Best solved puzzle

Shinichiro Hamaji
shinichiro.hamaji@gmail.com

bsoup

Judges' comments:

To build:

make hamaji

To run:

./hamaji < a_nonogram_file

Try:

./hamaji < dragon.nono
./hamaji < random.nono
./hamaji < soccer.nono
./hamaji < codegolf.nono
./hamaji < face.nono
./hamaji < penguin.nono
./hamaji < samurai.nono

Selected Judges Remarks:

This program solves “Nonogram”, which also known as “Paint by Numbers”:

http://en.wikipedia.org/wiki/Nonogram

This entry roar as the Year of the Dragon: A year considered some to be good luck. Well it was certainly true that the IOCCC was lucky to get such a fine submission!

The source code is a solution to a Paint by Numbers puzzle when you paint by C. Try your hand at solving a few of the other .nono files and then compare your solution with this program!

NOTE: NONOGRAM ® is a registered Trademark and is used with permission. See http://nonogram.com/ for more information.

Author’s comments:

The format of the input files are as same as this site:

http://codegolf.com/paint-by-numbers

For example, for a simple problem which shows a glider of conway’s lifegame.

the input should be

The width and the height of an input puzzle should be less than or equal to 60.

When this solver finds the problem has no solution, it prints “invalid”. This means the problem is indeed broken. For example, this happens for an input like

2
-
2

When this solver cannot find a solution, it prints “failed”. This failure can happen when the problem has more than one solution or the solver cannot solve the input problem. Here is an example input of the former:

There are two solutions:

and

Detail

Solving Nonogram is known NP-complete, so we cannot make a solver which solves all puzzles. However, we can write a solver which solves most problems solvable by humans. There are a bunch of ways to write such programs. The two most obvious approaches are:

Check all possible solutions with backtracking.
Implement all techniques which humans use.

The former, brute-force approach is easy to implement. However, it works only for smaller problems (width and height should be at most 30 or so). Similarly, using generic probabilistic metaheuristics such as simulated annealing doesn’t work for big problems.

http://ccl.northwestern.edu/netlogo/models/community/Nonogram

By its definition, the latter works for all human-solvable problems. The issue of the latter approach is this kind of programs tends to be lengthy. This approach still requires some recursions. The most difficult problems which are considered human-solvable may require some “guesses”. For this kind of problems, we need to assume a cell is a space or a box, go ahead some steps with simple methods, mark the tried cell as a box (or a space if we assumed the tried cell is a box) if a contradiction is found.

My program uses an approach like “Set based solver” introduced in http://www.haskell.org/haskellwiki/Nonogram . This should be equivalent to the combinations of simple methods (no advanced reasoning with guess). My solver also runs non-recursive guess for each undetermined cells. So, my program should be able to solve most human-solvable problems (assuming humans cannot execute deeper recursions).

Obfuscation

First of all, as you see, my program looks like a nonogram puzzle which results a dragon. This puzzle is “dragon.nono” in this directory. Note that whitespaces in character literals, a string literal, and around the last comment are considered as a boxed cell.

My program is based on bit operations for speed. Many values are stored using negative logic (0 means on and 1 means off) to utilize the initial value of global variables. Thus, it might be difficult to figure out how it works.

My program is decently shortened, because the size of source code is limited by IOCCC’s rule (2048 non-whitespace characters) and a lot of characters are just wasted for numbers around the picture.

The three requirements (dragon, speed, and size) naturally ended up with the well obfuscated code.

Portability

This program uses two C99 features. One line comments and “long long”. GCC’s “-ansi -pedantic” check should pass:

$ gcc -ansi -pedantic -std=c99 hamaji.c

I checked my program with gcc-4.6.2 on linux, llvm-gcc-4.2 on mac, and clang-3.0 on mac.

Note that the behavior of scanf differs on linux and mac, but this program supports both semantics.

This program should not depend on sizeof(int), sizeof(void*), ASCII, memory layout, undefined evaluation order (e.g., a[i++]=i), etc. This program won’t work if the size of long long literal is less than 8. By grepping the source code of gcc

grep LONG_LONG_TYPE_SIZE gcc/config/*/*

I found long long may be 32bit integers only on AVR.

A few more

This code is considered as a new year’s postcard in Japan. We chose the dragon because 2012 is the year of dragon.

http://en.wikipedia.org/wiki/Japanese_New_Year#Postcards

The second author wrote the picture, and the first author did everything else.

The dragon.nono, samurai.nono, and penguin.nono was written by the second author. The soccer.nono was from a wikipedia entry. This file is important because it requires some guesses. The random.nono file was randomly generated. The codegolf.nono and face.nono have borrowed from this site:

http://codegolf.com/paint-by-numbers

You can reproduce dragon.nono from hamaji.c using c2nono.pl:

perl c2nono.pl hamaji.c | ./hamaji