align RTL: tage window method, block resolveQ if update failed due to bank conflict.

src/cpu/o3/fetch.cc

@@ -1535,8 +1535,13 @@ Fetch::handleIEWSignals()
         for (const auto resolvedInstPC : entry.resolvedInstPC) {
             dbpbtb->markCFIResolved(stream_id, resolvedInstPC);
         }
-        dbpbtb->resolveUpdate(stream_id);
-        resolveQueue.pop_front();
+        bool success = dbpbtb->resolveUpdate(stream_id);
+        if (success) {
+            dbpbtb->notifyResolveSuccess();
+            resolveQueue.pop_front();
+        } else {
+            dbpbtb->notifyResolveFailure();
+        }
     }
 }
 

src/cpu/pred/BranchPredictor.py

@@ -1171,3 +1171,4 @@ class DecoupledBPUWithBTB(BranchPredictor):
     enableLoopBuffer = Param.Bool(False, "Enable loop buffer to supply inst for loops")
     enableLoopPredictor = Param.Bool(False, "Use loop predictor to predict loop exit")
     enableJumpAheadPredictor = Param.Bool(False, "Use jump ahead predictor to skip no-need-to-predict blocks")
+    resolveBlockThreshold = Param.Unsigned(8, "Consecutive resolve dequeue failures before blocking prediction once")

src/cpu/pred/btb/btb_tage.cc

@@ -137,7 +137,6 @@ tageStats(this, p.numPredictors, p.numBanks)
 
     // initialize use_alt_on_na table
     useAlt.resize(useAltOnNaSize, 0);
-
 #ifndef UNIT_TEST
     hasDB = true;
     switch (getDelay()) {
@@ -640,56 +639,58 @@ BTBTAGE::handleNewEntryAllocation(const Addr &startPC,
 }
 
 /**
- * @brief Updates the TAGE predictor state based on actual branch execution results
- * 
- * @param stream The fetch stream containing branch execution information
+ * @brief Probe resolved update for bank conflicts without mutating state.
+ * Returns false if the update cannot proceed due to a bank conflict.
  */
-void
-BTBTAGE::update(const FetchStream &stream) {
+bool
+BTBTAGE::canResolveUpdate(const FetchStream &stream) {
     Addr startAddr = stream.getRealStartPC();
     unsigned updateBank = getBankId(startAddr);
 
-    DPRINTF(TAGE, "update startAddr: %#lx, bank: %u\n", startAddr, updateBank);
-
 #ifndef UNIT_TEST
-    // Record update access per bank
+    // Record attempted update access per bank (even if it conflicts)
     tageStats.updateAccessPerBank[updateBank]++;
 #endif
 
-    // ========== Bank Conflict Detection ==========
-    // Check if this update conflicts with the last prediction
-    // Only perform conflict detection if enableBankConflict is true
-    if (enableBankConflict) {
-        // Conflict happens when:
-        //   1. There was a valid prediction in this or previous tick (predBankValid)
-        //   2. Update and prediction target the same bank (updateBank == lastPredBankId)
-        //
-        // Note: This assumes one update() call per tick (guaranteed by fetch stage)
-        if (predBankValid && updateBank == lastPredBankId) {
-            tageStats.updateBankConflict++;
-            tageStats.updateDroppedDueToConflict++;
-
+    if (enableBankConflict && predBankValid && updateBank == lastPredBankId) {
+        tageStats.updateBankConflict++;
+        tageStats.updateDeferredDueToConflict++;
 #ifndef UNIT_TEST
-            // Record conflict for this specific bank
-            tageStats.updateBankConflictPerBank[updateBank]++;
+        tageStats.updateBankConflictPerBank[updateBank]++;
 #endif
+        DPRINTF(TAGE, "Bank conflict detected: update bank %u conflicts with prediction bank %u, "
+                      "deferring this update (will retry after blocking prediction)\n",
+                      updateBank, lastPredBankId);
+        predBankValid = false;
+        return false;
+    }
 
-            DPRINTF(TAGE, "Bank conflict detected: update bank %u conflicts with "
-                          "prediction bank %u, dropping this update\n",
-                          updateBank, lastPredBankId);
-
-            // Clear prediction state after consuming it
-            predBankValid = false;
-            return;  // Drop this update entirely
-        }
+    return true;
+}
 
-        // If no conflict, clear prediction state (prediction has been consumed)
-        if (predBankValid) {
-            DPRINTF(TAGE, "No bank conflict: update bank %u != prediction bank %u\n",
-                          updateBank, lastPredBankId);
-        }
+/**
+ * @brief Perform resolved update after probe success.
+ */
+void
+BTBTAGE::doResolveUpdate(const FetchStream &stream) {
+    if (enableBankConflict && predBankValid) {
+        // Prediction consumed; clear bank tag for next cycle
         predBankValid = false;
     }
+    update(stream);
+}
+
+/**
+ * @brief Updates the TAGE predictor state based on actual branch execution results
+ * 
+ * @param stream The fetch stream containing branch execution information
+ */
+void
+BTBTAGE::update(const FetchStream &stream) {
+    Addr startAddr = stream.getRealStartPC();
+    unsigned updateBank = getBankId(startAddr);
+
+    DPRINTF(TAGE, "update startAddr: %#lx, bank: %u\n", startAddr, updateBank);
 
     // ========== Normal Update Logic ==========
     // Prepare BTB entries to update
@@ -1023,7 +1024,7 @@ BTBTAGE::TageStats::TageStats(statistics::Group* parent, int numPredictors, int
     ADD_STAT(updateMispred, statistics::units::Count::get(), "mispred when update"),
     ADD_STAT(updateResetU, statistics::units::Count::get(), "reset u when update"),
     ADD_STAT(updateBankConflict, statistics::units::Count::get(), "number of bank conflicts detected"),
-    ADD_STAT(updateDroppedDueToConflict, statistics::units::Count::get(), "number of updates dropped due to bank conflict"),
+    ADD_STAT(updateDeferredDueToConflict, statistics::units::Count::get(), "number of updates deferred due to bank conflict (retried later)"),
     ADD_STAT(updateBankConflictPerBank, statistics::units::Count::get(), "bank conflicts per bank"),
     ADD_STAT(updateAccessPerBank, statistics::units::Count::get(), "update accesses per bank"),
     ADD_STAT(predAccessPerBank, statistics::units::Count::get(), "prediction accesses per bank"),

src/cpu/pred/btb/btb_tage.hh

@@ -150,6 +150,8 @@ class BTBTAGE : public TimedBaseBTBPredictor
 
     // Update predictor state based on actual branch outcomes
     void update(const FetchStream &entry) override;
+    bool canResolveUpdate(const FetchStream &entry) override;
+    void doResolveUpdate(const FetchStream &entry) override;
 
 #ifndef UNIT_TEST
     void commitBranch(const FetchStream &stream, const DynInstPtr &inst) override;
@@ -349,7 +351,7 @@ class BTBTAGE : public TimedBaseBTBPredictor
 
         // Bank conflict statistics
         Scalar updateBankConflict;           // Number of bank conflicts detected
-        Scalar updateDroppedDueToConflict;   // Number of updates dropped due to bank conflict
+        Scalar updateDeferredDueToConflict;  // Number of updates deferred due to bank conflict (retried later)
 
 #ifndef UNIT_TEST
         // Fine-grained per-bank statistics

src/cpu/pred/btb/decoupled_bpred.cc

@@ -44,6 +44,7 @@ DecoupledBPUWithBTB::DecoupledBPUWithBTB(const DecoupledBPUWithBTBParams &p)
       bpDBSwitches(p.bpDBSwitches),
       numStages(p.numStages),
       historyManager(16), // TODO: fix this
+      resolveBlockThreshold(p.resolveBlockThreshold),
       dbpBtbStats(this, p.numStages, p.fsq_size, maxInstsNum)
 {
     if (bpDBSwitches.size() > 0) {
@@ -139,8 +140,14 @@ DecoupledBPUWithBTB::tick()
 
     // 1. Request new prediction if FSQ not full and we are idle
     if (bpuState == BpuState::IDLE && !streamQueueFull()) {
-        requestNewPrediction();
-        bpuState = BpuState::PREDICTOR_DONE;
+        if (blockPredictionPending) {
+            DPRINTF(Override, "Prediction blocked to prioritize resolve update\n");
+            dbpBtbStats.predictionBlockedForUpdate++;
+            blockPredictionPending = false;
+        } else {
+            requestNewPrediction();
+            bpuState = BpuState::PREDICTOR_DONE;
+        }
     }
 
     // 2. Handle pending prediction if available
@@ -620,26 +627,61 @@ DecoupledBPUWithBTB::update(unsigned stream_id, ThreadID tid)
     historyManager.commit(stream_id);
 }
 
-void
+bool
 DecoupledBPUWithBTB::resolveUpdate(unsigned &stream_id)
 {
     auto stream_it = fetchStreamQueue.find(stream_id);
     if (stream_it == fetchStreamQueue.end()) {
         DPRINTF(DecoupleBP, "Stream id %u not found in fetchStreamQueue, cannot update predictors\n", stream_id);
-        return;
+        return true;
     }
 
     auto &stream = stream_it->second;
 
     // Update predictor components only if the stream is hit or taken
-    if (stream.isHit || stream.exeTaken) {
-        // Update predictor components
-        for (int i = 0; i < numComponents; ++i) {
-            if (components[i]->getResolvedUpdate()) {
-                components[i]->update(stream);
+    if (!(stream.isHit || stream.exeTaken)) {
+        return true;
+    }
+
+    // Phase 1: probe all resolved-update components to ensure no blocker
+    for (int i = 0; i < numComponents; ++i) {
+        if (components[i]->getResolvedUpdate()) {
+            if (!components[i]->canResolveUpdate(stream)) {
+                return false;
             }
         }
     }
+
+    // Phase 2: all clear, perform updates once
+    for (int i = 0; i < numComponents; ++i) {
+        if (components[i]->getResolvedUpdate()) {
+            components[i]->doResolveUpdate(stream);
+        }
+    }
+
+    return true;
+}
+
+void
+DecoupledBPUWithBTB::notifyResolveSuccess()
+{
+    resolveDequeueFailCounter = 0;
+}
+
+void
+DecoupledBPUWithBTB::notifyResolveFailure()
+{
+    resolveDequeueFailCounter++;
+    if (resolveDequeueFailCounter >= resolveBlockThreshold) {
+        blockPredictionOnce();
+        resolveDequeueFailCounter = 0;
+    }
+}
+
+void
+DecoupledBPUWithBTB::blockPredictionOnce()
+{
+    blockPredictionPending = true;
 }
 
 void

src/cpu/pred/btb/decoupled_bpred.hh

@@ -165,6 +165,9 @@ class DecoupledBPUWithBTB : public BPredUnit
     bool squashing{false};
 
     HistoryManager historyManager;
+    bool blockPredictionPending{false};
+    unsigned resolveDequeueFailCounter{0};
+    const unsigned resolveBlockThreshold;
 
     unsigned numOverrideBubbles{0};
 
@@ -362,6 +365,9 @@ class DecoupledBPUWithBTB : public BPredUnit
         statistics::Scalar predFalseHit;
         statistics::Scalar commitFalseHit;
 
+        // Window blocking statistics
+        statistics::Scalar predictionBlockedForUpdate;  // Times prediction was blocked for update priority
+
         DBPBTBStats(statistics::Group* parent, unsigned numStages, unsigned fsqSize, unsigned maxInstsNum);
     } dbpBtbStats;
 
@@ -815,11 +821,14 @@ class DecoupledBPUWithBTB : public BPredUnit
     void resetPC(Addr new_pc);
 
     // Helper functions for update
-    void resolveUpdate(unsigned &stream_id);
+    bool resolveUpdate(unsigned &stream_id);
     void prepareResolveUpdateEntries(unsigned &stream_id);
     void markCFIResolved(unsigned &stream_id, uint64_t resolvedInstPC);
     void updatePredictorComponents(FetchStream &stream);
     void updateStatistics(const FetchStream &stream);
+    void notifyResolveSuccess();
+    void notifyResolveFailure();
+    void blockPredictionOnce();
 
     // Helper function to process FTQ entry completion
     void processFetchTargetCompletion(const FtqEntry &target_to_fetch);

src/cpu/pred/btb/decoupled_bpred_stats.cc

@@ -496,7 +496,8 @@ DecoupledBPUWithBTB::DBPBTBStats::DBPBTBStats(
     ADD_STAT(btbEntriesWithDifferentStart, statistics::units::Count::get(), "number of btb entries with different start PC"),
     ADD_STAT(btbEntriesWithOnlyOneJump, statistics::units::Count::get(), "number of btb entries with different start PC starting with a jump"),
     ADD_STAT(predFalseHit, statistics::units::Count::get(), "false hit detected at pred"),
-    ADD_STAT(commitFalseHit, statistics::units::Count::get(), "false hit detected at commit")
+    ADD_STAT(commitFalseHit, statistics::units::Count::get(), "false hit detected at commit"),
+    ADD_STAT(predictionBlockedForUpdate, statistics::units::Count::get(), "prediction blocked for update priority")
 {
     predsOfEachStage.init(numStages);
     commitPredsFromEachStage.init(numStages+1);

src/cpu/pred/btb/test/btb_tage.test.cc

@@ -901,10 +901,11 @@ TEST_F(BTBTAGETest, BankConflict) {
         setupTageEntry(bankTage, 0xa0, 0, 1, false);
 
         uint64_t conflicts_before = bankTage->tageStats.updateBankConflict;
-        bankTage->update(stream);
+        bool can_update = bankTage->canResolveUpdate(stream);
 
-        // Should detect conflict and drop update
+        // Should detect conflict and defer update
         EXPECT_EQ(bankTage->tageStats.updateBankConflict, conflicts_before + 1);
+        EXPECT_FALSE(can_update);
         EXPECT_FALSE(bankTage->predBankValid);
     }
 
@@ -918,7 +919,9 @@ TEST_F(BTBTAGETest, BankConflict) {
         FetchStream stream = createStream(0x104, createBTBEntry(0x104), true, meta);
 
         uint64_t conflicts_before = bankTage->tageStats.updateBankConflict;
-        bankTage->update(stream);
+        bool can_update = bankTage->canResolveUpdate(stream);
+        ASSERT_TRUE(can_update);
+        bankTage->doResolveUpdate(stream);
 
         // Should not detect conflict
         EXPECT_EQ(bankTage->tageStats.updateBankConflict, conflicts_before);
@@ -936,7 +939,9 @@ TEST_F(BTBTAGETest, BankConflict) {
         setupTageEntry(bankTage, 0xa0, 0, 1, false);
 
         uint64_t conflicts_before = bankTage->tageStats.updateBankConflict;
-        bankTage->update(stream);
+        bool can_update = bankTage->canResolveUpdate(stream);
+        ASSERT_TRUE(can_update);
+        bankTage->doResolveUpdate(stream);
 
         // No conflict even with same bank
         EXPECT_EQ(bankTage->tageStats.updateBankConflict, conflicts_before);

src/cpu/pred/btb/timed_base_pred.hh

@@ -75,6 +75,9 @@ class TimedBaseBTBPredictor: public SimObject
     virtual void update(const FetchStream &entry) {}
     virtual unsigned getDelay() {return numDelay;}
     virtual bool getResolvedUpdate() {return resolvedUpdate;}
+    // Two-phase resolved update: probe first, then apply
+    virtual bool canResolveUpdate(const FetchStream &entry) { return true; }
+    virtual void doResolveUpdate(const FetchStream &entry) { update(entry); }
 #ifndef UNIT_TEST
     // do some statistics on a per-branch and per-predictor basis
     virtual void commitBranch(const FetchStream &entry, const DynInstPtr &inst) {}