Copyright © 2010 David Schmidt

Chapter 9:
Domain-Specific Languages


9.1 Domain-specific software architecture
9.2 Domain-specific language (DSL) and domain-specific programming language (DSPL)
    9.2.1 From DSL to DSPL
    9.2.2 DSPLs as ``little languages''
9.3 Top-down (``external'') DSPL
    9.3.1 Developing a top-down DSPL
    9.3.2 Example top-down language: Gate-layout language
    9.3.3 Script versus projecting editor
    9.3.4 Summary
9.4 Bottom-up (''internal'') DSPL
    9.4.1 A bottom-up DSPL evolves over time
    9.4.2 Developing a bottom-up DSPL
    9.4.3 Example: Grid-GUI patterns
    9.4.4 Example: Observer design pattern
    9.4.5 Summary
9.5 Hybrid DSPL
    9.5.1 Implementation techniques
9.6 Further reading


It is unlikely that you will ever design a general-use language like Fortran, C++, ML, or Prolog, but if you become a professional software engineer or software architect, it is highly likely that you will specialize in some problem area, like telecommunications, aviation, website management , banking, or gaming. You will become expert at building systems in your problem area, and you may well design a notation, a language, that helps you and others write solutions to problems in this area. In this case, you are a designer of a domain-specific language that is used to build domain-specific software architectures.

This chapter introduces these concepts, applying the concepts already learned.


9.1 Domain-specific software architecture

Every large system is built from software and hardware components; the pattern of layout and connection of the components is called its architecture. A software architecture is the layout of software components. The software architecture is deployed (installed) on the hardware architecture.

Specific problem areas, e.g., flight-control or telecommunications or banking, use specific hardware architectures, and they also use specific software architectures. When a new model of airplane is designed, the hardware architecture (the airplane hardware, including its computers) is based on a hardware design that has succeeded in the past. (It is too great of a risk to start from scratch; it is also better to build on and refine what is known to work.) The software architecture for the plane will also be based on some standard layout that is known to work well.

Software architects use a collection of concepts, techniques, and patterns to build a new system in an established problem area; this collection is called a domain-specific software architecture:

  1. domain-specific language: the language of the problem area: (i) concepts and terminology (words, phrases, and actions that the clients, designers, and builders use to discuss the problem and design the solution); (ii) customer requirements (what the system must do) and reference requirements (what the system can do --- see the next numbered item below); (iii) scenarios (examples of behaviors); (iv) configuration models (blueprinting techniques for the system and its operation --- entity-relationship ("class" or "dependency") diagrams, data flow ("sequence") diagrams, deployment diagrams, etc.)

  2. reference requirements: the ``features'' or ``customizations'' or ``attributes'' or ``ordering options'' that clients (customers/users) select to configure the desired system. (Think about all the choices you make when you order a brand new car from an auto dealer --- colors, engine options, accessories --- these are the reference requirements for the car you want.)

    The reference requirements are part of the terminology of the application domain, but they are often specially identified because they are often treated specially in the implementation methodology.

  3. reference architecture: the software and hardware architectures ("platforms") used as starting points for building the implementation, typically mapped out by blueprints. (These can be based on earlier versions/releases.)

  4. supporting environment/infrastructure: the available hardware platforms and software languages, libraries, frameworks, and development tools.

  5. a methodology for designing, implementing, and evaluating the system using the above items.
We can't study all these topics, so we focus on the first one: The language of a domain is called a domain-specific language (DSL).


9.2 Domain-specific language (DSL) and domain-specific programming language (DSPL)

English is a general-purpose language. Legal English is a special-purpose language, dedicated to writing contracts and laws --- it is specific to the domain of contracts and laws. Algebra is a domain-specific language for stating numerical relationships.

A language for discussing problems, behaviors, and solutions within a problem domain is a domain-specific language (DSL). The language's vocabulary includes concepts and notation from the domain --- the nouns, pronouns, adjectives, verbs, and adverbs of the language. The language lets participants (people and machines) discuss and implement solutions in the domain. Because its vocabulary is limited to the specific domain, a DSL is often useless for discussing and solving problems outside the domain.

A DSL uses concepts familiar to people who work in the domain. Here are two examples:

Spreadsheets:

To an accountant who must prepare spreadsheets each day, the "spreadsheet domain" is a little world of its own. Its domain consists of grids, cells, numbers, words. These are the "entities" or "nouns" of spreadsheet-language.

The entities can have features/attributes (``adjectives''): e.g., a word can be a label or data (in a cell). A grid has dimensions (rows and columns). A number can be data or a total value.

There are certainly operations on the entities --- inserting data into a cell, totalling the values in a row or column, printing a table.

A sequence or "script" of operations in some pattern or order, perhaps triggered by an event, is called an action. (Actions are "sentences" or commands.) Example: "When a number is inserted into Row 9, update the total for Row 9 using Equation 9 and redisplay the updated grid."

The accountant thinks in the language of the spreadsheet domain when building a spreadsheet, whether or not a computer program is helping to assemble and display the spreadsheet. But if the computer is doing the speadsheet layout, the DSL for spreadsheets becomes a programming language, e.g., Excel. (See below.)

A sample behavior or "test case" of a spreadsheet is a scenario. For example,

"I have constructed a spreadsheet that models an order form, and I insert a new item to my order: in the first empty row of my order form, I typed the product number, the quantity, and cost per item. My typing (it's an "event") triggers these "actions" by the spreadsheet: (i) it computes, in the rightmost column of the row, the total cost of the item; (ii) it adds the item's total cost to the other costs in the rightmost column and updates the cell in the last row, last column, with the new total cost of my order."

Scenarios like this help the software architect implement the spreadsheet application for the accountant to use.

Alarm systems:

Say you make money by installing alarm systems in office buildings. You must think and talk and work in the "alarm-system domain" with the building's owners and employees. A DSL for sensor-alarm networks would discuss

domains (''nouns''): sites (building, floor, hallway, room), devices (alarm, movement detector, camera, badge), people (employee, guard, police, intruder). These are the ``nouns'' of the DSL.

Elements have features/attributes (``adjectives'') and operations (''verbs'').

Actions (``sentences'') are initiated by events.

Here is a scenario, stated in the DSL:

``when a movement detector detects an intruder in a room (this is an event), it sends an alarm to the guard's remote monitor, it switches on the lights in the room, and it activates the room's camera.''

A collection of such scenarios helps you install an alarm system with the desired detection and reporting devices that are used with the building and people. If a computer is involved in the domain, then some of the DSL about alarms will be a computer language that you can program in.

Compare the lingo of sensor alarms to the lingo you write in Java --- in the latter, the ``nouns'' are numbers, arrays, objects, and variables that name numbers, arrays, objects, etc. The ``adjectives'' are data types and other declaration modifiers. The ``operations'' are arithmetic, data-structure indexing, method call, etc. ``Actions'' are commands, or groups of commands. ``Events'' can be GUI events or a call to a method to start execution. Java is a ``DSL'' for computation on numbers and arrays and objects.

A DSL lets stakeholders (the participants in a systems project) communicate their ideas (needs, suggestions, solutions, implementations, orders). The DSL is a is a modelling language that lets us discuss models, structures, and behaviors specialized to a problem domain like telecommunications, banking, transportation, gaming, algebra, typesetting, etc.

If the computer is a ``participant,'' that is, we can use the DSL to tell the computer what to do --- we can program the computer --- then the DSL is a domain-specific programming language (DSPL).


9.2.1 From DSL to DSPL

The previous point is critical. Consider the spreadsheet scenario again: If we build a spreadsheet application from scratch, say in Java, then we use the scenario to help us design a "use-case realization" --- an explanation of the Java components that will be required to make the scenario come to life on a computer. We study a lot of scenarios; we write lots of use-case realizations; we design a lot of object diagrams, class diagrams, sequence diagrams, etc.; we code a lot of Java components; and we have a Java-coded system --- it's a lot of work, maybe weeks or months of work.

Now, say we learn Excel (or a similar spreadsheet application). Excel understands the DSL of spreadsheets, and we "talk" (program) directly in Excel, implementing the scenarios and the spreadsheet application in a matter of hours or days. This is because Excel is a DSPL for spreadsheets --- we say what we want in the language of Excel and we have the application we want in minimal time and effort. This is the big payoff of having a DSPL.

This observation drives everything in this chapter.

Other problem domains and their DSLs

Domain-specific languages are especially useful for describing reactive systems --- alarm systems, telecommunications systems, vending machines, multi-player games, single-user applications that use a GUI, and communications protocols --- hence the classification of the DSL into events, actions, features, nouns, and operations.

But not all computational mechanisms are reactive. For example, the equational language of algebra is a DSL, and the computation underlying its equation sets are simplification laws.

Yet another variation is a domain related to constraint solving, such as crossword puzzles or Sudoku or database queries, where the domain language is a set of clues or constraints that must be solved.

In these cases, the appropriate DSL might be less``event-action oriented'', but in any case, it will certainly remain as the appropriate language that the stakeholders use to discuss their problems and the solutions.

General-purpose languages

Why are languages like C, Java, and Prolog called ``general purpose'' languages? After all, each such language is specific to data domains like numbers, strings, tables, structs, objects, relations, and so on.

One might argue that a general-purpose computer language is ``domain un-specific'' because it favors no one application domain very much over another. (A cynic would say that a general-purpose language is a ``no-domain language,'' since there is no real-life application domain that matches it!)

A user of a general-purpose language must become an expert modeller of real-life application domains in the domains of the general-purpose language --- This is core computer science: how to model real-life domains and domain-specific language within general-purpose computing machines and general-purpose computer language.

On the other side of the coin, we can argue that a language like C is domain-specific to the domain of von Neumann machines, and a language like Java is domain-specific to heap-based object machines. Such machines are used to mimick/model other computational domains, and this is why we use C or Java to mimick/model other DSLs.

When the complexities in domain modelling become too great, the general-purpose programming language must be abandoned for a domain-specific one.


9.2.2 DSPLs as ``little languages''

The previous section suggests you define a Domain-Specific Programming Language (DSPL) by extracting from a DSL its computational part. This treats the computer as one stakeholder in the community of participants.

But there is another origin of DSPLs that comes totally from within the programming world: It is inconvenient to drag out a general-purpose language to code a solution to something small and simple. For example, do you code a Java program each time you do some calculator arithmetic? No --- you use a calculator language instead. It is always better to use a smaller, simpler language --- a DSPL --- that matches the problem you face.

For this reason, programmers sometimes call DSPLs little languages (e.g., ``here is a little language for drawing figures''; ``here is a little language for linking files''.) Here is a short list of ''little language'' DSPLs that have/had wide use:

  1. Make --- for linking files
  2. Matlab (and Mathematica) --- for doing linear algebra
  3. SQL --- for doing queries and updates to databases
  4. VHDL, Verilog, and VHSIC -- for laying out hardware circuits
  5. Yacc, Bison, Antlr --- for programming parsers
  6. Excel --- for computing spreadsheets
  7. HTML, CSS --- for generating web-browser documents
  8. groff and LaTex (and even Word90) --- for typesetting documents
  9. eqn --- for typesetting math formulas
Admittedly, the current versions of many of these ``little languages'' aren't so little anymore, but almost all them came about because someone thought,
''It would be nice to have a little language to help me do ...this little job....''
So, that person designed a little language to do the little job.

In terms of domain-specific software architecture, someone might ask you,

''It would be nice to have a little language to help ...somebody... do ...some little job in this domain.... Can you put together something for us?''

For example, ''It would be nice to have a little language to help us lay out the wiring and sensors for a building's alarm system.''.

Or, ''It would be nice to have a little language to help us write the protocols for how the movement detectors send/receive messages to/from the other devices and people in the network.''

This kind of wishful thinking can lead to a domain-specific programming language, in particular, a top-down domain-specific programming language.


9.3 Top-down (``external'') DSPL

Each of the languages in the list in the previous section does one thing well in one application domain; by no means should any of these languages be used for general-purpose computing. All the examples in the list are called top-down (or ``external'') DSPLs because they are designed as stand-alone languages that implement domain concepts and nothing more. Since a top-down DSL is a ``little language,'' it should be easier to learn and use than a general-purpose language. (If it isn't, then it is a failure!) In many cases, less-experienced and maybe even non-programmers should be able to use a top-down DSPL to write solutions.

Let's review some of the little languages in the above list.

There is one critical standard for the success of a top-down DSPL:

Programmers must ``see'' their domain and the actions within it in the DSPL.
That is, the top-down DSPL lets the programmer think and talk directly in the problem domain; there are no distracting complications. (For example, a typical Excel user sees (figuratively and literally) a spreadsheet and does not want to write for-loops that compute on columns of numbers in the spreadsheet!)

Upon first hearing, it sounds like top-down DSPLs are wonderful --- a language for just my problem that lets me say exactly what I want! --- but in reality, a top-down DSPL is a ``mixed bag'' of assets and drawbacks:

+ non-programmers can discuss and use the DPSL
+ the DSPL encourages standard patterns of design, implementation, and optimization
+ there is fast development of programs that fall squarely in the problem domain
- staff must be trained to use the DSPL, which is a brand-new language to them
- the interaction of DSPL-generated software with other software components can be difficult
- there is a high cost in designing, developing, and maintaining a DSPL
For these reasons, top-down DSPLs are not always the best tool for solving domain-specific problems.


9.3.1 Developing a top-down DSPL

The starting point is this: become an expert in the problem domain: learn the vocabulary --- nouns, verbs, and adjectives. Develop many scenarios (case studies) within the domain. Extract from the scenarios patterns or schemes of structure, behavior, computation. Build computer systems in the domain; learn about their

Keep this directive in mind, always:

The programmers must ``see'' their domain and the actions within it in the DSPL!

If the DSPL's users are forced to code in notation and concepts that lie outside their problem domain, the users will get lost. (That's why non-programmers don't use Java as a DSPL for spreadsheet building!) In a serious development effort, you will design an IDE-like tool as well.

Implementation

A top-down DSPL is usually built as a stand-alone parser-interpreter, which reads programs in the little language, parses and executes them. The language in which the interpreter is built is hidden from the DSPL user --- users believe they are "talking" directly to the computer, which "understands" the DSPL "directly".


9.3.2 Example top-down language: Gate-layout language

Here is a classic example. People who design chips from gates once drew huge wiring diagrams, which were photo-reduced to tiny templates from which circuit board were fabricated. This is tedious, error-prone, and expensive. A DSPL would be a better solution.

The DSL talks about ``gates'', which have features like input connections (ports/wires) and output connections. Gates have a function (feature): AND, OR, NOT, etc. Gates are assembled into subassemblies (e.g., one can build an XOR-subsassembly), and subassemblies must be connected and grouped on a board. Gates and subassembles might be annotated with power and space requirements.

The DSL is not event-or-action driven, like an alarm system or GUI, so scenarios are more like puzzles:

There is no need for a general-purpose language, with assignments and loops, for expressing these scenarios --- an equation-like layout language with a few key operations and a few assembly constructions will suffice. (Or, a predicate-based, definition language might be the way to go!)

We might design a language whose scripts handle scenarios like the ones above, e.g., the first scenario is programmed like this:

ASSEMBLY A1 : inputs IN1, IN2;  outputs OUT1.
   w1 = AND(w1, w2).
   w2 = OR(w1, IN2).
   OUT1 = NOT(w2).
ENDASSEMBLY
The connections are coded as equations, where the left-hand side of each equation is the name of a wire. The second scenario might be programmed like this:
ASSEMBLE A2 : inputs IN1, IN2;  outputs OUT1, OUT2.
  solve for           1     1              1     0;
                      1     0              1     1;
                      0     1              1     1;
                      0     0              0     0.
  using {NAND}
ENDASSEMBLY
The new operation, solve for _ using {_}, accepts the tabular input and generates a solution that is named A2 that has the required functionality and behavior.

The third might go like this:

ASSEMBLY A3 : inputs IN1;  outputs OUT1; 
              suchthat w2.voltage < (2 mv).
   w1 = AND(In1, w2).
   w2 = NOT(w1)
   OUT1 = w2
ENDASSEMBLY
Here, the constraint on a wire's voltage is listed as part of the assembly's output specification. The user does not code how to solve the constraint --- the implementation knows the details.

At this point, think about what operations you might add to this little language that connects assemblies like A1 to A2 or A3 --- you will want a little linking language, and it should probably look like equations that "equate" inputs to outputs. Try it.

Of course, there are many hardware languages that can do the above and then some.

Next, consider how a parser and interpreter would be defined to read programs in this DSPL and execute them (in this case, generate circuit-diagram layouts for a board).


9.3.3 Script versus projecting editor

Computer programmers treat language as text that must be typed with a keyboard. This view is outdated.

Indeed, what is a program? Is it

All of the above are means of inputting the program into a computer. Actually, a program is a data structure holding semantic information.

Indeed, programmers use development tools, such as text editors, IDEs, and debuggers, for inputting their programs. Software engineers don't write programs as script --- they interact with the IDE, choosing menu options and completing templates until the IDE announces that a program is completed.

Users of top-down DSPLs are even more ``IDE-dependent'' than programmers. For example, an Excel user will interact with Excel's GUI to insert data into cells of a spreadsheet and write equations that are embedded into the spreadsheet's ``logic'' so that the row-and-column totals are correctly computed and displayed. Exactly where is the ``program''? Indeed, the ''program'' is completely intertwined with the Excel development environment and its internal data structures and event handlers.

In a note at http://martinfowler.com/articles/languageWorkbench.html Martin Fowler coins the term, ``projecting editor'' for DSPL that is programmed and understood solely within an IDE:


The projecting editor keeps an abstract representation of the user's "program" that is filled in bit by bit, not necessarily sequentially, not necessarily as script. (The "program" is data structures, event handlers, and GUIs held within the projecting editor!). The program's abstract representation might include a parse-tree-plus-symbol-table along with data/control/component structures that define the program's semantic intent.

IMPORTANT: The projecting editor must have a back end that can interpret the complex "program" or can generate a script ("target code") that can be interpreted. (The ``storage representation'' in the diagram is some file format that archives ("pickles") the abstract representation at the end of the IDE session.)

If you are an Emacs or a vi or an Eclipse or a Visual Studio user, you are using a simple projecting editor/DSPL for document generation. The most extreme view is that any user interface for any application is a DSPL.

If you are developing a top-down DSPL for non-programmer users, then you are almost certainly forced to develop a projecting editor to go with it.


9.3.4 Summary

A top-down (external) DSPL is a stand-alone, ''little language'' that has a limited domain of application. Examples are SQL for data base queries, Excel for spreadsheet construction, and Matlab for linear algebra.

A top-down DSPL lets you write small, simple programs in which you ``see'' the elements, features, and actions of the application domain. Some top-down DSPLs are meant for use by non-programmers (e.g., Excel).

A top-down DSPL has high overhead in maintaining the language and its implementation. The implementation will include a parser and interpreter, and when non-programmers are users there should be an IDE/projecting editor that helps users ``see'' the application domain and ``see'' their programs.


9.4 Bottom-up (''internal'') DSPL

There is another variant of DSPL, one that is used by an experienced programmer who wants to ``extend'' a general-purpose language with concepts specific to a problem domain. In this situation, the DSPL is added to the general-purpose, host language, so that programs are a mix of host-language code and the DSPL.

Over time, more and more DSPL constructions are added; the host language is "covered over" by the DSPL; and programs are written (almost) entirely in the DSPL.

This is called a bottom-up (or ``internal'') DSPL.

Background: GUI-building frameworks

DSPLs often arise as extensions of frameworks.

Recall that a framework is a collection of components that help someone more quickly implement solutions in a problem domain. (Think of javax.swing or a library from .NET.) Frameworks are ``not-quite bottom-up DSPLs.'' (We will explain the remark later.)

Consider these libraries for GUI-building:

Each library is married to a host language, because a GUI by itself is useless --- the GUI must be connected to components that do something.

These libraries are called frameworks, because each has their own collection of components that implement nouns (''window,'' ''frame,'' ''button,'' ''layout,'' ...) and verbs (''setTitle,'' ''getText,'' ''paint,'' ...) of the GUI domain. They usually come with sample programs that suggest patterns for assembling and calling the components. But they are implemented in their host languages, and a programmer must write (lots of) code in the host language to assemble a working GUI from the GUI framework. (Often, an IDE is used to help with writing the host-language code.)

A GUI framework is an almost-DSPL for GUIs, because it is a library that implements GUI concepts, but there is no ``programming language'' for GUI building, only the components and some example assemblies, where the assemblies are written in host-language code.

An application is a mixture of components from the GUI-framework, calls to the GUI-framework, and coding written from scratch in the host language.

A GUI framework is often ``married'' to its host language by a visual editor, e.g., Visual Basic, Visual C++, Visual Studio, and Eclipse. The visual editor tries to fill the gap between the framework and host language. It usually isn't enough, because there is no language for GUI building, only bits and pieces.


9.4.1 A bottom-up DSPL evolves over time

If you build GUIs for a living, you will not be satisfied with just a GUI framework and an IDE --- you will develop and use design patterns, macros, custom components and shortcuts that reduce your time and effort. You are developing your own little language, your own bottom-up DSPL, for GUI building.

Experienced programmers naturally become bottom-up DSPL designers, because over time they assemble a library of components and patterns that express domain concepts that they use over and over to solve problems in the domain.

Eventually, the programs these people write consist almost totally of their library code. The host programming language --- when it is used at all --- acts merely as minimal ``glue code'' for connecting the components and patterns.

At this point, the host language plus the library or components and patterns is a bottom-up DSPL, because the library has become ``more important'' to the problem solving than the host language itself. What has happened is this:

The programmer has extended the host language ``upwards'' towards the problems to be solved.
This makes the host(glue)-language-plus-its-library a bottom-up DSPL.

The custom-written library for the problem area is written in the host language, and it is oriented towards encoding ``domain-concepts-as-code'' (nouns as data-structure patterns, verbs as operations/control-structure patterns, features as attributes, sentences and paragraphs as assembly/design patterns) so that the scenarios discussed in the DSL are readily converted into code. Experienced programmers have good instincts for coding domain concepts as code and saving them as libraries. It is almost a matter of survival --- there is never enough time to build a new solution completely from scratch!

Many of the ideas from object-oriented design and design patterns apply to bottom-up DSPL development: classes, methods, and design patterns implement domain concepts. You code them, save them, reuse them --- you have a language. Languages like Scheme (via lambda abstraction and hygienic macros) and Smalltalk and Ruby (via blocks and macros) let a programmer easily define design patterns as custom templates directly in source-code syntax. The Scala language even lets you alter its own parser so that it can be extended to parse you new syntax patterns! These are ways of extending the host language upwards towards the application domain.

But any general-purpose language can serve as a host language. Usually the host language is whatever language in which the starting libraries and frameworks are written.

A bottom-up DSPL has its strengths and weaknesses also:

+ It integrates well with applications written in its host language
+ Its development and maintenance is managed naturally over the time period when it is used, since it is an ``organic library'' that grows and adapts to its applications.
- Its range of users is largely limited to experienced programmers, perhaps just to the people who develop and maintain the library.
- Its ease of use is connected to the ease of use of the underlying host language.

Implementation

As just noted, a library of components, macros, and syntax extensions are added to a host language to implement the DSPL. Macro definitions are explained in detail later in the chapter; a macro is the name of a pattern that, when used in a program, is extracted and replaced ("expanded") into source code in the host language. It is a naive use of procedure-call copy-rule semantics.

A separate parser-translator tool might be written to translate DSPL patterns into "target code" in the host language.

An implementation might also use ``meta-programming'' features provided by the host language. One classic meta-programming feature is eval, used in Lisp and scripting languages to parse and execute strings and data that are concatenated into what is meant to be code. A variant of eval that allows one to inspect, alter and then execute data structure, even the executing program's data structures, is ``reflection/reification.'' Javabeans is an example of this feature.


9.4.2 Developing a bottom-up DSPL

Some of the following was already stated but bears repeating:

Experienced programmers are the natural users of a bottom-up DSPL because they design it themselves, over time, as a library of implemented components and patterns. Eventually the host programming language acts merely as ``glue'' for connecting the components selected from the library: The programmer has extended the host language ``upwards'' towards the problems to be solved.

Bottom-up design might go like this:

  1. Use a framework to write lots of systems in your application domain. Add more and more components (classes, modules) to the framework.
  2. Notice which control and assembly patterns you are copying-and-pasting into the systems you build. Code these patterns into macros or parameterized procedures/control structures, that is, find some way to extend your host language with the patterns.
  3. Repeat the previous steps until the components and patterns you implemented have names and computation power that match the entities/features/operations/events/actions in the DSL that you think in and talk in.
  4. Most importantly, force yourself to use your library (and improve it!) as much as possible, instead of writing new code from scratch. You should program by selecting code from your library and ``gluing'' it together with minimal code from the underlying host language. (Research has shown this is the most difficult step for software engineers to do!)

Your ultimate goal is to make the bottom-up-DSPL library ``stand-alone,'' so that you write programs just with your library and with almost zero new code from the host language. This means you use the host language only as a as an ``interface language'' to contact external components that you have not written or to ``escape'' from the problem-domain area to execute code from some other library or application.

Implicit in the previous paragraphs are the notions of framework and product line from mainstream Software Engineering:

When you implement design or assembly patterns as new constructions, you want to have a nice syntax for the pattern. Some languages, e.g., Ruby, have a built-in "macro processor" for defining new patterns. Other languages, e.g., Java, are less helpful. Here are some possible combinations to use:
Host languagePattern language that links to host
Java Scala (provides ML-like front end) or Groovy (provides Python-like front end)
C# F# (provides ML-like front end)
PythonPython's re module (provides macroprocessor)
CGPP or m4 (macroprocessors)

Unlike "little languages" (top-down DSPLs), bottom-up DSPLs are "big", because they start with a host language and a framework and get bigger and bigger with extensions until the DSPL library is completed. So, it is difficult to give simple examples of a bottom-up DSPL. But we will try.


9.4.3 Example: Grid-GUI patterns

Say that you use Tkinter to program lots of GUIs that are grids. The grids always turn out to be matrices of buttons that look and behave the same. (Spreadsheets and game boards work like this!) This means you are copying-and-pasting many patterns of definition, assembly, and control from existing applications to new ones. It would be much better to code the patterns as classes, functions, templates, and macros that are inserted into your Tkinter programs. The resulting programs would be easier to code and read and would work reliably (because your patterns are implemented correctly once and for all).

There isn't time or space here to present lots of grid-GUIs, but here's one, a game board, where some common grid-GUI coding patterns are marked by #****.

===================================================

#Game board for "Pente" game:     

from Tkinter import *
import PenteBoard   # the model subassembly --- holds the gameboard data

### the CONTROLLER module --- this should be placed in a separate file.

#1 ***********************************
def makeHandler(myrow, mycolumn) :
    """makeHandler constructs a handler function for a new button.
       parameters: 
          myrow - an int, the row coordinate where the new button lives
          mycolumn  - an int, the column coordinate where the new button lives
       returns: the handler function customized for a new button
    """
    def handleButtonPress() :
        """handleButtonPress is the constructed handler function.
           It makes the move for the human who pressed
           this button (which is at position myrow,mycolumn).
           The updated board is then painted.
        """
        if PenteBoard.game_on() :
            PenteBoard.makeMove(myrow, mycolumn)
        repaintGUI()

    return handleButtonPress
# END ***********************************

### the VIEW module starts here:

def repaintGUI() :
    """repaintGUI  repaints the foreground text of all the buttons on the GUI,
       it also updates the displayed count of captures, and if there is a
       winner, it prints a message as to who won.
    """
    global buttons, label1

    #2 ****************************
    for i in range(size) :
        for j in range(size) :
             buttons[i][j].configure(text = PenteBoard.contents(i,j))
             buttons[i][j].configure(bg = "white")
    #END *****************************

    label1.configure(text = "Your captures " + " = " \
                      + str(PenteBoard.getCaptures()))

#3 *********************************
window = Tk()
window.title("Pente")
size = PenteBoard.size
window.geometry(str(50 * size) + "x" + str(50 * (size)))
frame = Frame(window)
frame.grid()
#END  *********************************

label1 = Label(frame,
              text = "Captures " + " = " \
                      + str(PenteBoard.getCaptures()),
              font=("Arial", 12, "bold") )
label1.grid(row = 0, column = 0, columnspan = 5)

#4 **********************************
buttons = []    # a nested list that remembers addresses of all button objects

for i in range(size) :
    button_row = []
    for j in range(size) :
        button = Button(frame,
                        font = ("Arial", 14, "bold"), fg = "blue", bg = "white",
                        width = 2, height = 1)
        button.configure(text = PenteBoard.contents(i,j))
        button.configure(command = makeHandler(i, j))
        button.grid(row = i+2, column = j)
        button_row = button_row + [button]
    buttons.append(button_row)
#END*********************************

window.mainloop()   # activate GUI

===================================================
There's a lot of ugly code here, and an IDE will not help you avoid the tedious coding of the nested loops for initializing and resetting the button grids. The event handlers must also be manually coded. The marked patterns above are based on simple concepts:
  1. Pattern 1 is code for defining a family of similar-behaving event-handling functions for a grid of buttons.
  2. Pattern 2 is code for repainting a grid of buttons when there has been an update to the GUI's model (in this case, when the model, Penteboad has been altered due to a move).
  3. Pattern 3 is "boilerplate" for allocating the main window.
  4. Pattern 4 configures the appearance of the buttons.
All the "patterns" are copy-and-paste coding from earlier applications. They should be converted into DSL concepts, enriching the GUI framework and its language of Frames and Buttons with "Grids".

With some macro-named patterns, the above code is simplified to the following, where the macros are prefixed by @-signs:

===================================================

from Tkinter import *
import PenteBoard   # the model subassembly

def repaintGUI() :
    """repaintGUI  repaints the foreground text of all the buttons on the GUI,
       it also updates the displayed count of captures.
    """
    global label1
    @repaintGrid from (PenteBoard)
    label1.configure(text = "Your captures " + " = " \
                      + str(PenteBoard.getCaptures()))

window, frame = @initializeFrame("Pente", PenteBoard)

label1 = Label(frame,
              text = "Captures " + " = " \
                      + str(PenteBoard.getCaptures()),
              font=("Arial", 12, "bold") )
label1.grid(row = 0, column = 0, columnspan = 5)

@configureGrid from (PenteBoard, frame)
  handler (lambda(myrow, mycolumn) =>
              if PenteBoard.game_on() :
                  PenteBoard.makeMove(myrow, mycolumn);
              repaintGUI() )  # THE HANDLER IS DEFINED AS CLOSURE CODE
  and (font = ("Arial", 14, "bold"),
       fg = "blue", bg = "white",
       width = 2, height = 1)

window.mainloop()

===================================================
We have a more readable mix of GUI-domain concepts and host-language code. The macros are the start to a bottom-up DSPL.

Here are the patterns that were used above:

  1. @repaintGrid from (MODEL) consults the MODEL object for the values of all the cells that are modelled and uses this information to repaint the grid GUI. The call expands into this template:
    global buttons
    size = MODEL.getSize()
    for i in range(size) :
            for j in range(size) :
                 buttons[i][j].configure(text = MODEL.contents(i,j))
                 buttons[i][j].configure(bg = "white")
    

  2. @initializeFrame(TITLE, MODEL) generates "boilerplate" code for configuring a window and frame large enough to display the MODEL object:
    window = Tk()
    window.title(TITLE)
    size = MODEL.getSize()
    window.geometry(str(50 * size) + "x" + str(50 * (size)))
    frame = Frame(window)
    frame.grid()
    

  3. @configureGrid from (MODEL, FRAME) handler (HANDLER) and (ATTRIBUTES) allocates a grid of buttons for the MODEL, binds each to a unique handler function generated from HANDLER and annotates each button with the optional ATTRIBUTES:
    buttons = []
    size = MODEL.getSize()
    for i in range(size) :
        button_row = []
        for j in range(size) :
            button = Button(frame, ATTRIBUTES)
            button.configure(text = MODEL.contents(i,j))
            button.configure(command = HANDLER(i, j))
            button.grid(row = i+2, column = j)
            button_row = button_row + [button]
        buttons.append(button_row)
    

Over time, more and more components and patterns will be named, saved, and reusued. The GUI programs will call more and more of the saved DSPL concepts and less and less of new code. A bottom-up DSPL evolves.... We will learn in a future section how to use a macroprocessor to declare and call the above macro patterns.


9.4.4 Example: Observer design pattern

Programmers who use object languages use design patterns that help them assemble systems faster. Say that you develop a lot of data structures (``models'' or ``entity objects'') that must be displayed by GUI objects (``observers'') and repainted each time the models are updated. The Observer Design Pattern is a well-known design pattern for this situation. If you use it a lot, you should define it as an assembly pattern and add it to your bottom-up DSPL library.

Here is one version of the Observer Design Pattern:

  1. The ConcreteObserver objects want to be told whenever the ConcreteSubject object is updated. But the ConcreteSubject never contacts ConcreteObservers directly --- this is too messy. So, the ConcreteObservers' ask to be registered (saved) in an Observable (sub)object (or wrapper object), which knows the identity of the ConcreteSubject. Each ConcreteObserver has a handle method, that is, it must implement the Observer abstract interface.
  2. All requests to setState(...) of the ConcreteSubject are in fact sent to the Observable object, which (i) forwards the request to the ConcreteSubject and (ii) then contacts all ConcreteObservers, by event broadcast, which indirectly calls the handle method.
  3. When a ConcreteObserver's handle() is called, the handler calls the ConcreteSubject.getState() to obtain the information it needs to repaint its GUI.
In this way, the ConcreteSubject is isolated from the overhead of observing it. This is the standard subassembly in GUI-based tools/toys; it is the key part of the ``Model-View-Controller'' software architecture.

Rather than recode the assembly each time, we define a macro pattern, @observed, that returns a handle to the Observable object that anchors the design pattern:

Observable omodel = @observed (SUBJECT) by (OBSERVER LIST);
The macro expands to code that declared the Observer-related event, allocates the Observable wrapper object, registers the OBSERVER LIST, and returns the handle that the rest of the system uses for contacting the SUBJECT:
===================================================

# declare a new event type, subjectUpdated, and bind it to its event handler:
public Observer event/delegate subjectUpdated;

# code for constructing wrapper object:
class Observable {
  ConcreteSubject model;
  Observer[] registered;

  public Observable(m, olist) {
    model = m; registered = olist
    foreach (obs in registered) { obs.setSubject(model);
                                  subjectUpdated.register(ob.handle); }
  }

  public setState(...) { model.SetState(...); signal subjectUpdated; }

  public getState() { return model.getState(); }
}

# return handle to observable object:
return new Observable(SUBJECT, OBSERVER LIST);

===================================================
The code isn't strict Java or C#, but the intent is clear. Now, the controllers that contact the model do so by calling omodel.setState(...), which triggers the update to the SUBJECT and an event broadcast that activates the handle method of each observer.

Note that the code above is more than one component --- it is components, declarations, and executed code; it is an extension of what Java/C# provides; it is a coded, reusable DSPL concept.


9.4.5 Summary

A bottom-up (internal) DSPL is a library of components, macros, and precoded patterns that extend a host language upwards to the domain of problems to be solved. Domain concepts become reusable code.

A bottom-up DSPL is developed incrementally, evolving so that programs in the application domain are eventually built (almost) entirely from the DSPL library; the host language is used merely as ``glue code'' for calling/connecting the DSPL components. The DSPL library becomes more important than the language in which the library is written.

In the end, the library is the DSPL, and a programmer programs in the library, not in the host language. The latter fades into the background --- it is the implementation language, no longer the programming language.

A bottom-up DSPL is implemented by components (modules and classes), macro patterns (that define new syntax patterns), and ``meta-programming'' features in the host language (e.g., eval to execute code assembled from strings and data). The sections that follow develop these techniques.


9.5 Hybrid DSPL

Most DSPLs use a mix of top-down and bottom-up concepts.

Mostly top-down

We might say that a DSPL is ``mostly top-down'' if it is designed to express DSL scenarios-in-code and has its own parser (or IDE editor) and interpreter/translator.

A mostly-top-down DSPL can appear like this: You use some framework or component library to build systems, and you develop insight about what the "dream language" (the DSL language!) truly is for writing the algorithms you regularly implement with host-language code and library calls.

So, you design the dream language: You write grammar rules for the syntax and you write semantic equations that map syntax into host language code and library calls. You build the translator. In this way, you have built a bridge from the DSL, at the top of your thinking, mostly downwards to the frameworks below.

A danger of any top-down DSPL is that it is isolated from other systems and implementations. Your mostly-top-down DSPL should let you call library components and execute code written in the implementation language. To do this, add a ``trap door'' to the DSPL so that the execution of the DSPL program can be paused and the implementation-language code can be executed instead. Many scripting languages provide such a trap door, in the guise of an eval operation, which takes as its argument a string that holds executable code --- eval runs the code. Here are three useful forms of trap door in Python:

  1. A eval-like function that executes a string as Python script: For example, exec("y = 2 + x; print y"). executes the program y = 2 + x; print y, using the variables that are visible at the position where exec appears. If one does not want the executed string to affect existing variables, one can invent a namespace exclusively for the string's use, like this: exec("y = 2 + x; print y") in {'x':0}.

    Here is a program that builds a string and runs it:

    x = 2;  y = 3;  z = 5
    invar = raw_input("Type name of variable (x, y, or z) to zero out: ")
    if invar in ("x", "y", "z") :
        code = invar + " = 0"
    else :
        code = "pass"
    exec(code)
    
    The exec command can also read and execute the contents of an opened text file:
    handleToCodefile = open("MyPythonProgram.py", "r")  # open a readable file
    exec(handleToCodefile)  # execute its contents 
    

  2. Python's os package contains procedures for querying the operating system and performing OS commands, e.g.,
    import os
    
    cwd = os.getcwd()     # get current working directory
    if os.path.basename(cwd) == "MyPictures": # is the lowest-level dir "MyPictures" ?
         # then, move up one level to parent directory:
         os.chdir(os.pardir)
    print "Current path is ", os.getcwd()
    
    os.system("ls -a")   # ANY OS command can be supplied as a string arg
    

  3. We can pause execution and execute any external program we wish:
    # run an external program from within Python code:
    import subprocess
    # general format:   subprocess.call(["program-name", "param1", "param2", ...])
    subprocess.call(["C:/Python26/Python.exe", "MyPythonPgm.py"])
    
A mostly-top-down DSPL includes some form of trap door so that bottom-up defined components written in the implementation language can be executed. Some implementation tricks are shown later in this chapter.

Mostly bottom-up

A DSPL is ``mostly bottom-up'' if it is developed as layers of host-language-coded components and macro-coded patterns that help model the problem domain. Perhaps the layers of components do not express directly and immediately the DSL --- there is still a "gap" between the code solutions and the solutions described in scenarios.

To close the gap, we add customized control- or linking-patterns that express the missing domain concepts, so that the concepts look like they are built into the host language. (In particular, we want to avoid writing ugly dot-notation, like packageName.objectName.methodName(arg1, arg2, ...), each time we use a custom-coded domain concept/pattern.)

A good host language will give you a technique to add custom patterns. Here is a simple example:

Say that your problem domain has lots of solutions that use the phrase, ``repeat ACTION until CONDITION holds'' so that this pattern should be added to the DSPL library. Some languages let you define higher-order functions (functions that take code/closures as parameters) in mix-fix keyword notation, like this:

def repeat(action)until(condition)holds :
    """executes the command,  action,  until expression,  condition,  is true"""
    action()         # do the action step
    if condition():  # finished ?
        return
    else:            # do it again:
        repeat(action)until(condition)holds 
This defines a function named, repeat..action..until. The function is used in a program like this:
...
repeat([x = x - 1])until([x == 0])end
...
The brackets, [..], are quoting the code .., that is, constructing a closure holding the code. Functional languages, like Scheme and Haskell, support this approach, as do Ruby and Smalltalk to a lesser degree.

For older programming languages, the traditional way to add custom control structures is with a macro processor (``preprocessor''). A macro processor is a program that reads as input a program in the host language that has the custom structures mixed into the code. The macro processor locates the occurrences of the custom structures and replaces them with the instructions in the host language that perform the intended operations.

C's preprocessor is a standard but not-too-exciting example. A segment of C code like this,

#define PI 3.14159
#define Double(x)  (x + x)
// now,  PI  and  Double  act like they are built-in C functions:
y = Double(PI * 5) ;
defines two macros, PI and double, which look like functions and can be called like functions. When the above code is input to C's preprocessor, this text is the output:
y = (3.14159 * 5 + 3.14159 * 5) ;
The macro definitions are removed, and the calls are replaced by C-text, giving a program in pure C.

When a macro is called, its arguments are text and not computed values! At a macro call, the text argument is bound to the parameter and the text is inserted for occurrences of the parameter in the macro body. The text computed by the macro's body is copied back in place ofthe macro call. In the example, y = Double(PI * 5) is rewritten to y = (PI * 5 + PI * 5), which is rewritten to y = (3.14159 * 5 + PI * 5), which is rewritten to y = (3.14159 * 5 + 3.14159 * 5). The example shows why the macro processor must be a separate program, run first, before the parser, interpreter or translator. There is a preprocessor, called GPP, that can be used stand-alone to process any program that contains C-like macros. Like C's preprocessor, GPP requires that a macro call look like a function call, of the form, MACRONAME(ARG1, ... ARGn). The m4 macroprocessor lets its user write macro definitions whose calls look somewhat like the mix-fix notation seen in the previous repeat..until..holds example.

In a future section, we will see how to use a language's regular-expression library to code a simple but useful macro processor.

Here are some references for existing macro processors:

Ruby supports a ``block'' construction (the [..] syntax) that makes it possible to code simple customized control structures directly in Ruby. There are some Ruby-implementation approaches at http://weblog.jamisbuck.org/2006/4/20/writing-domain-specific-languages


9.5.1 Implementation techniques

Perhaps this is obvious, but the first question to ask is: What language is understood by the hardware platform that you use in the problem domain? If the hardware language lets you code a parser and interpreter, then you can readily implement a top-down DSPL on the hardware. (Note: a hardware platform might ``understand'' several languages, if there already exist quality interpreters or compilers for the languages on the hardware.)

If the hardware language is not expressive enough, or it is limited in space and speed, you must protoype the top-down DSPL interpreter in a different language and then convert the interpreter into a compiler that translates into the hardware language. Do this as a last resort, since compiler development and maintenance are expensive tasks.

In the case of a bottom-up DSPL, you should select a host language that either (i) is directly understood by the hardware or (ii) has an efficient compiler from the host language to the hardware language. In all cases, the chosen host language must support components and libraries, so that you can extend the host language bottom up.

Although it is almost never done, it is an excellent project to implement a DSPL one way and then use the acquired knowledge to implement it the ``inverse way.'' That is, if you designed a DSPL top-down, try to extract from its interpreter the parts that become components for a bottom-up implementation. Dually, if you first built a bottom-up DSPL, then next use the components as the ``logic'' within a top-down, interpreter implementation. The second version of the language might be the one that you prefer!

Top-down DSPLs and trap doors

If you have designed a top-down DSPL, you should add a ``trap door'' so that code in the implementation language can be embedded in the programs you write. The simplest way to do this is to use an implementation language that has an eval/exec operation.

Here is a small example. Perhaps you have designed a game-app for a cell phone, where a child can tell birds to eat bugs. The game has a GUI front-end, but the mouse moves and clicks on the GUI generate code in this syntax:

===================================================

CL : CommandList       A : Atom
C : Command            S : String

CL ::=  C |  C . CL
C  ::=  A1 eats A2 | do S
A  ::=  bird  |  bug
S  is a quoted string

===================================================
An example program that the GUI might generate is
bird eats bug.
bird eats bug.
bug eats bird
The game has limited functionality (haha), but notice the do command, which is a trap door that lets a programmer insert Python code that directly manipulates the language's interpreter, say, like this:
bird eats bug.
do "census['cat'] = 1\ncensus['bird'] = 0\nprint 'uh oh!'"
The string holds Python code:
census['cat'] = 1
census['bird'] = 0
print 'uh oh!'
Here is the interpreter for the bird-cage language:

===================================================

"""Interpreter for mini top-down DSL for bird-cage domain of birds and bugs.
   Includes trap-door operation,  do S,   for embedding Python source code.  
   Source language syntax to be parsed:
     CL : CommandList       A : Atom
      C : Command           S : String 
           CL ::=  C |  C . CL
           C  ::=  A1 eats A2 | do S
           A  ::=  bird  |  bug
           S  is a quoted string

   Operator-tree structures resulting from the parser:
      CLIST ::=  [ C* ]
      CTREE ::=  ["eat", A1, A2 ]  |  ["do", S]
      A     ::=  "bird"  |  "bug"  
      S     ::=  a quoted string
"""
# Global variable: remembers count of entities in bird cage:
census = {"bird": 9,  "bug": 99}

def interpretCLIST(p) :
    """interprets CLIST  p"""
    for command in p :
        interpretCTREE(command)

def interpretCTREE(c) :
    """interprets CTREE  c"""
    operator = c[0]
    if operator == "eat" :
        eater = c[1] 
        lunch = c[2]
        if census[eater] > 0 and census[lunch] > 0 :
            census[lunch] = census[lunch] - 1
    elif operator == "do" :  # trap-door ``eval'' operation ---
        exec(c[1])    # executes  c[1]  as python code.  Can affect  census,
                      # add new global variables to interpreter's namespace,
                      # print trace information, etc.
    else :  
        crash("invalid command")

def crash(message) :
    print message + "! crash! core dump: ", census
    raise Exception  

def main(program) :
    """interprets the operator tree,  program"""
    interpretCLIST(program)
    print "final census =", census

===================================================
Here are some sample uses of the interpreter:
python -i top.py
>>> main([["eat", "bird", "bug"]])
final census = {'bird': 9, 'bug': 98}

>>> main([["eat", "bird", "bug"], ["do", "census['cat'] = 1\ncensus['bird'] = 0\nprint 'uh oh!'"]])
uh oh!
final census = {'bird': 0, 'bug': 97, 'cat': 1}
The do command lets a programmer escape from the limited functionality of the DSPL and use the operations of the implementation language.

Bottom-up DSPLs and macro expansion

If you have developed a bottom-up DSPL, you should also define control-structure patterns and linking patterns for your DSPL library. It is always best to use host-language facilities to do this.

Some host languages (e.g., Scheme and C) come with their own macro processors. Others (e.g., Smalltalk and Ruby) have flexible procedure-call syntax for defining new patterns. Others (e.g., Perl, PHP, Python, Ruby) supply regular-expression libraries that have powerful pattern-matching operations that you can use to write your own macro processor.

Here is an example of using regular-expression string matching in Python. We use Python's regular-expression module, re, to define a pattern, match the pattern in a string, and replace it. The comments in the code explain how this operates:

===================================================

import re    # re  is the module of regular-expression operations

# Here is a pattern that matches strings of form,  
#    @DOUBLE alpha END
# where  alpha  is some substring that holds no occurrences of  @ :
#     "(\\s*)@DOUBLE\\b([^@]*?)\\bEND\\b"
# where 
#     \\s  means a whitespace character
#     \\b  means a word boundary
#     E*   means match  E  zero or more times as much as possible for success
#     E*?  means match  E  zero or more times as little as possible for success
#     [^c] means match any character that is NOT character  c
# The parens mark _groups_ that are used below.

# p  is a string-matching object compiled from the pattern string:
p = re.compile("(\\s*)@DOUBLE\\b([^@]*?)\\bEND\\b")

# try this multi-line example:
source = """
x = 0
x =  @DOUBLE x END
print x
"""
print "source text ="
print source

# search for compiled pattern  p  in  source:
m = p.search(source)
print

# if the match succeeds,  m  is an object; else  m = None
print "match result =", m
# m  holds a list of substrings that matched parenthesized groups in the pattern:
print "matched groups =", m.groups()

# m  also holds the start and end indexes of the matched string:
print "span of matched text =", m.span()
# the start and end indexes can be referenced individually, too:
print "matched text =", source[m.start() : m.end()]
print

# let's replace the matched string by something else:
matches = m.groups()
source = source[:m.start()]  \
       + matches[0] + "(2 * " + matches[1] + ")" \
       + source[m.end():]
print "updated text ="
print source

# We have completed a simple macro-expansion of  "!DOUBLE alpha END"
# into  "(2 * alpha )", preserving any leading spacing

===================================================
Here is the output from the above script:
source text =

x = 0
x =  @DOUBLE x END
print x


match result = <_sre.SRE_Match object at 0x7ff3d6e0>
matched groups = ('  ', ' x ')
span of matched text = (10, 25)
matched text =   @DOUBLE x END

updated text =

x = 0
x =  (2 *  x )
print x
The example shows that patterns can be complex. There is a tutorial on writing patterns at http://docs.python.org/howto/regex.html and there is a mostly complete listing of pattern options at http://docs.python.org/library/re.html.

We now use the ideas in the example to write a macroprocessor in Python that searches for macro-call patterns and replaces them with expansions. Here are the two macro calls the processor will perform:

@REPEAT Code FOR Expr TIMES  ===>  newvar = Expr
                                   while newvar > 0 :
                                       Code
                                       newvar = newvar - 1

@DOUBLE Expr END  ===>  ((Expr) * 2) 
Each macro call on the left is coded as a pattern string, and each translation is done by a Python-coded handler function. The macro processor's main data structure is a list of (compiled-pattern, handler-function) pairs.

Here is the macro processor:

===================================================

"""Simplistic macroprocessor based on regular expressions.

   main data structure:
      macrotable : list of (COMPILED_PATTERN, HANDLER) pairs

   Example:
   macrotable = [ (re.compile("(\\s*)@REPEAT\\b(\\s*)([^@]*?)\\bFOR\\b([^@]*?)\\bTIMES\\b")
                   translateREPEAT),
                  (re.compile("@DOUBLE\\b([^@]*?)\\bEND\\b"), translateDOUBLE) ]

      holds these two macro definitions:
         indent1 @REPEAT indent2 alpha FOR beta TIMES  
                        =>  translateREPEAT(indent1,indent2,alpha,beta)
         @DOUBLE alpha END  =>  translateDOUBLE(alpha,)

   Compiled patterns, as written above, match macro-call symbol, @,
   followed by keywords (which are required to be separate words by  \\b )
   such that included text arguments do not include any call symbols, @.  

   Note that  E*?  denotes the minimal match of  E*  such that
   the overall pattern match succeeds.  Thus, the macro processor computes
   inside-out processing of macro calls so that nested calls are never confused.

   The pattern for @REPEAT  also records the amount of indentations via  (\\s*).

   Macro-processor algorithm:
   read  source
   repeat until no more macro matches:
       search  source  for each compiled pattern in  macrotable
       if successful match,
          then call accompanying  handler  function,
            which assembles appropriate translation
            insert translation in place of matched pattern in  source
   write source
"""
### This portion should be a separate module that holds the translation
### functions.  It is embedded here for simplicity.

#GENSYM function:
var_count = 0   # count of new names generated for expanded macros
def genNewVar() :
    """genNewVar is a gensym function, generating unique new names
       returns: a string of form, "_varN", where N is a unique nonneg int
    """
    global var_count
    newvar = "_var" + str(var_count)
    var_count = var_count + 1
    return newvar


def translateREPEAT(args) :
    """translateREPEAT  expands this macro call:
       indent1 @REPEAT 
               indent2 Code
               FOR Expr TIMES  =into=>  indent1 newvar = Expr
                                        indent1 while newvar > 0 :
                                                indent2 Code
                                                indent2 newvar = newvar - 1

       where indent1 = args[0]  and  indent2 = args[1]
             Code = args[2]     and  Expr = args[3]
       (indent1  and  indent2  are leading white-space)
       returns: ans, a string holding the macro-expanded call
    """
    indent1 = args[0]
    indent2 = args[1]
    bodycode = args[2]
    exprcode = args[3]
    # the call to REPEAT is replaced by this python code, as documented above:
    newvar = genNewVar()
    ans = indent1 + newvar + " = " + exprcode  \
          + indent1 +  "while " + newvar + " > 0:"  \
          + indent2 + bodycode  \
          + indent2 + newvar + " = " + newvar + " - 1"
    return ans


def translateDOUBLE(arg) :
    """translates   @DOUBLE(arg,) =into=>  '((arg) * 2)' """
    ans = "((" + arg[0] + ") * 2)"
    return ans

### END OF HANDLER-FUNCTION MODULE

### MACRO PROCESSOR CONTROL ALGORITHM:

import re   # import regular-expression module
# initialize macrotable:
macrotable = [ (re.compile("(\\s*)@REPEAT\\b(\\s*)([^@]*?)\\bFOR\\b([^@]*?)\\bTIMES\\b"),
                translateREPEAT), 
               (re.compile("@DOUBLE\\b([^@]*?)\\bEND\\b"), translateDOUBLE)
             ]

# read source:
import sys
if len(sys.argv) < 2 : 
    inputfilename = raw_input("Type input file to copy: ")
else :
    inputfilename = sys.argv[1]
input = open(inputfilename, "r")
source = input.read()
input.close()  

# replace all macro calls:
still_matching = True
while still_matching :
    still_matching = False
    for (pattern, handler) in macrotable :
        match = pattern.search(source)
        if match :  # != None
            replacement = handler(match.groups())
            source = source[:match.start()] + replacement + source[match.end():]
            still_matching = True

# write source:
index = inputfilename.find(".py")
outputfilename = inputfilename[:index] + "out" + ".py"
output = open(outputfilename, "w")
output.write(source)
output.close()

print
print "Contents of " + outputfilename + ":"
print source

===================================================
Say we have this file, test.py, whose contents are:
x = 0
@REPEAT
    x = @DOUBLE x + 1 END
    @REPEAT
        pass
    FOR 2 TIMES
FOR 3 TIMES
print x
When we use the macroprocessor to rewrite the file (python macrop.py test.py), we get this report:
Contents of testout.py:
x = 0
_var1 =  3 
while _var1 > 0:
    x = (( x + 1 ) * 2)
    _var0 =  2 
    while _var0 > 0:
        pass
    
        _var0 = _var0 - 1

    _var1 = _var1 - 1
print x
All macro calls are expanded.


9.6 Further reading

It is tough finding good tutorial material about DSL development. (The people who do these things are in industry, and they have little time to write scholarly papers and books. The few academics who have tried to research the topic have found themselves swallowed up by the application domains, and they either give up or they never come back.) Have fun!