The Text Editor sam

Rob Pike
rob@plan9.bell-labs.com

ABSTRACT

Sam is an interactive multi-file text editor intended for bitmap displays. A textual command language supplements the mouse-driven, cut-and-paste interface to make complex or repetitive editing tasks easy to specify. The language is characterized by the composition of regular expressions to describe the structure of the text being modified. The treatment of files as a database, with changes logged as atomic transactions, guides the implementation and makes a general `undo' mechanism straightforward.

Sam is implemented as two processes connected by a low-bandwidth stream, one process handling the display and the other the editing algorithms. Therefore it can run with the display process in a bitmap terminal and the editor on a local host, with both processes on a bitmap-equipped host, or with the display process in the terminal and the editor in a remote host. By suppressing the display process, it can even run without a bitmap terminal.

This paper is reprinted from Software­Practice and Experience, Vol 17, number 11, pp. 813-845, November 1987. The paper has not been updated for the Plan 9 manuals. Although Sam has not changed much since the paper was written, the system around it certainly has. Nonetheless, the description here still stands as the best introduction to the editor.

Introduction

Sam edits uninterpreted ASCII text. It has no facilities for multiple fonts, graphics or tables, unlike MacWrite,² Bravo,³ Tioga⁴ or Lara.⁵ Also unlike them, it has a rich command language. (Throughout this paper, the phrase command language refers to textual commands; commands activated from the mouse form the mouse language.) Sam developed as an editor for use by programmers, and tries to join the styles of the UNIX text editor ed^6,7 with that of interactive cut-and-paste editors by providing a comfortable mouse-driven interface to a program with a solid command language driven by regular expressions. The command language developed more than the mouse language, and acquired a notation for describing the structure of files more richly than as a sequence of lines, using a dataflow-like syntax for specifying changes.

The interactive style was influenced by jim,¹ an early cut-and-paste editor for the Blit, and by mux,⁸ the Blit window system. Mux merges the original Blit window system, mpx,¹ with cut-and-paste editing, forming something like a multiplexed version of jim that edits the output of (and input to) command sessions rather than files.

The first part of this paper describes the command language, then the mouse language, and explains how they interact. That is followed by a description of the implementation, first of the host part, then of the terminal part. A principle that influenced the design of sam is that it should have no explicit limits, such as upper limits on file size or line length. A secondary consideration is that it be efficient. To honor these two goals together requires a method for efficiently manipulating huge strings (files) without breaking them into lines, perhaps while making thousands of changes under control of the command language. Sam's method is to treat the file as a transaction database, implementing changes as atomic updates. These updates may be unwound easily to `undo' changes. Efficiency is achieved through a collection of caches that minimizes disc traffic and data motion, both within the two parts of the program and between them.

The terminal part of sam is fairly straightforward. More interesting is how the two halves of the editor stay synchronized when either half may initiate a change. This is achieved through a data structure that organizes the communications and is maintained in parallel by both halves.

The last part of the paper chronicles the writing of sam and discusses the lessons that were learned through its development and use.

The paper is long, but is composed largely of two papers of reasonable length: a description of the user interface of sam and a discussion of its implementation. They are combined because the implementation is strongly influenced by the user interface, and vice versa.

The Interface

sam file1 file2 ...

Sam is usually connected to a bitmap display that presents a cut-and-paste editor driven by the mouse. In this mode, the command language is still available: text typed in a special window, called the sam window, is interpreted as commands to be executed in the current file. Cut-and-paste editing may be used in any window ­ even in the sam window to construct commands. The other mode of operation, invoked by starting sam with the option -d (for `no download'), does not use the mouse or bitmap display, but still permits editing using the textual command language, even on an ordinary terminal, interactively or from a script.

The following sections describe first the command language (under sam -d and in the sam window), and then the mouse interface. These two languages are nearly independent, but connect through the current text, described below.

The Command Language

The current file is the file to which editing commands refer. The current text is therefore dot in the current file. If a command doesn't explicitly name a particular file or piece of text, the command is assumed to apply to the current text. For the moment, ignore the presence of multiple files and consider editing a single file.


Figure 1. A typical sam screen, with the editing menu presented. The sam (command language) window is in the middle, with file windows above and below. (The user interface makes it easy to create these abutting windows.) The partially obscured window is a third file window. The uppermost window is that to which typing and mouse operations apply, as indicated by its heavy border. Each window has its current text highlighted in reverse video. The sam window's current text is the null string on the last visible line, indicated by a vertical bar. See also Figure 2.

Commands have one-letter names. Except for non-editing commands such as writing the file to disc, most commands make some change to the text in dot and leave dot set to the text resulting from the change. For example, the delete command, d, deletes the text in dot, replacing it by the null string and setting dot to the result. The change command, c, replaces dot by text delimited by an arbitrary punctuation character, conventionally a slash. Thus,

c/Peter/

a/Peter/

i/Peter/

Newlines are part of the syntax of commands: the newline character lexically terminates a command. Within the inserted text, however, newlines are never implicit. But since it is often convenient to insert multiple lines of text, sam has a special syntax for that case:

a
some lines of text
to be inserted in the file,
terminated by a period
on a line by itself
.

c/\n/

Sam also has a substitute command, s:

s/expression/replacement/

s/t/st/

s/Peter/Oh, &, &, &, &!/

There are also three commands that apply programs to text:

< UNIX program

| sort

One more command: p prints the contents of dot. Table I summarizes sam's commands.

The value of dot may be changed by specifying an address for the command. The simplest address is a line number:

3

3d

An address is always a single substring of the file. Thus, the address 3 addresses the characters after the second newline of the file through the third newline of the file. A compound address is constructed by the comma operator

address1,address2

These addresses are all absolute positions in the file, but sam also has relative addresses, indicated by + or -. For example,

$-3

.+1

The + operator acts relative to the end of its first argument, while the - operator acts relative to the beginning. Thus .+1 addresses the first line after dot, .- addresses the first line before dot, and +- refers to the line containing the end of dot. (Dot may span multiple lines, and + selects the line after the end of dot, then - backs up one line.)

The final type of address is a regular expression, which addresses the text matched by the expression. The expression is enclosed in slashes, as in

/expression/

/x/

/xx*/

/x|Peter/

/.*/

Regular expressions are always relative addresses. The direction is forwards by default, so /Peter/ is really an abbreviation for +/Peter/. The search can be reversed with a minus sign, so

-/Peter/

The language discussed so far will not seem novel to people who use UNIX text editors such as ed or vi.⁹ Moreover, the kinds of editing operations these commands allow, with the exception of regular expressions and line numbers, are clearly more conveniently handled by a mouse-based interface. Indeed, sam's mouse language (discussed at length below) is the means by which simple changes are usually made. For large or repetitive changes, however, a textual language outperforms a manual interface.

Imagine that, instead of deleting just one occurrence of the string Peter, we wanted to eliminate every Peter. What's needed is an iterator that runs a command for each occurrence of some text. Sam's iterator is called x, for extract:

x/expression/ command

0,$ x/Peter/ d

, x/Peter/ p

Of course, the text extracted by x may be selected by a regular expression, which complicates deciding what set of matches is chosen ­ matches may overlap. This is resolved by generating the matches starting from the beginning of dot using the leftmost-longest rule, and searching for each match starting from the end of the previous one. Regular expressions may also match null strings, but a null match adjacent to a non-null match is never selected; at least one character must intervene. For example,

, c/AAA/
x/B*/ c/-/
, p

-A-A-A-

The x command has a complement, y, with similar syntax, that executes the command with dot set to the text between the matches of the expression. For example,

, c/AAA/
y/A/ c/-/
, p

The x and y commands are looping constructs, and sam has a pair of conditional commands to go with them. They have similar syntax:

g/expression/ command

, x/Peter/ d

, g/Peter/ d

These control-structure-like commands may be composed to construct more involved operations. For example, to print those lines of text that contain the string Peter:

, x/.*\n/ g/Peter/ p

Composition may be continued indefinitely.

, x/.*\n/ g/Peter/ v/SaltPeter/ p

Structural Regular Expressions

Herbert Tic
44 Turnip Ave., Endive, NJ
201-5555642

Norbert Twinge
16 Potato St., Cabbagetown, NJ
201-5553145

...

(.+\n)+

, x/(.+\n)+/ g/^Herbert Tic$/ p

, x/(.+\n)+/ g/^Herbert Tic$/ x/^[0-9]*-[0-9]*\n/ p

A more involved problem is that of renaming a particular variable, say n, to num in a C program. The obvious first attempt,

, x/n/ c/num/

, x/[A-Za-z_][A-Za-z_0-9]*/ g/n/ v/../ c/num/

The point of these composed commands is successive refinement. A simple version of the command is tried, and if it's not good enough, it can be honed by adding a clause or two. (Mistakes can be undone; see below. Also, the mouse language makes it unnecessary to retype the command each time.) The resulting chains of commands are somewhat reminiscent of shell pipelines.⁷ Unlike pipelines, though, which pass along modified data, sam commands pass a view of the data. The text at each step of the command is the same, but which pieces are selected is refined step by step until the correct piece is available to the final step of the command line, which ultimately makes the change.

In other UNIX programs, regular expressions are used only for selection, as in the sam g command, never for extraction as in the x or y command. For example, patterns in awk⁷ are used to select lines to be operated on, but cannot be used to describe the format of the input text, or to handle newline-free text. The use of regular expressions to describe the structure of a piece of text rather than its contents, as in the x command, has been given a name: structural regular expressions. When they are composed, as in the above example, they are pleasantly expressive. Their use is discussed at greater length elsewhere.¹⁰

Multiple files

e discfilename

w discfilename

r discfilename

f discfilename

'-. discfilename

Sam may be started with a set of disc files (such as all the source for a program) by invoking it with a list of file names as arguments, and more may be added or deleted on demand.

B discfile1 discfile2 ...

D discfile1 discfile2 ...

B <echo *.c

B <grep -l variable *.c

There are two ways to change which file is current:

b filename

"expression" address

"peter.c" 3

0,$ t "peter.c" 0

The X command is a looping construct, like x, that refers to files instead of strings:

X/expression/ command

X/'/ w

Y/\.c/ D

Braces allow commands to be grouped, so

{
	command1
	command2
}

X/\.c/ ,g/variable/ {
	f
	, x/.*\n/ g/variable/ p
}

Finally, the undo command, u, undoes the last command, no matter how many files were affected. Multiple undo operations move further back in time, so

u
u

e filename
u

Mouse Interface

Each file may have zero or more windows open on the display. At any time, only one window in all of sam is the current window, that is, the window to which typing and mouse actions refer; this may be the sam window (that in which commands may be typed) or one of the file windows. When a file has multiple windows, the image of the file in each window is always kept up to date. The current file is the last file affected by a command, so if the sam window is current, the current window is not a window on the current file. However, each window on a file has its own value of dot, and when switching between windows on a single file, the file's value of dot is changed to that of the window. Thus, flipping between windows behaves in the obvious, convenient way.

The mouse on the Blit has three buttons, numbered left to right. Button 3 has a list of commands to manipulate windows, followed by a list of `menu lines' exactly as printed by the f command, one per file (not one per window). These menu lines are sorted by file name. If the list is long, the Blit menu software will make it more manageable by generating a scrolling menu instead of an unwieldy long list. Using the menu to select a file from the list makes that file the current file, and the most recently current window in that file the current window. But if that file is already current, selecting it in the menu cycles through the windows on the file; this simple trick avoids a special menu to choose windows on a file. If there is no window open on the file, sam changes the mouse cursor to prompt the user to create one.

The commands on the button 3 menu are straightforward (see Figure 3), and are like the commands to manipulate windows in mux,⁸ the Blit's window system. New makes a new file, and gives it one empty window, whose size is determined by a rectangle swept by the mouse. Zerox prompts for a window to be selected, and makes a clone of that window; this is how multiple windows are created on one file. Reshape changes the size of the indicated window, and close deletes it. If that is the last window open on the file, close first does a D command on the file. Write is identical to a w command on the file; it is in the menu purely for convenience. Finally, ~~sam~~ is a menu item that appears between the commands and the file names. Selecting it makes the sam window the current window, causing subsequent typing to be interpreted as commands.


Figure 3. The menu on button 3. The black rectangle on the left is a scroll bar; the menu is limited to the length shown to prevent its becoming unwieldy. Above the ~~sam~~ line is a list of commands; beneath it is a list of files, presented exactly as with the f command.

When sam requests that a window be swept, in response to new, zerox or reshape, it changes the mouse cursor from the usual arrow to a box with a small arrow. In this state, the mouse may be used to indicate an arbitrary rectangle by pressing button 3 at one corner and releasing it at the opposite corner. More conveniently, button 3 may simply be clicked, whereupon sam creates the maximal rectangle that contains the cursor and abuts the sam window. By placing the sam window in the middle of the screen, the user can define two regions (one above, one below) in which stacked fully-overlapping windows can be created with minimal fuss (see Figure 1). This simple user interface trick makes window creation noticeably easier.

The cut-and-paste editor is essentially the same as that in Smalltalk-80.¹¹ The text in dot is always highlighted on the screen. When a character is typed it replaces dot, and sets dot to the null string after the character. Thus, ordinary typing inserts text. Button 1 is used for selection: pressing the button, moving the mouse, and lifting the button selects (sets dot to) the text between the points where the button was pressed and released. Pressing and releasing at the same point selects a null string; this is called clicking. Clicking twice quickly, or double clicking, selects larger objects; for example, double clicking in a word selects the word, double clicking just inside an opening bracket selects the text contained in the brackets (handling nested brackets correctly), and similarly for parentheses, quotes, and so on. The double-clicking rules reflect a bias toward programmers. If sam were intended more for word processing, double-clicks would probably select linguistic structures such as sentences.

If button 1 is pressed outside the current window, it makes the indicated window current. This is the easiest way to switch between windows and files.

Pressing button 2 brings up a menu of editing functions (see Figure 4). These mostly apply to the selected text: cut deletes the selected text, and remembers it in a hidden buffer called the snarf buffer, paste replaces the selected text by the contents of the snarf buffer, snarf just copies the selected text to the snarf buffer, look searches forward for the next literal occurrence of the selected text, and exchanges snarf buffers with the window system in which sam is running. Finally, the last regular expression used appears as a menu entry to search forward for the next occurrence of a match for the expression.


Figure 4. The menu on button 2. The bottom entry tracks the most recently used regular expression, which may be literal text.

The relationship between the command language and the mouse language is entirely due to the equality of dot and the selected text chosen with button 1 on the mouse. For example, to make a set of changes in a C subroutine, dot can be set by double clicking on the left brace that begins the subroutine, which sets dot for the command language. An address-free command then typed in the sam window will apply only to the text between the opening and closing braces of the function. The idea is to select what you want, and then say what you want to do with it, whether invoked by a menu selection or by a typed command. And of course, the value of dot is highlighted on the display after the command completes. This relationship between mouse interface and command language is clumsy to explain, but comfortable, even natural, in practice.

The Implementation

Parsing and execution

x/.*\n/ g/Peter/ a
one line about Peter
another line about Peter
.

The other advantage is specific to the two-process structure of sam. The host process must decide when a command is completed so the command interpreter can be called. This problem is easily resolved by having the lexical analyzer read the single stream of events from the terminal, directly executing all typing and mouse commands, but passing to the parser characters typed to the sam command window. This scheme is slightly complicated by the availability of cut-and-paste editing in the sam window, but that difficulty is resolved by applying the rules used in mux: when a newline is typed to the sam window, all text between the newline and the previously typed newline is made available to the parser. This permits arbitrary editing to be done to a command before typing newline and thereby requesting execution.

The parser is driven by a table because the syntax of addresses and commands is regular enough to be encoded compactly. There are few special cases, such as the replacement text in a substitution, so the syntax of almost all commands can be encoded with a few flags. These include whether the command allows an address (for example, e does not), whether it takes a regular expression (as in x and s), whether it takes replacement text (as in c or i), which may be multi-line, and so on. The internal syntax of regular expressions is handled by a separate parser; a regular expression is a leaf of the command parse tree. Regular expressions are discussed fully in the next section.

The parser table also has information about defaults, so the interpreter is always called with a complete tree. For example, the parser fills in the implicit 0 and $ in the abbreviated address , (comma), inserts a + to the left of an unadorned regular expression in an address, and provides the usual default address . (dot) for commands that expect an address but are not given one.

Once a complete command is parsed, the evaluation is easy. The address is evaluated left-to-right starting from the value of dot, with a mostly ordinary expression evaluator. Addresses, like many of the data structures in sam, are held in a C structure and passed around by value:

typedef long Posn;    /* Position in a file */
typedef struct Range{
        Posn    p1, p2;
}Range;
typedef struct Address{
        Range   r;
        File    *f;
}Address;

The address interpreter is an Address-valued function that traverses the parse tree describing an address (the parse tree for the address has type Addrtree):

Address
address(ap, a, sign)
	Addrtree *ap;
	Address a;
	int sign;
{
	Address a2;
	do
		switch(ap->type){
		case '.':
			a=a.f->dot;
			break;
		case '$':
			a.r.p1=a.r.p2=a.f->nbytes;
			break;
		case '"':	
			a=matchfile(a, ap->aregexp)->dot; 
			break;
		case ',':
			a2=address(ap->right, a, 0);
			a=address(ap->left, a, 0);
			if(a.f!=a2.f || a2.r.p2<a.r.p1)
				error(Eorder);
			a.r.p2=a2.r.p2;
			return a;
		/* and so on */
		}
	while((ap=ap->right)!=0);
	return a;
}

Throughout, errors are handled by a non-local goto (a setjmp/longjmp in C terminology) hidden in a routine called error that immediately aborts the execution, retracts any partially made changes (see the section below on `undoing'), and returns to the top level of the parser. The argument to error is an enumeration type that is translated to a terse but possibly helpful message such as `?addresses out of order.' Very common messages are kept short; for example the message for a failed regular expression search is `?search.'

Character addresses such as #3 are trivial to implement, as the File data structure is accessible by character number. However, sam keeps no information about the position of newlines ­ it is too expensive to track dynamically ­ so line addresses are computed by reading the file, counting newlines. Except in very large files, this has proven acceptable: file access is fast enough to make the technique practical, and lines are not central to the structure of the command language.

The command interpreter, called cmdexec, is also straightforward. The parse table includes a function to call to interpret a particular command. That function receives as arguments the calculated address for the command and the command tree (of type Cmdtree), which may contain information such as the subtree for compound commands. Here, for example, is the function for the g and v commands:

int
g_cmd(a, cp)
	Address a;
	Cmdtree *cp;
{
	compile(cp->regexp);
	if(execute(a.f, a.r.p1, a.r.p2)!=(cp->cmdchar=='v')){
		a.f->dot=a;
		return cmdexec(a, cp->subcmd);
	}
	return TRUE;	/* cause execution to continue */
}

Regular expressions

The expressions are compiled by a routine, compile, that generates the description of the non-deterministic finite state machine. A second routine, execute, interprets the machine to generate the leftmost-longest match of the expression in a substring of the file. The algorithm is described elsewhere.^12,13 Execute reports whether a match was found, and sets a global variable, of type Range, to the substring matched.

A trick is required to evaluate the expression in reverse, such as when searching backwards for an expression. For example,

-/P.*r/

Execute generates only one match each time it is called. To interpret looping constructs such as the x command, sam must therefore synchronize between calls of execute to avoid problems with null matches. For example, even given the leftmost-longest rule, the expression a* matches three times in the string ab (the character a, the null string between the a and b, and the final null string). After returning a match for the a, sam must not match the null string before the b. The algorithm starts execute at the end of its previous match, and if the match it returns is null and abuts the previous match, rejects the match and advances the initial position one character.

Memory allocation

Objects to be allocated in sam are of two flavors: the first is C structs, which are small and often addressed by pointer variables; the second is variable-sized arrays of characters or integers whose base pointer is always used to access them. The memory allocator in sam is therefore in two parts: first, a traditional first-fit allocator that provides fixed storage for structs; and second, a garbage-compacting allocator that reduces storage overhead for variable-sized objects, at the cost of some bookkeeping. The two types of objects are allocated from adjoining arenas, with the garbage-compacting allocator controlling the arena with higher addresses. Separating into two arenas simplifies compaction and prevents fragmentation due to immovable objects. The access rules for garbage-compactable objects (discussed in the next paragraph) allow them to be relocated, so when the first-fit arena needs space, it moves the garbage-compacted arena to higher addresses to make room. Storage is therefore created only at successively higher addresses, either when more garbage-compacted space is needed or when the first-fit arena pushes up the other arena.

Objects that may be compacted declare to the allocator a cell that is guaranteed to be the sole repository of the address of the object whenever a compaction can occur. The compactor can then update the address when the object is moved. For example, the implementation of type List (really a variable-length array) is:

typedef struct List{
        int     nused;
        long    *ptr;
}List;

A related matter of programming style: sam frequently passes structures by value, which simplifies the code. Traditionally, C programs have passed structures by reference, but implicit allocation on the stack is easier to use. Structure passing is a relatively new feature of C (it is not in the standard reference manual for C¹⁴), and is poorly supported in most commercial C compilers. It's convenient and expressive, though, and simplifies memory management by avoiding the allocator altogether and eliminating pointer aliases.

Data structures for manipulating files

The File data type is constructed from three simpler data structures that hold arrays of characters. Each of these types has an insertion and deletion operator, and the insertion and deletion operators of the File type itself are constructed from them.

The simplest type is the String, which is used to hold strings in main memory. The code that manages Strings guarantees that they will never be longer than some moderate size, and in practice they are rarely larger than 8 Kbytes. Strings have two purposes: they hold short strings like file names with little overhead, and because they are deliberately small, they are efficient to modify. They are therefore used as the data structure for in-memory caches.

The disc copy of the file is managed by a data structure called a Disc, which corresponds to a temporary file. A Disc has no storage in main memory other than bookkeeping information; the actual data being held is all on the disc. To reduce the number of open files needed, sam opens a dozen temporary UNIX files and multiplexes the Discs upon them. This permits many files to be edited; the entire sam source (48 files) may be edited comfortably with a single instance of sam. Allocating one temporary file per Disc would strain the operating system's limit on the number of open files. Also, spreading the traffic among temporary files keeps the files shorter, and shorter files are more efficiently implemented by the UNIX I/O subsystem.

A Disc is an array of fixed-length blocks, each of which contains between 1 and 4096 characters of active data. (The block size of our UNIX file system is 4096 bytes.) The block addresses within the temporary file and the length of each block are stored in a List. When changes are made the live part of blocks may change size. Blocks are created and coalesced when necessary to try to keep the sizes between 2048 and 4096 bytes. An actively changing part of the Disc therefore typically has about a kilobyte of slop that can be inserted or deleted without changing more than one block or affecting the block order. When an insertion would overflow a block, the block is split, a new one is allocated to receive the overflow, and the memory-resident list of blocks is rearranged to reflect the insertion of the new block.

Obviously, going to the disc for every modification to the file is prohibitively expensive. The data type Buffer consists of a Disc to hold the data and a String that acts as a cache. This is the first of a series of caches throughout the data structures in sam. The caches not only improve performance, they provide a way to organize the flow of data, particularly in the communication between the host and terminal. This idea is developed below, in the section on communications.

To reduce disc traffic, changes to a Buffer are mediated by a variable-length string, in memory, that acts as a cache. When an insertion or deletion is made to a Buffer, if the change can be accommodated by the cache, it is done there. If the cache becomes bigger than a block because of an insertion, some of it is written to the Disc and deleted from the cache. If the change does not intersect the cache, the cache is flushed. The cache is only loaded at the new position if the change is smaller than a block; otherwise, it is sent directly to the Disc. This is because large changes are typically sequential, whereupon the next change is unlikely to overlap the current one.

A File comprises a String to hold the file name and some ancillary data such as dot and the modified bit. The most important components, though, are a pair of Buffers, one called the transcript and the other the contents. Their use is described in the next section.

The overall structure is shown in Figure 5. Although it may seem that the data is touched many times on its way from the Disc, it is read (by one UNIX system call) directly into the cache of the associated Buffer; no extra copy is done. Similarly, when flushing the cache, the text is written directly from the cache to disc. Most operations act directly on the text in the cache. A principle applied throughout sam is that the fewer times the data is copied, the faster the program will run (see also the paper by Waite¹⁵).

The contents of a File are accessed by a routine that copies to a buffer a substring of a file starting at a specified offset. To read a byte at a time, a per-File array is loaded starting from a specified initial position, and bytes may then be read from the array. The implementation is done by a macro similar to the C standard I/O getc macro.¹⁴ Because the reading may be done at any address, a minor change to the macro allows the file to be read backwards. This array is read-only; there is no putc.

Doing and undoing

Sam uses a two-pass algorithm for making changes, and treats each file as a database against which transactions are registered. Changes are not made directly to the contents. Instead, when a command is started, a `mark' containing a sequence number is placed in the transcript Buffer, and each change made to the file, either an insertion or deletion or a change to the file name, is appended to the end of the transcript. When the command is complete, the transcript is rewound to the mark and applied to the contents.

One reason for separating evaluation from application in this way is to simplify tracking the addresses of changes made in the middle of a long sequence. The two-pass algorithm also allows all changes to apply to the original data: no change can affect another change made in the same command. This is particularly important when evaluating an x command because it prevents regular expression matches from stumbling over changes made earlier in the execution. Also, the two-pass algorithm is cleaner than the way other UNIX editors allow changes to affect each other; for example, ed's idioms to do things like delete every other line depend critically on the implementation. Instead, sam's simple model, in which all changes in a command occur effectively simultaneously, is easy to explain and to understand.

The records in the transcript are of the form ``delete substring from locations 123 to 456'' and ``insert 11 characters `hello there' at location 789.'' (It is an error if the changes are not at monotonically greater positions through the file.) While the update is occurring, these numbers must be offset by earlier changes, but that is straightforward and local to the update routine; moreover, all the numbers have been computed before the first is examined.

Treating the file as a transaction system has another advantage: undo is trivial. All it takes is to invert the transcript after it has been implemented, converting insertions into deletions and vice versa, and saving them in a holding Buffer. The `do' transcript can then be deleted from the transcript Buffer and replaced by the `undo' transcript. If an undo is requested, the transcript is rewound and the undo transcript executed. Because the transcript Buffer is not truncated after each command, it accumulates successive changes. A sequence of undo commands can therefore back up the file arbitrarily, which is more helpful than the more commonly implemented self-inverse form of undo. (Sam provides no way to undo an undo, but if it were desired, it would be easy to provide by re-interpreting the `do' transcript.) Each mark in the transcript contains a sequence number and the offset into the transcript of the previous mark, to aid in unwinding the transcript. Marks also contain the value of dot and the modified bit so these can be restored easily. Undoing multiple files is easy; it merely demands undoing all files whose latest change has the same sequence number as the current file.

Another benefit of having a transcript is that errors encountered in the middle of a complicated command need not leave the files in an intermediate state. By rewinding the transcript to the mark beginning the command, the partial command can be trivially undone.

When the update algorithm was first implemented, it was unacceptably slow, so a cache was added to coalesce nearby changes, replacing multiple small changes by a single larger one. This reduced the number of insertions into the transaction Buffer, and made a dramatic improvement in performance, but made it impossible to handle changes in non-monotonic order in the file; the caching method only works if changes don't overlap. Before the cache was added, the transaction could in principle be sorted if the changes were out of order, although this was never done. The current status is therefore acceptable performance with a minor restriction on global changes, which is sometimes, but rarely, an annoyance.

The update algorithm obviously paws the data more than simpler algorithms, but it is not prohibitively expensive; the caches help. (The principle of avoiding copying the data is still honored here, although not as piously: the data is moved from contents' cache to the transcript's all at once and through only one internal buffer.) Performance figures confirm the efficiency. To read from a dead start a hundred kilobyte file on a VAX-11/750 takes 1.4 seconds of user time, 2.5 seconds of system time, and 5 seconds of real time. Reading the same file in ed takes 6.0 seconds of user time, 1.7 seconds of system time, and 8 seconds of real time. Sam uses about half the CPU time. A more interesting example is the one stated above: inserting a character between every pair of characters in the file. The sam command is

,y/@/ a/x/

Communications

There are several ways sam's host and terminal parts may be connected. The first and simplest is to forgo the terminal part and use the host part's command language to edit text on an ordinary terminal. This mode is invoked by starting sam with the -d option. With no options, sam runs separate host and terminal programs, communicating with a message protocol over the physical connection that joins them. Typically, the connection is an RS-232 link between a Blit (the prototypical display for sam) and a host running the Ninth Edition of the UNIX operating system.⁸ (This is the version of the system used in the Computing Sciences Research Center at AT&T Bell Laboratories [now Lucent Technologies, Bell Labs], where I work. Its relevant aspects are discussed in the Blit paper.¹) The implementation of sam for the SUN computer runs both processes on the same machine and connects them by a pipe.

The low bandwidth of an RS-232 link necessitated the split between the two programs. The division is a mixed blessing: a program in two parts is much harder to write and to debug than a self-contained one, but the split makes several unusual configurations possible. The terminal may be physically separated from the host, allowing the conveniences of a mouse and bitmap display to be taken home while leaving the files at work. It is also possible to run the host part on a remote machine:

sam -r host

The host and terminal exchange messages asynchronously (rather than, say, as remote procedure calls) but there is no error detection or correction because, whatever the configuration, the connection is reliable. Because the terminal handles mundane interaction tasks such as popping up menus and interpreting the responses, the messages are about data, not actions. For example, the host knows nothing about what is displayed on the screen, and when the user types a character, the message sent to the host says ``insert a one-byte string at location 123 in file 7,'' not ``a character was typed at the current position in the current file.'' In other words, the messages look very much like the transaction records in the transcripts.

Either the host or terminal part of sam may initiate a change to a file. The command language operates on the host, while typing and some mouse operations are executed directly in the terminal to optimize response. Changes initiated by the host program must be transmitted to the terminal, and vice versa. (A token is exchanged to determine which end is in control, which means that characters typed while a time-consuming command runs must be buffered and do not appear until the command is complete.) To maintain consistent information, the host and terminal track changes through a per-file data structure that records what portions of the file the terminal has received. The data structure, called a Rasp (a weak pun: it's a file with holes) is held and updated by both the host and terminal. A Rasp is a list of Strings holding those parts of the file known to the terminal, separated by counts of the number of bytes in the interstices. Of course, the host doesn't keep a separate copy of the data (it only needs the lengths of the various pieces), but the structure is the same on both ends.

The Rasp in the terminal doubles as a cache. Since the terminal keeps the text for portions of the file it has displayed, it need not request data from the host when revisiting old parts of the file or redrawing obscured windows, which speeds things up considerably over low-speed links.

It's trivial for the terminal to maintain its Rasp, because all changes made on the terminal apply to parts of the file already loaded there. Changes made by the host are compared against the Rasp during the update sequence after each command. Small changes to pieces of the file loaded in the terminal are sent in their entirety. Larger changes, and changes that fall entirely in the holes, are transmitted as messages without literal data: only the lengths of the deleted and inserted strings are transmitted. When a command is completed, the terminal examines its visible windows to see if any holes in their Rasps intersect the visible portion of the file. It then requests the missing data from the host, along with up to 512 bytes of surrounding data, to minimize the number of messages when visiting a new portion of the file. This technique provides a kind of two-level lazy evaluation for the terminal. The first level sends a minimum of information about parts of the file not being edited interactively; the second level waits until a change is displayed before transmitting the new data. Of course, performance is also helped by having the terminal respond immediately to typing and simple mouse requests. Except for small changes to active pieces of the file, which are transmitted to the terminal without negotiation, the terminal is wholly responsible for deciding what is displayed; the host uses the Rasp only to tell the terminal what might be relevant.

When a change is initiated by the host, the messages to the terminal describing the change are generated by the routine that applies the transcript of the changes to the contents of the File. Since changes are undone by the same update routine, undoing requires no extra code in the communications; the usual messages describing changes to the file are sufficient to back up the screen image.

The Rasp is a particularly good example of the way caches are used in sam. First, it facilitates access to the active portion of the text by placing the busy text in main memory. In so doing, it provides efficient access to a large data structure that does not fit in memory. Since the form of data is to be imposed by the user, not by the program, and because characters will frequently be scanned sequentially, files are stored as flat objects. Caches help keep performance good and linear when working with such data.

Second, the Rasp and several of the other caches have some read-ahead; that is, the cache is loaded with more information than is needed for the job immediately at hand. When manipulating linear structures, the accesses are usually sequential, and read-ahead can significantly reduce the average time to access the next element of the object. Sequential access is a common mode for people as well as programs; consider scrolling through a document while looking for something.

Finally, like any good data structure, the cache guides the algorithm, or at least the implementation. The Rasp was actually invented to control the communications between the host and terminal parts, but I realized very early that it was also a form of cache. Other caches were more explicitly intended to serve a double purpose: for example, the caches in Files that coalesce updates not only reduce traffic to the transcript and contents Buffers, they also clump screen updates so that complicated changes to the screen are achieved in just a few messages to the terminal. This saved me considerable work: I did not need to write special code to optimize the message traffic to the terminal. Caches pay off in surprising ways. Also, they tend to be independent, so their performance improvements are multiplicative.

Data structures in the terminal

Sam typically has far more windows than does mux, the window system within which its Blit implementation runs. Mux has a fairly small number of asynchronously updated windows; sam needs a large number of synchronously updated windows that are usually static and often fully obscured. The different tradeoffs guided sam away from the memory-intensive implementation of windows, called Layers,¹⁷ used in mux. Rather than depending on a complete bitmap image of the display for each window, sam regenerates the image from its in-memory text (stored in the Rasp) when necessary, although it will use such an image if it is available. Like Layers, though, sam uses the screen bitmap as active storage in which to update the image using bitblt.^18,19 The resulting organization, pictured in Figure 6, has a global array of windows, called Flayers, each of which holds an image of a piece of text held in a data structure called a Frame, which in turn represents a rectangular window full of text displayed in some Bitmap. Each Flayer appears in a global list that orders them all front-to-back on the display, and simultaneously as an element of a per-file array that holds all the open windows for that file. The complement in the terminal of the File on the host is called a Text; each connects its Flayers to the associated Rasp.

The Bitmap for a Frame contains the image of the text. For a fully visible window, the Bitmap will be the screen (or at least the Layer in which sam is being run), while for partially obscured windows the Bitmap will be off-screen. If the window is fully obscured, the Bitmap will be null.

The Bitmap is a kind of cache. When making changes to the display, most of the original image will look the same in the final image, and the update algorithms exploit this. The Frame software updates the image in the Bitmap incrementally; the Bitmap is not just an image, it is a data structure.^18,19 The job of the software that updates the display is therefore to use as much as possible of the existing image (converting the text from ASCII characters to pixels is expensive) in a sort of two-dimensional string insertion algorithm. The details of this process are described in the next section.

The Frame software has no code to support overlapping windows; its job is to keep a single Bitmap up to date. It falls to the Flayer software to multiplex the various Bitmaps onto the screen. The problem of maintaining overlapping Flayers is easier than for Layers¹⁷ because changes are made synchronously and because the contents of the window can be reconstructed from the data stored in the Frame; the Layers software makes no such assumptions. In sam, the window being changed is almost always fully visible, because the current window is always fully visible, by construction. However, when multi-file changes are being made, or when more than one window is open on a file, it may be necessary to update partially obscured windows.

There are three cases: the window is fully visible, invisible (fully obscured), or partially visible. If fully visible, the Bitmap is part of the screen, so when the Flayer update routine calls the Frame update routine, the screen will be updated directly. If the window is invisible, there is no associated Bitmap, and all that is necessary is to update the Frame data structure, not the image. If the window is partially visible, the Frame routine is called to update the image in the off-screen Bitmap, which may require regenerating it from the text of the window. The Flayer code then clips this Bitmap against the Bitmaps of all Frames in front of the Frame being modified, and the remainder is copied to the display.

This is much faster than recreating the image off-screen for every change, or clipping all the changes made to the image during its update. Unfortunately, these caches can also consume prohibitive amounts of memory, so they are freed fairly liberally ­ after every change to the front-to-back order of the Flayers. The result is that the off-screen Bitmaps exist only while multi-window changes are occurring, which is the only time the performance improvement they provide is needed. Also, the user interface causes fully-obscured windows to be the easiest to make ­ creating a canonically sized and placed window requires only a button click ­ which reduces the need for caching still further.

Screen update

Two special kinds of Boxes contain a single character: either a newline or a tab. Newlines and tabs are white space. A newline Box always extends to the right edge of the window, forcing the following Box to the next line. The width of a tab depends on where it is located: it forces the next Box to begin at a tab location. Tabs also have a minimum width equivalent to a blank (blanks are drawn by string and are not treated specially); newlines have a minimum width of zero.

The update algorithms always use the Bitmap image of the text (either the display or cache Bitmap); they never examine the characters within a Box except when the Box needs to be split in two. Before a change, the window consists of a tiling of Boxes; after the change the window is tiled differently. The update algorithms rearrange the tiles in place, without backup storage. The algorithms are not strictly optimal ­ for example, they can clear a pixel that is later going to be written upon ­ but they never move a tile that doesn't need to be moved, and they move each tile at most once. Frinsert on a Blit can absorb over a thousand characters a second if the strings being inserted are a few tens of characters long.

Consider frdelete. Its job is to delete a substring from a Frame and restore the image of the Frame. The image of a substring has a peculiar shape (see Figure 2) comprising possibly a partial line, zero or more full lines, and possibly a final partial line. For reference, call this the Z-shape. Frdelete begins by splitting, if necessary, the Boxes containing the ends of the substring so the substring begins and ends on Box boundaries. Because the substring is being deleted, its image is not needed, so the Z-shape is then cleared. Then, tiles (that is, the images of Boxes) are copied, using bitblt, from immediately after the Z-shape to the beginning of the Z-shape, resulting in a new Z-shape. (Boxes whose contents would span two lines in the new position must first be split.)

Copying the remainder of the Frame tile by tile this way will clearly accomplish the deletion but eventually, typically when the copying algorithm encounters a tab or newline, the old and new x coordinates of the tile to be copied are the same. This correspondence implies that the Z-shape has its beginning and ending edges aligned vertically, and a sequence of at most two bitblts can be used to copy the remaining tiles. The last step is to clear out the resulting empty space at the bottom of the window; the number of lines to be cleared is the number of complete lines in the Z-shape closed by the final bitblts. The final step is to merge horizontally adjacent Boxes of plain text. The complete source to frdelete is less than 100 lines of C.

frinsert is more complicated because it must do four passes: one to construct the Box list for the inserted string, one to reconnoitre, one to copy (in opposite order to frdelete) the Boxes to make the hole for the new text, and finally one to copy the new text into place. Overall, though, frinsert has a similar flavor to frdelete, and needn't be described further. Frinsert and its subsidiary routines comprise 211 lines of C.

The terminal source code is 3024 lines of C, and the host source is 5797 lines.

Discussion

History

Jim was written primarily as a vehicle for experimenting with a mouse-based interface to text, and the experiment was successful. Jim had some spin-offs: mux, the second window system for the Blit, is essentially a multiplexed version of the terminal part of jim; and the debugger pi's user interface²⁰ was closely modeled on jim's. But after a couple of years, jim had become difficult to maintain and limiting to use, and its replacement was overdue.

I began the design of sam by asking jim customers what they wanted. This was probably a mistake; the answers were essentially a list of features to be found in other editors, which did not provide any of the guiding principles I was seeking. For instance, one common request was for a ``global substitute,'' but no one suggested how to provide it within a cut-and-paste editor. I was looking for a scheme that would support such specialized features comfortably in the context of some general command language. Ideas were not forthcoming, though, particularly given my insistence on removing all limits on file sizes, line lengths and so on. Even worse, I recognized that, since the mouse could easily indicate a region of the screen that was not an integral number of lines, the command language would best forget about newlines altogether, and that meant the command language had to treat the file as a single string, not an array of lines.

Eventually, I decided that thinking was not getting me very far and it was time to try building. I knew that the terminal part could be built easily ­ that part of jim behaved acceptably well ­ and that most of the hard work was going to be in the host part: the file interface, command interpreter and so on. Moreover, I had some ideas about how the architecture of jim could be improved without destroying its basic structure, which I liked in principle but which hadn't worked out as well as I had hoped. So I began by designing the file data structure, starting with the way jim worked ­ comparable to a single structure merging Disc and Buffer, which I split to make the cache more general ­ and thinking about how global substitute could be implemented. The answer was clearly that it had to be done in two passes, and the transcript-oriented implementation fell out naturally.

Sam was written bottom-up, starting from the data structures and algorithms for manipulating text, through the command language and up to the code for maintaining the display. In retrospect, it turned out well, but this implementation method is not recommended in general. There were several times when I had a large body of interesting code assembled and no clue how to proceed with it. The command language, in particular, took almost a year to figure out, but can be implemented (given what was there at the beginning of that year) in a day or two. Similarly, inventing the Rasp data structure delayed the connection of the host and terminal pieces by another few months. Sam took about two years to write, although only about four months were spent actually working on it.

Part of the design process was unusual: the subset of the protocol that maintains the Rasp was simulated, debugged and verified by an automatic protocol analyzer,²¹ and was bug-free from the start. The rest of the protocol, concerned mostly with keeping menus up to date, was unfortunately too unwieldy for such analysis, and was debugged by more traditional methods, primarily by logging in a file all messages in and out of the host.

Reflections

NOTE: ¿The people who criticize ed as an interactive program often forget that it and its close relative sed⁷ still thrive as programmable editors. The strength of these programs is independent of their convenience for interactive editing.

To be honest, ed still gets occasional use, but usually when something quick needs to be done and the overhead of downloading the terminal part of sam isn't worth the trouble. Also, as a `line' editor, sam -d is a bit odd; when using a good old ASCII terminal, it's comforting to have a true line editor. But it is fair to say that sam's command language has displaced ed's for most of the complicated editing that has kept line editors (that is, command-driven editors) with us.

Sam's command language is even fancier than ed's, and most sam customers don't come near to using all its capabilities. Does it need to be so sophisticated? I think the answer is yes, for two reasons.

First, the model for sam's command language is really relatively simple, and certainly simpler than that of ed. For instance, there is only one kind of textual loop in sam ­ the x command ­ while ed has three (the g command, the global flag on substitutions, and the implicit loop over lines in multi-line substitutions). Also, ed's substitute command is necessary to make changes within lines, but in sam the s command is more of a familiar convenience than a necessity; c and t can do all the work.

Second, given a community that expects an editor to be about as powerful as ed, it's hard to see how sam could really be much simpler and still satisfy that expectation. People want to do ``global substitutes,'' and most are content to have the recipe for that and a few other fancy changes. The sophistication of the command language is really just a veneer over a design that makes it possible to do global substitutes in a screen editor. Some people will always want something more, however, and it's gratifying to be able to provide it. The real power of sam's command language comes from composability of the operators, which is by nature orthogonal to the underlying model. In other words, sam is not itself complex, but it makes complex things possible. If you don't want to do anything complex, you can ignore the complexity altogether, and many people do so.

Sometimes I am asked the opposite question: why didn't I just make sam a real programmable editor, with macros and variables and so on? The main reason is a matter of taste: I like the editor to be the same every time I use it. There is one technical reason, though: programmability in editors is largely a workaround for insufficient interactivity. Programmable editors are used to make particular, usually short-term, things easy to do, such as by providing shorthands for common actions. If things are generally easy to do in the first place, shorthands are not as helpful. Sam makes common editing operations very easy, and the solutions to complex editing problems seem commensurate with the problems themselves. Also, the ability to edit the sam window makes it easy to repeat commands ­ it only takes a mouse button click to execute a command again.

Pros and cons

Another of sam's good things is its undo capability. I had never before used an editor with a true undo, but I would never go back now. Undo must be done well, but if it is, it can be relied on. For example, it's safe to experiment if you're not sure how to write some intricate command, because if you make a mistake, it can be fixed simply and reliably. I learned two things about undo from writing sam: first, it's easy to provide if you design it in from the beginning, and second, it's necessary, particularly if the system has some subtle properties that may be unfamiliar or error-prone for users.

Sam's lack of internal limits and sizes is a virtue. Because it avoids all fixed-size tables and data structures, sam is able to make global changes to files that some of our other tools cannot even read. Moreover, the design keeps the performance linear when doing such operations, although I must admit sam does get slow when editing a huge file.

Now, the problems. Externally, the most obvious is that it is poorly integrated into the surrounding window system. By design, the user interface in sam feels almost identical to that of mux, but a thick wall separates text in sam from the programs running in mux. For instance, the `snarf buffer' in sam must be maintained separately from that in mux. This is regrettable, but probably necessary given the unusual configuration of the system, with a programmable terminal on the far end of an RS-232 link.

Sam is reliable; otherwise, people wouldn't use it. But it was written over such a long time, and has so many new (to me) ideas in it, that I would like to see it done over again to clean up the code and remove many of the lingering problems in the implementation. The worst part is in the interconnection of the host and terminal parts, which might even be able to go away in a redesign for a more conventional window system. The program must be split in two to use the terminal effectively, but the low bandwidth of the connection forces the separation to occur in an inconvenient part of the design if performance is to be acceptable. A simple remote procedure call protocol driven by the host, emitting only graphics commands, would be easy to write but wouldn't have nearly the necessary responsiveness. On the other hand, if the terminal were in control and requested much simpler file services from the host, regular expression searches would require that the terminal read the entire file over its RS-232 link, which would be unreasonably slow. A compromise in which either end can take control is necessary. In retrospect, the communications protocol should have been designed and verified formally, although I do not know of any tool that can adequately relate the protocol to its implementation.

Not all of sam's users are comfortable with its command language, and few are adept. Some (venerable) people use a sort of ed `` subset'' of sam's command language, and even ask why sam's command language is not exactly ed's. (The reason, of course, is that sam's model for text does not include newlines, which are central to ed. Making the text an array of newlines to the command language would be too much of a break from the seamless model provided by the mouse. Some editors, such as vi, are willing to make this break, though.) The difficulty is that sam's syntax is so close to ed's that people believe it should be the same. I thought, with some justification in hindsight, that making sam similar to ed would make it easier to learn and to accept. But I may have overstepped and raised the users' expectations too much. It's hard to decide which way to resolve this problem.

Finally, there is a tradeoff in sam that was decided by the environment in which it runs: sam is a multi-file editor, although in a different system there might instead be multiple single-file editors. The decision was made primarily because starting a new program in a Blit is time-consuming. If the choice could be made freely, however, I would still choose the multi-file architecture, because it allows groups of files to be handled as a unit; the usefulness of the multi-file commands is incontrovertible. It is delightful to have the source to an entire program available at your fingertips.

Acknowledgements

REFERENCES

1.

R. Pike, `The Blit: a multiplexed graphics terminal,' AT&T Bell Labs. Tech. J., 63, (8), 1607-1631 (1984).

2.

L. Johnson, MacWrite, Apple Computer Inc., Cupertino, Calif. 1983.

3.

B. Lampson, `Bravo Manual,' in Alto User's Handbook, pp. 31-62, Xerox Palo Alto Research Center, Palo Alto, Calif. 1979.

4.

W. Teitelman, `A tour through Cedar,' IEEE Software, 1 (2), 44-73 (1984).

5.

J. Gutknecht, `Concepts of the text editor Lara,' Comm. ACM, 28, (9), 942-960 (1985).

6.

Bell Telephone Laboratories, UNIX Programmer's Manual, Holt, Rinehart and Winston, New York 1983.

7.

B. W. Kernighan and R. Pike, The Unix Programming Environment, Prentice-Hall, Englewood Cliffs, New Jersey 1984.

8.

Unix Time-Sharing System Programmer's Manual, Research Version, Ninth Edition, Volume 1, AT&T Bell Laboratories, Murray Hill, New Jersey 1986.

9.

Unix Time-Sharing System Programmer's Manual, 4.1 Berkeley Software Distribution, Volumes 1 and 2C, University of California, Berkeley, Calif. 1981.

10.

R. Pike, `Structural Regular Expressions,' Proc. EUUG Spring Conf., Helsinki 1987, Eur. Unix User's Group, Buntingford, Herts, UK 1987.

11.

A. Goldberg, Smalltalk-80 ¿ The Interactive Programming Environment, Addison-Wesley, Reading, Mass. 1984.

12.

K. Thompson, `Regular expression search algorithm,' Comm. ACM, 11, (6), 419-422 (1968).

13.

A. V. Aho, J. E. Hopcroft and J. D. Ullman, The Design and Analysis of Computer Algorithms, Addison-Wesley, Reading, Mass. 1974.

14.

B. W. Kernighan and D. M. Ritchie, The C Programming Language, Prentice-Hall, Englewood Cliffs, New Jersey 1978.

15.

W. M. Waite, `The cost of lexical analysis,' Softw. Pract. Exp., 16, (5), 473-488 (1986).

16.

C. W. Fraser, `A generalized text editor,' Comm. ACM, 23, (3), 154-158 (1980).

17.

R. Pike, `Graphics in overlapping bitmap layers,' ACM Trans. on Graph., 2, (2) 135-160 (1983).

18.

L. J. Guibas and J. Stolfi, `A language for bitmap manipulation,' ACM Trans. on Graph., 1, (3), 191-214 (1982).

19.

R. Pike, B. Locanthi and J. Reiser, `Hardware/software trade-offs for bitmap graphics on the Blit,' Softw. Pract. Exp., 15, (2), 131-151 (1985).

20.

T. A. Cargill, `The feel of Pi,' Winter USENIX Conference Proceedings, Denver 1986, 62-71, USENIX Assoc., El Cerrito, CA.

21.

G. J. Holzmann, `Tracing protocols,' AT&T Tech. J., 64, (10), 2413-2434 (1985).