@regehr Leave things implementation-defined. Let the x86s have their segmented memory models. Let the source code be machine and implementation specific. So what? We already know how to #ifdef for porting between different OSs and platforms, and the need for adaptation to different platforms will never go away. So wishing for perfect source portability on an ideal abstract machine is futile anyway. It's an invalid justification for these definedness rules.

@regehr The concept of definedness isn't even sensible. Saying the compiler is free to do anything it wants for undefined behavior is bogus. It can't take a division by zero sequence and decide to emit a completely unrelated code sequence instead. The result of that arithmetic is undefined, but it still has to issue the instruction. It may not know at compile time that the divisor will be zero, so it has to just emit the code as requested.

@regehr phooey. I'll stick to C89 then, before they went overboard with the abstract machine. We don't have CPUs with 1s complement math any more, nor with 9 bit bytes. We don't have to worry about invisible bits in 36bit integers. The compiler's job is to turn C code into assembly code. If it's gratuitously setting tags into the low bits of pointers, that the programmer didn't ask for, it's a broken compiler.