(Wangle) Allocate lambda space inside Core instead of inside std::function

Summary:
Taking this trick that is used in the fibers library. We can keep 64
bytes of space inside `Core` to allocate the callback lambda into,
instead of having `std::function` do another `malloc`. This seems to
greatly improve the synthethic benchmark, and hopefully production
workloads, too, by reducing the number of mallocs.

64 bytes were picked because it fits all the lambdas in the futures
tests. We might want to adjust this based on production data...?

https://fb.intern.facebook.com/groups/715931878455430/permalink/837898842925399/

Test Plan: Run all the tests for all platforms, compilers, and Windtunnel.

Reviewed By: hans@fb.com

Subscribers: chalfant, meisner, folly-diffs@, jsedgwick, yfeldblum

FB internal diff: D1931620

Signature: t1:1931620:1427841595:6ec667b58980be232dfb116bc316148bb67de4fc

folly/futures/detail/Core.h

@@ -80,7 +80,7 @@ class Core {
   /// off-hand, I'm punting.
   Core() {}
   ~Core() {
-    assert(detached_ == 2);
+    assert(attached_ == 0);
   }
 
   // not copyable
@@ -119,13 +119,33 @@ class Core {
     }
   }
 
+  template <typename F>
+  class LambdaBufHelper {
+   public:
+    explicit LambdaBufHelper(F&& func) : func_(std::forward<F>(func)) {}
+    void operator()(Try<T>&& t) {
+      SCOPE_EXIT { this->~LambdaBufHelper(); };
+      func_(std::move(t));
+    }
+   private:
+    F func_;
+  };
+
   /// Call only from Future thread.
   template <typename F>
   void setCallback(F func) {
     bool transitionToArmed = false;
     auto setCallback_ = [&]{
       context_ = RequestContext::saveContext();
-      callback_ = std::move(func);
+
+      // Move the lambda into the Core if it fits
+      if (sizeof(LambdaBufHelper<F>) <= lambdaBufSize) {
+        auto funcLoc = static_cast<LambdaBufHelper<F>*>((void*)lambdaBuf_);
+        new (funcLoc) LambdaBufHelper<F>(std::forward<F>(func));
+        callback_ = std::ref(*funcLoc);
+      } else {
+        callback_ = std::move(func);
+      }
     };
 
     FSM_START(fsm_)
@@ -265,29 +285,33 @@ class Core {
     // TODO(6115514) semantic race on reading executor_ and setExecutor()
     Executor* x = executor_;
     if (x) {
-      MoveWrapper<std::function<void(Try<T>&&)>> cb(std::move(callback_));
-      MoveWrapper<Try<T>> val(std::move(*result_));
-      x->add([cb, val]() mutable { (*cb)(std::move(*val)); });
+      ++attached_; // keep Core alive until executor did its thing
+      x->add([this]() mutable {
+        SCOPE_EXIT { detachOne(); };
+        callback_(std::move(*result_));
+      });
     } else {
       callback_(std::move(*result_));
     }
   }
 
   void detachOne() {
-    auto d = ++detached_;
-    assert(d >= 1);
-    assert(d <= 2);
-    if (d == 2) {
+    auto a = --attached_;
+    assert(a >= 0);
+    assert(a <= 2);
+    if (a == 0) {
       delete this;
     }
   }
 
   FSM<State> fsm_ {State::Start};
-  std::atomic<unsigned char> detached_ {0};
+  std::atomic<unsigned char> attached_ {2};
   std::atomic<bool> active_ {true};
   folly::MicroSpinLock interruptLock_ {0};
   folly::Optional<Try<T>> result_ {};
   std::function<void(Try<T>&&)> callback_ {nullptr};
+  static constexpr size_t lambdaBufSize = 8 * sizeof(void*);
+  char lambdaBuf_[lambdaBufSize];
   std::shared_ptr<RequestContext> context_ {nullptr};
   std::atomic<Executor*> executor_ {nullptr};
   std::unique_ptr<exception_wrapper> interrupt_ {};

folly/futures/test/FutureTest.cpp

@@ -63,7 +63,9 @@ class ThreadExecutor : public Executor {
 
  public:
   explicit ThreadExecutor(size_t n = 1024)
-    : funcs(n), worker(std::bind(&ThreadExecutor::work, this)) {}
+    : funcs(n) {
+    worker = std::thread(std::bind(&ThreadExecutor::work, this));
+  }
 
   ~ThreadExecutor() {
     done = true;
@@ -88,7 +90,7 @@ static eggs_t eggs("eggs");
 TEST(Future, coreSize) {
   // If this number goes down, it's fine!
   // If it goes up, please seek professional advice ;-)
-  EXPECT_EQ(128, sizeof(detail::Core<void>));
+  EXPECT_EQ(192, sizeof(detail::Core<void>));
 }
 
 // Future