Tutorial: Let's build a Compiler! - Part IV: Interpreters
Jon A. Lambert
Jon.A.Lambert at ix.netcom.com
Fri Feb 27 18:27:28 CET 1998
LET'S BUILD A COMPILER!
By
Jack W. Crenshaw, Ph.D.
24 July 1988
Part IV: INTERPRETERS
*****************************************************************
* *
* COPYRIGHT NOTICE *
* *
* Copyright (C) 1988 Jack W. Crenshaw. All rights reserved. *
* *
*****************************************************************
INTRODUCTION
In the first three installments of this series, we've looked at
parsing and compiling math expressions, and worked our way grad-
ually and methodically from dealing with very simple one-term,
one-character "expressions" up through more general ones, finally
arriving at a very complete parser that could parse and translate
complete assignment statements, with multi-character tokens,
embedded white space, and function calls. This time, I'm going
to walk you through the process one more time, only with the goal
of interpreting rather than compiling object code.
Since this is a series on compilers, why should we bother with
interpreters? Simply because I want you to see how the nature of
the parser changes as we change the goals. I also want to unify
the concepts of the two types of translators, so that you can see
not only the differences, but also the similarities.
Consider the assignment statement
x = 2 * y + 3
In a compiler, we want the target CPU to execute this assignment
at EXECUTION time. The translator itself doesn't do any arith-
metic ... it only issues the object code that will cause the CPU
to do it when the code is executed. For the example above, the
compiler would issue code to compute the expression and store the
results in variable x.
For an interpreter, on the other hand, no object code is gen-
erated. Instead, the arithmetic is computed immediately, as the
parsing is going on. For the example, by the time parsing of the
statement is complete, x will have a new value.
The approach we've been taking in this whole series is called
"syntax-driven translation." As you are aware by now, the struc-
ture of the parser is very closely tied to the syntax of the
productions we parse. We have built Pascal procedures that rec-
ognize every language construct. Associated with each of these
constructs (and procedures) is a corresponding "action," which
does whatever makes sense to do once a construct has been
recognized. In our compiler so far, every action involves
emitting object code, to be executed later at execution time. In
an interpreter, every action involves something to be done im-
mediately.
What I'd like you to see here is that the layout ... the struc-
ture ... of the parser doesn't change. It's only the actions
that change. So if you can write an interpreter for a given
language, you can also write a compiler, and vice versa. Yet, as
you will see, there ARE differences, and significant ones.
Because the actions are different, the procedures that do the
recognizing end up being written differently. Specifically, in
the interpreter the recognizing procedures end up being coded as
FUNCTIONS that return numeric values to their callers. None of
the parsing routines for our compiler did that.
Our compiler, in fact, is what we might call a "pure" compiler.
Each time a construct is recognized, the object code is emitted
IMMEDIATELY. (That's one reason the code is not very efficient.)
The interpreter we'll be building here is a pure interpreter, in
the sense that there is no translation, such as "tokenizing,"
performed on the source code. These represent the two extremes
of translation. In the real world, translators are rarely so
pure, but tend to have bits of each technique.
I can think of several examples. I've already mentioned one:
most interpreters, such as Microsoft BASIC, for example, trans-
late the source code (tokenize it) into an intermediate form so
that it'll be easier to parse real time.
Another example is an assembler. The purpose of an assembler, of
course, is to produce object code, and it normally does that on a
one-to-one basis: one object instruction per line of source code.
But almost every assembler also permits expressions as arguments.
In this case, the expressions are always constant expressions,
and so the assembler isn't supposed to issue object code for
them. Rather, it "interprets" the expressions and computes the
corresponding constant result, which is what it actually emits as
object code.
As a matter of fact, we could use a bit of that ourselves. The
translator we built in the previous installment will dutifully
spit out object code for complicated expressions, even though
every term in the expression is a constant. In that case it
would be far better if the translator behaved a bit more like an
interpreter, and just computed the equivalent constant result.
There is a concept in compiler theory called "lazy" translation.
The idea is that you typically don't just emit code at every
action. In fact, at the extreme you don't emit anything at all,
until you absolutely have to. To accomplish this, the actions
associated with the parsing routines typically don't just emit
code. Sometimes they do, but often they simply return in-
formation back to the caller. Armed with such information, the
caller can then make a better choice of what to do.
For example, given the statement
x = x + 3 - 2 - (5 - 4) ,
our compiler will dutifully spit out a stream of 18 instructions
to load each parameter into registers, perform the arithmetic,
and store the result. A lazier evaluation would recognize that
the arithmetic involving constants can be evaluated at compile
time, and would reduce the expression to
x = x + 0 .
An even lazier evaluation would then be smart enough to figure
out that this is equivalent to
x = x ,
which calls for no action at all. We could reduce 18 in-
structions to zero!
Note that there is no chance of optimizing this way in our trans-
lator as it stands, because every action takes place immediately.
Lazy expression evaluation can produce significantly better
object code than we have been able to so far. I warn you,
though: it complicates the parser code considerably, because each
routine now has to make decisions as to whether to emit object
code or not. Lazy evaluation is certainly not named that because
it's easier on the compiler writer!
Since we're operating mainly on the KISS principle here, I won't
go into much more depth on this subject. I just want you to be
aware that you can get some code optimization by combining the
techniques of compiling and interpreting. In particular, you
should know that the parsing routines in a smarter translator
will generally return things to their caller, and sometimes
expect things as well. That's the main reason for going over
interpretation in this installment.
THE INTERPRETER
OK, now that you know WHY we're going into all this, let's do it.
Just to give you practice, we're going to start over with a bare
cradle and build up the translator all over again. This time, of
course, we can go a bit faster.
Since we're now going to do arithmetic, the first thing we need
to do is to change function GetNum, which up till now has always
returned a character (or string). Now, it's better for it to
return an integer. MAKE A COPY of the cradle (for goodness's
sake, don't change the version in Cradle itself!!) and modify
GetNum as follows:
{--------------------------------------------------------------}
{ Get a Number }
function GetNum: integer;
begin
if not IsDigit(Look) then Expected('Integer');
GetNum := Ord(Look) - Ord('0');
GetChar;
end;
{--------------------------------------------------------------}
Now, write the following version of Expression:
{---------------------------------------------------------------}
{ Parse and Translate an Expression }
function Expression: integer;
begin
Expression := GetNum;
end;
{--------------------------------------------------------------}
Finally, insert the statement
Writeln(Expression);
at the end of the main program. Now compile and test.
All this program does is to "parse" and translate a single
integer "expression." As always, you should make sure that it
does that with the digits 0..9, and gives an error message for
anything else. Shouldn't take you very long!
OK, now let's extend this to include addops. Change Expression
to read:
{---------------------------------------------------------------}
{ Parse and Translate an Expression }
function Expression: integer;
var Value: integer;
begin
if IsAddop(Look) then
Value := 0
else
Value := GetNum;
while IsAddop(Look) do begin
case Look of
'+': begin
Match('+');
Value := Value + GetNum;
end;
'-': begin
Match('-');
Value := Value - GetNum;
end;
end;
end;
Expression := Value;
end;
{--------------------------------------------------------------}
The structure of Expression, of course, parallels what we did
before, so we shouldn't have too much trouble debugging it.
There's been a SIGNIFICANT development, though, hasn't there?
Procedures Add and Subtract went away! The reason is that the
action to be taken requires BOTH arguments of the operation. I
could have chosen to retain the procedures and pass into them the
value of the expression to date, which is Value. But it seemed
cleaner to me to keep Value as strictly a local variable, which
meant that the code for Add and Subtract had to be moved in line.
This result suggests that, while the structure we had developed
was nice and clean for our simple-minded translation scheme, it
probably wouldn't do for use with lazy evaluation. That's a
little tidbit we'll probably want to keep in mind for later.
OK, did the translator work? Then let's take the next step.
It's not hard to figure out what procedure Term should now look
like. Change every call to GetNum in function Expression to a
call to Term, and then enter the following form for Term:
{---------------------------------------------------------------}
{ Parse and Translate a Math Term }
function Term: integer;
var Value: integer;
begin
Value := GetNum;
while Look in ['*', '/'] do begin
case Look of
'*': begin
Match('*');
Value := Value * GetNum;
end;
'/': begin
Match('/');
Value := Value div GetNum;
end;
end;
end;
Term := Value;
end;
{--------------------------------------------------------------}
Now, try it out. Don't forget two things: first, we're dealing
with integer division, so, for example, 1/3 should come out zero.
Second, even though we can output multi-digit results, our input
is still restricted to single digits.
That seems like a silly restriction at this point, since we have
already seen how easily function GetNum can be extended. So
let's go ahead and fix it right now. The new version is
{--------------------------------------------------------------}
{ Get a Number }
function GetNum: integer;
var Value: integer;
begin
Value := 0;
if not IsDigit(Look) then Expected('Integer');
while IsDigit(Look) do begin
Value := 10 * Value + Ord(Look) - Ord('0');
GetChar;
end;
GetNum := Value;
end;
{--------------------------------------------------------------}
If you've compiled and tested this version of the interpreter,
the next step is to install function Factor, complete with pa-
renthesized expressions. We'll hold off a bit longer on the
variable names. First, change the references to GetNum, in
function Term, so that they call Factor instead. Now code the
following version of Factor:
{---------------------------------------------------------------}
{ Parse and Translate a Math Factor }
function Expression: integer; Forward;
function Factor: integer;
begin
if Look = '(' then begin
Match('(');
Factor := Expression;
Match(')');
end
else
Factor := GetNum;
end;
{---------------------------------------------------------------}
That was pretty easy, huh? We're rapidly closing in on a useful
interpreter.
A LITTLE PHILOSOPHY
Before going any further, there's something I'd like to call to
your attention. It's a concept that we've been making use of in
all these sessions, but I haven't explicitly mentioned it up till
now. I think it's time, because it's a concept so useful, and so
powerful, that it makes all the difference between a parser
that's trivially easy, and one that's too complex to deal with.
In the early days of compiler technology, people had a terrible
time figuring out how to deal with things like operator prece-
dence ... the way that multiply and divide operators take
precedence over add and subtract, etc. I remember a colleague of
some thirty years ago, and how excited he was to find out how to
do it. The technique used involved building two stacks, upon
which you pushed each operator or operand. Associated with each
operator was a precedence level, and the rules required that you
only actually performed an operation ("reducing" the stack) if
the precedence level showing on top of the stack was correct. To
make life more interesting, an operator like ')' had different
precedence levels, depending upon whether or not it was already
on the stack. You had to give it one value before you put it on
the stack, and another to decide when to take it off. Just for
the experience, I worked all of this out for myself a few years
ago, and I can tell you that it's very tricky.
We haven't had to do anything like that. In fact, by now the
parsing of an arithmetic statement should seem like child's play.
How did we get so lucky? And where did the precedence stacks go?
A similar thing is going on in our interpreter above. You just
KNOW that in order for it to do the computation of arithmetic
statements (as opposed to the parsing of them), there have to be
numbers pushed onto a stack somewhere. But where is the stack?
Finally, in compiler textbooks, there are a number of places
where stacks and other structures are discussed. In the other
leading parsing method (LR), an explicit stack is used. In fact,
the technique is very much like the old way of doing arithmetic
expressions. Another concept is that of a parse tree. Authors
like to draw diagrams of the tokens in a statement, connected
into a tree with operators at the internal nodes. Again, where
are the trees and stacks in our technique? We haven't seen any.
The answer in all cases is that the structures are implicit, not
explicit. In any computer language, there is a stack involved
every time you call a subroutine. Whenever a subroutine is
called, the return address is pushed onto the CPU stack. At the
end of the subroutine, the address is popped back off and control
is transferred there. In a recursive language such as Pascal,
there can also be local data pushed onto the stack, and it, too,
returns when it's needed.
For example, function Expression contains a local parameter
called Value, which it fills by a call to Term. Suppose, in its
next call to Term for the second argument, that Term calls
Factor, which recursively calls Expression again. That "in-
stance" of Expression gets another value for its copy of Value.
What happens to the first Value? Answer: it's still on the
stack, and will be there again when we return from our call
sequence.
In other words, the reason things look so simple is that we've
been making maximum use of the resources of the language. The
hierarchy levels and the parse trees are there, all right, but
they're hidden within the structure of the parser, and they're
taken care of by the order with which the various procedures are
called. Now that you've seen how we do it, it's probably hard to
imagine doing it any other way. But I can tell you that it took
a lot of years for compiler writers to get that smart. The early
compilers were too complex too imagine. Funny how things get
easier with a little practice.
The reason I've brought all this up is as both a lesson and a
warning. The lesson: things can be easy when you do them right.
The warning: take a look at what you're doing. If, as you branch
out on your own, you begin to find a real need for a separate
stack or tree structure, it may be time to ask yourself if you're
looking at things the right way. Maybe you just aren't using the
facilities of the language as well as you could be.
The next step is to add variable names. Now, though, we have a
slight problem. For the compiler, we had no problem in dealing
with variable names ... we just issued the names to the assembler
and let the rest of the program take care of allocating storage
for them. Here, on the other hand, we need to be able to fetch
the values of the variables and return them as the return values
of Factor. We need a storage mechanism for these variables.
Back in the early days of personal computing, Tiny BASIC lived.
It had a grand total of 26 possible variables: one for each
letter of the alphabet. This fits nicely with our concept of
single-character tokens, so we'll try the same trick. In the
beginning of your interpreter, just after the declaration of
variable Look, insert the line:
Table: Array['A'..'Z'] of integer;
We also need to initialize the array, so add this procedure:
{---------------------------------------------------------------}
{ Initialize the Variable Area }
procedure InitTable;
var i: char;
begin
for i := 'A' to 'Z' do
Table[i] := 0;
end;
{---------------------------------------------------------------}
You must also insert a call to InitTable, in procedure Init.
DON'T FORGET to do that, or the results may surprise you!
Now that we have an array of variables, we can modify Factor to
use it. Since we don't have a way (so far) to set the variables,
Factor will always return zero values for them, but let's go
ahead and extend it anyway. Here's the new version:
{---------------------------------------------------------------}
{ Parse and Translate a Math Factor }
function Expression: integer; Forward;
function Factor: integer;
begin
if Look = '(' then begin
Match('(');
Factor := Expression;
Match(')');
end
else if IsAlpha(Look) then
Factor := Table[GetName]
else
Factor := GetNum;
end;
{---------------------------------------------------------------}
As always, compile and test this version of the program. Even
though all the variables are now zeros, at least we can correctly
parse the complete expressions, as well as catch any badly formed
expressions.
I suppose you realize the next step: we need to do an assignment
statement so we can put something INTO the variables. For now,
let's stick to one-liners, though we will soon be handling
multiple statements.
The assignment statement parallels what we did before:
{--------------------------------------------------------------}
{ Parse and Translate an Assignment Statement }
procedure Assignment;
var Name: char;
begin
Name := GetName;
Match('=');
Table[Name] := Expression;
end;
{--------------------------------------------------------------}
To test this, I added a temporary write statement in the main
program, to print out the value of A. Then I tested it with
various assignments to it.
Of course, an interpretive language that can only accept a single
line of program is not of much value. So we're going to want to
handle multiple statements. This merely means putting a loop
around the call to Assignment. So let's do that now. But what
should be the loop exit criterion? Glad you asked, because it
brings up a point we've been able to ignore up till now.
One of the most tricky things to handle in any translator is to
determine when to bail out of a given construct and go look for
something else. This hasn't been a problem for us so far because
we've only allowed for a single kind of construct ... either an
expression or an assignment statement. When we start adding
loops and different kinds of statements, you'll find that we have
to be very careful that things terminate properly. If we put our
interpreter in a loop, we need a way to quit. Terminating on a
newline is no good, because that's what sends us back for another
line. We could always let an unrecognized character take us out,
but that would cause every run to end in an error message, which
certainly seems uncool.
What we need is a termination character. I vote for Pascal's
ending period ('.'). A minor complication is that Turbo ends
every normal line with TWO characters, the carriage return (CR)
and line feed (LF). At the end of each line, we need to eat
these characters before processing the next one. A natural way
to do this would be with procedure Match, except that Match's
error message prints the character, which of course for the CR
and/or LF won't look so great. What we need is a special proce-
dure for this, which we'll no doubt be using over and over. Here
it is:
{--------------------------------------------------------------}
{ Recognize and Skip Over a Newline }
procedure NewLine;
begin
if Look = CR then begin
GetChar;
if Look = LF then
GetChar;
end;
end;
{--------------------------------------------------------------}
Insert this procedure at any convenient spot ... I put mine just
after Match. Now, rewrite the main program to look like this:
{--------------------------------------------------------------}
{ Main Program }
begin
Init;
repeat
Assignment;
NewLine;
until Look = '.';
end.
{--------------------------------------------------------------}
Note that the test for a CR is now gone, and that there are also
no error tests within NewLine itself. That's OK, though ...
whatever is left over in terms of bogus characters will be caught
at the beginning of the next assignment statement.
Well, we now have a functioning interpreter. It doesn't do us a
lot of good, however, since we have no way to read data in or
write it out. Sure would help to have some I/O!
Let's wrap this session up, then, by adding the I/O routines.
Since we're sticking to single-character tokens, I'll use '?' to
stand for a read statement, and '!' for a write, with the char-
acter immediately following them to be used as a one-token
"parameter list." Here are the routines:
{--------------------------------------------------------------}
{ Input Routine }
procedure Input;
begin
Match('?');
Read(Table[GetName]);
end;
{--------------------------------------------------------------}
{ Output Routine }
procedure Output;
begin
Match('!');
WriteLn(Table[GetName]);
end;
{--------------------------------------------------------------}
They aren't very fancy, I admit ... no prompt character on input,
for example ... but they get the job done.
The corresponding changes in the main program are shown below.
Note that we use the usual trick of a case statement based upon
the current lookahead character, to decide what to do.
{--------------------------------------------------------------}
{ Main Program }
begin
Init;
repeat
case Look of
'?': Input;
'!': Output;
else Assignment;
end;
NewLine;
until Look = '.';
end.
{--------------------------------------------------------------}
You have now completed a real, working interpreter. It's pretty
sparse, but it works just like the "big boys." It includes three
kinds of program statements (and can tell the difference!), 26
variables, and I/O statements. The only things that it lacks,
really, are control statements, subroutines, and some kind of
program editing function. The program editing part, I'm going to
pass on. After all, we're not here to build a product, but to
learn things. The control statements, we'll cover in the next
installment, and the subroutines soon after. I'm anxious to get
on with that, so we'll leave the interpreter as it stands.
I hope that by now you're convinced that the limitation of sin-
gle-character names and the processing of white space are easily
taken care of, as we did in the last session. This time, if
you'd like to play around with these extensions, be my guest ...
they're "left as an exercise for the student." See you next
time.
*****************************************************************
* *
* COPYRIGHT NOTICE *
* *
* Copyright (C) 1988 Jack W. Crenshaw. All rights reserved. *
* *
*****************************************************************
--
--/*\ Jon A. Lambert - TychoMUD Internet:jlsysinc at ix.netcom.com /*\--
--/*\ Mud Server Developer's Page <http://www.netcom.com/~jlsysinc> /*\--
--/*\ "Everything that deceives may be said to enchant" - Plato /*\--
More information about the mud-dev-archive
mailing list