diff --git a/CMakeLists.txt b/CMakeLists.txt
index 4c5cbe90..72640926 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -16,7 +16,7 @@ if("${CMAKE_BUILD_TYPE}" STREQUAL "")
 	set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-Wall -g -O2 -fPIC")
 endif()
 
-if(DEFINED USE_WASM)
+if(DEFINED USE_WASM)  # wasm build needs OPENCV_DIR defined
 	set(DISABLE_TESTS true)
 	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DGF256_TARGET_MOBILE")
 	set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "-Os")
@@ -28,10 +28,13 @@ if(DEFINED USE_WASM)
 		${OPENCV_DIR}/opencv-build-wasm/build_wasm/
 		${opencv_include_modules}
 	)
+else()  # if not wasm, go find opencv. 3 or 4 should both work
+	find_package(OpenCV REQUIRED)
+	include_directories(${OpenCV_INCLUDE_DIRS})
 endif()
 
 if(NOT DEFINED OPENCV_LIBS)
-	set(OPENCV_LIBS "opencv_core" "opencv_imgcodecs" "opencv_imgproc" "opencv_photo")
+	set(OPENCV_LIBS "opencv_calib3d" "opencv_imgcodecs" "opencv_imgproc" "opencv_photo" "opencv_core")
 endif()
 
 if(NOT DEFINED CPPFILESYSTEM)
diff --git a/src/exe/cimbar_send/send.cpp b/src/exe/cimbar_send/send.cpp
index f82c9e67..4cb3c404 100644
--- a/src/exe/cimbar_send/send.cpp
+++ b/src/exe/cimbar_send/send.cpp
@@ -70,7 +70,7 @@ int main(int argc, char** argv)
 	if (!initialize_GL(window_size, window_size))
 	{
 		std::cerr << "failed to create window :(" << std::endl;
-		return 50;
+		return 70;
 	}
 
 	configure(colorBits, ecc, compressionLevel);
diff --git a/src/lib/cimb_translator/CellPositions.h b/src/lib/cimb_translator/CellPositions.h
index ca446d1e..ad8654b6 100644
--- a/src/lib/cimb_translator/CellPositions.h
+++ b/src/lib/cimb_translator/CellPositions.h
@@ -1,8 +1,9 @@
 /* This code is subject to the terms of the Mozilla Public License, v.2.0. http://mozilla.org/MPL/2.0/. */
 #pragma once
 
-#include <vector>
+#include <cstddef>
 #include <utility>
+#include <vector>
 
 class CellPositions
 {
diff --git a/src/lib/compression/zstd_compressor.h b/src/lib/compression/zstd_compressor.h
index 8e32dd87..6eee6e79 100644
--- a/src/lib/compression/zstd_compressor.h
+++ b/src/lib/compression/zstd_compressor.h
@@ -32,7 +32,7 @@ class zstd_compressor : public STREAM
 			size_t compressedBytes = ZSTD_compressCCtx(_cctx, _compBuff.data(), _compBuff.size(), data, writeLen, _compressionLevel);
 			if (ZSTD_isError(compressedBytes))
 			{
-				std::cout << "error? " << ZSTD_getErrorName(compressedBytes) << std::endl;
+				std::cerr << "error? " << ZSTD_getErrorName(compressedBytes) << std::endl;
 				return false;
 			}
 			STREAM::write(_compBuff.data(), compressedBytes);
diff --git a/src/lib/encoder/test/DecoderTest.cpp b/src/lib/encoder/test/DecoderTest.cpp
index 3895dc0e..d88fd4cd 100644
--- a/src/lib/encoder/test/DecoderTest.cpp
+++ b/src/lib/encoder/test/DecoderTest.cpp
@@ -56,5 +56,8 @@ TEST_CASE( "DecoderTest/testDecode.Sample", "[unit]" )
 	unsigned bytesDecoded = dec.decode(TestCimbar::getSample("6bit/4_30_f0_627_extract.jpg"), decodedFile);
 	assertEquals( 9300, bytesDecoded );
 
-	assertEquals( "3de927c8aa0221807a2784210160cdc17567eb587bf01233d166900aadf14bf5", get_hash(decodedFile) );
+	if (CV_VERSION_MAJOR == 3)
+		assertEquals( "3de927c8aa0221807a2784210160cdc17567eb587bf01233d166900aadf14bf5", get_hash(decodedFile) );
+	else // # cv4
+		assertEquals( "59ddb2516b4ff5a528aebe538a22b736a6714263a454d20e146e1ffbba36c5ae", get_hash(decodedFile) );
 }
diff --git a/src/lib/fountain/FountainMetadata.h b/src/lib/fountain/FountainMetadata.h
index 5ba0ce8f..2d64ddfe 100644
--- a/src/lib/fountain/FountainMetadata.h
+++ b/src/lib/fountain/FountainMetadata.h
@@ -15,7 +15,7 @@ class FountainMetadata
 	}
 
 public:
-	FountainMetadata(uint64_t id)
+	FountainMetadata(uint32_t id)
 	    : _data(id)
 	{
 	}
@@ -33,7 +33,7 @@ class FountainMetadata
 		to_uint8_arr(encode_id, size, d);
 	}
 
-	unsigned id() const
+	uint32_t id() const
 	{
 		return _data;
 	}
@@ -65,5 +65,5 @@ class FountainMetadata
 	}
 
 protected:
-	uint64_t _data; // might invert this and only generate the uint64_t when we need it
+	uint32_t _data; // might invert this and only generate the uint32_t when we need it
 };
diff --git a/src/lib/fountain/fountain_decoder_sink.h b/src/lib/fountain/fountain_decoder_sink.h
index 711538c1..3fcb10dc 100644
--- a/src/lib/fountain/fountain_decoder_sink.h
+++ b/src/lib/fountain/fountain_decoder_sink.h
@@ -39,10 +39,10 @@ class fountain_decoder_sink
 		return true;
 	}
 
-	void mark_done(uint64_t id)
+	void mark_done(const FountainMetadata& md)
 	{
-		_done.insert(id);
-		auto it = _streams.find(id);
+		_done.insert(md.id());
+		auto it = _streams.find(stream_slot(md));
 		if (it != _streams.end())
 			_streams.erase(it);
 	}
@@ -60,7 +60,7 @@ class fountain_decoder_sink
 	std::vector<std::string> get_done() const
 	{
 		std::vector<std::string> done;
-		for (uint64_t id : _done)
+		for (uint32_t id : _done)
 			done.push_back( get_filename(FountainMetadata(id)) );
 		return done;
 	}
@@ -77,7 +77,7 @@ class fountain_decoder_sink
 		return progress;
 	}
 
-	bool is_done(uint64_t id) const
+	bool is_done(uint32_t id) const
 	{
 		return _done.find(id) != _done.end();
 	}
@@ -106,7 +106,7 @@ class fountain_decoder_sink
 			return false;
 
 		if (store(md, *finished))
-			mark_done(md.id());
+			mark_done(md);
 		return true;
 	}
 
@@ -122,7 +122,7 @@ class fountain_decoder_sink
 	}
 
 protected:
-	// streams is limited to at most 8 decoders at a time. Current, we just use the lower bits of the encode_id.
+	// streams is limited to at most 8 decoders at a time. Currently, we just use the lower bits of the encode_id.
 	uint8_t stream_slot(const FountainMetadata& md) const
 	{
 		return md.encode_id() & 0x7;
@@ -137,7 +137,10 @@ class fountain_decoder_sink
 	std::string _dataDir;
 	unsigned _chunkSize;
 
+	// maybe instead of unordered_map+set, something where we can "age out" old streams?
+	// e.g. most recent 16/8, or something?
+	// question is what happens to _done/_streams when we wrap for continuous data streaming...
 	std::unordered_map<uint8_t, fountain_decoder_stream> _streams;
-	// track the uint64_t combo of (encode_id,size) to avoid redundant work
-	std::set<uint64_t> _done;
+	// track the uint32_t combo of (encode_id,size) to avoid redundant work
+	std::set<uint32_t> _done;
 };
diff --git a/src/lib/image_hash/ahash_result.h b/src/lib/image_hash/ahash_result.h
index d4c8e2c1..6a90cb46 100644
--- a/src/lib/image_hash/ahash_result.h
+++ b/src/lib/image_hash/ahash_result.h
@@ -2,7 +2,7 @@
 #pragma once
 
 #include "bit_extractor.h"
-#include "intx/int128.hpp"
+#include "intx/intx.hpp"
 #include <array>
 #include <iostream>
 #include <utility>
diff --git a/src/lib/image_hash/average_hash.h b/src/lib/image_hash/average_hash.h
index 06706a83..3eb9c33d 100644
--- a/src/lib/image_hash/average_hash.h
+++ b/src/lib/image_hash/average_hash.h
@@ -6,7 +6,7 @@
 #include "bit_file/bitmatrix.h"
 #include "cimb_translator/Cell.h"
 
-#include "intx/int128.hpp"
+#include "intx/intx.hpp"
 #include <opencv2/opencv.hpp>
 
 #include <array>
diff --git a/src/lib/image_hash/test/bitExtractorTest.cpp b/src/lib/image_hash/test/bitExtractorTest.cpp
index f5f65844..9848b955 100644
--- a/src/lib/image_hash/test/bitExtractorTest.cpp
+++ b/src/lib/image_hash/test/bitExtractorTest.cpp
@@ -2,7 +2,7 @@
 #include "unittest.h"
 
 #include "bit_extractor.h"
-#include "intx/int128.hpp"
+#include "intx/intx.hpp"
 
 #include <bitset>
 #include <iostream>
@@ -36,7 +36,9 @@ TEST_CASE( "bitExtractorTest/testLargerValue.1", "[unit]" )
 
 TEST_CASE( "bitExtractorTest/testLargerValue.2", "[unit]" )
 {
-	intx::uint128 bits{0xFFBFCFE3FULL, 0xF83C0E030080000ULL};
+	intx::uint128 bits{0xF83C0E030080000ULL, 0xFFBFCFE3FULL};
+	assertEquals( "ffbfcfe3f0f83c0e030080000", intx::hex(bits) );  // sanity check
+
 	bit_extractor<intx::uint128, 100> be(bits);
 	uint64_t res = be.extract(1, 11, 21, 31);
 	assertEquals( 0xfffefcf8, res );
diff --git a/src/third_party_lib/PicoSHA2/picosha2.h b/src/third_party_lib/PicoSHA2/picosha2.h
index bc00c743..a921736b 100644
--- a/src/third_party_lib/PicoSHA2/picosha2.h
+++ b/src/third_party_lib/PicoSHA2/picosha2.h
@@ -93,7 +93,7 @@ void hash256_block(RaIter1 message_digest, RaIter2 first, RaIter2 last) {
     assert(first + 64 == last);
     static_cast<void>(last);  // for avoiding unused-variable warning
     word_t w[64];
-    std::fill(w, w + 64, 0);
+    std::fill(w, w + 64, word_t(0));
     for (std::size_t i = 0; i < 16; ++i) {
         w[i] = (static_cast<word_t>(mask_8bit(*(first + i * 4))) << 24) |
                (static_cast<word_t>(mask_8bit(*(first + i * 4 + 1))) << 16) |
@@ -185,7 +185,7 @@ class hash256_one_by_one {
 
     void init() {
         buffer_.clear();
-        std::fill(data_length_digits_, data_length_digits_ + 4, 0);
+        std::fill(data_length_digits_, data_length_digits_ + 4, word_t(0));
         std::copy(detail::initial_message_digest,
                   detail::initial_message_digest + 8, h_);
     }
@@ -204,17 +204,17 @@ class hash256_one_by_one {
 
     void finish() {
         byte_t temp[64];
-        std::fill(temp, temp + 64, 0);
+        std::fill(temp, temp + 64, byte_t(0));
         std::size_t remains = buffer_.size();
         std::copy(buffer_.begin(), buffer_.end(), temp);
         temp[remains] = 0x80;
 
         if (remains > 55) {
-            std::fill(temp + remains + 1, temp + 64, 0);
+            std::fill(temp + remains + 1, temp + 64, byte_t(0));
             detail::hash256_block(h_, temp, temp + 64);
-            std::fill(temp, temp + 64 - 4, 0);
+            std::fill(temp, temp + 64 - 4, byte_t(0));
         } else {
-            std::fill(temp + remains + 1, temp + 64 - 4, 0);
+            std::fill(temp + remains + 1, temp + 64 - 4, byte_t(0));
         }
 
         write_data_bit_length(&(temp[56]));
diff --git a/src/third_party_lib/concurrentqueue/LICENSE.md b/src/third_party_lib/concurrentqueue/LICENSE.md
index 47efd439..51933897 100644
--- a/src/third_party_lib/concurrentqueue/LICENSE.md
+++ b/src/third_party_lib/concurrentqueue/LICENSE.md
@@ -1,6 +1,7 @@
 This license file applies to everything in this repository except that which
 is explicitly annotated as being written by other authors, i.e. the Boost
 queue (included in the benchmarks for comparison), Intel's TBB library (ditto),
+dlib::pipe (ditto),
 the CDSChecker tool (used for verification), the Relacy model checker (ditto),
 and Jeff Preshing's semaphore implementation (used in the blocking queue) which
 has a zlib license (embedded in lightweightsempahore.h).
diff --git a/src/third_party_lib/concurrentqueue/concurrentqueue.h b/src/third_party_lib/concurrentqueue/concurrentqueue.h
index 830b8193..ff3156f7 100644
--- a/src/third_party_lib/concurrentqueue/concurrentqueue.h
+++ b/src/third_party_lib/concurrentqueue/concurrentqueue.h
@@ -1688,7 +1688,7 @@ class ConcurrentQueue
 		{
 		}
 		
-		virtual ~ProducerBase() { };
+		virtual ~ProducerBase() { }
 		
 		template<typename U>
 		inline bool dequeue(U& element)
@@ -1897,7 +1897,7 @@ class ConcurrentQueue
 					++pr_blockIndexSlotsUsed;
 				}
 
-				MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new ((T*)nullptr) T(std::forward<U>(element)))) {
+				MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast<T*>(nullptr)) T(std::forward<U>(element)))) {
 					// The constructor may throw. We want the element not to appear in the queue in
 					// that case (without corrupting the queue):
 					MOODYCAMEL_TRY {
@@ -1923,7 +1923,7 @@ class ConcurrentQueue
 				blockIndex.load(std::memory_order_relaxed)->front.store(pr_blockIndexFront, std::memory_order_release);
 				pr_blockIndexFront = (pr_blockIndexFront + 1) & (pr_blockIndexSize - 1);
 				
-				MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new ((T*)nullptr) T(std::forward<U>(element)))) {
+				MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast<T*>(nullptr)) T(std::forward<U>(element)))) {
 					this->tailIndex.store(newTailIndex, std::memory_order_release);
 					return true;
 				}
@@ -2139,7 +2139,7 @@ class ConcurrentQueue
 					block = block->next;
 				}
 				
-				MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))) {
+				MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast<T*>(nullptr)) T(details::deref_noexcept(itemFirst)))) {
 					blockIndex.load(std::memory_order_relaxed)->front.store((pr_blockIndexFront - 1) & (pr_blockIndexSize - 1), std::memory_order_release);
 				}
 			}
@@ -2158,7 +2158,7 @@ class ConcurrentQueue
 				if (details::circular_less_than<index_t>(newTailIndex, stopIndex)) {
 					stopIndex = newTailIndex;
 				}
-				MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))) {
+				MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast<T*>(nullptr)) T(details::deref_noexcept(itemFirst)))) {
 					while (currentTailIndex != stopIndex) {
 						new ((*this->tailBlock)[currentTailIndex++]) T(*itemFirst++);
 					}
@@ -2173,7 +2173,7 @@ class ConcurrentQueue
 							// may only define a (noexcept) move constructor, and so calls to the
 							// cctor will not compile, even if they are in an if branch that will never
 							// be executed
-							new ((*this->tailBlock)[currentTailIndex]) T(details::nomove_if<(bool)!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))>::eval(*itemFirst));
+							new ((*this->tailBlock)[currentTailIndex]) T(details::nomove_if<!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast<T*>(nullptr)) T(details::deref_noexcept(itemFirst)))>::eval(*itemFirst));
 							++currentTailIndex;
 							++itemFirst;
 						}
@@ -2220,7 +2220,7 @@ class ConcurrentQueue
 				this->tailBlock = this->tailBlock->next;
 			}
 			
-			MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))) {
+			MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast<T*>(nullptr)) T(details::deref_noexcept(itemFirst)))) {
 				if (firstAllocatedBlock != nullptr)
 					blockIndex.load(std::memory_order_relaxed)->front.store((pr_blockIndexFront - 1) & (pr_blockIndexSize - 1), std::memory_order_release);
 			}
@@ -2239,7 +2239,7 @@ class ConcurrentQueue
 				desiredCount = desiredCount < max ? desiredCount : max;
 				std::atomic_thread_fence(std::memory_order_acquire);
 				
-				auto myDequeueCount = this->dequeueOptimisticCount.fetch_add(desiredCount, std::memory_order_relaxed);;
+				auto myDequeueCount = this->dequeueOptimisticCount.fetch_add(desiredCount, std::memory_order_relaxed);
 				
 				tail = this->tailIndex.load(std::memory_order_acquire);
 				auto actualCount = static_cast<size_t>(tail - (myDequeueCount - overcommit));
@@ -2501,7 +2501,7 @@ class ConcurrentQueue
 #endif
 				newBlock->ConcurrentQueue::Block::template reset_empty<implicit_context>();
 
-				MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new ((T*)nullptr) T(std::forward<U>(element)))) {
+				MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast<T*>(nullptr)) T(std::forward<U>(element)))) {
 					// May throw, try to insert now before we publish the fact that we have this new block
 					MOODYCAMEL_TRY {
 						new ((*newBlock)[currentTailIndex]) T(std::forward<U>(element));
@@ -2519,7 +2519,7 @@ class ConcurrentQueue
 				
 				this->tailBlock = newBlock;
 				
-				MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new ((T*)nullptr) T(std::forward<U>(element)))) {
+				MOODYCAMEL_CONSTEXPR_IF (!MOODYCAMEL_NOEXCEPT_CTOR(T, U, new (static_cast<T*>(nullptr)) T(std::forward<U>(element)))) {
 					this->tailIndex.store(newTailIndex, std::memory_order_release);
 					return true;
 				}
@@ -2697,7 +2697,7 @@ class ConcurrentQueue
 				if (details::circular_less_than<index_t>(newTailIndex, stopIndex)) {
 					stopIndex = newTailIndex;
 				}
-				MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))) {
+				MOODYCAMEL_CONSTEXPR_IF (MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast<T*>(nullptr)) T(details::deref_noexcept(itemFirst)))) {
 					while (currentTailIndex != stopIndex) {
 						new ((*this->tailBlock)[currentTailIndex++]) T(*itemFirst++);
 					}
@@ -2705,7 +2705,7 @@ class ConcurrentQueue
 				else {
 					MOODYCAMEL_TRY {
 						while (currentTailIndex != stopIndex) {
-							new ((*this->tailBlock)[currentTailIndex]) T(details::nomove_if<(bool)!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new ((T*)nullptr) T(details::deref_noexcept(itemFirst)))>::eval(*itemFirst));
+							new ((*this->tailBlock)[currentTailIndex]) T(details::nomove_if<!MOODYCAMEL_NOEXCEPT_CTOR(T, decltype(*itemFirst), new (static_cast<T*>(nullptr)) T(details::deref_noexcept(itemFirst)))>::eval(*itemFirst));
 							++currentTailIndex;
 							++itemFirst;
 						}
@@ -3459,7 +3459,7 @@ class ConcurrentQueue
 					}
 					
 					auto newHash = new (raw) ImplicitProducerHash;
-					newHash->capacity = (size_t)newCapacity;
+					newHash->capacity = static_cast<size_t>(newCapacity);
 					newHash->entries = reinterpret_cast<ImplicitProducerKVP*>(details::align_for<ImplicitProducerKVP>(raw + sizeof(ImplicitProducerHash)));
 					for (size_t i = 0; i != newCapacity; ++i) {
 						new (newHash->entries + i) ImplicitProducerKVP;
@@ -3698,7 +3698,7 @@ ConsumerToken::ConsumerToken(ConcurrentQueue<T, Traits>& queue)
 	: itemsConsumedFromCurrent(0), currentProducer(nullptr), desiredProducer(nullptr)
 {
 	initialOffset = queue.nextExplicitConsumerId.fetch_add(1, std::memory_order_release);
-	lastKnownGlobalOffset = (std::uint32_t)-1;
+	lastKnownGlobalOffset = static_cast<std::uint32_t>(-1);
 }
 
 template<typename T, typename Traits>
@@ -3706,7 +3706,7 @@ ConsumerToken::ConsumerToken(BlockingConcurrentQueue<T, Traits>& queue)
 	: itemsConsumedFromCurrent(0), currentProducer(nullptr), desiredProducer(nullptr)
 {
 	initialOffset = reinterpret_cast<ConcurrentQueue<T, Traits>*>(&queue)->nextExplicitConsumerId.fetch_add(1, std::memory_order_release);
-	lastKnownGlobalOffset = (std::uint32_t)-1;
+	lastKnownGlobalOffset = static_cast<std::uint32_t>(-1);
 }
 
 template<typename T, typename Traits>
diff --git a/src/third_party_lib/fmt/core.h b/src/third_party_lib/fmt/core.h
index 0a81e0cc..92a7aa1d 100644
--- a/src/third_party_lib/fmt/core.h
+++ b/src/third_party_lib/fmt/core.h
@@ -1,4 +1,4 @@
-// Formatting library for C++ - the core API
+// Formatting library for C++ - the core API for char/UTF-8
 //
 // Copyright (c) 2012 - present, Victor Zverovich
 // All rights reserved.
@@ -8,40 +8,45 @@
 #ifndef FMT_CORE_H_
 #define FMT_CORE_H_
 
-#include <cstdio>  // std::FILE
+#include <cstddef>  // std::byte
+#include <cstdio>   // std::FILE
 #include <cstring>
-#include <functional>
 #include <iterator>
-#include <memory>
+#include <limits>
 #include <string>
 #include <type_traits>
-#include <vector>
 
 // The fmt library version in the form major * 10000 + minor * 100 + patch.
-#define FMT_VERSION 70103
+#define FMT_VERSION 80101
 
-#ifdef __clang__
+#if defined(__clang__) && !defined(__ibmxl__)
 #  define FMT_CLANG_VERSION (__clang_major__ * 100 + __clang_minor__)
 #else
 #  define FMT_CLANG_VERSION 0
 #endif
 
-#if defined(__GNUC__) && !defined(__clang__)
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) && \
+    !defined(__NVCOMPILER)
 #  define FMT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
 #else
 #  define FMT_GCC_VERSION 0
 #endif
 
-#if defined(__INTEL_COMPILER)
-#  define FMT_ICC_VERSION __INTEL_COMPILER
-#else
-#  define FMT_ICC_VERSION 0
+#ifndef FMT_GCC_PRAGMA
+// Workaround _Pragma bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59884.
+#  if FMT_GCC_VERSION >= 504
+#    define FMT_GCC_PRAGMA(arg) _Pragma(arg)
+#  else
+#    define FMT_GCC_PRAGMA(arg)
+#  endif
 #endif
 
-#if __cplusplus >= 201103L || defined(__GXX_EXPERIMENTAL_CXX0X__)
-#  define FMT_HAS_GXX_CXX11 FMT_GCC_VERSION
+#ifdef __ICL
+#  define FMT_ICC_VERSION __ICL
+#elif defined(__INTEL_COMPILER)
+#  define FMT_ICC_VERSION __INTEL_COMPILER
 #else
-#  define FMT_HAS_GXX_CXX11 0
+#  define FMT_ICC_VERSION 0
 #endif
 
 #ifdef __NVCC__
@@ -52,10 +57,10 @@
 
 #ifdef _MSC_VER
 #  define FMT_MSC_VER _MSC_VER
-#  define FMT_SUPPRESS_MSC_WARNING(n) __pragma(warning(suppress : n))
+#  define FMT_MSC_WARNING(...) __pragma(warning(__VA_ARGS__))
 #else
 #  define FMT_MSC_VER 0
-#  define FMT_SUPPRESS_MSC_WARNING(n)
+#  define FMT_MSC_WARNING(...)
 #endif
 
 #ifdef __has_feature
@@ -64,7 +69,8 @@
 #  define FMT_HAS_FEATURE(x) 0
 #endif
 
-#if defined(__has_include) && !defined(__INTELLISENSE__) && \
+#if defined(__has_include) &&                             \
+    (!defined(__INTELLISENSE__) || FMT_MSC_VER > 1900) && \
     (!FMT_ICC_VERSION || FMT_ICC_VERSION >= 1600)
 #  define FMT_HAS_INCLUDE(x) __has_include(x)
 #else
@@ -77,17 +83,23 @@
 #  define FMT_HAS_CPP_ATTRIBUTE(x) 0
 #endif
 
+#ifdef _MSVC_LANG
+#  define FMT_CPLUSPLUS _MSVC_LANG
+#else
+#  define FMT_CPLUSPLUS __cplusplus
+#endif
+
 #define FMT_HAS_CPP14_ATTRIBUTE(attribute) \
-  (__cplusplus >= 201402L && FMT_HAS_CPP_ATTRIBUTE(attribute))
+  (FMT_CPLUSPLUS >= 201402L && FMT_HAS_CPP_ATTRIBUTE(attribute))
 
 #define FMT_HAS_CPP17_ATTRIBUTE(attribute) \
-  (__cplusplus >= 201703L && FMT_HAS_CPP_ATTRIBUTE(attribute))
+  (FMT_CPLUSPLUS >= 201703L && FMT_HAS_CPP_ATTRIBUTE(attribute))
 
 // Check if relaxed C++14 constexpr is supported.
 // GCC doesn't allow throw in constexpr until version 6 (bug 67371).
 #ifndef FMT_USE_CONSTEXPR
 #  define FMT_USE_CONSTEXPR                                           \
-    (FMT_HAS_FEATURE(cxx_relaxed_constexpr) || FMT_MSC_VER >= 1910 || \
+    (FMT_HAS_FEATURE(cxx_relaxed_constexpr) || FMT_MSC_VER >= 1912 || \
      (FMT_GCC_VERSION >= 600 && __cplusplus >= 201402L)) &&           \
         !FMT_NVCC && !FMT_ICC_VERSION
 #endif
@@ -95,17 +107,32 @@
 #  define FMT_CONSTEXPR constexpr
 #  define FMT_CONSTEXPR_DECL constexpr
 #else
-#  define FMT_CONSTEXPR inline
+#  define FMT_CONSTEXPR
 #  define FMT_CONSTEXPR_DECL
 #endif
 
-#ifndef FMT_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
-#    define FMT_OVERRIDE
+#if ((__cplusplus >= 202002L) &&                              \
+     (!defined(_GLIBCXX_RELEASE) || _GLIBCXX_RELEASE > 9)) || \
+    (__cplusplus >= 201709L && FMT_GCC_VERSION >= 1002)
+#  define FMT_CONSTEXPR20 constexpr
+#else
+#  define FMT_CONSTEXPR20
+#endif
+
+// Check if constexpr std::char_traits<>::compare,length is supported.
+#if defined(__GLIBCXX__)
+#  if __cplusplus >= 201703L && defined(_GLIBCXX_RELEASE) && \
+      _GLIBCXX_RELEASE >= 7  // GCC 7+ libstdc++ has _GLIBCXX_RELEASE.
+#    define FMT_CONSTEXPR_CHAR_TRAITS constexpr
 #  endif
+#elif defined(_LIBCPP_VERSION) && __cplusplus >= 201703L && \
+    _LIBCPP_VERSION >= 4000
+#  define FMT_CONSTEXPR_CHAR_TRAITS constexpr
+#elif FMT_MSC_VER >= 1914 && _MSVC_LANG >= 201703L
+#  define FMT_CONSTEXPR_CHAR_TRAITS constexpr
+#endif
+#ifndef FMT_CONSTEXPR_CHAR_TRAITS
+#  define FMT_CONSTEXPR_CHAR_TRAITS
 #endif
 
 // Check if exceptions are disabled.
@@ -124,7 +151,7 @@
 #endif
 
 #if FMT_USE_NOEXCEPT || FMT_HAS_FEATURE(cxx_noexcept) || \
-    (FMT_GCC_VERSION >= 408 && FMT_HAS_GXX_CXX11) || FMT_MSC_VER >= 1900
+    FMT_GCC_VERSION >= 408 || FMT_MSC_VER >= 1900
 #  define FMT_DETECTED_NOEXCEPT noexcept
 #  define FMT_HAS_CXX11_NOEXCEPT 1
 #else
@@ -149,25 +176,39 @@
 #  define FMT_NORETURN
 #endif
 
-#ifndef FMT_DEPRECATED
-#  if FMT_HAS_CPP14_ATTRIBUTE(deprecated) || FMT_MSC_VER >= 1900
-#    define FMT_DEPRECATED [[deprecated]]
+#if __cplusplus == 201103L || __cplusplus == 201402L
+#  if defined(__INTEL_COMPILER) || defined(__PGI)
+#    define FMT_FALLTHROUGH
+#  elif defined(__clang__)
+#    define FMT_FALLTHROUGH [[clang::fallthrough]]
+#  elif FMT_GCC_VERSION >= 700 && \
+      (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 520)
+#    define FMT_FALLTHROUGH [[gnu::fallthrough]]
 #  else
-#    if (defined(__GNUC__) && !defined(__LCC__)) || defined(__clang__)
-#      define FMT_DEPRECATED __attribute__((deprecated))
-#    elif FMT_MSC_VER
-#      define FMT_DEPRECATED __declspec(deprecated)
-#    else
-#      define FMT_DEPRECATED /* deprecated */
-#    endif
+#    define FMT_FALLTHROUGH
 #  endif
+#elif FMT_HAS_CPP17_ATTRIBUTE(fallthrough)
+#  define FMT_FALLTHROUGH [[fallthrough]]
+#else
+#  define FMT_FALLTHROUGH
 #endif
 
-// Workaround broken [[deprecated]] in the Intel, PGI and NVCC compilers.
-#if FMT_ICC_VERSION || defined(__PGI) || FMT_NVCC
-#  define FMT_DEPRECATED_ALIAS
-#else
-#  define FMT_DEPRECATED_ALIAS FMT_DEPRECATED
+#ifndef FMT_NODISCARD
+#  if FMT_HAS_CPP17_ATTRIBUTE(nodiscard)
+#    define FMT_NODISCARD [[nodiscard]]
+#  else
+#    define FMT_NODISCARD
+#  endif
+#endif
+
+#ifndef FMT_USE_FLOAT
+#  define FMT_USE_FLOAT 1
+#endif
+#ifndef FMT_USE_DOUBLE
+#  define FMT_USE_DOUBLE 1
+#endif
+#ifndef FMT_USE_LONG_DOUBLE
+#  define FMT_USE_LONG_DOUBLE 1
 #endif
 
 #ifndef FMT_INLINE
@@ -178,61 +219,55 @@
 #  endif
 #endif
 
-#ifndef FMT_USE_INLINE_NAMESPACES
-#  if FMT_HAS_FEATURE(cxx_inline_namespaces) || FMT_GCC_VERSION >= 404 || \
-      (FMT_MSC_VER >= 1900 && !_MANAGED)
-#    define FMT_USE_INLINE_NAMESPACES 1
+#ifndef FMT_DEPRECATED
+#  if FMT_HAS_CPP14_ATTRIBUTE(deprecated) || FMT_MSC_VER >= 1900
+#    define FMT_DEPRECATED [[deprecated]]
 #  else
-#    define FMT_USE_INLINE_NAMESPACES 0
+#    if (defined(__GNUC__) && !defined(__LCC__)) || defined(__clang__)
+#      define FMT_DEPRECATED __attribute__((deprecated))
+#    elif FMT_MSC_VER
+#      define FMT_DEPRECATED __declspec(deprecated)
+#    else
+#      define FMT_DEPRECATED /* deprecated */
+#    endif
 #  endif
 #endif
 
 #ifndef FMT_BEGIN_NAMESPACE
-#  if FMT_USE_INLINE_NAMESPACES
-#    define FMT_INLINE_NAMESPACE inline namespace
-#    define FMT_END_NAMESPACE \
-      }                       \
-      }
-#  else
-#    define FMT_INLINE_NAMESPACE namespace
-#    define FMT_END_NAMESPACE \
-      }                       \
-      using namespace v7;     \
-      }
-#  endif
 #  define FMT_BEGIN_NAMESPACE \
     namespace fmt {           \
-    FMT_INLINE_NAMESPACE v7 {
+    inline namespace v8 {
+#  define FMT_END_NAMESPACE \
+    }                       \
+    }
+#endif
+
+#ifndef FMT_MODULE_EXPORT
+#  define FMT_MODULE_EXPORT
+#  define FMT_MODULE_EXPORT_BEGIN
+#  define FMT_MODULE_EXPORT_END
+#  define FMT_BEGIN_DETAIL_NAMESPACE namespace detail {
+#  define FMT_END_DETAIL_NAMESPACE }
 #endif
 
 #if !defined(FMT_HEADER_ONLY) && defined(_WIN32)
-#  define FMT_CLASS_API FMT_SUPPRESS_MSC_WARNING(4275)
+#  define FMT_CLASS_API FMT_MSC_WARNING(suppress : 4275)
 #  ifdef FMT_EXPORT
 #    define FMT_API __declspec(dllexport)
-#    define FMT_EXTERN_TEMPLATE_API FMT_API
-#    define FMT_EXPORTED
 #  elif defined(FMT_SHARED)
 #    define FMT_API __declspec(dllimport)
-#    define FMT_EXTERN_TEMPLATE_API FMT_API
 #  endif
 #else
 #  define FMT_CLASS_API
+#  if defined(FMT_EXPORT) || defined(FMT_SHARED)
+#    if defined(__GNUC__) || defined(__clang__)
+#      define FMT_API __attribute__((visibility("default")))
+#    endif
+#  endif
 #endif
 #ifndef FMT_API
 #  define FMT_API
 #endif
-#ifndef FMT_EXTERN_TEMPLATE_API
-#  define FMT_EXTERN_TEMPLATE_API
-#endif
-#ifndef FMT_INSTANTIATION_DEF_API
-#  define FMT_INSTANTIATION_DEF_API FMT_API
-#endif
-
-#ifndef FMT_HEADER_ONLY
-#  define FMT_EXTERN extern
-#else
-#  define FMT_EXTERN
-#endif
 
 // libc++ supports string_view in pre-c++17.
 #if (FMT_HAS_INCLUDE(<string_view>) &&                       \
@@ -248,16 +283,43 @@
 #ifndef FMT_UNICODE
 #  define FMT_UNICODE !FMT_MSC_VER
 #endif
-#if FMT_UNICODE && FMT_MSC_VER
-#  pragma execution_character_set("utf-8")
+
+#ifndef FMT_CONSTEVAL
+#  if ((FMT_GCC_VERSION >= 1000 || FMT_CLANG_VERSION >= 1101) &&      \
+       __cplusplus > 201703L && !defined(__apple_build_version__)) || \
+      (defined(__cpp_consteval) &&                                    \
+       (!FMT_MSC_VER || _MSC_FULL_VER >= 193030704))
+// consteval is broken in MSVC before VS2022 and Apple clang 13.
+#    define FMT_CONSTEVAL consteval
+#    define FMT_HAS_CONSTEVAL
+#  else
+#    define FMT_CONSTEVAL
+#  endif
+#endif
+
+#ifndef FMT_USE_NONTYPE_TEMPLATE_PARAMETERS
+#  if defined(__cpp_nontype_template_args) &&                \
+      ((FMT_GCC_VERSION >= 903 && __cplusplus >= 201709L) || \
+       __cpp_nontype_template_args >= 201911L)
+#    define FMT_USE_NONTYPE_TEMPLATE_PARAMETERS 1
+#  else
+#    define FMT_USE_NONTYPE_TEMPLATE_PARAMETERS 0
+#  endif
+#endif
+
+// Enable minimal optimizations for more compact code in debug mode.
+FMT_GCC_PRAGMA("GCC push_options")
+#ifndef __OPTIMIZE__
+FMT_GCC_PRAGMA("GCC optimize(\"Og\")")
 #endif
 
 FMT_BEGIN_NAMESPACE
+FMT_MODULE_EXPORT_BEGIN
 
 // Implementations of enable_if_t and other metafunctions for older systems.
-template <bool B, class T = void>
+template <bool B, typename T = void>
 using enable_if_t = typename std::enable_if<B, T>::type;
-template <bool B, class T, class F>
+template <bool B, typename T, typename F>
 using conditional_t = typename std::conditional<B, T, F>::type;
 template <bool B> using bool_constant = std::integral_constant<bool, B>;
 template <typename T>
@@ -269,17 +331,40 @@ using remove_cvref_t = typename std::remove_cv<remove_reference_t<T>>::type;
 template <typename T> struct type_identity { using type = T; };
 template <typename T> using type_identity_t = typename type_identity<T>::type;
 
-struct monostate {};
+struct monostate {
+  constexpr monostate() {}
+};
 
 // An enable_if helper to be used in template parameters which results in much
 // shorter symbols: https://godbolt.org/z/sWw4vP. Extra parentheses are needed
 // to workaround a bug in MSVC 2019 (see #1140 and #1186).
-#define FMT_ENABLE_IF(...) enable_if_t<(__VA_ARGS__), int> = 0
+#ifdef FMT_DOC
+#  define FMT_ENABLE_IF(...)
+#else
+#  define FMT_ENABLE_IF(...) enable_if_t<(__VA_ARGS__), int> = 0
+#endif
+
+FMT_BEGIN_DETAIL_NAMESPACE
 
-namespace detail {
+// Suppress "unused variable" warnings with the method described in
+// https://herbsutter.com/2009/10/18/mailbag-shutting-up-compiler-warnings/.
+// (void)var does not work on many Intel compilers.
+template <typename... T> FMT_CONSTEXPR void ignore_unused(const T&...) {}
+
+constexpr FMT_INLINE auto is_constant_evaluated(bool default_value = false)
+    FMT_NOEXCEPT -> bool {
+#ifdef __cpp_lib_is_constant_evaluated
+  ignore_unused(default_value);
+  return std::is_constant_evaluated();
+#else
+  return default_value;
+#endif
+}
 
-// A helper function to suppress "conditional expression is constant" warnings.
-template <typename T> constexpr T const_check(T value) { return value; }
+// A function to suppress "conditional expression is constant" warnings.
+template <typename T> constexpr FMT_INLINE auto const_check(T value) -> T {
+  return value;
+}
 
 FMT_NORETURN FMT_API void assert_fail(const char* file, int line,
                                       const char* message);
@@ -287,7 +372,8 @@ FMT_NORETURN FMT_API void assert_fail(const char* file, int line,
 #ifndef FMT_ASSERT
 #  ifdef NDEBUG
 // FMT_ASSERT is not empty to avoid -Werror=empty-body.
-#    define FMT_ASSERT(condition, message) ((void)0)
+#    define FMT_ASSERT(condition, message) \
+      ::fmt::detail::ignore_unused((condition), (message))
 #  else
 #    define FMT_ASSERT(condition, message)                                    \
       ((condition) /* void() fails with -Winvalid-constexpr on clang 4.0.1 */ \
@@ -296,6 +382,12 @@ FMT_NORETURN FMT_API void assert_fail(const char* file, int line,
 #  endif
 #endif
 
+#ifdef __cpp_lib_byte
+using byte = std::byte;
+#else
+enum class byte : unsigned char {};
+#endif
+
 #if defined(FMT_USE_STRING_VIEW)
 template <typename Char> using std_string_view = std::basic_string_view<Char>;
 #elif defined(FMT_USE_EXPERIMENTAL_STRING_VIEW)
@@ -312,38 +404,39 @@ template <typename T> struct std_string_view {};
 #  define FMT_USE_INT128 1
 using int128_t = __int128_t;
 using uint128_t = __uint128_t;
+template <typename T> inline auto convert_for_visit(T value) -> T {
+  return value;
+}
 #else
 #  define FMT_USE_INT128 0
 #endif
 #if !FMT_USE_INT128
-struct int128_t {};
-struct uint128_t {};
+enum class int128_t {};
+enum class uint128_t {};
+// Reduce template instantiations.
+template <typename T> inline auto convert_for_visit(T) -> monostate {
+  return {};
+}
 #endif
 
 // Casts a nonnegative integer to unsigned.
 template <typename Int>
-FMT_CONSTEXPR typename std::make_unsigned<Int>::type to_unsigned(Int value) {
+FMT_CONSTEXPR auto to_unsigned(Int value) ->
+    typename std::make_unsigned<Int>::type {
   FMT_ASSERT(value >= 0, "negative value");
   return static_cast<typename std::make_unsigned<Int>::type>(value);
 }
 
-FMT_SUPPRESS_MSC_WARNING(4566) constexpr unsigned char micro[] = "\u00B5";
+FMT_MSC_WARNING(suppress : 4566) constexpr unsigned char micro[] = "\u00B5";
 
-template <typename Char> constexpr bool is_unicode() {
-  return FMT_UNICODE || sizeof(Char) != 1 ||
-         (sizeof(micro) == 3 && micro[0] == 0xC2 && micro[1] == 0xB5);
+constexpr auto is_utf8() -> bool {
+  // Avoid buggy sign extensions in MSVC's constant evaluation mode.
+  // https://developercommunity.visualstudio.com/t/C-difference-in-behavior-for-unsigned/1233612
+  using uchar = unsigned char;
+  return FMT_UNICODE || (sizeof(micro) == 3 && uchar(micro[0]) == 0xC2 &&
+                         uchar(micro[1]) == 0xB5);
 }
-
-#ifdef __cpp_char8_t
-using char8_type = char8_t;
-#else
-enum char8_type : unsigned char {};
-#endif
-}  // namespace detail
-
-#ifdef FMT_USE_INTERNAL
-namespace internal = detail;  // DEPRECATED
-#endif
+FMT_END_DETAIL_NAMESPACE
 
 /**
   An implementation of ``std::basic_string_view`` for pre-C++17. It provides a
@@ -374,11 +467,14 @@ template <typename Char> class basic_string_view {
     the size with ``std::char_traits<Char>::length``.
     \endrst
    */
-#if __cplusplus >= 201703L  // C++17's char_traits::length() is constexpr.
-  FMT_CONSTEXPR
-#endif
+  FMT_CONSTEXPR_CHAR_TRAITS
+  FMT_INLINE
   basic_string_view(const Char* s)
-      : data_(s), size_(std::char_traits<Char>::length(s)) {}
+      : data_(s),
+        size_(detail::const_check(std::is_same<Char, char>::value &&
+                                  !detail::is_constant_evaluated(true))
+                  ? std::strlen(reinterpret_cast<const char*>(s))
+                  : std::char_traits<Char>::length(s)) {}
 
   /** Constructs a string reference from a ``std::basic_string`` object. */
   template <typename Traits, typename Alloc>
@@ -393,23 +489,25 @@ template <typename Char> class basic_string_view {
                                                       size_(s.size()) {}
 
   /** Returns a pointer to the string data. */
-  constexpr const Char* data() const { return data_; }
+  constexpr auto data() const FMT_NOEXCEPT -> const Char* { return data_; }
 
   /** Returns the string size. */
-  constexpr size_t size() const { return size_; }
+  constexpr auto size() const FMT_NOEXCEPT -> size_t { return size_; }
 
-  constexpr iterator begin() const { return data_; }
-  constexpr iterator end() const { return data_ + size_; }
+  constexpr auto begin() const FMT_NOEXCEPT -> iterator { return data_; }
+  constexpr auto end() const FMT_NOEXCEPT -> iterator { return data_ + size_; }
 
-  constexpr const Char& operator[](size_t pos) const { return data_[pos]; }
+  constexpr auto operator[](size_t pos) const FMT_NOEXCEPT -> const Char& {
+    return data_[pos];
+  }
 
-  FMT_CONSTEXPR void remove_prefix(size_t n) {
+  FMT_CONSTEXPR void remove_prefix(size_t n) FMT_NOEXCEPT {
     data_ += n;
     size_ -= n;
   }
 
   // Lexicographically compare this string reference to other.
-  int compare(basic_string_view other) const {
+  FMT_CONSTEXPR_CHAR_TRAITS auto compare(basic_string_view other) const -> int {
     size_t str_size = size_ < other.size_ ? size_ : other.size_;
     int result = std::char_traits<Char>::compare(data_, other.data_, str_size);
     if (result == 0)
@@ -417,72 +515,53 @@ template <typename Char> class basic_string_view {
     return result;
   }
 
-  friend bool operator==(basic_string_view lhs, basic_string_view rhs) {
+  FMT_CONSTEXPR_CHAR_TRAITS friend auto operator==(basic_string_view lhs,
+                                                   basic_string_view rhs)
+      -> bool {
     return lhs.compare(rhs) == 0;
   }
-  friend bool operator!=(basic_string_view lhs, basic_string_view rhs) {
+  friend auto operator!=(basic_string_view lhs, basic_string_view rhs) -> bool {
     return lhs.compare(rhs) != 0;
   }
-  friend bool operator<(basic_string_view lhs, basic_string_view rhs) {
+  friend auto operator<(basic_string_view lhs, basic_string_view rhs) -> bool {
     return lhs.compare(rhs) < 0;
   }
-  friend bool operator<=(basic_string_view lhs, basic_string_view rhs) {
+  friend auto operator<=(basic_string_view lhs, basic_string_view rhs) -> bool {
     return lhs.compare(rhs) <= 0;
   }
-  friend bool operator>(basic_string_view lhs, basic_string_view rhs) {
+  friend auto operator>(basic_string_view lhs, basic_string_view rhs) -> bool {
     return lhs.compare(rhs) > 0;
   }
-  friend bool operator>=(basic_string_view lhs, basic_string_view rhs) {
+  friend auto operator>=(basic_string_view lhs, basic_string_view rhs) -> bool {
     return lhs.compare(rhs) >= 0;
   }
 };
 
 using string_view = basic_string_view<char>;
-using wstring_view = basic_string_view<wchar_t>;
 
 /** Specifies if ``T`` is a character type. Can be specialized by users. */
 template <typename T> struct is_char : std::false_type {};
 template <> struct is_char<char> : std::true_type {};
-template <> struct is_char<wchar_t> : std::true_type {};
-template <> struct is_char<detail::char8_type> : std::true_type {};
-template <> struct is_char<char16_t> : std::true_type {};
-template <> struct is_char<char32_t> : std::true_type {};
-
-/**
-  \rst
-  Returns a string view of `s`. In order to add custom string type support to
-  {fmt} provide an overload of `to_string_view` for it in the same namespace as
-  the type for the argument-dependent lookup to work.
 
-  **Example**::
-
-    namespace my_ns {
-    inline string_view to_string_view(const my_string& s) {
-      return {s.data(), s.length()};
-    }
-    }
-    std::string message = fmt::format(my_string("The answer is {}"), 42);
-  \endrst
- */
+// Returns a string view of `s`.
 template <typename Char, FMT_ENABLE_IF(is_char<Char>::value)>
-inline basic_string_view<Char> to_string_view(const Char* s) {
+FMT_INLINE auto to_string_view(const Char* s) -> basic_string_view<Char> {
   return s;
 }
-
 template <typename Char, typename Traits, typename Alloc>
-inline basic_string_view<Char> to_string_view(
-    const std::basic_string<Char, Traits, Alloc>& s) {
+inline auto to_string_view(const std::basic_string<Char, Traits, Alloc>& s)
+    -> basic_string_view<Char> {
   return s;
 }
-
 template <typename Char>
-inline basic_string_view<Char> to_string_view(basic_string_view<Char> s) {
+constexpr auto to_string_view(basic_string_view<Char> s)
+    -> basic_string_view<Char> {
   return s;
 }
-
 template <typename Char,
           FMT_ENABLE_IF(!std::is_empty<detail::std_string_view<Char>>::value)>
-inline basic_string_view<Char> to_string_view(detail::std_string_view<Char> s) {
+inline auto to_string_view(detail::std_string_view<Char> s)
+    -> basic_string_view<Char> {
   return s;
 }
 
@@ -494,13 +573,15 @@ template <typename S>
 struct is_compile_string : std::is_base_of<compile_string, S> {};
 
 template <typename S, FMT_ENABLE_IF(is_compile_string<S>::value)>
-constexpr basic_string_view<typename S::char_type> to_string_view(const S& s) {
-  return s;
+constexpr auto to_string_view(const S& s)
+    -> basic_string_view<typename S::char_type> {
+  return basic_string_view<typename S::char_type>(s);
 }
 
-namespace detail {
+FMT_BEGIN_DETAIL_NAMESPACE
+
 void to_string_view(...);
-using fmt::v7::to_string_view;
+using fmt::to_string_view;
 
 // Specifies whether S is a string type convertible to fmt::basic_string_view.
 // It should be a constexpr function but MSVC 2017 fails to compile it in
@@ -527,6 +608,8 @@ FMT_INLINE void check_format_string(const S&) {
 template <typename..., typename S, FMT_ENABLE_IF(is_compile_string<S>::value)>
 void check_format_string(S);
 
+FMT_NORETURN FMT_API void throw_format_error(const char* message);
+
 struct error_handler {
   constexpr error_handler() = default;
   constexpr error_handler(const error_handler&) = default;
@@ -534,7 +617,7 @@ struct error_handler {
   // This function is intentionally not constexpr to give a compile-time error.
   FMT_NORETURN FMT_API void on_error(const char* message);
 };
-}  // namespace detail
+FMT_END_DETAIL_NAMESPACE
 
 /** String's character type. */
 template <typename S> using char_t = typename detail::char_t_impl<S>::type;
@@ -543,16 +626,7 @@ template <typename S> using char_t = typename detail::char_t_impl<S>::type;
   \rst
   Parsing context consisting of a format string range being parsed and an
   argument counter for automatic indexing.
-
-  You can use one of the following type aliases for common character types:
-
-  +-----------------------+-------------------------------------+
-  | Type                  | Definition                          |
-  +=======================+=====================================+
-  | format_parse_context  | basic_format_parse_context<char>    |
-  +-----------------------+-------------------------------------+
-  | wformat_parse_context | basic_format_parse_context<wchar_t> |
-  +-----------------------+-------------------------------------+
+  You can use the ``format_parse_context`` type alias for ``char`` instead.
   \endrst
  */
 template <typename Char, typename ErrorHandler = detail::error_handler>
@@ -574,12 +648,16 @@ class basic_format_parse_context : private ErrorHandler {
     Returns an iterator to the beginning of the format string range being
     parsed.
    */
-  constexpr iterator begin() const FMT_NOEXCEPT { return format_str_.begin(); }
+  constexpr auto begin() const FMT_NOEXCEPT -> iterator {
+    return format_str_.begin();
+  }
 
   /**
     Returns an iterator past the end of the format string range being parsed.
    */
-  constexpr iterator end() const FMT_NOEXCEPT { return format_str_.end(); }
+  constexpr auto end() const FMT_NOEXCEPT -> iterator {
+    return format_str_.end();
+  }
 
   /** Advances the begin iterator to ``it``. */
   FMT_CONSTEXPR void advance_to(iterator it) {
@@ -590,7 +668,7 @@ class basic_format_parse_context : private ErrorHandler {
     Reports an error if using the manual argument indexing; otherwise returns
     the next argument index and switches to the automatic indexing.
    */
-  FMT_CONSTEXPR int next_arg_id() {
+  FMT_CONSTEXPR auto next_arg_id() -> int {
     // Don't check if the argument id is valid to avoid overhead and because it
     // will be checked during formatting anyway.
     if (next_arg_id_ >= 0) return next_arg_id_++;
@@ -615,11 +693,10 @@ class basic_format_parse_context : private ErrorHandler {
     ErrorHandler::on_error(message);
   }
 
-  constexpr ErrorHandler error_handler() const { return *this; }
+  constexpr auto error_handler() const -> ErrorHandler { return *this; }
 };
 
 using format_parse_context = basic_format_parse_context<char>;
-using wformat_parse_context = basic_format_parse_context<wchar_t>;
 
 template <typename Context> class basic_format_arg;
 template <typename Context> class basic_format_args;
@@ -643,11 +720,30 @@ template <typename T> struct is_contiguous : std::false_type {};
 template <typename Char>
 struct is_contiguous<std::basic_string<Char>> : std::true_type {};
 
-namespace detail {
+class appender;
+
+FMT_BEGIN_DETAIL_NAMESPACE
+
+template <typename Context, typename T>
+constexpr auto has_const_formatter_impl(T*)
+    -> decltype(typename Context::template formatter_type<T>().format(
+                    std::declval<const T&>(), std::declval<Context&>()),
+                true) {
+  return true;
+}
+template <typename Context>
+constexpr auto has_const_formatter_impl(...) -> bool {
+  return false;
+}
+template <typename T, typename Context>
+constexpr auto has_const_formatter() -> bool {
+  return has_const_formatter_impl<Context>(static_cast<T*>(nullptr));
+}
 
 // Extracts a reference to the container from back_insert_iterator.
 template <typename Container>
-inline Container& get_container(std::back_insert_iterator<Container> it) {
+inline auto get_container(std::back_insert_iterator<Container> it)
+    -> Container& {
   using bi_iterator = std::back_insert_iterator<Container>;
   struct accessor : bi_iterator {
     accessor(bi_iterator iter) : bi_iterator(iter) {}
@@ -656,6 +752,23 @@ inline Container& get_container(std::back_insert_iterator<Container> it) {
   return *accessor(it).container;
 }
 
+template <typename Char, typename InputIt, typename OutputIt>
+FMT_CONSTEXPR auto copy_str(InputIt begin, InputIt end, OutputIt out)
+    -> OutputIt {
+  while (begin != end) *out++ = static_cast<Char>(*begin++);
+  return out;
+}
+
+template <typename Char, typename T, typename U,
+          FMT_ENABLE_IF(
+              std::is_same<remove_const_t<T>, U>::value&& is_char<U>::value)>
+FMT_CONSTEXPR auto copy_str(T* begin, T* end, U* out) -> U* {
+  if (is_constant_evaluated()) return copy_str<Char, T*, U*>(begin, end, out);
+  auto size = to_unsigned(end - begin);
+  memcpy(out, begin, size * sizeof(U));
+  return out + size;
+}
+
 /**
   \rst
   A contiguous memory buffer with an optional growing ability. It is an internal
@@ -670,24 +783,25 @@ template <typename T> class buffer {
 
  protected:
   // Don't initialize ptr_ since it is not accessed to save a few cycles.
-  FMT_SUPPRESS_MSC_WARNING(26495)
+  FMT_MSC_WARNING(suppress : 26495)
   buffer(size_t sz) FMT_NOEXCEPT : size_(sz), capacity_(sz) {}
 
-  buffer(T* p = nullptr, size_t sz = 0, size_t cap = 0) FMT_NOEXCEPT
-      : ptr_(p),
-        size_(sz),
-        capacity_(cap) {}
+  FMT_CONSTEXPR20 buffer(T* p = nullptr, size_t sz = 0,
+                         size_t cap = 0) FMT_NOEXCEPT : ptr_(p),
+                                                        size_(sz),
+                                                        capacity_(cap) {}
 
-  ~buffer() = default;
+  FMT_CONSTEXPR20 ~buffer() = default;
+  buffer(buffer&&) = default;
 
   /** Sets the buffer data and capacity. */
-  void set(T* buf_data, size_t buf_capacity) FMT_NOEXCEPT {
+  FMT_CONSTEXPR void set(T* buf_data, size_t buf_capacity) FMT_NOEXCEPT {
     ptr_ = buf_data;
     capacity_ = buf_capacity;
   }
 
   /** Increases the buffer capacity to hold at least *capacity* elements. */
-  virtual void grow(size_t capacity) = 0;
+  virtual FMT_CONSTEXPR20 void grow(size_t capacity) = 0;
 
  public:
   using value_type = T;
@@ -696,30 +810,30 @@ template <typename T> class buffer {
   buffer(const buffer&) = delete;
   void operator=(const buffer&) = delete;
 
-  T* begin() FMT_NOEXCEPT { return ptr_; }
-  T* end() FMT_NOEXCEPT { return ptr_ + size_; }
+  auto begin() FMT_NOEXCEPT -> T* { return ptr_; }
+  auto end() FMT_NOEXCEPT -> T* { return ptr_ + size_; }
 
-  const T* begin() const FMT_NOEXCEPT { return ptr_; }
-  const T* end() const FMT_NOEXCEPT { return ptr_ + size_; }
+  auto begin() const FMT_NOEXCEPT -> const T* { return ptr_; }
+  auto end() const FMT_NOEXCEPT -> const T* { return ptr_ + size_; }
 
   /** Returns the size of this buffer. */
-  size_t size() const FMT_NOEXCEPT { return size_; }
+  constexpr auto size() const FMT_NOEXCEPT -> size_t { return size_; }
 
   /** Returns the capacity of this buffer. */
-  size_t capacity() const FMT_NOEXCEPT { return capacity_; }
+  constexpr auto capacity() const FMT_NOEXCEPT -> size_t { return capacity_; }
 
   /** Returns a pointer to the buffer data. */
-  T* data() FMT_NOEXCEPT { return ptr_; }
+  FMT_CONSTEXPR auto data() FMT_NOEXCEPT -> T* { return ptr_; }
 
   /** Returns a pointer to the buffer data. */
-  const T* data() const FMT_NOEXCEPT { return ptr_; }
+  FMT_CONSTEXPR auto data() const FMT_NOEXCEPT -> const T* { return ptr_; }
 
   /** Clears this buffer. */
   void clear() { size_ = 0; }
 
   // Tries resizing the buffer to contain *count* elements. If T is a POD type
   // the new elements may not be initialized.
-  void try_resize(size_t count) {
+  FMT_CONSTEXPR20 void try_resize(size_t count) {
     try_reserve(count);
     size_ = count <= capacity_ ? count : capacity_;
   }
@@ -728,11 +842,11 @@ template <typename T> class buffer {
   // 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) {
+  FMT_CONSTEXPR20 void try_reserve(size_t new_capacity) {
     if (new_capacity > capacity_) grow(new_capacity);
   }
 
-  void push_back(const T& value) {
+  FMT_CONSTEXPR20 void push_back(const T& value) {
     try_reserve(size_ + 1);
     ptr_[size_++] = value;
   }
@@ -740,16 +854,19 @@ template <typename T> class buffer {
   /** Appends data to the end of the buffer. */
   template <typename U> void append(const U* begin, const U* end);
 
-  template <typename I> T& operator[](I index) { return ptr_[index]; }
-  template <typename I> const T& operator[](I index) const {
+  template <typename I> FMT_CONSTEXPR auto operator[](I index) -> T& {
+    return ptr_[index];
+  }
+  template <typename I>
+  FMT_CONSTEXPR auto operator[](I index) const -> const T& {
     return ptr_[index];
   }
 };
 
 struct buffer_traits {
   explicit buffer_traits(size_t) {}
-  size_t count() const { return 0; }
-  size_t limit(size_t size) { return size; }
+  auto count() const -> size_t { return 0; }
+  auto limit(size_t size) -> size_t { return size; }
 };
 
 class fixed_buffer_traits {
@@ -759,8 +876,8 @@ class fixed_buffer_traits {
 
  public:
   explicit fixed_buffer_traits(size_t limit) : limit_(limit) {}
-  size_t count() const { return count_; }
-  size_t limit(size_t size) {
+  auto count() const -> size_t { return count_; }
+  auto limit(size_t size) -> size_t {
     size_t n = limit_ > count_ ? limit_ - count_ : 0;
     count_ += size;
     return size < n ? size : n;
@@ -776,33 +893,84 @@ class iterator_buffer final : public Traits, public buffer<T> {
   T data_[buffer_size];
 
  protected:
-  void grow(size_t) final FMT_OVERRIDE {
+  FMT_CONSTEXPR20 void grow(size_t) override {
     if (this->size() == buffer_size) flush();
   }
-  void flush();
+
+  void flush() {
+    auto size = this->size();
+    this->clear();
+    out_ = copy_str<T>(data_, data_ + this->limit(size), out_);
+  }
 
  public:
   explicit iterator_buffer(OutputIt out, size_t n = buffer_size)
-      : Traits(n),
-        buffer<T>(data_, 0, buffer_size),
-        out_(out) {}
+      : Traits(n), buffer<T>(data_, 0, buffer_size), out_(out) {}
+  iterator_buffer(iterator_buffer&& other)
+      : Traits(other), buffer<T>(data_, 0, buffer_size), out_(other.out_) {}
+  ~iterator_buffer() { flush(); }
+
+  auto out() -> OutputIt {
+    flush();
+    return out_;
+  }
+  auto count() const -> size_t { return Traits::count() + this->size(); }
+};
+
+template <typename T>
+class iterator_buffer<T*, T, fixed_buffer_traits> final
+    : public fixed_buffer_traits,
+      public buffer<T> {
+ private:
+  T* out_;
+  enum { buffer_size = 256 };
+  T data_[buffer_size];
+
+ protected:
+  FMT_CONSTEXPR20 void grow(size_t) override {
+    if (this->size() == this->capacity()) flush();
+  }
+
+  void flush() {
+    size_t n = this->limit(this->size());
+    if (this->data() == out_) {
+      out_ += n;
+      this->set(data_, buffer_size);
+    }
+    this->clear();
+  }
+
+ public:
+  explicit iterator_buffer(T* out, size_t n = buffer_size)
+      : fixed_buffer_traits(n), buffer<T>(out, 0, n), out_(out) {}
+  iterator_buffer(iterator_buffer&& other)
+      : fixed_buffer_traits(other),
+        buffer<T>(std::move(other)),
+        out_(other.out_) {
+    if (this->data() != out_) {
+      this->set(data_, buffer_size);
+      this->clear();
+    }
+  }
   ~iterator_buffer() { flush(); }
 
-  OutputIt out() {
+  auto out() -> T* {
     flush();
     return out_;
   }
-  size_t count() const { return Traits::count() + this->size(); }
+  auto count() const -> size_t {
+    return fixed_buffer_traits::count() + this->size();
+  }
 };
 
 template <typename T> class iterator_buffer<T*, T> final : public buffer<T> {
  protected:
-  void grow(size_t) final FMT_OVERRIDE {}
+  FMT_CONSTEXPR20 void grow(size_t) override {}
 
  public:
   explicit iterator_buffer(T* out, size_t = 0) : buffer<T>(out, 0, ~size_t()) {}
 
-  T* out() { return &*this->end(); }
+  auto out() -> T* { return &*this->end(); }
 };
 
 // A buffer that writes to a container with the contiguous storage.
@@ -815,7 +983,7 @@ class iterator_buffer<std::back_insert_iterator<Container>,
   Container& container_;
 
  protected:
-  void grow(size_t capacity) final FMT_OVERRIDE {
+  FMT_CONSTEXPR20 void grow(size_t capacity) override {
     container_.resize(capacity);
     this->set(&container_[0], capacity);
   }
@@ -825,7 +993,7 @@ class iterator_buffer<std::back_insert_iterator<Container>,
       : 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() {
+  auto out() -> std::back_insert_iterator<Container> {
     return std::back_inserter(container_);
   }
 };
@@ -838,7 +1006,7 @@ template <typename T = char> class counting_buffer final : public buffer<T> {
   size_t count_ = 0;
 
  protected:
-  void grow(size_t) final FMT_OVERRIDE {
+  FMT_CONSTEXPR20 void grow(size_t) override {
     if (this->size() != buffer_size) return;
     count_ += this->size();
     this->clear();
@@ -847,48 +1015,24 @@ template <typename T = char> class counting_buffer final : public buffer<T> {
  public:
   counting_buffer() : buffer<T>(data_, 0, buffer_size) {}
 
-  size_t count() { return count_ + this->size(); }
+  auto count() -> size_t { 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) {}
+using buffer_appender = conditional_t<std::is_same<T, char>::value, appender,
+                                      std::back_insert_iterator<buffer<T>>>;
 
-  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.
+// Maps an output iterator to 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);
+auto get_buffer(OutputIt out) -> iterator_buffer<OutputIt, T> {
+  return iterator_buffer<OutputIt, T>(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) {
+template <typename T> auto get_iterator(buffer<T>& buf) -> buffer_appender<T> {
   return buffer_appender<T>(buf);
 }
 
@@ -898,9 +1042,9 @@ struct fallback_formatter {
 };
 
 // Specifies if T has an enabled fallback_formatter specialization.
-template <typename T, typename Context>
+template <typename T, typename Char>
 using has_fallback_formatter =
-    std::is_constructible<fallback_formatter<T, typename Context::char_type>>;
+    std::is_constructible<fallback_formatter<T, Char>>;
 
 struct view {};
 
@@ -925,8 +1069,8 @@ struct arg_data {
   template <typename... U>
   arg_data(const U&... init) : args_{T(named_args_, NUM_NAMED_ARGS), init...} {}
   arg_data(const arg_data& other) = delete;
-  const T* args() const { return args_ + 1; }
-  named_arg_info<Char>* named_args() { return named_args_; }
+  auto args() const -> const T* { return args_ + 1; }
+  auto named_args() -> named_arg_info<Char>* { return named_args_; }
 };
 
 template <typename T, typename Char, size_t NUM_ARGS>
@@ -935,45 +1079,55 @@ struct arg_data<T, Char, NUM_ARGS, 0> {
   T args_[NUM_ARGS != 0 ? NUM_ARGS : +1];
 
   template <typename... U>
-  FMT_INLINE arg_data(const U&... init) : args_{init...} {}
-  FMT_INLINE const T* args() const { return args_; }
-  FMT_INLINE std::nullptr_t named_args() { return nullptr; }
+  FMT_CONSTEXPR FMT_INLINE arg_data(const U&... init) : args_{init...} {}
+  FMT_CONSTEXPR FMT_INLINE auto args() const -> const T* { return args_; }
+  FMT_CONSTEXPR FMT_INLINE auto named_args() -> std::nullptr_t {
+    return nullptr;
+  }
 };
 
 template <typename Char>
 inline void init_named_args(named_arg_info<Char>*, int, int) {}
 
-template <typename Char, typename T, typename... Tail>
+template <typename T> struct is_named_arg : std::false_type {};
+template <typename T> struct is_statically_named_arg : std::false_type {};
+
+template <typename T, typename Char>
+struct is_named_arg<named_arg<Char, T>> : std::true_type {};
+
+template <typename Char, typename T, typename... Tail,
+          FMT_ENABLE_IF(!is_named_arg<T>::value)>
 void init_named_args(named_arg_info<Char>* named_args, int arg_count,
                      int named_arg_count, const T&, const Tail&... args) {
   init_named_args(named_args, arg_count + 1, named_arg_count, args...);
 }
 
-template <typename Char, typename T, typename... Tail>
+template <typename Char, typename T, typename... Tail,
+          FMT_ENABLE_IF(is_named_arg<T>::value)>
 void init_named_args(named_arg_info<Char>* named_args, int arg_count,
-                     int named_arg_count, const named_arg<Char, T>& arg,
-                     const Tail&... args) {
+                     int named_arg_count, const T& arg, const Tail&... args) {
   named_args[named_arg_count++] = {arg.name, arg_count};
   init_named_args(named_args, arg_count + 1, named_arg_count, args...);
 }
 
 template <typename... Args>
-FMT_INLINE void init_named_args(std::nullptr_t, int, int, const Args&...) {}
-
-template <typename T> struct is_named_arg : std::false_type {};
-
-template <typename T, typename Char>
-struct is_named_arg<named_arg<Char, T>> : std::true_type {};
+FMT_CONSTEXPR FMT_INLINE void init_named_args(std::nullptr_t, int, int,
+                                              const Args&...) {}
 
-template <bool B = false> constexpr size_t count() { return B ? 1 : 0; }
-template <bool B1, bool B2, bool... Tail> constexpr size_t count() {
+template <bool B = false> constexpr auto count() -> size_t { return B ? 1 : 0; }
+template <bool B1, bool B2, bool... Tail> constexpr auto count() -> size_t {
   return (B1 ? 1 : 0) + count<B2, Tail...>();
 }
 
-template <typename... Args> constexpr size_t count_named_args() {
+template <typename... Args> constexpr auto count_named_args() -> size_t {
   return count<is_named_arg<Args>::value...>();
 }
 
+template <typename... Args>
+constexpr auto count_statically_named_args() -> size_t {
+  return count<is_statically_named_arg<Args>::value...>();
+}
+
 enum class type {
   none_type,
   // Integer types should go first,
@@ -1029,6 +1183,11 @@ constexpr bool is_arithmetic_type(type t) {
   return t > type::none_type && t <= type::last_numeric_type;
 }
 
+struct unformattable {};
+struct unformattable_char : unformattable {};
+struct unformattable_const : unformattable {};
+struct unformattable_pointer : unformattable {};
+
 template <typename Char> struct string_value {
   const Char* data;
   size_t size;
@@ -1041,8 +1200,8 @@ template <typename Char> struct named_arg_value {
 
 template <typename Context> struct custom_value {
   using parse_context = typename Context::parse_context_type;
-  const void* value;
-  void (*format)(const void* arg, parse_context& parse_ctx, Context& ctx);
+  void* value;
+  void (*format)(void* arg, parse_context& parse_ctx, Context& ctx);
 };
 
 // A formatting argument value.
@@ -1051,6 +1210,7 @@ template <typename Context> class value {
   using char_type = typename Context::char_type;
 
   union {
+    monostate no_value;
     int int_value;
     unsigned uint_value;
     long long long_long_value;
@@ -1068,19 +1228,23 @@ template <typename Context> class value {
     named_arg_value<char_type> named_args;
   };
 
-  constexpr FMT_INLINE value(int val = 0) : int_value(val) {}
+  constexpr FMT_INLINE value() : no_value() {}
+  constexpr FMT_INLINE value(int val) : int_value(val) {}
   constexpr FMT_INLINE value(unsigned val) : uint_value(val) {}
-  FMT_INLINE value(long long val) : long_long_value(val) {}
-  FMT_INLINE value(unsigned long long val) : ulong_long_value(val) {}
+  constexpr FMT_INLINE value(long long val) : long_long_value(val) {}
+  constexpr FMT_INLINE value(unsigned long long val) : ulong_long_value(val) {}
   FMT_INLINE value(int128_t val) : int128_value(val) {}
   FMT_INLINE value(uint128_t val) : uint128_value(val) {}
-  FMT_INLINE value(float val) : float_value(val) {}
-  FMT_INLINE value(double val) : double_value(val) {}
+  constexpr FMT_INLINE value(float val) : float_value(val) {}
+  constexpr FMT_INLINE value(double val) : double_value(val) {}
   FMT_INLINE value(long double val) : long_double_value(val) {}
-  FMT_INLINE value(bool val) : bool_value(val) {}
-  FMT_INLINE value(char_type val) : char_value(val) {}
-  FMT_INLINE value(const char_type* val) { string.data = val; }
-  FMT_INLINE value(basic_string_view<char_type> val) {
+  constexpr FMT_INLINE value(bool val) : bool_value(val) {}
+  constexpr FMT_INLINE value(char_type val) : char_value(val) {}
+  FMT_CONSTEXPR FMT_INLINE value(const char_type* val) {
+    string.data = val;
+    if (is_constant_evaluated()) string.size = {};
+  }
+  FMT_CONSTEXPR FMT_INLINE value(basic_string_view<char_type> val) {
     string.data = val.data();
     string.size = val.size();
   }
@@ -1088,31 +1252,39 @@ template <typename Context> class value {
   FMT_INLINE value(const named_arg_info<char_type>* args, size_t size)
       : named_args{args, size} {}
 
-  template <typename T> FMT_INLINE value(const T& val) {
-    custom.value = &val;
+  template <typename T> FMT_CONSTEXPR FMT_INLINE value(T& val) {
+    using value_type = remove_cvref_t<T>;
+    custom.value = const_cast<value_type*>(&val);
     // Get the formatter type through the context to allow different contexts
     // have different extension points, e.g. `formatter<T>` for `format` and
     // `printf_formatter<T>` for `printf`.
     custom.format = format_custom_arg<
-        T, conditional_t<has_formatter<T, Context>::value,
-                         typename Context::template formatter_type<T>,
-                         fallback_formatter<T, char_type>>>;
+        value_type,
+        conditional_t<has_formatter<value_type, Context>::value,
+                      typename Context::template formatter_type<value_type>,
+                      fallback_formatter<value_type, char_type>>>;
   }
+  value(unformattable);
+  value(unformattable_char);
+  value(unformattable_const);
+  value(unformattable_pointer);
 
  private:
   // Formats an argument of a custom type, such as a user-defined class.
   template <typename T, typename Formatter>
-  static void format_custom_arg(const void* arg,
+  static void format_custom_arg(void* arg,
                                 typename Context::parse_context_type& parse_ctx,
                                 Context& ctx) {
-    Formatter f;
+    auto f = Formatter();
     parse_ctx.advance_to(f.parse(parse_ctx));
-    ctx.advance_to(f.format(*static_cast<const T*>(arg), ctx));
+    using qualified_type =
+        conditional_t<has_const_formatter<T, Context>(), const T, T>;
+    ctx.advance_to(f.format(*static_cast<qualified_type*>(arg), ctx));
   }
 };
 
 template <typename Context, typename T>
-FMT_CONSTEXPR basic_format_arg<Context> make_arg(const T& value);
+FMT_CONSTEXPR auto make_arg(const T& value) -> basic_format_arg<Context>;
 
 // To minimize the number of types we need to deal with, long is translated
 // either to int or to long long depending on its size.
@@ -1120,52 +1292,84 @@ 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.
+// arg_mapper reports errors by returning unformattable instead of using
+// static_assert because it's used in the is_formattable trait.
 template <typename Context> struct arg_mapper {
   using char_type = typename Context::char_type;
 
-  FMT_CONSTEXPR int map(signed char val) { return val; }
-  FMT_CONSTEXPR unsigned map(unsigned char val) { return val; }
-  FMT_CONSTEXPR int map(short val) { return val; }
-  FMT_CONSTEXPR unsigned map(unsigned short val) { return val; }
-  FMT_CONSTEXPR int map(int val) { return val; }
-  FMT_CONSTEXPR unsigned map(unsigned val) { return val; }
-  FMT_CONSTEXPR long_type map(long val) { return val; }
-  FMT_CONSTEXPR ulong_type map(unsigned long val) { return val; }
-  FMT_CONSTEXPR long long map(long long val) { return val; }
-  FMT_CONSTEXPR unsigned long long map(unsigned long long val) { return val; }
-  FMT_CONSTEXPR int128_t map(int128_t val) { return val; }
-  FMT_CONSTEXPR uint128_t map(uint128_t val) { return val; }
-  FMT_CONSTEXPR bool map(bool val) { return val; }
-
-  template <typename T, FMT_ENABLE_IF(is_char<T>::value)>
-  FMT_CONSTEXPR char_type map(T val) {
-    static_assert(
-        std::is_same<T, char>::value || std::is_same<T, char_type>::value,
-        "mixing character types is disallowed");
+  FMT_CONSTEXPR FMT_INLINE auto map(signed char val) -> int { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned char val) -> unsigned {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(short val) -> int { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned short val) -> unsigned {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(int val) -> int { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned val) -> unsigned { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(long val) -> long_type { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned long val) -> ulong_type {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(long long val) -> long long { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(unsigned long long val)
+      -> unsigned long long {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(int128_t val) -> int128_t { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(uint128_t val) -> uint128_t { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(bool val) -> bool { return val; }
+
+  template <typename T, FMT_ENABLE_IF(std::is_same<T, char>::value ||
+                                      std::is_same<T, char_type>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(T val) -> char_type {
     return val;
   }
+  template <typename T, enable_if_t<(std::is_same<T, wchar_t>::value ||
+#ifdef __cpp_char8_t
+                                     std::is_same<T, char8_t>::value ||
+#endif
+                                     std::is_same<T, char16_t>::value ||
+                                     std::is_same<T, char32_t>::value) &&
+                                        !std::is_same<T, char_type>::value,
+                                    int> = 0>
+  FMT_CONSTEXPR FMT_INLINE auto map(T) -> unformattable_char {
+    return {};
+  }
 
-  FMT_CONSTEXPR float map(float val) { return val; }
-  FMT_CONSTEXPR double map(double val) { return val; }
-  FMT_CONSTEXPR long double map(long double val) { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(float val) -> float { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(double val) -> double { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(long double val) -> long double {
+    return val;
+  }
 
-  FMT_CONSTEXPR const char_type* map(char_type* val) { return val; }
-  FMT_CONSTEXPR const char_type* map(const char_type* val) { return val; }
-  template <typename T, FMT_ENABLE_IF(is_string<T>::value)>
-  FMT_CONSTEXPR basic_string_view<char_type> map(const T& val) {
-    static_assert(std::is_same<char_type, char_t<T>>::value,
-                  "mixing character types is disallowed");
+  FMT_CONSTEXPR FMT_INLINE auto map(char_type* val) -> const char_type* {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(const char_type* val) -> const char_type* {
+    return val;
+  }
+  template <typename T,
+            FMT_ENABLE_IF(is_string<T>::value && !std::is_pointer<T>::value &&
+                          std::is_same<char_type, char_t<T>>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T& val)
+      -> basic_string_view<char_type> {
     return to_string_view(val);
   }
+  template <typename T,
+            FMT_ENABLE_IF(is_string<T>::value && !std::is_pointer<T>::value &&
+                          !std::is_same<char_type, char_t<T>>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T&) -> unformattable_char {
+    return {};
+  }
   template <typename T,
             FMT_ENABLE_IF(
                 std::is_constructible<basic_string_view<char_type>, T>::value &&
                 !is_string<T>::value && !has_formatter<T, Context>::value &&
-                !has_fallback_formatter<T, Context>::value)>
-  FMT_CONSTEXPR basic_string_view<char_type> map(const T& val) {
+                !has_fallback_formatter<T, char_type>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T& val)
+      -> basic_string_view<char_type> {
     return basic_string_view<char_type>(val);
   }
   template <
@@ -1174,63 +1378,113 @@ template <typename Context> struct arg_mapper {
           std::is_constructible<std_string_view<char_type>, T>::value &&
           !std::is_constructible<basic_string_view<char_type>, T>::value &&
           !is_string<T>::value && !has_formatter<T, Context>::value &&
-          !has_fallback_formatter<T, Context>::value)>
-  FMT_CONSTEXPR basic_string_view<char_type> map(const T& val) {
+          !has_fallback_formatter<T, char_type>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T& val)
+      -> basic_string_view<char_type> {
     return std_string_view<char_type>(val);
   }
-  FMT_CONSTEXPR const char* map(const signed char* val) {
-    static_assert(std::is_same<char_type, char>::value, "invalid string type");
-    return reinterpret_cast<const char*>(val);
+
+  using cstring_result = conditional_t<std::is_same<char_type, char>::value,
+                                       const char*, unformattable_pointer>;
+
+  FMT_DEPRECATED FMT_CONSTEXPR FMT_INLINE auto map(const signed char* val)
+      -> cstring_result {
+    return map(reinterpret_cast<const char*>(val));
   }
-  FMT_CONSTEXPR const char* map(const unsigned char* val) {
-    static_assert(std::is_same<char_type, char>::value, "invalid string type");
-    return reinterpret_cast<const char*>(val);
+  FMT_DEPRECATED FMT_CONSTEXPR FMT_INLINE auto map(const unsigned char* val)
+      -> cstring_result {
+    return map(reinterpret_cast<const char*>(val));
   }
-  FMT_CONSTEXPR const char* map(signed char* val) {
-    const auto* const_val = val;
-    return map(const_val);
+  FMT_DEPRECATED FMT_CONSTEXPR FMT_INLINE auto map(signed char* val)
+      -> cstring_result {
+    return map(reinterpret_cast<const char*>(val));
   }
-  FMT_CONSTEXPR const char* map(unsigned char* val) {
-    const auto* const_val = val;
-    return map(const_val);
+  FMT_DEPRECATED FMT_CONSTEXPR FMT_INLINE auto map(unsigned char* val)
+      -> cstring_result {
+    return map(reinterpret_cast<const char*>(val));
   }
 
-  FMT_CONSTEXPR const void* map(void* val) { return val; }
-  FMT_CONSTEXPR const void* map(const void* val) { return val; }
-  FMT_CONSTEXPR const void* map(std::nullptr_t val) { return val; }
-  template <typename T> FMT_CONSTEXPR int map(const T*) {
-    // Formatting of arbitrary pointers is disallowed. If you want to output
-    // a pointer cast it to "void *" or "const void *". In particular, this
-    // forbids formatting of "[const] volatile char *" which is printed as bool
-    // by iostreams.
-    static_assert(!sizeof(T), "formatting of non-void pointers is disallowed");
-    return 0;
+  FMT_CONSTEXPR FMT_INLINE auto map(void* val) -> const void* { return val; }
+  FMT_CONSTEXPR FMT_INLINE auto map(const void* val) -> const void* {
+    return val;
+  }
+  FMT_CONSTEXPR FMT_INLINE auto map(std::nullptr_t val) -> const void* {
+    return val;
+  }
+
+  // We use SFINAE instead of a const T* parameter to avoid conflicting with
+  // the C array overload.
+  template <
+      typename T,
+      FMT_ENABLE_IF(
+          std::is_member_pointer<T>::value ||
+          std::is_function<typename std::remove_pointer<T>::type>::value ||
+          (std::is_convertible<const T&, const void*>::value &&
+           !std::is_convertible<const T&, const char_type*>::value))>
+  FMT_CONSTEXPR auto map(const T&) -> unformattable_pointer {
+    return {};
+  }
+
+  template <typename T, std::size_t N,
+            FMT_ENABLE_IF(!std::is_same<T, wchar_t>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T (&values)[N]) -> const T (&)[N] {
+    return values;
   }
 
   template <typename T,
-            FMT_ENABLE_IF(std::is_enum<T>::value &&
-                          !has_formatter<T, Context>::value &&
-                          !has_fallback_formatter<T, Context>::value)>
-  FMT_CONSTEXPR auto map(const T& val)
+            FMT_ENABLE_IF(
+                std::is_enum<T>::value&& std::is_convertible<T, int>::value &&
+                !has_formatter<T, Context>::value &&
+                !has_fallback_formatter<T, char_type>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T& val)
       -> decltype(std::declval<arg_mapper>().map(
           static_cast<typename std::underlying_type<T>::type>(val))) {
     return map(static_cast<typename std::underlying_type<T>::type>(val));
   }
-  template <typename T,
-            FMT_ENABLE_IF(!is_string<T>::value && !is_char<T>::value &&
-                          (has_formatter<T, Context>::value ||
-                           has_fallback_formatter<T, Context>::value))>
-  FMT_CONSTEXPR const T& map(const T& val) {
+
+  FMT_CONSTEXPR FMT_INLINE auto map(detail::byte val) -> unsigned {
+    return map(static_cast<unsigned char>(val));
+  }
+
+  template <typename T, typename U = remove_cvref_t<T>>
+  struct formattable
+      : bool_constant<has_const_formatter<U, Context>() ||
+                      !std::is_const<remove_reference_t<T>>::value ||
+                      has_fallback_formatter<U, char_type>::value> {};
+
+#if FMT_MSC_VER != 0 && FMT_MSC_VER < 1910
+  // Workaround a bug in MSVC.
+  template <typename T> FMT_CONSTEXPR FMT_INLINE auto do_map(T&& val) -> T& {
+    return val;
+  }
+#else
+  template <typename T, FMT_ENABLE_IF(formattable<T>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto do_map(T&& val) -> T& {
     return val;
   }
+  template <typename T, FMT_ENABLE_IF(!formattable<T>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto do_map(T&&) -> unformattable_const {
+    return {};
+  }
+#endif
+
+  template <typename T, typename U = remove_cvref_t<T>,
+            FMT_ENABLE_IF(!is_string<U>::value && !is_char<U>::value &&
+                          !std::is_array<U>::value &&
+                          (has_formatter<U, Context>::value ||
+                           has_fallback_formatter<U, char_type>::value))>
+  FMT_CONSTEXPR FMT_INLINE auto map(T&& val)
+      -> decltype(this->do_map(std::forward<T>(val))) {
+    return do_map(std::forward<T>(val));
+  }
 
-  template <typename T>
-  FMT_CONSTEXPR auto map(const named_arg<char_type, T>& val)
-      -> decltype(std::declval<arg_mapper>().map(val.value)) {
-    return map(val.value);
+  template <typename T, FMT_ENABLE_IF(is_named_arg<T>::value)>
+  FMT_CONSTEXPR FMT_INLINE auto map(const T& named_arg)
+      -> decltype(std::declval<arg_mapper>().map(named_arg.value)) {
+    return map(named_arg.value);
   }
 
-  unformattable map(...) { return {}; }
+  auto map(...) -> unformattable { return {}; }
 };
 
 // A type constant after applying arg_mapper<Context>.
@@ -1244,7 +1498,28 @@ enum { packed_arg_bits = 4 };
 enum { max_packed_args = 62 / packed_arg_bits };
 enum : unsigned long long { is_unpacked_bit = 1ULL << 63 };
 enum : unsigned long long { has_named_args_bit = 1ULL << 62 };
-}  // namespace detail
+
+FMT_END_DETAIL_NAMESPACE
+
+// An output iterator that appends to a buffer.
+// It is used to reduce symbol sizes for the common case.
+class appender : public std::back_insert_iterator<detail::buffer<char>> {
+  using base = std::back_insert_iterator<detail::buffer<char>>;
+
+  template <typename T>
+  friend auto get_buffer(appender out) -> detail::buffer<char>& {
+    return detail::get_container(out);
+  }
+
+ public:
+  using std::back_insert_iterator<detail::buffer<char>>::back_insert_iterator;
+  appender(base it) FMT_NOEXCEPT : base(it) {}
+  using _Unchecked_type = appender;  // Mark iterator as checked.
+
+  auto operator++() FMT_NOEXCEPT -> appender& { return *this; }
+
+  auto operator++(int) FMT_NOEXCEPT -> appender { return *this; }
+};
 
 // A formatting argument. It is a trivially copyable/constructible type to
 // allow storage in basic_memory_buffer.
@@ -1254,8 +1529,8 @@ template <typename Context> class basic_format_arg {
   detail::type type_;
 
   template <typename ContextType, typename T>
-  friend FMT_CONSTEXPR basic_format_arg<ContextType> detail::make_arg(
-      const T& value);
+  friend FMT_CONSTEXPR auto detail::make_arg(const T& value)
+      -> basic_format_arg<ContextType>;
 
   template <typename Visitor, typename Ctx>
   friend FMT_CONSTEXPR auto visit_format_arg(Visitor&& vis,
@@ -1293,10 +1568,12 @@ template <typename Context> class basic_format_arg {
     return type_ != detail::type::none_type;
   }
 
-  detail::type type() const { return type_; }
+  auto type() const -> detail::type { return type_; }
 
-  bool is_integral() const { return detail::is_integral_type(type_); }
-  bool is_arithmetic() const { return detail::is_arithmetic_type(type_); }
+  auto is_integral() const -> bool { return detail::is_integral_type(type_); }
+  auto is_arithmetic() const -> bool {
+    return detail::is_arithmetic_type(type_);
+  }
 };
 
 /**
@@ -1307,9 +1584,8 @@ template <typename Context> class basic_format_arg {
   \endrst
  */
 template <typename Visitor, typename Context>
-FMT_CONSTEXPR_DECL FMT_INLINE auto visit_format_arg(
+FMT_CONSTEXPR FMT_INLINE auto visit_format_arg(
     Visitor&& vis, const basic_format_arg<Context>& arg) -> decltype(vis(0)) {
-  using char_type = typename Context::char_type;
   switch (arg.type_) {
   case detail::type::none_type:
     break;
@@ -1321,16 +1597,10 @@ FMT_CONSTEXPR_DECL FMT_INLINE auto visit_format_arg(
     return vis(arg.value_.long_long_value);
   case detail::type::ulong_long_type:
     return vis(arg.value_.ulong_long_value);
-#if FMT_USE_INT128
   case detail::type::int128_type:
-    return vis(arg.value_.int128_value);
+    return vis(detail::convert_for_visit(arg.value_.int128_value));
   case detail::type::uint128_type:
-    return vis(arg.value_.uint128_value);
-#else
-  case detail::type::int128_type:
-  case detail::type::uint128_type:
-    break;
-#endif
+    return vis(detail::convert_for_visit(arg.value_.uint128_value));
   case detail::type::bool_type:
     return vis(arg.value_.bool_value);
   case detail::type::char_type:
@@ -1344,8 +1614,8 @@ FMT_CONSTEXPR_DECL FMT_INLINE auto visit_format_arg(
   case detail::type::cstring_type:
     return vis(arg.value_.string.data);
   case detail::type::string_type:
-    return vis(basic_string_view<char_type>(arg.value_.string.data,
-                                            arg.value_.string.size));
+    using sv = basic_string_view<typename Context::char_type>;
+    return vis(sv(arg.value_.string.data, arg.value_.string.size));
   case detail::type::pointer_type:
     return vis(arg.value_.pointer);
   case detail::type::custom_type:
@@ -1354,14 +1624,22 @@ FMT_CONSTEXPR_DECL FMT_INLINE auto visit_format_arg(
   return vis(monostate());
 }
 
-template <typename T> struct formattable : std::false_type {};
+FMT_BEGIN_DETAIL_NAMESPACE
 
-namespace detail {
+template <typename Char, typename InputIt>
+auto copy_str(InputIt begin, InputIt end, appender out) -> appender {
+  get_container(out).append(begin, end);
+  return out;
+}
 
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 500
 // 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;
+#else
+template <typename...> using void_t = void;
+#endif
 
 template <typename It, typename T, typename Enable = void>
 struct is_output_iterator : std::false_type {};
@@ -1384,9 +1662,8 @@ 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 {};
+template <>
+struct is_contiguous_back_insert_iterator<appender> : std::true_type {};
 
 // A type-erased reference to an std::locale to avoid heavy <locale> include.
 class locale_ref {
@@ -1394,97 +1671,72 @@ class locale_ref {
   const void* locale_;  // A type-erased pointer to std::locale.
 
  public:
-  locale_ref() : locale_(nullptr) {}
+  constexpr locale_ref() : locale_(nullptr) {}
   template <typename Locale> explicit locale_ref(const Locale& loc);
 
   explicit operator bool() const FMT_NOEXCEPT { return locale_ != nullptr; }
 
-  template <typename Locale> Locale get() const;
+  template <typename Locale> auto get() const -> Locale;
 };
 
-template <typename> constexpr unsigned long long encode_types() { return 0; }
-
+template <typename> constexpr auto encode_types() -> unsigned long long {
+  return 0;
+}
+
 template <typename Context, typename Arg, typename... Args>
-constexpr unsigned long long encode_types() {
+constexpr auto encode_types() -> unsigned long long {
   return static_cast<unsigned>(mapped_type_constant<Arg, Context>::value) |
          (encode_types<Context, Args...>() << packed_arg_bits);
 }
 
 template <typename Context, typename T>
-FMT_CONSTEXPR basic_format_arg<Context> make_arg(const T& value) {
+FMT_CONSTEXPR auto make_arg(const T& value) -> basic_format_arg<Context> {
   basic_format_arg<Context> arg;
   arg.type_ = mapped_type_constant<T, Context>::value;
   arg.value_ = arg_mapper<Context>().map(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 check<T>(arg_mapper<Context>().map(val));
+FMT_CONSTEXPR FMT_INLINE auto make_arg(T&& val) -> value<Context> {
+  const auto& arg = arg_mapper<Context>().map(std::forward<T>(val));
+
+  constexpr bool formattable_char =
+      !std::is_same<decltype(arg), const unformattable_char&>::value;
+  static_assert(formattable_char, "Mixing character types is disallowed.");
+
+  constexpr bool formattable_const =
+      !std::is_same<decltype(arg), const unformattable_const&>::value;
+  static_assert(formattable_const, "Cannot format a const argument.");
+
+  // Formatting of arbitrary pointers is disallowed. If you want to output
+  // a pointer cast it to "void *" or "const void *". In particular, this
+  // forbids formatting of "[const] volatile char *" which is printed as bool
+  // by iostreams.
+  constexpr bool formattable_pointer =
+      !std::is_same<decltype(arg), const unformattable_pointer&>::value;
+  static_assert(formattable_pointer,
+                "Formatting of non-void pointers is disallowed.");
+
+  constexpr bool formattable =
+      !std::is_same<decltype(arg), const unformattable&>::value;
+  static_assert(
+      formattable,
+      "Cannot format an argument. To make type T formattable provide a "
+      "formatter<T> specialization: https://fmt.dev/latest/api.html#udt");
+  return {arg};
 }
 
 template <bool IS_PACKED, typename Context, type, typename T,
           FMT_ENABLE_IF(!IS_PACKED)>
-inline basic_format_arg<Context> make_arg(const T& value) {
+inline auto make_arg(const T& value) -> basic_format_arg<Context> {
   return make_arg<Context>(value);
 }
-
-template <typename T> struct is_reference_wrapper : std::false_type {};
-template <typename T>
-struct is_reference_wrapper<std::reference_wrapper<T>> : std::true_type {};
-
-template <typename T> const T& unwrap(const T& v) { return v; }
-template <typename T> const T& unwrap(const std::reference_wrapper<T>& v) {
-  return static_cast<const T&>(v);
-}
-
-class dynamic_arg_list {
-  // Workaround for clang's -Wweak-vtables. Unlike for regular classes, for
-  // templates it doesn't complain about inability to deduce single translation
-  // unit for placing vtable. So storage_node_base is made a fake template.
-  template <typename = void> struct node {
-    virtual ~node() = default;
-    std::unique_ptr<node<>> next;
-  };
-
-  template <typename T> struct typed_node : node<> {
-    T value;
-
-    template <typename Arg>
-    FMT_CONSTEXPR typed_node(const Arg& arg) : value(arg) {}
-
-    template <typename Char>
-    FMT_CONSTEXPR typed_node(const basic_string_view<Char>& arg)
-        : value(arg.data(), arg.size()) {}
-  };
-
-  std::unique_ptr<node<>> head_;
-
- public:
-  template <typename T, typename Arg> const T& push(const Arg& arg) {
-    auto new_node = std::unique_ptr<typed_node<T>>(new typed_node<T>(arg));
-    auto& value = new_node->value;
-    new_node->next = std::move(head_);
-    head_ = std::move(new_node);
-    return value;
-  }
-};
-}  // namespace detail
+FMT_END_DETAIL_NAMESPACE
 
 // Formatting context.
 template <typename OutputIt, typename Char> class basic_format_context {
@@ -1503,46 +1755,59 @@ template <typename OutputIt, typename Char> class basic_format_context {
   using parse_context_type = basic_format_parse_context<Char>;
   template <typename T> using formatter_type = formatter<T, char_type>;
 
+  basic_format_context(basic_format_context&&) = default;
   basic_format_context(const basic_format_context&) = delete;
   void operator=(const basic_format_context&) = delete;
   /**
    Constructs a ``basic_format_context`` object. References to the arguments are
    stored in the object so make sure they have appropriate lifetimes.
    */
-  basic_format_context(OutputIt out,
-                       basic_format_args<basic_format_context> ctx_args,
-                       detail::locale_ref loc = detail::locale_ref())
+  constexpr basic_format_context(
+      OutputIt out, basic_format_args<basic_format_context> ctx_args,
+      detail::locale_ref loc = detail::locale_ref())
       : out_(out), args_(ctx_args), loc_(loc) {}
 
-  format_arg arg(int id) const { return args_.get(id); }
-  format_arg arg(basic_string_view<char_type> name) { return args_.get(name); }
-  int arg_id(basic_string_view<char_type> name) { return args_.get_id(name); }
-  const basic_format_args<basic_format_context>& args() const { return args_; }
+  constexpr auto arg(int id) const -> format_arg { return args_.get(id); }
+  FMT_CONSTEXPR auto arg(basic_string_view<char_type> name) -> format_arg {
+    return args_.get(name);
+  }
+  FMT_CONSTEXPR auto arg_id(basic_string_view<char_type> name) -> int {
+    return args_.get_id(name);
+  }
+  auto args() const -> const basic_format_args<basic_format_context>& {
+    return args_;
+  }
 
-  detail::error_handler error_handler() { return {}; }
+  FMT_CONSTEXPR auto error_handler() -> detail::error_handler { return {}; }
   void on_error(const char* message) { error_handler().on_error(message); }
 
   // Returns an iterator to the beginning of the output range.
-  iterator out() { return out_; }
+  FMT_CONSTEXPR auto out() -> iterator { return out_; }
 
   // Advances the begin iterator to ``it``.
   void advance_to(iterator it) {
     if (!detail::is_back_insert_iterator<iterator>()) out_ = it;
   }
 
-  detail::locale_ref locale() { return loc_; }
+  FMT_CONSTEXPR auto locale() -> detail::locale_ref { return loc_; }
 };
 
 template <typename Char>
 using buffer_context =
     basic_format_context<detail::buffer_appender<Char>, Char>;
 using format_context = buffer_context<char>;
-using wformat_context = buffer_context<wchar_t>;
 
 // Workaround an alias issue: https://stackoverflow.com/q/62767544/471164.
 #define FMT_BUFFER_CONTEXT(Char) \
   basic_format_context<detail::buffer_appender<Char>, Char>
 
+template <typename T, typename Char = char>
+using is_formattable = bool_constant<
+    !std::is_base_of<detail::unformattable,
+                     decltype(detail::arg_mapper<buffer_context<Char>>().map(
+                         std::declval<T>()))>::value &&
+    !detail::has_fallback_formatter<T, Char>::value>;
+
 /**
   \rst
   An array of references to arguments. It can be implicitly converted into
@@ -1579,14 +1844,16 @@ class format_arg_store
            : 0);
 
  public:
-  format_arg_store(const Args&... args)
+  template <typename... T>
+  FMT_CONSTEXPR FMT_INLINE format_arg_store(T&&... args)
       :
 #if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
         basic_format_args<Context>(*this),
 #endif
         data_{detail::make_arg<
             is_packed, Context,
-            detail::mapped_type_constant<Args, Context>::value>(args)...} {
+            detail::mapped_type_constant<remove_cvref_t<T>, Context>::value>(
+            std::forward<T>(args))...} {
     detail::init_named_args(data_.named_args(), 0, 0, args...);
   }
 };
@@ -1600,36 +1867,16 @@ class format_arg_store
   \endrst
  */
 template <typename Context = format_context, typename... Args>
-inline format_arg_store<Context, Args...> make_format_args(
-    const Args&... args) {
-  return {args...};
-}
-
-/**
-  \rst
-  Constructs a `~fmt::format_arg_store` object that contains references
-  to arguments and can be implicitly converted to `~fmt::format_args`.
-  If ``format_str`` is a compile-time string then `make_args_checked` checks
-  its validity at compile time.
-  \endrst
- */
-template <typename... Args, typename S, typename Char = char_t<S>>
-inline auto make_args_checked(const S& format_str,
-                              const remove_reference_t<Args>&... args)
-    -> format_arg_store<buffer_context<Char>, remove_reference_t<Args>...> {
-  static_assert(
-      detail::count<(
-              std::is_base_of<detail::view, remove_reference_t<Args>>::value &&
-              std::is_reference<Args>::value)...>() == 0,
-      "passing views as lvalues is disallowed");
-  detail::check_format_string<Args...>(format_str);
-  return {args...};
+constexpr auto make_format_args(Args&&... args)
+    -> format_arg_store<Context, remove_cvref_t<Args>...> {
+  return {std::forward<Args>(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.
+  Returns a named argument to be used in a formatting function.
+  It should only be used in a call to a formatting function or
+  `dynamic_format_arg_store::push_back`.
 
   **Example**::
 
@@ -1637,184 +1884,11 @@ inline auto make_args_checked(const S& format_str,
   \endrst
  */
 template <typename Char, typename T>
-inline detail::named_arg<Char, T> arg(const Char* name, const T& arg) {
+inline auto arg(const Char* name, const T& arg) -> detail::named_arg<Char, T> {
   static_assert(!detail::is_named_arg<T>(), "nested named arguments");
   return {name, arg};
 }
 
-/**
-  \rst
-  A dynamic version of `fmt::format_arg_store`.
-  It's equipped with a storage to potentially temporary objects which lifetimes
-  could be shorter than the format arguments object.
-
-  It can be implicitly converted into `~fmt::basic_format_args` for passing
-  into type-erased formatting functions such as `~fmt::vformat`.
-  \endrst
- */
-template <typename Context>
-class dynamic_format_arg_store
-#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
-    // Workaround a GCC template argument substitution bug.
-    : public basic_format_args<Context>
-#endif
-{
- private:
-  using char_type = typename Context::char_type;
-
-  template <typename T> struct need_copy {
-    static constexpr detail::type mapped_type =
-        detail::mapped_type_constant<T, Context>::value;
-
-    enum {
-      value = !(detail::is_reference_wrapper<T>::value ||
-                std::is_same<T, basic_string_view<char_type>>::value ||
-                std::is_same<T, detail::std_string_view<char_type>>::value ||
-                (mapped_type != detail::type::cstring_type &&
-                 mapped_type != detail::type::string_type &&
-                 mapped_type != detail::type::custom_type))
-    };
-  };
-
-  template <typename T>
-  using stored_type = conditional_t<detail::is_string<T>::value,
-                                    std::basic_string<char_type>, T>;
-
-  // Storage of basic_format_arg must be contiguous.
-  std::vector<basic_format_arg<Context>> data_;
-  std::vector<detail::named_arg_info<char_type>> named_info_;
-
-  // Storage of arguments not fitting into basic_format_arg must grow
-  // without relocation because items in data_ refer to it.
-  detail::dynamic_arg_list dynamic_args_;
-
-  friend class basic_format_args<Context>;
-
-  unsigned long long get_types() const {
-    return detail::is_unpacked_bit | data_.size() |
-           (named_info_.empty()
-                ? 0ULL
-                : static_cast<unsigned long long>(detail::has_named_args_bit));
-  }
-
-  const basic_format_arg<Context>* data() const {
-    return named_info_.empty() ? data_.data() : data_.data() + 1;
-  }
-
-  template <typename T> void emplace_arg(const T& arg) {
-    data_.emplace_back(detail::make_arg<Context>(arg));
-  }
-
-  template <typename T>
-  void emplace_arg(const detail::named_arg<char_type, T>& arg) {
-    if (named_info_.empty()) {
-      constexpr const detail::named_arg_info<char_type>* zero_ptr{nullptr};
-      data_.insert(data_.begin(), {zero_ptr, 0});
-    }
-    data_.emplace_back(detail::make_arg<Context>(detail::unwrap(arg.value)));
-    auto pop_one = [](std::vector<basic_format_arg<Context>>* data) {
-      data->pop_back();
-    };
-    std::unique_ptr<std::vector<basic_format_arg<Context>>, decltype(pop_one)>
-        guard{&data_, pop_one};
-    named_info_.push_back({arg.name, static_cast<int>(data_.size() - 2u)});
-    data_[0].value_.named_args = {named_info_.data(), named_info_.size()};
-    guard.release();
-  }
-
- public:
-  /**
-    \rst
-    Adds an argument into the dynamic store for later passing to a formatting
-    function.
-
-    Note that custom types and string types (but not string views) are copied
-    into the store dynamically allocating memory if necessary.
-
-    **Example**::
-
-      fmt::dynamic_format_arg_store<fmt::format_context> store;
-      store.push_back(42);
-      store.push_back("abc");
-      store.push_back(1.5f);
-      std::string result = fmt::vformat("{} and {} and {}", store);
-    \endrst
-  */
-  template <typename T> void push_back(const T& arg) {
-    if (detail::const_check(need_copy<T>::value))
-      emplace_arg(dynamic_args_.push<stored_type<T>>(arg));
-    else
-      emplace_arg(detail::unwrap(arg));
-  }
-
-  /**
-    \rst
-    Adds a reference to the argument into the dynamic store for later passing to
-    a formatting function. Supports named arguments wrapped in
-    ``std::reference_wrapper`` via ``std::ref()``/``std::cref()``.
-
-    **Example**::
-
-      fmt::dynamic_format_arg_store<fmt::format_context> store;
-      char str[] = "1234567890";
-      store.push_back(std::cref(str));
-      int a1_val{42};
-      auto a1 = fmt::arg("a1_", a1_val);
-      store.push_back(std::cref(a1));
-
-      // Changing str affects the output but only for string and custom types.
-      str[0] = 'X';
-
-      std::string result = fmt::vformat("{} and {a1_}");
-      assert(result == "X234567890 and 42");
-    \endrst
-  */
-  template <typename T> void push_back(std::reference_wrapper<T> arg) {
-    static_assert(
-        detail::is_named_arg<typename std::remove_cv<T>::type>::value ||
-            need_copy<T>::value,
-        "objects of built-in types and string views are always copied");
-    emplace_arg(arg.get());
-  }
-
-  /**
-    Adds named argument into the dynamic store for later passing to a formatting
-    function. ``std::reference_wrapper`` is supported to avoid copying of the
-    argument.
-  */
-  template <typename T>
-  void push_back(const detail::named_arg<char_type, T>& arg) {
-    const char_type* arg_name =
-        dynamic_args_.push<std::basic_string<char_type>>(arg.name).c_str();
-    if (detail::const_check(need_copy<T>::value)) {
-      emplace_arg(
-          fmt::arg(arg_name, dynamic_args_.push<stored_type<T>>(arg.value)));
-    } else {
-      emplace_arg(fmt::arg(arg_name, arg.value));
-    }
-  }
-
-  /** Erase all elements from the store */
-  void clear() {
-    data_.clear();
-    named_info_.clear();
-    dynamic_args_ = detail::dynamic_arg_list();
-  }
-
-  /**
-    \rst
-    Reserves space to store at least *new_cap* arguments including
-    *new_cap_named* named arguments.
-    \endrst
-  */
-  void reserve(size_t new_cap, size_t new_cap_named) {
-    FMT_ASSERT(new_cap >= new_cap_named,
-               "Set of arguments includes set of named arguments");
-    data_.reserve(new_cap);
-    named_info_.reserve(new_cap_named);
-  }
-};
-
 /**
   \rst
   A view of a collection of formatting arguments. To avoid lifetime issues it
@@ -1846,25 +1920,27 @@ template <typename Context> class basic_format_args {
     const format_arg* args_;
   };
 
-  bool is_packed() const { return (desc_ & detail::is_unpacked_bit) == 0; }
-  bool has_named_args() const {
+  constexpr auto is_packed() const -> bool {
+    return (desc_ & detail::is_unpacked_bit) == 0;
+  }
+  auto has_named_args() const -> bool {
     return (desc_ & detail::has_named_args_bit) != 0;
   }
 
-  detail::type type(int index) const {
+  FMT_CONSTEXPR auto type(int index) const -> detail::type {
     int shift = index * detail::packed_arg_bits;
     unsigned int mask = (1 << detail::packed_arg_bits) - 1;
     return static_cast<detail::type>((desc_ >> shift) & mask);
   }
 
-  basic_format_args(unsigned long long desc,
-                    const detail::value<Context>* values)
+  constexpr FMT_INLINE basic_format_args(unsigned long long desc,
+                                         const detail::value<Context>* values)
       : desc_(desc), values_(values) {}
-  basic_format_args(unsigned long long desc, const format_arg* args)
+  constexpr basic_format_args(unsigned long long desc, const format_arg* args)
       : desc_(desc), args_(args) {}
 
  public:
-  basic_format_args() : desc_(0) {}
+  constexpr basic_format_args() : desc_(0), args_(nullptr) {}
 
   /**
    \rst
@@ -1872,8 +1948,10 @@ template <typename Context> class basic_format_args {
    \endrst
    */
   template <typename... Args>
-  FMT_INLINE basic_format_args(const format_arg_store<Context, Args...>& store)
-      : basic_format_args(store.desc, store.data_.args()) {}
+  constexpr FMT_INLINE basic_format_args(
+      const format_arg_store<Context, Args...>& store)
+      : basic_format_args(format_arg_store<Context, Args...>::desc,
+                          store.data_.args()) {}
 
   /**
    \rst
@@ -1881,7 +1959,8 @@ template <typename Context> class basic_format_args {
    `~fmt::dynamic_format_arg_store`.
    \endrst
    */
-  FMT_INLINE basic_format_args(const dynamic_format_arg_store<Context>& store)
+  constexpr FMT_INLINE basic_format_args(
+      const dynamic_format_arg_store<Context>& store)
       : basic_format_args(store.get_types(), store.data()) {}
 
   /**
@@ -1889,12 +1968,12 @@ template <typename Context> class basic_format_args {
    Constructs a `basic_format_args` object from a dynamic set of arguments.
    \endrst
    */
-  basic_format_args(const format_arg* args, int count)
+  constexpr basic_format_args(const format_arg* args, int count)
       : basic_format_args(detail::is_unpacked_bit | detail::to_unsigned(count),
                           args) {}
 
   /** Returns the argument with the specified id. */
-  format_arg get(int id) const {
+  FMT_CONSTEXPR auto get(int id) const -> format_arg {
     format_arg arg;
     if (!is_packed()) {
       if (id < max_size()) arg = args_[id];
@@ -1907,12 +1986,14 @@ template <typename Context> class basic_format_args {
     return arg;
   }
 
-  template <typename Char> format_arg get(basic_string_view<Char> name) const {
+  template <typename Char>
+  auto get(basic_string_view<Char> name) const -> format_arg {
     int id = get_id(name);
     return id >= 0 ? get(id) : format_arg();
   }
 
-  template <typename Char> int get_id(basic_string_view<Char> name) const {
+  template <typename Char>
+  auto get_id(basic_string_view<Char> name) const -> int {
     if (!has_named_args()) return -1;
     const auto& named_args =
         (is_packed() ? values_[-1] : args_[-1].value_).named_args;
@@ -1922,87 +2003,1149 @@ template <typename Context> class basic_format_args {
     return -1;
   }
 
-  int max_size() const {
+  auto max_size() const -> int {
     unsigned long long max_packed = detail::max_packed_args;
     return static_cast<int>(is_packed() ? max_packed
                                         : desc_ & ~detail::is_unpacked_bit);
   }
 };
 
-#ifdef FMT_ARM_ABI_COMPATIBILITY
 /** An alias to ``basic_format_args<format_context>``. */
-// Separate types would result in shorter symbols but break ABI compatibility
+// A separate type would result in shorter symbols but break ABI compatibility
 // between clang and gcc on ARM (#1919).
 using format_args = basic_format_args<format_context>;
-using wformat_args = basic_format_args<wformat_context>;
-#else
-// DEPRECATED! These are kept for ABI compatibility.
-// It is a separate type rather than an alias to make symbols readable.
-struct format_args : basic_format_args<format_context> {
-  template <typename... Args>
-  FMT_INLINE format_args(const Args&... args) : basic_format_args(args...) {}
+
+// We cannot use enum classes as bit fields because of a gcc bug
+// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61414.
+namespace align {
+enum type { none, left, right, center, numeric };
+}
+using align_t = align::type;
+namespace sign {
+enum type { none, minus, plus, space };
+}
+using sign_t = sign::type;
+
+FMT_BEGIN_DETAIL_NAMESPACE
+
+// Workaround an array initialization issue in gcc 4.8.
+template <typename Char> struct fill_t {
+ private:
+  enum { max_size = 4 };
+  Char data_[max_size] = {Char(' '), Char(0), Char(0), Char(0)};
+  unsigned char size_ = 1;
+
+ public:
+  FMT_CONSTEXPR void operator=(basic_string_view<Char> s) {
+    auto size = s.size();
+    if (size > max_size) return throw_format_error("invalid fill");
+    for (size_t i = 0; i < size; ++i) data_[i] = s[i];
+    size_ = static_cast<unsigned char>(size);
+  }
+
+  constexpr auto size() const -> size_t { return size_; }
+  constexpr auto data() const -> const Char* { return data_; }
+
+  FMT_CONSTEXPR auto operator[](size_t index) -> Char& { return data_[index]; }
+  FMT_CONSTEXPR auto operator[](size_t index) const -> const Char& {
+    return data_[index];
+  }
+};
+FMT_END_DETAIL_NAMESPACE
+
+enum class presentation_type : unsigned char {
+  none,
+  // Integer types should go first,
+  dec,             // 'd'
+  oct,             // 'o'
+  hex_lower,       // 'x'
+  hex_upper,       // 'X'
+  bin_lower,       // 'b'
+  bin_upper,       // 'B'
+  hexfloat_lower,  // 'a'
+  hexfloat_upper,  // 'A'
+  exp_lower,       // 'e'
+  exp_upper,       // 'E'
+  fixed_lower,     // 'f'
+  fixed_upper,     // 'F'
+  general_lower,   // 'g'
+  general_upper,   // 'G'
+  chr,             // 'c'
+  string,          // 's'
+  pointer          // 'p'
+};
+
+// Format specifiers for built-in and string types.
+template <typename Char> struct basic_format_specs {
+  int width;
+  int precision;
+  presentation_type type;
+  align_t align : 4;
+  sign_t sign : 3;
+  bool alt : 1;  // Alternate form ('#').
+  bool localized : 1;
+  detail::fill_t<Char> fill;
+
+  constexpr basic_format_specs()
+      : width(0),
+        precision(-1),
+        type(presentation_type::none),
+        align(align::none),
+        sign(sign::none),
+        alt(false),
+        localized(false) {}
+};
+
+using format_specs = basic_format_specs<char>;
+
+FMT_BEGIN_DETAIL_NAMESPACE
+
+enum class arg_id_kind { none, index, name };
+
+// An argument reference.
+template <typename Char> struct arg_ref {
+  FMT_CONSTEXPR arg_ref() : kind(arg_id_kind::none), val() {}
+
+  FMT_CONSTEXPR explicit arg_ref(int index)
+      : kind(arg_id_kind::index), val(index) {}
+  FMT_CONSTEXPR explicit arg_ref(basic_string_view<Char> name)
+      : kind(arg_id_kind::name), val(name) {}
+
+  FMT_CONSTEXPR auto operator=(int idx) -> arg_ref& {
+    kind = arg_id_kind::index;
+    val.index = idx;
+    return *this;
+  }
+
+  arg_id_kind kind;
+  union value {
+    FMT_CONSTEXPR value(int id = 0) : index{id} {}
+    FMT_CONSTEXPR value(basic_string_view<Char> n) : name(n) {}
+
+    int index;
+    basic_string_view<Char> name;
+  } val;
+};
+
+// Format specifiers with width and precision resolved at formatting rather
+// than parsing time to allow re-using the same parsed specifiers with
+// different sets of arguments (precompilation of format strings).
+template <typename Char>
+struct dynamic_format_specs : basic_format_specs<Char> {
+  arg_ref<Char> width_ref;
+  arg_ref<Char> precision_ref;
+};
+
+struct auto_id {};
+
+// A format specifier handler that sets fields in basic_format_specs.
+template <typename Char> class specs_setter {
+ protected:
+  basic_format_specs<Char>& specs_;
+
+ public:
+  explicit FMT_CONSTEXPR specs_setter(basic_format_specs<Char>& specs)
+      : specs_(specs) {}
+
+  FMT_CONSTEXPR specs_setter(const specs_setter& other)
+      : specs_(other.specs_) {}
+
+  FMT_CONSTEXPR void on_align(align_t align) { specs_.align = align; }
+  FMT_CONSTEXPR void on_fill(basic_string_view<Char> fill) {
+    specs_.fill = fill;
+  }
+  FMT_CONSTEXPR void on_sign(sign_t s) { specs_.sign = s; }
+  FMT_CONSTEXPR void on_hash() { specs_.alt = true; }
+  FMT_CONSTEXPR void on_localized() { specs_.localized = true; }
+
+  FMT_CONSTEXPR void on_zero() {
+    if (specs_.align == align::none) specs_.align = align::numeric;
+    specs_.fill[0] = Char('0');
+  }
+
+  FMT_CONSTEXPR void on_width(int width) { specs_.width = width; }
+  FMT_CONSTEXPR void on_precision(int precision) {
+    specs_.precision = precision;
+  }
+  FMT_CONSTEXPR void end_precision() {}
+
+  FMT_CONSTEXPR void on_type(presentation_type type) { specs_.type = type; }
+};
+
+// Format spec handler that saves references to arguments representing dynamic
+// width and precision to be resolved at formatting time.
+template <typename ParseContext>
+class dynamic_specs_handler
+    : public specs_setter<typename ParseContext::char_type> {
+ public:
+  using char_type = typename ParseContext::char_type;
+
+  FMT_CONSTEXPR dynamic_specs_handler(dynamic_format_specs<char_type>& specs,
+                                      ParseContext& ctx)
+      : specs_setter<char_type>(specs), specs_(specs), context_(ctx) {}
+
+  FMT_CONSTEXPR dynamic_specs_handler(const dynamic_specs_handler& other)
+      : specs_setter<char_type>(other),
+        specs_(other.specs_),
+        context_(other.context_) {}
+
+  template <typename Id> FMT_CONSTEXPR void on_dynamic_width(Id arg_id) {
+    specs_.width_ref = make_arg_ref(arg_id);
+  }
+
+  template <typename Id> FMT_CONSTEXPR void on_dynamic_precision(Id arg_id) {
+    specs_.precision_ref = make_arg_ref(arg_id);
+  }
+
+  FMT_CONSTEXPR void on_error(const char* message) {
+    context_.on_error(message);
+  }
+
+ private:
+  dynamic_format_specs<char_type>& specs_;
+  ParseContext& context_;
+
+  using arg_ref_type = arg_ref<char_type>;
+
+  FMT_CONSTEXPR auto make_arg_ref(int arg_id) -> arg_ref_type {
+    context_.check_arg_id(arg_id);
+    return arg_ref_type(arg_id);
+  }
+
+  FMT_CONSTEXPR auto make_arg_ref(auto_id) -> arg_ref_type {
+    return arg_ref_type(context_.next_arg_id());
+  }
+
+  FMT_CONSTEXPR auto make_arg_ref(basic_string_view<char_type> arg_id)
+      -> arg_ref_type {
+    context_.check_arg_id(arg_id);
+    basic_string_view<char_type> format_str(
+        context_.begin(), to_unsigned(context_.end() - context_.begin()));
+    return arg_ref_type(arg_id);
+  }
+};
+
+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 auto to_ascii(Char value) -> Char {
+  return value;
+}
+template <typename Char, FMT_ENABLE_IF(std::is_enum<Char>::value)>
+constexpr auto to_ascii(Char value) ->
+    typename std::underlying_type<Char>::type {
+  return value;
+}
+
+template <typename Char>
+FMT_CONSTEXPR auto code_point_length(const Char* begin) -> int {
+  if (const_check(sizeof(Char) != 1)) return 1;
+  auto 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";
+  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;
+}
+
+// Return the result via the out param to workaround gcc bug 77539.
+template <bool IS_CONSTEXPR, typename T, typename Ptr = const T*>
+FMT_CONSTEXPR auto find(Ptr first, Ptr last, T value, Ptr& out) -> bool {
+  for (out = first; out != last; ++out) {
+    if (*out == value) return true;
+  }
+  return false;
+}
+
+template <>
+inline auto find<false, char>(const char* first, const char* last, char value,
+                              const char*& out) -> bool {
+  out = static_cast<const char*>(
+      std::memchr(first, value, to_unsigned(last - first)));
+  return out != nullptr;
+}
+
+// Parses the range [begin, end) as an unsigned integer. This function assumes
+// that the range is non-empty and the first character is a digit.
+template <typename Char>
+FMT_CONSTEXPR auto parse_nonnegative_int(const Char*& begin, const Char* end,
+                                         int error_value) noexcept -> int {
+  FMT_ASSERT(begin != end && '0' <= *begin && *begin <= '9', "");
+  unsigned value = 0, prev = 0;
+  auto p = begin;
+  do {
+    prev = value;
+    value = value * 10 + unsigned(*p - '0');
+    ++p;
+  } while (p != end && '0' <= *p && *p <= '9');
+  auto num_digits = p - begin;
+  begin = p;
+  if (num_digits <= std::numeric_limits<int>::digits10)
+    return static_cast<int>(value);
+  // Check for overflow.
+  const unsigned max = to_unsigned((std::numeric_limits<int>::max)());
+  return num_digits == std::numeric_limits<int>::digits10 + 1 &&
+                 prev * 10ull + unsigned(p[-1] - '0') <= max
+             ? static_cast<int>(value)
+             : error_value;
+}
+
+// Parses fill and alignment.
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto parse_align(const Char* begin, const Char* end,
+                               Handler&& handler) -> const Char* {
+  FMT_ASSERT(begin != end, "");
+  auto align = align::none;
+  auto p = begin + code_point_length(begin);
+  if (p >= end) p = begin;
+  for (;;) {
+    switch (to_ascii(*p)) {
+    case '<':
+      align = align::left;
+      break;
+    case '>':
+      align = align::right;
+      break;
+    case '^':
+      align = align::center;
+      break;
+    default:
+      break;
+    }
+    if (align != align::none) {
+      if (p != begin) {
+        auto c = *begin;
+        if (c == '{')
+          return handler.on_error("invalid fill character '{'"), begin;
+        handler.on_fill(basic_string_view<Char>(begin, to_unsigned(p - begin)));
+        begin = p + 1;
+      } else
+        ++begin;
+      handler.on_align(align);
+      break;
+    } else if (p == begin) {
+      break;
+    }
+    p = begin;
+  }
+  return begin;
+}
+
+template <typename Char> FMT_CONSTEXPR bool is_name_start(Char c) {
+  return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || '_' == c;
+}
+
+template <typename Char, typename IDHandler>
+FMT_CONSTEXPR auto do_parse_arg_id(const Char* begin, const Char* end,
+                                   IDHandler&& handler) -> const Char* {
+  FMT_ASSERT(begin != end, "");
+  Char c = *begin;
+  if (c >= '0' && c <= '9') {
+    int index = 0;
+    if (c != '0')
+      index =
+          parse_nonnegative_int(begin, end, (std::numeric_limits<int>::max)());
+    else
+      ++begin;
+    if (begin == end || (*begin != '}' && *begin != ':'))
+      handler.on_error("invalid format string");
+    else
+      handler(index);
+    return begin;
+  }
+  if (!is_name_start(c)) {
+    handler.on_error("invalid format string");
+    return begin;
+  }
+  auto it = begin;
+  do {
+    ++it;
+  } while (it != end && (is_name_start(c = *it) || ('0' <= c && c <= '9')));
+  handler(basic_string_view<Char>(begin, to_unsigned(it - begin)));
+  return it;
+}
+
+template <typename Char, typename IDHandler>
+FMT_CONSTEXPR FMT_INLINE auto parse_arg_id(const Char* begin, const Char* end,
+                                           IDHandler&& handler) -> const Char* {
+  Char c = *begin;
+  if (c != '}' && c != ':') return do_parse_arg_id(begin, end, handler);
+  handler();
+  return begin;
+}
+
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto parse_width(const Char* begin, const Char* end,
+                               Handler&& handler) -> const Char* {
+  using detail::auto_id;
+  struct width_adapter {
+    Handler& handler;
+
+    FMT_CONSTEXPR void operator()() { handler.on_dynamic_width(auto_id()); }
+    FMT_CONSTEXPR void operator()(int id) { handler.on_dynamic_width(id); }
+    FMT_CONSTEXPR void operator()(basic_string_view<Char> id) {
+      handler.on_dynamic_width(id);
+    }
+    FMT_CONSTEXPR void on_error(const char* message) {
+      if (message) handler.on_error(message);
+    }
+  };
+
+  FMT_ASSERT(begin != end, "");
+  if ('0' <= *begin && *begin <= '9') {
+    int width = parse_nonnegative_int(begin, end, -1);
+    if (width != -1)
+      handler.on_width(width);
+    else
+      handler.on_error("number is too big");
+  } else if (*begin == '{') {
+    ++begin;
+    if (begin != end) begin = parse_arg_id(begin, end, width_adapter{handler});
+    if (begin == end || *begin != '}')
+      return handler.on_error("invalid format string"), begin;
+    ++begin;
+  }
+  return begin;
+}
+
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto parse_precision(const Char* begin, const Char* end,
+                                   Handler&& handler) -> const Char* {
+  using detail::auto_id;
+  struct precision_adapter {
+    Handler& handler;
+
+    FMT_CONSTEXPR void operator()() { handler.on_dynamic_precision(auto_id()); }
+    FMT_CONSTEXPR void operator()(int id) { handler.on_dynamic_precision(id); }
+    FMT_CONSTEXPR void operator()(basic_string_view<Char> id) {
+      handler.on_dynamic_precision(id);
+    }
+    FMT_CONSTEXPR void on_error(const char* message) {
+      if (message) handler.on_error(message);
+    }
+  };
+
+  ++begin;
+  auto c = begin != end ? *begin : Char();
+  if ('0' <= c && c <= '9') {
+    auto precision = parse_nonnegative_int(begin, end, -1);
+    if (precision != -1)
+      handler.on_precision(precision);
+    else
+      handler.on_error("number is too big");
+  } else if (c == '{') {
+    ++begin;
+    if (begin != end)
+      begin = parse_arg_id(begin, end, precision_adapter{handler});
+    if (begin == end || *begin++ != '}')
+      return handler.on_error("invalid format string"), begin;
+  } else {
+    return handler.on_error("missing precision specifier"), begin;
+  }
+  handler.end_precision();
+  return begin;
+}
+
+template <typename Char>
+FMT_CONSTEXPR auto parse_presentation_type(Char type) -> presentation_type {
+  switch (to_ascii(type)) {
+  case 'd':
+    return presentation_type::dec;
+  case 'o':
+    return presentation_type::oct;
+  case 'x':
+    return presentation_type::hex_lower;
+  case 'X':
+    return presentation_type::hex_upper;
+  case 'b':
+    return presentation_type::bin_lower;
+  case 'B':
+    return presentation_type::bin_upper;
+  case 'a':
+    return presentation_type::hexfloat_lower;
+  case 'A':
+    return presentation_type::hexfloat_upper;
+  case 'e':
+    return presentation_type::exp_lower;
+  case 'E':
+    return presentation_type::exp_upper;
+  case 'f':
+    return presentation_type::fixed_lower;
+  case 'F':
+    return presentation_type::fixed_upper;
+  case 'g':
+    return presentation_type::general_lower;
+  case 'G':
+    return presentation_type::general_upper;
+  case 'c':
+    return presentation_type::chr;
+  case 's':
+    return presentation_type::string;
+  case 'p':
+    return presentation_type::pointer;
+  default:
+    return presentation_type::none;
+  }
+}
+
+// Parses standard format specifiers and sends notifications about parsed
+// components to handler.
+template <typename Char, typename SpecHandler>
+FMT_CONSTEXPR FMT_INLINE auto parse_format_specs(const Char* begin,
+                                                 const Char* end,
+                                                 SpecHandler&& handler)
+    -> const Char* {
+  if (1 < end - begin && begin[1] == '}' && is_ascii_letter(*begin) &&
+      *begin != 'L') {
+    presentation_type type = parse_presentation_type(*begin++);
+    if (type == presentation_type::none)
+      handler.on_error("invalid type specifier");
+    handler.on_type(type);
+    return begin;
+  }
+
+  if (begin == end) return begin;
+
+  begin = parse_align(begin, end, handler);
+  if (begin == end) return begin;
+
+  // Parse sign.
+  switch (to_ascii(*begin)) {
+  case '+':
+    handler.on_sign(sign::plus);
+    ++begin;
+    break;
+  case '-':
+    handler.on_sign(sign::minus);
+    ++begin;
+    break;
+  case ' ':
+    handler.on_sign(sign::space);
+    ++begin;
+    break;
+  default:
+    break;
+  }
+  if (begin == end) return begin;
+
+  if (*begin == '#') {
+    handler.on_hash();
+    if (++begin == end) return begin;
+  }
+
+  // Parse zero flag.
+  if (*begin == '0') {
+    handler.on_zero();
+    if (++begin == end) return begin;
+  }
+
+  begin = parse_width(begin, end, handler);
+  if (begin == end) return begin;
+
+  // Parse precision.
+  if (*begin == '.') {
+    begin = parse_precision(begin, end, handler);
+    if (begin == end) return begin;
+  }
+
+  if (*begin == 'L') {
+    handler.on_localized();
+    ++begin;
+  }
+
+  // Parse type.
+  if (begin != end && *begin != '}') {
+    presentation_type type = parse_presentation_type(*begin++);
+    if (type == presentation_type::none)
+      handler.on_error("invalid type specifier");
+    handler.on_type(type);
+  }
+  return begin;
+}
+
+template <typename Char, typename Handler>
+FMT_CONSTEXPR auto parse_replacement_field(const Char* begin, const Char* end,
+                                           Handler&& handler) -> const Char* {
+  struct id_adapter {
+    Handler& handler;
+    int arg_id;
+
+    FMT_CONSTEXPR void operator()() { arg_id = handler.on_arg_id(); }
+    FMT_CONSTEXPR void operator()(int id) { arg_id = handler.on_arg_id(id); }
+    FMT_CONSTEXPR void operator()(basic_string_view<Char> id) {
+      arg_id = handler.on_arg_id(id);
+    }
+    FMT_CONSTEXPR void on_error(const char* message) {
+      if (message) handler.on_error(message);
+    }
+  };
+
+  ++begin;
+  if (begin == end) return handler.on_error("invalid format string"), end;
+  if (*begin == '}') {
+    handler.on_replacement_field(handler.on_arg_id(), begin);
+  } else if (*begin == '{') {
+    handler.on_text(begin, begin + 1);
+  } else {
+    auto adapter = id_adapter{handler, 0};
+    begin = parse_arg_id(begin, end, adapter);
+    Char c = begin != end ? *begin : Char();
+    if (c == '}') {
+      handler.on_replacement_field(adapter.arg_id, begin);
+    } else if (c == ':') {
+      begin = handler.on_format_specs(adapter.arg_id, begin + 1, end);
+      if (begin == end || *begin != '}')
+        return handler.on_error("unknown format specifier"), end;
+    } else {
+      return handler.on_error("missing '}' in format string"), end;
+    }
+  }
+  return begin + 1;
+}
+
+template <bool IS_CONSTEXPR, typename Char, typename Handler>
+FMT_CONSTEXPR FMT_INLINE void parse_format_string(
+    basic_string_view<Char> format_str, Handler&& handler) {
+  // Workaround a name-lookup bug in MSVC's modules implementation.
+  using detail::find;
+
+  auto begin = format_str.data();
+  auto end = begin + format_str.size();
+  if (end - begin < 32) {
+    // Use a simple loop instead of memchr for small strings.
+    const Char* p = begin;
+    while (p != end) {
+      auto c = *p++;
+      if (c == '{') {
+        handler.on_text(begin, p - 1);
+        begin = p = parse_replacement_field(p - 1, end, handler);
+      } else if (c == '}') {
+        if (p == end || *p != '}')
+          return handler.on_error("unmatched '}' in format string");
+        handler.on_text(begin, p);
+        begin = ++p;
+      }
+    }
+    handler.on_text(begin, end);
+    return;
+  }
+  struct writer {
+    FMT_CONSTEXPR void operator()(const Char* pbegin, const Char* pend) {
+      if (pbegin == pend) return;
+      for (;;) {
+        const Char* p = nullptr;
+        if (!find<IS_CONSTEXPR>(pbegin, pend, Char('}'), p))
+          return handler_.on_text(pbegin, pend);
+        ++p;
+        if (p == pend || *p != '}')
+          return handler_.on_error("unmatched '}' in format string");
+        handler_.on_text(pbegin, p);
+        pbegin = p + 1;
+      }
+    }
+    Handler& handler_;
+  } write{handler};
+  while (begin != end) {
+    // Doing two passes with memchr (one for '{' and another for '}') is up to
+    // 2.5x faster than the naive one-pass implementation on big format strings.
+    const Char* p = begin;
+    if (*begin != '{' && !find<IS_CONSTEXPR>(begin + 1, end, Char('{'), p))
+      return write(begin, end);
+    write(begin, p);
+    begin = parse_replacement_field(p, end, handler);
+  }
+}
+
+template <typename T, typename ParseContext>
+FMT_CONSTEXPR auto parse_format_specs(ParseContext& ctx)
+    -> decltype(ctx.begin()) {
+  using char_type = typename ParseContext::char_type;
+  using context = buffer_context<char_type>;
+  using mapped_type = conditional_t<
+      mapped_type_constant<T, context>::value != type::custom_type,
+      decltype(arg_mapper<context>().map(std::declval<const T&>())), T>;
+  auto f = conditional_t<has_formatter<mapped_type, context>::value,
+                         formatter<mapped_type, char_type>,
+                         fallback_formatter<T, char_type>>();
+  return f.parse(ctx);
+}
+
+// A parse context with extra argument id checks. It is only used at compile
+// time because adding checks at runtime would introduce substantial overhead
+// and would be redundant since argument ids are checked when arguments are
+// retrieved anyway.
+template <typename Char, typename ErrorHandler = error_handler>
+class compile_parse_context
+    : public basic_format_parse_context<Char, ErrorHandler> {
+ private:
+  int num_args_;
+  using base = basic_format_parse_context<Char, ErrorHandler>;
+
+ public:
+  explicit FMT_CONSTEXPR compile_parse_context(
+      basic_string_view<Char> format_str,
+      int num_args = (std::numeric_limits<int>::max)(), ErrorHandler eh = {})
+      : base(format_str, eh), num_args_(num_args) {}
+
+  FMT_CONSTEXPR auto next_arg_id() -> int {
+    int id = base::next_arg_id();
+    if (id >= num_args_) this->on_error("argument not found");
+    return id;
+  }
+
+  FMT_CONSTEXPR void check_arg_id(int id) {
+    base::check_arg_id(id);
+    if (id >= num_args_) this->on_error("argument not found");
+  }
+  using base::check_arg_id;
 };
-struct wformat_args : basic_format_args<wformat_context> {
-  using basic_format_args::basic_format_args;
+
+template <typename ErrorHandler>
+FMT_CONSTEXPR void check_int_type_spec(presentation_type type,
+                                       ErrorHandler&& eh) {
+  if (type > presentation_type::bin_upper && type != presentation_type::chr)
+    eh.on_error("invalid type specifier");
+}
+
+// Checks char specs and returns true if the type spec is char (and not int).
+template <typename Char, typename ErrorHandler = error_handler>
+FMT_CONSTEXPR auto check_char_specs(const basic_format_specs<Char>& specs,
+                                    ErrorHandler&& eh = {}) -> bool {
+  if (specs.type != presentation_type::none &&
+      specs.type != presentation_type::chr) {
+    check_int_type_spec(specs.type, eh);
+    return false;
+  }
+  if (specs.align == align::numeric || specs.sign != sign::none || specs.alt)
+    eh.on_error("invalid format specifier for char");
+  return true;
+}
+
+// A floating-point presentation format.
+enum class float_format : unsigned char {
+  general,  // General: exponent notation or fixed point based on magnitude.
+  exp,      // Exponent notation with the default precision of 6, e.g. 1.2e-3.
+  fixed,    // Fixed point with the default precision of 6, e.g. 0.0012.
+  hex
+};
+
+struct float_specs {
+  int precision;
+  float_format format : 8;
+  sign_t sign : 8;
+  bool upper : 1;
+  bool locale : 1;
+  bool binary32 : 1;
+  bool fallback : 1;
+  bool showpoint : 1;
+};
+
+template <typename ErrorHandler = error_handler, typename Char>
+FMT_CONSTEXPR auto parse_float_type_spec(const basic_format_specs<Char>& specs,
+                                         ErrorHandler&& eh = {})
+    -> float_specs {
+  auto result = float_specs();
+  result.showpoint = specs.alt;
+  result.locale = specs.localized;
+  switch (specs.type) {
+  case presentation_type::none:
+    result.format = float_format::general;
+    break;
+  case presentation_type::general_upper:
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case presentation_type::general_lower:
+    result.format = float_format::general;
+    break;
+  case presentation_type::exp_upper:
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case presentation_type::exp_lower:
+    result.format = float_format::exp;
+    result.showpoint |= specs.precision != 0;
+    break;
+  case presentation_type::fixed_upper:
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case presentation_type::fixed_lower:
+    result.format = float_format::fixed;
+    result.showpoint |= specs.precision != 0;
+    break;
+  case presentation_type::hexfloat_upper:
+    result.upper = true;
+    FMT_FALLTHROUGH;
+  case presentation_type::hexfloat_lower:
+    result.format = float_format::hex;
+    break;
+  default:
+    eh.on_error("invalid type specifier");
+    break;
+  }
+  return result;
+}
+
+template <typename ErrorHandler = error_handler>
+FMT_CONSTEXPR auto check_cstring_type_spec(presentation_type type,
+                                           ErrorHandler&& eh = {}) -> bool {
+  if (type == presentation_type::none || type == presentation_type::string)
+    return true;
+  if (type != presentation_type::pointer) eh.on_error("invalid type specifier");
+  return false;
+}
+
+template <typename ErrorHandler = error_handler>
+FMT_CONSTEXPR void check_string_type_spec(presentation_type type,
+                                          ErrorHandler&& eh = {}) {
+  if (type != presentation_type::none && type != presentation_type::string)
+    eh.on_error("invalid type specifier");
+}
+
+template <typename ErrorHandler>
+FMT_CONSTEXPR void check_pointer_type_spec(presentation_type type,
+                                           ErrorHandler&& eh) {
+  if (type != presentation_type::none && type != presentation_type::pointer)
+    eh.on_error("invalid type specifier");
+}
+
+// A parse_format_specs handler that checks if specifiers are consistent with
+// the argument type.
+template <typename Handler> class specs_checker : public Handler {
+ private:
+  detail::type arg_type_;
+
+  FMT_CONSTEXPR void require_numeric_argument() {
+    if (!is_arithmetic_type(arg_type_))
+      this->on_error("format specifier requires numeric argument");
+  }
+
+ public:
+  FMT_CONSTEXPR specs_checker(const Handler& handler, detail::type arg_type)
+      : Handler(handler), arg_type_(arg_type) {}
+
+  FMT_CONSTEXPR void on_align(align_t align) {
+    if (align == align::numeric) require_numeric_argument();
+    Handler::on_align(align);
+  }
+
+  FMT_CONSTEXPR void on_sign(sign_t s) {
+    require_numeric_argument();
+    if (is_integral_type(arg_type_) && arg_type_ != type::int_type &&
+        arg_type_ != type::long_long_type && arg_type_ != type::char_type) {
+      this->on_error("format specifier requires signed argument");
+    }
+    Handler::on_sign(s);
+  }
+
+  FMT_CONSTEXPR void on_hash() {
+    require_numeric_argument();
+    Handler::on_hash();
+  }
+
+  FMT_CONSTEXPR void on_localized() {
+    require_numeric_argument();
+    Handler::on_localized();
+  }
+
+  FMT_CONSTEXPR void on_zero() {
+    require_numeric_argument();
+    Handler::on_zero();
+  }
+
+  FMT_CONSTEXPR void end_precision() {
+    if (is_integral_type(arg_type_) || arg_type_ == type::pointer_type)
+      this->on_error("precision not allowed for this argument type");
+  }
 };
+
+constexpr int invalid_arg_index = -1;
+
+#if FMT_USE_NONTYPE_TEMPLATE_PARAMETERS
+template <int N, typename T, typename... Args, typename Char>
+constexpr auto get_arg_index_by_name(basic_string_view<Char> name) -> int {
+  if constexpr (detail::is_statically_named_arg<T>()) {
+    if (name == T::name) return N;
+  }
+  if constexpr (sizeof...(Args) > 0)
+    return get_arg_index_by_name<N + 1, Args...>(name);
+  (void)name;  // Workaround an MSVC bug about "unused" parameter.
+  return invalid_arg_index;
+}
+#endif
+
+template <typename... Args, typename Char>
+FMT_CONSTEXPR auto get_arg_index_by_name(basic_string_view<Char> name) -> int {
+#if FMT_USE_NONTYPE_TEMPLATE_PARAMETERS
+  if constexpr (sizeof...(Args) > 0)
+    return get_arg_index_by_name<0, Args...>(name);
+#endif
+  (void)name;
+  return invalid_arg_index;
+}
+
+template <typename Char, typename ErrorHandler, typename... Args>
+class format_string_checker {
+ private:
+  using parse_context_type = compile_parse_context<Char, ErrorHandler>;
+  enum { num_args = sizeof...(Args) };
+
+  // Format specifier parsing function.
+  using parse_func = const Char* (*)(parse_context_type&);
+
+  parse_context_type context_;
+  parse_func parse_funcs_[num_args > 0 ? num_args : 1];
+
+ public:
+  explicit FMT_CONSTEXPR format_string_checker(
+      basic_string_view<Char> format_str, ErrorHandler eh)
+      : context_(format_str, num_args, eh),
+        parse_funcs_{&parse_format_specs<Args, parse_context_type>...} {}
+
+  FMT_CONSTEXPR void on_text(const Char*, const Char*) {}
+
+  FMT_CONSTEXPR auto on_arg_id() -> int { return context_.next_arg_id(); }
+  FMT_CONSTEXPR auto on_arg_id(int id) -> int {
+    return context_.check_arg_id(id), id;
+  }
+  FMT_CONSTEXPR auto on_arg_id(basic_string_view<Char> id) -> int {
+#if FMT_USE_NONTYPE_TEMPLATE_PARAMETERS
+    auto index = get_arg_index_by_name<Args...>(id);
+    if (index == invalid_arg_index) on_error("named argument is not found");
+    return context_.check_arg_id(index), index;
+#else
+    (void)id;
+    on_error("compile-time checks for named arguments require C++20 support");
+    return 0;
 #endif
+  }
+
+  FMT_CONSTEXPR void on_replacement_field(int, const Char*) {}
 
-namespace detail {
+  FMT_CONSTEXPR auto on_format_specs(int id, const Char* begin, const Char*)
+      -> const Char* {
+    context_.advance_to(context_.begin() + (begin - &*context_.begin()));
+    // id >= 0 check is a workaround for gcc 10 bug (#2065).
+    return id >= 0 && id < num_args ? parse_funcs_[id](context_) : begin;
+  }
 
-template <typename Char, FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
-std::basic_string<Char> vformat(
-    basic_string_view<Char> format_str,
-    basic_format_args<buffer_context<type_identity_t<Char>>> args);
+  FMT_CONSTEXPR void on_error(const char* message) {
+    context_.on_error(message);
+  }
+};
 
-FMT_API std::string vformat(string_view format_str, format_args args);
+template <typename... Args, typename S,
+          enable_if_t<(is_compile_string<S>::value), int>>
+void check_format_string(S format_str) {
+  FMT_CONSTEXPR auto s = to_string_view(format_str);
+  using checker = format_string_checker<typename S::char_type, error_handler,
+                                        remove_cvref_t<Args>...>;
+  FMT_CONSTEXPR bool invalid_format =
+      (parse_format_string<true>(s, checker(s, {})), true);
+  ignore_unused(invalid_format);
+}
 
 template <typename Char>
 void vformat_to(
-    buffer<Char>& buf, basic_string_view<Char> format_str,
+    buffer<Char>& buf, basic_string_view<Char> fmt,
     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)>
-inline void vprint_mojibake(std::FILE*, basic_string_view<Char>, const Args&) {}
+    locale_ref loc = {});
 
 FMT_API void vprint_mojibake(std::FILE*, string_view, format_args);
 #ifndef _WIN32
 inline void vprint_mojibake(std::FILE*, string_view, format_args) {}
 #endif
-}  // namespace detail
+FMT_END_DETAIL_NAMESPACE
+
+// A formatter specialization for the core types corresponding to detail::type
+// constants.
+template <typename T, typename Char>
+struct formatter<T, Char,
+                 enable_if_t<detail::type_constant<T, Char>::value !=
+                             detail::type::custom_type>> {
+ private:
+  detail::dynamic_format_specs<Char> specs_;
+
+ public:
+  // Parses format specifiers stopping either at the end of the range or at the
+  // terminating '}'.
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    auto begin = ctx.begin(), end = ctx.end();
+    if (begin == end) return begin;
+    using handler_type = detail::dynamic_specs_handler<ParseContext>;
+    auto type = detail::type_constant<T, Char>::value;
+    auto checker =
+        detail::specs_checker<handler_type>(handler_type(specs_, ctx), type);
+    auto it = detail::parse_format_specs(begin, end, checker);
+    auto eh = ctx.error_handler();
+    switch (type) {
+    case detail::type::none_type:
+      FMT_ASSERT(false, "invalid argument type");
+      break;
+    case detail::type::bool_type:
+      if (specs_.type == presentation_type::none ||
+          specs_.type == presentation_type::string) {
+        break;
+      }
+      FMT_FALLTHROUGH;
+    case detail::type::int_type:
+    case detail::type::uint_type:
+    case detail::type::long_long_type:
+    case detail::type::ulong_long_type:
+    case detail::type::int128_type:
+    case detail::type::uint128_type:
+      detail::check_int_type_spec(specs_.type, eh);
+      break;
+    case detail::type::char_type:
+      detail::check_char_specs(specs_, eh);
+      break;
+    case detail::type::float_type:
+      if (detail::const_check(FMT_USE_FLOAT))
+        detail::parse_float_type_spec(specs_, eh);
+      else
+        FMT_ASSERT(false, "float support disabled");
+      break;
+    case detail::type::double_type:
+      if (detail::const_check(FMT_USE_DOUBLE))
+        detail::parse_float_type_spec(specs_, eh);
+      else
+        FMT_ASSERT(false, "double support disabled");
+      break;
+    case detail::type::long_double_type:
+      if (detail::const_check(FMT_USE_LONG_DOUBLE))
+        detail::parse_float_type_spec(specs_, eh);
+      else
+        FMT_ASSERT(false, "long double support disabled");
+      break;
+    case detail::type::cstring_type:
+      detail::check_cstring_type_spec(specs_.type, eh);
+      break;
+    case detail::type::string_type:
+      detail::check_string_type_spec(specs_.type, eh);
+      break;
+    case detail::type::pointer_type:
+      detail::check_pointer_type_spec(specs_.type, eh);
+      break;
+    case detail::type::custom_type:
+      // Custom format specifiers are checked in parse functions of
+      // formatter specializations.
+      break;
+    }
+    return it;
+  }
+
+  template <typename FormatContext>
+  FMT_CONSTEXPR auto format(const T& val, FormatContext& ctx) const
+      -> decltype(ctx.out());
+};
+
+template <typename Char> struct basic_runtime { basic_string_view<Char> str; };
+
+/** A compile-time format string. */
+template <typename Char, typename... Args> class basic_format_string {
+ private:
+  basic_string_view<Char> str_;
+
+ public:
+  template <typename S,
+            FMT_ENABLE_IF(
+                std::is_convertible<const S&, basic_string_view<Char>>::value)>
+  FMT_CONSTEVAL FMT_INLINE basic_format_string(const S& s) : str_(s) {
+    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");
+#ifdef FMT_HAS_CONSTEVAL
+    if constexpr (detail::count_named_args<Args...>() ==
+                  detail::count_statically_named_args<Args...>()) {
+      using checker = detail::format_string_checker<Char, detail::error_handler,
+                                                    remove_cvref_t<Args>...>;
+      detail::parse_format_string<true>(str_, checker(s, {}));
+    }
+#else
+    detail::check_format_string<Args...>(s);
+#endif
+  }
+  basic_format_string(basic_runtime<Char> r) : str_(r.str) {}
+
+  FMT_INLINE operator basic_string_view<Char>() const { return str_; }
+};
+
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+// Workaround broken conversion on older gcc.
+template <typename... Args> using format_string = string_view;
+template <typename S> auto runtime(const S& s) -> basic_string_view<char_t<S>> {
+  return s;
+}
+#else
+template <typename... Args>
+using format_string = basic_format_string<char, type_identity_t<Args>...>;
+/**
+  \rst
+  Creates a runtime format string.
+
+  **Example**::
+
+    // Check format string at runtime instead of compile-time.
+    fmt::print(fmt::runtime("{:d}"), "I am not a number");
+  \endrst
+ */
+template <typename S> auto runtime(const S& s) -> basic_runtime<char_t<S>> {
+  return {{s}};
+}
+#endif
+
+FMT_API auto vformat(string_view fmt, format_args args) -> std::string;
+
+/**
+  \rst
+  Formats ``args`` according to specifications in ``fmt`` and returns the result
+  as a string.
+
+  **Example**::
+
+    #include <fmt/core.h>
+    std::string message = fmt::format("The answer is {}.", 42);
+  \endrst
+*/
+template <typename... T>
+FMT_NODISCARD FMT_INLINE auto format(format_string<T...> fmt, T&&... args)
+    -> std::string {
+  return vformat(fmt, fmt::make_format_args(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>,
-          bool enable = detail::is_output_iterator<OutputIt, Char>::value>
-auto vformat_to(OutputIt out, const S& format_str,
-                basic_format_args<buffer_context<type_identity_t<Char>>> args)
-    -> typename std::enable_if<enable, OutputIt>::type {
-  decltype(detail::get_buffer<Char>(out)) buf(detail::get_buffer_init(out));
-  detail::vformat_to(buf, to_string_view(format_str), args);
+template <typename OutputIt,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+auto vformat_to(OutputIt out, string_view fmt, format_args args) -> OutputIt {
+  using detail::get_buffer;
+  auto&& buf = get_buffer<char>(out);
+  detail::vformat_to(buf, fmt, args, {});
   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.
+ Formats ``args`` according to specifications in ``fmt``, writes the result to
+ the output iterator ``out`` and returns the iterator past the end of the output
+ range. `format_to` does not append a terminating null character.
 
  **Example**::
 
-   std::vector<char> out;
+   auto out = std::vector<char>();
    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, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+FMT_INLINE auto format_to(OutputIt out, format_string<T...> fmt, T&&... args)
+    -> OutputIt {
+  return vformat_to(out, fmt, fmt::make_format_args(args...));
 }
 
 template <typename OutputIt> struct format_to_n_result {
@@ -2012,111 +3155,82 @@ template <typename OutputIt> struct format_to_n_result {
   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);
+template <typename OutputIt, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+auto vformat_to_n(OutputIt out, size_t n, string_view fmt, format_args args)
+    -> format_to_n_result<OutputIt> {
+  using traits = detail::fixed_buffer_traits;
+  auto buf = detail::iterator_buffer<OutputIt, char, traits>(out, n);
+  detail::vformat_to(buf, fmt, 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
+  \rst
+  Formats ``args`` according to specifications in ``fmt``, writes up to ``n``
+  characters of the result to the output iterator ``out`` and returns the total
+  (not truncated) output size and the iterator past the end of the output range.
+  `format_to_n` does not append a terminating null character.
+  \endrst
  */
-template <typename OutputIt, typename S, typename... Args,
-          bool enable = detail::is_output_iterator<OutputIt, char_t<S>>::value>
-inline auto format_to_n(OutputIt out, size_t n, const S& format_str,
-                        const Args&... args) ->
-    typename std::enable_if<enable, format_to_n_result<OutputIt>>::type {
-  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
-  return vformat_to_n(out, n, to_string_view(format_str), vargs);
+template <typename OutputIt, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value)>
+FMT_INLINE auto format_to_n(OutputIt out, size_t n, format_string<T...> fmt,
+                            T&&... args) -> format_to_n_result<OutputIt> {
+  return vformat_to_n(out, n, fmt, fmt::make_format_args(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);
+/** Returns the number of chars in the output of ``format(fmt, args...)``. */
+template <typename... T>
+FMT_NODISCARD FMT_INLINE auto formatted_size(format_string<T...> fmt,
+                                             T&&... args) -> size_t {
+  auto buf = detail::counting_buffer<>();
+  detail::vformat_to(buf, string_view(fmt), fmt::make_format_args(args...), {});
   return buf.count();
 }
 
-template <typename S, typename Char = char_t<S>>
-FMT_INLINE std::basic_string<Char> vformat(
-    const S& format_str,
-    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
-  return detail::vformat(to_string_view(format_str), args);
-}
+FMT_API void vprint(string_view fmt, format_args args);
+FMT_API void vprint(std::FILE* f, string_view fmt, format_args args);
 
 /**
   \rst
-  Formats arguments and returns the result as a string.
+  Formats ``args`` according to specifications in ``fmt`` and writes the output
+  to ``stdout``.
 
   **Example**::
 
-    #include <fmt/core.h>
-    std::string message = fmt::format("The answer is {}", 42);
+    fmt::print("Elapsed time: {0:.2f} seconds", 1.23);
   \endrst
-*/
-// Pass char_t as a default template parameter instead of using
-// std::basic_string<char_t<S>> to reduce the symbol size.
-template <typename S, typename... Args, typename Char = char_t<S>>
-FMT_INLINE std::basic_string<Char> format(const S& format_str, Args&&... args) {
-  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
-  return detail::vformat(to_string_view(format_str), vargs);
+ */
+template <typename... T>
+FMT_INLINE void print(format_string<T...> fmt, T&&... args) {
+  const auto& vargs = fmt::make_format_args(args...);
+  return detail::is_utf8() ? vprint(fmt, vargs)
+                           : detail::vprint_mojibake(stdout, fmt, vargs);
 }
 
-FMT_API void vprint(string_view, format_args);
-FMT_API void vprint(std::FILE*, string_view, format_args);
-
 /**
   \rst
-  Formats ``args`` according to specifications in ``format_str`` and writes the
-  output to the file ``f``. Strings are assumed to be Unicode-encoded unless the
-  ``FMT_UNICODE`` macro is set to 0.
+  Formats ``args`` according to specifications in ``fmt`` and writes the
+  output to the file ``f``.
 
   **Example**::
 
     fmt::print(stderr, "Don't {}!", "panic");
   \endrst
  */
-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 = 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);
+template <typename... T>
+FMT_INLINE void print(std::FILE* f, format_string<T...> fmt, T&&... args) {
+  const auto& vargs = fmt::make_format_args(args...);
+  return detail::is_utf8() ? vprint(f, fmt, vargs)
+                           : detail::vprint_mojibake(f, fmt, vargs);
 }
 
-/**
-  \rst
-  Formats ``args`` according to specifications in ``format_str`` and writes
-  the output to ``stdout``. Strings are assumed to be Unicode-encoded unless
-  the ``FMT_UNICODE`` macro is set to 0.
-
-  **Example**::
-
-    fmt::print("Elapsed time: {0:.2f} seconds", 1.23);
-  \endrst
- */
-template <typename S, typename... Args, typename Char = char_t<S>>
-inline void print(const S& format_str, Args&&... args) {
-  const auto& vargs = fmt::make_args_checked<Args...>(format_str, args...);
-  return detail::is_unicode<Char>()
-             ? vprint(to_string_view(format_str), vargs)
-             : detail::vprint_mojibake(stdout, to_string_view(format_str),
-                                       vargs);
-}
+FMT_MODULE_EXPORT_END
+FMT_GCC_PRAGMA("GCC pop_options")
 FMT_END_NAMESPACE
 
+#ifdef FMT_HEADER_ONLY
+#  include "format.h"
+#endif
 #endif  // FMT_CORE_H_
diff --git a/src/third_party_lib/fmt/format-inl.h b/src/third_party_lib/fmt/format-inl.h
index 8f2fe735..2c51c50a 100644
--- a/src/third_party_lib/fmt/format-inl.h
+++ b/src/third_party_lib/fmt/format-inl.h
@@ -8,8 +8,9 @@
 #ifndef FMT_FORMAT_INL_H_
 #define FMT_FORMAT_INL_H_
 
-#include <cassert>
+#include <algorithm>
 #include <cctype>
+#include <cerrno>  // errno
 #include <climits>
 #include <cmath>
 #include <cstdarg>
@@ -27,11 +28,6 @@
 
 #include "format.h"
 
-// Dummy implementations of strerror_r and strerror_s called if corresponding
-// system functions are not available.
-inline fmt::detail::null<> strerror_r(int, char*, ...) { return {}; }
-inline fmt::detail::null<> strerror_s(char*, size_t, ...) { return {}; }
-
 FMT_BEGIN_NAMESPACE
 namespace detail {
 
@@ -44,6 +40,10 @@ FMT_FUNC void assert_fail(const char* file, int line, const char* message) {
   std::terminate();
 }
 
+FMT_FUNC void throw_format_error(const char* message) {
+  FMT_THROW(format_error(message));
+}
+
 #ifndef _MSC_VER
 #  define FMT_SNPRINTF snprintf
 #else  // _MSC_VER
@@ -57,76 +57,6 @@ inline int fmt_snprintf(char* buffer, size_t size, const char* format, ...) {
 #  define FMT_SNPRINTF fmt_snprintf
 #endif  // _MSC_VER
 
-// A portable thread-safe version of strerror.
-// Sets buffer to point to a string describing the error code.
-// This can be either a pointer to a string stored in buffer,
-// or a pointer to some static immutable string.
-// Returns one of the following values:
-//   0      - success
-//   ERANGE - buffer is not large enough to store the error message
-//   other  - failure
-// Buffer should be at least of size 1.
-inline int safe_strerror(int error_code, char*& buffer,
-                         size_t buffer_size) FMT_NOEXCEPT {
-  FMT_ASSERT(buffer != nullptr && buffer_size != 0, "invalid buffer");
-
-  class dispatcher {
-   private:
-    int error_code_;
-    char*& buffer_;
-    size_t buffer_size_;
-
-    // A noop assignment operator to avoid bogus warnings.
-    void operator=(const dispatcher&) {}
-
-    // Handle the result of XSI-compliant version of strerror_r.
-    int handle(int result) {
-      // glibc versions before 2.13 return result in errno.
-      return result == -1 ? errno : result;
-    }
-
-    // Handle the result of GNU-specific version of strerror_r.
-    FMT_MAYBE_UNUSED
-    int handle(char* message) {
-      // If the buffer is full then the message is probably truncated.
-      if (message == buffer_ && strlen(buffer_) == buffer_size_ - 1)
-        return ERANGE;
-      buffer_ = message;
-      return 0;
-    }
-
-    // Handle the case when strerror_r is not available.
-    FMT_MAYBE_UNUSED
-    int handle(detail::null<>) {
-      return fallback(strerror_s(buffer_, buffer_size_, error_code_));
-    }
-
-    // Fallback to strerror_s when strerror_r is not available.
-    FMT_MAYBE_UNUSED
-    int fallback(int result) {
-      // If the buffer is full then the message is probably truncated.
-      return result == 0 && strlen(buffer_) == buffer_size_ - 1 ? ERANGE
-                                                                : result;
-    }
-
-#if !FMT_MSC_VER
-    // Fallback to strerror if strerror_r and strerror_s are not available.
-    int fallback(detail::null<>) {
-      errno = 0;
-      buffer_ = strerror(error_code_);
-      return errno;
-    }
-#endif
-
-   public:
-    dispatcher(int err_code, char*& buf, size_t buf_size)
-        : error_code_(err_code), buffer_(buf), buffer_size_(buf_size) {}
-
-    int run() { return handle(strerror_r(error_code_, buffer_, buffer_size_)); }
-  };
-  return dispatcher(error_code, buffer, buffer_size).run();
-}
-
 FMT_FUNC void format_error_code(detail::buffer<char>& out, int error_code,
                                 string_view message) FMT_NOEXCEPT {
   // Report error code making sure that the output fits into
@@ -145,18 +75,18 @@ FMT_FUNC void format_error_code(detail::buffer<char>& out, int error_code,
   error_code_size += detail::to_unsigned(detail::count_digits(abs_value));
   auto it = buffer_appender<char>(out);
   if (message.size() <= inline_buffer_size - error_code_size)
-    format_to(it, "{}{}", message, SEP);
-  format_to(it, "{}{}", ERROR_STR, error_code);
-  assert(out.size() <= inline_buffer_size);
+    format_to(it, FMT_STRING("{}{}"), message, SEP);
+  format_to(it, FMT_STRING("{}{}"), ERROR_STR, error_code);
+  FMT_ASSERT(out.size() <= inline_buffer_size, "");
 }
 
 FMT_FUNC void report_error(format_func func, int error_code,
-                           string_view message) FMT_NOEXCEPT {
+                           const char* message) FMT_NOEXCEPT {
   memory_buffer full_message;
   func(full_message, error_code, message);
   // Don't use fwrite_fully because the latter may throw.
-  (void)std::fwrite(full_message.data(), full_message.size(), 1, stderr);
-  std::fputc('\n', stderr);
+  if (std::fwrite(full_message.data(), full_message.size(), 1, stderr) > 0)
+    std::fputc('\n', stderr);
 }
 
 // A wrapper around fwrite that throws on error.
@@ -165,11 +95,8 @@ inline void fwrite_fully(const void* ptr, size_t size, size_t count,
   size_t written = std::fwrite(ptr, size, count, stream);
   if (written < count) FMT_THROW(system_error(errno, "cannot write to file"));
 }
-}  // namespace detail
-
-#if !defined(FMT_STATIC_THOUSANDS_SEPARATOR)
-namespace detail {
 
+#ifndef FMT_STATIC_THOUSANDS_SEPARATOR
 template <typename Locale>
 locale_ref::locale_ref(const Locale& loc) : locale_(&loc) {
   static_assert(std::is_same<Locale, std::locale>::value, "");
@@ -180,41 +107,36 @@ template <typename Locale> Locale locale_ref::get() const {
   return locale_ ? *static_cast<const std::locale*>(locale_) : std::locale();
 }
 
-template <typename Char> FMT_FUNC std::string grouping_impl(locale_ref loc) {
-  return std::use_facet<std::numpunct<Char>>(loc.get<std::locale>()).grouping();
-}
-template <typename Char> FMT_FUNC Char thousands_sep_impl(locale_ref loc) {
-  return std::use_facet<std::numpunct<Char>>(loc.get<std::locale>())
-      .thousands_sep();
+template <typename Char>
+FMT_FUNC auto thousands_sep_impl(locale_ref loc) -> thousands_sep_result<Char> {
+  auto& facet = std::use_facet<std::numpunct<Char>>(loc.get<std::locale>());
+  auto grouping = facet.grouping();
+  auto thousands_sep = grouping.empty() ? Char() : facet.thousands_sep();
+  return {std::move(grouping), thousands_sep};
 }
 template <typename Char> FMT_FUNC Char decimal_point_impl(locale_ref loc) {
   return std::use_facet<std::numpunct<Char>>(loc.get<std::locale>())
       .decimal_point();
 }
-}  // namespace detail
 #else
 template <typename Char>
-FMT_FUNC std::string detail::grouping_impl(locale_ref) {
-  return "\03";
-}
-template <typename Char> FMT_FUNC Char detail::thousands_sep_impl(locale_ref) {
-  return FMT_STATIC_THOUSANDS_SEPARATOR;
+FMT_FUNC auto thousands_sep_impl(locale_ref) -> thousands_sep_result<Char> {
+  return {"\03", FMT_STATIC_THOUSANDS_SEPARATOR};
 }
-template <typename Char> FMT_FUNC Char detail::decimal_point_impl(locale_ref) {
+template <typename Char> FMT_FUNC Char decimal_point_impl(locale_ref) {
   return '.';
 }
 #endif
+}  // namespace detail
 
+#if !FMT_MSC_VER
 FMT_API FMT_FUNC format_error::~format_error() FMT_NOEXCEPT = default;
-FMT_API FMT_FUNC system_error::~system_error() FMT_NOEXCEPT = default;
+#endif
 
-FMT_FUNC void system_error::init(int err_code, string_view format_str,
-                                 format_args args) {
-  error_code_ = err_code;
-  memory_buffer buffer;
-  format_system_error(buffer, err_code, vformat(format_str, args));
-  std::runtime_error& base = *this;
-  base = std::runtime_error(to_string(buffer));
+FMT_FUNC std::system_error vsystem_error(int error_code, string_view format_str,
+                                         format_args args) {
+  auto ec = std::error_code(error_code, std::generic_category());
+  return std::system_error(ec, vformat(format_str, args));
 }
 
 namespace detail {
@@ -227,946 +149,153 @@ template <> FMT_FUNC int count_digits<4>(detail::fallback_uintptr n) {
   return i >= 0 ? i * char_digits + count_digits<4, unsigned>(n.value[i]) : 1;
 }
 
-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'}};
-
-template <typename T>
-const char basic_data<T>::hex_digits[] = "0123456789abcdef";
-
-#define FMT_POWERS_OF_10(factor)                                             \
-  factor * 10, (factor)*100, (factor)*1000, (factor)*10000, (factor)*100000, \
-      (factor)*1000000, (factor)*10000000, (factor)*100000000,               \
-      (factor)*1000000000
-
-template <typename T>
-const uint64_t basic_data<T>::powers_of_10_64[] = {
-    1, FMT_POWERS_OF_10(1), FMT_POWERS_OF_10(1000000000ULL),
-    10000000000000000000ULL};
-
-template <typename T>
-const uint32_t basic_data<T>::zero_or_powers_of_10_32[] = {0,
-                                                           FMT_POWERS_OF_10(1)};
-template <typename T>
-const uint64_t basic_data<T>::zero_or_powers_of_10_64[] = {
-    0, FMT_POWERS_OF_10(1), FMT_POWERS_OF_10(1000000000ULL),
-    10000000000000000000ULL};
-
-template <typename T>
-const uint32_t basic_data<T>::zero_or_powers_of_10_32_new[] = {
-    0, 0, FMT_POWERS_OF_10(1)};
-
-template <typename T>
-const uint64_t basic_data<T>::zero_or_powers_of_10_64_new[] = {
-    0, 0, FMT_POWERS_OF_10(1), FMT_POWERS_OF_10(1000000000ULL),
-    10000000000000000000ULL};
-
-// Normalized 64-bit significands of pow(10, k), for k = -348, -340, ..., 340.
-// These are generated by support/compute-powers.py.
-template <typename T>
-const uint64_t basic_data<T>::grisu_pow10_significands[] = {
-    0xfa8fd5a0081c0288, 0xbaaee17fa23ebf76, 0x8b16fb203055ac76,
-    0xcf42894a5dce35ea, 0x9a6bb0aa55653b2d, 0xe61acf033d1a45df,
-    0xab70fe17c79ac6ca, 0xff77b1fcbebcdc4f, 0xbe5691ef416bd60c,
-    0x8dd01fad907ffc3c, 0xd3515c2831559a83, 0x9d71ac8fada6c9b5,
-    0xea9c227723ee8bcb, 0xaecc49914078536d, 0x823c12795db6ce57,
-    0xc21094364dfb5637, 0x9096ea6f3848984f, 0xd77485cb25823ac7,
-    0xa086cfcd97bf97f4, 0xef340a98172aace5, 0xb23867fb2a35b28e,
-    0x84c8d4dfd2c63f3b, 0xc5dd44271ad3cdba, 0x936b9fcebb25c996,
-    0xdbac6c247d62a584, 0xa3ab66580d5fdaf6, 0xf3e2f893dec3f126,
-    0xb5b5ada8aaff80b8, 0x87625f056c7c4a8b, 0xc9bcff6034c13053,
-    0x964e858c91ba2655, 0xdff9772470297ebd, 0xa6dfbd9fb8e5b88f,
-    0xf8a95fcf88747d94, 0xb94470938fa89bcf, 0x8a08f0f8bf0f156b,
-    0xcdb02555653131b6, 0x993fe2c6d07b7fac, 0xe45c10c42a2b3b06,
-    0xaa242499697392d3, 0xfd87b5f28300ca0e, 0xbce5086492111aeb,
-    0x8cbccc096f5088cc, 0xd1b71758e219652c, 0x9c40000000000000,
-    0xe8d4a51000000000, 0xad78ebc5ac620000, 0x813f3978f8940984,
-    0xc097ce7bc90715b3, 0x8f7e32ce7bea5c70, 0xd5d238a4abe98068,
-    0x9f4f2726179a2245, 0xed63a231d4c4fb27, 0xb0de65388cc8ada8,
-    0x83c7088e1aab65db, 0xc45d1df942711d9a, 0x924d692ca61be758,
-    0xda01ee641a708dea, 0xa26da3999aef774a, 0xf209787bb47d6b85,
-    0xb454e4a179dd1877, 0x865b86925b9bc5c2, 0xc83553c5c8965d3d,
-    0x952ab45cfa97a0b3, 0xde469fbd99a05fe3, 0xa59bc234db398c25,
-    0xf6c69a72a3989f5c, 0xb7dcbf5354e9bece, 0x88fcf317f22241e2,
-    0xcc20ce9bd35c78a5, 0x98165af37b2153df, 0xe2a0b5dc971f303a,
-    0xa8d9d1535ce3b396, 0xfb9b7cd9a4a7443c, 0xbb764c4ca7a44410,
-    0x8bab8eefb6409c1a, 0xd01fef10a657842c, 0x9b10a4e5e9913129,
-    0xe7109bfba19c0c9d, 0xac2820d9623bf429, 0x80444b5e7aa7cf85,
-    0xbf21e44003acdd2d, 0x8e679c2f5e44ff8f, 0xd433179d9c8cb841,
-    0x9e19db92b4e31ba9, 0xeb96bf6ebadf77d9, 0xaf87023b9bf0ee6b,
-};
-
-// Binary exponents of pow(10, k), for k = -348, -340, ..., 340, corresponding
-// to significands above.
-template <typename T>
-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,
-    -343,  -316,  -289,  -263,  -236,  -210,  -183,  -157,  -130,  -103, -77,
-    -50,   -24,   3,     30,    56,    83,    109,   136,   162,   189,  216,
-    242,   269,   295,   322,   348,   375,   402,   428,   455,   481,  508,
-    534,   561,   588,   614,   641,   667,   694,   720,   747,   774,  800,
-    827,   853,   880,   907,   933,   960,   986,   1013,  1039,  1066};
-
-template <typename T>
-const divtest_table_entry<uint32_t> basic_data<T>::divtest_table_for_pow5_32[] =
-    {{0x00000001, 0xffffffff}, {0xcccccccd, 0x33333333},
-     {0xc28f5c29, 0x0a3d70a3}, {0x26e978d5, 0x020c49ba},
-     {0x3afb7e91, 0x0068db8b}, {0x0bcbe61d, 0x0014f8b5},
-     {0x68c26139, 0x000431bd}, {0xae8d46a5, 0x0000d6bf},
-     {0x22e90e21, 0x00002af3}, {0x3a2e9c6d, 0x00000897},
-     {0x3ed61f49, 0x000001b7}};
-
-template <typename T>
-const divtest_table_entry<uint64_t> basic_data<T>::divtest_table_for_pow5_64[] =
-    {{0x0000000000000001, 0xffffffffffffffff},
-     {0xcccccccccccccccd, 0x3333333333333333},
-     {0x8f5c28f5c28f5c29, 0x0a3d70a3d70a3d70},
-     {0x1cac083126e978d5, 0x020c49ba5e353f7c},
-     {0xd288ce703afb7e91, 0x0068db8bac710cb2},
-     {0x5d4e8fb00bcbe61d, 0x0014f8b588e368f0},
-     {0x790fb65668c26139, 0x000431bde82d7b63},
-     {0xe5032477ae8d46a5, 0x0000d6bf94d5e57a},
-     {0xc767074b22e90e21, 0x00002af31dc46118},
-     {0x8e47ce423a2e9c6d, 0x0000089705f4136b},
-     {0x4fa7f60d3ed61f49, 0x000001b7cdfd9d7b},
-     {0x0fee64690c913975, 0x00000057f5ff85e5},
-     {0x3662e0e1cf503eb1, 0x000000119799812d},
-     {0xa47a2cf9f6433fbd, 0x0000000384b84d09},
-     {0x54186f653140a659, 0x00000000b424dc35},
-     {0x7738164770402145, 0x0000000024075f3d},
-     {0xe4a4d1417cd9a041, 0x000000000734aca5},
-     {0xc75429d9e5c5200d, 0x000000000170ef54},
-     {0xc1773b91fac10669, 0x000000000049c977},
-     {0x26b172506559ce15, 0x00000000000ec1e4},
-     {0xd489e3a9addec2d1, 0x000000000002f394},
-     {0x90e860bb892c8d5d, 0x000000000000971d},
-     {0x502e79bf1b6f4f79, 0x0000000000001e39},
-     {0xdcd618596be30fe5, 0x000000000000060b}};
-
-template <typename T>
-const uint64_t basic_data<T>::dragonbox_pow10_significands_64[] = {
-    0x81ceb32c4b43fcf5, 0xa2425ff75e14fc32, 0xcad2f7f5359a3b3f,
-    0xfd87b5f28300ca0e, 0x9e74d1b791e07e49, 0xc612062576589ddb,
-    0xf79687aed3eec552, 0x9abe14cd44753b53, 0xc16d9a0095928a28,
-    0xf1c90080baf72cb2, 0x971da05074da7bef, 0xbce5086492111aeb,
-    0xec1e4a7db69561a6, 0x9392ee8e921d5d08, 0xb877aa3236a4b44a,
-    0xe69594bec44de15c, 0x901d7cf73ab0acda, 0xb424dc35095cd810,
-    0xe12e13424bb40e14, 0x8cbccc096f5088cc, 0xafebff0bcb24aaff,
-    0xdbe6fecebdedd5bf, 0x89705f4136b4a598, 0xabcc77118461cefd,
-    0xd6bf94d5e57a42bd, 0x8637bd05af6c69b6, 0xa7c5ac471b478424,
-    0xd1b71758e219652c, 0x83126e978d4fdf3c, 0xa3d70a3d70a3d70b,
-    0xcccccccccccccccd, 0x8000000000000000, 0xa000000000000000,
-    0xc800000000000000, 0xfa00000000000000, 0x9c40000000000000,
-    0xc350000000000000, 0xf424000000000000, 0x9896800000000000,
-    0xbebc200000000000, 0xee6b280000000000, 0x9502f90000000000,
-    0xba43b74000000000, 0xe8d4a51000000000, 0x9184e72a00000000,
-    0xb5e620f480000000, 0xe35fa931a0000000, 0x8e1bc9bf04000000,
-    0xb1a2bc2ec5000000, 0xde0b6b3a76400000, 0x8ac7230489e80000,
-    0xad78ebc5ac620000, 0xd8d726b7177a8000, 0x878678326eac9000,
-    0xa968163f0a57b400, 0xd3c21bcecceda100, 0x84595161401484a0,
-    0xa56fa5b99019a5c8, 0xcecb8f27f4200f3a, 0x813f3978f8940984,
-    0xa18f07d736b90be5, 0xc9f2c9cd04674ede, 0xfc6f7c4045812296,
-    0x9dc5ada82b70b59d, 0xc5371912364ce305, 0xf684df56c3e01bc6,
-    0x9a130b963a6c115c, 0xc097ce7bc90715b3, 0xf0bdc21abb48db20,
-    0x96769950b50d88f4, 0xbc143fa4e250eb31, 0xeb194f8e1ae525fd,
-    0x92efd1b8d0cf37be, 0xb7abc627050305ad, 0xe596b7b0c643c719,
-    0x8f7e32ce7bea5c6f, 0xb35dbf821ae4f38b, 0xe0352f62a19e306e};
+// log10(2) = 0x0.4d104d427de7fbcc...
+static constexpr uint64_t log10_2_significand = 0x4d104d427de7fbcc;
+
+template <typename T = void> struct basic_impl_data {
+  // Normalized 64-bit significands of pow(10, k), for k = -348, -340, ..., 340.
+  // These are generated by support/compute-powers.py.
+  static constexpr uint64_t pow10_significands[87] = {
+      0xfa8fd5a0081c0288, 0xbaaee17fa23ebf76, 0x8b16fb203055ac76,
+      0xcf42894a5dce35ea, 0x9a6bb0aa55653b2d, 0xe61acf033d1a45df,
+      0xab70fe17c79ac6ca, 0xff77b1fcbebcdc4f, 0xbe5691ef416bd60c,
+      0x8dd01fad907ffc3c, 0xd3515c2831559a83, 0x9d71ac8fada6c9b5,
+      0xea9c227723ee8bcb, 0xaecc49914078536d, 0x823c12795db6ce57,
+      0xc21094364dfb5637, 0x9096ea6f3848984f, 0xd77485cb25823ac7,
+      0xa086cfcd97bf97f4, 0xef340a98172aace5, 0xb23867fb2a35b28e,
+      0x84c8d4dfd2c63f3b, 0xc5dd44271ad3cdba, 0x936b9fcebb25c996,
+      0xdbac6c247d62a584, 0xa3ab66580d5fdaf6, 0xf3e2f893dec3f126,
+      0xb5b5ada8aaff80b8, 0x87625f056c7c4a8b, 0xc9bcff6034c13053,
+      0x964e858c91ba2655, 0xdff9772470297ebd, 0xa6dfbd9fb8e5b88f,
+      0xf8a95fcf88747d94, 0xb94470938fa89bcf, 0x8a08f0f8bf0f156b,
+      0xcdb02555653131b6, 0x993fe2c6d07b7fac, 0xe45c10c42a2b3b06,
+      0xaa242499697392d3, 0xfd87b5f28300ca0e, 0xbce5086492111aeb,
+      0x8cbccc096f5088cc, 0xd1b71758e219652c, 0x9c40000000000000,
+      0xe8d4a51000000000, 0xad78ebc5ac620000, 0x813f3978f8940984,
+      0xc097ce7bc90715b3, 0x8f7e32ce7bea5c70, 0xd5d238a4abe98068,
+      0x9f4f2726179a2245, 0xed63a231d4c4fb27, 0xb0de65388cc8ada8,
+      0x83c7088e1aab65db, 0xc45d1df942711d9a, 0x924d692ca61be758,
+      0xda01ee641a708dea, 0xa26da3999aef774a, 0xf209787bb47d6b85,
+      0xb454e4a179dd1877, 0x865b86925b9bc5c2, 0xc83553c5c8965d3d,
+      0x952ab45cfa97a0b3, 0xde469fbd99a05fe3, 0xa59bc234db398c25,
+      0xf6c69a72a3989f5c, 0xb7dcbf5354e9bece, 0x88fcf317f22241e2,
+      0xcc20ce9bd35c78a5, 0x98165af37b2153df, 0xe2a0b5dc971f303a,
+      0xa8d9d1535ce3b396, 0xfb9b7cd9a4a7443c, 0xbb764c4ca7a44410,
+      0x8bab8eefb6409c1a, 0xd01fef10a657842c, 0x9b10a4e5e9913129,
+      0xe7109bfba19c0c9d, 0xac2820d9623bf429, 0x80444b5e7aa7cf85,
+      0xbf21e44003acdd2d, 0x8e679c2f5e44ff8f, 0xd433179d9c8cb841,
+      0x9e19db92b4e31ba9, 0xeb96bf6ebadf77d9, 0xaf87023b9bf0ee6b,
+  };
 
-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}
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+#  pragma GCC diagnostic push
+#  pragma GCC diagnostic ignored "-Wnarrowing"
 #endif
+  // Binary exponents of pow(10, k), for k = -348, -340, ..., 340, corresponding
+  // to significands above.
+  static constexpr int16_t pow10_exponents[87] = {
+      -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,
+      -343,  -316,  -289,  -263,  -236,  -210,  -183,  -157,  -130,  -103, -77,
+      -50,   -24,   3,     30,    56,    83,    109,   136,   162,   189,  216,
+      242,   269,   295,   322,   348,   375,   402,   428,   455,   481,  508,
+      534,   561,   588,   614,   641,   667,   694,   720,   747,   774,  800,
+      827,   853,   880,   907,   933,   960,   986,   1013,  1039,  1066};
+#if FMT_GCC_VERSION && FMT_GCC_VERSION < 409
+#  pragma GCC diagnostic pop
+#endif
+
+  static constexpr uint64_t power_of_10_64[20] = {
+      1, FMT_POWERS_OF_10(1ULL), FMT_POWERS_OF_10(1000000000ULL),
+      10000000000000000000ULL};
 };
 
-#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};
+// This is a struct rather than an alias to avoid shadowing warnings in gcc.
+struct impl_data : basic_impl_data<> {};
 
+#if __cplusplus < 201703L
 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};
+constexpr uint64_t basic_impl_data<T>::pow10_significands[];
+template <typename T> constexpr int16_t basic_impl_data<T>::pow10_exponents[];
+template <typename T> constexpr uint64_t basic_impl_data<T>::power_of_10_64[];
 #endif
 
-template <typename T>
-const char basic_data<T>::foreground_color[] = "\x1b[38;2;";
-template <typename T>
-const char basic_data<T>::background_color[] = "\x1b[48;2;";
-template <typename T> const char basic_data<T>::reset_color[] = "\x1b[0m";
-template <typename T> const wchar_t basic_data<T>::wreset_color[] = L"\x1b[0m";
-template <typename T> const char basic_data<T>::signs[] = {0, '-', '+', ' '};
-template <typename T>
-const char basic_data<T>::left_padding_shifts[] = {31, 31, 0, 1, 0};
-template <typename T>
-const char basic_data<T>::right_padding_shifts[] = {0, 31, 0, 1, 0};
-
 template <typename T> struct bits {
   static FMT_CONSTEXPR_DECL const int value =
       static_cast<int>(sizeof(T) * std::numeric_limits<unsigned char>::digits);
 };
 
-class fp;
-template <int SHIFT = 0> fp normalize(fp value);
-
-// Lower (upper) boundary is a value half way between a floating-point value
-// and its predecessor (successor). Boundaries have the same exponent as the
-// value so only significands are stored.
-struct boundaries {
-  uint64_t lower;
-  uint64_t upper;
-};
-
-// A handmade floating-point number f * pow(2, e).
-class fp {
- private:
-  using significand_type = uint64_t;
-
-  template <typename Float>
-  using is_supported_float = bool_constant<sizeof(Float) == sizeof(uint64_t) ||
-                                           sizeof(Float) == sizeof(uint32_t)>;
+// Returns the number of significand bits in Float excluding the implicit bit.
+template <typename Float> constexpr int num_significand_bits() {
+  // Subtract 1 to account for an implicit most significant bit in the
+  // normalized form.
+  return std::numeric_limits<Float>::digits - 1;
+}
 
- public:
-  significand_type f;
+// A floating-point number f * pow(2, e).
+struct fp {
+  uint64_t f;
   int e;
 
-  // All sizes are in bits.
-  // Subtract 1 to account for an implicit most significant bit in the
-  // normalized form.
-  static FMT_CONSTEXPR_DECL const int double_significand_size =
-      std::numeric_limits<double>::digits - 1;
-  static FMT_CONSTEXPR_DECL const uint64_t implicit_bit =
-      1ULL << double_significand_size;
-  static FMT_CONSTEXPR_DECL const int significand_size =
-      bits<significand_type>::value;
+  static constexpr const int num_significand_bits = bits<decltype(f)>::value;
 
-  fp() : f(0), e(0) {}
-  fp(uint64_t f_val, int e_val) : f(f_val), e(e_val) {}
+  constexpr fp() : f(0), e(0) {}
+  constexpr fp(uint64_t f_val, int e_val) : f(f_val), e(e_val) {}
 
-  // Constructs fp from an IEEE754 double. It is a template to prevent compile
-  // errors on platforms where double is not IEEE754.
-  template <typename Double> explicit fp(Double d) { assign(d); }
+  // Constructs fp from an IEEE754 floating-point number. It is a template to
+  // prevent compile errors on systems where n is not IEEE754.
+  template <typename Float> explicit FMT_CONSTEXPR fp(Float n) { assign(n); }
+
+  template <typename Float>
+  using is_supported = bool_constant<sizeof(Float) == sizeof(uint64_t) ||
+                                     sizeof(Float) == sizeof(uint32_t)>;
 
   // Assigns d to this and return true iff predecessor is closer than successor.
-  template <typename Float, FMT_ENABLE_IF(is_supported_float<Float>::value)>
-  bool assign(Float d) {
+  template <typename Float, FMT_ENABLE_IF(is_supported<Float>::value)>
+  FMT_CONSTEXPR bool assign(Float n) {
     // 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<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;
+    const int num_float_significand_bits =
+        detail::num_significand_bits<Float>();
+    const uint64_t implicit_bit = 1ULL << num_float_significand_bits;
+    const uint64_t significand_mask = implicit_bit - 1;
     constexpr bool is_double = sizeof(Float) == sizeof(uint64_t);
-    auto u = bit_cast<conditional_t<is_double, uint64_t, uint32_t>>(d);
+    auto u = bit_cast<conditional_t<is_double, uint64_t, uint32_t>>(n);
     f = u & significand_mask;
+    const uint64_t exponent_mask = (~0ULL >> 1) & ~significand_mask;
     int biased_e =
-        static_cast<int>((u & exponent_mask) >> float_significand_size);
-    // Predecessor is closer if d is a normalized power of 2 (f == 0) other than
-    // the smallest normalized number (biased_e > 1).
+        static_cast<int>((u & exponent_mask) >> num_float_significand_bits);
+    // The predecessor is closer if n 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 += float_implicit_bit;
+      f += implicit_bit;
     else
       biased_e = 1;  // Subnormals use biased exponent 1 (min exponent).
-    e = biased_e - exponent_bias - float_significand_size;
+    const int exponent_bias = std::numeric_limits<Float>::max_exponent - 1;
+    e = biased_e - exponent_bias - num_float_significand_bits;
     return is_predecessor_closer;
   }
 
-  template <typename Float, FMT_ENABLE_IF(!is_supported_float<Float>::value)>
+  template <typename Float, FMT_ENABLE_IF(!is_supported<Float>::value)>
   bool assign(Float) {
-    *this = fp();
+    FMT_ASSERT(false, "");
     return false;
   }
 };
 
 // Normalizes the value converted from double and multiplied by (1 << SHIFT).
-template <int SHIFT> fp normalize(fp value) {
+template <int SHIFT = 0> FMT_CONSTEXPR fp normalize(fp value) {
   // Handle subnormals.
-  const auto shifted_implicit_bit = fp::implicit_bit << SHIFT;
+  const uint64_t implicit_bit = 1ULL << num_significand_bits<double>();
+  const auto shifted_implicit_bit = implicit_bit << SHIFT;
   while ((value.f & shifted_implicit_bit) == 0) {
     value.f <<= 1;
     --value.e;
   }
   // Subtract 1 to account for hidden bit.
   const auto offset =
-      fp::significand_size - fp::double_significand_size - SHIFT - 1;
+      fp::num_significand_bits - num_significand_bits<double>() - SHIFT - 1;
   value.f <<= offset;
   value.e -= offset;
   return value;
@@ -1175,7 +304,7 @@ template <int SHIFT> fp normalize(fp value) {
 inline bool operator==(fp x, fp y) { return x.f == y.f && x.e == y.e; }
 
 // Computes lhs * rhs / pow(2, 64) rounded to nearest with half-up tie breaking.
-inline uint64_t multiply(uint64_t lhs, uint64_t rhs) {
+FMT_CONSTEXPR inline uint64_t multiply(uint64_t lhs, uint64_t rhs) {
 #if FMT_USE_INT128
   auto product = static_cast<__uint128_t>(lhs) * rhs;
   auto f = static_cast<uint64_t>(product >> 64);
@@ -1192,15 +321,18 @@ inline uint64_t multiply(uint64_t lhs, uint64_t rhs) {
 #endif
 }
 
-inline fp operator*(fp x, fp y) { return {multiply(x.f, y.f), x.e + y.e + 64}; }
+FMT_CONSTEXPR 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) {
+FMT_CONSTEXPR inline fp get_cached_power(int min_exponent,
+                                         int& pow10_exponent) {
   const int shift = 32;
-  const auto significand = static_cast<int64_t>(data::log10_2_significand);
+  const auto significand = static_cast<int64_t>(log10_2_significand);
   int index = static_cast<int>(
-      ((min_exponent + fp::significand_size - 1) * (significand >> shift) +
+      ((min_exponent + fp::num_significand_bits - 1) * (significand >> shift) +
        ((int64_t(1) << shift) - 1))  // ceil
       >> 32                          // arithmetic shift
   );
@@ -1210,8 +342,8 @@ 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::grisu_pow10_significands[index],
-          data::grisu_pow10_exponents[index]};
+  return {impl_data::pow10_significands[index],
+          impl_data::pow10_exponents[index]};
 }
 
 // A simple accumulator to hold the sums of terms in bigint::square if uint128_t
@@ -1220,15 +352,17 @@ struct accumulator {
   uint64_t lower;
   uint64_t upper;
 
-  accumulator() : lower(0), upper(0) {}
-  explicit operator uint32_t() const { return static_cast<uint32_t>(lower); }
+  constexpr accumulator() : lower(0), upper(0) {}
+  constexpr explicit operator uint32_t() const {
+    return static_cast<uint32_t>(lower);
+  }
 
-  void operator+=(uint64_t n) {
+  FMT_CONSTEXPR void operator+=(uint64_t n) {
     lower += n;
     if (lower < n) ++upper;
   }
-  void operator>>=(int shift) {
-    assert(shift == 32);
+  FMT_CONSTEXPR void operator>>=(int shift) {
+    FMT_ASSERT(shift == 32, "");
     (void)shift;
     lower = (upper << 32) | (lower >> 32);
     upper >>= 32;
@@ -1245,27 +379,31 @@ class bigint {
   basic_memory_buffer<bigit, bigits_capacity> bigits_;
   int exp_;
 
-  bigit operator[](int index) const { return bigits_[to_unsigned(index)]; }
-  bigit& operator[](int index) { return bigits_[to_unsigned(index)]; }
+  FMT_CONSTEXPR20 bigit operator[](int index) const {
+    return bigits_[to_unsigned(index)];
+  }
+  FMT_CONSTEXPR20 bigit& operator[](int index) {
+    return bigits_[to_unsigned(index)];
+  }
 
   static FMT_CONSTEXPR_DECL const int bigit_bits = bits<bigit>::value;
 
   friend struct formatter<bigint>;
 
-  void subtract_bigits(int index, bigit other, bigit& borrow) {
+  FMT_CONSTEXPR20 void subtract_bigits(int index, bigit other, bigit& borrow) {
     auto result = static_cast<double_bigit>((*this)[index]) - other - borrow;
     (*this)[index] = static_cast<bigit>(result);
     borrow = static_cast<bigit>(result >> (bigit_bits * 2 - 1));
   }
 
-  void remove_leading_zeros() {
+  FMT_CONSTEXPR20 void remove_leading_zeros() {
     int num_bigits = static_cast<int>(bigits_.size()) - 1;
     while (num_bigits > 0 && (*this)[num_bigits] == 0) --num_bigits;
     bigits_.resize(to_unsigned(num_bigits + 1));
   }
 
   // Computes *this -= other assuming aligned bigints and *this >= other.
-  void subtract_aligned(const bigint& other) {
+  FMT_CONSTEXPR20 void subtract_aligned(const bigint& other) {
     FMT_ASSERT(other.exp_ >= exp_, "unaligned bigints");
     FMT_ASSERT(compare(*this, other) >= 0, "");
     bigit borrow = 0;
@@ -1276,7 +414,7 @@ class bigint {
     remove_leading_zeros();
   }
 
-  void multiply(uint32_t value) {
+  FMT_CONSTEXPR20 void multiply(uint32_t value) {
     const double_bigit wide_value = value;
     bigit carry = 0;
     for (size_t i = 0, n = bigits_.size(); i < n; ++i) {
@@ -1287,7 +425,7 @@ class bigint {
     if (carry != 0) bigits_.push_back(carry);
   }
 
-  void multiply(uint64_t value) {
+  FMT_CONSTEXPR20 void multiply(uint64_t value) {
     const bigit mask = ~bigit(0);
     const double_bigit lower = value & mask;
     const double_bigit upper = value >> bigit_bits;
@@ -1305,14 +443,16 @@ class bigint {
   }
 
  public:
-  bigint() : exp_(0) {}
+  FMT_CONSTEXPR20 bigint() : exp_(0) {}
   explicit bigint(uint64_t n) { assign(n); }
-  ~bigint() { assert(bigits_.capacity() <= bigits_capacity); }
+  FMT_CONSTEXPR20 ~bigint() {
+    FMT_ASSERT(bigits_.capacity() <= bigits_capacity, "");
+  }
 
   bigint(const bigint&) = delete;
   void operator=(const bigint&) = delete;
 
-  void assign(const bigint& other) {
+  FMT_CONSTEXPR20 void assign(const bigint& other) {
     auto size = other.bigits_.size();
     bigits_.resize(size);
     auto data = other.bigits_.data();
@@ -1320,7 +460,7 @@ class bigint {
     exp_ = other.exp_;
   }
 
-  void assign(uint64_t n) {
+  FMT_CONSTEXPR20 void assign(uint64_t n) {
     size_t num_bigits = 0;
     do {
       bigits_[num_bigits++] = n & ~bigit(0);
@@ -1330,10 +470,12 @@ class bigint {
     exp_ = 0;
   }
 
-  int num_bigits() const { return static_cast<int>(bigits_.size()) + exp_; }
+  FMT_CONSTEXPR20 int num_bigits() const {
+    return static_cast<int>(bigits_.size()) + exp_;
+  }
 
-  FMT_NOINLINE bigint& operator<<=(int shift) {
-    assert(shift >= 0);
+  FMT_NOINLINE FMT_CONSTEXPR20 bigint& operator<<=(int shift) {
+    FMT_ASSERT(shift >= 0, "");
     exp_ += shift / bigit_bits;
     shift %= bigit_bits;
     if (shift == 0) return *this;
@@ -1347,13 +489,13 @@ class bigint {
     return *this;
   }
 
-  template <typename Int> bigint& operator*=(Int value) {
+  template <typename Int> FMT_CONSTEXPR20 bigint& operator*=(Int value) {
     FMT_ASSERT(value > 0, "");
     multiply(uint32_or_64_or_128_t<Int>(value));
     return *this;
   }
 
-  friend int compare(const bigint& lhs, const bigint& rhs) {
+  friend FMT_CONSTEXPR20 int compare(const bigint& lhs, const bigint& rhs) {
     int num_lhs_bigits = lhs.num_bigits(), num_rhs_bigits = rhs.num_bigits();
     if (num_lhs_bigits != num_rhs_bigits)
       return num_lhs_bigits > num_rhs_bigits ? 1 : -1;
@@ -1370,8 +512,8 @@ class bigint {
   }
 
   // Returns compare(lhs1 + lhs2, rhs).
-  friend int add_compare(const bigint& lhs1, const bigint& lhs2,
-                         const bigint& rhs) {
+  friend FMT_CONSTEXPR20 int add_compare(const bigint& lhs1, const bigint& lhs2,
+                                         const bigint& rhs) {
     int max_lhs_bigits = (std::max)(lhs1.num_bigits(), lhs2.num_bigits());
     int num_rhs_bigits = rhs.num_bigits();
     if (max_lhs_bigits + 1 < num_rhs_bigits) return -1;
@@ -1394,8 +536,8 @@ class bigint {
   }
 
   // Assigns pow(10, exp) to this bigint.
-  void assign_pow10(int exp) {
-    assert(exp >= 0);
+  FMT_CONSTEXPR20 void assign_pow10(int exp) {
+    FMT_ASSERT(exp >= 0, "");
     if (exp == 0) return assign(1);
     // Find the top bit.
     int bitmask = 1;
@@ -1413,10 +555,10 @@ class bigint {
     *this <<= exp;  // Multiply by pow(2, exp) by shifting.
   }
 
-  void square() {
-    basic_memory_buffer<bigit, bigits_capacity> n(std::move(bigits_));
+  FMT_CONSTEXPR20 void square() {
     int num_bigits = static_cast<int>(bigits_.size());
     int num_result_bigits = 2 * num_bigits;
+    basic_memory_buffer<bigit, bigits_capacity> n(std::move(bigits_));
     bigits_.resize(to_unsigned(num_result_bigits));
     using accumulator_t = conditional_t<FMT_USE_INT128, uint128_t, accumulator>;
     auto sum = accumulator_t();
@@ -1438,14 +580,13 @@ class bigint {
       (*this)[bigit_index] = static_cast<bigit>(sum);
       sum >>= bits<bigit>::value;
     }
-    --num_result_bigits;
     remove_leading_zeros();
     exp_ *= 2;
   }
 
   // If this bigint has a bigger exponent than other, adds trailing zero to make
   // exponents equal. This simplifies some operations such as subtraction.
-  void align(const bigint& other) {
+  FMT_CONSTEXPR20 void align(const bigint& other) {
     int exp_difference = exp_ - other.exp_;
     if (exp_difference <= 0) return;
     int num_bigits = static_cast<int>(bigits_.size());
@@ -1458,7 +599,7 @@ class bigint {
 
   // Divides this bignum by divisor, assigning the remainder to this and
   // returning the quotient.
-  int divmod_assign(const bigint& divisor) {
+  FMT_CONSTEXPR20 int divmod_assign(const bigint& divisor) {
     FMT_ASSERT(this != &divisor, "");
     if (compare(*this, divisor) < 0) return 0;
     FMT_ASSERT(divisor.bigits_[divisor.bigits_.size() - 1u] != 0, "");
@@ -1478,8 +619,9 @@ enum class round_direction { unknown, up, down };
 // some number v and the error, returns whether v should be rounded up, down, or
 // whether the rounding direction can't be determined due to error.
 // error should be less than divisor / 2.
-inline round_direction get_round_direction(uint64_t divisor, uint64_t remainder,
-                                           uint64_t error) {
+FMT_CONSTEXPR inline round_direction get_round_direction(uint64_t divisor,
+                                                         uint64_t remainder,
+                                                         uint64_t error) {
   FMT_ASSERT(remainder < divisor, "");  // divisor - remainder won't overflow.
   FMT_ASSERT(error < divisor, "");      // divisor - error won't overflow.
   FMT_ASSERT(error < divisor - error, "");  // error * 2 won't overflow.
@@ -1502,12 +644,52 @@ enum result {
 };
 }
 
+struct gen_digits_handler {
+  char* buf;
+  int size;
+  int precision;
+  int exp10;
+  bool fixed;
+
+  FMT_CONSTEXPR digits::result on_digit(char digit, uint64_t divisor,
+                                        uint64_t remainder, uint64_t error,
+                                        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.
+      // The check is not needed for the integral part because error = 1
+      // and divisor > (1 << 32) there.
+      if (error >= divisor || error >= divisor - error) return digits::error;
+    } else {
+      FMT_ASSERT(error == 1 && divisor > 2, "");
+    }
+    auto dir = get_round_direction(divisor, remainder, error);
+    if (dir != round_direction::up)
+      return dir == round_direction::down ? digits::done : digits::error;
+    ++buf[size - 1];
+    for (int i = size - 1; i > 0 && buf[i] > '9'; --i) {
+      buf[i] = '0';
+      ++buf[i - 1];
+    }
+    if (buf[0] > '9') {
+      buf[0] = '1';
+      if (fixed)
+        buf[size++] = '0';
+      else
+        ++exp10;
+    }
+    return digits::done;
+  }
+};
+
 // 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).
-template <typename Handler>
-FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
-                                                  int& exp, Handler& handler) {
+FMT_INLINE FMT_CONSTEXPR20 digits::result grisu_gen_digits(
+    fp value, uint64_t error, int& exp, gen_digits_handler& handler) {
   const fp one(1ULL << -value.e, value.e);
   // The integral part of scaled value (p1 in Grisu) = value / one. It cannot be
   // zero because it contains a product of two 64-bit numbers with MSB set (due
@@ -1518,10 +700,28 @@ FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
   // The fractional part of scaled value (p2 in Grisu) c = value % one.
   uint64_t fractional = value.f & (one.f - 1);
   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);
-  if (result != digits::more) return result;
+  // Non-fixed formats require at least one digit and no precision adjustment.
+  if (handler.fixed) {
+    // Adjust fixed precision by exponent because it is relative to decimal
+    // point.
+    int precision_offset = exp + handler.exp10;
+    if (precision_offset > 0 &&
+        handler.precision > max_value<int>() - precision_offset) {
+      FMT_THROW(format_error("number is too big"));
+    }
+    handler.precision += precision_offset;
+    // Check if precision is satisfied just by leading zeros, e.g.
+    // format("{:.2f}", 0.001) gives "0.00" without generating any digits.
+    if (handler.precision <= 0) {
+      if (handler.precision < 0) return digits::done;
+      // Divide by 10 to prevent overflow.
+      uint64_t divisor = impl_data::power_of_10_64[exp - 1] << -one.e;
+      auto dir = get_round_direction(divisor, value.f / 10, error * 10);
+      if (dir == round_direction::unknown) return digits::error;
+      handler.buf[handler.size++] = dir == round_direction::up ? '1' : '0';
+      return digits::done;
+    }
+  }
   // Generate digits for the integral part. This can produce up to 10 digits.
   do {
     uint32_t digit = 0;
@@ -1568,9 +768,9 @@ FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
     }
     --exp;
     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);
+    auto result = handler.on_digit(static_cast<char>('0' + digit),
+                                   impl_data::power_of_10_64[exp] << -one.e,
+                                   remainder, error, true);
     if (result != digits::more) return result;
   } while (exp > 0);
   // Generate digits for the fractional part.
@@ -1580,74 +780,63 @@ FMT_ALWAYS_INLINE digits::result grisu_gen_digits(fp value, uint64_t error,
     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);
+    auto result = handler.on_digit(digit, one.f, fractional, error, false);
     if (result != digits::more) return result;
   }
 }
 
-// The fixed precision digit handler.
-struct fixed_handler {
-  char* buf;
-  int size;
-  int precision;
-  int exp10;
-  bool fixed;
+// A 128-bit integer type used internally,
+struct uint128_wrapper {
+  uint128_wrapper() = default;
 
-  digits::result on_start(uint64_t divisor, uint64_t remainder, uint64_t error,
-                          int& exp) {
-    // Non-fixed formats require at least one digit and no precision adjustment.
-    if (!fixed) return digits::more;
-    // Adjust fixed precision by exponent because it is relative to decimal
-    // point.
-    precision += exp + exp10;
-    // Check if precision is satisfied just by leading zeros, e.g.
-    // format("{:.2f}", 0.001) gives "0.00" without generating any digits.
-    if (precision > 0) return digits::more;
-    if (precision < 0) return digits::done;
-    auto dir = get_round_direction(divisor, remainder, error);
-    if (dir == round_direction::unknown) return digits::error;
-    buf[size++] = dir == round_direction::up ? '1' : '0';
-    return digits::done;
+#if FMT_USE_INT128
+  uint128_t internal_;
+
+  constexpr uint128_wrapper(uint64_t high, uint64_t low) FMT_NOEXCEPT
+      : internal_{static_cast<uint128_t>(low) |
+                  (static_cast<uint128_t>(high) << 64)} {}
+
+  constexpr uint128_wrapper(uint128_t u) : internal_{u} {}
+
+  constexpr uint64_t high() const FMT_NOEXCEPT {
+    return uint64_t(internal_ >> 64);
   }
+  constexpr uint64_t low() const FMT_NOEXCEPT { return uint64_t(internal_); }
 
-  digits::result on_digit(char digit, uint64_t divisor, uint64_t remainder,
-                          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.
-      // The check is not needed for the integral part because error = 1
-      // and divisor > (1 << 32) there.
-      if (error >= divisor || error >= divisor - error) return digits::error;
-    } else {
-      FMT_ASSERT(error == 1 && divisor > 2, "");
-    }
-    auto dir = get_round_direction(divisor, remainder, error);
-    if (dir != round_direction::up)
-      return dir == round_direction::down ? digits::done : digits::error;
-    ++buf[size - 1];
-    for (int i = size - 1; i > 0 && buf[i] > '9'; --i) {
-      buf[i] = '0';
-      ++buf[i - 1];
-    }
-    if (buf[0] > '9') {
-      buf[0] = '1';
-      if (fixed)
-        buf[size++] = '0';
-      else
-        ++exp10;
-    }
-    return digits::done;
+  uint128_wrapper& operator+=(uint64_t n) FMT_NOEXCEPT {
+    internal_ += n;
+    return *this;
   }
+#else
+  uint64_t high_;
+  uint64_t low_;
+
+  constexpr uint128_wrapper(uint64_t high, uint64_t low) FMT_NOEXCEPT
+      : high_{high},
+        low_{low} {}
+
+  constexpr uint64_t high() const FMT_NOEXCEPT { return high_; }
+  constexpr 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
 };
 
 // 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 {
+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)
@@ -1675,8 +864,7 @@ FMT_SAFEBUFFERS inline uint128_wrapper umul128(uint64_t x,
 }
 
 // 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 {
+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);
@@ -1689,8 +877,7 @@ FMT_SAFEBUFFERS inline uint64_t umul128_upper64(uint64_t x,
 
 // 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 {
+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();
@@ -1704,8 +891,7 @@ inline uint32_t umul96_upper32(uint32_t x, uint64_t y) FMT_NOEXCEPT {
 
 // 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 {
+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;
@@ -1722,8 +908,7 @@ inline uint64_t umul96_lower64(uint32_t x, uint64_t y) FMT_NOEXCEPT {
 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;
+  return (e * static_cast<int>(log10_2_significand >> (64 - shift))) >> shift;
 }
 
 // Various fast log computations.
@@ -1741,8 +926,7 @@ 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)) -
+  return (e * static_cast<int>(log10_2_significand >> (64 - shift_amount)) -
           static_cast<int>(log10_4_over_3_fractional_digits >>
                            (64 - shift_amount))) >>
          shift_amount;
@@ -1768,16 +952,52 @@ inline bool divisible_by_power_of_2(uint64_t x, int exp) FMT_NOEXCEPT {
 #endif
 }
 
+// Table entry type for divisibility test.
+template <typename T> struct divtest_table_entry {
+  T mod_inv;
+  T max_quotient;
+};
+
 // 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;
+  static constexpr const divtest_table_entry<uint32_t> divtest_table[] = {
+      {0x00000001, 0xffffffff}, {0xcccccccd, 0x33333333},
+      {0xc28f5c29, 0x0a3d70a3}, {0x26e978d5, 0x020c49ba},
+      {0x3afb7e91, 0x0068db8b}, {0x0bcbe61d, 0x0014f8b5},
+      {0x68c26139, 0x000431bd}, {0xae8d46a5, 0x0000d6bf},
+      {0x22e90e21, 0x00002af3}, {0x3a2e9c6d, 0x00000897},
+      {0x3ed61f49, 0x000001b7}};
+  return x * divtest_table[exp].mod_inv <= divtest_table[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;
+  static constexpr const divtest_table_entry<uint64_t> divtest_table[] = {
+      {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}};
+  return x * divtest_table[exp].mod_inv <= divtest_table[exp].max_quotient;
 }
 
 // Replaces n by floor(n / pow(5, N)) returning true if and only if n is
@@ -1831,7 +1051,34 @@ template <> struct cache_accessor<float> {
   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 constexpr const uint64_t pow10_significands[] = {
+        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};
+    return pow10_significands[k - float_info<float>::min_k];
   }
 
   static carrier_uint compute_mul(carrier_uint u,
@@ -1885,10 +1132,679 @@ template <> struct cache_accessor<double> {
     FMT_ASSERT(k >= float_info<double>::min_k && k <= float_info<double>::max_k,
                "k is out of range");
 
+    static constexpr const uint128_wrapper pow10_significands[] = {
+#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
-    return data::dragonbox_pow10_significands_128[k -
-                                                  float_info<double>::min_k];
+    return pow10_significands[k - float_info<double>::min_k];
 #else
+    static constexpr const uint64_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};
+
+    static constexpr const uint32_t 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};
+
     static const int compression_ratio = 27;
 
     // Compute base index.
@@ -1897,8 +1813,7 @@ template <> struct cache_accessor<double> {
     int offset = k - kb;
 
     // Get base cache.
-    uint128_wrapper base_cache =
-        data::dragonbox_pow10_significands_128[cache_index];
+    uint128_wrapper base_cache = pow10_significands[cache_index];
     if (offset == 0) return base_cache;
 
     // Compute the required amount of bit-shift.
@@ -1906,7 +1821,7 @@ template <> struct cache_accessor<double> {
     FMT_ASSERT(alpha > 0 && alpha < 64, "shifting error detected");
 
     // Try to recover the real cache.
-    uint64_t pow5 = data::powers_of_5_64[offset];
+    uint64_t pow5 = powers_of_5_64[offset];
     uint128_wrapper recovered_cache = umul128(base_cache.high(), pow5);
     uint128_wrapper middle_low =
         umul128(base_cache.low() - (kb < 0 ? 1u : 0u), pow5);
@@ -1924,7 +1839,7 @@ template <> struct cache_accessor<double> {
 
     // Get error.
     int error_idx = (k - float_info<double>::min_k) / 16;
-    uint32_t error = (data::dragonbox_pow10_recovery_errors[error_idx] >>
+    uint32_t error = (pow10_recovery_errors[error_idx] >>
                       ((k - float_info<double>::min_k) % 16) * 2) &
                      0x3;
 
@@ -2010,7 +1925,7 @@ bool is_center_integer(typename float_info<T>::carrier_uint two_f, int exponent,
 }
 
 // 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 {
+FMT_INLINE int remove_trailing_zeros(uint32_t& n) FMT_NOEXCEPT {
 #ifdef FMT_BUILTIN_CTZ
   int t = FMT_BUILTIN_CTZ(n);
 #else
@@ -2038,7 +1953,7 @@ FMT_ALWAYS_INLINE int remove_trailing_zeros(uint32_t& n) FMT_NOEXCEPT {
 }
 
 // Removes trailing zeros and returns the number of zeros removed (double)
-FMT_ALWAYS_INLINE int remove_trailing_zeros(uint64_t& n) FMT_NOEXCEPT {
+FMT_INLINE int remove_trailing_zeros(uint64_t& n) FMT_NOEXCEPT {
 #ifdef FMT_BUILTIN_CTZLL
   int t = FMT_BUILTIN_CTZLL(n);
 #else
@@ -2124,8 +2039,7 @@ FMT_ALWAYS_INLINE int remove_trailing_zeros(uint64_t& n) FMT_NOEXCEPT {
 
 // 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 {
+FMT_INLINE 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);
@@ -2171,8 +2085,7 @@ FMT_ALWAYS_INLINE FMT_SAFEBUFFERS decimal_fp<T> shorter_interval_case(
   return ret_value;
 }
 
-template <typename T>
-FMT_SAFEBUFFERS decimal_fp<T> to_decimal(T x) FMT_NOEXCEPT {
+template <typename T> 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;
@@ -2306,24 +2219,21 @@ FMT_SAFEBUFFERS decimal_fp<T> to_decimal(T x) FMT_NOEXCEPT {
 }
 }  // namespace dragonbox
 
-// Formats value using a variation of the Fixed-Precision Positive
-// Floating-Point Printout ((FPP)^2) algorithm by Steele & White:
-// https://fmt.dev/p372-steele.pdf.
-template <typename Double>
-void fallback_format(Double d, int num_digits, bool binary32, buffer<char>& buf,
-                     int& exp10) {
+// Formats a floating-point number using a variation of the Fixed-Precision
+// Positive Floating-Point Printout ((FPP)^2) algorithm by Steele & White:
+// https://fmt.dev/papers/p372-steele.pdf.
+FMT_CONSTEXPR20 inline void format_dragon(fp value, bool is_predecessor_closer,
+                                          int num_digits, buffer<char>& buf,
+                                          int& exp10) {
   bigint numerator;    // 2 * R in (FPP)^2.
   bigint denominator;  // 2 * S in (FPP)^2.
   // lower and upper are differences between value and corresponding boundaries.
   bigint lower;             // (M^- in (FPP)^2).
   bigint upper_store;       // upper's value if different from lower.
   bigint* upper = nullptr;  // (M^+ in (FPP)^2).
-  fp value;
   // Shift numerator and denominator by an extra bit or two (if lower boundary
   // is closer) to make lower and upper integers. This eliminates multiplication
   // by 2 during later computations.
-  const bool is_predecessor_closer =
-      binary32 ? value.assign(static_cast<float>(d)) : value.assign(d);
   int shift = is_predecessor_closer ? 2 : 1;
   uint64_t significand = value.f << shift;
   if (value.e >= 0) {
@@ -2393,9 +2303,9 @@ void fallback_format(Double d, int num_digits, bool binary32, buffer<char>& buf,
   // 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';
+    auto digit = add_compare(numerator, numerator, denominator) > 0 ? '1' : '0';
+    buf.push_back(digit);
     return;
   }
   buf.try_resize(to_unsigned(num_digits));
@@ -2426,9 +2336,12 @@ void fallback_format(Double d, int num_digits, bool binary32, buffer<char>& buf,
   buf[num_digits - 1] = static_cast<char>('0' + digit);
 }
 
-template <typename T>
-int format_float(T value, int precision, float_specs specs, buffer<char>& buf) {
-  static_assert(!std::is_same<T, float>::value, "");
+template <typename Float>
+FMT_HEADER_ONLY_CONSTEXPR20 int format_float(Float value, int precision,
+                                             float_specs specs,
+                                             buffer<char>& buf) {
+  // float is passed as double to reduce the number of instantiations.
+  static_assert(!std::is_same<Float, float>::value, "");
   FMT_ASSERT(value >= 0, "value is negative");
 
   const bool fixed = specs.format == float_format::fixed;
@@ -2438,13 +2351,13 @@ int format_float(T value, int precision, float_specs specs, buffer<char>& buf) {
       return 0;
     }
     buf.try_resize(to_unsigned(precision));
-    std::uninitialized_fill_n(buf.data(), precision, '0');
+    fill_n(buf.data(), precision, '0');
     return -precision;
   }
 
-  if (!specs.use_grisu) return snprintf_float(value, precision, specs, buf);
+  if (specs.fallback) return snprintf_float(value, precision, specs, buf);
 
-  if (precision < 0) {
+  if (!is_constant_evaluated() && precision < 0) {
     // Use Dragonbox for the shortest format.
     if (specs.binary32) {
       auto dec = dragonbox::to_decimal(static_cast<float>(value));
@@ -2456,26 +2369,37 @@ int format_float(T value, int precision, float_specs specs, buffer<char>& buf) {
     return dec.exponent;
   }
 
-  // Use Grisu + Dragon4 for the given precision:
-  // https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf.
   int exp = 0;
-  const int min_exp = -60;  // alpha in Grisu.
-  int cached_exp10 = 0;     // K in Grisu.
-  fp normalized = normalize(fp(value));
-  const auto cached_pow = get_cached_power(
-      min_exp - (normalized.e + fp::significand_size), cached_exp10);
-  normalized = normalized * cached_pow;
-  // Limit precision to the maximum possible number of significant digits in an
-  // IEEE754 double because we don't need to generate zeros.
-  const int max_double_digits = 767;
-  if (precision > max_double_digits) precision = max_double_digits;
-  fixed_handler handler{buf.data(), 0, precision, -cached_exp10, fixed};
-  if (grisu_gen_digits(normalized, 1, exp, handler) == digits::error) {
-    exp += handler.size - cached_exp10 - 1;
-    fallback_format(value, handler.precision, specs.binary32, buf, exp);
-  } else {
-    exp += handler.exp10;
-    buf.try_resize(to_unsigned(handler.size));
+  bool use_dragon = true;
+  if (is_fast_float<Float>()) {
+    // Use Grisu + Dragon4 for the given precision:
+    // https://www.cs.tufts.edu/~nr/cs257/archive/florian-loitsch/printf.pdf.
+    const int min_exp = -60;  // alpha in Grisu.
+    int cached_exp10 = 0;     // K in Grisu.
+    fp normalized = normalize(fp(value));
+    const auto cached_pow = get_cached_power(
+        min_exp - (normalized.e + fp::num_significand_bits), cached_exp10);
+    normalized = normalized * cached_pow;
+    gen_digits_handler handler{buf.data(), 0, precision, -cached_exp10, fixed};
+    if (grisu_gen_digits(normalized, 1, exp, handler) != digits::error &&
+        !is_constant_evaluated()) {
+      exp += handler.exp10;
+      buf.try_resize(to_unsigned(handler.size));
+      use_dragon = false;
+    } else {
+      exp += handler.size - cached_exp10 - 1;
+      precision = handler.precision;
+    }
+  }
+  if (use_dragon) {
+    auto f = fp();
+    bool is_predecessor_closer =
+        specs.binary32 ? f.assign(static_cast<float>(value)) : f.assign(value);
+    // 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;
+    format_dragon(f, is_predecessor_closer, precision, buf, exp);
   }
   if (!fixed && !specs.showpoint) {
     // Remove trailing zeros.
@@ -2487,7 +2411,7 @@ int format_float(T value, int precision, float_specs specs, buffer<char>& buf) {
     buf.try_resize(num_digits);
   }
   return exp;
-}  // namespace detail
+}
 
 template <typename T>
 int snprintf_float(T value, int precision, float_specs specs,
@@ -2571,11 +2495,11 @@ int snprintf_float(T value, int precision, float_specs specs,
       --exp_pos;
     } while (*exp_pos != 'e');
     char sign = exp_pos[1];
-    assert(sign == '+' || sign == '-');
+    FMT_ASSERT(sign == '+' || sign == '-', "");
     int exp = 0;
     auto p = exp_pos + 2;  // Skip 'e' and sign.
     do {
-      assert(is_digit(*p));
+      FMT_ASSERT(is_digit(*p), "");
       exp = exp * 10 + (*p++ - '0');
     } while (p != end);
     if (sign == '-') exp = -exp;
@@ -2592,71 +2516,11 @@ int snprintf_float(T value, int precision, float_specs specs,
     return exp - fraction_size;
   }
 }
-
-// A public domain branchless UTF-8 decoder by Christopher Wellons:
-// https://github.com/skeeto/branchless-utf8
-/* Decode the next character, c, from buf, reporting errors in e.
- *
- * Since this is a branchless decoder, four bytes will be read from the
- * buffer regardless of the actual length of the next character. This
- * means the buffer _must_ have at least three bytes of zero padding
- * following the end of the data stream.
- *
- * Errors are reported in e, which will be non-zero if the parsed
- * character was somehow invalid: invalid byte sequence, non-canonical
- * encoding, or a surrogate half.
- *
- * The function returns a pointer to the next character. When an error
- * occurs, this pointer will be a guess that depends on the particular
- * error, but it will always advance at least one byte.
- */
-inline const char* utf8_decode(const char* buf, uint32_t* c, int* e) {
-  static const int masks[] = {0x00, 0x7f, 0x1f, 0x0f, 0x07};
-  static const uint32_t mins[] = {4194304, 0, 128, 2048, 65536};
-  static const int shiftc[] = {0, 18, 12, 6, 0};
-  static const int shifte[] = {0, 6, 4, 2, 0};
-
-  int len = code_point_length(buf);
-  const char* next = buf + len;
-
-  // Assume a four-byte character and load four bytes. Unused bits are
-  // shifted out.
-  auto s = reinterpret_cast<const unsigned char*>(buf);
-  *c = uint32_t(s[0] & masks[len]) << 18;
-  *c |= uint32_t(s[1] & 0x3f) << 12;
-  *c |= uint32_t(s[2] & 0x3f) << 6;
-  *c |= uint32_t(s[3] & 0x3f) << 0;
-  *c >>= shiftc[len];
-
-  // Accumulate the various error conditions.
-  *e = (*c < mins[len]) << 6;       // non-canonical encoding
-  *e |= ((*c >> 11) == 0x1b) << 7;  // surrogate half?
-  *e |= (*c > 0x10FFFF) << 8;       // out of range?
-  *e |= (s[1] & 0xc0) >> 2;
-  *e |= (s[2] & 0xc0) >> 4;
-  *e |= (s[3]) >> 6;
-  *e ^= 0x2a;  // top two bits of each tail byte correct?
-  *e >>= shifte[len];
-
-  return next;
-}
-
-struct stringifier {
-  template <typename T> FMT_INLINE std::string operator()(T value) const {
-    return to_string(value);
-  }
-  std::string operator()(basic_format_arg<format_context>::handle h) const {
-    memory_buffer buf;
-    format_parse_context parse_ctx({});
-    format_context format_ctx(buffer_appender<char>(buf), {}, {});
-    h.format(parse_ctx, format_ctx);
-    return to_string(buf);
-  }
-};
 }  // namespace detail
 
 template <> struct formatter<detail::bigint> {
-  format_parse_context::iterator parse(format_parse_context& ctx) {
+  FMT_CONSTEXPR format_parse_context::iterator parse(
+      format_parse_context& ctx) {
     return ctx.begin();
   }
 
@@ -2667,24 +2531,22 @@ template <> struct formatter<detail::bigint> {
     for (auto i = n.bigits_.size(); i > 0; --i) {
       auto value = n.bigits_[i - 1u];
       if (first) {
-        out = format_to(out, "{:x}", value);
+        out = format_to(out, FMT_STRING("{:x}"), value);
         first = false;
         continue;
       }
-      out = format_to(out, "{:08x}", value);
+      out = format_to(out, FMT_STRING("{:08x}"), value);
     }
     if (n.exp_ > 0)
-      out = format_to(out, "p{}", n.exp_ * detail::bigint::bigit_bits);
+      out = format_to(out, FMT_STRING("p{}"),
+                      n.exp_ * detail::bigint::bigit_bits);
     return out;
   }
 };
 
 FMT_FUNC detail::utf8_to_utf16::utf8_to_utf16(string_view s) {
-  auto transcode = [this](const char* p) {
-    auto cp = uint32_t();
-    auto error = 0;
-    p = utf8_decode(p, &cp, &error);
-    if (error != 0) FMT_THROW(std::runtime_error("invalid utf8"));
+  for_each_codepoint(s, [this](uint32_t cp, string_view) {
+    if (cp == invalid_code_point) FMT_THROW(std::runtime_error("invalid utf8"));
     if (cp <= 0xFFFF) {
       buffer_.push_back(static_cast<wchar_t>(cp));
     } else {
@@ -2692,64 +2554,38 @@ FMT_FUNC detail::utf8_to_utf16::utf8_to_utf16(string_view s) {
       buffer_.push_back(static_cast<wchar_t>(0xD800 + (cp >> 10)));
       buffer_.push_back(static_cast<wchar_t>(0xDC00 + (cp & 0x3FF)));
     }
-    return p;
-  };
-  auto p = s.data();
-  const size_t block_size = 4;  // utf8_decode always reads blocks of 4 chars.
-  if (s.size() >= block_size) {
-    for (auto end = p + s.size() - block_size + 1; p < end;) p = transcode(p);
-  }
-  if (auto num_chars_left = s.data() + s.size() - p) {
-    char buf[2 * block_size - 1] = {};
-    memcpy(buf, p, to_unsigned(num_chars_left));
-    p = buf;
-    do {
-      p = transcode(p);
-    } while (p - buf < num_chars_left);
-  }
+    return true;
+  });
   buffer_.push_back(0);
 }
 
 FMT_FUNC void format_system_error(detail::buffer<char>& out, int error_code,
-                                  string_view message) FMT_NOEXCEPT {
+                                  const char* message) FMT_NOEXCEPT {
   FMT_TRY {
-    memory_buffer buf;
-    buf.resize(inline_buffer_size);
-    for (;;) {
-      char* system_message = &buf[0];
-      int result =
-          detail::safe_strerror(error_code, system_message, buf.size());
-      if (result == 0) {
-        format_to(detail::buffer_appender<char>(out), "{}: {}", message,
-                  system_message);
-        return;
-      }
-      if (result != ERANGE)
-        break;  // Can't get error message, report error code instead.
-      buf.resize(buf.size() * 2);
-    }
+    auto ec = std::error_code(error_code, std::generic_category());
+    write(std::back_inserter(out), std::system_error(ec, message).what());
+    return;
   }
   FMT_CATCH(...) {}
   format_error_code(out, error_code, message);
 }
 
-FMT_FUNC void detail::error_handler::on_error(const char* message) {
-  FMT_THROW(format_error(message));
-}
-
 FMT_FUNC void report_system_error(int error_code,
-                                  fmt::string_view message) FMT_NOEXCEPT {
+                                  const char* message) FMT_NOEXCEPT {
   report_error(format_system_error, error_code, message);
 }
 
-FMT_FUNC std::string detail::vformat(string_view format_str, format_args args) {
-  if (format_str.size() == 2 && equal2(format_str.data(), "{}")) {
-    auto arg = args.get(0);
-    if (!arg) error_handler().on_error("argument not found");
-    return visit_format_arg(stringifier(), arg);
-  }
-  memory_buffer buffer;
-  detail::vformat_to(buffer, format_str, args);
+// DEPRECATED!
+// This function is defined here and not inline for ABI compatiblity.
+FMT_FUNC void detail::error_handler::on_error(const char* message) {
+  throw_format_error(message);
+}
+
+FMT_FUNC std::string vformat(string_view fmt, format_args args) {
+  // Don't optimize the "{}" case to keep the binary size small and because it
+  // can be better optimized in fmt::format anyway.
+  auto buffer = memory_buffer();
+  detail::vformat_to(buffer, fmt, args);
   return to_string(buffer);
 }
 
@@ -2761,24 +2597,30 @@ extern "C" __declspec(dllimport) int __stdcall WriteConsoleW(  //
 }  // 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,
-                     basic_format_args<buffer_context<char>>(args));
+namespace detail {
+FMT_FUNC void print(std::FILE* f, string_view text) {
 #ifdef _WIN32
   auto fd = _fileno(f);
   if (_isatty(fd)) {
-    detail::utf8_to_utf16 u16(string_view(buffer.data(), buffer.size()));
+    detail::utf8_to_utf16 u16(string_view(text.data(), text.size()));
     auto written = detail::dword();
-    if (!detail::WriteConsoleW(reinterpret_cast<void*>(_get_osfhandle(fd)),
-                               u16.c_str(), static_cast<uint32_t>(u16.size()),
-                               &written, nullptr)) {
-      FMT_THROW(format_error("failed to write to console"));
+    if (detail::WriteConsoleW(reinterpret_cast<void*>(_get_osfhandle(fd)),
+                              u16.c_str(), static_cast<uint32_t>(u16.size()),
+                              &written, nullptr)) {
+      return;
     }
-    return;
+    // Fallback to fwrite on failure. It can happen if the output has been
+    // redirected to NUL.
   }
 #endif
-  detail::fwrite_fully(buffer.data(), 1, buffer.size(), f);
+  detail::fwrite_fully(text.data(), 1, text.size(), f);
+}
+}  // namespace detail
+
+FMT_FUNC void vprint(std::FILE* f, string_view format_str, format_args args) {
+  memory_buffer buffer;
+  detail::vformat_to(buffer, format_str, args);
+  detail::print(f, {buffer.data(), buffer.size()});
 }
 
 #ifdef _WIN32
diff --git a/src/third_party_lib/fmt/format.h b/src/third_party_lib/fmt/format.h
index 1a037b02..ee69651c 100644
--- a/src/third_party_lib/fmt/format.h
+++ b/src/third_party_lib/fmt/format.h
@@ -33,22 +33,24 @@
 #ifndef FMT_FORMAT_H_
 #define FMT_FORMAT_H_
 
-#include <algorithm>
-#include <cerrno>
-#include <cmath>
-#include <cstdint>
-#include <limits>
-#include <memory>
-#include <stdexcept>
+#include <cmath>         // std::signbit
+#include <cstdint>       // uint32_t
+#include <limits>        // std::numeric_limits
+#include <memory>        // std::uninitialized_copy
+#include <stdexcept>     // std::runtime_error
+#include <system_error>  // std::system_error
+#include <utility>       // std::swap
+
+#ifdef __cpp_lib_bit_cast
+#  include <bit>  // std::bitcast
+#endif
 
 #include "core.h"
 
-#ifdef __INTEL_COMPILER
-#  define FMT_ICC_VERSION __INTEL_COMPILER
-#elif defined(__ICL)
-#  define FMT_ICC_VERSION __ICL
+#if FMT_GCC_VERSION
+#  define FMT_GCC_VISIBILITY_HIDDEN __attribute__((visibility("hidden")))
 #else
-#  define FMT_ICC_VERSION 0
+#  define FMT_GCC_VISIBILITY_HIDDEN
 #endif
 
 #ifdef __NVCC__
@@ -69,30 +71,10 @@
 #  define FMT_NOINLINE
 #endif
 
-#if __cplusplus == 201103L || __cplusplus == 201402L
-#  if defined(__INTEL_COMPILER) || defined(__PGI)
-#    define FMT_FALLTHROUGH
-#  elif defined(__clang__)
-#    define FMT_FALLTHROUGH [[clang::fallthrough]]
-#  elif FMT_GCC_VERSION >= 700 && \
-      (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 520)
-#    define FMT_FALLTHROUGH [[gnu::fallthrough]]
-#  else
-#    define FMT_FALLTHROUGH
-#  endif
-#elif FMT_HAS_CPP17_ATTRIBUTE(fallthrough) || \
-    (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
-#  define FMT_FALLTHROUGH [[fallthrough]]
+#if FMT_MSC_VER
+#  define FMT_MSC_DEFAULT = default
 #else
-#  define FMT_FALLTHROUGH
-#endif
-
-#ifndef FMT_MAYBE_UNUSED
-#  if FMT_HAS_CPP17_ATTRIBUTE(maybe_unused)
-#    define FMT_MAYBE_UNUSED [[maybe_unused]]
-#  else
-#    define FMT_MAYBE_UNUSED
-#  endif
+#  define FMT_MSC_DEFAULT
 #endif
 
 #ifndef FMT_THROW
@@ -113,10 +95,9 @@ FMT_END_NAMESPACE
 #      define FMT_THROW(x) throw x
 #    endif
 #  else
-#    define FMT_THROW(x)              \
-      do {                            \
-        static_cast<void>(sizeof(x)); \
-        FMT_ASSERT(false, "");        \
+#    define FMT_THROW(x)               \
+      do {                             \
+        FMT_ASSERT(false, (x).what()); \
       } while (false)
 #  endif
 #endif
@@ -129,6 +110,21 @@ FMT_END_NAMESPACE
 #  define FMT_CATCH(x) if (false)
 #endif
 
+#ifndef FMT_MAYBE_UNUSED
+#  if FMT_HAS_CPP17_ATTRIBUTE(maybe_unused)
+#    define FMT_MAYBE_UNUSED [[maybe_unused]]
+#  else
+#    define FMT_MAYBE_UNUSED
+#  endif
+#endif
+
+// Workaround broken [[deprecated]] in the Intel, PGI and NVCC compilers.
+#if FMT_ICC_VERSION || defined(__PGI) || FMT_NVCC
+#  define FMT_DEPRECATED_ALIAS
+#else
+#  define FMT_DEPRECATED_ALIAS FMT_DEPRECATED
+#endif
+
 #ifndef FMT_USE_USER_DEFINED_LITERALS
 // EDG based compilers (Intel, NVIDIA, Elbrus, etc), GCC and MSVC support UDLs.
 #  if (FMT_HAS_FEATURE(cxx_user_literals) || FMT_GCC_VERSION >= 407 || \
@@ -140,53 +136,34 @@ FMT_END_NAMESPACE
 #  endif
 #endif
 
-#ifndef FMT_USE_UDL_TEMPLATE
-// EDG frontend based compilers (icc, nvcc, PGI, etc) and GCC < 6.4 do not
-// properly support UDL templates and GCC >= 9 warns about them.
-#  if FMT_USE_USER_DEFINED_LITERALS &&                         \
-      (!defined(__EDG_VERSION__) || __EDG_VERSION__ >= 501) && \
-      ((FMT_GCC_VERSION >= 604 && __cplusplus >= 201402L) ||   \
-       FMT_CLANG_VERSION >= 304) &&                            \
-      !defined(__PGI) && !defined(__NVCC__)
-#    define FMT_USE_UDL_TEMPLATE 1
-#  else
-#    define FMT_USE_UDL_TEMPLATE 0
-#  endif
-#endif
-
-#ifndef FMT_USE_FLOAT
-#  define FMT_USE_FLOAT 1
-#endif
-
-#ifndef FMT_USE_DOUBLE
-#  define FMT_USE_DOUBLE 1
-#endif
-
-#ifndef FMT_USE_LONG_DOUBLE
-#  define FMT_USE_LONG_DOUBLE 1
-#endif
-
 // Defining FMT_REDUCE_INT_INSTANTIATIONS to 1, will reduce the number of
-// int_writer template instances to just one by only using the largest integer
-// type. This results in a reduction in binary size but will cause a decrease in
-// integer formatting performance.
+// integer formatter template instantiations 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
-#  define FMT_BUILTIN_CLZ(n) __builtin_clz(n)
-#endif
-#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)
+// https://github.com/fmtlib/fmt/issues/519.
+#if !FMT_MSC_VER
+#  if FMT_HAS_BUILTIN(__builtin_clz) || FMT_GCC_VERSION || FMT_ICC_VERSION
+#    define FMT_BUILTIN_CLZ(n) __builtin_clz(n)
+#  endif
+#  if FMT_HAS_BUILTIN(__builtin_clzll) || FMT_GCC_VERSION || FMT_ICC_VERSION
+#    define FMT_BUILTIN_CLZLL(n) __builtin_clzll(n)
+#  endif
 #endif
-#if (FMT_GCC_VERSION || FMT_HAS_BUILTIN(__builtin_ctzll))
-#  define FMT_BUILTIN_CTZLL(n) __builtin_ctzll(n)
+
+// __builtin_ctz is broken in Intel Compiler Classic on Windows:
+// https://github.com/fmtlib/fmt/issues/2510.
+#ifndef __ICL
+#  if FMT_HAS_BUILTIN(__builtin_ctz) || FMT_GCC_VERSION || FMT_ICC_VERSION
+#    define FMT_BUILTIN_CTZ(n) __builtin_ctz(n)
+#  endif
+#  if FMT_HAS_BUILTIN(__builtin_ctzll) || FMT_GCC_VERSION || FMT_ICC_VERSION
+#    define FMT_BUILTIN_CTZLL(n) __builtin_ctzll(n)
+#  endif
 #endif
 
 #if FMT_MSC_VER
@@ -196,33 +173,32 @@ FMT_END_NAMESPACE
 // 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(FMT_BUILTIN_CTZLL) && !defined(_MANAGED)
+#if FMT_MSC_VER && !defined(FMT_BUILTIN_CLZLL) && !defined(FMT_BUILTIN_CTZLL)
 FMT_BEGIN_NAMESPACE
 namespace detail {
 // Avoid Clang with Microsoft CodeGen's -Wunknown-pragmas warning.
-#  ifndef __clang__
+#  if !defined(__clang__)
 #    pragma intrinsic(_BitScanForward)
 #    pragma intrinsic(_BitScanReverse)
-#  endif
-#  if defined(_WIN64) && !defined(__clang__)
-#    pragma intrinsic(_BitScanForward64)
-#    pragma intrinsic(_BitScanReverse64)
+#    if defined(_WIN64)
+#      pragma intrinsic(_BitScanForward64)
+#      pragma intrinsic(_BitScanReverse64)
+#    endif
 #  endif
 
-inline int clz(uint32_t x) {
+inline auto clz(uint32_t x) -> int {
   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)
+  FMT_MSC_WARNING(suppress : 6102)
   return 31 ^ static_cast<int>(r);
 }
 #  define FMT_BUILTIN_CLZ(n) detail::clz(n)
 
-inline int clzll(uint64_t x) {
+inline auto clzll(uint64_t x) -> int {
   unsigned long r = 0;
 #  ifdef _WIN64
   _BitScanReverse64(&r, x);
@@ -233,24 +209,24 @@ inline int clzll(uint64_t x) {
   _BitScanReverse(&r, static_cast<uint32_t>(x));
 #  endif
   FMT_ASSERT(x != 0, "");
-  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+  FMT_MSC_WARNING(suppress : 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) {
+inline auto ctz(uint32_t x) -> int {
   unsigned long r = 0;
   _BitScanForward(&r, x);
   FMT_ASSERT(x != 0, "");
-  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+  FMT_MSC_WARNING(suppress : 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) {
+inline auto ctzll(uint64_t x) -> int {
   unsigned long r = 0;
   FMT_ASSERT(x != 0, "");
-  FMT_SUPPRESS_MSC_WARNING(6102)  // Suppress a bogus static analysis warning.
+  FMT_MSC_WARNING(suppress : 6102)  // Suppress a bogus static analysis warning.
 #  ifdef _WIN64
   _BitScanForward64(&r, x);
 #  else
@@ -267,31 +243,71 @@ inline int ctzll(uint64_t x) {
 FMT_END_NAMESPACE
 #endif
 
-// Enable the deprecated numeric alignment.
-#ifndef FMT_DEPRECATED_NUMERIC_ALIGN
-#  define FMT_DEPRECATED_NUMERIC_ALIGN 0
+#ifdef FMT_HEADER_ONLY
+#  define FMT_HEADER_ONLY_CONSTEXPR20 FMT_CONSTEXPR20
+#else
+#  define FMT_HEADER_ONLY_CONSTEXPR20
 #endif
 
 FMT_BEGIN_NAMESPACE
 namespace detail {
 
-// An equivalent of `*reinterpret_cast<Dest*>(&source)` that doesn't have
-// undefined behavior (e.g. due to type aliasing).
-// Example: uint64_t d = bit_cast<uint64_t>(2.718);
-template <typename Dest, typename Source>
-inline Dest bit_cast(const Source& source) {
-  static_assert(sizeof(Dest) == sizeof(Source), "size mismatch");
-  Dest dest;
-  std::memcpy(&dest, &source, sizeof(dest));
-  return dest;
+template <typename Streambuf> class formatbuf : public Streambuf {
+ private:
+  using char_type = typename Streambuf::char_type;
+  using streamsize = decltype(std::declval<Streambuf>().sputn(nullptr, 0));
+  using int_type = typename Streambuf::int_type;
+  using traits_type = typename Streambuf::traits_type;
+
+  buffer<char_type>& buffer_;
+
+ public:
+  explicit formatbuf(buffer<char_type>& buf) : buffer_(buf) {}
+
+ protected:
+  // The put area is always empty. This makes the implementation simpler and has
+  // the advantage that the streambuf and the buffer are always in sync and
+  // sputc never writes into uninitialized memory. A disadvantage is that each
+  // call to sputc always results in a (virtual) call to overflow. There is no
+  // disadvantage here for sputn since this always results in a call to xsputn.
+
+  auto overflow(int_type ch) -> int_type override {
+    if (!traits_type::eq_int_type(ch, traits_type::eof()))
+      buffer_.push_back(static_cast<char_type>(ch));
+    return ch;
+  }
+
+  auto xsputn(const char_type* s, streamsize count) -> streamsize override {
+    buffer_.append(s, s + count);
+    return count;
+  }
+};
+
+// Implementation of std::bit_cast for pre-C++20.
+template <typename To, typename From>
+FMT_CONSTEXPR20 auto bit_cast(const From& from) -> To {
+  static_assert(sizeof(To) == sizeof(From), "size mismatch");
+#ifdef __cpp_lib_bit_cast
+  if (is_constant_evaluated()) return std::bit_cast<To>(from);
+#endif
+  auto to = To();
+  std::memcpy(&to, &from, sizeof(to));
+  return to;
 }
 
-inline bool is_big_endian() {
-  const auto u = 1u;
+inline auto is_big_endian() -> bool {
+#ifdef _WIN32
+  return false;
+#elif defined(__BIG_ENDIAN__)
+  return true;
+#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__)
+  return __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__;
+#else
   struct bytes {
-    char data[sizeof(u)];
+    char data[sizeof(int)];
   };
-  return bit_cast<bytes>(u).data[0] == 0;
+  return bit_cast<bytes>(1).data[0] == 0;
+#endif
 }
 
 // A fallback implementation of uintptr_t for systems that lack it.
@@ -301,7 +317,7 @@ struct fallback_uintptr {
   fallback_uintptr() = default;
   explicit fallback_uintptr(const void* p) {
     *this = bit_cast<fallback_uintptr>(p);
-    if (is_big_endian()) {
+    if (const_check(is_big_endian())) {
       for (size_t i = 0, j = sizeof(void*) - 1; i < j; ++i, --j)
         std::swap(value[i], value[j]);
     }
@@ -309,26 +325,28 @@ struct fallback_uintptr {
 };
 #ifdef UINTPTR_MAX
 using uintptr_t = ::uintptr_t;
-inline uintptr_t to_uintptr(const void* p) { return bit_cast<uintptr_t>(p); }
+inline auto to_uintptr(const void* p) -> uintptr_t {
+  return bit_cast<uintptr_t>(p);
+}
 #else
 using uintptr_t = fallback_uintptr;
-inline fallback_uintptr to_uintptr(const void* p) {
+inline auto to_uintptr(const void* p) -> fallback_uintptr {
   return fallback_uintptr(p);
 }
 #endif
 
 // Returns the largest possible value for type T. Same as
 // std::numeric_limits<T>::max() but shorter and not affected by the max macro.
-template <typename T> constexpr T max_value() {
+template <typename T> constexpr auto max_value() -> T {
   return (std::numeric_limits<T>::max)();
 }
-template <typename T> constexpr int num_bits() {
+template <typename T> constexpr auto num_bits() -> int {
   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>() {
+template <> constexpr auto num_bits<int128_t>() -> int { return 128; }
+template <> constexpr auto num_bits<uint128_t>() -> int { return 128; }
+template <> constexpr auto num_bits<fallback_uintptr>() -> int {
   return static_cast<int>(sizeof(void*) *
                           std::numeric_limits<unsigned char>::digits);
 }
@@ -346,31 +364,38 @@ using iterator_t = decltype(std::begin(std::declval<T&>()));
 template <typename T> using sentinel_t = decltype(std::end(std::declval<T&>()));
 
 // A workaround for std::string not having mutable data() until C++17.
-template <typename Char> inline Char* get_data(std::basic_string<Char>& s) {
+template <typename Char>
+inline auto get_data(std::basic_string<Char>& s) -> Char* {
   return &s[0];
 }
 template <typename Container>
-inline typename Container::value_type* get_data(Container& c) {
+inline auto get_data(Container& c) -> typename Container::value_type* {
   return c.data();
 }
 
 #if defined(_SECURE_SCL) && _SECURE_SCL
 // Make a checked iterator to avoid MSVC warnings.
 template <typename T> using checked_ptr = stdext::checked_array_iterator<T*>;
-template <typename T> checked_ptr<T> make_checked(T* p, size_t size) {
+template <typename T>
+constexpr auto make_checked(T* p, size_t size) -> checked_ptr<T> {
   return {p, size};
 }
 #else
 template <typename T> using checked_ptr = T*;
-template <typename T> inline T* make_checked(T* p, size_t) { return p; }
+template <typename T> constexpr auto make_checked(T* p, size_t) -> T* {
+  return p;
+}
 #endif
 
+// Attempts to reserve space for n extra characters in the output range.
+// Returns a pointer to the reserved range or a reference to it.
 template <typename Container, FMT_ENABLE_IF(is_contiguous<Container>::value)>
-#if FMT_CLANG_VERSION
+#if FMT_CLANG_VERSION >= 307 && !FMT_ICC_VERSION
 __attribute__((no_sanitize("undefined")))
 #endif
-inline checked_ptr<typename Container::value_type>
-reserve(std::back_insert_iterator<Container> it, size_t n) {
+inline auto
+reserve(std::back_insert_iterator<Container> it, size_t n)
+    -> checked_ptr<typename Container::value_type> {
   Container& c = get_container(it);
   size_t size = c.size();
   c.resize(size + n);
@@ -378,21 +403,26 @@ reserve(std::back_insert_iterator<Container> it, size_t n) {
 }
 
 template <typename T>
-inline buffer_appender<T> reserve(buffer_appender<T> it, size_t n) {
+inline auto reserve(buffer_appender<T> it, size_t n) -> buffer_appender<T> {
   buffer<T>& buf = get_container(it);
   buf.try_reserve(buf.size() + n);
   return it;
 }
 
-template <typename Iterator> inline Iterator& reserve(Iterator& it, size_t) {
+template <typename Iterator>
+constexpr auto reserve(Iterator& it, size_t) -> Iterator& {
   return it;
 }
 
+template <typename OutputIt>
+using reserve_iterator =
+    remove_reference_t<decltype(reserve(std::declval<OutputIt&>(), 0))>;
+
 template <typename T, typename OutputIt>
-constexpr T* to_pointer(OutputIt, size_t) {
+constexpr auto to_pointer(OutputIt, size_t) -> T* {
   return nullptr;
 }
-template <typename T> T* to_pointer(buffer_appender<T> it, size_t n) {
+template <typename T> auto to_pointer(buffer_appender<T> it, size_t n) -> T* {
   buffer<T>& buf = get_container(it);
   auto size = buf.size();
   if (buf.capacity() < size + n) return nullptr;
@@ -401,195 +431,195 @@ template <typename T> T* to_pointer(buffer_appender<T> it, size_t n) {
 }
 
 template <typename Container, FMT_ENABLE_IF(is_contiguous<Container>::value)>
-inline std::back_insert_iterator<Container> base_iterator(
-    std::back_insert_iterator<Container>& it,
-    checked_ptr<typename Container::value_type>) {
+inline auto base_iterator(std::back_insert_iterator<Container>& it,
+                          checked_ptr<typename Container::value_type>)
+    -> std::back_insert_iterator<Container> {
   return it;
 }
 
 template <typename Iterator>
-inline Iterator base_iterator(Iterator, Iterator it) {
+constexpr auto base_iterator(Iterator, Iterator it) -> Iterator {
   return it;
 }
 
-// An output iterator that counts the number of objects written to it and
-// discards them.
-class counting_iterator {
- private:
-  size_t count_;
-
- public:
-  using iterator_category = std::output_iterator_tag;
-  using difference_type = std::ptrdiff_t;
-  using pointer = void;
-  using reference = void;
-  using _Unchecked_type = counting_iterator;  // Mark iterator as checked.
-
-  struct value_type {
-    template <typename T> void operator=(const T&) {}
-  };
-
-  counting_iterator() : count_(0) {}
-
-  size_t count() const { return count_; }
-
-  counting_iterator& operator++() {
-    ++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 {}; }
-};
-
-template <typename OutputIt> class truncating_iterator_base {
- protected:
-  OutputIt out_;
-  size_t limit_;
-  size_t count_;
-
-  truncating_iterator_base(OutputIt out, size_t limit)
-      : out_(out), limit_(limit), count_(0) {}
-
- public:
-  using iterator_category = std::output_iterator_tag;
-  using value_type = typename std::iterator_traits<OutputIt>::value_type;
-  using difference_type = void;
-  using pointer = void;
-  using reference = void;
-  using _Unchecked_type =
-      truncating_iterator_base;  // Mark iterator as checked.
-
-  OutputIt base() const { return out_; }
-  size_t count() const { return count_; }
-};
-
-// An output iterator that truncates the output and counts the number of objects
-// written to it.
-template <typename OutputIt,
-          typename Enable = typename std::is_void<
-              typename std::iterator_traits<OutputIt>::value_type>::type>
-class truncating_iterator;
-
-template <typename OutputIt>
-class truncating_iterator<OutputIt, std::false_type>
-    : public truncating_iterator_base<OutputIt> {
-  mutable typename truncating_iterator_base<OutputIt>::value_type blackhole_;
-
- public:
-  using value_type = typename truncating_iterator_base<OutputIt>::value_type;
-
-  truncating_iterator(OutputIt out, size_t limit)
-      : truncating_iterator_base<OutputIt>(out, limit) {}
-
-  truncating_iterator& operator++() {
-    if (this->count_++ < this->limit_) ++this->out_;
-    return *this;
+// <algorithm> is spectacularly slow to compile in C++20 so use a simple fill_n
+// instead (#1998).
+template <typename OutputIt, typename Size, typename T>
+FMT_CONSTEXPR auto fill_n(OutputIt out, Size count, const T& value)
+    -> OutputIt {
+  for (Size i = 0; i < count; ++i) *out++ = value;
+  return out;
+}
+template <typename T, typename Size>
+FMT_CONSTEXPR20 auto fill_n(T* out, Size count, char value) -> T* {
+  if (is_constant_evaluated()) {
+    return fill_n<T*, Size, T>(out, count, value);
   }
+  std::memset(out, value, to_unsigned(count));
+  return out + count;
+}
 
-  truncating_iterator operator++(int) {
-    auto it = *this;
-    ++*this;
-    return it;
-  }
+#ifdef __cpp_char8_t
+using char8_type = char8_t;
+#else
+enum char8_type : unsigned char {};
+#endif
 
-  value_type& operator*() const {
-    return this->count_ < this->limit_ ? *this->out_ : blackhole_;
+template <typename OutChar, typename InputIt, typename OutputIt>
+FMT_CONSTEXPR FMT_NOINLINE auto copy_str_noinline(InputIt begin, InputIt end,
+                                                  OutputIt out) -> OutputIt {
+  return copy_str<OutChar>(begin, end, out);
+}
+
+// A public domain branchless UTF-8 decoder by Christopher Wellons:
+// https://github.com/skeeto/branchless-utf8
+/* Decode the next character, c, from s, reporting errors in e.
+ *
+ * Since this is a branchless decoder, four bytes will be read from the
+ * buffer regardless of the actual length of the next character. This
+ * means the buffer _must_ have at least three bytes of zero padding
+ * following the end of the data stream.
+ *
+ * Errors are reported in e, which will be non-zero if the parsed
+ * character was somehow invalid: invalid byte sequence, non-canonical
+ * encoding, or a surrogate half.
+ *
+ * The function returns a pointer to the next character. When an error
+ * occurs, this pointer will be a guess that depends on the particular
+ * error, but it will always advance at least one byte.
+ */
+FMT_CONSTEXPR inline auto utf8_decode(const char* s, uint32_t* c, int* e)
+    -> const char* {
+  constexpr const int masks[] = {0x00, 0x7f, 0x1f, 0x0f, 0x07};
+  constexpr const uint32_t mins[] = {4194304, 0, 128, 2048, 65536};
+  constexpr const int shiftc[] = {0, 18, 12, 6, 0};
+  constexpr const int shifte[] = {0, 6, 4, 2, 0};
+
+  int len = code_point_length(s);
+  const char* next = s + len;
+
+  // Assume a four-byte character and load four bytes. Unused bits are
+  // shifted out.
+  *c = uint32_t(s[0] & masks[len]) << 18;
+  *c |= uint32_t(s[1] & 0x3f) << 12;
+  *c |= uint32_t(s[2] & 0x3f) << 6;
+  *c |= uint32_t(s[3] & 0x3f) << 0;
+  *c >>= shiftc[len];
+
+  // Accumulate the various error conditions.
+  using uchar = unsigned char;
+  *e = (*c < mins[len]) << 6;       // non-canonical encoding
+  *e |= ((*c >> 11) == 0x1b) << 7;  // surrogate half?
+  *e |= (*c > 0x10FFFF) << 8;       // out of range?
+  *e |= (uchar(s[1]) & 0xc0) >> 2;
+  *e |= (uchar(s[2]) & 0xc0) >> 4;
+  *e |= uchar(s[3]) >> 6;
+  *e ^= 0x2a;  // top two bits of each tail byte correct?
+  *e >>= shifte[len];
+
+  return next;
+}
+
+constexpr uint32_t invalid_code_point = ~uint32_t();
+
+// Invokes f(cp, sv) for every code point cp in s with sv being the string view
+// corresponding to the code point. cp is invalid_code_point on error.
+template <typename F>
+FMT_CONSTEXPR void for_each_codepoint(string_view s, F f) {
+  auto decode = [f](const char* buf_ptr, const char* ptr) {
+    auto cp = uint32_t();
+    auto error = 0;
+    auto end = utf8_decode(buf_ptr, &cp, &error);
+    bool result = f(error ? invalid_code_point : cp,
+                    string_view(ptr, to_unsigned(end - buf_ptr)));
+    return result ? end : nullptr;
+  };
+  auto p = s.data();
+  const size_t block_size = 4;  // utf8_decode always reads blocks of 4 chars.
+  if (s.size() >= block_size) {
+    for (auto end = p + s.size() - block_size + 1; p < end;) {
+      p = decode(p, p);
+      if (!p) return;
+    }
   }
-};
-
-template <typename OutputIt>
-class truncating_iterator<OutputIt, std::true_type>
-    : public truncating_iterator_base<OutputIt> {
- public:
-  truncating_iterator(OutputIt out, size_t limit)
-      : truncating_iterator_base<OutputIt>(out, limit) {}
-
-  template <typename T> truncating_iterator& operator=(T val) {
-    if (this->count_++ < this->limit_) *this->out_++ = val;
-    return *this;
+  if (auto num_chars_left = s.data() + s.size() - p) {
+    char buf[2 * block_size - 1] = {};
+    copy_str<char>(p, p + num_chars_left, buf);
+    const char* buf_ptr = buf;
+    do {
+      auto end = decode(buf_ptr, p);
+      if (!end) return;
+      p += end - buf_ptr;
+      buf_ptr = end;
+    } while (buf_ptr - buf < num_chars_left);
   }
-
-  truncating_iterator& operator++() { return *this; }
-  truncating_iterator& operator++(int) { return *this; }
-  truncating_iterator& operator*() { return *this; }
-};
+}
 
 template <typename Char>
-inline size_t count_code_points(basic_string_view<Char> s) {
+inline auto compute_width(basic_string_view<Char> s) -> size_t {
   return s.size();
 }
 
-// Counts the number of code points in a UTF-8 string.
-inline size_t count_code_points(basic_string_view<char> s) {
-  const char* data = s.data();
+// Computes approximate display width of a UTF-8 string.
+FMT_CONSTEXPR inline size_t compute_width(string_view s) {
   size_t num_code_points = 0;
-  for (size_t i = 0, size = s.size(); i != size; ++i) {
-    if ((data[i] & 0xc0) != 0x80) ++num_code_points;
-  }
+  // It is not a lambda for compatibility with C++14.
+  struct count_code_points {
+    size_t* count;
+    FMT_CONSTEXPR auto operator()(uint32_t cp, string_view) const -> bool {
+      *count += detail::to_unsigned(
+          1 +
+          (cp >= 0x1100 &&
+           (cp <= 0x115f ||  // Hangul Jamo init. consonants
+            cp == 0x2329 ||  // LEFT-POINTING ANGLE BRACKET
+            cp == 0x232a ||  // RIGHT-POINTING ANGLE BRACKET
+            // CJK ... Yi except IDEOGRAPHIC HALF FILL SPACE:
+            (cp >= 0x2e80 && cp <= 0xa4cf && cp != 0x303f) ||
+            (cp >= 0xac00 && cp <= 0xd7a3) ||    // Hangul Syllables
+            (cp >= 0xf900 && cp <= 0xfaff) ||    // CJK Compatibility Ideographs
+            (cp >= 0xfe10 && cp <= 0xfe19) ||    // Vertical Forms
+            (cp >= 0xfe30 && cp <= 0xfe6f) ||    // CJK Compatibility Forms
+            (cp >= 0xff00 && cp <= 0xff60) ||    // Fullwidth Forms
+            (cp >= 0xffe0 && cp <= 0xffe6) ||    // Fullwidth Forms
+            (cp >= 0x20000 && cp <= 0x2fffd) ||  // CJK
+            (cp >= 0x30000 && cp <= 0x3fffd) ||
+            // Miscellaneous Symbols and Pictographs + Emoticons:
+            (cp >= 0x1f300 && cp <= 0x1f64f) ||
+            // Supplemental Symbols and Pictographs:
+            (cp >= 0x1f900 && cp <= 0x1f9ff))));
+      return true;
+    }
+  };
+  for_each_codepoint(s, count_code_points{&num_code_points});
   return num_code_points;
 }
 
-inline size_t count_code_points(basic_string_view<char8_type> s) {
-  return count_code_points(basic_string_view<char>(
+inline auto compute_width(basic_string_view<char8_type> s) -> size_t {
+  return compute_width(basic_string_view<char>(
       reinterpret_cast<const char*>(s.data()), s.size()));
 }
 
 template <typename Char>
-inline size_t code_point_index(basic_string_view<Char> s, size_t n) {
+inline auto code_point_index(basic_string_view<Char> s, size_t n) -> size_t {
   size_t size = s.size();
   return n < size ? n : size;
 }
 
 // Calculates the index of the nth code point in a UTF-8 string.
-inline size_t code_point_index(basic_string_view<char8_type> s, size_t n) {
+inline auto code_point_index(basic_string_view<char8_type> s, size_t n)
+    -> size_t {
   const char8_type* data = s.data();
   size_t num_code_points = 0;
   for (size_t i = 0, size = s.size(); i != size; ++i) {
-    if ((data[i] & 0xc0) != 0x80 && ++num_code_points > n) {
-      return i;
-    }
+    if ((data[i] & 0xc0) != 0x80 && ++num_code_points > n) return i;
   }
   return s.size();
 }
 
-template <typename InputIt, typename OutChar>
-using needs_conversion = bool_constant<
-    std::is_same<typename std::iterator_traits<InputIt>::value_type,
-                 char>::value &&
-    std::is_same<OutChar, char8_type>::value>;
-
-template <typename OutChar, typename InputIt, typename OutputIt,
-          FMT_ENABLE_IF(!needs_conversion<InputIt, OutChar>::value)>
-OutputIt copy_str(InputIt begin, InputIt end, OutputIt it) {
-  return std::copy(begin, end, it);
-}
-
-template <typename OutChar, typename InputIt, typename OutputIt,
-          FMT_ENABLE_IF(needs_conversion<InputIt, OutChar>::value)>
-OutputIt copy_str(InputIt begin, InputIt end, OutputIt it) {
-  return std::transform(begin, end, it,
-                        [](char c) { return static_cast<char8_type>(c); });
-}
-
-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)>;
+template <typename T, bool = std::is_floating_point<T>::value>
+struct is_fast_float : bool_constant<std::numeric_limits<T>::is_iec559 &&
+                                     sizeof(T) <= sizeof(double)> {};
+template <typename T> struct is_fast_float<T, false> : std::false_type {};
 
 #ifndef FMT_USE_FULL_CACHE_DRAGONBOX
 #  define FMT_USE_FULL_CACHE_DRAGONBOX 0
@@ -598,7 +628,7 @@ using is_fast_float = bool_constant<std::numeric_limits<T>::is_iec559 &&
 template <typename T>
 template <typename U>
 void buffer<T>::append(const U* begin, const U* end) {
-  do {
+  while (begin != end) {
     auto count = to_unsigned(end - begin);
     try_reserve(size_ + count);
     auto free_cap = capacity_ - size_;
@@ -606,16 +636,17 @@ void buffer<T>::append(const U* begin, const U* end) {
     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();
-}
+template <typename T, typename Enable = void>
+struct is_locale : std::false_type {};
+template <typename T>
+struct is_locale<T, void_t<decltype(T::classic())>> : std::true_type {};
 }  // namespace detail
 
+FMT_MODULE_EXPORT_BEGIN
+
 // The number of characters to store in the basic_memory_buffer object itself
 // to avoid dynamic memory allocation.
 enum { inline_buffer_size = 500 };
@@ -625,20 +656,12 @@ enum { inline_buffer_size = 500 };
   A dynamically growing memory buffer for trivially copyable/constructible types
   with the first ``SIZE`` elements stored in the object itself.
 
-  You can use one of the following type aliases for common character types:
-
-  +----------------+------------------------------+
-  | Type           | Definition                   |
-  +================+==============================+
-  | memory_buffer  | basic_memory_buffer<char>    |
-  +----------------+------------------------------+
-  | wmemory_buffer | basic_memory_buffer<wchar_t> |
-  +----------------+------------------------------+
+  You can use the ``memory_buffer`` type alias for ``char`` instead.
 
   **Example**::
 
-     fmt::memory_buffer out;
-     format_to(out, "The answer is {}.", 42);
+     auto out = fmt::memory_buffer();
+     format_to(std::back_inserter(out), "The answer is {}.", 42);
 
   This will append the following output to the ``out`` object:
 
@@ -659,34 +682,43 @@ class basic_memory_buffer final : public detail::buffer<T> {
   Allocator alloc_;
 
   // Deallocate memory allocated by the buffer.
-  void deallocate() {
+  FMT_CONSTEXPR20 void deallocate() {
     T* data = this->data();
     if (data != store_) alloc_.deallocate(data, this->capacity());
   }
 
  protected:
-  void grow(size_t size) final FMT_OVERRIDE;
+  FMT_CONSTEXPR20 void grow(size_t size) override;
 
  public:
   using value_type = T;
   using const_reference = const T&;
 
-  explicit basic_memory_buffer(const Allocator& alloc = Allocator())
+  FMT_CONSTEXPR20 explicit basic_memory_buffer(
+      const Allocator& alloc = Allocator())
       : alloc_(alloc) {
     this->set(store_, SIZE);
+    if (detail::is_constant_evaluated()) {
+      detail::fill_n(store_, SIZE, T{});
+    }
   }
-  ~basic_memory_buffer() { deallocate(); }
+  FMT_CONSTEXPR20 ~basic_memory_buffer() { deallocate(); }
 
  private:
   // Move data from other to this buffer.
-  void move(basic_memory_buffer& other) {
+  FMT_CONSTEXPR20 void move(basic_memory_buffer& other) {
     alloc_ = std::move(other.alloc_);
     T* data = other.data();
     size_t size = other.size(), capacity = other.capacity();
     if (data == other.store_) {
       this->set(store_, capacity);
-      std::uninitialized_copy(other.store_, other.store_ + size,
-                              detail::make_checked(store_, capacity));
+      if (detail::is_constant_evaluated()) {
+        detail::copy_str<T>(other.store_, other.store_ + size,
+                            detail::make_checked(store_, capacity));
+      } else {
+        std::uninitialized_copy(other.store_, other.store_ + size,
+                                detail::make_checked(store_, capacity));
+      }
     } else {
       this->set(data, capacity);
       // Set pointer to the inline array so that delete is not called
@@ -703,14 +735,18 @@ class basic_memory_buffer final : public detail::buffer<T> {
     of the other object to it.
     \endrst
    */
-  basic_memory_buffer(basic_memory_buffer&& other) FMT_NOEXCEPT { move(other); }
+  FMT_CONSTEXPR20 basic_memory_buffer(basic_memory_buffer&& other)
+      FMT_NOEXCEPT {
+    move(other);
+  }
 
   /**
     \rst
     Moves the content of the other ``basic_memory_buffer`` object to this one.
     \endrst
    */
-  basic_memory_buffer& operator=(basic_memory_buffer&& other) FMT_NOEXCEPT {
+  auto operator=(basic_memory_buffer&& other) FMT_NOEXCEPT
+      -> basic_memory_buffer& {
     FMT_ASSERT(this != &other, "");
     deallocate();
     move(other);
@@ -718,13 +754,13 @@ class basic_memory_buffer final : public detail::buffer<T> {
   }
 
   // Returns a copy of the allocator associated with this buffer.
-  Allocator get_allocator() const { return alloc_; }
+  auto get_allocator() const -> Allocator { 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); }
+  FMT_CONSTEXPR20 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); }
@@ -738,13 +774,18 @@ class basic_memory_buffer final : public detail::buffer<T> {
 };
 
 template <typename T, size_t SIZE, typename Allocator>
-void basic_memory_buffer<T, SIZE, Allocator>::grow(size_t size) {
+FMT_CONSTEXPR20 void basic_memory_buffer<T, SIZE, Allocator>::grow(
+    size_t size) {
 #ifdef FMT_FUZZ
   if (size > 5000) throw std::runtime_error("fuzz mode - won't grow that much");
 #endif
+  const size_t max_size = std::allocator_traits<Allocator>::max_size(alloc_);
   size_t old_capacity = this->capacity();
   size_t new_capacity = old_capacity + old_capacity / 2;
-  if (size > new_capacity) new_capacity = size;
+  if (size > new_capacity)
+    new_capacity = size;
+  else if (new_capacity > max_size)
+    new_capacity = size > max_size ? size : max_size;
   T* old_data = this->data();
   T* new_data =
       std::allocator_traits<Allocator>::allocate(alloc_, new_capacity);
@@ -759,12 +800,15 @@ void basic_memory_buffer<T, SIZE, Allocator>::grow(size_t size) {
 }
 
 using memory_buffer = basic_memory_buffer<char>;
-using wmemory_buffer = basic_memory_buffer<wchar_t>;
 
 template <typename T, size_t SIZE, typename Allocator>
 struct is_contiguous<basic_memory_buffer<T, SIZE, Allocator>> : std::true_type {
 };
 
+namespace detail {
+FMT_API void print(std::FILE*, string_view);
+}
+
 /** A formatting error such as invalid format string. */
 FMT_CLASS_API
 class FMT_API format_error : public std::runtime_error {
@@ -776,10 +820,62 @@ class FMT_API format_error : public std::runtime_error {
   format_error& operator=(const format_error&) = default;
   format_error(format_error&&) = default;
   format_error& operator=(format_error&&) = default;
-  ~format_error() FMT_NOEXCEPT FMT_OVERRIDE;
+  ~format_error() FMT_NOEXCEPT override FMT_MSC_DEFAULT;
 };
 
-namespace detail {
+/**
+  \rst
+  Constructs a `~fmt::format_arg_store` object that contains references
+  to arguments and can be implicitly converted to `~fmt::format_args`.
+  If ``fmt`` 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>>
+FMT_INLINE auto make_args_checked(const S& fmt,
+                                  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...>(fmt);
+  return {args...};
+}
+
+// compile-time support
+namespace detail_exported {
+#if FMT_USE_NONTYPE_TEMPLATE_PARAMETERS
+template <typename Char, size_t N> struct fixed_string {
+  constexpr fixed_string(const Char (&str)[N]) {
+    detail::copy_str<Char, const Char*, Char*>(static_cast<const Char*>(str),
+                                               str + N, data);
+  }
+  Char data[N]{};
+};
+#endif
+
+// Converts a compile-time string to basic_string_view.
+template <typename Char, size_t N>
+constexpr auto compile_string_to_view(const Char (&s)[N])
+    -> basic_string_view<Char> {
+  // Remove trailing NUL character if needed. Won't be present if this is used
+  // with a raw character array (i.e. not defined as a string).
+  return {s, N - (std::char_traits<Char>::to_int_type(s[N - 1]) == 0 ? 1 : 0)};
+}
+template <typename Char>
+constexpr auto compile_string_to_view(detail::std_string_view<Char> s)
+    -> basic_string_view<Char> {
+  return {s.data(), s.size()};
+}
+}  // namespace detail_exported
+
+FMT_BEGIN_DETAIL_NAMESPACE
+
+template <typename T> struct is_integral : std::is_integral<T> {};
+template <> struct is_integral<int128_t> : std::true_type {};
+template <> struct is_integral<uint128_t> : std::true_type {};
 
 template <typename T>
 using is_signed =
@@ -789,16 +885,16 @@ using is_signed =
 // 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(is_signed<T>::value)>
-FMT_CONSTEXPR bool is_negative(T value) {
+FMT_CONSTEXPR auto is_negative(T value) -> bool {
   return value < 0;
 }
 template <typename T, FMT_ENABLE_IF(!is_signed<T>::value)>
-FMT_CONSTEXPR bool is_negative(T) {
+FMT_CONSTEXPR auto is_negative(T) -> bool {
   return false;
 }
 
 template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
-FMT_CONSTEXPR bool is_supported_floating_point(T) {
+FMT_CONSTEXPR auto is_supported_floating_point(T) -> uint16_t {
   return (std::is_same<T, float>::value && FMT_USE_FLOAT) ||
          (std::is_same<T, double>::value && FMT_USE_DOUBLE) ||
          (std::is_same<T, long double>::value && FMT_USE_LONG_DOUBLE);
@@ -811,121 +907,33 @@ 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_new[];
-  static const uint64_t zero_or_powers_of_10_64_new[];
-  static const uint64_t grisu_pow10_significands[];
-  static const int16_t grisu_pow10_exponents[];
-  static const divtest_table_entry<uint32_t> divtest_table_for_pow5_32[];
-  static const divtest_table_entry<uint64_t> divtest_table_for_pow5_64[];
-  static const uint64_t dragonbox_pow10_significands_64[];
-  static const uint128_wrapper dragonbox_pow10_significands_128[];
-  // log10(2) = 0x0.4d104d427de7fbcc...
-  static const uint64_t log10_2_significand = 0x4d104d427de7fbcc;
-#if !FMT_USE_FULL_CACHE_DRAGONBOX
-  static const uint64_t powers_of_5_64[];
-  static const uint32_t dragonbox_pow10_recovery_errors[];
+template <typename T>
+using uint64_or_128_t = conditional_t<num_bits<T>() <= 64, uint64_t, uint128_t>;
+
+#define FMT_POWERS_OF_10(factor)                                             \
+  factor * 10, (factor)*100, (factor)*1000, (factor)*10000, (factor)*100000, \
+      (factor)*1000000, (factor)*10000000, (factor)*100000000,               \
+      (factor)*1000000000
+
+// Converts value in the range [0, 100) to a string.
+constexpr const char* digits2(size_t value) {
+  // GCC generates slightly better code when value is pointer-size.
+  return &"0001020304050607080910111213141516171819"
+         "2021222324252627282930313233343536373839"
+         "4041424344454647484950515253545556575859"
+         "6061626364656667686970717273747576777879"
+         "8081828384858687888990919293949596979899"[value * 2];
+}
+
+// Sign is a template parameter to workaround a bug in gcc 4.8.
+template <typename Char, typename Sign> constexpr Char sign(Sign s) {
+#if !FMT_GCC_VERSION || FMT_GCC_VERSION >= 604
+  static_assert(std::is_same<Sign, sign_t>::value, "");
 #endif
-  // GCC generates slightly better code for pairs than chars.
-  using digit_pair = char[2];
-  static const digit_pair digits[];
-  static const char hex_digits[];
-  static const char foreground_color[];
-  static const char background_color[];
-  static const char reset_color[5];
-  static const wchar_t wreset_color[5];
-  static const char signs[];
-  static const char left_padding_shifts[5];
-  static const char right_padding_shifts[5];
-
-  // DEPRECATED! These are for ABI compatibility.
-  static const uint32_t zero_or_powers_of_10_32[];
-  static const uint64_t zero_or_powers_of_10_64[];
-};
-
-// Maps bsr(n) to ceil(log10(pow(2, bsr(n) + 1) - 1)).
-// This is a function instead of an array to workaround a bug in GCC10 (#1810).
-FMT_INLINE uint16_t bsr2log10(int bsr) {
-  static constexpr uint16_t data[] = {
-      1,  1,  1,  2,  2,  2,  3,  3,  3,  4,  4,  4,  4,  5,  5,  5,
-      6,  6,  6,  7,  7,  7,  7,  8,  8,  8,  9,  9,  9,  10, 10, 10,
-      10, 11, 11, 11, 12, 12, 12, 13, 13, 13, 13, 14, 14, 14, 15, 15,
-      15, 16, 16, 16, 16, 17, 17, 17, 18, 18, 18, 19, 19, 19, 19, 20};
-  return data[bsr];
+  return static_cast<Char>("\0-+ "[s]);
 }
 
-#ifndef FMT_EXPORTED
-FMT_EXTERN template struct basic_data<void>;
-#endif
-
-// This is a struct rather than an alias to avoid shadowing warnings in gcc.
-struct data : basic_data<> {};
-
-#ifdef FMT_BUILTIN_CLZLL
-// 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) {
-  // https://github.com/fmtlib/format-benchmark/blob/master/digits10
-  auto t = bsr2log10(FMT_BUILTIN_CLZLL(n | 1) ^ 63);
-  return t - (n < data::zero_or_powers_of_10_64_new[t]);
-}
-#else
-// Fallback version of count_digits used when __builtin_clz is not available.
-inline int count_digits(uint64_t n) {
+template <typename T> FMT_CONSTEXPR auto count_digits_fallback(T n) -> int {
   int count = 1;
   for (;;) {
     // Integer division is slow so do it for a group of four digits instead
@@ -939,103 +947,152 @@ inline int count_digits(uint64_t n) {
     count += 4;
   }
 }
-#endif
-
 #if FMT_USE_INT128
-inline int count_digits(uint128_t n) {
-  int count = 1;
-  for (;;) {
-    // Integer division is slow so do it for a group of four digits instead
-    // of for every digit. The idea comes from the talk by Alexandrescu
-    // "Three Optimization Tips for C++". See speed-test for a comparison.
-    if (n < 10) return count;
-    if (n < 100) return count + 1;
-    if (n < 1000) return count + 2;
-    if (n < 10000) return count + 3;
-    n /= 10000U;
-    count += 4;
-  }
+FMT_CONSTEXPR inline auto count_digits(uint128_t n) -> int {
+  return count_digits_fallback(n);
 }
 #endif
 
-// Counts the number of digits in n. BITS = log2(radix).
-template <unsigned BITS, typename UInt> inline int count_digits(UInt n) {
-  int num_digits = 0;
-  do {
-    ++num_digits;
-  } while ((n >>= BITS) != 0);
-  return num_digits;
+#ifdef FMT_BUILTIN_CLZLL
+// It is a separate function rather than a part of count_digits to workaround
+// the lack of static constexpr in constexpr functions.
+inline auto do_count_digits(uint64_t n) -> int {
+  // This has comparable performance to the version by Kendall Willets
+  // (https://github.com/fmtlib/format-benchmark/blob/master/digits10)
+  // but uses smaller tables.
+  // Maps bsr(n) to ceil(log10(pow(2, bsr(n) + 1) - 1)).
+  static constexpr uint8_t bsr2log10[] = {
+      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};
+  auto t = bsr2log10[FMT_BUILTIN_CLZLL(n | 1) ^ 63];
+  static constexpr const uint64_t zero_or_powers_of_10[] = {
+      0, 0, FMT_POWERS_OF_10(1U), FMT_POWERS_OF_10(1000000000ULL),
+      10000000000000000000ULL};
+  return t - (n < zero_or_powers_of_10[t]);
 }
+#endif
 
-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
+// Returns the number of decimal digits in n. Leading zeros are not counted
+// except for n == 0 in which case count_digits returns 1.
+FMT_CONSTEXPR20 inline auto count_digits(uint64_t n) -> int {
+#ifdef FMT_BUILTIN_CLZLL
+  if (!is_constant_evaluated()) {
+    return do_count_digits(n);
+  }
 #endif
+  return count_digits_fallback(n);
+}
 
-// To suppress unnecessary security cookie checks
-#if FMT_MSC_VER && !FMT_CLANG_VERSION
-#  define FMT_SAFEBUFFERS __declspec(safebuffers)
-#else
-#  define FMT_SAFEBUFFERS
+// Counts the number of digits in n. BITS = log2(radix).
+template <int BITS, typename UInt>
+FMT_CONSTEXPR auto count_digits(UInt n) -> int {
+#ifdef FMT_BUILTIN_CLZ
+  if (num_bits<UInt>() == 32)
+    return (FMT_BUILTIN_CLZ(static_cast<uint32_t>(n) | 1) ^ 31) / BITS + 1;
 #endif
+  // Lambda avoids unreachable code warnings from NVHPC.
+  return [](UInt m) {
+    int num_digits = 0;
+    do {
+      ++num_digits;
+    } while ((m >>= BITS) != 0);
+    return num_digits;
+  }(n);
+}
+
+template <> auto count_digits<4>(detail::fallback_uintptr n) -> int;
 
 #ifdef FMT_BUILTIN_CLZ
-// Optional version of count_digits for better performance on 32-bit platforms.
-inline int count_digits(uint32_t n) {
-  auto t = bsr2log10(FMT_BUILTIN_CLZ(n | 1) ^ 31);
-  return t - (n < data::zero_or_powers_of_10_32_new[t]);
+// It is a separate function rather than a part of count_digits to workaround
+// the lack of static constexpr in constexpr functions.
+FMT_INLINE auto do_count_digits(uint32_t n) -> int {
+// An optimization by Kendall Willets from https://bit.ly/3uOIQrB.
+// This increments the upper 32 bits (log10(T) - 1) when >= T is added.
+#  define FMT_INC(T) (((sizeof(#  T) - 1ull) << 32) - T)
+  static constexpr uint64_t table[] = {
+      FMT_INC(0),          FMT_INC(0),          FMT_INC(0),           // 8
+      FMT_INC(10),         FMT_INC(10),         FMT_INC(10),          // 64
+      FMT_INC(100),        FMT_INC(100),        FMT_INC(100),         // 512
+      FMT_INC(1000),       FMT_INC(1000),       FMT_INC(1000),        // 4096
+      FMT_INC(10000),      FMT_INC(10000),      FMT_INC(10000),       // 32k
+      FMT_INC(100000),     FMT_INC(100000),     FMT_INC(100000),      // 256k
+      FMT_INC(1000000),    FMT_INC(1000000),    FMT_INC(1000000),     // 2048k
+      FMT_INC(10000000),   FMT_INC(10000000),   FMT_INC(10000000),    // 16M
+      FMT_INC(100000000),  FMT_INC(100000000),  FMT_INC(100000000),   // 128M
+      FMT_INC(1000000000), FMT_INC(1000000000), FMT_INC(1000000000),  // 1024M
+      FMT_INC(1000000000), FMT_INC(1000000000)                        // 4B
+  };
+  auto inc = table[FMT_BUILTIN_CLZ(n | 1) ^ 31];
+  return static_cast<int>((n + inc) >> 32);
 }
 #endif
 
-template <typename Int> constexpr int digits10() FMT_NOEXCEPT {
-  return std::numeric_limits<Int>::digits10;
+// Optional version of count_digits for better performance on 32-bit platforms.
+FMT_CONSTEXPR20 inline auto count_digits(uint32_t n) -> int {
+#ifdef FMT_BUILTIN_CLZ
+  if (!is_constant_evaluated()) {
+    return do_count_digits(n);
+  }
+#endif
+  return count_digits_fallback(n);
 }
-template <> constexpr int digits10<int128_t>() FMT_NOEXCEPT { return 38; }
-template <> constexpr int digits10<uint128_t>() FMT_NOEXCEPT { return 38; }
 
-template <typename Char> FMT_API std::string grouping_impl(locale_ref loc);
-template <typename Char> inline std::string grouping(locale_ref loc) {
-  return grouping_impl<char>(loc);
+template <typename Int> constexpr auto digits10() FMT_NOEXCEPT -> int {
+  return std::numeric_limits<Int>::digits10;
+}
+template <> constexpr auto digits10<int128_t>() FMT_NOEXCEPT -> int {
+  return 38;
 }
-template <> inline std::string grouping<wchar_t>(locale_ref loc) {
-  return grouping_impl<wchar_t>(loc);
+template <> constexpr auto digits10<uint128_t>() FMT_NOEXCEPT -> int {
+  return 38;
 }
 
-template <typename Char> FMT_API Char thousands_sep_impl(locale_ref loc);
-template <typename Char> inline Char thousands_sep(locale_ref loc) {
-  return Char(thousands_sep_impl<char>(loc));
+template <typename Char> struct thousands_sep_result {
+  std::string grouping;
+  Char thousands_sep;
+};
+
+template <typename Char>
+FMT_API auto thousands_sep_impl(locale_ref loc) -> thousands_sep_result<Char>;
+template <typename Char>
+inline auto thousands_sep(locale_ref loc) -> thousands_sep_result<Char> {
+  auto result = thousands_sep_impl<char>(loc);
+  return {result.grouping, Char(result.thousands_sep)};
 }
-template <> inline wchar_t thousands_sep(locale_ref loc) {
+template <>
+inline auto thousands_sep(locale_ref loc) -> thousands_sep_result<wchar_t> {
   return thousands_sep_impl<wchar_t>(loc);
 }
 
-template <typename Char> FMT_API Char decimal_point_impl(locale_ref loc);
-template <typename Char> inline Char decimal_point(locale_ref loc) {
+template <typename Char>
+FMT_API auto decimal_point_impl(locale_ref loc) -> Char;
+template <typename Char> inline auto decimal_point(locale_ref loc) -> Char {
   return Char(decimal_point_impl<char>(loc));
 }
-template <> inline wchar_t decimal_point(locale_ref loc) {
+template <> inline auto decimal_point(locale_ref loc) -> wchar_t {
   return decimal_point_impl<wchar_t>(loc);
 }
 
 // Compares two characters for equality.
-template <typename Char> bool equal2(const Char* lhs, const char* rhs) {
-  return lhs[0] == rhs[0] && lhs[1] == rhs[1];
+template <typename Char> auto equal2(const Char* lhs, const char* rhs) -> bool {
+  return lhs[0] == Char(rhs[0]) && lhs[1] == Char(rhs[1]);
 }
-inline bool equal2(const char* lhs, const char* rhs) {
+inline auto equal2(const char* lhs, const char* rhs) -> bool {
   return memcmp(lhs, rhs, 2) == 0;
 }
 
 // Copies two characters from src to dst.
-template <typename Char> void copy2(Char* dst, const char* src) {
+template <typename Char>
+FMT_CONSTEXPR20 FMT_INLINE void copy2(Char* dst, const char* src) {
+  if (!is_constant_evaluated() && sizeof(Char) == sizeof(char)) {
+    memcpy(dst, src, 2);
+    return;
+  }
   *dst++ = static_cast<Char>(*src++);
   *dst = static_cast<Char>(*src);
 }
-FMT_INLINE void copy2(char* dst, const char* src) { memcpy(dst, src, 2); }
 
 template <typename Iterator> struct format_decimal_result {
   Iterator begin;
@@ -1046,8 +1103,8 @@ template <typename Iterator> struct format_decimal_result {
 // buffer of specified size. The caller must ensure that the buffer is large
 // enough.
 template <typename Char, typename UInt>
-inline format_decimal_result<Char*> format_decimal(Char* out, UInt value,
-                                                   int size) {
+FMT_CONSTEXPR20 auto format_decimal(Char* out, UInt value, int size)
+    -> format_decimal_result<Char*> {
   FMT_ASSERT(size >= count_digits(value), "invalid digit count");
   out += size;
   Char* end = out;
@@ -1056,7 +1113,7 @@ inline format_decimal_result<Char*> format_decimal(Char* out, UInt value,
     // of for every digit. The idea comes from the talk by Alexandrescu
     // "Three Optimization Tips for C++". See speed-test for a comparison.
     out -= 2;
-    copy2(out, data::digits[value % 100]);
+    copy2(out, digits2(static_cast<size_t>(value % 100)));
     value /= 100;
   }
   if (value < 10) {
@@ -1064,27 +1121,27 @@ inline format_decimal_result<Char*> format_decimal(Char* out, UInt value,
     return {out, end};
   }
   out -= 2;
-  copy2(out, data::digits[value]);
+  copy2(out, digits2(static_cast<size_t>(value)));
   return {out, end};
 }
 
 template <typename Char, typename UInt, typename Iterator,
           FMT_ENABLE_IF(!std::is_pointer<remove_cvref_t<Iterator>>::value)>
-inline format_decimal_result<Iterator> format_decimal(Iterator out, UInt value,
-                                                      int size) {
+inline auto format_decimal(Iterator out, UInt value, int size)
+    -> format_decimal_result<Iterator> {
   // 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)};
+  return {out, detail::copy_str_noinline<Char>(buffer, end, out)};
 }
 
 template <unsigned BASE_BITS, typename Char, typename UInt>
-inline Char* format_uint(Char* buffer, UInt value, int num_digits,
-                         bool upper = false) {
+FMT_CONSTEXPR auto format_uint(Char* buffer, UInt value, int num_digits,
+                               bool upper = false) -> Char* {
   buffer += num_digits;
   Char* end = buffer;
   do {
-    const char* digits = upper ? "0123456789ABCDEF" : data::hex_digits;
+    const char* digits = upper ? "0123456789ABCDEF" : "0123456789abcdef";
     unsigned digit = (value & ((1 << BASE_BITS) - 1));
     *--buffer = static_cast<Char>(BASE_BITS < 4 ? static_cast<char>('0' + digit)
                                                 : digits[digit]);
@@ -1093,8 +1150,8 @@ inline Char* format_uint(Char* buffer, UInt value, int num_digits,
 }
 
 template <unsigned BASE_BITS, typename Char>
-Char* format_uint(Char* buffer, detail::fallback_uintptr n, int num_digits,
-                  bool = false) {
+auto format_uint(Char* buffer, detail::fallback_uintptr n, int num_digits,
+                 bool = false) -> Char* {
   auto char_digits = std::numeric_limits<unsigned char>::digits / 4;
   int start = (num_digits + char_digits - 1) / char_digits - 1;
   if (int start_digits = num_digits % char_digits) {
@@ -1107,7 +1164,7 @@ Char* format_uint(Char* buffer, detail::fallback_uintptr n, int num_digits,
     auto p = buffer;
     for (int i = 0; i < char_digits; ++i) {
       unsigned digit = (value & ((1 << BASE_BITS) - 1));
-      *--p = static_cast<Char>(data::hex_digits[digit]);
+      *--p = static_cast<Char>("0123456789abcdef"[digit]);
       value >>= BASE_BITS;
     }
   }
@@ -1115,7 +1172,8 @@ 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) {
+inline auto format_uint(It out, UInt value, int num_digits, bool upper = false)
+    -> It {
   if (auto ptr = to_pointer<Char>(out, to_unsigned(num_digits))) {
     format_uint<BASE_BITS>(ptr, value, num_digits, upper);
     return out;
@@ -1123,90 +1181,26 @@ inline It format_uint(It out, UInt value, int num_digits, bool upper = false) {
   // 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);
-  return detail::copy_str<Char>(buffer, buffer + num_digits, out);
+  return detail::copy_str_noinline<Char>(buffer, buffer + num_digits, out);
 }
 
 // A converter from UTF-8 to UTF-16.
 class utf8_to_utf16 {
  private:
-  wmemory_buffer buffer_;
+  basic_memory_buffer<wchar_t> buffer_;
 
  public:
   FMT_API explicit utf8_to_utf16(string_view s);
-  operator wstring_view() const { return {&buffer_[0], size()}; }
-  size_t size() const { return buffer_.size() - 1; }
-  const wchar_t* c_str() const { return &buffer_[0]; }
-  std::wstring str() const { return {&buffer_[0], size()}; }
+  operator basic_string_view<wchar_t>() const { return {&buffer_[0], size()}; }
+  auto size() const -> size_t { return buffer_.size() - 1; }
+  auto c_str() const -> const wchar_t* { return &buffer_[0]; }
+  auto str() const -> std::wstring { return {&buffer_[0], size()}; }
 };
 
-template <typename T = void> struct null {};
+namespace dragonbox {
 
-// Workaround an array initialization issue in gcc 4.8.
-template <typename Char> struct fill_t {
- private:
-  enum { max_size = 4 };
-  Char data_[max_size] = {Char(' '), Char(0), Char(0), Char(0)};
-  unsigned char size_ = 1;
-
- public:
-  FMT_CONSTEXPR void operator=(basic_string_view<Char> s) {
-    auto size = s.size();
-    if (size > max_size) {
-      FMT_THROW(format_error("invalid fill"));
-      return;
-    }
-    for (size_t i = 0; i < size; ++i) data_[i] = s[i];
-    size_ = static_cast<unsigned char>(size);
-  }
-
-  size_t size() const { return size_; }
-  const Char* data() const { return data_; }
-
-  FMT_CONSTEXPR Char& operator[](size_t index) { return data_[index]; }
-  FMT_CONSTEXPR const Char& operator[](size_t index) const {
-    return data_[index];
-  }
-};
-}  // namespace detail
-
-// We cannot use enum classes as bit fields because of a gcc bug
-// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61414.
-namespace align {
-enum type { none, left, right, center, numeric };
-}
-using align_t = align::type;
-
-namespace sign {
-enum type { none, minus, plus, space };
-}
-using sign_t = sign::type;
-
-// Format specifiers for built-in and string types.
-template <typename Char> struct basic_format_specs {
-  int width;
-  int precision;
-  char type;
-  align_t align : 4;
-  sign_t sign : 3;
-  bool alt : 1;  // Alternate form ('#').
-  detail::fill_t<Char> fill;
-
-  constexpr basic_format_specs()
-      : width(0),
-        precision(-1),
-        type(0),
-        align(align::none),
-        sign(sign::none),
-        alt(false) {}
-};
-
-using format_specs = basic_format_specs<char>;
-
-namespace detail {
-namespace dragonbox {
-
-// Type-specific information that Dragonbox uses.
-template <class T> struct float_info;
+// Type-specific information that Dragonbox uses.
+template <class T> struct float_info;
 
 template <> struct float_info<float> {
   using carrier_uint = uint32_t;
@@ -1266,37 +1260,21 @@ template <typename T> struct decimal_fp {
   int exponent;
 };
 
-template <typename T> FMT_API decimal_fp<T> to_decimal(T x) FMT_NOEXCEPT;
+template <typename T>
+FMT_API auto to_decimal(T x) FMT_NOEXCEPT -> decimal_fp<T>;
 }  // namespace dragonbox
 
 template <typename T>
-constexpr typename dragonbox::float_info<T>::carrier_uint exponent_mask() {
+constexpr auto exponent_mask() ->
+    typename dragonbox::float_info<T>::carrier_uint {
   using uint = typename dragonbox::float_info<T>::carrier_uint;
   return ((uint(1) << dragonbox::float_info<T>::exponent_bits) - 1)
          << dragonbox::float_info<T>::significand_bits;
 }
 
-// A floating-point presentation format.
-enum class float_format : unsigned char {
-  general,  // General: exponent notation or fixed point based on magnitude.
-  exp,      // Exponent notation with the default precision of 6, e.g. 1.2e-3.
-  fixed,    // Fixed point with the default precision of 6, e.g. 0.0012.
-  hex
-};
-
-struct float_specs {
-  int precision;
-  float_format format : 8;
-  sign_t sign : 8;
-  bool upper : 1;
-  bool locale : 1;
-  bool binary32 : 1;
-  bool use_grisu : 1;
-  bool showpoint : 1;
-};
-
 // Writes the exponent exp in the form "[+-]d{2,3}" to buffer.
-template <typename Char, typename It> It write_exponent(int exp, It it) {
+template <typename Char, typename It>
+FMT_CONSTEXPR auto write_exponent(int exp, It it) -> It {
   FMT_ASSERT(-10000 < exp && exp < 10000, "exponent out of range");
   if (exp < 0) {
     *it++ = static_cast<Char>('-');
@@ -1305,185 +1283,42 @@ template <typename Char, typename It> It write_exponent(int exp, It it) {
     *it++ = static_cast<Char>('+');
   }
   if (exp >= 100) {
-    const char* top = data::digits[exp / 100];
+    const char* top = digits2(to_unsigned(exp / 100));
     if (exp >= 1000) *it++ = static_cast<Char>(top[0]);
     *it++ = static_cast<Char>(top[1]);
     exp %= 100;
   }
-  const char* d = data::digits[exp];
+  const char* d = digits2(to_unsigned(exp));
   *it++ = static_cast<Char>(d[0]);
   *it++ = static_cast<Char>(d[1]);
   return it;
 }
 
 template <typename T>
-int format_float(T value, int precision, float_specs specs, buffer<char>& buf);
+FMT_HEADER_ONLY_CONSTEXPR20 auto format_float(T value, int precision,
+                                              float_specs specs,
+                                              buffer<char>& buf) -> int;
 
 // Formats a floating-point number with snprintf.
 template <typename T>
-int snprintf_float(T value, int precision, float_specs specs,
-                   buffer<char>& buf);
-
-template <typename T> T promote_float(T value) { return value; }
-inline double promote_float(float value) { return static_cast<double>(value); }
-
-template <typename Handler>
-FMT_CONSTEXPR void handle_int_type_spec(char spec, Handler&& handler) {
-  switch (spec) {
-  case 0:
-  case 'd':
-    handler.on_dec();
-    break;
-  case 'x':
-  case 'X':
-    handler.on_hex();
-    break;
-  case 'b':
-  case 'B':
-    handler.on_bin();
-    break;
-  case 'o':
-    handler.on_oct();
-    break;
-#ifdef FMT_DEPRECATED_N_SPECIFIER
-  case 'n':
-#endif
-  case 'L':
-    handler.on_num();
-    break;
-  case 'c':
-    handler.on_chr();
-    break;
-  default:
-    handler.on_error();
-  }
-}
-
-template <typename ErrorHandler = error_handler, typename Char>
-FMT_CONSTEXPR float_specs parse_float_type_spec(
-    const basic_format_specs<Char>& specs, ErrorHandler&& eh = {}) {
-  auto result = float_specs();
-  result.showpoint = specs.alt;
-  switch (specs.type) {
-  case 0:
-    result.format = float_format::general;
-    result.showpoint |= specs.precision > 0;
-    break;
-  case 'G':
-    result.upper = true;
-    FMT_FALLTHROUGH;
-  case 'g':
-    result.format = float_format::general;
-    break;
-  case 'E':
-    result.upper = true;
-    FMT_FALLTHROUGH;
-  case 'e':
-    result.format = float_format::exp;
-    result.showpoint |= specs.precision != 0;
-    break;
-  case 'F':
-    result.upper = true;
-    FMT_FALLTHROUGH;
-  case 'f':
-    result.format = float_format::fixed;
-    result.showpoint |= specs.precision != 0;
-    break;
-  case 'A':
-    result.upper = true;
-    FMT_FALLTHROUGH;
-  case 'a':
-    result.format = float_format::hex;
-    break;
-#ifdef FMT_DEPRECATED_N_SPECIFIER
-  case 'n':
-#endif
-  case 'L':
-    result.locale = true;
-    break;
-  default:
-    eh.on_error("invalid type specifier");
-    break;
-  }
-  return result;
-}
-
-template <typename Char, typename Handler>
-FMT_CONSTEXPR void handle_char_specs(const basic_format_specs<Char>* specs,
-                                     Handler&& handler) {
-  if (!specs) return handler.on_char();
-  if (specs->type && specs->type != 'c') return handler.on_int();
-  if (specs->align == align::numeric || specs->sign != sign::none || specs->alt)
-    handler.on_error("invalid format specifier for char");
-  handler.on_char();
-}
-
-template <typename Char, typename Handler>
-FMT_CONSTEXPR void handle_cstring_type_spec(Char spec, Handler&& handler) {
-  if (spec == 0 || spec == 's')
-    handler.on_string();
-  else if (spec == 'p')
-    handler.on_pointer();
-  else
-    handler.on_error("invalid type specifier");
-}
+auto snprintf_float(T value, int precision, float_specs specs,
+                    buffer<char>& buf) -> int;
 
-template <typename Char, typename ErrorHandler>
-FMT_CONSTEXPR void check_string_type_spec(Char spec, ErrorHandler&& eh) {
-  if (spec != 0 && spec != 's') eh.on_error("invalid type specifier");
+template <typename T> constexpr auto promote_float(T value) -> T {
+  return value;
 }
-
-template <typename Char, typename ErrorHandler>
-FMT_CONSTEXPR void check_pointer_type_spec(Char spec, ErrorHandler&& eh) {
-  if (spec != 0 && spec != 'p') eh.on_error("invalid type specifier");
+constexpr auto promote_float(float value) -> double {
+  return static_cast<double>(value);
 }
 
-template <typename ErrorHandler> class int_type_checker : private ErrorHandler {
- public:
-  FMT_CONSTEXPR explicit int_type_checker(ErrorHandler eh) : ErrorHandler(eh) {}
-
-  FMT_CONSTEXPR void on_dec() {}
-  FMT_CONSTEXPR void on_hex() {}
-  FMT_CONSTEXPR void on_bin() {}
-  FMT_CONSTEXPR void on_oct() {}
-  FMT_CONSTEXPR void on_num() {}
-  FMT_CONSTEXPR void on_chr() {}
-
-  FMT_CONSTEXPR void on_error() {
-    ErrorHandler::on_error("invalid type specifier");
-  }
-};
-
-template <typename ErrorHandler>
-class char_specs_checker : public ErrorHandler {
- private:
-  char type_;
-
- public:
-  FMT_CONSTEXPR char_specs_checker(char type, ErrorHandler eh)
-      : ErrorHandler(eh), type_(type) {}
-
-  FMT_CONSTEXPR void on_int() {
-    handle_int_type_spec(type_, int_type_checker<ErrorHandler>(*this));
-  }
-  FMT_CONSTEXPR void on_char() {}
-};
-
-template <typename ErrorHandler>
-class cstring_type_checker : public ErrorHandler {
- public:
-  FMT_CONSTEXPR explicit cstring_type_checker(ErrorHandler eh)
-      : ErrorHandler(eh) {}
-
-  FMT_CONSTEXPR void on_string() {}
-  FMT_CONSTEXPR void on_pointer() {}
-};
-
 template <typename OutputIt, typename Char>
-FMT_NOINLINE OutputIt fill(OutputIt it, size_t n, const fill_t<Char>& fill) {
+FMT_NOINLINE FMT_CONSTEXPR auto fill(OutputIt it, size_t n,
+                                     const fill_t<Char>& fill) -> OutputIt {
   auto fill_size = fill.size();
-  if (fill_size == 1) return std::fill_n(it, n, fill[0]);
-  for (size_t i = 0; i < n; ++i) it = std::copy_n(fill.data(), fill_size, it);
+  if (fill_size == 1) return detail::fill_n(it, n, fill[0]);
+  auto data = fill.data();
+  for (size_t i = 0; i < n; ++i)
+    it = copy_str<Char>(data, data + fill_size, it);
   return it;
 }
 
@@ -1492,39 +1327,73 @@ FMT_NOINLINE OutputIt fill(OutputIt it, size_t n, const fill_t<Char>& fill) {
 // width: output display width in (terminal) column positions.
 template <align::type align = align::left, typename OutputIt, typename Char,
           typename F>
-inline OutputIt write_padded(OutputIt out,
-                             const basic_format_specs<Char>& specs, size_t size,
-                             size_t width, F&& f) {
+FMT_CONSTEXPR auto write_padded(OutputIt out,
+                                const basic_format_specs<Char>& specs,
+                                size_t size, size_t width, F&& f) -> OutputIt {
   static_assert(align == align::left || align == align::right, "");
   unsigned spec_width = to_unsigned(specs.width);
   size_t padding = spec_width > width ? spec_width - width : 0;
-  auto* shifts = align == align::left ? data::left_padding_shifts
-                                      : data::right_padding_shifts;
+  // Shifts are encoded as string literals because static constexpr is not
+  // supported in constexpr functions.
+  auto* shifts = align == align::left ? "\x1f\x1f\x00\x01" : "\x00\x1f\x00\x01";
   size_t left_padding = padding >> shifts[specs.align];
+  size_t right_padding = padding - left_padding;
   auto it = reserve(out, size + padding * specs.fill.size());
-  it = fill(it, left_padding, specs.fill);
+  if (left_padding != 0) it = fill(it, left_padding, specs.fill);
   it = f(it);
-  it = fill(it, padding - left_padding, specs.fill);
+  if (right_padding != 0) it = fill(it, right_padding, specs.fill);
   return base_iterator(out, it);
 }
 
 template <align::type align = align::left, typename OutputIt, typename Char,
           typename F>
-inline OutputIt write_padded(OutputIt out,
-                             const basic_format_specs<Char>& specs, size_t size,
-                             F&& f) {
+constexpr auto write_padded(OutputIt out, const basic_format_specs<Char>& specs,
+                            size_t size, F&& f) -> OutputIt {
   return write_padded<align>(out, specs, size, size, f);
 }
 
+template <align::type align = align::left, typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write_bytes(OutputIt out, string_view bytes,
+                               const basic_format_specs<Char>& specs)
+    -> OutputIt {
+  return write_padded<align>(
+      out, specs, bytes.size(), [bytes](reserve_iterator<OutputIt> it) {
+        const char* data = bytes.data();
+        return copy_str<Char>(data, data + bytes.size(), it);
+      });
+}
+
+template <typename Char, typename OutputIt, typename UIntPtr>
+auto write_ptr(OutputIt out, UIntPtr value,
+               const basic_format_specs<Char>* specs) -> OutputIt {
+  int num_digits = count_digits<4>(value);
+  auto size = to_unsigned(num_digits) + size_t(2);
+  auto write = [=](reserve_iterator<OutputIt> it) {
+    *it++ = static_cast<Char>('0');
+    *it++ = static_cast<Char>('x');
+    return format_uint<4, Char>(it, value, num_digits);
+  };
+  return specs ? write_padded<align::right>(out, *specs, size, write)
+               : base_iterator(out, write(reserve(out, size)));
+}
+
 template <typename Char, typename OutputIt>
-OutputIt write_bytes(OutputIt out, string_view bytes,
-                     const basic_format_specs<Char>& specs) {
-  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
-  return write_padded(out, specs, bytes.size(), [bytes](iterator it) {
-    const char* data = bytes.data();
-    return copy_str<Char>(data, data + bytes.size(), it);
+FMT_CONSTEXPR auto write_char(OutputIt out, Char value,
+                              const basic_format_specs<Char>& specs)
+    -> OutputIt {
+  return write_padded(out, specs, 1, [=](reserve_iterator<OutputIt> it) {
+    *it++ = value;
+    return it;
   });
 }
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, Char value,
+                         const basic_format_specs<Char>& specs,
+                         locale_ref loc = {}) -> OutputIt {
+  return check_char_specs(specs)
+             ? write_char(out, value, specs)
+             : write(out, static_cast<int>(value), specs, loc);
+}
 
 // Data for write_int that doesn't depend on output iterator type. It is used to
 // avoid template code bloat.
@@ -1532,9 +1401,9 @@ template <typename Char> struct write_int_data {
   size_t size;
   size_t padding;
 
-  write_int_data(int num_digits, string_view prefix,
-                 const basic_format_specs<Char>& specs)
-      : size(prefix.size() + to_unsigned(num_digits)), padding(0) {
+  FMT_CONSTEXPR write_int_data(int num_digits, unsigned prefix,
+                               const basic_format_specs<Char>& specs)
+      : size((prefix >> 24) + to_unsigned(num_digits)), padding(0) {
     if (specs.align == align::numeric) {
       auto width = to_unsigned(specs.width);
       if (width > size) {
@@ -1542,7 +1411,7 @@ template <typename Char> struct write_int_data {
         size = width;
       }
     } else if (specs.precision > num_digits) {
-      size = prefix.size() + to_unsigned(specs.precision);
+      size = (prefix >> 24) + to_unsigned(specs.precision);
       padding = to_unsigned(specs.precision - num_digits);
     }
   }
@@ -1550,183 +1419,280 @@ template <typename Char> struct write_int_data {
 
 // Writes an integer in the format
 //   <left-padding><prefix><numeric-padding><digits><right-padding>
-// where <digits> are written by f(it).
-template <typename OutputIt, typename Char, typename F>
-OutputIt write_int(OutputIt out, int num_digits, string_view prefix,
-                   const basic_format_specs<Char>& specs, F f) {
+// where <digits> are written by write_digits(it).
+// prefix contains chars in three lower bytes and the size in the fourth byte.
+template <typename OutputIt, typename Char, typename W>
+FMT_CONSTEXPR FMT_INLINE auto write_int(OutputIt out, int num_digits,
+                                        unsigned prefix,
+                                        const basic_format_specs<Char>& specs,
+                                        W write_digits) -> OutputIt {
+  // Slightly faster check for specs.width == 0 && specs.precision == -1.
+  if ((specs.width | (specs.precision + 1)) == 0) {
+    auto it = reserve(out, to_unsigned(num_digits) + (prefix >> 24));
+    if (prefix != 0) {
+      for (unsigned p = prefix & 0xffffff; p != 0; p >>= 8)
+        *it++ = static_cast<Char>(p & 0xff);
+    }
+    return base_iterator(out, write_digits(it));
+  }
   auto data = write_int_data<Char>(num_digits, prefix, specs);
-  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
-  return write_padded<align::right>(out, specs, data.size, [=](iterator it) {
-    if (prefix.size() != 0)
-      it = copy_str<Char>(prefix.begin(), prefix.end(), it);
-    it = std::fill_n(it, data.padding, static_cast<Char>('0'));
-    return f(it);
-  });
+  return write_padded<align::right>(
+      out, specs, data.size, [=](reserve_iterator<OutputIt> it) {
+        for (unsigned p = prefix & 0xffffff; p != 0; p >>= 8)
+          *it++ = static_cast<Char>(p & 0xff);
+        it = detail::fill_n(it, data.padding, static_cast<Char>('0'));
+        return write_digits(it);
+      });
 }
 
-template <typename StrChar, typename Char, typename OutputIt>
-OutputIt write(OutputIt out, basic_string_view<StrChar> s,
-               const basic_format_specs<Char>& specs) {
-  auto data = s.data();
-  auto size = s.size();
-  if (specs.precision >= 0 && to_unsigned(specs.precision) < size)
-    size = code_point_index(s, to_unsigned(specs.precision));
-  auto width = specs.width != 0
-                   ? count_code_points(basic_string_view<StrChar>(data, size))
-                   : 0;
-  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
-  return write_padded(out, specs, size, width, [=](iterator it) {
-    return copy_str<Char>(data, data + size, it);
-  });
-}
+template <typename Char> class digit_grouping {
+ private:
+  thousands_sep_result<Char> sep_;
 
-// The handle_int_type_spec handler that writes an integer.
-template <typename OutputIt, typename Char, typename UInt> struct int_writer {
-  OutputIt out;
-  locale_ref locale;
-  const basic_format_specs<Char>& specs;
-  UInt abs_value;
-  char prefix[4];
-  unsigned prefix_size;
-
-  using iterator =
-      remove_reference_t<decltype(reserve(std::declval<OutputIt&>(), 0))>;
-
-  string_view get_prefix() const { return string_view(prefix, prefix_size); }
-
-  template <typename Int>
-  int_writer(OutputIt output, locale_ref loc, Int value,
-             const basic_format_specs<Char>& s)
-      : out(output),
-        locale(loc),
-        specs(s),
-        abs_value(static_cast<UInt>(value)),
-        prefix_size(0) {
-    static_assert(std::is_same<uint32_or_64_or_128_t<Int>, UInt>::value, "");
-    if (is_negative(value)) {
-      prefix[0] = '-';
-      ++prefix_size;
-      abs_value = 0 - abs_value;
-    } else if (specs.sign != sign::none && specs.sign != sign::minus) {
-      prefix[0] = specs.sign == sign::plus ? '+' : ' ';
-      ++prefix_size;
-    }
-  }
+  struct next_state {
+    std::string::const_iterator group;
+    int pos;
+  };
+  next_state initial_state() const { return {sep_.grouping.begin(), 0}; }
 
-  void on_dec() {
-    auto num_digits = count_digits(abs_value);
-    out = write_int(
-        out, num_digits, get_prefix(), specs, [this, num_digits](iterator it) {
-          return format_decimal<Char>(it, abs_value, num_digits).end;
-        });
+  // Returns the next digit group separator position.
+  int next(next_state& state) const {
+    if (!sep_.thousands_sep) return max_value<int>();
+    if (state.group == sep_.grouping.end())
+      return state.pos += sep_.grouping.back();
+    if (*state.group <= 0 || *state.group == max_value<char>())
+      return max_value<int>();
+    state.pos += *state.group++;
+    return state.pos;
   }
 
-  void on_hex() {
-    if (specs.alt) {
-      prefix[prefix_size++] = '0';
-      prefix[prefix_size++] = specs.type;
-    }
-    int num_digits = count_digits<4>(abs_value);
-    out = write_int(out, num_digits, get_prefix(), specs,
-                    [this, num_digits](iterator it) {
-                      return format_uint<4, Char>(it, abs_value, num_digits,
-                                                  specs.type != 'x');
-                    });
+ public:
+  explicit digit_grouping(locale_ref loc, bool localized = true) {
+    if (localized)
+      sep_ = thousands_sep<Char>(loc);
+    else
+      sep_.thousands_sep = Char();
   }
+  explicit digit_grouping(thousands_sep_result<Char> sep) : sep_(sep) {}
 
-  void on_bin() {
-    if (specs.alt) {
-      prefix[prefix_size++] = '0';
-      prefix[prefix_size++] = static_cast<char>(specs.type);
-    }
-    int num_digits = count_digits<1>(abs_value);
-    out = write_int(out, num_digits, get_prefix(), specs,
-                    [this, num_digits](iterator it) {
-                      return format_uint<1, Char>(it, abs_value, num_digits);
-                    });
+  Char separator() const { return sep_.thousands_sep; }
+
+  int count_separators(int num_digits) const {
+    int count = 0;
+    auto state = initial_state();
+    while (num_digits > next(state)) ++count;
+    return count;
   }
 
-  void on_oct() {
-    int num_digits = count_digits<3>(abs_value);
-    if (specs.alt && specs.precision <= num_digits && abs_value != 0) {
-      // Octal prefix '0' is counted as a digit, so only add it if precision
-      // is not greater than the number of digits.
-      prefix[prefix_size++] = '0';
-    }
-    out = write_int(out, num_digits, get_prefix(), specs,
-                    [this, num_digits](iterator it) {
-                      return format_uint<3, Char>(it, abs_value, num_digits);
-                    });
-  }
-
-  enum { sep_size = 1 };
-
-  void on_num() {
-    std::string groups = grouping<Char>(locale);
-    if (groups.empty()) return on_dec();
-    auto sep = thousands_sep<Char>(locale);
-    if (!sep) return on_dec();
-    int num_digits = count_digits(abs_value);
-    int size = num_digits, n = num_digits;
-    std::string::const_iterator group = groups.cbegin();
-    while (group != groups.cend() && n > *group && *group > 0 &&
-           *group != max_value<char>()) {
-      size += sep_size;
-      n -= *group;
-      ++group;
+  // Applies grouping to digits and write the output to out.
+  template <typename Out, typename C>
+  Out apply(Out out, basic_string_view<C> digits) const {
+    auto num_digits = static_cast<int>(digits.size());
+    auto separators = basic_memory_buffer<int>();
+    separators.push_back(0);
+    auto state = initial_state();
+    while (int i = next(state)) {
+      if (i >= num_digits) break;
+      separators.push_back(i);
     }
-    if (group == groups.cend()) size += sep_size * ((n - 1) / groups.back());
-    char digits[40];
-    format_decimal(digits, abs_value, num_digits);
-    basic_memory_buffer<Char> buffer;
-    size += static_cast<int>(prefix_size);
-    const auto usize = to_unsigned(size);
-    buffer.resize(usize);
-    basic_string_view<Char> s(&sep, sep_size);
-    // Index of a decimal digit with the least significant digit having index 0.
-    int digit_index = 0;
-    group = groups.cbegin();
-    auto p = buffer.data() + size - 1;
-    for (int i = num_digits - 1; i > 0; --i) {
-      *p-- = static_cast<Char>(digits[i]);
-      if (*group <= 0 || ++digit_index % *group != 0 ||
-          *group == max_value<char>())
-        continue;
-      if (group + 1 != groups.cend()) {
-        digit_index = 0;
-        ++group;
+    for (int i = 0, sep_index = static_cast<int>(separators.size() - 1);
+         i < num_digits; ++i) {
+      if (num_digits - i == separators[sep_index]) {
+        *out++ = separator();
+        --sep_index;
       }
-      std::uninitialized_copy(s.data(), s.data() + s.size(),
-                              make_checked(p, s.size()));
-      p -= s.size();
+      *out++ = static_cast<Char>(digits[to_unsigned(i)]);
     }
-    *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); });
+    return out;
   }
+};
 
-  void on_chr() { *out++ = static_cast<Char>(abs_value); }
+template <typename OutputIt, typename UInt, typename Char>
+auto write_int_localized(OutputIt out, UInt value, unsigned prefix,
+                         const basic_format_specs<Char>& specs,
+                         const digit_grouping<Char>& grouping) -> OutputIt {
+  static_assert(std::is_same<uint64_or_128_t<UInt>, UInt>::value, "");
+  int num_digits = count_digits(value);
+  char digits[40];
+  format_decimal(digits, value, num_digits);
+  unsigned size = to_unsigned((prefix != 0 ? 1 : 0) + num_digits +
+                              grouping.count_separators(num_digits));
+  return write_padded<align::right>(
+      out, specs, size, size, [&](reserve_iterator<OutputIt> it) {
+        if (prefix != 0) *it++ = static_cast<Char>(prefix);
+        return grouping.apply(it, string_view(digits, to_unsigned(num_digits)));
+      });
+}
+
+template <typename OutputIt, typename UInt, typename Char>
+auto write_int_localized(OutputIt& out, UInt value, unsigned prefix,
+                         const basic_format_specs<Char>& specs, locale_ref loc)
+    -> bool {
+  auto grouping = digit_grouping<Char>(loc);
+  out = write_int_localized(out, value, prefix, specs, grouping);
+  return true;
+}
+
+FMT_CONSTEXPR inline void prefix_append(unsigned& prefix, unsigned value) {
+  prefix |= prefix != 0 ? value << 8 : value;
+  prefix += (1u + (value > 0xff ? 1 : 0)) << 24;
+}
+
+template <typename UInt> struct write_int_arg {
+  UInt abs_value;
+  unsigned prefix;
+};
 
-  FMT_NORETURN void on_error() {
-    FMT_THROW(format_error("invalid type specifier"));
+template <typename T>
+FMT_CONSTEXPR auto make_write_int_arg(T value, sign_t sign)
+    -> write_int_arg<uint32_or_64_or_128_t<T>> {
+  auto prefix = 0u;
+  auto abs_value = static_cast<uint32_or_64_or_128_t<T>>(value);
+  if (is_negative(value)) {
+    prefix = 0x01000000 | '-';
+    abs_value = 0 - abs_value;
+  } else {
+    constexpr const unsigned prefixes[4] = {0, 0, 0x1000000u | '+',
+                                            0x1000000u | ' '};
+    prefix = prefixes[sign];
   }
-};
+  return {abs_value, prefix};
+}
+
+template <typename Char, typename OutputIt, typename T>
+FMT_CONSTEXPR FMT_INLINE auto write_int(OutputIt out, write_int_arg<T> arg,
+                                        const basic_format_specs<Char>& specs,
+                                        locale_ref loc) -> OutputIt {
+  static_assert(std::is_same<T, uint32_or_64_or_128_t<T>>::value, "");
+  auto abs_value = arg.abs_value;
+  auto prefix = arg.prefix;
+  switch (specs.type) {
+  case presentation_type::none:
+  case presentation_type::dec: {
+    if (specs.localized &&
+        write_int_localized(out, static_cast<uint64_or_128_t<T>>(abs_value),
+                            prefix, specs, loc)) {
+      return out;
+    }
+    auto num_digits = count_digits(abs_value);
+    return write_int(
+        out, num_digits, prefix, specs, [=](reserve_iterator<OutputIt> it) {
+          return format_decimal<Char>(it, abs_value, num_digits).end;
+        });
+  }
+  case presentation_type::hex_lower:
+  case presentation_type::hex_upper: {
+    bool upper = specs.type == presentation_type::hex_upper;
+    if (specs.alt)
+      prefix_append(prefix, unsigned(upper ? 'X' : 'x') << 8 | '0');
+    int num_digits = count_digits<4>(abs_value);
+    return write_int(
+        out, num_digits, prefix, specs, [=](reserve_iterator<OutputIt> it) {
+          return format_uint<4, Char>(it, abs_value, num_digits, upper);
+        });
+  }
+  case presentation_type::bin_lower:
+  case presentation_type::bin_upper: {
+    bool upper = specs.type == presentation_type::bin_upper;
+    if (specs.alt)
+      prefix_append(prefix, unsigned(upper ? 'B' : 'b') << 8 | '0');
+    int num_digits = count_digits<1>(abs_value);
+    return write_int(out, num_digits, prefix, specs,
+                     [=](reserve_iterator<OutputIt> it) {
+                       return format_uint<1, Char>(it, abs_value, num_digits);
+                     });
+  }
+  case presentation_type::oct: {
+    int num_digits = count_digits<3>(abs_value);
+    // Octal prefix '0' is counted as a digit, so only add it if precision
+    // is not greater than the number of digits.
+    if (specs.alt && specs.precision <= num_digits && abs_value != 0)
+      prefix_append(prefix, '0');
+    return write_int(out, num_digits, prefix, specs,
+                     [=](reserve_iterator<OutputIt> it) {
+                       return format_uint<3, Char>(it, abs_value, num_digits);
+                     });
+  }
+  case presentation_type::chr:
+    return write_char(out, static_cast<Char>(abs_value), specs);
+  default:
+    throw_format_error("invalid type specifier");
+  }
+  return out;
+}
+template <typename Char, typename OutputIt, typename T>
+FMT_CONSTEXPR FMT_NOINLINE auto write_int_noinline(
+    OutputIt out, write_int_arg<T> arg, const basic_format_specs<Char>& specs,
+    locale_ref loc) -> OutputIt {
+  return write_int(out, arg, specs, loc);
+}
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_integral<T>::value &&
+                        !std::is_same<T, bool>::value &&
+                        std::is_same<OutputIt, buffer_appender<Char>>::value)>
+FMT_CONSTEXPR FMT_INLINE auto write(OutputIt out, T value,
+                                    const basic_format_specs<Char>& specs,
+                                    locale_ref loc) -> OutputIt {
+  return write_int_noinline(out, make_write_int_arg(value, specs.sign), specs,
+                            loc);
+}
+// An inlined version of write used in format string compilation.
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_integral<T>::value &&
+                        !std::is_same<T, bool>::value &&
+                        !std::is_same<OutputIt, buffer_appender<Char>>::value)>
+FMT_CONSTEXPR FMT_INLINE auto write(OutputIt out, T value,
+                                    const basic_format_specs<Char>& specs,
+                                    locale_ref loc) -> OutputIt {
+  return write_int(out, make_write_int_arg(value, specs.sign), specs, loc);
+}
 
 template <typename Char, typename OutputIt>
-OutputIt write_nonfinite(OutputIt out, bool isinf,
-                         const basic_format_specs<Char>& specs,
-                         const float_specs& fspecs) {
+FMT_CONSTEXPR auto write(OutputIt out, basic_string_view<Char> s,
+                         const basic_format_specs<Char>& specs) -> OutputIt {
+  auto data = s.data();
+  auto size = s.size();
+  if (specs.precision >= 0 && to_unsigned(specs.precision) < size)
+    size = code_point_index(s, to_unsigned(specs.precision));
+  auto width =
+      specs.width != 0 ? compute_width(basic_string_view<Char>(data, size)) : 0;
+  return write_padded(out, specs, size, width,
+                      [=](reserve_iterator<OutputIt> it) {
+                        return copy_str<Char>(data, data + size, it);
+                      });
+}
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out,
+                         basic_string_view<type_identity_t<Char>> s,
+                         const basic_format_specs<Char>& specs, locale_ref)
+    -> OutputIt {
+  check_string_type_spec(specs.type);
+  return write(out, s, specs);
+}
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR auto write(OutputIt out, const Char* s,
+                         const basic_format_specs<Char>& specs, locale_ref)
+    -> OutputIt {
+  return check_cstring_type_spec(specs.type)
+             ? write(out, basic_string_view<Char>(s), specs, {})
+             : write_ptr<Char>(out, to_uintptr(s), &specs);
+}
+
+template <typename Char, typename OutputIt>
+FMT_CONSTEXPR20 auto write_nonfinite(OutputIt out, bool isinf,
+                                     basic_format_specs<Char> specs,
+                                     const float_specs& fspecs) -> OutputIt {
   auto str =
       isinf ? (fspecs.upper ? "INF" : "inf") : (fspecs.upper ? "NAN" : "nan");
   constexpr size_t str_size = 3;
   auto sign = fspecs.sign;
   auto size = str_size + (sign ? 1 : 0);
-  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
-  return write_padded(out, specs, size, [=](iterator it) {
-    if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+  // Replace '0'-padding with space for non-finite values.
+  const bool is_zero_fill =
+      specs.fill.size() == 1 && *specs.fill.data() == static_cast<Char>('0');
+  if (is_zero_fill) specs.fill[0] = static_cast<Char>(' ');
+  return write_padded(out, specs, size, [=](reserve_iterator<OutputIt> it) {
+    if (sign) *it++ = detail::sign<Char>(sign);
     return copy_str<Char>(str, str + str_size, it);
   });
 }
@@ -1738,74 +1704,118 @@ struct big_decimal_fp {
   int exponent;
 };
 
-inline int get_significand_size(const big_decimal_fp& fp) {
+constexpr auto get_significand_size(const big_decimal_fp& fp) -> int {
   return fp.significand_size;
 }
 template <typename T>
-inline int get_significand_size(const dragonbox::decimal_fp<T>& fp) {
+inline auto get_significand_size(const dragonbox::decimal_fp<T>& fp) -> int {
   return count_digits(fp.significand);
 }
 
 template <typename Char, typename OutputIt>
-inline OutputIt write_significand(OutputIt out, const char* significand,
-                                  int& significand_size) {
+constexpr auto write_significand(OutputIt out, const char* significand,
+                                 int significand_size) -> OutputIt {
   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) {
+inline auto write_significand(OutputIt out, UInt significand,
+                              int significand_size) -> OutputIt {
   return format_decimal<Char>(out, significand, significand_size).end;
 }
+template <typename Char, typename OutputIt, typename T, typename Grouping>
+FMT_CONSTEXPR20 auto write_significand(OutputIt out, T significand,
+                                       int significand_size, int exponent,
+                                       const Grouping& grouping) -> OutputIt {
+  if (!grouping.separator()) {
+    out = write_significand<Char>(out, significand, significand_size);
+    return detail::fill_n(out, exponent, static_cast<Char>('0'));
+  }
+  auto buffer = memory_buffer();
+  write_significand<char>(appender(buffer), significand, significand_size);
+  detail::fill_n(appender(buffer), exponent, '0');
+  return grouping.apply(out, string_view(buffer.data(), buffer.size()));
+}
 
 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) {
+inline auto write_significand(Char* out, UInt significand, int significand_size,
+                              int integral_size, Char decimal_point) -> Char* {
   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;
+  out += significand_size + 1;
+  Char* end = out;
+  int floating_size = significand_size - integral_size;
+  for (int i = floating_size / 2; i > 0; --i) {
+    out -= 2;
+    copy2(out, digits2(significand % 100));
+    significand /= 100;
+  }
+  if (floating_size % 2 != 0) {
+    *--out = static_cast<Char>('0' + significand % 10);
+    significand /= 10;
+  }
+  *--out = decimal_point;
+  format_decimal(out - integral_size, significand, integral_size);
   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) {
+inline auto write_significand(OutputIt out, UInt significand,
+                              int significand_size, int integral_size,
+                              Char decimal_point) -> OutputIt {
   // 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);
+  return detail::copy_str_noinline<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);
+FMT_CONSTEXPR auto write_significand(OutputIt out, const char* significand,
+                                     int significand_size, int integral_size,
+                                     Char decimal_point) -> OutputIt {
+  out = detail::copy_str_noinline<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);
+  return detail::copy_str_noinline<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) {
+template <typename OutputIt, typename Char, typename T, typename Grouping>
+FMT_CONSTEXPR20 auto write_significand(OutputIt out, T significand,
+                                       int significand_size, int integral_size,
+                                       Char decimal_point,
+                                       const Grouping& grouping) -> OutputIt {
+  if (!grouping.separator()) {
+    return write_significand(out, significand, significand_size, integral_size,
+                             decimal_point);
+  }
+  auto buffer = basic_memory_buffer<Char>();
+  write_significand(buffer_appender<Char>(buffer), significand,
+                    significand_size, integral_size, decimal_point);
+  grouping.apply(
+      out, basic_string_view<Char>(buffer.data(), to_unsigned(integral_size)));
+  return detail::copy_str_noinline<Char>(buffer.data() + integral_size,
+                                         buffer.end(), out);
+}
+
+template <typename OutputIt, typename DecimalFP, typename Char,
+          typename Grouping = digit_grouping<Char>>
+FMT_CONSTEXPR20 auto do_write_float(OutputIt out, const DecimalFP& fp,
+                                    const basic_format_specs<Char>& specs,
+                                    float_specs fspecs, locale_ref loc)
+    -> OutputIt {
   auto significand = fp.significand;
   int significand_size = get_significand_size(fp);
-  static const Char zero = static_cast<Char>('0');
+  constexpr 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))>;
+  using iterator = reserve_iterator<OutputIt>;
+
+  Char decimal_point =
+      fspecs.locale ? detail::decimal_point<Char>(loc) : static_cast<Char>('.');
 
   int output_exp = fp.exponent + significand_size - 1;
   auto use_exp_format = [=]() {
@@ -1820,7 +1830,8 @@ OutputIt write_float(OutputIt out, const DecimalFP& fp,
   if (use_exp_format()) {
     int num_zeros = 0;
     if (fspecs.showpoint) {
-      num_zeros = (std::max)(fspecs.precision - significand_size, 0);
+      num_zeros = fspecs.precision - significand_size;
+      if (num_zeros < 0) num_zeros = 0;
       size += to_unsigned(num_zeros);
     } else if (significand_size == 1) {
       decimal_point = Char();
@@ -1832,11 +1843,11 @@ OutputIt write_float(OutputIt out, const DecimalFP& fp,
     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]);
+      if (sign) *it++ = detail::sign<Char>(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);
+      if (num_zeros > 0) it = detail::fill_n(it, num_zeros, zero);
       *it++ = static_cast<Char>(exp_char);
       return write_exponent<Char>(output_exp, it);
     };
@@ -1855,25 +1866,29 @@ OutputIt write_float(OutputIt out, const DecimalFP& fp,
 #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);
+      if (num_zeros > 0) size += to_unsigned(num_zeros) + 1;
     }
+    auto grouping = Grouping(loc, fspecs.locale);
+    size += to_unsigned(grouping.count_separators(significand_size));
     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 (sign) *it++ = detail::sign<Char>(sign);
+      it = write_significand<Char>(it, significand, significand_size,
+                                   fp.exponent, grouping);
       if (!fspecs.showpoint) return it;
       *it++ = decimal_point;
-      return num_zeros > 0 ? std::fill_n(it, num_zeros, zero) : it;
+      return num_zeros > 0 ? detail::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);
+    auto grouping = Grouping(loc, fspecs.locale);
+    size += to_unsigned(grouping.count_separators(significand_size));
     return write_padded<align::right>(out, specs, size, [&](iterator it) {
-      if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+      if (sign) *it++ = detail::sign<Char>(sign);
       it = write_significand(it, significand, significand_size, exp,
-                             decimal_point);
-      return num_zeros > 0 ? std::fill_n(it, num_zeros, zero) : it;
+                             decimal_point, grouping);
+      return num_zeros > 0 ? detail::fill_n(it, num_zeros, zero) : it;
     });
   }
   // 1234e-6 -> 0.001234
@@ -1882,37 +1897,109 @@ OutputIt write_float(OutputIt out, const DecimalFP& fp,
       fspecs.precision < num_zeros) {
     num_zeros = fspecs.precision;
   }
-  size += 2 + to_unsigned(num_zeros);
+  bool pointy = num_zeros != 0 || significand_size != 0 || fspecs.showpoint;
+  size += 1 + (pointy ? 1 : 0) + to_unsigned(num_zeros);
   return write_padded<align::right>(out, specs, size, [&](iterator it) {
-    if (sign) *it++ = static_cast<Char>(data::signs[sign]);
+    if (sign) *it++ = detail::sign<Char>(sign);
     *it++ = zero;
-    if (num_zeros == 0 && significand_size == 0 && !fspecs.showpoint) return it;
+    if (!pointy) return it;
     *it++ = decimal_point;
-    it = std::fill_n(it, num_zeros, zero);
+    it = detail::fill_n(it, num_zeros, zero);
     return write_significand<Char>(it, significand, significand_size);
   });
 }
 
+template <typename Char> class fallback_digit_grouping {
+ public:
+  constexpr fallback_digit_grouping(locale_ref, bool) {}
+
+  constexpr Char separator() const { return Char(); }
+
+  constexpr int count_separators(int) const { return 0; }
+
+  template <typename Out, typename C>
+  constexpr Out apply(Out out, basic_string_view<C>) const {
+    return out;
+  }
+};
+
+template <typename OutputIt, typename DecimalFP, typename Char>
+FMT_CONSTEXPR20 auto write_float(OutputIt out, const DecimalFP& fp,
+                                 const basic_format_specs<Char>& specs,
+                                 float_specs fspecs, locale_ref loc)
+    -> OutputIt {
+  if (is_constant_evaluated()) {
+    return do_write_float<OutputIt, DecimalFP, Char,
+                          fallback_digit_grouping<Char>>(out, fp, specs, fspecs,
+                                                         loc);
+  } else {
+    return do_write_float(out, fp, specs, fspecs, loc);
+  }
+}
+
+template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+FMT_CONSTEXPR20 bool isinf(T value) {
+  if (is_constant_evaluated()) {
+#if defined(__cpp_if_constexpr)
+    if constexpr (std::numeric_limits<double>::is_iec559) {
+      auto bits = detail::bit_cast<uint64_t>(static_cast<double>(value));
+      constexpr auto significand_bits =
+          dragonbox::float_info<double>::significand_bits;
+      return (bits & exponent_mask<double>()) &&
+             !(bits & ((uint64_t(1) << significand_bits) - 1));
+    }
+#endif
+  }
+  return std::isinf(value);
+}
+
+template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+FMT_CONSTEXPR20 bool isfinite(T value) {
+  if (is_constant_evaluated()) {
+#if defined(__cpp_if_constexpr)
+    if constexpr (std::numeric_limits<double>::is_iec559) {
+      auto bits = detail::bit_cast<uint64_t>(static_cast<double>(value));
+      return (bits & exponent_mask<double>()) != exponent_mask<double>();
+    }
+#endif
+  }
+  return std::isfinite(value);
+}
+
+template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
+FMT_INLINE FMT_CONSTEXPR bool signbit(T value) {
+  if (is_constant_evaluated()) {
+#ifdef __cpp_if_constexpr
+    if constexpr (std::numeric_limits<double>::is_iec559) {
+      auto bits = detail::bit_cast<uint64_t>(static_cast<double>(value));
+      return (bits & (uint64_t(1) << (num_bits<uint64_t>() - 1))) != 0;
+    }
+#endif
+  }
+  return std::signbit(value);
+}
+
 template <typename Char, typename OutputIt, typename T,
           FMT_ENABLE_IF(std::is_floating_point<T>::value)>
-OutputIt write(OutputIt out, T value, basic_format_specs<Char> specs,
-               locale_ref loc = {}) {
+FMT_CONSTEXPR20 auto write(OutputIt out, T value,
+                           basic_format_specs<Char> specs, locale_ref loc = {})
+    -> OutputIt {
   if (const_check(!is_supported_floating_point(value))) return out;
   float_specs fspecs = parse_float_type_spec(specs);
   fspecs.sign = specs.sign;
-  if (std::signbit(value)) {  // value < 0 is false for NaN so use signbit.
+  if (detail::signbit(value)) {  // value < 0 is false for NaN so use signbit.
     fspecs.sign = sign::minus;
     value = -value;
   } else if (fspecs.sign == sign::minus) {
     fspecs.sign = sign::none;
   }
 
-  if (!std::isfinite(value))
-    return write_nonfinite(out, std::isinf(value), specs, fspecs);
+  if (!detail::isfinite(value))
+    return write_nonfinite(out, detail::isinf(value), specs, fspecs);
 
   if (specs.align == align::numeric && fspecs.sign) {
     auto it = reserve(out, 1);
-    *it++ = static_cast<Char>(data::signs[fspecs.sign]);
+    *it++ = detail::sign<Char>(fspecs.sign);
     out = base_iterator(out, it);
     fspecs.sign = sign::none;
     if (specs.width != 0) --specs.width;
@@ -1920,30 +2007,35 @@ OutputIt write(OutputIt out, T value, basic_format_specs<Char> specs,
 
   memory_buffer buffer;
   if (fspecs.format == float_format::hex) {
-    if (fspecs.sign) buffer.push_back(data::signs[fspecs.sign]);
+    if (fspecs.sign) buffer.push_back(detail::sign<char>(fspecs.sign));
     snprintf_float(promote_float(value), specs.precision, fspecs, buffer);
-    return write_bytes(out, {buffer.data(), buffer.size()}, specs);
+    return write_bytes<align::right>(out, {buffer.data(), buffer.size()},
+                                     specs);
   }
-  int precision = specs.precision >= 0 || !specs.type ? specs.precision : 6;
+  int precision = specs.precision >= 0 || specs.type == presentation_type::none
+                      ? specs.precision
+                      : 6;
   if (fspecs.format == float_format::exp) {
     if (precision == max_value<int>())
-      FMT_THROW(format_error("number is too big"));
+      throw_format_error("number is too big");
     else
       ++precision;
   }
   if (const_check(std::is_same<T, float>())) fspecs.binary32 = true;
-  fspecs.use_grisu = is_fast_float<T>();
+  if (!is_fast_float<T>()) fspecs.fallback = true;
   int exp = format_float(promote_float(value), precision, fspecs, buffer);
   fspecs.precision = precision;
-  Char point =
-      fspecs.locale ? decimal_point<Char>(loc) : static_cast<Char>('.');
   auto fp = big_decimal_fp{buffer.data(), static_cast<int>(buffer.size()), exp};
-  return write_float(out, fp, specs, fspecs, point);
+  return write_float(out, fp, specs, fspecs, loc);
 }
 
 template <typename Char, typename OutputIt, typename T,
           FMT_ENABLE_IF(is_fast_float<T>::value)>
-OutputIt write(OutputIt out, T value) {
+FMT_CONSTEXPR20 auto write(OutputIt out, T value) -> OutputIt {
+  if (is_constant_evaluated()) {
+    return write(out, value, basic_format_specs<Char>());
+  }
+
   if (const_check(!is_supported_floating_point(value))) return out;
 
   using floaty = conditional_t<std::is_same<T, long double>::value, double, T>;
@@ -1951,90 +2043,53 @@ OutputIt write(OutputIt out, T value) {
   auto bits = bit_cast<uint>(value);
 
   auto fspecs = float_specs();
-  auto sign_bit = bits & (uint(1) << (num_bits<uint>() - 1));
-  if (sign_bit != 0) {
+  if (detail::signbit(value)) {
     fspecs.sign = sign::minus;
     value = -value;
   }
 
-  static const auto specs = basic_format_specs<Char>();
+  constexpr auto specs = basic_format_specs<Char>();
   uint mask = exponent_mask<floaty>();
   if ((bits & mask) == mask)
     return write_nonfinite(out, std::isinf(value), specs, fspecs);
 
   auto dec = dragonbox::to_decimal(static_cast<floaty>(value));
-  return write_float(out, dec, specs, fspecs, static_cast<Char>('.'));
+  return write_float(out, dec, specs, fspecs, {});
 }
 
 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) {
+inline auto write(OutputIt out, T value) -> OutputIt {
   return write(out, value, basic_format_specs<Char>());
 }
 
 template <typename Char, typename OutputIt>
-OutputIt write_char(OutputIt out, Char value,
-                    const basic_format_specs<Char>& specs) {
-  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
-  return write_padded(out, specs, 1, [=](iterator it) {
-    *it++ = value;
-    return it;
-  });
-}
-
-template <typename Char, typename OutputIt, typename UIntPtr>
-OutputIt write_ptr(OutputIt out, UIntPtr value,
-                   const basic_format_specs<Char>* specs) {
-  int num_digits = count_digits<4>(value);
-  auto size = to_unsigned(num_digits) + size_t(2);
-  using iterator = remove_reference_t<decltype(reserve(out, 0))>;
-  auto write = [=](iterator it) {
-    *it++ = static_cast<Char>('0');
-    *it++ = static_cast<Char>('x');
-    return format_uint<4, Char>(it, value, num_digits);
-  };
-  return specs ? write_padded<align::right>(out, *specs, size, write)
-               : base_iterator(out, write(reserve(out, size)));
-}
-
-template <typename T> struct is_integral : std::is_integral<T> {};
-template <> struct is_integral<int128_t> : std::true_type {};
-template <> struct is_integral<uint128_t> : std::true_type {};
-
-template <typename Char, typename OutputIt>
-OutputIt write(OutputIt out, monostate) {
+auto write(OutputIt out, monostate, basic_format_specs<Char> = {},
+           locale_ref = {}) -> OutputIt {
   FMT_ASSERT(false, "");
   return out;
 }
 
-template <typename Char, typename OutputIt,
-          FMT_ENABLE_IF(!std::is_same<Char, char>::value)>
-OutputIt write(OutputIt out, string_view value) {
-  auto it = reserve(out, value.size());
-  it = copy_str<Char>(value.begin(), value.end(), it);
-  return base_iterator(out, it);
-}
-
 template <typename Char, typename OutputIt>
-OutputIt write(OutputIt out, basic_string_view<Char> value) {
+FMT_CONSTEXPR auto write(OutputIt out, basic_string_view<Char> value)
+    -> OutputIt {
   auto it = reserve(out, value.size());
-  it = std::copy(value.begin(), value.end(), it);
+  it = copy_str_noinline<Char>(value.begin(), value.end(), it);
   return base_iterator(out, it);
 }
 
-template <typename Char>
-buffer_appender<Char> write(buffer_appender<Char> out,
-                            basic_string_view<Char> value) {
-  get_container(out).append(value.begin(), value.end());
-  return out;
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(is_string<T>::value)>
+constexpr auto write(OutputIt out, const T& value) -> OutputIt {
+  return write<Char>(out, to_string_view(value));
 }
 
 template <typename Char, typename OutputIt, typename T,
           FMT_ENABLE_IF(is_integral<T>::value &&
                         !std::is_same<T, bool>::value &&
                         !std::is_same<T, Char>::value)>
-OutputIt write(OutputIt out, T value) {
+FMT_CONSTEXPR auto write(OutputIt out, T value) -> OutputIt {
   auto abs_value = static_cast<uint32_or_64_or_128_t<T>>(value);
   bool negative = is_negative(value);
   // Don't do -abs_value since it trips unsigned-integer-overflow sanitizer.
@@ -2052,336 +2107,130 @@ OutputIt write(OutputIt out, T value) {
   return base_iterator(out, it);
 }
 
-template <typename Char, typename OutputIt>
-OutputIt write(OutputIt out, bool value) {
-  return write<Char>(out, string_view(value ? "true" : "false"));
+// FMT_ENABLE_IF() condition separated to workaround an MSVC bug.
+template <
+    typename Char, typename OutputIt, typename T,
+    bool check =
+        std::is_enum<T>::value && !std::is_same<T, Char>::value &&
+        mapped_type_constant<T, basic_format_context<OutputIt, Char>>::value !=
+            type::custom_type,
+    FMT_ENABLE_IF(check)>
+FMT_CONSTEXPR auto write(OutputIt out, T value) -> OutputIt {
+  return write<Char>(
+      out, static_cast<typename std::underlying_type<T>::type>(value));
+}
+
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(std::is_same<T, bool>::value)>
+FMT_CONSTEXPR auto write(OutputIt out, T value,
+                         const basic_format_specs<Char>& specs = {},
+                         locale_ref = {}) -> OutputIt {
+  return specs.type != presentation_type::none &&
+                 specs.type != presentation_type::string
+             ? write(out, value ? 1 : 0, specs, {})
+             : write_bytes(out, value ? "true" : "false", specs);
 }
 
 template <typename Char, typename OutputIt>
-OutputIt write(OutputIt out, Char value) {
+FMT_CONSTEXPR auto write(OutputIt out, Char value) -> OutputIt {
   auto it = reserve(out, 1);
   *it++ = value;
   return base_iterator(out, it);
 }
 
 template <typename Char, typename OutputIt>
-OutputIt write(OutputIt out, const Char* value) {
+FMT_CONSTEXPR_CHAR_TRAITS auto write(OutputIt out, const Char* value)
+    -> OutputIt {
   if (!value) {
-    FMT_THROW(format_error("string pointer is null"));
+    throw_format_error("string pointer is null");
   } else {
-    auto length = std::char_traits<Char>::length(value);
-    out = write(out, basic_string_view<Char>(value, length));
+    out = write(out, basic_string_view<Char>(value));
   }
   return out;
 }
 
-template <typename Char, typename OutputIt>
-OutputIt write(OutputIt out, const void* value) {
-  return write_ptr<Char>(out, to_uintptr(value), nullptr);
+template <typename Char, typename OutputIt, typename T,
+          FMT_ENABLE_IF(std::is_same<T, void>::value)>
+auto write(OutputIt out, const T* value,
+           const basic_format_specs<Char>& specs = {}, locale_ref = {})
+    -> OutputIt {
+  check_pointer_type_spec(specs.type, error_handler());
+  return write_ptr<Char>(out, to_uintptr(value), &specs);
 }
 
-template <typename Char, typename OutputIt, typename T>
-auto write(OutputIt out, const T& value) -> typename std::enable_if<
-    mapped_type_constant<T, basic_format_context<OutputIt, Char>>::value ==
-        type::custom_type,
-    OutputIt>::type {
-  using context_type = basic_format_context<OutputIt, Char>;
+// A write overload that handles implicit conversions.
+template <typename Char, typename OutputIt, typename T,
+          typename Context = basic_format_context<OutputIt, Char>>
+FMT_CONSTEXPR auto write(OutputIt out, const T& value) -> enable_if_t<
+    std::is_class<T>::value && !is_string<T>::value &&
+        !std::is_same<T, Char>::value &&
+        !std::is_same<const T&,
+                      decltype(arg_mapper<Context>().map(value))>::value,
+    OutputIt> {
+  return write<Char>(out, arg_mapper<Context>().map(value));
+}
+
+template <typename Char, typename OutputIt, typename T,
+          typename Context = basic_format_context<OutputIt, Char>>
+FMT_CONSTEXPR auto write(OutputIt out, const T& value)
+    -> enable_if_t<mapped_type_constant<T, Context>::value == type::custom_type,
+                   OutputIt> {
   using formatter_type =
-      conditional_t<has_formatter<T, context_type>::value,
-                    typename context_type::template formatter_type<T>,
+      conditional_t<has_formatter<T, Context>::value,
+                    typename Context::template formatter_type<T>,
                     fallback_formatter<T, Char>>;
-  context_type ctx(out, {}, {});
+  auto ctx = Context(out, {}, {});
   return formatter_type().format(value, ctx);
 }
 
 // An argument visitor that formats the argument and writes it via the output
 // iterator. It's a class and not a generic lambda for compatibility with C++11.
-template <typename OutputIt, typename Char> struct default_arg_formatter {
-  using context = basic_format_context<OutputIt, Char>;
+template <typename Char> struct default_arg_formatter {
+  using iterator = buffer_appender<Char>;
+  using context = buffer_context<Char>;
 
-  OutputIt out;
+  iterator out;
   basic_format_args<context> args;
   locale_ref loc;
 
-  template <typename T> OutputIt operator()(T value) {
+  template <typename T> auto operator()(T value) -> iterator {
     return write<Char>(out, value);
   }
-
-  OutputIt operator()(typename basic_format_arg<context>::handle handle) {
+  auto operator()(typename basic_format_arg<context>::handle h) -> iterator {
     basic_format_parse_context<Char> parse_ctx({});
-    basic_format_context<OutputIt, Char> format_ctx(out, args, loc);
-    handle.format(parse_ctx, format_ctx);
+    context format_ctx(out, args, loc);
+    h.format(parse_ctx, format_ctx);
     return format_ctx.out();
   }
 };
 
-template <typename OutputIt, typename Char,
-          typename ErrorHandler = error_handler>
-class arg_formatter_base {
- public:
-  using iterator = OutputIt;
-  using char_type = Char;
-  using format_specs = basic_format_specs<Char>;
+template <typename Char> struct arg_formatter {
+  using iterator = buffer_appender<Char>;
+  using context = buffer_context<Char>;
 
- private:
-  iterator out_;
-  locale_ref locale_;
-  format_specs* specs_;
+  iterator out;
+  const basic_format_specs<Char>& specs;
+  locale_ref locale;
 
-  // Attempts to reserve space for n extra characters in the output range.
-  // Returns a pointer to the reserved range or a reference to out_.
-  auto reserve(size_t n) -> decltype(detail::reserve(out_, n)) {
-    return detail::reserve(out_, n);
+  template <typename T>
+  FMT_CONSTEXPR FMT_INLINE auto operator()(T value) -> iterator {
+    return detail::write(out, value, specs, locale);
   }
-
-  using reserve_iterator = remove_reference_t<decltype(
-      detail::reserve(std::declval<iterator&>(), 0))>;
-
-  template <typename T> void write_int(T value, const format_specs& spec) {
-    using uint_type = uint32_or_64_or_128_t<T>;
-    int_writer<iterator, Char, uint_type> w(out_, locale_, value, spec);
-    handle_int_type_spec(spec.type, w);
-    out_ = w.out;
-  }
-
-  void write(char value) {
-    auto&& it = reserve(1);
-    *it++ = value;
-  }
-
-  template <typename Ch, FMT_ENABLE_IF(std::is_same<Ch, Char>::value)>
-  void write(Ch value) {
-    out_ = detail::write<Char>(out_, value);
-  }
-
-  void write(string_view value) {
-    auto&& it = reserve(value.size());
-    it = copy_str<Char>(value.begin(), value.end(), it);
-  }
-  void write(wstring_view value) {
-    static_assert(std::is_same<Char, wchar_t>::value, "");
-    auto&& it = reserve(value.size());
-    it = std::copy(value.begin(), value.end(), it);
-  }
-
-  template <typename Ch>
-  void write(const Ch* s, size_t size, const format_specs& specs) {
-    auto width = specs.width != 0
-                     ? count_code_points(basic_string_view<Ch>(s, size))
-                     : 0;
-    out_ = write_padded(out_, specs, size, width, [=](reserve_iterator it) {
-      return copy_str<Char>(s, s + size, it);
-    });
-  }
-
-  template <typename Ch>
-  void write(basic_string_view<Ch> s, const format_specs& specs = {}) {
-    out_ = detail::write(out_, s, specs);
-  }
-
-  void write_pointer(const void* p) {
-    out_ = write_ptr<char_type>(out_, to_uintptr(p), specs_);
-  }
-
-  struct char_spec_handler : ErrorHandler {
-    arg_formatter_base& formatter;
-    Char value;
-
-    char_spec_handler(arg_formatter_base& f, Char val)
-        : formatter(f), value(val) {}
-
-    void on_int() {
-      // char is only formatted as int if there are specs.
-      formatter.write_int(static_cast<int>(value), *formatter.specs_);
-    }
-    void on_char() {
-      if (formatter.specs_)
-        formatter.out_ = write_char(formatter.out_, value, *formatter.specs_);
-      else
-        formatter.write(value);
-    }
-  };
-
-  struct cstring_spec_handler : error_handler {
-    arg_formatter_base& formatter;
-    const Char* value;
-
-    cstring_spec_handler(arg_formatter_base& f, const Char* val)
-        : formatter(f), value(val) {}
-
-    void on_string() { formatter.write(value); }
-    void on_pointer() { formatter.write_pointer(value); }
-  };
-
- protected:
-  iterator out() { return out_; }
-  format_specs* specs() { return specs_; }
-
-  void write(bool value) {
-    if (specs_)
-      write(string_view(value ? "true" : "false"), *specs_);
-    else
-      out_ = detail::write<Char>(out_, value);
-  }
-
-  void write(const Char* value) {
-    if (!value) {
-      FMT_THROW(format_error("string pointer is null"));
-    } else {
-      auto length = std::char_traits<char_type>::length(value);
-      basic_string_view<char_type> sv(value, length);
-      specs_ ? write(sv, *specs_) : write(sv);
-    }
-  }
-
- public:
-  arg_formatter_base(OutputIt out, format_specs* s, locale_ref loc)
-      : out_(out), locale_(loc), specs_(s) {}
-
-  iterator operator()(monostate) {
-    FMT_ASSERT(false, "invalid argument type");
-    return out_;
-  }
-
-  template <typename T, FMT_ENABLE_IF(is_integral<T>::value)>
-  FMT_INLINE iterator operator()(T value) {
-    if (specs_)
-      write_int(value, *specs_);
-    else
-      out_ = detail::write<Char>(out_, value);
-    return out_;
-  }
-
-  iterator operator()(Char value) {
-    handle_char_specs(specs_,
-                      char_spec_handler(*this, static_cast<Char>(value)));
-    return out_;
-  }
-
-  iterator operator()(bool value) {
-    if (specs_ && specs_->type) return (*this)(value ? 1 : 0);
-    write(value != 0);
-    return out_;
-  }
-
-  template <typename T, FMT_ENABLE_IF(std::is_floating_point<T>::value)>
-  iterator operator()(T value) {
-    auto specs = specs_ ? *specs_ : format_specs();
-    if (const_check(is_supported_floating_point(value)))
-      out_ = detail::write(out_, value, specs, locale_);
-    else
-      FMT_ASSERT(false, "unsupported float argument type");
-    return out_;
-  }
-
-  iterator operator()(const Char* value) {
-    if (!specs_) return write(value), out_;
-    handle_cstring_type_spec(specs_->type, cstring_spec_handler(*this, value));
-    return out_;
-  }
-
-  iterator operator()(basic_string_view<Char> value) {
-    if (specs_) {
-      check_string_type_spec(specs_->type, error_handler());
-      write(value, *specs_);
-    } else {
-      write(value);
-    }
-    return out_;
-  }
-
-  iterator operator()(const void* value) {
-    if (specs_) check_pointer_type_spec(specs_->type, error_handler());
-    write_pointer(value);
-    return out_;
+  auto operator()(typename basic_format_arg<context>::handle) -> iterator {
+    // User-defined types are handled separately because they require access
+    // to the parse context.
+    return out;
   }
 };
 
-/** The default argument formatter. */
-template <typename OutputIt, typename Char>
-class arg_formatter : public arg_formatter_base<OutputIt, Char> {
- private:
-  using char_type = Char;
-  using base = arg_formatter_base<OutputIt, Char>;
-  using context_type = basic_format_context<OutputIt, Char>;
+template <typename Char> struct custom_formatter {
+  basic_format_parse_context<Char>& parse_ctx;
+  buffer_context<Char>& ctx;
 
-  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();
+  void operator()(
+      typename basic_format_arg<buffer_context<Char>>::handle h) const {
+    h.format(parse_ctx, ctx);
   }
-};
-
-template <typename Char> FMT_CONSTEXPR bool is_name_start(Char c) {
-  return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || '_' == c;
-}
-
-// Parses the range [begin, end) as an unsigned integer. This function assumes
-// that the range is non-empty and the first character is a digit.
-template <typename Char, typename ErrorHandler>
-FMT_CONSTEXPR int parse_nonnegative_int(const Char*& begin, const Char* end,
-                                        ErrorHandler&& eh) {
-  FMT_ASSERT(begin != end && '0' <= *begin && *begin <= '9', "");
-  unsigned value = 0;
-  // Convert to unsigned to prevent a warning.
-  constexpr unsigned max_int = max_value<int>();
-  unsigned big = max_int / 10;
-  do {
-    // Check for overflow.
-    if (value > big) {
-      value = max_int + 1;
-      break;
-    }
-    value = value * 10 + unsigned(*begin - '0');
-    ++begin;
-  } while (begin != end && '0' <= *begin && *begin <= '9');
-  if (value > max_int) eh.on_error("number is too big");
-  return static_cast<int>(value);
-}
-
-template <typename Context> class custom_formatter {
- private:
-  using char_type = typename Context::char_type;
-
-  basic_format_parse_context<char_type>& parse_ctx_;
-  Context& ctx_;
-
- public:
-  explicit custom_formatter(basic_format_parse_context<char_type>& parse_ctx,
-                            Context& ctx)
-      : parse_ctx_(parse_ctx), ctx_(ctx) {}
-
-  void operator()(typename basic_format_arg<Context>::handle h) const {
-    h.format(parse_ctx_, ctx_);
-  }
-
   template <typename T> void operator()(T) const {}
 };
 
@@ -2396,13 +2245,13 @@ template <typename ErrorHandler> class width_checker {
   explicit FMT_CONSTEXPR width_checker(ErrorHandler& eh) : handler_(eh) {}
 
   template <typename T, FMT_ENABLE_IF(is_integer<T>::value)>
-  FMT_CONSTEXPR unsigned long long operator()(T value) {
+  FMT_CONSTEXPR auto operator()(T value) -> unsigned long long {
     if (is_negative(value)) handler_.on_error("negative width");
     return static_cast<unsigned long long>(value);
   }
 
   template <typename T, FMT_ENABLE_IF(!is_integer<T>::value)>
-  FMT_CONSTEXPR unsigned long long operator()(T) {
+  FMT_CONSTEXPR auto operator()(T) -> unsigned long long {
     handler_.on_error("width is not integer");
     return 0;
   }
@@ -2416,13 +2265,13 @@ template <typename ErrorHandler> class precision_checker {
   explicit FMT_CONSTEXPR precision_checker(ErrorHandler& eh) : handler_(eh) {}
 
   template <typename T, FMT_ENABLE_IF(is_integer<T>::value)>
-  FMT_CONSTEXPR unsigned long long operator()(T value) {
+  FMT_CONSTEXPR auto operator()(T value) -> unsigned long long {
     if (is_negative(value)) handler_.on_error("negative precision");
     return static_cast<unsigned long long>(value);
   }
 
   template <typename T, FMT_ENABLE_IF(!is_integer<T>::value)>
-  FMT_CONSTEXPR unsigned long long operator()(T) {
+  FMT_CONSTEXPR auto operator()(T) -> unsigned long long {
     handler_.on_error("precision is not integer");
     return 0;
   }
@@ -2431,779 +2280,94 @@ template <typename ErrorHandler> class precision_checker {
   ErrorHandler& handler_;
 };
 
-// A format specifier handler that sets fields in basic_format_specs.
-template <typename Char> class specs_setter {
- public:
-  explicit FMT_CONSTEXPR specs_setter(basic_format_specs<Char>& specs)
-      : specs_(specs) {}
-
-  FMT_CONSTEXPR specs_setter(const specs_setter& other)
-      : specs_(other.specs_) {}
-
-  FMT_CONSTEXPR void on_align(align_t align) { specs_.align = align; }
-  FMT_CONSTEXPR void on_fill(basic_string_view<Char> fill) {
-    specs_.fill = fill;
-  }
-  FMT_CONSTEXPR void on_plus() { specs_.sign = sign::plus; }
-  FMT_CONSTEXPR void on_minus() { specs_.sign = sign::minus; }
-  FMT_CONSTEXPR void on_space() { specs_.sign = sign::space; }
-  FMT_CONSTEXPR void on_hash() { specs_.alt = true; }
-
-  FMT_CONSTEXPR void on_zero() {
-    specs_.align = align::numeric;
-    specs_.fill[0] = Char('0');
-  }
-
-  FMT_CONSTEXPR void on_width(int width) { specs_.width = width; }
-  FMT_CONSTEXPR void on_precision(int precision) {
-    specs_.precision = precision;
-  }
-  FMT_CONSTEXPR void end_precision() {}
-
-  FMT_CONSTEXPR void on_type(Char type) {
-    specs_.type = static_cast<char>(type);
-  }
-
- protected:
-  basic_format_specs<Char>& specs_;
-};
-
-template <typename ErrorHandler> class numeric_specs_checker {
- public:
-  FMT_CONSTEXPR numeric_specs_checker(ErrorHandler& eh, detail::type arg_type)
-      : error_handler_(eh), arg_type_(arg_type) {}
-
-  FMT_CONSTEXPR void require_numeric_argument() {
-    if (!is_arithmetic_type(arg_type_))
-      error_handler_.on_error("format specifier requires numeric argument");
-  }
-
-  FMT_CONSTEXPR void check_sign() {
-    require_numeric_argument();
-    if (is_integral_type(arg_type_) && arg_type_ != type::int_type &&
-        arg_type_ != type::long_long_type && arg_type_ != type::char_type) {
-      error_handler_.on_error("format specifier requires signed argument");
-    }
-  }
-
-  FMT_CONSTEXPR void check_precision() {
-    if (is_integral_type(arg_type_) || arg_type_ == type::pointer_type)
-      error_handler_.on_error("precision not allowed for this argument type");
-  }
-
- private:
-  ErrorHandler& error_handler_;
-  detail::type arg_type_;
-};
-
-// A format specifier handler that checks if specifiers are consistent with the
-// argument type.
-template <typename Handler> class specs_checker : public Handler {
- private:
-  numeric_specs_checker<Handler> checker_;
-
-  // Suppress an MSVC warning about using this in initializer list.
-  FMT_CONSTEXPR Handler& error_handler() { return *this; }
-
- public:
-  FMT_CONSTEXPR specs_checker(const Handler& handler, detail::type arg_type)
-      : Handler(handler), checker_(error_handler(), arg_type) {}
-
-  FMT_CONSTEXPR specs_checker(const specs_checker& other)
-      : Handler(other), checker_(error_handler(), other.arg_type_) {}
-
-  FMT_CONSTEXPR void on_align(align_t align) {
-    if (align == align::numeric) checker_.require_numeric_argument();
-    Handler::on_align(align);
-  }
-
-  FMT_CONSTEXPR void on_plus() {
-    checker_.check_sign();
-    Handler::on_plus();
-  }
-
-  FMT_CONSTEXPR void on_minus() {
-    checker_.check_sign();
-    Handler::on_minus();
-  }
-
-  FMT_CONSTEXPR void on_space() {
-    checker_.check_sign();
-    Handler::on_space();
-  }
-
-  FMT_CONSTEXPR void on_hash() {
-    checker_.require_numeric_argument();
-    Handler::on_hash();
-  }
-
-  FMT_CONSTEXPR void on_zero() {
-    checker_.require_numeric_argument();
-    Handler::on_zero();
-  }
-
-  FMT_CONSTEXPR void end_precision() { checker_.check_precision(); }
-};
-
 template <template <typename> class Handler, typename FormatArg,
           typename ErrorHandler>
-FMT_CONSTEXPR int get_dynamic_spec(FormatArg arg, ErrorHandler eh) {
+FMT_CONSTEXPR auto get_dynamic_spec(FormatArg arg, ErrorHandler eh) -> int {
   unsigned long long value = visit_format_arg(Handler<ErrorHandler>(eh), arg);
   if (value > to_unsigned(max_value<int>())) eh.on_error("number is too big");
   return static_cast<int>(value);
 }
 
-struct auto_id {};
-
 template <typename Context, typename ID>
-FMT_CONSTEXPR typename Context::format_arg get_arg(Context& ctx, ID id) {
+FMT_CONSTEXPR auto get_arg(Context& ctx, ID id) ->
+    typename Context::format_arg {
   auto arg = ctx.arg(id);
   if (!arg) ctx.on_error("argument not found");
   return arg;
 }
 
 // The standard format specifier handler with checking.
-template <typename ParseContext, typename Context>
-class specs_handler : public specs_setter<typename Context::char_type> {
- public:
-  using char_type = typename Context::char_type;
-
-  FMT_CONSTEXPR specs_handler(basic_format_specs<char_type>& specs,
-                              ParseContext& parse_ctx, Context& ctx)
-      : specs_setter<char_type>(specs),
-        parse_context_(parse_ctx),
-        context_(ctx) {}
-
-  template <typename Id> FMT_CONSTEXPR void on_dynamic_width(Id arg_id) {
-    this->specs_.width = get_dynamic_spec<width_checker>(
-        get_arg(arg_id), context_.error_handler());
-  }
-
-  template <typename Id> FMT_CONSTEXPR void on_dynamic_precision(Id arg_id) {
-    this->specs_.precision = get_dynamic_spec<precision_checker>(
-        get_arg(arg_id), context_.error_handler());
-  }
-
-  void on_error(const char* message) { context_.on_error(message); }
-
+template <typename Char> class specs_handler : public specs_setter<Char> {
  private:
+  basic_format_parse_context<Char>& parse_context_;
+  buffer_context<Char>& context_;
+
   // This is only needed for compatibility with gcc 4.4.
-  using format_arg = typename Context::format_arg;
+  using format_arg = basic_format_arg<buffer_context<Char>>;
 
-  FMT_CONSTEXPR format_arg get_arg(auto_id) {
+  FMT_CONSTEXPR auto get_arg(auto_id) -> format_arg {
     return detail::get_arg(context_, parse_context_.next_arg_id());
   }
 
-  FMT_CONSTEXPR format_arg get_arg(int arg_id) {
+  FMT_CONSTEXPR auto get_arg(int arg_id) -> format_arg {
     parse_context_.check_arg_id(arg_id);
     return detail::get_arg(context_, arg_id);
   }
 
-  FMT_CONSTEXPR format_arg get_arg(basic_string_view<char_type> arg_id) {
+  FMT_CONSTEXPR auto get_arg(basic_string_view<Char> arg_id) -> format_arg {
     parse_context_.check_arg_id(arg_id);
     return detail::get_arg(context_, arg_id);
   }
 
-  ParseContext& parse_context_;
-  Context& context_;
-};
-
-enum class arg_id_kind { none, index, name };
-
-// An argument reference.
-template <typename Char> struct arg_ref {
-  FMT_CONSTEXPR arg_ref() : kind(arg_id_kind::none), val() {}
-
-  FMT_CONSTEXPR explicit arg_ref(int index)
-      : kind(arg_id_kind::index), val(index) {}
-  FMT_CONSTEXPR explicit arg_ref(basic_string_view<Char> name)
-      : kind(arg_id_kind::name), val(name) {}
-
-  FMT_CONSTEXPR arg_ref& operator=(int idx) {
-    kind = arg_id_kind::index;
-    val.index = idx;
-    return *this;
-  }
-
-  arg_id_kind kind;
-  union value {
-    FMT_CONSTEXPR value(int id = 0) : index{id} {}
-    FMT_CONSTEXPR value(basic_string_view<Char> n) : name(n) {}
-
-    int index;
-    basic_string_view<Char> name;
-  } val;
-};
-
-// Format specifiers with width and precision resolved at formatting rather
-// than parsing time to allow re-using the same parsed specifiers with
-// different sets of arguments (precompilation of format strings).
-template <typename Char>
-struct dynamic_format_specs : basic_format_specs<Char> {
-  arg_ref<Char> width_ref;
-  arg_ref<Char> precision_ref;
-};
-
-// Format spec handler that saves references to arguments representing dynamic
-// width and precision to be resolved at formatting time.
-template <typename ParseContext>
-class dynamic_specs_handler
-    : public specs_setter<typename ParseContext::char_type> {
  public:
-  using char_type = typename ParseContext::char_type;
-
-  FMT_CONSTEXPR dynamic_specs_handler(dynamic_format_specs<char_type>& specs,
-                                      ParseContext& ctx)
-      : specs_setter<char_type>(specs), specs_(specs), context_(ctx) {}
-
-  FMT_CONSTEXPR dynamic_specs_handler(const dynamic_specs_handler& other)
-      : specs_setter<char_type>(other),
-        specs_(other.specs_),
-        context_(other.context_) {}
+  FMT_CONSTEXPR specs_handler(basic_format_specs<Char>& specs,
+                              basic_format_parse_context<Char>& parse_ctx,
+                              buffer_context<Char>& ctx)
+      : specs_setter<Char>(specs), parse_context_(parse_ctx), context_(ctx) {}
 
   template <typename Id> FMT_CONSTEXPR void on_dynamic_width(Id arg_id) {
-    specs_.width_ref = make_arg_ref(arg_id);
+    this->specs_.width = get_dynamic_spec<width_checker>(
+        get_arg(arg_id), context_.error_handler());
   }
 
   template <typename Id> FMT_CONSTEXPR void on_dynamic_precision(Id arg_id) {
-    specs_.precision_ref = make_arg_ref(arg_id);
-  }
-
-  FMT_CONSTEXPR void on_error(const char* message) {
-    context_.on_error(message);
-  }
-
- private:
-  using arg_ref_type = arg_ref<char_type>;
-
-  FMT_CONSTEXPR arg_ref_type make_arg_ref(int arg_id) {
-    context_.check_arg_id(arg_id);
-    return arg_ref_type(arg_id);
-  }
-
-  FMT_CONSTEXPR arg_ref_type make_arg_ref(auto_id) {
-    return arg_ref_type(context_.next_arg_id());
-  }
-
-  FMT_CONSTEXPR arg_ref_type make_arg_ref(basic_string_view<char_type> arg_id) {
-    context_.check_arg_id(arg_id);
-    basic_string_view<char_type> format_str(
-        context_.begin(), to_unsigned(context_.end() - context_.begin()));
-    return arg_ref_type(arg_id);
-  }
-
-  dynamic_format_specs<char_type>& specs_;
-  ParseContext& context_;
-};
-
-template <typename Char, typename IDHandler>
-FMT_CONSTEXPR const Char* parse_arg_id(const Char* begin, const Char* end,
-                                       IDHandler&& handler) {
-  FMT_ASSERT(begin != end, "");
-  Char c = *begin;
-  if (c == '}' || c == ':') {
-    handler();
-    return begin;
-  }
-  if (c >= '0' && c <= '9') {
-    int index = 0;
-    if (c != '0')
-      index = parse_nonnegative_int(begin, end, handler);
-    else
-      ++begin;
-    if (begin == end || (*begin != '}' && *begin != ':'))
-      handler.on_error("invalid format string");
-    else
-      handler(index);
-    return begin;
-  }
-  if (!is_name_start(c)) {
-    handler.on_error("invalid format string");
-    return begin;
-  }
-  auto it = begin;
-  do {
-    ++it;
-  } while (it != end && (is_name_start(c = *it) || ('0' <= c && c <= '9')));
-  handler(basic_string_view<Char>(begin, to_unsigned(it - begin)));
-  return it;
-}
-
-// Adapts SpecHandler to IDHandler API for dynamic width.
-template <typename SpecHandler, typename Char> struct width_adapter {
-  explicit FMT_CONSTEXPR width_adapter(SpecHandler& h) : handler(h) {}
-
-  FMT_CONSTEXPR void operator()() { handler.on_dynamic_width(auto_id()); }
-  FMT_CONSTEXPR void operator()(int id) { handler.on_dynamic_width(id); }
-  FMT_CONSTEXPR void operator()(basic_string_view<Char> id) {
-    handler.on_dynamic_width(id);
-  }
-
-  FMT_CONSTEXPR void on_error(const char* message) {
-    handler.on_error(message);
-  }
-
-  SpecHandler& handler;
-};
-
-// Adapts SpecHandler to IDHandler API for dynamic precision.
-template <typename SpecHandler, typename Char> struct precision_adapter {
-  explicit FMT_CONSTEXPR precision_adapter(SpecHandler& h) : handler(h) {}
-
-  FMT_CONSTEXPR void operator()() { handler.on_dynamic_precision(auto_id()); }
-  FMT_CONSTEXPR void operator()(int id) { handler.on_dynamic_precision(id); }
-  FMT_CONSTEXPR void operator()(basic_string_view<Char> id) {
-    handler.on_dynamic_precision(id);
-  }
-
-  FMT_CONSTEXPR void on_error(const char* message) {
-    handler.on_error(message);
+    this->specs_.precision = get_dynamic_spec<precision_checker>(
+        get_arg(arg_id), context_.error_handler());
   }
 
-  SpecHandler& handler;
+  void on_error(const char* message) { context_.on_error(message); }
 };
 
-template <typename Char>
-FMT_CONSTEXPR int code_point_length(const Char* begin) {
-  if (const_check(sizeof(Char) != 1)) return 1;
-  constexpr char lengths[] = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
-                              0, 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 2, 3, 3, 4, 0};
-  int len = lengths[static_cast<unsigned char>(*begin) >> 3];
-
-  // Compute the pointer to the next character early so that the next
-  // iteration can start working on the next character. Neither Clang
-  // nor GCC figure out this reordering on their own.
-  return len + !len;
-}
-
-template <typename Char> constexpr bool is_ascii_letter(Char c) {
-  return (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z');
-}
-
-// Converts a character to ASCII. Returns a number > 127 on conversion failure.
-template <typename Char, FMT_ENABLE_IF(std::is_integral<Char>::value)>
-constexpr Char to_ascii(Char value) {
-  return value;
-}
-template <typename Char, FMT_ENABLE_IF(std::is_enum<Char>::value)>
-constexpr typename std::underlying_type<Char>::type to_ascii(Char value) {
-  return value;
-}
-
-// Parses fill and alignment.
-template <typename Char, typename Handler>
-FMT_CONSTEXPR const Char* parse_align(const Char* begin, const Char* end,
-                                      Handler&& handler) {
-  FMT_ASSERT(begin != end, "");
-  auto align = align::none;
-  auto p = begin + code_point_length(begin);
-  if (p >= end) p = begin;
-  for (;;) {
-    switch (to_ascii(*p)) {
-    case '<':
-      align = align::left;
-      break;
-    case '>':
-      align = align::right;
-      break;
-#if FMT_DEPRECATED_NUMERIC_ALIGN
-    case '=':
-      align = align::numeric;
-      break;
-#endif
-    case '^':
-      align = align::center;
-      break;
-    }
-    if (align != align::none) {
-      if (p != begin) {
-        auto c = *begin;
-        if (c == '{')
-          return handler.on_error("invalid fill character '{'"), begin;
-        handler.on_fill(basic_string_view<Char>(begin, to_unsigned(p - begin)));
-        begin = p + 1;
-      } else
-        ++begin;
-      handler.on_align(align);
-      break;
-    } else if (p == begin) {
-      break;
-    }
-    p = begin;
-  }
-  return begin;
-}
-
-template <typename Char, typename Handler>
-FMT_CONSTEXPR const Char* parse_width(const Char* begin, const Char* end,
-                                      Handler&& handler) {
-  FMT_ASSERT(begin != end, "");
-  if ('0' <= *begin && *begin <= '9') {
-    handler.on_width(parse_nonnegative_int(begin, end, handler));
-  } else if (*begin == '{') {
-    ++begin;
-    if (begin != end)
-      begin = parse_arg_id(begin, end, width_adapter<Handler, Char>(handler));
-    if (begin == end || *begin != '}')
-      return handler.on_error("invalid format string"), begin;
-    ++begin;
-  }
-  return begin;
-}
-
-template <typename Char, typename Handler>
-FMT_CONSTEXPR const Char* parse_precision(const Char* begin, const Char* end,
-                                          Handler&& handler) {
-  ++begin;
-  auto c = begin != end ? *begin : Char();
-  if ('0' <= c && c <= '9') {
-    handler.on_precision(parse_nonnegative_int(begin, end, handler));
-  } else if (c == '{') {
-    ++begin;
-    if (begin != end) {
-      begin =
-          parse_arg_id(begin, end, precision_adapter<Handler, Char>(handler));
-    }
-    if (begin == end || *begin++ != '}')
-      return handler.on_error("invalid format string"), begin;
-  } else {
-    return handler.on_error("missing precision specifier"), begin;
-  }
-  handler.end_precision();
-  return begin;
-}
-
-// Parses standard format specifiers and sends notifications about parsed
-// components to handler.
-template <typename Char, typename SpecHandler>
-FMT_CONSTEXPR const Char* parse_format_specs(const Char* begin, const Char* end,
-                                             SpecHandler&& handler) {
-  if (begin == end) return begin;
-
-  begin = parse_align(begin, end, handler);
-  if (begin == end) return begin;
-
-  // Parse sign.
-  switch (to_ascii(*begin)) {
-  case '+':
-    handler.on_plus();
-    ++begin;
-    break;
-  case '-':
-    handler.on_minus();
-    ++begin;
+template <template <typename> class Handler, typename Context>
+FMT_CONSTEXPR void handle_dynamic_spec(int& value,
+                                       arg_ref<typename Context::char_type> ref,
+                                       Context& ctx) {
+  switch (ref.kind) {
+  case arg_id_kind::none:
     break;
-  case ' ':
-    handler.on_space();
-    ++begin;
+  case arg_id_kind::index:
+    value = detail::get_dynamic_spec<Handler>(ctx.arg(ref.val.index),
+                                              ctx.error_handler());
+    break;
+  case arg_id_kind::name:
+    value = detail::get_dynamic_spec<Handler>(ctx.arg(ref.val.name),
+                                              ctx.error_handler());
     break;
   }
-  if (begin == end) return begin;
-
-  if (*begin == '#') {
-    handler.on_hash();
-    if (++begin == end) return begin;
-  }
-
-  // Parse zero flag.
-  if (*begin == '0') {
-    handler.on_zero();
-    if (++begin == end) return begin;
-  }
-
-  begin = parse_width(begin, end, handler);
-  if (begin == end) return begin;
-
-  // Parse precision.
-  if (*begin == '.') {
-    begin = parse_precision(begin, end, handler);
-  }
-
-  // Parse type.
-  if (begin != end && *begin != '}') handler.on_type(*begin++);
-  return begin;
-}
-
-// Return the result via the out param to workaround gcc bug 77539.
-template <bool IS_CONSTEXPR, typename T, typename Ptr = const T*>
-FMT_CONSTEXPR bool find(Ptr first, Ptr last, T value, Ptr& out) {
-  for (out = first; out != last; ++out) {
-    if (*out == value) return true;
-  }
-  return false;
-}
-
-template <>
-inline bool find<false, char>(const char* first, const char* last, char value,
-                              const char*& out) {
-  out = static_cast<const char*>(
-      std::memchr(first, value, detail::to_unsigned(last - first)));
-  return out != nullptr;
-}
-
-template <typename Handler, typename Char> struct id_adapter {
-  Handler& handler;
-  int arg_id;
-
-  FMT_CONSTEXPR void operator()() { arg_id = handler.on_arg_id(); }
-  FMT_CONSTEXPR void operator()(int id) { arg_id = handler.on_arg_id(id); }
-  FMT_CONSTEXPR void operator()(basic_string_view<Char> id) {
-    arg_id = handler.on_arg_id(id);
-  }
-  FMT_CONSTEXPR void on_error(const char* message) {
-    handler.on_error(message);
-  }
-};
-
-template <typename Char, typename Handler>
-FMT_CONSTEXPR const Char* parse_replacement_field(const Char* begin,
-                                                  const Char* end,
-                                                  Handler&& handler) {
-  ++begin;
-  if (begin == end) return handler.on_error("invalid format string"), end;
-  if (*begin == '}') {
-    handler.on_replacement_field(handler.on_arg_id(), begin);
-  } else if (*begin == '{') {
-    handler.on_text(begin, begin + 1);
-  } else {
-    auto adapter = id_adapter<Handler, Char>{handler, 0};
-    begin = parse_arg_id(begin, end, adapter);
-    Char c = begin != end ? *begin : Char();
-    if (c == '}') {
-      handler.on_replacement_field(adapter.arg_id, begin);
-    } else if (c == ':') {
-      begin = handler.on_format_specs(adapter.arg_id, begin + 1, end);
-      if (begin == end || *begin != '}')
-        return handler.on_error("unknown format specifier"), end;
-    } else {
-      return handler.on_error("missing '}' in format string"), end;
-    }
-  }
-  return begin + 1;
-}
-
-template <bool IS_CONSTEXPR, typename Char, typename Handler>
-FMT_CONSTEXPR_DECL FMT_INLINE void parse_format_string(
-    basic_string_view<Char> format_str, Handler&& handler) {
-  auto begin = format_str.data();
-  auto end = begin + format_str.size();
-  if (end - begin < 32) {
-    // Use a simple loop instead of memchr for small strings.
-    const Char* p = begin;
-    while (p != end) {
-      auto c = *p++;
-      if (c == '{') {
-        handler.on_text(begin, p - 1);
-        begin = p = parse_replacement_field(p - 1, end, handler);
-      } else if (c == '}') {
-        if (p == end || *p != '}')
-          return handler.on_error("unmatched '}' in format string");
-        handler.on_text(begin, p);
-        begin = ++p;
-      }
-    }
-    handler.on_text(begin, end);
-    return;
-  }
-  struct writer {
-    FMT_CONSTEXPR void operator()(const Char* pbegin, const Char* pend) {
-      if (pbegin == pend) return;
-      for (;;) {
-        const Char* p = nullptr;
-        if (!find<IS_CONSTEXPR>(pbegin, pend, '}', p))
-          return handler_.on_text(pbegin, pend);
-        ++p;
-        if (p == pend || *p != '}')
-          return handler_.on_error("unmatched '}' in format string");
-        handler_.on_text(pbegin, p);
-        pbegin = p + 1;
-      }
-    }
-    Handler& handler_;
-  } write{handler};
-  while (begin != end) {
-    // Doing two passes with memchr (one for '{' and another for '}') is up to
-    // 2.5x faster than the naive one-pass implementation on big format strings.
-    const Char* p = begin;
-    if (*begin != '{' && !find<IS_CONSTEXPR>(begin + 1, end, '{', p))
-      return write(begin, end);
-    write(begin, p);
-    begin = parse_replacement_field(p, end, handler);
-  }
-}
-
-template <typename T, typename ParseContext>
-FMT_CONSTEXPR const typename ParseContext::char_type* parse_format_specs(
-    ParseContext& ctx) {
-  using char_type = typename ParseContext::char_type;
-  using context = buffer_context<char_type>;
-  using mapped_type =
-      conditional_t<detail::mapped_type_constant<T, context>::value !=
-                        type::custom_type,
-                    decltype(arg_mapper<context>().map(std::declval<T>())), T>;
-  auto f = conditional_t<has_formatter<mapped_type, context>::value,
-                         formatter<mapped_type, char_type>,
-                         detail::fallback_formatter<T, char_type>>();
-  return f.parse(ctx);
-}
-
-template <typename OutputIt, typename Char, typename Context>
-struct format_handler : detail::error_handler {
-  basic_format_parse_context<Char> parse_context;
-  Context context;
-
-  format_handler(OutputIt out, basic_string_view<Char> str,
-                 basic_format_args<Context> format_args, detail::locale_ref loc)
-      : parse_context(str), context(out, format_args, loc) {}
-
-  void on_text(const Char* begin, const Char* end) {
-    auto size = to_unsigned(end - begin);
-    auto out = context.out();
-    auto&& it = reserve(out, size);
-    it = std::copy_n(begin, size, it);
-    context.advance_to(out);
-  }
-
-  int on_arg_id() { return parse_context.next_arg_id(); }
-  int on_arg_id(int id) { return parse_context.check_arg_id(id), id; }
-  int on_arg_id(basic_string_view<Char> id) {
-    int arg_id = context.arg_id(id);
-    if (arg_id < 0) on_error("argument not found");
-    return arg_id;
-  }
-
-  FMT_INLINE void on_replacement_field(int id, const Char*) {
-    auto arg = get_arg(context, id);
-    context.advance_to(visit_format_arg(
-        default_arg_formatter<OutputIt, Char>{context.out(), context.args(),
-                                              context.locale()},
-        arg));
-  }
-
-  const Char* on_format_specs(int id, const Char* begin, const Char* end) {
-    auto arg = get_arg(context, id);
-    if (arg.type() == type::custom_type) {
-      advance_to(parse_context, begin);
-      visit_format_arg(custom_formatter<Context>(parse_context, context), arg);
-      return parse_context.begin();
-    }
-    auto specs = basic_format_specs<Char>();
-    if (begin + 1 < end && begin[1] == '}' && is_ascii_letter(*begin)) {
-      specs.type = static_cast<char>(*begin++);
-    } else {
-      using parse_context_t = basic_format_parse_context<Char>;
-      specs_checker<specs_handler<parse_context_t, Context>> handler(
-          specs_handler<parse_context_t, Context>(specs, parse_context,
-                                                  context),
-          arg.type());
-      begin = parse_format_specs(begin, end, handler);
-      if (begin == end || *begin != '}')
-        on_error("missing '}' in format string");
-    }
-    context.advance_to(visit_format_arg(
-        arg_formatter<OutputIt, Char>(context, &parse_context, &specs), arg));
-    return begin;
-  }
-};
-
-// A parse context with extra argument id checks. It is only used at compile
-// time because adding checks at runtime would introduce substantial overhead
-// and would be redundant since argument ids are checked when arguments are
-// retrieved anyway.
-template <typename Char, typename ErrorHandler = error_handler>
-class compile_parse_context
-    : public basic_format_parse_context<Char, ErrorHandler> {
- private:
-  int num_args_;
-  using base = basic_format_parse_context<Char, ErrorHandler>;
-
- public:
-  explicit FMT_CONSTEXPR compile_parse_context(
-      basic_string_view<Char> format_str, int num_args = max_value<int>(),
-      ErrorHandler eh = {})
-      : base(format_str, eh), num_args_(num_args) {}
-
-  FMT_CONSTEXPR int next_arg_id() {
-    int id = base::next_arg_id();
-    if (id >= num_args_) this->on_error("argument not found");
-    return id;
-  }
-
-  FMT_CONSTEXPR void check_arg_id(int id) {
-    base::check_arg_id(id);
-    if (id >= num_args_) this->on_error("argument not found");
-  }
-  using base::check_arg_id;
-};
-
-template <typename Char, typename ErrorHandler, typename... Args>
-class format_string_checker {
- public:
-  explicit FMT_CONSTEXPR format_string_checker(
-      basic_string_view<Char> format_str, ErrorHandler eh)
-      : context_(format_str, num_args, eh),
-        parse_funcs_{&parse_format_specs<Args, parse_context_type>...} {}
-
-  FMT_CONSTEXPR void on_text(const Char*, const Char*) {}
-
-  FMT_CONSTEXPR int on_arg_id() { return context_.next_arg_id(); }
-  FMT_CONSTEXPR int on_arg_id(int id) { return context_.check_arg_id(id), id; }
-  FMT_CONSTEXPR int on_arg_id(basic_string_view<Char>) {
-    on_error("compile-time checks don't support named arguments");
-    return 0;
-  }
-
-  FMT_CONSTEXPR void on_replacement_field(int, const Char*) {}
-
-  FMT_CONSTEXPR const Char* on_format_specs(int id, const Char* begin,
-                                            const Char*) {
-    advance_to(context_, begin);
-    return id < num_args ? parse_funcs_[id](context_) : begin;
-  }
-
-  FMT_CONSTEXPR void on_error(const char* message) {
-    context_.on_error(message);
-  }
-
- private:
-  using parse_context_type = compile_parse_context<Char, ErrorHandler>;
-  enum { num_args = sizeof...(Args) };
-
-  // Format specifier parsing function.
-  using parse_func = const Char* (*)(parse_context_type&);
-
-  parse_context_type context_;
-  parse_func parse_funcs_[num_args > 0 ? num_args : 1];
-};
-
-// Converts string literals to basic_string_view.
-template <typename Char, size_t N>
-FMT_CONSTEXPR basic_string_view<Char> compile_string_to_view(
-    const Char (&s)[N]) {
-  // Remove trailing null character if needed. Won't be present if this is used
-  // with raw character array (i.e. not defined as a string).
-  return {s,
-          N - ((std::char_traits<Char>::to_int_type(s[N - 1]) == 0) ? 1 : 0)};
 }
 
-// Converts string_view to basic_string_view.
-template <typename Char>
-FMT_CONSTEXPR basic_string_view<Char> compile_string_to_view(
-    const std_string_view<Char>& s) {
-  return {s.data(), s.size()};
-}
-
-#define FMT_STRING_IMPL(s, base)                                  \
-  [] {                                                            \
-    /* Use a macro-like name to avoid shadowing warnings. */      \
-    struct FMT_COMPILE_STRING : base {                            \
-      using char_type = fmt::remove_cvref_t<decltype(s[0])>;      \
-      FMT_MAYBE_UNUSED FMT_CONSTEXPR                              \
-      operator fmt::basic_string_view<char_type>() const {        \
-        return fmt::detail::compile_string_to_view<char_type>(s); \
-      }                                                           \
-    };                                                            \
-    return FMT_COMPILE_STRING();                                  \
+#define FMT_STRING_IMPL(s, base, explicit)                                 \
+  [] {                                                                     \
+    /* Use the hidden visibility as a workaround for a GCC bug (#1973). */ \
+    /* Use a macro-like name to avoid shadowing warnings. */               \
+    struct FMT_GCC_VISIBILITY_HIDDEN FMT_COMPILE_STRING : base {           \
+      using char_type = fmt::remove_cvref_t<decltype(s[0])>;               \
+      FMT_MAYBE_UNUSED FMT_CONSTEXPR explicit                              \
+      operator fmt::basic_string_view<char_type>() const {                 \
+        return fmt::detail_exported::compile_string_to_view<char_type>(s); \
+      }                                                                    \
+    };                                                                     \
+    return FMT_COMPILE_STRING();                                           \
   }()
 
 /**
@@ -3216,119 +2380,122 @@ FMT_CONSTEXPR basic_string_view<Char> compile_string_to_view(
     std::string s = fmt::format(FMT_STRING("{:d}"), "foo");
   \endrst
  */
-#define FMT_STRING(s) FMT_STRING_IMPL(s, fmt::compile_string)
+#define FMT_STRING(s) FMT_STRING_IMPL(s, fmt::compile_string, )
 
-template <typename... Args, typename S,
-          enable_if_t<(is_compile_string<S>::value), int>>
-void check_format_string(S format_str) {
-  FMT_CONSTEXPR_DECL auto s = to_string_view(format_str);
-  using checker = format_string_checker<typename S::char_type, error_handler,
-                                        remove_cvref_t<Args>...>;
-  FMT_CONSTEXPR_DECL bool invalid_format =
-      (parse_format_string<true>(s, checker(s, {})), true);
-  (void)invalid_format;
-}
+#if FMT_USE_USER_DEFINED_LITERALS
+template <typename Char> struct udl_formatter {
+  basic_string_view<Char> str;
 
-template <template <typename> class Handler, typename Context>
-void handle_dynamic_spec(int& value, arg_ref<typename Context::char_type> ref,
-                         Context& ctx) {
-  switch (ref.kind) {
-  case arg_id_kind::none:
-    break;
-  case arg_id_kind::index:
-    value = detail::get_dynamic_spec<Handler>(ctx.arg(ref.val.index),
-                                              ctx.error_handler());
-    break;
-  case arg_id_kind::name:
-    value = detail::get_dynamic_spec<Handler>(ctx.arg(ref.val.name),
-                                              ctx.error_handler());
-    break;
+  template <typename... T>
+  auto operator()(T&&... args) const -> std::basic_string<Char> {
+    return vformat(str, fmt::make_args_checked<T...>(str, args...));
+  }
+};
+
+#  if FMT_USE_NONTYPE_TEMPLATE_PARAMETERS
+template <typename T, typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct statically_named_arg : view {
+  static constexpr auto name = Str.data;
+
+  const T& value;
+  statically_named_arg(const T& v) : value(v) {}
+};
+
+template <typename T, typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct is_named_arg<statically_named_arg<T, Char, N, Str>> : std::true_type {};
+
+template <typename T, typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct is_statically_named_arg<statically_named_arg<T, Char, N, Str>>
+    : std::true_type {};
+
+template <typename Char, size_t N,
+          fmt::detail_exported::fixed_string<Char, N> Str>
+struct udl_arg {
+  template <typename T> auto operator=(T&& value) const {
+    return statically_named_arg<T, Char, N, Str>(std::forward<T>(value));
+  }
+};
+#  else
+template <typename Char> struct udl_arg {
+  const Char* str;
+
+  template <typename T> auto operator=(T&& value) const -> named_arg<Char, T> {
+    return {str, std::forward<T>(value)};
   }
+};
+#  endif
+#endif  // FMT_USE_USER_DEFINED_LITERALS
+
+template <typename Locale, typename Char>
+auto vformat(const Locale& loc, basic_string_view<Char> format_str,
+             basic_format_args<buffer_context<type_identity_t<Char>>> args)
+    -> std::basic_string<Char> {
+  basic_memory_buffer<Char> buffer;
+  detail::vformat_to(buffer, format_str, args, detail::locale_ref(loc));
+  return {buffer.data(), buffer.size()};
 }
 
-using format_func = void (*)(detail::buffer<char>&, int, string_view);
+using format_func = void (*)(detail::buffer<char>&, int, const char*);
 
 FMT_API void format_error_code(buffer<char>& out, int error_code,
                                string_view message) FMT_NOEXCEPT;
 
 FMT_API void report_error(format_func func, int error_code,
-                          string_view message) FMT_NOEXCEPT;
-}  // namespace detail
+                          const char* message) FMT_NOEXCEPT;
+FMT_END_DETAIL_NAMESPACE
 
-template <typename OutputIt, typename Char>
-using arg_formatter FMT_DEPRECATED_ALIAS =
-    detail::arg_formatter<OutputIt, Char>;
+FMT_API auto vsystem_error(int error_code, string_view format_str,
+                           format_args args) -> std::system_error;
 
 /**
- An error returned by an operating system or a language runtime,
- for example a file opening error.
-*/
-FMT_CLASS_API
-class FMT_API system_error : public std::runtime_error {
- private:
-  void init(int err_code, string_view format_str, format_args args);
-
- protected:
-  int error_code_;
+ \rst
+ Constructs :class:`std::system_error` with a message formatted with
+ ``fmt::format(fmt, args...)``.
+  *error_code* is a system error code as given by ``errno``.
 
-  system_error() : std::runtime_error(""), error_code_(0) {}
+ **Example**::
 
- public:
-  /**
-   \rst
-   Constructs a :class:`fmt::system_error` object with a description
-   formatted with `fmt::format_system_error`. *message* and additional
-   arguments passed into the constructor are formatted similarly to
-   `fmt::format`.
-
-   **Example**::
-
-     // This throws a system_error with the description
-     //   cannot open file 'madeup': No such file or directory
-     // or similar (system message may vary).
-     const char *filename = "madeup";
-     std::FILE *file = std::fopen(filename, "r");
-     if (!file)
-       throw fmt::system_error(errno, "cannot open file '{}'", filename);
-   \endrst
-  */
-  template <typename... Args>
-  system_error(int error_code, string_view message, const Args&... args)
-      : std::runtime_error("") {
-    init(error_code, message, make_format_args(args...));
-  }
-  system_error(const system_error&) = default;
-  system_error& operator=(const system_error&) = default;
-  system_error(system_error&&) = default;
-  system_error& operator=(system_error&&) = default;
-  ~system_error() FMT_NOEXCEPT FMT_OVERRIDE;
-
-  int error_code() const { return error_code_; }
-};
+   // This throws std::system_error with the description
+   //   cannot open file 'madeup': No such file or directory
+   // or similar (system message may vary).
+   const char* filename = "madeup";
+   std::FILE* file = std::fopen(filename, "r");
+   if (!file)
+     throw fmt::system_error(errno, "cannot open file '{}'", filename);
+ \endrst
+*/
+template <typename... T>
+auto system_error(int error_code, format_string<T...> fmt, T&&... args)
+    -> std::system_error {
+  return vsystem_error(error_code, fmt, fmt::make_format_args(args...));
+}
 
 /**
   \rst
-  Formats an error returned by an operating system or a language runtime,
-  for example a file opening error, and writes it to *out* in the following
-  form:
+  Formats an error message for an error returned by an operating system or a
+  language runtime, for example a file opening error, and writes it to *out*.
+  The format is the same as the one used by ``std::system_error(ec, message)``
+  where ``ec`` is ``std::error_code(error_code, std::generic_category()})``.
+  It is implementation-defined but normally looks like:
 
   .. parsed-literal::
      *<message>*: *<system-message>*
 
-  where *<message>* is the passed message and *<system-message>* is
-  the system message corresponding to the error code.
+  where *<message>* is the passed message and *<system-message>* is the system
+  message corresponding to the error code.
   *error_code* is a system error code as given by ``errno``.
-  If *error_code* is not a valid error code such as -1, the system message
-  may look like "Unknown error -1" and is platform-dependent.
   \endrst
  */
 FMT_API void format_system_error(detail::buffer<char>& out, int error_code,
-                                 string_view message) FMT_NOEXCEPT;
+                                 const char* message) FMT_NOEXCEPT;
 
 // Reports a system error without throwing an exception.
 // Can be used to report errors from destructors.
 FMT_API void report_system_error(int error_code,
-                                 string_view message) FMT_NOEXCEPT;
+                                 const char* message) FMT_NOEXCEPT;
 
 /** Fast integer formatter. */
 class format_int {
@@ -3339,12 +2506,12 @@ class format_int {
   mutable char buffer_[buffer_size];
   char* str_;
 
-  template <typename UInt> char* format_unsigned(UInt value) {
+  template <typename UInt> auto format_unsigned(UInt value) -> char* {
     auto n = static_cast<detail::uint32_or_64_or_128_t<UInt>>(value);
     return detail::format_decimal(buffer_, n, buffer_size - 1).begin;
   }
 
-  template <typename Int> char* format_signed(Int value) {
+  template <typename Int> auto format_signed(Int value) -> char* {
     auto abs_value = static_cast<detail::uint32_or_64_or_128_t<Int>>(value);
     bool negative = value < 0;
     if (negative) abs_value = 0 - abs_value;
@@ -3363,7 +2530,7 @@ class format_int {
       : str_(format_unsigned(value)) {}
 
   /** Returns the number of characters written to the output buffer. */
-  size_t size() const {
+  auto size() const -> size_t {
     return detail::to_unsigned(buffer_ - str_ + buffer_size - 1);
   }
 
@@ -3371,13 +2538,13 @@ class format_int {
     Returns a pointer to the output buffer content. No terminating null
     character is appended.
    */
-  const char* data() const { return str_; }
+  auto data() const -> const char* { return str_; }
 
   /**
     Returns a pointer to the output buffer content with terminating null
     character appended.
    */
-  const char* c_str() const {
+  auto c_str() const -> const char* {
     buffer_[buffer_size - 1] = '\0';
     return str_;
   }
@@ -3387,104 +2554,37 @@ class format_int {
     Returns the content of the output buffer as an ``std::string``.
     \endrst
    */
-  std::string str() const { return std::string(str_, size()); }
+  auto str() const -> std::string { return std::string(str_, size()); }
 };
 
-// A formatter specialization for the core types corresponding to detail::type
-// constants.
 template <typename T, typename Char>
-struct formatter<T, Char,
-                 enable_if_t<detail::type_constant<T, Char>::value !=
-                             detail::type::custom_type>> {
-  FMT_CONSTEXPR formatter() = default;
-
-  // Parses format specifiers stopping either at the end of the range or at the
-  // terminating '}'.
-  template <typename ParseContext>
-  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
-    using handler_type = detail::dynamic_specs_handler<ParseContext>;
-    auto type = detail::type_constant<T, Char>::value;
-    detail::specs_checker<handler_type> handler(handler_type(specs_, ctx),
-                                                type);
-    auto it = parse_format_specs(ctx.begin(), ctx.end(), handler);
-    auto eh = ctx.error_handler();
-    switch (type) {
-    case detail::type::none_type:
-      FMT_ASSERT(false, "invalid argument type");
-      break;
-    case detail::type::int_type:
-    case detail::type::uint_type:
-    case detail::type::long_long_type:
-    case detail::type::ulong_long_type:
-    case detail::type::int128_type:
-    case detail::type::uint128_type:
-    case detail::type::bool_type:
-      handle_int_type_spec(specs_.type,
-                           detail::int_type_checker<decltype(eh)>(eh));
-      break;
-    case detail::type::char_type:
-      handle_char_specs(
-          &specs_, detail::char_specs_checker<decltype(eh)>(specs_.type, eh));
-      break;
-    case detail::type::float_type:
-      if (detail::const_check(FMT_USE_FLOAT))
-        detail::parse_float_type_spec(specs_, eh);
-      else
-        FMT_ASSERT(false, "float support disabled");
-      break;
-    case detail::type::double_type:
-      if (detail::const_check(FMT_USE_DOUBLE))
-        detail::parse_float_type_spec(specs_, eh);
-      else
-        FMT_ASSERT(false, "double support disabled");
-      break;
-    case detail::type::long_double_type:
-      if (detail::const_check(FMT_USE_LONG_DOUBLE))
-        detail::parse_float_type_spec(specs_, eh);
-      else
-        FMT_ASSERT(false, "long double support disabled");
-      break;
-    case detail::type::cstring_type:
-      detail::handle_cstring_type_spec(
-          specs_.type, detail::cstring_type_checker<decltype(eh)>(eh));
-      break;
-    case detail::type::string_type:
-      detail::check_string_type_spec(specs_.type, eh);
-      break;
-    case detail::type::pointer_type:
-      detail::check_pointer_type_spec(specs_.type, eh);
-      break;
-    case detail::type::custom_type:
-      // Custom format specifiers should be checked in parse functions of
-      // formatter specializations.
-      break;
-    }
-    return it;
-  }
-
-  template <typename FormatContext>
-  auto format(const T& val, FormatContext& ctx) -> decltype(ctx.out()) {
-    detail::handle_dynamic_spec<detail::width_checker>(specs_.width,
-                                                       specs_.width_ref, ctx);
+template <typename FormatContext>
+FMT_CONSTEXPR FMT_INLINE auto
+formatter<T, Char,
+          enable_if_t<detail::type_constant<T, Char>::value !=
+                      detail::type::custom_type>>::format(const T& val,
+                                                          FormatContext& ctx)
+    const -> decltype(ctx.out()) {
+  if (specs_.width_ref.kind != detail::arg_id_kind::none ||
+      specs_.precision_ref.kind != detail::arg_id_kind::none) {
+    auto specs = specs_;
+    detail::handle_dynamic_spec<detail::width_checker>(specs.width,
+                                                       specs.width_ref, ctx);
     detail::handle_dynamic_spec<detail::precision_checker>(
-        specs_.precision, specs_.precision_ref, ctx);
-    using af = detail::arg_formatter<typename FormatContext::iterator,
-                                     typename FormatContext::char_type>;
-    return visit_format_arg(af(ctx, nullptr, &specs_),
-                            detail::make_arg<FormatContext>(val));
+        specs.precision, specs.precision_ref, ctx);
+    return detail::write<Char>(ctx.out(), val, specs, ctx.locale());
   }
+  return detail::write<Char>(ctx.out(), val, specs_, ctx.locale());
+}
 
- private:
-  detail::dynamic_format_specs<Char> specs_;
-};
-
-#define FMT_FORMAT_AS(Type, Base)                                             \
-  template <typename Char>                                                    \
-  struct formatter<Type, Char> : formatter<Base, Char> {                      \
-    template <typename FormatContext>                                         \
-    auto format(Type const& val, FormatContext& ctx) -> decltype(ctx.out()) { \
-      return formatter<Base, Char>::format(val, ctx);                         \
-    }                                                                         \
+#define FMT_FORMAT_AS(Type, Base)                                        \
+  template <typename Char>                                               \
+  struct formatter<Type, Char> : formatter<Base, Char> {                 \
+    template <typename FormatContext>                                    \
+    auto format(Type const& val, FormatContext& ctx) const               \
+        -> decltype(ctx.out()) {                                         \
+      return formatter<Base, Char>::format(static_cast<Base>(val), ctx); \
+    }                                                                    \
   }
 
 FMT_FORMAT_AS(signed char, int);
@@ -3496,12 +2596,13 @@ FMT_FORMAT_AS(unsigned long, unsigned long long);
 FMT_FORMAT_AS(Char*, const Char*);
 FMT_FORMAT_AS(std::basic_string<Char>, basic_string_view<Char>);
 FMT_FORMAT_AS(std::nullptr_t, const void*);
+FMT_FORMAT_AS(detail::byte, unsigned char);
 FMT_FORMAT_AS(detail::std_string_view<Char>, basic_string_view<Char>);
 
 template <typename Char>
 struct formatter<void*, Char> : formatter<const void*, Char> {
   template <typename FormatContext>
-  auto format(void* val, FormatContext& ctx) -> decltype(ctx.out()) {
+  auto format(void* val, FormatContext& ctx) const -> decltype(ctx.out()) {
     return formatter<const void*, Char>::format(val, ctx);
   }
 };
@@ -3509,7 +2610,8 @@ struct formatter<void*, Char> : formatter<const void*, Char> {
 template <typename Char, size_t N>
 struct formatter<Char[N], Char> : formatter<basic_string_view<Char>, Char> {
   template <typename FormatContext>
-  auto format(const Char* val, FormatContext& ctx) -> decltype(ctx.out()) {
+  FMT_CONSTEXPR auto format(const Char* val, FormatContext& ctx) const
+      -> decltype(ctx.out()) {
     return formatter<basic_string_view<Char>, Char>::format(val, ctx);
   }
 };
@@ -3528,21 +2630,29 @@ struct formatter<Char[N], Char> : formatter<basic_string_view<Char>, Char> {
 //   };
 template <typename Char = char> class dynamic_formatter {
  private:
+  detail::dynamic_format_specs<Char> specs_;
+  const Char* format_str_;
+
   struct null_handler : detail::error_handler {
     void on_align(align_t) {}
-    void on_plus() {}
-    void on_minus() {}
-    void on_space() {}
+    void on_sign(sign_t) {}
     void on_hash() {}
   };
 
+  template <typename Context> void handle_specs(Context& ctx) {
+    detail::handle_dynamic_spec<detail::width_checker>(specs_.width,
+                                                       specs_.width_ref, ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs_.precision, specs_.precision_ref, ctx);
+  }
+
  public:
   template <typename ParseContext>
-  auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
     format_str_ = ctx.begin();
     // Checks are deferred to formatting time when the argument type is known.
     detail::dynamic_specs_handler<ParseContext> handler(specs_, ctx);
-    return parse_format_specs(ctx.begin(), ctx.end(), handler);
+    return detail::parse_format_specs(ctx.begin(), ctx.end(), handler);
   }
 
   template <typename T, typename FormatContext>
@@ -3551,46 +2661,13 @@ template <typename Char = char> class dynamic_formatter {
     detail::specs_checker<null_handler> checker(
         null_handler(), detail::mapped_type_constant<T, FormatContext>::value);
     checker.on_align(specs_.align);
-    switch (specs_.sign) {
-    case sign::none:
-      break;
-    case sign::plus:
-      checker.on_plus();
-      break;
-    case sign::minus:
-      checker.on_minus();
-      break;
-    case sign::space:
-      checker.on_space();
-      break;
-    }
+    if (specs_.sign != sign::none) checker.on_sign(specs_.sign);
     if (specs_.alt) checker.on_hash();
     if (specs_.precision >= 0) checker.end_precision();
-    using af = detail::arg_formatter<typename FormatContext::iterator,
-                                     typename FormatContext::char_type>;
-    visit_format_arg(af(ctx, nullptr, &specs_),
-                     detail::make_arg<FormatContext>(val));
-    return ctx.out();
-  }
-
- private:
-  template <typename Context> void handle_specs(Context& ctx) {
-    detail::handle_dynamic_spec<detail::width_checker>(specs_.width,
-                                                       specs_.width_ref, ctx);
-    detail::handle_dynamic_spec<detail::precision_checker>(
-        specs_.precision, specs_.precision_ref, ctx);
+    return detail::write<Char>(ctx.out(), val, specs_, ctx.locale());
   }
-
-  detail::dynamic_format_specs<Char> specs_;
-  const Char* format_str_;
 };
 
-template <typename Char, typename ErrorHandler>
-FMT_CONSTEXPR void advance_to(
-    basic_format_parse_context<Char, ErrorHandler>& ctx, const Char* p) {
-  ctx.advance_to(ctx.begin() + (p - &*ctx.begin()));
-}
-
 /**
   \rst
   Converts ``p`` to ``const void*`` for pointer formatting.
@@ -3600,11 +2677,14 @@ FMT_CONSTEXPR void advance_to(
     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) {
+template <typename T> auto ptr(T p) -> const void* {
+  static_assert(std::is_pointer<T>::value, "");
+  return detail::bit_cast<const void*>(p);
+}
+template <typename T> auto ptr(const std::unique_ptr<T>& p) -> const void* {
   return p.get();
 }
-template <typename T> inline const void* ptr(const std::shared_ptr<T>& p) {
+template <typename T> auto ptr(const std::shared_ptr<T>& p) -> const void* {
   return p.get();
 }
 
@@ -3642,31 +2722,114 @@ template <> struct formatter<bytes> {
   }
 };
 
-template <typename It, typename Sentinel, typename Char>
-struct arg_join : detail::view {
+// group_digits_view is not derived from view because it copies the argument.
+template <typename T> struct group_digits_view { T value; };
+
+/**
+  \rst
+  Returns a view that formats an integer value using ',' as a locale-independent
+  thousands separator.
+
+  **Example**::
+
+    fmt::print("{}", fmt::group_digits(12345));
+    // Output: "12,345"
+  \endrst
+ */
+template <typename T> auto group_digits(T value) -> group_digits_view<T> {
+  return {value};
+}
+
+template <typename T> struct formatter<group_digits_view<T>> : formatter<T> {
+ private:
+  detail::dynamic_format_specs<char> specs_;
+
+ public:
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    using handler_type = detail::dynamic_specs_handler<ParseContext>;
+    detail::specs_checker<handler_type> handler(handler_type(specs_, ctx),
+                                                detail::type::int_type);
+    auto it = parse_format_specs(ctx.begin(), ctx.end(), handler);
+    detail::check_string_type_spec(specs_.type, ctx.error_handler());
+    return it;
+  }
+
+  template <typename FormatContext>
+  auto format(group_digits_view<T> t, FormatContext& ctx)
+      -> decltype(ctx.out()) {
+    detail::handle_dynamic_spec<detail::width_checker>(specs_.width,
+                                                       specs_.width_ref, ctx);
+    detail::handle_dynamic_spec<detail::precision_checker>(
+        specs_.precision, specs_.precision_ref, ctx);
+    return detail::write_int_localized(
+        ctx.out(), static_cast<detail::uint64_or_128_t<T>>(t.value), 0, specs_,
+        detail::digit_grouping<char>({"\3", ','}));
+  }
+};
+
+template <typename It, typename Sentinel, typename Char = char>
+struct join_view : detail::view {
   It begin;
   Sentinel end;
   basic_string_view<Char> sep;
 
-  arg_join(It b, Sentinel e, basic_string_view<Char> s)
+  join_view(It b, Sentinel e, basic_string_view<Char> s)
       : begin(b), end(e), sep(s) {}
 };
 
 template <typename It, typename Sentinel, typename Char>
-struct formatter<arg_join<It, Sentinel, Char>, Char>
-    : formatter<typename std::iterator_traits<It>::value_type, Char> {
+using arg_join FMT_DEPRECATED_ALIAS = join_view<It, Sentinel, Char>;
+
+template <typename It, typename Sentinel, typename Char>
+struct formatter<join_view<It, Sentinel, Char>, Char> {
+ private:
+  using value_type =
+#ifdef __cpp_lib_ranges
+      std::iter_value_t<It>;
+#else
+      typename std::iterator_traits<It>::value_type;
+#endif
+  using context = buffer_context<Char>;
+  using mapper = detail::arg_mapper<context>;
+
+  template <typename T, FMT_ENABLE_IF(has_formatter<T, context>::value)>
+  static auto map(const T& value) -> const T& {
+    return value;
+  }
+  template <typename T, FMT_ENABLE_IF(!has_formatter<T, context>::value)>
+  static auto map(const T& value) -> decltype(mapper().map(value)) {
+    return mapper().map(value);
+  }
+
+  using formatter_type =
+      conditional_t<is_formattable<value_type, Char>::value,
+                    formatter<remove_cvref_t<decltype(map(
+                                  std::declval<const value_type&>()))>,
+                              Char>,
+                    detail::fallback_formatter<value_type, Char>>;
+
+  formatter_type value_formatter_;
+
+ public:
+  template <typename ParseContext>
+  FMT_CONSTEXPR auto parse(ParseContext& ctx) -> decltype(ctx.begin()) {
+    return value_formatter_.parse(ctx);
+  }
+
   template <typename FormatContext>
-  auto format(const arg_join<It, Sentinel, Char>& value, FormatContext& ctx)
+  auto format(const join_view<It, Sentinel, Char>& value, FormatContext& ctx)
       -> decltype(ctx.out()) {
-    using base = formatter<typename std::iterator_traits<It>::value_type, Char>;
     auto it = value.begin;
     auto out = ctx.out();
     if (it != value.end) {
-      out = base::format(*it++, ctx);
+      out = value_formatter_.format(map(*it), ctx);
+      ++it;
       while (it != value.end) {
-        out = std::copy(value.sep.begin(), value.sep.end(), out);
+        out = detail::copy_str<Char>(value.sep.begin(), value.sep.end(), out);
         ctx.advance_to(out);
-        out = base::format(*it++, ctx);
+        out = value_formatter_.format(map(*it), ctx);
+        ++it;
       }
     }
     return out;
@@ -3674,22 +2837,17 @@ struct formatter<arg_join<It, Sentinel, Char>, Char>
 };
 
 /**
-  Returns an object that formats the iterator range `[begin, end)` with elements
+  Returns a view that formats the iterator range `[begin, end)` with elements
   separated by `sep`.
  */
 template <typename It, typename Sentinel>
-arg_join<It, Sentinel, char> join(It begin, Sentinel end, string_view sep) {
-  return {begin, end, sep};
-}
-
-template <typename It, typename Sentinel>
-arg_join<It, Sentinel, wchar_t> join(It begin, Sentinel end, wstring_view sep) {
+auto join(It begin, Sentinel end, string_view sep) -> join_view<It, Sentinel> {
   return {begin, end, sep};
 }
 
 /**
   \rst
-  Returns an object that formats `range` with elements separated by `sep`.
+  Returns a view that formats `range` with elements separated by `sep`.
 
   **Example**::
 
@@ -3704,14 +2862,8 @@ arg_join<It, Sentinel, wchar_t> join(It begin, Sentinel end, wstring_view sep) {
   \endrst
  */
 template <typename Range>
-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<Range>, detail::sentinel_t<Range>, wchar_t> join(
-    Range&& range, wstring_view sep) {
+auto join(Range&& range, string_view sep)
+    -> join_view<detail::iterator_t<Range>, detail::sentinel_t<Range>> {
   return join(std::begin(range), std::end(range), sep);
 }
 
@@ -3727,229 +2879,221 @@ arg_join<detail::iterator_t<Range>, detail::sentinel_t<Range>, wchar_t> join(
   \endrst
  */
 template <typename T, FMT_ENABLE_IF(!std::is_integral<T>::value)>
-inline std::string to_string(const T& value) {
-  std::string result;
+inline auto to_string(const T& value) -> std::string {
+  auto result = std::string();
   detail::write<char>(std::back_inserter(result), value);
   return result;
 }
 
 template <typename T, FMT_ENABLE_IF(std::is_integral<T>::value)>
-inline std::string to_string(T value) {
-  // The buffer should be large enough to store the number including the sign or
-  // "false" for bool.
+FMT_NODISCARD inline auto to_string(T value) -> std::string {
+  // The buffer should be large enough to store the number including the sign
+  // or "false" for bool.
   constexpr int max_size = detail::digits10<T>() + 2;
   char buffer[max_size > 5 ? static_cast<unsigned>(max_size) : 5];
   char* begin = buffer;
   return std::string(begin, detail::write<char>(begin, value));
 }
 
-/**
-  Converts *value* to ``std::wstring`` using the default format for type *T*.
- */
-template <typename T> inline std::wstring to_wstring(const T& value) {
-  return format(L"{}", value);
-}
-
 template <typename Char, size_t SIZE>
-std::basic_string<Char> to_string(const basic_memory_buffer<Char, SIZE>& buf) {
+FMT_NODISCARD auto to_string(const basic_memory_buffer<Char, SIZE>& buf)
+    -> std::basic_string<Char> {
   auto size = buf.size();
   detail::assume(size < std::basic_string<Char>().max_size());
   return std::basic_string<Char>(buf.data(), size);
 }
 
+FMT_BEGIN_DETAIL_NAMESPACE
+
 template <typename Char>
-void detail::vformat_to(
-    detail::buffer<Char>& buf, basic_string_view<Char> format_str,
-    basic_format_args<buffer_context<type_identity_t<Char>>> args,
-    detail::locale_ref loc) {
-  using iterator = typename buffer_context<Char>::iterator;
+void vformat_to(
+    buffer<Char>& buf, basic_string_view<Char> fmt,
+    basic_format_args<FMT_BUFFER_CONTEXT(type_identity_t<Char>)> args,
+    locale_ref loc) {
+  // workaround for msvc bug regarding name-lookup in module
+  // link names into function scope
+  using detail::arg_formatter;
+  using detail::buffer_appender;
+  using detail::custom_formatter;
+  using detail::default_arg_formatter;
+  using detail::get_arg;
+  using detail::locale_ref;
+  using detail::parse_format_specs;
+  using detail::specs_checker;
+  using detail::specs_handler;
+  using detail::to_unsigned;
+  using detail::type;
+  using detail::write;
   auto out = buffer_appender<Char>(buf);
-  if (format_str.size() == 2 && equal2(format_str.data(), "{}")) {
+  if (fmt.size() == 2 && equal2(fmt.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);
+    visit_format_arg(default_arg_formatter<Char>{out, args, loc}, arg);
     return;
   }
-  format_handler<iterator, Char, buffer_context<Char>> h(out, format_str, args,
-                                                         loc);
-  parse_format_string<false>(format_str, h);
-}
-
-#ifndef FMT_HEADER_ONLY
-extern template void detail::vformat_to(detail::buffer<char>&, string_view,
-                                        basic_format_args<format_context>,
-                                        detail::locale_ref);
-namespace detail {
 
-extern template FMT_API std::string grouping_impl<char>(locale_ref loc);
-extern template FMT_API std::string grouping_impl<wchar_t>(locale_ref loc);
-extern template FMT_API char thousands_sep_impl<char>(locale_ref loc);
-extern template FMT_API wchar_t thousands_sep_impl<wchar_t>(locale_ref loc);
-extern template FMT_API char decimal_point_impl(locale_ref loc);
-extern template FMT_API wchar_t decimal_point_impl(locale_ref loc);
-extern template int format_float<double>(double value, int precision,
-                                         float_specs specs, buffer<char>& buf);
-extern template int format_float<long double>(long double value, int precision,
-                                              float_specs specs,
-                                              buffer<char>& buf);
-int snprintf_float(float value, int precision, float_specs specs,
-                   buffer<char>& buf) = delete;
-extern template int snprintf_float<double>(double value, int precision,
-                                           float_specs specs,
-                                           buffer<char>& buf);
-extern template int snprintf_float<long double>(long double value,
-                                                int precision,
-                                                float_specs specs,
-                                                buffer<char>& buf);
-}  // namespace detail
-#endif
-
-template <typename S, typename Char = char_t<S>,
-          FMT_ENABLE_IF(detail::is_string<S>::value)>
-inline void vformat_to(
-    detail::buffer<Char>& buf, const S& format_str,
-    basic_format_args<FMT_BUFFER_CONTEXT(type_identity_t<Char>)> args) {
-  return detail::vformat_to(buf, to_string_view(format_str), args);
-}
-
-template <typename S, typename... Args, 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) {
-  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>
-using format_context_t = basic_format_context<OutputIt, Char>;
+  struct format_handler : error_handler {
+    basic_format_parse_context<Char> parse_context;
+    buffer_context<Char> context;
 
-template <typename OutputIt, typename Char = char>
-using format_args_t = basic_format_args<format_context_t<OutputIt, Char>>;
+    format_handler(buffer_appender<Char> out, basic_string_view<Char> str,
+                   basic_format_args<buffer_context<Char>> args, locale_ref loc)
+        : parse_context(str), context(out, args, loc) {}
 
-template <typename OutputIt, typename Char = typename OutputIt::value_type>
-using format_to_n_context FMT_DEPRECATED_ALIAS = buffer_context<Char>;
-
-template <typename OutputIt, typename Char = typename OutputIt::value_type>
-using format_to_n_args FMT_DEPRECATED_ALIAS =
-    basic_format_args<buffer_context<Char>>;
+    void on_text(const Char* begin, const Char* end) {
+      auto text = basic_string_view<Char>(begin, to_unsigned(end - begin));
+      context.advance_to(write<Char>(context.out(), text));
+    }
 
-template <typename OutputIt, typename Char, typename... Args>
-FMT_DEPRECATED format_arg_store<buffer_context<Char>, Args...>
-make_format_to_n_args(const Args&... args) {
-  return format_arg_store<buffer_context<Char>, Args...>(args...);
-}
+    FMT_CONSTEXPR auto on_arg_id() -> int {
+      return parse_context.next_arg_id();
+    }
+    FMT_CONSTEXPR auto on_arg_id(int id) -> int {
+      return parse_context.check_arg_id(id), id;
+    }
+    FMT_CONSTEXPR auto on_arg_id(basic_string_view<Char> id) -> int {
+      int arg_id = context.arg_id(id);
+      if (arg_id < 0) on_error("argument not found");
+      return arg_id;
+    }
 
-template <typename Char, enable_if_t<(!std::is_same<Char, char>::value), int>>
-std::basic_string<Char> detail::vformat(
-    basic_string_view<Char> format_str,
-    basic_format_args<buffer_context<type_identity_t<Char>>> args) {
-  basic_memory_buffer<Char> buffer;
-  detail::vformat_to(buffer, format_str, args);
-  return to_string(buffer);
-}
+    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<Char>{context.out(), context.args(),
+                                      context.locale()},
+          arg));
+    }
 
-template <typename Char, FMT_ENABLE_IF(std::is_same<Char, wchar_t>::value)>
-void vprint(std::FILE* f, basic_string_view<Char> format_str,
-            wformat_args args) {
-  wmemory_buffer buffer;
-  detail::vformat_to(buffer, format_str, args);
-  buffer.push_back(L'\0');
-  if (std::fputws(buffer.data(), f) == -1)
-    FMT_THROW(system_error(errno, "cannot write to file"));
+    auto on_format_specs(int id, const Char* begin, const Char* end)
+        -> const Char* {
+      auto arg = get_arg(context, id);
+      if (arg.type() == type::custom_type) {
+        parse_context.advance_to(parse_context.begin() +
+                                 (begin - &*parse_context.begin()));
+        visit_format_arg(custom_formatter<Char>{parse_context, context}, arg);
+        return parse_context.begin();
+      }
+      auto specs = basic_format_specs<Char>();
+      specs_checker<specs_handler<Char>> handler(
+          specs_handler<Char>(specs, parse_context, context), arg.type());
+      begin = parse_format_specs(begin, end, handler);
+      if (begin == end || *begin != '}')
+        on_error("missing '}' in format string");
+      auto f = arg_formatter<Char>{context.out(), specs, context.locale()};
+      context.advance_to(visit_format_arg(f, arg));
+      return begin;
+    }
+  };
+  detail::parse_format_string<false>(fmt, format_handler(out, fmt, args, loc));
 }
 
-template <typename Char, FMT_ENABLE_IF(std::is_same<Char, wchar_t>::value)>
-void vprint(basic_string_view<Char> format_str, wformat_args args) {
-  vprint(stdout, format_str, args);
-}
+#ifndef FMT_HEADER_ONLY
+extern template FMT_API auto thousands_sep_impl<char>(locale_ref)
+    -> thousands_sep_result<char>;
+extern template FMT_API auto thousands_sep_impl<wchar_t>(locale_ref)
+    -> thousands_sep_result<wchar_t>;
+extern template FMT_API auto decimal_point_impl(locale_ref) -> char;
+extern template FMT_API auto decimal_point_impl(locale_ref) -> wchar_t;
+extern template auto format_float<double>(double value, int precision,
+                                          float_specs specs, buffer<char>& buf)
+    -> int;
+extern template auto format_float<long double>(long double value, int precision,
+                                               float_specs specs,
+                                               buffer<char>& buf) -> int;
+void snprintf_float(float, int, float_specs, buffer<char>&) = delete;
+extern template auto snprintf_float<double>(double value, int precision,
+                                            float_specs specs,
+                                            buffer<char>& buf) -> int;
+extern template auto snprintf_float<long double>(long double value,
+                                                 int precision,
+                                                 float_specs specs,
+                                                 buffer<char>& buf) -> int;
+#endif  // FMT_HEADER_ONLY
+
+FMT_END_DETAIL_NAMESPACE
 
 #if FMT_USE_USER_DEFINED_LITERALS
-namespace detail {
-
-#  if FMT_USE_UDL_TEMPLATE
-template <typename Char, Char... CHARS> class udl_formatter {
- public:
-  template <typename... Args>
-  std::basic_string<Char> operator()(Args&&... args) const {
-    static FMT_CONSTEXPR_DECL Char s[] = {CHARS..., '\0'};
-    return format(FMT_STRING(s), std::forward<Args>(args)...);
-  }
-};
-#  else
-template <typename Char> struct udl_formatter {
-  basic_string_view<Char> str;
-
-  template <typename... Args>
-  std::basic_string<Char> operator()(Args&&... args) const {
-    return format(str, std::forward<Args>(args)...);
-  }
-};
-#  endif  // FMT_USE_UDL_TEMPLATE
-
-template <typename Char> struct udl_arg {
-  const Char* str;
-
-  template <typename T> named_arg<Char, T> operator=(T&& value) const {
-    return {str, std::forward<T>(value)};
-  }
-};
-}  // namespace detail
-
 inline namespace literals {
-#  if FMT_USE_UDL_TEMPLATE
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wpedantic"
-#    if FMT_CLANG_VERSION
-#      pragma GCC diagnostic ignored "-Wgnu-string-literal-operator-template"
-#    endif
-template <typename Char, Char... CHARS>
-FMT_CONSTEXPR detail::udl_formatter<Char, CHARS...> operator""_format() {
-  return {};
-}
-#    pragma GCC diagnostic pop
-#  else
 /**
   \rst
-  User-defined literal equivalent of :func:`fmt::format`.
+  User-defined literal equivalent of :func:`fmt::arg`.
 
   **Example**::
 
     using namespace fmt::literals;
-    std::string message = "The answer is {}"_format(42);
+    fmt::print("Elapsed time: {s:.2f} seconds", "s"_a=1.23);
   \endrst
  */
-FMT_CONSTEXPR detail::udl_formatter<char> operator"" _format(const char* s,
-                                                             size_t n) {
-  return {{s, n}};
+#  if FMT_USE_NONTYPE_TEMPLATE_PARAMETERS
+template <detail_exported::fixed_string Str>
+constexpr auto operator""_a()
+    -> detail::udl_arg<remove_cvref_t<decltype(Str.data[0])>,
+                       sizeof(Str.data) / sizeof(decltype(Str.data[0])), Str> {
+  return {};
+}
+#  else
+constexpr auto operator"" _a(const char* s, size_t) -> detail::udl_arg<char> {
+  return {s};
 }
-FMT_CONSTEXPR detail::udl_formatter<wchar_t> operator"" _format(
-    const wchar_t* s, size_t n) {
+#  endif
+
+// DEPRECATED!
+// User-defined literal equivalent of fmt::format.
+FMT_DEPRECATED constexpr auto operator"" _format(const char* s, size_t n)
+    -> detail::udl_formatter<char> {
   return {{s, n}};
 }
-#  endif  // FMT_USE_UDL_TEMPLATE
+}  // namespace literals
+#endif  // FMT_USE_USER_DEFINED_LITERALS
 
-/**
-  \rst
-  User-defined literal equivalent of :func:`fmt::arg`.
+template <typename Locale, FMT_ENABLE_IF(detail::is_locale<Locale>::value)>
+inline auto vformat(const Locale& loc, string_view fmt, format_args args)
+    -> std::string {
+  return detail::vformat(loc, fmt, args);
+}
 
-  **Example**::
+template <typename Locale, typename... T,
+          FMT_ENABLE_IF(detail::is_locale<Locale>::value)>
+inline auto format(const Locale& loc, format_string<T...> fmt, T&&... args)
+    -> std::string {
+  return vformat(loc, string_view(fmt), fmt::make_format_args(args...));
+}
 
-    using namespace fmt::literals;
-    fmt::print("Elapsed time: {s:.2f} seconds", "s"_a=1.23);
-  \endrst
- */
-FMT_CONSTEXPR detail::udl_arg<char> operator"" _a(const char* s, size_t) {
-  return {s};
+template <typename... T, size_t SIZE, typename Allocator>
+FMT_DEPRECATED auto format_to(basic_memory_buffer<char, SIZE, Allocator>& buf,
+                              format_string<T...> fmt, T&&... args)
+    -> appender {
+  detail::vformat_to(buf, string_view(fmt), fmt::make_format_args(args...));
+  return appender(buf);
 }
-FMT_CONSTEXPR detail::udl_arg<wchar_t> operator"" _a(const wchar_t* s, size_t) {
-  return {s};
+
+template <typename OutputIt, typename Locale,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value&&
+                            detail::is_locale<Locale>::value)>
+auto vformat_to(OutputIt out, const Locale& loc, string_view fmt,
+                format_args args) -> OutputIt {
+  using detail::get_buffer;
+  auto&& buf = get_buffer<char>(out);
+  detail::vformat_to(buf, fmt, args, detail::locale_ref(loc));
+  return detail::get_iterator(buf);
 }
-}  // namespace literals
-#endif  // FMT_USE_USER_DEFINED_LITERALS
+
+template <typename OutputIt, typename Locale, typename... T,
+          FMT_ENABLE_IF(detail::is_output_iterator<OutputIt, char>::value&&
+                            detail::is_locale<Locale>::value)>
+FMT_INLINE auto format_to(OutputIt out, const Locale& loc,
+                          format_string<T...> fmt, T&&... args) -> OutputIt {
+  return vformat_to(out, loc, fmt, fmt::make_format_args(args...));
+}
+
+FMT_MODULE_EXPORT_END
 FMT_END_NAMESPACE
 
+#ifdef FMT_DEPRECATED_INCLUDE_XCHAR
+#  include "xchar.h"
+#endif
+
 #ifdef FMT_HEADER_ONLY
 #  define FMT_FUNC inline
 #  include "format-inl.h"
diff --git a/src/third_party_lib/intx/CMakeLists.txt b/src/third_party_lib/intx/CMakeLists.txt
index dd0c0919..52bfe5bd 100644
--- a/src/third_party_lib/intx/CMakeLists.txt
+++ b/src/third_party_lib/intx/CMakeLists.txt
@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.10)
 
 set(SOURCES
-	int128.hpp
+	intx.hpp
 )
 
 add_library(intx INTERFACE)
diff --git a/src/third_party_lib/intx/int128.hpp b/src/third_party_lib/intx/int128.hpp
deleted file mode 100644
index 9577e72e..00000000
--- a/src/third_party_lib/intx/int128.hpp
+++ /dev/null
@@ -1,885 +0,0 @@
-// intx: extended precision integer library.
-// Copyright 2019-2020 Pawel Bylica.
-// Licensed under the Apache License, Version 2.0.
-
-#pragma once
-
-#include <algorithm>
-#include <climits>
-#include <cstdint>
-#include <limits>
-#include <stdexcept>
-#include <string>
-#include <tuple>
-#include <type_traits>
-
-#ifdef _MSC_VER
-    #include <intrin.h>
-#endif
-
-#ifdef _MSC_VER
-    #define INTX_UNREACHABLE __assume(0)
-#else
-    #define INTX_UNREACHABLE __builtin_unreachable()
-#endif
-
-#ifdef _MSC_VER
-    #define INTX_UNLIKELY(EXPR) (bool{EXPR})
-#else
-    #define INTX_UNLIKELY(EXPR) __builtin_expect(bool{EXPR}, false)
-#endif
-
-#ifdef NDEBUG
-    #define INTX_REQUIRE(X) (X) ? (void)0 : INTX_UNREACHABLE
-#else
-    #include <cassert>
-    #define INTX_REQUIRE assert
-#endif
-
-namespace intx
-{
-template <unsigned N>
-struct uint;
-
-/// The 128-bit unsigned integer.
-///
-/// This type is defined as a specialization of uint<> to easier integration with full intx package,
-/// however, uint128 may be used independently.
-template <>
-struct uint<128>
-{
-    static constexpr unsigned num_bits = 128;
-
-    uint64_t lo = 0;
-    uint64_t hi = 0;
-
-    constexpr uint() noexcept = default;
-
-    constexpr uint(uint64_t high, uint64_t low) noexcept : lo{low}, hi{high} {}
-
-    template <typename T,
-        typename = typename std::enable_if_t<std::is_convertible<T, uint64_t>::value>>
-    constexpr uint(T x) noexcept : lo(static_cast<uint64_t>(x))  // NOLINT
-    {}
-
-#ifdef __SIZEOF_INT128__
-    #pragma GCC diagnostic push
-    #pragma GCC diagnostic ignored "-Wpedantic"
-    constexpr uint(unsigned __int128 x) noexcept  // NOLINT
-      : lo{uint64_t(x)}, hi{uint64_t(x >> 64)}
-    {}
-
-    constexpr explicit operator unsigned __int128() const noexcept
-    {
-        return (static_cast<unsigned __int128>(hi) << 64) | lo;
-    }
-    #pragma GCC diagnostic pop
-#endif
-
-    constexpr explicit operator bool() const noexcept { return hi | lo; }
-
-    /// Explicit converting operator for all builtin integral types.
-    template <typename Int, typename = typename std::enable_if<std::is_integral<Int>::value>::type>
-    constexpr explicit operator Int() const noexcept
-    {
-        return static_cast<Int>(lo);
-    }
-};
-
-using uint128 = uint<128>;
-
-
-/// Contains result of add/sub/etc with a carry flag.
-template <typename T>
-struct result_with_carry
-{
-    T value;
-    bool carry;
-
-    /// Conversion to tuple of references, to allow usage with std::tie().
-    constexpr operator std::tuple<T&, bool&>() noexcept { return {value, carry}; }
-};
-
-
-/// Linear arithmetic operators.
-/// @{
-
-constexpr inline result_with_carry<uint64_t> add_with_carry(
-    uint64_t x, uint64_t y, bool carry = false) noexcept
-{
-    const auto s = x + y;
-    const auto carry1 = s < x;
-    const auto t = s + carry;
-    const auto carry2 = t < s;
-    return {t, carry1 || carry2};
-}
-
-template <unsigned N>
-constexpr result_with_carry<uint<N>> add_with_carry(
-    const uint<N>& a, const uint<N>& b, bool carry = false) noexcept
-{
-    const auto lo = add_with_carry(a.lo, b.lo, carry);
-    const auto hi = add_with_carry(a.hi, b.hi, lo.carry);
-    return {{hi.value, lo.value}, hi.carry};
-}
-
-constexpr inline uint128 operator+(uint128 x, uint128 y) noexcept
-{
-    return add_with_carry(x, y).value;
-}
-
-constexpr inline uint128 operator+(uint128 x) noexcept
-{
-    return x;
-}
-
-constexpr inline result_with_carry<uint64_t> sub_with_carry(
-    uint64_t x, uint64_t y, bool carry = false) noexcept
-{
-    const auto d = x - y;
-    const auto carry1 = d > x;
-    const auto e = d - carry;
-    const auto carry2 = e > d;
-    return {e, carry1 || carry2};
-}
-
-/// Performs subtraction of two unsigned numbers and returns the difference
-/// and the carry bit (aka borrow, overflow).
-template <unsigned N>
-constexpr inline result_with_carry<uint<N>> sub_with_carry(
-    const uint<N>& a, const uint<N>& b, bool carry = false) noexcept
-{
-    const auto lo = sub_with_carry(a.lo, b.lo, carry);
-    const auto hi = sub_with_carry(a.hi, b.hi, lo.carry);
-    return {{hi.value, lo.value}, hi.carry};
-}
-
-constexpr inline uint128 operator-(uint128 x, uint128 y) noexcept
-{
-    return sub_with_carry(x, y).value;
-}
-
-constexpr inline uint128 operator-(uint128 x) noexcept
-{
-    // Implementing as subtraction is better than ~x + 1.
-    // Clang9: Perfect.
-    // GCC8: Does something weird.
-    return 0 - x;
-}
-
-inline uint128& operator++(uint128& x) noexcept
-{
-    return x = x + 1;
-}
-
-inline uint128& operator--(uint128& x) noexcept
-{
-    return x = x - 1;
-}
-
-inline uint128 operator++(uint128& x, int) noexcept
-{
-    auto ret = x;
-    ++x;
-    return ret;
-}
-
-inline uint128 operator--(uint128& x, int) noexcept
-{
-    auto ret = x;
-    --x;
-    return ret;
-}
-
-/// Optimized addition.
-///
-/// This keeps the multiprecision addition until CodeGen so the pattern is not
-/// broken during other optimizations.
-constexpr uint128 fast_add(uint128 x, uint128 y) noexcept
-{
-#ifdef __SIZEOF_INT128__
-    #pragma GCC diagnostic push
-    #pragma GCC diagnostic ignored "-Wpedantic"
-    using uint128_native = unsigned __int128;
-    return uint128_native{x} + uint128_native{y};
-    #pragma GCC diagnostic pop
-#else
-    // Fallback to regular addition.
-    return x + y;
-#endif
-}
-
-/// @}
-
-
-/// Comparison operators.
-///
-/// In all implementations bitwise operators are used instead of logical ones
-/// to avoid branching.
-///
-/// @{
-
-constexpr bool operator==(uint128 x, uint128 y) noexcept
-{
-    // Clang7: generates perfect xor based code,
-    //         much better than __int128 where it uses vector instructions.
-    // GCC8: generates a bit worse cmp based code
-    //       although it generates the xor based one for __int128.
-    return (x.lo == y.lo) & (x.hi == y.hi);
-}
-
-constexpr bool operator!=(uint128 x, uint128 y) noexcept
-{
-    // Analogous to ==, but == not used directly, because that confuses GCC 8-9.
-    return (x.lo != y.lo) | (x.hi != y.hi);
-}
-
-constexpr bool operator<(uint128 x, uint128 y) noexcept
-{
-    // OPT: This should be implemented by checking the borrow of x - y,
-    //      but compilers (GCC8, Clang7)
-    //      have problem with properly optimizing subtraction.
-    return (x.hi < y.hi) | ((x.hi == y.hi) & (x.lo < y.lo));
-}
-
-constexpr bool operator<=(uint128 x, uint128 y) noexcept
-{
-    return !(y < x);
-}
-
-constexpr bool operator>(uint128 x, uint128 y) noexcept
-{
-    return y < x;
-}
-
-constexpr bool operator>=(uint128 x, uint128 y) noexcept
-{
-    return !(x < y);
-}
-
-/// @}
-
-
-/// Bitwise operators.
-/// @{
-
-constexpr uint128 operator~(uint128 x) noexcept
-{
-    return {~x.hi, ~x.lo};
-}
-
-constexpr uint128 operator|(uint128 x, uint128 y) noexcept
-{
-    // Clang7: perfect.
-    // GCC8: stupidly uses a vector instruction in all bitwise operators.
-    return {x.hi | y.hi, x.lo | y.lo};
-}
-
-constexpr uint128 operator&(uint128 x, uint128 y) noexcept
-{
-    return {x.hi & y.hi, x.lo & y.lo};
-}
-
-constexpr uint128 operator^(uint128 x, uint128 y) noexcept
-{
-    return {x.hi ^ y.hi, x.lo ^ y.lo};
-}
-
-constexpr uint128 operator<<(uint128 x, unsigned shift) noexcept
-{
-    return (shift < 64) ?
-               // Find the part moved from lo to hi.
-               // For shift == 0 right shift by (64 - shift) is invalid so
-               // split it into 2 shifts by 1 and (63 - shift).
-               uint128{(x.hi << shift) | ((x.lo >> 1) >> (63 - shift)), x.lo << shift} :
-
-               // Guarantee "defined" behavior for shifts larger than 128.
-               (shift < 128) ? uint128{x.lo << (shift - 64), 0} : 0;
-}
-
-constexpr uint128 operator<<(uint128 x, uint128 shift) noexcept
-{
-    if (shift < 128)
-        return x << unsigned(shift);
-    return 0;
-}
-
-constexpr uint128 operator>>(uint128 x, unsigned shift) noexcept
-{
-    return (shift < 64) ?
-               // Find the part moved from lo to hi.
-               // For shift == 0 left shift by (64 - shift) is invalid so
-               // split it into 2 shifts by 1 and (63 - shift).
-               uint128{x.hi >> shift, (x.lo >> shift) | ((x.hi << 1) << (63 - shift))} :
-
-               // Guarantee "defined" behavior for shifts larger than 128.
-               (shift < 128) ? uint128{0, x.hi >> (shift - 64)} : 0;
-}
-
-constexpr uint128 operator>>(uint128 x, uint128 shift) noexcept
-{
-    if (shift < 128)
-        return x >> unsigned(shift);
-    return 0;
-}
-
-
-/// @}
-
-
-/// Multiplication
-/// @{
-
-/// Portable full unsigned multiplication 64 x 64 -> 128.
-constexpr uint128 constexpr_umul(uint64_t x, uint64_t y) noexcept
-{
-    uint64_t xl = x & 0xffffffff;
-    uint64_t xh = x >> 32;
-    uint64_t yl = y & 0xffffffff;
-    uint64_t yh = y >> 32;
-
-    uint64_t t0 = xl * yl;
-    uint64_t t1 = xh * yl;
-    uint64_t t2 = xl * yh;
-    uint64_t t3 = xh * yh;
-
-    uint64_t u1 = t1 + (t0 >> 32);
-    uint64_t u2 = t2 + (u1 & 0xffffffff);
-
-    uint64_t lo = (u2 << 32) | (t0 & 0xffffffff);
-    uint64_t hi = t3 + (u2 >> 32) + (u1 >> 32);
-    return {hi, lo};
-}
-
-/// Full unsigned multiplication 64 x 64 -> 128.
-inline uint128 umul(uint64_t x, uint64_t y) noexcept
-{
-#if defined(__SIZEOF_INT128__)
-    #pragma GCC diagnostic push
-    #pragma GCC diagnostic ignored "-Wpedantic"
-    const auto p = static_cast<unsigned __int128>(x) * y;
-    return {uint64_t(p >> 64), uint64_t(p)};
-    #pragma GCC diagnostic pop
-#elif defined(_MSC_VER)
-    unsigned __int64 hi;
-    const auto lo = _umul128(x, y, &hi);
-    return {hi, lo};
-#else
-    return constexpr_umul(x, y);
-#endif
-}
-
-inline uint128 operator*(uint128 x, uint128 y) noexcept
-{
-    auto p = umul(x.lo, y.lo);
-    p.hi += (x.lo * y.hi) + (x.hi * y.lo);
-    return {p.hi, p.lo};
-}
-
-constexpr uint128 constexpr_mul(uint128 x, uint128 y) noexcept
-{
-    auto p = constexpr_umul(x.lo, y.lo);
-    p.hi += (x.lo * y.hi) + (x.hi * y.lo);
-    return {p.hi, p.lo};
-}
-
-/// @}
-
-
-/// Assignment operators.
-/// @{
-
-constexpr uint128& operator+=(uint128& x, uint128 y) noexcept
-{
-    return x = x + y;
-}
-
-constexpr uint128& operator-=(uint128& x, uint128 y) noexcept
-{
-    return x = x - y;
-}
-
-inline uint128& operator*=(uint128& x, uint128 y) noexcept
-{
-    return x = x * y;
-}
-
-constexpr uint128& operator|=(uint128& x, uint128 y) noexcept
-{
-    return x = x | y;
-}
-
-constexpr uint128& operator&=(uint128& x, uint128 y) noexcept
-{
-    return x = x & y;
-}
-
-constexpr uint128& operator^=(uint128& x, uint128 y) noexcept
-{
-    return x = x ^ y;
-}
-
-constexpr uint128& operator<<=(uint128& x, unsigned shift) noexcept
-{
-    return x = x << shift;
-}
-
-constexpr uint128& operator>>=(uint128& x, unsigned shift) noexcept
-{
-    return x = x >> shift;
-}
-
-/// @}
-
-
-constexpr unsigned clz_generic(uint32_t x) noexcept
-{
-    unsigned n = 32;
-    for (int i = 4; i >= 0; --i)
-    {
-        const auto s = unsigned{1} << i;
-        const auto hi = x >> s;
-        if (hi != 0)
-        {
-            n -= s;
-            x = hi;
-        }
-    }
-    return n - x;
-}
-
-constexpr unsigned clz_generic(uint64_t x) noexcept
-{
-    unsigned n = 64;
-    for (int i = 5; i >= 0; --i)
-    {
-        const auto s = unsigned{1} << i;
-        const auto hi = x >> s;
-        if (hi != 0)
-        {
-            n -= s;
-            x = hi;
-        }
-    }
-    return n - static_cast<unsigned>(x);
-}
-
-constexpr inline unsigned clz(uint32_t x) noexcept
-{
-#ifdef _MSC_VER
-    return clz_generic(x);
-#else
-    return x != 0 ? unsigned(__builtin_clz(x)) : 32;
-#endif
-}
-
-constexpr inline unsigned clz(uint64_t x) noexcept
-{
-#ifdef _MSC_VER
-    return clz_generic(x);
-#else
-    return x != 0 ? unsigned(__builtin_clzll(x)) : 64;
-#endif
-}
-
-constexpr inline unsigned clz(uint128 x) noexcept
-{
-    // In this order `h == 0` we get less instructions than in case of `h != 0`.
-    return x.hi == 0 ? clz(x.lo) + 64 : clz(x.hi);
-}
-
-
-inline uint64_t bswap(uint64_t x) noexcept
-{
-#ifdef _MSC_VER
-    return _byteswap_uint64(x);
-#else
-    return __builtin_bswap64(x);
-#endif
-}
-
-inline uint128 bswap(uint128 x) noexcept
-{
-    return {bswap(x.lo), bswap(x.hi)};
-}
-
-
-/// Division.
-/// @{
-
-template <typename QuotT, typename RemT = QuotT>
-struct div_result
-{
-    QuotT quot;
-    RemT rem;
-
-    /// Conversion to tuple of references, to allow usage with std::tie().
-    constexpr operator std::tuple<QuotT&, RemT&>() noexcept { return {quot, rem}; }
-};
-
-namespace internal
-{
-constexpr uint16_t reciprocal_table_item(uint8_t d9) noexcept
-{
-    return uint16_t(0x7fd00 / (0x100 | d9));
-}
-
-#define REPEAT4(x)                                                  \
-    reciprocal_table_item((x) + 0), reciprocal_table_item((x) + 1), \
-        reciprocal_table_item((x) + 2), reciprocal_table_item((x) + 3)
-
-#define REPEAT32(x)                                                                         \
-    REPEAT4((x) + 4 * 0), REPEAT4((x) + 4 * 1), REPEAT4((x) + 4 * 2), REPEAT4((x) + 4 * 3), \
-        REPEAT4((x) + 4 * 4), REPEAT4((x) + 4 * 5), REPEAT4((x) + 4 * 6), REPEAT4((x) + 4 * 7)
-
-#define REPEAT256()                                                                           \
-    REPEAT32(32 * 0), REPEAT32(32 * 1), REPEAT32(32 * 2), REPEAT32(32 * 3), REPEAT32(32 * 4), \
-        REPEAT32(32 * 5), REPEAT32(32 * 6), REPEAT32(32 * 7)
-
-/// Reciprocal lookup table.
-constexpr uint16_t reciprocal_table[] = {REPEAT256()};
-
-#undef REPEAT4
-#undef REPEAT32
-#undef REPEAT256
-}  // namespace internal
-
-/// Computes the reciprocal (2^128 - 1) / d - 2^64 for normalized d.
-///
-/// Based on Algorithm 2 from "Improved division by invariant integers".
-inline uint64_t reciprocal_2by1(uint64_t d) noexcept
-{
-    INTX_REQUIRE(d & 0x8000000000000000);  // Must be normalized.
-
-    const uint64_t d9 = d >> 55;
-    const uint32_t v0 = internal::reciprocal_table[d9 - 256];
-
-    const uint64_t d40 = (d >> 24) + 1;
-    const uint64_t v1 = (v0 << 11) - uint32_t(v0 * v0 * d40 >> 40) - 1;
-
-    const uint64_t v2 = (v1 << 13) + (v1 * (0x1000000000000000 - v1 * d40) >> 47);
-
-    const uint64_t d0 = d & 1;
-    const uint64_t d63 = (d >> 1) + d0;  // ceil(d/2)
-    const uint64_t e = ((v2 >> 1) & (0 - d0)) - v2 * d63;
-    const uint64_t v3 = (umul(v2, e).hi >> 1) + (v2 << 31);
-
-    const uint64_t v4 = v3 - (umul(v3, d) + d).hi - d;
-    return v4;
-}
-
-inline uint64_t reciprocal_3by2(uint128 d) noexcept
-{
-    auto v = reciprocal_2by1(d.hi);
-    auto p = d.hi * v;
-    p += d.lo;
-    if (p < d.lo)
-    {
-        --v;
-        if (p >= d.hi)
-        {
-            --v;
-            p -= d.hi;
-        }
-        p -= d.hi;
-    }
-
-    const auto t = umul(v, d.lo);
-
-    p += t.hi;
-    if (p < t.hi)
-    {
-        --v;
-        if (p >= d.hi)
-        {
-            if (p > d.hi || t.lo >= d.lo)
-                --v;
-        }
-    }
-    return v;
-}
-
-inline div_result<uint64_t> udivrem_2by1(uint128 u, uint64_t d, uint64_t v) noexcept
-{
-    auto q = umul(v, u.hi);
-    q = fast_add(q, u);
-
-    ++q.hi;
-
-    auto r = u.lo - q.hi * d;
-
-    if (r > q.lo)
-    {
-        --q.hi;
-        r += d;
-    }
-
-    if (r >= d)
-    {
-        ++q.hi;
-        r -= d;
-    }
-
-    return {q.hi, r};
-}
-
-inline div_result<uint64_t, uint128> udivrem_3by2(
-    uint64_t u2, uint64_t u1, uint64_t u0, uint128 d, uint64_t v) noexcept
-{
-    auto q = umul(v, u2);
-    q = fast_add(q, {u2, u1});
-
-    auto r1 = u1 - q.hi * d.hi;
-
-    auto t = umul(d.lo, q.hi);
-
-    auto r = uint128{r1, u0} - t - d;
-    r1 = r.hi;
-
-    ++q.hi;
-
-    if (r1 >= q.lo)
-    {
-        --q.hi;
-        r += d;
-    }
-
-    if (r >= d)
-    {
-        ++q.hi;
-        r -= d;
-    }
-
-    return {q.hi, r};
-}
-
-inline div_result<uint128> udivrem(uint128 x, uint128 y) noexcept
-{
-    if (y.hi == 0)
-    {
-        INTX_REQUIRE(y.lo != 0);  // Division by 0.
-
-        const auto lsh = clz(y.lo);
-        const auto rsh = (64 - lsh) % 64;
-        const auto rsh_mask = uint64_t{lsh == 0} - 1;
-
-        const auto yn = y.lo << lsh;
-        const auto xn_lo = x.lo << lsh;
-        const auto xn_hi = (x.hi << lsh) | ((x.lo >> rsh) & rsh_mask);
-        const auto xn_ex = (x.hi >> rsh) & rsh_mask;
-
-        const auto v = reciprocal_2by1(yn);
-        const auto res1 = udivrem_2by1({xn_ex, xn_hi}, yn, v);
-        const auto res2 = udivrem_2by1({res1.rem, xn_lo}, yn, v);
-        return {{res1.quot, res2.quot}, res2.rem >> lsh};
-    }
-
-    if (y.hi > x.hi)
-        return {0, x};
-
-    const auto lsh = clz(y.hi);
-    if (lsh == 0)
-    {
-        const auto q = unsigned{y.hi < x.hi} | unsigned{y.lo <= x.lo};
-        return {q, x - (q ? y : 0)};
-    }
-
-    const auto rsh = 64 - lsh;
-
-    const auto yn_lo = y.lo << lsh;
-    const auto yn_hi = (y.hi << lsh) | (y.lo >> rsh);
-    const auto xn_lo = x.lo << lsh;
-    const auto xn_hi = (x.hi << lsh) | (x.lo >> rsh);
-    const auto xn_ex = x.hi >> rsh;
-
-    const auto v = reciprocal_3by2({yn_hi, yn_lo});
-    const auto res = udivrem_3by2(xn_ex, xn_hi, xn_lo, {yn_hi, yn_lo}, v);
-
-    return {res.quot, res.rem >> lsh};
-}
-
-inline div_result<uint128> sdivrem(uint128 x, uint128 y) noexcept
-{
-    constexpr auto sign_mask = uint128{1} << 127;
-    const auto x_is_neg = (x & sign_mask) != 0;
-    const auto y_is_neg = (y & sign_mask) != 0;
-
-    const auto x_abs = x_is_neg ? -x : x;
-    const auto y_abs = y_is_neg ? -y : y;
-
-    const auto q_is_neg = x_is_neg ^ y_is_neg;
-
-    const auto res = udivrem(x_abs, y_abs);
-
-    return {q_is_neg ? -res.quot : res.quot, x_is_neg ? -res.rem : res.rem};
-}
-
-inline uint128 operator/(uint128 x, uint128 y) noexcept
-{
-    return udivrem(x, y).quot;
-}
-
-inline uint128 operator%(uint128 x, uint128 y) noexcept
-{
-    return udivrem(x, y).rem;
-}
-
-inline uint128& operator/=(uint128& x, uint128 y) noexcept
-{
-    return x = x / y;
-}
-
-inline uint128& operator%=(uint128& x, uint128 y) noexcept
-{
-    return x = x % y;
-}
-
-/// @}
-
-}  // namespace intx
-
-
-namespace std
-{
-template <unsigned N>
-struct numeric_limits<intx::uint<N>>
-{
-    using type = intx::uint<N>;
-
-    static constexpr bool is_specialized = true;
-    static constexpr bool is_integer = true;
-    static constexpr bool is_signed = false;
-    static constexpr bool is_exact = true;
-    static constexpr bool has_infinity = false;
-    static constexpr bool has_quiet_NaN = false;
-    static constexpr bool has_signaling_NaN = false;
-    static constexpr float_denorm_style has_denorm = denorm_absent;
-    static constexpr bool has_denorm_loss = false;
-    static constexpr float_round_style round_style = round_toward_zero;
-    static constexpr bool is_iec559 = false;
-    static constexpr bool is_bounded = true;
-    static constexpr bool is_modulo = true;
-    static constexpr int digits = CHAR_BIT * sizeof(type);
-    static constexpr int digits10 = int(0.3010299956639812 * digits);
-    static constexpr int max_digits10 = 0;
-    static constexpr int radix = 2;
-    static constexpr int min_exponent = 0;
-    static constexpr int min_exponent10 = 0;
-    static constexpr int max_exponent = 0;
-    static constexpr int max_exponent10 = 0;
-    static constexpr bool traps = std::numeric_limits<unsigned>::traps;
-    static constexpr bool tinyness_before = false;
-
-    static constexpr type min() noexcept { return 0; }
-    static constexpr type lowest() noexcept { return min(); }
-    static constexpr type max() noexcept { return ~type{0}; }
-    static constexpr type epsilon() noexcept { return 0; }
-    static constexpr type round_error() noexcept { return 0; }
-    static constexpr type infinity() noexcept { return 0; }
-    static constexpr type quiet_NaN() noexcept { return 0; }
-    static constexpr type signaling_NaN() noexcept { return 0; }
-    static constexpr type denorm_min() noexcept { return 0; }
-};
-}  // namespace std
-
-namespace intx
-{
-template <typename T>
-[[noreturn]] inline void throw_(const char* what)
-{
-#if __cpp_exceptions
-    throw T{what};
-#else
-    std::fputs(what, stderr);
-    std::abort();
-#endif
-}
-
-constexpr inline int from_dec_digit(char c)
-{
-    if (c < '0' || c > '9')
-        throw_<std::invalid_argument>("invalid digit");
-    return c - '0';
-}
-
-constexpr inline int from_hex_digit(char c)
-{
-    if (c >= 'a' && c <= 'f')
-        return c - ('a' - 10);
-    if (c >= 'A' && c <= 'F')
-        return c - ('A' - 10);
-    return from_dec_digit(c);
-}
-
-template <typename Int>
-constexpr Int from_string(const char* str)
-{
-    auto s = str;
-    auto x = Int{};
-    int num_digits = 0;
-
-    if (s[0] == '0' && s[1] == 'x')
-    {
-        s += 2;
-        while (const auto c = *s++)
-        {
-            if (++num_digits > int{sizeof(x) * 2})
-                throw_<std::out_of_range>(str);
-            x = (x << 4) | from_hex_digit(c);
-        }
-        return x;
-    }
-
-    while (const auto c = *s++)
-    {
-        if (num_digits++ > std::numeric_limits<Int>::digits10)
-            throw_<std::out_of_range>(str);
-
-        const auto d = from_dec_digit(c);
-        x = constexpr_mul(x, Int{10}) + d;
-        if (x < d)
-            throw_<std::out_of_range>(str);
-    }
-    return x;
-}
-
-template <typename Int>
-constexpr Int from_string(const std::string& s)
-{
-    return from_string<Int>(s.c_str());
-}
-
-constexpr uint128 operator""_u128(const char* s)
-{
-    return from_string<uint128>(s);
-}
-
-template <unsigned N>
-inline std::string to_string(uint<N> x, int base = 10)
-{
-    if (base < 2 || base > 36)
-        throw_<std::invalid_argument>("invalid base");
-
-    if (x == 0)
-        return "0";
-
-    auto s = std::string{};
-    while (x != 0)
-    {
-        // TODO: Use constexpr udivrem_1?
-        const auto res = udivrem(x, uint<N>{base});
-        const auto d = int(res.rem);
-        const auto c = d < 10 ? '0' + d : 'a' + d - 10;
-        s.push_back(char(c));
-        x = res.quot;
-    }
-    std::reverse(s.begin(), s.end());
-    return s;
-}
-
-template <unsigned N>
-inline std::string hex(uint<N> x)
-{
-    return to_string(x, 16);
-}
-}  // namespace intx
diff --git a/src/third_party_lib/intx/intx.hpp b/src/third_party_lib/intx/intx.hpp
new file mode 100644
index 00000000..650a3dc9
--- /dev/null
+++ b/src/third_party_lib/intx/intx.hpp
@@ -0,0 +1,2124 @@
+// intx: extended precision integer library.
+// Copyright 2019 Pawel Bylica.
+// Licensed under the Apache License, Version 2.0.
+
+#pragma once
+
+#include <algorithm>
+#include <cassert>
+#include <climits>
+#include <cstdint>
+#include <cstring>
+#include <limits>
+#include <stdexcept>
+#include <string>
+#include <tuple>
+#include <type_traits>
+
+#ifndef __has_builtin
+    #define __has_builtin(NAME) 0
+#endif
+
+#ifdef _MSC_VER
+    #include <intrin.h>
+#endif
+
+#if !defined(__has_builtin)
+    #define __has_builtin(NAME) 0
+#endif
+
+#if !defined(__has_feature)
+    #define __has_feature(NAME) 0
+#endif
+
+#if !defined(NDEBUG)
+    #define INTX_UNREACHABLE() assert(false)
+#elif __has_builtin(__builtin_unreachable)
+    #define INTX_UNREACHABLE() __builtin_unreachable()
+#elif defined(_MSC_VER)
+    #define INTX_UNREACHABLE() __assume(0)
+#else
+    #define INTX_UNREACHABLE() (void)0
+#endif
+
+
+#if __has_builtin(__builtin_expect)
+    #define INTX_UNLIKELY(EXPR) __builtin_expect(bool{EXPR}, false)
+#else
+    #define INTX_UNLIKELY(EXPR) (bool{EXPR})
+#endif
+
+#if !defined(NDEBUG)
+    #define INTX_REQUIRE assert
+#else
+    #define INTX_REQUIRE(X) (X) ? (void)0 : INTX_UNREACHABLE()
+#endif
+
+
+// Detect compiler support for 128-bit integer __int128
+#if defined(__SIZEOF_INT128__)
+    #define INTX_HAS_BUILTIN_INT128 1
+#else
+    #define INTX_HAS_BUILTIN_INT128 0
+#endif
+
+namespace intx
+{
+#if INTX_HAS_BUILTIN_INT128
+    #pragma GCC diagnostic push
+    #pragma GCC diagnostic ignored "-Wpedantic"  // Usage of __int128 triggers a pedantic warning.
+
+/// Alias for the compiler supported unsigned __int128 type.
+using builtin_uint128 = unsigned __int128;
+
+    #pragma GCC diagnostic pop
+#endif
+
+template <unsigned N>
+struct uint;
+
+/// The 128-bit unsigned integer.
+///
+/// This type is defined as a specialization of uint<> to easier integration with full intx package,
+/// however, uint128 may be used independently.
+template <>
+struct uint<128>
+{
+    using word_type = uint64_t;
+    static constexpr auto word_num_bits = sizeof(word_type) * 8;
+    static constexpr unsigned num_bits = 128;
+    static constexpr auto num_words = num_bits / word_num_bits;
+
+private:
+    uint64_t words_[2]{};
+
+public:
+    constexpr uint() noexcept = default;
+
+    constexpr uint(uint64_t low, uint64_t high) noexcept : words_{low, high} {}
+
+    template <typename T,
+        typename = typename std::enable_if_t<std::is_convertible<T, uint64_t>::value>>
+    constexpr uint(T x) noexcept : words_{static_cast<uint64_t>(x), 0}  // NOLINT
+    {}
+
+#if INTX_HAS_BUILTIN_INT128
+    constexpr uint(builtin_uint128 x) noexcept  // NOLINT
+      : words_{uint64_t(x), uint64_t(x >> 64)}
+    {}
+
+    constexpr explicit operator builtin_uint128() const noexcept
+    {
+        return (builtin_uint128{words_[1]} << 64) | words_[0];
+    }
+#endif
+
+    constexpr uint64_t& operator[](size_t i) noexcept { return words_[i]; }
+    constexpr const uint64_t& operator[](size_t i) const noexcept { return words_[i]; }
+
+    constexpr explicit operator bool() const noexcept { return (words_[0] | words_[1]) != 0; }
+
+    /// Explicit converting operator for all builtin integral types.
+    template <typename Int, typename = typename std::enable_if<std::is_integral<Int>::value>::type>
+    constexpr explicit operator Int() const noexcept
+    {
+        return static_cast<Int>(words_[0]);
+    }
+};
+
+using uint128 = uint<128>;
+
+
+inline constexpr bool is_constant_evaluated() noexcept
+{
+#if __has_builtin(__builtin_is_constant_evaluated) || (defined(_MSC_VER) && _MSC_VER >= 1925)
+    return __builtin_is_constant_evaluated();
+#else
+    return true;
+#endif
+}
+
+
+/// Contains result of add/sub/etc with a carry flag.
+template <typename T>
+struct result_with_carry
+{
+    T value;
+    bool carry;
+
+    /// Conversion to tuple of references, to allow usage with std::tie().
+    constexpr operator std::tuple<T&, bool&>() noexcept { return {value, carry}; }
+};
+
+
+/// Linear arithmetic operators.
+/// @{
+
+inline constexpr result_with_carry<uint64_t> add_with_carry(
+    uint64_t x, uint64_t y, bool carry = false) noexcept
+{
+    const auto s = x + y;
+    const auto carry1 = s < x;
+    const auto t = s + carry;
+    const auto carry2 = t < s;
+    return {t, carry1 || carry2};
+}
+
+template <unsigned N>
+inline constexpr result_with_carry<uint<N>> add_with_carry(
+    const uint<N>& x, const uint<N>& y, bool carry = false) noexcept
+{
+    uint<N> s;
+    bool k = carry;
+    for (size_t i = 0; i < uint<N>::num_words; ++i)
+    {
+        s[i] = x[i] + y[i];
+        const auto k1 = s[i] < x[i];
+        s[i] += k;
+        k = (s[i] < uint64_t{k}) || k1;
+    }
+    return {s, k};
+}
+
+inline constexpr uint128 operator+(uint128 x, uint128 y) noexcept
+{
+    return add_with_carry(x, y).value;
+}
+
+inline constexpr uint128 operator+(uint128 x) noexcept
+{
+    return x;
+}
+
+inline constexpr result_with_carry<uint64_t> sub_with_carry(
+    uint64_t x, uint64_t y, bool carry = false) noexcept
+{
+    const auto d = x - y;
+    const auto carry1 = x < y;
+    const auto e = d - carry;
+    const auto carry2 = d < uint64_t{carry};
+    return {e, carry1 || carry2};
+}
+
+/// Performs subtraction of two unsigned numbers and returns the difference
+/// and the carry bit (aka borrow, overflow).
+template <unsigned N>
+inline constexpr result_with_carry<uint<N>> sub_with_carry(
+    const uint<N>& x, const uint<N>& y, bool carry = false) noexcept
+{
+    uint<N> z;
+    bool k = carry;
+    for (size_t i = 0; i < uint<N>::num_words; ++i)
+    {
+        z[i] = x[i] - y[i];
+        const auto k1 = x[i] < y[i];
+        const auto k2 = z[i] < uint64_t{k};
+        z[i] -= k;
+        k = k1 || k2;
+    }
+    return {z, k};
+}
+
+inline constexpr uint128 operator-(uint128 x, uint128 y) noexcept
+{
+    return sub_with_carry(x, y).value;
+}
+
+inline constexpr uint128 operator-(uint128 x) noexcept
+{
+    // Implementing as subtraction is better than ~x + 1.
+    // Clang9: Perfect.
+    // GCC8: Does something weird.
+    return 0 - x;
+}
+
+inline uint128& operator++(uint128& x) noexcept
+{
+    return x = x + 1;
+}
+
+inline uint128& operator--(uint128& x) noexcept
+{
+    return x = x - 1;
+}
+
+inline uint128 operator++(uint128& x, int) noexcept
+{
+    auto ret = x;
+    ++x;
+    return ret;
+}
+
+inline uint128 operator--(uint128& x, int) noexcept
+{
+    auto ret = x;
+    --x;
+    return ret;
+}
+
+/// Optimized addition.
+///
+/// This keeps the multiprecision addition until CodeGen so the pattern is not
+/// broken during other optimizations.
+inline constexpr uint128 fast_add(uint128 x, uint128 y) noexcept
+{
+#if INTX_HAS_BUILTIN_INT128
+    return builtin_uint128{x} + builtin_uint128{y};
+#else
+    return x + y;  // Fallback to generic addition.
+#endif
+}
+
+/// @}
+
+
+/// Comparison operators.
+///
+/// In all implementations bitwise operators are used instead of logical ones
+/// to avoid branching.
+///
+/// @{
+
+inline constexpr bool operator==(uint128 x, uint128 y) noexcept
+{
+    return ((x[0] ^ y[0]) | (x[1] ^ y[1])) == 0;
+}
+
+inline constexpr bool operator!=(uint128 x, uint128 y) noexcept
+{
+    return !(x == y);
+}
+
+inline constexpr bool operator<(uint128 x, uint128 y) noexcept
+{
+    // OPT: This should be implemented by checking the borrow of x - y,
+    //      but compilers (GCC8, Clang7)
+    //      have problem with properly optimizing subtraction.
+#if INTX_HAS_BUILTIN_INT128
+    return builtin_uint128{x} < builtin_uint128{y};
+#else
+    return (unsigned{x[1] < y[1]} | (unsigned{x[1] == y[1]} & unsigned{x[0] < y[0]})) != 0;
+#endif
+}
+
+inline constexpr bool operator<=(uint128 x, uint128 y) noexcept
+{
+    return !(y < x);
+}
+
+inline constexpr bool operator>(uint128 x, uint128 y) noexcept
+{
+    return y < x;
+}
+
+inline constexpr bool operator>=(uint128 x, uint128 y) noexcept
+{
+    return !(x < y);
+}
+
+/// @}
+
+
+/// Bitwise operators.
+/// @{
+
+inline constexpr uint128 operator~(uint128 x) noexcept
+{
+    return {~x[0], ~x[1]};
+}
+
+inline constexpr uint128 operator|(uint128 x, uint128 y) noexcept
+{
+    // Clang7: perfect.
+    // GCC8: stupidly uses a vector instruction in all bitwise operators.
+    return {x[0] | y[0], x[1] | y[1]};
+}
+
+inline constexpr uint128 operator&(uint128 x, uint128 y) noexcept
+{
+    return {x[0] & y[0], x[1] & y[1]};
+}
+
+inline constexpr uint128 operator^(uint128 x, uint128 y) noexcept
+{
+    return {x[0] ^ y[0], x[1] ^ y[1]};
+}
+
+inline constexpr uint128 operator<<(uint128 x, uint64_t shift) noexcept
+{
+    return (shift < 64) ?
+               // Find the part moved from lo to hi.
+               // For shift == 0 right shift by (64 - shift) is invalid so
+               // split it into 2 shifts by 1 and (63 - shift).
+               uint128{x[0] << shift, (x[1] << shift) | ((x[0] >> 1) >> (63 - shift))} :
+
+               // Guarantee "defined" behavior for shifts larger than 128.
+               (shift < 128) ? uint128{0, x[0] << (shift - 64)} : 0;
+}
+
+inline constexpr uint128 operator<<(uint128 x, uint128 shift) noexcept
+{
+    if (INTX_UNLIKELY(shift[1] != 0))
+        return 0;
+
+    return x << shift[0];
+}
+
+inline constexpr uint128 operator>>(uint128 x, uint64_t shift) noexcept
+{
+    return (shift < 64) ?
+               // Find the part moved from lo to hi.
+               // For shift == 0 left shift by (64 - shift) is invalid so
+               // split it into 2 shifts by 1 and (63 - shift).
+               uint128{(x[0] >> shift) | ((x[1] << 1) << (63 - shift)), x[1] >> shift} :
+
+               // Guarantee "defined" behavior for shifts larger than 128.
+               (shift < 128) ? uint128{x[1] >> (shift - 64)} : 0;
+}
+
+inline constexpr uint128 operator>>(uint128 x, uint128 shift) noexcept
+{
+    if (INTX_UNLIKELY(shift[1] != 0))
+        return 0;
+
+    return x >> static_cast<uint64_t>(shift);
+}
+
+/// @}
+
+
+/// Multiplication
+/// @{
+
+/// Full unsigned multiplication 64 x 64 -> 128.
+inline constexpr uint128 umul(uint64_t x, uint64_t y) noexcept
+{
+#if INTX_HAS_BUILTIN_INT128
+    return builtin_uint128{x} * builtin_uint128{y};
+#elif defined(_MSC_VER) && _MSC_VER >= 1925
+    if (!is_constant_evaluated())
+    {
+        unsigned __int64 hi = 0;
+        const auto lo = _umul128(x, y, &hi);
+        return {lo, hi};
+    }
+    // For constexpr fallback to portable variant.
+#endif
+
+    // Portable full unsigned multiplication 64 x 64 -> 128.
+    uint64_t xl = x & 0xffffffff;
+    uint64_t xh = x >> 32;
+    uint64_t yl = y & 0xffffffff;
+    uint64_t yh = y >> 32;
+
+    uint64_t t0 = xl * yl;
+    uint64_t t1 = xh * yl;
+    uint64_t t2 = xl * yh;
+    uint64_t t3 = xh * yh;
+
+    uint64_t u1 = t1 + (t0 >> 32);
+    uint64_t u2 = t2 + (u1 & 0xffffffff);
+
+    uint64_t lo = (u2 << 32) | (t0 & 0xffffffff);
+    uint64_t hi = t3 + (u2 >> 32) + (u1 >> 32);
+    return {lo, hi};
+}
+
+inline constexpr uint128 operator*(uint128 x, uint128 y) noexcept
+{
+    auto p = umul(x[0], y[0]);
+    p[1] += (x[0] * y[1]) + (x[1] * y[0]);
+    return {p[0], p[1]};
+}
+
+/// @}
+
+
+/// Assignment operators.
+/// @{
+
+inline constexpr uint128& operator+=(uint128& x, uint128 y) noexcept
+{
+    return x = x + y;
+}
+
+inline constexpr uint128& operator-=(uint128& x, uint128 y) noexcept
+{
+    return x = x - y;
+}
+
+inline uint128& operator*=(uint128& x, uint128 y) noexcept
+{
+    return x = x * y;
+}
+
+inline constexpr uint128& operator|=(uint128& x, uint128 y) noexcept
+{
+    return x = x | y;
+}
+
+inline constexpr uint128& operator&=(uint128& x, uint128 y) noexcept
+{
+    return x = x & y;
+}
+
+inline constexpr uint128& operator^=(uint128& x, uint128 y) noexcept
+{
+    return x = x ^ y;
+}
+
+inline constexpr uint128& operator<<=(uint128& x, uint64_t shift) noexcept
+{
+    return x = x << shift;
+}
+
+inline constexpr uint128& operator>>=(uint128& x, uint64_t shift) noexcept
+{
+    return x = x >> shift;
+}
+
+/// @}
+
+
+inline constexpr unsigned clz_generic(uint32_t x) noexcept
+{
+    unsigned n = 32;
+    for (int i = 4; i >= 0; --i)
+    {
+        const auto s = unsigned{1} << i;
+        const auto hi = x >> s;
+        if (hi != 0)
+        {
+            n -= s;
+            x = hi;
+        }
+    }
+    return n - x;
+}
+
+inline constexpr unsigned clz_generic(uint64_t x) noexcept
+{
+    unsigned n = 64;
+    for (int i = 5; i >= 0; --i)
+    {
+        const auto s = unsigned{1} << i;
+        const auto hi = x >> s;
+        if (hi != 0)
+        {
+            n -= s;
+            x = hi;
+        }
+    }
+    return n - static_cast<unsigned>(x);
+}
+
+inline constexpr unsigned clz(uint32_t x) noexcept
+{
+#ifdef _MSC_VER
+    return clz_generic(x);
+#else
+    return x != 0 ? unsigned(__builtin_clz(x)) : 32;
+#endif
+}
+
+inline constexpr unsigned clz(uint64_t x) noexcept
+{
+#ifdef _MSC_VER
+    return clz_generic(x);
+#else
+    return x != 0 ? unsigned(__builtin_clzll(x)) : 64;
+#endif
+}
+
+inline constexpr unsigned clz(uint128 x) noexcept
+{
+    // In this order `h == 0` we get less instructions than in case of `h != 0`.
+    return x[1] == 0 ? clz(x[0]) + 64 : clz(x[1]);
+}
+
+template <typename T>
+T bswap(T x) noexcept = delete;  // Disable type auto promotion
+
+inline constexpr uint8_t bswap(uint8_t x) noexcept
+{
+    return x;
+}
+
+inline constexpr uint16_t bswap(uint16_t x) noexcept
+{
+#if __has_builtin(__builtin_bswap16)
+    return __builtin_bswap16(x);
+#else
+    #ifdef _MSC_VER
+    if (!is_constant_evaluated())
+        return _byteswap_ushort(x);
+    #endif
+    return static_cast<uint16_t>((x << 8) | (x >> 8));
+#endif
+}
+
+inline constexpr uint32_t bswap(uint32_t x) noexcept
+{
+#if __has_builtin(__builtin_bswap32)
+    return __builtin_bswap32(x);
+#else
+    #ifdef _MSC_VER
+    if (!is_constant_evaluated())
+        return _byteswap_ulong(x);
+    #endif
+    const auto a = ((x << 8) & 0xFF00FF00) | ((x >> 8) & 0x00FF00FF);
+    return (a << 16) | (a >> 16);
+#endif
+}
+
+inline constexpr uint64_t bswap(uint64_t x) noexcept
+{
+#if __has_builtin(__builtin_bswap64)
+    return __builtin_bswap64(x);
+#else
+    #ifdef _MSC_VER
+    if (!is_constant_evaluated())
+        return _byteswap_uint64(x);
+    #endif
+    const auto a = ((x << 8) & 0xFF00FF00FF00FF00) | ((x >> 8) & 0x00FF00FF00FF00FF);
+    const auto b = ((a << 16) & 0xFFFF0000FFFF0000) | ((a >> 16) & 0x0000FFFF0000FFFF);
+    return (b << 32) | (b >> 32);
+#endif
+}
+
+inline constexpr uint128 bswap(uint128 x) noexcept
+{
+    return {bswap(x[1]), bswap(x[0])};
+}
+
+
+/// Division.
+/// @{
+
+template <typename QuotT, typename RemT = QuotT>
+struct div_result
+{
+    QuotT quot;
+    RemT rem;
+
+    /// Conversion to tuple of references, to allow usage with std::tie().
+    constexpr operator std::tuple<QuotT&, RemT&>() noexcept { return {quot, rem}; }
+};
+
+namespace internal
+{
+inline constexpr uint16_t reciprocal_table_item(uint8_t d9) noexcept
+{
+    return uint16_t(0x7fd00 / (0x100 | d9));
+}
+
+#define REPEAT4(x)                                                  \
+    reciprocal_table_item((x) + 0), reciprocal_table_item((x) + 1), \
+        reciprocal_table_item((x) + 2), reciprocal_table_item((x) + 3)
+
+#define REPEAT32(x)                                                                         \
+    REPEAT4((x) + 4 * 0), REPEAT4((x) + 4 * 1), REPEAT4((x) + 4 * 2), REPEAT4((x) + 4 * 3), \
+        REPEAT4((x) + 4 * 4), REPEAT4((x) + 4 * 5), REPEAT4((x) + 4 * 6), REPEAT4((x) + 4 * 7)
+
+#define REPEAT256()                                                                           \
+    REPEAT32(32 * 0), REPEAT32(32 * 1), REPEAT32(32 * 2), REPEAT32(32 * 3), REPEAT32(32 * 4), \
+        REPEAT32(32 * 5), REPEAT32(32 * 6), REPEAT32(32 * 7)
+
+/// Reciprocal lookup table.
+constexpr uint16_t reciprocal_table[] = {REPEAT256()};
+
+#undef REPEAT4
+#undef REPEAT32
+#undef REPEAT256
+}  // namespace internal
+
+/// Computes the reciprocal (2^128 - 1) / d - 2^64 for normalized d.
+///
+/// Based on Algorithm 2 from "Improved division by invariant integers".
+inline uint64_t reciprocal_2by1(uint64_t d) noexcept
+{
+    INTX_REQUIRE(d & 0x8000000000000000);  // Must be normalized.
+
+    const uint64_t d9 = d >> 55;
+    const uint32_t v0 = internal::reciprocal_table[d9 - 256];
+
+    const uint64_t d40 = (d >> 24) + 1;
+    const uint64_t v1 = (v0 << 11) - uint32_t(v0 * v0 * d40 >> 40) - 1;
+
+    const uint64_t v2 = (v1 << 13) + (v1 * (0x1000000000000000 - v1 * d40) >> 47);
+
+    const uint64_t d0 = d & 1;
+    const uint64_t d63 = (d >> 1) + d0;  // ceil(d/2)
+    const uint64_t e = ((v2 >> 1) & (0 - d0)) - v2 * d63;
+    const uint64_t v3 = (umul(v2, e)[1] >> 1) + (v2 << 31);
+
+    const uint64_t v4 = v3 - (umul(v3, d) + d)[1] - d;
+    return v4;
+}
+
+inline uint64_t reciprocal_3by2(uint128 d) noexcept
+{
+    auto v = reciprocal_2by1(d[1]);
+    auto p = d[1] * v;
+    p += d[0];
+    if (p < d[0])
+    {
+        --v;
+        if (p >= d[1])
+        {
+            --v;
+            p -= d[1];
+        }
+        p -= d[1];
+    }
+
+    const auto t = umul(v, d[0]);
+
+    p += t[1];
+    if (p < t[1])
+    {
+        --v;
+        if (p >= d[1])
+        {
+            if (p > d[1] || t[0] >= d[0])
+                --v;
+        }
+    }
+    return v;
+}
+
+inline div_result<uint64_t> udivrem_2by1(uint128 u, uint64_t d, uint64_t v) noexcept
+{
+    auto q = umul(v, u[1]);
+    q = fast_add(q, u);
+
+    ++q[1];
+
+    auto r = u[0] - q[1] * d;
+
+    if (r > q[0])
+    {
+        --q[1];
+        r += d;
+    }
+
+    if (r >= d)
+    {
+        ++q[1];
+        r -= d;
+    }
+
+    return {q[1], r};
+}
+
+inline div_result<uint64_t, uint128> udivrem_3by2(
+    uint64_t u2, uint64_t u1, uint64_t u0, uint128 d, uint64_t v) noexcept
+{
+    auto q = umul(v, u2);
+    q = fast_add(q, {u1, u2});
+
+    auto r1 = u1 - q[1] * d[1];
+
+    auto t = umul(d[0], q[1]);
+
+    auto r = uint128{u0, r1} - t - d;
+    r1 = r[1];
+
+    ++q[1];
+
+    if (r1 >= q[0])
+    {
+        --q[1];
+        r += d;
+    }
+
+    if (r >= d)
+    {
+        ++q[1];
+        r -= d;
+    }
+
+    return {q[1], r};
+}
+
+inline div_result<uint128> udivrem(uint128 x, uint128 y) noexcept
+{
+    if (y[1] == 0)
+    {
+        INTX_REQUIRE(y[0] != 0);  // Division by 0.
+
+        const auto lsh = clz(y[0]);
+        const auto rsh = (64 - lsh) % 64;
+        const auto rsh_mask = uint64_t{lsh == 0} - 1;
+
+        const auto yn = y[0] << lsh;
+        const auto xn_lo = x[0] << lsh;
+        const auto xn_hi = (x[1] << lsh) | ((x[0] >> rsh) & rsh_mask);
+        const auto xn_ex = (x[1] >> rsh) & rsh_mask;
+
+        const auto v = reciprocal_2by1(yn);
+        const auto res1 = udivrem_2by1({xn_hi, xn_ex}, yn, v);
+        const auto res2 = udivrem_2by1({xn_lo, res1.rem}, yn, v);
+        return {{res2.quot, res1.quot}, res2.rem >> lsh};
+    }
+
+    if (y[1] > x[1])
+        return {0, x};
+
+    const auto lsh = clz(y[1]);
+    if (lsh == 0)
+    {
+        const auto q = unsigned{y[1] < x[1]} | unsigned{y[0] <= x[0]};
+        return {q, x - (q ? y : 0)};
+    }
+
+    const auto rsh = 64 - lsh;
+
+    const auto yn_lo = y[0] << lsh;
+    const auto yn_hi = (y[1] << lsh) | (y[0] >> rsh);
+    const auto xn_lo = x[0] << lsh;
+    const auto xn_hi = (x[1] << lsh) | (x[0] >> rsh);
+    const auto xn_ex = x[1] >> rsh;
+
+    const auto v = reciprocal_3by2({yn_lo, yn_hi});
+    const auto res = udivrem_3by2(xn_ex, xn_hi, xn_lo, {yn_lo, yn_hi}, v);
+
+    return {res.quot, res.rem >> lsh};
+}
+
+inline div_result<uint128> sdivrem(uint128 x, uint128 y) noexcept
+{
+    constexpr auto sign_mask = uint128{1} << 127;
+    const auto x_is_neg = (x & sign_mask) != 0;
+    const auto y_is_neg = (y & sign_mask) != 0;
+
+    const auto x_abs = x_is_neg ? -x : x;
+    const auto y_abs = y_is_neg ? -y : y;
+
+    const auto q_is_neg = x_is_neg ^ y_is_neg;
+
+    const auto res = udivrem(x_abs, y_abs);
+
+    return {q_is_neg ? -res.quot : res.quot, x_is_neg ? -res.rem : res.rem};
+}
+
+inline uint128 operator/(uint128 x, uint128 y) noexcept
+{
+    return udivrem(x, y).quot;
+}
+
+inline uint128 operator%(uint128 x, uint128 y) noexcept
+{
+    return udivrem(x, y).rem;
+}
+
+inline uint128& operator/=(uint128& x, uint128 y) noexcept
+{
+    return x = x / y;
+}
+
+inline uint128& operator%=(uint128& x, uint128 y) noexcept
+{
+    return x = x % y;
+}
+
+/// @}
+
+}  // namespace intx
+
+
+namespace std
+{
+template <unsigned N>
+struct numeric_limits<intx::uint<N>>
+{
+    using type = intx::uint<N>;
+
+    static constexpr bool is_specialized = true;
+    static constexpr bool is_integer = true;
+    static constexpr bool is_signed = false;
+    static constexpr bool is_exact = true;
+    static constexpr bool has_infinity = false;
+    static constexpr bool has_quiet_NaN = false;
+    static constexpr bool has_signaling_NaN = false;
+    static constexpr float_denorm_style has_denorm = denorm_absent;
+    static constexpr bool has_denorm_loss = false;
+    static constexpr float_round_style round_style = round_toward_zero;
+    static constexpr bool is_iec559 = false;
+    static constexpr bool is_bounded = true;
+    static constexpr bool is_modulo = true;
+    static constexpr int digits = CHAR_BIT * sizeof(type);
+    static constexpr int digits10 = int(0.3010299956639812 * digits);
+    static constexpr int max_digits10 = 0;
+    static constexpr int radix = 2;
+    static constexpr int min_exponent = 0;
+    static constexpr int min_exponent10 = 0;
+    static constexpr int max_exponent = 0;
+    static constexpr int max_exponent10 = 0;
+    static constexpr bool traps = std::numeric_limits<unsigned>::traps;
+    static constexpr bool tinyness_before = false;
+
+    static constexpr type min() noexcept { return 0; }
+    static constexpr type lowest() noexcept { return min(); }
+    static constexpr type max() noexcept { return ~type{0}; }
+    static constexpr type epsilon() noexcept { return 0; }
+    static constexpr type round_error() noexcept { return 0; }
+    static constexpr type infinity() noexcept { return 0; }
+    static constexpr type quiet_NaN() noexcept { return 0; }
+    static constexpr type signaling_NaN() noexcept { return 0; }
+    static constexpr type denorm_min() noexcept { return 0; }
+};
+}  // namespace std
+
+namespace intx
+{
+template <typename T>
+[[noreturn]] inline void throw_(const char* what)
+{
+#if __cpp_exceptions
+    throw T{what};
+#else
+    std::fputs(what, stderr);
+    std::abort();
+#endif
+}
+
+inline constexpr int from_dec_digit(char c)
+{
+    if (c < '0' || c > '9')
+        throw_<std::invalid_argument>("invalid digit");
+    return c - '0';
+}
+
+inline constexpr int from_hex_digit(char c)
+{
+    if (c >= 'a' && c <= 'f')
+        return c - ('a' - 10);
+    if (c >= 'A' && c <= 'F')
+        return c - ('A' - 10);
+    return from_dec_digit(c);
+}
+
+template <typename Int>
+inline constexpr Int from_string(const char* str)
+{
+    auto s = str;
+    auto x = Int{};
+    int num_digits = 0;
+
+    if (s[0] == '0' && s[1] == 'x')
+    {
+        s += 2;
+        while (const auto c = *s++)
+        {
+            if (++num_digits > int{sizeof(x) * 2})
+                throw_<std::out_of_range>(str);
+            x = (x << uint64_t{4}) | from_hex_digit(c);
+        }
+        return x;
+    }
+
+    while (const auto c = *s++)
+    {
+        if (num_digits++ > std::numeric_limits<Int>::digits10)
+            throw_<std::out_of_range>(str);
+
+        const auto d = from_dec_digit(c);
+        x = x * Int{10} + d;
+        if (x < d)
+            throw_<std::out_of_range>(str);
+    }
+    return x;
+}
+
+template <typename Int>
+inline constexpr Int from_string(const std::string& s)
+{
+    return from_string<Int>(s.c_str());
+}
+
+inline constexpr uint128 operator""_u128(const char* s)
+{
+    return from_string<uint128>(s);
+}
+
+template <unsigned N>
+inline std::string to_string(uint<N> x, int base = 10)
+{
+    if (base < 2 || base > 36)
+        throw_<std::invalid_argument>("invalid base");
+
+    if (x == 0)
+        return "0";
+
+    auto s = std::string{};
+    while (x != 0)
+    {
+        // TODO: Use constexpr udivrem_1?
+        const auto res = udivrem(x, uint<N>{base});
+        const auto d = int(res.rem);
+        const auto c = d < 10 ? '0' + d : 'a' + d - 10;
+        s.push_back(char(c));
+        x = res.quot;
+    }
+    std::reverse(s.begin(), s.end());
+    return s;
+}
+
+template <unsigned N>
+inline std::string hex(uint<N> x)
+{
+    return to_string(x, 16);
+}
+
+template <unsigned N>
+struct uint
+{
+    using word_type = uint64_t;
+    static constexpr auto word_num_bits = sizeof(word_type) * 8;
+    static constexpr auto num_bits = N;
+    static constexpr auto num_words = num_bits / word_num_bits;
+
+    static_assert(N >= 2 * word_num_bits, "Number of bits must be at lest 128");
+    static_assert(N % word_num_bits == 0, "Number of bits must be a multiply of 64");
+
+private:
+    uint64_t words_[num_words]{};
+
+public:
+    constexpr uint() noexcept = default;
+
+    /// Implicit converting constructor for any smaller uint type.
+    template <unsigned M, typename = typename std::enable_if_t<(M < N)>>
+    constexpr uint(const uint<M>& x) noexcept
+    {
+        for (size_t i = 0; i < uint<M>::num_words; ++i)
+            words_[i] = x[i];
+    }
+
+    template <typename... T,
+        typename = std::enable_if_t<std::conjunction_v<std::is_convertible<T, uint64_t>...>>>
+    constexpr uint(T... v) noexcept : words_{static_cast<uint64_t>(v)...}
+    {}
+
+    constexpr uint64_t& operator[](size_t i) noexcept { return words_[i]; }
+
+    constexpr const uint64_t& operator[](size_t i) const noexcept { return words_[i]; }
+
+    constexpr explicit operator bool() const noexcept { return *this != uint{}; }
+
+    template <unsigned M, typename = typename std::enable_if_t<(M < N)>>
+    explicit operator uint<M>() const noexcept
+    {
+        uint<M> r;
+        for (size_t i = 0; i < uint<M>::num_words; ++i)
+            r[i] = words_[i];
+        return r;
+    }
+
+    /// Explicit converting operator for all builtin integral types.
+    template <typename Int, typename = typename std::enable_if_t<std::is_integral_v<Int>>>
+    explicit operator Int() const noexcept
+    {
+        static_assert(sizeof(Int) <= sizeof(uint64_t));
+        return static_cast<Int>(words_[0]);
+    }
+};
+
+using uint192 = uint<192>;
+using uint256 = uint<256>;
+using uint320 = uint<320>;
+using uint384 = uint<384>;
+using uint512 = uint<512>;
+
+template <unsigned N>
+inline constexpr bool operator==(const uint<N>& x, const uint<N>& y) noexcept
+{
+    uint64_t folded = 0;
+    for (size_t i = 0; i < uint<N>::num_words; ++i)
+        folded |= (x[i] ^ y[i]);
+    return folded == 0;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator==(const uint<N>& x, const T& y) noexcept
+{
+    return x == uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator==(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(y) == x;
+}
+
+
+template <unsigned N>
+inline constexpr bool operator!=(const uint<N>& x, const uint<N>& y) noexcept
+{
+    return !(x == y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator!=(const uint<N>& x, const T& y) noexcept
+{
+    return x != uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator!=(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) != y;
+}
+
+#if !defined(_MSC_VER) || _MSC_VER < 1916  // This kills MSVC 2017 compiler.
+inline constexpr bool operator<(const uint256& x, const uint256& y) noexcept
+{
+    const auto xhi = uint128{x[2], x[3]};
+    const auto xlo = uint128{x[0], x[1]};
+    const auto yhi = uint128{y[2], y[3]};
+    const auto ylo = uint128{y[0], y[1]};
+    return (unsigned(xhi < yhi) | (unsigned(xhi == yhi) & unsigned(xlo < ylo))) != 0;
+}
+#endif
+
+template <unsigned N>
+inline constexpr bool operator<(const uint<N>& x, const uint<N>& y) noexcept
+{
+    return sub_with_carry(x, y).carry;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator<(const uint<N>& x, const T& y) noexcept
+{
+    return x < uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator<(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) < y;
+}
+
+
+template <unsigned N>
+inline constexpr bool operator>(const uint<N>& x, const uint<N>& y) noexcept
+{
+    return y < x;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator>(const uint<N>& x, const T& y) noexcept
+{
+    return x > uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator>(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) > y;
+}
+
+
+template <unsigned N>
+inline constexpr bool operator>=(const uint<N>& x, const uint<N>& y) noexcept
+{
+    return !(x < y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator>=(const uint<N>& x, const T& y) noexcept
+{
+    return x >= uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator>=(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) >= y;
+}
+
+
+template <unsigned N>
+inline constexpr bool operator<=(const uint<N>& x, const uint<N>& y) noexcept
+{
+    return !(y < x);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator<=(const uint<N>& x, const T& y) noexcept
+{
+    return x <= uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr bool operator<=(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) <= y;
+}
+
+/// Signed less than comparison.
+///
+/// Interprets the arguments as two's complement signed integers
+/// and checks the "less than" relation.
+template <unsigned N>
+inline constexpr bool slt(const uint<N>& x, const uint<N>& y) noexcept
+{
+    constexpr auto top_word_idx = uint<N>::num_words - 1;
+    const auto x_neg = static_cast<int64_t>(x[top_word_idx]) < 0;
+    const auto y_neg = static_cast<int64_t>(y[top_word_idx]) < 0;
+    return ((x_neg ^ y_neg) != 0) ? x_neg : x < y;
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator|(const uint<N>& x, const uint<N>& y) noexcept
+{
+    uint<N> z;
+    for (size_t i = 0; i < uint<N>::num_words; ++i)
+        z[i] = x[i] | y[i];
+    return z;
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator&(const uint<N>& x, const uint<N>& y) noexcept
+{
+    uint<N> z;
+    for (size_t i = 0; i < uint<N>::num_words; ++i)
+        z[i] = x[i] & y[i];
+    return z;
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator^(const uint<N>& x, const uint<N>& y) noexcept
+{
+    uint<N> z;
+    for (size_t i = 0; i < uint<N>::num_words; ++i)
+        z[i] = x[i] ^ y[i];
+    return z;
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator~(const uint<N>& x) noexcept
+{
+    uint<N> z;
+    for (size_t i = 0; i < uint<N>::num_words; ++i)
+        z[i] = ~x[i];
+    return z;
+}
+
+
+inline constexpr uint256 operator<<(const uint256& x, uint64_t shift) noexcept
+{
+    if (INTX_UNLIKELY(shift >= uint256::num_bits))
+        return 0;
+
+    constexpr auto num_bits = uint256::num_bits;
+    constexpr auto half_bits = num_bits / 2;
+
+    const auto xlo = uint128{x[0], x[1]};
+
+    if (shift < half_bits)
+    {
+        const auto lo = xlo << shift;
+
+        const auto xhi = uint128{x[2], x[3]};
+
+        // Find the part moved from lo to hi.
+        // The shift right here can be invalid:
+        // for shift == 0 => rshift == half_bits.
+        // Split it into 2 valid shifts by (rshift - 1) and 1.
+        const auto rshift = half_bits - shift;
+        const auto lo_overflow = (xlo >> (rshift - 1)) >> 1;
+        const auto hi = (xhi << shift) | lo_overflow;
+        return {lo[0], lo[1], hi[0], hi[1]};
+    }
+
+    const auto hi = xlo << (shift - half_bits);
+    return {0, 0, hi[0], hi[1]};
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator<<(const uint<N>& x, uint64_t shift) noexcept
+{
+    if (INTX_UNLIKELY(shift >= uint<N>::num_bits))
+        return 0;
+
+    constexpr auto word_bits = sizeof(uint64_t) * 8;
+
+    const auto s = shift % word_bits;
+    const auto skip = static_cast<size_t>(shift / word_bits);
+
+    uint<N> r;
+    uint64_t carry = 0;
+    for (size_t i = 0; i < (uint<N>::num_words - skip); ++i)
+    {
+        r[i + skip] = (x[i] << s) | carry;
+        carry = (x[i] >> (word_bits - s - 1)) >> 1;
+    }
+    return r;
+}
+
+
+inline constexpr uint256 operator>>(const uint256& x, uint64_t shift) noexcept
+{
+    if (INTX_UNLIKELY(shift >= uint256::num_bits))
+        return 0;
+
+    constexpr auto num_bits = uint256::num_bits;
+    constexpr auto half_bits = num_bits / 2;
+
+    const auto xhi = uint128{x[2], x[3]};
+
+    if (shift < half_bits)
+    {
+        const auto hi = xhi >> shift;
+
+        const auto xlo = uint128{x[0], x[1]};
+
+        // Find the part moved from hi to lo.
+        // The shift left here can be invalid:
+        // for shift == 0 => lshift == half_bits.
+        // Split it into 2 valid shifts by (lshift - 1) and 1.
+        const auto lshift = half_bits - shift;
+        const auto hi_overflow = (xhi << (lshift - 1)) << 1;
+        const auto lo = (xlo >> shift) | hi_overflow;
+        return {lo[0], lo[1], hi[0], hi[1]};
+    }
+
+    const auto lo = xhi >> (shift - half_bits);
+    return {lo[0], lo[1], 0, 0};
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator>>(const uint<N>& x, uint64_t shift) noexcept
+{
+    if (INTX_UNLIKELY(shift >= uint<N>::num_bits))
+        return 0;
+
+    constexpr auto num_words = uint<N>::num_words;
+    constexpr auto word_bits = sizeof(uint64_t) * 8;
+
+    const auto s = shift % word_bits;
+    const auto skip = static_cast<size_t>(shift / word_bits);
+
+    uint<N> r;
+    uint64_t carry = 0;
+    for (size_t i = 0; i < (num_words - skip); ++i)
+    {
+        r[num_words - 1 - i - skip] = (x[num_words - 1 - i] >> s) | carry;
+        carry = (x[num_words - 1 - i] << (word_bits - s - 1)) << 1;
+    }
+    return r;
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator<<(const uint<N>& x, const uint<N>& shift) noexcept
+{
+    uint64_t high_words_fold = 0;
+    for (size_t i = 1; i < uint<N>::num_words; ++i)
+        high_words_fold |= shift[i];
+
+    if (INTX_UNLIKELY(high_words_fold != 0))
+        return 0;
+
+    return x << shift[0];
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator>>(const uint<N>& x, const uint<N>& shift) noexcept
+{
+    uint64_t high_words_fold = 0;
+    for (size_t i = 1; i < uint<N>::num_words; ++i)
+        high_words_fold |= shift[i];
+
+    if (INTX_UNLIKELY(high_words_fold != 0))
+        return 0;
+
+    return x >> shift[0];
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator<<(const uint<N>& x, const T& shift) noexcept
+{
+    if (shift < T{sizeof(x) * 8})
+        return x << static_cast<uint64_t>(shift);
+    return 0;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator>>(const uint<N>& x, const T& shift) noexcept
+{
+    if (shift < T{sizeof(x) * 8})
+        return x >> static_cast<uint64_t>(shift);
+    return 0;
+}
+
+template <unsigned N>
+inline constexpr uint<N>& operator>>=(uint<N>& x, uint64_t shift) noexcept
+{
+    return x = x >> shift;
+}
+
+
+inline constexpr uint64_t* as_words(uint128& x) noexcept
+{
+    return &x[0];
+}
+
+inline constexpr const uint64_t* as_words(const uint128& x) noexcept
+{
+    return &x[0];
+}
+
+template <unsigned N>
+inline constexpr uint64_t* as_words(uint<N>& x) noexcept
+{
+    return &x[0];
+}
+
+template <unsigned N>
+inline constexpr const uint64_t* as_words(const uint<N>& x) noexcept
+{
+    return &x[0];
+}
+
+template <unsigned N>
+inline uint8_t* as_bytes(uint<N>& x) noexcept
+{
+    return reinterpret_cast<uint8_t*>(as_words(x));
+}
+
+template <unsigned N>
+inline const uint8_t* as_bytes(const uint<N>& x) noexcept
+{
+    return reinterpret_cast<const uint8_t*>(as_words(x));
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator+(const uint<N>& x, const uint<N>& y) noexcept
+{
+    return add_with_carry(x, y).value;
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator-(const uint<N>& x) noexcept
+{
+    return ~x + uint<N>{1};
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator-(const uint<N>& x, const uint<N>& y) noexcept
+{
+    return sub_with_carry(x, y).value;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator+=(uint<N>& x, const T& y) noexcept
+{
+    return x = x + y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator-=(uint<N>& x, const T& y) noexcept
+{
+    return x = x - y;
+}
+
+template <unsigned N>
+inline constexpr uint<2 * N> umul(const uint<N>& x, const uint<N>& y) noexcept
+{
+    constexpr auto num_words = uint<N>::num_words;
+
+    uint<2 * N> p;
+    for (size_t j = 0; j < num_words; ++j)
+    {
+        uint64_t k = 0;
+        for (size_t i = 0; i < num_words; ++i)
+        {
+            const auto t = umul(x[i], y[j]) + p[i + j] + k;
+            p[i + j] = t[0];
+            k = t[1];
+        }
+        p[j + num_words] = k;
+    }
+    return p;
+}
+
+/// Multiplication implementation using word access
+/// and discarding the high part of the result product.
+template <unsigned N>
+inline constexpr uint<N> operator*(const uint<N>& x, const uint<N>& y) noexcept
+{
+    constexpr auto num_words = uint<N>::num_words;
+
+    uint<N> p;
+    for (size_t j = 0; j < num_words; j++)
+    {
+        uint64_t k = 0;
+        for (size_t i = 0; i < (num_words - j - 1); i++)
+        {
+            const auto t = umul(x[i], y[j]) + p[i + j] + k;
+            p[i + j] = t[0];
+            k = t[1];
+        }
+        p[num_words - 1] += x[num_words - j - 1] * y[j] + k;
+    }
+    return p;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator*=(uint<N>& x, const T& y) noexcept
+{
+    return x = x * y;
+}
+
+template <unsigned N>
+inline constexpr uint<N> exp(uint<N> base, uint<N> exponent) noexcept
+{
+    auto result = uint<N>{1};
+    if (base == 2)
+        return result << exponent;
+
+    while (exponent != 0)
+    {
+        if ((exponent & 1) != 0)
+            result *= base;
+        base *= base;
+        exponent >>= 1;
+    }
+    return result;
+}
+
+template <unsigned N>
+inline constexpr unsigned count_significant_words(const uint<N>& x) noexcept
+{
+    for (size_t i = uint<N>::num_words; i > 0; --i)
+    {
+        if (x[i - 1] != 0)
+            return static_cast<unsigned>(i);
+    }
+    return 0;
+}
+
+inline constexpr unsigned count_significant_bytes(uint64_t x) noexcept
+{
+    return (64 - clz(x) + 7) / 8;
+}
+
+template <unsigned N>
+inline constexpr unsigned count_significant_bytes(const uint<N>& x) noexcept
+{
+    const auto w = count_significant_words(x);
+    return (w != 0) ? count_significant_bytes(x[w - 1]) + (w - 1) * 8 : 0;
+}
+
+template <unsigned N>
+inline constexpr unsigned clz(const uint<N>& x) noexcept
+{
+    constexpr unsigned num_words = uint<N>::num_words;
+    const auto s = count_significant_words(x);
+    if (s == 0)
+        return num_words * 64;
+    return clz(x[s - 1]) + (num_words - s) * 64;
+}
+
+namespace internal
+{
+/// Counts the number of zero leading bits in nonzero argument x.
+inline constexpr unsigned clz_nonzero(uint64_t x) noexcept
+{
+    INTX_REQUIRE(x != 0);
+#ifdef _MSC_VER
+    return clz_generic(x);
+#else
+    return unsigned(__builtin_clzll(x));
+#endif
+}
+
+template <unsigned M, unsigned N>
+struct normalized_div_args
+{
+    uint<N> divisor;
+    uint<M + 64> numerator;
+    int num_divisor_words;
+    int num_numerator_words;
+    unsigned shift;
+};
+
+template <unsigned M, unsigned N>
+[[gnu::always_inline]] inline normalized_div_args<M, N> normalize(
+    const uint<M>& numerator, const uint<N>& denominator) noexcept
+{
+    // FIXME: Make the implementation type independent
+    static constexpr auto num_numerator_words = uint<M>::num_words;
+    static constexpr auto num_denominator_words = uint<N>::num_words;
+
+    auto* u = as_words(numerator);
+    auto* v = as_words(denominator);
+
+    normalized_div_args<M, N> na;
+    auto* un = as_words(na.numerator);
+    auto* vn = as_words(na.divisor);
+
+    auto& m = na.num_numerator_words;
+    for (m = num_numerator_words; m > 0 && u[m - 1] == 0; --m)
+        ;
+
+    auto& n = na.num_divisor_words;
+    for (n = num_denominator_words; n > 0 && v[n - 1] == 0; --n)
+        ;
+
+    na.shift = clz_nonzero(v[n - 1]);  // Use clz_nonzero() to avoid clang analyzer's warning.
+    if (na.shift)
+    {
+        for (int i = num_denominator_words - 1; i > 0; --i)
+            vn[i] = (v[i] << na.shift) | (v[i - 1] >> (64 - na.shift));
+        vn[0] = v[0] << na.shift;
+
+        un[num_numerator_words] = u[num_numerator_words - 1] >> (64 - na.shift);
+        for (int i = num_numerator_words - 1; i > 0; --i)
+            un[i] = (u[i] << na.shift) | (u[i - 1] >> (64 - na.shift));
+        un[0] = u[0] << na.shift;
+    }
+    else
+    {
+        na.numerator = numerator;
+        na.divisor = denominator;
+    }
+
+    // Skip the highest word of numerator if not significant.
+    if (un[m] != 0 || un[m - 1] >= vn[n - 1])
+        ++m;
+
+    return na;
+}
+
+/// Divides arbitrary long unsigned integer by 64-bit unsigned integer (1 word).
+/// @param u    The array of a normalized numerator words. It will contain
+///             the quotient after execution.
+/// @param len  The number of numerator words.
+/// @param d    The normalized divisor.
+/// @return     The remainder.
+inline uint64_t udivrem_by1(uint64_t u[], int len, uint64_t d) noexcept
+{
+    INTX_REQUIRE(len >= 2);
+
+    const auto reciprocal = reciprocal_2by1(d);
+
+    auto rem = u[len - 1];  // Set the top word as remainder.
+    u[len - 1] = 0;         // Reset the word being a part of the result quotient.
+
+    auto it = &u[len - 2];
+    do
+    {
+        std::tie(*it, rem) = udivrem_2by1({*it, rem}, d, reciprocal);
+    } while (it-- != &u[0]);
+
+    return rem;
+}
+
+/// Divides arbitrary long unsigned integer by 128-bit unsigned integer (2 words).
+/// @param u    The array of a normalized numerator words. It will contain the
+///             quotient after execution.
+/// @param len  The number of numerator words.
+/// @param d    The normalized divisor.
+/// @return     The remainder.
+inline uint128 udivrem_by2(uint64_t u[], int len, uint128 d) noexcept
+{
+    INTX_REQUIRE(len >= 3);
+
+    const auto reciprocal = reciprocal_3by2(d);
+
+    auto rem = uint128{u[len - 2], u[len - 1]};  // Set the 2 top words as remainder.
+    u[len - 1] = u[len - 2] = 0;  // Reset these words being a part of the result quotient.
+
+    auto it = &u[len - 3];
+    do
+    {
+        std::tie(*it, rem) = udivrem_3by2(rem[1], rem[0], *it, d, reciprocal);
+    } while (it-- != &u[0]);
+
+    return rem;
+}
+
+/// s = x + y.
+inline bool add(uint64_t s[], const uint64_t x[], const uint64_t y[], int len) noexcept
+{
+    // OPT: Add MinLen template parameter and unroll first loop iterations.
+    INTX_REQUIRE(len >= 2);
+
+    bool carry = false;
+    for (int i = 0; i < len; ++i)
+        std::tie(s[i], carry) = add_with_carry(x[i], y[i], carry);
+    return carry;
+}
+
+/// r = x - multiplier * y.
+inline uint64_t submul(
+    uint64_t r[], const uint64_t x[], const uint64_t y[], int len, uint64_t multiplier) noexcept
+{
+    // OPT: Add MinLen template parameter and unroll first loop iterations.
+    INTX_REQUIRE(len >= 1);
+
+    uint64_t borrow = 0;
+    for (int i = 0; i < len; ++i)
+    {
+        const auto s = sub_with_carry(x[i], borrow);
+        const auto p = umul(y[i], multiplier);
+        const auto t = sub_with_carry(s.value, p[0]);
+        r[i] = t.value;
+        borrow = p[1] + s.carry + t.carry;
+    }
+    return borrow;
+}
+
+inline void udivrem_knuth(
+    uint64_t q[], uint64_t u[], int ulen, const uint64_t d[], int dlen) noexcept
+{
+    INTX_REQUIRE(dlen >= 3);
+    INTX_REQUIRE(ulen >= dlen);
+
+    const auto divisor = uint128{d[dlen - 2], d[dlen - 1]};
+    const auto reciprocal = reciprocal_3by2(divisor);
+    for (int j = ulen - dlen - 1; j >= 0; --j)
+    {
+        const auto u2 = u[j + dlen];
+        const auto u1 = u[j + dlen - 1];
+        const auto u0 = u[j + dlen - 2];
+
+        uint64_t qhat;
+        if (INTX_UNLIKELY((uint128{u1, u2}) == divisor))  // Division overflows.
+        {
+            qhat = ~uint64_t{0};
+
+            u[j + dlen] = u2 - submul(&u[j], &u[j], d, dlen, qhat);
+        }
+        else
+        {
+            uint128 rhat;
+            std::tie(qhat, rhat) = udivrem_3by2(u2, u1, u0, divisor, reciprocal);
+
+            bool carry;
+            const auto overflow = submul(&u[j], &u[j], d, dlen - 2, qhat);
+            std::tie(u[j + dlen - 2], carry) = sub_with_carry(rhat[0], overflow);
+            std::tie(u[j + dlen - 1], carry) = sub_with_carry(rhat[1], carry);
+
+            if (INTX_UNLIKELY(carry))
+            {
+                --qhat;
+                u[j + dlen - 1] += divisor[1] + add(&u[j], &u[j], d, dlen - 1);
+            }
+        }
+
+        q[j] = qhat;  // Store quotient digit.
+    }
+}
+
+}  // namespace internal
+
+template <unsigned M, unsigned N>
+div_result<uint<M>, uint<N>> udivrem(const uint<M>& u, const uint<N>& v) noexcept
+{
+    auto na = internal::normalize(u, v);
+
+    if (na.num_numerator_words <= na.num_divisor_words)
+        return {0, static_cast<uint<N>>(u)};
+
+    if (na.num_divisor_words == 1)
+    {
+        const auto r = internal::udivrem_by1(
+            as_words(na.numerator), na.num_numerator_words, as_words(na.divisor)[0]);
+        return {static_cast<uint<M>>(na.numerator), r >> na.shift};
+    }
+
+    if (na.num_divisor_words == 2)
+    {
+        const auto d = as_words(na.divisor);
+        const auto r =
+            internal::udivrem_by2(as_words(na.numerator), na.num_numerator_words, {d[0], d[1]});
+        return {static_cast<uint<M>>(na.numerator), r >> na.shift};
+    }
+
+    auto un = as_words(na.numerator);  // Will be modified.
+
+    uint<M> q;
+    internal::udivrem_knuth(
+        as_words(q), &un[0], na.num_numerator_words, as_words(na.divisor), na.num_divisor_words);
+
+    uint<N> r;
+    auto rw = as_words(r);
+    for (int i = 0; i < na.num_divisor_words - 1; ++i)
+        rw[i] = na.shift ? (un[i] >> na.shift) | (un[i + 1] << (64 - na.shift)) : un[i];
+    rw[na.num_divisor_words - 1] = un[na.num_divisor_words - 1] >> na.shift;
+
+    return {q, r};
+}
+
+template <unsigned N>
+inline constexpr div_result<uint<N>> sdivrem(const uint<N>& u, const uint<N>& v) noexcept
+{
+    const auto sign_mask = uint<N>{1} << (sizeof(u) * 8 - 1);
+    auto u_is_neg = (u & sign_mask) != 0;
+    auto v_is_neg = (v & sign_mask) != 0;
+
+    auto u_abs = u_is_neg ? -u : u;
+    auto v_abs = v_is_neg ? -v : v;
+
+    auto q_is_neg = u_is_neg ^ v_is_neg;
+
+    auto res = udivrem(u_abs, v_abs);
+
+    return {q_is_neg ? -res.quot : res.quot, u_is_neg ? -res.rem : res.rem};
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator/(const uint<N>& x, const uint<N>& y) noexcept
+{
+    return udivrem(x, y).quot;
+}
+
+template <unsigned N>
+inline constexpr uint<N> operator%(const uint<N>& x, const uint<N>& y) noexcept
+{
+    return udivrem(x, y).rem;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator/=(uint<N>& x, const T& y) noexcept
+{
+    return x = x / y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator%=(uint<N>& x, const T& y) noexcept
+{
+    return x = x % y;
+}
+
+template <unsigned N>
+inline constexpr uint<N> bswap(const uint<N>& x) noexcept
+{
+    constexpr auto num_words = uint<N>::num_words;
+    uint<N> z;
+    for (size_t i = 0; i < num_words; ++i)
+        z[num_words - 1 - i] = bswap(x[i]);
+    return z;
+}
+
+
+// Support for type conversions for binary operators.
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator+(const uint<N>& x, const T& y) noexcept
+{
+    return x + uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator+(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) + y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator-(const uint<N>& x, const T& y) noexcept
+{
+    return x - uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator-(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) - y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator*(const uint<N>& x, const T& y) noexcept
+{
+    return x * uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator*(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) * y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator/(const uint<N>& x, const T& y) noexcept
+{
+    return x / uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator/(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) / y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator%(const uint<N>& x, const T& y) noexcept
+{
+    return x % uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator%(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) % y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator|(const uint<N>& x, const T& y) noexcept
+{
+    return x | uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator|(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) | y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator&(const uint<N>& x, const T& y) noexcept
+{
+    return x & uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator&(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) & y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator^(const uint<N>& x, const T& y) noexcept
+{
+    return x ^ uint<N>(y);
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N> operator^(const T& x, const uint<N>& y) noexcept
+{
+    return uint<N>(x) ^ y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator|=(uint<N>& x, const T& y) noexcept
+{
+    return x = x | y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator&=(uint<N>& x, const T& y) noexcept
+{
+    return x = x & y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator^=(uint<N>& x, const T& y) noexcept
+{
+    return x = x ^ y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator<<=(uint<N>& x, const T& y) noexcept
+{
+    return x = x << y;
+}
+
+template <unsigned N, typename T,
+    typename = typename std::enable_if<std::is_convertible<T, uint<N>>::value>::type>
+inline constexpr uint<N>& operator>>=(uint<N>& x, const T& y) noexcept
+{
+    return x = x >> y;
+}
+
+
+inline uint256 addmod(const uint256& x, const uint256& y, const uint256& mod) noexcept
+{
+    // Fast path for mod >= 2^192, with x and y at most slightly bigger than mod.
+    // This is always the case when x and y are already reduced modulo mod.
+    // Based on https://github.com/holiman/uint256/pull/86.
+    if ((mod[3] != 0) && (x[3] <= mod[3]) && (y[3] <= mod[3]))
+    {
+        auto s = sub_with_carry(x, mod);
+        if (s.carry)
+            s.value = x;
+
+        auto t = sub_with_carry(y, mod);
+        if (t.carry)
+            t.value = y;
+
+        s = add_with_carry(s.value, t.value);
+        t = sub_with_carry(s.value, mod);
+        return (s.carry || !t.carry) ? t.value : s.value;
+    }
+
+    const auto s = add_with_carry(x, y);
+    uint<256 + 64> n = s.value;
+    n[4] = s.carry;
+    return udivrem(n, mod).rem;
+}
+
+inline uint256 mulmod(const uint256& x, const uint256& y, const uint256& mod) noexcept
+{
+    return udivrem(umul(x, y), mod).rem;
+}
+
+
+inline constexpr uint256 operator"" _u256(const char* s)
+{
+    return from_string<uint256>(s);
+}
+
+inline constexpr uint512 operator"" _u512(const char* s)
+{
+    return from_string<uint512>(s);
+}
+
+namespace le  // Conversions to/from LE bytes.
+{
+template <typename IntT, unsigned M>
+inline IntT load(const uint8_t (&bytes)[M]) noexcept
+{
+    static_assert(M == IntT::num_bits / 8,
+        "the size of source bytes must match the size of the destination uint");
+    auto x = IntT{};
+    std::memcpy(&x, bytes, sizeof(x));
+    return x;
+}
+
+template <unsigned N>
+inline void store(uint8_t (&dst)[N / 8], const intx::uint<N>& x) noexcept
+{
+    std::memcpy(dst, &x, sizeof(x));
+}
+
+}  // namespace le
+
+
+namespace be  // Conversions to/from BE bytes.
+{
+/// Loads an uint value from bytes of big-endian order.
+/// If the size of bytes is smaller than the result uint, the value is zero-extended.
+template <typename IntT, unsigned M>
+inline IntT load(const uint8_t (&bytes)[M]) noexcept
+{
+    static_assert(M <= IntT::num_bits / 8,
+        "the size of source bytes must not exceed the size of the destination uint");
+    auto x = IntT{};
+    std::memcpy(&as_bytes(x)[IntT::num_bits / 8 - M], bytes, M);
+    return bswap(x);
+}
+
+template <typename IntT, typename T>
+inline IntT load(const T& t) noexcept
+{
+    return load<IntT>(t.bytes);
+}
+
+/// Stores an uint value in a bytes array in big-endian order.
+template <unsigned N>
+inline void store(uint8_t (&dst)[N / 8], const intx::uint<N>& x) noexcept
+{
+    const auto d = bswap(x);
+    std::memcpy(dst, &d, sizeof(d));
+}
+
+/// Stores an uint value in .bytes field of type T. The .bytes must be an array of uint8_t
+/// of the size matching the size of uint.
+template <typename T, unsigned N>
+inline T store(const intx::uint<N>& x) noexcept
+{
+    T r{};
+    store(r.bytes, x);
+    return r;
+}
+
+/// Stores the truncated value of an uint in a bytes array.
+/// Only the least significant bytes from big-endian representation of the uint
+/// are stored in the result bytes array up to array's size.
+template <unsigned M, unsigned N>
+inline void trunc(uint8_t (&dst)[M], const intx::uint<N>& x) noexcept
+{
+    static_assert(M < N / 8, "destination must be smaller than the source value");
+    const auto d = bswap(x);
+    const auto b = as_bytes(d);
+    std::memcpy(dst, &b[sizeof(d) - M], M);
+}
+
+/// Stores the truncated value of an uint in the .bytes field of an object of type T.
+template <typename T, unsigned N>
+inline T trunc(const intx::uint<N>& x) noexcept
+{
+    T r{};
+    trunc(r.bytes, x);
+    return r;
+}
+
+namespace unsafe
+{
+/// Loads an uint value from a buffer. The user must make sure
+/// that the provided buffer is big enough. Therefore marked "unsafe".
+template <typename IntT>
+inline IntT load(const uint8_t* bytes) noexcept
+{
+    auto x = IntT{};
+    std::memcpy(&x, bytes, sizeof(x));
+    return bswap(x);
+}
+
+/// Stores an uint value at the provided pointer in big-endian order. The user must make sure
+/// that the provided buffer is big enough to fit the value. Therefore marked "unsafe".
+template <unsigned N>
+inline void store(uint8_t* dst, const intx::uint<N>& x) noexcept
+{
+    const auto d = bswap(x);
+    std::memcpy(dst, &d, sizeof(d));
+}
+}  // namespace unsafe
+
+}  // namespace be
+
+}  // namespace intx
diff --git a/src/third_party_lib/stb/stb_image.h b/src/third_party_lib/stb/stb_image.h
index 1b7e7b02..d60371b9 100644
--- a/src/third_party_lib/stb/stb_image.h
+++ b/src/third_party_lib/stb/stb_image.h
@@ -1,4 +1,4 @@
-/* stb_image - v2.26 - public domain image loader - http://nothings.org/stb
+/* stb_image - v2.27 - public domain image loader - http://nothings.org/stb
                                   no warranty implied; use at your own risk
 
    Do this:
@@ -48,6 +48,7 @@ LICENSE
 
 RECENT REVISION HISTORY:
 
+      2.27  (2021-07-11) document stbi_info better, 16-bit PNM support, bug fixes
       2.26  (2020-07-13) many minor fixes
       2.25  (2020-02-02) fix warnings
       2.24  (2020-02-02) fix warnings; thread-local failure_reason and flip_vertically
@@ -89,7 +90,7 @@ RECENT REVISION HISTORY:
                                            Jeremy Sawicki (handle all ImageNet JPGs)
  Optimizations & bugfixes                  Mikhail Morozov (1-bit BMP)
     Fabian "ryg" Giesen                    Anael Seghezzi (is-16-bit query)
-    Arseny Kapoulkine
+    Arseny Kapoulkine                      Simon Breuss (16-bit PNM)
     John-Mark Allen
     Carmelo J Fdez-Aguera
 
@@ -102,7 +103,7 @@ RECENT REVISION HISTORY:
     Thomas Ruf              Ronny Chevalier                         github:rlyeh
     Janez Zemva             John Bartholomew   Michal Cichon        github:romigrou
     Jonathan Blow           Ken Hamada         Tero Hanninen        github:svdijk
-                            Laurent Gomila     Cort Stratton        github:snagar
+    Eugene Golushkov        Laurent Gomila     Cort Stratton        github:snagar
     Aruelien Pocheville     Sergio Gonzalez    Thibault Reuille     github:Zelex
     Cass Everitt            Ryamond Barbiero                        github:grim210
     Paul Du Bois            Engin Manap        Aldo Culquicondor    github:sammyhw
@@ -110,11 +111,13 @@ RECENT REVISION HISTORY:
     Josh Tobin                                 Matthew Gregan       github:poppolopoppo
     Julian Raschke          Gregory Mullen     Christian Floisand   github:darealshinji
     Baldur Karlsson         Kevin Schmidt      JR Smith             github:Michaelangel007
-                            Brad Weinberger    Matvey Cherevko      [reserved]
+                            Brad Weinberger    Matvey Cherevko      github:mosra
     Luca Sas                Alexander Veselov  Zack Middleton       [reserved]
     Ryan C. Gordon          [reserved]                              [reserved]
                      DO NOT ADD YOUR NAME HERE
 
+                     Jacko Dirks
+
   To add your name to the credits, pick a random blank space in the middle and fill it.
   80% of merge conflicts on stb PRs are due to people adding their name at the end
   of the credits.
@@ -176,6 +179,32 @@ RECENT REVISION HISTORY:
 //
 // Paletted PNG, BMP, GIF, and PIC images are automatically depalettized.
 //
+// To query the width, height and component count of an image without having to
+// decode the full file, you can use the stbi_info family of functions:
+//
+//   int x,y,n,ok;
+//   ok = stbi_info(filename, &x, &y, &n);
+//   // returns ok=1 and sets x, y, n if image is a supported format,
+//   // 0 otherwise.
+//
+// Note that stb_image pervasively uses ints in its public API for sizes,
+// including sizes of memory buffers. This is now part of the API and thus
+// hard to change without causing breakage. As a result, the various image
+// loaders all have certain limits on image size; these differ somewhat
+// by format but generally boil down to either just under 2GB or just under
+// 1GB. When the decoded image would be larger than this, stb_image decoding
+// will fail.
+//
+// Additionally, stb_image will reject image files that have any of their
+// dimensions set to a larger value than the configurable STBI_MAX_DIMENSIONS,
+// which defaults to 2**24 = 16777216 pixels. Due to the above memory limit,
+// the only way to have an image with such dimensions load correctly
+// is for it to have a rather extreme aspect ratio. Either way, the
+// assumption here is that such larger images are likely to be malformed
+// or malicious. If you do need to load an image with individual dimensions
+// larger than that, and it still fits in the overall size limit, you can
+// #define STBI_MAX_DIMENSIONS on your own to be something larger.
+//
 // ===========================================================================
 //
 // UNICODE:
@@ -281,11 +310,10 @@ RECENT REVISION HISTORY:
 //
 // iPhone PNG support:
 //
-// By default we convert iphone-formatted PNGs back to RGB, even though
-// they are internally encoded differently. You can disable this conversion
-// by calling stbi_convert_iphone_png_to_rgb(0), in which case
-// you will always just get the native iphone "format" through (which
-// is BGR stored in RGB).
+// We optionally support converting iPhone-formatted PNGs (which store
+// premultiplied BGRA) back to RGB, even though they're internally encoded
+// differently. To enable this conversion, call
+// stbi_convert_iphone_png_to_rgb(1).
 //
 // Call stbi_set_unpremultiply_on_load(1) as well to force a divide per
 // pixel to remove any premultiplied alpha *only* if the image file explicitly
@@ -489,6 +517,8 @@ STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip);
 // as above, but only applies to images loaded on the thread that calls the function
 // this function is only available if your compiler supports thread-local variables;
 // calling it will fail to link if your compiler doesn't
+STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply);
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert);
 STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip);
 
 // ZLIB client - used by PNG, available for other purposes
@@ -634,7 +664,7 @@ typedef unsigned char validate_uint32[sizeof(stbi__uint32)==4 ? 1 : -1];
 #ifdef STBI_HAS_LROTL
    #define stbi_lrot(x,y)  _lrotl(x,y)
 #else
-   #define stbi_lrot(x,y)  (((x) << (y)) | ((x) >> (32 - (y))))
+   #define stbi_lrot(x,y)  (((x) << (y)) | ((x) >> (-(y) & 31)))
 #endif
 
 #if defined(STBI_MALLOC) && defined(STBI_FREE) && (defined(STBI_REALLOC) || defined(STBI_REALLOC_SIZED))
@@ -748,9 +778,12 @@ static int stbi__sse2_available(void)
 
 #ifdef STBI_NEON
 #include <arm_neon.h>
-// assume GCC or Clang on ARM targets
+#ifdef _MSC_VER
+#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
+#else
 #define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
 #endif
+#endif
 
 #ifndef STBI_SIMD_ALIGN
 #define STBI_SIMD_ALIGN(type, name) type name
@@ -924,6 +957,7 @@ static int      stbi__gif_info(stbi__context *s, int *x, int *y, int *comp);
 static int      stbi__pnm_test(stbi__context *s);
 static void    *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
 static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__pnm_is16(stbi__context *s);
 #endif
 
 static
@@ -998,7 +1032,7 @@ static int stbi__mad3sizes_valid(int a, int b, int c, int add)
 }
 
 // returns 1 if "a*b*c*d + add" has no negative terms/factors and doesn't overflow
-#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR)
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
 static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add)
 {
    return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) &&
@@ -1021,7 +1055,7 @@ static void *stbi__malloc_mad3(int a, int b, int c, int add)
    return stbi__malloc(a*b*c + add);
 }
 
-#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR)
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
 static void *stbi__malloc_mad4(int a, int b, int c, int d, int add)
 {
    if (!stbi__mad4sizes_valid(a, b, c, d, add)) return NULL;
@@ -1087,9 +1121,8 @@ static void *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int re
    ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order
    ri->num_channels = 0;
 
-   #ifndef STBI_NO_JPEG
-   if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp, ri);
-   #endif
+   // test the formats with a very explicit header first (at least a FOURCC
+   // or distinctive magic number first)
    #ifndef STBI_NO_PNG
    if (stbi__png_test(s))  return stbi__png_load(s,x,y,comp,req_comp, ri);
    #endif
@@ -1107,6 +1140,13 @@ static void *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int re
    #ifndef STBI_NO_PIC
    if (stbi__pic_test(s))  return stbi__pic_load(s,x,y,comp,req_comp, ri);
    #endif
+
+   // then the formats that can end up attempting to load with just 1 or 2
+   // bytes matching expectations; these are prone to false positives, so
+   // try them later
+   #ifndef STBI_NO_JPEG
+   if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp, ri);
+   #endif
    #ifndef STBI_NO_PNM
    if (stbi__pnm_test(s))  return stbi__pnm_load(s,x,y,comp,req_comp, ri);
    #endif
@@ -1262,12 +1302,12 @@ static void stbi__float_postprocess(float *result, int *x, int *y, int *comp, in
 
 #ifndef STBI_NO_STDIO
 
-#if defined(_MSC_VER) && defined(STBI_WINDOWS_UTF8)
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
 STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide);
 STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default);
 #endif
 
-#if defined(_MSC_VER) && defined(STBI_WINDOWS_UTF8)
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
 STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input)
 {
 	return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL);
@@ -1277,16 +1317,16 @@ STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wch
 static FILE *stbi__fopen(char const *filename, char const *mode)
 {
    FILE *f;
-#if defined(_MSC_VER) && defined(STBI_WINDOWS_UTF8)
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
    wchar_t wMode[64];
    wchar_t wFilename[1024];
-	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)))
+	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename)))
       return 0;
 
-	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)))
+	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode)))
       return 0;
 
-#if _MSC_VER >= 1400
+#if defined(_MSC_VER) && _MSC_VER >= 1400
 	if (0 != _wfopen_s(&f, wFilename, wMode))
 		f = 0;
 #else
@@ -1662,7 +1702,8 @@ static int stbi__get16le(stbi__context *s)
 static stbi__uint32 stbi__get32le(stbi__context *s)
 {
    stbi__uint32 z = stbi__get16le(s);
-   return z + (stbi__get16le(s) << 16);
+   z += (stbi__uint32)stbi__get16le(s) << 16;
+   return z;
 }
 #endif
 
@@ -2090,13 +2131,12 @@ stbi_inline static int stbi__extend_receive(stbi__jpeg *j, int n)
    int sgn;
    if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
 
-   sgn = (stbi__int32)j->code_buffer >> 31; // sign bit is always in MSB
+   sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative)
    k = stbi_lrot(j->code_buffer, n);
-   if (n < 0 || n >= (int) (sizeof(stbi__bmask)/sizeof(*stbi__bmask))) return 0;
    j->code_buffer = k & ~stbi__bmask[n];
    k &= stbi__bmask[n];
    j->code_bits -= n;
-   return k + (stbi__jbias[n] & ~sgn);
+   return k + (stbi__jbias[n] & (sgn - 1));
 }
 
 // get some unsigned bits
@@ -2146,7 +2186,7 @@ static int stbi__jpeg_decode_block(stbi__jpeg *j, short data[64], stbi__huffman
 
    if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
    t = stbi__jpeg_huff_decode(j, hdc);
-   if (t < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+   if (t < 0 || t > 15) return stbi__err("bad huffman code","Corrupt JPEG");
 
    // 0 all the ac values now so we can do it 32-bits at a time
    memset(data,0,64*sizeof(data[0]));
@@ -2203,12 +2243,12 @@ static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg *j, short data[64], stbi__
       // first scan for DC coefficient, must be first
       memset(data,0,64*sizeof(data[0])); // 0 all the ac values now
       t = stbi__jpeg_huff_decode(j, hdc);
-      if (t == -1) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+      if (t < 0 || t > 15) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
       diff = t ? stbi__extend_receive(j, t) : 0;
 
       dc = j->img_comp[b].dc_pred + diff;
       j->img_comp[b].dc_pred = dc;
-      data[0] = (short) (dc << j->succ_low);
+      data[0] = (short) (dc * (1 << j->succ_low));
    } else {
       // refinement scan for DC coefficient
       if (stbi__jpeg_get_bit(j))
@@ -2245,7 +2285,7 @@ static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__
             j->code_buffer <<= s;
             j->code_bits -= s;
             zig = stbi__jpeg_dezigzag[k++];
-            data[zig] = (short) ((r >> 8) << shift);
+            data[zig] = (short) ((r >> 8) * (1 << shift));
          } else {
             int rs = stbi__jpeg_huff_decode(j, hac);
             if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
@@ -2263,7 +2303,7 @@ static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__
             } else {
                k += r;
                zig = stbi__jpeg_dezigzag[k++];
-               data[zig] = (short) (stbi__extend_receive(j,s) << shift);
+               data[zig] = (short) (stbi__extend_receive(j,s) * (1 << shift));
             }
          }
       } while (k <= j->spec_end);
@@ -3227,6 +3267,13 @@ static int stbi__process_frame_header(stbi__jpeg *z, int scan)
       if (z->img_comp[i].v > v_max) v_max = z->img_comp[i].v;
    }
 
+   // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios
+   // and I've never seen a non-corrupted JPEG file actually use them
+   for (i=0; i < s->img_n; ++i) {
+      if (h_max % z->img_comp[i].h != 0) return stbi__err("bad H","Corrupt JPEG");
+      if (v_max % z->img_comp[i].v != 0) return stbi__err("bad V","Corrupt JPEG");
+   }
+
    // compute interleaved mcu info
    z->img_h_max = h_max;
    z->img_v_max = v_max;
@@ -3782,6 +3829,10 @@ static stbi_uc *load_jpeg_image(stbi__jpeg *z, int *out_x, int *out_y, int *comp
    else
       decode_n = z->s->img_n;
 
+   // nothing to do if no components requested; check this now to avoid
+   // accessing uninitialized coutput[0] later
+   if (decode_n <= 0) { stbi__cleanup_jpeg(z); return NULL; }
+
    // resample and color-convert
    {
       int k;
@@ -3924,6 +3975,7 @@ static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int re
 {
    unsigned char* result;
    stbi__jpeg* j = (stbi__jpeg*) stbi__malloc(sizeof(stbi__jpeg));
+   if (!j) return stbi__errpuc("outofmem", "Out of memory");
    STBI_NOTUSED(ri);
    j->s = s;
    stbi__setup_jpeg(j);
@@ -3936,6 +3988,7 @@ static int stbi__jpeg_test(stbi__context *s)
 {
    int r;
    stbi__jpeg* j = (stbi__jpeg*)stbi__malloc(sizeof(stbi__jpeg));
+   if (!j) return stbi__err("outofmem", "Out of memory");
    j->s = s;
    stbi__setup_jpeg(j);
    r = stbi__decode_jpeg_header(j, STBI__SCAN_type);
@@ -3960,6 +4013,7 @@ static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp)
 {
    int result;
    stbi__jpeg* j = (stbi__jpeg*) (stbi__malloc(sizeof(stbi__jpeg)));
+   if (!j) return stbi__err("outofmem", "Out of memory");
    j->s = s;
    result = stbi__jpeg_info_raw(j, x, y, comp);
    STBI_FREE(j);
@@ -3979,6 +4033,7 @@ static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp)
 // fast-way is faster to check than jpeg huffman, but slow way is slower
 #define STBI__ZFAST_BITS  9 // accelerate all cases in default tables
 #define STBI__ZFAST_MASK  ((1 << STBI__ZFAST_BITS) - 1)
+#define STBI__ZNSYMS 288 // number of symbols in literal/length alphabet
 
 // zlib-style huffman encoding
 // (jpegs packs from left, zlib from right, so can't share code)
@@ -3988,8 +4043,8 @@ typedef struct
    stbi__uint16 firstcode[16];
    int maxcode[17];
    stbi__uint16 firstsymbol[16];
-   stbi_uc  size[288];
-   stbi__uint16 value[288];
+   stbi_uc  size[STBI__ZNSYMS];
+   stbi__uint16 value[STBI__ZNSYMS];
 } stbi__zhuffman;
 
 stbi_inline static int stbi__bitreverse16(int n)
@@ -4120,7 +4175,7 @@ static int stbi__zhuffman_decode_slowpath(stbi__zbuf *a, stbi__zhuffman *z)
    if (s >= 16) return -1; // invalid code!
    // code size is s, so:
    b = (k >> (16-s)) - z->firstcode[s] + z->firstsymbol[s];
-   if (b >= sizeof (z->size)) return -1; // some data was corrupt somewhere!
+   if (b >= STBI__ZNSYMS) return -1; // some data was corrupt somewhere!
    if (z->size[b] != s) return -1;  // was originally an assert, but report failure instead.
    a->code_buffer >>= s;
    a->num_bits -= s;
@@ -4317,7 +4372,7 @@ static int stbi__parse_zlib_header(stbi__zbuf *a)
    return 1;
 }
 
-static const stbi_uc stbi__zdefault_length[288] =
+static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] =
 {
    8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
    8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
@@ -4363,7 +4418,7 @@ static int stbi__parse_zlib(stbi__zbuf *a, int parse_header)
       } else {
          if (type == 1) {
             // use fixed code lengths
-            if (!stbi__zbuild_huffman(&a->z_length  , stbi__zdefault_length  , 288)) return 0;
+            if (!stbi__zbuild_huffman(&a->z_length  , stbi__zdefault_length  , STBI__ZNSYMS)) return 0;
             if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance,  32)) return 0;
          } else {
             if (!stbi__compute_huffman_codes(a)) return 0;
@@ -4759,6 +4814,7 @@ static int stbi__create_png_image(stbi__png *a, stbi_uc *image_data, stbi__uint3
 
    // de-interlacing
    final = (stbi_uc *) stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0);
+   if (!final) return stbi__err("outofmem", "Out of memory");
    for (p=0; p < 7; ++p) {
       int xorig[] = { 0,4,0,2,0,1,0 };
       int yorig[] = { 0,0,4,0,2,0,1 };
@@ -4879,19 +4935,46 @@ static int stbi__expand_png_palette(stbi__png *a, stbi_uc *palette, int len, int
    return 1;
 }
 
-static int stbi__unpremultiply_on_load = 0;
-static int stbi__de_iphone_flag = 0;
+static int stbi__unpremultiply_on_load_global = 0;
+static int stbi__de_iphone_flag_global = 0;
 
 STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply)
 {
-   stbi__unpremultiply_on_load = flag_true_if_should_unpremultiply;
+   stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply;
 }
 
 STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert)
 {
-   stbi__de_iphone_flag = flag_true_if_should_convert;
+   stbi__de_iphone_flag_global = flag_true_if_should_convert;
+}
+
+#ifndef STBI_THREAD_LOCAL
+#define stbi__unpremultiply_on_load  stbi__unpremultiply_on_load_global
+#define stbi__de_iphone_flag  stbi__de_iphone_flag_global
+#else
+static STBI_THREAD_LOCAL int stbi__unpremultiply_on_load_local, stbi__unpremultiply_on_load_set;
+static STBI_THREAD_LOCAL int stbi__de_iphone_flag_local, stbi__de_iphone_flag_set;
+
+STBIDEF void stbi__unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply)
+{
+   stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply;
+   stbi__unpremultiply_on_load_set = 1;
+}
+
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert)
+{
+   stbi__de_iphone_flag_local = flag_true_if_should_convert;
+   stbi__de_iphone_flag_set = 1;
 }
 
+#define stbi__unpremultiply_on_load  (stbi__unpremultiply_on_load_set           \
+                                       ? stbi__unpremultiply_on_load_local      \
+                                       : stbi__unpremultiply_on_load_global)
+#define stbi__de_iphone_flag  (stbi__de_iphone_flag_set                         \
+                                ? stbi__de_iphone_flag_local                    \
+                                : stbi__de_iphone_flag_global)
+#endif // STBI_THREAD_LOCAL
+
 static void stbi__de_iphone(stbi__png *z)
 {
    stbi__context *s = z->s;
@@ -5272,6 +5355,32 @@ typedef struct
    int extra_read;
 } stbi__bmp_data;
 
+static int stbi__bmp_set_mask_defaults(stbi__bmp_data *info, int compress)
+{
+   // BI_BITFIELDS specifies masks explicitly, don't override
+   if (compress == 3)
+      return 1;
+
+   if (compress == 0) {
+      if (info->bpp == 16) {
+         info->mr = 31u << 10;
+         info->mg = 31u <<  5;
+         info->mb = 31u <<  0;
+      } else if (info->bpp == 32) {
+         info->mr = 0xffu << 16;
+         info->mg = 0xffu <<  8;
+         info->mb = 0xffu <<  0;
+         info->ma = 0xffu << 24;
+         info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0
+      } else {
+         // otherwise, use defaults, which is all-0
+         info->mr = info->mg = info->mb = info->ma = 0;
+      }
+      return 1;
+   }
+   return 0; // error
+}
+
 static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info)
 {
    int hsz;
@@ -5299,6 +5408,8 @@ static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info)
    if (hsz != 12) {
       int compress = stbi__get32le(s);
       if (compress == 1 || compress == 2) return stbi__errpuc("BMP RLE", "BMP type not supported: RLE");
+      if (compress >= 4) return stbi__errpuc("BMP JPEG/PNG", "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes
+      if (compress == 3 && info->bpp != 16 && info->bpp != 32) return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel
       stbi__get32le(s); // discard sizeof
       stbi__get32le(s); // discard hres
       stbi__get32le(s); // discard vres
@@ -5313,17 +5424,7 @@ static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info)
          }
          if (info->bpp == 16 || info->bpp == 32) {
             if (compress == 0) {
-               if (info->bpp == 32) {
-                  info->mr = 0xffu << 16;
-                  info->mg = 0xffu <<  8;
-                  info->mb = 0xffu <<  0;
-                  info->ma = 0xffu << 24;
-                  info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0
-               } else {
-                  info->mr = 31u << 10;
-                  info->mg = 31u <<  5;
-                  info->mb = 31u <<  0;
-               }
+               stbi__bmp_set_mask_defaults(info, compress);
             } else if (compress == 3) {
                info->mr = stbi__get32le(s);
                info->mg = stbi__get32le(s);
@@ -5338,6 +5439,7 @@ static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info)
                return stbi__errpuc("bad BMP", "bad BMP");
          }
       } else {
+         // V4/V5 header
          int i;
          if (hsz != 108 && hsz != 124)
             return stbi__errpuc("bad BMP", "bad BMP");
@@ -5345,6 +5447,8 @@ static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info)
          info->mg = stbi__get32le(s);
          info->mb = stbi__get32le(s);
          info->ma = stbi__get32le(s);
+         if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs
+            stbi__bmp_set_mask_defaults(info, compress);
          stbi__get32le(s); // discard color space
          for (i=0; i < 12; ++i)
             stbi__get32le(s); // discard color space parameters
@@ -5394,8 +5498,7 @@ static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req
          psize = (info.offset - info.extra_read - info.hsz) >> 2;
    }
    if (psize == 0) {
-      STBI_ASSERT(info.offset == s->callback_already_read + (int) (s->img_buffer - s->img_buffer_original));
-      if (info.offset != s->callback_already_read + (s->img_buffer - s->buffer_start)) {
+      if (info.offset != s->callback_already_read + (s->img_buffer - s->img_buffer_original)) {
         return stbi__errpuc("bad offset", "Corrupt BMP");
       }
    }
@@ -6342,6 +6445,7 @@ static void *stbi__pic_load(stbi__context *s,int *px,int *py,int *comp,int req_c
 
    // intermediate buffer is RGBA
    result = (stbi_uc *) stbi__malloc_mad3(x, y, 4, 0);
+   if (!result) return stbi__errpuc("outofmem", "Out of memory");
    memset(result, 0xff, x*y*4);
 
    if (!stbi__pic_load_core(s,x,y,comp, result)) {
@@ -6457,6 +6561,7 @@ static int stbi__gif_header(stbi__context *s, stbi__gif *g, int *comp, int is_in
 static int stbi__gif_info_raw(stbi__context *s, int *x, int *y, int *comp)
 {
    stbi__gif* g = (stbi__gif*) stbi__malloc(sizeof(stbi__gif));
+   if (!g) return stbi__err("outofmem", "Out of memory");
    if (!stbi__gif_header(s, g, comp, 1)) {
       STBI_FREE(g);
       stbi__rewind( s );
@@ -6766,6 +6871,17 @@ static stbi_uc *stbi__gif_load_next(stbi__context *s, stbi__gif *g, int *comp, i
    }
 }
 
+static void *stbi__load_gif_main_outofmem(stbi__gif *g, stbi_uc *out, int **delays)
+{
+   STBI_FREE(g->out);
+   STBI_FREE(g->history);
+   STBI_FREE(g->background);
+
+   if (out) STBI_FREE(out);
+   if (delays && *delays) STBI_FREE(*delays);
+   return stbi__errpuc("outofmem", "Out of memory");
+}
+
 static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp)
 {
    if (stbi__gif_test(s)) {
@@ -6777,6 +6893,10 @@ static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y,
       int stride;
       int out_size = 0;
       int delays_size = 0;
+
+      STBI_NOTUSED(out_size);
+      STBI_NOTUSED(delays_size);
+
       memset(&g, 0, sizeof(g));
       if (delays) {
          *delays = 0;
@@ -6794,26 +6914,29 @@ static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y,
 
             if (out) {
                void *tmp = (stbi_uc*) STBI_REALLOC_SIZED( out, out_size, layers * stride );
-               if (NULL == tmp) {
-                  STBI_FREE(g.out);
-                  STBI_FREE(g.history);
-                  STBI_FREE(g.background);
-                  return stbi__errpuc("outofmem", "Out of memory");
-               }
+               if (!tmp)
+                  return stbi__load_gif_main_outofmem(&g, out, delays);
                else {
                    out = (stbi_uc*) tmp;
                    out_size = layers * stride;
                }
 
                if (delays) {
-                  *delays = (int*) STBI_REALLOC_SIZED( *delays, delays_size, sizeof(int) * layers );
+                  int *new_delays = (int*) STBI_REALLOC_SIZED( *delays, delays_size, sizeof(int) * layers );
+                  if (!new_delays)
+                     return stbi__load_gif_main_outofmem(&g, out, delays);
+                  *delays = new_delays;
                   delays_size = layers * sizeof(int);
                }
             } else {
                out = (stbi_uc*)stbi__malloc( layers * stride );
+               if (!out)
+                  return stbi__load_gif_main_outofmem(&g, out, delays);
                out_size = layers * stride;
                if (delays) {
                   *delays = (int*) stbi__malloc( layers * sizeof(int) );
+                  if (!*delays)
+                     return stbi__load_gif_main_outofmem(&g, out, delays);
                   delays_size = layers * sizeof(int);
                }
             }
@@ -7138,9 +7261,10 @@ static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp)
 
    info.all_a = 255;
    p = stbi__bmp_parse_header(s, &info);
-   stbi__rewind( s );
-   if (p == NULL)
+   if (p == NULL) {
+      stbi__rewind( s );
       return 0;
+   }
    if (x) *x = s->img_x;
    if (y) *y = s->img_y;
    if (comp) {
@@ -7206,8 +7330,8 @@ static int stbi__psd_is16(stbi__context *s)
        stbi__rewind( s );
        return 0;
    }
-   (void) stbi__get32be(s);
-   (void) stbi__get32be(s);
+   STBI_NOTUSED(stbi__get32be(s));
+   STBI_NOTUSED(stbi__get32be(s));
    depth = stbi__get16be(s);
    if (depth != 16) {
        stbi__rewind( s );
@@ -7286,7 +7410,6 @@ static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp)
 // Known limitations:
 //    Does not support comments in the header section
 //    Does not support ASCII image data (formats P2 and P3)
-//    Does not support 16-bit-per-channel
 
 #ifndef STBI_NO_PNM
 
@@ -7307,7 +7430,8 @@ static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req
    stbi_uc *out;
    STBI_NOTUSED(ri);
 
-   if (!stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n))
+   ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n);
+   if (ri->bits_per_channel == 0)
       return 0;
 
    if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
@@ -7317,12 +7441,12 @@ static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req
    *y = s->img_y;
    if (comp) *comp = s->img_n;
 
-   if (!stbi__mad3sizes_valid(s->img_n, s->img_x, s->img_y, 0))
+   if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0))
       return stbi__errpuc("too large", "PNM too large");
 
-   out = (stbi_uc *) stbi__malloc_mad3(s->img_n, s->img_x, s->img_y, 0);
+   out = (stbi_uc *) stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0);
    if (!out) return stbi__errpuc("outofmem", "Out of memory");
-   stbi__getn(s, out, s->img_n * s->img_x * s->img_y);
+   stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8));
 
    if (req_comp && req_comp != s->img_n) {
       out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y);
@@ -7398,11 +7522,19 @@ static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp)
    stbi__pnm_skip_whitespace(s, &c);
 
    maxv = stbi__pnm_getinteger(s, &c);  // read max value
-
-   if (maxv > 255)
-      return stbi__err("max value > 255", "PPM image not 8-bit");
+   if (maxv > 65535)
+      return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images");
+   else if (maxv > 255)
+      return 16;
    else
-      return 1;
+      return 8;
+}
+
+static int stbi__pnm_is16(stbi__context *s)
+{
+   if (stbi__pnm_info(s, NULL, NULL, NULL) == 16)
+	   return 1;
+   return 0;
 }
 #endif
 
@@ -7458,6 +7590,9 @@ static int stbi__is_16_main(stbi__context *s)
    if (stbi__psd_is16(s))  return 1;
    #endif
 
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_is16(s))  return 1;
+   #endif
    return 0;
 }
 
@@ -7760,4 +7895,3 @@ ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 ------------------------------------------------------------------------------
 */
-
diff --git a/src/third_party_lib/wirehair/CMakeLists.txt b/src/third_party_lib/wirehair/CMakeLists.txt
index 49a4a95c..b3633236 100644
--- a/src/third_party_lib/wirehair/CMakeLists.txt
+++ b/src/third_party_lib/wirehair/CMakeLists.txt
@@ -61,9 +61,6 @@ endif()
 if(MSVC)
 else()
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${ARCH_NATIVE}")
-    if (CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64")
-        add_definitions(-DLINUX_ARM=1)
-    endif()
 endif()
 
 add_library(wirehair ${LIB_SOURCE_FILES})
diff --git a/src/third_party_lib/wirehair/WirehairTools.cpp b/src/third_party_lib/wirehair/WirehairTools.cpp
index dfd65856..df5bdd37 100644
--- a/src/third_party_lib/wirehair/WirehairTools.cpp
+++ b/src/third_party_lib/wirehair/WirehairTools.cpp
@@ -30,6 +30,7 @@
 #include "WirehairTools.h"
 
 #include <cmath>
+#include <cstdlib>
 
 #ifdef _MSC_VER
 #include <intrin.h> // _BitScanReverse
diff --git a/src/third_party_lib/wirehair/gf256.cpp b/src/third_party_lib/wirehair/gf256.cpp
index 510146b8..66aa9f01 100644
--- a/src/third_party_lib/wirehair/gf256.cpp
+++ b/src/third_party_lib/wirehair/gf256.cpp
@@ -200,7 +200,7 @@ static bool gf256_self_test()
 #endif
 
 #if defined(GF256_TRY_NEON)
-# if defined(IOS) && defined(__ARM_NEON__)
+# if defined(IOS) && (defined(__ARM_NEON) || defined(__ARM_NEON__))
 // Requires iPhone 5S or newer
 static const bool CpuHasNeon = true;
 static const bool CpuHasNeon64 = true;
diff --git a/src/third_party_lib/wirehair/gf256.h b/src/third_party_lib/wirehair/gf256.h
index c618c0d2..8396a44e 100644
--- a/src/third_party_lib/wirehair/gf256.h
+++ b/src/third_party_lib/wirehair/gf256.h
@@ -53,11 +53,17 @@
 //------------------------------------------------------------------------------
 // Platform/Architecture
 
+#if defined(__ARM_ARCH) || defined(__ARM_NEON) || defined(__ARM_NEON__)
+    #if !defined IOS
+        #define LINUX_ARM
+    #endif
+#endif
+
 #if defined(ANDROID) || defined(IOS) || defined(LINUX_ARM) || defined(__powerpc__) || defined(__s390__)
     #define GF256_TARGET_MOBILE
 #endif // ANDROID
 
-#if defined(__AVX2__) || (defined (_MSC_VER) && _MSC_VER >= 1900)
+#if defined(__AVX2__) && (!defined (_MSC_VER) || _MSC_VER >= 1900)
     #define GF256_TRY_AVX2 /* 256-bit */
     #include <immintrin.h>
     #define GF256_ALIGN_BYTES 32
@@ -66,36 +72,28 @@
 #endif // __AVX2__
 
 #if !defined(GF256_TARGET_MOBILE)
-    // Note: MSVC currently only supports SSSE3 but not AVX2
     #include <tmmintrin.h> // SSSE3: _mm_shuffle_epi8
     #include <emmintrin.h> // SSE2
 #endif // GF256_TARGET_MOBILE
 
-#if defined(HAVE_ARM_NEON_H)
+#if defined(__ARM_NEON) || defined(__ARM_NEON__)
     #include <arm_neon.h>
-#endif // HAVE_ARM_NEON_H
+    #define GF256_TRY_NEON
+#endif
 
+// Compiler-specific 128-bit SIMD register keyword
 #if defined(GF256_TARGET_MOBILE)
-
-    #define GF256_ALIGNED_ACCESSES /* Inputs must be aligned to GF256_ALIGN_BYTES */
-
-# if defined(HAVE_ARM_NEON_H)
-    // Compiler-specific 128-bit SIMD register keyword
+#if defined(GF256_TRY_NEON)
     #define GF256_M128 uint8x16_t
-    #define GF256_TRY_NEON
 #else
     #define GF256_M128 uint64_t
-# endif
-
+#endif // GF256_TRY_NEON
 #else // GF256_TARGET_MOBILE
-
-    // Compiler-specific 128-bit SIMD register keyword
     #define GF256_M128 __m128i
-
 #endif // GF256_TARGET_MOBILE
 
+// Compiler-specific 256-bit SIMD register keyword
 #ifdef GF256_TRY_AVX2
-    // Compiler-specific 256-bit SIMD register keyword
     #define GF256_M256 __m256i
 #endif
 
@@ -272,10 +270,6 @@ static GF256_FORCE_INLINE void gf256_div_mem(void * GF256_RESTRICT vz,
 //------------------------------------------------------------------------------
 // Misc Operations
 
-/// Swap two memory buffers in-place
-extern void gf256_memswap(void * GF256_RESTRICT vx, void * GF256_RESTRICT vy, int bytes);
-
-
 #ifdef __cplusplus
 }
 #endif // __cplusplus
diff --git a/src/third_party_lib/wirehair/wirehair.cpp b/src/third_party_lib/wirehair/wirehair.cpp
index 28c663cc..6c84acd6 100644
--- a/src/third_party_lib/wirehair/wirehair.cpp
+++ b/src/third_party_lib/wirehair/wirehair.cpp
@@ -153,7 +153,7 @@ WIREHAIR_EXPORT WirehairCodec wirehair_decoder_create(
 )
 {
     // If input is invalid:
-    if (messageBytes < 1 || blockBytes < 1) {
+    if (!m_init || messageBytes < 1 || blockBytes < 1) {
         return nullptr;
     }
 
diff --git a/test/catch.hpp b/test/catch.hpp
index 2a2d77a2..db1fed3b 100644
--- a/test/catch.hpp
+++ b/test/catch.hpp
@@ -1,9 +1,9 @@
 /*
- *  Catch v2.13.3
- *  Generated: 2020-10-31 18:20:31.045274
+ *  Catch v2.13.8
+ *  Generated: 2022-01-03 21:20:09.589503
  *  ----------------------------------------------------------
  *  This file has been merged from multiple headers. Please don't edit it directly
- *  Copyright (c) 2020 Two Blue Cubes Ltd. All rights reserved.
+ *  Copyright (c) 2022 Two Blue Cubes Ltd. All rights reserved.
  *
  *  Distributed under the Boost Software License, Version 1.0. (See accompanying
  *  file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
@@ -15,7 +15,7 @@
 
 #define CATCH_VERSION_MAJOR 2
 #define CATCH_VERSION_MINOR 13
-#define CATCH_VERSION_PATCH 3
+#define CATCH_VERSION_PATCH 8
 
 #ifdef __clang__
 #    pragma clang system_header
@@ -66,13 +66,16 @@
 #if !defined(CATCH_CONFIG_IMPL_ONLY)
 // start catch_platform.h
 
+// See e.g.:
+// https://opensource.apple.com/source/CarbonHeaders/CarbonHeaders-18.1/TargetConditionals.h.auto.html
 #ifdef __APPLE__
-# include <TargetConditionals.h>
-# if TARGET_OS_OSX == 1
-#  define CATCH_PLATFORM_MAC
-# elif TARGET_OS_IPHONE == 1
-#  define CATCH_PLATFORM_IPHONE
-# endif
+#  include <TargetConditionals.h>
+#  if (defined(TARGET_OS_OSX) && TARGET_OS_OSX == 1) || \
+      (defined(TARGET_OS_MAC) && TARGET_OS_MAC == 1)
+#    define CATCH_PLATFORM_MAC
+#  elif (defined(TARGET_OS_IPHONE) && TARGET_OS_IPHONE == 1)
+#    define CATCH_PLATFORM_IPHONE
+#  endif
 
 #elif defined(linux) || defined(__linux) || defined(__linux__)
 #  define CATCH_PLATFORM_LINUX
@@ -132,9 +135,9 @@ namespace Catch {
 
 #endif
 
-// We have to avoid both ICC and Clang, because they try to mask themselves
-// as gcc, and we want only GCC in this block
-#if defined(__GNUC__) && !defined(__clang__) && !defined(__ICC) && !defined(__CUDACC__)
+// Only GCC compiler should be used in this block, so other compilers trying to
+// mask themselves as GCC should be ignored.
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__ICC) && !defined(__CUDACC__) && !defined(__LCC__)
 #    define CATCH_INTERNAL_START_WARNINGS_SUPPRESSION _Pragma( "GCC diagnostic push" )
 #    define CATCH_INTERNAL_STOP_WARNINGS_SUPPRESSION  _Pragma( "GCC diagnostic pop" )
 
@@ -237,9 +240,6 @@ namespace Catch {
 // Visual C++
 #if defined(_MSC_VER)
 
-#  define CATCH_INTERNAL_START_WARNINGS_SUPPRESSION __pragma( warning(push) )
-#  define CATCH_INTERNAL_STOP_WARNINGS_SUPPRESSION  __pragma( warning(pop) )
-
 // Universal Windows platform does not support SEH
 // Or console colours (or console at all...)
 #  if defined(WINAPI_FAMILY) && (WINAPI_FAMILY == WINAPI_FAMILY_APP)
@@ -248,13 +248,18 @@ namespace Catch {
 #    define CATCH_INTERNAL_CONFIG_WINDOWS_SEH
 #  endif
 
+#  if !defined(__clang__) // Handle Clang masquerading for msvc
+
 // MSVC traditional preprocessor needs some workaround for __VA_ARGS__
 // _MSVC_TRADITIONAL == 0 means new conformant preprocessor
 // _MSVC_TRADITIONAL == 1 means old traditional non-conformant preprocessor
-#  if !defined(__clang__) // Handle Clang masquerading for msvc
 #    if !defined(_MSVC_TRADITIONAL) || (defined(_MSVC_TRADITIONAL) && _MSVC_TRADITIONAL)
 #      define CATCH_INTERNAL_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
 #    endif // MSVC_TRADITIONAL
+
+// Only do this if we're not using clang on Windows, which uses `diagnostic push` & `diagnostic pop`
+#    define CATCH_INTERNAL_START_WARNINGS_SUPPRESSION __pragma( warning(push) )
+#    define CATCH_INTERNAL_STOP_WARNINGS_SUPPRESSION  __pragma( warning(pop) )
 #  endif // __clang__
 
 #endif // _MSC_VER
@@ -323,7 +328,7 @@ namespace Catch {
   // Check if byte is available and usable
   #  if __has_include(<cstddef>) && defined(CATCH_CPP17_OR_GREATER)
   #    include <cstddef>
-  #    if __cpp_lib_byte > 0
+  #    if defined(__cpp_lib_byte) && (__cpp_lib_byte > 0)
   #      define CATCH_INTERNAL_CONFIG_CPP17_BYTE
   #    endif
   #  endif // __has_include(<cstddef>) && defined(CATCH_CPP17_OR_GREATER)
@@ -1007,34 +1012,34 @@ struct AutoReg : NonCopyable {
 
     #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
         #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION(Name, Tags, ...) \
-            INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, typename TestType, __VA_ARGS__ )
+            INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, typename TestType, __VA_ARGS__ )
     #else
         #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION(Name, Tags, ...) \
-            INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, typename TestType, __VA_ARGS__ ) )
+            INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, typename TestType, __VA_ARGS__ ) )
     #endif
 
     #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
         #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_SIG_NO_REGISTRATION(Name, Tags, Signature, ...) \
-            INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, Signature, __VA_ARGS__ )
+            INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, Signature, __VA_ARGS__ )
     #else
         #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_SIG_NO_REGISTRATION(Name, Tags, Signature, ...) \
-            INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, Signature, __VA_ARGS__ ) )
+            INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, Signature, __VA_ARGS__ ) )
     #endif
 
     #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
         #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION( ClassName, Name, Tags,... ) \
-            INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____C_L_A_S_S____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ) , ClassName, Name, Tags, typename T, __VA_ARGS__ )
+            INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_C_L_A_S_S_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ) , ClassName, Name, Tags, typename T, __VA_ARGS__ )
     #else
         #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION( ClassName, Name, Tags,... ) \
-            INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____C_L_A_S_S____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ) , ClassName, Name, Tags, typename T, __VA_ARGS__ ) )
+            INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_C_L_A_S_S_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ) , ClassName, Name, Tags, typename T, __VA_ARGS__ ) )
     #endif
 
     #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
         #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_SIG_NO_REGISTRATION( ClassName, Name, Tags, Signature, ... ) \
-            INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____C_L_A_S_S____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ) , ClassName, Name, Tags, Signature, __VA_ARGS__ )
+            INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_C_L_A_S_S_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ) , ClassName, Name, Tags, Signature, __VA_ARGS__ )
     #else
         #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_SIG_NO_REGISTRATION( ClassName, Name, Tags, Signature, ... ) \
-            INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____C_L_A_S_S____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ) , ClassName, Name, Tags, Signature, __VA_ARGS__ ) )
+            INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_NO_REGISTRATION_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_C_L_A_S_S_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ) , ClassName, Name, Tags, Signature, __VA_ARGS__ ) )
     #endif
 #endif
 
@@ -1047,7 +1052,7 @@ struct AutoReg : NonCopyable {
         CATCH_INTERNAL_STOP_WARNINGS_SUPPRESSION \
         static void TestName()
     #define INTERNAL_CATCH_TESTCASE( ... ) \
-        INTERNAL_CATCH_TESTCASE2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ), __VA_ARGS__ )
+        INTERNAL_CATCH_TESTCASE2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ), __VA_ARGS__ )
 
     ///////////////////////////////////////////////////////////////////////////////
     #define INTERNAL_CATCH_METHOD_AS_TEST_CASE( QualifiedMethod, ... ) \
@@ -1069,7 +1074,7 @@ struct AutoReg : NonCopyable {
         CATCH_INTERNAL_STOP_WARNINGS_SUPPRESSION \
         void TestName::test()
     #define INTERNAL_CATCH_TEST_CASE_METHOD( ClassName, ... ) \
-        INTERNAL_CATCH_TEST_CASE_METHOD2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ), ClassName, __VA_ARGS__ )
+        INTERNAL_CATCH_TEST_CASE_METHOD2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ), ClassName, __VA_ARGS__ )
 
     ///////////////////////////////////////////////////////////////////////////////
     #define INTERNAL_CATCH_REGISTER_TESTCASE( Function, ... ) \
@@ -1110,18 +1115,18 @@ struct AutoReg : NonCopyable {
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
     #define INTERNAL_CATCH_TEMPLATE_TEST_CASE(Name, Tags, ...) \
-        INTERNAL_CATCH_TEMPLATE_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, typename TestType, __VA_ARGS__ )
+        INTERNAL_CATCH_TEMPLATE_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, typename TestType, __VA_ARGS__ )
 #else
     #define INTERNAL_CATCH_TEMPLATE_TEST_CASE(Name, Tags, ...) \
-        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, typename TestType, __VA_ARGS__ ) )
+        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, typename TestType, __VA_ARGS__ ) )
 #endif
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
     #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_SIG(Name, Tags, Signature, ...) \
-        INTERNAL_CATCH_TEMPLATE_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, Signature, __VA_ARGS__ )
+        INTERNAL_CATCH_TEMPLATE_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, Signature, __VA_ARGS__ )
 #else
     #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_SIG(Name, Tags, Signature, ...) \
-        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, Signature, __VA_ARGS__ ) )
+        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, Signature, __VA_ARGS__ ) )
 #endif
 
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE2(TestName, TestFuncName, Name, Tags, Signature, TmplTypes, TypesList) \
@@ -1159,18 +1164,18 @@ struct AutoReg : NonCopyable {
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE(Name, Tags, ...)\
-        INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE2(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, typename T,__VA_ARGS__)
+        INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE2(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, typename T,__VA_ARGS__)
 #else
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE(Name, Tags, ...)\
-        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, typename T, __VA_ARGS__ ) )
+        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, typename T, __VA_ARGS__ ) )
 #endif
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_SIG(Name, Tags, Signature, ...)\
-        INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE2(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, Signature, __VA_ARGS__)
+        INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE2(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, Signature, __VA_ARGS__)
 #else
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_SIG(Name, Tags, Signature, ...)\
-        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, Signature, __VA_ARGS__ ) )
+        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, Signature, __VA_ARGS__ ) )
 #endif
 
     #define INTERNAL_CATCH_TEMPLATE_LIST_TEST_CASE_2(TestName, TestFunc, Name, Tags, TmplList)\
@@ -1201,7 +1206,7 @@ struct AutoReg : NonCopyable {
         static void TestFunc()
 
     #define INTERNAL_CATCH_TEMPLATE_LIST_TEST_CASE(Name, Tags, TmplList) \
-        INTERNAL_CATCH_TEMPLATE_LIST_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), Name, Tags, TmplList )
+        INTERNAL_CATCH_TEMPLATE_LIST_TEST_CASE_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), Name, Tags, TmplList )
 
     #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_2( TestNameClass, TestName, ClassName, Name, Tags, Signature, ... ) \
         CATCH_INTERNAL_START_WARNINGS_SUPPRESSION \
@@ -1234,18 +1239,18 @@ struct AutoReg : NonCopyable {
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
     #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD( ClassName, Name, Tags,... ) \
-        INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____C_L_A_S_S____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ) , ClassName, Name, Tags, typename T, __VA_ARGS__ )
+        INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_C_L_A_S_S_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ) , ClassName, Name, Tags, typename T, __VA_ARGS__ )
 #else
     #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD( ClassName, Name, Tags,... ) \
-        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____C_L_A_S_S____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ) , ClassName, Name, Tags, typename T, __VA_ARGS__ ) )
+        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_C_L_A_S_S_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ) , ClassName, Name, Tags, typename T, __VA_ARGS__ ) )
 #endif
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
     #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_SIG( ClassName, Name, Tags, Signature, ... ) \
-        INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____C_L_A_S_S____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ) , ClassName, Name, Tags, Signature, __VA_ARGS__ )
+        INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_C_L_A_S_S_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ) , ClassName, Name, Tags, Signature, __VA_ARGS__ )
 #else
     #define INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_SIG( ClassName, Name, Tags, Signature, ... ) \
-        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____C_L_A_S_S____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ) , ClassName, Name, Tags, Signature, __VA_ARGS__ ) )
+        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_C_L_A_S_S_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ) , ClassName, Name, Tags, Signature, __VA_ARGS__ ) )
 #endif
 
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_2(TestNameClass, TestName, ClassName, Name, Tags, Signature, TmplTypes, TypesList)\
@@ -1286,18 +1291,18 @@ struct AutoReg : NonCopyable {
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD( ClassName, Name, Tags, ... )\
-        INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), ClassName, Name, Tags, typename T, __VA_ARGS__ )
+        INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), ClassName, Name, Tags, typename T, __VA_ARGS__ )
 #else
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD( ClassName, Name, Tags, ... )\
-        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), ClassName, Name, Tags, typename T,__VA_ARGS__ ) )
+        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), ClassName, Name, Tags, typename T,__VA_ARGS__ ) )
 #endif
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_SIG( ClassName, Name, Tags, Signature, ... )\
-        INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), ClassName, Name, Tags, Signature, __VA_ARGS__ )
+        INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), ClassName, Name, Tags, Signature, __VA_ARGS__ )
 #else
     #define INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_SIG( ClassName, Name, Tags, Signature, ... )\
-        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), ClassName, Name, Tags, Signature,__VA_ARGS__ ) )
+        INTERNAL_CATCH_EXPAND_VARGS( INTERNAL_CATCH_TEMPLATE_PRODUCT_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), ClassName, Name, Tags, Signature,__VA_ARGS__ ) )
 #endif
 
     #define INTERNAL_CATCH_TEMPLATE_LIST_TEST_CASE_METHOD_2( TestNameClass, TestName, ClassName, Name, Tags, TmplList) \
@@ -1331,7 +1336,7 @@ struct AutoReg : NonCopyable {
         void TestName<TestType>::test()
 
 #define INTERNAL_CATCH_TEMPLATE_LIST_TEST_CASE_METHOD(ClassName, Name, Tags, TmplList) \
-        INTERNAL_CATCH_TEMPLATE_LIST_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____ ), INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_M_P_L_A_T_E____T_E_S_T____F_U_N_C____ ), ClassName, Name, Tags, TmplList )
+        INTERNAL_CATCH_TEMPLATE_LIST_TEST_CASE_METHOD_2( INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_ ), INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_M_P_L_A_T_E_T_E_S_T_F_U_N_C_ ), ClassName, Name, Tags, TmplList )
 
 // end catch_test_registry.h
 // start catch_capture.hpp
@@ -3088,7 +3093,7 @@ namespace Detail {
         Approx operator-() const;
 
         template <typename T, typename = typename std::enable_if<std::is_constructible<double, T>::value>::type>
-        Approx operator()( T const& value ) {
+        Approx operator()( T const& value ) const {
             Approx approx( static_cast<double>(value) );
             approx.m_epsilon = m_epsilon;
             approx.m_margin = m_margin;
@@ -4160,7 +4165,7 @@ namespace Generators {
             if (!m_predicate(m_generator.get())) {
                 // It might happen that there are no values that pass the
                 // filter. In that case we throw an exception.
-                auto has_initial_value = next();
+                auto has_initial_value = nextImpl();
                 if (!has_initial_value) {
                     Catch::throw_exception(GeneratorException("No valid value found in filtered generator"));
                 }
@@ -4172,6 +4177,11 @@ namespace Generators {
         }
 
         bool next() override {
+            return nextImpl();
+        }
+
+    private:
+        bool nextImpl() {
             bool success = m_generator.next();
             if (!success) {
                 return false;
@@ -5455,6 +5465,8 @@ namespace Catch {
 } // namespace Catch
 
 // end catch_outlier_classification.hpp
+
+#include <iterator>
 #endif // CATCH_CONFIG_ENABLE_BENCHMARKING
 
 #include <string>
@@ -6339,9 +6351,10 @@ namespace Catch {
 
         void writeTestCase(TestCaseNode const& testCaseNode);
 
-        void writeSection(std::string const& className,
-                          std::string const& rootName,
-                          SectionNode const& sectionNode);
+        void writeSection( std::string const& className,
+                           std::string const& rootName,
+                           SectionNode const& sectionNode,
+                           bool testOkToFail );
 
         void writeAssertions(SectionNode const& sectionNode);
         void writeAssertion(AssertionStats const& stats);
@@ -6876,7 +6889,7 @@ namespace Catch {
                     }
                     iters *= 2;
                 }
-                throw optimized_away_error{};
+                Catch::throw_exception(optimized_away_error{});
             }
         } // namespace Detail
     } // namespace Benchmark
@@ -6884,6 +6897,7 @@ namespace Catch {
 
 // end catch_run_for_at_least.hpp
 #include <algorithm>
+#include <iterator>
 
 namespace Catch {
     namespace Benchmark {
@@ -7054,8 +7068,8 @@ namespace Catch {
                 double b2 = bias - z1;
                 double a1 = a(b1);
                 double a2 = a(b2);
-                auto lo = std::max(cumn(a1), 0);
-                auto hi = std::min(cumn(a2), n - 1);
+                auto lo = (std::max)(cumn(a1), 0);
+                auto hi = (std::min)(cumn(a2), n - 1);
 
                 return { point, resample[lo], resample[hi], confidence_level };
             }
@@ -7124,7 +7138,9 @@ namespace Catch {
             }
             template <typename Clock>
             EnvironmentEstimate<FloatDuration<Clock>> estimate_clock_cost(FloatDuration<Clock> resolution) {
-                auto time_limit = std::min(resolution * clock_cost_estimation_tick_limit, FloatDuration<Clock>(clock_cost_estimation_time_limit));
+                auto time_limit = (std::min)(
+                    resolution * clock_cost_estimation_tick_limit,
+                    FloatDuration<Clock>(clock_cost_estimation_time_limit));
                 auto time_clock = [](int k) {
                     return Detail::measure<Clock>([k] {
                         for (int i = 0; i < k; ++i) {
@@ -7771,7 +7787,7 @@ namespace Catch {
                 double sb = stddev.point;
                 double mn = mean.point / n;
                 double mg_min = mn / 2.;
-                double sg = std::min(mg_min / 4., sb / std::sqrt(n));
+                double sg = (std::min)(mg_min / 4., sb / std::sqrt(n));
                 double sg2 = sg * sg;
                 double sb2 = sb * sb;
 
@@ -7790,7 +7806,7 @@ namespace Catch {
                     return (nc / n) * (sb2 - nc * sg2);
                 };
 
-                return std::min(var_out(1), var_out(std::min(c_max(0.), c_max(mg_min)))) / sb2;
+                return (std::min)(var_out(1), var_out((std::min)(c_max(0.), c_max(mg_min)))) / sb2;
             }
 
             bootstrap_analysis analyse_samples(double confidence_level, int n_resamples, std::vector<double>::iterator first, std::vector<double>::iterator last) {
@@ -7980,86 +7996,58 @@ namespace Catch {
 
 // start catch_fatal_condition.h
 
-// start catch_windows_h_proxy.h
-
-
-#if defined(CATCH_PLATFORM_WINDOWS)
-
-#if !defined(NOMINMAX) && !defined(CATCH_CONFIG_NO_NOMINMAX)
-#  define CATCH_DEFINED_NOMINMAX
-#  define NOMINMAX
-#endif
-#if !defined(WIN32_LEAN_AND_MEAN) && !defined(CATCH_CONFIG_NO_WIN32_LEAN_AND_MEAN)
-#  define CATCH_DEFINED_WIN32_LEAN_AND_MEAN
-#  define WIN32_LEAN_AND_MEAN
-#endif
-
-#ifdef __AFXDLL
-#include <AfxWin.h>
-#else
-#include <windows.h>
-#endif
-
-#ifdef CATCH_DEFINED_NOMINMAX
-#  undef NOMINMAX
-#endif
-#ifdef CATCH_DEFINED_WIN32_LEAN_AND_MEAN
-#  undef WIN32_LEAN_AND_MEAN
-#endif
-
-#endif // defined(CATCH_PLATFORM_WINDOWS)
-
-// end catch_windows_h_proxy.h
-#if defined( CATCH_CONFIG_WINDOWS_SEH )
+#include <cassert>
 
 namespace Catch {
 
-    struct FatalConditionHandler {
-
-        static LONG CALLBACK handleVectoredException(PEXCEPTION_POINTERS ExceptionInfo);
+    // Wrapper for platform-specific fatal error (signals/SEH) handlers
+    //
+    // Tries to be cooperative with other handlers, and not step over
+    // other handlers. This means that unknown structured exceptions
+    // are passed on, previous signal handlers are called, and so on.
+    //
+    // Can only be instantiated once, and assumes that once a signal
+    // is caught, the binary will end up terminating. Thus, there
+    class FatalConditionHandler {
+        bool m_started = false;
+
+        // Install/disengage implementation for specific platform.
+        // Should be if-defed to work on current platform, can assume
+        // engage-disengage 1:1 pairing.
+        void engage_platform();
+        void disengage_platform();
+    public:
+        // Should also have platform-specific implementations as needed
         FatalConditionHandler();
-        static void reset();
         ~FatalConditionHandler();
 
-    private:
-        static bool isSet;
-        static ULONG guaranteeSize;
-        static PVOID exceptionHandlerHandle;
-    };
-
-} // namespace Catch
-
-#elif defined ( CATCH_CONFIG_POSIX_SIGNALS )
-
-#include <signal.h>
-
-namespace Catch {
-
-    struct FatalConditionHandler {
-
-        static bool isSet;
-        static struct sigaction oldSigActions[];
-        static stack_t oldSigStack;
-        static char altStackMem[];
-
-        static void handleSignal( int sig );
+        void engage() {
+            assert(!m_started && "Handler cannot be installed twice.");
+            m_started = true;
+            engage_platform();
+        }
 
-        FatalConditionHandler();
-        ~FatalConditionHandler();
-        static void reset();
+        void disengage() {
+            assert(m_started && "Handler cannot be uninstalled without being installed first");
+            m_started = false;
+            disengage_platform();
+        }
     };
 
-} // namespace Catch
-
-#else
-
-namespace Catch {
-    struct FatalConditionHandler {
-        void reset();
+    //! Simple RAII guard for (dis)engaging the FatalConditionHandler
+    class FatalConditionHandlerGuard {
+        FatalConditionHandler* m_handler;
+    public:
+        FatalConditionHandlerGuard(FatalConditionHandler* handler):
+            m_handler(handler) {
+            m_handler->engage();
+        }
+        ~FatalConditionHandlerGuard() {
+            m_handler->disengage();
+        }
     };
-}
 
-#endif
+} // end namespace Catch
 
 // end catch_fatal_condition.h
 #include <string>
@@ -8185,6 +8173,7 @@ namespace Catch {
         std::vector<SectionEndInfo> m_unfinishedSections;
         std::vector<ITracker*> m_activeSections;
         TrackerContext m_trackerContext;
+        FatalConditionHandler m_fatalConditionhandler;
         bool m_lastAssertionPassed = false;
         bool m_shouldReportUnexpected = true;
         bool m_includeSuccessfulResults;
@@ -10057,6 +10046,36 @@ namespace Catch {
 }
 
 // end catch_errno_guard.h
+// start catch_windows_h_proxy.h
+
+
+#if defined(CATCH_PLATFORM_WINDOWS)
+
+#if !defined(NOMINMAX) && !defined(CATCH_CONFIG_NO_NOMINMAX)
+#  define CATCH_DEFINED_NOMINMAX
+#  define NOMINMAX
+#endif
+#if !defined(WIN32_LEAN_AND_MEAN) && !defined(CATCH_CONFIG_NO_WIN32_LEAN_AND_MEAN)
+#  define CATCH_DEFINED_WIN32_LEAN_AND_MEAN
+#  define WIN32_LEAN_AND_MEAN
+#endif
+
+#ifdef __AFXDLL
+#include <AfxWin.h>
+#else
+#include <windows.h>
+#endif
+
+#ifdef CATCH_DEFINED_NOMINMAX
+#  undef NOMINMAX
+#endif
+#ifdef CATCH_DEFINED_WIN32_LEAN_AND_MEAN
+#  undef WIN32_LEAN_AND_MEAN
+#endif
+
+#endif // defined(CATCH_PLATFORM_WINDOWS)
+
+// end catch_windows_h_proxy.h
 #include <sstream>
 
 namespace Catch {
@@ -10573,7 +10592,7 @@ namespace Catch {
             // Extracts the actual name part of an enum instance
             // In other words, it returns the Blue part of Bikeshed::Colour::Blue
             StringRef extractInstanceName(StringRef enumInstance) {
-                // Find last occurence of ":"
+                // Find last occurrence of ":"
                 size_t name_start = enumInstance.size();
                 while (name_start > 0 && enumInstance[name_start - 1] != ':') {
                     --name_start;
@@ -10735,25 +10754,47 @@ namespace Catch {
 // end catch_exception_translator_registry.cpp
 // start catch_fatal_condition.cpp
 
-#if defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wmissing-field-initializers"
-#endif
+#include <algorithm>
+
+#if !defined( CATCH_CONFIG_WINDOWS_SEH ) && !defined( CATCH_CONFIG_POSIX_SIGNALS )
+
+namespace Catch {
+
+    // If neither SEH nor signal handling is required, the handler impls
+    // do not have to do anything, and can be empty.
+    void FatalConditionHandler::engage_platform() {}
+    void FatalConditionHandler::disengage_platform() {}
+    FatalConditionHandler::FatalConditionHandler() = default;
+    FatalConditionHandler::~FatalConditionHandler() = default;
+
+} // end namespace Catch
+
+#endif // !CATCH_CONFIG_WINDOWS_SEH && !CATCH_CONFIG_POSIX_SIGNALS
+
+#if defined( CATCH_CONFIG_WINDOWS_SEH ) && defined( CATCH_CONFIG_POSIX_SIGNALS )
+#error "Inconsistent configuration: Windows' SEH handling and POSIX signals cannot be enabled at the same time"
+#endif // CATCH_CONFIG_WINDOWS_SEH && CATCH_CONFIG_POSIX_SIGNALS
 
 #if defined( CATCH_CONFIG_WINDOWS_SEH ) || defined( CATCH_CONFIG_POSIX_SIGNALS )
 
 namespace {
-    // Report the error condition
+    //! Signals fatal error message to the run context
     void reportFatal( char const * const message ) {
         Catch::getCurrentContext().getResultCapture()->handleFatalErrorCondition( message );
     }
-}
 
-#endif // signals/SEH handling
+    //! Minimal size Catch2 needs for its own fatal error handling.
+    //! Picked anecdotally, so it might not be sufficient on all
+    //! platforms, and for all configurations.
+    constexpr std::size_t minStackSizeForErrors = 32 * 1024;
+} // end unnamed namespace
+
+#endif // CATCH_CONFIG_WINDOWS_SEH || CATCH_CONFIG_POSIX_SIGNALS
 
 #if defined( CATCH_CONFIG_WINDOWS_SEH )
 
 namespace Catch {
+
     struct SignalDefs { DWORD id; const char* name; };
 
     // There is no 1-1 mapping between signals and windows exceptions.
@@ -10766,7 +10807,7 @@ namespace Catch {
         { static_cast<DWORD>(EXCEPTION_INT_DIVIDE_BY_ZERO), "Divide by zero error" },
     };
 
-    LONG CALLBACK FatalConditionHandler::handleVectoredException(PEXCEPTION_POINTERS ExceptionInfo) {
+    static LONG CALLBACK handleVectoredException(PEXCEPTION_POINTERS ExceptionInfo) {
         for (auto const& def : signalDefs) {
             if (ExceptionInfo->ExceptionRecord->ExceptionCode == def.id) {
                 reportFatal(def.name);
@@ -10777,38 +10818,50 @@ namespace Catch {
         return EXCEPTION_CONTINUE_SEARCH;
     }
 
+    // Since we do not support multiple instantiations, we put these
+    // into global variables and rely on cleaning them up in outlined
+    // constructors/destructors
+    static PVOID exceptionHandlerHandle = nullptr;
+
+    // For MSVC, we reserve part of the stack memory for handling
+    // memory overflow structured exception.
     FatalConditionHandler::FatalConditionHandler() {
-        isSet = true;
-        // 32k seems enough for Catch to handle stack overflow,
-        // but the value was found experimentally, so there is no strong guarantee
-        guaranteeSize = 32 * 1024;
-        exceptionHandlerHandle = nullptr;
+        ULONG guaranteeSize = static_cast<ULONG>(minStackSizeForErrors);
+        if (!SetThreadStackGuarantee(&guaranteeSize)) {
+            // We do not want to fully error out, because needing
+            // the stack reserve should be rare enough anyway.
+            Catch::cerr()
+                << "Failed to reserve piece of stack."
+                << " Stack overflows will not be reported successfully.";
+        }
+    }
+
+    // We do not attempt to unset the stack guarantee, because
+    // Windows does not support lowering the stack size guarantee.
+    FatalConditionHandler::~FatalConditionHandler() = default;
+
+    void FatalConditionHandler::engage_platform() {
         // Register as first handler in current chain
         exceptionHandlerHandle = AddVectoredExceptionHandler(1, handleVectoredException);
-        // Pass in guarantee size to be filled
-        SetThreadStackGuarantee(&guaranteeSize);
+        if (!exceptionHandlerHandle) {
+            CATCH_RUNTIME_ERROR("Could not register vectored exception handler");
+        }
     }
 
-    void FatalConditionHandler::reset() {
-        if (isSet) {
-            RemoveVectoredExceptionHandler(exceptionHandlerHandle);
-            SetThreadStackGuarantee(&guaranteeSize);
-            exceptionHandlerHandle = nullptr;
-            isSet = false;
+    void FatalConditionHandler::disengage_platform() {
+        if (!RemoveVectoredExceptionHandler(exceptionHandlerHandle)) {
+            CATCH_RUNTIME_ERROR("Could not unregister vectored exception handler");
         }
+        exceptionHandlerHandle = nullptr;
     }
 
-    FatalConditionHandler::~FatalConditionHandler() {
-        reset();
-    }
+} // end namespace Catch
 
-bool FatalConditionHandler::isSet = false;
-ULONG FatalConditionHandler::guaranteeSize = 0;
-PVOID FatalConditionHandler::exceptionHandlerHandle = nullptr;
+#endif // CATCH_CONFIG_WINDOWS_SEH
 
-} // namespace Catch
+#if defined( CATCH_CONFIG_POSIX_SIGNALS )
 
-#elif defined( CATCH_CONFIG_POSIX_SIGNALS )
+#include <signal.h>
 
 namespace Catch {
 
@@ -10817,10 +10870,6 @@ namespace Catch {
         const char* name;
     };
 
-    // 32kb for the alternate stack seems to be sufficient. However, this value
-    // is experimentally determined, so that's not guaranteed.
-    static constexpr std::size_t sigStackSize = 32768 >= MINSIGSTKSZ ? 32768 : MINSIGSTKSZ;
-
     static SignalDefs signalDefs[] = {
         { SIGINT,  "SIGINT - Terminal interrupt signal" },
         { SIGILL,  "SIGILL - Illegal instruction signal" },
@@ -10830,7 +10879,32 @@ namespace Catch {
         { SIGABRT, "SIGABRT - Abort (abnormal termination) signal" }
     };
 
-    void FatalConditionHandler::handleSignal( int sig ) {
+// Older GCCs trigger -Wmissing-field-initializers for T foo = {}
+// which is zero initialization, but not explicit. We want to avoid
+// that.
+#if defined(__GNUC__)
+#    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wmissing-field-initializers"
+#endif
+
+    static char* altStackMem = nullptr;
+    static std::size_t altStackSize = 0;
+    static stack_t oldSigStack{};
+    static struct sigaction oldSigActions[sizeof(signalDefs) / sizeof(SignalDefs)]{};
+
+    static void restorePreviousSignalHandlers() {
+        // We set signal handlers back to the previous ones. Hopefully
+        // nobody overwrote them in the meantime, and doesn't expect
+        // their signal handlers to live past ours given that they
+        // installed them after ours..
+        for (std::size_t i = 0; i < sizeof(signalDefs) / sizeof(SignalDefs); ++i) {
+            sigaction(signalDefs[i].id, &oldSigActions[i], nullptr);
+        }
+        // Return the old stack
+        sigaltstack(&oldSigStack, nullptr);
+    }
+
+    static void handleSignal( int sig ) {
         char const * name = "<unknown signal>";
         for (auto const& def : signalDefs) {
             if (sig == def.id) {
@@ -10838,16 +10912,33 @@ namespace Catch {
                 break;
             }
         }
-        reset();
-        reportFatal(name);
+        // We need to restore previous signal handlers and let them do
+        // their thing, so that the users can have the debugger break
+        // when a signal is raised, and so on.
+        restorePreviousSignalHandlers();
+        reportFatal( name );
         raise( sig );
     }
 
     FatalConditionHandler::FatalConditionHandler() {
-        isSet = true;
+        assert(!altStackMem && "Cannot initialize POSIX signal handler when one already exists");
+        if (altStackSize == 0) {
+            altStackSize = std::max(static_cast<size_t>(SIGSTKSZ), minStackSizeForErrors);
+        }
+        altStackMem = new char[altStackSize]();
+    }
+
+    FatalConditionHandler::~FatalConditionHandler() {
+        delete[] altStackMem;
+        // We signal that another instance can be constructed by zeroing
+        // out the pointer.
+        altStackMem = nullptr;
+    }
+
+    void FatalConditionHandler::engage_platform() {
         stack_t sigStack;
         sigStack.ss_sp = altStackMem;
-        sigStack.ss_size = sigStackSize;
+        sigStack.ss_size = altStackSize;
         sigStack.ss_flags = 0;
         sigaltstack(&sigStack, &oldSigStack);
         struct sigaction sa = { };
@@ -10859,40 +10950,17 @@ namespace Catch {
         }
     }
 
-    FatalConditionHandler::~FatalConditionHandler() {
-        reset();
-    }
+#if defined(__GNUC__)
+#    pragma GCC diagnostic pop
+#endif
 
-    void FatalConditionHandler::reset() {
-        if( isSet ) {
-            // Set signals back to previous values -- hopefully nobody overwrote them in the meantime
-            for( std::size_t i = 0; i < sizeof(signalDefs)/sizeof(SignalDefs); ++i ) {
-                sigaction(signalDefs[i].id, &oldSigActions[i], nullptr);
-            }
-            // Return the old stack
-            sigaltstack(&oldSigStack, nullptr);
-            isSet = false;
-        }
+    void FatalConditionHandler::disengage_platform() {
+        restorePreviousSignalHandlers();
     }
 
-    bool FatalConditionHandler::isSet = false;
-    struct sigaction FatalConditionHandler::oldSigActions[sizeof(signalDefs)/sizeof(SignalDefs)] = {};
-    stack_t FatalConditionHandler::oldSigStack = {};
-    char FatalConditionHandler::altStackMem[sigStackSize] = {};
-
-} // namespace Catch
-
-#else
-
-namespace Catch {
-    void FatalConditionHandler::reset() {}
-}
-
-#endif // signals/SEH handling
+} // end namespace Catch
 
-#if defined(__GNUC__)
-#    pragma GCC diagnostic pop
-#endif
+#endif // CATCH_CONFIG_POSIX_SIGNALS
 // end catch_fatal_condition.cpp
 // start catch_generators.cpp
 
@@ -11447,7 +11515,8 @@ namespace {
             return lhs == rhs;
         }
 
-        auto ulpDiff = std::abs(lc - rc);
+        // static cast as a workaround for IBM XLC
+        auto ulpDiff = std::abs(static_cast<FP>(lc - rc));
         return static_cast<uint64_t>(ulpDiff) <= maxUlpDiff;
     }
 
@@ -11621,7 +11690,6 @@ Floating::WithinRelMatcher WithinRel(float target) {
 
 } // namespace Matchers
 } // namespace Catch
-
 // end catch_matchers_floating.cpp
 // start catch_matchers_generic.cpp
 
@@ -12955,9 +13023,8 @@ namespace Catch {
     }
 
     void RunContext::invokeActiveTestCase() {
-        FatalConditionHandler fatalConditionHandler; // Handle signals
+        FatalConditionHandlerGuard _(&m_fatalConditionhandler);
         m_activeTestCase->invoke();
-        fatalConditionHandler.reset();
     }
 
     void RunContext::handleUnfinishedSections() {
@@ -14126,24 +14193,28 @@ namespace Catch {
 
     namespace {
         struct TestHasher {
-            explicit TestHasher(Catch::SimplePcg32& rng_instance) {
-                basis = rng_instance();
-                basis <<= 32;
-                basis |= rng_instance();
-            }
+            using hash_t = uint64_t;
 
-            uint64_t basis;
+            explicit TestHasher( hash_t hashSuffix ):
+                m_hashSuffix{ hashSuffix } {}
 
-            uint64_t operator()(TestCase const& t) const {
-                // Modified FNV-1a hash
-                static constexpr uint64_t prime = 1099511628211;
-                uint64_t hash = basis;
-                for (const char c : t.name) {
+            uint32_t operator()( TestCase const& t ) const {
+                // FNV-1a hash with multiplication fold.
+                const hash_t prime = 1099511628211u;
+                hash_t hash = 14695981039346656037u;
+                for ( const char c : t.name ) {
                     hash ^= c;
                     hash *= prime;
                 }
-                return hash;
+                hash ^= m_hashSuffix;
+                hash *= prime;
+                const uint32_t low{ static_cast<uint32_t>( hash ) };
+                const uint32_t high{ static_cast<uint32_t>( hash >> 32 ) };
+                return low * high;
             }
+
+        private:
+            hash_t m_hashSuffix;
         };
     } // end unnamed namespace
 
@@ -14161,9 +14232,9 @@ namespace Catch {
 
             case RunTests::InRandomOrder: {
                 seedRng( config );
-                TestHasher h( rng() );
+                TestHasher h{ config.rngSeed() };
 
-                using hashedTest = std::pair<uint64_t, TestCase const*>;
+                using hashedTest = std::pair<TestHasher::hash_t, TestCase const*>;
                 std::vector<hashedTest> indexed_tests;
                 indexed_tests.reserve( unsortedTestCases.size() );
 
@@ -15316,7 +15387,7 @@ namespace Catch {
     }
 
     Version const& libraryVersion() {
-        static Version version( 2, 13, 3, "", 0 );
+        static Version version( 2, 13, 8, "", 0 );
         return version;
     }
 
@@ -16729,6 +16800,7 @@ CATCH_REGISTER_REPORTER("console", ConsoleReporter)
 #include <sstream>
 #include <ctime>
 #include <algorithm>
+#include <iomanip>
 
 namespace Catch {
 
@@ -16756,7 +16828,7 @@ namespace Catch {
 #else
             std::strftime(timeStamp, timeStampSize, fmt, timeInfo);
 #endif
-            return std::string(timeStamp);
+            return std::string(timeStamp, timeStampSize-1);
         }
 
         std::string fileNameTag(const std::vector<std::string> &tags) {
@@ -16767,6 +16839,17 @@ namespace Catch {
                 return it->substr(1);
             return std::string();
         }
+
+        // Formats the duration in seconds to 3 decimal places.
+        // This is done because some genius defined Maven Surefire schema
+        // in a way that only accepts 3 decimal places, and tools like
+        // Jenkins use that schema for validation JUnit reporter output.
+        std::string formatDuration( double seconds ) {
+            ReusableStringStream rss;
+            rss << std::fixed << std::setprecision( 3 ) << seconds;
+            return rss.str();
+        }
+
     } // anonymous namespace
 
     JunitReporter::JunitReporter( ReporterConfig const& _config )
@@ -16836,7 +16919,7 @@ namespace Catch {
         if( m_config->showDurations() == ShowDurations::Never )
             xml.writeAttribute( "time", "" );
         else
-            xml.writeAttribute( "time", suiteTime );
+            xml.writeAttribute( "time", formatDuration( suiteTime ) );
         xml.writeAttribute( "timestamp", getCurrentTimestamp() );
 
         // Write properties if there are any
@@ -16881,12 +16964,13 @@ namespace Catch {
         if ( !m_config->name().empty() )
             className = m_config->name() + "." + className;
 
-        writeSection( className, "", rootSection );
+        writeSection( className, "", rootSection, stats.testInfo.okToFail() );
     }
 
-    void JunitReporter::writeSection(  std::string const& className,
-                        std::string const& rootName,
-                        SectionNode const& sectionNode ) {
+    void JunitReporter::writeSection( std::string const& className,
+                                      std::string const& rootName,
+                                      SectionNode const& sectionNode,
+                                      bool testOkToFail) {
         std::string name = trim( sectionNode.stats.sectionInfo.name );
         if( !rootName.empty() )
             name = rootName + '/' + name;
@@ -16903,13 +16987,18 @@ namespace Catch {
                 xml.writeAttribute( "classname", className );
                 xml.writeAttribute( "name", name );
             }
-            xml.writeAttribute( "time", ::Catch::Detail::stringify( sectionNode.stats.durationInSeconds ) );
+            xml.writeAttribute( "time", formatDuration( sectionNode.stats.durationInSeconds ) );
             // This is not ideal, but it should be enough to mimic gtest's
             // junit output.
             // Ideally the JUnit reporter would also handle `skipTest`
             // events and write those out appropriately.
             xml.writeAttribute( "status", "run" );
 
+            if (sectionNode.stats.assertions.failedButOk) {
+                xml.scopedElement("skipped")
+                    .writeAttribute("message", "TEST_CASE tagged with !mayfail");
+            }
+
             writeAssertions( sectionNode );
 
             if( !sectionNode.stdOut.empty() )
@@ -16919,9 +17008,9 @@ namespace Catch {
         }
         for( auto const& childNode : sectionNode.childSections )
             if( className.empty() )
-                writeSection( name, "", *childNode );
+                writeSection( name, "", *childNode, testOkToFail );
             else
-                writeSection( className, name, *childNode );
+                writeSection( className, name, *childNode, testOkToFail );
     }
 
     void JunitReporter::writeAssertions( SectionNode const& sectionNode ) {
@@ -17566,9 +17655,9 @@ int main (int argc, char * const argv[]) {
 
 #if defined(CATCH_CONFIG_ENABLE_BENCHMARKING)
 #define CATCH_BENCHMARK(...) \
-    INTERNAL_CATCH_BENCHMARK(INTERNAL_CATCH_UNIQUE_NAME(____C_A_T_C_H____B_E_N_C_H____), INTERNAL_CATCH_GET_1_ARG(__VA_ARGS__,,), INTERNAL_CATCH_GET_2_ARG(__VA_ARGS__,,))
+    INTERNAL_CATCH_BENCHMARK(INTERNAL_CATCH_UNIQUE_NAME(C_A_T_C_H_B_E_N_C_H_), INTERNAL_CATCH_GET_1_ARG(__VA_ARGS__,,), INTERNAL_CATCH_GET_2_ARG(__VA_ARGS__,,))
 #define CATCH_BENCHMARK_ADVANCED(name) \
-    INTERNAL_CATCH_BENCHMARK_ADVANCED(INTERNAL_CATCH_UNIQUE_NAME(____C_A_T_C_H____B_E_N_C_H____), name)
+    INTERNAL_CATCH_BENCHMARK_ADVANCED(INTERNAL_CATCH_UNIQUE_NAME(C_A_T_C_H_B_E_N_C_H_), name)
 #endif // CATCH_CONFIG_ENABLE_BENCHMARKING
 
 // If CATCH_CONFIG_PREFIX_ALL is not defined then the CATCH_ prefix is not required
@@ -17670,9 +17759,9 @@ int main (int argc, char * const argv[]) {
 
 #if defined(CATCH_CONFIG_ENABLE_BENCHMARKING)
 #define BENCHMARK(...) \
-    INTERNAL_CATCH_BENCHMARK(INTERNAL_CATCH_UNIQUE_NAME(____C_A_T_C_H____B_E_N_C_H____), INTERNAL_CATCH_GET_1_ARG(__VA_ARGS__,,), INTERNAL_CATCH_GET_2_ARG(__VA_ARGS__,,))
+    INTERNAL_CATCH_BENCHMARK(INTERNAL_CATCH_UNIQUE_NAME(C_A_T_C_H_B_E_N_C_H_), INTERNAL_CATCH_GET_1_ARG(__VA_ARGS__,,), INTERNAL_CATCH_GET_2_ARG(__VA_ARGS__,,))
 #define BENCHMARK_ADVANCED(name) \
-    INTERNAL_CATCH_BENCHMARK_ADVANCED(INTERNAL_CATCH_UNIQUE_NAME(____C_A_T_C_H____B_E_N_C_H____), name)
+    INTERNAL_CATCH_BENCHMARK_ADVANCED(INTERNAL_CATCH_UNIQUE_NAME(C_A_T_C_H_B_E_N_C_H_), name)
 #endif // CATCH_CONFIG_ENABLE_BENCHMARKING
 
 using Catch::Detail::Approx;
@@ -17719,8 +17808,8 @@ using Catch::Detail::Approx;
 #define CATCH_WARN( msg )          (void)(0)
 #define CATCH_CAPTURE( msg )       (void)(0)
 
-#define CATCH_TEST_CASE( ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ))
-#define CATCH_TEST_CASE_METHOD( className, ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ))
+#define CATCH_TEST_CASE( ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ))
+#define CATCH_TEST_CASE_METHOD( className, ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ))
 #define CATCH_METHOD_AS_TEST_CASE( method, ... )
 #define CATCH_REGISTER_TEST_CASE( Function, ... ) (void)(0)
 #define CATCH_SECTION( ... )
@@ -17729,7 +17818,7 @@ using Catch::Detail::Approx;
 #define CATCH_FAIL_CHECK( ... ) (void)(0)
 #define CATCH_SUCCEED( ... ) (void)(0)
 
-#define CATCH_ANON_TEST_CASE() INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ))
+#define CATCH_ANON_TEST_CASE() INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ))
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
 #define CATCH_TEMPLATE_TEST_CASE( ... ) INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION(__VA_ARGS__)
@@ -17752,8 +17841,8 @@ using Catch::Detail::Approx;
 #endif
 
 // "BDD-style" convenience wrappers
-#define CATCH_SCENARIO( ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ))
-#define CATCH_SCENARIO_METHOD( className, ... ) INTERNAL_CATCH_TESTCASE_METHOD_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ), className )
+#define CATCH_SCENARIO( ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ))
+#define CATCH_SCENARIO_METHOD( className, ... ) INTERNAL_CATCH_TESTCASE_METHOD_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ), className )
 #define CATCH_GIVEN( desc )
 #define CATCH_AND_GIVEN( desc )
 #define CATCH_WHEN( desc )
@@ -17803,8 +17892,8 @@ using Catch::Detail::Approx;
 #define WARN( msg ) (void)(0)
 #define CAPTURE( msg ) (void)(0)
 
-#define TEST_CASE( ... )  INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ))
-#define TEST_CASE_METHOD( className, ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ))
+#define TEST_CASE( ... )  INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ))
+#define TEST_CASE_METHOD( className, ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ))
 #define METHOD_AS_TEST_CASE( method, ... )
 #define REGISTER_TEST_CASE( Function, ... ) (void)(0)
 #define SECTION( ... )
@@ -17812,7 +17901,7 @@ using Catch::Detail::Approx;
 #define FAIL( ... ) (void)(0)
 #define FAIL_CHECK( ... ) (void)(0)
 #define SUCCEED( ... ) (void)(0)
-#define ANON_TEST_CASE() INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ))
+#define ANON_TEST_CASE() INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ))
 
 #ifndef CATCH_CONFIG_TRADITIONAL_MSVC_PREPROCESSOR
 #define TEMPLATE_TEST_CASE( ... ) INTERNAL_CATCH_TEMPLATE_TEST_CASE_NO_REGISTRATION(__VA_ARGS__)
@@ -17842,8 +17931,8 @@ using Catch::Detail::Approx;
 #define CATCH_TRANSLATE_EXCEPTION( signature ) INTERNAL_CATCH_TRANSLATE_EXCEPTION_NO_REG( INTERNAL_CATCH_UNIQUE_NAME( catch_internal_ExceptionTranslator ), signature )
 
 // "BDD-style" convenience wrappers
-#define SCENARIO( ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ) )
-#define SCENARIO_METHOD( className, ... ) INTERNAL_CATCH_TESTCASE_METHOD_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( ____C_A_T_C_H____T_E_S_T____ ), className )
+#define SCENARIO( ... ) INTERNAL_CATCH_TESTCASE_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ) )
+#define SCENARIO_METHOD( className, ... ) INTERNAL_CATCH_TESTCASE_METHOD_NO_REGISTRATION(INTERNAL_CATCH_UNIQUE_NAME( C_A_T_C_H_T_E_S_T_ ), className )
 
 #define GIVEN( desc )
 #define AND_GIVEN( desc )