fixup for #1602 - deprecate fixed cycle check

[MartinNowak]

• only print a deprecation warning for the fixed cycle detection
  cases (cycles over modules without ctor/dtor)

[schveiguy]

Looks close, but the issue is not that cycles with a non-ctor are a problem, the issue is with a cycle that goes through the same non-ctor module twice. The previous algorithm skipped that path.

So for instance with module a and c having ctors, and module b not, the following cycle was previously correctly detected:

a -> b -> c -> a

But this was not:

a -> b -> c -> b -> a.

I think that's the only thing to look for, a module cycle with 2 of the same non-ctor module in it. There could be more cases, but I think it's close enough to be acceptable.

If you want to add this extra detection, I will merge it.

[MartinNowak]

That doesn't fix Higgs, I still get another fatal cycle after restricting the deprecation to the first one.

Deprecated: Cyclic dependency between module ir.inlining and ir.iir
ir.inlining* ->
ir.ast ->
ir.iir* ->
ir.ops ->
jit.ops* ->
ir.ast ->
ir.inlining*
object.Error@src/rt/minfo.d(346): Cyclic dependency between module jit.ops and ir.iir
jit.ops* ->
ir.ast ->
ir.iir* ->
ir.ops ->
jit.ops*

The problem was more generally that modules w/o ctors were immediately sorted in before all their deps were checked, so the module could already occur outside of the cycle.
a -> b* -> a -> c* -> d -> b*
Doesn't seem like we can detect this from the current implementation. Also running somewhat out of time on this.

[MartinNowak]

An additional bitvector deprecatedCtorDone for non-relevant modules would work, but not that elegant.

[schveiguy]

There also is the issue of the ordering of the modules. I believe if you reordered the modules, in some cases a cycle was then detected.

I think it's probably inelegant, but we may want to just run the old algorithm if we see any cycles, and act accordingly.

[MartinNowak]

I think it's probably inelegant, but we may want to just run the old algorithm if we see any cycles, and act accordingly.

Yes, that was what I wanted to do, but then had this idea. Would you have a bit time for that?

[schveiguy]

I will look into doing this later tonight if I can.

[schveiguy]

OK, I think this should do it: #1720