Re: Optimising assignment using std::swap

From: Dean Sturtevant (deansturtevant_at_[hidden])
Date: 2001-01-21 21:27:24

I find it hard to imagine a case where std::swap is more efficient 
than assignment, unless the assignment is coded inefficiently. Can 
you provide an example where that might not be the case?

--- In boost_at_[hidden], "Paul Moore" <gustav_at_m...> wrote:
> I have a small optimisation issue - in the rational class, the gcd 
> operation (one of the key hotspots) repeatedly executes
> 
>     IntType r(n%m);
>     n = m;
>     m = r;
> 
> This could be replaced with
> 
>     IntType r(n%m);
>     std::swap(n,m)
>     std::swap(m,r);
> 
> For built-in types, this is notably less efficient, as it will do 6 
> assignments as opposed to 2. However, for a user-defined type 
> which may have a highly optimised swap, it could be far more 
> efficient.
> 
> Is there any way of using the most efficient version in each case - 
> or alternatively would it make sense to use the swap version in all 
> cases? (I feel that this is a bad case of premature optimisation, 
as 
> we don't yet have a common candidate for an unlimited-precision 
> integer type to use with rational<>)
> 
> [[ As usual, this could be handled by explicit specialisation - but 
I 
> don't feel that is appropriate here. ]]
> 
> This came to me when I was writing a section for the 
> documentation covering performance issues. If I don't use swap(), I 
> feel that it might be worth pointing this out in this section.
> 
> BTW, I suspect this means that rational<> should provide a swap() 
> specialisation, to delegate to (potential) swap() specialisations 
for 
> IntType. Can anyone remind me what the canonical form of swap() 
> is?
> 
> Paul