Since the year-end mini-update, progress has continued on cppfront. (If you don’t know what this personal project is, please see the CppCon 2022 talk on YouTube.)

This update covers Acknowledgments, and highlights of what’s new in the compiler and language since last time, including:

Acknowledgments: 267 issues, 128 pull requests, and new collaborators

I want to say a big “thank you” again to everyone who has participated in the cppfront repo. Since the last update, I’ve merged PRs from Jo Bates, Gabriel Gerlero, jarzec, Greg Marr, Pierre Renaux, Filip Sajdak and Nick Treleaven. Thanks also to many great issues opened by (as alphabetically as I can): Abhinav00, Robert Adam, Adam, Aaron Albers, Alex, Graham Asher, Peter Barnett, Sean Baxter, Jan Bielak, Simon Buchan, Michael Clausen, ct-clmsn, Joshua Dahl, Denis, Matthew Deweese, dmicsa, dobkeratops, Deven Dranaga, Konstantin F, Igor Ferreira, Stefano Fiorentino, fknauf, Robert Fry, Artie Fuffkin, Gabriel Gerlero, Matt Godbolt, William Gooch, ILoveGoulash, Víctor M. González, Terence J. Grant, GrigorenkoPV, HALL9kv0, Morten Hattesen, Neil Henderson, Michael Hermier, h-vetinari, Stefan Isak, Tim Keitt, Vanya Khodor, Hugo Lindström, Ferenc Nandor Janky, jarzec, jgarvin, Dominik Kaszewski, kelbon, Marek Knápek, Emilia Kond, Vladimir Kraus, Ahmed Al Lakani, Junyoung Lee, Greg Marr, megajocke, Thomas Neumann, Niel, Jim Northrup, Daniel Oberhoff, Jussi Pakkanen, PaTiToMaSteR, Johel Ernesto Guerrero Peña, Bastien Penavayre, Daniel Pfeifer, Piotr, Davide Pomi, Andrei Rabusov, rconde01, realgdman, Alex Reinking, Pierre Renaux, Alexey Rochev, RPeschke, Sadeq, Filip Sajdak, satu, Wolf Seifert, Tor Shepherd, Luke Shore, Zenja Solaja, Francis Grizzly Smit, Sören Sprößig, Benjamin Summerton, Hypatia of Sva, SwitchBlade, Ramy Tarchichy, tkielan, Marzo Sette Torres Junior, Nick Treleaven, Jan Tusil, userxfce, Ezekiel Warren, Kayla Washburn, Tyler Weaver, Will Wray, and thanks also to many others who participated on PR reviews and comment threads.

These contributors represent people from high school and undergrad students to full professors, from commercial developers to conference speakers, and from every continent except Antarctica. Thank you!

Next, here are some highlights of things added to the cppfront compiler in the four months since the previous update linked at top.

Simple, mathematically safe, and efficient chained comparisons (commit)

P0515 “Consistent comparison” (aka “spaceship”) was the first feature derived from this Cpp2 work to be adopted into the ISO Standard for C++, in C++20. That means cppfront didn’t have to do much to implement operator<=> and its generative semantics, because C++20 compilers already do so, which is great. Thank you again to everyone who helped land this Cpp2 feature in the ISO C++ Standard.

However, one part of P0515 isn’t yet merged into ISO C++: chained comparisons from P0515 section 3.3, such as min <= index < max. See also Barry Revzin‘s great followup ISO C++ proposal paper P0893 “Chaining comparisons.” The cppfront compiler now implements this as described in those ISO proposal papers, and:

I think this is a great example to demonstrate that “simple,” “safe,” and “fast” are often not in tension, and how it’s often possible to get all three at the same time without compromises.

Named break and continue (commit)

This feature further expands the Cpp2 “name :” way of introducing all names, to also support introducing loop names. Examples like the following now work… see test file pure2-break-continue.cpp2 for more examples.

outer: while i<M next i++ {      // loop named "outer"
    // ...
    inner: while j<N next j++ {  // loop named "inner"
        // ...
        if something() {
            continue inner;      // continue the inner loop
        }
        // ...
        if something_else() {
            break outer;         // break the outer loop
        }
        // ...
    }
    // ...
}

“Simple and safe” starts with . . . main

main can now be defined to return nothing, and/or as main: (args) to have a single argument of type std::vector<std::string_view>.

For example, here is a complete compilable and runnable program (in -pure-cpp2 mode, no #include is needed to use the C++ standard library)… 

main: (args) =
    std::cout << "This program's name is (args[0])$";

Yes, this really is 100% C++ under the covers as you can see on Godbolt Compiler Explorer… “just nicer”:

type: User-defined types

User-defined types are written using the same name : kind = value syntax as everything in Cpp2:

mytype: type =
{
    // data members are private by default
    x: std::string;

    // functions are public by default
    protected f: (this) = { do_something_with(x); }

    // ...
}

Here are some highlights…

First, types are order-independent. Cpp2 still has no forward declarations, and you can just write types that refer to each other. For example, see the test case pure2-types-order-independence-and-nesting.cpp2.

The this parameter is explicit, and has special sauce:

For example, here is how to write const member function named print that takes a const string value and prints this object’s data value and the string message (yes, everything in Cpp2 is const by default except for local-scope variables):

mytype: type =
{
    data: i32;   // some data member (private by default)

    print: (this, msg: std::string) = {
        std::cout << data << msg;
                 // "data" is shorthand for "this.data"
    }

    // ...
}

All Cpp1 special member functions (including construction, assignment, destruction) and conversions are unified as operator=, default to memberwise semantics and safe “explicit” by default, and there’s a special that parameter that makes writing copy/move in particular simpler and safer. On the cppfront wiki, see the Design Note “operator=, this & that” for details. Briefly summarizing here:

Note: Because generated functions are always opt-in, you can never get a generated function that’s wrong for your type, and so Cpp2 doesn’t need to support “=delete” for the purpose of suppressing unwanted generated functions.

Virtual functions and base classes are all about “this”:

Alias support: Type, namespace, function, and object aliases (commit)

Cpp2 already defines every new entity using the syntax “name : kind = value“.

So how should it declare aliases, which declare not a new entity but a synonym for an existing entity? I considered several alternatives, and decided to try out the identical declaration syntax except changing = (which connotes value setting) to == (which connotes sameness):

// Namespace alias
lit: namespace == ::std::literals;

// Type alias
pmr_vec: <T> type
    == std::vector<T, std::pmr::polymorphic_allocator<T>>;

// Function alias
func :== some_original::inconvenient::function_name;

// Object alias
vec :== my_vector;  // note: const&, aliases are never mutable

Note again the const default. For now, Cpp2 supports only read-only aliases, not read-write aliases.

Header reflect.h: Initial support for reflection API, and implementing the first 10 working metafunctions from P0707… interface, polymorphic_base, ordered, weakly_ordered, partially_ordered, basic_value, value, weakly_ordered_value, partially_ordered_value, struct

Disclaimer: I have not yet implemented a reflection operator that Cpp2 code can invoke, or written a Cpp2 interpreter to run inside the compiler. But I am doing everything else needed for type metafunctions: cppfront has started a usable reflection metatype API, and has started getting working metafunctions that are compile-time code that uses that metatype API… the only thing missing is that those functions aren’t run through an interpreter (yet).

For example, cppfront now supports code like the following… importantly, “value” and “interface” are not built-in types hardwired into the language as they are in Java and C# and other languages, but rather each is a function that uses the reflection API to apply requirements and defaults to the type (C++ class) being written:

// Point2D is declaratively a value type: it is guaranteed to have
// default/copy/move construction and <=> std::strong_ordering
// comparison (each generated with memberwise semantics
// if the user didn't write their own, because "@value" explicitly
// opts in to ask for these functions), a public destructor, and
// no protected or virtual functions... the word "value" carries
// all that meaning as a convenient and readable opt-in, but
// without hardwiring "value" specially into the language
//
Point2D: @value type = {
    x: i32 = 0;  // data members (private by default)
    y: i32 = 0;  // with default values
    // ...
}

// Shape is declaratively an abstract base class having only public
// and pure virtual functions (with "public" and "virtual" applied
// by default if the user didn't write an access specifier on a
// function, because "@interface" explicitly opts in to ask for
// these defaults), and a public pure virtual destructor (generated
// by default if not user-written)... the word "interface" carries
// all that meaning as a convenient and readable opt-in, but
// without hardwiring "interface" specially into the language
//
Shape: @interface type = {
    draw: (this);
    move: (inout this, offset: Point2D);
}

At compile time, cppfront parses the type’s body and then invokes the compile-time metafunction (here value or interface), which enforces requirements and applies defaults and generates functions, such as… well, I can just paste the actual code for interface from reflect.h, it’s pretty readable:

Note: For now I wrote the code in today’s Cpp1 syntax, which works fine as Cpp2 is just a fully compatible alternate syntax for the same true C++… later this year I aim to start self-hosting and writing more of cppfront itself in Cpp2 syntax, including functions like these.

//-----------------------------------------------------------------------
//  Some common metafunction helpers (metafunctions are just
//  functions, so they can be factored as usual)
//
auto add_virtual_destructor(meta::type_declaration& t)
    -> void
{
    t.require( t.add_member( "operator=: (virtual move this) = { }"),
               "could not add virtual destructor");
}

//-----------------------------------------------------------------------
// 
//      "... an abstract base class defines an interface ..."
// 
//          -- Stroustrup (The Design and Evolution of C++, 12.3.1)
//
//-----------------------------------------------------------------------
// 
//  interface
//
//  an abstract base class having only pure virtual functions
//
auto interface(meta::type_declaration& t)
    -> void
{
    auto has_dtor = false;

    for (auto m : t.get_members())
    {
        m.require( !m.is_object(),
                   "interfaces may not contain data objects");
        if (m.is_function()) {
            auto mf = m.as_function();
            mf.require( !mf.is_copy_or_move(),
                        "interfaces may not copy or move; consider a virtual clone() instead");
            mf.require( !mf.has_initializer(),
                        "interface functions must not have a function body; remove the '=' initializer");
            mf.require( mf.make_public(),
                        "interface functions must be public");
            mf.make_virtual();
            has_dtor |= mf.is_destructor();
        }
    }

    if (!has_dtor) {
        add_virtual_destructor(t);
    }
}

Note a few things that are demonstrated here:

Today in cppfront, metafunctions like value and interface are legitimately doing everything envisioned for them in P0707 except for being run through an interpreter — the metafunctions are using the meta:: API and exercising it so I can learn how that API should expand and become richer, cppfront is spinning up a new lexer and parser when a metafunction asks to do code generation to add a member, and then cppfront is stitching the generated result into the parse tree as if it had been written by the user explicitly… this implementation is doing everything I envisioned for it in P0707 except for being run through an interpreter.

As of this writing, here are the currently implemented metafunctions in reflect.h are as described in P0707 section 3, sometimes with a minor name change… and including links to the function source code…

interface: An abstract class having only pure virtual functions.

polymorphic_base (in P0707, originally named base_class): A pure polymorphic base type that has no instance data, is not copyable, and whose destructor is either public and virtual or protected and nonvirtual.

ordered: A totally ordered type with operator<=> that implements std::strong_ordering.

Similarly, weakly_ordered and partially_ordered do the same for std::weak_ordering and std::partial_ordering respectively. I chose to call the strongly-ordered one “ordered,” not “strong_ordered,” because I think the one that should be encouraged as the default should get the nice name.

basic_value: A type that is copyable and has value semantics. It must have all-public default construction, copy/move construction/assignment, and destruction, all of which are generated by default if not user-written; and it must not have any protected or virtual functions (including the destructor).

value: A basic_value that is totally ordered.

Note: Many of you would call this a “regular” type… but I recognize that there’s a difference of opinion about whether “regular” includes ordering. That’s one reason I’ve avoided the word “regular” here, and this way we can all separately talk about a basic_value (which may not include ordering) or a value (which does include strong total ordering; see next paragraph for weaker orderings) and we can know we’re all talking about the same thing.

Similarly, weakly_ordered_value and partially_ordered_value do the same for weakly_ordered and partially_ordered respectively. I again chose to call the strongly-ordered one “value,” not “strongly_ordered_value,” because I think the one that should be encouraged as the default should get the nice name.

struct (in P0707, originally named plain_struct because struct is a reserved word in Cpp1… but struct isn’t a reserved word in Cpp2): A basic_value where all members are public, there are no virtual functions, and there are no user-written (non-default operator=) constructors, assignment operators, or destructors.

Local statement/block parameters (commit)

I had long intended to support the following unification of functions and blocks, where cppfront already provided all of these except only the third case:

f:(x: int = init) = { ... }     // x is a parameter to the function
f:(x: int = init) = statement;  // same, { } is implicit

 :(x: int = init) = { ... }     // x is a parameter to the lambda
 :(x: int = init) = statement;  // same, { } is implicit

  (x: int = init)   { ... }     // x is a parameter to the block
  (x: int = init)   statement;  // same, { } is implicit

                    { ... }     // x is a parameter to the block
                    statement;  // same, { } is implicit

(Recall that in Cpp2 : always and only means “declaring a new thing,” and therefore also always has an = immediately or eventually to set the value of that new thing.)

The idea is to treat functions and blocks/statements uniformly, as syntactic and semantic subsets of each other:

In this model, the third (just now implemented) option above allows a block parameter list, which does the same work as “let” variables in other languages, but without a “let” keyword. This would subsume all the Cpp1 loop/branch scope variables (and more generally than in Cpp1 today, because you could declare multiple parameters easily which you can’t currently do with the Cpp1 loop/branch scope variables).

So this now works, pasting from test case pure2-statement-scope-parameters.cpp2:

main: (args) = 
{
    local_int := 42;

    //  'in' statement scope variable
    // declares read-only access to local_int via i
    (i := local_int) for args do (arg) {
        std::cout << i << "\n";       // prints 42
    }

    //  'inout' statement scope variable
    // declares read-write access to local_int via i
    (inout i := local_int) {
        i++;
    }
    std::cout << local_int << "\n";   // prints 43
}

Note that block parameters enable us to use the same declarative data-flow for local statements and blocks as for functions: Above, we declare a block (a statement, in this case a single loop, is implicitly treated as a block) that is read-only with respect to the local variable, and declare another to be read-write with respect to that variable. Being able to declare data flow is important for writing correct and safe code.

Corollary: Removed : and = from for

Eagle-eyed readers of the above example will notice a change: As a result of unifying functions and blocks, I realized that the for loop syntax should use the third syntax, not the first or second, because the loop body is a parameterized block, not a local function. So changed the for syntax from this

// previous syntax
for items do: (item) = {
    x := local + item;
    // ...
}

to this, which is the same except that it removes : and =

// current syntax
for items do (item) {
    x := local + item;
    // ...
}

Note that what follows for ... do is exactly a local block, just the parameter item doesn’t write an initializer because it is implicitly initialized by the for loop with each successive value in the range.

By the way, this is the first breaking change from code that I’ve shown publicly, so cppfront also includes a diagnostic for the old syntax to steer you to the new syntax. Compatibility!

Other features

Also implemented since last time:

What’s next

Well, that’s all so far.

For cppfront, over the summer and fall I plan to:

For conferences: