[PartitionAlloc] Make SpinningFutex's spinning path more like SpinLock.

That is:
- Increase the number of spinning iterations
- Do a relaxed atomic check before the CAS in the spin loop.

Bug: 1125999, 1061437
Change-Id: I4f33376b55206dff9557a9ee59109aeaa06b5a56
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/2424310Reviewed-by: Kentaro Hara <haraken@chromium.org>
Commit-Queue: Benoit L <lizeb@chromium.org>
Cr-Commit-Position: refs/heads/master@{#809714}

base/allocator/partition_allocator/spinning_futex_linux.h

@@ -56,8 +56,9 @@ class BASE_EXPORT SpinningFutex {
   static constexpr int kLockedUncontended = 1;
   static constexpr int kLockedContended = 2;
 
-  // Same as SpinLock, not scientifically calibrated.
-  static constexpr int kSpinCount = 10;
+  // Same as SpinLock, not scientifically calibrated. Consider lowering later,
+  // as the slow path has better characteristics than SpinLocks's.
+  static constexpr int kSpinCount = 1000;
 
   std::atomic<int32_t> state_{kUnlocked};
 };
@@ -78,7 +79,8 @@ ALWAYS_INLINE void SpinningFutex::Acquire() {
 
 ALWAYS_INLINE bool SpinningFutex::Try() {
   int expected = kUnlocked;
-  return state_.compare_exchange_strong(expected, kLockedUncontended,
+  return (state_.load(std::memory_order_relaxed) == expected) &&
+         state_.compare_exchange_strong(expected, kLockedUncontended,
                                         std::memory_order_acquire,
                                         std::memory_order_relaxed);
 }