Merge pull request #515 from ucb-bar/cache-cork

Cache cork

regression/Makefile

@@ -59,6 +59,7 @@ endif
 ifeq ($(SUITE),GroundtestSuiteA)
 PROJECT=groundtest
 CONFIGS=MemtestConfig MemtestBufferlessConfig MemtestStatelessConfig
+# FancyMemtestConfig takes too long to compile
 endif
 
 ifeq ($(SUITE),GroundtestSuiteB)

src/main/scala/coreplex/Configs.scala

@@ -143,12 +143,11 @@ class WithBufferlessBroadcastHub extends Config((site, here, up) => {
  * DO NOT use this configuration.
  */
 class WithStatelessBridge extends Config((site, here, up) => {
-/* !!! FIXME
+  case BankedL2Config => up(BankedL2Config, site).copy(coherenceManager = { case q =>
-    case BankedL2Config => up(BankedL2Config, site).copy(coherenceManager = { case (_, _) =>
+    implicit val p = q
-      val pass = LazyModule(new TLBuffer(0)(site))
+    val cork = LazyModule(new TLCacheCork(unsafe = true))
-      (pass.node, pass.node)
+    (cork.node, TLWidthWidget(p(L1toL2Config).beatBytes)(cork.node))
-    })
+  })
-*/
   case DCacheKey => up(DCacheKey, site).copy(nMSHRs = 0)
 })
 

src/main/scala/coreplex/CoreplexNetwork.scala

@@ -55,7 +55,7 @@ trait CoreplexNetworkModule extends HasCoreplexParameters {
     val prot = (if (manager.supportsGet)     "R" else "") +
                (if (manager.supportsPutFull) "W" else "") +
                (if (manager.executable)      "X" else "") +
-               (if (manager.supportsAcquire) " [C]" else "")
+               (if (manager.supportsAcquireB) " [C]" else "")
     manager.address.foreach { a =>
       println(f"\t${manager.name}%s ${a.base}%x - ${a.base+a.mask+1}%x, $prot")
     }
@@ -70,27 +70,23 @@ trait BankedL2CoherenceManagers extends CoreplexNetwork {
   require (isPow2(l2Config.nBanksPerChannel))
   require (isPow2(l1tol2_lineBytes))
 
-  val mem = Seq.fill(l2Config.nMemoryChannels) {
+  val mem = TLOutputNode()
+  for (i <- 0 until l2Config.nMemoryChannels) {
     val bankBar = LazyModule(new TLXbar)
-    val output = TLOutputNode()
 
-    output := bankBar.node
+    mem := bankBar.node
     val mask = ~BigInt((l2Config.nBanksPerChannel-1) * l1tol2_lineBytes)
     for (i <- 0 until l2Config.nBanksPerChannel) {
       val (in, out) = l2Config.coherenceManager(p)
       in := TLFilter(AddressSet(i * l1tol2_lineBytes, mask))(l1tol2.node)
       bankBar.node := out
     }
-
-    output
   }
 }
 
 trait BankedL2CoherenceManagersBundle extends CoreplexNetworkBundle {
   val outer: BankedL2CoherenceManagers
-
+  val mem = outer.mem.bundleOut
-  require (l2Config.nMemoryChannels <= 1, "Seq in Chisel Bundle needed to support > 1") // !!!
-  val mem = outer.mem.map(_.bundleOut).toList.headOption // .headOption should be removed !!!
 }
 
 trait BankedL2CoherenceManagersModule extends CoreplexNetworkModule {

src/main/scala/diplomacy/Nodes.scala

@@ -224,29 +224,29 @@ class SinkNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(pi: P
   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)
+class BlindOutputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(pi: Seq[PI])
-  extends SimpleNode(imp)({case (0, _) => Seq()}, {case (n, Seq()) => Seq.fill(n)(pi)}, 0 to 0, 1 to 1)
+  extends SimpleNode(imp)({case (0, _) => Seq()}, {case (_, Seq()) => pi}, 0 to 0, pi.size to pi.size)
 {
   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)
+class BlindInputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(po: Seq[PO])
-  extends SimpleNode(imp)({case (n, Seq()) => Seq.fill(n)(po)}, {case (0, _) => Seq()}, 1 to 1, 0 to 0)
+  extends SimpleNode(imp)({case (_, Seq()) => po}, {case (0, _) => Seq()}, po.size to po.size, 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)
+class InternalOutputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(pi: Seq[PI])
-  extends SimpleNode(imp)({case (0, _) => Seq()}, {case (n, Seq()) => Seq.fill(n)(pi)}, 0 to 0, 1 to 1)
+  extends SimpleNode(imp)({case (0, _) => Seq()}, {case (_, Seq()) => pi}, 0 to 0, pi.size to pi.size)
 {
   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)
+class InternalInputNode[PO, PI, EO, EI, B <: Data](imp: NodeImp[PO, PI, EO, EI, B])(po: Seq[PO])
-  extends SimpleNode(imp)({case (n, Seq()) => Seq.fill(n)(po)}, {case (0, _) => Seq()}, 1 to 1, 0 to 0)
+  extends SimpleNode(imp)({case (_, Seq()) => po}, {case (0, _) => Seq()}, po.size to po.size, 0 to 0)
 {
   override val wire = true
   override lazy val bundleIn = bundleOut

src/main/scala/groundtest/Configs.scala

@@ -40,7 +40,7 @@ class MemtestStatelessConfig extends Config(
 // Test ALL the things
 class FancyMemtestConfig extends Config(
   new WithNGenerators(1, 2) ++ new WithNCores(2) ++ new WithMemtest ++
-  new WithNMemoryChannels(1) ++ new WithNBanksPerMemChannel(4) ++ // !!! waiting on Chisel3 support for 2 channels
+  new WithNMemoryChannels(2) ++ new WithNBanksPerMemChannel(4) ++
   new WithL2Cache ++ new GroundTestConfig)
 
 class CacheFillTestConfig extends Config(

src/main/scala/groundtest/TestHarness.scala

@@ -16,11 +16,6 @@ class TestHarness(implicit p: Parameters) extends Module {
   val dut = Module(LazyModule(new GroundTestTop).module)
   io.success := dut.io.success
 
-  if (dut.io.mem_axi4.nonEmpty) {
+  val channels = p(coreplex.BankedL2Config).nMemoryChannels
-    val memSize = p(ExtMem).size
+  if (channels > 0) Module(LazyModule(new SimAXIMem(channels)).module).io.axi4 <> dut.io.mem_axi4
-    require(memSize % dut.io.mem_axi4.size == 0)
-    for (axi4 <- dut.io.mem_axi4) {
-      Module(LazyModule(new SimAXIMem(memSize / dut.io.mem_axi4.size)).module).io.axi4 <> axi4
-    }
-  }
 }

src/main/scala/groundtest/Top.scala

@@ -17,7 +17,7 @@ class GroundTestTop(implicit p: Parameters) extends BaseTop
 
   socBus.node := coreplex.mmio
   coreplex.mmioInt := intBus.intnode
-  (mem zip coreplex.mem) foreach { case (m, c) => m := c }
+  mem.foreach { _ := coreplex.mem }
 }
 
 class GroundTestTopBundle[+L <: GroundTestTop](_outer: L) extends BaseTopBundle(_outer)

src/main/scala/rocket/DCache.scala

@@ -253,7 +253,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   val access_address = s2_req.addr
   val a_size = s2_req.typ(MT_SZ-2, 0)
   val a_data = Fill(beatWords, pstore1_storegen.data)
-  val acquire = if (edge.manager.anySupportAcquire) {
+  val acquire = if (edge.manager.anySupportAcquireB) {
     edge.Acquire(a_source, acquire_address, lgCacheBlockBytes, s2_grow_param)._2 // Cacheability checked by tlb
   } else {
     Wire(new TLBundleA(edge.bundle))
@@ -373,7 +373,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
                                   b = probe_bits,
                                   reportPermissions = TLPermissions.NtoN)
 
-  val voluntaryReleaseMessage = if (edge.manager.anySupportAcquire) {
+  val voluntaryReleaseMessage = if (edge.manager.anySupportAcquireB) {
                                 edge.Release(
                                   fromSource = UInt(maxUncachedInFlight - 1),
                                   toAddress = probe_bits.address,

src/main/scala/rocket/NBDcache.scala

@@ -332,7 +332,7 @@ class MSHRFile(implicit edge: TLEdgeOut, cfg: DCacheConfig, p: Parameters) exten
   }
 
   // determine if the request is cacheable or not
-  val cacheable = edge.manager.supportsAcquireFast(io.req.bits.addr, lgCacheBlockBytes)
+  val cacheable = edge.manager.supportsAcquireBFast(io.req.bits.addr, lgCacheBlockBytes)
 
   val sdq_val = Reg(init=Bits(0, cfg.nSDQ))
   val sdq_alloc_id = PriorityEncoder(~sdq_val(cfg.nSDQ-1,0))

src/main/scala/rocket/TLB.scala

@@ -73,16 +73,19 @@ class TLB(implicit edge: TLEdgeOut, val p: Parameters) extends Module with HasTL
   val prot_r = fastCheck(_.supportsGet)
   val prot_w = fastCheck(_.supportsPutFull)
   val prot_x = fastCheck(_.executable)
-  val cacheable = fastCheck(_.supportsAcquire)
+  val cacheable = fastCheck(_.supportsAcquireB)
+  val xferSizes = TransferSizes(cacheBlockBytes, cacheBlockBytes)
   val allSizes = TransferSizes(1, cacheBlockBytes)
   val amoSizes = TransferSizes(1, xLen/8)
   edge.manager.managers.foreach { m =>
     require (m.minAlignment >= 4096, s"MemoryMap region ${m.name} must be page-aligned (is ${m.minAlignment})")
-    require (!m.supportsGet     || m.supportsGet    .contains(allSizes), s"MemoryMap region ${m.name} only supports ${m.supportsGet} Get, but must support ${allSizes}")
+    require (!m.supportsGet        || m.supportsGet       .contains(allSizes),  s"MemoryMap region ${m.name} only supports ${m.supportsGet} Get, but must support ${allSizes}")
-    require (!m.supportsPutFull || m.supportsPutFull.contains(allSizes), s"MemoryMap region ${m.name} only supports ${m.supportsPutFull} PutFull, but must support ${allSizes}")
+    require (!m.supportsPutFull    || m.supportsPutFull   .contains(allSizes),  s"MemoryMap region ${m.name} only supports ${m.supportsPutFull} PutFull, but must support ${allSizes}")
-    require (!m.supportsAcquire || m.supportsAcquire.contains(allSizes), s"MemoryMap region ${m.name} only supports ${m.supportsAcquire} Acquire, but must support ${allSizes}")
+    require (!m.supportsAcquireB   || m.supportsAcquireB  .contains(xferSizes), s"MemoryMap region ${m.name} only supports ${m.supportsAcquireB} AcquireB, but must support ${xferSizes}")
-    require (!m.supportsLogical || m.supportsLogical.contains(amoSizes), s"MemoryMap region ${m.name} only supports ${m.supportsLogical} Logical, but must support ${amoSizes}")
+    require (!m.supportsAcquireT   || m.supportsAcquireT  .contains(xferSizes), s"MemoryMap region ${m.name} only supports ${m.supportsAcquireT} AcquireT, but must support ${xferSizes}")
-    require (!m.supportsArithmetic || m.supportsArithmetic.contains(amoSizes), s"MemoryMap region ${m.name} only supports ${m.supportsArithmetic} Arithmetic, but must support ${amoSizes}")
+    require (!m.supportsLogical    || m.supportsLogical   .contains(amoSizes),  s"MemoryMap region ${m.name} only supports ${m.supportsLogical} Logical, but must support ${amoSizes}")
+    require (!m.supportsArithmetic || m.supportsArithmetic.contains(amoSizes),  s"MemoryMap region ${m.name} only supports ${m.supportsArithmetic} Arithmetic, but must support ${amoSizes}")
+    require (m.supportsAcquireT || !m.supportsPutFull || !m.supportsAcquireB,   s"MemoryMap region ${m.name} supports PutFull and AcquireB but not AcquireT")
   }
 
   val lookup_tag = Cat(io.ptw.ptbr.asid, io.req.bits.vpn(vpnBits-1,0))

src/main/scala/rocketchip/Periphery.scala

@@ -87,11 +87,11 @@ trait PeripheryMasterAXI4Mem {
   private val config = p(ExtMem)
   private val channels = p(BankedL2Config).nMemoryChannels
 
-  val mem_axi4 = Seq.tabulate(channels) { i =>
+  val mem_axi4 = AXI4BlindOutputNode(Seq.tabulate(channels) { i =>
     val c_size = config.size/channels
     val c_base = config.base + c_size*i
 
-    AXI4BlindOutputNode(AXI4SlavePortParameters(
+    AXI4SlavePortParameters(
       slaves = Seq(AXI4SlaveParameters(
         address       = List(AddressSet(c_base, c_size-1)),
         regionType    = RegionType.UNCACHED,   // cacheable
@@ -99,12 +99,12 @@ trait PeripheryMasterAXI4Mem {
         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 = config.beatBytes))
+      beatBytes = config.beatBytes)
-  }
+  })
 
-  val mem = mem_axi4.map { node =>
+  val mem = Seq.fill(channels) {
     val converter = LazyModule(new TLToAXI4(config.idBits))
-    node := AXI4Buffer()(converter.node)
+    mem_axi4 := AXI4Buffer()(converter.node)
     converter.node
   }
 }
@@ -113,7 +113,7 @@ trait PeripheryMasterAXI4MemBundle {
   this: TopNetworkBundle {
     val outer: PeripheryMasterAXI4Mem
   } =>
-  val mem_axi4 = outer.mem_axi4.map(_.bundleOut).toList.headOption // !!! remove headOption when Seq supported
+  val mem_axi4 = outer.mem_axi4.bundleOut
 }
 
 trait PeripheryMasterAXI4MemModule {
@@ -130,14 +130,14 @@ trait PeripheryMasterAXI4MMIO {
   this: TopNetwork =>
 
   private val config = p(ExtBus)
-  val mmio_axi4 = AXI4BlindOutputNode(AXI4SlavePortParameters(
+  val mmio_axi4 = AXI4BlindOutputNode(Seq(AXI4SlavePortParameters(
     slaves = Seq(AXI4SlaveParameters(
       address       = List(AddressSet(BigInt(config.base), config.size-1)),
       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 = config.beatBytes))
+    beatBytes = config.beatBytes)))
 
   mmio_axi4 :=
     AXI4Buffer()(
@@ -167,9 +167,9 @@ trait PeripheryMasterAXI4MMIOModule {
 // PeripherySlaveAXI4 is an example, make your own cake pattern like this one.
 trait PeripherySlaveAXI4 extends L2Crossbar {
   private val config = p(ExtIn)
-  val l2_axi4 = AXI4BlindInputNode(AXI4MasterPortParameters(
+  val l2_axi4 = AXI4BlindInputNode(Seq(AXI4MasterPortParameters(
     masters = Seq(AXI4MasterParameters(
-      id = IdRange(0, 1 << config.idBits)))))
+      id = IdRange(0, 1 << config.idBits))))))
 
   l2.node :=
     TLSourceShrinker(1 << config.sourceBits)(
@@ -197,14 +197,14 @@ trait PeripheryMasterTLMMIO {
   this: TopNetwork =>
 
   private val config = p(ExtBus)
-  val mmio_tl = TLBlindOutputNode(TLManagerPortParameters(
+  val mmio_tl = TLBlindOutputNode(Seq(TLManagerPortParameters(
     managers = Seq(TLManagerParameters(
       address            = List(AddressSet(BigInt(config.base), config.size-1)),
       executable         = true,
       supportsGet        = TransferSizes(1, cacheBlockBytes),
       supportsPutFull    = TransferSizes(1, cacheBlockBytes),
       supportsPutPartial = TransferSizes(1, cacheBlockBytes))),
-    beatBytes = config.beatBytes))
+    beatBytes = config.beatBytes)))
 
   mmio_tl :=
     TLBuffer()(
@@ -233,9 +233,9 @@ trait PeripheryMasterTLMMIOModule {
 // NOTE: this port is NOT allowed to issue Acquires
 trait PeripherySlaveTL extends L2Crossbar {
   private val config = p(ExtIn)
-  val l2_tl = TLBlindInputNode(TLClientPortParameters(
+  val l2_tl = TLBlindInputNode(Seq(TLClientPortParameters(
     clients = Seq(TLClientParameters(
-      sourceId = IdRange(0, 1 << config.idBits)))))
+      sourceId = IdRange(0, 1 << config.idBits))))))
 
   l2.node :=
     TLSourceShrinker(1 << config.sourceBits)(

src/main/scala/rocketchip/RocketPlexMaster.scala

@@ -19,7 +19,7 @@ trait RocketPlexMaster extends L2Crossbar {
   coreplex.l2in := l2.node
   socBus.node := coreplex.mmio
   coreplex.mmioInt := intBus.intnode
-  (mem zip coreplex.mem) foreach { case (m, c) => m := c }
+  mem.foreach { _ := coreplex.mem }
 }
 
 trait RocketPlexMasterBundle extends L2CrossbarBundle {

src/main/scala/rocketchip/TestHarness.scala

@@ -18,13 +18,8 @@ class TestHarness()(implicit p: Parameters) extends Module {
   for (int <- dut.io.interrupts(0))
     int := Bool(false)
 
-  if (dut.io.mem_axi4.nonEmpty) {
+  val channels = p(coreplex.BankedL2Config).nMemoryChannels
-    val memSize = p(ExtMem).size
+  if (channels > 0) Module(LazyModule(new SimAXIMem(channels)).module).io.axi4 <> dut.io.mem_axi4
-    require(memSize % dut.io.mem_axi4.size == 0)
-    for (axi4 <- dut.io.mem_axi4) {
-      Module(LazyModule(new SimAXIMem(memSize / dut.io.mem_axi4.size)).module).io.axi4 <> axi4
-    }
-  }
 
   if (!p(IncludeJtagDTM)) {
     val dtm = Module(new SimDTM).connect(clock, reset, dut.io.debug.get, io.success)
@@ -32,7 +27,7 @@ class TestHarness()(implicit p: Parameters) extends Module {
      val jtag = Module(new JTAGVPI).connect(dut.io.jtag.get, reset, io.success)		
   }
 
-  val mmio_sim = Module(LazyModule(new SimAXIMem(4096)).module)
+  val mmio_sim = Module(LazyModule(new SimAXIMem(1, 4096)).module)
   mmio_sim.io.axi4 <> dut.io.mmio_axi4
 
   val l2_axi4 = dut.io.l2_axi4(0)
@@ -43,12 +38,19 @@ class TestHarness()(implicit p: Parameters) extends Module {
   l2_axi4.b .ready := Bool(true)
 }
 
-class SimAXIMem(size: BigInt)(implicit p: Parameters) extends LazyModule {
+class SimAXIMem(channels: Int, forceSize: BigInt = 0)(implicit p: Parameters) extends LazyModule {
   val config = p(ExtMem)
+  val totalSize = if (forceSize > 0) forceSize else BigInt(config.size)
+  val size = totalSize / channels
+  require(totalSize % channels == 0)
 
-  val node = AXI4BlindInputNode(AXI4MasterPortParameters(Seq(AXI4MasterParameters(IdRange(0, 1 << config.idBits)))))
+  val node = AXI4BlindInputNode(Seq.fill(channels) {
-  val sram = LazyModule(new AXI4RAM(AddressSet(0, size-1), beatBytes = config.beatBytes))
+    AXI4MasterPortParameters(Seq(AXI4MasterParameters(IdRange(0, 1 << config.idBits))))})
-  sram.node := AXI4Buffer()(AXI4Fragmenter(maxInFlight = 4)(node))
+
+  for (i <- 0 until channels) {
+    val sram = LazyModule(new AXI4RAM(AddressSet(0, size-1), beatBytes = config.beatBytes))
+    sram.node := AXI4Buffer()(AXI4Fragmenter(maxInFlight = 4)(node))
+  }
 
   lazy val module = new LazyModuleImp(this) {
     val io = new Bundle {

src/main/scala/uncore/ahb/Nodes.scala

@@ -11,14 +11,9 @@ object AHBImp extends NodeImp[AHBMasterPortParameters, AHBSlavePortParameters, A
 {
   def edgeO(pd: AHBMasterPortParameters, pu: AHBSlavePortParameters): AHBEdgeParameters = AHBEdgeParameters(pd, pu)
   def edgeI(pd: AHBMasterPortParameters, pu: AHBSlavePortParameters): AHBEdgeParameters = AHBEdgeParameters(pd, pu)
-  def bundleO(eo: Seq[AHBEdgeParameters]): Vec[AHBBundle] = {
+
-    require (!eo.isEmpty)
+  def bundleO(eo: Seq[AHBEdgeParameters]): Vec[AHBBundle] = Vec(eo.size, AHBBundle(AHBBundleParameters.union(eo.map(_.bundle))))
-    Vec(eo.size, AHBBundle(eo.map(_.bundle).reduce(_.union(_))))
+  def bundleI(ei: Seq[AHBEdgeParameters]): Vec[AHBBundle] = Vec(ei.size, AHBBundle(AHBBundleParameters.union(ei.map(_.bundle))))
-  }
-  def bundleI(ei: Seq[AHBEdgeParameters]): Vec[AHBBundle] = {
-    require (!ei.isEmpty)
-    Vec(ei.size, AHBBundle(ei.map(_.bundle).reduce(_.union(_))))
-  }
 
   def colour = "#00ccff" // bluish
   override def labelI(ei: AHBEdgeParameters) = (ei.slave.beatBytes * 8).toString
@@ -52,8 +47,8 @@ case class AHBOutputNode() extends OutputNode(AHBImp)
 case class AHBInputNode() extends InputNode(AHBImp)
 
 // Nodes used for external ports
-case class AHBBlindOutputNode(portParams: AHBSlavePortParameters) extends BlindOutputNode(AHBImp)(portParams)
+case class AHBBlindOutputNode(portParams: Seq[AHBSlavePortParameters]) extends BlindOutputNode(AHBImp)(portParams)
-case class AHBBlindInputNode(portParams: AHBMasterPortParameters) extends BlindInputNode(AHBImp)(portParams)
+case class AHBBlindInputNode(portParams: Seq[AHBMasterPortParameters]) extends BlindInputNode(AHBImp)(portParams)
 
-case class AHBInternalOutputNode(portParams: AHBSlavePortParameters) extends InternalOutputNode(AHBImp)(portParams)
+case class AHBInternalOutputNode(portParams: Seq[AHBSlavePortParameters]) extends InternalOutputNode(AHBImp)(portParams)
-case class AHBInternalInputNode(portParams: AHBMasterPortParameters) extends InternalInputNode(AHBImp)(portParams)
+case class AHBInternalInputNode(portParams: Seq[AHBMasterPortParameters]) extends InternalInputNode(AHBImp)(portParams)

src/main/scala/uncore/ahb/Parameters.scala

@@ -85,6 +85,9 @@ case class AHBBundleParameters(
 
 object AHBBundleParameters
 {
+  val emptyBundleParams = AHBBundleParameters(addrBits = 1, dataBits = 8)
+  def union(x: Seq[AHBBundleParameters]) = x.foldLeft(emptyBundleParams)((x,y) => x.union(y))
+
   def apply(master: AHBMasterPortParameters, slave: AHBSlavePortParameters) =
     new AHBBundleParameters(
       addrBits = log2Up(slave.maxAddress+1),

src/main/scala/uncore/apb/Nodes.scala

@@ -11,14 +11,9 @@ object APBImp extends NodeImp[APBMasterPortParameters, APBSlavePortParameters, A
 {
   def edgeO(pd: APBMasterPortParameters, pu: APBSlavePortParameters): APBEdgeParameters = APBEdgeParameters(pd, pu)
   def edgeI(pd: APBMasterPortParameters, pu: APBSlavePortParameters): APBEdgeParameters = APBEdgeParameters(pd, pu)
-  def bundleO(eo: Seq[APBEdgeParameters]): Vec[APBBundle] = {
+
-    require (!eo.isEmpty)
+  def bundleO(eo: Seq[APBEdgeParameters]): Vec[APBBundle] = Vec(eo.size, APBBundle(APBBundleParameters.union(eo.map(_.bundle))))
-    Vec(eo.size, APBBundle(eo.map(_.bundle).reduce(_.union(_))))
+  def bundleI(ei: Seq[APBEdgeParameters]): Vec[APBBundle] = Vec(ei.size, APBBundle(APBBundleParameters.union(ei.map(_.bundle))))
-  }
-  def bundleI(ei: Seq[APBEdgeParameters]): Vec[APBBundle] = {
-    require (!ei.isEmpty)
-    Vec(ei.size, APBBundle(ei.map(_.bundle).reduce(_.union(_))))
-  }
 
   def colour = "#00ccff" // bluish
   override def labelI(ei: APBEdgeParameters) = (ei.slave.beatBytes * 8).toString
@@ -52,8 +47,8 @@ case class APBOutputNode() extends OutputNode(APBImp)
 case class APBInputNode() extends InputNode(APBImp)
 
 // Nodes used for external ports
-case class APBBlindOutputNode(portParams: APBSlavePortParameters) extends BlindOutputNode(APBImp)(portParams)
+case class APBBlindOutputNode(portParams: Seq[APBSlavePortParameters]) extends BlindOutputNode(APBImp)(portParams)
-case class APBBlindInputNode(portParams: APBMasterPortParameters) extends BlindInputNode(APBImp)(portParams)
+case class APBBlindInputNode(portParams: Seq[APBMasterPortParameters]) extends BlindInputNode(APBImp)(portParams)
 
-case class APBInternalOutputNode(portParams: APBSlavePortParameters) extends InternalOutputNode(APBImp)(portParams)
+case class APBInternalOutputNode(portParams: Seq[APBSlavePortParameters]) extends InternalOutputNode(APBImp)(portParams)
-case class APBInternalInputNode(portParams: APBMasterPortParameters) extends InternalInputNode(APBImp)(portParams)
+case class APBInternalInputNode(portParams: Seq[APBMasterPortParameters]) extends InternalInputNode(APBImp)(portParams)

src/main/scala/uncore/apb/Parameters.scala

@@ -72,6 +72,9 @@ case class APBBundleParameters(
 
 object APBBundleParameters
 {
+  val emptyBundleParams = APBBundleParameters(addrBits = 1, dataBits = 8)
+  def union(x: Seq[APBBundleParameters]) = x.foldLeft(emptyBundleParams)((x,y) => x.union(y))
+
   def apply(master: APBMasterPortParameters, slave: APBSlavePortParameters) =
     new APBBundleParameters(
       addrBits = log2Up(slave.maxAddress+1),

src/main/scala/uncore/axi4/Nodes.scala

@@ -11,14 +11,9 @@ object AXI4Imp extends NodeImp[AXI4MasterPortParameters, AXI4SlavePortParameters
 {
   def edgeO(pd: AXI4MasterPortParameters, pu: AXI4SlavePortParameters): AXI4EdgeParameters = AXI4EdgeParameters(pd, pu)
   def edgeI(pd: AXI4MasterPortParameters, pu: AXI4SlavePortParameters): AXI4EdgeParameters = AXI4EdgeParameters(pd, pu)
-  def bundleO(eo: Seq[AXI4EdgeParameters]): Vec[AXI4Bundle] = {
+
-    require (!eo.isEmpty)
+  def bundleO(eo: Seq[AXI4EdgeParameters]): Vec[AXI4Bundle] = Vec(eo.size, AXI4Bundle(AXI4BundleParameters.union(eo.map(_.bundle))))
-    Vec(eo.size, AXI4Bundle(eo.map(_.bundle).reduce(_.union(_))))
+  def bundleI(ei: Seq[AXI4EdgeParameters]): Vec[AXI4Bundle] = Vec(ei.size, AXI4Bundle(AXI4BundleParameters.union(ei.map(_.bundle))))
-  }
-  def bundleI(ei: Seq[AXI4EdgeParameters]): Vec[AXI4Bundle] = {
-    require (!ei.isEmpty)
-    Vec(ei.size, AXI4Bundle(ei.map(_.bundle).reduce(_.union(_))))
-  }
 
   def colour = "#00ccff" // bluish
   override def labelI(ei: AXI4EdgeParameters) = (ei.slave.beatBytes * 8).toString
@@ -52,8 +47,8 @@ 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 AXI4BlindOutputNode(portParams: Seq[AXI4SlavePortParameters]) extends BlindOutputNode(AXI4Imp)(portParams)
-case class AXI4BlindInputNode(portParams: AXI4MasterPortParameters) extends BlindInputNode(AXI4Imp)(portParams)
+case class AXI4BlindInputNode(portParams: Seq[AXI4MasterPortParameters]) extends BlindInputNode(AXI4Imp)(portParams)
 
-case class AXI4InternalOutputNode(portParams: AXI4SlavePortParameters) extends InternalOutputNode(AXI4Imp)(portParams)
+case class AXI4InternalOutputNode(portParams: Seq[AXI4SlavePortParameters]) extends InternalOutputNode(AXI4Imp)(portParams)
-case class AXI4InternalInputNode(portParams: AXI4MasterPortParameters) extends InternalInputNode(AXI4Imp)(portParams)
+case class AXI4InternalInputNode(portParams: Seq[AXI4MasterPortParameters]) extends InternalInputNode(AXI4Imp)(portParams)

src/main/scala/uncore/axi4/Parameters.scala

@@ -101,6 +101,9 @@ case class AXI4BundleParameters(
 
 object AXI4BundleParameters
 {
+  val emptyBundleParams = AXI4BundleParameters(addrBits=1, dataBits=8, idBits=1)
+  def union(x: Seq[AXI4BundleParameters]) = x.foldLeft(emptyBundleParams)((x,y) => x.union(y))
+
   def apply(master: AXI4MasterPortParameters, slave: AXI4SlavePortParameters) =
     new AXI4BundleParameters(
       addrBits = log2Up(slave.maxAddress+1),

src/main/scala/uncore/tilelink2/AtomicAutomata.scala

@@ -253,7 +253,7 @@ class TLAtomicAutomata(logical: Boolean = true, arithmetic: Boolean = true, conc
       in.d.bits := out.d.bits
     }
 
-    if (edgeOut.manager.anySupportAcquire && edgeIn.client.anySupportProbe) {
+    if (edgeOut.manager.anySupportAcquireB && edgeIn.client.anySupportProbe) {
       in.b.valid := out.b.valid
       out.b.ready := in.b.ready
       in.b.bits := out.b.bits

src/main/scala/uncore/tilelink2/Broadcast.scala

@@ -22,22 +22,24 @@ class TLBroadcast(lineBytes: Int, numTrackers: Int = 4, bufferless: Boolean = fa
         endSinkId  = numTrackers,
         managers   = mp.managers.map { m =>
           // We are the last level manager
-          require (m.regionType != RegionType.CACHED)
+          require (!m.supportsAcquireB)
-          require (m.regionType != RegionType.TRACKED)
-          require (!m.supportsAcquire)
           // We only manage addresses which are uncached
-          if (m.regionType == RegionType.UNCACHED && m.supportsGet) {
+          if (m.regionType == RegionType.UNCACHED) {
             // The device had better support line transfers
             val lowerBound = max(m.supportsPutFull.min, m.supportsGet.min)
             require (!m.supportsPutFull || m.supportsPutFull.contains(lineBytes))
             require (!m.supportsGet     || m.supportsGet    .contains(lineBytes))
             m.copy(
               regionType         = RegionType.TRACKED,
-              supportsAcquire    = TransferSizes(lowerBound, lineBytes),
+              supportsAcquireB   = TransferSizes(lowerBound, lineBytes),
+              supportsAcquireT   = if (m.supportsPutFull) TransferSizes(lowerBound, lineBytes) else TransferSizes.none,
               // truncate supported accesses to lineBytes (we only ever probe for one line)
               supportsPutFull    = TransferSizes(m.supportsPutFull   .min, min(m.supportsPutFull   .max, lineBytes)),
               supportsPutPartial = TransferSizes(m.supportsPutPartial.min, min(m.supportsPutPartial.max, lineBytes)),
               supportsGet        = TransferSizes(m.supportsGet       .min, min(m.supportsGet       .max, lineBytes)),
+              supportsHint       = TransferSizes(m.supportsHint      .min, min(m.supportsHint      .max, lineBytes)),
+              supportsArithmetic = TransferSizes(m.supportsArithmetic.min, min(m.supportsArithmetic.max, lineBytes)),
+              supportsLogical    = TransferSizes(m.supportsLogical   .min, min(m.supportsLogical   .max, lineBytes)),
               fifoId             = None // trackers do not respond in FIFO order!
             )
           } else {

src/main/scala/uncore/tilelink2/Buffer.scala

@@ -31,7 +31,7 @@ class TLBuffer(a: Int = 2, b: Int = 2, c: Int = 2, d: Int = 2, e: Int = 2, pipe:
       if (a>0) { out.a <> Queue(in .a, a, pipe && a<2) } else { out.a <> in.a }
       if (d>0) { in .d <> Queue(out.d, d, pipe && d<2) } else { in.d <> out.d }
 
-      if (edgeOut.manager.anySupportAcquire && edgeOut.client.anySupportProbe) {
+      if (edgeOut.manager.anySupportAcquireB && edgeOut.client.anySupportProbe) {
         if (b>0) { in .b <> Queue(out.b, b, pipe && b<2) } else { in.b <> out.b }
         if (c>0) { out.c <> Queue(in .c, c, pipe && c<2) } else { out.c <> in.c }
         if (e>0) { out.e <> Queue(in .e, e, pipe && e<2) } else { out.e <> in.e }

src/main/scala/uncore/tilelink2/CacheCork.scala

@@ -0,0 +1,128 @@
+// See LICENSE.SiFive for license details.
+
+package uncore.tilelink2
+
+import Chisel._
+import chisel3.internal.sourceinfo.SourceInfo
+import config._
+import diplomacy._
+import scala.math.{min,max}
+import TLMessages._
+
+class TLCacheCork(unsafe: Boolean = false)(implicit p: Parameters) extends LazyModule
+{
+  val node = TLAdapterNode(
+    clientFn  = { case Seq(cp) =>
+      cp.copy(clients = cp.clients.map { c => c.copy(
+        sourceId = IdRange(c.sourceId.start*2, c.sourceId.end*2))})},
+    managerFn = { case Seq(mp) =>
+      mp.copy(managers = mp.managers.map { m => m.copy(
+        regionType         = if (m.regionType == RegionType.UNCACHED) RegionType.TRACKED else m.regionType,
+        supportsAcquireB   = m.supportsGet,
+        supportsAcquireT   = m.supportsPutFull)})})
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val in  = node.bundleIn
+      val out = node.bundleOut
+    }
+
+    val edgeIn = node.edgesIn(0)
+    val edgeOut = node.edgesOut(0)
+
+    require (edgeIn.client.clients.size == 1 || unsafe, "Only one client can safely use a TLCacheCork")
+    require (edgeIn.client.clients.filter(_.supportsProbe).size == 1, "Only one caching client allowed")
+    edgeOut.manager.managers.foreach { case m =>
+      require (!m.supportsAcquireB, "Cannot support caches beyond the Cork")
+    }
+
+    val out = io.out(0)
+    val in = io.in(0)
+
+    // The Cork turns [Acquire=>Get] => [AccessAckData=>GrantData]
+    //            and [ReleaseData=>PutFullData] => [AccessAck=>ReleaseAck]
+    // We need to encode information sufficient to reverse the transformation in output.
+    // A caveat is that we get Acquire+Release with the same source and must keep the
+    // source unique after transformation onto the A channel.
+    // The coding scheme is:
+    //   Put: 1, Release: 0 => AccessAck
+    //   *: 0, Acquire: 1 => AccessAckData
+
+    // Take requests from A to A
+    val isPut = in.a.bits.opcode === PutFullData || in.a.bits.opcode === PutPartialData
+    val a_a = Wire(out.a)
+    a_a <> in.a
+    a_a.bits.source := in.a.bits.source << 1 | Mux(isPut, UInt(1), UInt(0))
+
+    // Transform Acquire into Get
+    when (in.a.bits.opcode === Acquire) {
+      a_a.bits.opcode := Get
+      a_a.bits.param  := UInt(0)
+      a_a.bits.source := in.a.bits.source << 1 | UInt(1)
+    }
+
+    // Take ReleaseData from C to A; Release from C to D
+    val c_a = Wire(out.a)
+    c_a.valid := in.c.valid && in.c.bits.opcode === ReleaseData
+    c_a.bits.opcode  := PutFullData
+    c_a.bits.param   := UInt(0)
+    c_a.bits.size    := in.c.bits.size
+    c_a.bits.source  := in.c.bits.source << 1
+    c_a.bits.address := in.c.bits.address
+    c_a.bits.mask    := edgeOut.mask(in.c.bits.address, in.c.bits.size)
+    c_a.bits.data    := in.c.bits.data
+
+    val c_d = Wire(in.d)
+    c_d.valid := in.c.valid && in.c.bits.opcode === Release
+    c_d.bits.opcode  := ReleaseAck
+    c_d.bits.param   := UInt(0)
+    c_d.bits.size    := in.c.bits.size
+    c_d.bits.source  := in.c.bits.source
+    c_d.bits.sink    := UInt(0)
+    c_d.bits.addr_lo := in.c.bits.address
+    c_d.bits.data    := UInt(0)
+    c_d.bits.error   := Bool(false)
+
+    assert (!in.c.valid || in.c.bits.opcode === Release || in.c.bits.opcode === ReleaseData)
+    in.c.ready := Mux(in.c.bits.opcode === Release, c_d.ready, c_a.ready)
+
+    // Discard E
+    in.e.ready := Bool(true)
+
+    // Block B; should never happen
+    out.b.ready := Bool(false)
+    assert (!out.b.valid)
+
+    // Take responses from D and transform them
+    val d_d = Wire(in.d)
+    d_d <> out.d
+    d_d.bits.source := out.d.bits.source >> 1
+
+    when (out.d.bits.opcode === AccessAckData && out.d.bits.source(0)) {
+      d_d.bits.opcode := GrantData
+      d_d.bits.param  := TLPermissions.toT
+    }
+    when (out.d.bits.opcode === AccessAck && !out.d.bits.source(0)) {
+      d_d.bits.opcode := ReleaseAck
+    }
+
+    // Combine the sources of messages into the channels
+    TLArbiter(TLArbiter.lowestIndexFirst)(out.a, (edgeOut.numBeats1(c_a.bits), c_a), (edgeOut.numBeats1(a_a.bits), a_a))
+    TLArbiter(TLArbiter.lowestIndexFirst)(in.d,  (edgeIn .numBeats1(d_d.bits), d_d), (UInt(0), Queue(c_d, 2)))
+
+    // Tie off unused ports
+    in.b.valid := Bool(false)
+    out.c.valid := Bool(false)
+    out.e.valid := Bool(false)
+  }
+}
+
+object TLCacheCork
+{
+  // applied to the TL source node; y.node := TLCacheCork()(x.node)
+  def apply(unsafe: Boolean = false)(x: TLOutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): TLOutwardNode = {
+    val cork = LazyModule(new TLCacheCork(unsafe))
+    cork.node := x
+    cork.node
+  }
+}

src/main/scala/uncore/tilelink2/Crossing.scala

@@ -20,7 +20,7 @@ class TLAsyncCrossingSource(sync: Int = 3)(implicit p: Parameters) extends LazyM
 
     ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
       val sink_reset_n = out.a.sink_reset_n
-      val bce = edgeIn.manager.anySupportAcquire && edgeIn.client.anySupportProbe
+      val bce = edgeIn.manager.anySupportAcquireB && edgeIn.client.anySupportProbe
       val depth = edgeOut.manager.depth
 
       out.a <> ToAsyncBundle(in.a, depth, sync)
@@ -54,7 +54,7 @@ class TLAsyncCrossingSink(depth: Int = 8, sync: Int = 3)(implicit p: Parameters)
 
     ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
       val source_reset_n = in.a.source_reset_n
-      val bce = edgeOut.manager.anySupportAcquire && edgeOut.client.anySupportProbe
+      val bce = edgeOut.manager.anySupportAcquireB && edgeOut.client.anySupportProbe
 
       out.a <> FromAsyncBundle(in.a, sync)
       in.d <> ToAsyncBundle(out.d, depth, sync)

src/main/scala/uncore/tilelink2/Edges.scala

@@ -27,7 +27,7 @@ class TLEdge(
         // Do there exist A messages with Data?
         val aDataYes = manager.anySupportArithmetic || manager.anySupportLogical || manager.anySupportPutFull || manager.anySupportPutPartial
         // Do there exist A messages without Data?
-        val aDataNo  = manager.anySupportAcquire || manager.anySupportGet || manager.anySupportHint
+        val aDataNo  = manager.anySupportAcquireB || manager.anySupportGet || manager.anySupportHint
         // Statically optimize the case where hasData is a constant
         if (!aDataYes) Some(false) else if (!aDataNo) Some(true) else None
       }
@@ -48,9 +48,9 @@ class TLEdge(
       }
       case _:TLBundleD => {
         // Do there eixst D messages with Data?
-        val dDataYes = manager.anySupportGet || manager.anySupportArithmetic || manager.anySupportLogical || manager.anySupportAcquire
+        val dDataYes = manager.anySupportGet || manager.anySupportArithmetic || manager.anySupportLogical || manager.anySupportAcquireB
         // Do there exist D messages without Data?
-        val dDataNo  = manager.anySupportPutFull || manager.anySupportPutPartial || manager.anySupportHint || manager.anySupportAcquire
+        val dDataNo  = manager.anySupportPutFull || manager.anySupportPutPartial || manager.anySupportHint || manager.anySupportAcquireT
         if (!dDataYes) Some(false) else if (!dDataNo) Some(true) else None
       }
       case _:TLBundleE => Some(false)
@@ -244,8 +244,8 @@ class TLEdgeOut(
 {
   // Transfers
   def Acquire(fromSource: UInt, toAddress: UInt, lgSize: UInt, growPermissions: UInt) = {
-    require (manager.anySupportAcquire)
+    require (manager.anySupportAcquireB)
-    val legal = manager.supportsAcquireFast(toAddress, lgSize)
+    val legal = manager.supportsAcquireBFast(toAddress, lgSize)
     val a = Wire(new TLBundleA(bundle))
     a.opcode  := TLMessages.Acquire
     a.param   := growPermissions
@@ -258,8 +258,8 @@ class TLEdgeOut(
   }
 
   def Release(fromSource: UInt, toAddress: UInt, lgSize: UInt, shrinkPermissions: UInt) = {
-    require (manager.anySupportAcquire)
+    require (manager.anySupportAcquireB)
-    val legal = manager.supportsAcquireFast(toAddress, lgSize)
+    val legal = manager.supportsAcquireBFast(toAddress, lgSize)
     val c = Wire(new TLBundleC(bundle))
     c.opcode  := TLMessages.Release
     c.param   := shrinkPermissions
@@ -272,8 +272,8 @@ class TLEdgeOut(
   }
 
   def Release(fromSource: UInt, toAddress: UInt, lgSize: UInt, shrinkPermissions: UInt, data: UInt) = {
-    require (manager.anySupportAcquire)
+    require (manager.anySupportAcquireB)
-    val legal = manager.supportsAcquireFast(toAddress, lgSize)
+    val legal = manager.supportsAcquireBFast(toAddress, lgSize)
     val c = Wire(new TLBundleC(bundle))
     c.opcode  := TLMessages.ReleaseData
     c.param   := shrinkPermissions

src/main/scala/uncore/tilelink2/Filter.scala

@@ -22,7 +22,8 @@ class TLFilter(select: AddressSet)(implicit p: Parameters) extends LazyModule
         if (filtered.isEmpty) { None } else {
           Some(m.copy(
             address            = filtered,
-            supportsAcquire    = m.supportsAcquire   .intersect(cap),
+            supportsAcquireT   = m.supportsAcquireT  .intersect(cap),
+            supportsAcquireB   = m.supportsAcquireB  .intersect(cap),
             supportsArithmetic = m.supportsArithmetic.intersect(cap),
             supportsLogical    = m.supportsLogical   .intersect(cap),
             supportsGet        = m.supportsGet       .intersect(cap),

src/main/scala/uncore/tilelink2/Fragmenter.scala

@@ -62,7 +62,7 @@ class TLFragmenter(val minSize: Int, val maxSize: Int, val alwaysMin: Boolean =
     // We don't support fragmenting to sub-beat accesses
     require (minSize >= beatBytes)
     // We can't support devices which are cached on both sides of us
-    require (!edgeOut.manager.anySupportAcquire || !edgeIn.client.anySupportProbe)
+    require (!edgeOut.manager.anySupportAcquireB || !edgeIn.client.anySupportProbe)
 
     /* The Fragmenter is a bit tricky, because there are 5 sizes in play:
      *   max  size -- the maximum transfer size possible

src/main/scala/uncore/tilelink2/HintHandler.scala

@@ -27,7 +27,7 @@ class TLHintHandler(supportManagers: Boolean = true, supportClients: Boolean = f
     val edgeOut = node.edgesOut(0)
 
     // Don't add support for clients if there is no BCE channel
-    val bce = edgeOut.manager.anySupportAcquire && edgeIn.client.anySupportProbe
+    val bce = edgeOut.manager.anySupportAcquireB && edgeIn.client.anySupportProbe
     require (!supportClients || bce)
 
     // Does it even make sense to add the HintHandler?

src/main/scala/uncore/tilelink2/IntNodes.scala

@@ -95,11 +95,11 @@ case class IntAdapterNode(
 case class IntOutputNode() extends OutputNode(IntImp)
 case class IntInputNode() extends InputNode(IntImp)
 
-case class IntBlindOutputNode() extends BlindOutputNode(IntImp)(IntSinkPortParameters(Seq(IntSinkParameters())))
+case class IntBlindOutputNode() extends BlindOutputNode(IntImp)(Seq(IntSinkPortParameters(Seq(IntSinkParameters()))))
-case class IntBlindInputNode(num: Int) extends BlindInputNode(IntImp)(IntSourcePortParameters(Seq(IntSourceParameters(num))))
+case class IntBlindInputNode(num: Int) extends BlindInputNode(IntImp)(Seq(IntSourcePortParameters(Seq(IntSourceParameters(num)))))
 
-case class IntInternalOutputNode() extends InternalOutputNode(IntImp)(IntSinkPortParameters(Seq(IntSinkParameters())))
+case class IntInternalOutputNode() extends InternalOutputNode(IntImp)(Seq(IntSinkPortParameters(Seq(IntSinkParameters()))))
-case class IntInternalInputNode(num: Int) extends InternalInputNode(IntImp)(IntSourcePortParameters(Seq(IntSourceParameters(num))))
+case class IntInternalInputNode(num: Int) extends InternalInputNode(IntImp)(Seq(IntSourcePortParameters(Seq(IntSourceParameters(num)))))
 
 class IntXbar()(implicit p: Parameters) extends LazyModule
 {

src/main/scala/uncore/tilelink2/Isolation.scala

@@ -57,7 +57,7 @@ class TLIsolation(fOut: (Bool, UInt) => UInt, fIn: (Bool, UInt) => UInt)(implici
       ABo(out.a, in .a)
       ABi(in .d, out.d)
 
-      if (edgeOut.manager.base.anySupportAcquire && edgeOut.client.base.anySupportProbe) {
+      if (edgeOut.manager.base.anySupportAcquireB && edgeOut.client.base.anySupportProbe) {
         ABi(in .b, out.b)
         ABo(out.c, in .c)
         ABo(out.e, in .e)

src/main/scala/uncore/tilelink2/Monitor.scala

@@ -42,7 +42,7 @@ class TLMonitor(args: TLMonitorArgs) extends TLMonitorBase(args)
     val mask = edge.full_mask(bundle)
 
     when (bundle.opcode === TLMessages.Acquire) {
-      assert (edge.manager.supportsAcquireSafe(edge.address(bundle), bundle.size), "'A' channel carries Acquire type unsupported by manager" + extra)
+      assert (edge.manager.supportsAcquireBSafe(edge.address(bundle), bundle.size), "'A' channel carries Acquire type unsupported by manager" + extra)
       assert (source_ok, "'A' channel Acquire carries invalid source ID" + extra)
       assert (bundle.size >= UInt(log2Ceil(edge.manager.beatBytes)), "'A' channel Acquire smaller than a beat" + extra)
       assert (is_aligned, "'A' channel Acquire address not aligned to size" + extra)
@@ -189,7 +189,7 @@ class TLMonitor(args: TLMonitorArgs) extends TLMonitorBase(args)
     }
 
     when (bundle.opcode === TLMessages.Release) {
-      assert (edge.manager.supportsAcquireSafe(edge.address(bundle), bundle.size), "'C' channel carries Release type unsupported by manager" + extra)
+      assert (edge.manager.supportsAcquireBSafe(edge.address(bundle), bundle.size), "'C' channel carries Release type unsupported by manager" + extra)
       assert (source_ok, "'C' channel Release carries invalid source ID" + extra)
       assert (bundle.size >= UInt(log2Ceil(edge.manager.beatBytes)), "'C' channel Release smaller than a beat" + extra)
       assert (is_aligned, "'C' channel Release address not aligned to size" + extra)
@@ -198,7 +198,7 @@ class TLMonitor(args: TLMonitorArgs) extends TLMonitorBase(args)
     }
 
     when (bundle.opcode === TLMessages.ReleaseData) {
-      assert (edge.manager.supportsAcquireSafe(edge.address(bundle), bundle.size), "'C' channel carries ReleaseData type unsupported by manager" + extra)
+      assert (edge.manager.supportsAcquireBSafe(edge.address(bundle), bundle.size), "'C' channel carries ReleaseData type unsupported by manager" + extra)
       assert (source_ok, "'C' channel ReleaseData carries invalid source ID" + extra)
       assert (bundle.size >= UInt(log2Ceil(edge.manager.beatBytes)), "'C' channel ReleaseData smaller than a beat" + extra)
       assert (is_aligned, "'C' channel ReleaseData address not aligned to size" + extra)
@@ -409,7 +409,7 @@ class TLMonitor(args: TLMonitorArgs) extends TLMonitorBase(args)
     val d_last = edge.last(bundle.d.bits, bundle.d.fire())
 
     if (edge.manager.minLatency > 0) {
-      assert(bundle.a.bits.source =/= bundle.d.bits.source || !bundle.a.valid || !bundle.d.valid, s"'A' and 'D' concurrent, despite minlatency ${edge.manager.minLatency}" + extra)
+      assert(bundle.d.bits.opcode === TLMessages.ReleaseAck || bundle.a.bits.source =/= bundle.d.bits.source || !bundle.a.valid || !bundle.d.valid, s"'A' and 'D' concurrent, despite minlatency ${edge.manager.minLatency}" + extra)
     }
 
     val a_set = Wire(init = UInt(0, width = edge.client.endSourceId))

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

@@ -16,14 +16,9 @@ object TLImp extends NodeImp[TLClientPortParameters, TLManagerPortParameters, TL
 {
   def edgeO(pd: TLClientPortParameters, pu: TLManagerPortParameters): TLEdgeOut = new TLEdgeOut(pd, pu)
   def edgeI(pd: TLClientPortParameters, pu: TLManagerPortParameters): TLEdgeIn  = new TLEdgeIn(pd, pu)
-  def bundleO(eo: Seq[TLEdgeOut]): Vec[TLBundle] = {
+
-    require (!eo.isEmpty)
+  def bundleO(eo: Seq[TLEdgeOut]): Vec[TLBundle] = Vec(eo.size, TLBundle(TLBundleParameters.union(eo.map(_.bundle))))
-    Vec(eo.size, TLBundle(eo.map(_.bundle).reduce(_.union(_))))
+  def bundleI(ei: Seq[TLEdgeIn]):  Vec[TLBundle] = Vec(ei.size, TLBundle(TLBundleParameters.union(ei.map(_.bundle))))
-  }
-  def bundleI(ei: Seq[TLEdgeIn]): Vec[TLBundle] = {
-    require (!ei.isEmpty)
-    Vec(ei.size, TLBundle(ei.map(_.bundle).reduce(_.union(_))))
-  }
 
   def colour = "#000000" // black
   override def labelI(ei: TLEdgeIn)  = (ei.manager.beatBytes * 8).toString
@@ -120,11 +115,11 @@ 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 TLBlindOutputNode(portParams: Seq[TLManagerPortParameters]) extends BlindOutputNode(TLImp)(portParams)
-case class TLBlindInputNode(portParams: TLClientPortParameters) extends BlindInputNode(TLImp)(portParams)
+case class TLBlindInputNode(portParams: Seq[TLClientPortParameters]) extends BlindInputNode(TLImp)(portParams)
 
-case class TLInternalOutputNode(portParams: TLManagerPortParameters) extends InternalOutputNode(TLImp)(portParams)
+case class TLInternalOutputNode(portParams: Seq[TLManagerPortParameters]) extends InternalOutputNode(TLImp)(portParams)
-case class TLInternalInputNode(portParams: TLClientPortParameters) extends InternalInputNode(TLImp)(portParams)
+case class TLInternalInputNode(portParams: Seq[TLClientPortParameters]) extends InternalInputNode(TLImp)(portParams)
 
 /** Synthesizeable unit tests */
 import unittest._
@@ -152,14 +147,9 @@ object TLAsyncImp extends NodeImp[TLAsyncClientPortParameters, TLAsyncManagerPor
 {
   def edgeO(pd: TLAsyncClientPortParameters, pu: TLAsyncManagerPortParameters): TLAsyncEdgeParameters = TLAsyncEdgeParameters(pd, pu)
   def edgeI(pd: TLAsyncClientPortParameters, pu: TLAsyncManagerPortParameters): TLAsyncEdgeParameters = TLAsyncEdgeParameters(pd, pu)
-  def bundleO(eo: Seq[TLAsyncEdgeParameters]): Vec[TLAsyncBundle] = {
+
-    require (eo.size == 1)
+  def bundleO(eo: Seq[TLAsyncEdgeParameters]): Vec[TLAsyncBundle] = Vec(eo.size, new TLAsyncBundle(TLAsyncBundleParameters.union(eo.map(_.bundle))))
-    Vec(eo.size, new TLAsyncBundle(eo(0).bundle))
+  def bundleI(ei: Seq[TLAsyncEdgeParameters]): Vec[TLAsyncBundle] = Vec(ei.size, new TLAsyncBundle(TLAsyncBundleParameters.union(ei.map(_.bundle))))
-  }
-  def bundleI(ei: Seq[TLAsyncEdgeParameters]): Vec[TLAsyncBundle] = {
-    require (ei.size == 1)
-    Vec(ei.size, new TLAsyncBundle(ei(0).bundle))
-  }
 
   def colour = "#ff0000" // red
   override def labelI(ei: TLAsyncEdgeParameters) = ei.manager.depth.toString

src/main/scala/uncore/tilelink2/Parameters.scala

@@ -12,7 +12,8 @@ case class TLManagerParameters(
   executable:         Boolean       = false, // processor can execute from this memory
   nodePath:           Seq[BaseNode] = Seq(),
   // Supports both Acquire+Release+Finish of these sizes
-  supportsAcquire:    TransferSizes = TransferSizes.none,
+  supportsAcquireT:   TransferSizes = TransferSizes.none,
+  supportsAcquireB:   TransferSizes = TransferSizes.none,
   supportsArithmetic: TransferSizes = TransferSizes.none,
   supportsLogical:    TransferSizes = TransferSizes.none,
   supportsGet:        TransferSizes = TransferSizes.none,
@@ -28,10 +29,20 @@ case class TLManagerParameters(
 
   address.combinations(2).foreach { case Seq(x,y) => require (!x.overlaps(y)) }
   require (supportsPutFull.contains(supportsPutPartial))
+  require (supportsPutFull.contains(supportsArithmetic))
+  require (supportsPutFull.contains(supportsLogical))
+  require (supportsGet.contains(supportsArithmetic))
+  require (supportsGet.contains(supportsLogical))
+  require (supportsAcquireB.contains(supportsAcquireT))
+
+  // Make sure that the regionType agrees with the capabilities
+  require ((regionType == RegionType.CACHED || regionType == RegionType.TRACKED) != supportsAcquireB.none)
+  require (regionType != RegionType.UNCACHED || supportsGet)
 
   // Largest support transfer of all types
   val maxTransfer = List(
-    supportsAcquire.max,
+    supportsAcquireT.max,
+    supportsAcquireB.max,
     supportsArithmetic.max,
     supportsLogical.max,
     supportsGet.max,
@@ -73,7 +84,8 @@ case class TLManagerPortParameters(
   def maxTransfer = managers.map(_.maxTransfer).max
   
   // Operation sizes supported by all outward Managers
-  val allSupportAcquire    = managers.map(_.supportsAcquire)   .reduce(_ intersect _)
+  val allSupportAcquireT   = managers.map(_.supportsAcquireT)  .reduce(_ intersect _)
+  val allSupportAcquireB   = managers.map(_.supportsAcquireB)  .reduce(_ intersect _)
   val allSupportArithmetic = managers.map(_.supportsArithmetic).reduce(_ intersect _)
   val allSupportLogical    = managers.map(_.supportsLogical)   .reduce(_ intersect _)
   val allSupportGet        = managers.map(_.supportsGet)       .reduce(_ intersect _)
@@ -82,7 +94,8 @@ case class TLManagerPortParameters(
   val allSupportHint       = managers.map(_.supportsHint)      .reduce(_ intersect _)
 
   // Operation supported by at least one outward Managers
-  val anySupportAcquire    = managers.map(!_.supportsAcquire.none)   .reduce(_ || _)
+  val anySupportAcquireT   = managers.map(!_.supportsAcquireT.none)  .reduce(_ || _)
+  val anySupportAcquireB   = managers.map(!_.supportsAcquireB.none)  .reduce(_ || _)
   val anySupportArithmetic = managers.map(!_.supportsArithmetic.none).reduce(_ || _)
   val anySupportLogical    = managers.map(!_.supportsLogical.none)   .reduce(_ || _)
   val anySupportGet        = managers.map(!_.supportsGet.none)       .reduce(_ || _)
@@ -122,7 +135,8 @@ case class TLManagerPortParameters(
   }
 
   // Check for support of a given operation at a specific address
-  val supportsAcquireSafe    = safe_helper(_.supportsAcquire) _
+  val supportsAcquireTSafe   = safe_helper(_.supportsAcquireT) _
+  val supportsAcquireBSafe   = safe_helper(_.supportsAcquireB) _
   val supportsArithmeticSafe = safe_helper(_.supportsArithmetic) _
   val supportsLogicalSafe    = safe_helper(_.supportsLogical) _
   val supportsGetSafe        = safe_helper(_.supportsGet) _
@@ -130,7 +144,8 @@ case class TLManagerPortParameters(
   val supportsPutPartialSafe = safe_helper(_.supportsPutPartial) _
   val supportsHintSafe       = safe_helper(_.supportsHint) _
 
-  val supportsAcquireFast    = fast_helper(_.supportsAcquire) _
+  val supportsAcquireTFast   = fast_helper(_.supportsAcquireT) _
+  val supportsAcquireBFast   = fast_helper(_.supportsAcquireB) _
   val supportsArithmeticFast = fast_helper(_.supportsArithmetic) _
   val supportsLogicalFast    = fast_helper(_.supportsLogical) _
   val supportsGetFast        = fast_helper(_.supportsGet) _
@@ -258,6 +273,15 @@ case class TLBundleParameters(
 
 object TLBundleParameters
 {
+  val emptyBundleParams = TLBundleParameters(
+    addressBits = 1,
+    dataBits    = 8,
+    sourceBits  = 1,
+    sinkBits    = 1,
+    sizeBits    = 1)
+
+  def union(x: Seq[TLBundleParameters]) = x.foldLeft(emptyBundleParams)((x,y) => x.union(y))
+
   def apply(client: TLClientPortParameters, manager: TLManagerPortParameters) =
     new TLBundleParameters(
       addressBits = log2Up(manager.maxAddress + 1),
@@ -282,7 +306,21 @@ case class TLEdgeParameters(
 
 case class TLAsyncManagerPortParameters(depth: Int, base: TLManagerPortParameters) { require (isPow2(depth)) }
 case class TLAsyncClientPortParameters(base: TLClientPortParameters)
-case class TLAsyncBundleParameters(depth: Int, base: TLBundleParameters) { require (isPow2(depth)) }
+
+case class TLAsyncBundleParameters(depth: Int, base: TLBundleParameters)
+{
+  require (isPow2(depth))
+  def union(x: TLAsyncBundleParameters) = TLAsyncBundleParameters(
+    depth = max(depth, x.depth),
+    base  = base.union(x.base))
+}
+
+object TLAsyncBundleParameters
+{
+  val emptyBundleParams = TLAsyncBundleParameters(depth = 1, base = TLBundleParameters.emptyBundleParams)
+  def union(x: Seq[TLAsyncBundleParameters]) = x.foldLeft(emptyBundleParams)((x,y) => x.union(y))
+}
+
 case class TLAsyncEdgeParameters(client: TLAsyncClientPortParameters, manager: TLAsyncManagerPortParameters)
 {
   val bundle = TLAsyncBundleParameters(manager.depth, TLBundleParameters(client.base, manager.base))
@@ -296,7 +334,8 @@ object ManagerUnification
       regionType:         RegionType.T,
       executable:         Boolean,
       lastNode:           BaseNode,
-      supportsAcquire:    TransferSizes,
+      supportsAcquireT:   TransferSizes,
+      supportsAcquireB:   TransferSizes,
       supportsArithmetic: TransferSizes,
       supportsLogical:    TransferSizes,
       supportsGet:        TransferSizes,
@@ -307,7 +346,8 @@ object ManagerUnification
       regionType         = x.regionType,
       executable         = x.executable,
       lastNode           = x.nodePath.last,
-      supportsAcquire    = x.supportsAcquire,
+      supportsAcquireT   = x.supportsAcquireT,
+      supportsAcquireB   = x.supportsAcquireB,
       supportsArithmetic = x.supportsArithmetic,
       supportsLogical    = x.supportsLogical,
       supportsGet        = x.supportsGet,

src/main/scala/uncore/tilelink2/SourceShrinker.scala

@@ -31,7 +31,7 @@ class TLSourceShrinker(maxInFlight: Int)(implicit p: Parameters) extends LazyMod
 
     // Acquires cannot pass this adapter; it makes Probes impossible
     require (!edgeIn.client.anySupportProbe || 
-             !edgeOut.manager.anySupportAcquire)
+             !edgeOut.manager.anySupportAcquireB)
 
     out.b.ready := Bool(true)
     out.c.valid := Bool(false)

src/main/scala/uncore/tilelink2/ToAXI4.scala

@@ -84,9 +84,9 @@ class TLToAXI4(idBits: Int, combinational: Boolean = true)(implicit p: Parameter
     val a_last    = edgeIn.last(in.a)
 
     // Make sure the fields are within the bounds we assumed
-    assert (a_source  < UInt(1 << sourceBits))
+    assert (a_source  < UInt(BigInt(1) << sourceBits))
-    assert (a_size    < UInt(1 << sizeBits))
+    assert (a_size    < UInt(BigInt(1) << sizeBits))
-    assert (a_addr_lo < UInt(1 << addrBits))
+    assert (a_addr_lo < UInt(BigInt(1) << addrBits))
 
     // Carefully pack/unpack fields into the state we send
     val baseEnd = 0

src/main/scala/uncore/tilelink2/WidthWidget.scala

@@ -147,7 +147,7 @@ class TLWidthWidget(innerBeatBytes: Int)(implicit p: Parameters) extends LazyMod
     splice(edgeIn,  in.a,  edgeOut, out.a)
     splice(edgeOut, out.d, edgeIn,  in.d)
 
-    if (edgeOut.manager.anySupportAcquire && edgeIn.client.anySupportProbe) {
+    if (edgeOut.manager.anySupportAcquireB && edgeIn.client.anySupportProbe) {
       splice(edgeOut, out.b, edgeIn,  in.b)
       splice(edgeIn,  in.c,  edgeOut, out.c)
       in.e.ready := out.e.ready

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

@@ -164,7 +164,7 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.lowestIndexFirst)(implicit p:
     for (o <- 0 until out.size) {
       val allowI = Seq.tabulate(in.size) { i =>
         node.edgesIn(i).client.anySupportProbe &&
-        node.edgesOut(o).manager.anySupportAcquire
+        node.edgesOut(o).manager.anySupportAcquireB
       }
       TLArbiter(policy)(out(o).a,       (beatsAI zip portsAOI(o)        ):_*)
       TLArbiter(policy)(out(o).c, filter(beatsCI zip portsCOI(o), allowI):_*)
@@ -174,7 +174,7 @@ class TLXbar(policy: TLArbiter.Policy = TLArbiter.lowestIndexFirst)(implicit p:
     for (i <- 0 until in.size) {
       val allowO = Seq.tabulate(out.size) { o =>
         node.edgesIn(i).client.anySupportProbe &&
-        node.edgesOut(o).manager.anySupportAcquire
+        node.edgesOut(o).manager.anySupportAcquireB
       }
       TLArbiter(policy)(in(i).b, filter(beatsBO zip portsBIO(i), allowO):_*)
       TLArbiter(policy)(in(i).d,       (beatsDO zip portsDIO(i)        ):_*)