CODEWITHSUNDEEP
c++boostclangboost-hana
jinscoe123
jinscoe123

Reputation: 1739

Clang compiler error when using boost::hana Y-combinator

I have the following implementation of an AST, built using C++17's std::variant type, on which I would like to apply a visitor recursively. I have done this with the help of some utilities from Boost's Hana library.

#include <iostream>
#include <memory>
#include <variant>

#include <boost/hana.hpp>

namespace hana = boost::hana;


struct AddExpr;
struct SubExpr;
struct MulExpr;
struct DivExpr;

using Expr = std::variant<int, AddExpr, SubExpr, MulExpr, DivExpr>;

struct BinaryExpr
{
    BinaryExpr(std::unique_ptr<Expr> lhs, std::unique_ptr<Expr> rhs) noexcept
        : lhs{ std::move(lhs) }, rhs{ std::move(rhs) } { }

    std::unique_ptr<Expr> lhs;
    std::unique_ptr<Expr> rhs;
};

struct AddExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };
struct SubExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };
struct MulExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };
struct DivExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };


int main()
{
    Expr root {
        MulExpr{
            std::make_unique<Expr>(AddExpr{
                std::make_unique<Expr>(1),
                std::make_unique<Expr>(2),
            }),
            std::make_unique<Expr>(SubExpr{
                std::make_unique<Expr>(3),
                std::make_unique<Expr>(4),
            }),
        }
    };

    constexpr auto printer = hana::fix([](auto visitor, auto const& arg) -> void {
        hana::overload(
            [&](int val) {
                std::cout << val;
            },
            [&](AddExpr const& exp) {
                std::cout << "(";
                std::visit(visitor, *exp.lhs);
                std::cout << "+";
                std::visit(visitor, *exp.rhs);
                std::cout << ")";
            },
            [&](SubExpr const& exp) {
                std::cout << "(";
                std::visit(visitor, *exp.lhs);
                std::cout << "-";
                std::visit(visitor, *exp.rhs);
                std::cout << ")";
            },
            [&](MulExpr const& exp) {
                std::cout << "(";
                std::visit(visitor, *exp.lhs);
                std::cout << "*";
                std::visit(visitor, *exp.rhs);
                std::cout << ")";
            },
            [&](DivExpr const& exp) {
                std::cout << "(";
                std::visit(visitor, *exp.lhs);
                std::cout << "/";
                std::visit(visitor, *exp.rhs);
                std::cout << ")";
            }
        )(arg);
    });

    std::visit(printer, root);

    std::cout << "\n";

    return 0;
}

The code compiles without any warnings when using GCC 7.1+, but Clang does not seem to be able to compile it, issuing the following error instead.

<source>:54:22: error: function 'visit<boost::hana::fix_t<(lambda at <source>:47:40)> &, std::variant<int, AddExpr, SubExpr, MulExpr, DivExpr> &>' with deduced return type cannot be used before it is defined
                std::visit(visitor, *exp.lhs);
...

Is this a bug in Clang? Or am I doing something wrong, and, if so, why does this work for GCC?

Also, is there an elegant way to do what I want in Clang? Particularly, I would prefer to be able to create the printer visitor from some overloaded lambdas, rather than having to create a PrinterVisitor class with operator() overloaded for each type. However, if that's not possible, I am open to other suggestions.

Upvotes: 2

Views: 152

Answers (2)

jinscoe123
jinscoe123

Reputation: 1739

The final example in the solution posted by @sehe is almost perfect, except that the number of memory allocations increases exponentially with the depth of the expression tree. This is because the copy constructor is called 2^l + 2^r times, where l and r are the depths of the left and right expression trees, respectively, passed to the constructor of BinaryExpr.

Thus, to construct the following expression tree, I found that it would take 68 allocations.

Expr root {
    AddExpr {
        MulExpr{ AddExpr{ 1, 2 }, SubExpr{ 3, 4 } },
        DivExpr{ SubExpr{ 5, 6 }, AddExpr{ 7, 8 } },
    }
};

However, we can easily avoid these memory allocations by moving (rather than copying) the expression trees like so.

struct AddExpr;
struct SubExpr;
struct MulExpr;
struct DivExpr;

using Expr = std::variant<int, AddExpr, SubExpr, MulExpr, DivExpr>;

struct BinaryExpr
{
    BinaryExpr(Expr lhs, Expr rhs) noexcept;
    BinaryExpr(BinaryExpr&&) = default; // Require the default move-constructor.
    BinaryExpr(BinaryExpr const&);

    std::unique_ptr<Expr> lhs;
    std::unique_ptr<Expr> rhs;
};

struct AddExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };
struct SubExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };
struct MulExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };
struct DivExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };

BinaryExpr::BinaryExpr(Expr lhs, Expr rhs) noexcept
    : lhs{ std::make_unique<Expr>(std::move(lhs)) }, // Call std::move() here.
      rhs{ std::make_unique<Expr>(std::move(rhs)) }  // Call std::move() here.
{ }

BinaryExpr::BinaryExpr(BinaryExpr const& exp)
    : lhs{ std::make_unique<Expr>(*exp.lhs) },
      rhs{ std::make_unique<Expr>(*exp.rhs) }
{ }

With this modification, constructing the example expression tree, shown above, only takes 14 allocations, instead of 68.

Upvotes: 1

sehe
sehe

Reputation: 393674

Your exact code doesn't work on GCC trunk either: Compiler Explorero, which is funny since the releases did work: Compiler Explorer.

My local messages are very similar: GCC

/home/sehe/custom/boost_1_75_0/boost/hana/functional/fix.hpp|74 col 19| error: use of ‘main()::<lambda(auto:39, const auto:40&)> [with auto:39 = boost::hana::fix_t<main()::<lambda(auto:39, const auto:40&)> >; auto:40 = int]’ before deduction of ‘auto’
||    74 |         { return f(fix(f), static_cast<X&&>(x)...); }

Clang:

/home/sehe/custom/boost_1_75_0/boost/hana/functional/fix.hpp|74 col 18| error: function 'operator()<boost::hana::fix_t<(lambda at /home/sehe/Projects/stackoverflow/test.cpp:44:40)>, int>' with deduced return type cannot be used before it is defined
||         { return f(fix(f), static_cast<X&&>(x)...); }

I remember when I wrote my own combinators that I couldn't get around having an implementation helper, like:

struct F {
    template <typename Self, typename T>
    void operator()(Self const& self, T const& arg) const {
        auto bin = [&](char op, BinaryExpr const& exp) {
            std::cout << "("; std::visit(self, *exp.lhs); std::cout << op; std::visit(self, *exp.rhs); std::cout << ")"; 
        };
        hana::overload(
            [&](int val) { std::cout << val; },
            [&](AddExpr const& exp) { bin('+', exp); },
            [&](SubExpr const& exp) { bin('-', exp); },
            [&](MulExpr const& exp) { bin('*', exp); },
            [&](DivExpr const& exp) { bin('/', exp); }
        )(arg);
    }
};

See it Live On Compiler Explorer

Explanation

I think it merely requires default constructible lambdas, which are new to C++20: http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0624r2.pdf

Source: https://www.bfilipek.com/2019/03/lambdas-story-part2.html

With C++20 we’ll get the following features:

  • Allow [=, this] as a lambda capture - P0409R2 and Deprecate implicit capture of this via [=] - P0806
  • Pack expansion in lambda init-capture: ...args = std::move(args)](){} - P0780
  • static, thread_local, and lambda capture for structured bindings - P1091
  • template lambdas (also with concepts) - P0428R2
  • Simplifying implicit lambda capture - P0588R1
  • Default constructible and assignable stateless lambdas - P0624R2
  • Lambdas in unevaluated contexts - P0315R4

This leads me to believe that it can work in c++20.

Counter-Offer

I have some simplifications and my preferred pattern for recursive visitors. with some benefits like

  • encapsulating the [recursive] visitation
  • making it easier to bind extra arguments (std::cout in this example)

I find the maintainabiliity outweighs the convenience of declaring the visitor on the call site.

EDIT I found a more compelling solution so I'll just put a link to this in case you're interested: Live On Compiler Explorer

HAVE YOUR CAKE AND EAT IT, Without Boost

I realized you can have your cake and eat it, without requiring any Boost at all:

template <typename... F>
struct RecursiveVisitor : F... {
    template <typename... Ts>
    void operator()(std::variant<Ts...> const& v) const {
        std::visit( std::bind(*this, std::cref(*this), std::placeholders::_1), v );
    } 

    using F::operator()...;
};

template <typename... F>
    RecursiveVisitor(F...) -> RecursiveVisitor<F...>;

Now you can pull off your mission with e.g.:

RecursiveVisitor const printer {
    [](auto const&, int const& v) { std::cout << v; },
    [](auto const& self, AddExpr const& exp) { self(self, '+', exp); },
    [](auto const& self, SubExpr const& exp) { self(self, '-', exp); }, 
    [](auto const& self, MulExpr const& exp) { self(self, '*', exp); },
    [](auto const& self, DivExpr const& exp) { self(self, '/', exp); },
    [](auto const& self, char op, BinaryExpr const& exp) {
        std::cout << "(";
        self(*exp.lhs);
        std::cout << op;
        self(*exp.rhs);
        std::cout << ")"; 
    }
};

Which does work across the board:

Live On Compiler Explorer

#include <iostream>
#include <memory>
#include <variant>
#include <functional>

struct AddExpr;
struct SubExpr;
struct MulExpr;
struct DivExpr;

using Expr = std::variant<int, AddExpr, SubExpr, MulExpr, DivExpr>;

struct BinaryExpr
{
    BinaryExpr(Expr l, Expr r);
    BinaryExpr(BinaryExpr const& o);
    std::unique_ptr<Expr> lhs, rhs;
};

struct AddExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };
struct SubExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };
struct MulExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };
struct DivExpr : BinaryExpr { using BinaryExpr::BinaryExpr; };

BinaryExpr::BinaryExpr(Expr l, Expr r)
: lhs{ std::make_unique<Expr>(l) }, rhs{ std::make_unique<Expr>(r) } { }

BinaryExpr::BinaryExpr(BinaryExpr const& o)
: lhs{ std::make_unique<Expr>(*o.lhs) }, rhs{ std::make_unique<Expr>(*o.rhs) } { }

template <typename... F>
struct RecursiveVisitor : F... {
    template <typename... Ts>
    void operator()(std::variant<Ts...> const& v) const {
        std::visit( std::bind(*this, std::cref(*this), std::placeholders::_1), v );
    } 

    using F::operator()...;
};

template <typename... F>
    RecursiveVisitor(F...) -> RecursiveVisitor<F...>;

int main()
{
    RecursiveVisitor const printer {
        [](auto const&, int const& v) { std::cout << v; },
        [](auto const& self, AddExpr const& exp) { self(self, '+', exp); },
        [](auto const& self, SubExpr const& exp) { self(self, '-', exp); }, 
        [](auto const& self, MulExpr const& exp) { self(self, '*', exp); },
        [](auto const& self, DivExpr const& exp) { self(self, '/', exp); },
        [](auto const& self, char op, BinaryExpr const& exp) {
            std::cout << "(";
            self(*exp.lhs);
            std::cout << op;
            self(*exp.rhs);
            std::cout << ")"; 
        }
    };

    for (Expr root : { Expr
            { MulExpr{ AddExpr{ 1, 2 }, SubExpr{ 3, 4 } } },
            { DivExpr{ 1, MulExpr{ 7, SubExpr{ 3, 4 } } } },
        }) 
    {
        printer(root);
        std::cout << "\n";
    }
}

Printing

((1+2)*(3-4))
(1/(7*(3-4)))

I realize this puts duplicates the Y-combinator pattern manually, but at least it removes all bottlenecks. Also, not how the shared char, BinExpr& overload is elegantly "just another overload".

Upvotes: 1

Related Questions