1 files changed, 32 insertions, 0 deletions
diff --git a/src/abstract_hardware_model.h b/src/abstract_hardware_model.h
index 713c51a..369459b 100644
--- a/src/abstract_hardware_model.h
+++ b/src/abstract_hardware_model.h
@@ -1315,6 +1315,17 @@ class warp_inst_t : public inst_t {
       }
     }
   }
+  // MEMCO v3 Model B (unified shared bank-conflict computation).
+  // Walks the primary's per-lane addresses (if primary is shared) plus each
+  // valid simd_set whose source is shared/sstarr (using captured
+  // per_thread_addrs + set_active_mask). Builds a single bank→word→count
+  // map across the union, computes the max bank-access count, and stores
+  // it via set_unified_shared_cycles. No subwarp subdivision (assumes a
+  // unified bank arbiter sees all participating lanes concurrently).
+  // No-op if env var MEMCOV3_SHARED_UNIFIED is unset, or composite has no
+  // shared work, or m_per_scalar_thread_valid is false. Records pre/post
+  // delta in two globals (g_memcov3_shared_unified_*) for diagnostics.
+  void compute_unified_shared_cycles();
   void merge_simd_sets(const std::vector<simd_set_info> &other_sets);
   void set_source_inst_on_sets(const inst_t *src);
   void set_split_id_on_sets(unsigned split_id);
@@ -1470,6 +1481,27 @@ class warp_inst_t : public inst_t {
 
   bool has_dispatch_delay() { return cycles > 0; }
 
+  // MEMCO v3 shared unified bank-conflict (Model B). When active, replaces
+  // per-set max with a single composite-level cycle counter computed across
+  // all participating lanes' (bank, word) pairs — matching real-HW bank
+  // arbiter that sees all co-issued lanes concurrently. Activated by
+  // compute_unified_shared_cycles() at ldst_unit::issue when the env var
+  // MEMCOV3_SHARED_UNIFIED=1 is set and the composite has shared work.
+  bool m_unified_shared_active = false;
+  unsigned m_unified_shared_cycles = 0;
+  bool dispatch_delay_unified() {
+    if (m_unified_shared_cycles > 0) m_unified_shared_cycles--;
+    return m_unified_shared_cycles > 0;
+  }
+  bool has_dispatch_delay_unified() const {
+    return m_unified_shared_cycles > 0;
+  }
+  bool unified_shared_active() const { return m_unified_shared_active; }
+  void set_unified_shared_cycles(unsigned c) {
+    m_unified_shared_cycles = c;
+    m_unified_shared_active = true;
+  }
+
   void print(FILE *fout) const;
   unsigned get_uid() const { return m_uid; }
   unsigned long long get_streamID() const { return m_streamID; }