Optimisation by search

Optimizers these days are pretty good. They can shuffle around opcodes far better than most mere mortals. The commercial packages that are currently available are probably going to be better at optimizing than anything anyone else could put together. I also don't think the "brute force" method is nessicary. Processors aren't random, they work on logic. If you have well defined rules, your optimizer should know what will be more efficient without testing it. A lot of those really fine detailes are processor specific too, so finding the "best case" isn't always that simple. Something might be a lot faster on a P4, but a lot slower on an Athlon, so a regular optimizer picks another option which lies in the middle ground.

Usually, you can get much bigger performance gains by modifying your algorithms on a higher level. Rather than worrying about what exact opcodes are used, and in what order they come, you're better off looking at if you really need to perform those operations at all. There are many different algorithms to perform the exact same task, some much more efficient than others. Sometimes, you can do with the data being in a different form too, such as dealing with squared values, rather than calculating the square root. Your time is far better spent on high-level optimizations.

There are a few noted exceptions to this rule. If you know you can make use of SIMD instructions, you'll have to code with them in mind. Also, eliminate branching inside loops where possible. The conditonal operator in C does give major performance bonuses if you're going to be executing that segment of code often.

Just something to look at:
The best log2 function we have is 5 clocks.

I'm pretty sure you could make an exhaustive search of that rather quickly.

From,
Nice coder