Export of internal Abseil changes

--
e2de21d54c02b6419c57c0f4e2a16b608deca260 by Evan Brown <ezb@google.com>:

Remove the InsertEnd benchmark.

This benchmark has significantly different possible behaviors that can result in misleading metrics. Specifically, we can have a case where we are deallocating the last node in the b-tree in the erase and then allocating a new node in the insert call repeatedly, whereas normally, we end up just inserting/erasing a value from the last node. Also, the name of the benchmark is misleading because it involves an erase and an insert, but the name only mentions the insert.

PiperOrigin-RevId: 360930639

--
51f6bb97b9cbdb809c31b77e93ce080ca3cba9ea by Benjamin Barenblat <bbaren@google.com>:

Stop testing with double-double random variables

On POWER, long double is often represented as a pair of doubles added
together (double-double arithmetic). We’ve already special-cased
double-double arithmetic in a number of tests, but compiler
bugs [1, 2, 3] have now triggered both false positives and false
negatives, which suggests testing with double doubles is unlikely to
yield useful signal. Remove the special casing and detect if we’re on a
double-double system; if so, just don’t test long doubles.

[1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99048
[2] https://bugs.llvm.org/show_bug.cgi?id=49131
[3] https://bugs.llvm.org/show_bug.cgi?id=49132

PiperOrigin-RevId: 360793161

--
07fb4d7932c2f5d711c480f759dacb0be60f975e by Abseil Team <absl-team@google.com>:

internal change

PiperOrigin-RevId: 360712825
GitOrigin-RevId: e2de21d54c02b6419c57c0f4e2a16b608deca260
Change-Id: I98389b5a8789dcc8f35abc00c767e909181665f0

14 files changed, 138 insertions, 113 deletions

diff --git a/CMake/AbseilDll.cmake b/CMake/AbseilDll.cmake
index 47707dfd..39f85f2f 100644
--- a/CMake/AbseilDll.cmake
+++ b/CMake/AbseilDll.cmake
@@ -131,6 +131,7 @@ set(ABSL_INTERNAL_DLL_FILES
   "numeric/int128.cc"
   "numeric/int128.h"
   "numeric/internal/bits.h"
+  "numeric/internal/representation.h"
   "random/bernoulli_distribution.h"
   "random/beta_distribution.h"
   "random/bit_gen_ref.h"
diff --git a/absl/container/btree_benchmark.cc b/absl/container/btree_benchmark.cc
index 46798676..41f13f52 100644
--- a/absl/container/btree_benchmark.cc
+++ b/absl/container/btree_benchmark.cc
@@ -101,39 +101,6 @@ void BM_InsertSorted(benchmark::State& state) {
   BM_InsertImpl<T>(state, true);
 }
 
-// container::insert sometimes returns a pair<iterator, bool> and sometimes
-// returns an iterator (for multi- containers).
-template <typename Iter>
-Iter GetIterFromInsert(const std::pair<Iter, bool>& pair) {
-  return pair.first;
-}
-template <typename Iter>
-Iter GetIterFromInsert(const Iter iter) {
-  return iter;
-}
-
-// Benchmark insertion of values into a container at the end.
-template <typename T>
-void BM_InsertEnd(benchmark::State& state) {
-  using V = typename remove_pair_const<typename T::value_type>::type;
-  typename KeyOfValue<typename T::key_type, V>::type key_of_value;
-
-  T container;
-  const int kSize = 10000;
-  for (int i = 0; i < kSize; ++i) {
-    container.insert(Generator<V>(kSize)(i));
-  }
-  V v = Generator<V>(kSize)(kSize - 1);
-  typename T::key_type k = key_of_value(v);
-
-  auto it = container.find(k);
-  while (state.KeepRunning()) {
-    // Repeatedly removing then adding v.
-    container.erase(it);
-    it = GetIterFromInsert(container.insert(v));
-  }
-}
-
 // Benchmark inserting the first few elements in a container. In b-tree, this is
 // when the root node grows.
 template <typename T>
@@ -513,7 +480,6 @@ BTREE_TYPES(Time);
 #define MY_BENCHMARK3(type)               \
   MY_BENCHMARK4(type, Insert);            \
   MY_BENCHMARK4(type, InsertSorted);      \
-  MY_BENCHMARK4(type, InsertEnd);         \
   MY_BENCHMARK4(type, InsertSmall);       \
   MY_BENCHMARK4(type, Lookup);            \
   MY_BENCHMARK4(type, FullLookup);        \
diff --git a/absl/numeric/BUILD.bazel b/absl/numeric/BUILD.bazel
index 5d7b1857..ea587bfa 100644
--- a/absl/numeric/BUILD.bazel
+++ b/absl/numeric/BUILD.bazel
@@ -101,3 +101,15 @@ cc_test(
         "@com_github_google_benchmark//:benchmark_main",
     ],
 )
+
+cc_library(
+    name = "representation",
+    hdrs = [
+        "internal/representation.h",
+    ],
+    copts = ABSL_DEFAULT_COPTS,
+    linkopts = ABSL_DEFAULT_LINKOPTS,
+    deps = [
+        "//absl/base:config",
+    ],
+)
diff --git a/absl/numeric/CMakeLists.txt b/absl/numeric/CMakeLists.txt
index be94352a..781987dc 100644
--- a/absl/numeric/CMakeLists.txt
+++ b/absl/numeric/CMakeLists.txt
@@ -86,3 +86,15 @@ absl_cc_library(
     absl::int128
   PUBLIC
 )
+
+absl_cc_library(
+  NAME
+    numeric_representation
+  HDRS
+    "internal/representation.h"
+  COPTS
+    ${ABSL_DEFAULT_COPTS}
+  DEPS
+    absl::config
+  PUBLIC
+)
diff --git a/absl/numeric/internal/representation.h b/absl/numeric/internal/representation.h
new file mode 100644
index 00000000..82d332fd
--- /dev/null
+++ b/absl/numeric/internal/representation.h
@@ -0,0 +1,55 @@
+// Copyright 2021 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.
+
+#ifndef ABSL_NUMERIC_INTERNAL_REPRESENTATION_H_
+#define ABSL_NUMERIC_INTERNAL_REPRESENTATION_H_
+
+#include <limits>
+
+#include "absl/base/config.h"
+
+namespace absl {
+ABSL_NAMESPACE_BEGIN
+namespace numeric_internal {
+
+// Returns true iff long double is represented as a pair of doubles added
+// together.
+inline constexpr bool IsDoubleDouble() {
+  // A double-double value always has exactly twice the precision of a double
+  // value--one double carries the high digits and one double carries the low
+  // digits. This property is not shared with any other common floating-point
+  // representation, so this test won't trigger false positives. For reference,
+  // this table gives the number of bits of precision of each common
+  // floating-point representation:
+  //
+  //                type     precision
+  //         IEEE single          24 b
+  //         IEEE double          53
+  //     x86 long double          64
+  //       double-double         106
+  //      IEEE quadruple         113
+  //
+  // Note in particular that a quadruple-precision float has greater precision
+  // than a double-double float despite taking up the same amount of memory; the
+  // quad has more of its bits allocated to the mantissa than the double-double
+  // has.
+  return std::numeric_limits<long double>::digits ==
+         2 * std::numeric_limits<double>::digits;
+}
+
+}  // namespace numeric_internal
+ABSL_NAMESPACE_END
+}  // namespace absl
+
+#endif  // ABSL_NUMERIC_INTERNAL_REPRESENTATION_H_
diff --git a/absl/random/BUILD.bazel b/absl/random/BUILD.bazel
index d97b2c4e..66ffcbc7 100644
--- a/absl/random/BUILD.bazel
+++ b/absl/random/BUILD.bazel
@@ -188,6 +188,7 @@ cc_test(
         ":distributions",
         ":random",
         "//absl/base:raw_logging_internal",
+        "//absl/numeric:representation",
         "//absl/random/internal:distribution_test_util",
         "//absl/random/internal:pcg_engine",
         "//absl/random/internal:sequence_urbg",
@@ -308,6 +309,7 @@ cc_test(
         ":random",
         "//absl/base:core_headers",
         "//absl/base:raw_logging_internal",
+        "//absl/numeric:representation",
         "//absl/random/internal:distribution_test_util",
         "//absl/random/internal:pcg_engine",
         "//absl/random/internal:sequence_urbg",
@@ -331,6 +333,7 @@ cc_test(
         ":random",
         "//absl/base:core_headers",
         "//absl/base:raw_logging_internal",
+        "//absl/numeric:representation",
         "//absl/random/internal:distribution_test_util",
         "//absl/random/internal:sequence_urbg",
         "//absl/strings",
@@ -377,6 +380,7 @@ cc_test(
         ":distributions",
         ":random",
         "//absl/base:raw_logging_internal",
+        "//absl/numeric:representation",
         "//absl/random/internal:distribution_test_util",
         "//absl/random/internal:pcg_engine",
         "//absl/random/internal:sequence_urbg",
diff --git a/absl/random/CMakeLists.txt b/absl/random/CMakeLists.txt
index 13093d6d..3009a034 100644
--- a/absl/random/CMakeLists.txt
+++ b/absl/random/CMakeLists.txt
@@ -259,6 +259,7 @@ absl_cc_test(
   LINKOPTS
     ${ABSL_DEFAULT_LINKOPTS}
   DEPS
+    absl::numeric_representation
     absl::random_distributions
     absl::random_random
     absl::random_internal_distribution_test_util
@@ -381,6 +382,7 @@ absl_cc_test(
     ${ABSL_DEFAULT_LINKOPTS}
   DEPS
     absl::core_headers
+    absl::numeric_representation
     absl::random_distributions
     absl::random_internal_distribution_test_util
     absl::random_internal_pcg_engine
@@ -404,6 +406,7 @@ absl_cc_test(
     ${ABSL_DEFAULT_LINKOPTS}
   DEPS
     absl::core_headers
+    absl::numeric_representation
     absl::random_distributions
     absl::random_internal_distribution_test_util
     absl::random_internal_sequence_urbg
@@ -446,6 +449,7 @@ absl_cc_test(
   LINKOPTS
     ${ABSL_DEFAULT_LINKOPTS}
   DEPS
+    absl::numeric_representation
     absl::random_distributions
     absl::random_internal_distribution_test_util
     absl::random_internal_pcg_engine
diff --git a/absl/random/beta_distribution_test.cc b/absl/random/beta_distribution_test.cc
index 277e4dc6..44cdfdd0 100644
--- a/absl/random/beta_distribution_test.cc
+++ b/absl/random/beta_distribution_test.cc
@@ -21,12 +21,14 @@
 #include <random>
 #include <sstream>
 #include <string>
+#include <type_traits>
 #include <unordered_map>
 #include <vector>
 
 #include "gmock/gmock.h"
 #include "gtest/gtest.h"
 #include "absl/base/internal/raw_logging.h"
+#include "absl/numeric/internal/representation.h"
 #include "absl/random/internal/chi_square.h"
 #include "absl/random/internal/distribution_test_util.h"
 #include "absl/random/internal/pcg_engine.h"
@@ -42,7 +44,15 @@ namespace {
 template <typename IntType>
 class BetaDistributionInterfaceTest : public ::testing::Test {};
 
-using RealTypes = ::testing::Types<float, double, long double>;
+// double-double arithmetic is not supported well by either GCC or Clang; see
+// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99048,
+// https://bugs.llvm.org/show_bug.cgi?id=49131, and
+// https://bugs.llvm.org/show_bug.cgi?id=49132. Don't bother running these tests
+// with double doubles until compiler support is better.
+using RealTypes =
+    std::conditional<absl::numeric_internal::IsDoubleDouble(),
+                     ::testing::Types<float, double>,
+                     ::testing::Types<float, double, long double>>::type;
 TYPED_TEST_CASE(BetaDistributionInterfaceTest, RealTypes);
 
 TYPED_TEST(BetaDistributionInterfaceTest, SerializeTest) {
@@ -53,9 +63,6 @@ TYPED_TEST(BetaDistributionInterfaceTest, SerializeTest) {
   const TypeParam kLargeA =
       std::exp(std::log((std::numeric_limits<TypeParam>::max)()) -
                std::log(std::log((std::numeric_limits<TypeParam>::max)())));
-  const TypeParam kLargeAPPC = std::exp(
-      std::log((std::numeric_limits<TypeParam>::max)()) -
-      std::log(std::log((std::numeric_limits<TypeParam>::max)())) - 10.0f);
   using param_type = typename absl::beta_distribution<TypeParam>::param_type;
 
   constexpr int kCount = 1000;
@@ -76,9 +83,6 @@ TYPED_TEST(BetaDistributionInterfaceTest, SerializeTest) {
       kLargeA,                                //
       std::nextafter(kLargeA, TypeParam(0)),  //
       std::nextafter(kLargeA, std::numeric_limits<TypeParam>::max()),
-      kLargeAPPC,  //
-      std::nextafter(kLargeAPPC, TypeParam(0)),
-      std::nextafter(kLargeAPPC, std::numeric_limits<TypeParam>::max()),
       // Boundary cases.
       std::numeric_limits<TypeParam>::max(),
       std::numeric_limits<TypeParam>::epsilon(),
@@ -125,28 +129,6 @@ TYPED_TEST(BetaDistributionInterfaceTest, SerializeTest) {
 
       ss >> after;
 
-#if defined(__powerpc64__) || defined(__PPC64__) || defined(__powerpc__) || \
-    defined(__ppc__) || defined(__PPC__)
-      if (std::is_same<TypeParam, long double>::value) {
-        // Roundtripping floating point values requires sufficient precision
-        // to reconstruct the exact value. It turns out that long double
-        // has some errors doing this on ppc.
-        if (alpha <= std::numeric_limits<double>::max() &&
-            alpha >= std::numeric_limits<double>::lowest()) {
-          EXPECT_EQ(static_cast<double>(before.alpha()),
-                    static_cast<double>(after.alpha()))
-              << ss.str();
-        }
-        if (beta <= std::numeric_limits<double>::max() &&
-            beta >= std::numeric_limits<double>::lowest()) {
-          EXPECT_EQ(static_cast<double>(before.beta()),
-                    static_cast<double>(after.beta()))
-              << ss.str();
-        }
-        continue;
-      }
-#endif
-
       EXPECT_EQ(before.alpha(), after.alpha());
       EXPECT_EQ(before.beta(), after.beta());
       EXPECT_EQ(before, after)           //
diff --git a/absl/random/exponential_distribution_test.cc b/absl/random/exponential_distribution_test.cc
index 5a8afde5..af11d61c 100644
--- a/absl/random/exponential_distribution_test.cc
+++ b/absl/random/exponential_distribution_test.cc
@@ -30,6 +30,7 @@
 #include "gtest/gtest.h"
 #include "absl/base/internal/raw_logging.h"
 #include "absl/base/macros.h"
+#include "absl/numeric/internal/representation.h"
 #include "absl/random/internal/chi_square.h"
 #include "absl/random/internal/distribution_test_util.h"
 #include "absl/random/internal/pcg_engine.h"
@@ -47,7 +48,15 @@ using absl::random_internal::kChiSquared;
 template <typename RealType>
 class ExponentialDistributionTypedTest : public ::testing::Test {};
 
-using RealTypes = ::testing::Types<float, double, long double>;
+// double-double arithmetic is not supported well by either GCC or Clang; see
+// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99048,
+// https://bugs.llvm.org/show_bug.cgi?id=49131, and
+// https://bugs.llvm.org/show_bug.cgi?id=49132. Don't bother running these tests
+// with double doubles until compiler support is better.
+using RealTypes =
+    std::conditional<absl::numeric_internal::IsDoubleDouble(),
+                     ::testing::Types<float, double>,
+                     ::testing::Types<float, double, long double>>::type;
 TYPED_TEST_CASE(ExponentialDistributionTypedTest, RealTypes);
 
 TYPED_TEST(ExponentialDistributionTypedTest, SerializeTest) {
@@ -126,23 +135,6 @@ TYPED_TEST(ExponentialDistributionTypedTest, SerializeTest) {
 
     ss >> after;
 
-#if defined(__powerpc64__) || defined(__PPC64__) || defined(__powerpc__) || \
-    defined(__ppc__) || defined(__PPC__)
-    if (std::is_same<TypeParam, long double>::value) {
-      // Roundtripping floating point values requires sufficient precision to
-      // reconstruct the exact value. It turns out that long double has some
-      // errors doing this on ppc, particularly for values
-      // near {1.0 +/- epsilon}.
-      if (lambda <= std::numeric_limits<double>::max() &&
-          lambda >= std::numeric_limits<double>::lowest()) {
-        EXPECT_EQ(static_cast<double>(before.lambda()),
-                  static_cast<double>(after.lambda()))
-            << ss.str();
-      }
-      continue;
-    }
-#endif
-
     EXPECT_EQ(before.lambda(), after.lambda())  //
         << ss.str() << " "                      //
         << (ss.good() ? "good " : "")           //
diff --git a/absl/random/gaussian_distribution_test.cc b/absl/random/gaussian_distribution_test.cc
index 2aa7caf4..c0bac2b0 100644
--- a/absl/random/gaussian_distribution_test.cc
+++ b/absl/random/gaussian_distribution_test.cc
@@ -21,12 +21,14 @@
 #include <iterator>
 #include <random>
 #include <string>
+#include <type_traits>
 #include <vector>
 
 #include "gmock/gmock.h"
 #include "gtest/gtest.h"
 #include "absl/base/internal/raw_logging.h"
 #include "absl/base/macros.h"
+#include "absl/numeric/internal/representation.h"
 #include "absl/random/internal/chi_square.h"
 #include "absl/random/internal/distribution_test_util.h"
 #include "absl/random/internal/sequence_urbg.h"
@@ -43,7 +45,15 @@ using absl::random_internal::kChiSquared;
 template <typename RealType>
 class GaussianDistributionInterfaceTest : public ::testing::Test {};
 
-using RealTypes = ::testing::Types<float, double, long double>;
+// double-double arithmetic is not supported well by either GCC or Clang; see
+// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99048,
+// https://bugs.llvm.org/show_bug.cgi?id=49131, and
+// https://bugs.llvm.org/show_bug.cgi?id=49132. Don't bother running these tests
+// with double doubles until compiler support is better.
+using RealTypes =
+    std::conditional<absl::numeric_internal::IsDoubleDouble(),
+                     ::testing::Types<float, double>,
+                     ::testing::Types<float, double, long double>>::type;
 TYPED_TEST_CASE(GaussianDistributionInterfaceTest, RealTypes);
 
 TYPED_TEST(GaussianDistributionInterfaceTest, SerializeTest) {
@@ -129,29 +139,6 @@ TYPED_TEST(GaussianDistributionInterfaceTest, SerializeTest) {
 
         ss >> after;
 
-#if defined(__powerpc64__) || defined(__PPC64__) || defined(__powerpc__) || \
-    defined(__ppc__) || defined(__PPC__)
-        if (std::is_same<TypeParam, long double>::value) {
-          // Roundtripping floating point values requires sufficient precision
-          // to reconstruct the exact value.  It turns out that long double
-          // has some errors doing this on ppc, particularly for values
-          // near {1.0 +/- epsilon}.
-          if (mean <= std::numeric_limits<double>::max() &&
-              mean >= std::numeric_limits<double>::lowest()) {
-            EXPECT_EQ(static_cast<double>(before.mean()),
-                      static_cast<double>(after.mean()))
-                << ss.str();
-          }
-          if (stddev <= std::numeric_limits<double>::max() &&
-              stddev >= std::numeric_limits<double>::lowest()) {
-            EXPECT_EQ(static_cast<double>(before.stddev()),
-                      static_cast<double>(after.stddev()))
-                << ss.str();
-          }
-          continue;
-        }
-#endif
-
         EXPECT_EQ(before.mean(), after.mean());
         EXPECT_EQ(before.stddev(), after.stddev())  //
             << ss.str() << " "                      //
diff --git a/absl/random/uniform_real_distribution_test.cc b/absl/random/uniform_real_distribution_test.cc
index 8cf874d6..18bcd3bc 100644
--- a/absl/random/uniform_real_distribution_test.cc
+++ b/absl/random/uniform_real_distribution_test.cc
@@ -20,11 +20,13 @@
 #include <random>
 #include <sstream>
 #include <string>
+#include <type_traits>
 #include <vector>
 
 #include "gmock/gmock.h"
 #include "gtest/gtest.h"
 #include "absl/base/internal/raw_logging.h"
+#include "absl/numeric/internal/representation.h"
 #include "absl/random/internal/chi_square.h"
 #include "absl/random/internal/distribution_test_util.h"
 #include "absl/random/internal/pcg_engine.h"
@@ -55,7 +57,15 @@ namespace {
 template <typename RealType>
 class UniformRealDistributionTest : public ::testing::Test {};
 
-using RealTypes = ::testing::Types<float, double, long double>;
+// double-double arithmetic is not supported well by either GCC or Clang; see
+// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99048,
+// https://bugs.llvm.org/show_bug.cgi?id=49131, and
+// https://bugs.llvm.org/show_bug.cgi?id=49132. Don't bother running these tests
+// with double doubles until compiler support is better.
+using RealTypes =
+    std::conditional<absl::numeric_internal::IsDoubleDouble(),
+                     ::testing::Types<float, double>,
+                     ::testing::Types<float, double, long double>>::type;
 
 TYPED_TEST_SUITE(UniformRealDistributionTest, RealTypes);
 
diff --git a/absl/strings/BUILD.bazel b/absl/strings/BUILD.bazel
index 5efaf896..123b5efb 100644
--- a/absl/strings/BUILD.bazel
+++ b/absl/strings/BUILD.bazel
@@ -709,6 +709,7 @@ cc_library(
         "//absl/meta:type_traits",
         "//absl/numeric:bits",
         "//absl/numeric:int128",
+        "//absl/numeric:representation",
         "//absl/types:optional",
         "//absl/types:span",
     ],
diff --git a/absl/strings/CMakeLists.txt b/absl/strings/CMakeLists.txt
index 12f322a9..3b7ae639 100644
--- a/absl/strings/CMakeLists.txt
+++ b/absl/strings/CMakeLists.txt
@@ -410,6 +410,7 @@ absl_cc_library(
     absl::strings
     absl::config
     absl::core_headers
+    absl::numeric_representation
     absl::type_traits
     absl::int128
     absl::span
diff --git a/absl/strings/internal/str_format/float_conversion.cc b/absl/strings/internal/str_format/float_conversion.cc
index 2b1fd8cb..b1c40684 100644
--- a/absl/strings/internal/str_format/float_conversion.cc
+++ b/absl/strings/internal/str_format/float_conversion.cc
@@ -29,6 +29,7 @@
 #include "absl/meta/type_traits.h"
 #include "absl/numeric/bits.h"
 #include "absl/numeric/int128.h"
+#include "absl/numeric/internal/representation.h"
 #include "absl/strings/numbers.h"
 #include "absl/types/optional.h"
 #include "absl/types/span.h"
@@ -39,6 +40,8 @@ namespace str_format_internal {
 
 namespace {
 
+using ::absl::numeric_internal::IsDoubleDouble;
+
 // The code below wants to avoid heap allocations.
 // To do so it needs to allocate memory on the stack.
 // `StackArray` will allocate memory on the stack in the form of a uint32_t
@@ -112,13 +115,6 @@ inline uint64_t DivideBy10WithCarry(uint64_t *v, uint64_t carry) {
   return next_carry % divisor;
 }
 
-constexpr bool IsDoubleDouble() {
-  // This is the `double-double` representation of `long double`.
-  // We do not handle it natively. Fallback to snprintf.
-  return std::numeric_limits<long double>::digits ==
-         2 * std::numeric_limits<double>::digits;
-}
-
 using MaxFloatType =
     typename std::conditional<IsDoubleDouble(), double, long double>::type;
 
@@ -1404,6 +1400,8 @@ bool FloatToSink(const Float v, const FormatConversionSpecImpl &conv,
 bool ConvertFloatImpl(long double v, const FormatConversionSpecImpl &conv,
                       FormatSinkImpl *sink) {
   if (IsDoubleDouble()) {
+    // This is the `double-double` representation of `long double`. We do not
+    // handle it natively. Fallback to snprintf.
     return FallbackToSnprintf(v, conv, sink);
   }