Some of the conflicts and issues in the grammar are documented here.
Why can't [t] be considered a valid atomic type? (A variant type.)
(This is related to MPR #3835.)
A class type that begins with [t] foo could continue as follows:
[t] foo -> <class_type>
Here t is understood as a variant type,
and is used as an actual parameter of the parameterized type 'a foo.
Or it could continue as follows:
[t] foo
Here t is a type (there is no variant type)
and is used as an actual parameter of the class ['a] foo.
After we have read the closing bracket and are looking ahead at foo,
we need to decide which of the above two situations we have. (The first
situation requires a reduction; the second situation requires shifting.)
But we cannot decide yet; we would need to look at the arrow -> beyond
foo in order to decide. In this example LR(2) is required; in general,
foo could be replaced with an arbitrary qualified name, so unbounded
lookahead is required.
As a result of this issue, we must abandon the idea that [t] could be
a well-formed variant type. In the syntax of atomic types, instead of:
atomic_type: LBRACKET row_field RBRACKET
we must use the more restricted form:
atomic_type: LBRACKET tag_field RBRACKET
In other words, we rule out exactly the following:
atomic_type: LBRACKET atomic_type RBRACKET