S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt. Recovery requirements of branch prediction storage structures in the presence of mispredicted-path execution. International Journal of Parallel Programming, 25(5):363--383, Oct. 1997.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....time, and HydraScalar models multiple layers of misprediction with full detail. Branch predictor updates take place at commit, but the stack and branch history are updated speculatively at fetch time with suitable repair mechanisms if the speculative updates are later found to be incorrect [18], 21] 29] 31] Branch mispredictions suffer at least a 7 cycle latency, but branches whose direction is correctly predicted and merely miss in the BTB suffer only a 2 cycle penalty. Indirect branch mispredictions suffer the full minimum 7 cycle latency. The predictor we model makes a ....

S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt. Recovery requirements of branch prediction storage structures in the presence of mispredicted-path execution. International Journal of Parallel Programming, 25(5):363--83, Oct. 1997.

....almost certain. For example, after a fork, both paths might encounter calls to printf( Both push a return address, even though only one return address belongs on the stack (only one call to printf( eventually commits) Neither mechanisms for repairing a returnaddress stack after mis speculation [10, 22] nor per path copies of the top of stack pointer can prevent this sort of corruption. Per path copies of the entire return address stack are the best solution. Multipath execution already requires path contexts; the return address stack is merely an additional element in the path context. Copying ....

....more detail in [22] 4.2 Speculative History Update and Fixup In addition to update timing for the return address stack, update timing also matters for the branch history in two level predictors. This was shown by Hao, Chang, and Patt in [6] A subsequent paper by Jourdan, Stark, Hsing, and Patt [10] also showed that update timing does not matter much for the table of two bit counters, because the two bit counters provide adequate hysteresis. The problem for branch history is that if the history is updated when branches resolve at the end of the execution stage, a long time may elapse ....

[Article contains additional citation context not shown here]

S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt. Recovery requirements of branch prediction storage structures in the presence of mispredicted-path execution. Int'l J. Parallel Programming, 25(5):363--83, Oct. 1997.

....and this variability makes speculative history update necessary (see Section 3. 1) Their results consider a GAp 1 scheme, a theoretical predictor that maintains global history but provides each branch with its own table of 2 bit counters (the pattern history table, or PHT) Jourdan et al. [15] confirm these results for a single gshare predictor, and briefly suggest some mechanisms for repairing histories that have been corrupted by mispredictions, primarily focusing on global history predictors. Their discussion mostly envisions storing necessary repair state in the instruction reorder ....

....Global history, Adaptive PHT, Per address PHT 2 SPECULATIVE UPDATES OF LOCAL AND GLOBAL BRANCH HISTORY:A QUANTITATIVE ANALYSIS allowed to be in flight at once. This detailed data substantially extends prior characterizations of speculative update s importance for global history based predictors [13, 14, 15], and to our knowledge is the first such published data for local history based predictors. Our results show that, as predictors grow larger, speculative update tends to matter less for global history predictors, but that it matters more for local history predictors. For both types of predictor, ....

[Article contains additional citation context not shown here]

S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt, "Recovery requirements of branch prediction storage structures in the presence of mispredicted-path execution," International Journal of Parallel Programming, vol. 25, pp. 363--83, Oct. 1997.

....different mechanisms for avoiding this corruption. Related work. The literature discusses procedure returns infrequently. The most relevant work, by Jourdan et al. 1 This paper uses stack and return address stack interchangeably, and never speaks of the actual program stack in memory. 259 [19], evaluates repair mechanisms for the return address stack as well as other branch prediction structures. As we do, they find stack repair to be necessary for effective return address prediction. They evaluate a self checkpointing mechanism that saves popped entries to avoid overwriting them with ....

....save an arbitrary number of return address stack entries this way; the extreme would be to checkpoint the entire return address stack each time a branch is predicted. Our results provide data for full stack checkpointing as an upper limit. The sophisticated scheme proposed by Jourdan, et al. [19], can have the effect of checkpointing the full stack, but requires a larger number of stack entries than the methods proposed here because it preserves popped entries. Each of these stack entries also requires more space, because in addition to storing the return address, a stack entry in this ....

[Article contains additional citation context not shown here]

S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt. Recovery requirements of branch prediction storage structures in the presence ff mispredicted-path execution. Int'l J. Parallel Programming, 25(5):363--83, Oct. 1997.

....are difficult to come by, it seems that early i.e. speculative history update is a recent innovation due to the fixup mechanisms required to undo corruption from incorrect updates. The importance of speculative update is discussed in [17] and mechanisms for implementing fixup appear in [24] and [52] The Alpha 21264 implements speculative update with fixup for the global history portion of its hybrid predictor [26] Many mispredicted indirect jumps are function returns. Since a function might be called from many different locations, a BTB often provides the wrong target for these ....

....and popped immediately after a call or return is fetched, i.e. speculatively in the fetch stage, so some pushes or pops may correspond to wrong path instructions. As with branch history, if the stack is not repaired, the wrong path pops or pushes may corrupt it, as Jourdan et al. have pointed out [24]. They propose a sophisticated, self checkpointing return address stack that saves popped entries to avoid overwriting them with future mis speculated pushes. So long as the stack does not overflow, this structure can return to any prior state. We have shown that simply saving the current ....

[Article contains additional citation context not shown here]

S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt. Recovery requirements of branch prediction storage structures in the presence of mispredictedpath execution. International Journal on Parallel Programming, 25(5):363--83, Oct. 1997.

....are di#cult to come by, it seems that early i.e. speculative history update is only a recent innovation because it requires fixup mechanisms to undo corruption from incorrect updates. The importance of speculative update is discussed in [6] and mechanisms for implementing fixup appear in [8] and [21] The only known example of speculative update appears in the Alpha 21264, which implements speculative update with fixup for just the global history portion of its hybrid predictor [10] Many mispredicted indirect jumps are function returns. Since a function might be called from many ....

S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt. Recovery requirements of branch prediction storage structures in the presence of mispredicted-path execution. International Journal of Parallel Programming, 25(5):363--83, Oct. 1997.

....are di#cult to come by, it seems that early i.e. speculative history update is only a recent innovation because it requires fixup mechanisms to undo corruption from incorrect updates. The importance of speculative update is discussed in [33] and mechanisms for implementing fixup appear in [44] and Chapter 5. The only known example of speculative update appears in the Alpha 21264, which implements speculative update with fixup for just the global history portion of its hybrid predictor [49] Many mispredicted indirect jumps are function returns. Since a function might be called from ....

....cycle by cycle simulations showing that the number of outstanding branches varies, and this variability makes speculative history update necessary. Their results consider a GAp scheme, a theoretical predictor that maintains global history but provides each branch with its own PHT. Jourdan et al. [44] confirm these results for a single gshare predictor, and briefly suggest some mechanisms for repairing histories that have been corrupted by mispredictions, primarily focusing on global history predictors. This discussion primarily envisions storing necessary repair state in the instruction ....

[Article contains additional citation context not shown here]

S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt. Recovery requirements of branch prediction storage structures in the presence of mispredicted-path execution. International Journal of Parallel Programming, 25(5):363--83, Oct. 1997.

....certain. For example, after a fork, both paths might encounter calls to printf( Both push a return address, even though only one return address belongs on the stack (only one call to printf( eventually commits) Neither mechanisms for repairing a returnaddress stack after mis speculation [7, 12] nor per path copies of the top of stack pointer can prevent this sort of corruption. Per path copies of the return address stack are the best solution [12] Multipath execution already requires path contexts; the return address stack is merely an additional element in the path context, and ....

S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt. Recovery requirements of branch prediction storage structures in the presence of mispredicted-path execution. Int'l J. Parallel Programming, 25(5):363--83, Oct. 1997.

No context found.

S. Jourdan, J. Stark, T.-H. Hsing, and Y. N. Patt. Recovery requirements of branch prediction storage structures in the presence of mispredicted-path execution. International Journal of Parallel Programming, 25(5):363--383, Oct. 1997.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC