[Performance issue] Dead code don't remove in flang front end

[vfdff]

Base the following code, all the pointer malloced are not used outside the function, so all of them are dead code. We'll see that
gfortran will delete the dead code the the very early tree-dump, and only leave 3 movl to retrun the arguments（reference to https://godbolt.org/z/3dMc1n1x1）. while flang doen't do similar optimization.

• gfortran -march=armv8.2-a -g -O3 -ffast-math -fivopts -ftree-vectorize -lm simple.f90 -o simple.gcc.s -S -v -save-temps -fdump-tree-all

subroutine test(ij, l, k, N)
    integer :: ij, k, l, N
	integer, parameter :: ADM_gall = 1156
    integer, parameter :: ADM_kall = 42
    integer, parameter :: ADM_lall = 2
    real(8), dimension(N,N,N) :: rhog, rhogvx, rhogvy, rhogvz, rhogw, rhoge, rhog_split
    real(8), dimension(N,N,N) :: rhogvx_split, rhogvy_split, rhogvz_split, rhogw_split, rhoge_split 
    !---------------------------------------------------------------------------
 
    !--- update prognostic variables
    do l = 1, ADM_lall
       do k  = 1, ADM_kall
       do ij = 1, ADM_gall
             rhog  (ij,k,l) = rhog  (ij,k,l) + rhog_split  (ij,k,l)
             rhogvx(ij,k,l) = rhogvx(ij,k,l) + rhogvx_split(ij,k,l)
             rhogvy(ij,k,l) = rhogvy(ij,k,l) + rhogvy_split(ij,k,l)
             rhogvz(ij,k,l) = rhogvz(ij,k,l) + rhogvz_split(ij,k,l)
             rhogw (ij,k,l) = rhogw (ij,k,l) + rhogw_split (ij,k,l)
             rhoge (ij,k,l) = rhoge (ij,k,l) + rhoge_split (ij,k,l)
       enddo
       enddo
    enddo
 end

[vfdff]

while gfortran doesn't delete the initializtion in following smart case . https://godbolt.org/z/oenTK4rMn

subroutine t(v, row, col)
  integer :: row
  integer :: col
  integer(8) :: v (row, col)

  v = 2 // initializtion all elements 
end