1. CS 7810 Lecture 8 Memory Dependence Prediction using Store Sets G.Z. Chrysos and J.S. Emer Proceedings of ISCA-25 1998

2. Lifetime of a Load

3. LSQ Basics Ld/StAddressDataCompleted StoreUnknown1000-- Loadx40000000-- Storex50000000-- Loadx50000000-- Loadx30000000-- An incomplete store stalls all future loads – No Speculation – the paper is overly conservative because it also waits for store values Most of these stalls are unnecessary – artificial dependences

4. Aggressive Approach Assume that loads do not conflict with earlier stores – all loads and stores execute out of order -- Naive Speculation When there is a conflict, the load behaves like a branch mispredict – all subsequent instructions are squashed and re-fetched  Expensive – 30-cycle penalty  Rename checkpoints for all instructions  Re-execute only the dependent instructions? – more complex, better performance

5. Ideal Model In the perfect model, loads only wait for conflicting stores – no artificial dependences and no memory-order violations

6. False Dependences and Violations

7. Store Sets Concept For every load, keep track of all stores that it has conflicted with in the past A load does not issue if members of its store set have not finished (dependences are introduced at the time of dispatch) The implementation is easy if  a load depends on only one store  a store is present in only one store set

8. Trivial Implementations Execution time normalized to an ideal store set implementation

9. Ideal Store Set Predictor An occasional memory-order violation can introduce many false dependencies – hence, use saturating counters

10. Implementation Overview Every ld/st depends on the last store in its set Causes serialized stores and false dependences st

11. Store Set Implementation Every load and store belong to one color – keep track of the last writer for each color – mpreds can pose problems Colors are merged as you discover m-o violations

12. Store Set Merging Store set merging improves performance by 12% Note that merging happens gradually – no need to instantly correct all entries in the table

13. Design Details Merging store sets To deal with occasional dependences and conflicts  clear the table every million cycles  use saturating counters for each entry The SSIT needs 4K entries and the LFST needs 128 entries

14. Results

15. Related Work Store barrier cache: identify stores that are likely to pose conflicts Keep track of all store-load conflict pairs and associatively check for dependences while dispatching instructions

16. Next Week’s Paper “Effective Hardware-Based Prefetching for High-Performance Microprocessors”, T.F. Chen and J.L. Baer, IEEE Transactions on Computers, May 1995

17. Title Bullet