Noop方法避免c＃编译器优化掉一个临时变量

Question

in a derivative calculation context I need to make sure x and x+h differ by an exactly representable number, so I have done this:

var temp = x + deltaX;
DoNothing(temp);
deltaX = temp - x;

DoNothing() method being there in order to prevent temp to be optimized away.

Is that correct way of handling this?

Answer 1

That compiler optimization you are worried about is invalid with .NET. It's known to be done in C++ with "fast math" compiler settings. .NET guarantees IEEE floating point behavior with the exception that excess precision is allowed (not a problem here). The optimization would be value-changing in general so it's forbidden.

If you want to clear compiler knowledge of a value I always use:

[MethodImplOptions(NoInlining)]
T F<T>(T x) { return x; }

This is not guaranteed to work but it does with all JITs that are in existence at the moment. The .NET JITs are so weak at optimizing that I doubt they will add the kind of data flow analysis that would make this trick stop working.