[egonelbre]

> Without generics you can't separate algorithms from data structures so everything is much less useful.

Generics is not the only way to separate data and algorithms. I agree that it is possible to sub-type/embed and rename - although looking at generic code, I rarely see that happen.

> That "concrete solution" not only has to be written, but it will also perform identically to an instantiated template.

When you know where exactly the code will be run, it can be possible to create a better package taking into account the context where it will be run. By concrete, I mean a package in a real-world problem.

Basically, abstract things are harder to optimize for a particular case.

[p0nce]

> Generics is not the only way to separate data and algorithms.

Do you see any way other than the 3 ones discussed in the OP article under the label "Generics"?

> When you know where exactly the code will be run, it can be possible to create a better package taking into account the context where it will be run. By concrete, I mean a package in a real-world problem. Basically, abstract things are harder to optimize for a particular case.

I could agree with that if time wasn't limited, but imho building custom data structure/algorithms for each problem will precisely leave less time to optimize for performance since reaching a working state first will be longer without reusable abstractions.

[egonelbre]

> Do you see any way other than the 3 ones discussed in the OP article under the label "Generics"?

The one I mainly meant was interfaces. Of course there are also DSLs and code generation.

> I could agree with that if time wasn't limited, but imho building custom data structure/algorithms for each problem will precisely leave less time to optimize for performance since reaching a working state first will be longer without reusable abstractions.

I mentioned that in the summary. It's about tradeoffs. Even if all the cons/pros are listed people could still come to different conclusions, because people value different things. And there's nothing wrong about it.

The amount you need to reimplement depends on the amount of available third-party packages - the more good packages there are the less you need to build those custom data structures.