Implicit cast to immutable

[Steven Schveighoffer]

On Wed, 19 Oct 2011 21:39:47 -0400, Daniel Murphy  
<yebblies at nospamgmail.com> wrote:

>
> "Steven Schveighoffer" <schveiguy at yahoo.com> wrote in message
> news:op.v3lug2q2eav7ka at localhost.localdomain...
>> On Tue, 18 Oct 2011 02:40:12 -0400, Daniel Murphy
>> <yebblies at nospamgmail.com> wrote:
>>
>>> "Steven Schveighoffer" <schveiguy at yahoo.com> wrote in message
>>> news:op.v3h06olweav7ka at localhost.localdomain...
>>>>
>>>> That sounds like an incorrect restriction.  The implicit cast to
>>>> immutable
>>>> should depend on whether the function being *called* qualifies, not if
>>>> the
>>>> function you are calling *from* qualifies.
>>>>
>>>
>>> I think you've misunderstood what I'm saying.  The patch I made
>>> implemented
>>> two ways to implicly convert to immutable: the result of a pure  
>>> function
>>> returning immutable and a return statement inside a pure function.
>>>
>>> 1)
>>> int[] fun() { return new int[]; }
>>> immutable x = fun(); // conversion happens here
>>>
>>> 2)
>>> immutable(int[]) fun() { return new int[]; } // conversion happens here
>>> immutable x  = fun();
>>
>> Neither of those is relevant, since they do not contain parameters (I'm
>> assuming you meant them both to be pure).
>>
>
> Yeah, they were supposed to be pure.  I was just demonstrating that
> conversion happens at the return statement, or the call site, but only
> currently for strongly pure functions.
>
>>>> Qualifying means the return type should be mutable, and cannot be
>>>> derived
>>>> from the parameters without requiring casts.  The easiest way to do  
>>>> this
>>>> is to ensure the parameters are all const, immutable, or implicitly  
>>>> cast
>>>> to immutable.  You could do funky things like assume for instance an
>>>> int[]
>>>> cannot possibly be implicit-casted to a char[], so therefore int[]
>>>> foo(char[] data) pure can be implicitly casted to immutable, but that
>>>> might be flirting with dangerous situations.
>>>>
>>>> The one exception should be allocating memory, which should always
>>>> qualify, even though it's not a pure function.
>>>>
>>>
>>> This is valid, but becomes very complicated with complex return  
>>> types.  I
>>> doubt this will ever make it into the language.
>>
>> No, it's not complicated, at least in my view.  The rules are:
>>
>> 1. if the pure function parameters are only immutable or
>> implicitly-convertible-to-immutable, the result is safe to cast to
>> immutable, regardless of what happens inside the function, or the type  
>> of
>> the result.
>> 2. if the pure function parameters contain any mutable data that is
>> mutable and *not* implicitly convertible to mutable, and the result is
>> only safe to cast to immutable if it can already implicitly cast to
>> immutable.
>> 3. if the pure function parameters are implicitly castible to immutable,
>> immutable, or are const, then:
>>       a) if the return type is implicitly castable to immutable,  
>> obviously
>> it can be cast.
>>       b) if the return type is not implicitly castable to immutable, but
>> contains only mutable references, then it's implicitly castable to
>> immutable.
>>       c) otherwise, it cannot be implicitly cast.
>>
>> You do not need the function body to determine this, and the decision
>> should be made by the compiler at the call site.
>>
>> Your case 2 where the conversion happens at the return statement is
>> already covered.
>>
>>> I've got the beginnings of a patch to enable a sort of 'pure  
>>> expression',
>>> such as new, array.dup and array concatenation expressions.  The result
>>> of a
>>> call to a const-pure function using immutable arguments can be  
>>> converted
>>> to
>>> immutable, while calling it with mutable or const arguments cannot,
>>> without
>>> searching the return type for anything the arguments can implicitly
>>> convert
>>> to (or create).
>>>
>>> Eg. I can't see a great way to detect situations like this:
>>>
>>> struct S { const void* p; }
>>
>> This struct could only be cast to immutable under my rule 1 above.  It
>> does not implicitly cast to immutable.
>>
>>> S[] fun(int[] arr)
>>> {
>>>     return [ S(arr.ptr) ];
>>> }
>>> immutable x = fun([1, 2, 3]);
>>
>> Invalid.  arr is not implicitly castable to immutable and is mutable, so
>> the result must already be implicitly castable (which it is not).  This
>> falls under rule 2 above, and fails the test.
>>
>> -Steve
>
> Ok, I think I meant to make fun pure and fun's parameter 'in', so it  
> would
> fall under rule 3 but be rejected by 3b as it contains a const member.
>
> It's not always simple to determine if an aggregate contains non-mutable
> members.  An example would be a class when you only have a base class
> reference.

I had not thought about classes (or interfaces), I was only thinking of  
concrete types.  I think in the case of classes, all parameters with  
classes must be marked as immutable to have an implicit cast of the result  
to immutable (i.e. fall under rule 1).  In reality, the rules I specify  
are enough, but they simply aren't explicit about how classes make it  
impossible to determine if any const members are aboard.

The same goes for void *, which could point to a type which has const  
members.

All cases where classes or void * pointers are involved must fall under  
rule 1, since you cannot determine whether such types are transitively  
mutable.

> Another problem (which we're already seeing with the existing  
> conversions)
> is that when you screw it up the conversion fails, but doesn't give you  
> any
> hint as to why it failed.  Making the rules more complicated is just  
> going
> to make this worse.

By letting the caller choose the return type, you are eliminating code  
duplication.

Consider the function:

int[] repeat(in int[] x, uint ntimes) pure
{
    int[] result;
    result.reserve(x.length * ntimes);
    foreach(uint i; 0..ntimes)
       result ~= x;
    return result;
}

Under my rules, you need only this one function to handle all 9 cases of  
converting between constancies of int

int[]        => int[]
int[]        => const(int)[]
int[]        => immutable(int)[]
const(int)[] => int[]
...

I think this is worth having.  In contrast, what we have now is quite  
limited, and inconsistent.

A long-standing request I've had is this:

http://d.puremagic.com/issues/show_bug.cgi?id=1654

I think this new feature of implicit immutable casting via pure is the key.

The more we make immutable data safer and easier to create and convert,  
the better off D (and pure functions) will be.

-Steve