In this part of our compiler writing journey, I fix up a few more small issues with the compiler.

Our compiler doesn't have a run-time -D flag to define a symbol to the pre-processor, and it would be somewhat complicated to add it in. But we use this in the Makefile to set the location of the directory where our header files are.

I've rewritten the Makefile to write this location into a new header file:

# Define the location of the include directory
INCDIR=/tmp/include
...

incdir.h:
        echo "#define INCDIR \"$(INCDIR)\"" > incdir.h

and in defs.h we now have:

#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <ctype.h>
#include "incdir.h"

This ensures that the location of this directory is known to the source code.

I've added these three external variables in include/stdio.h:

extern FILE *stdin;
extern FILE *stdout;
extern FILE *stderr;

but when I tried to use them they were being treated as local variables! It turns out my logic to choose a global variable was wrong. In genAST() in gen.c, we now have:

    case A_IDENT:
      // Load our value if we are an rvalue
      // or we are being dereferenced
      if (n->rvalue || parentASTop == A_DEREF) {
        if (n->sym->class == C_GLOBAL || n->sym->class == C_STATIC
            || n->sym->class == C_EXTERN) {
          return (cgloadglob(n->sym, n->op));
        } else {
          return (cgloadlocal(n->sym, n->op));
        }

with the C_EXTERN alternative being added.

Way back in part 3 of this journey, I introduced the Pratt parser which has a table of precedence values associated with each token. We've been using it ever since as it works.

However, I've introduced tokens that don't get parsed by the Pratt parser: prefix operators, postfix operators, casts, array element access etc. And along the way I broke the chain that ensures the Pratt parser knows the precedence of the previous operator token.

Here is the basic Pratt algorihm again, as shown by the code in binexpr() in expr.c:

  // Get the tree on the left.
  // Fetch the next token at the same time.
  left = prefix();
  tokentype = Token.token;

  // While the precedence of this token is more than that of the
  // previous token precedence, or it's right associative and
  // equal to the previous token's precedence
  while ((op_precedence(tokentype) > ptp) ||
         (rightassoc(tokentype) && op_precedence(tokentype) == ptp)) {
    // Fetch in the next integer literal
    scan(&Token);

    // Recursively call binexpr() with the
    // precedence of our token to build a sub-tree
    right = binexpr(OpPrec[tokentype]);

    // Join that sub-tree with ours (code not given)

    // Update the details of the current token.
    // Leave the loop if a terminating token (code not given)
    tokentype = Token.token;
  }

  // Return the tree we have when the precedence
  // is the same or lower
  return (left);

We must ensure that binexpr() gets called with the precedence of the previous token. Now let's look at how this got broken.

Consider this expression that checks if three pointers are valid:

  if (a == NULL || b == NULL || c == NULL)

The == operator has higher precedence that the || operator, so the Pratt parser should treat this the same as:

  if ((a == NULL) || (b == NULL) || (c == NULL))

Now, NULL is defined as this expression, and it includes a cast:

#define NULL (void *)0

So let's look at the call chain of the IF line above:

• binexpr(0) is called from if_statement()
• binexpr(0) parses the == (which has precedence 40) and calls binexpr(40)
• binexpr(40) calls prefix()
• prefix() calls postfix()
• postfix() calls primary()
• primary() sees the left parenthesis at the start of the (void *)0 and calls paren_expression()
• paren_expression() sees the void token and calls parse_cast(). Once the cast is parsed, it calls binexpr(0) to parse the 0.

And that's the problem. The value of NULL, i.e. 0 should still be at precedence level 40, but paren_expression() just reset it back to zero.

This means that we will now parse NULL || b, making an AST tree out of it instead of parsing a == NULL and building that AST tree.

The solution is to ensure that the previous token precedence is passed through the call chain all the way from binexpr() up to paren_expression(). This means that:

• prefix(), postfix(), primary() and paren_expression()

all now take an int ptp argument and this is passed on.

The program tests/input143.c checks that this change now works for if (a==NULL || b==NULL || c==NULL).

A while back, I realised that if we were adding an integer value to a pointer, we needed to scale the integer by the type size that the pointer points at. For example:

int list[]= {3, 5, 7, 9, 11, 13, 15};
int *lptr;

int main() {
  lptr= list;
  printf("%d\n", *lptr);
  lptr= lptr   1; printf("%d\n", *lptr);
}

should print the value at the base of list, i.e. 3. The lptr should be incremented by the size of int, i.e. 4, so that it now points at the next element in the list.

Now, we do this for the and - operators, but I forgot to implement it for the = and -= operators. Fortunately this was easy to fix. At the bottom of modify_type() in types.c, we now have:

  // We can scale only on add and subtract operations
  if (op == A_ADD || op == A_SUBTRACT ||
      op == A_ASPLUS || op == A_ASMINUS) {

    // Left is int type, right is pointer type and the size
    // of the original type is >1: scale the left
    if (inttype(ltype) && ptrtype(rtype)) {
      rsize = genprimsize(value_at(rtype));
      if (rsize > 1)
        return (mkastunary(A_SCALE, rtype, rctype, tree, NULL, rsize));
      else
        return (tree);          // Size 1, no need to scale
    }
  }

You can see I've added A_ASPLUS and A_ASMINUS to the list of operations where we can scale an int value.

That's enough mopping up for now. When I fixed up the = and -= problem, it highlighted a big issue with the and -- operators (prefix and postfix) as applied to pointers.

In the next part of our compiler writing journey, I will tackle this issue. Next step