Sild - header files and a refactoring

July 03, 2016

I’ve now implemented the following built in functions:

And they work. Great! We’re getting closer to something useful, but before moving on, this is a great time to stop for a refactoring!

Up until now, the entire program has lived in one big file sild.c, which has everything I’ve written from top to bottom in rough dependency order and a main() function at the end. The file is 475 lines long, which is pretty long! I can do better; I need to find a way to separate this file into logical units that #include each other, and the .c file that contains main shouldn’t have that much else inside of it.

I struggled with this one for quite a while, actually! There are a lot of ways to get C code into the final executable that are bogus, on one iteration I was inline including .c files into the files they depend on, which totally works, but is a major giant antipattern for lots of reasons that I had no idea about. I was thinking like Ruby, where you just require a file, and it is read in, and everything is fine. C doesn’t work that way! To start with, the whole concept of header files was new to me- I had touched them in Objective-C and was taught that they “define an interface” to a library. This is true! But also pretty vague! Does a .c file always include its own header? do all functions need to be defined in the header, or just the ones you want to expose? Do you fully define structs and unions in the header file, or simply typedef them? Do you initialize global variables in the header file? How the hell does all this get linked together, really? Lots of questions, I had. I’ll skip the details of a lot of my mis-adventures, and instead focus on what I eventually found to be a reasonable set of rules of thumb for good compilation practices.

There is a lot of weird info on the interwebs about this, too… and nothing was one hundred percent clearly the “best way” to factor out code into libraries. I found this to be helpful:

C header file guidelines.

and this:

Best C coding practices for header files

But ultimately the set of golden rules came from friend Andrew Kelley, and they boiled down to something like this.

  1. Each .o file is produced independently from all other .o files via a separate invocation of the compiler…

.o stands for “object” file. An object file is compiled C code, and is non-executable. Let’s say I have a .c file with some functions inside of it, and call

$ cc myfile.c

By default I’m going to get a file called a.out that is an executable. I can explicitly set a target with the -o flag, which specifies the output file name:

$ cc myfile.c -o myfile

This is the only line that has been in my makefile this entire time, as a matter of fact. (I’ll get way deep into makefiles in a while!)

sild: sild.c
	cc sild.c -o sild

And when I run make, it compiles sild.c into sild as an executable, and I can run it, yay!

BUT, if I take out the main function, and try to compile that, I get a nasty compiler error:

Undefined symbols for architecture x86_64:
  "_main", referenced from:
     implicit entry/start for main executable
ld: symbol(s) not found for architecture x86_64
clang: error: linker command failed with exit code 1 (use -v to see invocation)

The compiler is trying to make an executable, but an executable needs to know where to start, which is implicitly a main function. If I want to compile arbitrary C code into a lump of machine instructions, what I want is an object file: a .o file! I could do that with the -c flag:

$ cc -c myfile.c

By default, this will compile to an object file of the same name: myfile.o which contains that arbitrary machine code and anything else you define in there, like constants or variables or whatever!

  1. … So really, in C, your goal is merely to produce a bunch of .o files to link together into a final library or executable. The reason you might have more than one .o file is for your own abstraction benefit.

Ok, so let’s see. I have already separated my code into some sections, and my goal is to abstract those sections into libraries. I’ll start with an easy one! At the very top of sild.c, I have my replacement strcmp() function that I called scmp().

/* ----------*/
/* utilities */
/* ----------*/

int scmp(char *str1, char *str2) {
    int i;
    for (i = 0; str1[i] != '\0'; i++) {
        if (str1[i] != str2[i] || str2[i] == '\0') {
            return 0;
        }
    }
    if (str2[i] == '\0') {
        return 1;
    } else {
        return 0;
    }
}

This is a great place to start since this doesn’t depend on any other standard libraries or even any functions within my own code. I can pull it straight out into another file called utils.c. This file will NOT have a main() function, and I will want to compile it on its own as an object file, with a command like:

cc utils.c -c

… which yields an object file `utils.o’. Since this doesn’t depend on anything, that’s all I have to do! Just the body of that function will successfully compile.

But when I go back and try to compile the sild.c file, as you would expect, I get an error:

Undefined symbols for architecture x86_64:
    "_scmp", referenced from:
    _eq in sild-XPBvI6.o
    _categorize in sild-XPBvI6.o
ld: symbol(s) not found for architecture x86_64
clang: error: linker command failed with exit code 1 (use -v to see invocation)

This is basically the same error as before: the compiler sees that functions inside of sild.c call scmp(), but it doesn’t know where that function is, it doesn’t know how to link them together. This is where I was tempted to do this in sild.c at the very top:

#include <stdio.h>
#include <stdlib.h>
#include "utils.c"

(Angled brackets search the system libraries path first, and quotes search the current directory the file is actually in first, for a matching name.)

This works! Because the #include direction instructs the compiler to simply insert the contents of that file right at that line before compiling the whole file together. (#include, like #define and other octothorpe beginning thingers, are compiler preprocessor commands.) You can just compile that main file and it will work as you expect.

cc sild.c -o sild

But, this is bad news bears, and scales terribly! Consider this:

#include <stdio.h>
#include <stdlib.h>
#include "utils.c"
#include "utils.c"

On a compilation attempt, gives the following error:

In file included from sild.c:5:
./utils.c:5:5: error: redefinition of 'scmp'
int scmp(char *str1, char *str2) {
    ^
./utils.c:5:5: note: previous definition is here
int scmp(char *str1, char *str2) {
    ^

And of course it does! It’s inlining the file twice! Just as if I had typed in the same function twice, C will complain that it has been redefined, which is a no no.

This may seem like a contrived example, but consider including utils.c in a file that also includes another file that also include utils.c, both for valid reasons (in that they both use the function in utils.c independently). Boom, suddenly you’ve got the very same problem! You’ve inlined the same code more than once and everything sucks. The answer to this problem, is header files!

As Andrew said:

Most C programmers would find this odd, typically you include .h files and compile .c files independently into .o files, then link them all together.

I didn’t really understand this advice fully for a while, so let me explain what that means.

Take utils.c for example! A corresponding header file for this library would be called utils.h, and it would include the function prototype for any code the library contains that will be called from outside that library’s source file. A prototype simple states the returning type, the name, and the argument types. The arguments can be unnamed at this point, or they can have their name in the prototype for clarity, but it doesn’t matter. Basically, you just copy the first line of the function.

int scmp(char*, char*);

The actual definition of the function still happens in the .c file, but declaring the function in the header allows the compiler to know about it.

We replace the #include "utils.c" with #include "utils.h" in sild.c`, and we get the following, same error as before, when we weren’t including anything at all and had just taken the function body out.

Undefined symbols for architecture x86_64:
    "_scmp", referenced from:
    _eq in sild-XPBvI6.o
    _categorize in sild-XPBvI6.o
ld: symbol(s) not found for architecture x86_64
clang: error: linker command failed with exit code 1 (use -v to see invocation)

And of course we do! The function prototype is not the function definition! This brings us to this foundational rule of thumb:

  1. Every .c and .h file should be able to be compiled into an object file on its own.

And what Andrew had to say about this was:

The idea is to include as little as possible (because this improves compile time) while still including what you depend on (this prevents shifting downstream dependencies from breaking upstream code). So, step 1, can you avoid including it? Then don’t include it. Step 2, can you only include it in the .c file and not in the .h file? Then only include it in the .c file. Step 3, looks like your .h file depends on it, so you’ll have to include it in your .h file.

Let’s say I try to compile sild.c into it’s own object file without include the header file from utils.h:

cc -c sild.c

I get a new, different error, actually a warning:

sild.c:225:14: warning: implicit declaration of function 'scmp' is invalid in C99 [-Wimplicit-function-declaration]
     scmp(operand->val.label, operand2->val.label)
      ^
      1 warning generated.

But if I compile sild.c into am object file and include utils.h, it works, and I get sild.o out of it.

Further, to pass this sniff test, both utils.c and utils.h should be able to be compiled into an object file. They both do, because there is only one function that depends on nothing else.

So great, but how do I get an executable out of this? I first compile the libraries into their object files, then link those all together and bundle them with the c file that has main() inside of it. Those commands look like this:

cc -c utils.c

yields utils.o, and

cc utils.o sild.c -o sild

Will link that library into the executable. Success!

What about the header file itself?

The header file itself doesn’t need to get bundled into this. Basically, when the header file states the function prototype, it tells the compiler that that function is going to exist, so the compiler assigns it an address, like setting aside a house for it. When the function is defined, it will move into that house. As long as the executable knows where to find that function when it is run, it will be fine!


Here are some other considerations. What if scmp() depended on some helper function that also would live in utils.c, but this helper function didn’t need to be exposed to the consumers of the utils library? What if it looked something like this:

// utils.c

int helper_function(char *string) {
    int result;
    // stuff
    return result
}

int scmp(char *str1, char *str2) {
    int thing_i_need = helper_function(str1);
    // stuff that uses thing_i_need
    // ...
    return // whatever
}

In this case, helper_function() does not need to live in the header file at all. It is an implementation detail that lives inside the black box that is utils.c, and as such, it should be marked static, like this:

static int helper_function(char *string) {
    int result;
    // stuff
    return result
}

Which tells the compiler that it is not needed outside of its defining module, effectively limiting that function’s scope to it’s translation unit (the object file that that c file is compiled into)!

  1. …oh yeah, and if a function is only used in this file, mark it as static and don’t give it a prototype in the .h file. this will optimize better and keeps things more encapsulated

Header guards

It is traditional to wrap .h files inside of header guards so that they are only processed once no matter how many files #include them. A header guard looks like this:

#ifndef UNIQUENAME_GUARD // if UNIQUENAME_GUARD is not defined
#define UNIQUENAME_GUARD // define it (to nothing, but it is still defined)

// ...header file prototypes, constants, etc

#endif // end the if block

All of the standard library headers have this, for example, stdio.h starts with:

#ifndef _STDIO_H_
// ...some stuff
#define _STDIO_H_
// ...some stuff
#endif /* _STDIO_H_ */

Guard constants that start with underscores are reserved for system library usage.

Not processing all these prototypes multiple times is good for compiler performance. It doesn’t really matter so much in a small project like mine, but in a bigger project with giant header files, that type of optimization can really make a difference, and it is good practice to do it.

So, utils.h will look like this:

#ifndef UTILS_GUARD
#define UTILS_GUARD

int scmp(char*, char*);

#endif

And that’s that! To recap:

  1. Every .c file gets a .h file that exposes the function prototypes of its public facing functions.
  2. Every .c and .h file should be able to compile into an object file independantly of all other files while using the -c flag.
  3. functions etc should be marked static if they are not being used outside of their compilation unit.
  4. .h files should include header guards as a general rule.

I went through sild.c and factored out 6 distinct units in this way.

  1. utils.c has scmp().
  2. cell.c has all the cell structures and con/destructors.
  3. eval.c has eval/apply and exposes eval.
  4. builtins.c has all of the builtin functions.
  5. print.c has all of the debug/print functions.
  6. read.c contains the reader and exposes read().

One thing I want to point out is that struct and union prototypes should include their member declarations if their internals need to be accessed outside of the compilation unit. In my case, they did, so cell.h has those, and cell.c includes its own .h file since it needs access to those definitions as well.


The only thing left in sild.c is the main() function, so I’ll rename that to main.c

Now, to get an executable, I would compile each library module (including main.c itself) into an object file and then link them all together, just like Andrew said originally.

cc utils.c -o
cc cell.c -o
cc eval.c -o
cc builtins.c -o
cc print.c -o
cc read.c -o
cc main.c -o
cc utils.o cell.o eval.o builtins.o print.o read.o main.o -o sild

These are a lot of commands to run every time I want to get a new executable! If only there were a way to further automate this build process…


Very big thanks to Andrew Kelley for his extremely well reasoned and cogently worded answers to my noob C questions. I was very surprised how much of a mind bender resolving dependencies in C turned out to be, and though I’m sure there is still a lot to learn about this, I feel very good about the practices presented in this post being a reasonable starting point. I can also see how undisciplined dependancy management can be a hell of a noose, an intuition that was born out in my research. This is a hard problem!