Pipeline数据冒险主要分为数据危害和控制危害。数据危害包括读写冲突（RAW），解决方法有检测并暂停或检测并转发。检测并暂停策略会在发现危害时使处理器暂停，直到危害消除，但会增加CPI。检测并转发则通过额外的数据路径和MUX控制来传递计算数据，减少暂停，但仍可能需要在某些情况下暂停，如加载指令后使用该寄存器的情况。
Pipeline Hazards
What can go wrong in pipelining?
- Data Hazards: Wrong data read before data is written.
- Control Hazards: Wrong instruction fetched before a taken branch.
- Exceptions
Data Hazards
read-after-write (RAW) dependences
add 1 2 3 //writes reg 3
nor 3 4 5 //reads reg 3
Two possible solutions.
Detect and Stall
If hazards exist, stall the processor until they go away.
-
Detect: Compare regA & regB with previous DestRegs
-
Stall:
- Keep current instructions in fetch and decode (IF/ID reg & pc unchanged)
- Pass a noop to execute
Time table for
最低0.47元/天 解锁文章
扫一扫
7231
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
