Store Vulnerability Window (SVW): A Filter and Potential Replacement for Load Re-Execution

Load scheduling and execution are performance critical aspects of dynamically-scheduled proce Several techniques employ speculation on loads with respect to older stores to improve some aspect processing. Speculative scheduling and speculative indexed store-load forwarding are two examples Speculative actions require verification. One simple mechanism that can verify any load specula in-order re-execution prior to commit. The drawback of load re-execution is data cache bandwidth con tion. If a given technique requires a sufficient fraction of the loads to re-execute, the resulting contentio severely compromise the intended benefit. Store Vulnerability Window (SVW) is an address-based filtering mechanism that significantly red the number of loads that must re-execute to verify a given speculative technique. The high-level idea is load need not re-execute if the address it reads has not been written to in a long time. SVW realizes th using a store sequence numbering scheme and an adaptation of Bloom filtering. An SVW implemen with a 1KB filter can reduce re-executions by a factor of 200 and virtually eliminate the overhead of re cution based verification. The same SVW implementation can be used as a complete replacement for cution with only 3% overhead.