How do "pure" member functions work?

[Jonathan M Davis]

On Monday, August 22, 2011 15:57 Timon Gehr wrote:
> On 08/22/2011 10:19 PM, Don wrote:
> > Timon Gehr wrote:
> >> On 08/21/2011 09:10 PM, Don wrote:
> >>> bearophile wrote:
> >>>> Sean Eskapp:
> >>>>> Oh, I see, thanks! This isn't documented in the function
> >>>>> documentation!
> >>>> 
> >>>> D purity implementation looks like a simple thing, but it's not
> >>>> simple, it has several parts that in the last months have be added to
> >>>> the language and compiler, and we are not done yet, there are few more
> >>>> things to add (like implicit conversion to immutable of the results of
> >>>> strongly pure functions). It will need several book pages to document
> >>>> all such necessary design details.
> >>>> 
> >>>> Bye,
> >>>> bearophile
> >>> 
> >>> It is actually very simple: a function marked as 'pure' is not allowed
> >>> to explicitly access any static variables.
> >>> Everything else is just compiler optimisation, and the programmer
> >>> shouldn't need to worry about it.
> >> 
> >> It can be of value to know that a function is pure as in mathematics
> >> if it is strongly pure, but can have restricted side-effects if it is
> >> weakly pure.
> > 
> > Well, from the compiler's point of view, it's more complicated than
> > that. There are const-pure as well as immutable-pure functions. A
> > const-pure function has no side-effects, but cannot be optimised as
> > strongly as an immutable-pure function.
> 
> What significant optimization do immutable-pure functions benefit from
> that const-pure functions cannot? Is it just that the compiler cannot do
> CSE in some cases if there is an impure call in between two const-pure
> calls? (which I think would be rather insignificant)

The _only_ time that subsequent calls to a pure function can be optimized out 
is when all of its arguments are immutable or implicitly convertible to 
immutable, otherwise their values could change between calls to the function. 
A function whose parameters are all immutable or implicitly convertible to 
immutable is always going to be able to be optimized that way, because you 
_know_ that it's arguments are immutable or implicitly convertible to 
immutable (they _have_ to be). In cases where a pure function whose parameters 
are const or implicitly convertible to immutable is passed immutable 
arguments, it could be optimized (though that doesn't currently happen), but 
if the arguments aren't immutable, then it can't be, because they could change 
between calls to the function.

> > BTW: The whole "weak pure"/"strong pure" naming was just something I
> > came up with, to convince Walter to relax the purity rules. I'd rather
> > those names disappeared, they aren't very helpful.
> 
> In some contexts they are in fact useful, otherwise the compiler
> implementation wouldn't have any use for them.
> 
> > But the basic point is, that knowing that there are no static variables
> > is hugely significant for reasoning about code. The strong pure/weak
> > pure distinction is not very interesting (you can distinguish them
> > using only the function signature).
> 
> Yes, but saying 'the function is weakly pure' is shorter than saying
> 'the pure function signature does include mutable references', just like
> having a flag set to PUREweak is more efficient than examining the
> function signature multiple times.

Ultimately, whether pure functions can be optimized or not becomes an 
implementation detail. pure functions can't access mutable global state. 
That's what pure means. The compiler then does what it can to optimize pure 
functions. Weak and strong purity are what determine whether such 
optimizations can take place, but I think that Don is basically saying that 
that should just be an implementation detail and that what should be focused 
on with regards to pure functions is the fact that they can't access mutable 
global state.

- Jonathan M Davis