double problem

[Nomak]

Hello,

i've lost about hours (at least) to find a problem. I use the
following utility function:

13 template <typename T>
14 inline
15 T
16 diffabs(T a, T b)
17 {
18 T ret;
19
20 if (a > b)
21 ret = a - b;
22 else
23 ret = b - a;
24
25 if (ret < 0)
26 cerr << "ERROR: diffabs(" << a << ", "
27 << b << ") = " << ret << endl;
28
29 assert(ret >= 0);
30 return ret;
31 }

The verification part have been added after founding the bug.

At runtime, i have the assertion wich fails, but no error message, so
i run gdb:

#7 0x08050747 in diffabs<double> (a=-nan(0x8000000000000), b=-1)
29 assert(ret >= 0);
(gdb) print ret
$1 = -nan(0x8000000000000)

I don't even know wich path is taken. But the error message test fails
(<0) while the assert test is ok (>=0).

Please tell me it's not the way it should be and it's a compiler bug.

If it's the standard way, do you know why? And how can i detect if a
double if negativ or -nan ?

TIA

 

[Nomak]

Hello,

i've lost about hours (at least) to find a problem. I use the
following utility function:

13 template <typename T>
14 inline
15 T
16 diffabs(T a, T b)
17 {
18 T ret;
19
20 if (a > b)
21 ret = a - b;
22 else
23 ret = b - a;
24
25 if (ret < 0)
26 cerr << "ERROR: diffabs(" << a << ", "
27 << b << ") = " << ret << endl;
28
29 assert(ret >= 0);
30 return ret;
31 }

The verification part have been added after founding the bug.

At runtime, i have the assertion wich fails, but no error message, so
i run gdb:

#7 0x08050747 in diffabs<double> (a=-nan(0x8000000000000), b=-1)
29 assert(ret >= 0);
(gdb) print ret
$1 = -nan(0x8000000000000)

I don't even know wich path is taken. But the error message test fails
(<0) while the assert test is ok (>=0).

Please tell me it's not the way it should be and it's a compiler bug.

If it's the standard way, do you know why? And how can i detect if a
double if negativ or -nan ?

TIA

ok, i've seen the faq, too bad

 

[Kai-Uwe Bux]

Hello,

i've lost about hours (at least) to find a problem. I use the
following utility function:

13 template <typename T>
14 inline
15 T
16 diffabs(T a, T b)
17 {
18 T ret;
19
20 if (a > b)
21 ret = a - b;
22 else
23 ret = b - a;
24
25 if (ret < 0)
26 cerr << "ERROR: diffabs(" << a << ", "
27 << b << ") = " << ret << endl;
28
29 assert(ret >= 0);
30 return ret;
31 }

The verification part have been added after founding the bug.

At runtime, i have the assertion wich fails, but no error message, so
i run gdb:

#7 0x08050747 in diffabs<double> (a=-nan(0x8000000000000), b=-1)
29 assert(ret >= 0);
(gdb) print ret
$1 = -nan(0x8000000000000)

The assertion fails and you do not see the error message. The reason is that
the value of ret is "not_a_number". This particular value does not compare
the usual way:

#include <limits>
#include <iostream>

int main ( void ) {
double nan = std::numeric_limits<double>::quiet_NaN();
std::cout << ( nan < 0 ) << "\n"
<< ( nan > 0 ) << "\n"
<< ( nan <= 0 ) << "\n"
<< ( nan >= 0 ) << "\n"
<< ( nan >= 0 ) << "\n"
<< ( nan < nan ) << "\n"
<< ( nan > nan ) << "\n"
<< ( nan == nan ) << "\n"
<< ( nan <= nan ) << "\n"
<< ( nan >= nan ) << "\n";
}

output:

0
0
0
0
0
0
0
0
0
0

And this output is expected.

I don't even know wich path is taken. But the error message test fails
(<0) while the assert test is ok (>=0).

Please tell me it's not the way it should be and it's a compiler bug.

If it's the standard way, do you know why? And how can i detect if a
double if negativ or -nan ?

TIA

One way to check for nan is to use the fact that !( nan == nan ) is true.


Best

Kai-Uwe Bux

 

[Matthew Hall]

Kai-Uwe Bux wrote:
.....

One way to check for nan is to use the fact that !( nan == nan ) is true.


Best

Kai-Uwe Bux

Just a note - there are compilers/platforms where (x==x) returns true
even if x is nan - the mipspro compiler (at least the 7.3.2.1m version
installed here) is one such example, even with all optimizations
disabled. If you care about portability to such machines, you may wish
to use 'isnan(x)' as defined in cmath (or math.h on Irix machines). I
have no idea of the relative speed of isnan to (x==x), but I do know
isnan can be slow (somewhat slower than floating point division on my
intel boxes w/linux).

-matt