Automata-based programming

Consider the task of reading a text from standard input line-by-line and writing the first word of each line to standard output. First we skip all leading whitespace characters, if any. Then we print all the characters of the first word. Finally we skip all the trailing characters until a newline character is encountered. Whenever a sequence of newline characters is encountered not at the beginning of the stream, we print only the first one and skip the remaining ones; else, we skip all. Next we restart the process at the following line. Upon encountering the end-of-file condition (regardless of the stage), we stop.

A traditional program in C which performs the above task could look like this:

#include <ctype.h>
#include <stdio.h>


int main(void) {
  int c;

  do {
    do {
      c = getchar();
    } while (isspace(c));

    while (!isspace(c) && c != EOF) {
      putchar(c);
      c = getchar();
    }
    
    while (c != '\n' && c != EOF) {
      c = getchar();
    }
    
    if (c == '\n') {
      putchar(c);
    }
  } while (c != EOF);

  return 0;
}

For instance, compiling and running the above program on this input:

$ clang program.c && (printf "\t\v\f\r \n\n\t\v\f\r foo bar baz\n\n\t\v\f\r qux quux corge" | ./a.out)

yields:

foo
qux

The same task can be solved by thinking in terms of finite-state machines. The parsing of a line has three stages: skipping the leading whitespace characters, printing the characters of the first word and skipping the trailing characters. Let's call these automaton states BEFORE, INSIDE and AFTER. An automata-based version of the program could look like this:

#include <ctype.h>
#include <stdio.h>

enum State {BEFORE, INSIDE, AFTER};


int main(void) {
  int c;
  enum State s = BEFORE;

  while ((c = getchar()) != EOF) {
    switch (s) {
      case BEFORE:
        if (!isspace(c)) {
          putchar(c);
          s = INSIDE;
        }
        
        break;
      case INSIDE:
        if (c == '\n') {
          putchar(c);
          s = BEFORE;
        } else if (isspace(c)) {
          s = AFTER;
        } else {
          putchar(c);
        }
          
        break;
      case AFTER:
        if (c == '\n') {
          putchar(c);
          s = BEFORE;
        }
        
        break;
    }
  }

  return 0;
}

Although the program now looks longer, it has at least one significant advantage: there is only one reading (that is, call to the getchar function) instruction. Besides that, there is only one loop instead of the four the traditional version had. The body of the while loop is the automaton step and the loop itself is the cycle of the automaton step. The program implements the work of a finite-state machine shown in the state diagram.

The most important property of the program is that the automaton step code section is clearly localized. With an explicit function step for the automation step, the program better demonstrates this property:

#include <ctype.h>
#include <stdio.h>

enum State {BEFORE, INSIDE, AFTER};


void step(enum State* const s, int const c) {
  switch (*s) {
    case BEFORE:
      if (!isspace(c)) {
        putchar(c);
        *s = INSIDE;
      }
      
      break;
    case INSIDE:
      if (c == '\n') {
        putchar(c);
        *s = BEFORE;
      } else if (isspace(c)) {
        *s = AFTER;
      } else {
        putchar(c);
      }
        
      break;
    case AFTER:
      if (c == '\n') {
        putchar(c);
        *s = BEFORE;
      }
      
      break;
  }
}


int main(void) {
  int c;
  enum State s = BEFORE;

  while ((c = getchar()) != EOF) {
    step(&s, c);
  }

  return 0;
}

The program now clearly demonstrates the basic properties of automata-based code:

• time periods of automaton step executions may not overlap;
• the only information passed from the previous step to the next is the explicitly specified automaton state.

A finite automaton can be defined by a state-transition table whose rows stand for the current states, columns stand for the inputs, and cells stand for the next states and actions to perform.

State diagram

The state diagram of a finite-state machine that prints the first word of each line of an input stream. The machine follows exactly one transition on each step, depending on the current state and the encountered character. The action accompanying a transition is either a no-operation or the printing of the encountered character, denoted with print.

Generally speaking, an automata-based program can naturally use this approach. With an explicit two-dimensional array transitions for the state-transition table, the program uses this approach:

#include <ctype.h>
#include <stdio.h>

enum State {BEFORE, INSIDE, AFTER};


void nop(int const c) {}


void print(int const c) {
  putchar(c);
}


struct Branch {
  enum State const next_state;
  void (*action)(int);
};


struct Branch const transitions[3][3] = {
  //   newline         whitespace         other             Inputs/States
  {{BEFORE,   &nop}, {BEFORE, &nop}, {INSIDE, &print}},  // before
  {{BEFORE, &print}, {AFTER,  &nop}, {INSIDE, &print}},  // inside
  {{BEFORE, &print}, {AFTER,  &nop}, {AFTER,    &nop}}   // after
};


void step(enum State* const s, int const c) {
  int const row = (*s == BEFORE) ? 0 : (*s == INSIDE) ? 1 : 2;
  int const column = (c == '\n') ? 0 : isspace(c) ? 1 : 2;
  struct Branch const* const b = &transitions[row][column];
  *s = b->next_state;
  b->action(c);
}


int main(void) {
  int c;
  enum State s = BEFORE;

  while ((c = getchar()) != EOF) {
    step(&s, c);
  }

  return 0;
}

If the implementation language supports object-oriented programming, a simple refactoring of the program is to encapsulate the automaton into an object, thus hiding its implementation details. The program in C++ using object-oriented style could look like this:

#include <ctype.h>
#include <stdio.h>

enum State {BEFORE, INSIDE, AFTER};


struct Branch {
  enum State const next_state;
  void (*action)(int);
};


class StateMachine {
  public:
    StateMachine();
    void feedChar(int);

  protected:
    static void nop(int);
    static void print(int);

  private:
    enum State _state;
    static struct Branch const _transitions[3][3];
};


StateMachine::StateMachine(): _state(BEFORE) {}


void StateMachine::feedChar(int const c) {
  int const row = (_state == BEFORE) ? 0 : (_state == INSIDE) ? 1 : 2;
  int const column = (c == '\n') ? 0 : isspace(c) ? 1 : 2;
  struct Branch const* const b = &_transitions[row][column];
  _state = b->next_state;
  b->action(c);
}


void StateMachine::nop(int const c) {}


void StateMachine::print(int const c) {
  putchar(c);
}


struct Branch const StateMachine::_transitions[3][3] = {
  //   newline         whitespace         other             Inputs/States
  {{BEFORE,   &nop}, {BEFORE, &nop}, {INSIDE, &print}},  // before
  {{BEFORE, &print}, {AFTER,  &nop}, {INSIDE, &print}},  // inside
  {{BEFORE, &print}, {AFTER,  &nop}, {AFTER,    &nop}}   // after
};


int main() {
  int c;
  StateMachine m;

  while ((c = getchar()) != EOF) {
    m.feedChar(c);
  }

  return 0;
}

To minimize changes not directly related to the subject of the article, the input/output getchar and putchar functions from the standard library of C are being used.

The state design pattern is a way for an object to change its behavior at runtime according to its internal state without resorting to large conditional statements or table lookups thanks to virtual function calls. Its main advantage over code using large conditional statements is that state-specific code is distributed across different objects rather than localized in a monolithic block, which improves maintainability. Its main advantages over code using state-transition tables are that virtual function calls are often more efficient than table lookups, that state-transition criteria are more explicit than in tabular format, and that it is easier to add actions accompanying state transitions. However it introduces a new problem: the number of classes makes the code less compact than the other approaches. The program using the state design pattern could look like this:

#include <ctype.h>
#include <stdio.h>

class StateMachine;


class State {
  public:
    virtual void feedChar(StateMachine*, int) const = 0;
};


class Before: public State {
  public:
    static State const* instantiate();
    virtual void feedChar(StateMachine*, int) const override;

  protected:
    Before() = default;

  private:
    static State const* _instance;
};


class Inside: public State {
  public:
    static State const* instantiate();
    virtual void feedChar(StateMachine*, int) const override;

  protected:
    Inside() = default;

  private:
    static State const* _instance;
};


class After: public State {
  public:
    static State const* instantiate();
    virtual void feedChar(StateMachine*, int) const override;

  protected:
    After() = default;

  private:
    static State const* _instance;
};


class StateMachine {
  public:
    StateMachine();
    void feedChar(int);

  protected:
    void setState(State const*);

  private:
    State const* _state;
    friend class Before;
    friend class Inside;
    friend class After;
};


State const* Before::instantiate() {
  if (!_instance) {
    _instance = new Before;
  }

  return _instance;
}


void Before::feedChar(StateMachine* const m, int const c) const {
  if (!isspace(c)) {
    putchar(c);
    m->setState(Inside::instantiate());
  }
}


State const* Before::_instance = nullptr;


State const* Inside::instantiate() {
  if (!_instance) {
    _instance = new Inside;
  }

  return _instance;
}


void Inside::feedChar(StateMachine* const m, int const c) const {
  if (c == '\n') {
    putchar(c);
    m->setState(Before::instantiate());
  } else if (isspace(c)) {
    m->setState(After::instantiate());
  } else {
    putchar(c);
  }
}


State const* Inside::_instance = nullptr;


State const* After::instantiate() {
  if (!_instance) {
    _instance = new After;
  }

  return _instance;
}


void After::feedChar(StateMachine* const m, int const c) const {
  if (c == '\n') {
    putchar(c);
    m->setState(Before::instantiate());
  }
}


State const* After::_instance = nullptr;


StateMachine::StateMachine(): _state(Before::instantiate()) {}


void StateMachine::feedChar(int const c) {
  _state->feedChar(this, c);
}


void StateMachine::setState(State const* const s) {
  _state = s;
}


int main() {
  int c;
  StateMachine m;

  while ((c = getchar()) != EOF) {
    m.feedChar(c);
  }

  return 0;
}

Automata-based programming indeed closely matches the programming needs found in the field of automation.

A production cycle is commonly modelled as:

• a sequence of stages stepping according to input data (from captors);
• a set of actions performed depending on the current stage.

Various dedicated programming languages allow expressing such a model in more or less sophisticated ways.

The example presented above could be expressed according to this view like in the following pseudo-code ('set' activates a logic variable, 'reset' inactivates a logic variable, ':' assigns a variable, and '=' tests for equality):

newline: '\n'
whitespace: ('\t', '\n', '\v', '\f', '\r', ' ')
states: (before, inside, after)


setState(c) {
  if before and (c != newline and c not in whitespace) then set inside
  if inside then (if c in whitespace then set after else if c = newline then set before)
  if after and c = newline then set before
}


doAction(c) {
  if before and (c != newline and c not in whitespace) then write(c)
  if inside and c not in whitespace then write(c)
  if after and c = newline then write(c)
}


cycle {
  set before

  loop until (c: readCharacter) = EOL {
    setState(c)
    doAction(c)
  }
}

The separation of routines expressing cycle progression on one side, and actual action on the other (matching input and output) allows clearer and simpler code.

In the field of automation, stepping from step to step depends on input data coming from the machine itself. This is represented in the program by reading characters from a text. In reality, those data inform about position, speed, temperature, etc. of critical elements of a machine.

Like in GUI programming, changes in the machine state can thus be considered as events causing the passage from a state to another, until the final one is reached. The combination of possible states can generate a wide variety of events, thus defining a more complex production cycle. As a consequence, cycles are usually far to be simple linear sequences. There are commonly parallel branches running together and alternatives selected according to different events, schematically represented below:

   s:stage   c:condition
   
   s1
   |
   |-c2
   |
   s2
   |
   ----------
   |        |
   |-c31    |-c32
   |        |
  s31       s32
   |        |
   |-c41    |-c42
   |        |
   ----------
   |
   s4