Replace Fypp directional #:for blocks with runtime direction dispatch via struct components

[sbryngelson]

Background

PR #1432 consolidated directional triplets across several modules into derived type instances with %x/%y/%z components (is_muscl, bc, bf, q_prim_rs, etc.). This is load-bearing groundwork for replacing the most prevalent code-duplication pattern in the simulation target.

The pattern to eliminate

Nearly every directional sweep in m_muscl.fpp, m_cbc.fpp, m_rhs.fpp, and m_riemann_solvers.fpp currently looks like:

#:for MUSCL_DIR, XYZ, X_BND, Y_BND, Z_BND in &
        [(1, 'x', 'is_muscl%x', 'is_muscl%y', 'is_muscl%z'), &
         (2, 'y', 'is_muscl%y', 'is_muscl%x', 'is_muscl%z'), &
         (3, 'z', 'is_muscl%z', 'is_muscl%y', 'is_muscl%x')]
    if (muscl_dir == ${MUSCL_DIR}$) then
        ! 50–200 lines of physics, repeated verbatim 3×
    end if
#:endfor

Fypp generates three near-identical blocks at compile time. Any logic bug must be fixed in three places. Any new feature must be added three times.

What the struct consolidation unlocks

With the triplets now in structs, the direction-specific roles (primary, secondary, tertiary sweep axis) can be selected at runtime using associate or a value copy on the CPU path, or via a Fypp loop that expands to a single runtime select case on the GPU path:

! CPU initialization / finalize path — no GPU needed
associate(is_x => is_muscl%x, is_y => is_muscl%y, is_z => is_muscl%z)
    select case (dir)
    case (1); is_primary => is_muscl%x; ...
    case (2); is_primary => is_muscl%y; ...
    case (3); is_primary => is_muscl%z; ...
    end select
end associate

For GPU kernel bodies, the Fypp #:for loop remains, but the loop body collapses from three separately maintained copies to one template — because the struct member names (%x, %y, %z) can be the Fypp iteration variable directly.

Concrete targets

File	Pattern	Current lines	Target
m_muscl.fpp	muscl_dir == 1/2/3 blocks for order-1 reconstruction	~40	single block
m_muscl.fpp	muscl_dir == 1/2/3 in order-2 limiter body	~60	single template
m_cbc.fpp	cbc_dir == 1/2/3 sweep dispatch	~200	single sweep
m_rhs.fpp	norm_dir dispatch blocks	~150	single block

What this enables next

Once directional dispatch is unified, the per-direction allocation/initialization code can also collapse — s_initialize_muscl_module currently sets is_muscl%x/y/z in three separate guarded blocks; a loop over [is_muscl%x, is_muscl%y, is_muscl%z] with dimension limits [m, n, p] replaces all of it.

Related

• Builds on PR refactor: consolidate grid, stretch, and body force params into derived types #1432 (struct consolidation groundwork)
• Coordinate with Refactor m_riemann_solvers: reduce duplication, improve structure, add schemes #1426 (Riemann solver directional refactor)