DES SBOX Output with Bitslice

Eli Biham's original algorithm to implement any 6 to 4 bit S-box as described in Matthew Kwan's paper is to

• Single out two input bits, say $i_1$ and $i_2$
• Build all $2^{(2^2)}=16$ single-bit functions of $i_1$ and $i_2$, say $f_0$ to $f_{15}$
• Describe each of the four outputs of the S-box as which of these functions $f_j$ needs to go to that output for each of the $2^4=16$ combinations of the four other input bits $i_3$ $i_4$ $i_5$ $i_6$ of the S-box, and implement that by using four layers of digital multiplexing for each output:
  • For each of the $2^3=8$ combinations of $i_4$ $i_5$ $i_6$, we select according to $i_3$ which $f_j$ is needed. E.g. if for a certain output and certain combination of $i_4$ $i_5$ $i_6$ we need to select $f_4$ when $i_3=0$ and $f_7$ when $i_3=1$, then we can do this as $(f_4\operatorname{NAND}\bar{i_3})\operatorname{NAND}(f_7\operatorname{NAND}i_3)$, costing $3$ gates (discounting cost of inverting $i_3$). Thus this stage will cost $4\times8\times3=96$ gates total (but see optimization 1 below).
  • For each of the $2^2=4$ combinations of $i_5$ $i_6$, we select according to $i_4$ which of two functions of the earlier stage is needed.
  • For each of the $2$ values of $i_6$, we select according to $i_5$ which of two functions of the earlier stage is needed.
  • We select according to $i_6$ which of the two functions of the earlier stage is needed.

The above does the multiplexing with $4\times(8+4+2+1)\times3=180$ $\operatorname{NAND}$ gates (plus $4$ inverters for $i_3$ $i_4$ $i_5$ $i_6$ if these needs to be accounted for).

Many optimizations are possible, including:

1. Using $\operatorname{XOR}$ which allows a multiplexing with two gates/instructions instead of three, e.g. we compute $((f_4\operatorname{XOR}f_7)\operatorname{AND}i_3)\operatorname{XOR} f_4$, noting that $f_4\operatorname{XOR}f_7$ comes for free since this is still a function of $i_1$ and $i_2$, thus an $f_j$, likely $f_3$ for some natural numbering; same for later multiplexing stages, by adjusting what earlier stages compute. This optimization is very effective in software. It's in Biham's implementation and in Kwan's account.
2. Computing $8$ rather than $16$ functions $f_j$, by adjusting polarity in the multiplexing.
3. On some occasions, reusing a function (beyond the $f_j$) across multiple S-box outputs.
4. On some occasions, not needing all functions $f_j$, because one happens not being used.
5. On some occasions, being able to remove a multiplexing stage, because the multiplexing input has no influence on the desired output.
6. On some occasions, being able to simplify a multiplexer because one of it's data input is constant.
7. Reordering things that can (the inputs $i_j$, the data inputs of multiplexers, the order of multiplexing bits $i_3$ $i_4$ $i_5$ $i_6$ for each output) to maximize occurrences of 3/4/5/6.