WIP | =sink is now =move

[Clyybber]

... which makes self-assignments work correctly.
Refs #11217

[Araq]

Note: I'm taking over, I really want this in 0.20.

[Clyybber]

TODO:

• handle explicit moves better??
• const temporaries
• adapt tests

EDIT: I think this elides destructors too agressively, investigating..

[Araq]

You need to patch the C codegen so that tySink can also be "pass by hidden pointer". Also this section

  if owner.kind in {skProc, skFunc, skMethod, skIterator, skConverter}:
    let params = owner.typ.n
    for i in 1 ..< params.len:
      let param = params[i].sym
      if isSinkTypeForParam(param.typ) and hasDestructor(param.typ.skipTypes({tySink})):
        c.destroys.add genDestroy(c, param.typ.skipTypes({tyGenericInst, tyAlias, tySink}), params[i])

needs to be removed, the caller is responsible for cleaning up what was passed as a sink parameter.

[Clyybber]

TODO:

• Make closures work
• Fix obj and tuple constructors of the backends: (affected test: tgcdestructors.nim)
  This is a bit tricky, see last commit for a dirty leaking unmergeable workaround.
  @Araq The problem is this:

type
  Obj* = object
    f*: seq[int]

var o = Obj(f: @[1, 2, 3])

will be transformed to:

eq__qdqSn882Av3rQxYVfnQyMg((&colontmpD_), TM_mDIVXahtYVExBX9bJLnfz0Q_76); //Copy, because const seq
T2_.f = colontmpD_;
eqmove__C1DzWNOtypwtWid2xBvXLg((&o), (&T2_));
//...
eqdestroy__39bIYiz25XeqkqfTm4wBXNg((&colontmpD_));
eqdestroy__9aQE9amztaLIv7YEpj9agHaSg((&o));

which will cause a double free segfault, since colontmpD_ is shallow copied to the field of T2_ / o.

[Clyybber]

This turned out to solve a problem that doesn't exist, and also hurt performance by 5% on avg.