Prepare for the winners of **IOCCC28**. > However these are not the 2024 droids you are looking for. :-) The HTML menu (both JavaScript and non-JavaScript) has been updated. The `Status -> Rules and guidelines' have been replaced with two separate menus `Status -> IOCCC rules` and `Status -> IOCCC guidelines`. Added a new top level `social.html` page: built from a new `social.md`. Sorted rules for building specific web pages in the top level `Makefile`. Added `make social` to make the `social.html` page from `social.md`. The `bin/gen-top-html.sh` now is able to process a argument of **social** to process the new top level `social.html` page. Changed `bin/gen-top-html.sh` VERSION from "2.0.0 2025-03-13" to "2.0.1 2025-07-25". Below `About -> Home page` a new `About -> IOCCC social media` menu item as added. Because the menu change touched every HTML file, we changed the general copyright date range for HTML files from "1984-2024" to "1984-2025" due to the extensive changes made for **IOCCC28**. Those markdown files that were modified in 2025 had their copyright date range also changed from "1984-2024" to "1984-2025". Rebuilt tarballs and sitemap due to the extensive HTML rebuilding. Ran `make update` to perform the above.
To build:
make all
There are two alternate versions, one that has an increased buffer size and one
that uses fgets(); neither are completely functional. See below section Bugs
and (mis)features for details and the Alternate
code section for how to use.
Bugs and (Mis)features:
The current status of this entry is:
STATUS: uses gets() - change to fgets() if possible
For more detailed information see 1994/schnitzi in bugs.html.
To use:
./schnitzi < textfile
Try:
./try.sh
Alternate code:
These alternate versions are for those who wish to help with getting the entry
to use fgets(3). See the
FAQ on "gets and fgets"
for why this is desired.
The first version, schnitzi.alt.c, uses fgets(3) but when fed its own code
it will generate code that still uses gets(3). What's more that code will not
compile. Other functionality of the alt version works fine.
The second version, schnitzi.alt2.c, has an increased buffer size but when fed
its own source code it will generate code that, although it can compile but, it
has the original buffer size. This is due to a comment which is explained in
more detail in 1994/schnitzi in bugs.html.
Alternate build:
make alt
Alternate use:
Use schnitzi.alt and schnitzi.alt2 as you would schnitzi above except that
feeding the source code to it will not work right.
Alternate try:
./try.alt.sh
Judges' remarks:
Try:
./schnitzi < schnitzi.c
Compare the output of the program with its source. Sure the output
is different, but there is more than meets the diff's eye. Can you
find out why?
You might also wish to redirect the output to another file, compile it and compare how it works to the original code.
NOTICE to those who wish for a greater challenge:
If you want a greater challenge, don't read any further: just try to understand the program via the source.
If you get stuck, come back and read below for additional hints and information.
Author's remarks:
This is a program which takes any text file as input and 'flips' the input file along a diagonal running from the top left corner down to the bottom right. In other words, the first row in the file ends up written down the first column of the result, the second row is written down the second column, etc. The program is run by redirecting a file into the executable on the command line; for instance, for the file 'info', the program would be run as follows:
./schnitzi < info
The program generates interesting results when its source file is used as input:
./schnitzi < schnitzi.c
The result of this command is a program which does the exact same thing -- in other words, both the program and the "flipped" version of itself work identically (you'll need to redirect the output of the above command into a separate file and compile it). You might notice some differences in the flipped version, though. First off, a secret message becomes visible that was not visible in the original program. Also, much of the code shows up in different places in the flipped program than it appeared in the first program. My first version of this program was perfectly symmetrical along the diagonal, but I later found out there were interesting ways to break the symmetry. The best way to see this is to load both the original and flipped versions of the program into an editor and switch back and forth between them rapidly.
C tokens longer than a character (such as main()) proved difficult
to use in both the original and flipped versions (after flipping,
they show up as a string of single letters on successive lines).
However, I found that it was possible to get around this through
a few well-placed comments. An earlier version of this program
used nearly every single-character token in the "shared" area;
however, it proved to be too large for contest guidelines (although
the code itself is only about 300 characters, the extra whitespace
needed to pad it put it over the limit -- it nearly filled a 79 by
79 area, adding up to over 4000 total characters).