Overhead of Object Expressions when used with inline function - f#

I loved the idea used in the accepted answer here. My understanding of it is that it is an elegant way of using an interface as a method parameter without having to implement the interface. SRTP is used for duck typing to adopt a type (excuse the OO term, happy to learn the FP one) to the interface:
let inline namedModel< ^T when ^T : (member Name : string)> (model:^T)=
{ new INamed with
member x.Name =
(^T : (member Name : string) model) }
What confuses me in the code above which I took from the accepted answer is the runtime behaviour of the inlined method.
It appears to me that this method will create and return a new implementation of the INamed every single time it is called. As I asked in the comments in the linked question, would not this lead to a lot of pressure on the garbage collector if namedModel method was called many times?
I'm really keen to use this approach but I cannot dare go ahead with lest my understanding of its memory consumption is correct.

Yes, it will create a new object on every call.
But keep in mind the first rule of optimization: first measure, then optimize. Are you sure that creating an object on every call would be prohibitively expensive in your case? Have you measured?
Another thing to keep in mind is that in normal code you are constantly creating and discarding objects, often without even thinking of it. The .NET garbage collector is specifically designed to handle this.

Related

Arrow KT: Reader Monad vs #extension for Dependency Injection

I've read about Reader Monad from this article by Jorge Castillo himself and I've also got this article by Paco. It seems that both tackles the idea of Dependency Injection just in a different way. (Or am I wrong?)
I'm really confused whether I understand the whole Reader Monad and how it relates to the Simple Depenency Injection that Paco is talking about.
Can anyone help me understand these two things? Would I ever need both of them in one project depending on situations?
Your doubt is understandable, since yes both approaches share the same outcome: Passing dependencies implicitly for you all the way across your call stack, so you don't need to pass them explicitly at every level. With both approaches you will pass your dependencies once from the outer edge, and that's it.
Let's say you have the functions a(), b(), c() and d(), and let's say each one calls the next one: a() -> b() -> c() -> d(). That is our program.
If you didn't use any of the mentioned mechanisms, and you needed some dependencies in d(), you would end up forwarding your dependencies (let's call them ctx) all the way down on every single level:
a(ctx) -> b(ctx) -> c(ctx) -> d(ctx)
While after using any of the mentioned two approaches, it'd be like:
a(ctx) -> b() -> c() -> d()
But still, and this is the important thing to remember, you'd have your dependencies accessible in the scope of each one of those functions. This is possible because with the described approaches you enable an enclosing context that automatically forwards them on every level, and that each one of the functions runs within. So, being into that context, the function gets visibility of those dependencies.
Reader: It's a data type. I encourage you to read and try to understand this glossary where data types are explained, since the difference between both approaches requires understanding what type classes and data types are, and how they play together:
https://arrow-kt.io/docs/patterns/glossary/
As a summary, data types represent a context for the program's data. In this case, Reader stands for a computation that requires some dependencies to run. I.e. a computation like (D) -> A. Thanks to it's flatMap / map / and other of its functions and how they are encoded, D will be passed implicitly on every level, and since you will define every one of your program functions as a Reader, you will always be operating within the Reader context hence get access to the required dependencies (ctx). I.e:
a(): Reader<D, A>
b(): Reader<D, A>
c(): Reader<D, A>
d(): Reader<D, A>
So chaining them with the Reader available combinators like flatMap or map you'll get D being implicitly passed all the way down and enabled (accessible) for each of those levels.
In the other hand, the approach described by Paco's post looks different, but ends up achieving the same. This approach is about leveraging Kotlin extension functions, since by defining a program to work over a receiver type (let's call it Context) at all levels will mean every level will have access to the mentioned context and its properties. I.e:
Context.a()
Context.b()
Context.c()
Context.d()
Note that an extension function receiver is a parameter that without extension function support you'd need to manually pass as an additional function argument on every call, so in that way is a dependency, or a "context" that the function requires to run. Understanding those this way and understanding how Kotlin interprets extension functions, the receiver will not need to be forwarded manually on every level but just passed to the entry edge:
ctx.a() -> b() -> c() -> d()
B, c, and d would be called implicitly without the need for you to explicitly call each level function over the receiver since each function is already running inside that context, hence it has access to its properties (dependencies) enabled automatically.
So once we understand both we'd need to pick one, or any other DI approach. That's quite subjective, since in the functional world there are also other alternatives for injecting dependencies like the tagless final approach which relies on type classes and their compile time resolution, or the EnvIO which is still not available in Arrow but will be soon (or an equivalent alternative). But I don't want to get you more confused here. In my opinion the Reader is a bit "noisy" in combination with other common data types like IO, and I usually aim for tagless final approaches, since those allow to keep program constraints determined by injected type classes and rely on IO runtime for the complete your program.
Hopefully this helped a bit, otherwise feel free to ask again and we'll be back to answer 👍

Delay the implementation of interface methods?

I've been programming in F# for some years and there's an "issue" that's been bothering me for some time and I have not been able to solve. It is not a bug, I think it is a design decision, but anyway, the problem is this: is there a way to delay (maybe that's not the correct word for this) the implementation of interfaces?, that is, not implementing them in the initial definition, but later, maybe in the same file after I have implemented a module for the type. I'll explain with a simplified example:
Suppose I have the following data structure:
type 'T MyCollection =
(*type definition*)
interface IEnumerable<'T> with
member this.GetEnumerator () =
(* I don't want to implement it here
because I still don't have the module
with a toSeq function *)
If I implemented the method right there, I would have to also implement all the functions as methods of the type and then the module would be just a "proxy" for calling the methods. This way I'm creating a OO-first data structure and then creating a module (overloaded with type annotations) to allow for a functional-first usage. I would prefer to write a functional-first data structure (cleaner since the type inference can work better) and then create a OO wrapper to allow a better intellisense support for languages like C#. That approach complies with what the design guidelines for F# tells us, but the interfaces can't be implemented anywhere but in the initial definition of the type. That restriction forces me to write the whole data structure with members.
I've been looking for examples and I've found that the list implementation in FSharp.Core list does exactly what I want, but I can't do that, the compiler won't let me.
I'm almost sure that this is a design decision, maybe to avoid encouraging bad practices, I don't know, but I don't consider my wish to be a bad practice. Also I'm well aware of the linear nature of the fsharp compiler.
Please if any of you know how to do what I want, I'll be glad if you tell me. Also, if any of you know why I should not follow this approach I'll be glad to know too. There must be a reason why this is not a problem for anyone else.
Thanks in advance.
I completely agree that this is unfortunate problem. The trick that is used in the source code of 'a list in the F# Core library is to define the implementation of the interface in a type augmentation. The compiler does not complain when you add members to a type in this way, but it says that adding implementation of an interface in this way is deprecated. However, it does not prevent you from doing this. The following compiles fine for me:
open System.Collections
open System.Collections.Generic
type MyCollection<'T> =
{ Data : 'T list }
interface IEnumerable<'T>
interface IEnumerable
let getEnumerator { Data = d } =
(d :> seq<_>).GetEnumerator()
type MyCollection<'T> with
interface IEnumerable<'T> with
member this.GetEnumerator() = getEnumerator this
interface IEnumerable with
member this.GetEnumerator() = (getEnumerator this) :> _
The fact that this is deprecated is a bit unfortunate. I quite like this style and I use it when it makes sense. You can start a discussion about this on F# user voice and perhaps it could be turned back into a normal accepted feature :-)

F# Instance Methods... should they return a new instance instead of altering the current object?

The problem is whether an instance method should in anyway alter the object that contains the method or should it return a new instance? I'm new to F# and the concept of full mmutability that is suggested for F#.
Just using psuedo code for now unless I need to be more specific.
First thought is just add the message to the message list on the object:
class Something
ctr(messages)
_messages.Add(messages)
AddMessage(message)
_messages.Add(message)
Second is to construct a new list that joins the old list and the new message. Then I would create a new instance altogther and send back.
class Something
ctr(messages)
_messages.Add(messages)
AddMessage(message)
newMessageList = _messages.Join(message)
return new Something(newMessageList)
Am I overthinking immutability?
In my opinion, the answer depends on your requirements. The immutable style is probably more idiomatic, and would be a sensible default. However, one nice thing about F# is that you can choose what to do based on your needs; there's nothing inherently wrong with code that uses mutation. Here are some things to consider:
Sometimes the mutable approach leads to better performance, particularly when used in a single-threaded context (but make sure to measure realistic scenarios to be sure!)
Sometimes the immutable approach lends itself better to use in multi-threaded scenarios
Sometimes you want to interface with libraries that are easier to use with imperitave code (e.g. an API taking a System.Action<_>).
Are you working on a team? If so, are they experienced C# developers? Experienced F# developers? What kind of code would they find easiest to read (perhaps the mutable style)? What kind of code will you find easiest to maintain (probably the immutable style)?
Are you just doing this as an exercise? Then practicing the immutable style may be worthwhile.
Stepping back even further, there are a few other points to consider:
Do you really even need an instance method? Often, using a let-bound function in a module is more idiomatic.
Do you really even need a new nominal type for what you're doing? If it's just a thin wrapper around a list, you might consider just using lists directly.
As you are doing "class based" programming which is one of the way (rather unfortunate) to do object oriented programming, you would be doing in place state modification rather than returning a new state (as that's what would be expected when you are doing OO).
In case you really want to go towards immutability then I would suggest you need to use more FP concepts like Modules, Functions (not methods which have you have in class based programming), recursive data types etc.
My answer is way too general and the appropriate answer lies in the fact that how this class of your will fit in the big picture of your application design.

XNA/C#: Entity Factories and typeof(T) performance

In our game (targeted at mobile) we have a few different entity types and I'm writing a factory/repository to handle instantiation of new entities. Each concrete entity type has its own factory implementation and these factories are managed by an EntityRepository.
I'd like to implement the repository as such:
Repository
{
private Dictionary <System.Type, IEntityFactory<IEntity>> factoryDict;
public T CreateEntity<T> (params) where T : IEntity
{
return factoryDict[typeof(T)].CreateEntity() as T;
}
}
usage example
var enemy = repo.CreateEntity<Enemy>();
but I am concerned about performance, specifically related to the typeof(T) operation in the above. It is my understanding that the compiler would not be able to determine T's type and it will have to be determined at runtime via reflection, is this correct? One alternative is:
Repository
{
private Dictionary <System.Type, IEntityFactory> factoryDict;
public IEntity CreateEntity (System.Type type, params)
{
return factoryDict[type].CreateEntity();
}
}
which will be used as
var enemy = (Enemy)repo.CreateEntity(typeof(Enemy), params);
in this case whenever typeof() is called, the type is on hand and can be determined by the compiler (right?) and performance should be better. Will there be a noteable difference? any other considerations? I know I can also just have a method such as CreateEnemy in the repository (we only have a few entity types) which would be faster but I would prefer to keep the repository as entity-unaware as possible.
EDIT:
I know that this may most likely not be a bottleneck, my concern is just that it is such a waste to use up time on reflecting when there is a slightly less sugared alternative available. And I think it's an interesting question :)
I did some benchmarking which proved quite interesting (and which seem to confirm my initial suspicions).
Using the performance measurement tool I found at
http://blogs.msdn.com/b/vancem/archive/2006/09/21/765648.aspx
(which runs a test method several times and displays metrics such as average time etc) I conducted a basic test, testing:
private static T GenFunc<T>() where T : class
{
return dict[typeof(T)] as T;
}
against
private static Object ParamFunc(System.Type type)
{
var d = dict[type];
return d;
}
called as
str = GenFunc<string>();
vs
str = (String)ParamFunc(typeof(String));
respectively. Paramfunc shows a remarkable improvement in performance (executes on average in 60-70% the time it takes GenFunc) but the test is quite rudimentary and I might be missing a few things. Specifically how the casting is performed in the generic function.
An interesting aside is that there is little (neglible) performance gained by 'caching' the type in a variable and passing it to ParamFunc vs using typeof() every time.
Generics in C# don't use or need reflection.
Internally types are passed around as RuntimeTypeHandle values. And the typeof operator maps to Type.GetTypeFromHandle (MSDN). Without looking at Rotor or Mono to check, I would expect GetTypeFromHandle to be O(1) and very fast (eg: an array lookup).
So in the generic (<T>) case you're essentially passing a RuntimeTypeHandle into your method and calling GetTypeFromHandle in your method. In your non-generic case you're calling GetTypeFromHandle first and then passing the resultant Type into your method. Performance should be near identical - and massively outweighed by other factors, like any places you're allocating memory (eg: if you're using the params keyword).
But it's a factory method anyway. Surely it won't be called more than a couple of times per second? Is it even worth optimising?
You always hear how slow reflection is, but in C#, there is actually fast reflection and slow reflection. typeof is fast-reflection - it has basically the overhead of method call, which is nearly infinitesimal.
I would bet a steak and lobster dinner that this isn't going to be a performance bottleneck in your application, so it's not even worth your (or our) time in trying to optimize it. It's been said a million times before, but it's worth saying again: "Premature optimization is the root of all evil."
So, finish writing the application, then profile to determine where your bottlenecks are. If this turns out to be one of them, then and only then spend time optimizing it. And let me know where you'd like to have dinner.
Also, my comment above is worth repeating, so you don't spend any more time reinventing the wheel: Any decent IoC container (such as AutoFac) can [create factory methods] automatically. If you use one of those, there is no need to write your own repository, or to write your own CreateEntity() methods, or even to call the CreateEntity() method yourself - the library does all of this for you.

Is there any good way for an inline function to access private or internal values?

I just ran into an issue: when I try to access a private or internal value from an inline function, I get the error "The value 'xxx' was marked inline but its implementation makes use of an internal or private function which is not sufficiently accessible". While logical, I wonder if anyone has a good work around. The best thing I can think to do is place the values publicly in a nested module and just hope no-one goes poking around (which I'm not too worried about anyway, since these values are immutable). I suppose reflection is an option, but without being able cache calls (using... private delegates) the performance hit is too much.
Short answer: no, since the value will be inserted inline into the call-site, it can't use private values and there's no real way to work around it.
Longer answer: if you don't mind writing incredibly ugly code and you can handle the overhead of a few method calls per use, one alternative would be to create a dynamic implementation (e.g. OperatorIntrinsics.AbsDynamicTableImpl in the core library), which can be private. You can then wrap the dynamic implementation in a public opaque generic method (e.g. OperatorIntrinsics.AbsDynamic<'T>), and then create an inline value which adds the proper type constraints and defers to the dynamic implementation (e.g. let inline abs< ^t when ^t : (static member Abs : ^t -> ^t)> x = AbsDynamic x). Now when you inline abs you just see a call to AbsDynamic but none of the further implementation details. In most cases I would expect this to be a much worse solution than just making your value public instead of private.

Resources