Agner`s CPU blog

how could gather/scatter be implemented in hardware?

I am using a trick when the data to gather are not too distant from each other in memory: Read contiguous data into the largest available vector register, and then use a permute instruction to get the data into the desired positions in the vector. We should of cause have efficient permute instructions that can move data from any vector position to any other vector position. The indexes for permutations are provided in another vector register. An index out of range should produce a zero, so that larger permutes can be produced by ORing the results of multiple permute instructions.