Labor of Division (Episode III): Faster Unsigned Division by Constants

As is well known (and seen in a previous post), compilers optimize unsigned division by constants into multiplication by a "magic number." But not all constants are created equal, and approximately 30% of divisors require magic numbers that are one bit too large, which necessitates special handling (read: are slower). Of these 30%, slightly less than half (46%) are even, which can be handled at a minimum of increased expense (see below); the remaining odd divisors (659 million, including well known celebrity divisors like 7) need a relatively expensive "fixup" after the multiplication. Or so we used to think. This post gives a variation on the usual algorithm that improves performance for these expensive divisors.