
     ++error_code_size;
   }
   error_code_size += detail::to_unsigned(detail::count_digits(abs_value));
-  auto it = std::back_inserter(out);
+  auto it = buffer_appender<char>(out);
   if (message.size() <= inline_buffer_size - error_code_size)
     format_to(it, "{}{}", message, SEP);
   format_to(it, "{}{}", ERROR_STR, error_code);
@@ -173,8 +160,8 @@ FMT_FUNC void report_error(format_func func, int error_code,
 }
 
 // A wrapper around fwrite that throws on error.
-FMT_FUNC void fwrite_fully(const void* ptr, size_t size, size_t count,
-                           FILE* stream) {
+inline void fwrite_fully(const void* ptr, size_t size, size_t count,
+                         FILE* stream) {
   size_t written = std::fwrite(ptr, size, count, stream);
   if (written < count) FMT_THROW(system_error(errno, "cannot write to file"));
 }
@@ -242,26 +229,23 @@ template <> FMT_FUNC int count_digits<4>(detail::fallback_uintptr n) {
 
 template <typename T>
 const typename basic_data<T>::digit_pair basic_data<T>::digits[] = {
-    {'0', '0'},  {'0', '1'},  {'0', '2'},  {'0', '3'},  {'0', '4'},
-    {'0', '5'},  {'0', '6'},  {'0', '7'},  {'0', '8'},  {'0', '9'},
-    {'1', '0'},  {'1', '1'},  {'1', '2'},  {'1', '3'},  {'1', '4'},
-    {'1', '5'},  {'1', '6'},  {'1', '7'},  {'1', '8'},  {'1', '9'},
-    {'2', '0'},  {'2', '1'},  {'2', '2'},  {'2', '3'},  {'2', '4'},
-    {'2', '5'},  {'2', '6'},  {'2', '7'},  {'2', '8'},  {'2', '9'},
-    {'3', '0'},  {'3', '1'},  {'3', '2'},  {'3', '3'},  {'3', '4'},
-    {'3', '5'},  {'3', '6'},  {'3', '7'},  {'3', '8'},  {'3', '9'},
-    {'4', '0'},  {'4', '1'},  {'4', '2'},  {'4', '3'},  {'4', '4'},
-    {'4', '5'},  {'4', '6'},  {'4', '7'},  {'4', '8'},  {'4', '9'},
-    {'5', '0'},  {'5', '1'},  {'5', '2'},  {'5', '3'},  {'5', '4'},
-    {'5', '5'},  {'5', '6'},  {'5', '7'},  {'5', '8'},  {'5', '9'},
-    {'6', '0'},  {'6', '1'},  {'6', '2'},  {'6', '3'},  {'6', '4'},
-    {'6', '5'},  {'6', '6'},  {'6', '7'},  {'6', '8'},  {'6', '9'},
-    {'7', '0'},  {'7', '1'},  {'7', '2'},  {'7', '3'},  {'7', '4'},
-    {'7', '5'},  {'7', '6'},  {'7', '7'},  {'7', '8'},  {'7', '9'},
-    {'8', '0'},  {'8', '1'},  {'8', '2'},  {'8', '3'},  {'8', '4'},
-    {'8', '5'},  {'8', '6'},  {'8', '7'},  {'8', '8'},  {'8', '9'},
-    {'9', '0'},  {'9', '1'},  {'9', '2'},  {'9', '3'},  {'9', '4'},
-    {'9', '5'},  {'9', '6'},  {'9', '7'},  {'9', '8'},  {'9', '9'}};
+    {'0', '0'}, {'0', '1'}, {'0', '2'}, {'0', '3'}, {'0', '4'}, {'0', '5'},
+    {'0', '6'}, {'0', '7'}, {'0', '8'}, {'0', '9'}, {'1', '0'}, {'1', '1'},
+    {'1', '2'}, {'1', '3'}, {'1', '4'}, {'1', '5'}, {'1', '6'}, {'1', '7'},
+    {'1', '8'}, {'1', '9'}, {'2', '0'}, {'2', '1'}, {'2', '2'}, {'2', '3'},
+    {'2', '4'}, {'2', '5'}, {'2', '6'}, {'2', '7'}, {'2', '8'}, {'2', '9'},
+    {'3', '0'}, {'3', '1'}, {'3', '2'}, {'3', '3'}, {'3', '4'}, {'3', '5'},
+    {'3', '6'}, {'3', '7'}, {'3', '8'}, {'3', '9'}, {'4', '0'}, {'4', '1'},
+    {'4', '2'}, {'4', '3'}, {'4', '4'}, {'4', '5'}, {'4', '6'}, {'4', '7'},
+    {'4', '8'}, {'4', '9'}, {'5', '0'}, {'5', '1'}, {'5', '2'}, {'5', '3'},
+    {'5', '4'}, {'5', '5'}, {'5', '6'}, {'5', '7'}, {'5', '8'}, {'5', '9'},
+    {'6', '0'}, {'6', '1'}, {'6', '2'}, {'6', '3'}, {'6', '4'}, {'6', '5'},
+    {'6', '6'}, {'6', '7'}, {'6', '8'}, {'6', '9'}, {'7', '0'}, {'7', '1'},
+    {'7', '2'}, {'7', '3'}, {'7', '4'}, {'7', '5'}, {'7', '6'}, {'7', '7'},
+    {'7', '8'}, {'7', '9'}, {'8', '0'}, {'8', '1'}, {'8', '2'}, {'8', '3'},
+    {'8', '4'}, {'8', '5'}, {'8', '6'}, {'8', '7'}, {'8', '8'}, {'8', '9'},
+    {'9', '0'}, {'9', '1'}, {'9', '2'}, {'9', '3'}, {'9', '4'}, {'9', '5'},
+    {'9', '6'}, {'9', '7'}, {'9', '8'}, {'9', '9'}};
 
 template <typename T>
 const char basic_data<T>::hex_digits[] = "0123456789abcdef";
@@ -279,16 +263,24 @@ const uint64_t basic_data<T>::powers_of_10_64[] = {
 template <typename T>
 const uint32_t basic_data<T>::zero_or_powers_of_10_32[] = {0,
                                                            FMT_POWERS_OF_10(1)};
-
 template <typename T>
 const uint64_t basic_data<T>::zero_or_powers_of_10_64[] = {
     0, FMT_POWERS_OF_10(1), FMT_POWERS_OF_10(1000000000ULL),
     10000000000000000000ULL};
 
+template <typename T>
+const uint32_t basic_data<T>::zero_or_powers_of_10_32_new[] = {
+    0, 0, FMT_POWERS_OF_10(1)};
+
+template <typename T>
+const uint64_t basic_data<T>::zero_or_powers_of_10_64_new[] = {
+    0, 0, FMT_POWERS_OF_10(1), FMT_POWERS_OF_10(1000000000ULL),
+    10000000000000000000ULL};
+
 // Normalized 64-bit significands of pow(10, k), for k = -348, -340, ..., 340.
 // These are generated by support/compute-powers.py.
 template <typename T>
-const uint64_t basic_data<T>::pow10_significands[] = {
+const uint64_t basic_data<T>::grisu_pow10_significands[] = {
     0xfa8fd5a0081c0288, 0xbaaee17fa23ebf76, 0x8b16fb203055ac76,
     0xcf42894a5dce35ea, 0x9a6bb0aa55653b2d, 0xe61acf033d1a45df,
     0xab70fe17c79ac6ca, 0xff77b1fcbebcdc4f, 0xbe5691ef416bd60c,
@@ -323,7 +315,7 @@ const uint64_t basic_data<T>::pow10_significands[] = {
 // Binary exponents of pow(10, k), for k = -348, -340, ..., 340, corresponding
 // to significands above.
 template <typename T>
-const int16_t basic_data<T>::pow10_exponents[] = {
+const int16_t basic_data<T>::grisu_pow10_exponents[] = {
     -1220, -1193, -1166, -1140, -1113, -1087, -1060, -1034, -1007, -980, -954,
     -927,  -901,  -874,  -847,  -821,  -794,  -768,  -741,  -715,  -688, -661,
     -635,  -608,  -582,  -555,  -529,  -502,  -475,  -449,  -422,  -396, -369,
@@ -333,6 +325,744 @@ const int16_t basic_data<T>::pow10_exponents[] = {
     534,   561,   588,   614,   641,   667,   694,   720,   747,   774,  800,
     827,   853,   880,   907,   933,   960,   986,   1013,  1039,  1066};
 
+template <typename T>
+const divtest_table_entry<uint32_t> basic_data<T>::divtest_table_for_pow5_32[] =
+    {{0x00000001, 0xffffffff}, {0xcccccccd, 0x33333333},
+     {0xc28f5c29, 0x0a3d70a3}, {0x26e978d5, 0x020c49ba},
+     {0x3afb7e91, 0x0068db8b}, {0x0bcbe61d, 0x0014f8b5},
+     {0x68c26139, 0x000431bd}, {0xae8d46a5, 0x0000d6bf},
+     {0x22e90e21, 0x00002af3}, {0x3a2e9c6d, 0x00000897},
+     {0x3ed61f49, 0x000001b7}};
+
+template <typename T>
+const divtest_table_entry<uint64_t> basic_data<T>::divtest_table_for_pow5_64[] =
+    {{0x0000000000000001, 0xffffffffffffffff},
+     {0xcccccccccccccccd, 0x3333333333333333},
+     {0x8f5c28f5c28f5c29, 0x0a3d70a3d70a3d70},
+     {0x1cac083126e978d5, 0x020c49ba5e353f7c},
+     {0xd288ce703afb7e91, 0x0068db8bac710cb2},
+     {0x5d4e8fb00bcbe61d, 0x0014f8b588e368f0},
+     {0x790fb65668c26139, 0x000431bde82d7b63},
+     {0xe5032477ae8d46a5, 0x0000d6bf94d5e57a},
+     {0xc767074b22e90e21, 0x00002af31dc46118},
+     {0x8e47ce423a2e9c6d, 0x0000089705f4136b},
+     {0x4fa7f60d3ed61f49, 0x000001b7cdfd9d7b},
+     {0x0fee64690c913975, 0x00000057f5ff85e5},
+     {0x3662e0e1cf503eb1, 0x000000119799812d},
+     {0xa47a2cf9f6433fbd, 0x0000000384b84d09},
+     {0x54186f653140a659, 0x00000000b424dc35},
+     {0x7738164770402145, 0x0000000024075f3d},
+     {0xe4a4d1417cd9a041, 0x000000000734aca5},
+     {0xc75429d9e5c5200d, 0x000000000170ef54},
+     {0xc1773b91fac10669, 0x000000000049c977},
+     {0x26b172506559ce15, 0x00000000000ec1e4},
+     {0xd489e3a9addec2d1, 0x000000000002f394},
+     {0x90e860bb892c8d5d, 0x000000000000971d},
+     {0x502e79bf1b6f4f79, 0x0000000000001e39},
+     {0xdcd618596be30fe5, 0x000000000000060b}};
+
+template <typename T>
+const uint64_t basic_data<T>::dragonbox_pow10_significands_64[] = {
+    0x81ceb32c4b43fcf5, 0xa2425ff75e14fc32, 0xcad2f7f5359a3b3f,
+    0xfd87b5f28300ca0e, 0x9e74d1b791e07e49, 0xc612062576589ddb,
+    0xf79687aed3eec552, 0x9abe14cd44753b53, 0xc16d9a0095928a28,
+    0xf1c90080baf72cb2, 0x971da05074da7bef, 0xbce5086492111aeb,
+    0xec1e4a7db69561a6, 0x9392ee8e921d5d08, 0xb877aa3236a4b44a,
+    0xe69594bec44de15c, 0x901d7cf73ab0acda, 0xb424dc35095cd810,
+    0xe12e13424bb40e14, 0x8cbccc096f5088cc, 0xafebff0bcb24aaff,
+    0xdbe6fecebdedd5bf, 0x89705f4136b4a598, 0xabcc77118461cefd,
+    0xd6bf94d5e57a42bd, 0x8637bd05af6c69b6, 0xa7c5ac471b478424,
+    0xd1b71758e219652c, 0x83126e978d4fdf3c, 0xa3d70a3d70a3d70b,
+    0xcccccccccccccccd, 0x8000000000000000, 0xa000000000000000,
+    0xc800000000000000, 0xfa00000000000000, 0x9c40000000000000,
+    0xc350000000000000, 0xf424000000000000, 0x9896800000000000,
+    0xbebc200000000000, 0xee6b280000000000, 0x9502f90000000000,
+    0xba43b74000000000, 0xe8d4a51000000000, 0x9184e72a00000000,
+    0xb5e620f480000000, 0xe35fa931a0000000, 0x8e1bc9bf04000000,
+    0xb1a2bc2ec5000000, 0xde0b6b3a76400000, 0x8ac7230489e80000,
+    0xad78ebc5ac620000, 0xd8d726b7177a8000, 0x878678326eac9000,
+    0xa968163f0a57b400, 0xd3c21bcecceda100, 0x84595161401484a0,
+    0xa56fa5b99019a5c8, 0xcecb8f27f4200f3a, 0x813f3978f8940984,
+    0xa18f07d736b90be5, 0xc9f2c9cd04674ede, 0xfc6f7c4045812296,
+    0x9dc5ada82b70b59d, 0xc5371912364ce305, 0xf684df56c3e01bc6,
+    0x9a130b963a6c115c, 0xc097ce7bc90715b3, 0xf0bdc21abb48db20,
+    0x96769950b50d88f4, 0xbc143fa4e250eb31, 0xeb194f8e1ae525fd,
+    0x92efd1b8d0cf37be, 0xb7abc627050305ad, 0xe596b7b0c643c719,
+    0x8f7e32ce7bea5c6f, 0xb35dbf821ae4f38b, 0xe0352f62a19e306e};
+
+template <typename T>
+const uint128_wrapper basic_data<T>::dragonbox_pow10_significands_128[] = {
+#if FMT_USE_FULL_CACHE_DRAGONBOX
+    {0xff77b1fcbebcdc4f, 0x25e8e89c13bb0f7b},
+    {0x9faacf3df73609b1, 0x77b191618c54e9ad},
+    {0xc795830d75038c1d, 0xd59df5b9ef6a2418},
+    {0xf97ae3d0d2446f25, 0x4b0573286b44ad1e},
+    {0x9becce62836ac577, 0x4ee367f9430aec33},
+    {0xc2e801fb244576d5, 0x229c41f793cda740},
+    {0xf3a20279ed56d48a, 0x6b43527578c11110},
+    {0x9845418c345644d6, 0x830a13896b78aaaa},
+    {0xbe5691ef416bd60c, 0x23cc986bc656d554},
+    {0xedec366b11c6cb8f, 0x2cbfbe86b7ec8aa9},
+    {0x94b3a202eb1c3f39, 0x7bf7d71432f3d6aa},
+    {0xb9e08a83a5e34f07, 0xdaf5ccd93fb0cc54},
+    {0xe858ad248f5c22c9, 0xd1b3400f8f9cff69},
+    {0x91376c36d99995be, 0x23100809b9c21fa2},
+    {0xb58547448ffffb2d, 0xabd40a0c2832a78b},
+    {0xe2e69915b3fff9f9, 0x16c90c8f323f516d},
+    {0x8dd01fad907ffc3b, 0xae3da7d97f6792e4},
+    {0xb1442798f49ffb4a, 0x99cd11cfdf41779d},
+    {0xdd95317f31c7fa1d, 0x40405643d711d584},
+    {0x8a7d3eef7f1cfc52, 0x482835ea666b2573},
+    {0xad1c8eab5ee43b66, 0xda3243650005eed0},
+    {0xd863b256369d4a40, 0x90bed43e40076a83},
+    {0x873e4f75e2224e68, 0x5a7744a6e804a292},
+    {0xa90de3535aaae202, 0x711515d0a205cb37},
+    {0xd3515c2831559a83, 0x0d5a5b44ca873e04},
+    {0x8412d9991ed58091, 0xe858790afe9486c3},
+    {0xa5178fff668ae0b6, 0x626e974dbe39a873},
+    {0xce5d73ff402d98e3, 0xfb0a3d212dc81290},
+    {0x80fa687f881c7f8e, 0x7ce66634bc9d0b9a},
+    {0xa139029f6a239f72, 0x1c1fffc1ebc44e81},
+    {0xc987434744ac874e, 0xa327ffb266b56221},
+    {0xfbe9141915d7a922, 0x4bf1ff9f0062baa9},
+    {0x9d71ac8fada6c9b5, 0x6f773fc3603db4aa},
+    {0xc4ce17b399107c22, 0xcb550fb4384d21d4},
+    {0xf6019da07f549b2b, 0x7e2a53a146606a49},
+    {0x99c102844f94e0fb, 0x2eda7444cbfc426e},
+    {0xc0314325637a1939, 0xfa911155fefb5309},
+    {0xf03d93eebc589f88, 0x793555ab7eba27cb},
+    {0x96267c7535b763b5, 0x4bc1558b2f3458df},
+    {0xbbb01b9283253ca2, 0x9eb1aaedfb016f17},
+    {0xea9c227723ee8bcb, 0x465e15a979c1cadd},
+    {0x92a1958a7675175f, 0x0bfacd89ec191eca},
+    {0xb749faed14125d36, 0xcef980ec671f667c},
+    {0xe51c79a85916f484, 0x82b7e12780e7401b},
+    {0x8f31cc0937ae58d2, 0xd1b2ecb8b0908811},
+    {0xb2fe3f0b8599ef07, 0x861fa7e6dcb4aa16},
+    {0xdfbdcece67006ac9, 0x67a791e093e1d49b},
+    {0x8bd6a141006042bd, 0xe0c8bb2c5c6d24e1},
+    {0xaecc49914078536d, 0x58fae9f773886e19},
+    {0xda7f5bf590966848, 0xaf39a475506a899f},
+    {0x888f99797a5e012d, 0x6d8406c952429604},
+    {0xaab37fd7d8f58178, 0xc8e5087ba6d33b84},
+    {0xd5605fcdcf32e1d6, 0xfb1e4a9a90880a65},
+    {0x855c3be0a17fcd26, 0x5cf2eea09a550680},
+    {0xa6b34ad8c9dfc06f, 0xf42faa48c0ea481f},
+    {0xd0601d8efc57b08b, 0xf13b94daf124da27},
+    {0x823c12795db6ce57, 0x76c53d08d6b70859},
+    {0xa2cb1717b52481ed, 0x54768c4b0c64ca6f},
+    {0xcb7ddcdda26da268, 0xa9942f5dcf7dfd0a},
+    {0xfe5d54150b090b02, 0xd3f93b35435d7c4d},
+    {0x9efa548d26e5a6e1, 0xc47bc5014a1a6db0},
+    {0xc6b8e9b0709f109a, 0x359ab6419ca1091c},
+    {0xf867241c8cc6d4c0, 0xc30163d203c94b63},
+    {0x9b407691d7fc44f8, 0x79e0de63425dcf1e},
+    {0xc21094364dfb5636, 0x985915fc12f542e5},
+    {0xf294b943e17a2bc4, 0x3e6f5b7b17b2939e},
+    {0x979cf3ca6cec5b5a, 0xa705992ceecf9c43},
+    {0xbd8430bd08277231, 0x50c6ff782a838354},
+    {0xece53cec4a314ebd, 0xa4f8bf5635246429},
+    {0x940f4613ae5ed136, 0x871b7795e136be9a},
+    {0xb913179899f68584, 0x28e2557b59846e40},
+    {0xe757dd7ec07426e5, 0x331aeada2fe589d0},
+    {0x9096ea6f3848984f, 0x3ff0d2c85def7622},
+    {0xb4bca50b065abe63, 0x0fed077a756b53aa},
+    {0xe1ebce4dc7f16dfb, 0xd3e8495912c62895},
+    {0x8d3360f09cf6e4bd, 0x64712dd7abbbd95d},
+    {0xb080392cc4349dec, 0xbd8d794d96aacfb4},
+    {0xdca04777f541c567, 0xecf0d7a0fc5583a1},
+    {0x89e42caaf9491b60, 0xf41686c49db57245},
+    {0xac5d37d5b79b6239, 0x311c2875c522ced6},
+    {0xd77485cb25823ac7, 0x7d633293366b828c},
+    {0x86a8d39ef77164bc, 0xae5dff9c02033198},
+    {0xa8530886b54dbdeb, 0xd9f57f830283fdfd},
+    {0xd267caa862a12d66, 0xd072df63c324fd7c},
+    {0x8380dea93da4bc60, 0x4247cb9e59f71e6e},
+    {0xa46116538d0deb78, 0x52d9be85f074e609},
+    {0xcd795be870516656, 0x67902e276c921f8c},
+    {0x806bd9714632dff6, 0x00ba1cd8a3db53b7},
+    {0xa086cfcd97bf97f3, 0x80e8a40eccd228a5},
+    {0xc8a883c0fdaf7df0, 0x6122cd128006b2ce},
+    {0xfad2a4b13d1b5d6c, 0x796b805720085f82},
+    {0x9cc3a6eec6311a63, 0xcbe3303674053bb1},
+    {0xc3f490aa77bd60fc, 0xbedbfc4411068a9d},
+    {0xf4f1b4d515acb93b, 0xee92fb5515482d45},
+    {0x991711052d8bf3c5, 0x751bdd152d4d1c4b},
+    {0xbf5cd54678eef0b6, 0xd262d45a78a0635e},
+    {0xef340a98172aace4, 0x86fb897116c87c35},
+    {0x9580869f0e7aac0e, 0xd45d35e6ae3d4da1},
+    {0xbae0a846d2195712, 0x8974836059cca10a},
+    {0xe998d258869facd7, 0x2bd1a438703fc94c},
+    {0x91ff83775423cc06, 0x7b6306a34627ddd0},
+    {0xb67f6455292cbf08, 0x1a3bc84c17b1d543},
+    {0xe41f3d6a7377eeca, 0x20caba5f1d9e4a94},
+    {0x8e938662882af53e, 0x547eb47b7282ee9d},
+    {0xb23867fb2a35b28d, 0xe99e619a4f23aa44},
+    {0xdec681f9f4c31f31, 0x6405fa00e2ec94d5},
+    {0x8b3c113c38f9f37e, 0xde83bc408dd3dd05},
+    {0xae0b158b4738705e, 0x9624ab50b148d446},
+    {0xd98ddaee19068c76, 0x3badd624dd9b0958},
+    {0x87f8a8d4cfa417c9, 0xe54ca5d70a80e5d7},
+    {0xa9f6d30a038d1dbc, 0x5e9fcf4ccd211f4d},
+    {0xd47487cc8470652b, 0x7647c32000696720},
+    {0x84c8d4dfd2c63f3b, 0x29ecd9f40041e074},
+    {0xa5fb0a17c777cf09, 0xf468107100525891},
+    {0xcf79cc9db955c2cc, 0x7182148d4066eeb5},
+    {0x81ac1fe293d599bf, 0xc6f14cd848405531},
+    {0xa21727db38cb002f, 0xb8ada00e5a506a7d},
+    {0xca9cf1d206fdc03b, 0xa6d90811f0e4851d},
+    {0xfd442e4688bd304a, 0x908f4a166d1da664},
+    {0x9e4a9cec15763e2e, 0x9a598e4e043287ff},
+    {0xc5dd44271ad3cdba, 0x40eff1e1853f29fe},
+    {0xf7549530e188c128, 0xd12bee59e68ef47d},
+    {0x9a94dd3e8cf578b9, 0x82bb74f8301958cf},
+    {0xc13a148e3032d6e7, 0xe36a52363c1faf02},
+    {0xf18899b1bc3f8ca1, 0xdc44e6c3cb279ac2},
+    {0x96f5600f15a7b7e5, 0x29ab103a5ef8c0ba},
+    {0xbcb2b812db11a5de, 0x7415d448f6b6f0e8},
+    {0xebdf661791d60f56, 0x111b495b3464ad22},
+    {0x936b9fcebb25c995, 0xcab10dd900beec35},
+    {0xb84687c269ef3bfb, 0x3d5d514f40eea743},
+    {0xe65829b3046b0afa, 0x0cb4a5a3112a5113},
+    {0x8ff71a0fe2c2e6dc, 0x47f0e785eaba72ac},
+    {0xb3f4e093db73a093, 0x59ed216765690f57},
+    {0xe0f218b8d25088b8, 0x306869c13ec3532d},
+    {0x8c974f7383725573, 0x1e414218c73a13fc},
+    {0xafbd2350644eeacf, 0xe5d1929ef90898fb},
+    {0xdbac6c247d62a583, 0xdf45f746b74abf3a},
+    {0x894bc396ce5da772, 0x6b8bba8c328eb784},
+    {0xab9eb47c81f5114f, 0x066ea92f3f326565},
+    {0xd686619ba27255a2, 0xc80a537b0efefebe},
+    {0x8613fd0145877585, 0xbd06742ce95f5f37},
+    {0xa798fc4196e952e7, 0x2c48113823b73705},
+    {0xd17f3b51fca3a7a0, 0xf75a15862ca504c6},
+    {0x82ef85133de648c4, 0x9a984d73dbe722fc},
+    {0xa3ab66580d5fdaf5, 0xc13e60d0d2e0ebbb},
+    {0xcc963fee10b7d1b3, 0x318df905079926a9},
+    {0xffbbcfe994e5c61f, 0xfdf17746497f7053},
+    {0x9fd561f1fd0f9bd3, 0xfeb6ea8bedefa634},
+    {0xc7caba6e7c5382c8, 0xfe64a52ee96b8fc1},
+    {0xf9bd690a1b68637b, 0x3dfdce7aa3c673b1},
+    {0x9c1661a651213e2d, 0x06bea10ca65c084f},
+    {0xc31bfa0fe5698db8, 0x486e494fcff30a63},
+    {0xf3e2f893dec3f126, 0x5a89dba3c3efccfb},
+    {0x986ddb5c6b3a76b7, 0xf89629465a75e01d},
+    {0xbe89523386091465, 0xf6bbb397f1135824},
+    {0xee2ba6c0678b597f, 0x746aa07ded582e2d},
+    {0x94db483840b717ef, 0xa8c2a44eb4571cdd},
+    {0xba121a4650e4ddeb, 0x92f34d62616ce414},
+    {0xe896a0d7e51e1566, 0x77b020baf9c81d18},
+    {0x915e2486ef32cd60, 0x0ace1474dc1d122f},
+    {0xb5b5ada8aaff80b8, 0x0d819992132456bb},
+    {0xe3231912d5bf60e6, 0x10e1fff697ed6c6a},
+    {0x8df5efabc5979c8f, 0xca8d3ffa1ef463c2},
+    {0xb1736b96b6fd83b3, 0xbd308ff8a6b17cb3},
+    {0xddd0467c64bce4a0, 0xac7cb3f6d05ddbdf},
+    {0x8aa22c0dbef60ee4, 0x6bcdf07a423aa96c},
+    {0xad4ab7112eb3929d, 0x86c16c98d2c953c7},
+    {0xd89d64d57a607744, 0xe871c7bf077ba8b8},
+    {0x87625f056c7c4a8b, 0x11471cd764ad4973},
+    {0xa93af6c6c79b5d2d, 0xd598e40d3dd89bd0},
+    {0xd389b47879823479, 0x4aff1d108d4ec2c4},
+    {0x843610cb4bf160cb, 0xcedf722a585139bb},
+    {0xa54394fe1eedb8fe, 0xc2974eb4ee658829},
+    {0xce947a3da6a9273e, 0x733d226229feea33},
+    {0x811ccc668829b887, 0x0806357d5a3f5260},
+    {0xa163ff802a3426a8, 0xca07c2dcb0cf26f8},
+    {0xc9bcff6034c13052, 0xfc89b393dd02f0b6},
+    {0xfc2c3f3841f17c67, 0xbbac2078d443ace3},
+    {0x9d9ba7832936edc0, 0xd54b944b84aa4c0e},
+    {0xc5029163f384a931, 0x0a9e795e65d4df12},
+    {0xf64335bcf065d37d, 0x4d4617b5ff4a16d6},
+    {0x99ea0196163fa42e, 0x504bced1bf8e4e46},
+    {0xc06481fb9bcf8d39, 0xe45ec2862f71e1d7},
+    {0xf07da27a82c37088, 0x5d767327bb4e5a4d},
+    {0x964e858c91ba2655, 0x3a6a07f8d510f870},
+    {0xbbe226efb628afea, 0x890489f70a55368c},
+    {0xeadab0aba3b2dbe5, 0x2b45ac74ccea842f},
+    {0x92c8ae6b464fc96f, 0x3b0b8bc90012929e},
+    {0xb77ada0617e3bbcb, 0x09ce6ebb40173745},
+    {0xe55990879ddcaabd, 0xcc420a6a101d0516},
+    {0x8f57fa54c2a9eab6, 0x9fa946824a12232e},
+    {0xb32df8e9f3546564, 0x47939822dc96abfa},
+    {0xdff9772470297ebd, 0x59787e2b93bc56f8},
+    {0x8bfbea76c619ef36, 0x57eb4edb3c55b65b},
+    {0xaefae51477a06b03, 0xede622920b6b23f2},
+    {0xdab99e59958885c4, 0xe95fab368e45ecee},
+    {0x88b402f7fd75539b, 0x11dbcb0218ebb415},
+    {0xaae103b5fcd2a881, 0xd652bdc29f26a11a},
+    {0xd59944a37c0752a2, 0x4be76d3346f04960},
+    {0x857fcae62d8493a5, 0x6f70a4400c562ddc},
+    {0xa6dfbd9fb8e5b88e, 0xcb4ccd500f6bb953},
+    {0xd097ad07a71f26b2, 0x7e2000a41346a7a8},
+    {0x825ecc24c873782f, 0x8ed400668c0c28c9},
+    {0xa2f67f2dfa90563b, 0x728900802f0f32fb},
+    {0xcbb41ef979346bca, 0x4f2b40a03ad2ffba},
+    {0xfea126b7d78186bc, 0xe2f610c84987bfa9},
+    {0x9f24b832e6b0f436, 0x0dd9ca7d2df4d7ca},
+    {0xc6ede63fa05d3143, 0x91503d1c79720dbc},
+    {0xf8a95fcf88747d94, 0x75a44c6397ce912b},
+    {0x9b69dbe1b548ce7c, 0xc986afbe3ee11abb},
+    {0xc24452da229b021b, 0xfbe85badce996169},
+    {0xf2d56790ab41c2a2, 0xfae27299423fb9c4},
+    {0x97c560ba6b0919a5, 0xdccd879fc967d41b},
+    {0xbdb6b8e905cb600f, 0x5400e987bbc1c921},
+    {0xed246723473e3813, 0x290123e9aab23b69},
+    {0x9436c0760c86e30b, 0xf9a0b6720aaf6522},
+    {0xb94470938fa89bce, 0xf808e40e8d5b3e6a},
+    {0xe7958cb87392c2c2, 0xb60b1d1230b20e05},
+    {0x90bd77f3483bb9b9, 0xb1c6f22b5e6f48c3},
+    {0xb4ecd5f01a4aa828, 0x1e38aeb6360b1af4},
+    {0xe2280b6c20dd5232, 0x25c6da63c38de1b1},
+    {0x8d590723948a535f, 0x579c487e5a38ad0f},
+    {0xb0af48ec79ace837, 0x2d835a9df0c6d852},
+    {0xdcdb1b2798182244, 0xf8e431456cf88e66},
+    {0x8a08f0f8bf0f156b, 0x1b8e9ecb641b5900},
+    {0xac8b2d36eed2dac5, 0xe272467e3d222f40},
+    {0xd7adf884aa879177, 0x5b0ed81dcc6abb10},
+    {0x86ccbb52ea94baea, 0x98e947129fc2b4ea},
+    {0xa87fea27a539e9a5, 0x3f2398d747b36225},
+    {0xd29fe4b18e88640e, 0x8eec7f0d19a03aae},
+    {0x83a3eeeef9153e89, 0x1953cf68300424ad},
+    {0xa48ceaaab75a8e2b, 0x5fa8c3423c052dd8},
+    {0xcdb02555653131b6, 0x3792f412cb06794e},
+    {0x808e17555f3ebf11, 0xe2bbd88bbee40bd1},
+    {0xa0b19d2ab70e6ed6, 0x5b6aceaeae9d0ec5},
+    {0xc8de047564d20a8b, 0xf245825a5a445276},
+    {0xfb158592be068d2e, 0xeed6e2f0f0d56713},
+    {0x9ced737bb6c4183d, 0x55464dd69685606c},
+    {0xc428d05aa4751e4c, 0xaa97e14c3c26b887},
+    {0xf53304714d9265df, 0xd53dd99f4b3066a9},
+    {0x993fe2c6d07b7fab, 0xe546a8038efe402a},
+    {0xbf8fdb78849a5f96, 0xde98520472bdd034},
+    {0xef73d256a5c0f77c, 0x963e66858f6d4441},
+    {0x95a8637627989aad, 0xdde7001379a44aa9},
+    {0xbb127c53b17ec159, 0x5560c018580d5d53},
+    {0xe9d71b689dde71af, 0xaab8f01e6e10b4a7},
+    {0x9226712162ab070d, 0xcab3961304ca70e9},
+    {0xb6b00d69bb55c8d1, 0x3d607b97c5fd0d23},
+    {0xe45c10c42a2b3b05, 0x8cb89a7db77c506b},
+    {0x8eb98a7a9a5b04e3, 0x77f3608e92adb243},
+    {0xb267ed1940f1c61c, 0x55f038b237591ed4},
+    {0xdf01e85f912e37a3, 0x6b6c46dec52f6689},
+    {0x8b61313bbabce2c6, 0x2323ac4b3b3da016},
+    {0xae397d8aa96c1b77, 0xabec975e0a0d081b},
+    {0xd9c7dced53c72255, 0x96e7bd358c904a22},
+    {0x881cea14545c7575, 0x7e50d64177da2e55},
+    {0xaa242499697392d2, 0xdde50bd1d5d0b9ea},
+    {0xd4ad2dbfc3d07787, 0x955e4ec64b44e865},
+    {0x84ec3c97da624ab4, 0xbd5af13bef0b113f},
+    {0xa6274bbdd0fadd61, 0xecb1ad8aeacdd58f},
+    {0xcfb11ead453994ba, 0x67de18eda5814af3},
+    {0x81ceb32c4b43fcf4, 0x80eacf948770ced8},
+    {0xa2425ff75e14fc31, 0xa1258379a94d028e},
+    {0xcad2f7f5359a3b3e, 0x096ee45813a04331},
+    {0xfd87b5f28300ca0d, 0x8bca9d6e188853fd},
+    {0x9e74d1b791e07e48, 0x775ea264cf55347e},
+    {0xc612062576589dda, 0x95364afe032a819e},
+    {0xf79687aed3eec551, 0x3a83ddbd83f52205},
+    {0x9abe14cd44753b52, 0xc4926a9672793543},
+    {0xc16d9a0095928a27, 0x75b7053c0f178294},
+    {0xf1c90080baf72cb1, 0x5324c68b12dd6339},
+    {0x971da05074da7bee, 0xd3f6fc16ebca5e04},
+    {0xbce5086492111aea, 0x88f4bb1ca6bcf585},
+    {0xec1e4a7db69561a5, 0x2b31e9e3d06c32e6},
+    {0x9392ee8e921d5d07, 0x3aff322e62439fd0},
+    {0xb877aa3236a4b449, 0x09befeb9fad487c3},
+    {0xe69594bec44de15b, 0x4c2ebe687989a9b4},
+    {0x901d7cf73ab0acd9, 0x0f9d37014bf60a11},
+    {0xb424dc35095cd80f, 0x538484c19ef38c95},
+    {0xe12e13424bb40e13, 0x2865a5f206b06fba},
+    {0x8cbccc096f5088cb, 0xf93f87b7442e45d4},
+    {0xafebff0bcb24aafe, 0xf78f69a51539d749},
+    {0xdbe6fecebdedd5be, 0xb573440e5a884d1c},
+    {0x89705f4136b4a597, 0x31680a88f8953031},
+    {0xabcc77118461cefc, 0xfdc20d2b36ba7c3e},
+    {0xd6bf94d5e57a42bc, 0x3d32907604691b4d},
+    {0x8637bd05af6c69b5, 0xa63f9a49c2c1b110},
+    {0xa7c5ac471b478423, 0x0fcf80dc33721d54},
+    {0xd1b71758e219652b, 0xd3c36113404ea4a9},
+    {0x83126e978d4fdf3b, 0x645a1cac083126ea},
+    {0xa3d70a3d70a3d70a, 0x3d70a3d70a3d70a4},
+    {0xcccccccccccccccc, 0xcccccccccccccccd},
+    {0x8000000000000000, 0x0000000000000000},
+    {0xa000000000000000, 0x0000000000000000},
+    {0xc800000000000000, 0x0000000000000000},
+    {0xfa00000000000000, 0x0000000000000000},
+    {0x9c40000000000000, 0x0000000000000000},
+    {0xc350000000000000, 0x0000000000000000},
+    {0xf424000000000000, 0x0000000000000000},
+    {0x9896800000000000, 0x0000000000000000},
+    {0xbebc200000000000, 0x0000000000000000},
+    {0xee6b280000000000, 0x0000000000000000},
+    {0x9502f90000000000, 0x0000000000000000},
+    {0xba43b74000000000, 0x0000000000000000},
+    {0xe8d4a51000000000, 0x0000000000000000},
+    {0x9184e72a00000000, 0x0000000000000000},
+    {0xb5e620f480000000, 0x0000000000000000},
+    {0xe35fa931a0000000, 0x0000000000000000},
+    {0x8e1bc9bf04000000, 0x0000000000000000},
+    {0xb1a2bc2ec5000000, 0x0000000000000000},
+    {0xde0b6b3a76400000, 0x0000000000000000},
+    {0x8ac7230489e80000, 0x0000000000000000},
+    {0xad78ebc5ac620000, 0x0000000000000000},
+    {0xd8d726b7177a8000, 0x0000000000000000},
+    {0x878678326eac9000, 0x0000000000000000},
+    {0xa968163f0a57b400, 0x0000000000000000},
+    {0xd3c21bcecceda100, 0x0000000000000000},
+    {0x84595161401484a0, 0x0000000000000000},
+    {0xa56fa5b99019a5c8, 0x0000000000000000},
+    {0xcecb8f27f4200f3a, 0x0000000000000000},
+    {0x813f3978f8940984, 0x4000000000000000},
+    {0xa18f07d736b90be5, 0x5000000000000000},
+    {0xc9f2c9cd04674ede, 0xa400000000000000},
+    {0xfc6f7c4045812296, 0x4d00000000000000},
+    {0x9dc5ada82b70b59d, 0xf020000000000000},
+    {0xc5371912364ce305, 0x6c28000000000000},
+    {0xf684df56c3e01bc6, 0xc732000000000000},
+    {0x9a130b963a6c115c, 0x3c7f400000000000},
+    {0xc097ce7bc90715b3, 0x4b9f100000000000},
+    {0xf0bdc21abb48db20, 0x1e86d40000000000},
+    {0x96769950b50d88f4, 0x1314448000000000},
+    {0xbc143fa4e250eb31, 0x17d955a000000000},
+    {0xeb194f8e1ae525fd, 0x5dcfab0800000000},
+    {0x92efd1b8d0cf37be, 0x5aa1cae500000000},
+    {0xb7abc627050305ad, 0xf14a3d9e40000000},
+    {0xe596b7b0c643c719, 0x6d9ccd05d0000000},
+    {0x8f7e32ce7bea5c6f, 0xe4820023a2000000},
+    {0xb35dbf821ae4f38b, 0xdda2802c8a800000},
+    {0xe0352f62a19e306e, 0xd50b2037ad200000},
+    {0x8c213d9da502de45, 0x4526f422cc340000},
+    {0xaf298d050e4395d6, 0x9670b12b7f410000},
+    {0xdaf3f04651d47b4c, 0x3c0cdd765f114000},
+    {0x88d8762bf324cd0f, 0xa5880a69fb6ac800},
+    {0xab0e93b6efee0053, 0x8eea0d047a457a00},
+    {0xd5d238a4abe98068, 0x72a4904598d6d880},
+    {0x85a36366eb71f041, 0x47a6da2b7f864750},
+    {0xa70c3c40a64e6c51, 0x999090b65f67d924},
+    {0xd0cf4b50cfe20765, 0xfff4b4e3f741cf6d},
+    {0x82818f1281ed449f, 0xbff8f10e7a8921a4},
+    {0xa321f2d7226895c7, 0xaff72d52192b6a0d},
+    {0xcbea6f8ceb02bb39, 0x9bf4f8a69f764490},
+    {0xfee50b7025c36a08, 0x02f236d04753d5b4},
+    {0x9f4f2726179a2245, 0x01d762422c946590},
+    {0xc722f0ef9d80aad6, 0x424d3ad2b7b97ef5},
+    {0xf8ebad2b84e0d58b, 0xd2e0898765a7deb2},
+    {0x9b934c3b330c8577, 0x63cc55f49f88eb2f},
+    {0xc2781f49ffcfa6d5, 0x3cbf6b71c76b25fb},
+    {0xf316271c7fc3908a, 0x8bef464e3945ef7a},
+    {0x97edd871cfda3a56, 0x97758bf0e3cbb5ac},
+    {0xbde94e8e43d0c8ec, 0x3d52eeed1cbea317},
+    {0xed63a231d4c4fb27, 0x4ca7aaa863ee4bdd},
+    {0x945e455f24fb1cf8, 0x8fe8caa93e74ef6a},
+    {0xb975d6b6ee39e436, 0xb3e2fd538e122b44},
+    {0xe7d34c64a9c85d44, 0x60dbbca87196b616},
+    {0x90e40fbeea1d3a4a, 0xbc8955e946fe31cd},
+    {0xb51d13aea4a488dd, 0x6babab6398bdbe41},
+    {0xe264589a4dcdab14, 0xc696963c7eed2dd1},
+    {0x8d7eb76070a08aec, 0xfc1e1de5cf543ca2},
+    {0xb0de65388cc8ada8, 0x3b25a55f43294bcb},
+    {0xdd15fe86affad912, 0x49ef0eb713f39ebe},
+    {0x8a2dbf142dfcc7ab, 0x6e3569326c784337},
+    {0xacb92ed9397bf996, 0x49c2c37f07965404},
+    {0xd7e77a8f87daf7fb, 0xdc33745ec97be906},
+    {0x86f0ac99b4e8dafd, 0x69a028bb3ded71a3},
+    {0xa8acd7c0222311bc, 0xc40832ea0d68ce0c},
+    {0xd2d80db02aabd62b, 0xf50a3fa490c30190},
+    {0x83c7088e1aab65db, 0x792667c6da79e0fa},
+    {0xa4b8cab1a1563f52, 0x577001b891185938},
+    {0xcde6fd5e09abcf26, 0xed4c0226b55e6f86},
+    {0x80b05e5ac60b6178, 0x544f8158315b05b4},
+    {0xa0dc75f1778e39d6, 0x696361ae3db1c721},
+    {0xc913936dd571c84c, 0x03bc3a19cd1e38e9},
+    {0xfb5878494ace3a5f, 0x04ab48a04065c723},
+    {0x9d174b2dcec0e47b, 0x62eb0d64283f9c76},
+    {0xc45d1df942711d9a, 0x3ba5d0bd324f8394},
+    {0xf5746577930d6500, 0xca8f44ec7ee36479},
+    {0x9968bf6abbe85f20, 0x7e998b13cf4e1ecb},
+    {0xbfc2ef456ae276e8, 0x9e3fedd8c321a67e},
+    {0xefb3ab16c59b14a2, 0xc5cfe94ef3ea101e},
+    {0x95d04aee3b80ece5, 0xbba1f1d158724a12},
+    {0xbb445da9ca61281f, 0x2a8a6e45ae8edc97},
+    {0xea1575143cf97226, 0xf52d09d71a3293bd},
+    {0x924d692ca61be758, 0x593c2626705f9c56},
+    {0xb6e0c377cfa2e12e, 0x6f8b2fb00c77836c},
+    {0xe498f455c38b997a, 0x0b6dfb9c0f956447},
+    {0x8edf98b59a373fec, 0x4724bd4189bd5eac},
+    {0xb2977ee300c50fe7, 0x58edec91ec2cb657},
+    {0xdf3d5e9bc0f653e1, 0x2f2967b66737e3ed},
+    {0x8b865b215899f46c, 0xbd79e0d20082ee74},
+    {0xae67f1e9aec07187, 0xecd8590680a3aa11},
+    {0xda01ee641a708de9, 0xe80e6f4820cc9495},
+    {0x884134fe908658b2, 0x3109058d147fdcdd},
+    {0xaa51823e34a7eede, 0xbd4b46f0599fd415},
+    {0xd4e5e2cdc1d1ea96, 0x6c9e18ac7007c91a},
+    {0x850fadc09923329e, 0x03e2cf6bc604ddb0},
+    {0xa6539930bf6bff45, 0x84db8346b786151c},
+    {0xcfe87f7cef46ff16, 0xe612641865679a63},
+    {0x81f14fae158c5f6e, 0x4fcb7e8f3f60c07e},
+    {0xa26da3999aef7749, 0xe3be5e330f38f09d},
+    {0xcb090c8001ab551c, 0x5cadf5bfd3072cc5},
+    {0xfdcb4fa002162a63, 0x73d9732fc7c8f7f6},
+    {0x9e9f11c4014dda7e, 0x2867e7fddcdd9afa},
+    {0xc646d63501a1511d, 0xb281e1fd541501b8},
+    {0xf7d88bc24209a565, 0x1f225a7ca91a4226},
+    {0x9ae757596946075f, 0x3375788de9b06958},
+    {0xc1a12d2fc3978937, 0x0052d6b1641c83ae},
+    {0xf209787bb47d6b84, 0xc0678c5dbd23a49a},
+    {0x9745eb4d50ce6332, 0xf840b7ba963646e0},
+    {0xbd176620a501fbff, 0xb650e5a93bc3d898},
+    {0xec5d3fa8ce427aff, 0xa3e51f138ab4cebe},
+    {0x93ba47c980e98cdf, 0xc66f336c36b10137},
+    {0xb8a8d9bbe123f017, 0xb80b0047445d4184},
+    {0xe6d3102ad96cec1d, 0xa60dc059157491e5},
+    {0x9043ea1ac7e41392, 0x87c89837ad68db2f},
+    {0xb454e4a179dd1877, 0x29babe4598c311fb},
+    {0xe16a1dc9d8545e94, 0xf4296dd6fef3d67a},
+    {0x8ce2529e2734bb1d, 0x1899e4a65f58660c},
+    {0xb01ae745b101e9e4, 0x5ec05dcff72e7f8f},
+    {0xdc21a1171d42645d, 0x76707543f4fa1f73},
+    {0x899504ae72497eba, 0x6a06494a791c53a8},
+    {0xabfa45da0edbde69, 0x0487db9d17636892},
+    {0xd6f8d7509292d603, 0x45a9d2845d3c42b6},
+    {0x865b86925b9bc5c2, 0x0b8a2392ba45a9b2},
+    {0xa7f26836f282b732, 0x8e6cac7768d7141e},
+    {0xd1ef0244af2364ff, 0x3207d795430cd926},
+    {0x8335616aed761f1f, 0x7f44e6bd49e807b8},
+    {0xa402b9c5a8d3a6e7, 0x5f16206c9c6209a6},
+    {0xcd036837130890a1, 0x36dba887c37a8c0f},
+    {0x802221226be55a64, 0xc2494954da2c9789},
+    {0xa02aa96b06deb0fd, 0xf2db9baa10b7bd6c},
+    {0xc83553c5c8965d3d, 0x6f92829494e5acc7},
+    {0xfa42a8b73abbf48c, 0xcb772339ba1f17f9},
+    {0x9c69a97284b578d7, 0xff2a760414536efb},
+    {0xc38413cf25e2d70d, 0xfef5138519684aba},
+    {0xf46518c2ef5b8cd1, 0x7eb258665fc25d69},
+    {0x98bf2f79d5993802, 0xef2f773ffbd97a61},
+    {0xbeeefb584aff8603, 0xaafb550ffacfd8fa},
+    {0xeeaaba2e5dbf6784, 0x95ba2a53f983cf38},
+    {0x952ab45cfa97a0b2, 0xdd945a747bf26183},
+    {0xba756174393d88df, 0x94f971119aeef9e4},
+    {0xe912b9d1478ceb17, 0x7a37cd5601aab85d},
+    {0x91abb422ccb812ee, 0xac62e055c10ab33a},
+    {0xb616a12b7fe617aa, 0x577b986b314d6009},
+    {0xe39c49765fdf9d94, 0xed5a7e85fda0b80b},
+    {0x8e41ade9fbebc27d, 0x14588f13be847307},
+    {0xb1d219647ae6b31c, 0x596eb2d8ae258fc8},
+    {0xde469fbd99a05fe3, 0x6fca5f8ed9aef3bb},
+    {0x8aec23d680043bee, 0x25de7bb9480d5854},
+    {0xada72ccc20054ae9, 0xaf561aa79a10ae6a},
+    {0xd910f7ff28069da4, 0x1b2ba1518094da04},
+    {0x87aa9aff79042286, 0x90fb44d2f05d0842},
+    {0xa99541bf57452b28, 0x353a1607ac744a53},
+    {0xd3fa922f2d1675f2, 0x42889b8997915ce8},
+    {0x847c9b5d7c2e09b7, 0x69956135febada11},
+    {0xa59bc234db398c25, 0x43fab9837e699095},
+    {0xcf02b2c21207ef2e, 0x94f967e45e03f4bb},
+    {0x8161afb94b44f57d, 0x1d1be0eebac278f5},
+    {0xa1ba1ba79e1632dc, 0x6462d92a69731732},
+    {0xca28a291859bbf93, 0x7d7b8f7503cfdcfe},
+    {0xfcb2cb35e702af78, 0x5cda735244c3d43e},
+    {0x9defbf01b061adab, 0x3a0888136afa64a7},
+    {0xc56baec21c7a1916, 0x088aaa1845b8fdd0},
+    {0xf6c69a72a3989f5b, 0x8aad549e57273d45},
+    {0x9a3c2087a63f6399, 0x36ac54e2f678864b},
+    {0xc0cb28a98fcf3c7f, 0x84576a1bb416a7dd},
+    {0xf0fdf2d3f3c30b9f, 0x656d44a2a11c51d5},
+    {0x969eb7c47859e743, 0x9f644ae5a4b1b325},
+    {0xbc4665b596706114, 0x873d5d9f0dde1fee},
+    {0xeb57ff22fc0c7959, 0xa90cb506d155a7ea},
+    {0x9316ff75dd87cbd8, 0x09a7f12442d588f2},
+    {0xb7dcbf5354e9bece, 0x0c11ed6d538aeb2f},
+    {0xe5d3ef282a242e81, 0x8f1668c8a86da5fa},
+    {0x8fa475791a569d10, 0xf96e017d694487bc},
+    {0xb38d92d760ec4455, 0x37c981dcc395a9ac},
+    {0xe070f78d3927556a, 0x85bbe253f47b1417},
+    {0x8c469ab843b89562, 0x93956d7478ccec8e},
+    {0xaf58416654a6babb, 0x387ac8d1970027b2},
+    {0xdb2e51bfe9d0696a, 0x06997b05fcc0319e},
+    {0x88fcf317f22241e2, 0x441fece3bdf81f03},
+    {0xab3c2fddeeaad25a, 0xd527e81cad7626c3},
+    {0xd60b3bd56a5586f1, 0x8a71e223d8d3b074},
+    {0x85c7056562757456, 0xf6872d5667844e49},
+    {0xa738c6bebb12d16c, 0xb428f8ac016561db},
+    {0xd106f86e69d785c7, 0xe13336d701beba52},
+    {0x82a45b450226b39c, 0xecc0024661173473},
+    {0xa34d721642b06084, 0x27f002d7f95d0190},
+    {0xcc20ce9bd35c78a5, 0x31ec038df7b441f4},
+    {0xff290242c83396ce, 0x7e67047175a15271},
+    {0x9f79a169bd203e41, 0x0f0062c6e984d386},
+    {0xc75809c42c684dd1, 0x52c07b78a3e60868},
+    {0xf92e0c3537826145, 0xa7709a56ccdf8a82},
+    {0x9bbcc7a142b17ccb, 0x88a66076400bb691},
+    {0xc2abf989935ddbfe, 0x6acff893d00ea435},
+    {0xf356f7ebf83552fe, 0x0583f6b8c4124d43},
+    {0x98165af37b2153de, 0xc3727a337a8b704a},
+    {0xbe1bf1b059e9a8d6, 0x744f18c0592e4c5c},
+    {0xeda2ee1c7064130c, 0x1162def06f79df73},
+    {0x9485d4d1c63e8be7, 0x8addcb5645ac2ba8},
+    {0xb9a74a0637ce2ee1, 0x6d953e2bd7173692},
+    {0xe8111c87c5c1ba99, 0xc8fa8db6ccdd0437},
+    {0x910ab1d4db9914a0, 0x1d9c9892400a22a2},
+    {0xb54d5e4a127f59c8, 0x2503beb6d00cab4b},
+    {0xe2a0b5dc971f303a, 0x2e44ae64840fd61d},
+    {0x8da471a9de737e24, 0x5ceaecfed289e5d2},
+    {0xb10d8e1456105dad, 0x7425a83e872c5f47},
+    {0xdd50f1996b947518, 0xd12f124e28f77719},
+    {0x8a5296ffe33cc92f, 0x82bd6b70d99aaa6f},
+    {0xace73cbfdc0bfb7b, 0x636cc64d1001550b},
+    {0xd8210befd30efa5a, 0x3c47f7e05401aa4e},
+    {0x8714a775e3e95c78, 0x65acfaec34810a71},
+    {0xa8d9d1535ce3b396, 0x7f1839a741a14d0d},
+    {0xd31045a8341ca07c, 0x1ede48111209a050},
+    {0x83ea2b892091e44d, 0x934aed0aab460432},
+    {0xa4e4b66b68b65d60, 0xf81da84d5617853f},
+    {0xce1de40642e3f4b9, 0x36251260ab9d668e},
+    {0x80d2ae83e9ce78f3, 0xc1d72b7c6b426019},
+    {0xa1075a24e4421730, 0xb24cf65b8612f81f},
+    {0xc94930ae1d529cfc, 0xdee033f26797b627},
+    {0xfb9b7cd9a4a7443c, 0x169840ef017da3b1},
+    {0x9d412e0806e88aa5, 0x8e1f289560ee864e},
+    {0xc491798a08a2ad4e, 0xf1a6f2bab92a27e2},
+    {0xf5b5d7ec8acb58a2, 0xae10af696774b1db},
+    {0x9991a6f3d6bf1765, 0xacca6da1e0a8ef29},
+    {0xbff610b0cc6edd3f, 0x17fd090a58d32af3},
+    {0xeff394dcff8a948e, 0xddfc4b4cef07f5b0},
+    {0x95f83d0a1fb69cd9, 0x4abdaf101564f98e},
+    {0xbb764c4ca7a4440f, 0x9d6d1ad41abe37f1},
+    {0xea53df5fd18d5513, 0x84c86189216dc5ed},
+    {0x92746b9be2f8552c, 0x32fd3cf5b4e49bb4},
+    {0xb7118682dbb66a77, 0x3fbc8c33221dc2a1},
+    {0xe4d5e82392a40515, 0x0fabaf3feaa5334a},
+    {0x8f05b1163ba6832d, 0x29cb4d87f2a7400e},
+    {0xb2c71d5bca9023f8, 0x743e20e9ef511012},
+    {0xdf78e4b2bd342cf6, 0x914da9246b255416},
+    {0x8bab8eefb6409c1a, 0x1ad089b6c2f7548e},
+    {0xae9672aba3d0c320, 0xa184ac2473b529b1},
+    {0xda3c0f568cc4f3e8, 0xc9e5d72d90a2741e},
+    {0x8865899617fb1871, 0x7e2fa67c7a658892},
+    {0xaa7eebfb9df9de8d, 0xddbb901b98feeab7},
+    {0xd51ea6fa85785631, 0x552a74227f3ea565},
+    {0x8533285c936b35de, 0xd53a88958f87275f},
+    {0xa67ff273b8460356, 0x8a892abaf368f137},
+    {0xd01fef10a657842c, 0x2d2b7569b0432d85},
+    {0x8213f56a67f6b29b, 0x9c3b29620e29fc73},
+    {0xa298f2c501f45f42, 0x8349f3ba91b47b8f},
+    {0xcb3f2f7642717713, 0x241c70a936219a73},
+    {0xfe0efb53d30dd4d7, 0xed238cd383aa0110},
+    {0x9ec95d1463e8a506, 0xf4363804324a40aa},
+    {0xc67bb4597ce2ce48, 0xb143c6053edcd0d5},
+    {0xf81aa16fdc1b81da, 0xdd94b7868e94050a},
+    {0x9b10a4e5e9913128, 0xca7cf2b4191c8326},
+    {0xc1d4ce1f63f57d72, 0xfd1c2f611f63a3f0},
+    {0xf24a01a73cf2dccf, 0xbc633b39673c8cec},
+    {0x976e41088617ca01, 0xd5be0503e085d813},
+    {0xbd49d14aa79dbc82, 0x4b2d8644d8a74e18},
+    {0xec9c459d51852ba2, 0xddf8e7d60ed1219e},
+    {0x93e1ab8252f33b45, 0xcabb90e5c942b503},
+    {0xb8da1662e7b00a17, 0x3d6a751f3b936243},
+    {0xe7109bfba19c0c9d, 0x0cc512670a783ad4},
+    {0x906a617d450187e2, 0x27fb2b80668b24c5},
+    {0xb484f9dc9641e9da, 0xb1f9f660802dedf6},
+    {0xe1a63853bbd26451, 0x5e7873f8a0396973},
+    {0x8d07e33455637eb2, 0xdb0b487b6423e1e8},
+    {0xb049dc016abc5e5f, 0x91ce1a9a3d2cda62},
+    {0xdc5c5301c56b75f7, 0x7641a140cc7810fb},
+    {0x89b9b3e11b6329ba, 0xa9e904c87fcb0a9d},
+    {0xac2820d9623bf429, 0x546345fa9fbdcd44},
+    {0xd732290fbacaf133, 0xa97c177947ad4095},
+    {0x867f59a9d4bed6c0, 0x49ed8eabcccc485d},
+    {0xa81f301449ee8c70, 0x5c68f256bfff5a74},
+    {0xd226fc195c6a2f8c, 0x73832eec6fff3111},
+    {0x83585d8fd9c25db7, 0xc831fd53c5ff7eab},
+    {0xa42e74f3d032f525, 0xba3e7ca8b77f5e55},
+    {0xcd3a1230c43fb26f, 0x28ce1bd2e55f35eb},
+    {0x80444b5e7aa7cf85, 0x7980d163cf5b81b3},
+    {0xa0555e361951c366, 0xd7e105bcc332621f},
+    {0xc86ab5c39fa63440, 0x8dd9472bf3fefaa7},
+    {0xfa856334878fc150, 0xb14f98f6f0feb951},
+    {0x9c935e00d4b9d8d2, 0x6ed1bf9a569f33d3},
+    {0xc3b8358109e84f07, 0x0a862f80ec4700c8},
+    {0xf4a642e14c6262c8, 0xcd27bb612758c0fa},
+    {0x98e7e9cccfbd7dbd, 0x8038d51cb897789c},
+    {0xbf21e44003acdd2c, 0xe0470a63e6bd56c3},
+    {0xeeea5d5004981478, 0x1858ccfce06cac74},
+    {0x95527a5202df0ccb, 0x0f37801e0c43ebc8},
+    {0xbaa718e68396cffd, 0xd30560258f54e6ba},
+    {0xe950df20247c83fd, 0x47c6b82ef32a2069},
+    {0x91d28b7416cdd27e, 0x4cdc331d57fa5441},
+    {0xb6472e511c81471d, 0xe0133fe4adf8e952},
+    {0xe3d8f9e563a198e5, 0x58180fddd97723a6},
+    {0x8e679c2f5e44ff8f, 0x570f09eaa7ea7648},
+    {0xb201833b35d63f73, 0x2cd2cc6551e513da},
+    {0xde81e40a034bcf4f, 0xf8077f7ea65e58d1},
+    {0x8b112e86420f6191, 0xfb04afaf27faf782},
+    {0xadd57a27d29339f6, 0x79c5db9af1f9b563},
+    {0xd94ad8b1c7380874, 0x18375281ae7822bc},
+    {0x87cec76f1c830548, 0x8f2293910d0b15b5},
+    {0xa9c2794ae3a3c69a, 0xb2eb3875504ddb22},
+    {0xd433179d9c8cb841, 0x5fa60692a46151eb},
+    {0x849feec281d7f328, 0xdbc7c41ba6bcd333},
+    {0xa5c7ea73224deff3, 0x12b9b522906c0800},
+    {0xcf39e50feae16bef, 0xd768226b34870a00},
+    {0x81842f29f2cce375, 0xe6a1158300d46640},
+    {0xa1e53af46f801c53, 0x60495ae3c1097fd0},
+    {0xca5e89b18b602368, 0x385bb19cb14bdfc4},
+    {0xfcf62c1dee382c42, 0x46729e03dd9ed7b5},
+    {0x9e19db92b4e31ba9, 0x6c07a2c26a8346d1},
+    {0xc5a05277621be293, 0xc7098b7305241885},
+    {0xf70867153aa2db38, 0xb8cbee4fc66d1ea7}
+#else
+    {0xff77b1fcbebcdc4f, 0x25e8e89c13bb0f7b},
+    {0xce5d73ff402d98e3, 0xfb0a3d212dc81290},
+    {0xa6b34ad8c9dfc06f, 0xf42faa48c0ea481f},
+    {0x86a8d39ef77164bc, 0xae5dff9c02033198},
+    {0xd98ddaee19068c76, 0x3badd624dd9b0958},
+    {0xafbd2350644eeacf, 0xe5d1929ef90898fb},
+    {0x8df5efabc5979c8f, 0xca8d3ffa1ef463c2},
+    {0xe55990879ddcaabd, 0xcc420a6a101d0516},
+    {0xb94470938fa89bce, 0xf808e40e8d5b3e6a},
+    {0x95a8637627989aad, 0xdde7001379a44aa9},
+    {0xf1c90080baf72cb1, 0x5324c68b12dd6339},
+    {0xc350000000000000, 0x0000000000000000},
+    {0x9dc5ada82b70b59d, 0xf020000000000000},
+    {0xfee50b7025c36a08, 0x02f236d04753d5b4},
+    {0xcde6fd5e09abcf26, 0xed4c0226b55e6f86},
+    {0xa6539930bf6bff45, 0x84db8346b786151c},
+    {0x865b86925b9bc5c2, 0x0b8a2392ba45a9b2},
+    {0xd910f7ff28069da4, 0x1b2ba1518094da04},
+    {0xaf58416654a6babb, 0x387ac8d1970027b2},
+    {0x8da471a9de737e24, 0x5ceaecfed289e5d2},
+    {0xe4d5e82392a40515, 0x0fabaf3feaa5334a},
+    {0xb8da1662e7b00a17, 0x3d6a751f3b936243},
+    {0x95527a5202df0ccb, 0x0f37801e0c43ebc8}
+#endif
+};
+
+#if !FMT_USE_FULL_CACHE_DRAGONBOX
+template <typename T>
+const uint64_t basic_data<T>::powers_of_5_64[] = {
+    0x0000000000000001, 0x0000000000000005, 0x0000000000000019,
+    0x000000000000007d, 0x0000000000000271, 0x0000000000000c35,
+    0x0000000000003d09, 0x000000000001312d, 0x000000000005f5e1,
+    0x00000000001dcd65, 0x00000000009502f9, 0x0000000002e90edd,
+    0x000000000e8d4a51, 0x0000000048c27395, 0x000000016bcc41e9,
+    0x000000071afd498d, 0x0000002386f26fc1, 0x000000b1a2bc2ec5,
+    0x000003782dace9d9, 0x00001158e460913d, 0x000056bc75e2d631,
+    0x0001b1ae4d6e2ef5, 0x000878678326eac9, 0x002a5a058fc295ed,
+    0x00d3c21bcecceda1, 0x0422ca8b0a00a425, 0x14adf4b7320334b9};
+
+template <typename T>
+const uint32_t basic_data<T>::dragonbox_pow10_recovery_errors[] = {
+    0x50001400, 0x54044100, 0x54014555, 0x55954415, 0x54115555, 0x00000001,
+    0x50000000, 0x00104000, 0x54010004, 0x05004001, 0x55555544, 0x41545555,
+    0x54040551, 0x15445545, 0x51555514, 0x10000015, 0x00101100, 0x01100015,
+    0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x04450514, 0x45414110,
+    0x55555145, 0x50544050, 0x15040155, 0x11054140, 0x50111514, 0x11451454,
+    0x00400541, 0x00000000, 0x55555450, 0x10056551, 0x10054011, 0x55551014,
+    0x69514555, 0x05151109, 0x00155555};
+#endif
+
 template <typename T>
 const char basic_data<T>::foreground_color[] = "\x1b[38;2;";
 template <typename T>
@@ -366,6 +1096,10 @@ class fp {
  private:
   using significand_type = uint64_t;
 
+  template <typename Float>
+  using is_supported_float = bool_constant<sizeof(Float) == sizeof(uint64_t) ||
+                                           sizeof(Float) == sizeof(uint32_t)>;
+
  public:
   significand_type f;
   int e;
@@ -388,63 +1122,38 @@ class fp {
   template <typename Double> explicit fp(Double d) { assign(d); }
 
   // Assigns d to this and return true iff predecessor is closer than successor.
-  template <typename Double, FMT_ENABLE_IF(sizeof(Double) == sizeof(uint64_t))>
-  bool assign(Double d) {
-    // Assume double is in the format [sign][exponent][significand].
-    using limits = std::numeric_limits<Double>;
+  template <typename Float, FMT_ENABLE_IF(is_supported_float<Float>::value)>
+  bool assign(Float d) {
+    // Assume float is in the format [sign][exponent][significand].
+    using limits = std::numeric_limits<Float>;
+    const int float_significand_size = limits::digits - 1;
     const int exponent_size =
-        bits<Double>::value - double_significand_size - 1;  // -1 for sign
-    const uint64_t significand_mask = implicit_bit - 1;
+        bits<Float>::value - float_significand_size - 1;  // -1 for sign
+    const uint64_t float_implicit_bit = 1ULL << float_significand_size;
+    const uint64_t significand_mask = float_implicit_bit - 1;
     const uint64_t exponent_mask = (~0ULL >> 1) & ~significand_mask;
     const int exponent_bias = (1 << exponent_size) - limits::max_exponent - 1;
-    auto u = bit_cast<uint64_t>(d);
+    constexpr bool is_double = sizeof(Float) == sizeof(uint64_t);
+    auto u = bit_cast<conditional_t<is_double, uint64_t, uint32_t>>(d);
     f = u & significand_mask;
     int biased_e =
-        static_cast<int>((u & exponent_mask) >> double_significand_size);
+        static_cast<int>((u & exponent_mask) >> float_significand_size);
     // Predecessor is closer if d is a normalized power of 2 (f == 0) other than
     // the smallest normalized number (biased_e > 1).
     bool is_predecessor_closer = f == 0 && biased_e > 1;
     if (biased_e != 0)
-      f += implicit_bit;
+      f += float_implicit_bit;
     else
       biased_e = 1;  // Subnormals use biased exponent 1 (min exponent).
-    e = biased_e - exponent_bias - double_significand_size;
+    e = biased_e - exponent_bias - float_significand_size;
     return is_predecessor_closer;
   }
 
-  template <typename Double, FMT_ENABLE_IF(sizeof(Double) != sizeof(uint64_t))>
-  bool assign(Double) {
+  template <typename Float, FMT_ENABLE_IF(!is_supported_float<Float>::value)>
+  bool assign(Float) {
     *this = fp();
     return false;
   }
-
-  // Assigns d to this together with computing lower and upper boundaries,
-  // where a boundary is a value half way between the number and its predecessor
-  // (lower) or successor (upper). The upper boundary is normalized and lower
-  // has the same exponent but may be not normalized.
-  template <typename Double> boundaries assign_with_boundaries(Double d) {
-    bool is_lower_closer = assign(d);
-    fp lower =
-        is_lower_closer ? fp((f << 2) - 1, e - 2) : fp((f << 1) - 1, e - 1);
-    // 1 in normalize accounts for the exponent shift above.
-    fp upper = normalize<1>(fp((f << 1) + 1, e - 1));
-    lower.f <<= lower.e - upper.e;
-    return boundaries{lower.f, upper.f};
-  }
-
-  template <typename Double> boundaries assign_float_with_boundaries(Double d) {
-    assign(d);
-    constexpr int min_normal_e = std::numeric_limits<float>::min_exponent -
-                                 std::numeric_limits<double>::digits;
-    significand_type half_ulp = 1 << (std::numeric_limits<double>::digits -
-                                      std::numeric_limits<float>::digits - 1);
-    if (min_normal_e > e) half_ulp <<= min_normal_e - e;
-    fp upper = normalize<0>(fp(f + half_ulp, e));
-    fp lower = fp(
-        f - (half_ulp >> ((f == implicit_bit && e > min_normal_e) ? 1 : 0)), e);
-    lower.f <<= lower.e - upper.e;
-    return boundaries{lower.f, upper.f};
-  }
 };
 
 // Normalizes the value converted from double and multiplied by (1 << SHIFT).
@@ -488,11 +1197,12 @@ inline fp operator*(fp x, fp y) { return {multiply(x.f, y.f), x.e + y.e + 64}; }
 // Returns a cached power of 10 `c_k = c_k.f * pow(2, c_k.e)` such that its
 // (binary) exponent satisfies `min_exponent <= c_k.e <= min_exponent + 28`.
 inline fp get_cached_power(int min_exponent, int& pow10_exponent) {
-  const int64_t one_over_log2_10 = 0x4d104d42;  // round(pow(2, 32) / log2(10))
+  const int shift = 32;
+  const auto significand = static_cast<int64_t>(data::log10_2_significand);
   int index = static_cast<int>(
-      ((min_exponent + fp::significand_size - 1) * one_over_log2_10 +
-       ((int64_t(1) << 32) - 1))  // ceil
-      >> 32                       // arithmetic shift
+      ((min_exponent + fp::significand_size - 1) * (significand >> shift) +
+       ((int64_t(1) << shift) - 1))  // ceil
+      >> 32                          // arithmetic shift
   );
   // Decimal exponent of the first (smallest) cached power of 10.
   const int first_dec_exp = -348;
@@ -500,7 +1210,8 @@ inline fp get_cached_power(int min_exponent, int& pow10_exponent) {
   const int dec_exp_step = 8;
   index = (index - first_dec_exp - 1) / dec_exp_step + 1;
   pow10_exponent = first_dec_exp + index * dec_exp_step;
-  return {data::pow10_significands[index], data::pow10_exponents[index]};
+  return {data::grisu_pow10_significands[index],
+          data::grisu_pow10_exponents[index]};
 }
 
 // A simple accumulator to hold the sums of terms in bigint::square if uint128_t
@@ -559,9 +1270,8 @@ class bigint {
     FMT_ASSERT(compare(*this, other) >= 0, "");
     bigit borrow = 0;
     int i = other.exp_ - exp_;
-    for (size_t j = 0, n = other.bigits_.size(); j != n; ++i, ++j) {
+    for (size_t j = 0, n = other.bigits_.size(); j != n; ++i, ++j)
       subtract_bigits(i, other.bigits_[j], borrow);
-    }
     while (borrow > 0) subtract_bigits(i, 0, borrow);
     remove_leading_zeros();
   }
@@ -733,22 +1443,26 @@ class bigint {
     exp_ *= 2;
   }
 
+  // If this bigint has a bigger exponent than other, adds trailing zero to make
+  // exponents equal. This simplifies some operations such as subtraction.
+  void align(const bigint& other) {
+    int exp_difference = exp_ - other.exp_;
+    if (exp_difference <= 0) return;
+    int num_bigits = static_cast<int>(bigits_.size());
+    bigits_.resize(to_unsigned(num_bigits + exp_difference));
+    for (int i = num_bigits - 1, j = i + exp_difference; i >= 0; --i, --j)
+      bigits_[j] = bigits_[i];
+    std::uninitialized_fill_n(bigits_.data(), exp_difference, 0);
+    exp_ -= exp_difference;
+  }
+
   // Divides this bignum by divisor, assigning the remainder to this and
   // returning the quotient.
   int divmod_assign(const bigint& divisor) {
     FMT_ASSERT(this != &divisor, "");
     if (compare(*this, divisor) < 0) return 0;
-    int num_bigits = static_cast<int>(bigits_.size());
     FMT_ASSERT(divisor.bigits_[divisor.bigits_.size() - 1u] != 0, "");
-    int exp_difference = exp_ - divisor.exp_;
-    if (exp_difference > 0) {
-      // Align bigints by adding trailing zeros to simplify subtraction.
-      bigits_.resize(to_unsigned(num_bigits + exp_difference));
-      for (int i = num_bigits - 1, j = i + exp_difference; i >= 0; --i, --j)
-        bigits_[j] = bigits_[i];
-      std::uninitialized_fill_n(bigits_.data(), exp_difference, 0);
-      exp_ -= exp_difference;
-    }
+    align(divisor);
     int quotient = 0;
     do {
       subtract_aligned(divisor);
@@ -788,20 +1502,6 @@ enum result {
 };
 }
 
-// A version of count_digits optimized for grisu_gen_digits.
-inline int grisu_count_digits(uint32_t n) {
-  if (n < 10) return 1;
-  if (n < 100) return 2;
-  if (n < 1000) return 3;
-  if (n < 10000) return 4;
-  if (n < 100000) return 5;
-  if (n < 1000000) return 6;
-  if (n < 10000000) return 7;
-  if (n < 100000000) return 8;
-  if (n < 1000000000) return 9;
-  return 10;
-}
-
 // Generates output using the Grisu digit-gen algorithm.
 // error: the size of the region (lower, upper) outside of which numbers
 // definitely do not round to value (Delta in Grisu3).
@@ -817,7 +1517,7 @@ FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
   FMT_ASSERT(integral == value.f >> -one.e, "");
   // The fractional part of scaled value (p2 in Grisu) c = value % one.
   uint64_t fractional = value.f & (one.f - 1);
-  exp = grisu_count_digits(integral);  // kappa in Grisu.
+  exp = count_digits(integral);  // kappa in Grisu.
   // Divide by 10 to prevent overflow.
   auto result = handler.on_start(data::powers_of_10_64[exp - 1] << -one.e,
                                  value.f / 10, error * 10, exp);
@@ -867,8 +1567,7 @@ FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
       FMT_ASSERT(false, "invalid number of digits");
     }
     --exp;
-    uint64_t remainder =
-        (static_cast<uint64_t>(integral) << -one.e) + fractional;
+    auto remainder = (static_cast<uint64_t>(integral) << -one.e) + fractional;
     result = handler.on_digit(static_cast<char>('0' + digit),
                               data::powers_of_10_64[exp] << -one.e, remainder,
                               error, exp, true);
@@ -878,8 +1577,7 @@ FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
   for (;;) {
     fractional *= 10;
     error *= 10;
-    char digit =
-        static_cast<char>('0' + static_cast<char>(fractional >> -one.e));
+    char digit = static_cast<char>('0' + (fractional >> -one.e));
     fractional &= one.f - 1;
     --exp;
     result = handler.on_digit(digit, one.f, fractional, error, exp, false);
@@ -916,6 +1614,7 @@ struct fixed_handler {
                           uint64_t error, int, bool integral) {
     FMT_ASSERT(remainder < divisor, "");
     buf[size++] = digit;
+    if (!integral && error >= remainder) return digits::error;
     if (size < precision) return digits::more;
     if (!integral) {
       // Check if error * 2 < divisor with overflow prevention.
@@ -935,59 +1634,684 @@ struct fixed_handler {
     }
     if (buf[0] > '9') {
       buf[0] = '1';
-      buf[size++] = '0';
+      if (fixed)
+        buf[size++] = '0';
+      else
+        ++exp10;
     }
     return digits::done;
   }
 };
 
-// The shortest representation digit handler.
-struct grisu_shortest_handler {
-  char* buf;
-  int size;
-  // Distance between scaled value and upper bound (wp_W in Grisu3).
-  uint64_t diff;
+// Implementation of Dragonbox algorithm: https://github.com/jk-jeon/dragonbox.
+namespace dragonbox {
+// Computes 128-bit result of multiplication of two 64-bit unsigned integers.
+FMT_SAFEBUFFERS inline uint128_wrapper umul128(uint64_t x,
+                                               uint64_t y) FMT_NOEXCEPT {
+#if FMT_USE_INT128
+  return static_cast<uint128_t>(x) * static_cast<uint128_t>(y);
+#elif defined(_MSC_VER) && defined(_M_X64)
+  uint128_wrapper result;
+  result.low_ = _umul128(x, y, &result.high_);
+  return result;
+#else
+  const uint64_t mask = (uint64_t(1) << 32) - uint64_t(1);
+
+  uint64_t a = x >> 32;
+  uint64_t b = x & mask;
+  uint64_t c = y >> 32;
+  uint64_t d = y & mask;
+
+  uint64_t ac = a * c;
+  uint64_t bc = b * c;
+  uint64_t ad = a * d;
+  uint64_t bd = b * d;
+
+  uint64_t intermediate = (bd >> 32) + (ad & mask) + (bc & mask);
+
+  return {ac + (intermediate >> 32) + (ad >> 32) + (bc >> 32),
+          (intermediate << 32) + (bd & mask)};
+#endif
+}
+
+// Computes upper 64 bits of multiplication of two 64-bit unsigned integers.
+FMT_SAFEBUFFERS inline uint64_t umul128_upper64(uint64_t x,
+                                                uint64_t y) FMT_NOEXCEPT {
+#if FMT_USE_INT128
+  auto p = static_cast<uint128_t>(x) * static_cast<uint128_t>(y);
+  return static_cast<uint64_t>(p >> 64);
+#elif defined(_MSC_VER) && defined(_M_X64)
+  return __umulh(x, y);
+#else
+  return umul128(x, y).high();
+#endif
+}
+
+// Computes upper 64 bits of multiplication of a 64-bit unsigned integer and a
+// 128-bit unsigned integer.
+FMT_SAFEBUFFERS inline uint64_t umul192_upper64(uint64_t x, uint128_wrapper y)
+    FMT_NOEXCEPT {
+  uint128_wrapper g0 = umul128(x, y.high());
+  g0 += umul128_upper64(x, y.low());
+  return g0.high();
+}
+
+// Computes upper 32 bits of multiplication of a 32-bit unsigned integer and a
+// 64-bit unsigned integer.
+inline uint32_t umul96_upper32(uint32_t x, uint64_t y) FMT_NOEXCEPT {
+  return static_cast<uint32_t>(umul128_upper64(x, y));
+}
+
+// Computes middle 64 bits of multiplication of a 64-bit unsigned integer and a
+// 128-bit unsigned integer.
+FMT_SAFEBUFFERS inline uint64_t umul192_middle64(uint64_t x, uint128_wrapper y)
+    FMT_NOEXCEPT {
+  uint64_t g01 = x * y.high();
+  uint64_t g10 = umul128_upper64(x, y.low());
+  return g01 + g10;
+}
+
+// Computes lower 64 bits of multiplication of a 32-bit unsigned integer and a
+// 64-bit unsigned integer.
+inline uint64_t umul96_lower64(uint32_t x, uint64_t y) FMT_NOEXCEPT {
+  return x * y;
+}
+
+// Computes floor(log10(pow(2, e))) for e in [-1700, 1700] using the method from
+// https://fmt.dev/papers/Grisu-Exact.pdf#page=5, section 3.4.
+inline int floor_log10_pow2(int e) FMT_NOEXCEPT {
+  FMT_ASSERT(e <= 1700 && e >= -1700, "too large exponent");
+  const int shift = 22;
+  return (e * static_cast<int>(data::log10_2_significand >> (64 - shift))) >>
+         shift;
+}
+
+// Various fast log computations.
+inline int floor_log2_pow10(int e) FMT_NOEXCEPT {
+  FMT_ASSERT(e <= 1233 && e >= -1233, "too large exponent");
+  const uint64_t log2_10_integer_part = 3;
+  const uint64_t log2_10_fractional_digits = 0x5269e12f346e2bf9;
+  const int shift_amount = 19;
+  return (e * static_cast<int>(
+                  (log2_10_integer_part << shift_amount) |
+                  (log2_10_fractional_digits >> (64 - shift_amount)))) >>
+         shift_amount;
+}
+inline int floor_log10_pow2_minus_log10_4_over_3(int e) FMT_NOEXCEPT {
+  FMT_ASSERT(e <= 1700 && e >= -1700, "too large exponent");
+  const uint64_t log10_4_over_3_fractional_digits = 0x1ffbfc2bbc780375;
+  const int shift_amount = 22;
+  return (e * static_cast<int>(data::log10_2_significand >>
+                               (64 - shift_amount)) -
+          static_cast<int>(log10_4_over_3_fractional_digits >>
+                           (64 - shift_amount))) >>
+         shift_amount;
+}
+
+// Returns true iff x is divisible by pow(2, exp).
+inline bool divisible_by_power_of_2(uint32_t x, int exp) FMT_NOEXCEPT {
+  FMT_ASSERT(exp >= 1, "");
+  FMT_ASSERT(x != 0, "");
+#ifdef FMT_BUILTIN_CTZ
+  return FMT_BUILTIN_CTZ(x) >= exp;
+#else
+  return exp < num_bits<uint32_t>() && x == ((x >> exp) << exp);
+#endif
+}
+inline bool divisible_by_power_of_2(uint64_t x, int exp) FMT_NOEXCEPT {
+  FMT_ASSERT(exp >= 1, "");
+  FMT_ASSERT(x != 0, "");
+#ifdef FMT_BUILTIN_CTZLL
+  return FMT_BUILTIN_CTZLL(x) >= exp;
+#else
+  return exp < num_bits<uint64_t>() && x == ((x >> exp) << exp);
+#endif
+}
+
+// Returns true iff x is divisible by pow(5, exp).
+inline bool divisible_by_power_of_5(uint32_t x, int exp) FMT_NOEXCEPT {
+  FMT_ASSERT(exp <= 10, "too large exponent");
+  return x * data::divtest_table_for_pow5_32[exp].mod_inv <=
+         data::divtest_table_for_pow5_32[exp].max_quotient;
+}
+inline bool divisible_by_power_of_5(uint64_t x, int exp) FMT_NOEXCEPT {
+  FMT_ASSERT(exp <= 23, "too large exponent");
+  return x * data::divtest_table_for_pow5_64[exp].mod_inv <=
+         data::divtest_table_for_pow5_64[exp].max_quotient;
+}
+
+// Replaces n by floor(n / pow(5, N)) returning true if and only if n is
+// divisible by pow(5, N).
+// Precondition: n <= 2 * pow(5, N + 1).
+template <int N>
+bool check_divisibility_and_divide_by_pow5(uint32_t& n) FMT_NOEXCEPT {
+  static constexpr struct {
+    uint32_t magic_number;
+    int bits_for_comparison;
+    uint32_t threshold;
+    int shift_amount;
+  } infos[] = {{0xcccd, 16, 0x3333, 18}, {0xa429, 8, 0x0a, 20}};
+  constexpr auto info = infos[N - 1];
+  n *= info.magic_number;
+  const uint32_t comparison_mask = (1u << info.bits_for_comparison) - 1;
+  bool result = (n & comparison_mask) <= info.threshold;
+  n >>= info.shift_amount;
+  return result;
+}
+
+// Computes floor(n / pow(10, N)) for small n and N.
+// Precondition: n <= pow(10, N + 1).
+template <int N> uint32_t small_division_by_pow10(uint32_t n) FMT_NOEXCEPT {
+  static constexpr struct {
+    uint32_t magic_number;
+    int shift_amount;
+    uint32_t divisor_times_10;
+  } infos[] = {{0xcccd, 19, 100}, {0xa3d8, 22, 1000}};
+  constexpr auto info = infos[N - 1];
+  FMT_ASSERT(n <= info.divisor_times_10, "n is too large");
+  return n * info.magic_number >> info.shift_amount;
+}
+
+// Computes floor(n / 10^(kappa + 1)) (float)
+inline uint32_t divide_by_10_to_kappa_plus_1(uint32_t n) FMT_NOEXCEPT {
+  return n / float_info<float>::big_divisor;
+}
+// Computes floor(n / 10^(kappa + 1)) (double)
+inline uint64_t divide_by_10_to_kappa_plus_1(uint64_t n) FMT_NOEXCEPT {
+  return umul128_upper64(n, 0x83126e978d4fdf3c) >> 9;
+}
+
+// Various subroutines using pow10 cache
+template <class T> struct cache_accessor;
+
+template <> struct cache_accessor<float> {
+  using carrier_uint = float_info<float>::carrier_uint;
+  using cache_entry_type = uint64_t;
+
+  static uint64_t get_cached_power(int k) FMT_NOEXCEPT {
+    FMT_ASSERT(k >= float_info<float>::min_k && k <= float_info<float>::max_k,
+               "k is out of range");
+    return data::dragonbox_pow10_significands_64[k - float_info<float>::min_k];
+  }
+
+  static carrier_uint compute_mul(carrier_uint u,
+                                  const cache_entry_type& cache) FMT_NOEXCEPT {
+    return umul96_upper32(u, cache);
+  }
+
+  static uint32_t compute_delta(const cache_entry_type& cache,
+                                int beta_minus_1) FMT_NOEXCEPT {
+    return static_cast<uint32_t>(cache >> (64 - 1 - beta_minus_1));
+  }
+
+  static bool compute_mul_parity(carrier_uint two_f,
+                                 const cache_entry_type& cache,
+                                 int beta_minus_1) FMT_NOEXCEPT {
+    FMT_ASSERT(beta_minus_1 >= 1, "");
+    FMT_ASSERT(beta_minus_1 < 64, "");
+
+    return ((umul96_lower64(two_f, cache) >> (64 - beta_minus_1)) & 1) != 0;
+  }
+
+  static carrier_uint compute_left_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return static_cast<carrier_uint>(
+        (cache - (cache >> (float_info<float>::significand_bits + 2))) >>
+        (64 - float_info<float>::significand_bits - 1 - beta_minus_1));
+  }
+
+  static carrier_uint compute_right_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return static_cast<carrier_uint>(
+        (cache + (cache >> (float_info<float>::significand_bits + 1))) >>
+        (64 - float_info<float>::significand_bits - 1 - beta_minus_1));
+  }
+
+  static carrier_uint compute_round_up_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return (static_cast<carrier_uint>(
+                cache >>
+                (64 - float_info<float>::significand_bits - 2 - beta_minus_1)) +
+            1) /
+           2;
+  }
+};
+
+template <> struct cache_accessor<double> {
+  using carrier_uint = float_info<double>::carrier_uint;
+  using cache_entry_type = uint128_wrapper;
+
+  static uint128_wrapper get_cached_power(int k) FMT_NOEXCEPT {
+    FMT_ASSERT(k >= float_info<double>::min_k && k <= float_info<double>::max_k,
+               "k is out of range");
+
+#if FMT_USE_FULL_CACHE_DRAGONBOX
+    return data::dragonbox_pow10_significands_128[k -
+                                                  float_info<double>::min_k];
+#else
+    static const int compression_ratio = 27;
+
+    // Compute base index.
+    int cache_index = (k - float_info<double>::min_k) / compression_ratio;
+    int kb = cache_index * compression_ratio + float_info<double>::min_k;
+    int offset = k - kb;
+
+    // Get base cache.
+    uint128_wrapper base_cache =
+        data::dragonbox_pow10_significands_128[cache_index];
+    if (offset == 0) return base_cache;
+
+    // Compute the required amount of bit-shift.
+    int alpha = floor_log2_pow10(kb + offset) - floor_log2_pow10(kb) - offset;
+    FMT_ASSERT(alpha > 0 && alpha < 64, "shifting error detected");
+
+    // Try to recover the real cache.
+    uint64_t pow5 = data::powers_of_5_64[offset];
+    uint128_wrapper recovered_cache = umul128(base_cache.high(), pow5);
+    uint128_wrapper middle_low =
+        umul128(base_cache.low() - (kb < 0 ? 1u : 0u), pow5);
+
+    recovered_cache += middle_low.high();
+
+    uint64_t high_to_middle = recovered_cache.high() << (64 - alpha);
+    uint64_t middle_to_low = recovered_cache.low() << (64 - alpha);
+
+    recovered_cache =
+        uint128_wrapper{(recovered_cache.low() >> alpha) | high_to_middle,
+                        ((middle_low.low() >> alpha) | middle_to_low)};
+
+    if (kb < 0) recovered_cache += 1;
+
+    // Get error.
+    int error_idx = (k - float_info<double>::min_k) / 16;
+    uint32_t error = (data::dragonbox_pow10_recovery_errors[error_idx] >>
+                      ((k - float_info<double>::min_k) % 16) * 2) &
+                     0x3;
+
+    // Add the error back.
+    FMT_ASSERT(recovered_cache.low() + error >= recovered_cache.low(), "");
+    return {recovered_cache.high(), recovered_cache.low() + error};
+#endif
+  }
+
+  static carrier_uint compute_mul(carrier_uint u,
+                                  const cache_entry_type& cache) FMT_NOEXCEPT {
+    return umul192_upper64(u, cache);
+  }
+
+  static uint32_t compute_delta(cache_entry_type const& cache,
+                                int beta_minus_1) FMT_NOEXCEPT {
+    return static_cast<uint32_t>(cache.high() >> (64 - 1 - beta_minus_1));
+  }
+
+  static bool compute_mul_parity(carrier_uint two_f,
+                                 const cache_entry_type& cache,
+                                 int beta_minus_1) FMT_NOEXCEPT {
+    FMT_ASSERT(beta_minus_1 >= 1, "");
+    FMT_ASSERT(beta_minus_1 < 64, "");
 
-  digits::result on_start(uint64_t, uint64_t, uint64_t, int&) {
-    return digits::more;
+    return ((umul192_middle64(two_f, cache) >> (64 - beta_minus_1)) & 1) != 0;
   }
 
-  // Decrement the generated number approaching value from above.
-  void round(uint64_t d, uint64_t divisor, uint64_t& remainder,
-             uint64_t error) {
-    while (
-        remainder < d && error - remainder >= divisor &&
-        (remainder + divisor < d || d - remainder >= remainder + divisor - d)) {
-      --buf[size - 1];
-      remainder += divisor;
+  static carrier_uint compute_left_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return (cache.high() -
+            (cache.high() >> (float_info<double>::significand_bits + 2))) >>
+           (64 - float_info<double>::significand_bits - 1 - beta_minus_1);
+  }
+
+  static carrier_uint compute_right_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return (cache.high() +
+            (cache.high() >> (float_info<double>::significand_bits + 1))) >>
+           (64 - float_info<double>::significand_bits - 1 - beta_minus_1);
+  }
+
+  static carrier_uint compute_round_up_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return ((cache.high() >>
+             (64 - float_info<double>::significand_bits - 2 - beta_minus_1)) +
+            1) /
+           2;
+  }
+};
+
+// Various integer checks
+template <class T>
+bool is_left_endpoint_integer_shorter_interval(int exponent) FMT_NOEXCEPT {
+  return exponent >=
+             float_info<
+                 T>::case_shorter_interval_left_endpoint_lower_threshold &&
+         exponent <=
+             float_info<T>::case_shorter_interval_left_endpoint_upper_threshold;
+}
+template <class T>
+bool is_endpoint_integer(typename float_info<T>::carrier_uint two_f,
+                         int exponent, int minus_k) FMT_NOEXCEPT {
+  if (exponent < float_info<T>::case_fc_pm_half_lower_threshold) return false;
+  // For k >= 0.
+  if (exponent <= float_info<T>::case_fc_pm_half_upper_threshold) return true;
+  // For k < 0.
+  if (exponent > float_info<T>::divisibility_check_by_5_threshold) return false;
+  return divisible_by_power_of_5(two_f, minus_k);
+}
+
+template <class T>
+bool is_center_integer(typename float_info<T>::carrier_uint two_f, int exponent,
+                       int minus_k) FMT_NOEXCEPT {
+  // Exponent for 5 is negative.
+  if (exponent > float_info<T>::divisibility_check_by_5_threshold) return false;
+  if (exponent > float_info<T>::case_fc_upper_threshold)
+    return divisible_by_power_of_5(two_f, minus_k);
+  // Both exponents are nonnegative.
+  if (exponent >= float_info<T>::case_fc_lower_threshold) return true;
+  // Exponent for 2 is negative.
+  return divisible_by_power_of_2(two_f, minus_k - exponent + 1);
+}
+
+// Remove trailing zeros from n and return the number of zeros removed (float)
+FMT_ALWAYS_INLINE int remove_trailing_zeros(uint32_t& n) FMT_NOEXCEPT {
+#ifdef FMT_BUILTIN_CTZ
+  int t = FMT_BUILTIN_CTZ(n);
+#else
+  int t = ctz(n);
+#endif
+  if (t > float_info<float>::max_trailing_zeros)
+    t = float_info<float>::max_trailing_zeros;
+
+  const uint32_t mod_inv1 = 0xcccccccd;
+  const uint32_t max_quotient1 = 0x33333333;
+  const uint32_t mod_inv2 = 0xc28f5c29;
+  const uint32_t max_quotient2 = 0x0a3d70a3;
+
+  int s = 0;
+  for (; s < t - 1; s += 2) {
+    if (n * mod_inv2 > max_quotient2) break;
+    n *= mod_inv2;
+  }
+  if (s < t && n * mod_inv1 <= max_quotient1) {
+    n *= mod_inv1;
+    ++s;
+  }
+  n >>= s;
+  return s;
+}
+
+// Removes trailing zeros and returns the number of zeros removed (double)
+FMT_ALWAYS_INLINE int remove_trailing_zeros(uint64_t& n) FMT_NOEXCEPT {
+#ifdef FMT_BUILTIN_CTZLL
+  int t = FMT_BUILTIN_CTZLL(n);
+#else
+  int t = ctzll(n);
+#endif
+  if (t > float_info<double>::max_trailing_zeros)
+    t = float_info<double>::max_trailing_zeros;
+  // Divide by 10^8 and reduce to 32-bits
+  // Since ret_value.significand <= (2^64 - 1) / 1000 < 10^17,
+  // both of the quotient and the r should fit in 32-bits
+
+  const uint32_t mod_inv1 = 0xcccccccd;
+  const uint32_t max_quotient1 = 0x33333333;
+  const uint64_t mod_inv8 = 0xc767074b22e90e21;
+  const uint64_t max_quotient8 = 0x00002af31dc46118;
+
+  // If the number is divisible by 1'0000'0000, work with the quotient
+  if (t >= 8) {
+    auto quotient_candidate = n * mod_inv8;
+
+    if (quotient_candidate <= max_quotient8) {
+      auto quotient = static_cast<uint32_t>(quotient_candidate >> 8);
+
+      int s = 8;
+      for (; s < t; ++s) {
+        if (quotient * mod_inv1 > max_quotient1) break;
+        quotient *= mod_inv1;
+      }
+      quotient >>= (s - 8);
+      n = quotient;
+      return s;
     }
   }
 
-  // Implements Grisu's round_weed.
-  digits::result on_digit(char digit, uint64_t divisor, uint64_t remainder,
-                          uint64_t error, int exp, bool integral) {
-    buf[size++] = digit;
-    if (remainder >= error) return digits::more;
-    uint64_t unit = integral ? 1 : data::powers_of_10_64[-exp];
-    uint64_t up = (diff - 1) * unit;  // wp_Wup
-    round(up, divisor, remainder, error);
-    uint64_t down = (diff + 1) * unit;  // wp_Wdown
-    if (remainder < down && error - remainder >= divisor &&
-        (remainder + divisor < down ||
-         down - remainder > remainder + divisor - down)) {
-      return digits::error;
+  // Otherwise, work with the remainder
+  auto quotient = static_cast<uint32_t>(n / 100000000);
+  auto remainder = static_cast<uint32_t>(n - 100000000 * quotient);
+
+  if (t == 0 || remainder * mod_inv1 > max_quotient1) {
+    return 0;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 1 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 1) + quotient * 10000000ull;
+    return 1;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 2 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 2) + quotient * 1000000ull;
+    return 2;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 3 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 3) + quotient * 100000ull;
+    return 3;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 4 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 4) + quotient * 10000ull;
+    return 4;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 5 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 5) + quotient * 1000ull;
+    return 5;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 6 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 6) + quotient * 100ull;
+    return 6;
+  }
+  remainder *= mod_inv1;
+
+  n = (remainder >> 7) + quotient * 10ull;
+  return 7;
+}
+
+// The main algorithm for shorter interval case
+template <class T>
+FMT_ALWAYS_INLINE FMT_SAFEBUFFERS decimal_fp<T> shorter_interval_case(
+    int exponent) FMT_NOEXCEPT {
+  decimal_fp<T> ret_value;
+  // Compute k and beta
+  const int minus_k = floor_log10_pow2_minus_log10_4_over_3(exponent);
+  const int beta_minus_1 = exponent + floor_log2_pow10(-minus_k);
+
+  // Compute xi and zi
+  using cache_entry_type = typename cache_accessor<T>::cache_entry_type;
+  const cache_entry_type cache = cache_accessor<T>::get_cached_power(-minus_k);
+
+  auto xi = cache_accessor<T>::compute_left_endpoint_for_shorter_interval_case(
+      cache, beta_minus_1);
+  auto zi = cache_accessor<T>::compute_right_endpoint_for_shorter_interval_case(
+      cache, beta_minus_1);
+
+  // If the left endpoint is not an integer, increase it
+  if (!is_left_endpoint_integer_shorter_interval<T>(exponent)) ++xi;
+
+  // Try bigger divisor
+  ret_value.significand = zi / 10;
+
+  // If succeed, remove trailing zeros if necessary and return
+  if (ret_value.significand * 10 >= xi) {
+    ret_value.exponent = minus_k + 1;
+    ret_value.exponent += remove_trailing_zeros(ret_value.significand);
+    return ret_value;
+  }
+
+  // Otherwise, compute the round-up of y
+  ret_value.significand =
+      cache_accessor<T>::compute_round_up_for_shorter_interval_case(
+          cache, beta_minus_1);
+  ret_value.exponent = minus_k;
+
+  // When tie occurs, choose one of them according to the rule
+  if (exponent >= float_info<T>::shorter_interval_tie_lower_threshold &&
+      exponent <= float_info<T>::shorter_interval_tie_upper_threshold) {
+    ret_value.significand = ret_value.significand % 2 == 0
+                                ? ret_value.significand
+                                : ret_value.significand - 1;
+  } else if (ret_value.significand < xi) {
+    ++ret_value.significand;
+  }
+  return ret_value;
+}
+
+template <typename T>
+FMT_SAFEBUFFERS decimal_fp<T> to_decimal(T x) FMT_NOEXCEPT {
+  // Step 1: integer promotion & Schubfach multiplier calculation.
+
+  using carrier_uint = typename float_info<T>::carrier_uint;
+  using cache_entry_type = typename cache_accessor<T>::cache_entry_type;
+  auto br = bit_cast<carrier_uint>(x);
+
+  // Extract significand bits and exponent bits.
+  const carrier_uint significand_mask =
+      (static_cast<carrier_uint>(1) << float_info<T>::significand_bits) - 1;
+  carrier_uint significand = (br & significand_mask);
+  int exponent = static_cast<int>((br & exponent_mask<T>()) >>
+                                  float_info<T>::significand_bits);
+
+  if (exponent != 0) {  // Check if normal.
+    exponent += float_info<T>::exponent_bias - float_info<T>::significand_bits;
+
+    // Shorter interval case; proceed like Schubfach.
+    if (significand == 0) return shorter_interval_case<T>(exponent);
+
+    significand |=
+        (static_cast<carrier_uint>(1) << float_info<T>::significand_bits);
+  } else {
+    // Subnormal case; the interval is always regular.
+    if (significand == 0) return {0, 0};
+    exponent = float_info<T>::min_exponent - float_info<T>::significand_bits;
+  }
+
+  const bool include_left_endpoint = (significand % 2 == 0);
+  const bool include_right_endpoint = include_left_endpoint;
+
+  // Compute k and beta.
+  const int minus_k = floor_log10_pow2(exponent) - float_info<T>::kappa;
+  const cache_entry_type cache = cache_accessor<T>::get_cached_power(-minus_k);
+  const int beta_minus_1 = exponent + floor_log2_pow10(-minus_k);
+
+  // Compute zi and deltai
+  // 10^kappa <= deltai < 10^(kappa + 1)
+  const uint32_t deltai = cache_accessor<T>::compute_delta(cache, beta_minus_1);
+  const carrier_uint two_fc = significand << 1;
+  const carrier_uint two_fr = two_fc | 1;
+  const carrier_uint zi =
+      cache_accessor<T>::compute_mul(two_fr << beta_minus_1, cache);
+
+  // Step 2: Try larger divisor; remove trailing zeros if necessary
+
+  // Using an upper bound on zi, we might be able to optimize the division
+  // better than the compiler; we are computing zi / big_divisor here
+  decimal_fp<T> ret_value;
+  ret_value.significand = divide_by_10_to_kappa_plus_1(zi);
+  uint32_t r = static_cast<uint32_t>(zi - float_info<T>::big_divisor *
+                                              ret_value.significand);
+
+  if (r > deltai) {
+    goto small_divisor_case_label;
+  } else if (r < deltai) {
+    // Exclude the right endpoint if necessary
+    if (r == 0 && !include_right_endpoint &&
+        is_endpoint_integer<T>(two_fr, exponent, minus_k)) {
+      --ret_value.significand;
+      r = float_info<T>::big_divisor;
+      goto small_divisor_case_label;
+    }
+  } else {
+    // r == deltai; compare fractional parts
+    // Check conditions in the order different from the paper
+    // to take advantage of short-circuiting
+    const carrier_uint two_fl = two_fc - 1;
+    if ((!include_left_endpoint ||
+         !is_endpoint_integer<T>(two_fl, exponent, minus_k)) &&
+        !cache_accessor<T>::compute_mul_parity(two_fl, cache, beta_minus_1)) {
+      goto small_divisor_case_label;
     }
-    return 2 * unit <= remainder && remainder <= error - 4 * unit
-               ? digits::done
-               : digits::error;
   }
-};
+  ret_value.exponent = minus_k + float_info<T>::kappa + 1;
+
+  // We may need to remove trailing zeros
+  ret_value.exponent += remove_trailing_zeros(ret_value.significand);
+  return ret_value;
+
+  // Step 3: Find the significand with the smaller divisor
+
+small_divisor_case_label:
+  ret_value.significand *= 10;
+  ret_value.exponent = minus_k + float_info<T>::kappa;
+
+  const uint32_t mask = (1u << float_info<T>::kappa) - 1;
+  auto dist = r - (deltai / 2) + (float_info<T>::small_divisor / 2);
+
+  // Is dist divisible by 2^kappa?
+  if ((dist & mask) == 0) {
+    const bool approx_y_parity =
+        ((dist ^ (float_info<T>::small_divisor / 2)) & 1) != 0;
+    dist >>= float_info<T>::kappa;
+
+    // Is dist divisible by 5^kappa?
+    if (check_divisibility_and_divide_by_pow5<float_info<T>::kappa>(dist)) {
+      ret_value.significand += dist;
+
+      // Check z^(f) >= epsilon^(f)
+      // We have either yi == zi - epsiloni or yi == (zi - epsiloni) - 1,
+      // where yi == zi - epsiloni if and only if z^(f) >= epsilon^(f)
+      // Since there are only 2 possibilities, we only need to care about the
+      // parity. Also, zi and r should have the same parity since the divisor
+      // is an even number
+      if (cache_accessor<T>::compute_mul_parity(two_fc, cache, beta_minus_1) !=
+          approx_y_parity) {
+        --ret_value.significand;
+      } else {
+        // If z^(f) >= epsilon^(f), we might have a tie
+        // when z^(f) == epsilon^(f), or equivalently, when y is an integer
+        if (is_center_integer<T>(two_fc, exponent, minus_k)) {
+          ret_value.significand = ret_value.significand % 2 == 0
+                                      ? ret_value.significand
+                                      : ret_value.significand - 1;
+        }
+      }
+    }
+    // Is dist not divisible by 5^kappa?
+    else {
+      ret_value.significand += dist;
+    }
+  }
+  // Is dist not divisible by 2^kappa?
+  else {
+    // Since we know dist is small, we might be able to optimize the division
+    // better than the compiler; we are computing dist / small_divisor here
+    ret_value.significand +=
+        small_division_by_pow10<float_info<T>::kappa>(dist);
+  }
+  return ret_value;
+}
+}  // namespace dragonbox
 
 // Formats value using a variation of the Fixed-Precision Positive
 // Floating-Point Printout ((FPP)^2) algorithm by Steele & White:
 // https://fmt.dev/p372-steele.pdf.
 template <typename Double>
-void fallback_format(Double d, buffer<char>& buf, int& exp10) {
+void fallback_format(Double d, int num_digits, bool binary32, buffer<char>& buf,
+                     int& exp10) {
   bigint numerator;    // 2 * R in (FPP)^2.
   bigint denominator;  // 2 * S in (FPP)^2.
   // lower and upper are differences between value and corresponding boundaries.
@@ -998,8 +2322,9 @@ void fallback_format(Double d, buffer<char>& buf, int& exp10) {
   // Shift numerator and denominator by an extra bit or two (if lower boundary
   // is closer) to make lower and upper integers. This eliminates multiplication
   // by 2 during later computations.
-  // TODO: handle float
-  int shift = value.assign(d) ? 2 : 1;
+  const bool is_predecessor_closer =
+      binary32 ? value.assign(static_cast<float>(d)) : value.assign(d);
+  int shift = is_predecessor_closer ? 2 : 1;
   uint64_t significand = value.f << shift;
   if (value.e >= 0) {
     numerator.assign(significand);
@@ -1012,7 +2337,7 @@ void fallback_format(Double d, buffer<char>& buf, int& exp10) {
       upper = &upper_store;
     }
     denominator.assign_pow10(exp10);
-    denominator <<= 1;
+    denominator <<= shift;
   } else if (exp10 < 0) {
     numerator.assign_pow10(-exp10);
     lower.assign(numerator);
@@ -1034,39 +2359,73 @@ void fallback_format(Double d, buffer<char>& buf, int& exp10) {
       upper = &upper_store;
     }
   }
-  if (!upper) upper = &lower;
   // Invariant: value == (numerator / denominator) * pow(10, exp10).
-  bool even = (value.f & 1) == 0;
-  int num_digits = 0;
-  char* data = buf.data();
-  for (;;) {
-    int digit = numerator.divmod_assign(denominator);
-    bool low = compare(numerator, lower) - even < 0;  // numerator <[=] lower.
-    // numerator + upper >[=] pow10:
-    bool high = add_compare(numerator, *upper, denominator) + even > 0;
-    data[num_digits++] = static_cast<char>('0' + digit);
-    if (low || high) {
-      if (!low) {
-        ++data[num_digits - 1];
-      } else if (high) {
-        int result = add_compare(numerator, numerator, denominator);
-        // Round half to even.
-        if (result > 0 || (result == 0 && (digit % 2) != 0))
+  if (num_digits < 0) {
+    // Generate the shortest representation.
+    if (!upper) upper = &lower;
+    bool even = (value.f & 1) == 0;
+    num_digits = 0;
+    char* data = buf.data();
+    for (;;) {
+      int digit = numerator.divmod_assign(denominator);
+      bool low = compare(numerator, lower) - even < 0;  // numerator <[=] lower.
+      // numerator + upper >[=] pow10:
+      bool high = add_compare(numerator, *upper, denominator) + even > 0;
+      data[num_digits++] = static_cast<char>('0' + digit);
+      if (low || high) {
+        if (!low) {
           ++data[num_digits - 1];
+        } else if (high) {
+          int result = add_compare(numerator, numerator, denominator);
+          // Round half to even.
+          if (result > 0 || (result == 0 && (digit % 2) != 0))
+            ++data[num_digits - 1];
+        }
+        buf.try_resize(to_unsigned(num_digits));
+        exp10 -= num_digits - 1;
+        return;
       }
-      buf.resize(to_unsigned(num_digits));
-      exp10 -= num_digits - 1;
-      return;
+      numerator *= 10;
+      lower *= 10;
+      if (upper != &lower) *upper *= 10;
     }
+  }
+  // Generate the given number of digits.
+  exp10 -= num_digits - 1;
+  if (num_digits == 0) {
+    buf.try_resize(1);
+    denominator *= 10;
+    buf[0] = add_compare(numerator, numerator, denominator) > 0 ? '1' : '0';
+    return;
+  }
+  buf.try_resize(to_unsigned(num_digits));
+  for (int i = 0; i < num_digits - 1; ++i) {
+    int digit = numerator.divmod_assign(denominator);
+    buf[i] = static_cast<char>('0' + digit);
     numerator *= 10;
-    lower *= 10;
-    if (upper != &lower) *upper *= 10;
   }
+  int digit = numerator.divmod_assign(denominator);
+  auto result = add_compare(numerator, numerator, denominator);
+  if (result > 0 || (result == 0 && (digit % 2) != 0)) {
+    if (digit == 9) {
+      const auto overflow = '0' + 10;
+      buf[num_digits - 1] = overflow;
+      // Propagate the carry.
+      for (int i = num_digits - 1; i > 0 && buf[i] == overflow; --i) {
+        buf[i] = '0';
+        ++buf[i - 1];
+      }
+      if (buf[0] == overflow) {
+        buf[0] = '1';
+        ++exp10;
+      }
+      return;
+    }
+    ++digit;
+  }
+  buf[num_digits - 1] = static_cast<char>('0' + digit);
 }
 
-// Formats value using the Grisu algorithm
-// (https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf)
-// if T is a IEEE754 binary32 or binary64 and snprintf otherwise.
 template <typename T>
 int format_float(T value, int precision, float_specs specs, buffer<char>& buf) {
   static_assert(!std::is_same<T, float>::value, "");
@@ -1078,66 +2437,57 @@ int format_float(T value, int precision, float_specs specs, buffer<char>& buf) {
       buf.push_back('0');
       return 0;
     }
-    buf.resize(to_unsigned(precision));
+    buf.try_resize(to_unsigned(precision));
     std::uninitialized_fill_n(buf.data(), precision, '0');
     return -precision;
   }
 
   if (!specs.use_grisu) return snprintf_float(value, precision, specs, buf);
 
+  if (precision < 0) {
+    // Use Dragonbox for the shortest format.
+    if (specs.binary32) {
+      auto dec = dragonbox::to_decimal(static_cast<float>(value));
+      write<char>(buffer_appender<char>(buf), dec.significand);
+      return dec.exponent;
+    }
+    auto dec = dragonbox::to_decimal(static_cast<double>(value));
+    write<char>(buffer_appender<char>(buf), dec.significand);
+    return dec.exponent;
+  }
+
+  // Use Grisu + Dragon4 for the given precision:
+  // https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf.
   int exp = 0;
   const int min_exp = -60;  // alpha in Grisu.
   int cached_exp10 = 0;     // K in Grisu.
-  if (precision < 0) {
-    fp fp_value;
-    auto boundaries = specs.binary32
-                          ? fp_value.assign_float_with_boundaries(value)
-                          : fp_value.assign_with_boundaries(value);
-    fp_value = normalize(fp_value);
-    // Find a cached power of 10 such that multiplying value by it will bring
-    // the exponent in the range [min_exp, -32].
-    const fp cached_pow = get_cached_power(
-        min_exp - (fp_value.e + fp::significand_size), cached_exp10);
-    // Multiply value and boundaries by the cached power of 10.
-    fp_value = fp_value * cached_pow;
-    boundaries.lower = multiply(boundaries.lower, cached_pow.f);
-    boundaries.upper = multiply(boundaries.upper, cached_pow.f);
-    assert(min_exp <= fp_value.e && fp_value.e <= -32);
-    --boundaries.lower;  // \tilde{M}^- - 1 ulp -> M^-_{\downarrow}.
-    ++boundaries.upper;  // \tilde{M}^+ + 1 ulp -> M^+_{\uparrow}.
-    // Numbers outside of (lower, upper) definitely do not round to value.
-    grisu_shortest_handler handler{buf.data(), 0,
-                                   boundaries.upper - fp_value.f};
-    auto result =
-        grisu_gen_digits(fp(boundaries.upper, fp_value.e),
-                         boundaries.upper - boundaries.lower, exp, handler);
-    if (result == digits::error) {
-      exp += handler.size - cached_exp10 - 1;
-      fallback_format(value, buf, exp);
-      return exp;
-    }
-    buf.resize(to_unsigned(handler.size));
+  fp normalized = normalize(fp(value));
+  const auto cached_pow = get_cached_power(
+      min_exp - (normalized.e + fp::significand_size), cached_exp10);
+  normalized = normalized * cached_pow;
+  // Limit precision to the maximum possible number of significant digits in an
+  // IEEE754 double because we don't need to generate zeros.
+  const int max_double_digits = 767;
+  if (precision > max_double_digits) precision = max_double_digits;
+  fixed_handler handler{buf.data(), 0, precision, -cached_exp10, fixed};
+  if (grisu_gen_digits(normalized, 1, exp, handler) == digits::error) {
+    exp += handler.size - cached_exp10 - 1;
+    fallback_format(value, handler.precision, specs.binary32, buf, exp);
   } else {
-    if (precision > 17) return snprintf_float(value, precision, specs, buf);
-    fp normalized = normalize(fp(value));
-    const auto cached_pow = get_cached_power(
-        min_exp - (normalized.e + fp::significand_size), cached_exp10);
-    normalized = normalized * cached_pow;
-    fixed_handler handler{buf.data(), 0, precision, -cached_exp10, fixed};
-    if (grisu_gen_digits(normalized, 1, exp, handler) == digits::error)
-      return snprintf_float(value, precision, specs, buf);
-    int num_digits = handler.size;
-    if (!fixed) {
-      // Remove trailing zeros.
-      while (num_digits > 0 && buf[num_digits - 1] == '0') {
-        --num_digits;
-        ++exp;
-      }
+    exp += handler.exp10;
+    buf.try_resize(to_unsigned(handler.size));
+  }
+  if (!fixed && !specs.showpoint) {
+    // Remove trailing zeros.
+    auto num_digits = buf.size();
+    while (num_digits > 0 && buf[num_digits - 1] == '0') {
+      --num_digits;
+      ++exp;
     }
-    buf.resize(to_unsigned(num_digits));
+    buf.try_resize(num_digits);
   }
-  return exp - cached_exp10;
-}
+  return exp;
+}  // namespace detail
 
 template <typename T>
 int snprintf_float(T value, int precision, float_specs specs,
@@ -1185,19 +2535,20 @@ int snprintf_float(T value, int precision, float_specs specs,
                      ? snprintf_ptr(begin, capacity, format, precision, value)
                      : snprintf_ptr(begin, capacity, format, value);
     if (result < 0) {
-      buf.reserve(buf.capacity() + 1);  // The buffer will grow exponentially.
+      // The buffer will grow exponentially.
+      buf.try_reserve(buf.capacity() + 1);
       continue;
     }
     auto size = to_unsigned(result);
     // Size equal to capacity means that the last character was truncated.
     if (size >= capacity) {
-      buf.reserve(size + offset + 1);  // Add 1 for the terminating '\0'.
+      buf.try_reserve(size + offset + 1);  // Add 1 for the terminating '\0'.
       continue;
     }
     auto is_digit = [](char c) { return c >= '0' && c <= '9'; };
     if (specs.format == float_format::fixed) {
       if (precision == 0) {
-        buf.resize(size);
+        buf.try_resize(size);
         return 0;
       }
       // Find and remove the decimal point.
@@ -1207,11 +2558,11 @@ int snprintf_float(T value, int precision, float_specs specs,
       } while (is_digit(*p));
       int fraction_size = static_cast<int>(end - p - 1);
       std::memmove(p, p + 1, to_unsigned(fraction_size));
-      buf.resize(size - 1);
+      buf.try_resize(size - 1);
       return -fraction_size;
     }
     if (specs.format == float_format::hex) {
-      buf.resize(size + offset);
+      buf.try_resize(size + offset);
       return 0;
     }
     // Find and parse the exponent.
@@ -1237,7 +2588,7 @@ int snprintf_float(T value, int precision, float_specs specs,
       fraction_size = static_cast<int>(fraction_end - begin - 1);
       std::memmove(begin + 1, begin + 2, to_unsigned(fraction_size));
     }
-    buf.resize(to_unsigned(fraction_size) + offset + 1);
+    buf.try_resize(to_unsigned(fraction_size) + offset + 1);
     return exp - fraction_size;
   }
 }
@@ -1259,25 +2610,18 @@ int snprintf_float(T value, int precision, float_specs specs,
  * occurs, this pointer will be a guess that depends on the particular
  * error, but it will always advance at least one byte.
  */
-FMT_FUNC const char* utf8_decode(const char* buf, uint32_t* c, int* e) {
-  static const char lengths[] = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
-                                 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0,
-                                 0, 0, 2, 2, 2, 2, 3, 3, 4, 0};
+inline const char* utf8_decode(const char* buf, uint32_t* c, int* e) {
   static const int masks[] = {0x00, 0x7f, 0x1f, 0x0f, 0x07};
   static const uint32_t mins[] = {4194304, 0, 128, 2048, 65536};
   static const int shiftc[] = {0, 18, 12, 6, 0};
   static const int shifte[] = {0, 6, 4, 2, 0};
 
-  auto s = reinterpret_cast<const unsigned char*>(buf);
-  int len = lengths[s[0] >> 3];
-
-  // Compute the pointer to the next character early so that the next
-  // iteration can start working on the next character. Neither Clang
-  // nor GCC figure out this reordering on their own.
-  const char* next = buf + len + !len;
+  int len = code_point_length(buf);
+  const char* next = buf + len;
 
   // Assume a four-byte character and load four bytes. Unused bits are
   // shifted out.
+  auto s = reinterpret_cast<const unsigned char*>(buf);
   *c = uint32_t(s[0] & masks[len]) << 18;
   *c |= uint32_t(s[1] & 0x3f) << 12;
   *c |= uint32_t(s[2] & 0x3f) << 6;
@@ -1296,6 +2640,19 @@ FMT_FUNC const char* utf8_decode(const char* buf, uint32_t* c, int* e) {
 
   return next;
 }
+
+struct stringifier {
+  template <typename T> FMT_INLINE std::string operator()(T value) const {
+    return to_string(value);
+  }
+  std::string operator()(basic_format_arg<format_context>::handle h) const {
+    memory_buffer buf;
+    format_parse_context parse_ctx({});
+    format_context format_ctx(buffer_appender<char>(buf), {}, {});
+    h.format(parse_ctx, format_ctx);
+    return to_string(buf);
+  }
+};
 }  // namespace detail
 
 template <> struct formatter<detail::bigint> {
@@ -1363,7 +2720,8 @@ FMT_FUNC void format_system_error(detail::buffer<char>& out, int error_code,
       int result =
           detail::safe_strerror(error_code, system_message, buf.size());
       if (result == 0) {
-        format_to(std::back_inserter(out), "{}: {}", message, system_message);
+        format_to(detail::buffer_appender<char>(out), "{}: {}", message,
+                  system_message);
         return;
       }
       if (result != ERANGE)
@@ -1384,20 +2742,6 @@ FMT_FUNC void report_system_error(int error_code,
   report_error(format_system_error, error_code, message);
 }
 
-struct stringifier {
-  template <typename T> FMT_INLINE std::string operator()(T value) const {
-    return to_string(value);
-  }
-  std::string operator()(basic_format_arg<format_context>::handle h) const {
-    memory_buffer buf;
-    detail::buffer<char>& base = buf;
-    format_parse_context parse_ctx({});
-    format_context format_ctx(std::back_inserter(base), {}, {});
-    h.format(parse_ctx, format_ctx);
-    return to_string(buf);
-  }
-};
-
 FMT_FUNC std::string detail::vformat(string_view format_str, format_args args) {
   if (format_str.size() == 2 && equal2(format_str.data(), "{}")) {
     auto arg = args.get(0);
@@ -1409,6 +2753,14 @@ FMT_FUNC std::string detail::vformat(string_view format_str, format_args args) {
   return to_string(buffer);
 }
 
+#ifdef _WIN32
+namespace detail {
+using dword = conditional_t<sizeof(long) == 4, unsigned long, unsigned>;
+extern "C" __declspec(dllimport) int __stdcall WriteConsoleW(  //
+    void*, const void*, dword, dword*, void*);
+}  // namespace detail
+#endif
+
 FMT_FUNC void vprint(std::FILE* f, string_view format_str, format_args args) {
   memory_buffer buffer;
   detail::vformat_to(buffer, format_str,
@@ -1417,10 +2769,10 @@ FMT_FUNC void vprint(std::FILE* f, string_view format_str, format_args args) {
   auto fd = _fileno(f);
   if (_isatty(fd)) {
     detail::utf8_to_utf16 u16(string_view(buffer.data(), buffer.size()));
-    auto written = DWORD();
-    if (!WriteConsoleW(reinterpret_cast<HANDLE>(_get_osfhandle(fd)),
-                       u16.c_str(), static_cast<DWORD>(u16.size()), &written,
-                       nullptr)) {
+    auto written = detail::dword();
+    if (!detail::WriteConsoleW(reinterpret_cast<void*>(_get_osfhandle(fd)),
+                               u16.c_str(), static_cast<uint32_t>(u16.size()),
+                               &written, nullptr)) {
       FMT_THROW(format_error("failed to write to console"));
     }
     return;
@@ -1446,8 +2798,4 @@ FMT_FUNC void vprint(string_view format_str, format_args args) {
 
 FMT_END_NAMESPACE
 
-#ifdef _MSC_VER
-#  pragma warning(pop)
-#endif
-
 #endif  // FMT_FORMAT_INL_H_

diff --git a/src/third_party/fmt/format.h b/src/third_party/fmt/format.h
index 17509b7b45dcac0907c24498f033bc03886b9a0c..1a037b02b7e475fbdff757594b40ffd3cfdcf5a7 100644 (file)
--- a/src/third_party/fmt/format.h
+++ b/src/third_party/fmt/format.h
@@ -70,9 +70,11 @@
 #endif
 
 #if __cplusplus == 201103L || __cplusplus == 201402L
-#  if defined(__clang__)
+#  if defined(__INTEL_COMPILER) || defined(__PGI)
+#    define FMT_FALLTHROUGH
+#  elif defined(__clang__)
 #    define FMT_FALLTHROUGH [[clang::fallthrough]]
-#  elif FMT_GCC_VERSION >= 700 && !defined(__PGI) && \
+#  elif FMT_GCC_VERSION >= 700 && \
       (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 520)
 #    define FMT_FALLTHROUGH [[gnu::fallthrough]]
 #  else
@@ -139,12 +141,13 @@ FMT_END_NAMESPACE
 #endif
 
 #ifndef FMT_USE_UDL_TEMPLATE
-// EDG frontend based compilers (icc, nvcc, etc) and GCC < 6.4 do not properly
-// support UDL templates and GCC >= 9 warns about them.
+// EDG frontend based compilers (icc, nvcc, PGI, etc) and GCC < 6.4 do not
+// properly support UDL templates and GCC >= 9 warns about them.
 #  if FMT_USE_USER_DEFINED_LITERALS &&                         \
       (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 501) && \
       ((FMT_GCC_VERSION >= 604 && __cplusplus >= 201402L) ||   \
-       FMT_CLANG_VERSION >= 304)
+       FMT_CLANG_VERSION >= 304) &&                            \
+      !defined(__PGI) && !defined(__NVCC__)
 #    define FMT_USE_UDL_TEMPLATE 1
 #  else
 #    define FMT_USE_UDL_TEMPLATE 0
@@ -163,6 +166,14 @@ FMT_END_NAMESPACE
 #  define FMT_USE_LONG_DOUBLE 1
 #endif
 
+// Defining FMT_REDUCE_INT_INSTANTIATIONS to 1, will reduce the number of
+// int_writer template instances to just one by only using the largest integer
+// type. This results in a reduction in binary size but will cause a decrease in
+// integer formatting performance.
+#if !defined(FMT_REDUCE_INT_INSTANTIATIONS)
+#  define FMT_REDUCE_INT_INSTANTIATIONS 0
+#endif
+
 // __builtin_clz is broken in clang with Microsoft CodeGen:
 // https://github.com/fmtlib/fmt/issues/519
 #if (FMT_GCC_VERSION || FMT_HAS_BUILTIN(__builtin_clz)) && !FMT_MSC_VER
@@ -171,56 +182,87 @@ FMT_END_NAMESPACE
 #if (FMT_GCC_VERSION || FMT_HAS_BUILTIN(__builtin_clzll)) && !FMT_MSC_VER
 #  define FMT_BUILTIN_CLZLL(n) __builtin_clzll(n)
 #endif
+#if (FMT_GCC_VERSION || FMT_HAS_BUILTIN(__builtin_ctz))
+#  define FMT_BUILTIN_CTZ(n) __builtin_ctz(n)
+#endif
+#if (FMT_GCC_VERSION || FMT_HAS_BUILTIN(__builtin_ctzll))
+#  define FMT_BUILTIN_CTZLL(n) __builtin_ctzll(n)
+#endif
+
+#if FMT_MSC_VER
+#  include <intrin.h>  // _BitScanReverse[64], _BitScanForward[64], _umul128
+#endif
 
 // Some compilers masquerade as both MSVC and GCC-likes or otherwise support
 // __builtin_clz and __builtin_clzll, so only define FMT_BUILTIN_CLZ using the
 // MSVC intrinsics if the clz and clzll builtins are not available.
-#if FMT_MSC_VER && !defined(FMT_BUILTIN_CLZLL) && !defined(_MANAGED)
-#  include <intrin.h>  // _BitScanReverse, _BitScanReverse64
-
+#if FMT_MSC_VER && !defined(FMT_BUILTIN_CLZLL) && \
+    !defined(FMT_BUILTIN_CTZLL) && !defined(_MANAGED)
 FMT_BEGIN_NAMESPACE
 namespace detail {
 // Avoid Clang with Microsoft CodeGen's -Wunknown-pragmas warning.
 #  ifndef __clang__
+#    pragma intrinsic(_BitScanForward)
 #    pragma intrinsic(_BitScanReverse)
 #  endif
-inline uint32_t clz(uint32_t x) {
+#  if defined(_WIN64) && !defined(__clang__)
+#    pragma intrinsic(_BitScanForward64)
+#    pragma intrinsic(_BitScanReverse64)
+#  endif
+
+inline int clz(uint32_t x) {
   unsigned long r = 0;
   _BitScanReverse(&r, x);
-
   FMT_ASSERT(x != 0, "");
   // Static analysis complains about using uninitialized data
   // "r", but the only way that can happen is if "x" is 0,
   // which the callers guarantee to not happen.
   FMT_SUPPRESS_MSC_WARNING(6102)
-  return 31 - r;
+  return 31 ^ static_cast<int>(r);
 }
 #  define FMT_BUILTIN_CLZ(n) detail::clz(n)
 
-#  if defined(_WIN64) && !defined(__clang__)
-#    pragma intrinsic(_BitScanReverse64)
-#  endif
-
-inline uint32_t clzll(uint64_t x) {
+inline int clzll(uint64_t x) {
   unsigned long r = 0;
 #  ifdef _WIN64
   _BitScanReverse64(&r, x);
 #  else
   // Scan the high 32 bits.
-  if (_BitScanReverse(&r, static_cast<uint32_t>(x >> 32))) return 63 - (r + 32);
-
+  if (_BitScanReverse(&r, static_cast<uint32_t>(x >> 32))) return 63 ^ (r + 32);
   // Scan the low 32 bits.
   _BitScanReverse(&r, static_cast<uint32_t>(x));
 #  endif
-
   FMT_ASSERT(x != 0, "");
-  // Static analysis complains about using uninitialized data
-  // "r", but the only way that can happen is if "x" is 0,
-  // which the callers guarantee to not happen.
-  FMT_SUPPRESS_MSC_WARNING(6102)
-  return 63 - r;
+  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+  return 63 ^ static_cast<int>(r);
 }
 #  define FMT_BUILTIN_CLZLL(n) detail::clzll(n)
+
+inline int ctz(uint32_t x) {
+  unsigned long r = 0;
+  _BitScanForward(&r, x);
+  FMT_ASSERT(x != 0, "");
+  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+  return static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CTZ(n) detail::ctz(n)
+
+inline int ctzll(uint64_t x) {
+  unsigned long r = 0;
+  FMT_ASSERT(x != 0, "");
+  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+#  ifdef _WIN64
+  _BitScanForward64(&r, x);
+#  else
+  // Scan the low 32 bits.
+  if (_BitScanForward(&r, static_cast<uint32_t>(x))) return static_cast<int>(r);
+  // Scan the high 32 bits.
+  _BitScanForward(&r, static_cast<uint32_t>(x >> 32));
+  r += 32;
+#  endif
+  return static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CTZLL(n) detail::ctzll(n)
 }  // namespace detail
 FMT_END_NAMESPACE
 #endif
@@ -298,50 +340,11 @@ FMT_INLINE void assume(bool condition) {
 #endif
 }
 
-// A workaround for gcc 4.8 to make void_t work in a SFINAE context.
-template <typename... Ts> struct void_t_impl { using type = void; };
-
-template <typename... Ts>
-using void_t = typename detail::void_t_impl<Ts...>::type;
-
 // An approximation of iterator_t for pre-C++20 systems.
 template <typename T>
 using iterator_t = decltype(std::begin(std::declval<T&>()));
 template <typename T> using sentinel_t = decltype(std::end(std::declval<T&>()));
 
-// Detect the iterator category of *any* given type in a SFINAE-friendly way.
-// Unfortunately, older implementations of std::iterator_traits are not safe
-// for use in a SFINAE-context.
-template <typename It, typename Enable = void>
-struct iterator_category : std::false_type {};
-
-template <typename T> struct iterator_category<T*> {
-  using type = std::random_access_iterator_tag;
-};
-
-template <typename It>
-struct iterator_category<It, void_t<typename It::iterator_category>> {
-  using type = typename It::iterator_category;
-};
-
-// Detect if *any* given type models the OutputIterator concept.
-template <typename It> class is_output_iterator {
-  // Check for mutability because all iterator categories derived from
-  // std::input_iterator_tag *may* also meet the requirements of an
-  // OutputIterator, thereby falling into the category of 'mutable iterators'
-  // [iterator.requirements.general] clause 4. The compiler reveals this
-  // property only at the point of *actually dereferencing* the iterator!
-  template <typename U>
-  static decltype(*(std::declval<U>())) test(std::input_iterator_tag);
-  template <typename U> static char& test(std::output_iterator_tag);
-  template <typename U> static const char& test(...);
-
-  using type = decltype(test<It>(typename iterator_category<It>::type{}));
-
- public:
-  enum { value = !std::is_const<remove_reference_t<type>>::value };
-};
-
 // A workaround for std::string not having mutable data() until C++17.
 template <typename Char> inline Char* get_data(std::basic_string<Char>& s) {
   return &s[0];
@@ -374,10 +377,29 @@ reserve(std::back_insert_iterator<Container> it, size_t n) {
   return make_checked(get_data(c) + size, n);
 }
 
+template <typename T>
+inline buffer_appender<T> reserve(buffer_appender<T> it, size_t n) {
+  buffer<T>& buf = get_container(it);
+  buf.try_reserve(buf.size() + n);
+  return it;
+}
+
 template <typename Iterator> inline Iterator& reserve(Iterator& it, size_t) {
   return it;
 }
 
+template <typename T, typename OutputIt>
+constexpr T* to_pointer(OutputIt, size_t) {
+  return nullptr;
+}
+template <typename T> T* to_pointer(buffer_appender<T> it, size_t n) {
+  buffer<T>& buf = get_container(it);
+  auto size = buf.size();
+  if (buf.capacity() < size + n) return nullptr;
+  buf.try_resize(size + n);
+  return buf.data() + size;
+}
+
 template <typename Container, FMT_ENABLE_IF(is_contiguous<Container>::value)>
 inline std::back_insert_iterator<Container> base_iterator(
     std::back_insert_iterator<Container>& it,
@@ -415,13 +437,17 @@ class counting_iterator {
     ++count_;
     return *this;
   }
-
   counting_iterator operator++(int) {
     auto it = *this;
     ++*this;
     return it;
   }
 
+  friend counting_iterator operator+(counting_iterator it, difference_type n) {
+    it.count_ += static_cast<size_t>(n);
+    return it;
+  }
+
   value_type operator*() const { return {}; }
 };
 
@@ -555,23 +581,38 @@ OutputIt copy_str(InputIt begin, InputIt end, OutputIt it) {
                         [](char c) { return static_cast<char8_type>(c); });
 }
 
-#ifndef FMT_USE_GRISU
-#  define FMT_USE_GRISU 1
-#endif
-
-template <typename T> constexpr bool use_grisu() {
-  return FMT_USE_GRISU && std::numeric_limits<double>::is_iec559 &&
-         sizeof(T) <= sizeof(double);
+template <typename Char, typename InputIt>
+inline counting_iterator copy_str(InputIt begin, InputIt end,
+                                  counting_iterator it) {
+  return it + (end - begin);
 }
 
+template <typename T>
+using is_fast_float = bool_constant<std::numeric_limits<T>::is_iec559 &&
+                                    sizeof(T) <= sizeof(double)>;
+
+#ifndef FMT_USE_FULL_CACHE_DRAGONBOX
+#  define FMT_USE_FULL_CACHE_DRAGONBOX 0
+#endif
+
 template <typename T>
 template <typename U>
 void buffer<T>::append(const U* begin, const U* end) {
-  size_t new_size = size_ + to_unsigned(end - begin);
-  reserve(new_size);
-  std::uninitialized_copy(begin, end,
-                          make_checked(ptr_ + size_, capacity_ - size_));
-  size_ = new_size;
+  do {
+    auto count = to_unsigned(end - begin);
+    try_reserve(size_ + count);
+    auto free_cap = capacity_ - size_;
+    if (free_cap < count) count = free_cap;
+    std::uninitialized_copy_n(begin, count, make_checked(ptr_ + size_, count));
+    size_ += count;
+    begin += count;
+  } while (begin != end);
+}
+
+template <typename OutputIt, typename T, typename Traits>
+void iterator_buffer<OutputIt, T, Traits>::flush() {
+  out_ = std::copy_n(data_, this->limit(this->size()), out_);
+  this->clear();
 }
 }  // namespace detail
 
@@ -610,7 +651,7 @@ enum { inline_buffer_size = 500 };
  */
 template <typename T, size_t SIZE = inline_buffer_size,
           typename Allocator = std::allocator<T>>
-class basic_memory_buffer : public detail::buffer<T> {
+class basic_memory_buffer final : public detail::buffer<T> {
  private:
   T store_[SIZE];
 
@@ -624,7 +665,7 @@ class basic_memory_buffer : public detail::buffer<T> {
   }
 
  protected:
-  void grow(size_t size) FMT_OVERRIDE;
+  void grow(size_t size) final FMT_OVERRIDE;
 
  public:
   using value_type = T;
@@ -634,7 +675,7 @@ class basic_memory_buffer : public detail::buffer<T> {
       : alloc_(alloc) {
     this->set(store_, SIZE);
   }
-  ~basic_memory_buffer() FMT_OVERRIDE { deallocate(); }
+  ~basic_memory_buffer() { deallocate(); }
 
  private:
   // Move data from other to this buffer.
@@ -678,6 +719,22 @@ class basic_memory_buffer : public detail::buffer<T> {
 
   // Returns a copy of the allocator associated with this buffer.
   Allocator get_allocator() const { return alloc_; }
+
+  /**
+    Resizes the buffer to contain *count* elements. If T is a POD type new
+    elements may not be initialized.
+   */
+  void resize(size_t count) { this->try_resize(count); }
+
+  /** Increases the buffer capacity to *new_capacity*. */
+  void reserve(size_t new_capacity) { this->try_reserve(new_capacity); }
+
+  // Directly append data into the buffer
+  using detail::buffer<T>::append;
+  template <typename ContiguousRange>
+  void append(const ContiguousRange& range) {
+    append(range.data(), range.data() + range.size());
+  }
 };
 
 template <typename T, size_t SIZE, typename Allocator>
@@ -748,19 +805,81 @@ FMT_CONSTEXPR bool is_supported_floating_point(T) {
 }
 
 // Smallest of uint32_t, uint64_t, uint128_t that is large enough to
-// represent all values of T.
+// represent all values of an integral type T.
 template <typename T>
 using uint32_or_64_or_128_t =
-    conditional_t<num_bits<T>() <= 32, uint32_t,
+    conditional_t<num_bits<T>() <= 32 && !FMT_REDUCE_INT_INSTANTIATIONS,
+                  uint32_t,
                   conditional_t<num_bits<T>() <= 64, uint64_t, uint128_t>>;
 
+// 128-bit integer type used internally
+struct FMT_EXTERN_TEMPLATE_API uint128_wrapper {
+  uint128_wrapper() = default;
+
+#if FMT_USE_INT128
+  uint128_t internal_;
+
+  uint128_wrapper(uint64_t high, uint64_t low) FMT_NOEXCEPT
+      : internal_{static_cast<uint128_t>(low) |
+                  (static_cast<uint128_t>(high) << 64)} {}
+
+  uint128_wrapper(uint128_t u) : internal_{u} {}
+
+  uint64_t high() const FMT_NOEXCEPT { return uint64_t(internal_ >> 64); }
+  uint64_t low() const FMT_NOEXCEPT { return uint64_t(internal_); }
+
+  uint128_wrapper& operator+=(uint64_t n) FMT_NOEXCEPT {
+    internal_ += n;
+    return *this;
+  }
+#else
+  uint64_t high_;
+  uint64_t low_;
+
+  uint128_wrapper(uint64_t high, uint64_t low) FMT_NOEXCEPT : high_{high},
+                                                              low_{low} {}
+
+  uint64_t high() const FMT_NOEXCEPT { return high_; }
+  uint64_t low() const FMT_NOEXCEPT { return low_; }
+
+  uint128_wrapper& operator+=(uint64_t n) FMT_NOEXCEPT {
+#  if defined(_MSC_VER) && defined(_M_X64)
+    unsigned char carry = _addcarry_u64(0, low_, n, &low_);
+    _addcarry_u64(carry, high_, 0, &high_);
+    return *this;
+#  else
+    uint64_t sum = low_ + n;
+    high_ += (sum < low_ ? 1 : 0);
+    low_ = sum;
+    return *this;
+#  endif
+  }
+#endif
+};
+
+// Table entry type for divisibility test used internally
+template <typename T> struct FMT_EXTERN_TEMPLATE_API divtest_table_entry {
+  T mod_inv;
+  T max_quotient;
+};
+
 // Static data is placed in this class template for the header-only config.
 template <typename T = void> struct FMT_EXTERN_TEMPLATE_API basic_data {
   static const uint64_t powers_of_10_64[];
-  static const uint32_t zero_or_powers_of_10_32[];
-  static const uint64_t zero_or_powers_of_10_64[];
-  static const uint64_t pow10_significands[];
-  static const int16_t pow10_exponents[];
+  static const uint32_t zero_or_powers_of_10_32_new[];
+  static const uint64_t zero_or_powers_of_10_64_new[];
+  static const uint64_t grisu_pow10_significands[];
+  static const int16_t grisu_pow10_exponents[];
+  static const divtest_table_entry<uint32_t> divtest_table_for_pow5_32[];
+  static const divtest_table_entry<uint64_t> divtest_table_for_pow5_64[];
+  static const uint64_t dragonbox_pow10_significands_64[];
+  static const uint128_wrapper dragonbox_pow10_significands_128[];
+  // log10(2) = 0x0.4d104d427de7fbcc...
+  static const uint64_t log10_2_significand = 0x4d104d427de7fbcc;
+#if !FMT_USE_FULL_CACHE_DRAGONBOX
+  static const uint64_t powers_of_5_64[];
+  static const uint32_t dragonbox_pow10_recovery_errors[];
+#endif
   // GCC generates slightly better code for pairs than chars.
   using digit_pair = char[2];
   static const digit_pair digits[];
@@ -772,8 +891,23 @@ template <typename T = void> struct FMT_EXTERN_TEMPLATE_API basic_data {
   static const char signs[];
   static const char left_padding_shifts[5];
   static const char right_padding_shifts[5];
+
+  // DEPRECATED! These are for ABI compatibility.
+  static const uint32_t zero_or_powers_of_10_32[];
+  static const uint64_t zero_or_powers_of_10_64[];
 };
 
+// Maps bsr(n) to ceil(log10(pow(2, bsr(n) + 1) - 1)).
+// This is a function instead of an array to workaround a bug in GCC10 (#1810).
+FMT_INLINE uint16_t bsr2log10(int bsr) {
+  static constexpr uint16_t data[] = {
+      1,  1,  1,  2,  2,  2,  3,  3,  3,  4,  4,  4,  4,  5,  5,  5,
+      6,  6,  6,  7,  7,  7,  7,  8,  8,  8,  9,  9,  9,  10, 10, 10,
+      10, 11, 11, 11, 12, 12, 12, 13, 13, 13, 13, 14, 14, 14, 15, 15,
+      15, 16, 16, 16, 16, 17, 17, 17, 18, 18, 18, 19, 19, 19, 19, 20};
+  return data[bsr];
+}
+
 #ifndef FMT_EXPORTED
 FMT_EXTERN template struct basic_data<void>;
 #endif
@@ -785,10 +919,9 @@ struct data : basic_data<> {};
 // Returns the number of decimal digits in n. Leading zeros are not counted
 // except for n == 0 in which case count_digits returns 1.
 inline int count_digits(uint64_t n) {
-  // Based on http://graphics.stanford.edu/~seander/bithacks.html#IntegerLog10
-  // and the benchmark https://github.com/localvoid/cxx-benchmark-count-digits.
-  int t = (64 - FMT_BUILTIN_CLZLL(n | 1)) * 1233 >> 12;
-  return t - (n < data::zero_or_powers_of_10_64[t]) + 1;
+  // https://github.com/fmtlib/format-benchmark/blob/master/digits10
+  auto t = bsr2log10(FMT_BUILTIN_CLZLL(n | 1) ^ 63);
+  return t - (n < data::zero_or_powers_of_10_64_new[t]);
 }
 #else
 // Fallback version of count_digits used when __builtin_clz is not available.
@@ -838,15 +971,24 @@ template <> int count_digits<4>(detail::fallback_uintptr n);
 
 #if FMT_GCC_VERSION || FMT_CLANG_VERSION
 #  define FMT_ALWAYS_INLINE inline __attribute__((always_inline))
+#elif FMT_MSC_VER
+#  define FMT_ALWAYS_INLINE __forceinline
 #else
-#  define FMT_ALWAYS_INLINE
+#  define FMT_ALWAYS_INLINE inline
+#endif
+
+// To suppress unnecessary security cookie checks
+#if FMT_MSC_VER && !FMT_CLANG_VERSION
+#  define FMT_SAFEBUFFERS __declspec(safebuffers)
+#else
+#  define FMT_SAFEBUFFERS
 #endif
 
 #ifdef FMT_BUILTIN_CLZ
 // Optional version of count_digits for better performance on 32-bit platforms.
 inline int count_digits(uint32_t n) {
-  int t = (32 - FMT_BUILTIN_CLZ(n | 1)) * 1233 >> 12;
-  return t - (n < data::zero_or_powers_of_10_32[t]) + 1;
+  auto t = bsr2log10(FMT_BUILTIN_CLZ(n | 1) ^ 31);
+  return t - (n < data::zero_or_powers_of_10_32_new[t]);
 }
 #endif
 
@@ -893,7 +1035,7 @@ template <typename Char> void copy2(Char* dst, const char* src) {
   *dst++ = static_cast<Char>(*src++);
   *dst = static_cast<Char>(*src);
 }
-inline void copy2(char* dst, const char* src) { memcpy(dst, src, 2); }
+FMT_INLINE void copy2(char* dst, const char* src) { memcpy(dst, src, 2); }
 
 template <typename Iterator> struct format_decimal_result {
   Iterator begin;
@@ -929,11 +1071,10 @@ inline format_decimal_result<Char*> format_decimal(Char* out, UInt value,
 template <typename Char, typename UInt, typename Iterator,
           FMT_ENABLE_IF(!std::is_pointer<remove_cvref_t<Iterator>>::value)>
 inline format_decimal_result<Iterator> format_decimal(Iterator out, UInt value,
-                                                      int num_digits) {
-  // Buffer should be large enough to hold all digits (<= digits10 + 1).
-  enum { max_size = digits10<UInt>() + 1 };
-  Char buffer[2 * max_size];
-  auto end = format_decimal(buffer, value, num_digits).end;
+                                                      int size) {
+  // Buffer is large enough to hold all digits (digits10 + 1).
+  Char buffer[digits10<UInt>() + 1];
+  auto end = format_decimal(buffer, value, size).end;
   return {out, detail::copy_str<Char>(buffer, end, out)};
 }
 
@@ -975,6 +1116,10 @@ Char* format_uint(Char* buffer, detail::fallback_uintptr n, int num_digits,
 
 template <unsigned BASE_BITS, typename Char, typename It, typename UInt>
 inline It format_uint(It out, UInt value, int num_digits, bool upper = false) {
+  if (auto ptr = to_pointer<Char>(out, to_unsigned(num_digits))) {
+    format_uint<BASE_BITS>(ptr, value, num_digits, upper);
+    return out;
+  }
   // Buffer should be large enough to hold all digits (digits / BASE_BITS + 1).
   char buffer[num_bits<UInt>() / BASE_BITS + 1];
   format_uint<BASE_BITS>(buffer, value, num_digits, upper);
@@ -1000,8 +1145,8 @@ template <typename T = void> struct null {};
 template <typename Char> struct fill_t {
  private:
   enum { max_size = 4 };
-  Char data_[max_size];
-  unsigned char size_;
+  Char data_[max_size] = {Char(' '), Char(0), Char(0), Char(0)};
+  unsigned char size_ = 1;
 
  public:
   FMT_CONSTEXPR void operator=(basic_string_view<Char> s) {
@@ -1021,13 +1166,6 @@ template <typename Char> struct fill_t {
   FMT_CONSTEXPR const Char& operator[](size_t index) const {
     return data_[index];
   }
-
-  static FMT_CONSTEXPR fill_t<Char> make() {
-    auto fill = fill_t<Char>();
-    fill[0] = Char(' ');
-    fill.size_ = 1;
-    return fill;
-  }
 };
 }  // namespace detail
 
@@ -1059,13 +1197,84 @@ template <typename Char> struct basic_format_specs {
         type(0),
         align(align::none),
         sign(sign::none),
-        alt(false),
-        fill(detail::fill_t<Char>::make()) {}
+        alt(false) {}
 };
 
 using format_specs = basic_format_specs<char>;
 
 namespace detail {
+namespace dragonbox {
+
+// Type-specific information that Dragonbox uses.
+template <class T> struct float_info;
+
+template <> struct float_info<float> {
+  using carrier_uint = uint32_t;
+  static const int significand_bits = 23;
+  static const int exponent_bits = 8;
+  static const int min_exponent = -126;
+  static const int max_exponent = 127;
+  static const int exponent_bias = -127;
+  static const int decimal_digits = 9;
+  static const int kappa = 1;
+  static const int big_divisor = 100;
+  static const int small_divisor = 10;
+  static const int min_k = -31;
+  static const int max_k = 46;
+  static const int cache_bits = 64;
+  static const int divisibility_check_by_5_threshold = 39;
+  static const int case_fc_pm_half_lower_threshold = -1;
+  static const int case_fc_pm_half_upper_threshold = 6;
+  static const int case_fc_lower_threshold = -2;
+  static const int case_fc_upper_threshold = 6;
+  static const int case_shorter_interval_left_endpoint_lower_threshold = 2;
+  static const int case_shorter_interval_left_endpoint_upper_threshold = 3;
+  static const int shorter_interval_tie_lower_threshold = -35;
+  static const int shorter_interval_tie_upper_threshold = -35;
+  static const int max_trailing_zeros = 7;
+};
+
+template <> struct float_info<double> {
+  using carrier_uint = uint64_t;
+  static const int significand_bits = 52;
+  static const int exponent_bits = 11;
+  static const int min_exponent = -1022;
+  static const int max_exponent = 1023;
+  static const int exponent_bias = -1023;
+  static const int decimal_digits = 17;
+  static const int kappa = 2;
+  static const int big_divisor = 1000;
+  static const int small_divisor = 100;
+  static const int min_k = -292;
+  static const int max_k = 326;
+  static const int cache_bits = 128;
+  static const int divisibility_check_by_5_threshold = 86;
+  static const int case_fc_pm_half_lower_threshold = -2;
+  static const int case_fc_pm_half_upper_threshold = 9;
+  static const int case_fc_lower_threshold = -4;
+  static const int case_fc_upper_threshold = 9;
+  static const int case_shorter_interval_left_endpoint_lower_threshold = 2;
+  static const int case_shorter_interval_left_endpoint_upper_threshold = 3;
+  static const int shorter_interval_tie_lower_threshold = -77;
+  static const int shorter_interval_tie_upper_threshold = -77;
+  static const int max_trailing_zeros = 16;
+};
+
+template <typename T> struct decimal_fp {
+  using significand_type = typename float_info<T>::carrier_uint;
+  significand_type significand;
+  int exponent;
+};
+
+template <typename T> FMT_API decimal_fp<T> to_decimal(T x) FMT_NOEXCEPT;
+}  // namespace dragonbox
+
+template <typename T>
+constexpr typename dragonbox::float_info<T>::carrier_uint exponent_mask() {
+  using uint = typename dragonbox::float_info<T>::carrier_uint;
+  return ((uint(1) << dragonbox::float_info<T>::exponent_bits) - 1)
+         << dragonbox::float_info<T>::significand_bits;
+}
 
 // A floating-point presentation format.
 enum class float_format : unsigned char {
@@ -1107,113 +1316,6 @@ template <typename Char, typename It> It write_exponent(int exp, It it) {
   return it;
 }
 
-template <typename Char> class float_writer {
- private:
-  // The number is given as v = digits_ * pow(10, exp_).
-  const char* digits_;
-  int num_digits_;
-  int exp_;
-  size_t size_;
-  float_specs specs_;
-  Char decimal_point_;
-
-  template <typename It> It prettify(It it) const {
-    // pow(10, full_exp - 1) <= v <= pow(10, full_exp).
-    int full_exp = num_digits_ + exp_;
-    if (specs_.format == float_format::exp) {
-      // Insert a decimal point after the first digit and add an exponent.
-      *it++ = static_cast<Char>(*digits_);
-      int num_zeros = specs_.precision - num_digits_;
-      if (num_digits_ > 1 || specs_.showpoint) *it++ = decimal_point_;
-      it = copy_str<Char>(digits_ + 1, digits_ + num_digits_, it);
-      if (num_zeros > 0 && specs_.showpoint)
-        it = std::fill_n(it, num_zeros, static_cast<Char>('0'));
-      *it++ = static_cast<Char>(specs_.upper ? 'E' : 'e');
-      return write_exponent<Char>(full_exp - 1, it);
-    }
-    if (num_digits_ <= full_exp) {
-      // 1234e7 -> 12340000000[.0+]
-      it = copy_str<Char>(digits_, digits_ + num_digits_, it);
-      it = std::fill_n(it, full_exp - num_digits_, static_cast<Char>('0'));
-      if (specs_.showpoint || specs_.precision < 0) {
-        *it++ = decimal_point_;
-        int num_zeros = specs_.precision - full_exp;
-        if (num_zeros <= 0) {
-          if (specs_.format != float_format::fixed)
-            *it++ = static_cast<Char>('0');
-          return it;
-        }
-#ifdef FMT_FUZZ
-        if (num_zeros > 5000)
-          throw std::runtime_error("fuzz mode - avoiding excessive cpu use");
-#endif
-        it = std::fill_n(it, num_zeros, static_cast<Char>('0'));
-      }
-    } else if (full_exp > 0) {
-      // 1234e-2 -> 12.34[0+]
-      it = copy_str<Char>(digits_, digits_ + full_exp, it);
-      if (!specs_.showpoint) {
-        // Remove trailing zeros.
-        int num_digits = num_digits_;
-        while (num_digits > full_exp && digits_[num_digits - 1] == '0')
-          --num_digits;
-        if (num_digits != full_exp) *it++ = decimal_point_;
-        return copy_str<Char>(digits_ + full_exp, digits_ + num_digits, it);
-      }
-      *it++ = decimal_point_;
-      it = copy_str<Char>(digits_ + full_exp, digits_ + num_digits_, it);
-      if (specs_.precision > num_digits_) {
-        // Add trailing zeros.
-        int num_zeros = specs_.precision - num_digits_;
-        it = std::fill_n(it, num_zeros, static_cast<Char>('0'));
-      }
-    } else {
-      // 1234e-6 -> 0.001234
-      *it++ = static_cast<Char>('0');
-      int num_zeros = -full_exp;
-      int num_digits = num_digits_;
-      if (num_digits == 0 && specs_.precision >= 0 &&
-          specs_.precision < num_zeros) {
-        num_zeros = specs_.precision;
-      }
-      // Remove trailing zeros.
-      if (!specs_.showpoint)
-        while (num_digits > 0 && digits_[num_digits - 1] == '0') --num_digits;
-      if (num_zeros != 0 || num_digits != 0 || specs_.showpoint) {
-        *it++ = decimal_point_;
-        it = std::fill_n(it, num_zeros, static_cast<Char>('0'));
-        it = copy_str<Char>(digits_, digits_ + num_digits, it);
-      }
-    }
-    return it;
-  }
-
- public:
-  float_writer(const char* digits, int num_digits, int exp, float_specs specs,
-               Char decimal_point)
-      : digits_(digits),
-        num_digits_(num_digits),
-        exp_(exp),
-        specs_(specs),
-        decimal_point_(decimal_point) {
-    int full_exp = num_digits + exp - 1;
-    int precision = specs.precision > 0 ? specs.precision : 16;
-    if (specs_.format == float_format::general &&
-        !(full_exp >= -4 && full_exp < precision)) {
-      specs_.format = float_format::exp;
-    }
-    size_ = prettify(counting_iterator()).count();
-    size_ += specs.sign ? 1 : 0;
-  }
-
-  size_t size() const { return size_; }
-
-  template <typename It> It operator()(It it) const {
-    if (specs_.sign) *it++ = static_cast<Char>(data::signs[specs_.sign]);
-    return prettify(it);
-  }
-};
-
 template <typename T>
 int format_float(T value, int precision, float_specs specs, buffer<char>& buf);
 
@@ -1392,7 +1494,7 @@ template <align::type align = align::left, typename OutputIt, typename Char,
           typename F>
 inline OutputIt write_padded(OutputIt out,
                              const basic_format_specs<Char>& specs, size_t size,
-                             size_t width, const F& f) {
+                             size_t width, F&& f) {
   static_assert(align == align::left || align == align::right, "");
   unsigned spec_width = to_unsigned(specs.width);
   size_t padding = spec_width > width ? spec_width - width : 0;
@@ -1410,7 +1512,7 @@ template <align::type align = align::left, typename OutputIt, typename Char,
           typename F>
 inline OutputIt write_padded(OutputIt out,
                              const basic_format_specs<Char>& specs, size_t size,
-                             const F& f) {
+                             F&& f) {
   return write_padded<align>(out, specs, size, size, f);
 }
 
@@ -1577,15 +1679,16 @@ template <typename OutputIt, typename Char, typename UInt> struct int_writer {
     char digits[40];
     format_decimal(digits, abs_value, num_digits);
     basic_memory_buffer<Char> buffer;
-    size += prefix_size;
-    buffer.resize(size);
+    size += static_cast<int>(prefix_size);
+    const auto usize = to_unsigned(size);
+    buffer.resize(usize);
     basic_string_view<Char> s(&sep, sep_size);
     // Index of a decimal digit with the least significant digit having index 0.
     int digit_index = 0;
     group = groups.cbegin();
-    auto p = buffer.data() + size;
-    for (int i = num_digits - 1; i >= 0; --i) {
-      *--p = static_cast<Char>(digits[i]);
+    auto p = buffer.data() + size - 1;
+    for (int i = num_digits - 1; i > 0; --i) {
+      *p-- = static_cast<Char>(digits[i]);
       if (*group <= 0 || ++digit_index % *group != 0 ||
           *group == max_value<char>())
         continue;
@@ -1593,16 +1696,16 @@ template <typename OutputIt, typename Char, typename UInt> struct int_writer {
         digit_index = 0;
         ++group;
       }
-      p -= s.size();
       std::uninitialized_copy(s.data(), s.data() + s.size(),
                               make_checked(p, s.size()));
+      p -= s.size();
     }
-    if (prefix_size != 0) p[-1] = static_cast<Char>('-');
-    using iterator = remove_reference_t<decltype(reserve(out, 0))>;
+    *p-- = static_cast<Char>(*digits);
+    if (prefix_size != 0) *p = static_cast<Char>('-');
     auto data = buffer.data();
-    out = write_padded<align::right>(out, specs, size, size, [=](iterator it) {
-      return copy_str<Char>(data, data + size, it);
-    });
+    out = write_padded<align::right>(
+        out, specs, usize, usize,
+        [=](iterator it) { return copy_str<Char>(data, data + size, it); });
   }
 
   void on_chr() { *out++ = static_cast<Char>(abs_value); }
@@ -1628,6 +1731,168 @@ OutputIt write_nonfinite(OutputIt out, bool isinf,
   });
 }
 
+// A decimal floating-point number significand * pow(10, exp).
+struct big_decimal_fp {
+  const char* significand;
+  int significand_size;
+  int exponent;
+};
+
+inline int get_significand_size(const big_decimal_fp& fp) {
+  return fp.significand_size;
+}
+template <typename T>
+inline int get_significand_size(const dragonbox::decimal_fp<T>& fp) {
+  return count_digits(fp.significand);
+}
+
+template <typename Char, typename OutputIt>
+inline OutputIt write_significand(OutputIt out, const char* significand,
+                                  int& significand_size) {
+  return copy_str<Char>(significand, significand + significand_size, out);
+}
+template <typename Char, typename OutputIt, typename UInt>
+inline OutputIt write_significand(OutputIt out, UInt significand,
+                                  int significand_size) {
+  return format_decimal<Char>(out, significand, significand_size).end;
+}
+
+template <typename Char, typename UInt,
+          FMT_ENABLE_IF(std::is_integral<UInt>::value)>
+inline Char* write_significand(Char* out, UInt significand,
+                               int significand_size, int integral_size,
+                               Char decimal_point) {
+  if (!decimal_point)
+    return format_decimal(out, significand, significand_size).end;
+  auto end = format_decimal(out + 1, significand, significand_size).end;
+  if (integral_size == 1)
+    out[0] = out[1];
+  else
+    std::copy_n(out + 1, integral_size, out);
+  out[integral_size] = decimal_point;
+  return end;
+}
+
+template <typename OutputIt, typename UInt, typename Char,
+          FMT_ENABLE_IF(!std::is_pointer<remove_cvref_t<OutputIt>>::value)>
+inline OutputIt write_significand(OutputIt out, UInt significand,
+                                  int significand_size, int integral_size,
+                                  Char decimal_point) {
+  // Buffer is large enough to hold digits (digits10 + 1) and a decimal point.
+  Char buffer[digits10<UInt>() + 2];
+  auto end = write_significand(buffer, significand, significand_size,
+                               integral_size, decimal_point);
+  return detail::copy_str<Char>(buffer, end, out);
+}
+
+template <typename OutputIt, typename Char>
+inline OutputIt write_significand(OutputIt out, const char* significand,
+                                  int significand_size, int integral_size,
+                                  Char decimal_point) {
+  out = detail::copy_str<Char>(significand, significand + integral_size, out);
+  if (!decimal_point) return out;
+  *out++ = decimal_point;
+  return detail::copy_str<Char>(significand + integral_size,
+                                significand + significand_size, out);
+}
+
+template <typename OutputIt, typename DecimalFP, typename Char>
+OutputIt write_float(OutputIt out, const DecimalFP& fp,
+                     const basic_format_specs<Char>& specs, float_specs fspecs,
+                     Char decimal_point) {
+  auto significand = fp.significand;
+  int significand_size = get_significand_size(fp);
+  static const Char zero = static_cast<Char>('0');
+  auto sign = fspecs.sign;
+  size_t size = to_unsigned(significand_size) + (sign ? 1 : 0);
+  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
+
+  int output_exp = fp.exponent + significand_size - 1;
+  auto use_exp_format = [=]() {
+    if (fspecs.format == float_format::exp) return true;
+    if (fspecs.format != float_format::general) return false;
+    // Use the fixed notation if the exponent is in [exp_lower, exp_upper),
+    // e.g. 0.0001 instead of 1e-04. Otherwise use the exponent notation.
+    const int exp_lower = -4, exp_upper = 16;
+    return output_exp < exp_lower ||
+           output_exp >= (fspecs.precision > 0 ? fspecs.precision : exp_upper);
+  };
+  if (use_exp_format()) {
+    int num_zeros = 0;
+    if (fspecs.showpoint) {
+      num_zeros = (std::max)(fspecs.precision - significand_size, 0);
+      size += to_unsigned(num_zeros);
+    } else if (significand_size == 1) {
+      decimal_point = Char();
+    }
+    auto abs_output_exp = output_exp >= 0 ? output_exp : -output_exp;
+    int exp_digits = 2;
+    if (abs_output_exp >= 100) exp_digits = abs_output_exp >= 1000 ? 4 : 3;
+
+    size += to_unsigned((decimal_point ? 1 : 0) + 2 + exp_digits);
+    char exp_char = fspecs.upper ? 'E' : 'e';
+    auto write = [=](iterator it) {
+      if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+      // Insert a decimal point after the first digit and add an exponent.
+      it = write_significand(it, significand, significand_size, 1,
+                             decimal_point);
+      if (num_zeros > 0) it = std::fill_n(it, num_zeros, zero);
+      *it++ = static_cast<Char>(exp_char);
+      return write_exponent<Char>(output_exp, it);
+    };
+    return specs.width > 0 ? write_padded<align::right>(out, specs, size, write)
+                           : base_iterator(out, write(reserve(out, size)));
+  }
+
+  int exp = fp.exponent + significand_size;
+  if (fp.exponent >= 0) {
+    // 1234e5 -> 123400000[.0+]
+    size += to_unsigned(fp.exponent);
+    int num_zeros = fspecs.precision - exp;
+#ifdef FMT_FUZZ
+    if (num_zeros > 5000)
+      throw std::runtime_error("fuzz mode - avoiding excessive cpu use");
+#endif
+    if (fspecs.showpoint) {
+      if (num_zeros <= 0 && fspecs.format != float_format::fixed) num_zeros = 1;
+      if (num_zeros > 0) size += to_unsigned(num_zeros);
+    }
+    return write_padded<align::right>(out, specs, size, [&](iterator it) {
+      if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+      it = write_significand<Char>(it, significand, significand_size);
+      it = std::fill_n(it, fp.exponent, zero);
+      if (!fspecs.showpoint) return it;
+      *it++ = decimal_point;
+      return num_zeros > 0 ? std::fill_n(it, num_zeros, zero) : it;
+    });
+  } else if (exp > 0) {
+    // 1234e-2 -> 12.34[0+]
+    int num_zeros = fspecs.showpoint ? fspecs.precision - significand_size : 0;
+    size += 1 + to_unsigned(num_zeros > 0 ? num_zeros : 0);
+    return write_padded<align::right>(out, specs, size, [&](iterator it) {
+      if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+      it = write_significand(it, significand, significand_size, exp,
+                             decimal_point);
+      return num_zeros > 0 ? std::fill_n(it, num_zeros, zero) : it;
+    });
+  }
+  // 1234e-6 -> 0.001234
+  int num_zeros = -exp;
+  if (significand_size == 0 && fspecs.precision >= 0 &&
+      fspecs.precision < num_zeros) {
+    num_zeros = fspecs.precision;
+  }
+  size += 2 + to_unsigned(num_zeros);
+  return write_padded<align::right>(out, specs, size, [&](iterator it) {
+    if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+    *it++ = zero;
+    if (num_zeros == 0 && significand_size == 0 && !fspecs.showpoint) return it;
+    *it++ = decimal_point;
+    it = std::fill_n(it, num_zeros, zero);
+    return write_significand<Char>(it, significand, significand_size);
+  });
+}
+
 template <typename Char, typename OutputIt, typename T,
           FMT_ENABLE_IF(std::is_floating_point<T>::value)>
 OutputIt write(OutputIt out, T value, basic_format_specs<Char> specs,
@@ -1667,39 +1932,45 @@ OutputIt write(OutputIt out, T value, basic_format_specs<Char> specs,
       ++precision;
   }
   if (const_check(std::is_same<T, float>())) fspecs.binary32 = true;
-  fspecs.use_grisu = use_grisu<T>();
+  fspecs.use_grisu = is_fast_float<T>();
   int exp = format_float(promote_float(value), precision, fspecs, buffer);
   fspecs.precision = precision;
   Char point =
       fspecs.locale ? decimal_point<Char>(loc) : static_cast<Char>('.');
-  float_writer<Char> w(buffer.data(), static_cast<int>(buffer.size()), exp,
-                       fspecs, point);
-  return write_padded<align::right>(out, specs, w.size(), w);
+  auto fp = big_decimal_fp{buffer.data(), static_cast<int>(buffer.size()), exp};
+  return write_float(out, fp, specs, fspecs, point);
 }
 
 template <typename Char, typename OutputIt, typename T,
-          FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+          FMT_ENABLE_IF(is_fast_float<T>::value)>
 OutputIt write(OutputIt out, T value) {
   if (const_check(!is_supported_floating_point(value))) return out;
+
+  using floaty = conditional_t<std::is_same<T, long double>::value, double, T>;
+  using uint = typename dragonbox::float_info<floaty>::carrier_uint;
+  auto bits = bit_cast<uint>(value);
+
   auto fspecs = float_specs();
-  if (std::signbit(value)) {  // value < 0 is false for NaN so use signbit.
+  auto sign_bit = bits & (uint(1) << (num_bits<uint>() - 1));
+  if (sign_bit != 0) {
     fspecs.sign = sign::minus;
     value = -value;
   }
 
-  auto specs = basic_format_specs<Char>();
-  if (!std::isfinite(value))
+  static const auto specs = basic_format_specs<Char>();
+  uint mask = exponent_mask<floaty>();
+  if ((bits & mask) == mask)
     return write_nonfinite(out, std::isinf(value), specs, fspecs);
 
-  memory_buffer buffer;
-  int precision = -1;
-  if (const_check(std::is_same<T, float>())) fspecs.binary32 = true;
-  fspecs.use_grisu = use_grisu<T>();
-  int exp = format_float(promote_float(value), precision, fspecs, buffer);
-  fspecs.precision = precision;
-  float_writer<Char> w(buffer.data(), static_cast<int>(buffer.size()), exp,
-                       fspecs, static_cast<Char>('.'));
-  return base_iterator(out, w(reserve(out, w.size())));
+  auto dec = dragonbox::to_decimal(static_cast<floaty>(value));
+  return write_float(out, dec, specs, fspecs, static_cast<Char>('.'));
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(std::is_floating_point<T>::value &&
+                        !is_fast_float<T>::value)>
+inline OutputIt write(OutputIt out, T value) {
+  return write(out, value, basic_format_specs<Char>());
 }
 
 template <typename Char, typename OutputIt>
@@ -1752,6 +2023,13 @@ OutputIt write(OutputIt out, basic_string_view<Char> value) {
   return base_iterator(out, it);
 }
 
+template <typename Char>
+buffer_appender<Char> write(buffer_appender<Char> out,
+                            basic_string_view<Char> value) {
+  get_container(out).append(value.begin(), value.end());
+  return out;
+}
+
 template <typename Char, typename OutputIt, typename T,
           FMT_ENABLE_IF(is_integral<T>::value &&
                         !std::is_same<T, bool>::value &&
@@ -1762,7 +2040,13 @@ OutputIt write(OutputIt out, T value) {
   // Don't do -abs_value since it trips unsigned-integer-overflow sanitizer.
   if (negative) abs_value = ~abs_value + 1;
   int num_digits = count_digits(abs_value);
-  auto it = reserve(out, (negative ? 1 : 0) + static_cast<size_t>(num_digits));
+  auto size = (negative ? 1 : 0) + static_cast<size_t>(num_digits);
+  auto it = reserve(out, size);
+  if (auto ptr = to_pointer<Char>(it, size)) {
+    if (negative) *ptr++ = static_cast<Char>('-');
+    format_decimal<Char>(ptr, abs_value, num_digits);
+    return out;
+  }
   if (negative) *it++ = static_cast<Char>('-');
   it = format_decimal<Char>(it, abs_value, num_digits).end;
   return base_iterator(out, it);
@@ -1801,8 +2085,13 @@ auto write(OutputIt out, const T& value) -> typename std::enable_if<
     mapped_type_constant<T, basic_format_context<OutputIt, Char>>::value ==
         type::custom_type,
     OutputIt>::type {
-  basic_format_context<OutputIt, Char> ctx(out, {}, {});
-  return formatter<T>().format(value, ctx);
+  using context_type = basic_format_context<OutputIt, Char>;
+  using formatter_type =
+      conditional_t<has_formatter<T, context_type>::value,
+                    typename context_type::template formatter_type<T>,
+                    fallback_formatter<T, Char>>;
+  context_type ctx(out, {}, {});
+  return formatter_type().format(value, ctx);
 }
 
 // An argument visitor that formats the argument and writes it via the output
@@ -2008,6 +2297,48 @@ class arg_formatter_base {
   }
 };
 
+/** The default argument formatter. */
+template <typename OutputIt, typename Char>
+class arg_formatter : public arg_formatter_base<OutputIt, Char> {
+ private:
+  using char_type = Char;
+  using base = arg_formatter_base<OutputIt, Char>;
+  using context_type = basic_format_context<OutputIt, Char>;
+
+  context_type& ctx_;
+  basic_format_parse_context<char_type>* parse_ctx_;
+  const Char* ptr_;
+
+ public:
+  using iterator = typename base::iterator;
+  using format_specs = typename base::format_specs;
+
+  /**
+    \rst
+    Constructs an argument formatter object.
+    *ctx* is a reference to the formatting context,
+    *specs* contains format specifier information for standard argument types.
+    \endrst
+   */
+  explicit arg_formatter(
+      context_type& ctx,
+      basic_format_parse_context<char_type>* parse_ctx = nullptr,
+      format_specs* specs = nullptr, const Char* ptr = nullptr)
+      : base(ctx.out(), specs, ctx.locale()),
+        ctx_(ctx),
+        parse_ctx_(parse_ctx),
+        ptr_(ptr) {}
+
+  using base::operator();
+
+  /** Formats an argument of a user-defined type. */
+  iterator operator()(typename basic_format_arg<context_type>::handle handle) {
+    if (ptr_) advance_to(*parse_ctx_, ptr_);
+    handle.format(*parse_ctx_, ctx_);
+    return ctx_.out();
+  }
+};
+
 template <typename Char> FMT_CONSTEXPR bool is_name_start(Char c) {
   return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || '_' == c;
 }
@@ -2047,12 +2378,11 @@ template <typename Context> class custom_formatter {
                             Context& ctx)
       : parse_ctx_(parse_ctx), ctx_(ctx) {}
 
-  bool operator()(typename basic_format_arg<Context>::handle h) const {
+  void operator()(typename basic_format_arg<Context>::handle h) const {
     h.format(parse_ctx_, ctx_);
-    return true;
   }
 
-  template <typename T> bool operator()(T) const { return false; }
+  template <typename T> void operator()(T) const {}
 };
 
 template <typename T>
@@ -2434,12 +2764,30 @@ template <typename SpecHandler, typename Char> struct precision_adapter {
 };
 
 template <typename Char>
-FMT_CONSTEXPR const Char* next_code_point(const Char* begin, const Char* end) {
-  if (const_check(sizeof(Char) != 1) || (*begin & 0x80) == 0) return begin + 1;
-  do {
-    ++begin;
-  } while (begin != end && (*begin & 0xc0) == 0x80);
-  return begin;
+FMT_CONSTEXPR int code_point_length(const Char* begin) {
+  if (const_check(sizeof(Char) != 1)) return 1;
+  constexpr char lengths[] = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+                              0, 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 2, 3, 3, 4, 0};
+  int len = lengths[static_cast<unsigned char>(*begin) >> 3];
+
+  // Compute the pointer to the next character early so that the next
+  // iteration can start working on the next character. Neither Clang
+  // nor GCC figure out this reordering on their own.
+  return len + !len;
+}
+
+template <typename Char> constexpr bool is_ascii_letter(Char c) {
+  return (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z');
+}
+
+// Converts a character to ASCII. Returns a number > 127 on conversion failure.
+template <typename Char, FMT_ENABLE_IF(std::is_integral<Char>::value)>
+constexpr Char to_ascii(Char value) {
+  return value;
+}
+template <typename Char, FMT_ENABLE_IF(std::is_enum<Char>::value)>
+constexpr typename std::underlying_type<Char>::type to_ascii(Char value) {
+  return value;
 }
 
 // Parses fill and alignment.
@@ -2448,10 +2796,10 @@ FMT_CONSTEXPR const Char* parse_align(const Char* begin, const Char* end,
                                       Handler&& handler) {
   FMT_ASSERT(begin != end, "");
   auto align = align::none;
-  auto p = next_code_point(begin, end);
-  if (p == end) p = begin;
+  auto p = begin + code_point_length(begin);
+  if (p >= end) p = begin;
   for (;;) {
-    switch (static_cast<char>(*p)) {
+    switch (to_ascii(*p)) {
     case '<':
       align = align::left;
       break;
@@ -2530,13 +2878,13 @@ FMT_CONSTEXPR const Char* parse_precision(const Char* begin, const Char* end,
 template <typename Char, typename SpecHandler>
 FMT_CONSTEXPR const Char* parse_format_specs(const Char* begin, const Char* end,
                                              SpecHandler&& handler) {
-  if (begin == end || *begin == '}') return begin;
+  if (begin == end) return begin;
 
   begin = parse_align(begin, end, handler);
   if (begin == end) return begin;
 
   // Parse sign.
-  switch (static_cast<char>(*begin)) {
+  switch (to_ascii(*begin)) {
   case '+':
     handler.on_plus();
     ++begin;
@@ -2613,7 +2961,7 @@ FMT_CONSTEXPR const Char* parse_replacement_field(const Char* begin,
                                                   Handler&& handler) {
   ++begin;
   if (begin == end) return handler.on_error("invalid format string"), end;
-  if (static_cast<char>(*begin) == '}') {
+  if (*begin == '}') {
     handler.on_replacement_field(handler.on_arg_id(), begin);
   } else if (*begin == '{') {
     handler.on_text(begin, begin + 1);
@@ -2658,17 +3006,17 @@ FMT_CONSTEXPR_DECL FMT_INLINE void parse_format_string(
     return;
   }
   struct writer {
-    FMT_CONSTEXPR void operator()(const Char* begin, const Char* end) {
-      if (begin == end) return;
+    FMT_CONSTEXPR void operator()(const Char* pbegin, const Char* pend) {
+      if (pbegin == pend) return;
       for (;;) {
         const Char* p = nullptr;
-        if (!find<IS_CONSTEXPR>(begin, end, '}', p))
-          return handler_.on_text(begin, end);
+        if (!find<IS_CONSTEXPR>(pbegin, pend, '}', p))
+          return handler_.on_text(pbegin, pend);
         ++p;
-        if (p == end || *p != '}')
+        if (p == pend || *p != '}')
           return handler_.on_error("unmatched '}' in format string");
-        handler_.on_text(begin, p);
-        begin = p + 1;
+        handler_.on_text(pbegin, p);
+        pbegin = p + 1;
       }
     }
     Handler& handler_;
@@ -2699,13 +3047,12 @@ FMT_CONSTEXPR const typename ParseContext::char_type* parse_format_specs(
   return f.parse(ctx);
 }
 
-template <typename ArgFormatter, typename Char, typename Context>
+template <typename OutputIt, typename Char, typename Context>
 struct format_handler : detail::error_handler {
   basic_format_parse_context<Char> parse_context;
   Context context;
 
-  format_handler(typename ArgFormatter::iterator out,
-                 basic_string_view<Char> str,
+  format_handler(OutputIt out, basic_string_view<Char> str,
                  basic_format_args<Context> format_args, detail::locale_ref loc)
       : parse_context(str), context(out, format_args, loc) {}
 
@@ -2728,26 +3075,33 @@ struct format_handler : detail::error_handler {
   FMT_INLINE void on_replacement_field(int id, const Char*) {
     auto arg = get_arg(context, id);
     context.advance_to(visit_format_arg(
-        default_arg_formatter<typename ArgFormatter::iterator, Char>{
-            context.out(), context.args(), context.locale()},
+        default_arg_formatter<OutputIt, Char>{context.out(), context.args(),
+                                              context.locale()},
         arg));
   }
 
   const Char* on_format_specs(int id, const Char* begin, const Char* end) {
-    advance_to(parse_context, begin);
     auto arg = get_arg(context, id);
-    custom_formatter<Context> f(parse_context, context);
-    if (visit_format_arg(f, arg)) return parse_context.begin();
-    basic_format_specs<Char> specs;
-    using parse_context_t = basic_format_parse_context<Char>;
-    specs_checker<specs_handler<parse_context_t, Context>> handler(
-        specs_handler<parse_context_t, Context>(specs, parse_context, context),
-        arg.type());
-    begin = parse_format_specs(begin, end, handler);
-    if (begin == end || *begin != '}') on_error("missing '}' in format string");
-    advance_to(parse_context, begin);
-    context.advance_to(
-        visit_format_arg(ArgFormatter(context, &parse_context, &specs), arg));
+    if (arg.type() == type::custom_type) {
+      advance_to(parse_context, begin);
+      visit_format_arg(custom_formatter<Context>(parse_context, context), arg);
+      return parse_context.begin();
+    }
+    auto specs = basic_format_specs<Char>();
+    if (begin + 1 < end && begin[1] == '}' && is_ascii_letter(*begin)) {
+      specs.type = static_cast<char>(*begin++);
+    } else {
+      using parse_context_t = basic_format_parse_context<Char>;
+      specs_checker<specs_handler<parse_context_t, Context>> handler(
+          specs_handler<parse_context_t, Context>(specs, parse_context,
+                                                  context),
+          arg.type());
+      begin = parse_format_specs(begin, end, handler);
+      if (begin == end || *begin != '}')
+        on_error("missing '}' in format string");
+    }
+    context.advance_to(visit_format_arg(
+        arg_formatter<OutputIt, Char>(context, &parse_context, &specs), arg));
     return begin;
   }
 };
@@ -2899,53 +3253,11 @@ FMT_API void format_error_code(buffer<char>& out, int error_code,
 
 FMT_API void report_error(format_func func, int error_code,
                           string_view message) FMT_NOEXCEPT;
-
-/** The default argument formatter. */
-template <typename OutputIt, typename Char>
-class arg_formatter : public arg_formatter_base<OutputIt, Char> {
- private:
-  using char_type = Char;
-  using base = arg_formatter_base<OutputIt, Char>;
-  using context_type = basic_format_context<OutputIt, Char>;
-
-  context_type& ctx_;
-  basic_format_parse_context<char_type>* parse_ctx_;
-  const Char* ptr_;
-
- public:
-  using iterator = typename base::iterator;
-  using format_specs = typename base::format_specs;
-
-  /**
-    \rst
-    Constructs an argument formatter object.
-    *ctx* is a reference to the formatting context,
-    *specs* contains format specifier information for standard argument types.
-    \endrst
-   */
-  explicit arg_formatter(
-      context_type& ctx,
-      basic_format_parse_context<char_type>* parse_ctx = nullptr,
-      format_specs* specs = nullptr, const Char* ptr = nullptr)
-      : base(ctx.out(), specs, ctx.locale()),
-        ctx_(ctx),
-        parse_ctx_(parse_ctx),
-        ptr_(ptr) {}
-
-  using base::operator();
-
-  /** Formats an argument of a user-defined type. */
-  iterator operator()(typename basic_format_arg<context_type>::handle handle) {
-    if (ptr_) advance_to(*parse_ctx_, ptr_);
-    handle.format(*parse_ctx_, ctx_);
-    return ctx_.out();
-  }
-};
 }  // namespace detail
 
 template <typename OutputIt, typename Char>
 using arg_formatter FMT_DEPRECATED_ALIAS =
-  detail::arg_formatter<OutputIt, Char>;
+    detail::arg_formatter<OutputIt, Char>;
 
 /**
  An error returned by an operating system or a language runtime,
@@ -3208,8 +3520,10 @@ struct formatter<Char[N], Char> : formatter<basic_string_view<Char>, Char> {
 //   using variant = std::variant<int, std::string>;
 //   template <>
 //   struct formatter<variant>: dynamic_formatter<> {
-//     void format(buffer &buf, const variant &v, context &ctx) {
-//       visit([&](const auto &val) { format(buf, val, ctx); }, v);
+//     auto format(const variant& v, format_context& ctx) {
+//       return visit([&](const auto& val) {
+//           return dynamic_formatter<>::format(val, ctx);
+//       }, v);
 //     }
 //   };
 template <typename Char = char> class dynamic_formatter {
@@ -3277,28 +3591,15 @@ FMT_CONSTEXPR void advance_to(
   ctx.advance_to(ctx.begin() + (p - &*ctx.begin()));
 }
 
-/** Formats arguments and writes the output to the range. */
-template <typename ArgFormatter, typename Char, typename Context>
-typename Context::iterator vformat_to(
-    typename ArgFormatter::iterator out, basic_string_view<Char> format_str,
-    basic_format_args<Context> args,
-    detail::locale_ref loc = detail::locale_ref()) {
-  if (format_str.size() == 2 && detail::equal2(format_str.data(), "{}")) {
-    auto arg = args.get(0);
-    if (!arg) detail::error_handler().on_error("argument not found");
-    using iterator = typename ArgFormatter::iterator;
-    return visit_format_arg(
-        detail::default_arg_formatter<iterator, Char>{out, args, loc}, arg);
-  }
-  detail::format_handler<ArgFormatter, Char, Context> h(out, format_str, args,
-                                                        loc);
-  detail::parse_format_string<false>(format_str, h);
-  return h.context.out();
-}
+/**
+  \rst
+  Converts ``p`` to ``const void*`` for pointer formatting.
 
-// Casts ``p`` to ``const void*`` for pointer formatting.
-// Example:
-//   auto s = format("{}", ptr(p));
+  **Example**::
+
+    auto s = fmt::format("{}", fmt::ptr(p));
+  \endrst
+ */
 template <typename T> inline const void* ptr(const T* p) { return p; }
 template <typename T> inline const void* ptr(const std::unique_ptr<T>& p) {
   return p.get();
@@ -3317,6 +3618,10 @@ class bytes {
 };
 
 template <> struct formatter<bytes> {
+ private:
+  detail::dynamic_format_specs<char> specs_;
+
+ public:
   template <typename ParseContext>
   FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
     using handler_type = detail::dynamic_specs_handler<ParseContext>;
@@ -3335,9 +3640,6 @@ template <> struct formatter<bytes> {
         specs_.precision, specs_.precision_ref, ctx);
     return detail::write_bytes(ctx.out(), b.data_, specs_);
   }
-
- private:
-  detail::dynamic_format_specs<char> specs_;
 };
 
 template <typename It, typename Sentinel, typename Char>
@@ -3402,15 +3704,14 @@ arg_join<It, Sentinel, wchar_t> join(It begin, Sentinel end, wstring_view sep) {
   \endrst
  */
 template <typename Range>
-arg_join<detail::iterator_t<const Range>, detail::sentinel_t<const Range>, char>
-join(const Range& range, string_view sep) {
+arg_join<detail::iterator_t<Range>, detail::sentinel_t<Range>, char> join(
+    Range&& range, string_view sep) {
   return join(std::begin(range), std::end(range), sep);
 }
 
 template <typename Range>
-arg_join<detail::iterator_t<const Range>, detail::sentinel_t<const Range>,
-         wchar_t>
-join(const Range& range, wstring_view sep) {
+arg_join<detail::iterator_t<Range>, detail::sentinel_t<Range>, wchar_t> join(
+    Range&& range, wstring_view sep) {
   return join(std::begin(range), std::end(range), sep);
 }
 
@@ -3437,7 +3738,7 @@ inline std::string to_string(T value) {
   // The buffer should be large enough to store the number including the sign or
   // "false" for bool.
   constexpr int max_size = detail::digits10<T>() + 2;
-  char buffer[max_size > 5 ? max_size : 5];
+  char buffer[max_size > 5 ? static_cast<unsigned>(max_size) : 5];
   char* begin = buffer;
   return std::string(begin, detail::write<char>(begin, value));
 }
@@ -3457,18 +3758,30 @@ std::basic_string<Char> to_string(const basic_memory_buffer<Char, SIZE>& buf) {
 }
 
 template <typename Char>
-typename buffer_context<Char>::iterator detail::vformat_to(
+void detail::vformat_to(
     detail::buffer<Char>& buf, basic_string_view<Char> format_str,
-    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
-  using af = arg_formatter<typename buffer_context<Char>::iterator, Char>;
-  return vformat_to<af>(std::back_inserter(buf), to_string_view(format_str),
-                        args);
+    basic_format_args<buffer_context<type_identity_t<Char>>> args,
+    detail::locale_ref loc) {
+  using iterator = typename buffer_context<Char>::iterator;
+  auto out = buffer_appender<Char>(buf);
+  if (format_str.size() == 2 && equal2(format_str.data(), "{}")) {
+    auto arg = args.get(0);
+    if (!arg) error_handler().on_error("argument not found");
+    visit_format_arg(default_arg_formatter<iterator, Char>{out, args, loc},
+                     arg);
+    return;
+  }
+  format_handler<iterator, Char, buffer_context<Char>> h(out, format_str, args,
+                                                         loc);
+  parse_format_string<false>(format_str, h);
 }
 
 #ifndef FMT_HEADER_ONLY
-extern template format_context::iterator detail::vformat_to(
-    detail::buffer<char>&, string_view, basic_format_args<format_context>);
+extern template void detail::vformat_to(detail::buffer<char>&, string_view,
+                                        basic_format_args<format_context>,
+                                        detail::locale_ref);
 namespace detail {
+
 extern template FMT_API std::string grouping_impl<char>(locale_ref loc);
 extern template FMT_API std::string grouping_impl<wchar_t>(locale_ref loc);
 extern template FMT_API char thousands_sep_impl<char>(locale_ref loc);
@@ -3494,7 +3807,7 @@ extern template int snprintf_float<long double>(long double value,
 
 template <typename S, typename Char = char_t<S>,
           FMT_ENABLE_IF(detail::is_string<S>::value)>
-inline typename FMT_BUFFER_CONTEXT(Char)::iterator vformat_to(
+inline void vformat_to(
     detail::buffer<Char>& buf, const S& format_str,
     basic_format_args<FMT_BUFFER_CONTEXT(type_identity_t<Char>)> args) {
   return detail::vformat_to(buf, to_string_view(format_str), args);
@@ -3504,10 +3817,9 @@ template <typename S, typename... Args, size_t SIZE = inline_buffer_size,
           typename Char = enable_if_t<detail::is_string<S>::value, char_t<S>>>
 inline typename buffer_context<Char>::iterator format_to(
     basic_memory_buffer<Char, SIZE>& buf, const S& format_str, Args&&... args) {
-  detail::check_format_string<Args...>(format_str);
-  using context = buffer_context<Char>;
-  return detail::vformat_to(buf, to_string_view(format_str),
-                            make_format_args<context>(args...));
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
+  detail::vformat_to(buf, to_string_view(format_str), vargs);
+  return detail::buffer_appender<Char>(buf);
 }
 
 template <typename OutputIt, typename Char = char>
@@ -3516,88 +3828,17 @@ using format_context_t = basic_format_context<OutputIt, Char>;
 template <typename OutputIt, typename Char = char>
 using format_args_t = basic_format_args<format_context_t<OutputIt, Char>>;
 
-template <
-    typename S, typename OutputIt, typename... Args,
-    FMT_ENABLE_IF(detail::is_output_iterator<OutputIt>::value &&
-                  !detail::is_contiguous_back_insert_iterator<OutputIt>::value)>
-inline OutputIt vformat_to(
-    OutputIt out, const S& format_str,
-    format_args_t<type_identity_t<OutputIt>, char_t<S>> args) {
-  using af = detail::arg_formatter<OutputIt, char_t<S>>;
-  return vformat_to<af>(out, to_string_view(format_str), args);
-}
-
-/**
- \rst
- Formats arguments, writes the result to the output iterator ``out`` and returns
- the iterator past the end of the output range.
-
- **Example**::
-
-   std::vector<char> out;
-   fmt::format_to(std::back_inserter(out), "{}", 42);
- \endrst
- */
-template <typename OutputIt, typename S, typename... Args,
-          FMT_ENABLE_IF(
-              detail::is_output_iterator<OutputIt>::value &&
-              !detail::is_contiguous_back_insert_iterator<OutputIt>::value &&
-              detail::is_string<S>::value)>
-inline OutputIt format_to(OutputIt out, const S& format_str, Args&&... args) {
-  detail::check_format_string<Args...>(format_str);
-  using context = format_context_t<OutputIt, char_t<S>>;
-  return vformat_to(out, to_string_view(format_str),
-                    make_format_args<context>(args...));
-}
-
-template <typename OutputIt> struct format_to_n_result {
-  /** Iterator past the end of the output range. */
-  OutputIt out;
-  /** Total (not truncated) output size. */
-  size_t size;
-};
-
 template <typename OutputIt, typename Char = typename OutputIt::value_type>
-using format_to_n_context =
-    format_context_t<detail::truncating_iterator<OutputIt>, Char>;
+using format_to_n_context FMT_DEPRECATED_ALIAS = buffer_context<Char>;
 
 template <typename OutputIt, typename Char = typename OutputIt::value_type>
-using format_to_n_args = basic_format_args<format_to_n_context<OutputIt, Char>>;
+using format_to_n_args FMT_DEPRECATED_ALIAS =
+    basic_format_args<buffer_context<Char>>;
 
 template <typename OutputIt, typename Char, typename... Args>
-inline format_arg_store<format_to_n_context<OutputIt, Char>, Args...>
+FMT_DEPRECATED format_arg_store<buffer_context<Char>, Args...>
 make_format_to_n_args(const Args&... args) {
-  return format_arg_store<format_to_n_context<OutputIt, Char>, Args...>(
-      args...);
-}
-
-template <typename OutputIt, typename Char, typename... Args,
-          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt>::value)>
-inline format_to_n_result<OutputIt> vformat_to_n(
-    OutputIt out, size_t n, basic_string_view<Char> format_str,
-    format_to_n_args<type_identity_t<OutputIt>, type_identity_t<Char>> args) {
-  auto it = vformat_to(detail::truncating_iterator<OutputIt>(out, n),
-                       format_str, args);
-  return {it.base(), it.count()};
-}
-
-/**
- \rst
- Formats arguments, writes up to ``n`` characters of the result to the output
- iterator ``out`` and returns the total output size and the iterator past the
- end of the output range.
- \endrst
- */
-template <typename OutputIt, typename S, typename... Args,
-          FMT_ENABLE_IF(detail::is_string<S>::value&&
-                            detail::is_output_iterator<OutputIt>::value)>
-inline format_to_n_result<OutputIt> format_to_n(OutputIt out, size_t n,
-                                                const S& format_str,
-                                                const Args&... args) {
-  detail::check_format_string<Args...>(format_str);
-  using context = format_to_n_context<OutputIt, char_t<S>>;
-  return vformat_to_n(out, n, to_string_view(format_str),
-                      make_format_args<context>(args...));
+  return format_arg_store<buffer_context<Char>, Args...>(args...);
 }
 
 template <typename Char, enable_if_t<(!std::is_same<Char, char>::value), int>>
@@ -3609,15 +3850,6 @@ std::basic_string<Char> detail::vformat(
   return to_string(buffer);
 }
 
-/**
-  Returns the number of characters in the output of
-  ``format(format_str, args...)``.
- */
-template <typename... Args>
-inline size_t formatted_size(string_view format_str, const Args&... args) {
-  return format_to(detail::counting_iterator(), format_str, args...).count();
-}
-
 template <typename Char, FMT_ENABLE_IF(std::is_same<Char, wchar_t>::value)>
 void vprint(std::FILE* f, basic_string_view<Char> format_str,
             wformat_args args) {
@@ -3642,8 +3874,7 @@ template <typename Char, Char... CHARS> class udl_formatter {
   template <typename... Args>
   std::basic_string<Char> operator()(Args&&... args) const {
     static FMT_CONSTEXPR_DECL Char s[] = {CHARS..., '\0'};
-    check_format_string<remove_cvref_t<Args>...>(FMT_STRING(s));
-    return format(s, std::forward<Args>(args)...);
+    return format(FMT_STRING(s), std::forward<Args>(args)...);
   }
 };
 #  else

diff --git a/src/third_party/format.cpp b/src/third_party/format.cpp
index a64a1f3893a59297899f13d8382f9c79838c1ba7..6141d964a791203273b32d89edf888b47b35590b 100644 (file)
--- a/src/third_party/format.cpp
+++ b/src/third_party/format.cpp
@@ -23,6 +23,36 @@ int format_float(char* buf, std::size_t size, const char* format, int precision,
   return precision < 0 ? snprintf_ptr(buf, size, format, value)
                        : snprintf_ptr(buf, size, format, precision, value);
 }
+
+template FMT_API dragonbox::decimal_fp<float> dragonbox::to_decimal(float x)
+    FMT_NOEXCEPT;
+template FMT_API dragonbox::decimal_fp<double> dragonbox::to_decimal(double x)
+    FMT_NOEXCEPT;
+
+// DEPRECATED! This function exists for ABI compatibility.
+template <typename Char>
+typename basic_format_context<std::back_insert_iterator<buffer<Char>>,
+                              Char>::iterator
+vformat_to(buffer<Char>& buf, basic_string_view<Char> format_str,
+           basic_format_args<basic_format_context<
+               std::back_insert_iterator<buffer<type_identity_t<Char>>>,
+               type_identity_t<Char>>>
+               args) {
+  using iterator = std::back_insert_iterator<buffer<char>>;
+  using context = basic_format_context<
+      std::back_insert_iterator<buffer<type_identity_t<Char>>>,
+      type_identity_t<Char>>;
+  auto out = iterator(buf);
+  format_handler<iterator, Char, context> h(out, format_str, args, {});
+  parse_format_string<false>(format_str, h);
+  return out;
+}
+template basic_format_context<std::back_insert_iterator<buffer<char>>,
+                              char>::iterator
+vformat_to(buffer<char>&, string_view,
+           basic_format_args<basic_format_context<
+               std::back_insert_iterator<buffer<type_identity_t<char>>>,
+               type_identity_t<char>>>);
 }  // namespace detail
 
 template struct FMT_INSTANTIATION_DEF_API detail::basic_data<void>;
@@ -44,9 +74,9 @@ template FMT_API char detail::decimal_point_impl(locale_ref);
 
 template FMT_API void detail::buffer<char>::append(const char*, const char*);
 
-template FMT_API FMT_BUFFER_CONTEXT(char)::iterator detail::vformat_to(
+template FMT_API void detail::vformat_to(
     detail::buffer<char>&, string_view,
-    basic_format_args<FMT_BUFFER_CONTEXT(char)>);
+    basic_format_args<FMT_BUFFER_CONTEXT(char)>, detail::locale_ref);
 
 template FMT_API int detail::snprintf_float(double, int, detail::float_specs,
                                             detail::buffer<char>&);