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

3 files changed, 79 insertions, 0 deletions

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_