how to write sane C




BASIC HYGIENE

H0. use C11 (or newer)
H1. do not use deprecated features (old-style declarations, trigraphs, gets,...)
H2. never use the following keywords: const inline register short "long long"
H3. never use the following includes: wchar.h wctype.h
H4. do not use casts
H5. do not typedef structs
H6. do not typedef pointer types
H7. POSIX features are OK, but hidden behind macros so that the code compiles
    without them (producing either slower binaries or error messages at runtime)
H8. OPENMP pragmata, includes and functions are hidden behind _OPENMP checks
H9. SSE and AVX stuff is also hidden behind macros, and semantically redundant



NAMING

N0. mixed-case symbols are forbidden
N1. all local variables are single letters (like in the rest of mathematics)
    - exception: letters like "alpha", "beta" (gcc does not yet grok UTF8)
    - exception: variations of a single letter like "x_in", "x_out"
    - exception: if a variable is only used once, it may have a longer name
N2. all global variables are self-documenting and have the form "global_*"
N3. all the symbols exposed by a translation unit share a prefix ending in _
N4. exposed function names are long, self-documenting and all-lowercase
N5. static functions may have shorter names and, if needed, use upper-case
N6. struct tags are all-lowercase
N7. macros are all-uppercase
N8. struct members are all-lowercase



FORMATTING

F0. indent with 1 tab or with 8 spaces (but not both on the same file)
F1. the maximum line width is 80 (1 indentation level counts as 8)
F2. functions have at most 20 or 25 lines, unless they have some symmetry
F3. no declaration without initialization
    - exception: when the initialization is via a pointer on the next line
F4. max about 7 variables in scope at the same time



COMMENTING

C0. each declaration on a separate line, with a descriptive comment
C1. the arguments of a function go on separate lines, with descriptive comments
C2. static functions with obvious arguments may be declared in one line
C3. avoid comments if possible: change "// compute stuff" to "compute_stuff();"
C4. do not include boilerplate at the top of each file



STYLE

S0. do not be afraid of the C language, master all the available features
S1. do not cast: either use implicit conversions or write conversion functions
S2. local variables are pronouns, comments are the names (see also N1 and F0)
S3. unsigned types are only used for bit fields
S4. all loops are exactly of the form "for (int i=0; i<n; i++)"
    - n is a constant or a variable whose value can be estimated at compile-time
S5. strive to make the body of each loop a single statement
S6. non-tail recursion is only allowed if the depth is O(log(input size))
S7. use pointers to VLA for simple array arithmetic
S8. use gotos only for function cleanup code, allowing a single exit point
S9. use many small functions that are called only once
S10. strive to avoid call diamonds
S11. try not to increment pointers (p++), better use p[i] and then increment i
S12. do not compose calls, use temporary variables for the intermediate results
S99. design things to be non-scalable
S100. write the less general code that solves the problem




ENUMS, STRUCTS, TYPEDEFS

E0. use enums to list all the possible cases of a switch statement
E1. use typedefs for basic types (e.g. float) that you want to abstract away
E2. use typedefs to declare pointers to functions
E3. do not use typedefs for anything else
E4. use structs to hold related variables together
E5. use structs to represent abstract data types
E6. do not use structs for anything else
E7. structs only need member functions if those are supposed to change
E8. avoid getters and setters, access the struct fields directly



INTERFACE

I0. separate interface and implementation
I1. the interface is easy on the user and hard on the programmer
I2. the interface does not expose macros, typedefs nor structs
I3. the interface only exposes functions of plain data types
I4. each file must be compilable stand-alone producing a useful CLI program
I5. each file must be compilable with others as a single translation unit
I6. use macros to enable different usages of the same file
I7. the interface does not allocate memory: it fills-in a user-given array
I8. allocate memory only inside main(), then you do not need to free it!
I9. prefer a single large file than several smaller files
    - exception: an archive of many functions that you want to link separately
I10. foo.c does never include foo.h



REFERENCES

R0. K&R
R1. Notes on programmin in C, Rob Pike, 1989
R2. the MISRA C style guide
R3. the JPL C style gude
R4. the Linux kernel coding style





EXAMPLES


X0. The following construction should be avoided:

	int num_vertices;
	int num_edges;

Use instead:

	int n; // number of edges in the graph
	int m; // number of vertices in the graph


X1. The following construction should be avoided:

	void my_function(...)
	{
		// declare the eschaton
		do_stuff_1;
		do_stuff_2;
		do_stuff_3;

		// immanentize the eschaton
		do_more_stuff_1;
		do_more_stuff_2;
		do_more_stuff_3;

		// mount the golden apple
		yet_more_stuff_1;
		yet_more_stuff_2;
		yet_more_stuff_3;
	}

Use instead:

	static void declare_the_eschaton(...)
	{
		do_stuff_1;
		do_stuff_2;
		do_stuff_3;
	}

	static void immanentize_the_eschaton(...)
	{
		do_more_stuff_1;
		do_more_stuff_2;
		do_more_stuff_3;
	}

	satic void mount_the_golden_apple(...)
	{
		yet_more_stuff_1;
		yet_more_stuff_2;
		yet_more_stuff_3;
	}

	void my_function(...)
	{
		declare_the_eschaton(...);
		immanentize_the_eschaton(...);
		mount_the_golden_apple(...);
	}


X2. The following construction should be avoided:

	for (int i = 0; i < n; i++)
	{
		do_stuff_1;
		do_stuff_2;
		do_stuff_3;
	}

Use instead:

	for (int i = 0; i < n; i++)
		do_all_stuff(..., i);

where "do_all_stuff" is a static function declared nearby.


X3. The following type of construction should be avoiced:

	// structure to store a graph as a list of vertex pairs
	struct graph {
		int number_of_vertices;
		int number_of_edges;
		int (*table_of_edges)[2];
		float *vertex_values;
		float *edge_weights;
	};

	// getters and setters

	int graph_get_num_edges(struct graph *g)
	{
		return g->number_of_edges;
	}

	int graph_get_num_vertices(struct graph *g)
	{
		return g->number_of_vertices;
	}

	int graph_edge_from(struct graph *g, int e)
	{
		return g->table_of_edges[e][0];
	}

	int graph_edge_to(struct graph *g, int e)
	{
		return g->table_of_edges[e][1];
	}

	float graph_get_vertex_value(struct graph *g, int v)
	{
		return g->vertex_values[v];
	}

	void graph_set_vertex_value(struct graph *g, int v, float x)
	{
		g->vertex_values[v] = x;
	}

	float graph_get_edge_weight(struct graph *g, int v)
	{
		return g->edge_weights[v];
	}

	void graph_set_edge_weight(struct graph *g, int v, float x)
	{
		g->edge_weights[v] = x;
	}

	// serious functions on graphs

	void graph_process(struct graph *g)
	{
		...
	}

	void my_function(...)
	{
		struct graph *g = graph_read_from_file("filename.txt");
		graph_process(g);
		int m = graph_get_num_edges(g);
		for (int i = 0; i < m; i++)
		{
			int a = graph_edge_from(g, i);
			int b = graph_edge_to(g, i);
			float v_a = graph_get_vertex_value(g, a);
			float v_b = graph_get_vertex_value(g, b);
			float w_i = exp(fabs(v_a - v_b));
			graph_set_edge_weight(g, i, w_i);
		}
	}

Use instead:

	// structure to store a graph as a list of vertex pairs
	struct graph {
		int n;       // number of vertices
		int m;       // number of edges
		int *e;      // table of edges (vertex index pairs)
		float *v;    // vertex values
		float *w;    // edge weights
	};

	// serious functions on graphs

	void graph_process(struct graph *g)
	{
		...
	}

	float compute_weight(float x, float y)
	{
		return exp(fabs(x-y));
	}

	void my_function(...)
	{
		struct graph g;
		graph_read_from_file(&g, "filename.txt");
		graph_process(&g);
		for (int i = 0; i < g.m; i++)
		{
			int *e = g->e + 2*i; // ith edge
			int a = e[0];        // ith edge origin vertex
			int b = e[1];        // ith edge destination vertex
			float v_a = g.v[a];  // value of origin vertex
			float v_b = g.v[b];  // value of destination vertex
			g.w[i] = compute_weight(v_a, v_b);
		}
	}

Notice that the second code is much simpler (good!), less general (good!), much
shorter (good!), less prone to trivial errors (good!) and just as readable.

There are some possible variations of the second code.  For example, if you
dilike taking references and prefer acessing struct fields by "->" instead of
".", or if you want to use fewer temporary variables, you can do

	void my_function(...)
	{
		struct graph g[1];
		graph_read_from_file(g, "filename.txt");
		graph_process(g);
		for (int i = 0; i < g->m; i++)
		{
			int *e = g->e + 2*i; // ith edge
			g->w[i] = compute_weight(g->v[e[0]], g->v[e[1]]);
		}
	}

or even, if you like this sort of thing (god forbid!)

	void my_function(...)
	{
		struct graph g[1];
		graph_read_from_file(g, "filename.txt");
		graph_process(g);
		for (int *e = g->e; e < g->e + 2*g->m; e += 2)
			g->w[i] = compute_weight(g->v[e[0]], g->v[e[1]]);
	}

This kind of code is simply not possible when you are forced to use getters and
setters.  Notice that since the "graph" struct is only visible on this file,
the documentation of the fields is readily available.



edit | history | main page