jxbforce reads uninitialized bsubs(ns,:) in the js = ns pass

[CharlesCNorton]

bsubs in jxbforce is a local array (jxbforce.f90:46). The radial loop (jxbforce.f90:168) fills it for js = 2..ns-1 only:

! jxbforce.f90:116
  bsubs(1,:) = 0

! jxbforce.f90:171-173
     IF (js.gt.1 .and. js.lt.ns) THEN
        bsubs(js,:) = cp5*(bsubsh(js,:) + bsubsh(js+1,:))
     END IF

The pass at js = ns reads row ns anyway, at jxbforce.f90:184 (bs1=>bsubs(js,:), into fsym_fft for lasym) and jxbforce.f90:192 (bsubs_s(:) = bsubs(js,:)), and Fourier-transforms it at 233. The edge value is only assigned after the loop (jxbforce.f90:522).

Built with

-fcheck=all -finit-real=snan -ffpe-trap=invalid,zero,overflow

every case aborts there: input.cth_like_fixed_bdy at jxbforce.f90:233, the same input with LASYM = T at fsym_fft.f90:52. Rows ns of bsubsu, bsubsv, bsubuv and bsubvu are never consumed (the current loop is js = 2, ns1, jxbforce.f90:594), so the wout and jxbout are unchanged to roundoff when bsubs = 0 is added at entry; with that one line the instrumented build runs every case of the VMEC++ test set without a trap.

The 2017 VMEC2000 8.52 source has the same local array and the same loop, and there the post-loop rule bsubs(ns,:) = 2*bsubs(ns,:) - bsubs(ns-1,:) consumed the uninitialized row as well. The rule at jxbforce.f90:522 avoids that; the pass inside the loop still reads it.