In this part of our compiler writing journey, I've added the ability to write function prototypes. In the process, I've had to rewrite some of the code that I'd just written in the previous parts; sorry about that. I didn't see far enough ahead!
So what do we want with function prototypes:
- the ability to declare a function prototype with no body
- the ability to declare a full function later on
- to keep the prototype in the global symbol table section, and the parameters as local variables in the local symbol table section
- error checking on the number and types of parameters against a previous function prototype
And here is what I'm not going to do, at least not yet:
function(void): this will be the same as thefunction()declaration- the declaration of a function with just the types, e.g.
function(int ,char, long);as this will make the parsing harder. We can do this later.
In a recent part of our journey I added the declaration of a function with parameters and a full function body. As we were parsing each parameter, I immediately added it to both the global symbol table (to form the prototype) and also the local symbol table (to be the function's parameters).
Now that we want to implement function prototypes, it's not always true that a parameter list will become the actual function's parameters. Consider this function prototype:
int fred(char a, int foo, long bar);We can only define fred as a function, and a, foo and bar as
three parameters in the global symbol table. We have to wait until the
full function declaration before we can add a, foo and bar to
the local symbol table.
I'll need to separate the definition of C_PARAM entries on the global symbol table and on the local symbol table.
Here is my quick design for the new function parsing mechanism which also deals with prototypes.
Get the identifier and '('.
Search for the identifier in the symbol table.
If it exists, there is already a prototype: get the id position of
the function and its parammeter count.
While parsing parameters:
- if a previous prototype, compare this param's type against the existing
one. Update the symbol's name in case this is a full function
- if no previous prototype, add the parameter to the symbol table
Ensure # of params matches any existing prototype.
Parse the ')'. If ';' is next, done.
If '{' is next, copy the parameter list from the global symtable to the
local sym table. Copy them in a loop so that they are put in reverse order
in the local sym table.
I got this done in the last few hours, so here are the code changes.
I've changed the parameter list of a couple of functions in sym.c:
int addglob(char *name, int type, int stype, int class, int endlabel, int size);
int addlocl(char *name, int type, int stype, int class, int size);Previously, we had addlocl() also call addglob() to add a C_PARAM
symbol to both symbol tables. Now that we are separating this function,
it makes sense to pass the actual class of the symbol to both functions.
There are calls to these functions in main.c and decl.c. I'll cover
the ones in decl.c later. The change in main.c is trivial.
Once we hit the declaration of a real function, we will need to copy
its parameter list from the global to the local symbol table. As this is
really something specific to the symbol table, I've added this function to
sym.c:
// Given a function's slot number, copy the global parameters
// from its prototype to be local parameters
void copyfuncparams(int slot) {
int i, id = slot + 1;
for (i = 0; i < Symtable[slot].nelems; i++, id++) {
addlocl(Symtable[id].name, Symtable[id].type, Symtable[id].stype,
Symtable[id].class, Symtable[id].size);
}
}Nearly all of the changes to the compiler are confined to decl.c. We'll
start with the small ones and work up to the big ones.
I've changed the parameter list to var_declaration() in the same way
that I did for the sym.c functions:
void var_declaration(int type, int class) {
...
addglob(Text, pointer_to(type), S_ARRAY, class, 0, Token.intvalue);
...
if (addlocl(Text, type, S_VARIABLE, class, 1) == -1)
...
addglob(Text, type, S_VARIABLE, class, 0, 1);
}We will use the ability to pass in the class in the other decl.c
functions.
We have big changes here, as we might already have a parameter list in the global symbol table as an existing prototype. If we do, we need to check the number and types in the new list against the prototype.
// Parse the parameters in parentheses after the function name.
// Add them as symbols to the symbol table and return the number
// of parameters. If id is not -1, there is an existing function
// prototype, and the function has this symbol slot number.
static int param_declaration(int id) {
int type, param_id;
int orig_paramcnt;
int paramcnt = 0;
// Add 1 to id so that it's either zero (no prototype), or
// it's the position of the zeroth existing parameter in
// the symbol table
param_id = id + 1;
// Get any existing prototype parameter count
if (param_id)
orig_paramcnt = Symtable[id].nelems;
// Loop until the final right parentheses
while (Token.token != T_RPAREN) {
// Get the type and identifier
// and add it to the symbol table
type = parse_type();
ident();
// We have an existing prototype.
// Check that this type matches the prototype.
if (param_id) {
if (type != Symtable[id].type)
fatald("Type doesn't match prototype for parameter", paramcnt + 1);
param_id++;
} else {
// Add a new parameter to the new prototype
var_declaration(type, C_PARAM);
}
paramcnt++;
// Must have a ',' or ')' at this point
switch (Token.token) {
case T_COMMA:
scan(&Token);
break;
case T_RPAREN:
break;
default:
fatald("Unexpected token in parameter list", Token.token);
}
}
// Check that the number of parameters in this list matches
// any existing prototype
if ((id != -1) && (paramcnt != orig_paramcnt))
fatals("Parameter count mismatch for function", Symtable[id].name);
// Return the count of parameters
return (paramcnt);
}Remember that the first parameter's global symbol table slot position is immediately after the slot for the function name's symbol. We get passed the slot position of an existing prototype, or -1 if there is no prototype.
It's a happy coincidence that we can add one to this to get the first parameter's slot number, or have 0 to indicate that there is no existing prototype.
We still loop parsing each new parameter, but now there is new code to either compare against the existing prototype, or to add the parameter to the global symbol table.
Once we exit the loop, we can compare the number of parameters in this list against the number in any existing prototype.
Right now, the code feels a bit ugly and I'm sure that if I leave it a while, I'll be able to see a way to refactor it a bit.
Previously, this was a fairly simple function: get the type and name, add a global symbol, read in the parameters, get the function's body and generate an AST tree for the function's code.
Now, we have to deal with the fact this this might only be a prototype, or it could be a full function. And we won't know until we parse either the ';' (for a prototype) or the '{' (for a full function). So let's take the exposition of the code in stages.
// Parse the declaration of function.
// The identifier has been scanned & we have the type.
struct ASTnode *function_declaration(int type) {
struct ASTnode *tree, *finalstmt;
int id;
int nameslot, endlabel, paramcnt;
// Text has the identifier's name. If this exists and is a
// function, get the id. Otherwise, set id to -1
if ((id = findsymbol(Text)) != -1)
if (Symtable[id].stype != S_FUNCTION)
id = -1;
// If this is a new function declaration, get a
// label-id for the end label, and add the function
// to the symbol table,
if (id == -1) {
endlabel = genlabel();
nameslot = addglob(Text, type, S_FUNCTION, C_GLOBAL, endlabel, 0);
}
// Scan in the '(', any parameters and the ')'.
// Pass in any existing function prototype symbol slot number
lparen();
paramcnt = param_declaration(id);
rparen();This is nearly the same as the previous version of the code, except that
id is now set to -1 when there is no previous prototype or a positive
number when there is a previous prototype. We only add the function's
name to the global symbol table if it's not already there.
// If this is a new function declaration, update the
// function symbol entry with the number of parameters
if (id == -1)
Symtable[nameslot].nelems = paramcnt;
// Declaration ends in a semicolon, only a prototype.
if (Token.token == T_SEMI) {
scan(&Token);
return (NULL);
}We've got the count of parameters. If no previous prototype, update this
prototype with this count. Now we can peek at the token after the end
of the parameter list. If it's a semicolon, this is just a prototype.
We now have no AST tree to return, so skip the token and return NULL.
I've had to slightly alter the code in global_declarations() to deal
with this NULL value: no big change.
If we continue on, we are now dealing with a full function declaration with a body.
// This is not just a prototype.
// Copy the global parameters to be local parameters
if (id == -1)
id = nameslot;
copyfuncparams(id);We now need to copy the parameters from the prototype to the local symbol
table. The id = nameslot code is there for when we have just added
the global symbols ourselves and there was no previous prototype.
The rest of the code in function_declaration() is the same as before and
I'll omit it. It checks that a non-void function does return a value, and
generates the AST tree with an A_FUNCTION root node.
One of the drawbacks of the tests/runtests script is that it assumes
the compiler will definitely produce an assembly output file out.s
which can be assembled and run. This prevents us from testing that the
compiler detects syntax and semantic errors.
A quick grep of decl.c shows these new errors are detected:
fatald("Type doesn't match prototype for parameter", paramcnt + 1);
fatals("Parameter count mismatch for function", Symtable[id].name);Thus, I'd better rewrite tests/runtests to verify that the compiler
does detect these errors on bad input.
We do have two new working test programs, input29.c and input30.c.
The first one is the same as input28.c except that I've put the
prototypes of all the functions at the top of the program:
int param8(int a, int b, int c, int d, int e, int f, int g, int h);
int fred(int a, int b, int c);
int main();This, and all previous test programs, still work. input30.c, though, is
probably the first non-trivial program that our compiler has been given.
It opens its own source file and prints it to standard output:
int open(char *pathname, int flags);
int read(int fd, char *buf, int count);
int write(int fd, void *buf, int count);
int close(int fd);
char *buf;
int main() {
int zin;
int cnt;
buf= " ";
zin = open("input30.c", 0);
if (zin == -1) {
return (1);
}
while ((cnt = read(zin, buf, 60)) > 0) {
write(1, buf, cnt);
}
close(zin);
return (0);
}We can't yet call the pre-processor, so we manually put in the prototypes
for the open(), read(), write() and close() functions. We also
have to use 0 instead of O_RDONLY in the open() call.
Right now, the compiler lets us declare a char buf[60]; but we can't use
buf itself as a char pointer. So I chose to assign a 60-character literal
string to a char pointer and we use this as the buffer.
We still also have to wrap IF and WHILE bodies with '{' ... '}' to make
them compound statements: I still haven't dealt with the dangling else
problem. Finally, we can't accept char *argv[] as a parameter declaration
for main yet, so I've had to hard-code the input file's name.
Still, we now have a very primitive cat(1) program which our compiler can compile! That's progress.
In the next part of our compiler writing journey, I'll follow up on a comment above and improve the testing of our compiler's functionality.