Copyright (C) 1989, Digital Equipment Corporation           
 All rights reserved.                                        
 See the file COPYRIGHT for a full description.              

Last modified on Fri Jun 18 17:55:24 PDT 1993 by wobber modified on Tue Jun 15 09:59:17 1993 by gnelson modified on Fri May 21 09:50:57 PDT 1993 by swart modified on Wed Nov 6 10:44:49 PST 1991 by kalsow modified on Thu Nov 2 21:55:19 1989 by muller

There is no end to the number of useful classes of readers and writers. Here are a few examples from SRC's standard libraries:

- Tee writers, which write copies of their stream to each of two other writers. The name comes from the Unix program ``tee'', which performs a similar function in the realm of pipes. The most common use is to write to a terminal and to a logfile at the same time.

- Various ways to make new readers from old readers, for example, by concatenation, subsequencing, duplication, filtering, and by making a seekable reader from an unseekable one.

- Split writers, which are intended for use by applications that use parallel threads writing to a single writer. Split writers keep the output from each thread separate; this creates the illusion that one thread writes all of its output before the next thread starts writing its output.

- Local pipes, which consist of a reader and a writer hooked together in such a way that the reader will return precisely the characters that have been written to the writer.

- Formatted writers, in which the client can mark the start and end of logical objects and specify desirable places to break the objects into lines. Formatted writers are basic tools for building pretty printers.

It is beyond the scope of this paper to describe these classes in detail. Instead we will describe the interfaces that allow you to define new classes.

The basic idea is that readers and writers are objects whose method suites are determined by their class. In the most naive version of this idea, a writer class's putChar method would determine the effect of Wr.PutChar for writers of the class:

PutChar(wr, ch) = wr.putChar(ch)

The putChar method for a terminal writer would send characters to the terminal; while the method for a disk file writer would send characters to the disk, etc.

There are two reasons for rejecting this naive version. The first reason is that it is inefficient to call a method for every PutChar. The second and more important reason is that most writers are buffered, and it is undesirable to force every client to reimplement buffering.

We implement PutChar and GetChar by class-independent code operating on a buffer; class-dependent code is invoked only when the buffer fills up (in the case of a writer) or empties (in the case of a reader).

In this section we define the WrClass interface, which reveals the buffer structure in a writer object. New writer classes are created by importing WrClass (to gain access to the buffer and the methods) and then defining a subclass of Wr.T whose methods provide the new class's behavior. The private fields that are needed by the class-independent code but are irrelevant to the buffer structure are lumped together into the opaque type Private.

FROM Thread IMPORT Alerted;

  Private <: ROOT;

           buff       : REF ARRAY OF CHAR;
           st         : CARDINAL;           (* index into buff *)
           lo, hi, cur: CARDINAL            := 0; (* indexes into c(wr) *)
           closed := TRUE;       (* Subtype's init method should set this
                                    to FALSE *)
           seekable, buffered: BOOLEAN
           seek      (n: CARDINAL) RAISES {Failure, Alerted};
           putString (READONLY a: ARRAY OF CHAR) RAISES {Failure, Alerted}
                                                         := PutStringDefault;
           length (): CARDINAL RAISES {Failure, Alerted} := LengthDefault;
           flush  () RAISES {Failure, Alerted}           := FlushDefault;
           close  () RAISES {Failure, Alerted}           := CloseDefault
Let wr be a writer, which abstractly is given by c(wr), target(wr), cur(wr), closed(wr), seekable(wr), buffered(wr). The actual representation of wr is an object of type Wr.T. The wr.cur, wr.closed, wr.seekable, and wr.buffered fields in the object represent the corresponding abstract attributes of wr. The wr.buff,, wr.lo, and wr.hi fields in the object represent a buffer containing the unflushed part of c(wr). The target of the writer is represented in some class-specific way, which is not specified by this interface.

More precisely, we say that the state of the writer object wr is valid if the following conditions V1 through V5 hold:

V1. the cur field and the booleans are correct: wr.cur = cur(wr) AND wr.closed = closed(wr) AND wr.buffered = buffered(wr) AND wr.seekable = seekable(wr)

V2. the indexes of any unflushed characters are in the range [lo..cur-1]. That is, for all i not in [wr.lo..wr.cur-1],

c(wr)[i] = target(wr)[i]

V3. the (possibly) unflushed characters are stored in buff starting with buff[st]. That is, for all i in [wr.lo..wr.cur-1],

c(wr)[i] = wr.buff[ + i - wr.lo]

(Usually st is zero. Non-zero values may be useful to satisfy buffer alignment constraints.)

V4. the current position is either contained in the buffer, or just past the buffer:

wr.lo <= wr.cur <= wr.hi

It is possible that buff = NIL in a valid state, since the range of i's in V3 can be empty; for example, in the case lo = hi = cur.

V5. if closed(wr) then wr.buff = NIL AND wr.lo = wr.hi

We say that the state is ready if the buffer contains the current position; that is, if

NOT wr.closed AND wr.buff # NIL AND wr.lo <= cur(wr) < wr.hi

If the state is ready, then Wr.PutChar can be implemented by storing into the buffer. The class-independent code does exactly this, until the buffer is full, at which point it calls a class method to consume the buffer and provide a new one. Together V4 and V5 imply that if wr.cur # wr.hi then wr.buff # NIL and NOT wr.closed. Therefore a valid writer is ready if wr.cur # wr.hi.

In general, the class-independent code modifies cur and buff[i] for i in the range [], but not the buff reference itself, st, lo, or hi (except that Wr.Close modifies wr.lo and wr.cur and sets wr.buff to NIL in order to maintain invariant V5). The class-independent code locks the writer before calling any methods; therefore, no two method activations initialized by the class-independent code will be concurrent. A method must not apply operations from the Wr interface to the writer itself, or deadlock will result.

Here are the specifications for the methods:

The method call treats n as a position to seek to, and moves the buffer to contain this position. More precisely:

- Given a valid state, must produce a valid ready state in which wr.cur = MIN(n, len(wr)) and c(wr) is unchanged.

An important special case is when n = wr.cur = wr.hi; that is, when the buffer has overflowed and the effect of the seek is simply to advance from the last character of a buffer to the first character of a new buffer. Every writer class (seekable or not) must provide a seek method that supports this special case. The method must support the general case only if the writer is seekable.

The putString method is used to implement Wr.PutString and is called with the writer lock held. Efficient implementations override this method to avoid unnecessary copying by transferring directly into the writer target, bypassing the writer buffer. The default implementation is correct for any class, but always copies through the writer buffer.

The flush method updates the underlying target of the writer. That is: Given a valid state, wr.flush() must produce a valid state in which c(wr) and cur(wr) are unchanged and target(wr) = c(wr). If a writer is unbuffered, the class-independent code will call the flush method after every modification to the buffer.

The close method releases all resources associated with a writer. That is: Given a valid state in which target(wr) = c(wr), the call wr.close() must release all resources associated with wr. The exact meaning is class-specific V. Wr.Close sets the buff field to NIL, so the method need not do this. Validity is not required when the method returns, since after it returns, the class-independent code will set the closed bit in the writer, which makes the rest of the state irrelevant, even if it is invalid.

The length method returns the length of the writer. That is: Given a valid state, wr.length() must return len(wr), leaving a valid state in which c(wr) and cur(wr) are unchanged.

Finally, here are two procedures that are essential for implementing class-specific operations.

The writer wr must be unlocked; lock it and make its state valid.

PROCEDURE Unlock(wr: Wr.T);
The writer wr must be locked and valid; unlock it and restore the private invariant of the writer implementation.
 A class-specific operation on a writer wr should use the following

Lock(wr); TRY ... FINALLY Unlock(wr) END

The methods don't have to do this, since the class-independent code automatically locks and unlocks the writer around method calls. The next section provides examples of the use of Lock and Unlock.

The last declarations in the interface are for the default methods:

    RAISES {Failure, Alerted};

PROCEDURE LengthDefault(wr: Wr.T): CARDINAL RAISES {};

PROCEDURE CloseDefault(wr: Wr.T) RAISES {};

PROCEDURE FlushDefault(wr: Wr.T) RAISES {};
LengthDefault returns wr.cur, while CloseDefault and FlushDefault are no-ops.

END WrClass.