How many instructions need to be killed on a miss-predict in a 6-stage scalar or superscalar MIPS?

我与影子孤独终老i 提交于 2021-01-27 14:31:32

问题


I am working on a pipeline with 6 stages: F D I X0 X1 W. I am asked how many instructions need to be killed when a branch miss-predict happens.

I have come up with 4. I think this because the branch resolution happens in X1 and we will need to kill all the instructions that came after the branch. In the pipeline diagram, it looks like it would require killing 4 instructions that are in the process of flowing through the pipeline. Is that correct?

I am also asked how many need to be killed if the pipeline is a three-wide superscalar. This one I am not sure on. I think that it would be 12 because you can fetch 3 instructions at a time. Is that correct?


回答1:


kill all the instructions that came after the branch

Not if this is a real MIPS. MIPS has one branch-delay slot: The instruction after a branch always executes whether the branch is taken or not. (jal's return address is the end of the delay slot so it doesn't execute twice.)

This was enough to fully hide the 1 cycle of branch latency on classic MIPS I (R2000), which used a scalar classic RISC 5-stage pipeline. It managed that 1 cycle branch latency by forwarding from the first half of an EX clock cycle to an IF starting in the 2nd half of a clock cycle. This is why MIPS branch conditions are all "simple" (don't need carry propagation through the whole word), like beq between two registers but only one-operand bgez / bltz against an implicit 0 for signed 2's complement comparisons. That only has to check the sign bit.

If your pipeline was well-designed, you'd expect it to resolve branches after X0 because the MIPS ISA is already limited to make low-latency branch decision easy for the ALU. But apparently your pipeline is not optimized and branch decisions aren't ready until the end of X1, defeating the purpose of making it run MIPS code instead of RISC-V or whatever other RISC instruction set.


I have come up with 4. I think this because the branch resolution happens in X1 and we will need to kill all the instructions that came after the branch.

I think 4 cycles looks right for a generic scalar pipeline without a branch delay slot.

At the end of that X1 cycle, there's an instruction in each of the previous 4 pipeline stages, waiting to move to the next stage on that clock edge. (Assuming no other pipeline bubbles). The delay-slot instruction is one of those and doesn't need to be killed.

(Unless there was an I-cache miss fetching the delay slot instruction, in which case the delay slot instruction might not even be in the pipeline yet. So it's not as simple as killing the 3 stages before X0, or even killing all but the oldest previous instruction in the pipeline. Delay slots are not free to implement, also complicating exception handling.)

So 0..3 instructions need to be killed in pipeline stages from F to I. (If it's possible for the delay-slot instruction to be in one of those stages, you have to detect that special case. If it isn't, e.g. I-cache miss latency long enough that it's either in X0 or still waiting to be fetched, then the pipeline can just kill those first 3 stages and do something based on X0 being a bubble or not.)


I think that it would be 12 because you can fetch 3 instructions at a time

No. Remember the branch itself is one of a group of 3 instructions that can go through the pipeline. In the predict-not-taken case, presumably the decode stage would have sent all 3 instructions in that fetch/decode group down the pipe.

The worst case is I think when the branch is the first (oldest in program order) instruction in a group. Then 1 (or 2 with no branch delay slot) instructions from that group in X1 have to be killed, as well as all instructions in previous stages. Then (assuming no bubbles) you're cancelling 13 (or 14) instructions, 3 in each previous stage.

The best case is when the branch is last (youngest in program order) in a group of 3. Then you're discarding 11 (or 12 with no delay slot).

So for a 3-wide version of this pipeline with no delay slot, depending on bubbles in previous pipeline stages, you're killing 0..14 instructions that are in the pipeline already.

Implementing a delay slot sucks; there's a reason newer ISAs don't expose that pipeline detail. Long-term pain for short-term gain.



来源:https://stackoverflow.com/questions/60781635/how-many-instructions-need-to-be-killed-on-a-miss-predict-in-a-6-stage-scalar-or

易学教程内所有资源均来自网络或用户发布的内容,如有违反法律规定的内容欢迎反馈
该文章没有解决你所遇到的问题?点击提问,说说你的问题,让更多的人一起探讨吧!