March 18, 2008
Hey all.
As I might have mentioned a while ago, I have a config library that I
wrapped in D classes a while ago.
While it works perfectly fine, it's currently running somewhat
inefficiently and I wish to fix this.

The problem: The library currently contains two classes. A root class and
a node class. To load/create a config, one would create a root class and
tie it to a file. Each key=value pair in the config can be accessed by
passing the value's "path" to the root object, which will return a node.
Herein lies the issue.
Nodes are very short lived objects, and are often created and discarded
very quickly. It would make far more sense if a node was a struct, as
classes being uses like this tend to needlessly waste memory. The problem
is I need to create the node in the root object (or in a parent node) and
pass it back to the caller which would be more efficient to do with
classes. In C++, what I would do is pass the back argument to the
struct's constructor, which would create the struct and pass it back
rather efficiently, and the node would be destroyed when scope is left. I
doubt this is possible in D since structs don't have constructors.
So what should I do?

Another issue: Nodes can contain different data types. As such, I need to
access their contents using an overloaded function 'getValue' which will
use the passed parameter to determine what type of data the node holds
and then pass it back. Yes, this sucks. Since DMD doesn't have implicit,
multiple opCast overloads it appears that the only other solution is to
have a different node type for each type of data. I guess this can be
done with templates but nodes are created at runtime, not compile time.
Different node types also mean I need to know what type of node the root
config will return and I can't always do that since configs are user
editable.
Is there anything I can do about this?


After writing this post out and looking at my situation I kind of get the feeling I'm just going to have to grit my teeth and bear it for now because it looks like the only workarounds are probably going to be elaborate hacks. But hey, if you can think of something I'm all ears.

Thanks,
Jarrod.

PS pleaseeee add multiple implicit opCasts and struct constructors to D soon, Walter. It would be really helpful, especially for making and using abstract data types.

PSS It's kind of late so I'm sorry if this post has any grammatical errors.
March 18, 2008
"Jarrod" <qwerty@ytre.wq> wrote in message news:frobsv$960$1@digitalmars.com...
> Hey all.
> As I might have mentioned a while ago, I have a config library that I
> wrapped in D classes a while ago.
> While it works perfectly fine, it's currently running somewhat
> inefficiently and I wish to fix this.
>
> The problem: The library currently contains two classes. A root class and
> a node class. To load/create a config, one would create a root class and
> tie it to a file. Each key=value pair in the config can be accessed by
> passing the value's "path" to the root object, which will return a node.
> Herein lies the issue.
> Nodes are very short lived objects, and are often created and discarded
> very quickly. It would make far more sense if a node was a struct, as
> classes being uses like this tend to needlessly waste memory. The problem
> is I need to create the node in the root object (or in a parent node) and
> pass it back to the caller which would be more efficient to do with
> classes. In C++, what I would do is pass the back argument to the
> struct's constructor, which would create the struct and pass it back
> rather efficiently, and the node would be destroyed when scope is left. I
> doubt this is possible in D since structs don't have constructors.
> So what should I do?

NRVO (Named Return Value Optimization) to the rescue.  There are no struct constructors, but it doesn't mean the compiler can't optimize this out. When you return a struct from a function, chances are very good that it will not be pushed onto the stack, but will be constructed in place.  So even though it looks like you're allocating (on the stack) and filling out a struct in the function, the compiler will rewrite the function so that a pointer to the struct is passed as a parameter and the function fills out the struct in place.

> Another issue: Nodes can contain different data types. As such, I need to
> access their contents using an overloaded function 'getValue' which will
> use the passed parameter to determine what type of data the node holds
> and then pass it back. Yes, this sucks. Since DMD doesn't have implicit,
> multiple opCast overloads it appears that the only other solution is to
> have a different node type for each type of data. I guess this can be
> done with templates but nodes are created at runtime, not compile time.
> Different node types also mean I need to know what type of node the root
> config will return and I can't always do that since configs are user
> editable.
> Is there anything I can do about this?

Well, what would you have to do in C++?  So you have multiple implicit cast overloads, but how is the node represented?  As a tagged union?  As a string which is converted to the correct type upon use?  Is there a member in the node which represents its type?  Don't you still need to check that the node is of the correct type before getting its value anyway, by which time you already know the type and the implicit cast is next to useless?

It sounds like your node is wanting to be a variant.  In D1, there's a variant type in std.boxer called Box, and in D2, there's std.variant.  (In Tango, there's tango.core.Variant.)  Or, if you want to implement it yourself, it's not that hard, it's basically a tagged union with some methods to check the type and convert it to a desired value.


March 18, 2008
"Jarrod" wrote
> Hey all.
> As I might have mentioned a while ago, I have a config library that I
> wrapped in D classes a while ago.
> While it works perfectly fine, it's currently running somewhat
> inefficiently and I wish to fix this.
>
> The problem: The library currently contains two classes. A root class and
> a node class. To load/create a config, one would create a root class and
> tie it to a file. Each key=value pair in the config can be accessed by
> passing the value's "path" to the root object, which will return a node.
> Herein lies the issue.
> Nodes are very short lived objects, and are often created and discarded
> very quickly. It would make far more sense if a node was a struct, as
> classes being uses like this tend to needlessly waste memory. The problem
> is I need to create the node in the root object (or in a parent node) and
> pass it back to the caller which would be more efficient to do with
> classes. In C++, what I would do is pass the back argument to the
> struct's constructor, which would create the struct and pass it back
> rather efficiently, and the node would be destroyed when scope is left. I
> doubt this is possible in D since structs don't have constructors.
> So what should I do?

Have you tried static opCall?  It's sort of like a struct constructor:

struct S
{
  int val;
  static S opCall(int arg)
  {
      S retval;
      retval.val = arg;
      return retval;
  }
}

S func()
{
    return S(1);
}

-Steve


March 18, 2008
On Tue, 18 Mar 2008 09:05:19 -0400, Jarrett Billingsley wrote:

> NRVO (Named Return Value Optimization) to the rescue.  There are no struct constructors, but it doesn't mean the compiler can't optimize this out. When you return a struct from a function, chances are very good that it will not be pushed onto the stack, but will be constructed in place.  So even though it looks like you're allocating (on the stack) and filling out a struct in the function, the compiler will rewrite the function so that a pointer to the struct is passed as a parameter and the function fills out the struct in place.

This is good to hear. I guess I won't feel (as) guilty about having to pass back a struct then.


> Well, what would you have to do in C++?  So you have multiple implicit cast overloads, but how is the node represented?  As a tagged union?  As a string which is converted to the correct type upon use?  Is there a member in the node which represents its type?  Don't you still need to check that the node is of the correct type before getting its value anyway, by which time you already know the type and the implicit cast is next to useless?
> 
> It sounds like your node is wanting to be a variant.  In D1, there's a variant type in std.boxer called Box, and in D2, there's std.variant. (In Tango, there's tango.core.Variant.)  Or, if you want to implement it yourself, it's not that hard, it's basically a tagged union with some methods to check the type and convert it to a desired value.

Well yeah the node acts much like a variant. And yes, it internally uses a union with an enum to keep track of the stored value type. It tries its best to convert the stored value to the desired type upon being called. My problem isn't with storing or handling the abstract data, it's that I need to use an overloaded function call to get the value. It's quite ugly and results in a lot of temporary variables being created when they don't need to be created, like for instance if I wanted to pass a number to a function:
>int temp;
>myNode.getValue(temp);
>someFunc(temp);

In C++ with overloaded return types I can just use:
>someFunc(myNode); //Sometimes a cast is needed, but it's still better

Okay, it's not exactly a show stopper, it's just a bit ugly and somewhat
inconvenient.
With that said, it looks like boxer/variant have workarounds for this
using templates. I'll look at their source and see how it's done.

Thanks for your reply, it was quite helpful.
March 18, 2008
On Tue, 18 Mar 2008 12:00:29 -0400, Steven Schveighoffer wrote:

> Have you tried static opCall?  It's sort of like a struct constructor:
> 
> struct S
> {
>   int val;
>   static S opCall(int arg)
>   {
>       S retval;
>       retval.val = arg;
>       return retval;
>   }
> }
> 
> S func()
> {
>     return S(1);
> }
> 
> -Steve

Hey, that is actually really handy. I'll have to remember it. Thanks Steve.
March 18, 2008
"Jarrod" <qwerty@ytre.wq> wrote in message news:frpi9g$1avh$1@digitalmars.com...

>
> Well yeah the node acts much like a variant. And yes, it internally uses
> a union with an enum to keep track of the stored value type.
> It tries its best to convert the stored value to the desired type upon
> being called. My problem isn't with storing or handling the abstract
> data, it's that I need to use an overloaded function call to get the
> value. It's quite ugly and results in a lot of temporary variables being
> created when they don't need to be created, like for instance if I wanted
> to pass a number to a function:
>>int temp;
>>myNode.getValue(temp);
>>someFunc(temp);

This is precisely where a templated method would come in, as you've noticed:

someFunc(myNode.to!(int));

Sure, you don't get the cast syntax, but it's actually less typing and doesn't require fancy languages support.

Implicit casting in this case would be nice though.  D2 is apparently getting overloadable opImplicitCast at some nebulous time in the future.


Top | Discussion index | About this forum | D home