I don't think this should be the case on the type of out of order superscalar pipeline that this core has, as far as I'm aware of it. If there's a bind on waiting for memory an instruction should sit in a reservation station until its data comes in and that should take the same amount of time in either case. (In addition, with trivial accesses like this, it is highly likely a prefetcher would predict the access and have it fetched from cache in time. Though perhaps prefetch behavior works better in the longer loop?)

My intuition is that this is more likely to be behavior related to the trace cache and rayiner's comment points out several good suggestions on what might be happening here.

On the other hand, given the likely difference in how these instructions may be represented in the trace cache, there may also be interactions between the order and groups of instructions dispatched from the trace cache and that may change how things occupy different slots in reservation stations. Which may change the amount of instruction level parallelism in the loop. It doesn't seem likely, but it may be the case.

Trace cache shouldn't be the issue because the loop is simply executing too slowly. Even with very pessimistic alignment, HSW should get well more than one instruction per clock from that loop.

The culprit is probably store-to-load forwarding. Memory ops cannot wait until the coast is clear like data ops do because the addresses are not known early enough in the pipeline, so they are dispatched well before it's even known that there will be a store-to-load hit. The system to maintain the memory model is very complex, not well understood outside Intel, and might conceivably include optimistic optimizations. In such cases, it's completely feasible that issuing another store in the pipeline might reduce the amount of work that is rolled back and needs to be redone.

> Even with very pessimistic alignment, HSW should get well more than one instruction per clock from that loop.

Why? In Haswell, a store needs an AGU (ports 2, 3, or 7) and the store data unit (port 4). Even with perfect store-load forwarding, you can only write one value per cycle into the store buffer.

Instruction per clock, not iteration per clock. Each iteration of the tight loop takes 6 cycles there.