Version française
Home     About     Download     Resources     Contact us    
Browse thread
Re: cl7
[ Home ] [ Index: by date | by threads ]
[ Search: ]

[ Message by date: previous | next ] [ Message in thread: previous | next ] [ Thread: previous | next ]
Date: -- (:)
From: Pierre Weis <weis@p...>
Subject: Re: cl7
There has been a lot of questions about the new type-checking
algorithm for mutable values incorporated in the new 0.7 release of
Caml Light. In this message, I try to explain the new type-checking
rule. This rule is sound and in some cases allows more programs than
the old one. However there exists programs that become ill-typed. This
situation rarely happens: the 7000 lignes of the Caml Light compiler
were type-checked without modifications by the new type-checker.

I'll give some hints to translate problematic programs into typable ones.

First, note that the scheme described here is the simplest known
scheme to get a sound type system in a polymorphic langage with
mutable variables.

The problem and its solution:
-----------------------------
The scheme changes the polymorphism, since it relies on a modification
of the generalisation rule for let bound identifiers. Let's recall
this old rule. When typing

let my-name = expr1 in expr2

expr1 is type-checked first, and the identifier my-name gets a
type-scheme (a polymorphic type) according to the type of expr1:
namely, the type variables introduced when type-checking expr1 that remain
unconstraint when expr1 is fully type-checked, are generalized, that is
become parameters of the type scheme assigned to my-name. For instance in

let id = function x -> x in ...

id gets type scheme: for all 'a. 'a -> 'a, thus can be used with
several incompatible types in the rest of the program. Fort instance,
with type int -> int and bool -> bool if the in part is id 1; id true

Unfortunately this rule is unsound when mutable data exist in the
language (such as references, vectors, or records with mutable
fields). In effect, consider

let bad_ref = ref [] in ...

with the old let-rule, bad_ref gets type scheme for all 'a. 'a list
ref. It thus can be used with different types in the rest of the
program, and this is boguous. For instance, if we can use bad_ref with
type int list ref and bool list ref, we can write:

let bad_ref = ref [] in
    bad_ref := [1];
    if hd (!bad_ref) then ...

    (bad_ref := [1] is correctly typed, since bad_ref can be used with
     type int list ref,
     hd (!bad_ref) is correctly typed, since bad_ref can be used with
     type bool list ref,)
This program leads to a bus error since after the assignment, bad_ref
contains a list of integers and not a list of booleans.

The new rule for let-bound identifiers:
---------------------------------------
In short, to be safe a type system for Caml must prohibit the
existence of polymorphic mutable values at runtime. This has been
intensively studied in the ML community, and many complex type-systems
have been proposed. Recently it appears that a very simple
modification of the let-rule is sufficient: when typing

let my-name = expr1 

we use the same old let-rule only when expr1 is a function, an
identifier, or a constant. Otherwise the identifier my-name is not
polymorphic (type variables are not generalized).

Which programs must change?
---------------------------
The main change with the old let-rule concerns definition of variables
which are bound to an application:
let bad_ref = ref [] in
bad_ref is not polymorphic, and this is good.

But some other programs are now monomorphic:

let good_old = map (function x -> x) in ...

since good_old is an application it is not generalized and cannot be
used with different types. Thus
let good_old = map (function x -> x) in good_old [1]; good_old [true]
is now ill typed (good_old [1] imposes to good_old to be of type int
list -> int list, and good_old [true] is ill-typed).

How to recover?
---------------
In this case (the most common one) it is very easy to get an
equivalent well-typed program: since good_old is in fact a function,
you just have to make it explicit for the type-checker by adding a
``function construct'' or an extra parameter (this transformation is
technically known as eta-expansion):

let good_old l = map (function x -> x) l in ...

or 

let good_old = function l -> map (function x -> x) l in ...

Strange types and messages from the type-checker
------------------------------------------------

When identifiers are bound to non-polymorphic types with unknown type
variables this type variables are instnaciated when the identifier is used.
Consider

let good_old = map (function x -> x) in
good_old [1]

After its definition good_old gets type 'a list -> 'a list (where 'a just
stands for an unknown type, this is NOT the type scheme for all 'a. 'a
list -> 'a list). Hence the type-checking of good_old [1] will resolve
this unknown type to be int (and good_old gets type int list -> int
list).

Unknowns or non-polymorphic type variables:
-------------------------------------------
For clarity, an unknown type variable is denoted by '_a instead of 'a
for a parameter of a type scheme.
If we break the preceding program into 2 top-level phrases, we get:

#let good_old = map (function x -> x);;
good_old : '_a list -> '_a list = <fun>
#good_old [1];;
- : int list = [1]

Once more the unknown type '_a is known to be int, and thus the type
of good_old becomes completely determined:

#good_old;;
- : int list -> int list = <fun>

For specialists only: why is this new rule correct?
----------------------------------------------
The new rule only generalizes identifiers bound to expressions that
cannot create polymorphic values. Since mutable values are always
monomorphic, it is not very difficult to prove that the type system is sound.

Pierre Weis
----------------------------------------------------------------------------
Projet Cristal
INRIA, BP 105, F-78153 Le Chesnay Cedex (France)
E-mail: Pierre.Weis@inria.fr
Telephone: +33 1 39 63 55 98
----------------------------------------------------------------------------