CPU Pipelines and Hazards

Purpose

Explain classic and modern CPU pipeline concepts: stages, data and control hazards, forwarding/bypassing, stalls, and branch handling — foundational for optimization and understanding microarchitecture counters.

When to Use

Interpreting pipeline stall metrics from skills/profilers/intel-vtune-amd-uprof
Teaching why instruction order affects throughput
Relating assembly scheduling to hardware behavior
Debugging unexpected performance cliffs in hot loops

Workflow

1. Five-stage classic pipeline (MIPS-style mental model)

IF → ID → EX → MEM → WB

Overlapped execution: instruction N in EX while N+1 in ID.

2. Data hazards

Type	Example	Mitigation
RAW (true)	`add r1,r2,r3` then `sub r4,r1,r5`	Forwarding from EX/MEM/WB
WAR / WAW	Rare in in-order; relevant in OoO rename	Register renaming

Without forwarding:

stall until writeback completes

3. Control hazards (branches)

Branch in ID → target unknown until EX
├── Predict taken/not-taken (static or dynamic)
├── Flush wrong-path instructions on mispredict
└── Penalty = pipeline depth (varies by CPU)

See skills/computer-architecture/branch-prediction-and-speculation.

4. Structural hazards

Limited functional units (single memory port) cause stalls even without dependencies.

5. Practical optimization hints

/* Bad — tight dependency chain */
for (int i = 0; i < n; i++)
    acc = acc + data[i];   /* each iter waits on acc */

/* Better — multiple accumulators (ILP) */
acc0 = acc1 = 0;
for (int i = 0; i < n; i += 2) {
    acc0 += data[i];
    acc1 += data[i+1];
}
acc = acc0 + acc1;

Pair with skills/low-level-programming/cpu-cache-opt — memory often dominates.

6. Reading uops / ports (x86)

perf stat -e instructions,cycles,stalls-frontend,stalls-backend ./app

VTune "Microarchitecture Exploration" maps to pipeline slots.

7. Agent usage

/cpu-pipelines-and-hazards Explain RAW hazard in this ARM assembly loop and how to break it

Common Problems

Symptom	Cause	Fix
High `stalls-frontend`	I-cache misses / branch mispredict	Align hot loop; see branch skill
No speedup from unroll	Memory bound	Profile loads; prefetch
Wrong cycle model	Ignored OoO execution	Use perf hardware counters
"NOP fixes it"	Timing-sensitive MMIO	Never tune device delays by NOP

Related Skills

skills/computer-architecture/branch-prediction-and-speculation — mispredict cost
skills/computer-architecture/memory-hierarchy-and-caches — load latency
skills/low-level-programming/cpu-cache-opt — cache-line effects
skills/profilers/intel-vtune-amd-uprof — pipeline analysis
skills/low-level-programming/assembly-arm — instruction scheduling