“Every developer should be able to recognize the opportunity and, most importantly, should not be scared of SIMD.”

SIMD has a reputation as black magic for compiler people. Hashimoto walks through the five-step shape that covers most of it: broadcast the constant, loop by vector width, operate, reduce, handle the remainder with scalar code. The Ghostty example gets 5x on AVX2. He also refuses to trust auto-vectorization, because a compiler update can silently turn your vector loop back into a scalar one and nothing will tell you.