module - Surprising behavior of transparent signature ascription -

i want define own abstract type 'a foo, that, 'a ref, eqtype if 'a isn't. example:

lolcathost% poly poly/ml 5.5.2 release > signature foo = # sig #   eqtype 'a foo #   val bar : real foo #   val qux : real foo # end; signature foo = sig val bar : real foo type 'a foo val qux : real foo end > structure foo : foo = # struct #   datatype 'a wat = wat of 'a #   type 'a foo = 'a wat ref #   val bar = ref (wat 0.0) #   val qux = ref (wat 0.0) # end; structure foo : foo > foo.bar = foo.qux; val = false: bool 

so far, good! now:

> !foo.bar; val = wat 0.0: real wat > !foo.qux; val = wat 0.0: real wat 

wait. isn't wat supposed hidden? why seeing values of type real wat?

isn't wat supposed hidden?

it's "hidden" in sense code cannot refer it; since human aware of it, there's no reason repl evasive it.

why seeing values of type real wat?

why not? identifier wat not in scope, type-name still exists. there's no reason there can't values of types involve it.

(the same thing possible without signatures; like

local    datatype 'a wat = wat of 'a in    val = wat 0.0 end 

is valid, , has similar effect.)

since don't want outside code know 'a foo internally 'a wat ref, means !foo.bar , !foo.qux should not typecheck in first place.

if don't want outside code know 'a foo internally 'a wat ref, shouldn't use transparent ascription; whole point of transparent ascription outside code still knows type if signature doesn't specify (as opposed opaque ascription, outside code has what's specified in signature).

i want define own abstract type 'a foo, that, 'a ref, eqtype if 'a isn't.

unfortunately, not possible. the definition of standard ml (revised) (which defines standard ml '97, part) lays out exact set of equality types in §4.4 "types , type functions", on page 17. constructed type admits equality if (1) type name , of type arguments admit equality or (2) type name 1 denoted ref. in case, since 'a isn't equality type, , foo not ref, 'a foo doesn't admit equality.

(note: should mention definition little bit wrong here, in the standard ml basis library specifies few additional type names have same special property ref, , compilers implement these well. extension adds more hardcoding; doesn't add way programmers create additional such type names.)