    1. store_src                      src = src.rgba
    2. store_dst                      dst = dst.rgba
    3. init_lane_masks                CondMask = LoopMask = RetMask = true
    4. copy_4_slots_unmasked          $0..3 = src
    5. copy_constant                  $4 = 0x3F800000 (1.0)
    6. copy_slot_unmasked             $5 = src(3)
    7. sub_float                      $4 -= $5
    8. swizzle_4                      $4..7 = ($4..7).xxxx
    9. copy_constant                  $8 = 0x3F800000 (1.0)
   10. swizzle_4                      $8..11 = ($8..11).xxxx
   11. copy_4_slots_unmasked          $12..15 = dst
   12. sub_4_floats                   $8..11 -= $12..15
   13. mul_4_floats                   $4..7 *= $8..11
   14. add_4_floats                   $0..3 += $4..7
   15. copy_4_slots_unmasked          [main].result = $0..3
   16. load_src                       src.rgba = [main].result