CppCon 2025

Floating-point arithmetic is easy to write. But, when it goes wrong, it is much harder to debug. It's even harder to define what "wrong" means: a game has to be fast, but a financial report has to be reasonable for humans. This talk explores how different applications impose implicit requirements on floating-point behavior, what classical issues arise, how to work with them, and why there is sometimes no workaround.

We will start by looking at how floating-point operations are implemented in different C++ contexts, and the counter-intuitive implications that follow. We will see how different environments find subtly different balances between correctness, predictability, and performance. We will then examine code examples that produce subtly incorrect results or even UB, but are still easily fixable.

Next, we will dive into more complicated calculations and understand where we are restricted or helped by IEEE-754, C++, and computational mathematics. The talk will conclude with practical recommendations: how to define realistic expectations, how to test and structure your code, and how libraries can and/or cannot help you.

If you've ever been surprised by floating-point behavior in C++, this talk will tell you what's going on.