Floating Exception taking zero to a power

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

I found what seems like a bug using pgf90 17.7.0 on a cray machine.

ftn --version
pgf90 17.7-0 64-bit target on x86-64 Linux -tp haswell-64

I’m taking 0.0 to a power, which gives a floating point exception. I’ve isolated this problem to a short fortran program:

program simple 
implicit none
real(kind=8) :: r
real(kind=8) :: a ,b
 a = 0.00000000
 b = 2.4
 write (6,*) "Equation r = ",a , " ** ", b
     r = a ** (b)
     write (6,*)"Debugging print r = ",  r

end program simple

This is what happens when I compile and run:

ftn -Ktrap=ovf simple.F90 && ./a.out
Equation r = 0.000000000000000 ** 2.400000095367432
Floating exception (core dumped)

Here’s where it gets wacky. If I change b = 3.4, the program runs to completion:

grep 3.4 simple.F90
b = 3.4
ftn -Ktrap=fp simple.F90 && ./a.out
Equation r = 0.000000000000000 ** 3.400000095367432
Debugging print r = 0.000000000000000

I’ve tried changing b to a bunch of different numbers. Sometimes the program crashes, and sometimes it runs to completion. Adding “d0” to the end of 0.0000d0 or 2.4d0 does not fix the problem, although that shouldn’t be necessary anyway.

Has anyone else seen this problem? Is it fixed in a later version of the pfg90 compiler?

2

Hi Tom,

Thanks for the report. I’ve reached out to our math library team and confirmed that it is indeed an issue with our pow routine.

The error does still occur in our current and upcoming 18.7 release, hence I filed a problem report (TPR#26157).

Thanks again,
Mat

3

Hi Tom,

The issue is resolved with release 18.10