There were two major bugs with the predictor.
First the update mechanism was completely wrong. The PHT is updated with the GHR that was used to lookup the prediction. PHT[GHR] = Sat2(PHT[GHR], branch outcome).
Second the GHR needs to be updated speculatively as the branch is predicted. This is important so that back to back branches' GHRs are not the same. The must be different to avoid aliasing. Speculation of the GHR update allows them to be different. On mis prediction the GHR must be reverted.
This implementation is a bit sloppy with names and now the GHR recovery is performed. Updates to follow.
Added performance counters to individually track
branches; jumps, jump register, jal, and jalr; return.
jump and jump register are special cases of jal and jalr.
Similarlly return is a special case of jalr.
Also added counters to track if the branch direction was wrong,
btb target wrong, or the ras target was wrong.
Finally added one more counter to track if the BP incorrectly predicts
a non-cfi instruction.
Fixed bug with the instruction class.
Most tests now pass. Only Wally-JAL and the compressed instruction tests fail. Currently the bpred does not support compressed. This will be in the next version.
Fixed some typos with the names of signals in the branch predictor. They were causing signals to be not set. Note there is a modelsim flag which prevents it from compiling if a logic is undefined.
I will look this up and add it to the compiler.
