At the office, we frequently do a scrambled word puzzle from the newspaper. Occasionally I overlook the answer for minutes at a time. In the past I've used various websites or programs to find solutions to these problems, but now I've written my own very fast one from scratch. It includes commandline and web interfaces (including ajax for lightning-fast updates).

You can try it or get the source.

I don't stake any claim to being "the fastest" anagrammer, but my program is pretty fast. I credit its speed to two main factors: first, a data structure that makes it very fast to exclude a word from being an anagram solution; and second, a binary dictionary format that trims 95% of the loading time compared to a text dictionary.

The data structure consists of two parts. First, a 32-bit integer in which each bit is 1 if the corresponding letter appears in the word and 0 otherwise; and second an array of 26 8-bit integers which record the actual number of occurrences of the letter. This makes the quick-exclusion test for 'word B can be spelled from letters A' simply (A.mask & B.mask) == B.mask, a quick operation to perform. If this step doesn't show that it's impossible to spell B with the letters from A, then 26 more comparisons are done to see whether for each letter that is in B, A has enough copies of that letter to suffice.

As a result, it's possible to generate 1000 anagrams of a typical-length input in about 140 million CPU cycles, or around .07 seconds per input. Using 'server mode' and limiting the number of results improves things a bit, but the initial fitering of the (typically ~90k word) list dominates.

All in all, a fun project (which should be just about done now!)