That's not quite true... the code (essentially) loops over itself infinitely; the only branch is a loop back to the beginning of the program. Since this branch has only one target, it can be a direct jump (as opposed to an indirect one), meaning it would not use the branch trace buffer, and would not cause speculative execution. The author used faulting for branches only to write the program in all mov instructions; if the last 'mov' is replaced with a direct jmp (which, it looks like, can be done with the --no-mov-loop flag to the compiler), no more speculative execution.

That's a pity, I was hoping that this could be converted to a SIMD architecture allowing you to run hundreds of staggeringly slow doom instances simultaneously on a GPU.

But the faulting mov instruction is the last instruction in the program. There are no instructions after it to speculatively execute.

One of the main ideas in OoO processors is that the processors are in many ways like dataflow machines, waiting for inputs to become available in order to start (potentially speculative) execution.

So, encoding logic in dependent movs does not inherently address the cache side channel issue.

History based branch prediction one difference with with explicit branches of course.