Re: How can one record component be local and another not?

[Niklas Holsti <niklas.holsti@tidorum.invalid>]

On 2020-05-06 21:28, Jere wrote:
> On Wednesday, May 6, 2020 at 1:30:58 PM UTC-4, Niklas Holsti wrote:
>> On 2020-05-05 20:32, hreba wrote:
>>> Ok, with all your hints I came to the following solution:
>>>
>>> -- 
>>> package Aux is
>>>      type Integer_P is access all Integer;
>>>      type Rec is limited record
>>>         a: aliased Integer;
>>>         p: Integer_P;
>>>      end record;
>>>
>>>      procedure Init (r: access Rec);
>>> end Aux;
>>> -- 
>>> package body Aux is
>>>      procedure Init (r: access Rec) is
>>>      begin
>>>         r.p:= r.a'Access;
>>>      end Init;
>>> end Aux;
>>> -- 
>>> with Aux;
>>> procedure Test is
>>>      r:    aliased Aux.Rec;
>>> begin
>>>      Aux.Init (r'Access);
>>> end Test;
>>> -- 
>>>
>> <SNIPPED>
>> Somewhat surprisingly to me, any call of the form
>> "Aux.Init (r'Unchecked_Access)" also works, whether "r" is a
>> library-level object or a local (limited-life) object, even if
>> "Aux.Init" itself uses "'Access" and not "'Unchecked_Access". Apparently
>> the accessibility level provided in the value of Unchecked_Access is
>> such as to make the check succeed, as for a library-level object.
>>
> Yeah, RM section 13.10 says it treats it as if the object were declared
> at library level for Unchecked_Access

That explains it clearly, thanks for the reference.

>> I also found that a call like "Aux.Init (new Aux.Rec)", which applies
>> "Aux.Init" to a heap-allocated "Rec" object, fails the accessibility
>> check, so this check is really strict. Of course, the heap object can be
>> deallocated at any time, so it may have a short life-time, but that can
>> be done (mainly) by Unchecked_Deallocation, so it involves "unchecked"
>> programming. Hm.
>>
> Is that because the heap allocated object is allocated using an
> anonymous access type instead of a named access type per chance?
> I know anonymous access types have different accessibility level
> rules.  It might be setting the accessibility level to the point
> of call maybe?

I made some more experiments:

a) anonymous access (Aux.Init (new Aux.Rec)): check fails

b) local named type "access Aux.Rec": check fails

c) library-level access type (access Aux.Rec declared in package Aux): 
works.

Interesting... I haven't developed a good intuition for this yet.

-- 
Niklas Holsti
Tidorum Ltd
niklas holsti tidorum fi
       .      @       .