In FIPS 186-3 appendix D.2 "implementation of Modular Arithmetic", they show shortcuts for solving $$B = A \mod m$$
for select Curves.
How would you go about determining a valid short cuts for the various curves? If I wanted to practice on using a small finite field (e.g. 23, 31) because it's something that could be solved by hand. And I wanted to use one of the mentioned shortcuts, how would I go about deriving it?
These prime numbers are called Solinas primes (because they were described by Jerome Solinas). The article details how they are found and how optimization works for them.
As a brief summary, consider a prime: $$p = \sum_{i=0}^{k} b_i 2^{iw} $$ where each $b_i$ is either $0$, $1$ or $-1$, and $w$ is your "word size" (typically $w = 32$ or $64$ will yield the best results on usual computers). You also want $b_0 \neq 0$ ($p$ must be prime, so in particular it must be odd) and $b_k = 1$: the prime value is close to $2^{wk}$. To make things even easier to implement, make sure that $p$ is slightly below $2^{wk}$ rather than slightly higher, meaning that the highest non-zero $b_i$ (for $i < k$) has value $-1$.
With such a prime, modular reduction is easy because, modulo $p$, you have: $$ 2^{wk} = \sum_{i=0}^{k-1} -b_i 2^{iw} \pmod p $$ from which you can infer: $$ -x 2^{w(k+j)} = \sum_{i=0}^{k-1} b_i x 2^{w(i+j)} \pmod p $$ for all integers $x$ and $j$. Therefore, if you have a $n$-word value $X$ (with $n > k$): $$ X = \sum_{i=0}^{n-1} x_i 2^{wi} \mathrm{\ \ \ \ \ \ where\ } 0 \leq x_i < 2^w $$ then you can compute the value $X'$: $$ X' = X - x_{n-1} 2^{w(n-1)} + \sum_{i=0}^{k-1} b_i x_{n-1} 2^{w(n-1-k+i)} $$ This addition is easily computed because all the $b_i$ are $0$, $1$ or $-1$, so this amounts to adding or subtracting word $x_{n-1}$ to some other, lower words. With the relation above, you have $X' = X \pmod p$. But $X'$ now consists of $n-1$ words. So, with a few addition and subtractions, you have removed one word from your problem. Iterate.
Whether Solinas primes really give that much an advantage over a random prime (with Montgomery multiplication) is disputed. It depends on the implementation architecture. It has also been argued that choosing $p = 2^m - z$ for a small $z$ yields more efficient computations. As for all performance things, there is no absolute answer; it must be tried and measured, and any new language, compiler, processor version or architecture can change the answers.
