Merge pull request #993 from freechipsproject/auto-diplomacy-bundles

Auto diplomacy bundles

macros/src/main/scala/ValName.scala

@@ -0,0 +1,23 @@
+// See LICENSE.SiFive for license details.
+
+package freechips.rocketchip.macros
+
+import scala.language.experimental.macros
+import scala.reflect.macros.blackbox.Context
+
+case class ValNameImpl(name: String)
+
+object ValNameImpl
+{
+  implicit def materialize: ValNameImpl = macro detail
+  def detail(c: Context): c.Expr[ValNameImpl] = {
+    import c.universe._
+    def allOwners(s: c.Symbol): Seq[c.Symbol] =
+      if (s == `NoSymbol`) Nil else s +: allOwners(s.owner)
+    val terms = allOwners(c.internal.enclosingOwner).filter(_.isTerm).map(_.asTerm)
+    terms.filter(_.isVal).map(_.name.toString).find(_(0) != '$').map { s =>
+      val trim = s.replaceAll("\\s", "")
+      c.Expr[ValNameImpl] { q"_root_.freechips.rocketchip.macros.ValNameImpl(${trim})" }
+    }.getOrElse(c.abort(c.enclosingPosition, "Not a valid application."))
+  }
+}

project/build.scala

@@ -21,7 +21,8 @@ object BuildSettings extends Build {
 
   lazy val chisel = project in file("chisel3")
   lazy val hardfloat  = project.dependsOn(chisel)
-  lazy val rocketchip = (project in file(".")).settings(chipSettings).dependsOn(chisel, hardfloat)
+  lazy val macros = project in file("macros")
+  lazy val rocketchip = (project in file(".")).settings(chipSettings).dependsOn(chisel, hardfloat, macros)
 
   lazy val addons = settingKey[Seq[String]]("list of addons used for this build")
   lazy val make = inputKey[Unit]("trigger backend-specific makefile command")

src/main/scala/amba/ahb/Nodes.scala

@@ -4,19 +4,14 @@ package freechips.rocketchip.amba.ahb
 
 import Chisel._
 import chisel3.internal.sourceinfo.SourceInfo
+import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 
-object AHBImp extends NodeImp[AHBMasterPortParameters, AHBSlavePortParameters, AHBEdgeParameters, AHBEdgeParameters, AHBBundle]
+object AHBImp extends SimpleNodeImp[AHBMasterPortParameters, AHBSlavePortParameters, AHBEdgeParameters, AHBBundle]
 {
-  def edgeO(pd: AHBMasterPortParameters, pu: AHBSlavePortParameters): AHBEdgeParameters = AHBEdgeParameters(pd, pu)
-  def edgeI(pd: AHBMasterPortParameters, pu: AHBSlavePortParameters): AHBEdgeParameters = AHBEdgeParameters(pd, pu)
-
-  def bundleO(eo: AHBEdgeParameters): AHBBundle = AHBBundle(eo.bundle)
-  def bundleI(ei: AHBEdgeParameters): AHBBundle = AHBBundle(ei.bundle)
-
-  def colour = "#00ccff" // bluish
-  override def labelI(ei: AHBEdgeParameters) = (ei.slave.beatBytes * 8).toString
-  override def labelO(eo: AHBEdgeParameters) = (eo.slave.beatBytes * 8).toString
+  def edge(pd: AHBMasterPortParameters, pu: AHBSlavePortParameters, p: Parameters, sourceInfo: SourceInfo) = AHBEdgeParameters(pd, pu, p, sourceInfo)
+  def bundle(e: AHBEdgeParameters) = AHBBundle(e.bundle)
+  def render(e: AHBEdgeParameters) = RenderedEdge(colour = "#00ccff" /* bluish */, label = (e.slave.beatBytes * 8).toString)
 
   override def mixO(pd: AHBMasterPortParameters, node: OutwardNode[AHBMasterPortParameters, AHBSlavePortParameters, AHBBundle]): AHBMasterPortParameters  =
    pd.copy(masters = pd.masters.map  { c => c.copy (nodePath = node +: c.nodePath) })
@@ -25,23 +20,14 @@ object AHBImp extends NodeImp[AHBMasterPortParameters, AHBSlavePortParameters, A
 }
 
 // Nodes implemented inside modules
-case class AHBIdentityNode() extends IdentityNode(AHBImp)
-case class AHBMasterNode(portParams: Seq[AHBMasterPortParameters]) extends SourceNode(AHBImp)(portParams)
-case class AHBSlaveNode(portParams: Seq[AHBSlavePortParameters]) extends SinkNode(AHBImp)(portParams)
+case class AHBMasterNode(portParams: Seq[AHBMasterPortParameters])(implicit valName: ValName) extends SourceNode(AHBImp)(portParams)
+case class AHBSlaveNode(portParams: Seq[AHBSlavePortParameters])(implicit valName: ValName) extends SinkNode(AHBImp)(portParams)
 case class AHBNexusNode(
   masterFn:       Seq[AHBMasterPortParameters] => AHBMasterPortParameters,
   slaveFn:        Seq[AHBSlavePortParameters]  => AHBSlavePortParameters,
   numMasterPorts: Range.Inclusive = 1 to 999,
-  numSlavePorts:  Range.Inclusive = 1 to 999)
+  numSlavePorts:  Range.Inclusive = 1 to 999)(
+  implicit valName: ValName)
   extends NexusNode(AHBImp)(masterFn, slaveFn, numMasterPorts, numSlavePorts)
 
-// Nodes passed from an inner module
-case class AHBOutputNode() extends OutputNode(AHBImp)
-case class AHBInputNode() extends InputNode(AHBImp)
-
-// Nodes used for external ports
-case class AHBBlindOutputNode(portParams: Seq[AHBSlavePortParameters]) extends BlindOutputNode(AHBImp)(portParams)
-case class AHBBlindInputNode(portParams: Seq[AHBMasterPortParameters]) extends BlindInputNode(AHBImp)(portParams)
-
-case class AHBInternalOutputNode(portParams: Seq[AHBSlavePortParameters]) extends InternalOutputNode(AHBImp)(portParams)
-case class AHBInternalInputNode(portParams: Seq[AHBMasterPortParameters]) extends InternalInputNode(AHBImp)(portParams)
+case class AHBIdentityNode()(implicit valName: ValName) extends IdentityNode(AHBImp)()

src/main/scala/amba/ahb/Parameters.scala

@@ -3,6 +3,8 @@
 package freechips.rocketchip.amba.ahb
 
 import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
+import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 import scala.math.max
 
@@ -90,7 +92,9 @@ object AHBBundleParameters
 
 case class AHBEdgeParameters(
   master: AHBMasterPortParameters,
-  slave:  AHBSlavePortParameters)
+  slave:  AHBSlavePortParameters,
+  params: Parameters,
+  sourceInfo: SourceInfo)
 {
   val bundle = AHBBundleParameters(master, slave)
 }

src/main/scala/amba/ahb/RegisterRouter.scala

@@ -10,8 +10,8 @@ import freechips.rocketchip.tilelink.{IntSourceNode, IntSourcePortSimple}
 import freechips.rocketchip.util.{HeterogeneousBag, MaskGen}
 import scala.math.{min,max}
 
-class AHBRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)
-  extends AHBSlaveNode(Seq(AHBSlavePortParameters(
+case class AHBRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)(implicit valName: ValName)
+  extends SinkNode(AHBImp)(Seq(AHBSlavePortParameters(
     Seq(AHBSlaveParameters(
       address       = Seq(address),
       executable    = executable,
@@ -24,7 +24,7 @@ class AHBRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
   // Calling this method causes the matching AHB bundle to be
   // configured to route all requests to the listed RegFields.
   def regmap(mapping: RegField.Map*) = {
-    val ahb = bundleIn(0)
+    val (ahb, _) = this.in(0)
 
     val indexBits = log2Up((address.mask+1)/beatBytes)
     val params = RegMapperParams(indexBits, beatBytes, 1)
@@ -67,12 +67,6 @@ class AHBRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
   }
 }
 
-object AHBRegisterNode
-{
-  def apply(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false) =
-    new AHBRegisterNode(address, concurrency, beatBytes, undefZero, executable)
-}
-
 // These convenience methods below combine to make it possible to create a AHB
 // register mapped device from a totally abstract register mapped device.
 
@@ -82,13 +76,11 @@ abstract class AHBRegisterRouterBase(address: AddressSet, interrupts: Int, concu
   val intnode = IntSourceNode(IntSourcePortSimple(num = interrupts))
 }
 
-case class AHBRegBundleArg(interrupts: HeterogeneousBag[Vec[Bool]], in: HeterogeneousBag[AHBBundle])(implicit val p: Parameters)
+case class AHBRegBundleArg()(implicit val p: Parameters)
 
 class AHBRegBundleBase(arg: AHBRegBundleArg) extends Bundle
 {
   implicit val p = arg.p
-  val interrupts = arg.interrupts
-  val in = arg.in
 }
 
 class AHBRegBundle[P](val params: P, arg: AHBRegBundleArg) extends AHBRegBundleBase(arg)
@@ -96,8 +88,8 @@ class AHBRegBundle[P](val params: P, arg: AHBRegBundleArg) extends AHBRegBundleB
 class AHBRegModule[P, B <: AHBRegBundleBase](val params: P, bundleBuilder: => B, router: AHBRegisterRouterBase)
   extends LazyModuleImp(router) with HasRegMap
 {
-  val io = bundleBuilder
-  val interrupts = if (io.interrupts.isEmpty) Vec(0, Bool()) else io.interrupts(0)
+  val io = IO(bundleBuilder)
+  val interrupts = if (router.intnode.out.isEmpty) Vec(0, Bool()) else router.intnode.out(0)._1
   def regmap(mapping: RegField.Map*) = router.node.regmap(mapping:_*)
 }
 
@@ -110,5 +102,5 @@ class AHBRegisterRouter[B <: AHBRegBundleBase, M <: LazyModuleImp]
   require (isPow2(size))
   // require (size >= 4096) ... not absolutely required, but highly recommended
 
-  lazy val module = moduleBuilder(bundleBuilder(AHBRegBundleArg(intnode.bundleOut, node.bundleIn)), this)
+  lazy val module = moduleBuilder(bundleBuilder(AHBRegBundleArg()), this)
 }

src/main/scala/amba/ahb/SRAM.scala

@@ -22,15 +22,11 @@ class AHBRAM(address: AddressSet, executable: Boolean = true, beatBytes: Int = 4
   require ((address.mask & (beatBytes-1)) == beatBytes-1)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-
     def bigBits(x: BigInt, tail: List[Boolean] = List.empty[Boolean]): List[Boolean] =
       if (x == 0) tail.reverse else bigBits(x >> 1, ((x & 1) == 1) :: tail)
     val mask = bigBits(address.mask >> log2Ceil(beatBytes))
 
-    val in = io.in(0)
+    val (in, _) = node.in(0)
 
     // The mask and address during the address phase
     val a_access    = in.htrans === AHBParameters.TRANS_NONSEQ || in.htrans === AHBParameters.TRANS_SEQ

src/main/scala/amba/ahb/Test.scala

@@ -21,7 +21,7 @@ class AHBFuzzNative(aFlow: Boolean, txns: Int)(implicit p: Parameters) extends L
 {
   val fuzz  = LazyModule(new TLFuzzer(txns))
   val model = LazyModule(new TLRAMModel("AHBFuzzNative"))
-  var xbar  = LazyModule(new AHBFanout)
+  val xbar  = LazyModule(new AHBFanout)
   val ram   = LazyModule(new AHBRAM(AddressSet(0x0, 0xff)))
   val gpio  = LazyModule(new RRTest0(0x100))
 
@@ -30,7 +30,7 @@ class AHBFuzzNative(aFlow: Boolean, txns: Int)(implicit p: Parameters) extends L
   ram.node  := xbar.node
   gpio.node := xbar.node
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }
@@ -42,7 +42,7 @@ class AHBNativeTest(aFlow: Boolean, txns: Int = 5000, timeout: Int = 500000)(imp
 
 class AHBFuzzMaster(aFlow: Boolean, txns: Int)(implicit p: Parameters) extends LazyModule
 {
-  val node  = AHBOutputNode()
+  val node  = AHBIdentityNode()
   val fuzz  = LazyModule(new TLFuzzer(txns))
   val model = LazyModule(new TLRAMModel("AHBFuzzMaster"))
 
@@ -55,10 +55,9 @@ class AHBFuzzMaster(aFlow: Boolean, txns: Int)(implicit p: Parameters) extends L
     model.node))))
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val out = node.bundleOut
+    val io = IO(new Bundle {
       val finished = Bool(OUTPUT)
-    }
+    })
 
     io.finished := fuzz.module.io.finished
   }
@@ -66,7 +65,7 @@ class AHBFuzzMaster(aFlow: Boolean, txns: Int)(implicit p: Parameters) extends L
 
 class AHBFuzzSlave()(implicit p: Parameters) extends LazyModule
 {
-  val node = AHBInputNode()
+  val node = AHBIdentityNode()
   val ram  = LazyModule(new TLTestRAM(AddressSet(0x0, 0xfff)))
 
   ram.node :=
@@ -77,11 +76,7 @@ class AHBFuzzSlave()(implicit p: Parameters) extends LazyModule
     AHBToTL()(
     node)))))
 
-  lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-  }
+  lazy val module = new LazyModuleImp(this) { }
 }
 
 class AHBFuzzBridge(aFlow: Boolean, txns: Int)(implicit p: Parameters) extends LazyModule
@@ -91,7 +86,7 @@ class AHBFuzzBridge(aFlow: Boolean, txns: Int)(implicit p: Parameters) extends L
 
   slave.node := master.node
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := master.module.io.finished
   }
 }

src/main/scala/amba/ahb/ToTL.scala

@@ -9,7 +9,7 @@ import freechips.rocketchip.diplomacy._
 import freechips.rocketchip.tilelink._
 import freechips.rocketchip.util.MaskGen
 
-case class AHBToTLNode() extends MixedAdapterNode(AHBImp, TLImp)(
+case class AHBToTLNode()(implicit valName: ValName) extends MixedAdapterNode(AHBImp, TLImp)(
   dFn = { case AHBMasterPortParameters(masters) =>
     TLClientPortParameters(clients = masters.map { m =>
       TLClientParameters(name = m.name, nodePath = m.nodePath)
@@ -41,12 +41,7 @@ class AHBToTL()(implicit p: Parameters) extends LazyModule
   val node = AHBToTLNode()
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val beatBytes = edgeOut.manager.beatBytes
 
       val d_send  = RegInit(Bool(false))

src/main/scala/amba/ahb/Xbar.scala

@@ -16,35 +16,31 @@ class AHBFanout()(implicit p: Parameters) extends LazyModule {
     slaveFn  = { seq => seq(0).copy(slaves = seq.flatMap(_.slaves)) })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-      val out = node.bundleOut
-    }
-
     // Require consistent bus widths
-    val port0 = node.edgesOut(0).slave
-    node.edgesOut.foreach { edge =>
+    val (io_out, edgesOut) = node.out.unzip
+    val port0 = edgesOut(0).slave
+    edgesOut.foreach { edge =>
       val port = edge.slave
       require (port.beatBytes == port0.beatBytes,
         s"${port.slaves.map(_.name)} ${port.beatBytes} vs ${port0.slaves.map(_.name)} ${port0.beatBytes}")
     }
 
-    val port_addrs = node.edgesOut.map(_.slave.slaves.map(_.address).flatten)
+    val port_addrs = edgesOut.map(_.slave.slaves.map(_.address).flatten)
     val routingMask = AddressDecoder(port_addrs)
     val route_addrs = port_addrs.map(_.map(_.widen(~routingMask)).distinct)
 
-    val in = io.in(0)
+    val (in, _) = node.in(0)
     val a_sel = Vec(route_addrs.map(seq => seq.map(_.contains(in.haddr)).reduce(_ || _)))
     val d_sel = Reg(a_sel)
 
     when (in.hready) { d_sel := a_sel }
-    (a_sel zip io.out) foreach { case (sel, out) =>
+    (a_sel zip io_out) foreach { case (sel, out) =>
       out := in
       out.hsel := in.hsel && sel
     }
 
-    in.hreadyout := !Mux1H(d_sel, io.out.map(!_.hreadyout))
-    in.hresp     :=  Mux1H(d_sel, io.out.map(_.hresp))
-    in.hrdata    :=  Mux1H(d_sel, io.out.map(_.hrdata))
+    in.hreadyout := !Mux1H(d_sel, io_out.map(!_.hreadyout))
+    in.hresp     :=  Mux1H(d_sel, io_out.map(_.hresp))
+    in.hrdata    :=  Mux1H(d_sel, io_out.map(_.hrdata))
   }
 }

src/main/scala/amba/apb/Nodes.scala

@@ -7,17 +7,11 @@ import chisel3.internal.sourceinfo.SourceInfo
 import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 
-object APBImp extends NodeImp[APBMasterPortParameters, APBSlavePortParameters, APBEdgeParameters, APBEdgeParameters, APBBundle]
+object APBImp extends SimpleNodeImp[APBMasterPortParameters, APBSlavePortParameters, APBEdgeParameters, APBBundle]
 {
-  def edgeO(pd: APBMasterPortParameters, pu: APBSlavePortParameters): APBEdgeParameters = APBEdgeParameters(pd, pu)
-  def edgeI(pd: APBMasterPortParameters, pu: APBSlavePortParameters): APBEdgeParameters = APBEdgeParameters(pd, pu)
-
-  def bundleO(eo: APBEdgeParameters): APBBundle = APBBundle(eo.bundle)
-  def bundleI(ei: APBEdgeParameters): APBBundle = APBBundle(ei.bundle)
-
-  def colour = "#00ccff" // bluish
-  override def labelI(ei: APBEdgeParameters) = (ei.slave.beatBytes * 8).toString
-  override def labelO(eo: APBEdgeParameters) = (eo.slave.beatBytes * 8).toString
+  def edge(pd: APBMasterPortParameters, pu: APBSlavePortParameters, p: Parameters, sourceInfo: SourceInfo) = APBEdgeParameters(pd, pu, p, sourceInfo)
+  def bundle(e: APBEdgeParameters) = APBBundle(e.bundle)
+  def render(e: APBEdgeParameters) = RenderedEdge(colour = "#00ccff" /* bluish */, (e.slave.beatBytes * 8).toString)
 
   override def mixO(pd: APBMasterPortParameters, node: OutwardNode[APBMasterPortParameters, APBSlavePortParameters, APBBundle]): APBMasterPortParameters  =
    pd.copy(masters = pd.masters.map  { c => c.copy (nodePath = node +: c.nodePath) })
@@ -25,24 +19,14 @@ object APBImp extends NodeImp[APBMasterPortParameters, APBSlavePortParameters, A
    pu.copy(slaves  = pu.slaves.map { m => m.copy (nodePath = node +: m.nodePath) })
 }
 
-// Nodes implemented inside modules
-case class APBIdentityNode() extends IdentityNode(APBImp)
-case class APBMasterNode(portParams: Seq[APBMasterPortParameters]) extends SourceNode(APBImp)(portParams)
-case class APBSlaveNode(portParams: Seq[APBSlavePortParameters]) extends SinkNode(APBImp)(portParams)
+case class APBMasterNode(portParams: Seq[APBMasterPortParameters])(implicit valName: ValName) extends SourceNode(APBImp)(portParams)
+case class APBSlaveNode(portParams: Seq[APBSlavePortParameters])(implicit valName: ValName) extends SinkNode(APBImp)(portParams)
 case class APBNexusNode(
   masterFn:       Seq[APBMasterPortParameters] => APBMasterPortParameters,
   slaveFn:        Seq[APBSlavePortParameters]  => APBSlavePortParameters,
   numMasterPorts: Range.Inclusive = 1 to 1,
-  numSlavePorts:  Range.Inclusive = 1 to 1)
+  numSlavePorts:  Range.Inclusive = 1 to 1)(
+  implicit valName: ValName)
   extends NexusNode(APBImp)(masterFn, slaveFn, numMasterPorts, numSlavePorts)
 
-// Nodes passed from an inner module
-case class APBOutputNode() extends OutputNode(APBImp)
-case class APBInputNode() extends InputNode(APBImp)
-
-// Nodes used for external ports
-case class APBBlindOutputNode(portParams: Seq[APBSlavePortParameters]) extends BlindOutputNode(APBImp)(portParams)
-case class APBBlindInputNode(portParams: Seq[APBMasterPortParameters]) extends BlindInputNode(APBImp)(portParams)
-
-case class APBInternalOutputNode(portParams: Seq[APBSlavePortParameters]) extends InternalOutputNode(APBImp)(portParams)
-case class APBInternalInputNode(portParams: Seq[APBMasterPortParameters]) extends InternalInputNode(APBImp)(portParams)
+case class APBIdentityNode()(implicit valName: ValName) extends IdentityNode(APBImp)()

src/main/scala/amba/apb/Parameters.scala

@@ -3,6 +3,7 @@
 package freechips.rocketchip.amba.apb
 
 import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
 import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 import scala.math.max
@@ -78,7 +79,9 @@ object APBBundleParameters
 
 case class APBEdgeParameters(
   master: APBMasterPortParameters,
-  slave:  APBSlavePortParameters)
+  slave:  APBSlavePortParameters,
+  params: Parameters,
+  sourceInfo: SourceInfo)
 {
   val bundle = APBBundleParameters(master, slave)
 }

src/main/scala/amba/apb/RegisterRouter.scala

@@ -10,8 +10,8 @@ import freechips.rocketchip.tilelink.{IntSourceNode, IntSourcePortSimple}
 import freechips.rocketchip.util.HeterogeneousBag
 import scala.math.{min,max}
 
-class APBRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)
-  extends APBSlaveNode(Seq(APBSlavePortParameters(
+case class APBRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)(implicit valName: ValName)
+  extends SinkNode(APBImp)(Seq(APBSlavePortParameters(
     Seq(APBSlaveParameters(
       address       = Seq(address),
       executable    = executable,
@@ -24,7 +24,7 @@ class APBRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
   // Calling this method causes the matching APB bundle to be
   // configured to route all requests to the listed RegFields.
   def regmap(mapping: RegField.Map*) = {
-    val apb = bundleIn(0)
+    val (apb, _) = this.in(0)
 
     val indexBits = log2Up((address.mask+1)/beatBytes)
     val params = RegMapperParams(indexBits, beatBytes, 1)
@@ -51,12 +51,6 @@ class APBRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
   }
 }
 
-object APBRegisterNode
-{
-  def apply(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false) =
-    new APBRegisterNode(address, concurrency, beatBytes, undefZero, executable)
-}
-
 // These convenience methods below combine to make it possible to create a APB
 // register mapped device from a totally abstract register mapped device.
 
@@ -66,13 +60,11 @@ abstract class APBRegisterRouterBase(address: AddressSet, interrupts: Int, concu
   val intnode = IntSourceNode(IntSourcePortSimple(num = interrupts))
 }
 
-case class APBRegBundleArg(interrupts: HeterogeneousBag[Vec[Bool]], in: HeterogeneousBag[APBBundle])(implicit val p: Parameters)
+case class APBRegBundleArg()(implicit val p: Parameters)
 
 class APBRegBundleBase(arg: APBRegBundleArg) extends Bundle
 {
   implicit val p = arg.p
-  val interrupts = arg.interrupts
-  val in = arg.in
 }
 
 class APBRegBundle[P](val params: P, arg: APBRegBundleArg) extends APBRegBundleBase(arg)
@@ -80,8 +72,8 @@ class APBRegBundle[P](val params: P, arg: APBRegBundleArg) extends APBRegBundleB
 class APBRegModule[P, B <: APBRegBundleBase](val params: P, bundleBuilder: => B, router: APBRegisterRouterBase)
   extends LazyModuleImp(router) with HasRegMap
 {
-  val io = bundleBuilder
-  val interrupts = if (io.interrupts.isEmpty) Vec(0, Bool()) else io.interrupts(0)
+  val io = IO(bundleBuilder)
+  val interrupts = if (router.intnode.out.isEmpty) Vec(0, Bool()) else router.intnode.out(0)._1
   def regmap(mapping: RegField.Map*) = router.node.regmap(mapping:_*)
 }
 
@@ -94,5 +86,5 @@ class APBRegisterRouter[B <: APBRegBundleBase, M <: LazyModuleImp]
   require (isPow2(size))
   // require (size >= 4096) ... not absolutely required, but highly recommended
 
-  lazy val module = moduleBuilder(bundleBuilder(APBRegBundleArg(intnode.bundleOut, node.bundleIn)), this)
+  lazy val module = moduleBuilder(bundleBuilder(APBRegBundleArg()), this)
 }

src/main/scala/amba/apb/SRAM.scala

@@ -22,11 +22,7 @@ class APBRAM(address: AddressSet, executable: Boolean = true, beatBytes: Int = 4
   require ((address.mask & (beatBytes-1)) == beatBytes-1)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-
-    val in = io.in(0)
+    val (in, _) = node.in(0)
 
     def bigBits(x: BigInt, tail: List[Boolean] = List.empty[Boolean]): List[Boolean] =
       if (x == 0) tail.reverse else bigBits(x >> 1, ((x & 1) == 1) :: tail)

src/main/scala/amba/apb/Test.scala

@@ -20,7 +20,7 @@ class APBFuzzBridge(aFlow: Boolean, txns: Int)(implicit p: Parameters) extends L
 {
   val fuzz  = LazyModule(new TLFuzzer(txns))
   val model = LazyModule(new TLRAMModel("APBFuzzMaster"))
-  var xbar  = LazyModule(new APBFanout)
+  val xbar  = LazyModule(new APBFanout)
   val ram   = LazyModule(new APBRAM(AddressSet(0x0, 0xff)))
   val gpio  = LazyModule(new RRTest0(0x100))
 
@@ -34,7 +34,7 @@ class APBFuzzBridge(aFlow: Boolean, txns: Int)(implicit p: Parameters) extends L
     TLDelayer(0.2)(
     model.node))))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }

src/main/scala/amba/apb/Xbar.scala

@@ -16,34 +16,30 @@ class APBFanout()(implicit p: Parameters) extends LazyModule {
     slaveFn  = { seq => seq(0).copy(slaves = seq.flatMap(_.slaves)) })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    val in = io.in(0)
+    val (in, _) = node.in(0)
 
     // Require consistent bus widths
-    val port0 = node.edgesOut(0).slave
-    node.edgesOut.foreach { edge =>
+    val (io_out, edgesOut) = node.out.unzip
+    val port0 = edgesOut(0).slave
+    edgesOut.foreach { edge =>
       val port = edge.slave
       require (port.beatBytes == port0.beatBytes,
         s"${port.slaves.map(_.name)} ${port.beatBytes} vs ${port0.slaves.map(_.name)} ${port0.beatBytes}")
     }
 
-    val port_addrs = node.edgesOut.map(_.slave.slaves.map(_.address).flatten)
+    val port_addrs = edgesOut.map(_.slave.slaves.map(_.address).flatten)
     val routingMask = AddressDecoder(port_addrs)
     val route_addrs = port_addrs.map(_.map(_.widen(~routingMask)).distinct)
 
     val sel = Vec(route_addrs.map(seq => seq.map(_.contains(in.paddr)).reduce(_ || _)))
-    (sel zip io.out) foreach { case (sel, out) =>
+    (sel zip io_out) foreach { case (sel, out) =>
       out := in
       out.psel    := sel && in.psel
       out.penable := sel && in.penable
     }
 
-    in.pready  := !Mux1H(sel, io.out.map(!_.pready))
-    in.pslverr :=  Mux1H(sel, io.out.map(_.pslverr))
-    in.prdata  :=  Mux1H(sel, io.out.map(_.prdata))
+    in.pready  := !Mux1H(sel, io_out.map(!_.pready))
+    in.pslverr :=  Mux1H(sel, io_out.map(_.pslverr))
+    in.prdata  :=  Mux1H(sel, io_out.map(_.prdata))
   }
 }

src/main/scala/amba/axi4/Buffer.scala

@@ -26,11 +26,6 @@ class AXI4Buffer(
     slaveFn  = { p => p.copy(minLatency = p.minLatency + min(aw.latency,ar.latency) + min(r.latency,b.latency)) })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
     def buffer[T <: Data](config: BufferParams, data: IrrevocableIO[T]): IrrevocableIO[T] = {
       if (config.isDefined) {
         Queue.irrevocable(data, config.depth, pipe=config.pipe, flow=config.flow)
@@ -39,7 +34,7 @@ class AXI4Buffer(
       }
     }
 
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       out.aw <> buffer(aw, in .aw)
       out.w  <> buffer(w,  in .w)
       in .b  <> buffer(b,  out.b)

src/main/scala/amba/axi4/Deinterleaver.scala

@@ -22,12 +22,7 @@ class AXI4Deinterleaver(maxReadBytes: Int)(implicit p: Parameters) extends LazyM
   })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val endId = edgeOut.master.endId
       val beatBytes = edgeOut.slave.beatBytes
       val beats = (maxReadBytes+beatBytes-1) / beatBytes

src/main/scala/amba/axi4/Fragmenter.scala

@@ -26,12 +26,7 @@ class AXI4Fragmenter()(implicit p: Parameters) extends LazyModule
     slaveFn  = { sp => sp.copy(slaves  = sp.slaves .map(s => mapSlave(s, sp.beatBytes))) })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val slave     = edgeOut.slave
       val slaves    = slave.slaves
       val beatBytes = slave.beatBytes

src/main/scala/amba/axi4/IdIndexer.scala

@@ -42,12 +42,7 @@ class AXI4IdIndexer(idBits: Int)(implicit p: Parameters) extends LazyModule
     })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
 
       // Leave everything mostly untouched
       out.ar <> in.ar

src/main/scala/amba/axi4/Nodes.scala

@@ -7,17 +7,11 @@ import chisel3.internal.sourceinfo.SourceInfo
 import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 
-object AXI4Imp extends NodeImp[AXI4MasterPortParameters, AXI4SlavePortParameters, AXI4EdgeParameters, AXI4EdgeParameters, AXI4Bundle]
+object AXI4Imp extends SimpleNodeImp[AXI4MasterPortParameters, AXI4SlavePortParameters, AXI4EdgeParameters, AXI4Bundle]
 {
-  def edgeO(pd: AXI4MasterPortParameters, pu: AXI4SlavePortParameters): AXI4EdgeParameters = AXI4EdgeParameters(pd, pu)
-  def edgeI(pd: AXI4MasterPortParameters, pu: AXI4SlavePortParameters): AXI4EdgeParameters = AXI4EdgeParameters(pd, pu)
-
-  def bundleO(eo: AXI4EdgeParameters): AXI4Bundle = AXI4Bundle(eo.bundle)
-  def bundleI(ei: AXI4EdgeParameters): AXI4Bundle = AXI4Bundle(ei.bundle)
-
-  def colour = "#00ccff" // bluish
-  override def labelI(ei: AXI4EdgeParameters) = (ei.slave.beatBytes * 8).toString
-  override def labelO(eo: AXI4EdgeParameters) = (eo.slave.beatBytes * 8).toString
+  def edge(pd: AXI4MasterPortParameters, pu: AXI4SlavePortParameters, p: Parameters, sourceInfo: SourceInfo) = AXI4EdgeParameters(pd, pu, p, sourceInfo)
+  def bundle(e: AXI4EdgeParameters) = AXI4Bundle(e.bundle)
+  def render(e: AXI4EdgeParameters) = RenderedEdge(colour = "#00ccff" /* bluish */, label  = (e.slave.beatBytes * 8).toString)
 
   override def mixO(pd: AXI4MasterPortParameters, node: OutwardNode[AXI4MasterPortParameters, AXI4SlavePortParameters, AXI4Bundle]): AXI4MasterPortParameters  =
    pd.copy(masters = pd.masters.map  { c => c.copy (nodePath = node +: c.nodePath) })
@@ -25,23 +19,12 @@ object AXI4Imp extends NodeImp[AXI4MasterPortParameters, AXI4SlavePortParameters
    pu.copy(slaves  = pu.slaves.map { m => m.copy (nodePath = node +: m.nodePath) })
 }
 
-// Nodes implemented inside modules
-case class AXI4IdentityNode() extends IdentityNode(AXI4Imp)
-case class AXI4MasterNode(portParams: Seq[AXI4MasterPortParameters]) extends SourceNode(AXI4Imp)(portParams)
-case class AXI4SlaveNode(portParams: Seq[AXI4SlavePortParameters]) extends SinkNode(AXI4Imp)(portParams)
+case class AXI4MasterNode(portParams: Seq[AXI4MasterPortParameters])(implicit valName: ValName) extends SourceNode(AXI4Imp)(portParams)
+case class AXI4SlaveNode(portParams: Seq[AXI4SlavePortParameters])(implicit valName: ValName) extends SinkNode(AXI4Imp)(portParams)
 case class AXI4AdapterNode(
   masterFn:  AXI4MasterPortParameters => AXI4MasterPortParameters,
   slaveFn:   AXI4SlavePortParameters  => AXI4SlavePortParameters,
-  numPorts:  Range.Inclusive = 0 to 999)
+  numPorts:  Range.Inclusive = 0 to 999)(
+  implicit valName: ValName)
   extends AdapterNode(AXI4Imp)(masterFn, slaveFn, numPorts)
-
-// Nodes passed from an inner module
-case class AXI4OutputNode() extends OutputNode(AXI4Imp)
-case class AXI4InputNode() extends InputNode(AXI4Imp)
-
-// Nodes used for external ports
-case class AXI4BlindOutputNode(portParams: Seq[AXI4SlavePortParameters]) extends BlindOutputNode(AXI4Imp)(portParams)
-case class AXI4BlindInputNode(portParams: Seq[AXI4MasterPortParameters]) extends BlindInputNode(AXI4Imp)(portParams)
-
-case class AXI4InternalOutputNode(portParams: Seq[AXI4SlavePortParameters]) extends InternalOutputNode(AXI4Imp)(portParams)
-case class AXI4InternalInputNode(portParams: Seq[AXI4MasterPortParameters]) extends InternalInputNode(AXI4Imp)(portParams)
+case class AXI4IdentityNode()(implicit valName: ValName) extends IdentityNode(AXI4Imp)()

src/main/scala/amba/axi4/Parameters.scala

@@ -3,6 +3,7 @@
 package freechips.rocketchip.amba.axi4
 
 import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
 import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 import scala.math.max
@@ -124,7 +125,9 @@ object AXI4BundleParameters
 
 case class AXI4EdgeParameters(
   master: AXI4MasterPortParameters,
-  slave:  AXI4SlavePortParameters)
+  slave:  AXI4SlavePortParameters,
+  params: Parameters,
+  sourceInfo: SourceInfo)
 {
   val bundle = AXI4BundleParameters(master, slave)
 }

src/main/scala/amba/axi4/RegisterRouter.scala

@@ -10,8 +10,8 @@ import freechips.rocketchip.tilelink.{IntSourceNode, IntSourcePortSimple}
 import freechips.rocketchip.util.{HeterogeneousBag, MaskGen}
 import scala.math.{min,max}
 
-class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)
-  extends AXI4SlaveNode(Seq(AXI4SlavePortParameters(
+case class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)(implicit valName: ValName)
+  extends SinkNode(AXI4Imp)(Seq(AXI4SlavePortParameters(
     Seq(AXI4SlaveParameters(
       address       = Seq(address),
       executable    = executable,
@@ -26,11 +26,12 @@ class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
   // Calling this method causes the matching AXI4 bundle to be
   // configured to route all requests to the listed RegFields.
   def regmap(mapping: RegField.Map*) = {
-    val ar = bundleIn(0).ar
-    val aw = bundleIn(0).aw
-    val w  = bundleIn(0).w
-    val r  = bundleIn(0).r
-    val b  = bundleIn(0).b
+    val (io, _) = this.in(0)
+    val ar = io.ar
+    val aw = io.aw
+    val w  = io.w
+    val r  = io.r
+    val b  = io.b
 
     val params = RegMapperParams(log2Up((address.mask+1)/beatBytes), beatBytes, ar.bits.params.idBits + ar.bits.params.userBits)
     val in = Wire(Decoupled(new RegMapperInput(params)))
@@ -77,12 +78,6 @@ class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
   }
 }
 
-object AXI4RegisterNode
-{
-  def apply(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false) =
-    new AXI4RegisterNode(address, concurrency, beatBytes, undefZero, executable)
-}
-
 // These convenience methods below combine to make it possible to create a AXI4
 // register mapped device from a totally abstract register mapped device.
 
@@ -92,13 +87,11 @@ abstract class AXI4RegisterRouterBase(address: AddressSet, interrupts: Int, conc
   val intnode = IntSourceNode(IntSourcePortSimple(num = interrupts))
 }
 
-case class AXI4RegBundleArg(interrupts: HeterogeneousBag[Vec[Bool]], in: HeterogeneousBag[AXI4Bundle])(implicit val p: Parameters)
+case class AXI4RegBundleArg()(implicit val p: Parameters)
 
 class AXI4RegBundleBase(arg: AXI4RegBundleArg) extends Bundle
 {
   implicit val p = arg.p
-  val interrupts = arg.interrupts
-  val in = arg.in
 }
 
 class AXI4RegBundle[P](val params: P, arg: AXI4RegBundleArg) extends AXI4RegBundleBase(arg)
@@ -106,8 +99,8 @@ class AXI4RegBundle[P](val params: P, arg: AXI4RegBundleArg) extends AXI4RegBund
 class AXI4RegModule[P, B <: AXI4RegBundleBase](val params: P, bundleBuilder: => B, router: AXI4RegisterRouterBase)
   extends LazyModuleImp(router) with HasRegMap
 {
-  val io = bundleBuilder
-  val interrupts = if (io.interrupts.isEmpty) Vec(0, Bool()) else io.interrupts(0)
+  val io = IO(bundleBuilder)
+  val interrupts = if (router.intnode.out.isEmpty) Vec(0, Bool()) else router.intnode.out(0)._1
   def regmap(mapping: RegField.Map*) = router.node.regmap(mapping:_*)
 }
 
@@ -120,5 +113,5 @@ class AXI4RegisterRouter[B <: AXI4RegBundleBase, M <: LazyModuleImp]
   require (isPow2(size))
   // require (size >= 4096) ... not absolutely required, but highly recommended
 
-  lazy val module = moduleBuilder(bundleBuilder(AXI4RegBundleArg(intnode.bundleOut, node.bundleIn)), this)
+  lazy val module = moduleBuilder(bundleBuilder(AXI4RegBundleArg()), this)
 }

src/main/scala/amba/axi4/SRAM.scala

@@ -24,15 +24,11 @@ class AXI4RAM(address: AddressSet, executable: Boolean = true, beatBytes: Int =
   require ((address.mask & (beatBytes-1)) == beatBytes-1)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-
     def bigBits(x: BigInt, tail: List[Boolean] = List.empty[Boolean]): List[Boolean] =
       if (x == 0) tail.reverse else bigBits(x >> 1, ((x & 1) == 1) :: tail)
     val mask = bigBits(address.mask >> log2Ceil(beatBytes))
 
-    val in = io.in(0)
+    val (in, _) = node.in(0)
     val mem = SeqMem(1 << mask.filter(b=>b).size, Vec(beatBytes, Bits(width = 8)))
 
     val r_addr = Cat((mask zip (in.ar.bits.addr >> log2Ceil(beatBytes)).toBools).filter(_._1).map(_._2).reverse)

src/main/scala/amba/axi4/Test.scala

@@ -30,7 +30,7 @@ class AXI4LiteFuzzRAM(txns: Int)(implicit p: Parameters) extends LazyModule
   ram.node   := AXI4UserYanker()(AXI4IdIndexer(0)(TLToAXI4(4, true )(TLFragmenter(4, 16)(xbar.node))))
   gpio.node  := AXI4UserYanker()(AXI4IdIndexer(0)(TLToAXI4(4, false)(TLFragmenter(4, 16)(xbar.node))))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }
@@ -53,7 +53,7 @@ class AXI4FullFuzzRAM(txns: Int)(implicit p: Parameters) extends LazyModule
   ram.node   := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(4,false)(xbar.node)))
   gpio.node  := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(4,true )(xbar.node)))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }
@@ -69,7 +69,7 @@ trait HasFuzzTarget {
 
 class AXI4FuzzMaster(txns: Int)(implicit p: Parameters) extends LazyModule with HasFuzzTarget
 {
-  val node  = AXI4OutputNode()
+  val node  = AXI4IdentityNode()
   val fuzz  = LazyModule(new TLFuzzer(txns, overrideAddress = Some(fuzzAddr)))
   val model = LazyModule(new TLRAMModel("AXI4FuzzMaster"))
 
@@ -84,10 +84,9 @@ class AXI4FuzzMaster(txns: Int)(implicit p: Parameters) extends LazyModule with
     model.node))))))
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val out = node.bundleOut
+    val io = IO(new Bundle {
       val finished = Bool(OUTPUT)
-    }
+    })
 
     io.finished := fuzz.module.io.finished
   }
@@ -95,7 +94,7 @@ class AXI4FuzzMaster(txns: Int)(implicit p: Parameters) extends LazyModule with
 
 class AXI4FuzzSlave()(implicit p: Parameters) extends LazyModule with HasFuzzTarget
 {
-  val node = AXI4InputNode()
+  val node = AXI4IdentityNode()
   val xbar = LazyModule(new TLXbar)
   val ram  = LazyModule(new TLRAM(fuzzAddr))
   val error= LazyModule(new TLError(ErrorParams(Seq(AddressSet(0x1800, 0xff)), maxTransfer = 256)))
@@ -114,11 +113,7 @@ class AXI4FuzzSlave()(implicit p: Parameters) extends LazyModule with HasFuzzTar
     AXI4IdIndexer(2)(
     node))))))))
 
-  lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-  }
+  lazy val module = new LazyModuleImp(this) { }
 }
 
 class AXI4FuzzBridge(txns: Int)(implicit p: Parameters) extends LazyModule
@@ -128,7 +123,7 @@ class AXI4FuzzBridge(txns: Int)(implicit p: Parameters) extends LazyModule
 
   slave.node := master.node
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := master.module.io.finished
   }
 }

src/main/scala/amba/axi4/ToTL.scala

@@ -9,7 +9,7 @@ import freechips.rocketchip.diplomacy._
 import freechips.rocketchip.tilelink._
 import freechips.rocketchip.util._
 
-case class AXI4ToTLNode() extends MixedAdapterNode(AXI4Imp, TLImp)(
+case class AXI4ToTLNode()(implicit valName: ValName) extends MixedAdapterNode(AXI4Imp, TLImp)(
   dFn = { case AXI4MasterPortParameters(masters, userBits) =>
     masters.foreach { m => require (m.maxFlight.isDefined, "AXI4 must include a transaction maximum per ID to convert to TL") }
     val maxFlight = masters.map(_.maxFlight.get).max
@@ -44,12 +44,7 @@ class AXI4ToTL()(implicit p: Parameters) extends LazyModule
   val node = AXI4ToTLNode()
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val numIds = edgeIn.master.endId
       val beatBytes = edgeOut.manager.beatBytes
       val beatCountBits = AXI4Parameters.lenBits + (1 << AXI4Parameters.sizeBits) - 1

src/main/scala/amba/axi4/UserYanker.scala

@@ -22,12 +22,7 @@ class AXI4UserYanker(capMaxFlight: Option[Int] = None)(implicit p: Parameters) e
     slaveFn = { sp => sp })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val bits = edgeIn.bundle.userBits
       val need_bypass = edgeOut.slave.minLatency < 1
       require (bits > 0) // useless UserYanker!

src/main/scala/config/Config.scala

@@ -10,8 +10,10 @@ abstract class Field[T] private (val default: Option[T])
 
 abstract class View {
   final def apply[T](pname: Field[T]): T = apply(pname, this)
-  final def apply[T](pname: Field[T], site: View): T = find(pname, site) match {
-    case Some(x) => x.asInstanceOf[T]
+  final def apply[T](pname: Field[T], site: View): T = {
+    val out = find(pname, site)
+    require (out.isDefined, s"Key ${pname} is not defined in Parameters")
+    out.get
   }
 
   final def lift[T](pname: Field[T]): Option[T] = lift(pname, this)

src/main/scala/coreplex/BaseCoreplex.scala

@@ -21,7 +21,7 @@ abstract class BareCoreplex(implicit p: Parameters) extends LazyModule with Bind
   lazy val json = JSON(bindingTree)
 }
 
-abstract class BareCoreplexModule[+L <: BareCoreplex](_outer: L) extends LazyMultiIOModuleImp(_outer) {
+abstract class BareCoreplexModule[+L <: BareCoreplex](_outer: L) extends LazyModuleImp(_outer) {
   val outer = _outer
   ElaborationArtefacts.add("graphml", outer.graphML)
   ElaborationArtefacts.add("dts", outer.dts)

src/main/scala/coreplex/Configs.scala

@@ -187,14 +187,21 @@ class WithRoccExample extends Config((site, here, up) => {
       Seq(
         RoCCParams(
           opcodes = OpcodeSet.custom0,
-          generator = (p: Parameters) => LazyModule(new AccumulatorExample()(p))),
+          generator = (p: Parameters) => {
+            val accumulator = LazyModule(new AccumulatorExample()(p))
+            accumulator}),
         RoCCParams(
           opcodes = OpcodeSet.custom1,
-          generator = (p: Parameters) => LazyModule(new TranslatorExample()(p)),
+          generator = (p: Parameters) => {
+            val translator = LazyModule(new TranslatorExample()(p))
+            translator},
           nPTWPorts = 1),
         RoCCParams(
           opcodes = OpcodeSet.custom2,
-          generator = (p: Parameters) => LazyModule(new CharacterCountExample()(p)))
+          generator = (p: Parameters) => {
+            val counter = LazyModule(new CharacterCountExample()(p))
+            counter
+          })
         ))
     }
 })
@@ -265,7 +272,7 @@ class WithNBitPeripheryBus(nBits: Int) extends Config ((site, here, up) => {
 })
 
 class WithoutTLMonitors extends Config ((site, here, up) => {
-  case TLMonitorBuilder => (args: TLMonitorArgs) => None
+  case MonitorsEnabled => false
 })
 
 class WithNExtTopInterrupts(nExtInts: Int) extends Config((site, here, up) => {

src/main/scala/coreplex/FrontBus.scala

@@ -53,8 +53,7 @@ trait HasFrontBus extends HasSystemBus {
   private val frontbusParams = p(FrontBusKey)
   val frontbusBeatBytes = frontbusParams.beatBytes
 
-  val fbus = new FrontBus(frontbusParams)
-
-  sbus.fromFrontBus := fbus.toSystemBus
+  val fbus = LazyModule(new FrontBus(frontbusParams))
 
+  FlipRendering { implicit p => sbus.fromFrontBus := fbus.toSystemBus }
 }

src/main/scala/coreplex/InterruptBus.scala

@@ -44,7 +44,7 @@ abstract trait HasExtInterrupts extends HasInterruptBus {
   }
 
   val nExtInterrupts = p(NExtTopInterrupts)
-  val extInterrupts = IntInternalInputNode(IntSourcePortSimple(num = nExtInterrupts, resources = device.int))
+  val extInterrupts = IntSourceNode(IntSourcePortSimple(num = nExtInterrupts, resources = device.int))
 }
 
 /** This trait should be used if the External Interrupts have NOT
@@ -77,9 +77,9 @@ trait HasExtInterruptsBundle {
 /** This trait performs the translation from a UInt IO into Diplomatic Interrupts.
   * The wiring must be done in the concrete LazyModuleImp. 
   */
-trait HasExtInterruptsModuleImp extends LazyMultiIOModuleImp with HasExtInterruptsBundle {
+trait HasExtInterruptsModuleImp extends LazyModuleImp with HasExtInterruptsBundle {
   val outer: HasExtInterrupts
   val interrupts = IO(UInt(INPUT, width = outer.nExtInterrupts))
 
-  outer.extInterrupts.bundleIn.flatten.zipWithIndex.foreach { case(o, i) => o := interrupts(i) }
+  outer.extInterrupts.in.map(_._1).flatten.zipWithIndex.foreach { case(o, i) => o := interrupts(i) }
 }

src/main/scala/coreplex/MemoryBus.scala

@@ -68,7 +68,7 @@ trait HasMemoryBus extends HasSystemBus with HasPeripheryBus with HasInterruptBu
 
   private val mask = ~BigInt((nBanks-1) * memBusBlockBytes)
   val memBuses = Seq.tabulate(nMemoryChannels) { channel =>
-    val mbus = new MemoryBus(mbusParams)
+    val mbus = LazyModule(new MemoryBus(mbusParams))
     for (bank <- 0 until nBanksPerChannel) {
       val offset = (bank * nMemoryChannels) + channel
       ForceFanout(a = true) { implicit p => in := sbus.toMemoryBus }

src/main/scala/coreplex/PeripheryBus.scala

@@ -45,7 +45,7 @@ trait HasPeripheryBus extends HasSystemBus {
   private val pbusParams = p(PeripheryBusKey)
   val pbusBeatBytes = pbusParams.beatBytes
 
-  val pbus = new PeripheryBus(pbusParams)
+  val pbus = LazyModule(new PeripheryBus(pbusParams))
 
   // The peripheryBus hangs off of systemBus; here we convert TL-UH -> TL-UL
   pbus.fromSystemBus := sbus.toPeripheryBus()

src/main/scala/coreplex/Ports.scala

@@ -33,7 +33,7 @@ trait HasMasterAXI4MemPort extends HasMemoryBus {
   private val params = p(ExtMem)
   private val device = new MemoryDevice
 
-  val mem_axi4 = AXI4BlindOutputNode(Seq.tabulate(nMemoryChannels) { channel =>
+  val mem_axi4 = AXI4SlaveNode(Seq.tabulate(nMemoryChannels) { channel =>
     val base = AddressSet(params.base, params.size-1)
     val filter = AddressSet(channel * cacheBlockBytes, ~((nMemoryChannels-1) * cacheBlockBytes))
 
@@ -69,14 +69,18 @@ trait HasMasterAXI4MemPortBundle {
   val mem_axi4: HeterogeneousBag[AXI4Bundle]
   val nMemoryChannels: Int
   def connectSimAXIMem(dummy: Int = 1) = {
-    if (nMemoryChannels > 0) Module(LazyModule(new SimAXIMem(nMemoryChannels)).module).io.axi4 <> mem_axi4
+    if (nMemoryChannels > 0)  {
+      val mem = LazyModule(new SimAXIMem(nMemoryChannels))
+      Module(mem.module).io.axi4 <> mem_axi4
+    }
   }
 }
 
 /** Actually generates the corresponding IO in the concrete Module */
-trait HasMasterAXI4MemPortModuleImp extends LazyMultiIOModuleImp with HasMasterAXI4MemPortBundle {
+trait HasMasterAXI4MemPortModuleImp extends LazyModuleImp with HasMasterAXI4MemPortBundle {
   val outer: HasMasterAXI4MemPort
-  val mem_axi4 = IO(outer.mem_axi4.bundleOut)
+  val mem_axi4 = IO(HeterogeneousBag.fromNode(outer.mem_axi4.in))
+  (mem_axi4 zip outer.mem_axi4.in) foreach { case (i, (o, _)) => i <> o }
   val nMemoryChannels = outer.nMemoryChannels
 }
 
@@ -84,7 +88,7 @@ trait HasMasterAXI4MemPortModuleImp extends LazyMultiIOModuleImp with HasMasterA
 trait HasMasterAXI4MMIOPort extends HasSystemBus {
   private val params = p(ExtBus)
   private val device = new SimpleBus("mmio", Nil)
-  val mmio_axi4 = AXI4BlindOutputNode(Seq(AXI4SlavePortParameters(
+  val mmio_axi4 = AXI4SlaveNode(Seq(AXI4SlavePortParameters(
     slaves = Seq(AXI4SlaveParameters(
       address       = List(AddressSet(params.base, params.size-1)),
       resources     = device.ranges,
@@ -107,20 +111,22 @@ trait HasMasterAXI4MMIOPortBundle {
   implicit val p: Parameters
   val mmio_axi4: HeterogeneousBag[AXI4Bundle]
   def connectSimAXIMMIO(dummy: Int = 1) {
-    Module(LazyModule(new SimAXIMem(1, 4096)).module).io.axi4 <> mmio_axi4
+    val mmio_mem = LazyModule(new SimAXIMem(1, 4096))
+    Module(mmio_mem.module).io.axi4 <> mmio_axi4
   }
 }
 
 /** Actually generates the corresponding IO in the concrete Module */
-trait HasMasterAXI4MMIOPortModuleImp extends LazyMultiIOModuleImp with HasMasterAXI4MMIOPortBundle {
+trait HasMasterAXI4MMIOPortModuleImp extends LazyModuleImp with HasMasterAXI4MMIOPortBundle {
   val outer: HasMasterAXI4MMIOPort
-  val mmio_axi4 = IO(outer.mmio_axi4.bundleOut)
+  val mmio_axi4 = IO(HeterogeneousBag.fromNode(outer.mmio_axi4.in))
+  (mmio_axi4 zip outer.mmio_axi4.in) foreach { case (i, (o, _)) => i <> o }
 }
 
 /** Adds an AXI4 port to the system intended to be a slave on an MMIO device bus */
 trait HasSlaveAXI4Port extends HasSystemBus {
   private val params = p(ExtIn)
-  val l2FrontendAXI4Node = AXI4BlindInputNode(Seq(AXI4MasterPortParameters(
+  val l2FrontendAXI4Node = AXI4MasterNode(Seq(AXI4MasterPortParameters(
     masters = Seq(AXI4MasterParameters(
       name = "AXI4 periphery",
       id   = IdRange(0, 1 << params.idBits))))))
@@ -151,16 +157,17 @@ trait HasSlaveAXI4PortBundle {
 }
 
 /** Actually generates the corresponding IO in the concrete Module */
-trait HasSlaveAXI4PortModuleImp extends LazyMultiIOModuleImp with HasSlaveAXI4PortBundle {
+trait HasSlaveAXI4PortModuleImp extends LazyModuleImp with HasSlaveAXI4PortBundle {
   val outer: HasSlaveAXI4Port
-  val l2_frontend_bus_axi4 = IO(outer.l2FrontendAXI4Node.bundleIn)
+  val l2_frontend_bus_axi4 = IO(HeterogeneousBag.fromNode(outer.l2FrontendAXI4Node.out).flip)
+  (outer.l2FrontendAXI4Node.out zip l2_frontend_bus_axi4) foreach { case ((i, _), o) => i <> o }
 }
 
 /** Adds a TileLink port to the system intended to master an MMIO device bus */
 trait HasMasterTLMMIOPort extends HasSystemBus {
   private val params = p(ExtBus)
   private val device = new SimpleBus("mmio", Nil)
-  val mmio_tl = TLBlindOutputNode(Seq(TLManagerPortParameters(
+  val mmio_tl = TLManagerNode(Seq(TLManagerPortParameters(
     managers = Seq(TLManagerParameters(
       address            = List(AddressSet(params.base, params.size-1)),
       resources          = device.ranges,
@@ -192,9 +199,10 @@ trait HasMasterTLMMIOPortBundle {
 }
 
 /** Actually generates the corresponding IO in the concrete Module */
-trait HasMasterTLMMIOPortModuleImp extends LazyMultiIOModuleImp with HasMasterTLMMIOPortBundle {
+trait HasMasterTLMMIOPortModuleImp extends LazyModuleImp with HasMasterTLMMIOPortBundle {
   val outer: HasMasterTLMMIOPort
-  val mmio_tl = IO(outer.mmio_tl.bundleOut)
+  val mmio_tl = IO(HeterogeneousBag.fromNode(outer.mmio_tl.in))
+  (mmio_tl zip outer.mmio_tl.out) foreach { case (i, (o, _)) => i <> o }
 }
 
 /** Adds an TL port to the system intended to be a slave on an MMIO device bus.
@@ -202,7 +210,7 @@ trait HasMasterTLMMIOPortModuleImp extends LazyMultiIOModuleImp with HasMasterTL
   */
 trait HasSlaveTLPort extends HasSystemBus {
   private val params = p(ExtIn)
-  val l2FrontendTLNode = TLBlindInputNode(Seq(TLClientPortParameters(
+  val l2FrontendTLNode = TLClientNode(Seq(TLClientPortParameters(
     clients = Seq(TLClientParameters(
       name     = "Front Port (TL)",
       sourceId = IdRange(0, 1 << params.idBits))))))
@@ -229,9 +237,10 @@ trait HasSlaveTLPortBundle {
 }
 
 /** Actually generates the corresponding IO in the concrete Module */
-trait HasSlaveTLPortModuleImp extends LazyMultiIOModuleImp with HasSlaveTLPortBundle {
+trait HasSlaveTLPortModuleImp extends LazyModuleImp with HasSlaveTLPortBundle {
   val outer: HasSlaveTLPort
-  val l2_frontend_bus_tl = IO(outer.l2FrontendTLNode.bundleIn)
+  val l2_frontend_bus_tl = IO(HeterogeneousBag.fromNode(outer.l2FrontendTLNode.out).flip)
+  (outer.l2FrontendTLNode.in zip l2_frontend_bus_tl) foreach { case ((i, _), o) => i <> o }
 }
 
 /** Memory with AXI port for use in elaboratable test harnesses. */
@@ -241,7 +250,7 @@ class SimAXIMem(channels: Int, forceSize: BigInt = 0)(implicit p: Parameters) ex
   val size = totalSize / channels
   require(totalSize % channels == 0)
 
-  val node = AXI4BlindInputNode(Seq.fill(channels) {
+  val node = AXI4MasterNode(Seq.fill(channels) {
     AXI4MasterPortParameters(Seq(AXI4MasterParameters(
       name = "dut",
       id   = IdRange(0, 1 << config.idBits)
@@ -254,8 +263,9 @@ class SimAXIMem(channels: Int, forceSize: BigInt = 0)(implicit p: Parameters) ex
   }
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val axi4 = node.bundleIn
-    }
+    val io = IO(new Bundle {
+      val axi4 = HeterogeneousBag.fromNode(node.out).flip
+    })
+    (node.out zip io.axi4) foreach { case ((i, _), o) => i <> o }
   }
 }

src/main/scala/coreplex/RTC.scala

@@ -3,10 +3,10 @@
 package freechips.rocketchip.coreplex
 
 import Chisel._
-import freechips.rocketchip.diplomacy.{LazyMultiIOModuleImp, DTSTimebase}
+import freechips.rocketchip.diplomacy.{LazyModuleImp, DTSTimebase}
 import freechips.rocketchip.devices.tilelink.HasPeripheryClint
 
-trait HasRTCModuleImp extends LazyMultiIOModuleImp {
+trait HasRTCModuleImp extends LazyModuleImp {
   val outer: HasPeripheryClint
   private val pbusFreq = outer.p(PeripheryBusKey).frequency
   private val rtcFreq = outer.p(DTSTimebase)

src/main/scala/coreplex/RocketCoreplex.scala

@@ -46,19 +46,19 @@ trait HasRocketTiles extends HasSystemBus
       case SynchronousCrossing(params) => {
         val wrapper = LazyModule(new SyncRocketTile(tp, i)(pWithExtra))
         sbus.fromSyncTiles(params, tp.externalMasterBuffers, tp.name) :=* wrapper.masterNode
-        wrapper.slaveNode :*= pbus.toSyncSlaves(tp.name, tp.externalSlaveBuffers)
+        FlipRendering { implicit p => wrapper.slaveNode :*= pbus.toSyncSlaves(tp.name, tp.externalSlaveBuffers) }
         wrapper
       }
       case AsynchronousCrossing(depth, sync) => {
         val wrapper = LazyModule(new AsyncRocketTile(tp, i)(pWithExtra))
         sbus.fromAsyncTiles(depth, sync, tp.externalMasterBuffers, tp.name) :=* wrapper.masterNode
-        wrapper.slaveNode :*= pbus.toAsyncSlaves(sync, tp.name, tp.externalSlaveBuffers)
+        FlipRendering { implicit p => wrapper.slaveNode :*= pbus.toAsyncSlaves(sync, tp.name, tp.externalSlaveBuffers) }
         wrapper
       }
       case RationalCrossing(direction) => {
         val wrapper = LazyModule(new RationalRocketTile(tp, i)(pWithExtra))
         sbus.fromRationalTiles(direction, tp.externalMasterBuffers, tp.name) :=* wrapper.masterNode
-        wrapper.slaveNode :*= pbus.toRationalSlaves(tp.name, tp.externalSlaveBuffers)
+        FlipRendering { implicit p => wrapper.slaveNode :*= pbus.toRationalSlaves(tp.name, tp.externalSlaveBuffers) }
         wrapper
       }
     }
@@ -86,7 +86,7 @@ trait HasRocketTiles extends HasSystemBus
 
     wrapper.intOutputNode.foreach { case int =>
       val rocketIntXing = LazyModule(new IntXing(wrapper.outputInterruptXingLatency))
-      rocketIntXing.intnode := int
+      FlipRendering { implicit p => rocketIntXing.intnode := int }
       plic.intnode := rocketIntXing.intnode
     }
 
@@ -102,7 +102,7 @@ trait HasRocketTilesBundle {
   val rocket_tile_inputs: Vec[ClockedRocketTileInputs]
 }
 
-trait HasRocketTilesModuleImp extends LazyMultiIOModuleImp
+trait HasRocketTilesModuleImp extends LazyModuleImp
     with HasRocketTilesBundle
     with HasResetVectorWire
     with HasPeripheryDebugModuleImp {

src/main/scala/coreplex/SystemBus.scala

@@ -22,7 +22,7 @@ class SystemBus(params: SystemBusParams)(implicit p: Parameters) extends TLBusWr
   private val master_splitter = LazyModule(new TLSplitter)  // Allows cycle-free connection to external networks
   master_splitter.suggestName(s"${busName}_master_TLSplitter")
   inwardNode :=* master_splitter.node
-  def busView = master_splitter.node.edgesIn.head
+  def busView = master_splitter.node.edges.in.head
 
   protected def inwardSplitNode: TLInwardNode = master_splitter.node
   protected def outwardSplitNode: TLOutwardNode = master_splitter.node
@@ -122,7 +122,7 @@ trait HasSystemBus extends HasInterruptBus {
   private val sbusParams = p(SystemBusKey)
   val sbusBeatBytes = sbusParams.beatBytes
 
-  val sbus = new SystemBus(sbusParams)
+  val sbus = LazyModule(new SystemBus(sbusParams))
 
   def sharedMemoryTLEdge: TLEdge = sbus.busView
   def paddrBits: Int = sbus.busView.bundle.addressBits

src/main/scala/devices/debug/Debug.scala

@@ -288,14 +288,12 @@ class TLDebugModuleOuter(device: Device)(implicit p: Parameters) extends LazyMod
 
   lazy val module = new LazyModuleImp(this) {
 
-    val nComponents = intnode.bundleOut.size
+    val nComponents = intnode.out.size
 
-    val io = new Bundle {
+    val io = IO(new Bundle {
       val ctrl = (new DebugCtrlBundle(nComponents))
-      val tlIn = dmiNode.bundleIn
-      val debugInterrupts = intnode.bundleOut
       val innerCtrl = new DecoupledIO(new DebugInternalBundle())
-    }
+    })
 
     //----DMCONTROL (The whole point of 'Outer' is to maintain this register on dmiClock (e.g. TCK) domain, so that it
     //               can be written even if 'Inner' is not being clocked or is in reset. This allows halting
@@ -355,8 +353,9 @@ class TLDebugModuleOuter(device: Device)(implicit p: Parameters) extends LazyMod
 
     debugIntNxt := debugIntRegs
 
+    val (intnode_out, _) = intnode.out.unzip
     for (component <- 0 until nComponents) {
-      io.debugInterrupts(component)(0) := debugIntRegs(component)
+      intnode_out(component)(0) := debugIntRegs(component)
     }
 
     // Halt request registers are set & cleared by writes to DMCONTROL.haltreq
@@ -393,27 +392,22 @@ class TLDebugModuleOuterAsync(device: Device)(implicit p: Parameters) extends La
   val dmiXbar = LazyModule (new TLXbar())
 
   val dmOuter = LazyModule( new TLDebugModuleOuter(device))
-  val intnode = IntOutputNode()
+  val intnode = dmOuter.intnode
 
-  val dmiInnerNode = TLAsyncOutputNode()
-
-  intnode :*= dmOuter.intnode
+  val dmiInnerNode = TLAsyncCrossingSource()(dmiXbar.node)
 
   dmiXbar.node := dmi2tl.node
   dmOuter.dmiNode := dmiXbar.node
-  dmiInnerNode := TLAsyncCrossingSource()(dmiXbar.node)
   
   lazy val module = new LazyModuleImp(this) {
 
-    val nComponents = intnode.bundleOut.size
+    val nComponents = intnode.out.size
 
-    val io = new Bundle {
+    val io = IO(new Bundle {
       val dmi   = new DMIIO()(p).flip()
-      val dmiInner = dmiInnerNode.bundleOut
       val ctrl = new DebugCtrlBundle(nComponents)
-      val debugInterrupts = intnode.bundleOut
       val innerCtrl = new AsyncBundle(depth=1, new DebugInternalBundle())
-    }
+    })
 
     dmi2tl.module.io.dmi <> io.dmi
 
@@ -447,13 +441,11 @@ class TLDebugModuleInner(device: Device, getNComponents: () => Int)(implicit p:
 
     val nComponents = getNComponents()
 
-    val io = new Bundle {
-      val hart_in = tlNode.bundleIn
-      val dmi_in = dmiNode.bundleIn
+    val io = IO(new Bundle {
       val dmactive = Bool(INPUT)
       val innerCtrl = (new DecoupledIO(new DebugInternalBundle())).flip
       val debugUnavail = Vec(nComponents, Bool()).asInput
-    }
+    })
 
     //--------------------------------------------------------------
     // Import constants for shorter variable names
@@ -1011,26 +1003,23 @@ class TLDebugModuleInner(device: Device, getNComponents: () => Int)(implicit p:
 // Also is the Sink side of hartsel & resumereq fields of DMCONTROL.
 class TLDebugModuleInnerAsync(device: Device, getNComponents: () => Int)(implicit p: Parameters) extends LazyModule{
 
-  val dmInner = LazyModule(new TLDebugModuleInner(device, getNComponents)(p))
-  val dmiNode = TLAsyncInputNode()
-  val tlNode = TLInputNode()
+  val dmInner = LazyModule(new TLDebugModuleInner(device, getNComponents))
+  val dmiXing = LazyModule(new TLAsyncCrossingSink(depth=1))
+  val dmiNode: TLAsyncInwardNode = dmiXing.node
+  val tlNode = dmInner.tlNode
 
-  dmInner.dmiNode := TLAsyncCrossingSink(depth=1)(dmiNode)
-  dmInner.tlNode  := tlNode
+  dmInner.dmiNode := dmiXing.node
 
   lazy val module = new LazyModuleImp(this) {
 
-    val io = new Bundle {
-      // this comes from tlClk domain.
-      val tl_in = tlNode.bundleIn
+    val io = IO(new Bundle {
       // These are all asynchronous and come from Outer
-      val dmi_in = dmiNode.bundleIn
       val dmactive = Bool(INPUT)
       val innerCtrl = new AsyncBundle(1, new DebugInternalBundle()).flip
       // This comes from tlClk domain.
       val debugUnavail    = Vec(getNComponents(), Bool()).asInput
       val psd = new PSDTestMode().asInput
-    }
+    })
 
     dmInner.module.io.innerCtrl := FromAsyncBundle(io.innerCtrl)
     dmInner.module.io.dmactive := ~ResetCatchAndSync(clock, ~io.dmactive, "dmactiveSync", io.psd)
@@ -1049,26 +1038,22 @@ class TLDebugModule(implicit p: Parameters) extends LazyModule {
     override val alwaysExtended = true
   }
 
-  val node = TLInputNode()
-  val intnode = IntOutputNode()
-
   val dmOuter = LazyModule(new TLDebugModuleOuterAsync(device)(p))
-  val dmInner = LazyModule(new TLDebugModuleInnerAsync(device, () => {intnode.bundleOut.size})(p))
+  val dmInner = LazyModule(new TLDebugModuleInnerAsync(device, () => {intnode.edges.out.size})(p))
+
+  val node = dmInner.tlNode
+  val intnode = dmOuter.intnode
 
   dmInner.dmiNode := dmOuter.dmiInnerNode
-  dmInner.tlNode := node
-  intnode :*= dmOuter.intnode
 
   lazy val module = new LazyModuleImp(this) {
-    val nComponents = intnode.bundleOut.size
+    val nComponents = intnode.out.size
 
-    val io = new Bundle {
+    val io = IO(new Bundle {
       val ctrl = new DebugCtrlBundle(nComponents)
       val dmi = new ClockedDMIIO().flip
-      val in = node.bundleIn
-      val debugInterrupts = intnode.bundleOut
       val psd = new PSDTestMode().asInput
-    }
+    })
 
     dmOuter.module.io.dmi <> io.dmi.dmi
     dmOuter.module.reset := io.dmi.dmiReset
@@ -1102,16 +1087,14 @@ class ClockedDMIIO(implicit val p: Parameters) extends ParameterizedBundle()(p){
 
 class DMIToTL(implicit p: Parameters) extends LazyModule {
 
-  val node = TLClientNode(TLClientParameters("debug"))
+  val node = TLClientNode(Seq(TLClientPortParameters(Seq(TLClientParameters("debug")))))
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
+    val io = IO(new Bundle {
       val dmi = new DMIIO()(p).flip()
-      val out = node.bundleOut
-    }
+    })
 
-    val tl = io.out(0)
-    val edge = node.edgesOut(0)
+    val (tl, edge) = node.out(0)
 
     val src  = Wire(init = 0.U)
     val addr = Wire(init = (io.dmi.req.bits.addr << 2))

src/main/scala/devices/debug/Periphery.scala

@@ -54,9 +54,9 @@ trait HasPeripheryDebugBundle {
     }
     debug.psd.foreach { _ <> psd }
   }
-
 }
-trait HasPeripheryDebugModuleImp extends LazyMultiIOModuleImp with HasPeripheryDebugBundle {
+
+trait HasPeripheryDebugModuleImp extends LazyModuleImp with HasPeripheryDebugBundle {
   val outer: HasPeripheryDebug
 
   val debug = IO(new DebugIO)

src/main/scala/devices/tilelink/BootROM.scala

@@ -23,25 +23,22 @@ case object BootROMParams extends Field[BootROMParams]
 class TLROM(val base: BigInt, val size: Int, contentsDelayed: => Seq[Byte], executable: Boolean = true, beatBytes: Int = 4,
   resources: Seq[Resource] = new SimpleDevice("rom", Seq("sifive,rom0")).reg("mem"))(implicit p: Parameters) extends LazyModule
 {
-  val node = TLManagerNode(beatBytes, TLManagerParameters (
+  val node = TLManagerNode(Seq(TLManagerPortParameters(
+    Seq(TLManagerParameters(
       address     = List(AddressSet(base, size-1)),
       resources   = resources,
       regionType  = RegionType.UNCACHED,
       executable  = executable,
       supportsGet = TransferSizes(1, beatBytes),
-    fifoId      = Some(0)))
+      fifoId      = Some(0))),
+    beatBytes = beatBytes)))
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-
     val contents = contentsDelayed
     val wrapSize = 1 << log2Ceil(contents.size)
     require (wrapSize <= size)
 
-    val in = io.in(0)
-    val edge = node.edgesIn(0)
+    val (in, edge) = node.in(0)
 
     val words = (contents ++ Seq.fill(wrapSize-contents.size)(0.toByte)).grouped(beatBytes).toSeq
     val bigs = words.map(_.foldRight(BigInt(0)){ case (x,y) => (x.toInt & 0xff) | y << 8})
@@ -78,7 +75,7 @@ trait HasPeripheryBootROM extends HasPeripheryBus {
 }
 
 /** Coreplex will power-on running at 0x10040 (BootROM) */
-trait HasPeripheryBootROMModuleImp extends LazyMultiIOModuleImp
+trait HasPeripheryBootROMModuleImp extends LazyModuleImp
     with HasResetVectorWire {
   val outer: HasPeripheryBootROM
   global_reset_vector := outer.resetVector.U

src/main/scala/devices/tilelink/BusBlocker.scala

@@ -81,20 +81,13 @@ class BusBlocker(params: BusBlockerParams)(implicit p: Parameters) extends TLBus
     beatBytes = params.controlBeatBytes)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val ctl = controlNode.bundleIn
-      val in = nodeIn.bundleIn
-      val out = nodeOut.bundleOut
-    }
-
     // We need to be able to represent +1 larger than the largest populated address
-    val addressBits = log2Ceil(nodeOut.edgesOut(0).manager.maxAddress+1+1)
+    val addressBits = log2Ceil(nodeOut.edges.out(0).manager.maxAddress+1+1)
     val pmps = RegInit(Vec.fill(params.pmpRegisters) { DevicePMP(addressBits, params.pageBits) })
     val blocks = pmps.tail.map(_.blockPriorAddress) :+ Bool(false)
     controlNode.regmap(0 -> (pmps zip blocks).map { case (p, b) => p.fields(b) }.toList.flatten)
 
-    val in = io.in(0)
-    val edge = nodeIn.edgesIn(0)
+    val (in, edge) = nodeIn.in(0)
 
     // Determine if a request is allowed
     val needW = in.a.bits.opcode =/= TLMessages.Get

src/main/scala/devices/tilelink/BusBypass.scala

@@ -12,14 +12,15 @@ import scala.math.min
 
 abstract class TLBusBypassBase(beatBytes: Int)(implicit p: Parameters) extends LazyModule
 {
-  protected val nodeIn = TLInputNode()
-  protected val nodeOut = TLOutputNode()
+  protected val nodeIn = TLIdentityNode()
+  protected val nodeOut = TLIdentityNode()
   val node = NodeHandle(nodeIn, nodeOut)
 
   protected val bar = LazyModule(new TLBusBypassBar)
   protected val everything = Seq(AddressSet(0, BigInt("ffffffffffffffffffffffffffffffff", 16))) // 128-bit
   protected val error = LazyModule(new TLError(ErrorParams(everything), beatBytes))
 
+  // order matters
   bar.node := nodeIn
   error.node := bar.node
   nodeOut := bar.node
@@ -28,11 +29,9 @@ abstract class TLBusBypassBase(beatBytes: Int)(implicit p: Parameters) extends L
 class TLBusBypass(beatBytes: Int)(implicit p: Parameters) extends TLBusBypassBase(beatBytes)
 {
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = nodeIn.bundleIn
-      val out = nodeOut.bundleOut
+    val io = IO(new Bundle {
       val bypass = Bool(INPUT)
-    }
+    })
     bar.module.io.bypass := io.bypass
   }
 }
@@ -47,17 +46,13 @@ class TLBusBypassBar(implicit p: Parameters) extends LazyModule
     managerFn = { seq => seq(1) })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
+    val io = IO(new Bundle {
       val bypass = Bool(INPUT)
-    }
+    })
 
-    val in   = io.in(0)
-    val out0 = io.out(0)
-    val out1 = io.out(1)
+    val (in, edge) = node.in(0)
+    val Seq((out0,_), (out1,_)) = node.out
 
-    val edge = node.edgesIn(0)
     val bce = edge.manager.anySupportAcquireB && edge.client.anySupportProbe
 
     // We need to be locked to the given bypass direction until all transactions stop

src/main/scala/devices/tilelink/Clint.scala

@@ -53,11 +53,9 @@ class CoreplexLocalInterrupter(params: ClintParams)(implicit p: Parameters) exte
     sinkFn         = { _ => IntSinkPortParameters(Seq(IntSinkParameters())) })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
+    val io = IO(new Bundle {
       val rtcTick = Bool(INPUT)
-      val int = intnode.bundleOut
-      val in = node.bundleIn
-    }
+    })
 
     val time = Seq.fill(timeWidth/regWidth)(Reg(init=UInt(0, width = regWidth)))
     when (io.rtcTick) {
@@ -66,11 +64,12 @@ class CoreplexLocalInterrupter(params: ClintParams)(implicit p: Parameters) exte
         reg := newTime >> i
     }
 
-    val nTiles = intnode.edgesOut.size
+    val nTiles = intnode.out.size
     val timecmp = Seq.fill(nTiles) { Seq.fill(timeWidth/regWidth)(Reg(UInt(width = regWidth))) }
     val ipi = Seq.fill(nTiles) { RegInit(UInt(0, width = 1)) }
 
-    io.int.zipWithIndex.foreach { case (int, i) =>
+    val (intnode_out, _) = intnode.out.unzip
+    intnode_out.zipWithIndex.foreach { case (int, i) =>
       int(0) := ShiftRegister(ipi(i)(0), params.intStages) // msip
       int(1) := ShiftRegister(time.asUInt >= timecmp(i).asUInt, params.intStages) // mtip
     }

src/main/scala/devices/tilelink/Error.scala

@@ -40,15 +40,10 @@ class TLError(params: ErrorParams, beatBytes: Int = 4)(implicit p: Parameters) e
     minLatency = 1))) // no bypass needed for this device
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-
     import TLMessages._
     import TLPermissions._
 
-    val edge = node.edgesIn(0)
-    val in = io.in(0)
+    val (in, edge) = node.in(0)
     val a = Queue(in.a, 1)
     val c = Queue(in.c, 1)
     val da = Wire(in.d)

src/main/scala/devices/tilelink/MaskROM.scala

@@ -24,21 +24,18 @@ trait HasPeripheryMaskROMSlave extends HasPeripheryBus {
 
 class TLMaskROM(c: MaskROMParams)(implicit p: Parameters) extends LazyModule {
   val beatBytes = c.width/8
-  val node = TLManagerNode(beatBytes, TLManagerParameters(
+  val node = TLManagerNode(Seq(TLManagerPortParameters(
+    Seq(TLManagerParameters(
       address            = AddressSet.misaligned(c.address, c.depth*beatBytes),
       resources          = new SimpleDevice("rom", Seq("sifive,maskrom0")).reg("mem"),
       regionType         = RegionType.UNCACHED,
       executable         = true,
       supportsGet        = TransferSizes(1, beatBytes),
-    fifoId             = Some(0))) // requests are handled in order
+      fifoId             = Some(0))), // requests are handled in order
+    beatBytes = beatBytes)))
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-
-    val in = io.in(0)
-    val edge = node.edgesIn(0)
+    val (in, edge)= node.in(0)
 
     val rom = ROMGenerator(ROMConfig(c.name, c.depth, c.width))
     rom.io.clock := clock

src/main/scala/devices/tilelink/Plic.scala

@@ -91,12 +91,12 @@ class TLPLIC(params: PLICParams)(implicit p: Parameters) extends LazyModule
     sinkFn         = { _ => IntSinkPortParameters(Seq(IntSinkParameters())) })
 
   /* Negotiated sizes */
-  def nDevices: Int = intnode.edgesIn.map(_.source.num).sum
+  def nDevices: Int = intnode.edges.in.map(_.source.num).sum
   def nPriorities = min(params.maxPriorities, nDevices)
-  def nHarts = intnode.edgesOut.map(_.source.num).sum
+  def nHarts = intnode.edges.out.map(_.source.num).sum
 
   // Assign all the devices unique ranges
-  lazy val sources = intnode.edgesIn.map(_.source)
+  lazy val sources = intnode.edges.in.map(_.source)
   lazy val flatSources = (sources zip sources.map(_.num).scanLeft(0)(_+_).init).map {
     case (s, o) => s.sources.map(z => z.copy(range = z.range.offset(o)))
   }.flatten
@@ -109,16 +109,13 @@ class TLPLIC(params: PLICParams)(implicit p: Parameters) extends LazyModule
   }
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val tl_in = node.bundleIn
-      val devices = intnode.bundleIn
-      val harts = intnode.bundleOut
-    }
+    val (io_devices, edgesIn) = intnode.in.unzip
+    val (io_harts, _) = intnode.out.unzip
 
     // Compact the interrupt vector the same way
-    val interrupts = (intnode.edgesIn zip io.devices).map { case (e, i) => i.take(e.source.num) }.flatten
+    val interrupts = intnode.in.map { case (i, e) => i.take(e.source.num) }.flatten
     // This flattens the harts into an MSMSMSMSMS... or MMMMM.... sequence
-    val harts = io.harts.flatten
+    val harts = io_harts.flatten
 
     println(s"Interrupt map (${nHarts} harts ${nDevices} interrupts):")
     flatSources.foreach { s =>

src/main/scala/devices/tilelink/TestRAM.scala

@@ -29,16 +29,11 @@ class TLTestRAM(address: AddressSet, executable: Boolean = true, beatBytes: Int
   require ((address.mask & (beatBytes-1)) == beatBytes-1)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-
     def bigBits(x: BigInt, tail: List[Boolean] = List.empty[Boolean]): List[Boolean] =
       if (x == 0) tail.reverse else bigBits(x >> 1, ((x & 1) == 1) :: tail)
     val mask = bigBits(address.mask >> log2Ceil(beatBytes))
 
-    val in = io.in(0)
-    val edge = node.edgesIn(0)
+    val (in, edge) = node.in(0)
 
     val addrBits = (mask zip edge.addr_hi(in.a.bits).toBools).filter(_._1).map(_._2)
     val memAddress = Cat(addrBits.reverse)
@@ -75,7 +70,7 @@ class TLRAMZeroDelay(ramBeatBytes: Int, txns: Int)(implicit p: Parameters) exten
   model.node := fuzz.node
   ram.node := TLDelayer(0.25)(model.node)
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }

src/main/scala/devices/tilelink/Zero.scala

@@ -24,12 +24,7 @@ class TLZero(address: AddressSet, resources: Seq[Resource], executable: Boolean
     minLatency = 1))) // no bypass needed for this device
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-
-    val in = io.in(0)
-    val edge = node.edgesIn(0)
+    val (in, edge) = node.in(0)
 
     val a = Queue(in.a, 2)
     val hasData = edge.hasData(a.bits)

src/main/scala/diplomacy/LazyModule.scala

@@ -6,6 +6,8 @@ import Chisel._
 import chisel3.experimental.{BaseModule, RawModule, MultiIOModule, withClockAndReset}
 import chisel3.internal.sourceinfo.{SourceInfo, SourceLine, UnlocatableSourceInfo}
 import freechips.rocketchip.config.Parameters
+import scala.collection.immutable.ListMap
+import scala.util.matching._
 
 abstract class LazyModule()(implicit val p: Parameters)
 {
@@ -13,9 +15,9 @@ abstract class LazyModule()(implicit val p: Parameters)
   protected[diplomacy] var children = List[LazyModule]()
   protected[diplomacy] var nodes = List[BaseNode]()
   protected[diplomacy] var info: SourceInfo = UnlocatableSourceInfo
-  protected[diplomacy] val parent = LazyModule.stack.headOption
+  protected[diplomacy] val parent = LazyModule.scope
 
-  LazyModule.stack = this :: LazyModule.stack
+  LazyModule.scope = Some(this)
   parent.foreach(p => p.children = this :: p.children)
 
   private var suggestedName: Option[String] = None
@@ -54,15 +56,6 @@ abstract class LazyModule()(implicit val p: Parameters)
 
   def module: LazyModuleImpLike
 
-  protected[diplomacy] def instantiate() = {
-    children.reverse.foreach { c =>
-      // !!! fix chisel3 so we can pass the desired sourceInfo
-      // implicit val sourceInfo = c.module.outer.info
-      Module(c.module)
-    }
-    bindings.reverse.foreach { f => f () }
-  }
-
   def omitGraphML: Boolean = !nodes.exists(!_.omitGraphML) && !children.exists(!_.omitGraphML)
   lazy val graphML: String = parent.map(_.graphML).getOrElse {
     val buf = new StringBuilder
@@ -95,10 +88,11 @@ abstract class LazyModule()(implicit val p: Parameters)
     buf ++= s"""${pad}</node>\n"""
   }
   private def edgesGraphML(buf: StringBuilder, pad: String) {
-    nodes.filter(!_.omitGraphML) foreach { n => n.outputs.filter(!_._1.omitGraphML).foreach { case (o, l) =>
+    nodes.filter(!_.omitGraphML) foreach { n => n.outputs.filter(!_._1.omitGraphML).foreach { case (o, edge) =>
+      val RenderedEdge(colour, label, flipped) = edge
       buf ++= pad
       buf ++= "<edge"
-      if (o.reverse) {
+      if (flipped) {
         buf ++= s""" target=\"${index}::${n.index}\""""
         buf ++= s""" source=\"${o.lazyModule.index}::${o.index}\">"""
       } else {
@@ -106,13 +100,13 @@ abstract class LazyModule()(implicit val p: Parameters)
         buf ++= s""" target=\"${o.lazyModule.index}::${o.index}\">"""
       }
       buf ++= s"""<data key=\"e\"><y:PolyLineEdge>"""
-      if (o.reverse) {
+      if (flipped) {
         buf ++= s"""<y:Arrows source=\"standard\" target=\"none\"/>"""
       } else {
         buf ++= s"""<y:Arrows source=\"none\" target=\"standard\"/>"""
       }
-      buf ++= s"""<y:LineStyle color=\"${o.colour}\" type=\"line\" width=\"1.0\"/>"""
-      buf ++= s"""<y:EdgeLabel modelName=\"centered\" rotationAngle=\"270.0\">${l}</y:EdgeLabel>"""
+      buf ++= s"""<y:LineStyle color=\"${colour}\" type=\"line\" width=\"1.0\"/>"""
+      buf ++= s"""<y:EdgeLabel modelName=\"centered\" rotationAngle=\"270.0\">${label}</y:EdgeLabel>"""
       buf ++= s"""</y:PolyLineEdge></data></edge>\n"""
     } }
     children.filter(!_.omitGraphML).foreach { c => c.edgesGraphML(buf, pad) }
@@ -126,16 +120,16 @@ abstract class LazyModule()(implicit val p: Parameters)
 
 object LazyModule
 {
-  protected[diplomacy] var stack = List[LazyModule]()
+  protected[diplomacy] var scope: Option[LazyModule] = None
   private var index = 0
 
   def apply[T <: LazyModule](bc: T)(implicit sourceInfo: SourceInfo): T = {
     // Make sure the user put LazyModule around modules in the correct order
     // If this require fails, probably some grandchild was missing a LazyModule
     // ... or you applied LazyModule twice
-    require (!stack.isEmpty, s"LazyModule() applied to ${bc.name} twice ${sourceLine(sourceInfo)}")
-    require (stack.head eq bc, s"LazyModule() applied to ${bc.name} before ${stack.head.name} ${sourceLine(sourceInfo)}")
-    stack = stack.tail
+    require (scope.isDefined, s"LazyModule() applied to ${bc.name} twice ${sourceLine(sourceInfo)}")
+    require (scope.get eq bc, s"LazyModule() applied to ${bc.name} before ${scope.get.name} ${sourceLine(sourceInfo)}")
+    scope = bc.parent
     bc.info = sourceInfo
     bc
   }
@@ -144,26 +138,96 @@ object LazyModule
 sealed trait LazyModuleImpLike extends BaseModule
 {
   val wrapper: LazyModule
+  val auto: AutoBundle
+  protected[diplomacy] val dangles: Seq[Dangle]
 
   // .module had better not be accessed while LazyModules are still being built!
-  require (LazyModule.stack.isEmpty, s"${wrapper.name}.module was constructed before LazyModule() was run on ${LazyModule.stack.head.name}")
+  require (!LazyModule.scope.isDefined, s"${wrapper.name}.module was constructed before LazyModule() was run on ${LazyModule.scope.get.name}")
 
   override def desiredName = wrapper.moduleName
   suggestName(wrapper.instanceName)
 
   implicit val p = wrapper.p
+
+  protected[diplomacy] def instantiate() = {
+    val childDangles = wrapper.children.reverse.flatMap { c =>
+      implicit val sourceInfo = c.info
+      Module(c.module).dangles
+    }
+    val nodeDangles = wrapper.nodes.reverse.flatMap(_.instantiate())
+    val allDangles = nodeDangles ++ childDangles
+    val done = Set() ++ allDangles.groupBy(_.source).values.filter(_.size == 2).map { case Seq(a, b) =>
+      require (a.flipped != b.flipped)
+      if (a.flipped) { a.data <> b.data } else { b.data <> a.data }
+      a.source
+    }
+    val forward = allDangles.filter(d => !done(d.source))
+    val auto = IO(new AutoBundle(forward.map { d => (d.name, d.data, d.flipped) }:_*))
+    val dangles = (forward zip auto.elements) map { case (d, (_, io)) =>
+      if (d.flipped) { d.data <> io } else { io <> d.data }
+      d.copy(data = io, name = wrapper.valName.getOrElse("anon") + "_" + d.name)
+    }
+    wrapper.bindings.reverse.foreach { f => f () }
+    (auto, dangles)
+  }
 }
 
-abstract class LazyModuleImp(val wrapper: LazyModule) extends Module with LazyModuleImpLike {
-  wrapper.instantiate()
+class LazyModuleImp(val wrapper: LazyModule) extends MultiIOModule with LazyModuleImpLike {
+  val (auto, dangles) = instantiate()
 }
 
-abstract class LazyMultiIOModuleImp(val wrapper: LazyModule) extends MultiIOModule with LazyModuleImpLike {
-  wrapper.instantiate()
+class LazyRawModuleImp(val wrapper: LazyModule) extends RawModule with LazyModuleImpLike {
+  val (auto, dangles) = withClockAndReset(Bool(false).asClock, Bool(true)) {
+    instantiate()
+  }
 }
 
-abstract class LazyRawModuleImp(val wrapper: LazyModule) extends RawModule with LazyModuleImpLike {
-  withClockAndReset(Bool(false).asClock, Bool(true)) {
-    wrapper.instantiate()
+class SimpleLazyModule(implicit p: Parameters) extends LazyModule
+{
+  lazy val module = new LazyModuleImp(this)
+}
+
+trait LazyScope
+{
+  this: LazyModule =>
+  def apply[T](body: => T)(implicit p: Parameters) = {
+    val saved = LazyModule.scope
+    LazyModule.scope = Some(this)
+    val out = body
+    require (LazyModule.scope.isDefined, s"LazyScope ${name} tried to exit, but scope was empty!")
+    require (LazyModule.scope.get eq this, s"LazyScope ${name} exited before LazyModule ${LazyModule.scope.get.name} was closed")
+    LazyModule.scope = saved
+    out
   }
 }
+
+object LazyScope
+{
+  def apply[T](name: String)(body: => T)(implicit p: Parameters) = {
+    val scope = LazyModule(new SimpleLazyModule with LazyScope)
+    scope.suggestName(name)
+    scope { body }
+  }
+}
+
+case class HalfEdge(serial: Int, index: Int)
+case class Dangle(source: HalfEdge, sink: HalfEdge, flipped: Boolean, name: String, data: Data)
+
+final class AutoBundle(elts: (String, Data, Boolean)*) extends Record {
+  // We need to preserve the order of elts, despite grouping by name to disambiguate things
+  val elements = ListMap() ++ elts.zipWithIndex.map(makeElements).groupBy(_._1).values.flatMap {
+    case Seq((key, element, i)) => Seq(i -> (key -> element))
+    case seq => seq.zipWithIndex.map { case ((key, element, i), j) => i -> (key + "_" + j -> element) }
+  }.toList.sortBy(_._1).map(_._2)
+  require (elements.size == elts.size)
+
+  private def makeElements(tuple: ((String, Data, Boolean), Int)) = {
+    val ((key, data, flip), i) = tuple
+    // trim trailing _0_1_2 stuff so that when we append _# we don't create collisions
+    val regex = new Regex("(_[0-9]+)*$")
+    val element = if (flip) data.cloneType.flip else data.cloneType
+    (regex.replaceAllIn(key, ""), element, i)
+  }
+
+  override def cloneType = (new AutoBundle(elts:_*)).asInstanceOf[this.type]
+}

src/main/scala/diplomacy/Monitor.scala

@@ -1,10 +0,0 @@
-// See LICENSE.SiFive for license details.
-
-package freechips.rocketchip.diplomacy
-
-import chisel3.internal.sourceinfo.{SourceInfo, SourceLine}
-import freechips.rocketchip.config.Parameters
-
-abstract class MonitorBase(implicit val sourceInfo: SourceInfo, p: Parameters) extends LazyModule()(p) {
-  override val module: LazyModuleImp
-}

src/main/scala/diplomacy/Nodes.scala

@@ -7,6 +7,7 @@ import chisel3.internal.sourceinfo.SourceInfo
 import freechips.rocketchip.config.{Parameters,Field}
 import freechips.rocketchip.util.HeterogeneousBag
 import scala.collection.mutable.ListBuffer
+import scala.util.matching._
 
 object CardinalityInferenceDirection {
   val cases = Seq(SOURCE_TO_SINK, SINK_TO_SOURCE, NO_INFERENCE)
@@ -26,7 +27,13 @@ object CardinalityInferenceDirection {
 private case object CardinalityInferenceDirectionKey extends
   Field[CardinalityInferenceDirection.T](CardinalityInferenceDirection.NO_INFERENCE)
 
-private case object MonitorsEnabled extends Field[Boolean](true)
+case object MonitorsEnabled extends Field[Boolean](true)
+case object RenderFlipped extends Field[Boolean](false)
+
+case class RenderedEdge(
+  colour:  String,
+  label:   String  = "",
+  flipped: Boolean = false) // prefer to draw the arrow pointing the opposite direction of other edges
 
 // DI = Downwards flowing Parameters received on the inner side of the node
 // UI = Upwards   flowing Parameters generated by the inner side of the node
@@ -34,19 +41,16 @@ private case object MonitorsEnabled extends Field[Boolean](true)
 // BI = Bundle type used when connecting to the inner side of the node
 trait InwardNodeImp[DI, UI, EI, BI <: Data]
 {
-  def edgeI(pd: DI, pu: UI): EI
+  def edgeI(pd: DI, pu: UI, p: Parameters, sourceInfo: SourceInfo): EI
   def bundleI(ei: EI): BI
-  def colour: String
-  def reverse: Boolean = false
-  def connect(edges: () => Seq[EI], bundles: () => Seq[(BI, BI)], enableMonitoring: Boolean)
-              (implicit p: Parameters, sourceInfo: SourceInfo): (Option[MonitorBase], () => Unit) = {
-    (None, () => bundles().foreach { case (i, o) => i <> o })
-  }
+
+  // Edge functions
+  def monitor(bundle: BI, edge: EI) {}
+  def render(e: EI): RenderedEdge
 
   // optional methods to track node graph
   def mixI(pu: UI, node: InwardNode[DI, UI, BI]): UI = pu // insert node into parameters
   def getO(pu: UI): Option[BaseNode] = None // most-outward common node
-  def labelI(ei: EI) = ""
 }
 
 // DO = Downwards flowing Parameters generated by the outer side of the node
@@ -55,57 +59,80 @@ trait InwardNodeImp[DI, UI, EI, BI <: Data]
 // BO = Bundle type used when connecting to the outer side of the node
 trait OutwardNodeImp[DO, UO, EO, BO <: Data]
 {
-  def edgeO(pd: DO, pu: UO): EO
+  def edgeO(pd: DO, pu: UO, p: Parameters, sourceInfo: SourceInfo): EO
   def bundleO(eo: EO): BO
 
   // optional methods to track node graph
   def mixO(pd: DO, node: OutwardNode[DO, UO, BO]): DO = pd // insert node into parameters
   def getI(pd: DO): Option[BaseNode] = None // most-inward common node
-  def labelO(eo: EO) = ""
 }
 
 abstract class NodeImp[D, U, EO, EI, B <: Data]
   extends Object with InwardNodeImp[D, U, EI, B] with OutwardNodeImp[D, U, EO, B]
 
-abstract class BaseNode
+// If your edges have the same direction, using this saves you some typing
+abstract class SimpleNodeImp[D, U, E, B <: Data]
+  extends NodeImp[D, U, E, E, B]
 {
-  require (!LazyModule.stack.isEmpty, "You cannot create a node outside a LazyModule!")
+  def edge(pd: D, pu: U, p: Parameters, sourceInfo: SourceInfo): E
+  def edgeO(pd: D, pu: U, p: Parameters, sourceInfo: SourceInfo) = edge(pd, pu, p, sourceInfo)
+  def edgeI(pd: D, pu: U, p: Parameters, sourceInfo: SourceInfo) = edge(pd, pu, p, sourceInfo)
+  def bundle(e: E): B
+  def bundleO(e: E) = bundle(e)
+  def bundleI(e: E) = bundle(e)
+}
 
-  val lazyModule = LazyModule.stack.head
+abstract class BaseNode(implicit val valName: ValName)
+{
+  require (LazyModule.scope.isDefined, "You cannot create a node outside a LazyModule!")
+
+  val lazyModule = LazyModule.scope.get
   val index = lazyModule.nodes.size
   lazyModule.nodes = this :: lazyModule.nodes
 
-  val externalIn: Boolean
-  val externalOut: Boolean
+  val serial = BaseNode.serial
+  BaseNode.serial = BaseNode.serial + 1
+  protected[diplomacy] def instantiate(): Seq[Dangle]
 
-  def nodename = getClass.getName.split('.').last
-  def name = lazyModule.name + "." + nodename
+  def name = lazyModule.name + "." + valName.name
   def omitGraphML = outputs.isEmpty && inputs.isEmpty
   lazy val nodedebugstring: String = ""
 
-  protected[diplomacy] def gci: Option[BaseNode] // greatest common inner
-  protected[diplomacy] def gco: Option[BaseNode] // greatest common outer
-  protected[diplomacy] def outputs: Seq[(BaseNode, String)]
-  protected[diplomacy] def inputs:  Seq[(BaseNode, String)]
-  protected[diplomacy] def colour:  String
-  protected[diplomacy] def reverse: Boolean
+  def wirePrefix = {
+    val camelCase = "([a-z])([A-Z])".r
+    val decamel = camelCase.replaceAllIn(valName.name, _ match { case camelCase(l, h) => l + "_" + h })
+    val trimNode = "_?node$".r
+    val name = trimNode.replaceFirstIn(decamel.toLowerCase, "")
+    if (name.isEmpty) "" else name + "_"
   }
 
+  protected[diplomacy] def gci: Option[BaseNode] // greatest common inner
+  protected[diplomacy] def gco: Option[BaseNode] // greatest common outer
+  protected[diplomacy] def inputs:  Seq[(BaseNode, RenderedEdge)]
+  protected[diplomacy] def outputs: Seq[(BaseNode, RenderedEdge)]
+}
+
+object BaseNode
+{
+  protected[diplomacy] var serial = 0
+}
+
+// !!! rename the nodes we bind?
 case class NodeHandle[DI, UI, BI <: Data, DO, UO, BO <: Data]
-  (inward: InwardNode[DI, UI, BI], outward: OutwardNode[DO, UO, BO])
+  (inwardHandle: InwardNodeHandle[DI, UI, BI], outwardHandle: OutwardNodeHandle[DO, UO, BO])
   extends Object with InwardNodeHandle[DI, UI, BI] with OutwardNodeHandle[DO, UO, BO]
+{
+  val inward = inwardHandle.inward
+  val outward = outwardHandle.outward
+}
 
 trait InwardNodeHandle[DI, UI, BI <: Data]
 {
   protected[diplomacy] val inward: InwardNode[DI, UI, BI]
-  def := (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo): Option[MonitorBase] =
-    inward.:=(h)(p, sourceInfo)
-  def :*= (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo): Option[MonitorBase] =
-    inward.:*=(h)(p, sourceInfo)
-  def :=* (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo): Option[MonitorBase] =
-    inward.:=*(h)(p, sourceInfo)
-  def :=? (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo): Option[MonitorBase] =
-    inward.:=?(h)(p, sourceInfo)
+  def := (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo) { inward.:=(h)(p, sourceInfo) }
+  def :*= (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo) { inward.:*=(h)(p, sourceInfo) }
+  def :=* (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo) { inward.:=*(h)(p, sourceInfo) }
+  def :=? (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo) { inward.:=?(h)(p, sourceInfo) }
 }
 
 sealed trait NodeBinding
@@ -121,7 +148,7 @@ trait InwardNode[DI, UI, BI <: Data] extends BaseNode with InwardNodeHandle[DI,
   require (!numPI.isEmpty, s"No number of inputs would be acceptable to ${name}${lazyModule.line}")
   require (numPI.start >= 0, s"${name} accepts a negative number of inputs${lazyModule.line}")
 
-  private val accPI = ListBuffer[(Int, OutwardNode[DI, UI, BI], NodeBinding, Parameters)]()
+  private val accPI = ListBuffer[(Int, OutwardNode[DI, UI, BI], NodeBinding, Parameters, SourceInfo)]()
   private var iRealized = false
 
   protected[diplomacy] def iPushed = accPI.size
@@ -129,8 +156,8 @@ trait InwardNode[DI, UI, BI <: Data] extends BaseNode with InwardNodeHandle[DI,
     val info = sourceLine(sourceInfo, " at ", "")
     val noIs = numPI.size == 1 && numPI.contains(0)
     require (!noIs, s"${name}${lazyModule.line} was incorrectly connected as a sink" + info)
-    require (!iRealized, s"${name}${lazyModule.line} was incorrectly connected as a sink after it's .module was used" + info)
-    accPI += ((index, node, binding, p))
+    require (!iRealized, s"${name}${lazyModule.line} was incorrectly connected as a sink after its .module was used" + info)
+    accPI += ((index, node, binding, p, sourceInfo))
   }
 
   protected[diplomacy] lazy val iBindings = { iRealized = true; accPI.result() }
@@ -138,7 +165,6 @@ trait InwardNode[DI, UI, BI <: Data] extends BaseNode with InwardNodeHandle[DI,
   protected[diplomacy] val iStar: Int
   protected[diplomacy] val iPortMapping: Seq[(Int, Int)]
   protected[diplomacy] val iParams: Seq[UI]
-  protected[diplomacy] val bundleIn: HeterogeneousBag[BI]
 }
 
 trait OutwardNodeHandle[DO, UO, BO <: Data]
@@ -154,7 +180,7 @@ trait OutwardNode[DO, UO, BO <: Data] extends BaseNode with OutwardNodeHandle[DO
   require (!numPO.isEmpty, s"No number of outputs would be acceptable to ${name}${lazyModule.line}")
   require (numPO.start >= 0, s"${name} accepts a negative number of outputs${lazyModule.line}")
 
-  private val accPO = ListBuffer[(Int, InwardNode [DO, UO, BO], NodeBinding, Parameters)]()
+  private val accPO = ListBuffer[(Int, InwardNode [DO, UO, BO], NodeBinding, Parameters, SourceInfo)]()
   private var oRealized = false
 
   protected[diplomacy] def oPushed = accPO.size
@@ -162,8 +188,8 @@ trait OutwardNode[DO, UO, BO <: Data] extends BaseNode with OutwardNodeHandle[DO
     val info = sourceLine(sourceInfo, " at ", "")
     val noOs = numPO.size == 1 && numPO.contains(0)
     require (!noOs, s"${name}${lazyModule.line} was incorrectly connected as a source" + info)
-    require (!oRealized, s"${name}${lazyModule.line} was incorrectly connected as a source after it's .module was used" + info)
-    accPO += ((index, node, binding, p))
+    require (!oRealized, s"${name}${lazyModule.line} was incorrectly connected as a source after its .module was used" + info)
+    accPO += ((index, node, binding, p, sourceInfo))
   }
 
   protected[diplomacy] lazy val oBindings = { oRealized = true; accPO.result() }
@@ -171,37 +197,46 @@ trait OutwardNode[DO, UO, BO <: Data] extends BaseNode with OutwardNodeHandle[DO
   protected[diplomacy] val oStar: Int
   protected[diplomacy] val oPortMapping: Seq[(Int, Int)]
   protected[diplomacy] val oParams: Seq[DO]
-  protected[diplomacy] val bundleOut: HeterogeneousBag[BO]
 }
 
-abstract class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
+abstract class CycleException(kind: String, loop: Seq[String]) extends Exception(s"Diplomatic ${kind} cycle detected involving ${loop}")
+case class StarCycleException(loop: Seq[String] = Nil) extends CycleException("star", loop)
+case class DownwardCycleException(loop: Seq[String] = Nil) extends CycleException("downward", loop)
+case class UpwardCycleException(loop: Seq[String] = Nil) extends CycleException("upward", loop)
+
+case class Edges[EI, EO](in: EI, out: EO)
+sealed abstract class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   inner: InwardNodeImp [DI, UI, EI, BI],
   outer: OutwardNodeImp[DO, UO, EO, BO])(
   protected[diplomacy] val numPO: Range.Inclusive,
-  protected[diplomacy] val numPI: Range.Inclusive)
+  protected[diplomacy] val numPI: Range.Inclusive)(
+  implicit valName: ValName)
   extends BaseNode with InwardNode[DI, UI, BI] with OutwardNode[DO, UO, BO]
 {
   protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStar: Int, oStar: Int): (Int, Int)
   protected[diplomacy] def mapParamsD(n: Int, p: Seq[DI]): Seq[DO]
   protected[diplomacy] def mapParamsU(n: Int, p: Seq[UO]): Seq[UI]
 
+  private var starCycleGuard = false
   protected[diplomacy] lazy val (oPortMapping, iPortMapping, oStar, iStar) = {
-    val oStars = oBindings.filter { case (_,_,b,_) => b == BIND_STAR }.size
-    val iStars = iBindings.filter { case (_,_,b,_) => b == BIND_STAR }.size
-    val oKnown = oBindings.map { case (_, n, b, _) => b match {
+    try {
+      if (starCycleGuard) throw StarCycleException()
+      val oStars = oBindings.filter { case (_,_,b,_,_) => b == BIND_STAR }.size
+      val iStars = iBindings.filter { case (_,_,b,_,_) => b == BIND_STAR }.size
+      val oKnown = oBindings.map { case (_, n, b, _, _) => b match {
         case BIND_ONCE  => 1
         case BIND_QUERY => n.iStar
         case BIND_STAR  => 0 }}.foldLeft(0)(_+_)
-    val iKnown = iBindings.map { case (_, n, b, _) => b match {
+      val iKnown = iBindings.map { case (_, n, b, _, _) => b match {
         case BIND_ONCE  => 1
         case BIND_QUERY => n.oStar
         case BIND_STAR  => 0 }}.foldLeft(0)(_+_)
       val (iStar, oStar) = resolveStar(iKnown, oKnown, iStars, oStars)
-    val oSum = oBindings.map { case (_, n, b, _) => b match {
+      val oSum = oBindings.map { case (_, n, b, _, _) => b match {
         case BIND_ONCE  => 1
         case BIND_QUERY => n.iStar
         case BIND_STAR  => oStar }}.scanLeft(0)(_+_)
-    val iSum = iBindings.map { case (_, n, b, _) => b match {
+      val iSum = iBindings.map { case (_, n, b, _, _) => b match {
         case BIND_ONCE  => 1
         case BIND_QUERY => n.oStar
         case BIND_STAR  => iStar }}.scanLeft(0)(_+_)
@@ -210,56 +245,102 @@ abstract class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
       require(numPO.contains(oTotal), s"${name} has ${oTotal} outputs, expected ${numPO}${lazyModule.line}")
       require(numPI.contains(iTotal), s"${name} has ${iTotal} inputs, expected ${numPI}${lazyModule.line}")
       (oSum.init zip oSum.tail, iSum.init zip iSum.tail, oStar, iStar)
+    } catch {
+      case c: StarCycleException => throw c.copy(loop = s"${name}${lazyModule.line}" +: c.loop)
+    }
   }
 
-  lazy val oPorts = oBindings.flatMap { case (i, n, _, _) =>
+  lazy val oPorts = oBindings.flatMap { case (i, n, _, p, s) =>
     val (start, end) = n.iPortMapping(i)
-    (start until end) map { j => (j, n) }
+    (start until end) map { j => (j, n, p, s) }
   }
-  lazy val iPorts = iBindings.flatMap { case (i, n, _, _) =>
+  lazy val iPorts = iBindings.flatMap { case (i, n, _, p, s) =>
     val (start, end) = n.oPortMapping(i)
-    (start until end) map { j => (j, n) }
+    (start until end) map { j => (j, n, p, s) }
   }
 
+  private var oParamsCycleGuard = false
   protected[diplomacy] lazy val oParams: Seq[DO] = {
-    val o = mapParamsD(oPorts.size, iPorts.map { case (i, n) => n.oParams(i) })
+    try {
+      if (oParamsCycleGuard) throw DownwardCycleException()
+      oParamsCycleGuard = true
+      val o = mapParamsD(oPorts.size, iPorts.map { case (i, n, _, _) => n.oParams(i) })
       require (o.size == oPorts.size, s"Bug in diplomacy; ${name} has ${o.size} != ${oPorts.size} down/up outer parameters${lazyModule.line}")
       o.map(outer.mixO(_, this))
+    } catch {
+      case c: DownwardCycleException => throw c.copy(loop = s"${name}${lazyModule.line}" +: c.loop)
     }
+  }
+
+  private var iParamsCycleGuard = false
   protected[diplomacy] lazy val iParams: Seq[UI] = {
-    val i = mapParamsU(iPorts.size, oPorts.map { case (o, n) => n.iParams(o) })
+    try {
+      if (iParamsCycleGuard) throw UpwardCycleException()
+      iParamsCycleGuard = true
+      val i = mapParamsU(iPorts.size, oPorts.map { case (o, n, _, _) => n.iParams(o) })
       require (i.size == iPorts.size, s"Bug in diplomacy; ${name} has ${i.size} != ${iPorts.size} up/down inner parameters${lazyModule.line}")
       i.map(inner.mixI(_, this))
+    } catch {
+      case c: UpwardCycleException => throw c.copy(loop = s"${name}${lazyModule.line}" +: c.loop)
+    }
   }
 
   protected[diplomacy] def gco = if (iParams.size != 1) None else inner.getO(iParams(0))
   protected[diplomacy] def gci = if (oParams.size != 1) None else outer.getI(oParams(0))
 
-  lazy val edgesOut = (oPorts zip oParams).map { case ((i, n), o) => outer.edgeO(o, n.iParams(i)) }
-  lazy val edgesIn  = (iPorts zip iParams).map { case ((o, n), i) => inner.edgeI(n.oParams(o), i) }
-  lazy val externalEdgesOut = if (externalOut) {edgesOut} else { Seq() }
-  lazy val externalEdgesIn = if (externalIn) {edgesIn} else { Seq() }
+  protected[diplomacy] lazy val edgesOut = (oPorts zip oParams).map { case ((i, n, p, s), o) => outer.edgeO(o, n.iParams(i), p, s) }
+  protected[diplomacy] lazy val edgesIn  = (iPorts zip iParams).map { case ((o, n, p, s), i) => inner.edgeI(n.oParams(o), i, p, s) }
 
-  lazy val paramsOut: Seq[Parameters] = (oPortMapping zip oBindings).flatMap { case ((s, e), b) => Seq.fill(e-s) { b._4 } }
-  lazy val paramsIn:  Seq[Parameters] = (iPortMapping zip iBindings).flatMap { case ((s, e), b) => Seq.fill(e-s) { b._4 } }
+  // If you need access to the edges of a foreign Node, use this method (in/out create bundles)
+  lazy val edges = Edges(edgesIn, edgesOut)
 
-  val flip = false // needed for blind nodes
-  private def flipO(b: HeterogeneousBag[BO]) = if (flip) b.flip else b
-  private def flipI(b: HeterogeneousBag[BI]) = if (flip) b      else b.flip
-  val wire = false // needed if you want to grab access to from inside a module
-  private def wireO(b: HeterogeneousBag[BO]) = if (wire) Wire(b) else b
-  private def wireI(b: HeterogeneousBag[BI]) = if (wire) Wire(b) else b
+  protected[diplomacy] lazy val bundleOut: Seq[BO] = edgesOut.map(e => Wire(outer.bundleO(e)))
+  protected[diplomacy] lazy val bundleIn:  Seq[BI] = edgesIn .map(e => Wire(inner.bundleI(e)))
 
-  lazy val bundleOut = wireO(flipO(HeterogeneousBag(edgesOut.map(outer.bundleO(_)))))
-  lazy val bundleIn  = wireI(flipI(HeterogeneousBag(edgesIn .map(inner.bundleI(_)))))
+  protected[diplomacy] def danglesOut: Seq[Dangle] = oPorts.zipWithIndex.map { case ((j, n, _, _), i) =>
+    Dangle(
+      source = HalfEdge(serial, i),
+      sink   = HalfEdge(n.serial, j),
+      flipped= false,
+      name   = wirePrefix + "out",
+      data   = bundleOut(i))
+  }
+  protected[diplomacy] def danglesIn: Seq[Dangle] = iPorts.zipWithIndex.map { case ((j, n, _, _), i) =>
+    Dangle(
+      source = HalfEdge(n.serial, j),
+      sink   = HalfEdge(serial, i),
+      flipped= true,
+      name   = wirePrefix + "in",
+      data   = bundleIn(i))
+  }
+
+  private var bundlesSafeNow = false
+  // Accessors to the result of negotiation to be used in LazyModuleImp:
+  def out: Seq[(BO, EO)] = {
+    require(bundlesSafeNow, s"${name}.out should only be called from the context of its module implementation")
+    bundleOut zip edgesOut
+  }
+  def in: Seq[(BI, EI)] = {
+    require(bundlesSafeNow, s"${name}.in should only be called from the context of its module implementation")
+    bundleIn zip edgesIn
+  }
+
+  // Used by LazyModules.module.instantiate
+  protected val identity = false
+  protected[diplomacy] def instantiate() = {
+    bundlesSafeNow = true
+    if (!identity) {
+      (iPorts zip in) foreach {
+        case ((_, _, p, _), (b, e)) => if (p(MonitorsEnabled)) inner.monitor(b, e)
+    } }
+    danglesOut ++ danglesIn
+  }
 
   // connects the outward part of a node with the inward part of this node
-  private def bind(h: OutwardNodeHandle[DI, UI, BI], binding: NodeBinding)
-                  (implicit p: Parameters, sourceInfo: SourceInfo): Option[MonitorBase] = {
+  private def bind(h: OutwardNodeHandle[DI, UI, BI], binding: NodeBinding)(implicit p: Parameters, sourceInfo: SourceInfo) {
     val x = this // x := y
     val y = h.outward
     val info = sourceLine(sourceInfo, " at ", "")
-    require (!LazyModule.stack.isEmpty, s"${y.name} cannot be connected to ${x.name} outside of LazyModule scope" + info)
     val i = x.iPushed
     val o = y.oPushed
     y.oPush(i, x, binding match {
@@ -267,29 +348,12 @@ abstract class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
       case BIND_STAR  => BIND_QUERY
       case BIND_QUERY => BIND_STAR })
     x.iPush(o, y, binding)
-    def edges() = {
-      val (iStart, iEnd) = x.iPortMapping(i)
-      val (oStart, oEnd) = y.oPortMapping(o)
-      require (iEnd - iStart == oEnd - oStart, s"Bug in diplomacy; ${iEnd-iStart} != ${oEnd-oStart} means port resolution failed")
-      Seq.tabulate(iEnd - iStart) { j => x.edgesIn(iStart+j) }
-    }
-    def bundles() = {
-      val (iStart, iEnd) = x.iPortMapping(i)
-      val (oStart, oEnd) = y.oPortMapping(o)
-      require (iEnd - iStart == oEnd - oStart, s"Bug in diplomacy; ${iEnd-iStart} != ${oEnd-oStart} means port resolution failed")
-      Seq.tabulate(iEnd - iStart) { j =>
-        (x.bundleIn(iStart+j), y.bundleOut(oStart+j))
-      }
-    }
-    val (out, newbinding) = inner.connect(edges _, bundles _, p(MonitorsEnabled))
-    LazyModule.stack.head.bindings = newbinding :: LazyModule.stack.head.bindings
-    out
   }
 
-  override def :=  (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo): Option[MonitorBase] = bind(h, BIND_ONCE)
-  override def :*= (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo): Option[MonitorBase] = bind(h, BIND_STAR)
-  override def :=* (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo): Option[MonitorBase] = bind(h, BIND_QUERY)
-  override def :=? (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo): Option[MonitorBase] = {
+  override def :=  (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo) = bind(h, BIND_ONCE)
+  override def :*= (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo) = bind(h, BIND_STAR)
+  override def :=* (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo) = bind(h, BIND_QUERY)
+  override def :=? (h: OutwardNodeHandle[DI, UI, BI])(implicit p: Parameters, sourceInfo: SourceInfo) = {
     p(CardinalityInferenceDirectionKey) match {
       case CardinalityInferenceDirection.SOURCE_TO_SINK => this :=* h
       case CardinalityInferenceDirection.SINK_TO_SOURCE => this :*= h
@@ -298,17 +362,19 @@ abstract class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   }
 
   // meta-data for printing the node graph
-  protected[diplomacy] def colour  = inner.colour
-  protected[diplomacy] def reverse = inner.reverse
-  protected[diplomacy] def outputs = oPorts.map(_._2) zip edgesOut.map(e => outer.labelO(e))
-  protected[diplomacy] def inputs  = iPorts.map(_._2) zip edgesIn .map(e => inner.labelI(e))
+  protected[diplomacy] def inputs = (iPorts zip edgesIn) map { case ((_, n, p, _), e) =>
+    val re = inner.render(e)
+    (n, re.copy(flipped = re.flipped != p(RenderFlipped)))
+  }
+  protected[diplomacy] def outputs = oPorts map { case (i, n, _, _) => (n, n.inputs(i)._2) }
 }
 
 abstract class MixedCustomNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   inner: InwardNodeImp [DI, UI, EI, BI],
   outer: OutwardNodeImp[DO, UO, EO, BO])(
   numPO: Range.Inclusive,
-  numPI: Range.Inclusive)
+  numPI: Range.Inclusive)(
+  implicit valName: ValName)
   extends MixedNode(inner, outer)(numPO, numPI)
 {
   def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int)
@@ -318,7 +384,8 @@ abstract class MixedCustomNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
 
 abstract class CustomNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(
   numPO: Range.Inclusive,
-  numPI: Range.Inclusive)
+  numPI: Range.Inclusive)(
+  implicit valName: ValName)
   extends MixedCustomNode(imp, imp)(numPO, numPI)
 
 class MixedAdapterNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
@@ -326,12 +393,10 @@ class MixedAdapterNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   outer: OutwardNodeImp[DO, UO, EO, BO])(
   dFn: DI => DO,
   uFn: UO => UI,
-  num: Range.Inclusive = 0 to 999)
+  num: Range.Inclusive = 0 to 999)(
+  implicit valName: ValName)
   extends MixedNode(inner, outer)(num, num)
 {
-  val externalIn: Boolean = true
-  val externalOut: Boolean = true
-
   protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
     require (oStars + iStars <= 1, s"${name} (an adapter) appears left of a :*= ${iStars} times and right of a :=* ${oStars} times; at most once is allowed${lazyModule.line}")
     if (oStars > 0) {
@@ -352,21 +417,38 @@ class MixedAdapterNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   }
 }
 
+class AdapterNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(
+  dFn: D => D,
+  uFn: U => U,
+  num: Range.Inclusive = 0 to 999)(
+  implicit valName: ValName)
+    extends MixedAdapterNode[D, U, EI, B, D, U, EO, B](imp, imp)(dFn, uFn, num)
+
+// IdentityNodes automatically connect their inputs to outputs
+class IdentityNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])()(implicit valName: ValName)
+  extends AdapterNode(imp)({ s => s }, { s => s })
+{
+  protected override val identity = true
+  override protected[diplomacy] def instantiate() = {
+    val dangles = super.instantiate()
+    (out zip in) map { case ((o, _), (i, _)) => o <> i }
+    dangles
+  } 
+}
+
 class MixedNexusNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   inner: InwardNodeImp [DI, UI, EI, BI],
   outer: OutwardNodeImp[DO, UO, EO, BO])(
   dFn: Seq[DI] => DO,
   uFn: Seq[UO] => UI,
   numPO: Range.Inclusive = 1 to 999,
-  numPI: Range.Inclusive = 1 to 999)
+  numPI: Range.Inclusive = 1 to 999)(
+  implicit valName: ValName)
   extends MixedNode(inner, outer)(numPO, numPI)
 {
 //  require (numPO.end >= 1, s"${name} does not accept outputs${lazyModule.line}")
 //  require (numPI.end >= 1, s"${name} does not accept inputs${lazyModule.line}")
 
-  val externalIn: Boolean = true
-  val externalOut: Boolean = true
-
   protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
     require (iStars == 0, s"${name} (a nexus) appears left of :*= (perhaps you should flip the '*' to :=*?)${lazyModule.line}")
     require (oStars == 0, s"${name} (a nexus) appears right of a :=* (perhaps you should flip the '*' to :*=?)${lazyModule.line}")
@@ -376,82 +458,18 @@ class MixedNexusNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   protected[diplomacy] def mapParamsU(n: Int, p: Seq[UO]): Seq[UI] = { val a = uFn(p); Seq.fill(n)(a) }
 }
 
-class AdapterNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(
-  dFn: D => D,
-  uFn: U => U,
-  num: Range.Inclusive = 0 to 999)
-    extends MixedAdapterNode[D, U, EI, B, D, U, EO, B](imp, imp)(dFn, uFn, num)
-
 class NexusNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(
   dFn: Seq[D] => D,
   uFn: Seq[U] => U,
   numPO: Range.Inclusive = 1 to 999,
-  numPI: Range.Inclusive = 1 to 999)
+  numPI: Range.Inclusive = 1 to 999)(
+  implicit valName: ValName)
     extends MixedNexusNode[D, U, EI, B, D, U, EO, B](imp, imp)(dFn, uFn, numPO, numPI)
 
-case class SplitterArg[T](newSize: Int, ports: Seq[T])
-class MixedSplitterNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
-  inner: InwardNodeImp [DI, UI, EI, BI],
-  outer: OutwardNodeImp[DO, UO, EO, BO])(
-  dFn: SplitterArg[DI] => Seq[DO],
-  uFn: SplitterArg[UO] => Seq[UI],
-  numPO: Range.Inclusive = 1 to 999,
-  numPI: Range.Inclusive = 1 to 999)
-  extends MixedNode(inner, outer)(numPO, numPI)
-{
-  override val externalIn: Boolean = true
-  override val externalOut: Boolean = true
-
-  protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
-    require (oKnown == 0, s"${name} (a splitter) appears right of a := or :*=; use a :=* instead${lazyModule.line}")
-    require (iStars == 0, s"${name} (a splitter) cannot appear left of a :*=; did you mean :=*?${lazyModule.line}")
-    (0, iKnown)
-  }
-  protected[diplomacy] def mapParamsD(n: Int, p: Seq[DI]): Seq[DO] = {
-    require (p.size == 0 || n % p.size == 0, s"Diplomacy bug; splitter inputs do not divide outputs")
-    val out = dFn(SplitterArg(n, p))
-    require (out.size == n, s"${name} created the wrong number of outputs from inputs${lazyModule.line}")
-    out
-  }
-  protected[diplomacy] def mapParamsU(n: Int, p: Seq[UO]): Seq[UI] = {
-    require (n == 0 || p.size % n == 0, s"Diplomacy bug; splitter outputs indivisable by inputs")
-    val out = uFn(SplitterArg(n, p))
-    require (out.size == n, s"${name} created the wrong number of inputs from outputs${lazyModule.line}")
-    out
-  }
-}
-
-class SplitterNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(
-  dFn: SplitterArg[D] => Seq[D],
-  uFn: SplitterArg[U] => Seq[U],
-  numPO: Range.Inclusive = 1 to 999,
-  numPI: Range.Inclusive = 1 to 999)
-    extends MixedSplitterNode[D, U, EI, B, D, U, EO, B](imp, imp)(dFn, uFn, numPO, numPI)
-
-class IdentityNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])
-  extends AdapterNode(imp)({s => s}, {s => s})
-
-class OutputNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B]) extends IdentityNode(imp)
-{
-  override val externalIn: Boolean = false
-  override val externalOut: Boolean = true
-  override lazy val bundleIn = bundleOut
-}
-
-class InputNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B]) extends IdentityNode(imp)
-{
-  override val externalIn: Boolean = true
-  override val externalOut: Boolean = false
-
-  override lazy val bundleOut = bundleIn
-}
-
-class SourceNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(po: Seq[D])
+// There are no Mixed SourceNodes
+class SourceNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(po: Seq[D])(implicit valName: ValName)
   extends MixedNode(imp, imp)(po.size to po.size, 0 to 0)
 {
-  override val externalIn: Boolean = false
-  override val externalOut: Boolean = true
-
   protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
     require (oStars <= 1, s"${name} (a source) appears right of a :=* ${oStars} times; at most once is allowed${lazyModule.line}")
     require (iStars == 0, s"${name} (a source) cannot appear left of a :*=${lazyModule.line}")
@@ -461,16 +479,12 @@ class SourceNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(po: Seq
   }
   protected[diplomacy] def mapParamsD(n: Int, p: Seq[D]): Seq[D] = po
   protected[diplomacy] def mapParamsU(n: Int, p: Seq[U]): Seq[U] = Seq()
-
-  override lazy val bundleIn = { require(false, s"${name} has no bundleIn; try bundleOut?"); bundleOut }
 }
 
-class SinkNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(pi: Seq[U])
+// There are no Mixed SinkNodes
+class SinkNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(pi: Seq[U])(implicit valName: ValName)
   extends MixedNode(imp, imp)(0 to 0, pi.size to pi.size)
 {
-  override val externalIn: Boolean = true
-  override val externalOut: Boolean = false
-
   protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
     require (iStars <= 1, s"${name} (a sink) appears left of a :*= ${iStars} times; at most once is allowed${lazyModule.line}")
     require (oStars == 0, s"${name} (a sink) cannot appear right of a :=*${lazyModule.line}")
@@ -480,40 +494,4 @@ class SinkNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(pi: Seq[U
   }
   protected[diplomacy] def mapParamsD(n: Int, p: Seq[D]): Seq[D] = Seq()
   protected[diplomacy] def mapParamsU(n: Int, p: Seq[U]): Seq[U] = pi
-
-  override lazy val bundleOut = { require(false, s"${name} has no bundleOut; try bundleIn?"); bundleIn }
-}
-
-class BlindOutputNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(pi: Seq[U])
-  extends SinkNode(imp)(pi)
-{
-  override val externalIn: Boolean = false
-  override val flip = true
-  override lazy val bundleOut = bundleIn
-}
-
-class BlindInputNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(po: Seq[D])
-  extends SourceNode(imp)(po)
-{
-  override val externalOut: Boolean = false
-  override val flip = true
-  override lazy val bundleIn = bundleOut
-}
-
-class InternalOutputNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(pi: Seq[U])
-  extends SinkNode(imp)(pi)
-{
-  override val externalIn: Boolean = false
-  override val externalOut: Boolean = false
-  override val wire = true
-  override lazy val bundleOut = bundleIn
-}
-
-class InternalInputNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(po: Seq[D])
-  extends SourceNode(imp)(po)
-{
-  override val externalIn: Boolean = false
-  override val externalOut: Boolean = false
-  override val wire = true
-  override lazy val bundleIn = bundleOut
 }

src/main/scala/diplomacy/Resources.scala

@@ -179,7 +179,7 @@ trait BindingScope
 
   private case class ExpandedValue(path: Seq[String], labels: Seq[String], value: Seq[ResourceValue])
   private lazy val eval: Unit = {
-    require (LazyModule.stack.isEmpty, "May not evaluate binding while still constructing LazyModules")
+    require (!LazyModule.scope.isDefined, "May not evaluate binding while still constructing LazyModules")
     parentScope.foreach { _.eval }
     resourceBindings = parentScope.map(_.resourceBindings).getOrElse(Nil)
     BindingScope.active = Some(this)
@@ -224,7 +224,7 @@ trait BindingScope
 object BindingScope
 {
   protected[diplomacy] var active: Option[BindingScope] = None
-  protected[diplomacy] def find(m: Option[LazyModule] = LazyModule.stack.headOption): Option[BindingScope] = m.flatMap {
+  protected[diplomacy] def find(m: Option[LazyModule] = LazyModule.scope): Option[BindingScope] = m.flatMap {
     case s: BindingScope => Some(s)
     case x => find(x.parent)
   }

src/main/scala/diplomacy/ValName.scala

@@ -0,0 +1,13 @@
+// See LICENSE.SiFive for license details.
+
+package freechips.rocketchip.diplomacy
+
+import scala.language.experimental.macros
+import freechips.rocketchip.macros.ValNameImpl
+
+case class ValName(name: String)
+
+object ValName
+{
+  implicit def materialize(implicit x: ValNameImpl): ValName = ValName(x.name)
+}

src/main/scala/diplomacy/package.scala

@@ -47,4 +47,7 @@ package object diplomacy
   def DisableMonitors[T](body: Parameters => T)(implicit p: Parameters) = body(p.alterPartial {
     case MonitorsEnabled => false
   })
+  def FlipRendering[T](body: Parameters => T)(implicit p: Parameters) = body(p.alterPartial {
+    case RenderFlipped => !p(RenderFlipped)
+  })
 }

src/main/scala/groundtest/Tile.scala

@@ -5,6 +5,7 @@ package freechips.rocketchip.groundtest
 
 import Chisel._
 import freechips.rocketchip.config._
+import freechips.rocketchip.diplomacy._
 import freechips.rocketchip.coreplex._
 import freechips.rocketchip.rocket.{HellaCache, RocketCoreParams}
 import freechips.rocketchip.tile._
@@ -29,7 +30,7 @@ case object GroundTestTilesKey extends Field[Seq[GroundTestTileParams]]
 
 abstract class GroundTestTile(params: GroundTestTileParams)(implicit p: Parameters) extends BaseTile(params)(p) {
   val slave = None
-  val dcacheOpt = params.dcache.map { dc => HellaCache(0, dc.nMSHRs == 0) }
+  val dcacheOpt = params.dcache.map { dc => LazyModule(HellaCache(0, dc.nMSHRs == 0)) }
   dcacheOpt.foreach { tileBus.node := _.node }
 
   override lazy val module = new GroundTestTileModule(this, () => new GroundTestTileBundle(this))

src/main/scala/rocket/BusErrorUnit.scala

@@ -37,11 +37,9 @@ class BusErrorUnit[T <: BusErrors](t: => T, params: BusErrorUnitParams)(implicit
     beatBytes = p(XLen)/8)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val tl = node.bundleIn
-      val interrupt = intNode.bundleOut
+    val io = IO(new Bundle {
       val errors = t.flip
-    }
+    })
 
     val sources = io.errors.toErrorList
     val mask = sources.map(_.nonEmpty.B).asUInt
@@ -61,7 +59,8 @@ class BusErrorUnit[T <: BusErrors](t: => T, params: BusErrorUnitParams)(implicit
       }
     }
 
-    io.interrupt.head(0) := (accrued & interrupt).orR
+    val (int_out, _) = intNode.out(0)
+    int_out(0) := (accrued & interrupt).orR
 
     def reg(r: UInt) = RegField(regWidth, r)
     def maskedReg(r: UInt, m: UInt) = RegField(regWidth, r, RegWriteFn((v, d) => { when (v) { r := d & m }; true }))

src/main/scala/rocket/Frontend.scala

@@ -56,28 +56,22 @@ class FrontendIO(implicit p: Parameters) extends CoreBundle()(p) {
 class Frontend(val icacheParams: ICacheParams, hartid: Int)(implicit p: Parameters) extends LazyModule {
   lazy val module = new FrontendModule(this)
   val icache = LazyModule(new ICache(icacheParams, hartid))
-  val masterNode = TLOutputNode()
-  val slaveNode = TLInputNode()
-
-  masterNode := icache.masterNode
-  // Avoid breaking tile dedup due to address constants in the monitor
-  DisableMonitors { implicit p => icache.slaveNode.map { _ := slaveNode } }
+  val masterNode = icache.masterNode
+  val slaveNode = icache.slaveNode
 }
 
 class FrontendBundle(outer: Frontend) extends CoreBundle()(outer.p)
     with HasExternallyDrivenTileConstants {
   val cpu = new FrontendIO().flip
   val ptw = new TLBPTWIO()
-  val tl_out = outer.masterNode.bundleOut
-  val tl_in = outer.slaveNode.bundleIn
   val errors = new ICacheErrors
 }
 
 class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
     with HasCoreParameters
     with HasL1ICacheParameters {
-  val io = new FrontendBundle(outer)
-  implicit val edge = outer.masterNode.edgesOut.head
+  val io = IO(new FrontendBundle(outer))
+  implicit val edge = outer.masterNode.edges.out(0)
   val icache = outer.icache.module
   require(fetchWidth*coreInstBytes == outer.icacheParams.fetchBytes)
 

src/main/scala/rocket/HellaCache.scala

@@ -176,15 +176,14 @@ class HellaCacheBundle(outer: HellaCache)(implicit p: Parameters) extends CoreBu
   val hartid = UInt(INPUT, hartIdLen)
   val cpu = (new HellaCacheIO).flip
   val ptw = new TLBPTWIO()
-  val mem = outer.node.bundleOut
   val errors = new DCacheErrors
 }
 
 class HellaCacheModule(outer: HellaCache) extends LazyModuleImp(outer)
     with HasL1HellaCacheParameters {
-  implicit val edge = outer.node.edgesOut(0)
-  val io = new HellaCacheBundle(outer)
-  val tl_out = io.mem(0)
+  implicit val edge = outer.node.edges.out(0)
+  val (tl_out, _) = outer.node.out(0)
+  val io = IO(new HellaCacheBundle(outer))
 
   private val fifoManagers = edge.manager.managers.filter(TLFIFOFixer.allUncacheable)
   fifoManagers.foreach { m =>
@@ -195,8 +194,7 @@ class HellaCacheModule(outer: HellaCache) extends LazyModuleImp(outer)
 
 object HellaCache {
   def apply(hartid: Int, blocking: Boolean, scratch: () => Option[AddressSet] = () => None)(implicit p: Parameters) = {
-    if (blocking) LazyModule(new DCache(hartid, scratch))
-    else LazyModule(new NonBlockingDCache(hartid))
+    if (blocking) new DCache(hartid, scratch) else new NonBlockingDCache(hartid)
   }
 }
 
@@ -208,7 +206,7 @@ trait HasHellaCache extends HasTileLinkMasterPort with HasTileParameters {
   def findScratchpadFromICache: Option[AddressSet]
   val hartid: Int
   var nDCachePorts = 0
-  val dcache = HellaCache(hartid, tileParams.dcache.get.nMSHRs == 0, findScratchpadFromICache _)
+  val dcache = LazyModule(HellaCache(hartid, tileParams.dcache.get.nMSHRs == 0, findScratchpadFromICache _))
   tileBus.node := dcache.node
 }
 

src/main/scala/rocket/ICache.scala

@@ -45,15 +45,15 @@ class ICacheErrors(implicit p: Parameters) extends CoreBundle()(p)
 
 class ICache(val icacheParams: ICacheParams, val hartid: Int)(implicit p: Parameters) extends LazyModule {
   lazy val module = new ICacheModule(this)
-  val masterNode = TLClientNode(TLClientParameters(
+  val masterNode = TLClientNode(Seq(TLClientPortParameters(Seq(TLClientParameters(
     sourceId = IdRange(0, 1 + icacheParams.prefetch.toInt), // 0=refill, 1=hint
-    name = s"Core ${hartid} ICache"))
+    name = s"Core ${hartid} ICache")))))
 
   val size = icacheParams.nSets * icacheParams.nWays * icacheParams.blockBytes
   val device = new SimpleDevice("itim", Seq("sifive,itim0"))
-  val slaveNode = icacheParams.itimAddr.map { itimAddr =>
-    val wordBytes = icacheParams.fetchBytes
-    TLManagerNode(Seq(TLManagerPortParameters(
+  private val wordBytes = icacheParams.fetchBytes
+  val slaveNode =
+    TLManagerNode(icacheParams.itimAddr.toSeq.map { itimAddr => TLManagerPortParameters(
       Seq(TLManagerParameters(
         address         = Seq(AddressSet(itimAddr, size-1)),
         resources       = device.reg("mem"),
@@ -64,8 +64,7 @@ class ICache(val icacheParams: ICacheParams, val hartid: Int)(implicit p: Parame
         supportsGet     = TransferSizes(1, wordBytes),
         fifoId          = Some(0))), // requests handled in FIFO order
       beatBytes = wordBytes,
-      minLatency = 1)))
-  }
+      minLatency = 1)})
 }
 
 class ICacheResp(outer: ICache) extends Bundle {
@@ -91,8 +90,6 @@ class ICacheBundle(outer: ICache) extends CoreBundle()(outer.p) {
 
   val resp = Valid(new ICacheResp(outer))
   val invalidate = Bool(INPUT)
-  val tl_out = outer.masterNode.bundleOut
-  val tl_in = outer.slaveNode.map(_.bundleIn)
 
   val errors = new ICacheErrors
   val perf = new ICachePerfEvents().asOutput
@@ -109,11 +106,10 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
     with HasL1ICacheParameters {
   override val cacheParams = outer.icacheParams // Use the local parameters
 
-  val io = new ICacheBundle(outer)
-  val edge_out = outer.masterNode.edgesOut.head
-  val tl_out = io.tl_out.head
-  val edge_in = outer.slaveNode.map(_.edgesIn.head)
-  val tl_in = io.tl_in.map(_.head)
+  val io = IO(new ICacheBundle(outer))
+  val (tl_out, edge_out) = outer.masterNode.out(0)
+  // Option.unzip does not exist :-(
+  val (tl_in, edge_in) = outer.slaveNode.in.headOption.unzip
 
   val tECC = cacheParams.tagECC
   val dECC = cacheParams.dataECC

src/main/scala/rocket/PTW.scala

@@ -283,7 +283,7 @@ trait CanHavePTW extends HasHellaCache {
 trait CanHavePTWModule extends HasHellaCacheModule {
   val outer: CanHavePTW
   val ptwPorts = ListBuffer(outer.dcache.module.io.ptw)
-  val ptw = Module(new PTW(outer.nPTWPorts)(outer.dcache.node.edgesOut(0), outer.p))
+  val ptw = Module(new PTW(outer.nPTWPorts)(outer.dcache.node.edges.out(0), outer.p))
   if (outer.usingPTW)
     dcachePorts += ptw.io.mem
 }

src/main/scala/rocket/ScratchpadSlavePort.scala

@@ -30,13 +30,11 @@ class ScratchpadSlavePort(address: AddressSet, coreDataBytes: Int, usingAtomics:
     minLatency = 1)))
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val tl_in = node.bundleIn
+    val io = IO(new Bundle {
       val dmem = new HellaCacheIO
-    }
+    })
 
-    val tl_in = io.tl_in(0)
-    val edge = node.edgesIn(0)
+    val (tl_in, edge) = node.in(0)
 
     val s_ready :: s_wait :: s_replay :: s_grant :: Nil = Enum(UInt(), 4)
     val state = Reg(init = s_ready)
@@ -104,7 +102,7 @@ trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend {
     LazyModule(new ScratchpadSlavePort(AddressSet(s, d.dataScratchpadBytes-1), xBytes, tileParams.core.useAtomics)))
   }
 
-  val intOutputNode = tileParams.core.tileControlAddr.map(dummy => IntOutputNode())
+  val intOutputNode = tileParams.core.tileControlAddr.map(dummy => IntIdentityNode())
   val busErrorUnit = tileParams.core.tileControlAddr map { a =>
     val beu = LazyModule(new BusErrorUnit(new L1BusErrors, BusErrorUnitParams(a)))
     intOutputNode.get := beu.intNode
@@ -112,7 +110,7 @@ trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend {
   }
 
   // connect any combination of ITIM, DTIM, and BusErrorUnit
-  val slaveNode = TLInputNode()
+  val slaveNode = TLIdentityNode()
   DisableMonitors { implicit p =>
     val xbarPorts =
       scratch.map(lm => (lm.node, xBytes)) ++
@@ -129,7 +127,7 @@ trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend {
   }
 
   def findScratchpadFromICache: Option[AddressSet] = scratch.map { s =>
-    val finalNode = frontend.masterNode.edgesOut.head.manager.managers.find(_.nodePath.last == s.node)
+    val finalNode = frontend.masterNode.edges.out.head.manager.managers.find(_.nodePath.last == s.node)
     require (finalNode.isDefined, "Could not find the scratch pad; not reachable via icache?")
     require (finalNode.get.address.size == 1, "Scratchpad address space was fragmented!")
     finalNode.get.address(0)
@@ -140,8 +138,6 @@ trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend {
 
 trait CanHaveScratchpadBundle extends HasHellaCacheBundle with HasICacheFrontendBundle {
   val outer: CanHaveScratchpad
-  val slave = outer.slaveNode.bundleIn
-  val intOutput = outer.intOutputNode.map(_.bundleOut)
 }
 
 trait CanHaveScratchpadModule extends HasHellaCacheModule with HasICacheFrontendModule {

src/main/scala/tile/BaseTile.scala

@@ -66,21 +66,20 @@ abstract class BareTileBundle[+L <: BareTile](_outer: L) extends GenericParamete
 
 abstract class BareTileModule[+L <: BareTile, +B <: BareTileBundle[L]](_outer: L, _io: () => B) extends LazyModuleImp(_outer) {
   val outer = _outer
-  val io = _io ()
+  val io = IO(_io ())
 }
 
 /** Uses TileLink master port to connect caches and accelerators to the coreplex */
 trait HasTileLinkMasterPort {
   implicit val p: Parameters
   val module: HasTileLinkMasterPortModule
-  val masterNode = TLOutputNode()
+  val masterNode = TLIdentityNode()
   val tileBus = LazyModule(new TLXbar) // TileBus xbar for cache backends to connect to
   masterNode := tileBus.node
 }
 
 trait HasTileLinkMasterPortBundle {
   val outer: HasTileLinkMasterPort
-  val master = outer.masterNode.bundleOut
 }
 
 trait HasTileLinkMasterPortModule {

src/main/scala/tile/Interrupts.scala

@@ -38,7 +38,6 @@ trait HasExternalInterrupts extends HasTileParameters {
 
 trait HasExternalInterruptsBundle {
   val outer: HasExternalInterrupts
-  val interrupts = outer.intNode.bundleIn
 }
 
 trait HasExternalInterruptsModule {
@@ -57,6 +56,7 @@ trait HasExternalInterruptsModule {
 
     val core_ips = core.lip
 
-    (async_ips ++ periph_ips ++ seip ++ core_ips).zip(io.interrupts(0)).foreach { case(c, i) => c := i }
+    val (interrupts, _) = outer.intNode.in(0)
+    (async_ips ++ periph_ips ++ seip ++ core_ips).zip(interrupts).foreach { case(c, i) => c := i }
   }
 }

src/main/scala/tile/LazyRoCC.scala

@@ -60,13 +60,11 @@ class RoCCCoreIO(implicit p: Parameters) extends CoreBundle()(p) {
 abstract class LazyRoCC(implicit p: Parameters) extends LazyModule {
   val module: LazyRoCCModule
 
-  val atlNode: TLMixedNode = TLOutputNode()
-  val tlNode: TLMixedNode = TLOutputNode()
+  val atlNode: TLMixedNode = TLIdentityNode()
+  val tlNode: TLMixedNode = TLIdentityNode()
 }
 
 class RoCCIO(outer: LazyRoCC)(implicit p: Parameters) extends RoCCCoreIO()(p) {
-  val atl = outer.atlNode.bundleOut
-  val tl = outer.tlNode.bundleOut
   // Should be handled differently, eventually
   val ptw = Vec(p(RoccNPTWPorts), new TLBPTWIO)
   val fpu_req = Decoupled(new FPInput)
@@ -74,7 +72,7 @@ class RoCCIO(outer: LazyRoCC)(implicit p: Parameters) extends RoCCCoreIO()(p) {
 }
 
 class LazyRoCCModule(outer: LazyRoCC) extends LazyModuleImp(outer) {
-  val io = new RoCCIO(outer)
+  val io = IO(new RoCCIO(outer))
 }
 
 /** Mixins for including RoCC **/
@@ -263,7 +261,7 @@ class TranslatorExampleModule(outer: TranslatorExample)(implicit p: Parameters)
 
 class  CharacterCountExample(implicit p: Parameters) extends LazyRoCC {
   override lazy val module = new CharacterCountExampleModule(this)
-  override val atlNode = TLClientNode(TLClientParameters("CharacterCountRoCC"))
+  override val atlNode = TLClientNode(Seq(TLClientPortParameters(Seq(TLClientParameters("CharacterCountRoCC")))))
 }
 
 class CharacterCountExampleModule(outer: CharacterCountExample)(implicit p: Parameters) extends LazyRoCCModule(outer)
@@ -286,7 +284,7 @@ class CharacterCountExampleModule(outer: CharacterCountExample)(implicit p: Para
   val s_idle :: s_acq :: s_gnt :: s_check :: s_resp :: Nil = Enum(Bits(), 5)
   val state = Reg(init = s_idle)
 
-  val tl_out = io.atl.head
+  val (tl_out, edgesOut) = outer.atlNode.out(0)
   val gnt = tl_out.d.bits
   val recv_data = Reg(UInt(width = cacheDataBits))
   val recv_beat = Reg(UInt(width = log2Up(cacheDataBeats+1)), init = UInt(0))
@@ -309,7 +307,7 @@ class CharacterCountExampleModule(outer: CharacterCountExample)(implicit p: Para
   io.resp.bits.rd := resp_rd
   io.resp.bits.data := count
   tl_out.a.valid := (state === s_acq)
-  tl_out.a.bits := outer.atlNode.edgesOut(0).Get(
+  tl_out.a.bits := edgesOut.Get(
                        fromSource = UInt(0),
                        toAddress = addr_block << blockOffset,
                        lgSize = UInt(lgCacheBlockBytes))._2

src/main/scala/tile/RocketTile.scala

@@ -57,7 +57,7 @@ class RocketTile(val rocketParams: RocketTileParams, val hartid: Int)(implicit p
       val dtim = scratch.map(d => Map(
         "sifive,dtim"          -> ofRef(d.device))).getOrElse(Map())
 
-      val itim = if (!frontend.icache.slaveNode.isDefined) Map() else Map(
+      val itim = if (frontend.icache.slaveNode.edges.in.isEmpty) Map() else Map(
         "sifive,itim"          -> ofRef(frontend.icache.device))
 
       val icache = rocketParams.icache.map(i => Map(
@@ -82,7 +82,7 @@ class RocketTile(val rocketParams: RocketTileParams, val hartid: Int)(implicit p
       }))
 
       // Find all the caches
-      val outer = masterNode.edgesOut
+      val outer = masterNode.edges.out
         .flatMap(_.manager.managers)
         .filter(_.supportsAcquireB)
         .flatMap(_.resources.headOption)
@@ -115,7 +115,7 @@ class RocketTile(val rocketParams: RocketTileParams, val hartid: Int)(implicit p
     Resource(cpuDevice, "reg").bind(ResourceInt(BigInt(hartid)))
     Resource(intcDevice, "reg").bind(ResourceInt(BigInt(hartid)))
 
-    intNode.edgesIn.flatMap(_.source.sources).map { case s =>
+    intNode.edges.in.flatMap(_.source.sources).map { case s =>
       for (i <- s.range.start until s.range.end) {
        csrIntMap.lift(i).foreach { j =>
           s.resources.foreach { r =>
@@ -181,12 +181,10 @@ class RocketTileModule(outer: RocketTile) extends BaseTileModule(outer, () => ne
 
 abstract class RocketTileWrapper(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends LazyModule {
   val rocket = LazyModule(new RocketTile(rtp, hartid))
-  val masterNode: OutputNode[_,_,_,_,_]
-  val slaveNode: InputNode[_,_,_,_,_]
-  val intOutputNode = rocket.intOutputNode.map(dummy => IntOutputNode())
-  val asyncIntNode   = IntInputNode()
-  val periphIntNode  = IntInputNode()
-  val coreIntNode    = IntInputNode()
+  val asyncIntNode   : IntInwardNode
+  val periphIntNode  : IntInwardNode
+  val coreIntNode    : IntInwardNode
+  val intOutputNode = rocket.intOutputNode
   val intXbar = LazyModule(new IntXbar)
 
   rocket.intNode := intXbar.intnode
@@ -201,33 +199,25 @@ abstract class RocketTileWrapper(rtp: RocketTileParams, hartid: Int)(implicit p:
     }
   }
 
-  def optionalSlaveBuffer(in: TLOutwardNode): TLOutwardNode = {
+  def optionalSlaveBuffer(out: TLInwardNode): TLInwardNode = {
     if (rtp.boundaryBuffers) {
       val sbuf = LazyModule(new TLBuffer(BufferParams.flow, BufferParams.none, BufferParams.none, BufferParams.none, BufferParams.none))
-      DisableMonitors { implicit p => sbuf.node :*= in }
+      DisableMonitors { implicit p => out :*= sbuf.node }
       sbuf.node
     } else {
-      in
+      out
     }
   }
 
   def outputInterruptXingLatency: Int
 
-  intOutputNode.foreach { _ := rocket.intOutputNode.get }
-
   lazy val module = new LazyModuleImp(this) {
-    val io = new CoreBundle
+    val io = IO(new CoreBundle
         with HasExternallyDrivenTileConstants
         with CanHaveInstructionTracePort
         with CanHaltAndCatchFire {
-      val master = masterNode.bundleOut
-      val slave = slaveNode.bundleIn
-      val outputInterrupts = intOutputNode.map(_.bundleOut)
-      val asyncInterrupts  = asyncIntNode.bundleIn
-      val periphInterrupts = periphIntNode.bundleIn
-      val coreInterrupts   = coreIntNode.bundleIn
       val halt_and_catch_fire = rocket.module.io.halt_and_catch_fire.map(_.cloneType)
-    }
+    })
     // signals that do not change based on crossing type:
     rocket.module.io.hartid := io.hartid
     rocket.module.io.reset_vector := io.reset_vector
@@ -237,17 +227,18 @@ abstract class RocketTileWrapper(rtp: RocketTileParams, hartid: Int)(implicit p:
 }
 
 class SyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends RocketTileWrapper(rtp, hartid) {
-  val masterNode = TLOutputNode()
-  masterNode :=* optionalMasterBuffer(rocket.masterNode)
-
-  val slaveNode = new TLInputNode() { override def reverse = true }
-  DisableMonitors { implicit p => rocket.slaveNode :*= optionalSlaveBuffer(slaveNode) }
+  val masterNode = optionalMasterBuffer(rocket.masterNode)
+  val slaveNode = optionalSlaveBuffer(rocket.slaveNode)
 
   // Fully async interrupts need synchronizers.
   // Others need no synchronization.
   val xing = LazyModule(new IntXing(3))
-  xing.intnode := asyncIntNode
+  val asyncIntNode = xing.intnode
 
+  val periphIntNode = IntIdentityNode()
+  val coreIntNode = IntIdentityNode()
+
+  // order here matters
   intXbar.intnode := xing.intnode
   intXbar.intnode := periphIntNode
   intXbar.intnode := coreIntNode
@@ -256,27 +247,24 @@ class SyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters)
 }
 
 class AsyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends RocketTileWrapper(rtp, hartid) {
-  val masterNode = TLAsyncOutputNode()
   val source = LazyModule(new TLAsyncCrossingSource)
   source.node :=* rocket.masterNode
-  masterNode :=* source.node
+  val masterNode = source.node
 
-  val slaveNode = new TLAsyncInputNode() { override def reverse = true }
   val sink = LazyModule(new TLAsyncCrossingSink)
-
-  DisableMonitors { implicit p =>
-    rocket.slaveNode :*= sink.node
-    sink.node :*= slaveNode
-  }
+  DisableMonitors { implicit p => rocket.slaveNode :*= sink.node }
+  val slaveNode = sink.node
 
   // Fully async interrupts need synchronizers,
   // as do those coming from the periphery clock.
   // Others need no synchronization.
   val asyncXing = LazyModule(new IntXing(3))
   val periphXing = LazyModule(new IntXing(3))
-  asyncXing.intnode := asyncIntNode
-  periphXing.intnode := periphIntNode
+  val asyncIntNode = asyncXing.intnode
+  val periphIntNode = periphXing.intnode
+  val coreIntNode = IntIdentityNode()
 
+  // order here matters
   intXbar.intnode := asyncXing.intnode
   intXbar.intnode := periphXing.intnode
   intXbar.intnode := coreIntNode
@@ -285,18 +273,13 @@ class AsyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters
 }
 
 class RationalRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends RocketTileWrapper(rtp, hartid) {
-  val masterNode = TLRationalOutputNode()
   val source = LazyModule(new TLRationalCrossingSource)
   source.node :=* optionalMasterBuffer(rocket.masterNode)
-  masterNode :=* source.node
+  val masterNode = source.node
 
-  val slaveNode = new TLRationalInputNode() { override def reverse = true }
   val sink = LazyModule(new TLRationalCrossingSink(SlowToFast))
-
-  DisableMonitors { implicit p =>
-    sink.node :*= slaveNode
-    rocket.slaveNode :*= optionalSlaveBuffer(sink.node)
-  }
+  DisableMonitors { implicit p => optionalSlaveBuffer(rocket.slaveNode) :*= sink.node }
+  val slaveNode = sink.node
 
   // Fully async interrupts need synchronizers.
   // Those coming from periphery clock need a
@@ -304,9 +287,11 @@ class RationalRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Paramet
   // Others need no synchronization.
   val asyncXing    = LazyModule(new IntXing(3))
   val periphXing = LazyModule(new IntXing(1))
-  asyncXing.intnode := asyncIntNode
-  periphXing.intnode := periphIntNode
+  val asyncIntNode = asyncXing.intnode
+  val periphIntNode = periphXing.intnode
+  val coreIntNode = IntIdentityNode()
 
+  // order here matters
   intXbar.intnode := asyncXing.intnode
   intXbar.intnode := periphXing.intnode
   intXbar.intnode := coreIntNode

src/main/scala/tilelink/AsyncCrossing.scala

@@ -13,12 +13,7 @@ class TLAsyncCrossingSource(sync: Int = 3)(implicit p: Parameters) extends LazyM
   val node = TLAsyncSourceNode(sync)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val sink_reset_n = out.a.sink_reset_n
       val bce = edgeIn.manager.anySupportAcquireB && edgeIn.client.anySupportProbe
       val depth = edgeOut.manager.depth
@@ -47,12 +42,7 @@ class TLAsyncCrossingSink(depth: Int = 8, sync: Int = 3)(implicit p: Parameters)
   val node = TLAsyncSinkNode(depth, sync)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val source_reset_n = in.a.source_reset_n
       val bce = edgeOut.manager.anySupportAcquireB && edgeOut.client.anySupportProbe
 
@@ -97,40 +87,24 @@ object TLAsyncCrossingSink
 
 class TLAsyncCrossing(depth: Int = 8, sync: Int = 3)(implicit p: Parameters) extends LazyModule
 {
-  val nodeIn = TLInputNode()
-  val nodeOut = TLOutputNode()
-  val node = NodeHandle(nodeIn, nodeOut)
-
   val source = LazyModule(new TLAsyncCrossingSource(sync))
   val sink = LazyModule(new TLAsyncCrossingSink(depth, sync))
+  val node = NodeHandle(source.node, sink.node)
 
-  val _    = (sink.node := source.node) // no monitor
-  val in   = (source.node := nodeIn)
-  val out  = (nodeOut := sink.node)
+  sink.node := source.node
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in        = nodeIn.bundleIn
+    val io = IO(new Bundle {
       val in_clock  = Clock(INPUT)
       val in_reset  = Bool(INPUT)
-      val out       = nodeOut.bundleOut
       val out_clock = Clock(INPUT)
       val out_reset = Bool(INPUT)
-    }
+    })
 
     source.module.clock := io.in_clock
     source.module.reset := io.in_reset
-    in.foreach { lm =>
-      lm.module.clock := io.in_clock
-      lm.module.reset := io.in_reset
-    }
-
     sink.module.clock := io.out_clock
     sink.module.reset := io.out_reset
-    out.foreach { lm =>
-      lm.module.clock := io.out_clock
-      lm.module.reset := io.out_reset
-    }
   }
 }
 
@@ -145,9 +119,9 @@ class TLRAMAsyncCrossing(txns: Int)(implicit p: Parameters) extends LazyModule {
 
   model.node := fuzz.node
   cross.node := TLFragmenter(4, 256)(TLDelayer(0.1)(model.node))
-  val monitor = (ram.node := cross.node)
+  ram.node := cross.node
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
 
     // Shove the RAM into another clock domain
@@ -159,12 +133,6 @@ class TLRAMAsyncCrossing(txns: Int)(implicit p: Parameters) extends LazyModule {
     cross.module.io.in_reset := reset
     cross.module.io.out_clock := clocks.io.clock_out
     cross.module.io.out_reset := reset
-
-    // Push the Monitor into the right clock domain
-    monitor.foreach { m =>
-      m.module.clock := clocks.io.clock_out
-      m.module.reset := reset
-    }
   }
 }
 

src/main/scala/tilelink/AtomicAutomata.scala

@@ -29,12 +29,7 @@ class TLAtomicAutomata(logical: Boolean = true, arithmetic: Boolean = true, conc
     })})
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val managers = edgeOut.manager.managers
       val beatBytes = edgeOut.manager.beatBytes
 
@@ -297,7 +292,7 @@ class TLRAMAtomicAutomata(txns: Int)(implicit p: Parameters) extends LazyModule
   model.node := fuzz.node
   ram.node := TLFragmenter(4, 256)(TLDelayer(0.1)(TLAtomicAutomata()(TLDelayer(0.1)(model.node))))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }

src/main/scala/tilelink/Broadcast.scala

@@ -53,12 +53,7 @@ class TLBroadcast(lineBytes: Int, numTrackers: Int = 4, bufferless: Boolean = fa
   )
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val clients = edgeIn.client.clients
       val managers = edgeOut.manager.managers
       val lineShift = log2Ceil(lineBytes)

src/main/scala/tilelink/Buffer.scala

@@ -13,7 +13,7 @@ class TLBufferNode (
   b: BufferParams,
   c: BufferParams,
   d: BufferParams,
-  e: BufferParams)(implicit p: Parameters) extends TLAdapterNode(
+  e: BufferParams)(implicit valName: ValName) extends TLAdapterNode(
     clientFn  = { p => p.copy(minLatency = p.minLatency + b.latency + c.latency) },
     managerFn = { p => p.copy(minLatency = p.minLatency + a.latency + d.latency) }
 ) {
@@ -35,12 +35,7 @@ class TLBuffer(
   val node = new TLBufferNode(a, b, c, d, e)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       out.a <> a(in .a)
       in .d <> d(out.d)
 
@@ -79,36 +74,25 @@ object TLBuffer
 }
 
 class TLBufferChain(depth: Int)(implicit p: Parameters) extends LazyModule {
-
-  val nodeIn = TLInputNode()
-  val nodeOut = TLOutputNode()
-  val node = NodeHandle(nodeIn, nodeOut)
-
-  val buf_chain = if (depth > 0) {
-    val chain = List.fill(depth)(LazyModule(new TLBuffer(BufferParams.default)))
-
-    (chain.init zip chain.tail) foreach { case(prev, next) => next.node :=? prev.node }
-    chain
+  val buf_chain = List.fill(depth)(LazyModule(new TLBuffer(BufferParams.default)))
+  val node = if (depth > 0) {
+    (buf_chain.init zip buf_chain.tail) foreach { case (prev, next) => next.node :=? prev.node }
+    NodeHandle(buf_chain.head.node, buf_chain.last.node)
   } else {
-    List(LazyModule(new TLBuffer(BufferParams.none)))
-  }
-
-  buf_chain.head.node :=? nodeIn
-  nodeOut :=? buf_chain.last.node
-
-  lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = nodeIn.bundleIn
-      val out = nodeOut.bundleOut
-    }
+    TLIdentityNode()
   }
+  lazy val module = new LazyModuleImp(this) { }
 }
 
 object TLBufferChain
 {
   def apply(depth: Int)(x: TLOutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): TLOutwardNode = {
+    if (depth > 0) {
       val buffer = LazyModule(new TLBufferChain(depth))
       buffer.node :=? x
       buffer.node
+    } else {
+      x
+    }
   }
 }

src/main/scala/tilelink/Bus.scala

@@ -21,7 +21,8 @@ trait TLBusParams {
   def blockOffset: Int = log2Up(blockBytes)
 }
 
-abstract class TLBusWrapper(params: TLBusParams, val busName: String)(implicit p: Parameters) extends TLBusParams {
+abstract class TLBusWrapper(params: TLBusParams, val busName: String)(implicit p: Parameters)
+    extends SimpleLazyModule with LazyScope with TLBusParams {
 
   val beatBytes = params.beatBytes
   val blockBytes = params.blockBytes
@@ -71,7 +72,7 @@ abstract class TLBusWrapper(params: TLBusParams, val busName: String)(implicit p
     SourceCardinality { implicit p =>
       val chain = LazyModule(new TLBufferChain(depth))
       name.foreach { n => chain.suggestName(s"${busName}_${n}_TLBufferChain")}
-      (chain.nodeIn, chain.nodeOut)
+      (chain.node, chain.node)
     }
   }
 

src/main/scala/tilelink/CacheCork.scala

@@ -24,12 +24,7 @@ class TLCacheCork(unsafe: Boolean = false)(implicit p: Parameters) extends LazyM
           supportsAcquireT = if (m.regionType == RegionType.UNCACHED) m.supportsPutFull else m.supportsAcquireT)})})
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val clients = edgeIn.client.clients
       val caches = clients.filter(_.supportsProbe)
       require (clients.size == 1 || caches.size == 0 || unsafe, "Only one client can safely use a TLCacheCork")

src/main/scala/tilelink/Delayer.scala

@@ -10,15 +10,10 @@ import freechips.rocketchip.diplomacy._
 // q is the probability to delay a request
 class TLDelayer(q: Double)(implicit p: Parameters) extends LazyModule
 {
-  val node = TLIdentityNode()
+  val node = TLAdapterNode()
   require (0.0 <= q && q < 1)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
     def feed[T <: Data](sink: DecoupledIO[T], source: DecoupledIO[T], noise: T) {
       val allow = UInt((q * 65535.0).toInt) <= LFSR16(source.valid)
       sink.valid := source.valid && allow
@@ -27,7 +22,7 @@ class TLDelayer(q: Double)(implicit p: Parameters) extends LazyModule
       when (!sink.valid) { sink.bits := noise }
     }
 
-    (io.in zip io.out) foreach { case (in, out) =>
+    (node.in zip node.out) foreach { case ((in, _), (out, _)) =>
       val anoise = Wire(in.a.bits)
       anoise.opcode  := LFSRNoiseMaker(3)
       anoise.param   := LFSRNoiseMaker(3)

src/main/scala/tilelink/Edges.scala

@@ -4,13 +4,16 @@ package freechips.rocketchip.tilelink
 
 import Chisel._
 import chisel3.internal.sourceinfo.SourceInfo
+import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 import freechips.rocketchip.util._
 
 class TLEdge(
   client:  TLClientPortParameters,
-  manager: TLManagerPortParameters)
-  extends TLEdgeParameters(client, manager)
+  manager: TLManagerPortParameters,
+  params:  Parameters,
+  sourceInfo: SourceInfo)
+  extends TLEdgeParameters(client, manager, params, sourceInfo)
 {
   def isAligned(address: UInt, lgSize: UInt): Bool = {
     if (maxLgSize == 0) Bool(true) else {
@@ -262,8 +265,10 @@ class TLEdge(
 
 class TLEdgeOut(
   client:  TLClientPortParameters,
-  manager: TLManagerPortParameters)
-  extends TLEdge(client, manager)
+  manager: TLManagerPortParameters,
+  params:  Parameters,
+  sourceInfo: SourceInfo)
+  extends TLEdge(client, manager, params, sourceInfo)
 {
   // Transfers
   def Acquire(fromSource: UInt, toAddress: UInt, lgSize: UInt, growPermissions: UInt) = {
@@ -476,8 +481,10 @@ class TLEdgeOut(
 
 class TLEdgeIn(
   client:  TLClientPortParameters,
-  manager: TLManagerPortParameters)
-  extends TLEdge(client, manager)
+  manager: TLManagerPortParameters,
+  params:  Parameters,
+  sourceInfo: SourceInfo)
+  extends TLEdge(client, manager, params, sourceInfo)
 {
   // Transfers
   def Probe(fromAddress: UInt, toSource: UInt, lgSize: UInt, capPermissions: UInt) = {

src/main/scala/tilelink/FIFOFixer.scala

@@ -41,12 +41,7 @@ class TLFIFOFixer(policy: TLFIFOFixer.Policy = TLFIFOFixer.all)(implicit p: Para
     })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val (fixMap, splatMap) = fifoMap(edgeOut.manager.managers)
 
       // Do we need to serialize the request to this manager?

src/main/scala/tilelink/Filter.scala

@@ -49,11 +49,9 @@ class TLFilter(
     })})
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
+      out <> in
     }
-    io.out <> io.in
   }
 }
 

src/main/scala/tilelink/Fragmenter.scala

@@ -52,12 +52,7 @@ class TLFragmenter(val minSize: Int, val maxSize: Int, val alwaysMin: Boolean =
     managerFn = { m => m.copy(managers = m.managers.map(mapManager)) })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       // All managers must share a common FIFO domain (responses might end up interleaved)
       val manager   = edgeOut.manager
       val managers  = manager.managers
@@ -249,7 +244,7 @@ class TLFragmenter(val minSize: Int, val maxSize: Int, val alwaysMin: Boolean =
       val aFrag = Mux(aOrig > limit, limit, aOrig)
       val aOrigOH1 = UIntToOH1(aOrig, log2Ceil(maxSize))
       val aFragOH1 = UIntToOH1(aFrag, log2Up(maxDownSize))
-      val aHasData = node.edgesIn(0).hasData(in_a.bits)
+      val aHasData = edgeIn.hasData(in_a.bits)
       val aMask = Mux(aHasData, UInt(0), aFragOH1)
 
       val gennum = RegInit(UInt(0, width = counterBits))
@@ -317,7 +312,7 @@ class TLRAMFragmenter(ramBeatBytes: Int, maxSize: Int, txns: Int)(implicit p: Pa
     TLBuffer(BufferParams.flow)(
     model.node)))))))))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }

src/main/scala/tilelink/Fuzzer.scala

@@ -107,13 +107,11 @@ class TLFuzzer(
   val node = TLClientNode(Seq(TLClientPortParameters(clientParams)))
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val out = node.bundleOut
+    val io = IO(new Bundle {
       val finished = Bool(OUTPUT)
-    }
+    })
 
-    val out = io.out(0)
-    val edge = node.edgesOut(0)
+    val (out, edge) = node.out(0)
 
     // Extract useful parameters from the TL edge
     val maxTransfer  = edge.manager.maxTransfer
@@ -242,10 +240,10 @@ class TLFuzzRAM(txns: Int)(implicit p: Parameters) extends LazyModule
   ram2.node := TLFragmenter(16, 256)(xbar2.node)
   xbar.node := TLWidthWidget(16)(TLHintHandler()(xbar2.node))
   cross.node := TLFragmenter(4, 256)(TLBuffer()(xbar.node))
-  val monitor = (ram.node := cross.node)
+  ram.node := cross.node
   gpio.node := TLFragmenter(4, 32)(TLBuffer()(xbar.node))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
 
     // Shove the RAM into another clock domain
@@ -257,12 +255,6 @@ class TLFuzzRAM(txns: Int)(implicit p: Parameters) extends LazyModule
     cross.module.io.in_reset := reset
     cross.module.io.out_clock := clocks.io.clock_out
     cross.module.io.out_reset := reset
-
-    // Push the Monitor into the right clock domain
-    monitor.foreach { m =>
-      m.module.clock := clocks.io.clock_out
-      m.module.reset := reset
-    }
   }
 }
 

src/main/scala/tilelink/HintHandler.scala

@@ -16,12 +16,7 @@ class TLHintHandler(supportManagers: Boolean = true, supportClients: Boolean = f
     managerFn = { m => if (!supportManagers) m else m.copy(minLatency = min(1, m.minLatency), managers = m.managers.map(_.copy(supportsHint = TransferSizes(1, m.maxTransfer)))) })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       // Don't add support for clients if there is no BCE channel
       val bce = edgeOut.manager.anySupportAcquireB && edgeIn.client.anySupportProbe
       require (!supportClients || bce)
@@ -118,7 +113,7 @@ class TLRAMHintHandler(txns: Int)(implicit p: Parameters) extends LazyModule {
   model.node := fuzz.node
   ram.node := TLFragmenter(4, 256)(TLDelayer(0.1)(TLHintHandler()(TLDelayer(0.1)(model.node))))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }

src/main/scala/tilelink/IntNodes.scala

@@ -61,26 +61,13 @@ object IntSinkPortSimple
     Seq.fill(ports)(IntSinkPortParameters(Seq.fill(sinks)(IntSinkParameters())))
 }
 
-case class IntEdge(source: IntSourcePortParameters, sink: IntSinkPortParameters)
+case class IntEdge(source: IntSourcePortParameters, sink: IntSinkPortParameters, params: Parameters, sourceInfo: SourceInfo)
 
-object IntImp extends NodeImp[IntSourcePortParameters, IntSinkPortParameters, IntEdge, IntEdge, Vec[Bool]]
+object IntImp extends SimpleNodeImp[IntSourcePortParameters, IntSinkPortParameters, IntEdge, Vec[Bool]]
 {
-  def edgeO(pd: IntSourcePortParameters, pu: IntSinkPortParameters): IntEdge = IntEdge(pd, pu)
-  def edgeI(pd: IntSourcePortParameters, pu: IntSinkPortParameters): IntEdge = IntEdge(pd, pu)
-  def bundleO(eo: IntEdge): Vec[Bool] = Vec(eo.source.num, Bool())
-  def bundleI(ei: IntEdge): Vec[Bool] = Vec(ei.source.num, Bool())
-
-  def colour = "#0000ff" // blue
-  override def reverse = true
-  override def labelI(ei: IntEdge) = ei.source.sources.map(_.range.size).sum.toString
-  override def labelO(eo: IntEdge) = eo.source.sources.map(_.range.size).sum.toString
-
-  def connect(bo: => Vec[Bool], bi: => Vec[Bool], ei: => IntEdge)(implicit p: Parameters, sourceInfo: SourceInfo): (Option[LazyModule], () => Unit) = {
-    (None, () => {
-      // Cannot use bulk connect, because the widths could differ
-      (bo zip bi) foreach { case (o, i) => i := o }
-    })
-  }
+  def edge(pd: IntSourcePortParameters, pu: IntSinkPortParameters, p: Parameters, sourceInfo: SourceInfo) = IntEdge(pd, pu, p, sourceInfo)
+  def bundle(e: IntEdge) = Vec(e.source.num, Bool())
+  def render(e: IntEdge) = RenderedEdge(colour = "#0000ff" /* blue */, label = e.source.sources.map(_.range.size).sum.toString, flipped = true)
 
   override def mixO(pd: IntSourcePortParameters, node: OutwardNode[IntSourcePortParameters, IntSinkPortParameters, Vec[Bool]]): IntSourcePortParameters =
    pd.copy(sources = pd.sources.map  { s => s.copy (nodePath = node +: s.nodePath) })
@@ -88,26 +75,24 @@ object IntImp extends NodeImp[IntSourcePortParameters, IntSinkPortParameters, In
    pu.copy(sinks   = pu.sinks.map    { s => s.copy (nodePath = node +: s.nodePath) })
 }
 
-case class IntIdentityNode() extends IdentityNode(IntImp)
-case class IntSourceNode(portParams: Seq[IntSourcePortParameters]) extends SourceNode(IntImp)(portParams)
-case class IntSinkNode(portParams: Seq[IntSinkPortParameters]) extends SinkNode(IntImp)(portParams)
+case class IntSourceNode(portParams: Seq[IntSourcePortParameters])(implicit valName: ValName) extends SourceNode(IntImp)(portParams)
+case class IntSinkNode(portParams: Seq[IntSinkPortParameters])(implicit valName: ValName) extends SinkNode(IntImp)(portParams)
+case class IntAdapterNode(
+  sourceFn: IntSourcePortParameters => IntSourcePortParameters = { s => s },
+  sinkFn:   IntSinkPortParameters   => IntSinkPortParameters   = { s => s },
+  num:      Range.Inclusive = 0 to 999)(
+  implicit valName: ValName)
+  extends AdapterNode(IntImp)(sourceFn, sinkFn, num)
+case class IntIdentityNode()(implicit valName: ValName) extends IdentityNode(IntImp)()
 
 case class IntNexusNode(
   sourceFn:       Seq[IntSourcePortParameters] => IntSourcePortParameters,
   sinkFn:         Seq[IntSinkPortParameters]   => IntSinkPortParameters,
   numSourcePorts: Range.Inclusive = 0 to 128,
-  numSinkPorts:   Range.Inclusive = 0 to 128)
+  numSinkPorts:   Range.Inclusive = 0 to 128)(
+  implicit valName: ValName)
   extends NexusNode(IntImp)(sourceFn, sinkFn, numSourcePorts, numSinkPorts)
 
-case class IntOutputNode() extends OutputNode(IntImp)
-case class IntInputNode() extends InputNode(IntImp)
-
-case class IntBlindOutputNode(portParams: Seq[IntSinkPortParameters]) extends BlindOutputNode(IntImp)(portParams)
-case class IntBlindInputNode(portParams: Seq[IntSourcePortParameters]) extends BlindInputNode(IntImp)(portParams)
-
-case class IntInternalOutputNode(portParams: Seq[IntSinkPortParameters]) extends InternalOutputNode(IntImp)(portParams)
-case class IntInternalInputNode(portParams: Seq[IntSourcePortParameters]) extends InternalInputNode(IntImp)(portParams)
-
 class IntXbar()(implicit p: Parameters) extends LazyModule
 {
   val intnode = IntNexusNode(
@@ -119,27 +104,17 @@ class IntXbar()(implicit p: Parameters) extends LazyModule
     })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = intnode.bundleIn
-      val out = intnode.bundleOut
-    }
-
-    val cat = (intnode.edgesIn zip io.in).map{ case (e, i) => i.take(e.source.num) }.flatten
-    io.out.foreach { _ := cat }
+    val cat = intnode.in.map { case (i, e) => i.take(e.source.num) }.flatten
+    intnode.out.foreach { case (o, _) => o := cat }
   }
 }
 
 class IntXing(sync: Int = 3)(implicit p: Parameters) extends LazyModule
 {
-  val intnode = IntIdentityNode()
+  val intnode = IntAdapterNode()
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = intnode.bundleIn
-      val out = intnode.bundleOut
-    }
-
-    (io.in zip io.out) foreach { case (in, out) =>
+    (intnode.in zip intnode.out) foreach { case ((in, _), (out, _)) =>
       out := SynchronizerShiftReg(in, sync)
     }
   }

src/main/scala/tilelink/Isolation.scala

@@ -11,15 +11,13 @@ import freechips.rocketchip.util.AsyncBundle
 // READ the comments in the TLIsolation object before you instantiate this module
 class TLIsolation(fOut: (Bool, UInt) => UInt, fIn: (Bool, UInt) => UInt)(implicit p: Parameters) extends LazyModule
 {
-  val node = TLAsyncIdentityNode()
+  val node = TLAsyncAdapterNode()
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
+    val io = IO(new Bundle {
       val iso_out = Bool(INPUT) // Isolate from client to manager
       val iso_in  = Bool(INPUT) // Isolate from manager to client
-    }
+    })
 
     def ISOo[T <: Data](x: T): T = x.fromBits(fOut(io.iso_out, x.asUInt))
     def ISOi[T <: Data](x: T): T = x.fromBits(fIn (io.iso_in,  x.asUInt))
@@ -53,7 +51,7 @@ class TLIsolation(fOut: (Bool, UInt) => UInt, fIn: (Bool, UInt) => UInt)(implici
       y.sink_reset_n   := Bool(false)
     }
 
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       ABo(out.a, in .a)
       ABi(in .d, out.d)
 

src/main/scala/tilelink/Map.scala

@@ -20,14 +20,8 @@ class TLMap(fn: AddressSet => BigInt)(implicit p: Parameters) extends LazyModule
           AddressSet(fn(a), a.mask)))))})
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    io.out <> io.in
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
+      out <> in
       val convert = edgeIn.manager.managers.flatMap(_.address) zip edgeOut.manager.managers.flatMap(_.address)
       def forward(x: UInt) =
         convert.map { case (i, o) => Mux(i.contains(x), UInt(o.base) | (x & UInt(o.mask)), UInt(0)) }.reduce(_ | _)

src/main/scala/tilelink/Monitor.scala

@@ -8,26 +8,22 @@ import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 import freechips.rocketchip.util.{HeterogeneousBag, PlusArg}
 
-case class TLMonitorArgs(edge: () => Seq[TLEdge], sourceInfo: SourceInfo, p: Parameters)
+case class TLMonitorArgs(edge: TLEdge)
 
-abstract class TLMonitorBase(args: TLMonitorArgs) extends MonitorBase()(args.sourceInfo, args.p)
+abstract class TLMonitorBase(args: TLMonitorArgs) extends Module
 {
-  def legalize(bundle: TLBundleSnoop, edge: TLEdge, reset: Bool): Unit
-
-  lazy val module = new LazyModuleImp(this) {
-    val edges = args.edge()
   val io = new Bundle {
-      val in = HeterogeneousBag(edges.map(p => new TLBundleSnoop(p.bundle))).flip
+    val in = new TLBundleSnoop(args.edge.bundle).flip
   }
 
-    (edges zip io.in).foreach { case (e, in) => legalize(in, e, reset) }
-  }
+  def legalize(bundle: TLBundleSnoop, edge: TLEdge, reset: Bool): Unit
+  legalize(io.in, args.edge, reset)
 }
 
 class TLMonitor(args: TLMonitorArgs) extends TLMonitorBase(args)
 {
   def extra = {
-    sourceInfo match {
+    args.edge.sourceInfo match {
       case SourceLine(filename, line, col) => s" (connected at $filename:$line:$col)"
       case _ => ""
     }

src/main/scala/tilelink/NodeNumberer.scala

@@ -7,11 +7,8 @@ import chisel3.internal.sourceinfo.SourceInfo
 import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 
-case class TLNodeNumbererNode(nodeAddressOffset: Option[Int] = None) extends TLCustomNode(0 to 999, 0 to 999)
+case class TLNodeNumbererNode(nodeAddressOffset: Option[Int] = None)(implicit valName: ValName) extends TLCustomNode(0 to 999, 0 to 999)
 {
-  val externalIn = true
-  val externalOut = true
-
   def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
     require (oStars + iStars <= 1, s"${name} (a custom adapter) appears left of a :*= ${iStars} times and right of a :=* ${oStars} times; at most once is allowed${lazyModule.line}")
     if (oStars > 0) {
@@ -46,15 +43,10 @@ class TLNodeNumberer(nodeAddressOffset: Option[Int] = None)(implicit p: Paramete
   val node = TLNodeNumbererNode(nodeAddressOffset)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    val minNodeOffset = log2Ceil(node.edgesOut.map(_.manager.maxAddress).max)
+    val minNodeOffset = log2Ceil(node.edges.out.map(_.manager.maxAddress).max)
     val nodeOffset = nodeAddressOffset.getOrElse(minNodeOffset)
 
-    (io.in zip io.out).zipWithIndex foreach { case ((in, out), i) =>
+    (node.in zip node.out).zipWithIndex foreach { case (((in, _), (out, _)), i) =>
       out <> in
       // a&c address already get truncated
       in.b.bits.address := (UInt(i+1) << nodeOffset) | out.b.bits.address

src/main/scala/tilelink/Nodes.scala

@@ -9,40 +9,21 @@ import freechips.rocketchip.diplomacy._
 import freechips.rocketchip.util.RationalDirection
 import scala.collection.mutable.ListBuffer
 
-case object TLMonitorBuilder extends Field[TLMonitorArgs => TLMonitorBase](args => LazyModule(new TLMonitor(args)))
-case object TLCombinationalCheck extends Field[Boolean](false)
+case object TLMonitorBuilder extends Field[TLMonitorArgs => TLMonitorBase](args => new TLMonitor(args))
 
 object TLImp extends NodeImp[TLClientPortParameters, TLManagerPortParameters, TLEdgeOut, TLEdgeIn, TLBundle]
 {
-  def edgeO(pd: TLClientPortParameters, pu: TLManagerPortParameters): TLEdgeOut = new TLEdgeOut(pd, pu)
-  def edgeI(pd: TLClientPortParameters, pu: TLManagerPortParameters): TLEdgeIn  = new TLEdgeIn(pd, pu)
+  def edgeO(pd: TLClientPortParameters, pu: TLManagerPortParameters, p: Parameters, sourceInfo: SourceInfo) = new TLEdgeOut(pd, pu, p, sourceInfo)
+  def edgeI(pd: TLClientPortParameters, pu: TLManagerPortParameters, p: Parameters, sourceInfo: SourceInfo) = new TLEdgeIn (pd, pu, p, sourceInfo)
 
-  def bundleO(eo: TLEdgeOut): TLBundle = TLBundle(eo.bundle)
-  def bundleI(ei: TLEdgeIn):  TLBundle = TLBundle(ei.bundle)
+  def bundleO(eo: TLEdgeOut) = TLBundle(eo.bundle)
+  def bundleI(ei: TLEdgeIn)  = TLBundle(ei.bundle)
 
-  def colour = "#000000" // black
-  override def labelI(ei: TLEdgeIn)  = (ei.manager.beatBytes * 8).toString
-  override def labelO(eo: TLEdgeOut) = (eo.manager.beatBytes * 8).toString
+  def render(ei: TLEdgeIn) = RenderedEdge(colour = "#000000" /* black */, label = (ei.manager.beatBytes * 8).toString)
 
-  override def connect(edges: () => Seq[TLEdgeIn], bundles: () => Seq[(TLBundle, TLBundle)], enableMonitoring: Boolean)
-                      (implicit p: Parameters, sourceInfo: SourceInfo): (Option[TLMonitorBase], () => Unit) = {
-    val monitor = if (enableMonitoring) Some(p(TLMonitorBuilder)(TLMonitorArgs(edges, sourceInfo, p))) else None
-    (monitor, () => {
-      val eval = bundles ()
-      monitor.foreach { m => (eval zip m.module.io.in) foreach { case ((i,o), m) => m := TLBundleSnoop(o,i) } }
-      eval.foreach { case (bi, bo) =>
-        bi <> bo
-        if (p(TLCombinationalCheck)) {
-          // It is forbidden for valid to depend on ready in TL2
-          // If someone did that, then this will create a detectable combinational loop
-          bo.a.ready := bi.a.ready && bo.a.valid
-          bi.b.ready := bo.b.ready && bi.b.valid
-          bo.c.ready := bi.c.ready && bo.c.valid
-          bi.d.ready := bo.d.ready && bi.d.valid
-          bo.e.ready := bi.e.ready && bo.e.valid
-        }
-      }
-    })
+  override def monitor(bundle: TLBundle, edge: TLEdgeIn) {
+    val monitor = Module(edge.params(TLMonitorBuilder)(TLMonitorArgs(edge)))
+    monitor.io.in := TLBundleSnoop(bundle, bundle)
   }
 
   override def mixO(pd: TLClientPortParameters, node: OutwardNode[TLClientPortParameters, TLManagerPortParameters, TLBundle]): TLClientPortParameters  =
@@ -63,92 +44,39 @@ object TLImp extends NodeImp[TLClientPortParameters, TLManagerPortParameters, TL
   }
 }
 
-// Nodes implemented inside modules
-case class TLIdentityNode() extends IdentityNode(TLImp)
-case class TLClientNode(portParams: Seq[TLClientPortParameters]) extends SourceNode(TLImp)(portParams)
-case class TLManagerNode(portParams: Seq[TLManagerPortParameters]) extends SinkNode(TLImp)(portParams)
-
-object TLClientNode
-{
-  def apply(params: TLClientParameters) =
-    new TLClientNode(Seq(TLClientPortParameters(Seq(params))))
-}
-
-object TLManagerNode
-{
-  def apply(beatBytes: Int, params: TLManagerParameters) =
-    new TLManagerNode(Seq(TLManagerPortParameters(Seq(params), beatBytes, minLatency = 0)))
-}
+case class TLClientNode(portParams: Seq[TLClientPortParameters])(implicit valName: ValName) extends SourceNode(TLImp)(portParams)
+case class TLManagerNode(portParams: Seq[TLManagerPortParameters])(implicit valName: ValName) extends SinkNode(TLImp)(portParams)
 
 case class TLAdapterNode(
-  clientFn:  TLClientPortParameters  => TLClientPortParameters,
-  managerFn: TLManagerPortParameters => TLManagerPortParameters,
-  num:       Range.Inclusive = 0 to 999)
+  clientFn:  TLClientPortParameters  => TLClientPortParameters  = { s => s },
+  managerFn: TLManagerPortParameters => TLManagerPortParameters = { s => s },
+  num:       Range.Inclusive = 0 to 999)(
+  implicit valName: ValName)
   extends AdapterNode(TLImp)(clientFn, managerFn, num)
 
+case class TLIdentityNode()(implicit valName: ValName) extends IdentityNode(TLImp)()
+
 case class TLNexusNode(
   clientFn:        Seq[TLClientPortParameters]  => TLClientPortParameters,
   managerFn:       Seq[TLManagerPortParameters] => TLManagerPortParameters,
   numClientPorts:  Range.Inclusive = 1 to 999,
-  numManagerPorts: Range.Inclusive = 1 to 999)
+  numManagerPorts: Range.Inclusive = 1 to 999)(
+  implicit valName: ValName)
   extends NexusNode(TLImp)(clientFn, managerFn, numClientPorts, numManagerPorts)
 
-case class TLSplitterNode(
-  clientFn:        SplitterArg[TLClientPortParameters]  => Seq[TLClientPortParameters],
-  managerFn:       SplitterArg[TLManagerPortParameters] => Seq[TLManagerPortParameters],
-  numClientPorts:  Range.Inclusive = 0 to 999,
-  numManagerPorts: Range.Inclusive = 0 to 999)
-  extends SplitterNode(TLImp)(clientFn, managerFn, numClientPorts, numManagerPorts)
-
 abstract class TLCustomNode(
   numClientPorts:  Range.Inclusive,
-  numManagerPorts: Range.Inclusive)
+  numManagerPorts: Range.Inclusive)(
+  implicit valName: ValName)
   extends CustomNode(TLImp)(numClientPorts, numManagerPorts)
 
-// Nodes passed from an inner module
-case class TLOutputNode() extends OutputNode(TLImp)
-case class TLInputNode() extends InputNode(TLImp)
+// Asynchronous crossings
 
-// Nodes used for external ports
-case class TLBlindOutputNode(portParams: Seq[TLManagerPortParameters]) extends BlindOutputNode(TLImp)(portParams)
-case class TLBlindInputNode(portParams: Seq[TLClientPortParameters]) extends BlindInputNode(TLImp)(portParams)
-
-case class TLInternalOutputNode(portParams: Seq[TLManagerPortParameters]) extends InternalOutputNode(TLImp)(portParams)
-case class TLInternalInputNode(portParams: Seq[TLClientPortParameters]) extends InternalInputNode(TLImp)(portParams)
-
-/** Synthesizeable unit tests */
-import freechips.rocketchip.unittest._
-
-class TLInputNodeTest(txns: Int = 5000, timeout: Int = 500000)(implicit p: Parameters) extends UnitTest(timeout) {
-  class Acceptor extends LazyModule {
-    val node = TLInputNode()
-    val tlram = LazyModule(new TLRAM(AddressSet(0x54321000, 0xfff)))
-    tlram.node := node
-
-    lazy val module = new LazyModuleImp(this) {
-      val io = new Bundle {
-        val in = node.bundleIn
-      }
-    }
-  }
-
-  val fuzzer = LazyModule(new TLFuzzer(txns))
-  LazyModule(new Acceptor).node := TLFragmenter(4, 64)(fuzzer.node)
-
-  io.finished := Module(fuzzer.module).io.finished
-}
-
-object TLAsyncImp extends NodeImp[TLAsyncClientPortParameters, TLAsyncManagerPortParameters, TLAsyncEdgeParameters, TLAsyncEdgeParameters, TLAsyncBundle]
+object TLAsyncImp extends SimpleNodeImp[TLAsyncClientPortParameters, TLAsyncManagerPortParameters, TLAsyncEdgeParameters, TLAsyncBundle]
 {
-  def edgeO(pd: TLAsyncClientPortParameters, pu: TLAsyncManagerPortParameters): TLAsyncEdgeParameters = TLAsyncEdgeParameters(pd, pu)
-  def edgeI(pd: TLAsyncClientPortParameters, pu: TLAsyncManagerPortParameters): TLAsyncEdgeParameters = TLAsyncEdgeParameters(pd, pu)
-
-  def bundleO(eo: TLAsyncEdgeParameters): TLAsyncBundle = new TLAsyncBundle(eo.bundle)
-  def bundleI(ei: TLAsyncEdgeParameters): TLAsyncBundle = new TLAsyncBundle(ei.bundle)
-
-  def colour = "#ff0000" // red
-  override def labelI(ei: TLAsyncEdgeParameters) = ei.manager.depth.toString
-  override def labelO(eo: TLAsyncEdgeParameters) = eo.manager.depth.toString
+  def edge(pd: TLAsyncClientPortParameters, pu: TLAsyncManagerPortParameters, p: Parameters, sourceInfo: SourceInfo) = TLAsyncEdgeParameters(pd, pu, p, sourceInfo)
+  def bundle(e: TLAsyncEdgeParameters) = new TLAsyncBundle(e.bundle)
+  def render(e: TLAsyncEdgeParameters) = RenderedEdge(colour = "#ff0000" /* red */, label = e.manager.depth.toString)
 
   override def mixO(pd: TLAsyncClientPortParameters, node: OutwardNode[TLAsyncClientPortParameters, TLAsyncManagerPortParameters, TLAsyncBundle]): TLAsyncClientPortParameters  =
    pd.copy(base = pd.base.copy(clients  = pd.base.clients.map  { c => c.copy (nodePath = node +: c.nodePath) }))
@@ -156,29 +84,32 @@ object TLAsyncImp extends NodeImp[TLAsyncClientPortParameters, TLAsyncManagerPor
    pu.copy(base = pu.base.copy(managers = pu.base.managers.map { m => m.copy (nodePath = node +: m.nodePath) }))
 }
 
-case class TLAsyncIdentityNode() extends IdentityNode(TLAsyncImp)
-case class TLAsyncOutputNode() extends OutputNode(TLAsyncImp)
-case class TLAsyncInputNode() extends InputNode(TLAsyncImp)
+case class TLAsyncAdapterNode(
+  clientFn:  TLAsyncClientPortParameters  => TLAsyncClientPortParameters  = { s => s },
+  managerFn: TLAsyncManagerPortParameters => TLAsyncManagerPortParameters = { s => s },
+  num:       Range.Inclusive = 0 to 999)(
+  implicit valName: ValName)
+  extends AdapterNode(TLAsyncImp)(clientFn, managerFn, num)
 
-case class TLAsyncSourceNode(sync: Int)
+case class TLAsyncIdentityNode()(implicit valName: ValName) extends IdentityNode(TLAsyncImp)()
+
+case class TLAsyncSourceNode(sync: Int)(implicit valName: ValName)
   extends MixedAdapterNode(TLImp, TLAsyncImp)(
     dFn = { p => TLAsyncClientPortParameters(p) },
     uFn = { p => p.base.copy(minLatency = sync+1) }) // discard cycles in other clock domain
 
-case class TLAsyncSinkNode(depth: Int, sync: Int)
+case class TLAsyncSinkNode(depth: Int, sync: Int)(implicit valName: ValName)
   extends MixedAdapterNode(TLAsyncImp, TLImp)(
     dFn = { p => p.base.copy(minLatency = sync+1) },
     uFn = { p => TLAsyncManagerPortParameters(depth, p) })
 
-object TLRationalImp extends NodeImp[TLRationalClientPortParameters, TLRationalManagerPortParameters, TLRationalEdgeParameters, TLRationalEdgeParameters, TLRationalBundle]
+// Rationally related crossings
+
+object TLRationalImp extends SimpleNodeImp[TLRationalClientPortParameters, TLRationalManagerPortParameters, TLRationalEdgeParameters, TLRationalBundle]
 {
-  def edgeO(pd: TLRationalClientPortParameters, pu: TLRationalManagerPortParameters): TLRationalEdgeParameters = TLRationalEdgeParameters(pd, pu)
-  def edgeI(pd: TLRationalClientPortParameters, pu: TLRationalManagerPortParameters): TLRationalEdgeParameters = TLRationalEdgeParameters(pd, pu)
-
-  def bundleO(eo: TLRationalEdgeParameters): TLRationalBundle = new TLRationalBundle(eo.bundle)
-  def bundleI(ei: TLRationalEdgeParameters): TLRationalBundle = new TLRationalBundle(ei.bundle)
-
-  def colour = "#00ff00" // green
+  def edge(pd: TLRationalClientPortParameters, pu: TLRationalManagerPortParameters, p: Parameters, sourceInfo: SourceInfo) = TLRationalEdgeParameters(pd, pu, p, sourceInfo)
+  def bundle(e: TLRationalEdgeParameters) = new TLRationalBundle(e.bundle)
+  def render(e: TLRationalEdgeParameters) = RenderedEdge(colour = "#00ff00" /* green */)
 
   override def mixO(pd: TLRationalClientPortParameters, node: OutwardNode[TLRationalClientPortParameters, TLRationalManagerPortParameters, TLRationalBundle]): TLRationalClientPortParameters  =
    pd.copy(base = pd.base.copy(clients  = pd.base.clients.map  { c => c.copy (nodePath = node +: c.nodePath) }))
@@ -186,16 +117,21 @@ object TLRationalImp extends NodeImp[TLRationalClientPortParameters, TLRationalM
    pu.copy(base = pu.base.copy(managers = pu.base.managers.map { m => m.copy (nodePath = node +: m.nodePath) }))
 }
 
-case class TLRationalIdentityNode() extends IdentityNode(TLRationalImp)
-case class TLRationalOutputNode() extends OutputNode(TLRationalImp)
-case class TLRationalInputNode() extends InputNode(TLRationalImp)
+case class TLRationalAdapterNode(
+  clientFn:  TLRationalClientPortParameters  => TLRationalClientPortParameters  = { s => s },
+  managerFn: TLRationalManagerPortParameters => TLRationalManagerPortParameters = { s => s },
+  num:       Range.Inclusive = 0 to 999)(
+  implicit valName: ValName)
+  extends AdapterNode(TLRationalImp)(clientFn, managerFn, num)
 
-case class TLRationalSourceNode()
+case class TLRationalIdentityNode()(implicit valName: ValName) extends IdentityNode(TLRationalImp)()
+
+case class TLRationalSourceNode()(implicit valName: ValName)
   extends MixedAdapterNode(TLImp, TLRationalImp)(
     dFn = { p => TLRationalClientPortParameters(p) },
     uFn = { p => p.base.copy(minLatency = 1) }) // discard cycles from other clock domain
 
-case class TLRationalSinkNode(direction: RationalDirection)
+case class TLRationalSinkNode(direction: RationalDirection)(implicit valName: ValName)
   extends MixedAdapterNode(TLRationalImp, TLImp)(
     dFn = { p => p.base.copy(minLatency = 1) },
     uFn = { p => TLRationalManagerPortParameters(direction, p) })

src/main/scala/tilelink/Parameters.scala

@@ -3,6 +3,8 @@
 package freechips.rocketchip.tilelink
 
 import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
+import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 import freechips.rocketchip.util.RationalDirection
 import scala.math.max
@@ -312,7 +314,9 @@ object TLBundleParameters
 
 case class TLEdgeParameters(
   client:  TLClientPortParameters,
-  manager: TLManagerPortParameters)
+  manager: TLManagerPortParameters,
+  params:  Parameters,
+  sourceInfo: SourceInfo)
 {
   val maxTransfer = max(client.maxTransfer, manager.maxTransfer)
   val maxLgSize = log2Ceil(maxTransfer)
@@ -340,7 +344,7 @@ object TLAsyncBundleParameters
   def union(x: Seq[TLAsyncBundleParameters]) = x.foldLeft(emptyBundleParams)((x,y) => x.union(y))
 }
 
-case class TLAsyncEdgeParameters(client: TLAsyncClientPortParameters, manager: TLAsyncManagerPortParameters)
+case class TLAsyncEdgeParameters(client: TLAsyncClientPortParameters, manager: TLAsyncManagerPortParameters, params: Parameters, sourceInfo: SourceInfo)
 {
   val bundle = TLAsyncBundleParameters(manager.depth, TLBundleParameters(client.base, manager.base))
 }
@@ -348,7 +352,7 @@ case class TLAsyncEdgeParameters(client: TLAsyncClientPortParameters, manager: T
 case class TLRationalManagerPortParameters(direction: RationalDirection, base: TLManagerPortParameters)
 case class TLRationalClientPortParameters(base: TLClientPortParameters)
 
-case class TLRationalEdgeParameters(client: TLRationalClientPortParameters, manager: TLRationalManagerPortParameters)
+case class TLRationalEdgeParameters(client: TLRationalClientPortParameters, manager: TLRationalManagerPortParameters, params: Parameters, sourceInfo: SourceInfo)
 {
   val bundle = TLBundleParameters(client.base, manager.base)
 }

src/main/scala/tilelink/PatternPusher.scala

@@ -38,13 +38,12 @@ class TLPatternPusher(name: String, pattern: Seq[Pattern])(implicit p: Parameter
   val node = TLClientNode(Seq(TLClientPortParameters(Seq(TLClientParameters(name = name)))))
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val tl_out = node.bundleOut
+    val io = IO(new Bundle {
       val run = Bool(INPUT)
       val done = Bool(OUTPUT)
-    }
+    })
 
-    val edgeOut = node.edgesOut(0)
+    val (tl_out, edgeOut) = node.out(0)
     pattern.foreach { p =>
       require (p.size <= log2Ceil(edgeOut.manager.beatBytes), "Patterns must fit in a single beat")
     }
@@ -56,8 +55,8 @@ class TLPatternPusher(name: String, pattern: Seq[Pattern])(implicit p: Parameter
     val end = step === UInt(pattern.size)
     io.done := end && !flight
 
-    val a = io.tl_out(0).a
-    val d = io.tl_out(0).d
+    val a = tl_out.a
+    val d = tl_out.d
 
     // Expected response?
     val check  = Vec(pattern.map(p => Bool(p.dataIn.isDefined)))(step) holdUnless a.fire()
@@ -80,8 +79,8 @@ class TLPatternPusher(name: String, pattern: Seq[Pattern])(implicit p: Parameter
     d.ready := Bool(true)
 
     // Tie off unused channels
-    io.tl_out(0).b.ready := Bool(true)
-    io.tl_out(0).c.valid := Bool(false)
-    io.tl_out(0).e.valid := Bool(false)
+    tl_out.b.ready := Bool(true)
+    tl_out.c.valid := Bool(false)
+    tl_out.e.valid := Bool(false)
   }
 }

src/main/scala/tilelink/RAMModel.scala

@@ -24,15 +24,10 @@ import freechips.rocketchip.util._
 
 class TLRAMModel(log: String = "")(implicit p: Parameters) extends LazyModule
 {
-  val node = TLIdentityNode()
+  val node = TLAdapterNode()
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val edge         = edgeIn
       val endAddress   = edge.manager.maxAddress + 1
       val endSourceId  = edge.client.endSourceId

src/main/scala/tilelink/RationalCrossing.scala

@@ -20,12 +20,7 @@ class TLRationalCrossingSource(implicit p: Parameters) extends LazyModule
   val node = TLRationalSourceNode()
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val bce = edgeIn.manager.anySupportAcquireB && edgeIn.client.anySupportProbe
       val direction = edgeOut.manager.direction
 
@@ -56,12 +51,7 @@ class TLRationalCrossingSink(direction: RationalDirection = Symmetric)(implicit
   val node = TLRationalSinkNode(direction)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val bce = edgeOut.manager.anySupportAcquireB && edgeOut.client.anySupportProbe
       val direction = edgeIn.manager.direction
 
@@ -109,40 +99,24 @@ object TLRationalCrossingSink
 
 class TLRationalCrossing(direction: RationalDirection = Symmetric)(implicit p: Parameters) extends LazyModule
 {
-  val nodeIn = TLInputNode()
-  val nodeOut = TLOutputNode()
-  val node = NodeHandle(nodeIn, nodeOut)
-
   val source = LazyModule(new TLRationalCrossingSource)
   val sink = LazyModule(new TLRationalCrossingSink(direction))
+  val node = NodeHandle(source.node, sink.node)
 
-  val _    = (sink.node := source.node) // no monitor
-  val in   = (source.node := nodeIn)
-  val out  = (nodeOut := sink.node)
+  sink.node := source.node
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in        = nodeIn.bundleIn
+    val io = IO(new Bundle {
       val in_clock  = Clock(INPUT)
       val in_reset  = Bool(INPUT)
-      val out       = nodeOut.bundleOut
       val out_clock = Clock(INPUT)
       val out_reset = Bool(INPUT)
-    }
+    })
 
     source.module.clock := io.in_clock
     source.module.reset := io.in_reset
-    in.foreach { lm =>
-      lm.module.clock := io.in_clock
-      lm.module.reset := io.in_reset
-    }
-
     sink.module.clock := io.out_clock
     sink.module.reset := io.out_reset
-    out.foreach { lm =>
-      lm.module.clock := io.out_clock
-      lm.module.reset := io.out_reset
-    }
   }
 }
 
@@ -150,7 +124,7 @@ class TLRationalCrossing(direction: RationalDirection = Symmetric)(implicit p: P
 import freechips.rocketchip.unittest._
 
 class TLRAMRationalCrossingSource(name: String, txns: Int)(implicit p: Parameters) extends LazyModule {
-  val node = TLRationalOutputNode()
+  val node = TLRationalIdentityNode()
   val fuzz  = LazyModule(new TLFuzzer(txns))
   val model = LazyModule(new TLRAMModel(name))
 
@@ -158,25 +132,20 @@ class TLRAMRationalCrossingSource(name: String, txns: Int)(implicit p: Parameter
   node := TLRationalCrossingSource()(TLDelayer(0.25)(model.node))
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
+    val io = IO(new Bundle {
       val finished = Bool(OUTPUT)
-      val out = node.bundleOut
-    }
+    })
     io.finished := fuzz.module.io.finished
   }
 }
 
 class TLRAMRationalCrossingSink(direction: RationalDirection)(implicit p: Parameters) extends LazyModule {
-  val node = TLRationalInputNode()
+  val node = TLRationalIdentityNode()
   val ram  = LazyModule(new TLRAM(AddressSet(0x0, 0x3ff)))
 
   ram.node := TLFragmenter(4, 256)(TLDelayer(0.25)(TLRationalCrossingSink(direction)(node)))
 
-  lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-  }
+  lazy val module = new LazyModuleImp(this) { }
 }
 
 class TLRAMRationalCrossing(txns: Int)(implicit p: Parameters) extends LazyModule {
@@ -196,7 +165,7 @@ class TLRAMRationalCrossing(txns: Int)(implicit p: Parameters) extends LazyModul
   val fix_fast_sink   = LazyModule(new TLRAMRationalCrossingSink(SlowToFast))
   fix_fast_sink.node := fix_slow_source.node
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished :=
       sym_fast_source.module.io.finished &&
       sym_slow_source.module.io.finished &&

src/main/scala/tilelink/RegisterRouter.scala

@@ -9,15 +9,16 @@ import freechips.rocketchip.regmapper._
 import freechips.rocketchip.util.HeterogeneousBag
 import scala.math.{min,max}
 
-class TLRegisterNode(
+case class TLRegisterNode(
     address:     Seq[AddressSet],
     device:      Device,
     deviceKey:   String  = "reg/control",
     concurrency: Int     = 0,
     beatBytes:   Int     = 4,
     undefZero:   Boolean = true,
-    executable:  Boolean = false)
-  extends TLManagerNode(Seq(TLManagerPortParameters(
+    executable:  Boolean = false)(
+    implicit valName: ValName)
+  extends SinkNode(TLImp)(Seq(TLManagerPortParameters(
     Seq(TLManagerParameters(
       address            = address,
       resources          = Seq(Resource(device, deviceKey)),
@@ -39,9 +40,9 @@ class TLRegisterNode(
   // Calling this method causes the matching TL2 bundle to be
   // configured to route all requests to the listed RegFields.
   def regmap(mapping: RegField.Map*) = {
-    val a = bundleIn(0).a
-    val d = bundleIn(0).d
-    val edge = edgesIn(0)
+    val (bundleIn, edge) = this.in(0)
+    val a = bundleIn.a
+    val d = bundleIn.d
 
     // Please forgive me ...
     val baseEnd = 0
@@ -75,26 +76,12 @@ class TLRegisterNode(
     d.bits.opcode := Mux(out.bits.read, TLMessages.AccessAckData, TLMessages.AccessAck)
 
     // Tie off unused channels
-    bundleIn(0).b.valid := Bool(false)
-    bundleIn(0).c.ready := Bool(true)
-    bundleIn(0).e.ready := Bool(true)
+    bundleIn.b.valid := Bool(false)
+    bundleIn.c.ready := Bool(true)
+    bundleIn.e.ready := Bool(true)
   }
 }
 
-object TLRegisterNode
-{
-  def apply(
-      address:     Seq[AddressSet],
-      device:      Device,
-      deviceKey:   String  = "reg/control",
-      concurrency: Int     = 0,
-      beatBytes:   Int     = 4,
-      undefZero:   Boolean = true,
-      executable:  Boolean = false) =
-    new TLRegisterNode(address, device, deviceKey, concurrency, beatBytes, undefZero, executable)
-}
-
-// These convenience methods below combine to make it possible to create a TL2 
 // register mapped device from a totally abstract register mapped device.
 // See GPIO.scala in this directory for an example
 
@@ -105,13 +92,11 @@ abstract class TLRegisterRouterBase(devname: String, devcompat: Seq[String], val
   val intnode = IntSourceNode(IntSourcePortSimple(num = interrupts, resources = Seq(Resource(device, "int"))))
 }
 
-case class TLRegBundleArg(interrupts: HeterogeneousBag[Vec[Bool]], in: HeterogeneousBag[TLBundle])(implicit val p: Parameters)
+case class TLRegBundleArg()(implicit val p: Parameters)
 
 class TLRegBundleBase(arg: TLRegBundleArg) extends Bundle
 {
   implicit val p = arg.p
-  val interrupts = arg.interrupts
-  val in = arg.in
 }
 
 class TLRegBundle[P](val params: P, arg: TLRegBundleArg)(implicit p: Parameters) extends TLRegBundleBase(arg)
@@ -119,8 +104,8 @@ class TLRegBundle[P](val params: P, arg: TLRegBundleArg)(implicit p: Parameters)
 class TLRegModule[P, B <: TLRegBundleBase](val params: P, bundleBuilder: => B, router: TLRegisterRouterBase)
   extends LazyModuleImp(router) with HasRegMap
 {
-  val io = bundleBuilder
-  val interrupts = if (io.interrupts.isEmpty) Vec(0, Bool()) else io.interrupts(0)
+  val io = IO(bundleBuilder)
+  val interrupts = if (router.intnode.out.isEmpty) Vec(0, Bool()) else router.intnode.out(0)._1
   val address = router.address
   def regmap(mapping: RegField.Map*) = router.node.regmap(mapping:_*)
 }
@@ -142,5 +127,7 @@ class TLRegisterRouter[B <: TLRegBundleBase, M <: LazyModuleImp](
   require (isPow2(size))
   // require (size >= 4096) ... not absolutely required, but highly recommended
 
-  lazy val module = moduleBuilder(bundleBuilder(TLRegBundleArg(intnode.bundleOut, node.bundleIn)), this)
+  lazy val module = moduleBuilder(bundleBuilder(TLRegBundleArg()), this)
 }
+
+// !!! eliminate third trait

src/main/scala/tilelink/RegisterRouterTest.scala

@@ -3,6 +3,7 @@
 package freechips.rocketchip.tilelink
 
 import Chisel._
+import chisel3.experimental.MultiIOModule
 import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 import freechips.rocketchip.regmapper._
@@ -222,7 +223,7 @@ trait RRTest1Bundle
 {
 }
 
-trait RRTest1Module extends Module with HasRegMap
+trait RRTest1Module extends MultiIOModule with HasRegMap
 {
   val clocks = Module(new Pow2ClockDivider(2))
 
@@ -261,7 +262,7 @@ class FuzzRRTest0(txns: Int)(implicit p: Parameters) extends LazyModule {
 
   rrtr.node := TLFragmenter(4, 32)(TLDelayer(0.1)(fuzz.node))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }
@@ -276,7 +277,7 @@ class FuzzRRTest1(txns: Int)(implicit p: Parameters) extends LazyModule {
 
   rrtr.node := TLFragmenter(4, 32)(TLDelayer(0.1)(fuzz.node))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }

src/main/scala/tilelink/SRAM.scala

@@ -29,18 +29,12 @@ class TLRAM(address: AddressSet, executable: Boolean = true, beatBytes: Int = 4,
   // We require the address range to include an entire beat (for the write mask)
   require ((address.mask & (beatBytes-1)) == beatBytes-1)
 
-  lazy val module = new Implementation
-  @chiselName class Implementation extends LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-    }
-
+  lazy val module = new LazyModuleImp(this) {
     def bigBits(x: BigInt, tail: List[Boolean] = List.empty[Boolean]): List[Boolean] =
       if (x == 0) tail.reverse else bigBits(x >> 1, ((x & 1) == 1) :: tail)
     val mask = bigBits(address.mask >> log2Ceil(beatBytes))
 
-    val in = io.in(0)
-    val edge = node.edgesIn(0)
+    val (in, edge) = node.in(0)
 
     val addrBits = (mask zip edge.addr_hi(in.a.bits).toBools).filter(_._1).map(_._2)
     val a_legal = address.contains(in.a.bits.address)
@@ -101,7 +95,7 @@ class TLRAMSimple(ramBeatBytes: Int, txns: Int)(implicit p: Parameters) extends
   model.node := fuzz.node
   ram.node := TLDelayer(0.25)(model.node)
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }

src/main/scala/tilelink/SourceShrinker.scala

@@ -23,12 +23,7 @@ class TLSourceShrinker(maxInFlight: Int)(implicit p: Parameters) extends LazyMod
     managerFn = { mp => mp.copy(managers = mp.managers.map(_.copy(fifoId = None)))  })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       // Acquires cannot pass this adapter; it makes Probes impossible
       require (!edgeIn.client.anySupportProbe || 
                !edgeOut.manager.anySupportAcquireB)

src/main/scala/tilelink/Splitter.scala

@@ -6,6 +6,34 @@ import Chisel._
 import freechips.rocketchip.config.Parameters
 import freechips.rocketchip.diplomacy._
 
+case class SplitterArg[T](newSize: Int, ports: Seq[T])
+case class TLSplitterNode(
+  clientFn:        SplitterArg[TLClientPortParameters]  => Seq[TLClientPortParameters],
+  managerFn:       SplitterArg[TLManagerPortParameters] => Seq[TLManagerPortParameters],
+  numClientPorts:  Range.Inclusive = 0 to 999,
+  numManagerPorts: Range.Inclusive = 0 to 999)(
+  implicit valName: ValName)
+  extends TLCustomNode(numClientPorts, numManagerPorts)
+{
+  def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
+    require (oKnown == 0, s"${name} (a splitter) appears right of a := or :*=; use a :=* instead${lazyModule.line}")
+    require (iStars == 0, s"${name} (a splitter) cannot appear left of a :*=; did you mean :=*?${lazyModule.line}")
+    (0, iKnown)
+  }
+  def mapParamsD(n: Int, p: Seq[TLClientPortParameters]): Seq[TLClientPortParameters] = {
+    require (p.size == 0 || n % p.size == 0, s"Diplomacy bug; splitter inputs do not divide outputs")
+    val out = clientFn(SplitterArg(n, p))
+    require (out.size == n, s"${name} created the wrong number of outputs from inputs${lazyModule.line}")
+    out
+  }
+  def mapParamsU(n: Int, p: Seq[TLManagerPortParameters]): Seq[TLManagerPortParameters] = {
+    require (n == 0 || p.size % n == 0, s"Diplomacy bug; splitter outputs indivisable by inputs")
+    val out = managerFn(SplitterArg(n, p))
+    require (out.size == n, s"${name} created the wrong number of inputs from outputs${lazyModule.line}")
+    out
+  }
+}
+
 class TLSplitter(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parameters) extends LazyModule
 {
   val node = TLSplitterNode(
@@ -34,18 +62,14 @@ class TLSplitter(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Pa
     })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
     def group[T](x: Seq[T]) =
-      if (x.isEmpty) Nil else x.grouped(node.edgesIn.size).toList.transpose
+      if (x.isEmpty) Nil else x.grouped(node.in.size).toList.transpose
 
-    if (node.edgesOut.size == node.edgesIn.size) {
-      io.out <> io.in
-    } else ((node.edgesIn zip io.in) zip (group(node.edgesOut) zip group(io.out))) foreach {
-      case ((edgeIn, io_in), (edgesOut, io_out)) =>
+    if (node.out.size == node.in.size) {
+      (node.in zip node.out) foreach { case ((i, _), (o, _)) => o <> i }
+    } else (node.in zip group(node.out)) foreach {
+      case ((io_in, edgeIn), seq) =>
+      val (io_out, edgesOut) = seq.unzip
 
       // Grab the port ID mapping
       val outputIdRanges = TLXbar.mapOutputIds(edgesOut.map(_.manager))

src/main/scala/tilelink/ToAHB.scala

@@ -11,7 +11,7 @@ import freechips.rocketchip.util._
 import scala.math.{min, max}
 import AHBParameters._
 
-case class TLToAHBNode() extends MixedAdapterNode(TLImp, AHBImp)(
+case class TLToAHBNode()(implicit valName: ValName) extends MixedAdapterNode(TLImp, AHBImp)(
   dFn = { case TLClientPortParameters(clients, unsafeAtomics, minLatency) =>
     val masters = clients.map { case c => AHBMasterParameters(name = c.name, nodePath = c.nodePath) }
     AHBMasterPortParameters(masters)
@@ -53,12 +53,7 @@ class TLToAHB(val aFlow: Boolean = false)(implicit p: Parameters) extends LazyMo
   val node = TLToAHBNode()
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-      val out = node.bundleOut
-    }
-
-   ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+   (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val beatBytes = edgeOut.slave.beatBytes
       val maxTransfer = edgeOut.slave.maxTransfer
       val lgMax = log2Ceil(maxTransfer)

src/main/scala/tilelink/ToAPB.scala

@@ -10,7 +10,7 @@ import freechips.rocketchip.amba.apb._
 import scala.math.{min, max}
 import APBParameters._
 
-case class TLToAPBNode() extends MixedAdapterNode(TLImp, APBImp)(
+case class TLToAPBNode()(implicit valName: ValName) extends MixedAdapterNode(TLImp, APBImp)(
   dFn = { case TLClientPortParameters(clients, unsafeAtomics, minLatency) =>
     val masters = clients.map { case c => APBMasterParameters(name = c.name, nodePath = c.nodePath) }
     APBMasterPortParameters(masters)
@@ -38,12 +38,7 @@ class TLToAPB(val aFlow: Boolean = true)(implicit p: Parameters) extends LazyMod
   val node = TLToAPBNode()
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val beatBytes = edgeOut.slave.beatBytes
       val lgBytes = log2Ceil(beatBytes)
 

src/main/scala/tilelink/ToAXI4.scala

@@ -10,7 +10,7 @@ import freechips.rocketchip.util._
 import freechips.rocketchip.amba.axi4._
 import scala.math.{min, max}
 
-case class TLToAXI4Node(beatBytes: Int, stripBits: Int = 0) extends MixedAdapterNode(TLImp, AXI4Imp)(
+case class TLToAXI4Node(beatBytes: Int, stripBits: Int = 0)(implicit valName: ValName) extends MixedAdapterNode(TLImp, AXI4Imp)(
   dFn = { p =>
     p.clients.foreach { c =>
       require (c.sourceId.start % (1 << stripBits) == 0 &&
@@ -53,12 +53,7 @@ class TLToAXI4(val beatBytes: Int, val combinational: Boolean = true, val adapte
   val node = TLToAXI4Node(beatBytes, stripBits)
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in = node.bundleIn
-      val out = node.bundleOut
-    }
-
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
       val slaves  = edgeOut.slave.slaves
 
       // All pairs of slaves must promise that they will never interleave data

src/main/scala/tilelink/WidthWidget.scala

@@ -17,11 +17,6 @@ class TLWidthWidget(innerBeatBytes: Int)(implicit p: Parameters) extends LazyMod
     managerFn = { case m => m.copy(beatBytes = innerBeatBytes) })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
-    }
-
     def merge[T <: TLDataChannel](edgeIn: TLEdge, in: DecoupledIO[T], edgeOut: TLEdge, out: DecoupledIO[T]) = {
       val inBytes = edgeIn.manager.beatBytes
       val outBytes = edgeOut.manager.beatBytes
@@ -154,7 +149,7 @@ class TLWidthWidget(innerBeatBytes: Int)(implicit p: Parameters) extends LazyMod
       }
     }
 
-    ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
+    (node.in zip node.out) foreach { case ((in, edgeIn), (out, edgeOut)) =>
 
       // If the master is narrower than the slave, the D channel must be narrowed.
       // This is tricky, because the D channel has no address data.
@@ -220,7 +215,7 @@ class TLRAMWidthWidget(first: Int, second: Int, txns: Int)(implicit p: Parameter
                   TLWidthWidget(second)(
                     TLWidthWidget(first)(TLDelayer(0.1)(model.node)))}))
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }

src/main/scala/tilelink/Xbar.scala

@@ -63,22 +63,21 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
     })
 
   lazy val module = new LazyModuleImp(this) {
-    val io = new Bundle {
-      val in  = node.bundleIn
-      val out = node.bundleOut
+    if ((node.in.size * node.out.size) > (8*32)) {
+      println (s"!!! WARNING !!!")
+      println (s" Your TLXbar ($name) is very large, with ${node.in.size} Masters and ${node.out.size} Slaves.")
+      println (s"!!! WARNING !!!")
     }
 
-    if ((io.in.size * io.out.size) > (8*32)) {
-      println (s"!!! WARNING !!!")
-      println (s" Your TLXbar ($name) is very large, with ${io.in.size} Masters and ${io.out.size} Slaves.")
-      println (s"!!! WARNING !!!")
-    }
+    val (io_in, edgesIn) = node.in.unzip
+    val (io_out, edgesOut) = node.out.unzip
+
     // Grab the port ID mapping
-    val inputIdRanges = TLXbar.mapInputIds(node.edgesIn.map(_.client))
-    val outputIdRanges = TLXbar.mapOutputIds(node.edgesOut.map(_.manager))
+    val inputIdRanges = TLXbar.mapInputIds(edgesIn.map(_.client))
+    val outputIdRanges = TLXbar.mapOutputIds(edgesOut.map(_.manager))
 
     // Find a good mask for address decoding
-    val port_addrs = node.edgesOut.map(_.manager.managers.map(_.address).flatten)
+    val port_addrs = edgesOut.map(_.manager.managers.map(_.address).flatten)
     val routingMask = AddressDecoder(port_addrs)
     val route_addrs = port_addrs.map(seq => AddressSet.unify(seq.map(_.widen(~routingMask)).distinct))
     val outputPorts = route_addrs.map(seq => (addr: UInt) => seq.map(_.contains(addr)).reduce(_ || _))
@@ -97,70 +96,70 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
     // Print the ID mapping
     if (false) {
       println(s"XBar ${name} mapping:")
-      (node.edgesIn zip inputIdRanges).zipWithIndex.foreach { case ((edge, id), i) =>
+      (edgesIn zip inputIdRanges).zipWithIndex.foreach { case ((edge, id), i) =>
         println(s"\t$i assigned ${id} for ${edge.client.clients.map(_.name).mkString(", ")}")
       }
       println("")
     }
 
     // We need an intermediate size of bundle with the widest possible identifiers
-    val wide_bundle = TLBundleParameters.union(io.in.map(_.params) ++ io.out.map(_.params))
+    val wide_bundle = TLBundleParameters.union(io_in.map(_.params) ++ io_out.map(_.params))
 
     // Handle size = 1 gracefully (Chisel3 empty range is broken)
     def trim(id: UInt, size: Int) = if (size <= 1) UInt(0) else id(log2Ceil(size)-1, 0)
 
     // Transform input bundle sources (sinks use global namespace on both sides)
-    val in = Wire(Vec(io.in.size, TLBundle(wide_bundle)))
+    val in = Wire(Vec(io_in.size, TLBundle(wide_bundle)))
     for (i <- 0 until in.size) {
       val r = inputIdRanges(i)
 
-      in(i).a <> io.in(i).a
-      io.in(i).d <> in(i).d
-      in(i).a.bits.source := io.in(i).a.bits.source | UInt(r.start)
-      io.in(i).d.bits.source := trim(in(i).d.bits.source, r.size)
+      in(i).a <> io_in(i).a
+      io_in(i).d <> in(i).d
+      in(i).a.bits.source := io_in(i).a.bits.source | UInt(r.start)
+      io_in(i).d.bits.source := trim(in(i).d.bits.source, r.size)
 
-      if (node.edgesIn(i).client.anySupportProbe && node.edgesOut.exists(_.manager.anySupportAcquireB)) {
-        in(i).c <> io.in(i).c
-        in(i).e <> io.in(i).e
-        io.in(i).b <> in(i).b
-        in(i).c.bits.source := io.in(i).c.bits.source | UInt(r.start)
-        io.in(i).b.bits.source := trim(in(i).b.bits.source, r.size)
+      if (edgesIn(i).client.anySupportProbe && edgesOut.exists(_.manager.anySupportAcquireB)) {
+        in(i).c <> io_in(i).c
+        in(i).e <> io_in(i).e
+        io_in(i).b <> in(i).b
+        in(i).c.bits.source := io_in(i).c.bits.source | UInt(r.start)
+        io_in(i).b.bits.source := trim(in(i).b.bits.source, r.size)
       } else {
         in(i).c.valid := Bool(false)
         in(i).e.valid := Bool(false)
         in(i).b.ready := Bool(false)
-        io.in(i).c.ready := Bool(true)
-        io.in(i).e.ready := Bool(true)
-        io.in(i).b.valid := Bool(false)
+        io_in(i).c.ready := Bool(true)
+        io_in(i).e.ready := Bool(true)
+        io_in(i).b.valid := Bool(false)
       }
     }
 
     // Transform output bundle sinks (sources use global namespace on both sides)
-    val out = Wire(Vec(io.out.size, TLBundle(wide_bundle)))
+    val out = Wire(Vec(io_out.size, TLBundle(wide_bundle)))
     for (i <- 0 until out.size) {
       val r = outputIdRanges(i)
 
-      io.out(i).a <> out(i).a
-      out(i).d <> io.out(i).d
-      out(i).d.bits.sink := io.out(i).d.bits.sink | UInt(r.map(_.start).getOrElse(0))
+      io_out(i).a <> out(i).a
+      out(i).d <> io_out(i).d
+      out(i).d.bits.sink := io_out(i).d.bits.sink | UInt(r.map(_.start).getOrElse(0))
 
-      if (node.edgesOut(i).manager.anySupportAcquireB && node.edgesIn.exists(_.client.anySupportProbe)) {
-        io.out(i).c <> out(i).c
-        io.out(i).e <> out(i).e
-        out(i).b <> io.out(i).b
-        io.out(i).e.bits.sink := trim(out(i).e.bits.sink, r.map(_.size).getOrElse(0))
+      if (edgesOut(i).manager.anySupportAcquireB && edgesIn.exists(_.client.anySupportProbe)) {
+        io_out(i).c <> out(i).c
+        io_out(i).e <> out(i).e
+        out(i).b <> io_out(i).b
+        io_out(i).e.bits.sink := trim(out(i).e.bits.sink, r.map(_.size).getOrElse(0))
       } else {
         out(i).c.ready := Bool(false)
         out(i).e.ready := Bool(false)
         out(i).b.valid := Bool(false)
-        io.out(i).c.valid := Bool(false)
-        io.out(i).e.valid := Bool(false)
-        io.out(i).b.ready := Bool(true)
+        io_out(i).c.valid := Bool(false)
+        io_out(i).e.valid := Bool(false)
+        io_out(i).b.ready := Bool(true)
       }
     }
 
-    val addressA = (in zip node.edgesIn) map { case (i, e) => e.address(i.a.bits) }
-    val addressC = (in zip node.edgesIn) map { case (i, e) => e.address(i.c.bits) }
+    val addressA = (in zip edgesIn) map { case (i, e) => e.address(i.a.bits) }
+    val addressC = (in zip edgesIn) map { case (i, e) => e.address(i.c.bits) }
 
     val requestAIO = Vec(addressA.map { i => Vec(outputPorts.map { o => o(i) }) })
     val requestCIO = Vec(addressC.map { i => Vec(outputPorts.map { o => o(i) }) })
@@ -168,28 +167,28 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
     val requestDOI = Vec(out.map { o => Vec(inputIdRanges.map  { i => i.contains(o.d.bits.source) }) })
     val requestEIO = Vec(in.map  { i => Vec(outputIdRanges.map { o => o.map(_.contains(i.e.bits.sink)).getOrElse(Bool(false)) }) })
 
-    val beatsAI = Vec((in  zip node.edgesIn)  map { case (i, e) => e.numBeats1(i.a.bits) })
-    val beatsBO = Vec((out zip node.edgesOut) map { case (o, e) => e.numBeats1(o.b.bits) })
-    val beatsCI = Vec((in  zip node.edgesIn)  map { case (i, e) => e.numBeats1(i.c.bits) })
-    val beatsDO = Vec((out zip node.edgesOut) map { case (o, e) => e.numBeats1(o.d.bits) })
-    val beatsEI = Vec((in  zip node.edgesIn)  map { case (i, e) => e.numBeats1(i.e.bits) })
+    val beatsAI = Vec((in  zip edgesIn)  map { case (i, e) => e.numBeats1(i.a.bits) })
+    val beatsBO = Vec((out zip edgesOut) map { case (o, e) => e.numBeats1(o.b.bits) })
+    val beatsCI = Vec((in  zip edgesIn)  map { case (i, e) => e.numBeats1(i.c.bits) })
+    val beatsDO = Vec((out zip edgesOut) map { case (o, e) => e.numBeats1(o.d.bits) })
+    val beatsEI = Vec((in  zip edgesIn)  map { case (i, e) => e.numBeats1(i.e.bits) })
 
     // Which pairs support support transfers
     def transpose[T](x: Seq[Seq[T]]) = Seq.tabulate(x(0).size) { i => Seq.tabulate(x.size) { j => x(j)(i) } }
     def filter[T](data: Seq[T], mask: Seq[Boolean]) = (data zip mask).filter(_._2).map(_._1)
 
     // Fanout the input sources to the output sinks
-    val portsAOI = transpose((in  zip requestAIO) map { case (i, r) => TLXbar.fanout(i.a, r, node.paramsOut.map(_(ForceFanoutKey).a)) })
-    val portsBIO = transpose((out zip requestBOI) map { case (o, r) => TLXbar.fanout(o.b, r, node.paramsIn .map(_(ForceFanoutKey).b)) })
-    val portsCOI = transpose((in  zip requestCIO) map { case (i, r) => TLXbar.fanout(i.c, r, node.paramsOut.map(_(ForceFanoutKey).c)) })
-    val portsDIO = transpose((out zip requestDOI) map { case (o, r) => TLXbar.fanout(o.d, r, node.paramsIn .map(_(ForceFanoutKey).d)) })
-    val portsEOI = transpose((in  zip requestEIO) map { case (i, r) => TLXbar.fanout(i.e, r, node.paramsOut.map(_(ForceFanoutKey).e)) })
+    val portsAOI = transpose((in  zip requestAIO) map { case (i, r) => TLXbar.fanout(i.a, r, edgesOut.map(_.params(ForceFanoutKey).a)) })
+    val portsBIO = transpose((out zip requestBOI) map { case (o, r) => TLXbar.fanout(o.b, r, edgesIn .map(_.params(ForceFanoutKey).b)) })
+    val portsCOI = transpose((in  zip requestCIO) map { case (i, r) => TLXbar.fanout(i.c, r, edgesOut.map(_.params(ForceFanoutKey).c)) })
+    val portsDIO = transpose((out zip requestDOI) map { case (o, r) => TLXbar.fanout(o.d, r, edgesIn .map(_.params(ForceFanoutKey).d)) })
+    val portsEOI = transpose((in  zip requestEIO) map { case (i, r) => TLXbar.fanout(i.e, r, edgesOut.map(_.params(ForceFanoutKey).e)) })
 
     // Arbitrate amongst the sources
     for (o <- 0 until out.size) {
       val allowI = Seq.tabulate(in.size) { i =>
-        node.edgesIn(i).client.anySupportProbe &&
-        node.edgesOut(o).manager.anySupportAcquireB
+        edgesIn(i).client.anySupportProbe &&
+        edgesOut(o).manager.anySupportAcquireB
       }
       TLArbiter(policy)(out(o).a,       (beatsAI zip portsAOI(o)        ):_*)
       TLArbiter(policy)(out(o).c, filter(beatsCI zip portsCOI(o), allowI):_*)
@@ -198,8 +197,8 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
 
     for (i <- 0 until in.size) {
       val allowO = Seq.tabulate(out.size) { o =>
-        node.edgesIn(i).client.anySupportProbe &&
-        node.edgesOut(o).manager.anySupportAcquireB
+        edgesIn(i).client.anySupportProbe &&
+        edgesOut(o).manager.anySupportAcquireB
       }
       TLArbiter(policy)(in(i).b, filter(beatsBO zip portsBIO(i), allowO):_*)
       TLArbiter(policy)(in(i).d,       (beatsDO zip portsDIO(i)        ):_*)
@@ -264,7 +263,7 @@ class TLRAMXbar(nManagers: Int, txns: Int)(implicit p: Parameters) extends LazyM
     ram.node := TLFragmenter(4, 256)(TLDelayer(0.1)(xbar.node))
   }
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzz.module.io.finished
   }
 }
@@ -287,7 +286,7 @@ class TLMulticlientXbar(nManagers: Int, nClients: Int, txns: Int)(implicit p: Pa
     ram.node := TLFragmenter(4, 256)(TLDelayer(0.1)(xbar.node))
   }
 
-  lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
+  lazy val module = new LazyModuleImp(this) with UnitTestModule {
     io.finished := fuzzers.last.module.io.finished
   }
 }

src/main/scala/unittest/UnitTest.scala

@@ -3,17 +3,28 @@
 package freechips.rocketchip.unittest
 
 import Chisel._
+import chisel3.experimental.MultiIOModule
 import freechips.rocketchip.config._
 import freechips.rocketchip.util.SimpleTimer
 
-trait HasUnitTestIO {
-  val io = new Bundle {
+trait UnitTestIO {
   val finished = Bool(OUTPUT)
   val start = Bool(INPUT)
 }
+
+trait HasUnitTestIO {
+  val io: UnitTestIO
 }
 
-abstract class UnitTest(val timeout: Int = 4096) extends Module with HasUnitTestIO {
+trait UnitTestLegacyModule extends HasUnitTestIO {
+  val io = new Bundle with UnitTestIO
+}
+
+trait UnitTestModule extends MultiIOModule with HasUnitTestIO {
+  val io = IO(new Bundle with UnitTestIO)
+}
+
+abstract class UnitTest(val timeout: Int = 4096) extends Module with UnitTestLegacyModule {
   val testName = this.getClass.getSimpleName
 
   when (io.start) { printf(s"Started UnitTest $testName\n") }

src/main/scala/util/HeterogeneousBag.scala

@@ -17,3 +17,8 @@ final case class HeterogeneousBag[T <: Data](elts: Seq[T]) extends Record with c
   override def hashCode: Int = super[Record].hashCode
   override def equals(that: Any): Boolean = super[Record].equals(that)
 }
+
+object HeterogeneousBag
+{
+  def fromNode[D <: Data, E](elts: Seq[(D, E)]) = new HeterogeneousBag(elts.map(_._1.cloneType))
+}