optimize bits of HistogramWrapper::add

Summary:
* Move deferred initialization in `ensureApplySpec` from the fast-path to the slow-path.
* Use `ThreadLocalPtr<TLHistogram>` v.s. `ThreadLocal<shared_ptr<TLHistogram>>` to reduce object size and to remove one level of indirection. `ThreadLocalPtr<T>` has special support for being assigned `shared_ptr<T>`.
* Put spec on heap and delete it after deferred init to reduce its memory consumption.

Reviewed By: Gownta

Differential Revision: D66590811

fbshipit-source-id: 10ed7c5219bd6cfde0ca7d6d18ce52040ab239bb

fb303/ThreadCachedServiceData.cpp

@@ -113,6 +113,19 @@ SubminuteMinuteOnlyTimeseriesWrapper::templateExportedStat() {
   return *obj.get();
 }
 
+FOLLY_NOINLINE void HistogramWrapper::doApplySpecLocked() {
+  std::exchange(spec_, {})->apply(key_, ServiceData::get());
+}
+
+FOLLY_NOINLINE ThreadCachedServiceData::TLHistogram*
+HistogramWrapper::tcHistogramSlow() {
+  folly::call_once(once_, &HistogramWrapper::doApplySpecLocked, this);
+  auto& local = ThreadCachedServiceData::getStatsThreadLocal();
+  auto const histogram = local->getHistogramSafe(key_);
+  tlHistogram_.reset(histogram);
+  return histogram.get();
+}
+
 ThreadCachedServiceData::StatsThreadLocal&
 ThreadCachedServiceData::getStatsThreadLocal() {
   static folly::Indestructible<ThreadCachedServiceData::StatsThreadLocal>

fb303/ThreadCachedServiceData.h

@@ -1204,39 +1204,30 @@ class HistogramWrapper {
       int64_t min,
       int64_t max,
       const Args&... args)
-      : key_(key), spec_(bucketWidth, min, max, args...) {}
+      : key_(key),
+        spec_(std::make_unique<internal::HistogramSpec>(
+            bucketWidth,
+            min,
+            max,
+            args...)) {}
 
   void add(int64_t value = 1) {
-    ensureApplySpec();
     tcHistogram()->addValue(value);
   }
 
  protected:
-  void doApplySpecLocked() {
-    spec_.apply(key_, ThreadCachedServiceData::get());
-  }
-  FOLLY_ALWAYS_INLINE void ensureApplySpec() {
-    // minimize inline slow path size by passing the callback this way
-    folly::call_once(once_, &HistogramWrapper::doApplySpecLocked, this);
-  }
-
-  inline ThreadCachedServiceData::TLHistogram* tcHistogram() {
-    ThreadCachedServiceData::TLHistogram* cached = tlHistogram_->get();
-    if (cached) {
-      return cached;
-    }
+  void doApplySpecLocked();
 
-    auto histogram =
-        ThreadCachedServiceData::getStatsThreadLocal()->getHistogramSafe(key_);
-    *tlHistogram_ = histogram;
-    return histogram.get();
+  FOLLY_ALWAYS_INLINE ThreadCachedServiceData::TLHistogram* tcHistogram() {
+    ThreadCachedServiceData::TLHistogram* cached = tlHistogram_.get();
+    return FOLLY_LIKELY(!!cached) ? cached : tcHistogramSlow();
   }
+  ThreadCachedServiceData::TLHistogram* tcHistogramSlow();
 
   folly::once_flag once_;
   std::string key_;
-  internal::HistogramSpec spec_;
-  folly::ThreadLocal<std::shared_ptr<ThreadCachedServiceData::TLHistogram>>
-      tlHistogram_;
+  std::unique_ptr<internal::HistogramSpec> spec_; // ptr for memory-size
+  folly::ThreadLocalPtr<ThreadCachedServiceData::TLHistogram> tlHistogram_;
 };
 
 class MinuteOnlyHistogram : public HistogramWrapper {