feat(Svpbmt): support WMO memory access and outstanding of NC

src/main/scala/xiangshan/Bundle.scala

@@ -351,9 +351,12 @@ class ResetPregStateReq(implicit p: Parameters) extends XSBundle {
 
 class DebugBundle(implicit p: Parameters) extends XSBundle {
   val isMMIO = Bool()
+  val isNC = Bool()
   val isPerfCnt = Bool()
   val paddr = UInt(PAddrBits.W)
   val vaddr = UInt(VAddrBits.W)
+
+  def isSkipDiff: Bool = isMMIO || isNC || isPerfCnt
   /* add L/S inst info in EXU */
   // val L1toL2TlbLatency = UInt(XLEN.W)
   // val levelTlbHit = UInt(2.W)

src/main/scala/xiangshan/XSCore.scala

@@ -208,7 +208,7 @@ class XSCoreImp(outer: XSCoreBase) extends LazyModuleImp(outer)
   memBlock.io.ooo_to_mem.tlbCsr := backend.io.mem.tlbCsr
   memBlock.io.ooo_to_mem.lsqio.lcommit          := backend.io.mem.robLsqIO.lcommit
   memBlock.io.ooo_to_mem.lsqio.scommit          := backend.io.mem.robLsqIO.scommit
-  memBlock.io.ooo_to_mem.lsqio.pendingUncacheld := backend.io.mem.robLsqIO.pendingUncacheld
+  memBlock.io.ooo_to_mem.lsqio.pendingMMIOld    := backend.io.mem.robLsqIO.pendingMMIOld
   memBlock.io.ooo_to_mem.lsqio.pendingld        := backend.io.mem.robLsqIO.pendingld
   memBlock.io.ooo_to_mem.lsqio.pendingst        := backend.io.mem.robLsqIO.pendingst
   memBlock.io.ooo_to_mem.lsqio.pendingVst       := backend.io.mem.robLsqIO.pendingVst

src/main/scala/xiangshan/backend/MemBlock.scala

@@ -89,7 +89,7 @@ class ooo_to_mem(implicit p: Parameters) extends MemBlockBundle {
   val lsqio = new Bundle {
     val lcommit = Input(UInt(log2Up(CommitWidth + 1).W))
     val scommit = Input(UInt(log2Up(CommitWidth + 1).W))
-    val pendingUncacheld = Input(Bool())
+    val pendingMMIOld = Input(Bool())
     val pendingld = Input(Bool())
     val pendingst = Input(Bool())
     val pendingVst = Input(Bool())
@@ -823,6 +823,7 @@ class MemBlockInlinedImp(outer: MemBlockInlined) extends LazyModuleImp(outer)
     // forward
     loadUnits(i).io.lsq.forward <> lsq.io.forward(i)
     loadUnits(i).io.sbuffer <> sbuffer.io.forward(i)
+    loadUnits(i).io.ubuffer <> uncache.io.forward(i)
     loadUnits(i).io.tl_d_channel := dcache.io.lsu.forward_D(i)
     loadUnits(i).io.forward_mshr <> dcache.io.lsu.forward_mshr(i)
     // ld-ld violation check
@@ -908,6 +909,7 @@ class MemBlockInlinedImp(outer: MemBlockInlined) extends LazyModuleImp(outer)
       loadUnits(i).io.lsq.uncache.bits := DontCare
     }
     lsq.io.ld_raw_data(i) <> loadUnits(i).io.lsq.ld_raw_data
+    lsq.io.ncOut(i) <> loadUnits(i).io.lsq.nc_ldin
     lsq.io.l2_hint.valid := l2_hint.valid
     lsq.io.l2_hint.bits.sourceId := l2_hint.bits.sourceId
     lsq.io.l2_hint.bits.isKeyword := l2_hint.bits.isKeyword
@@ -965,6 +967,7 @@ class MemBlockInlinedImp(outer: MemBlockInlined) extends LazyModuleImp(outer)
     hybridUnits(i).io.ldu_io.lsq.forward <> lsq.io.forward(LduCnt + i)
     // forward
     hybridUnits(i).io.ldu_io.sbuffer <> sbuffer.io.forward(LduCnt + i)
+    hybridUnits(i).io.ldu_io.ubuffer <> uncache.io.forward(LduCnt + i)
     // hybridUnits(i).io.ldu_io.vec_forward <> vsFlowQueue.io.forward(LduCnt + i)
     hybridUnits(i).io.ldu_io.vec_forward := DontCare
     hybridUnits(i).io.ldu_io.tl_d_channel := dcache.io.lsu.forward_D(LduCnt + i)
@@ -1038,6 +1041,8 @@ class MemBlockInlinedImp(outer: MemBlockInlined) extends LazyModuleImp(outer)
 
 
     // passdown to lsq (load s2)
+    hybridUnits(i).io.ldu_io.lsq.nc_ldin.valid := false.B
+    hybridUnits(i).io.ldu_io.lsq.nc_ldin.bits := DontCare
     lsq.io.ldu.ldin(LduCnt + i) <> hybridUnits(i).io.ldu_io.lsq.ldin
     // Lsq to sta unit
     lsq.io.sta.storeMaskIn(StaCnt + i) <> hybridUnits(i).io.stu_io.st_mask_out
@@ -1077,7 +1082,7 @@ class MemBlockInlinedImp(outer: MemBlockInlined) extends LazyModuleImp(outer)
   loadMisalignBuffer.io.redirect                <> redirect
   loadMisalignBuffer.io.rob.lcommit             := io.ooo_to_mem.lsqio.lcommit
   loadMisalignBuffer.io.rob.scommit             := io.ooo_to_mem.lsqio.scommit
-  loadMisalignBuffer.io.rob.pendingUncacheld    := io.ooo_to_mem.lsqio.pendingUncacheld
+  loadMisalignBuffer.io.rob.pendingMMIOld       := io.ooo_to_mem.lsqio.pendingMMIOld
   loadMisalignBuffer.io.rob.pendingld           := io.ooo_to_mem.lsqio.pendingld
   loadMisalignBuffer.io.rob.pendingst           := io.ooo_to_mem.lsqio.pendingst
   loadMisalignBuffer.io.rob.pendingVst          := io.ooo_to_mem.lsqio.pendingVst
@@ -1090,7 +1095,7 @@ class MemBlockInlinedImp(outer: MemBlockInlined) extends LazyModuleImp(outer)
   storeMisalignBuffer.io.redirect               <> redirect
   storeMisalignBuffer.io.rob.lcommit            := io.ooo_to_mem.lsqio.lcommit
   storeMisalignBuffer.io.rob.scommit            := io.ooo_to_mem.lsqio.scommit
-  storeMisalignBuffer.io.rob.pendingUncacheld   := io.ooo_to_mem.lsqio.pendingUncacheld
+  storeMisalignBuffer.io.rob.pendingMMIOld      := io.ooo_to_mem.lsqio.pendingMMIOld
   storeMisalignBuffer.io.rob.pendingld          := io.ooo_to_mem.lsqio.pendingld
   storeMisalignBuffer.io.rob.pendingst          := io.ooo_to_mem.lsqio.pendingst
   storeMisalignBuffer.io.rob.pendingVst         := io.ooo_to_mem.lsqio.pendingVst
@@ -1291,7 +1296,7 @@ class MemBlockInlinedImp(outer: MemBlockInlined) extends LazyModuleImp(outer)
   io.mem_to_ooo.lsqio.uop        := lsq.io.rob.uop
   lsq.io.rob.lcommit             := io.ooo_to_mem.lsqio.lcommit
   lsq.io.rob.scommit             := io.ooo_to_mem.lsqio.scommit
-  lsq.io.rob.pendingUncacheld    := io.ooo_to_mem.lsqio.pendingUncacheld
+  lsq.io.rob.pendingMMIOld       := io.ooo_to_mem.lsqio.pendingMMIOld
   lsq.io.rob.pendingld           := io.ooo_to_mem.lsqio.pendingld
   lsq.io.rob.pendingst           := io.ooo_to_mem.lsqio.pendingst
   lsq.io.rob.pendingVst          := io.ooo_to_mem.lsqio.pendingVst
@@ -1313,7 +1318,7 @@ class MemBlockInlinedImp(outer: MemBlockInlined) extends LazyModuleImp(outer)
     )).andR))
     resultOnehot
   }
-  val allRedirect = loadUnits.map(_.io.rollback) ++ hybridUnits.map(_.io.ldu_io.rollback) ++ Seq(lsq.io.nack_rollback) ++ lsq.io.nuke_rollback
+  val allRedirect = loadUnits.map(_.io.rollback) ++ hybridUnits.map(_.io.ldu_io.rollback) ++ lsq.io.nack_rollback ++ lsq.io.nuke_rollback
   val oldestOneHot = selectOldestRedirect(allRedirect)
   val oldestRedirect = WireDefault(Mux1H(oldestOneHot, allRedirect))
   // memory replay would not cause IAF/IPF/IGPF
@@ -1344,8 +1349,7 @@ class MemBlockInlinedImp(outer: MemBlockInlined) extends LazyModuleImp(outer)
     is (s_idle) {
       when (uncacheReq.fire) {
         when (lsq.io.uncache.req.valid) {
-          val isStore = lsq.io.uncache.req.bits.cmd === MemoryOpConstants.M_XWR
-          when (!isStore || !io.ooo_to_mem.csrCtrl.uncache_write_outstanding_enable) {
+          when (!lsq.io.uncache.req.bits.nc || !io.ooo_to_mem.csrCtrl.uncache_write_outstanding_enable) {
             uncacheState := s_scalar_uncache
           }
         }.otherwise {

src/main/scala/xiangshan/backend/rob/Rob.scala

@@ -530,7 +530,7 @@ class RobImp(override val wrapper: Rob)(implicit p: Parameters, params: BackendP
       XSInfo(true.B,
         p"writebacked pc 0x${Hexadecimal(debug_Uop.pc)} wen ${debug_Uop.rfWen} " +
           p"data 0x${Hexadecimal(wb.bits.data(0))} ldst ${debug_Uop.ldest} pdst ${debug_Uop.pdest} " +
-          p"skip ${wb.bits.debug.isMMIO} robIdx: ${wb.bits.robIdx}\n"
+          p"skip ${wb.bits.debug.isSkipDiff} robIdx: ${wb.bits.robIdx}\n"
       )
     }
   }
@@ -799,7 +799,7 @@ class RobImp(override val wrapper: Rob)(implicit p: Parameters, params: BackendP
   io.lsq.lcommit := RegNext(Mux(io.commits.isCommit, PopCount(ldCommitVec), 0.U))
   io.lsq.scommit := RegNext(Mux(io.commits.isCommit, PopCount(stCommitVec), 0.U))
   // indicate a pending load or store
-  io.lsq.pendingUncacheld := RegNext(io.commits.isCommit && io.commits.info(0).commitType === CommitType.LOAD && robEntries(deqPtr.value).valid && robEntries(deqPtr.value).mmio)
+  io.lsq.pendingMMIOld := RegNext(io.commits.isCommit && io.commits.info(0).commitType === CommitType.LOAD && robEntries(deqPtr.value).valid && robEntries(deqPtr.value).mmio)
   io.lsq.pendingld := RegNext(io.commits.isCommit && io.commits.info(0).commitType === CommitType.LOAD && robEntries(deqPtr.value).valid)
   // TODO: Check if need deassert pendingst when it is vst
   io.lsq.pendingst := RegNext(io.commits.isCommit && io.commits.info(0).commitType === CommitType.STORE && robEntries(deqPtr.value).valid)
@@ -1436,7 +1436,7 @@ class RobImp(override val wrapper: Rob)(implicit p: Parameters, params: BackendP
       val isRVC = dt_isRVC(ptr)
 
       val difftest = DifftestModule(new DiffInstrCommit(MaxPhyRegs), delay = 3, dontCare = true)
-      val dt_skip = Mux(eliminatedMove, false.B, exuOut.isMMIO || exuOut.isPerfCnt)
+      val dt_skip = Mux(eliminatedMove, false.B, exuOut.isSkipDiff)
       difftest.coreid := io.hartId
       difftest.index := i.U
       difftest.valid := io.commits.commitValid(i) && io.commits.isCommit

src/main/scala/xiangshan/backend/rob/RobBundles.scala

@@ -231,7 +231,7 @@ class RobCSRIO(implicit p: Parameters) extends XSBundle {
 class RobLsqIO(implicit p: Parameters) extends XSBundle {
   val lcommit = Output(UInt(log2Up(CommitWidth + 1).W))
   val scommit = Output(UInt(log2Up(CommitWidth + 1).W))
-  val pendingUncacheld = Output(Bool())
+  val pendingMMIOld = Output(Bool())
   val pendingld = Output(Bool())
   val pendingst = Output(Bool())
   // set when vector store at the head of ROB

src/main/scala/xiangshan/cache/dcache/DCacheWrapper.scala

@@ -513,11 +513,13 @@ class UncacheWordReq(implicit p: Parameters) extends DCacheBundle
 {
   val cmd  = UInt(M_SZ.W)
   val addr = UInt(PAddrBits.W)
+  val vaddr = UInt(VAddrBits.W) // for uncache buffer forwarding
   val data = UInt(XLEN.W)
   val mask = UInt((XLEN/8).W)
   val id   = UInt(uncacheIdxBits.W)
   val instrtype = UInt(sourceTypeWidth.W)
   val atomic = Bool()
+  val nc = Bool()
   val isFirstIssue = Bool()
   val replayCarry = new ReplayCarry(nWays)
 
@@ -531,7 +533,9 @@ class UncacheWordResp(implicit p: Parameters) extends DCacheBundle
 {
   val data      = UInt(XLEN.W)
   val data_delayed = UInt(XLEN.W)
-  val id        = UInt(uncacheIdxBits.W)
+  val id        = UInt(uncacheIdxBits.W) // resp identified signals
+  val nc        = Bool() // resp identified signals
+  val is2lq     = Bool() // resp identified signals
   val miss      = Bool()
   val replay    = Bool()
   val tag_error = Bool()