Local Stall Propagation.PDF

Microprocessors commonly encounter data hazards that must be resolved by stalling the pipeline until all inputs become available. The stall is broadcast globally to previous stages so no tokens are lost. As propagation delays of cross-chip wires exceed 1 cycle, it is no longer possible to stop the processor “on a dime.” Local Stall Propagation is a systematic approach to solving the problem. In such a system, each pipeline register contains storage for two tokens. On a stall, the pipeline retains its primary token, but accepts a secondary token from the previous stage that has not yet received the stall. An entire cycle is available to communicate the stall to the previous stage. Flip-flops, transparent latches, or pulsed latches may be modified to hold two tokens per stage with minimal additional hardware or latency and while maintaining scan capability. Valid bits provide a natural mechanism for automatic power reduction.