GVM dispatch is notoriously slow(-ish), yeah. But it does not require JIT. Otherwise it wouldn't work with NativeAOT :) (the latter can also auto-seal methods and unconditionally devirtualize few-implementation members which does a good job, guarded devirtualization with JIT does this even better however)
I remember when this feature was specifically not available with NativeAOT.
It's good that it is now, but how can it be implemented in a way that has truly separate instantiations of generics at runtime, when calls cross assembly boundaries? There's no single good place to generate a specialization when virtual method body is in one assembly while the type parameter passed to it is a type in another assembly.
> how can it be implemented in a way that has truly separate instantiations of generics at runtime, when calls cross assembly boundaries
There are no assembly boundaries under NativeAOT :)
Even with JIT compilation - the main concern, and what requires special handling, are collectible assemblies. In either case it just JITs the implementation. The cost comes from the lookup - you have to look up a virtual member implementation and then specific generic instantiation of it, which is what makes it more expensive. NativeAOT has the definitive knowledge of all generic instantiations that exist, since it must compile all code and the final binary does not have JIT.
