Export of internal Abseil changes

--
739f9fb80212c21c015fec473e9e29803a156ef9 by Derek Mauro <dmauro@google.com>:

Define FlagStateInterface::~FlagStateInterface() in the translation
unit in which it is actually declared

Fixes #717

PiperOrigin-RevId: 319083605

--
913ef1f23113268b22d636d3ae3b992862efdb1a by Derek Mauro <dmauro@google.com>:

Fix ABSL_LOCK_RETURNED statement

PiperOrigin-RevId: 319078667

--
a43b1413da1770d638147c73e7e1693cfaf869c7 by Derek Mauro <dmauro@google.com>:

Fix redeclaration ‘absl::Cord::InlineRep::kMaxInline’, which differs in ‘constexpr’
Fixes #725

PiperOrigin-RevId: 319060910

--
1ad7d491a80f6c9de78a6fc20f09b7765d224503 by Abseil Team <absl-team@google.com>:

Make absl SpinLock trivially destructible when possible

PiperOrigin-RevId: 319049456

--
659ecad3578dfa669854a82279fa590002bdb37f by Derek Mauro <dmauro@google.com>:

Remove the static initialization of global variables used by absl::Mutex
as requested by Chromium

PiperOrigin-RevId: 319031204

--
609c491d8bb4f8bb3b44c5a4c0bee51c583df24c by Abseil Team <absl-team@google.com>:

Add implementation of %a and %A to absl::StrFormat.

Prior to this it just fell back to sprintf.

PiperOrigin-RevId: 318888039

--
5e8ae6392bcd135248aac14c4b9f2a5116868678 by Abseil Team <absl-team@google.com>:

Google-internal changes only.

PiperOrigin-RevId: 318857077

--
4a2578e33e8442954e29e5f0380ddfcf0f033f0d by Greg Falcon <gfalcon@google.com>:

Change of enum constants to accommodate internal change.

PiperOrigin-RevId: 318844716

--
4b578b102816260c213675759f4c15911735578a by Abseil Team <absl-team@google.com>:

Internal change

PiperOrigin-RevId: 318704453

--
0ee82fd24d548b260c9229fa1f54571dae1dfa24 by Gennadiy Rozental <rogeeff@google.com>:

Allow lookup of retired flags.

At the moment we issue warning on attempt to find a retired flag. This way we can't even check if flag is retired without issuing the warning.

With this change we will only issue the warning if one tries to access any functionality of retired flag but it's name "is retired" status, and type.

PiperOrigin-RevId: 318605017

--
3e35fe9b4c79f636fa328c59e2aabb93e29b7c99 by Abseil Team <absl-team@google.com>:

Fix error return from InstallSymbolDecorator().

PiperOrigin-RevId: 318490405

--
ae46063f3eb2998cb961f62a359d932e5908a4bc by Abseil Team <absl-team@google.com>:

Do not make copies of iterated collection elements into the loop variable.

PiperOrigin-RevId: 318423139

--
d06a075a12aab5f6ab98474677ce50d588b21de3 by Abseil Team <absl-team@google.com>:

add missing word making the error code better English

PiperOrigin-RevId: 318335052
GitOrigin-RevId: 739f9fb80212c21c015fec473e9e29803a156ef9
Change-Id: Id77a0a4b1959036b00555deef40e82d51884fbc1

22 files changed, 618 insertions, 94 deletions

diff --git a/absl/base/internal/low_level_scheduling.h b/absl/base/internal/low_level_scheduling.h
index 31261298..ed0b4bfa 100644
--- a/absl/base/internal/low_level_scheduling.h
+++ b/absl/base/internal/low_level_scheduling.h
@@ -29,9 +29,6 @@ extern "C" void __google_enable_rescheduling(bool disable_result);
 
 namespace absl {
 ABSL_NAMESPACE_BEGIN
-class CondVar;
-class Mutex;
-
 namespace base_internal {
 
 class SchedulingHelper;  // To allow use of SchedulingGuard.
@@ -80,8 +77,6 @@ class SchedulingGuard {
   };
 
   // Access to SchedulingGuard is explicitly permitted.
-  friend class absl::CondVar;
-  friend class absl::Mutex;
   friend class SchedulingHelper;
   friend class SpinLock;
 
diff --git a/absl/base/internal/spinlock.h b/absl/base/internal/spinlock.h
index 2222398b..e6ac9e64 100644
--- a/absl/base/internal/spinlock.h
+++ b/absl/base/internal/spinlock.h
@@ -64,7 +64,14 @@ class ABSL_LOCKABLE SpinLock {
   constexpr SpinLock(absl::ConstInitType, base_internal::SchedulingMode mode)
       : lockword_(IsCooperative(mode) ? kSpinLockCooperative : 0) {}
 
+  // For global SpinLock instances prefer trivial destructor when possible.
+  // Default but non-trivial destructor in some build configurations causes an
+  // extra static initializer.
+#ifdef ABSL_INTERNAL_HAVE_TSAN_INTERFACE
   ~SpinLock() { ABSL_TSAN_MUTEX_DESTROY(this, __tsan_mutex_not_static); }
+#else
+  ~SpinLock() = default;
+#endif
 
   // Acquire this SpinLock.
   inline void Lock() ABSL_EXCLUSIVE_LOCK_FUNCTION() {
diff --git a/absl/base/spinlock_test_common.cc b/absl/base/spinlock_test_common.cc
index b68c51a1..b33c54ba 100644
--- a/absl/base/spinlock_test_common.cc
+++ b/absl/base/spinlock_test_common.cc
@@ -20,6 +20,7 @@
 #include <limits>
 #include <random>
 #include <thread>  // NOLINT(build/c++11)
+#include <type_traits>
 #include <vector>
 
 #include "gtest/gtest.h"
@@ -103,6 +104,10 @@ static void ThreadedTest(SpinLock* spinlock) {
   }
 }
 
+#ifndef THREAD_SANITIZER
+static_assert(std::is_trivially_destructible<SpinLock>(), "");
+#endif
+
 TEST(SpinLock, StackNonCooperativeDisablesScheduling) {
   SpinLock spinlock(base_internal::SCHEDULE_KERNEL_ONLY);
   spinlock.Lock();
diff --git a/absl/container/internal/hash_function_defaults_test.cc b/absl/container/internal/hash_function_defaults_test.cc
index 2d05a0b7..59576b8e 100644
--- a/absl/container/internal/hash_function_defaults_test.cc
+++ b/absl/container/internal/hash_function_defaults_test.cc
@@ -337,11 +337,11 @@ ABSL_NAMESPACE_END
 }  // namespace absl
 
 enum Hash : size_t {
-  kStd = 0x2,       // std::hash
+  kStd = 0x1,       // std::hash
 #ifdef _MSC_VER
   kExtension = kStd,  // In MSVC, std::hash == ::hash
 #else                 // _MSC_VER
-  kExtension = 0x4,  // ::hash (GCC extension)
+  kExtension = 0x2,  // ::hash (GCC extension)
 #endif                // _MSC_VER
 };
 
diff --git a/absl/debugging/symbolize_elf.inc b/absl/debugging/symbolize_elf.inc
index ed77159e..328869f1 100644
--- a/absl/debugging/symbolize_elf.inc
+++ b/absl/debugging/symbolize_elf.inc
@@ -1376,7 +1376,7 @@ int InstallSymbolDecorator(SymbolDecorator decorator, void *arg) {
 
   if (!g_decorators_mu.TryLock()) {
     // Someone else is using decorators. Get out.
-    return false;
+    return -2;
   }
   int ret = ticket;
   if (g_num_decorators >= kMaxDecorators) {
diff --git a/absl/flags/BUILD.bazel b/absl/flags/BUILD.bazel
index 006911fd..524e7027 100644
--- a/absl/flags/BUILD.bazel
+++ b/absl/flags/BUILD.bazel
@@ -106,6 +106,9 @@ cc_library(
 
 cc_library(
     name = "commandlineflag_internal",
+    srcs = [
+        "internal/commandlineflag.cc",
+    ],
     hdrs = [
         "internal/commandlineflag.h",
     ],
diff --git a/absl/flags/CMakeLists.txt b/absl/flags/CMakeLists.txt
index ef75db8e..343774de 100644
--- a/absl/flags/CMakeLists.txt
+++ b/absl/flags/CMakeLists.txt
@@ -95,6 +95,8 @@ absl_cc_library(
 absl_cc_library(
   NAME
     flags_commandlineflag_internal
+  SRCS
+    "internal/commandlineflag.cc"
   HDRS
     "internal/commandlineflag.h"
   COPTS
diff --git a/absl/flags/commandlineflag.cc b/absl/flags/commandlineflag.cc
index 217b2d87..9f3b4a5a 100644
--- a/absl/flags/commandlineflag.cc
+++ b/absl/flags/commandlineflag.cc
@@ -30,9 +30,5 @@ bool CommandLineFlag::ParseFrom(absl::string_view value, std::string* error) {
                    flags_internal::kProgrammaticChange, *error);
 }
 
-namespace flags_internal {
-FlagStateInterface::~FlagStateInterface() {}
-}  // namespace flags_internal
 ABSL_NAMESPACE_END
 }  // namespace absl
-
diff --git a/absl/flags/commandlineflag.h b/absl/flags/commandlineflag.h
index 7e21d05d..f2fa0897 100644
--- a/absl/flags/commandlineflag.h
+++ b/absl/flags/commandlineflag.h
@@ -108,6 +108,10 @@ class CommandLineFlag {
     U u;
 
     Read(&u.value);
+    // allow retired flags to be "read", so we can report invalid access.
+    if (IsRetired()) {
+      return absl::nullopt;
+    }
     return std::move(u.value);
   }
 
diff --git a/absl/flags/internal/commandlineflag.cc b/absl/flags/internal/commandlineflag.cc
new file mode 100644
index 00000000..4482955c
--- /dev/null
+++ b/absl/flags/internal/commandlineflag.cc
@@ -0,0 +1,26 @@
+//
+// Copyright 2020 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      https://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/flags/internal/commandlineflag.h"
+
+namespace absl {
+ABSL_NAMESPACE_BEGIN
+namespace flags_internal {
+
+FlagStateInterface::~FlagStateInterface() {}
+
+}  // namespace flags_internal
+ABSL_NAMESPACE_END
+}  // namespace absl
diff --git a/absl/flags/internal/flag.h b/absl/flags/internal/flag.h
index 2cc44e00..89e43ad7 100644
--- a/absl/flags/internal/flag.h
+++ b/absl/flags/internal/flag.h
@@ -482,7 +482,8 @@ class FlagImpl final : public CommandLineFlag {
   friend class FlagState;
 
   // Ensures that `data_guard_` is initialized and returns it.
-  absl::Mutex* DataGuard() const ABSL_LOCK_RETURNED((absl::Mutex*)&data_guard_);
+  absl::Mutex* DataGuard() const
+      ABSL_LOCK_RETURNED(reinterpret_cast<absl::Mutex*>(data_guard_));
   // Returns heap allocated value of type T initialized with default value.
   std::unique_ptr<void, DynValueDeleter> MakeInitValue() const
       ABSL_EXCLUSIVE_LOCKS_REQUIRED(*DataGuard());
diff --git a/absl/flags/internal/usage.cc b/absl/flags/internal/usage.cc
index 35b6427b..0805df31 100644
--- a/absl/flags/internal/usage.cc
+++ b/absl/flags/internal/usage.cc
@@ -250,14 +250,14 @@ void FlagsHelpImpl(std::ostream& out, flags_internal::FlagKindFilter filter_cb,
       matching_flags;
 
   flags_internal::ForEachFlag([&](absl::CommandLineFlag& flag) {
-    std::string flag_filename = flag.Filename();
-
     // Ignore retired flags.
     if (flag.IsRetired()) return;
 
     // If the flag has been stripped, pretend that it doesn't exist.
     if (flag.Help() == flags_internal::kStrippedFlagHelp) return;
 
+    std::string flag_filename = flag.Filename();
+
     // Make sure flag satisfies the filter
     if (!filter_cb || !filter_cb(flag_filename)) return;
 
diff --git a/absl/flags/parse.cc b/absl/flags/parse.cc
index e2c88ff8..4f4bb3d5 100644
--- a/absl/flags/parse.cc
+++ b/absl/flags/parse.cc
@@ -729,12 +729,13 @@ std::vector<char*> ParseCommandLineImpl(int argc, char* argv[],
     }
 
     // 100. Set the located flag to a new new value, unless it is retired.
-    // Setting retired flag fails, but we ignoring it here.
-    if (flag->IsRetired()) continue;
-
+    // Setting retired flag fails, but we ignoring it here while also reporting
+    // access to retired flag.
     std::string error;
     if (!flags_internal::PrivateHandleAccessor::ParseFrom(
             *flag, value, SET_FLAGS_VALUE, kCommandLine, error)) {
+      if (flag->IsRetired()) continue;
+
       flags_internal::ReportUsageError(error, true);
       success = false;
     } else {
diff --git a/absl/flags/reflection.cc b/absl/flags/reflection.cc
index 5fc945f2..02b7c06a 100644
--- a/absl/flags/reflection.cc
+++ b/absl/flags/reflection.cc
@@ -88,12 +88,6 @@ CommandLineFlag* FlagRegistry::FindFlagLocked(absl::string_view name) {
   if (i == flags_.end()) {
     return nullptr;
   }
-
-  if (i->second->IsRetired()) {
-    flags_internal::ReportUsageError(
-        absl::StrCat("Accessing retired flag '", name, "'"), false);
-  }
-
   return i->second;
 }
 
@@ -155,7 +149,7 @@ void FlagRegistry::RegisterFlag(CommandLineFlag& flag) {
     } else {
       flags_internal::ReportUsageError(
           absl::StrCat(
-              "Something wrong with flag '", flag.Name(), "' in file '",
+              "Something is wrong with flag '", flag.Name(), "' in file '",
               flag.Filename(), "'. One possibility: file '", flag.Filename(),
               "' is being linked both statically and dynamically into this "
               "executable. e.g. some files listed as srcs to a test and also "
@@ -206,16 +200,34 @@ class RetiredFlagObj final : public CommandLineFlag {
 
  private:
   absl::string_view Name() const override { return name_; }
-  std::string Filename() const override { return "RETIRED"; }
+  std::string Filename() const override {
+    OnAccess();
+    return "RETIRED";
+  }
   FlagFastTypeId TypeId() const override { return type_id_; }
-  std::string Help() const override { return ""; }
+  std::string Help() const override {
+    OnAccess();
+    return "";
+  }
   bool IsRetired() const override { return true; }
-  bool IsSpecifiedOnCommandLine() const override { return false; }
-  std::string DefaultValue() const override { return ""; }
-  std::string CurrentValue() const override { return ""; }
+  bool IsSpecifiedOnCommandLine() const override {
+    OnAccess();
+    return false;
+  }
+  std::string DefaultValue() const override {
+    OnAccess();
+    return "";
+  }
+  std::string CurrentValue() const override {
+    OnAccess();
+    return "";
+  }
 
   // Any input is valid
-  bool ValidateInputValue(absl::string_view) const override { return true; }
+  bool ValidateInputValue(absl::string_view) const override {
+    OnAccess();
+    return true;
+  }
 
   std::unique_ptr<flags_internal::FlagStateInterface> SaveState() override {
     return nullptr;
@@ -223,12 +235,18 @@ class RetiredFlagObj final : public CommandLineFlag {
 
   bool ParseFrom(absl::string_view, flags_internal::FlagSettingMode,
                  flags_internal::ValueSource, std::string&) override {
+    OnAccess();
     return false;
   }
 
-  void CheckDefaultValueParsingRoundtrip() const override {}
+  void CheckDefaultValueParsingRoundtrip() const override { OnAccess(); }
 
-  void Read(void*) const override {}
+  void Read(void*) const override { OnAccess(); }
+
+  void OnAccess() const {
+    flags_internal::ReportUsageError(
+        absl::StrCat("Accessing retired flag '", name_, "'"), false);
+  }
 
   // Data members
   const char* const name_;
diff --git a/absl/random/BUILD.bazel b/absl/random/BUILD.bazel
index 694331c2..81e150e6 100644
--- a/absl/random/BUILD.bazel
+++ b/absl/random/BUILD.bazel
@@ -199,6 +199,7 @@ cc_test(
 cc_test(
     name = "distributions_test",
     size = "small",
+    timeout = "moderate",
     srcs = [
         "distributions_test.cc",
     ],
diff --git a/absl/strings/BUILD.bazel b/absl/strings/BUILD.bazel
index 8220896d..ef412639 100644
--- a/absl/strings/BUILD.bazel
+++ b/absl/strings/BUILD.bazel
@@ -720,6 +720,7 @@ cc_test(
     visibility = ["//visibility:private"],
     deps = [
         ":str_format_internal",
+        ":strings",
         "//absl/base:raw_logging_internal",
         "//absl/types:optional",
         "@com_google_googletest//:gtest_main",
diff --git a/absl/strings/CMakeLists.txt b/absl/strings/CMakeLists.txt
index c0ea0c8e..d7237231 100644
--- a/absl/strings/CMakeLists.txt
+++ b/absl/strings/CMakeLists.txt
@@ -475,6 +475,7 @@ absl_cc_test(
   COPTS
     ${ABSL_TEST_COPTS}
   DEPS
+    absl::strings
     absl::str_format_internal
     absl::raw_logging_internal
     absl::int128
diff --git a/absl/strings/cord.cc b/absl/strings/cord.cc
index 68f53987..8ecffc4b 100644
--- a/absl/strings/cord.cc
+++ b/absl/strings/cord.cc
@@ -495,7 +495,7 @@ static CordRep* NewSubstring(CordRep* child, size_t offset, size_t length) {
 
 // This will trigger LNK2005 in MSVC.
 #ifndef COMPILER_MSVC
-const unsigned char Cord::InlineRep::kMaxInline;
+constexpr unsigned char Cord::InlineRep::kMaxInline;
 #endif  // COMPILER_MSVC
 
 inline void Cord::InlineRep::set_data(const char* data, size_t n,
diff --git a/absl/strings/internal/str_format/convert_test.cc b/absl/strings/internal/str_format/convert_test.cc
index e37d0546..488d4cd4 100644
--- a/absl/strings/internal/str_format/convert_test.cc
+++ b/absl/strings/internal/str_format/convert_test.cc
@@ -12,6 +12,7 @@
 #include "gtest/gtest.h"
 #include "absl/base/internal/raw_logging.h"
 #include "absl/strings/internal/str_format/bind.h"
+#include "absl/strings/match.h"
 #include "absl/types/optional.h"
 
 namespace absl {
@@ -19,6 +20,13 @@ ABSL_NAMESPACE_BEGIN
 namespace str_format_internal {
 namespace {
 
+struct NativePrintfTraits {
+  bool hex_float_has_glibc_rounding;
+  bool hex_float_prefers_denormal_repr;
+  bool hex_float_uses_minimal_precision_when_not_specified;
+  bool hex_float_optimizes_leading_digit_bit_count;
+};
+
 template <typename T, size_t N>
 size_t ArraySize(T (&)[N]) {
   return N;
@@ -118,6 +126,63 @@ std::string StrPrint(const char *format, ...) {
   return result;
 }
 
+NativePrintfTraits VerifyNativeImplementationImpl() {
+  NativePrintfTraits result;
+
+  // >>> hex_float_has_glibc_rounding. To have glibc's rounding behavior we need
+  // to meet three requirements:
+  //
+  //   - The threshold for rounding up is 8 (for e.g. MSVC uses 9).
+  //   - If the digits lower than than the 8 are non-zero then we round up.
+  //   - If the digits lower than the 8 are all zero then we round toward even.
+  //
+  // The numbers below represent all the cases covering {below,at,above} the
+  // threshold (8) with both {zero,non-zero} lower bits and both {even,odd}
+  // preceding digits.
+  const double d0079 = 65657.0;  // 0x1.0079p+16
+  const double d0179 = 65913.0;  // 0x1.0179p+16
+  const double d0080 = 65664.0;  // 0x1.0080p+16
+  const double d0180 = 65920.0;  // 0x1.0180p+16
+  const double d0081 = 65665.0;  // 0x1.0081p+16
+  const double d0181 = 65921.0;  // 0x1.0181p+16
+  result.hex_float_has_glibc_rounding =
+      StartsWith(StrPrint("%.2a", d0079), "0x1.00") &&
+      StartsWith(StrPrint("%.2a", d0179), "0x1.01") &&
+      StartsWith(StrPrint("%.2a", d0080), "0x1.00") &&
+      StartsWith(StrPrint("%.2a", d0180), "0x1.02") &&
+      StartsWith(StrPrint("%.2a", d0081), "0x1.01") &&
+      StartsWith(StrPrint("%.2a", d0181), "0x1.02");
+
+  // >>> hex_float_prefers_denormal_repr. Formatting `denormal` on glibc yields
+  // "0x0.0000000000001p-1022", whereas on std libs that don't use denormal
+  // representation it would either be 0x1p-1074 or 0x1.0000000000000-1074.
+  const double denormal = std::numeric_limits<double>::denorm_min();
+  result.hex_float_prefers_denormal_repr =
+      StartsWith(StrPrint("%a", denormal), "0x0.0000000000001");
+
+  // >>> hex_float_uses_minimal_precision_when_not_specified. Some (non-glibc)
+  // libs will format the following as "0x1.0079000000000p+16".
+  result.hex_float_uses_minimal_precision_when_not_specified =
+      (StrPrint("%a", d0079) == "0x1.0079p+16");
+
+  // >>> hex_float_optimizes_leading_digit_bit_count. The number 1.5, when
+  // formatted by glibc should yield "0x1.8p+0" for `double` and "0xcp-3" for
+  // `long double`, i.e., number of bits in the leading digit is adapted to the
+  // number of bits in the mantissa.
+  const double d_15 = 1.5;
+  const long double ld_15 = 1.5;
+  result.hex_float_optimizes_leading_digit_bit_count =
+      StartsWith(StrPrint("%a", d_15), "0x1.8") &&
+      StartsWith(StrPrint("%La", ld_15), "0xc");
+
+  return result;
+}
+
+const NativePrintfTraits &VerifyNativeImplementation() {
+  static NativePrintfTraits native_traits = VerifyNativeImplementationImpl();
+  return native_traits;
+}
+
 class FormatConvertTest : public ::testing::Test { };
 
 template <typename T>
@@ -476,6 +541,7 @@ TEST_F(FormatConvertTest, Uint128) {
 
 template <typename Floating>
 void TestWithMultipleFormatsHelper(const std::vector<Floating> &floats) {
+  const NativePrintfTraits &native_traits = VerifyNativeImplementation();
   // Reserve the space to ensure we don't allocate memory in the output itself.
   std::string str_format_result;
   str_format_result.reserve(1 << 20);
@@ -493,13 +559,23 @@ void TestWithMultipleFormatsHelper(const std::vector<Floating> &floats) {
                    'e', 'E'}) {
       std::string fmt_str = std::string(fmt) + f;
 
-      if (fmt == absl::string_view("%.5000") && f != 'f' && f != 'F') {
+      if (fmt == absl::string_view("%.5000") && f != 'f' && f != 'F' &&
+          f != 'a' && f != 'A') {
         // This particular test takes way too long with snprintf.
         // Disable for the case we are not implementing natively.
         continue;
       }
 
+      if ((f == 'a' || f == 'A') &&
+          !native_traits.hex_float_has_glibc_rounding) {
+        continue;
+      }
+
       for (Floating d : floats) {
+        if (!native_traits.hex_float_prefers_denormal_repr &&
+            (f == 'a' || f == 'A') && std::fpclassify(d) == FP_SUBNORMAL) {
+          continue;
+        }
         int i = -10;
         FormatArgImpl args[2] = {FormatArgImpl(d), FormatArgImpl(i)};
         UntypedFormatSpecImpl format(fmt_str);
@@ -766,6 +842,111 @@ TEST_F(FormatConvertTest, DoubleRound) {
             "1837869002408041296803276054561138153076171875");
 }
 
+TEST_F(FormatConvertTest, DoubleRoundA) {
+  const NativePrintfTraits &native_traits = VerifyNativeImplementation();
+  std::string s;
+  const auto format = [&](const char *fmt, double d) -> std::string & {
+    s.clear();
+    FormatArgImpl args[1] = {FormatArgImpl(d)};
+    AppendPack(&s, UntypedFormatSpecImpl(fmt), absl::MakeSpan(args));
+    if (native_traits.hex_float_has_glibc_rounding) {
+      EXPECT_EQ(StrPrint(fmt, d), s);
+    }
+    return s;
+  };
+
+  // 0x1.00018000p+100
+  const double on_boundary_odd = 1267679614447900152596896153600.0;
+  EXPECT_EQ(format("%.0a", on_boundary_odd), "0x1p+100");
+  EXPECT_EQ(format("%.1a", on_boundary_odd), "0x1.0p+100");
+  EXPECT_EQ(format("%.2a", on_boundary_odd), "0x1.00p+100");
+  EXPECT_EQ(format("%.3a", on_boundary_odd), "0x1.000p+100");
+  EXPECT_EQ(format("%.4a", on_boundary_odd), "0x1.0002p+100");  // round
+  EXPECT_EQ(format("%.5a", on_boundary_odd), "0x1.00018p+100");
+  EXPECT_EQ(format("%.6a", on_boundary_odd), "0x1.000180p+100");
+
+  // 0x1.00028000p-2
+  const double on_boundary_even = 0.250009536743164062500;
+  EXPECT_EQ(format("%.0a", on_boundary_even), "0x1p-2");
+  EXPECT_EQ(format("%.1a", on_boundary_even), "0x1.0p-2");
+  EXPECT_EQ(format("%.2a", on_boundary_even), "0x1.00p-2");
+  EXPECT_EQ(format("%.3a", on_boundary_even), "0x1.000p-2");
+  EXPECT_EQ(format("%.4a", on_boundary_even), "0x1.0002p-2");  // no round
+  EXPECT_EQ(format("%.5a", on_boundary_even), "0x1.00028p-2");
+  EXPECT_EQ(format("%.6a", on_boundary_even), "0x1.000280p-2");
+
+  // 0x1.00018001p+1
+  const double slightly_over = 2.00004577683284878730773925781250;
+  EXPECT_EQ(format("%.0a", slightly_over), "0x1p+1");
+  EXPECT_EQ(format("%.1a", slightly_over), "0x1.0p+1");
+  EXPECT_EQ(format("%.2a", slightly_over), "0x1.00p+1");
+  EXPECT_EQ(format("%.3a", slightly_over), "0x1.000p+1");
+  EXPECT_EQ(format("%.4a", slightly_over), "0x1.0002p+1");
+  EXPECT_EQ(format("%.5a", slightly_over), "0x1.00018p+1");
+  EXPECT_EQ(format("%.6a", slightly_over), "0x1.000180p+1");
+
+  // 0x1.00017fffp+0
+  const double slightly_under = 1.000022887950763106346130371093750;
+  EXPECT_EQ(format("%.0a", slightly_under), "0x1p+0");
+  EXPECT_EQ(format("%.1a", slightly_under), "0x1.0p+0");
+  EXPECT_EQ(format("%.2a", slightly_under), "0x1.00p+0");
+  EXPECT_EQ(format("%.3a", slightly_under), "0x1.000p+0");
+  EXPECT_EQ(format("%.4a", slightly_under), "0x1.0001p+0");
+  EXPECT_EQ(format("%.5a", slightly_under), "0x1.00018p+0");
+  EXPECT_EQ(format("%.6a", slightly_under), "0x1.000180p+0");
+  EXPECT_EQ(format("%.7a", slightly_under), "0x1.0001800p+0");
+
+  // 0x1.1b3829ac28058p+3
+  const double hex_value = 8.85060580848964661981881363317370414733886718750;
+  EXPECT_EQ(format("%.0a", hex_value), "0x1p+3");
+  EXPECT_EQ(format("%.1a", hex_value), "0x1.2p+3");
+  EXPECT_EQ(format("%.2a", hex_value), "0x1.1bp+3");
+  EXPECT_EQ(format("%.3a", hex_value), "0x1.1b4p+3");
+  EXPECT_EQ(format("%.4a", hex_value), "0x1.1b38p+3");
+  EXPECT_EQ(format("%.5a", hex_value), "0x1.1b383p+3");
+  EXPECT_EQ(format("%.6a", hex_value), "0x1.1b382ap+3");
+  EXPECT_EQ(format("%.7a", hex_value), "0x1.1b3829bp+3");
+  EXPECT_EQ(format("%.8a", hex_value), "0x1.1b3829acp+3");
+  EXPECT_EQ(format("%.9a", hex_value), "0x1.1b3829ac3p+3");
+  EXPECT_EQ(format("%.10a", hex_value), "0x1.1b3829ac28p+3");
+  EXPECT_EQ(format("%.11a", hex_value), "0x1.1b3829ac280p+3");
+  EXPECT_EQ(format("%.12a", hex_value), "0x1.1b3829ac2806p+3");
+  EXPECT_EQ(format("%.13a", hex_value), "0x1.1b3829ac28058p+3");
+  EXPECT_EQ(format("%.14a", hex_value), "0x1.1b3829ac280580p+3");
+  EXPECT_EQ(format("%.15a", hex_value), "0x1.1b3829ac2805800p+3");
+  EXPECT_EQ(format("%.16a", hex_value), "0x1.1b3829ac28058000p+3");
+  EXPECT_EQ(format("%.17a", hex_value), "0x1.1b3829ac280580000p+3");
+  EXPECT_EQ(format("%.18a", hex_value), "0x1.1b3829ac2805800000p+3");
+  EXPECT_EQ(format("%.19a", hex_value), "0x1.1b3829ac28058000000p+3");
+  EXPECT_EQ(format("%.20a", hex_value), "0x1.1b3829ac280580000000p+3");
+  EXPECT_EQ(format("%.21a", hex_value), "0x1.1b3829ac2805800000000p+3");
+
+  // 0x1.0818283848586p+3
+  const double hex_value2 = 8.2529488658208371987257123691961169242858886718750;
+  EXPECT_EQ(format("%.0a", hex_value2), "0x1p+3");
+  EXPECT_EQ(format("%.1a", hex_value2), "0x1.1p+3");
+  EXPECT_EQ(format("%.2a", hex_value2), "0x1.08p+3");
+  EXPECT_EQ(format("%.3a", hex_value2), "0x1.082p+3");
+  EXPECT_EQ(format("%.4a", hex_value2), "0x1.0818p+3");
+  EXPECT_EQ(format("%.5a", hex_value2), "0x1.08183p+3");
+  EXPECT_EQ(format("%.6a", hex_value2), "0x1.081828p+3");
+  EXPECT_EQ(format("%.7a", hex_value2), "0x1.0818284p+3");
+  EXPECT_EQ(format("%.8a", hex_value2), "0x1.08182838p+3");
+  EXPECT_EQ(format("%.9a", hex_value2), "0x1.081828385p+3");
+  EXPECT_EQ(format("%.10a", hex_value2), "0x1.0818283848p+3");
+  EXPECT_EQ(format("%.11a", hex_value2), "0x1.08182838486p+3");
+  EXPECT_EQ(format("%.12a", hex_value2), "0x1.081828384858p+3");
+  EXPECT_EQ(format("%.13a", hex_value2), "0x1.0818283848586p+3");
+  EXPECT_EQ(format("%.14a", hex_value2), "0x1.08182838485860p+3");
+  EXPECT_EQ(format("%.15a", hex_value2), "0x1.081828384858600p+3");
+  EXPECT_EQ(format("%.16a", hex_value2), "0x1.0818283848586000p+3");
+  EXPECT_EQ(format("%.17a", hex_value2), "0x1.08182838485860000p+3");
+  EXPECT_EQ(format("%.18a", hex_value2), "0x1.081828384858600000p+3");
+  EXPECT_EQ(format("%.19a", hex_value2), "0x1.0818283848586000000p+3");
+  EXPECT_EQ(format("%.20a", hex_value2), "0x1.08182838485860000000p+3");
+  EXPECT_EQ(format("%.21a", hex_value2), "0x1.081828384858600000000p+3");
+}
+
 // We don't actually store the results. This is just to exercise the rest of the
 // machinery.
 struct NullSink {
@@ -797,6 +978,7 @@ TEST_F(FormatConvertTest, LongDouble) {
   // implementation against the native one there.
   return;
 #endif  // _MSC_VER
+  const NativePrintfTraits &native_traits = VerifyNativeImplementation();
   const char *const kFormats[] = {"%",    "%.3", "%8.5", "%9",  "%.5000",
                                   "%.60", "%+",  "% ",   "%-10"};
 
@@ -839,12 +1021,20 @@ TEST_F(FormatConvertTest, LongDouble) {
                    'e', 'E'}) {
       std::string fmt_str = std::string(fmt) + 'L' + f;
 
-      if (fmt == absl::string_view("%.5000") && f != 'f' && f != 'F') {
+      if (fmt == absl::string_view("%.5000") && f != 'f' && f != 'F' &&
+          f != 'a' && f != 'A') {
         // This particular test takes way too long with snprintf.
         // Disable for the case we are not implementing natively.
         continue;
       }
 
+      if (f == 'a' || f == 'A') {
+        if (!native_traits.hex_float_has_glibc_rounding ||
+            !native_traits.hex_float_optimizes_leading_digit_bit_count) {
+          continue;
+        }
+      }
+
       for (auto d : doubles) {
         FormatArgImpl arg(d);
         UntypedFormatSpecImpl format(fmt_str);
@@ -860,6 +1050,7 @@ TEST_F(FormatConvertTest, LongDouble) {
 }
 
 TEST_F(FormatConvertTest, IntAsDouble) {
+  const NativePrintfTraits &native_traits = VerifyNativeImplementation();
   const int kMin = std::numeric_limits<int>::min();
   const int kMax = std::numeric_limits<int>::max();
   const int ia[] = {
@@ -875,14 +1066,16 @@ TEST_F(FormatConvertTest, IntAsDouble) {
       const char *fmt;
     };
     const double dx = static_cast<double>(fx);
-    const Expectation kExpect[] = {
-      { __LINE__, StrPrint("%f", dx), "%f" },
-      { __LINE__, StrPrint("%12f", dx), "%12f" },
-      { __LINE__, StrPrint("%.12f", dx), "%.12f" },
-      { __LINE__, StrPrint("%12a", dx), "%12a" },
-      { __LINE__, StrPrint("%.12a", dx), "%.12a" },
+    std::vector<Expectation> expect = {
+        {__LINE__, StrPrint("%f", dx), "%f"},
+        {__LINE__, StrPrint("%12f", dx), "%12f"},
+        {__LINE__, StrPrint("%.12f", dx), "%.12f"},
+        {__LINE__, StrPrint("%.12a", dx), "%.12a"},
     };
-    for (const Expectation &e : kExpect) {
+    if (native_traits.hex_float_uses_minimal_precision_when_not_specified) {
+      expect.push_back({__LINE__, StrPrint("%12a", dx), "%12a"});
+    }
+    for (const Expectation &e : expect) {
       SCOPED_TRACE(e.line);
       SCOPED_TRACE(e.fmt);
       UntypedFormatSpecImpl format(e.fmt);
@@ -927,6 +1120,25 @@ TEST_F(FormatConvertTest, ExpectedFailures) {
   EXPECT_TRUE(FormatFails("%*d", ""));
 }
 
+// Sanity check to make sure that we are testing what we think we're testing on
+// e.g. the x86_64+glibc platform.
+TEST_F(FormatConvertTest, GlibcHasCorrectTraits) {
+#if !defined(__GLIBC__) || !defined(__x86_64__)
+  return;
+#endif
+  const NativePrintfTraits &native_traits = VerifyNativeImplementation();
+  // If one of the following tests break then it is either because the above PP
+  // macro guards failed to exclude a new platform (likely) or because something
+  // has changed in the implemention of glibc sprintf float formatting behavior.
+  // If the latter, then the code that computes these flags needs to be
+  // revisited and/or possibly the StrFormat implementation.
+  EXPECT_TRUE(native_traits.hex_float_has_glibc_rounding);
+  EXPECT_TRUE(native_traits.hex_float_prefers_denormal_repr);
+  EXPECT_TRUE(
+      native_traits.hex_float_uses_minimal_precision_when_not_specified);
+  EXPECT_TRUE(native_traits.hex_float_optimizes_leading_digit_bit_count);
+}
+
 }  // namespace
 }  // namespace str_format_internal
 ABSL_NAMESPACE_END
diff --git a/absl/strings/internal/str_format/float_conversion.cc b/absl/strings/internal/str_format/float_conversion.cc
index 39fc5f60..6eb7b9fc 100644
--- a/absl/strings/internal/str_format/float_conversion.cc
+++ b/absl/strings/internal/str_format/float_conversion.cc
@@ -15,6 +15,7 @@
 #include "absl/functional/function_ref.h"
 #include "absl/meta/type_traits.h"
 #include "absl/numeric/int128.h"
+#include "absl/strings/numbers.h"
 #include "absl/types/optional.h"
 #include "absl/types/span.h"
 
@@ -453,26 +454,31 @@ Padding ExtraWidthToPadding(size_t total_size, const FormatState &state) {
   }
 }
 
-void FinalPrint(absl::string_view data, int trailing_zeros,
-                const FormatState &state) {
+void FinalPrint(const FormatState &state, absl::string_view data,
+                int padding_offset, int trailing_zeros,
+                absl::string_view data_postfix) {
   if (state.conv.width() < 0) {
     // No width specified. Fast-path.
     if (state.sign_char != '\0') state.sink->Append(1, state.sign_char);
     state.sink->Append(data);
     state.sink->Append(trailing_zeros, '0');
+    state.sink->Append(data_postfix);
     return;
   }
 
-  auto padding =
-      ExtraWidthToPadding((state.sign_char != '\0' ? 1 : 0) + data.size() +
-                              static_cast<size_t>(trailing_zeros),
-                          state);
+  auto padding = ExtraWidthToPadding((state.sign_char != '\0' ? 1 : 0) +
+                                         data.size() + data_postfix.size() +
+                                         static_cast<size_t>(trailing_zeros),
+                                     state);
 
   state.sink->Append(padding.left_spaces, ' ');
   if (state.sign_char != '\0') state.sink->Append(1, state.sign_char);
+  // Padding in general needs to be inserted somewhere in the middle of `data`.
+  state.sink->Append(data.substr(0, padding_offset));
   state.sink->Append(padding.zeros, '0');
-  state.sink->Append(data);
+  state.sink->Append(data.substr(padding_offset));
   state.sink->Append(trailing_zeros, '0');
+  state.sink->Append(data_postfix);
   state.sink->Append(padding.right_spaces, ' ');
 }
 
@@ -525,10 +531,11 @@ void FormatFFast(Int v, int exp, const FormatState &state) {
   // In `alt` mode (flag #) we keep the `.` even if there are no fractional
   // digits. In non-alt mode, we strip it.
   if (!state.ShouldPrintDot()) --size;
-  FinalPrint(absl::string_view(integral_digits_start, size),
+  FinalPrint(state, absl::string_view(integral_digits_start, size),
+             /*padding_offset=*/0,
              static_cast<int>(state.precision - (fractional_digits_end -
                                                  fractional_digits_start)),
-             state);
+             /*data_postfix=*/"");
 }
 
 // Slow %f formatter for when the shifted value does not fit in a uint128, and
@@ -655,6 +662,257 @@ void FormatF(Int mantissa, int exp, const FormatState &state) {
   return FormatFFast(mantissa, exp, state);
 }
 
+// Grab the group of four bits (nibble) from `n`. E.g., nibble 1 corresponds to
+// bits 4-7.
+template <typename Int>
+uint8_t GetNibble(Int n, int nibble_index) {
+  constexpr Int mask_low_nibble = Int{0xf};
+  int shift = nibble_index * 4;
+  n &= mask_low_nibble << shift;
+  return static_cast<uint8_t>((n >> shift) & 0xf);
+}
+
+// Add one to the given nibble, applying carry to higher nibbles. Returns true
+// if overflow, false otherwise.
+template <typename Int>
+bool IncrementNibble(int nibble_index, Int *n) {
+  constexpr int kShift = sizeof(Int) * 8 - 1;
+  constexpr int kNumNibbles = sizeof(Int) * 8 / 4;
+  Int before = *n >> kShift;
+  // Here we essentially want to take the number 1 and move it into the requsted
+  // nibble, then add it to *n to effectively increment the nibble. However,
+  // ASan will complain if we try to shift the 1 beyond the limits of the Int,
+  // i.e., if the nibble_index is out of range. So therefore we check for this
+  // and if we are out of range we just add 0 which leaves *n unchanged, which
+  // seems like the reasonable thing to do in that case.
+  *n +=
+      ((nibble_index * 4 >= sizeof(Int) * 8) ? 0
+                                             : (Int{1} << (nibble_index * 4)));
+  Int after = *n >> kShift;
+  return (before && !after) || (nibble_index >= kNumNibbles);
+}
+
+// Return a mask with 1's in the given nibble and all lower nibbles.
+template <typename Int>
+Int MaskUpToNibbleInclusive(int nibble_index) {
+  constexpr int kNumNibbles = sizeof(Int) * 8 / 4;
+  static const Int ones = ~Int{0};
+  return ones >> std::max(0, 4 * (kNumNibbles - nibble_index - 1));
+}
+
+// Return a mask with 1's below the given nibble.
+template <typename Int>
+Int MaskUpToNibbleExclusive(int nibble_index) {
+  return nibble_index <= 0 ? 0 : MaskUpToNibbleInclusive<Int>(nibble_index - 1);
+}
+
+template <typename Int>
+Int MoveToNibble(uint8_t nibble, int nibble_index) {
+  return Int{nibble} << (4 * nibble_index);
+}
+
+// Given mantissa size, find optimal # of mantissa bits to put in initial digit.
+//
+// In the hex representation we keep a single hex digit to the left of the dot.
+// However, the question as to how many bits of the mantissa should be put into
+// that hex digit in theory is arbitrary, but in practice it is optimal to
+// choose based on the size of the mantissa. E.g., for a `double`, there are 53
+// mantissa bits, so that means that we should put 1 bit to the left of the dot,
+// thereby leaving 52 bits to the right, which is evenly divisible by four and
+// thus all fractional digits represent actual precision. For a `long double`,
+// on the other hand, there are 64 bits of mantissa, thus we can use all four
+// bits for the initial hex digit and still have a number left over (60) that is
+// a multiple of four. Once again, the goal is to have all fractional digits
+// represent real precision.
+template <typename Float>
+constexpr int HexFloatLeadingDigitSizeInBits() {
+  return std::numeric_limits<Float>::digits % 4 > 0
+             ? std::numeric_limits<Float>::digits % 4
+             : 4;
+}
+
+// This function captures the rounding behavior of glibc for hex float
+// representations. E.g. when rounding 0x1.ab800000 to a precision of .2
+// ("%.2a") glibc will round up because it rounds toward the even number (since
+// 0xb is an odd number, it will round up to 0xc). However, when rounding at a
+// point that is not followed by 800000..., it disregards the parity and rounds
+// up if > 8 and rounds down if < 8.
+template <typename Int>
+bool HexFloatNeedsRoundUp(Int mantissa, int final_nibble_displayed) {
+  // If the last nibble (hex digit) to be displayed is the lowest on in the
+  // mantissa then that means that we don't have any further nibbles to inform
+  // rounding, so don't round.
+  if (final_nibble_displayed <= 0) {
+    return false;
+  }
+  int rounding_nibble_idx = final_nibble_displayed - 1;
+  constexpr int kTotalNibbles = sizeof(Int) * 8 / 4;
+  assert(final_nibble_displayed <= kTotalNibbles);
+  Int mantissa_up_to_rounding_nibble_inclusive =
+      mantissa & MaskUpToNibbleInclusive<Int>(rounding_nibble_idx);
+  Int eight = MoveToNibble<Int>(8, rounding_nibble_idx);
+  if (mantissa_up_to_rounding_nibble_inclusive != eight) {
+    return mantissa_up_to_rounding_nibble_inclusive > eight;
+  }
+  // Nibble in question == 8.
+  uint8_t should_round_at_8 =
+      (final_nibble_displayed >= kTotalNibbles)
+          ? true
+          : (GetNibble(mantissa, final_nibble_displayed) % 2 == 1);
+  return should_round_at_8;
+}
+
+// Stores values associated with a Float type needed by the FormatA
+// implementation in order to avoid templatizing that function by the Float
+// type.
+struct HexFloatTypeParams {
+  template <typename Float>
+  explicit HexFloatTypeParams(Float)
+      : min_exponent(std::numeric_limits<Float>::min_exponent - 1),
+        leading_digit_size_bits(HexFloatLeadingDigitSizeInBits<Float>()) {
+    assert(leading_digit_size_bits >= 1 && leading_digit_size_bits <= 4);
+  }
+
+  int min_exponent;
+  int leading_digit_size_bits;
+};
+
+// Hex Float Rounding. First check if we need to round; if so, then we do that
+// by manipulating (incrementing) the mantissa, that way we can later print the
+// mantissa digits by iterating through them in the same way regardless of
+// whether a rounding happened.
+template <typename Int>
+void FormatARound(bool precision_specified, const FormatState &state,
+                  uint8_t *leading, Int *mantissa, int *exp) {
+  constexpr int kTotalNibbles = sizeof(Int) * 8 / 4;
+  // Index of the last nibble that we could display given precision.
+  int final_nibble_displayed =
+      precision_specified ? std::max(0, (kTotalNibbles - state.precision)) : 0;
+  if (HexFloatNeedsRoundUp(*mantissa, final_nibble_displayed)) {
+    // Need to round up.
+    bool overflow = IncrementNibble(final_nibble_displayed, mantissa);
+    *leading += (overflow ? 1 : 0);
+    if (ABSL_PREDICT_FALSE(*leading > 15)) {
+      // We have overflowed the leading digit. This would mean that we would
+      // need two hex digits to the left of the dot, which is not allowed. So
+      // adjust the mantissa and exponent so that the result is always 1.0eXXX.
+      *leading = 1;
+      *mantissa = 0;
+      *exp += 4;
+    }
+  }
+  // Now that we have handled a possible round-up we can go ahead and zero out
+  // all the nibbles of the mantissa that we won't need.
+  if (precision_specified) {
+    *mantissa &= ~MaskUpToNibbleExclusive<Int>(final_nibble_displayed);
+  }
+}
+
+template <typename Int>
+void FormatANormalize(const HexFloatTypeParams float_traits, uint8_t *leading,
+                      Int *mantissa, int *exp) {
+  constexpr int kIntBits = sizeof(Int) * 8;
+  static const Int kHighIntBit = Int{1} << (kIntBits - 1);
+  const int kLeadDigitBitsCount = float_traits.leading_digit_size_bits;
+  // Normalize mantissa so that highest bit set is in MSB position, unless we
+  // get interrupted by the exponent threshold.
+  while (*mantissa && !(*mantissa & kHighIntBit)) {
+    if (ABSL_PREDICT_FALSE(*exp - 1 < float_traits.min_exponent)) {
+      *mantissa >>= (float_traits.min_exponent - *exp);
+      *exp = float_traits.min_exponent;
+      return;
+    }
+    *mantissa <<= 1;
+    --*exp;
+  }
+  // Extract bits for leading digit then shift them away leaving the
+  // fractional part.
+  *leading =
+      static_cast<uint8_t>(*mantissa >> (kIntBits - kLeadDigitBitsCount));
+  *exp -= (*mantissa != 0) ? kLeadDigitBitsCount : *exp;
+  *mantissa <<= kLeadDigitBitsCount;
+}
+
+template <typename Int>
+void FormatA(const HexFloatTypeParams float_traits, Int mantissa, int exp,
+             bool uppercase, const FormatState &state) {
+  // Int properties.
+  constexpr int kIntBits = sizeof(Int) * 8;
+  constexpr int kTotalNibbles = sizeof(Int) * 8 / 4;
+  // Did the user specify a precision explicitly?
+  const bool precision_specified = state.conv.precision() >= 0;
+
+  // ========== Normalize/Denormalize ==========
+  exp += kIntBits;  // make all digits fractional digits.
+  // This holds the (up to four) bits of leading digit, i.e., the '1' in the
+  // number 0x1.e6fp+2. It's always > 0 unless number is zero or denormal.
+  uint8_t leading = 0;
+  FormatANormalize(float_traits, &leading, &mantissa, &exp);
+
+  // =============== Rounding ==================
+  // Check if we need to round; if so, then we do that by manipulating
+  // (incrementing) the mantissa before beginning to print characters.
+  FormatARound(precision_specified, state, &leading, &mantissa, &exp);
+
+  // ============= Format Result ===============
+  // This buffer holds the "0x1.ab1de3" portion of "0x1.ab1de3pe+2". Compute the
+  // size with long double which is the largest of the floats.
+  constexpr size_t kBufSizeForHexFloatRepr =
+      2                                               // 0x
+      + std::numeric_limits<long double>::digits / 4  // number of hex digits
+      + 1                                             // round up
+      + 1;                                            // "." (dot)
+  char digits_buffer[kBufSizeForHexFloatRepr];
+  char *digits_iter = digits_buffer;
+  const char *const digits =
+      static_cast<const char *>("0123456789ABCDEF0123456789abcdef") +
+      (uppercase ? 0 : 16);
+
+  // =============== Hex Prefix ================
+  *digits_iter++ = '0';
+  *digits_iter++ = uppercase ? 'X' : 'x';
+
+  // ========== Non-Fractional Digit ===========
+  *digits_iter++ = digits[leading];
+
+  // ================== Dot ====================
+  // There are three reasons we might need a dot. Keep in mind that, at this
+  // point, the mantissa holds only the fractional part.
+  if ((precision_specified && state.precision > 0) ||
+      (!precision_specified && mantissa > 0) || state.conv.has_alt_flag()) {
+    *digits_iter++ = '.';
+  }
+
+  // ============ Fractional Digits ============
+  int digits_emitted = 0;
+  while (mantissa > 0) {
+    *digits_iter++ = digits[GetNibble(mantissa, kTotalNibbles - 1)];
+    mantissa <<= 4;
+    ++digits_emitted;
+  }
+  int trailing_zeros =
+      precision_specified ? state.precision - digits_emitted : 0;
+  assert(trailing_zeros >= 0);
+  auto digits_result = string_view(digits_buffer, digits_iter - digits_buffer);
+
+  // =============== Exponent ==================
+  constexpr size_t kBufSizeForExpDecRepr =
+      numbers_internal::kFastToBufferSize  // requred for FastIntToBuffer
+      + 1                                  // 'p' or 'P'
+      + 1;                                 // '+' or '-'
+  char exp_buffer[kBufSizeForExpDecRepr];
+  exp_buffer[0] = uppercase ? 'P' : 'p';
+  exp_buffer[1] = exp >= 0 ? '+' : '-';
+  numbers_internal::FastIntToBuffer(exp < 0 ? -exp : exp, exp_buffer + 2);
+
+  // ============ Assemble Result ==============
+  FinalPrint(state,           //
+             digits_result,   // 0xN.NNN...
+             2,               // offset in `data` to start padding if needed.
+             trailing_zeros,  // num remaining mantissa padding zeros
+             exp_buffer);     // exponent
+}
+
 char *CopyStringTo(absl::string_view v, char *out) {
   std::memcpy(out, v.data(), v.size());
   return out + v.size();
@@ -1103,7 +1361,10 @@ bool FloatToSink(const Float v, const FormatConversionSpecImpl &conv,
     }
   } else if (c == FormatConversionCharInternal::a ||
              c == FormatConversionCharInternal::A) {
-    return FallbackToSnprintf(v, conv, sink);
+    bool uppercase = (c == FormatConversionCharInternal::A);
+    FormatA(HexFloatTypeParams(Float{}), decomposed.mantissa,
+            decomposed.exponent, uppercase, {sign_char, precision, conv, sink});
+    return true;
   } else {
     return false;
   }
diff --git a/absl/strings/str_cat.cc b/absl/strings/str_cat.cc
index d9afe2f3..dd5d25b0 100644
--- a/absl/strings/str_cat.cc
+++ b/absl/strings/str_cat.cc
@@ -141,12 +141,12 @@ namespace strings_internal {
 std::string CatPieces(std::initializer_list<absl::string_view> pieces) {
   std::string result;
   size_t total_size = 0;
-  for (const absl::string_view piece : pieces) total_size += piece.size();
+  for (const absl::string_view& piece : pieces) total_size += piece.size();
   strings_internal::STLStringResizeUninitialized(&result, total_size);
 
   char* const begin = &result[0];
   char* out = begin;
-  for (const absl::string_view piece : pieces) {
+  for (const absl::string_view& piece : pieces) {
     const size_t this_size = piece.size();
     if (this_size != 0) {
       memcpy(out, piece.data(), this_size);
@@ -170,7 +170,7 @@ void AppendPieces(std::string* dest,
                   std::initializer_list<absl::string_view> pieces) {
   size_t old_size = dest->size();
   size_t total_size = old_size;
-  for (const absl::string_view piece : pieces) {
+  for (const absl::string_view& piece : pieces) {
     ASSERT_NO_OVERLAP(*dest, piece);
     total_size += piece.size();
   }
@@ -178,7 +178,7 @@ void AppendPieces(std::string* dest,
 
   char* const begin = &(*dest)[0];
   char* out = begin + old_size;
-  for (const absl::string_view piece : pieces) {
+  for (const absl::string_view& piece : pieces) {
     const size_t this_size = piece.size();
     if (this_size != 0) {
       memcpy(out, piece.data(), this_size);
diff --git a/absl/synchronization/mutex.cc b/absl/synchronization/mutex.cc
index 05f5c041..c7968f06 100644
--- a/absl/synchronization/mutex.cc
+++ b/absl/synchronization/mutex.cc
@@ -39,6 +39,7 @@
 #include <thread>  // NOLINT(build/c++11)
 
 #include "absl/base/attributes.h"
+#include "absl/base/call_once.h"
 #include "absl/base/config.h"
 #include "absl/base/dynamic_annotations.h"
 #include "absl/base/internal/atomic_hook.h"
@@ -58,7 +59,6 @@
 
 using absl::base_internal::CurrentThreadIdentityIfPresent;
 using absl::base_internal::PerThreadSynch;
-using absl::base_internal::SchedulingGuard;
 using absl::base_internal::ThreadIdentity;
 using absl::synchronization_internal::GetOrCreateCurrentThreadIdentity;
 using absl::synchronization_internal::GraphCycles;
@@ -86,28 +86,6 @@ ABSL_CONST_INIT std::atomic<OnDeadlockCycle> synch_deadlock_detection(
     kDeadlockDetectionDefault);
 ABSL_CONST_INIT std::atomic<bool> synch_check_invariants(false);
 
-// ------------------------------------------ spinlock support
-
-// Make sure read-only globals used in the Mutex code are contained on the
-// same cacheline and cacheline aligned to eliminate any false sharing with
-// other globals from this and other modules.
-static struct MutexGlobals {
-  MutexGlobals() {
-    // Find machine-specific data needed for Delay() and
-    // TryAcquireWithSpinning(). This runs in the global constructor
-    // sequence, and before that zeros are safe values.
-    num_cpus = absl::base_internal::NumCPUs();
-    spinloop_iterations = num_cpus > 1 ? 1500 : 0;
-  }
-  int num_cpus;
-  int spinloop_iterations;
-  // Pad this struct to a full cacheline to prevent false sharing.
-  char padding[ABSL_CACHELINE_SIZE - 2 * sizeof(int)];
-} ABSL_CACHELINE_ALIGNED mutex_globals;
-static_assert(
-    sizeof(MutexGlobals) == ABSL_CACHELINE_SIZE,
-    "MutexGlobals must occupy an entire cacheline to prevent false sharing");
-
 ABSL_INTERNAL_ATOMIC_HOOK_ATTRIBUTES
     absl::base_internal::AtomicHook<void (*)(int64_t wait_cycles)>
         submit_profile_data;
@@ -144,7 +122,22 @@ void RegisterSymbolizer(bool (*fn)(const void *pc, char *out, int out_size)) {
   symbolizer.Store(fn);
 }
 
-// spinlock delay on iteration c.  Returns new c.
+struct ABSL_CACHELINE_ALIGNED MutexGlobals {
+  absl::once_flag once;
+  int num_cpus = 0;
+  int spinloop_iterations = 0;
+};
+
+static const MutexGlobals& GetMutexGlobals() {
+  ABSL_CONST_INIT static MutexGlobals data;
+  absl::base_internal::LowLevelCallOnce(&data.once, [&]() {
+    data.num_cpus = absl::base_internal::NumCPUs();
+    data.spinloop_iterations = data.num_cpus > 1 ? 1500 : 0;
+  });
+  return data;
+}
+
+// Spinlock delay on iteration c.  Returns new c.
 namespace {
   enum DelayMode { AGGRESSIVE, GENTLE };
 };
@@ -154,22 +147,25 @@ static int Delay(int32_t c, DelayMode mode) {
   // gentle then spin only a few times before yielding.  Aggressive spinning is
   // used to ensure that an Unlock() call, which  must get the spin lock for
   // any thread to make progress gets it without undue delay.
-  int32_t limit = (mutex_globals.num_cpus > 1) ?
-      ((mode == AGGRESSIVE) ? 5000 : 250) : 0;
+  const int32_t limit =
+      GetMutexGlobals().num_cpus > 1 ? (mode == AGGRESSIVE ? 5000 : 250) : 0;
   if (c < limit) {
-    c++;               // spin
+    // Spin.
+    c++;
   } else {
     ABSL_TSAN_MUTEX_PRE_DIVERT(nullptr, 0);
-    if (c == limit) {  // yield once
+    if (c == limit) {
+      // Yield once.
       AbslInternalMutexYield();
       c++;
-    } else {           // then wait
+    } else {
+      // Then wait.
       absl::SleepFor(absl::Microseconds(10));
       c = 0;
     }
     ABSL_TSAN_MUTEX_POST_DIVERT(nullptr, 0);
   }
-  return (c);
+  return c;
 }
 
 // --------------------------Generic atomic ops
@@ -1055,7 +1051,6 @@ static PerThreadSynch *DequeueAllWakeable(PerThreadSynch *head,
 // Try to remove thread s from the list of waiters on this mutex.
 // Does nothing if s is not on the waiter list.
 void Mutex::TryRemove(PerThreadSynch *s) {
-  SchedulingGuard::ScopedDisable disable_rescheduling;
   intptr_t v = mu_.load(std::memory_order_relaxed);
   // acquire spinlock & lock
   if ((v & (kMuWait | kMuSpin | kMuWriter | kMuReader)) == kMuWait &&
@@ -1439,7 +1434,7 @@ void Mutex::AssertNotHeld() const {
 // Attempt to acquire *mu, and return whether successful.  The implementation
 // may spin for a short while if the lock cannot be acquired immediately.
 static bool TryAcquireWithSpinning(std::atomic<intptr_t>* mu) {
-  int c = mutex_globals.spinloop_iterations;
+  int c = GetMutexGlobals().spinloop_iterations;
   do {  // do/while somewhat faster on AMD
     intptr_t v = mu->load(std::memory_order_relaxed);
     if ((v & (kMuReader|kMuEvent)) != 0) {
@@ -1899,7 +1894,6 @@ static void CheckForMutexCorruption(intptr_t v, const char* label) {
 }
 
 void Mutex::LockSlowLoop(SynchWaitParams *waitp, int flags) {
-  SchedulingGuard::ScopedDisable disable_rescheduling;
   int c = 0;
   intptr_t v = mu_.load(std::memory_order_relaxed);
   if ((v & kMuEvent) != 0) {
@@ -2019,7 +2013,6 @@ void Mutex::LockSlowLoop(SynchWaitParams *waitp, int flags) {
 // or it is in the process of blocking on a condition variable; it must requeue
 // itself on the mutex/condvar to wait for its condition to become true.
 ABSL_ATTRIBUTE_NOINLINE void Mutex::UnlockSlow(SynchWaitParams *waitp) {
-  SchedulingGuard::ScopedDisable disable_rescheduling;
   intptr_t v = mu_.load(std::memory_order_relaxed);
   this->AssertReaderHeld();
   CheckForMutexCorruption(v, "Unlock");
@@ -2335,7 +2328,6 @@ void Mutex::Trans(MuHow how) {
 // It will later acquire the mutex with high probability.  Otherwise, we
 // enqueue thread w on this mutex.
 void Mutex::Fer(PerThreadSynch *w) {
-  SchedulingGuard::ScopedDisable disable_rescheduling;
   int c = 0;
   ABSL_RAW_CHECK(w->waitp->cond == nullptr,
                  "Mutex::Fer while waiting on Condition");
@@ -2434,7 +2426,6 @@ CondVar::~CondVar() {
 
 // Remove thread s from the list of waiters on this condition variable.
 void CondVar::Remove(PerThreadSynch *s) {
-  SchedulingGuard::ScopedDisable disable_rescheduling;
   intptr_t v;
   int c = 0;
   for (v = cv_.load(std::memory_order_relaxed);;
@@ -2595,7 +2586,6 @@ void CondVar::Wakeup(PerThreadSynch *w) {
 }
 
 void CondVar::Signal() {
-  SchedulingGuard::ScopedDisable disable_rescheduling;
   ABSL_TSAN_MUTEX_PRE_SIGNAL(nullptr, 0);
   intptr_t v;
   int c = 0;