[valyala]

The sync.Map is a good example , which could benefit from switching to generics. The problem is that this is the only useful case for generics. This case could be implemented in the way similar to built-in map[k]v generic type, which is available in Go since the first public release. And this could have prevent from opening the Pandora box with the generics.

[0x62c1b43e]

It’s not the only one. Some other packages that used workarounds like interface{} or other things to work around the lack of generics were container/{heap,list,ring}, sort, golang.org/x/sync/singleflight, /x/exp/{maps,slices}, etc. And people will want to write their own patterns at other times of course too. It wouldn’t be reasonable for these all to become builtin types like map. These standard library packages that already exist will also become more efficient and potentially reduce allocations (when using primitives) as well.

[valyala]

The container/list and container/ring is one of the least useful packages included in the standard Go library, since they aren't used widely in Go programs. It is better from performance and readability PoV to use an ordinary slices instead of these packages in most cases.

The container/heap is more useful, but it could benefit more from adding an optimization for inlining interface method calls when Go compiler knows the underlying implementation behind the interface.

The golang.org/x/exp/maps is useless and may be harmful [1].

The golang.org/x/exp/slices is mostly useless, except of Sort*() functions. But it would be better to apply the optimization mentioned above to standard sort.* functions instead of forcing users to switch to different Sort*() implementations in other packages.

[1] https://news.ycombinator.com/item?id=34622393

[0x62c1b43e]

> The container/heap is more useful, but it could benefit more from adding an optimization for inlining interface method calls when Go compiler knows the underlying implementation behind the interface.

This is exactly what generics do. With e.g. a heap.Heap[uint32] the compiler knows the implementation and there’s no interface method call overhead.

In order for the compiler to do this optimization, it has to know that you don’t e.g. pass a *heap.Heap[uint32] to a function expecting *heap.Heap[uint64], so the type system is what allows it to optimize.

And on top of that, now the user also gets assurance at compile time that heap.Heap[uint32].Pop returns a uint32, preventing bugs from type confusion and also so you don’t have to add type assertions everywhere you use the heap.

So now heap, sort, etc. can benefit from this improved performance; users don’t have to write wrapper types and interface implementations just so their type can be sorted; and bugs are prevented at compile time.

For [1] I posted a reply. It’s true that there are overheads with some slice-returning routines but I explained how in the reply how I viewed the tradeoffs.

[valyala]

In theory the compiler can inline interface method calls without the need to introduce generics. For example, it can detect that the customStruct is passed to the sort.Sort() in the code below, and then instantiate the sort.Sort() code for the given inlined Less(), Swap() and Len() interface calls:

    type customStruct struct { ... }
    func (cs *customStruct) Less(i, j int) bool { ... }
    func (cs *customStruct) Swap(i, j int) { ... }
    func (cs *customStruct) Len() int { ... }

    func sortMyCustomStruct(cs *customStruct) {
      sort.Sort(cs)
    }

The tricky part here is that the compiler should be careful when instantiating such calls for different interface implementations, in order to avoid generated code bloat. For example, if sort.Sort() is used for a thousand different sort.Interface implementations, then it may be not a great decision to create a thousand of distinct sort.Sort() instances for every sort.Interface implementation. But this should work OK for a dozen of distinct implementations.

[jen20]

> since they aren't used widely in Go programs

They aren't widely used because the ergonomics suck, because they aren't generic yet.