closes #8895 add std/once

[ringabout]

fix #8895
add thread-safe once to std/once, for ORC.

[timotheecour]

what about re-entrant code? do we need a ronce / reentrantOnce (rlocks), or can once be made re-entrant safe without overhead?

[dom96]

Instead of implementing a new module for this, can't we fix system.once instead?

[timotheecour]

more idiomatic:

for t in mitems(thr):
  createThread(t, threadFunc)

(and Example code in

Nim/lib/system/threads.nim

Line 41 in 73299b0

## for i in 0..high(thr):

should be updated as well in future work)

ditto below

[timotheecour]

Instead of implementing a new module for this, can't we fix system.once instead?

if it can be done transparently for user code and it doesn't add new dependencies to system.nim, yes, but with the way it's done in this PR at least, this isn't the case as it requires an extra parameter Once so would require users to change their code anyways. So it's better to delegate this to another module and avoid bloating system.nim further.

• However existing system.once should be marked as deprecated in the same PR, just like it was the case for addQuitProc in addQuitProc now works with closures, and c, js(node/browser) backend; fix some bugs in testament #14342

[timotheecour]

else: static doAssert false
(prevents silent misuse)

[timotheecour]

more idiomatic:

• var thr: array[5, Thread[void]]
• and as https://github.com/nim-lang/Nim/pull/16192/files#r540547170 upstream code in threads.nim should be updated

[timotheecour]

doAssert count == thr.len

[timotheecour]

here you actually don't need matrix for that since you can combine both:

block:
  var thr: array[0..4, Thread[void]]
  ...

block:
  var thr: array[0..4, Thread[void]]
  ...

but instead, use matrix for:

matrix: "--threads:off; --threads:on"

and adjust code to work with and without threads with compileOption. There are ways to do it without any code duplication.

[timotheecour]

js needs to be supported because system.once already works in js (not considering edge case of webworkers)

(and leave a TODO for supporting VM because VM already doesn't work with system.once)

when true:
  var count = 0
  var countCT {.compileTime.} = 0
  proc fn(n: int) =
    once:
      when nimvm:
        countCT.inc
      else:
        count.inc
      echo (n,)
    if n > 1:
      fn(n-1)
  proc main() =
    fn(5)
    fn(4)
    when nimvm:
      echo countCT
      # doAssert countCT == 1 # fails
    else:
      doAssert count == 1
  static: main()
  main()

[timotheecour]

we also need a test where the code in once block raises

[timotheecour]

I'm wondering whether we should customize whether cond.finished.store(true, moRelease) happens in a finally block or only after a successful body that didn't raise.

there are use cases for both.
eg:

var db: MyDataBase
proc fn=
  once(Once, retryOnFailure = true):
    db = newMyDataBaseFlaky() # suppose this fails 10% of the times
  # we want to make sure `db` is always initialized here
  db.call() # if `newMyDataBase` raised in another thread, `db` will be nil

without retryOnFailure flag, how would you emulate retryOnFailure in user code calling std/once ?

[juancarlospaco]

Should be named std/onces, maybe it can have more than one once overload, etc.

[juancarlospaco]

Can be func ?.

[konsumlamm]

Should be named std/onces, maybe it can have more than one once overload, etc.

onces doesn't make much sense, it's not even a word (once is an adverb, not a subject).

[juancarlospaco]

What word is jsffi, cgi, coro, etc 🤷

[planetis-m]

My gripe with this implementation is that a user needs to init once explicitly. This is unusual and can cause UB if its forgotten. An easy way is using a spinlock More advanced implementations are in rust and abseil