Purity

[Don]

Steven Schveighoffer wrote:
> On Fri, 17 Dec 2010 09:39:19 -0500, Don <nospam at nospam.com> wrote:
> 
>> A function which has immutably pure parameters can undergo *some* 
>> optimisation, even if the return value is a mutable pointer.
>> For example, if the parameters are identical for both calls, you can 
>> do a deepdup of the first return value instead of calling the function 
>> again.
> 
> Yes, I agree with this.  However, I still believe statement 1 is still 
> correct, because you really have just introduced a third class of pure 
> functions that I was not aware of :)
> 
> So we have:
> 
> weak-pure functions : functions which can accept and return any type of 
> values, but can only call pure functions and cannot access global state.
> middle-pure ? functions : weak-pure functions which accept only 
> immutable or implicitly convertable to immutable values, but returns a 
> mutable reference.


> strong-pure funtions : pure functions which accept and return only 
> immutable or implicitly convertible to immutable values.
> 
> There actually could be another class, const-pure functions: pure 
> functions which accept and return only const or implicitly convertible 
> to const values.

The problem with these ones, is there could be aliasing between the 
input and the output. Do they have any interesting properties?


Another interesting one is a function with all-const parameters, that 
returns a mutable reference that has only mutable or immutable members.
(ie, nothing const). Like the one you've called middle-pure, this 
returns a unique result, because you're guaranteed there is no aliasing.


> I suspect the compiler will need to classify things as strong or weak 
> pure, and store various attributes on weak-pure functions to see what 
> optimizations can be had.

The compiler does that already.