diff --git a/src/main/scala/coreplex/BaseCoreplex.scala b/src/main/scala/coreplex/BaseCoreplex.scala
index 22d26ca1..89b0cb99 100644
--- a/src/main/scala/coreplex/BaseCoreplex.scala
+++ b/src/main/scala/coreplex/BaseCoreplex.scala
@@ -23,7 +23,7 @@ case object BankIdLSB extends Field[Int]
 /** Function for building some kind of coherence manager agent */
 case object BuildL2CoherenceManager extends Field[(Int, Parameters) => CoherenceAgent]
 /** Function for building some kind of tile connected to a reset signal */
-case object BuildTiles extends Field[Seq[(Bool, Parameters) => Tile]]
+case object BuildTiles extends Field[Seq[Parameters => LazyTile]]
 /** The file to read the BootROM contents from */
 case object BootROMFile extends Field[String]
 
@@ -35,50 +35,137 @@ trait HasCoreplexParameters {
   lazy val outerMemParams = p.alterPartial({ case TLId => "L2toMC" })
   lazy val outerMMIOParams = p.alterPartial({ case TLId => "L2toMMIO" })
   lazy val globalAddrMap = p(rocketchip.GlobalAddrMap)
+  lazy val nTiles = p(uncore.devices.NTiles)
+  lazy val nSlaves = p(rocketchip.NCoreplexExtClients)
+  lazy val nMemChannels = p(NMemoryChannels)
+  lazy val hasSupervisor = p(rocket.UseVM)
 }
 
-case class CoreplexConfig(
-    nTiles: Int,
-    nExtInterrupts: Int,
-    nSlaves: Int,
-    nMemChannels: Int,
-    hasSupervisor: Boolean)
-{
-  val nInterruptPriorities = if (nExtInterrupts <= 1) 0 else (nExtInterrupts min 7)
-  val plicKey = PLICConfig(nTiles, hasSupervisor, nExtInterrupts, nInterruptPriorities)
+case class CoreplexParameters(implicit val p: Parameters) extends HasCoreplexParameters
+
+abstract class BareCoreplex(implicit val p: Parameters) extends LazyModule
+abstract class BareCoreplexBundle[+L <: BareCoreplex](_outer: L) extends Bundle {
+  val outer = _outer
+}
+abstract class BareCoreplexModule[+L <: BareCoreplex, +B <: BareCoreplexBundle[L]](_outer: L, _io: () => B) extends LazyModuleImp(_outer) {
+  val outer = _outer
+  val io = _io ()
 }
 
-abstract class BaseCoreplex(c: CoreplexConfig)(implicit p: Parameters) extends LazyModule
+trait CoreplexNetwork extends HasCoreplexParameters {
+  this: BareCoreplex =>
 
-abstract class BaseCoreplexBundle(val c: CoreplexConfig)(implicit val p: Parameters) extends Bundle with HasCoreplexParameters {
-  val master = new Bundle {
-    val mem = Vec(c.nMemChannels, new ClientUncachedTileLinkIO()(outerMemParams))
-    val mmio = new ClientUncachedTileLinkIO()(outerMMIOParams)
-  }
-  val slave = Vec(c.nSlaves, new ClientUncachedTileLinkIO()(innerParams)).flip
-  val interrupts = Vec(c.nExtInterrupts, Bool()).asInput
-  val debug = new DebugBusIO()(p).flip
-  val clint = Vec(c.nTiles, new CoreplexLocalInterrupts).asInput
-  val resetVector = UInt(INPUT, p(XLen))
-  val success = Bool(OUTPUT) // used for testing
+  val l1tol2 = LazyModule(new TLXbar)
+  val l1tol2_beatBytes = p(TLKey("L2toMMIO")).dataBitsPerBeat/8
+  val l1tol2_lineBytes = p(CacheBlockBytes)
 
-  override def cloneType = this.getClass.getConstructors.head.newInstance(c, p).asInstanceOf[this.type]
+  val cbus = LazyModule(new TLXbar)
+  val cbus_beatBytes = p(XLen)/8
+  val cbus_lineBytes = l1tol2_lineBytes
+
+  val mmio = TLOutputNode()
+  val mmioInt = IntInputNode()
+
+  cbus.node :=
+    TLAtomicAutomata(arithmetic = true)( // disable once TLB uses TL2 metadata
+    TLWidthWidget(l1tol2_beatBytes)(
+    TLBuffer()(
+    l1tol2.node)))
+
+  mmio :=
+    TLBuffer()(
+    TLWidthWidget(l1tol2_beatBytes)(
+    l1tol2.node))
 }
 
-abstract class BaseCoreplexModule[+L <: BaseCoreplex, +B <: BaseCoreplexBundle](
-    c: CoreplexConfig, l: L, b: => B)(implicit val p: Parameters) extends LazyModuleImp(l) with HasCoreplexParameters {
-  val outer: L = l
-  val io: B = b
+trait CoreplexNetworkBundle extends HasCoreplexParameters {
+  this: {
+    val outer: CoreplexNetwork
+  } =>
+
+  implicit val p = outer.p
+  val mmio = outer.mmio.bundleOut
+  val interrupts = outer.mmioInt.bundleIn
+}
+
+trait CoreplexNetworkModule extends HasCoreplexParameters {
+    this: BareCoreplexModule[BareCoreplex, BareCoreplexBundle[BareCoreplex]] =>
+  implicit val p = outer.p
+}
+
+trait CoreplexRISCVPlatform {
+    this: CoreplexNetwork =>
 
   // Build a set of Tiles
-  val tiles = p(BuildTiles) map { _(reset, p) }
+  val lazyTiles = p(BuildTiles) map { _(p) }
+  val legacy = LazyModule(new TLLegacy()(outerMMIOParams))
+  val tileIntNodes = lazyTiles.map { _ => IntInternalOutputNode() } // this should be moved into the Tile...
+
+  val debug = LazyModule(new TLDebugModule())
+  val plic  = LazyModule(new TLPLIC(hasSupervisor, maxPriorities = 7))
+  val clint = LazyModule(new CoreplexLocalInterrupter)
+
+  // Kill this once we move TL2 into rocket
+  l1tol2.node :=
+    TLHintHandler()(
+    legacy.node)
+
+  debug.node := TLFragmenter(cbus_beatBytes, cbus_lineBytes)(cbus.node)
+  plic.node  := TLFragmenter(cbus_beatBytes, cbus_lineBytes)(cbus.node)
+  clint.node := TLFragmenter(cbus_beatBytes, cbus_lineBytes)(cbus.node)
+
+  plic.intnode := mmioInt
+  tileIntNodes.foreach { _ := plic.intnode }
+}
+
+trait CoreplexRISCVPlatformBundle {
+  this: CoreplexNetworkBundle {
+    val outer: CoreplexRISCVPlatform
+  } =>
+
+  val mem = Vec(nMemChannels, new ClientUncachedTileLinkIO()(outerMemParams))
+  val slave = Vec(nSlaves, new ClientUncachedTileLinkIO()(innerParams)).flip
+  val debug = new DebugBusIO().flip
+  val rtcTick = Bool(INPUT)
+  val resetVector = UInt(INPUT, p(XLen))
+  val success = Bool(OUTPUT) // used for testing
+}
+
+trait CoreplexRISCVPlatformModule {
+  this: CoreplexNetworkModule {
+    val outer: CoreplexNetwork with CoreplexRISCVPlatform
+    val io: CoreplexRISCVPlatformBundle
+  } =>
+
+  val tiles = outer.lazyTiles.map(_.module)
   val uncoreTileIOs = (tiles zipWithIndex) map { case (tile, i) => Wire(tile.io) }
 
+  println("\nGenerated Address Map")
+  for (entry <- p(rocketchip.GlobalAddrMap).flatten) {
+    val name = entry.name
+    val start = entry.region.start
+    val end = entry.region.start + entry.region.size - 1
+    val prot = entry.region.attr.prot
+    val protStr = (if ((prot & AddrMapProt.R) > 0) "R" else "") +
+                  (if ((prot & AddrMapProt.W) > 0) "W" else "") +
+                  (if ((prot & AddrMapProt.X) > 0) "X" else "")
+    val cacheable = if (entry.region.attr.cacheable) " [C]" else ""
+    println(f"\t$name%s $start%x - $end%x, $protStr$cacheable")
+  }
+
+  // Create and export the ConfigString
+  val managers = outer.l1tol2.node.edgesIn(0).manager.managers
+  val configString = rocketchip.GenerateConfigString(p, outer.clint, outer.plic, managers)
+  // Allow something else to have override the config string
+  if (!ConfigStringOutput.contents.isDefined) {
+    ConfigStringOutput.contents = Some(configString)
+  }
+  println(s"\nGenerated Configuration String\n${ConfigStringOutput.contents.get}")
+
   val nCachedPorts = tiles.map(tile => tile.io.cached.size).reduce(_ + _)
   val nUncachedPorts = tiles.map(tile => tile.io.uncached.size).reduce(_ + _)
-  val nBanks = c.nMemChannels * nBanksPerMemChannel
+  val nBanks = nMemChannels * nBanksPerMemChannel
 
-  // Build an uncore backing the Tiles
   buildUncore(p.alterPartial({
     case HastiId => "TL"
     case TLId => "L1toL2"
@@ -86,7 +173,7 @@ abstract class BaseCoreplexModule[+L <: BaseCoreplex, +B <: BaseCoreplexBundle](
     case NUncachedTileLinkPorts => nUncachedPorts
   }))
 
-  def buildUncore(implicit p: Parameters) = {
+  def buildUncore(implicit p: Parameters) {
     // Create a simple L1toL2 NoC between the tiles and the banks of outer memory
     // Cached ports are first in client list, making sharerToClientId just an indentity function
     // addrToBank is sed to hash physical addresses (of cache blocks) to banks (and thereby memory channels)
@@ -112,8 +199,9 @@ abstract class BaseCoreplexModule[+L <: BaseCoreplex, +B <: BaseCoreplexBundle](
     l1tol2net.io.clients_cached <> uncoreTileIOs.map(_.cached).flatten
     l1tol2net.io.clients_uncached <> uncoreTileIOs.map(_.uncached).flatten ++ io.slave
     l1tol2net.io.managers <> managerEndpoints.map(_.innerTL) :+ mmioManager.io.inner
+    outer.legacy.module.io.legacy <>  mmioManager.io.outer 
 
-    val mem_ic = Module(new TileLinkMemoryInterconnect(nBanksPerMemChannel, c.nMemChannels)(outerMemParams))
+    val mem_ic = Module(new TileLinkMemoryInterconnect(nBanksPerMemChannel, nMemChannels)(outerMemParams))
 
     val backendBuffering = TileLinkDepths(0,0,0,0,0)
     for ((bank, icPort) <- managerEndpoints zip mem_ic.io.in) {
@@ -121,46 +209,36 @@ abstract class BaseCoreplexModule[+L <: BaseCoreplex, +B <: BaseCoreplexBundle](
       icPort <> TileLinkIOUnwrapper(enqueued)
     }
 
-    io.master.mem <> mem_ic.io.out
-
-    buildMMIONetwork(TileLinkEnqueuer(mmioManager.io.outer, 1))(outerMMIOParams)
+    io.mem <> mem_ic.io.out
   }
 
-  def buildMMIONetwork(mmio: ClientUncachedTileLinkIO)(implicit p: Parameters) = {
-    val ioAddrMap = globalAddrMap.subMap("io")
-
-    val cBus = Module(new TileLinkRecursiveInterconnect(1, ioAddrMap))
-    cBus.io.in.head <> mmio
-
-    val plic = Module(new PLIC(c.plicKey))
-    plic.io.tl <> cBus.port("cbus:plic")
-    for (i <- 0 until io.interrupts.size) {
-      val gateway = Module(new LevelGateway)
-      gateway.io.interrupt := io.interrupts(i)
-      plic.io.devices(i) <> gateway.io.plic
-    }
-
-    val debugModule = Module(new DebugModule)
-    debugModule.io.tl <> cBus.port("cbus:debug")
-    debugModule.io.db <> io.debug
-
-    // connect coreplex-internal interrupts to tiles
-    for ((tile, i) <- (uncoreTileIOs zipWithIndex)) {
-      tile.interrupts <> io.clint(i)
-      tile.interrupts.meip := plic.io.harts(plic.cfg.context(i, 'M'))
-      tile.interrupts.seip.foreach(_ := plic.io.harts(plic.cfg.context(i, 'S')))
-      tile.interrupts.debug := debugModule.io.debugInterrupts(i)
-      tile.hartid := UInt(i)
-      tile.resetVector := io.resetVector
-    }
-
-    val tileSlavePorts = (0 until c.nTiles) map (i => s"cbus:dmem$i") filter (ioAddrMap contains _)
-    for ((t, m) <- (uncoreTileIOs.map(_.slave).flatten) zip (tileSlavePorts map (cBus port _)))
-      t <> m
-
-    io.master.mmio <> cBus.port("pbus")
+  // connect coreplex-internal interrupts to tiles
+  for ((tile, i) <- (uncoreTileIOs zipWithIndex)) {
+    tile.hartid := UInt(i)
+    tile.resetVector := io.resetVector
+    tile.interrupts := outer.clint.module.io.tiles(i)
+    tile.interrupts.debug := outer.debug.module.io.debugInterrupts(i)
+    tile.interrupts.meip := outer.tileIntNodes(i).bundleOut(0)(0)
+    tile.interrupts.seip.foreach(_ := outer.tileIntNodes(i).bundleOut(0)(1))
   }
 
+  outer.debug.module.io.db <> io.debug
+  outer.clint.module.io.rtcTick := io.rtcTick
+
   // Coreplex doesn't know when to stop running
   io.success := Bool(false)
 }
+
+class BaseCoreplex(implicit p: Parameters) extends BareCoreplex
+    with CoreplexNetwork
+    with CoreplexRISCVPlatform {
+  override lazy val module = new BaseCoreplexModule(this, () => new BaseCoreplexBundle(this))
+}
+
+class BaseCoreplexBundle[+L <: BaseCoreplex](_outer: L) extends BareCoreplexBundle(_outer)
+    with CoreplexNetworkBundle
+    with CoreplexRISCVPlatformBundle
+
+class BaseCoreplexModule[+L <: BaseCoreplex, +B <: BaseCoreplexBundle[L]](_outer: L, _io: () => B) extends BareCoreplexModule(_outer, _io)
+    with CoreplexNetworkModule
+    with CoreplexRISCVPlatformModule
diff --git a/src/main/scala/coreplex/Configs.scala b/src/main/scala/coreplex/Configs.scala
index 4fea63f1..baff8e7b 100644
--- a/src/main/scala/coreplex/Configs.scala
+++ b/src/main/scala/coreplex/Configs.scala
@@ -4,6 +4,7 @@ package coreplex
 
 import Chisel._
 import junctions._
+import diplomacy._
 import uncore.tilelink._
 import uncore.coherence._
 import uncore.agents._
@@ -69,8 +70,8 @@ class BaseCoreplexConfig extends Config (
       case NUncachedTileLinkPorts => 1
       //Tile Constants
       case BuildTiles => {
-        List.tabulate(site(NTiles)){ i => (r: Bool, p: Parameters) =>
-          Module(new RocketTile(resetSignal = r)(p.alterPartial({
+        List.tabulate(site(NTiles)){ i => (p: Parameters) =>
+          LazyModule(new RocketTile()(p.alterPartial({
             case TileId => i
             case TLId => "L1toL2"
             case NUncachedTileLinkPorts => 1 + site(RoccNMemChannels)
diff --git a/src/main/scala/coreplex/Coreplex.scala b/src/main/scala/coreplex/Coreplex.scala
index bbd39367..ad8952d2 100644
--- a/src/main/scala/coreplex/Coreplex.scala
+++ b/src/main/scala/coreplex/Coreplex.scala
@@ -3,14 +3,20 @@ package coreplex
 import Chisel._
 import cde.{Parameters, Field}
 import junctions._
+import diplomacy._
 import uncore.tilelink._
+import uncore.tilelink2._
 import uncore.util._
 import util._
 import rocket._
 
 trait DirectConnection {
-  val tiles: Seq[Tile]
-  val uncoreTileIOs: Seq[TileIO]
+    this: CoreplexNetwork with CoreplexRISCVPlatform =>
+  lazyTiles.map(_.slave).flatten.foreach { scratch => scratch := cbus.node }
+}
+
+trait DirectConnectionModule {
+    this: CoreplexNetworkModule with CoreplexRISCVPlatformModule =>
 
   val tlBuffering = TileLinkDepths(1,1,2,2,0)
   val ultBuffering = UncachedTileLinkDepths(1,2)
@@ -18,7 +24,6 @@ trait DirectConnection {
   (tiles zip uncoreTileIOs) foreach { case (tile, uncore) =>
     (uncore.cached zip tile.io.cached) foreach { case (u, t) => u <> TileLinkEnqueuer(t, tlBuffering) }
     (uncore.uncached zip tile.io.uncached) foreach { case (u, t) => u <> TileLinkEnqueuer(t, ultBuffering) }
-    tile.io.slave.foreach { _ <> TileLinkEnqueuer(uncore.slave.get, 1) }
 
     tile.io.interrupts <> uncore.interrupts
 
@@ -27,30 +32,54 @@ trait DirectConnection {
   }
 }
 
-class DefaultCoreplex(c: CoreplexConfig)(implicit p: Parameters) extends BaseCoreplex(c)(p) {
-  override lazy val module = Module(new DefaultCoreplexModule(c, this, new DefaultCoreplexBundle(c)(p))(p))
+class DefaultCoreplex(implicit p: Parameters) extends BaseCoreplex
+    with DirectConnection {
+  override lazy val module = new DefaultCoreplexModule(this, () => new DefaultCoreplexBundle(this))
 }
 
-class DefaultCoreplexBundle(c: CoreplexConfig)(implicit p: Parameters) extends BaseCoreplexBundle(c)(p)
+class DefaultCoreplexBundle[+L <: DefaultCoreplex](_outer: L) extends BaseCoreplexBundle(_outer)
 
-class DefaultCoreplexModule[+L <: DefaultCoreplex, +B <: DefaultCoreplexBundle](
-    c: CoreplexConfig, l: L, b: => B)(implicit p: Parameters) extends BaseCoreplexModule(c, l, b)(p)
-    with DirectConnection
+class DefaultCoreplexModule[+L <: DefaultCoreplex, +B <: DefaultCoreplexBundle[L]](_outer: L, _io: () => B) extends BaseCoreplexModule(_outer, _io)
+    with DirectConnectionModule
 
 /////
 
-trait TileClockResetBundle {
-  val c: CoreplexConfig
-  val tcrs = Vec(c.nTiles, new Bundle {
+trait AsyncConnection {
+    this: CoreplexNetwork with CoreplexRISCVPlatform =>
+  val crossings = lazyTiles.map(_.slave).map(_.map { scratch =>
+    val crossing = LazyModule(new TLAsyncCrossing)
+    crossing.node := cbus.node
+    val monitor = (scratch := crossing.node)
+    (crossing, monitor)
+  })
+}
+
+trait AsyncConnectionBundle {
+    this: CoreplexNetworkBundle with CoreplexRISCVPlatformBundle =>
+  val tcrs = Vec(nTiles, new Bundle {
     val clock = Clock(INPUT)
     val reset = Bool(INPUT)
   })
 }
 
-trait AsyncConnection {
-  val io: TileClockResetBundle
-  val tiles: Seq[Tile]
-  val uncoreTileIOs: Seq[TileIO]
+trait AsyncConnectionModule {
+  this: Module with CoreplexNetworkModule with CoreplexRISCVPlatformModule {
+    val outer: AsyncConnection
+    val io: AsyncConnectionBundle
+  } =>
+
+  (outer.crossings zip io.tcrs) foreach { case (slaves, tcr) =>
+    slaves.foreach { case (crossing, monitor) =>
+      crossing.module.io.in_clock  := clock
+      crossing.module.io.in_reset  := reset
+      crossing.module.io.out_clock := tcr.clock
+      crossing.module.io.out_reset := tcr.reset
+      monitor.foreach { m =>
+        m.module.clock := tcr.clock
+        m.module.reset := tcr.reset
+      }
+    }
+  }
 
   (tiles, uncoreTileIOs, io.tcrs).zipped foreach { case (tile, uncore, tcr) =>
     tile.clock := tcr.clock
@@ -58,7 +87,6 @@ trait AsyncConnection {
 
     (uncore.cached zip tile.io.cached) foreach { case (u, t) => u <> AsyncTileLinkFrom(tcr.clock, tcr.reset, t) }
     (uncore.uncached zip tile.io.uncached) foreach { case (u, t) => u <> AsyncUTileLinkFrom(tcr.clock, tcr.reset, t) }
-    tile.io.slave.foreach { _ <> AsyncUTileLinkTo(tcr.clock, tcr.reset, uncore.slave.get)}
 
     val ti = tile.io.interrupts
     val ui = uncore.interrupts
@@ -73,13 +101,13 @@ trait AsyncConnection {
   }
 }
 
-class MultiClockCoreplex(c: CoreplexConfig)(implicit p: Parameters) extends BaseCoreplex(c)(p) {
-  override lazy val module = Module(new MultiClockCoreplexModule(c, this, new MultiClockCoreplexBundle(c)(p))(p))
+class MultiClockCoreplex(implicit p: Parameters) extends BaseCoreplex
+    with AsyncConnection {
+  override lazy val module = new MultiClockCoreplexModule(this, () => new MultiClockCoreplexBundle(this))
 }
 
-class MultiClockCoreplexBundle(c: CoreplexConfig)(implicit p: Parameters) extends BaseCoreplexBundle(c)(p)
-    with TileClockResetBundle
+class MultiClockCoreplexBundle[+L <: MultiClockCoreplex](_outer: L) extends BaseCoreplexBundle(_outer)
+    with AsyncConnectionBundle
 
-class MultiClockCoreplexModule[+L <: MultiClockCoreplex, +B <: MultiClockCoreplexBundle](
-    c: CoreplexConfig, l: L, b: => B)(implicit p: Parameters) extends BaseCoreplexModule(c, l, b)(p)
-    with AsyncConnection
+class MultiClockCoreplexModule[+L <: MultiClockCoreplex, +B <: MultiClockCoreplexBundle[L]](_outer: L, _io: () => B) extends BaseCoreplexModule(_outer, _io)
+    with AsyncConnectionModule
diff --git a/src/main/scala/diplomacy/LazyModule.scala b/src/main/scala/diplomacy/LazyModule.scala
index a8e2c10b..5de4877c 100644
--- a/src/main/scala/diplomacy/LazyModule.scala
+++ b/src/main/scala/diplomacy/LazyModule.scala
@@ -62,7 +62,7 @@ abstract class LazyModule
 
   private def nodesGraphML(buf: StringBuilder, pad: String) {
     buf ++= s"""${pad}<node id=\"${index}\">\n"""
-    buf ++= s"""${pad}  <data key=\"n\"><y:ShapeNode><y:NodeLabel modelName=\"sides\" modelPosition=\"w\" fontSize=\"10\" borderDistance=\"1.0\" rotationAngle=\"270.0\">${module.name}</y:NodeLabel></y:ShapeNode></data>\n"""
+    buf ++= s"""${pad}  <data key=\"n\"><y:ShapeNode><y:NodeLabel modelName=\"sides\" modelPosition=\"w\" rotationAngle=\"270.0\">${module.instanceName}</y:NodeLabel></y:ShapeNode></data>\n"""
     buf ++= s"""${pad}  <graph id=\"${index}::\" edgedefault=\"directed\">\n"""
     nodes.filter(!_.omitGraphML).foreach { n =>
       buf ++= s"""${pad}    <node id=\"${index}::${n.index}\"/>\n"""
@@ -72,11 +72,15 @@ abstract class LazyModule
     buf ++= s"""${pad}</node>\n"""
   }
   private def edgesGraphML(buf: StringBuilder, pad: String) {
-    nodes.filter(!_.omitGraphML) foreach { n => n.outputs.filter(!_.omitGraphML).foreach { o =>
+    nodes.filter(!_.omitGraphML) foreach { n => n.outputs.filter(!_._1.omitGraphML).foreach { case (o, l) =>
       buf ++= pad
       buf ++= "<edge"
       buf ++= s""" source=\"${index}::${n.index}\""""
-      buf ++= s""" target=\"${o.lazyModule.index}::${o.index}\"><data key=\"e\"><y:PolyLineEdge><y:Arrows source=\"none\" target=\"standard\"/><y:LineStyle color=\"${o.colour}\" type=\"line\" width=\"1.0\"/></y:PolyLineEdge></data></edge>\n"""
+      buf ++= s""" target=\"${o.lazyModule.index}::${o.index}\"><data key=\"e\"><y:PolyLineEdge>"""
+      buf ++= s"""<y:Arrows source=\"none\" target=\"standard\"/>"""
+      buf ++= s"""<y:LineStyle color=\"${o.colour}\" type=\"line\" width=\"1.0\"/>"""
+      buf ++= s"""<y:EdgeLabel modelName=\"centered\" rotationAngle=\"270.0\">${l}</y:EdgeLabel>"""
+      buf ++= s"""</y:PolyLineEdge></data></edge>\n"""
     } }
     children.filter(!_.omitGraphML).foreach { c => c.edgesGraphML(buf, pad) }
   }
diff --git a/src/main/scala/diplomacy/Nodes.scala b/src/main/scala/diplomacy/Nodes.scala
index db6b3513..83e0c80a 100644
--- a/src/main/scala/diplomacy/Nodes.scala
+++ b/src/main/scala/diplomacy/Nodes.scala
@@ -20,6 +20,7 @@ trait InwardNodeImp[DI, UI, EI, BI <: Data]
   // 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
@@ -34,6 +35,7 @@ trait OutwardNodeImp[DO, UO, EO, BO <: Data]
   // 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]
@@ -54,8 +56,8 @@ abstract class BaseNode
 
   protected[diplomacy] def gci: Option[BaseNode] // greatest common inner
   protected[diplomacy] def gco: Option[BaseNode] // greatest common outer
-  protected[diplomacy] def outputs: Seq[BaseNode]
-  protected[diplomacy] def inputs:  Seq[BaseNode]
+  protected[diplomacy] def outputs: Seq[(BaseNode, String)]
+  protected[diplomacy] def inputs:  Seq[(BaseNode, String)]
   protected[diplomacy] def colour:  String
 }
 
@@ -94,7 +96,7 @@ trait InwardNode[DI, UI, BI <: Data] extends BaseNode with InwardNodeHandle[DI,
   protected[diplomacy] lazy val iPorts = { iRealized = true; reqI(); accPI.result() }
 
   protected[diplomacy] val iParams: Seq[UI]
-  protected[diplomacy] def iConnect: Vec[BI]
+  val bundleIn: Vec[BI]
 }
 
 trait OutwardNodeHandle[DO, UO, BO <: Data]
@@ -126,7 +128,7 @@ trait OutwardNode[DO, UO, BO <: Data] extends BaseNode with OutwardNodeHandle[DO
   protected[diplomacy] lazy val oPorts = { oRealized = true; reqO(); accPO.result() }
 
   protected[diplomacy] val oParams: Seq[DO]
-  protected[diplomacy] def oConnect: Vec[BO]
+  val bundleOut: Vec[BO]
 }
 
 class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
@@ -140,8 +142,8 @@ 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 outputs = oPorts.map(_._2)
-  protected[diplomacy] def inputs  = iPorts.map(_._2)
+  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))
 
   private def reqE(o: Int, i: Int) = require(i == o, s"${name} has ${i} inputs and ${o} outputs; they must match${lazyModule.line}")
   protected[diplomacy] lazy val oParams: Seq[DO] = {
@@ -161,11 +163,15 @@ class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   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 bundleOut = outer.bundleO(edgesOut)
-  lazy val bundleIn  = inner.bundleI(edgesIn)
+  val flip = false // needed for blind nodes
+  private def flipO(b: Vec[BO]) = if (flip) b.flip else b
+  private def flipI(b: Vec[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: Vec[BO]) = if (wire) Wire(b) else b
+  private def wireI(b: Vec[BI]) = if (wire) Wire(b) else b
 
-  def oConnect = bundleOut
-  def iConnect = bundleIn
+  lazy val bundleOut = wireO(flipO(outer.bundleO(edgesOut)))
+  lazy val bundleIn  = wireI(flipI(inner.bundleI(edgesIn)))
 
   // connects the outward part of a node with the inward part of this node
   override def := (h: OutwardNodeHandle[DI, UI, BI])(implicit sourceInfo: SourceInfo): Option[LazyModule] = {
@@ -177,7 +183,7 @@ class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
     val o = y.oPushed
     y.oPush(i, x)
     x.iPush(o, y)
-    val (out, binding) = inner.connect(y.oConnect(o), x.iConnect(i), x.edgesIn(i))
+    val (out, binding) = inner.connect(y.bundleOut(o), x.bundleIn(i), x.edgesIn(i))
     LazyModule.stack.head.bindings = binding :: LazyModule.stack.head.bindings
     out
   }
@@ -195,26 +201,54 @@ class IdentityNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])
 
 class OutputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B]) extends IdentityNode(imp)
 {
-  override def oConnect = bundleOut
-  override def iConnect = bundleOut
+  override lazy val bundleIn = bundleOut
 }
 
 class InputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B]) extends IdentityNode(imp)
 {
-  override def oConnect = bundleIn
-  override def iConnect = bundleIn
+  override lazy val bundleOut = bundleIn
 }
 
 class SourceNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(po: PO, num: Range.Inclusive = 1 to 1)
   extends SimpleNode(imp)({case (n, Seq()) => Seq.fill(n)(po)}, {case (0, _) => Seq()}, num, 0 to 0)
 {
   require (num.end >= 1, s"${name} is a source which does not accept outputs${lazyModule.line}")
+  override lazy val bundleIn = { require(false, s"${name} has no bundleIn; try bundleOut?"); bundleOut }
 }
 
 class SinkNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(pi: PI, num: Range.Inclusive = 1 to 1)
   extends SimpleNode(imp)({case (0, _) => Seq()}, {case (n, Seq()) => Seq.fill(n)(pi)}, 0 to 0, num)
 {
   require (num.end >= 1, s"${name} is a sink which does not accept inputs${lazyModule.line}")
+  override lazy val bundleOut = { require(false, s"${name} has no bundleOut; try bundleIn?"); bundleIn }
+}
+
+class BlindOutputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(pi: PI)
+  extends SimpleNode(imp)({case (0, _) => Seq()}, {case (n, Seq()) => Seq.fill(n)(pi)}, 0 to 0, 1 to 1)
+{
+  override val flip = true
+  override lazy val bundleOut = bundleIn
+}
+
+class BlindInputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(po: PO)
+  extends SimpleNode(imp)({case (n, Seq()) => Seq.fill(n)(po)}, {case (0, _) => Seq()}, 1 to 1, 0 to 0)
+{
+  override val flip = true
+  override lazy val bundleIn = bundleOut
+}
+
+class InternalOutputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(pi: PI)
+  extends SimpleNode(imp)({case (0, _) => Seq()}, {case (n, Seq()) => Seq.fill(n)(pi)}, 0 to 0, 1 to 1)
+{
+  override val wire = true
+  override lazy val bundleOut = bundleIn
+}
+
+class InternalInputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(po: PO)
+  extends SimpleNode(imp)({case (n, Seq()) => Seq.fill(n)(po)}, {case (0, _) => Seq()}, 1 to 1, 0 to 0)
+{
+  override val wire = true
+  override lazy val bundleIn = bundleOut
 }
 
 class InteriorNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])
diff --git a/src/main/scala/groundtest/Configs.scala b/src/main/scala/groundtest/Configs.scala
index 7f5bc0eb..5503e5a4 100644
--- a/src/main/scala/groundtest/Configs.scala
+++ b/src/main/scala/groundtest/Configs.scala
@@ -72,8 +72,6 @@ class Edge32BitMemtestConfig extends Config(
 /* Composable Configs to set individual parameters */
 class WithGroundTest extends Config(
   (pname, site, here) => pname match {
-    case BuildCoreplex =>
-      (c: CoreplexConfig, p: Parameters) => LazyModule(new GroundTestCoreplex(c)(p)).module
     case TLKey("L1toL2") => {
       val useMEI = site(NTiles) <= 1 && site(NCachedTileLinkPorts) <= 1
       val dataBeats = (8 * site(CacheBlockBytes)) / site(XLen)
@@ -95,8 +93,8 @@ class WithGroundTest extends Config(
     case BuildTiles => {
       (0 until site(NTiles)).map { i =>
         val tileSettings = site(GroundTestKey)(i)
-        (r: Bool, p: Parameters) => {
-          Module(new GroundTestTile(resetSignal = r)(p.alterPartial({
+        (p: Parameters) => {
+          LazyModule(new GroundTestTile()(p.alterPartial({
             case TLId => "L1toL2"
             case TileId => i
             case NCachedTileLinkPorts => if(tileSettings.cached > 0) 1 else 0
@@ -106,7 +104,7 @@ class WithGroundTest extends Config(
       }
     }
     case BuildExampleTop =>
-      (p: Parameters) => LazyModule(new ExampleTopWithTestRAM(p))
+      (p: Parameters) => LazyModule(new ExampleTopWithTestRAM(new GroundTestCoreplex()(_))(p))
     case FPUKey => None
     case UseAtomics => false
     case UseCompressed => false
@@ -116,12 +114,12 @@ class WithGroundTest extends Config(
 class WithComparator extends Config(
   (pname, site, here) => pname match {
     case GroundTestKey => Seq.fill(site(NTiles)) {
-      GroundTestTileSettings(uncached = site(ComparatorKey).targets.size)
+      GroundTestTileSettings(uncached = 2)
     }
     case BuildGroundTest =>
       (p: Parameters) => Module(new ComparatorCore()(p))
     case ComparatorKey => ComparatorParameters(
-      targets    = Seq("mem", "io:pbus:TL2:testram").map(name =>
+      targets    = Seq("mem", "TL2:testram").map(name =>
                     site(GlobalAddrMap)(name).start.longValue),
       width      = 8,
       operations = 1000,
diff --git a/src/main/scala/groundtest/Coreplex.scala b/src/main/scala/groundtest/Coreplex.scala
index e4492e1c..738b02f8 100644
--- a/src/main/scala/groundtest/Coreplex.scala
+++ b/src/main/scala/groundtest/Coreplex.scala
@@ -4,13 +4,14 @@ import Chisel._
 import cde.{Parameters}
 import coreplex._
 
-class GroundTestCoreplex(c: CoreplexConfig)(implicit p: Parameters) extends BaseCoreplex(c)(p) {
-  override lazy val module = Module(new GroundTestCoreplexModule(c, this, new GroundTestCoreplexBundle(c)(p))(p))
+class GroundTestCoreplex(implicit p: Parameters) extends BaseCoreplex
+    with DirectConnection {
+  override lazy val module = new GroundTestCoreplexModule(this, () => new GroundTestCoreplexBundle(this))
 }
 
-class GroundTestCoreplexBundle(c: CoreplexConfig)(implicit p: Parameters) extends BaseCoreplexBundle(c)(p)
+class GroundTestCoreplexBundle[+L <: GroundTestCoreplex](_outer: L) extends BaseCoreplexBundle(_outer)
 
-class GroundTestCoreplexModule[+L <: GroundTestCoreplex, +B <: GroundTestCoreplexBundle](
-    c: CoreplexConfig, l: L, b: => B)(implicit p: Parameters) extends BaseCoreplexModule(c, l, b)(p) with DirectConnection {
+class GroundTestCoreplexModule[+L <: GroundTestCoreplex, +B <: GroundTestCoreplexBundle[L]](_outer: L, _io: () => B) extends BaseCoreplexModule(_outer, _io)
+    with DirectConnectionModule {
   io.success := tiles.flatMap(_.io.elements get "success").map(_.asInstanceOf[Bool]).reduce(_&&_)
 }
diff --git a/src/main/scala/groundtest/Generator.scala b/src/main/scala/groundtest/Generator.scala
index f5c6bb27..62a1f38c 100644
--- a/src/main/scala/groundtest/Generator.scala
+++ b/src/main/scala/groundtest/Generator.scala
@@ -5,6 +5,7 @@ package groundtest
 object Generator extends util.GeneratorApp {
   val longName = names.topModuleProject + "." + names.configs
   generateFirrtl
+  generateGraphML
   generateTestSuiteMakefrags // TODO: Needed only for legacy make targets
   generateParameterDump // TODO: Needed only for legacy make targets
 }
diff --git a/src/main/scala/groundtest/Regression.scala b/src/main/scala/groundtest/Regression.scala
index cf0fefe7..3db5ad74 100644
--- a/src/main/scala/groundtest/Regression.scala
+++ b/src/main/scala/groundtest/Regression.scala
@@ -71,7 +71,7 @@ class IOGetAfterPutBlockRegression(implicit p: Parameters) extends Regression()(
   io.mem.grant.ready := Bool(true)
 
   io.cache.req.valid := !get_sent && started
-  io.cache.req.bits.addr := UInt(addrMap("io:pbus:TL2:bootrom").start)
+  io.cache.req.bits.addr := UInt(addrMap("TL2:bootrom").start)
   io.cache.req.bits.typ := MT_WU
   io.cache.req.bits.cmd := M_XRD
   io.cache.req.bits.tag := UInt(0)
diff --git a/src/main/scala/groundtest/Tile.scala b/src/main/scala/groundtest/Tile.scala
index 5a550fb3..f0973dc4 100644
--- a/src/main/scala/groundtest/Tile.scala
+++ b/src/main/scala/groundtest/Tile.scala
@@ -96,48 +96,47 @@ abstract class GroundTest(implicit val p: Parameters) extends Module
   val io = new GroundTestIO
 }
 
-class GroundTestTile(resetSignal: Bool)
-                    (implicit val p: Parameters)
-                    extends Tile(resetSignal = resetSignal)(p)
-                    with HasGroundTestParameters {
-
-  override val io = new TileIO(bc) {
-    val success = Bool(OUTPUT)
-  }
-
-  val test = p(BuildGroundTest)(dcacheParams)
-
-  val ptwPorts = ListBuffer.empty ++= test.io.ptw
-  val memPorts = ListBuffer.empty ++= test.io.mem
-
-  if (nCached > 0) {
-    val dcache_io = HellaCache(p(DCacheKey))(dcacheParams)
-    val dcacheArb = Module(new HellaCacheArbiter(nCached)(dcacheParams))
-
-    dcacheArb.io.requestor.zip(test.io.cache).foreach {
-      case (requestor, cache) =>
-        val dcacheIF = Module(new SimpleHellaCacheIF()(dcacheParams))
-        dcacheIF.io.requestor <> cache
-        requestor <> dcacheIF.io.cache
+class GroundTestTile(implicit val p: Parameters) extends LazyTile {
+  val slave = None
+  lazy val module = new TileImp(this) with HasGroundTestParameters {
+    val io = new TileIO(bc) {
+      val success = Bool(OUTPUT)
     }
-    dcache_io.cpu <> dcacheArb.io.mem
-    io.cached.head <> dcache_io.mem
 
-    // SimpleHellaCacheIF leaves invalidate_lr dangling, so we wire it to false
-    dcache_io.cpu.invalidate_lr := Bool(false)
+    val test = p(BuildGroundTest)(dcacheParams)
 
-    ptwPorts += dcache_io.ptw
+    val ptwPorts = ListBuffer.empty ++= test.io.ptw
+    val memPorts = ListBuffer.empty ++= test.io.mem
+
+    if (nCached > 0) {
+      val dcache_io = HellaCache(p(DCacheKey))(dcacheParams)
+      val dcacheArb = Module(new HellaCacheArbiter(nCached)(dcacheParams))
+
+      dcacheArb.io.requestor.zip(test.io.cache).foreach {
+        case (requestor, cache) =>
+          val dcacheIF = Module(new SimpleHellaCacheIF()(dcacheParams))
+          dcacheIF.io.requestor <> cache
+          requestor <> dcacheIF.io.cache
+      }
+      dcache_io.cpu <> dcacheArb.io.mem
+      io.cached.head <> dcache_io.mem
+
+      // SimpleHellaCacheIF leaves invalidate_lr dangling, so we wire it to false
+      dcache_io.cpu.invalidate_lr := Bool(false)
+
+      ptwPorts += dcache_io.ptw
+    }
+
+    if (ptwPorts.size > 0) {
+      val ptw = Module(new DummyPTW(ptwPorts.size))
+      ptw.io.requestors <> ptwPorts
+    }
+
+    require(memPorts.size == io.uncached.size)
+    if (memPorts.size > 0) {
+      io.uncached <> memPorts
+    }
+
+    io.success := test.io.status.finished
   }
-
-  if (ptwPorts.size > 0) {
-    val ptw = Module(new DummyPTW(ptwPorts.size))
-    ptw.io.requestors <> ptwPorts
-  }
-
-  require(memPorts.size == io.uncached.size)
-  if (memPorts.size > 0) {
-    io.uncached <> memPorts
-  }
-
-  io.success := test.io.status.finished
 }
diff --git a/src/main/scala/junctions/addrmap.scala b/src/main/scala/junctions/addrmap.scala
index 56158001..86fdd276 100644
--- a/src/main/scala/junctions/addrmap.scala
+++ b/src/main/scala/junctions/addrmap.scala
@@ -76,12 +76,7 @@ class AddrMap(
     var cacheable = true
     for (AddrMapEntry(name, r) <- entriesIn) {
       require (!mapping.contains(name))
-
-      if (r.start != 0) {
-        base = r.start
-      } else {
-        base = (base + r.size - 1) / r.size * r.size
-      }
+      base = r.start
 
       r match {
         case r: AddrMap =>
diff --git a/src/main/scala/rocket/dcache.scala b/src/main/scala/rocket/dcache.scala
index 9bd89541..8e908dcd 100644
--- a/src/main/scala/rocket/dcache.scala
+++ b/src/main/scala/rocket/dcache.scala
@@ -4,7 +4,9 @@ package rocket
 
 import Chisel._
 import junctions._
+import diplomacy._
 import uncore.tilelink._
+import uncore.tilelink2._
 import uncore.agents._
 import uncore.coherence._
 import uncore.constants._
@@ -125,7 +127,7 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
       require(nWays == 1)
       metaWriteArb.io.out.ready := true
       metaReadArb.io.out.ready := !metaWriteArb.io.out.valid
-      val inScratchpad = addrMap(s"io:cbus:dmem${tileId}").containsAddress(s1_paddr)
+      val inScratchpad = addrMap(s"TL2:dmem${tileId}").containsAddress(s1_paddr)
       val hitState = Mux(inScratchpad, ClientMetadata.onReset.onHit(M_XWR), ClientMetadata.onReset)
       (inScratchpad, hitState, L1Metadata(UInt(0), ClientMetadata.onReset))
     } else {
@@ -495,60 +497,97 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   }
 }
 
-class ScratchpadSlavePort(implicit p: Parameters) extends CoreModule()(p) {
-  val io = new Bundle {
-    val tl = new ClientUncachedTileLinkIO().flip
-    val dmem = new HellaCacheIO
+class ScratchpadSlavePort(implicit val p: Parameters) extends LazyModule with HasCoreParameters {
+  val node = TLManagerNode(TLManagerPortParameters(
+    Seq(TLManagerParameters(
+      address            = List(AddressSet(0x80000000L, BigInt(p(DataScratchpadSize)-1))),
+      regionType         = RegionType.UNCACHED,
+      executable         = true,
+      supportsArithmetic = if (p(UseAtomics)) TransferSizes(1, coreDataBytes) else TransferSizes.none,
+      supportsLogical    = if (p(UseAtomics)) TransferSizes(1, coreDataBytes) else TransferSizes.none,
+      supportsPutPartial = TransferSizes(1, coreDataBytes),
+      supportsPutFull    = TransferSizes(1, coreDataBytes),
+      supportsGet        = TransferSizes(1, coreDataBytes),
+      fifoId             = Some(0))), // requests handled in FIFO order
+    beatBytes = coreDataBytes,
+    minLatency = 1))
+
+  // Make sure this ends up with the same name as before
+  override def name = "dmem0"
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val tl_in = node.bundleIn
+      val dmem = new HellaCacheIO
+    }
+
+    val tl_in = io.tl_in(0)
+    val edge = node.edgesIn(0)
+
+    require(usingDataScratchpad)
+
+    val s_ready :: s_wait :: s_replay :: s_grant :: Nil = Enum(UInt(), 4)
+    val state = Reg(init = s_ready)
+    when (io.dmem.resp.valid) { state := s_grant }
+    when (tl_in.d.fire()) { state := s_ready }
+    when (io.dmem.s2_nack) { state := s_replay }
+    when (io.dmem.req.fire()) { state := s_wait }
+
+    val acq = Reg(tl_in.a.bits)
+    when (io.dmem.resp.valid) { acq.data := io.dmem.resp.bits.data }
+    when (tl_in.a.fire()) { acq := tl_in.a.bits }
+
+    val isWrite = acq.opcode === TLMessages.PutFullData || acq.opcode === TLMessages.PutPartialData
+    val isRead = !edge.hasData(acq)
+
+    def formCacheReq(acq: TLBundleA) = {
+      val req = Wire(new HellaCacheReq)
+      req.cmd := MuxLookup(acq.opcode, Wire(M_XRD), Array(
+        TLMessages.PutFullData    -> M_XWR,
+        TLMessages.PutPartialData -> M_XWR,
+        TLMessages.ArithmeticData -> MuxLookup(acq.param, Wire(M_XRD), Array(
+          TLAtomics.MIN           -> M_XA_MIN,
+          TLAtomics.MAX           -> M_XA_MAX,
+          TLAtomics.MINU          -> M_XA_MINU,
+          TLAtomics.MAXU          -> M_XA_MAXU,
+          TLAtomics.ADD           -> M_XA_ADD)),
+        TLMessages.LogicalData    -> MuxLookup(acq.param, Wire(M_XRD), Array(
+          TLAtomics.XOR           -> M_XA_XOR,
+          TLAtomics.OR            -> M_XA_OR,
+          TLAtomics.AND           -> M_XA_AND,
+          TLAtomics.SWAP          -> M_XA_SWAP)),
+        TLMessages.Get            -> M_XRD))
+      // treat all loads as full words, so bytes appear in correct lane
+      req.typ := Mux(isRead, log2Ceil(coreDataBytes), acq.size)
+      req.addr := Mux(isRead, ~(~acq.address | (coreDataBytes-1)), acq.address)
+      req.tag := UInt(0)
+      req
+    }
+
+    val ready = state === s_ready || tl_in.d.fire()
+    io.dmem.req.valid := (tl_in.a.valid && ready) || state === s_replay
+    tl_in.a.ready := io.dmem.req.ready && ready
+    io.dmem.req.bits := formCacheReq(Mux(state === s_replay, acq, tl_in.a.bits))
+    // the TL data is already in the correct byte lane, but the D$
+    // expects right-justified store data, so that it can steer the bytes.
+    io.dmem.s1_data := new LoadGen(acq.size, Bool(false), acq.address(log2Ceil(coreDataBytes)-1,0), acq.data, Bool(false), coreDataBytes).data
+    io.dmem.s1_kill := false
+    io.dmem.invalidate_lr := false
+
+    // place AMO data in correct word lane
+    val minAMOBytes = 4
+    val grantData = Mux(io.dmem.resp.valid, io.dmem.resp.bits.data, acq.data)
+    val alignedGrantData = Mux(acq.size <= log2Ceil(minAMOBytes), Fill(coreDataBytes/minAMOBytes, grantData(8*minAMOBytes-1, 0)), grantData)
+
+    tl_in.d.valid := io.dmem.resp.valid || state === s_grant
+    tl_in.d.bits := Mux(isWrite,
+      edge.AccessAck(acq, UInt(0)),
+      edge.AccessAck(acq, UInt(0), UInt(0)))
+    tl_in.d.bits.data := alignedGrantData
+
+    // Tie off unused channels
+    tl_in.b.valid := Bool(false)
+    tl_in.c.ready := Bool(true)
+    tl_in.e.ready := Bool(true)
   }
-
-  val s_ready :: s_wait :: s_replay :: s_grant :: Nil = Enum(UInt(), 4)
-  val state = Reg(init = s_ready)
-  when (io.dmem.resp.valid) { state := s_grant }
-  when (io.tl.grant.fire()) { state := s_ready }
-  when (io.dmem.s2_nack) { state := s_replay }
-  when (io.dmem.req.fire()) { state := s_wait }
-
-  val acq = Reg(io.tl.acquire.bits)
-  when (io.dmem.resp.valid) { acq.data := io.dmem.resp.bits.data }
-  when (io.tl.acquire.fire()) { acq := io.tl.acquire.bits }
-
-  val isRead = acq.isBuiltInType(Acquire.getType)
-  val isWrite = acq.isBuiltInType(Acquire.putType)
-  assert(state === s_ready || isRead || isWrite)
-  require(coreDataBits == acq.tlDataBits)
-  require(usingDataScratchpad)
-
-  def formCacheReq(acq: Acquire) = {
-    val req = Wire(new HellaCacheReq)
-    // treat all loads as full words, so bytes appear in correct lane
-    req.typ := Mux(isRead, log2Ceil(acq.tlDataBytes), acq.op_size())
-    req.cmd := acq.op_code()
-    req.addr := Mux(isRead, ~(~acq.full_addr() | (acq.tlDataBytes-1)), acq.full_addr())
-    req.tag := UInt(0)
-    req
-  }
-
-  val ready = state === s_ready || io.tl.grant.fire()
-  io.dmem.req.valid := (io.tl.acquire.valid && ready) || state === s_replay
-  io.tl.acquire.ready := io.dmem.req.ready && ready
-  io.dmem.req.bits := formCacheReq(Mux(state === s_replay, acq, io.tl.acquire.bits))
-  // this blows.  the TL data is already in the correct byte lane, but the D$
-  // expects right-justified store data, so that it can steer the bytes.
-  io.dmem.s1_data := new LoadGen(acq.op_size(), Bool(false), acq.addr_byte(), acq.data, Bool(false), acq.tlDataBytes).data
-  io.dmem.s1_kill := false
-  io.dmem.invalidate_lr := false
-
-  // place AMO data in correct word lane
-  val minAMOBytes = 4
-  val grantData = Mux(io.dmem.resp.valid, io.dmem.resp.bits.data, acq.data)
-  val alignedGrantData = Mux(acq.op_size() <= log2Ceil(minAMOBytes), Fill(coreDataBytes/minAMOBytes, grantData(8*minAMOBytes-1, 0)), grantData)
-
-  io.tl.grant.valid := io.dmem.resp.valid || state === s_grant
-  io.tl.grant.bits := Grant(
-    is_builtin_type = Bool(true),
-    g_type = acq.getBuiltInGrantType(),
-    client_xact_id = acq.client_xact_id,
-    manager_xact_id = UInt(0),
-    addr_beat = acq.addr_beat,
-    data = alignedGrantData)
 }
diff --git a/src/main/scala/rocket/tile.scala b/src/main/scala/rocket/tile.scala
index 39dd5a35..c497a98e 100644
--- a/src/main/scala/rocket/tile.scala
+++ b/src/main/scala/rocket/tile.scala
@@ -3,7 +3,9 @@
 package rocket
 
 import Chisel._
+import diplomacy._
 import uncore.tilelink._
+import uncore.tilelink2._
 import uncore.agents._
 import uncore.converters._
 import uncore.devices._
@@ -25,136 +27,141 @@ case class RoccParameters(
 case class TileBundleConfig(
   nCachedTileLinkPorts: Int,
   nUncachedTileLinkPorts: Int,
-  xLen: Int,
-  hasSlavePort: Boolean)
+  xLen: Int)
 
-class TileIO(c: TileBundleConfig)(implicit p: Parameters) extends Bundle {
+class TileIO(c: TileBundleConfig, node: Option[TLInwardNode] = None)(implicit p: Parameters) extends Bundle {
   val cached = Vec(c.nCachedTileLinkPorts, new ClientTileLinkIO)
   val uncached = Vec(c.nUncachedTileLinkPorts, new ClientUncachedTileLinkIO)
   val hartid = UInt(INPUT, c.xLen)
   val interrupts = new TileInterrupts().asInput
-  val slave = c.hasSlavePort.option(new ClientUncachedTileLinkIO().flip)
+  val slave = node.map(_.inward.bundleIn)
   val resetVector = UInt(INPUT, c.xLen)
 
   override def cloneType = new TileIO(c).asInstanceOf[this.type]
 }
 
-abstract class Tile(clockSignal: Clock = null, resetSignal: Bool = null)
-    (implicit p: Parameters) extends Module(Option(clockSignal), Option(resetSignal)) {
+abstract class TileImp(l: LazyTile)(implicit val p: Parameters) extends LazyModuleImp(l) {
+  val io: TileIO
+}
+
+abstract class LazyTile(implicit p: Parameters) extends LazyModule {
   val nCachedTileLinkPorts = p(NCachedTileLinkPorts)
   val nUncachedTileLinkPorts = p(NUncachedTileLinkPorts)
   val dcacheParams = p.alterPartial({ case CacheName => "L1D" })
   val bc = TileBundleConfig(
     nCachedTileLinkPorts = nCachedTileLinkPorts,
     nUncachedTileLinkPorts = nUncachedTileLinkPorts,
-    xLen = p(XLen),
-    hasSlavePort = p(DataScratchpadSize) > 0)
+    xLen = p(XLen))
 
-  val io = new TileIO(bc)
+  val module: TileImp
+  val slave: Option[TLInputNode]
 }
 
-class RocketTile(clockSignal: Clock = null, resetSignal: Bool = null)
-    (implicit p: Parameters) extends Tile(clockSignal, resetSignal)(p) {
-  val buildRocc = p(BuildRoCC)
-  val usingRocc = !buildRocc.isEmpty
-  val nRocc = buildRocc.size
-  val nFPUPorts = buildRocc.filter(_.useFPU).size
+class RocketTile(implicit p: Parameters) extends LazyTile {
+  val slave = if (p(DataScratchpadSize) == 0) None else Some(TLInputNode())
+  val scratch = if (p(DataScratchpadSize) == 0) None else Some(LazyModule(new ScratchpadSlavePort()(dcacheParams)))
 
-  val core = Module(new Rocket)
-  val icache = Module(new Frontend()(p.alterPartial({ case CacheName => "L1I" })))
-  val dcache = HellaCache(p(DCacheKey))(dcacheParams)
+  (slave zip scratch) foreach { case (node, lm) => lm.node := TLFragmenter(p(XLen)/8, p(CacheBlockBytes))(node) }
 
-  val ptwPorts = collection.mutable.ArrayBuffer(icache.io.ptw, dcache.ptw)
-  val dcPorts = collection.mutable.ArrayBuffer(core.io.dmem)
-  val uncachedArbPorts = collection.mutable.ArrayBuffer(icache.io.mem)
-  val uncachedPorts = collection.mutable.ArrayBuffer[ClientUncachedTileLinkIO]()
-  val cachedPorts = collection.mutable.ArrayBuffer(dcache.mem)
-  core.io.interrupts := io.interrupts
-  core.io.hartid := io.hartid
-  icache.io.cpu <> core.io.imem
-  icache.io.resetVector := io.resetVector
+  lazy val module = new TileImp(this) {
+    val io = new TileIO(bc, slave)
+    val buildRocc = p(BuildRoCC)
+    val usingRocc = !buildRocc.isEmpty
+    val nRocc = buildRocc.size
+    val nFPUPorts = buildRocc.filter(_.useFPU).size
 
-  val fpuOpt = p(FPUKey).map(cfg => Module(new FPU(cfg)))
-  fpuOpt.foreach(fpu => core.io.fpu <> fpu.io)
+    val core = Module(new Rocket)
+    val icache = Module(new Frontend()(p.alterPartial({ case CacheName => "L1I" })))
+    val dcache = HellaCache(p(DCacheKey))(dcacheParams)
 
-  if (usingRocc) {
-    val respArb = Module(new RRArbiter(new RoCCResponse, nRocc))
-    core.io.rocc.resp <> respArb.io.out
+    val ptwPorts = collection.mutable.ArrayBuffer(icache.io.ptw, dcache.ptw)
+    val dcPorts = collection.mutable.ArrayBuffer(core.io.dmem)
+    val uncachedArbPorts = collection.mutable.ArrayBuffer(icache.io.mem)
+    val uncachedPorts = collection.mutable.ArrayBuffer[ClientUncachedTileLinkIO]()
+    val cachedPorts = collection.mutable.ArrayBuffer(dcache.mem)
+    core.io.interrupts := io.interrupts
+    core.io.hartid := io.hartid
+    icache.io.cpu <> core.io.imem
+    icache.io.resetVector := io.resetVector
 
-    val roccOpcodes = buildRocc.map(_.opcodes)
-    val cmdRouter = Module(new RoccCommandRouter(roccOpcodes))
-    cmdRouter.io.in <> core.io.rocc.cmd
+    val fpuOpt = p(FPUKey).map(cfg => Module(new FPU(cfg)))
+    fpuOpt.foreach(fpu => core.io.fpu <> fpu.io)
 
-    val roccs = buildRocc.zipWithIndex.map { case (accelParams, i) =>
-      val rocc = accelParams.generator(p.alterPartial({
-        case RoccNMemChannels => accelParams.nMemChannels
-        case RoccNPTWPorts => accelParams.nPTWPorts
-      }))
-      val dcIF = Module(new SimpleHellaCacheIF()(dcacheParams))
-      rocc.io.cmd <> cmdRouter.io.out(i)
-      rocc.io.exception := core.io.rocc.exception
-      dcIF.io.requestor <> rocc.io.mem
-      dcPorts += dcIF.io.cache
-      uncachedArbPorts += rocc.io.autl
-      rocc
+    if (usingRocc) {
+      val respArb = Module(new RRArbiter(new RoCCResponse, nRocc))
+      core.io.rocc.resp <> respArb.io.out
+
+      val roccOpcodes = buildRocc.map(_.opcodes)
+      val cmdRouter = Module(new RoccCommandRouter(roccOpcodes))
+      cmdRouter.io.in <> core.io.rocc.cmd
+
+      val roccs = buildRocc.zipWithIndex.map { case (accelParams, i) =>
+        val rocc = accelParams.generator(p.alterPartial({
+          case RoccNMemChannels => accelParams.nMemChannels
+          case RoccNPTWPorts => accelParams.nPTWPorts
+        }))
+        val dcIF = Module(new SimpleHellaCacheIF()(dcacheParams))
+        rocc.io.cmd <> cmdRouter.io.out(i)
+        rocc.io.exception := core.io.rocc.exception
+        dcIF.io.requestor <> rocc.io.mem
+        dcPorts += dcIF.io.cache
+        uncachedArbPorts += rocc.io.autl
+        rocc
+      }
+
+      if (nFPUPorts > 0) {
+        fpuOpt.foreach { fpu =>
+          val fpArb = Module(new InOrderArbiter(new FPInput, new FPResult, nFPUPorts))
+          val fp_roccs = roccs.zip(buildRocc)
+            .filter { case (_, params) => params.useFPU }
+            .map { case (rocc, _) => rocc.io }
+          fpArb.io.in_req <> fp_roccs.map(_.fpu_req)
+          fp_roccs.zip(fpArb.io.in_resp).foreach {
+            case (rocc, fpu_resp) => rocc.fpu_resp <> fpu_resp
+          }
+          fpu.io.cp_req <> fpArb.io.out_req
+          fpArb.io.out_resp <> fpu.io.cp_resp
+        }
+      }
+
+      core.io.rocc.busy := cmdRouter.io.busy || roccs.map(_.io.busy).reduce(_ || _)
+      core.io.rocc.interrupt := roccs.map(_.io.interrupt).reduce(_ || _)
+      respArb.io.in <> roccs.map(rocc => Queue(rocc.io.resp))
+
+      ptwPorts ++= roccs.flatMap(_.io.ptw)
+      uncachedPorts ++= roccs.flatMap(_.io.utl)
     }
 
-    if (nFPUPorts > 0) {
+    val uncachedArb = Module(new ClientUncachedTileLinkIOArbiter(uncachedArbPorts.size))
+    uncachedArb.io.in <> uncachedArbPorts
+    uncachedArb.io.out +=: uncachedPorts
+
+    // Connect the caches and RoCC to the outer memory system
+    io.uncached <> uncachedPorts
+    io.cached <> cachedPorts
+    // TODO remove nCached/nUncachedTileLinkPorts parameters and these assertions
+    require(uncachedPorts.size == nUncachedTileLinkPorts)
+    require(cachedPorts.size == nCachedTileLinkPorts)
+
+    if (p(UseVM)) {
+      val ptw = Module(new PTW(ptwPorts.size)(dcacheParams))
+      ptw.io.requestor <> ptwPorts
+      ptw.io.mem +=: dcPorts
+      core.io.ptw <> ptw.io.dpath
+    }
+
+    scratch.foreach { lm => lm.module.io.dmem +=: dcPorts }
+
+    require(dcPorts.size == core.dcacheArbPorts)
+    val dcArb = Module(new HellaCacheArbiter(dcPorts.size)(dcacheParams))
+    dcArb.io.requestor <> dcPorts
+    dcache.cpu <> dcArb.io.mem
+
+    if (nFPUPorts == 0) {
       fpuOpt.foreach { fpu =>
-        val fpArb = Module(new InOrderArbiter(new FPInput, new FPResult, nFPUPorts))
-        val fp_roccs = roccs.zip(buildRocc)
-          .filter { case (_, params) => params.useFPU }
-          .map { case (rocc, _) => rocc.io }
-        fpArb.io.in_req <> fp_roccs.map(_.fpu_req)
-        fp_roccs.zip(fpArb.io.in_resp).foreach {
-          case (rocc, fpu_resp) => rocc.fpu_resp <> fpu_resp
-        }
-        fpu.io.cp_req <> fpArb.io.out_req
-        fpArb.io.out_resp <> fpu.io.cp_resp
+        fpu.io.cp_req.valid := Bool(false)
+        fpu.io.cp_resp.ready := Bool(false)
       }
     }
-
-    core.io.rocc.busy := cmdRouter.io.busy || roccs.map(_.io.busy).reduce(_ || _)
-    core.io.rocc.interrupt := roccs.map(_.io.interrupt).reduce(_ || _)
-    respArb.io.in <> roccs.map(rocc => Queue(rocc.io.resp))
-
-    ptwPorts ++= roccs.flatMap(_.io.ptw)
-    uncachedPorts ++= roccs.flatMap(_.io.utl)
-  }
-
-  val uncachedArb = Module(new ClientUncachedTileLinkIOArbiter(uncachedArbPorts.size))
-  uncachedArb.io.in <> uncachedArbPorts
-  uncachedArb.io.out +=: uncachedPorts
-
-  // Connect the caches and RoCC to the outer memory system
-  io.uncached <> uncachedPorts
-  io.cached <> cachedPorts
-  // TODO remove nCached/nUncachedTileLinkPorts parameters and these assertions
-  require(uncachedPorts.size == nUncachedTileLinkPorts)
-  require(cachedPorts.size == nCachedTileLinkPorts)
-
-  if (p(UseVM)) {
-    val ptw = Module(new PTW(ptwPorts.size)(dcacheParams))
-    ptw.io.requestor <> ptwPorts
-    ptw.io.mem +=: dcPorts
-    core.io.ptw <> ptw.io.dpath
-  }
-
-  io.slave foreach { case slavePort =>
-    val adapter = Module(new ScratchpadSlavePort()(dcacheParams))
-    adapter.io.tl <> TileLinkFragmenter(slavePort)
-    adapter.io.dmem +=: dcPorts
-  }
-
-  require(dcPorts.size == core.dcacheArbPorts)
-  val dcArb = Module(new HellaCacheArbiter(dcPorts.size)(dcacheParams))
-  dcArb.io.requestor <> dcPorts
-  dcache.cpu <> dcArb.io.mem
-
-  if (nFPUPorts == 0) {
-    fpuOpt.foreach { fpu =>
-      fpu.io.cp_req.valid := Bool(false)
-      fpu.io.cp_resp.ready := Bool(false)
-    }
   }
 }
diff --git a/src/main/scala/rocketchip/BaseTop.scala b/src/main/scala/rocketchip/BaseTop.scala
index a7d9836f..99a5521e 100644
--- a/src/main/scala/rocketchip/BaseTop.scala
+++ b/src/main/scala/rocketchip/BaseTop.scala
@@ -15,103 +15,97 @@ import coreplex._
 
 // the following parameters will be refactored properly with TL2
 case object GlobalAddrMap extends Field[AddrMap]
-case object ConfigString extends Field[String]
 case object NCoreplexExtClients extends Field[Int]
 /** Enable or disable monitoring of Diplomatic buses */
 case object TLEmitMonitors extends Field[Bool]
-/** Function for building Coreplex */
-case object BuildCoreplex extends Field[(CoreplexConfig, Parameters) => BaseCoreplexModule[BaseCoreplex, BaseCoreplexBundle]]
 
-/** Base Top with no Periphery */
-abstract class BaseTop(q: Parameters) extends LazyModule {
-  // the following variables will be refactored properly with TL2
-  val pInterrupts = new RangeManager
+abstract class BareTop[+C <: BaseCoreplex](_coreplex: Parameters => C)(implicit val q: Parameters) extends LazyModule {
+  // Fill in the TL1 legacy parameters; remove these once rocket/groundtest/unittest are TL2
   val pBusMasters = new RangeManager
-  val pDevices = new ResourceManager[AddrMapEntry]
-
-  TLImp.emitMonitors = q(TLEmitMonitors)
-
-  // Add a SoC and peripheral bus
-  val socBus = LazyModule(new TLXbar)
-  val peripheryBus = LazyModule(new TLXbar)
-  lazy val peripheryManagers = socBus.node.edgesIn(0).manager.managers
-
-  lazy val c = CoreplexConfig(
-    nTiles = q(NTiles),
-    nExtInterrupts = pInterrupts.sum,
-    nSlaves = pBusMasters.sum,
-    nMemChannels = q(NMemoryChannels),
-    hasSupervisor = q(UseVM)
-  )
-
-  lazy val genGlobalAddrMap = GenerateGlobalAddrMap(q, pDevices.get, peripheryManagers)
-  private val qWithMap = q.alterPartial({case GlobalAddrMap => genGlobalAddrMap})
-
-  lazy val genConfigString = GenerateConfigString(qWithMap, c, pDevices.get, peripheryManagers)
-  implicit val p = qWithMap.alterPartial({
-    case ConfigString => genConfigString
-    case NCoreplexExtClients => pBusMasters.sum})
-
-  val legacy = LazyModule(new TLLegacy()(p.alterPartial({ case TLId => "L2toMMIO" })))
-
-  peripheryBus.node :=
-    TLWidthWidget(p(SOCBusKey).beatBytes)(
-    TLBuffer()(
-    TLAtomicAutomata(arithmetic = p(PeripheryBusKey).arithAMO)(
-    socBus.node)))
-
-  socBus.node :=
-    TLWidthWidget(legacy.tlDataBytes)(
-    TLHintHandler()(
-    legacy.node))
+  lazy val legacyAddrMap = GenerateGlobalAddrMap(q, coreplex.l1tol2.node.edgesIn(0).manager.managers)
+  val coreplex : C = LazyModule(_coreplex(q.alterPartial {
+    case NCoreplexExtClients => pBusMasters.sum
+    case GlobalAddrMap => legacyAddrMap
+  }))
 
   TopModule.contents = Some(this)
 }
 
-abstract class BaseTopBundle(val p: Parameters) extends Bundle {
+abstract class BareTopBundle[+L <: BareTop[BaseCoreplex]](_outer: L) extends Bundle {
+  val outer = _outer
+}
+
+abstract class BareTopModule[+L <: BareTop[BaseCoreplex], +B <: BareTopBundle[L]](_outer: L, _io: () => B) extends LazyModuleImp(_outer) {
+  val outer = _outer
+  val io = _io ()
+}
+
+/** Base Top with no Periphery */
+trait TopNetwork extends HasPeripheryParameters {
+  this: BareTop[BaseCoreplex] =>
+  implicit val p = q
+  TLImp.emitMonitors = p(TLEmitMonitors)
+
+  // Add a SoC and peripheral bus
+  val socBus = LazyModule(new TLXbar)
+  val peripheryBus = LazyModule(new TLXbar)
+  val intBus = LazyModule(new IntXbar)
+
+  peripheryBus.node :=
+    TLWidthWidget(p(SOCBusKey).beatBytes)(
+    TLAtomicAutomata(arithmetic = p(PeripheryBusKey).arithAMO)(
+    socBus.node))
+}
+
+trait TopNetworkBundle extends HasPeripheryParameters {
+  this: BareTopBundle[BareTop[BaseCoreplex]] =>
+  implicit val p = outer.q
   val success = Bool(OUTPUT)
 }
 
-abstract class BaseTopModule[+L <: BaseTop, +B <: BaseTopBundle](
-    val p: Parameters, l: L, b: => B) extends LazyModuleImp(l) {
-  val outer: L = l
-  val io: B = b
+trait TopNetworkModule extends HasPeripheryParameters {
+  this: {
+    val outer: BareTop[BaseCoreplex] with TopNetwork
+    val io: TopNetworkBundle
+  } =>
+  implicit val p = outer.p
 
-  val coreplex = p(BuildCoreplex)(outer.c, p)
-  val coreplexIO = Wire(coreplex.io)
+  val coreplexMem  : Vec[ClientUncachedTileLinkIO] = Wire(outer.coreplex.module.io.mem)
+  val coreplexSlave: Vec[ClientUncachedTileLinkIO] = Wire(outer.coreplex.module.io.slave)
+  val coreplexDebug: DebugBusIO                    = Wire(outer.coreplex.module.io.debug)
+  val coreplexRtc  : Bool                          = Wire(outer.coreplex.module.io.rtcTick)
 
-  val pBus =
-    Module(new TileLinkRecursiveInterconnect(1, p(GlobalAddrMap).subMap("io:pbus"))(
-      p.alterPartial({ case TLId => "L2toMMIO" })))
-  pBus.io.in.head <> coreplexIO.master.mmio
-  outer.legacy.module.io.legacy <> pBus.port("TL2")
+  io.success := outer.coreplex.module.io.success
 
-  println("Generated Address Map")
-  for (entry <- p(GlobalAddrMap).flatten) {
-    val name = entry.name
-    val start = entry.region.start
-    val end = entry.region.start + entry.region.size - 1
-    val prot = entry.region.attr.prot
-    val protStr = (if ((prot & AddrMapProt.R) > 0) "R" else "") +
-                  (if ((prot & AddrMapProt.W) > 0) "W" else "") +
-                  (if ((prot & AddrMapProt.X) > 0) "X" else "")
-    val cacheable = if (entry.region.attr.cacheable) " [C]" else ""
-    println(f"\t$name%s $start%x - $end%x, $protStr$cacheable")
-  }
-
-  println("\nGenerated Interrupt Vector")
-  outer.pInterrupts.print
-
-  println("\nGenerated Configuration String")
-  println(p(ConfigString))
-  ConfigStringOutput.contents = Some(p(ConfigString))
-
-  io.success := coreplexIO.success
+  outer.coreplex.module.io.rtcTick := coreplexRtc
+  coreplexRtc := Counter(p(rocketchip.RTCPeriod)).inc()
 }
 
+/** Base Top with no Periphery */
+class BaseTop[+C <: BaseCoreplex](_coreplex: Parameters => C)(implicit p: Parameters) extends BareTop(_coreplex)
+    with TopNetwork {
+  override lazy val module = new BaseTopModule(this, () => new BaseTopBundle(this))
+}
+
+class BaseTopBundle[+L <: BaseTop[BaseCoreplex]](_outer: L) extends BareTopBundle(_outer)
+    with TopNetworkBundle
+
+class BaseTopModule[+L <: BaseTop[BaseCoreplex], +B <: BaseTopBundle[L]](_outer: L, _io: () => B) extends BareTopModule(_outer, _io)
+    with TopNetworkModule
+
 trait DirectConnection {
-  val coreplexIO: BaseCoreplexBundle
-  val coreplex: BaseCoreplexModule[BaseCoreplex, BaseCoreplexBundle]
+  this: BareTop[BaseCoreplex] with TopNetwork =>
 
-  coreplexIO <> coreplex.io
+  socBus.node := coreplex.mmio
+  coreplex.mmioInt := intBus.intnode
+}
+
+trait DirectConnectionModule {
+  this: TopNetworkModule {
+    val outer: BaseTop[BaseCoreplex]
+  } =>
+
+  coreplexMem <> outer.coreplex.module.io.mem
+  outer.coreplex.module.io.slave <> coreplexSlave
+  outer.coreplex.module.io.debug <> coreplexDebug
 }
diff --git a/src/main/scala/rocketchip/Configs.scala b/src/main/scala/rocketchip/Configs.scala
index e253bfb5..92643944 100644
--- a/src/main/scala/rocketchip/Configs.scala
+++ b/src/main/scala/rocketchip/Configs.scala
@@ -40,19 +40,12 @@ class BasePlatformConfig extends Config(
           site(TLKey("L2toMC")).copy(dataBeats = edgeDataBeats)
         case TLKey("MMIOtoEdge") =>
           site(TLKey("L2toMMIO")).copy(dataBeats = edgeDataBeats)
-        case BuildCoreplex =>
-          (c: CoreplexConfig, p: Parameters) => LazyModule(new DefaultCoreplex(c)(p)).module
         case NExtTopInterrupts => 2
         case SOCBusKey => SOCBusConfig(beatBytes = site(TLKey("L2toMMIO")).dataBitsPerBeat/8)
         case PeripheryBusKey => PeripheryBusConfig(arithAMO = true, beatBytes = 4)
         // Note that PLIC asserts that this is > 0.
         case AsyncDebugBus => false
         case IncludeJtagDTM => false
-        case AsyncMMIOChannels => false
-        case ExtMMIOPorts => Nil
-        case NExtMMIOAXIChannels => 0
-        case NExtMMIOAHBChannels => 0
-        case NExtMMIOTLChannels  => 0
         case AsyncBusChannels => false
         case NExtBusAXIChannels => 0
         case HastiId => "Ext"
@@ -70,7 +63,7 @@ class BasePlatformConfig extends Config(
         case ExtMemSize => Dump("MEM_SIZE", 0x10000000L)
         case RTCPeriod => 100 // gives 10 MHz RTC assuming 1 GHz uncore clock
         case BuildExampleTop =>
-          (p: Parameters) => LazyModule(new ExampleTop(p))
+          (p: Parameters) => LazyModule(new ExampleTop(new DefaultCoreplex()(_))(p))
         case SimMemLatency => 0
         case _ => throw new CDEMatchError
       }
@@ -110,17 +103,6 @@ class WithExtMemSize(n: Long) extends Config(
     case _ => throw new CDEMatchError
   }
 )
-class WithAHB extends Config(
-  (pname, site, here) => pname match {
-    case TMemoryChannels     => BusType.AHB
-    case NExtMMIOAHBChannels => 1
-  })
-
-class WithTL extends Config(
-  (pname, site, here) => pname match {
-    case TMemoryChannels     => BusType.TL
-    case NExtMMIOTLChannels  => 1
-  })
 
 class WithScratchpads extends Config(new WithNMemoryChannels(0) ++ new WithDataScratchpad(16384))
 
diff --git a/src/main/scala/rocketchip/ExampleTop.scala b/src/main/scala/rocketchip/ExampleTop.scala
index aef3dce9..5b17b242 100644
--- a/src/main/scala/rocketchip/ExampleTop.scala
+++ b/src/main/scala/rocketchip/ExampleTop.scala
@@ -9,45 +9,43 @@ import coreplex._
 import rocketchip._
 
 /** Example Top with Periphery */
-class ExampleTop(q: Parameters) extends BaseTop(q)
+class ExampleTop[+C <: BaseCoreplex](_coreplex: Parameters => C)(implicit p: Parameters) extends BaseTop(_coreplex)
     with PeripheryBootROM
     with PeripheryDebug
     with PeripheryExtInterrupts
-    with PeripheryCoreplexLocalInterrupter
     with PeripheryMasterMem
-    with PeripheryMasterMMIO
-    with PeripherySlave {
-  override lazy val module = Module(new ExampleTopModule(p, this, new ExampleTopBundle(p)))
+    with PeripheryMasterAXI4MMIO
+    with PeripherySlave
+    with DirectConnection {
+  override lazy val module = new ExampleTopModule(this, () => new ExampleTopBundle(this))
 }
 
-class ExampleTopBundle(p: Parameters) extends BaseTopBundle(p)
+class ExampleTopBundle[+L <: ExampleTop[BaseCoreplex]](_outer: L) extends BaseTopBundle(_outer)
     with PeripheryBootROMBundle
     with PeripheryDebugBundle
     with PeripheryExtInterruptsBundle
-    with PeripheryCoreplexLocalInterrupterBundle
     with PeripheryMasterMemBundle
-    with PeripheryMasterMMIOBundle
+    with PeripheryMasterAXI4MMIOBundle
     with PeripherySlaveBundle
 
-class ExampleTopModule[+L <: ExampleTop, +B <: ExampleTopBundle](p: Parameters, l: L, b: => B) extends BaseTopModule(p, l, b)
+class ExampleTopModule[+L <: ExampleTop[BaseCoreplex], +B <: ExampleTopBundle[L]](_outer: L, _io: () => B) extends BaseTopModule(_outer, _io)
     with PeripheryBootROMModule
     with PeripheryDebugModule
     with PeripheryExtInterruptsModule
-    with PeripheryCoreplexLocalInterrupterModule
     with PeripheryMasterMemModule
-    with PeripheryMasterMMIOModule
+    with PeripheryMasterAXI4MMIOModule
     with PeripherySlaveModule
     with HardwiredResetVector
-    with DirectConnection
+    with DirectConnectionModule
 
 /** Example Top with TestRAM */
-class ExampleTopWithTestRAM(q: Parameters) extends ExampleTop(q)
+class ExampleTopWithTestRAM[+C <: BaseCoreplex](_coreplex: Parameters => C)(implicit p: Parameters) extends ExampleTop(_coreplex)
     with PeripheryTestRAM {
-  override lazy val module = Module(new ExampleTopWithTestRAMModule(p, this, new ExampleTopWithTestRAMBundle(p)))
+  override lazy val module = new ExampleTopWithTestRAMModule(this, () => new ExampleTopWithTestRAMBundle(this))
 }
 
-class ExampleTopWithTestRAMBundle(p: Parameters) extends ExampleTopBundle(p)
+class ExampleTopWithTestRAMBundle[+L <: ExampleTopWithTestRAM[BaseCoreplex]](_outer: L) extends ExampleTopBundle(_outer)
     with PeripheryTestRAMBundle
 
-class ExampleTopWithTestRAMModule[+L <: ExampleTopWithTestRAM, +B <: ExampleTopWithTestRAMBundle](p: Parameters, l: L, b: => B) extends ExampleTopModule(p, l, b)
+class ExampleTopWithTestRAMModule[+L <: ExampleTopWithTestRAM[BaseCoreplex], +B <: ExampleTopWithTestRAMBundle[L]](_outer: L, _io: () => B) extends ExampleTopModule(_outer, _io)
     with PeripheryTestRAMModule
diff --git a/src/main/scala/rocketchip/Periphery.scala b/src/main/scala/rocketchip/Periphery.scala
index d0db2c2a..0b6783d8 100644
--- a/src/main/scala/rocketchip/Periphery.scala
+++ b/src/main/scala/rocketchip/Periphery.scala
@@ -9,6 +9,7 @@ import junctions.NastiConstants._
 import diplomacy._
 import uncore.tilelink._
 import uncore.tilelink2._
+import uncore.axi4._
 import uncore.converters._
 import uncore.devices._
 import uncore.agents._
@@ -29,19 +30,12 @@ object BusType {
 
 /** Memory channel controls */
 case object TMemoryChannels extends Field[BusType.EnumVal]
-/** External MMIO controls */
-case object NExtMMIOAXIChannels extends Field[Int]
-case object NExtMMIOAHBChannels extends Field[Int]
-case object NExtMMIOTLChannels  extends Field[Int]
 /** External Bus controls */
 case object NExtBusAXIChannels extends Field[Int]
 /** Async configurations */
 case object AsyncBusChannels extends Field[Boolean]
 case object AsyncDebugBus extends Field[Boolean]
 case object AsyncMemChannels extends Field[Boolean]
-case object AsyncMMIOChannels extends Field[Boolean]
-/** External address map settings */
-case object ExtMMIOPorts extends Field[Seq[AddrMapEntry]]
 /** Specifies the size of external memory */
 case object ExtMemSize extends Field[Long]
 /** Specifies the number of external interrupts */
@@ -89,10 +83,8 @@ trait HasPeripheryParameters {
   lazy val nMemAXIChannels = if (tMemChannels == BusType.AXI) nMemChannels else 0
   lazy val nMemAHBChannels = if (tMemChannels == BusType.AHB) nMemChannels else 0
   lazy val nMemTLChannels  = if (tMemChannels == BusType.TL)  nMemChannels else 0
-  lazy val outerMMIOParams = p.alterPartial({ case TLId => "L2toMMIO" })
   lazy val edgeSlaveParams = p.alterPartial({ case TLId => "EdgetoSlave" })
   lazy val edgeMemParams = p.alterPartial({ case TLId => "MCtoEdge" })
-  lazy val edgeMMIOParams = p.alterPartial({ case TLId => "MMIOtoEdge" })
   lazy val peripheryBusConfig = p(PeripheryBusKey)
   lazy val socBusConfig = p(SOCBusKey)
   lazy val cacheBlockBytes = p(CacheBlockBytes)
@@ -100,12 +92,14 @@ trait HasPeripheryParameters {
 
 /////
 
-trait PeripheryDebug extends LazyModule {
-  implicit val p: Parameters
+trait PeripheryDebug {
+  this: TopNetwork =>
 }
 
 trait PeripheryDebugBundle {
-  implicit val p: Parameters
+  this: TopNetworkBundle {
+    val outer: PeripheryDebug
+  } =>
   val debug_clk = (p(AsyncDebugBus) && !p(IncludeJtagDTM)).option(Clock(INPUT))
   val debug_rst = (p(AsyncDebugBus) && !p(IncludeJtagDTM)).option(Bool(INPUT))
   val debug = (!p(IncludeJtagDTM)).option(new DebugBusIO()(p).flip)
@@ -113,19 +107,19 @@ trait PeripheryDebugBundle {
 }
 
 trait PeripheryDebugModule {
-  implicit val p: Parameters
-  val outer: PeripheryDebug
-  val io: PeripheryDebugBundle
-  val coreplexIO: BaseCoreplexBundle
+  this: TopNetworkModule {
+    val outer: PeripheryDebug
+    val io: PeripheryDebugBundle
+  } =>
 
   if (p(IncludeJtagDTM)) {
     // JtagDTMWithSync is a wrapper which
     // handles the synchronization as well.
     val dtm = Module (new JtagDTMWithSync()(p))
     dtm.io.jtag <> io.jtag.get
-    coreplexIO.debug <> dtm.io.debug
+    coreplexDebug <> dtm.io.debug
   } else {
-    coreplexIO.debug <>
+    coreplexDebug <>
       (if (p(AsyncDebugBus)) AsyncDebugBusFrom(io.debug_clk.get, io.debug_rst.get, io.debug.get)
       else io.debug.get)
   }
@@ -133,40 +127,40 @@ trait PeripheryDebugModule {
 
 /////
 
-trait PeripheryExtInterrupts extends LazyModule {
-  implicit val p: Parameters
-  val pInterrupts: RangeManager
+trait PeripheryExtInterrupts {
+  this: TopNetwork =>
 
-  pInterrupts.add("ext", p(NExtTopInterrupts))
+  val extInterrupts = IntBlindInputNode(p(NExtTopInterrupts))
+  val extInterruptXing = LazyModule(new IntXing)
+
+  intBus.intnode := extInterruptXing.intnode
+  extInterruptXing.intnode := extInterrupts
 }
 
 trait PeripheryExtInterruptsBundle {
-  implicit val p: Parameters
-  val interrupts = Vec(p(NExtTopInterrupts), Bool()).asInput
+  this: TopNetworkBundle {
+    val outer: PeripheryExtInterrupts
+  } =>
+  val interrupts = outer.extInterrupts.bundleIn
 }
 
 trait PeripheryExtInterruptsModule {
-  implicit val p: Parameters
-  val outer: PeripheryExtInterrupts
-  val io: PeripheryExtInterruptsBundle
-  val coreplexIO: BaseCoreplexBundle
-
-  {
-    val r = outer.pInterrupts.range("ext")
-    ((r._1 until r._2) zipWithIndex) foreach { case (c, i) =>
-      coreplexIO.interrupts(c) := io.interrupts(i)
-    }
-  }
+  this: TopNetworkModule {
+    val outer: PeripheryExtInterrupts
+    val io: PeripheryExtInterruptsBundle
+  } =>
 }
 
 /////
 
-trait PeripheryMasterMem extends LazyModule {
-  implicit val p: Parameters
+trait PeripheryMasterMem {
+  this: TopNetwork =>
 }
 
-trait PeripheryMasterMemBundle extends HasPeripheryParameters {
-  implicit val p: Parameters
+trait PeripheryMasterMemBundle {
+  this: TopNetworkBundle {
+    val outer: PeripheryMasterMem
+  } =>
   val mem_clk = p(AsyncMemChannels).option(Vec(nMemChannels, Clock(INPUT)))
   val mem_rst = p(AsyncMemChannels).option(Vec(nMemChannels, Bool (INPUT)))
   val mem_axi = Vec(nMemAXIChannels, new NastiIO)
@@ -174,13 +168,13 @@ trait PeripheryMasterMemBundle extends HasPeripheryParameters {
   val mem_tl = Vec(nMemTLChannels, new ClientUncachedTileLinkIO()(edgeMemParams))
 }
 
-trait PeripheryMasterMemModule extends HasPeripheryParameters {
-  implicit val p: Parameters
-  val outer: PeripheryMasterMem
-  val io: PeripheryMasterMemBundle
-  val coreplexIO: BaseCoreplexBundle
+trait PeripheryMasterMemModule {
+  this: TopNetworkModule {
+    val outer: PeripheryMasterMem
+    val io: PeripheryMasterMemBundle
+  } =>
 
-  val edgeMem = coreplexIO.master.mem.map(TileLinkWidthAdapter(_, edgeMemParams))
+  val edgeMem = coreplexMem.map(TileLinkWidthAdapter(_, edgeMemParams))
 
   // Abuse the fact that zip takes the shorter of the two lists
   ((io.mem_axi zip edgeMem) zipWithIndex) foreach { case ((axi, mem), idx) =>
@@ -204,78 +198,65 @@ trait PeripheryMasterMemModule extends HasPeripheryParameters {
 
 /////
 
-trait PeripheryMasterMMIO extends LazyModule {
-  implicit val p: Parameters
+// PeripheryMasterAXI4MMIO is an example, make your own cake pattern like this one.
+trait PeripheryMasterAXI4MMIO {
+  this: TopNetwork =>
+
+  val mmio_axi4 = AXI4BlindOutputNode(AXI4SlavePortParameters(
+    slaves = Seq(AXI4SlaveParameters(
+      address       = List(AddressSet(0x60000000L, 0x1fffffffL)),
+      executable    = true,                  // Can we run programs on this memory?
+      supportsWrite = TransferSizes(1, 256), // The slave supports 1-256 byte transfers
+      supportsRead  = TransferSizes(1, 256),
+      interleavedId = Some(0))),             // slave does not interleave read responses
+    beatBytes = 8)) // 64-bit AXI interface
+
+  mmio_axi4 :=
+    // AXI4Fragmenter(lite=false, maxInFlight = 20)( // beef device up to support awlen = 0xff
+    TLToAXI4(idBits = 4)(                  // use idBits = 0 for AXI4-Lite
+    TLWidthWidget(socBusConfig.beatBytes)( // convert width before attaching to socBus
+    socBus.node))
 }
 
-trait PeripheryMasterMMIOBundle extends HasPeripheryParameters {
-  implicit val p: Parameters
-  val mmio_clk = p(AsyncMMIOChannels).option(Vec(p(NExtMMIOAXIChannels), Clock(INPUT)))
-  val mmio_rst = p(AsyncMMIOChannels).option(Vec(p(NExtMMIOAXIChannels), Bool (INPUT)))
-  val mmio_axi = Vec(p(NExtMMIOAXIChannels), new NastiIO)
-  val mmio_ahb = Vec(p(NExtMMIOAHBChannels), new HastiMasterIO)
-  val mmio_tl = Vec(p(NExtMMIOTLChannels), new ClientUncachedTileLinkIO()(edgeMMIOParams))
+trait PeripheryMasterAXI4MMIOBundle {
+  this: TopNetworkBundle {
+    val outer: PeripheryMasterAXI4MMIO
+  } =>
+  val mmio_axi = outer.mmio_axi4.bundleOut
 }
 
-trait PeripheryMasterMMIOModule extends HasPeripheryParameters {
-  implicit val p: Parameters
-  val outer: PeripheryMasterMMIO
-  val io: PeripheryMasterMMIOBundle
-  val pBus: TileLinkRecursiveInterconnect
-
-  val mmio_ports = p(ExtMMIOPorts) map { port =>
-    TileLinkWidthAdapter(pBus.port(port.name), edgeMMIOParams)
-  }
-
-  val mmio_axi_start = 0
-  val mmio_axi_end   = mmio_axi_start + p(NExtMMIOAXIChannels)
-  val mmio_ahb_start = mmio_axi_end
-  val mmio_ahb_end   = mmio_ahb_start + p(NExtMMIOAHBChannels)
-  val mmio_tl_start  = mmio_ahb_end
-  val mmio_tl_end    = mmio_tl_start  + p(NExtMMIOTLChannels)
-  require (mmio_tl_end == mmio_ports.size)
-
-  for (i <- 0 until mmio_ports.size) {
-    if (mmio_axi_start <= i && i < mmio_axi_end) {
-      val idx = i-mmio_axi_start
-      val axi_sync = PeripheryUtils.convertTLtoAXI(mmio_ports(i))
-      io.mmio_axi(idx) <> (
-        if (!p(AsyncMMIOChannels)) axi_sync
-        else AsyncNastiTo(io.mmio_clk.get(idx), io.mmio_rst.get(idx), axi_sync)
-      )
-    } else if (mmio_ahb_start <= i && i < mmio_ahb_end) {
-      val idx = i-mmio_ahb_start
-      io.mmio_ahb(idx) <> PeripheryUtils.convertTLtoAHB(mmio_ports(i), atomics = true)
-    } else if (mmio_tl_start <= i && i < mmio_tl_end) {
-      val idx = i-mmio_tl_start
-      io.mmio_tl(idx) <> TileLinkEnqueuer(mmio_ports(i), 2)
-    } else {
-      require(false, "Unconnected external MMIO port")
-    }
-  }
+trait PeripheryMasterAXI4MMIOModule {
+  this: TopNetworkModule {
+    val outer: PeripheryMasterAXI4MMIO
+    val io: PeripheryMasterAXI4MMIOBundle
+  } =>
+  // nothing to do
 }
 
 /////
 
-trait PeripherySlave extends LazyModule {
-  implicit val p: Parameters
-  val pBusMasters: RangeManager
+trait PeripherySlave {
+  this: TopNetwork {
+    val pBusMasters: RangeManager
+  } =>
 
   if (p(NExtBusAXIChannels) > 0) pBusMasters.add("ext", 1) // NExtBusAXIChannels are arbitrated into one TL port
 }
 
-trait PeripherySlaveBundle extends HasPeripheryParameters {
-  implicit val p: Parameters
+trait PeripherySlaveBundle {
+  this: TopNetworkBundle {
+    val outer: PeripherySlave
+  } =>
   val bus_clk = p(AsyncBusChannels).option(Vec(p(NExtBusAXIChannels), Clock(INPUT)))
   val bus_rst = p(AsyncBusChannels).option(Vec(p(NExtBusAXIChannels), Bool (INPUT)))
   val bus_axi = Vec(p(NExtBusAXIChannels), new NastiIO).flip
 }
 
-trait PeripherySlaveModule extends HasPeripheryParameters {
-  implicit val p: Parameters
-  val outer: PeripherySlave
-  val io: PeripherySlaveBundle
-  val coreplexIO: BaseCoreplexBundle
+trait PeripherySlaveModule {
+  this: TopNetworkModule {
+    val outer: PeripherySlave { val pBusMasters: RangeManager }
+    val io: PeripherySlaveBundle
+  } =>
 
   if (p(NExtBusAXIChannels) > 0) {
     val arb = Module(new NastiArbiter(p(NExtBusAXIChannels)))
@@ -290,105 +271,82 @@ trait PeripherySlaveModule extends HasPeripheryParameters {
 
     val (r_start, r_end) = outer.pBusMasters.range("ext")
     require(r_end - r_start == 1, "RangeManager should return 1 slot")
-    TileLinkWidthAdapter(coreplexIO.slave(r_start), conv.io.tl)
+    TileLinkWidthAdapter(coreplexSlave(r_start), conv.io.tl)
   }
 }
 
 /////
 
-trait PeripheryCoreplexLocalInterrupter extends LazyModule with HasPeripheryParameters {
-  implicit val p: Parameters
-  val peripheryBus: TLXbar
+trait PeripheryBootROM {
+  this: TopNetwork =>
 
-  // CoreplexLocalInterrupter must be at least 64b if XLen >= 64
-  val beatBytes = max((outerMMIOParams(XLen) min 64) / 8, peripheryBusConfig.beatBytes)
-  val clintConfig = CoreplexLocalInterrupterConfig(beatBytes)
-  val clint = LazyModule(new CoreplexLocalInterrupter(clintConfig)(outerMMIOParams))
-  // The periphery bus is 32-bit, so we may need to adapt its width to XLen
-  clint.node := TLFragmenter(beatBytes, cacheBlockBytes)(TLWidthWidget(peripheryBusConfig.beatBytes)(peripheryBus.node))
-}
-
-trait PeripheryCoreplexLocalInterrupterBundle {
-  implicit val p: Parameters
-}
-
-trait PeripheryCoreplexLocalInterrupterModule extends HasPeripheryParameters {
-  implicit val p: Parameters
-  val outer: PeripheryCoreplexLocalInterrupter
-  val io: PeripheryCoreplexLocalInterrupterBundle
-  val coreplexIO: BaseCoreplexBundle
-
-  outer.clint.module.io.rtcTick := Counter(p(RTCPeriod)).inc()
-  coreplexIO.clint <> outer.clint.module.io.tiles
-}
-
-/////
-
-trait PeripheryBootROM extends LazyModule with HasPeripheryParameters {
-  implicit val p: Parameters
-  val peripheryBus: TLXbar
-
-  val address = 0x1000
-  val size = 0x1000
-  val bootrom = LazyModule(new TLROM(address, size, GenerateBootROM(p, address), true, peripheryBusConfig.beatBytes))
+  val bootrom_address = 0x1000
+  val bootrom_size = 0x1000
+  val bootrom = LazyModule(new TLROM(bootrom_address, bootrom_size, GenerateBootROM(p, bootrom_address), true, peripheryBusConfig.beatBytes))
   bootrom.node := TLFragmenter(peripheryBusConfig.beatBytes, cacheBlockBytes)(peripheryBus.node)
 }
 
 trait PeripheryBootROMBundle {
-  implicit val p: Parameters
+  this: TopNetworkBundle {
+    val outer: PeripheryBootROM
+  } =>
 }
 
-trait PeripheryBootROMModule extends HasPeripheryParameters {
-  implicit val p: Parameters
-  val outer: PeripheryBootROM
-  val io: PeripheryBootROMBundle
+trait PeripheryBootROMModule {
+  this: TopNetworkModule {
+    val outer: PeripheryBootROM
+    val io: PeripheryBootROMBundle
+  } =>
 }
 
 /////
 
-trait PeripheryTestRAM extends LazyModule with HasPeripheryParameters {
-  implicit val p: Parameters
-  val peripheryBus: TLXbar
+trait PeripheryTestRAM {
+  this: TopNetwork =>
 
-  val ramBase = 0x52000000
-  val ramSize = 0x1000
-
-  val sram = LazyModule(new TLRAM(AddressSet(ramBase, ramSize-1), true, peripheryBusConfig.beatBytes)
-                        { override def name = "testram" })
-  sram.node := TLFragmenter(peripheryBusConfig.beatBytes, cacheBlockBytes)(peripheryBus.node)
+  val testram = LazyModule(new TLRAM(AddressSet(0x52000000, 0xfff), true, peripheryBusConfig.beatBytes))
+  testram.node := TLFragmenter(peripheryBusConfig.beatBytes, cacheBlockBytes)(peripheryBus.node)
 }
 
 trait PeripheryTestRAMBundle {
-  implicit val p: Parameters
+  this: TopNetworkBundle {
+    val outer: PeripheryTestRAM
+  } =>
 }
 
-trait PeripheryTestRAMModule extends HasPeripheryParameters {
-  implicit val p: Parameters
-  val outer: PeripheryTestRAM
+trait PeripheryTestRAMModule {
+  this: TopNetworkModule {
+    val outer: PeripheryTestRAM
+    val io: PeripheryTestRAMBundle
+  } =>
 }
 
 /////
 
-trait PeripheryTestBusMaster extends LazyModule {
-  implicit val p: Parameters
-  val peripheryBus: TLXbar
-
+trait PeripheryTestBusMaster {
+  this: TopNetwork =>
   val fuzzer = LazyModule(new TLFuzzer(5000))
   peripheryBus.node := fuzzer.node
 }
 
 trait PeripheryTestBusMasterBundle {
-  implicit val p: Parameters
+  this: TopNetworkBundle {
+    val outer: PeripheryTestBusMaster
+  } =>
 }
 
 trait PeripheryTestBusMasterModule {
-  implicit val p: Parameters
-  val outer: PeripheryTestBusMaster
+  this: TopNetworkModule {
+    val outer: PeripheryTestBusMaster
+    val io: PeripheryTestBusMasterBundle
+  } =>
 }
 
 /////
 
 trait HardwiredResetVector {
-  val coreplexIO: BaseCoreplexBundle
-  coreplexIO.resetVector := UInt(0x1000) // boot ROM
+  this: TopNetworkModule {
+    val outer: BaseTop[BaseCoreplex]
+  } =>
+  outer.coreplex.module.io.resetVector := UInt(0x1000) // boot ROM
 }
diff --git a/src/main/scala/rocketchip/TestHarness.scala b/src/main/scala/rocketchip/TestHarness.scala
index f09c9b50..5d07d1cd 100644
--- a/src/main/scala/rocketchip/TestHarness.scala
+++ b/src/main/scala/rocketchip/TestHarness.scala
@@ -8,14 +8,14 @@ import junctions._
 import junctions.NastiConstants._
 import util.LatencyPipe
 
-case object BuildExampleTop extends Field[Parameters => ExampleTop]
+case object BuildExampleTop extends Field[Parameters => ExampleTop[coreplex.BaseCoreplex]]
 case object SimMemLatency extends Field[Int]
 
 class TestHarness(q: Parameters) extends Module {
   val io = new Bundle {
     val success = Bool(OUTPUT)
   }
-  val dut = q(BuildExampleTop)(q).module
+  val dut = Module(q(BuildExampleTop)(q).module)
   implicit val p = dut.p
 
   // This test harness isn't especially flexible yet
@@ -25,16 +25,12 @@ class TestHarness(q: Parameters) extends Module {
   require(dut.io.mem_tl.isEmpty)
   require(dut.io.bus_clk.isEmpty)
   require(dut.io.bus_rst.isEmpty)
-  require(dut.io.mmio_clk.isEmpty)
-  require(dut.io.mmio_rst.isEmpty)
-  require(dut.io.mmio_ahb.isEmpty)
-  require(dut.io.mmio_tl.isEmpty)
 
-  for (int <- dut.io.interrupts)
+  for (int <- dut.io.interrupts(0))
     int := Bool(false)
 
   if (dut.io.mem_axi.nonEmpty) {
-    val memSize = p(GlobalAddrMap)("mem").size
+    val memSize = p(ExtMemSize)
     require(memSize % dut.io.mem_axi.size == 0)
     for (axi <- dut.io.mem_axi) {
       val mem = Module(new SimAXIMem(memSize / dut.io.mem_axi.size))
diff --git a/src/main/scala/rocketchip/Utils.scala b/src/main/scala/rocketchip/Utils.scala
index 69a7d338..d4839120 100644
--- a/src/main/scala/rocketchip/Utils.scala
+++ b/src/main/scala/rocketchip/Utils.scala
@@ -53,20 +53,8 @@ class GlobalVariable[T] {
 }
 
 object GenerateGlobalAddrMap {
-  def apply(p: Parameters, pDevicesEntries: Seq[AddrMapEntry], peripheryManagers: Seq[TLManagerParameters]) = {
-    lazy val cBusIOAddrMap: AddrMap = {
-      val entries = collection.mutable.ArrayBuffer[AddrMapEntry]()
-      entries += AddrMapEntry("debug", MemSize(4096, MemAttr(AddrMapProt.RWX)))
-      entries += AddrMapEntry("plic", MemRange(0x0C000000, 0x4000000, MemAttr(AddrMapProt.RW)))
-      if (p(DataScratchpadSize) > 0) { // TODO heterogeneous tiles
-        require(p(NTiles) == 1) // TODO relax this
-        require(p(NMemoryChannels) == 0) // TODO allow both scratchpad & DRAM
-        entries += AddrMapEntry("dmem0", MemRange(0x80000000L, BigInt(p(DataScratchpadSize)), MemAttr(AddrMapProt.RWX)))
-      }
-      new AddrMap(entries)
-    }
-
-    lazy val tl2Devices = peripheryManagers.map { manager =>
+  def apply(p: Parameters, peripheryManagers: Seq[TLManagerParameters]) = {
+    val tl2Devices = peripheryManagers.map { manager =>
       val cacheable = manager.regionType match {
         case RegionType.CACHED   => true
         case RegionType.TRACKED  => true
@@ -84,41 +72,28 @@ object GenerateGlobalAddrMap {
       }
     }.flatten
 
-    lazy val uniquelyNamedTL2Devices =
+    val uniquelyNamedTL2Devices =
       tl2Devices.groupBy(_.name).values.map(_.zipWithIndex.map {
         case (e, i) => if (i == 0) e else e.copy(name = e.name + "_" + i)
       }).flatten.toList
 
-    lazy val tl2AddrMap = new AddrMap(uniquelyNamedTL2Devices, collapse = true)
-    lazy val pBusIOAddrMap = new AddrMap(AddrMapEntry("TL2", tl2AddrMap) +: (p(ExtMMIOPorts) ++ pDevicesEntries), collapse = true)
-
     val memBase = 0x80000000L
     val memSize = p(ExtMemSize)
     Dump("MEM_BASE", memBase)
 
-    val cBus = AddrMapEntry("cbus", cBusIOAddrMap)
-    val pBus = AddrMapEntry("pbus", pBusIOAddrMap)
-    val io = AddrMapEntry("io", AddrMap((cBus +: (!pBusIOAddrMap.isEmpty).option(pBus).toSeq):_*))
+    val tl2 = AddrMapEntry("TL2", new AddrMap(uniquelyNamedTL2Devices, collapse = true))
     val mem = AddrMapEntry("mem", MemRange(memBase, memSize, MemAttr(AddrMapProt.RWX, true)))
-    AddrMap((io +: (p(NMemoryChannels) > 0).option(mem).toSeq):_*)
+    AddrMap((tl2 +: (p(NMemoryChannels) > 0).option(mem).toSeq):_*)
   }
 }
 
 object GenerateConfigString {
-  def apply(p: Parameters, c: CoreplexConfig, pDevicesEntries: Seq[AddrMapEntry], peripheryManagers: Seq[TLManagerParameters]) = {
+  def apply(p: Parameters, clint: CoreplexLocalInterrupter, plic: TLPLIC, peripheryManagers: Seq[TLManagerParameters]) = {
+    val c = CoreplexParameters()(p)
     val addrMap = p(GlobalAddrMap)
-    val plicAddr = addrMap("io:cbus:plic").start
-    val clint = CoreplexLocalInterrupterConfig(0, addrMap("io:pbus:TL2:clint").start)
-    val xLen = p(XLen)
     val res = new StringBuilder
-    res append  "plic {\n"
-    res append s"  priority 0x${plicAddr.toString(16)};\n"
-    res append s"  pending 0x${(plicAddr + c.plicKey.pendingBase).toString(16)};\n"
-    res append s"  ndevs ${c.plicKey.nDevices};\n"
-    res append  "};\n"
-    res append  "rtc {\n"
-    res append s"  addr 0x${clint.timeAddress.toString(16)};\n"
-    res append  "};\n"
+    res append plic.module.globalConfigString
+    res append clint.module.globalConfigString
     if (addrMap contains "mem") {
       res append  "ram {\n"
       res append  "  0 {\n"
@@ -140,33 +115,12 @@ object GenerateConfigString {
       res append s"  $i {\n"
       res append  "    0 {\n"
       res append s"      isa $isa;\n"
-      res append s"      timecmp 0x${clint.timecmpAddress(i).toString(16)};\n"
-      res append s"      ipi 0x${clint.msipAddress(i).toString(16)};\n"
-      res append s"      plic {\n"
-      res append s"        m {\n"
-      res append s"         ie 0x${(plicAddr + c.plicKey.enableAddr(i, 'M')).toString(16)};\n"
-      res append s"         thresh 0x${(plicAddr + c.plicKey.threshAddr(i, 'M')).toString(16)};\n"
-      res append s"         claim 0x${(plicAddr + c.plicKey.claimAddr(i, 'M')).toString(16)};\n"
-      res append s"        };\n"
-      if (c.hasSupervisor) {
-        res append s"        s {\n"
-        res append s"         ie 0x${(plicAddr + c.plicKey.enableAddr(i, 'S')).toString(16)};\n"
-        res append s"         thresh 0x${(plicAddr + c.plicKey.threshAddr(i, 'S')).toString(16)};\n"
-        res append s"         claim 0x${(plicAddr + c.plicKey.claimAddr(i, 'S')).toString(16)};\n"
-        res append s"        };\n"
-      }
-      res append  "      };\n"
+      res append clint.module.hartConfigStrings(i)
+      res append plic.module.hartConfigStrings(i)
       res append  "    };\n"
       res append  "  };\n"
     }
     res append  "};\n"
-    pDevicesEntries foreach { entry =>
-      val region = addrMap("io:pbus:" + entry.name)
-      res append s"${entry.name} {\n"
-      res append s"  addr 0x${region.start.toString(16)};\n"
-      res append s"  size 0x${region.size.toString(16)}; \n"
-      res append  "}\n"
-    }
     peripheryManagers.foreach { manager => res append manager.dts }
     res append '\u0000'
     res.toString
@@ -185,6 +139,6 @@ object GenerateBootROM {
     require(rom.getInt(12) == 0,
       "Config string address position should not be occupied by code")
     rom.putInt(12, configStringAddr)
-    rom.array() ++ (p(ConfigString).getBytes.toSeq)
+    rom.array() ++ (ConfigStringOutput.contents.get.getBytes.toSeq)
   }
 }
diff --git a/src/main/scala/uncore/axi4/Fragmenter.scala b/src/main/scala/uncore/axi4/Fragmenter.scala
index 939555cb..e933270f 100644
--- a/src/main/scala/uncore/axi4/Fragmenter.scala
+++ b/src/main/scala/uncore/axi4/Fragmenter.scala
@@ -10,7 +10,7 @@ import scala.math.{min,max}
 import uncore.tilelink2.{leftOR, rightOR, UIntToOH1, OH1ToOH}
 
 // lite: masters all use only one ID => reads will not be interleaved
-class AXI4Fragmenter(lite: Boolean = false, maxInFlight: Int = 32, combinational: Boolean = true) extends LazyModule
+class AXI4Fragmenter(lite: Boolean = false, maxInFlight: => Int = 32, combinational: Boolean = true) extends LazyModule
 {
   val maxBeats = 1 << AXI4Parameters.lenBits
   def expandTransfer(x: TransferSizes, beatBytes: Int, alignment: BigInt) =
@@ -287,7 +287,7 @@ class AXI4FragmenterSideband(maxInFlight: Int, flow: Boolean = false) extends Mo
 object AXI4Fragmenter
 {
   // applied to the AXI4 source node; y.node := AXI4Fragmenter()(x.node)
-  def apply(lite: Boolean = false, maxInFlight: Int = 32, combinational: Boolean = true)(x: AXI4OutwardNode)(implicit sourceInfo: SourceInfo): AXI4OutwardNode = {
+  def apply(lite: Boolean = false, maxInFlight: => Int = 32, combinational: Boolean = true)(x: AXI4OutwardNode)(implicit sourceInfo: SourceInfo): AXI4OutwardNode = {
     val fragmenter = LazyModule(new AXI4Fragmenter(lite, maxInFlight, combinational))
     fragmenter.node := x
     fragmenter.node
diff --git a/src/main/scala/uncore/axi4/Nodes.scala b/src/main/scala/uncore/axi4/Nodes.scala
index 0613a70b..b8fb1efd 100644
--- a/src/main/scala/uncore/axi4/Nodes.scala
+++ b/src/main/scala/uncore/axi4/Nodes.scala
@@ -16,10 +16,13 @@ object AXI4Imp extends NodeImp[AXI4MasterPortParameters, AXI4SlavePortParameters
   }
   def bundleI(ei: Seq[AXI4EdgeParameters]): Vec[AXI4Bundle] = {
     require (!ei.isEmpty)
-    Vec(ei.size, AXI4Bundle(ei.map(_.bundle).reduce(_.union(_)))).flip
+    Vec(ei.size, AXI4Bundle(ei.map(_.bundle).reduce(_.union(_))))
   }
 
   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 connect(bo: => AXI4Bundle, bi: => AXI4Bundle, ei: => AXI4EdgeParameters)(implicit sourceInfo: SourceInfo): (Option[LazyModule], () => Unit) = {
     (None, () => { bi <> bo })
   }
@@ -30,18 +33,23 @@ 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 AXI4OutputNode() extends OutputNode(AXI4Imp)
-case class AXI4InputNode() extends InputNode(AXI4Imp)
-
 case class AXI4MasterNode(portParams: AXI4MasterPortParameters, numPorts: Range.Inclusive = 1 to 1)
   extends SourceNode(AXI4Imp)(portParams, numPorts)
 case class AXI4SlaveNode(portParams: AXI4SlavePortParameters, numPorts: Range.Inclusive = 1 to 1)
   extends SinkNode(AXI4Imp)(portParams, numPorts)
-
 case class AXI4AdapterNode(
   masterFn:       Seq[AXI4MasterPortParameters]  => AXI4MasterPortParameters,
   slaveFn:        Seq[AXI4SlavePortParameters] => AXI4SlavePortParameters,
   numMasterPorts: Range.Inclusive = 1 to 1,
   numSlavePorts:  Range.Inclusive = 1 to 1)
   extends InteriorNode(AXI4Imp)(masterFn, slaveFn, numMasterPorts, numSlavePorts)
+
+// 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: AXI4SlavePortParameters) extends BlindOutputNode(AXI4Imp)(portParams)
+case class AXI4BlindInputNode(portParams: AXI4MasterPortParameters) extends BlindInputNode(AXI4Imp)(portParams)
diff --git a/src/main/scala/uncore/axi4/RegisterRouter.scala b/src/main/scala/uncore/axi4/RegisterRouter.scala
index 97b598d5..9f408c9a 100644
--- a/src/main/scala/uncore/axi4/RegisterRouter.scala
+++ b/src/main/scala/uncore/axi4/RegisterRouter.scala
@@ -7,10 +7,11 @@ import diplomacy._
 import regmapper._
 import scala.math.{min,max}
 
-class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true)
+class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)
   extends AXI4SlaveNode(AXI4SlavePortParameters(
     Seq(AXI4SlaveParameters(
       address       = Seq(address),
+      executable    = executable,
       supportsWrite = TransferSizes(1, beatBytes),
       supportsRead  = TransferSizes(1, beatBytes),
       interleavedId = Some(0))),
@@ -69,16 +70,16 @@ class AXI4RegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int
 
 object AXI4RegisterNode
 {
-  def apply(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true) =
-    new AXI4RegisterNode(address, concurrency, beatBytes, undefZero)
+  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.
 
-abstract class AXI4RegisterRouterBase(address: AddressSet, interrupts: Int, concurrency: Int, beatBytes: Int, undefZero: Boolean) extends LazyModule
+abstract class AXI4RegisterRouterBase(address: AddressSet, interrupts: Int, concurrency: Int, beatBytes: Int, undefZero: Boolean, executable: Boolean) extends LazyModule
 {
-  val node = AXI4RegisterNode(address, concurrency, beatBytes, undefZero)
+  val node = AXI4RegisterNode(address, concurrency, beatBytes, undefZero, executable)
   val intnode = uncore.tilelink2.IntSourceNode(interrupts)
 }
 
@@ -101,10 +102,10 @@ class AXI4RegModule[P, B <: AXI4RegBundleBase](val params: P, bundleBuilder: =>
 }
 
 class AXI4RegisterRouter[B <: AXI4RegBundleBase, M <: LazyModuleImp]
-   (val base: BigInt, val interrupts: Int = 0, val size: BigInt = 4096, val concurrency: Int = 0, val beatBytes: Int = 4, undefZero: Boolean = true)
+   (val base: BigInt, val interrupts: Int = 0, val size: BigInt = 4096, val concurrency: Int = 0, val beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)
    (bundleBuilder: AXI4RegBundleArg => B)
    (moduleBuilder: (=> B, AXI4RegisterRouterBase) => M)
-  extends AXI4RegisterRouterBase(AddressSet(base, size-1), interrupts, concurrency, beatBytes, undefZero)
+  extends AXI4RegisterRouterBase(AddressSet(base, size-1), interrupts, concurrency, beatBytes, undefZero, executable)
 {
   require (isPow2(size))
   // require (size >= 4096) ... not absolutely required, but highly recommended
diff --git a/src/main/scala/uncore/devices/Debug.scala b/src/main/scala/uncore/devices/Debug.scala
index 9997f493..3ee83e9a 100644
--- a/src/main/scala/uncore/devices/Debug.scala
+++ b/src/main/scala/uncore/devices/Debug.scala
@@ -4,8 +4,9 @@ package uncore.devices
 
 import Chisel._
 import junctions._
-import uncore.tilelink._
 import util._
+import regmapper._
+import uncore.tilelink2._
 import cde.{Parameters, Config, Field}
 
 // *****************************************
@@ -15,6 +16,7 @@ import cde.{Parameters, Config, Field}
 
 object DbRegAddrs{
 
+  def DMRAMBASE    = UInt(0x0)
   def DMCONTROL    = UInt(0x10)
 
   def DMINFO       = UInt(0x11)
@@ -69,7 +71,7 @@ object DsbBusConsts {
   // See $RISCV/riscv-tools/riscv-isa-sim/debug_rom/debug_rom.h/S
   // The code assumes 64 bytes of Debug RAM.
 
-  def defaultRomContents : Array[Byte] = Array(
+  def xlenAnyRomContents : Array[Byte] = Array(
     0x6f, 0x00, 0xc0, 0x04, 0x6f, 0x00, 0xc0, 0x00, 0x13, 0x04, 0xf0, 0xff,
     0x6f, 0x00, 0x80, 0x00, 0x13, 0x04, 0x00, 0x00, 0x0f, 0x00, 0xf0, 0x0f,
     0xf3, 0x24, 0x00, 0xf1, 0x63, 0xc6, 0x04, 0x00, 0x83, 0x24, 0xc0, 0x43,
@@ -121,10 +123,10 @@ object DsbBusConsts {
 
 object DsbRegAddrs{
 
-  def CLEARDEBINT  = UInt(0x100)
-  def SETHALTNOT   = UInt(0x10C)
-  def SERINFO      = UInt(0x110)
-  def SERBASE      = UInt(0x114)
+  def CLEARDEBINT  = 0x100
+  def SETHALTNOT   = 0x10C
+  def SERINFO      = 0x110
+  def SERBASE      = 0x114
   // For each serial, there are
   // 3 registers starting here:
   // SERSEND0
@@ -132,9 +134,11 @@ object DsbRegAddrs{
   // SERSTATUS0
   // ...
   // SERSTATUS7
-  def SERTX_OFFSET   = UInt(0)
-  def SERRX_OFFSET   = UInt(4)
-  def SERSTAT_OFFSET = UInt(8)
+  def SERTX_OFFSET   = 0
+  def SERRX_OFFSET   = 4
+  def SERSTAT_OFFSET = 8
+  def RAMBASE      = 0x400
+  def ROMBASE      = 0x800
 
 }
 
@@ -302,6 +306,24 @@ class DebugBusIO(implicit val p: cde.Parameters) extends ParameterizedBundle()(p
   val resp = new DecoupledIO(new DebugBusResp).flip()
 }
 
+trait HasDebugModuleParameters {
+  val params : Parameters
+  implicit val p = params
+  val cfg = p(DMKey)
+}
+
+/** Debug Module I/O, with the exclusion of the RegisterRouter
+  *  Access interface.
+  */
+
+trait DebugModuleBundle extends Bundle with HasDebugModuleParameters {
+  val db = new DebugBusIO()(p).flip()
+  val debugInterrupts = Vec(cfg.nComponents, Bool()).asOutput
+  val ndreset    = Bool(OUTPUT)
+  val fullreset  = Bool(OUTPUT)
+}
+
+
 // *****************************************
 // The Module 
 // 
@@ -313,7 +335,7 @@ class DebugBusIO(implicit val p: cde.Parameters) extends ParameterizedBundle()(p
   *  DebugModule is a slave to two masters:
   *    The Debug Bus -- implemented as a generic Decoupled IO with request
   *                   and response channels
-  *    The System Bus -- implemented as Uncached Tile Link.
+  *    The System Bus -- implemented as generic RegisterRouter
   *  
   *  DebugModule is responsible for holding registers, RAM, and ROM
   *      to support debug interactions, as well as driving interrupts
@@ -321,10 +343,9 @@ class DebugBusIO(implicit val p: cde.Parameters) extends ParameterizedBundle()(p
   *      It is also responsible for some reset lines.
   */
 
-class DebugModule ()(implicit val p:cde.Parameters)
-    extends Module
-    with HasTileLinkParameters { 
-  val cfg = p(DMKey)
+trait DebugModule extends Module with HasDebugModuleParameters with HasRegMap {
+
+  val io: DebugModuleBundle
 
   //--------------------------------------------------------------
   // Import constants for shorter variable names
@@ -344,6 +365,7 @@ class DebugModule ()(implicit val p:cde.Parameters)
   require (cfg.hasBusMaster == false)
   require (cfg.nDebugRamBytes <= 64)
   require (cfg.authType == DebugModuleAuthType.None)
+  require((DbBusConsts.dbRamWordBits % 8) == 0)
 
   //--------------------------------------------------------------
   // Private Classes (Register Fields)
@@ -403,17 +425,6 @@ class DebugModule ()(implicit val p:cde.Parameters)
     
   }
 
-  //--------------------------------------------------------------
-  // Module I/O
-  //--------------------------------------------------------------
-
-  val io = new Bundle {
-    val db = new DebugBusIO()(p).flip()
-    val debugInterrupts = Vec(cfg.nComponents, Bool()).asOutput
-    val tl = new ClientUncachedTileLinkIO().flip 
-    val ndreset    = Bool(OUTPUT)
-    val fullreset  = Bool(OUTPUT) 
-  }
 
   //--------------------------------------------------------------
   // Register & Wire Declarations
@@ -455,46 +466,23 @@ class DebugModule ()(implicit val p:cde.Parameters)
 
   // --- Debug RAM
 
-  // Since the access size from Debug Bus and System Bus may not be consistent,
-  // use the maximum to build the RAM, and then select as needed for the smaller
-  // size.
+  val ramDataWidth    = DbBusConsts.dbRamWordBits
+  val ramDataBytes    = ramDataWidth / 8;
+  val ramAddrWidth    = log2Up(cfg.nDebugRamBytes / ramDataBytes)
 
-  val dbRamDataWidth    = DbBusConsts.dbRamWordBits
-  val sbRamDataWidth    = tlDataBits
-  val dbRamAddrWidth    = log2Up((cfg.nDebugRamBytes * 8) / dbRamDataWidth) 
-  val sbRamAddrWidth    = log2Up((cfg.nDebugRamBytes * 8) / sbRamDataWidth)
-  val sbRamAddrOffset   = log2Up(tlDataBits/8)
+  val ramMem    = Reg(init = Vec.fill(cfg.nDebugRamBytes){UInt(0, width = 8)})
 
-  val ramDataWidth = dbRamDataWidth max sbRamDataWidth
-  val ramAddrWidth  = dbRamAddrWidth min sbRamAddrWidth
-  val ramMem    = Mem(1 << ramAddrWidth , UInt(width=ramDataWidth))
-  val ramAddr   = Wire(UInt(width=ramAddrWidth))
-  val ramRdData = Wire(UInt(width=ramDataWidth))
-  val ramWrData = Wire(UInt(width=ramDataWidth))
-  val ramWrMask = Wire(UInt(width=ramDataWidth))
-  val ramWrEn   = Wire(Bool())
-
-  val dbRamAddr   = Wire(UInt(width=dbRamAddrWidth))
+  val dbRamAddr   = Wire(UInt(width=ramAddrWidth))
   val dbRamAddrValid   = Wire(Bool())
-  val dbRamRdData = Wire (UInt(width=dbRamDataWidth))
-  val dbRamWrData = Wire(UInt(width=dbRamDataWidth))
+  val dbRamRdData = Wire (UInt(width=ramDataWidth))
+  val dbRamWrData = Wire(UInt(width=ramDataWidth))
   val dbRamWrEn   = Wire(Bool())
   val dbRamRdEn   = Wire(Bool())
   val dbRamWrEnFinal   = Wire(Bool())
   val dbRamRdEnFinal   = Wire(Bool())
+  require((cfg.nDebugRamBytes % ramDataBytes) == 0)
+  val dbRamDataOffset = log2Up(ramDataBytes)
 
-  val sbRamAddr   = Wire(UInt(width=sbRamAddrWidth))
-  val sbRamAddrValid = Wire(Bool())
-  val sbRamRdData = Wire (UInt(width=sbRamDataWidth))
-  val sbRamWrData = Wire(UInt(width=sbRamDataWidth))
-  val sbRamWrEn   = Wire(Bool())
-  val sbRamRdEn   = Wire(Bool())
-  val sbRamWrEnFinal   = Wire(Bool())
-  val sbRamRdEnFinal   = Wire(Bool())
-
-
-  val sbRomRdData       = Wire(UInt(width=tlDataBits))
-  val sbRomAddrOffset   = log2Up(tlDataBits/8)
 
   // --- Debug Bus Accesses
 
@@ -513,16 +501,6 @@ class DebugModule ()(implicit val p:cde.Parameters)
   val rdCondWrFailure = Wire(Bool())
   val dbWrNeeded = Wire(Bool())
 
-  // --- System Bus Access 
-  val sbAddr   = Wire(UInt(width=sbAddrWidth))
-  val sbRdData = Wire(UInt(width=tlDataBits))
-  val sbWrData = Wire(UInt(width=tlDataBits))
-  val sbWrMask = Wire(UInt(width=tlDataBits))
-  val sbWrEn   = Wire(Bool())
-  val sbRdEn   = Wire(Bool())
-
-  val stallFromDb = Wire(Bool())
-  val stallFromSb = Wire(Bool())
   //--------------------------------------------------------------
   // Interrupt Registers
   //--------------------------------------------------------------
@@ -622,63 +600,29 @@ class DebugModule ()(implicit val p:cde.Parameters)
   HALTSUMRdData.acks        := haltnotSummary
 
   //--------------------------------------------------------------
-  // Debug RAM Access (Debug Bus & System Bus)
+  // Debug RAM Access (Debug Bus ... System Bus can override)
   //--------------------------------------------------------------
 
   dbReq := io.db.req.bits
   // Debug Bus RAM Access
   // From Specification: Debug RAM is 0x00 - 0x0F
   //                                  0x40 - 0x6F Not Implemented
-  dbRamAddr   := dbReq.addr( dbRamAddrWidth-1 , 0)
+  dbRamAddr   := dbReq.addr( ramAddrWidth-1 , 0)
   dbRamWrData := dbReq.data
-  dbRamAddrValid := Bool(true)
-  if (dbRamAddrWidth < 4){
-    dbRamAddrValid := (dbReq.addr(3, dbRamAddrWidth) === UInt(0))
+  dbRamAddrValid := (dbReq.addr(3,0) <= UInt((cfg.nDebugRamBytes/ramDataBytes)))
+
+  val dbRamRdDataFields = List.tabulate(cfg.nDebugRamBytes / ramDataBytes) { ii =>
+    val slice = ramMem.slice(ii * ramDataBytes, (ii+1)*ramDataBytes)
+    slice.reduce[UInt]{ case (x: UInt, y: UInt) => Cat(y, x)}
   }
 
-  sbRamAddr   := sbAddr(sbRamAddrWidth + sbRamAddrOffset - 1, sbRamAddrOffset)   
-  sbRamWrData := sbWrData
-  sbRamAddrValid := Bool(true)
-  // From Specification: Debug RAM is 0x400 - 0x4FF
-  if ((sbRamAddrWidth + sbRamAddrOffset) < 8){
-    sbRamAddrValid := (sbAddr(7, sbRamAddrWidth + sbRamAddrOffset) === UInt(0))
+  dbRamRdData := dbRamRdDataFields(dbRamAddr)
+
+  when (dbRamWrEnFinal) {
+    for (ii <- 0 until ramDataBytes) {
+      ramMem((dbRamAddr << UInt(dbRamDataOffset)) + UInt(ii)) := dbRamWrData((8*(ii+1)-1), (8*ii))
+    }
   }
-
-  require (dbRamAddrWidth >= ramAddrWidth)    // SB accesses less than 32 bits Not Implemented.
-  val dbRamWrMask = Wire(init=Vec.fill(1 << (dbRamAddrWidth - ramAddrWidth)){Fill(dbRamDataWidth, UInt(1, width=1))})
-
-  if (dbRamDataWidth < ramDataWidth){
-
-    val dbRamSel = dbRamAddr(dbRamAddrWidth - ramAddrWidth - 1 , 0)
-    val rdDataWords = Vec.tabulate(1 << (dbRamAddrWidth - ramAddrWidth)){ ii =>
-      ramRdData((ii+1)*dbRamDataWidth - 1 , ii*dbRamDataWidth)}
-
-    dbRamWrMask := Vec.fill(1 << (dbRamAddrWidth - ramAddrWidth)){UInt(0, width = dbRamDataWidth)}
-    dbRamWrMask(dbRamSel) := Fill(dbRamDataWidth, UInt(1, width=1))
-    dbRamRdData := rdDataWords(dbRamSel)
-  } else {
-    dbRamRdData    := ramRdData
-  }
-
-  sbRamRdData := ramRdData
-
-  ramWrMask := Mux(sbRamWrEn, sbWrMask, dbRamWrMask.asUInt)
-
-  assert (!((dbRamWrEn | dbRamRdEn) & (sbRamRdEn | sbRamWrEn)), "Stall logic should have prevented concurrent SB/DB RAM Access")
-
-  // Make copies of DB RAM data before writing.
-  val dbRamWrDataVec = Fill(1 << (dbRamAddrWidth - ramAddrWidth), dbRamWrData)
-  ramWrData := Mux(sbRamWrEn,
-    (ramWrMask & sbRamWrData   ) | (~ramWrMask & ramRdData),
-    (ramWrMask & dbRamWrDataVec) | (~ramWrMask & ramRdData))
-
-  ramAddr   := Mux(sbRamWrEn | sbRamRdEn, sbRamAddr,
-    dbRamAddr >> (dbRamAddrWidth - ramAddrWidth))
-
-  ramRdData := ramMem(ramAddr)
-  when (ramWrEn) { ramMem(ramAddr) := ramWrData }
-
-  ramWrEn := sbRamWrEnFinal | dbRamWrEnFinal
   
   //--------------------------------------------------------------
   // Debug Bus Access
@@ -813,8 +757,8 @@ class DebugModule ()(implicit val p:cde.Parameters)
 
   // -----------------------------------------
   // DB Access State Machine Decode (Combo)
-  io.db.req.ready := !stallFromSb && ((dbStateReg === s_DB_READY) ||
-    (dbStateReg === s_DB_RESP && io.db.resp.fire()))
+  io.db.req.ready := (dbStateReg === s_DB_READY) ||
+  (dbStateReg === s_DB_RESP && io.db.resp.fire())
 
   io.db.resp.valid := (dbStateReg === s_DB_RESP)
   io.db.resp.bits  := dbRespReg
@@ -844,172 +788,28 @@ class DebugModule ()(implicit val p:cde.Parameters)
   // Debug ROM
   //--------------------------------------------------------------
 
-  sbRomRdData := UInt(0)
-  if (cfg.hasDebugRom) {
-    // Inspired by ROMSlave
-    val romContents = cfg.debugRomContents.get
-    val romByteWidth = tlDataBits / 8
-    val romRows = (romContents.size + romByteWidth - 1)/romByteWidth
-    val romMem = Vec.tabulate(romRows) { ii =>
-      val slice = romContents.slice(ii*romByteWidth, (ii+1)*romByteWidth)
-      UInt(slice.foldRight(BigInt(0)) { case (x,y) => ((y << 8) + (x.toInt & 0xFF))}, width = romByteWidth*8)
-    }
-
-    val sbRomRdAddr = Wire(UInt())
-
-    if (romRows == 1) {
-      sbRomRdAddr := UInt(0)
-    } else {
-      sbRomRdAddr := sbAddr(log2Up(romRows) + sbRomAddrOffset - 1, sbRomAddrOffset)
-    }
-    sbRomRdData := romMem (sbRomRdAddr)
+  val romRegFields  =  if (cfg.hasDebugRom) {
+    cfg.debugRomContents.get.map( x => RegField.r(8, UInt(x.toInt & 0xFF)))
+  } else {
+    Seq(RegField(8))
   }
 
   //--------------------------------------------------------------
   // System Bus Access
   //--------------------------------------------------------------
 
-
-  // -----------------------------------------
-  // SB Access Write Decoder
-
-  sbRamWrEn  := Bool(false)
-  sbRamWrEnFinal := Bool(false)
-  SETHALTNOTWrEn := Bool(false)
-  CLEARDEBINTWrEn := Bool(false)
-
-  if (tlDataBits == 32) {
-    SETHALTNOTWrData := sbWrData
-    CLEARDEBINTWrData := sbWrData
-    when (sbAddr(11, 8)   === UInt(4)){ // 0x400-0x4ff is Debug RAM
-      sbRamWrEn := sbWrEn
-      sbRamRdEn := sbRdEn
-      when (sbRamAddrValid) {
-        sbRamWrEnFinal := sbWrEn
-        sbRamRdEnFinal := sbRdEn
-      }
-    }.elsewhen (sbAddr === SETHALTNOT){
-      SETHALTNOTWrEn := sbWrEn
-    }.elsewhen (sbAddr === CLEARDEBINT){
-      CLEARDEBINTWrEn := sbWrEn
-    }.otherwise {
-      //Other registers/RAM are Not Implemented.
-    }
-  } else {
-
-    // Pick out the correct word based on the address.
-    val sbWrDataWords = Vec.tabulate (tlDataBits / 32) {ii => sbWrData((ii+1)*32 - 1, ii*32)}
-    val sbWrMaskWords = Vec.tabulate (tlDataBits / 32) {ii => sbWrMask ((ii+1)*32 -1, ii*32)}
-
-    val sbWrSelTop = log2Up(tlDataBits/8) - 1
-    val sbWrSelBottom = 2
-
-    SETHALTNOTWrData  := sbWrDataWords(SETHALTNOT(sbWrSelTop, sbWrSelBottom))
-    CLEARDEBINTWrData := sbWrDataWords(CLEARDEBINT(sbWrSelTop, sbWrSelBottom))
-
-    when (sbAddr(11,8) === UInt(4)){ //0x400-0x4ff is Debug RAM
-      sbRamWrEn := sbWrEn
-      sbRamRdEn := sbRdEn
-      when (sbRamAddrValid){
-        sbRamWrEnFinal := sbWrEn
-        sbRamRdEnFinal := sbRdEn
-      }
-    }
-
-    SETHALTNOTWrEn := sbAddr(sbAddrWidth - 1, sbWrSelTop + 1) === SETHALTNOT(sbAddrWidth-1, sbWrSelTop + 1) &&
-    (sbWrMaskWords(SETHALTNOT(sbWrSelTop, sbWrSelBottom))).orR &&
-    sbWrEn
-
-    CLEARDEBINTWrEn := sbAddr(sbAddrWidth - 1, sbWrSelTop + 1) === CLEARDEBINT(sbAddrWidth-1, sbWrSelTop + 1) &&
-    (sbWrMaskWords(CLEARDEBINT(sbWrSelTop, sbWrSelBottom))).orR &&
-    sbWrEn
-
+  // Local reg mapper function : Notify when written, but give the value.
+  def wValue (n: Int, value: UInt, set: Bool) : RegField = {
+    RegField(n, value, RegWriteFn((valid, data) => {set := valid ; value := data; Bool(true)}))
   }
 
-  // -----------------------------------------
-  // SB Access Read Mux
- 
-  sbRdData := UInt(0)
-  sbRamRdEn      := Bool(false)
-  sbRamRdEnFinal := Bool(false)
-
-  when (sbAddr(11, 8) === UInt(4)) {                                 //0x400-0x4FF Debug RAM
-    sbRamRdEn := sbRdEn
-    when (sbRamAddrValid) {
-      sbRdData  := sbRamRdData
-      sbRamRdEnFinal := sbRdEn
-    }
-  }.elsewhen (sbAddr(11,8).isOneOf(UInt(8), UInt(9))){ //0x800-0x9FF Debug ROM
-    if (cfg.hasDebugRom) {
-      sbRdData := sbRomRdData
-    } else {
-      sbRdData := UInt(0)
-    }
-  }. otherwise {
-    // All readable registers are Not Implemented.
-    sbRdData := UInt(0)
-  }
-
-  // -----------------------------------------
-  // SB Access State Machine -- based on BRAM Slave
-
-  val sbAcqReg = Reg(io.tl.acquire.bits)
-  val sbAcqValidReg = Reg(init = Bool(false))
-
-  val (sbReg_get :: sbReg_getblk :: sbReg_put :: sbReg_putblk :: Nil) = Seq(
-    Acquire.getType, Acquire.getBlockType, Acquire.putType, Acquire.putBlockType
-  ).map(sbAcqReg.isBuiltInType _)
-
-  val sbMultibeat = sbReg_getblk & sbAcqValidReg;
-
-  val sbBeatInc1 = sbAcqReg.addr_beat + UInt(1)
- 
-  val sbLast = (sbAcqReg.addr_beat === UInt(tlDataBeats - 1))
-
-  sbAddr := sbAcqReg.full_addr()
-  sbRdEn := (sbAcqValidReg && (sbReg_get || sbReg_getblk))
-  sbWrEn := (sbAcqValidReg && (sbReg_put || sbReg_putblk))
-  sbWrData := sbAcqReg.data
-  sbWrMask := sbAcqReg.full_wmask()
-
-  // -----------------------------------------
-  // SB Access State Machine Update (Seq)
-
-  when (io.tl.acquire.fire()){
-    sbAcqReg       := io.tl.acquire.bits
-    sbAcqValidReg  := Bool(true)
-  } .elsewhen (io.tl.grant.fire()) {
-    when (sbMultibeat){
-      sbAcqReg.addr_beat := sbBeatInc1
-      when (sbLast) {
-        sbAcqValidReg := Bool(false)
-      }
-    } . otherwise {
-      sbAcqValidReg := Bool(false)
-    }
-  }
-  
-
-  io.tl.grant.valid := sbAcqValidReg
-  io.tl.grant.bits := Grant(
-    is_builtin_type = Bool(true),
-    g_type = sbAcqReg.getBuiltInGrantType(),
-    client_xact_id = sbAcqReg.client_xact_id,
-    manager_xact_id = UInt(0),
-    addr_beat = sbAcqReg.addr_beat,
-    data = sbRdData
+  regmap(
+    CLEARDEBINT -> Seq(wValue(sbIdWidth, CLEARDEBINTWrData, CLEARDEBINTWrEn)),
+    SETHALTNOT  -> Seq(wValue(sbIdWidth, SETHALTNOTWrData, SETHALTNOTWrEn)),
+    RAMBASE     -> ramMem.map(x => RegField(8, x)),
+    ROMBASE     -> romRegFields
   )
 
-  stallFromDb := Bool(false) // SB always wins, and DB latches its read data so it is not necessary for SB to wait
-
-  stallFromSb := sbRamRdEn || sbRamWrEn // pessimistically assume that DB/SB are going to conflict on the RAM,
-                                        // and SB doesn't latch its read data to it is necessary for DB hold
-                                        // off while SB is accessing the RAM and waiting to send its result.
-
-  val sbStall = (sbMultibeat & !sbLast) || (io.tl.grant.valid  && !io.tl.grant.ready) || stallFromDb
-
-  io.tl.acquire.ready := !sbStall
-
   //--------------------------------------------------------------
   // Misc. Outputs
   //--------------------------------------------------------------
@@ -1019,6 +819,21 @@ class DebugModule ()(implicit val p:cde.Parameters)
 
 }
 
+/** Create a concrete TL2 Slave for the DebugModule RegMapper interface.
+  *  
+  */
+
+class TLDebugModule(address: BigInt = 0)(implicit p: Parameters)
+  extends TLRegisterRouter(address, beatBytes=p(rocket.XLen)/8, executable=true)(
+  new TLRegBundle(p, _ )    with DebugModuleBundle)(
+  new TLRegModule(p, _, _)  with DebugModule)
+
+
+/** Synchronizers for DebugBus
+  *  
+  */
+
+
 object AsyncDebugBusCrossing {
   // takes from_source from the 'from' clock domain to the 'to' clock domain
   def apply(from_clock: Clock, from_reset: Bool, from_source: DebugBusIO, to_clock: Clock, to_reset: Bool, depth: Int = 1, sync: Int = 3) = {
@@ -1029,6 +844,7 @@ object AsyncDebugBusCrossing {
   }
 }
 
+
 object AsyncDebugBusFrom { // OutsideClockDomain
   // takes from_source from the 'from' clock domain and puts it into your clock domain
   def apply(from_clock: Clock, from_reset: Bool, from_source: DebugBusIO, depth: Int = 1, sync: Int = 3): DebugBusIO = {
diff --git a/src/main/scala/uncore/devices/Plic.scala b/src/main/scala/uncore/devices/Plic.scala
index 8a180d99..830cd88a 100644
--- a/src/main/scala/uncore/devices/Plic.scala
+++ b/src/main/scala/uncore/devices/Plic.scala
@@ -6,8 +6,11 @@ import Chisel._
 import Chisel.ImplicitConversions._
 
 import junctions._
-import uncore.tilelink._
+import diplomacy._
+import regmapper._
+import uncore.tilelink2._
 import cde.Parameters
+import scala.math.min
 
 class GatewayPLICIO extends Bundle {
   val valid = Bool(OUTPUT)
@@ -27,161 +30,183 @@ class LevelGateway extends Module {
   io.plic.valid := io.interrupt && !inFlight
 }
 
-case class PLICConfig(nHartsIn: Int, supervisor: Boolean, nDevices: Int, nPriorities: Int) {
-  def contextsPerHart = if (supervisor) 2 else 1
-  def nHarts = contextsPerHart * nHartsIn
-  def context(i: Int, mode: Char) = mode match {
-    case 'M' => i * contextsPerHart
-    case 'S' => require(supervisor); i * contextsPerHart + 1
-  }
-  def claimAddr(i: Int, mode: Char) = hartBase + hartOffset(context(i, mode)) + claimOffset
-  def threshAddr(i: Int, mode: Char) = hartBase + hartOffset(context(i, mode))
-  def enableAddr(i: Int, mode: Char) = enableBase + enableOffset(context(i, mode))
-  def size = hartBase + hartOffset(maxHarts)
-
+object PLICConsts
+{
   def maxDevices = 1023
   def maxHarts = 15872
+  def priorityBase = 0x0
   def pendingBase = 0x1000
   def enableBase = 0x2000
   def hartBase = 0x200000
-  require(hartBase >= enableBase + enableOffset(maxHarts))
 
-  def enableOffset(i: Int) = i * ((maxDevices+7)/8)
-  def hartOffset(i: Int) = i * 0x1000
   def claimOffset = 4
   def priorityBytes = 4
 
-  require(nDevices <= maxDevices)
-  require(nHarts > 0 && nHarts <= maxHarts)
-  require(nPriorities >= 0 && nPriorities <= nDevices)
+  def enableOffset(i: Int) = i * ((maxDevices+7)/8)
+  def hartOffset(i: Int) = i * 0x1000
+  def enableBase(i: Int):Int = enableOffset(i) + enableBase
+  def hartBase(i: Int):Int = hartOffset(i) + hartBase
+
+  def size = hartBase(maxHarts)
+  require(hartBase >= enableBase(maxHarts))
 }
 
 /** Platform-Level Interrupt Controller */
-class PLIC(val cfg: PLICConfig)(implicit val p: Parameters) extends Module
-    with HasTileLinkParameters
-    with HasAddrMapParameters {
-  val io = new Bundle {
-    val devices = Vec(cfg.nDevices, new GatewayPLICIO).flip
-    val harts = Vec(cfg.nHarts, Bool()).asOutput
-    val tl = new ClientUncachedTileLinkIO().flip
-  }
+class TLPLIC(supervisor: Boolean, maxPriorities: Int, address: BigInt = 0xC000000)(implicit val p: Parameters) extends LazyModule
+{
+  val contextsPerHart = if (supervisor) 2 else 1
+  require (maxPriorities >= 0)
 
-  val priority =
-    if (cfg.nPriorities > 0) Reg(Vec(cfg.nDevices+1, UInt(width=log2Up(cfg.nPriorities+1))))
-    else Wire(init=Vec.fill(cfg.nDevices+1)(UInt(1)))
-  val threshold =
-    if (cfg.nPriorities > 0) Reg(Vec(cfg.nHarts, UInt(width = log2Up(cfg.nPriorities+1))))
-    else Wire(init=Vec.fill(cfg.nHarts)(UInt(0)))
-  val pending = Reg(init=Vec.fill(cfg.nDevices+1){Bool(false)})
-  val enables = Reg(Vec(cfg.nHarts, Vec(cfg.nDevices+1, Bool())))
+  val node = TLRegisterNode(
+    address   = AddressSet(address, PLICConsts.size-1),
+    beatBytes = p(rocket.XLen)/8,
+    undefZero = false)
 
-  for ((p, g) <- pending.tail zip io.devices) {
-    g.ready := !p
-    g.complete := false
-    when (g.valid) { p := true }
-  }
+  val intnode = IntAdapterNode(
+    numSourcePorts = 0 to 1024,
+    numSinkPorts   = 0 to 1024,
+    sourceFn       = { _ => IntSourcePortParameters(Seq(IntSourceParameters(contextsPerHart))) },
+    sinkFn         = { _ => IntSinkPortParameters(Seq(IntSinkParameters())) })
 
-  def findMax(x: Seq[UInt]): (UInt, UInt) = {
-    if (x.length > 1) {
-      val half = 1 << (log2Ceil(x.length) - 1)
-      val lMax = findMax(x take half)
-      val rMax = findMax(x drop half)
-      val useLeft = lMax._1 >= rMax._1
-      (Mux(useLeft, lMax._1, rMax._1), Mux(useLeft, lMax._2, UInt(half) + rMax._2))
-    } else (x.head, UInt(0))
-  }
-
-  val maxDevs = Wire(Vec(cfg.nHarts, UInt(width = log2Up(pending.size))))
-  for (hart <- 0 until cfg.nHarts) {
-    val effectivePriority =
-      for (((p, en), pri) <- (pending zip enables(hart) zip priority).tail)
-        yield Cat(p && en, pri)
-    val (maxPri, maxDev) = findMax((UInt(1) << priority(0).getWidth) +: effectivePriority)
-
-    maxDevs(hart) := Reg(next = maxDev)
-    io.harts(hart) := Reg(next = maxPri) > Cat(UInt(1), threshold(hart))
-  }
-
-  val acq = Queue(io.tl.acquire, 1)
-  val read = acq.fire() && acq.bits.isBuiltInType(Acquire.getType)
-  val write = acq.fire() && acq.bits.isBuiltInType(Acquire.putType)
-  assert(!acq.fire() || read || write, "unsupported PLIC operation")
-  val addr = acq.bits.full_addr()(log2Up(cfg.size)-1,0)
-
-  val claimant =
-    if (cfg.nHarts == 1) UInt(0)
-    else (addr - cfg.hartBase)(log2Up(cfg.hartOffset(cfg.nHarts))-1,log2Up(cfg.hartOffset(1)))
-  val hart = Wire(init = claimant)
-  val myMaxDev = maxDevs(claimant)
-  val myEnables = enables(hart)
-  val rdata = Wire(init = UInt(0, tlDataBits))
-  val masked_wdata = (acq.bits.data & acq.bits.full_wmask()) | (rdata & ~acq.bits.full_wmask())
-
-  if (cfg.nDevices > 0) when (addr >= cfg.hartBase) {
-    val word =
-      if (tlDataBytes > cfg.claimOffset) UInt(0)
-      else addr(log2Up(cfg.claimOffset),log2Up(tlDataBytes))
-    rdata := Cat(myMaxDev, UInt(0, 8*cfg.priorityBytes-threshold(0).getWidth), threshold(claimant)) >> (word * tlDataBits)
-
-    when (read && addr(log2Ceil(cfg.claimOffset))) {
-      pending(myMaxDev) := false
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val tl_in = node.bundleIn
+      val devices = intnode.bundleIn
+      val harts = intnode.bundleOut
     }
-    when (write) {
-      when (if (tlDataBytes > cfg.claimOffset) acq.bits.wmask()(cfg.claimOffset) else addr(log2Ceil(cfg.claimOffset))) {
-        val dev = (acq.bits.data >> ((8 * cfg.claimOffset) % tlDataBits))(log2Up(pending.size)-1,0)
-        when (myEnables(dev)) { io.devices(dev-1).complete := true }
-      }.otherwise {
-        if (cfg.nPriorities > 0) threshold(claimant) := acq.bits.data
-      }
+
+    // Assign all the devices unique ranges
+    val sources = intnode.edgesIn.map(_.source)
+    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
+    // Compact the interrupt vector the same way
+    val interrupts = (intnode.edgesIn zip io.devices).map { case (e, i) => i.take(e.source.num) }.flatten
+    // This flattens the harts into an MSMSMSMSMS... or MMMMM.... sequence
+    val harts = io.harts.flatten
+
+    println("\nInterrupt map:")
+    flatSources.foreach { s =>
+      // +1 because 0 is reserved, +1-1 because the range is half-open
+      println(s"  [${s.range.start+1}, ${s.range.end}] => ${s.name}")
     }
-  }.elsewhen (addr >= cfg.enableBase) {
-    val enableHart =
-      if (cfg.nHarts > 1) (addr - cfg.enableBase)(log2Up(cfg.enableOffset(cfg.nHarts))-1,log2Up(cfg.enableOffset(1)))
-      else UInt(0)
-    hart := enableHart
-    val word =
-      if (tlDataBits >= myEnables.size) UInt(0)
-      else addr(log2Ceil((myEnables.size-1)/tlDataBits+1) + tlByteAddrBits - 1, tlByteAddrBits)
-    for (i <- 0 until myEnables.size by tlDataBits) {
-      when (word === i/tlDataBits) {
-        rdata := Cat(myEnables.slice(i, i + tlDataBits).reverse)
-        for (j <- 0 until (tlDataBits min (myEnables.size - i))) {
-          when (write) { enables(enableHart)(i+j) := masked_wdata(j) }
-        }
-      }
+
+    val nDevices = interrupts.size
+    val nPriorities = min(maxPriorities, nDevices)
+    val nHarts = harts.size
+
+    require(nDevices <= PLICConsts.maxDevices)
+    require(nHarts > 0 && nHarts <= PLICConsts.maxHarts)
+
+    def context(i: Int, mode: Char) = mode match {
+      case 'M' => i * contextsPerHart
+      case 'S' => require(supervisor); i * contextsPerHart + 1
     }
-  }.elsewhen (addr >= cfg.pendingBase) {
-    val word =
-      if (tlDataBytes >= pending.size) UInt(0)
-      else addr(log2Up(pending.size)-1,log2Up(tlDataBytes))
-    rdata := pending.asUInt >> (word * tlDataBits)
-  }.otherwise {
-    val regsPerBeat = tlDataBytes >> log2Up(cfg.priorityBytes)
-    val word =
-      if (regsPerBeat >= priority.size) UInt(0)
-      else addr(log2Up(priority.size*cfg.priorityBytes)-1,log2Up(tlDataBytes))
-    for (i <- 0 until priority.size by regsPerBeat) {
-      when (word === i/regsPerBeat) {
-        rdata := Cat(priority.slice(i, i + regsPerBeat).map(p => Cat(UInt(0, 8*cfg.priorityBytes-p.getWidth), p)).reverse)
-        for (j <- 0 until (regsPerBeat min (priority.size - i))) {
-          if (cfg.nPriorities > 0) when (write) { priority(i+j) := masked_wdata >> (j * 8 * cfg.priorityBytes) }
-        }
-      }
+    def claimAddr(i: Int, mode: Char)  = address + PLICConsts.hartBase(context(i, mode)) + PLICConsts.claimOffset
+    def threshAddr(i: Int, mode: Char) = address + PLICConsts.hartBase(context(i, mode))
+    def enableAddr(i: Int, mode: Char) = address + PLICConsts.enableBase(context(i, mode))
+
+    // Create the global PLIC config string
+    val globalConfigString = Seq(
+      s"plic {\n",
+      s"  priority 0x${address.toString(16)};\n",
+      s"  pending 0x${(address + PLICConsts.pendingBase).toString(16)};\n",
+      s"  ndevs ${nDevices};\n",
+      s"};\n").mkString
+
+    // Create the per-Hart config string
+    val hartConfigStrings = io.harts.zipWithIndex.map { case (_, i) => (Seq(
+      s"      plic {\n",
+      s"        m {\n",
+      s"         ie 0x${enableAddr(i, 'M').toString(16)};\n",
+      s"         thresh 0x${threshAddr(i, 'M').toString(16)};\n",
+      s"         claim 0x${claimAddr(i, 'M').toString(16)};\n",
+      s"        };\n") ++ (if (!supervisor) Seq() else Seq(
+      s"        s {\n",
+      s"         ie 0x${enableAddr(i, 'S').toString(16)};\n",
+      s"         thresh 0x${threshAddr(i, 'S').toString(16)};\n",
+      s"         claim 0x${claimAddr(i, 'S').toString(16)};\n",
+      s"        };\n")) ++ Seq(
+      s"      };\n")).mkString
     }
+
+    // For now, use LevelGateways for all TL2 interrupts
+    val gateways = Vec(interrupts.map { case i =>
+      val gateway = Module(new LevelGateway)
+      gateway.io.interrupt := i
+      gateway.io.plic
+    })
+
+    val priority =
+      if (nPriorities > 0) Reg(Vec(nDevices+1, UInt(width=log2Up(nPriorities+1))))
+      else Wire(init=Vec.fill(nDevices+1)(UInt(1)))
+    val threshold =
+      if (nPriorities > 0) Reg(Vec(nHarts, UInt(width = log2Up(nPriorities+1))))
+      else Wire(init=Vec.fill(nHarts)(UInt(0)))
+    val pending = Reg(init=Vec.fill(nDevices+1){Bool(false)})
+    val enables = Reg(Vec(nHarts, Vec(nDevices+1, Bool())))
+
+    for ((p, g) <- pending.tail zip gateways) {
+      g.ready := !p
+      g.complete := false
+      when (g.valid) { p := true }
+    }
+
+    def findMax(x: Seq[UInt]): (UInt, UInt) = {
+      if (x.length > 1) {
+        val half = 1 << (log2Ceil(x.length) - 1)
+        val lMax = findMax(x take half)
+        val rMax = findMax(x drop half)
+        val useLeft = lMax._1 >= rMax._1
+        (Mux(useLeft, lMax._1, rMax._1), Mux(useLeft, lMax._2, UInt(half) | rMax._2))
+      } else (x.head, UInt(0))
+    }
+
+    val maxDevs = Reg(Vec(nHarts, UInt(width = log2Up(pending.size))))
+    for (hart <- 0 until nHarts) {
+      val effectivePriority =
+        for (((p, en), pri) <- (pending zip enables(hart) zip priority).tail)
+          yield Cat(p && en, pri)
+      val (maxPri, maxDev) = findMax((UInt(1) << priority(0).getWidth) +: effectivePriority)
+
+      maxDevs(hart) := maxDev
+      harts(hart) := Reg(next = maxPri) > Cat(UInt(1), threshold(hart))
+    }
+
+    def priorityRegField(x: UInt) = if (nPriorities > 0) RegField(32, x) else RegField.r(32, x)
+    val priorityRegFields = Seq(PLICConsts.priorityBase -> priority.map(p => priorityRegField(p)))
+    val pendingRegFields = Seq(PLICConsts.pendingBase  -> pending .map(b => RegField.r(1, b)))
+
+    val enableRegFields = enables.zipWithIndex.map { case (e, i) =>
+      PLICConsts.enableBase(i) -> e.map(b => RegField(1, b))
+    }
+
+    val hartRegFields = Seq.tabulate(nHarts) { i =>
+      PLICConsts.hartBase(i) -> Seq(
+        priorityRegField(threshold(i)),
+        RegField(32,
+          RegReadFn { valid =>
+            when (valid) {
+              pending(maxDevs(i)) := Bool(false)
+              maxDevs(i) := UInt(0) // flush pipeline
+            }
+            (Bool(true), maxDevs(i))
+          },
+          RegWriteFn { (valid, data) =>
+            when (valid && enables(i)(data)) {
+              gateways(data - UInt(1)).complete := Bool(true)
+            }
+            Bool(true)
+          }
+        )
+      )
+    }
+
+    node.regmap((priorityRegFields ++ pendingRegFields ++ enableRegFields ++ hartRegFields):_*)
+
+    priority(0) := 0
+    pending(0) := false
+    for (e <- enables)
+      e(0) := false
   }
-
-  priority(0) := 0
-  pending(0) := false
-  for (e <- enables)
-    e(0) := false
-
-  io.tl.grant.valid := acq.valid
-  acq.ready := io.tl.grant.ready
-  io.tl.grant.bits := Grant(
-    is_builtin_type = Bool(true),
-    g_type = acq.bits.getBuiltInGrantType(),
-    client_xact_id = acq.bits.client_xact_id,
-    manager_xact_id = UInt(0),
-    addr_beat = UInt(0),
-    data = rdata)
 }
diff --git a/src/main/scala/uncore/devices/Prci.scala b/src/main/scala/uncore/devices/Prci.scala
index ab3ce91a..604c89ab 100644
--- a/src/main/scala/uncore/devices/Prci.scala
+++ b/src/main/scala/uncore/devices/Prci.scala
@@ -6,6 +6,7 @@ import Chisel._
 import junctions._
 import junctions.NastiConstants._
 import regmapper._
+import diplomacy._
 import uncore.tilelink2._
 import uncore.util._
 import util._
@@ -20,22 +21,19 @@ class CoreplexLocalInterrupts extends Bundle {
   val msip = Bool()
 }
 
-case class CoreplexLocalInterrupterConfig(beatBytes: Int, address: BigInt = 0x02000000) {
+object ClintConsts
+{
   def msipOffset(hart: Int) = hart * msipBytes
-  def msipAddress(hart: Int) = address + msipOffset(hart)
   def timecmpOffset(hart: Int) = 0x4000 + hart * timecmpBytes
-  def timecmpAddress(hart: Int) = address + timecmpOffset(hart)
   def timeOffset = 0xbff8
-  def timeAddress = address + timeOffset
   def msipBytes = 4
   def timecmpBytes = 8
   def size = 0x10000
 }
 
 trait MixCoreplexLocalInterrupterParameters {
-  val params: (CoreplexLocalInterrupterConfig, Parameters)
-  val c = params._1
-  implicit val p = params._2
+  val params: Parameters
+  implicit val p = params
 }
 
 trait CoreplexLocalInterrupterBundle extends Bundle with MixCoreplexLocalInterrupterParameters {
@@ -45,11 +43,10 @@ trait CoreplexLocalInterrupterBundle extends Bundle with MixCoreplexLocalInterru
 
 trait CoreplexLocalInterrupterModule extends Module with HasRegMap with MixCoreplexLocalInterrupterParameters {
   val io: CoreplexLocalInterrupterBundle
+  val address: AddressSet
 
   val timeWidth = 64
   val regWidth = 32
-  // demand atomic accesses for RV64
-  require(c.beatBytes >= (p(rocket.XLen) min timeWidth)/8)
 
   val time = Seq.fill(timeWidth/regWidth)(Reg(init=UInt(0, width = regWidth)))
   when (io.rtcTick) {
@@ -66,6 +63,15 @@ trait CoreplexLocalInterrupterModule extends Module with HasRegMap with MixCorep
     tile.mtip := time.asUInt >= timecmp(i).asUInt
   }
 
+  val globalConfigString = Seq(
+    s"rtc {\n",
+    s"  addr 0x${(address.base + ClintConsts.timeOffset).toString(16)};\n",
+    s"};\n").mkString
+  val hartConfigStrings = (0 until p(NTiles)).map { i => Seq(
+    s"      timecmp 0x${(address.base + ClintConsts.timecmpOffset(i)).toString(16)};\n",
+    s"      ipi 0x${(address.base + ClintConsts.msipOffset(i)).toString(16)};\n").mkString
+  }
+
   /* 0000 msip hart 0
    * 0004 msip hart 1
    * 4000 mtimecmp hart 0 lo
@@ -77,16 +83,16 @@ trait CoreplexLocalInterrupterModule extends Module with HasRegMap with MixCorep
    */
 
   regmap(
-    0                  -> makeRegFields(ipi),
-    c.timecmpOffset(0) -> makeRegFields(timecmp.flatten),
-    c.timeOffset       -> makeRegFields(time))
+    0                            -> makeRegFields(ipi),
+    ClintConsts.timecmpOffset(0) -> makeRegFields(timecmp.flatten),
+    ClintConsts.timeOffset       -> makeRegFields(time))
 
   def makeRegFields(s: Seq[UInt]) = s.map(r => RegField(regWidth, r))
 }
 
 /** Power, Reset, Clock, Interrupt */
 // Magic TL2 Incantation to create a TL2 Slave
-class CoreplexLocalInterrupter(c: CoreplexLocalInterrupterConfig)(implicit val p: Parameters)
-  extends TLRegisterRouter(c.address, 0, c.size, 0, c.beatBytes, false)(
-  new TLRegBundle((c, p), _)    with CoreplexLocalInterrupterBundle)(
-  new TLRegModule((c, p), _, _) with CoreplexLocalInterrupterModule)
+class CoreplexLocalInterrupter(address: BigInt = 0x02000000)(implicit val p: Parameters)
+  extends TLRegisterRouter(address, size = ClintConsts.size, beatBytes = p(rocket.XLen)/8, undefZero = false)(
+  new TLRegBundle(p, _)    with CoreplexLocalInterrupterBundle)(
+  new TLRegModule(p, _, _) with CoreplexLocalInterrupterModule)
diff --git a/src/main/scala/uncore/tilelink2/AtomicAutomata.scala b/src/main/scala/uncore/tilelink2/AtomicAutomata.scala
index 70bacae2..1a471334 100644
--- a/src/main/scala/uncore/tilelink2/AtomicAutomata.scala
+++ b/src/main/scala/uncore/tilelink2/AtomicAutomata.scala
@@ -94,7 +94,7 @@ class TLAtomicAutomata(logical: Boolean = true, arithmetic: Boolean = true, conc
       val cam_free   = cam_s.map(_.state === FREE)
       val cam_amo    = cam_s.map(_.state === AMO)
       val cam_abusy  = cam_s.map(e => e.state === GET || e.state === AMO) // A is blocked
-      val cam_dmatch = cam_s.map(e => e.state === GET || e.state === ACK) // D should inspect these entries
+      val cam_dmatch = cam_s.map(e => e.state =/= FREE) // D should inspect these entries
 
       // Can the manager already handle this message?
       val a_size = edgeIn.size(in.a.bits)
@@ -211,7 +211,7 @@ class TLAtomicAutomata(logical: Boolean = true, arithmetic: Boolean = true, conc
       val d_cam_sel_match = (d_cam_sel_raw zip cam_dmatch) map { case (a,b) => a&&b }
       val d_cam_data = Mux1H(d_cam_sel_match, cam_d.map(_.data))
       val d_cam_sel_bypass = if (edgeOut.manager.minLatency > 0) Bool(false) else
-                             out.d.bits.source === in.a.bits.source && in.a.valid && out.d.valid && !a_isSupported
+                             out.d.bits.source === in.a.bits.source && in.a.valid && !a_isSupported
       val d_cam_sel = (a_cam_sel_free zip d_cam_sel_match) map { case (a,d) => Mux(d_cam_sel_bypass, a, d) }
       val d_cam_sel_any = d_cam_sel_bypass || d_cam_sel_match.reduce(_ || _)
       val d_ackd = out.d.bits.opcode === TLMessages.AccessAckData
diff --git a/src/main/scala/uncore/tilelink2/HintHandler.scala b/src/main/scala/uncore/tilelink2/HintHandler.scala
index 16d36a8e..3ba187e0 100644
--- a/src/main/scala/uncore/tilelink2/HintHandler.scala
+++ b/src/main/scala/uncore/tilelink2/HintHandler.scala
@@ -2,6 +2,7 @@
 
 package uncore.tilelink2
 
+import scala.math.min
 import Chisel._
 import chisel3.internal.sourceinfo.SourceInfo
 import diplomacy._
@@ -9,10 +10,9 @@ import diplomacy._
 // Acks Hints for managers that don't support them or Acks all Hints if !passthrough
 class TLHintHandler(supportManagers: Boolean = true, supportClients: Boolean = false, passthrough: Boolean = true) extends LazyModule
 {
-  // HintAcks can come back combinationally => minLatency=0
   val node = TLAdapterNode(
-    clientFn  = { case Seq(c) => if (!supportClients)  c else c.copy(minLatency = 0, clients  = c.clients .map(_.copy(supportsHint = TransferSizes(1, c.maxTransfer)))) },
-    managerFn = { case Seq(m) => if (!supportManagers) m else m.copy(minLatency = 0, managers = m.managers.map(_.copy(supportsHint = TransferSizes(1, m.maxTransfer)))) })
+    clientFn  = { case Seq(c) => if (!supportClients)  c else c.copy(minLatency = min(1, c.minLatency), clients  = c.clients .map(_.copy(supportsHint = TransferSizes(1, c.maxTransfer)))) },
+    managerFn = { case Seq(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 {
@@ -46,7 +46,7 @@ class TLHintHandler(supportManagers: Boolean = true, supportClients: Boolean = f
       hint.bits := edgeIn.HintAck(in.a.bits, edgeOut.manager.findIdStartFast(address))
       out.a.bits := in.a.bits
 
-      TLArbiter(TLArbiter.lowestIndexFirst)(in.d, (edgeOut.numBeats1(out.d.bits), out.d), (UInt(0), hint))
+      TLArbiter(TLArbiter.lowestIndexFirst)(in.d, (edgeOut.numBeats1(out.d.bits), out.d), (UInt(0), Queue(hint, 1)))
     } else {
       out.a.valid := in.a.valid
       in.a.ready := out.a.ready
@@ -69,7 +69,7 @@ class TLHintHandler(supportManagers: Boolean = true, supportClients: Boolean = f
       hint.bits := edgeOut.HintAck(out.b.bits)
       in.b.bits := out.b.bits
 
-      TLArbiter(TLArbiter.lowestIndexFirst)(out.c, (edgeIn.numBeats1(in.c.bits), in.c), (UInt(0), hint))
+      TLArbiter(TLArbiter.lowestIndexFirst)(out.c, (edgeIn.numBeats1(in.c.bits), in.c), (UInt(0), Queue(hint, 1)))
     } else if (bce) {
       in.b.valid := out.b.valid
       out.b.ready := in.b.ready
diff --git a/src/main/scala/uncore/tilelink2/IntNodes.scala b/src/main/scala/uncore/tilelink2/IntNodes.scala
index 639d03ed..d98e93e7 100644
--- a/src/main/scala/uncore/tilelink2/IntNodes.scala
+++ b/src/main/scala/uncore/tilelink2/IntNodes.scala
@@ -58,10 +58,13 @@ object IntImp extends NodeImp[IntSourcePortParameters, IntSinkPortParameters, In
   }
   def bundleI(ei: Seq[IntEdge]): Vec[Vec[Bool]] = {
     require (!ei.isEmpty)
-    Vec(ei.size, Vec(ei.map(_.source.num).max, Bool())).flip
+    Vec(ei.size, Vec(ei.map(_.source.num).max, Bool()))
   }
 
   def colour = "#0000ff" // blue
+  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 sourceInfo: SourceInfo): (Option[LazyModule], () => Unit) = {
     (None, () => {
       // Cannot use bulk connect, because the widths could differ
@@ -76,9 +79,6 @@ object IntImp extends NodeImp[IntSourcePortParameters, IntSinkPortParameters, In
 }
 
 case class IntIdentityNode() extends IdentityNode(IntImp)
-case class IntOutputNode() extends OutputNode(IntImp)
-case class IntInputNode() extends InputNode(IntImp)
-
 case class IntSourceNode(num: Int) extends SourceNode(IntImp)(
   IntSourcePortParameters(Seq(IntSourceParameters(num))), (if (num == 0) 0 else 1) to 1)
 case class IntSinkNode() extends SinkNode(IntImp)(
@@ -91,11 +91,20 @@ case class IntAdapterNode(
   numSinkPorts:   Range.Inclusive = 1 to 1)
   extends InteriorNode(IntImp)(sourceFn, sinkFn, numSourcePorts, numSinkPorts)
 
+case class IntOutputNode() extends OutputNode(IntImp)
+case class IntInputNode() extends InputNode(IntImp)
+
+case class IntBlindOutputNode() extends BlindOutputNode(IntImp)(IntSinkPortParameters(Seq(IntSinkParameters())))
+case class IntBlindInputNode(num: Int) extends BlindInputNode(IntImp)(IntSourcePortParameters(Seq(IntSourceParameters(num))))
+
+case class IntInternalOutputNode() extends InternalOutputNode(IntImp)(IntSinkPortParameters(Seq(IntSinkParameters())))
+case class IntInternalInputNode(num: Int) extends InternalInputNode(IntImp)(IntSourcePortParameters(Seq(IntSourceParameters(num))))
+
 class IntXbar extends LazyModule
 {
   val intnode = IntAdapterNode(
     numSourcePorts = 1 to 1, // does it make sense to have more than one interrupt sink?
-    numSinkPorts   = 1 to 128,
+    numSinkPorts   = 0 to 128,
     sinkFn         = { _ => IntSinkPortParameters(Seq(IntSinkParameters())) },
     sourceFn       = { seq =>
       IntSourcePortParameters((seq zip seq.map(_.num).scanLeft(0)(_+_).init).map {
@@ -113,3 +122,17 @@ class IntXbar extends LazyModule
     io.out.foreach { _ := cat }
   }
 }
+
+class IntXing extends LazyModule
+{
+  val intnode = IntIdentityNode()
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val in = intnode.bundleIn
+      val out = intnode.bundleOut
+    }
+
+    io.out := RegNext(RegNext(RegNext(io.in)))
+  }
+}
diff --git a/src/main/scala/uncore/tilelink2/Legacy.scala b/src/main/scala/uncore/tilelink2/Legacy.scala
index 252d77ea..3fa75774 100644
--- a/src/main/scala/uncore/tilelink2/Legacy.scala
+++ b/src/main/scala/uncore/tilelink2/Legacy.scala
@@ -38,11 +38,9 @@ class TLLegacy(implicit val p: Parameters) extends LazyModule with HasTileLinkPa
         require (m.supportsPutPartial.contains(TransferSizes(1, tlDataBytes)))
         require (m.supportsPutPartial.contains(TransferSizes(tlDataBeats * tlDataBytes)))
       }
-      // Any atomic support => must support 32-bit up to beat size of all types
-      if (m.supportsArithmetic || m.supportsLogical) {
-        require (m.supportsArithmetic.contains(TransferSizes(4, tlDataBytes)))
-        require (m.supportsLogical   .contains(TransferSizes(4, tlDataBytes)))
-      }
+      // Any atomic support => must support 32-bit size
+      if (m.supportsArithmetic) { require (m.supportsArithmetic.contains(TransferSizes(4))) }
+      if (m.supportsLogical)    { require (m.supportsLogical   .contains(TransferSizes(4))) }
       // We straight-up require hints
       require (edge.manager.allSupportHint)
     }
diff --git a/src/main/scala/uncore/tilelink2/Nodes.scala b/src/main/scala/uncore/tilelink2/Nodes.scala
index 24668401..962517ba 100644
--- a/src/main/scala/uncore/tilelink2/Nodes.scala
+++ b/src/main/scala/uncore/tilelink2/Nodes.scala
@@ -17,7 +17,7 @@ object TLImp extends NodeImp[TLClientPortParameters, TLManagerPortParameters, TL
   }
   def bundleI(ei: Seq[TLEdgeIn]): Vec[TLBundle] = {
     require (!ei.isEmpty)
-    Vec(ei.size, TLBundle(ei.map(_.bundle).reduce(_.union(_)))).flip
+    Vec(ei.size, TLBundle(ei.map(_.bundle).reduce(_.union(_))))
   }
 
   var emitMonitors = true
@@ -25,6 +25,9 @@ object TLImp extends NodeImp[TLClientPortParameters, TLManagerPortParameters, TL
   var combinationalCheck = false
 
   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 connect(bo: => TLBundle, bi: => TLBundle, ei: => TLEdgeIn)(implicit sourceInfo: SourceInfo): (Option[LazyModule], () => Unit) = {
     val monitor = if (emitMonitors) {
       Some(LazyModule(new TLMonitor(() => new TLBundleSnoop(bo.params), () => ei, sourceInfo)))
@@ -89,10 +92,8 @@ object TLImp extends NodeImp[TLClientPortParameters, TLManagerPortParameters, TL
   }
 }
 
+// Nodes implemented inside modules
 case class TLIdentityNode() extends IdentityNode(TLImp)
-case class TLOutputNode() extends OutputNode(TLImp)
-case class TLInputNode() extends InputNode(TLImp)
-
 case class TLClientNode(portParams: TLClientPortParameters, numPorts: Range.Inclusive = 1 to 1)
   extends SourceNode(TLImp)(portParams, numPorts)
 case class TLManagerNode(portParams: TLManagerPortParameters, numPorts: Range.Inclusive = 1 to 1)
@@ -117,6 +118,14 @@ case class TLAdapterNode(
   numManagerPorts: Range.Inclusive = 1 to 1)
   extends InteriorNode(TLImp)(clientFn, managerFn, numClientPorts, numManagerPorts)
 
+// Nodes passed from an inner module
+case class TLOutputNode() extends OutputNode(TLImp)
+case class TLInputNode() extends InputNode(TLImp)
+
+// Nodes used for external ports
+case class TLBlindOutputNode(portParams: TLManagerPortParameters) extends BlindOutputNode(TLImp)(portParams)
+case class TLBlindInputNode(portParams: TLClientPortParameters) extends BlindInputNode(TLImp)(portParams)
+
 /** Synthesizeable unit tests */
 import unittest._
 
@@ -149,10 +158,13 @@ object TLAsyncImp extends NodeImp[TLAsyncClientPortParameters, TLAsyncManagerPor
   }
   def bundleI(ei: Seq[TLAsyncEdgeParameters]): Vec[TLAsyncBundle] = {
     require (ei.size == 1)
-    Vec(ei.size, new TLAsyncBundle(ei(0).bundle)).flip
+    Vec(ei.size, new TLAsyncBundle(ei(0).bundle))
   }
 
   def colour = "#ff0000" // red
+  override def labelI(ei: TLAsyncEdgeParameters) = ei.manager.depth.toString
+  override def labelO(eo: TLAsyncEdgeParameters) = eo.manager.depth.toString
+
   def connect(bo: => TLAsyncBundle, bi: => TLAsyncBundle, ei: => TLAsyncEdgeParameters)(implicit sourceInfo: SourceInfo): (Option[LazyModule], () => Unit) = {
     (None, () => { bi <> bo })
   }
diff --git a/src/main/scala/uncore/tilelink2/RegisterRouter.scala b/src/main/scala/uncore/tilelink2/RegisterRouter.scala
index f8ab2fc3..3004ea1f 100644
--- a/src/main/scala/uncore/tilelink2/RegisterRouter.scala
+++ b/src/main/scala/uncore/tilelink2/RegisterRouter.scala
@@ -7,10 +7,11 @@ import diplomacy._
 import regmapper._
 import scala.math.{min,max}
 
-class TLRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true)
+class TLRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)
   extends TLManagerNode(TLManagerPortParameters(
     Seq(TLManagerParameters(
       address            = Seq(address),
+      executable         = executable,
       supportsGet        = TransferSizes(1, beatBytes),
       supportsPutPartial = TransferSizes(1, beatBytes),
       supportsPutFull    = TransferSizes(1, beatBytes),
@@ -70,17 +71,17 @@ class TLRegisterNode(address: AddressSet, concurrency: Int = 0, beatBytes: Int =
 
 object TLRegisterNode
 {
-  def apply(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true) =
-    new TLRegisterNode(address, concurrency, beatBytes, undefZero)
+  def apply(address: AddressSet, concurrency: Int = 0, beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false) =
+    new TLRegisterNode(address, 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
 
-abstract class TLRegisterRouterBase(address: AddressSet, interrupts: Int, concurrency: Int, beatBytes: Int, undefZero: Boolean) extends LazyModule
+abstract class TLRegisterRouterBase(val address: AddressSet, interrupts: Int, concurrency: Int, beatBytes: Int, undefZero: Boolean, executable: Boolean) extends LazyModule
 {
-  val node = TLRegisterNode(address, concurrency, beatBytes, undefZero)
+  val node = TLRegisterNode(address, concurrency, beatBytes, undefZero, executable)
   val intnode = IntSourceNode(interrupts)
 }
 
@@ -99,14 +100,15 @@ class TLRegModule[P, B <: TLRegBundleBase](val params: P, bundleBuilder: => B, r
 {
   val io = bundleBuilder
   val interrupts = if (io.interrupts.isEmpty) Vec(0, Bool()) else io.interrupts(0)
+  val address = router.address
   def regmap(mapping: RegField.Map*) = router.node.regmap(mapping:_*)
 }
 
 class TLRegisterRouter[B <: TLRegBundleBase, M <: LazyModuleImp]
-   (val base: BigInt, val interrupts: Int = 0, val size: BigInt = 4096, val concurrency: Int = 0, val beatBytes: Int = 4, undefZero: Boolean = true)
+   (val base: BigInt, val interrupts: Int = 0, val size: BigInt = 4096, val concurrency: Int = 0, val beatBytes: Int = 4, undefZero: Boolean = true, executable: Boolean = false)
    (bundleBuilder: TLRegBundleArg => B)
    (moduleBuilder: (=> B, TLRegisterRouterBase) => M)
-  extends TLRegisterRouterBase(AddressSet(base, size-1), interrupts, concurrency, beatBytes, undefZero)
+  extends TLRegisterRouterBase(AddressSet(base, size-1), interrupts, concurrency, beatBytes, undefZero, executable)
 {
   require (isPow2(size))
   // require (size >= 4096) ... not absolutely required, but highly recommended
diff --git a/src/main/scala/uncore/tilelink2/package.scala b/src/main/scala/uncore/tilelink2/package.scala
index 55ff6b27..3fcd3fc5 100644
--- a/src/main/scala/uncore/tilelink2/package.scala
+++ b/src/main/scala/uncore/tilelink2/package.scala
@@ -5,9 +5,11 @@ import diplomacy._
 
 package object tilelink2
 {
+  type TLInwardNode = InwardNodeHandle[TLClientPortParameters, TLManagerPortParameters, TLBundle]
   type TLOutwardNode = OutwardNodeHandle[TLClientPortParameters, TLManagerPortParameters, TLBundle]
   type TLAsyncOutwardNode = OutwardNodeHandle[TLAsyncClientPortParameters, TLAsyncManagerPortParameters, TLAsyncBundle]
   type IntOutwardNode = OutwardNodeHandle[IntSourcePortParameters, IntSinkPortParameters, Vec[Bool]]
+
   def OH1ToOH(x: UInt) = (x << 1 | UInt(1)) & ~Cat(UInt(0, width=1), x)
   def OH1ToUInt(x: UInt) = OHToUInt(OH1ToOH(x))
   def UIntToOH1(x: UInt, width: Int) = ~(SInt(-1, width=width).asUInt << x)(width-1, 0)
diff --git a/src/main/scala/unittest/Generator.scala b/src/main/scala/unittest/Generator.scala
index 95e114d1..36d64779 100644
--- a/src/main/scala/unittest/Generator.scala
+++ b/src/main/scala/unittest/Generator.scala
@@ -5,6 +5,7 @@ package unittest
 object Generator extends util.GeneratorApp {
   val longName = names.topModuleProject + "." + names.configs
   generateFirrtl
+  generateGraphML
   generateTestSuiteMakefrags // TODO: Needed only for legacy make targets
   generateParameterDump // TODO: Needed only for legacy make targets
 }