From: Sean Parent (sparent_at_[hidden])
Date: 2006-07-14 14:30:45
> I think I have a basic understanding of your logic, but let's think
> practicality for a moment. I'm not saying logical and philosophical
> are not appropriate for this discussion, but let's try and not
> think about
> them for a second.
I haven't heard mathematics described as "philosophical reasons" but
> The ordering defined by the operator< overload for shared_ptr has the
> following important property: it is independent of shared_ptr's
> parameter. You can order shared_ptr<void> objects, if you like. Or, a
> shared_ptr<void> and a shared_ptr<int>.
I understand, total ordering is important and useful. The question
is, what are the semantics of operator < ? If we fail to establish
such principals then generic programming is a futile exercise. A
generic algorithm is a theorem - it's correctness depends on a
collection of semantic requirements (Concepts), these are the axioms
for the theorem.
Does it matter if share_ptr provides an operator <? No, but it isn't
useful for generic code unless such an operation is tied to semantics
which are generally accepted. T* does not provide such an operator <
and I see no semantic difference between the two. The two should be
consistent - I don't really much care which way we go but there are
important semantics at play here for operator <. Specifically we're
pinning down whether operator < should or should not guarantee a
stable ordering across executions of the application.
We can say that the C and C++ standard got the rules for T* wrong and
lobby to change them, or that shared_ptr<> shouldn't have a less than
operator. We might even still provide it under the claim that it
shouldn't be defined but is so that std::less<> will work
consistently with T* (assuming you don't want to just specialize
std::less<> - which I think would be a better solution).
Since we can't reasonably check semantic requirements, it is
important for code correctness that we associate the requirements
with the name of the operation. The signature is the only thing I can
> Can you think of any problems that operator< could cause, in its
> definition? In my mind, since you can't use T in your ordering
> you have no reasons to define your own ordering for shared_ptr.
> you have no other use for operator<, and therefore if the one
> defined by the
> library causes no problems, we should keep it because it is practical.
Not any realistic one - in general providing such an operator only
breaks concept checks (when we are able to check if a function
exists). But without consistent semantics it also doesn't allow me to
develop any useful algorithms which rely on this.
> > Christoph Ludwig posted the correct reason why std::complex should
> > not have an operator < () defined - sorry, my follow up using
> > addition was poorly constructed.
> > What I am proposing is that std::less should be defined for
> > std::complex.
> Disagree. That could potentially break other template programming
> using std::less, to support std::unary_negate etc. Suddenly this would
> "work" for std::complex, when it shouldn't.
Allowing code which otherwise would not be allowed is hardly breaking
anything. I can see the argument that std::less<> should also be
consistent with other operators. I think it is then important to
define a general total order operation for this case and develop the
rules. Please propose a set of rules for when operator<() should and
should not be defined.
> > Why would we have a separate default relation for set/map then
> > std::less? std::less is intended to map to operator <() on the type
> > _or_ to a total ordering for the type if operator <() is not
> > defined.
> No, that's just for pointers. 20.5.7/8 says:
> "For templates greater, less, greater_equal, and less_equal, the
> specializations for any pointer type yield a total
> order, even if the built-in operators <, >, <=, >= do not."
> It doesn't mention std::complex or any other composite types.
That still doesn't answer the question of why have another relation
for set/map then std::less? And as for the intent for STL I can
assure you that the intent of the designer was std::less was intended
to provide a total ordering on all types. That standard also does not
prohibit us from specializing std::less for use defined types.
> > For pair<shared_ptr<>, int> - what makes shared_ptr special?
> The fact that it is a pointer? I don't care much if it doesn't have an
Then it shouldn't be different from T*.
> > std::less is supposed to be operator <, unless operator < is not
> > defined in which case std::less provides a total ordering. It isn't
> > ambiguous.
> Where does the standard say that?
I have the advantage of having Alex Stepanov on my team. I asked him.
> If you want to have an ordered collection of otherwise unordered data,
> the relation must be an application specific attribute. So define an
> appropriate functor, and use that with std::map, etc.
If we want to introduce a third (or more) way to define a standard
ordering other than operator <, std::less I'm fine with that. Just
specify the semantics of them.
---- > That is also how I interpret his intent based on what is currently > in the > standard, and that is why I maintain that *under the current > standard*, the > appropriate way to make a type ordered (even if only for the > associatives) > is via defining operator<. If we do this for shared_ptr it should be with a clear statement that we expect the rules for T* should be changed (or consider T* to be a legacy exception which cannot be changed). > > Why would we have a separate default relation for set/map then > > std::less? > > and: > > > As for what I expect std::greater<> et al. to do - I had that in my > > earlier 6 point note - they _better_ be defined consistently with > > std::less. > > If the standard was written with this in mind, greater<>, > less_equal<> and > greater_equal<> would have been defined in terms of less<>. It is not > reasonable to provide four customization points that the user is > supposed to > keep in sync. For me the fact that less_equal returns x <= y is strong > evidence that it is not supposed to be a customization point, but a > function > object alias of operator<=. You can make the same arguments that operator < is not intended to be a customization point because the compiler doesn't automatically provide a consistent >, >=, <= (and == and !=). The standard is a legacy that must be dealt with - this should be done by providing a consistent set of rules and then lobbying for how the standard should change to be compliant. Ideally the rules should be as consistent with the standard as possible, but the important thing is that we are always striving for a correct solution. > > For pair<shared_ptr<>, int> - what makes shared_ptr special? > > Why should it be special? If a type K is ordered, pair<K, int> > should also > be. That isn't true for T*. I contend that either both should have operator < () defined or neither. I believe for shared_ptr is is probably best to try to be consistent with T* then to lobby to get T* changed. Note that this likely still means that you want to define std::less on pair<> to be less of each of the elements. Either way, the standard is flawed. Sean
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk