Hello there, thanks for your interest in Nim!

A quick review of your code:

Padding

Instead of padding with arrays

you can do

type
  SPSCQueueInterface* = object
    head* {.align: CACHELINE_BYTES.}: Atomic[uint]
    tail* {.align: CACHELINE_BYTES.}: Atomic[uint]

I recommend that you pad the start of the data structure as well since queues are often used in an arrays of queues:

Sizing

Currently your buffer requires being a power of 2, this wastes memory for large buffers
Instead, to distinguish between the empty and the full queue, you can have your capacity run from 0 ..< 2*RealCapacity

See https://github.com/mratsim/weave/blob/master/weave/cross_thread_com/channels_spsc.md#litterature

• http://www.vitorian.com/x1/archives/370
• p3: http://www.sunburst-design.com/papers/CummingsSNUG2002SJ_FIFO1.pdf

And implementation in Nim for a serial queue: https://github.com/mratsim/weave/blob/2e528bd2/weave/datatypes/bounded_queues.nim

Misc

You don't need to return a seq here, you can return an array, it will be copied

Though the state might have torn copies.

hello, first of thank you for making this! Ive noticed you use moRelaxed everywhere, this is obviously not correct, as it leads to data races. One way to test this is with ThreadSanitizer, an example. An oversimplification, but a good starting point is to make all loads moAcquire and stores moRelease. Basically you need to do the publishing protocol. You can use moRelaxed when outside or inside a "critical section" (I think but that might also be an oversimplification), that is code between a load and a store.