Hacker News Reader

> It also states that it's more work for the compiler, stating that the template resolution seems more complex, and suggesting it's likely to bloat compile times. That's a more persuasive point (hard numbers are provided at the end of the article)

I take the author's point about the cost of compiling a template, but the benchmark is odd: a thousand functions with a single call to emplace_back(). C++ programmers are (in)famously willing to trade compile-time performance for run-time performance. Complement the compile benchmark with a realistic program benchmark. In a typical program, I would expect emplace_back() to be called many more times than it's compiled.

Well I'm willing to trade compile-time performance for run-time performance on parts where it does matter. But these parts are quite rare, most of the code of an application is not a performance hot spot where using emplace_back instead of push_back will make a measurable performance difference. So I think the author about compile time is interesting.

When adding items to maps it becomes less obvious (when learning from examples at least) whether map.emplace({Key(args), Value(args)}) avoids copies for Key, Value, pair<Key, Value>, or all of them.

> To say that something isn't magic is an empty statement

I think this is said to express "simply replacing that function call/method/... without changing your calls/methods/coding style is not going to magically solve your problems". Which is true in the push_back/emplace_back case (you need to adjust your call).

GC, on the other hand, is magic pixie dust, if my definition is correct ;)

> I think this is said to express "simply replacing that function call/method/... without changing your calls/methods/coding style is not going to magically solve your problems"

I agree, but you're doing all the work in conjuring this interpretation. Saying it's not magic doesn't say anything substantial.

If it were possible to mechanically replace all references to push_back by emplace_back for an easy performance boost, then we wouldn't need emplace_back to be its own member function, it would just be a compiler optimisation. You have to make shallow and localised changes to make proper use of emplace_back, but isn't that the best we can hope for? Far easier than parallelism, say, where sweeping architectural changes may be necessary.

> auto w = Widget(1,2,3);

Note that unlike in Java or C# this is not the right way to initialize a variable in C++.

Great point, I'd missed that! It should be:

    Widget w(1,2,3);

Herb Sutter offers another viewpoint:

In particular: #5 on that page: "we see that the right-hand style is not only more robust and maintainable for the reasons already given, but also arguably cleaner and more regular with the type consistently on the right when it is mentioned"

[1] https://herbsutter.com/2013/08/12/gotw-94-solution-aaa-style...

I don't think Herb's points apply to an ordinary declare-an-object-and-immediately-initialize-using-the-constructor, do they?

He gives an example using aggregate initialization, but that's not the same thing:

    // Classic C++ declaration order     // Modern C++ style

    employee e{ empid };                 auto e = employee{ empid };
    widget w{ 12, 34 };                  auto w = widget{ 12, 34 };

He makes the case that it's nicer to use auto pretty much wherever we can, to avoid repeating ourselves regarding type, and to avoid unexpected conversions. The auto keyword doesn't apply in our case, there's no way to use auto to simplify a statement like:

    Widget w(1,2,3);

Exactly. auto is only useful in contexts where the type is expected to be inferred (by the compiler).

Actually, he does discuss this in section 4b. He doesn't come out as strongly in favor of it, "The jury is still out on whether to recommend this one wholesale," although he does list what he considers some advantages. He also acknowledges in section 6 that it doesn't work for non-moveable types.

Despite my tendencies towards almost-always auto, I can't say that I consistently use auto x = type{init}, but there is certainly nothing wrong with it.

> He also acknowledges in section 6 that it doesn't work for non-moveable types.

Actually these days it does. Copy/Move elision is mandated by the language in this case and no valid copy/move constructor is required. Similarly is no possible to return non-movable non-copyable types from functions which opens the possibility of interesting code patternes.

Well, that piece of code is a perfectly valid way to initialize a variable in C++.

It is a completely valid way of initializing a variable in modern C++. Modern compilers will not create a copy assignment, if that is what you are afraid of.

> Modern compilers will not create a copy assignment, if that is what you are afraid of.

Probably, but using the canonical syntax removes all doubt, following the C++ language specification. That's a small but real disadvantage in robustness when using the unnecessary assignment, and I'm not aware of any reason to favour using it.

Copy elision in this example is also required by the language specification (if you're using a modern version of C++).

Neat, I didn't know that.

I still don't like it though. I'd rather express what I want to happen than express something I don't want to happen in the knowledge that the standard requires my compiler to be sufficiently smart to repair my statement. Consistent behaviour across different language versions is another plus.

If you'll permit me a moment's snark:

Most OOP languages: You can't copy objects. You may copy references to objects. Copying objects is one road to madness. Move semantics is another.

Old C++: You can copy objects, and specify your own copy constructors (which are permitted to side-effect), but the compiler is permitted to optimise away copy operations under certain circumstances. You shouldn't assume your copy constructor will be invoked whenever you see what looks like a copy.

New C++: You can copy objects, and specify your own copy constructors (which are permitted to side-effect), but the compiler is permitted to optimise away copy operations under certain circumstances, and beyond that, is now required to optimise away copy operations in some specific circumstances. You shouldn't assume your copy constructor will be invoked whenever you see what looks like a copy. When you see a copy operation in certain specific contexts, you are permitted to assume your copy constructor will not be invoked.

As with so many discussions of C++, I'm reminded of two quotes:

> Within C++, there is a much smaller and cleaner language struggling to get out.

- Bjarne Stroustrup [0]

> Do you know how the Orcs first came into being? They were elves once, taken by the dark powers, tortured and mutilated. A ruined and terrible form of life. And now... perfected.

- Saruman

[0] https://en.wikiquote.org/wiki/Bjarne_Stroustrup