write system.abs in nim, not magic

[timotheecour]

refs nim-lang#16432 (comment)

func abs*[T: SomeSignedInt](x: T): T {.inline.} =
  ## Returns the absolute value of `x`.
  ##
  ## If `x` is ``low(x)`` (that is -MININT for its type),
  ## an overflow exception is thrown (if overflow checking is turned on).
  if x < 0: -x else: x

and remove mAbsI magic, and then, in subsequent work, investigate whether the branchless bit hack https://graphics.stanford.edu/~seander/bithacks.html#IntegerAbs can improve performance in at least some cases ?

[ringabout]

Implementing them as magics makes them could be folded in compile time

  of mAbsI: result = foldAbs(getInt(a), n, g)

[timotheecour]

does it matter though? why would abs be singled-out as being special? I have a feeling this is for historical reasons that aren't relevant anymore; I have a WIP branch but no PR yet.

[ringabout]

abs(-4) is folded into 4?
Besides this micro optimization, mAbs could be removed safely.

[ringabout]

system.min and system.max are also magics, do I need to clean them?

[timotheecour]

indeed, const folding does happen for max/min as can be seen by looking at codegen for:

when true:
  func max2[T](a, b: T): T {.inline.} =
    if a<b: b else: a
  proc main()=
    let a1 = max(11+1,12+2)
    let a2 = max2(11+1,12+2)
    echo (a1, a2)
  main()

=> codegen shows (with nim r -d:danger main)

	a1 = ((NI) 14);
	a2 = max2__g9bYy9c5l5m9b6EyjAm9ac4Y9bQt11587(((NI) 12), ((NI) 14));

It feels very odd to have these hard-coded cases in the compiler.

Even if const folding is desired (which should be justified; C compiler should already take care of this; I can't see how this can possibly generate different optimized binary code), why should a magic be used for that? Instead of, say, a flag (maybe enabled by --opt:speed|size or preferably a dedicated one eg --constfolding) which would control whether expressions can safely be const folded without changing semantics, as should be the case with max2.

I think for abs the cleanup doesn't require an RFC (abs is relatively rare) but maybe it's needed for other magics-that-shouldn't-be-magics like max,min etc. I'm curious in fact how many of them there are.

[timotheecour]

regarding symbols that trigger const folding, IIUC it's handled by ctfeWhitelist in compiler

# things that we can evaluate safely at compile time, even if not asked for it:
const
  ctfeWhitelist* = {mNone, mSucc, ...

[ringabout]

what's the progress?