Speculative execution

We understand now how pipelining keeps the CPU busy and how, by predicting the results of conditional branches and executing the expected code speculatively, before we know for sure that it must be executed, we allow the conditional code to be pipelined. Figure 3.21 illustrates this approach: by assuming that the end of the loop condition is not going to happen after the current iteration, we can interleave the instruction from the next iteration with those of the current one, so we have more instructions to execute in parallel.

Sooner or later, our prediction will be wrong, but all we have to do is discard some results that should have never been computed in the first place and make it look like they were indeed never computed. This costs us some time, but we more than make up for it by speeding up the pipeline when the branch prediction is correct. But is this really all we have to do to cover up the fact that we tried to execute some code that doesn&apos...