Value Profiling, Prediction and Influence of Compiler Optimizations

1
Value Profiling, Prediction and Influence of
Compiler Optimizations

• Amithash E. Prasad
• Jagan Annur Kannappan

2
Value Predictability

• 32 bit registers can represent over 4 billion
  values.
• Value Profiling suggests that values are
  predictable.
• What about compiler optimizations?
• Does value prediction add any value when compiler
  optimizations are turned on?

3
(No Transcript)
4
Load Value Profiling

• Load values were profiled and organized based on
  Instruction pointer and register names.
• These were fed into a database and queried to
  find useful information on value invariance.
• 10 of the SPEC2006 integer benchmarks were
  profiled (limited to 10 billion loads)

5
(No Transcript)
6
Value Prediction
7
Methodology

• PIN was used to instrument the benchmarks.
• The Instruction Pointer and the Write Value was
  passed to the Analysis Function
• 1024 entries with 3 way associative tables were
  used to store the last value and the stride (for
  strided predictor)
• Confidence based prediction (Selective Value
  Prediction)

8
Load Value Predictor

• Loads tend to be the most cycle consuming,
  stalling dependent instructions.
• Only Loads were instrumented.
• The value predictor was run on benchmarks
  compiled with O0, O2 and O3 levels of
  optimizations.

9
Load Value Predictor (Last Value)
10
Load Value Predictor (Strided)
11
Register Value Predictor

• Strided Value Predictor.
• All instructions except branch are instrumented
  for write values.
• Only instructions which write to GPRs are
  instrumented

12
Register Value Predictor (Strided)
13
Effective Address Predictor

• Predicts EA of instructions.
• Used to drive a hardware based Pre-fetcher.
• Mis-Predicts tend to clog memory bandwidth and
  does not cause immediate performance degradation.

14
Effective Address Predictor (Strided)
Pre-Fetcher
15
Conclusion

• Value Prediction can make or break performance.
• Hides cache misses on correct prediction.
• High prediction accuracies do not necessarily
  imply improved performance. (Application
  dependent).

16
Future Work

• Incorporate and study the effect on floating
  point registers.
• What if the correctly predicted value is anyway
  in the cache?
• Incorporate value prediction into simulators like
  PSatSim.