Pass-by-reference hinders gcc from tail call elimination

问题

See BlendingTable::create and BlendingTable::print. Both have the same form of tail recursion, but while create will be optimized as a loop, print will not and cause a stack overflow.

Go down to see a fix, which I got from a hint from one of the gcc devs on my bug report of this problem.

#include <cstdlib>
#include <iostream>
#include <memory>
#include <array>
#include <limits>

class System {
public:
    template<typename T, typename... Ts>
    static void print(const T& t, const Ts&... ts) {
        std::cout << t << std::flush;
        print(ts...);
    }

    static void print() {}

    template<typename... Ts>
    static void printLine(const Ts&... ts) {
        print(ts..., '\n');
    }
};

template<typename T, int dimension = 1>
class Array {
private:
    std::unique_ptr<T[]> pointer;
    std::array<int, dimension> sizes;
    int realSize;

public:
    Array() {}

    template<typename... Ns>
    Array(Ns... ns):
    realSize(1) {
        checkArguments(ns...);
        create(1, ns...);
    }

private:
    template<typename... Ns>
    static void checkArguments(Ns...) {
        static_assert(sizeof...(Ns) == dimension, "dimension mismatch");
    }

    template<typename... Ns>
    void create(int d, int n, Ns... ns) {
        realSize *= n;
        sizes[d - 1] = n;
        create(d + 1, ns...);
    }

    void create(int) {
        pointer = std::unique_ptr<T[]>(new T[realSize]);
    }

    int computeSubSize(int d) const {
        if (d == dimension) {
            return 1;
        }
        return sizes[d] * computeSubSize(d + 1);
    }

    template<typename... Ns>
    int getIndex(int d, int n, Ns... ns) const {
        return n * computeSubSize(d) + getIndex(d + 1, ns...);
    }

    int getIndex(int) const {
        return 0;
    }

public:
    template<typename... Ns>
    T& operator()(Ns... ns) const {
        checkArguments(ns...);
        return pointer[getIndex(1, ns...)];
    }

    int getSize(int d = 1) const {
        return sizes[d - 1];
    }
};

class BlendingTable : public Array<unsigned char, 3> {
private:
    enum {
        SIZE = 0x100,
        FF = SIZE - 1,
    };

public:
    BlendingTable():
    Array<unsigned char, 3>(SIZE, SIZE, SIZE) {
        static_assert(std::numeric_limits<unsigned char>::max() == FF, "unsupported byte format");
        create(FF, FF, FF);
    }

private:
    void create(int dst, int src, int a) {
        (*this)(dst, src, a) = (src * a + dst * (FF - a)) / FF;
        if (a > 0) {
            create(dst, src, a - 1);
        } else if (src > 0) {
            create(dst, src - 1, FF);
        } else if (dst > 0) {
            create(dst - 1, FF, FF);
        } else {
            return;
        }
    }

    void print(int dst, int src, int a) const {
        System::print(static_cast<int>((*this)(FF - dst, FF - src, FF - a)), ' ');
        if (a > 0) {
            print(dst, src, a - 1);
        } else if (src > 0) {
            print(dst, src - 1, FF);
        } else if (dst > 0) {
            print(dst - 1, FF, FF);
        } else {
            System::printLine();
            return;
        }
    }

public:
    void print() const {
        print(FF, FF, FF);
    }
};

int main() {
    BlendingTable().print();
    return EXIT_SUCCESS;
}

Changing the class definition of System from

class System {
public:
    template<typename T, typename... Ts>
    static void print(const T& t, const Ts&... ts) {
        std::cout << t << std::flush;
        print(ts...);
    }

    static void print() {}

    template<typename... Ts>
    static void printLine(const Ts&... ts) {
        print(ts..., '\n');
    }
};

to

class System {
public:
    template<typename T, typename... Ts>
    static void print(T t, Ts... ts) {
        std::cout << t << std::flush;
        print(ts...);
    }

    static void print() {}

    template<typename... Ts>
    static void printLine(Ts... ts) {
        print(ts..., '\n');
    }
};

magically allows gcc to eliminate the tail calls.

Why does 'whether or not passing function arguments by reference' make such a big difference in gcc's behaviour? Semantically they both look the same to me in this case.

回答1:

As it is noted by @jxh the cast static_cast<int>() creates a temporary whose reference is passed to the print function. Without such cast the tail recursion is optimized correctly.

The issue is very similar to the old case Why isn't g++ tail call optimizing while gcc is? and the workaround may be similar to https://stackoverflow.com/a/31793391/4023446.

It is still possible to use System with the arguments passed by reference if call to System::print will be moved to a separate private helper function SystemPrint:

class BlendingTable : public Array<unsigned char, 3> {

//...

private:
    void SystemPrint(int dst, int src, int a) const
    {
        System::print(static_cast<int>((*this)(FF - dst, FF - src, FF - a)), ' ');
    }

    void print(int dst, int src, int a) const {
        SystemPrint(dst, src, a);
        if (a > 0) {
            print(dst, src, a - 1);
        } else if (src > 0) {
            print(dst, src - 1, FF);
        } else if (dst > 0) {
            print(dst - 1, FF, FF);
        } else {
            System::printLine();
            return;
        }
    }

// ...

}

Now the tail call optimization works (g++ (Ubuntu/Linaro 4.7.2-2ubuntu1) 4.7.2 with the optimization option -O2) and the print does not cause a stack overflow.

Update

I verified it with other compilers:

• the original code without any change is perfectly optimized by clang++ Apple LLVM version 5.1 (clang-503.0.40) (based on LLVM 3.4svn) with -O1 optimization
• g++ (Ubuntu 4.8.4-2ubuntu1~14.04) 4.8.4 fails to perform TCO even without the cast or with the wrapping function SystemPrint workaround; here only the workaround with System::print arguments by values works.

So, the issue is very specific to compiler versions.

来源：https://stackoverflow.com/questions/31279447/pass-by-reference-hinders-gcc-from-tail-call-elimination