[clean-list] Serious bug in the type system

[John van Groningen]

Edsko de Vries wrote:
>...
> > For unique arrays with unboxed values this may not be necessary, it may be
> > correct to type these as observers.
>
>I don't understand your last remark; with unboxed values, 'ohoh' is
>already rejected by the compiler:
>
>  ohoh :: *{Char} -> *(*{Char}, {Char})
>  ohoh arr
>      #! arr` = arr
>      = (update arr 0 'a', arr`)
>
>gives
>
>  Uniqueness error [t.icl,5,ohoh]: "arr" demanded attribute cannot be
>  offered by shared object

There are no arrays with unboxed values in this example. {#Char} is an array
of unboxed evaluated characters. {Char} is an array of boxed, possibly
unevaluated characters.

>This makes sense, since unboxed values can have thunks inside them, thus
>violating the hyperstrict condition -- no?

Yes, your example violates the hyperstrict condition, and is therefore
rejected by the compiler.

An example of a unique array with (only) unboxed values for which typing
as an observer may be correct is:

copy2 :: !*{#Int} -> (!*{#Int},!*{#Int})
copy2 a1
    #! a2=copy a1
    = (a2,a1)

copy :: {#Int} -> *{#Int}
copy a = {e\\e<-:a}

The Clean 2.2 compiler allows this because of the bug. After fixing the bug
it is rejected. This is not necessary, because the result of copy must be a
newly created array, because no unique array is passed to function copy.

Kind regards.

John van Groningen