cpu-o3: branchPred: parameterize predictWidth/fetchBlockSize (#227)

Change-Id: I94561ec64feebdcebb6d8837ff9f6f64d51b09bd

src/cpu/pred/BranchPredictor.py

@@ -861,6 +861,7 @@ class TimedBaseFTBPredictor(SimObject):
     cxx_header = "cpu/pred/ftb/timed_base_pred.hh"
 
     numBr = Param.Unsigned(Parent.numBr, "Number of maximum branches per entry")
+    predictWidth = Param.Unsigned(Parent.predictWidth, "The width of prediction/fetch block")
     # subclass are encouraged to explicitly declare latency as numDelay
     numDelay = Param.Unsigned(1000, "Number of bubbles to put on a prediction")
 
@@ -943,6 +944,7 @@ class DecoupledBPUWithFTB(BranchPredictor):
     fsq_size = Param.Unsigned(64, "Fetch stream queue size")
     maxHistLen = Param.Unsigned(970, "The length of history")
     numBr = Param.Unsigned(2, "Number of maximum branches per entry")
+    predictWidth = Param.Unsigned(32, "The width of prediction/fetch block")
     numStages = Param.Unsigned(3, "Number of stages in the pipeline")
     ftb = Param.DefaultFTB(DefaultFTB(), "FTB")
     tage = Param.FTBTAGE(FTBTAGE(), "TAGE predictor")

src/cpu/pred/SConscript

@@ -84,6 +84,7 @@ Source('stream/stream_common.cc')
 Source('ftb/decoupled_bpred.cc')
 Source('ftb/ftb.cc')
 Source('ftb/stream_common.cc')
+Source('ftb/stream_struct.cc')
 Source('ftb/timed_base_pred.cc')
 Source('ftb/fetch_target_queue.cc')
 Source('ftb/ftb_tage.cc')

src/cpu/pred/ftb/decoupled_bpred.cc

@@ -30,6 +30,7 @@ DecoupledBPUWithFTB::DecoupledBPUWithFTB(const DecoupledBPUWithFTBParams &p)
       fetchTargetQueue(p.ftq_size),
       fetchStreamQueueSize(p.fsq_size),
       numBr(p.numBr),
+      predictWidth(p.predictWidth),
       historyBits(p.maxHistLen),
       uftb(p.uftb),
       ftb(p.ftb),
@@ -43,6 +44,7 @@ DecoupledBPUWithFTB::DecoupledBPUWithFTB(const DecoupledBPUWithFTBParams &p)
       historyManager(p.numBr),
       dbpFtbStats(this, p.numStages, p.fsq_size)
 {
+    gem5::branch_prediction::ftb_pred::predictWidth = p.predictWidth;
     if (bpDBSwitches.size() > 0) {
 
         bpdb.init_db();
@@ -2098,7 +2100,7 @@ DecoupledBPUWithFTB::tryEnqFetchTarget()
          ftq_enq_state.pc, end);
     }
 
-    assert(ftq_enq_state.pc <= end || (end < 0x20 && (ftq_enq_state.pc + 0x20 < 0x20)));
+    assert(ftq_enq_state.pc <= end || (end < predictWidth && (ftq_enq_state.pc + predictWidth < predictWidth)));
 
     // create a new target entry
     FtqEntry ftq_entry;
@@ -2114,7 +2116,7 @@ DecoupledBPUWithFTB::tryEnqFetchTarget()
         bool jaHit = stream_to_enq.jaHit;
         if (jaHit) {
             int &currentSentBlock = stream_to_enq.currentSentBlock;
-            thisFtqEntryShouldEndPC = stream_to_enq.startPC + (currentSentBlock + 1) * 0x20;
+            thisFtqEntryShouldEndPC = stream_to_enq.startPC + (currentSentBlock + 1) * predictWidth;
             currentSentBlock++;
         }
     }
@@ -2289,9 +2291,9 @@ DecoupledBPUWithFTB::makeNewPrediction(bool create_new_stream)
             entry.isHit = false;
             entry.falseHit = true;
             entry.predTaken = false;
-            entry.predEndPC = entry.startPC + 32;
+            entry.predEndPC = entry.startPC + predictWidth;
             entry.predFTBEntry = FTBEntry();
-            s0PC = entry.startPC + 32; // TODO: parameterize
+            s0PC = entry.startPC + predictWidth;
             // TODO: when false hit, act like a miss, do not update history
         }
 

src/cpu/pred/ftb/decoupled_bpred.hh

@@ -209,6 +209,8 @@ class DecoupledBPUWithFTB : public BPredUnit
 
     unsigned numBr;
 
+    unsigned predictWidth;
+
     unsigned cacheLineOffsetBits{6};  // TODO: parameterize this
     unsigned cacheLineSize{64};
 

src/cpu/pred/ftb/ftb.cc

@@ -48,6 +48,7 @@ DefaultFTB::DefaultFTB(const Params &p)
     tagBits(p.tagBits),
     instShiftAmt(p.instShiftAmt),
     log2NumThreads(floorLog2(p.numThreads)),
+    predictWidth(p.predictWidth),
     numBr(p.numBr),
     numWays(p.numWays),
     numSets(numEntries / numWays),
@@ -264,7 +265,7 @@ DefaultFTB::getAndSetNewFTBEntry(FetchStream &stream)
                 new_entry.fallThruAddr = branch_info.getEnd();
                 incNonL0Stat(ftbStats.newEntryWithUncond);
             } else {
-                new_entry.fallThruAddr = startPC + 32;
+                new_entry.fallThruAddr = startPC + predictWidth;
                 incNonL0Stat(ftbStats.newEntryWithCond);
             }
             entry_to_write = new_entry;

src/cpu/pred/ftb/ftb.hh

@@ -247,6 +247,8 @@ class DefaultFTB : public TimedBaseFTBPredictor
     /** Log2 NumThreads used for hashing threadid */
     unsigned log2NumThreads;
 
+    unsigned predictWidth;
+
     unsigned numBr;
 
     unsigned numWays;

src/cpu/pred/ftb/stream_struct.cc

@@ -0,0 +1,17 @@
+#include "cpu/pred/ftb/stream_struct.hh"
+
+namespace gem5 {
+
+namespace branch_prediction {
+
+namespace ftb_pred {
+
+unsigned predictWidth{0};
+// global variable, set in decoupled_bpred.cc
+// used in stream_struct.hh
+
+}
+
+}
+
+}

src/cpu/pred/ftb/stream_struct.hh

@@ -20,6 +20,8 @@ namespace branch_prediction {
 
 namespace ftb_pred {
 
+extern unsigned predictWidth;
+
 enum EndType {
     END_CALL=0,
     END_RET,
@@ -189,7 +191,7 @@ typedef struct FTBEntry
     // every 
     bool isReasonable(Addr start) {
         Addr min = start;
-        Addr max = start+34;
+        Addr max = start + predictWidth + 2;
         bool reasonable = true;
         for (auto &slot : slots) {
             if (slot.pc < min || slot.pc > max) {
@@ -375,8 +377,8 @@ struct FetchStream
         if (jaHit && squashType == SQUASH_CTRL) {
             Addr realStart = startPC;
             Addr squashBranchPC = exeBranchInfo.pc;
-            while (realStart + 0x20 <= squashBranchPC) {
-                realStart += 0x20;
+            while (realStart + predictWidth <= squashBranchPC) {
+                realStart += predictWidth;
             }
             return realStart;
         } else {
@@ -473,7 +475,7 @@ typedef struct FullFTBPrediction
             }
 
         } else {
-            target = bbStart + 32; //TODO: +predictWidth
+            target = bbStart + predictWidth;
         }
         return target;
     }
@@ -493,7 +495,7 @@ typedef struct FullFTBPrediction
         if (valid) {
             return ftbEntry.fallThruAddr;
         } else {
-            return bbStart + 32; //TODO: +predictWidth
+            return bbStart + predictWidth;
         }
     }