26.7. The More Things Change, the More They Stay the Same

You might expect that performance attributes of systems would have changed somewhat in the 10 years since I wrote the first edition of Code Complete, and in some ways they have. Computers are dramatically faster and memory is more plentiful. In the first edition, I ran most of the tests in this chapter 10,000 to 50,000 times to get meaningful, measurable results. For this edition I had to run most tests 1 million to 100 million times. When you have to run a test 100 million times to get measurable results, you have to ask whether anyone will ever notice the impact in a real program. Computers have become so powerful that for many common kinds of programs, the level of performance optimization discussed in this chapter has become irrelevant.

In other ways, performance issues have hardly changed at all. People writing desktop applications may not need this information, but people writing software for embedded systems, real-time systems, and other systems with strict speed or space restrictions can still benefit from it.

The need to measure the impact of each and every attempt at code tuning has been a constant since Donald Knuth published his study of Fortran programs in 1971. According to the measurements in this chapter, the effect of any specific optimization is actually less predictable than it was 10 years ago. The effect of each code tuning is affected by the programming language, compiler, compiler version, code libraries, library versions, and compiler settings, among other things.

Code tuning invariably involves tradeoffs among complexity, readability, simplicity, and maintainability on the one hand and a desire to improve performance on the other. It introduces a high degree of maintenance overhead because of all the reprofiling that's required.

I have found that insisting on measurable improvement is a good way to resist the temptation to optimize prematurely and a good way to enforce a bias toward clear, straightforward code. If an optimization is important enough to haul out the profiler and measure the optimization's effect, then it's probably important enough to allow as long as it works. But if an optimization isn't important enough to haul out the profiling machinery, it isn't important enough to degrade readability, maintainability, and other code characteristics. The impact of unmeasured code tuning on performance is speculative at best, whereas the impact on readability is as certain as it is detrimental.