3.3: Reducing Branch Costs with Advanced Branch Prediction

Review of Branch Prediction from Appendix C

• Static Branch Prediction:
  • Run the program.
  • For each branch, record how often it is taken.
  • Use this info to predict that branch on later runs.
• Dynamic Prediction using prediction buffers
  • Use an array of size k, each containing a branch predictor.
  • A branch at address n uses the entry at position n mod k.
  • A 2-bit predictor only changes the prediction if branch does the same thing (taken or not) 2 times in a row.
• These predictors can keep misprediction rates below 7% for some SPEC benchmarks.

Correlating Branch Predictors

• A 2-bit predictor uses recent behavior of a single branch.
• Can do better if we also look at behavior of nearby branches.
• Correlating branch predictors are also called 2-level predictors.
• Use multiple 2-bit predictors, one for each combination of the results of the last n branches.
• Example: a (1,2) predictor uses the last (one) branch to choose one of 2 2-bit predictors.
• An (m,n) predictor uses the last m branches (2^m possibilities) to choose among 2^m n-bit predictors.
• These predictors can keep misprediction rates below 4% for some SPEC benchmarks.

Tournament Predictor

• Use multiple predictors (usually both local and global).
• Choose among them dynamically based on which predictors are working the best.
• These predictors can keep misprediction rates below 3% for some SPEC benchmarks.