The model
Branches make program control depend on data. A modern CPU predicts likely directions so it can keep fetching instructions before the condition is fully resolved. Correct predictions preserve pipeline flow; incorrect ones require discarding speculative work and refilling parts of the pipeline.
A concrete walk-through
A loop over sorted values may have a highly predictable branch, while a branch on randomized values may be difficult to predict. Some code can be expressed with conditional operations or grouped data to reduce unpredictable control flow, but those rewrites can increase instruction count or complicate correctness.
Costs and failure cases
Instruction-level parallelism is limited by dependencies, branches, and memory stalls. Removing a branch is not automatically faster; the new operations may cost more or prevent vectorization. Benchmark realistic input distributions and check that the compiler generated the intended code.
Check your understanding
You have a filter with a branch that is true for almost every element. Explain why that may perform differently from a branch that is true half the time, and name one measurement you would collect.