A load executed past an older store whose address is unresolved

An out-of-order core encounters, in program order, a store to address *p whose address isn't fully resolved yet (p is itself the result of a not-yet-complete computation), immediately followed by a load from *q (q's address ready sooner). Rather than stall the load until the older store's address is known and definitively ruled out as aliasing, the core's memory disambiguation predictor speculatively guesses that the load does not alias the pending store, and lets the load execute early against memory/cache. What happens if p and q later turn out to actually be equal (a true RAW dependency the core bet against)?

30 of 47 in Computer Architecture