Hi all.

I'm currently playing around with template metaprogramming in D. I've encountered a problem with implicit instantiation of templates that I don't really understand, and I thought somebody here could shed some light on it. Consider the following code:

template MyType(T)
{
	alias T MyType;
}

void mytest(T)(MyType!(T) x)
{
}

void main()
{
	MyType!(int) x;
	mytest(x);
}

This does not work. I would expect the template mechanism to recognize that the argument to mytest is an instance of MyType!(int) and hence instantiate T with int, but instead I get the errors
template mytest(T) does not match any template declaration
template mytest(T) cannot deduce template function from argument types (int)

I've also tried replacing alias by typedef, with the same results. Instantiating the template explicitly seems to work, but I would first like to know *why* I can't do it like this before I use any workarounds. Any pointers? Thanks in advance.

Oh, and this is DMD v1.029.

Regards,
Clemens.

Clemens Hofreither wrote:
> Hi all.
> 
> I'm currently playing around with template metaprogramming in D. I've encountered a problem with implicit instantiation of templates that I don't really understand, and I thought somebody here could shed some light on it. Consider the following code:
> 
> template MyType(T)
> {
> 	alias T MyType;
> }
> 
> void mytest(T)(MyType!(T) x)
> {
> }
> 
> void main()
> {
> 	MyType!(int) x;
> 	mytest(x);
> }
> 
> This does not work. I would expect the template mechanism to recognize that the argument to mytest is an instance of MyType!(int) and hence instantiate T with int, but instead I get the errors
> template mytest(T) does not match any template declaration
> template mytest(T) cannot deduce template function from argument types (int)
> 
> I've also tried replacing alias by typedef, with the same results. Instantiating the template explicitly seems to work, but I would first like to know *why* I can't do it like this before I use any workarounds. Any pointers? Thanks in advance.
> 
> Oh, and this is DMD v1.029.

It's just not that smart.  It's a specialization so you have to use specialization syntax.  Something like:

void mytest(T : MyType!(S), S) {}

But note that using specialization also disables IFTI.
In D2 you should be able to use a constraint to do something like

void mytest(T) if(is(T : MyType!(S))) {}

My syntax is probably all wrong here, but that's the idea anyway.


Here's a bug marked fixed with a similar goal to yours:
    http://d.puremagic.com/issues/show_bug.cgi?id=1661
The discussion may be useful.

--bb

"Bill Baxter" <dnewsgroup@billbaxter.com> wrote in message news:g4s05a$2l0l$1@digitalmars.com...
> Clemens Hofreither wrote:
>> Hi all.
>>
>> I'm currently playing around with template metaprogramming in D. I've encountered a problem with implicit instantiation of templates that I don't really understand, and I thought somebody here could shed some light on it. Consider the following code:
>>
>> template MyType(T)
>> {
>> alias T MyType;
>> }
>>
>> void mytest(T)(MyType!(T) x)
>> {
>> }
>>
>> void main()
>> {
>> MyType!(int) x;
>> mytest(x);
>> }
>>
>> This does not work. I would expect the template mechanism to recognize
>> that the argument to mytest is an instance of MyType!(int) and hence
>> instantiate T with int, but instead I get the errors
>> template mytest(T) does not match any template declaration
>> template mytest(T) cannot deduce template function from argument types
>> (int)
>>
>> I've also tried replacing alias by typedef, with the same results. Instantiating the template explicitly seems to work, but I would first like to know *why* I can't do it like this before I use any workarounds. Any pointers? Thanks in advance.
>>
>> Oh, and this is DMD v1.029.
>
> It's just not that smart.  It's a specialization so you have to use specialization syntax.  Something like:
>
> void mytest(T : MyType!(S), S) {}
>
> But note that using specialization also disables IFTI.
> In D2 you should be able to use a constraint to do something like
>
> void mytest(T) if(is(T : MyType!(S))) {}
>
> My syntax is probably all wrong here, but that's the idea anyway.
>
>
> Here's a bug marked fixed with a similar goal to yours:
>     http://d.puremagic.com/issues/show_bug.cgi?id=1661
> The discussion may be useful.

I have an open enhancement request on this:

http://d.puremagic.com/issues/show_bug.cgi?id=1653

I think the main issue is really that there is no 'reverse' path from the alias to the parameter to a specific template.  That is given the type MyType!(T), which is really an alias for T, What template should the compiler use to deduce the argument?  What if there were multiple templates that aliased to T, should those also be assumed to be MyType!(T)?  What if you pass in just a T?  Technically it's the same thing.

I think these are the issues that makes it tough to do the IFTI.  I think we are spoiled because things like:

void f(T)(T[] x)

but that's a much simpler case to solve.

-Steve

Steven Schveighoffer Wrote:

> 
> "Bill Baxter" <dnewsgroup@billbaxter.com> wrote in message news:g4s05a$2l0l$1@digitalmars.com...
> > Clemens Hofreither wrote:
> >> Hi all.
> >>
> >> I'm currently playing around with template metaprogramming in D. I've encountered a problem with implicit instantiation of templates that I don't really understand, and I thought somebody here could shed some light on it. Consider the following code:
> >>
> >> template MyType(T)
> >> {
> >> alias T MyType;
> >> }
> >>
> >> void mytest(T)(MyType!(T) x)
> >> {
> >> }
> >>
> >> void main()
> >> {
> >> MyType!(int) x;
> >> mytest(x);
> >> }
> >>
> >> This does not work. I would expect the template mechanism to recognize
> >> that the argument to mytest is an instance of MyType!(int) and hence
> >> instantiate T with int, but instead I get the errors
> >> template mytest(T) does not match any template declaration
> >> template mytest(T) cannot deduce template function from argument types
> >> (int)
> >>
> >> I've also tried replacing alias by typedef, with the same results. Instantiating the template explicitly seems to work, but I would first like to know *why* I can't do it like this before I use any workarounds. Any pointers? Thanks in advance.
> >>
> >> Oh, and this is DMD v1.029.
> >
> > It's just not that smart.  It's a specialization so you have to use specialization syntax.  Something like:
> >
> > void mytest(T : MyType!(S), S) {}
> >
> > But note that using specialization also disables IFTI.
> > In D2 you should be able to use a constraint to do something like
> >
> > void mytest(T) if(is(T : MyType!(S))) {}
> >
> > My syntax is probably all wrong here, but that's the idea anyway.
> >
> >
> > Here's a bug marked fixed with a similar goal to yours:
> >     http://d.puremagic.com/issues/show_bug.cgi?id=1661
> > The discussion may be useful.
> 
> I have an open enhancement request on this:
> 
> http://d.puremagic.com/issues/show_bug.cgi?id=1653
> 
> I think the main issue is really that there is no 'reverse' path from the alias to the parameter to a specific template.  That is given the type MyType!(T), which is really an alias for T, What template should the compiler use to deduce the argument?  What if there were multiple templates that aliased to T, should those also be assumed to be MyType!(T)?  What if you pass in just a T?  Technically it's the same thing.
> 
> I think these are the issues that makes it tough to do the IFTI.  I think we are spoiled because things like:
> 
> void f(T)(T[] x)
> 
> but that's a much simpler case to solve.
> 
> -Steve
> 
> 

I understand the issue. On the other hand, if I use a typedef instead of an alias, shouldn't the compiler be (conceptually) able to know exactly whether the passed in type is of the right type, because typedefs are uniquely identified? In other words, would you think the compiler could possibly be extended to accept my program if I use typedef instead of alias? It seems like a useful feature to have, if it is possible.

Clemens

"Clemens Hofreither" wrote
> Steven Schveighoffer Wrote:
>
>>
>> "Bill Baxter" <dnewsgroup@billbaxter.com> wrote in message news:g4s05a$2l0l$1@digitalmars.com...
>> > Clemens Hofreither wrote:
>> >> Hi all.
>> >>
>> >> I'm currently playing around with template metaprogramming in D. I've encountered a problem with implicit instantiation of templates that I don't really understand, and I thought somebody here could shed some light on it. Consider the following code:
>> >>
>> >> template MyType(T)
>> >> {
>> >> alias T MyType;
>> >> }
>> >>
>> >> void mytest(T)(MyType!(T) x)
>> >> {
>> >> }
>> >>
>> >> void main()
>> >> {
>> >> MyType!(int) x;
>> >> mytest(x);
>> >> }
>> >>
>> >> This does not work. I would expect the template mechanism to recognize
>> >> that the argument to mytest is an instance of MyType!(int) and hence
>> >> instantiate T with int, but instead I get the errors
>> >> template mytest(T) does not match any template declaration
>> >> template mytest(T) cannot deduce template function from argument types
>> >> (int)
>> >>
>> >> I've also tried replacing alias by typedef, with the same results.
>> >> Instantiating the template explicitly seems to work, but I would first
>> >> like to know *why* I can't do it like this before I use any
>> >> workarounds.
>> >> Any pointers? Thanks in advance.
>> >>
>> >> Oh, and this is DMD v1.029.
>> >
>> > It's just not that smart.  It's a specialization so you have to use specialization syntax.  Something like:
>> >
>> > void mytest(T : MyType!(S), S) {}
>> >
>> > But note that using specialization also disables IFTI.
>> > In D2 you should be able to use a constraint to do something like
>> >
>> > void mytest(T) if(is(T : MyType!(S))) {}
>> >
>> > My syntax is probably all wrong here, but that's the idea anyway.
>> >
>> >
>> > Here's a bug marked fixed with a similar goal to yours:
>> >     http://d.puremagic.com/issues/show_bug.cgi?id=1661
>> > The discussion may be useful.
>>
>> I have an open enhancement request on this:
>>
>> http://d.puremagic.com/issues/show_bug.cgi?id=1653
>>
>> I think the main issue is really that there is no 'reverse' path from the
>> alias to the parameter to a specific template.  That is given the type
>> MyType!(T), which is really an alias for T, What template should the
>> compiler use to deduce the argument?  What if there were multiple
>> templates
>> that aliased to T, should those also be assumed to be MyType!(T)?  What
>> if
>> you pass in just a T?  Technically it's the same thing.
>>
>> I think these are the issues that makes it tough to do the IFTI.  I think
>> we
>> are spoiled because things like:
>>
>> void f(T)(T[] x)
>>
>> but that's a much simpler case to solve.
>>
>> -Steve
>>
>>
>
> I understand the issue. On the other hand, if I use a typedef instead of an alias, shouldn't the compiler be (conceptually) able to know exactly whether the passed in type is of the right type, because typedefs are uniquely identified? In other words, would you think the compiler could possibly be extended to accept my program if I use typedef instead of alias? It seems like a useful feature to have, if it is possible.

Yes, typedef would be easier to infer, because the original template that constructed the type would be embedded in the type.  However, you lose a lot of usefulness by using typedef instead of alias.  But I doubt Walter will add the functionality just for typedef, as it is not the common case.

In any case, here is a good counter example for the alias:

template MyType!(T)
{
    static if(is(T == int))
       alias long MyType!(T);
    else
       alias T MyType!(T);
}

T f(T)(MyType!(T) t) {}

If f is passed a long, how should it infer the original type?  As a MyType!(int) or MyType!(long)?  It is important because the function needs to know which is correct so it can return the correct type.  Maybe this just becomes an error (if the requested functionality is implemented), similar to ambiguous overload.

I'll put this counter-example in the bugzilla report.

-Steve

For what it's worth, I have circumvented my original problem by making
the templated type in question a class. Since the code seems to work
nicely now, I figured I could as well post it; it's a generic memoize function
which works (type-safely) on functions with any return type and any
number and type of arguments (well, as long as they can be AA indices). I
would be interested in any improvement suggestions. Also, I'd really like to
know why I have to use "inout" for the AA argument in store(); I thought
AAs had reference semantics, but without inout it doesn't work as
expected.

Well, here's the code:


/*****************************************/
import tango.io.Stdout;

class MultiMap(S, T...)
{
	private template MultiMapT(R, Q...)
	{
		static if (Q.length <= 1)
			alias R[Q[0]] MultiMapT;
		else
			alias MultiMapT!(R,Q[1..length])[Q[0]] MultiMapT;
	}

	private alias MultiMapT!(S,T) M;
	private M map;

	private static S *lookup(R, Q...)(R[Q[0]] map, Q key)
	{
		static if (key.length <= 1)
			return key[0] in map;
		else {
			auto inner = key[0] in map;
			if (inner != null)
				return lookup!(typeof(R[Q[0]]), Q[1..length])(*inner, key[1..length]);
			else
				return null;
		}
	}

	private static void store(R, Q...)(inout R[Q[0]] map, S value, Q key)
	{
		static if (key.length <= 1)
			map[key[0]] = value;
		else
			store!(typeof(R[Q[0]]), Q[1..length])(map[key[0]], value, key[1..length]);
	}

	S *get(T key) {
		return lookup!(typeof(M[T[0]]), T)(map, key);
	}
	void set(S value, T key) {
		store!(typeof(M[T[0]]), T)(map, value, key);
	}
}



class Memo(R, P...)
{
	private R function(P) func;
	private MultiMap!(R,P) values;

	this(R function(P) f)
	{
		func = f;
		values = new MultiMap!(R,P);
	}

	R opCall(P args)
	{
		R *val = values.get(args);
		if (val != null)
			return *val;
		else {
			R v = func(args);
			values.set(v, args);
			return v;
		}
	}
}

Memo!(R, P) memoize(R, P...)(R function(P) myfunc)
{
	return new Memo!(R,P)(myfunc);
}


float foo(float x, float y)
{
	Stdout.format("foo called.\n");
	return x / y;
}

void main()
{
	auto f = memoize(&foo);
	Stdout.format("{}\n", f(3f, 4f));
	Stdout.format("{}\n", f(3f, 4f));
}
/*****************************************/

-Clemens