Overriding iteration
From what I understand, when you override iteration, you can either implement the basic range primitives, permitting foreach to destructively iterate over your object, or you can implement a custom method that's called, and that must perform the iteration. The destructiveness of the first option can, of course, be mitigated if you use a struct rather than a class, and make sure that anything that would be destroyed by popFront() is copied. What I'm wondering is whether there is a way to do what Python does -- to construct/return an iterator (or, in this case, a range) that is used during the iteration, rather than the object itself? I'm thinking about when you iterate directly over the object here. As far as I can see, the solution used in the std.container is to use opSlice() for this functionality. In other words, in order to iterate over container foo, you need to use foreach(e; foo[])? Is there no way to get this functionality directly (i.e., for foreach(e; foo))? -- Magnus Lie Hetland http://hetland.org
Re: Overriding iteration
Magnus Lie Hetland wrote: From what I understand, when you override iteration, you can either implement the basic range primitives, permitting foreach to destructively iterate over your object, or you can implement a custom method that's called, and that must perform the iteration. The destructiveness of the first option can, of course, be mitigated if you use a struct rather than a class, and make sure that anything that would be destroyed by popFront() is copied. What I'm wondering is whether there is a way to do what Python does -- to construct/return an iterator (or, in this case, a range) that is used during the iteration, rather than the object itself? I'm thinking about when you iterate directly over the object here. As far as I can see, the solution used in the std.container is to use opSlice() for this functionality. In other words, in order to iterate over container foo, you need to use foreach(e; foo[])? Is there no way to get this functionality directly (i.e., for foreach(e; foo))? foreach ( e; foo ) {} Should work. I believe there is a bug already filed on it not working. -- Simen
Re: Overriding iteration
On Fri, 04 Mar 2011 11:29:08 -0500, Magnus Lie Hetland wrote: From what I understand, when you override iteration, you can either implement the basic range primitives, permitting foreach to destructively iterate over your object, or you can implement a custom method that's called, and that must perform the iteration. The destructiveness of the first option can, of course, be mitigated if you use a struct rather than a class, and make sure that anything that would be destroyed by popFront() is copied. What I'm wondering is whether there is a way to do what Python does -- to construct/return an iterator (or, in this case, a range) that is used during the iteration, rather than the object itself? That's exactly how to do it. I'm thinking about when you iterate directly over the object here. As far as I can see, the solution used in the std.container is to use opSlice() for this functionality. In other words, in order to iterate over container foo, you need to use foreach(e; foo[])? Is there no way to get this functionality directly (i.e., for foreach(e; foo))? I believe someone has filed a bug for this, because TDPL has said this should be possible. But with the current compiler, you can use opApply to achieve that behavior. -Steve
Re: Overriding iteration
Simen kjaeraas wrote: Magnus Lie Hetland wrote: From what I understand, when you override iteration, you can either implement the basic range primitives, permitting foreach to destructively iterate over your object, or you can implement a custom method that's called, and that must perform the iteration. The destructiveness of the first option can, of course, be mitigated if you use a struct rather than a class, and make sure that anything that would be destroyed by popFront() is copied. What I'm wondering is whether there is a way to do what Python does -- to construct/return an iterator (or, in this case, a range) that is used during the iteration, rather than the object itself? I'm thinking about when you iterate directly over the object here. As far as I can see, the solution used in the std.container is to use opSlice() for this functionality. In other words, in order to iterate over container foo, you need to use foreach(e; foo[])? Is there no way to get this functionality directly (i.e., for foreach(e; foo))? foreach ( e; foo ) {} Should work. I believe there is a bug already filed on it not working. Found it: http://d.puremagic.com/issues/show_bug.cgi?id=5605 -- Simen
Re: Overriding iteration
On 03/04/2011 05:43 PM, Steven Schveighoffer wrote: On Fri, 04 Mar 2011 11:29:08 -0500, Magnus Lie Hetland wrote: From what I understand, when you override iteration, you can either implement the basic range primitives, permitting foreach to destructively iterate over your object, or you can implement a custom method that's called, and that must perform the iteration. The destructiveness of the first option can, of course, be mitigated if you use a struct rather than a class, and make sure that anything that would be destroyed by popFront() is copied. What I'm wondering is whether there is a way to do what Python does -- to construct/return an iterator (or, in this case, a range) that is used during the iteration, rather than the object itself? That's exactly how to do it. I'm thinking about when you iterate directly over the object here. As far as I can see, the solution used in the std.container is to use opSlice() for this functionality. In other words, in order to iterate over container foo, you need to use foreach(e; foo[])? Is there no way to get this functionality directly (i.e., for foreach(e; foo))? I believe someone has filed a bug for this, because TDPL has said this should be possible. But with the current compiler, you can use opApply to achieve that behavior. opApply should work but it is supposed to be slower. Defining range primitives directly on the object/container cannot work as of now, unfortunately, because of a pair of bugs (conflicts in formatValue template definitions between struct/class on one hand and ranges on the other). Denis -- _ vita es estrany spir.wikidot.com
Re: Overriding iteration
On Friday, March 04, 2011 09:13:34 spir wrote: > On 03/04/2011 05:43 PM, Steven Schveighoffer wrote: > > On Fri, 04 Mar 2011 11:29:08 -0500, Magnus Lie Hetland wrote: > >> From what I understand, when you override iteration, you can either > >> implement the basic range primitives, permitting foreach to > >> destructively iterate over your object, or you can implement a custom > >> method that's called, and that must perform the iteration. The > >> destructiveness of the first option can, of course, be mitigated if you > >> use a struct rather than a class, and make sure that anything that > >> would be destroyed by popFront() is copied. > >> > >> What I'm wondering is whether there is a way to do what Python does -- > >> to construct/return an iterator (or, in this case, a range) that is > >> used during the iteration, rather than the object itself? > > > > That's exactly how to do it. > > > >> I'm thinking about when you iterate directly over the object here. As > >> far as I can see, the solution used in the std.container is to use > >> opSlice() for this functionality. In other words, in order to iterate > >> over container foo, you need to use foreach(e; foo[])? Is there no way > >> to get this functionality directly (i.e., for foreach(e; foo))? > > > > I believe someone has filed a bug for this, because TDPL has said this > > should be possible. > > > > But with the current compiler, you can use opApply to achieve that > > behavior. > > opApply should work but it is supposed to be slower. > Defining range primitives directly on the object/container cannot work as > of now, unfortunately, because of a pair of bugs (conflicts in formatValue > template definitions between struct/class on one hand and ranges on the > other). You don't _want_ range primitives directly on the container. That would mean that everything in your container goes away when you process it. Every popFront() call would be removing an element from your container. So, for insteance, you try and call find() on your container and everything before what you were looking isn't in the container anymore - and if it isn't there at all, you have an empty container. You _want_ to have a separate type which is a slice of our container and has the range primitives. Now, it could very well be that foreach(v; container) should be calling opSlice on the container, allowing you to feed the container to foreach directly instead of having to slice it yourself foreach(v; container[]) but that's just syntactic sugar. You don't want to actually treat your container like a range. Ranges should be slices of containers, not containers themselves. - Jonathan M Davis
Re: Overriding iteration
On Fri, 04 Mar 2011 12:13:34 -0500, spir wrote: On 03/04/2011 05:43 PM, Steven Schveighoffer wrote: But with the current compiler, you can use opApply to achieve that behavior. opApply should work but it is supposed to be slower. It depends on the application and aggregate you are trying to iterate. If inlining is possible, ranges can be extremely fast. However, there are certain applications that are better suited or only work with opApply: * iterating polymorphic types (i.e. classes or interfaces) * iterating non-linearly (e.g. iterating a tree) * iterating multiple items in foreach (i.e. foreach(i, v; arr) ) In addition, LDC is able to inline opApply delegates that are compiler-generated, making opApply pretty much as fast as a range iteration. I hope some day dmd can do this too. Defining range primitives directly on the object/container cannot work as of now, unfortunately, because of a pair of bugs (conflicts in formatValue template definitions between struct/class on one hand and ranges on the other). It is not a good idea to define range primitives on a container. This would mean that iterating the container destroys the data. What you want is a range on the container, and to iterate that range. Think of a range as a view of the data in the container. Think of the container as the owner of the data. A confusing aspect is that builtin arrays are often thought of as containers. They are not containers, they are ranges. The owner of the data is actually the GC. -Steve
Re: Overriding iteration
On 2011-03-04 17:46:39 +0100, Simen kjaeraas said: Simen kjaeraas wrote: [snip] Found it: http://d.puremagic.com/issues/show_bug.cgi?id=5605 Oo -- nice :) (That it should work, that is; not that it doesn't ;) -- Magnus Lie Hetland http://hetland.org
Re: Overriding iteration
On 2011-03-04 19:06:34 +0100, Jonathan M Davis said: On Friday, March 04, 2011 09:13:34 spir wrote: On 03/04/2011 05:43 PM, Steven Schveighoffer wrote: [snip] opApply should work but it is supposed to be slower. Defining range primitives directly on the object/container cannot work as of now, unfortunately, because of a pair of bugs (conflicts in formatValue template definitions between struct/class on one hand and ranges on the other). You don't _want_ range primitives directly on the container. That would mean that everything in your container goes away when you process it. That was the point of my original question, yes :) In TDPL, where Andrei discusses overloading foreach, he has two main examples -- one using opApply, and one using such a self-destructive container. I think (as I said in my post) that what saves that list is that it's a struct, so it's copied by the initial assignment of the foreach statement. You _want_ to have a separate type which is a slice of our container and has the range primitives. Exactly. That was what I was asking for. Now, it could very well be that foreach(v; container) should be calling opSlice on the container, allowing you to feed the container to foreach directly instead of having to slice it yourself foreach(v; container[]) but that's just syntactic sugar. Sure. And judging from the other responses (and from TDPL), the fact that this doesn't currently work is just a bug. You don't want to actually treat your container like a range. Ranges should be slices of containers, not containers themselves. Well, it would still be nice to have things be consistent -- and in order for the opSlice approach to be consistent with the opApply approach (so a client needn't know how iteration is implemented for a given container), it seems reasonable to me to have foreach directly run on a slice of your container (i.e., implicitly calling []). But as this seems to be the way it is (save for the current bug), I guess it's sort of a moot point. I certainly agree with your point, though. In Python, too, iterators (i.e., ranges) and iterables (i.e., containers) are separate concepts. You can iterate over an iterable, and the loop then automatically extracts an iterator. As this is The Way to Go, it makes sense to me that it's automatic/implicit. -- Magnus Lie Hetland http://hetland.org
Re: Overriding iteration
On 03/04/2011 07:06 PM, Jonathan M Davis wrote: On Friday, March 04, 2011 09:13:34 spir wrote: On 03/04/2011 05:43 PM, Steven Schveighoffer wrote: On Fri, 04 Mar 2011 11:29:08 -0500, Magnus Lie Hetland wrote: From what I understand, when you override iteration, you can either implement the basic range primitives, permitting foreach to destructively iterate over your object, or you can implement a custom method that's called, and that must perform the iteration. The destructiveness of the first option can, of course, be mitigated if you use a struct rather than a class, and make sure that anything that would be destroyed by popFront() is copied. What I'm wondering is whether there is a way to do what Python does -- to construct/return an iterator (or, in this case, a range) that is used during the iteration, rather than the object itself? That's exactly how to do it. I'm thinking about when you iterate directly over the object here. As far as I can see, the solution used in the std.container is to use opSlice() for this functionality. In other words, in order to iterate over container foo, you need to use foreach(e; foo[])? Is there no way to get this functionality directly (i.e., for foreach(e; foo))? I believe someone has filed a bug for this, because TDPL has said this should be possible. But with the current compiler, you can use opApply to achieve that behavior. opApply should work but it is supposed to be slower. Defining range primitives directly on the object/container cannot work as of now, unfortunately, because of a pair of bugs (conflicts in formatValue template definitions between struct/class on one hand and ranges on the other). You don't _want_ range primitives directly on the container. That would mean that everything in your container goes away when you process it. Every popFront() call would be removing an element from your container. So, for insteance, you try and call find() on your container and everything before what you were looking isn't in the container anymore - and if it isn't there at all, you have an empty container. You _want_ to have a separate type which is a slice of our container and has the range primitives. Certainly, as long as, on an array-like container, you implement popFront as this = this[1..$]; or this.elements = this.elements[1..$]; then, yes, iterating on it shrinks it. (Note this works only on array-like containers; how would you shrink a tree?) But I prefere using a private index an have popFront do: ++ this.index; This is a more general iteration mechanism solution, based on current state of the object beeing iterated on. For many kinds of sequences, you needs state anyway. How else iterate over the sequence of multiples of 3, or squares of natural numbers? Now, it could very well be that foreach(v; container) should be calling opSlice on the container, allowing you to feed the container to foreach directly instead of having to slice it yourself foreach(v; container[]) but that's just syntactic sugar. You don't want to actually treat your container like a range. Ranges should be slices of containers, not containers themselves. I agree slices should be an alternate iteration mechanism (as said in TDPL). But one cannot slice a tree that easily :-) (where's my chain saw?) Denis -- _ vita es estrany spir.wikidot.com
Re: Overriding iteration
On Friday, March 04, 2011 13:30:39 spir wrote: > On 03/04/2011 07:06 PM, Jonathan M Davis wrote: > > On Friday, March 04, 2011 09:13:34 spir wrote: > >> On 03/04/2011 05:43 PM, Steven Schveighoffer wrote: > >>> On Fri, 04 Mar 2011 11:29:08 -0500, Magnus Lie > >>> Hetland > > > > wrote: > From what I understand, when you override iteration, you can either > > implement the basic range primitives, permitting foreach to > destructively iterate over your object, or you can implement a custom > method that's called, and that must perform the iteration. The > destructiveness of the first option can, of course, be mitigated if > you use a struct rather than a class, and make sure that anything > that would be destroyed by popFront() is copied. > > What I'm wondering is whether there is a way to do what Python does -- > to construct/return an iterator (or, in this case, a range) that is > used during the iteration, rather than the object itself? > >>> > >>> That's exactly how to do it. > >>> > I'm thinking about when you iterate directly over the object here. As > far as I can see, the solution used in the std.container is to use > opSlice() for this functionality. In other words, in order to iterate > over container foo, you need to use foreach(e; foo[])? Is there no way > to get this functionality directly (i.e., for foreach(e; foo))? > >>> > >>> I believe someone has filed a bug for this, because TDPL has said this > >>> should be possible. > >>> > >>> But with the current compiler, you can use opApply to achieve that > >>> behavior. > >> > >> opApply should work but it is supposed to be slower. > >> Defining range primitives directly on the object/container cannot work > >> as of now, unfortunately, because of a pair of bugs (conflicts in > >> formatValue template definitions between struct/class on one hand and > >> ranges on the other). > > > > You don't _want_ range primitives directly on the container. That would > > mean that everything in your container goes away when you process it. > > Every popFront() call would be removing an element from your container. > > So, for insteance, you try and call find() on your container and > > everything before what you were looking isn't in the container anymore - > > and if it isn't there at all, you have an empty container. You _want_ to > > have a separate type which is a slice of our container and has the range > > primitives. > > Certainly, as long as, on an array-like container, you implement popFront > as this = this[1..$]; > or > this.elements = this.elements[1..$]; > then, yes, iterating on it shrinks it. (Note this works only on array-like > containers; how would you shrink a tree?) But I prefere using a private > index an have popFront do: > ++ this.index; > > This is a more general iteration mechanism solution, based on current state > of the object beeing iterated on. For many kinds of sequences, you needs > state anyway. How else iterate over the sequence of multiples of 3, or > squares of natural numbers? > > > Now, it could very well be that > > > > foreach(v; container) > > > > should be calling opSlice on the container, allowing you to feed the > > container to foreach directly instead of having to slice it yourself > > > > foreach(v; container[]) > > > > but that's just syntactic sugar. You don't want to actually treat your > > container like a range. Ranges should be slices of containers, not > > containers themselves. > > I agree slices should be an alternate iteration mechanism (as said in > TDPL). But one cannot slice a tree that easily :-) (where's my chain saw?) ??? Go take a look RedBlackTree. It's sliced. And it's based on the current state of the container. _All_ slices are based on the current state of the container. They're a view into that container. You need to not think too much about arrays when thinking about slices. They're not normal. As Steve likes to point out, they're not really containers. They're slices. It's more like the GC is their container. So, arrays are actually a pretty bad example when containers and ranges. And ranges _are_ a general iteration mechanism. I don't know what you would think is not general about that. - Jonathan M Davis