Optimizer bug

[mecej4]

This is a report on an optimizer bug that is shown by the current 32-bit compiler (Update 7) for Windows, w_cprof_p_11.1.067, and the 32-bit compiler (Update 7) for Linux, l_cprof_p_11.1.073.

Starting with the longer test program reported a few days ago by Yaqi Wang, I merged and simplified his code and removed all real variables to obtain the reproducer given below.

The bug is encountered with the compiler option -O3, but not with -O2, -Ot, -Od or -fast. The correct output is -1 repeated 5 times. The incorrect output, produced with -O3, is 1 repeated 5 times.

The bug is sensitive to changes in the source code. Activating the WRITE statement on Line-54, or merging subroutines INIT and IDENT make the bug disappear. Similarly, replacing qrt(1) by qrt throughout the program makes the bug go away, as does replacing qrt(1)%a(1:qm) by qrt(1)%a on Line-55.

[fortran]MODULE DataKind
  IMPLICIT NONE
  INTEGER, PARAMETER :: zero =  0
  INTEGER, PARAMETER :: one  =  1
  INTEGER, PARAMETER :: mOne = -1
END MODULE DataKind

MODULE SparseMatrix
  USE DataKind
  IMPLICIT NONE

  TYPE CSR
    INTEGER :: n
    INTEGER, POINTER :: a(:) =>null()
  END TYPE CSR

 CONTAINS

  SUBROUTINE init(a, n)
    IMPLICIT NONE
    TYPE(CSR) :: a
    INTEGER :: n
    ALLOCATE(a%a(n))
    a%n = n
    a%a(1:n) = zero
    RETURN
  END SUBROUTINE init

  SUBROUTINE ident(n, a)
    IMPLICIT NONE
    INTEGER :: n
    TYPE(CSR) :: a

    CALL init(a, n)
    a%a(1:n) = one
    RETURN
  END SUBROUTINE ident

END MODULE SparseMatrix

MODULE quad1
  USE DataKind
  USE SparseMatrix
  IMPLICIT NONE
  INTEGER, PARAMETER :: qm = 5

  TYPE(CSR), SAVE :: qrt(1)

 CONTAINS

  SUBROUTINE quad()
    IMPLICIT NONE
    CALL ident(qm, qrt(1))
!   write(*,'(1x,10I5)')qrt(1)%a !                          <====== activate to make optimizer bug go away
    qrt(1)%a(1:qm) = qrt(1)%a(1:qm) * mOne
    RETURN
  END SUBROUTINE quad

END MODULE quad1

program test
  use quad1
  implicit none

  call quad()
  write(*,'(1x,10I5)') qrt(1)%a
  stop

end program test[/fortran]

[Kevin_D_Intel]

Greatly appreciate the time/effort to investigate and isolate this issue into a convenient reproducer, mecej4. I will test with some internal developmentcompilers and update the post when I have more information.