replay: send redirect when replay

src/main/scala/xiangshan/backend/CtrlBlock.scala

@@ -2,6 +2,7 @@ package xiangshan.backend
 
 import chisel3._
 import chisel3.util._
+import utils._
 import xiangshan._
 import xiangshan.backend.decode.{DecodeBuffer, DecodeStage}
 import xiangshan.backend.rename.{Rename, BusyTable}
@@ -37,7 +38,7 @@ class CtrlToLsBlockIO extends XSBundle {
   val redirect = ValidIO(new Redirect)
 }
 
-class CtrlBlock extends XSModule {
+class CtrlBlock extends XSModule with HasCircularQueuePtrHelper {
   val io = IO(new Bundle {
     val frontend = Flipped(new FrontendToBackendIO)
     val fromIntBlock = Flipped(new IntBlockToCtrlIO)
@@ -70,18 +71,14 @@ class CtrlBlock extends XSModule {
 
   val roq = Module(new Roq(roqWbSize))
 
-  val redirect = Mux(
-    roq.io.redirect.valid,
-    roq.io.redirect,
-    Mux(
-      brq.io.redirect.valid,
-      brq.io.redirect,
-      io.fromLsBlock.replay
-    )
-  )
+  val lsqIsAfterBrq = isAfter(io.fromLsBlock.replay.bits.roqIdx, brq.io.redirect.bits.roqIdx)
+  val redirectArb = Mux(brq.io.redirect.valid && (!io.fromLsBlock.replay.valid || lsqIsAfterBrq),
+    brq.io.redirect.bits, io.fromLsBlock.replay.bits)
+  val redirectValid = roq.io.redirect.valid || brq.io.redirect.valid || io.fromLsBlock.replay.valid
+  val redirect = Mux(roq.io.redirect.valid, roq.io.redirect.bits, redirectArb)
 
-  io.frontend.redirect.valid := roq.io.redirect.valid || brq.io.redirect.valid
-  io.frontend.redirect.bits := Mux(brq.io.redirect.valid, brq.io.redirect.bits.target, roq.io.redirect.bits.target)
+  io.frontend.redirect.valid := redirectValid
+  io.frontend.redirect.bits := Mux(roq.io.redirect.valid, roq.io.redirect.bits.target, redirectArb.target)
   io.frontend.outOfOrderBrInfo <> brq.io.outOfOrderBrInfo
   io.frontend.inOrderBrInfo <> brq.io.inOrderBrInfo
 
@@ -97,15 +94,18 @@ class CtrlBlock extends XSModule {
   brq.io.exuRedirect <> io.fromIntBlock.exuRedirect
 
   decBuf.io.isWalking := roq.io.commits(0).valid && roq.io.commits(0).bits.isWalk
-  decBuf.io.redirect <> redirect
+  decBuf.io.redirect.valid <> redirectValid
+  decBuf.io.redirect.bits <> redirect
   decBuf.io.out <> rename.io.in
 
-  rename.io.redirect <> redirect
+  rename.io.redirect.valid <> redirectValid
+  rename.io.redirect.bits <> redirect
   rename.io.roqCommits <> roq.io.commits
   rename.io.out <> dispatch.io.fromRename
   rename.io.renameBypass <> dispatch.io.renameBypass
 
-  dispatch.io.redirect <> redirect
+  dispatch.io.redirect.valid <> redirectValid
+  dispatch.io.redirect.bits <> redirect
   dispatch.io.enqRoq <> roq.io.enq
   dispatch.io.enqLsq <> io.toLsBlock.enqLsq
   dispatch.io.dequeueRoqIndex.valid := roq.io.commitRoqIndex.valid || io.oldestStore.valid
@@ -126,7 +126,7 @@ class CtrlBlock extends XSModule {
   dispatch.io.enqIQData <> io.toIntBlock.enqIqData ++ io.toFpBlock.enqIqData ++ io.toLsBlock.enqIqData
 
 
-  val flush = redirect.valid && (redirect.bits.isException || redirect.bits.isFlushPipe)
+  val flush = redirectValid && (redirect.isException || redirect.isFlushPipe)
   fpBusyTable.io.flush := flush
   intBusyTable.io.flush := flush
   for((wb, setPhyRegRdy) <- io.fromIntBlock.wbRegs.zip(intBusyTable.io.wbPregs)){
@@ -148,8 +148,10 @@ class CtrlBlock extends XSModule {
     fpBusyTable.io.replayPregs(i).bits := dispatch.io.replayPregReq(i).preg
   }
 
-  roq.io.memRedirect <> io.fromLsBlock.replay
-  roq.io.brqRedirect <> brq.io.redirect
+  roq.io.memRedirect := DontCare
+  roq.io.memRedirect.valid := false.B
+  roq.io.brqRedirect.valid := brq.io.redirect.valid || io.fromLsBlock.replay.valid
+  roq.io.brqRedirect.bits <> redirectArb
   roq.io.exeWbResults.take(roqWbSize-1).zip(
     io.fromIntBlock.wbRegs ++ io.fromFpBlock.wbRegs ++ io.fromLsBlock.stOut
   ).foreach{
@@ -159,9 +161,12 @@ class CtrlBlock extends XSModule {
   }
   roq.io.exeWbResults.last := brq.io.out
 
-  io.toIntBlock.redirect := redirect
-  io.toFpBlock.redirect := redirect
-  io.toLsBlock.redirect := redirect
+  io.toIntBlock.redirect.valid := redirectValid
+  io.toIntBlock.redirect.bits := redirect
+  io.toFpBlock.redirect.valid := redirectValid
+  io.toFpBlock.redirect.bits := redirect
+  io.toLsBlock.redirect.valid := redirectValid
+  io.toLsBlock.redirect.bits := redirect
 
   // roq to int block
   io.roqio.toCSR <> roq.io.csr

src/main/scala/xiangshan/backend/brq/Brq.scala

@@ -224,16 +224,21 @@ class Brq extends XSModule with HasCircularQueuePtrHelper {
     // misprediction or replay
     stateQueue.zipWithIndex.foreach({case(s, i) =>
       val ptr = BrqPtr(brQueue(i).ptrFlag, i.U)
-      when(s.isWb && brQueue(i).exuOut.uop.roqIdx.needFlush(io.memRedirect)){
+      when(io.memRedirect.valid && ptr.needBrFlush(io.memRedirect.bits.brTag)){
         s := s_idle
       }
       when(io.redirect.valid && ptr.needBrFlush(io.redirect.bits.brTag)){
         s := s_invalid
       }
     })
-    when(io.redirect.valid){ // Only Br Mispred reset tailPtr, replay does not
-      tailPtr := io.redirect.bits.brTag + true.B
+    when(io.redirect.valid || io.memRedirect.valid){
+      when (!io.memRedirect.valid || isAfter(io.memRedirect.bits.brTag, io.redirect.bits.brTag)) {
+        tailPtr := io.redirect.bits.brTag + 1.U
+      }.otherwise {
+        tailPtr := io.memRedirect.bits.brTag + 1.U
+      }
     }
+
   }
 
 

src/main/scala/xiangshan/backend/decode/DecodeBuffer.scala

@@ -24,7 +24,7 @@ class DecodeBuffer extends XSModule {
     })
   )
 
-  val flush = io.redirect.valid && !io.redirect.bits.isReplay
+  val flush = io.redirect.valid// && !io.redirect.bits.isReplay
 
   for( i <- 0 until RenameWidth){
     when(io.out(i).fire()){

src/main/scala/xiangshan/backend/dispatch/Dispatch.scala

@@ -67,7 +67,7 @@ class Dispatch extends XSModule {
 
   // pipeline between rename and dispatch
   // accepts all at once
-  val redirectValid = io.redirect.valid && !io.redirect.bits.isReplay
+  val redirectValid = io.redirect.valid// && !io.redirect.bits.isReplay
   for (i <- 0 until RenameWidth) {
     PipelineConnect(io.fromRename(i), dispatch1.io.fromRename(i), dispatch1.io.recv(i), redirectValid)
   }

src/main/scala/xiangshan/backend/dispatch/Dispatch1.scala

@@ -122,7 +122,7 @@ class Dispatch1 extends XSModule {
     *   acquire ROQ (all), LSQ (load/store only) and dispatch queue slots
     *   only set valid when all of them provides enough entries
     */
-  val redirectValid = io.redirect.valid && !io.redirect.bits.isReplay
+  val redirectValid = io.redirect.valid// && !io.redirect.bits.isReplay
   val allResourceReady = io.enqLsq.canAccept && io.enqRoq.canAccept && io.toIntDqReady && io.toFpDqReady && io.toLsDqReady
 
   // Instructions should enter dispatch queues in order.

src/main/scala/xiangshan/backend/dispatch/DispatchQueue.scala

@@ -62,7 +62,7 @@ class DispatchQueue(size: Int, enqnum: Int, deqnum: Int, replayWidth: Int) exten
 
   val allWalkDone = !io.inReplayWalk && io.otherWalkDone
   val canEnqueue = validEntries <= (size - enqnum).U && allWalkDone
-  val canActualEnqueue = canEnqueue && !(io.redirect.valid && !io.redirect.bits.isReplay)
+  val canActualEnqueue = canEnqueue && !(io.redirect.valid /*&& !io.redirect.bits.isReplay*/)
 
   /**
     * Part 1: update states and uops when enqueue, dequeue, commit, redirect/replay
@@ -109,7 +109,7 @@ class DispatchQueue(size: Int, enqnum: Int, deqnum: Int, replayWidth: Int) exten
   }
 
   // redirect: cancel uops currently in the queue
-  val mispredictionValid = io.redirect.valid && io.redirect.bits.isMisPred
+  val mispredictionValid = io.redirect.valid //&& io.redirect.bits.isMisPred
   val exceptionValid = io.redirect.valid && io.redirect.bits.isException
   val flushPipeValid = io.redirect.valid && io.redirect.bits.isFlushPipe
   val roqNeedFlush = Wire(Vec(size, Bool()))
@@ -128,7 +128,7 @@ class DispatchQueue(size: Int, enqnum: Int, deqnum: Int, replayWidth: Int) exten
   }
 
   // replay: from s_dispatched to s_valid
-  val replayValid = io.redirect.valid && io.redirect.bits.isReplay
+  val replayValid = false.B//io.redirect.valid && io.redirect.bits.isReplay
   val needReplay = Wire(Vec(size, Bool()))
   for (i <- 0 until size) {
     needReplay(i) := roqNeedFlush(i) && stateEntries(i) === s_dispatched && replayValid

src/main/scala/xiangshan/backend/roq/Roq.scala

@@ -425,14 +425,14 @@ class Roq(numWbPorts: Int) extends XSModule with HasCircularQueuePtrHelper {
   }
   // rollback: write all 
   // when rollback, reset writebacked entry to valid
-  when(io.memRedirect.valid) { // TODO: opt timing
-    for (i <- 0 until RoqSize) {
-      val recRoqIdx = RoqPtr(flagBkup(i), i.U)
-      when (valid(i) && isAfter(recRoqIdx, io.memRedirect.bits.roqIdx)) {
-        writebacked(i) := false.B
-      }
-    }
-  }
+  // when(io.memRedirect.valid) { // TODO: opt timing
+  //   for (i <- 0 until RoqSize) {
+  //     val recRoqIdx = RoqPtr(flagBkup(i), i.U)
+  //     when (valid(i) && isAfter(recRoqIdx, io.memRedirect.bits.roqIdx)) {
+  //       writebacked(i) := false.B
+  //     }
+  //   }
+  // }
 
   // read
   // deqPtrWritebacked

src/main/scala/xiangshan/mem/lsqueue/LoadQueue.scala

@@ -433,6 +433,8 @@ class LoadQueue extends XSModule with HasDCacheParameters with HasCircularQueueP
       rollback(i).bits.isMisPred := false.B
       rollback(i).bits.isException := false.B
       rollback(i).bits.isFlushPipe := false.B
+      rollback(i).bits.target := rollbackUop.cf.pc
+      rollback(i).bits.brTag := rollbackUop.brTag
 
       XSDebug(
         l1Violation,
@@ -529,15 +531,15 @@ class LoadQueue extends XSModule with HasDCacheParameters with HasCircularQueueP
   for (i <- 0 until LoadQueueSize) {
     needCancel(i) := uop(i).roqIdx.needFlush(io.brqRedirect) && allocated(i) && !commited(i)
     when(needCancel(i)) {
-      when(io.brqRedirect.bits.isReplay){
-        valid(i) := false.B
-        writebacked(i) := false.B
-        listening(i) := false.B
-        miss(i) := false.B
-        pending(i) := false.B
-      }.otherwise{
+      // when(io.brqRedirect.bits.isReplay){
+      //   valid(i) := false.B
+      //   writebacked(i) := false.B
+      //   listening(i) := false.B
+      //   miss(i) := false.B
+      //   pending(i) := false.B
+      // }.otherwise{
         allocated(i) := false.B
-      }
+      // }
     }
   }
   when (io.brqRedirect.valid && io.brqRedirect.bits.isMisPred) {

src/main/scala/xiangshan/mem/lsqueue/StoreQueue.scala

@@ -340,13 +340,13 @@ class StoreQueue extends XSModule with HasDCacheParameters with HasCircularQueue
   for (i <- 0 until StoreQueueSize) {
     needCancel(i) := uop(i).roqIdx.needFlush(io.brqRedirect) && allocated(i) && !commited(i)
     when(needCancel(i)) {
-      when(io.brqRedirect.bits.isReplay){
-        datavalid(i) := false.B
-        writebacked(i) := false.B
-        pending(i) := false.B
-      }.otherwise{
+      // when(io.brqRedirect.bits.isReplay){
+      //   datavalid(i) := false.B
+      //   writebacked(i) := false.B
+      //   pending(i) := false.B
+      // }.otherwise{
         allocated(i) := false.B
-      }
+      // }
     }
   }
   when (io.brqRedirect.valid && io.brqRedirect.bits.isMisPred) {