Specialising when swap / iter_swap exists

[chris <caj at cs dot york dot ac dot uk>]

Some algorithms in libstdc++, such as reverse, use only iter_swap when 
operating on elements.

However, some other algorithms, including rotate, sort, etc. don't use 
swap but instead use assignment to move variables around.

Changing all these to swap would be more efficent in for those cases 
where an efficent swap or iter_swap exists, but using iter_swap instead 
of assignment where there is no efficent iter_swap (which includes the 
cases where the iterator's value_type is something like char or int) 
then (as two tests) sort and rotate appear to take about twice as long.

I have 4 broad ideas on what to do about this...

1) Use iter_swap except on a fixed set of types (int,char,etc.)
2) Only use iter_swap on a fixed set of types (vector<T>, list<T>, etc.)
3) Consider introducing a general method for marking if a type has an 
efficent type
4) Introduce a new templated function:

template<T>
void move(T& a,T& b); with a condition something like "assigns b the 
value of a, leaves a with a valid but undefined value".

Where the normal definition of move would be simply "b=a", but for types 
where swapping is more efficent than assignment use "swap(a,b)". This 
would fit in more naturally with the existing definitions of the 
algorithm, but would be very slightly less efficent than just using swap 
in the algorithms.

I suspect that someone has already thought about this kind of thing, but 
have been unable to find any references to it. I am interested on any 
thoughts on this, because whatever decision (if any) is made, it will 
require some quite substansal edits to the algorithms..

Thank you,

Chris