Possible bug in cshift?

[pgfogli]

It seems that CSHIFT intrinsic function has a bug


when at least one of the array bounds is zero or less.


The following program shows this weird behaviour:





####################################################

PROGRAM test_cshift

IMPLICIT NONE

  ! Allocatable array later allocated with negative bounds
  INTEGER, ALLOCATABLE :: pa(:)
  ! Array with lower bound = 0
  INTEGER :: ps(0:4) = (/ 1, 2, 3, 4, 5 /)
  INTEGER :: k=1

  WRITE(*,*) 'k: ', k

  WRITE(*,*) 'ps: ', ps
  ps=CSHIFT(ps,k)
  WRITE(*,*) 'ps: ', ps

  ALLOCATE(pa(-31:-29))
  pa = (/ 10, 11, 12 /)
  WRITE(*,*) 'pa: ', pa
  pa=CSHIFT(pa,k)
  WRITE(*,*) 'pa: ', pa

  DEALLOCATE(pa)
  
  STOP

END PROGRAM test_cshift

####################################################





Follows the output:

$> ifort -o test test_cshift.f90
$> ./test
 k:            1
 ps:            1           2           3           4           5
 ps:            3           4           5           0           2
 pa:           10          11          12
 pa:            0           0           0

####################################################





System informations:





Intel Fortran Compiler for 32-bit applications, Version 9.0 Build 20051201Z


ID: l_fc_c_9.0.031


Slackware GNU/Linux 10.2


glibc 2.3.6


kernel 2.4.32


Mobile Intel Pentium 4 - M CPU 2.00GHz





The above test program gives the correct results


with g95 (gcc version 4.0.1 Jan 9 2006) on the


same platform, and with Sun Fortran 95 8.0 on


Solaris 8 (ULTRASPARC 9):

$> g95 -o test test_cshift.f90
$> ./test
 k:  1
 ps:  1 2 3 4 5
 ps:  2 3 4 5 1
 pa:  10 11 12
 pa:  11 12 10

#> f90 -o test test_cshift.f90
#> ./test
 k:  1
 ps:  1 2 3 4 5
 ps:  2 3 4 5 1
 pa:  10 11 12
 pa:  11 12 10

Is this really a bug?





Thanks,


P. Fogli

[TimP]

I see the same bad result with ifort 9.0.028 on SuSE 9.2 32-bit, while 9.1.017 produces the correct result.

[pgfogli]

Is 9.1 already released?
The web page still report version 9.0
(9.0.031 on the ftp).

I feel that version 8.1 too has this bug
but I can't test it myself because
I no longer have that version.

BTW, if version 9.1 has fixed the bug a preprocessor conditional like the following one should be sufficient:

#if defined(__INTEL_COMPILER) && (__INTEL_COMPILER <= 900)
! work around code
#else
! normal code
#endif

[Steven_L_Intel1]

Wait for l_fc_c_9.0.032, should be available in the next few days. The bug is fixed in that version.

9.1 is still in beta.