Merge remote-tracking branch 'origin/master' into bump-submodules

.travis.yml

@@ -35,6 +35,7 @@ env:
   matrix:
     - SUITE=RocketSuiteA
     - SUITE=RocketSuiteB
+    - SUITE=RocketSuiteC
     - SUITE=GroundtestSuiteA
     - SUITE=GroundtestSuiteB
     - SUITE=UnittestSuite

regression/Makefile

@@ -43,24 +43,28 @@ endif
 
 ifeq ($(SUITE),RocketSuiteA)
 PROJECT=rocketchip
-CONFIGS=DefaultConfig DefaultBufferlessConfig
+CONFIGS=DefaultConfig
 endif
 
 ifeq ($(SUITE),RocketSuiteB)
 PROJECT=rocketchip
+CONFIGS=DefaultBufferlessConfig
+endif
+
+ifeq ($(SUITE),RocketSuiteC)
+PROJECT=rocketchip
 CONFIGS=DefaultL2Config TinyConfig
 endif
 
 ifeq ($(SUITE),GroundtestSuiteA)
 PROJECT=groundtest
-CONFIGS=MemtestConfig MemtestBufferlessConfig MemtestStatelessConfig FancyMemtestConfig \
+CONFIGS=MemtestConfig MemtestBufferlessConfig MemtestStatelessConfig FancyMemtestConfig
-	BroadcastRegressionTestConfig BufferlessRegressionTestConfig
 endif
 
 ifeq ($(SUITE),GroundtestSuiteB)
 PROJECT=groundtest
-CONFIGS=CacheRegressionTestConfig \
+CONFIGS=BroadcastRegressionTestConfig BufferlessRegressionTestConfig CacheRegressionTestConfig \
-	ComparatorConfig ComparatorBufferlessConfig ComparatorL2Config ComparatorStatelessConfig
+	ComparatorConfig ComparatorBufferlessConfig ComparatorStatelessConfig
 endif
 
 ifeq ($(SUITE),UnittestSuite)

scripts/tracegen.py

@@ -52,6 +52,8 @@ def main():
       numFinished = numFinished + 1
       if numFinished == total:
         break
+    elif line[0:12] == "using random":
+      continue
     else:
       print line,
 

src/main/scala/coreplex/Configs.scala

@@ -11,6 +11,7 @@ import uncore.coherence._
 import uncore.agents._
 import uncore.devices._
 import uncore.converters._
+import uncore.util._
 import rocket._
 import util._
 import util.ConfigUtils._
@@ -108,7 +109,7 @@ class BaseCoreplexConfig extends Config (
       case BroadcastConfig => BroadcastConfig()
       case BankedL2Config => BankedL2Config()
       case CacheBlockBytes => 64
-      case CacheBlockOffsetBits => log2Up(here(CacheBlockBytes))
+      case CacheBlockOffsetBits => log2Up(site(CacheBlockBytes))
       case EnableL2Logging => false
       case _ => throw new CDEMatchError
     }
@@ -123,39 +124,39 @@ class WithNCores(n: Int) extends Config(
 
 class WithNBanksPerMemChannel(n: Int) extends Config(
   (pname, site, here, up) => pname match {
-    case BankedL2Config => up(BankedL2Config).copy(nBanksPerChannel = n)
+    case BankedL2Config => up(BankedL2Config, site).copy(nBanksPerChannel = n)
     case _ => throw new CDEMatchError
   })
 
 class WithNTrackersPerBank(n: Int) extends Config(
   (pname, site, here, up) => pname match {
-    case BroadcastConfig => up(BroadcastConfig).copy(nTrackers = n)
+    case BroadcastConfig => up(BroadcastConfig, site).copy(nTrackers = n)
     case _ => throw new CDEMatchError
   })
 
 // This is the number of sets **per way**
 class WithL1ICacheSets(sets: Int) extends Config(
   (pname, site, here, up) => pname match {
-    case CacheName("L1I") => up(CacheName("L1I")).copy(nSets = sets)
+    case CacheName("L1I") => up(CacheName("L1I"), site).copy(nSets = sets)
     case _ => throw new CDEMatchError
   })
 
 // This is the number of sets **per way**
 class WithL1DCacheSets(sets: Int) extends Config(
   (pname, site, here, up) => pname match {
-    case CacheName("L1D") => up(CacheName("L1D")).copy(nSets = sets)
+    case CacheName("L1D") => up(CacheName("L1D"), site).copy(nSets = sets)
     case _ => throw new CDEMatchError
   })
 
 class WithL1ICacheWays(ways: Int) extends Config(
   (pname, site, here, up) => pname match {
-    case CacheName("L1I") => up(CacheName("L1I")).copy(nWays = ways)
+    case CacheName("L1I") => up(CacheName("L1I"), site).copy(nWays = ways)
     case _ => throw new CDEMatchError
   })
 
 class WithL1DCacheWays(ways: Int) extends Config(
   (pname, site, here, up) => pname match {
-    case CacheName("L1D") => up(CacheName("L1D")).copy(nWays = ways)
+    case CacheName("L1D") => up(CacheName("L1D"), site).copy(nWays = ways)
     case _ => throw new CDEMatchError
   })
 
@@ -168,7 +169,7 @@ class WithCacheBlockBytes(linesize: Int) extends Config(
 class WithDataScratchpad(n: Int) extends Config(
   (pname,site,here,up) => pname match {
     case DataScratchpadSize => n
-    case CacheName("L1D") => up(CacheName("L1D")).copy(nSets = n / site(CacheBlockBytes))
+    case CacheName("L1D") => up(CacheName("L1D"), site).copy(nSets = n / site(CacheBlockBytes))
     case _ => throw new CDEMatchError
   })
 
@@ -187,7 +188,7 @@ class WithL2Cache extends Config(
 
 class WithBufferlessBroadcastHub extends Config(
   (pname, site, here, up) => pname match {
-    case BroadcastConfig => up(BroadcastConfig).copy(bufferless = true)
+    case BroadcastConfig => up(BroadcastConfig, site).copy(bufferless = true)
   })
 
 /**
@@ -205,12 +206,12 @@ class WithBufferlessBroadcastHub extends Config(
 class WithStatelessBridge extends Config(
   (pname,site,here,up) => pname match {
 /* !!! FIXME
-    case BankedL2Config => up(BankedL2Config).copy(coherenceManager = { case (_, _) =>
+    case BankedL2Config => up(BankedL2Config, site).copy(coherenceManager = { case (_, _) =>
       val pass = LazyModule(new TLBuffer(0))
       (pass.node, pass.node)
     })
 */
-    case DCacheKey => up(DCacheKey).copy(nMSHRs = 0)
+    case DCacheKey => up(DCacheKey, site).copy(nMSHRs = 0)
     case _ => throw new CDEMatchError
   })
 
@@ -226,7 +227,7 @@ class WithL2Capacity(size_kb: Int) extends Config(
 
 class WithNL2Ways(n: Int) extends Config(
   (pname,site,here,up) => pname match {
-    case CacheName("L2") => up(CacheName("L2")).copy(nWays = n)
+    case CacheName("L2") => up(CacheName("L2"), site).copy(nWays = n)
   })
 
 class WithRV32 extends Config(
@@ -238,7 +239,7 @@ class WithRV32 extends Config(
 
 class WithBlockingL1 extends Config(
   (pname,site,here,up) => pname match {
-    case DCacheKey => up(DCacheKey).copy(nMSHRs = 0)
+    case DCacheKey => up(DCacheKey, site).copy(nMSHRs = 0)
     case _ => throw new CDEMatchError
   })
 
@@ -249,9 +250,9 @@ class WithSmallCores extends Config(
     case UseVM => false
     case BtbKey => BtbParameters(nEntries = 0)
     case NAcquireTransactors => 2
-    case CacheName("L1D") => up(CacheName("L1D")).copy(nSets = 64, nWays = 1, nTLBEntries = 4)
+    case CacheName("L1D") => up(CacheName("L1D"), site).copy(nSets = 64, nWays = 1, nTLBEntries = 4)
-    case CacheName("L1I") => up(CacheName("L1I")).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).copy(nMSHRs = 0)
+    case DCacheKey => up(DCacheKey, site).copy(nMSHRs = 0)
     case _ => throw new CDEMatchError
   })
 

src/main/scala/coreplex/RISCVPlatform.scala

@@ -29,7 +29,9 @@ trait CoreplexRISCVPlatform extends CoreplexNetwork {
 
   lazy val configString = {
     val managers = l1tol2.node.edgesIn(0).manager.managers
-    rocketchip.GenerateConfigString(p, clint, plic, managers)
+    // Use the existing config string if the user overrode it
+    ConfigStringOutput.contents.getOrElse(
+      rocketchip.GenerateConfigString(p, clint, plic, managers))
   }
 }
 

src/main/scala/coreplex/RocketTiles.scala

@@ -57,7 +57,7 @@ class AsyncRocketTile(tileId: Int)(implicit p: Parameters) extends LazyModule {
       val uncached = uncachedOut.bundleOut
       val slave = slaveNode.map(_.bundleIn)
       val hartid = UInt(INPUT, p(XLen))
-      val interrupts = new TileInterrupts().asInput
+      val interrupts = new TileInterrupts()(rocket.coreParams).asInput
       val resetVector = UInt(INPUT, p(XLen))
     }
     rocket.module.io.interrupts := ShiftRegister(io.interrupts, 3)

src/main/scala/groundtest/CacheFillTest.scala

@@ -3,7 +3,7 @@ package groundtest
 import Chisel._
 import uncore.tilelink._
 import uncore.constants._
-import uncore.agents._
+import uncore.util._
 import util._
 import config._
 

src/main/scala/groundtest/Configs.scala

@@ -6,6 +6,7 @@ import diplomacy._
 import uncore.tilelink._
 import uncore.coherence._
 import uncore.agents._
+import uncore.util._
 import uncore.devices.NTiles
 import junctions._
 import config._
@@ -122,7 +123,7 @@ class WithAtomics extends Config(
 
 class WithPrefetches extends Config(
   (pname, site, here, up) => pname match {
-    case ComparatorKey => up(ComparatorKey).copy(prefetches = true)
+    case ComparatorKey => up(ComparatorKey, site).copy(prefetches = true)
     case _ => throw new CDEMatchError
   })
 
@@ -201,6 +202,6 @@ class WithTraceGen extends Config(
       }.flatten
     }
     case UseAtomics => true
-    case CacheName("L1D") => up(CacheName("L1D")).copy(nSets = 16, nWays = 1)
+    case CacheName("L1D") => up(CacheName("L1D"), site).copy(nSets = 16, nWays = 1)
     case _ => throw new CDEMatchError
   })

src/main/scala/groundtest/Coreplex.scala

@@ -6,9 +6,10 @@ import diplomacy._
 import coreplex._
 import uncore.devices.NTiles
 import uncore.tilelink2._
-import rocket.TileId
 import uncore.tilelink.TLId
 
+case object TileId extends Field[Int]
+
 class GroundTestCoreplex(implicit p: Parameters) extends BaseCoreplex {
   val tiles = List.tabulate(p(NTiles)) { i =>
     LazyModule(new GroundTestTile()(p.alterPartial({

src/main/scala/groundtest/Regression.scala

@@ -3,7 +3,7 @@ package groundtest
 import Chisel._
 import uncore.tilelink._
 import uncore.constants._
-import uncore.agents._
+import uncore.util._
 import util._
 import rocket._
 import rocketchip._

src/main/scala/groundtest/Tile.scala

@@ -3,7 +3,7 @@ package groundtest
 import Chisel._
 import rocket._
 import uncore.tilelink._
-import uncore.agents.CacheName
+import uncore.util.CacheName
 import uncore.tilelink2._
 import rocketchip.ExtMem
 import diplomacy._

src/main/scala/rocket/Dcache.scala

@@ -5,7 +5,6 @@ package rocket
 import Chisel._
 import Chisel.ImplicitConversions._
 import diplomacy._
-import uncore.agents._
 import uncore.constants._
 import uncore.tilelink2._
 import uncore.util._
@@ -252,7 +251,7 @@ class DCacheModule(outer: DCache)(implicit p: Parameters) extends HellaCacheModu
   val a_source = PriorityEncoder(~uncachedInFlight.asUInt)
   val acquire_address = s2_req_block_addr
   val access_address = s2_req.addr
-  val a_size = s2_req.typ
+  val a_size = s2_req.typ(MT_SZ-2, 0)
   val a_data = Fill(beatWords, pstore1_storegen.data)
   val acquire = if (edge.manager.anySupportAcquire) {
     edge.Acquire(a_source, acquire_address, lgCacheBlockBytes, s2_grow_param)._2 // Cacheability checked by tlb

src/main/scala/rocket/HellaCache.scala

@@ -6,7 +6,7 @@ import Chisel._
 import config.{Parameters, Field}
 import diplomacy._
 import uncore.tilelink2._
-import uncore.agents._
+import uncore.util._
 import uncore.constants._
 import uncore.tilelink.{TLKey, TLId}
 import util.ParameterizedBundle

src/main/scala/rocket/NBDcache.scala

@@ -5,7 +5,6 @@ package rocket
 import Chisel._
 import Chisel.ImplicitConversions._
 import diplomacy._
-import uncore.agents._
 import uncore.constants._
 import uncore.tilelink._
 import uncore.tilelink2._
@@ -176,15 +175,18 @@ class MSHR(id: Int, edge: TLEdgeOut)(implicit cfg: DCacheConfig, p: Parameters)
   val req_tag = req.addr >> untagBits
   val req_block_addr = (req.addr >> blockOffBits) << blockOffBits
   val idx_match = req_idx === io.req_bits.addr(untagBits-1,blockOffBits)
+
+  val new_coh = Reg(init=ClientMetadata.onReset)
+  val (_, shrink_param, coh_on_clear)    = req.old_meta.coh.onCacheControl(M_FLUSH)
+  val grow_param                                  = new_coh.onAccess(req.cmd)._2
+  val coh_on_grant                                = new_coh.onGrant(req.cmd, io.mem_grant.bits.param)
   // We only accept secondary misses if we haven't yet sent an Acquire to outer memory
   // or if the Acquire that was sent will obtain a Grant with sufficient permissions
   // to let us replay this new request. I.e. we don't handle multiple outstanding
   // Acquires on the same block for now.
-  val cmd_requires_second_acquire = 
+  val (cmd_requires_second_acquire, is_hit_again, _, dirtier_coh, dirtier_cmd) =
-    req.old_meta.coh.requiresAcquireOnSecondaryMiss(req.cmd, io.req_bits.cmd)
+    new_coh.onSecondaryAccess(req.cmd, io.req_bits.cmd)
-  // Track whether or not a secondary acquire will cause the coherence state
+
-  // to go from clean to dirty.
-  val dirties_coh = Reg(Bool())
   val states_before_refill = Seq(s_wb_req, s_wb_resp, s_meta_clear)
   val (_, _, refill_done, refill_address_inc) = edge.addr_inc(io.mem_grant)
   val sec_rdy = idx_match &&
@@ -197,19 +199,6 @@ class MSHR(id: Int, edge: TLEdgeOut)(implicit cfg: DCacheConfig, p: Parameters)
   rpq.io.enq.bits := io.req_bits
   rpq.io.deq.ready := (io.replay.ready && state === s_drain_rpq) || state === s_invalid
 
-  // TODO clean all this coh state business up
-  val new_coh_state = Reg(init=ClientMetadata.onReset)
-  val grow_param = Reg(init=UInt(0))
-  val coh_on_grant = Mux(dirties_coh,
-                          ClientMetadata.maximum,
-                          req.old_meta.coh.onGrant(req.cmd, io.mem_grant.bits.param))
-  val (is_hit, missed_param, coh_on_hit)   = io.req_bits.old_meta.coh.onAccess(io.req_bits.cmd)
-  val (needs_wb, _, _)     = io.req_bits.old_meta.coh.onCacheControl(M_FLUSH)
-  val (_, shrink_param, _)  = req.old_meta.coh.onCacheControl(M_FLUSH)
-  val (hit_again, missed_again_param, _) = req.old_meta.coh.onCacheControl(io.req_bits.cmd)
-  val (_, _, clear_coh_state) = req.old_meta.coh.onCacheControl(M_FLUSH)
-  val (_, after_wb_param, _) = ClientMetadata.onReset.onAccess(req.cmd)
-
   when (state === s_drain_rpq && !rpq.io.deq.valid) {
     state := s_invalid
   }
@@ -221,14 +210,13 @@ class MSHR(id: Int, edge: TLEdgeOut)(implicit cfg: DCacheConfig, p: Parameters)
     state := s_meta_write_resp
   }
   when (state === s_refill_resp && refill_done) {
+    new_coh := coh_on_grant
     state := s_meta_write_req
-    new_coh_state := coh_on_grant
   }
   when (io.mem_acquire.fire()) { // s_refill_req
     state := s_refill_resp
   }
   when (state === s_meta_clear && io.meta_write.ready) {
-    grow_param := after_wb_param
     state := s_refill_req
   }
   when (state === s_wb_resp && io.mem_grant.valid) {
@@ -241,24 +229,26 @@ class MSHR(id: Int, edge: TLEdgeOut)(implicit cfg: DCacheConfig, p: Parameters)
     //If we get a secondary miss that needs more permissions before we've sent
     //  out the primary miss's Acquire, we can upgrade the permissions we're 
     //  going to ask for in s_refill_req
-    when(cmd_requires_second_acquire) {
+    req.cmd := dirtier_cmd
-      req.cmd := io.req_bits.cmd
+    when (is_hit_again) {
-      when(!hit_again) { grow_param := missed_again_param }
+      new_coh := dirtier_coh
     }
-    dirties_coh := dirties_coh || isWrite(io.req_bits.cmd)
   }
   when (io.req_pri_val && io.req_pri_rdy) {
     req := io.req_bits
-    dirties_coh := isWrite(io.req_bits.cmd)
+    val old_coh = io.req_bits.old_meta.coh
+    val needs_wb = old_coh.onCacheControl(M_FLUSH)._1
+    val (is_hit, _, coh_on_hit) = old_coh.onAccess(io.req_bits.cmd)
     when (io.req_bits.tag_match) {
       when (is_hit) { // set dirty bit
+        new_coh := coh_on_hit
         state := s_meta_write_req
-        new_coh_state := coh_on_hit
       }.otherwise { // upgrade permissions
+        new_coh := old_coh
         state := s_refill_req
-        grow_param := missed_param
       }
     }.otherwise { // writback if necessary and refill
+      new_coh := ClientMetadata.onReset
       state := Mux(needs_wb, s_wb_req, s_meta_clear)
     }
   }
@@ -285,7 +275,7 @@ class MSHR(id: Int, edge: TLEdgeOut)(implicit cfg: DCacheConfig, p: Parameters)
 
   io.meta_write.valid := state.isOneOf(s_meta_write_req, s_meta_clear)
   io.meta_write.bits.idx := req_idx
-  io.meta_write.bits.data.coh := Mux(state === s_meta_clear, clear_coh_state, new_coh_state)
+  io.meta_write.bits.data.coh := Mux(state === s_meta_clear, coh_on_clear, new_coh)
   io.meta_write.bits.data.tag := io.tag
   io.meta_write.bits.way_en := req.way_en
 

src/main/scala/rocket/btb.scala

@@ -6,7 +6,7 @@ import Chisel._
 import config._
 import util._
 import Chisel.ImplicitConversions._
-import uncore.agents.PseudoLRU
+import uncore.util.PseudoLRU
 
 case object BtbKey extends Field[BtbParameters]
 

src/main/scala/rocket/ptw.scala

@@ -3,7 +3,7 @@
 package rocket
 
 import Chisel._
-import uncore.agents._
+import uncore.util.PseudoLRU
 import uncore.constants._
 import util._
 import Chisel.ImplicitConversions._

src/main/scala/rocket/rocc.scala

@@ -5,7 +5,7 @@ package rocket
 import Chisel._
 import uncore.tilelink._
 import uncore.constants._
-import uncore.agents.CacheName
+import uncore.util.CacheName
 import util._
 import Chisel.ImplicitConversions._
 import config._

src/main/scala/rocket/rocket.scala

@@ -4,7 +4,7 @@ package rocket
 
 import Chisel._
 import uncore.devices._
-import uncore.agents.CacheName
+import uncore.util.CacheName
 import uncore.constants._
 import uncore.tilelink2._
 import util._

src/main/scala/rocket/tile.scala

@@ -6,7 +6,7 @@ import Chisel._
 import diplomacy._
 import uncore.tilelink._
 import uncore.tilelink2._
-import uncore.agents._
+import uncore.util.{CacheName, CacheBlockBytes}
 import uncore.converters._
 import uncore.devices._
 import util._
@@ -14,7 +14,6 @@ import config._
 import scala.collection.mutable.ListBuffer
 
 case object BuildRoCC extends Field[Seq[RoccParameters]]
-case object TileId extends Field[Int]
 
 case class RoccParameters(
   opcodes: OpcodeSet,
@@ -30,7 +29,6 @@ class RocketTile(tileId: Int)(implicit p: Parameters) extends LazyModule {
   val dcacheParams = coreParams.alterPartial({
     case CacheName => CacheName("L1D")
     case TLId => "L1toL2"
-    case TileId => tileId // TODO using this messes with Heirarchical P&R: change to io.hartid?
   })
   val icacheParams = coreParams.alterPartial({
     case CacheName => CacheName("L1I")

src/main/scala/rocket/tlb.scala

@@ -8,8 +8,7 @@ import Chisel.ImplicitConversions._
 import scala.math._
 import config._
 import diplomacy._
-import uncore.agents._
+import uncore.util._
-import uncore.coherence._
 import uncore.tilelink2._
 
 case object PgLevels extends Field[Int]
@@ -66,8 +65,10 @@ class TLB(implicit val p: Parameters) extends Module with HasTLBParameters {
   val refill_ppn = io.ptw.resp.bits.pte.ppn(ppnBits-1, 0)
   val do_refill = Bool(usingVM) && io.ptw.resp.valid
   val mpu_ppn = Mux(do_refill, refill_ppn, passthrough_ppn)
+  val mpu_physaddr = mpu_ppn << pgIdxBits
+  val legal_address = edge.manager.findSafe(mpu_physaddr).reduce(_||_)
   def fastCheck(member: TLManagerParameters => Boolean) =
-    Mux1H(edge.manager.findFast(mpu_ppn << pgIdxBits), edge.manager.managers.map(m => Bool(member(m))))
+    legal_address && Mux1H(edge.manager.findFast(mpu_physaddr), edge.manager.managers.map(m => Bool(member(m))))
   val prot_r = fastCheck(_.supportsGet)
   val prot_w = fastCheck(_.supportsPutFull)
   val prot_x = fastCheck(_.executable)

src/main/scala/rocketchip/Configs.scala

@@ -28,8 +28,9 @@ class BasePlatformConfig extends Config(
     case PeripheryBusArithmetic => true
     // Note that PLIC asserts that this is > 0.
     case IncludeJtagDTM => false
-    case ExtMem => AXIMasterConfig(0x80000000L, 0x10000000L, 8, 4)
+    case ExtMem => MasterConfig(base=0x80000000L, size=0x10000000L, beatBytes=8, idBits=4)
-    case ExtBus => AXIMasterConfig(0x60000000L, 0x20000000L, 8, 4)
+    case ExtBus => MasterConfig(base=0x60000000L, size=0x20000000L, beatBytes=8, idBits=4)
+    case ExtIn  => SlaveConfig(beatBytes=8, idBits=8, sourceBits=2)
     case RTCPeriod => 100 // gives 10 MHz RTC assuming 1 GHz uncore clock
     case _ => throw new CDEMatchError
   })
@@ -56,14 +57,14 @@ class PLRUL2Config extends Config(new WithPLRU ++ new DefaultL2Config)
 
 class WithNMemoryChannels(n: Int) extends Config(
   (pname,site,here,up) => pname match {
-    case BankedL2Config => up(BankedL2Config).copy(nMemoryChannels = n)
+    case BankedL2Config => up(BankedL2Config, site).copy(nMemoryChannels = n)
     case _ => throw new CDEMatchError
   }
 )
 
 class WithExtMemSize(n: Long) extends Config(
   (pname,site,here,up) => pname match {
-    case ExtMem => up(ExtMem).copy(size = n)
+    case ExtMem => up(ExtMem, site).copy(size = n)
     case _ => throw new CDEMatchError
   }
 )
@@ -94,7 +95,7 @@ class RoccExampleConfig extends Config(new WithRoccExample ++ new BaseConfig)
 
 class WithEdgeDataBits(dataBits: Int) extends Config(
   (pname, site, here, up) => pname match {
-    case ExtMem => up(ExtMem).copy(beatBytes = dataBits/8)
+    case ExtMem => up(ExtMem, site).copy(beatBytes = dataBits/8)
     case _ => throw new CDEMatchError
   })
 
@@ -168,3 +169,9 @@ class WithNBreakpoints(hwbp: Int) extends Config (
     case _ => throw new CDEMatchError
   }
 )
+
+class WithRTCPeriod(nCycles: Int) extends Config(
+  (pname, site, here) => pname match {
+    case RTCPeriod => nCycles
+    case _ => throw new CDEMatchError
+  })

src/main/scala/rocketchip/DebugTransport.scala

@@ -38,32 +38,25 @@ case object IncludeJtagDTM extends Field[Boolean]
  * 
  */
 
-class JtagDTMWithSync(depth: Int = 1, sync: Int = 3)(implicit val p: Parameters)
+class JtagDTMWithSync(implicit val p: Parameters) extends Module {
-    extends Module {
-
   // io.DebugBusIO <-> Sync <-> DebugBusIO <-> UInt <-> DTM Black Box
 
   val io = new Bundle {
-
     val jtag = new JTAGIO(true).flip
     val debug = new AsyncDebugBusIO
-
   }
 
   val req_width = io.debug.req.mem(0).getWidth
   val resp_width = io.debug.resp.mem(0).getWidth
 
-  val jtag_dtm = Module (new DebugTransportModuleJtag(req_width, resp_width))
+  val jtag_dtm = Module(new DebugTransportModuleJtag(req_width, resp_width))
-
   jtag_dtm.io.jtag <> io.jtag
 
-  val dtm_req = Wire(new DecoupledIO(UInt(width = req_width)))
-  val dtm_resp = Wire(new DecoupledIO(UInt(width = resp_width)))
-
   val io_debug_bus = Wire (new DebugBusIO)
+  io.debug <> ToAsyncDebugBus(io_debug_bus)
 
-  io.debug.req <> ToAsyncBundle(io_debug_bus.req)
+  val dtm_req = jtag_dtm.io.dtm_req
-  io_debug_bus.resp <> FromAsyncBundle(io.debug.resp)
+  val dtm_resp = jtag_dtm.io.dtm_resp
 
   // Translate from straight 'bits' interface of the blackboxes
   // into the Resp/Req data structures.
@@ -74,20 +67,12 @@ class JtagDTMWithSync(depth: Int = 1, sync: Int = 3)(implicit val p: Parameters)
   dtm_resp.valid := io_debug_bus.resp.valid
   dtm_resp.bits  := io_debug_bus.resp.bits.asUInt
   io_debug_bus.resp.ready  := dtm_resp.ready
-
-  dtm_req <> jtag_dtm.io.dtm_req
-  jtag_dtm.io.dtm_resp <> dtm_resp
 }
 
 class DebugTransportModuleJtag(reqSize : Int, respSize : Int)(implicit val p: Parameters)  extends BlackBox {
-
   val io = new Bundle {
     val jtag = new JTAGIO(true).flip()
-
     val dtm_req = new DecoupledIO(UInt(width = reqSize))
-
     val dtm_resp = new DecoupledIO(UInt(width = respSize)).flip()
-
   }
-
 }

src/main/scala/rocketchip/Periphery.scala

@@ -20,9 +20,11 @@ import scala.math.max
 import coreplex._
 
 /** Specifies the size of external memory */
-case class AXIMasterConfig(base: Long, size: Long, beatBytes: Int, idBits: Int)
+case class MasterConfig(base: Long, size: Long, beatBytes: Int, idBits: Int)
-case object ExtMem extends Field[AXIMasterConfig]
+case object ExtMem extends Field[MasterConfig]
-case object ExtBus extends Field[AXIMasterConfig]
+case object ExtBus extends Field[MasterConfig]
+case class SlaveConfig(beatBytes: Int, idBits: Int, sourceBits: Int)
+case object ExtIn extends Field[SlaveConfig]
 /** Specifies the number of external interrupts */
 case object NExtTopInterrupts extends Field[Int]
 /** Source of RTC. First bundle is TopIO.extra, Second bundle is periphery.io.extra  **/
@@ -70,6 +72,14 @@ trait PeripheryExtInterruptsModule {
 
 /////
 
+trait PeripheryNoMem extends TopNetwork {
+  private val channels = p(BankedL2Config).nMemoryChannels
+  require (channels == 0)
+  val mem = Seq()
+}
+
+/////
+
 trait PeripheryMasterAXI4Mem {
   this: TopNetwork =>
   val module: PeripheryMasterAXI4MemModule
@@ -155,14 +165,17 @@ trait PeripheryMasterAXI4MMIOModule {
 
 // PeripherySlaveAXI4 is an example, make your own cake pattern like this one.
 trait PeripherySlaveAXI4 extends L2Crossbar {
-  private val axiIdBits = 8
+  private val config = p(ExtIn)
-  private val tlIdBits = 2 // at most 4 AXI requets inflight at a time
-
   val l2_axi4 = AXI4BlindInputNode(AXI4MasterPortParameters(
     masters = Seq(AXI4MasterParameters(
-      id = IdRange(0, 1 << axiIdBits)))))
+      id = IdRange(0, 1 << config.idBits)))))
 
-  l2.node := TLSourceShrinker(1 << tlIdBits)(AXI4ToTL()(AXI4Fragmenter()(l2_axi4)))
+  l2.node :=
+    TLSourceShrinker(1 << config.sourceBits)(
+    TLWidthWidget(config.beatBytes)(
+    AXI4ToTL()(
+    AXI4Fragmenter()(
+    l2_axi4))))
 }
 
 trait PeripherySlaveAXI4Bundle extends L2CrossbarBundle {
@@ -178,6 +191,69 @@ trait PeripherySlaveAXI4Module extends L2CrossbarModule {
 
 /////
 
+// Add an external TL-UL slave
+trait PeripheryMasterTLMMIO {
+  this: TopNetwork =>
+
+  private val config = p(ExtBus)
+  val mmio_tl = TLBlindOutputNode(TLManagerPortParameters(
+    managers = Seq(TLManagerParameters(
+      address            = List(AddressSet(BigInt(config.base), config.size-1)),
+      executable         = true,
+      supportsGet        = TransferSizes(1, cacheBlockBytes),
+      supportsPutFull    = TransferSizes(1, cacheBlockBytes),
+      supportsPutPartial = TransferSizes(1, cacheBlockBytes))),
+    beatBytes = config.beatBytes))
+
+  mmio_tl :=
+    TLSourceShrinker(config.idBits)(
+    TLWidthWidget(socBusConfig.beatBytes)(
+    socBus.node))
+}
+
+trait PeripheryMasterTLMMIOBundle {
+  this: TopNetworkBundle {
+    val outer: PeripheryMasterTLMMIO
+  } =>
+  val mmio_tl = outer.mmio_tl.bundleOut
+}
+
+trait PeripheryMasterTLMMIOModule {
+  this: TopNetworkModule {
+    val outer: PeripheryMasterTLMMIO
+    val io: PeripheryMasterTLMMIOBundle
+  } =>
+  // nothing to do
+}
+
+/////
+
+// NOTE: this port is NOT allowed to issue Acquires
+trait PeripherySlaveTL extends L2Crossbar {
+  private val config = p(ExtIn)
+  val l2_tl = TLBlindInputNode(TLClientPortParameters(
+    clients = Seq(TLClientParameters(
+      sourceId = IdRange(0, 1 << config.idBits)))))
+
+  l2.node :=
+    TLSourceShrinker(1 << config.sourceBits)(
+    TLWidthWidget(config.beatBytes)(
+    l2_tl))
+}
+
+trait PeripherySlaveTLBundle extends L2CrossbarBundle {
+  val outer: PeripherySlaveTL
+  val l2_tl = outer.l2_tl.bundleIn
+}
+
+trait PeripherySlaveTLModule extends L2CrossbarModule {
+  val outer: PeripherySlaveTL
+  val io: PeripherySlaveTLBundle
+  // nothing to do
+}
+
+/////
+
 trait PeripheryBootROM {
   this: TopNetwork =>
   val coreplex: CoreplexRISCVPlatform

src/main/scala/rocketchip/RISCVPlatform.scala

@@ -11,6 +11,8 @@ import util._
 import junctions.JTAGIO
 import coreplex._
 
+/// Core with JTAG for debug only
+
 trait PeripheryJTAG extends TopNetwork {
   val module: PeripheryJTAGModule
   val coreplex: CoreplexRISCVPlatform
@@ -34,6 +36,8 @@ trait PeripheryJTAGModule extends TopNetworkModule {
   dtm.reset := io.jtag.TRST
 }
 
+/// Core with DTM for debug only
+
 trait PeripheryDTM extends TopNetwork {
   val module: PeripheryDTMModule
   val coreplex: CoreplexRISCVPlatform
@@ -52,6 +56,36 @@ trait PeripheryDTMModule extends TopNetworkModule {
   outer.coreplex.module.io.debug <> ToAsyncDebugBus(io.debug)
 }
 
+/// Core with DTM or JTAG based on a parameter
+
+trait PeripheryDebug extends TopNetwork {
+  val module: PeripheryDebugModule
+  val coreplex: CoreplexRISCVPlatform
+}
+
+trait PeripheryDebugBundle extends TopNetworkBundle {
+  val outer: PeripheryDebug
+
+  val debug = (!p(IncludeJtagDTM)).option(new DebugBusIO().flip)
+  val jtag = (p(IncludeJtagDTM)).option(new JTAGIO(true).flip)
+}
+
+trait PeripheryDebugModule extends TopNetworkModule {
+  val outer: PeripheryDebug
+  val io: PeripheryDebugBundle
+
+  io.debug.foreach { dbg => outer.coreplex.module.io.debug <> ToAsyncDebugBus(dbg) }
+  io.jtag.foreach { jtag =>
+    val dtm = Module (new JtagDTMWithSync)
+    dtm.clock := jtag.TCK
+    dtm.reset := jtag.TRST
+    dtm.io.jtag <> jtag
+    outer.coreplex.module.io.debug <> dtm.io.debug
+  }
+}
+
+/// Real-time clock is based on RTCPeriod relative to Top clock
+
 trait PeripheryCounter extends TopNetwork {
   val module: PeripheryCounterModule
   val coreplex: CoreplexRISCVPlatform
@@ -75,6 +109,8 @@ trait PeripheryCounterModule extends TopNetworkModule {
   }
 }
 
+/// Coreplex will power-on running at 0x1000 (BootROM)
+
 trait HardwiredResetVector extends TopNetwork {
   val module: HardwiredResetVectorModule
   val coreplex: CoreplexRISCVPlatform

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

@@ -5,186 +5,14 @@ package uncore.agents
 import Chisel._
 import scala.reflect.ClassTag
 import junctions.PAddrBits
-import util.ParameterizedBundle
-import uncore.util.AMOALU
 import uncore.coherence._
 import uncore.tilelink._
 import uncore.constants._
 import uncore.util._
 import util._
-import config._
+import config.{Parameters, Field}
 
-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]]
 case object CacheId extends Field[Int]
-
-trait HasCacheParameters {
-  implicit val p: Parameters
-  val cacheConfig = p(p(CacheName))
-  val nSets = cacheConfig.nSets
-  val blockOffBits = p(CacheBlockOffsetBits)
-  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
-}
-
 case object L2DirectoryRepresentation extends Field[DirectoryRepresentation]
 
 trait HasOuterCacheParameters extends HasCacheParameters with HasCoherenceAgentParameters {

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

@@ -53,3 +53,6 @@ case class AXI4InputNode() extends InputNode(AXI4Imp)
 // Nodes used for external ports
 case class AXI4BlindOutputNode(portParams: AXI4SlavePortParameters) extends BlindOutputNode(AXI4Imp)(portParams)
 case class AXI4BlindInputNode(portParams: AXI4MasterPortParameters) extends BlindInputNode(AXI4Imp)(portParams)
+
+case class AXI4InternalOutputNode(portParams: AXI4SlavePortParameters) extends InternalOutputNode(AXI4Imp)(portParams)
+case class AXI4InternalInputNode(portParams: AXI4MasterPortParameters) extends InternalInputNode(AXI4Imp)(portParams)

src/main/scala/uncore/tilelink2/Metadata.scala

@@ -88,12 +88,6 @@ class ClientMetadata extends Bundle {
       Cat(wr, toT)   -> Dirty))
   }
 
-  /** Does a secondary miss on the block require another Acquire message */
-  def requiresAcquireOnSecondaryMiss(first_cmd: UInt, second_cmd: UInt): Bool = {
-    import MemoryOpCategories._
-    isWriteIntent(second_cmd) && !isWriteIntent(first_cmd)
-  }
-
   /** Does this cache have permissions on this block sufficient to perform op,
     * and what to do next (Acquire message param or updated metadata). */
   def onAccess(cmd: UInt): (Bool, UInt, ClientMetadata) = {
@@ -101,6 +95,20 @@ class ClientMetadata extends Bundle {
     (r._1, r._2, ClientMetadata(r._2))
   }
 
+  /** Does a secondary miss on the block require another Acquire message */
+  def onSecondaryAccess(first_cmd: UInt, second_cmd: UInt): (Bool, Bool, UInt, ClientMetadata, UInt) = {
+    import MemoryOpCategories._
+    val r1 = growStarter(first_cmd)
+    val r2 = growStarter(second_cmd)
+    val needs_second_acq = isWriteIntent(second_cmd) && !isWriteIntent(first_cmd)
+    val hit_again = r1._1 && r2._1
+    val dirties = categorize(second_cmd) === wr
+    val biggest_grow_param = Mux(dirties, r2._2, r1._2)
+    val dirtiest_state = ClientMetadata(biggest_grow_param)
+    val dirtiest_cmd = Mux(dirties, second_cmd, first_cmd)
+    (needs_second_acq, hit_again, biggest_grow_param, dirtiest_state, dirtiest_cmd)
+  }
+
   /** Metadata change on a returned Grant */
   def onGrant(cmd: UInt, param: UInt): ClientMetadata = ClientMetadata(growFinisher(cmd, param))
 

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

@@ -126,6 +126,9 @@ case class TLInputNode() extends InputNode(TLImp)
 case class TLBlindOutputNode(portParams: TLManagerPortParameters) extends BlindOutputNode(TLImp)(portParams)
 case class TLBlindInputNode(portParams: TLClientPortParameters) extends BlindInputNode(TLImp)(portParams)
 
+case class TLInternalOutputNode(portParams: TLManagerPortParameters) extends InternalOutputNode(TLImp)(portParams)
+case class TLInternalInputNode(portParams: TLClientPortParameters) extends InternalInputNode(TLImp)(portParams)
+
 /** Synthesizeable unit tests */
 import unittest._
 

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

@@ -0,0 +1,180 @@
+// See LICENSE for license details.
+
+package uncore.util
+
+import Chisel._
+import config.{Parameters, Field}
+import junctions.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/uncore/util/Ecc.scala

@@ -1,6 +1,6 @@
 // See LICENSE for license details.
 
-package uncore.agents
+package uncore.util
 
 import Chisel._
 import util._

src/main/scala/util/AsyncBundle.scala

@@ -9,8 +9,8 @@ final class AsyncBundle[T <: Data](val depth: Int, gen: T) extends Bundle
 {
   require (isPow2(depth))
   val mem  = Vec(depth, gen)
-  val ridx = UInt(width = log2Up(depth)+1).flip
+  val ridx = UInt(width = log2Ceil(depth)+1).flip
-  val widx = UInt(width = log2Up(depth)+1)
+  val widx = UInt(width = log2Ceil(depth)+1)
   val ridx_valid = Bool().flip
   val widx_valid = Bool()
   val source_reset_n = Bool()

src/main/scala/util/Config.scala

@@ -6,7 +6,8 @@ class CDEMatchError() extends Exception {
 }
 
 abstract class View {
-  final def apply[T](pname: Field[T]): T = find(pname, this).asInstanceOf[T]
+  final def apply[T](pname: Field[T]): T = apply(pname, this)
+  final def apply[T](pname: Field[T], site: View): T = find(pname, site).asInstanceOf[T]
 
   protected[config] def find(pname: Any, site: View): Any
 }

vsim/Makefrag-verilog

@@ -12,9 +12,9 @@ $(generated_dir)/%.fir $(generated_dir)/%.prm $(generated_dir)/%.d: $(FIRRTL_JAR
 	mkdir -p $(dir $@)
 	cd $(base_dir) && $(SBT) "run-main $(PROJECT).Generator $(generated_dir) $(PROJECT) $(MODEL) $(CFG_PROJECT) $(CONFIG)"
 
-$(generated_dir)/$(long_name).v $(generated_dir)/$(long_name).conf : $(firrtl) $(FIRRTL_JAR)
+$(generated_dir)/%.v $(generated_dir)/%.conf: $(generated_dir)/%.fir $(FIRRTL_JAR)
 	mkdir -p $(dir $@)
-	$(FIRRTL) -i $< -o $(generated_dir)/$(long_name).v -X verilog --infer-rw $(MODEL) --repl-seq-mem -c:$(MODEL):-o:$(generated_dir)/$(long_name).conf
+	$(FIRRTL) -i $< -o $(generated_dir)/$*.v -X verilog --infer-rw $(MODEL) --repl-seq-mem -c:$(MODEL):-o:$(generated_dir)/$*.conf
 
 $(generated_dir)/$(long_name).behav_srams.v : $(generated_dir)/$(long_name).conf $(mem_gen)
 	cd $(generated_dir) && \