Pipeline Architecture - 2024.2 English

MicroBlaze Processor Reference Guide (UG984)

Document ID
UG984
Release Date
2024-11-27
Version
2024.2 English

MicroBlaze instruction execution is pipelined. For most instructions, each stage takes one clock cycle to complete. Consequently, the number of clock cycles necessary for a specific instruction to complete is equal to the number of pipeline stages, and one instruction is completed on every cycle in the absence of data, control, or structural hazards.

A data hazard occurs when the result of an instruction is needed by a subsequent instruction. This can result in stalling the pipeline, unless the result can be forwarded to the subsequent instruction. The MicroBlaze GNU Compiler attempts to avoid data hazards by reordering instructions during optimization.

A control hazard occurs when a branch is taken, and the next instruction is not immediately available. This results in stalling the pipeline. MicroBlaze provides delay slot branches and the optional branch target cache to reduce the number of stall cycles.

A structural hazard occurs for a few instructions that require multiple clock cycles in the execute stage or a later stage to complete. This is achieved by stalling the pipeline.

Load and store instructions accessing slower memory might take multiple cycles. The pipeline is stalled until the access completes. MicroBlaze provides the optional data cache to improve the average latency of slower memory.

When executing from slower memory, instruction fetches might take multiple cycles. This additional latency directly affects the efficiency of the pipeline. MicroBlaze implements an instruction prefetch buffer that reduces the impact of such multi-cycle instruction memory latency. While the pipeline is stalled for any other reason, the prefetch buffer continues to load sequential instructions speculatively. When the pipeline resumes execution, the fetch stage can load new instructions directly from the prefetch buffer instead of waiting for the instruction memory access to complete.

If instructions are modified during execution (for example with self-modifying code), the prefetch buffer should be emptied before executing the modified instructions, to ensure that it does not contain the old unmodified instructions.

MicroBlaze also provides the optional instruction cache to improve the average instruction fetch latency of slower memory.

All hazards are independent, and can potentially occur simultaneously. In such cases, the number of cycles the pipeline is stalled is defined by the hazard with the longest stall duration.