Very true and worth evangelizing to others. I have unknowingly violated -fstrict-aliasing in some part of my code only to discover later that it is benign at -O0 and metastasized at -O3.
This freaks me out the most in networking code, where there is all kinds of casting of structs (esp. if you blindly copy-paste examples from StackOverflow) and performance usually matters. Rust has inspired me to take more time to profile C code to see whether strict aliasing (strict overflow, etc.) actually make a significant enough improvement to merit the UB-risk, review time, and acid in the stomach.
This freaks me out the most in networking code, where there is all kinds of casting of structs (esp. if you blindly copy-paste examples from StackOverflow) and performance usually matters. Rust has inspired me to take more time to profile C code to see whether strict aliasing (strict overflow, etc.) actually make a significant enough improvement to merit the UB-risk, review time, and acid in the stomach.