CPP Abuse:
To build:
To run
Important! This program, if it compiles at all, is mis-compiled by many
compilers due to compiler bugs. It could be the "least likely
to compile and execute" in conjunction with preprocessor abuse.
When run, it is likely to uncover bugs in your system headers.
The program is run without arguments. If your compiler is buggy,
the most likely result is no output and/or an exit status of 1.
If Amendment One headers are missing you will see
<iso646.h>:
gcc: /usr/include/iso646.h: No such file or directory
gcc: No input files specified
<wctype.h>:
gcc: /usr/include/wctype.h: No such file or directory
gcc: No input files specified
<wchar.h>:
gcc: /usr/include/wchar.h: No such file or directory
gcc: No input files specified
Bug your vendor for an upgrade.
The format of the info file is straightforward and described in the file
itself. Just load it in your favorite editor.
Requirements
------------
Gcc (the GNU C compiler) must be available at runtime. It is assumed
that your C implementation keeps headers as files in the /usr/include
directory. The info file must be readable and reside in the current
working directory. The current working directory must be writable in
order to create a temporary file (which is removed upon program
termination). In case you don't have gcc at runtime, not all is lost
if your compiler or preprocessor can produce a list of defined macros
in the format output by gcc -dM, i.e. lines of the form
#define MACRO value
Edit the source at line 55 in this case.
Background
----------
ISO C has a very strict idea of visibility of identifiers. All possible
macros are explicitly enumerated. In a compliant implementation no other
macros can be defined, because you could write strictly conforming
programs that may fail to compile due to syntax errors: suppose
<stdio.h> defines PIPE_BUF, then the conforming
#include <assert.h>
#include <stdio.h> /* <- or where the bogus macro is defined */
#include <string.h>
#define STR(x) #x
#define XSTR(x) STR(x)
int main (void)
{
int PIPE_BUF = 0;
assert (strcmp ("PIPE_BUF", XSTR (PIPE_BUF)) == 0);
return 0;
}
is expected to compile and meet the assertion. If it does not, your
compiler compiles some other language than ISO C.
Why I think my program is obfuscated
------------------------------------
Not only is the program's purpose a thorough standard conformance
test. The source itself is a beast and has uncovered bugs in many
C compilers.
1. The token sequence after #include can consist of pp-tokens and
"are processed just as in normal text." [ISO 6.8.2] The result
must resemble one of the canonical forms, "foo.h" or <bar.h>.
This lets us write things like
#define HEADER "foo.h"
#include HEADER
I could not find a compiler rejecting this. However, if we're
torturing the preprocessor a little more, by using token pasting,
at least one compiler falls over.
#define H(x) <st##x##.h>
#include H(dio) /* expands to <stdio.h> */
is rejected by tcc version 4.1.2, the TenDRA compiler:
"t.c", line 3: Error:
[ISO 6.8.3.3]: Invalid result for '##' operator in macro 'H'.
To be honest, I don't know if the rules for pp-tokens and token
pasting don't forbid what I do (and tcc is correct in rejecting it).
In this case, all other compilers I tried are buggy, or the Standard
itself :-)
2. The #line directive allows to set a line number, and optionally
a file name, i.e. modify __LINE__ and __FILE__. Strange things
happen if the file name contains a NUL byte -- which the Standard
allows: in #line num "file" the "file" is an s-char-sequence per
ISO 6.8.4. ISO 6.1.4 defines string literals as s-char-sequences,
which may include octal escape sequences. Let's look at
#include <stdio.h>
#line 42 "foo\0bar"
int main (void) {
printf ("%s %d\n", __FILE__, (int)sizeof __FILE__);
return 0;
}
This is supposed to output 'foo 8'. Here's what happens in RL:
foo 4 gcc 2.7.2.1, lcc
foo\0bar 9 tcc 4.1.2, Sunsoft cc turn "\0" into "\\0", ugh!
(I have written a bug report for gcc. Newer versions and egcs are ok.)
3. The %: and %:%: digraphs test for conformance to Amendment One.
Ask your local guru if C allows the same case label in the same switch
statement to appear more than once. Ask him/her to think real hard.
The answer will be "no". The true guru will cite a constraint in
ISO 6.6.4.2. Then make fun of the guru's answer by waving 'case __LINE__:'
under his/her nose. Easy money from a bet! Make sure your guru has
no chance to use Standardese weasel words as an escape: can I have
case <some_token>:
case <some_token>:
in the same switch? (Note that you need the [invisible] newline. You
can generously allow the additional constraint that no #undefs or
#defines are allowed between the two cases.)
Lots of integer constants and string literals come into the source
via __LINE__ and __FILE__ which are redefined at various places.
Sometimes the __FILE__ contains \0 bytes; different strings or
characters are then accessed with offsets, say 10+__FILE__ or
__LINE__+__FILE__ or 2[__FILE__]. One integral constant in octal.
This makes the source and header input depend on ASCII.
A few old obfuscations, like one character identifiers, not too
many macros, a goto O, "needless" assignments to satisfy lint
with its "function value ignored" warnings. My lint has nothing
to complain.
There's more whitespace in the Standard than just space, tab and
newline. In particular, there are vertical tab and formfeed that
can be used in certain places (outside of preprocessing directives,
i.e. they are not allowed from just after the # up to the final
newline. See 6.8 for details.)
On many implementations these appear as ^L and ^K. My source starts
with an extremely uncommon 32bit word consisting of ^L, ^K, % and :.
I have grepped old winning entries for ^K, none of them contains one,
so this is something new -- after 14 years of IOCCC. Be sure to have
enough paper in your printer when you make a hardcopy of the source...
While I'm at it, the rules state that only space, tab and newline
are ignored for the count (plus '{', '}', ';' followed by
whitespace). The mkentry.c program, however, uses isspace() which
returns nonzero for vtab and formfeed and other characters as well.
I could have used a lot more ^K and ^L probably undetected by your
counter but decided to err on the side of safety. I use a perl
script to compute the character count according to the rules.
The advantages of an independent clean room approach...
#!/usr/bin/perl -w
$/ = "\0";
$_ = <>;
s/[;{}][ \t\n]/\n/g;
s/[ \t\n]+//g;
print length ($_),"\n";
I have tried, as suggested in the guidelines, to let the code look
like ordinary C code. Apart from a few long lines I think I left
the indentation like I would in RL. You should however not try to
use indent on the source. The code is extremely fragile because of the
myriads of __LINE__ macros -- indenting is a sure way to break
the program. Don't even think of maintaining that beast; I've
had my share of core dumps during development :-) A test suite
was used after every minor change to find out if the program still
does what it should.
From the goals: "To stress C compilers with unusual code."
That describes exactly my modest attempts...
