[generator] less aggressive obfuscation marking.

[atsushieno]

This is a fixup for c51469e, as well as bugfix for:
https://bugzilla.xamarin.com/show_bug.cgi?id=51337

What's happening behind bug #51337 was this:

BINDINGSGENERATOR:  warning BG8800: Unknown parameter
type org.osmdroid.ResourceProxy.bitmap in method GetBitmap
in managed type OsmDroid.DefaultResourceProxyImpl.

This was, because, "bitmap" is all lowercased and was marked as obfuscated
(because, WHY NAME A CLASS ALL IN LOWERCASE!?).

Such classes should not exist, or should be marked as "obfuscated='false'".

However what bug #51337 implies is that people are not going to make
this additional markup for generator improvements.

What this generator change does is then - a hack. A hack to mark "classes
with very short names" as obfuscated, instead of "all lowercase or number".

As commented on c51469e, there isn't good way to check API element
siblings (it can check sibling names every time at all expensive
calculation). So we just count the sibling nodes and if name length is
short enough to fit within the number of classes, we mark as obfucated.

[jonpryor]

I don't understand the logic/rationale here, as 26**(last.Length-1) is possibly huge.

If last is c, then it's 26**0 or 1, so unless threshold is 0, all types with names of one character will be allowed. Perhaps that's desirable?

If last is ex, then it's 26**1 or 26, so a package would need 27 types for ex to be considered obfuscated. Huh?

If last is bitmap, then it's 26**5, or 11881376, at which point it's fair to say that no types will be considered to be obfuscated, as threshold will always be less than 11881376.

we just count the sibling nodes and if name length is short enough to fit within the number of classes, we mark as obfucated.

I'm not sure that's what this code does -- though I must admit I'm very confused by it -- and if it does do that, I'm not sure that's necessarily useful?

Perhaps a better obfuscation would be "all class names entirely lowercase that are <= 2 characters in length." Do we know of any obfuscated types that are longer than two characters?

[atsushieno]

I'm stupid at math. What I wanted was, always treat one-letter name as obfuscation result, treat two-letter names as obfuscated only when there are 26+ classes in the package, same for three-letter names only when there are 26^2+ classes, and so on... Therefore the condition should be:

(!(threshold >= Math.Pow (26, length - 1))

Note that this condition is to filter out non-obfuscated case ("return false").

I avoided simple length <= 2 because there can be FooBar.id or FooBar.in (possibly with FooBar.out). IMO that's no-go.

But you're right, that calculating Math.Pow() for lengthy string doesn't make sense (moreover, there's possible overflow...!). It should be trimmed down to, maybe 3? There wouldn't be 17576 types in a package, even if the library contains a lot of automatically generated types (I'm hesitant to set threshold at 100 or 1000 because of possibly generated types in a single package, 1000 is possible).

(The calculation above then covers "id" ? No, but when there are aa, ab, ... ic, id, ie etc. then this "id" wouldn't be distinct. At least we can say we assumed so).

[atsushieno]

Seealso: "id" in the real world use. https://developer.android.com/reference/android/R.id.html

[jonpryor]

For a heuristic, I'd suggest these as "obfuscated types":

1. All class names of length <= 2, in which the entire identifier is lowercase, e.g. aa, ab, etc.
2. All class names that start with zz (lowercase). This would capture e.g. zze, zzma, etc.

[jonpryor]

Seen in email, but not this thread (!):

Seealso: "id" in real world use. https://developer.android.com/reference/android/R.id.html

That is a good point, but the heuristic wouldn't necessarily hit this, because we already special case a class name prefix of R., so R.id would be skipped (unless you remove the R. check, but I don't think that would be a good idea).

[atsushieno]

Anything that happened in R or Manifest could happen in any namespace.

[atsushieno]

My comment was hidden under "show outdated" link :p #121 (comment)

[jonpryor]

This should also include tests. :-)

It looks like it should be possible to use tools/generator/Tests-Core/api.xml, in that modifying that file locally for me to include a <class name="aa" .../> caused aa to be skipped, so I believe that this will work.

[atsushieno]

Tests are added.