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

1 files changed, 10 insertions, 18 deletions

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 " : "")           //