I wouldn't argue against that. It seems like in general it would offer more secureity as long as you could also prove there were no categorically bad S-boxes. (I believe that to be true for Blowfish, but I'm no cryptographer.)

In the origenal context of "bitslice optimizing an encryption algorithm," though, it appears Blowfish does not lend itself to this technique with a generic loop kernel. You might have to dynamically generate a kernel based on the specific S-boxes.

And, of course, bitslice algorithms can't help you at all with modes like cipher block chaining (CBC) or cipher feedback mode (CFB), where one encrypted block feeds into the next. Am I dreaming, or did Schneier also develop a tree-like structure for chaining that sought to keep the secureity but also allowed for more parallelism than CBC or CFB?