[boost] Re: Constructor of iterator_adaptor takes parameter by value

From: Daniel Frey (daniel.frey_at_[hidden])
Date: 2004-01-12 04:46:40

Howard Hinnant wrote:
> On Jan 9, 2004, at 9:27 PM, David Abrahams wrote:
> 
>> "Peter Dimov" <pdimov_at_[hidden]> writes:
>>
>>> Executive summary: the reference wins. It doesn't really matter. You 
>>> knew
>>> that already.
>>
>> Yep, my tests show the same.  That's a relief to me, actually.  I
>> wonder what Andrei was on about?
> 
> Andrei was making a point that Peter's test does not elucidate.
> 
> If you must modify the argument internally, then it is better to pass 
> the argument by value and modify the argument directly, rather than pass 
> by const reference, and make a copy of the argument which you can then 
> modify:

Consider:

struct X {
   X() { cout << "X()\n"; }
   X( const X& ) { cout << "X(X)\n"; }
   X& operator+=( const X& ) { cout << "+=\n"; }
};

X operator+( const X& lhs, const X& rhs ) {
   X nrv( lhs );
   nrv += rhs;
   return nrv;
}

X operator-( X lhs, const X& rhs ) {
   lhs += rhs;
   return lhs;
}

int main() {
   {
     X a,b,c;
     cout << "--\n";
     a + b + c;
     cout << "--\n";
     a - b - c;
     cout << "--\n";
     a + ( b + c );
     cout << "--\n";
     a - ( b - c );
     cout << "--\n";
   }
}

Try this on a compiler that implements the NRVO. (GCC3.1+, Intel7+ 
(->EDG), MSVC8+). Conclusion: Even when you need a copy of the argument, 
the const reference can save you some copies. See also 
<http://tinyurl.com/yvabj>. IIRC, Andrei's point was that not all 
compilers implement the NRVO, but the situation is getting better and 
better... :)

Regards, Daniel

-- 
Daniel Frey
aixigo AG - financial solutions & technology
Schloß-Rahe-Straße 15, 52072 Aachen, Germany
fon: [phone], fax: [phone]
eMail: daniel.frey_at_[hidden], web: http://www.aixigo.de