Implementation:

• You must check that the prediction actually refers to the current PC (cf tag check in cache)
• BTB also includes branch prediction state bits, and is updated when branch prediction changes
• Must minimise cost of misprediction. E.g. also fetch the predicted-not-taken instruction
• Reduces branch delay to zero
  • Can also predict virtual function calls
  • Predicting function returns is harder; see H&P pp.277