Pattern for (or possibility of) tuple.set ?

[JaimeGensler]

Hello! As far as I've been able to figure, if I have a variable that's a tuple type, the only way I could update a single part of the tuple is to remake the entire thing with the single member adjusted (I use the JS wrapper currently). Is there any possibility of e.g. a tuple.set instruction in the future to update a single member of a tuple? Or is there already a better pattern to accomplish this that I'm just not familiar with / is this discouraged in general for some reason?

[tlively]

The problem with tuples and tuple instructions is that they don't exist in the underlying WebAssembly, so Binaryen's binary writer has to do a complicated lowering involving scratch locals to emit them. Because of this complexity, there's a high bar for adding new tuple instructions. The best reason for adding a new tuple instruction would be if it allowed us to emit better WebAssembly for common patterns.

I'm assuming that tuple.set would take an immediate index, a tuple operand, and the replacement value, and would return the modified tuple. The higher the replaced index is, the better the code generated from tuple.set would be compared to reconstructing the whole tuple, but when replacing multiple indices, it would generally be better to reconstruct the tuple just once. I doubt this would be worth the complexity of adding the new instruction.

An alternative would be a local.tuple.set instruction that takes the index of a tuple-typed local, an index in the tuple, and a single replacement value. This would generate very efficient code (just a single local.set), but you could achieve the same effect by storing the tuple elementwise in multiple locals to begin with. I'm not sure this is worth it either.

In general we don't want to incentivize using tuples by adding a bunch of complexity to make them easier to use. Binaryen can't optimize tuples very well, so it's best to use them only when absolutely necessary.

[JaimeGensler]

Thank you for the thoughtful answer! Yes, that's roughly what I had in mind for tuple.set (or local.tuple.set), but I see why it'd be difficult/not a good idea to support.

One follow-up question: if I store tuples into multiple locals, is there a recommended pattern for deconstructing a multivalue return from a function (e.g. a func (result f32 f32 f23) where I want to put members into locals x, y, and z) using the binaryen wrapper APIs / s-expression format? In plain linear-instruction-list WASM I'd think the ideal way would be to call the function and then local.set z, local.set y, local.set x, but for Binaryen's s-expressions the best approach I can think of is creating a local variable that's an (f32 f32 f32) to hold the entire return and then extracting each tuple element into the corresponding local? Am I missing something obvious here?

Cheers, and thanks for your time 🙂

[tlively]

No, you're not missing anything obvious. The pattern you described is the best you can do with the functionality we have today.