This email thread is driving me nuts because so much of it is wrong. *Speculative execution has nothing to do with caching.* Caching is about keeping often used data as close to the CPU execution unit as possible so that the CPU execution unit doesn't stall for lack of data. But the cache is clearly outside of the execution unit.
Some CPUs will speculatively prefetch memory through the cache when they predict that memory will be in use soon. For example, if you are running through a program in a straight line loading the next 32 or 128 bytes of memory ahead of time is a no brainer. Of if you are operating on a large data structure the cache subsystem might even decide to start loading sequential memory by itself. But that has nothing to do with the speculative execution inside of a CPU execution unit. *Pipelining in a CPU is not a form of speculative execution.* Pipelining is a method of introducing parallelism in a CPU's execution unit. Instead of having large functional blocks of logic that run at relatively slow clock speed you break those up into smaller units, allowing each of them to do less work and complete their task more quickly. And that partial work gets passed to the next unit of the pipeline. Pipelining was essential for enabling higher clock speeds on chips. There is nothing speculative about pipelining by itself. Branch prediction is what I think you are referring to. The CPU can monitor individual branch behavior and make a good guess about what the branch instruction will do the next time it is seen. When combined with pipelining it allows the pipeline to stay full when it sees a branch instead of just pausing (stalling) and waiting to see the outcome of the branch. If the CPU predicts incorrectly then the partial results in the pipeline have to be purged and started again. That is mildly speculative, but the speculative part is due to the branch prediction, not the pipelining. *The prefetch instruction queue that you refer to on the 8088 is a part of every CPU architecture, and it was not replaced by a cache on the Pentium.* Every CPU needs to have a functional part that fetches the next instruction and feeds it to the next step of the execution pipeline. The memory subsystem (often a cache being the first thing the CPU interfaces with) is entirely different. -Mike
_______________________________________________ Freedos-devel mailing list Freedos-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/freedos-devel
