Hans Boehm’s interpretation is much, much worse than the one used by compiler vendors: if the termination mentioned in the second requirement is supposed to refer to termination of the actual program, then a conforming implementation is permitted to compile any C code into any machine code it likes, so long as the first and third requirements are met and the resulting program does not terminate.

I think the more interesting question here is when a compiler is permitted to move instructions across a loop. If a loop has no side effects, is the compiler permitted to move a side-effecting instruction from after the loop to run before the loop? Here I’m assuming a loop such that the compiler can not prove how many iterations will run. It’s easy to imagine that in some cases moving the instruction would give better instruction scheduling and a better result over all if it happens that the loop does not run for long. However, moving the instruction will give surprising results if the programmer is writing a simple-minded delay loop on an embedded system.

To put it another way, in a real program that is not using loops for timing or to wait for some sort of signal, moving the instruction is on average better. But there are unusual but plausible programs which will break if the instruction is moved.

I think the standard permits the instruction to be moved. These issues do not arise in most other languages which run farther from the hardware.