Branch target prediction

• Predicting whether a branch is taken is not much use in DLX because we have already moved the test and target address calculation into ID
• There is still a one-cycle branch delay
• The nullifying delayed branch idea uses static prediction
• How can we avoid the delay when dynamic prediction is correct?
• Problem:

  

  • Need to know the address from which to load the next instruction
  • Before the branch has been decoded!