It's a neat trick, but I think a linked list (with the very specific layout where nodes are allocated in order) is the only situation where this trick could possibly be useful?
And I think it only works if Spectre mitigations are disabled anyway?
What the trick does is replace sequential fetches (where each fetch address depends on the result of the previous fetch because, well, linked lists) with parallel fetches. It takes the minimum fetch-to-fetch latency from a L1 cache hit (roughly 3 cycles IIRC) to a cycle or less (most CPUs can do multiple parallel fetches per cycle).
If your data is stored in a vector or a B-tree, accesses are already parallel by default and you'll never need this trick.
And I think it only works if Spectre mitigations are disabled anyway?
What the trick does is replace sequential fetches (where each fetch address depends on the result of the previous fetch because, well, linked lists) with parallel fetches. It takes the minimum fetch-to-fetch latency from a L1 cache hit (roughly 3 cycles IIRC) to a cycle or less (most CPUs can do multiple parallel fetches per cycle).
If your data is stored in a vector or a B-tree, accesses are already parallel by default and you'll never need this trick.