-Mbounds reports false positive for array slices passed to DOT_PRODUCT intrinsic

[zaikunzhang]

This is a bug that affects all flang-based compilers: classic flang 7.0.1, Huawei Bisheng flang 1.3.3, NVIDIA nvfortran 22.3, and AOCC 3.2.0 flang. They raise a false positive error of out-bound subscripts when invoked with the -Mbounds flag.

Here is a minimal working example. The latest version of the code is available at my GitHub repo dedicated to testing Fortran compilers.

! test_solve.f90
module solve_mod
implicit none
private
public :: solve


contains


function solve(A, b) result(x)  ! A naive solver for upper-triangle linear systems
implicit none
real, intent(in) :: A(:, :), b(:)
real :: x(size(A, 2))
integer :: i, n

n = size(b) 

do i = n, 1, -1
    x(i) = (b(i) - inprod(A(i, i + 1:n), x(i + 1:n))) / A(i, i)
    !x(i) = (b(i) - dot_product(A(i, i + 1:n), x(i + 1:n))) / A(i, i)  ! No problem will arise
end do

end function solve


function inprod(x, y) result(z)
real, intent(in) :: x(:), y(:)
real :: z
z = dot_product(x, y)
end function inprod

end module solve_mod


program test_solve
use, non_intrinsic :: solve_mod, only : solve
implicit none

real :: A(1, 1), b(1)

A = 1.0
b = 1.0

write (*, *) solve(A, b)

end program test_solve

The error message looks like the following.

$ flang -Mbounds test_solve.f90 && ./a.out
0: Subscript out of range for array x (test_solve.f90: 19)
    subscript=2, lower bound=1, upper bound=1, dimension=1

Thank you for having a look at it.

[zaikunzhang]

To be clear, when i = n = 1, the Fortran standard says that A(i, i + 1:n) and x(i + 1:n) are valid and they should be empty arrays. So this is a false positive.

Note that the false positive occurs only if we invoke the compilers with -Mbounds.

For your reference, see what I reported to NVIDIA before in “Bug of nvfortran 22.2-0: array subscript triplet handled wrongly”, in case they are related. [Update 20230526: They turn out to be different bugs.]

[bryanpkc]

@mleair Do you have any information on the NVIDIA bug? Is that fix applicable to open-source Flang?

[bryanpkc]

@d-parks

[mleair]

The fix checks for an empty section under the A_SUBSCR case in collapse_assignment() in transfrm.c. The check looks like this:

case A_SUBSCR:
if (!contiguous_section(rhs) || is_empty_section(rhs))
      return 0;

Originally, the check looked like this:

case A_SUBSCR:
if (!contiguous_section(rhs))
      return 0;

The is_empty_section() function checks for a zero size array section.

-Mark

[bryanpkc]

@mleair Thanks a lot Mark! Would it be possible for you guys to upstream that fix? 😄

[mleair]

@bryanpkc Sorry, I am too busy right now to implement, test, and upstream the fix.

[mleair]

@bryanpkc I am available to answer any additional questions, though.

[zaikunzhang]

FYI, the problem remains in flang 15.0.3.

[zaikunzhang]

This bug is the same as the following one in nvfortran 22.3, which has been fixed now by 23.5.

https://forums.developer.nvidia.com/t/bug-in-nvfortran-22-3-false-positive-of-out-bound-subscripts/209936

N.B.: It is DIFFERENT from the following one, which was mentioned earlier by me in this thread. nvfortran 22.5 fixed the following one, but did not fix the one under discussion.

https://forums.developer.nvidia.com/t/bug-of-nvfortran-22-2-0-array-subscript-triplet-handled-wrongly/208395

[jschueller]

hello @mleair, would it be possible to post the is_empty_section function code ? I would like to try to upstream a patch

[bryanpkc]

I looked at this problem briefly. I am not sure that Mark's proposed solution is relevant. The crux of the problem is that Classic Flang lacks proper bounds checking for array sections; a triple expression is handled incorrectly as if it is just a single subscript. An extension of the complicated logic in the A_SUBSCR case in lower_ast() is needed.