R – Faster CompareText implementation for D2009


I'm extensively using hash map data structures in my program. I'm using a hash map implementation by Barry Kelly posted on the Codegear forums. That implementation internally uses RTL's CompareText function. Profiling made me realize that A LOT of time is spent in SysUtils CompareText function.

I had a look at the

Fastcode site

and found some faster implementations of CompareText. Unfortunately they seem not to work for D2009 and its unicode strings.

Now for the question: Is there a similar faster version that supports D2009 strings? The CompareText functions seems to be called a lot when using hash maps (at least in the implemenation I'm currently using), so little performance improvements could really make a difference. Or should the implementations presented there also work for unicode strings?

Best Solution

Many of the FastCode functions will probably compile and appear to work just fine in Delphi 2009, but they won't be right for all input. The ones that are implemented in assembler will fail because they assume characters are just one byte each. The ones implemented in Delphi will fare a little better, but they'll still return incorrect results sometimes because the old CompareText's notion of "case-insensitive" is based on ASCII whereas the new one should be based on Unicode. The rules for which characters are considered the same save for case are much different for Unicode from how they are for ASCII.

Andreas says in a comment below that Unicode CompareText still uses the ASCII case-comparison rules, so a number of the FastCode functions should work fine. Just look them over before using them to make sure they're not making any character-size assumptions. I seem to recall that some FastCode functions were incorporated into the Delphi RTL already. I have no idea whether CompareText was one of them.

If you're calling CompareText a lot in a hash table, then that suggests your hash table isn't doing a very good job. CompareText should only get called when the hash of the thing you're searching for designated a non-empty bucket in the hash table. From there, a hash table will often use a linear search to find the right item in the bucket, and it will call CompareText for every item during that search. I don't know whether that's how the one you're using works.

You might solve this by using a different hash function that distributes its results more evenly over the available buckets. If your buckets are already evenly filled, then you may need more buckets (and then make sure the hash function still distributes evenly over that number as well).

If the hash-map class you're using is based on TBucketList, then there is room for improvement in the bucket storage. That class doesn't calculate a hash on the entire input. It uses the input only to determine the bucket to use. If the class would also keep track of the full hash computed for a string, then comparisons during the linear search could go much faster. Just compare the hashes, and only compare the strings when the hashes match completely. (For a 256-bucket bucket-list, the largest supported size, only one byte of the input determines the bucket, and the rest of the bytes are ignored.) I've written about TBucketList here before.