Value-based replay eliminates the need for content-addressable memories in the load queue, removing one barrier to scalable outof-order instruction windows. Instead, correct memory ordering is maintained by simply re-executing certain load instructions in program order. A set of novel filtering heuristics reduces the average additional cache bandwidth demanded by value-based replay to less than 3.5 percent.
Citation:
Harold W. Cain, Mikko H. Lipasti, "Memory Ordering: A Value-Based Approach," IEEE Micro, vol. 24, no. 6, pp. 110-117, Nov./Dec. 2004, doi:10.1109/MM.2004.81