Programming skills for optimizing program performance

After reading 《 Deep understanding of computer systems 》, Our programming is not just simple code . Instead, we should combine the foundation of the system to write more optimized code . We write code by understanding how computers work . Look at our code at a lower level , Perhaps a simple loop can change the overall efficiency of the code .

Modern compilers are complex , It does a lot of work for the program , The goal is to make the program safer and faster . For example, our compilation optimization options -Og -O1 -O2 -O3 etc. , But also due to optimization , Our machine instructions are arranged “ Beyond recognition ”, If you don't believe it, you can use the optimized code to disassemble , It will be much more than you think .

Here are a few guidelines that apply

1. Eliminate the inefficiency of the cycle , Never add extra calculations , What should be moved outside the circulation is moved outside the circulation

2. Reduce procedure calls , Of course, this is a double-edged sword , For example, the use of inline functions , But if you reduce it too much , It will make our code logic messy

3. Reduce unnecessary memory references , The best examples here are copy constructors and move constructors . After all, deposit access and cpu Compared with the execution cycle of , Too slow .

4. Try to improve cache shooting , This requires writing code , Try to put frequently accessed data into one cacheline in , Or do cacheline Alignment of , Here is another double-edged sword , It may waste a lot of space after alignment , After all cache Still valuable .

5. Loop unrolling , Use circular blocking , Reduce the number of iterations of the loop , Increase the number of elements calculated per iteration . Reduces the number of operations that do not directly contribute to program results , Provide some ways to optimize your code , Reduce the number of operations on the critical path in the whole calculation .

6. Improve parallelism , Contemporary computer systems are basically multi-core architecture , Different cores can process different data , Of course, the splitting of data here is very important , In particular, the correlation of data has a great impact on parallelism .

7. Write code suitable for conditional delivery , After all, branch prediction is miss The price is relatively high , Need to fill in the assembly line again .

8. Understand memory performance , Memory bus bandwidth , If you can use register variables, you don't need memory variables .

9. By using multiple accumulation variables and recombination techniques , Find ways to improve the parallelism of instruction sets .

10. Try to reduce cache line Of miss and cache The shaking of .

11. Disassemble the code of high-level language , Find ways to optimize from the assembly .