Complex Matrix Assign Crashes in IFORT 16.0.3

[John_Y_]

Hi,

Attached is a simple test case that crashes on a complex matrix assign in ifort 16.0.3.  It works with all previous ifort versions I have used.  The crash occurs when using the cmplx intrinsic and does not occur when not using the intrinsic.  For example,

          zz = cmplx(0._db,0._db,db) ! db defined in test file

crashes but

         zz = 0._db ! db defined in test file

does not.  My compile line is

       ifort -traceback testZ8.F90

Thanks,

John

[Kevin_D_Intel]

Thank you for the report and reproducer John. We'll look into this.

[Kevin_D_Intel]

This appears related to the creation of array stack temporaries. The program runs with pre-16.0 compilers without any aid and runs with 16.0 compilers with the aid of either -heap-arrays or elevating the shell stack limit (i.e. ulimit –s unlimited).

I need to consult others whether the changed behavior was anticipated with 16.0. I'll update after learning more.

For convenience I included your test case below.

!======================================================================
program test
  implicit none

  integer, parameter :: db = kind(1.0d0)
  integer, parameter :: sg = kind(1.0)

  integer, parameter :: stdout = 6

  ! local variables
  integer :: itmp, i, ierr, m

  complex(db), allocatable :: zz(:,:)
  complex(sg), allocatable :: cc(:,:)

  m = 5868

  allocate(zz(m,m),stat=ierr)
  write(stdout,*) 'IERR: ', ierr
  allocate(cc(m,m),stat=ierr)
  write(stdout,*) 'IERR: ', ierr

  zz = 0._db ! This line ok
  zz = cmplx(0._db,0._db,db) ! This line crashes

  cc = 0._sg ! This line ok
  cc = cmplx(0._sg,0._sg,sg) ! This line crashes

  deallocate(zz,cc)

end program test

[John_Y_]

Kevin,

Using "-heap-arrays" is a simple enough workaround.  It's interesting the behavior changed though.

Thanks,

John

[Steven_L_Intel1]

In version 15, the compiler did a simple element-by-element initialization in a loop. In version 16 it created an array temp for the whole initialized array and copied it - this triggered the stack overflow. In the upcoming version 17, it generates a vectorized loop to more efficiently set the values and doesn't use a temp.