Hi, On 2022-06-27 18:12:13 +0700, John Naylor wrote: > Another thought: for non-x86 platforms, the SIMD nodes degenerate to > "simple loop", and looping over up to 32 elements is not great > (although possibly okay). We could do binary search, but that has bad > branch prediction.
I'd be quite quite surprised if binary search were cheaper. Particularly on less fancy platforms. - Andres
