[tmyklebu]

In modern gcc, std::variant doesn't look like a competent replacement for old-fashioned tagged unions.

On x86_64 Linux, it looks like a function with signature `void f(std::variant<int, char>)` expects its (8-byte) argument to be passed by reference, whereas `void f2(tagged_union_of_int_and_char)` passes its argument in rdi.

gcc (-O3) also generates miserable code for calling f(42):

  4004f0:       48 83 ec 18             sub    $0x18,%rsp
  4004f4:       48 89 e7                mov    %rsp,%rdi
  4004f7:       c7 04 24 2a 00 00 00    movl   $0x2a,(%rsp)
  4004fe:       c6 44 24 04 00          movb   $0x0,0x4(%rsp)
  400503:       e8 c8 ff ff ff          callq  4004d0 <f>
  400508:       48 83 c4 18             add    $0x18,%rsp
  40050c:       c3                      retq

as compared with calling f2(42):

  400510:       48 bf 00 00 00 00 2a    movabs $0x2a00000000,%rdi
  400517:       00 00 00
  40051a:       eb c4                   jmp    4004e0 <f2>

std::optional seems to have the same disease; the ABI is different from a plain struct containing a bool and the value and you get worse initialisation code. There's also a build time penalty to using std::optional; on my machine, including the code that uses optionals makes my trivial test program take 320ms to compile and link rather than 50ms.

These look like yet more new C++ features that are worse than what they're trying to replace.

[ambrop7]

I suspect the issue is that variant<int,char> is not trivially copy constructible (at least when I tested on godbolt), so the ABI believes it is not allowed to pass the object by value. (this is a guess only, correct me if you know better!)

Now why it is not trivially copy constructible is another question. But the latest standard draft does not require it [variant.ctor] though it does require the destructor to be trivial if all type destructors are trivial [variant.dtor]. The GCC implementation also seems to ensure only trivial destructor not copy constructor (the code is here [1] if anyone likes to see).

EDIT: I did a small experiment that agrees with this - adding a copy constructor causes passing by pointer. https://godbolt.org/g/UG0YXJ

EDIT2: I filed a bug report for gcc, https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80187

[1] https://github.com/gcc-mirror/gcc/blob/master/libstdc%2B%2B-...

[gpderetta]

Unfortunately the abi requires non-trivially copyable/destructible types to be passed by (hidden) reference.

It should be possible to implement std::variant in such a way that it is trivially constructible and destructible as long as all elements​ are, but apparently stdlibc++ doesn't. This is GCC choice and other compilers apparently do differently. Unless GCC changes it before finalising their c++17 implementation, the choice will be unfortunately set in stone for them as they guarantee abi stability of the library.

P0602R0 is a proposal to require trivial copy/move (but strangely not destruct), don't know whether it has been accepted though.

[ambrop7]

Right, this is basically what I said in my comment :) Thanks for the proposal reference.

[yongjik]

std::optional is nice if you have a class with invariants that you want to always maintain (i.e., you don't want to deal with the pesky "uninitialized" state). You can build a class that always requires proper parameter in its constructor, and then you apply optional<> only when you defer initialization.

In most cases, I imagine the performance penalty would be dwarfed by the cost of actually using the class.

[typon]

Does anyone know what kind of assembly is generated when using std::option in Rust when make a similar function call?

[steveklabnik]

  void f(std::variant<int, char>)

would be

  f(x: IntOrChar)

roughly in Rust, I'm too lazy to make the types _exactly_ the same, but https://godbolt.org/g/pZde2A is roughly it.

Actually, https://godbolt.org/g/1vcMeG might be even better to compare; I'm not an expert, but looks like the only difference is in the size.

last one, adding inline(never) so you can see the calls in main: https://godbolt.org/g/cVDjQH

  movabs  rdi, 21474836480
  call    example::f1@PLT

vs

  mov     edi, 5
  pop     rbp
  jmp     example::f2@PLT

so, yup.

[typon]

Thank you! This is really cool.

[cfallin]

Another cool feature, on top of steveklabnik's comment: rustc is actually smart enough to collapse `Option<T>` where `T` is a borrowed reference (pointer) to a single word. Basically, the compiler knows that the pointer can never be null (borrows are always valid when in scope), so it can use the non-null values for `Some(...)` and the null value for `None`.

In other words, the compiler can turn an Option into a null-pointer convention by reasoning from first principles.

(I think this optimization works for enums in general, and is somehow related to the `nonzero::NonZero` type, but I could be wrong about that.)

Example (look at the calls to f1 and f2 in example::main): https://godbolt.org/g/SSs6X2

[steveklabnik]

Yes. It's when the type is NonZero.

[chubot]

Is this just a gcc implementation detail or is it required by the ABI?

(I know nothing about std::variant yet.)

[Sharlin]

Eh, I get that one of the points of C++ is to minimize abstraction penalty even when you do use abstractions, but these things are compromises. In almost all real-world software development the advantages of abstraction far outweigh micro-optimizations that shave off individual cycles...

[tmyklebu]

Writing ad-hoc tagged unions and two-element 'optional' structs is a completely straightforward mechanical task---exactly the sort of thing that a programming language ought to automate. Why does there need to be a "compromise" between boilerplate and performance here? Why should I have to pay performance to use standard tagged unions and optionals, or conversely, pay by writing boilerplate in order to get performance?

[wolf550e]

If the individual cycles don't matter, you can use a safer language than C++ which compiles faster and provides more static and dynamic guarantees. If you choose C++ for a new project in 2017 it's because you need all the cycles and would have chosen assembly if you thought you could write the asm fast and cheaply enough.

[nightski]

First of all, the world isn't that black and white. Second, the advantage of C++ is that you are not forced to use these abstractions. You can pick and choose when to use them and pay the performance penalty.

[wolf550e]

For the parts of the codebase where you can use abstractions that have runtime costs, you can also choose a safer language that imposes similar runtime costs. I bet it would be easier to write maintainable code in the safer language than in the version+subset of C++ on which your team agrees.

For the parts of the codebase where you count cycles, you often want to carefully lay out the data and then use hand coded SIMD. Possibly use a specialized code generator that writes the SIMD asm for you, with less chance of error. If you can't do SIMD, I would just use C, but use a safety-critical code standard for the C part.

Existing codebases and existing investment into learning C++ arcana complicates these decisions, but I truly believe C++ is not worth it for new projects.

[fsloth]

"I truly believe C++ is not worth it for new projects."

So what would you use if the constraints for the project are of the usual C++ kind? I.e. language has industrial support, proven toolchain, loads of potential candidates who can write the language, the code will compile 20 years in the future, can tightly integrate with a gui ... etc. Honestly, if there is a better language where I percieve C++ to be the best language I would dearly like to know. Half of the reasons I use C++ have nothing to do with the language itself but of the modern computing ecosystem.

[pekk]

Many other languages have industrial support, a proven toolchain, and are known by a lot of candidates: Java, C#, modern C... this is if we restrict ourselves to only the most popular languages with compile-time type enforcement, which is fairly arbitrary.

It's okay to admit that you like C++, it's what you know best, and that it's not unsuitable for the things you want to do. The same as someone else says that about F# or Lua.

[wolf550e]

It would really depend on what the project is (3d shooter game vs database server vs high frequency trading). But I think many projects can be broken into two parts:

One is a limited number of "computation kernels" (processing packets, processing transactions, computing a photoshop filter) that need to be very fast but can be carefully generated/handcoded in C or SIMD asm or whatever that has very simple behavior/API, for example it should not allocate memory.

The other part is the complicated part that has all the complex behavior (a lot of code, changing business requirements, opportunity for abstraction) that configures the kernels and feed data into them and interacts with the user and the network but takes a small part of the CPU time.

The second part must be organized so that it will not inhibit the first part in any way, so that means the second part can't use a language that won't let you layout stuff in memory in exactly the way you need or that has poor cffi performance.

The second part can be done in something as slow as cpython! (though it doesn't have to be quite that slow).

[blub]

Your whole argument falls apart when one considers that performance needs aren't split just between SIMD + careful data layout and anything goes.

C++ isn't used only for generating the fastest possible code, but also to generate code with predictable performance, code that is default-fast and code that can be optimized without the fear of having to change programming languages to reach a performance goal.

[pekk]

And the disadvantage is that you have to pick and choose them after learning about all of them, and you have to deal with the choices made by whatever other code you have to work with, and you have to reason about all the combinations. For most tasks, this is overkill, which is why C and languages other than C++ persist even though C++ has everything.

[xwvvvvwx]

I think the point is not that the abstractions have some cost, but that the cost is higher than it needs to be.

I would also agree that pure performance and deterministic memory management are the two main reasons to use c++ for a new project nowadays.

[blub]

I've seen this kind of misconception before and it should be corrected. Your premise is wrong, there are other reasons to pick C++ besides performance:

* it's platform-native on the major platforms and many others

* it's an international standard and has excellent backwards compatibility. i.e: it's not controlled by some corporation and it's very unlikely that a 2v3 Python-style fiasco would happen.

* it has a huge ecosystem. The only comparable ones are Java and maybe C#, but while there's overlap, each ecosystem has its focus. Java's is focused on the back-end and enterprise development.

* it's a flexible and powerful language

* it's basically the default option in certain domains