Merge pull request #757 from freechipsproject/isp-port

Inter-System-Port

src/main/scala/coreplex/CoreplexNetwork.scala

@@ -13,6 +13,8 @@ trait CoreplexNetwork extends HasCoreplexParameters {
   val module: CoreplexNetworkModule
   def bindingTree: ResourceMap
 
+  val tile_splitter = LazyModule(new TLSplitter)
+
   val l1tol2 = LazyModule(new TLXbar)
   val l1tol2_beatBytes = l1tol2Config.beatBytes
   val l1tol2_lineBytes = p(CacheBlockBytes)
@@ -34,6 +36,7 @@ trait CoreplexNetwork extends HasCoreplexParameters {
   private val l2in_buffer = LazyModule(new TLBuffer)
   private val l2in_fifo = LazyModule(new TLFIFOFixer)
   l1tol2.node :=* l2in_fifo.node
+  l1tol2.node :=* tile_splitter.node
   l2in_fifo.node :=* l2in_buffer.node
   l2in_buffer.node :=* l2in
 
@@ -83,7 +86,7 @@ trait CoreplexNetwork extends HasCoreplexParameters {
   }
 
   // Make topManagers an Option[] so as to avoid LM name reflection evaluating it...
-  lazy val topManagers = Some(ManagerUnification(l1tol2.node.edgesIn.headOption.map(_.manager.managers).getOrElse(Nil)))
+  lazy val topManagers = Some(ManagerUnification(tile_splitter.node.edgesIn.headOption.map(_.manager.managers).getOrElse(Nil)))
   ResourceBinding {
     val managers = topManagers.get
     val max = managers.flatMap(_.address).map(_.max).max

src/main/scala/coreplex/ISPPort.scala

@@ -0,0 +1,44 @@
+// See LICENSE.SiFive for license details.
+
+package coreplex
+
+import Chisel._
+import config._
+import diplomacy._
+import rocket._
+import tile._
+import uncore.tilelink2._
+import util._
+
+trait HasISPPort extends CoreplexNetwork {
+  val module: HasISPPortModule
+
+  // TODO: use ChipLink instead of AsyncTileLink
+  val isp_in = TLAsyncInputNode()
+  val isp_out = TLAsyncOutputNode()
+
+  private val out_xbar = LazyModule(new TLXbar)
+  private val out_nums = LazyModule(new TLNodeNumberer)
+  private val out_async = LazyModule(new TLAsyncCrossingSource)
+  out_xbar.node :=* tile_splitter.node
+  out_nums.node :*= out_xbar.node
+  out_async.node :*= out_nums.node
+  isp_out :*= out_async.node
+
+  private val in_async = LazyModule(new TLAsyncCrossingSink)
+  in_async.node :=* isp_in
+  l1tol2.node :=* in_async.node
+}
+
+trait HasISPPortBundle extends CoreplexNetworkBundle {
+  val outer: HasISPPort
+
+  // TODO: move to IO(...) in Module?
+  val isp_in = outer.isp_in.bundleIn
+  val isp_out = outer.isp_out.bundleOut
+}
+
+trait HasISPPortModule extends CoreplexNetworkModule {
+  val outer: HasISPPort
+  val io: HasISPPortBundle
+}

src/main/scala/coreplex/RocketTiles.scala

@@ -36,7 +36,7 @@ trait HasRocketTiles extends CoreplexRISCVPlatform {
     val pWithExtra = p.alterPartial {
       case TileKey => c
       case BuildRoCC => c.rocc
-      case SharedMemoryTLEdge => l1tol2.node.edgesIn(0)
+      case SharedMemoryTLEdge => tile_splitter.node.edgesIn(0)
     }
 
     val asyncIntXbar  = LazyModule(new IntXbar)
@@ -64,7 +64,7 @@ trait HasRocketTiles extends CoreplexRISCVPlatform {
         val fixer = LazyModule(new TLFIFOFixer)
         buffer.node :=* wrapper.masterNode
         fixer.node :=* buffer.node
-        l1tol2.node :=* fixer.node
+        tile_splitter.node :=* fixer.node
         wrapper.slaveNode :*= cbus.node
         wrapper.asyncIntNode  := asyncIntXbar.intnode
         wrapper.periphIntNode := periphIntXbar.intnode
@@ -82,7 +82,7 @@ trait HasRocketTiles extends CoreplexRISCVPlatform {
         val fixer = LazyModule(new TLFIFOFixer)
         sink.node :=* wrapper.masterNode
         fixer.node :=* sink.node
-        l1tol2.node :=* fixer.node
+        tile_splitter.node :=* fixer.node
         wrapper.slaveNode :*= source.node
         wrapper.asyncIntNode  := asyncIntXbar.intnode
         wrapper.periphIntNode := periphIntXbar.intnode
@@ -102,7 +102,7 @@ trait HasRocketTiles extends CoreplexRISCVPlatform {
         val fixer = LazyModule(new TLFIFOFixer)
         sink.node :=* wrapper.masterNode
         fixer.node :=* sink.node
-        l1tol2.node :=* fixer.node
+        tile_splitter.node :=* fixer.node
         wrapper.slaveNode :*= source.node
         wrapper.asyncIntNode := asyncIntXbar.intnode
         wrapper.periphIntNode := periphIntXbar.intnode

src/main/scala/diplomacy/Nodes.scala

@@ -159,15 +159,13 @@ abstract class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   protected[diplomacy] val numPI: Range.Inclusive)
   extends BaseNode with InwardNode[DI, UI, BI] with OutwardNode[DO, UO, BO]
 {
-  protected[diplomacy] def resolveStarO(i: Int, o: Int): Int
+  protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStar: Int, oStar: Int): (Int, Int)
-  protected[diplomacy] def resolveStarI(i: Int, o: Int): Int
   protected[diplomacy] def mapParamsD(n: Int, p: Seq[DI]): Seq[DO]
   protected[diplomacy] def mapParamsU(n: Int, p: Seq[UO]): Seq[UI]
 
   protected[diplomacy] lazy val (oPortMapping, iPortMapping, oStar, iStar) = {
     val oStars = oBindings.filter { case (_,_,b) => b == BIND_STAR }.size
     val iStars = iBindings.filter { case (_,_,b) => b == BIND_STAR }.size
-    require (oStars + iStars <= 1, s"${name} appears beside a :*= ${iStars} times and a :=* ${oStars} times; at most once is allowed${lazyModule.line}")
     val oKnown = oBindings.map { case (_, n, b) => b match {
       case BIND_ONCE  => 1
       case BIND_QUERY => n.iStar
@@ -176,8 +174,7 @@ abstract class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
       case BIND_ONCE  => 1
       case BIND_QUERY => n.oStar
       case BIND_STAR  => 0 }}.foldLeft(0)(_+_)
-    val oStar = if (oStars > 0) resolveStarO(iKnown, oKnown) else 0
+    val (iStar, oStar) = resolveStar(iKnown, oKnown, iStars, oStars)
-    val iStar = if (iStars > 0) resolveStarI(iKnown, oKnown) else 0
     val oSum = oBindings.map { case (_, n, b) => b match {
       case BIND_ONCE  => 1
       case BIND_QUERY => n.iStar
@@ -274,6 +271,23 @@ abstract class MixedNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   protected[diplomacy] def inputs  = iPorts.map(_._2) zip edgesIn .map(e => inner.labelI(e))
 }
 
+abstract class MixedCustomNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
+  inner: InwardNodeImp [DI, UI, EI, BI],
+  outer: OutwardNodeImp[DO, UO, EO, BO])(
+  numPO: Range.Inclusive,
+  numPI: Range.Inclusive)
+  extends MixedNode(inner, outer)(numPO, numPI)
+{
+  def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int)
+  def mapParamsD(n: Int, p: Seq[DI]): Seq[DO]
+  def mapParamsU(n: Int, p: Seq[UO]): Seq[UI]
+}
+
+abstract class CustomNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(
+  numPO: Range.Inclusive,
+  numPI: Range.Inclusive)
+  extends MixedCustomNode(imp, imp)(numPO, numPI)
+
 class MixedAdapterNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   inner: InwardNodeImp [DI, UI, EI, BI],
   outer: OutwardNodeImp[DO, UO, EO, BO])(
@@ -285,13 +299,15 @@ class MixedAdapterNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   val externalIn: Boolean = true
   val externalOut: Boolean = true
 
-  protected[diplomacy] def resolveStarO(i: Int, o: Int): Int = {
+  protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
-    require (i >= o, s"${name} has ${o} outputs and ${i} inputs; cannot assign ${i-o} edges to resolve :=*${lazyModule.line}")
+    require (oStars + iStars <= 1, s"${name} (an adapter) appears left of a :*= ${iStars} times and right of a :=* ${oStars} times; at most once is allowed${lazyModule.line}")
-    i - o
+    if (oStars > 0) {
+      require (iKnown >= oKnown, s"${name} (an adapter) has ${oKnown} outputs and ${iKnown} inputs; cannot assign ${iKnown-oKnown} edges to resolve :=*${lazyModule.line}")
+      (0, iKnown - oKnown)
+    } else {
+      require (oKnown >= iKnown, s"${name} (an adapter) has ${oKnown} outputs and ${iKnown} inputs; cannot assign ${oKnown-iKnown} edges to resolve :*=${lazyModule.line}")
+      (oKnown - iKnown, 0)
     }
-  protected[diplomacy] def resolveStarI(i: Int, o: Int): Int = {
-    require (o >= i, s"${name} has ${o} outputs and ${i} inputs; cannot assign ${o-i} edges to resolve :*=${lazyModule.line}")
-    o - i
   }
   protected[diplomacy] def mapParamsD(n: Int, p: Seq[DI]): Seq[DO] = {
     require(n == p.size, s"${name} has ${p.size} inputs and ${n} outputs; they must match${lazyModule.line}")
@@ -318,13 +334,10 @@ class MixedNexusNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
   val externalIn: Boolean = true
   val externalOut: Boolean = true
 
-  protected[diplomacy] def resolveStarO(i: Int, o: Int): Int = {
+  protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
-    require (false, "${name} cannot resolve :=*${lazyModule.line}")
+    require (iStars == 0, s"${name} (a nexus) appears left of :*= (perhaps you should flip the '*' to :=*?)${lazyModule.line}")
-    0
+    require (oStars == 0, s"${name} (a nexus) appears right of a :=* (perhaps you should flip the '*' to :*=?)${lazyModule.line}")
-  }
+    (0, 0)
-  protected[diplomacy] def resolveStarI(i: Int, o: Int): Int = {
-    require (false, s"${name} cannot resolve :*=${lazyModule.line}")
-    0
   }
   protected[diplomacy] def mapParamsD(n: Int, p: Seq[DI]): Seq[DO] = Seq.fill(n) { dFn(p) }
   protected[diplomacy] def mapParamsU(n: Int, p: Seq[UO]): Seq[UI] = Seq.fill(n) { uFn(p) }
@@ -343,6 +356,45 @@ class NexusNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(
   numPI: Range.Inclusive = 1 to 999)
     extends MixedNexusNode[D, U, EI, B, D, U, EO, B](imp, imp)(dFn, uFn, numPO, numPI)
 
+case class SplitterArg[T](newSize: Int, ports: Seq[T])
+class MixedSplitterNode[DI, UI, EI, BI <: Data, DO, UO, EO, BO <: Data](
+  inner: InwardNodeImp [DI, UI, EI, BI],
+  outer: OutwardNodeImp[DO, UO, EO, BO])(
+  dFn: SplitterArg[DI] => Seq[DO],
+  uFn: SplitterArg[UO] => Seq[UI],
+  numPO: Range.Inclusive = 1 to 999,
+  numPI: Range.Inclusive = 1 to 999)
+  extends MixedNode(inner, outer)(numPO, numPI)
+{
+  override val externalIn: Boolean = true
+  override val externalOut: Boolean = true
+
+  protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
+    require (oKnown == 0, s"${name} (a splitter) appears right of a := or :*=; use a :=* instead${lazyModule.line}")
+    require (iStars == 0, s"${name} (a splitter) cannot appear left of a :*=; did you mean :=*?${lazyModule.line}")
+    (0, iKnown)
+  }
+  protected[diplomacy] def mapParamsD(n: Int, p: Seq[DI]): Seq[DO] = {
+    require (p.size == 0 || n % p.size == 0, s"Diplomacy bug; splitter inputs do not divide outputs")
+    val out = dFn(SplitterArg(n, p))
+    require (out.size == n, s"${name} created the wrong number of outputs from inputs${lazyModule.line}")
+    out
+  }
+  protected[diplomacy] def mapParamsU(n: Int, p: Seq[UO]): Seq[UI] = {
+    require (n == 0 || p.size % n == 0, s"Diplomacy bug; splitter outputs indivisable by inputs")
+    val out = uFn(SplitterArg(n, p))
+    require (out.size == n, s"${name} created the wrong number of inputs from outputs${lazyModule.line}")
+    out
+  }
+}
+
+class SplitterNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(
+  dFn: SplitterArg[D] => Seq[D],
+  uFn: SplitterArg[U] => Seq[U],
+  numPO: Range.Inclusive = 1 to 999,
+  numPI: Range.Inclusive = 1 to 999)
+    extends MixedSplitterNode[D, U, EI, B, D, U, EO, B](imp, imp)(dFn, uFn, numPO, numPI)
+
 class IdentityNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])
   extends AdapterNode(imp)({s => s}, {s => s})
 
@@ -367,13 +419,12 @@ class SourceNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(po: Seq
   override val externalIn: Boolean = false
   override val externalOut: Boolean = true
 
-  protected[diplomacy] def resolveStarO(i: Int, o: Int): Int = {
+  protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
-    require (po.size >= o, s"${name} has ${o} outputs out of ${po.size}; cannot assign ${po.size - o} edges to resolve :=*${lazyModule.line}")
+    require (oStars <= 1, s"${name} (a source) appears right of a :=* ${oStars} times; at most once is allowed${lazyModule.line}")
-    po.size - o
+    require (iStars == 0, s"${name} (a source) cannot appear left of a :*=${lazyModule.line}")
-  }
+    require (iKnown == 0, s"${name} (a source) cannot appear left of a :=${lazyModule.line}")
-  protected[diplomacy] def resolveStarI(i: Int, o: Int): Int = {
+    require (po.size >= oKnown, s"${name} (a source) has ${oKnown} outputs out of ${po.size}; cannot assign ${po.size - oKnown} edges to resolve :=*${lazyModule.line}")
-    require (false, s"${name} cannot resolve :*=${lazyModule.line}")
+    (0, po.size - oKnown)
-    0
   }
   protected[diplomacy] def mapParamsD(n: Int, p: Seq[D]): Seq[D] = po
   protected[diplomacy] def mapParamsU(n: Int, p: Seq[U]): Seq[U] = Seq()
@@ -387,13 +438,12 @@ class SinkNode[D, U, EO, EI, B <: Data](imp: NodeImp[D, U, EO, EI, B])(pi: Seq[U
   override val externalIn: Boolean = true
   override val externalOut: Boolean = false
 
-  protected[diplomacy] def resolveStarO(i: Int, o: Int): Int = {
+  protected[diplomacy] def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
-    require (false, s"${name} cannot resolve :=*${lazyModule.line}")
+    require (iStars <= 1, s"${name} (a sink) appears left of a :*= ${iStars} times; at most once is allowed${lazyModule.line}")
-    0
+    require (oStars == 0, s"${name} (a sink) cannot appear right of a :=*${lazyModule.line}")
-  }
+    require (oKnown == 0, s"${name} (a sink) cannot appear right of a :=${lazyModule.line}")
-  protected[diplomacy] def resolveStarI(i: Int, o: Int): Int = {
+    require (pi.size >= iKnown, s"${name} (a sink) has ${iKnown} inputs out of ${pi.size}; cannot assign ${pi.size - iKnown} edges to resolve :*=${lazyModule.line}")
-    require (pi.size >= i, s"${name} has ${i} inputs out of ${pi.size}; cannot assign ${pi.size - i} edges to resolve :*=${lazyModule.line}")
+    (pi.size - iKnown, 0)
-    pi.size - i
   }
   protected[diplomacy] def mapParamsD(n: Int, p: Seq[D]): Seq[D] = Seq()
   protected[diplomacy] def mapParamsU(n: Int, p: Seq[U]): Seq[U] = pi

src/main/scala/groundtest/Coreplex.scala

@@ -24,7 +24,7 @@ class GroundTestCoreplex(implicit p: Parameters) extends BaseCoreplex {
       case TileId => i
       case CacheBlockOffsetBits => log2Up(site(CacheBlockBytes))
       case AmoAluOperandBits => site(XLen)
-      case SharedMemoryTLEdge => l1tol2.node.edgesIn(0)
+      case SharedMemoryTLEdge => tile_splitter.node.edgesIn(0)
       case TLId => "L1toL2"
       case TLKey("L1toL2") =>
         TileLinkParameters(
@@ -41,7 +41,7 @@ class GroundTestCoreplex(implicit p: Parameters) extends BaseCoreplex {
   }
 
   val fixer = LazyModule(new TLFIFOFixer)
-  l1tol2.node :=* fixer.node
+  tile_splitter.node :=* fixer.node
   tiles.foreach { fixer.node :=* _.masterNode }
 
   val cbusRAM = LazyModule(new TLRAM(AddressSet(testRamAddr, 0xffff), false, cbus_beatBytes))

src/main/scala/uncore/tilelink2/NodeNumberer.scala

@@ -0,0 +1,73 @@
+// See LICENSE.SiFive for license details.
+
+package uncore.tilelink2
+
+import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
+import config._
+import diplomacy._
+
+case class TLNodeNumbererNode(nodeAddressOffset: Option[Int] = None) extends TLCustomNode(0 to 999, 0 to 999)
+{
+  val externalIn = true
+  val externalOut = true
+
+  def resolveStar(iKnown: Int, oKnown: Int, iStars: Int, oStars: Int): (Int, Int) = {
+    require (oStars + iStars <= 1, s"${name} (a custom adapter) appears left of a :*= ${iStars} times and right of a :=* ${oStars} times; at most once is allowed${lazyModule.line}")
+    if (oStars > 0) {
+      require (iKnown >= oKnown, s"${name} (a custom adapter) has ${oKnown} outputs and ${iKnown} inputs; cannot assign ${iKnown-oKnown} edges to resolve :=*${lazyModule.line}")
+      (0, iKnown - oKnown)
+    } else {
+      require (oKnown >= iKnown, s"${name} (a custom adapter) has ${oKnown} outputs and ${iKnown} inputs; cannot assign ${oKnown-iKnown} edges to resolve :*=${lazyModule.line}")
+      (oKnown - iKnown, 0)
+    }
+  }
+
+  def mapParamsD(n: Int, p: Seq[TLClientPortParameters]): Seq[TLClientPortParameters] = {
+    require(n == p.size, s"${name} has ${p.size} inputs and ${n} outputs; they must match${lazyModule.line}")
+    p
+  }
+
+  def mapParamsU(n: Int, p: Seq[TLManagerPortParameters]): Seq[TLManagerPortParameters] = {
+    require(n == p.size, s"${name} has ${n} inputs and ${p.size} outputs; they must match${lazyModule.line}")
+    val minNodeOffset = log2Ceil(p.map(_.maxAddress).max)
+    val nodeOffset = nodeAddressOffset.getOrElse(minNodeOffset)
+    require (nodeOffset >= minNodeOffset)
+
+    p.zipWithIndex.map { case (mp, i) =>
+      val nodeIndex = BigInt(i+1) << nodeOffset
+      mp.copy(managers = mp.managers.map(m => m.copy(address = m.address.map(a => a.copy(base = a.base | nodeIndex)))))
+    }
+  }
+}
+
+class TLNodeNumberer(nodeAddressOffset: Option[Int] = None)(implicit p: Parameters) extends LazyModule
+{
+  val node = TLNodeNumbererNode(nodeAddressOffset)
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val in  = node.bundleIn
+      val out = node.bundleOut
+    }
+
+    val minNodeOffset = log2Ceil(node.edgesOut.map(_.manager.maxAddress).max)
+    val nodeOffset = nodeAddressOffset.getOrElse(minNodeOffset)
+
+    (io.in zip io.out).zipWithIndex foreach { case ((in, out), i) =>
+      out <> in
+      // a&c address already get truncated
+      in.b.bits.address := (UInt(i+1) << nodeOffset) | out.b.bits.address
+    }
+  }
+}
+
+object TLNodeNumberer
+{
+  // applied to the TL source node; y.node := TLBuffer(x.node)
+  def apply(nodeAddressOffset: Option[Int] = None)(x: TLOutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): TLOutwardNode = {
+    val numberer = LazyModule(new TLNodeNumberer(nodeAddressOffset))
+    numberer.node := x
+    numberer.node
+  }
+}

src/main/scala/uncore/tilelink2/Nodes.scala

@@ -92,6 +92,18 @@ case class TLNexusNode(
   numManagerPorts: Range.Inclusive = 1 to 999)
   extends NexusNode(TLImp)(clientFn, managerFn, numClientPorts, numManagerPorts)
 
+case class TLSplitterNode(
+  clientFn:        SplitterArg[TLClientPortParameters]  => Seq[TLClientPortParameters],
+  managerFn:       SplitterArg[TLManagerPortParameters] => Seq[TLManagerPortParameters],
+  numClientPorts:  Range.Inclusive = 0 to 999,
+  numManagerPorts: Range.Inclusive = 0 to 999)
+  extends SplitterNode(TLImp)(clientFn, managerFn, numClientPorts, numManagerPorts)
+
+abstract class TLCustomNode(
+  numClientPorts:  Range.Inclusive,
+  numManagerPorts: Range.Inclusive)
+  extends CustomNode(TLImp)(numClientPorts, numManagerPorts)
+
 // Nodes passed from an inner module
 case class TLOutputNode() extends OutputNode(TLImp)
 case class TLInputNode() extends InputNode(TLImp)

src/main/scala/uncore/tilelink2/Splitter.scala

@@ -0,0 +1,118 @@
+// See LICENSE.SiFive for license details.
+
+package uncore.tilelink2
+
+import Chisel._
+import config._
+import diplomacy._
+
+class TLSplitter(policy: TLArbiter.Policy = TLArbiter.lowestIndexFirst)(implicit p: Parameters) extends LazyModule
+{
+  val node = TLSplitterNode(
+    clientFn  = { case SplitterArg(newSize, ports) =>
+      if (newSize == 0) Nil else
+      Seq.fill(newSize / ports.size) { ports }.flatten
+    },
+    managerFn = { case SplitterArg(newSize, ports) =>
+      if (newSize == 0) Nil else
+      ports.grouped(newSize).toList.transpose.map { seq =>
+        val fifoIdFactory = TLXbar.relabeler()
+        val outputIdRanges = TLXbar.mapOutputIds(seq)
+        seq(0).copy(
+          minLatency = seq.map(_.minLatency).min,
+          endSinkId = outputIdRanges.map(_.map(_.end).getOrElse(0)).max,
+          managers = ManagerUnification(seq.zipWithIndex.flatMap { case (port, i) =>
+            require (port.beatBytes == seq(0).beatBytes,
+              s"Splitter data widths don't match: ${port.managers.map(_.name)} has ${port.beatBytes}B vs ${seq(0).managers.map(_.name)} has ${seq(0).beatBytes}B")
+            val fifoIdMapper = fifoIdFactory()
+            port.managers map { manager => manager.copy(
+              fifoId = manager.fifoId.map(fifoIdMapper(_))
+            )}
+          })
+        )
+      }
+    })
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val in  = node.bundleIn
+      val out = node.bundleOut
+    }
+
+    def group[T](x: Seq[T]) =
+      if (x.isEmpty) Nil else x.grouped(node.edgesIn.size).toList.transpose
+
+    ((node.edgesIn zip io.in) zip (group(node.edgesOut) zip group(io.out))) foreach {
+      case ((edgeIn, io_in), (edgesOut, io_out)) =>
+
+      // Grab the port ID mapping
+      val outputIdRanges = TLXbar.mapOutputIds(edgesOut.map(_.manager))
+
+      // Find a good mask for address decoding
+      val port_addrs = edgesOut.map(_.manager.managers.map(_.address).flatten)
+      val routingMask = AddressDecoder(port_addrs)
+      val route_addrs = port_addrs.map(_.map(_.widen(~routingMask)).distinct)
+      val outputPorts = route_addrs.map(seq => (addr: UInt) => seq.map(_.contains(addr)).reduce(_ || _))
+
+      // We need an intermediate size of bundle with the widest possible identifiers
+      val wide_bundle = TLBundleParameters.union(Seq(io_in.params) ++ io_out.map(_.params))
+
+      // Transform input bundle sources (sinks use global namespace on both sides)
+      val in = Wire(TLBundle(wide_bundle))
+      in.a <> io_in.a
+      io_in.d <> in.d
+
+      if (edgeIn.client.anySupportProbe && edgesOut.exists(_.manager.anySupportAcquireB)) {
+        in.c <> io_in.c
+        in.e <> io_in.e
+        io_in.b <> in.b
+      } else {
+        in.c.valid := Bool(false)
+        in.e.valid := Bool(false)
+        in.b.ready := Bool(false)
+        io_in.c.ready := Bool(true)
+        io_in.e.ready := Bool(true)
+        io_in.b.valid := Bool(false)
+      }
+
+      // Handle size = 1 gracefully (Chisel3 empty range is broken)
+      def trim(id: UInt, size: Int) = if (size <= 1) UInt(0) else id(log2Ceil(size)-1, 0)
+
+      // Transform output bundle sinks (sources use global namespace on both sides)
+      val out = Wire(Vec(io_out.size, TLBundle(wide_bundle)))
+      for (i <- 0 until out.size) {
+        val r = outputIdRanges(i)
+
+        io_out(i).a <> out(i).a
+        out(i).d <> io_out(i).d
+        out(i).d.bits.sink := io_out(i).d.bits.sink | UInt(r.map(_.start).getOrElse(0))
+
+        if (edgesOut(i).manager.anySupportAcquireB && edgeIn.client.anySupportProbe) {
+          io_out(i).c <> out(i).c
+          io_out(i).e <> out(i).e
+          out(i).b <> io_out(i).b
+          io_out(i).e.bits.sink := trim(out(i).e.bits.sink, r.map(_.size).getOrElse(0))
+        } else {
+          out(i).c.ready := Bool(false)
+          out(i).e.ready := Bool(false)
+          out(i).b.valid := Bool(false)
+          io_out(i).c.valid := Bool(false)
+          io_out(i).e.valid := Bool(false)
+          io_out(i).b.ready := Bool(true)
+        }
+      }
+
+      val requestA = Vec(outputPorts.map { o => o(in.a.bits.address) })
+      val requestC = Vec(outputPorts.map { o => o(in.c.bits.address) })
+      val requestE = Vec(outputIdRanges.map { o => o.map(_.contains(in.e.bits.sink)).getOrElse(Bool(false)) })
+      (out.map(_.a) zip TLXbar.fanout(in.a, requestA)) foreach { case (o, i) => o <> i }
+      (out.map(_.c) zip TLXbar.fanout(in.c, requestC)) foreach { case (o, i) => o <> i }
+      (out.map(_.e) zip TLXbar.fanout(in.e, requestE)) foreach { case (o, i) => o <> i }
+
+      val beatsB = Vec((out zip edgesOut) map { case (o, e) => e.numBeats1(o.b.bits) })
+      val beatsD = Vec((out zip edgesOut) map { case (o, e) => e.numBeats1(o.d.bits) })
+      TLArbiter(policy)(in.b, (beatsB zip out.map(_.b)):_*)
+      TLArbiter(policy)(in.d, (beatsD zip out.map(_.d)):_*)
+    }
+  }
+}

src/main/scala/uncore/tilelink2/Xbar.scala

@@ -8,34 +8,6 @@ import diplomacy._
 
 class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parameters) extends LazyModule
 {
-  def mapInputIds (ports: Seq[TLClientPortParameters ]) = assignRanges(ports.map(_.endSourceId)).map(_.get)
-  def mapOutputIds(ports: Seq[TLManagerPortParameters]) = assignRanges(ports.map(_.endSinkId))
-
-  def assignRanges(sizes: Seq[Int]) = {
-    val pow2Sizes = sizes.map { z => if (z == 0) 0 else 1 << log2Ceil(z) }
-    val tuples = pow2Sizes.zipWithIndex.sortBy(_._1) // record old index, then sort by increasing size
-    val starts = tuples.scanRight(0)(_._1 + _).tail // suffix-sum of the sizes = the start positions
-    val ranges = (tuples zip starts) map { case ((sz, i), st) =>
-      (if (sz == 0) None else Some(IdRange(st, st+sz)), i)
-    }
-    ranges.sortBy(_._2).map(_._1) // Restore orignal order
-  }
-
-  def relabeler() = {
-    var idFactory = 0
-    () => {
-      val fifoMap = scala.collection.mutable.HashMap.empty[Int, Int]
-      (x: Int) => {
-        if (fifoMap.contains(x)) fifoMap(x) else {
-          val out = idFactory
-          idFactory = idFactory + 1
-          fifoMap += (x -> out)
-          out
-        }
-      }
-    }
-  }
-
   val node = TLNexusNode(
     numClientPorts  = 1 to 32,
     numManagerPorts = 1 to 32,
@@ -44,7 +16,7 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
         "An unsafe atomic port can not be combined with any other!")
       seq(0).copy(
         minLatency = seq.map(_.minLatency).min,
-        clients = (mapInputIds(seq) zip seq) flatMap { case (range, port) =>
+        clients = (TLXbar.mapInputIds(seq) zip seq) flatMap { case (range, port) =>
           port.clients map { client => client.copy(
             sourceId = client.sourceId.shift(range.start)
           )}
@@ -52,8 +24,8 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
       )
     },
     managerFn = { seq =>
-      val fifoIdFactory = relabeler()
+      val fifoIdFactory = TLXbar.relabeler()
-      val outputIdRanges = mapOutputIds(seq)
+      val outputIdRanges = TLXbar.mapOutputIds(seq)
       seq(0).copy(
         minLatency = seq.map(_.minLatency).min,
         endSinkId = outputIdRanges.map(_.map(_.end).getOrElse(0)).max,
@@ -75,8 +47,8 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
     }
 
     // Grab the port ID mapping
-    val inputIdRanges = mapInputIds(node.edgesIn.map(_.client))
+    val inputIdRanges = TLXbar.mapInputIds(node.edgesIn.map(_.client))
-    val outputIdRanges = mapOutputIds(node.edgesOut.map(_.manager))
+    val outputIdRanges = TLXbar.mapOutputIds(node.edgesOut.map(_.manager))
 
     // Find a good mask for address decoding
     val port_addrs = node.edgesOut.map(_.manager.managers.map(_.address).flatten)
@@ -105,7 +77,7 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
     }
 
     // We need an intermediate size of bundle with the widest possible identifiers
-    val wide_bundle = io.in(0).params.union(io.out(0).params)
+    val wide_bundle = TLBundleParameters.union(io.in.map(_.params) ++ io.out.map(_.params))
 
     // Handle size = 1 gracefully (Chisel3 empty range is broken)
     def trim(id: UInt, size: Int) = if (size <= 1) UInt(0) else id(log2Ceil(size)-1, 0)
@@ -179,23 +151,12 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
     def transpose[T](x: Seq[Seq[T]]) = Seq.tabulate(x(0).size) { i => Seq.tabulate(x.size) { j => x(j)(i) } }
     def filter[T](data: Seq[T], mask: Seq[Boolean]) = (data zip mask).filter(_._2).map(_._1)
 
-    // Replicate an input port to each output port
-    def fanout[T <: TLChannel](input: DecoupledIO[T], select: Seq[Bool]) = {
-      val filtered = Wire(Vec(select.size, input))
-      for (i <- 0 until select.size) {
-        filtered(i).bits := input.bits
-        filtered(i).valid := input.valid && select(i)
-      }
-      input.ready := Mux1H(select, filtered.map(_.ready))
-      filtered
-    }
-
     // Fanout the input sources to the output sinks
-    val portsAOI = transpose((in  zip requestAIO) map { case (i, r) => fanout(i.a, r) })
+    val portsAOI = transpose((in  zip requestAIO) map { case (i, r) => TLXbar.fanout(i.a, r) })
-    val portsBIO = transpose((out zip requestBOI) map { case (o, r) => fanout(o.b, r) })
+    val portsBIO = transpose((out zip requestBOI) map { case (o, r) => TLXbar.fanout(o.b, r) })
-    val portsCOI = transpose((in  zip requestCIO) map { case (i, r) => fanout(i.c, r) })
+    val portsCOI = transpose((in  zip requestCIO) map { case (i, r) => TLXbar.fanout(i.c, r) })
-    val portsDIO = transpose((out zip requestDOI) map { case (o, r) => fanout(o.d, r) })
+    val portsDIO = transpose((out zip requestDOI) map { case (o, r) => TLXbar.fanout(o.d, r) })
-    val portsEOI = transpose((in  zip requestEIO) map { case (i, r) => fanout(i.e, r) })
+    val portsEOI = transpose((in  zip requestEIO) map { case (i, r) => TLXbar.fanout(i.e, r) })
 
     // Arbitrate amongst the sources
     for (o <- 0 until out.size) {
@@ -219,6 +180,48 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.roundRobin)(implicit p: Parame
   }
 }
 
+object TLXbar
+{
+  def mapInputIds (ports: Seq[TLClientPortParameters ]) = assignRanges(ports.map(_.endSourceId)).map(_.get)
+  def mapOutputIds(ports: Seq[TLManagerPortParameters]) = assignRanges(ports.map(_.endSinkId))
+
+  def assignRanges(sizes: Seq[Int]) = {
+    val pow2Sizes = sizes.map { z => if (z == 0) 0 else 1 << log2Ceil(z) }
+    val tuples = pow2Sizes.zipWithIndex.sortBy(_._1) // record old index, then sort by increasing size
+    val starts = tuples.scanRight(0)(_._1 + _).tail // suffix-sum of the sizes = the start positions
+    val ranges = (tuples zip starts) map { case ((sz, i), st) =>
+      (if (sz == 0) None else Some(IdRange(st, st+sz)), i)
+    }
+    ranges.sortBy(_._2).map(_._1) // Restore orignal order
+  }
+
+  def relabeler() = {
+    var idFactory = 0
+    () => {
+      val fifoMap = scala.collection.mutable.HashMap.empty[Int, Int]
+      (x: Int) => {
+        if (fifoMap.contains(x)) fifoMap(x) else {
+          val out = idFactory
+          idFactory = idFactory + 1
+          fifoMap += (x -> out)
+          out
+        }
+      }
+    }
+  }
+
+  // Replicate an input port to each output port
+  def fanout[T <: TLChannel](input: DecoupledIO[T], select: Seq[Bool]) = {
+    val filtered = Wire(Vec(select.size, input))
+    for (i <- 0 until select.size) {
+      filtered(i).bits := input.bits
+      filtered(i).valid := input.valid && select(i)
+    }
+    input.ready := Mux1H(select, filtered.map(_.ready))
+    filtered
+  }
+}
+
 /** Synthesizeable unit tests */
 import unittest._