Improving the performance of C code

Question

What is the most unorthodox way to improve the performance of C code? This is no-holds-barred! Everything goes including changing loop structures to gotos, hardcoding anything and everything, using case statements in bizarre ways etc. Don't worry at all about maintainability, readability etc.

p.s. This is practical... and I'm well aware of how to improve the performance of code in reasonable ways (improve algorithms, profile before you optimize etc)

Answer 1

In my experience the most unorthodox way of optimizing C code is to profile the application, identify slowly performing structures and or db hits and then design reasonable solutions around them using Big O analysis.

Answer 2

Duff's Device is the canonical example. It's so weird that Tom Duff admitted, "This code forms some sort of argument in [the debate about fall-through in case statements], but I'm not sure whether it's for or against".

Answer 3

Use inline assembly?

Seriously though, if by merely changing the C code you can improve performance, chances are you can do it cleanly.

A few exceptions:

1) If you are relying on alignment semantics for pointers of different types often you can perform block operations on pointers that technically expose your application to a bounds overrun condition, but in practice does not because of your system's alignment characteristics. So a memory copy can be performed, by aligning initial char's then the inner block can be done using a long* pointer.

2) It may be possible to copy stack frames in clever ways if you know the memory order in which your compiler assigns local variables. This may allow you to implement co-routines which the language doesn't otherwise support. Coroutines are often a simpler and faster way of implementing some kinds of loop control.

3) Unions are always a bit "hacky" however you use them. Its a way of implementing polymorphism with fairly loose type checking.

4) Use of the C preprocessor as a way of auto-generating code is usually very difficult to debug and read. As such people tend to avoid this.