UDAs for function arguments

[IgorStepanov]

Implementation of subj.

This topic was partially discussed ~year ago in NG.
The general idea:

class MyController
{
     View foo(@postParam("a") int p_a, @getParam("b") int p_b);
}

....
static if (is(typeof(MyController.foo) PAR == __parameters))
{
   enum attr_0 = __traits(getAttributes, PAR[0 .. 1]);
   enum attr_1 = __traits(getAttributes, PAR[1 .. 2]);
   static assert(attr_0 == postParam("a"));
   static assert(attr_1 == getParam("b"));
}

This feature is very usefull for different bindings generators like Web MVC or REST bindings like Java Spring MVC.

[schuetzm]

test12x vs test13x?

[IgorStepanov]

Omg, this test was totally wrong. Fixed.

[schuetzm]

I'm afraid I don't understand that static if() construction. typeof(MyController.foo) is probably something like View delegate(int, int). We would that ever be equal to __parameters?

Urgh... After some searching I found this:
http://forum.dlang.org/post/mailman.1197.1379014886.1719.digitalmars-d-learn@puremagic.com
I'm horrified 😱

[IgorStepanov]

I'm afraid I don't understand that static if() construction. typeof(MyController.foo) is probably something like View delegate(int, int). We would that ever be equal to __parameters?
Urgh... After some searching I found this:
http://forum.dlang.org/post/mailman.1197.1379014886.1719.digitalmars-d-learn@puremagic.com
I'm horrified 😱

Yes, __parameters result looks strange.

You should write __traits(identifier, PT[0 .. 1]) to get correct result. If you write __traits(identifier, PT[0]) you will get a __traits(identifier, int) result.
See phobos soultion for more info.
Anywhere, this feature should be wrapped into high-level phobos template like getArgumentUDAs!(func, <arg num>)

[Geod24]

Having to emulate UDAs on parameters is doable, although it's a pain.

I'd really like to have something in the language to support it (as long as it doesn't conflict with other features / grammar), but __parameters is an horrible hack, and so are some of the things built on top of it, such as ParameterDefaultValueTuple which is at the moment totally broken for meta programming purposes.

Having a magical tuple that can hold things not expressible in any other part of the language (and that vanishes as soon as you put them through a tuple) is already bad enough, please, don't put more magic in it.

[MetaLang]

I think it's important to note that D can currently introduce new storage classes without breaking any code, due to the fact that UDAs on parameters are not allowed. This pull request would make that no longer true. For example, we could add a new storage class today and be confident that no code would be broken:

struct Test { /*...*/ }

void acceptRvals(@rval ref Test t); //@rval currently doesn't conflict with UDAs

Although I doubt that is much of a concern at this point.

[IgorStepanov]

Having to emulate UDAs on parameters is doable, although it's a pain.

Yes, current __parameters interface looks like hack. However it is used for identifier or default value access now, and may be used for UDA access via a special Phobos template.
This PR is about UDA, not about __parameters. If another way to access to parameters data will be introduced in the language, this UDA will work through it.

I think it's important to note that D can currently introduce new storage classes without breaking any code, due to the fact that UDAs on parameters are not allowed.

The same point for global UDAs. However, when @nogc was being implemented, UDAs was exist and potentially may cause a conflict with the new attribute.

[9rnsr]

@IgorStepanov Additionally you need to tweak stripDefaultArgs in mtype.c. Because:

enum uda;
void foo(@uda int v) {}

alias x = X!(typeof(foo));

template X(F) {
    // in here, the function type F must not have any udas.
}

[IgorStepanov]

@9rnsr Done. Also I've added the corresponding test.
@CyberShadow What is the trouble with doc tester? How can I fix it?

[CyberShadow]

@CyberShadow What is the trouble with doc tester? How can I fix it?

Git repository corruption I think. Fixed for now, hopefully it won't recur

[IgorStepanov]

@9rnsr Any other objections?

Git repository corruption I think. Fixed for now, hopefully it won't recur

Thanks.

[9rnsr]

Better wording: s/argument/parameter/

[IgorStepanov]

fixed

[9rnsr]

Excepting nitpicks for the diagnostic messages, the implementation looks good to me.

[IgorStepanov]

Excepting nitpicks for the diagnostic messages, the implementation looks good to me.

I've fixed messages. Thanks!

[Geod24]

This PR is about UDA, not about __parameters. If another way to access to parameters data will be introduced in the language, this UDA will work through it.

My point was that as soon as it's introduced, we'll have to support it (unless it doesn't make it into a release). But since this would ease my life a lot, I won't complain too much :-)

Is there a corresponding documentation change open ?

Also, any chance you can add a test for the following scenarios:

• C variadic parameters;
• UDAs on the RHS of the identifier (fails with a meaningfull error message);
• Taking aggregates / using a function return value (like in Vibe.d);
• Types UDAs (including one with an enum Foo;);
• UDAs on function / delegate parameters (including ones with @nogc or @safe);

[IgorStepanov]

@Geod24

My point was that as soon as it's introduced, we'll have to support it (unless it doesn't make it into a release).

We will use phobos wrapper template in high-level code. Similar to ParameterIdentifierTuple and other. See the simple example of this template in the test runnable/uda.d.
I am also not happy with the current situation with __parameters, but if we start changing this situation now, we will finish this work near Christmas. Maybe. :)

Also, any chance you can add a test for the following scenarios:

Yes, done.

C variadic parameters;
UDAs on the RHS of the identifier (fails with a meaningfull error message);

See fail_compilation/udaparams.d

Taking aggregates / using a function return value (like in Vibe.d);
Types UDAs (including one with an enum Foo;);
UDAs on function / delegate parameters (including ones with @nogc or @safe);

See runnable/uda.d

Any other ideas?

Is there a corresponding documentation change open ?

It is not made. But I'll do it.

[IgorStepanov]

@9rnsr I've changed parse.c and func.c code. (Added error for postfix uda declaration and passing uda for parameter VarDeclaration* in FuncDeclaration.
Is it ok?

[IgorStepanov]

@9rnsr Please, revalidate this. I've added two changes, declared in previous post.

[9rnsr]

s/cannot has/cannot have/

[9rnsr]

Looks good excepting one diagnostic nitpick.

[MetaLang]

This is inconsistent with the error message on the very next line, which uses user defined attributes as opposed to UDAs. You should use either one or the other, but not both.

[IgorStepanov]

Ok, will change to "user defined attributes".

[TheFlyingFiddle]

What's the status on this?

[wilzbach]

Revived in #7576