SIGILL on Apple silicon

[Chengcheng-Xiao]

Has been reproduced on

OS/CPU: macOS Tahoe 26.5, M4 and M2 CPU
Compiler: GCC (gfortran) 15.2.0
OpenBLAS 0.3.32 (installed via macports, build formula)

and
OS/CPU: - / M2 CPU
Compiler: GCC (gfortran) 15.2.0
OpenBLAS 0.3.33 (installed via homebrew, build formula)

Minimum test code:

program test_zgetrf
    implicit none

    integer, parameter :: n = 4
    ! integer, parameter :: n = 3
    integer :: i, j, info
    integer :: ipiv(n)
    complex(kind=kind(1.0d0)) :: A(n,n), L(n,n), U(n,n), row(n)

    do j=1,n
       do i=1,n
          ! this should trigger a divide by zero
          ! A(i,j) = i+j
          ! this should be fine
          A(i,j) = mod(i+j, n)
       end do
    end do

    ! print *, "Original matrix A:"
    ! do i = 1, n
    !     print *, A(i,:)
    ! end do

    ! LU factorization
    call zgetrf(n, n, A, n, ipiv, info)

    if (info /= 0) then
        print *, "zgetrf failed with info =", info
        stop
    end if

    print *, "LU-factorized matrix A:"
    do i = 1, n
        print *, A(i,:)
    end do

    print *, "Pivot indices:"
    print *, ipiv

    L(:,:) = 0.d0; U(:,:) = 0.d0
    do j=1,n
       L(j,j) = 1.d0
       do i=1,j
          U(i,j) = A(i,j)
       end do
       do i=j+1,n
          L(i,j) = A(i,j)
       end do
    end do
    call zgemm('N', 'N', n, n, n, (1.d0, 0.d0), L, n, U, n, (0.d0, 0.d0), A, n)
    do i=n,1,-1
       row(:) = A(i,:)
       A(i,:) = A(ipiv(i),:)
       A(ipiv(i),:) = row(:)
    end do

    print *, "P.L.U="
    do i = 1, n
        print *, A(i,:)
    end do

end program test_zgetrf

When compiled with gfortran (15.2.0)

gfortran -ffpe-trap=zero -L/opt/local/lib -lopenblas test_zgetrf.F90

produces

Program received signal SIGILL: Illegal instruction.

Note that this only happens with -ffpe-trap=zero. Without it, everything works fine.

Switching to apple's accelerate framework then it works.

[martin-frbg]

This appears to be a misoptimization of the zgetf2 C code by clang - the error goes away when compiling with gcc, or when adding -g to the clang options

[Chengcheng-Xiao]

@martin-frbg Thanks! I can confirm this is a misoptimisation by clang. Adding CFLAGS='-O0' or swapping clang with gcc for the CMake build system fixes it. Will report this to macports and homebrew.