1' Improving Branch Predictors by Correlating on Data Values 2' A Language for Describing Predictors

1
1. Improving Branch Predictors by Correlating on
Data Values2. A Language for Describing
Predictors and its Application to Automatic
Synthesis
2
First Generation
Feedback
PC
history
Prediction
PC CondJmpEq Ra, Rb, Target
3
Improvement 1
Feedback

history
PC CondJmpEq Ra, Rb, Target
4
Improvement 2

history
Prediction

history
PC
Feedback
PC CondJmpEq Ra, Rb, Target
5
Improvement 3
PC CondJmpEq Ra, Rb, Target
Reduce table interference through more
intelligent table indexing scheme.
6
Input to the Predictor

• PC, GBH
• What else?
• PC CondJmpEq Ra, Rb, Target
• Ra, Rb?
• What is the main obstacle to this strategy?
• Solution predict the values of Ra and Rb.
• Still impractical?
• Solution predict the difference between Ra and
  Rb.
• Still impractical?
• Solution store only a few bits of the difference
  values.

7
Branch Difference Predictor
8
Rare Event Predictor
REP
PC
Prediction
value pattern
counter
GBH

value pattern
VHT
PC CondJmpEq Ra, Rb, Target
9
Value Pattern
Value Pattern Branch count, Branch difference
10
Experimental Results

• REP
• 6K bytes
• VHT
• BCT, 64 bytes
• BDC, 100 bytes

11
(No Transcript)
12
(No Transcript)
13
Discussion

• BDP can be used with other predictors.
• Why?
• Significant improvement?
• BDP removes 13 to 9 of the misprediction over
  gShare.
• and 12 to 8 of the misprediction over Bi-Mode.
• 1 of overall improvement?
• Can BDP be applied to any architecture?
• Set instructions in the MIPS-based SimpleScalar
• BEQ Ra, Target in the Alpha family.