To me the real problem is that compilers these days do too much. Nobody asked for these optimizations, if you want optimized code, use C++, or Rust, or whatever

To me the real problem is that compilers these days do too much. Nobody asked for these optimizations, if you want optimized code, use C++, or Rust, or whatever other "modern" language that has higher level constructs and thus can optimize the code better. The reason to use C is to have an "high level assembler", and there is no reason to use C if I can't rely on the output of the compiler, and if the compiler eliminates some code that I write. We all know that happens when a signed integer overflows, we all know that most architectures are little endian, etc. Even if something is "undefined behaviour" for the standard, it usually has some well defined behaviour on the platform that you are targeting. Unfortunately, for these reasons gcc is dangerous to use, it's not a problem of desktop application (if it crashes, who cares), but I talk about critical systems. It would be unacceptable if a machine ends up killing someone because at gcc they though that they could remove a check that they thought it was useless. And so you have to spend thousand of $ just to get reliable compilers that doesn't do silly optimizations. I think that they should make a sort of "gcc lite" that does exactly what (at least to me) a C compiler has to do: translate the code from C to assembly, without changing the semantic of the code. If there is an undefined behaviour that behaviour will be translated to the assembly code and you will let the processor to deal with it. Also, we talk about optimizations that could be made by exploiting undefined behaviour, fine, but also the programmer usually optimizes exploiting the same undefined behaviours by hand.