IME gcc is extraordinarily bad at anything but the most basic of auto-vectorizations, but its output from SIMD intrinsics is decent -- not usually quite as good

IME gcc is extraordinarily bad at anything but the most basic of auto-vectorizations, but its output from SIMD intrinsics is decent -- not usually quite as good as hand-rolled assembly, but much faster to write (for me anyway, YMMV) and still within 2x performance.