diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/chrono.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/chrono.h
index c965cf781..1a3b8d5e5 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/chrono.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/chrono.h
@@ -8,35 +8,285 @@
 #ifndef FMT_CHRONO_H_
 #define FMT_CHRONO_H_
 
-#include "format.h"
-#include "locale.h"
-
 #include <chrono>
 #include <ctime>
 #include <locale>
 #include <sstream>
 
-// enable safe chrono durations, unless explicitly disabled
+#include "format.h"
+#include "locale.h"
+
+FMT_BEGIN_NAMESPACE
+
+// Enable safe chrono durations, unless explicitly disabled.
 #ifndef FMT_SAFE_DURATION_CAST
 #  define FMT_SAFE_DURATION_CAST 1
 #endif
-
 #if FMT_SAFE_DURATION_CAST
-#  include "safe-duration-cast.h"
-#endif
 
-FMT_BEGIN_NAMESPACE
+// For conversion between std::chrono::durations without undefined
+// behaviour or erroneous results.
+// This is a stripped down version of duration_cast, for inclusion in fmt.
+// See https://github.com/pauldreik/safe_duration_cast
+//
+// Copyright Paul Dreik 2019
+namespace safe_duration_cast {
+
+template <typename To, typename From,
+          FMT_ENABLE_IF(!std::is_same<From, To>::value &&
+                        std::numeric_limits<From>::is_signed ==
+                            std::numeric_limits<To>::is_signed)>
+FMT_CONSTEXPR To lossless_integral_conversion(const From from, int& ec) {
+  ec = 0;
+  using F = std::numeric_limits<From>;
+  using T = std::numeric_limits<To>;
+  static_assert(F::is_integer, "From must be integral");
+  static_assert(T::is_integer, "To must be integral");
+
+  // A and B are both signed, or both unsigned.
+  if (F::digits <= T::digits) {
+    // From fits in To without any problem.
+  } else {
+    // From does not always fit in To, resort to a dynamic check.
+    if (from < (T::min)() || from > (T::max)()) {
+      // outside range.
+      ec = 1;
+      return {};
+    }
+  }
+  return static_cast<To>(from);
+}
+
+/**
+ * converts From to To, without loss. If the dynamic value of from
+ * can't be converted to To without loss, ec is set.
+ */
+template <typename To, typename From,
+          FMT_ENABLE_IF(!std::is_same<From, To>::value &&
+                        std::numeric_limits<From>::is_signed !=
+                            std::numeric_limits<To>::is_signed)>
+FMT_CONSTEXPR To lossless_integral_conversion(const From from, int& ec) {
+  ec = 0;
+  using F = std::numeric_limits<From>;
+  using T = std::numeric_limits<To>;
+  static_assert(F::is_integer, "From must be integral");
+  static_assert(T::is_integer, "To must be integral");
+
+  if (detail::const_check(F::is_signed && !T::is_signed)) {
+    // From may be negative, not allowed!
+    if (fmt::detail::is_negative(from)) {
+      ec = 1;
+      return {};
+    }
+    // From is positive. Can it always fit in To?
+    if (F::digits > T::digits &&
+        from > static_cast<From>(detail::max_value<To>())) {
+      ec = 1;
+      return {};
+    }
+  }
+
+  if (!F::is_signed && T::is_signed && F::digits >= T::digits &&
+      from > static_cast<From>(detail::max_value<To>())) {
+    ec = 1;
+    return {};
+  }
+  return static_cast<To>(from);  // Lossless conversion.
+}
+
+template <typename To, typename From,
+          FMT_ENABLE_IF(std::is_same<From, To>::value)>
+FMT_CONSTEXPR To lossless_integral_conversion(const From from, int& ec) {
+  ec = 0;
+  return from;
+}  // function
+
+// clang-format off
+/**
+ * converts From to To if possible, otherwise ec is set.
+ *
+ * input                            |    output
+ * ---------------------------------|---------------
+ * NaN                              | NaN
+ * Inf                              | Inf
+ * normal, fits in output           | converted (possibly lossy)
+ * normal, does not fit in output   | ec is set
+ * subnormal                        | best effort
+ * -Inf                             | -Inf
+ */
+// clang-format on
+template <typename To, typename From,
+          FMT_ENABLE_IF(!std::is_same<From, To>::value)>
+FMT_CONSTEXPR To safe_float_conversion(const From from, int& ec) {
+  ec = 0;
+  using T = std::numeric_limits<To>;
+  static_assert(std::is_floating_point<From>::value, "From must be floating");
+  static_assert(std::is_floating_point<To>::value, "To must be floating");
+
+  // catch the only happy case
+  if (std::isfinite(from)) {
+    if (from >= T::lowest() && from <= (T::max)()) {
+      return static_cast<To>(from);
+    }
+    // not within range.
+    ec = 1;
+    return {};
+  }
+
+  // nan and inf will be preserved
+  return static_cast<To>(from);
+}  // function
+
+template <typename To, typename From,
+          FMT_ENABLE_IF(std::is_same<From, To>::value)>
+FMT_CONSTEXPR To safe_float_conversion(const From from, int& ec) {
+  ec = 0;
+  static_assert(std::is_floating_point<From>::value, "From must be floating");
+  return from;
+}
+
+/**
+ * safe duration cast between integral durations
+ */
+template <typename To, typename FromRep, typename FromPeriod,
+          FMT_ENABLE_IF(std::is_integral<FromRep>::value),
+          FMT_ENABLE_IF(std::is_integral<typename To::rep>::value)>
+To safe_duration_cast(std::chrono::duration<FromRep, FromPeriod> from,
+                      int& ec) {
+  using From = std::chrono::duration<FromRep, FromPeriod>;
+  ec = 0;
+  // the basic idea is that we need to convert from count() in the from type
+  // to count() in the To type, by multiplying it with this:
+  struct Factor
+      : std::ratio_divide<typename From::period, typename To::period> {};
+
+  static_assert(Factor::num > 0, "num must be positive");
+  static_assert(Factor::den > 0, "den must be positive");
+
+  // the conversion is like this: multiply from.count() with Factor::num
+  // /Factor::den and convert it to To::rep, all this without
+  // overflow/underflow. let's start by finding a suitable type that can hold
+  // both To, From and Factor::num
+  using IntermediateRep =
+      typename std::common_type<typename From::rep, typename To::rep,
+                                decltype(Factor::num)>::type;
+
+  // safe conversion to IntermediateRep
+  IntermediateRep count =
+      lossless_integral_conversion<IntermediateRep>(from.count(), ec);
+  if (ec) return {};
+  // multiply with Factor::num without overflow or underflow
+  if (detail::const_check(Factor::num != 1)) {
+    const auto max1 = detail::max_value<IntermediateRep>() / Factor::num;
+    if (count > max1) {
+      ec = 1;
+      return {};
+    }
+    const auto min1 =
+        (std::numeric_limits<IntermediateRep>::min)() / Factor::num;
+    if (count < min1) {
+      ec = 1;
+      return {};
+    }
+    count *= Factor::num;
+  }
+
+  if (detail::const_check(Factor::den != 1)) count /= Factor::den;
+  auto tocount = lossless_integral_conversion<typename To::rep>(count, ec);
+  return ec ? To() : To(tocount);
+}
+
+/**
+ * safe duration_cast between floating point durations
+ */
+template <typename To, typename FromRep, typename FromPeriod,
+          FMT_ENABLE_IF(std::is_floating_point<FromRep>::value),
+          FMT_ENABLE_IF(std::is_floating_point<typename To::rep>::value)>
+To safe_duration_cast(std::chrono::duration<FromRep, FromPeriod> from,
+                      int& ec) {
+  using From = std::chrono::duration<FromRep, FromPeriod>;
+  ec = 0;
+  if (std::isnan(from.count())) {
+    // nan in, gives nan out. easy.
+    return To{std::numeric_limits<typename To::rep>::quiet_NaN()};
+  }
+  // maybe we should also check if from is denormal, and decide what to do about
+  // it.
+
+  // +-inf should be preserved.
+  if (std::isinf(from.count())) {
+    return To{from.count()};
+  }
+
+  // the basic idea is that we need to convert from count() in the from type
+  // to count() in the To type, by multiplying it with this:
+  struct Factor
+      : std::ratio_divide<typename From::period, typename To::period> {};
+
+  static_assert(Factor::num > 0, "num must be positive");
+  static_assert(Factor::den > 0, "den must be positive");
+
+  // the conversion is like this: multiply from.count() with Factor::num
+  // /Factor::den and convert it to To::rep, all this without
+  // overflow/underflow. let's start by finding a suitable type that can hold
+  // both To, From and Factor::num
+  using IntermediateRep =
+      typename std::common_type<typename From::rep, typename To::rep,
+                                decltype(Factor::num)>::type;
+
+  // force conversion of From::rep -> IntermediateRep to be safe,
+  // even if it will never happen be narrowing in this context.
+  IntermediateRep count =
+      safe_float_conversion<IntermediateRep>(from.count(), ec);
+  if (ec) {
+    return {};
+  }
+
+  // multiply with Factor::num without overflow or underflow
+  if (Factor::num != 1) {
+    constexpr auto max1 = detail::max_value<IntermediateRep>() /
+                          static_cast<IntermediateRep>(Factor::num);
+    if (count > max1) {
+      ec = 1;
+      return {};
+    }
+    constexpr auto min1 = std::numeric_limits<IntermediateRep>::lowest() /
+                          static_cast<IntermediateRep>(Factor::num);
+    if (count < min1) {
+      ec = 1;
+      return {};
+    }
+    count *= static_cast<IntermediateRep>(Factor::num);
+  }
+
+  // this can't go wrong, right? den>0 is checked earlier.
+  if (Factor::den != 1) {
+    using common_t = typename std::common_type<IntermediateRep, intmax_t>::type;
+    count /= static_cast<common_t>(Factor::den);
+  }
+
+  // convert to the to type, safely
+  using ToRep = typename To::rep;
+
+  const ToRep tocount = safe_float_conversion<ToRep>(count, ec);
+  if (ec) {
+    return {};
+  }
+  return To{tocount};
+}
+}  // namespace safe_duration_cast
+#endif
 
 // Prevents expansion of a preceding token as a function-style macro.
 // Usage: f FMT_NOMACRO()
 #define FMT_NOMACRO
 
-namespace internal {
+namespace detail {
 inline null<> localtime_r FMT_NOMACRO(...) { return null<>(); }
 inline null<> localtime_s(...) { return null<>(); }
 inline null<> gmtime_r(...) { return null<>(); }
 inline null<> gmtime_s(...) { return null<>(); }
-}  // namespace internal
+}  // namespace detail
 
 // Thread-safe replacement for std::localtime
 inline std::tm localtime(std::time_t time) {
@@ -47,22 +297,22 @@ inline std::tm localtime(std::time_t time) {
     dispatcher(std::time_t t) : time_(t) {}
 
     bool run() {
-      using namespace fmt::internal;
+      using namespace fmt::detail;
       return handle(localtime_r(&time_, &tm_));
     }
 
     bool handle(std::tm* tm) { return tm != nullptr; }
 
-    bool handle(internal::null<>) {
-      using namespace fmt::internal;
+    bool handle(detail::null<>) {
+      using namespace fmt::detail;
       return fallback(localtime_s(&tm_, &time_));
     }
 
     bool fallback(int res) { return res == 0; }
 
 #if !FMT_MSC_VER
-    bool fallback(internal::null<>) {
-      using namespace fmt::internal;
+    bool fallback(detail::null<>) {
+      using namespace fmt::detail;
       std::tm* tm = std::localtime(&time_);
       if (tm) tm_ = *tm;
       return tm != nullptr;
@@ -75,6 +325,11 @@ inline std::tm localtime(std::time_t time) {
   return lt.tm_;
 }
 
+inline std::tm localtime(
+    std::chrono::time_point<std::chrono::system_clock> time_point) {
+  return localtime(std::chrono::system_clock::to_time_t(time_point));
+}
+
 // Thread-safe replacement for std::gmtime
 inline std::tm gmtime(std::time_t time) {
   struct dispatcher {
@@ -84,21 +339,21 @@ inline std::tm gmtime(std::time_t time) {
     dispatcher(std::time_t t) : time_(t) {}
 
     bool run() {
-      using namespace fmt::internal;
+      using namespace fmt::detail;
       return handle(gmtime_r(&time_, &tm_));
     }
 
     bool handle(std::tm* tm) { return tm != nullptr; }
 
-    bool handle(internal::null<>) {
-      using namespace fmt::internal;
+    bool handle(detail::null<>) {
+      using namespace fmt::detail;
       return fallback(gmtime_s(&tm_, &time_));
     }
 
     bool fallback(int res) { return res == 0; }
 
 #if !FMT_MSC_VER
-    bool fallback(internal::null<>) {
+    bool fallback(detail::null<>) {
       std::tm* tm = std::gmtime(&time_);
       if (tm) tm_ = *tm;
       return tm != nullptr;
@@ -111,17 +366,33 @@ inline std::tm gmtime(std::time_t time) {
   return gt.tm_;
 }
 
-namespace internal {
-inline std::size_t strftime(char* str, std::size_t count, const char* format,
-                            const std::tm* time) {
+inline std::tm gmtime(
+    std::chrono::time_point<std::chrono::system_clock> time_point) {
+  return gmtime(std::chrono::system_clock::to_time_t(time_point));
+}
+
+namespace detail {
+inline size_t strftime(char* str, size_t count, const char* format,
+                       const std::tm* time) {
   return std::strftime(str, count, format, time);
 }
 
-inline std::size_t strftime(wchar_t* str, std::size_t count,
-                            const wchar_t* format, const std::tm* time) {
+inline size_t strftime(wchar_t* str, size_t count, const wchar_t* format,
+                       const std::tm* time) {
   return std::wcsftime(str, count, format, time);
 }
-}  // namespace internal
+}  // namespace detail
+
+template <typename Char>
+struct formatter<std::chrono::time_point<std::chrono::system_clock>, Char>
+    : formatter<std::tm, Char> {
+  template <typename FormatContext>
+  auto format(std::chrono::time_point<std::chrono::system_clock> val,
+              FormatContext& ctx) -> decltype(ctx.out()) {
+    std::tm time = localtime(val);
+    return formatter<std::tm, Char>::format(time, ctx);
+  }
+};
 
 template <typename Char> struct formatter<std::tm, Char> {
   template <typename ParseContext>
@@ -130,7 +401,7 @@ template <typename Char> struct formatter<std::tm, Char> {
     if (it != ctx.end() && *it == ':') ++it;
     auto end = it;
     while (end != ctx.end() && *end != '}') ++end;
-    tm_format.reserve(internal::to_unsigned(end - it + 1));
+    tm_format.reserve(detail::to_unsigned(end - it + 1));
     tm_format.append(it, end);
     tm_format.push_back('\0');
     return end;
@@ -139,11 +410,10 @@ template <typename Char> struct formatter<std::tm, Char> {
   template <typename FormatContext>
   auto format(const std::tm& tm, FormatContext& ctx) -> decltype(ctx.out()) {
     basic_memory_buffer<Char> buf;
-    std::size_t start = buf.size();
+    size_t start = buf.size();
     for (;;) {
-      std::size_t size = buf.capacity() - start;
-      std::size_t count =
-          internal::strftime(&buf[start], size, &tm_format[0], &tm);
+      size_t size = buf.capacity() - start;
+      size_t count = detail::strftime(&buf[start], size, &tm_format[0], &tm);
       if (count != 0) {
         buf.resize(start + count);
         break;
@@ -155,7 +425,7 @@ template <typename Char> struct formatter<std::tm, Char> {
         // https://github.com/fmtlib/fmt/issues/367
         break;
       }
-      const std::size_t MIN_GROWTH = 10;
+      const size_t MIN_GROWTH = 10;
       buf.reserve(buf.capacity() + (size > MIN_GROWTH ? size : MIN_GROWTH));
     }
     return std::copy(buf.begin(), buf.end(), ctx.out());
@@ -164,7 +434,7 @@ template <typename Char> struct formatter<std::tm, Char> {
   basic_memory_buffer<Char> tm_format;
 };
 
-namespace internal {
+namespace detail {
 template <typename Period> FMT_CONSTEXPR const char* get_units() {
   return nullptr;
 }
@@ -220,12 +490,12 @@ FMT_CONSTEXPR const Char* parse_chrono_format(const Char* begin,
       handler.on_text(ptr - 1, ptr);
       break;
     case 'n': {
-      const char newline[] = "\n";
+      const Char newline[] = {'\n'};
       handler.on_text(newline, newline + 1);
       break;
     }
     case 't': {
-      const char tab[] = "\t";
+      const Char tab[] = {'\t'};
       handler.on_text(tab, tab + 1);
       break;
     }
@@ -421,7 +691,7 @@ inline int to_nonnegative_int(T value, int upper) {
 
 template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
 inline T mod(T x, int y) {
-  return x % y;
+  return x % static_cast<T>(y);
 }
 template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
 inline T mod(T x, int y) {
@@ -484,18 +754,36 @@ inline std::chrono::duration<Rep, std::milli> get_milliseconds(
   return std::chrono::duration<Rep, std::milli>(static_cast<Rep>(ms));
 }
 
-template <typename Rep, typename OutputIt>
-OutputIt format_chrono_duration_value(OutputIt out, Rep val, int precision) {
-  if (precision >= 0) return format_to(out, "{:.{}f}", val, precision);
-  return format_to(out, std::is_floating_point<Rep>::value ? "{:g}" : "{}",
+template <typename Char, typename Rep, typename OutputIt>
+OutputIt format_duration_value(OutputIt out, Rep val, int precision) {
+  const Char pr_f[] = {'{', ':', '.', '{', '}', 'f', '}', 0};
+  if (precision >= 0) return format_to(out, pr_f, val, precision);
+  const Char fp_f[] = {'{', ':', 'g', '}', 0};
+  const Char format[] = {'{', '}', 0};
+  return format_to(out, std::is_floating_point<Rep>::value ? fp_f : format,
                    val);
 }
+template <typename Char, typename OutputIt>
+OutputIt copy_unit(string_view unit, OutputIt out, Char) {
+  return std::copy(unit.begin(), unit.end(), out);
+}
 
-template <typename Period, typename OutputIt>
-static OutputIt format_chrono_duration_unit(OutputIt out) {
-  if (const char* unit = get_units<Period>()) return format_to(out, "{}", unit);
-  if (Period::den == 1) return format_to(out, "[{}]s", Period::num);
-  return format_to(out, "[{}/{}]s", Period::num, Period::den);
+template <typename OutputIt>
+OutputIt copy_unit(string_view unit, OutputIt out, wchar_t) {
+  // This works when wchar_t is UTF-32 because units only contain characters
+  // that have the same representation in UTF-16 and UTF-32.
+  utf8_to_utf16 u(unit);
+  return std::copy(u.c_str(), u.c_str() + u.size(), out);
+}
+
+template <typename Char, typename Period, typename OutputIt>
+OutputIt format_duration_unit(OutputIt out) {
+  if (const char* unit = get_units<Period>())
+    return copy_unit(string_view(unit), out, Char());
+  const Char num_f[] = {'[', '{', '}', ']', 's', 0};
+  if (const_check(Period::den == 1)) return format_to(out, num_f, Period::num);
+  const Char num_def_f[] = {'[', '{', '}', '/', '{', '}', ']', 's', 0};
+  return format_to(out, num_def_f, Period::num, Period::den);
 }
 
 template <typename FormatContext, typename OutputIt, typename Rep,
@@ -518,7 +806,10 @@ struct chrono_formatter {
 
   explicit chrono_formatter(FormatContext& ctx, OutputIt o,
                             std::chrono::duration<Rep, Period> d)
-      : context(ctx), out(o), val(d.count()), negative(false) {
+      : context(ctx),
+        out(o),
+        val(static_cast<rep>(d.count())),
+        negative(false) {
     if (d.count() < 0) {
       val = 0 - val;
       negative = true;
@@ -582,24 +873,24 @@ struct chrono_formatter {
   void write(Rep value, int width) {
     write_sign();
     if (isnan(value)) return write_nan();
-    uint32_or_64_t<int> n = to_unsigned(
-        to_nonnegative_int(value, (std::numeric_limits<int>::max)()));
-    int num_digits = internal::count_digits(n);
+    uint32_or_64_or_128_t<int> n =
+        to_unsigned(to_nonnegative_int(value, max_value<int>()));
+    int num_digits = detail::count_digits(n);
     if (width > num_digits) out = std::fill_n(out, width - num_digits, '0');
-    out = format_decimal<char_type>(out, n, num_digits);
+    out = format_decimal<char_type>(out, n, num_digits).end;
   }
 
   void write_nan() { std::copy_n("nan", 3, out); }
   void write_pinf() { std::copy_n("inf", 3, out); }
   void write_ninf() { std::copy_n("-inf", 4, out); }
 
-  void format_localized(const tm& time, const char* format) {
+  void format_localized(const tm& time, char format, char modifier = 0) {
     if (isnan(val)) return write_nan();
     auto locale = context.locale().template get<std::locale>();
     auto& facet = std::use_facet<std::time_put<char_type>>(locale);
     std::basic_ostringstream<char_type> os;
     os.imbue(locale);
-    facet.put(os, os, ' ', &time, format, format + std::strlen(format));
+    facet.put(os, os, ' ', &time, format, modifier);
     auto str = os.str();
     std::copy(str.begin(), str.end(), out);
   }
@@ -629,7 +920,7 @@ struct chrono_formatter {
     if (ns == numeric_system::standard) return write(hour(), 2);
     auto time = tm();
     time.tm_hour = to_nonnegative_int(hour(), 24);
-    format_localized(time, "%OH");
+    format_localized(time, 'H', 'O');
   }
 
   void on_12_hour(numeric_system ns) {
@@ -638,7 +929,7 @@ struct chrono_formatter {
     if (ns == numeric_system::standard) return write(hour12(), 2);
     auto time = tm();
     time.tm_hour = to_nonnegative_int(hour12(), 12);
-    format_localized(time, "%OI");
+    format_localized(time, 'I', 'O');
   }
 
   void on_minute(numeric_system ns) {
@@ -647,7 +938,7 @@ struct chrono_formatter {
     if (ns == numeric_system::standard) return write(minute(), 2);
     auto time = tm();
     time.tm_min = to_nonnegative_int(minute(), 60);
-    format_localized(time, "%OM");
+    format_localized(time, 'M', 'O');
   }
 
   void on_second(numeric_system ns) {
@@ -672,13 +963,12 @@ struct chrono_formatter {
     }
     auto time = tm();
     time.tm_sec = to_nonnegative_int(second(), 60);
-    format_localized(time, "%OS");
+    format_localized(time, 'S', 'O');
   }
 
   void on_12_hour_time() {
     if (handle_nan_inf()) return;
-
-    format_localized(time(), "%r");
+    format_localized(time(), 'r');
   }
 
   void on_24_hour_time() {
@@ -702,25 +992,27 @@ struct chrono_formatter {
 
   void on_am_pm() {
     if (handle_nan_inf()) return;
-    format_localized(time(), "%p");
+    format_localized(time(), 'p');
   }
 
   void on_duration_value() {
     if (handle_nan_inf()) return;
     write_sign();
-    out = format_chrono_duration_value(out, val, precision);
+    out = format_duration_value<char_type>(out, val, precision);
   }
 
-  void on_duration_unit() { out = format_chrono_duration_unit<Period>(out); }
+  void on_duration_unit() {
+    out = format_duration_unit<char_type, Period>(out);
+  }
 };
-}  // namespace internal
+}  // namespace detail
 
 template <typename Rep, typename Period, typename Char>
 struct formatter<std::chrono::duration<Rep, Period>, Char> {
  private:
   basic_format_specs<Char> specs;
   int precision;
-  using arg_ref_type = internal::arg_ref<Char>;
+  using arg_ref_type = detail::arg_ref<Char>;
   arg_ref_type width_ref;
   arg_ref_type precision_ref;
   mutable basic_string_view<Char> format_str;
@@ -728,7 +1020,7 @@ struct formatter<std::chrono::duration<Rep, Period>, Char> {
 
   struct spec_handler {
     formatter& f;
-    basic_parse_context<Char>& context;
+    basic_format_parse_context<Char>& context;
     basic_string_view<Char> format_str;
 
     template <typename Id> FMT_CONSTEXPR arg_ref_type make_arg_ref(Id arg_id) {
@@ -738,19 +1030,18 @@ struct formatter<std::chrono::duration<Rep, Period>, Char> {
 
     FMT_CONSTEXPR arg_ref_type make_arg_ref(basic_string_view<Char> arg_id) {
       context.check_arg_id(arg_id);
-      const auto str_val = internal::string_view_metadata(format_str, arg_id);
-      return arg_ref_type(str_val);
+      return arg_ref_type(arg_id);
     }
 
-    FMT_CONSTEXPR arg_ref_type make_arg_ref(internal::auto_id) {
+    FMT_CONSTEXPR arg_ref_type make_arg_ref(detail::auto_id) {
       return arg_ref_type(context.next_arg_id());
     }
 
     void on_error(const char* msg) { FMT_THROW(format_error(msg)); }
-    void on_fill(Char fill) { f.specs.fill[0] = fill; }
+    void on_fill(basic_string_view<Char> fill) { f.specs.fill = fill; }
     void on_align(align_t align) { f.specs.align = align; }
-    void on_width(unsigned width) { f.specs.width = width; }
-    void on_precision(unsigned precision) { f.precision = precision; }
+    void on_width(int width) { f.specs.width = width; }
+    void on_precision(int _precision) { f.precision = _precision; }
     void end_precision() {}
 
     template <typename Id> void on_dynamic_width(Id arg_id) {
@@ -762,38 +1053,38 @@ struct formatter<std::chrono::duration<Rep, Period>, Char> {
     }
   };
 
-  using iterator = typename basic_parse_context<Char>::iterator;
+  using iterator = typename basic_format_parse_context<Char>::iterator;
   struct parse_range {
     iterator begin;
     iterator end;
   };
 
-  FMT_CONSTEXPR parse_range do_parse(basic_parse_context<Char>& ctx) {
+  FMT_CONSTEXPR parse_range do_parse(basic_format_parse_context<Char>& ctx) {
     auto begin = ctx.begin(), end = ctx.end();
     if (begin == end || *begin == '}') return {begin, begin};
     spec_handler handler{*this, ctx, format_str};
-    begin = internal::parse_align(begin, end, handler);
+    begin = detail::parse_align(begin, end, handler);
     if (begin == end) return {begin, begin};
-    begin = internal::parse_width(begin, end, handler);
+    begin = detail::parse_width(begin, end, handler);
     if (begin == end) return {begin, begin};
     if (*begin == '.') {
       if (std::is_floating_point<Rep>::value)
-        begin = internal::parse_precision(begin, end, handler);
+        begin = detail::parse_precision(begin, end, handler);
       else
         handler.on_error("precision not allowed for this argument type");
     }
-    end = parse_chrono_format(begin, end, internal::chrono_format_checker());
+    end = parse_chrono_format(begin, end, detail::chrono_format_checker());
     return {begin, end};
   }
 
  public:
   formatter() : precision(-1) {}
 
-  FMT_CONSTEXPR auto parse(basic_parse_context<Char>& ctx)
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
       -> decltype(ctx.begin()) {
     auto range = do_parse(ctx);
     format_str = basic_string_view<Char>(
-        &*range.begin, internal::to_unsigned(range.end - range.begin));
+        &*range.begin, detail::to_unsigned(range.end - range.begin));
     return range.end;
   }
 
@@ -804,23 +1095,21 @@ struct formatter<std::chrono::duration<Rep, Period>, Char> {
     // is not specified.
     basic_memory_buffer<Char> buf;
     auto out = std::back_inserter(buf);
-    using range = internal::output_range<decltype(ctx.out()), Char>;
-    internal::basic_writer<range> w(range(ctx.out()));
-    internal::handle_dynamic_spec<internal::width_checker>(
-        specs.width, width_ref, ctx, format_str.begin());
-    internal::handle_dynamic_spec<internal::precision_checker>(
-        precision, precision_ref, ctx, format_str.begin());
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width, width_ref,
+                                                       ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(precision,
+                                                           precision_ref, ctx);
     if (begin == end || *begin == '}') {
-      out = internal::format_chrono_duration_value(out, d.count(), precision);
-      internal::format_chrono_duration_unit<Period>(out);
+      out = detail::format_duration_value<Char>(out, d.count(), precision);
+      detail::format_duration_unit<Char, Period>(out);
     } else {
-      internal::chrono_formatter<FormatContext, decltype(out), Rep, Period> f(
+      detail::chrono_formatter<FormatContext, decltype(out), Rep, Period> f(
           ctx, out, d);
       f.precision = precision;
       parse_chrono_format(begin, end, f);
     }
-    w.write(buf.data(), buf.size(), specs);
-    return w.out();
+    return detail::write(
+        ctx.out(), basic_string_view<Char>(buf.data(), buf.size()), specs);
   }
 };
 
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/color.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/color.h
index d9d315599..94e3419d1 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/color.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/color.h
@@ -198,7 +198,7 @@ struct rgb {
   uint8_t b;
 };
 
-namespace internal {
+namespace detail {
 
 // color is a struct of either a rgb color or a terminal color.
 struct color_type {
@@ -221,7 +221,7 @@ struct color_type {
     uint32_t rgb_color;
   } value;
 };
-}  // namespace internal
+}  // namespace detail
 
 // Experimental text formatting support.
 class text_style {
@@ -298,22 +298,22 @@ class text_style {
   FMT_CONSTEXPR bool has_emphasis() const FMT_NOEXCEPT {
     return static_cast<uint8_t>(ems) != 0;
   }
-  FMT_CONSTEXPR internal::color_type get_foreground() const FMT_NOEXCEPT {
-    assert(has_foreground() && "no foreground specified for this style");
+  FMT_CONSTEXPR detail::color_type get_foreground() const FMT_NOEXCEPT {
+    FMT_ASSERT(has_foreground(), "no foreground specified for this style");
     return foreground_color;
   }
-  FMT_CONSTEXPR internal::color_type get_background() const FMT_NOEXCEPT {
-    assert(has_background() && "no background specified for this style");
+  FMT_CONSTEXPR detail::color_type get_background() const FMT_NOEXCEPT {
+    FMT_ASSERT(has_background(), "no background specified for this style");
     return background_color;
   }
   FMT_CONSTEXPR emphasis get_emphasis() const FMT_NOEXCEPT {
-    assert(has_emphasis() && "no emphasis specified for this style");
+    FMT_ASSERT(has_emphasis(), "no emphasis specified for this style");
     return ems;
   }
 
  private:
   FMT_CONSTEXPR text_style(bool is_foreground,
-                           internal::color_type text_color) FMT_NOEXCEPT
+                           detail::color_type text_color) FMT_NOEXCEPT
       : set_foreground_color(),
         set_background_color(),
         ems() {
@@ -326,23 +326,23 @@ class text_style {
     }
   }
 
-  friend FMT_CONSTEXPR_DECL text_style fg(internal::color_type foreground)
+  friend FMT_CONSTEXPR_DECL text_style fg(detail::color_type foreground)
       FMT_NOEXCEPT;
-  friend FMT_CONSTEXPR_DECL text_style bg(internal::color_type background)
+  friend FMT_CONSTEXPR_DECL text_style bg(detail::color_type background)
       FMT_NOEXCEPT;
 
-  internal::color_type foreground_color;
-  internal::color_type background_color;
+  detail::color_type foreground_color;
+  detail::color_type background_color;
   bool set_foreground_color;
   bool set_background_color;
   emphasis ems;
 };
 
-FMT_CONSTEXPR text_style fg(internal::color_type foreground) FMT_NOEXCEPT {
+FMT_CONSTEXPR text_style fg(detail::color_type foreground) FMT_NOEXCEPT {
   return text_style(/*is_foreground=*/true, foreground);
 }
 
-FMT_CONSTEXPR text_style bg(internal::color_type background) FMT_NOEXCEPT {
+FMT_CONSTEXPR text_style bg(detail::color_type background) FMT_NOEXCEPT {
   return text_style(/*is_foreground=*/false, background);
 }
 
@@ -350,21 +350,21 @@ FMT_CONSTEXPR text_style operator|(emphasis lhs, emphasis rhs) FMT_NOEXCEPT {
   return text_style(lhs) | rhs;
 }
 
-namespace internal {
+namespace detail {
 
 template <typename Char> struct ansi_color_escape {
-  FMT_CONSTEXPR ansi_color_escape(internal::color_type text_color,
+  FMT_CONSTEXPR ansi_color_escape(detail::color_type text_color,
                                   const char* esc) FMT_NOEXCEPT {
     // If we have a terminal color, we need to output another escape code
     // sequence.
     if (!text_color.is_rgb) {
-      bool is_background = esc == internal::data::background_color;
+      bool is_background = esc == detail::data::background_color;
       uint32_t value = text_color.value.term_color;
       // Background ASCII codes are the same as the foreground ones but with
       // 10 more.
       if (is_background) value += 10u;
 
-      std::size_t index = 0;
+      size_t index = 0;
       buffer[index++] = static_cast<Char>('\x1b');
       buffer[index++] = static_cast<Char>('[');
 
@@ -398,7 +398,7 @@ template <typename Char> struct ansi_color_escape {
     if (em_bits & static_cast<uint8_t>(emphasis::strikethrough))
       em_codes[3] = 9;
 
-    std::size_t index = 0;
+    size_t index = 0;
     for (int i = 0; i < 4; ++i) {
       if (!em_codes[i]) continue;
       buffer[index++] = static_cast<Char>('\x1b');
@@ -412,7 +412,7 @@ template <typename Char> struct ansi_color_escape {
 
   FMT_CONSTEXPR const Char* begin() const FMT_NOEXCEPT { return buffer; }
   FMT_CONSTEXPR const Char* end() const FMT_NOEXCEPT {
-    return buffer + std::strlen(buffer);
+    return buffer + std::char_traits<Char>::length(buffer);
   }
 
  private:
@@ -429,14 +429,14 @@ template <typename Char> struct ansi_color_escape {
 
 template <typename Char>
 FMT_CONSTEXPR ansi_color_escape<Char> make_foreground_color(
-    internal::color_type foreground) FMT_NOEXCEPT {
-  return ansi_color_escape<Char>(foreground, internal::data::foreground_color);
+    detail::color_type foreground) FMT_NOEXCEPT {
+  return ansi_color_escape<Char>(foreground, detail::data::foreground_color);
 }
 
 template <typename Char>
 FMT_CONSTEXPR ansi_color_escape<Char> make_background_color(
-    internal::color_type background) FMT_NOEXCEPT {
-  return ansi_color_escape<Char>(background, internal::data::background_color);
+    detail::color_type background) FMT_NOEXCEPT {
+  return ansi_color_escape<Char>(background, detail::data::background_color);
 }
 
 template <typename Char>
@@ -455,90 +455,71 @@ inline void fputs<wchar_t>(const wchar_t* chars, FILE* stream) FMT_NOEXCEPT {
 }
 
 template <typename Char> inline void reset_color(FILE* stream) FMT_NOEXCEPT {
-  fputs(internal::data::reset_color, stream);
+  fputs(detail::data::reset_color, stream);
 }
 
 template <> inline void reset_color<wchar_t>(FILE* stream) FMT_NOEXCEPT {
-  fputs(internal::data::wreset_color, stream);
+  fputs(detail::data::wreset_color, stream);
 }
 
 template <typename Char>
-inline void reset_color(basic_memory_buffer<Char>& buffer) FMT_NOEXCEPT {
+inline void reset_color(buffer<Char>& buffer) FMT_NOEXCEPT {
   const char* begin = data::reset_color;
   const char* end = begin + sizeof(data::reset_color) - 1;
   buffer.append(begin, end);
 }
 
 template <typename Char>
-std::basic_string<Char> vformat(const text_style& ts,
-                                basic_string_view<Char> format_str,
-                                basic_format_args<buffer_context<Char> > args) {
-  basic_memory_buffer<Char> buffer;
+void vformat_to(buffer<Char>& buf, const text_style& ts,
+                basic_string_view<Char> format_str,
+                basic_format_args<buffer_context<type_identity_t<Char>>> args) {
   bool has_style = false;
   if (ts.has_emphasis()) {
     has_style = true;
-    ansi_color_escape<Char> escape = make_emphasis<Char>(ts.get_emphasis());
-    buffer.append(escape.begin(), escape.end());
+    auto emphasis = detail::make_emphasis<Char>(ts.get_emphasis());
+    buf.append(emphasis.begin(), emphasis.end());
   }
   if (ts.has_foreground()) {
     has_style = true;
-    ansi_color_escape<Char> escape =
-        make_foreground_color<Char>(ts.get_foreground());
-    buffer.append(escape.begin(), escape.end());
+    auto foreground = detail::make_foreground_color<Char>(ts.get_foreground());
+    buf.append(foreground.begin(), foreground.end());
   }
   if (ts.has_background()) {
     has_style = true;
-    ansi_color_escape<Char> escape =
-        make_background_color<Char>(ts.get_background());
-    buffer.append(escape.begin(), escape.end());
+    auto background = detail::make_background_color<Char>(ts.get_background());
+    buf.append(background.begin(), background.end());
   }
-  internal::vformat_to(buffer, format_str, args);
-  if (has_style) {
-    reset_color<Char>(buffer);
-  }
-  return fmt::to_string(buffer);
+  detail::vformat_to(buf, format_str, args);
+  if (has_style) detail::reset_color<Char>(buf);
 }
-}  // namespace internal
+}  // namespace detail
 
-template <typename S, typename Char = char_t<S> >
+template <typename S, typename Char = char_t<S>>
 void vprint(std::FILE* f, const text_style& ts, const S& format,
-            basic_format_args<buffer_context<Char> > args) {
-  bool has_style = false;
-  if (ts.has_emphasis()) {
-    has_style = true;
-    internal::fputs<Char>(internal::make_emphasis<Char>(ts.get_emphasis()), f);
-  }
-  if (ts.has_foreground()) {
-    has_style = true;
-    internal::fputs<Char>(
-        internal::make_foreground_color<Char>(ts.get_foreground()), f);
-  }
-  if (ts.has_background()) {
-    has_style = true;
-    internal::fputs<Char>(
-        internal::make_background_color<Char>(ts.get_background()), f);
-  }
-  vprint(f, format, args);
-  if (has_style) {
-    internal::reset_color<Char>(f);
-  }
+            basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  basic_memory_buffer<Char> buf;
+  detail::vformat_to(buf, ts, to_string_view(format), args);
+  buf.push_back(Char(0));
+  detail::fputs(buf.data(), f);
 }
 
 /**
+  \rst
   Formats a string and prints it to the specified file stream using ANSI
   escape sequences to specify text formatting.
-  Example:
+
+  **Example**::
+
     fmt::print(fmt::emphasis::bold | fg(fmt::color::red),
                "Elapsed time: {0:.2f} seconds", 1.23);
+  \endrst
  */
 template <typename S, typename... Args,
-          FMT_ENABLE_IF(internal::is_string<S>::value)>
+          FMT_ENABLE_IF(detail::is_string<S>::value)>
 void print(std::FILE* f, const text_style& ts, const S& format_str,
            const Args&... args) {
-  internal::check_format_string<Args...>(format_str);
-  using context = buffer_context<char_t<S> >;
-  format_arg_store<context, Args...> as{args...};
-  vprint(f, ts, format_str, basic_format_args<context>(as));
+  vprint(f, ts, format_str,
+         fmt::make_args_checked<Args...>(format_str, args...));
 }
 
 /**
@@ -549,16 +530,18 @@ void print(std::FILE* f, const text_style& ts, const S& format_str,
                "Elapsed time: {0:.2f} seconds", 1.23);
  */
 template <typename S, typename... Args,
-          FMT_ENABLE_IF(internal::is_string<S>::value)>
+          FMT_ENABLE_IF(detail::is_string<S>::value)>
 void print(const text_style& ts, const S& format_str, const Args&... args) {
   return print(stdout, ts, format_str, args...);
 }
 
-template <typename S, typename Char = char_t<S> >
+template <typename S, typename Char = char_t<S>>
 inline std::basic_string<Char> vformat(
     const text_style& ts, const S& format_str,
-    basic_format_args<buffer_context<Char> > args) {
-  return internal::vformat(ts, to_string_view(format_str), args);
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  basic_memory_buffer<Char> buf;
+  detail::vformat_to(buf, ts, to_string_view(format_str), args);
+  return fmt::to_string(buf);
 }
 
 /**
@@ -573,11 +556,46 @@ inline std::basic_string<Char> vformat(
                                       "The answer is {}", 42);
   \endrst
 */
-template <typename S, typename... Args, typename Char = char_t<S> >
+template <typename S, typename... Args, typename Char = char_t<S>>
 inline std::basic_string<Char> format(const text_style& ts, const S& format_str,
                                       const Args&... args) {
-  return internal::vformat(ts, to_string_view(format_str),
-                           {internal::make_args_checked(format_str, args...)});
+  return vformat(ts, to_string_view(format_str),
+                 fmt::make_args_checked<Args...>(format_str, args...));
+}
+
+/**
+  Formats a string with the given text_style and writes the output to ``out``.
+ */
+template <typename OutputIt, typename Char,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value)>
+OutputIt vformat_to(
+    OutputIt out, const text_style& ts, basic_string_view<Char> format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  decltype(detail::get_buffer<Char>(out)) buf(detail::get_buffer_init(out));
+  detail::vformat_to(buf, ts, format_str, args);
+  return detail::get_iterator(buf);
+}
+
+/**
+  \rst
+  Formats arguments with the given text_style, 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),
+                   fmt::emphasis::bold | fg(fmt::color::red), "{}", 42);
+  \endrst
+*/
+template <typename OutputIt, typename S, typename... Args,
+          bool enable = detail::is_output_iterator<OutputIt, char_t<S>>::value&&
+              detail::is_string<S>::value>
+inline auto format_to(OutputIt out, const text_style& ts, const S& format_str,
+                      Args&&... args) ->
+    typename std::enable_if<enable, OutputIt>::type {
+  return vformat_to(out, ts, to_string_view(format_str),
+                    fmt::make_args_checked<Args...>(format_str, args...));
 }
 
 FMT_END_NAMESPACE
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/compile.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/compile.h
index 82625bbc6..3a33b0201 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/compile.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/compile.h
@@ -9,255 +9,76 @@
 #define FMT_COMPILE_H_
 
 #include <vector>
+
 #include "format.h"
 
 FMT_BEGIN_NAMESPACE
-namespace internal {
+namespace detail {
 
+// A compile-time string which is compiled into fast formatting code.
+class compiled_string {};
+
+template <typename S>
+struct is_compiled_string : std::is_base_of<compiled_string, S> {};
+
+/**
+  \rst
+  Converts a string literal *s* into a format string that will be parsed at
+  compile time and converted into efficient formatting code. Requires C++17
+  ``constexpr if`` compiler support.
+
+  **Example**::
+
+    // Converts 42 into std::string using the most efficient method and no
+    // runtime format string processing.
+    std::string s = fmt::format(FMT_COMPILE("{}"), 42);
+  \endrst
+ */
+#define FMT_COMPILE(s) FMT_STRING_IMPL(s, fmt::detail::compiled_string)
+
+template <typename T, typename... Tail>
+const T& first(const T& value, const Tail&...) {
+  return value;
+}
+
+// Part of a compiled format string. It can be either literal text or a
+// replacement field.
 template <typename Char> struct format_part {
- public:
-  struct named_argument_id {
-    FMT_CONSTEXPR named_argument_id(internal::string_view_metadata id)
-        : id(id) {}
-    internal::string_view_metadata id;
+  enum class kind { arg_index, arg_name, text, replacement };
+
+  struct replacement {
+    arg_ref<Char> arg_id;
+    dynamic_format_specs<Char> specs;
   };
 
-  struct argument_id {
-    FMT_CONSTEXPR argument_id() : argument_id(0u) {}
-
-    FMT_CONSTEXPR argument_id(unsigned id)
-        : which(which_arg_id::index), val(id) {}
-
-    FMT_CONSTEXPR argument_id(internal::string_view_metadata id)
-        : which(which_arg_id::named_index), val(id) {}
-
-    enum class which_arg_id { index, named_index };
-
-    which_arg_id which;
-
-    union value {
-      FMT_CONSTEXPR value() : index(0u) {}
-      FMT_CONSTEXPR value(unsigned id) : index(id) {}
-      FMT_CONSTEXPR value(internal::string_view_metadata id)
-          : named_index(id) {}
-
-      unsigned index;
-      internal::string_view_metadata named_index;
-    } val;
-  };
-
-  struct specification {
-    FMT_CONSTEXPR specification() : arg_id(0u) {}
-    FMT_CONSTEXPR specification(unsigned id) : arg_id(id) {}
-
-    FMT_CONSTEXPR specification(internal::string_view_metadata id)
-        : arg_id(id) {}
-
-    argument_id arg_id;
-    internal::dynamic_format_specs<Char> parsed_specs;
-  };
-
-  FMT_CONSTEXPR format_part()
-      : which(kind::argument_id), end_of_argument_id(0u), val(0u) {}
-
-  FMT_CONSTEXPR format_part(internal::string_view_metadata text)
-      : which(kind::text), end_of_argument_id(0u), val(text) {}
-
-  FMT_CONSTEXPR format_part(unsigned id)
-      : which(kind::argument_id), end_of_argument_id(0u), val(id) {}
-
-  FMT_CONSTEXPR format_part(named_argument_id arg_id)
-      : which(kind::named_argument_id), end_of_argument_id(0u), val(arg_id) {}
-
-  FMT_CONSTEXPR format_part(specification spec)
-      : which(kind::specification), end_of_argument_id(0u), val(spec) {}
-
-  enum class kind { argument_id, named_argument_id, text, specification };
-
-  kind which;
-  std::size_t end_of_argument_id;
+  kind part_kind;
   union value {
-    FMT_CONSTEXPR value() : arg_id(0u) {}
-    FMT_CONSTEXPR value(unsigned id) : arg_id(id) {}
-    FMT_CONSTEXPR value(named_argument_id named_id)
-        : named_arg_id(named_id.id) {}
-    FMT_CONSTEXPR value(internal::string_view_metadata t) : text(t) {}
-    FMT_CONSTEXPR value(specification s) : spec(s) {}
-    unsigned arg_id;
-    internal::string_view_metadata named_arg_id;
-    internal::string_view_metadata text;
-    specification spec;
+    int arg_index;
+    basic_string_view<Char> str;
+    replacement repl;
+
+    FMT_CONSTEXPR value(int index = 0) : arg_index(index) {}
+    FMT_CONSTEXPR value(basic_string_view<Char> s) : str(s) {}
+    FMT_CONSTEXPR value(replacement r) : repl(r) {}
   } val;
-};
+  // Position past the end of the argument id.
+  const Char* arg_id_end = nullptr;
 
-template <typename Char, typename PartsContainer>
-class format_preparation_handler : public internal::error_handler {
- private:
-  using part = format_part<Char>;
+  FMT_CONSTEXPR format_part(kind k = kind::arg_index, value v = {})
+      : part_kind(k), val(v) {}
 
- public:
-  using iterator = typename basic_string_view<Char>::iterator;
-
-  FMT_CONSTEXPR format_preparation_handler(basic_string_view<Char> format,
-                                           PartsContainer& parts)
-      : parts_(parts), format_(format), parse_context_(format) {}
-
-  FMT_CONSTEXPR void on_text(const Char* begin, const Char* end) {
-    if (begin == end) return;
-    const auto offset = begin - format_.data();
-    const auto size = end - begin;
-    parts_.push_back(part(string_view_metadata(offset, size)));
+  static FMT_CONSTEXPR format_part make_arg_index(int index) {
+    return format_part(kind::arg_index, index);
   }
-
-  FMT_CONSTEXPR void on_arg_id() {
-    parts_.push_back(part(parse_context_.next_arg_id()));
+  static FMT_CONSTEXPR format_part make_arg_name(basic_string_view<Char> name) {
+    return format_part(kind::arg_name, name);
   }
-
-  FMT_CONSTEXPR void on_arg_id(unsigned id) {
-    parse_context_.check_arg_id(id);
-    parts_.push_back(part(id));
+  static FMT_CONSTEXPR format_part make_text(basic_string_view<Char> text) {
+    return format_part(kind::text, text);
   }
-
-  FMT_CONSTEXPR void on_arg_id(basic_string_view<Char> id) {
-    const auto view = string_view_metadata(format_, id);
-    const auto arg_id = typename part::named_argument_id(view);
-    parts_.push_back(part(arg_id));
+  static FMT_CONSTEXPR format_part make_replacement(replacement repl) {
+    return format_part(kind::replacement, repl);
   }
-
-  FMT_CONSTEXPR void on_replacement_field(const Char* ptr) {
-    parts_.back().end_of_argument_id = ptr - format_.begin();
-  }
-
-  FMT_CONSTEXPR const Char* on_format_specs(const Char* begin,
-                                            const Char* end) {
-    const auto specs_offset = to_unsigned(begin - format_.begin());
-
-    using parse_context = basic_parse_context<Char>;
-    internal::dynamic_format_specs<Char> parsed_specs;
-    dynamic_specs_handler<parse_context> handler(parsed_specs, parse_context_);
-    begin = parse_format_specs(begin, end, handler);
-
-    if (*begin != '}') on_error("missing '}' in format string");
-
-    auto& last_part = parts_.back();
-    auto specs = last_part.which == part::kind::argument_id
-                     ? typename part::specification(last_part.val.arg_id)
-                     : typename part::specification(last_part.val.named_arg_id);
-    specs.parsed_specs = parsed_specs;
-    last_part = part(specs);
-    last_part.end_of_argument_id = specs_offset;
-    return begin;
-  }
-
- private:
-  PartsContainer& parts_;
-  basic_string_view<Char> format_;
-  basic_parse_context<Char> parse_context_;
-};
-
-template <typename Format, typename PreparedPartsProvider, typename... Args>
-class prepared_format {
- public:
-  using char_type = char_t<Format>;
-  using format_part_t = format_part<char_type>;
-
-  constexpr prepared_format(Format f)
-      : format_(std::move(f)), parts_provider_(to_string_view(format_)) {}
-
-  prepared_format() = delete;
-
-  using context = buffer_context<char_type>;
-
-  template <typename Range, typename Context>
-  auto vformat_to(Range out, basic_format_args<Context> args) const ->
-      typename Context::iterator {
-    const auto format_view = internal::to_string_view(format_);
-    basic_parse_context<char_type> parse_ctx(format_view);
-    Context ctx(out.begin(), args);
-
-    const auto& parts = parts_provider_.parts();
-    for (auto part_it = parts.begin(); part_it != parts.end(); ++part_it) {
-      const auto& part = *part_it;
-      const auto& value = part.val;
-
-      switch (part.which) {
-      case format_part_t::kind::text: {
-        const auto text = value.text.to_view(format_view.data());
-        auto output = ctx.out();
-        auto&& it = internal::reserve(output, text.size());
-        it = std::copy_n(text.begin(), text.size(), it);
-        ctx.advance_to(output);
-      } break;
-
-      case format_part_t::kind::argument_id: {
-        advance_parse_context_to_specification(parse_ctx, part);
-        format_arg<Range>(parse_ctx, ctx, value.arg_id);
-      } break;
-
-      case format_part_t::kind::named_argument_id: {
-        advance_parse_context_to_specification(parse_ctx, part);
-        const auto named_arg_id =
-            value.named_arg_id.to_view(format_view.data());
-        format_arg<Range>(parse_ctx, ctx, named_arg_id);
-      } break;
-      case format_part_t::kind::specification: {
-        const auto& arg_id_value = value.spec.arg_id.val;
-        const auto arg = value.spec.arg_id.which ==
-                                 format_part_t::argument_id::which_arg_id::index
-                             ? ctx.arg(arg_id_value.index)
-                             : ctx.arg(arg_id_value.named_index.to_view(
-                                   to_string_view(format_).data()));
-
-        auto specs = value.spec.parsed_specs;
-
-        handle_dynamic_spec<internal::width_checker>(
-            specs.width, specs.width_ref, ctx, format_view.begin());
-        handle_dynamic_spec<internal::precision_checker>(
-            specs.precision, specs.precision_ref, ctx, format_view.begin());
-
-        check_prepared_specs(specs, arg.type());
-        advance_parse_context_to_specification(parse_ctx, part);
-        ctx.advance_to(
-            visit_format_arg(arg_formatter<Range>(ctx, nullptr, &specs), arg));
-      } break;
-      }
-    }
-
-    return ctx.out();
-  }
-
- private:
-  void advance_parse_context_to_specification(
-      basic_parse_context<char_type>& parse_ctx,
-      const format_part_t& part) const {
-    const auto view = to_string_view(format_);
-    const auto specification_begin = view.data() + part.end_of_argument_id;
-    advance_to(parse_ctx, specification_begin);
-  }
-
-  template <typename Range, typename Context, typename Id>
-  void format_arg(basic_parse_context<char_type>& parse_ctx, Context& ctx,
-                  Id arg_id) const {
-    parse_ctx.check_arg_id(arg_id);
-    const auto stopped_at =
-        visit_format_arg(arg_formatter<Range>(ctx), ctx.arg(arg_id));
-    ctx.advance_to(stopped_at);
-  }
-
-  template <typename Char>
-  void check_prepared_specs(const basic_format_specs<Char>& specs,
-                            internal::type arg_type) const {
-    internal::error_handler h;
-    numeric_specs_checker<internal::error_handler> checker(h, arg_type);
-    if (specs.align == align::numeric) checker.require_numeric_argument();
-    if (specs.sign != sign::none) checker.check_sign();
-    if (specs.alt) checker.require_numeric_argument();
-    if (specs.precision >= 0) checker.check_precision();
-  }
-
- private:
-  Format format_;
-  PreparedPartsProvider parts_provider_;
 };
 
 template <typename Char> struct part_counter {
@@ -267,22 +88,24 @@ template <typename Char> struct part_counter {
     if (begin != end) ++num_parts;
   }
 
-  FMT_CONSTEXPR void on_arg_id() { ++num_parts; }
-  FMT_CONSTEXPR void on_arg_id(unsigned) { ++num_parts; }
-  FMT_CONSTEXPR void on_arg_id(basic_string_view<Char>) { ++num_parts; }
+  FMT_CONSTEXPR int on_arg_id() { return ++num_parts, 0; }
+  FMT_CONSTEXPR int on_arg_id(int) { return ++num_parts, 0; }
+  FMT_CONSTEXPR int on_arg_id(basic_string_view<Char>) {
+    return ++num_parts, 0;
+  }
 
-  FMT_CONSTEXPR void on_replacement_field(const Char*) {}
+  FMT_CONSTEXPR void on_replacement_field(int, const Char*) {}
 
-  FMT_CONSTEXPR const Char* on_format_specs(const Char* begin,
+  FMT_CONSTEXPR const Char* on_format_specs(int, const Char* begin,
                                             const Char* end) {
     // Find the matching brace.
-    unsigned braces_counter = 0;
+    unsigned brace_counter = 0;
     for (; begin != end; ++begin) {
       if (*begin == '{') {
-        ++braces_counter;
+        ++brace_counter;
       } else if (*begin == '}') {
-        if (braces_counter == 0u) break;
-        --braces_counter;
+        if (brace_counter == 0u) break;
+        --brace_counter;
       }
     }
     return begin;
@@ -291,174 +114,586 @@ template <typename Char> struct part_counter {
   FMT_CONSTEXPR void on_error(const char*) {}
 };
 
-template <typename Format> class compiletime_prepared_parts_type_provider {
- private:
-  using char_type = char_t<Format>;
+// Counts the number of parts in a format string.
+template <typename Char>
+FMT_CONSTEXPR unsigned count_parts(basic_string_view<Char> format_str) {
+  part_counter<Char> counter;
+  parse_format_string<true>(format_str, counter);
+  return counter.num_parts;
+}
 
-  static FMT_CONSTEXPR unsigned count_parts() {
-    FMT_CONSTEXPR_DECL const auto text = to_string_view(Format{});
-    part_counter<char_type> counter;
-    internal::parse_format_string</*IS_CONSTEXPR=*/true>(text, counter);
-    return counter.num_parts;
+template <typename Char, typename PartHandler>
+class format_string_compiler : public error_handler {
+ private:
+  using part = format_part<Char>;
+
+  PartHandler handler_;
+  part part_;
+  basic_string_view<Char> format_str_;
+  basic_format_parse_context<Char> parse_context_;
+
+ public:
+  FMT_CONSTEXPR format_string_compiler(basic_string_view<Char> format_str,
+                                       PartHandler handler)
+      : handler_(handler),
+        format_str_(format_str),
+        parse_context_(format_str) {}
+
+  FMT_CONSTEXPR void on_text(const Char* begin, const Char* end) {
+    if (begin != end)
+      handler_(part::make_text({begin, to_unsigned(end - begin)}));
   }
 
-// Workaround for old compilers. Compiletime parts preparation will not be
-// performed with them anyway.
+  FMT_CONSTEXPR int on_arg_id() {
+    part_ = part::make_arg_index(parse_context_.next_arg_id());
+    return 0;
+  }
+
+  FMT_CONSTEXPR int on_arg_id(int id) {
+    parse_context_.check_arg_id(id);
+    part_ = part::make_arg_index(id);
+    return 0;
+  }
+
+  FMT_CONSTEXPR int on_arg_id(basic_string_view<Char> id) {
+    part_ = part::make_arg_name(id);
+    return 0;
+  }
+
+  FMT_CONSTEXPR void on_replacement_field(int, const Char* ptr) {
+    part_.arg_id_end = ptr;
+    handler_(part_);
+  }
+
+  FMT_CONSTEXPR const Char* on_format_specs(int, const Char* begin,
+                                            const Char* end) {
+    auto repl = typename part::replacement();
+    dynamic_specs_handler<basic_format_parse_context<Char>> handler(
+        repl.specs, parse_context_);
+    auto it = parse_format_specs(begin, end, handler);
+    if (*it != '}') on_error("missing '}' in format string");
+    repl.arg_id = part_.part_kind == part::kind::arg_index
+                      ? arg_ref<Char>(part_.val.arg_index)
+                      : arg_ref<Char>(part_.val.str);
+    auto part = part::make_replacement(repl);
+    part.arg_id_end = begin;
+    handler_(part);
+    return it;
+  }
+};
+
+// Compiles a format string and invokes handler(part) for each parsed part.
+template <bool IS_CONSTEXPR, typename Char, typename PartHandler>
+FMT_CONSTEXPR void compile_format_string(basic_string_view<Char> format_str,
+                                         PartHandler handler) {
+  parse_format_string<IS_CONSTEXPR>(
+      format_str,
+      format_string_compiler<Char, PartHandler>(format_str, handler));
+}
+
+template <typename OutputIt, typename Context, typename Id>
+void format_arg(
+    basic_format_parse_context<typename Context::char_type>& parse_ctx,
+    Context& ctx, Id arg_id) {
+  ctx.advance_to(visit_format_arg(
+      arg_formatter<OutputIt, typename Context::char_type>(ctx, &parse_ctx),
+      ctx.arg(arg_id)));
+}
+
+// vformat_to is defined in a subnamespace to prevent ADL.
+namespace cf {
+template <typename Context, typename OutputIt, typename CompiledFormat>
+auto vformat_to(OutputIt out, CompiledFormat& cf,
+                basic_format_args<Context> args) -> typename Context::iterator {
+  using char_type = typename Context::char_type;
+  basic_format_parse_context<char_type> parse_ctx(
+      to_string_view(cf.format_str_));
+  Context ctx(out, args);
+
+  const auto& parts = cf.parts();
+  for (auto part_it = std::begin(parts); part_it != std::end(parts);
+       ++part_it) {
+    const auto& part = *part_it;
+    const auto& value = part.val;
+
+    using format_part_t = format_part<char_type>;
+    switch (part.part_kind) {
+    case format_part_t::kind::text: {
+      const auto text = value.str;
+      auto output = ctx.out();
+      auto&& it = reserve(output, text.size());
+      it = std::copy_n(text.begin(), text.size(), it);
+      ctx.advance_to(output);
+      break;
+    }
+
+    case format_part_t::kind::arg_index:
+      advance_to(parse_ctx, part.arg_id_end);
+      detail::format_arg<OutputIt>(parse_ctx, ctx, value.arg_index);
+      break;
+
+    case format_part_t::kind::arg_name:
+      advance_to(parse_ctx, part.arg_id_end);
+      detail::format_arg<OutputIt>(parse_ctx, ctx, value.str);
+      break;
+
+    case format_part_t::kind::replacement: {
+      const auto& arg_id_value = value.repl.arg_id.val;
+      const auto arg = value.repl.arg_id.kind == arg_id_kind::index
+                           ? ctx.arg(arg_id_value.index)
+                           : ctx.arg(arg_id_value.name);
+
+      auto specs = value.repl.specs;
+
+      handle_dynamic_spec<width_checker>(specs.width, specs.width_ref, ctx);
+      handle_dynamic_spec<precision_checker>(specs.precision,
+                                             specs.precision_ref, ctx);
+
+      error_handler h;
+      numeric_specs_checker<error_handler> checker(h, arg.type());
+      if (specs.align == align::numeric) checker.require_numeric_argument();
+      if (specs.sign != sign::none) checker.check_sign();
+      if (specs.alt) checker.require_numeric_argument();
+      if (specs.precision >= 0) checker.check_precision();
+
+      advance_to(parse_ctx, part.arg_id_end);
+      ctx.advance_to(
+          visit_format_arg(arg_formatter<OutputIt, typename Context::char_type>(
+                               ctx, nullptr, &specs),
+                           arg));
+      break;
+    }
+    }
+  }
+  return ctx.out();
+}
+}  // namespace cf
+
+struct basic_compiled_format {};
+
+template <typename S, typename = void>
+struct compiled_format_base : basic_compiled_format {
+  using char_type = char_t<S>;
+  using parts_container = std::vector<detail::format_part<char_type>>;
+
+  parts_container compiled_parts;
+
+  explicit compiled_format_base(basic_string_view<char_type> format_str) {
+    compile_format_string<false>(format_str,
+                                 [this](const format_part<char_type>& part) {
+                                   compiled_parts.push_back(part);
+                                 });
+  }
+
+  const parts_container& parts() const { return compiled_parts; }
+};
+
+template <typename Char, unsigned N> struct format_part_array {
+  format_part<Char> data[N] = {};
+  FMT_CONSTEXPR format_part_array() = default;
+};
+
+template <typename Char, unsigned N>
+FMT_CONSTEXPR format_part_array<Char, N> compile_to_parts(
+    basic_string_view<Char> format_str) {
+  format_part_array<Char, N> parts;
+  unsigned counter = 0;
+  // This is not a lambda for compatibility with older compilers.
+  struct {
+    format_part<Char>* parts;
+    unsigned* counter;
+    FMT_CONSTEXPR void operator()(const format_part<Char>& part) {
+      parts[(*counter)++] = part;
+    }
+  } collector{parts.data, &counter};
+  compile_format_string<true>(format_str, collector);
+  if (counter < N) {
+    parts.data[counter] =
+        format_part<Char>::make_text(basic_string_view<Char>());
+  }
+  return parts;
+}
+
+template <typename T> constexpr const T& constexpr_max(const T& a, const T& b) {
+  return (a < b) ? b : a;
+}
+
+template <typename S>
+struct compiled_format_base<S, enable_if_t<is_compile_string<S>::value>>
+    : basic_compiled_format {
+  using char_type = char_t<S>;
+
+  FMT_CONSTEXPR explicit compiled_format_base(basic_string_view<char_type>) {}
+
+// Workaround for old compilers. Format string compilation will not be
+// performed there anyway.
 #if FMT_USE_CONSTEXPR
-  static FMT_CONSTEXPR_DECL const unsigned number_of_format_parts =
-      compiletime_prepared_parts_type_provider::count_parts();
+  static FMT_CONSTEXPR_DECL const unsigned num_format_parts =
+      constexpr_max(count_parts(to_string_view(S())), 1u);
 #else
-  static const unsigned number_of_format_parts = 0u;
+  static const unsigned num_format_parts = 1;
 #endif
 
- public:
-  template <unsigned N> struct format_parts_array {
-    using value_type = format_part<char_type>;
+  using parts_container = format_part<char_type>[num_format_parts];
 
-    FMT_CONSTEXPR format_parts_array() : arr{} {}
-
-    FMT_CONSTEXPR value_type& operator[](unsigned ind) { return arr[ind]; }
-
-    FMT_CONSTEXPR const value_type* begin() const { return arr; }
-    FMT_CONSTEXPR const value_type* end() const { return begin() + N; }
-
-   private:
-    value_type arr[N];
-  };
-
-  struct empty {
-    // Parts preparator will search for it
-    using value_type = format_part<char_type>;
-  };
-
-  using type = conditional_t<number_of_format_parts != 0,
-                             format_parts_array<number_of_format_parts>, empty>;
-};
-
-template <typename Parts> class compiletime_prepared_parts_collector {
- private:
-  using format_part = typename Parts::value_type;
-
- public:
-  FMT_CONSTEXPR explicit compiletime_prepared_parts_collector(Parts& parts)
-      : parts_{parts}, counter_{0u} {}
-
-  FMT_CONSTEXPR void push_back(format_part part) { parts_[counter_++] = part; }
-
-  FMT_CONSTEXPR format_part& back() { return parts_[counter_ - 1]; }
-
- private:
-  Parts& parts_;
-  unsigned counter_;
-};
-
-template <typename PartsContainer, typename Char>
-FMT_CONSTEXPR PartsContainer prepare_parts(basic_string_view<Char> format) {
-  PartsContainer parts;
-  internal::parse_format_string</*IS_CONSTEXPR=*/false>(
-      format, format_preparation_handler<Char, PartsContainer>(format, parts));
-  return parts;
-}
-
-template <typename PartsContainer, typename Char>
-FMT_CONSTEXPR PartsContainer
-prepare_compiletime_parts(basic_string_view<Char> format) {
-  using collector = compiletime_prepared_parts_collector<PartsContainer>;
-
-  PartsContainer parts;
-  collector c(parts);
-  internal::parse_format_string</*IS_CONSTEXPR=*/true>(
-      format, format_preparation_handler<Char, collector>(format, c));
-  return parts;
-}
-
-template <typename PartsContainer> class runtime_parts_provider {
- public:
-  runtime_parts_provider() = delete;
-  template <typename Char>
-  runtime_parts_provider(basic_string_view<Char> format)
-      : parts_(prepare_parts<PartsContainer>(format)) {}
-
-  const PartsContainer& parts() const { return parts_; }
-
- private:
-  PartsContainer parts_;
-};
-
-template <typename Format, typename PartsContainer>
-struct compiletime_parts_provider {
-  compiletime_parts_provider() = delete;
-  template <typename Char>
-  FMT_CONSTEXPR compiletime_parts_provider(basic_string_view<Char>) {}
-
-  const PartsContainer& parts() const {
-    static FMT_CONSTEXPR_DECL const PartsContainer prepared_parts =
-        prepare_compiletime_parts<PartsContainer>(
-            internal::to_string_view(Format{}));
-
-    return prepared_parts;
+  const parts_container& parts() const {
+    static FMT_CONSTEXPR_DECL const auto compiled_parts =
+        compile_to_parts<char_type, num_format_parts>(
+            detail::to_string_view(S()));
+    return compiled_parts.data;
   }
 };
-}  // namespace internal
 
-#if FMT_USE_CONSTEXPR
+template <typename S, typename... Args>
+class compiled_format : private compiled_format_base<S> {
+ public:
+  using typename compiled_format_base<S>::char_type;
+
+ private:
+  basic_string_view<char_type> format_str_;
+
+  template <typename Context, typename OutputIt, typename CompiledFormat>
+  friend auto cf::vformat_to(OutputIt out, CompiledFormat& cf,
+                             basic_format_args<Context> args) ->
+      typename Context::iterator;
+
+ public:
+  compiled_format() = delete;
+  explicit constexpr compiled_format(basic_string_view<char_type> format_str)
+      : compiled_format_base<S>(format_str), format_str_(format_str) {}
+};
+
+#ifdef __cpp_if_constexpr
+template <typename... Args> struct type_list {};
+
+// Returns a reference to the argument at index N from [first, rest...].
+template <int N, typename T, typename... Args>
+constexpr const auto& get([[maybe_unused]] const T& first,
+                          [[maybe_unused]] const Args&... rest) {
+  static_assert(N < 1 + sizeof...(Args), "index is out of bounds");
+  if constexpr (N == 0)
+    return first;
+  else
+    return get<N - 1>(rest...);
+}
+
+template <int N, typename> struct get_type_impl;
+
+template <int N, typename... Args> struct get_type_impl<N, type_list<Args...>> {
+  using type = remove_cvref_t<decltype(get<N>(std::declval<Args>()...))>;
+};
+
+template <int N, typename T>
+using get_type = typename get_type_impl<N, T>::type;
+
+template <typename T> struct is_compiled_format : std::false_type {};
+
+template <typename Char> struct text {
+  basic_string_view<Char> data;
+  using char_type = Char;
+
+  template <typename OutputIt, typename... Args>
+  OutputIt format(OutputIt out, const Args&...) const {
+    return write<Char>(out, data);
+  }
+};
+
+template <typename Char>
+struct is_compiled_format<text<Char>> : std::true_type {};
+
+template <typename Char>
+constexpr text<Char> make_text(basic_string_view<Char> s, size_t pos,
+                               size_t size) {
+  return {{&s[pos], size}};
+}
+
+template <typename Char> struct code_unit {
+  Char value;
+  using char_type = Char;
+
+  template <typename OutputIt, typename... Args>
+  OutputIt format(OutputIt out, const Args&...) const {
+    return write<Char>(out, value);
+  }
+};
+
+template <typename Char>
+struct is_compiled_format<code_unit<Char>> : std::true_type {};
+
+// A replacement field that refers to argument N.
+template <typename Char, typename T, int N> struct field {
+  using char_type = Char;
+
+  template <typename OutputIt, typename... Args>
+  OutputIt format(OutputIt out, const Args&... args) const {
+    // This ensures that the argument type is convertile to `const T&`.
+    const T& arg = get<N>(args...);
+    return write<Char>(out, arg);
+  }
+};
+
+template <typename Char, typename T, int N>
+struct is_compiled_format<field<Char, T, N>> : std::true_type {};
+
+// A replacement field that refers to argument N and has format specifiers.
+template <typename Char, typename T, int N> struct spec_field {
+  using char_type = Char;
+  mutable formatter<T, Char> fmt;
+
+  template <typename OutputIt, typename... Args>
+  OutputIt format(OutputIt out, const Args&... args) const {
+    // This ensures that the argument type is convertile to `const T&`.
+    const T& arg = get<N>(args...);
+    const auto& vargs =
+        make_format_args<basic_format_context<OutputIt, Char>>(args...);
+    basic_format_context<OutputIt, Char> ctx(out, vargs);
+    return fmt.format(arg, ctx);
+  }
+};
+
+template <typename Char, typename T, int N>
+struct is_compiled_format<spec_field<Char, T, N>> : std::true_type {};
+
+template <typename L, typename R> struct concat {
+  L lhs;
+  R rhs;
+  using char_type = typename L::char_type;
+
+  template <typename OutputIt, typename... Args>
+  OutputIt format(OutputIt out, const Args&... args) const {
+    out = lhs.format(out, args...);
+    return rhs.format(out, args...);
+  }
+};
+
+template <typename L, typename R>
+struct is_compiled_format<concat<L, R>> : std::true_type {};
+
+template <typename L, typename R>
+constexpr concat<L, R> make_concat(L lhs, R rhs) {
+  return {lhs, rhs};
+}
+
+struct unknown_format {};
+
+template <typename Char>
+constexpr size_t parse_text(basic_string_view<Char> str, size_t pos) {
+  for (size_t size = str.size(); pos != size; ++pos) {
+    if (str[pos] == '{' || str[pos] == '}') break;
+  }
+  return pos;
+}
+
+template <typename Args, size_t POS, int ID, typename S>
+constexpr auto compile_format_string(S format_str);
+
+template <typename Args, size_t POS, int ID, typename T, typename S>
+constexpr auto parse_tail(T head, S format_str) {
+  if constexpr (POS !=
+                basic_string_view<typename S::char_type>(format_str).size()) {
+    constexpr auto tail = compile_format_string<Args, POS, ID>(format_str);
+    if constexpr (std::is_same<remove_cvref_t<decltype(tail)>,
+                               unknown_format>())
+      return tail;
+    else
+      return make_concat(head, tail);
+  } else {
+    return head;
+  }
+}
+
+template <typename T, typename Char> struct parse_specs_result {
+  formatter<T, Char> fmt;
+  size_t end;
+  int next_arg_id;
+};
+
+template <typename T, typename Char>
+constexpr parse_specs_result<T, Char> parse_specs(basic_string_view<Char> str,
+                                                  size_t pos, int arg_id) {
+  str.remove_prefix(pos);
+  auto ctx = basic_format_parse_context<Char>(str, {}, arg_id + 1);
+  auto f = formatter<T, Char>();
+  auto end = f.parse(ctx);
+  return {f, pos + (end - str.data()) + 1, ctx.next_arg_id()};
+}
+
+// Compiles a non-empty format string and returns the compiled representation
+// or unknown_format() on unrecognized input.
+template <typename Args, size_t POS, int ID, typename S>
+constexpr auto compile_format_string(S format_str) {
+  using char_type = typename S::char_type;
+  constexpr basic_string_view<char_type> str = format_str;
+  if constexpr (str[POS] == '{') {
+    if (POS + 1 == str.size())
+      throw format_error("unmatched '{' in format string");
+    if constexpr (str[POS + 1] == '{') {
+      return parse_tail<Args, POS + 2, ID>(make_text(str, POS, 1), format_str);
+    } else if constexpr (str[POS + 1] == '}') {
+      using type = get_type<ID, Args>;
+      return parse_tail<Args, POS + 2, ID + 1>(field<char_type, type, ID>(),
+                                               format_str);
+    } else if constexpr (str[POS + 1] == ':') {
+      using type = get_type<ID, Args>;
+      constexpr auto result = parse_specs<type>(str, POS + 2, ID);
+      return parse_tail<Args, result.end, result.next_arg_id>(
+          spec_field<char_type, type, ID>{result.fmt}, format_str);
+    } else {
+      return unknown_format();
+    }
+  } else if constexpr (str[POS] == '}') {
+    if (POS + 1 == str.size())
+      throw format_error("unmatched '}' in format string");
+    return parse_tail<Args, POS + 2, ID>(make_text(str, POS, 1), format_str);
+  } else {
+    constexpr auto end = parse_text(str, POS + 1);
+    if constexpr (end - POS > 1) {
+      return parse_tail<Args, end, ID>(make_text(str, POS, end - POS),
+                                       format_str);
+    } else {
+      return parse_tail<Args, end, ID>(code_unit<char_type>{str[POS]},
+                                       format_str);
+    }
+  }
+}
+
+template <typename... Args, typename S,
+          FMT_ENABLE_IF(is_compile_string<S>::value ||
+                        detail::is_compiled_string<S>::value)>
+constexpr auto compile(S format_str) {
+  constexpr basic_string_view<typename S::char_type> str = format_str;
+  if constexpr (str.size() == 0) {
+    return detail::make_text(str, 0, 0);
+  } else {
+    constexpr auto result =
+        detail::compile_format_string<detail::type_list<Args...>, 0, 0>(
+            format_str);
+    if constexpr (std::is_same<remove_cvref_t<decltype(result)>,
+                               detail::unknown_format>()) {
+      return detail::compiled_format<S, Args...>(to_string_view(format_str));
+    } else {
+      return result;
+    }
+  }
+}
+#else
 template <typename... Args, typename S,
           FMT_ENABLE_IF(is_compile_string<S>::value)>
-FMT_CONSTEXPR auto compile(S format_str) -> internal::prepared_format<
-    S,
-    internal::compiletime_parts_provider<
-        S,
-        typename internal::compiletime_prepared_parts_type_provider<S>::type>,
-    Args...> {
-  return format_str;
+constexpr auto compile(S format_str) -> detail::compiled_format<S, Args...> {
+  return detail::compiled_format<S, Args...>(to_string_view(format_str));
 }
-#endif
+#endif  // __cpp_if_constexpr
 
+// Compiles the format string which must be a string literal.
 template <typename... Args, typename Char, size_t N>
-auto compile(const Char (&format_str)[N]) -> internal::prepared_format<
-    std::basic_string<Char>,
-    internal::runtime_parts_provider<std::vector<internal::format_part<Char>>>,
-    Args...> {
-  return std::basic_string<Char>(format_str, N - 1);
+auto compile(const Char (&format_str)[N])
+    -> detail::compiled_format<const Char*, Args...> {
+  return detail::compiled_format<const Char*, Args...>(
+      basic_string_view<Char>(format_str, N - 1));
 }
+}  // namespace detail
+
+// DEPRECATED! use FMT_COMPILE instead.
+template <typename... Args>
+FMT_DEPRECATED auto compile(const Args&... args)
+    -> decltype(detail::compile(args...)) {
+  return detail::compile(args...);
+}
+
+#if FMT_USE_CONSTEXPR
+#  ifdef __cpp_if_constexpr
 
 template <typename CompiledFormat, typename... Args,
-          typename Char = typename CompiledFormat::char_type>
-std::basic_string<Char> format(const CompiledFormat& cf, const Args&... args) {
+          typename Char = typename CompiledFormat::char_type,
+          FMT_ENABLE_IF(detail::is_compiled_format<CompiledFormat>::value)>
+FMT_INLINE std::basic_string<Char> format(const CompiledFormat& cf,
+                                          const Args&... args) {
   basic_memory_buffer<Char> buffer;
-  using range = internal::buffer_range<Char>;
-  using context = buffer_context<Char>;
-  cf.template vformat_to<range, context>(range(buffer),
-                                         {make_format_args<context>(args...)});
+  cf.format(detail::buffer_appender<Char>(buffer), args...);
   return to_string(buffer);
 }
 
-template <typename OutputIt, typename CompiledFormat, typename... Args>
+template <typename OutputIt, typename CompiledFormat, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_format<CompiledFormat>::value)>
 OutputIt format_to(OutputIt out, const CompiledFormat& cf,
                    const Args&... args) {
-  using char_type = typename CompiledFormat::char_type;
-  using range = internal::output_range<OutputIt, char_type>;
-  using context = format_context_t<OutputIt, char_type>;
-  return cf.template vformat_to<range, context>(
-      range(out), {make_format_args<context>(args...)});
+  return cf.format(out, args...);
+}
+#  endif  // __cpp_if_constexpr
+#endif    // FMT_USE_CONSTEXPR
+
+template <typename CompiledFormat, typename... Args,
+          typename Char = typename CompiledFormat::char_type,
+          FMT_ENABLE_IF(std::is_base_of<detail::basic_compiled_format,
+                                        CompiledFormat>::value)>
+std::basic_string<Char> format(const CompiledFormat& cf, const Args&... args) {
+  basic_memory_buffer<Char> buffer;
+  using context = buffer_context<Char>;
+  detail::cf::vformat_to<context>(detail::buffer_appender<Char>(buffer), cf,
+                                  make_format_args<context>(args...));
+  return to_string(buffer);
+}
+
+template <typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+FMT_INLINE std::basic_string<typename S::char_type> format(const S&,
+                                                           Args&&... args) {
+#ifdef __cpp_if_constexpr
+  if constexpr (std::is_same<typename S::char_type, char>::value) {
+    constexpr basic_string_view<typename S::char_type> str = S();
+    if (str.size() == 2 && str[0] == '{' && str[1] == '}')
+      return fmt::to_string(detail::first(args...));
+  }
+#endif
+  constexpr auto compiled = detail::compile<Args...>(S());
+  return format(compiled, std::forward<Args>(args)...);
 }
 
 template <typename OutputIt, typename CompiledFormat, typename... Args,
-          FMT_ENABLE_IF(internal::is_output_iterator<OutputIt>::value)>
-format_to_n_result<OutputIt> format_to_n(OutputIt out, size_t n,
-                                         const CompiledFormat& cf,
-                                         const Args&... args) {
+          FMT_ENABLE_IF(std::is_base_of<detail::basic_compiled_format,
+                                        CompiledFormat>::value)>
+OutputIt format_to(OutputIt out, const CompiledFormat& cf,
+                   const Args&... args) {
+  using char_type = typename CompiledFormat::char_type;
+  using context = format_context_t<OutputIt, char_type>;
+  return detail::cf::vformat_to<context>(out, cf,
+                                         make_format_args<context>(args...));
+}
+
+template <typename OutputIt, typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+OutputIt format_to(OutputIt out, const S&, const Args&... args) {
+  constexpr auto compiled = detail::compile<Args...>(S());
+  return format_to(out, compiled, args...);
+}
+
+template <typename OutputIt, typename CompiledFormat, typename... Args>
+auto format_to_n(OutputIt out, size_t n, const CompiledFormat& cf,
+                 const Args&... args) ->
+    typename std::enable_if<
+        detail::is_output_iterator<OutputIt,
+                                   typename CompiledFormat::char_type>::value &&
+            std::is_base_of<detail::basic_compiled_format,
+                            CompiledFormat>::value,
+        format_to_n_result<OutputIt>>::type {
   auto it =
-      format_to(internal::truncating_iterator<OutputIt>(out, n), cf, args...);
+      format_to(detail::truncating_iterator<OutputIt>(out, n), cf, args...);
+  return {it.base(), it.count()};
+}
+
+template <typename OutputIt, typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_compiled_string<S>::value)>
+format_to_n_result<OutputIt> format_to_n(OutputIt out, size_t n, const S&,
+                                         const Args&... args) {
+  constexpr auto compiled = detail::compile<Args...>(S());
+  auto it = format_to(detail::truncating_iterator<OutputIt>(out, n), compiled,
+                      args...);
   return {it.base(), it.count()};
 }
 
 template <typename CompiledFormat, typename... Args>
-std::size_t formatted_size(const CompiledFormat& cf, const Args&... args) {
-  return fmt::format_to(
-             internal::counting_iterator<typename CompiledFormat::char_type>(),
-             cf, args...)
-      .count();
+size_t formatted_size(const CompiledFormat& cf, const Args&... args) {
+  return format_to(detail::counting_iterator(), cf, args...).count();
 }
 
 FMT_END_NAMESPACE
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/core.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/core.h
index 29a1281bb..0a81e0ccd 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/core.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/core.h
@@ -8,15 +8,55 @@
 #ifndef FMT_CORE_H_
 #define FMT_CORE_H_
 
-#include <cassert>
 #include <cstdio>  // std::FILE
 #include <cstring>
+#include <functional>
 #include <iterator>
+#include <memory>
 #include <string>
 #include <type_traits>
+#include <vector>
 
 // The fmt library version in the form major * 10000 + minor * 100 + patch.
-#define FMT_VERSION 60000
+#define FMT_VERSION 70103
+
+#ifdef __clang__
+#  define FMT_CLANG_VERSION (__clang_major__ * 100 + __clang_minor__)
+#else
+#  define FMT_CLANG_VERSION 0
+#endif
+
+#if defined(__GNUC__) && !defined(__clang__)
+#  define FMT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
+#else
+#  define FMT_GCC_VERSION 0
+#endif
+
+#if defined(__INTEL_COMPILER)
+#  define FMT_ICC_VERSION __INTEL_COMPILER
+#else
+#  define FMT_ICC_VERSION 0
+#endif
+
+#if __cplusplus >= 201103L || defined(__GXX_EXPERIMENTAL_CXX0X__)
+#  define FMT_HAS_GXX_CXX11 FMT_GCC_VERSION
+#else
+#  define FMT_HAS_GXX_CXX11 0
+#endif
+
+#ifdef __NVCC__
+#  define FMT_NVCC __NVCC__
+#else
+#  define FMT_NVCC 0
+#endif
+
+#ifdef _MSC_VER
+#  define FMT_MSC_VER _MSC_VER
+#  define FMT_SUPPRESS_MSC_WARNING(n) __pragma(warning(suppress : n))
+#else
+#  define FMT_MSC_VER 0
+#  define FMT_SUPPRESS_MSC_WARNING(n)
+#endif
 
 #ifdef __has_feature
 #  define FMT_HAS_FEATURE(x) __has_feature(x)
@@ -25,7 +65,7 @@
 #endif
 
 #if defined(__has_include) && !defined(__INTELLISENSE__) && \
-    !(defined(__INTEL_COMPILER) && __INTEL_COMPILER < 1600)
+    (!FMT_ICC_VERSION || FMT_ICC_VERSION >= 1600)
 #  define FMT_HAS_INCLUDE(x) __has_include(x)
 #else
 #  define FMT_HAS_INCLUDE(x) 0
@@ -37,30 +77,19 @@
 #  define FMT_HAS_CPP_ATTRIBUTE(x) 0
 #endif
 
-#if defined(__GNUC__) && !defined(__clang__)
-#  define FMT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
-#else
-#  define FMT_GCC_VERSION 0
-#endif
+#define FMT_HAS_CPP14_ATTRIBUTE(attribute) \
+  (__cplusplus >= 201402L && FMT_HAS_CPP_ATTRIBUTE(attribute))
 
-#if __cplusplus >= 201103L || defined(__GXX_EXPERIMENTAL_CXX0X__)
-#  define FMT_HAS_GXX_CXX11 FMT_GCC_VERSION
-#else
-#  define FMT_HAS_GXX_CXX11 0
-#endif
-
-#ifdef _MSC_VER
-#  define FMT_MSC_VER _MSC_VER
-#else
-#  define FMT_MSC_VER 0
-#endif
+#define FMT_HAS_CPP17_ATTRIBUTE(attribute) \
+  (__cplusplus >= 201703L && FMT_HAS_CPP_ATTRIBUTE(attribute))
 
 // Check if relaxed C++14 constexpr is supported.
 // GCC doesn't allow throw in constexpr until version 6 (bug 67371).
 #ifndef FMT_USE_CONSTEXPR
 #  define FMT_USE_CONSTEXPR                                           \
     (FMT_HAS_FEATURE(cxx_relaxed_constexpr) || FMT_MSC_VER >= 1910 || \
-     (FMT_GCC_VERSION >= 600 && __cplusplus >= 201402L))
+     (FMT_GCC_VERSION >= 600 && __cplusplus >= 201402L)) &&           \
+        !FMT_NVCC && !FMT_ICC_VERSION
 #endif
 #if FMT_USE_CONSTEXPR
 #  define FMT_CONSTEXPR constexpr
@@ -71,7 +100,7 @@
 #endif
 
 #ifndef FMT_OVERRIDE
-#  if FMT_HAS_FEATURE(cxx_override) || \
+#  if FMT_HAS_FEATURE(cxx_override_control) || \
       (FMT_GCC_VERSION >= 408 && FMT_HAS_GXX_CXX11) || FMT_MSC_VER >= 1900
 #    define FMT_OVERRIDE override
 #  else
@@ -111,19 +140,20 @@
 #  endif
 #endif
 
-// [[noreturn]] is disabled on MSVC because of bogus unreachable code warnings.
-#if FMT_EXCEPTIONS && FMT_HAS_CPP_ATTRIBUTE(noreturn) && !FMT_MSC_VER
+// [[noreturn]] is disabled on MSVC and NVCC because of bogus unreachable code
+// warnings.
+#if FMT_EXCEPTIONS && FMT_HAS_CPP_ATTRIBUTE(noreturn) && !FMT_MSC_VER && \
+    !FMT_NVCC
 #  define FMT_NORETURN [[noreturn]]
 #else
 #  define FMT_NORETURN
 #endif
 
 #ifndef FMT_DEPRECATED
-#  if (FMT_HAS_CPP_ATTRIBUTE(deprecated) && __cplusplus >= 201402L) || \
-      FMT_MSC_VER >= 1900
+#  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)
@@ -132,16 +162,33 @@
 #    endif
 #  endif
 #endif
-// Workaround broken [[deprecated]] in the Intel compiler.
-#if defined(__INTEL_COMPILER) || defined(__NVCC__)
+
+// Workaround broken [[deprecated]] in the Intel, PGI and NVCC compilers.
+#if FMT_ICC_VERSION || defined(__PGI) || FMT_NVCC
 #  define FMT_DEPRECATED_ALIAS
 #else
 #  define FMT_DEPRECATED_ALIAS FMT_DEPRECATED
 #endif
 
-#ifndef FMT_BEGIN_NAMESPACE
+#ifndef FMT_INLINE
+#  if FMT_GCC_VERSION || FMT_CLANG_VERSION
+#    define FMT_INLINE inline __attribute__((always_inline))
+#  else
+#    define FMT_INLINE inline
+#  endif
+#endif
+
+#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 \
       }                       \
@@ -150,21 +197,26 @@
 #    define FMT_INLINE_NAMESPACE namespace
 #    define FMT_END_NAMESPACE \
       }                       \
-      using namespace v6;     \
+      using namespace v7;     \
       }
 #  endif
 #  define FMT_BEGIN_NAMESPACE \
     namespace fmt {           \
-    FMT_INLINE_NAMESPACE v6 {
+    FMT_INLINE_NAMESPACE v7 {
 #endif
 
 #if !defined(FMT_HEADER_ONLY) && defined(_WIN32)
+#  define FMT_CLASS_API FMT_SUPPRESS_MSC_WARNING(4275)
 #  ifdef FMT_EXPORT
 #    define FMT_API __declspec(dllexport)
+#    define FMT_EXTERN_TEMPLATE_API FMT_API
+#    define FMT_EXPORTED
 #  elif defined(FMT_SHARED)
 #    define FMT_API __declspec(dllimport)
 #    define FMT_EXTERN_TEMPLATE_API FMT_API
 #  endif
+#else
+#  define FMT_CLASS_API
 #endif
 #ifndef FMT_API
 #  define FMT_API
@@ -172,6 +224,9 @@
 #ifndef FMT_EXTERN_TEMPLATE_API
 #  define FMT_EXTERN_TEMPLATE_API
 #endif
+#ifndef FMT_INSTANTIATION_DEF_API
+#  define FMT_INSTANTIATION_DEF_API FMT_API
+#endif
 
 #ifndef FMT_HEADER_ONLY
 #  define FMT_EXTERN extern
@@ -179,10 +234,6 @@
 #  define FMT_EXTERN
 #endif
 
-#ifndef FMT_ASSERT
-#  define FMT_ASSERT(condition, message) assert((condition) && message)
-#endif
-
 // libc++ supports string_view in pre-c++17.
 #if (FMT_HAS_INCLUDE(<string_view>) &&                       \
      (__cplusplus > 201402L || defined(_LIBCPP_VERSION))) || \
@@ -194,9 +245,16 @@
 #  define FMT_USE_EXPERIMENTAL_STRING_VIEW
 #endif
 
+#ifndef FMT_UNICODE
+#  define FMT_UNICODE !FMT_MSC_VER
+#endif
+#if FMT_UNICODE && FMT_MSC_VER
+#  pragma execution_character_set("utf-8")
+#endif
+
 FMT_BEGIN_NAMESPACE
 
-// Implementations of enable_if_t and other types for pre-C++14 systems.
+// Implementations of enable_if_t and other metafunctions for older systems.
 template <bool B, class T = void>
 using enable_if_t = typename std::enable_if<B, T>::type;
 template <bool B, class T, class F>
@@ -206,6 +264,10 @@ template <typename T>
 using remove_reference_t = typename std::remove_reference<T>::type;
 template <typename T>
 using remove_const_t = typename std::remove_const<T>::type;
+template <typename T>
+using remove_cvref_t = typename std::remove_cv<remove_reference_t<T>>::type;
+template <typename T> struct type_identity { using type = T; };
+template <typename T> using type_identity_t = typename type_identity<T>::type;
 
 struct monostate {};
 
@@ -214,10 +276,25 @@ struct monostate {};
 // to workaround a bug in MSVC 2019 (see #1140 and #1186).
 #define FMT_ENABLE_IF(...) enable_if_t<(__VA_ARGS__), int> = 0
 
-namespace internal {
+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; };
+// 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,
+                                      const char* message);
+
+#ifndef FMT_ASSERT
+#  ifdef NDEBUG
+// FMT_ASSERT is not empty to avoid -Werror=empty-body.
+#    define FMT_ASSERT(condition, message) ((void)0)
+#  else
+#    define FMT_ASSERT(condition, message)                                    \
+      ((condition) /* void() fails with -Winvalid-constexpr on clang 4.0.1 */ \
+           ? (void)0                                                          \
+           : ::fmt::detail::assert_fail(__FILE__, __LINE__, (message)))
+#  endif
+#endif
 
 #if defined(FMT_USE_STRING_VIEW)
 template <typename Char> using std_string_view = std::basic_string_view<Char>;
@@ -228,16 +305,45 @@ using std_string_view = std::experimental::basic_string_view<Char>;
 template <typename T> struct std_string_view {};
 #endif
 
-// Casts nonnegative integer to unsigned.
+#ifdef FMT_USE_INT128
+// Do nothing.
+#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;
+#else
+#  define FMT_USE_INT128 0
+#endif
+#if !FMT_USE_INT128
+struct int128_t {};
+struct uint128_t {};
+#endif
+
+// Casts a nonnegative integer to unsigned.
 template <typename Int>
 FMT_CONSTEXPR typename std::make_unsigned<Int>::type to_unsigned(Int value) {
   FMT_ASSERT(value >= 0, "negative value");
   return static_cast<typename std::make_unsigned<Int>::type>(value);
 }
-}  // namespace internal
 
-template <typename... Ts>
-using void_t = typename internal::void_t_impl<Ts...>::type;
+FMT_SUPPRESS_MSC_WARNING(4566) constexpr unsigned char micro[] = "\u00B5";
+
+template <typename Char> constexpr bool is_unicode() {
+  return FMT_UNICODE || sizeof(Char) != 1 ||
+         (sizeof(micro) == 3 && micro[0] == 0xC2 && micro[1] == 0xB5);
+}
+
+#ifdef __cpp_char8_t
+using char8_type = char8_t;
+#else
+enum char8_type : unsigned char {};
+#endif
+}  // namespace detail
+
+#ifdef FMT_USE_INTERNAL
+namespace internal = detail;  // DEPRECATED
+#endif
 
 /**
   An implementation of ``std::basic_string_view`` for pre-C++17. It provides a
@@ -252,13 +358,13 @@ template <typename Char> class basic_string_view {
   size_t size_;
 
  public:
-  using char_type = Char;
+  using value_type = Char;
   using iterator = const Char*;
 
-  FMT_CONSTEXPR basic_string_view() FMT_NOEXCEPT : data_(nullptr), size_(0) {}
+  constexpr basic_string_view() FMT_NOEXCEPT : data_(nullptr), size_(0) {}
 
   /** Constructs a string reference object from a C string and a size. */
-  FMT_CONSTEXPR basic_string_view(const Char* s, size_t count) FMT_NOEXCEPT
+  constexpr basic_string_view(const Char* s, size_t count) FMT_NOEXCEPT
       : data_(s),
         size_(count) {}
 
@@ -268,29 +374,34 @@ template <typename Char> class basic_string_view {
     the size with ``std::char_traits<Char>::length``.
     \endrst
    */
+#if __cplusplus >= 201703L  // C++17's char_traits::length() is constexpr.
+  FMT_CONSTEXPR
+#endif
   basic_string_view(const Char* s)
       : data_(s), size_(std::char_traits<Char>::length(s)) {}
 
   /** Constructs a string reference from a ``std::basic_string`` object. */
-  template <typename Alloc>
-  FMT_CONSTEXPR basic_string_view(const std::basic_string<Char, Alloc>& s)
-      FMT_NOEXCEPT : data_(s.data()),
-                     size_(s.size()) {}
+  template <typename Traits, typename Alloc>
+  FMT_CONSTEXPR basic_string_view(
+      const std::basic_string<Char, Traits, Alloc>& s) FMT_NOEXCEPT
+      : data_(s.data()),
+        size_(s.size()) {}
 
-  template <
-      typename S,
-      FMT_ENABLE_IF(std::is_same<S, internal::std_string_view<Char>>::value)>
+  template <typename S, FMT_ENABLE_IF(std::is_same<
+                                      S, detail::std_string_view<Char>>::value)>
   FMT_CONSTEXPR basic_string_view(S s) FMT_NOEXCEPT : data_(s.data()),
                                                       size_(s.size()) {}
 
   /** Returns a pointer to the string data. */
-  FMT_CONSTEXPR const Char* data() const { return data_; }
+  constexpr const Char* data() const { return data_; }
 
   /** Returns the string size. */
-  FMT_CONSTEXPR size_t size() const { return size_; }
+  constexpr size_t size() const { return size_; }
 
-  FMT_CONSTEXPR iterator begin() const { return data_; }
-  FMT_CONSTEXPR iterator end() const { return data_ + size_; }
+  constexpr iterator begin() const { return data_; }
+  constexpr iterator end() const { return data_ + size_; }
+
+  constexpr const Char& operator[](size_t pos) const { return data_[pos]; }
 
   FMT_CONSTEXPR void remove_prefix(size_t n) {
     data_ += n;
@@ -329,16 +440,11 @@ template <typename Char> class basic_string_view {
 using string_view = basic_string_view<char>;
 using wstring_view = basic_string_view<wchar_t>;
 
-#ifndef __cpp_char8_t
-// A UTF-8 code unit type.
-enum char8_t : unsigned char {};
-#endif
-
 /** Specifies if ``T`` is a character type. Can be specialized by users. */
 template <typename T> struct is_char : std::false_type {};
 template <> struct is_char<char> : std::true_type {};
 template <> struct is_char<wchar_t> : std::true_type {};
-template <> struct is_char<char8_t> : std::true_type {};
+template <> struct is_char<detail::char8_type> : std::true_type {};
 template <> struct is_char<char16_t> : std::true_type {};
 template <> struct is_char<char32_t> : std::true_type {};
 
@@ -363,10 +469,10 @@ inline basic_string_view<Char> to_string_view(const Char* s) {
   return s;
 }
 
-template <typename Char, typename Traits, typename Allocator>
+template <typename Char, typename Traits, typename Alloc>
 inline basic_string_view<Char> to_string_view(
-    const std::basic_string<Char, Traits, Allocator>& s) {
-  return {s.data(), s.size()};
+    const std::basic_string<Char, Traits, Alloc>& s) {
+  return s;
 }
 
 template <typename Char>
@@ -375,14 +481,13 @@ inline basic_string_view<Char> to_string_view(basic_string_view<Char> s) {
 }
 
 template <typename Char,
-          FMT_ENABLE_IF(!std::is_empty<internal::std_string_view<Char>>::value)>
-inline basic_string_view<Char> to_string_view(
-    internal::std_string_view<Char> s) {
+          FMT_ENABLE_IF(!std::is_empty<detail::std_string_view<Char>>::value)>
+inline basic_string_view<Char> to_string_view(detail::std_string_view<Char> s) {
   return s;
 }
 
 // A base class for compile-time strings. It is defined in the fmt namespace to
-// make formatting functions visible via ADL, e.g. format(fmt("{}"), 42).
+// make formatting functions visible via ADL, e.g. format(FMT_STRING("{}"), 42).
 struct compile_string {};
 
 template <typename S>
@@ -393,9 +498,9 @@ constexpr basic_string_view<typename S::char_type> to_string_view(const S& s) {
   return s;
 }
 
-namespace internal {
+namespace detail {
 void to_string_view(...);
-using fmt::v6::to_string_view;
+using fmt::v7::to_string_view;
 
 // Specifies whether S is a string type convertible to fmt::basic_string_view.
 // It should be a constexpr function but MSVC 2017 fails to compile it in
@@ -407,25 +512,51 @@ struct is_string : std::is_class<decltype(to_string_view(std::declval<S>()))> {
 template <typename S, typename = void> struct char_t_impl {};
 template <typename S> struct char_t_impl<S, enable_if_t<is_string<S>::value>> {
   using result = decltype(to_string_view(std::declval<S>()));
-  using type = typename result::char_type;
+  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 {
-  FMT_CONSTEXPR error_handler() {}
-  FMT_CONSTEXPR error_handler(const error_handler&) {}
+  constexpr error_handler() = default;
+  constexpr error_handler(const error_handler&) = default;
 
   // This function is intentionally not constexpr to give a compile-time error.
   FMT_NORETURN FMT_API void on_error(const char* message);
 };
-}  // namespace internal
+}  // namespace detail
 
 /** String's character type. */
-template <typename S> using char_t = typename internal::char_t_impl<S>::type;
+template <typename S> using char_t = typename detail::char_t_impl<S>::type;
 
-// Parsing context consisting of a format string range being parsed and an
-// argument counter for automatic indexing.
-template <typename Char, typename ErrorHandler = internal::error_handler>
-class basic_parse_context : private ErrorHandler {
+/**
+  \rst
+  Parsing context consisting of a format string range being parsed and an
+  argument counter for automatic indexing.
+
+  You can use one of the following type aliases for common character types:
+
+  +-----------------------+-------------------------------------+
+  | Type                  | Definition                          |
+  +=======================+=====================================+
+  | format_parse_context  | basic_format_parse_context<char>    |
+  +-----------------------+-------------------------------------+
+  | wformat_parse_context | basic_format_parse_context<wchar_t> |
+  +-----------------------+-------------------------------------+
+  \endrst
+ */
+template <typename Char, typename ErrorHandler = detail::error_handler>
+class basic_format_parse_context : private ErrorHandler {
  private:
   basic_string_view<Char> format_str_;
   int next_arg_id_;
@@ -434,38 +565,48 @@ class basic_parse_context : private ErrorHandler {
   using char_type = Char;
   using iterator = typename basic_string_view<Char>::iterator;
 
-  explicit FMT_CONSTEXPR basic_parse_context(basic_string_view<Char> format_str,
-                                             ErrorHandler eh = ErrorHandler())
-      : ErrorHandler(eh), format_str_(format_str), next_arg_id_(0) {}
+  explicit constexpr basic_format_parse_context(
+      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
-  // parsed.
-  FMT_CONSTEXPR iterator begin() const FMT_NOEXCEPT {
-    return format_str_.begin();
-  }
+  /**
+    Returns an iterator to the beginning of the format string range being
+    parsed.
+   */
+  constexpr iterator begin() const FMT_NOEXCEPT { return format_str_.begin(); }
 
-  // Returns an iterator past the end of the format string range being parsed.
-  FMT_CONSTEXPR iterator end() const FMT_NOEXCEPT { return format_str_.end(); }
+  /**
+    Returns an iterator past the end of the format string range being parsed.
+   */
+  constexpr iterator end() const FMT_NOEXCEPT { return format_str_.end(); }
 
-  // Advances the begin iterator to ``it``.
+  /** Advances the begin iterator to ``it``. */
   FMT_CONSTEXPR void advance_to(iterator it) {
-    format_str_.remove_prefix(internal::to_unsigned(it - begin()));
+    format_str_.remove_prefix(detail::to_unsigned(it - begin()));
   }
 
-  // Returns the next argument index.
+  /**
+    Reports an error if using the manual argument indexing; otherwise returns
+    the next argument index and switches to the automatic indexing.
+   */
   FMT_CONSTEXPR int next_arg_id() {
+    // Don't check if the argument id is valid to avoid overhead and because it
+    // will be checked during formatting anyway.
     if (next_arg_id_ >= 0) return next_arg_id_++;
     on_error("cannot switch from manual to automatic argument indexing");
     return 0;
   }
 
-  FMT_CONSTEXPR bool check_arg_id(int) {
-    if (next_arg_id_ > 0) {
+  /**
+    Reports an error if using the automatic argument indexing; otherwise
+    switches to the manual indexing.
+   */
+  FMT_CONSTEXPR void check_arg_id(int) {
+    if (next_arg_id_ > 0)
       on_error("cannot switch from automatic to manual argument indexing");
-      return false;
-    }
-    next_arg_id_ = -1;
-    return true;
+    else
+      next_arg_id_ = -1;
   }
 
   FMT_CONSTEXPR void check_arg_id(basic_string_view<Char>) {}
@@ -474,17 +615,15 @@ class basic_parse_context : private ErrorHandler {
     ErrorHandler::on_error(message);
   }
 
-  FMT_CONSTEXPR ErrorHandler error_handler() const { return *this; }
+  constexpr ErrorHandler error_handler() const { return *this; }
 };
 
-using format_parse_context = basic_parse_context<char>;
-using wformat_parse_context = basic_parse_context<wchar_t>;
-
-using parse_context FMT_DEPRECATED_ALIAS = basic_parse_context<char>;
-using wparse_context FMT_DEPRECATED_ALIAS = basic_parse_context<wchar_t>;
+using format_parse_context = basic_format_parse_context<char>;
+using wformat_parse_context = basic_format_parse_context<wchar_t>;
 
 template <typename Context> class basic_format_arg;
 template <typename Context> class basic_format_args;
+template <typename Context> class dynamic_format_arg_store;
 
 // A formatter for objects of type T.
 template <typename T, typename Char = char, typename Enable = void>
@@ -493,112 +632,18 @@ struct formatter {
   formatter() = delete;
 };
 
-template <typename T, typename Char, typename Enable = void>
-struct FMT_DEPRECATED convert_to_int
-    : bool_constant<!std::is_arithmetic<T>::value &&
-                    std::is_convertible<T, int>::value> {};
-
-namespace internal {
-
 // Specifies if T has an enabled formatter specialization. A type can be
 // formattable even if it doesn't have a formatter e.g. via a conversion.
 template <typename T, typename Context>
 using has_formatter =
     std::is_constructible<typename Context::template formatter_type<T>>;
 
-/** A contiguous memory buffer with an optional growing ability. */
-template <typename T> class buffer {
- private:
-  buffer(const buffer&) = delete;
-  void operator=(const buffer&) = delete;
+// 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 {};
 
-  T* ptr_;
-  std::size_t size_;
-  std::size_t capacity_;
-
- protected:
-  // Don't initialize ptr_ since it is not accessed to save a few cycles.
-  buffer(std::size_t sz) FMT_NOEXCEPT : size_(sz), capacity_(sz) {}
-
-  buffer(T* p = nullptr, std::size_t sz = 0, std::size_t cap = 0) FMT_NOEXCEPT
-      : ptr_(p),
-        size_(sz),
-        capacity_(cap) {}
-
-  /** Sets the buffer data and capacity. */
-  void set(T* buf_data, std::size_t buf_capacity) FMT_NOEXCEPT {
-    ptr_ = buf_data;
-    capacity_ = buf_capacity;
-  }
-
-  /** Increases the buffer capacity to hold at least *capacity* elements. */
-  virtual void grow(std::size_t capacity) = 0;
-
- public:
-  using value_type = T;
-  using const_reference = const T&;
-
-  virtual ~buffer() {}
-
-  T* begin() FMT_NOEXCEPT { return ptr_; }
-  T* end() FMT_NOEXCEPT { return ptr_ + size_; }
-
-  /** Returns the size of this buffer. */
-  std::size_t size() const FMT_NOEXCEPT { return size_; }
-
-  /** Returns the capacity of this buffer. */
-  std::size_t capacity() const FMT_NOEXCEPT { return capacity_; }
-
-  /** Returns a pointer to the buffer data. */
-  T* data() FMT_NOEXCEPT { return ptr_; }
-
-  /** 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(std::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(std::size_t new_capacity) {
-    if (new_capacity > capacity_) grow(new_capacity);
-  }
-
-  void push_back(const T& value) {
-    reserve(size_ + 1);
-    ptr_[size_++] = value;
-  }
-
-  /** Appends data to the end of the buffer. */
-  template <typename U> void append(const U* begin, const U* end);
-
-  T& operator[](std::size_t index) { return ptr_[index]; }
-  const T& operator[](std::size_t index) const { return ptr_[index]; }
-};
-
-// A container-backed buffer.
-template <typename Container>
-class container_buffer : public buffer<typename Container::value_type> {
- private:
-  Container& container_;
-
- protected:
-  void grow(std::size_t capacity) FMT_OVERRIDE {
-    container_.resize(capacity);
-    this->set(&container_[0], capacity);
-  }
-
- public:
-  explicit container_buffer(Container& c)
-      : buffer<typename Container::value_type>(c.size()), container_(c) {}
-};
+namespace detail {
 
 // Extracts a reference to the container from back_insert_iterator.
 template <typename Container>
@@ -611,6 +656,242 @@ inline Container& get_container(std::back_insert_iterator<Container> it) {
   return *accessor(it).container;
 }
 
+/**
+  \rst
+  A contiguous memory buffer with an optional growing ability. It is an internal
+  class and shouldn't be used directly, only via `~fmt::basic_memory_buffer`.
+  \endrst
+ */
+template <typename T> class buffer {
+ private:
+  T* ptr_;
+  size_t size_;
+  size_t capacity_;
+
+ protected:
+  // Don't initialize ptr_ since it is not accessed to save a few cycles.
+  FMT_SUPPRESS_MSC_WARNING(26495)
+  buffer(size_t sz) FMT_NOEXCEPT : size_(sz), capacity_(sz) {}
+
+  buffer(T* p = nullptr, size_t sz = 0, size_t cap = 0) FMT_NOEXCEPT
+      : ptr_(p),
+        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;
+    capacity_ = buf_capacity;
+  }
+
+  /** Increases the buffer capacity to hold at least *capacity* elements. */
+  virtual void grow(size_t capacity) = 0;
+
+ public:
+  using value_type = T;
+  using const_reference = const T&;
+
+  buffer(const buffer&) = delete;
+  void operator=(const buffer&) = delete;
+
+  T* begin() FMT_NOEXCEPT { return ptr_; }
+  T* end() FMT_NOEXCEPT { return ptr_ + size_; }
+
+  const T* begin() const FMT_NOEXCEPT { return ptr_; }
+  const T* end() const FMT_NOEXCEPT { return ptr_ + size_; }
+
+  /** Returns the size of this buffer. */
+  size_t size() const FMT_NOEXCEPT { return size_; }
+
+  /** Returns the capacity of this buffer. */
+  size_t capacity() const FMT_NOEXCEPT { return capacity_; }
+
+  /** Returns a pointer to the buffer data. */
+  T* data() FMT_NOEXCEPT { return ptr_; }
+
+  /** Returns a pointer to the buffer data. */
+  const T* data() const FMT_NOEXCEPT { return ptr_; }
+
+  /** Clears this buffer. */
+  void clear() { size_ = 0; }
+
+  // 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) {
+    try_reserve(size_ + 1);
+    ptr_[size_++] = value;
+  }
+
+  /** Appends data to the end of the buffer. */
+  template <typename U> void append(const U* begin, const U* end);
+
+  template <typename I> T& operator[](I index) { return ptr_[index]; }
+  template <typename I> const T& operator[](I index) const {
+    return ptr_[index];
+  }
+};
+
+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 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) final FMT_OVERRIDE {
+    container_.resize(capacity);
+    this->set(&container_[0], capacity);
+  }
+
+ public:
+  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_);
+  }
+};
+
+// 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>
 struct fallback_formatter {
   fallback_formatter() = delete;
@@ -621,21 +902,92 @@ template <typename T, typename Context>
 using has_fallback_formatter =
     std::is_constructible<fallback_formatter<T, typename Context::char_type>>;
 
-template <typename Char> struct named_arg_base;
-template <typename T, typename Char> struct named_arg;
+struct view {};
 
-enum type {
+template <typename Char, typename T> struct named_arg : view {
+  const Char* name;
+  const T& value;
+  named_arg(const Char* n, const T& v) : name(n), value(v) {}
+};
+
+template <typename Char> struct named_arg_info {
+  const Char* name;
+  int id;
+};
+
+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.
+  // +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>
+  arg_data(const U&... init) : args_{T(named_args_, NUM_NAMED_ARGS), init...} {}
+  arg_data(const arg_data& other) = delete;
+  const T* args() const { return args_ + 1; }
+  named_arg_info<Char>* named_args() { return named_args_; }
+};
+
+template <typename T, typename Char, size_t NUM_ARGS>
+struct arg_data<T, Char, NUM_ARGS, 0> {
+  // +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...} {}
+  FMT_INLINE const T* args() const { return args_; }
+  FMT_INLINE std::nullptr_t named_args() { return nullptr; }
+};
+
+template <typename Char>
+inline void init_named_args(named_arg_info<Char>*, int, int) {}
+
+template <typename Char, typename T, typename... Tail>
+void init_named_args(named_arg_info<Char>* named_args, int arg_count,
+                     int named_arg_count, const T&, const Tail&... args) {
+  init_named_args(named_args, arg_count + 1, named_arg_count, args...);
+}
+
+template <typename Char, typename T, typename... Tail>
+void init_named_args(named_arg_info<Char>* named_args, int arg_count,
+                     int named_arg_count, const named_arg<Char, T>& arg,
+                     const Tail&... args) {
+  named_args[named_arg_count++] = {arg.name, arg_count};
+  init_named_args(named_args, arg_count + 1, named_arg_count, args...);
+}
+
+template <typename... Args>
+FMT_INLINE void init_named_args(std::nullptr_t, int, int, const Args&...) {}
+
+template <typename T> struct is_named_arg : std::false_type {};
+
+template <typename T, typename Char>
+struct is_named_arg<named_arg<Char, T>> : std::true_type {};
+
+template <bool B = false> constexpr size_t count() { return B ? 1 : 0; }
+template <bool B1, bool B2, bool... Tail> constexpr size_t count() {
+  return (B1 ? 1 : 0) + count<B2, Tail...>();
+}
+
+template <typename... Args> constexpr size_t count_named_args() {
+  return count<is_named_arg<Args>::value...>();
+}
+
+enum class type {
   none_type,
-  named_arg_type,
   // Integer types should go first,
   int_type,
   uint_type,
   long_long_type,
   ulong_long_type,
+  int128_type,
+  uint128_type,
   bool_type,
   char_type,
   last_integer_type = char_type,
   // followed by floating-point types.
+  float_type,
   double_type,
   long_double_type,
   last_numeric_type = long_double_type,
@@ -647,42 +999,48 @@ enum type {
 
 // Maps core type T to the corresponding type enum constant.
 template <typename T, typename Char>
-struct type_constant : std::integral_constant<type, custom_type> {};
+struct type_constant : std::integral_constant<type, type::custom_type> {};
 
 #define FMT_TYPE_CONSTANT(Type, constant) \
   template <typename Char>                \
-  struct type_constant<Type, Char> : std::integral_constant<type, constant> {}
+  struct type_constant<Type, Char>        \
+      : std::integral_constant<type, type::constant> {}
 
-FMT_TYPE_CONSTANT(const named_arg_base<Char>&, named_arg_type);
 FMT_TYPE_CONSTANT(int, int_type);
 FMT_TYPE_CONSTANT(unsigned, uint_type);
 FMT_TYPE_CONSTANT(long long, long_long_type);
 FMT_TYPE_CONSTANT(unsigned long long, ulong_long_type);
+FMT_TYPE_CONSTANT(int128_t, int128_type);
+FMT_TYPE_CONSTANT(uint128_t, uint128_type);
 FMT_TYPE_CONSTANT(bool, bool_type);
 FMT_TYPE_CONSTANT(Char, char_type);
+FMT_TYPE_CONSTANT(float, float_type);
 FMT_TYPE_CONSTANT(double, double_type);
 FMT_TYPE_CONSTANT(long double, long_double_type);
 FMT_TYPE_CONSTANT(const Char*, cstring_type);
 FMT_TYPE_CONSTANT(basic_string_view<Char>, string_type);
 FMT_TYPE_CONSTANT(const void*, pointer_type);
 
-FMT_CONSTEXPR bool is_integral(type t) {
-  FMT_ASSERT(t != named_arg_type, "invalid argument type");
-  return t > none_type && t <= last_integer_type;
+constexpr bool is_integral_type(type t) {
+  return t > type::none_type && t <= type::last_integer_type;
 }
 
-FMT_CONSTEXPR bool is_arithmetic(type t) {
-  FMT_ASSERT(t != named_arg_type, "invalid argument type");
-  return t > none_type && t <= last_numeric_type;
+constexpr bool is_arithmetic_type(type t) {
+  return t > type::none_type && t <= type::last_numeric_type;
 }
 
 template <typename Char> struct string_value {
   const Char* data;
-  std::size_t size;
+  size_t size;
+};
+
+template <typename Char> struct named_arg_value {
+  const named_arg_info<Char>* data;
+  size_t size;
 };
 
 template <typename Context> struct custom_value {
-  using parse_context = basic_parse_context<typename Context::char_type>;
+  using parse_context = typename Context::parse_context_type;
   const void* value;
   void (*format)(const void* arg, parse_context& parse_ctx, Context& ctx);
 };
@@ -697,32 +1055,40 @@ template <typename Context> class value {
     unsigned uint_value;
     long long long_long_value;
     unsigned long long ulong_long_value;
+    int128_t int128_value;
+    uint128_t uint128_value;
     bool bool_value;
     char_type char_value;
+    float float_value;
     double double_value;
     long double long_double_value;
     const void* pointer;
     string_value<char_type> string;
     custom_value<Context> custom;
-    const named_arg_base<char_type>* named_arg;
+    named_arg_value<char_type> named_args;
   };
 
-  FMT_CONSTEXPR value(int val = 0) : int_value(val) {}
-  FMT_CONSTEXPR value(unsigned val) : uint_value(val) {}
-  value(long long val) : long_long_value(val) {}
-  value(unsigned long long val) : ulong_long_value(val) {}
-  value(double val) : double_value(val) {}
-  value(long double val) : long_double_value(val) {}
-  value(bool val) : bool_value(val) {}
-  value(char_type val) : char_value(val) {}
-  value(const char_type* val) { string.data = val; }
-  value(basic_string_view<char_type> val) {
+  constexpr FMT_INLINE value(int val = 0) : int_value(val) {}
+  constexpr FMT_INLINE value(unsigned val) : uint_value(val) {}
+  FMT_INLINE value(long long val) : long_long_value(val) {}
+  FMT_INLINE value(unsigned long long val) : ulong_long_value(val) {}
+  FMT_INLINE value(int128_t val) : int128_value(val) {}
+  FMT_INLINE value(uint128_t val) : uint128_value(val) {}
+  FMT_INLINE value(float val) : float_value(val) {}
+  FMT_INLINE value(double val) : double_value(val) {}
+  FMT_INLINE value(long double val) : long_double_value(val) {}
+  FMT_INLINE value(bool val) : bool_value(val) {}
+  FMT_INLINE value(char_type val) : char_value(val) {}
+  FMT_INLINE value(const char_type* val) { string.data = val; }
+  FMT_INLINE value(basic_string_view<char_type> val) {
     string.data = val.data();
     string.size = val.size();
   }
-  value(const void* val) : pointer(val) {}
+  FMT_INLINE value(const void* val) : pointer(val) {}
+  FMT_INLINE value(const named_arg_info<char_type>* args, size_t size)
+      : named_args{args, size} {}
 
-  template <typename T> value(const T& val) {
+  template <typename T> FMT_INLINE value(const T& val) {
     custom.value = &val;
     // Get the formatter type through the context to allow different contexts
     // have different extension points, e.g. `formatter<T>` for `format` and
@@ -733,13 +1099,11 @@ template <typename Context> class value {
                          fallback_formatter<T, char_type>>>;
   }
 
-  value(const named_arg_base<char_type>& val) { named_arg = &val; }
-
  private:
   // Formats an argument of a custom type, such as a user-defined class.
   template <typename T, typename Formatter>
   static void format_custom_arg(const void* arg,
-                                basic_parse_context<char_type>& parse_ctx,
+                                typename Context::parse_context_type& parse_ctx,
                                 Context& ctx) {
     Formatter f;
     parse_ctx.advance_to(f.parse(parse_ctx));
@@ -756,6 +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;
@@ -770,6 +1136,8 @@ template <typename Context> struct arg_mapper {
   FMT_CONSTEXPR ulong_type map(unsigned long val) { return val; }
   FMT_CONSTEXPR long long map(long long val) { return val; }
   FMT_CONSTEXPR unsigned long long map(unsigned long long val) { return val; }
+  FMT_CONSTEXPR int128_t map(int128_t val) { return val; }
+  FMT_CONSTEXPR uint128_t map(uint128_t val) { return val; }
   FMT_CONSTEXPR bool map(bool val) { return val; }
 
   template <typename T, FMT_ENABLE_IF(is_char<T>::value)>
@@ -780,7 +1148,7 @@ template <typename Context> struct arg_mapper {
     return val;
   }
 
-  FMT_CONSTEXPR double map(float val) { return static_cast<double>(val); }
+  FMT_CONSTEXPR float map(float val) { return val; }
   FMT_CONSTEXPR double map(double val) { return val; }
   FMT_CONSTEXPR long double map(long double val) { return val; }
 
@@ -795,10 +1163,21 @@ template <typename Context> struct arg_mapper {
   template <typename T,
             FMT_ENABLE_IF(
                 std::is_constructible<basic_string_view<char_type>, T>::value &&
-                !is_string<T>::value)>
+                !is_string<T>::value && !has_formatter<T, Context>::value &&
+                !has_fallback_formatter<T, Context>::value)>
   FMT_CONSTEXPR basic_string_view<char_type> map(const T& val) {
     return basic_string_view<char_type>(val);
   }
+  template <
+      typename T,
+      FMT_ENABLE_IF(
+          std::is_constructible<std_string_view<char_type>, T>::value &&
+          !std::is_constructible<basic_string_view<char_type>, T>::value &&
+          !is_string<T>::value && !has_formatter<T, Context>::value &&
+          !has_fallback_formatter<T, Context>::value)>
+  FMT_CONSTEXPR basic_string_view<char_type> map(const T& val) {
+    return std_string_view<char_type>(val);
+  }
   FMT_CONSTEXPR const char* map(const signed char* val) {
     static_assert(std::is_same<char_type, char>::value, "invalid string type");
     return reinterpret_cast<const char*>(val);
@@ -807,6 +1186,14 @@ template <typename Context> struct arg_mapper {
     static_assert(std::is_same<char_type, char>::value, "invalid string type");
     return reinterpret_cast<const char*>(val);
   }
+  FMT_CONSTEXPR const char* map(signed char* val) {
+    const auto* const_val = val;
+    return map(const_val);
+  }
+  FMT_CONSTEXPR const char* map(unsigned char* val) {
+    const auto* const_val = val;
+    return map(const_val);
+  }
 
   FMT_CONSTEXPR const void* map(void* val) { return val; }
   FMT_CONSTEXPR const void* map(const void* val) { return val; }
@@ -824,8 +1211,10 @@ template <typename Context> struct arg_mapper {
             FMT_ENABLE_IF(std::is_enum<T>::value &&
                           !has_formatter<T, Context>::value &&
                           !has_fallback_formatter<T, Context>::value)>
-  FMT_CONSTEXPR int map(const T& val) {
-    return static_cast<int>(val);
+  FMT_CONSTEXPR auto map(const T& val)
+      -> decltype(std::declval<arg_mapper>().map(
+          static_cast<typename std::underlying_type<T>::type>(val))) {
+    return map(static_cast<typename std::underlying_type<T>::type>(val));
   }
   template <typename T,
             FMT_ENABLE_IF(!is_string<T>::value && !is_char<T>::value &&
@@ -836,36 +1225,36 @@ template <typename Context> struct arg_mapper {
   }
 
   template <typename T>
-  FMT_CONSTEXPR const named_arg_base<char_type>& map(
-      const named_arg<T, char_type>& val) {
-    auto arg = make_arg<Context>(val.value);
-    std::memcpy(val.data, &arg, sizeof(arg));
-    return val;
+  FMT_CONSTEXPR auto map(const named_arg<char_type, T>& val)
+      -> decltype(std::declval<arg_mapper>().map(val.value)) {
+    return map(val.value);
   }
+
+  unformattable map(...) { return {}; }
 };
 
 // A type constant after applying arg_mapper<Context>.
 template <typename T, typename Context>
 using mapped_type_constant =
-    type_constant<decltype(arg_mapper<Context>().map(std::declval<T>())),
+    type_constant<decltype(arg_mapper<Context>().map(std::declval<const T&>())),
                   typename Context::char_type>;
 
+enum { packed_arg_bits = 4 };
 // Maximum number of arguments with packed types.
-enum { max_packed_args = 15 };
-enum : unsigned long long { is_unpacked_bit = 1ull << 63 };
-
-template <typename Context> class arg_map;
-}  // namespace internal
+enum { max_packed_args = 62 / packed_arg_bits };
+enum : unsigned long long { is_unpacked_bit = 1ULL << 63 };
+enum : unsigned long long { has_named_args_bit = 1ULL << 62 };
+}  // namespace detail
 
 // A formatting argument. It is a trivially copyable/constructible type to
 // allow storage in basic_memory_buffer.
 template <typename Context> class basic_format_arg {
  private:
-  internal::value<Context> value_;
-  internal::type type_;
+  detail::value<Context> value_;
+  detail::type type_;
 
   template <typename ContextType, typename T>
-  friend FMT_CONSTEXPR basic_format_arg<ContextType> internal::make_arg(
+  friend FMT_CONSTEXPR basic_format_arg<ContextType> detail::make_arg(
       const T& value);
 
   template <typename Visitor, typename Ctx>
@@ -874,33 +1263,40 @@ template <typename Context> class basic_format_arg {
       -> decltype(vis(0));
 
   friend class basic_format_args<Context>;
-  friend class internal::arg_map<Context>;
+  friend class dynamic_format_arg_store<Context>;
 
   using char_type = typename Context::char_type;
 
+  template <typename T, typename Char, size_t NUM_ARGS, size_t NUM_NAMED_ARGS>
+  friend struct detail::arg_data;
+
+  basic_format_arg(const detail::named_arg_info<char_type>* args, size_t size)
+      : value_(args, size) {}
+
  public:
   class handle {
    public:
-    explicit handle(internal::custom_value<Context> custom) : custom_(custom) {}
+    explicit handle(detail::custom_value<Context> custom) : custom_(custom) {}
 
-    void format(basic_parse_context<char_type>& parse_ctx, Context& ctx) const {
+    void format(typename Context::parse_context_type& parse_ctx,
+                Context& ctx) const {
       custom_.format(custom_.value, parse_ctx, ctx);
     }
 
    private:
-    internal::custom_value<Context> custom_;
+    detail::custom_value<Context> custom_;
   };
 
-  FMT_CONSTEXPR basic_format_arg() : type_(internal::none_type) {}
+  constexpr basic_format_arg() : type_(detail::type::none_type) {}
 
-  FMT_CONSTEXPR explicit operator bool() const FMT_NOEXCEPT {
-    return type_ != internal::none_type;
+  constexpr explicit operator bool() const FMT_NOEXCEPT {
+    return type_ != detail::type::none_type;
   }
 
-  internal::type type() const { return type_; }
+  detail::type type() const { return type_; }
 
-  bool is_integral() const { return internal::is_integral(type_); }
-  bool is_arithmetic() const { return internal::is_arithmetic(type_); }
+  bool is_integral() const { return detail::is_integral_type(type_); }
+  bool is_arithmetic() const { return detail::is_arithmetic_type(type_); }
 };
 
 /**
@@ -911,81 +1307,86 @@ template <typename Context> class basic_format_arg {
   \endrst
  */
 template <typename Visitor, typename Context>
-FMT_CONSTEXPR auto visit_format_arg(Visitor&& vis,
-                                    const basic_format_arg<Context>& arg)
-    -> decltype(vis(0)) {
+FMT_CONSTEXPR_DECL FMT_INLINE auto visit_format_arg(
+    Visitor&& vis, const basic_format_arg<Context>& arg) -> decltype(vis(0)) {
   using char_type = typename Context::char_type;
   switch (arg.type_) {
-  case internal::none_type:
+  case detail::type::none_type:
     break;
-  case internal::named_arg_type:
-    FMT_ASSERT(false, "invalid argument type");
-    break;
-  case internal::int_type:
+  case detail::type::int_type:
     return vis(arg.value_.int_value);
-  case internal::uint_type:
+  case detail::type::uint_type:
     return vis(arg.value_.uint_value);
-  case internal::long_long_type:
+  case detail::type::long_long_type:
     return vis(arg.value_.long_long_value);
-  case internal::ulong_long_type:
+  case detail::type::ulong_long_type:
     return vis(arg.value_.ulong_long_value);
-  case internal::bool_type:
+#if FMT_USE_INT128
+  case detail::type::int128_type:
+    return vis(arg.value_.int128_value);
+  case detail::type::uint128_type:
+    return vis(arg.value_.uint128_value);
+#else
+  case detail::type::int128_type:
+  case detail::type::uint128_type:
+    break;
+#endif
+  case detail::type::bool_type:
     return vis(arg.value_.bool_value);
-  case internal::char_type:
+  case detail::type::char_type:
     return vis(arg.value_.char_value);
-  case internal::double_type:
+  case detail::type::float_type:
+    return vis(arg.value_.float_value);
+  case detail::type::double_type:
     return vis(arg.value_.double_value);
-  case internal::long_double_type:
+  case detail::type::long_double_type:
     return vis(arg.value_.long_double_value);
-  case internal::cstring_type:
+  case detail::type::cstring_type:
     return vis(arg.value_.string.data);
-  case internal::string_type:
+  case detail::type::string_type:
     return vis(basic_string_view<char_type>(arg.value_.string.data,
                                             arg.value_.string.size));
-  case internal::pointer_type:
+  case detail::type::pointer_type:
     return vis(arg.value_.pointer);
-  case internal::custom_type:
+  case detail::type::custom_type:
     return vis(typename basic_format_arg<Context>::handle(arg.value_.custom));
   }
   return vis(monostate());
 }
 
-namespace internal {
-// A map from argument names to their values for named arguments.
-template <typename Context> class arg_map {
- private:
-  arg_map(const arg_map&) = delete;
-  void operator=(const arg_map&) = delete;
+template <typename T> struct formattable : std::false_type {};
 
-  using char_type = typename Context::char_type;
+namespace detail {
 
-  struct entry {
-    basic_string_view<char_type> name;
-    basic_format_arg<Context> arg;
-  };
+// 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;
 
-  entry* map_;
-  unsigned size_;
+template <typename It, typename T, typename Enable = void>
+struct is_output_iterator : std::false_type {};
 
-  void push_back(value<Context> val) {
-    const auto& named = *val.named_arg;
-    map_[size_] = {named.name, named.template deserialize<Context>()};
-    ++size_;
-  }
+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 {};
 
- public:
-  arg_map() : map_(nullptr), size_(0) {}
-  void init(const basic_format_args<Context>& args);
-  ~arg_map() { delete[] map_; }
+template <typename OutputIt>
+struct is_back_insert_iterator : std::false_type {};
+template <typename Container>
+struct is_back_insert_iterator<std::back_insert_iterator<Container>>
+    : std::true_type {};
 
-  basic_format_arg<Context> find(basic_string_view<char_type> name) const {
-    // The list is unsorted, so just return the first matching name.
-    for (entry *it = map_, *end = map_ + size_; it != end; ++it) {
-      if (it->name == name) return it->arg;
-    }
-    return {};
-  }
-};
+template <typename OutputIt>
+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 {
@@ -996,6 +1397,8 @@ class locale_ref {
   locale_ref() : locale_(nullptr) {}
   template <typename Locale> explicit locale_ref(const Locale& loc);
 
+  explicit operator bool() const FMT_NOEXCEPT { return locale_ != nullptr; }
+
   template <typename Locale> Locale get() const;
 };
 
@@ -1003,8 +1406,8 @@ template <typename> constexpr unsigned long long encode_types() { return 0; }
 
 template <typename Context, typename Arg, typename... Args>
 constexpr unsigned long long encode_types() {
-  return mapped_type_constant<Arg, Context>::value |
-         (encode_types<Context, Args...>() << 4);
+  return static_cast<unsigned>(mapped_type_constant<Arg, Context>::value) |
+         (encode_types<Context, Args...>() << packed_arg_bits);
 }
 
 template <typename Context, typename T>
@@ -1015,18 +1418,73 @@ FMT_CONSTEXPR basic_format_arg<Context> make_arg(const T& value) {
   return arg;
 }
 
-template <bool IS_PACKED, typename Context, typename T,
-          FMT_ENABLE_IF(IS_PACKED)>
-inline value<Context> make_arg(const T& val) {
-  return arg_mapper<Context>().map(val);
+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;
 }
 
-template <bool IS_PACKED, typename Context, typename T,
+// 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 check<T>(arg_mapper<Context>().map(val));
+}
+
+template <bool IS_PACKED, typename Context, type, typename T,
           FMT_ENABLE_IF(!IS_PACKED)>
 inline basic_format_arg<Context> make_arg(const T& value) {
   return make_arg<Context>(value);
 }
-}  // namespace internal
+
+template <typename T> struct is_reference_wrapper : std::false_type {};
+template <typename T>
+struct is_reference_wrapper<std::reference_wrapper<T>> : std::true_type {};
+
+template <typename T> const T& unwrap(const T& v) { return v; }
+template <typename T> const T& unwrap(const std::reference_wrapper<T>& v) {
+  return static_cast<const T&>(v);
+}
+
+class dynamic_arg_list {
+  // Workaround for clang's -Wweak-vtables. Unlike for regular classes, for
+  // templates it doesn't complain about inability to deduce single translation
+  // unit for placing vtable. So storage_node_base is made a fake template.
+  template <typename = void> struct node {
+    virtual ~node() = default;
+    std::unique_ptr<node<>> next;
+  };
+
+  template <typename T> struct typed_node : node<> {
+    T value;
+
+    template <typename Arg>
+    FMT_CONSTEXPR typed_node(const Arg& arg) : value(arg) {}
+
+    template <typename Char>
+    FMT_CONSTEXPR typed_node(const basic_string_view<Char>& arg)
+        : value(arg.data(), arg.size()) {}
+  };
+
+  std::unique_ptr<node<>> head_;
+
+ public:
+  template <typename T, typename Arg> const T& push(const Arg& arg) {
+    auto new_node = std::unique_ptr<typed_node<T>>(new typed_node<T>(arg));
+    auto& value = new_node->value;
+    new_node->next = std::move(head_);
+    head_ = std::move(new_node);
+    return value;
+  }
+};
+}  // namespace detail
 
 // Formatting context.
 template <typename OutputIt, typename Char> class basic_format_context {
@@ -1037,51 +1495,54 @@ template <typename OutputIt, typename Char> class basic_format_context {
  private:
   OutputIt out_;
   basic_format_args<basic_format_context> args_;
-  internal::arg_map<basic_format_context> map_;
-  internal::locale_ref loc_;
-
-  basic_format_context(const basic_format_context&) = delete;
-  void operator=(const basic_format_context&) = delete;
+  detail::locale_ref loc_;
 
  public:
   using iterator = OutputIt;
   using format_arg = basic_format_arg<basic_format_context>;
+  using parse_context_type = basic_format_parse_context<Char>;
   template <typename T> using formatter_type = formatter<T, char_type>;
 
+  basic_format_context(const basic_format_context&) = delete;
+  void operator=(const basic_format_context&) = delete;
   /**
    Constructs a ``basic_format_context`` object. References to the arguments are
    stored in the object so make sure they have appropriate lifetimes.
    */
   basic_format_context(OutputIt out,
                        basic_format_args<basic_format_context> ctx_args,
-                       internal::locale_ref loc = internal::locale_ref())
+                       detail::locale_ref loc = detail::locale_ref())
       : out_(out), args_(ctx_args), loc_(loc) {}
 
   format_arg arg(int id) const { return args_.get(id); }
+  format_arg arg(basic_string_view<char_type> name) { return args_.get(name); }
+  int arg_id(basic_string_view<char_type> name) { return args_.get_id(name); }
+  const basic_format_args<basic_format_context>& args() const { return args_; }
 
-  // Checks if manual indexing is used and returns the argument with the
-  // specified name.
-  format_arg arg(basic_string_view<char_type> name);
-
-  internal::error_handler error_handler() { return {}; }
+  detail::error_handler error_handler() { return {}; }
   void on_error(const char* message) { error_handler().on_error(message); }
 
   // Returns an iterator to the beginning of the output range.
   iterator out() { return out_; }
 
   // Advances the begin iterator to ``it``.
-  void advance_to(iterator it) { out_ = it; }
+  void advance_to(iterator it) {
+    if (!detail::is_back_insert_iterator<iterator>()) out_ = it;
+  }
 
-  internal::locale_ref locale() { return loc_; }
+  detail::locale_ref locale() { return loc_; }
 };
 
 template <typename Char>
 using buffer_context =
-    basic_format_context<std::back_insert_iterator<internal::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 an alias issue: https://stackoverflow.com/q/62767544/471164.
+#define FMT_BUFFER_CONTEXT(Char) \
+  basic_format_context<detail::buffer_appender<Char>, Char>
+
 /**
   \rst
   An array of references to arguments. It can be implicitly converted into
@@ -1089,32 +1550,50 @@ using wformat_context = buffer_context<wchar_t>;
   such as `~fmt::vformat`.
   \endrst
  */
-template <typename Context, typename... Args> class format_arg_store {
+template <typename Context, typename... Args>
+class format_arg_store
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+    // Workaround a GCC template argument substitution bug.
+    : public basic_format_args<Context>
+#endif
+{
  private:
   static const size_t num_args = sizeof...(Args);
-  static const bool is_packed = num_args < internal::max_packed_args;
+  static const size_t num_named_args = detail::count_named_args<Args...>();
+  static const bool is_packed = num_args <= detail::max_packed_args;
 
-  using value_type = conditional_t<is_packed, internal::value<Context>,
+  using value_type = conditional_t<is_packed, detail::value<Context>,
                                    basic_format_arg<Context>>;
 
-  // If the arguments are not packed, add one more element to mark the end.
-  value_type data_[num_args + (num_args == 0 ? 1 : 0)];
+  detail::arg_data<value_type, typename Context::char_type, num_args,
+                   num_named_args>
+      data_;
 
   friend class basic_format_args<Context>;
 
- public:
-  static constexpr unsigned long long types =
-      is_packed ? internal::encode_types<Context, Args...>()
-                : internal::is_unpacked_bit | num_args;
-  FMT_DEPRECATED static constexpr unsigned long long TYPES = types;
+  static constexpr unsigned long long desc =
+      (is_packed ? detail::encode_types<Context, Args...>()
+                 : detail::is_unpacked_bit | num_args) |
+      (num_named_args != 0
+           ? static_cast<unsigned long long>(detail::has_named_args_bit)
+           : 0);
 
+ public:
   format_arg_store(const Args&... args)
-      : data_{internal::make_arg<is_packed, Context>(args)...} {}
+      :
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+        basic_format_args<Context>(*this),
+#endif
+        data_{detail::make_arg<
+            is_packed, Context,
+            detail::mapped_type_constant<Args, Context>::value>(args)...} {
+    detail::init_named_args(data_.named_args(), 0, 0, args...);
+  }
 };
 
 /**
   \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.
@@ -1126,55 +1605,266 @@ inline format_arg_store<Context, Args...> make_format_args(
   return {args...};
 }
 
-/** Formatting arguments. */
+/**
+  \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
+  be used in a call to a formatting function.
+
+  **Example**::
+
+    fmt::print("Elapsed time: {s:.2f} seconds", fmt::arg("s", 1.23));
+  \endrst
+ */
+template <typename Char, typename T>
+inline detail::named_arg<Char, T> arg(const Char* name, const T& arg) {
+  static_assert(!detail::is_named_arg<T>(), "nested named arguments");
+  return {name, arg};
+}
+
+/**
+  \rst
+  A dynamic version of `fmt::format_arg_store`.
+  It's equipped with a storage to potentially temporary objects which lifetimes
+  could be shorter than the format arguments object.
+
+  It can be implicitly converted into `~fmt::basic_format_args` for passing
+  into type-erased formatting functions such as `~fmt::vformat`.
+  \endrst
+ */
+template <typename Context>
+class dynamic_format_arg_store
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+    // Workaround a GCC template argument substitution bug.
+    : public basic_format_args<Context>
+#endif
+{
+ private:
+  using char_type = typename Context::char_type;
+
+  template <typename T> struct need_copy {
+    static constexpr detail::type mapped_type =
+        detail::mapped_type_constant<T, Context>::value;
+
+    enum {
+      value = !(detail::is_reference_wrapper<T>::value ||
+                std::is_same<T, basic_string_view<char_type>>::value ||
+                std::is_same<T, detail::std_string_view<char_type>>::value ||
+                (mapped_type != detail::type::cstring_type &&
+                 mapped_type != detail::type::string_type &&
+                 mapped_type != detail::type::custom_type))
+    };
+  };
+
+  template <typename T>
+  using stored_type = conditional_t<detail::is_string<T>::value,
+                                    std::basic_string<char_type>, T>;
+
+  // Storage of basic_format_arg must be contiguous.
+  std::vector<basic_format_arg<Context>> data_;
+  std::vector<detail::named_arg_info<char_type>> named_info_;
+
+  // Storage of arguments not fitting into basic_format_arg must grow
+  // without relocation because items in data_ refer to it.
+  detail::dynamic_arg_list dynamic_args_;
+
+  friend class basic_format_args<Context>;
+
+  unsigned long long get_types() const {
+    return detail::is_unpacked_bit | data_.size() |
+           (named_info_.empty()
+                ? 0ULL
+                : static_cast<unsigned long long>(detail::has_named_args_bit));
+  }
+
+  const basic_format_arg<Context>* data() const {
+    return named_info_.empty() ? data_.data() : data_.data() + 1;
+  }
+
+  template <typename T> void emplace_arg(const T& arg) {
+    data_.emplace_back(detail::make_arg<Context>(arg));
+  }
+
+  template <typename T>
+  void emplace_arg(const detail::named_arg<char_type, T>& arg) {
+    if (named_info_.empty()) {
+      constexpr const detail::named_arg_info<char_type>* zero_ptr{nullptr};
+      data_.insert(data_.begin(), {zero_ptr, 0});
+    }
+    data_.emplace_back(detail::make_arg<Context>(detail::unwrap(arg.value)));
+    auto pop_one = [](std::vector<basic_format_arg<Context>>* data) {
+      data->pop_back();
+    };
+    std::unique_ptr<std::vector<basic_format_arg<Context>>, decltype(pop_one)>
+        guard{&data_, pop_one};
+    named_info_.push_back({arg.name, static_cast<int>(data_.size() - 2u)});
+    data_[0].value_.named_args = {named_info_.data(), named_info_.size()};
+    guard.release();
+  }
+
+ public:
+  /**
+    \rst
+    Adds an argument into the dynamic store for later passing to a formatting
+    function.
+
+    Note that custom types and string types (but not string views) are copied
+    into the store dynamically allocating memory if necessary.
+
+    **Example**::
+
+      fmt::dynamic_format_arg_store<fmt::format_context> store;
+      store.push_back(42);
+      store.push_back("abc");
+      store.push_back(1.5f);
+      std::string result = fmt::vformat("{} and {} and {}", store);
+    \endrst
+  */
+  template <typename T> void push_back(const T& arg) {
+    if (detail::const_check(need_copy<T>::value))
+      emplace_arg(dynamic_args_.push<stored_type<T>>(arg));
+    else
+      emplace_arg(detail::unwrap(arg));
+  }
+
+  /**
+    \rst
+    Adds a reference to the argument into the dynamic store for later passing to
+    a formatting function. Supports named arguments wrapped in
+    ``std::reference_wrapper`` via ``std::ref()``/``std::cref()``.
+
+    **Example**::
+
+      fmt::dynamic_format_arg_store<fmt::format_context> store;
+      char str[] = "1234567890";
+      store.push_back(std::cref(str));
+      int a1_val{42};
+      auto a1 = fmt::arg("a1_", a1_val);
+      store.push_back(std::cref(a1));
+
+      // Changing str affects the output but only for string and custom types.
+      str[0] = 'X';
+
+      std::string result = fmt::vformat("{} and {a1_}");
+      assert(result == "X234567890 and 42");
+    \endrst
+  */
+  template <typename T> void push_back(std::reference_wrapper<T> arg) {
+    static_assert(
+        detail::is_named_arg<typename std::remove_cv<T>::type>::value ||
+            need_copy<T>::value,
+        "objects of built-in types and string views are always copied");
+    emplace_arg(arg.get());
+  }
+
+  /**
+    Adds named argument into the dynamic store for later passing to a formatting
+    function. ``std::reference_wrapper`` is supported to avoid copying of the
+    argument.
+  */
+  template <typename T>
+  void push_back(const detail::named_arg<char_type, T>& arg) {
+    const char_type* arg_name =
+        dynamic_args_.push<std::basic_string<char_type>>(arg.name).c_str();
+    if (detail::const_check(need_copy<T>::value)) {
+      emplace_arg(
+          fmt::arg(arg_name, dynamic_args_.push<stored_type<T>>(arg.value)));
+    } else {
+      emplace_arg(fmt::arg(arg_name, arg.value));
+    }
+  }
+
+  /** Erase all elements from the store */
+  void clear() {
+    data_.clear();
+    named_info_.clear();
+    dynamic_args_ = detail::dynamic_arg_list();
+  }
+
+  /**
+    \rst
+    Reserves space to store at least *new_cap* arguments including
+    *new_cap_named* named arguments.
+    \endrst
+  */
+  void reserve(size_t new_cap, size_t new_cap_named) {
+    FMT_ASSERT(new_cap >= new_cap_named,
+               "Set of arguments includes set of named arguments");
+    data_.reserve(new_cap);
+    named_info_.reserve(new_cap_named);
+  }
+};
+
+/**
+  \rst
+  A view of a collection of formatting arguments. To avoid lifetime issues it
+  should only be used as a parameter type in type-erased functions such as
+  ``vformat``::
+
+    void vlog(string_view format_str, format_args args);  // OK
+    format_args args = make_format_args(42);  // Error: dangling reference
+  \endrst
+ */
 template <typename Context> class basic_format_args {
  public:
   using size_type = int;
   using format_arg = basic_format_arg<Context>;
 
  private:
-  // To reduce compiled code size per formatting function call, types of first
-  // max_packed_args arguments are passed in the types_ field.
-  unsigned long long types_;
+  // A descriptor that contains information about formatting arguments.
+  // If the number of arguments is less or equal to max_packed_args then
+  // argument types are passed in the descriptor. This reduces binary code size
+  // per formatting function call.
+  unsigned long long desc_;
   union {
-    // If the number of arguments is less than max_packed_args, the argument
-    // values are stored in values_, otherwise they are stored in args_.
-    // This is done to reduce compiled code size as storing larger objects
+    // If is_packed() returns true then argument values are stored in values_;
+    // otherwise they are stored in args_. This is done to improve cache
+    // locality and reduce compiled code size since storing larger objects
     // may require more code (at least on x86-64) even if the same amount of
     // data is actually copied to stack. It saves ~10% on the bloat test.
-    const internal::value<Context>* values_;
+    const detail::value<Context>* values_;
     const format_arg* args_;
   };
 
-  bool is_packed() const { return (types_ & internal::is_unpacked_bit) == 0; }
-
-  internal::type type(int index) const {
-    int shift = index * 4;
-    return static_cast<internal::type>((types_ & (0xfull << shift)) >> shift);
+  bool is_packed() const { return (desc_ & detail::is_unpacked_bit) == 0; }
+  bool has_named_args() const {
+    return (desc_ & detail::has_named_args_bit) != 0;
   }
 
-  friend class internal::arg_map<Context>;
-
-  void set_data(const internal::value<Context>* values) { values_ = values; }
-  void set_data(const format_arg* args) { args_ = args; }
-
-  format_arg do_get(int index) const {
-    format_arg arg;
-    if (!is_packed()) {
-      auto num_args = max_size();
-      if (index < num_args) arg = args_[index];
-      return arg;
-    }
-    if (index > internal::max_packed_args) return arg;
-    arg.type_ = type(index);
-    if (arg.type_ == internal::none_type) return arg;
-    internal::value<Context>& val = arg.value_;
-    val = values_[index];
-    return arg;
+  detail::type type(int index) const {
+    int shift = index * detail::packed_arg_bits;
+    unsigned int mask = (1 << detail::packed_arg_bits) - 1;
+    return static_cast<detail::type>((desc_ >> shift) & mask);
   }
 
+  basic_format_args(unsigned long long desc,
+                    const detail::value<Context>* values)
+      : desc_(desc), values_(values) {}
+  basic_format_args(unsigned long long desc, const format_arg* args)
+      : desc_(desc), args_(args) {}
+
  public:
-  basic_format_args() : types_(0) {}
+  basic_format_args() : desc_(0) {}
 
   /**
    \rst
@@ -1182,10 +1872,17 @@ template <typename Context> class basic_format_args {
    \endrst
    */
   template <typename... Args>
-  basic_format_args(const format_arg_store<Context, Args...>& store)
-      : types_(static_cast<unsigned long long>(store.types)) {
-    set_data(store.data_);
-  }
+  FMT_INLINE basic_format_args(const format_arg_store<Context, Args...>& store)
+      : basic_format_args(store.desc, store.data_.args()) {}
+
+  /**
+   \rst
+   Constructs a `basic_format_args` object from
+   `~fmt::dynamic_format_arg_store`.
+   \endrst
+   */
+  FMT_INLINE basic_format_args(const dynamic_format_arg_store<Context>& store)
+      : basic_format_args(store.get_types(), store.data()) {}
 
   /**
    \rst
@@ -1193,168 +1890,172 @@ template <typename Context> class basic_format_args {
    \endrst
    */
   basic_format_args(const format_arg* args, int count)
-      : types_(internal::is_unpacked_bit | internal::to_unsigned(count)) {
-    set_data(args);
+      : basic_format_args(detail::is_unpacked_bit | detail::to_unsigned(count),
+                          args) {}
+
+  /** Returns the argument with the specified id. */
+  format_arg get(int id) const {
+    format_arg arg;
+    if (!is_packed()) {
+      if (id < max_size()) arg = args_[id];
+      return arg;
+    }
+    if (id >= detail::max_packed_args) return arg;
+    arg.type_ = type(id);
+    if (arg.type_ == detail::type::none_type) return arg;
+    arg.value_ = values_[id];
+    return arg;
   }
 
-  /** Returns the argument at specified index. */
-  format_arg get(int index) const {
-    format_arg arg = do_get(index);
-    if (arg.type_ == internal::named_arg_type)
-      arg = arg.value_.named_arg->template deserialize<Context>();
-    return arg;
+  template <typename Char> format_arg get(basic_string_view<Char> name) const {
+    int id = get_id(name);
+    return id >= 0 ? get(id) : format_arg();
+  }
+
+  template <typename Char> int get_id(basic_string_view<Char> name) const {
+    if (!has_named_args()) return -1;
+    const auto& named_args =
+        (is_packed() ? values_[-1] : args_[-1].value_).named_args;
+    for (size_t i = 0; i < named_args.size; ++i) {
+      if (named_args.data[i].name == name) return named_args.data[i].id;
+    }
+    return -1;
   }
 
   int max_size() const {
-    unsigned long long max_packed = internal::max_packed_args;
+    unsigned long long max_packed = detail::max_packed_args;
     return static_cast<int>(is_packed() ? max_packed
-                                        : types_ & ~internal::is_unpacked_bit);
+                                        : desc_ & ~detail::is_unpacked_bit);
   }
 };
 
-/** 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>
-  format_args(Args&&... args)
-      : basic_format_args<format_context>(std::forward<Args>(args)...) {}
+  FMT_INLINE format_args(const Args&... args) : basic_format_args(args...) {}
 };
 struct wformat_args : basic_format_args<wformat_context> {
-  template <typename... Args>
-  wformat_args(Args&&... args)
-      : basic_format_args<wformat_context>(std::forward<Args>(args)...) {}
+  using basic_format_args::basic_format_args;
 };
-
-template <typename Container> 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<internal::buffer<Char>> : std::true_type {};
-
-namespace internal {
-
-template <typename OutputIt>
-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 named_arg_base {
-  basic_string_view<Char> name;
-
-  // Serialized value<context>.
-  mutable char data[sizeof(basic_format_arg<buffer_context<Char>>)];
-
-  named_arg_base(basic_string_view<Char> nm) : name(nm) {}
-
-  template <typename Context> basic_format_arg<Context> deserialize() const {
-    basic_format_arg<Context> arg;
-    std::memcpy(&arg, data, sizeof(basic_format_arg<Context>));
-    return arg;
-  }
-};
-
-template <typename T, typename Char> struct named_arg : named_arg_base<Char> {
-  const T& value;
-
-  named_arg(basic_string_view<Char> name, const T& val)
-      : named_arg_base<Char>(name), value(val) {}
-};
-
-template <typename..., typename S, FMT_ENABLE_IF(!is_compile_string<S>::value)>
-inline void check_format_string(const S&) {
-#if defined(FMT_ENFORCE_COMPILE_STRING)
-  static_assert(is_compile_string<S>::value,
-                "FMT_ENFORCE_COMPILE_STRING requires all format strings to "
-                "utilize FMT_STRING() or fmt().");
 #endif
-}
-template <typename..., typename S, FMT_ENABLE_IF(is_compile_string<S>::value)>
-void check_format_string(S);
 
-struct view {};
-template <bool...> struct bool_pack;
-template <bool... Args>
-using all_true =
-    std::is_same<bool_pack<Args..., true>, bool_pack<true, Args...>>;
+namespace detail {
 
-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(all_true<(!std::is_base_of<view, remove_reference_t<Args>>() ||
-                          !std::is_reference<Args>())...>::value,
-                "passing views as lvalues is disallowed");
-  check_format_string<remove_const_t<remove_reference_t<Args>>...>(format_str);
-  return {args...};
-}
+template <typename Char, FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+std::basic_string<Char> vformat(
+    basic_string_view<Char> format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args);
+
+FMT_API std::string vformat(string_view format_str, format_args args);
 
 template <typename Char>
-std::basic_string<Char> vformat(basic_string_view<Char> format_str,
-                                basic_format_args<buffer_context<Char>> args);
-
-template <typename Char>
-typename buffer_context<Char>::iterator vformat_to(
+void vformat_to(
     buffer<Char>& buf, basic_string_view<Char> format_str,
-    basic_format_args<buffer_context<Char>> args);
-}  // namespace internal
+    basic_format_args<FMT_BUFFER_CONTEXT(type_identity_t<Char>)> args,
+    detail::locale_ref loc = {});
 
-/**
-  \rst
-  Returns a named argument to be used in a formatting function.
+template <typename Char, typename Args,
+          FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+inline void vprint_mojibake(std::FILE*, basic_string_view<Char>, const Args&) {}
 
-  The named argument holds a reference and does not extend the lifetime
-  of its arguments.
-  Consequently, a dangling reference can accidentally be created.
-  The user should take care to only pass this function temporaries when
-  the named argument is itself a temporary, as per the following example.
-
-  **Example**::
-
-    fmt::print("Elapsed time: {s:.2f} seconds", fmt::arg("s", 1.23));
-  \endrst
- */
-template <typename S, typename T, typename Char = char_t<S>>
-inline internal::named_arg<T, Char> arg(const S& name, const T& arg) {
-  static_assert(internal::is_string<S>::value, "");
-  return {name, arg};
-}
-
-// Disable nested named arguments, e.g. ``arg("a", arg("b", 42))``.
-template <typename S, typename T, typename Char>
-void arg(S, internal::named_arg<T, Char>) = delete;
+FMT_API void vprint_mojibake(std::FILE*, string_view, format_args);
+#ifndef _WIN32
+inline void vprint_mojibake(std::FILE*, string_view, format_args) {}
+#endif
+}  // namespace detail
 
 /** 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(
-              internal::is_contiguous_back_insert_iterator<OutputIt>::value)>
-OutputIt vformat_to(OutputIt out, const S& format_str,
-                    basic_format_args<buffer_context<Char>> args) {
-  using container = remove_reference_t<decltype(internal::get_container(out))>;
-  internal::container_buffer<container> buf((internal::get_container(out)));
-  internal::vformat_to(buf, to_string_view(format_str), args);
-  return out;
+          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 detail::get_iterator(buf);
 }
 
-template <typename Container, typename S, typename... Args,
-          FMT_ENABLE_IF(
-              is_contiguous<Container>::value&& internal::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),
-      {internal::make_args_checked<Args...>(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
+ */
+// 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);
+}
+
+/**
+  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>>
-inline std::basic_string<Char> vformat(
-    const S& format_str, basic_format_args<buffer_context<Char>> args) {
-  return internal::vformat(to_string_view(format_str), args);
+FMT_INLINE std::basic_string<Char> vformat(
+    const S& format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  return detail::vformat(to_string_view(format_str), args);
 }
 
 /**
@@ -1370,50 +2071,51 @@ inline std::basic_string<Char> vformat(
 // Pass char_t as a default template parameter instead of using
 // std::basic_string<char_t<S>> to reduce the symbol size.
 template <typename S, typename... Args, typename Char = char_t<S>>
-inline std::basic_string<Char> format(const S& format_str, Args&&... args) {
-  return internal::vformat(
-      to_string_view(format_str),
-      {internal::make_args_checked<Args...>(format_str, args...)});
+FMT_INLINE std::basic_string<Char> format(const S& format_str, Args&&... args) {
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
+  return detail::vformat(to_string_view(format_str), vargs);
 }
 
-FMT_API void vprint(std::FILE* f, string_view format_str, format_args args);
-FMT_API void vprint(std::FILE* f, wstring_view format_str, wformat_args args);
+FMT_API void vprint(string_view, format_args);
+FMT_API void vprint(std::FILE*, string_view, format_args);
 
 /**
   \rst
-  Prints formatted data to the file *f*. For wide format strings,
-  *f* should be in wide-oriented mode set via ``fwide(f, 1)`` or
-  ``_setmode(_fileno(f), _O_U8TEXT)`` on Windows.
+  Formats ``args`` according to specifications in ``format_str`` and writes the
+  output to the file ``f``. Strings are assumed to be Unicode-encoded unless the
+  ``FMT_UNICODE`` macro is set to 0.
 
   **Example**::
 
     fmt::print(stderr, "Don't {}!", "panic");
   \endrst
  */
-template <typename S, typename... Args,
-          FMT_ENABLE_IF(internal::is_string<S>::value)>
+template <typename S, typename... Args, typename Char = char_t<S>>
 inline void print(std::FILE* f, const S& format_str, Args&&... args) {
-  vprint(f, to_string_view(format_str),
-         internal::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);
 }
 
-FMT_API void vprint(string_view format_str, format_args args);
-FMT_API void vprint(wstring_view format_str, wformat_args args);
-
 /**
   \rst
-  Prints formatted data to ``stdout``.
+  Formats ``args`` according to specifications in ``format_str`` and writes
+  the output to ``stdout``. Strings are assumed to be Unicode-encoded unless
+  the ``FMT_UNICODE`` macro is set to 0.
 
   **Example**::
 
     fmt::print("Elapsed time: {0:.2f} seconds", 1.23);
   \endrst
  */
-template <typename S, typename... Args,
-          FMT_ENABLE_IF(internal::is_string<S>::value)>
+template <typename S, typename... Args, typename Char = char_t<S>>
 inline void print(const S& format_str, Args&&... args) {
-  vprint(to_string_view(format_str),
-         internal::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),
+                                       vargs);
 }
 FMT_END_NAMESPACE
 
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/format-inl.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/format-inl.h
index 147062fe5..8f2fe7354 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/format-inl.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/format-inl.h
@@ -1,4 +1,4 @@
-// Formatting library for C++
+// Formatting library for C++ - implementation
 //
 // Copyright (c) 2012 - 2016, Victor Zverovich
 // All rights reserved.
@@ -8,63 +8,41 @@
 #ifndef FMT_FORMAT_INL_H_
 #define FMT_FORMAT_INL_H_
 
-#include "format.h"
-
-#include <string.h>
-
+#include <cassert>
 #include <cctype>
-#include <cerrno>
 #include <climits>
 #include <cmath>
 #include <cstdarg>
-#include <cstddef>  // for std::ptrdiff_t
-#include <cstring>  // for std::memmove
+#include <cstring>  // std::memmove
 #include <cwchar>
-#if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
+#include <exception>
+
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
 #  include <locale>
 #endif
 
-#if FMT_USE_WINDOWS_H
-#  if !defined(FMT_HEADER_ONLY) && !defined(WIN32_LEAN_AND_MEAN)
-#    define WIN32_LEAN_AND_MEAN
-#  endif
-#  if defined(NOMINMAX) || defined(FMT_WIN_MINMAX)
-#    include <windows.h>
-#  else
-#    define NOMINMAX
-#    include <windows.h>
-#    undef NOMINMAX
-#  endif
+#ifdef _WIN32
+#  include <io.h>  // _isatty
 #endif
 
-#if FMT_EXCEPTIONS
-#  define FMT_TRY try
-#  define FMT_CATCH(x) catch (x)
-#else
-#  define FMT_TRY if (true)
-#  define FMT_CATCH(x) if (false)
-#endif
-
-#ifdef _MSC_VER
-#  pragma warning(push)
-#  pragma warning(disable : 4127)  // conditional expression is constant
-#  pragma warning(disable : 4702)  // unreachable code
-// Disable deprecation warning for strerror. The latter is not called but
-// MSVC fails to detect it.
-#  pragma warning(disable : 4996)
-#endif
+#include "format.h"
 
 // Dummy implementations of strerror_r and strerror_s called if corresponding
 // system functions are not available.
-inline fmt::internal::null<> strerror_r(int, char*, ...) {
-  return fmt::internal::null<>();
-}
-inline fmt::internal::null<> strerror_s(char*, std::size_t, ...) {
-  return fmt::internal::null<>();
-}
+inline fmt::detail::null<> strerror_r(int, char*, ...) { return {}; }
+inline fmt::detail::null<> strerror_s(char*, size_t, ...) { return {}; }
 
 FMT_BEGIN_NAMESPACE
-namespace internal {
+namespace detail {
+
+FMT_FUNC void assert_fail(const char* file, int line, const char* message) {
+  // Use unchecked std::fprintf to avoid triggering another assertion when
+  // writing to stderr fails
+  std::fprintf(stderr, "%s:%d: assertion failed: %s", file, line, message);
+  // Chosen instead of std::abort to satisfy Clang in CUDA mode during device
+  // code pass.
+  std::terminate();
+}
 
 #ifndef _MSC_VER
 #  define FMT_SNPRINTF snprintf
@@ -79,9 +57,7 @@ inline int fmt_snprintf(char* buffer, size_t size, const char* format, ...) {
 #  define FMT_SNPRINTF fmt_snprintf
 #endif  // _MSC_VER
 
-using format_func = void (*)(internal::buffer<char>&, int, string_view);
-
-// Portable thread-safe version of strerror.
+// A portable thread-safe version of strerror.
 // Sets buffer to point to a string describing the error code.
 // This can be either a pointer to a string stored in buffer,
 // or a pointer to some static immutable string.
@@ -90,15 +66,15 @@ using format_func = void (*)(internal::buffer<char>&, int, string_view);
 //   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,
-                           std::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 {
    private:
     int error_code_;
     char*& buffer_;
-    std::size_t buffer_size_;
+    size_t buffer_size_;
 
     // A noop assignment operator to avoid bogus warnings.
     void operator=(const dispatcher&) {}
@@ -110,6 +86,7 @@ FMT_FUNC int safe_strerror(int error_code, char*& buffer,
     }
 
     // Handle the result of GNU-specific version of strerror_r.
+    FMT_MAYBE_UNUSED
     int handle(char* message) {
       // If the buffer is full then the message is probably truncated.
       if (message == buffer_ && strlen(buffer_) == buffer_size_ - 1)
@@ -119,11 +96,13 @@ FMT_FUNC int safe_strerror(int error_code, char*& buffer,
     }
 
     // Handle the case when strerror_r is not available.
-    int handle(internal::null<>) {
+    FMT_MAYBE_UNUSED
+    int handle(detail::null<>) {
       return fallback(strerror_s(buffer_, buffer_size_, error_code_));
     }
 
     // Fallback to strerror_s when strerror_r is not available.
+    FMT_MAYBE_UNUSED
     int fallback(int result) {
       // If the buffer is full then the message is probably truncated.
       return result == 0 && strlen(buffer_) == buffer_size_ - 1 ? ERANGE
@@ -132,7 +111,7 @@ FMT_FUNC int safe_strerror(int error_code, char*& buffer,
 
 #if !FMT_MSC_VER
     // Fallback to strerror if strerror_r and strerror_s are not available.
-    int fallback(internal::null<>) {
+    int fallback(detail::null<>) {
       errno = 0;
       buffer_ = strerror(error_code_);
       return errno;
@@ -140,7 +119,7 @@ FMT_FUNC int safe_strerror(int error_code, char*& buffer,
 #endif
 
    public:
-    dispatcher(int err_code, char*& buf, std::size_t buf_size)
+    dispatcher(int err_code, char*& buf, size_t buf_size)
         : error_code_(err_code), buffer_(buf), buffer_size_(buf_size) {}
 
     int run() { return handle(strerror_r(error_code_, buffer_, buffer_size_)); }
@@ -148,41 +127,29 @@ FMT_FUNC int safe_strerror(int error_code, char*& buffer,
   return dispatcher(error_code, buffer, buffer_size).run();
 }
 
-FMT_FUNC void format_error_code(internal::buffer<char>& out, int error_code,
+FMT_FUNC void format_error_code(detail::buffer<char>& out, int error_code,
                                 string_view message) FMT_NOEXCEPT {
   // 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.
-  std::size_t error_code_size = sizeof(SEP) + sizeof(ERROR_STR) - 2;
-  auto abs_value = static_cast<uint32_or_64_t<int>>(error_code);
-  if (internal::is_negative(error_code)) {
+  size_t error_code_size = sizeof(SEP) + sizeof(ERROR_STR) - 2;
+  auto abs_value = static_cast<uint32_or_64_or_128_t<int>>(error_code);
+  if (detail::is_negative(error_code)) {
     abs_value = 0 - abs_value;
     ++error_code_size;
   }
-  error_code_size += internal::to_unsigned(internal::count_digits(abs_value));
-  internal::writer w(out);
-  if (message.size() <= inline_buffer_size - error_code_size) {
-    w.write(message);
-    w.write(SEP);
-  }
-  w.write(ERROR_STR);
-  w.write(error_code);
+  error_code_size += detail::to_unsigned(detail::count_digits(abs_value));
+  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);
   assert(out.size() <= inline_buffer_size);
 }
 
-// A wrapper around fwrite that throws on error.
-FMT_FUNC 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"));
-  }
-}
-
 FMT_FUNC void report_error(format_func func, int error_code,
                            string_view message) FMT_NOEXCEPT {
   memory_buffer full_message;
@@ -191,10 +158,17 @@ FMT_FUNC void report_error(format_func func, int error_code,
   (void)std::fwrite(full_message.data(), full_message.size(), 1, stderr);
   std::fputc('\n', stderr);
 }
-}  // namespace internal
+
+// A wrapper around fwrite that throws on error.
+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"));
+}
+}  // namespace detail
 
 #if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
-namespace internal {
+namespace detail {
 
 template <typename Locale>
 locale_ref::locale_ref(const Locale& loc) : locale_(&loc) {
@@ -206,6 +180,9 @@ template <typename Locale> Locale locale_ref::get() const {
   return locale_ ? *static_cast<const std::locale*>(locale_) : std::locale();
 }
 
+template <typename Char> FMT_FUNC std::string grouping_impl(locale_ref loc) {
+  return std::use_facet<std::numpunct<Char>>(loc.get<std::locale>()).grouping();
+}
 template <typename Char> FMT_FUNC Char thousands_sep_impl(locale_ref loc) {
   return std::use_facet<std::numpunct<Char>>(loc.get<std::locale>())
       .thousands_sep();
@@ -214,20 +191,22 @@ template <typename Char> FMT_FUNC Char decimal_point_impl(locale_ref loc) {
   return std::use_facet<std::numpunct<Char>>(loc.get<std::locale>())
       .decimal_point();
 }
-}  // namespace internal
+}  // namespace detail
 #else
 template <typename Char>
-FMT_FUNC Char internal::thousands_sep_impl(locale_ref) {
+FMT_FUNC std::string detail::grouping_impl(locale_ref) {
+  return "\03";
+}
+template <typename Char> FMT_FUNC Char detail::thousands_sep_impl(locale_ref) {
   return FMT_STATIC_THOUSANDS_SEPARATOR;
 }
-template <typename Char>
-FMT_FUNC Char internal::decimal_point_impl(locale_ref) {
+template <typename Char> FMT_FUNC Char detail::decimal_point_impl(locale_ref) {
   return '.';
 }
 #endif
 
-FMT_API FMT_FUNC format_error::~format_error() FMT_NOEXCEPT {}
-FMT_API FMT_FUNC system_error::~system_error() FMT_NOEXCEPT {}
+FMT_API FMT_FUNC format_error::~format_error() FMT_NOEXCEPT = default;
+FMT_API FMT_FUNC system_error::~system_error() FMT_NOEXCEPT = default;
 
 FMT_FUNC void system_error::init(int err_code, string_view format_str,
                                  format_args args) {
@@ -238,10 +217,10 @@ FMT_FUNC void system_error::init(int err_code, string_view format_str,
   base = std::runtime_error(to_string(buffer));
 }
 
-namespace internal {
+namespace detail {
 
-template <> FMT_FUNC int count_digits<4>(internal::fallback_uintptr n) {
-  // Assume little endian; pointer formatting is implementation-defined anyway.
+template <> FMT_FUNC int count_digits<4>(detail::fallback_uintptr n) {
+  // fallback_uintptr is always stored in little endian.
   int i = static_cast<int>(sizeof(void*)) - 1;
   while (i > 0 && n.value[i] == 0) --i;
   auto char_digits = std::numeric_limits<unsigned char>::digits / 4;
@@ -249,53 +228,59 @@ template <> FMT_FUNC int count_digits<4>(internal::fallback_uintptr n) {
 }
 
 template <typename T>
-int format_float(char* buf, std::size_t size, const char* format, int precision,
-                 T value) {
-#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
-  if (precision > 100000)
-    throw std::runtime_error(
-        "fuzz mode - avoid large allocation inside snprintf");
-#endif
-  // Suppress the warning about nonliteral format string.
-  auto snprintf_ptr = FMT_SNPRINTF;
-  return precision < 0 ? snprintf_ptr(buf, size, format, value)
-                       : snprintf_ptr(buf, size, format, precision, value);
-}
-
-template <typename T>
-const char basic_data<T>::digits[] =
-    "0001020304050607080910111213141516171819"
-    "2021222324252627282930313233343536373839"
-    "4041424344454647484950515253545556575859"
-    "6061626364656667686970717273747576777879"
-    "8081828384858687888990919293949596979899";
+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'}};
 
 template <typename T>
 const char basic_data<T>::hex_digits[] = "0123456789abcdef";
 
 #define FMT_POWERS_OF_10(factor)                                             \
-  factor * 10, factor * 100, factor * 1000, factor * 10000, factor * 100000, \
-      factor * 1000000, factor * 10000000, factor * 100000000,               \
-      factor * 1000000000
+  factor * 10, (factor)*100, (factor)*1000, (factor)*10000, (factor)*100000, \
+      (factor)*1000000, (factor)*10000000, (factor)*100000000,               \
+      (factor)*1000000000
 
 template <typename T>
 const uint64_t basic_data<T>::powers_of_10_64[] = {
-    1, FMT_POWERS_OF_10(1), FMT_POWERS_OF_10(1000000000ull),
-    10000000000000000000ull};
+    1, 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[] = {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};
+    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,
@@ -330,7 +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,
@@ -340,35 +325,792 @@ 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>
 const char basic_data<T>::background_color[] = "\x1b[48;2;";
 template <typename T> const char basic_data<T>::reset_color[] = "\x1b[0m";
 template <typename T> const wchar_t basic_data<T>::wreset_color[] = L"\x1b[0m";
+template <typename T> const char basic_data<T>::signs[] = {0, '-', '+', ' '};
+template <typename T>
+const char basic_data<T>::left_padding_shifts[] = {31, 31, 0, 1, 0};
+template <typename T>
+const char basic_data<T>::right_padding_shifts[] = {0, 31, 0, 1, 0};
 
 template <typename T> struct bits {
   static FMT_CONSTEXPR_DECL const int value =
       static_cast<int>(sizeof(T) * std::numeric_limits<unsigned char>::digits);
 };
 
+class fp;
+template <int SHIFT = 0> fp normalize(fp value);
+
+// Lower (upper) boundary is a value half way between a floating-point value
+// and its predecessor (successor). Boundaries have the same exponent as the
+// value so only significands are stored.
+struct boundaries {
+  uint64_t lower;
+  uint64_t upper;
+};
+
 // A handmade floating-point number f * pow(2, e).
 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;
+
   // All sizes are in bits.
   // Subtract 1 to account for an implicit most significant bit in the
   // normalized form.
   static FMT_CONSTEXPR_DECL const int double_significand_size =
       std::numeric_limits<double>::digits - 1;
   static FMT_CONSTEXPR_DECL const uint64_t implicit_bit =
-      1ull << double_significand_size;
-
- public:
-  significand_type f;
-  int e;
-
+      1ULL << double_significand_size;
   static FMT_CONSTEXPR_DECL const int significand_size =
       bits<significand_type>::value;
 
@@ -377,96 +1119,360 @@ class fp {
 
   // Constructs fp from an IEEE754 double. It is a template to prevent compile
   // errors on platforms where double is not IEEE754.
-  template <typename Double> explicit fp(Double d) {
-    // Assume double is in the format [sign][exponent][significand].
-    using limits = std::numeric_limits<Double>;
+  template <typename Double> explicit fp(Double d) { assign(d); }
+
+  // Assigns d to this and return true iff predecessor is closer than successor.
+  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;
-    const uint64_t exponent_mask = (~0ull >> 1) & ~significand_mask;
+        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);
-    auto biased_e = (u & exponent_mask) >> double_significand_size;
+    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) >> 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 = static_cast<int>(biased_e - exponent_bias - double_significand_size);
+    e = biased_e - exponent_bias - float_significand_size;
+    return is_predecessor_closer;
   }
 
-  // Normalizes the value converted from double and multiplied by (1 << SHIFT).
-  template <int SHIFT = 0> void normalize() {
-    // Handle subnormals.
-    auto shifted_implicit_bit = implicit_bit << SHIFT;
-    while ((f & shifted_implicit_bit) == 0) {
-      f <<= 1;
-      --e;
-    }
-    // Subtract 1 to account for hidden bit.
-    auto offset = significand_size - double_significand_size - SHIFT - 1;
-    f <<= offset;
-    e -= offset;
-  }
-
-  // Compute lower and upper boundaries (m^- and m^+ in the Grisu paper), 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.
-  void compute_boundaries(fp& lower, fp& upper) const {
-    lower =
-        f == implicit_bit ? fp((f << 2) - 1, e - 2) : fp((f << 1) - 1, e - 1);
-    upper = fp((f << 1) + 1, e - 1);
-    upper.normalize<1>();  // 1 is to account for the exponent shift above.
-    lower.f <<= lower.e - upper.e;
-    lower.e = upper.e;
+  template <typename Float, FMT_ENABLE_IF(!is_supported_float<Float>::value)>
+  bool assign(Float) {
+    *this = fp();
+    return false;
   }
 };
 
-// Returns an fp number representing x - y. Result may not be normalized.
-inline fp operator-(fp x, fp y) {
-  FMT_ASSERT(x.f >= y.f && x.e == y.e, "invalid operands");
-  return fp(x.f - y.f, x.e);
+// Normalizes the value converted from double and multiplied by (1 << SHIFT).
+template <int SHIFT> fp normalize(fp value) {
+  // Handle subnormals.
+  const auto shifted_implicit_bit = fp::implicit_bit << SHIFT;
+  while ((value.f & shifted_implicit_bit) == 0) {
+    value.f <<= 1;
+    --value.e;
+  }
+  // Subtract 1 to account for hidden bit.
+  const auto offset =
+      fp::significand_size - fp::double_significand_size - SHIFT - 1;
+  value.f <<= offset;
+  value.e -= offset;
+  return value;
 }
 
-// Computes an fp number r with r.f = x.f * y.f / pow(2, 64) rounded to nearest
-// with half-up tie breaking, r.e = x.e + y.e + 64. Result may not be
-// normalized.
-FMT_FUNC fp operator*(fp x, fp y) {
-  int exp = x.e + y.e + 64;
+inline bool operator==(fp x, fp y) { return x.f == y.f && x.e == y.e; }
+
+// Computes lhs * rhs / pow(2, 64) rounded to nearest with half-up tie breaking.
+inline uint64_t multiply(uint64_t lhs, uint64_t rhs) {
 #if FMT_USE_INT128
-  auto product = static_cast<__uint128_t>(x.f) * y.f;
+  auto product = static_cast<__uint128_t>(lhs) * rhs;
   auto f = static_cast<uint64_t>(product >> 64);
-  if ((static_cast<uint64_t>(product) & (1ULL << 63)) != 0) ++f;
-  return fp(f, exp);
+  return (static_cast<uint64_t>(product) & (1ULL << 63)) != 0 ? f + 1 : f;
 #else
   // Multiply 32-bit parts of significands.
   uint64_t mask = (1ULL << 32) - 1;
-  uint64_t a = x.f >> 32, b = x.f & mask;
-  uint64_t c = y.f >> 32, d = y.f & mask;
+  uint64_t a = lhs >> 32, b = lhs & mask;
+  uint64_t c = rhs >> 32, d = rhs & mask;
   uint64_t ac = a * c, bc = b * c, ad = a * d, bd = b * d;
   // Compute mid 64-bit of result and round.
   uint64_t mid = (bd >> 32) + (ad & mask) + (bc & mask) + (1U << 31);
-  return fp(ac + (ad >> 32) + (bc >> 32) + (mid >> 32), exp);
+  return ac + (ad >> 32) + (bc >> 32) + (mid >> 32);
 #endif
 }
 
-// Returns 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.
-FMT_FUNC fp get_cached_power(int min_exponent, int& pow10_exponent) {
-  const double one_over_log2_10 = 0.30102999566398114;  // 1 / log2(10)
+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 int shift = 32;
+  const auto significand = static_cast<int64_t>(data::log10_2_significand);
   int index = static_cast<int>(
-      std::ceil((min_exponent + fp::significand_size - 1) * one_over_log2_10));
+      ((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;
   // Difference between 2 consecutive decimal exponents in cached powers of 10.
   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 fp(data::pow10_significands[index], data::pow10_exponents[index]);
+  return {data::grisu_pow10_significands[index],
+          data::grisu_pow10_exponents[index]};
 }
 
-enum round_direction { unknown, up, down };
+// A simple accumulator to hold the sums of terms in bigint::square if uint128_t
+// is not available.
+struct accumulator {
+  uint64_t lower;
+  uint64_t upper;
+
+  accumulator() : lower(0), upper(0) {}
+  explicit operator uint32_t() const { return static_cast<uint32_t>(lower); }
+
+  void operator+=(uint64_t n) {
+    lower += n;
+    if (lower < n) ++upper;
+  }
+  void operator>>=(int shift) {
+    assert(shift == 32);
+    (void)shift;
+    lower = (upper << 32) | (lower >> 32);
+    upper >>= 32;
+  }
+};
+
+class bigint {
+ private:
+  // A bigint is stored as an array of bigits (big digits), with bigit at index
+  // 0 being the least significant one.
+  using bigit = uint32_t;
+  using double_bigit = uint64_t;
+  enum { bigits_capacity = 32 };
+  basic_memory_buffer<bigit, bigits_capacity> bigits_;
+  int exp_;
+
+  bigit operator[](int index) const { return bigits_[to_unsigned(index)]; }
+  bigit& operator[](int index) { return bigits_[to_unsigned(index)]; }
+
+  static FMT_CONSTEXPR_DECL const int bigit_bits = bits<bigit>::value;
+
+  friend struct formatter<bigint>;
+
+  void subtract_bigits(int index, bigit other, bigit& borrow) {
+    auto result = static_cast<double_bigit>((*this)[index]) - other - borrow;
+    (*this)[index] = static_cast<bigit>(result);
+    borrow = static_cast<bigit>(result >> (bigit_bits * 2 - 1));
+  }
+
+  void remove_leading_zeros() {
+    int num_bigits = static_cast<int>(bigits_.size()) - 1;
+    while (num_bigits > 0 && (*this)[num_bigits] == 0) --num_bigits;
+    bigits_.resize(to_unsigned(num_bigits + 1));
+  }
+
+  // Computes *this -= other assuming aligned bigints and *this >= other.
+  void subtract_aligned(const bigint& other) {
+    FMT_ASSERT(other.exp_ >= exp_, "unaligned bigints");
+    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)
+      subtract_bigits(i, other.bigits_[j], borrow);
+    while (borrow > 0) subtract_bigits(i, 0, borrow);
+    remove_leading_zeros();
+  }
+
+  void multiply(uint32_t value) {
+    const double_bigit wide_value = value;
+    bigit carry = 0;
+    for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
+      double_bigit result = bigits_[i] * wide_value + carry;
+      bigits_[i] = static_cast<bigit>(result);
+      carry = static_cast<bigit>(result >> bigit_bits);
+    }
+    if (carry != 0) bigits_.push_back(carry);
+  }
+
+  void multiply(uint64_t value) {
+    const bigit mask = ~bigit(0);
+    const double_bigit lower = value & mask;
+    const double_bigit upper = value >> bigit_bits;
+    double_bigit carry = 0;
+    for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
+      double_bigit result = bigits_[i] * lower + (carry & mask);
+      carry =
+          bigits_[i] * upper + (result >> bigit_bits) + (carry >> bigit_bits);
+      bigits_[i] = static_cast<bigit>(result);
+    }
+    while (carry != 0) {
+      bigits_.push_back(carry & mask);
+      carry >>= bigit_bits;
+    }
+  }
+
+ public:
+  bigint() : exp_(0) {}
+  explicit bigint(uint64_t n) { assign(n); }
+  ~bigint() { assert(bigits_.capacity() <= bigits_capacity); }
+
+  bigint(const bigint&) = delete;
+  void operator=(const bigint&) = delete;
+
+  void assign(const bigint& other) {
+    auto size = other.bigits_.size();
+    bigits_.resize(size);
+    auto data = other.bigits_.data();
+    std::copy(data, data + size, make_checked(bigits_.data(), size));
+    exp_ = other.exp_;
+  }
+
+  void assign(uint64_t n) {
+    size_t num_bigits = 0;
+    do {
+      bigits_[num_bigits++] = n & ~bigit(0);
+      n >>= bigit_bits;
+    } while (n != 0);
+    bigits_.resize(num_bigits);
+    exp_ = 0;
+  }
+
+  int num_bigits() const { return static_cast<int>(bigits_.size()) + exp_; }
+
+  FMT_NOINLINE bigint& operator<<=(int shift) {
+    assert(shift >= 0);
+    exp_ += shift / bigit_bits;
+    shift %= bigit_bits;
+    if (shift == 0) return *this;
+    bigit carry = 0;
+    for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
+      bigit c = bigits_[i] >> (bigit_bits - shift);
+      bigits_[i] = (bigits_[i] << shift) + carry;
+      carry = c;
+    }
+    if (carry != 0) bigits_.push_back(carry);
+    return *this;
+  }
+
+  template <typename Int> bigint& operator*=(Int value) {
+    FMT_ASSERT(value > 0, "");
+    multiply(uint32_or_64_or_128_t<Int>(value));
+    return *this;
+  }
+
+  friend int compare(const bigint& lhs, const bigint& rhs) {
+    int num_lhs_bigits = lhs.num_bigits(), num_rhs_bigits = rhs.num_bigits();
+    if (num_lhs_bigits != num_rhs_bigits)
+      return num_lhs_bigits > num_rhs_bigits ? 1 : -1;
+    int i = static_cast<int>(lhs.bigits_.size()) - 1;
+    int j = static_cast<int>(rhs.bigits_.size()) - 1;
+    int end = i - j;
+    if (end < 0) end = 0;
+    for (; i >= end; --i, --j) {
+      bigit lhs_bigit = lhs[i], rhs_bigit = rhs[j];
+      if (lhs_bigit != rhs_bigit) return lhs_bigit > rhs_bigit ? 1 : -1;
+    }
+    if (i != j) return i > j ? 1 : -1;
+    return 0;
+  }
+
+  // Returns compare(lhs1 + lhs2, rhs).
+  friend int add_compare(const bigint& lhs1, const bigint& lhs2,
+                         const bigint& rhs) {
+    int max_lhs_bigits = (std::max)(lhs1.num_bigits(), lhs2.num_bigits());
+    int num_rhs_bigits = rhs.num_bigits();
+    if (max_lhs_bigits + 1 < num_rhs_bigits) return -1;
+    if (max_lhs_bigits > num_rhs_bigits) return 1;
+    auto get_bigit = [](const bigint& n, int i) -> bigit {
+      return i >= n.exp_ && i < n.num_bigits() ? n[i - n.exp_] : 0;
+    };
+    double_bigit borrow = 0;
+    int min_exp = (std::min)((std::min)(lhs1.exp_, lhs2.exp_), rhs.exp_);
+    for (int i = num_rhs_bigits - 1; i >= min_exp; --i) {
+      double_bigit sum =
+          static_cast<double_bigit>(get_bigit(lhs1, i)) + get_bigit(lhs2, i);
+      bigit rhs_bigit = get_bigit(rhs, i);
+      if (sum > rhs_bigit + borrow) return 1;
+      borrow = rhs_bigit + borrow - sum;
+      if (borrow > 1) return -1;
+      borrow <<= bigit_bits;
+    }
+    return borrow != 0 ? -1 : 0;
+  }
+
+  // Assigns pow(10, exp) to this bigint.
+  void assign_pow10(int exp) {
+    assert(exp >= 0);
+    if (exp == 0) return assign(1);
+    // Find the top bit.
+    int bitmask = 1;
+    while (exp >= bitmask) bitmask <<= 1;
+    bitmask >>= 1;
+    // pow(10, exp) = pow(5, exp) * pow(2, exp). First compute pow(5, exp) by
+    // repeated squaring and multiplication.
+    assign(5);
+    bitmask >>= 1;
+    while (bitmask != 0) {
+      square();
+      if ((exp & bitmask) != 0) *this *= 5;
+      bitmask >>= 1;
+    }
+    *this <<= exp;  // Multiply by pow(2, exp) by shifting.
+  }
+
+  void square() {
+    basic_memory_buffer<bigit, bigits_capacity> n(std::move(bigits_));
+    int num_bigits = static_cast<int>(bigits_.size());
+    int num_result_bigits = 2 * num_bigits;
+    bigits_.resize(to_unsigned(num_result_bigits));
+    using accumulator_t = conditional_t<FMT_USE_INT128, uint128_t, accumulator>;
+    auto sum = accumulator_t();
+    for (int bigit_index = 0; bigit_index < num_bigits; ++bigit_index) {
+      // Compute bigit at position bigit_index of the result by adding
+      // cross-product terms n[i] * n[j] such that i + j == bigit_index.
+      for (int i = 0, j = bigit_index; j >= 0; ++i, --j) {
+        // Most terms are multiplied twice which can be optimized in the future.
+        sum += static_cast<double_bigit>(n[i]) * n[j];
+      }
+      (*this)[bigit_index] = static_cast<bigit>(sum);
+      sum >>= bits<bigit>::value;  // Compute the carry.
+    }
+    // Do the same for the top half.
+    for (int bigit_index = num_bigits; bigit_index < num_result_bigits;
+         ++bigit_index) {
+      for (int j = num_bigits - 1, i = bigit_index - j; i < num_bigits;)
+        sum += static_cast<double_bigit>(n[i++]) * n[j--];
+      (*this)[bigit_index] = static_cast<bigit>(sum);
+      sum >>= bits<bigit>::value;
+    }
+    --num_result_bigits;
+    remove_leading_zeros();
+    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;
+    FMT_ASSERT(divisor.bigits_[divisor.bigits_.size() - 1u] != 0, "");
+    align(divisor);
+    int quotient = 0;
+    do {
+      subtract_aligned(divisor);
+      ++quotient;
+    } while (compare(*this, divisor) >= 0);
+    return quotient;
+  }
+};
+
+enum class round_direction { unknown, up, down };
 
 // Given the divisor (normally a power of 10), the remainder = v % divisor for
 // some number v and the error, returns whether v should be rounded up, down, or
@@ -479,13 +1485,13 @@ inline round_direction get_round_direction(uint64_t divisor, uint64_t remainder,
   FMT_ASSERT(error < divisor - error, "");  // error * 2 won't overflow.
   // Round down if (remainder + error) * 2 <= divisor.
   if (remainder <= divisor - remainder && error * 2 <= divisor - remainder * 2)
-    return down;
+    return round_direction::down;
   // Round up if (remainder - error) * 2 >= divisor.
   if (remainder >= error &&
       remainder - error >= divisor - (remainder - error)) {
-    return up;
+    return round_direction::up;
   }
-  return unknown;
+  return round_direction::unknown;
 }
 
 namespace digits {
@@ -500,13 +1506,13 @@ enum result {
 // error: the size of the region (lower, upper) outside of which numbers
 // definitely do not round to value (Delta in Grisu3).
 template <typename Handler>
-digits::result grisu_gen_digits(fp value, uint64_t error, int& exp,
-                                Handler& handler) {
-  fp one(1ull << -value.e, value.e);
+FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
+                                                  int& exp, Handler& handler) {
+  const fp one(1ULL << -value.e, value.e);
   // The integral part of scaled value (p1 in Grisu) = value / one. It cannot be
   // zero because it contains a product of two 64-bit numbers with MSB set (due
   // to normalization) - 1, shifted right by at most 60 bits.
-  uint32_t integral = static_cast<uint32_t>(value.f >> -one.e);
+  auto integral = static_cast<uint32_t>(value.f >> -one.e);
   FMT_ASSERT(integral != 0, "");
   FMT_ASSERT(integral == value.f >> -one.e, "");
   // The fractional part of scaled value (p2 in Grisu) c = value % one.
@@ -519,44 +1525,39 @@ digits::result grisu_gen_digits(fp value, uint64_t error, int& exp,
   // Generate digits for the integral part. This can produce up to 10 digits.
   do {
     uint32_t digit = 0;
-    // This optimization by miloyip reduces the number of integer divisions by
+    auto divmod_integral = [&](uint32_t divisor) {
+      digit = integral / divisor;
+      integral %= divisor;
+    };
+    // This optimization by Milo Yip reduces the number of integer divisions by
     // one per iteration.
     switch (exp) {
     case 10:
-      digit = integral / 1000000000;
-      integral %= 1000000000;
+      divmod_integral(1000000000);
       break;
     case 9:
-      digit = integral / 100000000;
-      integral %= 100000000;
+      divmod_integral(100000000);
       break;
     case 8:
-      digit = integral / 10000000;
-      integral %= 10000000;
+      divmod_integral(10000000);
       break;
     case 7:
-      digit = integral / 1000000;
-      integral %= 1000000;
+      divmod_integral(1000000);
       break;
     case 6:
-      digit = integral / 100000;
-      integral %= 100000;
+      divmod_integral(100000);
       break;
     case 5:
-      digit = integral / 10000;
-      integral %= 10000;
+      divmod_integral(10000);
       break;
     case 4:
-      digit = integral / 1000;
-      integral %= 1000;
+      divmod_integral(1000);
       break;
     case 3:
-      digit = integral / 100;
-      integral %= 100;
+      divmod_integral(100);
       break;
     case 2:
-      digit = integral / 10;
-      integral %= 10;
+      divmod_integral(10);
       break;
     case 1:
       digit = integral;
@@ -566,8 +1567,7 @@ digits::result grisu_gen_digits(fp value, uint64_t error, int& exp,
       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);
@@ -577,8 +1577,7 @@ digits::result grisu_gen_digits(fp value, uint64_t error, int& exp,
   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);
@@ -606,8 +1605,8 @@ struct fixed_handler {
     if (precision > 0) return digits::more;
     if (precision < 0) return digits::done;
     auto dir = get_round_direction(divisor, remainder, error);
-    if (dir == unknown) return digits::error;
-    buf[size++] = dir == up ? '1' : '0';
+    if (dir == round_direction::unknown) return digits::error;
+    buf[size++] = dir == round_direction::up ? '1' : '0';
     return digits::done;
   }
 
@@ -615,6 +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.
@@ -625,7 +1625,8 @@ struct fixed_handler {
       FMT_ASSERT(error == 1 && divisor > 2, "");
     }
     auto dir = get_round_direction(divisor, remainder, error);
-    if (dir != up) return dir == down ? digits::done : digits::error;
+    if (dir != round_direction::up)
+      return dir == round_direction::down ? digits::done : digits::error;
     ++buf[size - 1];
     for (int i = size - 1; i > 0 && buf[i] > '9'; --i) {
       buf[i] = '0';
@@ -633,300 +1634,1083 @@ 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.
-template <int GRISU_VERSION> 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);
 
-  digits::result on_start(uint64_t, uint64_t, uint64_t, int&) {
-    return digits::more;
+  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];
   }
 
-  // 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_mul(carrier_uint u,
+                                  const cache_entry_type& cache) FMT_NOEXCEPT {
+    return umul96_upper32(u, cache);
   }
 
-  // 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;
-    if (GRISU_VERSION != 3) {
-      uint64_t d = integral ? diff : diff * data::powers_of_10_64[-exp];
-      round(d, divisor, remainder, error);
-      return digits::done;
-    }
-    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;
-    }
-    return 2 * unit <= remainder && remainder <= error - 4 * unit
-               ? digits::done
-               : digits::error;
+  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 <typename Double,
-          enable_if_t<(sizeof(Double) == sizeof(uint64_t)), int>>
-FMT_API bool grisu_format(Double value, buffer<char>& buf, int precision,
-                          unsigned options, int& exp) {
-  FMT_ASSERT(value >= 0, "value is negative");
-  bool fixed = (options & grisu_options::fixed) != 0;
-  if (value <= 0) {  // <= instead of == to silence a warning.
-    if (precision <= 0 || !fixed) {
-      exp = 0;
-      buf.push_back('0');
-    } else {
-      exp = -precision;
-      buf.resize(precision);
-      std::uninitialized_fill_n(buf.data(), precision, '0');
-    }
-    return true;
+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
   }
 
-  fp fp_value(value);
-  const int min_exp = -60;  // alpha in Grisu.
-  int cached_exp10 = 0;     // K in Grisu.
-  if (precision != -1) {
-    if (precision > 17) return false;
-    fp_value.normalize();
-    auto cached_pow = get_cached_power(
-        min_exp - (fp_value.e + fp::significand_size), cached_exp10);
-    fp_value = fp_value * cached_pow;
-    fixed_handler handler{buf.data(), 0, precision, -cached_exp10, fixed};
-    if (grisu_gen_digits(fp_value, 1, exp, handler) == digits::error)
-      return false;
-    buf.resize(to_unsigned(handler.size));
-  } else {
-    fp lower, upper;  // w^- and w^+ in the Grisu paper.
-    fp_value.compute_boundaries(lower, upper);
-    // Find a cached power of 10 such that multiplying upper by it will bring
-    // the exponent in the range [min_exp, -32].
-    auto cached_pow = get_cached_power(  // \tilde{c}_{-k} in Grisu.
-        min_exp - (upper.e + fp::significand_size), cached_exp10);
-    fp_value.normalize();
-    fp_value = fp_value * cached_pow;
-    lower = lower * cached_pow;  // \tilde{M}^- in Grisu.
-    upper = upper * cached_pow;  // \tilde{M}^+ in Grisu.
-    assert(min_exp <= upper.e && upper.e <= -32);
-    auto result = digits::result();
-    int size = 0;
-    if ((options & grisu_options::grisu3) != 0) {
-      --lower.f;  // \tilde{M}^- - 1 ulp -> M^-_{\downarrow}.
-      ++upper.f;  // \tilde{M}^+ + 1 ulp -> M^+_{\uparrow}.
-      // Numbers outside of (lower, upper) definitely do not round to value.
-      grisu_shortest_handler<3> handler{buf.data(), 0, (upper - fp_value).f};
-      result = grisu_gen_digits(upper, upper.f - lower.f, exp, handler);
-      size = handler.size;
-    } else {
-      ++lower.f;  // \tilde{M}^- + 1 ulp -> M^-_{\uparrow}.
-      --upper.f;  // \tilde{M}^+ - 1 ulp -> M^+_{\downarrow}.
-      grisu_shortest_handler<2> handler{buf.data(), 0, (upper - fp_value).f};
-      result = grisu_gen_digits(upper, upper.f - lower.f, exp, handler);
-      size = handler.size;
-    }
-    if (result == digits::error) return false;
-    buf.resize(to_unsigned(size));
+  static carrier_uint compute_mul(carrier_uint u,
+                                  const cache_entry_type& cache) FMT_NOEXCEPT {
+    return umul192_upper64(u, cache);
   }
-  exp -= cached_exp10;
-  return true;
+
+  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, "");
+
+    return ((umul192_middle64(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 (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 <typename Double>
-char* sprintf_format(Double value, internal::buffer<char>& buf,
-                     sprintf_specs specs) {
-  // Buffer capacity must be non-zero, otherwise MSVC's vsnprintf_s will fail.
-  FMT_ASSERT(buf.capacity() != 0, "empty buffer");
+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);
+}
 
-  // Build format string.
-  enum { max_format_size = 10 };  // longest format: %#-*.*Lg
+// 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;
+    }
+  }
+
+  // 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;
+    }
+  }
+  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, 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.
+  bigint lower;             // (M^- in (FPP)^2).
+  bigint upper_store;       // upper's value if different from lower.
+  bigint* upper = nullptr;  // (M^+ in (FPP)^2).
+  fp value;
+  // 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.
+  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);
+    numerator <<= value.e;
+    lower.assign(1);
+    lower <<= value.e;
+    if (shift != 1) {
+      upper_store.assign(1);
+      upper_store <<= value.e + 1;
+      upper = &upper_store;
+    }
+    denominator.assign_pow10(exp10);
+    denominator <<= shift;
+  } else if (exp10 < 0) {
+    numerator.assign_pow10(-exp10);
+    lower.assign(numerator);
+    if (shift != 1) {
+      upper_store.assign(numerator);
+      upper_store <<= 1;
+      upper = &upper_store;
+    }
+    numerator *= significand;
+    denominator.assign(1);
+    denominator <<= shift - value.e;
+  } else {
+    numerator.assign(significand);
+    denominator.assign_pow10(exp10);
+    denominator <<= shift - value.e;
+    lower.assign(1);
+    if (shift != 1) {
+      upper_store.assign(1ULL << 1);
+      upper = &upper_store;
+    }
+  }
+  // Invariant: value == (numerator / denominator) * pow(10, exp10).
+  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;
+      }
+      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;
+  }
+  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);
+}
+
+template <typename T>
+int format_float(T value, int precision, float_specs specs, buffer<char>& buf) {
+  static_assert(!std::is_same<T, float>::value, "");
+  FMT_ASSERT(value >= 0, "value is negative");
+
+  const bool fixed = specs.format == float_format::fixed;
+  if (value <= 0) {  // <= instead of == to silence a warning.
+    if (precision <= 0 || !fixed) {
+      buf.push_back('0');
+      return 0;
+    }
+    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.
+  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 {
+    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.try_resize(num_digits);
+  }
+  return exp;
+}  // namespace detail
+
+template <typename T>
+int snprintf_float(T value, int precision, float_specs specs,
+                   buffer<char>& buf) {
+  // Buffer capacity must be non-zero, otherwise MSVC's vsnprintf_s will fail.
+  FMT_ASSERT(buf.capacity() > buf.size(), "empty buffer");
+  static_assert(!std::is_same<T, float>::value, "");
+
+  // Subtract 1 to account for the difference in precision since we use %e for
+  // both general and exponent format.
+  if (specs.format == float_format::general ||
+      specs.format == float_format::exp)
+    precision = (precision >= 0 ? precision : 6) - 1;
+
+  // Build the format string.
+  enum { max_format_size = 7 };  // The longest format is "%#.*Le".
   char format[max_format_size];
   char* format_ptr = format;
   *format_ptr++ = '%';
-  if (specs.alt || !specs.type) *format_ptr++ = '#';
-  if (specs.precision >= 0) {
+  if (specs.showpoint && specs.format == float_format::hex) *format_ptr++ = '#';
+  if (precision >= 0) {
     *format_ptr++ = '.';
     *format_ptr++ = '*';
   }
-  if (std::is_same<Double, long double>::value) *format_ptr++ = 'L';
-
-  char type = specs.type;
-
-  if (type == '%')
-    type = 'f';
-  else if (type == 0 || type == 'n')
-    type = 'g';
-#if FMT_MSC_VER
-  if (type == 'F') {
-    // MSVC's printf doesn't support 'F'.
-    type = 'f';
-  }
-#endif
-  *format_ptr++ = type;
+  if (std::is_same<T, long double>()) *format_ptr++ = 'L';
+  *format_ptr++ = specs.format != float_format::hex
+                      ? (specs.format == float_format::fixed ? 'f' : 'e')
+                      : (specs.upper ? 'A' : 'a');
   *format_ptr = '\0';
 
   // Format using snprintf.
-  char* start = nullptr;
-  char* decimal_point_pos = nullptr;
+  auto offset = buf.size();
   for (;;) {
-    std::size_t buffer_size = buf.capacity();
-    start = &buf[0];
-    int result =
-        format_float(start, buffer_size, format, specs.precision, value);
-    if (result >= 0) {
-      unsigned n = internal::to_unsigned(result);
-      if (n < buf.capacity()) {
-        // Find the decimal point.
-        auto p = buf.data(), end = p + n;
-        if (*p == '+' || *p == '-') ++p;
-        if (specs.type != 'a' && specs.type != 'A') {
-          while (p < end && *p >= '0' && *p <= '9') ++p;
-          if (p < end && *p != 'e' && *p != 'E') {
-            decimal_point_pos = p;
-            if (!specs.type) {
-              // Keep only one trailing zero after the decimal point.
-              ++p;
-              if (*p == '0') ++p;
-              while (p != end && *p >= '1' && *p <= '9') ++p;
-              char* where = p;
-              while (p != end && *p == '0') ++p;
-              if (p == end || *p < '0' || *p > '9') {
-                if (p != end) std::memmove(where, p, to_unsigned(end - p));
-                n -= static_cast<unsigned>(p - where);
-              }
-            }
-          }
-        }
-        buf.resize(n);
-        break;  // The buffer is large enough - continue with formatting.
+    auto begin = buf.data() + offset;
+    auto capacity = buf.capacity() - offset;
+#ifdef FMT_FUZZ
+    if (precision > 100000)
+      throw std::runtime_error(
+          "fuzz mode - avoid large allocation inside snprintf");
+#endif
+    // Suppress the warning about a nonliteral format string.
+    // Cannot use auto because of a bug in MinGW (#1532).
+    int (*snprintf_ptr)(char*, size_t, const char*, ...) = FMT_SNPRINTF;
+    int result = precision >= 0
+                     ? snprintf_ptr(begin, capacity, format, precision, value)
+                     : snprintf_ptr(begin, capacity, format, value);
+    if (result < 0) {
+      // 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.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.try_resize(size);
+        return 0;
       }
-      buf.reserve(n + 1);
+      // Find and remove the decimal point.
+      auto end = begin + size, p = end;
+      do {
+        --p;
+      } while (is_digit(*p));
+      int fraction_size = static_cast<int>(end - p - 1);
+      std::memmove(p, p + 1, to_unsigned(fraction_size));
+      buf.try_resize(size - 1);
+      return -fraction_size;
+    }
+    if (specs.format == float_format::hex) {
+      buf.try_resize(size + offset);
+      return 0;
+    }
+    // Find and parse the exponent.
+    auto end = begin + size, exp_pos = end;
+    do {
+      --exp_pos;
+    } while (*exp_pos != 'e');
+    char sign = exp_pos[1];
+    assert(sign == '+' || sign == '-');
+    int exp = 0;
+    auto p = exp_pos + 2;  // Skip 'e' and sign.
+    do {
+      assert(is_digit(*p));
+      exp = exp * 10 + (*p++ - '0');
+    } while (p != end);
+    if (sign == '-') exp = -exp;
+    int fraction_size = 0;
+    if (exp_pos != begin + 1) {
+      // Remove trailing zeros.
+      auto fraction_end = exp_pos - 1;
+      while (*fraction_end == '0') --fraction_end;
+      // Move the fractional part left to get rid of the decimal point.
+      fraction_size = static_cast<int>(fraction_end - begin - 1);
+      std::memmove(begin + 1, begin + 2, to_unsigned(fraction_size));
+    }
+    buf.try_resize(to_unsigned(fraction_size) + offset + 1);
+    return exp - fraction_size;
+  }
+}
+
+// A public domain branchless UTF-8 decoder by Christopher Wellons:
+// https://github.com/skeeto/branchless-utf8
+/* Decode the next character, c, from buf, reporting errors in e.
+ *
+ * Since this is a branchless decoder, four bytes will be read from the
+ * buffer regardless of the actual length of the next character. This
+ * means the buffer _must_ have at least three bytes of zero padding
+ * following the end of the data stream.
+ *
+ * Errors are reported in e, which will be non-zero if the parsed
+ * character was somehow invalid: invalid byte sequence, non-canonical
+ * encoding, or a surrogate half.
+ *
+ * The function returns a pointer to the next character. When an error
+ * occurs, this pointer will be a guess that depends on the particular
+ * error, but it will always advance at least one byte.
+ */
+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};
+
+  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;
+  *c |= uint32_t(s[3] & 0x3f) << 0;
+  *c >>= shiftc[len];
+
+  // Accumulate the various error conditions.
+  *e = (*c < mins[len]) << 6;       // non-canonical encoding
+  *e |= ((*c >> 11) == 0x1b) << 7;  // surrogate half?
+  *e |= (*c > 0x10FFFF) << 8;       // out of range?
+  *e |= (s[1] & 0xc0) >> 2;
+  *e |= (s[2] & 0xc0) >> 4;
+  *e |= (s[3]) >> 6;
+  *e ^= 0x2a;  // top two bits of each tail byte correct?
+  *e >>= shifte[len];
+
+  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> {
+  format_parse_context::iterator parse(format_parse_context& ctx) {
+    return ctx.begin();
+  }
+
+  format_context::iterator format(const detail::bigint& n,
+                                  format_context& ctx) {
+    auto out = ctx.out();
+    bool first = true;
+    for (auto i = n.bigits_.size(); i > 0; --i) {
+      auto value = n.bigits_[i - 1u];
+      if (first) {
+        out = format_to(out, "{:x}", value);
+        first = false;
+        continue;
+      }
+      out = format_to(out, "{:08x}", value);
+    }
+    if (n.exp_ > 0)
+      out = format_to(out, "p{}", n.exp_ * detail::bigint::bigit_bits);
+    return out;
+  }
+};
+
+FMT_FUNC detail::utf8_to_utf16::utf8_to_utf16(string_view s) {
+  auto transcode = [this](const char* p) {
+    auto cp = uint32_t();
+    auto error = 0;
+    p = utf8_decode(p, &cp, &error);
+    if (error != 0) FMT_THROW(std::runtime_error("invalid utf8"));
+    if (cp <= 0xFFFF) {
+      buffer_.push_back(static_cast<wchar_t>(cp));
     } else {
-      // If result is negative we ask to increase the capacity by at least 1,
-      // but as std::vector, the buffer grows exponentially.
-      buf.reserve(buf.capacity() + 1);
+      cp -= 0x10000;
+      buffer_.push_back(static_cast<wchar_t>(0xD800 + (cp >> 10)));
+      buffer_.push_back(static_cast<wchar_t>(0xDC00 + (cp & 0x3FF)));
     }
+    return p;
+  };
+  auto p = s.data();
+  const size_t block_size = 4;  // utf8_decode always reads blocks of 4 chars.
+  if (s.size() >= block_size) {
+    for (auto end = p + s.size() - block_size + 1; p < end;) p = transcode(p);
   }
-  return decimal_point_pos;
-}
-}  // namespace internal
-
-#if FMT_USE_WINDOWS_H
-
-FMT_FUNC internal::utf8_to_utf16::utf8_to_utf16(string_view s) {
-  static const char ERROR_MSG[] = "cannot convert string from UTF-8 to UTF-16";
-  if (s.size() > INT_MAX)
-    FMT_THROW(windows_error(ERROR_INVALID_PARAMETER, ERROR_MSG));
-  int s_size = static_cast<int>(s.size());
-  if (s_size == 0) {
-    // MultiByteToWideChar does not support zero length, handle separately.
-    buffer_.resize(1);
-    buffer_[0] = 0;
-    return;
+  if (auto num_chars_left = s.data() + s.size() - p) {
+    char buf[2 * block_size - 1] = {};
+    memcpy(buf, p, to_unsigned(num_chars_left));
+    p = buf;
+    do {
+      p = transcode(p);
+    } while (p - buf < num_chars_left);
   }
-
-  int length = MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, s.data(),
-                                   s_size, nullptr, 0);
-  if (length == 0) FMT_THROW(windows_error(GetLastError(), ERROR_MSG));
-  buffer_.resize(length + 1);
-  length = MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, s.data(), s_size,
-                               &buffer_[0], length);
-  if (length == 0) FMT_THROW(windows_error(GetLastError(), ERROR_MSG));
-  buffer_[length] = 0;
+  buffer_.push_back(0);
 }
 
-FMT_FUNC internal::utf16_to_utf8::utf16_to_utf8(wstring_view s) {
-  if (int error_code = convert(s)) {
-    FMT_THROW(windows_error(error_code,
-                            "cannot convert string from UTF-16 to UTF-8"));
-  }
-}
-
-FMT_FUNC int internal::utf16_to_utf8::convert(wstring_view s) {
-  if (s.size() > INT_MAX) return ERROR_INVALID_PARAMETER;
-  int s_size = static_cast<int>(s.size());
-  if (s_size == 0) {
-    // WideCharToMultiByte does not support zero length, handle separately.
-    buffer_.resize(1);
-    buffer_[0] = 0;
-    return 0;
-  }
-
-  int length = WideCharToMultiByte(CP_UTF8, 0, s.data(), s_size, nullptr, 0,
-                                   nullptr, nullptr);
-  if (length == 0) return GetLastError();
-  buffer_.resize(length + 1);
-  length = WideCharToMultiByte(CP_UTF8, 0, s.data(), s_size, &buffer_[0],
-                               length, nullptr, nullptr);
-  if (length == 0) return GetLastError();
-  buffer_[length] = 0;
-  return 0;
-}
-
-FMT_FUNC void windows_error::init(int err_code, string_view format_str,
-                                  format_args args) {
-  error_code_ = err_code;
-  memory_buffer buffer;
-  internal::format_windows_error(buffer, err_code, vformat(format_str, args));
-  std::runtime_error& base = *this;
-  base = std::runtime_error(to_string(buffer));
-}
-
-FMT_FUNC void internal::format_windows_error(internal::buffer<char>& out,
-                                             int error_code,
-                                             string_view message) FMT_NOEXCEPT {
-  FMT_TRY {
-    wmemory_buffer buf;
-    buf.resize(inline_buffer_size);
-    for (;;) {
-      wchar_t* system_message = &buf[0];
-      int result = FormatMessageW(
-          FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS, nullptr,
-          error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), system_message,
-          static_cast<uint32_t>(buf.size()), nullptr);
-      if (result != 0) {
-        utf16_to_utf8 utf8_message;
-        if (utf8_message.convert(system_message) == ERROR_SUCCESS) {
-          internal::writer w(out);
-          w.write(message);
-          w.write(": ");
-          w.write(utf8_message);
-          return;
-        }
-        break;
-      }
-      if (GetLastError() != ERROR_INSUFFICIENT_BUFFER)
-        break;  // Can't get error message, report error code instead.
-      buf.resize(buf.size() * 2);
-    }
-  }
-  FMT_CATCH(...) {}
-  format_error_code(out, error_code, message);
-}
-
-#endif  // FMT_USE_WINDOWS_H
-
-FMT_FUNC void format_system_error(internal::buffer<char>& out, int error_code,
+FMT_FUNC void format_system_error(detail::buffer<char>& out, int error_code,
                                   string_view message) FMT_NOEXCEPT {
   FMT_TRY {
     memory_buffer buf;
@@ -934,12 +2718,10 @@ FMT_FUNC void format_system_error(internal::buffer<char>& out, int error_code,
     for (;;) {
       char* system_message = &buf[0];
       int result =
-          internal::safe_strerror(error_code, system_message, buf.size());
+          detail::safe_strerror(error_code, system_message, buf.size());
       if (result == 0) {
-        internal::writer w(out);
-        w.write(message);
-        w.write(": ");
-        w.write(system_message);
+        format_to(detail::buffer_appender<char>(out), "{}: {}", message,
+                  system_message);
         return;
       }
       if (result != ERANGE)
@@ -951,7 +2733,7 @@ FMT_FUNC void format_system_error(internal::buffer<char>& out, int error_code,
   format_error_code(out, error_code, message);
 }
 
-FMT_FUNC void internal::error_handler::on_error(const char* message) {
+FMT_FUNC void detail::error_handler::on_error(const char* message) {
   FMT_THROW(format_error(message));
 }
 
@@ -960,41 +2742,60 @@ FMT_FUNC void report_system_error(int error_code,
   report_error(format_system_error, error_code, message);
 }
 
-#if FMT_USE_WINDOWS_H
-FMT_FUNC void report_windows_error(int error_code,
-                                   fmt::string_view message) FMT_NOEXCEPT {
-  report_error(internal::format_windows_error, error_code, message);
+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);
+    if (!arg) error_handler().on_error("argument not found");
+    return visit_format_arg(stringifier(), arg);
+  }
+  memory_buffer buffer;
+  detail::vformat_to(buffer, format_str, 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;
-  internal::vformat_to(buffer, format_str,
-                       basic_format_args<buffer_context<char>>(args));
-  internal::fwrite_fully(buffer.data(), 1, buffer.size(), f);
+  detail::vformat_to(buffer, format_str,
+                     basic_format_args<buffer_context<char>>(args));
+#ifdef _WIN32
+  auto fd = _fileno(f);
+  if (_isatty(fd)) {
+    detail::utf8_to_utf16 u16(string_view(buffer.data(), buffer.size()));
+    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;
+  }
+#endif
+  detail::fwrite_fully(buffer.data(), 1, buffer.size(), f);
 }
 
-FMT_FUNC void vprint(std::FILE* f, wstring_view format_str, wformat_args args) {
-  wmemory_buffer buffer;
-  internal::vformat_to(buffer, format_str, args);
-  buffer.push_back(L'\0');
-  if (std::fputws(buffer.data(), f) == -1) {
-    FMT_THROW(system_error(errno, "cannot write to file"));
-  }
+#ifdef _WIN32
+// Print assuming legacy (non-Unicode) encoding.
+FMT_FUNC void detail::vprint_mojibake(std::FILE* f, string_view format_str,
+                                      format_args args) {
+  memory_buffer buffer;
+  detail::vformat_to(buffer, format_str,
+                     basic_format_args<buffer_context<char>>(args));
+  fwrite_fully(buffer.data(), 1, buffer.size(), f);
 }
+#endif
 
 FMT_FUNC void vprint(string_view format_str, format_args args) {
   vprint(stdout, format_str, args);
 }
 
-FMT_FUNC void vprint(wstring_view format_str, wformat_args args) {
-  vprint(stdout, format_str, args);
-}
-
 FMT_END_NAMESPACE
 
-#ifdef _MSC_VER
-#  pragma warning(pop)
-#endif
-
 #endif  // FMT_FORMAT_INL_H_
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/format.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/format.h
index ce5d0695e..1de9c3879 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/format.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/format.h
@@ -36,23 +36,15 @@
 #define FMT_HEADER_ONLY     // Added by diy for header-only usage
 
 #include <algorithm>
-#include <cassert>
+#include <cerrno>
 #include <cmath>
 #include <cstdint>
-#include <cstring>
-#include <iterator>
 #include <limits>
 #include <memory>
 #include <stdexcept>
 
 #include "core.h"
 
-#ifdef __clang__
-#  define FMT_CLANG_VERSION (__clang_major__ * 100 + __clang_minor__)
-#else
-#  define FMT_CLANG_VERSION 0
-#endif
-
 #ifdef __INTEL_COMPILER
 #  define FMT_ICC_VERSION __INTEL_COMPILER
 #elif defined(__ICL)
@@ -73,20 +65,52 @@
 #  define FMT_HAS_BUILTIN(x) 0
 #endif
 
+#if FMT_GCC_VERSION || FMT_CLANG_VERSION
+#  define FMT_NOINLINE __attribute__((noinline))
+#else
+#  define FMT_NOINLINE
+#endif
+
+#if __cplusplus == 201103L || __cplusplus == 201402L
+#  if defined(__INTEL_COMPILER) || defined(__PGI)
+#    define FMT_FALLTHROUGH
+#  elif defined(__clang__)
+#    define FMT_FALLTHROUGH [[clang::fallthrough]]
+#  elif FMT_GCC_VERSION >= 700 && \
+      (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 520)
+#    define FMT_FALLTHROUGH [[gnu::fallthrough]]
+#  else
+#    define FMT_FALLTHROUGH
+#  endif
+#elif FMT_HAS_CPP17_ATTRIBUTE(fallthrough) || \
+    (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
+#  define FMT_FALLTHROUGH [[fallthrough]]
+#else
+#  define FMT_FALLTHROUGH
+#endif
+
+#ifndef FMT_MAYBE_UNUSED
+#  if FMT_HAS_CPP17_ATTRIBUTE(maybe_unused)
+#    define FMT_MAYBE_UNUSED [[maybe_unused]]
+#  else
+#    define FMT_MAYBE_UNUSED
+#  endif
+#endif
+
 #ifndef FMT_THROW
 #  if FMT_EXCEPTIONS
-#    if FMT_MSC_VER
+#    if FMT_MSC_VER || FMT_NVCC
 FMT_BEGIN_NAMESPACE
-namespace internal {
+namespace detail {
 template <typename Exception> inline void do_throw(const Exception& x) {
-  // Silence unreachable code warnings in MSVC because these are nearly
-  // impossible to fix in a generic code.
+  // Silence unreachable code warnings in MSVC and NVCC because these
+  // are nearly impossible to fix in a generic code.
   volatile bool b = true;
   if (b) throw x;
 }
-}  // namespace internal
+}  // namespace detail
 FMT_END_NAMESPACE
-#      define FMT_THROW(x) fmt::internal::do_throw(x)
+#      define FMT_THROW(x) detail::do_throw(x)
 #    else
 #      define FMT_THROW(x) throw x
 #    endif
@@ -94,17 +118,24 @@ FMT_END_NAMESPACE
 #    define FMT_THROW(x)              \
       do {                            \
         static_cast<void>(sizeof(x)); \
-        assert(false);                \
+        FMT_ASSERT(false, "");        \
       } while (false)
 #  endif
 #endif
 
+#if FMT_EXCEPTIONS
+#  define FMT_TRY try
+#  define FMT_CATCH(x) catch (x)
+#else
+#  define FMT_TRY if (true)
+#  define FMT_CATCH(x) if (false)
+#endif
+
 #ifndef FMT_USE_USER_DEFINED_LITERALS
-// For Intel and NVIDIA compilers both they and the system gcc/msc support UDLs.
-#  if (FMT_HAS_FEATURE(cxx_user_literals) || FMT_GCC_VERSION >= 407 ||      \
-       FMT_MSC_VER >= 1900) &&                                              \
-      (!(FMT_ICC_VERSION || FMT_CUDA_VERSION) || FMT_ICC_VERSION >= 1500 || \
-       FMT_CUDA_VERSION >= 700)
+// EDG based compilers (Intel, NVIDIA, Elbrus, etc), GCC and MSVC support UDLs.
+#  if (FMT_HAS_FEATURE(cxx_user_literals) || FMT_GCC_VERSION >= 407 || \
+       FMT_MSC_VER >= 1900) &&                                         \
+      (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= /* UDL feature */ 480)
 #    define FMT_USE_USER_DEFINED_LITERALS 1
 #  else
 #    define FMT_USE_USER_DEFINED_LITERALS 0
@@ -112,25 +143,37 @@ FMT_END_NAMESPACE
 #endif
 
 #ifndef FMT_USE_UDL_TEMPLATE
-// EDG front end based compilers (icc, nvcc) do not support UDL templates yet
-// and GCC 9 warns about them.
-#  if FMT_USE_USER_DEFINED_LITERALS && FMT_ICC_VERSION == 0 && \
-      FMT_CUDA_VERSION == 0 &&                                 \
-      ((FMT_GCC_VERSION >= 600 && FMT_GCC_VERSION <= 900 &&    \
-        __cplusplus >= 201402L) ||                             \
-       FMT_CLANG_VERSION >= 304)
+// 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) &&                            \
+      !defined(__PGI) && !defined(__NVCC__)
 #    define FMT_USE_UDL_TEMPLATE 1
 #  else
 #    define FMT_USE_UDL_TEMPLATE 0
 #  endif
 #endif
 
-#ifdef FMT_USE_INT128
-// Do nothing.
-#elif defined(__SIZEOF_INT128__)
-#  define FMT_USE_INT128 1
-#else
-#  define FMT_USE_INT128 0
+#ifndef FMT_USE_FLOAT
+#  define FMT_USE_FLOAT 1
+#endif
+
+#ifndef FMT_USE_DOUBLE
+#  define FMT_USE_DOUBLE 1
+#endif
+
+#ifndef FMT_USE_LONG_DOUBLE
+#  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:
@@ -141,74 +184,100 @@ 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 internal {
+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) {
-  unsigned long r = 0;
-  _BitScanReverse(&r, x);
-
-  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.
-#  pragma warning(suppress : 6102)
-  return 31 - r;
-}
-#  define FMT_BUILTIN_CLZ(n) fmt::internal::clz(n)
-
 #  if defined(_WIN64) && !defined(__clang__)
+#    pragma intrinsic(_BitScanForward64)
 #    pragma intrinsic(_BitScanReverse64)
 #  endif
 
-inline uint32_t clzll(uint64_t x) {
+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 ^ static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CLZ(n) detail::clz(n)
+
+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
-
-  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.
-#  pragma warning(suppress : 6102)
-  return 63 - r;
+  FMT_ASSERT(x != 0, "");
+  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+  return 63 ^ static_cast<int>(r);
 }
-#  define FMT_BUILTIN_CLZLL(n) fmt::internal::clzll(n)
-}  // namespace internal
+#  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
 
-FMT_BEGIN_NAMESPACE
-namespace internal {
-
-// A fallback implementation of uintptr_t for systems that lack it.
-struct fallback_uintptr {
-  unsigned char value[sizeof(void*)];
-};
-#ifdef UINTPTR_MAX
-using uintptr_t = ::uintptr_t;
-#else
-using uintptr_t = fallback_uintptr;
+// Enable the deprecated numeric alignment.
+#ifndef FMT_DEPRECATED_NUMERIC_ALIGN
+#  define FMT_DEPRECATED_NUMERIC_ALIGN 0
 #endif
 
-// An equivalent of `*reinterpret_cast<Dest*>(&source)` that doesn't produce
+FMT_BEGIN_NAMESPACE
+namespace detail {
+
+// An equivalent of `*reinterpret_cast<Dest*>(&source)` that doesn't have
 // undefined behavior (e.g. due to type aliasing).
 // Example: uint64_t d = bit_cast<uint64_t>(2.718);
 template <typename Dest, typename Source>
@@ -219,42 +288,64 @@ inline Dest bit_cast(const Source& source) {
   return dest;
 }
 
+inline bool is_big_endian() {
+  const auto u = 1u;
+  struct bytes {
+    char data[sizeof(u)];
+  };
+  return bit_cast<bytes>(u).data[0] == 0;
+}
+
+// A fallback implementation of uintptr_t for systems that lack it.
+struct fallback_uintptr {
+  unsigned char value[sizeof(void*)];
+
+  fallback_uintptr() = default;
+  explicit fallback_uintptr(const void* p) {
+    *this = bit_cast<fallback_uintptr>(p);
+    if (is_big_endian()) {
+      for (size_t i = 0, j = sizeof(void*) - 1; i < j; ++i, --j)
+        std::swap(value[i], value[j]);
+    }
+  }
+};
+#ifdef UINTPTR_MAX
+using uintptr_t = ::uintptr_t;
+inline uintptr_t to_uintptr(const void* p) { return bit_cast<uintptr_t>(p); }
+#else
+using uintptr_t = fallback_uintptr;
+inline fallback_uintptr to_uintptr(const void* p) {
+  return fallback_uintptr(p);
+}
+#endif
+
+// Returns the largest possible value for type T. Same as
+// std::numeric_limits<T>::max() but shorter and not affected by the max macro.
+template <typename T> constexpr T max_value() {
+  return (std::numeric_limits<T>::max)();
+}
+template <typename T> constexpr int num_bits() {
+  return std::numeric_limits<T>::digits;
+}
+// std::numeric_limits<T>::digits may return 0 for 128-bit ints.
+template <> constexpr int num_bits<int128_t>() { return 128; }
+template <> constexpr int num_bits<uint128_t>() { return 128; }
+template <> constexpr int num_bits<fallback_uintptr>() {
+  return static_cast<int>(sizeof(void*) *
+                          std::numeric_limits<unsigned char>::digits);
+}
+
+FMT_INLINE void assume(bool condition) {
+  (void)condition;
+#if FMT_HAS_BUILTIN(__builtin_assume)
+  __builtin_assume(condition);
+#endif
+}
+
 // An approximation of iterator_t for pre-C++20 systems.
 template <typename T>
 using iterator_t = decltype(std::begin(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:
-  static const bool value = !std::is_const<remove_reference_t<type>>::value;
-};
+template <typename T> using sentinel_t = decltype(std::end(std::declval<T&>()));
 
 // A workaround for std::string not having mutable data() until C++17.
 template <typename Char> inline Char* get_data(std::basic_string<Char>& s) {
@@ -265,75 +356,115 @@ inline typename Container::value_type* get_data(Container& c) {
   return c.data();
 }
 
-#ifdef _SECURE_SCL
+#if defined(_SECURE_SCL) && _SECURE_SCL
 // Make a checked iterator to avoid MSVC warnings.
 template <typename T> using checked_ptr = stdext::checked_array_iterator<T*>;
-template <typename T> checked_ptr<T> make_checked(T* p, std::size_t size) {
+template <typename T> checked_ptr<T> make_checked(T* p, size_t size) {
   return {p, size};
 }
 #else
 template <typename T> using checked_ptr = T*;
-template <typename T> inline T* make_checked(T* p, std::size_t) { return p; }
+template <typename T> inline T* make_checked(T* p, size_t) { return p; }
 #endif
 
 template <typename Container, FMT_ENABLE_IF(is_contiguous<Container>::value)>
-inline checked_ptr<typename Container::value_type> reserve(
-    std::back_insert_iterator<Container>& it, std::size_t n) {
+#if FMT_CLANG_VERSION
+__attribute__((no_sanitize("undefined")))
+#endif
+inline checked_ptr<typename Container::value_type>
+reserve(std::back_insert_iterator<Container> it, size_t n) {
   Container& c = get_container(it);
-  std::size_t size = c.size();
+  size_t size = c.size();
   c.resize(size + 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,
+    checked_ptr<typename Container::value_type>) {
+  return it;
+}
+
 template <typename Iterator>
-inline Iterator& reserve(Iterator& it, std::size_t) {
+inline Iterator base_iterator(Iterator, Iterator it) {
   return it;
 }
 
 // An output iterator that counts the number of objects written to it and
 // discards them.
-template <typename T> class counting_iterator {
+class counting_iterator {
  private:
-  std::size_t count_;
-  mutable T blackhole_;
+  size_t count_;
 
  public:
   using iterator_category = std::output_iterator_tag;
-  using value_type = T;
   using difference_type = std::ptrdiff_t;
-  using pointer = T*;
-  using reference = T&;
+  using pointer = void;
+  using reference = void;
   using _Unchecked_type = counting_iterator;  // Mark iterator as checked.
 
+  struct value_type {
+    template <typename T> void operator=(const T&) {}
+  };
+
   counting_iterator() : count_(0) {}
 
-  std::size_t count() const { return count_; }
+  size_t count() const { return count_; }
 
   counting_iterator& operator++() {
     ++count_;
     return *this;
   }
-
   counting_iterator operator++(int) {
     auto it = *this;
     ++*this;
     return it;
   }
 
-  T& operator*() const { return blackhole_; }
+  friend counting_iterator operator+(counting_iterator it, difference_type n) {
+    it.count_ += static_cast<size_t>(n);
+    return it;
+  }
+
+  value_type operator*() const { return {}; }
 };
 
 template <typename OutputIt> class truncating_iterator_base {
  protected:
   OutputIt out_;
-  std::size_t limit_;
-  std::size_t count_;
+  size_t limit_;
+  size_t count_;
 
-  truncating_iterator_base(OutputIt out, std::size_t limit)
+  truncating_iterator_base(OutputIt out, size_t limit)
       : out_(out), limit_(limit), count_(0) {}
 
  public:
   using iterator_category = std::output_iterator_tag;
+  using value_type = typename std::iterator_traits<OutputIt>::value_type;
   using difference_type = void;
   using pointer = void;
   using reference = void;
@@ -341,7 +472,7 @@ template <typename OutputIt> class truncating_iterator_base {
       truncating_iterator_base;  // Mark iterator as checked.
 
   OutputIt base() const { return out_; }
-  std::size_t count() const { return count_; }
+  size_t count() const { return count_; }
 };
 
 // An output iterator that truncates the output and counts the number of objects
@@ -354,14 +485,12 @@ class truncating_iterator;
 template <typename OutputIt>
 class truncating_iterator<OutputIt, std::false_type>
     : public truncating_iterator_base<OutputIt> {
-  using traits = std::iterator_traits<OutputIt>;
-
-  mutable typename traits::value_type blackhole_;
+  mutable typename truncating_iterator_base<OutputIt>::value_type blackhole_;
 
  public:
-  using value_type = typename traits::value_type;
+  using value_type = typename truncating_iterator_base<OutputIt>::value_type;
 
-  truncating_iterator(OutputIt out, std::size_t limit)
+  truncating_iterator(OutputIt out, size_t limit)
       : truncating_iterator_base<OutputIt>(out, limit) {}
 
   truncating_iterator& operator++() {
@@ -384,13 +513,11 @@ template <typename OutputIt>
 class truncating_iterator<OutputIt, std::true_type>
     : public truncating_iterator_base<OutputIt> {
  public:
-  using value_type = typename OutputIt::container_type::value_type;
-
-  truncating_iterator(OutputIt out, std::size_t limit)
+  truncating_iterator(OutputIt out, size_t limit)
       : truncating_iterator_base<OutputIt>(out, limit) {}
 
-  truncating_iterator& operator=(value_type val) {
-    if (this->count_++ < this->limit_) this->out_ = val;
+  template <typename T> truncating_iterator& operator=(T val) {
+    if (this->count_++ < this->limit_) *this->out_++ = val;
     return *this;
   }
 
@@ -399,41 +526,14 @@ class truncating_iterator<OutputIt, std::true_type>
   truncating_iterator& operator*() { return *this; }
 };
 
-// A range with the specified output iterator and value type.
-template <typename OutputIt, typename T = typename OutputIt::value_type>
-class output_range {
- private:
-  OutputIt it_;
-
- public:
-  using value_type = T;
-  using iterator = OutputIt;
-  struct sentinel {};
-
-  explicit output_range(OutputIt it) : it_(it) {}
-  OutputIt begin() const { return it_; }
-  sentinel end() const { return {}; }  // Sentinel is not used yet.
-};
-
-// A range with an iterator appending to a buffer.
-template <typename T>
-class buffer_range
-    : public output_range<std::back_insert_iterator<buffer<T>>, T> {
- public:
-  using iterator = std::back_insert_iterator<buffer<T>>;
-  using output_range<iterator, T>::output_range;
-  buffer_range(buffer<T>& buf)
-      : output_range<iterator, T>(std::back_inserter(buf)) {}
-};
-
 template <typename Char>
 inline size_t count_code_points(basic_string_view<Char> s) {
   return s.size();
 }
 
 // Counts the number of code points in a UTF-8 string.
-inline size_t count_code_points(basic_string_view<char8_t> s) {
-  const char8_t* data = s.data();
+inline size_t count_code_points(basic_string_view<char> s) {
+  const char* data = s.data();
   size_t num_code_points = 0;
   for (size_t i = 0, size = s.size(); i != size; ++i) {
     if ((data[i] & 0xc0) != 0x80) ++num_code_points;
@@ -441,13 +541,34 @@ inline size_t count_code_points(basic_string_view<char8_t> s) {
   return num_code_points;
 }
 
-inline char8_t to_char8_t(char c) { return static_cast<char8_t>(c); }
+inline size_t count_code_points(basic_string_view<char8_type> s) {
+  return count_code_points(basic_string_view<char>(
+      reinterpret_cast<const char*>(s.data()), s.size()));
+}
+
+template <typename Char>
+inline size_t code_point_index(basic_string_view<Char> s, size_t n) {
+  size_t size = s.size();
+  return n < size ? n : size;
+}
+
+// Calculates the index of the nth code point in a UTF-8 string.
+inline size_t code_point_index(basic_string_view<char8_type> s, size_t n) {
+  const char8_type* data = s.data();
+  size_t num_code_points = 0;
+  for (size_t i = 0, size = s.size(); i != size; ++i) {
+    if ((data[i] & 0xc0) != 0x80 && ++num_code_points > n) {
+      return i;
+    }
+  }
+  return s.size();
+}
 
 template <typename InputIt, typename OutChar>
 using needs_conversion = bool_constant<
     std::is_same<typename std::iterator_traits<InputIt>::value_type,
                  char>::value &&
-    std::is_same<OutChar, char8_t>::value>;
+    std::is_same<OutChar, char8_type>::value>;
 
 template <typename OutChar, typename InputIt, typename OutputIt,
           FMT_ENABLE_IF(!needs_conversion<InputIt, OutChar>::value)>
@@ -458,45 +579,44 @@ OutputIt copy_str(InputIt begin, InputIt end, OutputIt it) {
 template <typename OutChar, typename InputIt, typename OutputIt,
           FMT_ENABLE_IF(needs_conversion<InputIt, OutChar>::value)>
 OutputIt copy_str(InputIt begin, InputIt end, OutputIt it) {
-  return std::transform(begin, end, it, to_char8_t);
+  return std::transform(begin, end, it,
+                        [](char c) { return static_cast<char8_type>(c); });
 }
 
-#ifndef FMT_USE_GRISU
-#  define FMT_USE_GRISU 0
+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> constexpr bool use_grisu() {
-  return FMT_USE_GRISU && std::numeric_limits<double>::is_iec559 &&
-         sizeof(T) <= sizeof(double);
-}
-
 template <typename T>
 template <typename U>
 void buffer<T>::append(const U* begin, const U* end) {
-  std::size_t new_size = size_ + to_unsigned(end - begin);
-  reserve(new_size);
-  std::uninitialized_copy(begin, end, make_checked(ptr_, 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);
 }
-}  // namespace internal
 
-// A UTF-8 string view.
-class u8string_view : public basic_string_view<char8_t> {
- public:
-  u8string_view(const char* s)
-      : basic_string_view<char8_t>(reinterpret_cast<const char8_t*>(s)) {}
-  u8string_view(const char* s, size_t count) FMT_NOEXCEPT
-      : basic_string_view<char8_t>(reinterpret_cast<const char8_t*>(s), count) {
-  }
-};
-
-#if FMT_USE_USER_DEFINED_LITERALS
-inline namespace literals {
-inline u8string_view operator"" _u(const char* s, std::size_t n) {
-  return {s, n};
+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 literals
-#endif
+}  // namespace detail
 
 // The number of characters to store in the basic_memory_buffer object itself
 // to avoid dynamic memory allocation.
@@ -531,27 +651,30 @@ enum { inline_buffer_size = 500 };
   The output can be converted to an ``std::string`` with ``to_string(out)``.
   \endrst
  */
-template <typename T, std::size_t SIZE = inline_buffer_size,
+template <typename T, size_t SIZE = inline_buffer_size,
           typename Allocator = std::allocator<T>>
-class basic_memory_buffer : private Allocator, public internal::buffer<T> {
+class basic_memory_buffer final : public detail::buffer<T> {
  private:
   T store_[SIZE];
 
+  // Don't inherit from Allocator avoid generating type_info for it.
+  Allocator alloc_;
+
   // Deallocate memory allocated by the buffer.
   void deallocate() {
     T* data = this->data();
-    if (data != store_) Allocator::deallocate(data, this->capacity());
+    if (data != store_) alloc_.deallocate(data, this->capacity());
   }
 
  protected:
-  void grow(std::size_t size) FMT_OVERRIDE;
+  void grow(size_t size) final FMT_OVERRIDE;
 
  public:
   using value_type = T;
   using const_reference = const T&;
 
   explicit basic_memory_buffer(const Allocator& alloc = Allocator())
-      : Allocator(alloc) {
+      : alloc_(alloc) {
     this->set(store_, SIZE);
   }
   ~basic_memory_buffer() { deallocate(); }
@@ -559,14 +682,13 @@ class basic_memory_buffer : private Allocator, public internal::buffer<T> {
  private:
   // Move data from other to this buffer.
   void move(basic_memory_buffer& other) {
-    Allocator &this_alloc = *this, &other_alloc = other;
-    this_alloc = std::move(other_alloc);
+    alloc_ = std::move(other.alloc_);
     T* data = other.data();
-    std::size_t size = other.size(), capacity = other.capacity();
+    size_t size = other.size(), capacity = other.capacity();
     if (data == other.store_) {
       this->set(store_, capacity);
       std::uninitialized_copy(other.store_, other.store_ + size,
-                              internal::make_checked(store_, capacity));
+                              detail::make_checked(store_, capacity));
     } else {
       this->set(data, capacity);
       // Set pointer to the inline array so that delete is not called
@@ -583,91 +705,214 @@ class basic_memory_buffer : private Allocator, public internal::buffer<T> {
     of the other object to it.
     \endrst
    */
-  basic_memory_buffer(basic_memory_buffer&& other) { move(other); }
+  basic_memory_buffer(basic_memory_buffer&& other) FMT_NOEXCEPT { move(other); }
 
   /**
     \rst
     Moves the content of the other ``basic_memory_buffer`` object to this one.
     \endrst
    */
-  basic_memory_buffer& operator=(basic_memory_buffer&& other) {
-    assert(this != &other);
+  basic_memory_buffer& operator=(basic_memory_buffer&& other) FMT_NOEXCEPT {
+    FMT_ASSERT(this != &other, "");
     deallocate();
     move(other);
     return *this;
   }
 
   // Returns a copy of the allocator associated with this buffer.
-  Allocator get_allocator() const { return *this; }
+  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, std::size_t SIZE, typename Allocator>
-void basic_memory_buffer<T, SIZE, Allocator>::grow(std::size_t size) {
-#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
-  if (size > 1000) throw std::runtime_error("fuzz mode - won't grow that much");
+template <typename T, size_t SIZE, typename Allocator>
+void basic_memory_buffer<T, SIZE, Allocator>::grow(size_t size) {
+#ifdef FMT_FUZZ
+  if (size > 5000) throw std::runtime_error("fuzz mode - won't grow that much");
 #endif
-  std::size_t old_capacity = this->capacity();
-  std::size_t new_capacity = old_capacity + old_capacity / 2;
+  size_t old_capacity = this->capacity();
+  size_t new_capacity = old_capacity + old_capacity / 2;
   if (size > new_capacity) new_capacity = size;
   T* old_data = this->data();
-  T* new_data = std::allocator_traits<Allocator>::allocate(*this, new_capacity);
+  T* new_data =
+      std::allocator_traits<Allocator>::allocate(alloc_, new_capacity);
   // The following code doesn't throw, so the raw pointer above doesn't leak.
   std::uninitialized_copy(old_data, old_data + this->size(),
-                          internal::make_checked(new_data, new_capacity));
+                          detail::make_checked(new_data, new_capacity));
   this->set(new_data, new_capacity);
   // deallocate must not throw according to the standard, but even if it does,
   // the buffer already uses the new storage and will deallocate it in
   // destructor.
-  if (old_data != store_) Allocator::deallocate(old_data, old_capacity);
+  if (old_data != store_) alloc_.deallocate(old_data, old_capacity);
 }
 
 using memory_buffer = basic_memory_buffer<char>;
 using wmemory_buffer = basic_memory_buffer<wchar_t>;
 
+template <typename T, size_t SIZE, typename Allocator>
+struct is_contiguous<basic_memory_buffer<T, SIZE, Allocator>> : std::true_type {
+};
+
 /** A formatting error such as invalid format string. */
+FMT_CLASS_API
 class FMT_API format_error : public std::runtime_error {
  public:
   explicit format_error(const char* message) : std::runtime_error(message) {}
   explicit format_error(const std::string& message)
       : std::runtime_error(message) {}
-  ~format_error() FMT_NOEXCEPT;
+  format_error(const format_error&) = default;
+  format_error& operator=(const format_error&) = default;
+  format_error(format_error&&) = default;
+  format_error& operator=(format_error&&) = default;
+  ~format_error() FMT_NOEXCEPT FMT_OVERRIDE;
 };
 
-namespace internal {
+namespace detail {
+
+template <typename T>
+using is_signed =
+    std::integral_constant<bool, std::numeric_limits<T>::is_signed ||
+                                     std::is_same<T, int128_t>::value>;
 
 // Returns true if value is negative, false otherwise.
 // Same as `value < 0` but doesn't produce warnings if T is an unsigned type.
-template <typename T, FMT_ENABLE_IF(std::numeric_limits<T>::is_signed)>
+template <typename T, FMT_ENABLE_IF(is_signed<T>::value)>
 FMT_CONSTEXPR bool is_negative(T value) {
   return value < 0;
 }
-template <typename T, FMT_ENABLE_IF(!std::numeric_limits<T>::is_signed)>
+template <typename T, FMT_ENABLE_IF(!is_signed<T>::value)>
 FMT_CONSTEXPR bool is_negative(T) {
   return false;
 }
 
-// Smallest of uint32_t and uint64_t that is large enough to represent all
-// values of T.
+template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+FMT_CONSTEXPR bool is_supported_floating_point(T) {
+  return (std::is_same<T, float>::value && FMT_USE_FLOAT) ||
+         (std::is_same<T, double>::value && FMT_USE_DOUBLE) ||
+         (std::is_same<T, long double>::value && FMT_USE_LONG_DOUBLE);
+}
+
+// Smallest of uint32_t, uint64_t, uint128_t that is large enough to
+// represent all values of an integral type T.
 template <typename T>
-using uint32_or_64_t =
-    conditional_t<std::numeric_limits<T>::digits <= 32, uint32_t, uint64_t>;
+using uint32_or_64_or_128_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 char digits[];
+  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[];
   static const char hex_digits[];
   static const char foreground_color[];
   static const char background_color[];
   static const char reset_color[5];
   static const wchar_t wreset_color[5];
+  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
 
 // This is a struct rather than an alias to avoid shadowing warnings in gcc.
 struct data : basic_data<> {};
@@ -676,10 +921,9 @@ 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.
@@ -699,6 +943,23 @@ inline int count_digits(uint64_t n) {
 }
 #endif
 
+#if FMT_USE_INT128
+inline int count_digits(uint128_t n) {
+  int count = 1;
+  for (;;) {
+    // Integer division is slow so do it for a group of four digits instead
+    // of for every digit. The idea comes from the talk by Alexandrescu
+    // "Three Optimization Tips for C++". See speed-test for a comparison.
+    if (n < 10) return count;
+    if (n < 100) return count + 1;
+    if (n < 1000) return count + 2;
+    if (n < 10000) return count + 3;
+    n /= 10000U;
+    count += 4;
+  }
+}
+#endif
+
 // Counts the number of digits in n. BITS = log2(radix).
 template <unsigned BITS, typename UInt> inline int count_digits(UInt n) {
   int num_digits = 0;
@@ -708,79 +969,45 @@ template <unsigned BITS, typename UInt> inline int count_digits(UInt n) {
   return num_digits;
 }
 
-template <> int count_digits<4>(internal::fallback_uintptr n);
+template <> int count_digits<4>(detail::fallback_uintptr n);
 
-#if FMT_HAS_CPP_ATTRIBUTE(always_inline)
-#  define FMT_ALWAYS_INLINE __attribute__((always_inline))
+#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
 
-template <typename Handler>
-inline char* lg(uint32_t n, Handler h) FMT_ALWAYS_INLINE;
-
-// Computes g = floor(log10(n)) and calls h.on<g>(n);
-template <typename Handler> inline char* lg(uint32_t n, Handler h) {
-  return n < 100 ? n < 10 ? h.template on<0>(n) : h.template on<1>(n)
-                 : n < 1000000
-                       ? n < 10000 ? n < 1000 ? h.template on<2>(n)
-                                              : h.template on<3>(n)
-                                   : n < 100000 ? h.template on<4>(n)
-                                                : h.template on<5>(n)
-                       : n < 100000000 ? n < 10000000 ? h.template on<6>(n)
-                                                      : h.template on<7>(n)
-                                       : n < 1000000000 ? h.template on<8>(n)
-                                                        : h.template on<9>(n);
-}
-
-// An lg handler that formats a decimal number.
-// Usage: lg(n, decimal_formatter(buffer));
-class decimal_formatter {
- private:
-  char* buffer_;
-
-  void write_pair(unsigned N, uint32_t index) {
-    std::memcpy(buffer_ + N, data::digits + index * 2, 2);
-  }
-
- public:
-  explicit decimal_formatter(char* buf) : buffer_(buf) {}
-
-  template <unsigned N> char* on(uint32_t u) {
-    if (N == 0) {
-      *buffer_ = static_cast<char>(u) + '0';
-    } else if (N == 1) {
-      write_pair(0, u);
-    } else {
-      // The idea of using 4.32 fixed-point numbers is based on
-      // https://github.com/jeaiii/itoa
-      unsigned n = N - 1;
-      unsigned a = n / 5 * n * 53 / 16;
-      uint64_t t =
-          ((1ULL << (32 + a)) / data::zero_or_powers_of_10_32[n] + 1 - n / 9);
-      t = ((t * u) >> a) + n / 5 * 4;
-      write_pair(0, t >> 32);
-      for (unsigned i = 2; i < N; i += 2) {
-        t = 100ULL * static_cast<uint32_t>(t);
-        write_pair(i, t >> 32);
-      }
-      if (N % 2 == 0) {
-        buffer_[N] =
-            static_cast<char>((10ULL * static_cast<uint32_t>(t)) >> 32) + '0';
-      }
-    }
-    return buffer_ += N + 1;
-  }
-};
+// 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
 
+template <typename Int> constexpr int digits10() FMT_NOEXCEPT {
+  return std::numeric_limits<Int>::digits10;
+}
+template <> constexpr int digits10<int128_t>() FMT_NOEXCEPT { return 38; }
+template <> constexpr int digits10<uint128_t>() FMT_NOEXCEPT { return 38; }
+
+template <typename Char> FMT_API std::string grouping_impl(locale_ref loc);
+template <typename Char> inline std::string grouping(locale_ref loc) {
+  return grouping_impl<char>(loc);
+}
+template <> inline std::string grouping<wchar_t>(locale_ref loc) {
+  return grouping_impl<wchar_t>(loc);
+}
+
 template <typename Char> FMT_API Char thousands_sep_impl(locale_ref loc);
 template <typename Char> inline Char thousands_sep(locale_ref loc) {
   return Char(thousands_sep_impl<char>(loc));
@@ -797,51 +1024,60 @@ template <> inline wchar_t decimal_point(locale_ref loc) {
   return decimal_point_impl<wchar_t>(loc);
 }
 
-// Formats a decimal unsigned integer value writing into buffer.
-// add_thousands_sep is called after writing each char to add a thousands
-// separator if necessary.
-template <typename UInt, typename Char, typename F>
-inline Char* format_decimal(Char* buffer, UInt value, int num_digits,
-                            F add_thousands_sep) {
-  FMT_ASSERT(num_digits >= 0, "invalid digit count");
-  buffer += num_digits;
-  Char* end = buffer;
+// Compares two characters for equality.
+template <typename Char> bool equal2(const Char* lhs, const char* rhs) {
+  return lhs[0] == rhs[0] && lhs[1] == rhs[1];
+}
+inline bool equal2(const char* lhs, const char* rhs) {
+  return memcmp(lhs, rhs, 2) == 0;
+}
+
+// Copies two characters from src to dst.
+template <typename Char> void copy2(Char* dst, const char* src) {
+  *dst++ = static_cast<Char>(*src++);
+  *dst = static_cast<Char>(*src);
+}
+FMT_INLINE void copy2(char* dst, const char* src) { memcpy(dst, src, 2); }
+
+template <typename Iterator> struct format_decimal_result {
+  Iterator begin;
+  Iterator end;
+};
+
+// Formats a decimal unsigned integer value writing into out pointing to a
+// buffer of specified size. The caller must ensure that the buffer is large
+// enough.
+template <typename Char, typename UInt>
+inline format_decimal_result<Char*> format_decimal(Char* out, UInt value,
+                                                   int size) {
+  FMT_ASSERT(size >= count_digits(value), "invalid digit count");
+  out += size;
+  Char* end = out;
   while (value >= 100) {
     // Integer division is slow so do it for a group of two digits instead
     // of for every digit. The idea comes from the talk by Alexandrescu
     // "Three Optimization Tips for C++". See speed-test for a comparison.
-    unsigned index = static_cast<unsigned>((value % 100) * 2);
+    out -= 2;
+    copy2(out, data::digits[value % 100]);
     value /= 100;
-    *--buffer = static_cast<Char>(data::digits[index + 1]);
-    add_thousands_sep(buffer);
-    *--buffer = static_cast<Char>(data::digits[index]);
-    add_thousands_sep(buffer);
   }
   if (value < 10) {
-    *--buffer = static_cast<Char>('0' + value);
-    return end;
+    *--out = static_cast<Char>('0' + value);
+    return {out, end};
   }
-  unsigned index = static_cast<unsigned>(value * 2);
-  *--buffer = static_cast<Char>(data::digits[index + 1]);
-  add_thousands_sep(buffer);
-  *--buffer = static_cast<Char>(data::digits[index]);
-  return end;
+  out -= 2;
+  copy2(out, data::digits[value]);
+  return {out, end};
 }
 
-template <typename Char, typename UInt, typename Iterator, typename F>
-inline Iterator format_decimal(Iterator out, UInt value, int num_digits,
-                               F add_thousands_sep) {
-  FMT_ASSERT(num_digits >= 0, "invalid digit count");
-  // Buffer should be large enough to hold all digits (<= digits10 + 1).
-  enum { max_size = std::numeric_limits<UInt>::digits10 + 1 };
-  Char buffer[max_size + max_size / 3];
-  auto end = format_decimal(buffer, value, num_digits, add_thousands_sep);
-  return internal::copy_str<Char>(buffer, end, out);
-}
-
-template <typename Char, typename It, typename UInt>
-inline It format_decimal(It out, UInt value, int num_digits) {
-  return format_decimal<Char>(out, value, num_digits, [](Char*) {});
+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 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)};
 }
 
 template <unsigned BASE_BITS, typename Char, typename UInt>
@@ -859,7 +1095,7 @@ inline Char* format_uint(Char* buffer, UInt value, int num_digits,
 }
 
 template <unsigned BASE_BITS, typename Char>
-Char* format_uint(Char* buffer, internal::fallback_uintptr n, int num_digits,
+Char* format_uint(Char* buffer, detail::fallback_uintptr n, int num_digits,
                   bool = false) {
   auto char_digits = std::numeric_limits<unsigned char>::digits / 4;
   int start = (num_digits + char_digits - 1) / char_digits - 1;
@@ -882,80 +1118,58 @@ Char* format_uint(Char* buffer, internal::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[std::numeric_limits<UInt>::digits / BASE_BITS + 1];
+  char buffer[num_bits<UInt>() / BASE_BITS + 1];
   format_uint<BASE_BITS>(buffer, value, num_digits, upper);
-  return internal::copy_str<Char>(buffer, buffer + num_digits, out);
+  return detail::copy_str<Char>(buffer, buffer + num_digits, out);
 }
 
-#ifndef _WIN32
-#  define FMT_USE_WINDOWS_H 0
-#elif !defined(FMT_USE_WINDOWS_H)
-#  define FMT_USE_WINDOWS_H 1
-#endif
-
-// Define FMT_USE_WINDOWS_H to 0 to disable use of windows.h.
-// All the functionality that relies on it will be disabled too.
-#if FMT_USE_WINDOWS_H
 // A converter from UTF-8 to UTF-16.
-// It is only provided for Windows since other systems support UTF-8 natively.
 class utf8_to_utf16 {
  private:
   wmemory_buffer buffer_;
 
  public:
   FMT_API explicit utf8_to_utf16(string_view s);
-  operator wstring_view() const { return wstring_view(&buffer_[0], size()); }
+  operator wstring_view() const { return {&buffer_[0], size()}; }
   size_t size() const { return buffer_.size() - 1; }
   const wchar_t* c_str() const { return &buffer_[0]; }
-  std::wstring str() const { return std::wstring(&buffer_[0], size()); }
+  std::wstring str() const { return {&buffer_[0], size()}; }
 };
 
-// A converter from UTF-16 to UTF-8.
-// It is only provided for Windows since other systems support UTF-8 natively.
-class utf16_to_utf8 {
- private:
-  memory_buffer buffer_;
-
- public:
-  utf16_to_utf8() {}
-  FMT_API explicit utf16_to_utf8(wstring_view s);
-  operator string_view() const { return string_view(&buffer_[0], size()); }
-  size_t size() const { return buffer_.size() - 1; }
-  const char* c_str() const { return &buffer_[0]; }
-  std::string str() const { return std::string(&buffer_[0], size()); }
-
-  // Performs conversion returning a system error code instead of
-  // throwing exception on conversion error. This method may still throw
-  // in case of memory allocation error.
-  FMT_API int convert(wstring_view s);
-};
-
-FMT_API void format_windows_error(fmt::internal::buffer<char>& out,
-                                  int error_code,
-                                  fmt::string_view message) FMT_NOEXCEPT;
-#endif
-
 template <typename T = void> struct null {};
 
 // Workaround an array initialization issue in gcc 4.8.
 template <typename Char> struct fill_t {
  private:
-  Char data_[6];
+  enum { max_size = 4 };
+  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) {
+    auto size = s.size();
+    if (size > max_size) {
+      FMT_THROW(format_error("invalid fill"));
+      return;
+    }
+    for (size_t i = 0; i < size; ++i) data_[i] = s[i];
+    size_ = static_cast<unsigned char>(size);
+  }
+
+  size_t size() const { return size_; }
+  const Char* data() const { return data_; }
+
   FMT_CONSTEXPR Char& operator[](size_t index) { return data_[index]; }
   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(' ');
-    return fill;
-  }
 };
-}  // namespace internal
+}  // namespace detail
 
 // We cannot use enum classes as bit fields because of a gcc bug
 // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61414.
@@ -977,7 +1191,7 @@ template <typename Char> struct basic_format_specs {
   align_t align : 4;
   sign_t sign : 3;
   bool alt : 1;  // Alternate form ('#').
-  internal::fill_t<Char> fill;
+  detail::fill_t<Char> fill;
 
   constexpr basic_format_specs()
       : width(0),
@@ -985,17 +1199,107 @@ template <typename Char> struct basic_format_specs {
         type(0),
         align(align::none),
         sign(sign::none),
-        alt(false),
-        fill(internal::fill_t<Char>::make()) {}
+        alt(false) {}
 };
 
 using format_specs = basic_format_specs<char>;
 
-namespace internal {
+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 {
+  general,  // General: exponent notation or fixed point based on magnitude.
+  exp,      // Exponent notation with the default precision of 6, e.g. 1.2e-3.
+  fixed,    // Fixed point with the default precision of 6, e.g. 0.0012.
+  hex
+};
+
+struct float_specs {
+  int precision;
+  float_format format : 8;
+  sign_t sign : 8;
+  bool upper : 1;
+  bool locale : 1;
+  bool binary32 : 1;
+  bool use_grisu : 1;
+  bool showpoint : 1;
+};
 
 // Writes the exponent exp in the form "[+-]d{2,3}" to buffer.
 template <typename Char, typename It> It write_exponent(int exp, It it) {
-  FMT_ASSERT(-1000 < exp && exp < 1000, "exponent out of range");
+  FMT_ASSERT(-10000 < exp && exp < 10000, "exponent out of range");
   if (exp < 0) {
     *it++ = static_cast<Char>('-');
     exp = -exp;
@@ -1003,112 +1307,27 @@ template <typename Char, typename It> It write_exponent(int exp, It it) {
     *it++ = static_cast<Char>('+');
   }
   if (exp >= 100) {
-    *it++ = static_cast<Char>(static_cast<char>('0' + exp / 100));
+    const char* top = data::digits[exp / 100];
+    if (exp >= 1000) *it++ = static_cast<Char>(top[0]);
+    *it++ = static_cast<Char>(top[1]);
     exp %= 100;
   }
-  const char* d = data::digits + exp * 2;
+  const char* d = data::digits[exp];
   *it++ = static_cast<Char>(d[0]);
   *it++ = static_cast<Char>(d[1]);
   return it;
 }
 
-struct gen_digits_params {
-  int num_digits;
-  bool fixed;
-  bool upper;
-  bool trailing_zeros;
-};
+template <typename T>
+int format_float(T value, int precision, float_specs specs, buffer<char>& buf);
 
-// The number is given as v = digits * pow(10, exp).
-template <typename Char, typename It>
-It grisu_prettify(const char* digits, int size, int exp, It it,
-                  gen_digits_params params, Char decimal_point) {
-  // pow(10, full_exp - 1) <= v <= pow(10, full_exp).
-  int full_exp = size + exp;
-  if (!params.fixed) {
-    // Insert a decimal point after the first digit and add an exponent.
-    *it++ = static_cast<Char>(*digits);
-    if (size > 1) *it++ = decimal_point;
-    exp += size - 1;
-    it = copy_str<Char>(digits + 1, digits + size, it);
-    if (size < params.num_digits)
-      it = std::fill_n(it, params.num_digits - size, static_cast<Char>('0'));
-    *it++ = static_cast<Char>(params.upper ? 'E' : 'e');
-    return write_exponent<Char>(exp, it);
-  }
-  if (size <= full_exp) {
-    // 1234e7 -> 12340000000[.0+]
-    it = copy_str<Char>(digits, digits + size, it);
-    it = std::fill_n(it, full_exp - size, static_cast<Char>('0'));
-    int num_zeros = (std::max)(params.num_digits - full_exp, 1);
-    if (params.trailing_zeros) {
-      *it++ = decimal_point;
-#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
-      if (num_zeros > 1000)
-        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 (!params.trailing_zeros) {
-      // Remove trailing zeros.
-      while (size > full_exp && digits[size - 1] == '0') --size;
-      if (size != full_exp) *it++ = decimal_point;
-      return copy_str<Char>(digits + full_exp, digits + size, it);
-    }
-    *it++ = decimal_point;
-    it = copy_str<Char>(digits + full_exp, digits + size, it);
-    if (params.num_digits > size) {
-      // Add trailing zeros.
-      int num_zeros = params.num_digits - size;
-      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;
-    if (params.num_digits >= 0 && params.num_digits < num_zeros)
-      num_zeros = params.num_digits;
-    if (!params.trailing_zeros)
-      while (size > 0 && digits[size - 1] == '0') --size;
-    if (num_zeros != 0 || size != 0) {
-      *it++ = decimal_point;
-      it = std::fill_n(it, num_zeros, static_cast<Char>('0'));
-      it = copy_str<Char>(digits, digits + size, it);
-    }
-  }
-  return it;
-}
+// Formats a floating-point number with snprintf.
+template <typename T>
+int snprintf_float(T value, int precision, float_specs specs,
+                   buffer<char>& buf);
 
-namespace grisu_options {
-enum { fixed = 1, grisu3 = 2 };
-}
-
-// Formats value using the Grisu algorithm:
-// https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf
-template <typename Double, FMT_ENABLE_IF(sizeof(Double) == sizeof(uint64_t))>
-FMT_API bool grisu_format(Double, buffer<char>&, int, unsigned, int&);
-template <typename Double, FMT_ENABLE_IF(sizeof(Double) != sizeof(uint64_t))>
-inline bool grisu_format(Double, buffer<char>&, int, unsigned, int&) {
-  return false;
-}
-
-struct sprintf_specs {
-  int precision;
-  char type;
-  bool alt : 1;
-
-  template <typename Char>
-  constexpr sprintf_specs(basic_format_specs<Char> specs)
-      : precision(specs.precision), type(specs.type), alt(specs.alt) {}
-
-  constexpr bool has_precision() const { return precision >= 0; }
-};
-
-template <typename Double>
-char* sprintf_format(Double, internal::buffer<char>&, sprintf_specs);
+template <typename T> T promote_float(T value) { return value; }
+inline double promote_float(float value) { return static_cast<double>(value); }
 
 template <typename Handler>
 FMT_CONSTEXPR void handle_int_type_spec(char spec, Handler&& handler) {
@@ -1128,44 +1347,67 @@ FMT_CONSTEXPR void handle_int_type_spec(char spec, Handler&& handler) {
   case 'o':
     handler.on_oct();
     break;
+#ifdef FMT_DEPRECATED_N_SPECIFIER
   case 'n':
+#endif
+  case 'L':
     handler.on_num();
     break;
+  case 'c':
+    handler.on_chr();
+    break;
   default:
     handler.on_error();
   }
 }
 
-template <typename Handler>
-FMT_CONSTEXPR void handle_float_type_spec(char spec, Handler&& handler) {
-  switch (spec) {
+template <typename ErrorHandler = error_handler, typename Char>
+FMT_CONSTEXPR float_specs parse_float_type_spec(
+    const basic_format_specs<Char>& specs, ErrorHandler&& eh = {}) {
+  auto result = float_specs();
+  result.showpoint = specs.alt;
+  switch (specs.type) {
   case 0:
-  case 'g':
+    result.format = float_format::general;
+    result.showpoint |= specs.precision > 0;
+    break;
   case 'G':
-    handler.on_general();
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case 'g':
+    result.format = float_format::general;
     break;
-  case 'e':
   case 'E':
-    handler.on_exp();
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case 'e':
+    result.format = float_format::exp;
+    result.showpoint |= specs.precision != 0;
     break;
-  case 'f':
   case 'F':
-    handler.on_fixed();
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case 'f':
+    result.format = float_format::fixed;
+    result.showpoint |= specs.precision != 0;
     break;
-  case '%':
-    handler.on_percent();
-    break;
-  case 'a':
   case 'A':
-    handler.on_hex();
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case 'a':
+    result.format = float_format::hex;
     break;
+#ifdef FMT_DEPRECATED_N_SPECIFIER
   case 'n':
-    handler.on_num();
+#endif
+  case 'L':
+    result.locale = true;
     break;
   default:
-    handler.on_error();
+    eh.on_error("invalid type specifier");
     break;
   }
+  return result;
 }
 
 template <typename Char, typename Handler>
@@ -1207,24 +1449,7 @@ template <typename ErrorHandler> class int_type_checker : private ErrorHandler {
   FMT_CONSTEXPR void on_bin() {}
   FMT_CONSTEXPR void on_oct() {}
   FMT_CONSTEXPR void on_num() {}
-
-  FMT_CONSTEXPR void on_error() {
-    ErrorHandler::on_error("invalid type specifier");
-  }
-};
-
-template <typename ErrorHandler>
-class float_type_checker : private ErrorHandler {
- public:
-  FMT_CONSTEXPR explicit float_type_checker(ErrorHandler eh)
-      : ErrorHandler(eh) {}
-
-  FMT_CONSTEXPR void on_general() {}
-  FMT_CONSTEXPR void on_exp() {}
-  FMT_CONSTEXPR void on_fixed() {}
-  FMT_CONSTEXPR void on_percent() {}
-  FMT_CONSTEXPR void on_hex() {}
-  FMT_CONSTEXPR void on_num() {}
+  FMT_CONSTEXPR void on_chr() {}
 
   FMT_CONSTEXPR void on_error() {
     ErrorHandler::on_error("invalid type specifier");
@@ -1256,592 +1481,863 @@ class cstring_type_checker : public ErrorHandler {
   FMT_CONSTEXPR void on_pointer() {}
 };
 
-template <typename Context>
-void arg_map<Context>::init(const basic_format_args<Context>& args) {
-  if (map_) return;
-  map_ = new entry[internal::to_unsigned(args.max_size())];
-  if (args.is_packed()) {
-    for (int i = 0;; ++i) {
-      internal::type arg_type = args.type(i);
-      if (arg_type == internal::none_type) return;
-      if (arg_type == internal::named_arg_type) push_back(args.values_[i]);
-    }
-  }
-  for (int i = 0, n = args.max_size(); i < n; ++i) {
-    auto type = args.args_[i].type_;
-    if (type == internal::named_arg_type) push_back(args.args_[i].value_);
-  }
+template <typename OutputIt, typename Char>
+FMT_NOINLINE OutputIt fill(OutputIt it, size_t n, const fill_t<Char>& fill) {
+  auto fill_size = fill.size();
+  if (fill_size == 1) return std::fill_n(it, n, fill[0]);
+  for (size_t i = 0; i < n; ++i) it = std::copy_n(fill.data(), fill_size, it);
+  return it;
 }
 
-// This template provides operations for formatting and writing data into a
-// character range.
-template <typename Range> class basic_writer {
+// Writes the output of f, padded according to format specifications in specs.
+// size: output size in code units.
+// width: output display width in (terminal) column positions.
+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, 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;
+  auto* shifts = align == align::left ? data::left_padding_shifts
+                                      : data::right_padding_shifts;
+  size_t left_padding = padding >> shifts[specs.align];
+  auto it = reserve(out, size + padding * specs.fill.size());
+  it = fill(it, left_padding, specs.fill);
+  it = f(it);
+  it = fill(it, padding - left_padding, specs.fill);
+  return base_iterator(out, it);
+}
+
+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,
+                             F&& f) {
+  return write_padded<align>(out, specs, size, size, f);
+}
+
+template <typename Char, typename OutputIt>
+OutputIt write_bytes(OutputIt out, string_view bytes,
+                     const basic_format_specs<Char>& specs) {
+  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
+  return write_padded(out, specs, bytes.size(), [bytes](iterator it) {
+    const char* data = bytes.data();
+    return copy_str<Char>(data, data + bytes.size(), it);
+  });
+}
+
+// Data for write_int that doesn't depend on output iterator type. It is used to
+// avoid template code bloat.
+template <typename Char> struct write_int_data {
+  size_t size;
+  size_t padding;
+
+  write_int_data(int num_digits, string_view prefix,
+                 const basic_format_specs<Char>& specs)
+      : size(prefix.size() + to_unsigned(num_digits)), padding(0) {
+    if (specs.align == align::numeric) {
+      auto width = to_unsigned(specs.width);
+      if (width > size) {
+        padding = width - size;
+        size = width;
+      }
+    } else if (specs.precision > num_digits) {
+      size = prefix.size() + to_unsigned(specs.precision);
+      padding = to_unsigned(specs.precision - num_digits);
+    }
+  }
+};
+
+// Writes an integer in the format
+//   <left-padding><prefix><numeric-padding><digits><right-padding>
+// where <digits> are written by f(it).
+template <typename OutputIt, typename Char, typename F>
+OutputIt write_int(OutputIt out, int num_digits, string_view prefix,
+                   const basic_format_specs<Char>& specs, F f) {
+  auto data = write_int_data<Char>(num_digits, prefix, specs);
+  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
+  return write_padded<align::right>(out, specs, data.size, [=](iterator it) {
+    if (prefix.size() != 0)
+      it = copy_str<Char>(prefix.begin(), prefix.end(), it);
+    it = std::fill_n(it, data.padding, static_cast<Char>('0'));
+    return f(it);
+  });
+}
+
+template <typename StrChar, typename Char, typename OutputIt>
+OutputIt write(OutputIt out, basic_string_view<StrChar> s,
+               const basic_format_specs<Char>& specs) {
+  auto data = s.data();
+  auto size = s.size();
+  if (specs.precision >= 0 && to_unsigned(specs.precision) < size)
+    size = code_point_index(s, to_unsigned(specs.precision));
+  auto width = specs.width != 0
+                   ? count_code_points(basic_string_view<StrChar>(data, size))
+                   : 0;
+  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
+  return write_padded(out, specs, size, width, [=](iterator it) {
+    return copy_str<Char>(data, data + size, it);
+  });
+}
+
+// The handle_int_type_spec handler that writes an integer.
+template <typename OutputIt, typename Char, typename UInt> struct int_writer {
+  OutputIt out;
+  locale_ref locale;
+  const basic_format_specs<Char>& specs;
+  UInt abs_value;
+  char prefix[4];
+  unsigned prefix_size;
+
+  using iterator =
+      remove_reference_t<decltype(reserve(std::declval<OutputIt&>(), 0))>;
+
+  string_view get_prefix() const { return string_view(prefix, prefix_size); }
+
+  template <typename Int>
+  int_writer(OutputIt output, locale_ref loc, Int value,
+             const basic_format_specs<Char>& s)
+      : out(output),
+        locale(loc),
+        specs(s),
+        abs_value(static_cast<UInt>(value)),
+        prefix_size(0) {
+    static_assert(std::is_same<uint32_or_64_or_128_t<Int>, UInt>::value, "");
+    if (is_negative(value)) {
+      prefix[0] = '-';
+      ++prefix_size;
+      abs_value = 0 - abs_value;
+    } else if (specs.sign != sign::none && specs.sign != sign::minus) {
+      prefix[0] = specs.sign == sign::plus ? '+' : ' ';
+      ++prefix_size;
+    }
+  }
+
+  void on_dec() {
+    auto num_digits = count_digits(abs_value);
+    out = write_int(
+        out, num_digits, get_prefix(), specs, [this, num_digits](iterator it) {
+          return format_decimal<Char>(it, abs_value, num_digits).end;
+        });
+  }
+
+  void on_hex() {
+    if (specs.alt) {
+      prefix[prefix_size++] = '0';
+      prefix[prefix_size++] = specs.type;
+    }
+    int num_digits = count_digits<4>(abs_value);
+    out = write_int(out, num_digits, get_prefix(), specs,
+                    [this, num_digits](iterator it) {
+                      return format_uint<4, Char>(it, abs_value, num_digits,
+                                                  specs.type != 'x');
+                    });
+  }
+
+  void on_bin() {
+    if (specs.alt) {
+      prefix[prefix_size++] = '0';
+      prefix[prefix_size++] = static_cast<char>(specs.type);
+    }
+    int num_digits = count_digits<1>(abs_value);
+    out = write_int(out, num_digits, get_prefix(), specs,
+                    [this, num_digits](iterator it) {
+                      return format_uint<1, Char>(it, abs_value, num_digits);
+                    });
+  }
+
+  void on_oct() {
+    int num_digits = count_digits<3>(abs_value);
+    if (specs.alt && specs.precision <= num_digits && abs_value != 0) {
+      // Octal prefix '0' is counted as a digit, so only add it if precision
+      // is not greater than the number of digits.
+      prefix[prefix_size++] = '0';
+    }
+    out = write_int(out, num_digits, get_prefix(), specs,
+                    [this, num_digits](iterator it) {
+                      return format_uint<3, Char>(it, abs_value, num_digits);
+                    });
+  }
+
+  enum { sep_size = 1 };
+
+  void on_num() {
+    std::string groups = grouping<Char>(locale);
+    if (groups.empty()) return on_dec();
+    auto sep = thousands_sep<Char>(locale);
+    if (!sep) return on_dec();
+    int num_digits = count_digits(abs_value);
+    int size = num_digits, n = num_digits;
+    std::string::const_iterator group = groups.cbegin();
+    while (group != groups.cend() && n > *group && *group > 0 &&
+           *group != max_value<char>()) {
+      size += sep_size;
+      n -= *group;
+      ++group;
+    }
+    if (group == groups.cend()) size += sep_size * ((n - 1) / groups.back());
+    char digits[40];
+    format_decimal(digits, abs_value, num_digits);
+    basic_memory_buffer<Char> buffer;
+    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 - 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;
+      if (group + 1 != groups.cend()) {
+        digit_index = 0;
+        ++group;
+      }
+      std::uninitialized_copy(s.data(), s.data() + s.size(),
+                              make_checked(p, s.size()));
+      p -= s.size();
+    }
+    *p-- = static_cast<Char>(*digits);
+    if (prefix_size != 0) *p = static_cast<Char>('-');
+    auto data = buffer.data();
+    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); }
+
+  FMT_NORETURN void on_error() {
+    FMT_THROW(format_error("invalid type specifier"));
+  }
+};
+
+template <typename Char, typename OutputIt>
+OutputIt write_nonfinite(OutputIt out, bool isinf,
+                         const basic_format_specs<Char>& specs,
+                         const float_specs& fspecs) {
+  auto str =
+      isinf ? (fspecs.upper ? "INF" : "inf") : (fspecs.upper ? "NAN" : "nan");
+  constexpr size_t str_size = 3;
+  auto sign = fspecs.sign;
+  auto size = str_size + (sign ? 1 : 0);
+  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
+  return write_padded(out, specs, size, [=](iterator it) {
+    if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+    return copy_str<Char>(str, str + str_size, it);
+  });
+}
+
+// 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,
+               locale_ref loc = {}) {
+  if (const_check(!is_supported_floating_point(value))) return out;
+  float_specs fspecs = parse_float_type_spec(specs);
+  fspecs.sign = specs.sign;
+  if (std::signbit(value)) {  // value < 0 is false for NaN so use signbit.
+    fspecs.sign = sign::minus;
+    value = -value;
+  } else if (fspecs.sign == sign::minus) {
+    fspecs.sign = sign::none;
+  }
+
+  if (!std::isfinite(value))
+    return write_nonfinite(out, std::isinf(value), specs, fspecs);
+
+  if (specs.align == align::numeric && fspecs.sign) {
+    auto it = reserve(out, 1);
+    *it++ = static_cast<Char>(data::signs[fspecs.sign]);
+    out = base_iterator(out, it);
+    fspecs.sign = sign::none;
+    if (specs.width != 0) --specs.width;
+  }
+
+  memory_buffer buffer;
+  if (fspecs.format == float_format::hex) {
+    if (fspecs.sign) buffer.push_back(data::signs[fspecs.sign]);
+    snprintf_float(promote_float(value), specs.precision, fspecs, buffer);
+    return write_bytes(out, {buffer.data(), buffer.size()}, specs);
+  }
+  int precision = specs.precision >= 0 || !specs.type ? specs.precision : 6;
+  if (fspecs.format == float_format::exp) {
+    if (precision == max_value<int>())
+      FMT_THROW(format_error("number is too big"));
+    else
+      ++precision;
+  }
+  if (const_check(std::is_same<T, float>())) fspecs.binary32 = true;
+  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>('.');
+  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(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();
+  auto sign_bit = bits & (uint(1) << (num_bits<uint>() - 1));
+  if (sign_bit != 0) {
+    fspecs.sign = sign::minus;
+    value = -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);
+
+  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>
+OutputIt write_char(OutputIt out, Char value,
+                    const basic_format_specs<Char>& specs) {
+  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
+  return write_padded(out, specs, 1, [=](iterator it) {
+    *it++ = value;
+    return it;
+  });
+}
+
+template <typename Char, typename OutputIt, typename UIntPtr>
+OutputIt write_ptr(OutputIt out, UIntPtr value,
+                   const basic_format_specs<Char>* specs) {
+  int num_digits = count_digits<4>(value);
+  auto size = to_unsigned(num_digits) + size_t(2);
+  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
+  auto write = [=](iterator it) {
+    *it++ = static_cast<Char>('0');
+    *it++ = static_cast<Char>('x');
+    return format_uint<4, Char>(it, value, num_digits);
+  };
+  return specs ? write_padded<align::right>(out, *specs, size, write)
+               : base_iterator(out, write(reserve(out, size)));
+}
+
+template <typename T> struct is_integral : std::is_integral<T> {};
+template <> struct is_integral<int128_t> : std::true_type {};
+template <> struct is_integral<uint128_t> : std::true_type {};
+
+template <typename Char, typename OutputIt>
+OutputIt write(OutputIt out, monostate) {
+  FMT_ASSERT(false, "");
+  return out;
+}
+
+template <typename Char, typename OutputIt,
+          FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
+OutputIt write(OutputIt out, string_view value) {
+  auto it = reserve(out, value.size());
+  it = copy_str<Char>(value.begin(), value.end(), it);
+  return base_iterator(out, it);
+}
+
+template <typename Char, typename OutputIt>
+OutputIt write(OutputIt out, basic_string_view<Char> value) {
+  auto it = reserve(out, value.size());
+  it = std::copy(value.begin(), value.end(), it);
+  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 &&
+                        !std::is_same<T, Char>::value)>
+OutputIt write(OutputIt out, T value) {
+  auto abs_value = static_cast<uint32_or_64_or_128_t<T>>(value);
+  bool negative = is_negative(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 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);
+}
+
+template <typename Char, typename OutputIt>
+OutputIt write(OutputIt out, bool value) {
+  return write<Char>(out, string_view(value ? "true" : "false"));
+}
+
+template <typename Char, typename OutputIt>
+OutputIt write(OutputIt out, Char value) {
+  auto it = reserve(out, 1);
+  *it++ = value;
+  return base_iterator(out, it);
+}
+
+template <typename Char, typename OutputIt>
+OutputIt write(OutputIt out, const Char* value) {
+  if (!value) {
+    FMT_THROW(format_error("string pointer is null"));
+  } else {
+    auto length = std::char_traits<Char>::length(value);
+    out = write(out, basic_string_view<Char>(value, length));
+  }
+  return out;
+}
+
+template <typename Char, typename OutputIt>
+OutputIt write(OutputIt out, const void* value) {
+  return write_ptr<Char>(out, to_uintptr(value), nullptr);
+}
+
+template <typename Char, typename OutputIt, typename T>
+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 {
+  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
+// iterator. It's a class and not a generic lambda for compatibility with C++11.
+template <typename OutputIt, typename Char> struct default_arg_formatter {
+  using context = basic_format_context<OutputIt, Char>;
+
+  OutputIt out;
+  basic_format_args<context> args;
+  locale_ref loc;
+
+  template <typename T> OutputIt operator()(T value) {
+    return write<Char>(out, value);
+  }
+
+  OutputIt operator()(typename basic_format_arg<context>::handle handle) {
+    basic_format_parse_context<Char> parse_ctx({});
+    basic_format_context<OutputIt, Char> format_ctx(out, args, loc);
+    handle.format(parse_ctx, format_ctx);
+    return format_ctx.out();
+  }
+};
+
+template <typename OutputIt, typename Char,
+          typename ErrorHandler = error_handler>
+class arg_formatter_base {
  public:
-  using char_type = typename Range::value_type;
-  using iterator = typename Range::iterator;
-  using format_specs = basic_format_specs<char_type>;
+  using iterator = OutputIt;
+  using char_type = Char;
+  using format_specs = basic_format_specs<Char>;
 
  private:
-  iterator out_;  // Output iterator.
-  internal::locale_ref locale_;
+  iterator out_;
+  locale_ref locale_;
+  format_specs* specs_;
 
   // Attempts to reserve space for n extra characters in the output range.
   // Returns a pointer to the reserved range or a reference to out_.
-  auto reserve(std::size_t n) -> decltype(internal::reserve(out_, n)) {
-    return internal::reserve(out_, n);
+  auto reserve(size_t n) -> decltype(detail::reserve(out_, n)) {
+    return detail::reserve(out_, n);
   }
 
-  template <typename F> struct padded_int_writer {
-    size_t size_;
-    string_view prefix;
-    char_type fill;
-    std::size_t padding;
-    F f;
+  using reserve_iterator = remove_reference_t<decltype(
+      detail::reserve(std::declval<iterator&>(), 0))>;
 
-    size_t size() const { return size_; }
-    size_t width() const { return size_; }
-
-    template <typename It> void operator()(It&& it) const {
-      if (prefix.size() != 0)
-        it = internal::copy_str<char_type>(prefix.begin(), prefix.end(), it);
-      it = std::fill_n(it, padding, fill);
-      f(it);
-    }
-  };
-
-  // Writes an integer in the format
-  //   <left-padding><prefix><numeric-padding><digits><right-padding>
-  // where <digits> are written by f(it).
-  template <typename F>
-  void write_int(int num_digits, string_view prefix, format_specs specs, F f) {
-    std::size_t size = prefix.size() + internal::to_unsigned(num_digits);
-    char_type fill = specs.fill[0];
-    std::size_t padding = 0;
-    if (specs.align == align::numeric) {
-      auto unsiged_width = internal::to_unsigned(specs.width);
-      if (unsiged_width > size) {
-        padding = unsiged_width - size;
-        size = unsiged_width;
-      }
-    } else if (specs.precision > num_digits) {
-      size = prefix.size() + internal::to_unsigned(specs.precision);
-      padding = internal::to_unsigned(specs.precision - num_digits);
-      fill = static_cast<char_type>('0');
-    }
-    if (specs.align == align::none) specs.align = align::right;
-    write_padded(specs, padded_int_writer<F>{size, prefix, fill, padding, f});
+  template <typename T> void write_int(T value, const format_specs& spec) {
+    using uint_type = uint32_or_64_or_128_t<T>;
+    int_writer<iterator, Char, uint_type> w(out_, locale_, value, spec);
+    handle_int_type_spec(spec.type, w);
+    out_ = w.out;
   }
 
-  // Writes a decimal integer.
-  template <typename Int> void write_decimal(Int value) {
-    auto abs_value = static_cast<uint32_or_64_t<Int>>(value);
-    bool is_negative = internal::is_negative(value);
-    if (is_negative) abs_value = 0 - abs_value;
-    int num_digits = internal::count_digits(abs_value);
-    auto&& it =
-        reserve((is_negative ? 1 : 0) + static_cast<size_t>(num_digits));
-    if (is_negative) *it++ = static_cast<char_type>('-');
-    it = internal::format_decimal<char_type>(it, abs_value, num_digits);
-  }
-
-  // The handle_int_type_spec handler that writes an integer.
-  template <typename Int, typename Specs> struct int_writer {
-    using unsigned_type = uint32_or_64_t<Int>;
-
-    basic_writer<Range>& writer;
-    const Specs& specs;
-    unsigned_type abs_value;
-    char prefix[4];
-    unsigned prefix_size;
-
-    string_view get_prefix() const { return string_view(prefix, prefix_size); }
-
-    int_writer(basic_writer<Range>& w, Int value, const Specs& s)
-        : writer(w),
-          specs(s),
-          abs_value(static_cast<unsigned_type>(value)),
-          prefix_size(0) {
-      if (internal::is_negative(value)) {
-        prefix[0] = '-';
-        ++prefix_size;
-        abs_value = 0 - abs_value;
-      } else if (specs.sign != sign::none && specs.sign != sign::minus) {
-        prefix[0] = specs.sign == sign::plus ? '+' : ' ';
-        ++prefix_size;
-      }
-    }
-
-    struct dec_writer {
-      unsigned_type abs_value;
-      int num_digits;
-
-      template <typename It> void operator()(It&& it) const {
-        it = internal::format_decimal<char_type>(it, abs_value, num_digits);
-      }
-    };
-
-    void on_dec() {
-      int num_digits = internal::count_digits(abs_value);
-      writer.write_int(num_digits, get_prefix(), specs,
-                       dec_writer{abs_value, num_digits});
-    }
-
-    struct hex_writer {
-      int_writer& self;
-      int num_digits;
-
-      template <typename It> void operator()(It&& it) const {
-        it = internal::format_uint<4, char_type>(it, self.abs_value, num_digits,
-                                                 self.specs.type != 'x');
-      }
-    };
-
-    void on_hex() {
-      if (specs.alt) {
-        prefix[prefix_size++] = '0';
-        prefix[prefix_size++] = specs.type;
-      }
-      int num_digits = internal::count_digits<4>(abs_value);
-      writer.write_int(num_digits, get_prefix(), specs,
-                       hex_writer{*this, num_digits});
-    }
-
-    template <int BITS> struct bin_writer {
-      unsigned_type abs_value;
-      int num_digits;
-
-      template <typename It> void operator()(It&& it) const {
-        it = internal::format_uint<BITS, char_type>(it, abs_value, num_digits);
-      }
-    };
-
-    void on_bin() {
-      if (specs.alt) {
-        prefix[prefix_size++] = '0';
-        prefix[prefix_size++] = static_cast<char>(specs.type);
-      }
-      int num_digits = internal::count_digits<1>(abs_value);
-      writer.write_int(num_digits, get_prefix(), specs,
-                       bin_writer<1>{abs_value, num_digits});
-    }
-
-    void on_oct() {
-      int num_digits = internal::count_digits<3>(abs_value);
-      if (specs.alt && specs.precision <= num_digits) {
-        // Octal prefix '0' is counted as a digit, so only add it if precision
-        // is not greater than the number of digits.
-        prefix[prefix_size++] = '0';
-      }
-      writer.write_int(num_digits, get_prefix(), specs,
-                       bin_writer<3>{abs_value, num_digits});
-    }
-
-    enum { sep_size = 1 };
-
-    struct num_writer {
-      unsigned_type abs_value;
-      int size;
-      char_type sep;
-
-      template <typename It> void operator()(It&& it) const {
-        basic_string_view<char_type> s(&sep, sep_size);
-        // Index of a decimal digit with the least significant digit having
-        // index 0.
-        unsigned digit_index = 0;
-        it = internal::format_decimal<char_type>(
-            it, abs_value, size, [s, &digit_index](char_type*& buffer) {
-              if (++digit_index % 3 != 0) return;
-              buffer -= s.size();
-              std::uninitialized_copy(s.data(), s.data() + s.size(),
-                                      internal::make_checked(buffer, s.size()));
-            });
-      }
-    };
-
-    void on_num() {
-      char_type sep = internal::thousands_sep<char_type>(writer.locale_);
-      if (!sep) return on_dec();
-      int num_digits = internal::count_digits(abs_value);
-      int size = num_digits + sep_size * ((num_digits - 1) / 3);
-      writer.write_int(size, get_prefix(), specs,
-                       num_writer{abs_value, size, sep});
-    }
-
-    FMT_NORETURN void on_error() {
-      FMT_THROW(format_error("invalid type specifier"));
-    }
-  };
-
-  enum { inf_size = 3 };  // This is an enum to workaround a bug in MSVC.
-
-  struct inf_or_nan_writer {
-    char sign;
-    bool as_percentage;
-    const char* str;
-
-    size_t size() const {
-      return static_cast<std::size_t>(inf_size + (sign ? 1 : 0) +
-                                      (as_percentage ? 1 : 0));
-    }
-    size_t width() const { return size(); }
-
-    template <typename It> void operator()(It&& it) const {
-      if (sign) *it++ = static_cast<char_type>(sign);
-      it = internal::copy_str<char_type>(
-          str, str + static_cast<std::size_t>(inf_size), it);
-      if (as_percentage) *it++ = static_cast<char_type>('%');
-    }
-  };
-
-  struct double_writer {
-    char sign;
-    internal::buffer<char>& buffer;
-    char* decimal_point_pos;
-    char_type decimal_point;
-
-    size_t size() const { return buffer.size() + (sign ? 1 : 0); }
-    size_t width() const { return size(); }
-
-    template <typename It> void operator()(It&& it) {
-      if (sign) *it++ = static_cast<char_type>(sign);
-      auto begin = buffer.begin();
-      if (decimal_point_pos) {
-        it = internal::copy_str<char_type>(begin, decimal_point_pos, it);
-        *it++ = decimal_point;
-        begin = decimal_point_pos + 1;
-      }
-      it = internal::copy_str<char_type>(begin, buffer.end(), it);
-    }
-  };
-
-  class grisu_writer {
-   private:
-    internal::buffer<char>& digits_;
-    size_t size_;
-    char sign_;
-    int exp_;
-    internal::gen_digits_params params_;
-    char_type decimal_point_;
-
-   public:
-    grisu_writer(char sign, internal::buffer<char>& digits, int exp,
-                 const internal::gen_digits_params& params,
-                 char_type decimal_point)
-        : digits_(digits),
-          sign_(sign),
-          exp_(exp),
-          params_(params),
-          decimal_point_(decimal_point) {
-      int num_digits = static_cast<int>(digits.size());
-      int full_exp = num_digits + exp - 1;
-      int precision = params.num_digits > 0 ? params.num_digits : 11;
-      params_.fixed |= full_exp >= -4 && full_exp < precision;
-      auto it = internal::grisu_prettify<char>(
-          digits.data(), num_digits, exp, internal::counting_iterator<char>(),
-          params_, '.');
-      size_ = it.count();
-    }
-
-    size_t size() const { return size_ + (sign_ ? 1 : 0); }
-    size_t width() const { return size(); }
-
-    template <typename It> void operator()(It&& it) {
-      if (sign_) *it++ = static_cast<char_type>(sign_);
-      int num_digits = static_cast<int>(digits_.size());
-      it = internal::grisu_prettify<char_type>(digits_.data(), num_digits, exp_,
-                                               it, params_, decimal_point_);
-    }
-  };
-
-  template <typename Char> struct str_writer {
-    const Char* s;
-    size_t size_;
-
-    size_t size() const { return size_; }
-    size_t width() const {
-      return internal::count_code_points(basic_string_view<Char>(s, size_));
-    }
-
-    template <typename It> void operator()(It&& it) const {
-      it = internal::copy_str<char_type>(s, s + size_, it);
-    }
-  };
-
-  template <typename UIntPtr> struct pointer_writer {
-    UIntPtr value;
-    int num_digits;
-
-    size_t size() const { return to_unsigned(num_digits) + 2; }
-    size_t width() const { return size(); }
-
-    template <typename It> void operator()(It&& it) const {
-      *it++ = static_cast<char_type>('0');
-      *it++ = static_cast<char_type>('x');
-      it = internal::format_uint<4, char_type>(it, value, num_digits);
-    }
-  };
-
- public:
-  /** Constructs a ``basic_writer`` object. */
-  explicit basic_writer(Range out,
-                        internal::locale_ref loc = internal::locale_ref())
-      : out_(out.begin()), locale_(loc) {}
-
-  iterator out() const { return out_; }
-
-  // Writes a value in the format
-  //   <left-padding><value><right-padding>
-  // where <value> is written by f(it).
-  template <typename F> void write_padded(const format_specs& specs, F&& f) {
-    // User-perceived width (in code points).
-    unsigned width = to_unsigned(specs.width);
-    size_t size = f.size();  // The number of code units.
-    size_t num_code_points = width != 0 ? f.width() : size;
-    if (width <= num_code_points) return f(reserve(size));
-    auto&& it = reserve(width + (size - num_code_points));
-    char_type fill = specs.fill[0];
-    std::size_t padding = width - num_code_points;
-    if (specs.align == align::right) {
-      it = std::fill_n(it, padding, fill);
-      f(it);
-    } else if (specs.align == align::center) {
-      std::size_t left_padding = padding / 2;
-      it = std::fill_n(it, left_padding, fill);
-      f(it);
-      it = std::fill_n(it, padding - left_padding, fill);
-    } else {
-      f(it);
-      it = std::fill_n(it, padding, fill);
-    }
-  }
-
-  void write(int value) { write_decimal(value); }
-  void write(long value) { write_decimal(value); }
-  void write(long long value) { write_decimal(value); }
-
-  void write(unsigned value) { write_decimal(value); }
-  void write(unsigned long value) { write_decimal(value); }
-  void write(unsigned long long value) { write_decimal(value); }
-
-  // Writes a formatted integer.
-  template <typename T, typename Spec>
-  void write_int(T value, const Spec& spec) {
-    internal::handle_int_type_spec(spec.type,
-                                   int_writer<T, Spec>(*this, value, spec));
-  }
-
-  void write(double value, const format_specs& specs = format_specs()) {
-    write_double(value, specs);
-  }
-
-  /**
-    \rst
-    Formats *value* using the general format for floating-point numbers
-    (``'g'``) and writes it to the buffer.
-    \endrst
-   */
-  void write(long double value, const format_specs& specs = format_specs()) {
-    write_double(value, specs);
-  }
-
-  // Formats a floating-point number (double or long double).
-  template <typename T, bool USE_GRISU = fmt::internal::use_grisu<T>()>
-  void write_double(T value, const format_specs& specs);
-
-  /** Writes a character to the buffer. */
   void write(char value) {
     auto&& it = reserve(1);
     *it++ = value;
   }
 
-  template <typename Char, FMT_ENABLE_IF(std::is_same<Char, char_type>::value)>
-  void write(Char value) {
-    auto&& it = reserve(1);
-    *it++ = value;
+  template <typename Ch, FMT_ENABLE_IF(std::is_same<Ch, Char>::value)>
+  void write(Ch value) {
+    out_ = detail::write<Char>(out_, value);
   }
 
-  /**
-    \rst
-    Writes *value* to the buffer.
-    \endrst
-   */
   void write(string_view value) {
     auto&& it = reserve(value.size());
-    it = internal::copy_str<char_type>(value.begin(), value.end(), it);
+    it = copy_str<Char>(value.begin(), value.end(), it);
   }
   void write(wstring_view value) {
-    static_assert(std::is_same<char_type, wchar_t>::value, "");
+    static_assert(std::is_same<Char, wchar_t>::value, "");
     auto&& it = reserve(value.size());
     it = std::copy(value.begin(), value.end(), it);
   }
 
-  // Writes a formatted string.
-  template <typename Char>
-  void write(const Char* s, std::size_t size, const format_specs& specs) {
-    write_padded(specs, str_writer<Char>{s, size});
+  template <typename Ch>
+  void write(const Ch* s, size_t size, const format_specs& specs) {
+    auto width = specs.width != 0
+                     ? count_code_points(basic_string_view<Ch>(s, size))
+                     : 0;
+    out_ = write_padded(out_, specs, size, width, [=](reserve_iterator it) {
+      return copy_str<Char>(s, s + size, it);
+    });
   }
 
-  template <typename Char>
-  void write(basic_string_view<Char> s,
-             const format_specs& specs = format_specs()) {
-    const Char* data = s.data();
-    std::size_t size = s.size();
-    if (specs.precision >= 0 && internal::to_unsigned(specs.precision) < size)
-      size = internal::to_unsigned(specs.precision);
-    write(data, size, specs);
-  }
-
-  template <typename UIntPtr>
-  void write_pointer(UIntPtr value, const format_specs* specs) {
-    int num_digits = internal::count_digits<4>(value);
-    auto pw = pointer_writer<UIntPtr>{value, num_digits};
-    if (!specs) return pw(reserve(to_unsigned(num_digits) + 2));
-    format_specs specs_copy = *specs;
-    if (specs_copy.align == align::none) specs_copy.align = align::right;
-    write_padded(specs_copy, pw);
-  }
-};
-
-using writer = basic_writer<buffer_range<char>>;
-
-template <typename Range, typename ErrorHandler = internal::error_handler>
-class arg_formatter_base {
- public:
-  using char_type = typename Range::value_type;
-  using iterator = typename Range::iterator;
-  using format_specs = basic_format_specs<char_type>;
-
- private:
-  using writer_type = basic_writer<Range>;
-  writer_type writer_;
-  format_specs* specs_;
-
-  struct char_writer {
-    char_type value;
-
-    size_t size() const { return 1; }
-    size_t width() const { return 1; }
-
-    template <typename It> void operator()(It&& it) const { *it++ = value; }
-  };
-
-  void write_char(char_type value) {
-    if (specs_)
-      writer_.write_padded(*specs_, char_writer{value});
-    else
-      writer_.write(value);
+  template <typename Ch>
+  void write(basic_string_view<Ch> s, const format_specs& specs = {}) {
+    out_ = detail::write(out_, s, specs);
   }
 
   void write_pointer(const void* p) {
-    writer_.write_pointer(internal::bit_cast<internal::uintptr_t>(p), specs_);
-  }
-
- protected:
-  writer_type& writer() { return writer_; }
-  FMT_DEPRECATED format_specs* spec() { return specs_; }
-  format_specs* specs() { return specs_; }
-  iterator out() { return writer_.out(); }
-
-  void write(bool value) {
-    string_view sv(value ? "true" : "false");
-    specs_ ? writer_.write(sv, *specs_) : writer_.write(sv);
-  }
-
-  void write(const char_type* value) {
-    if (!value) {
-      FMT_THROW(format_error("string pointer is null"));
-    } else {
-      auto length = std::char_traits<char_type>::length(value);
-      basic_string_view<char_type> sv(value, length);
-      specs_ ? writer_.write(sv, *specs_) : writer_.write(sv);
-    }
-  }
-
- public:
-  arg_formatter_base(Range r, format_specs* s, locale_ref loc)
-      : writer_(r, loc), specs_(s) {}
-
-  iterator operator()(monostate) {
-    FMT_ASSERT(false, "invalid argument type");
-    return out();
-  }
-
-  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
-  iterator operator()(T value) {
-    if (specs_)
-      writer_.write_int(value, *specs_);
-    else
-      writer_.write(value);
-    return out();
-  }
-
-  iterator operator()(char_type value) {
-    internal::handle_char_specs(
-        specs_, char_spec_handler(*this, static_cast<char_type>(value)));
-    return out();
-  }
-
-  iterator operator()(bool value) {
-    if (specs_ && specs_->type) return (*this)(value ? 1 : 0);
-    write(value != 0);
-    return out();
-  }
-
-  template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
-  iterator operator()(T value) {
-    writer_.write_double(value, specs_ ? *specs_ : format_specs());
-    return out();
+    out_ = write_ptr<char_type>(out_, to_uintptr(p), specs_);
   }
 
   struct char_spec_handler : ErrorHandler {
     arg_formatter_base& formatter;
-    char_type value;
+    Char value;
 
-    char_spec_handler(arg_formatter_base& f, char_type val)
+    char_spec_handler(arg_formatter_base& f, Char val)
         : formatter(f), value(val) {}
 
     void on_int() {
-      if (formatter.specs_)
-        formatter.writer_.write_int(value, *formatter.specs_);
-      else
-        formatter.writer_.write(value);
+      // char is only formatted as int if there are specs.
+      formatter.write_int(static_cast<int>(value), *formatter.specs_);
+    }
+    void on_char() {
+      if (formatter.specs_)
+        formatter.out_ = write_char(formatter.out_, value, *formatter.specs_);
+      else
+        formatter.write(value);
     }
-    void on_char() { formatter.write_char(value); }
   };
 
-  struct cstring_spec_handler : internal::error_handler {
+  struct cstring_spec_handler : error_handler {
     arg_formatter_base& formatter;
-    const char_type* value;
+    const Char* value;
 
-    cstring_spec_handler(arg_formatter_base& f, const char_type* val)
+    cstring_spec_handler(arg_formatter_base& f, const Char* val)
         : formatter(f), value(val) {}
 
     void on_string() { formatter.write(value); }
     void on_pointer() { formatter.write_pointer(value); }
   };
 
-  iterator operator()(const char_type* value) {
-    if (!specs_) return write(value), out();
-    internal::handle_cstring_type_spec(specs_->type,
-                                       cstring_spec_handler(*this, value));
-    return out();
+ protected:
+  iterator out() { return out_; }
+  format_specs* specs() { return specs_; }
+
+  void write(bool value) {
+    if (specs_)
+      write(string_view(value ? "true" : "false"), *specs_);
+    else
+      out_ = detail::write<Char>(out_, value);
   }
 
-  iterator operator()(basic_string_view<char_type> value) {
-    if (specs_) {
-      internal::check_string_type_spec(specs_->type, internal::error_handler());
-      writer_.write(value, *specs_);
+  void write(const Char* value) {
+    if (!value) {
+      FMT_THROW(format_error("string pointer is null"));
     } else {
-      writer_.write(value);
+      auto length = std::char_traits<char_type>::length(value);
+      basic_string_view<char_type> sv(value, length);
+      specs_ ? write(sv, *specs_) : write(sv);
     }
-    return out();
+  }
+
+ public:
+  arg_formatter_base(OutputIt out, format_specs* s, locale_ref loc)
+      : out_(out), locale_(loc), specs_(s) {}
+
+  iterator operator()(monostate) {
+    FMT_ASSERT(false, "invalid argument type");
+    return out_;
+  }
+
+  template <typename T, FMT_ENABLE_IF(is_integral<T>::value)>
+  FMT_INLINE iterator operator()(T value) {
+    if (specs_)
+      write_int(value, *specs_);
+    else
+      out_ = detail::write<Char>(out_, value);
+    return out_;
+  }
+
+  iterator operator()(Char value) {
+    handle_char_specs(specs_,
+                      char_spec_handler(*this, static_cast<Char>(value)));
+    return out_;
+  }
+
+  iterator operator()(bool value) {
+    if (specs_ && specs_->type) return (*this)(value ? 1 : 0);
+    write(value != 0);
+    return out_;
+  }
+
+  template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+  iterator operator()(T value) {
+    auto specs = specs_ ? *specs_ : format_specs();
+    if (const_check(is_supported_floating_point(value)))
+      out_ = detail::write(out_, value, specs, locale_);
+    else
+      FMT_ASSERT(false, "unsupported float argument type");
+    return out_;
+  }
+
+  iterator operator()(const Char* value) {
+    if (!specs_) return write(value), out_;
+    handle_cstring_type_spec(specs_->type, cstring_spec_handler(*this, value));
+    return out_;
+  }
+
+  iterator operator()(basic_string_view<Char> value) {
+    if (specs_) {
+      check_string_type_spec(specs_->type, error_handler());
+      write(value, *specs_);
+    } else {
+      write(value);
+    }
+    return out_;
   }
 
   iterator operator()(const void* value) {
-    if (specs_)
-      check_pointer_type_spec(specs_->type, internal::error_handler());
+    if (specs_) check_pointer_type_spec(specs_->type, error_handler());
     write_pointer(value);
-    return out();
+    return out_;
+  }
+};
+
+/** 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();
   }
 };
 
@@ -1854,14 +2350,10 @@ template <typename Char> FMT_CONSTEXPR bool is_name_start(Char c) {
 template <typename Char, typename ErrorHandler>
 FMT_CONSTEXPR int parse_nonnegative_int(const Char*& begin, const Char* end,
                                         ErrorHandler&& eh) {
-  assert(begin != end && '0' <= *begin && *begin <= '9');
-  if (*begin == '0') {
-    ++begin;
-    return 0;
-  }
+  FMT_ASSERT(begin != end && '0' <= *begin && *begin <= '9', "");
   unsigned value = 0;
   // Convert to unsigned to prevent a warning.
-  constexpr unsigned max_int = (std::numeric_limits<int>::max)();
+  constexpr unsigned max_int = max_value<int>();
   unsigned big = max_int / 10;
   do {
     // Check for overflow.
@@ -1880,25 +2372,24 @@ template <typename Context> class custom_formatter {
  private:
   using char_type = typename Context::char_type;
 
-  basic_parse_context<char_type>& parse_ctx_;
+  basic_format_parse_context<char_type>& parse_ctx_;
   Context& ctx_;
 
  public:
-  explicit custom_formatter(basic_parse_context<char_type>& parse_ctx,
+  explicit custom_formatter(basic_format_parse_context<char_type>& parse_ctx,
                             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>
 using is_integer =
-    bool_constant<std::is_integral<T>::value && !std::is_same<T, bool>::value &&
+    bool_constant<is_integral<T>::value && !std::is_same<T, bool>::value &&
                   !std::is_same<T, char>::value &&
                   !std::is_same<T, wchar_t>::value>;
 
@@ -1952,7 +2443,9 @@ template <typename Char> class specs_setter {
       : specs_(other.specs_) {}
 
   FMT_CONSTEXPR void on_align(align_t align) { specs_.align = align; }
-  FMT_CONSTEXPR void on_fill(Char fill) { specs_.fill[0] = fill; }
+  FMT_CONSTEXPR void on_fill(basic_string_view<Char> fill) {
+    specs_.fill = fill;
+  }
   FMT_CONSTEXPR void on_plus() { specs_.sign = sign::plus; }
   FMT_CONSTEXPR void on_minus() { specs_.sign = sign::minus; }
   FMT_CONSTEXPR void on_space() { specs_.sign = sign::space; }
@@ -1979,41 +2472,47 @@ template <typename Char> class specs_setter {
 
 template <typename ErrorHandler> class numeric_specs_checker {
  public:
-  FMT_CONSTEXPR numeric_specs_checker(ErrorHandler& eh, internal::type arg_type)
+  FMT_CONSTEXPR numeric_specs_checker(ErrorHandler& eh, detail::type arg_type)
       : error_handler_(eh), arg_type_(arg_type) {}
 
   FMT_CONSTEXPR void require_numeric_argument() {
-    if (!is_arithmetic(arg_type_))
+    if (!is_arithmetic_type(arg_type_))
       error_handler_.on_error("format specifier requires numeric argument");
   }
 
   FMT_CONSTEXPR void check_sign() {
     require_numeric_argument();
-    if (is_integral(arg_type_) && arg_type_ != int_type &&
-        arg_type_ != long_long_type && arg_type_ != internal::char_type) {
+    if (is_integral_type(arg_type_) && arg_type_ != type::int_type &&
+        arg_type_ != type::long_long_type && arg_type_ != type::char_type) {
       error_handler_.on_error("format specifier requires signed argument");
     }
   }
 
   FMT_CONSTEXPR void check_precision() {
-    if (is_integral(arg_type_) || arg_type_ == internal::pointer_type)
+    if (is_integral_type(arg_type_) || arg_type_ == type::pointer_type)
       error_handler_.on_error("precision not allowed for this argument type");
   }
 
  private:
   ErrorHandler& error_handler_;
-  internal::type arg_type_;
+  detail::type arg_type_;
 };
 
 // A format specifier handler that checks if specifiers are consistent with the
 // argument type.
 template <typename Handler> class specs_checker : public Handler {
+ private:
+  numeric_specs_checker<Handler> checker_;
+
+  // Suppress an MSVC warning about using this in initializer list.
+  FMT_CONSTEXPR Handler& error_handler() { return *this; }
+
  public:
-  FMT_CONSTEXPR specs_checker(const Handler& handler, internal::type arg_type)
-      : Handler(handler), checker_(*this, arg_type) {}
+  FMT_CONSTEXPR specs_checker(const Handler& handler, detail::type arg_type)
+      : Handler(handler), checker_(error_handler(), arg_type) {}
 
   FMT_CONSTEXPR specs_checker(const specs_checker& other)
-      : Handler(other), checker_(*this, other.arg_type_) {}
+      : Handler(other), checker_(error_handler(), other.arg_type_) {}
 
   FMT_CONSTEXPR void on_align(align_t align) {
     if (align == align::numeric) checker_.require_numeric_argument();
@@ -2046,27 +2545,22 @@ template <typename Handler> class specs_checker : public Handler {
   }
 
   FMT_CONSTEXPR void end_precision() { checker_.check_precision(); }
-
- private:
-  numeric_specs_checker<Handler> checker_;
 };
 
-template <template <typename> class Handler, typename T, typename FormatArg,
+template <template <typename> class Handler, typename FormatArg,
           typename ErrorHandler>
-FMT_CONSTEXPR void set_dynamic_spec(T& value, FormatArg arg, ErrorHandler eh) {
-  unsigned long long big_value =
-      visit_format_arg(Handler<ErrorHandler>(eh), arg);
-  if (big_value > to_unsigned((std::numeric_limits<int>::max)()))
-    eh.on_error("number is too big");
-  value = static_cast<T>(big_value);
+FMT_CONSTEXPR int get_dynamic_spec(FormatArg arg, ErrorHandler eh) {
+  unsigned long long value = visit_format_arg(Handler<ErrorHandler>(eh), arg);
+  if (value > to_unsigned(max_value<int>())) eh.on_error("number is too big");
+  return static_cast<int>(value);
 }
 
 struct auto_id {};
 
-template <typename Context>
-FMT_CONSTEXPR typename Context::format_arg get_arg(Context& ctx, int id) {
+template <typename Context, typename ID>
+FMT_CONSTEXPR typename Context::format_arg get_arg(Context& ctx, ID id) {
   auto arg = ctx.arg(id);
-  if (!arg) ctx.on_error("argument index out of range");
+  if (!arg) ctx.on_error("argument not found");
   return arg;
 }
 
@@ -2083,13 +2577,13 @@ class specs_handler : public specs_setter<typename Context::char_type> {
         context_(ctx) {}
 
   template <typename Id> FMT_CONSTEXPR void on_dynamic_width(Id arg_id) {
-    set_dynamic_spec<width_checker>(this->specs_.width, get_arg(arg_id),
-                                    context_.error_handler());
+    this->specs_.width = get_dynamic_spec<width_checker>(
+        get_arg(arg_id), context_.error_handler());
   }
 
   template <typename Id> FMT_CONSTEXPR void on_dynamic_precision(Id arg_id) {
-    set_dynamic_spec<precision_checker>(this->specs_.precision, get_arg(arg_id),
-                                        context_.error_handler());
+    this->specs_.precision = get_dynamic_spec<precision_checker>(
+        get_arg(arg_id), context_.error_handler());
   }
 
   void on_error(const char* message) { context_.on_error(message); }
@@ -2099,49 +2593,32 @@ class specs_handler : public specs_setter<typename Context::char_type> {
   using format_arg = typename Context::format_arg;
 
   FMT_CONSTEXPR format_arg get_arg(auto_id) {
-    return internal::get_arg(context_, parse_context_.next_arg_id());
+    return detail::get_arg(context_, parse_context_.next_arg_id());
   }
 
   FMT_CONSTEXPR format_arg get_arg(int arg_id) {
     parse_context_.check_arg_id(arg_id);
-    return internal::get_arg(context_, arg_id);
+    return detail::get_arg(context_, arg_id);
   }
 
   FMT_CONSTEXPR format_arg get_arg(basic_string_view<char_type> arg_id) {
     parse_context_.check_arg_id(arg_id);
-    return context_.arg(arg_id);
+    return detail::get_arg(context_, arg_id);
   }
 
   ParseContext& parse_context_;
   Context& context_;
 };
 
-struct string_view_metadata {
-  FMT_CONSTEXPR string_view_metadata() : offset_(0u), size_(0u) {}
-  template <typename Char>
-  FMT_CONSTEXPR string_view_metadata(basic_string_view<Char> primary_string,
-                                     basic_string_view<Char> view)
-      : offset_(to_unsigned(view.data() - primary_string.data())),
-        size_(view.size()) {}
-  FMT_CONSTEXPR string_view_metadata(std::size_t offset, std::size_t size)
-      : offset_(offset), size_(size) {}
-  template <typename Char>
-  FMT_CONSTEXPR basic_string_view<Char> to_view(const Char* str) const {
-    return {str + offset_, size_};
-  }
-
-  std::size_t offset_;
-  std::size_t size_;
-};
-
 enum class arg_id_kind { none, index, name };
 
 // An argument reference.
 template <typename Char> struct arg_ref {
   FMT_CONSTEXPR arg_ref() : kind(arg_id_kind::none), val() {}
+
   FMT_CONSTEXPR explicit arg_ref(int index)
       : kind(arg_id_kind::index), val(index) {}
-  FMT_CONSTEXPR explicit arg_ref(string_view_metadata name)
+  FMT_CONSTEXPR explicit arg_ref(basic_string_view<Char> name)
       : kind(arg_id_kind::name), val(name) {}
 
   FMT_CONSTEXPR arg_ref& operator=(int idx) {
@@ -2152,12 +2629,11 @@ template <typename Char> struct arg_ref {
 
   arg_id_kind kind;
   union value {
-    FMT_CONSTEXPR value() : index(0u) {}
-    FMT_CONSTEXPR value(int id) : index(id) {}
-    FMT_CONSTEXPR value(string_view_metadata n) : name(n) {}
+    FMT_CONSTEXPR value(int id = 0) : index{id} {}
+    FMT_CONSTEXPR value(basic_string_view<Char> n) : name(n) {}
 
     int index;
-    string_view_metadata name;
+    basic_string_view<Char> name;
   } val;
 };
 
@@ -2215,8 +2691,7 @@ class dynamic_specs_handler
     context_.check_arg_id(arg_id);
     basic_string_view<char_type> format_str(
         context_.begin(), to_unsigned(context_.end() - context_.begin()));
-    const auto id_metadata = string_view_metadata(format_str, arg_id);
-    return arg_ref_type(id_metadata);
+    return arg_ref_type(arg_id);
   }
 
   dynamic_format_specs<char_type>& specs_;
@@ -2226,18 +2701,28 @@ class dynamic_specs_handler
 template <typename Char, typename IDHandler>
 FMT_CONSTEXPR const Char* parse_arg_id(const Char* begin, const Char* end,
                                        IDHandler&& handler) {
-  assert(begin != end);
+  FMT_ASSERT(begin != end, "");
   Char c = *begin;
-  if (c == '}' || c == ':') return handler(), begin;
-  if (c >= '0' && c <= '9') {
-    int index = parse_nonnegative_int(begin, end, handler);
-    if (begin == end || (*begin != '}' && *begin != ':'))
-      return handler.on_error("invalid format string"), begin;
-    handler(index);
+  if (c == '}' || c == ':') {
+    handler();
+    return begin;
+  }
+  if (c >= '0' && c <= '9') {
+    int index = 0;
+    if (c != '0')
+      index = parse_nonnegative_int(begin, end, handler);
+    else
+      ++begin;
+    if (begin == end || (*begin != '}' && *begin != ':'))
+      handler.on_error("invalid format string");
+    else
+      handler(index);
+    return begin;
+  }
+  if (!is_name_start(c)) {
+    handler.on_error("invalid format string");
     return begin;
   }
-  if (!is_name_start(c))
-    return handler.on_error("invalid format string"), begin;
   auto it = begin;
   do {
     ++it;
@@ -2280,42 +2765,74 @@ template <typename SpecHandler, typename Char> struct precision_adapter {
   SpecHandler& handler;
 };
 
+template <typename Char>
+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.
 template <typename Char, typename Handler>
 FMT_CONSTEXPR const Char* parse_align(const Char* begin, const Char* end,
                                       Handler&& handler) {
   FMT_ASSERT(begin != end, "");
   auto align = align::none;
-  int i = 0;
-  if (begin + 1 != end) ++i;
-  do {
-    switch (static_cast<char>(begin[i])) {
+  auto p = begin + code_point_length(begin);
+  if (p >= end) p = begin;
+  for (;;) {
+    switch (to_ascii(*p)) {
     case '<':
       align = align::left;
       break;
     case '>':
       align = align::right;
       break;
+#if FMT_DEPRECATED_NUMERIC_ALIGN
     case '=':
       align = align::numeric;
       break;
+#endif
     case '^':
       align = align::center;
       break;
     }
     if (align != align::none) {
-      if (i > 0) {
+      if (p != begin) {
         auto c = *begin;
         if (c == '{')
           return handler.on_error("invalid fill character '{'"), begin;
-        begin += 2;
-        handler.on_fill(c);
+        handler.on_fill(basic_string_view<Char>(begin, to_unsigned(p - begin)));
+        begin = p + 1;
       } else
         ++begin;
       handler.on_align(align);
       break;
+    } else if (p == begin) {
+      break;
     }
-  } while (i-- > 0);
+    p = begin;
+  }
   return begin;
 }
 
@@ -2363,13 +2880,13 @@ 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;
@@ -2422,71 +2939,98 @@ template <>
 inline bool find<false, char>(const char* first, const char* last, char value,
                               const char*& out) {
   out = static_cast<const char*>(
-      std::memchr(first, value, internal::to_unsigned(last - first)));
+      std::memchr(first, value, detail::to_unsigned(last - first)));
   return out != nullptr;
 }
 
 template <typename Handler, typename Char> struct id_adapter {
-  FMT_CONSTEXPR void operator()() { handler.on_arg_id(); }
-  FMT_CONSTEXPR void operator()(int id) { handler.on_arg_id(id); }
+  Handler& handler;
+  int arg_id;
+
+  FMT_CONSTEXPR void operator()() { arg_id = handler.on_arg_id(); }
+  FMT_CONSTEXPR void operator()(int id) { arg_id = handler.on_arg_id(id); }
   FMT_CONSTEXPR void operator()(basic_string_view<Char> id) {
-    handler.on_arg_id(id);
+    arg_id = handler.on_arg_id(id);
   }
   FMT_CONSTEXPR void on_error(const char* message) {
     handler.on_error(message);
   }
-  Handler& handler;
 };
 
+template <typename Char, typename Handler>
+FMT_CONSTEXPR const Char* parse_replacement_field(const Char* begin,
+                                                  const Char* end,
+                                                  Handler&& handler) {
+  ++begin;
+  if (begin == end) return handler.on_error("invalid format string"), end;
+  if (*begin == '}') {
+    handler.on_replacement_field(handler.on_arg_id(), begin);
+  } else if (*begin == '{') {
+    handler.on_text(begin, begin + 1);
+  } else {
+    auto adapter = id_adapter<Handler, Char>{handler, 0};
+    begin = parse_arg_id(begin, end, adapter);
+    Char c = begin != end ? *begin : Char();
+    if (c == '}') {
+      handler.on_replacement_field(adapter.arg_id, begin);
+    } else if (c == ':') {
+      begin = handler.on_format_specs(adapter.arg_id, begin + 1, end);
+      if (begin == end || *begin != '}')
+        return handler.on_error("unknown format specifier"), end;
+    } else {
+      return handler.on_error("missing '}' in format string"), end;
+    }
+  }
+  return begin + 1;
+}
+
 template <bool IS_CONSTEXPR, typename Char, typename Handler>
-FMT_CONSTEXPR void parse_format_string(basic_string_view<Char> format_str,
-                                       Handler&& handler) {
+FMT_CONSTEXPR_DECL FMT_INLINE void parse_format_string(
+    basic_string_view<Char> format_str, Handler&& handler) {
+  auto begin = format_str.data();
+  auto end = begin + format_str.size();
+  if (end - begin < 32) {
+    // Use a simple loop instead of memchr for small strings.
+    const Char* p = begin;
+    while (p != end) {
+      auto c = *p++;
+      if (c == '{') {
+        handler.on_text(begin, p - 1);
+        begin = p = parse_replacement_field(p - 1, end, handler);
+      } else if (c == '}') {
+        if (p == end || *p != '}')
+          return handler.on_error("unmatched '}' in format string");
+        handler.on_text(begin, p);
+        begin = ++p;
+      }
+    }
+    handler.on_text(begin, end);
+    return;
+  }
   struct pfs_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_;
   } write{handler};
-  auto begin = format_str.data();
-  auto end = begin + format_str.size();
   while (begin != end) {
     // Doing two passes with memchr (one for '{' and another for '}') is up to
     // 2.5x faster than the naive one-pass implementation on big format strings.
     const Char* p = begin;
-    if (*begin != '{' && !find<IS_CONSTEXPR>(begin, end, '{', p))
+    if (*begin != '{' && !find<IS_CONSTEXPR>(begin + 1, end, '{', p))
       return write(begin, end);
     write(begin, p);
-    ++p;
-    if (p == end) return handler.on_error("invalid format string");
-    if (static_cast<char>(*p) == '}') {
-      handler.on_arg_id();
-      handler.on_replacement_field(p);
-    } else if (*p == '{') {
-      handler.on_text(p, p + 1);
-    } else {
-      p = parse_arg_id(p, end, id_adapter<Handler, Char>{handler});
-      Char c = p != end ? *p : Char();
-      if (c == '}') {
-        handler.on_replacement_field(p);
-      } else if (c == ':') {
-        p = handler.on_format_specs(p + 1, end);
-        if (p == end || *p != '}')
-          return handler.on_error("unknown format specifier");
-      } else {
-        return handler.on_error("missing '}' in format string");
-      }
-    }
-    begin = p + 1;
+    begin = parse_replacement_field(p, end, handler);
   }
 }
 
@@ -2496,45 +3040,127 @@ FMT_CONSTEXPR const typename ParseContext::char_type* parse_format_specs(
   using char_type = typename ParseContext::char_type;
   using context = buffer_context<char_type>;
   using mapped_type =
-      conditional_t<internal::mapped_type_constant<T, context>::value !=
-                        internal::custom_type,
+      conditional_t<detail::mapped_type_constant<T, context>::value !=
+                        type::custom_type,
                     decltype(arg_mapper<context>().map(std::declval<T>())), T>;
-  conditional_t<has_formatter<mapped_type, context>::value,
-                formatter<mapped_type, char_type>,
-                internal::fallback_formatter<T, char_type>>
-      f;
+  auto f = conditional_t<has_formatter<mapped_type, context>::value,
+                         formatter<mapped_type, char_type>,
+                         detail::fallback_formatter<T, char_type>>();
   return f.parse(ctx);
 }
 
+template <typename OutputIt, typename Char, typename Context>
+struct format_handler : detail::error_handler {
+  basic_format_parse_context<Char> parse_context;
+  Context context;
+
+  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) {}
+
+  void on_text(const Char* begin, const Char* end) {
+    auto size = to_unsigned(end - begin);
+    auto out = context.out();
+    auto&& it = reserve(out, size);
+    it = std::copy_n(begin, size, it);
+    context.advance_to(out);
+  }
+
+  int on_arg_id() { return parse_context.next_arg_id(); }
+  int on_arg_id(int id) { return parse_context.check_arg_id(id), id; }
+  int on_arg_id(basic_string_view<Char> id) {
+    int arg_id = context.arg_id(id);
+    if (arg_id < 0) on_error("argument not found");
+    return arg_id;
+  }
+
+  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<OutputIt, Char>{context.out(), context.args(),
+                                              context.locale()},
+        arg));
+  }
+
+  const Char* on_format_specs(int id, const Char* begin, const Char* end) {
+    auto arg = get_arg(context, id);
+    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;
+  }
+};
+
+// A parse context with extra argument id checks. It is only used at compile
+// time because adding checks at runtime would introduce substantial overhead
+// and would be redundant since argument ids are checked when arguments are
+// retrieved anyway.
+template <typename Char, typename ErrorHandler = error_handler>
+class compile_parse_context
+    : public basic_format_parse_context<Char, ErrorHandler> {
+ private:
+  int num_args_;
+  using base = basic_format_parse_context<Char, ErrorHandler>;
+
+ public:
+  explicit FMT_CONSTEXPR compile_parse_context(
+      basic_string_view<Char> format_str, int num_args = max_value<int>(),
+      ErrorHandler eh = {})
+      : base(format_str, eh), num_args_(num_args) {}
+
+  FMT_CONSTEXPR int next_arg_id() {
+    int id = base::next_arg_id();
+    if (id >= num_args_) this->on_error("argument not found");
+    return id;
+  }
+
+  FMT_CONSTEXPR void check_arg_id(int id) {
+    base::check_arg_id(id);
+    if (id >= num_args_) this->on_error("argument not found");
+  }
+  using base::check_arg_id;
+};
+
 template <typename Char, typename ErrorHandler, typename... Args>
 class format_string_checker {
  public:
   explicit FMT_CONSTEXPR format_string_checker(
       basic_string_view<Char> format_str, ErrorHandler eh)
-      : arg_id_((std::numeric_limits<unsigned>::max)()),
-        context_(format_str, eh),
+      : context_(format_str, num_args, eh),
         parse_funcs_{&parse_format_specs<Args, parse_context_type>...} {}
 
   FMT_CONSTEXPR void on_text(const Char*, const Char*) {}
 
-  FMT_CONSTEXPR void on_arg_id() {
-    arg_id_ = context_.next_arg_id();
-    check_arg_id();
-  }
-  FMT_CONSTEXPR void on_arg_id(int id) {
-    arg_id_ = id;
-    context_.check_arg_id(id);
-    check_arg_id();
-  }
-  FMT_CONSTEXPR void on_arg_id(basic_string_view<Char>) {
+  FMT_CONSTEXPR int on_arg_id() { return context_.next_arg_id(); }
+  FMT_CONSTEXPR int on_arg_id(int id) { return context_.check_arg_id(id), id; }
+  FMT_CONSTEXPR int on_arg_id(basic_string_view<Char>) {
     on_error("compile-time checks don't support named arguments");
+    return 0;
   }
 
-  FMT_CONSTEXPR void on_replacement_field(const Char*) {}
+  FMT_CONSTEXPR void on_replacement_field(int, const Char*) {}
 
-  FMT_CONSTEXPR const Char* on_format_specs(const Char* begin, const Char*) {
+  FMT_CONSTEXPR const Char* on_format_specs(int id, const Char* begin,
+                                            const Char*) {
     advance_to(context_, begin);
-    return arg_id_ < num_args ? parse_funcs_[arg_id_](context_) : begin;
+    return id < num_args ? parse_funcs_[id](context_) : begin;
   }
 
   FMT_CONSTEXPR void on_error(const char* message) {
@@ -2542,109 +3168,104 @@ class format_string_checker {
   }
 
  private:
-  using parse_context_type = basic_parse_context<Char, ErrorHandler>;
+  using parse_context_type = compile_parse_context<Char, ErrorHandler>;
   enum { num_args = sizeof...(Args) };
 
-  FMT_CONSTEXPR void check_arg_id() {
-    if (arg_id_ >= num_args) context_.on_error("argument index out of range");
-  }
-
   // Format specifier parsing function.
   using parse_func = const Char* (*)(parse_context_type&);
 
-  unsigned arg_id_;
   parse_context_type context_;
   parse_func parse_funcs_[num_args > 0 ? num_args : 1];
 };
 
-template <typename Char, typename ErrorHandler, typename... Args>
-FMT_CONSTEXPR bool do_check_format_string(basic_string_view<Char> s,
-                                          ErrorHandler eh = ErrorHandler()) {
-  format_string_checker<Char, ErrorHandler, Args...> checker(s, eh);
-  parse_format_string<true>(s, checker);
-  return true;
+// Converts string literals to basic_string_view.
+template <typename Char, size_t N>
+FMT_CONSTEXPR basic_string_view<Char> compile_string_to_view(
+    const Char (&s)[N]) {
+  // Remove trailing null character if needed. Won't be present if this is used
+  // with raw character array (i.e. not defined as a string).
+  return {s,
+          N - ((std::char_traits<Char>::to_int_type(s[N - 1]) == 0) ? 1 : 0)};
 }
 
+// Converts string_view to basic_string_view.
+template <typename Char>
+FMT_CONSTEXPR basic_string_view<Char> compile_string_to_view(
+    const std_string_view<Char>& s) {
+  return {s.data(), s.size()};
+}
+
+#define FMT_STRING_IMPL(s, base)                                  \
+  [] {                                                            \
+    /* Use a macro-like name to avoid shadowing warnings. */      \
+    struct FMT_COMPILE_STRING : base {                            \
+      using char_type = fmt::remove_cvref_t<decltype(s[0])>;      \
+      FMT_MAYBE_UNUSED FMT_CONSTEXPR                              \
+      operator fmt::basic_string_view<char_type>() const {        \
+        return fmt::detail::compile_string_to_view<char_type>(s); \
+      }                                                           \
+    };                                                            \
+    return FMT_COMPILE_STRING();                                  \
+  }()
+
+/**
+  \rst
+  Constructs a compile-time format string from a string literal *s*.
+
+  **Example**::
+
+    // A compile-time error because 'd' is an invalid specifier for strings.
+    std::string s = fmt::format(FMT_STRING("{:d}"), "foo");
+  \endrst
+ */
+#define FMT_STRING(s) FMT_STRING_IMPL(s, fmt::compile_string)
+
 template <typename... Args, typename S,
           enable_if_t<(is_compile_string<S>::value), int>>
 void check_format_string(S format_str) {
+  FMT_CONSTEXPR_DECL auto s = to_string_view(format_str);
+  using checker = format_string_checker<typename S::char_type, error_handler,
+                                        remove_cvref_t<Args>...>;
   FMT_CONSTEXPR_DECL bool invalid_format =
-      internal::do_check_format_string<typename S::char_type,
-                                       internal::error_handler, Args...>(
-          to_string_view(format_str));
+      (parse_format_string<true>(s, checker(s, {})), true);
   (void)invalid_format;
 }
 
-template <template <typename> class Handler, typename Spec, typename Context>
-void handle_dynamic_spec(Spec& value, arg_ref<typename Context::char_type> ref,
-                         Context& ctx,
-                         const typename Context::char_type* format_str) {
+template <template <typename> class Handler, typename Context>
+void handle_dynamic_spec(int& value, arg_ref<typename Context::char_type> ref,
+                         Context& ctx) {
   switch (ref.kind) {
   case arg_id_kind::none:
     break;
   case arg_id_kind::index:
-    internal::set_dynamic_spec<Handler>(value, ctx.arg(ref.val.index),
-                                        ctx.error_handler());
+    value = detail::get_dynamic_spec<Handler>(ctx.arg(ref.val.index),
+                                              ctx.error_handler());
     break;
-  case arg_id_kind::name: {
-    const auto arg_id = ref.val.name.to_view(format_str);
-    internal::set_dynamic_spec<Handler>(value, ctx.arg(arg_id),
-                                        ctx.error_handler());
+  case arg_id_kind::name:
+    value = detail::get_dynamic_spec<Handler>(ctx.arg(ref.val.name),
+                                              ctx.error_handler());
     break;
   }
-  }
 }
-}  // namespace internal
 
-template <typename Range>
-using basic_writer FMT_DEPRECATED_ALIAS = internal::basic_writer<Range>;
-using writer FMT_DEPRECATED_ALIAS = internal::writer;
-using wwriter FMT_DEPRECATED_ALIAS =
-    internal::basic_writer<internal::buffer_range<wchar_t>>;
+using format_func = void (*)(detail::buffer<char>&, int, string_view);
 
-/** The default argument formatter. */
-template <typename Range>
-class arg_formatter : public internal::arg_formatter_base<Range> {
- private:
-  using char_type = typename Range::value_type;
-  using base = internal::arg_formatter_base<Range>;
-  using context_type = basic_format_context<typename base::iterator, char_type>;
+FMT_API void format_error_code(buffer<char>& out, int error_code,
+                               string_view message) FMT_NOEXCEPT;
 
-  context_type& ctx_;
-  basic_parse_context<char_type>* parse_ctx_;
+FMT_API void report_error(format_func func, int error_code,
+                          string_view message) FMT_NOEXCEPT;
+}  // namespace detail
 
- public:
-  using range = Range;
-  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_parse_context<char_type>* parse_ctx = nullptr,
-                         format_specs* specs = nullptr)
-      : base(Range(ctx.out()), specs, ctx.locale()),
-        ctx_(ctx),
-        parse_ctx_(parse_ctx) {}
-
-  using base::operator();
-
-  /** Formats an argument of a user-defined type. */
-  iterator operator()(typename basic_format_arg<context_type>::handle handle) {
-    handle.format(*parse_ctx_, ctx_);
-    return this->out();
-  }
-};
+template <typename OutputIt, typename Char>
+using arg_formatter FMT_DEPRECATED_ALIAS =
+    detail::arg_formatter<OutputIt, Char>;
 
 /**
  An error returned by an operating system or a language runtime,
  for example a file opening error.
 */
+FMT_CLASS_API
 class FMT_API system_error : public std::runtime_error {
  private:
   void init(int err_code, string_view format_str, format_args args);
@@ -2678,7 +3299,11 @@ class FMT_API system_error : public std::runtime_error {
       : std::runtime_error("") {
     init(error_code, message, make_format_args(args...));
   }
-  ~system_error() FMT_NOEXCEPT;
+  system_error(const system_error&) = default;
+  system_error& operator=(const system_error&) = default;
+  system_error(system_error&&) = default;
+  system_error& operator=(system_error&&) = default;
+  ~system_error() FMT_NOEXCEPT FMT_OVERRIDE;
 
   int error_code() const { return error_code_; }
 };
@@ -2699,182 +3324,14 @@ class FMT_API system_error : public std::runtime_error {
   may look like "Unknown error -1" and is platform-dependent.
   \endrst
  */
-FMT_API void format_system_error(internal::buffer<char>& out, int error_code,
-                                 fmt::string_view message) FMT_NOEXCEPT;
-
-struct float_spec_handler {
-  char type;
-  bool upper;
-  bool fixed;
-  bool as_percentage;
-  bool use_locale;
-
-  explicit float_spec_handler(char t)
-      : type(t),
-        upper(false),
-        fixed(false),
-        as_percentage(false),
-        use_locale(false) {}
-
-  void on_general() {
-    if (type == 'G') upper = true;
-  }
-
-  void on_exp() {
-    if (type == 'E') upper = true;
-  }
-
-  void on_fixed() {
-    fixed = true;
-    if (type == 'F') upper = true;
-  }
-
-  void on_percent() {
-    fixed = true;
-    as_percentage = true;
-  }
-
-  void on_hex() {
-    if (type == 'A') upper = true;
-  }
-
-  void on_num() { use_locale = true; }
-
-  FMT_NORETURN void on_error() {
-    FMT_THROW(format_error("invalid type specifier"));
-  }
-};
-
-template <typename Range>
-template <typename T, bool USE_GRISU>
-void internal::basic_writer<Range>::write_double(T value,
-                                                 const format_specs& specs) {
-  // Check type.
-  float_spec_handler handler(static_cast<char>(specs.type));
-  internal::handle_float_type_spec(handler.type, handler);
-
-  char sign = 0;
-  // Use signbit instead of value < 0 since the latter is always false for NaN.
-  if (std::signbit(value)) {
-    sign = '-';
-    value = -value;
-  } else if (specs.sign != sign::none) {
-    if (specs.sign == sign::plus)
-      sign = '+';
-    else if (specs.sign == sign::space)
-      sign = ' ';
-  }
-
-  if (!std::isfinite(value)) {
-    // Format infinity and NaN ourselves because sprintf's output is not
-    // consistent across platforms.
-    const char* str = std::isinf(value) ? (handler.upper ? "INF" : "inf")
-                                        : (handler.upper ? "NAN" : "nan");
-    return write_padded(specs,
-                        inf_or_nan_writer{sign, handler.as_percentage, str});
-  }
-
-  if (handler.as_percentage) value *= 100;
-
-  memory_buffer buffer;
-  int exp = 0;
-  int precision = specs.precision >= 0 || !specs.type ? specs.precision : 6;
-  unsigned options = handler.fixed ? internal::grisu_options::fixed : 0;
-  bool use_grisu = USE_GRISU &&
-                   (specs.type != 'a' && specs.type != 'A' &&
-                    specs.type != 'e' && specs.type != 'E') &&
-                   internal::grisu_format(static_cast<double>(value), buffer,
-                                          precision, options, exp);
-  char* decimal_point_pos = nullptr;
-  if (!use_grisu)
-    decimal_point_pos = internal::sprintf_format(value, buffer, specs);
-
-  if (handler.as_percentage) {
-    buffer.push_back('%');
-    --exp;  // Adjust decimal place position.
-  }
-  format_specs as = specs;
-  if (specs.align == align::numeric) {
-    if (sign) {
-      auto&& it = reserve(1);
-      *it++ = static_cast<char_type>(sign);
-      sign = 0;
-      if (as.width) --as.width;
-    }
-    as.align = align::right;
-  } else if (specs.align == align::none) {
-    as.align = align::right;
-  }
-  char_type decimal_point = handler.use_locale
-                                ? internal::decimal_point<char_type>(locale_)
-                                : static_cast<char_type>('.');
-  if (use_grisu) {
-    auto params = internal::gen_digits_params();
-    params.fixed = handler.fixed;
-    params.num_digits = precision;
-    params.trailing_zeros =
-        (precision != 0 && (handler.fixed || !specs.type)) || specs.alt;
-    write_padded(as, grisu_writer(sign, buffer, exp, params, decimal_point));
-  } else {
-    write_padded(as,
-                 double_writer{sign, buffer, decimal_point_pos, decimal_point});
-  }
-}
+FMT_API void format_system_error(detail::buffer<char>& out, int error_code,
+                                 string_view message) FMT_NOEXCEPT;
 
 // Reports a system error without throwing an exception.
 // Can be used to report errors from destructors.
 FMT_API void report_system_error(int error_code,
                                  string_view message) FMT_NOEXCEPT;
 
-#if FMT_USE_WINDOWS_H
-
-/** A Windows error. */
-class windows_error : public system_error {
- private:
-  FMT_API void init(int error_code, string_view format_str, format_args args);
-
- public:
-  /**
-   \rst
-   Constructs a :class:`fmt::windows_error` object with the description
-   of the form
-
-   .. parsed-literal::
-     *<message>*: *<system-message>*
-
-   where *<message>* is the formatted message and *<system-message>* is the
-   system message corresponding to the error code.
-   *error_code* is a Windows error code as given by ``GetLastError``.
-   If *error_code* is not a valid error code such as -1, the system message
-   will look like "error -1".
-
-   **Example**::
-
-     // This throws a windows_error with the description
-     //   cannot open file 'madeup': The system cannot find the file specified.
-     // or similar (system message may vary).
-     const char *filename = "madeup";
-     LPOFSTRUCT of = LPOFSTRUCT();
-     HFILE file = OpenFile(filename, &of, OF_READ);
-     if (file == HFILE_ERROR) {
-       throw fmt::windows_error(GetLastError(),
-                                "cannot open file '{}'", filename);
-     }
-   \endrst
-  */
-  template <typename... Args>
-  windows_error(int error_code, string_view message, const Args&... args) {
-    init(error_code, message, make_format_args(args...));
-  }
-};
-
-// Reports a Windows error without throwing an exception.
-// Can be used to report errors from destructors.
-FMT_API void report_windows_error(int error_code,
-                                  string_view message) FMT_NOEXCEPT;
-
-#endif
-
 /** Fast integer formatter. */
 class format_int {
  private:
@@ -2884,47 +3341,32 @@ class format_int {
   mutable char buffer_[buffer_size];
   char* str_;
 
-  // Formats value in reverse and returns a pointer to the beginning.
-  char* format_decimal(unsigned long long value) {
-    char* ptr = buffer_ + (buffer_size - 1);  // Parens to workaround MSVC bug.
-    while (value >= 100) {
-      // Integer division is slow so do it for a group of two digits instead
-      // of for every digit. The idea comes from the talk by Alexandrescu
-      // "Three Optimization Tips for C++". See speed-test for a comparison.
-      unsigned index = static_cast<unsigned>((value % 100) * 2);
-      value /= 100;
-      *--ptr = internal::data::digits[index + 1];
-      *--ptr = internal::data::digits[index];
-    }
-    if (value < 10) {
-      *--ptr = static_cast<char>('0' + value);
-      return ptr;
-    }
-    unsigned index = static_cast<unsigned>(value * 2);
-    *--ptr = internal::data::digits[index + 1];
-    *--ptr = internal::data::digits[index];
-    return ptr;
+  template <typename UInt> char* format_unsigned(UInt value) {
+    auto n = static_cast<detail::uint32_or_64_or_128_t<UInt>>(value);
+    return detail::format_decimal(buffer_, n, buffer_size - 1).begin;
   }
 
-  void format_signed(long long value) {
-    unsigned long long abs_value = static_cast<unsigned long long>(value);
+  template <typename Int> char* format_signed(Int value) {
+    auto abs_value = static_cast<detail::uint32_or_64_or_128_t<Int>>(value);
     bool negative = value < 0;
     if (negative) abs_value = 0 - abs_value;
-    str_ = format_decimal(abs_value);
-    if (negative) *--str_ = '-';
+    auto begin = format_unsigned(abs_value);
+    if (negative) *--begin = '-';
+    return begin;
   }
 
  public:
-  explicit format_int(int value) { format_signed(value); }
-  explicit format_int(long value) { format_signed(value); }
-  explicit format_int(long long value) { format_signed(value); }
-  explicit format_int(unsigned value) : str_(format_decimal(value)) {}
-  explicit format_int(unsigned long value) : str_(format_decimal(value)) {}
-  explicit format_int(unsigned long long value) : str_(format_decimal(value)) {}
+  explicit format_int(int value) : str_(format_signed(value)) {}
+  explicit format_int(long value) : str_(format_signed(value)) {}
+  explicit format_int(long long value) : str_(format_signed(value)) {}
+  explicit format_int(unsigned value) : str_(format_unsigned(value)) {}
+  explicit format_int(unsigned long value) : str_(format_unsigned(value)) {}
+  explicit format_int(unsigned long long value)
+      : str_(format_unsigned(value)) {}
 
   /** Returns the number of characters written to the output buffer. */
-  std::size_t size() const {
-    return internal::to_unsigned(buffer_ - str_ + buffer_size - 1);
+  size_t size() const {
+    return detail::to_unsigned(buffer_ - str_ + buffer_size - 1);
   }
 
   /**
@@ -2950,58 +3392,71 @@ class format_int {
   std::string str() const { return std::string(str_, size()); }
 };
 
-// A formatter specialization for the core types corresponding to internal::type
+// A formatter specialization for the core types corresponding to detail::type
 // constants.
 template <typename T, typename Char>
 struct formatter<T, Char,
-                 enable_if_t<internal::type_constant<T, Char>::value !=
-                             internal::custom_type>> {
-  FMT_CONSTEXPR formatter() : format_str_(nullptr) {}
+                 enable_if_t<detail::type_constant<T, Char>::value !=
+                             detail::type::custom_type>> {
+  FMT_CONSTEXPR formatter() = default;
 
   // Parses format specifiers stopping either at the end of the range or at the
   // terminating '}'.
   template <typename ParseContext>
   FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
-    format_str_ = ctx.begin();
-    using handler_type = internal::dynamic_specs_handler<ParseContext>;
-    auto type = internal::type_constant<T, Char>::value;
-    internal::specs_checker<handler_type> handler(handler_type(specs_, ctx),
-                                                  type);
+    using handler_type = detail::dynamic_specs_handler<ParseContext>;
+    auto type = detail::type_constant<T, Char>::value;
+    detail::specs_checker<handler_type> handler(handler_type(specs_, ctx),
+                                                type);
     auto it = parse_format_specs(ctx.begin(), ctx.end(), handler);
     auto eh = ctx.error_handler();
     switch (type) {
-    case internal::none_type:
-    case internal::named_arg_type:
+    case detail::type::none_type:
       FMT_ASSERT(false, "invalid argument type");
       break;
-    case internal::int_type:
-    case internal::uint_type:
-    case internal::long_long_type:
-    case internal::ulong_long_type:
-    case internal::bool_type:
+    case detail::type::int_type:
+    case detail::type::uint_type:
+    case detail::type::long_long_type:
+    case detail::type::ulong_long_type:
+    case detail::type::int128_type:
+    case detail::type::uint128_type:
+    case detail::type::bool_type:
       handle_int_type_spec(specs_.type,
-                           internal::int_type_checker<decltype(eh)>(eh));
+                           detail::int_type_checker<decltype(eh)>(eh));
       break;
-    case internal::char_type:
+    case detail::type::char_type:
       handle_char_specs(
-          &specs_, internal::char_specs_checker<decltype(eh)>(specs_.type, eh));
+          &specs_, detail::char_specs_checker<decltype(eh)>(specs_.type, eh));
       break;
-    case internal::double_type:
-    case internal::long_double_type:
-      handle_float_type_spec(specs_.type,
-                             internal::float_type_checker<decltype(eh)>(eh));
+    case detail::type::float_type:
+      if (detail::const_check(FMT_USE_FLOAT))
+        detail::parse_float_type_spec(specs_, eh);
+      else
+        FMT_ASSERT(false, "float support disabled");
       break;
-    case internal::cstring_type:
-      internal::handle_cstring_type_spec(
-          specs_.type, internal::cstring_type_checker<decltype(eh)>(eh));
+    case detail::type::double_type:
+      if (detail::const_check(FMT_USE_DOUBLE))
+        detail::parse_float_type_spec(specs_, eh);
+      else
+        FMT_ASSERT(false, "double support disabled");
       break;
-    case internal::string_type:
-      internal::check_string_type_spec(specs_.type, eh);
+    case detail::type::long_double_type:
+      if (detail::const_check(FMT_USE_LONG_DOUBLE))
+        detail::parse_float_type_spec(specs_, eh);
+      else
+        FMT_ASSERT(false, "long double support disabled");
       break;
-    case internal::pointer_type:
-      internal::check_pointer_type_spec(specs_.type, eh);
+    case detail::type::cstring_type:
+      detail::handle_cstring_type_spec(
+          specs_.type, detail::cstring_type_checker<decltype(eh)>(eh));
       break;
-    case internal::custom_type:
+    case detail::type::string_type:
+      detail::check_string_type_spec(specs_.type, eh);
+      break;
+    case detail::type::pointer_type:
+      detail::check_pointer_type_spec(specs_.type, eh);
+      break;
+    case detail::type::custom_type:
       // Custom format specifiers should be checked in parse functions of
       // formatter specializations.
       break;
@@ -3011,27 +3466,25 @@ struct formatter<T, Char,
 
   template <typename FormatContext>
   auto format(const T& val, FormatContext& ctx) -> decltype(ctx.out()) {
-    internal::handle_dynamic_spec<internal::width_checker>(
-        specs_.width, specs_.width_ref, ctx, format_str_);
-    internal::handle_dynamic_spec<internal::precision_checker>(
-        specs_.precision, specs_.precision_ref, ctx, format_str_);
-    using range_type =
-        internal::output_range<typename FormatContext::iterator,
-                               typename FormatContext::char_type>;
-    return visit_format_arg(arg_formatter<range_type>(ctx, nullptr, &specs_),
-                            internal::make_arg<FormatContext>(val));
+    detail::handle_dynamic_spec<detail::width_checker>(specs_.width,
+                                                       specs_.width_ref, ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs_.precision, specs_.precision_ref, ctx);
+    using af = detail::arg_formatter<typename FormatContext::iterator,
+                                     typename FormatContext::char_type>;
+    return visit_format_arg(af(ctx, nullptr, &specs_),
+                            detail::make_arg<FormatContext>(val));
   }
 
  private:
-  internal::dynamic_format_specs<Char> specs_;
-  const Char* format_str_;
+  detail::dynamic_format_specs<Char> specs_;
 };
 
 #define FMT_FORMAT_AS(Type, Base)                                             \
   template <typename Char>                                                    \
   struct formatter<Type, Char> : formatter<Base, Char> {                      \
     template <typename FormatContext>                                         \
-    auto format(const Type& val, FormatContext& ctx) -> decltype(ctx.out()) { \
+    auto format(Type const& val, FormatContext& ctx) -> decltype(ctx.out()) { \
       return formatter<Base, Char>::format(val, ctx);                         \
     }                                                                         \
   }
@@ -3042,11 +3495,10 @@ FMT_FORMAT_AS(short, int);
 FMT_FORMAT_AS(unsigned short, unsigned);
 FMT_FORMAT_AS(long, long long);
 FMT_FORMAT_AS(unsigned long, unsigned long long);
-FMT_FORMAT_AS(float, double);
 FMT_FORMAT_AS(Char*, const Char*);
 FMT_FORMAT_AS(std::basic_string<Char>, basic_string_view<Char>);
 FMT_FORMAT_AS(std::nullptr_t, const void*);
-FMT_FORMAT_AS(internal::std_string_view<Char>, basic_string_view<Char>);
+FMT_FORMAT_AS(detail::std_string_view<Char>, basic_string_view<Char>);
 
 template <typename Char>
 struct formatter<void*, Char> : formatter<const void*, Char> {
@@ -3070,13 +3522,15 @@ 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 {
  private:
-  struct null_handler : internal::error_handler {
+  struct null_handler : detail::error_handler {
     void on_align(align_t) {}
     void on_plus() {}
     void on_minus() {}
@@ -3089,16 +3543,15 @@ template <typename Char = char> class dynamic_formatter {
   auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
     format_str_ = ctx.begin();
     // Checks are deferred to formatting time when the argument type is known.
-    internal::dynamic_specs_handler<ParseContext> handler(specs_, ctx);
+    detail::dynamic_specs_handler<ParseContext> handler(specs_, ctx);
     return parse_format_specs(ctx.begin(), ctx.end(), handler);
   }
 
   template <typename T, typename FormatContext>
   auto format(const T& val, FormatContext& ctx) -> decltype(ctx.out()) {
     handle_specs(ctx);
-    internal::specs_checker<null_handler> checker(
-        null_handler(),
-        internal::mapped_type_constant<T, FormatContext>::value);
+    detail::specs_checker<null_handler> checker(
+        null_handler(), detail::mapped_type_constant<T, FormatContext>::value);
     checker.on_align(specs_.align);
     switch (specs_.sign) {
     case sign::none:
@@ -3115,111 +3568,40 @@ template <typename Char = char> class dynamic_formatter {
     }
     if (specs_.alt) checker.on_hash();
     if (specs_.precision >= 0) checker.end_precision();
-    using range = internal::output_range<typename FormatContext::iterator,
-                                         typename FormatContext::char_type>;
-    visit_format_arg(arg_formatter<range>(ctx, nullptr, &specs_),
-                     internal::make_arg<FormatContext>(val));
+    using af = detail::arg_formatter<typename FormatContext::iterator,
+                                     typename FormatContext::char_type>;
+    visit_format_arg(af(ctx, nullptr, &specs_),
+                     detail::make_arg<FormatContext>(val));
     return ctx.out();
   }
 
  private:
   template <typename Context> void handle_specs(Context& ctx) {
-    internal::handle_dynamic_spec<internal::width_checker>(
-        specs_.width, specs_.width_ref, ctx, format_str_);
-    internal::handle_dynamic_spec<internal::precision_checker>(
-        specs_.precision, specs_.precision_ref, ctx, format_str_);
+    detail::handle_dynamic_spec<detail::width_checker>(specs_.width,
+                                                       specs_.width_ref, ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs_.precision, specs_.precision_ref, ctx);
   }
 
-  internal::dynamic_format_specs<Char> specs_;
+  detail::dynamic_format_specs<Char> specs_;
   const Char* format_str_;
 };
 
-template <typename Range, typename Char>
-typename basic_format_context<Range, Char>::format_arg
-basic_format_context<Range, Char>::arg(basic_string_view<char_type> name) {
-  map_.init(args_);
-  format_arg arg = map_.find(name);
-  if (arg.type() == internal::none_type) this->on_error("argument not found");
-  return arg;
-}
-
 template <typename Char, typename ErrorHandler>
-FMT_CONSTEXPR void advance_to(basic_parse_context<Char, ErrorHandler>& ctx,
-                              const Char* p) {
+FMT_CONSTEXPR void advance_to(
+    basic_format_parse_context<Char, ErrorHandler>& ctx, const Char* p) {
   ctx.advance_to(ctx.begin() + (p - &*ctx.begin()));
 }
 
-template <typename ArgFormatter, typename Char, typename Context>
-struct format_handler : internal::error_handler {
-  using range = typename ArgFormatter::range;
+/**
+  \rst
+  Converts ``p`` to ``const void*`` for pointer formatting.
 
-  format_handler(range r, basic_string_view<Char> str,
-                 basic_format_args<Context> format_args,
-                 internal::locale_ref loc)
-      : parse_context(str), context(r.begin(), format_args, loc) {}
+  **Example**::
 
-  void on_text(const Char* begin, const Char* end) {
-    auto size = internal::to_unsigned(end - begin);
-    auto out = context.out();
-    auto&& it = internal::reserve(out, size);
-    it = std::copy_n(begin, size, it);
-    context.advance_to(out);
-  }
-
-  void get_arg(int id) { arg = internal::get_arg(context, id); }
-
-  void on_arg_id() { get_arg(parse_context.next_arg_id()); }
-  void on_arg_id(int id) {
-    parse_context.check_arg_id(id);
-    get_arg(id);
-  }
-  void on_arg_id(basic_string_view<Char> id) { arg = context.arg(id); }
-
-  void on_replacement_field(const Char* p) {
-    advance_to(parse_context, p);
-    internal::custom_formatter<Context> f(parse_context, context);
-    if (!visit_format_arg(f, arg))
-      context.advance_to(
-          visit_format_arg(ArgFormatter(context, &parse_context), arg));
-  }
-
-  const Char* on_format_specs(const Char* begin, const Char* end) {
-    advance_to(parse_context, begin);
-    internal::custom_formatter<Context> f(parse_context, context);
-    if (visit_format_arg(f, arg)) return parse_context.begin();
-    basic_format_specs<Char> specs;
-    using internal::specs_handler;
-    using parse_context_t = basic_parse_context<Char>;
-    internal::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));
-    return begin;
-  }
-
-  basic_parse_context<Char> parse_context;
-  Context context;
-  basic_format_arg<Context> arg;
-};
-
-/** Formats arguments and writes the output to the range. */
-template <typename ArgFormatter, typename Char, typename Context>
-typename Context::iterator vformat_to(
-    typename ArgFormatter::range out, basic_string_view<Char> format_str,
-    basic_format_args<Context> args,
-    internal::locale_ref loc = internal::locale_ref()) {
-  format_handler<ArgFormatter, Char, Context> h(out, format_str, args, loc);
-  internal::parse_format_string<false>(format_str, h);
-  return h.context.out();
-}
-
-// Casts ``p`` to ``const void*`` for pointer formatting.
-// Example:
-//   auto s = format("{}", ptr(p));
+    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();
@@ -3228,19 +3610,55 @@ template <typename T> inline const void* ptr(const std::shared_ptr<T>& p) {
   return p.get();
 }
 
-template <typename It, typename Char> struct arg_join : internal::view {
-  It begin;
-  It end;
-  basic_string_view<Char> sep;
+class bytes {
+ private:
+  string_view data_;
+  friend struct formatter<bytes>;
 
-  arg_join(It b, It e, basic_string_view<Char> s) : begin(b), end(e), sep(s) {}
+ public:
+  explicit bytes(string_view data) : data_(data) {}
 };
 
-template <typename It, typename Char>
-struct formatter<arg_join<It, Char>, Char>
+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>;
+    detail::specs_checker<handler_type> handler(handler_type(specs_, ctx),
+                                                detail::type::string_type);
+    auto it = parse_format_specs(ctx.begin(), ctx.end(), handler);
+    detail::check_string_type_spec(specs_.type, ctx.error_handler());
+    return it;
+  }
+
+  template <typename FormatContext>
+  auto format(bytes b, FormatContext& ctx) -> decltype(ctx.out()) {
+    detail::handle_dynamic_spec<detail::width_checker>(specs_.width,
+                                                       specs_.width_ref, ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs_.precision, specs_.precision_ref, ctx);
+    return detail::write_bytes(ctx.out(), b.data_, specs_);
+  }
+};
+
+template <typename It, typename Sentinel, typename Char>
+struct arg_join : detail::view {
+  It begin;
+  Sentinel end;
+  basic_string_view<Char> sep;
+
+  arg_join(It b, Sentinel e, basic_string_view<Char> s)
+      : begin(b), end(e), sep(s) {}
+};
+
+template <typename It, typename Sentinel, typename Char>
+struct formatter<arg_join<It, Sentinel, Char>, Char>
     : formatter<typename std::iterator_traits<It>::value_type, Char> {
   template <typename FormatContext>
-  auto format(const arg_join<It, Char>& value, FormatContext& ctx)
+  auto format(const arg_join<It, Sentinel, Char>& value, FormatContext& ctx)
       -> decltype(ctx.out()) {
     using base = formatter<typename std::iterator_traits<It>::value_type, Char>;
     auto it = value.begin;
@@ -3261,13 +3679,13 @@ struct formatter<arg_join<It, Char>, Char>
   Returns an object that formats the iterator range `[begin, end)` with elements
   separated by `sep`.
  */
-template <typename It>
-arg_join<It, char> join(It begin, It end, string_view sep) {
+template <typename It, typename Sentinel>
+arg_join<It, Sentinel, char> join(It begin, Sentinel end, string_view sep) {
   return {begin, end, sep};
 }
 
-template <typename It>
-arg_join<It, wchar_t> join(It begin, It end, wstring_view sep) {
+template <typename It, typename Sentinel>
+arg_join<It, Sentinel, wchar_t> join(It begin, Sentinel end, wstring_view sep) {
   return {begin, end, sep};
 }
 
@@ -3280,24 +3698,28 @@ arg_join<It, wchar_t> join(It begin, It end, wstring_view sep) {
     std::vector<int> v = {1, 2, 3};
     fmt::print("{}", fmt::join(v, ", "));
     // Output: "1, 2, 3"
+
+  ``fmt::join`` applies passed format specifiers to the range elements::
+
+    fmt::print("{:02}", fmt::join(v, ", "));
+    // Output: "01, 02, 03"
   \endrst
  */
 template <typename Range>
-arg_join<internal::iterator_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<internal::iterator_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);
 }
 
 /**
   \rst
   Converts *value* to ``std::string`` using the default format for type *T*.
-  It doesn't support user-defined types with custom formatters.
 
   **Example**::
 
@@ -3306,8 +3728,21 @@ arg_join<internal::iterator_t<const Range>, wchar_t> join(const Range& range,
     std::string answer = fmt::to_string(42);
   \endrst
  */
-template <typename T> inline std::string to_string(const T& value) {
-  return format("{}", value);
+template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
+inline std::string to_string(const T& value) {
+  std::string result;
+  detail::write<char>(std::back_inserter(result), value);
+  return result;
+}
+
+template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+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 ? static_cast<unsigned>(max_size) : 5];
+  char* begin = buffer;
+  return std::string(begin, detail::write<char>(begin, value));
 }
 
 /**
@@ -3317,36 +3752,76 @@ template <typename T> inline std::wstring to_wstring(const T& value) {
   return format(L"{}", value);
 }
 
-template <typename Char, std::size_t SIZE>
+template <typename Char, size_t SIZE>
 std::basic_string<Char> to_string(const basic_memory_buffer<Char, SIZE>& buf) {
-  return std::basic_string<Char>(buf.data(), buf.size());
+  auto size = buf.size();
+  detail::assume(size < std::basic_string<Char>().max_size());
+  return std::basic_string<Char>(buf.data(), size);
 }
 
 template <typename Char>
-typename buffer_context<Char>::iterator internal::vformat_to(
-    internal::buffer<Char>& buf, basic_string_view<Char> format_str,
-    basic_format_args<buffer_context<Char>> args) {
-  using range = buffer_range<Char>;
-  return vformat_to<arg_formatter<range>>(buf, to_string_view(format_str),
-                                          args);
+void detail::vformat_to(
+    detail::buffer<Char>& buf, basic_string_view<Char> format_str,
+    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 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);
+extern template FMT_API wchar_t thousands_sep_impl<wchar_t>(locale_ref loc);
+extern template FMT_API char decimal_point_impl(locale_ref loc);
+extern template FMT_API wchar_t decimal_point_impl(locale_ref loc);
+extern template int format_float<double>(double value, int precision,
+                                         float_specs specs, buffer<char>& buf);
+extern template int format_float<long double>(long double value, int precision,
+                                              float_specs specs,
+                                              buffer<char>& buf);
+int snprintf_float(float value, int precision, float_specs specs,
+                   buffer<char>& buf) = delete;
+extern template int snprintf_float<double>(double value, int precision,
+                                           float_specs specs,
+                                           buffer<char>& buf);
+extern template int snprintf_float<long double>(long double value,
+                                                int precision,
+                                                float_specs specs,
+                                                buffer<char>& buf);
+}  // namespace detail
+#endif
+
 template <typename S, typename Char = char_t<S>,
-          FMT_ENABLE_IF(internal::is_string<S>::value)>
-inline typename buffer_context<Char>::iterator vformat_to(
-    internal::buffer<Char>& buf, const S& format_str,
-    basic_format_args<buffer_context<Char>> args) {
-  return internal::vformat_to(buf, to_string_view(format_str), args);
+          FMT_ENABLE_IF(detail::is_string<S>::value)>
+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);
 }
 
-template <typename S, typename... Args, std::size_t SIZE = inline_buffer_size,
-          typename Char = enable_if_t<internal::is_string<S>::value, char_t<S>>>
+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) {
-  internal::check_format_string<Args...>(format_str);
-  using context = buffer_context<Char>;
-  return internal::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>
@@ -3355,123 +3830,53 @@ 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(
-              internal::is_output_iterator<OutputIt>::value &&
-              !internal::is_contiguous_back_insert_iterator<OutputIt>::value)>
-inline OutputIt vformat_to(OutputIt out, const S& format_str,
-                           format_args_t<OutputIt, char_t<S>> args) {
-  using range = internal::output_range<OutputIt, char_t<S>>;
-  return vformat_to<arg_formatter<range>>(range(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(
-              internal::is_output_iterator<OutputIt>::value &&
-              !internal::is_contiguous_back_insert_iterator<OutputIt>::value &&
-              internal::is_string<S>::value)>
-inline OutputIt format_to(OutputIt out, const S& format_str, Args&&... args) {
-  internal::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. */
-  std::size_t size;
-};
+template <typename OutputIt, typename Char = typename OutputIt::value_type>
+using format_to_n_context FMT_DEPRECATED_ALIAS = buffer_context<Char>;
 
 template <typename OutputIt, typename Char = typename OutputIt::value_type>
-using format_to_n_context =
-    format_context_t<fmt::internal::truncating_iterator<OutputIt>, 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...);
+  return format_arg_store<buffer_context<Char>, Args...>(args...);
 }
 
-template <typename OutputIt, typename Char, typename... Args,
-          FMT_ENABLE_IF(internal::is_output_iterator<OutputIt>::value)>
-inline format_to_n_result<OutputIt> vformat_to_n(
-    OutputIt out, std::size_t n, basic_string_view<Char> format_str,
-    format_to_n_args<OutputIt, Char> args) {
-  auto it = vformat_to(internal::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(internal::is_string<S>::value&&
-                            internal::is_output_iterator<OutputIt>::value)>
-inline format_to_n_result<OutputIt> format_to_n(OutputIt out, std::size_t n,
-                                                const S& format_str,
-                                                const Args&... args) {
-  internal::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...)});
-}
-
-template <typename Char>
-inline std::basic_string<Char> internal::vformat(
+template <typename Char, enable_if_t<(!std::is_same<Char, char>::value), int>>
+std::basic_string<Char> detail::vformat(
     basic_string_view<Char> format_str,
-    basic_format_args<buffer_context<Char>> args) {
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
   basic_memory_buffer<Char> buffer;
-  internal::vformat_to(buffer, format_str, args);
-  return fmt::to_string(buffer);
+  detail::vformat_to(buffer, format_str, args);
+  return to_string(buffer);
 }
 
-/**
-  Returns the number of characters in the output of
-  ``format(format_str, args...)``.
- */
-template <typename... Args>
-inline std::size_t formatted_size(string_view format_str, const Args&... args) {
-  auto it = format_to(internal::counting_iterator<char>(), format_str, args...);
-  return it.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) {
+  wmemory_buffer buffer;
+  detail::vformat_to(buffer, format_str, args);
+  buffer.push_back(L'\0');
+  if (std::fputws(buffer.data(), f) == -1)
+    FMT_THROW(system_error(errno, "cannot write to file"));
+}
+
+template <typename Char, FMT_ENABLE_IF(std::is_same<Char, wchar_t>::value)>
+void vprint(basic_string_view<Char> format_str, wformat_args args) {
+  vprint(stdout, format_str, args);
 }
 
 #if FMT_USE_USER_DEFINED_LITERALS
-namespace internal {
+namespace detail {
 
 #  if FMT_USE_UDL_TEMPLATE
 template <typename Char, Char... CHARS> class udl_formatter {
  public:
   template <typename... Args>
   std::basic_string<Char> operator()(Args&&... args) const {
-    FMT_CONSTEXPR_DECL Char s[] = {CHARS..., '\0'};
-    FMT_CONSTEXPR_DECL bool invalid_format =
-        do_check_format_string<Char, error_handler, Args...>(
-            basic_string_view<Char>(s, sizeof...(CHARS)));
-    (void)invalid_format;
-    return format(s, std::forward<Args>(args)...);
+    static FMT_CONSTEXPR_DECL Char s[] = {CHARS..., '\0'};
+    return format(FMT_STRING(s), std::forward<Args>(args)...);
   }
 };
 #  else
@@ -3486,23 +3891,23 @@ template <typename Char> struct udl_formatter {
 #  endif  // FMT_USE_UDL_TEMPLATE
 
 template <typename Char> struct udl_arg {
-  basic_string_view<Char> str;
+  const Char* str;
 
-  template <typename T> named_arg<T, Char> operator=(T&& value) const {
+  template <typename T> named_arg<Char, T> operator=(T&& value) const {
     return {str, std::forward<T>(value)};
   }
 };
-
-}  // namespace internal
+}  // namespace detail
 
 inline namespace literals {
 #  if FMT_USE_UDL_TEMPLATE
 #    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wpedantic"
 #    if FMT_CLANG_VERSION
 #      pragma GCC diagnostic ignored "-Wgnu-string-literal-operator-template"
 #    endif
 template <typename Char, Char... CHARS>
-FMT_CONSTEXPR internal::udl_formatter<Char, CHARS...> operator""_format() {
+FMT_CONSTEXPR detail::udl_formatter<Char, CHARS...> operator""_format() {
   return {};
 }
 #    pragma GCC diagnostic pop
@@ -3517,12 +3922,12 @@ FMT_CONSTEXPR internal::udl_formatter<Char, CHARS...> operator""_format() {
     std::string message = "The answer is {}"_format(42);
   \endrst
  */
-FMT_CONSTEXPR internal::udl_formatter<char> operator"" _format(const char* s,
-                                                               std::size_t n) {
+FMT_CONSTEXPR detail::udl_formatter<char> operator"" _format(const char* s,
+                                                             size_t n) {
   return {{s, n}};
 }
-FMT_CONSTEXPR internal::udl_formatter<wchar_t> operator"" _format(
-    const wchar_t* s, std::size_t n) {
+FMT_CONSTEXPR detail::udl_formatter<wchar_t> operator"" _format(
+    const wchar_t* s, size_t n) {
   return {{s, n}};
 }
 #  endif  // FMT_USE_UDL_TEMPLATE
@@ -3537,61 +3942,16 @@ FMT_CONSTEXPR internal::udl_formatter<wchar_t> operator"" _format(
     fmt::print("Elapsed time: {s:.2f} seconds", "s"_a=1.23);
   \endrst
  */
-FMT_CONSTEXPR internal::udl_arg<char> operator"" _a(const char* s,
-                                                    std::size_t n) {
-  return {{s, n}};
+FMT_CONSTEXPR detail::udl_arg<char> operator"" _a(const char* s, size_t) {
+  return {s};
 }
-FMT_CONSTEXPR internal::udl_arg<wchar_t> operator"" _a(const wchar_t* s,
-                                                       std::size_t n) {
-  return {{s, n}};
+FMT_CONSTEXPR detail::udl_arg<wchar_t> operator"" _a(const wchar_t* s, size_t) {
+  return {s};
 }
 }  // namespace literals
 #endif  // FMT_USE_USER_DEFINED_LITERALS
 FMT_END_NAMESPACE
 
-/**
-  \rst
-  Constructs a compile-time format string.
-
-  **Example**::
-
-    // A compile-time error because 'd' is an invalid specifier for strings.
-    std::string s = format(FMT_STRING("{:d}"), "foo");
-  \endrst
- */
-#define FMT_STRING(s)                                                    \
-  [] {                                                                   \
-    struct str : fmt::compile_string {                                   \
-      using char_type = typename std::remove_cv<std::remove_pointer<     \
-          typename std::decay<decltype(s)>::type>::type>::type;          \
-      FMT_CONSTEXPR operator fmt::basic_string_view<char_type>() const { \
-        return {s, sizeof(s) / sizeof(char_type) - 1};                   \
-      }                                                                  \
-    } result;                                                            \
-    /* Suppress Qt Creator warning about unused operator. */             \
-    (void)static_cast<fmt::basic_string_view<typename str::char_type>>(  \
-        result);                                                         \
-    return result;                                                       \
-  }()
-
-#if defined(FMT_STRING_ALIAS) && FMT_STRING_ALIAS
-/**
-  \rst
-  Constructs a compile-time format string. This macro is disabled by default to
-  prevent potential name collisions. To enable it define ``FMT_STRING_ALIAS`` to
-  1 before including ``fmt/format.h``.
-
-  **Example**::
-
-    #define FMT_STRING_ALIAS 1
-    #include <fmt/format.h>
-    // A compile-time error because 'd' is an invalid specifier for strings.
-    std::string s = format(fmt("{:d}"), "foo");
-  \endrst
- */
-#  define fmt(s) FMT_STRING(s)
-#endif
-
 #ifdef FMT_HEADER_ONLY
 #  define FMT_FUNC inline
 #  include "format-inl.h"
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/locale.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/locale.h
index 7c13656e4..7301bf92a 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/locale.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/locale.h
@@ -9,67 +9,54 @@
 #define FMT_LOCALE_H_
 
 #include <locale>
+
 #include "format.h"
 
 FMT_BEGIN_NAMESPACE
 
-namespace internal {
+namespace detail {
 template <typename Char>
-typename buffer_context<Char>::iterator vformat_to(
-    const std::locale& loc, buffer<Char>& buf,
-    basic_string_view<Char> format_str,
-    basic_format_args<buffer_context<Char>> args) {
-  using range = buffer_range<Char>;
-  return vformat_to<arg_formatter<range>>(buf, to_string_view(format_str), args,
-                                          internal::locale_ref(loc));
-}
-
-template <typename Char>
-std::basic_string<Char> vformat(const std::locale& loc,
-                                basic_string_view<Char> format_str,
-                                basic_format_args<buffer_context<Char>> args) {
+std::basic_string<Char> vformat(
+    const std::locale& loc, basic_string_view<Char> format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
   basic_memory_buffer<Char> buffer;
-  internal::vformat_to(loc, buffer, format_str, args);
+  detail::vformat_to(buffer, format_str, args, detail::locale_ref(loc));
   return fmt::to_string(buffer);
 }
-}  // namespace internal
+}  // namespace detail
 
 template <typename S, typename Char = char_t<S>>
 inline std::basic_string<Char> vformat(
     const std::locale& loc, const S& format_str,
-    basic_format_args<buffer_context<Char>> args) {
-  return internal::vformat(loc, to_string_view(format_str), args);
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  return detail::vformat(loc, to_string_view(format_str), args);
 }
 
 template <typename S, typename... Args, typename Char = char_t<S>>
 inline std::basic_string<Char> format(const std::locale& loc,
                                       const S& format_str, Args&&... args) {
-  return internal::vformat(
-      loc, to_string_view(format_str),
-      {internal::make_args_checked<Args...>(format_str, args...)});
+  return detail::vformat(loc, to_string_view(format_str),
+                         fmt::make_args_checked<Args...>(format_str, args...));
 }
 
 template <typename S, typename OutputIt, typename... Args,
-          typename Char = enable_if_t<
-              internal::is_output_iterator<OutputIt>::value, char_t<S>>>
-inline OutputIt vformat_to(OutputIt out, const std::locale& loc,
-                           const S& format_str,
-                           format_args_t<OutputIt, Char> args) {
-  using range = internal::output_range<OutputIt, Char>;
-  return vformat_to<arg_formatter<range>>(
-      range(out), to_string_view(format_str), args, internal::locale_ref(loc));
+          typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value)>
+inline OutputIt vformat_to(
+    OutputIt out, const std::locale& loc, const S& format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  decltype(detail::get_buffer<Char>(out)) buf(detail::get_buffer_init(out));
+  vformat_to(buf, to_string_view(format_str), args, detail::locale_ref(loc));
+  return detail::get_iterator(buf);
 }
 
 template <typename OutputIt, typename S, typename... Args,
-          FMT_ENABLE_IF(internal::is_output_iterator<OutputIt>::value&&
-                            internal::is_string<S>::value)>
-inline OutputIt format_to(OutputIt out, const std::locale& loc,
-                          const S& format_str, Args&&... args) {
-  internal::check_format_string<Args...>(format_str);
-  using context = format_context_t<OutputIt, char_t<S>>;
-  format_arg_store<context, Args...> as{args...};
-  return vformat_to(out, loc, to_string_view(format_str),
-                    basic_format_args<context>(as));
+          bool enable = detail::is_output_iterator<OutputIt, char_t<S>>::value>
+inline auto format_to(OutputIt out, const std::locale& loc,
+                      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, loc, to_string_view(format_str), vargs);
 }
 
 FMT_END_NAMESPACE
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/os.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/os.h
new file mode 100644
index 000000000..d44ea0c90
--- /dev/null
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/os.h
@@ -0,0 +1,480 @@
+// Formatting library for C++ - optional OS-specific functionality
+//
+// Copyright (c) 2012 - present, Victor Zverovich
+// All rights reserved.
+//
+// For the license information refer to format.h.
+
+#ifndef FMT_OS_H_
+#define FMT_OS_H_
+
+#if defined(__MINGW32__) || defined(__CYGWIN__)
+// Workaround MinGW bug https://sourceforge.net/p/mingw/bugs/2024/.
+#  undef __STRICT_ANSI__
+#endif
+
+#include <cerrno>
+#include <clocale>  // for locale_t
+#include <cstddef>
+#include <cstdio>
+#include <cstdlib>  // for strtod_l
+
+#if defined __APPLE__ || defined(__FreeBSD__)
+#  include <xlocale.h>  // for LC_NUMERIC_MASK on OS X
+#endif
+
+#include "format.h"
+
+// UWP doesn't provide _pipe.
+#if FMT_HAS_INCLUDE("winapifamily.h")
+#  include <winapifamily.h>
+#endif
+#if (FMT_HAS_INCLUDE(<fcntl.h>) || defined(__APPLE__) || \
+     defined(__linux__)) &&                              \
+    (!defined(WINAPI_FAMILY) || (WINAPI_FAMILY == WINAPI_FAMILY_DESKTOP_APP))
+#  include <fcntl.h>  // for O_RDONLY
+#  define FMT_USE_FCNTL 1
+#else
+#  define FMT_USE_FCNTL 0
+#endif
+
+#ifndef FMT_POSIX
+#  if defined(_WIN32) && !defined(__MINGW32__)
+// Fix warnings about deprecated symbols.
+#    define FMT_POSIX(call) _##call
+#  else
+#    define FMT_POSIX(call) call
+#  endif
+#endif
+
+// Calls to system functions are wrapped in FMT_SYSTEM for testability.
+#ifdef FMT_SYSTEM
+#  define FMT_POSIX_CALL(call) FMT_SYSTEM(call)
+#else
+#  define FMT_SYSTEM(call) ::call
+#  ifdef _WIN32
+// Fix warnings about deprecated symbols.
+#    define FMT_POSIX_CALL(call) ::_##call
+#  else
+#    define FMT_POSIX_CALL(call) ::call
+#  endif
+#endif
+
+// Retries the expression while it evaluates to error_result and errno
+// equals to EINTR.
+#ifndef _WIN32
+#  define FMT_RETRY_VAL(result, expression, error_result) \
+    do {                                                  \
+      (result) = (expression);                            \
+    } while ((result) == (error_result) && errno == EINTR)
+#else
+#  define FMT_RETRY_VAL(result, expression, error_result) result = (expression)
+#endif
+
+#define FMT_RETRY(result, expression) FMT_RETRY_VAL(result, expression, -1)
+
+FMT_BEGIN_NAMESPACE
+
+/**
+  \rst
+  A reference to a null-terminated string. It can be constructed from a C
+  string or ``std::string``.
+
+  You can use one of the following type aliases for common character types:
+
+  +---------------+-----------------------------+
+  | Type          | Definition                  |
+  +===============+=============================+
+  | cstring_view  | basic_cstring_view<char>    |
+  +---------------+-----------------------------+
+  | wcstring_view | basic_cstring_view<wchar_t> |
+  +---------------+-----------------------------+
+
+  This class is most useful as a parameter type to allow passing
+  different types of strings to a function, for example::
+
+    template <typename... Args>
+    std::string format(cstring_view format_str, const Args & ... args);
+
+    format("{}", 42);
+    format(std::string("{}"), 42);
+  \endrst
+ */
+template <typename Char> class basic_cstring_view {
+ private:
+  const Char* data_;
+
+ public:
+  /** Constructs a string reference object from a C string. */
+  basic_cstring_view(const Char* s) : data_(s) {}
+
+  /**
+    \rst
+    Constructs a string reference from an ``std::string`` object.
+    \endrst
+   */
+  basic_cstring_view(const std::basic_string<Char>& s) : data_(s.c_str()) {}
+
+  /** Returns the pointer to a C string. */
+  const Char* c_str() const { return data_; }
+};
+
+using cstring_view = basic_cstring_view<char>;
+using wcstring_view = basic_cstring_view<wchar_t>;
+
+// An error code.
+class error_code {
+ private:
+  int value_;
+
+ public:
+  explicit error_code(int value = 0) FMT_NOEXCEPT : value_(value) {}
+
+  int get() const FMT_NOEXCEPT { return value_; }
+};
+
+#ifdef _WIN32
+namespace detail {
+// A converter from UTF-16 to UTF-8.
+// It is only provided for Windows since other systems support UTF-8 natively.
+class utf16_to_utf8 {
+ private:
+  memory_buffer buffer_;
+
+ public:
+  utf16_to_utf8() {}
+  FMT_API explicit utf16_to_utf8(wstring_view s);
+  operator string_view() const { return string_view(&buffer_[0], size()); }
+  size_t size() const { return buffer_.size() - 1; }
+  const char* c_str() const { return &buffer_[0]; }
+  std::string str() const { return std::string(&buffer_[0], size()); }
+
+  // Performs conversion returning a system error code instead of
+  // throwing exception on conversion error. This method may still throw
+  // in case of memory allocation error.
+  FMT_API int convert(wstring_view s);
+};
+
+FMT_API void format_windows_error(buffer<char>& out, int error_code,
+                                  string_view message) FMT_NOEXCEPT;
+}  // namespace detail
+
+/** A Windows error. */
+class windows_error : public system_error {
+ private:
+  FMT_API void init(int error_code, string_view format_str, format_args args);
+
+ public:
+  /**
+   \rst
+   Constructs a :class:`fmt::windows_error` object with the description
+   of the form
+
+   .. parsed-literal::
+     *<message>*: *<system-message>*
+
+   where *<message>* is the formatted message and *<system-message>* is the
+   system message corresponding to the error code.
+   *error_code* is a Windows error code as given by ``GetLastError``.
+   If *error_code* is not a valid error code such as -1, the system message
+   will look like "error -1".
+
+   **Example**::
+
+     // This throws a windows_error with the description
+     //   cannot open file 'madeup': The system cannot find the file specified.
+     // or similar (system message may vary).
+     const char *filename = "madeup";
+     LPOFSTRUCT of = LPOFSTRUCT();
+     HFILE file = OpenFile(filename, &of, OF_READ);
+     if (file == HFILE_ERROR) {
+       throw fmt::windows_error(GetLastError(),
+                                "cannot open file '{}'", filename);
+     }
+   \endrst
+  */
+  template <typename... Args>
+  windows_error(int error_code, string_view message, const Args&... args) {
+    init(error_code, message, make_format_args(args...));
+  }
+};
+
+// Reports a Windows error without throwing an exception.
+// Can be used to report errors from destructors.
+FMT_API void report_windows_error(int error_code,
+                                  string_view message) FMT_NOEXCEPT;
+#endif  // _WIN32
+
+// A buffered file.
+class buffered_file {
+ private:
+  FILE* file_;
+
+  friend class file;
+
+  explicit buffered_file(FILE* f) : file_(f) {}
+
+ public:
+  buffered_file(const buffered_file&) = delete;
+  void operator=(const buffered_file&) = delete;
+
+  // Constructs a buffered_file object which doesn't represent any file.
+  buffered_file() FMT_NOEXCEPT : file_(nullptr) {}
+
+  // Destroys the object closing the file it represents if any.
+  FMT_API ~buffered_file() FMT_NOEXCEPT;
+
+ public:
+  buffered_file(buffered_file&& other) FMT_NOEXCEPT : file_(other.file_) {
+    other.file_ = nullptr;
+  }
+
+  buffered_file& operator=(buffered_file&& other) {
+    close();
+    file_ = other.file_;
+    other.file_ = nullptr;
+    return *this;
+  }
+
+  // Opens a file.
+  FMT_API buffered_file(cstring_view filename, cstring_view mode);
+
+  // Closes the file.
+  FMT_API void close();
+
+  // Returns the pointer to a FILE object representing this file.
+  FILE* get() const FMT_NOEXCEPT { return file_; }
+
+  // We place parentheses around fileno to workaround a bug in some versions
+  // of MinGW that define fileno as a macro.
+  FMT_API int(fileno)() const;
+
+  void vprint(string_view format_str, format_args args) {
+    fmt::vprint(file_, format_str, args);
+  }
+
+  template <typename... Args>
+  inline void print(string_view format_str, const Args&... args) {
+    vprint(format_str, make_format_args(args...));
+  }
+};
+
+#if FMT_USE_FCNTL
+// A file. Closed file is represented by a file object with descriptor -1.
+// Methods that are not declared with FMT_NOEXCEPT may throw
+// fmt::system_error in case of failure. Note that some errors such as
+// closing the file multiple times will cause a crash on Windows rather
+// than an exception. You can get standard behavior by overriding the
+// invalid parameter handler with _set_invalid_parameter_handler.
+class file {
+ private:
+  int fd_;  // File descriptor.
+
+  // Constructs a file object with a given descriptor.
+  explicit file(int fd) : fd_(fd) {}
+
+ public:
+  // Possible values for the oflag argument to the constructor.
+  enum {
+    RDONLY = FMT_POSIX(O_RDONLY),  // Open for reading only.
+    WRONLY = FMT_POSIX(O_WRONLY),  // Open for writing only.
+    RDWR = FMT_POSIX(O_RDWR),      // Open for reading and writing.
+    CREATE = FMT_POSIX(O_CREAT),   // Create if the file doesn't exist.
+    APPEND = FMT_POSIX(O_APPEND)   // Open in append mode.
+  };
+
+  // Constructs a file object which doesn't represent any file.
+  file() FMT_NOEXCEPT : fd_(-1) {}
+
+  // Opens a file and constructs a file object representing this file.
+  FMT_API file(cstring_view path, int oflag);
+
+ public:
+  file(const file&) = delete;
+  void operator=(const file&) = delete;
+
+  file(file&& other) FMT_NOEXCEPT : fd_(other.fd_) { other.fd_ = -1; }
+
+  file& operator=(file&& other) FMT_NOEXCEPT {
+    close();
+    fd_ = other.fd_;
+    other.fd_ = -1;
+    return *this;
+  }
+
+  // Destroys the object closing the file it represents if any.
+  FMT_API ~file() FMT_NOEXCEPT;
+
+  // Returns the file descriptor.
+  int descriptor() const FMT_NOEXCEPT { return fd_; }
+
+  // Closes the file.
+  FMT_API void close();
+
+  // Returns the file size. The size has signed type for consistency with
+  // stat::st_size.
+  FMT_API long long size() const;
+
+  // Attempts to read count bytes from the file into the specified buffer.
+  FMT_API size_t read(void* buffer, size_t count);
+
+  // Attempts to write count bytes from the specified buffer to the file.
+  FMT_API size_t write(const void* buffer, size_t count);
+
+  // Duplicates a file descriptor with the dup function and returns
+  // the duplicate as a file object.
+  FMT_API static file dup(int fd);
+
+  // Makes fd be the copy of this file descriptor, closing fd first if
+  // necessary.
+  FMT_API void dup2(int fd);
+
+  // Makes fd be the copy of this file descriptor, closing fd first if
+  // necessary.
+  FMT_API void dup2(int fd, error_code& ec) FMT_NOEXCEPT;
+
+  // Creates a pipe setting up read_end and write_end file objects for reading
+  // and writing respectively.
+  FMT_API static void pipe(file& read_end, file& write_end);
+
+  // Creates a buffered_file object associated with this file and detaches
+  // this file object from the file.
+  FMT_API buffered_file fdopen(const char* mode);
+};
+
+// Returns the memory page size.
+long getpagesize();
+
+namespace detail {
+
+struct buffer_size {
+  size_t value = 0;
+  buffer_size operator=(size_t val) const {
+    auto bs = buffer_size();
+    bs.value = val;
+    return bs;
+  }
+};
+
+struct ostream_params {
+  int oflag = file::WRONLY | file::CREATE;
+  size_t buffer_size = BUFSIZ > 32768 ? BUFSIZ : 32768;
+
+  ostream_params() {}
+
+  template <typename... T>
+  ostream_params(T... params, int oflag) : ostream_params(params...) {
+    this->oflag = oflag;
+  }
+
+  template <typename... T>
+  ostream_params(T... params, detail::buffer_size bs)
+      : ostream_params(params...) {
+    this->buffer_size = bs.value;
+  }
+};
+}  // namespace detail
+
+static constexpr detail::buffer_size buffer_size;
+
+// A fast output stream which is not thread-safe.
+class ostream final : private detail::buffer<char> {
+ private:
+  file file_;
+
+  void flush() {
+    if (size() == 0) return;
+    file_.write(data(), size());
+    clear();
+  }
+
+  FMT_API void grow(size_t) override final;
+
+  ostream(cstring_view path, const detail::ostream_params& params)
+      : file_(path, params.oflag) {
+    set(new char[params.buffer_size], params.buffer_size);
+  }
+
+ public:
+  ostream(ostream&& other)
+      : detail::buffer<char>(other.data(), other.size(), other.capacity()),
+        file_(std::move(other.file_)) {
+    other.set(nullptr, 0);
+  }
+  ~ostream() {
+    flush();
+    delete[] data();
+  }
+
+  template <typename... T>
+  friend ostream output_file(cstring_view path, T... params);
+
+  void close() {
+    flush();
+    file_.close();
+  }
+
+  template <typename S, typename... Args>
+  void print(const S& format_str, const Args&... args) {
+    format_to(detail::buffer_appender<char>(*this), format_str, args...);
+  }
+};
+
+/**
+  Opens a file for writing. Supported parameters passed in `params`:
+  * ``<integer>``: Output flags (``file::WRONLY | file::CREATE`` by default)
+  * ``buffer_size=<integer>``: Output buffer size
+ */
+template <typename... T>
+inline ostream output_file(cstring_view path, T... params) {
+  return {path, detail::ostream_params(params...)};
+}
+#endif  // FMT_USE_FCNTL
+
+#ifdef FMT_LOCALE
+// A "C" numeric locale.
+class locale {
+ private:
+#  ifdef _WIN32
+  using locale_t = _locale_t;
+
+  static void freelocale(locale_t loc) { _free_locale(loc); }
+
+  static double strtod_l(const char* nptr, char** endptr, _locale_t loc) {
+    return _strtod_l(nptr, endptr, loc);
+  }
+#  endif
+
+  locale_t locale_;
+
+ public:
+  using type = locale_t;
+  locale(const locale&) = delete;
+  void operator=(const locale&) = delete;
+
+  locale() {
+#  ifndef _WIN32
+    locale_ = FMT_SYSTEM(newlocale(LC_NUMERIC_MASK, "C", nullptr));
+#  else
+    locale_ = _create_locale(LC_NUMERIC, "C");
+#  endif
+    if (!locale_) FMT_THROW(system_error(errno, "cannot create locale"));
+  }
+  ~locale() { freelocale(locale_); }
+
+  type get() const { return locale_; }
+
+  // Converts string to floating-point number and advances str past the end
+  // of the parsed input.
+  double strtod(const char*& str) const {
+    char* end = nullptr;
+    double result = strtod_l(str, &end, locale_);
+    str = end;
+    return result;
+  }
+};
+using Locale FMT_DEPRECATED_ALIAS = locale;
+#endif  // FMT_LOCALE
+FMT_END_NAMESPACE
+
+#endif  // FMT_OS_H_
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/ostream.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/ostream.h
index 69bac0e24..29c58ec13 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/ostream.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/ostream.h
@@ -9,10 +9,15 @@
 #define FMT_OSTREAM_H_
 
 #include <ostream>
+
 #include "format.h"
 
 FMT_BEGIN_NAMESPACE
-namespace internal {
+
+template <typename Char> class basic_printf_parse_context;
+template <typename OutputIt, typename Char> class basic_printf_context;
+
+namespace detail {
 
 template <class Char> class formatbuf : public std::basic_streambuf<Char> {
  private:
@@ -44,21 +49,35 @@ template <class Char> class formatbuf : public std::basic_streambuf<Char> {
   }
 };
 
+struct converter {
+  template <typename T, FMT_ENABLE_IF(is_integral<T>::value)> converter(T);
+};
+
 template <typename Char> struct test_stream : std::basic_ostream<Char> {
  private:
-  struct null;
-  // Hide all operator<< from std::basic_ostream<Char>.
-  void operator<<(null);
+  void_t<> operator<<(converter);
 };
 
+// Hide insertion operators for built-in types.
+template <typename Char, typename Traits>
+void_t<> operator<<(std::basic_ostream<Char, Traits>&, Char);
+template <typename Char, typename Traits>
+void_t<> operator<<(std::basic_ostream<Char, Traits>&, char);
+template <typename Traits>
+void_t<> operator<<(std::basic_ostream<char, Traits>&, char);
+template <typename Traits>
+void_t<> operator<<(std::basic_ostream<char, Traits>&, signed char);
+template <typename Traits>
+void_t<> operator<<(std::basic_ostream<char, Traits>&, unsigned char);
+
 // Checks if T has a user-defined operator<< (e.g. not a member of
 // std::ostream).
 template <typename T, typename Char> class is_streamable {
  private:
   template <typename U>
-  static decltype((void)(std::declval<test_stream<Char>&>()
-                         << std::declval<U>()),
-                  std::true_type())
+  static bool_constant<!std::is_same<decltype(std::declval<test_stream<Char>&>()
+                                              << std::declval<U>()),
+                                     void_t<>>::value>
   test(int);
 
   template <typename> static std::false_type test(...);
@@ -71,12 +90,11 @@ template <typename T, typename Char> class is_streamable {
 
 // Write the content of buf to os.
 template <typename Char>
-void write(std::basic_ostream<Char>& os, buffer<Char>& buf) {
+void write_buffer(std::basic_ostream<Char>& os, buffer<Char>& buf) {
   const Char* buf_data = buf.data();
   using unsigned_streamsize = std::make_unsigned<std::streamsize>::type;
   unsigned_streamsize size = buf.size();
-  unsigned_streamsize max_size =
-      to_unsigned((std::numeric_limits<std::streamsize>::max)());
+  unsigned_streamsize max_size = to_unsigned(max_value<std::streamsize>());
   do {
     unsigned_streamsize n = size <= max_size ? size : max_size;
     os.write(buf_data, static_cast<std::streamsize>(n));
@@ -86,34 +104,57 @@ void write(std::basic_ostream<Char>& os, buffer<Char>& buf) {
 }
 
 template <typename Char, typename T>
-void format_value(buffer<Char>& buf, const T& value) {
+void format_value(buffer<Char>& buf, const T& value,
+                  locale_ref loc = locale_ref()) {
   formatbuf<Char> format_buf(buf);
   std::basic_ostream<Char> output(&format_buf);
-  output.exceptions(std::ios_base::failbit | std::ios_base::badbit);
+#if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
+  if (loc) output.imbue(loc.get<std::locale>());
+#endif
   output << value;
-  buf.resize(buf.size());
+  output.exceptions(std::ios_base::failbit | std::ios_base::badbit);
+  buf.try_resize(buf.size());
 }
 
 // Formats an object of type T that has an overloaded ostream operator<<.
 template <typename T, typename Char>
 struct fallback_formatter<T, Char, enable_if_t<is_streamable<T, Char>::value>>
-    : formatter<basic_string_view<Char>, Char> {
-  template <typename Context>
-  auto format(const T& value, Context& ctx) -> decltype(ctx.out()) {
+    : private formatter<basic_string_view<Char>, Char> {
+  FMT_CONSTEXPR auto parse(basic_format_parse_context<Char>& ctx)
+      -> decltype(ctx.begin()) {
+    return formatter<basic_string_view<Char>, Char>::parse(ctx);
+  }
+  template <typename ParseCtx,
+            FMT_ENABLE_IF(std::is_same<
+                          ParseCtx, basic_printf_parse_context<Char>>::value)>
+  auto parse(ParseCtx& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename OutputIt>
+  auto format(const T& value, basic_format_context<OutputIt, Char>& ctx)
+      -> OutputIt {
     basic_memory_buffer<Char> buffer;
-    format_value(buffer, value);
+    format_value(buffer, value, ctx.locale());
     basic_string_view<Char> str(buffer.data(), buffer.size());
     return formatter<basic_string_view<Char>, Char>::format(str, ctx);
   }
+  template <typename OutputIt>
+  auto format(const T& value, basic_printf_context<OutputIt, Char>& ctx)
+      -> OutputIt {
+    basic_memory_buffer<Char> buffer;
+    format_value(buffer, value, ctx.locale());
+    return std::copy(buffer.begin(), buffer.end(), ctx.out());
+  }
 };
-}  // namespace internal
+}  // namespace detail
 
 template <typename Char>
 void vprint(std::basic_ostream<Char>& os, basic_string_view<Char> format_str,
-            basic_format_args<buffer_context<Char>> args) {
+            basic_format_args<buffer_context<type_identity_t<Char>>> args) {
   basic_memory_buffer<Char> buffer;
-  internal::vformat_to(buffer, format_str, args);
-  internal::write(os, buffer);
+  detail::vformat_to(buffer, format_str, args);
+  detail::write_buffer(os, buffer);
 }
 
 /**
@@ -126,10 +167,10 @@ void vprint(std::basic_ostream<Char>& os, basic_string_view<Char> format_str,
   \endrst
  */
 template <typename S, typename... Args,
-          typename Char = enable_if_t<internal::is_string<S>::value, char_t<S>>>
+          typename Char = enable_if_t<detail::is_string<S>::value, char_t<S>>>
 void print(std::basic_ostream<Char>& os, const S& format_str, Args&&... args) {
   vprint(os, to_string_view(format_str),
-         {internal::make_args_checked<Args...>(format_str, args...)});
+         fmt::make_args_checked<Args...>(format_str, args...));
 }
 FMT_END_NAMESPACE
 
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/posix.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/posix.h
index 6b2d7f8e4..da19e9d53 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/posix.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/posix.h
@@ -1,311 +1,2 @@
-// A C++ interface to POSIX functions.
-//
-// Copyright (c) 2012 - 2016, Victor Zverovich
-// All rights reserved.
-//
-// For the license information refer to format.h.
-
-#ifndef FMT_POSIX_H_
-#define FMT_POSIX_H_
-
-#if defined(__MINGW32__) || defined(__CYGWIN__)
-// Workaround MinGW bug https://sourceforge.net/p/mingw/bugs/2024/.
-#  undef __STRICT_ANSI__
-#endif
-
-#include <errno.h>
-#include <fcntl.h>   // for O_RDONLY
-#include <locale.h>  // for locale_t
-#include <stdio.h>
-#include <stdlib.h>  // for strtod_l
-
-#include <cstddef>
-
-#if defined __APPLE__ || defined(__FreeBSD__)
-#  include <xlocale.h>  // for LC_NUMERIC_MASK on OS X
-#endif
-
-#include "format.h"
-
-#ifndef FMT_POSIX
-#  if defined(_WIN32) && !defined(__MINGW32__)
-// Fix warnings about deprecated symbols.
-#    define FMT_POSIX(call) _##call
-#  else
-#    define FMT_POSIX(call) call
-#  endif
-#endif
-
-// Calls to system functions are wrapped in FMT_SYSTEM for testability.
-#ifdef FMT_SYSTEM
-#  define FMT_POSIX_CALL(call) FMT_SYSTEM(call)
-#else
-#  define FMT_SYSTEM(call) call
-#  ifdef _WIN32
-// Fix warnings about deprecated symbols.
-#    define FMT_POSIX_CALL(call) ::_##call
-#  else
-#    define FMT_POSIX_CALL(call) ::call
-#  endif
-#endif
-
-// Retries the expression while it evaluates to error_result and errno
-// equals to EINTR.
-#ifndef _WIN32
-#  define FMT_RETRY_VAL(result, expression, error_result) \
-    do {                                                  \
-      result = (expression);                              \
-    } while (result == error_result && errno == EINTR)
-#else
-#  define FMT_RETRY_VAL(result, expression, error_result) result = (expression)
-#endif
-
-#define FMT_RETRY(result, expression) FMT_RETRY_VAL(result, expression, -1)
-
-FMT_BEGIN_NAMESPACE
-
-/**
-  \rst
-  A reference to a null-terminated string. It can be constructed from a C
-  string or ``std::string``.
-
-  You can use one of the following type aliases for common character types:
-
-  +---------------+-----------------------------+
-  | Type          | Definition                  |
-  +===============+=============================+
-  | cstring_view  | basic_cstring_view<char>    |
-  +---------------+-----------------------------+
-  | wcstring_view | basic_cstring_view<wchar_t> |
-  +---------------+-----------------------------+
-
-  This class is most useful as a parameter type to allow passing
-  different types of strings to a function, for example::
-
-    template <typename... Args>
-    std::string format(cstring_view format_str, const Args & ... args);
-
-    format("{}", 42);
-    format(std::string("{}"), 42);
-  \endrst
- */
-template <typename Char> class basic_cstring_view {
- private:
-  const Char* data_;
-
- public:
-  /** Constructs a string reference object from a C string. */
-  basic_cstring_view(const Char* s) : data_(s) {}
-
-  /**
-    \rst
-    Constructs a string reference from an ``std::string`` object.
-    \endrst
-   */
-  basic_cstring_view(const std::basic_string<Char>& s) : data_(s.c_str()) {}
-
-  /** Returns the pointer to a C string. */
-  const Char* c_str() const { return data_; }
-};
-
-using cstring_view = basic_cstring_view<char>;
-using wcstring_view = basic_cstring_view<wchar_t>;
-
-// An error code.
-class error_code {
- private:
-  int value_;
-
- public:
-  explicit error_code(int value = 0) FMT_NOEXCEPT : value_(value) {}
-
-  int get() const FMT_NOEXCEPT { return value_; }
-};
-
-// A buffered file.
-class buffered_file {
- private:
-  FILE* file_;
-
-  friend class file;
-
-  explicit buffered_file(FILE* f) : file_(f) {}
-
- public:
-  // Constructs a buffered_file object which doesn't represent any file.
-  buffered_file() FMT_NOEXCEPT : file_(nullptr) {}
-
-  // Destroys the object closing the file it represents if any.
-  FMT_API ~buffered_file() FMT_NOEXCEPT;
-
- private:
-  buffered_file(const buffered_file&) = delete;
-  void operator=(const buffered_file&) = delete;
-
- public:
-  buffered_file(buffered_file&& other) FMT_NOEXCEPT : file_(other.file_) {
-    other.file_ = nullptr;
-  }
-
-  buffered_file& operator=(buffered_file&& other) {
-    close();
-    file_ = other.file_;
-    other.file_ = nullptr;
-    return *this;
-  }
-
-  // Opens a file.
-  FMT_API buffered_file(cstring_view filename, cstring_view mode);
-
-  // Closes the file.
-  FMT_API void close();
-
-  // Returns the pointer to a FILE object representing this file.
-  FILE* get() const FMT_NOEXCEPT { return file_; }
-
-  // We place parentheses around fileno to workaround a bug in some versions
-  // of MinGW that define fileno as a macro.
-  FMT_API int(fileno)() const;
-
-  void vprint(string_view format_str, format_args args) {
-    fmt::vprint(file_, format_str, args);
-  }
-
-  template <typename... Args>
-  inline void print(string_view format_str, const Args&... args) {
-    vprint(format_str, make_format_args(args...));
-  }
-};
-
-// A file. Closed file is represented by a file object with descriptor -1.
-// Methods that are not declared with FMT_NOEXCEPT may throw
-// fmt::system_error in case of failure. Note that some errors such as
-// closing the file multiple times will cause a crash on Windows rather
-// than an exception. You can get standard behavior by overriding the
-// invalid parameter handler with _set_invalid_parameter_handler.
-class file {
- private:
-  int fd_;  // File descriptor.
-
-  // Constructs a file object with a given descriptor.
-  explicit file(int fd) : fd_(fd) {}
-
- public:
-  // Possible values for the oflag argument to the constructor.
-  enum {
-    RDONLY = FMT_POSIX(O_RDONLY),  // Open for reading only.
-    WRONLY = FMT_POSIX(O_WRONLY),  // Open for writing only.
-    RDWR = FMT_POSIX(O_RDWR)       // Open for reading and writing.
-  };
-
-  // Constructs a file object which doesn't represent any file.
-  file() FMT_NOEXCEPT : fd_(-1) {}
-
-  // Opens a file and constructs a file object representing this file.
-  FMT_API file(cstring_view path, int oflag);
-
- private:
-  file(const file&) = delete;
-  void operator=(const file&) = delete;
-
- public:
-  file(file&& other) FMT_NOEXCEPT : fd_(other.fd_) { other.fd_ = -1; }
-
-  file& operator=(file&& other) {
-    close();
-    fd_ = other.fd_;
-    other.fd_ = -1;
-    return *this;
-  }
-
-  // Destroys the object closing the file it represents if any.
-  FMT_API ~file() FMT_NOEXCEPT;
-
-  // Returns the file descriptor.
-  int descriptor() const FMT_NOEXCEPT { return fd_; }
-
-  // Closes the file.
-  FMT_API void close();
-
-  // Returns the file size. The size has signed type for consistency with
-  // stat::st_size.
-  FMT_API long long size() const;
-
-  // Attempts to read count bytes from the file into the specified buffer.
-  FMT_API std::size_t read(void* buffer, std::size_t count);
-
-  // Attempts to write count bytes from the specified buffer to the file.
-  FMT_API std::size_t write(const void* buffer, std::size_t count);
-
-  // Duplicates a file descriptor with the dup function and returns
-  // the duplicate as a file object.
-  FMT_API static file dup(int fd);
-
-  // Makes fd be the copy of this file descriptor, closing fd first if
-  // necessary.
-  FMT_API void dup2(int fd);
-
-  // Makes fd be the copy of this file descriptor, closing fd first if
-  // necessary.
-  FMT_API void dup2(int fd, error_code& ec) FMT_NOEXCEPT;
-
-  // Creates a pipe setting up read_end and write_end file objects for reading
-  // and writing respectively.
-  FMT_API static void pipe(file& read_end, file& write_end);
-
-  // Creates a buffered_file object associated with this file and detaches
-  // this file object from the file.
-  FMT_API buffered_file fdopen(const char* mode);
-};
-
-// Returns the memory page size.
-long getpagesize();
-
-#ifdef FMT_LOCALE
-// A "C" numeric locale.
-class Locale {
- private:
-#  ifdef _WIN32
-  using locale_t = _locale_t;
-
-  enum { LC_NUMERIC_MASK = LC_NUMERIC };
-
-  static locale_t newlocale(int category_mask, const char* locale, locale_t) {
-    return _create_locale(category_mask, locale);
-  }
-
-  static void freelocale(locale_t locale) { _free_locale(locale); }
-
-  static double strtod_l(const char* nptr, char** endptr, _locale_t locale) {
-    return _strtod_l(nptr, endptr, locale);
-  }
-#  endif
-
-  locale_t locale_;
-
-  Locale(const Locale&) = delete;
-  void operator=(const Locale&) = delete;
-
- public:
-  using type = locale_t;
-
-  Locale() : locale_(newlocale(LC_NUMERIC_MASK, "C", nullptr)) {
-    if (!locale_) FMT_THROW(system_error(errno, "cannot create locale"));
-  }
-  ~Locale() { freelocale(locale_); }
-
-  type get() const { return locale_; }
-
-  // Converts string to floating-point number and advances str past the end
-  // of the parsed input.
-  double strtod(const char*& str) const {
-    char* end = nullptr;
-    double result = strtod_l(str, &end, locale_);
-    str = end;
-    return result;
-  }
-};
-#endif  // FMT_LOCALE
-FMT_END_NAMESPACE
-
-#endif  // FMT_POSIX_H_
+#include "os.h"
+#warning "fmt/posix.h is deprecated; use fmt/os.h instead"
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/printf.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/printf.h
index c803aa952..8c28ac232 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/printf.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/printf.h
@@ -1,4 +1,4 @@
-// Formatting library for C++
+// Formatting library for C++ - legacy printf implementation
 //
 // Copyright (c) 2012 - 2016, Victor Zverovich
 // All rights reserved.
@@ -8,23 +8,19 @@
 #ifndef FMT_PRINTF_H_
 #define FMT_PRINTF_H_
 
-#include <algorithm>  // std::fill_n
+#include <algorithm>  // std::max
 #include <limits>     // std::numeric_limits
 
 #include "ostream.h"
 
 FMT_BEGIN_NAMESPACE
-namespace internal {
-
-// A helper function to suppress bogus "conditional expression is constant"
-// warnings.
-template <typename T> inline T const_check(T value) { return value; }
+namespace detail {
 
 // Checks if a value fits in int - used to avoid warnings about comparing
 // signed and unsigned integers.
 template <bool IsSigned> struct int_checker {
   template <typename T> static bool fits_in_int(T value) {
-    unsigned max = std::numeric_limits<int>::max();
+    unsigned max = max_value<int>();
     return value <= max;
   }
   static bool fits_in_int(bool) { return true; }
@@ -32,8 +28,8 @@ template <bool IsSigned> struct int_checker {
 
 template <> struct int_checker<true> {
   template <typename T> static bool fits_in_int(T value) {
-    return value >= std::numeric_limits<int>::min() &&
-           value <= std::numeric_limits<int>::max();
+    return value >= (std::numeric_limits<int>::min)() &&
+           value <= max_value<int>();
   }
   static bool fits_in_int(int) { return true; }
 };
@@ -94,11 +90,11 @@ template <typename T, typename Context> class arg_converter {
     if (const_check(sizeof(target_type) <= sizeof(int))) {
       // Extra casts are used to silence warnings.
       if (is_signed) {
-        arg_ = internal::make_arg<Context>(
+        arg_ = detail::make_arg<Context>(
             static_cast<int>(static_cast<target_type>(value)));
       } else {
         using unsigned_type = typename make_unsigned_or_bool<target_type>::type;
-        arg_ = internal::make_arg<Context>(
+        arg_ = detail::make_arg<Context>(
             static_cast<unsigned>(static_cast<unsigned_type>(value)));
       }
     } else {
@@ -106,9 +102,9 @@ template <typename T, typename Context> class arg_converter {
         // glibc's printf doesn't sign extend arguments of smaller types:
         //   std::printf("%lld", -42);  // prints "4294967254"
         // but we don't have to do the same because it's a UB.
-        arg_ = internal::make_arg<Context>(static_cast<long long>(value));
+        arg_ = detail::make_arg<Context>(static_cast<long long>(value));
       } else {
-        arg_ = internal::make_arg<Context>(
+        arg_ = detail::make_arg<Context>(
             static_cast<typename make_unsigned_or_bool<U>::type>(value));
       }
     }
@@ -137,7 +133,7 @@ template <typename Context> class char_converter {
 
   template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
   void operator()(T value) {
-    arg_ = internal::make_arg<Context>(
+    arg_ = detail::make_arg<Context>(
         static_cast<typename Context::char_type>(value));
   }
 
@@ -145,6 +141,13 @@ template <typename Context> class char_converter {
   void operator()(T) {}  // No conversion needed for non-integral types.
 };
 
+// An argument visitor that return a pointer to a C string if argument is a
+// string or null otherwise.
+template <typename Char> struct get_cstring {
+  template <typename T> const Char* operator()(T) { return nullptr; }
+  const Char* operator()(const Char* s) { return s; }
+};
+
 // Checks if an argument is a valid printf width specifier and sets
 // left alignment if it is negative.
 template <typename Char> class printf_width_handler {
@@ -158,12 +161,12 @@ template <typename Char> class printf_width_handler {
 
   template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
   unsigned operator()(T value) {
-    auto width = static_cast<uint32_or_64_t<T>>(value);
-    if (internal::is_negative(value)) {
+    auto width = static_cast<uint32_or_64_or_128_t<T>>(value);
+    if (detail::is_negative(value)) {
       specs_.align = align::left;
       width = 0 - width;
     }
-    unsigned int_max = std::numeric_limits<int>::max();
+    unsigned int_max = max_value<int>();
     if (width > int_max) FMT_THROW(format_error("number is too big"));
     return static_cast<unsigned>(width);
   }
@@ -176,23 +179,25 @@ template <typename Char> class printf_width_handler {
 };
 
 template <typename Char, typename Context>
-void printf(buffer<Char>& buf, basic_string_view<Char> format,
-            basic_format_args<Context> args) {
-  Context(std::back_inserter(buf), format, args).format();
+void vprintf(buffer<Char>& buf, basic_string_view<Char> format,
+             basic_format_args<Context> args) {
+  Context(buffer_appender<Char>(buf), format, args).format();
 }
+}  // namespace detail
 
-template <typename OutputIt, typename Char, typename Context>
-internal::truncating_iterator<OutputIt> printf(
-    internal::truncating_iterator<OutputIt> it, basic_string_view<Char> format,
-    basic_format_args<Context> args) {
-  return Context(it, format, args).format();
+// For printing into memory_buffer.
+template <typename Char, typename Context>
+FMT_DEPRECATED void printf(detail::buffer<Char>& buf,
+                           basic_string_view<Char> format,
+                           basic_format_args<Context> args) {
+  return detail::vprintf(buf, format, args);
 }
-}  // namespace internal
-
-using internal::printf;  // For printing into memory_buffer.
-
-template <typename Range> class printf_arg_formatter;
+using detail::vprintf;
 
+template <typename Char>
+class basic_printf_parse_context : public basic_format_parse_context<Char> {
+  using basic_format_parse_context<Char>::basic_format_parse_context;
+};
 template <typename OutputIt, typename Char> class basic_printf_context;
 
 /**
@@ -200,15 +205,15 @@ template <typename OutputIt, typename Char> class basic_printf_context;
   The ``printf`` argument formatter.
   \endrst
  */
-template <typename Range>
-class printf_arg_formatter : public internal::arg_formatter_base<Range> {
+template <typename OutputIt, typename Char>
+class printf_arg_formatter : public detail::arg_formatter_base<OutputIt, Char> {
  public:
-  using iterator = typename Range::iterator;
+  using iterator = OutputIt;
 
  private:
-  using char_type = typename Range::value_type;
-  using base = internal::arg_formatter_base<Range>;
-  using context_type = basic_printf_context<iterator, char_type>;
+  using char_type = Char;
+  using base = detail::arg_formatter_base<OutputIt, Char>;
+  using context_type = basic_printf_context<OutputIt, Char>;
 
   context_type& context_;
 
@@ -233,9 +238,9 @@ class printf_arg_formatter : public internal::arg_formatter_base<Range> {
     \endrst
    */
   printf_arg_formatter(iterator iter, format_specs& specs, context_type& ctx)
-      : base(Range(iter), &specs, internal::locale_ref()), context_(ctx) {}
+      : base(iter, &specs, detail::locale_ref()), context_(ctx) {}
 
-  template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
+  template <typename T, FMT_ENABLE_IF(fmt::detail::is_integral<T>::value)>
   iterator operator()(T value) {
     // MSVC2013 fails to compile separate overloads for bool and char_type so
     // use std::is_same instead.
@@ -250,7 +255,11 @@ class printf_arg_formatter : public internal::arg_formatter_base<Range> {
         return (*this)(static_cast<int>(value));
       fmt_specs.sign = sign::none;
       fmt_specs.alt = false;
-      fmt_specs.align = align::right;
+      fmt_specs.fill[0] = ' ';  // Ignore '0' flag for char types.
+      // align::numeric needs to be overwritten here since the '0' flag is
+      // ignored for non-numeric types
+      if (fmt_specs.align == align::none || fmt_specs.align == align::numeric)
+        fmt_specs.align = align::right;
       return base::operator()(value);
     } else {
       return base::operator()(value);
@@ -307,6 +316,8 @@ class printf_arg_formatter : public internal::arg_formatter_base<Range> {
 };
 
 template <typename T> struct printf_formatter {
+  printf_formatter() = delete;
+
   template <typename ParseContext>
   auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
     return ctx.begin();
@@ -314,17 +325,21 @@ template <typename T> struct printf_formatter {
 
   template <typename FormatContext>
   auto format(const T& value, FormatContext& ctx) -> decltype(ctx.out()) {
-    internal::format_value(internal::get_container(ctx.out()), value);
+    detail::format_value(detail::get_container(ctx.out()), value);
     return ctx.out();
   }
 };
 
-/** This template formats data and writes the output to a writer. */
+/**
+ This template formats data and writes the output through an output iterator.
+ */
 template <typename OutputIt, typename Char> class basic_printf_context {
  public:
   /** The character type for the output. */
   using char_type = Char;
+  using iterator = OutputIt;
   using format_arg = basic_format_arg<basic_printf_context>;
+  using parse_context_type = basic_printf_parse_context<Char>;
   template <typename T> using formatter_type = printf_formatter<T>;
 
  private:
@@ -332,24 +347,23 @@ template <typename OutputIt, typename Char> class basic_printf_context {
 
   OutputIt out_;
   basic_format_args<basic_printf_context> args_;
-  basic_parse_context<Char> parse_ctx_;
+  parse_context_type parse_ctx_;
 
   static void parse_flags(format_specs& specs, const Char*& it,
                           const Char* end);
 
-  // Returns the argument with specified index or, if arg_index is equal
-  // to the maximum unsigned value, the next argument.
-  format_arg get_arg(unsigned arg_index = std::numeric_limits<unsigned>::max());
+  // Returns the argument with specified index or, if arg_index is -1, the next
+  // argument.
+  format_arg get_arg(int arg_index = -1);
 
   // Parses argument index, flags and width and returns the argument index.
-  unsigned parse_header(const Char*& it, const Char* end, format_specs& specs);
+  int parse_header(const Char*& it, const Char* end, format_specs& specs);
 
  public:
   /**
    \rst
-   Constructs a ``printf_context`` object. References to the arguments and
-   the writer are stored in the context object so make sure they have
-   appropriate lifetimes.
+   Constructs a ``printf_context`` object. References to the arguments are
+   stored in the context object so make sure they have appropriate lifetimes.
    \endrst
    */
   basic_printf_context(OutputIt out, basic_string_view<char_type> format_str,
@@ -359,17 +373,18 @@ template <typename OutputIt, typename Char> class basic_printf_context {
   OutputIt out() { return out_; }
   void advance_to(OutputIt it) { out_ = it; }
 
-  format_arg arg(unsigned id) const { return args_.get(id); }
+  detail::locale_ref locale() { return {}; }
 
-  basic_parse_context<Char>& parse_context() { return parse_ctx_; }
+  format_arg arg(int id) const { return args_.get(id); }
+
+  parse_context_type& parse_context() { return parse_ctx_; }
 
   FMT_CONSTEXPR void on_error(const char* message) {
     parse_ctx_.on_error(message);
   }
 
   /** Formats stored arguments and writes the output to the range. */
-  template <typename ArgFormatter =
-                printf_arg_formatter<internal::buffer_range<Char>>>
+  template <typename ArgFormatter = printf_arg_formatter<OutputIt, Char>>
   OutputIt format();
 };
 
@@ -389,7 +404,9 @@ void basic_printf_context<OutputIt, Char>::parse_flags(format_specs& specs,
       specs.fill[0] = '0';
       break;
     case ' ':
-      specs.sign = sign::space;
+      if (specs.sign != sign::plus) {
+        specs.sign = sign::space;
+      }
       break;
     case '#':
       specs.alt = true;
@@ -402,24 +419,25 @@ void basic_printf_context<OutputIt, Char>::parse_flags(format_specs& specs,
 
 template <typename OutputIt, typename Char>
 typename basic_printf_context<OutputIt, Char>::format_arg
-basic_printf_context<OutputIt, Char>::get_arg(unsigned arg_index) {
-  if (arg_index == std::numeric_limits<unsigned>::max())
+basic_printf_context<OutputIt, Char>::get_arg(int arg_index) {
+  if (arg_index < 0)
     arg_index = parse_ctx_.next_arg_id();
   else
     parse_ctx_.check_arg_id(--arg_index);
-  return internal::get_arg(*this, arg_index);
+  return detail::get_arg(*this, arg_index);
 }
 
 template <typename OutputIt, typename Char>
-unsigned basic_printf_context<OutputIt, Char>::parse_header(
-    const Char*& it, const Char* end, format_specs& specs) {
-  unsigned arg_index = std::numeric_limits<unsigned>::max();
+int basic_printf_context<OutputIt, Char>::parse_header(const Char*& it,
+                                                       const Char* end,
+                                                       format_specs& specs) {
+  int arg_index = -1;
   char_type c = *it;
   if (c >= '0' && c <= '9') {
     // Parse an argument index (if followed by '$') or a width possibly
     // preceded with '0' flag(s).
-    internal::error_handler eh;
-    unsigned value = parse_nonnegative_int(it, end, eh);
+    detail::error_handler eh;
+    int value = parse_nonnegative_int(it, end, eh);
     if (it != end && *it == '$') {  // value is an argument index
       ++it;
       arg_index = value;
@@ -437,12 +455,12 @@ unsigned basic_printf_context<OutputIt, Char>::parse_header(
   // Parse width.
   if (it != end) {
     if (*it >= '0' && *it <= '9') {
-      internal::error_handler eh;
+      detail::error_handler eh;
       specs.width = parse_nonnegative_int(it, end, eh);
     } else if (*it == '*') {
       ++it;
-      specs.width = visit_format_arg(
-          internal::printf_width_handler<char_type>(specs), get_arg());
+      specs.width = static_cast<int>(visit_format_arg(
+          detail::printf_width_handler<char_type>(specs), get_arg()));
     }
   }
   return arg_index;
@@ -469,38 +487,53 @@ OutputIt basic_printf_context<OutputIt, Char>::format() {
     specs.align = align::right;
 
     // Parse argument index, flags and width.
-    unsigned arg_index = parse_header(it, end, specs);
+    int arg_index = parse_header(it, end, specs);
+    if (arg_index == 0) on_error("argument not found");
 
     // Parse precision.
     if (it != end && *it == '.') {
       ++it;
       c = it != end ? *it : 0;
       if ('0' <= c && c <= '9') {
-        internal::error_handler eh;
-        specs.precision = static_cast<int>(parse_nonnegative_int(it, end, eh));
+        detail::error_handler eh;
+        specs.precision = parse_nonnegative_int(it, end, eh);
       } else if (c == '*') {
         ++it;
-        specs.precision =
-            visit_format_arg(internal::printf_precision_handler(), get_arg());
+        specs.precision = static_cast<int>(
+            visit_format_arg(detail::printf_precision_handler(), get_arg()));
       } else {
         specs.precision = 0;
       }
     }
 
     format_arg arg = get_arg(arg_index);
-    if (specs.alt && visit_format_arg(internal::is_zero_int(), arg))
+    // For d, i, o, u, x, and X conversion specifiers, if a precision is
+    // specified, the '0' flag is ignored
+    if (specs.precision >= 0 && arg.is_integral())
+      specs.fill[0] =
+          ' ';  // Ignore '0' flag for non-numeric types or if '-' present.
+    if (specs.precision >= 0 && arg.type() == detail::type::cstring_type) {
+      auto str = visit_format_arg(detail::get_cstring<Char>(), arg);
+      auto str_end = str + specs.precision;
+      auto nul = std::find(str, str_end, Char());
+      arg = detail::make_arg<basic_printf_context>(basic_string_view<Char>(
+          str,
+          detail::to_unsigned(nul != str_end ? nul - str : specs.precision)));
+    }
+    if (specs.alt && visit_format_arg(detail::is_zero_int(), arg))
       specs.alt = false;
     if (specs.fill[0] == '0') {
-      if (arg.is_arithmetic())
+      if (arg.is_arithmetic() && specs.align != align::left)
         specs.align = align::numeric;
       else
-        specs.fill[0] = ' ';  // Ignore '0' flag for non-numeric types.
+        specs.fill[0] = ' ';  // Ignore '0' flag for non-numeric types or if '-'
+                              // flag is also present.
     }
 
     // Parse length and convert the argument to the required type.
     c = it != end ? *it++ : 0;
     char_type t = it != end ? *it : 0;
-    using internal::convert_arg;
+    using detail::convert_arg;
     switch (c) {
     case 'h':
       if (t == 'h') {
@@ -524,7 +557,7 @@ OutputIt basic_printf_context<OutputIt, Char>::format() {
       convert_arg<intmax_t>(arg, t);
       break;
     case 'z':
-      convert_arg<std::size_t>(arg, t);
+      convert_arg<size_t>(arg, t);
       break;
     case 't':
       convert_arg<std::ptrdiff_t>(arg, t);
@@ -549,7 +582,7 @@ OutputIt basic_printf_context<OutputIt, Char>::format() {
         specs.type = 'd';
         break;
       case 'c':
-        visit_format_arg(internal::char_converter<basic_printf_context>(arg),
+        visit_format_arg(detail::char_converter<basic_printf_context>(arg),
                          arg);
         break;
       }
@@ -558,15 +591,14 @@ OutputIt basic_printf_context<OutputIt, Char>::format() {
     start = it;
 
     // Format argument.
-    visit_format_arg(ArgFormatter(out, specs, *this), arg);
+    out = visit_format_arg(ArgFormatter(out, specs, *this), arg);
   }
   return std::copy(start, it, out);
 }
 
 template <typename Char>
 using basic_printf_context_t =
-    basic_printf_context<std::back_insert_iterator<internal::buffer<Char>>,
-                         Char>;
+    basic_printf_context<detail::buffer_appender<Char>, Char>;
 
 using printf_context = basic_printf_context_t<char>;
 using wprintf_context = basic_printf_context_t<wchar_t>;
@@ -600,9 +632,10 @@ inline format_arg_store<wprintf_context, Args...> make_wprintf_args(
 
 template <typename S, typename Char = char_t<S>>
 inline std::basic_string<Char> vsprintf(
-    const S& format, basic_format_args<basic_printf_context_t<Char>> args) {
+    const S& format,
+    basic_format_args<basic_printf_context_t<type_identity_t<Char>>> args) {
   basic_memory_buffer<Char> buffer;
-  printf(buffer, to_string_view(format), args);
+  vprintf(buffer, to_string_view(format), args);
   return to_string(buffer);
 }
 
@@ -616,18 +649,19 @@ inline std::basic_string<Char> vsprintf(
   \endrst
 */
 template <typename S, typename... Args,
-          typename Char = enable_if_t<internal::is_string<S>::value, char_t<S>>>
+          typename Char = enable_if_t<detail::is_string<S>::value, char_t<S>>>
 inline std::basic_string<Char> sprintf(const S& format, const Args&... args) {
   using context = basic_printf_context_t<Char>;
-  return vsprintf(to_string_view(format), {make_format_args<context>(args...)});
+  return vsprintf(to_string_view(format), make_format_args<context>(args...));
 }
 
 template <typename S, typename Char = char_t<S>>
-inline int vfprintf(std::FILE* f, const S& format,
-                    basic_format_args<basic_printf_context_t<Char>> args) {
+inline int vfprintf(
+    std::FILE* f, const S& format,
+    basic_format_args<basic_printf_context_t<type_identity_t<Char>>> args) {
   basic_memory_buffer<Char> buffer;
-  printf(buffer, to_string_view(format), args);
-  std::size_t size = buffer.size();
+  vprintf(buffer, to_string_view(format), args);
+  size_t size = buffer.size();
   return std::fwrite(buffer.data(), sizeof(Char), size, f) < size
              ? -1
              : static_cast<int>(size);
@@ -643,16 +677,17 @@ inline int vfprintf(std::FILE* f, const S& format,
   \endrst
  */
 template <typename S, typename... Args,
-          typename Char = enable_if_t<internal::is_string<S>::value, char_t<S>>>
+          typename Char = enable_if_t<detail::is_string<S>::value, char_t<S>>>
 inline int fprintf(std::FILE* f, const S& format, const Args&... args) {
   using context = basic_printf_context_t<Char>;
   return vfprintf(f, to_string_view(format),
-                  {make_format_args<context>(args...)});
+                  make_format_args<context>(args...));
 }
 
 template <typename S, typename Char = char_t<S>>
-inline int vprintf(const S& format,
-                   basic_format_args<basic_printf_context_t<Char>> args) {
+inline int vprintf(
+    const S& format,
+    basic_format_args<basic_printf_context_t<type_identity_t<Char>>> args) {
   return vfprintf(stdout, to_string_view(format), args);
 }
 
@@ -666,19 +701,20 @@ inline int vprintf(const S& format,
   \endrst
  */
 template <typename S, typename... Args,
-          FMT_ENABLE_IF(internal::is_string<S>::value)>
+          FMT_ENABLE_IF(detail::is_string<S>::value)>
 inline int printf(const S& format_str, const Args&... args) {
   using context = basic_printf_context_t<char_t<S>>;
   return vprintf(to_string_view(format_str),
-                 {make_format_args<context>(args...)});
+                 make_format_args<context>(args...));
 }
 
 template <typename S, typename Char = char_t<S>>
-inline int vfprintf(std::basic_ostream<Char>& os, const S& format,
-                    basic_format_args<basic_printf_context_t<Char>> args) {
+inline int vfprintf(
+    std::basic_ostream<Char>& os, const S& format,
+    basic_format_args<basic_printf_context_t<type_identity_t<Char>>> args) {
   basic_memory_buffer<Char> buffer;
-  printf(buffer, to_string_view(format), args);
-  internal::write(os, buffer);
+  vprintf(buffer, to_string_view(format), args);
+  detail::write_buffer(os, buffer);
   return static_cast<int>(buffer.size());
 }
 
@@ -686,9 +722,9 @@ inline int vfprintf(std::basic_ostream<Char>& os, const S& format,
 template <typename ArgFormatter, typename Char,
           typename Context =
               basic_printf_context<typename ArgFormatter::iterator, Char>>
-typename ArgFormatter::iterator vprintf(internal::buffer<Char>& out,
-                                        basic_string_view<Char> format_str,
-                                        basic_format_args<Context> args) {
+typename ArgFormatter::iterator vprintf(
+    detail::buffer<Char>& out, basic_string_view<Char> format_str,
+    basic_format_args<type_identity_t<Context>> args) {
   typename ArgFormatter::iterator iter(out);
   Context(iter, format_str, args).template format<ArgFormatter>();
   return iter;
@@ -708,7 +744,7 @@ inline int fprintf(std::basic_ostream<Char>& os, const S& format_str,
                    const Args&... args) {
   using context = basic_printf_context_t<Char>;
   return vfprintf(os, to_string_view(format_str),
-                  {make_format_args<context>(args...)});
+                  make_format_args<context>(args...));
 }
 FMT_END_NAMESPACE
 
diff --git a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/ranges.h b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/ranges.h
index cf0d41aaa..632f04949 100644
--- a/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/ranges.h
+++ b/vtkm/thirdparty/diy/vtkmdiy/include/vtkmdiy/thirdparty/fmt/ranges.h
@@ -12,7 +12,9 @@
 #ifndef FMT_RANGES_H_
 #define FMT_RANGES_H_
 
+#include <initializer_list>
 #include <type_traits>
+
 #include "format.h"
 
 // output only up to N items from the range.
@@ -31,7 +33,7 @@ template <typename Char> struct formatting_base {
 
 template <typename Char, typename Enable = void>
 struct formatting_range : formatting_base<Char> {
-  static FMT_CONSTEXPR_DECL const std::size_t range_length_limit =
+  static FMT_CONSTEXPR_DECL const size_t range_length_limit =
       FMT_RANGE_OUTPUT_LENGTH_LIMIT;  // output only up to N items from the
                                       // range.
   Char prefix;
@@ -52,7 +54,7 @@ struct formatting_tuple : formatting_base<Char> {
   static FMT_CONSTEXPR_DECL const bool add_prepostfix_space = false;
 };
 
-namespace internal {
+namespace detail {
 
 template <typename RangeT, typename OutputIterator>
 OutputIterator copy(const RangeT& range, OutputIterator out) {
@@ -104,10 +106,7 @@ struct is_range_<
 /// tuple_size and tuple_element check.
 template <typename T> class is_tuple_like_ {
   template <typename U>
-  static auto check(U* p)
-      -> decltype(std::tuple_size<U>::value,
-                  (void)std::declval<typename std::tuple_element<0, U>::type>(),
-                  int());
+  static auto check(U* p) -> decltype(std::tuple_size<U>::value, int());
   template <typename> static void check(...);
 
  public:
@@ -119,26 +118,24 @@ template <typename T> class is_tuple_like_ {
 #if defined(__cpp_lib_integer_sequence) || FMT_MSC_VER >= 1900
 template <typename T, T... N>
 using integer_sequence = std::integer_sequence<T, N...>;
-template <std::size_t... N> using index_sequence = std::index_sequence<N...>;
-template <std::size_t N>
-using make_index_sequence = std::make_index_sequence<N>;
+template <size_t... N> using index_sequence = std::index_sequence<N...>;
+template <size_t N> using make_index_sequence = std::make_index_sequence<N>;
 #else
 template <typename T, T... N> struct integer_sequence {
   using value_type = T;
 
-  static FMT_CONSTEXPR std::size_t size() { return sizeof...(N); }
+  static FMT_CONSTEXPR size_t size() { return sizeof...(N); }
 };
 
-template <std::size_t... N>
-using index_sequence = integer_sequence<std::size_t, N...>;
+template <size_t... N> using index_sequence = integer_sequence<size_t, N...>;
 
-template <typename T, std::size_t N, T... Ns>
+template <typename T, size_t N, T... Ns>
 struct make_integer_sequence : make_integer_sequence<T, N - 1, N - 1, Ns...> {};
 template <typename T, T... Ns>
 struct make_integer_sequence<T, 0, Ns...> : integer_sequence<T, Ns...> {};
 
-template <std::size_t N>
-using make_index_sequence = make_integer_sequence<std::size_t, N>;
+template <size_t N>
+using make_index_sequence = make_integer_sequence<size_t, N>;
 #endif
 
 template <class Tuple, class F, size_t... Is>
@@ -160,6 +157,9 @@ template <class Tuple, class F> void for_each(Tuple&& tup, F&& f) {
   for_each(indexes, std::forward<Tuple>(tup), std::forward<F>(f));
 }
 
+template <typename Range>
+using value_type = remove_cvref_t<decltype(*std::declval<Range>().begin())>;
+
 template <typename Arg, FMT_ENABLE_IF(!is_like_std_string<
                                       typename std::decay<Arg>::type>::value)>
 FMT_CONSTEXPR const char* format_str_quoted(bool add_space, const Arg&) {
@@ -185,12 +185,11 @@ FMT_CONSTEXPR const char* format_str_quoted(bool add_space, const char) {
 FMT_CONSTEXPR const wchar_t* format_str_quoted(bool add_space, const wchar_t) {
   return add_space ? L" '{}'" : L"'{}'";
 }
-
-}  // namespace internal
+}  // namespace detail
 
 template <typename T> struct is_tuple_like {
   static FMT_CONSTEXPR_DECL const bool value =
-      internal::is_tuple_like_<T>::value && !internal::is_range_<T>::value;
+      detail::is_tuple_like_<T>::value && !detail::is_range_<T>::value;
 };
 
 template <typename TupleT, typename Char>
@@ -203,17 +202,17 @@ struct formatter<TupleT, Char, enable_if_t<fmt::is_tuple_like<TupleT>::value>> {
         if (formatting.add_prepostfix_space) {
           *out++ = ' ';
         }
-        out = internal::copy(formatting.delimiter, out);
+        out = detail::copy(formatting.delimiter, out);
       }
       out = format_to(out,
-                      internal::format_str_quoted(
+                      detail::format_str_quoted(
                           (formatting.add_delimiter_spaces && i > 0), v),
                       v);
       ++i;
     }
 
     formatting_tuple<Char>& formatting;
-    std::size_t& i;
+    size_t& i;
     typename std::add_lvalue_reference<decltype(
         std::declval<FormatContext>().out())>::type out;
   };
@@ -229,14 +228,14 @@ struct formatter<TupleT, Char, enable_if_t<fmt::is_tuple_like<TupleT>::value>> {
   template <typename FormatContext = format_context>
   auto format(const TupleT& values, FormatContext& ctx) -> decltype(ctx.out()) {
     auto out = ctx.out();
-    std::size_t i = 0;
-    internal::copy(formatting.prefix, out);
+    size_t i = 0;
+    detail::copy(formatting.prefix, out);
 
-    internal::for_each(values, format_each<FormatContext>{formatting, i, out});
+    detail::for_each(values, format_each<FormatContext>{formatting, i, out});
     if (formatting.add_prepostfix_space) {
       *out++ = ' ';
     }
-    internal::copy(formatting.postfix, out);
+    detail::copy(formatting.postfix, out);
 
     return ctx.out();
   }
@@ -244,14 +243,23 @@ struct formatter<TupleT, Char, enable_if_t<fmt::is_tuple_like<TupleT>::value>> {
 
 template <typename T, typename Char> struct is_range {
   static FMT_CONSTEXPR_DECL const bool value =
-      internal::is_range_<T>::value &&
-      !internal::is_like_std_string<T>::value &&
-      !std::is_convertible<T, std::basic_string<Char>>::value;
+      detail::is_range_<T>::value && !detail::is_like_std_string<T>::value &&
+      !std::is_convertible<T, std::basic_string<Char>>::value &&
+      !std::is_constructible<detail::std_string_view<Char>, T>::value;
 };
 
-template <typename RangeT, typename Char>
-struct formatter<RangeT, Char,
-                 enable_if_t<fmt::is_range<RangeT, Char>::value>> {
+template <typename T, typename Char>
+struct formatter<
+    T, Char,
+    enable_if_t<fmt::is_range<T, Char>::value
+// Workaround a bug in MSVC 2017 and earlier.
+#if !FMT_MSC_VER || FMT_MSC_VER >= 1927
+                &&
+                (has_formatter<detail::value_type<T>, format_context>::value ||
+                 detail::has_fallback_formatter<detail::value_type<T>,
+                                                format_context>::value)
+#endif
+                >> {
   formatting_range<Char> formatting;
 
   template <typename ParseContext>
@@ -260,17 +268,18 @@ struct formatter<RangeT, Char,
   }
 
   template <typename FormatContext>
-  typename FormatContext::iterator format(const RangeT& values,
-                                          FormatContext& ctx) {
-    auto out = internal::copy(formatting.prefix, ctx.out());
-    std::size_t i = 0;
-    for (auto it = values.begin(), end = values.end(); it != end; ++it) {
+  typename FormatContext::iterator format(const T& values, FormatContext& ctx) {
+    auto out = detail::copy(formatting.prefix, ctx.out());
+    size_t i = 0;
+    auto it = values.begin();
+    auto end = values.end();
+    for (; it != end; ++it) {
       if (i > 0) {
         if (formatting.add_prepostfix_space) *out++ = ' ';
-        out = internal::copy(formatting.delimiter, out);
+        out = detail::copy(formatting.delimiter, out);
       }
       out = format_to(out,
-                      internal::format_str_quoted(
+                      detail::format_str_quoted(
                           (formatting.add_delimiter_spaces && i > 0), *it),
                       *it);
       if (++i > formatting.range_length_limit) {
@@ -279,10 +288,109 @@ struct formatter<RangeT, Char,
       }
     }
     if (formatting.add_prepostfix_space) *out++ = ' ';
-    return internal::copy(formatting.postfix, out);
+    return detail::copy(formatting.postfix, out);
   }
 };
 
+template <typename Char, typename... T> struct tuple_arg_join : detail::view {
+  const std::tuple<T...>& tuple;
+  basic_string_view<Char> sep;
+
+  tuple_arg_join(const std::tuple<T...>& t, basic_string_view<Char> s)
+      : tuple{t}, sep{s} {}
+};
+
+template <typename Char, typename... T>
+struct formatter<tuple_arg_join<Char, T...>, Char> {
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename FormatContext>
+  typename FormatContext::iterator format(
+      const tuple_arg_join<Char, T...>& value, FormatContext& ctx) {
+    return format(value, ctx, detail::make_index_sequence<sizeof...(T)>{});
+  }
+
+ private:
+  template <typename FormatContext, size_t... N>
+  typename FormatContext::iterator format(
+      const tuple_arg_join<Char, T...>& value, FormatContext& ctx,
+      detail::index_sequence<N...>) {
+    return format_args(value, ctx, std::get<N>(value.tuple)...);
+  }
+
+  template <typename FormatContext>
+  typename FormatContext::iterator format_args(
+      const tuple_arg_join<Char, T...>&, FormatContext& ctx) {
+    // NOTE: for compilers that support C++17, this empty function instantiation
+    // can be replaced with a constexpr branch in the variadic overload.
+    return ctx.out();
+  }
+
+  template <typename FormatContext, typename Arg, typename... Args>
+  typename FormatContext::iterator format_args(
+      const tuple_arg_join<Char, T...>& value, FormatContext& ctx,
+      const Arg& arg, const Args&... args) {
+    using base = formatter<typename std::decay<Arg>::type, Char>;
+    auto out = ctx.out();
+    out = base{}.format(arg, ctx);
+    if (sizeof...(Args) > 0) {
+      out = std::copy(value.sep.begin(), value.sep.end(), out);
+      ctx.advance_to(out);
+      return format_args(value, ctx, args...);
+    }
+    return out;
+  }
+};
+
+/**
+  \rst
+  Returns an object that formats `tuple` with elements separated by `sep`.
+
+  **Example**::
+
+    std::tuple<int, char> t = {1, 'a'};
+    fmt::print("{}", fmt::join(t, ", "));
+    // Output: "1, a"
+  \endrst
+ */
+template <typename... T>
+FMT_CONSTEXPR tuple_arg_join<char, T...> join(const std::tuple<T...>& tuple,
+                                              string_view sep) {
+  return {tuple, sep};
+}
+
+template <typename... T>
+FMT_CONSTEXPR tuple_arg_join<wchar_t, T...> join(const std::tuple<T...>& tuple,
+                                                 wstring_view sep) {
+  return {tuple, sep};
+}
+
+/**
+  \rst
+  Returns an object that formats `initializer_list` with elements separated by
+  `sep`.
+
+  **Example**::
+
+    fmt::print("{}", fmt::join({1, 2, 3}, ", "));
+    // Output: "1, 2, 3"
+  \endrst
+ */
+template <typename T>
+arg_join<const T*, const T*, char> join(std::initializer_list<T> list,
+                                        string_view sep) {
+  return join(std::begin(list), std::end(list), sep);
+}
+
+template <typename T>
+arg_join<const T*, const T*, wchar_t> join(std::initializer_list<T> list,
+                                           wstring_view sep) {
+  return join(std::begin(list), std::end(list), sep);
+}
+
 FMT_END_NAMESPACE
 
 #endif  // FMT_RANGES_H_