I'm not sure it's perfect, but it fixes the original issue and all the tests
pass.
The problem was in entering with locate (or some version of it) the same
span twice. So the algorithm saw an element with identical enclosing element
and the ‘commentFollowsElt’ function got confused.
The solution is two first augment registration of enclosing spans to allow
us to keep more than one item there. Then we can filter out spans which are
just copies of current reference span to get to the “real” enclosing span.
It seems to make sense to attach comments to child element if it starts
exactly at the same position of parent element, hence the distance between
start of enclosing/reference span and start of comment is the same, thus
changing ‘>’ to ‘>=’ in ‘commentFollowsElt’.
‘ghc-exactprint’ (or perhaps lexer of GHC itself) does a fairly poor job at
associating comments with elements of AST. In many cases the result is not
what you'd expect. We ran into insuperable problems with that to the effect
that correct comment placement were impossible.
The new approach is to exploit the raw position information provided by the
GHC lexer, that is, spans attached to AST elements and comments. This
allowed us to place comments in output in a very precise and satisfactory
fashion.
