Use compact_once_flag in DelayedInit only if size can be reduced

Summary: If `T` is greater than 4-byte aligned, then `compact_once_flag` brings no size improvements. We can use `once_flag` in these cases, which has a better mutex (`folly::SharedMutex` rather than `folly::MicroLock`).

Reviewed By: yfeldblum

Differential Revision: D26214262

fbshipit-source-id: ea15376ea892df42cc9d3a6eac3da458fe0564a8

folly/synchronization/DelayedInit.h

@@ -149,11 +149,16 @@ struct DelayedInit {
     }
   }
 
+  using OnceFlag = std::conditional_t<
+      alignof(T) >= sizeof(once_flag),
+      once_flag,
+      compact_once_flag>;
+
   struct StorageTriviallyDestructible {
     union {
       std::remove_const_t<T> value;
     };
-    compact_once_flag init;
+    OnceFlag init;
 
     StorageTriviallyDestructible() {}
   };
@@ -162,7 +167,7 @@ struct DelayedInit {
     union {
       std::remove_const_t<T> value;
     };
-    compact_once_flag init;
+    OnceFlag init;
 
     StorageNonTriviallyDestructible() {}
     ~StorageNonTriviallyDestructible() {

folly/synchronization/test/DelayedInitTest.cpp

@@ -232,6 +232,26 @@ TEST(DelayedInit, ConstType) {
   EXPECT_EQ(lazy->value, 12);
 }
 
+namespace {
+template <typename T>
+class DelayedInitSizeTest : public testing::Test {};
+
+template <typename T>
+struct WithOneByte {
+  T t;
+  char c;
+};
+} // namespace
+
+using DelayedInitSizeTestTypes =
+    testing::Types<char, short, int, long, long long, char[3], short[2]>;
+TYPED_TEST_CASE(DelayedInitSizeTest, DelayedInitSizeTestTypes);
+
+TYPED_TEST(DelayedInitSizeTest, Size) {
+  // DelayedInit should not add more than 1-byte size overhead (modulo padding)
+  EXPECT_EQ(sizeof(DelayedInit<TypeParam>), sizeof(WithOneByte<TypeParam>));
+}
+
 TEST(DelayedInit, Concurrent) {
   CtorCounts counts;
   DelayedInit<CtorCounts::Tracker> lazy;