better emplace implementation

TurboList.hpp

@@ -29,7 +29,7 @@ class TurboList {
 
 	uint32_t capacity;
 
-	TL_NOINLINE void grow_and_insert(T elem) noexcept {
+	TL_NOINLINE T& grow_and_insert(T elem) noexcept {
 			// assert(mid == 0);
 			if(old) FREE(old);
 			old = nex;
@@ -38,7 +38,7 @@ class TurboList {
 			nex = (int *) MALLOC(this->capacity * sizeof(T));
 
 			// Will go into the INSERT code path here
-			insert(elem);
+			return insert(elem);
 	}
 
 public:
@@ -72,7 +72,7 @@ public:
 		}
 	}
 
-	inline void insert(T elem) noexcept {
+	inline T& insert(T elem) noexcept {
 		if(TL_LIKELY(end < capacity)) {
 
 			// INSERT
@@ -87,17 +87,43 @@ public:
 			mid -= hasmid;
 			nex[mid] = hasmid ? old[mid] : nex[mid];
 
-			nex[end++] = elem;
+			return (nex[end++] = elem);
 		} else {
 			// GROW
-			grow_and_insert(elem);
+			return grow_and_insert(elem);
 		}
 	}
 
 	template<typename... Args>
-	inline void emplace(Args&&... args) {
-		// FIXME: placement new needed
-	    insert(T(std::forward<Args>(args)...));
+	inline T& emplace(Args&&... args) {
+		if(TL_LIKELY(end < capacity)) {
+
+			// INSERT
+
+			/* Same as this:
+			if(mid > 0) {
+				nex[mid - 1] = old[mid - 1];
+				--mid;
+			}
+			*/
+			bool hasmid = (mid > 0);
+			mid -= hasmid;
+			nex[mid] = hasmid ? old[mid] : nex[mid];
+
+			// Placement new
+			return *new (nex + end++) T(std::forward<Args>(args)...);
+		} else {
+			// GROW
+			//
+			// Rem.: I just chose this to be less optimized than
+			//       it is possible by making a copy and reusing
+			//       the existing grow and insert code instead of
+			//       writing a new "grow_and_emplace" again.
+			//
+			//       This happens rarely so its probably fine and
+			//       makes less template instantiations, smaller binary.
+			return grow_and_insert(T(std::forward<Args>(args)...));
+		}
 	}
 
 	inline uint32_t size() noexcept {