Export of internal Abseil changes

--
1620e8ffaa93ef24510ca60c7fff2a07248ac9f6 by Abseil Team <absl-team@google.com>:

Update comment.

PiperOrigin-RevId: 382858259

--
20db116f28469149d10e0f7f8b976cb903dd4879 by Gennadiy Rozental <rogeeff@google.com>:

Add benchmark running on multiple flags.

Update size_tester to include cost of absl::GetFlag call.
Add size_tester invocation for bool flag.

New benchmark better represent GetFlag usage.

PiperOrigin-RevId: 382820341

--
2e097ad3811c4e329f75b98877a5e74c1d3d84fd by Abseil Team <absl-team@google.com>:

Avoid 64x64->128 multiplication in absl::Hash's mix on AArch64

On AArch64, calculating a 128-bit product is inefficient, because it requires a sequence of two instructions to calculate the upper and lower halves of the result. So calculate a 64-bit product instead.

Making MultType 64-bits means the upper 32 bits of the result do not participate in shift/xor, but the add/multiply gives us sufficient mixing.

PiperOrigin-RevId: 382625931

--
f3ae3f32cb53168c8dc91b766f2932dc87cec503 by Abseil Team <absl-team@google.com>:

Remove homegrown Round implementation

absl/time/duration.cc defined a Round implementation to accommodate old
versions of MSVC that lacked std::round(long double). Abseil no longer
supports those MSVCs, so we don’t need the homegrown implementation
anymore. Remove it, and replace calls to it with std::rint.

PiperOrigin-RevId: 382605191

--
a13631c91bf5478289e1a512ce215c85501a26f7 by Martijn Vels <mvels@google.com>:

Move the Consume() conversion functions out of cord_rep_ring into cord_rep_consume.

This makes these functions generic, so we can repurpose these for the new Btree conversion functions.

PiperOrigin-RevId: 382594902

--
7394c737500c2d8371fcf913b21ad1b321ba499d by Benjamin Barenblat <bbaren@google.com>:

Remove homegrown Round implementation

absl/time/duration.cc defined a Round implementation to accommodate old
versions of MSVC that lacked std::round(long double). Abseil no longer
supports those MSVCs, so we don’t need the homegrown implementation
anymore. Remove it, and replace calls to it with std::rint.

PiperOrigin-RevId: 382569900

--
d72a761f43dc5c9b9510c3a1363177ed26646b5d by Abseil Team <absl-team@google.com>:

Prefer `getentropy` for Emscripten.

It needs a different header, so I've separated it out from the GLIBC
check above.

PiperOrigin-RevId: 382332475

--
74e261dbb467741b2ddd8b490e04c531fdd2f559 by Martijn Vels <mvels@google.com>:

Add BTREE tag for CordRepNode implementing a Btree cord.

This change only forward declared the CordRepBtree class (not implemented yet) and defines the enum value BTREE. While RING and BTREE should never co-exist, we define a new value for BTREE so as not to make transitioning between RING and BTREE harder than it needs to be. This changes shifts the FLAT value / computation from FLAT = 4 to FLAT =5

PiperOrigin-RevId: 382326710
GitOrigin-RevId: 1620e8ffaa93ef24510ca60c7fff2a07248ac9f6
Change-Id: Ia8f99dde3874808f56062bd37ab3e63764099734

1 files changed, 7 insertions, 0 deletions

diff --git a/absl/hash/internal/hash.h b/absl/hash/internal/hash.h
index 69dbbc6b..90627e00 100644
--- a/absl/hash/internal/hash.h
+++ b/absl/hash/internal/hash.h
@@ -856,8 +856,15 @@ class ABSL_DLL MixingHashState : public HashStateBase<MixingHashState> {
   }
 
   ABSL_ATTRIBUTE_ALWAYS_INLINE static uint64_t Mix(uint64_t state, uint64_t v) {
+#if defined(__aarch64__)
+    // On AArch64, calculating a 128-bit product is inefficient, because it
+    // requires a sequence of two instructions to calculate the upper and lower
+    // halves of the result.
+    using MultType = uint64_t;
+#else
     using MultType =
         absl::conditional_t<sizeof(size_t) == 4, uint64_t, uint128>;
+#endif
     // We do the addition in 64-bit space to make sure the 128-bit
     // multiplication is fast. If we were to do it as MultType the compiler has
     // to assume that the high word is non-zero and needs to perform 2