Creating Black Holes: Division By Zero In Practice

Dividing by zero — the fundamental no-can-do of arithmetic. It is somewhat surrounded by mystery, and is a constant source for internet humor, whether it involves exploding microcontrollers, the collapse of the universe, or crashing your own world by having Siri tell you that you have no friends.

It’s also one of the few things gcc will warn you about by default, which caused a rather vivid discussion with interesting insights when I recently wrote about compiler warnings. And if you’re running a modern operating system, it might even send you a signal that something’s gone wrong and let you handle it in your code. Dividing by zero is more than theoretical, and serves as a great introduction to signals, so let’s have a closer look at it.

Chances are, the first time you heard about division itself back in elementary school, it was taught that dividing by zero is strictly forbidden — and obviously you didn’t want your teacher call the cops on you, so you obeyed and refrained from it. But as with many other things in life, the older you get, the less restrictive they become, and dividing by zero eventually turned from forbidden into simply being impossible and yielding an undefined result.

And indeed, if a = b/0, it would mean in reverse that a×0 = b. If b itself was zero, the equation would be true for every single number there is, making it impossible to define a concrete value for a. And if b was any other value, no single value multiplied by zero could result in anything non-zero. Once we move into the realms of calculus, we will learn that infinity appears to be the answer, but that’s in the end just replacing one abstract, mind-boggling concept with another one. And it won’t answer one question: how does all this play out in a processor?

Floating Point Division By Zero

Let’s start with floating point numbers. Nowadays, they are usually represented and stored in memory using the IEEE 754 format, splitting the value itself into separate fields for the sign, exponent, and fraction. If we look at a float or double in memory, it won’t make too much sense, as its actual value is constructed from those three separate fields. To demonstrate that, we can cast a float into an int and print it. Note that we have to do so by pointer conversion, otherwise we’d just end up with the integer part of the floating point number.

float fval = 12.34f;
int *iptr  = (int *) &fval;
printf("%f -> 0x%08x\n", fval, *iptr);
// output: 12.340000 -> 0x414570a4

Neither 0x414570a4 nor 1095069860 would give us any hint that this is the number 12.34. However, this form of representation leaves room for some special cases like infinity and not a number, something dividing zero by zero will result in, since that equation has no single answer and therefore cannot be represented by a regular number. That means we can just go ahead and divide by zero all we want, the IEEE 754 format has us covered.

fval = 1 / 0.0f;
printf("%f -> 0x%08x\n", fval, *iptr);
// output: inf -> 0x7f800000

fval = 0 / 0.0f;
printf("%f -> 0x%08x\n", fval, *iptr);
// output: -nan -> 0xffc00000

In other words, floating point numbers have built-in mechanisms to deal with division by zero without wreaking havoc. Not that it really helps, we cannot do much with either inf or nan, and arithmetic operations on infinity either remain infinity (maybe with changed signedness), or turn into nan, which is a dead end. No arithmetic operation can turn nan into anything else ever again.

Integer Division By Zero

If you tried out the previous example for yourself, you may have noticed that after all the talk about compiler warnings that led us here in the first place, you didn’t see a single one of them. Since floating point numbers have their own, well-defined way to handle division by zero, it doesn’t pose a threat that the compiler would have to warn about. This is all great, provided we have an FPU in place, or enough resources to emulate floating point operations in software. However, if we’re working with integers, we have one major problem: the concept of infinity doesn’t exist in the limited range of values we have available. Even if we ended up with a data type with an enormous amount of bits, best we could represent is a really, really large number, but never infinity. This time, the compiler will warn about an obvious attempt to divide by zero.

// zerodiv.c
int main(void) {
    int i = 10/0;
    return 0;
}
$ gcc -o zerodiv zerodiv.c
zerodiv.c: In function ‘main’:
zerodiv.c:3:15: warning: division by zero [-Wdiv-by-zero]
     int i = 10/0;
               ^
$

So what’s going to happen if we still do it? After all, it’s just a warning, we got a functional executable from the compiler, and nothing is going to stop us running it. Well, let’s do it and see what happens on x86_64:

$ ./zerodiv
Floating point exception (core dumped)
$

There we go, since we cannot represent the result in any way, the program simply crashed. But what caused that crash?

In more complex processors, the instruction set offers dedicated division opcodes, and the division itself is performed in hardware. This allows the processor to detect a zero divisor before the operation itself is executed, causing a hardware exception. In our case, the operating system caught the exception and raised the floating point exception signal SIGFPE — don’t mind the somewhat misleading name of it. So just like with floating point numbers, the hardware division instruction has means in place to avoid dealing with actually dividing by zero. But what about processors without such dedicated hardware, like an 8-bit AVR or ARM Cortex-M0? After all, division isn’t a particularly new concept that was only made possible by modern processor technology.

If you think back to your school days, pen-and-paper division was mainly a combination of shifts, comparison, and subtraction in a loop. These are all basic instructions available in even the simplest processors, which allows them to replace division with a series of other instructions. AN0964 describes the concept for AVR if you’re curious. While there are also more sophisticated approaches, in some cases the division instruction isn’t any different behind the scenes, it just has dedicated hardware for it to wrap it in a single opcode.

However, for the processor, there is no telling that a particular series of regular operations are actually performing a division that require to keep an eye on the divisor not being zero. We’d have to manually check that ourselves before the division. If we don’t, it just keeps happily on looping and comparing for all eternity (or 1267+ years) until it finds a number that fulfills the impossible, i.e. it simply gets stuck in an infinite loop. Mechanical calculators are great devices to demonstrate this.

While this will leave the universe intact, it essentially renders your program unresponsive just like any other endless loop, which is obviously bad since it’s most likely supposed to handle other things. If you’re therefore performing division or modulo operations on an architecture without a hardware divider, and the divisor isn’t from a predefined set that guarantees it won’t be zero, make sure you check the value beforehand. Well, in fact, it’s probably a good idea even with the right hardware support. An instant crash might be better than a possibly undetected endless loop, but either way, your code won’t be doing what it’s supposed to do.

Back To That SIGFPE

One difference with receiving a hardware exception that either turns into an interrupt or a signal like SIGFPE, is that we can act on it. Not to go too deep into the details: signals are a way to notify our running program that some certain event has happened — segmentation fault (SIGSEGV), program termination (SIGTERM and the more drastic SIGKILL), or the previously encountered floating point exception SIGFPE, to name a few. We can catch most of these signals in our code and act upon them however needed. For example, if we catch SIGINT, we can shut down gracefully when CTRL+C is pressed, or simply ignore it.

This means, we could catch a division by zero occurrence and for example save our current program state, or write some log file entries that lets us reproduce how we got here in the first place, which might help us avoiding the situation in the future. Well, time to write a signal handler then. Note that the code is slightly simplified, the sigaction man page is a good source for more information about masks and error handling.

// zerodiv.c
#include <stdio.h>
#include <signal.h>

// SIGFPE callback function
void sigfpe_handler(int sig, siginfo_t *si, void *arg) {
    // print some info and see if it was division by zero that got us here
    printf("SIGFPE received at %p due to %s\n", si->si_addr,
            ((si->si_code == FPE_INTDIV) ? "div by zero" : "other reasons"));
    // do whatever should be done
}

int main(void) {
    int i = 123;
    struct sigaction sa;

    // set sigfpe_handler() as our handler function
    sa.sa_sigaction = sigfpe_handler;
    // make sure we get the siginfo_t content and reset the original handler
    sa.sa_flags = SA_SIGINFO | SA_RESETHAND;

    // set up the signal handling for SIGFPE
    sigaction(SIGFPE, &sa, NULL);

    printf("before: %d\n", i);
    i /= 0; // doing the nasty
    printf("after:  %d\n", i);

    return 0;
}

If we compile it and run it, we can expect the division by zero warning, and then get an output like this:

$ ./zerodiv
before: 123
SIGFPE received at 0x55f2c333b208 due to div by zero
Floating point exception (core dumped)
$

Since we added the SA_RESETHAND flag, the program gets still terminated with the original exception. If we omit the flag, it won’t, but that doesn’t mean we would have successfully worked around the problem. On x86_64, the signal handler simply ends up in an endless loop, printing the message over and over again. We’d have to explicitly terminate the process by calling exit(int) in the signal handler.

On a side note, it appears that the ARM Cortex-A53 processor (the one you find on a Raspberry Pi) automatically resets the exception flag once handled, and therefore the program continues after returning from the signal handler, displaying after: 0. This suggests that the division by zero is defined to result in zero. I did not succeed resetting the flag on x86_64, hence the endless loop, but that doesn’t necessarily mean it’s not possible, I simply wasn’t able to achieve it myself. However, there’s one other thing we can do on x86_64: skip the division.

Skipping The Division

Note the third void *arg parameter in the signal handler callback? It will contain the context of the moment the signal was received, which gives us access to the saved registers, including the instruction pointer, which will be restored when we leave the signal handler. If we disassemble our code, we will see that our division in this particular example is a 2-byte instruction:

$ objdump -d ./zerodiv |grep -2 div
    125b:       b9 00 00 00 00          mov    $0x0,%ecx
    1260:       99                      cltd   
    1261:   --> f7 f9                   idiv   %ecx
    1263:       89 85 5c ff ff ff       mov    %eax,-0xa4(%rbp)
    1269:       8b 85 5c ff ff ff       mov    -0xa4(%rbp),%eax
$

If we increased the instruction pointer register RIP (or EIP on 32-bit x86) by two bytes, the execution would continue with the mov instruction after the idiv when we return from the signal handler. This kind of register tweaking sounds like a really bad idea, so you probably shouldn’t be doing something like this:

#include <stdio.h>
#define __USE_GNU // this and the file order is important to succeed
#include <ucontext.h>
#include <signal.h>

// adjusted signal handler
void sigfpe_handler(int sig, siginfo_t *si, void *arg) {
    // cast arg to context struct pointer holding the registers
    ucontext_t *ctx = arg;
    // print some info
    printf("SIGFPE received at %p\n", si->si_addr);
    // add 2 bytes to the instruction pointer register
    ctx->uc_mcontext.gregs[REG_RIP] += 2;
}

// main() remained as it was
$ ./zerodiv 
before: 123
SIGFPE received at 0x555c1ef5b243
after:  123
$

Tadaa — the division was skipped and the program survived. The value itself remained the same as before the attempted division. Obviously, any subsequent operation relying on the division’s result will most likely be useless and/or has unknown consequences. Same case with the Raspberry Pi earlier that yielded zero as result. Just because a random outcome was defined for an otherwise undefined situation doesn’t mean the underlying math was solved and everything makes suddenly sense. Failing fast and determined is often times the better option.

So there we have it. You won’t create black holes when dividing by zero, but you won’t get much useful out of it either.

30 thoughts on “Creating Black Holes: Division By Zero In Practice

  1. One might consider the Matrix coding has need to remove components to reduce processing overhead so the so called Black Holes sequester the permutation such as to reduce overal computational load on the program that runs the universe ;-)

  2. That was an interesting read, thank you!
    TL;DR: You can get trapped in an infinite loop if you divide by zero – I didn’t know that, might be usefull one day when debugging with small µC.

  3. When I learned about division back when I were young, the teacher explained it as:
    “Division is like multiplication, but instead of adding together the same number a set amount of times, we instead look at how many times we need to subtract a number from another to reach zero.”

    This explanation then states that a/b=k would be stated as 0= a-b*k
    And if we make b = 0, then k now can be what ever number, but the answer of b*k will still equal 0, therefor the equation of 0 = a-b*k will never give an answer.

    So in my case, the teacher never told me I couldn’t divide by zero, but rather told my how division works and let me explore on my own. (Rather advanced for being in like 9 years old at the time.)

  4. Infinity seems like an ok answer, since the result approaches infinity as the divisor get closer to zero, but the result approaches negative infinity if the divisor is negative and gets arbitrarily small.

    Oh, fun fact. If you use signed division, INT_MIN/-1 also results in a SIGFPE. A lot of programs check for zero to avoid exceptions, but not INT_MIN/-1.

    1. And of course, there’s other desired answers as well. For instance, the sinc function (sin(x)/x) – when x = 0 you want a result of 1. But of course that’s squarely in the realm of 0 divided by 0 equalling anything at all – with the particular anything desired being determined by the limit.

    2. Not strictly. The LIMIT of a/x -> +/- infinity as x-> for constant a, and 0/x -> 0 as x-> 0. ax/bx -> a/b as x-> 0.

      Basically, there’s an annoyingly large number of solutions, you can’t generalize. It gets even worse when you have complex functions like sinusoids.

  5. “Tadaa — the division was skipped and the program survived. The value itself remained the same as before the attempted division.”

    Why didn’t the people who came up with mathematics ages ago simply say that dividing by zero should be treated the same as dividing by one? Same result. Divide by 1 the result is . Divide anything by nothing and you’re left with the starting amount of anything.

    1. Hmm, the issues of; asymptotes, limits and most importantly Calculus hit that squarely on the head well over a hundred years ago and especially so when dealing with more than a couple of dimensions. At the very least; Newton, Gauss and Einstein would be rolling in their graves and at a very high rate too leading to even more spam fluff these days in so called ‘free energy zero point devices’ ;-)

        1. Well sure, its an issue of diminishing returns or in respect to the so called zero point energy con-artist advocates interpreted as tantalising greedy accelerating returns as the capital cost appears fixed whilst the claimed continuous energy value produced successively improves wildest ROI – the facile propaganda of the illusory “no brainer” of the uneducated emotionally pliable (welcome to all religions) drives the attraction to their commercial doom (Barnum).
          My point in relation to division by zero as if by 1 if there were no change to numerator were adopted as some dogma would galvanise so many to go down all sorts of blind alleys wasting huge resources and most importantly diminishing intellectual time.
          iow. It cannot work in the real world as well proven long ago and for now since we live in those most interesting times in which mathematics has shown itself to be staggeringly successful in describing the real world far better than most ever imagined or expected and with the most successful theory of all time bar none being Quantum Mechanics (QM) with its nonassociative operators by far at the top of the tree !
          ie we can replace mere things very easily fwiw as money easier to make than ever before with basic psychological pre-requisites anyone can learn yet time can never be replaced (presently afawk) though with attention to health we can definitely (wide and probabilistically) extend life span seriously.
          Also fwiw 99% of all the top mathematicians who ever lived in all of history are alive today and continue a most sophisticated ‘work in progress’ divergent path with so many advanced tools far beyond most public’s appreciation let alone basic awareness. So the gestalt of the momentum of mathematics as it progresses will further likely sidestep the facile operative numeracy of messing with the long ago settled issue of division by (the idea of a nothing being) zero.

          Though I might add the potential for zero divided by zero in various arenas in a quantified QM setting is ‘most interesting indeed’ and one might argue in some theoretical physics paradigms possibly provokec by scifi or just plain private study into anomalies – is just the sort of conundrum that;
          You Do Really Want To Aim For and for all sorts of reasons beyond physics :D

          1. Lots of thinking about mathematics is a wonderful thing. Still it is best to remember that mathematics in its purest form has nothing to do with reality. That’s what makes it useful. We may talk about whether we must avoid dividing by zero or not in our wonderful fantasy world of mathematics but the universe just does what it does with no guarantee that it will match up with the “shapes” we create in our mathematical models. The fact that we have built many models in the world of mathematics that match closely to what we observe in the real world is a true source of wonder, but the worlds are independent with the only connection being people tweaking their mathematical models to match what is actually in reality. Because of this one cannot divide by zero and create a black hole somewhere. In fact there isn’t a single mathematical incantation that can change a single event in the real world. …and that’s what makes mathematics so wonderful and beautiful. Its the perfect “ledger” for understanding reality. In the world of mathematics, WE are the music makers, and WE are the dreamers of dreams…and such wonderful dreams when we manage to steer clear of dividing by zero. :-) As for the real universe its-self we are NOT the Author. We can do things within it but we cannot do anything that is outside our given nature. An equation is as real in mathematics as a soul is in the real world. Both can be “Fearfully and Wonderfully Made”. Both may have many undiscovered meanings and potentials. Both are loved by their Author.

    2. a/b=c is the same as a=bc. If b=1, then a/1=c, a=c*1 (thus a=c.) But you can’t do that with 0. a/0 =/= a*0 for any non-zero a. So no, you can’t simply say that “dividing by zero [is to] be treated the same as dividing by one” without breaking the “a/b=c is the same as a=bc” for all non-zero/non-one values of a, b or c.

    3. Hmm. So 12/1 = 12 and 12/0=12. 0 still does not equal 1. Dividing by zero is kind of a double-negative, english-wise… because divinding by nothing is not dividing… it never happened.

  6. Black holes are in fact the summation of millions upon millions of solar masses that were nearly divided by zero…the result is an aperture for the infinite focusing into the finite, which explains the acceleration of the universe.
    The same mechanisms are responsible for “flyback” or B/CEMF collapsing magnetic fields (pulsed DC). Delta Work/Delta Time….closer and closer you can get to Toff=0 the higher the inverted voltage spike. If a “perfect switch” could be created, one could translate the most minute power into the infinite, no matter how brief, infinite is infinite…. all we need is 1.21 gigawatts!

  7. I think the first code example is quite wrong, because it violates the strict aliasing rules. It also will fail on platforms where sizeof(int) != sizeof(float) (e.g. AVR). A better variant would be:

    float fval = 12.34f;

    printf(“%f -> 0x”, fval);
    for (size_t i = 0; i < sizeof(float); ++i)
    printf("%02x", ((unsigned char*) &fval) [i]);

    printf ("\n");

    It's still not 100% perfect, as a platform might have CHAR_BIT != 8 resulting in an incorrect hexadecimal number. Accessing individual bytes of any data type is usually a bad idea anyways.

    1. >It also will fail on platforms where sizeof(int) != sizeof(float) (e.g. AVR).
      Good catch. However as you said accessing individual bytes of a float is usually a bad idea, this code is just to show that floats are encoded in a “strange” way (well, IEEE754) and so i think its acceptable that this will only run on a PC. However i can think of an use case for decomposing float into bytes: send it over some wired or wireless communication channel to put it back together on the other side (but yes, you could transform it to ASCII before).

      1. Converting a float to ASCII for transmission will introduce more errors and require more bandwidth than transmitting the IEEE754 bytes as raw data. OTOH not all bit patterns are valid floats, and if you make a mistake and try to interpret random garbage data as a float, you are quite likely to generate an exception from the math coprocessor. I have worked with one industrial device where this generates a hard crash because the manufacturer didn’t bother to install a handler for floating point exceptions.

  8. Tangential question, regarding the aesthetics of the code snippets – is there gvim colorscheme for that? Or if it came from (gawdsforbid!) some other editor, what editor was it and is there a config file for that? Also, what’s the font?

      1. You would be doing zero lots of eight or zero lots of zero, it’s still zero. Eight boxes with numbers written in them. You add them up fine. Now you make say three lots of those eight boxes. You can still add them all up. Now make those boxes all contain a zero. You can still add them all up. They just sum to be zero even with infinite boxes.

  9. Often a plus or minus infinity is exactly the answer you want if you have compliant floating point handling. It’s worth the check to see if the algorithm works with inifinities and if so just leave them in. Sure you can’t do arithmetic on them in a useful fashion but inequalities work great for example.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.