Scripting: Higher Level Programming
for the 21st Century

John K. Ousterhout

Tcl Developer Xchange
2593 Coast Ave.
Mountain View, CA 94043
[email protected]

(This article appears in IEEE Computer magazine, March 1998)

Abstract
Scripting languages such as Perl and Tcl represent a very different style of programming than system programming languages such as C or JavaTM. Scripting languages are designed for "gluing" applications; they use typeless approaches to achieve a higher level of programming and more rapid application development than system programming languages. Increases in computer speed and changes in the application mix are making scripting languages more and more important for applications of the future.

Keywords: component frameworks, object-oriented programming, scripting, strong typing, system programming.

1 Introduction

For the last fifteen years a fundamental change has been occurring in the way people write computer programs. The change is a transition from system programming languages such as C or C++ to scripting languages such as Perl or Tcl. Although many people are participating in the change, few people realize that it is occurring and even fewer people know why it is happening. This article is an opinion piece that explains why scripting languages will handle many of the programming tasks of the next century better than system programming languages.

Scripting languages are designed for different tasks than system programming languages, and this leads to fundamental differences in the languages. System programming languages were designed for building data structures and algorithms from scratch, starting from the most primitive computer elements such as words of memory. In contrast, scripting languages are designed for gluing: they assume the existence of a set of powerful components and are intended primarily for connecting components together. System programming languages are strongly typed to help manage complexity, while scripting languages are typeless to simplify connections between components and provide rapid application development.

Scripting languages and system programming languages are complementary, and most major computing platforms since the 1960's have provided both kinds of languages. The languages are typically used together in component frameworks, where components are created with system programming languages and glued together with scripting languages. However, several recent trends, such as faster machines, better scripting languages, the increasing importance of graphical user interfaces and component architectures, and the growth of the Internet, have greatly increased the applicability of scripting languages. These trends will continue over the next decade, with more and more new applications written entirely in scripting languages and system programming languages used primarily for creating components.

2 System programming languages

In order to understand the differences between scripting languages and system programming languages, it is important to understand how system programming languages evolved. System programming languages were introduced as an alternative to assembly languages. In assembly languages, virtually every aspect of the machine is reflected in the program. Each statement represents a single machine instruction and programmers must deal with low-level details such as register allocation and procedure calling sequences. As a result, it is difficult to write and maintain large programs in assembly language.

By the late 1950's higher level languages such as Lisp, Fortran, and Algol began to appear. In these languages statements no longer correspond exactly to machine instructions; a compiler translates each statement in the source program into a sequence of binary instructions. Over time a series of system programming languages evolved from Algol, including such languages as PL/1, Pascal, C, C++, and Java. System programming languages are less efficient then assembly languages but they allow applications to be developed much more quickly. As a result, they have almost completely replaced assembly languages for the development of large applications.

System programming languages differ from assembly languages in two ways: they are higher level and they are strongly typed. The term "higher level" means that many details are handled automatically so that programmers can write less code to get the same job done. For example:

Register allocation is handled by the compiler so that programmers need not write code to move information between registers and memory.
Procedure calling sequences are generated automatically: programmers need not worry about moving arguments to and from the call stack.
Programmers can use simple keywords such as while and if for control structures; the compiler generates all the detailed instructions to implement the control structures.

On average, each line of code in a system programming language translates to about five machine instructions, compared to one instruction per line in assembly language (in an informal analysis of eight C files written by five different people, I found that the ratio ranged from about 3 to 7 instructions per line [7]; in a study of numerous languages Capers Jones found that for a given task, assembly languages require about 3-6 times as many lines of code as system programming languages[3]). Programmers can write roughly the same number of lines of code per year regardless of language[1], so system programming languages allow applications to be written much more quickly than assembly language.

The second difference between assembly language and system programming languages is typing. I use the term "typing" to refer to the degree to which the meaning of information is specified in advance of its use. In a strongly typed language the programmer declares how each piece of information will be used and the language prevents the information from being used in any other way. In a weakly typed language there are no a priori restrictions on how information can be used: the meaning of information is determined solely by the way it is used, not by any initial promises.¹

Modern computers are fundamentally typeless: any word in memory can hold any kind of value, such as an integer, a floating-point number, a pointer, or an instruction. The meaning of a value is determined by how it is used: if the program counter points at a word of memory then it is treated as an instruction; if a word is referenced by an integer add instruction then it is treated as an integer; and so on. The same word can be used in different ways at different times.

In contrast, today's system programming languages are strongly typed. For example:

Each variable in a system programming language must be declared with a particular type such as integer or pointer to string, and it must be used in ways that are appropriate for the type.
Data and code are totally segregated: it is difficult or impossible to create new code on the fly.
Variables can be collected into structures or objects with well-defined substructure and procedures or methods to manipulate them; an object of one type cannot be used where an object of a different type is expected.

Typing has several advantages. First, it makes large programs more manageable by clarifying how things are used and differentiating between things that must be treated differently. Second, compilers can use type information to detect certain kinds of errors, such as an attempt to use a floating-point value as a pointer. Third, typing improves performance by allowing compilers to generate specialized code. For example, if a compiler knows that a variable always holds an integer value then it can generate integer instructions to manipulate the variable; if the compiler doesn't know the type of a variable then it must generate additional instructions to check the variable's type at runtime.

To summarize, system programming languages are designed to handle the same tasks as assembly languages, namely creating applications from scratch. System programming languages are higher level and much more strongly typed than assembly languages. This allows applications to be created more rapidly and managed more easily with only a slight loss in performance. See Figure 1 for a graphical comparison of assembly language and several system programming languages.

3 Scripting languages

Scripting languages such as Perl [9], Python[4], Rexx[6], Tcl[8], Visual Basic, and the Unix shells represent a very different style of programming than system programming languages. Scripting languages assume that there already exists a collection of useful components written in other languages. Scripting languages aren't intended for writing applications from scratch; they are intended primarily for plugging together components. For example, Tcl and Visual Basic can be used to arrange collections of user interface controls on the screen, and Unix shell scripts are used to assemble filter programs into pipelines. Scripting languages are often used to extend the features of components but they are rarely used for complex algorithms and data structures; features like these are usually provided by the components. Scripting languages are sometimes referred to as glue languages or system integration languages.

In order to simplify the task of connecting components, scripting languages tend to be typeless: all things look and behave the same so that they are interchangeable. For example, in Tcl or Visual Basic a variable can hold a string one moment and an integer the next. Code and data are often interchangeable, so that a program can write another program and then execute it on the fly. Scripting languages are often string-oriented, since this provides a uniform representation for many different things.

A typeless language makes it much easier to hook together components. There are no a priori restrictions on how things can be used, and all components and values are represented in a uniform fashion. Thus any component or value can be used in any situation; components designed for one purpose can be used for totally different purposes never foreseen by the designer. For example, in the Unix shells, all filter programs read a stream of bytes from an input and write a string of bytes to an output; any two programs can be connected together by attaching the output of one program to the input of the other. The following shell command stacks three filters together to count the number of lines in the selection that contain the word "scripting":

select | grep scripting | wc

The select program reads the text that is currently selected on the display and prints it on its output; the grep program reads its input and prints on its output the lines containing "scripting"; the wc program counts the number of lines on its input. Each of these programs can be used in numerous other situations to perform different tasks.

The strongly typed nature of system programming languages discourages reuse. Typing encourages programmers to create a variety of incompatible interfaces ("interfaces are good; more interfaces are better"). Each interface requires objects of specific types and the compiler prevents any other types of objects from being used with the interface, even if that would be useful. In order to use a new object with an existing interface, conversion code must be written to translate between the type of the object and the type expected by the interface. This in turn requires recompiling part or all of the application, which isn't possible in the common case where the application is distributed in binary form.

To see the advantages of a typeless language, consider the following Tcl command:

button .b -text Hello! -font {Times 16} -command {puts hello}

This command creates a new button control that displays a text string in a 16-point Times font and prints a short message when the user clicks on the control. It mixes six different types of things in a single statement: a command name (button), a button control (.b), property names (-text, -font, and -command), simple strings (Hello! and hello), a font name (Times 16) that includes a typeface name (Times) and a size in points (16), and a Tcl script (puts hello). Tcl represents all of these things uniformly with strings. In this example the properties may be specified in any order and unspecified properties are given default values; more than 20 properties were left unspecified in the example.

The same example requires 7 lines of code in two methods when implemented in Java. With C++ and Microsoft Foundation Classes, it requires about 25 lines of code in three procedures (see [7] for the code for these examples). Just setting the font requires several lines of code in Microsoft Foundation Classes:

CFont *fontPtr = new CFont();

fontPtr->CreateFont(16, 0, 0,0,700, 0, 0, 0, ANSI_CHARSET,

    OUT_DEFAULT_PRECIS,CLIP_DEFAULT_PRECIS, DEFAULT_QUALITY,

    DEFAULT_PITCH|FF_DONTCARE, "Times New Roman");

buttonPtr->SetFont(fontPtr);

Much of this code is a consequence of the strong typing. In order to set the font of a button, its SetFont method must be invoked, but this method must be passed a pointer to a CFont object. This in turn requires a new object to be declared and initialized. In order to initialize the CFont object its CreateFont method must be invoked, but CreateFont has a rigid interface that requires 14 different arguments to be specified. In Tcl, the essential characteristics of the font (typeface Times, size 16 points) can be used immediately with no declarations or conversions. Furthermore, Tcl allows the behavior for the button to be included directly in the command that creates the button, while C++ and Java require it to be placed in a separately declared method.

(In practice, a trivial example like this would probably be handled with a graphical development environment that hides the complexity of the underlying language: the user enters property values in a form and the development environment outputs the code. However, in more complex situations such as conditional assignment of property values or interfaces generated programmatically, the developer must write code in the underlying language.)

It might seem that the typeless nature of scripting languages could allow errors to go undetected, but in practice scripting languages are just as safe as system programming languages. For example, an error will occur if the font size specified for the button example above is a non-integer string such as xyz. The difference is that scripting languages do their error checking at the last possible moment, when a value is used. Strong typing allows errors to be detected at compile-time, so the cost of run-time checks is avoided. However, the price to be paid for this efficiency is restrictions on how information can be used: this results in more code and less flexible programs.

Another key difference between scripting languages and system programming languages is that scripting languages are usually interpreted whereas system programming languages are usually compiled. Interpreted languages provide rapid turnaround during development by eliminating compile times. Interpreters also make applications more flexible by allowing users to program the applications at run-time. For example, many synthesis and analysis tools for integrated circuits include a Tcl interpreter; users of the programs write Tcl scripts to specify their designs and control the operation of the tools. Interpreters also allow powerful effects to be achieved by generating code on the fly. For example, a Tcl-based Web browser can parse a Web page by translating the HTML for the page into a Tcl script using a few regular expression substitutions. It then executes the Tcl script to render the page on the screen.

Scripting languages are less efficient than system programming languages, in part because they use interpreters instead of compilers but also because their basic components are chosen for power and ease of use rather than an efficient mapping onto the underlying hardware. For example, scripting languages often use variable-length strings in situations where a system programming language would use a binary value that fits in a single machine word, and scripting languages often use hash tables where system programming languages use indexed arrays.

Fortunately, the performance of a scripting language isn't usually a major issue. Applications for scripting languages are generally smaller than applications for system programming languages, and the performance of a scripting application tends to be dominated by the performance of the components, which are typically implemented in a system programming language.

Scripting languages are higher level than system programming languages, in the sense that a single statement does more work on average. A typical statement in a scripting language executes hundreds or thousands of machine instructions, whereas a typical statement in a system programming language executes about five machine instructions (see Figure 1). Part of this difference is because scripting languages use interpreters, which are less efficient than the compiled code for system programming languages. But much of the difference is because the primitive operations in scripting languages have greater functionality. For example, in Perl it is about as easy to invoke a regular expression substitution as it is to invoke an integer addition. In Tcl, a variable can have traces associated with it so that setting the variable causes side effects; for example, a trace might be used to keep the variable's value updated continuously on the screen.

Because of the features described above, scripting languages allow very rapid development for applications that are gluing-oriented. Table 1 provides anecdotal support for this claim. It describes several applications that were implemented in a system programming language and then reimplemented in a scripting language, or vice versa.

In every case the scripting version required less code and development time than the system programming version; the difference varied from a factor of 2 to a factor of 60. Scripting languages provided less benefit when they were used for the first implementation; this suggests that any reimplementation benefits substantially from the experiences of the first implementation and that the true difference between scripting and system programming is more like a factor of 5-10x than the extreme points of the table. The benefits of scripting also depend on the application. In the last example of the table the GUI part of the application is gluing-oriented but the simulator part isn't; this may explain why the application benefited less from scripting than other applications.

To summarize, scripting languages are designed for gluing applications. They provide a higher level of programming than assembly or system programming languages, much weaker typing than system programming languages, and an interpreted development environment. Scripting languages sacrifice execution speed to improve development speed.