Type safety

Type systems
Dynamic vs. Static Inferred vs. Manifest Nominal vs. Structural Strong vs. Weak Dependent typing Duck typing Latent typing Substructural typing Type safety Uniqueness typing

In computer science, type safety is the extent to which a programming language discourages or prevents type errors. A type error is erroneous or undesirable program behaviour caused by a discrepancy between differing data types for the program's constants, variables, and methods (functions), e.g., treating an integer (int) as a floating-point number (float). Type safety is sometimes alternatively considered to be a property of a computer program rather than the language in which that program is written; that is, some languages have type-safe facilities that can be circumvented by programmers who adopt practices that exhibit poor type safety. The formal type-theoretic definition of type safety is considerably stronger than what is understood by most programmers.

Type enforcement can be static, catching potential errors at compile time, or dynamic, associating type information with values at run-time and consulting them as needed to detect imminent errors, or a combination of both.

The behaviors classified as type errors by a given programming language are usually those that result from attempts to perform operations on values that are not of the appropriate data type. This classification is partly based on opinion; some language designers and programmers argue that any operation not leading to program crashes, security flaws or other obvious failures is legitimate and need not be considered an error, while others consider any contravention of the programmer's explicit intent (as communicated via typing annotations) to be erroneous and not "type-safe."

In the context of static (compile-time) type systems, type safety usually involves (among other things) a guarantee that the eventual value of any expression will be a legitimate member of that expression's static type. The precise requirement is more subtle than this — see, for example, subtype and polymorphism for complications.

Type safety is closely linked to memory safety, a restriction on the ability to copy arbitrary bit patterns from one memory location to another. For instance, in an implementation of a language that has some type $t$, such that some sequence of bits (of the appropriate length) does not represent a legitimate member of $t$, if that language allows data to be copied into a variable of type $t$, then it is not type-safe because such an operation might assign a non-$t$ value to that variable. Conversely, if the language is type-unsafe to the extent of allowing an arbitrary integer to be used as a pointer, then it is clearly not memory-safe.

Most statically typed languages provide a degree of type safety that is strictly stronger than memory safety, because their type systems enforce the proper use of abstract data types defined by programmers even when this is not strictly necessary for memory safety or for the prevention of any kind of catastrophic failure.

Definitions

Type-safe code accesses only the memory locations it is authorized to access. (For this discussion, type safety specifically refers to memory type safety and should not be confused with type safety in a broader respect.) For example, type-safe code cannot read values from another object's private fields.

Robin Milner provided the following slogan to describe type safety:

Well-typed programs cannot "go wrong".

The appropriate formalization of this slogan depends on the style of formal semantics used for a particular language. In the context of denotational semantics, type safety means that the value of an expression that is well-typed, say with type τ, is a bona fide member of the set corresponding to τ.

In 1994, Andrew Wright and Matthias Felleisen formulated what is now the standard definition and proof technique for type safety in languages defined by operational semantics. Under this approach, type safety is determined by two properties of the semantics of the programming language:

(Type-) preservation or subject reduction

"Well typedness" of programs remains invariant under the transition rules (i.e. evaluation rules or reduction rules) of the language.

Progress

A well typed program never gets "stuck", i.e., never gets into an undefined state where no further transitions are possible.

These properties do not exist in a vacuum; they are linked to the semantics of the programming language they describe, and there is a large space of varied languages that can fit these criteria, since the notion of "well typed" program is part of the static semantics of the programming language and the notion of "getting stuck" (or "going wrong") is a property of its dynamic semantics.

Vijay Saraswat provides the following definition:

"A language is type-safe if the only operations that can be performed on data in the language are those sanctioned by the type of the data."

Relation to other forms of safety

Type safety is ultimately aimed at excluding other problems, e.g.:-

• Prevention of illegal operations. For example, we can identify an expression 3 / "Hello, World" as invalid, because the rules of arithmetic do not specify how to divide an integer by a string. As discussed below, strong typing offers more safety, but generally does not guarantee complete safety.
• Memory safety
  • Wild pointers can arise when a pointer to one type object is treated as a pointer to another type. For instance, the size of an object depends on the type, so if a pointer is incremented under the wrong credentials, it will end up pointing at some random area of memory.
  • Buffer overflow - Out-of bound writes can corrupt the contents of objects already present on the heap. This can occur when a larger object of one type is crudely copied into smaller object of another type.
• Logic errors originating in the semantics of different types. For instance, inches and millimeters may both be stored as integers, but should not be substituted for each other or added. A type system can enforce two different types of integer for them.

Type-safe and type-unsafe languages

Type safety is usually a requirement for any toy language proposed in academic programming language research. Many languages, on the other hand, are too big for human-generated type-safety proofs, as they often require checking thousands of cases. Nevertheless, some languages such as Standard ML, which has rigorously defined semantics, have been proved to meet one definition of type safety. Some other languages such as Haskell are believed to meet some definition of type safety, provided certain "escape" features are not used (for example Haskell's unsafePerformIO, used to "escape" from the usual restricted environment in which I/O is possible, circumvents the type system and so can be used to break type safety.) Type punning is another example of such an "escape" feature. Regardless of the properties of the language definition, certain errors may occur at run-time due to bugs in the implementation, or in linked libraries written in other languages; such errors could render a given implementation type unsafe in certain circumstances. An early version of Sun's Java Virtual Machine was vulnerable to this sort of problem.

Type safety and strong typing

Type safety is synonymous with one of the many definitions of strong typing; but type safety and dynamic typing are mutually compatible. A dynamically typed language such as Smalltalk can be seen as a strongly typed language with a very permissive type system where any syntactically correct program is well-typed; as long as its dynamic semantics ensures that no such program ever "goes wrong" in an appropriate sense, it satisfies the definition above and can be called type-safe.

Type safety issues in specific languages

Ada

Ada was designed to be suitable for embedded systems, device drivers and other forms of system programming, but also to encourage type safe programming. To resolve these conflicting goals, Ada confines type-unsafety to a certain set of special constructs whose names usually begin with the string Unchecked_. Unchecked_Deallocation can be effectively banned from a unit of Ada text by applying pragma Pure to this unit. It is expected that programmers will use Unchecked_ constructs very carefully and only when necessary; programs that do not use them are type safe.

The SPARK programming language is a subset of Ada eliminating all its potential ambiguities and insecurities while at the same time adding statically checked contracts to the language features available. SPARK avoids the issues with dangling pointers by disallowing allocation at run time entirely.

C

The C programming language is typesafe in limited contexts; for example, a compile-time error is generated when an attempt is made to convert a pointer to one type of structure to a pointer to another type of structure, unless an explicit cast is used. However, a number of very common operations are non-typesafe; for example, the usual way to print an integer is something like printf("%d", 12), where the %d tells printf at run-time to expect an integer argument. (Something like printf("%s", 12), which erroneously tells the function to expect a pointer to a character-string, will be accepted by compilers, but will produce undefined results.) This is partially mitigated by some compilers (such as gcc) checking type correspondences between printf arguments and format strings.

In addition, C, like Ada, provides unspecified or undefined explicit conversions; and unlike in Ada, idioms that use these conversions are very common, and have helped to give C a type-unsafe reputation. For example, the standard way to allocate memory on the heap is to invoke a memory allocation function, such as malloc, with an argument indicating how many bytes are required. The function returns an untyped pointer (type void *), which the calling code must cast to the appropriate pointer type. Older C specifications required an explicit cast to do so, therefore the code (struct foo *) malloc(sizeof(struct foo)) became the accepted practice. However, this practice is discouraged in ANSI C as it can mask a failure to include the header file in which malloc is defined, resulting in downstream errors on machines where the int and pointer types are of different sizes, such as the now-ubiquitous x86_64 architecture. A conflict arises in code that is required to compile as C++, since the cast is necessary in that language.

C++

Some features of C++ that promote more type-safe code:

• The new operator returns a pointer of type based on operand, whereas malloc returns a void pointer.
• C++ code can use virtual functions and templates to achieve polymorphism without void pointers.
• Preprocessor constants (without type) can be rewritten as const variables (typed).
• Preprocessor macro functions (without type) can be rewritten as inline functions (typed). The flexibility of accepting and returning different types can still be obtained by function overloading.
• Safer casting operators, such as dynamic_cast that performs run-time type checking.

C#

C# is type-safe (but not statically type-safe). It has support for untyped pointers, but this must be accessed using the "unsafe" keyword which can be prohibited at the compiler level. It has inherent support for run-time cast validation. Casts can be validated by using the "as" keyword that will return a null reference if the cast is invalid, or by using a C-style cast that will throw an exception if the cast is invalid. See C Sharp Conversion Operators.

Undue reliance on the object type (from which all other types are derived) runs the risk of defeating the purpose of the C# type system. It is usually better practice to abandon object references in favour of generics, similar to templates in C++ and generics in Java.

Cyclone

Cyclone is a type-safe language. It does not require a virtual machine or garbage collection to achieve type safety during runtime. The syntax is very similar to C.

Standard ML

SML has rigorously defined semantics and is known to be type-safe. However, some implementations of SML, including Standard ML of New Jersey (SML/NJ), its syntactic variant Mythryl and Mlton, provide libraries that offer certain unsafe operations. These facilities are often used in conjunction with those implementations' foreign function interfaces to interact with non-ML code (such as C libraries) that may require data laid out in specific ways. Another example is the SML/NJ interactive toplevel itself, which must use unsafe operations to execute ML code entered by the user.

Pascal

Pascal has had a number of type safety requirements, some of which are kept in some compilers. Where a Pascal compiler dictates "strict typing", two variables cannot be assigned to each other unless they are either compatible (such as conversion of integer to real) or assigned to the identical subtype. For example, if you have the following code fragment:

type
  TwoTypes = record
     I: Integer;
     Q: Real;
  end;
 
  DualTypes = record
    I: Integer;
    Q: Real;
  end;
 
var
  T1, T2:  TwoTypes;
  D1, D2:  DualTypes;

Under strict typing, a variable defined as TwoTypes is not compatible with DualTypes (because they are not identical, even though the components of that user defined type are identical) and an assignment of T1 := D2; is illegal. An assignment of T1 := T2; would be legal because the subtypes they are defined to are identical. However, an assignment such as T1.Q := D1.Q; would be legal.

Common Lisp

Although Common Lisp is generally dynamically typed and type-safe, the language allows to explicitely declare type annotations on variables (strong typing), and to use machine words (fixnum)[1]. Declarations can also explicitely lower the type safety level for wanted code sections[2]. When such declarations are present and the compiler of an implementation of Common Lisp observes them, operations can become as unsafe as they could be in C or machine language.

C++ Examples

The following examples illustrates how C++ cast operators can break type safety when used incorrectly. The first example shows how basic data types can be incorrectly casted:

#include <iostream>
using namespace std;
 
int main () {
    int   ival = 5;                              // integer value
    float fval = reinterpret_cast<float&>(ival); // reinterpret bit pattern
    cout << fval << endl;                        // output integer as float
    return 0;
}

In this example, reinterpret_cast explicitly prevents the compiler from performing a safe conversion from integer to floating-point value. When the program runs it will output a garbage floating-point value. The problem could have been avoided by instead writing float fval = ival;

The next example shows how object references can be incorrectly downcasted:

#include <iostream>
using namespace std;
 
class Parent {
public:
    virtual ~Parent() {} // virtual destructor for RTTI
};
 
class Child1 : public Parent {
public:
    int a;
};
 
class Child2 : public Parent {
public:
    double b;
};
 
int main () {
    Child1 c1;
    c1.a = 5;
    Parent & p = c1;                     // upcast always safe
    Child2 & c2 = static_cast<Child2&>(p); // invalid downcast
    cout << c2.b << endl;          // will output garbage data
    return 0;
}

The two child classes have members of different types. When downcasting a parent class pointer to a child class pointer, then the resulting pointer may not point to a valid object of correct type. In the example, this leads to garbage value being printed. The problem could have been avoided by replacing static_cast with dynamic_cast that throws an exception on invalid casts.

See also

• Type theory

References

• Pierce, Benjamin C. (2002). Types and Programming Languages. MIT Press. ISBN 0-262-16209-1. 
• "Type Safe". Portland Pattern Repository Wiki. 
• Wright, Andrew K.; Matthias Felleisen (1994). "A Syntactic Approach to Type Soundness". Information and Computation 115 (1): 38–94. doi:10.1006/inco.1994.1093. 
• Macrakis, Stavros (April 1982). "Safety and power" (requires subscription). ACM SIGSOFT Software Engineering Notes 7 (2): 25–26. doi:10.1145/1005937.1005941.