Branchless programming is a technique that eliminates unpredictable CPU branch mispredictions by replacing conditional control flow with arithmetic. Using a filter-over-floats benchmark in Rust, the author demonstrates why a standard iterator-based filter is slowest at 50% selectivity on shuffled data — the branch predictor is essentially coin-flipping, causing ~500K pipeline flushes. By rewriting the filter to always write each element and use the boolean comparison result as a numeric index increment, the branch is converted from a control dependency to a data dependency. This makes the worst case nearly 4x faster and produces consistent timing regardless of data distribution. The trade-off: the best case (highly skewed data) gets slightly slower due to unconditional writes. The technique is recommended only for measured hot paths with unpredictable branches.
