Would you be able to make this work if your function signature took a pointer to the output? Like:

shuffle(mask, data []byte, out *byte)

I may be able to extend the gotypes.Component interface to make this possible. I'd need to think about this some more.

https://pkg.go.dev/github.com/mmcloughlin/avo@v0.0.0-20200425082757-f891fe8d9749/gotypes?tab=doc#Component

@mmcloughlin Thanks for the quick response. For now, I would prefer use Store(xmm, ReturnIndex(0).Index(0)) if it works, since [4]uint32 is a value type.

@klauspost has encountered the same problem:

I am trying to generate a function in avo that has func(data [n][16]byte) signature, and I'm trying to load the [16]byte into YMM registers.
My problem is that if I use Load(Param("data").Index(i), YMM()) (and XMM) I get component is not primitive and I cannot find a way to get it back to a VecVirtual anyway. I would like to send values directly on the stack, since slices are bigger and pointers is just an extra indirection.
I also tried to resolve the address so I could just use a MOVOU myself, but couldn't make that work either.

https://gophers.slack.com/archives/C6WDZJ70S/p1589142748014600

As discussed on Slack, probably the solution here is to provide an accessor method Component.Addr() for:

The only question in my mind is what to do with errors that can occur in method chaining, like when i is too large in Param("data").Index(i).Addr(). Should the signature be:

1. Addr() operand.Mem which panics when there is an error
2. Addr() (operand.Mem, error)

If it returns an error, a Must function might be useful. This could possibly belong in the operand package.

@mmcloughlin Thanks for the update. I did what you suggested for the time being.

func(masks, data []byte, offset int, out []uint32)

https://github.com/rleiwang/svb/blob/10f63e6855cff78e14a0371425cbf569e8456590/asm.go#L11