Devirtualization and Static Polymorphism
david.alvarezrosa.com
david.alvarezrosa.com
Coming from C++ I assumed this was the only way but Rust has an interesting approach where the single objects do not pay any cost because virtual dispatch is handled by fat pointers. So you carry around the `vptr` in fat pointers (`&dyn MyTrait`) only when needed, not in every instance.
Do you know how is this exactly deduced?
A specific type/reference to a type will always use static dispatch.
fn foo(bar: &Baz) { bar.thing(); }
A dyn trait reference will always use dynamic dispatch and carry around the vtable pointer.
fn foo(bar: &dyn BazTrait) { bar.thing(); }
And you can use each approach with the same type at different points in the code - even for the same function. It just depends on you local knowledge of the concrete type.
No, I don't think the way Rust implements dynamic dispatch has much, if anything, to do with trying to avoid code bloat. It's just a different way to implement dynamic dispatch with its own set of tradeoffs.
let a: i32 = 12
let b = &a as &dyn std::string::ToString; // i32 implements the ToString trait
let c = a.to_string(); // Static dispatch
let d = b.to_string(); // Dynamic dispatch through dyn reference
Note that there's not really any polymorphic objects in rust. All polymorphism in this case goes through the dyn reference which contains a pointer to a vtable for a specific trait.
Additionally, going from a dyn reference to a type-specific reference is not easy. Also, certain methods and traits are not dyn-compatible, mostly due to generic parameters.
The main use comes in with various libraries. Doing dynamic dispatch on a specific type is not very useful, but your library might expose a trait which you then call some methods on. If you accept a generic parameter (eg. impl Trait) each such invocation will cause monomorphization (the function body is compiled separately for each generic type combination). This can obviously bloat compile times.
Using a dyn reference in your API will result in only a single version being compiled. The downside is the inability to inline or optimize based on the type.
One additional use I found is that you can sometimes get around the divergent expression type in match expressions. Say you need to print out some values of different types:
let value: &dyn Display = match foo { A(numeric_id) => &numeric_id, B(string_name) => &string_name, C => &"static str", };
This would not work without dyn as each value has a different type.
Does this mean that the Rust frontend is spitting out the intermediate representation such that it doesn't allow the de-duplication at the linking phase? I see that Rust now has its own linker as of Sep 25' but it still works with normal linkers that are used in C and C++ too - gnu ld, lld, mold, ...
I'm not sure how useful deduplication at the linker level is in practice. Though I don't think Rust does anything different here than C++. The main issue I imagine is that the types used in generic code have different sizes and layouts. This seems to me like it would prevent deduplication for most functions.
One of the papers I had bookmarked when toying with my own language design was someone that had worked out how to make interfaces as fast or faster than vtables by using perfect hashing and using the vtable as a hash table instead of a list.
You can also, when inlining a polymorphic call, put a conditional block in that bounces back to full dispatch if the call occasionally doesn’t match the common case. The problem with polymorphic inlining though is that it quickly resembles the exact sort of code we delete and replace with polymorphic dispatch:
if (typeof arg1 == “string”) {
} else if typeof arg1 === …) {
} else if {
} else if {
} else {
}One caveat with "hash vtables" is that you only really see a performance win when the interface has a lot of specializations.
For instance, if you treat some collections as read only, you can define comprehensions across them with a single implementation. But that means the mutators have to be contained in another type, which a subset will implement, and may have covariant inputs.
Could you explain this a bit more? The word "list" makes me think you might be thinking that virtual method lookup iterates over each element of the vtable, doing comparisons until it finds a match -- but I'm certain that this is not how virtual method invocation works in C++. The vtable is constructed at compile time and is already the simplest possible "perfect hashtable": a short, dense array with each virtual method mapping to a function pointer at a statically known index.
So these guys essentially assigned a hashcode to every function of every interface and then you would do dispatch instead of obj.vtable[12] you would do modular math x = singature.hash % len(obj.vtable) and call that.
I believe this was sometime around 2005-2008 and they found that it was fast enough on hardware of that era to be usable.
But maybe I don't get it, since this would require knowledge of all interfaces, and as soon as you require that, it's straightforward to build a minimal-size mapping from method name+signature to integer index: e.g., just form the union of all method declarations appearing in any interface, sort them lexicographically, and use a method's position in this sorted list as its index. Lookups in this map are only ever done at compile time so there's no runtime inefficiency to worry about.
Although really it isn’t because the JVM is strongly typed but evaluates some things at load or first invocation so it allows some languages that run in the JVM to be a bit tricky. They first generics implementation on the JVM, called Pizza, leveraged this load time concretization to do its thing.
But if you have a language that can resolve the type system at link time then you can do this trick. Alternatively you could switch to cuckoo hashing and if you next module load starts causing collisions, then so be it.
It's a fair amount of extra work, but in a hot loop it's sometimes worth it. "You can often solve correctness problems (tricky corner cases) by adding an extra layer of indirection. You can solve any performance problem by removing a layer of indirection."
There's absolutely nothing wrong with this code. It's just that it's not as extensible
It's a 'closed world' representation where the code assumes it knows about every possibility. This make extension more difficult
The code itself is extraordinarily good and performant.
With the polymorphic approach, I just have to create the new subtype, and all the users can do the right thing (if they were written with polymorphism in mind, anyway - if they use virtual functions on the base class).
Still... doing it the C++ way, I can just declare the sub type as deriving from the super type, and I don't have to fix up the super type.
For me, the key insight was from the last paragraph of the article:
C++23 introduces "deducing this", which is a way to avoid the performance cost of dynamic dispatch without needing to use tricks like CRRT, by writing:
class Base {
public:
auto foo(this auto&& self) -> int { return 77 + self.bar(); }
};
class Derived : public Base {
public:
auto bar() -> int { return 88; }
};
With concepts, templates and compile time execution, there is no need for CRTP, and in addition it can cover for better error messages regarding what methods to dispatch to.
But still CRTP is widely used in low-latency environments :)