www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - can't initialize .outer in inner class

reply Andrei Alexandrescu <SeeWebsiteForEmail erdani.org> writes:
Consider:

void fun()
{
     int x;
     class C
     {
         ...
     }
}

Objects of type C have access to x because they have an .outer 
pseudo-member.

Problem is, emplace() and any other in-situ initialization techniques 
fail (e.g. emplace() will fail with inner classes).

This seems to be a compiler issue - there's no way to initialize outer 
without calling new. What would be a proper fix to this?


Thanks,

Andrei
May 06 2015
parent reply Timon Gehr <timon.gehr gmx.ch> writes:
On 05/06/2015 10:01 AM, Andrei Alexandrescu wrote:

 Consider:

 void fun()
 {
      int x;
      class C
      {
          ...
      }
 }

 Objects of type C have access to x because they have an .outer
 pseudo-member.

 Problem is, emplace() and any other in-situ initialization techniques
 fail (e.g. emplace() will fail with inner classes).

 This seems to be a compiler issue - there's no way to initialize outer
 without calling new.



The following workaround seems to do it (I didn't test it thoroughly 
though, in particular, I didn't test whether escape analysis always 
works correctly for this implementation):

T nestedEmplace(T,alias x,S...)(void[] mem,S args){
     auto res=cast(T)mem.ptr;
     enum siz=__traits(classInstanceSize, T);
     (cast(byte[])mem)[0..siz]=typeid(T).init[];
     auto dg=(){ return x; };
     res.outer=dg.ptr;
     static if(is(typeof(res.__ctor(args)))) res.__ctor(args);
     else assert(!is(typeof(&res.__ctor))&&args.length==0);
     return res;
}

void main(){
     int x=12345;
     class C{
         this(){}
         int foo(){
             return x;
         }
     }
     void[__traits(classInstanceSize,C)] mem=void;
     auto c=nestedEmplace!(C,x)(mem);
     assert(c.foo()==12345);
     x=3;
     assert(c.foo()==3);
}



 What would be a proper fix to this?


Probably nested types should invoke nested template instantiation, as 
the context pointer is runtime data. Then add a way to access the 
context pointer associated with some symbol (e.g. __traits).


 Thanks,

 Andrei
May 06 2015
next sibling parent Andrei Alexandrescu <SeeWebsiteForEmail erdani.org> writes:
On 5/6/15 3:31 AM, Timon Gehr wrote:

 On 05/06/2015 10:01 AM, Andrei Alexandrescu wrote:

 Consider:

 void fun()
 {
      int x;
      class C
      {
          ...
      }
 }

 Objects of type C have access to x because they have an .outer
 pseudo-member.

 Problem is, emplace() and any other in-situ initialization techniques
 fail (e.g. emplace() will fail with inner classes).

 This seems to be a compiler issue - there's no way to initialize outer
 without calling new.



 The following workaround seems to do it (I didn't test it thoroughly
 though, in particular, I didn't test whether escape analysis always
 works correctly for this implementation):

 T nestedEmplace(T,alias x,S...)(void[] mem,S args){
      auto res=cast(T)mem.ptr;
      enum siz=__traits(classInstanceSize, T);
      (cast(byte[])mem)[0..siz]=typeid(T).init[];
      auto dg=(){ return x; };
      res.outer=dg.ptr;
      static if(is(typeof(res.__ctor(args)))) res.__ctor(args);
      else assert(!is(typeof(&res.__ctor))&&args.length==0);
      return res;
 }

 void main(){
      int x=12345;
      class C{
          this(){}
          int foo(){
              return x;
          }
      }
      void[__traits(classInstanceSize,C)] mem=void;
      auto c=nestedEmplace!(C,x)(mem);
      assert(c.foo()==12345);
      x=3;
      assert(c.foo()==3);
 }


Thanks, that's quite the tour de force. I tried to massage things in 
various places to avoid changing emplace()'s signature, no avail. I 
think we have a problem here. -- Andrei
May 06 2015
prev sibling parent Kenji Hara via Digitalmars-d <digitalmars-d puremagic.com> writes:
2015/05/06 19:35 "Timon Gehr via Digitalmars-d" <digitalmars-d puremagic.com

:

 Probably nested types should invoke nested template instantiation, as the


context pointer is runtime data. Then add a way to access the context
pointer associated with some symbol (e.g. __traits).

Related: https://issues.dlang.org/show_bug.cgi?id=8863

Kenji Hara
May 06 2015