A couple of parts ago, I introduced pointers and implemented some code to check type compatibility. At the time, I realised that, for code like:
int c;
int *e;
e= &c + 1;
the addition of one to the pointer calculated by &c would need to be
converted into the size of c, to ensure we skip to the next int
in memory after c. In other words, we would have to scale the integer.
We need to do this for pointer, and later on we will need to do this for arrays. Consider the code:
int list[10];
int x= list[3];
To do this, we need to find the base address of list[], then add on
three times the size of int to find the element at index position 3.
At the time, I'd written a function in types.c called type_compatible()
to determine if two types were compatible, and to indicate if we needed to
"widen" a small integer type so that it was the same size as a larger integer
type. This widening, though, was performed elsewhere. In fact, it ended up
being done in three places in the compiler.
If type_compatible() indicated so, we would A_WIDEN a tree to match
a larger integer type. Now we need to A_SCALE a tree so that its value is
scaled by the size of a type. And I want to refactor the duplicate
widening code.
To this end, I've thrown out type_compatible() and replaced it. This took
quite a bit of thinking, and I probably will have to tweak or extend it again.
Let's look at the design.
The existing type_compatible():
- took two type values as arguments, plus an optional direction,
- returned true if the types were compatible,
- returned A_WIDEN on the left or right if either side needed to be widened,
- didn't actually add the A_WIDEN node to the tree,
- returned false if the types were not compatible, and
- didn't handle pointer types
Now let's look at the use cases for type comparisons:
- when performing a binary operation on two expressions, are their types compatible and do we need to widen or scale one?
- when doing a
printstatement, is the expression an integer and does it need widening? - when doing an assignment statement, does the expression need widening and does it match the lvalue type?
- when doing a
returnstatement, does the expression need widening and does it match the return type of the function?
In only one of these use cases do we have two expressions. Therefore, I've chosen to write a new function that takes one AST tree and the type we want it to become. For the binary operation use case, we will call it twice and see what happens for each call.
modify_type() in types.c is the replacement code for type_compatible().
The API for the function is:
// Given an AST tree and a type which we want it to become,
// possibly modify the tree by widening or scaling so that
// it is compatible with this type. Return the original tree
// if no changes occurred, a modified tree, or NULL if the
// tree is not compatible with the given type.
// If this will be part of a binary operation, the AST op is not zero.
struct ASTnode *modify_type(struct ASTnode *tree, int rtype, int op);
Question: why do we need whatever binary operation is being performed on the tree and some other tree? The answer is that we can only add to or subtract from pointers. We can't do anything else, e.g.
int x;
int *ptr;
x= *ptr; // OK
x= *(ptr + 2); // Two ints up from where ptr is pointing
x= *(ptr * 4); // Does not make sense
x= *(ptr / 13); // Does not make sense either
Here is the code for now. There are lots of specific tests, and at present I can't see a way to rationalise all the possible tests. Also, it will need to be extended later.
struct ASTnode *modify_type(struct ASTnode *tree, int rtype, int op) {
int ltype;
int lsize, rsize;
ltype = tree->type;
// Compare scalar int types
if (inttype(ltype) && inttype(rtype)) {
// Both types same, nothing to do
if (ltype == rtype) return (tree);
// Get the sizes for each type
lsize = genprimsize(ltype);
rsize = genprimsize(rtype);
// Tree's size is too big
if (lsize > rsize) return (NULL);
// Widen to the right
if (rsize > lsize) return (mkuastunary(A_WIDEN, rtype, tree, 0));
}
// For pointers on the left
if (ptrtype(ltype)) {
// OK is same type on right and not doing a binary op
if (op == 0 && ltype == rtype) return (tree);
}
// We can scale only on A_ADD or A_SUBTRACT operation
if (op == A_ADD || op == A_SUBTRACT) {
// 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 (mkuastunary(A_SCALE, rtype, tree, rsize));
}
}
// If we get here, the types are not compatible
return (NULL);
}
The code to add the AST A_WIDEN and A_SCALE operations are now done
here in one place only. The A_WIDEN operation converts the child's
type to the parent's type. The A_SCALE operation multiplies the child's
value by the size which is store in the new struct ASTnode union field
(in defs.h):
// Abstract Syntax Tree structure
struct ASTnode {
...
union {
int size; // For A_SCALE, the size to scale by
} v;
};
With this new API, we can remove the duplicated code to A_WIDEN which is
in stmt.c and expr.c. However, this new function only takes one tree.
This is fine when we indeed only have one tree. There are three calls now
to modify_type() in stmt.c. They are all similar, so here is the one
from assignment_statement():
// Make the AST node for the assignment lvalue
right = mkastleaf(A_LVIDENT, Gsym[id].type, id);
...
// Parse the following expression
left = binexpr(0);
// Ensure the two types are compatible.
left = modify_type(left, right->type, 0);
if (left == NULL) fatal("Incompatible expression in assignment");
which is so much cleaner than the code we had before.
But in binexpr() in expr.c, we now need to combine two AST trees
with a binary operations like addition, multiplication etc. Here, we
try to modify_type() each tree with the other tree's type. Now, one
may widen: this also implies that the other will fail and return NULL.
Thus, we can't just see if one result from modify_type() is NULL: we
need to see both be NULL to assume a type incompatibility. Here's the
new comparison code in binexpr():
struct ASTnode *binexpr(int ptp) {
struct ASTnode *left, *right;
struct ASTnode *ltemp, *rtemp;
int ASTop;
int tokentype;
...
// Get the tree on the left.
// Fetch the next token at the same time.
left = prefix();
tokentype = Token.token;
...
// Recursively call binexpr() with the
// precedence of our token to build a sub-tree
right = binexpr(OpPrec[tokentype]);
// Ensure the two types are compatible by trying
// to modify each tree to match the other's type.
ASTop = arithop(tokentype);
ltemp = modify_type(left, right->type, ASTop);
rtemp = modify_type(right, left->type, ASTop);
if (ltemp == NULL && rtemp == NULL)
fatal("Incompatible types in binary expression");
// Update any trees that were widened or scaled
if (ltemp != NULL) left = ltemp;
if (rtemp != NULL) right = rtemp;
The code is a little bit messy but no more that what was previously there, and it now deals with A_SCALE as well as A_WIDEN.
We have added the A_SCALE to the list of AST node operations in defs.h.
Now we need to implement it.
As I mentioned before, the A_SCALE operation
multiplies the child's value by the size which is store in the new struct ASTnode union field. For all of our integer types, this will be a multiple
of two. Because of this fact, we can multiply the child's value with a
shift left of a certain number of bits.
Later on, we will have structs that have a size which isn't a power of two. So we can do a shift optimisation when the scale factor is suitable, but we also need to implement a multiply for a more general scale factor.
Here is the new code that does this in genAST() in gen.c:
case A_SCALE:
// Small optimisation: use shift if the
// scale value is a known power of two
switch (n->v.size) {
case 2: return(cgshlconst(leftreg, 1));
case 4: return(cgshlconst(leftreg, 2));
case 8: return(cgshlconst(leftreg, 3));
default:
// Load a register with the size and
// multiply the leftreg by this size
rightreg= cgloadint(n->v.size, P_INT);
return (cgmul(leftreg, rightreg));
We now need a cgshlconst() function to shift a register value left
by a constant. When we add the C '<<' operator later, I will write a
more general shift left function. For now, we can use the salq instruction
with an integer literal value:
// Shift a register left by a constant
int cgshlconst(int r, int val) {
fprintf(Outfile, "\tsalq\t$%d, %s\n", val, reglist[r]);
return(r);
}
My test program for the scaling functionality is tests/input16.c:
int c;
int d;
int *e;
int f;
int main() {
c= 12; d=18; printint(c);
e= &c + 1; f= *e; printint(f);
return(0);
}
I was hoping that d would be placed immediately after c by the assembler
when we generate these assembly directives:
.comm c,1,1
.comm d,4,4
But when I compiled the assembly output and checked, they were not adjacent:
$ cc -o out out.s lib/printint.c
$ nm -n out | grep 'B '
0000000000201018 B d
0000000000201020 B b
0000000000201028 B f
0000000000201030 B e
0000000000201038 B c
d actually comes before c! I had to work out a way to ensure
the adjacency, so I looked at the code that SubC generates here,
and changed our compiler to now generate this:
.data
.globl c
c: .long 0 # Four byte integer
.globl d
d: .long 0
.globl e
e: .quad 0 # Eight byte pointer
.globl f
f: .long 0
Now when we run our input16.c test, e= &c + 1; f= *e; gets the address
of the integer one up from c and stores that integer's value in f. As
we declared:
int c;
int d;
...
c= 12; d=18; printint(c);
e= &c + 1; f= *e; printint(f);
we will print out both numbers:
cc -o comp1 -g -Wall cg.c decl.c expr.c gen.c main.c misc.c
scan.c stmt.c sym.c tree.c types.c
./comp1 tests/input16.c
cc -o out out.s lib/printint.c
./out
12
18
I feel a lot happier with the code that converts between types. Behind
the scenes, I wrote some test code that supplied all possible type
values to modify_type(), as well as a binary operation and zero for
the operation. I eyeballed the output and it seems to be what I want.
Only time will tell.
In the next part of our compiler writing journey, I don't know what I will do!