The current implementation of the instruction cache only services requests when not doing a refill. This introduces unneeded latency: for example, an instruction cache could return the first instruction as soon as it becomes available. Refilling starting from the requested address would ensure that the additional latency introduced by the cache is minimal.