Merge "Upgrade fmtlib to 7.1.0" am: 55900b4aee am: 3a9202ea4e am: ea11d1de5a am: 4f0d8dd00a

Original change: https://android-review.googlesource.com/c/platform/external/fmtlib/+/1479386

Change-Id: Iedca3cf7b8d81285b1fcef1c4adccc841f4ccad2

64 files changed, 4769 insertions, 2535 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index d8dd862c..f21cf456 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,10 +1,8 @@
-cmake_minimum_required(VERSION 3.1.0)
+cmake_minimum_required(VERSION 3.1...3.18)
 
-# Use newer policies if available, up to most recent tested version of CMake.
-if(${CMAKE_VERSION} VERSION_LESS 3.11)
+# Fallback for using newer policies on CMake <3.12.
+if(${CMAKE_VERSION} VERSION_LESS 3.12)
   cmake_policy(VERSION ${CMAKE_MAJOR_VERSION}.${CMAKE_MINOR_VERSION})
-else()
-  cmake_policy(VERSION 3.11)
 endif()
 
 # Determine if fmt is built as a subproject (using add_subdirectory)
@@ -24,15 +22,23 @@ function(join result_var)
   set(${result_var} "${result}" PARENT_SCOPE)
 endfunction()
 
+include(CMakeParseArguments)
+
 # Sets a cache variable with a docstring joined from multiple arguments:
 #   set(<variable> <value>... CACHE <type> <docstring>...)
 # This allows splitting a long docstring for readability.
 function(set_verbose)
-  cmake_parse_arguments(SET_VERBOSE "" "" "CACHE" ${ARGN})
-  list(GET SET_VERBOSE_CACHE 0 type)
-  list(REMOVE_AT SET_VERBOSE_CACHE 0)
-  join(doc ${SET_VERBOSE_CACHE})
-  set(${SET_VERBOSE_UNPARSED_ARGUMENTS} CACHE ${type} ${doc})
+  # cmake_parse_arguments is broken in CMake 3.4 (cannot parse CACHE) so use
+  # list instead.
+  list(GET ARGN 0 var)
+  list(REMOVE_AT ARGN 0)
+  list(GET ARGN 0 val)
+  list(REMOVE_AT ARGN 0)
+  list(REMOVE_AT ARGN 0)
+  list(GET ARGN 0 type)
+  list(REMOVE_AT ARGN 0)
+  join(doc ${ARGN})
+  set(${var} ${val} CACHE ${type} ${doc})
 endfunction()
 
 # Set the default CMAKE_BUILD_TYPE to Release.
@@ -47,8 +53,8 @@ endif ()
 project(FMT CXX)
 include(GNUInstallDirs)
 set_verbose(FMT_INC_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE STRING
-            "Installation directory for include files, a relative path "
-            "that will be joined to ${CMAKE_INSTALL_PREFIX}, or an arbitrary absolute path.")
+            "Installation directory for include files, a relative path that "
+            "will be joined with ${CMAKE_INSTALL_PREFIX} or an absolute path.")
 
 option(FMT_PEDANTIC "Enable extra warnings and expensive tests." OFF)
 option(FMT_WERROR "Halt the compilation with an error on compiler warnings."
@@ -101,11 +107,11 @@ if (CMAKE_CXX_COMPILER_ID MATCHES "GNU")
       -Wold-style-cast -Wundef
       -Wredundant-decls -Wwrite-strings -Wpointer-arith
       -Wcast-qual -Wformat=2 -Wmissing-include-dirs
-      -Wcast-align -Wnon-virtual-dtor
+      -Wcast-align
       -Wctor-dtor-privacy -Wdisabled-optimization
       -Winvalid-pch -Woverloaded-virtual
-      -Wconversion -Wswitch-enum
-      -Wno-ctor-dtor-privacy -Wno-format-nonliteral -Wno-shadow)
+      -Wconversion -Wswitch-enum -Wundef
+      -Wno-ctor-dtor-privacy -Wno-format-nonliteral)
   if (NOT CMAKE_CXX_COMPILER_VERSION VERSION_LESS 4.6)
       set(PEDANTIC_COMPILE_FLAGS ${PEDANTIC_COMPILE_FLAGS} -Wnoexcept
          -Wno-dangling-else -Wno-unused-local-typedefs)
@@ -123,8 +129,8 @@ if (CMAKE_CXX_COMPILER_ID MATCHES "GNU")
 endif ()
 
 if (CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-  set(PEDANTIC_COMPILE_FLAGS -Wall -Wextra -pedantic -Wconversion
-      -Wno-sign-conversion -Wdeprecated -Wweak-vtables)
+  set(PEDANTIC_COMPILE_FLAGS -Wall -Wextra -pedantic -Wconversion -Wundef
+      -Wdeprecated -Wweak-vtables)
   check_cxx_compiler_flag(-Wzero-as-null-pointer-constant HAS_NULLPTR_WARNING)
   if (HAS_NULLPTR_WARNING)
     set(PEDANTIC_COMPILE_FLAGS ${PEDANTIC_COMPILE_FLAGS}
@@ -193,7 +199,10 @@ if (HAVE_STRTOD_L)
 endif ()
 
 if (MINGW)
-  target_compile_options(fmt PUBLIC "-Wa,-mbig-obj")
+  check_cxx_compiler_flag("Wa,-mbig-obj" FMT_HAS_MBIG_OBJ)
+  if (${FMT_HAS_MBIG_OBJ})
+    target_compile_options(fmt PUBLIC "-Wa,-mbig-obj")
+  endif()
 endif ()
 
 if (FMT_WERROR)
@@ -223,7 +232,8 @@ if (CMAKE_BUILD_TYPE STREQUAL "Debug")
 endif ()
 
 if (BUILD_SHARED_LIBS)
-  if (UNIX AND NOT APPLE AND NOT ${CMAKE_SYSTEM_NAME} MATCHES "SunOS" AND NOT EMSCRIPTEN)
+  if (UNIX AND NOT APPLE AND NOT ${CMAKE_SYSTEM_NAME} MATCHES "SunOS" AND
+      NOT EMSCRIPTEN)
     # Fix rpmlint warning:
     # unused-direct-shlib-dependency /usr/lib/libformat.so.1.1.0 /lib/libm.so.6.
     target_link_libraries(fmt -Wl,--as-needed)
@@ -248,20 +258,22 @@ target_include_directories(fmt-header-only INTERFACE
 if (FMT_INSTALL)
   include(CMakePackageConfigHelpers)
   set_verbose(FMT_CMAKE_DIR ${CMAKE_INSTALL_LIBDIR}/cmake/fmt CACHE STRING
-              "Installation directory for cmake files, a relative path "
-              "that will be joined to ${CMAKE_INSTALL_PREFIX}, or an arbitrary absolute path.")
+              "Installation directory for cmake files, a relative path that "
+              "will be joined with ${CMAKE_INSTALL_PREFIX} or an absolute "
+              "path.")
   set(version_config ${PROJECT_BINARY_DIR}/fmt-config-version.cmake)
   set(project_config ${PROJECT_BINARY_DIR}/fmt-config.cmake)
   set(pkgconfig ${PROJECT_BINARY_DIR}/fmt.pc)
   set(targets_export_name fmt-targets)
 
   set_verbose(FMT_LIB_DIR ${CMAKE_INSTALL_LIBDIR} CACHE STRING
-              "Installation directory for libraries, a relative path "
-              "that will be joined to ${CMAKE_INSTALL_PREFIX}, or an arbitrary absolute path.")
+              "Installation directory for libraries, a relative path that "
+              "will be joined to ${CMAKE_INSTALL_PREFIX} or an absolute path.")
 
   set_verbose(FMT_PKGCONFIG_DIR ${CMAKE_INSTALL_LIBDIR}/pkgconfig CACHE PATH
-              "Installation directory for pkgconfig (.pc) files, a relative path "
-              "that will be joined to ${CMAKE_INSTALL_PREFIX}, or an arbitrary absolute path.")
+              "Installation directory for pkgconfig (.pc) files, a relative "
+              "path that will be joined with ${CMAKE_INSTALL_PREFIX} or an "
+              "absolute path.")
 
   # Generate the version, config and target files into the build directory.
   write_basic_package_version_file(
@@ -318,6 +330,12 @@ endif ()
 # Control fuzzing independent of the unit tests.
 if (FMT_FUZZ)
   add_subdirectory(test/fuzzing)
+
+  # The FMT_FUZZ macro is used to prevent resource exhaustion in fuzzing
+  # mode and make fuzzing practically possible. It is similar to
+  # FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION but uses a different name to
+  # avoid interfering with fuzzing of projects that use {fmt}.
+  # See also https://llvm.org/docs/LibFuzzer.html#fuzzer-friendly-build-mode.
   target_compile_definitions(fmt PUBLIC FMT_FUZZ)
 endif ()
 
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 3532bd11..b82f1450 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -14,4 +14,7 @@ exceptions:
 * snake_case should be used instead of UpperCamelCase for function and type
   names
 
+All documentation must adhere to the [Google Developer Documentation Style
+Guide](https://developers.google.com/style).
+
 Thanks for contributing!
diff --git a/ChangeLog.rst b/ChangeLog.rst
index 3e24c7d9..f9a2d8eb 100644
--- a/ChangeLog.rst
+++ b/ChangeLog.rst
@@ -1,3 +1,411 @@
+7.1.0 - 2020-10-25
+------------------
+
+* Switched from `Grisu3
+  <https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf>`_
+  to `Dragonbox <https://github.com/jk-jeon/dragonbox>`_ for the default
+  floating-point formatting which gives the shortest decimal representation
+  with round-trip guarantee and correct rounding
+  (`#1882 <https://github.com/fmtlib/fmt/pull/1882>`_,
+  `#1887 <https://github.com/fmtlib/fmt/pull/1887>`_,
+  `#1894 <https://github.com/fmtlib/fmt/pull/1894>`_). This makes {fmt} up to
+  20-30x faster than common implementations of ``std::ostringstream`` and
+  ``sprintf`` on `dtoa-benchmark <https://github.com/fmtlib/dtoa-benchmark>`_
+  and faster than double-conversion and Ryū:
+
+  .. image:: https://user-images.githubusercontent.com/576385/
+             95684665-11719600-0ba8-11eb-8e5b-972ff4e49428.png
+
+  It is possible to get even better performance at the cost of larger binary
+  size by compiling with the ``FMT_USE_FULL_CACHE_DRAGONBOX`` macro set to 1.
+
+  Thanks `@jk-jeon (Junekey Jeon) <https://github.com/jk-jeon>`_.
+
+* Added an experimental unsynchronized file output API which, together with
+  `format string compilation <https://fmt.dev/latest/api.html#compile-api>`_,
+  can give `5-9 times speed up compared to fprintf
+  <https://www.zverovich.net/2020/08/04/optimal-file-buffer-size.html>`_
+  on common platforms (`godbolt <https://godbolt.org/z/nsTcG8>`__):
+
+  .. code:: c++
+
+     #include <fmt/os.h>
+
+     int main() {
+       auto f = fmt::output_file("guide");
+       f.print("The answer is {}.", 42);
+     }
+
+* Added a formatter for ``std::chrono::time_point<system_clock>``
+  (`#1819 <https://github.com/fmtlib/fmt/issues/1819>`_,
+  `#1837 <https://github.com/fmtlib/fmt/pull/1837>`_). For example
+  (`godbolt <https://godbolt.org/z/c4M6fh>`__):
+
+  .. code:: c++
+
+     #include <fmt/chrono.h>
+
+     int main() {
+       auto now = std::chrono::system_clock::now();
+       fmt::print("The time is {:%H:%M:%S}.\n", now);
+     }
+
+  Thanks `@adamburgess (Adam Burgess) <https://github.com/adamburgess>`_.
+
+* Added support for ranges with non-const ``begin``/``end`` to ``fmt::join``
+  (`#1784 <https://github.com/fmtlib/fmt/issues/1784>`_,
+  `#1786 <https://github.com/fmtlib/fmt/pull/1786>`_). For example
+  (`godbolt <https://godbolt.org/z/jP63Tv>`__):
+
+  .. code:: c++
+
+     #include <fmt/ranges.h>
+     #include <range/v3/view/filter.hpp>
+
+     int main() {
+       using std::literals::string_literals::operator""s;
+       auto strs = std::array{"a"s, "bb"s, "ccc"s};
+       auto range = strs | ranges::views::filter(
+         [] (const std::string &x) { return x.size() != 2; }
+       );
+       fmt::print("{}\n", fmt::join(range, ""));
+     }
+
+  prints "accc".
+
+  Thanks `@tonyelewis (Tony E Lewis) <https://github.com/tonyelewis>`_.
+
+* Added a ``memory_buffer::append`` overload that takes a range
+  (`#1806 <https://github.com/fmtlib/fmt/pull/1806>`_).
+  Thanks `@BRevzin (Barry Revzin) <https://github.com/BRevzin>`_.
+
+* Improved handling of single code units in ``FMT_COMPILE``. For example:
+
+  .. code:: c++
+
+     #include <fmt/compile.h>
+
+     char* f(char* buf) {
+       return fmt::format_to(buf, FMT_COMPILE("x{}"), 42);
+     }
+
+  compiles to just (`godbolt <https://godbolt.org/z/5vncz3>`__):
+
+  .. code:: asm
+
+     _Z1fPc:
+       movb $120, (%rdi)
+       xorl %edx, %edx
+       cmpl $42, _ZN3fmt2v76detail10basic_dataIvE23zero_or_powers_of_10_32E+8(%rip)
+       movl $3, %eax
+       seta %dl
+       subl %edx, %eax
+       movzwl _ZN3fmt2v76detail10basic_dataIvE6digitsE+84(%rip), %edx
+       cltq
+       addq %rdi, %rax
+       movw %dx, -2(%rax)
+       ret
+
+  Here a single ``mov`` instruction writes ``'x'`` (``$120``) to the output
+  buffer.
+
+* Added dynamic width support to format string compilation
+  (`#1809 <https://github.com/fmtlib/fmt/issues/1809>`_).
+
+* Improved error reporting for unformattable types: now you'll get the type name
+  directly in the error message instead of the note:
+
+  .. code:: c++
+
+     #include <fmt/core.h>
+
+     struct how_about_no {};
+
+     int main() {
+       fmt::print("{}", how_about_no());
+     }
+
+  Error (`godbolt <https://godbolt.org/z/GoxM4e>`__):
+
+  ``fmt/core.h:1438:3: error: static_assert failed due to requirement
+  'fmt::v7::formattable<how_about_no>()' "Cannot format an argument.
+  To make type T formattable provide a formatter<T> specialization:
+  https://fmt.dev/latest/api.html#udt"
+  ...``
+
+* Added the `make_args_checked <https://fmt.dev/7.1.0/api.html#argument-lists>`_
+  function template that allows you to write formatting functions with
+  compile-time format string checks and avoid binary code bloat
+  (`godbolt <https://godbolt.org/z/PEf9qr>`__):
+
+  .. code:: c++
+
+     void vlog(const char* file, int line, fmt::string_view format,
+               fmt::format_args args) {
+       fmt::print("{}: {}: ", file, line);
+       fmt::vprint(format, args);
+     }
+
+     template <typename S, typename... Args>
+     void log(const char* file, int line, const S& format, Args&&... args) {
+       vlog(file, line, format,
+           fmt::make_args_checked<Args...>(format, args...));
+     }
+
+     #define MY_LOG(format, ...) \
+       log(__FILE__, __LINE__, FMT_STRING(format), __VA_ARGS__)
+
+     MY_LOG("invalid squishiness: {}", 42);
+
+* Replaced ``snprintf`` fallback with a faster internal IEEE 754 ``float`` and
+  ``double`` formatter for arbitrary precision. For example
+  (`godbolt <https://godbolt.org/z/dPhWvj>`__):
+
+  .. code:: c++
+
+     #include <fmt/core.h>
+
+     int main() {
+       fmt::print("{:.500}\n", 4.9406564584124654E-324);
+     }
+
+  prints
+
+  ``4.9406564584124654417656879286822137236505980261432476442558568250067550727020875186529983636163599237979656469544571773092665671035593979639877479601078187812630071319031140452784581716784898210368871863605699873072305000638740915356498438731247339727316961514003171538539807412623856559117102665855668676818703956031062493194527159149245532930545654440112748012970999954193198940908041656332452475714786901472678015935523861155013480352649347201937902681071074917033322268447533357208324319360923829e-324``.
+
+* Made ``format_to_n`` and ``formatted_size`` part of the `core API
+  <https://fmt.dev/latest/api.html#core-api>`__
+  (`godbolt <https://godbolt.org/z/sPjY1K>`__):
+
+  .. code:: c++
+
+     #include <fmt/core.h>
+
+     int main() {
+       char buffer[10];
+       auto result = fmt::format_to_n(buffer, sizeof(buffer), "{}", 42);
+     }
+
+* Added ``fmt::format_to_n`` overload with format string compilation
+  (`#1764 <https://github.com/fmtlib/fmt/issues/1764>`_,
+  `#1767 <https://github.com/fmtlib/fmt/pull/1767>`_,
+  `#1869 <https://github.com/fmtlib/fmt/pull/1869>`_). For example
+  (`godbolt <https://godbolt.org/z/93h86q>`__):
+
+  .. code:: c++
+
+     #include <fmt/compile.h>
+
+     int main() {
+       char buffer[8];
+       fmt::format_to_n(buffer, sizeof(buffer), FMT_COMPILE("{}"), 42);
+     }
+
+  Thanks `@Kurkin (Dmitry Kurkin) <https://github.com/Kurkin>`_,
+  `@alexezeder (Alexey Ochapov) <https://github.com/alexezeder>`_.
+
+* Added ``fmt::format_to`` overload that take ``text_style``
+  (`#1593 <https://github.com/fmtlib/fmt/issues/1593>`_,
+  `#1842 <https://github.com/fmtlib/fmt/issues/1842>`_,
+  `#1843 <https://github.com/fmtlib/fmt/pull/1843>`_). For example
+  (`godbolt <https://godbolt.org/z/91153r>`__):
+
+  .. code:: c++
+
+     #include <fmt/color.h>
+
+     int main() {
+       std::string out;
+       fmt::format_to(std::back_inserter(out),
+                      fmt::emphasis::bold | fg(fmt::color::red),
+                      "The answer is {}.", 42);
+     }
+
+  Thanks `@Naios (Denis Blank) <https://github.com/Naios>`_.
+
+* Made the ``#`` specifier emit trailing zeros in addition to the decimal point
+  (`#1797 <https://github.com/fmtlib/fmt/issues/1797>`_). For example
+  (`godbolt <https://godbolt.org/z/bhdcW9>`__):
+
+  .. code:: c++
+
+     #include <fmt/core.h>
+
+     int main() {
+       fmt::print("{:#.2g}", 0.5);
+     }
+
+  prints ``0.50``.
+
+* Changed the default floating point format to not include ``.0`` for
+  consistency with ``std::format`` and ``std::to_chars``
+  (`#1893 <https://github.com/fmtlib/fmt/issues/1893>`_,
+  `#1943 <https://github.com/fmtlib/fmt/issues/1943>`_). It is possible to get
+  the decimal point and trailing zero with the ``#`` specifier.
+
+* Fixed an issue with floating-point formatting that could result in addition of
+  a non-significant trailing zero in rare cases e.g. ``1.00e-34`` instead of
+  ``1.0e-34`` (`#1873 <https://github.com/fmtlib/fmt/issues/1873>`_,
+  `#1917 <https://github.com/fmtlib/fmt/issues/1917>`_).
+
+* Made ``fmt::to_string`` fallback on ``ostream`` insertion operator if
+  the ``formatter`` specialization is not provided
+  (`#1815 <https://github.com/fmtlib/fmt/issues/1815>`_,
+  `#1829 <https://github.com/fmtlib/fmt/pull/1829>`_).
+  Thanks `@alexezeder (Alexey Ochapov) <https://github.com/alexezeder>`_.
+
+* Added support for the append mode to the experimental file API and
+  improved ``fcntl.h`` detection.
+  (`#1847 <https://github.com/fmtlib/fmt/pull/1847>`_,
+  `#1848 <https://github.com/fmtlib/fmt/pull/1848>`_).
+  Thanks `@t-wiser <https://github.com/t-wiser>`_.
+
+* Fixed handling of types that have both an implicit conversion operator and
+  an overloaded ``ostream`` insertion operator
+  (`#1766 <https://github.com/fmtlib/fmt/issues/1766>`_).
+
+* Fixed a slicing issue in an internal iterator type
+  (`#1822 <https://github.com/fmtlib/fmt/pull/1822>`_).
+  Thanks `@BRevzin (Barry Revzin) <https://github.com/BRevzin>`_.
+
+* Fixed an issue in locale-specific integer formatting
+  (`#1927 <https://github.com/fmtlib/fmt/issues/1927>`_).
+
+* Fixed handling of exotic code unit types
+  (`#1870 <https://github.com/fmtlib/fmt/issues/1870>`_,
+  `#1932 <https://github.com/fmtlib/fmt/issues/1932>`_).
+
+* Improved ``FMT_ALWAYS_INLINE``
+  (`#1878 <https://github.com/fmtlib/fmt/pull/1878>`_).
+  Thanks `@jk-jeon (Junekey Jeon) <https://github.com/jk-jeon>`_.
+
+* Removed dependency on ``windows.h``
+  (`#1900 <https://github.com/fmtlib/fmt/pull/1900>`_).
+  Thanks `@bernd5 (Bernd Baumanns) <https://github.com/bernd5>`_.
+
+* Optimized counting of decimal digits on MSVC
+  (`#1890 <https://github.com/fmtlib/fmt/pull/1890>`_).
+  Thanks `@mwinterb <https://github.com/mwinterb>`_.
+
+* Improved documentation
+  (`#1772 <https://github.com/fmtlib/fmt/issues/1772>`_,
+  `#1775 <https://github.com/fmtlib/fmt/pull/1775>`_,
+  `#1792 <https://github.com/fmtlib/fmt/pull/1792>`_,
+  `#1838 <https://github.com/fmtlib/fmt/pull/1838>`_,
+  `#1888 <https://github.com/fmtlib/fmt/pull/1888>`_,
+  `#1918 <https://github.com/fmtlib/fmt/pull/1918>`_,
+  `#1939 <https://github.com/fmtlib/fmt/pull/1939>`_).
+  Thanks `@leolchat (Léonard Gérard) <https://github.com/leolchat>`_,
+  `@pepsiman (Malcolm Parsons) <https://github.com/pepsiman>`_,
+  `@Klaim (Joël Lamotte) <https://github.com/Klaim>`_,
+  `@ravijanjam (Ravi J) <https://github.com/ravijanjam>`_,
+  `@francesco-st <https://github.com/francesco-st>`_,
+  `@udnaan (Adnan) <https://github.com/udnaan>`_.
+
+* Added the ``FMT_REDUCE_INT_INSTANTIATIONS`` CMake option that reduces the
+  binary code size at the cost of some integer formatting performance. This can
+  be useful for extremely memory-constrained embedded systems
+  (`#1778 <https://github.com/fmtlib/fmt/issues/1778>`_,
+  `#1781 <https://github.com/fmtlib/fmt/pull/1781>`_).
+  Thanks `@kammce (Khalil Estell) <https://github.com/kammce>`_.
+
+* Added the ``FMT_USE_INLINE_NAMESPACES`` macro to control usage of inline
+  namespaces (`#1945 <https://github.com/fmtlib/fmt/pull/1945>`_).
+  Thanks `@darklukee <https://github.com/darklukee>`_.
+
+* Improved build configuration
+  (`#1760 <https://github.com/fmtlib/fmt/pull/1760>`_,
+  `#1770 <https://github.com/fmtlib/fmt/pull/1770>`_,
+  `#1779 <https://github.com/fmtlib/fmt/issues/1779>`_,
+  `#1783 <https://github.com/fmtlib/fmt/pull/1783>`_,
+  `#1823 <https://github.com/fmtlib/fmt/pull/1823>`_).
+  Thanks `@dvetutnev (Dmitriy Vetutnev) <https://github.com/dvetutnev>`_,
+  `@xvitaly (Vitaly Zaitsev) <https://github.com/xvitaly>`_,
+  `@tambry (Raul Tambre) <https://github.com/tambry>`_,
+  `@medithe <https://github.com/medithe>`_,
+  `@martinwuehrer (Martin Wührer) <https://github.com/martinwuehrer>`_.
+
+* Fixed various warnings and compilation issues
+  (`#1790 <https://github.com/fmtlib/fmt/pull/1790>`_,
+  `#1802 <https://github.com/fmtlib/fmt/pull/1802>`_,
+  `#1808 <https://github.com/fmtlib/fmt/pull/1808>`_,
+  `#1810 <https://github.com/fmtlib/fmt/issues/1810>`_,
+  `#1811 <https://github.com/fmtlib/fmt/issues/1811>`_,
+  `#1812 <https://github.com/fmtlib/fmt/pull/1812>`_,
+  `#1814 <https://github.com/fmtlib/fmt/pull/1814>`_,
+  `#1816 <https://github.com/fmtlib/fmt/pull/1816>`_,
+  `#1817 <https://github.com/fmtlib/fmt/pull/1817>`_,
+  `#1818 <https://github.com/fmtlib/fmt/pull/1818>`_,
+  `#1825 <https://github.com/fmtlib/fmt/issues/1825>`_,
+  `#1836 <https://github.com/fmtlib/fmt/pull/1836>`_,
+  `#1855 <https://github.com/fmtlib/fmt/pull/1855>`_,
+  `#1856 <https://github.com/fmtlib/fmt/pull/1856>`_,
+  `#1860 <https://github.com/fmtlib/fmt/pull/1860>`_,
+  `#1877 <https://github.com/fmtlib/fmt/pull/1877>`_,
+  `#1879 <https://github.com/fmtlib/fmt/pull/1879>`_,
+  `#1880 <https://github.com/fmtlib/fmt/pull/1880>`_,
+  `#1896 <https://github.com/fmtlib/fmt/issues/1896>`_,
+  `#1897 <https://github.com/fmtlib/fmt/pull/1897>`_,
+  `#1898 <https://github.com/fmtlib/fmt/pull/1898>`_,
+  `#1904 <https://github.com/fmtlib/fmt/issues/1904>`_,
+  `#1908 <https://github.com/fmtlib/fmt/pull/1908>`_,
+  `#1911 <https://github.com/fmtlib/fmt/issues/1911>`_,
+  `#1912 <https://github.com/fmtlib/fmt/issues/1912>`_,
+  `#1928 <https://github.com/fmtlib/fmt/issues/1928>`_,
+  `#1929 <https://github.com/fmtlib/fmt/pull/1929>`_,
+  `#1935 <https://github.com/fmtlib/fmt/issues/1935>`_
+  `#1937 <https://github.com/fmtlib/fmt/pull/1937>`_,
+  `#1942 <https://github.com/fmtlib/fmt/pull/1942>`_,
+  `#1949 <https://github.com/fmtlib/fmt/issues/1949>`_).
+  Thanks `@TheQwertiest <https://github.com/TheQwertiest>`_,
+  `@medithe <https://github.com/medithe>`_,
+  `@martinwuehrer (Martin Wührer) <https://github.com/martinwuehrer>`_,
+  `@n16h7hunt3r <https://github.com/n16h7hunt3r>`_,
+  `@Othereum (Seokjin Lee) <https://github.com/Othereum>`_,
+  `@gsjaardema (Greg Sjaardema) <https://github.com/gsjaardema>`_,
+  `@AlexanderLanin (Alexander Lanin) <https://github.com/AlexanderLanin>`_,
+  `@gcerretani (Giovanni Cerretani) <https://github.com/gcerretani>`_,
+  `@chronoxor (Ivan Shynkarenka) <https://github.com/chronoxor>`_,
+  `@noizefloor (Jan Schwers) <https://github.com/noizefloor>`_,
+  `@akohlmey (Axel Kohlmeyer) <https://github.com/akohlmey>`_,
+  `@jk-jeon (Junekey Jeon) <https://github.com/jk-jeon>`_,
+  `@rimathia <https://github.com/rimathia>`_,
+  `@rglarix (Riccardo Ghetta (larix)) <https://github.com/rglarix>`_,
+  `@moiwi <https://github.com/moiwi>`_,
+  `@heckad (Kazantcev Andrey) <https://github.com/heckad>`_,
+  `@MarcDirven <https://github.com/MarcDirven>`_.
+  `@BartSiwek (Bart Siwek) <https://github.com/BartSiwek>`_,
+  `@darklukee <https://github.com/darklukee>`_.
+
+7.0.3 - 2020-08-06
+------------------
+
+* Worked around broken ``numeric_limits`` for 128-bit integers
+  (`#1787 <https://github.com/fmtlib/fmt/issues/1787>`_).
+
+* Added error reporting on missing named arguments
+  (`#1796 <https://github.com/fmtlib/fmt/issues/1796>`_).
+
+* Stopped using 128-bit integers with clang-cl
+  (`#1800 <https://github.com/fmtlib/fmt/pull/1800>`_).
+  Thanks `@Kingcom <https://github.com/Kingcom>`_.
+
+* Fixed issues in locale-specific integer formatting
+  (`#1782 <https://github.com/fmtlib/fmt/issues/1782>`_,
+  `#1801 <https://github.com/fmtlib/fmt/issues/1801>`_).
+
+7.0.2 - 2020-07-29
+------------------
+
+* Worked around broken ``numeric_limits`` for 128-bit integers
+  (`#1725 <https://github.com/fmtlib/fmt/issues/1725>`_).
+
+* Fixed compatibility with CMake 3.4
+  (`#1779 <https://github.com/fmtlib/fmt/issues/1779>`_).
+
+* Fixed handling of digit separators in locale-specific formatting
+  (`#1782 <https://github.com/fmtlib/fmt/issues/1782>`_).
+
 7.0.1 - 2020-07-07
 ------------------
 
@@ -20,7 +428,7 @@
   <http://www.zverovich.net/2020/05/21/reducing-library-size.html>`_.
 
 * Added a simpler and more efficient `format string compilation API
-  <https://fmt.dev/dev/api.html#compile-api>`_:
+  <https://fmt.dev/7.0.0/api.html#compile-api>`_:
 
   .. code:: c++
 
@@ -159,7 +567,7 @@
 
   Thanks `@BRevzin (Barry Revzin) <https://github.com/BRevzin>`_.
 
-* Added support for named args, ``clear`` and ``reserve`` to
+* Added support for named arguments, ``clear`` and ``reserve`` to
   ``dynamic_format_arg_store``
   (`#1655 <https://github.com/fmtlib/fmt/issues/1655>`_,
   `#1663 <https://github.com/fmtlib/fmt/pull/1663>`_,
@@ -1676,7 +2084,7 @@
 
 * Implemented ``constexpr`` parsing of format strings and `compile-time format
   string checks
-  <https://fmt.dev/dev/api.html#compile-time-format-string-checks>`_. For
+  <https://fmt.dev/latest/api.html#compile-time-format-string-checks>`_. For
   example
 
   .. code:: c++
@@ -1737,7 +2145,7 @@
             throw format_error("invalid specifier");
 
 * Added `iterator support
-  <https://fmt.dev/dev/api.html#output-iterator-support>`_:
+  <https://fmt.dev/latest/api.html#output-iterator-support>`_:
 
   .. code:: c++
 
@@ -1748,7 +2156,7 @@
      fmt::format_to(std::back_inserter(out), "{}", 42);
 
 * Added the `format_to_n
-  <https://fmt.dev/dev/api.html#_CPPv2N3fmt11format_to_nE8OutputItNSt6size_tE11string_viewDpRK4Args>`_
+  <https://fmt.dev/latest/api.html#_CPPv2N3fmt11format_to_nE8OutputItNSt6size_tE11string_viewDpRK4Args>`_
   function that restricts the output to the specified number of characters
   (`#298 <https://github.com/fmtlib/fmt/issues/298>`_):
 
@@ -1759,7 +2167,7 @@
      // out == "1234" (without terminating '\0')
 
 * Added the `formatted_size
-  <https://fmt.dev/dev/api.html#_CPPv2N3fmt14formatted_sizeE11string_viewDpRK4Args>`_
+  <https://fmt.dev/latest/api.html#_CPPv2N3fmt14formatted_sizeE11string_viewDpRK4Args>`_
   function for computing the output size:
 
   .. code:: c++
@@ -1769,7 +2177,7 @@
      auto size = fmt::formatted_size("{}", 12345); // size == 5
 
 * Improved compile times by reducing dependencies on standard headers and
-  providing a lightweight `core API <https://fmt.dev/dev/api.html#core-api>`_:
+  providing a lightweight `core API <https://fmt.dev/latest/api.html#core-api>`_:
 
   .. code:: c++
 
@@ -1781,7 +2189,7 @@
   <https://github.com/fmtlib/fmt#compile-time-and-code-bloat>`_.
 
 * Added the `make_format_args
-  <https://fmt.dev/dev/api.html#_CPPv2N3fmt16make_format_argsEDpRK4Args>`_
+  <https://fmt.dev/latest/api.html#_CPPv2N3fmt16make_format_argsEDpRK4Args>`_
   function for capturing formatting arguments:
 
   .. code:: c++
@@ -1863,7 +2271,7 @@
      fmt::format("{} {two}", 1, fmt::arg("two", 2));
 
 * Removed the write API in favor of the `format API
-  <https://fmt.dev/dev/api.html#format-api>`_ with compile-time handling of
+  <https://fmt.dev/latest/api.html#format-api>`_ with compile-time handling of
   format strings.
 
 * Disallowed formatting of multibyte strings into a wide character target
diff --git a/METADATA b/METADATA
index 7cb87e9e..c9052184 100644
--- a/METADATA
+++ b/METADATA
@@ -9,11 +9,11 @@ third_party {
     type: GIT
     value: "https://github.com/fmtlib/fmt.git"
   }
-  version: "7.0.1"
+  version: "7.1.0"
   license_type: NOTICE
   last_upgrade_date {
     year: 2020
-    month: 7
-    day: 10
+    month: 10
+    day: 28
   }
 }
diff --git a/README.rst b/README.rst
index 722a65eb..acddc70e 100644
--- a/README.rst
+++ b/README.rst
@@ -7,48 +7,57 @@
 .. image:: https://ci.appveyor.com/api/projects/status/ehjkiefde6gucy1v
    :target: https://ci.appveyor.com/project/vitaut/fmt
 
-.. image:: https://oss-fuzz-build-logs.storage.googleapis.com/badges/libfmt.svg
-   :alt: fmt is continuously fuzzed att oss-fuzz
+.. image:: https://oss-fuzz-build-logs.storage.googleapis.com/badges/fmt.svg
+   :alt: fmt is continuously fuzzed at oss-fuzz
    :target: https://bugs.chromium.org/p/oss-fuzz/issues/list?\
             colspec=ID%20Type%20Component%20Status%20Proj%20Reported%20Owner%20\
-            Summary&q=proj%3Dlibfmt&can=1
+            Summary&q=proj%3Dfmt&can=1
 
 .. image:: https://img.shields.io/badge/stackoverflow-fmt-blue.svg
    :alt: Ask questions at StackOverflow with the tag fmt
    :target: https://stackoverflow.com/questions/tagged/fmt
 
-**{fmt}** is an open-source formatting library for C++.
-It can be used as a safe and fast alternative to (s)printf and iostreams.
+**{fmt}** is an open-source formatting library providing a fast and safe
+alternative to C stdio and C++ iostreams.
 
-`Documentation <https://fmt.dev/latest/>`__
+If you like this project, please consider donating to BYSOL,
+an initiative to help victims of political repressions in Belarus:
+https://www.facebook.com/donate/759400044849707/108388587646909/.
+
+`Documentation <https://fmt.dev>`__
 
 Q&A: ask questions on `StackOverflow with the tag fmt
 <https://stackoverflow.com/questions/tagged/fmt>`_.
 
+Try {fmt} in `Compiler Explorer <https://godbolt.org/z/Eq5763>`_.
+
 Features
 --------
 
-* Simple `format API <https://fmt.dev/dev/api.html>`_ with positional arguments
+* Simple `format API <https://fmt.dev/latest/api.html>`_ with positional arguments
   for localization
 * Implementation of `C++20 std::format
   <https://en.cppreference.com/w/cpp/utility/format>`__
-* `Format string syntax <https://fmt.dev/dev/syntax.html>`_ similar to the one
-  of Python's
+* `Format string syntax <https://fmt.dev/latest/syntax.html>`_ similar to Python's
   `format <https://docs.python.org/3/library/stdtypes.html#str.format>`_
+* Fast IEEE 754 floating-point formatter with correct rounding, shortness and
+  round-trip guarantees
 * Safe `printf implementation
-  <https://fmt.dev/latest/api.html#printf-formatting>`_ including
-  the POSIX extension for positional arguments
-* Extensibility: support for user-defined types
+  <https://fmt.dev/latest/api.html#printf-formatting>`_ including the POSIX
+  extension for positional arguments
+* Extensibility: `support for user-defined types
+  <https://fmt.dev/latest/api.html#formatting-user-defined-types>`_
 * High performance: faster than common standard library implementations of
-  `printf <https://en.cppreference.com/w/cpp/io/c/fprintf>`_,
-  iostreams, ``to_string`` and ``to_chars``, see `Speed tests`_ and
-  `Converting a hundred million integers to strings per second
+  ``(s)printf``, iostreams, ``to_string`` and ``to_chars``, see `Speed tests`_
+  and `Converting a hundred million integers to strings per second
   <http://www.zverovich.net/2020/06/13/fast-int-to-string-revisited.html>`_
-* Small code size both in terms of source code (the minimum configuration
-  consists of just three header files, ``core.h``, ``format.h`` and
-  ``format-inl.h``) and compiled code. See `Compile time and code bloat`_
-* Reliability: the library has an extensive set of `unit tests
-  <https://github.com/fmtlib/fmt/tree/master/test>`_ and is continuously fuzzed
+* Small code size both in terms of source code with the minimum configuration
+  consisting of just three files, ``core.h``, ``format.h`` and ``format-inl.h``,
+  and compiled code; see `Compile time and code bloat`_
+* Reliability: the library has an extensive set of `tests
+  <https://github.com/fmtlib/fmt/tree/master/test>`_ and is `continuously fuzzed
+  <https://bugs.chromium.org/p/oss-fuzz/issues/list?colspec=ID%20Type%20
+  Component%20Status%20Proj%20Reported%20Owner%20Summary&q=proj%3Dfmt&can=1>`_
 * Safety: the library is fully type safe, errors in format strings can be
   reported at compile time, automatic memory management prevents buffer overflow
   errors
@@ -57,18 +66,17 @@ Features
   <https://github.com/fmtlib/fmt/blob/master/LICENSE.rst>`_
 * `Portability <https://fmt.dev/latest/index.html#portability>`_ with
   consistent output across platforms and support for older compilers
-* Clean warning-free codebase even on high warning levels
-  (``-Wall -Wextra -pedantic``)
+* Clean warning-free codebase even on high warning levels such as
+  ``-Wall -Wextra -pedantic``
 * Locale-independence by default
-* Support for wide strings
 * Optional header-only configuration enabled with the ``FMT_HEADER_ONLY`` macro
 
-See the `documentation <https://fmt.dev/latest/>`_ for more details.
+See the `documentation <https://fmt.dev>`_ for more details.
 
 Examples
 --------
 
-Print ``Hello, world!`` to ``stdout``:
+**Print to stdout** (`run <https://godbolt.org/z/Tevcjh>`_)
 
 .. code:: c++
 
@@ -78,100 +86,95 @@ Print ``Hello, world!`` to ``stdout``:
       fmt::print("Hello, world!\n");
     }
 
-Format a string:
+**Format a string** (`run <https://godbolt.org/z/oK8h33>`_)
 
 .. code:: c++
 
     std::string s = fmt::format("The answer is {}.", 42);
     // s == "The answer is 42."
 
-Format a string using positional arguments:
+**Format a string using positional arguments** (`run <https://godbolt.org/z/Yn7Txe>`_)
 
 .. code:: c++
 
     std::string s = fmt::format("I'd rather be {1} than {0}.", "right", "happy");
     // s == "I'd rather be happy than right."
 
-Print a chrono duration:
+**Print chrono durations** (`run <https://godbolt.org/z/K8s4Mc>`_)
 
 .. code:: c++
 
     #include <fmt/chrono.h>
 
     int main() {
-      using namespace std::chrono_literals;
-      fmt::print("Elapsed time: {}", 42ms);
+      using namespace std::literals::chrono_literals;
+      fmt::print("Default format: {} {}\n", 42s, 100ms);
+      fmt::print("strftime-like format: {:%H:%M:%S}\n", 3h + 15min + 30s);
     }
 
-prints "Elapsed time: 42ms".
+Output::
 
-Check a format string at compile time:
+    Default format: 42s 100ms
+    strftime-like format: 03:15:30
 
-.. code:: c++
+**Print a container** (`run <https://godbolt.org/z/MjsY7c>`_)
 
-    // test.cc
-    #include <fmt/format.h>
-    std::string s = format(FMT_STRING("{:d}"), "hello");
+.. code:: c++
 
-gives a compile-time error because ``d`` is an invalid format specifier for a
-string.
+    #include <vector>
+    #include <fmt/ranges.h>
 
-Use {fmt} as a safe portable replacement for ``itoa``
-(`godbolt <https://godbolt.org/g/NXmpU4>`_):
+    int main() {
+      std::vector<int> v = {1, 2, 3};
+      fmt::print("{}\n", v);
+    }
 
-.. code:: c++
+Output::
 
-    fmt::memory_buffer buf;
-    format_to(buf, "{}", 42);    // replaces itoa(42, buffer, 10)
-    format_to(buf, "{:x}", 42);  // replaces itoa(42, buffer, 16)
-    // access the string with to_string(buf) or buf.data()
+    {1, 2, 3}
 
-Format objects of user-defined types via a simple `extension API
-<https://fmt.dev/latest/api.html#formatting-user-defined-types>`_:
+**Check a format string at compile time**
 
 .. code:: c++
 
-    #include <fmt/format.h>
+    std::string s = fmt::format(FMT_STRING("{:d}"), "don't panic");
 
-    struct date {
-      int year, month, day;
-    };
+This gives a compile-time error because ``d`` is an invalid format specifier for
+a string.
 
-    template <>
-    struct fmt::formatter<date> {
-      constexpr auto parse(format_parse_context& ctx) { return ctx.begin(); }
+**Write a file from a single thread**
 
-      template <typename FormatContext>
-      auto format(const date& d, FormatContext& ctx) {
-        return format_to(ctx.out(), "{}-{}-{}", d.year, d.month, d.day);
-      }
-    };
+.. code:: c++
 
-    std::string s = fmt::format("The date is {}", date{2012, 12, 9});
-    // s == "The date is 2012-12-9"
+    #include <fmt/os.h>
 
-Create your own functions similar to `format
-<https://fmt.dev/latest/api.html#format>`_ and
-`print <https://fmt.dev/latest/api.html#print>`_
-which take arbitrary arguments (`godbolt <https://godbolt.org/g/MHjHVf>`_):
+    int main() {
+      auto out = fmt::output_file("guide.txt");
+      out.print("Don't {}", "Panic");
+    }
+
+This can be `5 to 9 times faster than fprintf
+<http://www.zverovich.net/2020/08/04/optimal-file-buffer-size.html>`_.
+
+**Print with colors and text styles**
 
 .. code:: c++
 
-    // Prints formatted error message.
-    void vreport_error(const char* format, fmt::format_args args) {
-      fmt::print("Error: ");
-      fmt::vprint(format, args);
-    }
-    template <typename... Args>
-    void report_error(const char* format, const Args & ... args) {
-      vreport_error(format, fmt::make_format_args(args...));
+    #include <fmt/color.h>
+
+    int main() {
+      fmt::print(fg(fmt::color::crimson) | fmt::emphasis::bold,
+                 "Hello, {}!\n", "world");
+      fmt::print(fg(fmt::color::floral_white) | bg(fmt::color::slate_gray) |
+                 fmt::emphasis::underline, "Hello, {}!\n", "мир");
+      fmt::print(fg(fmt::color::steel_blue) | fmt::emphasis::italic,
+                 "Hello, {}!\n", "世界");
     }
 
-    report_error("file not found: {}", path);
+Output on a modern terminal:
 
-Note that ``vreport_error`` is not parameterized on argument types which can
-improve compile times and reduce code size compared to a fully parameterized
-version.
+.. image:: https://user-images.githubusercontent.com/
+           576385/88485597-d312f600-cf2b-11ea-9cbe-61f535a86e28.png
 
 Benchmarks
 ----------
@@ -198,12 +201,14 @@ or equivalent is filled 2,000,000 times with output sent to ``/dev/null``; for
 further details refer to the `source
 <https://github.com/fmtlib/format-benchmark/blob/master/tinyformat_test.cpp>`_.
 
-{fmt} is up to 10x faster than ``std::ostringstream`` and ``sprintf`` on
+{fmt} is up to 20-30x faster than ``std::ostringstream`` and ``sprintf`` on
 floating-point formatting (`dtoa-benchmark <https://github.com/fmtlib/dtoa-benchmark>`_)
-and faster than `double-conversion <https://github.com/google/double-conversion>`_:
+and faster than `double-conversion <https://github.com/google/double-conversion>`_ and
+`ryu <https://github.com/ulfjack/ryu>`_:
 
-.. image:: https://user-images.githubusercontent.com/576385/69767160-cdaca400-112f-11ea-9fc5-347c9f83caad.png
-   :target: https://fmt.dev/unknown_mac64_clang10.0.html
+.. image:: https://user-images.githubusercontent.com/576385/
+           95684665-11719600-0ba8-11eb-8e5b-972ff4e49428.png
+   :target: https://fmt.dev/unknown_mac64_clang12.0.html
 
 Compile time and code bloat
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -282,30 +287,33 @@ or the bloat test::
 Projects using this library
 ---------------------------
 
-* `0 A.D. <https://play0ad.com/>`_: A free, open-source, cross-platform
+* `0 A.D. <https://play0ad.com/>`_: a free, open-source, cross-platform
   real-time strategy game
 
 * `AMPL/MP <https://github.com/ampl/mp>`_:
-  An open-source library for mathematical programming
+  an open-source library for mathematical programming
 
 * `Aseprite <https://github.com/aseprite/aseprite>`_:
-  Animated sprite editor & pixel art tool 
+  animated sprite editor & pixel art tool 
 
-* `AvioBook <https://www.aviobook.aero/en>`_: A comprehensive aircraft
+* `AvioBook <https://www.aviobook.aero/en>`_: a comprehensive aircraft
   operations suite
   
-* `Celestia <https://celestia.space/>`_: Real-time 3D visualization of space
+* `Blizzard Battle.net <https://battle.net/>`_: an online gaming platform
+  
+* `Celestia <https://celestia.space/>`_: real-time 3D visualization of space
 
-* `Ceph <https://ceph.com/>`_: A scalable distributed storage system
+* `Ceph <https://ceph.com/>`_: a scalable distributed storage system
 
-* `ccache <https://ccache.dev/>`_: A compiler cache
+* `ccache <https://ccache.dev/>`_: a compiler cache
 
-* `ClickHouse <https://github.com/ClickHouse/ClickHouse>`_: analytical database management system
+* `ClickHouse <https://github.com/ClickHouse/ClickHouse>`_: analytical database
+  management system
 
 * `CUAUV <http://cuauv.org/>`_: Cornell University's autonomous underwater
   vehicle
 
-* `Drake <https://drake.mit.edu/>`_: A planning, control, and analysis toolbox
+* `Drake <https://drake.mit.edu/>`_: a planning, control, and analysis toolbox
   for nonlinear dynamical systems (MIT)
 
 * `Envoy <https://lyft.github.io/envoy/>`_: C++ L7 proxy and communication bus
@@ -318,66 +326,70 @@ Projects using this library
 * `HarpyWar/pvpgn <https://github.com/pvpgn/pvpgn-server>`_:
   Player vs Player Gaming Network with tweaks
 
-* `KBEngine <https://kbengine.org/>`_: An open-source MMOG server engine
+* `KBEngine <https://github.com/kbengine/kbengine>`_: an open-source MMOG server
+  engine
 
-* `Keypirinha <https://keypirinha.com/>`_: A semantic launcher for Windows
+* `Keypirinha <https://keypirinha.com/>`_: a semantic launcher for Windows
 
-* `Kodi <https://kodi.tv/>`_ (formerly xbmc): Home theater software
+* `Kodi <https://kodi.tv/>`_ (formerly xbmc): home theater software
 
-* `Knuth <https://kth.cash/>`_: High-performance Bitcoin full-node
+* `Knuth <https://kth.cash/>`_: high-performance Bitcoin full-node
 
 * `Microsoft Verona <https://github.com/microsoft/verona>`_:
-  Research programming language for concurrent ownership
+  research programming language for concurrent ownership
 
-* `MongoDB <https://mongodb.com/>`_: Distributed document database
+* `MongoDB <https://mongodb.com/>`_: distributed document database
 
-* `MongoDB Smasher <https://github.com/duckie/mongo_smasher>`_: A small tool to
+* `MongoDB Smasher <https://github.com/duckie/mongo_smasher>`_: a small tool to
   generate randomized datasets
 
-* `OpenSpace <https://openspaceproject.com/>`_: An open-source
+* `OpenSpace <https://openspaceproject.com/>`_: an open-source
   astrovisualization framework
 
 * `PenUltima Online (POL) <https://www.polserver.com/>`_:
-  An MMO server, compatible with most Ultima Online clients
+  an MMO server, compatible with most Ultima Online clients
 
-* `PyTorch <https://github.com/pytorch/pytorch>`_: An open-source machine
+* `PyTorch <https://github.com/pytorch/pytorch>`_: an open-source machine
   learning library
 
-* `quasardb <https://www.quasardb.net/>`_: A distributed, high-performance,
+* `quasardb <https://www.quasardb.net/>`_: a distributed, high-performance,
   associative database
+  
+* `Quill <https://github.com/odygrd/quill>`_: asynchronous low-latency logging library
 
-* `readpe <https://bitbucket.org/sys_dev/readpe>`_: Read Portable Executable
+* `QKW <https://github.com/ravijanjam/qkw>`_: generalizing aliasing to simplify
+  navigation, and executing complex multi-line terminal command sequences
 
-* `redis-cerberus <https://github.com/HunanTV/redis-cerberus>`_: A Redis cluster
+* `redis-cerberus <https://github.com/HunanTV/redis-cerberus>`_: a Redis cluster
   proxy
 
-* `redpanda <https://vectorized.io/redpanda>`_: A 10x faster Kafka® replacement
+* `redpanda <https://vectorized.io/redpanda>`_: a 10x faster Kafka® replacement
   for mission critical systems written in C++
 
-* `rpclib <http://rpclib.net/>`_: A modern C++ msgpack-RPC server and client
+* `rpclib <http://rpclib.net/>`_: a modern C++ msgpack-RPC server and client
   library
 
 * `Salesforce Analytics Cloud
   <https://www.salesforce.com/analytics-cloud/overview/>`_:
-  Business intelligence software
+  business intelligence software
 
-* `Scylla <https://www.scylladb.com/>`_: A Cassandra-compatible NoSQL data store
+* `Scylla <https://www.scylladb.com/>`_: a Cassandra-compatible NoSQL data store
   that can handle 1 million transactions per second on a single server
 
-* `Seastar <http://www.seastar-project.org/>`_: An advanced, open-source C++
+* `Seastar <http://www.seastar-project.org/>`_: an advanced, open-source C++
   framework for high-performance server applications on modern hardware
 
-* `spdlog <https://github.com/gabime/spdlog>`_: Super fast C++ logging library
+* `spdlog <https://github.com/gabime/spdlog>`_: super fast C++ logging library
 
-* `Stellar <https://www.stellar.org/>`_: Financial platform
+* `Stellar <https://www.stellar.org/>`_: financial platform
 
-* `Touch Surgery <https://www.touchsurgery.com/>`_: Surgery simulator
+* `Touch Surgery <https://www.touchsurgery.com/>`_: surgery simulator
 
-* `TrinityCore <https://github.com/TrinityCore/TrinityCore>`_: Open-source
+* `TrinityCore <https://github.com/TrinityCore/TrinityCore>`_: open-source
   MMORPG framework
 
-* `Windows Terminal <https://github.com/microsoft/terminal>`_: The new Windows
-  Terminal
+* `Windows Terminal <https://github.com/microsoft/terminal>`_: the new Windows
+  terminal
 
 `More... <https://github.com/search?q=fmtlib&type=Code>`_
 
@@ -435,7 +447,7 @@ Boost Format
 
 This is a very powerful library which supports both ``printf``-like format
 strings and positional arguments. Its main drawback is performance. According to
-various benchmarks it is much slower than other methods considered here. Boost
+various, benchmarks it is much slower than other methods considered here. Boost
 Format also has excessive build times and severe code bloat issues (see
 `Benchmarks`_).
 
diff --git a/doc/CMakeLists.txt b/doc/CMakeLists.txt
index 108aa71e..06848450 100644
--- a/doc/CMakeLists.txt
+++ b/doc/CMakeLists.txt
@@ -4,10 +4,14 @@ if (NOT DOXYGEN)
   return ()
 endif ()
 
+find_package(PythonInterp QUIET REQUIRED)
+
 add_custom_target(doc
-  COMMAND ${CMAKE_CURRENT_SOURCE_DIR}/build.py ${FMT_VERSION}
+  COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/build.py
+                               ${FMT_VERSION}
   SOURCES api.rst syntax.rst usage.rst build.py conf.py _templates/layout.html)
 
+include(GNUInstallDirs)
 install(DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}/html/
-        DESTINATION share/doc/fmt OPTIONAL
+        DESTINATION ${CMAKE_INSTALL_DATAROOTDIR}/doc/fmt OPTIONAL
         PATTERN ".doctrees" EXCLUDE)
diff --git a/doc/api.rst b/doc/api.rst
index ca016895..9b6c0f3c 100644
--- a/doc/api.rst
+++ b/doc/api.rst
@@ -15,6 +15,7 @@ The {fmt} library API consists of the following parts:
   and tuples
 * :ref:`fmt/chrono.h <chrono-api>`: date and time formatting
 * :ref:`fmt/compile.h <compile-api>`: format string compilation
+* :ref:`fmt/color.h <color-api>`: terminal color and text style
 * :ref:`fmt/ostream.h <ostream-api>`: ``std::ostream`` support
 * :ref:`fmt/printf.h <printf-api>`: ``printf`` formatting
 
@@ -47,6 +48,10 @@ participate in an overload resolution if the latter is not a string.
 .. doxygenfunction:: format(const S&, Args&&...)
 .. doxygenfunction:: vformat(const S&, basic_format_args<buffer_context<type_identity_t<Char>>>)
 
+.. doxygenfunction:: fmt::format_to(OutputIt, const S&, Args&&...)
+.. doxygenfunction:: fmt::format_to_n(OutputIt, size_t, const S&, const Args&...)
+.. doxygenfunction:: fmt::formatted_size(string_view, Args&&...)
+
 .. _print:
 
 .. doxygenfunction:: print(const S&, Args&&...)
@@ -65,6 +70,35 @@ Named arguments are not supported in compile-time checks at the moment.
 Argument Lists
 --------------
 
+You can create your own formatting function with compile-time checks and small
+binary footprint, for example (https://godbolt.org/z/oba4Mc):
+
+.. code:: c++
+
+    #include <fmt/format.h>
+
+    void vlog(const char* file, int line, fmt::string_view format,
+              fmt::format_args args) {
+      fmt::print("{}: {}: ", file, line);
+      fmt::vprint(format, args);
+    }
+
+    template <typename S, typename... Args>
+    void log(const char* file, int line, const S& format, Args&&... args) {
+      vlog(file, line, format,
+          fmt::make_args_checked<Args...>(format, args...));
+    }
+
+    #define MY_LOG(format, ...) \
+      log(__FILE__, __LINE__, FMT_STRING(format), __VA_ARGS__)
+
+    MY_LOG("invalid squishiness: {}", 42);
+
+Note that ``vlog`` is not parameterized on argument types which improves compile
+times and reduces binary code size compared to a fully parameterized version.
+
+.. doxygenfunction:: fmt::make_args_checked(const S&, const remove_reference_t<Args>&...)
+
 .. doxygenfunction:: fmt::make_format_args(const Args&...)
 
 .. doxygenclass:: fmt::format_arg_store
@@ -93,7 +127,7 @@ Compatibility
 Locale
 ------
 
-All formatting is locale-independent by default. Use the ``'n'`` format
+All formatting is locale-independent by default. Use the ``'L'`` format
 specifier to insert the appropriate number separator characters from the
 locale::
 
@@ -114,12 +148,14 @@ string checks, wide string, output iterator and user-defined type support.
 Compile-time Format String Checks
 ---------------------------------
 
-Compile-time checks are supported for built-in and string types as well as
-user-defined types with ``constexpr`` ``parse`` functions in their ``formatter``
-specializations.
+Compile-time checks are enabled when using ``FMT_STRING``. They support built-in
+and string types as well as user-defined types with ``constexpr`` ``parse``
+functions in their ``formatter`` specializations.
 
 .. doxygendefine:: FMT_STRING
 
+.. _udt:
+
 Formatting User-defined Types
 -----------------------------
 
@@ -237,6 +273,10 @@ You can also write a formatter for a hierarchy of classes::
     fmt::print("{}", a); // prints "B"
   }
 
+If a type provides both a ``formatter`` specialization and an implicit
+conversion to a formattable type, the specialization takes precedence over the
+conversion.
+
 .. doxygenclass:: fmt::basic_format_parse_context
    :members:
 
@@ -264,7 +304,9 @@ Utilities
 
 .. doxygentypedef:: fmt::char_t
 
-.. doxygenfunction:: fmt::formatted_size(string_view, const Args&...)
+.. doxygenfunction:: fmt::ptr(const T *)
+.. doxygenfunction:: fmt::ptr(const std::unique_ptr<T>&)
+.. doxygenfunction:: fmt::ptr(const std::shared_ptr<T>&)
 
 .. doxygenfunction:: fmt::to_string(const T&)
 
@@ -272,7 +314,7 @@ Utilities
 
 .. doxygenfunction:: fmt::to_string_view(const Char *)
 
-.. doxygenfunction:: fmt::join(const Range&, string_view)
+.. doxygenfunction:: fmt::join(Range&&, string_view)
 
 .. doxygenfunction:: fmt::join(It, Sentinel, string_view)
 
@@ -298,8 +340,6 @@ the value of ``errno`` being preserved by library functions.
 .. doxygenclass:: fmt::windows_error
    :members:
 
-.. _formatstrings:
-
 Custom Allocators
 -----------------
 
@@ -330,10 +370,10 @@ allocator::
       return vformat(alloc, format_str, fmt::make_format_args(args...));
     }
 
-The allocator will be used for the output container only. If you are using named
-arguments, the container that stores pointers to them will be allocated using
-the default allocator. Also floating-point formatting falls back on ``sprintf``
-which may do allocations.
+The allocator will be used for the output container only. Formatting functions
+normally don't do any allocations for built-in and string types except for
+non-default floating-point formatting that occasionally falls back on
+``sprintf``.
 
 .. _ranges-api:
 
@@ -383,9 +423,10 @@ The format string syntax is described in the documentation of
 Format string compilation
 =========================
 
-``fmt/compile.h`` provides format string compilation support. Format strings
-are parsed at compile time and converted into efficient formatting code. This
-supports arguments of built-in and string types as well as user-defined types
+``fmt/compile.h`` provides format string compilation support when using
+``FMT_COMPILE``. Format strings are parsed, checked and converted
+into efficient formatting code at compile-time.
+This supports arguments of built-in and string types as well as user-defined types
 with ``constexpr`` ``parse`` functions in their ``formatter`` specializations.
 Format string compilation can generate more binary code compared to the default
 API and is only recommended in places where formatting is a performance
@@ -393,6 +434,15 @@ bottleneck.
 
 .. doxygendefine:: FMT_COMPILE
 
+.. _color-api:
+
+Terminal color and text style
+=============================
+
+``fmt/color.h`` provides support for terminal color and text style output.
+
+.. doxygenfunction:: print(const text_style&, const S&, const Args&...)
+
 .. _ostream-api:
 
 ``std::ostream`` Support
diff --git a/doc/build.py b/doc/build.py
index 25d7885a..7152638b 100755
--- a/doc/build.py
+++ b/doc/build.py
@@ -6,20 +6,10 @@ import errno, os, shutil, sys, tempfile
 from subprocess import check_call, check_output, CalledProcessError, Popen, PIPE
 from distutils.version import LooseVersion
 
-versions = ['1.0.0', '1.1.0', '2.0.0', '3.0.2', '4.0.0', '4.1.0', '5.0.0', '5.1.0', '5.2.0', '5.2.1', '5.3.0', '6.0.0', '6.1.0', '6.1.1', '6.1.2', '6.2.0', '6.2.1', '7.0.0', '7.0.1']
+versions = ['1.0.0', '1.1.0', '2.0.0', '3.0.2', '4.0.0', '4.1.0', '5.0.0', '5.1.0', '5.2.0', '5.2.1', '5.3.0', '6.0.0', '6.1.0', '6.1.1', '6.1.2', '6.2.0', '6.2.1', '7.0.0', '7.0.1', '7.0.2', '7.0.3', '7.1.0']
 
 def pip_install(package, commit=None, **kwargs):
   "Install package using pip."
-  min_version = kwargs.get('min_version')
-  if min_version:
-    from pkg_resources import get_distribution, DistributionNotFound
-    try:
-      installed_version = get_distribution(os.path.basename(package)).version
-      if LooseVersion(installed_version) >= min_version:
-        print('{} {} already installed'.format(package, min_version))
-        return
-    except DistributionNotFound:
-      pass
   if commit:
     package = 'git+https://github.com/{0}.git@{1}'.format(package, commit)
   print('Installing {0}'.format(package))
@@ -52,12 +42,11 @@ def create_build_env(dirname='virtualenv'):
       check_call(['pip', 'install', '--upgrade', 'distribute'])
   except DistributionNotFound:
     pass
-  # Install Sphinx and Breathe.
-  pip_install('sphinx-doc/sphinx', '12b83372ac9316e8cbe86e7fed889296a4cc29ee',
-              min_version='1.4.1.dev20160531')
+  # Install Sphinx and Breathe. Require the exact version of Sphinx which is
+  # compatible with Breathe.
+  pip_install('sphinx-doc/sphinx', '12b83372ac9316e8cbe86e7fed889296a4cc29ee')
   pip_install('michaeljones/breathe',
-              '129222318f7c8f865d2631e7da7b033567e7f56a',
-              min_version='4.2.0')
+              '129222318f7c8f865d2631e7da7b033567e7f56a')
 
 def build_docs(version='dev', **kwargs):
   doc_dir = kwargs.get('doc_dir', os.path.dirname(os.path.realpath(__file__)))
@@ -74,8 +63,8 @@ def build_docs(version='dev', **kwargs):
       GENERATE_MAN      = NO
       GENERATE_RTF      = NO
       CASE_SENSE_NAMES  = NO
-      INPUT             = {0}/core.h {0}/compile.h {0}/format.h {0}/os.h \
-                          {0}/ostream.h {0}/printf.h {0}/time.h
+      INPUT             = {0}/chrono.h {0}/color.h {0}/core.h {0}/compile.h \
+                          {0}/format.h {0}/os.h {0}/ostream.h {0}/printf.h
       QUIET             = YES
       JAVADOC_AUTOBRIEF = YES
       AUTOLINK_SUPPORT  = NO
@@ -95,7 +84,8 @@ def build_docs(version='dev', **kwargs):
                           "FMT_END_NAMESPACE=}}" \
                           "FMT_STRING_ALIAS=1" \
                           "FMT_ENABLE_IF(B)="
-      EXCLUDE_SYMBOLS   = fmt::internal::* StringValue write_str
+      EXCLUDE_SYMBOLS   = fmt::formatter fmt::printf_formatter fmt::arg_join \
+                          fmt::basic_format_arg::handle
     '''.format(include_dir, doxyxml_dir).encode('UTF-8'))
   if p.returncode != 0:
     raise CalledProcessError(p.returncode, cmd)
diff --git a/doc/index.rst b/doc/index.rst
index 59dc6652..58f29295 100644
--- a/doc/index.rst
+++ b/doc/index.rst
@@ -23,7 +23,7 @@ Format API
 
 The format API is similar in spirit to the C ``printf`` family of function but
 is safer, simpler and several times `faster
-<http://zverovich.net/2013/09/07/integer-to-string-conversion-in-cplusplus.html>`_
+<https://www.zverovich.net/2020/06/13/fast-int-to-string-revisited.html>`_
 than common standard library implementations.
 The `format string syntax <syntax.html>`_ is similar to the one used by
 `str.format <http://docs.python.org/3/library/stdtypes.html#str.format>`_ in
@@ -48,21 +48,19 @@ The ``fmt::print`` function performs formatting and writes the result to a strea
 
   fmt::print(stderr, "System error code = {}\n", errno);
 
-The file argument can be omitted in which case the function prints to
-``stdout``:
+If you omit the file argument the function will print to ``stdout``:
 
 .. code:: c++
 
   fmt::print("Don't {}\n", "panic");
 
-The Format API also supports positional arguments useful for localization:
+The format API also supports positional arguments useful for localization:
 
 .. code:: c++
 
   fmt::print("I'd rather be {1} than {0}.", "right", "happy");
 
-Named arguments can be created with ``fmt::arg``. This makes it easier to track 
-what goes where when multiple arguments are being formatted:
+You can pass named arguments with ``fmt::arg``:
 
 .. code:: c++
 
@@ -91,16 +89,17 @@ time. For example, the code
 
   fmt::format("The answer is {:d}", "forty-two");
 
-throws a ``format_error`` exception with description "unknown format code 'd' for
-string", because the argument ``"forty-two"`` is a string while the format code
-``d`` only applies to integers, while
+throws the ``format_error`` exception because the argument ``"forty-two"`` is a
+string while the format code ``d`` only applies to integers.
+
+The code
 
 .. code:: c++
 
   format(FMT_STRING("The answer is {:d}"), "forty-two");
 
-reports a compile-time error for the same reason on compilers that support
-relaxed ``constexpr``. See `here <api.html#c.fmt>`_ for details.
+reports a compile-time error on compilers that support relaxed ``constexpr``.
+See `here <api.html#c.fmt>`_ for details.
 
 The following code
 
@@ -117,13 +116,13 @@ formatted into a narrow string. You can use a wide format string instead:
 
 For comparison, writing a wide character to ``std::ostream`` results in
 its numeric value being written to the stream (i.e. 1070 instead of letter 'ю'
-which is represented by ``L'\x42e'`` if we use Unicode) which is rarely what is
-needed.
+which is represented by ``L'\x42e'`` if we use Unicode) which is rarely
+desirable.
 
 Compact Binary Code
 -------------------
 
-The library is designed to produce compact per-call compiled code. For example
+The library produces compact per-call compiled code. For example
 (`godbolt <https://godbolt.org/g/TZU4KF>`_),
 
 .. code:: c++
@@ -144,8 +143,8 @@ compiles to just
      mov rcx, rsp
      mov edi, offset .L.str
      mov esi, 17
-     mov edx, 2
-     call fmt::v5::vprint(fmt::v5::basic_string_view<char>, fmt::v5::format_args)
+     mov edx, 1
+     call fmt::v7::vprint(fmt::v7::basic_string_view<char>, fmt::v7::format_args)
      xor eax, eax
      add rsp, 24
      ret
@@ -167,20 +166,19 @@ The library is highly portable and relies only on a small set of C++11 features:
 * deleted functions
 * alias templates
 
-These are available since GCC 4.8, Clang 3.0 and MSVC 19.0 (2015). For older
-compilers use {fmt} `version 4.x
-<https://github.com/fmtlib/fmt/releases/tag/4.1.0>`_ which continues to be
-maintained and only requires C++98.
+These are available in GCC 4.8, Clang 3.0, MSVC 19.0 (2015) and more recent
+compiler version. For older compilers use {fmt} `version 4.x
+<https://github.com/fmtlib/fmt/releases/tag/4.1.0>`_ which is maintained and
+only requires C++98.
 
-The output of all formatting functions is consistent across platforms. In
-particular, formatting a floating-point infinity always gives ``inf`` while the
-output of ``printf`` is platform-dependent. For example,
+The output of all formatting functions is consistent across platforms.
+For example,
 
 .. code::
 
   fmt::print("{}", std::numeric_limits<double>::infinity());
 
-always prints ``inf``.
+always prints ``inf`` while the output of ``printf`` is platform-dependent.
 
 .. _ease-of-use:
 
diff --git a/doc/usage.rst b/doc/usage.rst
index d42719e0..435afd53 100644
--- a/doc/usage.rst
+++ b/doc/usage.rst
@@ -52,6 +52,14 @@ To build a `shared library`__ set the ``BUILD_SHARED_LIBS`` CMake variable to
 
 __ http://en.wikipedia.org/wiki/Library_%28computing%29#Shared_libraries
 
+
+To build a `static library` with position independent code (required if the main
+consumer of the fmt library is a shared library i.e. a Python extension) set the
+``CMAKE_POSITION_INDEPENDENT_CODE`` CMake variable to ``TRUE``::
+
+  cmake -DCMAKE_POSITION_INDEPENDENT_CODE=TRUE ...
+
+
 Installing the Library
 ======================
 
@@ -83,6 +91,49 @@ Setting up your target to use a header-only version of ``fmt`` is equally easy::
 
    target_link_libraries(<your-target> PRIVATE fmt::fmt-header-only)
 
+Usage with build2
+=================
+
+You can use `build2 <https://build2.org>`_, a dependency manager and a
+build-system combined, to use ``fmt``.
+
+Currently this package is available in these package repositories:
+
+- **https://cppget.org/fmt/** for released and published versions.
+- `The git repository with the sources of the build2 package of fmt <https://github.com/build2-packaging/fmt.git>`_
+  for unreleased or custom revisions of ``fmt``.
+
+**Usage:**
+
+- ``build2`` package name: ``fmt``
+- Library target name : ``lib{fmt}``
+
+For example, to make your ``build2`` project depend on ``fmt``:
+
+- Add one of the repositories to your configurations, or in your
+  ``repositories.manifest``, if not already there::
+
+    :
+    role: prerequisite
+    location: https://pkg.cppget.org/1/stable
+
+- Add this package as a dependency to your ``./manifest`` file
+  (example for ``v7.0.x``)::
+
+    depends: fmt ~7.0.0
+
+- Import the target and use it as a prerequisite to your own target
+  using `fmt` in the appropriate ``buildfile``::
+
+    import fmt = fmt%lib{fmt}
+    lib{mylib} : cxx{**} ... $fmt
+
+Then build your project as usual with `b` or `bdep update`.
+
+For ``build2`` newcomers or to get more details and use cases, you can read the
+``build2``
+`toolchain introduction <https://build2.org/build2-toolchain/doc/build2-toolchain-intro.xhtml>`_.
+
 Building the Documentation
 ==========================
 
@@ -130,6 +181,18 @@ The fmt port in vcpkg is kept up to date by Microsoft team members and community
 contributors. If the version is out of date, please `create an issue or pull
 request <https://github.com/Microsoft/vcpkg>`__ on the vcpkg repository.
 
+LHelper
+=======
+
+You can download and install fmt using
+`lhelper <https://github.com/franko/lhelper>`__ dependency manager::
+
+  lhelper activate <some-environment>
+  lhelper install fmt
+
+All the recipes for lhelper are kept in the
+`lhelper's recipe <https://github.com/franko/lhelper-recipes>`__ repository.
+
 Android NDK
 ===========
 
diff --git a/include/fmt/chrono.h b/include/fmt/chrono.h
index e70b8053..1a3b8d5e 100644
--- a/include/fmt/chrono.h
+++ b/include/fmt/chrono.h
@@ -72,43 +72,27 @@ FMT_CONSTEXPR To lossless_integral_conversion(const From from, int& ec) {
   static_assert(F::is_integer, "From must be integral");
   static_assert(T::is_integer, "To must be integral");
 
-  if (F::is_signed && !T::is_signed) {
+  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) {
-      // yes, From always fits in To.
-    } else {
-      // from may not fit in To, we have to do a dynamic check
-      if (from > static_cast<From>((T::max)())) {
-        ec = 1;
-        return {};
-      }
+    if (F::digits > T::digits &&
+        from > static_cast<From>(detail::max_value<To>())) {
+      ec = 1;
+      return {};
     }
   }
 
-  if (!F::is_signed && T::is_signed) {
-    // can from be held in To?
-    if (F::digits < T::digits) {
-      // yes, From always fits in To.
-    } else {
-      // from may not fit in To, we have to do a dynamic check
-      if (from > static_cast<From>((T::max)())) {
-        // outside range.
-        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 {};
   }
-
-  // reaching here means all is ok for lossless conversion.
-  return static_cast<To>(from);
-
-}  // function
+  return static_cast<To>(from);  // Lossless conversion.
+}
 
 template <typename To, typename From,
           FMT_ENABLE_IF(std::is_same<From, To>::value)>
@@ -190,11 +174,9 @@ To safe_duration_cast(std::chrono::duration<FromRep, FromPeriod> from,
   // safe conversion to IntermediateRep
   IntermediateRep count =
       lossless_integral_conversion<IntermediateRep>(from.count(), ec);
-  if (ec) {
-    return {};
-  }
+  if (ec) return {};
   // multiply with Factor::num without overflow or underflow
-  if (Factor::num != 1) {
+  if (detail::const_check(Factor::num != 1)) {
     const auto max1 = detail::max_value<IntermediateRep>() / Factor::num;
     if (count > max1) {
       ec = 1;
@@ -209,17 +191,9 @@ To safe_duration_cast(std::chrono::duration<FromRep, FromPeriod> from,
     count *= Factor::num;
   }
 
-  // this can't go wrong, right? den>0 is checked earlier.
-  if (Factor::den != 1) {
-    count /= Factor::den;
-  }
-  // convert to the to type, safely
-  using ToRep = typename To::rep;
-  const ToRep tocount = lossless_integral_conversion<ToRep>(count, ec);
-  if (ec) {
-    return {};
-  }
-  return To{tocount};
+  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);
 }
 
 /**
@@ -351,6 +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 {
@@ -387,6 +366,11 @@ inline std::tm gmtime(std::time_t time) {
   return gt.tm_;
 }
 
+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) {
@@ -399,6 +383,17 @@ inline size_t strftime(wchar_t* str, size_t count, const wchar_t* format,
 }
 }  // 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>
   auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
diff --git a/include/fmt/color.h b/include/fmt/color.h
index b65f892a..78910589 100644
--- a/include/fmt/color.h
+++ b/include/fmt/color.h
@@ -463,16 +463,16 @@ template <> inline void reset_color<wchar_t>(FILE* stream) FMT_NOEXCEPT {
 }
 
 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>
-void vformat_to(basic_memory_buffer<Char>& buf, const text_style& ts,
+void vformat_to(buffer<Char>& buf, const text_style& ts,
                 basic_string_view<Char> format_str,
-                basic_format_args<buffer_context<Char>> args) {
+                basic_format_args<buffer_context<type_identity_t<Char>>> args) {
   bool has_style = false;
   if (ts.has_emphasis()) {
     has_style = true;
@@ -496,7 +496,7 @@ void vformat_to(basic_memory_buffer<Char>& buf, const text_style& ts,
 
 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) {
+            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));
@@ -504,20 +504,22 @@ void vprint(std::FILE* f, const text_style& ts, const S& format,
 }
 
 /**
+  \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(detail::is_string<S>::value)>
 void print(std::FILE* f, const text_style& ts, const S& format_str,
            const Args&... args) {
-  detail::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...));
 }
 
 /**
@@ -558,7 +560,41 @@ 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 vformat(ts, to_string_view(format_str),
-                 detail::make_args_checked<Args...>(format_str, args...));
+                 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,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char_t<S>>::value&&
+                            detail::is_string<S>::value)>
+inline OutputIt format_to(OutputIt out, const text_style& ts,
+                          const S& format_str, Args&&... args) {
+  return vformat_to(out, ts, to_string_view(format_str),
+                    fmt::make_args_checked<Args...>(format_str, args...));
 }
 
 FMT_END_NAMESPACE
diff --git a/include/fmt/compile.h b/include/fmt/compile.h
index d7e6449e..7db610d9 100644
--- a/include/fmt/compile.h
+++ b/include/fmt/compile.h
@@ -368,7 +368,8 @@ 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(const T& first, const Args&... rest) {
+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;
@@ -406,6 +407,19 @@ constexpr text<Char> make_text(basic_string_view<Char> s, size_t pos,
   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;
@@ -430,7 +444,9 @@ template <typename Char, typename T, int N> struct spec_field {
   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...);
-    basic_format_context<OutputIt, Char> ctx(out, {});
+    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);
   }
 };
@@ -489,16 +505,17 @@ constexpr auto parse_tail(T head, S format_str) {
 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) {
+                                                  size_t pos, int arg_id) {
   str.remove_prefix(pos);
-  auto ctx = basic_format_parse_context<Char>(str);
+  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};
+  return {f, pos + (end - str.data()) + 1, ctx.next_arg_id()};
 }
 
 // Compiles a non-empty format string and returns the compiled representation
@@ -518,8 +535,8 @@ constexpr auto compile_format_string(S format_str) {
                                                format_str);
     } else if constexpr (str[POS + 1] == ':') {
       using type = get_type<ID, Args>;
-      constexpr auto result = parse_specs<type>(str, POS + 2);
-      return parse_tail<Args, result.end, ID + 1>(
+      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();
@@ -530,8 +547,13 @@ constexpr auto compile_format_string(S format_str) {
     return parse_tail<Args, POS + 2, ID>(make_text(str, POS, 1), format_str);
   } else {
     constexpr auto end = parse_text(str, POS + 1);
-    return parse_tail<Args, end, ID>(make_text(str, POS, end - POS),
-                                     format_str);
+    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);
+    }
   }
 }
 
@@ -587,8 +609,7 @@ template <typename CompiledFormat, typename... Args,
 FMT_INLINE std::basic_string<Char> format(const CompiledFormat& cf,
                                           const Args&... args) {
   basic_memory_buffer<Char> buffer;
-  detail::buffer<Char>& base = buffer;
-  cf.format(std::back_inserter(base), args...);
+  cf.format(detail::buffer_appender<Char>(buffer), args...);
   return to_string(buffer);
 }
 
@@ -608,8 +629,7 @@ template <typename CompiledFormat, typename... Args,
 std::basic_string<Char> format(const CompiledFormat& cf, const Args&... args) {
   basic_memory_buffer<Char> buffer;
   using context = buffer_context<Char>;
-  detail::buffer<Char>& base = buffer;
-  detail::cf::vformat_to<context>(std::back_inserter(base), cf,
+  detail::cf::vformat_to<context>(detail::buffer_appender<Char>(buffer), cf,
                                   make_format_args<context>(args...));
   return to_string(buffer);
 }
@@ -618,9 +638,13 @@ 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) {
-  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...));
+#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)...);
 }
@@ -643,10 +667,11 @@ OutputIt format_to(OutputIt out, const S&, const Args&... args) {
   return format_to(out, compiled, args...);
 }
 
-template <
-    typename OutputIt, typename CompiledFormat, typename... Args,
-    FMT_ENABLE_IF(detail::is_output_iterator<OutputIt>::value&& std::is_base_of<
-                  detail::basic_compiled_format, CompiledFormat>::value)>
+template <typename OutputIt, typename CompiledFormat, typename... Args,
+          FMT_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> format_to_n(OutputIt out, size_t n,
                                          const CompiledFormat& cf,
                                          const Args&... args) {
@@ -655,6 +680,16 @@ format_to_n_result<OutputIt> format_to_n(OutputIt out, size_t n,
   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>
 size_t formatted_size(const CompiledFormat& cf, const Args&... args) {
   return format_to(detail::counting_iterator(), cf, args...).count();
diff --git a/include/fmt/core.h b/include/fmt/core.h
index 761162eb..2903be5a 100644
--- a/include/fmt/core.h
+++ b/include/fmt/core.h
@@ -18,7 +18,7 @@
 #include <vector>
 
 // The fmt library version in the form major * 10000 + minor * 100 + patch.
-#define FMT_VERSION 70001
+#define FMT_VERSION 70100
 
 #ifdef __clang__
 #  define FMT_CLANG_VERSION (__clang_major__ * 100 + __clang_minor__)
@@ -57,6 +57,7 @@
 #  define FMT_MSC_VER 0
 #  define FMT_SUPPRESS_MSC_WARNING(n)
 #endif
+
 #ifdef __has_feature
 #  define FMT_HAS_FEATURE(x) __has_feature(x)
 #else
@@ -64,7 +65,7 @@
 #endif
 
 #if defined(__has_include) && !defined(__INTELLISENSE__) && \
-    !(FMT_ICC_VERSION && FMT_ICC_VERSION < 1600)
+    (!FMT_ICC_VERSION || FMT_ICC_VERSION >= 1600)
 #  define FMT_HAS_INCLUDE(x) __has_include(x)
 #else
 #  define FMT_HAS_INCLUDE(x) 0
@@ -99,7 +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
@@ -152,7 +153,7 @@
 #  if FMT_HAS_CPP14_ATTRIBUTE(deprecated) || FMT_MSC_VER >= 1900
 #    define FMT_DEPRECATED [[deprecated]]
 #  else
-#    if defined(__GNUC__) || defined(__clang__)
+#    if (defined(__GNUC__) && !defined(__LCC__)) || defined(__clang__)
 #      define FMT_DEPRECATED __attribute__((deprecated))
 #    elif FMT_MSC_VER
 #      define FMT_DEPRECATED __declspec(deprecated)
@@ -177,9 +178,17 @@
 #  endif
 #endif
 
-#ifndef FMT_BEGIN_NAMESPACE
+#ifndef FMT_USE_INLINE_NAMESPACES
 #  if FMT_HAS_FEATURE(cxx_inline_namespaces) || FMT_GCC_VERSION >= 404 || \
-      FMT_MSC_VER >= 1900
+      (FMT_MSC_VER >= 1900 && !_MANAGED)
+#    define FMT_USE_INLINE_NAMESPACES 1
+#  else
+#    define FMT_USE_INLINE_NAMESPACES 0
+#  endif
+#endif
+
+#ifndef FMT_BEGIN_NAMESPACE
+#  if FMT_USE_INLINE_NAMESPACES
 #    define FMT_INLINE_NAMESPACE inline namespace
 #    define FMT_END_NAMESPACE \
       }                       \
@@ -269,8 +278,7 @@ struct monostate {};
 
 namespace detail {
 
-// A helper function to suppress bogus "conditional expression is constant"
-// warnings.
+// A helper function to suppress "conditional expression is constant" warnings.
 template <typename T> constexpr T const_check(T value) { return value; }
 
 FMT_NORETURN FMT_API void assert_fail(const char* file, int line,
@@ -299,7 +307,8 @@ template <typename T> struct std_string_view {};
 
 #ifdef FMT_USE_INT128
 // Do nothing.
-#elif defined(__SIZEOF_INT128__) && !FMT_NVCC
+#elif defined(__SIZEOF_INT128__) && !FMT_NVCC && \
+    !(FMT_CLANG_VERSION && FMT_MSC_VER)
 #  define FMT_USE_INT128 1
 using int128_t = __int128_t;
 using uint128_t = __uint128_t;
@@ -506,6 +515,18 @@ template <typename S> struct char_t_impl<S, enable_if_t<is_string<S>::value>> {
   using type = typename result::value_type;
 };
 
+// Reports a compile-time error if S is not a valid format string.
+template <typename..., typename S, FMT_ENABLE_IF(!is_compile_string<S>::value)>
+FMT_INLINE void check_format_string(const S&) {
+#ifdef FMT_ENFORCE_COMPILE_STRING
+  static_assert(is_compile_string<S>::value,
+                "FMT_ENFORCE_COMPILE_STRING requires all format strings to use "
+                "FMT_STRING.");
+#endif
+}
+template <typename..., typename S, FMT_ENABLE_IF(is_compile_string<S>::value)>
+void check_format_string(S);
+
 struct error_handler {
   constexpr error_handler() = default;
   constexpr error_handler(const error_handler&) = default;
@@ -545,8 +566,9 @@ class basic_format_parse_context : private ErrorHandler {
   using iterator = typename basic_string_view<Char>::iterator;
 
   explicit constexpr basic_format_parse_context(
-      basic_string_view<Char> format_str, ErrorHandler eh = {})
-      : ErrorHandler(eh), format_str_(format_str), next_arg_id_(0) {}
+      basic_string_view<Char> format_str, ErrorHandler eh = {},
+      int next_arg_id = 0)
+      : ErrorHandler(eh), format_str_(format_str), next_arg_id_(next_arg_id) {}
 
   /**
     Returns an iterator to the beginning of the format string range being
@@ -616,8 +638,24 @@ template <typename T, typename Context>
 using has_formatter =
     std::is_constructible<typename Context::template formatter_type<T>>;
 
+// Checks whether T is a container with contiguous storage.
+template <typename T> struct is_contiguous : std::false_type {};
+template <typename Char>
+struct is_contiguous<std::basic_string<Char>> : std::true_type {};
+
 namespace detail {
 
+// Extracts a reference to the container from back_insert_iterator.
+template <typename Container>
+inline Container& get_container(std::back_insert_iterator<Container> it) {
+  using bi_iterator = std::back_insert_iterator<Container>;
+  struct accessor : bi_iterator {
+    accessor(bi_iterator iter) : bi_iterator(iter) {}
+    using bi_iterator::container;
+  };
+  return *accessor(it).container;
+}
+
 /**
   \rst
   A contiguous memory buffer with an optional growing ability. It is an internal
@@ -640,6 +678,8 @@ template <typename T> class buffer {
         size_(sz),
         capacity_(cap) {}
 
+  ~buffer() = default;
+
   /** Sets the buffer data and capacity. */
   void set(T* buf_data, size_t buf_capacity) FMT_NOEXCEPT {
     ptr_ = buf_data;
@@ -655,7 +695,6 @@ template <typename T> class buffer {
 
   buffer(const buffer&) = delete;
   void operator=(const buffer&) = delete;
-  virtual ~buffer() = default;
 
   T* begin() FMT_NOEXCEPT { return ptr_; }
   T* end() FMT_NOEXCEPT { return ptr_ + size_; }
@@ -675,24 +714,26 @@ template <typename T> class buffer {
   /** Returns a pointer to the buffer data. */
   const T* data() const FMT_NOEXCEPT { return ptr_; }
 
-  /**
-    Resizes the buffer. If T is a POD type new elements may not be initialized.
-   */
-  void resize(size_t new_size) {
-    reserve(new_size);
-    size_ = new_size;
-  }
-
   /** Clears this buffer. */
   void clear() { size_ = 0; }
 
-  /** Reserves space to store at least *capacity* elements. */
-  void reserve(size_t new_capacity) {
+  // Tries resizing the buffer to contain *count* elements. If T is a POD type
+  // the new elements may not be initialized.
+  void try_resize(size_t count) {
+    try_reserve(count);
+    size_ = count <= capacity_ ? count : capacity_;
+  }
+
+  // Tries increasing the buffer capacity to *new_capacity*. It can increase the
+  // capacity by a smaller amount than requested but guarantees there is space
+  // for at least one additional element either by increasing the capacity or by
+  // flushing the buffer if it is full.
+  void try_reserve(size_t new_capacity) {
     if (new_capacity > capacity_) grow(new_capacity);
   }
 
   void push_back(const T& value) {
-    reserve(size_ + 1);
+    try_reserve(size_ + 1);
     ptr_[size_++] = value;
   }
 
@@ -705,32 +746,150 @@ template <typename T> class buffer {
   }
 };
 
-// A container-backed buffer.
+struct buffer_traits {
+  explicit buffer_traits(size_t) {}
+  size_t count() const { return 0; }
+  size_t limit(size_t size) { return size; }
+};
+
+class fixed_buffer_traits {
+ private:
+  size_t count_ = 0;
+  size_t limit_;
+
+ public:
+  explicit fixed_buffer_traits(size_t limit) : limit_(limit) {}
+  size_t count() const { return count_; }
+  size_t limit(size_t size) {
+    size_t n = limit_ - count_;
+    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, n < size_t(buffer_size) ? n : size_t(buffer_size)),
+        out_(out) {}
+  ~iterator_buffer() { flush(); }
+
+  OutputIt out() {
+    flush();
+    return out_;
+  }
+  size_t count() const { return Traits::count() + this->size(); }
+};
+
+template <typename T> class iterator_buffer<T*, T> final : public buffer<T> {
+ protected:
+  void grow(size_t) final FMT_OVERRIDE {}
+
+ public:
+  explicit iterator_buffer(T* out, size_t = 0) : buffer<T>(out, 0, ~size_t()) {}
+
+  T* out() { return &*this->end(); }
+};
+
+// A buffer that writes to a container with the contiguous storage.
 template <typename Container>
-class container_buffer : public buffer<typename Container::value_type> {
+class iterator_buffer<std::back_insert_iterator<Container>,
+                      enable_if_t<is_contiguous<Container>::value,
+                                  typename Container::value_type>>
+    final : public buffer<typename Container::value_type> {
  private:
   Container& container_;
 
  protected:
-  void grow(size_t capacity) FMT_OVERRIDE {
+  void grow(size_t capacity) final FMT_OVERRIDE {
     container_.resize(capacity);
     this->set(&container_[0], capacity);
   }
 
  public:
-  explicit container_buffer(Container& c)
+  explicit iterator_buffer(Container& c)
       : buffer<typename Container::value_type>(c.size()), container_(c) {}
+  explicit iterator_buffer(std::back_insert_iterator<Container> out, size_t = 0)
+      : iterator_buffer(get_container(out)) {}
+  std::back_insert_iterator<Container> out() {
+    return std::back_inserter(container_);
+  }
 };
 
-// Extracts a reference to the container from back_insert_iterator.
-template <typename Container>
-inline Container& get_container(std::back_insert_iterator<Container> it) {
-  using bi_iterator = std::back_insert_iterator<Container>;
-  struct accessor : bi_iterator {
-    accessor(bi_iterator iter) : bi_iterator(iter) {}
-    using bi_iterator::container;
-  };
-  return *accessor(it).container;
+// A buffer that counts the number of code units written discarding the output.
+template <typename T = char> class counting_buffer final : public buffer<T> {
+ private:
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+  size_t count_ = 0;
+
+ protected:
+  void grow(size_t) final FMT_OVERRIDE {
+    if (this->size() != buffer_size) return;
+    count_ += this->size();
+    this->clear();
+  }
+
+ public:
+  counting_buffer() : buffer<T>(data_, 0, buffer_size) {}
+
+  size_t count() { return count_ + this->size(); }
+};
+
+// An output iterator that appends to the buffer.
+// It is used to reduce symbol sizes for the common case.
+template <typename T>
+class buffer_appender : public std::back_insert_iterator<buffer<T>> {
+  using base = std::back_insert_iterator<buffer<T>>;
+
+ public:
+  explicit buffer_appender(buffer<T>& buf) : base(buf) {}
+  buffer_appender(base it) : base(it) {}
+
+  buffer_appender& operator++() {
+    base::operator++();
+    return *this;
+  }
+
+  buffer_appender operator++(int) {
+    buffer_appender tmp = *this;
+    ++*this;
+    return tmp;
+  }
+};
+
+// Maps an output iterator into a buffer.
+template <typename T, typename OutputIt>
+iterator_buffer<OutputIt, T> get_buffer(OutputIt);
+template <typename T> buffer<T>& get_buffer(buffer_appender<T>);
+
+template <typename OutputIt> OutputIt get_buffer_init(OutputIt out) {
+  return out;
+}
+template <typename T> buffer<T>& get_buffer_init(buffer_appender<T> out) {
+  return get_container(out);
+}
+
+template <typename Buffer>
+auto get_iterator(Buffer& buf) -> decltype(buf.out()) {
+  return buf.out();
+}
+template <typename T> buffer_appender<T> get_iterator(buffer<T>& buf) {
+  return buffer_appender<T>(buf);
 }
 
 template <typename T, typename Char = char, typename Enable = void>
@@ -759,7 +918,8 @@ template <typename Char> struct named_arg_info {
 template <typename T, typename Char, size_t NUM_ARGS, size_t NUM_NAMED_ARGS>
 struct arg_data {
   // args_[0].named_args points to named_args_ to avoid bloating format_args.
-  T args_[1 + (NUM_ARGS != 0 ? NUM_ARGS : 1)];
+  // +1 to workaround a bug in gcc 7.5 that causes duplicated-branches warning.
+  T args_[1 + (NUM_ARGS != 0 ? NUM_ARGS : +1)];
   named_arg_info<Char> named_args_[NUM_NAMED_ARGS];
 
   template <typename... U>
@@ -771,7 +931,8 @@ struct arg_data {
 
 template <typename T, typename Char, size_t NUM_ARGS>
 struct arg_data<T, Char, NUM_ARGS, 0> {
-  T args_[NUM_ARGS != 0 ? NUM_ARGS : 1];
+  // +1 to workaround a bug in gcc 7.5 that causes duplicated-branches warning.
+  T args_[NUM_ARGS != 0 ? NUM_ARGS : +1];
 
   template <typename... U>
   FMT_INLINE arg_data(const U&... init) : args_{init...} {}
@@ -959,6 +1120,8 @@ enum { long_short = sizeof(long) == sizeof(int) };
 using long_type = conditional_t<long_short, int, long long>;
 using ulong_type = conditional_t<long_short, unsigned, unsigned long long>;
 
+struct unformattable {};
+
 // Maps formatting arguments to core types.
 template <typename Context> struct arg_mapper {
   using char_type = typename Context::char_type;
@@ -1067,15 +1230,7 @@ template <typename Context> struct arg_mapper {
     return map(val.value);
   }
 
-  int map(...) {
-    constexpr bool formattable = sizeof(Context) == 0;
-    static_assert(
-        formattable,
-        "Cannot format argument. To make type T formattable provide a "
-        "formatter<T> specialization: "
-        "https://fmt.dev/latest/api.html#formatting-user-defined-types");
-    return 0;
-  }
+  unformattable map(...) { return {}; }
 };
 
 // A type constant after applying arg_mapper<Context>.
@@ -1199,15 +1354,25 @@ FMT_CONSTEXPR_DECL FMT_INLINE auto visit_format_arg(
   return vis(monostate());
 }
 
-// Checks whether T is a container with contiguous storage.
-template <typename T> struct is_contiguous : std::false_type {};
-template <typename Char>
-struct is_contiguous<std::basic_string<Char>> : std::true_type {};
-template <typename Char>
-struct is_contiguous<detail::buffer<Char>> : std::true_type {};
+template <typename T> struct formattable : std::false_type {};
 
 namespace detail {
 
+// A workaround for gcc 4.8 to make void_t work in a SFINAE context.
+template <typename... Ts> struct void_t_impl { using type = void; };
+template <typename... Ts>
+using void_t = typename detail::void_t_impl<Ts...>::type;
+
+template <typename It, typename T, typename Enable = void>
+struct is_output_iterator : std::false_type {};
+
+template <typename It, typename T>
+struct is_output_iterator<
+    It, T,
+    void_t<typename std::iterator_traits<It>::iterator_category,
+           decltype(*std::declval<It>() = std::declval<T>())>>
+    : std::true_type {};
+
 template <typename OutputIt>
 struct is_back_insert_iterator : std::false_type {};
 template <typename Container>
@@ -1219,6 +1384,9 @@ struct is_contiguous_back_insert_iterator : std::false_type {};
 template <typename Container>
 struct is_contiguous_back_insert_iterator<std::back_insert_iterator<Container>>
     : is_contiguous<Container> {};
+template <typename Char>
+struct is_contiguous_back_insert_iterator<buffer_appender<Char>>
+    : std::true_type {};
 
 // A type-erased reference to an std::locale to avoid heavy <locale> include.
 class locale_ref {
@@ -1250,13 +1418,24 @@ FMT_CONSTEXPR basic_format_arg<Context> make_arg(const T& value) {
   return arg;
 }
 
+template <typename T> int check(unformattable) {
+  static_assert(
+      formattable<T>(),
+      "Cannot format an argument. To make type T formattable provide a "
+      "formatter<T> specialization: https://fmt.dev/latest/api.html#udt");
+  return 0;
+}
+template <typename T, typename U> inline const U& check(const U& val) {
+  return val;
+}
+
 // The type template parameter is there to avoid an ODR violation when using
 // a fallback formatter in one translation unit and an implicit conversion in
 // another (not recommended).
 template <bool IS_PACKED, typename Context, type, typename T,
           FMT_ENABLE_IF(IS_PACKED)>
 inline value<Context> make_arg(const T& val) {
-  return arg_mapper<Context>().map(val);
+  return check<T>(arg_mapper<Context>().map(val));
 }
 
 template <bool IS_PACKED, typename Context, type, typename T,
@@ -1356,13 +1535,13 @@ template <typename OutputIt, typename Char> class basic_format_context {
 
 template <typename Char>
 using buffer_context =
-    basic_format_context<std::back_insert_iterator<detail::buffer<Char>>, Char>;
+    basic_format_context<detail::buffer_appender<Char>, Char>;
 using format_context = buffer_context<char>;
 using wformat_context = buffer_context<wchar_t>;
 
-// Workaround a bug in gcc: https://stackoverflow.com/q/62767544/471164.
+// Workaround an alias issue: https://stackoverflow.com/q/62767544/471164.
 #define FMT_BUFFER_CONTEXT(Char) \
-  basic_format_context<std::back_insert_iterator<detail::buffer<Char>>, Char>
+  basic_format_context<detail::buffer_appender<Char>, Char>
 
 /**
   \rst
@@ -1414,7 +1593,7 @@ class format_arg_store
 
 /**
   \rst
-  Constructs an `~fmt::format_arg_store` object that contains references to
+  Constructs a `~fmt::format_arg_store` object that contains references to
   arguments and can be implicitly converted to `~fmt::format_args`. `Context`
   can be omitted in which case it defaults to `~fmt::context`.
   See `~fmt::arg` for lifetime considerations.
@@ -1428,6 +1607,27 @@ inline format_arg_store<Context, Args...> make_format_args(
 
 /**
   \rst
+  Constructs a `~fmt::format_arg_store` object that contains references
+  to arguments and can be implicitly converted to `~fmt::format_args`.
+  If ``format_str`` is a compile-time string then `make_args_checked` checks
+  its validity at compile time.
+  \endrst
+ */
+template <typename... Args, typename S, typename Char = char_t<S>>
+inline auto make_args_checked(const S& format_str,
+                              const remove_reference_t<Args>&... args)
+    -> format_arg_store<buffer_context<Char>, remove_reference_t<Args>...> {
+  static_assert(
+      detail::count<(
+              std::is_base_of<detail::view, remove_reference_t<Args>>::value &&
+              std::is_reference<Args>::value)...>() == 0,
+      "passing views as lvalues is disallowed");
+  detail::check_format_string<Args...>(format_str);
+  return {args...};
+}
+
+/**
+  \rst
   Returns a named argument to be used in a formatting function. It should only
   be used in a call to a formatting function.
 
@@ -1713,7 +1913,7 @@ template <typename Context> class basic_format_args {
   }
 
   template <typename Char> int get_id(basic_string_view<Char> name) const {
-    if (!has_named_args()) return {};
+    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) {
@@ -1741,29 +1941,6 @@ struct wformat_args : basic_format_args<wformat_context> {
 
 namespace detail {
 
-// Reports a compile-time error if S is not a valid format string.
-template <typename..., typename S, FMT_ENABLE_IF(!is_compile_string<S>::value)>
-FMT_INLINE void check_format_string(const S&) {
-#ifdef FMT_ENFORCE_COMPILE_STRING
-  static_assert(is_compile_string<S>::value,
-                "FMT_ENFORCE_COMPILE_STRING requires all format strings to use "
-                "FMT_STRING.");
-#endif
-}
-template <typename..., typename S, FMT_ENABLE_IF(is_compile_string<S>::value)>
-void check_format_string(S);
-
-template <typename... Args, typename S, typename Char = char_t<S>>
-inline format_arg_store<buffer_context<Char>, remove_reference_t<Args>...>
-make_args_checked(const S& format_str,
-                  const remove_reference_t<Args>&... args) {
-  static_assert(count<(std::is_base_of<view, remove_reference_t<Args>>::value &&
-                       std::is_reference<Args>::value)...>() == 0,
-                "passing views as lvalues is disallowed");
-  check_format_string<Args...>(format_str);
-  return {args...};
-}
-
 template <typename Char, FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
 std::basic_string<Char> vformat(
     basic_string_view<Char> format_str,
@@ -1772,9 +1949,10 @@ std::basic_string<Char> vformat(
 FMT_API std::string vformat(string_view format_str, format_args args);
 
 template <typename Char>
-typename FMT_BUFFER_CONTEXT(Char)::iterator vformat_to(
+void vformat_to(
     buffer<Char>& buf, basic_string_view<Char> format_str,
-    basic_format_args<FMT_BUFFER_CONTEXT(type_identity_t<Char>)> args);
+    basic_format_args<FMT_BUFFER_CONTEXT(type_identity_t<Char>)> args,
+    detail::locale_ref loc = {});
 
 template <typename Char, typename Args,
           FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
@@ -1789,26 +1967,80 @@ inline void vprint_mojibake(std::FILE*, string_view, format_args) {}
 /** Formats a string and writes the output to ``out``. */
 // GCC 8 and earlier cannot handle std::back_insert_iterator<Container> with
 // vformat_to<ArgFormatter>(...) overload, so SFINAE on iterator type instead.
-template <
-    typename OutputIt, typename S, typename Char = char_t<S>,
-    FMT_ENABLE_IF(detail::is_contiguous_back_insert_iterator<OutputIt>::value)>
+template <typename OutputIt, typename S, typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value)>
 OutputIt vformat_to(
     OutputIt out, const S& format_str,
     basic_format_args<buffer_context<type_identity_t<Char>>> args) {
-  auto& c = detail::get_container(out);
-  detail::container_buffer<remove_reference_t<decltype(c)>> buf(c);
+  decltype(detail::get_buffer<Char>(out)) buf(detail::get_buffer_init(out));
   detail::vformat_to(buf, to_string_view(format_str), args);
-  return out;
+  return detail::get_iterator(buf);
+}
+
+/**
+ \rst
+ Formats arguments, writes the result to the output iterator ``out`` and returns
+ the iterator past the end of the output range.
+
+ **Example**::
+
+   std::vector<char> out;
+   fmt::format_to(std::back_inserter(out), "{}", 42);
+ \endrst
+ */
+// We cannot use FMT_ENABLE_IF because of a bug in gcc 8.3.
+template <typename OutputIt, typename S, typename... Args,
+          bool enable = detail::is_output_iterator<OutputIt, char_t<S>>::value>
+inline auto format_to(OutputIt out, const S& format_str, Args&&... args) ->
+    typename std::enable_if<enable, OutputIt>::type {
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
+  return vformat_to(out, to_string_view(format_str), vargs);
+}
+
+template <typename OutputIt> struct format_to_n_result {
+  /** Iterator past the end of the output range. */
+  OutputIt out;
+  /** Total (not truncated) output size. */
+  size_t size;
+};
+
+template <typename OutputIt, typename Char, typename... Args,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value)>
+inline format_to_n_result<OutputIt> vformat_to_n(
+    OutputIt out, size_t n, basic_string_view<Char> format_str,
+    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
+  detail::iterator_buffer<OutputIt, Char, detail::fixed_buffer_traits> buf(out,
+                                                                           n);
+  detail::vformat_to(buf, format_str, args);
+  return {buf.out(), buf.count()};
+}
+
+/**
+ \rst
+ Formats arguments, writes up to ``n`` characters of the result to the output
+ iterator ``out`` and returns the total output size and the iterator past the
+ end of the output range.
+ \endrst
+ */
+template <typename OutputIt, typename S, typename... Args,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char_t<S>>::value)>
+inline format_to_n_result<OutputIt> format_to_n(OutputIt out, size_t n,
+                                                const S& format_str,
+                                                const Args&... args) {
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
+  return vformat_to_n(out, n, to_string_view(format_str), vargs);
 }
 
-template <typename Container, typename S, typename... Args,
-          FMT_ENABLE_IF(
-              is_contiguous<Container>::value&& detail::is_string<S>::value)>
-inline std::back_insert_iterator<Container> format_to(
-    std::back_insert_iterator<Container> out, const S& format_str,
-    Args&&... args) {
-  return vformat_to(out, to_string_view(format_str),
-                    detail::make_args_checked<Args...>(format_str, args...));
+/**
+  Returns the number of characters in the output of
+  ``format(format_str, args...)``.
+ */
+template <typename... Args>
+inline size_t formatted_size(string_view format_str, Args&&... args) {
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
+  detail::counting_buffer<> buf;
+  detail::vformat_to(buf, format_str, vargs);
+  return buf.count();
 }
 
 template <typename S, typename Char = char_t<S>>
@@ -1832,7 +2064,7 @@ FMT_INLINE std::basic_string<Char> vformat(
 // std::basic_string<char_t<S>> to reduce the symbol size.
 template <typename S, typename... Args, typename Char = char_t<S>>
 FMT_INLINE std::basic_string<Char> format(const S& format_str, Args&&... args) {
-  const auto& vargs = detail::make_args_checked<Args...>(format_str, args...);
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
   return detail::vformat(to_string_view(format_str), vargs);
 }
 
@@ -1852,7 +2084,7 @@ FMT_API void vprint(std::FILE*, string_view, format_args);
  */
 template <typename S, typename... Args, typename Char = char_t<S>>
 inline void print(std::FILE* f, const S& format_str, Args&&... args) {
-  const auto& vargs = detail::make_args_checked<Args...>(format_str, args...);
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
   return detail::is_unicode<Char>()
              ? vprint(f, to_string_view(format_str), vargs)
              : detail::vprint_mojibake(f, to_string_view(format_str), vargs);
@@ -1871,7 +2103,7 @@ inline void print(std::FILE* f, const S& format_str, Args&&... args) {
  */
 template <typename S, typename... Args, typename Char = char_t<S>>
 inline void print(const S& format_str, Args&&... args) {
-  const auto& vargs = detail::make_args_checked<Args...>(format_str, args...);
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
   return detail::is_unicode<Char>()
              ? vprint(to_string_view(format_str), vargs)
              : detail::vprint_mojibake(stdout, to_string_view(format_str),
diff --git a/include/fmt/format-inl.h b/include/fmt/format-inl.h
index d8c9c8a5..b7cb3209 100644
--- a/include/fmt/format-inl.h
+++ b/include/fmt/format-inl.h
@@ -13,32 +13,19 @@
 #include <climits>
 #include <cmath>
 #include <cstdarg>
-#include <cstring>  // for std::memmove
+#include <cstring>  // std::memmove
 #include <cwchar>
 #include <exception>
 
-#include "format.h"
-#if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
 #  include <locale>
 #endif
 
 #ifdef _WIN32
-#  if !defined(NOMINMAX) && !defined(WIN32_LEAN_AND_MEAN)
-#    define NOMINMAX
-#    define WIN32_LEAN_AND_MEAN
-#    include <windows.h>
-#    undef WIN32_LEAN_AND_MEAN
-#    undef NOMINMAX
-#  else
-#    include <windows.h>
-#  endif
-#  include <io.h>
+#  include <io.h>  // _isatty
 #endif
 
-#ifdef _MSC_VER
-#  pragma warning(push)
-#  pragma warning(disable : 4702)  // unreachable code
-#endif
+#include "format.h"
 
 // Dummy implementations of strerror_r and strerror_s called if corresponding
 // system functions are not available.
@@ -79,8 +66,8 @@ inline int fmt_snprintf(char* buffer, size_t size, const char* format, ...) {
 //   ERANGE - buffer is not large enough to store the error message
 //   other  - failure
 // Buffer should be at least of size 1.
-FMT_FUNC int safe_strerror(int error_code, char*& buffer,
-                           size_t buffer_size) FMT_NOEXCEPT {
+inline int safe_strerror(int error_code, char*& buffer,
+                         size_t buffer_size) FMT_NOEXCEPT {
   FMT_ASSERT(buffer != nullptr && buffer_size != 0, "invalid buffer");
 
   class dispatcher {
@@ -145,7 +132,7 @@ FMT_FUNC void format_error_code(detail::buffer<char>& out, int error_code,
   // Report error code making sure that the output fits into
   // inline_buffer_size to avoid dynamic memory allocation and potential
   // bad_alloc.
-  out.resize(0);
+  out.try_resize(0);
   static const char SEP[] = ": ";
   static const char ERROR_STR[] = "error ";
   // Subtract 2 to account for terminating null characters in SEP and ERROR_STR.
@@ -156,7 +143,7 @@ FMT_FUNC void format_error_code(detail::buffer<char>& out, int error_code,
     ++error_code_size;
   }
   error_code_size += detail::to_unsigned(detail::count_digits(abs_value));
-  auto it = std::back_inserter(out);
+  auto it = buffer_appender<char>(out);
   if (message.size() <= inline_buffer_size - error_code_size)
     format_to(it, "{}{}", message, SEP);
   format_to(it, "{}{}", ERROR_STR, error_code);
@@ -173,8 +160,8 @@ FMT_FUNC void report_error(format_func func, int error_code,
 }
 
 // A wrapper around fwrite that throws on error.
-FMT_FUNC void fwrite_fully(const void* ptr, size_t size, size_t count,
-                           FILE* stream) {
+inline void fwrite_fully(const void* ptr, size_t size, size_t count,
+                         FILE* stream) {
   size_t written = std::fwrite(ptr, size, count, stream);
   if (written < count) FMT_THROW(system_error(errno, "cannot write to file"));
 }
@@ -242,26 +229,23 @@ template <> FMT_FUNC int count_digits<4>(detail::fallback_uintptr n) {
 
 template <typename T>
 const typename basic_data<T>::digit_pair basic_data<T>::digits[] = {
-    {'0', '0'},  {'0', '1'},  {'0', '2'},  {'0', '3'},  {'0', '4'},
-    {'0', '5'},  {'0', '6'},  {'0', '7'},  {'0', '8'},  {'0', '9'},
-    {'1', '0'},  {'1', '1'},  {'1', '2'},  {'1', '3'},  {'1', '4'},
-    {'1', '5'},  {'1', '6'},  {'1', '7'},  {'1', '8'},  {'1', '9'},
-    {'2', '0'},  {'2', '1'},  {'2', '2'},  {'2', '3'},  {'2', '4'},
-    {'2', '5'},  {'2', '6'},  {'2', '7'},  {'2', '8'},  {'2', '9'},
-    {'3', '0'},  {'3', '1'},  {'3', '2'},  {'3', '3'},  {'3', '4'},
-    {'3', '5'},  {'3', '6'},  {'3', '7'},  {'3', '8'},  {'3', '9'},
-    {'4', '0'},  {'4', '1'},  {'4', '2'},  {'4', '3'},  {'4', '4'},
-    {'4', '5'},  {'4', '6'},  {'4', '7'},  {'4', '8'},  {'4', '9'},
-    {'5', '0'},  {'5', '1'},  {'5', '2'},  {'5', '3'},  {'5', '4'},
-    {'5', '5'},  {'5', '6'},  {'5', '7'},  {'5', '8'},  {'5', '9'},
-    {'6', '0'},  {'6', '1'},  {'6', '2'},  {'6', '3'},  {'6', '4'},
-    {'6', '5'},  {'6', '6'},  {'6', '7'},  {'6', '8'},  {'6', '9'},
-    {'7', '0'},  {'7', '1'},  {'7', '2'},  {'7', '3'},  {'7', '4'},
-    {'7', '5'},  {'7', '6'},  {'7', '7'},  {'7', '8'},  {'7', '9'},
-    {'8', '0'},  {'8', '1'},  {'8', '2'},  {'8', '3'},  {'8', '4'},
-    {'8', '5'},  {'8', '6'},  {'8', '7'},  {'8', '8'},  {'8', '9'},
-    {'9', '0'},  {'9', '1'},  {'9', '2'},  {'9', '3'},  {'9', '4'},
-    {'9', '5'},  {'9', '6'},  {'9', '7'},  {'9', '8'},  {'9', '9'}};
+    {'0', '0'}, {'0', '1'}, {'0', '2'}, {'0', '3'}, {'0', '4'}, {'0', '5'},
+    {'0', '6'}, {'0', '7'}, {'0', '8'}, {'0', '9'}, {'1', '0'}, {'1', '1'},
+    {'1', '2'}, {'1', '3'}, {'1', '4'}, {'1', '5'}, {'1', '6'}, {'1', '7'},
+    {'1', '8'}, {'1', '9'}, {'2', '0'}, {'2', '1'}, {'2', '2'}, {'2', '3'},
+    {'2', '4'}, {'2', '5'}, {'2', '6'}, {'2', '7'}, {'2', '8'}, {'2', '9'},
+    {'3', '0'}, {'3', '1'}, {'3', '2'}, {'3', '3'}, {'3', '4'}, {'3', '5'},
+    {'3', '6'}, {'3', '7'}, {'3', '8'}, {'3', '9'}, {'4', '0'}, {'4', '1'},
+    {'4', '2'}, {'4', '3'}, {'4', '4'}, {'4', '5'}, {'4', '6'}, {'4', '7'},
+    {'4', '8'}, {'4', '9'}, {'5', '0'}, {'5', '1'}, {'5', '2'}, {'5', '3'},
+    {'5', '4'}, {'5', '5'}, {'5', '6'}, {'5', '7'}, {'5', '8'}, {'5', '9'},
+    {'6', '0'}, {'6', '1'}, {'6', '2'}, {'6', '3'}, {'6', '4'}, {'6', '5'},
+    {'6', '6'}, {'6', '7'}, {'6', '8'}, {'6', '9'}, {'7', '0'}, {'7', '1'},
+    {'7', '2'}, {'7', '3'}, {'7', '4'}, {'7', '5'}, {'7', '6'}, {'7', '7'},
+    {'7', '8'}, {'7', '9'}, {'8', '0'}, {'8', '1'}, {'8', '2'}, {'8', '3'},
+    {'8', '4'}, {'8', '5'}, {'8', '6'}, {'8', '7'}, {'8', '8'}, {'8', '9'},
+    {'9', '0'}, {'9', '1'}, {'9', '2'}, {'9', '3'}, {'9', '4'}, {'9', '5'},
+    {'9', '6'}, {'9', '7'}, {'9', '8'}, {'9', '9'}};
 
 template <typename T>
 const char basic_data<T>::hex_digits[] = "0123456789abcdef";
@@ -277,18 +261,18 @@ const uint64_t basic_data<T>::powers_of_10_64[] = {
     10000000000000000000ULL};
 
 template <typename T>
-const uint32_t basic_data<T>::zero_or_powers_of_10_32[] = {0,
+const uint32_t basic_data<T>::zero_or_powers_of_10_32[] = {0, 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),
+    0, 0, FMT_POWERS_OF_10(1), FMT_POWERS_OF_10(1000000000ULL),
     10000000000000000000ULL};
 
 // Normalized 64-bit significands of pow(10, k), for k = -348, -340, ..., 340.
 // These are generated by support/compute-powers.py.
 template <typename T>
-const uint64_t basic_data<T>::pow10_significands[] = {
+const uint64_t basic_data<T>::grisu_pow10_significands[] = {
     0xfa8fd5a0081c0288, 0xbaaee17fa23ebf76, 0x8b16fb203055ac76,
     0xcf42894a5dce35ea, 0x9a6bb0aa55653b2d, 0xe61acf033d1a45df,
     0xab70fe17c79ac6ca, 0xff77b1fcbebcdc4f, 0xbe5691ef416bd60c,
@@ -323,7 +307,7 @@ 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,
@@ -334,6 +318,744 @@ const int16_t basic_data<T>::pow10_exponents[] = {
     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;";
@@ -366,6 +1088,10 @@ class fp {
  private:
   using significand_type = uint64_t;
 
+  template <typename Float>
+  using is_supported_float = bool_constant<sizeof(Float) == sizeof(uint64_t) ||
+                                           sizeof(Float) == sizeof(uint32_t)>;
+
  public:
   significand_type f;
   int e;
@@ -388,63 +1114,38 @@ class fp {
   template <typename Double> explicit fp(Double d) { assign(d); }
 
   // Assigns d to this and return true iff predecessor is closer than successor.
-  template <typename Double, FMT_ENABLE_IF(sizeof(Double) == sizeof(uint64_t))>
-  bool assign(Double d) {
-    // Assume double is in the format [sign][exponent][significand].
-    using limits = std::numeric_limits<Double>;
+  template <typename Float, FMT_ENABLE_IF(is_supported_float<Float>::value)>
+  bool assign(Float d) {
+    // Assume float is in the format [sign][exponent][significand].
+    using limits = std::numeric_limits<Float>;
+    const int float_significand_size = limits::digits - 1;
     const int exponent_size =
-        bits<Double>::value - double_significand_size - 1;  // -1 for sign
-    const uint64_t significand_mask = implicit_bit - 1;
+        bits<Float>::value - float_significand_size - 1;  // -1 for sign
+    const uint64_t float_implicit_bit = 1ULL << float_significand_size;
+    const uint64_t significand_mask = float_implicit_bit - 1;
     const uint64_t exponent_mask = (~0ULL >> 1) & ~significand_mask;
     const int exponent_bias = (1 << exponent_size) - limits::max_exponent - 1;
-    auto u = bit_cast<uint64_t>(d);
+    constexpr bool is_double = sizeof(Float) == sizeof(uint64_t);
+    auto u = bit_cast<conditional_t<is_double, uint64_t, uint32_t>>(d);
     f = u & significand_mask;
     int biased_e =
-        static_cast<int>((u & exponent_mask) >> double_significand_size);
+        static_cast<int>((u & exponent_mask) >> float_significand_size);
     // Predecessor is closer if d is a normalized power of 2 (f == 0) other than
     // the smallest normalized number (biased_e > 1).
     bool is_predecessor_closer = f == 0 && biased_e > 1;
     if (biased_e != 0)
-      f += implicit_bit;
+      f += float_implicit_bit;
     else
       biased_e = 1;  // Subnormals use biased exponent 1 (min exponent).
-    e = biased_e - exponent_bias - double_significand_size;
+    e = biased_e - exponent_bias - float_significand_size;
     return is_predecessor_closer;
   }
 
-  template <typename Double, FMT_ENABLE_IF(sizeof(Double) != sizeof(uint64_t))>
-  bool assign(Double) {
+  template <typename Float, FMT_ENABLE_IF(!is_supported_float<Float>::value)>
+  bool assign(Float) {
     *this = fp();
     return false;
   }
-
-  // Assigns d to this together with computing lower and upper boundaries,
-  // where a boundary is a value half way between the number and its predecessor
-  // (lower) or successor (upper). The upper boundary is normalized and lower
-  // has the same exponent but may be not normalized.
-  template <typename Double> boundaries assign_with_boundaries(Double d) {
-    bool is_lower_closer = assign(d);
-    fp lower =
-        is_lower_closer ? fp((f << 2) - 1, e - 2) : fp((f << 1) - 1, e - 1);
-    // 1 in normalize accounts for the exponent shift above.
-    fp upper = normalize<1>(fp((f << 1) + 1, e - 1));
-    lower.f <<= lower.e - upper.e;
-    return boundaries{lower.f, upper.f};
-  }
-
-  template <typename Double> boundaries assign_float_with_boundaries(Double d) {
-    assign(d);
-    constexpr int min_normal_e = std::numeric_limits<float>::min_exponent -
-                                 std::numeric_limits<double>::digits;
-    significand_type half_ulp = 1 << (std::numeric_limits<double>::digits -
-                                      std::numeric_limits<float>::digits - 1);
-    if (min_normal_e > e) half_ulp <<= min_normal_e - e;
-    fp upper = normalize<0>(fp(f + half_ulp, e));
-    fp lower = fp(
-        f - (half_ulp >> ((f == implicit_bit && e > min_normal_e) ? 1 : 0)), e);
-    lower.f <<= lower.e - upper.e;
-    return boundaries{lower.f, upper.f};
-  }
 };
 
 // Normalizes the value converted from double and multiplied by (1 << SHIFT).
@@ -488,11 +1189,12 @@ inline fp operator*(fp x, fp y) { return {multiply(x.f, y.f), x.e + y.e + 64}; }
 // Returns a cached power of 10 `c_k = c_k.f * pow(2, c_k.e)` such that its
 // (binary) exponent satisfies `min_exponent <= c_k.e <= min_exponent + 28`.
 inline fp get_cached_power(int min_exponent, int& pow10_exponent) {
-  const int64_t one_over_log2_10 = 0x4d104d42;  // round(pow(2, 32) / log2(10))
+  const int shift = 32;
+  const auto significand = static_cast<int64_t>(data::log10_2_significand);
   int index = static_cast<int>(
-      ((min_exponent + fp::significand_size - 1) * one_over_log2_10 +
-       ((int64_t(1) << 32) - 1))  // ceil
-      >> 32                       // arithmetic shift
+      ((min_exponent + fp::significand_size - 1) * (significand >> shift) +
+       ((int64_t(1) << shift) - 1))  // ceil
+      >> 32                          // arithmetic shift
   );
   // Decimal exponent of the first (smallest) cached power of 10.
   const int first_dec_exp = -348;
@@ -500,7 +1202,8 @@ inline fp get_cached_power(int min_exponent, int& pow10_exponent) {
   const int dec_exp_step = 8;
   index = (index - first_dec_exp - 1) / dec_exp_step + 1;
   pow10_exponent = first_dec_exp + index * dec_exp_step;
-  return {data::pow10_significands[index], data::pow10_exponents[index]};
+  return {data::grisu_pow10_significands[index],
+          data::grisu_pow10_exponents[index]};
 }
 
 // A simple accumulator to hold the sums of terms in bigint::square if uint128_t
@@ -559,9 +1262,8 @@ class bigint {
     FMT_ASSERT(compare(*this, other) >= 0, "");
     bigit borrow = 0;
     int i = other.exp_ - exp_;
-    for (size_t j = 0, n = other.bigits_.size(); j != n; ++i, ++j) {
+    for (size_t j = 0, n = other.bigits_.size(); j != n; ++i, ++j)
       subtract_bigits(i, other.bigits_[j], borrow);
-    }
     while (borrow > 0) subtract_bigits(i, 0, borrow);
     remove_leading_zeros();
   }
@@ -733,22 +1435,26 @@ class bigint {
     exp_ *= 2;
   }
 
+  // If this bigint has a bigger exponent than other, adds trailing zero to make
+  // exponents equal. This simplifies some operations such as subtraction.
+  void align(const bigint& other) {
+    int exp_difference = exp_ - other.exp_;
+    if (exp_difference <= 0) return;
+    int num_bigits = static_cast<int>(bigits_.size());
+    bigits_.resize(to_unsigned(num_bigits + exp_difference));
+    for (int i = num_bigits - 1, j = i + exp_difference; i >= 0; --i, --j)
+      bigits_[j] = bigits_[i];
+    std::uninitialized_fill_n(bigits_.data(), exp_difference, 0);
+    exp_ -= exp_difference;
+  }
+
   // Divides this bignum by divisor, assigning the remainder to this and
   // returning the quotient.
   int divmod_assign(const bigint& divisor) {
     FMT_ASSERT(this != &divisor, "");
     if (compare(*this, divisor) < 0) return 0;
-    int num_bigits = static_cast<int>(bigits_.size());
     FMT_ASSERT(divisor.bigits_[divisor.bigits_.size() - 1u] != 0, "");
-    int exp_difference = exp_ - divisor.exp_;
-    if (exp_difference > 0) {
-      // Align bigints by adding trailing zeros to simplify subtraction.
-      bigits_.resize(to_unsigned(num_bigits + exp_difference));
-      for (int i = num_bigits - 1, j = i + exp_difference; i >= 0; --i, --j)
-        bigits_[j] = bigits_[i];
-      std::uninitialized_fill_n(bigits_.data(), exp_difference, 0);
-      exp_ -= exp_difference;
-    }
+    align(divisor);
     int quotient = 0;
     do {
       subtract_aligned(divisor);
@@ -788,20 +1494,6 @@ enum result {
 };
 }
 
-// A version of count_digits optimized for grisu_gen_digits.
-inline int grisu_count_digits(uint32_t n) {
-  if (n < 10) return 1;
-  if (n < 100) return 2;
-  if (n < 1000) return 3;
-  if (n < 10000) return 4;
-  if (n < 100000) return 5;
-  if (n < 1000000) return 6;
-  if (n < 10000000) return 7;
-  if (n < 100000000) return 8;
-  if (n < 1000000000) return 9;
-  return 10;
-}
-
 // Generates output using the Grisu digit-gen algorithm.
 // error: the size of the region (lower, upper) outside of which numbers
 // definitely do not round to value (Delta in Grisu3).
@@ -817,7 +1509,7 @@ FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
   FMT_ASSERT(integral == value.f >> -one.e, "");
   // The fractional part of scaled value (p2 in Grisu) c = value % one.
   uint64_t fractional = value.f & (one.f - 1);
-  exp = grisu_count_digits(integral);  // kappa in Grisu.
+  exp = count_digits(integral);  // kappa in Grisu.
   // Divide by 10 to prevent overflow.
   auto result = handler.on_start(data::powers_of_10_64[exp - 1] << -one.e,
                                  value.f / 10, error * 10, exp);
@@ -867,8 +1559,7 @@ FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
       FMT_ASSERT(false, "invalid number of digits");
     }
     --exp;
-    uint64_t remainder =
-        (static_cast<uint64_t>(integral) << -one.e) + fractional;
+    auto remainder = (static_cast<uint64_t>(integral) << -one.e) + fractional;
     result = handler.on_digit(static_cast<char>('0' + digit),
                               data::powers_of_10_64[exp] << -one.e, remainder,
                               error, exp, true);
@@ -878,8 +1569,7 @@ FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
   for (;;) {
     fractional *= 10;
     error *= 10;
-    char digit =
-        static_cast<char>('0' + static_cast<char>(fractional >> -one.e));
+    char digit = static_cast<char>('0' + (fractional >> -one.e));
     fractional &= one.f - 1;
     --exp;
     result = handler.on_digit(digit, one.f, fractional, error, exp, false);
@@ -916,6 +1606,7 @@ struct fixed_handler {
                           uint64_t error, int, bool integral) {
     FMT_ASSERT(remainder < divisor, "");
     buf[size++] = digit;
+    if (!integral && error >= remainder) return digits::error;
     if (size < precision) return digits::more;
     if (!integral) {
       // Check if error * 2 < divisor with overflow prevention.
@@ -935,59 +1626,684 @@ struct fixed_handler {
     }
     if (buf[0] > '9') {
       buf[0] = '1';
-      buf[size++] = '0';
+      if (fixed)
+        buf[size++] = '0';
+      else
+        ++exp10;
     }
     return digits::done;
   }
 };
 
-// The shortest representation digit handler.
-struct grisu_shortest_handler {
-  char* buf;
-  int size;
-  // Distance between scaled value and upper bound (wp_W in Grisu3).
-  uint64_t diff;
+// Implementation of Dragonbox algorithm: https://github.com/jk-jeon/dragonbox.
+namespace dragonbox {
+// Computes 128-bit result of multiplication of two 64-bit unsigned integers.
+FMT_SAFEBUFFERS inline uint128_wrapper umul128(uint64_t x,
+                                               uint64_t y) FMT_NOEXCEPT {
+#if FMT_USE_INT128
+  return static_cast<uint128_t>(x) * static_cast<uint128_t>(y);
+#elif defined(_MSC_VER) && defined(_M_X64)
+  uint128_wrapper result;
+  result.low_ = _umul128(x, y, &result.high_);
+  return result;
+#else
+  const uint64_t mask = (uint64_t(1) << 32) - uint64_t(1);
+
+  uint64_t a = x >> 32;
+  uint64_t b = x & mask;
+  uint64_t c = y >> 32;
+  uint64_t d = y & mask;
+
+  uint64_t ac = a * c;
+  uint64_t bc = b * c;
+  uint64_t ad = a * d;
+  uint64_t bd = b * d;
+
+  uint64_t intermediate = (bd >> 32) + (ad & mask) + (bc & mask);
+
+  return {ac + (intermediate >> 32) + (ad >> 32) + (bc >> 32),
+          (intermediate << 32) + (bd & mask)};
+#endif
+}
+
+// Computes upper 64 bits of multiplication of two 64-bit unsigned integers.
+FMT_SAFEBUFFERS inline uint64_t umul128_upper64(uint64_t x,
+                                                uint64_t y) FMT_NOEXCEPT {
+#if FMT_USE_INT128
+  auto p = static_cast<uint128_t>(x) * static_cast<uint128_t>(y);
+  return static_cast<uint64_t>(p >> 64);
+#elif defined(_MSC_VER) && defined(_M_X64)
+  return __umulh(x, y);
+#else
+  return umul128(x, y).high();
+#endif
+}
+
+// Computes upper 64 bits of multiplication of a 64-bit unsigned integer and a
+// 128-bit unsigned integer.
+FMT_SAFEBUFFERS inline uint64_t umul192_upper64(uint64_t x, uint128_wrapper y)
+    FMT_NOEXCEPT {
+  uint128_wrapper g0 = umul128(x, y.high());
+  g0 += umul128_upper64(x, y.low());
+  return g0.high();
+}
+
+// Computes upper 32 bits of multiplication of a 32-bit unsigned integer and a
+// 64-bit unsigned integer.
+inline uint32_t umul96_upper32(uint32_t x, uint64_t y) FMT_NOEXCEPT {
+  return static_cast<uint32_t>(umul128_upper64(x, y));
+}
+
+// Computes middle 64 bits of multiplication of a 64-bit unsigned integer and a
+// 128-bit unsigned integer.
+FMT_SAFEBUFFERS inline uint64_t umul192_middle64(uint64_t x, uint128_wrapper y)
+    FMT_NOEXCEPT {
+  uint64_t g01 = x * y.high();
+  uint64_t g10 = umul128_upper64(x, y.low());
+  return g01 + g10;
+}
+
+// Computes lower 64 bits of multiplication of a 32-bit unsigned integer and a
+// 64-bit unsigned integer.
+inline uint64_t umul96_lower64(uint32_t x, uint64_t y) FMT_NOEXCEPT {
+  return x * y;
+}
+
+// Computes floor(log10(pow(2, e))) for e in [-1700, 1700] using the method from
+// https://fmt.dev/papers/Grisu-Exact.pdf#page=5, section 3.4.
+inline int floor_log10_pow2(int e) FMT_NOEXCEPT {
+  FMT_ASSERT(e <= 1700 && e >= -1700, "too large exponent");
+  const int shift = 22;
+  return (e * static_cast<int>(data::log10_2_significand >> (64 - shift))) >>
+         shift;
+}
+
+// Various fast log computations.
+inline int floor_log2_pow10(int e) FMT_NOEXCEPT {
+  FMT_ASSERT(e <= 1233 && e >= -1233, "too large exponent");
+  const uint64_t log2_10_integer_part = 3;
+  const uint64_t log2_10_fractional_digits = 0x5269e12f346e2bf9;
+  const int shift_amount = 19;
+  return (e * static_cast<int>(
+                  (log2_10_integer_part << shift_amount) |
+                  (log2_10_fractional_digits >> (64 - shift_amount)))) >>
+         shift_amount;
+}
+inline int floor_log10_pow2_minus_log10_4_over_3(int e) FMT_NOEXCEPT {
+  FMT_ASSERT(e <= 1700 && e >= -1700, "too large exponent");
+  const uint64_t log10_4_over_3_fractional_digits = 0x1ffbfc2bbc780375;
+  const int shift_amount = 22;
+  return (e * static_cast<int>(data::log10_2_significand >>
+                               (64 - shift_amount)) -
+          static_cast<int>(log10_4_over_3_fractional_digits >>
+                           (64 - shift_amount))) >>
+         shift_amount;
+}
+
+// Returns true iff x is divisible by pow(2, exp).
+inline bool divisible_by_power_of_2(uint32_t x, int exp) FMT_NOEXCEPT {
+  FMT_ASSERT(exp >= 1, "");
+  FMT_ASSERT(x != 0, "");
+#ifdef FMT_BUILTIN_CTZ
+  return FMT_BUILTIN_CTZ(x) >= exp;
+#else
+  return exp < num_bits<uint32_t>() && x == ((x >> exp) << exp);
+#endif
+}
+inline bool divisible_by_power_of_2(uint64_t x, int exp) FMT_NOEXCEPT {
+  FMT_ASSERT(exp >= 1, "");
+  FMT_ASSERT(x != 0, "");
+#ifdef FMT_BUILTIN_CTZLL
+  return FMT_BUILTIN_CTZLL(x) >= exp;
+#else
+  return exp < num_bits<uint64_t>()) && x == ((x >> exp) << exp);
+#endif
+}
+
+// Returns true iff x is divisible by pow(5, exp).
+inline bool divisible_by_power_of_5(uint32_t x, int exp) FMT_NOEXCEPT {
+  FMT_ASSERT(exp <= 10, "too large exponent");
+  return x * data::divtest_table_for_pow5_32[exp].mod_inv <=
+         data::divtest_table_for_pow5_32[exp].max_quotient;
+}
+inline bool divisible_by_power_of_5(uint64_t x, int exp) FMT_NOEXCEPT {
+  FMT_ASSERT(exp <= 23, "too large exponent");
+  return x * data::divtest_table_for_pow5_64[exp].mod_inv <=
+         data::divtest_table_for_pow5_64[exp].max_quotient;
+}
+
+// Replaces n by floor(n / pow(5, N)) returning true if and only if n is
+// divisible by pow(5, N).
+// Precondition: n <= 2 * pow(5, N + 1).
+template <int N>
+bool check_divisibility_and_divide_by_pow5(uint32_t& n) FMT_NOEXCEPT {
+  static constexpr struct {
+    uint32_t magic_number;
+    int bits_for_comparison;
+    uint32_t threshold;
+    int shift_amount;
+  } infos[] = {{0xcccd, 16, 0x3333, 18}, {0xa429, 8, 0x0a, 20}};
+  constexpr auto info = infos[N - 1];
+  n *= info.magic_number;
+  const uint32_t comparison_mask = (1u << info.bits_for_comparison) - 1;
+  bool result = (n & comparison_mask) <= info.threshold;
+  n >>= info.shift_amount;
+  return result;
+}
+
+// Computes floor(n / pow(10, N)) for small n and N.
+// Precondition: n <= pow(10, N + 1).
+template <int N> uint32_t small_division_by_pow10(uint32_t n) FMT_NOEXCEPT {
+  static constexpr struct {
+    uint32_t magic_number;
+    int shift_amount;
+    uint32_t divisor_times_10;
+  } infos[] = {{0xcccd, 19, 100}, {0xa3d8, 22, 1000}};
+  constexpr auto info = infos[N - 1];
+  FMT_ASSERT(n <= info.divisor_times_10, "n is too large");
+  return n * info.magic_number >> info.shift_amount;
+}
+
+// Computes floor(n / 10^(kappa + 1)) (float)
+inline uint32_t divide_by_10_to_kappa_plus_1(uint32_t n) FMT_NOEXCEPT {
+  return n / float_info<float>::big_divisor;
+}
+// Computes floor(n / 10^(kappa + 1)) (double)
+inline uint64_t divide_by_10_to_kappa_plus_1(uint64_t n) FMT_NOEXCEPT {
+  return umul128_upper64(n, 0x83126e978d4fdf3c) >> 9;
+}
+
+// Various subroutines using pow10 cache
+template <class T> struct cache_accessor;
+
+template <> struct cache_accessor<float> {
+  using carrier_uint = float_info<float>::carrier_uint;
+  using cache_entry_type = uint64_t;
+
+  static uint64_t get_cached_power(int k) FMT_NOEXCEPT {
+    FMT_ASSERT(k >= float_info<float>::min_k && k <= float_info<float>::max_k,
+               "k is out of range");
+    return data::dragonbox_pow10_significands_64[k - float_info<float>::min_k];
+  }
+
+  static carrier_uint compute_mul(carrier_uint u,
+                                  const cache_entry_type& cache) FMT_NOEXCEPT {
+    return umul96_upper32(u, cache);
+  }
+
+  static uint32_t compute_delta(const cache_entry_type& cache,
+                                int beta_minus_1) FMT_NOEXCEPT {
+    return static_cast<uint32_t>(cache >> (64 - 1 - beta_minus_1));
+  }
+
+  static bool compute_mul_parity(carrier_uint two_f,
+                                 const cache_entry_type& cache,
+                                 int beta_minus_1) FMT_NOEXCEPT {
+    FMT_ASSERT(beta_minus_1 >= 1, "");
+    FMT_ASSERT(beta_minus_1 < 64, "");
+
+    return ((umul96_lower64(two_f, cache) >> (64 - beta_minus_1)) & 1) != 0;
+  }
+
+  static carrier_uint compute_left_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return static_cast<carrier_uint>(
+        (cache - (cache >> (float_info<float>::significand_bits + 2))) >>
+        (64 - float_info<float>::significand_bits - 1 - beta_minus_1));
+  }
+
+  static carrier_uint compute_right_endpoint_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return static_cast<carrier_uint>(
+        (cache + (cache >> (float_info<float>::significand_bits + 1))) >>
+        (64 - float_info<float>::significand_bits - 1 - beta_minus_1));
+  }
+
+  static carrier_uint compute_round_up_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return (static_cast<carrier_uint>(
+                cache >>
+                (64 - float_info<float>::significand_bits - 2 - beta_minus_1)) +
+            1) /
+           2;
+  }
+};
+
+template <> struct cache_accessor<double> {
+  using carrier_uint = float_info<double>::carrier_uint;
+  using cache_entry_type = uint128_wrapper;
+
+  static uint128_wrapper get_cached_power(int k) FMT_NOEXCEPT {
+    FMT_ASSERT(k >= float_info<double>::min_k && k <= float_info<double>::max_k,
+               "k is out of range");
+
+#if FMT_USE_FULL_CACHE_DRAGONBOX
+    return data::dragonbox_pow10_significands_128[k -
+                                                  float_info<double>::min_k];
+#else
+    static const int compression_ratio = 27;
+
+    // Compute base index.
+    int cache_index = (k - float_info<double>::min_k) / compression_ratio;
+    int kb = cache_index * compression_ratio + float_info<double>::min_k;
+    int offset = k - kb;
+
+    // Get base cache.
+    uint128_wrapper base_cache =
+        data::dragonbox_pow10_significands_128[cache_index];
+    if (offset == 0) return base_cache;
+
+    // Compute the required amount of bit-shift.
+    int alpha = floor_log2_pow10(kb + offset) - floor_log2_pow10(kb) - offset;
+    FMT_ASSERT(alpha > 0 && alpha < 64, "shifting error detected");
+
+    // Try to recover the real cache.
+    uint64_t pow5 = data::powers_of_5_64[offset];
+    uint128_wrapper recovered_cache = umul128(base_cache.high(), pow5);
+    uint128_wrapper middle_low =
+        umul128(base_cache.low() - (kb < 0 ? 1 : 0), pow5);
+
+    recovered_cache += middle_low.high();
+
+    uint64_t high_to_middle = recovered_cache.high() << (64 - alpha);
+    uint64_t middle_to_low = recovered_cache.low() << (64 - alpha);
+
+    recovered_cache =
+        uint128_wrapper{(recovered_cache.low() >> alpha) | high_to_middle,
+                        ((middle_low.low() >> alpha) | middle_to_low)};
+
+    if (kb < 0) recovered_cache += 1;
+
+    // Get error.
+    int error_idx = (k - float_info<double>::min_k) / 16;
+    uint32_t error = (data::dragonbox_pow10_recovery_errors[error_idx] >>
+                      ((k - float_info<double>::min_k) % 16) * 2) &
+                     0x3;
+
+    // Add the error back.
+    FMT_ASSERT(recovered_cache.low() + error >= recovered_cache.low(), "");
+    return {recovered_cache.high(), recovered_cache.low() + error};
+#endif
+  }
+
+  static carrier_uint compute_mul(carrier_uint u,
+                                  const cache_entry_type& cache) FMT_NOEXCEPT {
+    return umul192_upper64(u, cache);
+  }
+
+  static uint32_t compute_delta(cache_entry_type const& cache,
+                                int beta_minus_1) FMT_NOEXCEPT {
+    return static_cast<uint32_t>(cache.high() >> (64 - 1 - beta_minus_1));
+  }
+
+  static bool compute_mul_parity(carrier_uint two_f,
+                                 const cache_entry_type& cache,
+                                 int beta_minus_1) FMT_NOEXCEPT {
+    FMT_ASSERT(beta_minus_1 >= 1, "");
+    FMT_ASSERT(beta_minus_1 < 64, "");
+
+    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);
+  }
 
-  digits::result on_start(uint64_t, uint64_t, uint64_t, int&) {
-    return digits::more;
+  static carrier_uint compute_round_up_for_shorter_interval_case(
+      const cache_entry_type& cache, int beta_minus_1) FMT_NOEXCEPT {
+    return ((cache.high() >>
+             (64 - float_info<double>::significand_bits - 2 - beta_minus_1)) +
+            1) /
+           2;
   }
+};
+
+// Various integer checks
+template <class T>
+bool is_left_endpoint_integer_shorter_interval(int exponent) FMT_NOEXCEPT {
+  return exponent >=
+             float_info<
+                 T>::case_shorter_interval_left_endpoint_lower_threshold &&
+         exponent <=
+             float_info<T>::case_shorter_interval_left_endpoint_upper_threshold;
+}
+template <class T>
+bool is_endpoint_integer(typename float_info<T>::carrier_uint two_f,
+                         int exponent, int minus_k) FMT_NOEXCEPT {
+  if (exponent < float_info<T>::case_fc_pm_half_lower_threshold) return false;
+  // For k >= 0.
+  if (exponent <= float_info<T>::case_fc_pm_half_upper_threshold) return true;
+  // For k < 0.
+  if (exponent > float_info<T>::divisibility_check_by_5_threshold) return false;
+  return divisible_by_power_of_5(two_f, minus_k);
+}
+
+template <class T>
+bool is_center_integer(typename float_info<T>::carrier_uint two_f, int exponent,
+                       int minus_k) FMT_NOEXCEPT {
+  // Exponent for 5 is negative.
+  if (exponent > float_info<T>::divisibility_check_by_5_threshold) return false;
+  if (exponent > float_info<T>::case_fc_upper_threshold)
+    return divisible_by_power_of_5(two_f, minus_k);
+  // Both exponents are nonnegative.
+  if (exponent >= float_info<T>::case_fc_lower_threshold) return true;
+  // Exponent for 2 is negative.
+  return divisible_by_power_of_2(two_f, minus_k - exponent + 1);
+}
+
+// Remove trailing zeros from n and return the number of zeros removed (float)
+FMT_ALWAYS_INLINE int remove_trailing_zeros(uint32_t& n) FMT_NOEXCEPT {
+#ifdef FMT_BUILTIN_CTZ
+  int t = FMT_BUILTIN_CTZ(n);
+#else
+  int t = ctz(n);
+#endif
+  if (t > float_info<float>::max_trailing_zeros)
+    t = float_info<float>::max_trailing_zeros;
+
+  const uint32_t mod_inv1 = 0xcccccccd;
+  const uint32_t max_quotient1 = 0x33333333;
+  const uint32_t mod_inv2 = 0xc28f5c29;
+  const uint32_t max_quotient2 = 0x0a3d70a3;
+
+  int s = 0;
+  for (; s < t - 1; s += 2) {
+    if (n * mod_inv2 > max_quotient2) break;
+    n *= mod_inv2;
+  }
+  if (s < t && n * mod_inv1 <= max_quotient1) {
+    n *= mod_inv1;
+    ++s;
+  }
+  n >>= s;
+  return s;
+}
 
-  // 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;
+// Removes trailing zeros and returns the number of zeros removed (double)
+FMT_ALWAYS_INLINE int remove_trailing_zeros(uint64_t& n) FMT_NOEXCEPT {
+#ifdef FMT_BUILTIN_CTZLL
+  int t = FMT_BUILTIN_CTZLL(n);
+#else
+  int t = ctzll(n);
+#endif
+  if (t > float_info<double>::max_trailing_zeros)
+    t = float_info<double>::max_trailing_zeros;
+  // Divide by 10^8 and reduce to 32-bits
+  // Since ret_value.significand <= (2^64 - 1) / 1000 < 10^17,
+  // both of the quotient and the r should fit in 32-bits
+
+  const uint32_t mod_inv1 = 0xcccccccd;
+  const uint32_t max_quotient1 = 0x33333333;
+  const uint64_t mod_inv8 = 0xc767074b22e90e21;
+  const uint64_t max_quotient8 = 0x00002af31dc46118;
+
+  // If the number is divisible by 1'0000'0000, work with the quotient
+  if (t >= 8) {
+    auto quotient_candidate = n * mod_inv8;
+
+    if (quotient_candidate <= max_quotient8) {
+      auto quotient = static_cast<uint32_t>(quotient_candidate >> 8);
+
+      int s = 8;
+      for (; s < t; ++s) {
+        if (quotient * mod_inv1 > max_quotient1) break;
+        quotient *= mod_inv1;
+      }
+      quotient >>= (s - 8);
+      n = quotient;
+      return s;
     }
   }
 
-  // Implements Grisu's round_weed.
-  digits::result on_digit(char digit, uint64_t divisor, uint64_t remainder,
-                          uint64_t error, int exp, bool integral) {
-    buf[size++] = digit;
-    if (remainder >= error) return digits::more;
-    uint64_t unit = integral ? 1 : data::powers_of_10_64[-exp];
-    uint64_t up = (diff - 1) * unit;  // wp_Wup
-    round(up, divisor, remainder, error);
-    uint64_t down = (diff + 1) * unit;  // wp_Wdown
-    if (remainder < down && error - remainder >= divisor &&
-        (remainder + divisor < down ||
-         down - remainder > remainder + divisor - down)) {
-      return digits::error;
+  // Otherwise, work with the remainder
+  auto quotient = static_cast<uint32_t>(n / 100000000);
+  auto remainder = static_cast<uint32_t>(n - 100000000 * quotient);
+
+  if (t == 0 || remainder * mod_inv1 > max_quotient1) {
+    return 0;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 1 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 1) + quotient * 10000000ull;
+    return 1;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 2 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 2) + quotient * 1000000ull;
+    return 2;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 3 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 3) + quotient * 100000ull;
+    return 3;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 4 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 4) + quotient * 10000ull;
+    return 4;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 5 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 5) + quotient * 1000ull;
+    return 5;
+  }
+  remainder *= mod_inv1;
+
+  if (t == 6 || remainder * mod_inv1 > max_quotient1) {
+    n = (remainder >> 6) + quotient * 100ull;
+    return 6;
+  }
+  remainder *= mod_inv1;
+
+  n = (remainder >> 7) + quotient * 10ull;
+  return 7;
+}
+
+// The main algorithm for shorter interval case
+template <class T>
+FMT_ALWAYS_INLINE FMT_SAFEBUFFERS decimal_fp<T> shorter_interval_case(
+    int exponent) FMT_NOEXCEPT {
+  decimal_fp<T> ret_value;
+  // Compute k and beta
+  const int minus_k = floor_log10_pow2_minus_log10_4_over_3(exponent);
+  const int beta_minus_1 = exponent + floor_log2_pow10(-minus_k);
+
+  // Compute xi and zi
+  using cache_entry_type = typename cache_accessor<T>::cache_entry_type;
+  const cache_entry_type cache = cache_accessor<T>::get_cached_power(-minus_k);
+
+  auto xi = cache_accessor<T>::compute_left_endpoint_for_shorter_interval_case(
+      cache, beta_minus_1);
+  auto zi = cache_accessor<T>::compute_right_endpoint_for_shorter_interval_case(
+      cache, beta_minus_1);
+
+  // If the left endpoint is not an integer, increase it
+  if (!is_left_endpoint_integer_shorter_interval<T>(exponent)) ++xi;
+
+  // Try bigger divisor
+  ret_value.significand = zi / 10;
+
+  // If succeed, remove trailing zeros if necessary and return
+  if (ret_value.significand * 10 >= xi) {
+    ret_value.exponent = minus_k + 1;
+    ret_value.exponent += remove_trailing_zeros(ret_value.significand);
+    return ret_value;
+  }
+
+  // Otherwise, compute the round-up of y
+  ret_value.significand =
+      cache_accessor<T>::compute_round_up_for_shorter_interval_case(
+          cache, beta_minus_1);
+  ret_value.exponent = minus_k;
+
+  // When tie occurs, choose one of them according to the rule
+  if (exponent >= float_info<T>::shorter_interval_tie_lower_threshold &&
+      exponent <= float_info<T>::shorter_interval_tie_upper_threshold) {
+    ret_value.significand = ret_value.significand % 2 == 0
+                                ? ret_value.significand
+                                : ret_value.significand - 1;
+  } else if (ret_value.significand < xi) {
+    ++ret_value.significand;
+  }
+  return ret_value;
+}
+
+template <typename T>
+FMT_SAFEBUFFERS decimal_fp<T> to_decimal(T x) FMT_NOEXCEPT {
+  // Step 1: integer promotion & Schubfach multiplier calculation.
+
+  using carrier_uint = typename float_info<T>::carrier_uint;
+  using cache_entry_type = typename cache_accessor<T>::cache_entry_type;
+  auto br = bit_cast<carrier_uint>(x);
+
+  // Extract significand bits and exponent bits.
+  const carrier_uint significand_mask =
+      (static_cast<carrier_uint>(1) << float_info<T>::significand_bits) - 1;
+  carrier_uint significand = (br & significand_mask);
+  int exponent = static_cast<int>((br & exponent_mask<T>()) >>
+                                  float_info<T>::significand_bits);
+
+  if (exponent != 0) {  // Check if normal.
+    exponent += float_info<T>::exponent_bias - float_info<T>::significand_bits;
+
+    // Shorter interval case; proceed like Schubfach.
+    if (significand == 0) return shorter_interval_case<T>(exponent);
+
+    significand |=
+        (static_cast<carrier_uint>(1) << float_info<T>::significand_bits);
+  } else {
+    // Subnormal case; the interval is always regular.
+    if (significand == 0) return {0, 0};
+    exponent = float_info<T>::min_exponent - float_info<T>::significand_bits;
+  }
+
+  const bool include_left_endpoint = (significand % 2 == 0);
+  const bool include_right_endpoint = include_left_endpoint;
+
+  // Compute k and beta.
+  const int minus_k = floor_log10_pow2(exponent) - float_info<T>::kappa;
+  const cache_entry_type cache = cache_accessor<T>::get_cached_power(-minus_k);
+  const int beta_minus_1 = exponent + floor_log2_pow10(-minus_k);
+
+  // Compute zi and deltai
+  // 10^kappa <= deltai < 10^(kappa + 1)
+  const uint32_t deltai = cache_accessor<T>::compute_delta(cache, beta_minus_1);
+  const carrier_uint two_fc = significand << 1;
+  const carrier_uint two_fr = two_fc | 1;
+  const carrier_uint zi =
+      cache_accessor<T>::compute_mul(two_fr << beta_minus_1, cache);
+
+  // Step 2: Try larger divisor; remove trailing zeros if necessary
+
+  // Using an upper bound on zi, we might be able to optimize the division
+  // better than the compiler; we are computing zi / big_divisor here
+  decimal_fp<T> ret_value;
+  ret_value.significand = divide_by_10_to_kappa_plus_1(zi);
+  uint32_t r = static_cast<uint32_t>(zi - float_info<T>::big_divisor *
+                                              ret_value.significand);
+
+  if (r > deltai) {
+    goto small_divisor_case_label;
+  } else if (r < deltai) {
+    // Exclude the right endpoint if necessary
+    if (r == 0 && !include_right_endpoint &&
+        is_endpoint_integer<T>(two_fr, exponent, minus_k)) {
+      --ret_value.significand;
+      r = float_info<T>::big_divisor;
+      goto small_divisor_case_label;
+    }
+  } else {
+    // r == deltai; compare fractional parts
+    // Check conditions in the order different from the paper
+    // to take advantage of short-circuiting
+    const carrier_uint two_fl = two_fc - 1;
+    if ((!include_left_endpoint ||
+         !is_endpoint_integer<T>(two_fl, exponent, minus_k)) &&
+        !cache_accessor<T>::compute_mul_parity(two_fl, cache, beta_minus_1)) {
+      goto small_divisor_case_label;
     }
-    return 2 * unit <= remainder && remainder <= error - 4 * unit
-               ? digits::done
-               : digits::error;
   }
-};
+  ret_value.exponent = minus_k + float_info<T>::kappa + 1;
+
+  // We may need to remove trailing zeros
+  ret_value.exponent += remove_trailing_zeros(ret_value.significand);
+  return ret_value;
+
+  // Step 3: Find the significand with the smaller divisor
+
+small_divisor_case_label:
+  ret_value.significand *= 10;
+  ret_value.exponent = minus_k + float_info<T>::kappa;
+
+  const uint32_t mask = (1u << float_info<T>::kappa) - 1;
+  auto dist = r - (deltai / 2) + (float_info<T>::small_divisor / 2);
+
+  // Is dist divisible by 2^kappa?
+  if ((dist & mask) == 0) {
+    const bool approx_y_parity =
+        ((dist ^ (float_info<T>::small_divisor / 2)) & 1) != 0;
+    dist >>= float_info<T>::kappa;
+
+    // Is dist divisible by 5^kappa?
+    if (check_divisibility_and_divide_by_pow5<float_info<T>::kappa>(dist)) {
+      ret_value.significand += dist;
+
+      // Check z^(f) >= epsilon^(f)
+      // We have either yi == zi - epsiloni or yi == (zi - epsiloni) - 1,
+      // where yi == zi - epsiloni if and only if z^(f) >= epsilon^(f)
+      // Since there are only 2 possibilities, we only need to care about the
+      // parity. Also, zi and r should have the same parity since the divisor
+      // is an even number
+      if (cache_accessor<T>::compute_mul_parity(two_fc, cache, beta_minus_1) !=
+          approx_y_parity) {
+        --ret_value.significand;
+      } else {
+        // If z^(f) >= epsilon^(f), we might have a tie
+        // when z^(f) == epsilon^(f), or equivalently, when y is an integer
+        if (is_center_integer<T>(two_fc, exponent, minus_k)) {
+          ret_value.significand = ret_value.significand % 2 == 0
+                                      ? ret_value.significand
+                                      : ret_value.significand - 1;
+        }
+      }
+    }
+    // Is dist not divisible by 5^kappa?
+    else {
+      ret_value.significand += dist;
+    }
+  }
+  // Is dist not divisible by 2^kappa?
+  else {
+    // Since we know dist is small, we might be able to optimize the division
+    // better than the compiler; we are computing dist / small_divisor here
+    ret_value.significand +=
+        small_division_by_pow10<float_info<T>::kappa>(dist);
+  }
+  return ret_value;
+}
+}  // namespace dragonbox
 
 // Formats value using a variation of the Fixed-Precision Positive
 // Floating-Point Printout ((FPP)^2) algorithm by Steele & White:
 // https://fmt.dev/p372-steele.pdf.
 template <typename Double>
-void fallback_format(Double d, buffer<char>& buf, int& exp10) {
+void fallback_format(Double d, int num_digits, bool binary32, buffer<char>& buf,
+                     int& exp10) {
   bigint numerator;    // 2 * R in (FPP)^2.
   bigint denominator;  // 2 * S in (FPP)^2.
   // lower and upper are differences between value and corresponding boundaries.
@@ -998,8 +2314,9 @@ void fallback_format(Double d, buffer<char>& buf, int& exp10) {
   // Shift numerator and denominator by an extra bit or two (if lower boundary
   // is closer) to make lower and upper integers. This eliminates multiplication
   // by 2 during later computations.
-  // TODO: handle float
-  int shift = value.assign(d) ? 2 : 1;
+  const bool is_predecessor_closer =
+      binary32 ? value.assign(static_cast<float>(d)) : value.assign(d);
+  int shift = is_predecessor_closer ? 2 : 1;
   uint64_t significand = value.f << shift;
   if (value.e >= 0) {
     numerator.assign(significand);
@@ -1034,39 +2351,73 @@ void fallback_format(Double d, buffer<char>& buf, int& exp10) {
       upper = &upper_store;
     }
   }
-  if (!upper) upper = &lower;
   // Invariant: value == (numerator / denominator) * pow(10, exp10).
-  bool even = (value.f & 1) == 0;
-  int num_digits = 0;
-  char* data = buf.data();
-  for (;;) {
-    int digit = numerator.divmod_assign(denominator);
-    bool low = compare(numerator, lower) - even < 0;  // numerator <[=] lower.
-    // numerator + upper >[=] pow10:
-    bool high = add_compare(numerator, *upper, denominator) + even > 0;
-    data[num_digits++] = static_cast<char>('0' + digit);
-    if (low || high) {
-      if (!low) {
-        ++data[num_digits - 1];
-      } else if (high) {
-        int result = add_compare(numerator, numerator, denominator);
-        // Round half to even.
-        if (result > 0 || (result == 0 && (digit % 2) != 0))
+  if (num_digits < 0) {
+    // Generate the shortest representation.
+    if (!upper) upper = &lower;
+    bool even = (value.f & 1) == 0;
+    num_digits = 0;
+    char* data = buf.data();
+    for (;;) {
+      int digit = numerator.divmod_assign(denominator);
+      bool low = compare(numerator, lower) - even < 0;  // numerator <[=] lower.
+      // numerator + upper >[=] pow10:
+      bool high = add_compare(numerator, *upper, denominator) + even > 0;
+      data[num_digits++] = static_cast<char>('0' + digit);
+      if (low || high) {
+        if (!low) {
           ++data[num_digits - 1];
+        } else if (high) {
+          int result = add_compare(numerator, numerator, denominator);
+          // Round half to even.
+          if (result > 0 || (result == 0 && (digit % 2) != 0))
+            ++data[num_digits - 1];
+        }
+        buf.try_resize(to_unsigned(num_digits));
+        exp10 -= num_digits - 1;
+        return;
       }
-      buf.resize(to_unsigned(num_digits));
-      exp10 -= num_digits - 1;
-      return;
+      numerator *= 10;
+      lower *= 10;
+      if (upper != &lower) *upper *= 10;
     }
+  }
+  // Generate the given number of digits.
+  exp10 -= num_digits - 1;
+  if (num_digits == 0) {
+    buf.try_resize(1);
+    denominator *= 10;
+    buf[0] = add_compare(numerator, numerator, denominator) > 0 ? '1' : '0';
+    return;
+  }
+  buf.try_resize(to_unsigned(num_digits));
+  for (int i = 0; i < num_digits - 1; ++i) {
+    int digit = numerator.divmod_assign(denominator);
+    buf[i] = static_cast<char>('0' + digit);
     numerator *= 10;
-    lower *= 10;
-    if (upper != &lower) *upper *= 10;
   }
+  int digit = numerator.divmod_assign(denominator);
+  auto result = add_compare(numerator, numerator, denominator);
+  if (result > 0 || (result == 0 && (digit % 2) != 0)) {
+    if (digit == 9) {
+      const auto overflow = '0' + 10;
+      buf[num_digits - 1] = overflow;
+      // Propagate the carry.
+      for (int i = num_digits - 1; i > 0 && buf[i] == overflow; --i) {
+        buf[i] = '0';
+        ++buf[i - 1];
+      }
+      if (buf[0] == overflow) {
+        buf[0] = '1';
+        ++exp10;
+      }
+      return;
+    }
+    ++digit;
+  }
+  buf[num_digits - 1] = static_cast<char>('0' + digit);
 }
 
-// Formats value using the Grisu algorithm
-// (https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf)
-// if T is a IEEE754 binary32 or binary64 and snprintf otherwise.
 template <typename T>
 int format_float(T value, int precision, float_specs specs, buffer<char>& buf) {
   static_assert(!std::is_same<T, float>::value, "");
@@ -1078,66 +2429,57 @@ int format_float(T value, int precision, float_specs specs, buffer<char>& buf) {
       buf.push_back('0');
       return 0;
     }
-    buf.resize(to_unsigned(precision));
+    buf.try_resize(to_unsigned(precision));
     std::uninitialized_fill_n(buf.data(), precision, '0');
     return -precision;
   }
 
   if (!specs.use_grisu) return snprintf_float(value, precision, specs, buf);
 
+  if (precision < 0) {
+    // Use Dragonbox for the shortest format.
+    if (specs.binary32) {
+      auto dec = dragonbox::to_decimal(static_cast<float>(value));
+      write<char>(buffer_appender<char>(buf), dec.significand);
+      return dec.exponent;
+    }
+    auto dec = dragonbox::to_decimal(static_cast<double>(value));
+    write<char>(buffer_appender<char>(buf), dec.significand);
+    return dec.exponent;
+  }
+
+  // Use Grisu + Dragon4 for the given precision:
+  // https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf.
   int exp = 0;
   const int min_exp = -60;  // alpha in Grisu.
   int cached_exp10 = 0;     // K in Grisu.
-  if (precision < 0) {
-    fp fp_value;
-    auto boundaries = specs.binary32
-                          ? fp_value.assign_float_with_boundaries(value)
-                          : fp_value.assign_with_boundaries(value);
-    fp_value = normalize(fp_value);
-    // Find a cached power of 10 such that multiplying value by it will bring
-    // the exponent in the range [min_exp, -32].
-    const fp cached_pow = get_cached_power(
-        min_exp - (fp_value.e + fp::significand_size), cached_exp10);
-    // Multiply value and boundaries by the cached power of 10.
-    fp_value = fp_value * cached_pow;
-    boundaries.lower = multiply(boundaries.lower, cached_pow.f);
-    boundaries.upper = multiply(boundaries.upper, cached_pow.f);
-    assert(min_exp <= fp_value.e && fp_value.e <= -32);
-    --boundaries.lower;  // \tilde{M}^- - 1 ulp -> M^-_{\downarrow}.
-    ++boundaries.upper;  // \tilde{M}^+ + 1 ulp -> M^+_{\uparrow}.
-    // Numbers outside of (lower, upper) definitely do not round to value.
-    grisu_shortest_handler handler{buf.data(), 0,
-                                   boundaries.upper - fp_value.f};
-    auto result =
-        grisu_gen_digits(fp(boundaries.upper, fp_value.e),
-                         boundaries.upper - boundaries.lower, exp, handler);
-    if (result == digits::error) {
-      exp += handler.size - cached_exp10 - 1;
-      fallback_format(value, buf, exp);
-      return exp;
-    }
-    buf.resize(to_unsigned(handler.size));
+  fp normalized = normalize(fp(value));
+  const auto cached_pow = get_cached_power(
+      min_exp - (normalized.e + fp::significand_size), cached_exp10);
+  normalized = normalized * cached_pow;
+  // Limit precision to the maximum possible number of significant digits in an
+  // IEEE754 double because we don't need to generate zeros.
+  const int max_double_digits = 767;
+  if (precision > max_double_digits) precision = max_double_digits;
+  fixed_handler handler{buf.data(), 0, precision, -cached_exp10, fixed};
+  if (grisu_gen_digits(normalized, 1, exp, handler) == digits::error) {
+    exp += handler.size - cached_exp10 - 1;
+    fallback_format(value, handler.precision, specs.binary32, buf, exp);
   } else {
-    if (precision > 17) return snprintf_float(value, precision, specs, buf);
-    fp normalized = normalize(fp(value));
-    const auto cached_pow = get_cached_power(
-        min_exp - (normalized.e + fp::significand_size), cached_exp10);
-    normalized = normalized * cached_pow;
-    fixed_handler handler{buf.data(), 0, precision, -cached_exp10, fixed};
-    if (grisu_gen_digits(normalized, 1, exp, handler) == digits::error)
-      return snprintf_float(value, precision, specs, buf);
-    int num_digits = handler.size;
-    if (!fixed) {
-      // Remove trailing zeros.
-      while (num_digits > 0 && buf[num_digits - 1] == '0') {
-        --num_digits;
-        ++exp;
-      }
+    exp += handler.exp10;
+    buf.try_resize(to_unsigned(handler.size));
+  }
+  if (!fixed && !specs.showpoint) {
+    // Remove trailing zeros.
+    auto num_digits = buf.size();
+    while (num_digits > 0 && buf[num_digits - 1] == '0') {
+      --num_digits;
+      ++exp;
     }
-    buf.resize(to_unsigned(num_digits));
+    buf.try_resize(num_digits);
   }
-  return exp - cached_exp10;
-}
+  return exp;
+}  // namespace detail
 
 template <typename T>
 int snprintf_float(T value, int precision, float_specs specs,
@@ -1185,19 +2527,20 @@ int snprintf_float(T value, int precision, float_specs specs,
                      ? snprintf_ptr(begin, capacity, format, precision, value)
                      : snprintf_ptr(begin, capacity, format, value);
     if (result < 0) {
-      buf.reserve(buf.capacity() + 1);  // The buffer will grow exponentially.
+      // The buffer will grow exponentially.
+      buf.try_reserve(buf.capacity() + 1);
       continue;
     }
     auto size = to_unsigned(result);
     // Size equal to capacity means that the last character was truncated.
     if (size >= capacity) {
-      buf.reserve(size + offset + 1);  // Add 1 for the terminating '\0'.
+      buf.try_reserve(size + offset + 1);  // Add 1 for the terminating '\0'.
       continue;
     }
     auto is_digit = [](char c) { return c >= '0' && c <= '9'; };
     if (specs.format == float_format::fixed) {
       if (precision == 0) {
-        buf.resize(size);
+        buf.try_resize(size);
         return 0;
       }
       // Find and remove the decimal point.
@@ -1207,11 +2550,11 @@ int snprintf_float(T value, int precision, float_specs specs,
       } while (is_digit(*p));
       int fraction_size = static_cast<int>(end - p - 1);
       std::memmove(p, p + 1, to_unsigned(fraction_size));
-      buf.resize(size - 1);
+      buf.try_resize(size - 1);
       return -fraction_size;
     }
     if (specs.format == float_format::hex) {
-      buf.resize(size + offset);
+      buf.try_resize(size + offset);
       return 0;
     }
     // Find and parse the exponent.
@@ -1237,7 +2580,7 @@ int snprintf_float(T value, int precision, float_specs specs,
       fraction_size = static_cast<int>(fraction_end - begin - 1);
       std::memmove(begin + 1, begin + 2, to_unsigned(fraction_size));
     }
-    buf.resize(to_unsigned(fraction_size) + offset + 1);
+    buf.try_resize(to_unsigned(fraction_size) + offset + 1);
     return exp - fraction_size;
   }
 }
@@ -1259,25 +2602,18 @@ int snprintf_float(T value, int precision, float_specs specs,
  * occurs, this pointer will be a guess that depends on the particular
  * error, but it will always advance at least one byte.
  */
-FMT_FUNC const char* utf8_decode(const char* buf, uint32_t* c, int* e) {
-  static const char lengths[] = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
-                                 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0,
-                                 0, 0, 2, 2, 2, 2, 3, 3, 4, 0};
+inline const char* utf8_decode(const char* buf, uint32_t* c, int* e) {
   static const int masks[] = {0x00, 0x7f, 0x1f, 0x0f, 0x07};
   static const uint32_t mins[] = {4194304, 0, 128, 2048, 65536};
   static const int shiftc[] = {0, 18, 12, 6, 0};
   static const int shifte[] = {0, 6, 4, 2, 0};
 
-  auto s = reinterpret_cast<const unsigned char*>(buf);
-  int len = lengths[s[0] >> 3];
-
-  // Compute the pointer to the next character early so that the next
-  // iteration can start working on the next character. Neither Clang
-  // nor GCC figure out this reordering on their own.
-  const char* next = buf + len + !len;
+  int len = code_point_length(buf);
+  const char* next = buf + len;
 
   // Assume a four-byte character and load four bytes. Unused bits are
   // shifted out.
+  auto s = reinterpret_cast<const unsigned char*>(buf);
   *c = uint32_t(s[0] & masks[len]) << 18;
   *c |= uint32_t(s[1] & 0x3f) << 12;
   *c |= uint32_t(s[2] & 0x3f) << 6;
@@ -1296,6 +2632,19 @@ FMT_FUNC const char* utf8_decode(const char* buf, uint32_t* c, int* e) {
 
   return next;
 }
+
+struct stringifier {
+  template <typename T> FMT_INLINE std::string operator()(T value) const {
+    return to_string(value);
+  }
+  std::string operator()(basic_format_arg<format_context>::handle h) const {
+    memory_buffer buf;
+    format_parse_context parse_ctx({});
+    format_context format_ctx(buffer_appender<char>(buf), {}, {});
+    h.format(parse_ctx, format_ctx);
+    return to_string(buf);
+  }
+};
 }  // namespace detail
 
 template <> struct formatter<detail::bigint> {
@@ -1363,7 +2712,8 @@ FMT_FUNC void format_system_error(detail::buffer<char>& out, int error_code,
       int result =
           detail::safe_strerror(error_code, system_message, buf.size());
       if (result == 0) {
-        format_to(std::back_inserter(out), "{}: {}", message, system_message);
+        format_to(detail::buffer_appender<char>(out), "{}: {}", message,
+                  system_message);
         return;
       }
       if (result != ERANGE)
@@ -1384,20 +2734,6 @@ FMT_FUNC void report_system_error(int error_code,
   report_error(format_system_error, error_code, message);
 }
 
-struct stringifier {
-  template <typename T> FMT_INLINE std::string operator()(T value) const {
-    return to_string(value);
-  }
-  std::string operator()(basic_format_arg<format_context>::handle h) const {
-    memory_buffer buf;
-    detail::buffer<char>& base = buf;
-    format_parse_context parse_ctx({});
-    format_context format_ctx(std::back_inserter(base), {}, {});
-    h.format(parse_ctx, format_ctx);
-    return to_string(buf);
-  }
-};
-
 FMT_FUNC std::string detail::vformat(string_view format_str, format_args args) {
   if (format_str.size() == 2 && equal2(format_str.data(), "{}")) {
     auto arg = args.get(0);
@@ -1409,6 +2745,14 @@ FMT_FUNC std::string detail::vformat(string_view format_str, format_args args) {
   return to_string(buffer);
 }
 
+#ifdef _WIN32
+namespace detail {
+using dword = conditional_t<sizeof(long) == 4, unsigned long, unsigned>;
+extern "C" __declspec(dllimport) int __stdcall WriteConsoleW(  //
+    void*, const void*, dword, dword*, void*);
+}  // namespace detail
+#endif
+
 FMT_FUNC void vprint(std::FILE* f, string_view format_str, format_args args) {
   memory_buffer buffer;
   detail::vformat_to(buffer, format_str,
@@ -1417,10 +2761,10 @@ FMT_FUNC void vprint(std::FILE* f, string_view format_str, format_args args) {
   auto fd = _fileno(f);
   if (_isatty(fd)) {
     detail::utf8_to_utf16 u16(string_view(buffer.data(), buffer.size()));
-    auto written = DWORD();
-    if (!WriteConsoleW(reinterpret_cast<HANDLE>(_get_osfhandle(fd)),
-                       u16.c_str(), static_cast<DWORD>(u16.size()), &written,
-                       nullptr)) {
+    auto written = detail::dword();
+    if (!detail::WriteConsoleW(reinterpret_cast<void*>(_get_osfhandle(fd)),
+                               u16.c_str(), static_cast<uint32_t>(u16.size()),
+                               &written, nullptr)) {
       FMT_THROW(format_error("failed to write to console"));
     }
     return;
@@ -1446,8 +2790,4 @@ FMT_FUNC void vprint(string_view format_str, format_args args) {
 
 FMT_END_NAMESPACE
 
-#ifdef _MSC_VER
-#  pragma warning(pop)
-#endif
-
 #endif  // FMT_FORMAT_INL_H_
diff --git a/include/fmt/format.h b/include/fmt/format.h
index ddf086f8..fbe50450 100644
--- a/include/fmt/format.h
+++ b/include/fmt/format.h
@@ -70,9 +70,11 @@
 #endif
 
 #if __cplusplus == 201103L || __cplusplus == 201402L
-#  if defined(__clang__)
+#  if defined(__INTEL_COMPILER) || defined(__PGI)
+#    define FMT_FALLTHROUGH
+#  elif defined(__clang__)
 #    define FMT_FALLTHROUGH [[clang::fallthrough]]
-#  elif FMT_GCC_VERSION >= 700 && !defined(__PGI) && \
+#  elif FMT_GCC_VERSION >= 700 && \
       (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 520)
 #    define FMT_FALLTHROUGH [[gnu::fallthrough]]
 #  else
@@ -139,12 +141,13 @@ FMT_END_NAMESPACE
 #endif
 
 #ifndef FMT_USE_UDL_TEMPLATE
-// EDG frontend based compilers (icc, nvcc, etc) and GCC < 6.4 do not properly
-// support UDL templates and GCC >= 9 warns about them.
+// EDG frontend based compilers (icc, nvcc, PGI, etc) and GCC < 6.4 do not
+// properly support UDL templates and GCC >= 9 warns about them.
 #  if FMT_USE_USER_DEFINED_LITERALS &&                         \
       (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 501) && \
       ((FMT_GCC_VERSION >= 604 && __cplusplus >= 201402L) ||   \
-       FMT_CLANG_VERSION >= 304)
+       FMT_CLANG_VERSION >= 304) &&                            \
+      !defined(__PGI) && !defined(__NVCC__)
 #    define FMT_USE_UDL_TEMPLATE 1
 #  else
 #    define FMT_USE_UDL_TEMPLATE 0
@@ -163,6 +166,14 @@ FMT_END_NAMESPACE
 #  define FMT_USE_LONG_DOUBLE 1
 #endif
 
+// Defining FMT_REDUCE_INT_INSTANTIATIONS to 1, will reduce the number of
+// int_writer template instances to just one by only using the largest integer
+// type. This results in a reduction in binary size but will cause a decrease in
+// integer formatting performance.
+#if !defined(FMT_REDUCE_INT_INSTANTIATIONS)
+#  define FMT_REDUCE_INT_INSTANTIATIONS 0
+#endif
+
 // __builtin_clz is broken in clang with Microsoft CodeGen:
 // https://github.com/fmtlib/fmt/issues/519
 #if (FMT_GCC_VERSION || FMT_HAS_BUILTIN(__builtin_clz)) && !FMT_MSC_VER
@@ -171,56 +182,87 @@ FMT_END_NAMESPACE
 #if (FMT_GCC_VERSION || FMT_HAS_BUILTIN(__builtin_clzll)) && !FMT_MSC_VER
 #  define FMT_BUILTIN_CLZLL(n) __builtin_clzll(n)
 #endif
+#if (FMT_GCC_VERSION || FMT_HAS_BUILTIN(__builtin_ctz))
+#  define FMT_BUILTIN_CTZ(n) __builtin_ctz(n)
+#endif
+#if (FMT_GCC_VERSION || FMT_HAS_BUILTIN(__builtin_ctzll))
+#  define FMT_BUILTIN_CTZLL(n) __builtin_ctzll(n)
+#endif
+
+#if FMT_MSC_VER
+#  include <intrin.h>  // _BitScanReverse[64], _BitScanForward[64], _umul128
+#endif
 
 // Some compilers masquerade as both MSVC and GCC-likes or otherwise support
 // __builtin_clz and __builtin_clzll, so only define FMT_BUILTIN_CLZ using the
 // MSVC intrinsics if the clz and clzll builtins are not available.
-#if FMT_MSC_VER && !defined(FMT_BUILTIN_CLZLL) && !defined(_MANAGED)
-#  include <intrin.h>  // _BitScanReverse, _BitScanReverse64
-
+#if FMT_MSC_VER && !defined(FMT_BUILTIN_CLZLL) && \
+    !defined(FMT_BUILTIN_CTZLL) && !defined(_MANAGED)
 FMT_BEGIN_NAMESPACE
 namespace detail {
 // Avoid Clang with Microsoft CodeGen's -Wunknown-pragmas warning.
 #  ifndef __clang__
+#    pragma intrinsic(_BitScanForward)
 #    pragma intrinsic(_BitScanReverse)
 #  endif
-inline uint32_t clz(uint32_t x) {
+#  if defined(_WIN64) && !defined(__clang__)
+#    pragma intrinsic(_BitScanForward64)
+#    pragma intrinsic(_BitScanReverse64)
+#  endif
+
+inline int clz(uint32_t x) {
   unsigned long r = 0;
   _BitScanReverse(&r, x);
-
   FMT_ASSERT(x != 0, "");
   // Static analysis complains about using uninitialized data
   // "r", but the only way that can happen is if "x" is 0,
   // which the callers guarantee to not happen.
   FMT_SUPPRESS_MSC_WARNING(6102)
-  return 31 - r;
+  return 31 ^ static_cast<int>(r);
 }
 #  define FMT_BUILTIN_CLZ(n) detail::clz(n)
 
-#  if defined(_WIN64) && !defined(__clang__)
-#    pragma intrinsic(_BitScanReverse64)
-#  endif
-
-inline uint32_t clzll(uint64_t x) {
+inline int clzll(uint64_t x) {
   unsigned long r = 0;
 #  ifdef _WIN64
   _BitScanReverse64(&r, x);
 #  else
   // Scan the high 32 bits.
-  if (_BitScanReverse(&r, static_cast<uint32_t>(x >> 32))) return 63 - (r + 32);
-
+  if (_BitScanReverse(&r, static_cast<uint32_t>(x >> 32))) return 63 ^ (r + 32);
   // Scan the low 32 bits.
   _BitScanReverse(&r, static_cast<uint32_t>(x));
 #  endif
-
   FMT_ASSERT(x != 0, "");
-  // Static analysis complains about using uninitialized data
-  // "r", but the only way that can happen is if "x" is 0,
-  // which the callers guarantee to not happen.
-  FMT_SUPPRESS_MSC_WARNING(6102)
-  return 63 - r;
+  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+  return 63 ^ static_cast<int>(r);
 }
 #  define FMT_BUILTIN_CLZLL(n) detail::clzll(n)
+
+inline int ctz(uint32_t x) {
+  unsigned long r = 0;
+  _BitScanForward(&r, x);
+  FMT_ASSERT(x != 0, "");
+  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+  return static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CTZ(n) detail::ctz(n)
+
+inline int ctzll(uint64_t x) {
+  unsigned long r = 0;
+  FMT_ASSERT(x != 0, "");
+  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+#  ifdef _WIN64
+  _BitScanForward64(&r, x);
+#  else
+  // Scan the low 32 bits.
+  if (_BitScanForward(&r, static_cast<uint32_t>(x))) return static_cast<int>(r);
+  // Scan the high 32 bits.
+  _BitScanForward(&r, static_cast<uint32_t>(x >> 32));
+  r += 32;
+#  endif
+  return static_cast<int>(r);
+}
+#  define FMT_BUILTIN_CTZLL(n) detail::ctzll(n)
 }  // namespace detail
 FMT_END_NAMESPACE
 #endif
@@ -283,6 +325,9 @@ template <typename T> constexpr T max_value() {
 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);
@@ -295,50 +340,11 @@ FMT_INLINE void assume(bool condition) {
 #endif
 }
 
-// A workaround for gcc 4.8 to make void_t work in a SFINAE context.
-template <typename... Ts> struct void_t_impl { using type = void; };
-
-template <typename... Ts>
-using void_t = typename detail::void_t_impl<Ts...>::type;
-
 // An approximation of iterator_t for pre-C++20 systems.
 template <typename T>
 using iterator_t = decltype(std::begin(std::declval<T&>()));
 template <typename T> using sentinel_t = decltype(std::end(std::declval<T&>()));
 
-// Detect the iterator category of *any* given type in a SFINAE-friendly way.
-// Unfortunately, older implementations of std::iterator_traits are not safe
-// for use in a SFINAE-context.
-template <typename It, typename Enable = void>
-struct iterator_category : std::false_type {};
-
-template <typename T> struct iterator_category<T*> {
-  using type = std::random_access_iterator_tag;
-};
-
-template <typename It>
-struct iterator_category<It, void_t<typename It::iterator_category>> {
-  using type = typename It::iterator_category;
-};
-
-// Detect if *any* given type models the OutputIterator concept.
-template <typename It> class is_output_iterator {
-  // Check for mutability because all iterator categories derived from
-  // std::input_iterator_tag *may* also meet the requirements of an
-  // OutputIterator, thereby falling into the category of 'mutable iterators'
-  // [iterator.requirements.general] clause 4. The compiler reveals this
-  // property only at the point of *actually dereferencing* the iterator!
-  template <typename U>
-  static decltype(*(std::declval<U>())) test(std::input_iterator_tag);
-  template <typename U> static char& test(std::output_iterator_tag);
-  template <typename U> static const char& test(...);
-
-  using type = decltype(test<It>(typename iterator_category<It>::type{}));
-
- public:
-  enum { value = !std::is_const<remove_reference_t<type>>::value };
-};
-
 // A workaround for std::string not having mutable data() until C++17.
 template <typename Char> inline Char* get_data(std::basic_string<Char>& s) {
   return &s[0];
@@ -371,10 +377,29 @@ reserve(std::back_insert_iterator<Container> it, size_t n) {
   return make_checked(get_data(c) + size, n);
 }
 
+template <typename T>
+inline buffer_appender<T> reserve(buffer_appender<T> it, size_t n) {
+  buffer<T>& buf = get_container(it);
+  buf.try_reserve(buf.size() + n);
+  return it;
+}
+
 template <typename Iterator> inline Iterator& reserve(Iterator& it, size_t) {
   return it;
 }
 
+template <typename T, typename OutputIt>
+constexpr T* to_pointer(OutputIt, size_t) {
+  return nullptr;
+}
+template <typename T> T* to_pointer(buffer_appender<T> it, size_t n) {
+  buffer<T>& buf = get_container(it);
+  auto size = buf.size();
+  if (buf.capacity() < size + n) return nullptr;
+  buf.try_resize(size + n);
+  return buf.data() + size;
+}
+
 template <typename Container, FMT_ENABLE_IF(is_contiguous<Container>::value)>
 inline std::back_insert_iterator<Container> base_iterator(
     std::back_insert_iterator<Container>& it,
@@ -412,13 +437,17 @@ class counting_iterator {
     ++count_;
     return *this;
   }
-
   counting_iterator operator++(int) {
     auto it = *this;
     ++*this;
     return it;
   }
 
+  friend counting_iterator operator+(counting_iterator it, difference_type n) {
+    it.count_ += static_cast<size_t>(n);
+    return it;
+  }
+
   value_type operator*() const { return {}; }
 };
 
@@ -552,23 +581,38 @@ OutputIt copy_str(InputIt begin, InputIt end, OutputIt it) {
                         [](char c) { return static_cast<char8_type>(c); });
 }
 
-#ifndef FMT_USE_GRISU
-#  define FMT_USE_GRISU 1
-#endif
-
-template <typename T> constexpr bool use_grisu() {
-  return FMT_USE_GRISU && std::numeric_limits<double>::is_iec559 &&
-         sizeof(T) <= sizeof(double);
+template <typename Char, typename InputIt>
+inline counting_iterator copy_str(InputIt begin, InputIt end,
+                                  counting_iterator it) {
+  return it + (end - begin);
 }
 
 template <typename T>
+using is_fast_float = bool_constant<std::numeric_limits<T>::is_iec559 &&
+                                    sizeof(T) <= sizeof(double)>;
+
+#ifndef FMT_USE_FULL_CACHE_DRAGONBOX
+#  define FMT_USE_FULL_CACHE_DRAGONBOX 0
+#endif
+
+template <typename T>
 template <typename U>
 void buffer<T>::append(const U* begin, const U* end) {
-  size_t new_size = size_ + to_unsigned(end - begin);
-  reserve(new_size);
-  std::uninitialized_copy(begin, end,
-                          make_checked(ptr_ + size_, capacity_ - size_));
-  size_ = new_size;
+  do {
+    auto count = to_unsigned(end - begin);
+    try_reserve(size_ + count);
+    auto free_cap = capacity_ - size_;
+    if (free_cap < count) count = free_cap;
+    std::uninitialized_copy_n(begin, count, make_checked(ptr_ + size_, count));
+    size_ += count;
+    begin += count;
+  } while (begin != end);
+}
+
+template <typename OutputIt, typename T, typename Traits>
+void iterator_buffer<OutputIt, T, Traits>::flush() {
+  out_ = std::copy_n(data_, this->limit(this->size()), out_);
+  this->clear();
 }
 }  // namespace detail
 
@@ -607,7 +651,7 @@ enum { inline_buffer_size = 500 };
  */
 template <typename T, size_t SIZE = inline_buffer_size,
           typename Allocator = std::allocator<T>>
-class basic_memory_buffer : public detail::buffer<T> {
+class basic_memory_buffer final : public detail::buffer<T> {
  private:
   T store_[SIZE];
 
@@ -621,7 +665,7 @@ class basic_memory_buffer : public detail::buffer<T> {
   }
 
  protected:
-  void grow(size_t size) FMT_OVERRIDE;
+  void grow(size_t size) final FMT_OVERRIDE;
 
  public:
   using value_type = T;
@@ -631,7 +675,7 @@ class basic_memory_buffer : public detail::buffer<T> {
       : alloc_(alloc) {
     this->set(store_, SIZE);
   }
-  ~basic_memory_buffer() FMT_OVERRIDE { deallocate(); }
+  ~basic_memory_buffer() { deallocate(); }
 
  private:
   // Move data from other to this buffer.
@@ -675,6 +719,22 @@ class basic_memory_buffer : public detail::buffer<T> {
 
   // Returns a copy of the allocator associated with this buffer.
   Allocator get_allocator() const { return alloc_; }
+
+  /**
+    Resizes the buffer to contain *count* elements. If T is a POD type new
+    elements may not be initialized.
+   */
+  void resize(size_t count) { this->try_resize(count); }
+
+  /** Increases the buffer capacity to *new_capacity*. */
+  void reserve(size_t new_capacity) { this->try_reserve(new_capacity); }
+
+  // Directly append data into the buffer
+  using detail::buffer<T>::append;
+  template <typename ContiguousRange>
+  void append(const ContiguousRange& range) {
+    append(range.data(), range.data() + range.size());
+  }
 };
 
 template <typename T, size_t SIZE, typename Allocator>
@@ -721,13 +781,18 @@ class FMT_API format_error : public std::runtime_error {
 
 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;
 }
@@ -740,19 +805,81 @@ FMT_CONSTEXPR bool is_supported_floating_point(T) {
 }
 
 // Smallest of uint32_t, uint64_t, uint128_t that is large enough to
-// represent all values of T.
+// represent all values of an integral type T.
 template <typename T>
-using uint32_or_64_or_128_t = conditional_t<
-    std::numeric_limits<T>::digits <= 32, uint32_t,
-    conditional_t<std::numeric_limits<T>::digits <= 64, uint64_t, uint128_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 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[];
@@ -766,6 +893,17 @@ template <typename T = void> struct FMT_EXTERN_TEMPLATE_API basic_data {
   static const char right_padding_shifts[5];
 };
 
+// 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
@@ -777,10 +915,9 @@ 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[t]);
 }
 #else
 // Fallback version of count_digits used when __builtin_clz is not available.
@@ -830,15 +967,24 @@ template <> int count_digits<4>(detail::fallback_uintptr n);
 
 #if FMT_GCC_VERSION || FMT_CLANG_VERSION
 #  define FMT_ALWAYS_INLINE inline __attribute__((always_inline))
+#elif FMT_MSC_VER
+#  define FMT_ALWAYS_INLINE __forceinline
+#else
+#  define FMT_ALWAYS_INLINE inline
+#endif
+
+// To suppress unnecessary security cookie checks
+#if FMT_MSC_VER && !FMT_CLANG_VERSION
+#  define FMT_SAFEBUFFERS __declspec(safebuffers)
 #else
-#  define FMT_ALWAYS_INLINE
+#  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[t]);
 }
 #endif
 
@@ -885,7 +1031,7 @@ template <typename Char> void copy2(Char* dst, const char* src) {
   *dst++ = static_cast<Char>(*src++);
   *dst = static_cast<Char>(*src);
 }
-inline void copy2(char* dst, const char* src) { memcpy(dst, src, 2); }
+FMT_INLINE void copy2(char* dst, const char* src) { memcpy(dst, src, 2); }
 
 template <typename Iterator> struct format_decimal_result {
   Iterator begin;
@@ -921,11 +1067,10 @@ inline format_decimal_result<Char*> format_decimal(Char* out, UInt value,
 template <typename Char, typename UInt, typename Iterator,
           FMT_ENABLE_IF(!std::is_pointer<remove_cvref_t<Iterator>>::value)>
 inline format_decimal_result<Iterator> format_decimal(Iterator out, UInt value,
-                                                      int num_digits) {
-  // Buffer should be large enough to hold all digits (<= digits10 + 1).
-  enum { max_size = digits10<UInt>() + 1 };
-  Char buffer[2 * max_size];
-  auto end = format_decimal(buffer, value, num_digits).end;
+                                                      int size) {
+  // Buffer is large enough to hold all digits (digits10 + 1).
+  Char buffer[digits10<UInt>() + 1];
+  auto end = format_decimal(buffer, value, size).end;
   return {out, detail::copy_str<Char>(buffer, end, out)};
 }
 
@@ -967,6 +1112,10 @@ Char* format_uint(Char* buffer, detail::fallback_uintptr n, int num_digits,
 
 template <unsigned BASE_BITS, typename Char, typename It, typename UInt>
 inline It format_uint(It out, UInt value, int num_digits, bool upper = false) {
+  if (auto ptr = to_pointer<Char>(out, to_unsigned(num_digits))) {
+    format_uint<BASE_BITS>(ptr, value, num_digits, upper);
+    return out;
+  }
   // Buffer should be large enough to hold all digits (digits / BASE_BITS + 1).
   char buffer[num_bits<UInt>() / BASE_BITS + 1];
   format_uint<BASE_BITS>(buffer, value, num_digits, upper);
@@ -1058,6 +1207,78 @@ template <typename Char> struct basic_format_specs {
 using format_specs = basic_format_specs<char>;
 
 namespace detail {
+namespace dragonbox {
+
+// Type-specific information that Dragonbox uses.
+template <class T> struct float_info;
+
+template <> struct float_info<float> {
+  using carrier_uint = uint32_t;
+  static const int significand_bits = 23;
+  static const int exponent_bits = 8;
+  static const int min_exponent = -126;
+  static const int max_exponent = 127;
+  static const int exponent_bias = -127;
+  static const int decimal_digits = 9;
+  static const int kappa = 1;
+  static const int big_divisor = 100;
+  static const int small_divisor = 10;
+  static const int min_k = -31;
+  static const int max_k = 46;
+  static const int cache_bits = 64;
+  static const int divisibility_check_by_5_threshold = 39;
+  static const int case_fc_pm_half_lower_threshold = -1;
+  static const int case_fc_pm_half_upper_threshold = 6;
+  static const int case_fc_lower_threshold = -2;
+  static const int case_fc_upper_threshold = 6;
+  static const int case_shorter_interval_left_endpoint_lower_threshold = 2;
+  static const int case_shorter_interval_left_endpoint_upper_threshold = 3;
+  static const int shorter_interval_tie_lower_threshold = -35;
+  static const int shorter_interval_tie_upper_threshold = -35;
+  static const int max_trailing_zeros = 7;
+};
+
+template <> struct float_info<double> {
+  using carrier_uint = uint64_t;
+  static const int significand_bits = 52;
+  static const int exponent_bits = 11;
+  static const int min_exponent = -1022;
+  static const int max_exponent = 1023;
+  static const int exponent_bias = -1023;
+  static const int decimal_digits = 17;
+  static const int kappa = 2;
+  static const int big_divisor = 1000;
+  static const int small_divisor = 100;
+  static const int min_k = -292;
+  static const int max_k = 326;
+  static const int cache_bits = 128;
+  static const int divisibility_check_by_5_threshold = 86;
+  static const int case_fc_pm_half_lower_threshold = -2;
+  static const int case_fc_pm_half_upper_threshold = 9;
+  static const int case_fc_lower_threshold = -4;
+  static const int case_fc_upper_threshold = 9;
+  static const int case_shorter_interval_left_endpoint_lower_threshold = 2;
+  static const int case_shorter_interval_left_endpoint_upper_threshold = 3;
+  static const int shorter_interval_tie_lower_threshold = -77;
+  static const int shorter_interval_tie_upper_threshold = -77;
+  static const int max_trailing_zeros = 16;
+};
+
+template <typename T> struct decimal_fp {
+  using significand_type = typename float_info<T>::carrier_uint;
+  significand_type significand;
+  int exponent;
+};
+
+template <typename T> 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 {
@@ -1099,113 +1320,6 @@ template <typename Char, typename It> It write_exponent(int exp, It it) {
   return it;
 }
 
-template <typename Char> class float_writer {
- private:
-  // The number is given as v = digits_ * pow(10, exp_).
-  const char* digits_;
-  int num_digits_;
-  int exp_;
-  size_t size_;
-  float_specs specs_;
-  Char decimal_point_;
-
-  template <typename It> It prettify(It it) const {
-    // pow(10, full_exp - 1) <= v <= pow(10, full_exp).
-    int full_exp = num_digits_ + exp_;
-    if (specs_.format == float_format::exp) {
-      // Insert a decimal point after the first digit and add an exponent.
-      *it++ = static_cast<Char>(*digits_);
-      int num_zeros = specs_.precision - num_digits_;
-      if (num_digits_ > 1 || specs_.showpoint) *it++ = decimal_point_;
-      it = copy_str<Char>(digits_ + 1, digits_ + num_digits_, it);
-      if (num_zeros > 0 && specs_.showpoint)
-        it = std::fill_n(it, num_zeros, static_cast<Char>('0'));
-      *it++ = static_cast<Char>(specs_.upper ? 'E' : 'e');
-      return write_exponent<Char>(full_exp - 1, it);
-    }
-    if (num_digits_ <= full_exp) {
-      // 1234e7 -> 12340000000[.0+]
-      it = copy_str<Char>(digits_, digits_ + num_digits_, it);
-      it = std::fill_n(it, full_exp - num_digits_, static_cast<Char>('0'));
-      if (specs_.showpoint || specs_.precision < 0) {
-        *it++ = decimal_point_;
-        int num_zeros = specs_.precision - full_exp;
-        if (num_zeros <= 0) {
-          if (specs_.format != float_format::fixed)
-            *it++ = static_cast<Char>('0');
-          return it;
-        }
-#ifdef FMT_FUZZ
-        if (num_zeros > 5000)
-          throw std::runtime_error("fuzz mode - avoiding excessive cpu use");
-#endif
-        it = std::fill_n(it, num_zeros, static_cast<Char>('0'));
-      }
-    } else if (full_exp > 0) {
-      // 1234e-2 -> 12.34[0+]
-      it = copy_str<Char>(digits_, digits_ + full_exp, it);
-      if (!specs_.showpoint) {
-        // Remove trailing zeros.
-        int num_digits = num_digits_;
-        while (num_digits > full_exp && digits_[num_digits - 1] == '0')
-          --num_digits;
-        if (num_digits != full_exp) *it++ = decimal_point_;
-        return copy_str<Char>(digits_ + full_exp, digits_ + num_digits, it);
-      }
-      *it++ = decimal_point_;
-      it = copy_str<Char>(digits_ + full_exp, digits_ + num_digits_, it);
-      if (specs_.precision > num_digits_) {
-        // Add trailing zeros.
-        int num_zeros = specs_.precision - num_digits_;
-        it = std::fill_n(it, num_zeros, static_cast<Char>('0'));
-      }
-    } else {
-      // 1234e-6 -> 0.001234
-      *it++ = static_cast<Char>('0');
-      int num_zeros = -full_exp;
-      int num_digits = num_digits_;
-      if (num_digits == 0 && specs_.precision >= 0 &&
-          specs_.precision < num_zeros) {
-        num_zeros = specs_.precision;
-      }
-      // Remove trailing zeros.
-      if (!specs_.showpoint)
-        while (num_digits > 0 && digits_[num_digits - 1] == '0') --num_digits;
-      if (num_zeros != 0 || num_digits != 0 || specs_.showpoint) {
-        *it++ = decimal_point_;
-        it = std::fill_n(it, num_zeros, static_cast<Char>('0'));
-        it = copy_str<Char>(digits_, digits_ + num_digits, it);
-      }
-    }
-    return it;
-  }
-
- public:
-  float_writer(const char* digits, int num_digits, int exp, float_specs specs,
-               Char decimal_point)
-      : digits_(digits),
-        num_digits_(num_digits),
-        exp_(exp),
-        specs_(specs),
-        decimal_point_(decimal_point) {
-    int full_exp = num_digits + exp - 1;
-    int precision = specs.precision > 0 ? specs.precision : 16;
-    if (specs_.format == float_format::general &&
-        !(full_exp >= -4 && full_exp < precision)) {
-      specs_.format = float_format::exp;
-    }
-    size_ = prettify(counting_iterator()).count();
-    size_ += specs.sign ? 1 : 0;
-  }
-
-  size_t size() const { return size_; }
-
-  template <typename It> It operator()(It it) const {
-    if (specs_.sign) *it++ = static_cast<Char>(data::signs[specs_.sign]);
-    return prettify(it);
-  }
-};
-
 template <typename T>
 int format_float(T value, int precision, float_specs specs, buffer<char>& buf);
 
@@ -1384,7 +1498,7 @@ template <align::type align = align::left, typename OutputIt, typename Char,
           typename F>
 inline OutputIt write_padded(OutputIt out,
                              const basic_format_specs<Char>& specs, size_t size,
-                             size_t width, const F& f) {
+                             size_t width, F&& f) {
   static_assert(align == align::left || align == align::right, "");
   unsigned spec_width = to_unsigned(specs.width);
   size_t padding = spec_width > width ? spec_width - width : 0;
@@ -1402,7 +1516,7 @@ template <align::type align = align::left, typename OutputIt, typename Char,
           typename F>
 inline OutputIt write_padded(OutputIt out,
                              const basic_format_specs<Char>& specs, size_t size,
-                             const F& f) {
+                             F&& f) {
   return write_padded<align>(out, specs, size, size, f);
 }
 
@@ -1559,7 +1673,7 @@ template <typename OutputIt, typename Char, typename UInt> struct int_writer {
     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() && num_digits > *group && *group > 0 &&
+    while (group != groups.cend() && n > *group && *group > 0 &&
            *group != max_value<char>()) {
       size += sep_size;
       n -= *group;
@@ -1569,15 +1683,16 @@ template <typename OutputIt, typename Char, typename UInt> struct int_writer {
     char digits[40];
     format_decimal(digits, abs_value, num_digits);
     basic_memory_buffer<Char> buffer;
-    size += prefix_size;
-    buffer.resize(size);
+    size += static_cast<int>(prefix_size);
+    const auto usize = to_unsigned(size);
+    buffer.resize(usize);
     basic_string_view<Char> s(&sep, sep_size);
     // Index of a decimal digit with the least significant digit having index 0.
     int digit_index = 0;
     group = groups.cbegin();
-    auto p = buffer.data() + size;
-    for (int i = num_digits - 1; i >= 0; --i) {
-      *--p = static_cast<Char>(digits[i]);
+    auto p = buffer.data() + size - 1;
+    for (int i = num_digits - 1; i > 0; --i) {
+      *p-- = static_cast<Char>(digits[i]);
       if (*group <= 0 || ++digit_index % *group != 0 ||
           *group == max_value<char>())
         continue;
@@ -1585,12 +1700,16 @@ template <typename OutputIt, typename Char, typename UInt> struct int_writer {
         digit_index = 0;
         ++group;
       }
-      p -= s.size();
       std::uninitialized_copy(s.data(), s.data() + s.size(),
                               make_checked(p, s.size()));
+      p -= s.size();
     }
-    if (prefix_size != 0) p[-1] = static_cast<Char>('-');
-    write(out, basic_string_view<Char>(buffer.data(), buffer.size()), specs);
+    *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); }
@@ -1616,6 +1735,168 @@ OutputIt write_nonfinite(OutputIt out, bool isinf,
   });
 }
 
+// A decimal floating-point number significand * pow(10, exp).
+struct big_decimal_fp {
+  const char* significand;
+  int significand_size;
+  int exponent;
+};
+
+inline int get_significand_size(const big_decimal_fp& fp) {
+  return fp.significand_size;
+}
+template <typename T>
+inline int get_significand_size(const dragonbox::decimal_fp<T>& fp) {
+  return count_digits(fp.significand);
+}
+
+template <typename Char, typename OutputIt>
+inline OutputIt write_significand(OutputIt out, const char* significand,
+                                  int& significand_size) {
+  return copy_str<Char>(significand, significand + significand_size, out);
+}
+template <typename Char, typename OutputIt, typename UInt>
+inline OutputIt write_significand(OutputIt out, UInt significand,
+                                  int significand_size) {
+  return format_decimal<Char>(out, significand, significand_size).end;
+}
+
+template <typename Char, typename UInt,
+          FMT_ENABLE_IF(std::is_integral<UInt>::value)>
+inline Char* write_significand(Char* out, UInt significand,
+                               int significand_size, int integral_size,
+                               Char decimal_point) {
+  if (!decimal_point)
+    return format_decimal(out, significand, significand_size).end;
+  auto end = format_decimal(out + 1, significand, significand_size).end;
+  if (integral_size == 1)
+    out[0] = out[1];
+  else
+    std::copy_n(out + 1, integral_size, out);
+  out[integral_size] = decimal_point;
+  return end;
+}
+
+template <typename OutputIt, typename UInt, typename Char,
+          FMT_ENABLE_IF(!std::is_pointer<remove_cvref_t<OutputIt>>::value)>
+inline OutputIt write_significand(OutputIt out, UInt significand,
+                                  int significand_size, int integral_size,
+                                  Char decimal_point) {
+  // Buffer is large enough to hold digits (digits10 + 1) and a decimal point.
+  Char buffer[digits10<UInt>() + 2];
+  auto end = write_significand(buffer, significand, significand_size,
+                               integral_size, decimal_point);
+  return detail::copy_str<Char>(buffer, end, out);
+}
+
+template <typename OutputIt, typename Char>
+inline OutputIt write_significand(OutputIt out, const char* significand,
+                                  int significand_size, int integral_size,
+                                  Char decimal_point) {
+  out = detail::copy_str<Char>(significand, significand + integral_size, out);
+  if (!decimal_point) return out;
+  *out++ = decimal_point;
+  return detail::copy_str<Char>(significand + integral_size,
+                                significand + significand_size, out);
+}
+
+template <typename OutputIt, typename DecimalFP, typename Char>
+OutputIt write_float(OutputIt out, const DecimalFP& fp,
+                     const basic_format_specs<Char>& specs, float_specs fspecs,
+                     Char decimal_point) {
+  auto significand = fp.significand;
+  int significand_size = get_significand_size(fp);
+  static const Char zero = static_cast<Char>('0');
+  auto sign = fspecs.sign;
+  size_t size = to_unsigned(significand_size) + (sign ? 1 : 0);
+  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
+
+  int output_exp = fp.exponent + significand_size - 1;
+  auto use_exp_format = [=]() {
+    if (fspecs.format == float_format::exp) return true;
+    if (fspecs.format != float_format::general) return false;
+    // Use the fixed notation if the exponent is in [exp_lower, exp_upper),
+    // e.g. 0.0001 instead of 1e-04. Otherwise use the exponent notation.
+    const int exp_lower = -4, exp_upper = 16;
+    return output_exp < exp_lower ||
+           output_exp >= (fspecs.precision > 0 ? fspecs.precision : exp_upper);
+  };
+  if (use_exp_format()) {
+    int num_zeros = 0;
+    if (fspecs.showpoint) {
+      num_zeros = (std::max)(fspecs.precision - significand_size, 0);
+      size += to_unsigned(num_zeros);
+    } else if (significand_size == 1) {
+      decimal_point = Char();
+    }
+    auto abs_output_exp = output_exp >= 0 ? output_exp : -output_exp;
+    int exp_digits = 2;
+    if (abs_output_exp >= 100) exp_digits = abs_output_exp >= 1000 ? 4 : 3;
+
+    size += to_unsigned((decimal_point ? 1 : 0) + 2 + exp_digits);
+    char exp_char = fspecs.upper ? 'E' : 'e';
+    auto write = [=](iterator it) {
+      if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+      // Insert a decimal point after the first digit and add an exponent.
+      it = write_significand(it, significand, significand_size, 1,
+                             decimal_point);
+      if (num_zeros > 0) it = std::fill_n(it, num_zeros, zero);
+      *it++ = static_cast<Char>(exp_char);
+      return write_exponent<Char>(output_exp, it);
+    };
+    return specs.width > 0 ? write_padded<align::right>(out, specs, size, write)
+                           : base_iterator(out, write(reserve(out, size)));
+  }
+
+  int exp = fp.exponent + significand_size;
+  if (fp.exponent >= 0) {
+    // 1234e5 -> 123400000[.0+]
+    size += to_unsigned(fp.exponent);
+    int num_zeros = fspecs.precision - exp;
+#ifdef FMT_FUZZ
+    if (num_zeros > 5000)
+      throw std::runtime_error("fuzz mode - avoiding excessive cpu use");
+#endif
+    if (fspecs.showpoint) {
+      if (num_zeros <= 0 && fspecs.format != float_format::fixed) num_zeros = 1;
+      if (num_zeros > 0) size += to_unsigned(num_zeros);
+    }
+    return write_padded<align::right>(out, specs, size, [&](iterator it) {
+      if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+      it = write_significand<Char>(it, significand, significand_size);
+      it = std::fill_n(it, fp.exponent, zero);
+      if (!fspecs.showpoint) return it;
+      *it++ = decimal_point;
+      return num_zeros > 0 ? std::fill_n(it, num_zeros, zero) : it;
+    });
+  } else if (exp > 0) {
+    // 1234e-2 -> 12.34[0+]
+    int num_zeros = fspecs.showpoint ? fspecs.precision - significand_size : 0;
+    size += 1 + to_unsigned(num_zeros > 0 ? num_zeros : 0);
+    return write_padded<align::right>(out, specs, size, [&](iterator it) {
+      if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+      it = write_significand(it, significand, significand_size, exp,
+                             decimal_point);
+      return num_zeros > 0 ? std::fill_n(it, num_zeros, zero) : it;
+    });
+  }
+  // 1234e-6 -> 0.001234
+  int num_zeros = -exp;
+  if (significand_size == 0 && fspecs.precision >= 0 &&
+      fspecs.precision < num_zeros) {
+    num_zeros = fspecs.precision;
+  }
+  size += 2 + to_unsigned(num_zeros);
+  return write_padded<align::right>(out, specs, size, [&](iterator it) {
+    if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+    *it++ = zero;
+    if (num_zeros == 0 && significand_size == 0 && !fspecs.showpoint) return it;
+    *it++ = decimal_point;
+    it = std::fill_n(it, num_zeros, zero);
+    return write_significand<Char>(it, significand, significand_size);
+  });
+}
+
 template <typename Char, typename OutputIt, typename T,
           FMT_ENABLE_IF(std::is_floating_point<T>::value)>
 OutputIt write(OutputIt out, T value, basic_format_specs<Char> specs,
@@ -1655,39 +1936,45 @@ OutputIt write(OutputIt out, T value, basic_format_specs<Char> specs,
       ++precision;
   }
   if (const_check(std::is_same<T, float>())) fspecs.binary32 = true;
-  fspecs.use_grisu = use_grisu<T>();
+  fspecs.use_grisu = is_fast_float<T>();
   int exp = format_float(promote_float(value), precision, fspecs, buffer);
   fspecs.precision = precision;
   Char point =
       fspecs.locale ? decimal_point<Char>(loc) : static_cast<Char>('.');
-  float_writer<Char> w(buffer.data(), static_cast<int>(buffer.size()), exp,
-                       fspecs, point);
-  return write_padded<align::right>(out, specs, w.size(), w);
+  auto fp = big_decimal_fp{buffer.data(), static_cast<int>(buffer.size()), exp};
+  return write_float(out, fp, specs, fspecs, point);
 }
 
 template <typename Char, typename OutputIt, typename T,
-          FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+          FMT_ENABLE_IF(is_fast_float<T>::value)>
 OutputIt write(OutputIt out, T value) {
   if (const_check(!is_supported_floating_point(value))) return out;
+
+  using floaty = conditional_t<std::is_same<T, long double>::value, double, T>;
+  using uint = typename dragonbox::float_info<floaty>::carrier_uint;
+  auto bits = bit_cast<uint>(value);
+
   auto fspecs = float_specs();
-  if (std::signbit(value)) {  // value < 0 is false for NaN so use signbit.
+  auto sign_bit = bits & (uint(1) << (num_bits<uint>() - 1));
+  if (sign_bit != 0) {
     fspecs.sign = sign::minus;
     value = -value;
   }
 
-  auto specs = basic_format_specs<Char>();
-  if (!std::isfinite(value))
+  static const auto specs = basic_format_specs<Char>();
+  uint mask = exponent_mask<floaty>();
+  if ((bits & mask) == mask)
     return write_nonfinite(out, std::isinf(value), specs, fspecs);
 
-  memory_buffer buffer;
-  int precision = -1;
-  if (const_check(std::is_same<T, float>())) fspecs.binary32 = true;
-  fspecs.use_grisu = use_grisu<T>();
-  int exp = format_float(promote_float(value), precision, fspecs, buffer);
-  fspecs.precision = precision;
-  float_writer<Char> w(buffer.data(), static_cast<int>(buffer.size()), exp,
-                       fspecs, static_cast<Char>('.'));
-  return base_iterator(out, w(reserve(out, w.size())));
+  auto dec = dragonbox::to_decimal(static_cast<floaty>(value));
+  return write_float(out, dec, specs, fspecs, static_cast<Char>('.'));
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(std::is_floating_point<T>::value &&
+                        !is_fast_float<T>::value)>
+inline OutputIt write(OutputIt out, T value) {
+  return write(out, value, basic_format_specs<Char>());
 }
 
 template <typename Char, typename OutputIt>
@@ -1740,6 +2027,13 @@ OutputIt write(OutputIt out, basic_string_view<Char> value) {
   return base_iterator(out, it);
 }
 
+template <typename Char>
+buffer_appender<Char> write(buffer_appender<Char> out,
+                            basic_string_view<Char> value) {
+  get_container(out).append(value.begin(), value.end());
+  return out;
+}
+
 template <typename Char, typename OutputIt, typename T,
           FMT_ENABLE_IF(is_integral<T>::value &&
                         !std::is_same<T, bool>::value &&
@@ -1750,7 +2044,13 @@ OutputIt write(OutputIt out, T value) {
   // Don't do -abs_value since it trips unsigned-integer-overflow sanitizer.
   if (negative) abs_value = ~abs_value + 1;
   int num_digits = count_digits(abs_value);
-  auto it = reserve(out, (negative ? 1 : 0) + static_cast<size_t>(num_digits));
+  auto size = (negative ? 1 : 0) + static_cast<size_t>(num_digits);
+  auto it = reserve(out, size);
+  if (auto ptr = to_pointer<Char>(it, size)) {
+    if (negative) *ptr++ = static_cast<Char>('-');
+    format_decimal<Char>(ptr, abs_value, num_digits);
+    return out;
+  }
   if (negative) *it++ = static_cast<Char>('-');
   it = format_decimal<Char>(it, abs_value, num_digits).end;
   return base_iterator(out, it);
@@ -1789,8 +2089,13 @@ auto write(OutputIt out, const T& value) -> typename std::enable_if<
     mapped_type_constant<T, basic_format_context<OutputIt, Char>>::value ==
         type::custom_type,
     OutputIt>::type {
-  basic_format_context<OutputIt, Char> ctx(out, {}, {});
-  return formatter<T>().format(value, ctx);
+  using context_type = basic_format_context<OutputIt, Char>;
+  using formatter_type =
+      conditional_t<has_formatter<T, context_type>::value,
+                    typename context_type::template formatter_type<T>,
+                    fallback_formatter<T, Char>>;
+  context_type ctx(out, {}, {});
+  return formatter_type().format(value, ctx);
 }
 
 // An argument visitor that formats the argument and writes it via the output
@@ -1996,6 +2301,48 @@ class arg_formatter_base {
   }
 };
 
+/** The default argument formatter. */
+template <typename OutputIt, typename Char>
+class arg_formatter : public arg_formatter_base<OutputIt, Char> {
+ private:
+  using char_type = Char;
+  using base = arg_formatter_base<OutputIt, Char>;
+  using context_type = basic_format_context<OutputIt, Char>;
+
+  context_type& ctx_;
+  basic_format_parse_context<char_type>* parse_ctx_;
+  const Char* ptr_;
+
+ public:
+  using iterator = typename base::iterator;
+  using format_specs = typename base::format_specs;
+
+  /**
+    \rst
+    Constructs an argument formatter object.
+    *ctx* is a reference to the formatting context,
+    *specs* contains format specifier information for standard argument types.
+    \endrst
+   */
+  explicit arg_formatter(
+      context_type& ctx,
+      basic_format_parse_context<char_type>* parse_ctx = nullptr,
+      format_specs* specs = nullptr, const Char* ptr = nullptr)
+      : base(ctx.out(), specs, ctx.locale()),
+        ctx_(ctx),
+        parse_ctx_(parse_ctx),
+        ptr_(ptr) {}
+
+  using base::operator();
+
+  /** Formats an argument of a user-defined type. */
+  iterator operator()(typename basic_format_arg<context_type>::handle handle) {
+    if (ptr_) advance_to(*parse_ctx_, ptr_);
+    handle.format(*parse_ctx_, ctx_);
+    return ctx_.out();
+  }
+};
+
 template <typename Char> FMT_CONSTEXPR bool is_name_start(Char c) {
   return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || '_' == c;
 }
@@ -2035,12 +2382,11 @@ template <typename Context> class custom_formatter {
                             Context& ctx)
       : parse_ctx_(parse_ctx), ctx_(ctx) {}
 
-  bool operator()(typename basic_format_arg<Context>::handle h) const {
+  void operator()(typename basic_format_arg<Context>::handle h) const {
     h.format(parse_ctx_, ctx_);
-    return true;
   }
 
-  template <typename T> bool operator()(T) const { return false; }
+  template <typename T> void operator()(T) const {}
 };
 
 template <typename T>
@@ -2422,12 +2768,30 @@ template <typename SpecHandler, typename Char> struct precision_adapter {
 };
 
 template <typename Char>
-FMT_CONSTEXPR const Char* next_code_point(const Char* begin, const Char* end) {
-  if (const_check(sizeof(Char) != 1) || (*begin & 0x80) == 0) return begin + 1;
-  do {
-    ++begin;
-  } while (begin != end && (*begin & 0xc0) == 0x80);
-  return begin;
+FMT_CONSTEXPR int code_point_length(const Char* begin) {
+  if (const_check(sizeof(Char) != 1)) return 1;
+  constexpr char lengths[] = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+                              0, 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 2, 3, 3, 4, 0};
+  int len = lengths[static_cast<unsigned char>(*begin) >> 3];
+
+  // Compute the pointer to the next character early so that the next
+  // iteration can start working on the next character. Neither Clang
+  // nor GCC figure out this reordering on their own.
+  return len + !len;
+}
+
+template <typename Char> constexpr bool is_ascii_letter(Char c) {
+  return (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z');
+}
+
+// Converts a character to ASCII. Returns a number > 127 on conversion failure.
+template <typename Char, FMT_ENABLE_IF(std::is_integral<Char>::value)>
+constexpr Char to_ascii(Char value) {
+  return value;
+}
+template <typename Char, FMT_ENABLE_IF(std::is_enum<Char>::value)>
+constexpr typename std::underlying_type<Char>::type to_ascii(Char value) {
+  return value;
 }
 
 // Parses fill and alignment.
@@ -2436,10 +2800,10 @@ FMT_CONSTEXPR const Char* parse_align(const Char* begin, const Char* end,
                                       Handler&& handler) {
   FMT_ASSERT(begin != end, "");
   auto align = align::none;
-  auto p = next_code_point(begin, end);
-  if (p == end) p = begin;
+  auto p = begin + code_point_length(begin);
+  if (p >= end) p = begin;
   for (;;) {
-    switch (static_cast<char>(*p)) {
+    switch (to_ascii(*p)) {
     case '<':
       align = align::left;
       break;
@@ -2518,13 +2882,13 @@ 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;
@@ -2601,7 +2965,7 @@ FMT_CONSTEXPR const Char* parse_replacement_field(const Char* begin,
                                                   Handler&& handler) {
   ++begin;
   if (begin == end) return handler.on_error("invalid format string"), end;
-  if (static_cast<char>(*begin) == '}') {
+  if (*begin == '}') {
     handler.on_replacement_field(handler.on_arg_id(), begin);
   } else if (*begin == '{') {
     handler.on_text(begin, begin + 1);
@@ -2646,17 +3010,17 @@ FMT_CONSTEXPR_DECL FMT_INLINE void parse_format_string(
     return;
   }
   struct writer {
-    FMT_CONSTEXPR void operator()(const Char* begin, const Char* end) {
-      if (begin == end) return;
+    FMT_CONSTEXPR void operator()(const Char* pbegin, const Char* pend) {
+      if (pbegin == pend) return;
       for (;;) {
         const Char* p = nullptr;
-        if (!find<IS_CONSTEXPR>(begin, end, '}', p))
-          return handler_.on_text(begin, end);
+        if (!find<IS_CONSTEXPR>(pbegin, pend, '}', p))
+          return handler_.on_text(pbegin, pend);
         ++p;
-        if (p == end || *p != '}')
+        if (p == pend || *p != '}')
           return handler_.on_error("unmatched '}' in format string");
-        handler_.on_text(begin, p);
-        begin = p + 1;
+        handler_.on_text(pbegin, p);
+        pbegin = p + 1;
       }
     }
     Handler& handler_;
@@ -2687,12 +3051,12 @@ FMT_CONSTEXPR const typename ParseContext::char_type* parse_format_specs(
   return f.parse(ctx);
 }
 
-template <typename ArgFormatter, typename Char, typename Context>
+template <typename OutputIt, typename Char, typename Context>
 struct format_handler : detail::error_handler {
   basic_format_parse_context<Char> parse_context;
   Context context;
 
-  format_handler(typename ArgFormatter::iterator out,
+  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) {}
@@ -2716,26 +3080,33 @@ struct format_handler : detail::error_handler {
   FMT_INLINE void on_replacement_field(int id, const Char*) {
     auto arg = get_arg(context, id);
     context.advance_to(visit_format_arg(
-        default_arg_formatter<typename ArgFormatter::iterator, Char>{
+        default_arg_formatter<OutputIt, Char>{
             context.out(), context.args(), context.locale()},
         arg));
   }
 
   const Char* on_format_specs(int id, const Char* begin, const Char* end) {
-    advance_to(parse_context, begin);
     auto arg = get_arg(context, id);
-    custom_formatter<Context> f(parse_context, context);
-    if (visit_format_arg(f, arg)) return parse_context.begin();
-    basic_format_specs<Char> specs;
-    using parse_context_t = basic_format_parse_context<Char>;
-    specs_checker<specs_handler<parse_context_t, Context>> handler(
-        specs_handler<parse_context_t, Context>(specs, parse_context, context),
-        arg.type());
-    begin = parse_format_specs(begin, end, handler);
-    if (begin == end || *begin != '}') on_error("missing '}' in format string");
-    advance_to(parse_context, begin);
+    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(ArgFormatter(context, &parse_context, &specs), arg));
+        visit_format_arg(arg_formatter<OutputIt, Char>(context, &parse_context, &specs), arg));
     return begin;
   }
 };
@@ -2887,53 +3258,11 @@ FMT_API void format_error_code(buffer<char>& out, int error_code,
 
 FMT_API void report_error(format_func func, int error_code,
                           string_view message) FMT_NOEXCEPT;
-
-/** The default argument formatter. */
-template <typename OutputIt, typename Char>
-class arg_formatter : public arg_formatter_base<OutputIt, Char> {
- private:
-  using char_type = Char;
-  using base = arg_formatter_base<OutputIt, Char>;
-  using context_type = basic_format_context<OutputIt, Char>;
-
-  context_type& ctx_;
-  basic_format_parse_context<char_type>* parse_ctx_;
-  const Char* ptr_;
-
- public:
-  using iterator = typename base::iterator;
-  using format_specs = typename base::format_specs;
-
-  /**
-    \rst
-    Constructs an argument formatter object.
-    *ctx* is a reference to the formatting context,
-    *specs* contains format specifier information for standard argument types.
-    \endrst
-   */
-  explicit arg_formatter(
-      context_type& ctx,
-      basic_format_parse_context<char_type>* parse_ctx = nullptr,
-      format_specs* specs = nullptr, const Char* ptr = nullptr)
-      : base(ctx.out(), specs, ctx.locale()),
-        ctx_(ctx),
-        parse_ctx_(parse_ctx),
-        ptr_(ptr) {}
-
-  using base::operator();
-
-  /** Formats an argument of a user-defined type. */
-  iterator operator()(typename basic_format_arg<context_type>::handle handle) {
-    if (ptr_) advance_to(*parse_ctx_, ptr_);
-    handle.format(*parse_ctx_, ctx_);
-    return ctx_.out();
-  }
-};
 }  // namespace detail
 
 template <typename OutputIt, typename Char>
 using arg_formatter FMT_DEPRECATED_ALIAS =
-  detail::arg_formatter<OutputIt, Char>;
+    detail::arg_formatter<OutputIt, Char>;
 
 /**
  An error returned by an operating system or a language runtime,
@@ -3196,8 +3525,10 @@ struct formatter<Char[N], Char> : formatter<basic_string_view<Char>, Char> {
 //   using variant = std::variant<int, std::string>;
 //   template <>
 //   struct formatter<variant>: dynamic_formatter<> {
-//     void format(buffer &buf, const variant &v, context &ctx) {
-//       visit([&](const auto &val) { format(buf, val, ctx); }, v);
+//     auto format(const variant& v, format_context& ctx) {
+//       return visit([&](const auto& val) {
+//           return dynamic_formatter<>::format(val, ctx);
+//       }, v);
 //     }
 //   };
 template <typename Char = char> class dynamic_formatter {
@@ -3265,28 +3596,15 @@ FMT_CONSTEXPR void advance_to(
   ctx.advance_to(ctx.begin() + (p - &*ctx.begin()));
 }
 
-/** Formats arguments and writes the output to the range. */
-template <typename ArgFormatter, typename Char, typename Context>
-typename Context::iterator vformat_to(
-    typename ArgFormatter::iterator out, basic_string_view<Char> format_str,
-    basic_format_args<Context> args,
-    detail::locale_ref loc = detail::locale_ref()) {
-  if (format_str.size() == 2 && detail::equal2(format_str.data(), "{}")) {
-    auto arg = args.get(0);
-    if (!arg) detail::error_handler().on_error("argument not found");
-    using iterator = typename ArgFormatter::iterator;
-    return visit_format_arg(
-        detail::default_arg_formatter<iterator, Char>{out, args, loc}, arg);
-  }
-  detail::format_handler<ArgFormatter, Char, Context> h(out, format_str, args,
-                                                        loc);
-  detail::parse_format_string<false>(format_str, h);
-  return h.context.out();
-}
+/**
+  \rst
+  Converts ``p`` to ``const void*`` for pointer formatting.
 
-// Casts ``p`` to ``const void*`` for pointer formatting.
-// Example:
-//   auto s = format("{}", ptr(p));
+  **Example**::
+
+    auto s = fmt::format("{}", fmt::ptr(p));
+  \endrst
+ */
 template <typename T> inline const void* ptr(const T* p) { return p; }
 template <typename T> inline const void* ptr(const std::unique_ptr<T>& p) {
   return p.get();
@@ -3305,6 +3623,10 @@ class bytes {
 };
 
 template <> struct formatter<bytes> {
+ private:
+  detail::dynamic_format_specs<char> specs_;
+
+ public:
   template <typename ParseContext>
   FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
     using handler_type = detail::dynamic_specs_handler<ParseContext>;
@@ -3323,9 +3645,6 @@ template <> struct formatter<bytes> {
         specs_.precision, specs_.precision_ref, ctx);
     return detail::write_bytes(ctx.out(), b.data_, specs_);
   }
-
- private:
-  detail::dynamic_format_specs<char> specs_;
 };
 
 template <typename It, typename Sentinel, typename Char>
@@ -3390,15 +3709,14 @@ arg_join<It, Sentinel, wchar_t> join(It begin, Sentinel end, wstring_view sep) {
   \endrst
  */
 template <typename Range>
-arg_join<detail::iterator_t<const Range>, detail::sentinel_t<const Range>, char>
-join(const Range& range, string_view sep) {
+arg_join<detail::iterator_t<Range>, detail::sentinel_t<Range>, char> join(
+    Range&& range, string_view sep) {
   return join(std::begin(range), std::end(range), sep);
 }
 
 template <typename Range>
-arg_join<detail::iterator_t<const Range>, detail::sentinel_t<const Range>,
-         wchar_t>
-join(const Range& range, wstring_view sep) {
+arg_join<detail::iterator_t<Range>, detail::sentinel_t<Range>, wchar_t> join(
+    Range&& range, wstring_view sep) {
   return join(std::begin(range), std::end(range), sep);
 }
 
@@ -3425,7 +3743,7 @@ inline std::string to_string(T value) {
   // The buffer should be large enough to store the number including the sign or
   // "false" for bool.
   constexpr int max_size = detail::digits10<T>() + 2;
-  char buffer[max_size > 5 ? max_size : 5];
+  char buffer[max_size > 5 ? static_cast<unsigned>(max_size) : 5];
   char* begin = buffer;
   return std::string(begin, detail::write<char>(begin, value));
 }
@@ -3445,17 +3763,28 @@ std::basic_string<Char> to_string(const basic_memory_buffer<Char, SIZE>& buf) {
 }
 
 template <typename Char>
-typename buffer_context<Char>::iterator detail::vformat_to(
+void detail::vformat_to(
     detail::buffer<Char>& buf, basic_string_view<Char> format_str,
-    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
-  using af = arg_formatter<typename buffer_context<Char>::iterator, Char>;
-  return vformat_to<af>(std::back_inserter(buf), to_string_view(format_str),
-                        args);
+    basic_format_args<buffer_context<type_identity_t<Char>>> args,
+    detail::locale_ref loc) {
+  using iterator = typename buffer_context<Char>::iterator;
+  auto out = buffer_appender<Char>(buf);
+  if (format_str.size() == 2 && equal2(format_str.data(), "{}")) {
+    auto arg = args.get(0);
+    if (!arg) error_handler().on_error("argument not found");
+    visit_format_arg(default_arg_formatter<iterator, Char>{out, args, loc},
+                     arg);
+    return;
+  }
+  format_handler<iterator, Char, buffer_context<Char>> h(
+      out, format_str, args, loc);
+  parse_format_string<false>(format_str, h);
 }
 
 #ifndef FMT_HEADER_ONLY
-extern template format_context::iterator detail::vformat_to(
-    detail::buffer<char>&, string_view, basic_format_args<format_context>);
+extern template void detail::vformat_to(detail::buffer<char>&, string_view,
+                                        basic_format_args<format_context>,
+                                        detail::locale_ref);
 namespace detail {
 extern template FMT_API std::string grouping_impl<char>(locale_ref loc);
 extern template FMT_API std::string grouping_impl<wchar_t>(locale_ref loc);
@@ -3482,7 +3811,7 @@ extern template int snprintf_float<long double>(long double value,
 
 template <typename S, typename Char = char_t<S>,
           FMT_ENABLE_IF(detail::is_string<S>::value)>
-inline typename FMT_BUFFER_CONTEXT(Char)::iterator vformat_to(
+inline void vformat_to(
     detail::buffer<Char>& buf, const S& format_str,
     basic_format_args<FMT_BUFFER_CONTEXT(type_identity_t<Char>)> args) {
   return detail::vformat_to(buf, to_string_view(format_str), args);
@@ -3492,10 +3821,9 @@ template <typename S, typename... Args, size_t SIZE = inline_buffer_size,
           typename Char = enable_if_t<detail::is_string<S>::value, char_t<S>>>
 inline typename buffer_context<Char>::iterator format_to(
     basic_memory_buffer<Char, SIZE>& buf, const S& format_str, Args&&... args) {
-  detail::check_format_string<Args...>(format_str);
-  using context = buffer_context<Char>;
-  return detail::vformat_to(buf, to_string_view(format_str),
-                            make_format_args<context>(args...));
+  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
+  detail::vformat_to(buf, to_string_view(format_str), vargs);
+  return detail::buffer_appender<Char>(buf);
 }
 
 template <typename OutputIt, typename Char = char>
@@ -3504,88 +3832,17 @@ using format_context_t = basic_format_context<OutputIt, Char>;
 template <typename OutputIt, typename Char = char>
 using format_args_t = basic_format_args<format_context_t<OutputIt, Char>>;
 
-template <
-    typename S, typename OutputIt, typename... Args,
-    FMT_ENABLE_IF(detail::is_output_iterator<OutputIt>::value &&
-                  !detail::is_contiguous_back_insert_iterator<OutputIt>::value)>
-inline OutputIt vformat_to(
-    OutputIt out, const S& format_str,
-    format_args_t<type_identity_t<OutputIt>, char_t<S>> args) {
-  using af = detail::arg_formatter<OutputIt, char_t<S>>;
-  return vformat_to<af>(out, to_string_view(format_str), args);
-}
-
-/**
- \rst
- Formats arguments, writes the result to the output iterator ``out`` and returns
- the iterator past the end of the output range.
-
- **Example**::
-
-   std::vector<char> out;
-   fmt::format_to(std::back_inserter(out), "{}", 42);
- \endrst
- */
-template <typename OutputIt, typename S, typename... Args,
-          FMT_ENABLE_IF(
-              detail::is_output_iterator<OutputIt>::value &&
-              !detail::is_contiguous_back_insert_iterator<OutputIt>::value &&
-              detail::is_string<S>::value)>
-inline OutputIt format_to(OutputIt out, const S& format_str, Args&&... args) {
-  detail::check_format_string<Args...>(format_str);
-  using context = format_context_t<OutputIt, char_t<S>>;
-  return vformat_to(out, to_string_view(format_str),
-                    make_format_args<context>(args...));
-}
-
-template <typename OutputIt> struct format_to_n_result {
-  /** Iterator past the end of the output range. */
-  OutputIt out;
-  /** Total (not truncated) output size. */
-  size_t size;
-};
-
 template <typename OutputIt, typename Char = typename OutputIt::value_type>
-using format_to_n_context =
-    format_context_t<detail::truncating_iterator<OutputIt>, Char>;
+using format_to_n_context FMT_DEPRECATED_ALIAS = buffer_context<Char>;
 
 template <typename OutputIt, typename Char = typename OutputIt::value_type>
-using format_to_n_args = basic_format_args<format_to_n_context<OutputIt, Char>>;
+using format_to_n_args FMT_DEPRECATED_ALIAS =
+    basic_format_args<buffer_context<Char>>;
 
 template <typename OutputIt, typename Char, typename... Args>
-inline format_arg_store<format_to_n_context<OutputIt, Char>, Args...>
+FMT_DEPRECATED format_arg_store<buffer_context<Char>, Args...>
 make_format_to_n_args(const Args&... args) {
-  return format_arg_store<format_to_n_context<OutputIt, Char>, Args...>(
-      args...);
-}
-
-template <typename OutputIt, typename Char, typename... Args,
-          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt>::value)>
-inline format_to_n_result<OutputIt> vformat_to_n(
-    OutputIt out, size_t n, basic_string_view<Char> format_str,
-    format_to_n_args<type_identity_t<OutputIt>, type_identity_t<Char>> args) {
-  auto it = vformat_to(detail::truncating_iterator<OutputIt>(out, n),
-                       format_str, args);
-  return {it.base(), it.count()};
-}
-
-/**
- \rst
- Formats arguments, writes up to ``n`` characters of the result to the output
- iterator ``out`` and returns the total output size and the iterator past the
- end of the output range.
- \endrst
- */
-template <typename OutputIt, typename S, typename... Args,
-          FMT_ENABLE_IF(detail::is_string<S>::value&&
-                            detail::is_output_iterator<OutputIt>::value)>
-inline format_to_n_result<OutputIt> format_to_n(OutputIt out, size_t n,
-                                                const S& format_str,
-                                                const Args&... args) {
-  detail::check_format_string<Args...>(format_str);
-  using context = format_to_n_context<OutputIt, char_t<S>>;
-  return vformat_to_n(out, n, to_string_view(format_str),
-                      make_format_args<context>(args...));
+  return format_arg_store<buffer_context<Char>, Args...>(args...);
 }
 
 template <typename Char, enable_if_t<(!std::is_same<Char, char>::value), int>>
@@ -3597,15 +3854,6 @@ std::basic_string<Char> detail::vformat(
   return to_string(buffer);
 }
 
-/**
-  Returns the number of characters in the output of
-  ``format(format_str, args...)``.
- */
-template <typename... Args>
-inline size_t formatted_size(string_view format_str, const Args&... args) {
-  return format_to(detail::counting_iterator(), format_str, args...).count();
-}
-
 template <typename Char, FMT_ENABLE_IF(std::is_same<Char, wchar_t>::value)>
 void vprint(std::FILE* f, basic_string_view<Char> format_str,
             wformat_args args) {
@@ -3630,8 +3878,7 @@ template <typename Char, Char... CHARS> class udl_formatter {
   template <typename... Args>
   std::basic_string<Char> operator()(Args&&... args) const {
     static FMT_CONSTEXPR_DECL Char s[] = {CHARS..., '\0'};
-    check_format_string<remove_cvref_t<Args>...>(FMT_STRING(s));
-    return format(s, std::forward<Args>(args)...);
+    return format(FMT_STRING(s), std::forward<Args>(args)...);
   }
 };
 #  else
diff --git a/include/fmt/locale.h b/include/fmt/locale.h
index 988d15cd..517f6505 100644
--- a/include/fmt/locale.h
+++ b/include/fmt/locale.h
@@ -16,21 +16,11 @@ FMT_BEGIN_NAMESPACE
 
 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<type_identity_t<Char>>> args) {
-  using af = arg_formatter<typename buffer_context<Char>::iterator, Char>;
-  return vformat_to<af>(std::back_inserter(buf), to_string_view(format_str),
-                        args, detail::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<type_identity_t<Char>>> args) {
   basic_memory_buffer<Char> buffer;
-  detail::vformat_to(loc, buffer, format_str, args);
+  detail::vformat_to(buffer, format_str, args, detail::locale_ref(loc));
   return fmt::to_string(buffer);
 }
 }  // namespace detail
@@ -45,32 +35,28 @@ inline std::basic_string<Char> vformat(
 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 detail::vformat(
-      loc, to_string_view(format_str),
-      detail::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<
-              detail::is_output_iterator<OutputIt>::value, char_t<S>>>
+          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,
-    format_args_t<type_identity_t<OutputIt>, Char> args) {
-  using af = detail::arg_formatter<OutputIt, Char>;
-  return vformat_to<af>(out, to_string_view(format_str), args,
-                        detail::locale_ref(loc));
+    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(detail::is_output_iterator<OutputIt>::value&&
-                            detail::is_string<S>::value)>
+          typename Char = char_t<S>,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, Char>::value)>
 inline OutputIt format_to(OutputIt out, const std::locale& loc,
                           const S& format_str, Args&&... args) {
-  detail::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));
+  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/include/fmt/os.h b/include/fmt/os.h
index a9517ef8..88151006 100644
--- a/include/fmt/os.h
+++ b/include/fmt/os.h
@@ -29,7 +29,8 @@
 #if FMT_HAS_INCLUDE("winapifamily.h")
 #  include <winapifamily.h>
 #endif
-#if FMT_HAS_INCLUDE("fcntl.h") && \
+#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
@@ -278,7 +279,8 @@ class file {
     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.
+    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.
@@ -343,62 +345,90 @@ class file {
 // Returns the memory page size.
 long getpagesize();
 
-class direct_buffered_file;
+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 S, typename... Args>
-void print(direct_buffered_file& f, const S& format_str,
-           const Args&... args);
+  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
 
-// A buffered file with a direct buffer access and no synchronization.
-class direct_buffered_file {
+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_;
 
-  enum { buffer_size = 4096 };
-  char buffer_[buffer_size];
-  int pos_;
-
   void flush() {
-    if (pos_ == 0) return;
-    file_.write(buffer_, pos_);
-    pos_ = 0;
+    if (size() == 0) return;
+    file_.write(data(), size());
+    clear();
   }
 
-  int free_capacity() const { return buffer_size - pos_; }
+  void grow(size_t) final;
 
- public:
-  direct_buffered_file(cstring_view path, int oflag)
-    : file_(path, oflag), pos_(0) {}
+  ostream(cstring_view path, const detail::ostream_params& params)
+      : file_(path, params.oflag) {
+    set(new char[params.buffer_size], params.buffer_size);
+  }
 
-  ~direct_buffered_file() {
+ 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>
-  friend void print(direct_buffered_file& f, const S& format_str,
-                    const Args&... args) {
-    // We could avoid double buffering.
-    auto buf = fmt::memory_buffer();
-    fmt::format_to(std::back_inserter(buf), format_str, args...);
-    auto remaining_pos = 0;
-    auto remaining_size = buf.size();
-    while (remaining_size > detail::to_unsigned(f.free_capacity())) {
-      auto size = f.free_capacity();
-      memcpy(f.buffer_ + f.pos_, buf.data() + remaining_pos, size);
-      f.pos_ += size;
-      f.flush();
-      remaining_pos += size;
-      remaining_size -= size;
-    }
-    memcpy(f.buffer_ + f.pos_, buf.data() + remaining_pos, remaining_size);
-    f.pos_ += static_cast<int>(remaining_size);
+  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
diff --git a/include/fmt/ostream.h b/include/fmt/ostream.h
index c16107f7..29c58ec1 100644
--- a/include/fmt/ostream.h
+++ b/include/fmt/ostream.h
@@ -49,17 +49,27 @@ 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:
-  // Hide all operator<< from std::basic_ostream<Char>.
-  void_t<> operator<<(null<>);
-  void_t<> operator<<(const Char*);
-
-  template <typename T, FMT_ENABLE_IF(std::is_convertible<T, int>::value &&
-                                      !std::is_enum<T>::value)>
-  void_t<> operator<<(T);
+  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 {
@@ -103,7 +113,7 @@ void format_value(buffer<Char>& buf, const T& value,
 #endif
   output << value;
   output.exceptions(std::ios_base::failbit | std::ios_base::badbit);
-  buf.resize(buf.size());
+  buf.try_resize(buf.size());
 }
 
 // Formats an object of type T that has an overloaded ostream operator<<.
@@ -160,7 +170,7 @@ template <typename S, typename... Args,
           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),
-         detail::make_args_checked<Args...>(format_str, args...));
+         fmt::make_args_checked<Args...>(format_str, args...));
 }
 FMT_END_NAMESPACE
 
diff --git a/include/fmt/printf.h b/include/fmt/printf.h
index d4440ed1..8c28ac23 100644
--- a/include/fmt/printf.h
+++ b/include/fmt/printf.h
@@ -181,7 +181,7 @@ template <typename Char> class printf_width_handler {
 template <typename Char, typename Context>
 void vprintf(buffer<Char>& buf, basic_string_view<Char> format,
              basic_format_args<Context> args) {
-  Context(std::back_inserter(buf), format, args).format();
+  Context(buffer_appender<Char>(buf), format, args).format();
 }
 }  // namespace detail
 
@@ -598,7 +598,7 @@ OutputIt basic_printf_context<OutputIt, Char>::format() {
 
 template <typename Char>
 using basic_printf_context_t =
-    basic_printf_context<std::back_insert_iterator<detail::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>;
diff --git a/include/fmt/ranges.h b/include/fmt/ranges.h
index c48f1727..b603d637 100644
--- a/include/fmt/ranges.h
+++ b/include/fmt/ranges.h
@@ -157,6 +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&) {
@@ -182,7 +185,6 @@ 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 detail
 
 template <typename T> struct is_tuple_like {
@@ -246,9 +248,15 @@ template <typename T, typename Char> struct is_range {
       !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
+#endif
+                >> {
   formatting_range<Char> formatting;
 
   template <typename ParseContext>
@@ -257,8 +265,7 @@ struct formatter<RangeT, Char,
   }
 
   template <typename FormatContext>
-  typename FormatContext::iterator format(const RangeT& values,
-                                          FormatContext& ctx) {
+  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();
diff --git a/.github/pull_request_template.md b/pull_request_template.md
index 548e30f2..548e30f2 100644
--- a/.github/pull_request_template.md
+++ b/pull_request_template.md
diff --git a/src/format.cc b/src/format.cc
index a64a1f38..72713416 100644
--- a/src/format.cc
+++ b/src/format.cc
@@ -44,9 +44,9 @@ template FMT_API char detail::decimal_point_impl(locale_ref);
 
 template FMT_API void detail::buffer<char>::append(const char*, const char*);
 
-template FMT_API FMT_BUFFER_CONTEXT(char)::iterator detail::vformat_to(
+template FMT_API void detail::vformat_to(
     detail::buffer<char>&, string_view,
-    basic_format_args<FMT_BUFFER_CONTEXT(char)>);
+    basic_format_args<FMT_BUFFER_CONTEXT(char)>, detail::locale_ref);
 
 template FMT_API int detail::snprintf_float(double, int, detail::float_specs,
                                             detail::buffer<char>&);
diff --git a/src/os.cc b/src/os.cc
index 386119db..a07e7824 100644
--- a/src/os.cc
+++ b/src/os.cc
@@ -62,7 +62,7 @@ using RWResult = int;
 inline unsigned convert_rwcount(std::size_t count) {
   return count <= UINT_MAX ? static_cast<unsigned>(count) : UINT_MAX;
 }
-#else
+#elif FMT_USE_FCNTL
 // Return type of read and write functions.
 using RWResult = ssize_t;
 
@@ -124,7 +124,8 @@ void detail::format_windows_error(detail::buffer<char>& out, int error_code,
       if (result != 0) {
         utf16_to_utf8 utf8_message;
         if (utf8_message.convert(system_message) == ERROR_SUCCESS) {
-          format_to(std::back_inserter(out), "{}: {}", message, utf8_message);
+          format_to(buffer_appender<char>(out), "{}: {}", message,
+                    utf8_message);
           return;
         }
         break;
@@ -288,12 +289,12 @@ void file::pipe(file& read_end, file& write_end) {
 }
 
 buffered_file file::fdopen(const char* mode) {
-  // Don't retry as fdopen doesn't return EINTR.
-  #if defined(__MINGW32__) && defined(_POSIX_)
+// Don't retry as fdopen doesn't return EINTR.
+#  if defined(__MINGW32__) && defined(_POSIX_)
   FILE* f = ::fdopen(fd_, mode);
-  #else
+#  else
   FILE* f = FMT_POSIX_CALL(fdopen(fd_, mode));
-  #endif
+#  endif
   if (!f)
     FMT_THROW(
         system_error(errno, "cannot associate stream with file descriptor"));
@@ -313,5 +314,9 @@ long getpagesize() {
   return size;
 #  endif
 }
+
+void ostream::grow(size_t) {
+  if (this->size() == this->capacity()) flush();
+}
 #endif  // FMT_USE_FCNTL
 FMT_END_NAMESPACE
diff --git a/support/README b/support/README
index e7fbacc7..468f5485 100644
--- a/support/README
+++ b/support/README
@@ -2,5 +2,3 @@ This directory contains build support files such as
 
 * CMake modules
 * Build scripts
-* qmake (static build with dynamic libc only)
-
diff --git a/support/fmt.pro b/support/fmt.pro
deleted file mode 100644
index c555d0b4..00000000
--- a/support/fmt.pro
+++ /dev/null
@@ -1,27 +0,0 @@
-# Staticlib configuration for qmake builds
-# For some reason qmake 3.1 fails to identify source dependencies and excludes format.cc and printf.cc
-# from compilation so it _MUST_ be called as qmake -nodepend
-# A workaround is implemented below: a custom compiler is defined which does not track dependencies
-
-TEMPLATE = lib
-
-TARGET = fmt
-
-QMAKE_EXT_CPP = .cc
-
-CONFIG = staticlib warn_on c++11
-
-FMT_SOURCES = \
-    ../src/format.cc \
-    ../src/posix.cc
-
-fmt.name = libfmt
-fmt.input = FMT_SOURCES
-fmt.output = ${QMAKE_FILE_BASE}$$QMAKE_EXT_OBJ
-fmt.clean = ${QMAKE_FILE_BASE}$$QMAKE_EXT_OBJ
-fmt.depends = ${QMAKE_FILE_IN}
-# QMAKE_RUN_CXX will not be expanded
-fmt.commands = $$QMAKE_CXX -c $$QMAKE_CXXFLAGS $$QMAKE_CXXFLAGS_WARN_ON $$QMAKE_CXXFLAGS_RELEASE_WITH_DEBUGINFO $$QMAKE_CXXFLAGS_CXX11 ${QMAKE_FILE_IN}
-fmt.variable_out = OBJECTS
-fmt.CONFIG = no_dependencies no_link
-QMAKE_EXTRA_COMPILERS += fmt
diff --git a/support/update-coverity-branch.py b/support/update-coverity-branch.py
deleted file mode 100755
index 519f5d00..00000000
--- a/support/update-coverity-branch.py
+++ /dev/null
@@ -1,30 +0,0 @@
-#!/usr/bin/env python
-# Update the coverity branch from the master branch.
-# It is not done automatically because Coverity Scan limits
-# the number of submissions per day.
-
-from __future__ import print_function
-import shutil, tempfile
-from subprocess import check_output, STDOUT
-
-class Git:
-    def __init__(self, dir):
-        self.dir = dir
-
-    def __call__(self, *args):
-        output = check_output(['git'] + list(args), cwd=self.dir, stderr=STDOUT)
-        print(output)
-        return output
-
-dir = tempfile.mkdtemp()
-try:
-    git = Git(dir)
-    git('clone', '-b', 'coverity', 'git@github.com:fmtlib/fmt.git', dir)
-    output = git('merge', '-X', 'theirs', '--no-commit', 'origin/master')
-    if 'Fast-forward' not in output:
-        git('reset', 'HEAD', '.travis.yml')
-        git('checkout', '--', '.travis.yml')
-        git('commit', '-m', 'Update coverity branch')
-    git('push')
-finally:
-    shutil.rmtree(dir)
diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt
index 89176633..7ae5659d 100644
--- a/test/CMakeLists.txt
+++ b/test/CMakeLists.txt
@@ -91,8 +91,6 @@ add_fmt_test(assert-test)
 add_fmt_test(chrono-test)
 add_fmt_test(color-test)
 add_fmt_test(core-test)
-add_fmt_test(grisu-test)
-target_compile_definitions(grisu-test PRIVATE FMT_USE_GRISU=1)
 add_fmt_test(gtest-extra-test)
 add_fmt_test(format-test mock-allocator.h)
 if (MSVC)
@@ -105,11 +103,21 @@ add_fmt_test(locale-test)
 add_fmt_test(ostream-test)
 add_fmt_test(compile-test)
 add_fmt_test(printf-test)
-add_fmt_test(custom-formatter-test)
 add_fmt_test(ranges-test)
 add_fmt_test(scan-test)
 
-if (NOT MSVC_BUILD_STATIC)
+if (NOT DEFINED MSVC_STATIC_RUNTIME AND MSVC)
+  foreach (flag_var
+			 CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+			 CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
+	if (${flag_var} MATCHES "^(/|-)(MT|MTd)")
+	  set(MSVC_STATIC_RUNTIME ON)
+	  break()
+	endif()
+  endforeach()
+endif()
+
+if (NOT MSVC_STATIC_RUNTIME)
   add_fmt_executable(posix-mock-test
     posix-mock-test.cc ../src/format.cc ${TEST_MAIN_SRC})
   target_include_directories(
diff --git a/test/add-subdirectory-test/CMakeLists.txt b/test/add-subdirectory-test/CMakeLists.txt
index db7054bd..9cc4b0e9 100644
--- a/test/add-subdirectory-test/CMakeLists.txt
+++ b/test/add-subdirectory-test/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.1.0)
+cmake_minimum_required(VERSION 3.1...3.18)
 
 project(fmt-test)
 
diff --git a/test/assert-test.cc b/test/assert-test.cc
index bc728b53..70efa381 100644
--- a/test/assert-test.cc
+++ b/test/assert-test.cc
@@ -1,4 +1,8 @@
-// Formatting library for C++ - assertion tests
+// Formatting library for C++ - FMT_ASSERT test
+//
+// It is a separate test to minimize the number of EXPECT_DEBUG_DEATH checks
+// which are slow on some platforms. In other tests FMT_ASSERT is made to throw
+// an exception which is much faster and easier to check.
 //
 // Copyright (c) 2012 - present, Victor Zverovich
 // All rights reserved.
diff --git a/test/chrono-test.cc b/test/chrono-test.cc
index b876c151..fa383c14 100644
--- a/test/chrono-test.cc
+++ b/test/chrono-test.cc
@@ -95,6 +95,17 @@ TEST(TimeTest, GMTime) {
   EXPECT_TRUE(EqualTime(tm, fmt::gmtime(t)));
 }
 
+TEST(TimeTest, TimePoint) {
+  std::chrono::system_clock::time_point point = std::chrono::system_clock::now();
+
+  std::time_t t = std::chrono::system_clock::to_time_t(point);
+  std::tm tm = *std::localtime(&t);
+  char strftime_output[256];
+  std::strftime(strftime_output, sizeof(strftime_output), "It is %Y-%m-%d %H:%M:%S", &tm);
+
+  EXPECT_EQ(strftime_output, fmt::format("It is {:%Y-%m-%d %H:%M:%S}", point));
+}
+
 #define EXPECT_TIME(spec, time, duration)                 \
   {                                                       \
     std::locale loc("ja_JP.utf8");                        \
diff --git a/test/color-test.cc b/test/color-test.cc
index 454a0660..30738085 100644
--- a/test/color-test.cc
+++ b/test/color-test.cc
@@ -7,6 +7,10 @@
 
 #include "fmt/color.h"
 
+#include <iterator>
+#include <string>
+#include <utility>
+
 #include "gtest-extra.h"
 
 TEST(ColorsTest, ColorsPrint) {
@@ -84,3 +88,12 @@ TEST(ColorsTest, Format) {
   EXPECT_EQ(fmt::format(fg(fmt::terminal_color::red), "{}", "foo"),
             "\x1b[31mfoo\x1b[0m");
 }
+
+TEST(ColorsTest, FormatToOutAcceptsTextStyle) {
+  fmt::text_style ts = fg(fmt::rgb(255, 20, 30));
+  std::string out;
+  fmt::format_to(std::back_inserter(out), ts, "rgb(255,20,30){}{}{}", 1, 2, 3);
+
+  EXPECT_EQ(fmt::to_string(out),
+            "\x1b[38;2;255;020;030mrgb(255,20,30)123\x1b[0m");
+}
diff --git a/test/compile-error-test/CMakeLists.txt b/test/compile-error-test/CMakeLists.txt
index 75a0c5a5..8202f279 100644
--- a/test/compile-error-test/CMakeLists.txt
+++ b/test/compile-error-test/CMakeLists.txt
@@ -1,6 +1,6 @@
 # Test if compile errors are produced where necessary.
 
-cmake_minimum_required(VERSION 3.1.0)
+cmake_minimum_required(VERSION 3.1...3.18)
 
 include(CheckCXXSourceCompiles)
 include(CheckCXXCompilerFlag)
diff --git a/test/compile-test.cc b/test/compile-test.cc
index a47cea54..c0dda50c 100644
--- a/test/compile-test.cc
+++ b/test/compile-test.cc
@@ -5,19 +5,10 @@
 //
 // For the license information refer to format.h.
 
-#include <stdint.h>
-
-#include <cctype>
-#include <cfloat>
-#include <climits>
-#include <cmath>
-#include <cstring>
-#include <deque>
-#include <list>
-#include <memory>
 #include <string>
+#include <type_traits>
 
-// Check if fmt/compile.h compiles with windows.h included before it.
+// Check that fmt/compile.h compiles with windows.h included before it.
 #ifdef _WIN32
 #  include <windows.h>
 #endif
@@ -25,16 +16,8 @@
 #include "fmt/compile.h"
 #include "gmock.h"
 #include "gtest-extra.h"
-#include "mock-allocator.h"
 #include "util.h"
 
-#undef ERROR
-#undef min
-#undef max
-
-using testing::Return;
-using testing::StrictMock;
-
 // compiletime_prepared_parts_type_provider is useful only with relaxed
 // constexpr.
 #if FMT_USE_CONSTEXPR
@@ -114,20 +97,20 @@ TEST(CompileTest, MultipleTypes) {
   EXPECT_EQ(fmt::format(f, 42, 42), "42 42");
 }
 
-struct formattable {};
+struct test_formattable {};
 
 FMT_BEGIN_NAMESPACE
-template <> struct formatter<formattable> : formatter<const char*> {
+template <> struct formatter<test_formattable> : formatter<const char*> {
   template <typename FormatContext>
-  auto format(formattable, FormatContext& ctx) -> decltype(ctx.out()) {
+  auto format(test_formattable, FormatContext& ctx) -> decltype(ctx.out()) {
     return formatter<const char*>::format("foo", ctx);
   }
 };
 FMT_END_NAMESPACE
 
 TEST(CompileTest, FormatUserDefinedType) {
-  auto f = fmt::detail::compile<formattable>("{}");
-  EXPECT_EQ(fmt::format(f, formattable()), "foo");
+  auto f = fmt::detail::compile<test_formattable>("{}");
+  EXPECT_EQ(fmt::format(f, test_formattable()), "foo");
 }
 
 TEST(CompileTest, EmptyFormatString) {
@@ -146,21 +129,45 @@ TEST(CompileTest, FormatDefault) {
   EXPECT_EQ("4.2", fmt::format(FMT_COMPILE("{}"), 4.2));
   EXPECT_EQ("foo", fmt::format(FMT_COMPILE("{}"), "foo"));
   EXPECT_EQ("foo", fmt::format(FMT_COMPILE("{}"), std::string("foo")));
-  EXPECT_EQ("foo", fmt::format(FMT_COMPILE("{}"), formattable()));
+  EXPECT_EQ("foo", fmt::format(FMT_COMPILE("{}"), test_formattable()));
+}
+
+TEST(CompileTest, FormatWideString) {
+  EXPECT_EQ(L"42", fmt::format(FMT_COMPILE(L"{}"), 42));
 }
 
 TEST(CompileTest, FormatSpecs) {
   EXPECT_EQ("42", fmt::format(FMT_COMPILE("{:x}"), 0x42));
 }
 
+TEST(CompileTest, DynamicWidth) {
+  EXPECT_EQ("  42foo  ",
+            fmt::format(FMT_COMPILE("{:{}}{:{}}"), 42, 4, "foo", 5));
+}
+
 TEST(CompileTest, FormatTo) {
   char buf[8];
   auto end = fmt::format_to(buf, FMT_COMPILE("{}"), 42);
   *end = '\0';
   EXPECT_STREQ("42", buf);
+  end = fmt::format_to(buf, FMT_COMPILE("{:x}"), 42);
+  *end = '\0';
+  EXPECT_STREQ("2a", buf);
+}
+
+TEST(CompileTest, FormatToNWithCompileMacro) {
+  constexpr auto buffer_size = 8;
+  char buffer[buffer_size];
+  auto res = fmt::format_to_n(buffer, buffer_size, FMT_COMPILE("{}"), 42);
+  *res.out = '\0';
+  EXPECT_STREQ("42", buffer);
+  res = fmt::format_to_n(buffer, buffer_size, FMT_COMPILE("{:x}"), 42);
+  *res.out = '\0';
+  EXPECT_STREQ("2a", buffer);
 }
 
 TEST(CompileTest, TextAndArg) {
   EXPECT_EQ(">>>42<<<", fmt::format(FMT_COMPILE(">>>{}<<<"), 42));
+  EXPECT_EQ("42!", fmt::format(FMT_COMPILE("{}!"), 42));
 }
 #endif
diff --git a/test/core-test.cc b/test/core-test.cc
index 8a1ea83b..9d88070d 100644
--- a/test/core-test.cc
+++ b/test/core-test.cc
@@ -31,21 +31,16 @@
 using fmt::basic_format_arg;
 using fmt::string_view;
 using fmt::detail::buffer;
+using fmt::detail::make_arg;
 using fmt::detail::value;
 
 using testing::_;
+using testing::Invoke;
+using testing::Return;
 using testing::StrictMock;
 
-namespace {
-
 struct test_struct {};
 
-template <typename Context, typename T>
-basic_format_arg<Context> make_arg(const T& value) {
-  return fmt::detail::make_arg<Context>(value);
-}
-}  // namespace
-
 FMT_BEGIN_NAMESPACE
 template <typename Char> struct formatter<test_struct, Char> {
   template <typename ParseContext>
@@ -53,10 +48,7 @@ template <typename Char> struct formatter<test_struct, Char> {
     return ctx.begin();
   }
 
-  typedef std::back_insert_iterator<buffer<Char>> iterator;
-
-  auto format(test_struct, basic_format_context<iterator, char>& ctx)
-      -> decltype(ctx.out()) {
+  auto format(test_struct, format_context& ctx) -> decltype(ctx.out()) {
     const Char* test = "test";
     return std::copy_n(test, std::strlen(test), ctx.out());
   }
@@ -81,22 +73,22 @@ TEST(BufferTest, Nonmoveable) {
 }
 #endif
 
-// A test buffer with a dummy grow method.
-template <typename T> struct test_buffer : buffer<T> {
-  void grow(size_t capacity) { this->set(nullptr, capacity); }
-};
+TEST(BufferTest, Indestructible) {
+  static_assert(!std::is_destructible<fmt::detail::buffer<int>>(),
+                "buffer's destructor is protected");
+}
 
-template <typename T> struct mock_buffer : buffer<T> {
-  MOCK_METHOD1(do_grow, void(size_t capacity));
+template <typename T> struct mock_buffer final : buffer<T> {
+  MOCK_METHOD1(do_grow, size_t(size_t capacity));
 
-  void grow(size_t capacity) {
-    this->set(this->data(), capacity);
-    do_grow(capacity);
-  }
+  void grow(size_t capacity) { this->set(this->data(), do_grow(capacity)); }
 
-  mock_buffer() {}
-  mock_buffer(T* data) { this->set(data, 0); }
-  mock_buffer(T* data, size_t capacity) { this->set(data, capacity); }
+  mock_buffer(T* data = nullptr, size_t capacity = 0) {
+    this->set(data, capacity);
+    ON_CALL(*this, do_grow(_)).WillByDefault(Invoke([](size_t capacity) {
+      return capacity;
+    }));
+  }
 };
 
 TEST(BufferTest, Ctor) {
@@ -123,24 +115,6 @@ TEST(BufferTest, Ctor) {
   }
 }
 
-struct dying_buffer : test_buffer<int> {
-  MOCK_METHOD0(die, void());
-  ~dying_buffer() { die(); }
-
- private:
-  virtual void avoid_weak_vtable();
-};
-
-void dying_buffer::avoid_weak_vtable() {}
-
-TEST(BufferTest, VirtualDtor) {
-  typedef StrictMock<dying_buffer> stict_mock_buffer;
-  stict_mock_buffer* mock_buffer = new stict_mock_buffer();
-  EXPECT_CALL(*mock_buffer, die());
-  buffer<int>* buffer = mock_buffer;
-  delete buffer;
-}
-
 TEST(BufferTest, Access) {
   char data[10];
   mock_buffer<char> buffer(data, sizeof(data));
@@ -152,30 +126,40 @@ TEST(BufferTest, Access) {
   EXPECT_EQ(42, const_buffer[3]);
 }
 
-TEST(BufferTest, Resize) {
+TEST(BufferTest, TryResize) {
   char data[123];
   mock_buffer<char> buffer(data, sizeof(data));
   buffer[10] = 42;
   EXPECT_EQ(42, buffer[10]);
-  buffer.resize(20);
+  buffer.try_resize(20);
   EXPECT_EQ(20u, buffer.size());
   EXPECT_EQ(123u, buffer.capacity());
   EXPECT_EQ(42, buffer[10]);
-  buffer.resize(5);
+  buffer.try_resize(5);
   EXPECT_EQ(5u, buffer.size());
   EXPECT_EQ(123u, buffer.capacity());
   EXPECT_EQ(42, buffer[10]);
-  // Check if resize calls grow.
+  // Check if try_resize calls grow.
   EXPECT_CALL(buffer, do_grow(124));
-  buffer.resize(124);
+  buffer.try_resize(124);
   EXPECT_CALL(buffer, do_grow(200));
-  buffer.resize(200);
+  buffer.try_resize(200);
+}
+
+TEST(BufferTest, TryResizePartial) {
+  char data[10];
+  mock_buffer<char> buffer(data, sizeof(data));
+  EXPECT_CALL(buffer, do_grow(20)).WillOnce(Return(15));
+  buffer.try_resize(20);
+  EXPECT_EQ(buffer.capacity(), 15);
+  EXPECT_EQ(buffer.size(), 15);
 }
 
 TEST(BufferTest, Clear) {
-  test_buffer<char> buffer;
-  buffer.resize(20);
-  buffer.resize(0);
+  mock_buffer<char> buffer;
+  EXPECT_CALL(buffer, do_grow(20));
+  buffer.try_resize(20);
+  buffer.try_resize(0);
   EXPECT_EQ(static_cast<size_t>(0), buffer.size());
   EXPECT_EQ(20u, buffer.capacity());
 }
@@ -183,11 +167,11 @@ TEST(BufferTest, Clear) {
 TEST(BufferTest, Append) {
   char data[15];
   mock_buffer<char> buffer(data, 10);
-  const char* test = "test";
+  auto test = "test";
   buffer.append(test, test + 5);
   EXPECT_STREQ(test, &buffer[0]);
   EXPECT_EQ(5u, buffer.size());
-  buffer.resize(10);
+  buffer.try_resize(10);
   EXPECT_CALL(buffer, do_grow(12));
   buffer.append(test, test + 2);
   EXPECT_EQ('t', buffer[10]);
@@ -195,17 +179,31 @@ TEST(BufferTest, Append) {
   EXPECT_EQ(12u, buffer.size());
 }
 
+TEST(BufferTest, AppendPartial) {
+  char data[10];
+  mock_buffer<char> buffer(data, sizeof(data));
+  testing::InSequence seq;
+  EXPECT_CALL(buffer, do_grow(15)).WillOnce(Return(10));
+  EXPECT_CALL(buffer, do_grow(15)).WillOnce(Invoke([&buffer](size_t) {
+    EXPECT_EQ(fmt::string_view(buffer.data(), buffer.size()), "0123456789");
+    buffer.clear();
+    return 10;
+  }));
+  auto test = "0123456789abcde";
+  buffer.append(test, test + 15);
+}
+
 TEST(BufferTest, AppendAllocatesEnoughStorage) {
   char data[19];
   mock_buffer<char> buffer(data, 10);
-  const char* test = "abcdefgh";
-  buffer.resize(10);
+  auto test = "abcdefgh";
+  buffer.try_resize(10);
   EXPECT_CALL(buffer, do_grow(19));
   buffer.append(test, test + 9);
 }
 
 TEST(ArgTest, FormatArgs) {
-  fmt::format_args args;
+  auto args = fmt::format_args();
   EXPECT_FALSE(args.get(1));
 }
 
@@ -233,7 +231,7 @@ struct custom_context {
 };
 
 TEST(ArgTest, MakeValueWithCustomContext) {
-  test_struct t;
+  auto t = test_struct();
   fmt::detail::value<custom_context> arg(
       fmt::detail::arg_mapper<custom_context>().map(t));
   custom_context ctx = {false, fmt::format_parse_context("")};
@@ -255,10 +253,10 @@ FMT_END_NAMESPACE
 struct test_result {};
 
 template <typename T> struct mock_visitor {
-  template <typename U> struct result { typedef test_result type; };
+  template <typename U> struct result { using type = test_result; };
 
   mock_visitor() {
-    ON_CALL(*this, visit(_)).WillByDefault(testing::Return(test_result()));
+    ON_CALL(*this, visit(_)).WillByDefault(Return(test_result()));
   }
 
   MOCK_METHOD1_T(visit, test_result(T value));
@@ -272,10 +270,10 @@ template <typename T> struct mock_visitor {
   }
 };
 
-template <typename T> struct visit_type { typedef T Type; };
+template <typename T> struct visit_type { using type = T; };
 
-#define VISIT_TYPE(Type_, visit_type_) \
-  template <> struct visit_type<Type_> { typedef visit_type_ Type; }
+#define VISIT_TYPE(type_, visit_type_) \
+  template <> struct visit_type<type_> { using type = visit_type_; }
 
 VISIT_TYPE(signed char, int);
 VISIT_TYPE(unsigned char, unsigned);
@@ -294,36 +292,34 @@ VISIT_TYPE(unsigned long, unsigned long long);
   {                                                                           \
     testing::StrictMock<mock_visitor<decltype(expected)>> visitor;            \
     EXPECT_CALL(visitor, visit(expected));                                    \
-    typedef std::back_insert_iterator<buffer<Char>> iterator;                 \
+    using iterator = std::back_insert_iterator<buffer<Char>>;                 \
     fmt::visit_format_arg(                                                    \
         visitor, make_arg<fmt::basic_format_context<iterator, Char>>(value)); \
   }
 
 #define CHECK_ARG(value, typename_)                          \
   {                                                          \
-    typedef decltype(value) value_type;                      \
-    typename_ visit_type<value_type>::Type expected = value; \
+    using value_type = decltype(value);                      \
+    typename_ visit_type<value_type>::type expected = value; \
     CHECK_ARG_(char, expected, value)                        \
     CHECK_ARG_(wchar_t, expected, value)                     \
   }
 
 template <typename T> class NumericArgTest : public testing::Test {};
 
-typedef ::testing::Types<bool, signed char, unsigned char, signed,
-                         unsigned short, int, unsigned, long, unsigned long,
-                         long long, unsigned long long, float, double,
-                         long double>
-    Types;
-TYPED_TEST_CASE(NumericArgTest, Types);
+using types =
+    ::testing::Types<bool, signed char, unsigned char, signed, unsigned short,
+                     int, unsigned, long, unsigned long, long long,
+                     unsigned long long, float, double, long double>;
+TYPED_TEST_CASE(NumericArgTest, types);
 
 template <typename T>
-typename std::enable_if<std::is_integral<T>::value, T>::type test_value() {
+fmt::enable_if_t<std::is_integral<T>::value, T> test_value() {
   return static_cast<T>(42);
 }
 
 template <typename T>
-typename std::enable_if<std::is_floating_point<T>::value, T>::type
-test_value() {
+fmt::enable_if_t<std::is_floating_point<T>::value, T> test_value() {
   return static_cast<T>(4.2);
 }
 
@@ -345,7 +341,7 @@ TEST(ArgTest, StringArg) {
   const char* cstr = str;
   CHECK_ARG_(char, cstr, str);
 
-  string_view sref(str);
+  auto sref = string_view(str);
   CHECK_ARG_(char, sref, std::string(str));
 }
 
@@ -372,14 +368,14 @@ TEST(ArgTest, PointerArg) {
 struct check_custom {
   test_result operator()(
       fmt::basic_format_arg<fmt::format_context>::handle h) const {
-    struct test_buffer : fmt::detail::buffer<char> {
+    struct test_buffer final : fmt::detail::buffer<char> {
       char data[10];
       test_buffer() : fmt::detail::buffer<char>(data, 0, 10) {}
       void grow(size_t) {}
     } buffer;
-    fmt::detail::buffer<char>& base = buffer;
     fmt::format_parse_context parse_ctx("");
-    fmt::format_context ctx(std::back_inserter(base), fmt::format_args());
+    fmt::format_context ctx{fmt::detail::buffer_appender<char>(buffer),
+                            fmt::format_args()};
     h.format(parse_ctx, ctx);
     EXPECT_EQ("test", std::string(buffer.data, buffer.size()));
     return test_result();
@@ -388,10 +384,10 @@ struct check_custom {
 
 TEST(ArgTest, CustomArg) {
   test_struct test;
-  typedef mock_visitor<fmt::basic_format_arg<fmt::format_context>::handle>
-      visitor;
+  using visitor =
+      mock_visitor<fmt::basic_format_arg<fmt::format_context>::handle>;
   testing::StrictMock<visitor> v;
-  EXPECT_CALL(v, visit(_)).WillOnce(testing::Invoke(check_custom()));
+  EXPECT_CALL(v, visit(_)).WillOnce(Invoke(check_custom()));
   fmt::visit_format_arg(v, make_arg<fmt::format_context>(test));
 }
 
@@ -407,9 +403,7 @@ TEST(FormatDynArgsTest, Basic) {
   store.push_back(42);
   store.push_back("abc1");
   store.push_back(1.5f);
-
-  std::string result = fmt::vformat("{} and {} and {}", store);
-  EXPECT_EQ("42 and abc1 and 1.5", result);
+  EXPECT_EQ("42 and abc1 and 1.5", fmt::vformat("{} and {} and {}", store));
 }
 
 TEST(FormatDynArgsTest, StringsAndRefs) {
@@ -451,7 +445,6 @@ TEST(FormatDynArgsTest, CustomFormat) {
   ++c.i;
   store.push_back(std::cref(c));
   ++c.i;
-
   std::string result = fmt::vformat("{} and {} and {}", store);
   EXPECT_EQ("cust=0 and cust=1 and cust=3", result);
 }
@@ -459,8 +452,7 @@ TEST(FormatDynArgsTest, CustomFormat) {
 TEST(FormatDynArgsTest, NamedInt) {
   fmt::dynamic_format_arg_store<fmt::format_context> store;
   store.push_back(fmt::arg("a1", 42));
-  std::string result = fmt::vformat("{a1}", store);
-  EXPECT_EQ("42", result);
+  EXPECT_EQ("42", fmt::vformat("{a1}", store));
 }
 
 TEST(FormatDynArgsTest, NamedStrings) {
@@ -469,10 +461,7 @@ TEST(FormatDynArgsTest, NamedStrings) {
   store.push_back(fmt::arg("a1", str));
   store.push_back(fmt::arg("a2", std::cref(str)));
   str[0] = 'X';
-
-  std::string result = fmt::vformat("{a1} and {a2}", store);
-
-  EXPECT_EQ("1234567890 and X234567890", result);
+  EXPECT_EQ("1234567890 and X234567890", fmt::vformat("{a1} and {a2}", store));
 }
 
 TEST(FormatDynArgsTest, NamedArgByRef) {
@@ -494,7 +483,6 @@ TEST(FormatDynArgsTest, NamedArgByRef) {
   store.push_back(std::cref(a1));
 
   std::string result = fmt::vformat("{a1_} and {} and {} and {}", store);
-
   EXPECT_EQ("42 and abc and 1.5 and 42", result);
 }
 
@@ -507,7 +495,6 @@ TEST(FormatDynArgsTest, NamedCustomFormat) {
   ++c.i;
   store.push_back(fmt::arg("c_ref", std::cref(c)));
   ++c.i;
-
   std::string result = fmt::vformat("{c1} and {c2} and {c_ref}", store);
   EXPECT_EQ("cust=0 and cust=1 and cust=3", result);
 }
@@ -663,14 +650,14 @@ TEST(CoreTest, FormatterOverridesImplicitConversion) {
 
 namespace my_ns {
 template <typename Char> class my_string {
+ private:
+  std::basic_string<Char> s_;
+
  public:
   my_string(const Char* s) : s_(s) {}
   const Char* data() const FMT_NOEXCEPT { return s_.data(); }
   size_t length() const FMT_NOEXCEPT { return s_.size(); }
   operator const Char*() const { return s_.c_str(); }
-
- private:
-  std::basic_string<Char> s_;
 };
 
 template <typename Char>
@@ -748,7 +735,7 @@ struct implicitly_convertible_to_string_view {
   operator fmt::string_view() const { return "foo"; }
 };
 
-TEST(FormatterTest, FormatImplicitlyConvertibleToStringView) {
+TEST(CoreTest, FormatImplicitlyConvertibleToStringView) {
   EXPECT_EQ("foo", fmt::format("{}", implicitly_convertible_to_string_view()));
 }
 
@@ -758,7 +745,7 @@ struct explicitly_convertible_to_string_view {
   explicit operator fmt::string_view() const { return "foo"; }
 };
 
-TEST(FormatterTest, FormatExplicitlyConvertibleToStringView) {
+TEST(CoreTest, FormatExplicitlyConvertibleToStringView) {
   EXPECT_EQ("foo", fmt::format("{}", explicitly_convertible_to_string_view()));
 }
 
@@ -767,7 +754,7 @@ struct explicitly_convertible_to_std_string_view {
   explicit operator std::string_view() const { return "foo"; }
 };
 
-TEST(FormatterTest, FormatExplicitlyConvertibleToStdStringView) {
+TEST(CoreTest, FormatExplicitlyConvertibleToStdStringView) {
   EXPECT_EQ("foo",
             fmt::format("{}", explicitly_convertible_to_std_string_view()));
 }
@@ -781,6 +768,6 @@ struct disabled_rvalue_conversion {
   operator const char*() && = delete;
 };
 
-TEST(FormatterTest, DisabledRValueConversion) {
+TEST(CoreTest, DisabledRValueConversion) {
   EXPECT_EQ("foo", fmt::format("{}", disabled_rvalue_conversion()));
 }
diff --git a/test/custom-formatter-test.cc b/test/custom-formatter-test.cc
deleted file mode 100644
index 1b0c1e13..00000000
--- a/test/custom-formatter-test.cc
+++ /dev/null
@@ -1,58 +0,0 @@
-// Formatting library for C++ - custom argument formatter tests
-//
-// Copyright (c) 2012 - present, Victor Zverovich
-// All rights reserved.
-//
-// For the license information refer to format.h.
-
-#ifndef _CRT_SECURE_NO_WARNINGS
-#define _CRT_SECURE_NO_WARNINGS
-#endif
-
-#include "fmt/format.h"
-#include "gtest-extra.h"
-
-// MSVC 2013 is known to be broken.
-#if !FMT_MSC_VER || FMT_MSC_VER > 1800
-
-// A custom argument formatter that doesn't print `-` for floating-point values
-// rounded to 0.
-class custom_arg_formatter
-    : public fmt::detail::arg_formatter<fmt::format_context::iterator, char> {
- public:
-  using base = fmt::detail::arg_formatter<fmt::format_context::iterator, char>;
-
-  custom_arg_formatter(fmt::format_context& ctx,
-                       fmt::format_parse_context* parse_ctx,
-                       fmt::format_specs* s = nullptr,
-                       const char* = nullptr)
-      : base(ctx, parse_ctx, s) {}
-
-  using base::operator();
-
-  iterator operator()(double value) {
-    // Comparing a float to 0.0 is safe.
-    if (round(value * pow(10, specs()->precision)) == 0.0) value = 0;
-    return base::operator()(value);
-  }
-};
-
-std::string custom_vformat(fmt::string_view format_str, fmt::format_args args) {
-  fmt::memory_buffer buffer;
-  fmt::detail::buffer<char>& base = buffer;
-  // Pass custom argument formatter as a template arg to vwrite.
-  fmt::vformat_to<custom_arg_formatter>(std::back_inserter(base), format_str,
-                                        args);
-  return std::string(buffer.data(), buffer.size());
-}
-
-template <typename... Args>
-std::string custom_format(const char* format_str, const Args&... args) {
-  auto va = fmt::make_format_args(args...);
-  return custom_vformat(format_str, va);
-}
-
-TEST(CustomFormatterTest, Format) {
-  EXPECT_EQ("0.00", custom_format("{:.2f}", -.00001));
-}
-#endif
diff --git a/test/find-package-test/CMakeLists.txt b/test/find-package-test/CMakeLists.txt
index 51357557..93d686e6 100644
--- a/test/find-package-test/CMakeLists.txt
+++ b/test/find-package-test/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.1.0)
+cmake_minimum_required(VERSION 3.1...3.18)
 
 project(fmt-test)
 
diff --git a/test/format b/test/format
index 11c58b96..76ac5476 100644
--- a/test/format
+++ b/test/format
@@ -38,9 +38,9 @@ namespace std {
   
   template<class Out, class charT> class basic_format_context;
   using format_context = basic_format_context<
-    /* unspecified */ std::back_insert_iterator<fmt::detail::buffer<char>>, char>;
+    /* unspecified */ fmt::detail::buffer_appender<char>, char>;
   using wformat_context = basic_format_context<
-    /* unspecified */ std::back_insert_iterator<fmt::detail::buffer<wchar_t>>, wchar_t>;
+    /* unspecified */ fmt::detail::buffer_appender<wchar_t>, wchar_t>;
 
   template<class T, class charT = char> struct formatter {
     formatter() = delete;
@@ -714,7 +714,7 @@ string vformat(string_view fmt, format_args args) {
   fmt::detail::buffer<char>& buf = mbuf;
   using af = detail::arg_formatter<fmt::format_context::iterator, char>;
   detail::format_handler<af, char, format_context>
-    h(std::back_inserter(buf), fmt, args, {});
+    h(fmt::detail::buffer_appender<char>(buf), fmt, args, {});
   fmt::detail::parse_format_string<false>(fmt::to_string_view(fmt), h);
   return to_string(mbuf);
 }
diff --git a/test/format-dyn-args-test.cc b/test/format-dyn-args-test.cc
deleted file mode 100644
index acc5ef78..00000000
--- a/test/format-dyn-args-test.cc
+++ /dev/null
@@ -1,6 +0,0 @@
-// Copyright (c) 2020 Vladimir Solontsov
-// SPDX-License-Identifier: MIT Licence
-
-#include <fmt/core.h>
-
-#include "gtest-extra.h"
diff --git a/test/format-impl-test.cc b/test/format-impl-test.cc
index 10e7a7ae..1290c077 100644
--- a/test/format-impl-test.cc
+++ b/test/format-impl-test.cc
@@ -19,7 +19,10 @@
 #include "gtest-extra.h"
 #include "util.h"
 
-#undef max
+#ifdef _WIN32
+#  include <windows.h>
+#  undef max
+#endif
 
 using fmt::detail::bigint;
 using fmt::detail::fp;
@@ -186,29 +189,6 @@ template <bool is_iec559> void run_double_tests() {
 template <> void run_double_tests<true>() {
   // Construct from double.
   EXPECT_EQ(fp(1.23), fp(0x13ae147ae147aeu, -52));
-
-  // Compute boundaries:
-  fp value;
-  // Normalized & not power of 2 - equidistant boundaries:
-  auto b = value.assign_with_boundaries(1.23);
-  EXPECT_EQ(value, fp(0x0013ae147ae147ae, -52));
-  EXPECT_EQ(b.lower, 0x9d70a3d70a3d6c00);
-  EXPECT_EQ(b.upper, 0x9d70a3d70a3d7400);
-  // Normalized power of 2 - lower boundary is closer:
-  b = value.assign_with_boundaries(1.9807040628566084e+28);  // 2**94
-  EXPECT_EQ(value, fp(0x0010000000000000, 42));
-  EXPECT_EQ(b.lower, 0x7ffffffffffffe00);
-  EXPECT_EQ(b.upper, 0x8000000000000400);
-  // Smallest normalized double - equidistant boundaries:
-  b = value.assign_with_boundaries(2.2250738585072014e-308);
-  EXPECT_EQ(value, fp(0x0010000000000000, -1074));
-  EXPECT_EQ(b.lower, 0x7ffffffffffffc00);
-  EXPECT_EQ(b.upper, 0x8000000000000400);
-  // Subnormal - equidistant boundaries:
-  b = value.assign_with_boundaries(4.9406564584124654e-324);
-  EXPECT_EQ(value, fp(0x0000000000000001, -1074));
-  EXPECT_EQ(b.lower, 0x4000000000000000);
-  EXPECT_EQ(b.upper, 0xc000000000000000);
 }
 
 TEST(FPTest, DoubleTests) {
@@ -222,33 +202,6 @@ TEST(FPTest, Normalize) {
   EXPECT_EQ(-6, normalized.e);
 }
 
-TEST(FPTest, ComputeFloatBoundaries) {
-  struct {
-    double x, lower, upper;
-  } tests[] = {
-      // regular
-      {1.5f, 1.4999999403953552, 1.5000000596046448},
-      // boundary
-      {1.0f, 0.9999999701976776, 1.0000000596046448},
-      // min normal
-      {1.1754944e-38f, 1.1754942807573643e-38, 1.1754944208872107e-38},
-      // max subnormal
-      {1.1754942e-38f, 1.1754941406275179e-38, 1.1754942807573643e-38},
-      // min subnormal
-      {1e-45f, 7.006492321624085e-46, 2.1019476964872256e-45},
-  };
-  for (auto test : tests) {
-    fp vlower = normalize(fp(test.lower));
-    fp vupper = normalize(fp(test.upper));
-    vlower.f >>= vupper.e - vlower.e;
-    vlower.e = vupper.e;
-    fp value;
-    auto b = value.assign_float_with_boundaries(test.x);
-    EXPECT_EQ(vlower.f, b.lower);
-    EXPECT_EQ(vupper.f, b.upper);
-  }
-}
-
 TEST(FPTest, Multiply) {
   auto v = fp(123ULL << 32, 4) * fp(56ULL << 32, 7);
   EXPECT_EQ(v.f, 123u * 56u);
@@ -259,17 +212,55 @@ TEST(FPTest, Multiply) {
 }
 
 TEST(FPTest, GetCachedPower) {
-  typedef std::numeric_limits<double> limits;
+  using limits = std::numeric_limits<double>;
   for (auto exp = limits::min_exponent; exp <= limits::max_exponent; ++exp) {
     int dec_exp = 0;
     auto fp = fmt::detail::get_cached_power(exp, dec_exp);
-    EXPECT_LE(exp, fp.e);
-    int dec_exp_step = 8;
-    EXPECT_LE(fp.e, exp + dec_exp_step * log2(10));
-    EXPECT_DOUBLE_EQ(pow(10, dec_exp), ldexp(static_cast<double>(fp.f), fp.e));
+    bigint exact, cache(fp.f);
+    if (dec_exp >= 0) {
+      exact.assign_pow10(dec_exp);
+      if (fp.e <= 0)
+        exact <<= -fp.e;
+      else
+        cache <<= fp.e;
+      exact.align(cache);
+      cache.align(exact);
+      auto exact_str = fmt::format("{}", exact);
+      auto cache_str = fmt::format("{}", cache);
+      EXPECT_EQ(exact_str.size(), cache_str.size());
+      EXPECT_EQ(exact_str.substr(0, 15), cache_str.substr(0, 15));
+      int diff = cache_str[15] - exact_str[15];
+      if (diff == 1)
+        EXPECT_GT(exact_str[16], '8');
+      else
+        EXPECT_EQ(diff, 0);
+    } else {
+      cache.assign_pow10(-dec_exp);
+      cache *= fp.f + 1;  // Inexact check.
+      exact.assign(1);
+      exact <<= -fp.e;
+      exact.align(cache);
+      auto exact_str = fmt::format("{}", exact);
+      auto cache_str = fmt::format("{}", cache);
+      EXPECT_EQ(exact_str.size(), cache_str.size());
+      EXPECT_EQ(exact_str.substr(0, 16), cache_str.substr(0, 16));
+    }
   }
 }
 
+TEST(FPTest, DragonboxMaxK) {
+  using fmt::detail::dragonbox::floor_log10_pow2;
+  using float_info = fmt::detail::dragonbox::float_info<float>;
+  EXPECT_EQ(fmt::detail::const_check(float_info::max_k),
+            float_info::kappa - floor_log10_pow2(float_info::min_exponent -
+                                                 float_info::significand_bits));
+  using double_info = fmt::detail::dragonbox::float_info<double>;
+  EXPECT_EQ(
+      fmt::detail::const_check(double_info::max_k),
+      double_info::kappa - floor_log10_pow2(double_info::min_exponent -
+                                            double_info::significand_bits));
+}
+
 TEST(FPTest, GetRoundDirection) {
   using fmt::detail::get_round_direction;
   using fmt::detail::round_direction;
@@ -307,7 +298,7 @@ TEST(FPTest, FixedHandler) {
   EXPECT_THROW(handler().on_digit('0', 100, 100, 0, exp, false),
                assertion_failure);
   namespace digits = fmt::detail::digits;
-  EXPECT_EQ(handler(1).on_digit('0', 100, 10, 10, exp, false), digits::done);
+  EXPECT_EQ(handler(1).on_digit('0', 100, 10, 10, exp, false), digits::error);
   // Check that divisor - error doesn't overflow.
   EXPECT_EQ(handler(1).on_digit('0', 100, 10, 101, exp, false), digits::error);
   // Check that 2 * error doesn't overflow.
@@ -349,14 +340,6 @@ TEST(FormatTest, ArgConverter) {
   EXPECT_EQ(value, fmt::visit_format_arg(value_extractor<long long>(), arg));
 }
 
-TEST(FormatTest, FormatNegativeNaN) {
-  double nan = std::numeric_limits<double>::quiet_NaN();
-  if (std::signbit(-nan))
-    EXPECT_EQ("-nan", fmt::format("{}", -nan));
-  else
-    fmt::print("Warning: compiler doesn't handle negative NaN correctly");
-}
-
 TEST(FormatTest, StrError) {
   char* message = nullptr;
   char buffer[BUFFER_SIZE];
@@ -454,3 +437,10 @@ TEST(UtilTest, WriteFallbackUIntPtr) {
       fmt::detail::fallback_uintptr(reinterpret_cast<void*>(0xface)), nullptr);
   EXPECT_EQ(s, "0xface");
 }
+
+#ifdef _WIN32
+TEST(UtilTest, WriteConsoleSignature) {
+  decltype(WriteConsoleW)* p = fmt::detail::WriteConsoleW;
+  (void)p;
+}
+#endif
diff --git a/test/format-test.cc b/test/format-test.cc
index 978feef2..128b57a2 100644
--- a/test/format-test.cc
+++ b/test/format-test.cc
@@ -24,7 +24,6 @@
 // Check if fmt/format.h compiles with the X11 index macro defined.
 #define index(x, y) no nice things
 
-#include "fmt/color.h"
 #include "fmt/format.h"
 
 #undef index
@@ -147,6 +146,7 @@ TEST(IteratorTest, CountingIterator) {
   auto prev = it++;
   EXPECT_EQ(prev.count(), 0);
   EXPECT_EQ(it.count(), 1);
+  EXPECT_EQ((it + 41).count(), 42);
 }
 
 TEST(IteratorTest, TruncatingIterator) {
@@ -169,20 +169,22 @@ TEST(IteratorTest, TruncatingBackInserter) {
 }
 
 TEST(IteratorTest, IsOutputIterator) {
-  EXPECT_TRUE(fmt::detail::is_output_iterator<char*>::value);
-  EXPECT_FALSE(fmt::detail::is_output_iterator<const char*>::value);
-  EXPECT_FALSE(fmt::detail::is_output_iterator<std::string>::value);
-  EXPECT_TRUE(fmt::detail::is_output_iterator<
-              std::back_insert_iterator<std::string>>::value);
-  EXPECT_TRUE(fmt::detail::is_output_iterator<std::string::iterator>::value);
-  EXPECT_FALSE(
-      fmt::detail::is_output_iterator<std::string::const_iterator>::value);
-  EXPECT_FALSE(fmt::detail::is_output_iterator<std::list<char>>::value);
+  EXPECT_TRUE((fmt::detail::is_output_iterator<char*, char>::value));
+  EXPECT_FALSE((fmt::detail::is_output_iterator<const char*, char>::value));
+  EXPECT_FALSE((fmt::detail::is_output_iterator<std::string, char>::value));
   EXPECT_TRUE(
-      fmt::detail::is_output_iterator<std::list<char>::iterator>::value);
-  EXPECT_FALSE(
-      fmt::detail::is_output_iterator<std::list<char>::const_iterator>::value);
-  EXPECT_FALSE(fmt::detail::is_output_iterator<uint32_pair>::value);
+      (fmt::detail::is_output_iterator<std::back_insert_iterator<std::string>,
+                                       char>::value));
+  EXPECT_TRUE(
+      (fmt::detail::is_output_iterator<std::string::iterator, char>::value));
+  EXPECT_FALSE((fmt::detail::is_output_iterator<std::string::const_iterator,
+                                                char>::value));
+  EXPECT_FALSE((fmt::detail::is_output_iterator<std::list<char>, char>::value));
+  EXPECT_TRUE((
+      fmt::detail::is_output_iterator<std::list<char>::iterator, char>::value));
+  EXPECT_FALSE((fmt::detail::is_output_iterator<std::list<char>::const_iterator,
+                                                char>::value));
+  EXPECT_FALSE((fmt::detail::is_output_iterator<uint32_pair, char>::value));
 }
 
 TEST(MemoryBufferTest, Ctor) {
@@ -236,7 +238,7 @@ TEST(MemoryBufferTest, MoveCtorInlineBuffer) {
   std::allocator<char> alloc;
   basic_memory_buffer<char, 5, TestAllocator> buffer((TestAllocator(&alloc)));
   const char test[] = "test";
-  buffer.append(test, test + 4);
+  buffer.append(string_view(test, 4));
   check_move_buffer("test", buffer);
   // Adding one more character fills the inline buffer, but doesn't cause
   // dynamic allocation.
@@ -293,12 +295,8 @@ TEST(MemoryBufferTest, MoveAssignment) {
 
 TEST(MemoryBufferTest, Grow) {
   typedef allocator_ref<mock_allocator<int>> Allocator;
-  typedef basic_memory_buffer<int, 10, Allocator> Base;
   mock_allocator<int> alloc;
-  struct TestMemoryBuffer : Base {
-    TestMemoryBuffer(Allocator alloc) : Base(alloc) {}
-    void grow(size_t size) { Base::grow(size); }
-  } buffer((Allocator(&alloc)));
+  basic_memory_buffer<int, 10, Allocator> buffer((Allocator(&alloc)));
   buffer.resize(7);
   using fmt::detail::to_unsigned;
   for (int i = 0; i < 7; ++i) buffer[to_unsigned(i)] = i * i;
@@ -306,7 +304,7 @@ TEST(MemoryBufferTest, Grow) {
   int mem[20];
   mem[7] = 0xdead;
   EXPECT_CALL(alloc, allocate(20)).WillOnce(Return(mem));
-  buffer.grow(20);
+  buffer.try_reserve(20);
   EXPECT_EQ(20u, buffer.capacity());
   // Check if size elements have been copied
   for (int i = 0; i < 7; ++i) EXPECT_EQ(i * i, buffer[to_unsigned(i)]);
@@ -543,7 +541,6 @@ TEST(FormatterTest, ManyArgs) {
 TEST(FormatterTest, NamedArg) {
   EXPECT_EQ("1/a/A", format("{_1}/{a_}/{A_}", fmt::arg("a_", 'a'),
                             fmt::arg("A_", "A"), fmt::arg("_1", 1)));
-  EXPECT_THROW_MSG(format("{a}"), format_error, "argument not found");
   EXPECT_EQ(" -42", format("{0:{width}}", -42, fmt::arg("width", 4)));
   EXPECT_EQ("st", format("{0:.{precision}}", "str", fmt::arg("precision", 2)));
   EXPECT_EQ("1 2", format("{} {two}", 1, fmt::arg("two", 2)));
@@ -553,6 +550,8 @@ TEST(FormatterTest, NamedArg) {
                          fmt::arg("i", 0), fmt::arg("j", 0), fmt::arg("k", 0),
                          fmt::arg("l", 0), fmt::arg("m", 0), fmt::arg("n", 0),
                          fmt::arg("o", 0), fmt::arg("p", 0)));
+  EXPECT_THROW_MSG(format("{a}"), format_error, "argument not found");
+  EXPECT_THROW_MSG(format("{a}", 42), format_error, "argument not found");
 }
 
 TEST(FormatterTest, AutoArgIndex) {
@@ -581,8 +580,8 @@ TEST(FormatterTest, LeftAlign) {
   EXPECT_EQ("42   ", format("{0:<5}", 42ul));
   EXPECT_EQ("-42  ", format("{0:<5}", -42ll));
   EXPECT_EQ("42   ", format("{0:<5}", 42ull));
-  EXPECT_EQ("-42.0  ", format("{0:<7}", -42.0));
-  EXPECT_EQ("-42.0  ", format("{0:<7}", -42.0l));
+  EXPECT_EQ("-42  ", format("{0:<5}", -42.0));
+  EXPECT_EQ("-42  ", format("{0:<5}", -42.0l));
   EXPECT_EQ("c    ", format("{0:<5}", 'c'));
   EXPECT_EQ("abc  ", format("{0:<5}", "abc"));
   EXPECT_EQ("0xface  ", format("{0:<8}", reinterpret_cast<void*>(0xface)));
@@ -598,8 +597,8 @@ TEST(FormatterTest, RightAlign) {
   EXPECT_EQ("   42", format("{0:>5}", 42ul));
   EXPECT_EQ("  -42", format("{0:>5}", -42ll));
   EXPECT_EQ("   42", format("{0:>5}", 42ull));
-  EXPECT_EQ("  -42.0", format("{0:>7}", -42.0));
-  EXPECT_EQ("  -42.0", format("{0:>7}", -42.0l));
+  EXPECT_EQ("  -42", format("{0:>5}", -42.0));
+  EXPECT_EQ("  -42", format("{0:>5}", -42.0l));
   EXPECT_EQ("    c", format("{0:>5}", 'c'));
   EXPECT_EQ("  abc", format("{0:>5}", "abc"));
   EXPECT_EQ("  0xface", format("{0:>8}", reinterpret_cast<void*>(0xface)));
@@ -619,8 +618,8 @@ TEST(FormatterTest, CenterAlign) {
   EXPECT_EQ(" 42  ", format("{0:^5}", 42ul));
   EXPECT_EQ(" -42 ", format("{0:^5}", -42ll));
   EXPECT_EQ(" 42  ", format("{0:^5}", 42ull));
-  EXPECT_EQ(" -42.0 ", format("{0:^7}", -42.0));
-  EXPECT_EQ(" -42.0 ", format("{0:^7}", -42.0l));
+  EXPECT_EQ(" -42 ", format("{0:^5}", -42.0));
+  EXPECT_EQ(" -42 ", format("{0:^5}", -42.0l));
   EXPECT_EQ("  c  ", format("{0:^5}", 'c'));
   EXPECT_EQ(" abc  ", format("{0:^6}", "abc"));
   EXPECT_EQ(" 0xface ", format("{0:^8}", reinterpret_cast<void*>(0xface)));
@@ -638,8 +637,8 @@ TEST(FormatterTest, Fill) {
   EXPECT_EQ("***42", format("{0:*>5}", 42ul));
   EXPECT_EQ("**-42", format("{0:*>5}", -42ll));
   EXPECT_EQ("***42", format("{0:*>5}", 42ull));
-  EXPECT_EQ("**-42.0", format("{0:*>7}", -42.0));
-  EXPECT_EQ("**-42.0", format("{0:*>7}", -42.0l));
+  EXPECT_EQ("**-42", format("{0:*>5}", -42.0));
+  EXPECT_EQ("**-42", format("{0:*>5}", -42.0l));
   EXPECT_EQ("c****", format("{0:*<5}", 'c'));
   EXPECT_EQ("abc**", format("{0:*<5}", "abc"));
   EXPECT_EQ("**0xface", format("{0:*>8}", reinterpret_cast<void*>(0xface)));
@@ -647,7 +646,7 @@ TEST(FormatterTest, Fill) {
   EXPECT_EQ(std::string("\0\0\0*", 4), format(string_view("{:\0>4}", 6), '*'));
   EXPECT_EQ("жж42", format("{0:ж>4}", 42));
   EXPECT_THROW_MSG(format("{:\x80\x80\x80\x80\x80>}", 0), format_error,
-                   "invalid fill");
+                   "missing '}' in format string");
 }
 
 TEST(FormatterTest, PlusSign) {
@@ -662,8 +661,8 @@ TEST(FormatterTest, PlusSign) {
   EXPECT_EQ("+42", format("{0:+}", 42ll));
   EXPECT_THROW_MSG(format("{0:+}", 42ull), format_error,
                    "format specifier requires signed argument");
-  EXPECT_EQ("+42.0", format("{0:+}", 42.0));
-  EXPECT_EQ("+42.0", format("{0:+}", 42.0l));
+  EXPECT_EQ("+42", format("{0:+}", 42.0));
+  EXPECT_EQ("+42", format("{0:+}", 42.0l));
   EXPECT_THROW_MSG(format("{0:+", 'c'), format_error,
                    "missing '}' in format string");
   EXPECT_THROW_MSG(format("{0:+}", 'c'), format_error,
@@ -686,8 +685,8 @@ TEST(FormatterTest, MinusSign) {
   EXPECT_EQ("42", format("{0:-}", 42ll));
   EXPECT_THROW_MSG(format("{0:-}", 42ull), format_error,
                    "format specifier requires signed argument");
-  EXPECT_EQ("42.0", format("{0:-}", 42.0));
-  EXPECT_EQ("42.0", format("{0:-}", 42.0l));
+  EXPECT_EQ("42", format("{0:-}", 42.0));
+  EXPECT_EQ("42", format("{0:-}", 42.0l));
   EXPECT_THROW_MSG(format("{0:-", 'c'), format_error,
                    "missing '}' in format string");
   EXPECT_THROW_MSG(format("{0:-}", 'c'), format_error,
@@ -710,8 +709,8 @@ TEST(FormatterTest, SpaceSign) {
   EXPECT_EQ(" 42", format("{0: }", 42ll));
   EXPECT_THROW_MSG(format("{0: }", 42ull), format_error,
                    "format specifier requires signed argument");
-  EXPECT_EQ(" 42.0", format("{0: }", 42.0));
-  EXPECT_EQ(" 42.0", format("{0: }", 42.0l));
+  EXPECT_EQ(" 42", format("{0: }", 42.0));
+  EXPECT_EQ(" 42", format("{0: }", 42.0l));
   EXPECT_THROW_MSG(format("{0: ", 'c'), format_error,
                    "missing '}' in format string");
   EXPECT_THROW_MSG(format("{0: }", 'c'), format_error,
@@ -722,6 +721,12 @@ TEST(FormatterTest, SpaceSign) {
                    "format specifier requires numeric argument");
 }
 
+TEST(FormatterTest, SignNotTruncated) {
+  wchar_t format_str[] = {L'{', L':',
+                          '+' | (1 << fmt::detail::num_bits<char>()), L'}', 0};
+  EXPECT_THROW(format(format_str, 42), format_error);
+}
+
 TEST(FormatterTest, HashFlag) {
   EXPECT_EQ("42", format("{0:#}", 42));
   EXPECT_EQ("-42", format("{0:#}", -42));
@@ -760,8 +765,8 @@ TEST(FormatterTest, HashFlag) {
   EXPECT_EQ("-42.0", format("{0:#}", -42.0l));
   EXPECT_EQ("4.e+01", format("{:#.0e}", 42.0));
   EXPECT_EQ("0.", format("{:#.0f}", 0.01));
-  auto s = format("{:#.0f}", 0.5);  // MSVC's printf uses wrong rounding mode.
-  EXPECT_TRUE(s == "0." || s == "1.");
+  EXPECT_EQ("0.50", format("{:#.2g}", 0.5));
+  EXPECT_EQ("0.", format("{:#.0f}", 0.5));
   EXPECT_THROW_MSG(format("{0:#", 'c'), format_error,
                    "missing '}' in format string");
   EXPECT_THROW_MSG(format("{0:#}", 'c'), format_error,
@@ -780,8 +785,8 @@ TEST(FormatterTest, ZeroFlag) {
   EXPECT_EQ("00042", format("{0:05}", 42ul));
   EXPECT_EQ("-0042", format("{0:05}", -42ll));
   EXPECT_EQ("00042", format("{0:05}", 42ull));
-  EXPECT_EQ("-0042.0", format("{0:07}", -42.0));
-  EXPECT_EQ("-0042.0", format("{0:07}", -42.0l));
+  EXPECT_EQ("-000042", format("{0:07}", -42.0));
+  EXPECT_EQ("-000042", format("{0:07}", -42.0l));
   EXPECT_THROW_MSG(format("{0:0", 'c'), format_error,
                    "missing '}' in format string");
   EXPECT_THROW_MSG(format("{0:05}", 'c'), format_error,
@@ -954,6 +959,9 @@ TEST(FormatterTest, Precision) {
   EXPECT_EQ("123.", format("{:#.0f}", 123.0));
   EXPECT_EQ("1.23", format("{:.02f}", 1.234));
   EXPECT_EQ("0.001", format("{:.1g}", 0.001));
+  EXPECT_EQ("1019666400", format("{}", 1019666432.0f));
+  EXPECT_EQ("1e+01", format("{:.0e}", 9.5));
+  EXPECT_EQ("1.0e-34", fmt::format("{:.1e}", 1e-34));
 
   EXPECT_THROW_MSG(format("{0:.2}", reinterpret_cast<void*>(0xcafe)),
                    format_error,
@@ -1235,17 +1243,20 @@ TEST(FormatterTest, FormatConvertibleToLongLong) {
 }
 
 TEST(FormatterTest, FormatFloat) {
+  EXPECT_EQ("0", format("{}", 0.0f));
   EXPECT_EQ("392.500000", format("{0:f}", 392.5f));
 }
 
 TEST(FormatterTest, FormatDouble) {
+  EXPECT_EQ("0", format("{}", 0.0));
   check_unknown_types(1.2, "eEfFgGaAnL%", "double");
-  EXPECT_EQ("0.0", format("{:}", 0.0));
+  EXPECT_EQ("0", format("{:}", 0.0));
   EXPECT_EQ("0.000000", format("{:f}", 0.0));
   EXPECT_EQ("0", format("{:g}", 0.0));
   EXPECT_EQ("392.65", format("{:}", 392.65));
   EXPECT_EQ("392.65", format("{:g}", 392.65));
   EXPECT_EQ("392.65", format("{:G}", 392.65));
+  EXPECT_EQ("4.9014e+06", format("{:g}", 4.9014e6));
   EXPECT_EQ("392.650000", format("{:f}", 392.65));
   EXPECT_EQ("392.650000", format("{:F}", 392.65));
   EXPECT_EQ("42", format("{:L}", 42.0));
@@ -1272,17 +1283,42 @@ TEST(FormatterTest, PrecisionRounding) {
   EXPECT_EQ("1.000", format("{:.3f}", 0.9999));
   EXPECT_EQ("0.00123", format("{:.3}", 0.00123));
   EXPECT_EQ("0.1", format("{:.16g}", 0.1));
-  // Trigger rounding error in Grisu by a carefully chosen number.
-  auto n = 3788512123356.985352;
-  char buffer[64];
-  safe_sprintf(buffer, "%f", n);
-  EXPECT_EQ(buffer, format("{:f}", n));
+  EXPECT_EQ("1", fmt::format("{:.0}", 1.0));
+  EXPECT_EQ("225.51575035152063720",
+            fmt::format("{:.17f}", 225.51575035152064));
+  EXPECT_EQ("-761519619559038.2", fmt::format("{:.1f}", -761519619559038.2));
+  EXPECT_EQ("1.9156918820264798e-56",
+            fmt::format("{}", 1.9156918820264798e-56));
+  EXPECT_EQ("0.0000", fmt::format("{:.4f}", 7.2809479766055470e-15));
+
+  // Trigger a rounding error in Grisu by a specially chosen number.
+  EXPECT_EQ("3788512123356.985352", format("{:f}", 3788512123356.985352));
+}
+
+TEST(FormatterTest, PrettifyFloat) {
+  EXPECT_EQ("0.0001", fmt::format("{}", 1e-4));
+  EXPECT_EQ("1e-05", fmt::format("{}", 1e-5));
+  EXPECT_EQ("1000000000000000", fmt::format("{}", 1e15));
+  EXPECT_EQ("1e+16", fmt::format("{}", 1e16));
+  EXPECT_EQ("9.999e-05", fmt::format("{}", 9.999e-5));
+  EXPECT_EQ("10000000000", fmt::format("{}", 1e10));
+  EXPECT_EQ("100000000000", fmt::format("{}", 1e11));
+  EXPECT_EQ("12340000000", fmt::format("{}", 1234e7));
+  EXPECT_EQ("12.34", fmt::format("{}", 1234e-2));
+  EXPECT_EQ("0.001234", fmt::format("{}", 1234e-6));
+  EXPECT_EQ("0.1", fmt::format("{}", 0.1f));
+  EXPECT_EQ("0.10000000149011612", fmt::format("{}", double(0.1f)));
+  EXPECT_EQ("1.3563156e-19", fmt::format("{}", 1.35631564e-19f));
 }
 
 TEST(FormatterTest, FormatNaN) {
   double nan = std::numeric_limits<double>::quiet_NaN();
   EXPECT_EQ("nan", format("{}", nan));
   EXPECT_EQ("+nan", format("{:+}", nan));
+  if (std::signbit(-nan))
+    EXPECT_EQ("-nan", format("{}", -nan));
+  else
+    fmt::print("Warning: compiler doesn't handle negative NaN correctly");
   EXPECT_EQ(" nan", format("{: }", nan));
   EXPECT_EQ("NAN", format("{:F}", nan));
   EXPECT_EQ("nan    ", format("{:<7}", nan));
@@ -1303,7 +1339,7 @@ TEST(FormatterTest, FormatInfinity) {
 }
 
 TEST(FormatterTest, FormatLongDouble) {
-  EXPECT_EQ("0.0", format("{0:}", 0.0l));
+  EXPECT_EQ("0", format("{0:}", 0.0l));
   EXPECT_EQ("0.000000", format("{0:f}", 0.0l));
   EXPECT_EQ("392.65", format("{0:}", 392.65l));
   EXPECT_EQ("392.65", format("{0:g}", 392.65l));
@@ -1520,6 +1556,7 @@ TEST(FormatterTest, WideFormatString) {
   EXPECT_EQ(L"4.2", format(L"{}", 4.2));
   EXPECT_EQ(L"abc", format(L"{}", L"abc"));
   EXPECT_EQ(L"z", format(L"{}", L'z'));
+  EXPECT_THROW(fmt::format(L"{:*\x343E}", 42), fmt::format_error);
 }
 
 TEST(FormatterTest, FormatStringFromSpeedTest) {
@@ -1805,59 +1842,6 @@ TEST(FormatTest, StrongEnum) {
 }
 #endif
 
-using buffer_iterator = fmt::format_context::iterator;
-
-class mock_arg_formatter
-    : public fmt::detail::arg_formatter_base<buffer_iterator, char> {
- private:
-#if FMT_USE_INT128
-  MOCK_METHOD1(call, void(__int128_t value));
-#else
-  MOCK_METHOD1(call, void(long long value));
-#endif
-
- public:
-  using base = fmt::detail::arg_formatter_base<buffer_iterator, char>;
-
-  mock_arg_formatter(fmt::format_context& ctx, fmt::format_parse_context*,
-                     fmt::format_specs* s = nullptr, const char* = nullptr)
-      : base(ctx.out(), s, ctx.locale()) {
-    EXPECT_CALL(*this, call(42));
-  }
-
-  template <typename T>
-  typename std::enable_if<fmt::detail::is_integral<T>::value, iterator>::type
-  operator()(T value) {
-    call(value);
-    return base::operator()(value);
-  }
-
-  template <typename T>
-  typename std::enable_if<!fmt::detail::is_integral<T>::value, iterator>::type
-  operator()(T value) {
-    return base::operator()(value);
-  }
-
-  iterator operator()(fmt::basic_format_arg<fmt::format_context>::handle) {
-    return base::operator()(fmt::monostate());
-  }
-};
-
-static void custom_vformat(fmt::string_view format_str, fmt::format_args args) {
-  fmt::memory_buffer buffer;
-  fmt::detail::buffer<char>& base = buffer;
-  fmt::vformat_to<mock_arg_formatter>(std::back_inserter(base), format_str,
-                                      args);
-}
-
-template <typename... Args>
-void custom_format(const char* format_str, const Args&... args) {
-  auto va = fmt::make_format_args(args...);
-  return custom_vformat(format_str, va);
-}
-
-TEST(FormatTest, CustomArgFormatter) { custom_format("{}", 42); }
-
 TEST(FormatTest, NonNullTerminatedFormatString) {
   EXPECT_EQ("42", format(string_view("{}foo", 2), 42));
 }
@@ -1946,6 +1930,12 @@ TEST(FormatTest, FormattedSize) {
   EXPECT_EQ(2u, fmt::formatted_size("{}", 42));
 }
 
+TEST(FormatTest, FormatTo) {
+  std::vector<char> v;
+  fmt::format_to(std::back_inserter(v), "{}", "foo");
+  EXPECT_EQ(string_view(v.data(), v.size()), "foo");
+}
+
 TEST(FormatTest, FormatToN) {
   char buffer[4];
   buffer[3] = 'x';
@@ -2412,12 +2402,6 @@ TEST(FormatTest, FmtStringInTemplate) {
 
 #endif  // FMT_USE_CONSTEXPR
 
-TEST(FormatTest, EmphasisNonHeaderOnly) {
-  // Ensure this compiles even if FMT_HEADER_ONLY is not defined.
-  EXPECT_EQ(fmt::format(fmt::emphasis::bold, "bold error"),
-            "\x1b[1mbold error\x1b[0m");
-}
-
 TEST(FormatTest, CharTraitsIsNotAmbiguous) {
   // Test that we don't inject detail names into the std namespace.
   using namespace std;
@@ -2430,24 +2414,26 @@ TEST(FormatTest, CharTraitsIsNotAmbiguous) {
 #endif
 }
 
-struct mychar {
+struct custom_char {
   int value;
-  mychar() = default;
+  custom_char() = default;
 
-  template <typename T> mychar(T val) : value(static_cast<int>(val)) {}
+  template <typename T> custom_char(T val) : value(static_cast<int>(val)) {}
 
   operator int() const { return value; }
 };
 
+int to_ascii(custom_char c) { return c; }
+
 FMT_BEGIN_NAMESPACE
-template <> struct is_char<mychar> : std::true_type {};
+template <> struct is_char<custom_char> : std::true_type {};
 FMT_END_NAMESPACE
 
 TEST(FormatTest, FormatCustomChar) {
-  const mychar format[] = {'{', '}', 0};
-  auto result = fmt::format(format, mychar('x'));
+  const custom_char format[] = {'{', '}', 0};
+  auto result = fmt::format(format, custom_char('x'));
   EXPECT_EQ(result.size(), 1);
-  EXPECT_EQ(result[0], mychar('x'));
+  EXPECT_EQ(result[0], custom_char('x'));
 }
 
 // Convert a char8_t string to std::string. Otherwise GTest will insist on
@@ -2466,3 +2452,27 @@ TEST(FormatTest, FormatUTF8Precision) {
   EXPECT_EQ(result.size(), 5);
   EXPECT_EQ(from_u8str(result), from_u8str(str.substr(0, 5)));
 }
+
+struct check_back_appender {};
+
+FMT_BEGIN_NAMESPACE
+template <> struct formatter<check_back_appender> {
+  template <typename ParseContext>
+  auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return ctx.begin();
+  }
+
+  template <typename Context>
+  auto format(check_back_appender, Context& ctx) -> decltype(ctx.out()) {
+    auto out = ctx.out();
+    static_assert(std::is_same<decltype(++out), decltype(out)&>::value,
+                  "needs to satisfy weakly_incrementable");
+    *out = 'y';
+    return ++out;
+  }
+};
+FMT_END_NAMESPACE
+
+TEST(FormatTest, BackInsertSlicing) {
+  EXPECT_EQ(fmt::format("{}", check_back_appender{}), "y");
+}
diff --git a/test/fuzzing/CMakeLists.txt b/test/fuzzing/CMakeLists.txt
index 31344fc5..2f716d83 100644
--- a/test/fuzzing/CMakeLists.txt
+++ b/test/fuzzing/CMakeLists.txt
@@ -1,38 +1,30 @@
 # Copyright (c) 2019, Paul Dreik
 # License: see LICENSE.rst in the fmt root directory
 
-# settings this links in a main. useful for reproducing,
-# kcov, gdb, afl, valgrind.
-# (note that libFuzzer can also reproduce, just pass it the files)
-option(FMT_FUZZ_LINKMAIN "enables the reproduce mode, instead of libFuzzer" On)
+# Link in the main function. Useful for reproducing, kcov, gdb, afl, valgrind.
+# (Note that libFuzzer can also reproduce, just pass it the files.)
+option(FMT_FUZZ_LINKMAIN "Enables the reproduce mode, instead of libFuzzer" On)
 
 # For oss-fuzz - insert $LIB_FUZZING_ENGINE into the link flags, but only for
-# the fuzz targets, otherwise the cmake configuration step fails.
+# the fuzz targets, otherwise the CMake configuration step fails.
 set(FMT_FUZZ_LDFLAGS "" CACHE STRING "LDFLAGS for the fuzz targets")
 
-# Find all fuzzers.
-set(SOURCES
-  chrono_duration.cpp
-  named_arg.cpp
-  one_arg.cpp
-  sprintf.cpp
-  two_args.cpp
-)
-
-macro(implement_fuzzer sourcefile)
-  get_filename_component(basename ${sourcefile} NAME_WE)
-  set(name fuzzer_${basename})
-  add_executable(${name} ${sourcefile} fuzzer_common.h)
+# Adds a binary for reproducing, i.e. no fuzzing, just enables replaying data
+# through the fuzzers.
+function(add_fuzzer source)
+  get_filename_component(basename ${source} NAME_WE)
+  set(name ${basename}-fuzzer)
+  add_executable(${name} ${source} fuzzer-common.h)
   if (FMT_FUZZ_LINKMAIN)
-      target_sources(${name} PRIVATE main.cpp)
+    target_sources(${name} PRIVATE main.cc)
   endif ()
   target_link_libraries(${name} PRIVATE fmt)
-if (FMT_FUZZ_LDFLAGS)
-  target_link_libraries(${name} PRIVATE ${FMT_FUZZ_LDFLAGS})
-endif ()
+  if (FMT_FUZZ_LDFLAGS)
+    target_link_libraries(${name} PRIVATE ${FMT_FUZZ_LDFLAGS})
+  endif ()
   target_compile_features(${name} PRIVATE cxx_generic_lambdas)
-endmacro ()
+endfunction()
 
-foreach (X IN ITEMS ${SOURCES})
-  implement_fuzzer(${X})
+foreach (source chrono-duration.cc float.cc named-arg.cc one-arg.cc two-args.cc)
+  add_fuzzer(${source})
 endforeach ()
diff --git a/test/fuzzing/README.md b/test/fuzzing/README.md
index 8f7a4536..bb3d0e04 100644
--- a/test/fuzzing/README.md
+++ b/test/fuzzing/README.md
@@ -1,27 +1,4 @@
-# FMT Fuzzer
-
-Fuzzing has revealed [several bugs](https://github.com/fmtlib/fmt/issues?&q=is%3Aissue+fuzz)
-in fmt. It is a part of the continous fuzzing at
-[oss-fuzz](https://github.com/google/oss-fuzz).
-
-The source code is modified to make the fuzzing possible without locking up on
-resource exhaustion:
-```cpp
-#ifdef FMT_FUZZ
-if(spec.precision>100000) {
-  throw std::runtime_error("fuzz mode - avoiding large precision");
-}
-#endif
-``` 
-This macro `FMT_FUZZ` is enabled on OSS-Fuzz builds and makes fuzzing
-practically possible. It is used in fmt code to prevent resource exhaustion in
-fuzzing mode.  
-The macro `FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION` is the
-defacto standard for making fuzzing practically possible to disable certain
-fuzzing-unfriendly features (for example, randomness), see [the libFuzzer
-documentation](https://llvm.org/docs/LibFuzzer.html#fuzzer-friendly-build-mode).
-
-## Running the fuzzers locally
+# Running the fuzzers locally
 
 There is a [helper script](build.sh) to build the fuzzers, which has only been
 tested on Debian and Ubuntu linux so far. There should be no problems fuzzing on
@@ -34,7 +11,7 @@ mkdir build
 cd build
 export CXX=clang++
 export CXXFLAGS="-fsanitize=fuzzer-no-link -DFUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION= -g"
-cmake ..  -DFMT_SAFE_DURATION_CAST=On -DFMT_FUZZ=On -DFMT_FUZZ_LINKMAIN=Off -DFMT_FUZZ_LDFLAGS="-fsanitize=fuzzer"
+cmake .. -DFMT_SAFE_DURATION_CAST=On -DFMT_FUZZ=On -DFMT_FUZZ_LINKMAIN=Off -DFMT_FUZZ_LDFLAGS="-fsanitize=fuzzer"
 cmake --build .
 ```
 should work to build the fuzzers for all platforms which clang supports.
@@ -44,5 +21,5 @@ Execute a fuzzer with for instance
 cd build
 export UBSAN_OPTIONS=halt_on_error=1
 mkdir out_chrono
-bin/fuzzer_chrono_duration  out_chrono
+bin/fuzzer_chrono_duration out_chrono
 ```
diff --git a/test/fuzzing/build.sh b/test/fuzzing/build.sh
index 141a50d9..28c50633 100755
--- a/test/fuzzing/build.sh
+++ b/test/fuzzing/build.sh
@@ -1,7 +1,6 @@
 #!/bin/sh
 #
 # Creates fuzzer builds of various kinds
-# - reproduce mode (no fuzzing, just enables replaying data through the fuzzers)
 # - oss-fuzz emulated mode (makes sure a simulated invocation by oss-fuzz works)
 # - libFuzzer build (you will need clang)
 # - afl build (you will need afl)
@@ -9,7 +8,7 @@
 #
 # Copyright (c) 2019 Paul Dreik
 #
-# License: see LICENSE.rst in the fmt root directory
+# For the license information refer to format.h.
 
 set -e
 me=$(basename $0)
@@ -23,16 +22,7 @@ here=$(pwd)
 CXXFLAGSALL="-DFUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION= -g"
 CMAKEFLAGSALL="$root -GNinja -DCMAKE_BUILD_TYPE=Debug -DFMT_DOC=Off -DFMT_TEST=Off -DFMT_FUZZ=On -DCMAKE_CXX_STANDARD=17"
 
-#builds the fuzzers as one would do if using afl or just making
-#binaries for reproducing.
-builddir=$here/build-fuzzers-reproduce
-mkdir -p $builddir
-cd $builddir
-CXX="ccache g++" CXXFLAGS="$CXXFLAGSALL" cmake \
-$CMAKEFLAGSALL
-cmake --build $builddir
-
-#for performance analysis of the fuzzers
+# For performance analysis of the fuzzers.
 builddir=$here/build-fuzzers-perfanalysis
 mkdir -p $builddir
 cd $builddir
@@ -43,7 +33,7 @@ $CMAKEFLAGSALL \
 
 cmake --build $builddir
 
-#builds the fuzzers as oss-fuzz does
+# Builds the fuzzers as oss-fuzz does.
 builddir=$here/build-fuzzers-ossfuzz
 mkdir -p $builddir
 cd $builddir
@@ -56,7 +46,7 @@ cmake $CMAKEFLAGSALL \
 cmake --build $builddir
 
 
-#builds fuzzers for local fuzzing with libfuzzer with asan+usan
+# Builds fuzzers for local fuzzing with libfuzzer with asan+usan.
 builddir=$here/build-fuzzers-libfuzzer
 mkdir -p $builddir
 cd $builddir
@@ -68,19 +58,7 @@ cmake $CMAKEFLAGSALL \
 
 cmake --build $builddir
 
-#builds fuzzers for local fuzzing with libfuzzer with asan only
-builddir=$here/build-fuzzers-libfuzzer-addr
-mkdir -p $builddir
-cd $builddir
-CXX="clang++" \
-CXXFLAGS="$CXXFLAGSALL -fsanitize=fuzzer-no-link,undefined" cmake \
-cmake $CMAKEFLAGSALL \
--DFMT_FUZZ_LINKMAIN=Off \
--DFMT_FUZZ_LDFLAGS="-fsanitize=fuzzer"
-
-cmake --build $builddir
-
-#builds a fast fuzzer for making coverage fast
+# Builds a fast fuzzer for making coverage fast.
 builddir=$here/build-fuzzers-fast
 mkdir -p $builddir
 cd $builddir
@@ -94,7 +72,7 @@ cmake $CMAKEFLAGSALL \
 cmake --build $builddir
 
 
-#builds fuzzers for local fuzzing with afl
+# Builds fuzzers for local fuzzing with afl.
 builddir=$here/build-fuzzers-afl
 mkdir -p $builddir
 cd $builddir
diff --git a/test/fuzzing/chrono-duration.cc b/test/fuzzing/chrono-duration.cc
new file mode 100644
index 00000000..fdad9894
--- /dev/null
+++ b/test/fuzzing/chrono-duration.cc
@@ -0,0 +1,135 @@
+// Copyright (c) 2019, Paul Dreik
+// For the license information refer to format.h.
+
+#include <cstdint>
+#include <fmt/chrono.h>
+
+#include "fuzzer-common.h"
+
+template <typename Period, typename Rep>
+void invoke_inner(fmt::string_view format_str, Rep rep) {
+  auto value = std::chrono::duration<Rep, Period>(rep);
+  try {
+#if FMT_FUZZ_FORMAT_TO_STRING
+    std::string message = fmt::format(format_str, value);
+#else
+    fmt::memory_buffer buf;
+    fmt::format_to(buf, format_str, value);
+#endif
+  } catch (std::exception&) {
+  }
+}
+
+// Rep is a duration's representation type.
+template <typename Rep>
+void invoke_outer(const uint8_t* data, size_t size, int period) {
+  // Always use a fixed location of the data.
+  static_assert(sizeof(Rep) <= fixed_size, "fixed size is too small");
+  if (size <= fixed_size + 1) return;
+
+  const Rep rep = assign_from_buf<Rep>(data);
+  data += fixed_size;
+  size -= fixed_size;
+
+  // data is already allocated separately in libFuzzer so reading past the end 
+  // will most likely be detected anyway.
+  const auto format_str = fmt::string_view(as_chars(data), size);
+
+  // yocto, zepto, zetta and yotta are not handled.
+  switch (period) {
+  case 1:
+    invoke_inner<std::atto>(format_str, rep);
+    break;
+  case 2:
+    invoke_inner<std::femto>(format_str, rep);
+    break;
+  case 3:
+    invoke_inner<std::pico>(format_str, rep);
+    break;
+  case 4:
+    invoke_inner<std::nano>(format_str, rep);
+    break;
+  case 5:
+    invoke_inner<std::micro>(format_str, rep);
+    break;
+  case 6:
+    invoke_inner<std::milli>(format_str, rep);
+    break;
+  case 7:
+    invoke_inner<std::centi>(format_str, rep);
+    break;
+  case 8:
+    invoke_inner<std::deci>(format_str, rep);
+    break;
+  case 9:
+    invoke_inner<std::deca>(format_str, rep);
+    break;
+  case 10:
+    invoke_inner<std::kilo>(format_str, rep);
+    break;
+  case 11:
+    invoke_inner<std::mega>(format_str, rep);
+    break;
+  case 12:
+    invoke_inner<std::giga>(format_str, rep);
+    break;
+  case 13:
+    invoke_inner<std::tera>(format_str, rep);
+    break;
+  case 14:
+    invoke_inner<std::peta>(format_str, rep);
+    break;
+  case 15:
+    invoke_inner<std::exa>(format_str, rep);
+    break;
+  }
+}
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) {
+  if (size <= 4)  return 0;
+
+  const auto representation = data[0];
+  const auto period = data[1];
+  data += 2;
+  size -= 2;
+
+  switch (representation) {
+  case 1:
+    invoke_outer<char>(data, size, period);
+    break;
+  case 2:
+    invoke_outer<signed char>(data, size, period);
+    break;
+  case 3:
+    invoke_outer<unsigned char>(data, size, period);
+    break;
+  case 4:
+    invoke_outer<short>(data, size, period);
+    break;
+  case 5:
+    invoke_outer<unsigned short>(data, size, period);
+    break;
+  case 6:
+    invoke_outer<int>(data, size, period);
+    break;
+  case 7:
+    invoke_outer<unsigned int>(data, size, period);
+    break;
+  case 8:
+    invoke_outer<long>(data, size, period);
+    break;
+  case 9:
+    invoke_outer<unsigned long>(data, size, period);
+    break;
+  case 10:
+    invoke_outer<float>(data, size, period);
+    break;
+  case 11:
+    invoke_outer<double>(data, size, period);
+    break;
+  case 12:
+    invoke_outer<long double>(data, size, period);
+    break;
+  }
+  return 0;
+}
diff --git a/test/fuzzing/chrono_duration.cpp b/test/fuzzing/chrono_duration.cpp
deleted file mode 100644
index 3f25f6be..00000000
--- a/test/fuzzing/chrono_duration.cpp
+++ /dev/null
@@ -1,152 +0,0 @@
-// Copyright (c) 2019, Paul Dreik
-// License: see LICENSE.rst in the fmt root directory
-
-#include <fmt/chrono.h>
-#include <cstdint>
-#include <limits>
-#include <stdexcept>
-#include <type_traits>
-#include <vector>
-#include "fuzzer_common.h"
-
-template <typename Item, typename Ratio>
-void invoke_inner(fmt::string_view formatstring, const Item item) {
-  const std::chrono::duration<Item, Ratio> value(item);
-  try {
-#if FMT_FUZZ_FORMAT_TO_STRING
-    std::string message = fmt::format(formatstring, value);
-#else
-    fmt::memory_buffer buf;
-    fmt::format_to(buf, formatstring, value);
-#endif
-  } catch (std::exception& /*e*/) {
-  }
-}
-
-// Item is the underlying type for duration (int, long etc)
-template <typename Item>
-void invoke_outer(const uint8_t* Data, size_t Size, const int scaling) {
-  // always use a fixed location of the data
-  using fmt_fuzzer::Nfixed;
-
-  constexpr auto N = sizeof(Item);
-  static_assert(N <= Nfixed, "fixed size is too small");
-  if (Size <= Nfixed + 1) {
-    return;
-  }
-
-  const Item item = fmt_fuzzer::assignFromBuf<Item>(Data);
-
-  // fast forward
-  Data += Nfixed;
-  Size -= Nfixed;
-
-  // Data is already allocated separately in libFuzzer so reading past
-  // the end will most likely be detected anyway
-  const auto formatstring = fmt::string_view(fmt_fuzzer::as_chars(Data), Size);
-
-  // doit_impl<Item,std::yocto>(buf.data(),item);
-  // doit_impl<Item,std::zepto>(buf.data(),item);
-  switch (scaling) {
-  case 1:
-    invoke_inner<Item, std::atto>(formatstring, item);
-    break;
-  case 2:
-    invoke_inner<Item, std::femto>(formatstring, item);
-    break;
-  case 3:
-    invoke_inner<Item, std::pico>(formatstring, item);
-    break;
-  case 4:
-    invoke_inner<Item, std::nano>(formatstring, item);
-    break;
-  case 5:
-    invoke_inner<Item, std::micro>(formatstring, item);
-    break;
-  case 6:
-    invoke_inner<Item, std::milli>(formatstring, item);
-    break;
-  case 7:
-    invoke_inner<Item, std::centi>(formatstring, item);
-    break;
-  case 8:
-    invoke_inner<Item, std::deci>(formatstring, item);
-    break;
-  case 9:
-    invoke_inner<Item, std::deca>(formatstring, item);
-    break;
-  case 10:
-    invoke_inner<Item, std::kilo>(formatstring, item);
-    break;
-  case 11:
-    invoke_inner<Item, std::mega>(formatstring, item);
-    break;
-  case 12:
-    invoke_inner<Item, std::giga>(formatstring, item);
-    break;
-  case 13:
-    invoke_inner<Item, std::tera>(formatstring, item);
-    break;
-  case 14:
-    invoke_inner<Item, std::peta>(formatstring, item);
-    break;
-  case 15:
-    invoke_inner<Item, std::exa>(formatstring, item);
-  }
-  // doit_impl<Item,std::zeta>(buf.data(),item);
-  // doit_impl<Item,std::yotta>(buf.data(),item);
-}
-
-extern "C" int LLVMFuzzerTestOneInput(const uint8_t* Data, size_t Size) {
-  if (Size <= 4) {
-    return 0;
-  }
-
-  const auto representation = Data[0];
-  const auto scaling = Data[1];
-  Data += 2;
-  Size -= 2;
-
-  switch (representation) {
-  case 1:
-    invoke_outer<char>(Data, Size, scaling);
-    break;
-  case 2:
-    invoke_outer<unsigned char>(Data, Size, scaling);
-    break;
-  case 3:
-    invoke_outer<signed char>(Data, Size, scaling);
-    break;
-  case 4:
-    invoke_outer<short>(Data, Size, scaling);
-    break;
-  case 5:
-    invoke_outer<unsigned short>(Data, Size, scaling);
-    break;
-  case 6:
-    invoke_outer<int>(Data, Size, scaling);
-    break;
-  case 7:
-    invoke_outer<unsigned int>(Data, Size, scaling);
-    break;
-  case 8:
-    invoke_outer<long>(Data, Size, scaling);
-    break;
-  case 9:
-    invoke_outer<unsigned long>(Data, Size, scaling);
-    break;
-  case 10:
-    invoke_outer<float>(Data, Size, scaling);
-    break;
-  case 11:
-    invoke_outer<double>(Data, Size, scaling);
-    break;
-  case 12:
-    invoke_outer<long double>(Data, Size, scaling);
-    break;
-  default:
-    break;
-  }
-
-  return 0;
-}
diff --git a/test/fuzzing/float.cc b/test/fuzzing/float.cc
new file mode 100644
index 00000000..a9a347ef
--- /dev/null
+++ b/test/fuzzing/float.cc
@@ -0,0 +1,34 @@
+// A fuzzer for floating-point formatter.
+// For the license information refer to format.h.
+
+#include <cstdint>
+#include <cstdlib>
+#include <stdexcept>
+#include <limits>
+#include <fmt/format.h>
+
+#include "fuzzer-common.h"
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) {
+  if (size <= sizeof(double) || !std::numeric_limits<double>::is_iec559)
+    return 0;
+
+  auto value = assign_from_buf<double>(data);
+  auto buffer = fmt::memory_buffer();
+  fmt::format_to(buffer, "{}", value);
+
+  // Check a round trip.
+  if (std::isnan(value)) {
+    auto nan = std::signbit(value) ? "-nan" : "nan";
+    if (fmt::string_view(buffer.data(), buffer.size()) != nan)
+      throw std::runtime_error("round trip failure");
+    return 0;
+  }
+  buffer.push_back('\0');
+  char* ptr = nullptr;
+  if (std::strtod(buffer.data(), &ptr) != value)
+    throw std::runtime_error("round trip failure");
+  if (ptr + 1 != buffer.end())
+    throw std::runtime_error("unparsed output");
+  return 0;
+}
diff --git a/test/fuzzing/fuzzer-common.h b/test/fuzzing/fuzzer-common.h
new file mode 100644
index 00000000..635a5d99
--- /dev/null
+++ b/test/fuzzing/fuzzer-common.h
@@ -0,0 +1,75 @@
+// Copyright (c) 2019, Paul Dreik
+// For the license information refer to format.h.
+
+#ifndef FUZZER_COMMON_H
+#define FUZZER_COMMON_H
+
+#include <cstdint>      // std::uint8_t
+#include <cstring>      // memcpy
+#include <vector>
+
+#include <fmt/core.h>
+
+// One can format to either a string, or a buffer. The latter is faster, but
+// one may be interested in formatting to a string instead to verify it works
+// as intended. To avoid a combinatoric explosion, select this at compile time
+// instead of dynamically from the fuzz data.
+#define FMT_FUZZ_FORMAT_TO_STRING 0
+
+// If {fmt} is given a buffer that is separately allocated, chances that address
+// sanitizer detects out of bound reads is much higher. However, it slows down
+// the fuzzing.
+#define FMT_FUZZ_SEPARATE_ALLOCATION 1
+
+// The size of the largest possible type in use.
+// To let the the fuzzer mutation be efficient at cross pollinating between
+// different types, use a fixed size format. The same bit pattern, interpreted
+// as another type, is likely interesting.
+constexpr auto fixed_size = 16;
+
+// Casts data to a char pointer.
+template <typename T> inline const char* as_chars(const T* data) {
+  return reinterpret_cast<const char*>(data);
+}
+
+// Casts data to a byte pointer.
+template <typename T> inline const std::uint8_t* as_bytes(const T* data) {
+  return reinterpret_cast<const std::uint8_t*>(data);
+}
+
+// Blits bytes from data to form an (assumed trivially constructible) object
+// of type Item.
+template <class Item> inline Item assign_from_buf(const std::uint8_t* data) {
+  auto item = Item();
+  std::memcpy(&item, data, sizeof(Item));
+  return item;
+}
+
+// Reads a boolean value by looking at the first byte from data.
+template <> inline bool assign_from_buf<bool>(const std::uint8_t* data) {
+  return *data != 0;
+}
+
+struct data_to_string {
+#if FMT_FUZZ_SEPARATE_ALLOCATION
+  std::vector<char> buffer;
+
+  data_to_string(const uint8_t* data, size_t size, bool add_terminator = false)
+      : buffer(size + (add_terminator ? 1 : 0)) {
+    std::memcpy(buffer.data(), data, size);
+  }
+
+  fmt::string_view get() const { return {buffer.data(), buffer.size()}; }
+#else
+  fmt::string_view sv;
+
+  data_to_string(const uint8_t* data, size_t size, bool = false)
+      : str(as_chars(data), size) {}
+
+  fmt::string_view get() const { return sv; }
+#endif
+
+  const char* data() const { return get().data(); }
+};
+
+#endif  // FUZZER_COMMON_H
diff --git a/test/fuzzing/fuzzer_common.h b/test/fuzzing/fuzzer_common.h
deleted file mode 100644
index c3d85619..00000000
--- a/test/fuzzing/fuzzer_common.h
+++ /dev/null
@@ -1,67 +0,0 @@
-#ifndef FUZZER_COMMON_H
-#define FUZZER_COMMON_H
-
-// Copyright (c) 2019, Paul Dreik
-// License: see LICENSE.rst in the fmt root directory
-
-#include <cstdint>      // std::uint8_t
-#include <cstring>      // memcpy
-#include <type_traits>  // trivially copyable
-
-// one can format to either a string, or a buf. buf is faster,
-// but one may be interested in formatting to a string instead to
-// verify it works as intended. to avoid a combinatoric explosion,
-// select this at compile time instead of dynamically from the fuzz data
-#define FMT_FUZZ_FORMAT_TO_STRING 0
-
-// if fmt is given a buffer that is separately allocated,
-// chances that address sanitizer detects out of bound reads is
-// much higher. However, it slows down the fuzzing.
-#define FMT_FUZZ_SEPARATE_ALLOCATION 1
-
-// To let the the fuzzer mutation be efficient at cross pollinating
-// between different types, use a fixed size format.
-// The same bit pattern, interpreted as another type,
-// is likely interesting.
-// For this, we must know the size of the largest possible type in use.
-
-// There are some problems on travis, claiming Nfixed is not a constant
-// expression which seems to be an issue with older versions of libstdc++
-#if _GLIBCXX_RELEASE >= 7
-#  include <algorithm>
-namespace fmt_fuzzer {
-constexpr auto Nfixed = std::max(sizeof(long double), sizeof(std::intmax_t));
-}
-#else
-namespace fmt_fuzzer {
-constexpr auto Nfixed = 16;
-}
-#endif
-
-namespace fmt_fuzzer {
-// view data as a c char pointer.
-template <typename T> inline const char* as_chars(const T* data) {
-  return static_cast<const char*>(static_cast<const void*>(data));
-}
-
-// view data as a byte pointer
-template <typename T> inline const std::uint8_t* as_bytes(const T* data) {
-  return static_cast<const std::uint8_t*>(static_cast<const void*>(data));
-}
-
-// blits bytes from Data to form an (assumed trivially constructible) object
-// of type Item
-template <class Item> inline Item assignFromBuf(const std::uint8_t* Data) {
-  Item item{};
-  std::memcpy(&item, Data, sizeof(Item));
-  return item;
-}
-
-// reads a boolean value by looking at the first byte from Data
-template <> inline bool assignFromBuf<bool>(const std::uint8_t* Data) {
-  return !!Data[0];
-}
-
-}  // namespace fmt_fuzzer
-
-#endif  // FUZZER_COMMON_H
diff --git a/test/fuzzing/main.cc b/test/fuzzing/main.cc
new file mode 100644
index 00000000..8f8c719b
--- /dev/null
+++ b/test/fuzzing/main.cc
@@ -0,0 +1,22 @@
+#include <cassert>
+#include <fstream>
+#include <vector>
+
+#include "fuzzer-common.h"
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size);
+
+int main(int argc, char** argv) {
+  for (int i = 1; i < argc; ++i) {
+    std::ifstream in(argv[i]);
+    assert(in);
+    in.seekg(0, std::ios_base::end);
+    const auto size = in.tellg();
+    assert(size >= 0);
+    in.seekg(0, std::ios_base::beg);
+    std::vector<char> buf(static_cast<size_t>(size));
+    in.read(buf.data(), size);
+    assert(in.gcount() == size);
+    LLVMFuzzerTestOneInput(as_bytes(buf.data()), buf.size());
+  }
+}
diff --git a/test/fuzzing/main.cpp b/test/fuzzing/main.cpp
deleted file mode 100644
index 1053eeeb..00000000
--- a/test/fuzzing/main.cpp
+++ /dev/null
@@ -1,21 +0,0 @@
-#include <cassert>
-#include <fstream>
-#include <sstream>
-#include <vector>
-#include "fuzzer_common.h"
-
-extern "C" int LLVMFuzzerTestOneInput(const uint8_t* Data, size_t Size);
-int main(int argc, char* argv[]) {
-  for (int i = 1; i < argc; ++i) {
-    std::ifstream in(argv[i]);
-    assert(in);
-    in.seekg(0, std::ios_base::end);
-    const auto pos = in.tellg();
-    assert(pos >= 0);
-    in.seekg(0, std::ios_base::beg);
-    std::vector<char> buf(static_cast<size_t>(pos));
-    in.read(buf.data(), static_cast<long>(buf.size()));
-    assert(in.gcount() == pos);
-    LLVMFuzzerTestOneInput(fmt_fuzzer::as_bytes(buf.data()), buf.size());
-  }
-}
diff --git a/test/fuzzing/named-arg.cc b/test/fuzzing/named-arg.cc
new file mode 100644
index 00000000..ffd8e903
--- /dev/null
+++ b/test/fuzzing/named-arg.cc
@@ -0,0 +1,100 @@
+// Copyright (c) 2019, Paul Dreik
+// For the license information refer to format.h.
+
+#include <cstdint>
+#include <type_traits>
+#include <vector>
+#include <fmt/chrono.h>
+
+#include "fuzzer-common.h"
+
+template <typename T>
+void invoke_fmt(const uint8_t* data, size_t size, unsigned arg_name_size) {
+  static_assert(sizeof(T) <= fixed_size, "fixed_size too small");
+  if (size <= fixed_size) return;
+  const T value = assign_from_buf<T>(data);
+  data += fixed_size;
+  size -= fixed_size;
+
+  if (arg_name_size <= 0 || arg_name_size >= size) return;
+  data_to_string arg_name(data, arg_name_size, true);
+  data += arg_name_size;
+  size -= arg_name_size;
+
+  data_to_string format_str(data, size);
+  try {
+#if FMT_FUZZ_FORMAT_TO_STRING
+    std::string message =
+      fmt::format(format_str.get(), fmt::arg(arg_name.data(), value));
+#else
+    fmt::memory_buffer out;
+    fmt::format_to(out, format_str.get(), fmt::arg(arg_name.data(), value));
+#endif
+  } catch (std::exception&) {
+  }
+}
+
+// For dynamic dispatching to an explicit instantiation.
+template <typename Callback> void invoke(int type, Callback callback) {
+  switch (type) {
+  case 0:
+    callback(bool());
+    break;
+  case 1:
+    callback(char());
+    break;
+  case 2:
+    using sc = signed char;
+    callback(sc());
+    break;
+  case 3:
+    using uc = unsigned char;
+    callback(uc());
+    break;
+  case 4:
+    callback(short());
+    break;
+  case 5:
+    using us = unsigned short;
+    callback(us());
+    break;
+  case 6:
+    callback(int());
+    break;
+  case 7:
+    callback(unsigned());
+    break;
+  case 8:
+    callback(long());
+    break;
+  case 9:
+    using ul = unsigned long;
+    callback(ul());
+    break;
+  case 10:
+    callback(float());
+    break;
+  case 11:
+    callback(double());
+    break;
+  case 12:
+    using LD = long double;
+    callback(LD());
+    break;
+  }
+}
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) {
+  if (size <= 3) return 0;
+
+  // Switch types depending on the first byte of the input.
+  const auto type = data[0] & 0x0F;
+  const unsigned arg_name_size = (data[0] & 0xF0) >> 4;
+  data++;
+  size--;
+
+  invoke(type, [=](auto arg) {
+    invoke_fmt<decltype(arg)>(data, size, arg_name_size);
+  });
+  return 0;
+}
diff --git a/test/fuzzing/named_arg.cpp b/test/fuzzing/named_arg.cpp
deleted file mode 100644
index bd0cb686..00000000
--- a/test/fuzzing/named_arg.cpp
+++ /dev/null
@@ -1,128 +0,0 @@
-// Copyright (c) 2019, Paul Dreik
-// License: see LICENSE.rst in the fmt root directory
-
-#include <fmt/chrono.h>
-#include <fmt/core.h>
-#include <cstdint>
-#include <stdexcept>
-#include <type_traits>
-#include <vector>
-#include "fuzzer_common.h"
-
-template <typename Item1>
-void invoke_fmt(const uint8_t* Data, size_t Size, unsigned int argsize) {
-  constexpr auto N1 = sizeof(Item1);
-  static_assert(N1 <= fmt_fuzzer::Nfixed, "Nfixed too small");
-  if (Size <= fmt_fuzzer::Nfixed) {
-    return;
-  }
-  const Item1 item1 = fmt_fuzzer::assignFromBuf<Item1>(Data);
-
-  Data += fmt_fuzzer::Nfixed;
-  Size -= fmt_fuzzer::Nfixed;
-
-  // how many chars should be used for the argument name?
-  if (argsize <= 0 || argsize >= Size) {
-    return;
-  }
-
-  // allocating buffers separately is slower, but increases chances
-  // of detecting memory errors
-#if FMT_FUZZ_SEPARATE_ALLOCATION
-  std::vector<char> argnamebuffer(argsize + 1);
-  std::memcpy(argnamebuffer.data(), Data, argsize);
-  auto argname = argnamebuffer.data();
-#else
-  auto argname = fmt_fuzzer::as_chars(Data);
-#endif
-  Data += argsize;
-  Size -= argsize;
-
-#if FMT_FUZZ_SEPARATE_ALLOCATION
-  // allocates as tight as possible, making it easier to catch buffer overruns.
-  std::vector<char> fmtstringbuffer(Size);
-  std::memcpy(fmtstringbuffer.data(), Data, Size);
-  auto fmtstring = fmt::string_view(fmtstringbuffer.data(), Size);
-#else
-  auto fmtstring = fmt::string_view(fmt_fuzzer::as_chars(Data), Size);
-#endif
-
-#if FMT_FUZZ_FORMAT_TO_STRING
-  std::string message = fmt::format(fmtstring, fmt::arg(argname, item1));
-#else
-  fmt::memory_buffer outbuf;
-  fmt::format_to(outbuf, fmtstring, fmt::arg(argname, item1));
-#endif
-}
-
-// for dynamic dispatching to an explicit instantiation
-template <typename Callback> void invoke(int index, Callback callback) {
-  switch (index) {
-  case 0:
-    callback(bool{});
-    break;
-  case 1:
-    callback(char{});
-    break;
-  case 2:
-    using sc = signed char;
-    callback(sc{});
-    break;
-  case 3:
-    using uc = unsigned char;
-    callback(uc{});
-    break;
-  case 4:
-    callback(short{});
-    break;
-  case 5:
-    using us = unsigned short;
-    callback(us{});
-    break;
-  case 6:
-    callback(int{});
-    break;
-  case 7:
-    callback(unsigned{});
-    break;
-  case 8:
-    callback(long{});
-    break;
-  case 9:
-    using ul = unsigned long;
-    callback(ul{});
-    break;
-  case 10:
-    callback(float{});
-    break;
-  case 11:
-    callback(double{});
-    break;
-  case 12:
-    using LD = long double;
-    callback(LD{});
-    break;
-  }
-}
-
-extern "C" int LLVMFuzzerTestOneInput(const uint8_t* Data, size_t Size) {
-  if (Size <= 3) {
-    return 0;
-  }
-
-  // switch types depending on the first byte of the input
-  const auto first = Data[0] & 0x0F;
-  const unsigned int second = (Data[0] & 0xF0) >> 4;
-  Data++;
-  Size--;
-
-  auto outerfcn = [=](auto param1) {
-    invoke_fmt<decltype(param1)>(Data, Size, second);
-  };
-
-  try {
-    invoke(first, outerfcn);
-  } catch (std::exception& /*e*/) {
-  }
-  return 0;
-}
diff --git a/test/fuzzing/one-arg.cc b/test/fuzzing/one-arg.cc
new file mode 100644
index 00000000..df173432
--- /dev/null
+++ b/test/fuzzing/one-arg.cc
@@ -0,0 +1,91 @@
+// Copyright (c) 2019, Paul Dreik
+// For the license information refer to format.h.
+
+#include <cstdint>
+#include <exception>
+#include <fmt/chrono.h>
+
+#include "fuzzer-common.h"
+
+template <typename T, typename Repr>
+const T* from_repr(const Repr& r) { return &r; }
+
+template <>
+const std::tm* from_repr<std::tm>(const std::time_t& t) {
+  return std::localtime(&t);
+}
+
+template <typename T, typename Repr = T>
+void invoke_fmt(const uint8_t* data, size_t size) {
+  static_assert(sizeof(Repr) <= fixed_size, "Nfixed is too small");
+  if (size <= fixed_size) return;
+  auto repr = assign_from_buf<Repr>(data);
+  const T* value = from_repr<T>(repr);
+  if (!value) return;
+  data += fixed_size;
+  size -= fixed_size;
+  data_to_string format_str(data, size);
+  try {
+#if FMT_FUZZ_FORMAT_TO_STRING
+    std::string message = fmt::format(format_str.get(), *value);
+#else
+    fmt::memory_buffer message;
+    fmt::format_to(message, format_str.get(), *value);
+#endif
+  } catch (std::exception&) {
+  }
+}
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) {
+  if (size <= 3) return 0;
+
+  const auto first = data[0];
+  data++;
+  size--;
+
+  switch (first) {
+  case 0:
+    invoke_fmt<bool>(data, size);
+    break;
+  case 1:
+    invoke_fmt<char>(data, size);
+    break;
+  case 2:
+    invoke_fmt<unsigned char>(data, size);
+    break;
+  case 3:
+    invoke_fmt<signed char>(data, size);
+    break;
+  case 4:
+    invoke_fmt<short>(data, size);
+    break;
+  case 5:
+    invoke_fmt<unsigned short>(data, size);
+    break;
+  case 6:
+    invoke_fmt<int>(data, size);
+    break;
+  case 7:
+    invoke_fmt<unsigned int>(data, size);
+    break;
+  case 8:
+    invoke_fmt<long>(data, size);
+    break;
+  case 9:
+    invoke_fmt<unsigned long>(data, size);
+    break;
+  case 10:
+    invoke_fmt<float>(data, size);
+    break;
+  case 11:
+    invoke_fmt<double>(data, size);
+    break;
+  case 12:
+    invoke_fmt<long double>(data, size);
+    break;
+  case 13:
+    invoke_fmt<std::tm, std::time_t>(data, size);
+    break;
+  }
+  return 0;
+}
diff --git a/test/fuzzing/one_arg.cpp b/test/fuzzing/one_arg.cpp
deleted file mode 100644
index 3a1bf5cc..00000000
--- a/test/fuzzing/one_arg.cpp
+++ /dev/null
@@ -1,131 +0,0 @@
-// Copyright (c) 2019, Paul Dreik
-// License: see LICENSE.rst in the fmt root directory
-
-#include <fmt/core.h>
-#include <cstdint>
-#include <stdexcept>
-#include <type_traits>
-#include <vector>
-
-#include <fmt/chrono.h>
-#include "fuzzer_common.h"
-
-using fmt_fuzzer::Nfixed;
-
-template <typename Item>
-void invoke_fmt(const uint8_t* Data, size_t Size) {
-  constexpr auto N = sizeof(Item);
-  static_assert(N <= Nfixed, "Nfixed is too small");
-  if (Size <= Nfixed) {
-    return;
-  }
-  const Item item = fmt_fuzzer::assignFromBuf<Item>(Data);
-  Data += Nfixed;
-  Size -= Nfixed;
-
-#if FMT_FUZZ_SEPARATE_ALLOCATION
-  // allocates as tight as possible, making it easier to catch buffer overruns.
-  std::vector<char> fmtstringbuffer(Size);
-  std::memcpy(fmtstringbuffer.data(), Data, Size);
-  auto fmtstring = fmt::string_view(fmtstringbuffer.data(), Size);
-#else
-  auto fmtstring = fmt::string_view(fmt_fuzzer::as_chars(Data), Size);
-#endif
-
-#if FMT_FUZZ_FORMAT_TO_STRING
-  std::string message = fmt::format(fmtstring, item);
-#else
-  fmt::memory_buffer message;
-  fmt::format_to(message, fmtstring, item);
-#endif
-}
-
-void invoke_fmt_time(const uint8_t* Data, size_t Size) {
-  using Item = std::time_t;
-  constexpr auto N = sizeof(Item);
-  static_assert(N <= Nfixed, "Nfixed too small");
-  if (Size <= Nfixed) {
-    return;
-  }
-  const Item item = fmt_fuzzer::assignFromBuf<Item>(Data);
-  Data += Nfixed;
-  Size -= Nfixed;
-#if FMT_FUZZ_SEPARATE_ALLOCATION
-  // allocates as tight as possible, making it easier to catch buffer overruns.
-  std::vector<char> fmtstringbuffer(Size);
-  std::memcpy(fmtstringbuffer.data(), Data, Size);
-  auto fmtstring = fmt::string_view(fmtstringbuffer.data(), Size);
-#else
-  auto fmtstring = fmt::string_view(fmt_fuzzer::as_chars(Data), Size);
-#endif
-  auto* b = std::localtime(&item);
-  if (b) {
-#if FMT_FUZZ_FORMAT_TO_STRING
-    std::string message = fmt::format(fmtstring, *b);
-#else
-    fmt::memory_buffer message;
-    fmt::format_to(message, fmtstring, *b);
-#endif
-  }
-}
-
-extern "C" int LLVMFuzzerTestOneInput(const uint8_t* Data, size_t Size) {
-  if (Size <= 3) {
-    return 0;
-  }
-
-  const auto first = Data[0];
-  Data++;
-  Size--;
-
-  try {
-    switch (first) {
-    case 0:
-      invoke_fmt<bool>(Data, Size);
-      break;
-    case 1:
-      invoke_fmt<char>(Data, Size);
-      break;
-    case 2:
-      invoke_fmt<unsigned char>(Data, Size);
-      break;
-    case 3:
-      invoke_fmt<signed char>(Data, Size);
-      break;
-    case 4:
-      invoke_fmt<short>(Data, Size);
-      break;
-    case 5:
-      invoke_fmt<unsigned short>(Data, Size);
-      break;
-    case 6:
-      invoke_fmt<int>(Data, Size);
-      break;
-    case 7:
-      invoke_fmt<unsigned int>(Data, Size);
-      break;
-    case 8:
-      invoke_fmt<long>(Data, Size);
-      break;
-    case 9:
-      invoke_fmt<unsigned long>(Data, Size);
-      break;
-    case 10:
-      invoke_fmt<float>(Data, Size);
-      break;
-    case 11:
-      invoke_fmt<double>(Data, Size);
-      break;
-    case 12:
-      invoke_fmt<long double>(Data, Size);
-      break;
-    case 13:
-      invoke_fmt_time(Data, Size);
-      break;
-    default:
-      break;
-    }
-  } catch (std::exception& /*e*/) {
-  }
-  return 0;
-}
diff --git a/test/fuzzing/sprintf.cpp b/test/fuzzing/sprintf.cpp
deleted file mode 100644
index aa028756..00000000
--- a/test/fuzzing/sprintf.cpp
+++ /dev/null
@@ -1,116 +0,0 @@
-// Copyright (c) 2019, Paul Dreik
-// License: see LICENSE.rst in the fmt root directory
-#include <fmt/format.h>
-#include <fmt/printf.h>
-#include <cstdint>
-#include <stdexcept>
-
-#include "fuzzer_common.h"
-
-using fmt_fuzzer::Nfixed;
-
-template <typename Item1, typename Item2>
-void invoke_fmt(const uint8_t* Data, size_t Size) {
-  constexpr auto N1 = sizeof(Item1);
-  constexpr auto N2 = sizeof(Item2);
-  static_assert(N1 <= Nfixed, "size1 exceeded");
-  static_assert(N2 <= Nfixed, "size2 exceeded");
-  if (Size <= Nfixed + Nfixed) {
-    return;
-  }
-  Item1 item1 = fmt_fuzzer::assignFromBuf<Item1>(Data);
-  Data += Nfixed;
-  Size -= Nfixed;
-
-  Item2 item2 = fmt_fuzzer::assignFromBuf<Item2>(Data);
-  Data += Nfixed;
-  Size -= Nfixed;
-
-  auto fmtstring = fmt::string_view(fmt_fuzzer::as_chars(Data), Size);
-
-#if FMT_FUZZ_FORMAT_TO_STRING
-  std::string message = fmt::format(fmtstring, item1, item2);
-#else
-  fmt::memory_buffer message;
-  fmt::format_to(message, fmtstring, item1, item2);
-#endif
-}
-
-// for dynamic dispatching to an explicit instantiation
-template <typename Callback> void invoke(int index, Callback callback) {
-  switch (index) {
-  case 0:
-    callback(bool{});
-    break;
-  case 1:
-    callback(char{});
-    break;
-  case 2:
-    using sc = signed char;
-    callback(sc{});
-    break;
-  case 3:
-    using uc = unsigned char;
-    callback(uc{});
-    break;
-  case 4:
-    callback(short{});
-    break;
-  case 5:
-    using us = unsigned short;
-    callback(us{});
-    break;
-  case 6:
-    callback(int{});
-    break;
-  case 7:
-    callback(unsigned{});
-    break;
-  case 8:
-    callback(long{});
-    break;
-  case 9:
-    using ul = unsigned long;
-    callback(ul{});
-    break;
-  case 10:
-    callback(float{});
-    break;
-  case 11:
-    callback(double{});
-    break;
-  case 12:
-    using LD = long double;
-    callback(LD{});
-    break;
-  case 13:
-    using ptr = void*;
-    callback(ptr{});
-    break;
-  }
-}
-
-extern "C" int LLVMFuzzerTestOneInput(const uint8_t* Data, size_t Size) {
-  if (Size <= 3) {
-    return 0;
-  }
-
-  // switch types depending on the first byte of the input
-  const auto first = Data[0] & 0x0F;
-  const auto second = (Data[0] & 0xF0) >> 4;
-  Data++;
-  Size--;
-
-  auto outer = [=](auto param1) {
-    auto inner = [=](auto param2) {
-      invoke_fmt<decltype(param1), decltype(param2)>(Data, Size);
-    };
-    invoke(second, inner);
-  };
-
-  try {
-    invoke(first, outer);
-  } catch (std::exception& /*e*/) {
-  }
-  return 0;
-}
diff --git a/test/fuzzing/two-args.cc b/test/fuzzing/two-args.cc
new file mode 100644
index 00000000..4d7d3453
--- /dev/null
+++ b/test/fuzzing/two-args.cc
@@ -0,0 +1,105 @@
+// Copyright (c) 2019, Paul Dreik
+// For the license information refer to format.h.
+
+#include <cstdint>
+#include <exception>
+#include <string>
+#include <fmt/format.h>
+
+#include "fuzzer-common.h"
+
+template <typename Item1, typename Item2>
+void invoke_fmt(const uint8_t* data, size_t size) {
+  static_assert(sizeof(Item1) <= fixed_size, "size1 exceeded");
+  static_assert(sizeof(Item2) <= fixed_size, "size2 exceeded");
+  if (size <= fixed_size + fixed_size) return;
+
+  const Item1 item1 = assign_from_buf<Item1>(data);
+  data += fixed_size;
+  size -= fixed_size;
+
+  const Item2 item2 = assign_from_buf<Item2>(data);
+  data += fixed_size;
+  size -= fixed_size;
+
+  auto format_str = fmt::string_view(as_chars(data), size);
+#if FMT_FUZZ_FORMAT_TO_STRING
+  std::string message = fmt::format(format_str, item1, item2);
+#else
+  fmt::memory_buffer message;
+  fmt::format_to(message, format_str, item1, item2);
+#endif
+}
+
+// For dynamic dispatching to an explicit instantiation.
+template <typename Callback> void invoke(int index, Callback callback) {
+  switch (index) {
+  case 0:
+    callback(bool());
+    break;
+  case 1:
+    callback(char());
+    break;
+  case 2:
+    using sc = signed char;
+    callback(sc());
+    break;
+  case 3:
+    using uc = unsigned char;
+    callback(uc());
+    break;
+  case 4:
+    callback(short());
+    break;
+  case 5:
+    using us = unsigned short;
+    callback(us());
+    break;
+  case 6:
+    callback(int());
+    break;
+  case 7:
+    callback(unsigned());
+    break;
+  case 8:
+    callback(long());
+    break;
+  case 9:
+    using ul = unsigned long;
+    callback(ul());
+    break;
+  case 10:
+    callback(float());
+    break;
+  case 11:
+    callback(double());
+    break;
+  case 12:
+    using LD = long double;
+    callback(LD());
+    break;
+  case 13:
+    using ptr = void*;
+    callback(ptr());
+    break;
+  }
+}
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) {
+  if (size <= 3) return 0;
+
+  // Switch types depending on the first byte of the input.
+  const auto type1 = data[0] & 0x0F;
+  const auto type2 = (data[0] & 0xF0) >> 4;
+  data++;
+  size--;
+  try {
+    invoke(type1, [=](auto param1) {
+      invoke(type2, [=](auto param2) {
+        invoke_fmt<decltype(param1), decltype(param2)>(data, size);
+      });
+    });
+  } catch (std::exception&) {
+  }
+  return 0;
+}
diff --git a/test/fuzzing/two_args.cpp b/test/fuzzing/two_args.cpp
deleted file mode 100644
index af3495ce..00000000
--- a/test/fuzzing/two_args.cpp
+++ /dev/null
@@ -1,112 +0,0 @@
-// Copyright (c) 2019, Paul Dreik
-// License: see LICENSE.rst in the fmt root directory
-#include <fmt/format.h>
-#include <cstdint>
-#include <stdexcept>
-#include <type_traits>
-
-#include "fuzzer_common.h"
-
-constexpr auto Nfixed = fmt_fuzzer::Nfixed;
-
-template <typename Item1, typename Item2>
-void invoke_fmt(const uint8_t* Data, size_t Size) {
-  constexpr auto N1 = sizeof(Item1);
-  constexpr auto N2 = sizeof(Item2);
-  static_assert(N1 <= Nfixed, "size1 exceeded");
-  static_assert(N2 <= Nfixed, "size2 exceeded");
-  if (Size <= Nfixed + Nfixed) {
-    return;
-  }
-  const Item1 item1 = fmt_fuzzer::assignFromBuf<Item1>(Data);
-  Data += Nfixed;
-  Size -= Nfixed;
-
-  const Item2 item2 = fmt_fuzzer::assignFromBuf<Item2>(Data);
-  Data += Nfixed;
-  Size -= Nfixed;
-
-  auto fmtstring = fmt::string_view(fmt_fuzzer::as_chars(Data), Size);
-
-#if FMT_FUZZ_FORMAT_TO_STRING
-  std::string message = fmt::format(fmtstring, item1, item2);
-#else
-  fmt::memory_buffer message;
-  fmt::format_to(message, fmtstring, item1, item2);
-#endif
-}
-
-// for dynamic dispatching to an explicit instantiation
-template <typename Callback> void invoke(int index, Callback callback) {
-  switch (index) {
-  case 0:
-    callback(bool{});
-    break;
-  case 1:
-    callback(char{});
-    break;
-  case 2:
-    using sc = signed char;
-    callback(sc{});
-    break;
-  case 3:
-    using uc = unsigned char;
-    callback(uc{});
-    break;
-  case 4:
-    callback(short{});
-    break;
-  case 5:
-    using us = unsigned short;
-    callback(us{});
-    break;
-  case 6:
-    callback(int{});
-    break;
-  case 7:
-    callback(unsigned{});
-    break;
-  case 8:
-    callback(long{});
-    break;
-  case 9:
-    using ul = unsigned long;
-    callback(ul{});
-    break;
-  case 10:
-    callback(float{});
-    break;
-  case 11:
-    callback(double{});
-    break;
-  case 12:
-    using LD = long double;
-    callback(LD{});
-    break;
-  }
-}
-
-extern "C" int LLVMFuzzerTestOneInput(const uint8_t* Data, size_t Size) {
-  if (Size <= 3) {
-    return 0;
-  }
-
-  // switch types depending on the first byte of the input
-  const auto first = Data[0] & 0x0F;
-  const auto second = (Data[0] & 0xF0) >> 4;
-  Data++;
-  Size--;
-
-  auto outer = [=](auto param1) {
-    auto inner = [=](auto param2) {
-      invoke_fmt<decltype(param1), decltype(param2)>(Data, Size);
-    };
-    invoke(second, inner);
-  };
-
-  try {
-    invoke(first, outer);
-  } catch (std::exception& /*e*/) {
-  }
-  return 0;
-}
diff --git a/test/grisu-test.cc b/test/grisu-test.cc
deleted file mode 100644
index 3fc670cd..00000000
--- a/test/grisu-test.cc
+++ /dev/null
@@ -1,75 +0,0 @@
-// Formatting library for C++ - Grisu tests
-//
-// Copyright (c) 2012 - present, Victor Zverovich
-// All rights reserved.
-//
-// For the license information refer to format.h.
-
-#include "fmt/format.h"
-#include "gtest.h"
-
-static bool reported_skipped;
-
-#undef TEST
-#define TEST(test_fixture, test_name)        \
-  void test_fixture##test_name();            \
-  GTEST_TEST(test_fixture, test_name) {      \
-    if (FMT_USE_GRISU) {                     \
-      test_fixture##test_name();             \
-    } else if (!reported_skipped) {          \
-      reported_skipped = true;               \
-      fmt::print("Skipping Grisu tests.\n"); \
-    }                                        \
-  }                                          \
-  void test_fixture##test_name()
-
-TEST(GrisuTest, NaN) {
-  auto nan = std::numeric_limits<double>::quiet_NaN();
-  EXPECT_EQ("nan", fmt::format("{}", nan));
-  EXPECT_EQ("-nan", fmt::format("{}", -nan));
-}
-
-TEST(GrisuTest, Inf) {
-  auto inf = std::numeric_limits<double>::infinity();
-  EXPECT_EQ("inf", fmt::format("{}", inf));
-  EXPECT_EQ("-inf", fmt::format("{}", -inf));
-}
-
-TEST(GrisuTest, Zero) { EXPECT_EQ("0.0", fmt::format("{}", 0.0)); }
-
-TEST(GrisuTest, Round) {
-  EXPECT_EQ("1.9156918820264798e-56",
-            fmt::format("{}", 1.9156918820264798e-56));
-  EXPECT_EQ("0.0000", fmt::format("{:.4f}", 7.2809479766055470e-15));
-}
-
-TEST(GrisuTest, Prettify) {
-  EXPECT_EQ("0.0001", fmt::format("{}", 1e-4));
-  EXPECT_EQ("1e-05", fmt::format("{}", 1e-5));
-  EXPECT_EQ("9.999e-05", fmt::format("{}", 9.999e-5));
-  EXPECT_EQ("10000000000.0", fmt::format("{}", 1e10));
-  EXPECT_EQ("100000000000.0", fmt::format("{}", 1e11));
-  EXPECT_EQ("12340000000.0", fmt::format("{}", 1234e7));
-  EXPECT_EQ("12.34", fmt::format("{}", 1234e-2));
-  EXPECT_EQ("0.001234", fmt::format("{}", 1234e-6));
-  EXPECT_EQ("0.1", fmt::format("{}", 0.1f));
-  EXPECT_EQ("0.10000000149011612", fmt::format("{}", double(0.1f)));
-}
-
-TEST(GrisuTest, ZeroPrecision) { EXPECT_EQ("1", fmt::format("{:.0}", 1.0)); }
-
-TEST(GrisuTest, Fallback) {
-  EXPECT_EQ("1e+23", fmt::format("{}", 1e23));
-  EXPECT_EQ("9e-265", fmt::format("{}", 9e-265));
-  EXPECT_EQ("5.423717798060526e+125",
-            fmt::format("{}", 5.423717798060526e+125));
-  EXPECT_EQ("1.372371880954233e-288",
-            fmt::format("{}", 1.372371880954233e-288));
-  EXPECT_EQ("55388492.622190244", fmt::format("{}", 55388492.622190244));
-  EXPECT_EQ("2.2506787569811123e-253",
-            fmt::format("{}", 2.2506787569811123e-253));
-  EXPECT_EQ("1103618912042992.8", fmt::format("{}", 1103618912042992.8));
-  // pow(2, -25) - assymetric boundaries:
-  EXPECT_EQ("2.9802322387695312e-08",
-            fmt::format("{}", 2.9802322387695312e-08));
-}
diff --git a/test/gtest-extra.h b/test/gtest-extra.h
index 3ed8052b..01c70ddb 100644
--- a/test/gtest-extra.h
+++ b/test/gtest-extra.h
@@ -145,7 +145,13 @@ std::string read(fmt::file& f, size_t count);
               read(file, fmt::string_view(expected_content).size()))
 
 #else
-#  define EXPECT_WRITE(file, statement, expected_output) SUCCEED()
+#  define EXPECT_WRITE(file, statement, expected_output) \
+    do {                                                 \
+      (void)(file);                                      \
+      (void)(statement);                                 \
+      (void)(expected_output);                           \
+      SUCCEED();                                         \
+    } while (false)
 #endif  // FMT_USE_FCNTL
 
 template <typename Mock> struct ScopedMock : testing::StrictMock<Mock> {
diff --git a/test/locale-test.cc b/test/locale-test.cc
index fcce0091..c7c30f2e 100644
--- a/test/locale-test.cc
+++ b/test/locale-test.cc
@@ -7,6 +7,8 @@
 
 #include "fmt/locale.h"
 
+#include <complex>
+
 #include "gmock.h"
 
 using fmt::detail::max_value;
@@ -50,6 +52,7 @@ TEST(LocaleTest, Format) {
   EXPECT_EQ("1234567", fmt::format(std::locale(), "{:L}", 1234567));
   EXPECT_EQ("1~234~567", fmt::format(loc, "{:L}", 1234567));
   EXPECT_EQ("-1~234~567", fmt::format(loc, "{:L}", -1234567));
+  EXPECT_EQ("-256", fmt::format(loc, "{:L}", -256));
   fmt::format_arg_store<fmt::format_context, int> as{1234567};
   EXPECT_EQ("1~234~567", fmt::vformat(loc, "{:L}", fmt::format_args(as)));
   std::string s;
@@ -61,12 +64,18 @@ TEST(LocaleTest, Format) {
 
   std::locale special_grouping_loc(std::locale(), new special_grouping<char>());
   EXPECT_EQ("1,23,45,678", fmt::format(special_grouping_loc, "{:L}", 12345678));
+  EXPECT_EQ("12,345", fmt::format(special_grouping_loc, "{:L}", 12345));
 
   std::locale small_grouping_loc(std::locale(), new small_grouping<char>());
   EXPECT_EQ("4,2,9,4,9,6,7,2,9,5",
             fmt::format(small_grouping_loc, "{:L}", max_value<uint32_t>()));
 }
 
+TEST(LocaleTest, FormatDetaultAlign) {
+  std::locale special_grouping_loc(std::locale(), new special_grouping<char>());
+  EXPECT_EQ("  12,345", fmt::format(special_grouping_loc, "{:8L}", 12345));
+}
+
 TEST(LocaleTest, WFormat) {
   std::locale loc(std::locale(), new numpunct<wchar_t>());
   EXPECT_EQ(L"1234567", fmt::format(std::locale(), L"{:L}", 1234567));
@@ -88,4 +97,64 @@ TEST(LocaleTest, WFormat) {
             fmt::format(small_grouping_loc, L"{:L}", max_value<uint32_t>()));
 }
 
+TEST(LocaleTest, DoubleFormatter) {
+  auto loc = std::locale(std::locale(), new special_grouping<char>());
+  auto f = fmt::formatter<int>();
+  auto parse_ctx = fmt::format_parse_context("L");
+  f.parse(parse_ctx);
+  char buf[10] = {};
+  fmt::basic_format_context<char*, char> format_ctx(
+      buf, {}, fmt::detail::locale_ref(loc));
+  *f.format(12345, format_ctx) = 0;
+  EXPECT_STREQ("12,345", buf);
+}
+
+FMT_BEGIN_NAMESPACE
+template <class charT> struct formatter<std::complex<double>, charT> {
+ private:
+  detail::dynamic_format_specs<char> specs_;
+
+ public:
+  typename basic_format_parse_context<charT>::iterator parse(
+      basic_format_parse_context<charT>& ctx) {
+    using handler_type =
+        detail::dynamic_specs_handler<basic_format_parse_context<charT>>;
+    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::parse_float_type_spec(specs_, ctx.error_handler());
+    return it;
+  }
+
+  template <class FormatContext>
+  typename FormatContext::iterator format(const std::complex<double>& c,
+                                          FormatContext& ctx) {
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs_.precision, specs_.precision_ref, ctx);
+    auto format_specs = std::string();
+    if (specs_.precision > 0)
+      format_specs = fmt::format(".{}", specs_.precision);
+    if (specs_.type)
+      format_specs += specs_.type;
+    auto real = fmt::format(ctx.locale().template get<std::locale>(),
+                            "{:" + format_specs + "}", c.real());
+    auto imag = fmt::format(ctx.locale().template get<std::locale>(),
+                            "{:" + format_specs + "}", c.imag());
+    auto fill_align_width = std::string();
+    if (specs_.width > 0)
+      fill_align_width = fmt::format(">{}", specs_.width);
+    return format_to(
+        ctx.out(), "{:" + fill_align_width + "}",
+        fmt::format(c.real() != 0 ? "({0}+{1}i)" : "{1}i", real, imag));
+  }
+};
+FMT_END_NAMESPACE
+
+TEST(FormatTest, Complex) {
+  std::string s = fmt::format("{}", std::complex<double>(1, 2));
+  EXPECT_EQ(s, "(1+2i)");
+  EXPECT_EQ(fmt::format("{:.2f}", std::complex<double>(1, 2)), "(1.00+2.00i)");
+  EXPECT_EQ(fmt::format("{:8}", std::complex<double>(1, 2)), "  (1+2i)");
+}
+
 #endif  // FMT_STATIC_THOUSANDS_SEPARATOR
diff --git a/test/os-test.cc b/test/os-test.cc
index 186198ef..359b5ff8 100644
--- a/test/os-test.cc
+++ b/test/os-test.cc
@@ -81,9 +81,9 @@ TEST(UtilTest, FormatWindowsError) {
   EXPECT_EQ(fmt::format("test: {}", utf8_message.str()),
             fmt::to_string(actual_message));
   actual_message.resize(0);
-  auto max_size = fmt::detail::max_value<size_t>();
+  auto max_size = fmt::detail::max_value<size_t>() / 2;
   fmt::detail::format_windows_error(actual_message, ERROR_FILE_EXISTS,
-                                    fmt::string_view(0, max_size));
+                                    fmt::string_view(nullptr, max_size));
   EXPECT_EQ(fmt::format("error {}", ERROR_FILE_EXISTS),
             fmt::to_string(actual_message));
 }
@@ -287,26 +287,38 @@ TEST(BufferedFileTest, Fileno) {
   EXPECT_READ(copy, FILE_CONTENT);
 }
 
-TEST(DirectBufferedFileTest, Print) {
-  fmt::direct_buffered_file out(
-    "test-file", fmt::file::WRONLY | fmt::file::CREATE);
-  fmt::print(out, "The answer is {}.\n", 42);
+TEST(OStreamTest, Move) {
+  fmt::ostream out = fmt::output_file("test-file");
+  fmt::ostream moved(std::move(out));
+  moved.print("hello");
+}
+
+TEST(OStreamTest, Print) {
+  fmt::ostream out = fmt::output_file("test-file");
+  out.print("The answer is {}.\n", 42);
   out.close();
   file in("test-file", file::RDONLY);
   EXPECT_READ(in, "The answer is 42.\n");
 }
 
-TEST(DirectBufferedFileTest, BufferBoundary) {
+TEST(OStreamTest, BufferBoundary) {
   auto str = std::string(4096, 'x');
-  fmt::direct_buffered_file out(
-    "test-file", fmt::file::WRONLY | fmt::file::CREATE);
-  fmt::print(out, "{}", str);
-  fmt::print(out, "{}", str);
+  fmt::ostream out = fmt::output_file("test-file");
+  out.print("{}", str);
+  out.print("{}", str);
   out.close();
   file in("test-file", file::RDONLY);
   EXPECT_READ(in, str + str);
 }
 
+TEST(OStreamTest, BufferSize) {
+  fmt::ostream out = fmt::output_file("test-file", fmt::buffer_size=1);
+  out.print("{}", "foo");
+  out.close();
+  file in("test-file", file::RDONLY);
+  EXPECT_READ(in, "foo");
+}
+
 TEST(FileTest, DefaultCtor) {
   file f;
   EXPECT_EQ(-1, f.descriptor());
diff --git a/test/ostream-test.cc b/test/ostream-test.cc
index 1c87d46d..ebf14210 100644
--- a/test/ostream-test.cc
+++ b/test/ostream-test.cc
@@ -75,8 +75,8 @@ struct test_arg_formatter
 
 TEST(OStreamTest, CustomArg) {
   fmt::memory_buffer buffer;
-  fmt::detail::buffer<char>& base = buffer;
-  fmt::format_context ctx(std::back_inserter(base), fmt::format_args());
+  fmt::format_context ctx(fmt::detail::buffer_appender<char>{buffer},
+                          fmt::format_args());
   fmt::format_specs spec;
   test_arg_formatter af(ctx, spec);
   fmt::visit_format_arg(
@@ -150,8 +150,9 @@ TEST(OStreamTest, WriteToOStreamMaxSize) {
   std::streamsize max_streamsize = fmt::detail::max_value<std::streamsize>();
   if (max_size <= fmt::detail::to_unsigned(max_streamsize)) return;
 
-  struct test_buffer : fmt::detail::buffer<char> {
-    explicit test_buffer(size_t size) { resize(size); }
+  struct test_buffer final : fmt::detail::buffer<char> {
+    explicit test_buffer(size_t size)
+      : fmt::detail::buffer<char>(nullptr, size, size) {}
     void grow(size_t) {}
   } buffer(max_size);
 
@@ -289,9 +290,20 @@ std::ostream& operator<<(std::ostream& os,
 TEST(OStreamTest, FormatExplicitlyConvertibleToStdStringView) {
   EXPECT_EQ("bar", fmt::format("{}", explicitly_convertible_to_string_like()));
 }
-
 #endif  // FMT_USE_STRING_VIEW
 
+struct streamable_and_convertible_to_bool {
+  operator bool() const { return true; }
+};
+
+std::ostream& operator<<(std::ostream& os, streamable_and_convertible_to_bool) {
+  return os << "foo";
+}
+
+TEST(OStreamTest, FormatConvertibleToBool) {
+  EXPECT_EQ("foo", fmt::format("{}", streamable_and_convertible_to_bool()));
+}
+
 struct copyfmt_test {};
 
 std::ostream& operator<<(std::ostream& os, copyfmt_test) {
@@ -307,3 +319,7 @@ TEST(OStreamTest, CopyFmt) {
 TEST(OStreamTest, CompileTimeString) {
   EXPECT_EQ("42", fmt::format(FMT_STRING("{}"), 42));
 }
+
+TEST(OStreamTest, ToString) {
+  EXPECT_EQ("ABC", fmt::to_string(fmt_test::ABC()));
+}
diff --git a/test/ranges-test.cc b/test/ranges-test.cc
index 46208e8d..63f9e6e6 100644
--- a/test/ranges-test.cc
+++ b/test/ranges-test.cc
@@ -54,7 +54,7 @@ TEST(RangesTest, FormatTuple) {
 TEST(RangesTest, JoinTuple) {
   // Value tuple args
   std::tuple<char, int, float> t1 = std::make_tuple('a', 1, 2.0f);
-  EXPECT_EQ("(a, 1, 2.0)", fmt::format("({})", fmt::join(t1, ", ")));
+  EXPECT_EQ("(a, 1, 2)", fmt::format("({})", fmt::join(t1, ", ")));
 
   // Testing lvalue tuple args
   int x = 4;
@@ -67,7 +67,7 @@ TEST(RangesTest, JoinTuple) {
 
   // Single element tuple
   std::tuple<float> t4{4.0f};
-  EXPECT_EQ("4.0", fmt::format("{}", fmt::join(t4, "/")));
+  EXPECT_EQ("4", fmt::format("{}", fmt::join(t4, "/")));
 }
 
 TEST(RangesTest, JoinInitializerList) {
@@ -141,15 +141,63 @@ TEST(RangesTest, FormatStringLike) {
 #endif  // FMT_USE_STRING_VIEW
 
 struct zstring_sentinel {};
+
 bool operator==(const char* p, zstring_sentinel) { return *p == '\0'; }
 bool operator!=(const char* p, zstring_sentinel) { return *p != '\0'; }
+
 struct zstring {
   const char* p;
   const char* begin() const { return p; }
   zstring_sentinel end() const { return {}; }
 };
+
 TEST(RangesTest, JoinSentinel) {
   zstring hello{"hello"};
   EXPECT_EQ("{'h', 'e', 'l', 'l', 'o'}", fmt::format("{}", hello));
   EXPECT_EQ("h_e_l_l_o", fmt::format("{}", fmt::join(hello, "_")));
 }
+
+// A range that provides non-const only begin()/end() to test fmt::join handles
+// that
+//
+// Some ranges (eg those produced by range-v3's views::filter()) can cache
+// information during iteration so they only provide non-const begin()/end().
+template <typename T> class non_const_only_range {
+ private:
+  std::vector<T> vec;
+
+ public:
+  using const_iterator = typename ::std::vector<T>::const_iterator;
+
+  template <typename... Args>
+  explicit non_const_only_range(Args&&... args)
+      : vec(::std::forward<Args>(args)...) {}
+
+  const_iterator begin() { return vec.begin(); }
+  const_iterator end() { return vec.end(); }
+};
+
+TEST(RangesTest, JoinRange) {
+  non_const_only_range<int> x(3u, 0);
+  EXPECT_EQ("0,0,0", fmt::format("{}", fmt::join(x, ",")));
+  EXPECT_EQ(
+      "0,0,0",
+      fmt::format("{}", fmt::join(non_const_only_range<int>(3u, 0), ",")));
+
+  std::vector<int> y(3u, 0);
+  EXPECT_EQ("0,0,0", fmt::format("{}", fmt::join(y, ",")));
+  EXPECT_EQ("0,0,0",
+            fmt::format("{}", fmt::join(std::vector<int>(3u, 0), ",")));
+
+  const std::vector<int> z(3u, 0);
+  EXPECT_EQ("0,0,0", fmt::format("{}", fmt::join(z, ",")));
+}
+
+#if !FMT_MSC_VER || FMT_MSC_VER >= 1927
+struct unformattable {};
+
+TEST(RangesTest, UnformattableRange) {
+  EXPECT_FALSE((fmt::has_formatter<std::vector<unformattable>,
+                                   fmt::format_context>::value));
+}
+#endif