Export of internal Abseil changes.

--
3d20ce6cd6541579abecaba169d4b8716d511272 by Jon Cohen <cohenjon@google.com>:

Only use LSAN for clang version >= 3.5.  This should fix https://github.com/abseil/abseil-cpp/issues/244

PiperOrigin-RevId: 234675129

--
e15bd4ec7a81aa95cc3d09fa1e0e81d58ae478fb by Conrad Parker <conradparker@google.com>:

Fix errors in apply() sample code

The following changes are made:
 * Make the example method public.
 * Give the two user functions different names to avoid confusion about
   whether apply() can select the correct overload of a function based
   on its tuple argument (it can't).
 * Pass tuple2 to the second example apply() invocation, instead of
   passing its contents individually.
 * Fix a s/tuple/tuple3/ typo in the third example apply() invocation.

PiperOrigin-RevId: 234223407

--
de0ed71e21bc76ddf9fe715fdbaef74cd0df95c7 by Abseil Team <absl-team@google.com>:

First test if a macro is defined to avoid -Wundef.

ABSL clients may need to compile their code with the -Wundef warning
flag. It will be helpful if ABSL header files can be compiled without
the -Wundef warning.

How to avoid the -Wundef warning: If a macro may be undefined, we need
to first test whether the macro is defined before testing its value. We
can't rely on the C preprocessor rule that an undefined macro has the
value 0L.

PiperOrigin-RevId: 234201123

--
fa484ad7dae0cac21140a96662809ecb0ec8eb5d by Abseil Team <absl-team@google.com>:

Internal change.

PiperOrigin-RevId: 234185697

--
d69b1baef681e27954b065375ecf9c2320463b2b by Samuel Benzaquen <sbenza@google.com>:

Mix pointers more thoroughly.
Some pointer alignments interact badly with the mixing constant. By mixing twice we reduce this problem.

PiperOrigin-RevId: 234178401

--
1041d0e474610f3a8fea0db90958857327d6da1c by Samuel Benzaquen <sbenza@google.com>:

Record rehashes in the hashtablez struct.
Only recording the probe length on insertion causes a huge overestimation of
the total probe length at any given time.

With natural growth, elements are inserted when the load factor is between
(max load/2, max load). However, after a rehash the majority of elements are
actually inserted when the load factor is less than max/2 and have a much lower
average probe length.

Also reset some values when the table is cleared.

PiperOrigin-RevId: 234013580

--
299205caf3c89c47339f7409bc831746602cea84 by Mark Barolak <mbar@google.com>:

Fix a sample code snippet that assumes `absl::string_view::const_iterator` is `const char*`.  This is generally true, however in C++17 builds, absl::string_view is an alias for std::string_view and on MSVC, the std::string_view::const_iterator is an object instead of just a pointer.

PiperOrigin-RevId: 233844595

--
af6c6370cf51a1e6c1469c79dfb2a486a4009136 by Abseil Team <absl-team@google.com>:

Internal change.

PiperOrigin-RevId: 233773470

--
6e59e4b8e2bb6101b448f0f32b0bea81fe399ccf by Abseil Team <absl-team@google.com>:

fix typo in {Starts|Ends}WithIgnoreCase comment in match.h

PiperOrigin-RevId: 233662951
GitOrigin-RevId: 3d20ce6cd6541579abecaba169d4b8716d511272
Change-Id: Ib9a29b1c38c6aedf5d9f3f7f00596e8d30e864dd

2 files changed, 36 insertions, 1 deletions

diff --git a/absl/hash/hash_test.cc b/absl/hash/hash_test.cc
index a5af93a..2c9e46b 100644
--- a/absl/hash/hash_test.cc
+++ b/absl/hash/hash_test.cc
@@ -135,6 +135,36 @@ TEST(HashValueTest, Pointer) {
       std::make_tuple(&i, ptr, nullptr, ptr + 1, n)));
 }
 
+TEST(HashValueTest, PointerAlignment) {
+  // We want to make sure that pointer alignment will not cause bits to be
+  // stuck.
+
+  constexpr size_t kTotalSize = 1 << 20;
+  std::unique_ptr<char[]> data(new char[kTotalSize]);
+  constexpr size_t kLog2NumValues = 5;
+  constexpr size_t kNumValues = 1 << kLog2NumValues;
+
+  for (size_t align = 1; align < kTotalSize / kNumValues;
+       align < 8 ? align += 1 : align < 1024 ? align += 8 : align += 32) {
+    SCOPED_TRACE(align);
+    ASSERT_LE(align * kNumValues, kTotalSize);
+
+    size_t bits_or = 0;
+    size_t bits_and = ~size_t{};
+
+    for (size_t i = 0; i < kNumValues; ++i) {
+      size_t hash = absl::Hash<void*>()(data.get() + i * align);
+      bits_or |= hash;
+      bits_and &= hash;
+    }
+
+    // Limit the scope to the bits we would be using for Swisstable.
+    constexpr size_t kMask = (1 << (kLog2NumValues + 7)) - 1;
+    size_t stuck_bits = (~bits_or | bits_and) & kMask;
+    EXPECT_EQ(stuck_bits, 0) << "0x" << std::hex << stuck_bits;
+  }
+}
+
 // TODO(EricWF): MSVC 15 has a bug that causes it to incorrectly evaluate the
 // SFINAE in internal/hash.h, causing this test to fail.
 #if !defined(_MSC_VER)
diff --git a/absl/hash/internal/hash.h b/absl/hash/internal/hash.h
index ba6d746..bd07677 100644
--- a/absl/hash/internal/hash.h
+++ b/absl/hash/internal/hash.h
@@ -264,7 +264,12 @@ H AbslHashValue(H hash_state, long double value) {
 // AbslHashValue() for hashing pointers
 template <typename H, typename T>
 H AbslHashValue(H hash_state, T* ptr) {
-  return hash_internal::hash_bytes(std::move(hash_state), ptr);
+  auto v = reinterpret_cast<uintptr_t>(ptr);
+  // Due to alignment, pointers tend to have low bits as zero, and the next few
+  // bits follow a pattern since they are also multiples of some base value.
+  // Mixing the pointer twice helps prevent stuck low bits for certain alignment
+  // values.
+  return H::combine(std::move(hash_state), v, v);
 }
 
 // AbslHashValue() for hashing nullptr_t