Best Documented

Jens Schweikhardt

Judges' comments:

To use:


./prog n

where n is a base 16 number of any size


./prog 19

Selected Judges Remarks:

This code is clean. When you compile with all warnings enabled, such as with clang:

cc -Weverything -pedantic -std=c11 -Dtyp=uint64_t -O3 prog.c -o prog

The code compiles without warnings on the systems that we tested!

Even more, the code was 100% clean when we ran it against various static source checkers.

This tool references a problem that David I. Bell once described as having one of the largest “yummo quotients” in number theory:

                     complexity of the solution
yummo quotient = -----------------------------------
                 complexity of the problem statement

Erdős privately told one of the IOCCC judges:

“Solving the Generalized Riemann hypothesis would be a good warmup exercise for someone to get ready to begin to work on the Collatz conjecture.”

You may explore this famous conjecture using this entry:

./prog 2410
./prog abfff
./prog 27f8cebf
./prog 246f8fddf
./prog 2e95ab51ffea9de
./prog e6a5c22fd7bde9f
./prog 1b7dd73a937485bf
./prog 7d3237680d190a77e53751b
./prog 302ab3d052fb87c06228d249581be0e4

When you first look at the source, the code looks fairly straightforward. But look again. Like the Collatz conjecture, simplicity is deceptive! Oh, and the variable names? They are not simple single letter variables, they are names of various Proteinogenic amino acid: which is yet another simple building block that can be used to build some very complex structures. :-)

p.s. We appreciated that apart from a few powers of 2, the source code is magic number free.

Author’s comments:

The TL;DR version

This is the cleanest program ever submitted. If for some input it enters a non-terminating loop or runs out of memory you will be insta-famous. It’s a bit like a lottery without the need to buy a ticket and you can play as often as you like. One notable historic person has even offered some prize money.

The program illustrates the bloat caused by adherence to too many rules, each of which may sound sane in isolation, but in their entirety lead to an obfuscated, hard to read and understand monster.

What this program does.

The program tests whether a given natural number satisfies the Collatz Conjecture:

Take any natural number n. If n is even, divide it by 2 to get n/2. If n is odd, multiply it by 3 and add 1 to obtain 3n + 1. Repeat the process indefinitely. The conjecture is that no matter what number you start with, you eventually reach 1.

Paul Erdős said about the Collatz conjecture: “Mathematics may not be ready for such problems.” He also offered $500 for its solution.

For example, the sequence of numbers for n = 6 is

6, 3, 10, 5, 16, 8, 4, 2, 1.

Continuing past one leads to the cycle 1, 4, 2, 1, 4, 2, 1, … Interesting factoid: if you allow negative start values, there are a few more cycles, each of different length:

−1, −2, −1

−5, −14, −7, −20, −10, −5

−17, −50, −25, −74, −37, −110, −55, −164, −82, −41, −122, −61, −182, −91, −272, −136, −68, −34, −17

The program computes the sequence for a given positive natural number and stops at 1. The number n is specified in hexadecimal (without 0x prefix) as the first argument. The program prints the given number in zero-padded hex and each iteration along with a line count in decimal. The example above looks like this (compiled with 64 bit word size):

$ ./prog 6
0000000000000003 1
000000000000000A 2
0000000000000005 3
0000000000000010 4
0000000000000008 5
0000000000000004 6
0000000000000002 7
0000000000000001 8

The size of n is only limited by the argument size limit of your shell/OS (the program implements arbitrary size bignums). To query this on your POSIX system, run

$ getconf ARG_MAX

which reports the limit in bytes, here 256KB. This allows to test whether a Googol, which is 10100, satisfies the conjecture. But what is a googol in hex? Fear not, bc(1) to the rescue:

$ printf 'obase=16;10^100\n' | bc
$ ./prog 1249AD2594C37CEB0B2784C4CE0BF38ACE408E211A7CAAB24308A82E8F10000000000000000000000000 \
| less

Observe how the first 100 iterations melt the zeros to the right. Can you explain why?

For a given n the program behavior is one of the following 3:

  1. The sequence stops at 1. No fame. No money. Thanks for playing. Computational mathematicians have tested all n < 4FFDD776055A0000 (~1018) so don’t try anything less than that.
  2. The chosen n leads to a sequence with ever bigger numbers, so that eventually the bignum cannot be stored in memory. If this happens, the program outputs laugh (more likely) or throw up (less likely) and stops. You might have found a number for which the sequence diverges. If confirmed, this disproves the conjecture.
  3. The chosen n leads to a cycle not including 1 (i.e. runs forever, repeating the same sequence over and over). You have disproved the conjecture and should certainly submit a paper to the nearest mathematical journal.

Design objectives

I gave myself the following objectives. Like in real world programming, not all of them can be met 100%. Think of them as a multidimensional continuum, where trade-offs have to be made.

  1. No arbitrary limits on the input number. 64 bits might be enough for everybody, but is not enough for exploring new Collatz-territory. Thus bignums are required.
  2. Ultra-portable. Must run on C89 systems and self-adapt to C99 and C11 features like exact width types.
  3. Super-efficient. Must be able to run with the widest type as the base of the bignums. Make the user select the widest type supported by the implementation and then crunch away. If a non-standard 128bit type is available, it should be usable.
  4. Pentagon level lint cleanliness and MISRA compliance.
  5. Easy to understand, self-documenting clear code. A joy for maintenance programmers. The epitome of best practice demonstration for all future textbooks on C.

Program Obfuscation

In the following I address all the tests as specified by your honors in the guidelines.

Look at the original source

Using one letter identifiers is quite old. I decided to use TLI (three letter identifiers). Not the random kind, rather with a connection to the meaning of life, the universe and the rest. Enter amino acids! Among the myriad of possible amino acids there are 23 from which proteins are built. In biochemistry, each is assigned a TLI (see Proteinogenic amino acid):

ala Alanine
cys Cysteine
asp Aspartic acid
glu Glutamic acid
phe Phenylalanine
gly Glycine
his Histidine
ile Isoleucine
lys Lysine
leu Leucine
met Methionine
asn Asparagine
pyl Pyrrolysine
pro Proline
gln Glutamine
arg Arginine
ser Serine
thr Threonine
sec Selenocysteine
val Valine
trp Tryptophan
tyr Tyrosine
asx Asparagine or Aspartic acid
glx Glutamic acid or Glutamine
xle Leucine or Isoleucine
unk Unknown

My own research results complete this list (not yet in Wikipedia due to the rule “No original research”):

and Androgynine
xor Xenoricine
not Notanamine
tla Triletramine

Interestingly, the TLI are the perfect mnemonics for C language source. For example, met is “Main’s Exit Type” (int), ala is “A Large Algebraic” (bignum), ile an “Incremented Local Entity” (index counter), gly means “Grow Larger memorY”, gln is “Grown Larger Now” (after realloc), not is a “Not Overflowing Type” (recursive!), unk is the “UNit (Known as 1)”, trp is the “Tabula Rasa Product” (zero), phe is “Print HEx” and so on.

Convert ANSI tri-graphs to ASCII


C pre-process the source ignoring ‘#include’ lines

Wow, an identity operation (except for the <stdint.h> and EOF + __STDC__ trivialities). Did you gain any insight through this?

C pre-process the source ignoring ‘#define’ and ‘#include’ lines

#define? Which #define directives? How many 4K source files do you see that neither use a single #define directive nor abuse the build file? Even though the “no #define” rule I submitted myself to made it hard, I could use __LINE__ and stdio macros EOF, L_tmpnam, BUFSIZ, FILENAME_MAX, TMP_MAX to obfuscate at least something.

Run it through a C beautifier

Another identity operation. I’ve done it for you already. Use this with FreeBSD indent:

$ cat
-bad   /* blank line after decls */
-bap   /* blank line after functions */
-br    /* braces on if line */
-i8    /* indent */
-l999  /* line length */
-npcs  /* no space after function call names */
-npsl  /* don't break procedure type */
-ut    /* use tabs */
-ce    /* cuddle else */
-nip   /* no parameter indentation */
-di1   /* declaration indent */

It looks like a perfect program should:

The program contains not a single magic number (only 0 and powers of 2, each power from 1 to 512) which are obvious to competent software engineers). How many programs have that property? The check for __STDC_VERSION__ was a bit tough to arrive at, since 199901L has too many bits set. But I realized that I only needed a number larger than 199409 and less than 199901. 199680 has only 4 bits set and writing it as (256 + 128 + 4 + 2) * 512 minimizes the character count. That’s what judges get when they don’t like programs that are longer than they need to be.

Examine the algorithm

Spoilers ahead. Duh!

Pseudocode, with comments matching those in the C source:

/* run */
if (non-NULL and nonempty argv[1]) {
   n = convert(argv[1])
   print n                       /* 2hx */
   while (n != 1) {              /* one */
      if (n is odd) {            /* odd */
         m = deep copy of n      /* cpy */
         n <<= 1                 /* shl */
         n += m                  /* add */
         increment n             /* inc */
      } else {                   /* eve */
         n >>= 1                 /* shr */
      print n                    /* 2hx */

Bignums are represented as the two member structs

typedef struct {
     size_t places;   /* number of places in base 2<sup>8*sizeof(type)</sup> */
     type  *number;   /* dynamically allocated memory for number */
} bignum

Compile it (with flags to enable all warnings)

Here’s where the program shines brighter than a gamma-ray burst!

I challenge you to throw all kinds of compilers, lints, checkers and analyzers at my program and make it find the slightest of issues.

Is clang -Wall -Wextra -Weverything -Dtyp=uint32_t prog.c all you can do? Clang has implemented a new warning? Bring it on!

Execute it

May you win the jackpot!

For least surprising results, the execution character set should be ASCII. The program computes four bits from every character in argv[1] and interprets them as a hex digit. On ASCII the characters a-f, A-F and 0-9 are converted as expected.

This means however, that the program accepts any string, turns it into a starting number (which is output as the first line), and starts crunching. Nothing stops you from executing

$ ./prog "$(cat prog.c)"            # Kind of quine?
$ ./prog "$(cat rules guidelines)"  # A jackpot? Maybe next year...
$ ./prog "$(cat /bin/ls)"           # Number cut short at first NUL byte.

In a certain way, the program is character set and encoding agnostic.

Assumptions made

While the program works best when bytes/characters are octets and the number of bits in a type is sizeof(typ) << 3, it will work correctly on 24bit or 36bit systems with 9 bits/byte, or systems where sizeof(typ) is 1 even for int and so on. On such systems, it will only use 8 * sizeof(typ) bits per place. It does not work when CHAR_BIT <= 7.

Results of various checkers


The open source static checker cppcheck checks various problems with respect to style, performance, portability and general fishiness. To enable all checks, run

$ cppcheck --enable=all --force -I/usr/include -Dtyp=uint32_t prog.c

No issues found.


The open source static checker flawfinder checks various problems with respect to security like buffer overflows, function arguments to known troublemaker functions and more. It doesn’t need to pre-process code, so can be run without the typ macro being defined:

$ flawfinder prog.c
Flawfinder version 1.31, (C) 2001-2014 David A. Wheeler.
Number of rules (primarily dangerous function names) in C/C++ ruleset: 169
Examining prog.c
No hits found.


The open source dynamic checker valgrind runs an executable and verifies all memory accesses and (de)allocations for proper bounds and memory leaks. Since the program frees all memory in all possible execution paths, even when it must bail out, valgrind should be happy. An early version of my program however reported this:

$ valgrind --leak-check=full --show-leak-kinds=all ./prog 6
==14615== HEAP SUMMARY:
==14615==     in use at exit: 4,096 bytes in 1 blocks
==14615==   total heap usage: 4 allocs, 3 frees, 4,108 bytes allocated
==14615== 4,096 bytes in 1 blocks are still reachable in loss record 1 of 1
==14615==    at 0x4C236C0: malloc (in /usr/local/lib/valgrind/
==14615==    by 0x4F62175: ??? (in /lib/
==14615==    by 0x4F62073: ??? (in /lib/
==14615==    by 0x4F514EE: ??? (in /lib/
==14615==    by 0x4F51265: vfprintf_l (in /lib/
==14615==    by 0x4F3E001: printf (in /lib/
==14615==    by 0x40101E: phe (in ./prog)
==14615==    by 0x400A9F: main (in ./prog)
==14615== LEAK SUMMARY:
==14615==    definitely lost: 0 bytes in 0 blocks
==14615==    indirectly lost: 0 bytes in 0 blocks
==14615==      possibly lost: 0 bytes in 0 blocks
==14615==    still reachable: 4,096 bytes in 1 blocks
==14615==         suppressed: 0 bytes in 0 blocks
==14615== For counts of detected and suppressed errors, rerun with: -v
==14615== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)

From which I concluded that printf allocated a single 4K block for which no matching free existed. But how to free memory allocated deep down in the guts of the Standard I/O library? After some serious head scratching it hit me. The only chance I have is telling the system I no longer want to do I/O, maybe then it would free that buffer. A reading of C99, “The fclose function”, was encouraging:

Whether or not the call succeeds, the stream is disassociated from the file and any buffer set by the setbuf or setvbuf function is disassociated from the stream (and deallocated if it was automatically allocated).

So I fclose(stdout) before returning and now:

$ valgrind --leak-check=full --show-leak-kinds=all ./prog 6
==14571== Memcheck, a memory error detector
==14571== Copyright (C) 2002-2013, and GNU GPL'd, by Julian Seward et al.
==14571== Using Valgrind-3.10.0 and LibVEX; rerun with -h for copyright info
==14571== Command: ./prog 6
0000000000000003 1
000000000000000A 2
0000000000000005 3
0000000000000010 4
0000000000000008 5
0000000000000004 6
0000000000000002 7
0000000000000001 8
==14571== HEAP SUMMARY:
==14571==     in use at exit: 0 bytes in 0 blocks
==14571==   total heap usage: 4 allocs, 4 frees, 4,120 bytes allocated
==14571== All heap blocks were freed -- no leaks are possible
==14571== For counts of detected and suppressed errors, rerun with: -v
==14571== ERROR SUMMARY: 0 errors from 0 contexts (suppressed: 0 from 0)

A squeaky clean valgrind result!


FlexeLint is a commercial lint tool by Gimpel Software. It supports nearly a thousand checks, broadly categorized into 4 levels,

  1. Syntax errors only (messages 1 to 399)
  2. Warnings (messages 400 to 699)
  3. Informational messages (700 to 899)
  4. Elective notes (900 to 1000 and > 9000)

There is an on-line demonstrator you can use for checking your C programs and I highly recommend trying it. For a start, paste the well know first program in the form and press “Analyze Code”. Note the FlexeLint configuration options in comments (no space between /* and lint).

/*lint -w4            turn on everything */
/*lint +esym(534,*)   no demonstrator defaults */
/*lint -e966          indirectly included header file not used */
/*lint -esym(9058,_*) tag not used outside typedef */
/*lint -misra(2)      */

#include <stdio.h>
int main(void)
    printf("hello, world\n");
    return 0;

Possible checks include those for MISRA 2004 compliance verification. The Motor Industry Software Reliability Association has produced a list of 121 required and 20 advisory rules for C programming. I am proud to report that my program fulfills almost all rules. To assess the few exceptions, one has to understand that MISRA rules are geared towards embedded systems used in the automotive industry. That’s why features like malloc and printf are right out. But that’s too restrictive for an IOCCC hosted application, so I ignored these:

I started out with enabling all messages using FlexeLint’s -w4 option and disabling all noise from system headers. Then I dealt with the remaining messages by addressing them or suppressing them in such a way that the set of suppressions was minimal. At the end of the day, this is what remained:

//  === Tested with FlexeLint 9.00L on FreeBSD 11 ===
//  Compiler:
//  "4.2.1 Compatible FreeBSD Clang 3.6.1 (tags/RELEASE_361/final 237755)"

-w4                          // Enable maximum pickiness
-passes(6)                   // Recommended in FlexeLint manual for torture testing
+fsc                         // Make string literals have type const char*
+fnr                         // ptr-returning functions may return NULL
-strong(AJX,and,pro,ser,ala) // strong types
-strong(AcJcm,thr)           // thr is strong, but most arithmetic is okay
-strong(AarJemorX,met)       // met is strong, but some use is okay
+libclass(angle)             // All <headers> are system headers
-wlib(1)                     // Only errors for system headers
-i/usr/include               // System headers
-i/usr/local/lib/supp/ansi   // Comes with FlexeLint
-d__GNUC__=4                 // Pretend...
-d__GNUC_MINOR__=2           // I'm...
-d__GNUC_PATCHLEVEL__=1      // someone else.
-d__builtin_va_list=char*    // Fake this compiler extension
-d__inline=                  // Delete this compiler extension
-d__attribute__(x)=          // Delete this compiler extension
+e900                        // Announce number of messages produced

//                   MISRA 2004 checking

/usr/local/lib/supp/lnt/au-misra2.lnt    // Enable all MISRA checking
-elib(960)       // Don't check FreeBSD system headers for required rules
-elib(961)       // Don't check FreeBSD system headers for advisory rules
-e829            // stdio.h used
-e522            // MISRA 14.2 Highest operation, lacks side-effects
-esym(960,17.4)  // pointer arithmetic other than array indexing used
-e971            // char without signed/unsigned
-e974            // stack usage report
// Use verboten functions, Req. Rules 20.4 malloc() et al., 20.11, exit()

-e911            // implicit promotion of smaller than int to int
-e921            // cast from integral to integral
-e925            // cast from pointer to pointer
-e958            // padding required in struct

+e9???           // Enable all 9xxx informational messages, except:
-e9079           // conversion from pointer to void to pointer to other type
-e9087           // cast performed between a pointer to object type and a
                 // pointer to a different object type
-e9141           // global declaration of symbol

//    Messages due to code in FreeBSD headers.
-dlint           // The "lint" macro is tested in x86/_types.h
-elib(537)       // Repeated include file
-e793            // external identifiers > 6 chars
-e935            // (unsigned) int within struct (actually the size_t)
-estring(960,_*) // Could be defined at block scope
-e964            // Header file not directly used
-e966            // Indirectly included header file not used
-elib(970)       // int outside typedef
-esym(9003,_*)   // could be defined at block scope
-elib(9047)      // FILE pointer dereferenced
-esym(9058,__*)  // tag not used outside typedef
-e9092           // NULL does not expand to a pointer (but plain 0)

Compulsory obfuscations

Which of the rules cause which obfuscation?

MISRA 6.1, “The plain char type shall be used only for the storage and use of character values.” This forbids using character constants in expressions other than assignments to char objects. A consequence is that printing digits with '0' + digit is not allowed (even though '0' is technically an int!) so I am forced to output hex digits with

printf("%c", (met)tyr + 32 + 16 + ((8 + EOF) * ((met)tyr / (8 + 2))));

because of the “no magic numbers other than powers of two” rule. How is this an improvement over printf("%c", '0' + tyr + 7 * (tyr / 10), MISRA?

MISRA 6.3 “typedefs that indicate size and signedness should be used in place of the basic types.” Well, if you can’t infer the size and signedness from typedef int met, you’re not a real C programmer.

MISRA 10.5, “If the bitwise operators ~ and << are applied to an operand of underlying type unsigned char or unsigned short, the result shall be immediately cast to the underlying type of the operand.” Because the program must work for any unsigned type chosen for the typ macro, including the narrow types enumerated in the rule, a lot of redundant casting ensues. It gets worse with the next rule…

MISRA 12.1, “Limited dependence should be placed on the C operator precedence rules in expressions.” This requires parentheses for almost all expressions involving more than one operator, especially those for which a cast is required, leading to hard to understand expressions such as

const ser glx = (ser)((asx > (ser)64u) ? (ser)((ser)asx + (ser)8u + (ser)1u) : (ser)asx);
not.not[leu] = (and)((and)not.not[leu] | (and)(((and)glx % (and)16u) << (and)lys));

MISRA 14.7, “A function shall have a single point of exit at the end of the function.” Sigh. Since I must use eloquent prototypes (8.1, 16.3, 16.4) and static functions (8.10, 8.11), I can only use a few functions. Everything usually written with an early return now cause another useless indent level. The first three if statements in main cause a silly 24 character indent. The maximum indent is forced to 8, which is way too high for a sane function.

MISRA 16.10, “If a function returns error information, then that error information shall be tested.” A cast to void would draw a lint warning, so I use the printf result in expressions,

lys -= 4 * printf("%c", (met)tyr + 32 + 16 + ((8 + EOF) * ((met)tyr / (8 + 2))));
val += printf("\n") / ((__LINE__ * L_tmpnam) + TMP_MAX);
val -= (printf(" %d\n", val) > BUFSIZ) ? FILENAME_MAX : EOF;

which are, in the absence of errors, equivalent to

lys -= 4;

MISRA 13.1, “Assignment operators shall not be used in expressions that yield a Boolean value.” Forbids the idiomatic if ((p = malloc(n)) == NULL) and requires separate statements, in other words, bloat.

MISRA 16.7, “A pointer parameter in a function prototype should be declared as pointer to const if the pointer is not used to modify the addressed object.” in conjunction with lint’s “Note 952: Parameter could be declared const” causes const-poisoning for all functions,

static void phe(const ala not);
static void gly(ala *const not, const and his);
met main(met val, const pro *const his[]);

and quite a number of automatic variables.

Why the “laugh” and “throw up” messages?

The guidelines state “We like programs that: make us laugh and/or throw up :-) (humor really helps!)”

If you have a 128 bit type

If your compiler supports a 128 bit wide type (e.g. clang, gcc) then you can use it via the typ macro:

clang -Dtyp=__uint128_t -o prog prog.c

Indeed, the program can use (and lints clean for) all of

clang -Dtyp=uint8_t  -o prog8  prog.c
clang -Dtyp=uint16_t -o prog16 prog.c
clang -Dtyp=uint32_t -o prog32 prog.c
clang -Dtyp=uint64_t -o prog64 prog.c
clang -std=c89 -Dtyp="unsigned long" -o prog89 prog.c

This works because the program has no need for corresponding SCN or PRI macros to do the scanning and printing of variables of these types. With a 128 bit type the program can represent numbers up to 340282366920938463463374607431768211457 (3.4 * 1039) as a single “place”, enough to explore yet untested numbers for the rest of your life.

How big a number can I test with 4 GB of RAM?

The number of hex digits in the start number is limited by ARG_MAX, probably minus some overhead for the environment variables (use env(1) to trim your environment). In bits this leaves you with 24 * ARG_MAX.

The algorithm requires two bignums in memory for addition. If you hit a divergent number, this will cause out-of-memory (“laugh”) somewhere near 216,000,000,000 (4GB/2 * 8bits/byte). Sadly, I don’t have a test case :-)

How long would it take to overflow the Not Overflowing Type? Lets assume we’re processing 2GB numbers. The program copies, shifts by one, adds and increments 2GB long bit strings, each time completely thrashing the data cache – at all levels. We have a fast machine that can do an iteration in one second, on average. To make things easy, we round up 3n + 1 to 4n, and assume we never need to divide by two (which is quite optimistic). Then each iteration shifts left by 2. This would take 8 billion seconds, or about 250 years. If you want to get famous, you better remove some of the RAM, use an ancient box, or reduce available memory resources with the ulimit(1) built-in of your shell.

For all intents and purposes, the Not Overflowing Type keeps the promise!

Creative Commons License

© Copyright 1984-2016, Leo Broukhis, Simon Cooper, Landon Curt Noll - All rights reserved
This work is licensed under a Creative Commons Attribution-ShareAlike 3.0 Unported License.