He did that because to measure the latency of a single instruction the "domain" (eg GP register, SIMD register memory, etc)) of the input(s) as to be the same as the output.
Stores have memory output but register/address input so it isn't really possible to measure the latency of a store in isolation. That's why Agners paired up loads and stores to get a latency figure and did the arbitrary split.
All that out of the way, it is possible to measure the latency of a load, at least for the load-to-address path like you mention with a pointer chasing load. That latency is 4 cycles minimum on all modern Intel. Add extra 1 cycle each for complex addressing modes or vector loads.
So yeah, a load latency of 2 is nonsense for modern Intel. Even talking about the latency is also nonsense since all the loads are independent so it's the throughput, not the latency that matters.
IIRC some chips have had 2 cycle L1 accesses: some P4 designs and one of the semi-recent POWER chips.
He does say he's reporting the minimum latency; I was thinking that no case would be faster than reading from L1 cache, but it makes sense that store-forwarding can be faster.
« According to Agner’s instruction_tables.pdf load instruction that I use has 2 cycles latency »
and indeed that's what Agner Fog's tables say (and they further say this figure is « the delay that the instruction generates in a dependency chain »).
But all other sources seem to agree that the L1 dcache latency on the main Intel processor family is 4 cycles (since Nehalem).
Does anyone know what the distinction is here? If I do a a simple pointer chase through L1 dcache I wouldn't make progress at 2 cycles per step, would I?