Implement exponential backoff in spinlock (#2391)

Summary:
X-link: facebook/hhvm#9592

Pull Request resolved: #2391

Implement exponential backoff in spinlocks runniong on ARM, as suggested by: https://community.arm.com/arm-community-blogs/b/architectures-and-processors-blog/posts/multi-threaded-applications-arm

We have noticed throughput almost doubles, on the 32 and 64 thread-count cases within the small locks benchmark.

In the below results, 'Sum' is throughput:

before:

------- folly::MicroSpinLock 32 threads
Sum: 67853388 Mean: 942408 stddev: 108821
Lock time stats in us: mean 3 stddev 63 max 17020
------- folly::MicroSpinLock 64 threads
Sum: 28845120 Mean: 400626 stddev: 61624
Lock time stats in us: mean 9 stddev 149 max 30879

after:

------- folly::MicroSpinLock 32 threads
Sum: 116020339 Mean: 1611393 stddev: 27242
Lock time stats in us: mean 1 stddev 3 max 8394
------- folly::MicroSpinLock 64 threads
Sum: 54264302 Mean: 753670 stddev: 30988
Lock time stats in us: mean 4 stddev 3 max 12924

Reviewed By: Gownta

Differential Revision: D70250664

fbshipit-source-id: 7277fee05b451dafb5e8d61240efd138636b4684

folly/synchronization/detail/Sleeper.h

@@ -36,7 +36,11 @@ class Sleeper {
   const std::chrono::nanoseconds delta;
   uint32_t spinCount;
 
-  static constexpr uint32_t kMaxActiveSpin = 4000;
+#if FOLLY_AARCH64 && !FOLLY_MOBILE
+  uint32_t spinCountLog = 1;
+#endif
+
+  static constexpr uint32_t kMaxActiveSpin = 4096;
 
  public:
   static constexpr std::chrono::nanoseconds kMinYieldingSleep =
@@ -49,8 +53,15 @@ class Sleeper {
 
   void wait() noexcept {
     if (spinCount < kMaxActiveSpin) {
+#if FOLLY_AARCH64 && !FOLLY_MOBILE
+      do {
+        asm_volatile_pause();
+      } while (++spinCount < spinCountLog);
+      spinCountLog <<= 1;
+#else
       ++spinCount;
       asm_volatile_pause();
+#endif
     } else {
       /* sleep override */
       std::this_thread::sleep_for(delta);