GitLab

Make it so that the substitution function preserves as much sharing as possible from the original type.

This reverts commit a8ba6ab6.

This commit introduces a regression where suprious type variables are introduced in the final type.

Seal the representation of SortedList.Make(X).t (by making the type private). Expose Var.Set as a SortedList.S

Perform a more aggressive memoization of substituted types and register the name if the original type was named.
Make the pretty-printer re-entrant.

remain regular. Within their recursive definitions, parametric types
must always be instantiated with their original parameters and all
types of mutually recursive definitions must have the same parameters.

We use Tarjan's strongly connected components algorithm to group type definitions accordingly.

[ 'abc'ab ] is now parsed as [ 'abc' ab ] to maintain backward compatibility.
Change the warning message accordingly.

correctly handle insantiated types in regular expressions.

suggest to put a space before the second quote.

Move the auxiliary 'warning' function from the typechecker to the
Cduce_loc module.

Re-indent some comments in ulexer.ml

"\n" after the ;; token.

to use two lexers (depending on whether we are between square brackets
or not) is too brittle (it crudely tries to parse
 ``( [whitespace] 'a  [whitespace] )'' as a variable, to force the user
to write the variable beetween parenthesis. However this does not scale
to types with two arguments (says [ t ('a, 'b) ]).

We use a simpler heuristic (with look ahead)

(1) try to see if the regular expression

' (anything but ', \n)* '(anything but the first letter of an identifier)

can be found. If so, we put back the lexeme in the buffer and parse it as as
a string.

(2) if (1) failed, try to parse it as a variable

(3) if (3) failed, try to parse it again as a string. We are
guaranteed to fail here but it means we have a malformed string, so we
parse as a string to get a proper error message.

The only thing this does not cover are cases like
type t = [ 'abcd'Int ]
which was tokenized before as [, 'abcd', Int, ]
and is now tokenized as [, 'abcd, 'Int, ]
It does not seem to be a problem in practice though (since in the code
I have seen thus far, people were at least putting a space).
it is easy to emmit a warning in this case, suggesting the user to add
a whitespace to get the old behaviour back.

requiring an abscence of whitespace between ``t'' and ``(''.  Outside
of regular expression contexts, ``t (t1, …, tn)'' is parsed with a
higher precedence than & and \, to allow one to write
``t (Int) & Bool'' without extra parentheses (i.e.
``(t (Int)) & Bool'').  Inside a regular expression, type
instantiation and sequencing become ambiguous, and there is no way to
distinguish syntactically: ``[ Int (Int, Int) ]'' from
``[ t (Int, Int) ]''. The former should resolve to a sequence while
the latter only makes sense as an instantiation (if ``t'' is a
parametric type). Both are treated as element sequencing and
disambiguated during identifier resolution (more precisely during the
"derecurse" phase, before typechecking).

Note that due to the lower precedence of sequencing w.r.t to other
regular expression constructs, a type ``[ t (Int)* ]'' will be parsed
correctly, but yield an error message saying that t is not fully
intantiated. One has to write ``[ (t (Int))* ]'' which is similar to
function applications for expressions.

Finally, we also re-order sequencing after typing to always group a
potential type instantiation with its argument, i.e. we transform
sequences such as
``[ t1 t2 t3 ... tn ]'' (which are parsed as
``[ (((t1 t2) t3) ... tn) ]'' because sequence concatenation is
left-associative) into ``[ ... (ti tj) ... ]'' if ``ti'' is an
identifier and ``tj'' is of the form ``(s1,...,sk)''. This is sound
because concatenation of regular expression is associative (and the
original sequence would fail, anyway).

containing variables but no toplevel variables would not be printed as
regular expressions but nested pairs.

Now functions such as
let fun f (x:Int):Int = ...
are parsed correctly (that is, its parameter is 'x' with type Int).

sequence type would not show up (e.g. 'a & [ Any *] would be displayed
as [ Any * ], while the 'a would still be present internally).

compiles with the patched camlp4 bundled with fedora 21)

Positive.solve(Positive.decompose t) was not the identity but added spurious types at toplevel.

Added more test cases in poly-ok.cd to cover such regressions