Information-Entropy-Guided Counterfactual Stability Modeling and Replay-Based Selective Override for Context-Fragile Branch Prediction

Modern out-of-order processors depend critically on TAGE family branch predictors, which often struggle with context-fragile branches whose outcomes are highly sensitive to slight perturbations in the most recent global history. Existing predictors are confined to single-path factual reasoning and lack any mechanism to explicitly probe the local stability of a prediction. We formalise this limitation from an information entropy perspective: a context-fragile branch corresponds to a high conditional entropy region in the joint space of recent history bits and branch outcomes, where a single bit flip in the youngest history can shift the posterior branch probability across the decision boundary. To address this, we propose Mirror-TAGE, a lightweight microarchitectural framework that integrates counterfactual stability modeling with replay-guided selective overrides. Building on a TAGE-SC-L baseline following Seznec, Mirror-TAGE injects controlled bit-level perturbations into the youngest global history bits to construct two mirrored views, screened by an entropy-based reliability filter. We define a local prediction entropy, derive an information-theoretic characterisation of override correctness via DRT-filtered agreement, and decompose uncertainty into aleatoric and epistemic components. Under a paired causally consistent learning framework on synthetic workloads isolating the context-fragile regime, Mirror-TAGE achieves a 0.82 pp gain (66% override correctness, 7.5% overhead). Of successful overrides, 78% occur at entropy > 0.85 bits, consistent with the entropy-guided criterion, though partly driven by the entry gate design. Six SPEC CPU 2017 traces confirm no accuracy degradation, with gains of up to 0.38 pp. These results support entropy-guided counterfactual stability modelling as a promising branch prediction paradigm.