shrink SingletonRelaxedCounter inline slow path

Summary: Persuade the compiler to emit a tail call in the `SingletonRelaxedCounter` inline slow path.

Differential Revision: D27383131

fbshipit-source-id: 6f58b1eb23477c00221515d7d6e55a91668f7c84

folly/experimental/SingletonRelaxedCounter.h

@@ -141,8 +141,7 @@ class SingletonRelaxedCounter {
     }
   };
 
-  FOLLY_ALWAYS_INLINE static void mutate(Signed v) {
-    auto cl = counter();
+  FOLLY_ERASE static void mutate(Signed v, CounterRefAndLocal cl) {
     auto& c = *cl.counter;
     if (cl.local) {
       // splitting load/store on the local counter is faster than fetch-and-add
@@ -153,6 +152,14 @@ class SingletonRelaxedCounter {
     }
   }
 
+  FOLLY_NOINLINE static void mutate_slow(Signed v) { mutate(v, counter()); }
+
+  FOLLY_ERASE static void mutate(Signed v, void (&slow)(Signed) = mutate_slow) {
+    auto const cache = local().cache; // a copy! null before/after LocalLifetime
+    // fun-ref to trick compiler into emitting a tail call
+    FOLLY_LIKELY(!!cache) ? mutate(v, {cache, true}) : slow(v);
+  }
+
   FOLLY_EXPORT FOLLY_ALWAYS_INLINE static CounterAndCache& local() {
     // this is a member function local instead of a class member because of
     // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66944