integer overflow

[Ashutosh Iddya]

Hi ,

I am performing an integer count of a particular operation in my program.
After a sufficiently large value an overflow occurs. At the moment I have
gone around the problem by declaring it as a double, even that has its
limits. Is there a method of preventing this overflow or some method of
recovering from it. Any help in this regard would be greatly appreciated.

Thanking you.

Ashutosh

 

[CBFalconer]

I am performing an integer count of a particular operation in my
program. After a sufficiently large value an overflow occurs. At
the moment I have gone around the problem by declaring it as a
double, even that has its limits. Is there a method of preventing
this overflow or some method of recovering from it. Any help in
this regard would be greatly appreciated.

If long won't do, then you can use long long (on C99 or gnu gcc
systems). Otherwise try:

if (n < INT_MAX) n++;
else {
overflows++;
n = 0;
}

and don't forget to #include <limits.h>

It would be simpler to use unsigned types, and then:

if (!(++n)) overflow++;

 

[Ashutosh Iddya]

thanks for that. I will try it out and let you know how it went
Ashutosh

 

[Darrell Grainger]

Hi ,

I am performing an integer count of a particular operation in my program.
After a sufficiently large value an overflow occurs. At the moment I have
gone around the problem by declaring it as a double, even that has its
limits. Is there a method of preventing this overflow or some method of
recovering from it. Any help in this regard would be greatly appreciated.

If overflow is happening with your data type you could detect it but
preventing it means switching to a different data type. For an integer
data type (long long) is your best bet. If you are only working with
unsigned numbers then (unsigned long long) is your best bet.

You might still get overflow. If this is the case you can try using double
as your data type. Another option is to search the web for "big number C
library". There are libraries you can use for numbers bigger than unsigned
long long.

If you go with the big number library it will probably be slower than
using long long. You might me able to create your own limited big number
library. If you are just counting then you only need addition and some way
of printing the number.

 

[Dan Pop]

I am performing an integer count of a particular operation in my program.
After a sufficiently large value an overflow occurs. At the moment I have
gone around the problem by declaring it as a double, even that has its
limits.

I have a hard time imagining you're going to reach the limits of the
double solution within a reasonable period of time, even if all you
do is counting. Try the following program and see how long it takes
to terminate normally.

#include <stdio.h>

int main()
{
double d = 0;
while (d + 1 != d) d++;
printf("%.16f\n", d);
return 0;
}

Assumming that the loop takes one CPU cycle per iteration and that the CPU
is running at 4.5 GHz, this proggie should run for about 1e6 seconds, if
using IEEE-754 doubles.

BTW, how about "optimising" the while loop like this?

while (d != ++d) ;

;-)

Dan

 

[Dan Pop]

If overflow is happening with your data type you could detect it but
preventing it means switching to a different data type. For an integer
data type (long long) is your best bet. If you are only working with
unsigned numbers then (unsigned long long) is your best bet.

Unless your C compiler, invoked in conforming mode, tells you that
"long long" is a syntax error ;-)

Are the latest Microsoft C compilers supporting long long?

Dan

 

[Grumble]

If overflow is happening with your data type you could detect it but
preventing it means switching to a different data type. For an integer
data type (long long) is your best bet. If you are only working with
unsigned numbers then (unsigned long long) is your best bet.

You might still get overflow.

C99's long long int is /at least/ 64 bits wide.

If the OP's counter were incremented 100 times every cycle on a 10 GHz
processor, it would take 213 days for the counter to overflow.

If this is the case you can try using double as your data type.

Bad advice.

An IEEE 754 double will only provide 53 bits of precision.

See DBL_MANT_DIG and FLT_RADIX in float.h

For my information, are there implementations where double is wider
than 64 bits?

 

[Eric Sosman]

[...]
For my information, are there implementations where double is wider
than 64 bits?

VAX H-format floating point had/has 128 bits; I don't recall
how they're divided up between exponent and fraction. Also,
when I was working with VAXen the C implementations were not
able to use H-format. (For the curious: `float' used the
32-bit F-format, and `double' used either D-format or G-format,
both 64 bits, at your option -- and if you accidentally mixed
G's and D's ... "The horror! The horror!")

 

[CBFalconer]

thanks for that. I will try it out and let you know how it went

Kindly DO NOT toppost. See sig.

 

[Darrell Grainger]

C99's long long int is /at least/ 64 bits wide.

I'd doubt the OP was using a C99 compiler. It is possible but my first
assumption would be a C89 compiler.

If the OP's counter were incremented 100 times every cycle on a 10 GHz
processor, it would take 213 days for the counter to overflow.

Man, slow day and my brain shuts down. Just a month ago I proved that
overflow on a cycle count profiler for a 1 GHz processor would take over
500 years to occur. I should have realized this.

Bad advice.

An IEEE 754 double will only provide 53 bits of precision.

See DBL_MANT_DIG and FLT_RADIX in float.h

I was thinking more of situations when (long long) would be 32 bit. On
older compilers I remember seeing support for (long long) such that
sizeof(long) == sizeof(long long). Essentially they made it so source with
(long long) would not be considered a syntax error but still only
supported 32 bit integers.

For my information, are there implementations where double is wider
than 64 bits?

Not that I have seen.

 

[Keith Thompson]

I am performing an integer count of a particular operation in my program.
After a sufficiently large value an overflow occurs. At the moment I have
gone around the problem by declaring it as a double, even that has its
limits. Is there a method of preventing this overflow or some method of
recovering from it. Any help in this regard would be greatly appreciated.

Using double is probably the wrong solution. If you repeatedly add 1
to a double variable, you'll never get an overflow, just a loss of
precision; eventually, (d + 1.0 == d).

The type "long long" is guaranteed to be at least 64 bits, which
should be more than enough for whatever it is you're counting. This
type is new in C99, but many C90 compilers support it as an extension.
(Since the C90 standard doesn't specify "long long", it's
theoretically possible that a C90 compiler could provide a "long long"
type that's smaller than 64 bits, but I don't think anyone has ever
done that.)

If 32 bits is enough, use long (or unsigned long).

If 32 bits isn't enough, and you don't care about portability to
systems that don't support long long, use long long (or
unsigned long long).

Be aware that a C90 implementation that supports long long as an
extension may not necessarily support everything that goes with it,
particularly the printf formats.

If you can't use long long, you might try using a pair of unsigned
longs (untested code follows):

unsigned long count_hi = 0;
unsigned long count_lo = 0;
...
for (...) {
count_lo ++;
if (count_lo == 0) count_hi ++;
}

Note the use of unsigned types. Overflow is well-defined for unsigned
types; it wraps around. Overflow for signed types causes undefined
behavior; wraparound is common, but not guaranteed.

 

[jacob navia]

Darrell Grainger wrote:

For my information, are there implementations where double is wider
than 64 bits?

There is the standard long double for that.
In lcc-win32 long double is 80 bits.

Then, you have the qfloat with 350 bits if you
really want big precision.

After that, the bignums is the only way, or
(much better) to look at the algorithm and
see why it is overflowing

http://www.cs.virginia.edu/~lcc-win32

 

[Keith Thompson]

For my information, are there implementations where double is wider
than 64 bits?

The standard allows it, but I doubt that any implementations actually
do so. If you have a floating-point type wider than 64 bits, you're
more likely to call it "long double" (which has been valid at least
since the C89/C90 standard).

 

[Keith Thompson]

I was thinking more of situations when (long long) would be 32 bit. On
older compilers I remember seeing support for (long long) such that
sizeof(long) == sizeof(long long). Essentially they made it so source with
(long long) would not be considered a syntax error but still only
supported 32 bit integers.

Ick!

Can you remember which compiler did this? Did it really implement
"long long" as a distinct type, or did it just allow the "long"
keyword to be repeated with no further effect (so "long", "long long",
and "long long long" would all be equivalent)?

 

[Peter Nilsson]

Hi ,

I am performing an integer count of a particular operation in my program.
After a sufficiently large value an overflow occurs. At the moment I have
gone around the problem by declaring it as a double, even that has its
limits. Is there a method of preventing this overflow or some method of
recovering from it. Any help in this regard would be greatly appreciated.

unsigned long counter[2] = { 0 };

for (;
{
if (++counter[0] && ++counter[1])
puts("64+ bit counter overflowed!");
}

This is trivially extendable to as much precision as you want. But if you need a counter
bigger than a minimum of 64-bits, then I'd love to know what machine you're using and
where I can pick one up!

 

[Arthur J. O'Dwyer]

I am performing an integer count of a particular operation in my program.
After a sufficiently large value an overflow occurs. At the moment I have
gone around the problem by declaring it as a double, even that has its
limits. Is there a method of preventing this overflow or some method of
recovering from it. Any help in this regard would be greatly appreciated.

unsigned long counter[2] = { 0 };

for (;
{
if (++counter[0] && ++counter[1])

ITYM if (!(++counter[0] || ++counter[1]))

You want to make sure they both *don't* go back to zero at once.

puts("64+ bit counter overflowed!");
}

-Arthur

 

[Peter Nilsson]

I am performing an integer count of a particular operation in my
program. After a sufficiently large value an overflow occurs. At
the moment I have gone around the problem by declaring it as a
double, even that has its limits. Is there a method of preventing
this overflow or some method of recovering from it. Any help in
this regard would be greatly appreciated.

unsigned long counter[2] = { 0 };

for (;
{
if (++counter[0] && ++counter[1])

ITYM if (!(++counter[0] || ++counter[1]))

Or...

if (++counter[0] == 0 && ++count[1] == 0)

You want to make sure they both *don't* go back to zero at once.

Yup, my bad. Thanks.

 

[Dan Pop]

I was thinking more of situations when (long long) would be 32 bit.

No such thing. With 32-bit int and long, no point in adding a 32-bit
long long.

On older compilers I remember seeing support for (long long) such that
sizeof(long) == sizeof(long long).

True, but those were compilers with 64-bit long's (e.g. the Digital Unix
compilers as well as gcc on any 64-bit Linux flavour).

Essentially they made it so source with
(long long) would not be considered a syntax error but still only
supported 32 bit integers.

A brain dead idea: the code will compile but silently generate wrong
results, because people using long long *really* expect more than 32 bits
(otherwise they would be using long in the first place).

Dan

 

[Dan Pop]

[...]
For my information, are there implementations where double is wider
than 64 bits? ^^^^^^

VAX H-format floating point had/has 128 bits; I don't recall
how they're divided up between exponent and fraction. Also,
when I was working with VAXen the C implementations were not
able to use H-format.

So, you're providing a non-example. long double would have been the
right type for the VAX H_FLOAT type, but I don't know if any VAX C
compiler actually supported it. VAX FORTRAN did support it as REAL*16.

Dan

 

[Dan Pop]

Using double is probably the wrong solution. If you repeatedly add 1
to a double variable, you'll never get an overflow, just a loss of
precision; eventually, (d + 1.0 == d).

No loss of precision until d + 1.0 == d, and the point where this happens
can be detected. On typical implementation, this should be around 2 ** 53
which should be enough for most applications needing to count something.

The type "long long" is guaranteed to be at least 64 bits, which
should be more than enough for whatever it is you're counting. This
type is new in C99, but many C90 compilers support it as an extension.

Not that many out of the Unix world. They typically support a 64-bit
integer type, but give it a name in the implementation name space. Even
on 64-bit hardware...

Dan