IFX error with complex array

[g_granucci]

This simple program gives a wrong result if compiled with ifx (version 2025.3, but also with previous versions):

program cx
  implicit none
  complex, allocatable, dimension(:,:) :: a
  integer :: i,n

  n=2
  allocate (a(-1:1,n))
  
  a(0,:)=Cmplx(0,1)
  a(-1,:)=Cmplx(4,5)
  a(1,:)=Cmplx(2,3)

  do i=1,n
    write(*,*)i,a(0,i)
  end do
end program cx

$ ./cx
           1 (4.000000,5.000000)
           2 (4.000000,5.000000)

instead of the expected result (0,1), obtained with ifort and gfortran.

[MarcGrodent]

Running the code in debug mode leads to correct results. The problem occurs in release mode, probably due to the code optimization...

I'm using version 2025.3.0

[Mentzer__Stuart]

Good catch!  As @MarcGrodent noted it is optimization related: the bug shows up at /O2 and /O3 but not /O1.

Dumping the whole array shows that a(0,:) is treated as a(-1,:), both in the assignment and the elementwise write: the contents of a(0,:) when printing the whole array can contain (0,0) or uninitialized garbage values.

[Igor_V_Intel]

This loop optimizer bug is still in the latest compiler builds. I have escalated it to be fixed.

LLVM bisect tool points to the hir-post-vec-complete-unroll optimization pass, causing the problem. Note that with O1 results are correct. Looks like a full unrolling has a bug. I also tried to disable it, but then values are set to zero, which is also not correct.

$ ifx -O2 -fno-unroll-loops test.f90

$ ./a.out

      1 (0.0000000E+00,0.0000000E+00)

      2 (0.0000000E+00,0.0000000E+00)