Cursed C

An attempt to create the worst C program ever by repeatedly abusing obscure features. See c.c for the core code. Note that there were no version restraints imposed and extensions are also included.

Rationale

it was fun
perhaps it will be educational for someone

Break down

File naming

c.c is the least descriptive name imaginable to man. It neither marks what project it belongs to or what purpose it has inside the project.

Trigraph characters

There used to be a time when there were computers widely in use without the ability to type some "normal" characters. So alternative representations were added to replace them. Basically built-in macros. Under gcc the -trigraphs flag is actually required to make them available.

The ones relevant here are:

??=    :    #
??<    :    {
??>    :    }

So for example ??=define really is just equivalent to #define.

$$$

The dollar sign is valid to be the first (and only the first) char of a symbol name. My guess is that it is so for stubborn people who prefer their variable names to start with a '$' as it has to in some other languages. Perhaps it's for ancient parsers/syntax highlighters? Create an issue if you know better.

Concatenation macro operator

## is a special operator available inside macro definitions. It will literally glue together the arguments on its 2 sides. Historically it has been used to generate symbol names.

include_next

Including is surprisingly high-level with search paths and precedence. Turns out this happens to open a hole in functionality, which is otherwise absent from C.

So, #include <myheader.h> will start trying to include the file 'myheader.h' searching for in whatever directories the linker considers to contain system header files. Then, #include "myheader.h" start from the current directory, but falls back to the <> behavior. In either case, whatever is found first to satisfy the file name is taken. Now, what happens if we do not want the first? Or rather, we must have another.

As the GNU documentation explains it through an example, say you wish to override a system header -say wrap it, but your own header depends on the original one. Standard C provides no way to resolve the situation.

For this reason #include_next was added as a widely accepted extension, which excludes the current path from the list of searchable paths to prohibit recursive inclusion.

iso646.h

From what I've seen, this perfectly standard header that tends to be relatively little known. If you know C++ and about its keyword operators, then this is what defines their equivalents in C. E.g and, or, etc.

Later down it's abused to get the address of ű:

bitand ű

Where -you guessed it- bitand is just macrod to be & at the end of the day.

To make it extra fitting, it uses the worst imaginable formatting, which is -fun fact- unreproducible with '<>'s as the preprocessor '""'s just so happens to use the exact grammar as fully featured C string literals do. Lucky us, every '<>' include is a valid '""' include with only a slight compile overhead (as the local include is path is checked first, then the system headers).

_Decimal32

It's exactly what it sounds like, a decimal fraction available in C. Here, as a GNU extension.

Dot dot dot

Not nearly as obscure as the rest, but worth including. ... signals to the compiler that any number of values may be pushed into the stack on any call to this function. Normally stdarg.h and it's va_list ("Variable Argument LIST") would be used or some pointer magic applied by hand to read said values, but here they are just left unused.

The most famous function using this feature has to be printf. In glibc it's defined as:

extern int printf (const char *__restrict __format, ...);

That's why it can handle bothprintf("%d", 1) and printf("%d %d", 1, 2).

const const const

The specifier const turns out to be infinitely stackable. In fact, it can be applied to the wrong spot too.

const const int i;

Reduces to:

const int i;

This program makes both the pointed value and the pointer constant, then reiterates it a few times for good measure.

With a bit of simplification:

   const char const * const;
//   ^          ^       ^
//   |          |       |
//   |          |  constant pointer
//   |          |
//   |  const... air? for no effect
//   |
// constant character value

Naming

Having multi-char long variable names is bloat. Naming everything a single letter is the way of the warrior.

Ű

'ű' is letter in the Hungarian alphabet. The reason it was chosen is because it is non-ascii. Don't like it? Perhaps switch it out for a '😁', which is also valid due to the same mechanic.

Haskel style semi-colons

I guess it's self-evident that C being whitespace agnostic, starting every line with ';' instead of closing them with it, is valid, however it's a rarely thought about possibility. Also, semi-colons are stackable.

Pro-tip: the next time someone proposes indenting with spaces to you, suggest indenting with semi-colons.

'a' array

When accessing an array element, a[n] is equivalent to &a + n. Note that the address is taken only figuratively, the array a decays to a pointer to the first element, so the more scientific notation is *(a + n)

Smiles

Like trigraphs, but they are too hard to type? Digraphs are for you! :>

%:    :    #
<%    :    {
%>    :    }
<:    :    [
:>    :    ]

Reverse reference

-1[$a];

No, we are not trying to access the $ath element of -1. Addition commutative, hence *(a + n) == *(n + a).

The above evaluates exactly like:

$a[-1];

Negative index

Since a + (-n) is a - n, it just werks™.

Puts call

Perhaps you noticed #include <stdio.h> missing. Don't worry, the compiler will implicitly handle it for us!

Return

Relative to Assembly, C is high-level. And in many cases, for Assembly programmers, the return statement high-level enough to be considered many ways similar to a function. For this reason they may stick to the following syntax:

return(0);

END comments

Large blocks being closed with their type/name explicitly mentioned tend to be more readable than otherwise. With C however a nice balance of size is attainable. Overshoot a bit and you get people pretending comments are significant. Wait 30 years and -looking for a more radical solution- they form a cult around DRY.

Attributes

Attributes tend to be left out from textbooks, but they are extremely cool. They provide ways to both aid the compiler (or make it shut up) and make code self-documenting.

Too bad that C initially botched the syntax. Consider the following:

Both of these specify that func will never return.

__attribute__((noreturn))

C++:

[[ noreturn ]]

But good news! With C23, C++ style attributes are being introduced.

What better way to celebrate this, than to use all syntaxes possible at the same time?

Google?

Google.

; https://google.com

URLs happen to be valid C. More precisely, they are -due to pure chance- a label plus a comment. You can only have one per protocol per function, choose wisely.

The address was chosen in particular, because hard coding google endpoints into source is as evil (or rather, no longer "don't be evil") as it gets.

?

If one asks a C programmer "can the ternary operator be used to conditionally declare a (const) variable" even if he never needed to use it, he will instinctively answer yes. Yet, no one seems to ever state this -truth be told, rare times, but- very useful feature. Instead, its main purpose is to serve as premature formatting.

While the functionality is not cursed, the poor operator itself seems to be.

Nested functions

You would be surprised to know how popular these things used to be in the past. And I'm not talking about C in particular.

In the case of nested procedures, the scope rules permit, for example, the definition of a procedure to include the definition of subsidiary procedures without concern for interference with other global procedures. This capability facilitates writing pro- cedures and often simplifies their usage.

-- David R. Hanson; Is Block Structure Necessary? (1981)

Yet, they have fallen out from the public's taste so much that C++ disallows it. Nested classes seem to be looked down upon too now days, even in bloody Java tutorials where providing a single file would be so damn much clearer and easier, but I digress.

Code in switch

A switch statement is just a fancy code block with a bunch of labels and a jump-table. So, in our:

    switch (h) {
        int a = 89;
        default: break;
    }

expression, the available value mappings (none) are compared to h, then execution jumps to default as no match was found. Now the question becomes: how is that a is defined to 89? Easy. It's not. a gets defined, but its assignment is skipped. It's because variable declarations always result in allocations at the start of the scope, no further questions asked. For all the compiler cares, int a; might as well have been written under default:. However, the value copy is skipped.

String literal concatenation

Adjacent string literals are always (compile time) concatenated. Therefore:

"Make it stop, cruel" " " "w" "o" "r" "l" "d" "!"

Costlessly becomes:

"Make it stop, cruel world!"

Why is it cursed? Ask who ever does this:

puts("Message line 1");
puts("Message line 2");

Instead of:

puts("Message line 1\n"
     "Message line 2");

Literal pointer arithmetics

A string literal "decays" into a const char pointer, therefore puts("//Please!" + 2) is just the heretic way of saying:

const char * s = "//Please!";
puts(s + 2);

Batch assignment

The assignment operator is not particularly "special" in C. It does the actual value copying, but then plainly "returns" a value of its own. Which happens to be defined as the one assigned.

"Tricks" like this and such:

    do {
        // ...
    } while(i = fun());

tend to be pretty neat.

However, for wholeness, it stacks.

i = h = j = k

Is evaluated right to left. I.e. it becomes "assign k to j; assign j to h; assign h to i;". So what value does i hold after the operation? No one knows, since k was never initialized!

Comma operator

The comma operator evaluates the expressions on both sides, then returns the value returned by the right-hand side. Again, it's stackable.

This -similar to the wise usage of assignments,- opens a little door for playing with loop conditions.

do {
    // ...
} while((fun(&a), a));

But also, if you ever wanted to make semi-colons jealous to make them love you more, here is your chance.

fun1(), fun2(), fun3(), fun4()

Do not use the comma operator.

Do not use the comma operator.

Do not use the comma operator.

[...]

Do not use the comma operator.

Do not use the comma operator.

Do not use the comma operator.

-- Repeater, https://cplusplus.com/forum/beginner/272615/

K&R arguments

The language design of C is truly genius. Brian Kernighan and Dennis Ritchie got so many things perfectly right. One clear exception is their original function declaration syntax. Not that it doesn't make sense or does not have historical reasons, but damn is it to ugly and confusing.

A reasonably cleaned up version of our main() would look like this:

void main(argc, argv, envp)
int argc;
char * * argv;
char * * envp;
{ ; }

Where our parameters are named first, then their type is specified. This original way, used to be the only way to do it. It may look like absolutely useless typing at first, but there are a few things to consider.

it roots from B syntax, which is typeless:

main(argc, argv, envp) {
    ;
}

One doesn't specify the types of your arguments until the function definition:

void main(argc, argv, envp);

void main(argc, argv, envp)
char * * argv;
int argc;
char * * envp;
{ ; }

Which implies that a) you don't have to keep your declarations and definitions in sync b) calls can compile without type checking

It blends with other syntax rules, allowing this:

void main(argc, argv, envp)
char * * argv;
int argc;
char * * envp;
{ ; }

And this:

void main(argc, argv, envp)
int argc;
char * * argv, * * envp;
{ ; }

I would like to declare the type of nothing

g() int; int; { int; }

GCC allows you have a typename on its own line. Not even a warning in sight.

Why does it allow it? Good question.

h.h

As you know, this project compiles under gcc, Still, catting out h.h should make you raise your eyebrow.

#error

Its impossible that that's the actual header included and I must have messed with the compiler, right? Well, yes, but not the way you think. If you wanted to prove me guilty of manipulating the include path by inspecting the build script, you could not:

gcc c.c -trigraphs -Wall -Wextra -Wpedantic

Tho, there is also something suspicious going on with h.h:

gcc detention/h.h -o ./h.h.gch

Where the contents of detention/h.h seems like something that would translate. That is a good lead, however detention/ is hardly a standard folder to override with.

The solution is in the .gch extension as in "GNU Compiled Header". Its what it sounds like, a header processed by gcc. After the header is compiled once, its code can be included in multiple source files without it being reprocessed. This trick can significantly speed up compile times. Well, when it works that is.

For example if it was 1 line lower in the source, it would make the operation #error out. The documentation lists 10 reasons why a compiled header could be rejected and iso646.h breaks one of those. Considering that's a standard header with around a dozen defines, you can guess how stable it is. Not to mention it fails silently. gcc provides no feedback whether it did or did not use a compiled header, let alone why. Pessimism aside, compiled headers are great as long as you can make sure they are at the top of the source (with an -include perhaps) and seem way more reliable with clang (because they have their own version too).

Empty parenthesis

How would you declare a function with no parameters? If you answered the following:

f();

Then you just found a footgun! That actually signals how the function takes... well, we don't know what it takes, so better allow anything! Take a look at g:

g() int; int; { int; }

Ignoring those ints, one could assume it takes 0 arguments, yet if you glimpse at how it's called:

g(i = h = j = k)

You will see how passing one int is valid, in fact, it doesn't even warrant a warning! This, again routes from the K&R style, where as established before, arguments where not expected to be type checked.

Sooo, is g() equivalent to g(...)? No! For one, ... required a named argument before itself until C2X. However, in the case of g() the compiler not required to perform type promotions. Also, the generated assembly is... strange. That's what tool/g().c and tool/g(...).c are here for. Here is the relevant part of their disassembly:

// @BAKE gcc -g -O0 "$@" -o "$*"                                |  // @BAKE gcc -g -O0 "$@" -o "$*" -Wall -Wpedantic

g() {                                                           |  g(...) {
    1129:       f3 0f 1e fa             endbr64                        1129:       f3 0f 1e fa             endbr64
    112d:       55                      push   %rbp                    112d:       55                      push   %rbp
    112e:       48 89 e5                mov    %rsp,%rbp               112e:       48 89 e5                mov    %rsp,%rbp
                                                                >      1131:       48 83 ec 38             sub    $0x38,%rsp
                                                                >      1135:       48 89 bd 50 ff ff ff    mov    %rdi,-0xb0(%rbp)
                                                                >      113c:       48 89 b5 58 ff ff ff    mov    %rsi,-0xa8(%rbp)
                                                                >      1143:       48 89 95 60 ff ff ff    mov    %rdx,-0xa0(%rbp)
                                                                >      114a:       48 89 8d 68 ff ff ff    mov    %rcx,-0x98(%rbp)
                                                                >      1151:       4c 89 85 70 ff ff ff    mov    %r8,-0x90(%rbp)
                                                                >      1158:       4c 89 8d 78 ff ff ff    mov    %r9,-0x88(%rbp)
                                                                >      115f:       84 c0                   test   %al,%al
                                                                >      1161:       74 20                   je     1183 <g+0x5a>
                                                                >      1163:       0f 29 45 80             movaps %xmm0,-0x80(%rbp
                                                                >      1167:       0f 29 4d 90             movaps %xmm1,-0x70(%rbp
                                                                >      116b:       0f 29 55 a0             movaps %xmm2,-0x60(%rbp
                                                                >      116f:       0f 29 5d b0             movaps %xmm3,-0x50(%rbp
                                                                >      1173:       0f 29 65 c0             movaps %xmm4,-0x40(%rbp
                                                                >      1177:       0f 29 6d d0             movaps %xmm5,-0x30(%rbp
                                                                >      117b:       0f 29 75 e0             movaps %xmm6,-0x20(%rbp
                                                                >      117f:       0f 29 7d f0             movaps %xmm7,-0x10(%rbp
        return 2;                                                          return 2;
    1131:       b8 02 00 00 00          mov    $0x2,%eax        |      1183:       b8 02 00 00 00          mov    $0x2,%eax
}                                                                  }
    1136:       5d                      pop    %rbp             |      1188:       c9                      leave
    1137:       c3                      ret                     |      1189:       c3                      ret

On the right side, g(...) actually reads a bunch of values from the stack into registers. What I think is happening is that the compiler would really like to optimize arguments being read from the stack directly, but since -O0 was on, it was not allowed to make any assumptions about, how g will be called, hence it loads as much as it can.

(If you know better, please correct me!)

Duff()

As you can see case statements work inside nested blocks of the corresponding switch. Which is funny by itself, but what that allows us to do is something called a "Duff's Device". This allows us to force loop unrolling within the bounds of C.

I can explain it best with an incremental example. Consider:

do {
    perform_stuff();
} while (counter++ < repeat_count);

Here, after every call to perform_stuff() we have to perform a jump instruction to the top of the loop for every repetition. With the standards of modern software, the cost is small, however there are circumstances where it could be a legitimate performance concern. An optimization would be:

do {
    perform_stuff();
    perform_stuff();
    perform_stuff();
    perform_stuff();
    counter += 4;
} while (counter < repeat_count);

Now, we only need a jump after every 4 calls! Great, except if repeat_count is not a multiple of 4, we actually call perform_stuff() the wrong number of times... This is where Duff's Device comes into the picture.

switch (repeat_count % 4) {
    case 0: do { perform_stuff(); counter++;
    case 1:      perform_stuff(); counter++;
    case 2:      perform_stuff(); counter++;
    case 3:      perform_stuff(); counter++;
    } while (counter < repeat_count);
}

The switch "eats away" some number of calls, guaranteeing (repeat_count - count) % 4 == 0 and peform_stuff() being called the desired number of times. (Unless it was 0, that's an edge case we do not account for in the example for simplicity.) In our duff() we apply this trick to copying a string.

It should also be mentioned, that due to modern compiler optimizations, Duff's Device may or may not result in actual performance gains (see here). You may read Duff's own thoughts regarding it here.

_sum(int argc, ...)

Back to variadic functions and their quirks. Who cares if they nuke type safety? Let's talk about the real problem: ergonomics.

Any variadic function has to have some way to tell how many arguments it was given. In practice this boils down to one of the following techniques:

pass the number of extra arguments as a named argument
encode the number of arguments in a named argument (e.g. format string)
choose a termination value for the extra arguments (e.g NULL, passed as the last argument)

Sometimes -when trying to create comfortable interfaces- the first option seems perfect, if not for passing this count by hand.

Thing is, the preprocessor can save our day. Take a look inside auto_vararg.h.

Macros can be variadic too. __VA_ARGS__ is a standard macro, it simply expands to the variadic arguments.

PP_128TH_ARG always expands to its 128th argument.

PP_RSEQ_N expands to 128 numbers in descending order, 0 inclusive.

If we pass PP_RSEQ_N to PP_128TH_ARG it will return 0 to us. If we were to place exactly one argument before PP_RSEQ_N, all of its values would be shifted to the right, resulting in 1 becoming the 128th argument of PP_128TH_ARG. More generally, passing N arguments before PP_RSEQ_N will offset it by N, shifting the correct value to the 128th argument of PP_128TH_ARG.

In effect, we have just automated passing the argument count!

Argument counter hack credits:

Challenge

Try to make the project worse

Completionists

SurmanPP

agvxov/cursed_c