@chicoxyzzy Petr Прочел заметки с последнего SIMD subgroup. И вот этот диалог:
MD: original SIMD proposal is pretty complete, but one area we can improve is to handle multiple values at once, e.g. extend mul of 2 vectors, produce both high and low, can be more efficient than current. Separate low and high part.
ZN: If we have multiple returns, what does it look like in the intrinsic functions?
MD: Neon has additional data types, Wasm intrinsics could also have v128x2 values
LH: looked at this in add with carry out, it becomes hairy because Wasm is a stack machine, you always get one result underneath and the other on top, sometimes that is a wrong order, so you have variants of instruction that produces results in one order or the other, to avoid moves. It’s not completely straightforward
ZN: We don’t have a swap, if we have that instruction, but maybe we could make it better?
LH: Not sure, but we definitely would need some experimentation.