From 37f6dbde93dddde9f701ebc4e231169a79aadde3 Mon Sep 17 00:00:00 2001
From: Jonathan Woodruff <jdw57@cam.ac.uk>
Date: Wed, 12 Jan 2022 16:08:33 +0000
Subject: [PATCH] Timing optimisation for Tournament Predictor. Start
 global-history-indexed lookups in the previous cycle. This is a general
 timing optimisation that should be pushed upstream at some point.

---
 src_Core/RISCY_OOO/procs/lib/Btb.bsv      |  5 ++++-
 src_Core/RISCY_OOO/procs/lib/TourPred.bsv | 17 +++++++++++++----
 2 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/src_Core/RISCY_OOO/procs/lib/Btb.bsv b/src_Core/RISCY_OOO/procs/lib/Btb.bsv
index a1d9059..2b92d72 100644
--- a/src_Core/RISCY_OOO/procs/lib/Btb.bsv
+++ b/src_Core/RISCY_OOO/procs/lib/Btb.bsv
@@ -131,7 +131,10 @@ module mkBtbCore(NextAddrPred#(hashSz))
         // Start SupSizeX2 BTB lookups, but ensure to lookup in the appropriate
         // bank for the alignment of each potential branch.
         for (Integer i = 0; i < valueOf(SupSizeX2); i = i + 1) begin
-            BtbAddr a = unpack(pack(getBtbAddr(pc)) + fromInteger(i));
+            // Only add lower bits for timing.
+            BtbAddr a = getBtbAddr(pc);
+            a = unpack({a.tag, {a.index,a.bank} + fromInteger(i)});
+            //BtbAddr a = unpack(pack(getBtbAddr(pc)) + fromInteger(i));
             fullRecords[a.bank].lookupStart(MapKeyIndex{key: hash(a.tag), index: a.index});
             compressedRecords[a.bank].lookupStart(MapKeyIndex{key: hash(a.tag), index: a.index});
         end
diff --git a/src_Core/RISCY_OOO/procs/lib/TourPred.bsv b/src_Core/RISCY_OOO/procs/lib/TourPred.bsv
index d751f0d..7cd680b 100644
--- a/src_Core/RISCY_OOO/procs/lib/TourPred.bsv
+++ b/src_Core/RISCY_OOO/procs/lib/TourPred.bsv
@@ -114,6 +114,8 @@ module mkTourPred(DirPredictor#(TourTrainInfo));
     endfunction
 
     TourGlobalHist curGHist = gHistReg.history; // global history: MSB is the latest branch
+    Reg#(Vector#(SupSize, Bool)) globalTakenVec <- mkRegU;
+    Reg#(Vector#(SupSize, Bool)) useLocalVec <- mkRegU;
 
     Vector#(SupSize, DirPred#(TourTrainInfo)) predIfc;
     for(Integer i = 0; i < valueof(SupSize); i = i+1) begin
@@ -127,12 +129,11 @@ module mkTourPred(DirPredictor#(TourTrainInfo));
                 // all previous branch in this cycle must be not taken
                 // otherwise this branch should be on wrong path
                 // because all inst in same cycle are fetched consecutively
-                TourGlobalHist globalHist = curGHist >> predCnt[i];
                 // get global prediction
-                Bool globalTaken = isTaken(globalBht.sub(globalHist));
+                Bool globalTaken = globalTakenVec[predCnt[i]];
 
                 // make choice
-                Bool useLocal = isTaken(choiceBht.sub(globalHist));
+                Bool useLocal = useLocalVec[predCnt[i]];
                 Bool taken = useLocal ? localTaken : globalTaken;
 
                 // record prediction
@@ -145,7 +146,7 @@ module mkTourPred(DirPredictor#(TourTrainInfo));
                 return DirPredResult {
                     taken: taken,
                     train: TourTrainInfo {
-                        globalHist: globalHist,
+                        globalHist: curGHist >> predCnt[i],
                         localHist: localHist,
                         globalTaken: globalTaken,
                         localTaken: localTaken
@@ -158,6 +159,14 @@ module mkTourPred(DirPredictor#(TourTrainInfo));
     (* fire_when_enabled, no_implicit_conditions *)
     rule canonGlobalHist;
         gHistReg.addHistory(predRes[valueof(SupSize)], predCnt[valueof(SupSize)]);
+        // Buffer useLocalVec
+        // Reproduce next history; this would ideally be done in GlobalBrHistReg to avoid duplicating logic.
+        TourGlobalHist nHist = truncate({predRes[valueof(SupSize)], curGHist} >> predCnt[valueof(SupSize)]);
+        function Bool globalTakenLookup (Integer i) = isTaken(globalBht.sub(nHist >> i));
+        function Bool useLocalLookup (Integer i) = isTaken(choiceBht.sub(nHist >> i));
+        globalTakenVec <= genWith(globalTakenLookup);
+        useLocalVec <= genWith(useLocalLookup);
+        // Reset counters and prediction.
         predRes[valueof(SupSize)] <= 0;
         predCnt[valueof(SupSize)] <= 0;
     endrule