RVO (let the compiler copy)

RVO (let the compiler copy)

[Henrik Mannerström]

This question came up on SO (
http://stackoverflow.com/questions/7320520/optimizing-the-number-of-constructor-calls)
and I was baffled since it went against the mantra "let the compiler do
the copying". Do you have an idea of what is going on? Why is the
commented version of operator+ worse? I compile with g++ -std=gnu++11
-Wall -Wextra -pedantic -O2  mwe.cc

BR,
Henrik

#include <iostream>

struct Imaginary {
  Imaginary(int a_, int b_) : a(a_), b(b_) {
    std::cout << "int/int ctor" << std::endl;
  };
 
  Imaginary(Imaginary const & orig) {
    a = orig.a;
    b = orig.b;
    std::cout << "Copy ctor (" << a << ',' << b << ')' << std::endl;
  }

  Imaginary & append(Imaginary const & rhs) {
    a = 10 * a + rhs.a;
    b = 10 * b + rhs.b;
    return *this;
  }

  int a;
  int b;
};

/* This calls the copy constructor once */
Imaginary operator+(Imaginary const & lhs, Imaginary const & rhs) {
  Imaginary tmp(lhs);
  tmp.append(rhs);
  return tmp;
}

/* This calls the copy constructor twice
Imaginary operator+(Imaginary lhs, Imaginary const & rhs) {
  lhs.append(rhs);
  return lhs;
}
*/

int main()
{
    Imaginary x(1, 1);
    Imaginary y(2, 1);

    Imaginary c = x + y;
    return c.a;
}

Re: RVO (let the compiler copy)

[Marc Glisse]

On Thu, 19 Dec 2013, Henrik Mannerström wrote:

> This question came up on SO (
> http://stackoverflow.com/questions/7320520/optimizing-the-number-of-constructor-calls)
> and I was baffled since it went against the mantra "let the compiler do
> the copying". Do you have an idea of what is going on? Why is the
> commented version of operator+ worse?

See:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57176

-- 
Marc Glisse

Re: RVO (let the compiler copy)

[Henrik Mannerström]

On 12/19/2013 12:41 PM, Marc Glisse wrote:
> On Thu, 19 Dec 2013, Henrik Mannerström wrote:
>
>> This question came up on SO (
>> http://stackoverflow.com/questions/7320520/optimizing-the-number-of-constructor-calls)
>>
>> and I was baffled since it went against the mantra "let the compiler do
>> the copying". Do you have an idea of what is going on? Why is the
>> commented version of operator+ worse?
>
> See:
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57176
>
Thank you!

I was completely blinded by "pass by value" propaganda
(http://cpp-next.com/archive/2009/08/want-speed-pass-by-value/). Now I
won't be so naive.