Heterogeneous Tiles (#550)

Fundamental new features:

* Added tile package: This package is intended to hold components re-usable across different types of tile. Will be the future location of TL2-RoCC accelerators and new diplomatic versions of intra-tile interfaces.
* Adopted [ModuleName]Params convention: Code base was very inconsistent about what to name case classes that provide parameters to modules. Settled on calling them [ModuleName]Params to distinguish them from config.Parameters and config.Config. So far applied mostly only to case classes defined within rocket and tile.
* Defined RocketTileParams: A nested case class containing case classes for all the components of a tile (L1 caches and core). Allows all such parameters to vary per-tile.
* Defined RocketCoreParams: All the parameters that can be varied per-core.
* Defined L1CacheParams: A trait defining the parameters common to L1 caches, made concrete in different derived case classes.
* Defined RocketTilesKey: A sequence of RocketTileParams, one for every tile to be created.
* Provided HeterogeneousDualCoreConfig: An example of making a heterogeneous chip with two cores, one big and one little.
* Changes to legacy code: ReplacementPolicy moved to package util. L1Metadata moved to package tile. Legacy L2 cache agent removed because it can no longer share the metadata array implementation with the L1. Legacy GroundTests on life support.

Additional changes that got rolled in along the way:

* rocket: 	Fix critical path through BTB for I$ index bits > pgIdxBits
* coreplex: tiles connected via :=*
* groundtest: updated to use TileParams
* tilelink: cache cork requirements are relaxed to allow more cacheless masters

.travis.yml

@@ -1,9 +1,35 @@
 git:
   submodules: false
 language: scala
-sudo: required
+# run on new infrastructure
-services:
+sudo: false
-  - docker
+cache:
+  apt: true
+  directories:
+    $HOME/.ivy2
+    regression/install
+    emulator/verilator
+
+# packages needed to build riscv-tools
+addons:
+  apt:
+    sources:
+      - ubuntu-toolchain-r-test
+    packages:
+      - gcc-4.8
+      - g++-4.8
+      - gperf
+      - autoconf
+      - automake
+      - autotools-dev
+      - libmpc-dev
+      - libmpfr-dev
+      - libgmp-dev
+      - gawk
+      - build-essential
+      - bison
+      - flex
+      - texinfo
 
 env:
   matrix:
@@ -21,8 +47,15 @@ branches:
     - boom
     - /^hurricane.*$/
 
+install:
+  - make tools verilator -C regression SUITE=none
+
 before_install:
-  - docker pull hcook/docker-riscv:chisel
+  - export CXX=g++-4.8 CC=gcc-4.8
 
 script:
-  - docker run -v $(pwd):/opt/riscv/test hcook/docker-riscv:chisel /bin/sh -c  "make emulator-ndebug -C regression SUITE=$SUITE; make emulator-regression-tests -C regression SUITE=$SUITE"
+  - make emulator-ndebug -C regression SUITE=$SUITE
+  - travis_wait make emulator-regression-tests -C regression SUITE=$SUITE
+
+before_cache:
+  - ls -t regression/install | tail -n+2 | sed s@^@regression/install/@ | xargs rm -rf

src/main/scala/coreplex/BaseCoreplex.scala

@@ -5,9 +5,9 @@ package coreplex
 import Chisel._
 import config._
 import diplomacy._
-import rocket.{TileInterrupts, XLen}
+import tile.XLen
+import tile.TileInterrupts
 import uncore.tilelink2._
-import uncore.util._
 import util._
 
 /** Widths of various points in the SoC */
@@ -15,6 +15,9 @@ case class TLBusConfig(beatBytes: Int)
 case object CBusConfig extends Field[TLBusConfig]
 case object L1toL2Config extends Field[TLBusConfig]
 
+// These parameters apply to all caches, for now
+case object CacheBlockBytes extends Field[Int]
+
 /** L2 Broadcast Hub configuration */
 case class BroadcastConfig(
   nTrackers:  Int     = 4,
@@ -42,10 +45,11 @@ case object BootROMFile extends Field[String]
 
 trait HasCoreplexParameters {
   implicit val p: Parameters
+  lazy val tilesParams = p(RocketTilesKey)
   lazy val cbusConfig = p(CBusConfig)
   lazy val l1tol2Config = p(L1toL2Config)
-  lazy val nTiles = p(uncore.devices.NTiles)
+  lazy val nTiles = tilesParams.size
-  lazy val hasSupervisor = p(rocket.UseVM)
+  lazy val hasSupervisor = tilesParams.exists(_.core.useVM) // TODO ask andrew about this
   lazy val l2Config = p(BankedL2Config)
 }
 

src/main/scala/coreplex/BaseTile.scala

@@ -1,56 +0,0 @@
-// See LICENSE.SiFive for license details.
-
-package coreplex
-
-import Chisel._
-import config._
-import diplomacy._
-import rocket.{TileInterrupts, XLen}
-import uncore.tilelink2._
-import util.GenericParameterizedBundle
-
-abstract class BareTile(implicit p: Parameters) extends LazyModule
-
-abstract class BareTileBundle[+L <: BareTile](_outer: L) extends GenericParameterizedBundle(_outer) {
-  val outer = _outer
-  implicit val p = outer.p
-}
-
-abstract class BareTileModule[+L <: BareTile, +B <: BareTileBundle[L]](_outer: L, _io: () => B) extends LazyModuleImp(_outer) {
-  val outer = _outer
-  val io = _io ()
-}
-
-// Uses a tile-internal crossbar to provide a single TileLink master port
-trait TileNetwork {
-  implicit val p: Parameters
-  val module: TileNetworkModule
-  val l1backend = LazyModule(new TLXbar)
-  val masterNodes = List(TLOutputNode())
-  masterNodes.head := l1backend.node
-}
-
-trait TileNetworkBundle {
-  val outer: TileNetwork
-  val master = outer.masterNodes.head.bundleOut
-}
-
-trait TileNetworkModule {
-  val outer: TileNetwork
-  val io: TileNetworkBundle
-}
-
-abstract class BaseTile(implicit p: Parameters) extends BareTile
-    with TileNetwork {
-  override lazy val module = new BaseTileModule(this, () => new BaseTileBundle(this))
-}
-
-class BaseTileBundle[+L <: BaseTile](_outer: L) extends BareTileBundle(_outer)
-    with TileNetworkBundle {
-  val hartid = UInt(INPUT, p(XLen))
-  val interrupts = new TileInterrupts()(p).asInput
-  val resetVector = UInt(INPUT, p(XLen))
-}
-
-class BaseTileModule[+L <: BaseTile, +B <: BaseTileBundle[L]](_outer: L, _io: () => B) extends BareTileModule(_outer, _io)
-    with TileNetworkModule

src/main/scala/coreplex/Configs.scala

@@ -7,6 +7,7 @@ import Chisel._
 import config._
 import diplomacy._
 import rocket._
+import tile._
 import uncore.converters._
 import uncore.devices._
 import uncore.tilelink2._
@@ -14,71 +15,42 @@ import uncore.util._
 import util._
 
 class BaseCoreplexConfig extends Config ((site, here, up) => {
-  //Memory Parameters
   case PAddrBits => 32
   case PgLevels => if (site(XLen) == 64) 3 /* Sv39 */ else 2 /* Sv32 */
   case ASIdBits => 7
+  case XLen => 64 // Applies to all cores
+  case BuildCore => (p: Parameters) => new Rocket()(p)
   case RocketCrossing => Synchronous
-  //Params used by all caches
+  case RocketTilesKey =>  Nil
-  case CacheName("L1I") => CacheConfig(
-    nSets         = 64,
-    nWays         = 4,
-    rowBits       = site(L1toL2Config).beatBytes*8,
-    nTLBEntries   = 8,
-    cacheIdBits   = 0,
-    splitMetadata = false)
-  case CacheName("L1D") => CacheConfig(
-    nSets         = 64,
-    nWays         = 4,
-    rowBits       = site(L1toL2Config).beatBytes*8,
-    nTLBEntries   = 8,
-    cacheIdBits   = 0,
-    splitMetadata = false)
-  case ECCCode => None
-  case Replacer => () => new RandomReplacement(site(site(CacheName)).nWays)
-  //L1InstCache
-  case BtbKey => BtbParameters()
-  //L1DataCache
-  case DCacheKey => DCacheConfig(nMSHRs = 2)
-  case DataScratchpadSize => 0
-  //Tile Constants
-  case BuildRoCC => Nil
-  //Rocket Core Constants
-  case CoreInstBits => if (site(UseCompressed)) 16 else 32
-  case FetchWidth => if (site(UseCompressed)) 2 else 1
-  case RetireWidth => 1
-  case UseVM => true
-  case UseUser => false
-  case UseDebug => true
-  case NBreakpoints => 1
-  case NPerfCounters => 0
-  case NPerfEvents => 0
-  case FastLoadWord => true
-  case FastLoadByte => false
-  case FastJAL => false
-  case XLen => 64
-  case FPUKey => Some(FPUConfig())
-  case MulDivKey => Some(MulDivConfig(mulUnroll = 8, mulEarlyOut = (site(XLen) > 32), divEarlyOut = true))
-  case UseAtomics => true
-  case UseCompressed => true
   case DMKey => new DefaultDebugModuleConfig(site(NTiles), site(XLen))
-  case NCustomMRWCSRs => 0
+  case NTiles => site(RocketTilesKey).size
-  case MtvecInit => Some(BigInt(0))
-  case MtvecWritable => true
-  //Uncore Paramters
   case CBusConfig => TLBusConfig(beatBytes = site(XLen)/8)
   case L1toL2Config => TLBusConfig(beatBytes = site(XLen)/8) // increase for more PCIe bandwidth
   case BootROMFile => "./bootrom/bootrom.img"
-  case NTiles => site(RocketConfigs).size
-  case RocketConfigs => List(RocketConfig(site(XLen)))
-  case BuildCore => (c: RocketConfig, p: Parameters) => new Rocket(c)(p)
   case BroadcastConfig => BroadcastConfig()
   case BankedL2Config => BankedL2Config()
   case CacheBlockBytes => 64
 })
 
-class WithNCores(n: Int) extends Config((site, here, up) => {
+class WithNBigCores(n: Int) extends Config((site, here, up) => {
-  case RocketConfigs => List.fill(n){ RocketConfig(site(XLen)) }
+  case RocketTilesKey => {
+    val big = RocketTileParams(
+      core   = RocketCoreParams(mulDiv = Some(MulDivParams(mulUnroll = 8, mulEarlyOut = true, divEarlyOut = true))),
+      dcache = Some(DCacheParams(rowBits = site(L1toL2Config).beatBytes*8, nMSHRs  = 2)),
+      icache = Some(ICacheParams(rowBits = site(L1toL2Config).beatBytes*8)))
+    List.fill(n)(big) ++ up(RocketTilesKey, site)
+  }
+})
+
+class WithNSmallCores(n: Int) extends Config((site, here, up) => {
+  case RocketTilesKey => {
+    val small = RocketTileParams(
+      core = RocketCoreParams(useVM = false, fpu = None),
+      btb = None,
+      dcache = Some(DCacheParams(rowBits = site(L1toL2Config).beatBytes*8, nSets = 64, nWays = 1, nTLBEntries = 4, nMSHRs = 0)),
+      icache = Some(ICacheParams(rowBits = site(L1toL2Config).beatBytes*8, nSets = 64, nWays = 1, nTLBEntries = 4)))
+    List.fill(n)(small) ++ up(RocketTilesKey, site)
+  }
 })
 
 class WithNBanksPerMemChannel(n: Int) extends Config((site, here, up) => {
@@ -89,43 +61,50 @@ class WithNTrackersPerBank(n: Int) extends Config((site, here, up) => {
   case BroadcastConfig => up(BroadcastConfig, site).copy(nTrackers = n)
 })
 
-// This is the number of sets **per way**
+// This is the number of icache sets for all Rocket tiles
 class WithL1ICacheSets(sets: Int) extends Config((site, here, up) => {
-  case CacheName("L1I") => up(CacheName("L1I"), site).copy(nSets = sets)
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(icache = r.icache.map(_.copy(nSets = sets))) }
 })
 
-// This is the number of sets **per way**
+// This is the number of icache sets for all Rocket tiles
 class WithL1DCacheSets(sets: Int) extends Config((site, here, up) => {
-  case CacheName("L1D") => up(CacheName("L1D"), site).copy(nSets = sets)
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(dcache = r.dcache.map(_.copy(nSets = sets))) }
 })
 
 class WithL1ICacheWays(ways: Int) extends Config((site, here, up) => {
-  case CacheName("L1I") => up(CacheName("L1I"), site).copy(nWays = ways)
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(icache = r.icache.map(_.copy(nWays = ways)))
+  }
 })
 
 class WithL1DCacheWays(ways: Int) extends Config((site, here, up) => {
-  case CacheName("L1D") => up(CacheName("L1D"), site).copy(nWays = ways)
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(dcache = r.dcache.map(_.copy(nWays = ways)))
+  }
 })
 
 class WithCacheBlockBytes(linesize: Int) extends Config((site, here, up) => {
   case CacheBlockBytes => linesize
 })
 
-class WithDataScratchpad(n: Int) extends Config((site, here, up) => {
+/** Warning: applies only to the most recently added tile.
-  case DataScratchpadSize => n
+  * TODO: For now, there can only be a single scratchpad in the design
-  case CacheName("L1D") => up(CacheName("L1D"), site).copy(nSets = n / site(CacheBlockBytes))
+  *       because its address is hardcoded.
+  */
+class WithDataScratchpad(size: Int) extends Config((site, here, up) => {
+  case RocketTilesKey => {
+    val prev = up(RocketTilesKey, site)
+    prev.head.copy(
+      dcache = prev.head.dcache.map(_.copy(nSets = size / site(CacheBlockBytes))),
+      dataScratchpadBytes = size) +: prev.tail
+  }
 })
 
-// TODO: re-add L2
+class WithL2Cache extends Config(Parameters.empty) // TODO: re-add L2
-class WithL2Cache extends Config((site, here, up) => {
+class WithL2Capacity(size_kb: Int) extends Config(Parameters.empty) // TODO: re-add L2
-  case CacheName("L2") => CacheConfig(
+class WithNL2Ways(n: Int) extends Config(Parameters.empty) // TODO: re-add L2
-    nSets         = 1024,
-    nWays         = 1,
-    rowBits       = site(L1toL2Config).beatBytes*8,
-    nTLBEntries   = 0,
-    cacheIdBits   = 1,
-    splitMetadata = false)
-})
 
 class WithBufferlessBroadcastHub extends Config((site, here, up) => {
   case BroadcastConfig => up(BroadcastConfig, site).copy(bufferless = true)
@@ -151,44 +130,39 @@ class WithStatelessBridge extends Config((site, here, up) => {
     ww.node :*= cork.node
     (cork.node, ww.node)
   })
-  case DCacheKey => up(DCacheKey, site).copy(nMSHRs = 0)
-})
-
-class WithL2Capacity(size_kb: Int) extends Config(Parameters.empty) // TODO
-
-class WithNL2Ways(n: Int) extends Config((site, here, up) => {
-  case CacheName("L2") => up(CacheName("L2"), site).copy(nWays = n)
 })
 
 class WithRV32 extends Config((site, here, up) => {
   case XLen => 32
-  case FPUKey => Some(FPUConfig(divSqrt = false))
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(core = r.core.copy(
+      mulDiv = Some(MulDivParams(mulUnroll = 8)),
+      fpu = r.core.fpu.map(_.copy(divSqrt = false))))
+  }
 })
 
 class WithBlockingL1 extends Config((site, here, up) => {
-  case DCacheKey => up(DCacheKey, site).copy(nMSHRs = 0)
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(dcache = r.dcache.map(_.copy(nMSHRs = 0)))
+  }
 })
 
-class WithSmallCores extends Config((site, here, up) => {
+class WithNBreakpoints(hwbp: Int) extends Config ((site, here, up) => {
-  case MulDivKey => Some(MulDivConfig())
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
-  case FPUKey => None
+    r.copy(core = r.core.copy(nBreakpoints = hwbp))
-  case UseVM => false
+  }
-  case BtbKey => BtbParameters(nEntries = 0)
-  case CacheName("L1D") => up(CacheName("L1D"), site).copy(nSets = 64, nWays = 1, nTLBEntries = 4)
-  case CacheName("L1I") => up(CacheName("L1I"), site).copy(nSets = 64, nWays = 1, nTLBEntries = 4)
-  case DCacheKey => up(DCacheKey, site).copy(nMSHRs = 0)
 })
 
 class WithRoccExample extends Config((site, here, up) => {
   case BuildRoCC => Seq(
-    RoccParameters(
+    RoCCParams(
       opcodes = OpcodeSet.custom0,
       generator = (p: Parameters) => Module(new AccumulatorExample()(p))),
-    RoccParameters(
+    RoCCParams(
       opcodes = OpcodeSet.custom1,
       generator = (p: Parameters) => Module(new TranslatorExample()(p)),
       nPTWPorts = 1),
-    RoccParameters(
+    RoCCParams(
       opcodes = OpcodeSet.custom2,
       generator = (p: Parameters) => Module(new CharacterCountExample()(p))))
 
@@ -196,23 +170,34 @@ class WithRoccExample extends Config((site, here, up) => {
 })
 
 class WithDefaultBtb extends Config((site, here, up) => {
-  case BtbKey => BtbParameters()
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(btb = Some(BTBParams()))
+  }
 })
 
 class WithFastMulDiv extends Config((site, here, up) => {
-  case MulDivKey => Some(MulDivConfig(mulUnroll = 8, mulEarlyOut = (site(XLen) > 32), divEarlyOut = true))
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(core = r.core.copy(mulDiv = Some(
+      MulDivParams(mulUnroll = 8, mulEarlyOut = (site(XLen) > 32), divEarlyOut = true)
+  )))}
 })
 
 class WithoutMulDiv extends Config((site, here, up) => {
-  case MulDivKey => None
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(core = r.core.copy(mulDiv = None))
+  }
 })
 
 class WithoutFPU extends Config((site, here, up) => {
-  case FPUKey => None
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(core = r.core.copy(fpu = None))
+  }
 })
 
 class WithFPUWithoutDivSqrt extends Config((site, here, up) => {
-  case FPUKey => Some(FPUConfig(divSqrt = false))
+  case RocketTilesKey => up(RocketTilesKey, site) map { r =>
+    r.copy(core = r.core.copy(fpu = r.core.fpu.map(_.copy(divSqrt = false))))
+  }
 })
 
 class WithBootROMFile(bootROMFile: String) extends Config((site, here, up) => {

src/main/scala/coreplex/RISCVPlatform.scala

@@ -6,14 +6,9 @@ import Chisel._
 import config._
 import junctions._
 import diplomacy._
-import uncore.tilelink._
+import tile._
 import uncore.tilelink2._
-import uncore.coherence._
-import uncore.agents._
 import uncore.devices._
-import uncore.util._
-import uncore.converters._
-import rocket._
 import util._
 
 trait CoreplexRISCVPlatform extends CoreplexNetwork {

src/main/scala/coreplex/RocketTiles.scala

@@ -6,6 +6,7 @@ import Chisel._
 import config._
 import diplomacy._
 import rocket._
+import tile._
 import uncore.tilelink2._
 
 sealed trait ClockCrossing
@@ -13,18 +14,14 @@ case object Synchronous extends ClockCrossing
 case object Rational extends ClockCrossing
 case class Asynchronous(depth: Int, sync: Int = 2) extends ClockCrossing
 
-case object RocketConfigs extends Field[Seq[RocketConfig]]
+case object RocketTilesKey extends Field[Seq[RocketTileParams]]
 case object RocketCrossing extends Field[ClockCrossing]
 
 trait HasRocketTiles extends CoreplexRISCVPlatform {
   val module: HasRocketTilesModule
 
   private val crossing = p(RocketCrossing)
-  private val configs = p(RocketConfigs)
+  private val configs = p(RocketTilesKey)
-  private val pWithExtra = p.alterPartial {
-    case SharedMemoryTLEdge => l1tol2.node.edgesIn(0)
-    case PAddrBits => l1tol2.node.edgesIn(0).bundle.addressBits
-  }
 
   private val rocketTileIntNodes = configs.map { _ => IntInternalOutputNode() }
   rocketTileIntNodes.foreach { _ := plic.intnode }
@@ -37,11 +34,20 @@ trait HasRocketTiles extends CoreplexRISCVPlatform {
   }
 
   val rocketWires: Seq[HasRocketTilesBundle => Unit] = configs.zipWithIndex.map { case (c, i) =>
+    val pWithExtra = p.alterPartial {
+      case TileKey => c
+      case BuildRoCC => c.rocc
+      case SharedMemoryTLEdge => l1tol2.node.edgesIn(0)
+      case PAddrBits => l1tol2.node.edgesIn(0).bundle.addressBits
+    }
+
     crossing match {
       case Synchronous => {
         val tile = LazyModule(new RocketTile(c)(pWithExtra))
-        tile.masterNodes.foreach { l1tol2.node := TLBuffer()(_) }
+        val buffer = LazyModule(new TLBuffer)
-        tile.slaveNode.foreach { _ := cbus.node }
+        buffer.node :=* tile.masterNode
+        l1tol2.node :=* buffer.node
+        tile.slaveNode :*= cbus.node
         (io: HasRocketTilesBundle) => {
           // leave clock as default (simpler for hierarchical PnR)
           tile.module.io.hartid := UInt(i)
@@ -51,8 +57,12 @@ trait HasRocketTiles extends CoreplexRISCVPlatform {
       }
       case Asynchronous(depth, sync) => {
         val wrapper = LazyModule(new AsyncRocketTile(c)(pWithExtra))
-        wrapper.masterNodes.foreach { l1tol2.node := TLAsyncCrossingSink(depth, sync)(_) }
+        val sink = LazyModule(new TLAsyncCrossingSink(depth, sync))
-        wrapper.slaveNode.foreach { _ := TLAsyncCrossingSource(sync)(cbus.node) }
+        val source = LazyModule(new TLAsyncCrossingSource(sync))
+        sink.node :=* wrapper.masterNode
+        l1tol2.node :=* sink.node
+        wrapper.slaveNode :*= source.node
+        source.node :*= cbus.node
         (io: HasRocketTilesBundle) => {
           wrapper.module.clock := io.tcrs(i).clock
           wrapper.module.reset := io.tcrs(i).reset
@@ -63,8 +73,12 @@ trait HasRocketTiles extends CoreplexRISCVPlatform {
       }
       case Rational => {
         val wrapper = LazyModule(new RationalRocketTile(c)(pWithExtra))
-        wrapper.masterNodes.foreach { l1tol2.node := TLRationalCrossingSink()(_) }
+        val sink = LazyModule(new TLRationalCrossingSink)
-        wrapper.slaveNode.foreach { _ := TLRationalCrossingSource()(cbus.node) }
+        val source = LazyModule(new TLRationalCrossingSource)
+        sink.node :=* wrapper.masterNode
+        l1tol2.node :=* sink.node
+        wrapper.slaveNode :*= source.node
+        source.node :*= cbus.node
         (io: HasRocketTilesBundle) => {
           wrapper.module.clock := io.tcrs(i).clock
           wrapper.module.reset := io.tcrs(i).reset
@@ -79,7 +93,7 @@ trait HasRocketTiles extends CoreplexRISCVPlatform {
 
 trait HasRocketTilesBundle extends CoreplexRISCVPlatformBundle {
   val outer: HasRocketTiles
-  val tcrs = Vec(p(RocketConfigs).size, new Bundle {
+  val tcrs = Vec(p(RocketTilesKey).size, new Bundle {
     val clock = Clock(INPUT)
     val reset = Bool(INPUT)
   })

src/main/scala/groundtest/CacheFillTest.scala

@@ -4,6 +4,7 @@
 package groundtest
 
 import Chisel._
+import coreplex.CacheBlockBytes
 import uncore.tilelink._
 import uncore.constants._
 import uncore.util._
@@ -12,8 +13,9 @@ import config._
 
 class CacheFillTest(implicit p: Parameters) extends GroundTest()(p)
     with HasTileLinkParameters {
-  val l2Config = p(CacheName("L2"))
+  //val l2Config = p(CacheName("L2"))
-  val capacityKb = l2Config.nSets * l2Config.nWays * l2Config.rowBits / (1024*8)
+  //val capacityKb = l2Config.nSets * l2Config.nWays * l2Config.rowBits / (1024*8)
+  val capacityKb = 1024 // TODO
   val nblocks = capacityKb * 1024 / p(CacheBlockBytes)
   val s_start :: s_prefetch :: s_retrieve :: s_finished :: Nil = Enum(Bits(), 4)
   val state = Reg(init = s_start)

src/main/scala/groundtest/Configs.scala

@@ -11,6 +11,7 @@ import uncore.coherence._
 import uncore.agents._
 import uncore.util._
 import uncore.devices.NTiles
+import tile.TileKey
 import junctions._
 import config._
 import coreplex._
@@ -18,10 +19,10 @@ import rocketchip._
 
 /** Actual testing target Configs */
 
-class GroundTestConfig extends Config(new WithGroundTest ++ new BaseConfig)
+class GroundTestConfig extends Config(new WithGroundTestTiles ++ new BaseConfig)
 
 class ComparatorConfig extends Config(
-  new WithComparator ++ new GroundTestConfig)
+  new WithComparator(1) ++ new GroundTestConfig)
 class ComparatorL2Config extends Config(
   new WithAtomics ++ new WithPrefetches ++
   new WithL2Cache ++ new ComparatorConfig)
@@ -30,31 +31,32 @@ class ComparatorBufferlessConfig extends Config(
 class ComparatorStatelessConfig extends Config(
   new WithStatelessBridge ++ new ComparatorConfig)
 
-class MemtestConfig extends Config(new WithMemtest ++ new GroundTestConfig)
+class MemtestConfig extends Config(new WithMemtest(1) ++ new GroundTestConfig)
 class MemtestL2Config extends Config(
   new WithL2Cache ++ new MemtestConfig)
 class MemtestBufferlessConfig extends Config(
   new WithBufferlessBroadcastHub ++ new MemtestConfig)
 class MemtestStatelessConfig extends Config(
-  new WithNGenerators(0, 1) ++ new WithStatelessBridge ++ new MemtestConfig)
+  new WithStatelessBridge ++ new MemtestConfig)
 // Test ALL the things
 class FancyMemtestConfig extends Config(
-  new WithNGenerators(1, 2) ++ new WithNCores(2) ++ new WithMemtest ++
+  new WithMemtest(2) ++
   new WithNMemoryChannels(2) ++ new WithNBanksPerMemChannel(4) ++
   new WithL2Cache ++ new GroundTestConfig)
 
 class CacheFillTestConfig extends Config(
-  new WithNL2Ways(4) ++ new WithL2Capacity(4) ++ new WithCacheFillTest ++ new WithL2Cache ++ new GroundTestConfig)
+  new WithNL2Ways(4) ++ new WithL2Capacity(4) ++ 
+  new WithCacheFillTest(1) ++ new WithL2Cache ++ new GroundTestConfig)
 
 class BroadcastRegressionTestConfig extends Config(
-  new WithBroadcastRegressionTest ++ new GroundTestConfig)
+  new WithBroadcastRegressionTest(1) ++ new GroundTestConfig)
 class BufferlessRegressionTestConfig extends Config(
   new WithBufferlessBroadcastHub ++ new BroadcastRegressionTestConfig)
 class CacheRegressionTestConfig extends Config(
-  new WithCacheRegressionTest ++ new WithL2Cache ++ new GroundTestConfig)
+  new WithCacheRegressionTest(1) ++ new WithL2Cache ++ new GroundTestConfig)
 
 class TraceGenConfig extends Config(
-  new WithNCores(2) ++ new WithTraceGen ++ new GroundTestConfig)
+  new WithTraceGen(2) ++ new GroundTestConfig)
 class TraceGenBufferlessConfig extends Config(
   new WithBufferlessBroadcastHub ++ new TraceGenConfig)
 class TraceGenL2Config extends Config(
@@ -72,15 +74,15 @@ class Edge32BitMemtestConfig extends Config(
   new WithEdgeDataBits(32) ++ new MemtestConfig)
 
 /* Composable Configs to set individual parameters */
-class WithGroundTest extends Config((site, here, up) => {
+
-  case FPUKey => None
+class WithGroundTestTiles extends Config((site, here, up) => {
-  case UseAtomics => false
+  case TileKey => site(GroundTestKey).head
-  case UseCompressed => false
+  case NTiles => site(GroundTestKey).size
 })
 
-class WithComparator extends Config((site, here, up) => {
+class WithComparator(n: Int) extends Config((site, here, up) => {
-  case GroundTestKey => Seq.fill(site(NTiles)) {
+  case GroundTestKey => Seq.fill(n) {
-    GroundTestTileSettings(uncached = 2)
+    GroundTestTileParams(uncached = 2, dcache = None)
   }
   case BuildGroundTest =>
     (p: Parameters) => Module(new ComparatorCore()(p))
@@ -88,23 +90,21 @@ class WithComparator extends Config((site, here, up) => {
     targets    = Seq(site(ExtMem).base, testRamAddr),
     width      = 8,
     operations = 1000,
-    atomics    = site(UseAtomics),
+    atomics    = false,
     prefetches = false)
-  case FPUConfig => None
-  case UseAtomics => false
 })
 
 class WithAtomics extends Config((site, here, up) => {
-  case UseAtomics => true
+  case ComparatorKey => up(ComparatorKey, site).copy(atomics = true)
 })
 
 class WithPrefetches extends Config((site, here, up) => {
   case ComparatorKey => up(ComparatorKey, site).copy(prefetches = true)
 })
 
-class WithMemtest extends Config((site, here, up) => {
+class WithMemtest(n: Int) extends Config((site, here, up) => {
-  case GroundTestKey => Seq.fill(site(NTiles)) {
+  case GroundTestKey => Seq.fill(n) {
-    GroundTestTileSettings(1, 1)
+    GroundTestTileParams(uncached = 1)
   }
   case GeneratorKey => TrafficGeneratorParameters(
     maxRequests = 128,
@@ -113,23 +113,17 @@ class WithMemtest extends Config((site, here, up) => {
     (p: Parameters) => Module(new GeneratorTest()(p))
 })
 
-class WithNGenerators(nUncached: Int, nCached: Int) extends Config((site, here, up) => {
+class WithCacheFillTest(n: Int) extends Config((site, here, up) => {
-  case GroundTestKey => Seq.fill(site(NTiles)) {
+  case GroundTestKey => Seq.fill(n) {
-    GroundTestTileSettings(nUncached, nCached)
+    GroundTestTileParams(uncached = 1, dcache = None)
-  }
-})
-
-class WithCacheFillTest extends Config((site, here, up) => {
-  case GroundTestKey => Seq.fill(site(NTiles)) {
-    GroundTestTileSettings(uncached = 1)
   }
   case BuildGroundTest =>
     (p: Parameters) => Module(new CacheFillTest()(p))
 })
 
-class WithBroadcastRegressionTest extends Config((site, here, up) => {
+class WithBroadcastRegressionTest(n: Int) extends Config((site, here, up) => {
-  case GroundTestKey => Seq.fill(site(NTiles)) {
+  case GroundTestKey => Seq.fill(n) {
-    GroundTestTileSettings(1, 1, maxXacts = 3)
+    GroundTestTileParams(uncached = 1, maxXacts = 3)
   }
   case BuildGroundTest =>
     (p: Parameters) => Module(new RegressionTest()(p))
@@ -137,9 +131,9 @@ class WithBroadcastRegressionTest extends Config((site, here, up) => {
     (p: Parameters) => RegressionTests.broadcastRegressions(p)
 })
 
-class WithCacheRegressionTest extends Config((site, here, up) => {
+class WithCacheRegressionTest(n: Int) extends Config((site, here, up) => {
-  case GroundTestKey => Seq.fill(site(NTiles)) {
+  case GroundTestKey => Seq.fill(n) {
-    GroundTestTileSettings(1, 1, maxXacts = 5)
+    GroundTestTileParams(uncached = 1, maxXacts = 5)
   }
   case BuildGroundTest =>
     (p: Parameters) => Module(new RegressionTest()(p))
@@ -147,9 +141,9 @@ class WithCacheRegressionTest extends Config((site, here, up) => {
     (p: Parameters) => RegressionTests.cacheRegressions(p)
 })
 
-class WithTraceGen extends Config((site, here, up) => {
+class WithTraceGen(n: Int) extends Config((site, here, up) => {
-  case GroundTestKey => Seq.fill(site(NTiles)) {
+  case GroundTestKey => Seq.fill(n) {
-    GroundTestTileSettings(uncached = 0, cached = 1)
+    GroundTestTileParams(dcache = Some(DCacheParams(nSets = 16, nWays = 1)))
   }
   case BuildGroundTest =>
     (p: Parameters) => Module(new GroundTestTraceGenerator()(p))
@@ -160,11 +154,9 @@ class WithTraceGen extends Config((site, here, up) => {
     val nSets = 2
     val nWays = 1
     val blockOffset = site(CacheBlockOffsetBits)
-    val nBeats = site(TLKey("L1toL2")).dataBeats
+    val nBeats = site(CacheBlockBytes)/site(L1toL2Config).beatBytes
     List.tabulate(4 * nWays) { i =>
       Seq.tabulate(nBeats) { j => BigInt((j * 8) + ((i * nSets) << blockOffset)) }
     }.flatten
   }
-  case UseAtomics => true
-  case CacheName("L1D") => up(CacheName("L1D"), site).copy(nSets = 16, nWays = 1)
 })

src/main/scala/groundtest/Coreplex.scala

@@ -7,6 +7,7 @@ import config._
 import diplomacy._
 import coreplex._
 import rocket._
+import tile._
 import uncore.agents._
 import uncore.coherence._
 import uncore.devices._
@@ -31,9 +32,7 @@ class GroundTestCoreplex(implicit p: Parameters) extends BaseCoreplex {
           nManagers = site(BankedL2Config).nBanks + 1,
           nCachingClients = 1,
           nCachelessClients = 1,
-          maxClientXacts = ((site(DCacheKey).nMSHRs + 1) +:
+          maxClientXacts = site(GroundTestKey).map(_.maxXacts).reduce(max(_, _)),
-                             site(GroundTestKey).map(_.maxXacts))
-                               .reduce(max(_, _)),
           maxClientsPerPort = site(GroundTestKey).map(_.uncached).sum,
           maxManagerXacts = 8,
           dataBeats = (8 * site(CacheBlockBytes)) / site(XLen),
@@ -41,10 +40,7 @@ class GroundTestCoreplex(implicit p: Parameters) extends BaseCoreplex {
     }}))
   }
 
-  tiles.foreach { lm =>
+  tiles.foreach { l1tol2.node :=* _.masterNode }
-    l1tol2.node := lm.cachedOut
-    l1tol2.node := lm.uncachedOut
-  }
 
   val cbusRAM = LazyModule(new TLRAM(AddressSet(testRamAddr, 0xffff), false, cbus_beatBytes))
   cbusRAM.node := TLFragmenter(cbus_beatBytes, cbus_lineBytes)(cbus.node)
@@ -52,8 +48,7 @@ class GroundTestCoreplex(implicit p: Parameters) extends BaseCoreplex {
   override lazy val module = new GroundTestCoreplexModule(this, () => new GroundTestCoreplexBundle(this))
 }
 
-class GroundTestCoreplexBundle[+L <: GroundTestCoreplex](_outer: L) extends BaseCoreplexBundle(_outer)
+class GroundTestCoreplexBundle[+L <: GroundTestCoreplex](_outer: L) extends BaseCoreplexBundle(_outer) {
-{
   val success = Bool(OUTPUT)
 }
 

src/main/scala/groundtest/DummyPTW.scala

@@ -0,0 +1,57 @@
+// See LICENSE.SiFive for license details.
+// See LICENSE.Berkeley for license details.
+
+package groundtest
+
+import Chisel._
+import config._
+import rocket._
+import tile._
+import util.ParameterizedBundle
+
+class DummyPTW(n: Int)(implicit p: Parameters) extends CoreModule()(p) {
+  val io = new Bundle {
+    val requestors = Vec(n, new TLBPTWIO).flip
+  }
+
+  val req_arb = Module(new RRArbiter(new PTWReq, n))
+  req_arb.io.in <> io.requestors.map(_.req)
+  req_arb.io.out.ready := Bool(true)
+
+  def vpn_to_ppn(vpn: UInt): UInt = vpn(ppnBits - 1, 0)
+
+  class QueueChannel extends ParameterizedBundle()(p) {
+    val ppn = UInt(width = ppnBits)
+    val chosen = UInt(width = log2Up(n))
+  }
+
+  val s1_ppn = vpn_to_ppn(req_arb.io.out.bits.addr)
+  val s2_ppn = RegEnable(s1_ppn, req_arb.io.out.valid)
+  val s2_chosen = RegEnable(req_arb.io.chosen, req_arb.io.out.valid)
+  val s2_valid = Reg(next = req_arb.io.out.valid)
+
+  val s2_resp = Wire(new PTWResp)
+  s2_resp.pte.ppn := s2_ppn
+  s2_resp.pte.reserved_for_software := UInt(0)
+  s2_resp.pte.d := Bool(true)
+  s2_resp.pte.a := Bool(false)
+  s2_resp.pte.g := Bool(false)
+  s2_resp.pte.u := Bool(true)
+  s2_resp.pte.r := Bool(true)
+  s2_resp.pte.w := Bool(true)
+  s2_resp.pte.x := Bool(false)
+  s2_resp.pte.v := Bool(true)
+
+  io.requestors.zipWithIndex.foreach { case (requestor, i) =>
+    requestor.resp.valid := s2_valid && s2_chosen === UInt(i)
+    requestor.resp.bits := s2_resp
+    requestor.status.vm := UInt("b01000")
+    requestor.status.prv := UInt(PRV.S)
+    requestor.status.debug := Bool(false)
+    requestor.status.mprv  := Bool(true)
+    requestor.status.mpp := UInt(0)
+    requestor.ptbr.asid := UInt(0)
+    requestor.ptbr.ppn := UInt(0)
+    requestor.invalidate := Bool(false)
+  }
+}

src/main/scala/groundtest/Regression.scala

@@ -103,14 +103,13 @@ class IOGetAfterPutBlockRegression(implicit p: Parameters) extends Regression()(
  * acknowledge both of them when the first one finished.
  * This caused the state to go funky since the next time around it would
  * start the put in the middle */
-class PutBlockMergeRegression(implicit p: Parameters)
+class PutBlockMergeRegression(nSets: Int)(implicit p: Parameters)
     extends Regression()(p) with HasTileLinkParameters {
   val s_idle :: s_put :: s_wait :: s_done :: Nil = Enum(Bits(), 4)
   val state = Reg(init = s_idle)
 
   disableCache()
 
-  val nSets = p(CacheName("L2")).nSets
   val addr_blocks = Vec(Seq(0, 0, nSets).map(num => UInt(num + memStartBlock)))
   val nSteps = addr_blocks.size
   val (acq_beat, acq_done) = Counter(io.mem.acquire.fire(), tlDataBeats)
@@ -391,12 +390,9 @@ class PrefetchHitRegression(implicit p: Parameters) extends Regression()(p) {
  * set. This assumes that there is only a single cache bank. If we want to
  * test multibank configurations, we'll have to think of some other way to
  * determine which banks are conflicting */
-class WritebackRegression(implicit p: Parameters) extends Regression()(p) {
+class WritebackRegression(nSets: Int, nWays: Int)(implicit p: Parameters) extends Regression()(p) {
   disableCache()
 
-  val nSets = p(CacheName("L2")).nSets
-  val nWays = p(CacheName("L2")).nWays
-
   val addr_blocks = Vec.tabulate(nWays + 1) { i => UInt(memStartBlock + i * nSets) }
   val data = Vec.tabulate(nWays + 1) { i => UInt((i + 1) * 1423) }
 
@@ -443,11 +439,9 @@ class WritebackRegression(implicit p: Parameters) extends Regression()(p) {
   io.errored := data_mismatch
 }
 
-class ReleaseRegression(implicit p: Parameters) extends Regression()(p) {
+class ReleaseRegression(nSets: Int, nWays: Int)(implicit p: Parameters) extends Regression()(p) {
   disableMem()
 
-  val nSets = p(CacheName("L1D")).nSets
-  val nWays = p(CacheName("L1D")).nWays
   val blockOffset = p(CacheBlockOffsetBits)
 
   val startBlock = memStartBlock + 10
@@ -530,12 +524,9 @@ class PutBeforePutBlockRegression(implicit p: Parameters) extends Regression()(p
  * Make sure that multiple gets to the same line and beat are merged
  * correctly, even if it is a cache miss.
  */
-class MergedGetRegression(implicit p: Parameters) extends Regression()(p) {
+class MergedGetRegression(nSets: Int, nWays: Int)(implicit p: Parameters) extends Regression()(p) {
   disableCache()
 
-  val nSets = p(CacheName("L2")).nSets
-  val nWays = p(CacheName("L2")).nWays
-
   val (s_idle :: s_put :: s_get :: s_done :: Nil) = Enum(Bits(), 4)
   val state = Reg(init = s_idle)
 
@@ -697,23 +688,27 @@ class PutAfterReleaseRegression(implicit p: Parameters) extends Regression()(p)
 }
 
 object RegressionTests {
+  val l1sets = 16 // TODO
+  val l1ways = 1  // TODO
+  val l2sets = 32 // TODO
+  val l2ways = 2  // TODO
   def cacheRegressions(implicit p: Parameters) = Seq(
-    Module(new PutBlockMergeRegression),
+    Module(new PutBlockMergeRegression(l2sets)),
     Module(new NoAllocPutHitRegression),
     Module(new RepeatedNoAllocPutRegression),
     Module(new WriteMaskedPutBlockRegression),
     Module(new PrefetchHitRegression),
-    Module(new WritebackRegression),
+    Module(new WritebackRegression(l2sets, l2ways)),
     Module(new PutBeforePutBlockRegression),
     Module(new MixedAllocPutRegression),
-    Module(new ReleaseRegression),
+    Module(new ReleaseRegression(l1sets, l1ways)),
-    Module(new MergedGetRegression),
+    Module(new MergedGetRegression(l2sets, l2ways)),
     Module(new MergedPutRegression))
   def broadcastRegressions(implicit p: Parameters) = Seq(
     Module(new IOGetAfterPutBlockRegression),
     Module(new WriteMaskedPutBlockRegression),
     Module(new PutBeforePutBlockRegression),
-    Module(new ReleaseRegression),
+    Module(new ReleaseRegression(l1sets, l1ways)),
     Module(new PutAfterReleaseRegression))
 }
 

src/main/scala/groundtest/Tile.scala

@@ -4,23 +4,33 @@
 package groundtest
 
 import Chisel._
-import coreplex.BareTile
+import config._
+import coreplex._
 import rocket._
+import tile._
 import uncore.tilelink._
-import uncore.util.CacheName
 import uncore.tilelink2._
 import rocketchip.ExtMem
 import diplomacy._
-import scala.util.Random
-import scala.collection.mutable.ListBuffer
 import util.ParameterizedBundle
-import config._
+
+import scala.collection.mutable.ListBuffer
 
 case object BuildGroundTest extends Field[Parameters => GroundTest]
 
-case class GroundTestTileSettings(
+case class GroundTestTileParams(
-  uncached: Int = 0, cached: Int = 0, ptw: Int = 0, maxXacts: Int = 1)
+    uncached: Int = 0,
-case object GroundTestKey extends Field[Seq[GroundTestTileSettings]]
+    ptw: Int = 0,
+    maxXacts: Int = 1,
+    dcache: Option[DCacheParams] = Some(DCacheParams())) extends TileParams {
+  val icache = None
+  val btb = None
+  val rocc = Nil
+  val core = rocket.RocketCoreParams() //TODO remove this
+  val cached = if(dcache.isDefined) 1 else 0
+  val dataScratchpadBytes = 0
+}
+case object GroundTestKey extends Field[Seq[GroundTestTileParams]]
 
 trait HasGroundTestConstants {
   val timeoutCodeBits = 4
@@ -29,61 +39,14 @@ trait HasGroundTestConstants {
 
 trait HasGroundTestParameters {
   implicit val p: Parameters
-  val tileSettings = p(GroundTestKey)(p(TileId))
+  val tileParams = p(GroundTestKey)(p(TileId))
-  val nUncached = tileSettings.uncached
+  val nUncached = tileParams.uncached
-  val nCached = tileSettings.cached
+  val nCached = tileParams.cached
-  val nPTW = tileSettings.ptw
+  val nPTW = tileParams.ptw
   val memStart = p(ExtMem).base
   val memStartBlock = memStart >> p(CacheBlockOffsetBits)
 }
 
-class DummyPTW(n: Int)(implicit p: Parameters) extends CoreModule()(p) {
-  val io = new Bundle {
-    val requestors = Vec(n, new TLBPTWIO).flip
-  }
-
-  val req_arb = Module(new RRArbiter(new PTWReq, n))
-  req_arb.io.in <> io.requestors.map(_.req)
-  req_arb.io.out.ready := Bool(true)
-
-  def vpn_to_ppn(vpn: UInt): UInt = vpn(ppnBits - 1, 0)
-
-  class QueueChannel extends ParameterizedBundle()(p) {
-    val ppn = UInt(width = ppnBits)
-    val chosen = UInt(width = log2Up(n))
-  }
-
-  val s1_ppn = vpn_to_ppn(req_arb.io.out.bits.addr)
-  val s2_ppn = RegEnable(s1_ppn, req_arb.io.out.valid)
-  val s2_chosen = RegEnable(req_arb.io.chosen, req_arb.io.out.valid)
-  val s2_valid = Reg(next = req_arb.io.out.valid)
-
-  val s2_resp = Wire(new PTWResp)
-  s2_resp.pte.ppn := s2_ppn
-  s2_resp.pte.reserved_for_software := UInt(0)
-  s2_resp.pte.d := Bool(true)
-  s2_resp.pte.a := Bool(false)
-  s2_resp.pte.g := Bool(false)
-  s2_resp.pte.u := Bool(true)
-  s2_resp.pte.r := Bool(true)
-  s2_resp.pte.w := Bool(true)
-  s2_resp.pte.x := Bool(false)
-  s2_resp.pte.v := Bool(true)
-
-  io.requestors.zipWithIndex.foreach { case (requestor, i) =>
-    requestor.resp.valid := s2_valid && s2_chosen === UInt(i)
-    requestor.resp.bits := s2_resp
-    requestor.status.vm := UInt("b01000")
-    requestor.status.prv := UInt(PRV.S)
-    requestor.status.debug := Bool(false)
-    requestor.status.mprv  := Bool(true)
-    requestor.status.mpp := UInt(0)
-    requestor.ptbr.asid := UInt(0)
-    requestor.ptbr.ppn := UInt(0)
-    requestor.invalidate := Bool(false)
-  }
-}
-
 class GroundTestStatus extends Bundle with HasGroundTestConstants {
   val finished = Bool(OUTPUT)
   val timeout = Valid(UInt(width = timeoutCodeBits))
@@ -103,38 +66,33 @@ abstract class GroundTest(implicit val p: Parameters) extends Module
   val io = new GroundTestIO
 }
 
-class GroundTestTile(implicit p: Parameters) extends LazyModule with HasGroundTestParameters {
+class GroundTestTile(implicit p: Parameters) extends LazyModule
-  val dcacheParams = p.alterPartial {
+    with HasGroundTestParameters {
-    case CacheName => CacheName("L1D")
-  }
   val slave = None
-  val dcache = HellaCache(p(DCacheKey))(dcacheParams)
+  val dcacheOpt = tileParams.dcache.map { dc => HellaCache(dc.nMSHRs == 0) }
   val ucLegacy = LazyModule(new TLLegacy)
 
-   val cachedOut = TLOutputNode()
+   val masterNode = TLOutputNode()
-   val uncachedOut = TLOutputNode()
+   dcacheOpt.foreach { masterNode := _.node }
-   cachedOut := dcache.node
+   masterNode := TLHintHandler()(ucLegacy.node)
-   uncachedOut := TLHintHandler()(ucLegacy.node)
-   val masterNodes = List(cachedOut, uncachedOut)
 
   lazy val module = new LazyModuleImp(this) {
     val io = new Bundle {
-      val cached = cachedOut.bundleOut
+      val out = masterNode.bundleOut
-      val uncached = uncachedOut.bundleOut
       val success = Bool(OUTPUT)
     }
 
-    val test = p(BuildGroundTest)(dcacheParams)
+    val test = p(BuildGroundTest)(p)
 
     val ptwPorts = ListBuffer.empty ++= test.io.ptw
     val uncachedArbPorts = ListBuffer.empty ++= test.io.mem
 
-    if (nCached > 0) {
+    dcacheOpt foreach { dcache =>
-      val dcacheArb = Module(new HellaCacheArbiter(nCached)(dcacheParams))
+      val dcacheArb = Module(new HellaCacheArbiter(nCached))
 
       dcacheArb.io.requestor.zip(test.io.cache).foreach {
         case (requestor, cache) =>
-          val dcacheIF = Module(new SimpleHellaCacheIF()(dcacheParams))
+          val dcacheIF = Module(new SimpleHellaCacheIF())
           dcacheIF.io.requestor <> cache
           requestor <> dcacheIF.io.cache
       }
@@ -147,7 +105,7 @@ class GroundTestTile(implicit p: Parameters) extends LazyModule with HasGroundTe
     }
 
     if (ptwPorts.size > 0) {
-      val ptw = Module(new DummyPTW(ptwPorts.size)(dcacheParams))
+      val ptw = Module(new DummyPTW(ptwPorts.size))
       ptw.io.requestors <> ptwPorts
     }
 

src/main/scala/groundtest/TraceGen.scala

@@ -24,6 +24,7 @@ import uncore.tilelink._
 import uncore.constants._
 import uncore.devices.NTiles
 import rocket._
+import tile._
 import util.{Timer, DynamicTimer}
 import scala.util.Random
 import config._
@@ -179,7 +180,7 @@ class TagMan(val logNumTags : Int) extends Module {
 // ===============
 
 class TraceGenerator(id: Int)
-    (implicit p: Parameters) extends L1HellaCacheModule()(p)
+    (implicit val p: Parameters) extends Module
                                 with HasTraceGenParams
                                 with HasGroundTestParameters {
   val io = new Bundle {

src/main/scala/groundtest/TrafficGenerator.scala

@@ -129,7 +129,7 @@ class UncachedTileLinkGenerator(id: Int)
 }
 
 class HellaCacheGenerator(id: Int)
-    (implicit p: Parameters) extends L1HellaCacheModule()(p) with HasTrafficGeneratorParameters {
+    (implicit val p: Parameters) extends Module with HasTrafficGeneratorParameters {
   val io = new Bundle {
     val mem = new HellaCacheIO
     val status = new GroundTestStatus

src/main/scala/rocket/ALU.scala

@@ -5,6 +5,7 @@ package rocket
 
 import Chisel._
 import config._
+import tile._
 import Instructions._
 
 object ALU

src/main/scala/rocket/BTB.scala

@@ -4,23 +4,22 @@
 package rocket
 
 import Chisel._
-import config._
-import util._
 import Chisel.ImplicitConversions._
-import uncore.util.PseudoLRU
+import config._
+import tile.HasCoreParameters
+import util._
 
-case object BtbKey extends Field[BtbParameters]
+case class BTBParams(
-
-case class BtbParameters(
   nEntries: Int = 40,
   nRAS: Int = 2,
   updatesOutOfOrder: Boolean = false)
 
-abstract trait HasBtbParameters extends HasCoreParameters {
+trait HasBtbParameters extends HasCoreParameters {
-  val matchBits = pgIdxBits
+  val btbParams = tileParams.btb.getOrElse(BTBParams(nEntries = 0))
-  val entries = p(BtbKey).nEntries
+  val matchBits = pgIdxBits max log2Ceil(p(coreplex.CacheBlockBytes) * tileParams.icache.get.nSets)
-  val nRAS = p(BtbKey).nRAS
+  val entries = btbParams.nEntries
-  val updatesOutOfOrder = p(BtbKey).updatesOutOfOrder
+  val nRAS = btbParams.nRAS
+  val updatesOutOfOrder = btbParams.updatesOutOfOrder
   val nPages = ((1 max(log2Up(entries)))+1)/2*2 // control logic assumes 2 divides pages
   val opaqueBits = log2Up(entries)
   val nBHT = 1 << log2Up(entries*2)

src/main/scala/rocket/Breakpoint.scala

@@ -3,9 +3,10 @@
 package rocket
 
 import Chisel._
-import util._
 import Chisel.ImplicitConversions._
 import config._
+import tile._
+import util._
 
 class BPControl(implicit p: Parameters) extends CoreBundle()(p) {
   val ttype = UInt(width = 4)

src/main/scala/rocket/CSR.scala

@@ -6,6 +6,7 @@ package rocket
 import Chisel._
 import Instructions._
 import config._
+import tile._
 import uncore.devices._
 import util._
 import Chisel.ImplicitConversions._
@@ -60,14 +61,6 @@ class DCSR extends Bundle {
   val prv = UInt(width = PRV.SZ)
 }
 
-class TileInterrupts(implicit p: Parameters) extends CoreBundle()(p) {
-  val debug = Bool()
-  val mtip = Bool()
-  val msip = Bool()
-  val meip = Bool()
-  val seip = usingVM.option(Bool())
-}
-
 class MIP extends Bundle {
   val rocc = Bool()
   val meip = Bool()
@@ -127,7 +120,8 @@ object CSR
   val nCtr = firstHPM + nHPM
 }
 
-class CSRFileIO(implicit p: Parameters) extends CoreBundle {
+class CSRFileIO(implicit p: Parameters) extends CoreBundle
+    with HasRocketCoreParameters {
   val interrupts = new TileInterrupts().asInput
   val hartid = UInt(INPUT, xLen)
   val rw = new Bundle {
@@ -163,7 +157,7 @@ class CSRFileIO(implicit p: Parameters) extends CoreBundle {
 }
 
 class CSRFile(implicit p: Parameters) extends CoreModule()(p)
-{
+    with HasRocketCoreParameters {
   val io = new CSRFileIO
 
   val reset_mstatus = Wire(init=new MStatus().fromBits(0))
@@ -227,7 +221,7 @@ class CSRFile(implicit p: Parameters) extends CoreModule()(p)
   val reg_mbadaddr = Reg(UInt(width = vaddrBitsExtended))
   val reg_mscratch = Reg(Bits(width = xLen))
   val mtvecWidth = paddrBits min xLen
-  val reg_mtvec = p(MtvecInit) match {
+  val reg_mtvec = mtvecInit match {
     case Some(addr) => Reg(init=UInt(addr, mtvecWidth))
     case None => Reg(UInt(width = mtvecWidth))
   }
@@ -547,7 +541,7 @@ class CSRFile(implicit p: Parameters) extends CoreModule()(p)
     when (decoded_addr(CSRs.mie))      { reg_mie := wdata & supported_interrupts }
     when (decoded_addr(CSRs.mepc))     { reg_mepc := formEPC(wdata) }
     when (decoded_addr(CSRs.mscratch)) { reg_mscratch := wdata }
-    if (p(MtvecWritable))
+    if (mtvecWritable)
       when (decoded_addr(CSRs.mtvec))  { reg_mtvec := wdata >> 2 << 2 }
     when (decoded_addr(CSRs.mcause))   { reg_mcause := wdata & UInt((BigInt(1) << (xLen-1)) + 31) /* only implement 5 LSBs and MSB */ }
     when (decoded_addr(CSRs.mbadaddr)) { reg_mbadaddr := wdata(vaddrBitsExtended-1,0) }

src/main/scala/rocket/Core.scala

@@ -1,93 +0,0 @@
-// See LICENSE.SiFive for license details.
-
-package rocket
-
-import Chisel._
-import config._
-import uncore.tilelink2.TLEdgeOut
-import uncore.util.{CacheName, CacheBlockBytes}
-import util._
-
-case object BuildCore extends Field[(RocketConfig, Parameters) => CoreModule with HasCoreIO]
-case object SharedMemoryTLEdge extends Field[TLEdgeOut]
-
-trait HasCoreParameters {
-  implicit val p: Parameters
-  val xLen = p(XLen)
-  val fLen = xLen // TODO relax this
-
-  val usingVM = p(UseVM)
-  val usingUser = p(UseUser) || usingVM
-  val usingDebug = p(UseDebug)
-  val usingMulDiv = p(MulDivKey).nonEmpty
-  val usingFPU = p(FPUKey).nonEmpty
-  val usingAtomics = p(UseAtomics)
-  val usingCompressed = p(UseCompressed)
-  val usingRoCC = !p(BuildRoCC).isEmpty
-  val fastLoadWord = p(FastLoadWord)
-  val fastLoadByte = p(FastLoadByte)
-  val fastJAL = p(FastJAL)
-  val nBreakpoints = p(NBreakpoints)
-  val nPerfCounters = p(NPerfCounters)
-  val nPerfEvents = p(NPerfEvents)
-  val usingDataScratchpad = p(DataScratchpadSize) > 0
-
-  val retireWidth = p(RetireWidth)
-  val fetchWidth = p(FetchWidth)
-  val coreInstBits = p(CoreInstBits)
-  val coreInstBytes = coreInstBits/8
-  val coreDataBits = xLen
-  val coreDataBytes = coreDataBits/8
-
-  val dcacheArbPorts = 1 + usingVM.toInt + usingDataScratchpad.toInt + p(BuildRoCC).size
-  val coreDCacheReqTagBits = 6
-  val dcacheReqTagBits = coreDCacheReqTagBits + log2Ceil(dcacheArbPorts)
-
-  def pgIdxBits = 12
-  def pgLevelBits = 10 - log2Ceil(xLen / 32)
-  def vaddrBits = pgIdxBits + pgLevels * pgLevelBits
-  val paddrBits = p(PAddrBits)
-  def ppnBits = paddrBits - pgIdxBits
-  def vpnBits = vaddrBits - pgIdxBits
-  val pgLevels = p(PgLevels)
-  val asIdBits = p(ASIdBits)
-  val vpnBitsExtended = vpnBits + (vaddrBits < xLen).toInt
-  val vaddrBitsExtended = vpnBitsExtended + pgIdxBits
-  val coreMaxAddrBits = paddrBits max vaddrBitsExtended
-  val nCustomMrwCsrs = p(NCustomMRWCSRs)
-
-  // fetchWidth doubled, but coreInstBytes halved, for RVC
-  val decodeWidth = fetchWidth / (if (usingCompressed) 2 else 1)
-
-  // Print out log of committed instructions and their writeback values.
-  // Requires post-processing due to out-of-order writebacks.
-  val enableCommitLog = false
-
-  val maxPAddrBits = xLen match {
-    case 32 => 34
-    case 64 => 50
-  }
-
-  require(paddrBits <= maxPAddrBits)
-  require(!fastLoadByte || fastLoadWord)
-}
-
-abstract class CoreModule(implicit val p: Parameters) extends Module
-  with HasCoreParameters
-
-abstract class CoreBundle(implicit val p: Parameters) extends ParameterizedBundle()(p)
-  with HasCoreParameters
-
-trait HasCoreIO {
-  implicit val p: Parameters
-  val io = new Bundle {
-    val interrupts = new TileInterrupts().asInput
-    val hartid = UInt(INPUT, p(XLen))
-    val imem  = new FrontendIO()(p.alterPartial({case CacheName => CacheName("L1I") }))
-    val dmem = new HellaCacheIO()(p.alterPartial({ case CacheName => CacheName("L1D") }))
-    val ptw = new DatapathPTWIO().flip
-    val fpu = new FPUCoreIO().flip
-    val rocc = new RoCCCoreIO().flip
-  }
-}
-

src/main/scala/rocket/DCache.scala

@@ -4,13 +4,13 @@ package rocket
 
 import Chisel._
 import Chisel.ImplicitConversions._
+import config._
 import diplomacy._
 import uncore.constants._
 import uncore.tilelink2._
 import uncore.util._
 import util._
 import TLMessages._
-import config._
 
 class DCacheDataReq(implicit p: Parameters) extends L1HellaCacheBundle()(p) {
   val addr = Bits(width = untagBits)
@@ -38,22 +38,19 @@ class DCacheDataArray(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   }
 }
 
-class DCache(cfg: DCacheConfig, val scratch: () => Option[AddressSet] = () => None)(implicit p: Parameters) extends HellaCache(cfg)(p) {
+class DCache(val scratch: () => Option[AddressSet] = () => None)(implicit p: Parameters) extends HellaCache()(p) {
   override lazy val module = new DCacheModule(this) 
 }
 
 class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
-
-  val maxUncachedInFlight = cfg.nMMIOs
-
   require(rowBits == encRowBits) // no ECC
 
   val grantackq = Module(new Queue(tl_out.e.bits,1)) // TODO don't need this in scratchpad mode
 
   // tags
-  val replacer = p(Replacer)()
+  val replacer = cacheParams.replacement
   def onReset = L1Metadata(UInt(0), ClientMetadata.onReset)
-  val metaReadArb = Module(new Arbiter(new MetaReadReq, 3))
+  val metaReadArb = Module(new Arbiter(new L1MetaReadReq, 3))
   val metaWriteArb = Module(new Arbiter(new L1MetaWriteReq, 3))
 
   // data
@@ -104,7 +101,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   when (!metaReadArb.io.in(2).ready) { io.cpu.req.ready := false }
 
   // address translation
-  val tlb = Module(new TLB)
+  val tlb = Module(new TLB(nTLBEntries))
   io.ptw <> tlb.io.ptw
   tlb.io.req.valid := s1_valid_masked && s1_readwrite
   tlb.io.req.bits.passthrough := s1_req.phys
@@ -126,7 +123,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
       val hitState = Mux(inScratchpad, ClientMetadata.maximum, ClientMetadata.onReset)
       (inScratchpad, hitState, L1Metadata(UInt(0), ClientMetadata.onReset))
     } else {
-      val meta = Module(new MetadataArray(onReset _))
+      val meta = Module(new L1MetadataArray(onReset _))
       meta.io.read <> metaReadArb.io.out
       meta.io.write <> metaWriteArb.io.out
       val s1_meta = meta.io.resp

src/main/scala/rocket/Frontend.scala

@@ -4,13 +4,13 @@
 package rocket
 
 import Chisel._
+import Chisel.ImplicitConversions._
 import config._
 import coreplex._
 import diplomacy._
 import uncore.tilelink2._
-import uncore.util.CacheName
+import tile._
 import util._
-import Chisel.ImplicitConversions._
 
 class FrontendReq(implicit p: Parameters) extends CoreBundle()(p) {
   val pc = UInt(width = vaddrBitsExtended)
@@ -54,12 +54,12 @@ class FrontendBundle(outer: Frontend) extends CoreBundle()(outer.p) {
 
 class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
     with HasCoreParameters
-    with HasL1CacheParameters {
+    with HasL1ICacheParameters {
   val io = new FrontendBundle(outer)
   implicit val edge = outer.node.edgesOut(0)
   val icache = outer.icache.module
 
-  val tlb = Module(new TLB)
+  val tlb = Module(new TLB(nTLBEntries))
 
   val s1_pc_ = Reg(UInt(width=vaddrBitsExtended))
   val s1_pc = ~(~s1_pc_ | (coreInstBytes-1)) // discard PC LSBS (this propagates down the pipeline)
@@ -106,7 +106,7 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
     s2_valid := Bool(false)
   }
 
-  if (p(BtbKey).nEntries > 0) {
+  if (usingBTB) {
     val btb = Module(new BTB)
     btb.io.req.valid := false
     btb.io.req.bits.addr := s1_pc_
@@ -153,21 +153,18 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
 }
 
 /** Mix-ins for constructing tiles that have an ICache-based pipeline frontend */
-trait HasICacheFrontend extends CanHavePTW with TileNetwork {
+trait HasICacheFrontend extends CanHavePTW with HasTileLinkMasterPort {
   val module: HasICacheFrontendModule
-  val frontend = LazyModule(new Frontend()(p.alterPartial({
+  val frontend = LazyModule(new Frontend)
-    case CacheName => CacheName("L1I")
+  masterNode := frontend.node
-  })))
-  l1backend.node := frontend.node
   nPTWPorts += 1
 }
 
-trait HasICacheFrontendBundle extends TileNetworkBundle {
+trait HasICacheFrontendBundle extends HasTileLinkMasterPortBundle {
   val outer: HasICacheFrontend
 }
 
-trait HasICacheFrontendModule extends CanHavePTWModule with TileNetworkModule {
+trait HasICacheFrontendModule extends CanHavePTWModule with HasTileLinkMasterPortModule {
   val outer: HasICacheFrontend
-  //val io: HasICacheFrontendBundle
   ptwPorts += outer.frontend.module.io.ptw
 }

src/main/scala/rocket/HellaCache.scala

@@ -7,43 +7,58 @@ import Chisel._
 import config.{Parameters, Field}
 import coreplex._
 import diplomacy._
+import tile._
 import uncore.constants._
 import uncore.tilelink2._
-import uncore.util._
+import uncore.util.Code
-import util.ParameterizedBundle
+import util.{ParameterizedBundle, RandomReplacement}
 import scala.collection.mutable.ListBuffer
 
-case class DCacheConfig(
+case class DCacheParams(
+    nSets: Int = 64,
+    nWays: Int = 4,
+    rowBits: Int = 64,
+    nTLBEntries: Int = 8,
+    splitMetadata: Boolean = false,
+    ecc: Option[Code] = None,
     nMSHRs: Int = 1,
     nSDQ: Int = 17,
     nRPQ: Int = 16,
-  nMMIOs: Int = 1)
+    nMMIOs: Int = 1) extends L1CacheParams {
+  def replacement = new RandomReplacement(nWays)
+}
 
-case object DCacheKey extends Field[DCacheConfig]
+trait HasL1HellaCacheParameters extends HasL1CacheParameters
+    with HasCoreParameters {
+  val cacheParams = tileParams.dcache.get
+  val cfg = cacheParams
 
-trait HasL1HellaCacheParameters extends HasL1CacheParameters {
+  def wordBits = xLen // really, xLen max 
-  val wordBits = xLen // really, xLen max 
+  def wordBytes = wordBits/8
-  val wordBytes = wordBits/8
+  def wordOffBits = log2Up(wordBytes)
-  val wordOffBits = log2Up(wordBytes)
+  def beatBytes = cacheBlockBytes / cacheDataBeats
-  val beatBytes = cacheBlockBytes / cacheDataBeats
+  def beatWords = beatBytes / wordBytes
-  val beatWords = beatBytes / wordBytes
+  def beatOffBits = log2Up(beatBytes)
-  val beatOffBits = log2Up(beatBytes)
+  def idxMSB = untagBits-1
-  val idxMSB = untagBits-1
+  def idxLSB = blockOffBits
-  val idxLSB = blockOffBits
+  def offsetmsb = idxLSB-1
-  val offsetmsb = idxLSB-1
+  def offsetlsb = wordOffBits
-  val offsetlsb = wordOffBits
+  def rowWords = rowBits/wordBits
-  val rowWords = rowBits/wordBits
+  def doNarrowRead = coreDataBits * nWays % rowBits == 0
-  val doNarrowRead = coreDataBits * nWays % rowBits == 0
+  def encDataBits = code.width(coreDataBits)
-  val encDataBits = code.width(coreDataBits)
+  def encRowBits = encDataBits*rowWords
-  val encRowBits = encDataBits*rowWords
+  def lrscCycles = 32 // ISA requires 16-insn LRSC sequences to succeed
-  val nIOMSHRs = 1
+  def nIOMSHRs = cacheParams.nMMIOs
-  val lrscCycles = 32 // ISA requires 16-insn LRSC sequences to succeed
+  def maxUncachedInFlight = cacheParams.nMMIOs
+  def dataScratchpadSize = tileParams.dataScratchpadBytes
 
-  require(isPow2(nSets))
+  require(isPow2(nSets), s"nSets($nSets) must be pow2")
-  require(rowBits >= coreDataBits)
+  require(rowBits >= coreDataBits, s"rowBits($rowBits) < coreDataBits($coreDataBits)")
-  require(rowBits == cacheDataBits) // TODO should rowBits even be seperably specifiable?
+  // TODO should rowBits even be seperably specifiable?
-  require(xLen <= cacheDataBits) // would need offset addr for puts if data width < xlen
+  require(rowBits == cacheDataBits, s"rowBits($rowBits) != cacheDataBits($cacheDataBits)") 
-  require(!usingVM || untagBits <= pgIdxBits)
+  // would need offset addr for puts if data width < xlen
+  require(xLen <= cacheDataBits, s"xLen($xLen) > cacheDataBits($cacheDataBits)")
+  require(!usingVM || untagBits <= pgIdxBits, s"untagBits($untagBits) > pgIdxBits($pgIdxBits)")
 }
 
 abstract class L1HellaCacheModule(implicit val p: Parameters) extends Module
@@ -52,25 +67,7 @@ abstract class L1HellaCacheModule(implicit val p: Parameters) extends Module
 abstract class L1HellaCacheBundle(implicit val p: Parameters) extends ParameterizedBundle()(p)
   with HasL1HellaCacheParameters
 
-class L1Metadata(implicit p: Parameters) extends Metadata()(p) with HasL1HellaCacheParameters {
+/** Bundle definitions for HellaCache interfaces */
-  val coh = new ClientMetadata
-}
-object L1Metadata {
-  def apply(tag: Bits, coh: ClientMetadata)(implicit p: Parameters) = {
-    val meta = Wire(new L1Metadata)
-    meta.tag := tag
-    meta.coh := coh
-    meta
-  }
-}
-
-class L1MetaReadReq(implicit p: Parameters) extends MetaReadReq {
-  val tag = Bits(width = tagBits)
-  override def cloneType = new L1MetaReadReq()(p).asInstanceOf[this.type] //TODO remove
-}
-
-class L1MetaWriteReq(implicit p: Parameters) extends 
-  MetaWriteReq[L1Metadata](new L1Metadata)
 
 trait HasCoreMemOp extends HasCoreParameters {
   val addr = UInt(width = coreMaxAddrBits)
@@ -83,6 +80,10 @@ trait HasCoreData extends HasCoreParameters {
   val data = Bits(width = coreDataBits)
 }
 
+class HellaCacheReqInternal(implicit p: Parameters) extends CoreBundle()(p) with HasCoreMemOp {
+  val phys = Bool()
+}
+
 class HellaCacheReq(implicit p: Parameters) extends HellaCacheReqInternal()(p) with HasCoreData
 
 class HellaCacheResp(implicit p: Parameters) extends CoreBundle()(p)
@@ -104,7 +105,6 @@ class HellaCacheExceptions extends Bundle {
   val pf = new AlignmentExceptions
 }
 
-
 // interface between D$ and processor/DTLB
 class HellaCacheIO(implicit p: Parameters) extends CoreBundle()(p) {
   val req = Decoupled(new HellaCacheReq)
@@ -119,10 +119,13 @@ class HellaCacheIO(implicit p: Parameters) extends CoreBundle()(p) {
   val ordered = Bool(INPUT)
 }
 
-abstract class HellaCache(val cfg: DCacheConfig)(implicit p: Parameters) extends LazyModule {
+/** Base classes for Diplomatic TL2 HellaCaches */
+
+abstract class HellaCache(implicit p: Parameters) extends LazyModule {
+  private val cfg = p(TileKey).dcache.get
   val node = TLClientNode(TLClientParameters(
     sourceId = IdRange(0, cfg.nMSHRs + cfg.nMMIOs),
-    supportsProbe = TransferSizes(p(CacheBlockBytes))))
+    supportsProbe = TransferSizes(1, p(CacheBlockBytes))))
   val module: HellaCacheModule
 }
 
@@ -135,38 +138,101 @@ class HellaCacheBundle(outer: HellaCache) extends Bundle {
 
 class HellaCacheModule(outer: HellaCache) extends LazyModuleImp(outer)
     with HasL1HellaCacheParameters {
-  implicit val cfg = outer.cfg
   implicit val edge = outer.node.edgesOut(0)
   val io = new HellaCacheBundle(outer)
   val tl_out = io.mem(0)
 }
 
 object HellaCache {
-  def apply(cfg: DCacheConfig, scratch: () => Option[AddressSet] = () => None)(implicit p: Parameters) = {
+  def apply(blocking: Boolean, scratch: () => Option[AddressSet] = () => None)(implicit p: Parameters) = {
-    if (cfg.nMSHRs == 0) LazyModule(new DCache(cfg, scratch))
+    if (blocking) LazyModule(new DCache(scratch))
-    else LazyModule(new NonBlockingDCache(cfg))
+    else LazyModule(new NonBlockingDCache)
   }
 }
 
 /** Mix-ins for constructing tiles that have a HellaCache */
-trait HasHellaCache extends TileNetwork {
+
+trait HasHellaCache extends HasTileLinkMasterPort {
   val module: HasHellaCacheModule
   implicit val p: Parameters
   def findScratchpadFromICache: Option[AddressSet]
   var nDCachePorts = 0
-  val dcacheParams = p.alterPartial({ case CacheName => CacheName("L1D") })
+  val dcache = HellaCache(tileParams.dcache.get.nMSHRs == 0, findScratchpadFromICache _)
-  val dcache = HellaCache(p(DCacheKey), findScratchpadFromICache _)(dcacheParams)
+  masterNode := dcache.node
-  l1backend.node := dcache.node
 }
 
-trait HasHellaCacheBundle extends TileNetworkBundle {
+trait HasHellaCacheBundle extends HasTileLinkMasterPortBundle {
   val outer: HasHellaCache
 }
 
-trait HasHellaCacheModule extends TileNetworkModule {
+trait HasHellaCacheModule extends HasTileLinkMasterPortModule {
   val outer: HasHellaCache
   //val io: HasHellaCacheBundle
   val dcachePorts = ListBuffer[HellaCacheIO]()
-  val dcacheArb = Module(new HellaCacheArbiter(outer.nDCachePorts)(outer.dcacheParams))
+  val dcacheArb = Module(new HellaCacheArbiter(outer.nDCachePorts)(outer.p))
   outer.dcache.module.io.cpu <> dcacheArb.io.mem
 }
+
+/** Metadata array used for all HellaCaches */
+
+class L1Metadata(implicit p: Parameters) extends L1HellaCacheBundle()(p) {
+  val coh = new ClientMetadata
+  val tag = UInt(width = tagBits)
+}
+
+object L1Metadata {
+  def apply(tag: Bits, coh: ClientMetadata)(implicit p: Parameters) = {
+    val meta = Wire(new L1Metadata)
+    meta.tag := tag
+    meta.coh := coh
+    meta
+  }
+}
+
+class L1MetaReadReq(implicit p: Parameters) extends L1HellaCacheBundle()(p) {
+  val idx    = UInt(width = idxBits)
+  val way_en = UInt(width = nWays)
+  val tag    = UInt(width = tagBits)
+}
+
+class L1MetaWriteReq(implicit p: Parameters) extends L1MetaReadReq()(p) {
+  val data = new L1Metadata
+}
+
+class L1MetadataArray[T <: L1Metadata](onReset: () => T)(implicit p: Parameters) extends L1HellaCacheModule()(p) {
+  val rstVal = onReset()
+  val io = new Bundle {
+    val read = Decoupled(new L1MetaReadReq).flip
+    val write = Decoupled(new L1MetaWriteReq).flip
+    val resp = Vec(nWays, rstVal.cloneType).asOutput
+  }
+  val rst_cnt = Reg(init=UInt(0, log2Up(nSets+1)))
+  val rst = rst_cnt < UInt(nSets)
+  val waddr = Mux(rst, rst_cnt, io.write.bits.idx)
+  val wdata = Mux(rst, rstVal, io.write.bits.data).asUInt
+  val wmask = Mux(rst || Bool(nWays == 1), SInt(-1), io.write.bits.way_en.asSInt).toBools
+  val rmask = Mux(rst || Bool(nWays == 1), SInt(-1), io.read.bits.way_en.asSInt).toBools
+  when (rst) { rst_cnt := rst_cnt+UInt(1) }
+
+  val metabits = rstVal.getWidth
+
+  if (hasSplitMetadata) {
+    val tag_arrs = List.fill(nWays){ SeqMem(nSets, UInt(width = metabits)) }
+    val tag_readout = Wire(Vec(nWays,rstVal.cloneType))
+    (0 until nWays).foreach { (i) =>
+      when (rst || (io.write.valid && wmask(i))) {
+        tag_arrs(i).write(waddr, wdata)
+      }
+      io.resp(i) := rstVal.fromBits(tag_arrs(i).read(io.read.bits.idx, io.read.valid && rmask(i)))
+    }
+  } else {
+    val tag_arr = SeqMem(nSets, Vec(nWays, UInt(width = metabits)))
+    when (rst || io.write.valid) {
+      tag_arr.write(waddr, Vec.fill(nWays)(wdata), wmask)
+    }
+    io.resp := tag_arr.read(io.read.bits.idx, io.read.valid).map(rstVal.fromBits(_))
+  }
+
+  io.read.ready := !rst && !io.write.valid // so really this could be a 6T RAM
+  io.write.ready := !rst
+}

src/main/scala/rocket/IBuf.scala

@@ -3,9 +3,10 @@
 package rocket
 
 import Chisel._
-import util._
 import Chisel.ImplicitConversions._
 import config._
+import tile._
+import util._
 
 class Instruction(implicit val p: Parameters) extends ParameterizedBundle with HasCoreParameters {
   val pf0 = Bool() // page fault on first half of instruction

src/main/scala/rocket/ICache.scala

@@ -6,25 +6,32 @@ package rocket
 import Chisel._
 import config._
 import diplomacy._
-import uncore.agents._
+import tile._
 import uncore.tilelink2._
-import uncore.util._
+import uncore.util.Code
 import util._
 import Chisel.ImplicitConversions._
 
-trait HasL1CacheParameters extends HasCacheParameters with HasCoreParameters {
+case class ICacheParams(
-  val cacheBlockBytes = p(CacheBlockBytes)
+    nSets: Int = 64,
-  val lgCacheBlockBytes = log2Up(cacheBlockBytes)
+    nWays: Int = 4,
-  val cacheDataBits = p(SharedMemoryTLEdge).bundle.dataBits
+    rowBits: Int = 128,
-  val cacheDataBeats = (cacheBlockBytes * 8) / cacheDataBits
+    nTLBEntries: Int = 8,
-  val refillCycles = cacheDataBeats
+    cacheIdBits: Int = 0,
+    splitMetadata: Boolean = false,
+    ecc: Option[Code] = None) extends L1CacheParams {
+  def replacement = new RandomReplacement(nWays)
 }
 
-class ICacheReq(implicit p: Parameters) extends CoreBundle()(p) with HasL1CacheParameters {
+trait HasL1ICacheParameters extends HasL1CacheParameters with HasCoreParameters {
+  val cacheParams = tileParams.icache.get
+}
+
+class ICacheReq(implicit p: Parameters) extends CoreBundle()(p) with HasL1ICacheParameters {
   val addr = UInt(width = vaddrBits)
 }
 
-class ICacheResp(implicit p: Parameters) extends CoreBundle()(p) with HasL1CacheParameters {
+class ICacheResp(implicit p: Parameters) extends CoreBundle()(p) with HasL1ICacheParameters {
   val data = Bits(width = coreInstBits)
   val datablock = Bits(width = rowBits)
 }
@@ -46,8 +53,7 @@ class ICacheBundle(outer: ICache) extends CoreBundle()(outer.p) {
 }
 
 class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
-    with HasCoreParameters
+    with HasL1ICacheParameters {
-    with HasL1CacheParameters {
   val io = new ICacheBundle(outer)
   val edge = outer.node.edgesOut(0)
   val tl_out = io.mem(0)

src/main/scala/rocket/IDecode.scala

@@ -4,12 +4,13 @@
 package rocket
 
 import Chisel._
+import Chisel.ImplicitConversions._
 import Instructions._
 import uncore.constants.MemoryOpConstants._
 import ALU._
 import config._
+import tile.HasCoreParameters
 import util._
-import Chisel.ImplicitConversions._
 
 abstract trait DecodeConstants extends HasCoreParameters
 {

src/main/scala/rocket/Multiplier.scala

@@ -29,13 +29,13 @@ class MultiplierIO(dataBits: Int, tagBits: Int) extends Bundle {
   val resp = Decoupled(new MultiplierResp(dataBits, tagBits))
 }
 
-case class MulDivConfig(
+case class MulDivParams(
   mulUnroll: Int = 1,
   mulEarlyOut: Boolean = false,
   divEarlyOut: Boolean = false
 )
 
-class MulDiv(cfg: MulDivConfig, width: Int, nXpr: Int = 32) extends Module {
+class MulDiv(cfg: MulDivParams, width: Int, nXpr: Int = 32) extends Module {
   val io = new MultiplierIO(width, log2Up(nXpr))
   val w = io.req.bits.in1.getWidth
   val mulw = (w + cfg.mulUnroll - 1) / cfg.mulUnroll * cfg.mulUnroll

src/main/scala/rocket/NBDcache.scala

@@ -5,13 +5,12 @@ package rocket
 
 import Chisel._
 import Chisel.ImplicitConversions._
+import config._
 import diplomacy._
 import uncore.constants._
-import uncore.tilelink._
 import uncore.tilelink2._
 import uncore.util._
 import util._
-import config._
 
 trait HasMissInfo extends HasL1HellaCacheParameters {
   val tag_match = Bool()
@@ -19,11 +18,6 @@ trait HasMissInfo extends HasL1HellaCacheParameters {
   val way_en = Bits(width = nWays)
 }
 
-class HellaCacheReqInternal(implicit p: Parameters) extends CoreBundle()(p)
-    with HasCoreMemOp {
-  val phys = Bool()
-}
-
 class L1DataReadReq(implicit p: Parameters) extends L1HellaCacheBundle()(p) {
   val way_en = Bits(width = nWays)
   val addr   = Bits(width = untagBits)
@@ -38,18 +32,14 @@ class L1RefillReq(implicit p: Parameters) extends L1DataReadReq()(p)
 
 class Replay(implicit p: Parameters) extends HellaCacheReqInternal()(p) with HasCoreData
 
-class ReplayInternal(cfg: DCacheConfig)(implicit p: Parameters) extends HellaCacheReqInternal()(p) {
+class ReplayInternal(implicit p: Parameters) extends HellaCacheReqInternal()(p)
+    with HasL1HellaCacheParameters {
   val sdq_id = UInt(width = log2Up(cfg.nSDQ))
-
-  override def cloneType = new ReplayInternal(cfg)(p).asInstanceOf[this.type]
 }
 
 class MSHRReq(implicit p: Parameters) extends Replay()(p) with HasMissInfo
 
-class MSHRReqInternal(cfg: DCacheConfig)(implicit p: Parameters)
+class MSHRReqInternal(implicit p: Parameters) extends ReplayInternal()(p) with HasMissInfo
-    extends ReplayInternal(cfg)(p) with HasMissInfo {
-  override def cloneType = new MSHRReqInternal(cfg)(p).asInstanceOf[this.type]
-}
 
 class WritebackReq(params: TLBundleParameters)(implicit p: Parameters) extends L1HellaCacheBundle()(p) {
   val tag = Bits(width = tagBits)
@@ -145,13 +135,13 @@ class IOMSHR(id: Int)(implicit edge: TLEdgeOut, p: Parameters) extends L1HellaCa
   }
 }
 
-class MSHR(id: Int)(implicit edge: TLEdgeOut, cfg: DCacheConfig, p: Parameters) extends L1HellaCacheModule()(p) {
+class MSHR(id: Int)(implicit edge: TLEdgeOut, p: Parameters) extends L1HellaCacheModule()(p) {
   val io = new Bundle {
     val req_pri_val    = Bool(INPUT)
     val req_pri_rdy    = Bool(OUTPUT)
     val req_sec_val    = Bool(INPUT)
     val req_sec_rdy    = Bool(OUTPUT)
-    val req_bits       = new MSHRReqInternal(cfg).asInput
+    val req_bits       = new MSHRReqInternal().asInput
 
     val idx_match       = Bool(OUTPUT)
     val tag             = Bits(OUTPUT, tagBits)
@@ -163,7 +153,7 @@ class MSHR(id: Int)(implicit edge: TLEdgeOut, cfg: DCacheConfig, p: Parameters)
     val refill = new L1RefillReq().asOutput // Data is bypassed
     val meta_read = Decoupled(new L1MetaReadReq)
     val meta_write = Decoupled(new L1MetaWriteReq)
-    val replay = Decoupled(new ReplayInternal(cfg))
+    val replay = Decoupled(new ReplayInternal)
     val wb_req = Decoupled(new WritebackReq(edge.bundle))
     val probe_rdy = Bool(OUTPUT)
   }
@@ -171,7 +161,7 @@ class MSHR(id: Int)(implicit edge: TLEdgeOut, cfg: DCacheConfig, p: Parameters)
   val s_invalid :: s_wb_req :: s_wb_resp :: s_meta_clear :: s_refill_req :: s_refill_resp :: s_meta_write_req :: s_meta_write_resp :: s_drain_rpq :: Nil = Enum(UInt(), 9)
   val state = Reg(init=s_invalid)
 
-  val req = Reg(new MSHRReqInternal(cfg))
+  val req = Reg(new MSHRReqInternal)
   val req_idx = req.addr(untagBits-1,blockOffBits)
   val req_tag = req.addr >> untagBits
   val req_block_addr = (req.addr >> blockOffBits) << blockOffBits
@@ -195,7 +185,7 @@ class MSHR(id: Int)(implicit edge: TLEdgeOut, cfg: DCacheConfig, p: Parameters)
                     (state.isOneOf(s_refill_req, s_refill_resp) &&
                       !cmd_requires_second_acquire && !refill_done))
 
-  val rpq = Module(new Queue(new ReplayInternal(cfg), cfg.nRPQ))
+  val rpq = Module(new Queue(new ReplayInternal, cfg.nRPQ))
   rpq.io.enq.valid := (io.req_pri_val && io.req_pri_rdy || io.req_sec_val && sec_rdy) && !isPrefetch(io.req_bits.cmd)
   rpq.io.enq.bits := io.req_bits
   rpq.io.deq.ready := (io.replay.ready && state === s_drain_rpq) || state === s_invalid
@@ -310,7 +300,7 @@ class MSHR(id: Int)(implicit edge: TLEdgeOut, cfg: DCacheConfig, p: Parameters)
   }
 }
 
-class MSHRFile(implicit edge: TLEdgeOut, cfg: DCacheConfig, p: Parameters) extends L1HellaCacheModule()(p) {
+class MSHRFile(implicit edge: TLEdgeOut, p: Parameters) extends L1HellaCacheModule()(p) {
   val io = new Bundle {
     val req = Decoupled(new MSHRReq).flip
     val resp = Decoupled(new HellaCacheResp)
@@ -350,7 +340,7 @@ class MSHRFile(implicit edge: TLEdgeOut, cfg: DCacheConfig, p: Parameters) exten
   val meta_read_arb = Module(new Arbiter(new L1MetaReadReq, cfg.nMSHRs))
   val meta_write_arb = Module(new Arbiter(new L1MetaWriteReq, cfg.nMSHRs))
   val wb_req_arb = Module(new Arbiter(new WritebackReq(edge.bundle), cfg.nMSHRs))
-  val replay_arb = Module(new Arbiter(new ReplayInternal(cfg), cfg.nMSHRs))
+  val replay_arb = Module(new Arbiter(new ReplayInternal, cfg.nMSHRs))
   val alloc_arb = Module(new Arbiter(Bool(), cfg.nMSHRs))
 
   var idx_match = Bool(false)
@@ -671,14 +661,14 @@ class DataArray(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   io.write.ready := Bool(true)
 }
 
-class NonBlockingDCache(cfg: DCacheConfig)(implicit p: Parameters) extends HellaCache(cfg)(p) {
+class NonBlockingDCache(implicit p: Parameters) extends HellaCache()(p) {
   override lazy val module = new NonBlockingDCacheModule(this) 
 }
 
 class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule(outer) {
 
   require(isPow2(nWays)) // TODO: relax this
-  require(p(DataScratchpadSize) == 0)
+  require(dataScratchpadSize == 0)
 
   val wb = Module(new WritebackUnit)
   val prober = Module(new ProbeUnit)
@@ -706,7 +696,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   val s1_write = isWrite(s1_req.cmd)
   val s1_readwrite = s1_read || s1_write || isPrefetch(s1_req.cmd)
 
-  val dtlb = Module(new TLB)
+  val dtlb = Module(new TLB(nTLBEntries))
   io.ptw <> dtlb.io.ptw
   dtlb.io.req.valid := s1_valid_masked && s1_readwrite
   dtlb.io.req.bits.passthrough := s1_req.phys
@@ -754,8 +744,8 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
 
   // tags
   def onReset = L1Metadata(UInt(0), ClientMetadata.onReset)
-  val meta = Module(new MetadataArray(onReset _))
+  val meta = Module(new L1MetadataArray(onReset _))
-  val metaReadArb = Module(new Arbiter(new MetaReadReq, 5))
+  val metaReadArb = Module(new Arbiter(new L1MetaReadReq, 5))
   val metaWriteArb = Module(new Arbiter(new L1MetaWriteReq, 2))
   meta.io.read <> metaReadArb.io.out
   meta.io.write <> metaWriteArb.io.out
@@ -852,7 +842,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   writeArb.io.in(0).bits.way_en :=  s3_way
 
   // replacement policy
-  val replacer = p(Replacer)()
+  val replacer = cacheParams.replacement
   val s1_replaced_way_en = UIntToOH(replacer.way)
   val s2_replaced_way_en = UIntToOH(RegEnable(replacer.way, s1_clk_en))
   val s2_repl_meta = Mux1H(s2_replaced_way_en, wayMap((w: Int) => RegEnable(meta.io.resp(w), s1_clk_en && s1_replaced_way_en(w))).toSeq)

src/main/scala/rocket/PTW.scala

@@ -4,11 +4,12 @@
 package rocket
 
 import Chisel._
-import config._
-import uncore.constants._
-import uncore.util.PseudoLRU
-import util._
 import Chisel.ImplicitConversions._
+import config._
+import tile._
+import uncore.constants._
+import util._
+
 import scala.collection.mutable.ListBuffer
 
 class PTWReq(implicit p: Parameters) extends CoreBundle()(p) {
@@ -224,11 +225,12 @@ trait CanHavePTW extends HasHellaCache {
   implicit val p: Parameters
   val module: CanHavePTWModule
   var nPTWPorts = 1
+  nDCachePorts += usingPTW.toInt
 }
 
 trait CanHavePTWModule extends HasHellaCacheModule {
   val outer: CanHavePTW
   val ptwPorts = ListBuffer(outer.dcache.module.io.ptw)
-  val ptwOpt = if (outer.p(UseVM)) { Some(Module(new PTW(outer.nPTWPorts)(outer.p))) } else None
+  val ptwOpt = if (outer.usingPTW) { Some(Module(new PTW(outer.nPTWPorts)(outer.p))) } else None
   ptwOpt foreach { ptw => dcachePorts += ptw.io.mem }
 }

src/main/scala/rocket/RVC.scala

@@ -4,8 +4,9 @@ package rocket
 
 import Chisel._
 import Chisel.ImplicitConversions._
-import util._
 import config._
+import tile._
+import util._
 
 class ExpandedInstruction extends Bundle {
   val bits = UInt(width = 32)
@@ -150,14 +151,14 @@ class RVCDecoder(x: UInt)(implicit p: Parameters) {
   }
 }
 
-class RVCExpander(implicit p: Parameters) extends Module {
+class RVCExpander(implicit val p: Parameters) extends Module with HasCoreParameters {
   val io = new Bundle {
     val in = UInt(INPUT, 32)
     val out = new ExpandedInstruction
     val rvc = Bool(OUTPUT)
   }
 
-  if (p(UseCompressed)) {
+  if (usingCompressed) {
     io.rvc := io.in(1,0) =/= UInt(3)
     io.out := new RVCDecoder(io.in).decode
   } else {

src/main/scala/rocket/Rocket.scala

@@ -5,76 +5,57 @@ package rocket
 
 import Chisel._
 import config._
+import tile._
 import uncore.constants._
 import util._
 import Chisel.ImplicitConversions._
 
-case class RocketConfig(xLen: Int)
+case class RocketCoreParams(
-// TODO replace some of below fields with above Config
+  useVM: Boolean = true,
-case object XLen extends Field[Int]
+  useUser: Boolean = false,
-case object FetchWidth extends Field[Int]
+  useDebug: Boolean = true,
-case object RetireWidth extends Field[Int]
+  useAtomics: Boolean = true,
-case object FPUKey extends Field[Option[FPUConfig]]
+  useCompressed: Boolean = true,
-case object MulDivKey extends Field[Option[MulDivConfig]]
+  nBreakpoints: Int = 1,
-case object UseVM extends Field[Boolean]
+  nPerfCounters: Int = 0,
-case object UseUser extends Field[Boolean]
+  nPerfEvents: Int = 0,
-case object UseDebug extends Field[Boolean]
+  nCustomMRWCSRs: Int = 0,
-case object UseAtomics extends Field[Boolean]
+  mtvecInit: Option[BigInt] = Some(BigInt(0)),
-case object UseCompressed extends Field[Boolean]
+  mtvecWritable: Boolean = true,
-case object FastLoadWord extends Field[Boolean]
+  fastLoadWord: Boolean = true,
-case object FastLoadByte extends Field[Boolean]
+  fastLoadByte: Boolean = false,
-case object FastJAL extends Field[Boolean]
+  fastJAL: Boolean = false,
-case object CoreInstBits extends Field[Int]
+  mulDiv: Option[MulDivParams] = Some(MulDivParams()),
-case object NCustomMRWCSRs extends Field[Int]
+  fpu: Option[FPUParams] = Some(FPUParams())
-case object MtvecWritable extends Field[Boolean]
+) extends CoreParams {
-case object MtvecInit extends Field[Option[BigInt]]
+  val fetchWidth: Int = if (useCompressed) 2 else 1
-case object NBreakpoints extends Field[Int]
+  //  fetchWidth doubled, but coreInstBytes halved, for RVC:
-case object NPerfCounters extends Field[Int]
+  val decodeWidth: Int = fetchWidth / (if (useCompressed) 2 else 1)
-case object NPerfEvents extends Field[Int]
+  val retireWidth: Int = 1
-case object DataScratchpadSize extends Field[Int]
+  val instBits: Int = if (useCompressed) 16 else 32
-
-class RegFile(n: Int, w: Int, zero: Boolean = false) {
-  private val rf = Mem(n, UInt(width = w))
-  private def access(addr: UInt) = rf(~addr(log2Up(n)-1,0))
-  private val reads = collection.mutable.ArrayBuffer[(UInt,UInt)]()
-  private var canRead = true
-  def read(addr: UInt) = {
-    require(canRead)
-    reads += addr -> Wire(UInt())
-    reads.last._2 := Mux(Bool(zero) && addr === UInt(0), UInt(0), access(addr))
-    reads.last._2
-  }
-  def write(addr: UInt, data: UInt) = {
-    canRead = false
-    when (addr =/= UInt(0)) {
-      access(addr) := data
-      for ((raddr, rdata) <- reads)
-        when (addr === raddr) { rdata := data }
-    }
-  }
 }
 
-object ImmGen {
+trait HasRocketCoreParameters extends HasCoreParameters {
-  def apply(sel: UInt, inst: UInt) = {
+  val rocketParams: RocketCoreParams = tileParams.core.asInstanceOf[RocketCoreParams]
-    val sign = Mux(sel === IMM_Z, SInt(0), inst(31).asSInt)
-    val b30_20 = Mux(sel === IMM_U, inst(30,20).asSInt, sign)
-    val b19_12 = Mux(sel =/= IMM_U && sel =/= IMM_UJ, sign, inst(19,12).asSInt)
-    val b11 = Mux(sel === IMM_U || sel === IMM_Z, SInt(0),
-              Mux(sel === IMM_UJ, inst(20).asSInt,
-              Mux(sel === IMM_SB, inst(7).asSInt, sign)))
-    val b10_5 = Mux(sel === IMM_U || sel === IMM_Z, Bits(0), inst(30,25))
-    val b4_1 = Mux(sel === IMM_U, Bits(0),
-               Mux(sel === IMM_S || sel === IMM_SB, inst(11,8),
-               Mux(sel === IMM_Z, inst(19,16), inst(24,21))))
-    val b0 = Mux(sel === IMM_S, inst(7),
-             Mux(sel === IMM_I, inst(20),
-             Mux(sel === IMM_Z, inst(15), Bits(0))))
 
-    Cat(sign, b30_20, b19_12, b11, b10_5, b4_1, b0).asSInt
+  val fastLoadWord = rocketParams.fastLoadWord
-  }
+  val fastLoadByte = rocketParams.fastLoadByte
+  val fastJAL = rocketParams.fastJAL
+  val nBreakpoints = rocketParams.nBreakpoints
+  val nPerfCounters = rocketParams.nPerfCounters
+  val nPerfEvents = rocketParams.nPerfEvents
+  val nCustomMrwCsrs = rocketParams.nCustomMRWCSRs
+  val mtvecInit = rocketParams.mtvecInit
+  val mtvecWritable = rocketParams.mtvecWritable
+
+  val mulDivParams = rocketParams.mulDiv.getOrElse(MulDivParams()) // TODO ask andrew about this
+
+  require(!fastLoadByte || fastLoadWord)
 }
 
-class Rocket(val c: RocketConfig)(implicit p: Parameters) extends CoreModule()(p) with HasCoreIO {
+class Rocket(implicit p: Parameters) extends CoreModule()(p)
+    with HasRocketCoreParameters
+    with HasCoreIO {
 
   val decode_table = {
     (if (usingMulDiv) new MDecode +: (xLen > 32).option(new M64Decode).toSeq else Nil) ++:
@@ -243,7 +224,7 @@ class Rocket(val c: RocketConfig)(implicit p: Parameters) extends CoreModule()(p
   alu.io.in1 := ex_op1.asUInt
   
   // multiplier and divider
-  val div = Module(new MulDiv(p(MulDivKey).getOrElse(MulDivConfig()), width = xLen))
+  val div = Module(new MulDiv(mulDivParams, width = xLen))
   div.io.req.valid := ex_reg_valid && ex_ctrl.div
   div.io.req.bits.dw := ex_ctrl.alu_dw
   div.io.req.bits.fn := ex_ctrl.alu_fn
@@ -324,9 +305,7 @@ class Rocket(val c: RocketConfig)(implicit p: Parameters) extends CoreModule()(p
   val mem_int_wdata = Mux(!mem_reg_xcpt && (mem_ctrl.jalr ^ mem_npc_misaligned), mem_br_target, mem_reg_wdata.asSInt).asUInt
   val mem_cfi = mem_ctrl.branch || mem_ctrl.jalr || mem_ctrl.jal
   val mem_cfi_taken = (mem_ctrl.branch && mem_br_taken) || mem_ctrl.jalr || (Bool(!fastJAL) && mem_ctrl.jal)
-  val mem_misprediction =
+  val mem_misprediction = if (usingBTB) mem_wrong_npc else mem_cfi_taken
-    if (p(BtbKey).nEntries == 0) mem_cfi_taken
-    else mem_wrong_npc
   take_pc_mem := mem_reg_valid && (mem_misprediction || mem_reg_flush_pipe)
 
   mem_reg_valid := !ctrl_killx
@@ -652,3 +631,44 @@ class Rocket(val c: RocketConfig)(implicit p: Parameters) extends CoreModule()(p
     }
   }
 }
+
+class RegFile(n: Int, w: Int, zero: Boolean = false) {
+  private val rf = Mem(n, UInt(width = w))
+  private def access(addr: UInt) = rf(~addr(log2Up(n)-1,0))
+  private val reads = collection.mutable.ArrayBuffer[(UInt,UInt)]()
+  private var canRead = true
+  def read(addr: UInt) = {
+    require(canRead)
+    reads += addr -> Wire(UInt())
+    reads.last._2 := Mux(Bool(zero) && addr === UInt(0), UInt(0), access(addr))
+    reads.last._2
+  }
+  def write(addr: UInt, data: UInt) = {
+    canRead = false
+    when (addr =/= UInt(0)) {
+      access(addr) := data
+      for ((raddr, rdata) <- reads)
+        when (addr === raddr) { rdata := data }
+    }
+  }
+}
+
+object ImmGen {
+  def apply(sel: UInt, inst: UInt) = {
+    val sign = Mux(sel === IMM_Z, SInt(0), inst(31).asSInt)
+    val b30_20 = Mux(sel === IMM_U, inst(30,20).asSInt, sign)
+    val b19_12 = Mux(sel =/= IMM_U && sel =/= IMM_UJ, sign, inst(19,12).asSInt)
+    val b11 = Mux(sel === IMM_U || sel === IMM_Z, SInt(0),
+              Mux(sel === IMM_UJ, inst(20).asSInt,
+              Mux(sel === IMM_SB, inst(7).asSInt, sign)))
+    val b10_5 = Mux(sel === IMM_U || sel === IMM_Z, Bits(0), inst(30,25))
+    val b4_1 = Mux(sel === IMM_U, Bits(0),
+               Mux(sel === IMM_S || sel === IMM_SB, inst(11,8),
+               Mux(sel === IMM_Z, inst(19,16), inst(24,21))))
+    val b0 = Mux(sel === IMM_S, inst(7),
+             Mux(sel === IMM_I, inst(20),
+             Mux(sel === IMM_Z, inst(15), Bits(0))))
+
+    Cat(sign, b30_20, b19_12, b11, b10_5, b4_1, b0).asSInt
+  }
+}

src/main/scala/rocket/ScratchpadSlavePort.scala

@@ -4,22 +4,24 @@ package rocket
 
 import Chisel._
 import Chisel.ImplicitConversions._
-import junctions._
-import diplomacy._
 import config._
+import coreplex.CacheBlockBytes
+import diplomacy._
+import tile._
 import uncore.constants._
 import uncore.tilelink2._
 import uncore.util._
+import util._
 
-class ScratchpadSlavePort(implicit p: Parameters) extends LazyModule {
+class ScratchpadSlavePort(sizeBytes: Int)(implicit p: Parameters) extends LazyModule
-  val coreDataBytes = p(XLen)/8
+    with HasCoreParameters {
   val node = TLManagerNode(Seq(TLManagerPortParameters(
     Seq(TLManagerParameters(
-      address            = List(AddressSet(0x80000000L, BigInt(p(DataScratchpadSize)-1))),
+      address            = List(AddressSet(0x80000000L, BigInt(sizeBytes-1))),
       regionType         = RegionType.UNCACHED,
       executable         = true,
-      supportsArithmetic = if (p(UseAtomics)) TransferSizes(1, coreDataBytes) else TransferSizes.none,
+      supportsArithmetic = if (usingAtomics) TransferSizes(1, coreDataBytes) else TransferSizes.none,
-      supportsLogical    = if (p(UseAtomics)) TransferSizes(1, coreDataBytes) else TransferSizes.none,
+      supportsLogical    = if (usingAtomics) TransferSizes(1, coreDataBytes) else TransferSizes.none,
       supportsPutPartial = TransferSizes(1, coreDataBytes),
       supportsPutFull    = TransferSizes(1, coreDataBytes),
       supportsGet        = TransferSizes(1, coreDataBytes),
@@ -106,10 +108,11 @@ class ScratchpadSlavePort(implicit p: Parameters) extends LazyModule {
 trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend {
   val module: CanHaveScratchpadModule
 
-  val slaveNode = if (p(DataScratchpadSize) == 0) None else Some(TLInputNode())
+  val sizeBytes = tileParams.dataScratchpadBytes
-  val scratch = if (p(DataScratchpadSize) == 0) None else Some(LazyModule(new ScratchpadSlavePort()(dcacheParams)))
+  val slaveNode = TLInputNode()
+  val scratch   = if (sizeBytes > 0) Some(LazyModule(new ScratchpadSlavePort(sizeBytes))) else None
 
-  (slaveNode zip scratch) foreach { case (node, lm) => lm.node := TLFragmenter(p(XLen)/8, p(CacheBlockBytes))(node) }
+  scratch foreach { lm => lm.node := TLFragmenter(p(XLen)/8, p(CacheBlockBytes))(slaveNode) }
 
   def findScratchpadFromICache: Option[AddressSet] = scratch.map { s =>
     val finalNode = frontend.node.edgesOut(0).manager.managers.find(_.nodePath.last == s.node)
@@ -118,12 +121,12 @@ trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend {
     finalNode.get.address(0)
   }
 
-  nDCachePorts += 1 // core TODO dcachePorts += () => module.io.dmem ??
+  nDCachePorts += (sizeBytes > 0).toInt
 }
 
 trait CanHaveScratchpadBundle extends HasHellaCacheBundle with HasICacheFrontendBundle {
   val outer: CanHaveScratchpad
-  val slave = outer.slaveNode.map(_.bundleIn)
+  val slave = outer.slaveNode.bundleIn
 }
 
 trait CanHaveScratchpadModule extends HasHellaCacheModule with HasICacheFrontendModule {

src/main/scala/rocket/TLB.scala

@@ -4,24 +4,18 @@
 package rocket
 
 import Chisel._
-import util._
 import Chisel.ImplicitConversions._
-import scala.math._
 import config._
 import diplomacy._
-import uncore.util._
+import coreplex.CacheBlockBytes
+import tile.{XLen, CoreModule, CoreBundle}
 import uncore.tilelink2._
+import util._
 
 case object PAddrBits extends Field[Int]
 case object PgLevels extends Field[Int]
 case object ASIdBits extends Field[Int]
 
-trait HasTLBParameters extends HasL1CacheParameters {
-  val entries = p(p(CacheName)).nTLBEntries
-  val camAddrBits = log2Ceil(entries)
-  val camTagBits = asIdBits + vpnBits
-}
-
 class TLBReq(implicit p: Parameters) extends CoreBundle()(p) {
   val vpn = UInt(width = vpnBitsExtended)
   val passthrough = Bool()
@@ -39,12 +33,15 @@ class TLBResp(implicit p: Parameters) extends CoreBundle()(p) {
   val cacheable = Bool(OUTPUT)
 }
 
-class TLB(implicit edge: TLEdgeOut, val p: Parameters) extends Module with HasTLBParameters {
+class TLB(entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(p) {
   val io = new Bundle {
     val req = Decoupled(new TLBReq).flip
     val resp = new TLBResp
     val ptw = new TLBPTWIO
   }
+  val cacheBlockBytes = p(CacheBlockBytes)
+  val camAddrBits = log2Ceil(entries)
+  val camTagBits = asIdBits + vpnBits
 
   val valid = Reg(init = UInt(0, entries))
   val ppns = Reg(Vec(entries, UInt(width = ppnBits)))
@@ -182,7 +179,7 @@ class TLB(implicit edge: TLEdgeOut, val p: Parameters) extends Module with HasTL
   }
 }
 
-class DecoupledTLB(implicit edge: TLEdgeOut, p: Parameters) extends Module {
+class DecoupledTLB(entries: Int)(implicit edge: TLEdgeOut, p: Parameters) extends Module {
   val io = new Bundle {
     val req = Decoupled(new TLBReq).flip
     val resp = Decoupled(new TLBResp)
@@ -191,7 +188,7 @@ class DecoupledTLB(implicit edge: TLEdgeOut, p: Parameters) extends Module {
 
   val req = Reg(new TLBReq)
   val resp = Reg(new TLBResp)
-  val tlb = Module(new TLB)
+  val tlb = Module(new TLB(entries))
 
   val s_idle :: s_tlb_req :: s_tlb_resp :: s_done :: Nil = Enum(Bits(), 4)
   val state = Reg(init = s_idle)

src/main/scala/rocket/Tile.scala

@@ -7,12 +7,23 @@ import Chisel._
 import config._
 import coreplex._
 import diplomacy._
-import uncore.converters._
+import tile._
 import uncore.devices._
 import uncore.tilelink2._
 import util._
 
-class RocketTile(val c: RocketConfig)(implicit p: Parameters) extends BaseTile()(p)
+case class RocketTileParams(
+    core: RocketCoreParams = RocketCoreParams(),
+    icache: Option[ICacheParams] = Some(ICacheParams()),
+    dcache: Option[DCacheParams] = Some(DCacheParams()),
+    rocc: Seq[RoCCParams] = Nil,
+    btb: Option[BTBParams] = Some(BTBParams()),
+    dataScratchpadBytes: Int = 0) extends TileParams {
+  require(icache.isDefined)
+  require(dcache.isDefined)
+}
+  
+class RocketTile(val rocketParams: RocketTileParams)(implicit p: Parameters) extends BaseTile(rocketParams)(p)
     with CanHaveLegacyRoccs  // implies CanHaveSharedFPU with CanHavePTW with HasHellaCache
     with CanHaveScratchpad { // implies CanHavePTW with HasHellaCache with HasICacheFrontend
 
@@ -28,7 +39,7 @@ class RocketTileModule(outer: RocketTile) extends BaseTileModule(outer, () => ne
     with CanHaveLegacyRoccsModule
     with CanHaveScratchpadModule {
 
-  val core = Module(p(BuildCore)(outer.c, outer.p))
+  val core = Module(p(BuildCore)(outer.p))
   core.io.interrupts := io.interrupts
   core.io.hartid := io.hartid
   outer.frontend.module.io.cpu <> core.io.imem
@@ -44,25 +55,34 @@ class RocketTileModule(outer: RocketTile) extends BaseTileModule(outer, () => ne
     core.io.rocc.interrupt := lr.module.io.core.interrupt
   }
 
+  // TODO eliminate this redundancy
+  val h = dcachePorts.size
+  val c = core.dcacheArbPorts
+  val o = outer.nDCachePorts
+  require(h == c, s"port list size was $h, core expected $c")
+  require(h == o, s"port list size was $h, outer counted $o")
   // TODO figure out how to move the below into their respective mix-ins
-  require(dcachePorts.size == core.dcacheArbPorts)
   dcacheArb.io.requestor <> dcachePorts
   ptwOpt foreach { ptw => ptw.io.requestor <> ptwPorts }
 }
 
-class AsyncRocketTile(c: RocketConfig)(implicit p: Parameters) extends LazyModule {
+class AsyncRocketTile(rtp: RocketTileParams)(implicit p: Parameters) extends LazyModule {
-  val rocket = LazyModule(new RocketTile(c))
+  val rocket = LazyModule(new RocketTile(rtp))
 
-  val masterNodes = rocket.masterNodes.map(_ => TLAsyncOutputNode())
+  val masterNode = TLAsyncOutputNode()
-  val slaveNode = rocket.slaveNode.map(_ => TLAsyncInputNode())
+  val source = LazyModule(new TLAsyncCrossingSource)
+  source.node :=* rocket.masterNode
+  masterNode :=* source.node
 
-  (rocket.masterNodes zip masterNodes) foreach { case (r,n) => n := TLAsyncCrossingSource()(r) }
+  val slaveNode = TLAsyncInputNode()
-  (rocket.slaveNode zip slaveNode) foreach { case (r,n) => r := TLAsyncCrossingSink()(n) }
+  val sink = LazyModule(new TLAsyncCrossingSink)
+  rocket.slaveNode :*= sink.node
+  sink.node :*= slaveNode
 
   lazy val module = new LazyModuleImp(this) {
     val io = new Bundle {
-      val master = masterNodes.head.bundleOut // TODO fix after Chisel #366
+      val master = masterNode.bundleOut
-      val slave = slaveNode.map(_.bundleIn)
+      val slave = slaveNode.bundleIn
       val hartid = UInt(INPUT, p(XLen))
       val interrupts = new TileInterrupts()(p).asInput
       val resetVector = UInt(INPUT, p(XLen))
@@ -74,19 +94,23 @@ class AsyncRocketTile(c: RocketConfig)(implicit p: Parameters) extends LazyModul
   }
 }
 
-class RationalRocketTile(c: RocketConfig)(implicit p: Parameters) extends LazyModule {
+class RationalRocketTile(rtp: RocketTileParams)(implicit p: Parameters) extends LazyModule {
-  val rocket = LazyModule(new RocketTile(c))
+  val rocket = LazyModule(new RocketTile(rtp))
 
-  val masterNodes = rocket.masterNodes.map(_ => TLRationalOutputNode())
+  val masterNode = TLRationalOutputNode()
-  val slaveNode = rocket.slaveNode.map(_ => TLRationalInputNode())
+  val source = LazyModule(new TLRationalCrossingSource)
+  source.node :=* rocket.masterNode
+  masterNode :=* source.node
 
-  (rocket.masterNodes zip masterNodes) foreach { case (r,n) => n := TLRationalCrossingSource()(r) }
+  val slaveNode = TLRationalInputNode()
-  (rocket.slaveNode zip slaveNode) foreach { case (r,n) => r := TLRationalCrossingSink()(n) }
+  val sink = LazyModule(new TLRationalCrossingSink)
+  rocket.slaveNode :*= sink.node
+  sink.node :*= slaveNode
 
   lazy val module = new LazyModuleImp(this) {
     val io = new Bundle {
-      val master = masterNodes.head.bundleOut // TODO fix after Chisel #366
+      val master = masterNode.bundleOut
-      val slave = slaveNode.map(_.bundleIn)
+      val slave = slaveNode.bundleIn
       val hartid = UInt(INPUT, p(XLen))
       val interrupts = new TileInterrupts()(p).asInput
       val resetVector = UInt(INPUT, p(XLen))

src/main/scala/rocketchip/Configs.scala

@@ -38,33 +38,21 @@ class BasePlatformConfig extends Config((site, here, up) => {
   case RTCPeriod => 100 // gives 10 MHz RTC assuming 1 GHz uncore clock
 })
 
+/** Actual elaboratable target Configs */
+
 class BaseConfig extends Config(new BaseCoreplexConfig ++ new BasePlatformConfig)
-class DefaultConfig extends Config(new WithBlockingL1 ++ new BaseConfig)
+class DefaultConfig extends Config(new WithBlockingL1 ++ new WithNBigCores(1) ++ new BaseConfig)
 
-class DefaultL2Config extends Config(new WithL2Cache ++ new BaseConfig)
+class DefaultL2Config extends Config(new WithL2Cache ++ new WithNBigCores(1) ++ new BaseConfig)
 class DefaultBufferlessConfig extends Config(
-  new WithBufferlessBroadcastHub ++ new BaseConfig)
+  new WithBufferlessBroadcastHub ++ new WithNBigCores(1) ++ new BaseConfig)
-
-class FPGAConfig extends Config ((site, here, up) => {
-  case NAcquireTransactors => 4
-})
 
+class FPGAConfig extends Config(Parameters.empty)
 class DefaultFPGAConfig extends Config(new FPGAConfig ++ new BaseConfig)
 class DefaultL2FPGAConfig extends Config(
   new WithL2Capacity(64) ++ new WithL2Cache ++ new DefaultFPGAConfig)
 
-class WithNMemoryChannels(n: Int) extends Config((site, here, up) => {
+class DefaultSmallConfig extends Config(new WithNSmallCores(1) ++ new BaseConfig)
-  case BankedL2Config => up(BankedL2Config, site).copy(nMemoryChannels = n)
-})
-
-class WithExtMemSize(n: Long) extends Config((site, here, up) => {
-  case ExtMem => up(ExtMem, site).copy(size = n)
-})
-
-class WithScratchpads extends Config(new WithNMemoryChannels(0) ++ new WithDataScratchpad(16384))
-
-class DefaultFPGASmallConfig extends Config(new WithSmallCores ++ new DefaultFPGAConfig)
-class DefaultSmallConfig extends Config(new WithSmallCores ++ new BaseConfig)
 class DefaultRV32Config extends Config(new WithRV32 ++ new DefaultConfig)
 
 class DualBankConfig extends Config(
@@ -83,12 +71,7 @@ class DualChannelDualBankL2Config extends Config(
   new WithNMemoryChannels(2) ++ new WithNBanksPerMemChannel(2) ++
   new WithL2Cache ++ new BaseConfig)
 
-class RoccExampleConfig extends Config(new WithRoccExample ++ new BaseConfig)
+class RoccExampleConfig extends Config(new WithRoccExample ++ new DefaultConfig)
-
-class WithEdgeDataBits(dataBits: Int) extends Config((site, here, up) => {
-  case ExtMem => up(ExtMem, site).copy(beatBytes = dataBits/8)
-  case ZeroConfig => up(ZeroConfig, site).copy(beatBytes = dataBits/8)
-})
 
 class Edge128BitConfig extends Config(
   new WithEdgeDataBits(128) ++ new BaseConfig)
@@ -107,12 +90,22 @@ class OctoChannelBenchmarkConfig extends Config(new WithNMemoryChannels(8) ++ ne
 class EightChannelConfig extends Config(new WithNMemoryChannels(8) ++ new BaseConfig)
 
 class DualCoreConfig extends Config(
-  new WithNCores(2) ++ new WithL2Cache ++ new BaseConfig)
+  new WithNBigCores(2) ++ new WithL2Cache ++ new BaseConfig)
+class HeterogeneousDualCoreConfig extends Config(
+  new WithNSmallCores(1) ++ new WithNBigCores(1) ++ new WithL2Cache ++ new BaseConfig)
 
 class TinyConfig extends Config(
-  new WithScratchpads ++
+  new WithScratchpad ++
-  new WithSmallCores ++ new WithRV32 ++
+  new WithRV32 ++
-  new WithStatelessBridge ++ new BaseConfig)
+  new WithStatelessBridge ++
+  new WithNSmallCores(1) ++ new BaseConfig)
+
+/* Composable partial function Configs to set individual parameters */
+
+class WithEdgeDataBits(dataBits: Int) extends Config((site, here, up) => {
+  case ExtMem => up(ExtMem, site).copy(beatBytes = dataBits/8)
+  case ZeroConfig => up(ZeroConfig, site).copy(beatBytes = dataBits/8)
+})
 
 class WithJtagDTM extends Config ((site, here, up) => {
   case IncludeJtagDTM => true
@@ -134,10 +127,18 @@ class WithNExtTopInterrupts(nExtInts: Int) extends Config((site, here, up) => {
   case NExtTopInterrupts => nExtInts
 })
 
-class WithNBreakpoints(hwbp: Int) extends Config ((site, here, up) => {
-  case NBreakpoints => hwbp
-})
-
 class WithRTCPeriod(nCycles: Int) extends Config((site, here, up) => {
   case RTCPeriod => nCycles
 })
+
+class WithNMemoryChannels(n: Int) extends Config((site, here, up) => {
+  case BankedL2Config => up(BankedL2Config, site).copy(nMemoryChannels = n)
+})
+
+class WithExtMemSize(n: Long) extends Config((site, here, up) => {
+  case ExtMem => up(ExtMem, site).copy(size = n)
+})
+
+class WithScratchpad extends Config(new WithNMemoryChannels(0) ++ new WithDataScratchpad(16384))
+
+class DefaultFPGASmallConfig extends Config(new WithNSmallCores(1) ++ new DefaultFPGAConfig)

src/main/scala/rocketchip/Generator.scala

@@ -2,7 +2,9 @@
 
 package rocketchip
 
-import rocket.{XLen, UseVM, UseAtomics, UseCompressed, FPUKey}
+import tile.XLen
+import coreplex.RocketTilesKey
+
 import scala.collection.mutable.LinkedHashSet
 
 /** A Generator for platforms containing Rocket Coreplexes */
@@ -47,9 +49,11 @@ object Generator extends util.GeneratorApp {
   override def addTestSuites {
     import DefaultTestSuites._
     val xlen = params(XLen)
-    val vm = params(UseVM)
+    // TODO: for now only generate tests for the first core in the first coreplex
+    val coreParams = params(RocketTilesKey).head.core
+    val vm = coreParams.useVM
     val env = if (vm) List("p","v") else List("p")
-    params(FPUKey) foreach { case cfg =>
+    coreParams.fpu foreach { case cfg =>
       if (xlen == 32) {
         TestGeneration.addSuites(env.map(rv32ufNoDiv))
       } else {
@@ -62,8 +66,8 @@ object Generator extends util.GeneratorApp {
         }
       }
     }
-    if (params(UseAtomics))    TestGeneration.addSuites(env.map(if (xlen == 64) rv64ua else rv32ua))
+    if (coreParams.useAtomics)    TestGeneration.addSuites(env.map(if (xlen == 64) rv64ua else rv32ua))
-    if (params(UseCompressed)) TestGeneration.addSuites(env.map(if (xlen == 64) rv64uc else rv32uc))
+    if (coreParams.useCompressed) TestGeneration.addSuites(env.map(if (xlen == 64) rv64uc else rv32uc))
     val (rvi, rvu) =
       if (xlen == 64) ((if (vm) rv64i else rv64pi), rv64u)
       else            ((if (vm) rv32i else rv32pi), rv32u)

src/main/scala/rocketchip/Periphery.scala

@@ -4,20 +4,16 @@ package rocketchip
 
 import Chisel._
 import config._
-import junctions._
+import coreplex._
-import junctions.NastiConstants._
 import diplomacy._
-import uncore.tilelink._
+import tile.XLen
 import uncore.tilelink2._
 import uncore.axi4._
 import uncore.converters._
 import uncore.devices._
-import uncore.agents._
 import uncore.util._
 import util._
-import rocket.XLen
 import scala.math.max
-import coreplex._
 
 /** Specifies the size of external memory */
 case class MasterConfig(base: Long, size: Long, beatBytes: Int, idBits: Int)

src/main/scala/rocketchip/Utils.scala

@@ -6,7 +6,7 @@ import config._
 import junctions._
 import diplomacy._
 import uncore.devices._
-import rocket._
+import tile.XLen
 import coreplex._
 import uncore.tilelink2._
 import util._
@@ -59,14 +59,15 @@ object GenerateConfigString {
     res append plic.globalConfigString
     res append clint.globalConfigString
     res append  "core {\n"
-    for (i <- 0 until c.nTiles) { // TODO heterogeneous tiles
+    c.tilesParams.zipWithIndex.map { case(t, i) => 
       val isa = {
-        val m = if (p(MulDivKey).nonEmpty) "m" else ""
+        val m = if (t.core.mulDiv.nonEmpty) "m" else ""
-        val a = if (p(UseAtomics)) "a" else ""
+        val a = if (t.core.useAtomics) "a" else ""
-        val f = if (p(FPUKey).nonEmpty) "f" else ""
+        val f = if (t.core.fpu.nonEmpty) "f" else ""
-        val d = if (p(FPUKey).nonEmpty && p(XLen) > 32) "d" else ""
+        val d = if (t.core.fpu.nonEmpty && p(XLen) > 32) "d" else ""
-        val s = if (c.hasSupervisor) "s" else ""
+        val c = if (t.core.useCompressed) "c" else ""
-        s"rv${p(XLen)}i$m$a$f$d$s"
+        val s = if (t.core.useVM) "s" else ""
+        s"rv${p(XLen)}i$m$a$f$d$c$s"
       }
       res append s"  $i {\n"
       res append  "    0 {\n"

src/main/scala/tile/BaseTile.scala

@@ -0,0 +1,81 @@
+// See LICENSE.SiFive for license details.
+
+package tile
+
+import Chisel._
+import config._
+import diplomacy._
+import rocket._
+import uncore.tilelink2._
+import util._
+
+case object SharedMemoryTLEdge extends Field[TLEdgeOut]
+case object TileKey extends Field[TileParams]
+
+trait TileParams {
+  val core: CoreParams
+  val icache: Option[ICacheParams]
+  val dcache: Option[DCacheParams]
+  val rocc: Seq[RoCCParams]
+  val btb: Option[BTBParams]
+  val dataScratchpadBytes: Int
+}
+
+trait HasTileParameters {
+  implicit val p: Parameters
+  val tileParams: TileParams = p(TileKey)
+
+  val usingVM = tileParams.core.useVM
+  val usingUser = tileParams.core.useUser || usingVM
+  val usingDebug = tileParams.core.useDebug
+  val usingRoCC = !tileParams.rocc.isEmpty
+  val usingBTB = tileParams.btb.isDefined && tileParams.btb.get.nEntries > 0
+  val usingPTW = usingVM
+  val usingDataScratchpad = tileParams.dcache.isDefined && tileParams.dataScratchpadBytes > 0
+
+  def dcacheArbPorts = 1 + usingVM.toInt + usingDataScratchpad.toInt + tileParams.rocc.size
+}
+
+abstract class BareTile(implicit p: Parameters) extends LazyModule
+
+abstract class BareTileBundle[+L <: BareTile](_outer: L) extends GenericParameterizedBundle(_outer) {
+  val outer = _outer
+  implicit val p = outer.p
+}
+
+abstract class BareTileModule[+L <: BareTile, +B <: BareTileBundle[L]](_outer: L, _io: () => B) extends LazyModuleImp(_outer) {
+  val outer = _outer
+  val io = _io ()
+}
+
+// Uses TileLink master port to connect caches and accelerators to the coreplex
+trait HasTileLinkMasterPort extends HasTileParameters {
+  implicit val p: Parameters
+  val module: HasTileLinkMasterPortModule
+  val masterNode = TLOutputNode()
+}
+
+trait HasTileLinkMasterPortBundle {
+  val outer: HasTileLinkMasterPort
+  val master = outer.masterNode.bundleOut
+}
+
+trait HasTileLinkMasterPortModule {
+  val outer: HasTileLinkMasterPort
+  val io: HasTileLinkMasterPortBundle
+}
+
+abstract class BaseTile(tileParams: TileParams)(implicit p: Parameters) extends BareTile
+    with HasTileLinkMasterPort {
+  override lazy val module = new BaseTileModule(this, () => new BaseTileBundle(this))
+}
+
+class BaseTileBundle[+L <: BaseTile](_outer: L) extends BareTileBundle(_outer)
+    with HasTileLinkMasterPortBundle {
+  val hartid = UInt(INPUT, p(XLen))
+  val interrupts = new TileInterrupts()(p).asInput
+  val resetVector = UInt(INPUT, p(XLen))
+}
+
+class BaseTileModule[+L <: BaseTile, +B <: BaseTileBundle[L]](_outer: L, _io: () => B) extends BareTileModule(_outer, _io)
+    with HasTileLinkMasterPortModule

src/main/scala/tile/Core.scala

@@ -0,0 +1,88 @@
+// See LICENSE.SiFive for license details.
+
+package tile
+
+import Chisel._
+import config._
+import rocket._
+import util._
+
+case object BuildCore extends Field[Parameters => CoreModule with HasCoreIO]
+case object XLen extends Field[Int]
+
+// These parameters can be varied per-core
+trait CoreParams {
+  val useVM: Boolean
+  val useUser: Boolean
+  val useDebug: Boolean
+  val useAtomics: Boolean
+  val useCompressed: Boolean
+  val mulDiv: Option[MulDivParams]
+  val fpu: Option[FPUParams]
+  val fetchWidth: Int
+  val decodeWidth: Int
+  val retireWidth: Int
+  val instBits: Int
+}
+
+trait HasCoreParameters extends HasTileParameters {
+  val coreParams: CoreParams = tileParams.core
+
+  val xLen = p(XLen)
+  val fLen = xLen // TODO relax this
+  require(xLen == 32 || xLen == 64)
+
+  val usingMulDiv = coreParams.mulDiv.nonEmpty
+  val usingFPU = coreParams.fpu.nonEmpty
+  val usingAtomics = coreParams.useAtomics
+  val usingCompressed = coreParams.useCompressed
+
+  val retireWidth = coreParams.retireWidth
+  val fetchWidth = coreParams.fetchWidth
+  val decodeWidth = coreParams.decodeWidth
+
+  val coreInstBits = coreParams.instBits
+  val coreInstBytes = coreInstBits/8
+  val coreDataBits = xLen
+  val coreDataBytes = coreDataBits/8
+
+  val coreDCacheReqTagBits = 6
+  val dcacheReqTagBits = coreDCacheReqTagBits + log2Ceil(dcacheArbPorts)
+
+  def pgIdxBits = 12
+  def pgLevelBits = 10 - log2Ceil(xLen / 32)
+  def vaddrBits = pgIdxBits + pgLevels * pgLevelBits
+  val paddrBits = 32//p(PAddrBits)
+  def ppnBits = paddrBits - pgIdxBits
+  def vpnBits = vaddrBits - pgIdxBits
+  val pgLevels = p(PgLevels)
+  val asIdBits = p(ASIdBits)
+  val vpnBitsExtended = vpnBits + (vaddrBits < xLen).toInt
+  val vaddrBitsExtended = vpnBitsExtended + pgIdxBits
+  val coreMaxAddrBits = paddrBits max vaddrBitsExtended
+  val maxPAddrBits = xLen match { case 32 => 34; case 64 => 50 }
+  require(paddrBits <= maxPAddrBits)
+
+  // Print out log of committed instructions and their writeback values.
+  // Requires post-processing due to out-of-order writebacks.
+  val enableCommitLog = false
+}
+
+abstract class CoreModule(implicit val p: Parameters) extends Module
+  with HasCoreParameters
+
+abstract class CoreBundle(implicit val p: Parameters) extends ParameterizedBundle()(p)
+  with HasCoreParameters
+
+trait HasCoreIO {
+  implicit val p: Parameters
+  val io = new Bundle {
+    val interrupts = new TileInterrupts().asInput
+    val hartid = UInt(INPUT, p(XLen))
+    val imem  = new FrontendIO()(p)
+    val dmem = new HellaCacheIO()(p)
+    val ptw = new DatapathPTWIO().flip
+    val fpu = new FPUCoreIO().flip
+    val rocc = new RoCCCoreIO().flip
+  }
+}

src/main/scala/tile/FPU.scala

@@ -1,17 +1,18 @@
 // See LICENSE.Berkeley for license details.
 // See LICENSE.SiFive for license details.
 
-package rocket
+package tile
 
 import Chisel._
-import Instructions._
-import util._
 import Chisel.ImplicitConversions._
 import FPConstants._
+import rocket.DecodeLogic
+import rocket.Instructions._
 import uncore.constants.MemoryOpConstants._
 import config._
+import util._
 
-case class FPUConfig(
+case class FPUParams(
   divSqrt: Boolean = true,
   sfmaLatency: Int = 3,
   dfmaLatency: Int = 4
@@ -496,7 +497,7 @@ class FPUFMAPipe(val latency: Int, expWidth: Int, sigWidth: Int)(implicit p: Par
   io.out := Pipe(valid, res, latency-1)
 }
 
-class FPU(cfg: FPUConfig)(implicit p: Parameters) extends FPUModule()(p) {
+class FPU(cfg: FPUParams)(implicit p: Parameters) extends FPUModule()(p) {
   val io = new FPUIO
 
   val ex_reg_valid = Reg(next=io.valid, init=Bool(false))
@@ -742,12 +743,10 @@ class FPU(cfg: FPUConfig)(implicit p: Parameters) extends FPUModule()(p) {
 }
 
 /** Mix-ins for constructing tiles that may have an FPU external to the core pipeline */
-trait CanHaveSharedFPU {
+trait CanHaveSharedFPU extends HasTileParameters
-  implicit val p: Parameters
-}
 
 trait CanHaveSharedFPUModule {
   val outer: CanHaveSharedFPU
-  val fpuOpt = outer.p(FPUKey).map(cfg => Module(new FPU(cfg)(outer.p)))
+  val fpuOpt = outer.tileParams.core.fpu.map(params => Module(new FPU(params)(outer.p)))
   // TODO fpArb could go here instead of inside LegacyRoccComplex
 }

src/main/scala/tile/Interrupts.scala

@@ -0,0 +1,15 @@
+// See LICENSE.SiFive for license details.
+
+package tile
+
+import Chisel._
+import config.Parameters
+import util._
+
+class TileInterrupts(implicit p: Parameters) extends CoreBundle()(p) {
+  val debug = Bool()
+  val mtip = Bool()
+  val msip = Bool()
+  val meip = Bool()
+  val seip = usingVM.option(Bool())
+}

src/main/scala/tile/L1Cache.scala

@@ -0,0 +1,52 @@
+// See LICENSE.SiFive for license details.
+
+package tile
+
+import Chisel._
+import config.{Parameters, Field}
+import coreplex.CacheBlockBytes
+import rocket.PAddrBits
+import uncore.tilelink2.ClientMetadata
+import uncore.util.{Code, IdentityCode}
+import util.ParameterizedBundle
+
+trait L1CacheParams {
+  def nSets:         Int
+  def nWays:         Int
+  def rowBits:       Int
+  def nTLBEntries:   Int
+  def splitMetadata: Boolean
+  def ecc:           Option[Code]
+}
+
+trait HasL1CacheParameters {
+  implicit val p: Parameters
+  val cacheParams: L1CacheParams
+
+  def cacheBlockBytes = p(CacheBlockBytes)
+  def lgCacheBlockBytes = log2Up(cacheBlockBytes)
+  def nSets = cacheParams.nSets
+  def blockOffBits = lgCacheBlockBytes
+  def idxBits = log2Up(cacheParams.nSets)
+  def untagBits = blockOffBits + idxBits
+  def tagBits = p(PAddrBits) - untagBits
+  def nWays = cacheParams.nWays
+  def wayBits = log2Up(nWays)
+  def isDM = nWays == 1
+  def rowBits = cacheParams.rowBits
+  def rowBytes = rowBits/8
+  def rowOffBits = log2Up(rowBytes)
+  def code = cacheParams.ecc.getOrElse(new IdentityCode)
+  def nTLBEntries = cacheParams.nTLBEntries
+  def hasSplitMetadata = cacheParams.splitMetadata
+
+  def cacheDataBits = p(SharedMemoryTLEdge).bundle.dataBits
+  def cacheDataBeats = (cacheBlockBytes * 8) / cacheDataBits
+  def refillCycles = cacheDataBeats
+}
+
+abstract class L1CacheModule(implicit val p: Parameters) extends Module
+  with HasL1CacheParameters
+
+abstract class L1CacheBundle(implicit val p: Parameters) extends ParameterizedBundle()(p)
+  with HasL1CacheParameters

src/main/scala/tile/LegacyRoCC.scala

@@ -1,13 +1,14 @@
 // See LICENSE.Berkeley for license details.
 // See LICENSE.SiFive for license details.
 
-package rocket
+package tile
 
 import Chisel._
 import Chisel.ImplicitConversions._
 import config._
 import coreplex._
 import diplomacy._
+import rocket._
 import uncore.constants._
 import uncore.agents._
 import uncore.coherence._
@@ -20,9 +21,9 @@ import util._
 case object RoccMaxTaggedMemXacts extends Field[Int]
 case object RoccNMemChannels extends Field[Int]
 case object RoccNPTWPorts extends Field[Int]
-case object BuildRoCC extends Field[Seq[RoccParameters]]
+case object BuildRoCC extends Field[Seq[RoCCParams]]
 
-trait CanHaveLegacyRoccs extends CanHaveSharedFPU with CanHavePTW with TileNetwork {
+trait CanHaveLegacyRoccs extends CanHaveSharedFPU with CanHavePTW with HasTileLinkMasterPort {
   val module: CanHaveLegacyRoccsModule
   val legacyRocc = if (p(BuildRoCC).isEmpty) None
     else Some(LazyModule(new LegacyRoccComplex()(p.alter { (site, here, up) => {
@@ -38,7 +39,7 @@ trait CanHaveLegacyRoccs extends CanHaveSharedFPU with CanHavePTW with TileNetwo
             nCachingClients = 1,
             nCachelessClients = 1,
             maxClientXacts = List(
-                site(DCacheKey).nMSHRs + 1 /* IOMSHR */,
+                tileParams.dcache.get.nMSHRs + 1 /* IOMSHR */,
                 if (site(BuildRoCC).isEmpty) 1 else site(RoccMaxTaggedMemXacts)).max,
             maxClientsPerPort = if (site(BuildRoCC).isEmpty) 1 else 2,
             maxManagerXacts = 8,
@@ -46,15 +47,16 @@ trait CanHaveLegacyRoccs extends CanHaveSharedFPU with CanHavePTW with TileNetwo
             dataBits = site(CacheBlockBytes)*8)
     }})))
 
-  // TODO for now, all legacy rocc mem ports mapped to one external node
   legacyRocc foreach { lr =>
-    lr.masterNodes.foreach { l1backend.node := _ }
+    masterNode := lr.masterNode
     nPTWPorts += lr.nPTWPorts
     nDCachePorts += lr.nRocc
   }
 }
 
-trait CanHaveLegacyRoccsModule extends CanHaveSharedFPUModule with CanHavePTWModule with TileNetworkModule {
+trait CanHaveLegacyRoccsModule extends CanHaveSharedFPUModule
+    with CanHavePTWModule
+    with HasTileLinkMasterPortModule {
   val outer: CanHaveLegacyRoccs
 
   fpuOpt foreach { fpu =>
@@ -84,13 +86,13 @@ class LegacyRoccComplex(implicit p: Parameters) extends LazyModule {
   val nPTWPorts = buildRocc.map(_.nPTWPorts).sum
   val roccOpcodes = buildRocc.map(_.opcodes)
 
+  val masterNode = TLOutputNode()
   val legacies = List.fill(nMemChannels) { LazyModule(new TLLegacy()(p.alterPartial({ case PAddrBits => 32 }))) }
-  val masterNodes = legacies.map(_ => TLOutputNode())
+  legacies.foreach { leg => masterNode := TLHintHandler()(leg.node) }
-  legacies.zip(masterNodes).foreach { case(l,m) => m := TLHintHandler()(l.node) }
 
   lazy val module = new LazyModuleImp(this) with HasCoreParameters {
     val io = new Bundle {
-      val tl = masterNodes.map(_.bundleOut)
+      val tl = masterNode.bundleOut
       val dcache = Vec(nRocc, new HellaCacheIO)
       val fpu = new Bundle {
         val cp_req = Decoupled(new FPInput())
@@ -117,7 +119,7 @@ class LegacyRoccComplex(implicit p: Parameters) extends LazyModule {
         case RoccNMemChannels => accelParams.nMemChannels
         case RoccNPTWPorts => accelParams.nPTWPorts
       }))
-      val dcIF = Module(new SimpleHellaCacheIF()(p.alterPartial({ case CacheName => CacheName("L1D") })))
+      val dcIF = Module(new SimpleHellaCacheIF)
       rocc.io.cmd <> cmdRouter.io.out(i)
       rocc.io.exception := io.core.exception
       dcIF.io.requestor <> rocc.io.mem
@@ -151,7 +153,7 @@ class LegacyRoccComplex(implicit p: Parameters) extends LazyModule {
   }
 }
 
-case class RoccParameters(
+case class RoCCParams(
   opcodes: OpcodeSet,
   generator: Parameters => RoCC,
   nMemChannels: Int = 0,

src/main/scala/tile/Package.scala

@@ -0,0 +1,3 @@
+// See LICENSE.SiFive for license details.
+
+package object tile extends rocket.constants.ScalarOpConstants

src/main/scala/uncore/agents/L2Cache.scala

@@ -1,980 +0,0 @@
-// See LICENSE.Berkeley for license details.
-// See LICENSE.SiFive for license details.
-
-package uncore.agents
-
-import Chisel._
-import scala.reflect.ClassTag
-import rocket.PAddrBits
-import uncore.coherence._
-import uncore.tilelink._
-import uncore.constants._
-import uncore.util._
-import util._
-import config.{Parameters, Field}
-
-case object CacheId extends Field[Int]
-case object L2DirectoryRepresentation extends Field[DirectoryRepresentation]
-
-trait HasOuterCacheParameters extends HasCacheParameters with HasCoherenceAgentParameters {
-  val cacheId = p(CacheId)
-  val idxLSB = cacheIdBits
-  val idxMSB = idxLSB + idxBits - 1
-  val tagLSB = idxLSB + idxBits
-  val tagMSB = tagLSB + tagBits - 1
-
-  def inSameSet(block_a: HasCacheBlockAddress, block_b: HasCacheBlockAddress): Bool =
-    inSameSet(block_a, block_b.addr_block)
-
-  def inSameSet(block: HasCacheBlockAddress, addr: UInt): Bool =
-    inSet(block, addr(idxMSB, idxLSB))
-
-  def inSet(block: HasCacheBlockAddress, idx: UInt): Bool =
-    block.addr_block(idxMSB,idxLSB) === idx
-
-  def haveSameTag(block: HasCacheBlockAddress, addr: UInt): Bool =
-    hasTag(block, addr(tagMSB, tagLSB))
-
-  def hasTag(block: HasCacheBlockAddress, tag: UInt): Bool =
-    block.addr_block(tagMSB, tagLSB) === tag
-
-  def isSameBlock(block: HasCacheBlockAddress, tag: UInt, idx: UInt) =
-    hasTag(block, tag) && inSet(block, idx)
-
-  //val blockAddrBits = p(TLBlockAddrBits)
-  val refillCyclesPerBeat = outerDataBits/rowBits
-  val refillCycles = refillCyclesPerBeat*outerDataBeats
-  val internalDataBeats = p(CacheBlockBytes)*8/rowBits
-  require(refillCyclesPerBeat == 1)
-  val amoAluOperandBits = p(AmoAluOperandBits)
-  require(amoAluOperandBits <= innerDataBits)
-  require(rowBits == innerDataBits) // TODO: relax this by improving s_data_* states
-  val nSecondaryMisses = p(NSecondaryMisses)
-  val isLastLevelCache = true
-  val alwaysWriteFullBeat = !p(ECCCode).isEmpty
-}
-
-abstract class L2HellaCacheModule(implicit val p: Parameters) extends Module
-    with HasOuterCacheParameters {
-  def doInternalOutputArbitration[T <: Data : ClassTag](
-      out: DecoupledIO[T],
-      ins: Seq[DecoupledIO[T]],
-      block_transfer: T => Bool = (t: T) => Bool(false)) {
-    val arb = Module(new RRArbiter(out.bits, ins.size))
-    out.valid := arb.io.out.valid && !block_transfer(arb.io.out.bits)
-    out.bits := arb.io.out.bits
-    arb.io.out.ready := out.ready && !block_transfer(arb.io.out.bits)
-    arb.io.in <> ins
-  }
-
-  def doInternalInputRouting[T <: Bundle with HasL2Id](in: ValidIO[T], outs: Seq[ValidIO[T]]) {
-    outs.map(_.bits := in.bits)
-    outs.zipWithIndex.map { case (o,i) => o.valid := in.valid && in.bits.id === UInt(i) }
-  }
-}
-
-abstract class L2HellaCacheBundle(implicit val p: Parameters) extends ParameterizedBundle()(p)
-  with HasOuterCacheParameters
-
-trait HasL2Id extends HasCoherenceAgentParameters {
-  val id = UInt(width  = log2Up(nTransactors + 1))
-}
-
-trait HasL2InternalRequestState extends HasOuterCacheParameters {
-  val tag_match = Bool()
-  val meta = new L2Metadata
-  val way_en = Bits(width = nWays)
-}
-
-trait HasL2BeatAddr extends HasOuterCacheParameters {
-  val addr_beat = UInt(width = log2Up(refillCycles))
-}
-
-trait HasL2Data extends HasOuterCacheParameters
-    with HasL2BeatAddr {
-  val data = UInt(width = rowBits)
-  def hasData(dummy: Int = 0) = Bool(true)
-  def hasMultibeatData(dummy: Int = 0) = Bool(refillCycles > 1)
-}
-
-class L2Metadata(implicit p: Parameters) extends Metadata()(p) with HasOuterCacheParameters {
-  val coh = new HierarchicalMetadata
-}
-
-object L2Metadata {
-  def apply(tag: Bits, coh: HierarchicalMetadata)
-      (implicit p: Parameters): L2Metadata = {
-    val meta = Wire(new L2Metadata)
-    meta.tag := tag
-    meta.coh := coh
-    meta
-  }
-
-  def apply(
-        tag: Bits,
-        inner: ManagerMetadata,
-        outer: ClientMetadata)(implicit p: Parameters): L2Metadata = {
-    val coh = Wire(new HierarchicalMetadata)
-    coh.inner := inner
-    coh.outer := outer
-    apply(tag, coh)
-  }
-}
-
-class L2MetaReadReq(implicit p: Parameters) extends MetaReadReq()(p) with HasL2Id {
-  val tag = Bits(width = tagBits)
-}
-
-class L2MetaWriteReq(implicit p: Parameters) extends MetaWriteReq[L2Metadata](new L2Metadata)(p)
-    with HasL2Id {
-  override def cloneType = new L2MetaWriteReq().asInstanceOf[this.type]
-}
-
-class L2MetaResp(implicit p: Parameters) extends L2HellaCacheBundle()(p)
-  with HasL2Id 
-  with HasL2InternalRequestState
-
-trait HasL2MetaReadIO extends HasOuterCacheParameters {
-  val read = Decoupled(new L2MetaReadReq)
-  val resp = Valid(new L2MetaResp).flip
-}
-
-trait HasL2MetaWriteIO extends HasOuterCacheParameters {
-  val write = Decoupled(new L2MetaWriteReq)
-}
-
-class L2MetaRWIO(implicit p: Parameters) extends L2HellaCacheBundle()(p)
-  with HasL2MetaReadIO
-  with HasL2MetaWriteIO
-
-class L2MetaReadOnlyIO(implicit p: Parameters) extends L2HellaCacheBundle()(p)
-  with HasL2MetaReadIO
-
-trait HasL2MetaRWIO extends HasOuterCacheParameters {
-  val meta = new L2MetaRWIO
-}
-
-class L2MetadataArray(implicit p: Parameters) extends L2HellaCacheModule()(p) {
-  val io = new L2MetaRWIO().flip
-
-  def onReset = L2Metadata(UInt(0), HierarchicalMetadata.onReset)
-  val meta = Module(new MetadataArray(onReset _))
-  meta.io.read <> io.read
-  meta.io.write <> io.write
-  val way_en_1h = UInt((BigInt(1) << nWays) - 1)
-  val s1_way_en_1h = RegEnable(way_en_1h, io.read.valid)
-  meta.io.read.bits.way_en := way_en_1h
-
-  val s1_tag = RegEnable(io.read.bits.tag, io.read.valid)
-  val s1_id = RegEnable(io.read.bits.id, io.read.valid)
-  def wayMap[T <: Data](f: Int => T) = Vec((0 until nWays).map(f))
-  val s1_clk_en = Reg(next = io.read.fire())
-  val s1_tag_eq_way = wayMap((w: Int) => meta.io.resp(w).tag === s1_tag)
-  val s1_tag_match_way = wayMap((w: Int) => s1_tag_eq_way(w) && meta.io.resp(w).coh.outer.isValid() && s1_way_en_1h(w).toBool).asUInt
-  val s1_idx = RegEnable(io.read.bits.idx, io.read.valid) // deal with stalls?
-  val s2_tag_match_way = RegEnable(s1_tag_match_way, s1_clk_en)
-  val s2_tag_match = s2_tag_match_way.orR
-  val s2_hit_coh = Mux1H(s2_tag_match_way, wayMap((w: Int) => RegEnable(meta.io.resp(w).coh, s1_clk_en)))
-
-  val replacer = p(L2Replacer)()
-  val s1_hit_way = Wire(Bits())
-  s1_hit_way := Bits(0)
-  (0 until nWays).foreach(i => when (s1_tag_match_way(i)) { s1_hit_way := Bits(i) })
-  replacer.access(io.read.bits.idx)
-  replacer.update(s1_clk_en, s1_tag_match_way.orR, s1_idx, s1_hit_way)
-
-  val s1_replaced_way_en = UIntToOH(replacer.way)
-  val s2_replaced_way_en = UIntToOH(RegEnable(replacer.way, s1_clk_en))
-  val s2_repl_meta = Mux1H(s2_replaced_way_en, wayMap((w: Int) => 
-    RegEnable(meta.io.resp(w), s1_clk_en && s1_replaced_way_en(w))).toSeq)
-
-  io.resp.valid := Reg(next = s1_clk_en)
-  io.resp.bits.id := RegEnable(s1_id, s1_clk_en)
-  io.resp.bits.tag_match := s2_tag_match
-  io.resp.bits.meta := Mux(s2_tag_match, 
-    L2Metadata(s2_repl_meta.tag, s2_hit_coh), 
-    s2_repl_meta)
-  io.resp.bits.way_en := Mux(s2_tag_match, s2_tag_match_way, s2_replaced_way_en)
-}
-
-class L2DataReadReq(implicit p: Parameters) extends L2HellaCacheBundle()(p)
-    with HasL2BeatAddr 
-    with HasL2Id {
-  val addr_idx = UInt(width = idxBits)
-  val way_en = Bits(width = nWays)
-}
-
-object L2DataReadReq {
-  def apply(
-        id: UInt,
-        way_en: UInt,
-        addr_idx: UInt,
-        addr_beat: UInt)(implicit p: Parameters) = {
-    val req = Wire(new L2DataReadReq)
-    req.id := id
-    req.way_en := way_en
-    req.addr_idx := addr_idx
-    req.addr_beat := addr_beat
-    req
-  }
-}
-
-class L2DataWriteReq(implicit p: Parameters) extends L2DataReadReq()(p)
-    with HasL2Data {
-  val wmask  = Bits(width = rowBits/8)
-}
-
-object L2DataWriteReq {
-  def apply(
-        id: UInt,
-        way_en: UInt,
-        addr_idx: UInt,
-        addr_beat: UInt,
-        wmask: UInt,
-        data: UInt)(implicit p: Parameters) = {
-    val req = Wire(new L2DataWriteReq)
-    req.id := id
-    req.way_en := way_en
-    req.addr_idx := addr_idx
-    req.addr_beat := addr_beat
-    req.wmask := wmask
-    req.data := data
-    req
-  }
-}
-
-class L2DataResp(implicit p: Parameters) extends L2HellaCacheBundle()(p)
-  with HasL2Id
-  with HasL2Data
-
-trait HasL2DataReadIO extends HasOuterCacheParameters { 
-  val read = Decoupled(new L2DataReadReq)
-  val resp = Valid(new L2DataResp).flip
-}
-
-trait HasL2DataWriteIO extends HasOuterCacheParameters { 
-  val write = Decoupled(new L2DataWriteReq)
-}
-
-class L2DataRWIO(implicit p: Parameters) extends L2HellaCacheBundle()(p)
-  with HasL2DataReadIO
-  with HasL2DataWriteIO
-
-trait HasL2DataRWIO extends HasOuterCacheParameters {
-  val data = new L2DataRWIO
-}
-
-class L2DataArray(delay: Int)(implicit p: Parameters) extends L2HellaCacheModule()(p) {
-  val io = new L2DataRWIO().flip
-
-  val array = SeqMem(nWays*nSets*refillCycles, Vec(rowBits/8, Bits(width=8)))
-  val ren = !io.write.valid && io.read.valid
-  val raddr = Cat(OHToUInt(io.read.bits.way_en), io.read.bits.addr_idx, io.read.bits.addr_beat)
-  val waddr = Cat(OHToUInt(io.write.bits.way_en), io.write.bits.addr_idx, io.write.bits.addr_beat)
-  val wdata = Vec.tabulate(rowBits/8)(i => io.write.bits.data(8*(i+1)-1,8*i))
-  val wmask = io.write.bits.wmask.toBools
-  when (io.write.valid) { array.write(waddr, wdata, wmask) }
-
-  val r_req = Pipe(io.read.fire(), io.read.bits)
-  io.resp := Pipe(r_req.valid, r_req.bits, delay)
-  io.resp.bits.data := Pipe(r_req.valid, array.read(raddr, ren).asUInt, delay).bits
-  io.read.ready := !io.write.valid
-  io.write.ready := Bool(true)
-}
-
-class L2HellaCacheBank(implicit p: Parameters) extends HierarchicalCoherenceAgent()(p)
-    with HasOuterCacheParameters {
-  require(isPow2(nSets))
-  require(isPow2(nWays)) 
-
-  val meta = Module(new L2MetadataArray) // TODO: add delay knob
-  val data = Module(new L2DataArray(1))
-  val tshrfile = Module(new TSHRFile)
-  io.inner <> tshrfile.io.inner
-  io.outer <> tshrfile.io.outer
-  tshrfile.io.incoherent <> io.incoherent
-  meta.io <> tshrfile.io.meta
-  data.io <> tshrfile.io.data
-
-  disconnectOuterProbeAndFinish()
-}
-
-class TSHRFileIO(implicit p: Parameters) extends HierarchicalTLIO()(p)
-  with HasL2MetaRWIO
-  with HasL2DataRWIO
-
-class TSHRFile(implicit p: Parameters) extends L2HellaCacheModule()(p)
-    with HasCoherenceAgentWiringHelpers {
-  val io = new TSHRFileIO
-
-  // Create TSHRs for outstanding transactions
-  val irelTrackerList =
-    (0 until nReleaseTransactors).map(id =>
-      Module(new CacheVoluntaryReleaseTracker(id)))
-  val iacqTrackerList = 
-    (nReleaseTransactors until nTransactors).map(id =>
-      Module(new CacheAcquireTracker(id)))
-  val trackerList = irelTrackerList ++ iacqTrackerList
-
-  // Don't allow a writeback request to go through if we are taking
-  // a voluntary release for the same block.
-  // The writeback can go forward once the voluntary release is handled
-  def writebackConflictsWithVolRelease(wb: L2WritebackReq): Bool =
-    irelTrackerList
-      .map(tracker =>
-          !tracker.io.alloc.idle &&
-          isSameBlock(tracker.io.alloc, wb.tag, wb.idx))
-      .reduce(_ || _) ||
-    (io.inner.release.valid &&
-      isSameBlock(io.inner.release.bits, wb.tag, wb.idx))
-
-  // WritebackUnit evicts data from L2, including invalidating L1s
-  val wb = Module(new L2WritebackUnit(nTransactors))
-  val trackerAndWbIOs = trackerList.map(_.io) :+ wb.io
-  doInternalOutputArbitration(wb.io.wb.req, trackerList.map(_.io.wb.req),
-    block_transfer = writebackConflictsWithVolRelease _)
-  doInternalInputRouting(wb.io.wb.resp, trackerList.map(_.io.wb.resp))
-
-  // Propagate incoherence flags
-  (trackerList.map(_.io.incoherent) :+ wb.io.incoherent) foreach { _ := io.incoherent }
-
-  // Handle acquire transaction initiation
-  val irel_vs_iacq_conflict =
-        io.inner.acquire.valid &&
-        io.inner.release.valid &&
-        inSameSet(io.inner.acquire.bits, io.inner.release.bits)
-  doInputRoutingWithAllocation(
-    in = io.inner.acquire,
-    outs = trackerList.map(_.io.inner.acquire),
-    allocs = trackerList.map(_.io.alloc.iacq),
-    allocOverride = Some(!irel_vs_iacq_conflict))
-
-  assert(PopCount(trackerList.map(_.io.alloc.iacq.should)) <= UInt(1),
-    "At most a single tracker should now be allocated for any given Acquire")
-
-  // Wire releases from clients
-  doInputRoutingWithAllocation(
-    in = io.inner.release,
-    outs = trackerAndWbIOs.map(_.inner.release),
-    allocs = trackerAndWbIOs.map(_.alloc.irel))
-
-  assert(PopCount(trackerAndWbIOs.map(_.alloc.irel.should)) <= UInt(1),
-    "At most a single tracker should now be allocated for any given Release")
-
-  // Wire probe requests and grant reply to clients, finish acks from clients
-  doOutputArbitration(io.inner.probe, trackerList.map(_.io.inner.probe) :+ wb.io.inner.probe)
-  doOutputArbitration(io.inner.grant, trackerList.map(_.io.inner.grant) :+ wb.io.inner.grant)
-  doInputRouting(io.inner.finish, trackerList.map(_.io.inner.finish))
-
-  // Create an arbiter for the one memory port
-  val outerList = trackerList.map(_.io.outer) :+ wb.io.outer
-  val outer_arb = Module(new ClientTileLinkIOArbiter(outerList.size)
-                                                    (p.alterPartial({ case TLId => p(OuterTLId)})))
-  outer_arb.io.in <> outerList
-  io.outer <> outer_arb.io.out
-
-  // Wire local memory arrays
-  doInternalOutputArbitration(io.meta.read, trackerList.map(_.io.meta.read) :+ wb.io.meta.read)
-  doInternalOutputArbitration(io.meta.write, trackerList.map(_.io.meta.write))
-  doInternalOutputArbitration(io.data.read, trackerList.map(_.io.data.read) :+ wb.io.data.read)
-  doInternalOutputArbitration(io.data.write, trackerList.map(_.io.data.write))
-  doInternalInputRouting(io.meta.resp, trackerList.map(_.io.meta.resp) :+ wb.io.meta.resp)
-  doInternalInputRouting(io.data.resp, trackerList.map(_.io.data.resp) :+ wb.io.data.resp)
-}
-
-
-class L2XactTrackerIO(implicit p: Parameters) extends HierarchicalXactTrackerIO()(p)
-  with HasL2DataRWIO
-  with HasL2MetaRWIO
-  with HasL2WritebackIO
-
-trait HasRowBeatCounters extends HasOuterCacheParameters with HasPendingBitHelpers {
-  def mergeData(dataBits: Int)(beat: UInt, incoming: UInt): Unit
-
-  def connectDataBeatCounter[S <: L2HellaCacheBundle](inc: Bool, data: S, beat: UInt, full_block: Bool) = {
-    if(data.refillCycles > 1) {
-      val (multi_cnt, multi_done) = Counter(full_block && inc, data.refillCycles)
-      (Mux(!full_block, beat, multi_cnt), Mux(!full_block, inc, multi_done))
-    } else { (UInt(0), inc) }
-  }
-
-  def connectInternalDataBeatCounter[T <: L2HellaCacheBundle with HasL2BeatAddr](
-      in: DecoupledIO[T],
-      beat: UInt = UInt(0),
-      full_block: Bool = Bool(true)): (UInt, Bool) = {
-    connectDataBeatCounter(in.fire(), in.bits, beat, full_block)
-  }
-
-  def connectInternalDataBeatCounter[T <: L2HellaCacheBundle with HasL2Data](
-      in: ValidIO[T],
-      full_block: Bool): Bool = {
-    connectDataBeatCounter(in.valid, in.bits, UInt(0), full_block)._2
-  }
-
-  def addPendingBitInternal[T <: L2HellaCacheBundle with HasL2BeatAddr](in: DecoupledIO[T]) =
-    Fill(in.bits.refillCycles, in.fire()) & UIntToOH(in.bits.addr_beat)
-
-  def addPendingBitInternal[T <: L2HellaCacheBundle with HasL2BeatAddr](in: ValidIO[T]) =
-    Fill(in.bits.refillCycles, in.valid) & UIntToOH(in.bits.addr_beat)
-
-  def dropPendingBit[T <: L2HellaCacheBundle with HasL2BeatAddr] (in: DecoupledIO[T]) =
-    ~Fill(in.bits.refillCycles, in.fire()) | ~UIntToOH(in.bits.addr_beat)
-
-  def dropPendingBitInternal[T <: L2HellaCacheBundle with HasL2BeatAddr] (in: ValidIO[T]) =
-    ~Fill(in.bits.refillCycles, in.valid) | ~UIntToOH(in.bits.addr_beat)
-
-  // TODO: Deal with the possibility that rowBits != tlDataBits
-  def mergeDataInternal[T <: L2HellaCacheBundle with HasL2Data with HasL2BeatAddr](in: ValidIO[T]) {
-    when(in.valid) { mergeData(rowBits)(in.bits.addr_beat, in.bits.data) }
-  }
-}
-
-trait ReadsFromOuterCacheDataArray extends HasCoherenceMetadataBuffer
-    with HasRowBeatCounters
-    with HasDataBuffer {
-  def io: HasL2DataRWIO
-
-  val pending_reads = Reg(init=Bits(0, width = innerDataBeats))
-  val pending_resps = Reg(init=Bits(0, width = innerDataBeats))
-  val curr_read_beat = PriorityEncoder(pending_reads)
-
-  def readDataArray(drop_pending_bit: UInt,
-                    add_pending_bit: UInt = UInt(0),
-                    block_pending_read: Bool = Bool(false),
-                    can_update_pending: Bool = Bool(true)) {
-    val port = io.data
-    when (can_update_pending) {
-      pending_reads := (pending_reads | add_pending_bit) &
-        dropPendingBit(port.read) & drop_pending_bit
-    }
-    port.read.valid := state === s_busy && pending_reads.orR && !block_pending_read
-    port.read.bits := L2DataReadReq(
-                        id = UInt(trackerId),
-                        way_en = xact_way_en,
-                        addr_idx = xact_addr_idx,
-                        addr_beat = curr_read_beat)
-
-    pending_resps := (pending_resps & dropPendingBitInternal(port.resp)) |
-                       addPendingBitInternal(port.read)
-
-    scoreboard += (pending_reads.orR, pending_resps.orR)
-
-    mergeDataInternal(port.resp)
-  }
-}
-
-trait WritesToOuterCacheDataArray extends HasCoherenceMetadataBuffer
-    with HasRowBeatCounters
-    with HasDataBuffer {
-  def io: HasL2DataRWIO
-
-  val pending_writes = Reg(init=Bits(0, width = innerDataBeats))
-  val curr_write_beat = PriorityEncoder(pending_writes)
-
-  def writeDataArray(add_pending_bit: UInt = UInt(0),
-                     block_pending_write: Bool = Bool(false),
-                     can_update_pending: Bool = Bool(true)) {
-    val port = io.data
-    when (can_update_pending) {
-      pending_writes := (pending_writes & dropPendingBit(port.write)) |
-                         add_pending_bit
-    }
-    port.write.valid := state === s_busy && pending_writes.orR && !block_pending_write
-    port.write.bits := L2DataWriteReq(
-                        id = UInt(trackerId),
-                        way_en = xact_way_en,
-                        addr_idx = xact_addr_idx,
-                        addr_beat = curr_write_beat,
-                        wmask = ~UInt(0, port.write.bits.wmask.getWidth),
-                        data = data_buffer(curr_write_beat))
-    
-    scoreboard += pending_writes.orR
-  }
-}
-
-trait HasAMOALU extends HasAcquireMetadataBuffer
-    with HasByteWriteMaskBuffer
-    with HasRowBeatCounters {
-  val io: L2XactTrackerIO
-
-  // Provide a single ALU per tracker to merge Puts and AMOs with data being
-  // refilled, written back, or extant in the cache
-  val amoalu = Module(new AMOALU(amoAluOperandBits, rhsIsAligned = true))
-  val amo_result = Reg(init = UInt(0, innerDataBits))
-  
-  def initializeAMOALUIOs() {
-    amoalu.io.addr := Cat(xact_addr_block, xact_addr_beat, xact_addr_byte)
-    amoalu.io.cmd := xact_op_code
-    amoalu.io.typ := xact_op_size
-    amoalu.io.lhs := io.data.resp.bits.data // default, overwritten by calls to mergeData
-    amoalu.io.rhs := data_buffer.head  // default, overwritten by calls to mergeData
-  }
-
-  // Utility function for applying any buffered stored data to the cache line
-  // before storing it back into the data array
-  override def mergeData(dataBits: Int)(beat: UInt, incoming: UInt) {
-    val old_data = incoming     // Refilled, written back, or de-cached data
-    val new_data = data_buffer(beat) // Newly Put data is already in the buffer
-    val amo_shift_bits = xact_amo_shift_bytes << 3
-    amoalu.io.lhs := old_data >> amo_shift_bits
-    amoalu.io.rhs := new_data >> amo_shift_bits
-    val wmask = FillInterleaved(8, wmask_buffer(beat))
-    data_buffer(beat) := ~wmask & old_data |
-                          wmask & Mux(xact_iacq.isAtomic(), amoalu.io.out << amo_shift_bits, new_data)
-    wmask_buffer(beat) := ~UInt(0, innerWriteMaskBits)
-    when(xact_iacq.isAtomic() && xact_addr_beat === beat) { amo_result := old_data }
-  }
-}
-
-trait HasCoherenceMetadataBuffer extends HasOuterCacheParameters
-    with HasBlockAddressBuffer
-    with HasXactTrackerStates {
-  def trackerId: Int
-
-  val xact_way_en = Reg{ Bits(width = nWays) }
-  val xact_old_meta = Reg{ new L2Metadata }
-  val pending_coh = Reg{ xact_old_meta.coh }
-  val pending_meta_write = Reg{ Bool() } // pending_meta_write has own state (s_meta_write)
-
-  val inner_coh = pending_coh.inner
-  val outer_coh = pending_coh.outer
-
-  val xact_addr_idx = xact_addr_block(idxMSB,idxLSB)
-  val xact_addr_tag = xact_addr_block >> UInt(tagLSB) 
-
-  // Utility function for updating the metadata that will be kept in this cache
-  def updatePendingCohWhen(flag: Bool, next: HierarchicalMetadata) {
-    when(flag && pending_coh =/= next) {
-      pending_meta_write := Bool(true)
-      pending_coh := next
-    }
-  }
-
-  def metaRead(port: HasL2MetaReadIO, next_state: UInt, way_en_known: Bool = Bool(false)) {
-    port.read.valid := state === s_meta_read
-    port.read.bits.id := UInt(trackerId)
-    port.read.bits.idx := xact_addr_idx
-    port.read.bits.tag := xact_addr_tag
-    port.read.bits.way_en := Mux(way_en_known, xact_way_en, ~UInt(0, nWays))
-
-    when(state === s_meta_read && port.read.ready) { state := s_meta_resp }
-
-    when(state === s_meta_resp && port.resp.valid) {
-      xact_old_meta := port.resp.bits.meta
-      when (!way_en_known) { xact_way_en := port.resp.bits.way_en }
-      state := next_state
-    }
-  }
-  
-  def metaWrite(port: HasL2MetaWriteIO, to_write: L2Metadata, next_state: UInt) {
-    port.write.valid := state === s_meta_write
-    port.write.bits.id := UInt(trackerId)
-    port.write.bits.idx := xact_addr_idx
-    port.write.bits.way_en := xact_way_en
-    port.write.bits.data := to_write
-
-    when(state === s_meta_write && port.write.ready) { state := next_state }
-  }
-}
-
-trait TriggersWritebacks extends HasCoherenceMetadataBuffer {
-  def triggerWriteback(wb: L2WritebackIO, next_state: UInt) {
-    wb.req.valid := state === s_wb_req
-    wb.req.bits.id := UInt(trackerId)
-    wb.req.bits.idx := xact_addr_idx
-    wb.req.bits.tag := xact_old_meta.tag
-    wb.req.bits.way_en := xact_way_en
-
-    when(state === s_wb_req && wb.req.ready) { state := s_wb_resp }
-    when(state === s_wb_resp && wb.resp.valid) { state := s_outer_acquire }
-  }
-}
-
-class CacheVoluntaryReleaseTracker(trackerId: Int)(implicit p: Parameters)
-    extends VoluntaryReleaseTracker(trackerId)(p)
-    with HasDataBuffer 
-    with WritesToOuterCacheDataArray {
-  val io = new L2XactTrackerIO
-  pinAllReadyValidLow(io)
-
-  // Avoid metatdata races with writebacks
-  routeInParent(
-    iacqMatches = inSameSet(_, xact_addr_block),
-    irelCanAlloc = Bool(true))
-
-  // Initialize and accept pending Release beats
-  innerRelease(
-    block_vol_ignt = pending_writes.orR,
-    next = s_meta_read)
-
-  io.inner.release.ready := state === s_idle || irel_can_merge || irel_same_xact
-
-  // Begin a transaction by getting the current block metadata
-  metaRead(io.meta, s_busy)
-
-  // Write the voluntarily written back data to this cache
-  writeDataArray(add_pending_bit = addPendingBitWhenBeatHasData(io.inner.release),
-    can_update_pending = state =/= s_idle || io.alloc.irel.should)
-
-  // End a transaction by updating the block metadata
-  metaWrite(
-    io.meta,
-    L2Metadata(
-      tag = xact_addr_tag,
-      inner = xact_old_meta.coh.inner.onRelease(xact_vol_irel),
-      outer = Mux(xact_vol_irel.hasData(),
-                  xact_old_meta.coh.outer.onHit(M_XWR),
-                  xact_old_meta.coh.outer)),
-    s_idle)
-
-  mergeDataInner(io.inner.release)
-
-  when(irel_is_allocating) {
-    pending_writes := addPendingBitWhenBeatHasData(io.inner.release)
-  }
-
-  quiesce(s_meta_write) {}
-
-  // Checks for illegal behavior
-  assert(!(state === s_meta_resp && io.meta.resp.valid && !io.meta.resp.bits.tag_match),
-    "VoluntaryReleaseTracker accepted Release for a block not resident in this cache!")
-}
-
-class CacheAcquireTracker(trackerId: Int)(implicit p: Parameters)
-    extends AcquireTracker(trackerId)(p)
-    with HasByteWriteMaskBuffer
-    with HasAMOALU
-    with TriggersWritebacks
-    with ReadsFromOuterCacheDataArray
-    with WritesToOuterCacheDataArray {
-  val io = new L2XactTrackerIO
-  pinAllReadyValidLow(io)
-  initializeAMOALUIOs()
-
-
-  val pending_coh_on_ognt = HierarchicalMetadata(
-                              ManagerMetadata.onReset,
-                              pending_coh.outer.onGrant(io.outer.grant.bits, xact_op_code))
-
-  val pending_coh_on_ignt = HierarchicalMetadata(
-                              pending_coh.inner.onGrant(io.ignt()),
-                              Mux(ognt_counter.down.done,
-                                pending_coh_on_ognt.outer,
-                                pending_coh.outer))
-
-  val pending_coh_on_irel = HierarchicalMetadata(
-                              pending_coh.inner.onRelease(io.irel()), // Drop sharer
-                              Mux(io.irel().hasData(), // Dirty writeback
-                                pending_coh.outer.onHit(M_XWR),
-                                pending_coh.outer))
-
-  val pending_coh_on_hit = HierarchicalMetadata(
-                              io.meta.resp.bits.meta.coh.inner,
-                              io.meta.resp.bits.meta.coh.outer.onHit(xact_op_code))
-
-  val pending_coh_on_miss = HierarchicalMetadata.onReset
-
-  val before_wb_req = state.isOneOf(s_meta_read, s_meta_resp)
-
-  routeInParent(
-    iacqMatches = inSameSet(_, xact_addr_block),
-    irelMatches = (irel: HasCacheBlockAddress) =>
-      Mux(before_wb_req, inSameSet(irel, xact_addr_block), exactAddrMatch(irel)),
-    iacqCanAlloc = Bool(true))
-
-  // TileLink allows for Gets-under-Get
-  // and Puts-under-Put, and either may also merge with a preceding prefetch
-  // that requested the correct permissions (via op_code)
-  def acquiresAreMergeable(sec: AcquireMetadata): Bool = {
-    val allowedTypes = List((Acquire.getType, Acquire.getType),
-                       (Acquire.putType, Acquire.putType),
-                       (Acquire.putBlockType, Acquire.putBlockType),
-                       (Acquire.getPrefetchType, Acquire.getPrefetchType),
-                       (Acquire.putPrefetchType, Acquire.putPrefetchType),
-                       (Acquire.getPrefetchType, Acquire.getType),
-                       (Acquire.putPrefetchType, Acquire.putType),
-                       (Acquire.putPrefetchType, Acquire.putBlockType))
-    allowedTypes.map { case(a, b) => xact_iacq.isBuiltInType(a) && sec.isBuiltInType(b) }.reduce(_||_) &&
-      xact_op_code === sec.op_code() &&
-      sec.conflicts(xact_addr_block) &&
-      xact_allocate
-  }
-
-  // First, take care of accpeting new acquires or secondary misses
-  // Handling of primary and secondary misses' data and write mask merging
-  def iacq_can_merge = acquiresAreMergeable(io.iacq()) &&
-                         state =/= s_idle &&
-                         state =/= s_meta_resp &&
-                         state =/= s_meta_write &&
-                         !all_pending_done &&
-                         !io.inner.release.fire() &&
-                         !io.outer.grant.fire() &&
-                         !io.data.resp.valid &&
-                         ignt_q.io.enq.ready && ignt_q.io.deq.valid
-
-  innerAcquire(
-    can_alloc = Bool(true),
-    next = s_meta_read)
-
-  io.inner.acquire.ready := state === s_idle || iacq_can_merge ||
-                            iacq_same_xact_multibeat
-
-  // Begin a transaction by getting the current block metadata
-  // Defined here because of Chisel default wire demands, used in s_meta_resp
-  val coh = io.meta.resp.bits.meta.coh
-  val tag_match = io.meta.resp.bits.tag_match
-  val is_hit = (if(!isLastLevelCache) tag_match && coh.outer.isHit(xact_op_code)
-                else tag_match && coh.outer.isValid())
-  val needs_writeback = !tag_match &&
-                        xact_allocate && 
-                        (coh.outer.requiresVoluntaryWriteback() ||
-                           coh.inner.requiresProbesOnVoluntaryWriteback())
-  val needs_inner_probes = tag_match && coh.inner.requiresProbes(xact_iacq)
-  val should_update_meta = !tag_match && xact_allocate ||
-                           is_hit && pending_coh_on_hit =/= coh
-  def full_representation = coh.inner.full()
-
-  metaRead(
-    io.meta,
-    Mux(needs_writeback, s_wb_req,
-      Mux(needs_inner_probes, s_inner_probe,
-        Mux(!is_hit, s_outer_acquire, s_busy))))
-
-  updatePendingCohWhen(
-    io.meta.resp.valid, 
-    Mux(is_hit, pending_coh_on_hit,
-      Mux(tag_match, coh, pending_coh_on_miss)))
-
-  // Issue a request to the writeback unit
-  triggerWriteback(io.wb, s_outer_acquire)
-
-  // Track which clients yet need to be probed and make Probe message
-  // If we're probing, we know the tag matches, so if this is the
-  // last level cache, we can use the data without upgrading permissions
-  val skip_outer_acquire = 
-    (if(!isLastLevelCache) xact_old_meta.coh.outer.isHit(xact_op_code)
-     else xact_old_meta.coh.outer.isValid())
-
-  innerProbe(
-    inner_coh.makeProbe(curr_probe_dst, xact_iacq, xact_addr_block),
-    Mux(!skip_outer_acquire, s_outer_acquire, s_busy))
-
-  // Handle incoming releases from clients, which may reduce sharer counts
-  // and/or write back dirty data, and may be unexpected voluntary releases
-
-  innerRelease() // Don't block on pending_writes because they won't happen until s_busy
-
-  def irel_can_merge = io.irel().conflicts(xact_addr_block) &&
-                         io.irel().isVoluntary() &&
-                         !state.isOneOf(s_idle, s_meta_read, s_meta_resp, s_meta_write) &&
-                         !all_pending_done &&
-                         !io.outer.grant.fire() &&
-                         !io.inner.grant.fire() &&
-                         !vol_ignt_counter.pending
-
-  io.inner.release.ready := irel_can_merge || irel_same_xact
-
-  updatePendingCohWhen(io.inner.release.fire(), pending_coh_on_irel)
-
-  mergeDataInner(io.inner.release)
-
-  // Send outer request
-  outerAcquire(
-    caching = xact_allocate,
-    coh = xact_old_meta.coh.outer, // TODO outer_coh?
-    data = data_buffer(ognt_counter.up.idx),
-    wmask = wmask_buffer(ognt_counter.up.idx),
-    next = s_busy)
-                                        
-  // Handle the response from outer memory
-  updatePendingCohWhen(ognt_counter.down.done, pending_coh_on_ognt)
-  mergeDataOuter(io.outer.grant)
-
-  // Send read request and get resp
-  // Going back to the original inner transaction:
-  // We read from the the cache at this level if data wasn't written back or refilled.
-  // We may still merge further Gets, requiring further beats to be read.
-  // If ECC requires a full writemask, we'll read out data on partial writes as well.
-  readDataArray(
-    drop_pending_bit = (dropPendingBitWhenBeatHasData(io.inner.release) &
-                         dropPendingBitWhenBeatHasData(io.outer.grant)),
-    add_pending_bit = addPendingBitWhenBeatNeedsRead(
-      io.inner.acquire,
-      always = Bool(alwaysWriteFullBeat),
-      unless = data_valid(io.iacq().addr_beat)),
-    block_pending_read = ognt_counter.pending,
-    can_update_pending = state =/= s_idle || io.alloc.irel.should)
-
-  // No override for first accepted acquire
-  val alloc_override = xact_allocate && (state =/= s_idle)
-
-  // Do write
-  // We write data to the cache at this level if it was Put here with allocate flag,
-  // written back dirty, or refilled from outer memory.
-  writeDataArray(
-    add_pending_bit = (addPendingBitWhenBeatHasDataAndAllocs(io.inner.acquire, alloc_override) |
-                        addPendingBitWhenBeatHasData(io.inner.release) |
-                        addPendingBitWhenBeatHasData(io.outer.grant, xact_allocate)),
-    block_pending_write = (ognt_counter.pending ||
-                            pending_put_data.orR ||
-                            pending_reads(curr_write_beat) ||
-                            pending_resps(curr_write_beat)),
-    can_update_pending = state =/= s_idle || io.alloc.iacq.should || io.alloc.irel.should)
-
-  // Acknowledge or respond with data
-  innerGrant(
-    data = Mux(xact_iacq.isAtomic(), amo_result, data_buffer(ignt_data_idx)),
-    external_pending = pending_writes.orR || ognt_counter.pending,
-    add_pending_bits = addPendingBitInternal(io.data.resp))
-
-  updatePendingCohWhen(io.inner.grant.fire() && io.ignt().last(), pending_coh_on_ignt)
-
-  // End a transaction by updating the block metadata
-  metaWrite(io.meta, L2Metadata(xact_addr_tag, pending_coh), s_idle)
-
-  // Initialization of some scoreboard logic based on the original
-  // Acquire message on on the results of the metadata read:
-  when(state === s_meta_resp && io.meta.resp.valid) {
-    // If some kind of Put is marked no-allocate but is already in the cache,
-    // we need to write its data to the data array
-    when(is_hit && !xact_allocate && xact_iacq.hasData()) {
-      pending_writes := addPendingBitsFromAcquire(xact_iacq)
-      xact_allocate := Bool(true)
-    }
-    when (needs_inner_probes) { initializeProbes() }
-    pending_meta_write := should_update_meta //TODO what edge case was this covering?
-  }
-
-  // Initialize more transaction metadata. Pla
-  when(iacq_is_allocating) {
-    amo_result := UInt(0)
-    pending_meta_write := Bool(false)
-    pending_reads := Mux( // Pick out the specific beats of data that need to be read
-      io.iacq().isBuiltInType(Acquire.getBlockType) || !io.iacq().isBuiltInType(),
-      ~UInt(0, width = innerDataBeats),
-      addPendingBitWhenBeatNeedsRead(io.inner.acquire, Bool(alwaysWriteFullBeat)))
-    pending_writes := addPendingBitWhenBeatHasDataAndAllocs(io.inner.acquire)
-    pending_resps := UInt(0)
-  }
-
-  initDataInner(io.inner.acquire, iacq_is_allocating || iacq_is_merging)
-
-  // Wait for everything to quiesce
-  quiesce(Mux(pending_meta_write, s_meta_write, s_idle)) { clearWmaskBuffer() }
-}
-
-class L2WritebackReq(implicit p: Parameters)
-    extends L2HellaCacheBundle()(p) with HasL2Id {
-  val tag  = Bits(width = tagBits)
-  val idx  = Bits(width = idxBits)
-  val way_en = Bits(width = nWays)
-}
-
-class L2WritebackResp(implicit p: Parameters) extends L2HellaCacheBundle()(p) with HasL2Id
-
-class L2WritebackIO(implicit p: Parameters) extends L2HellaCacheBundle()(p) {
-  val req = Decoupled(new L2WritebackReq)
-  val resp = Valid(new L2WritebackResp).flip
-}
-
-trait HasL2WritebackIO extends HasOuterCacheParameters {
-  val wb = new L2WritebackIO()
-}
-
-class L2WritebackUnitIO(implicit p: Parameters)
-    extends HierarchicalXactTrackerIO()(p) with HasL2DataRWIO {
-  val wb = new L2WritebackIO().flip()
-  val meta = new L2MetaReadOnlyIO
-}
-
-class L2WritebackUnit(val trackerId: Int)(implicit p: Parameters) extends XactTracker()(p)
-    with AcceptsVoluntaryReleases
-    with EmitsVoluntaryReleases
-    with EmitsInnerProbes
-    with ReadsFromOuterCacheDataArray
-    with RoutesInParent {
-  val io = new L2WritebackUnitIO
-  pinAllReadyValidLow(io)
-
-  val xact_id = Reg{ io.wb.req.bits.id }
-
-  val pending_coh_on_irel = HierarchicalMetadata(
-                              inner_coh.onRelease(io.irel()), // Drop sharer
-                              Mux(io.irel().hasData(), // Dirty writeback
-                                outer_coh.onHit(M_XWR),
-                                outer_coh))
-
-  routeInParent()
-
-  // Start the writeback sub-transaction
-  io.wb.req.ready := state === s_idle
-
-  val coh = io.meta.resp.bits.meta.coh
-  val needs_inner_probes = coh.inner.requiresProbesOnVoluntaryWriteback()
-  val needs_outer_release = coh.outer.requiresVoluntaryWriteback()
-  def full_representation = coh.inner.full()
-
-  // Even though we already read the metadata in the acquire tracker that
-  // sent the writeback request, we have to read it again in the writeback
-  // unit, since it may have been updated in the meantime.
-  metaRead(io.meta,
-    next_state = Mux(needs_inner_probes, s_inner_probe, s_busy),
-    way_en_known = Bool(true))
-
-  // Track which clients yet need to be probed and make Probe message
-  innerProbe(
-    inner_coh.makeProbeForVoluntaryWriteback(curr_probe_dst, xact_addr_block),
-    s_busy)
-
-  // Handle incoming releases from clients, which may reduce sharer counts
-  // and/or write back dirty data
-  innerRelease()
-
-  def irel_can_merge = io.irel().conflicts(xact_addr_block) &&
-                         io.irel().isVoluntary() &&
-                         !state.isOneOf(s_idle, s_meta_read, s_meta_resp) &&
-                         !(state === s_busy && all_pending_done) &&
-                         !vol_ignt_counter.pending &&
-                         !blockInnerRelease()
-
-  io.inner.release.ready := irel_can_merge || irel_same_xact
-
-  updatePendingCohWhen(io.inner.release.fire(), pending_coh_on_irel)
-
-  mergeDataInner(io.inner.release)
-
-  // If a release didn't write back data, have to read it from data array
-  readDataArray(
-    drop_pending_bit = dropPendingBitWhenBeatHasData(io.inner.release))
-
-  // Once the data is buffered we can write it back to outer memory
-  outerRelease(
-    coh = outer_coh,
-    data = data_buffer(vol_ognt_counter.up.idx),
-    add_pending_data_bits = addPendingBitInternal(io.data.resp),
-    add_pending_send_bit = io.meta.resp.valid && needs_outer_release)
-
-  // Respond to the initiating transaction handler signalling completion of the writeback
-  io.wb.resp.valid := state === s_busy && all_pending_done
-  io.wb.resp.bits.id := xact_id
-
-  quiesce() {}
-
-  // State machine updates and transaction handler metadata intialization
-  when(state === s_idle && io.wb.req.valid) {
-    xact_id := io.wb.req.bits.id
-    xact_way_en := io.wb.req.bits.way_en
-    xact_addr_block := (if (cacheIdBits == 0) Cat(io.wb.req.bits.tag, io.wb.req.bits.idx)
-                        else Cat(io.wb.req.bits.tag, io.wb.req.bits.idx, UInt(cacheId, cacheIdBits)))
-    state := s_meta_read
-  }
-
-  when (state === s_meta_resp && io.meta.resp.valid) {
-    pending_reads := Fill(innerDataBeats, needs_outer_release)
-    pending_coh := coh
-    when(needs_inner_probes) { initializeProbes() }
-  }
-
-  assert(!io.meta.resp.valid || io.meta.resp.bits.tag_match,
-    "L2 requested Writeback for block not present in cache")
-}

src/main/scala/uncore/devices/Debug.scala

@@ -6,6 +6,7 @@ import Chisel._
 import junctions._
 import util._
 import regmapper._
+import tile.XLen
 import uncore.tilelink2._
 import config._
 
@@ -848,7 +849,7 @@ trait DebugModule extends Module with HasDebugModuleParameters with HasRegMap {
   */
 
 class TLDebugModule(address: BigInt = 0)(implicit p: Parameters)
-  extends TLRegisterRouter(address, beatBytes=p(rocket.XLen)/8, executable=true)(
+  extends TLRegisterRouter(address, beatBytes=p(XLen)/8, executable=true)(
   new TLRegBundle((), _ )    with DebugModuleBundle)(
   new TLRegModule((), _, _)  with DebugModule)
 

src/main/scala/uncore/devices/Plic.scala

@@ -11,6 +11,7 @@ import regmapper._
 import uncore.tilelink2._
 import config._
 import scala.math.min
+import tile.XLen
 
 class GatewayPLICIO extends Bundle {
   val valid = Bool(OUTPUT)
@@ -59,7 +60,7 @@ class TLPLIC(supervisor: Boolean, maxPriorities: Int, address: BigInt = 0xC00000
 
   val node = TLRegisterNode(
     address   = AddressSet(address, PLICConsts.size-1),
-    beatBytes = p(rocket.XLen)/8,
+    beatBytes = p(XLen)/8,
     undefZero = false)
 
   val intnode = IntNexusNode(

src/main/scala/uncore/devices/Prci.scala

@@ -12,6 +12,7 @@ import uncore.util._
 import util._
 import scala.math.{min,max}
 import config._
+import tile.XLen
 
 /** Number of tiles */
 case object NTiles extends Field[Int]
@@ -83,7 +84,7 @@ trait CoreplexLocalInterrupterModule extends Module with HasRegMap with MixCorep
 /** Power, Reset, Clock, Interrupt */
 // Magic TL2 Incantation to create a TL2 Slave
 class CoreplexLocalInterrupter(address: BigInt = 0x02000000)(implicit p: Parameters)
-  extends TLRegisterRouter(address, size = ClintConsts.size, beatBytes = p(rocket.XLen)/8, undefZero = true)(
+  extends TLRegisterRouter(address, size = ClintConsts.size, beatBytes = p(XLen)/8, undefZero = true)(
   new TLRegBundle((), _)    with CoreplexLocalInterrupterBundle)(
   new TLRegModule((), _, _) with CoreplexLocalInterrupterModule)
 {

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

@@ -28,8 +28,10 @@ class TLCacheCork(unsafe: Boolean = false)(implicit p: Parameters) extends LazyM
     }
 
     ((io.in zip io.out) zip (node.edgesIn zip node.edgesOut)) foreach { case ((in, out), (edgeIn, edgeOut)) =>
-      require (edgeIn.client.clients.size == 1 || unsafe, "Only one client can safely use a TLCacheCork")
+      val clients = edgeIn.client.clients
-      require (edgeIn.client.clients.filter(_.supportsProbe).size == 1, "Only one caching client allowed")
+      val caches = clients.filter(_.supportsProbe)
+      require (clients.size == 1 || caches.size == 0 || unsafe, "Only one client can safely use a TLCacheCork")
+      require (caches.size <= 1, "Only one caching client allowed")
       edgeOut.manager.managers.foreach { case m =>
         require (!m.supportsAcquireB, "Cannot support caches beyond the Cork")
       }

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

@@ -27,7 +27,7 @@ case class TLManagerParameters(
   require (!address.isEmpty)
   address.foreach { a => require (a.finite) }
 
-  address.combinations(2).foreach { case Seq(x,y) => require (!x.overlaps(y)) }
+  address.combinations(2).foreach { case Seq(x,y) => require (!x.overlaps(y), s"$x and $y overlap.") }
   require (supportsPutFull.contains(supportsPutPartial))
   require (supportsPutFull.contains(supportsArithmetic))
   require (supportsPutFull.contains(supportsLogical))

src/main/scala/uncore/util/Cache.scala

@@ -1,181 +0,0 @@
-// See LICENSE.Berkeley for license details.
-// See LICENSE.SiFive for license details.
-
-package uncore.util
-
-import Chisel._
-import config.{Parameters, Field}
-import rocket.PAddrBits
-import util.ParameterizedBundle
-import uncore.constants._
-
-case class CacheConfig(
-  nSets:         Int,
-  nWays:         Int,
-  rowBits:       Int,
-  nTLBEntries:   Int,
-  cacheIdBits:   Int,
-  splitMetadata: Boolean)
-case class CacheName(id: String) extends Field[CacheConfig]
-case object CacheName extends Field[CacheName]
-
-case object Replacer extends Field[() => ReplacementPolicy]
-case object L2Replacer extends Field[() => SeqReplacementPolicy]
-case object NPrimaryMisses extends Field[Int]
-case object NSecondaryMisses extends Field[Int]
-case object CacheBlockBytes extends Field[Int]
-case object ECCCode extends Field[Option[Code]]
-
-trait HasCacheParameters {
-  implicit val p: Parameters
-  val cacheConfig = p(p(CacheName))
-  val nSets = cacheConfig.nSets
-  val blockOffBits = log2Up(p(CacheBlockBytes))
-  val cacheIdBits = cacheConfig.cacheIdBits
-  val idxBits = log2Up(cacheConfig.nSets)
-  val untagBits = blockOffBits + cacheIdBits + idxBits
-  val tagBits = p(PAddrBits) - untagBits
-  val nWays = cacheConfig.nWays
-  val wayBits = log2Up(nWays)
-  val isDM = nWays == 1
-  val rowBits = cacheConfig.rowBits
-  val rowBytes = rowBits/8
-  val rowOffBits = log2Up(rowBytes)
-  val code = p(ECCCode).getOrElse(new IdentityCode)
-  val hasSplitMetadata = cacheConfig.splitMetadata
-}
-
-abstract class CacheModule(implicit val p: Parameters) extends Module
-  with HasCacheParameters
-abstract class CacheBundle(implicit val p: Parameters) extends ParameterizedBundle()(p)
-  with HasCacheParameters
-
-abstract class ReplacementPolicy {
-  def way: UInt
-  def miss: Unit
-  def hit: Unit
-}
-
-class RandomReplacement(ways: Int) extends ReplacementPolicy {
-  private val replace = Wire(Bool())
-  replace := Bool(false)
-  val lfsr = LFSR16(replace)
-
-  def way = if(ways == 1) UInt(0) else lfsr(log2Up(ways)-1,0)
-  def miss = replace := Bool(true)
-  def hit = {}
-}
-
-abstract class SeqReplacementPolicy {
-  def access(set: UInt): Unit
-  def update(valid: Bool, hit: Bool, set: UInt, way: UInt): Unit
-  def way: UInt
-}
-
-class SeqRandom(n_ways: Int) extends SeqReplacementPolicy {
-  val logic = new RandomReplacement(n_ways)
-  def access(set: UInt) = { }
-  def update(valid: Bool, hit: Bool, set: UInt, way: UInt) = {
-    when (valid && !hit) { logic.miss }
-  }
-  def way = logic.way
-}
-
-class PseudoLRU(n: Int)
-{
-  require(isPow2(n))
-  val state_reg = Reg(Bits(width = n))
-  def access(way: UInt) {
-    state_reg := get_next_state(state_reg,way)
-  }
-  def get_next_state(state: UInt, way: UInt) = {
-    var next_state = state
-    var idx = UInt(1,1)
-    for (i <- log2Up(n)-1 to 0 by -1) {
-      val bit = way(i)
-      next_state = next_state.bitSet(idx, !bit)
-      idx = Cat(idx, bit)
-    }
-    next_state
-  }
-  def replace = get_replace_way(state_reg)
-  def get_replace_way(state: Bits) = {
-    var idx = UInt(1,1)
-    for (i <- 0 until log2Up(n))
-      idx = Cat(idx, state(idx))
-    idx(log2Up(n)-1,0)
-  }
-}
-
-class SeqPLRU(n_sets: Int, n_ways: Int) extends SeqReplacementPolicy {
-  val state = SeqMem(n_sets, Bits(width = n_ways-1))
-  val logic = new PseudoLRU(n_ways)
-  val current_state = Wire(Bits())
-  val plru_way = logic.get_replace_way(current_state)
-  val next_state = Wire(Bits())
-
-  def access(set: UInt) = {
-    current_state := Cat(state.read(set), Bits(0, width = 1))
-  }
-
-  def update(valid: Bool, hit: Bool, set: UInt, way: UInt) = {
-    val update_way = Mux(hit, way, plru_way)
-    next_state := logic.get_next_state(current_state, update_way)
-    when (valid) { state.write(set, next_state(n_ways-1,1)) }
-  }
-
-  def way = plru_way
-}
-
-abstract class Metadata(implicit p: Parameters) extends CacheBundle()(p) {
-  val tag = Bits(width = tagBits)
-}
-
-class MetaReadReq(implicit p: Parameters) extends CacheBundle()(p) {
-  val idx  = Bits(width = idxBits)
-  val way_en = Bits(width = nWays)
-}
-
-class MetaWriteReq[T <: Metadata](gen: T)(implicit p: Parameters) extends MetaReadReq()(p) {
-  val data = gen.cloneType
-  override def cloneType = new MetaWriteReq(gen)(p).asInstanceOf[this.type]
-}
-
-class MetadataArray[T <: Metadata](onReset: () => T)(implicit p: Parameters) extends CacheModule()(p) {
-  val rstVal = onReset()
-  val io = new Bundle {
-    val read = Decoupled(new MetaReadReq).flip
-    val write = Decoupled(new MetaWriteReq(rstVal)).flip
-    val resp = Vec(nWays, rstVal.cloneType).asOutput
-  }
-  val rst_cnt = Reg(init=UInt(0, log2Up(nSets+1)))
-  val rst = rst_cnt < UInt(nSets)
-  val waddr = Mux(rst, rst_cnt, io.write.bits.idx)
-  val wdata = Mux(rst, rstVal, io.write.bits.data).asUInt
-  val wmask = Mux(rst || Bool(nWays == 1), SInt(-1), io.write.bits.way_en.asSInt).toBools
-  val rmask = Mux(rst || Bool(nWays == 1), SInt(-1), io.read.bits.way_en.asSInt).toBools
-  when (rst) { rst_cnt := rst_cnt+UInt(1) }
-
-  val metabits = rstVal.getWidth
-
-  if (hasSplitMetadata) {
-    val tag_arrs = List.fill(nWays){ SeqMem(nSets, UInt(width = metabits)) }
-    val tag_readout = Wire(Vec(nWays,rstVal.cloneType))
-    (0 until nWays).foreach { (i) =>
-      when (rst || (io.write.valid && wmask(i))) {
-        tag_arrs(i).write(waddr, wdata)
-      }
-      io.resp(i) := rstVal.fromBits(tag_arrs(i).read(io.read.bits.idx, io.read.valid && rmask(i)))
-    }
-  } else {
-    val tag_arr = SeqMem(nSets, Vec(nWays, UInt(width = metabits)))
-    when (rst || io.write.valid) {
-      tag_arr.write(waddr, Vec.fill(nWays)(wdata), wmask)
-    }
-    io.resp := tag_arr.read(io.read.bits.idx, io.read.valid).map(rstVal.fromBits(_))
-  }
-
-  io.read.ready := !rst && !io.write.valid // so really this could be a 6T RAM
-  io.write.ready := !rst
-}
-

src/main/scala/util/Replacement.scala

@@ -0,0 +1,83 @@
+// See LICENSE.Berkeley for license details.
+// See LICENSE.SiFive for license details.
+
+package util
+
+import Chisel._
+
+abstract class ReplacementPolicy {
+  def way: UInt
+  def miss: Unit
+  def hit: Unit
+}
+
+class RandomReplacement(ways: Int) extends ReplacementPolicy {
+  private val replace = Wire(Bool())
+  replace := Bool(false)
+  val lfsr = LFSR16(replace)
+
+  def way = if(ways == 1) UInt(0) else lfsr(log2Up(ways)-1,0)
+  def miss = replace := Bool(true)
+  def hit = {}
+}
+
+abstract class SeqReplacementPolicy {
+  def access(set: UInt): Unit
+  def update(valid: Bool, hit: Bool, set: UInt, way: UInt): Unit
+  def way: UInt
+}
+
+class SeqRandom(n_ways: Int) extends SeqReplacementPolicy {
+  val logic = new RandomReplacement(n_ways)
+  def access(set: UInt) = { }
+  def update(valid: Bool, hit: Bool, set: UInt, way: UInt) = {
+    when (valid && !hit) { logic.miss }
+  }
+  def way = logic.way
+}
+
+class PseudoLRU(n: Int)
+{
+  require(isPow2(n))
+  val state_reg = Reg(Bits(width = n))
+  def access(way: UInt) {
+    state_reg := get_next_state(state_reg,way)
+  }
+  def get_next_state(state: UInt, way: UInt) = {
+    var next_state = state
+    var idx = UInt(1,1)
+    for (i <- log2Up(n)-1 to 0 by -1) {
+      val bit = way(i)
+      next_state = next_state.bitSet(idx, !bit)
+      idx = Cat(idx, bit)
+    }
+    next_state
+  }
+  def replace = get_replace_way(state_reg)
+  def get_replace_way(state: Bits) = {
+    var idx = UInt(1,1)
+    for (i <- 0 until log2Up(n))
+      idx = Cat(idx, state(idx))
+    idx(log2Up(n)-1,0)
+  }
+}
+
+class SeqPLRU(n_sets: Int, n_ways: Int) extends SeqReplacementPolicy {
+  val state = SeqMem(n_sets, Bits(width = n_ways-1))
+  val logic = new PseudoLRU(n_ways)
+  val current_state = Wire(Bits())
+  val plru_way = logic.get_replace_way(current_state)
+  val next_state = Wire(Bits())
+
+  def access(set: UInt) = {
+    current_state := Cat(state.read(set), Bits(0, width = 1))
+  }
+
+  def update(valid: Bool, hit: Bool, set: UInt, way: UInt) = {
+    val update_way = Mux(hit, way, plru_way)
+    next_state := logic.get_next_state(current_state, update_way)
+    when (valid) { state.write(set, next_state(n_ways-1,1)) }
+  }
+
+  def way = plru_way
+}