Branch prediction

SimEng’s fetch unit is supplied with an instance of the abstract BranchPredictor class to enable speculative execution.

Access to the BranchPredictor is supported through the predict, update, and flush functions. predict provides a branch prediction, both target and direction, update updates an instructions’ prediction, and flush provides optional algorithm specific flushing functionality.

The predict function is passed an instruction address, branch type, and a possible known target. The branch type argument currently supports the following types:

• Conditional

• LoopClosing

• Return

• SubroutineCall

• Unconditional

The usage of these parameters within a branch predictor’s predict function is algorithm specific.

The update function is passed the branch outcome, the instruction address, and the branch type. From this information, any algorithms or branch structures may be updated.

Generic Predictor

The algorithm(s) held within a BranchPredictor class instance can be model-specific, however, SimEng provides a GenericPredictor which contains the following logic.

Global History

For indexing relevant prediction structures, a global history can be utilised. The global history value uses n-bits to store the n most recent branch direction outcomes, with the left-most bit being the oldest.

Branch Target Buffer (BTB)

For each entry, the BTB stores the most recent target along with an n-bit saturating counter for an associated direction. The indexing of this structure uses the lower bits of an instruction address XOR’ed with the current global branch history value.

If the supplied branch type is Unconditional, then the predicted direction is overridden to be taken. If the supplied branch type is Conditional and the predicted direction is not taken, then the predicted target is overridden to be the next sequential instruction.

Return Address Stack (RAS)

Identified through the supplied branch type, Return instructions pop values off of the RAS to get their branch target whilst Branch-and-Link instructions push values onto the RAS, for later use by the Branch-and-Link instruction’s corresponding Return instruction.

Static Prediction

Based on the chosen static prediction method of “always taken” or “always not taken”, the n-bit saturating counter value in the initial entries of the BTB structure are filled with the weakest variant of taken or not-taken respectively.

Perceptron Predictor

The PerceptronPredictor has the same overall structure as the GenericPredictor but replaces the saturating counter as a means for direction prediction with a perceptron. The PerceptronPredictor contains the following logic.

Global History

For indexing relevant prediction structures and for retrieving a direction from the perceptrons, a global history can be utilised. The global history value uses n-bits to store the n most recent branch direction outcomes, with the left-most bit being the oldest.

Branch Target Buffer (BTB)

For each entry, the BTB stores the most recent target along with a perceptron for an associated direction. The indexing of this structure uses the lower, non-zero bits of an instruction address XOR’ed with the current global branch history value.

The direction prediction is obtained from the perceptron by taking its dot-product with the global history. The prediction is not taken if this is negative, or taken otherwise. The perceptron is updated when its prediction is wrong or when the magnitude of the dot-product is below a pre-determined threshold (i.e., the confidence of the prediction is low). To update, each ith weight of the perceptron is incremented if the actual outcome of the branch is the same as the ith bit of globalHistory_, and decremented otherwise.

If the supplied branch type is Unconditional, then the predicted direction is overridden to be taken. If the supplied branch type is Conditional and the predicted direction is not taken, then the predicted target is overridden to be the next sequential instruction.

Return Address Stack (RAS)

Identified through the supplied branch type, Return instructions pop values off of the RAS to get their branch target whilst Branch-and-Link instructions push values onto the RAS, for later use by the Branch-and-Link instruction’s corresponding Return instruction.