diff --git a/src/main/scala/coreplex/Configs.scala b/src/main/scala/coreplex/Configs.scala
index c320390c..0d932c04 100644
--- a/src/main/scala/coreplex/Configs.scala
+++ b/src/main/scala/coreplex/Configs.scala
@@ -45,7 +45,7 @@ class WithNBigCores(n: Int) extends Config((site, here, up) => {
       icache = Some(ICacheParams(
         rowBits = site(SystemBusKey).beatBits,
         blockBytes = site(CacheBlockBytes))))
-    List.fill(n)(big) ++ up(RocketTilesKey, site)
+    List.tabulate(n)(i => big.copy(hartid = i))
   }
 })
 
@@ -67,7 +67,7 @@ class WithNSmallCores(n: Int) extends Config((site, here, up) => {
         nWays = 1,
         nTLBEntries = 4,
         blockBytes = site(CacheBlockBytes))))
-    List.fill(n)(small) ++ up(RocketTilesKey, site)
+    List.tabulate(n)(i => small.copy(hartid = i))
   }
 })
 
@@ -94,7 +94,7 @@ class WithNTinyCores(n: Int) extends Config((site, here, up) => {
           nWays = 1,
           nTLBEntries = 4,
           blockBytes = site(CacheBlockBytes))))
-    List.fill(n)(tiny) ++ up(RocketTilesKey, site)
+    List.tabulate(n)(i => tiny.copy(hartid = i))
   }
 })
 
diff --git a/src/main/scala/coreplex/RocketCoreplex.scala b/src/main/scala/coreplex/RocketCoreplex.scala
index b4992a3c..e7009316 100644
--- a/src/main/scala/coreplex/RocketCoreplex.scala
+++ b/src/main/scala/coreplex/RocketCoreplex.scala
@@ -24,6 +24,7 @@ trait HasRocketTiles extends HasSystemBus
   private val crossing = p(RocketCrossing)
   private val tileParams = p(RocketTilesKey)
   val nRocketTiles = tileParams.size
+  val hartIdList = tileParams.map(_.hartid)
 
   // Handle interrupts to be routed directly into each tile
   // TODO: figure out how to merge the localIntNodes and coreIntXbar below
@@ -34,8 +35,8 @@ trait HasRocketTiles extends HasSystemBus
 
   // Make a wrapper for each tile that will wire it to coreplex devices and crossbars,
   // according to the specified type of clock crossing.
-  val wiringTuple = localIntNodes.zip(tileParams).zipWithIndex
-  val rocket_tiles: Seq[RocketTileWrapper] = wiringTuple.map { case ((lip, tp), i) =>
+  val wiringTuple = localIntNodes.zip(tileParams)
+  val rocket_tiles: Seq[RocketTileWrapper] = wiringTuple.map { case (lip, tp) =>
     val pWithExtra = p.alterPartial {
       case TileKey => tp
       case BuildRoCC => tp.rocc
@@ -44,19 +45,19 @@ trait HasRocketTiles extends HasSystemBus
 
     val wrapper = crossing match {
       case SynchronousCrossing(params) => {
-        val wrapper = LazyModule(new SyncRocketTile(tp, i)(pWithExtra))
+        val wrapper = LazyModule(new SyncRocketTile(tp)(pWithExtra))
         sbus.fromSyncTiles(params, tp.externalMasterBuffers, tp.name) :=* wrapper.masterNode
         FlipRendering { implicit p => wrapper.slaveNode :*= pbus.toSyncSlaves(tp.name, tp.externalSlaveBuffers) }
         wrapper
       }
       case AsynchronousCrossing(depth, sync) => {
-        val wrapper = LazyModule(new AsyncRocketTile(tp, i)(pWithExtra))
+        val wrapper = LazyModule(new AsyncRocketTile(tp)(pWithExtra))
         sbus.fromAsyncTiles(depth, sync, tp.externalMasterBuffers, tp.name) :=* wrapper.masterNode
         FlipRendering { implicit p => wrapper.slaveNode :*= pbus.toAsyncSlaves(sync, tp.name, tp.externalSlaveBuffers) }
         wrapper
       }
       case RationalCrossing(direction) => {
-        val wrapper = LazyModule(new RationalRocketTile(tp, i)(pWithExtra))
+        val wrapper = LazyModule(new RationalRocketTile(tp)(pWithExtra))
         sbus.fromRationalTiles(direction, tp.externalMasterBuffers, tp.name) :=* wrapper.masterNode
         FlipRendering { implicit p => wrapper.slaveNode :*= pbus.toRationalSlaves(tp.name, tp.externalSlaveBuffers) }
         wrapper
@@ -108,8 +109,12 @@ trait HasRocketTilesModuleImp extends LazyModuleImp
     with HasPeripheryDebugModuleImp {
   val outer: HasRocketTiles
 
-  // TODO make this less gross and/or support tiles with differently sized reset vectors
-  def resetVectorBits: Int = outer.paddrBits
+  def resetVectorBits: Int = {
+    // Consider using the minimum over all widths, rather than enforcing homogeneity
+    val vectors = outer.rocket_tiles.map(_.module.io.reset_vector)
+    require(vectors.tail.forall(_.getWidth == vectors.head.getWidth))
+    vectors.head.getWidth
+  }
   val rocket_tile_inputs = Wire(Vec(outer.nRocketTiles, new ClockedRocketTileInputs()(p.alterPartial {
     case SharedMemoryTLEdge => outer.sharedMemoryTLEdge
   })))
@@ -123,7 +128,7 @@ trait HasRocketTilesModuleImp extends LazyModuleImp
   }
 
   // Default values for tile inputs; may be overriden in other traits
-  rocket_tile_inputs.zipWithIndex.foreach { case(wire, i) =>
+  rocket_tile_inputs.zip(outer.hartIdList).foreach { case(wire, i) =>
     wire.clock := clock
     wire.reset := reset
     wire.hartid := UInt(i)
diff --git a/src/main/scala/coreplex/SystemBus.scala b/src/main/scala/coreplex/SystemBus.scala
index 6721ead8..7096b3e5 100644
--- a/src/main/scala/coreplex/SystemBus.scala
+++ b/src/main/scala/coreplex/SystemBus.scala
@@ -125,5 +125,4 @@ trait HasSystemBus extends HasInterruptBus {
   val sbus = LazyModule(new SystemBus(sbusParams))
 
   def sharedMemoryTLEdge: TLEdge = sbus.busView
-  def paddrBits: Int = sbus.busView.bundle.addressBits
 }
diff --git a/src/main/scala/devices/tilelink/Error.scala b/src/main/scala/devices/tilelink/Error.scala
index a1a4d945..cdbb95e0 100644
--- a/src/main/scala/devices/tilelink/Error.scala
+++ b/src/main/scala/devices/tilelink/Error.scala
@@ -59,7 +59,7 @@ class TLError(params: ErrorParams, beatBytes: Int = 4)(implicit p: Parameters) e
 
     val a_opcodes = Vec(AccessAck, AccessAck, AccessAckData, AccessAckData, AccessAckData, HintAck, Grant)
     da.bits.opcode  := a_opcodes(a.bits.opcode)
-    da.bits.param   := UInt(0)
+    da.bits.param   := UInt(0) // toT, but error grants must be handled transiently (ie: you don't keep permissions)
     da.bits.size    := a.bits.size
     da.bits.source  := a.bits.source
     da.bits.sink    := UInt(0)
@@ -70,7 +70,7 @@ class TLError(params: ErrorParams, beatBytes: Int = 4)(implicit p: Parameters) e
     dc.valid := c.valid && c_last
 
     dc.bits.opcode := ReleaseAck
-    dc.bits.param  := Vec(toN, toN, toB)(c.bits.param)
+    dc.bits.param  := Vec(toB, toN, toN)(c.bits.param)
     dc.bits.size   := c.bits.size
     dc.bits.source := c.bits.source
     dc.bits.sink   := UInt(0)
diff --git a/src/main/scala/diplomacy/Resources.scala b/src/main/scala/diplomacy/Resources.scala
index d62f829e..4a76f87b 100644
--- a/src/main/scala/diplomacy/Resources.scala
+++ b/src/main/scala/diplomacy/Resources.scala
@@ -6,11 +6,33 @@ import Chisel.log2Ceil
 import scala.collection.immutable.{ListMap,SortedMap}
 
 sealed trait ResourceValue
+
+/** Permission of an address space.
+  * @param r            readable.
+  * @param w            writable.
+  * @param x            executable.
+  * @param c            cacheable.
+  */
 case class ResourcePermissions(r: Boolean, w: Boolean, x: Boolean, c: Boolean) // Not part of DTS
+
+/** An address space description.
+  * @param address      the address space.
+  * @param permissions  the permission attributes of this space. See [[freechips.rocketchip.diplomacy.ResourcePermissions]].
+  */
 final case class ResourceAddress(address: Seq[AddressSet], permissions: ResourcePermissions) extends ResourceValue
+
+/** A mapped address space (eg: when map a device to a bus).
+  * @param address      the address space.
+  * @param offset       the address offset of the mapped device (eg: base address of the bus).
+  * @param permissions  the permission attributes of this space. See [[freechips.rocketchip.diplomacy.ResourcePermissions]].
+  */
 final case class ResourceMapping(address: Seq[AddressSet], offset: BigInt, permissions: ResourcePermissions) extends ResourceValue
 final case class ResourceInt(value: BigInt) extends ResourceValue
 final case class ResourceString(value: String) extends ResourceValue
+
+/** A reference pointing to another device in DTS (eg: interrupt to interrupt controller).
+  * @param value        the label (String) of the device.
+  */
 final case class ResourceReference(value: String) extends ResourceValue
 final case class ResourceMap(value: Map[String, Seq[ResourceValue]], labels: Seq[String] = Nil) extends ResourceValue
 
@@ -21,12 +43,17 @@ case class ResourceBindings(map: Map[String, Seq[Binding]])
   def apply(key: String): Seq[Binding] = map.getOrElse(key, Nil)
 }
 
+/** A serializable description of a device.
+  * @param name         the resolved name of this device. See [[freechips.rocketchip.diplomacy.DeviceRegName]].
+  * @param mapping      the property map of this device.
+  */
 case class Description(name: String, mapping: Map[String, Seq[ResourceValue]])
 
 abstract class Device
 {
   def describe(resources: ResourceBindings): Description
 
+  /** make sure all derived devices have an unique label */
   val label = "L" + Device.index.toString
   Device.index = Device.index + 1
 }
@@ -36,9 +63,12 @@ object Device
   private var index: Int = 0
 }
 
+/** A trait for devices that generate interrupts. */
 trait DeviceInterrupts
 {
   this: Device =>
+
+  /** Whether to always use the expanded interrupt description in DTS: "interrupts-extended" */
   val alwaysExtended = false
   def describeInterrupts(resources: ResourceBindings): Map[String, Seq[ResourceValue]] = {
     val int = resources("int")
@@ -64,6 +94,7 @@ trait DeviceInterrupts
   def int = Seq(Resource(this, "int"))
 }
 
+/** A trait for resolving the name of a device. */
 trait DeviceRegName
 {
   this: Device =>
@@ -94,14 +125,18 @@ trait DeviceRegName
   }
 }
 
+/** A simple device descriptor for devices that may support interrupts and address spaces.
+  * @param devname      the base device named used in device name generation.
+  * @param devcompat    a list of compatible devices. See device tree property "compatible".
+  */
 class SimpleDevice(devname: String, devcompat: Seq[String]) extends Device with DeviceInterrupts with DeviceRegName
 {
   def describe(resources: ResourceBindings): Description = {
-    val name = describeName(devname, resources)
-    val int = describeInterrupts(resources)
+    val name = describeName(devname, resources)  // the generated device name in device tree
+    val int = describeInterrupts(resources)      // interrupt description
 
     def optDef(x: String, seq: Seq[ResourceValue]) = if (seq.isEmpty) None else Some(x -> seq)
-    val compat = optDef("compatible", devcompat.map(ResourceString(_)))
+    val compat = optDef("compatible", devcompat.map(ResourceString(_))) // describe the list of compatiable devices
 
     val reg = resources.map.filterKeys(regFilter)
     val (named, bulk) = reg.partition { case (k, v) => regName(k).isDefined }
@@ -114,13 +149,18 @@ class SimpleDevice(devname: String, devcompat: Seq[String]) extends Device with
         require (false, s"DTS device $name has $k = $seq, must be a single ResourceAddress!")
     }
 
-    val names = optDef("reg-names", named.map(x => ResourceString(regName(x._1).get)).toList)
-    val regs = optDef("reg", (named ++ bulk).flatMap(_._2.map(_.value)).toList)
+    val names = optDef("reg-names", named.map(x => ResourceString(regName(x._1).get)).toList) // names of the named address space
+    val regs = optDef("reg", (named ++ bulk).flatMap(_._2.map(_.value)).toList) // address ranges of all spaces (named and bulk)
 
     Description(name, ListMap() ++ compat ++ int ++ names ++ regs)
   }
 }
 
+/** A simple bus
+  * @param devname      the base device named used in device name generation.
+  * @param devcompat    a list of compatible devices. See device tree property "compatible".
+  * @param offset       the base address of this bus.
+  */
 class SimpleBus(devname: String, devcompat: Seq[String], offset: BigInt = 0) extends SimpleDevice(devname, devcompat ++ Seq("simple-bus"))
 {
   override def describe(resources: ResourceBindings): Description = {
@@ -147,6 +187,7 @@ class SimpleBus(devname: String, devcompat: Seq[String], offset: BigInt = 0) ext
   def ranges = Seq(Resource(this, "ranges"))
 }
 
+/** A generic memory block. */
 class MemoryDevice extends Device with DeviceRegName
 {
   override val prefix = ""
@@ -169,12 +210,13 @@ case class Resource(owner: Device, key: String)
   }
 }
 
+/** The resource binding scope for a LazyModule that generates a device tree (currently Coreplex only). */
 trait BindingScope
 {
   this: LazyModule =>
 
   private val parentScope = BindingScope.find(parent)
-  protected[diplomacy] var resourceBindingFns: Seq[() => Unit] = Nil
+  protected[diplomacy] var resourceBindingFns: Seq[() => Unit] = Nil // callback functions to resolve resource binding during elaboration
   protected[diplomacy] var resourceBindings: Seq[(Resource, Option[Device], ResourceValue)] = Nil
 
   private case class ExpandedValue(path: Seq[String], labels: Seq[String], value: Seq[ResourceValue])
@@ -208,6 +250,7 @@ trait BindingScope
     }
   }
 
+  /** Generate the device tree. */
   def bindingTree: ResourceMap = {
     eval
     val map: Map[Device, ResourceBindings] =
@@ -232,6 +275,9 @@ object BindingScope
 
 object ResourceBinding
 {
+  /** Add a resource callback function to the callback list BindingScope.resourceBindingFns.
+    * @param block      the callback function to be added.
+    */
   def apply(block: => Unit) {
     val scope = BindingScope.find()
     require (scope.isDefined, "ResourceBinding must be called from within a BindingScope")
diff --git a/src/main/scala/rocket/CSR.scala b/src/main/scala/rocket/CSR.scala
index c3a05712..ab136f95 100644
--- a/src/main/scala/rocket/CSR.scala
+++ b/src/main/scala/rocket/CSR.scala
@@ -63,7 +63,7 @@ class DCSR extends Bundle {
 }
 
 class MIP(implicit p: Parameters) extends CoreBundle()(p)
-    with HasRocketCoreParameters {
+    with HasCoreParameters {
   val lip = Vec(coreParams.nLocalInterrupts, Bool())
   val zero2 = Bool()
   val debug = Bool() // keep in sync with CSR.debugIntCause
@@ -144,7 +144,7 @@ object CSR
 }
 
 class PerfCounterIO(implicit p: Parameters) extends CoreBundle
-    with HasRocketCoreParameters {
+    with HasCoreParameters {
   val eventSel = UInt(OUTPUT, xLen)
   val inc = UInt(INPUT, log2Ceil(1+retireWidth))
 }
@@ -161,7 +161,7 @@ class TracedInstruction(implicit p: Parameters) extends CoreBundle {
 }
 
 class CSRFileIO(implicit p: Parameters) extends CoreBundle
-    with HasRocketCoreParameters {
+    with HasCoreParameters {
   val interrupts = new TileInterrupts().asInput
   val hartid = UInt(INPUT, hartIdLen)
   val rw = new Bundle {
@@ -190,7 +190,6 @@ class CSRFileIO(implicit p: Parameters) extends CoreBundle
   val evec = UInt(OUTPUT, vaddrBitsExtended)
   val exception = Bool(INPUT)
   val retire = UInt(INPUT, log2Up(1+retireWidth))
-  val custom_mrw_csrs = Vec(nCustomMrwCsrs, UInt(INPUT, xLen))
   val cause = UInt(INPUT, xLen)
   val pc = UInt(INPUT, vaddrBitsExtended)
   val badaddr = UInt(INPUT, vaddrBitsExtended)
@@ -208,7 +207,7 @@ class CSRFileIO(implicit p: Parameters) extends CoreBundle
 }
 
 class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Parameters) extends CoreModule()(p)
-    with HasRocketCoreParameters {
+    with HasCoreParameters {
   val io = new CSRFileIO
 
   val reset_mstatus = Wire(init=new MStatus().fromBits(0))
@@ -223,7 +222,6 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
   reset_dcsr.xdebugver := 1
   reset_dcsr.prv := PRV.M
   val reg_dcsr = Reg(init=reset_dcsr)
-  val reg_debugint = Reg(Bool())
 
   val (supported_interrupts, delegable_interrupts) = {
     val sup = Wire(new MIP)
@@ -305,13 +303,17 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
   val hpm_mask = reg_mcounteren & Mux((!usingVM).B || reg_mstatus.prv === PRV.S, delegable_counters.U, reg_scounteren)
 
   val mip = Wire(init=reg_mip)
+  mip.lip := (io.interrupts.lip: Seq[Bool])
+  mip.mtip := io.interrupts.mtip
+  mip.msip := io.interrupts.msip
+  mip.meip := io.interrupts.meip
   // seip is the OR of reg_mip.seip and the actual line from the PLIC
   io.interrupts.seip.foreach { mip.seip := reg_mip.seip || RegNext(_) }
   mip.rocc := io.rocc_interrupt
   val read_mip = mip.asUInt & supported_interrupts
 
   val pending_interrupts = read_mip & reg_mie
-  val d_interrupts = reg_debugint << CSR.debugIntCause
+  val d_interrupts = io.interrupts.debug << CSR.debugIntCause
   val m_interrupts = Mux(reg_mstatus.prv <= PRV.S || (reg_mstatus.prv === PRV.M && reg_mstatus.mie), pending_interrupts & ~reg_mideleg, UInt(0))
   val s_interrupts = Mux(reg_mstatus.prv < PRV.S || (reg_mstatus.prv === PRV.S && reg_mstatus.sie), pending_interrupts & reg_mideleg, UInt(0))
   val (anyInterrupt, whichInterrupt) = chooseInterrupt(Seq(s_interrupts, m_interrupts, d_interrupts))
@@ -343,8 +345,6 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
     CSRs.mimpid -> UInt(0),
     CSRs.marchid -> UInt(0),
     CSRs.mvendorid -> UInt(0),
-    CSRs.mcycle -> reg_cycle,
-    CSRs.minstret -> reg_instret,
     CSRs.misa -> reg_misa,
     CSRs.mstatus -> read_mstatus,
     CSRs.mtvec -> reg_mtvec,
@@ -372,14 +372,34 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
   if (usingFPU)
     read_mapping ++= fp_csrs
 
-  for (((e, c), i) <- (reg_hpmevent.padTo(CSR.nHPM, UInt(0))
-                       zip reg_hpmcounter.map(x => x: UInt).padTo(CSR.nHPM, UInt(0))) zipWithIndex) {
-    read_mapping += (i + CSR.firstHPE) -> e // mhpmeventN
-    read_mapping += (i + CSR.firstMHPC) -> c // mhpmcounterN
-    if (usingUser) read_mapping += (i + CSR.firstHPC) -> c // hpmcounterN
+  if (coreParams.haveBasicCounters) {
+    read_mapping += CSRs.mcycle -> reg_cycle
+    read_mapping += CSRs.minstret -> reg_instret
+
+    for (((e, c), i) <- (reg_hpmevent.padTo(CSR.nHPM, UInt(0))
+                         zip reg_hpmcounter.map(x => x: UInt).padTo(CSR.nHPM, UInt(0))) zipWithIndex) {
+      read_mapping += (i + CSR.firstHPE) -> e // mhpmeventN
+      read_mapping += (i + CSR.firstMHPC) -> c // mhpmcounterN
+      if (usingUser) read_mapping += (i + CSR.firstHPC) -> c // hpmcounterN
+      if (xLen == 32) {
+        read_mapping += (i + CSR.firstMHPCH) -> c // mhpmcounterNh
+        if (usingUser) read_mapping += (i + CSR.firstHPCH) -> c // hpmcounterNh
+      }
+    }
+
+    if (usingUser) {
+      read_mapping += CSRs.mcounteren -> reg_mcounteren
+      read_mapping += CSRs.cycle -> reg_cycle
+      read_mapping += CSRs.instret -> reg_instret
+    }
+
     if (xLen == 32) {
-      read_mapping += (i + CSR.firstMHPCH) -> c // mhpmcounterNh
-      if (usingUser) read_mapping += (i + CSR.firstHPCH) -> c // hpmcounterNh
+      read_mapping += CSRs.mcycleh -> (reg_cycle >> 32)
+      read_mapping += CSRs.minstreth -> (reg_instret >> 32)
+      if (usingUser) {
+        read_mapping += CSRs.cycleh -> (reg_cycle >> 32)
+        read_mapping += CSRs.instreth -> (reg_instret >> 32)
+      }
     }
   }
 
@@ -412,21 +432,6 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
     read_mapping += CSRs.medeleg -> reg_medeleg
   }
 
-  if (usingUser) {
-    read_mapping += CSRs.mcounteren -> reg_mcounteren
-    read_mapping += CSRs.cycle -> reg_cycle
-    read_mapping += CSRs.instret -> reg_instret
-  }
-
-  if (xLen == 32) {
-    read_mapping += CSRs.mcycleh -> (reg_cycle >> 32)
-    read_mapping += CSRs.minstreth -> (reg_instret >> 32)
-    if (usingUser) {
-      read_mapping += CSRs.cycleh -> (reg_cycle >> 32)
-      read_mapping += CSRs.instreth -> (reg_instret >> 32)
-    }
-  }
-
   val pmpCfgPerCSR = xLen / new PMPConfig().getWidth
   def pmpCfgIndex(i: Int) = (xLen / 32) * (i / pmpCfgPerCSR)
   if (reg_pmp.nonEmpty) {
@@ -438,13 +443,6 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
       read_mapping += (CSRs.pmpaddr0 + i) -> pmp.addr
   }
 
-  for (i <- 0 until nCustomMrwCsrs) {
-    val addr = 0xff0 + i
-    require(addr < (1 << CSR.ADDRSZ))
-    require(!read_mapping.contains(addr), "custom MRW CSR address " + i + " is already in use")
-    read_mapping += addr -> io.custom_mrw_csrs(i)
-  }
-
   val decoded_addr = read_mapping map { case (k, v) => k -> (io.rw.addr === k) }
   val wdata = readModifyWriteCSR(io.rw.cmd, io.rw.rdata, io.rw.wdata)
 
@@ -511,7 +509,7 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
   assert(PopCount(insn_ret :: insn_call :: insn_break :: io.exception :: Nil) <= 1, "these conditions must be mutually exclusive")
 
   when (insn_wfi && !io.singleStep && !reg_debug) { reg_wfi := true }
-  when (pending_interrupts.orR || exception || reg_debugint) { reg_wfi := false }
+  when (pending_interrupts.orR || exception || io.interrupts.debug) { reg_wfi := false }
   assert(!reg_wfi || io.retire === UInt(0))
 
   when (io.retire(0) || exception) { reg_singleStepped := true }
@@ -611,7 +609,8 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
     when (decoded_addr(CSRs.misa)) {
       val mask = UInt(isaStringToMask(isaMaskString), xLen)
       val f = wdata('f' - 'a')
-      reg_misa := ~(~wdata | (!f << ('d' - 'a'))) & mask | reg_misa & ~mask
+      if (coreParams.misaWritable)
+        reg_misa := ~(~wdata | (!f << ('d' - 'a'))) & mask | reg_misa & ~mask
     }
     when (decoded_addr(CSRs.mip)) {
       // MIP should be modified based on the value in reg_mip, not the value
@@ -637,8 +636,10 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
       writeCounter(i + CSR.firstMHPC, c, wdata)
       when (decoded_addr(i + CSR.firstHPE)) { e := perfEventSets.maskEventSelector(wdata) }
     }
-    writeCounter(CSRs.mcycle, reg_cycle, wdata)
-    writeCounter(CSRs.minstret, reg_instret, wdata)
+    if (coreParams.haveBasicCounters) {
+      writeCounter(CSRs.mcycle, reg_cycle, wdata)
+      writeCounter(CSRs.minstret, reg_instret, wdata)
+    }
 
     if (usingFPU) {
       when (decoded_addr(CSRs.fflags)) { reg_fflags := wdata }
@@ -721,12 +722,6 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
     }
   }
 
-  reg_mip.lip := (io.interrupts.lip: Seq[Bool])
-  reg_mip.mtip := io.interrupts.mtip
-  reg_mip.msip := io.interrupts.msip
-  reg_mip.meip := io.interrupts.meip
-  reg_debugint := io.interrupts.debug
-
   if (!usingVM) {
     reg_mideleg := 0
     reg_medeleg := 0
diff --git a/src/main/scala/rocket/DCache.scala b/src/main/scala/rocket/DCache.scala
index 7f3b3b63..e3daa973 100644
--- a/src/main/scala/rocket/DCache.scala
+++ b/src/main/scala/rocket/DCache.scala
@@ -384,7 +384,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   val a_size = mtSize(s2_req.typ)
   val a_data = Fill(beatWords, pstore1_data)
   val acquire = if (edge.manager.anySupportAcquireB) {
-    edge.Acquire(UInt(0), acquire_address, lgCacheBlockBytes, s2_grow_param)._2 // Cacheability checked by tlb
+    edge.AcquireBlock(UInt(0), acquire_address, lgCacheBlockBytes, s2_grow_param)._2 // Cacheability checked by tlb
   } else {
     Wire(new TLBundleA(edge.bundle))
   }
diff --git a/src/main/scala/rocket/Frontend.scala b/src/main/scala/rocket/Frontend.scala
index 7319e7ea..408088c3 100644
--- a/src/main/scala/rocket/Frontend.scala
+++ b/src/main/scala/rocket/Frontend.scala
@@ -75,7 +75,6 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   val icache = outer.icache.module
   require(fetchWidth*coreInstBytes == outer.icacheParams.fetchBytes)
 
-  val fetchBytes = coreInstBytes * fetchWidth
   val tlb = Module(new TLB(true, log2Ceil(fetchBytes), nTLBEntries))
   val fq = withReset(reset || io.cpu.req.valid) { Module(new ShiftQueue(new FrontendResp, 5, flow = true)) }
 
diff --git a/src/main/scala/rocket/NBDcache.scala b/src/main/scala/rocket/NBDcache.scala
index 8d9899d4..cf0fa9bb 100644
--- a/src/main/scala/rocket/NBDcache.scala
+++ b/src/main/scala/rocket/NBDcache.scala
@@ -276,7 +276,7 @@ class MSHR(id: Int)(implicit edge: TLEdgeOut, p: Parameters) extends L1HellaCach
   io.wb_req.bits.voluntary := Bool(true)
 
   io.mem_acquire.valid := state === s_refill_req && grantackq.io.enq.ready
-  io.mem_acquire.bits := edge.Acquire(
+  io.mem_acquire.bits := edge.AcquireBlock(
                                 fromSource = UInt(id),
                                 toAddress = Cat(io.tag, req_idx) << blockOffBits,
                                 lgSize = lgCacheBlockBytes,
diff --git a/src/main/scala/rocket/PMP.scala b/src/main/scala/rocket/PMP.scala
index 2e277a77..02d69aab 100644
--- a/src/main/scala/rocket/PMP.scala
+++ b/src/main/scala/rocket/PMP.scala
@@ -127,7 +127,7 @@ class PMPHomogeneityChecker(pmps: Seq[PMP])(implicit p: Parameters) {
 }
 
 class PMPChecker(lgMaxSize: Int)(implicit p: Parameters) extends CoreModule()(p)
-    with HasRocketCoreParameters {
+    with HasCoreParameters {
   val io = new Bundle {
     val prv = UInt(INPUT, PRV.SZ)
     val pmp = Vec(nPMPs, new PMP).asInput
diff --git a/src/main/scala/rocket/PTW.scala b/src/main/scala/rocket/PTW.scala
index 4618715f..70ec7b12 100644
--- a/src/main/scala/rocket/PTW.scala
+++ b/src/main/scala/rocket/PTW.scala
@@ -24,7 +24,7 @@ class PTWResp(implicit p: Parameters) extends CoreBundle()(p) {
 }
 
 class TLBPTWIO(implicit p: Parameters) extends CoreBundle()(p)
-    with HasRocketCoreParameters {
+    with HasCoreParameters {
   val req = Decoupled(new PTWReq)
   val resp = Valid(new PTWResp).flip
   val ptbr = new PTBR().asInput
@@ -37,7 +37,7 @@ class PTWPerfEvents extends Bundle {
 }
 
 class DatapathPTWIO(implicit p: Parameters) extends CoreBundle()(p)
-    with HasRocketCoreParameters {
+    with HasCoreParameters {
   val ptbr = new PTBR().asInput
   val sfence = Valid(new SFenceReq).flip
   val status = new MStatus().asInput
diff --git a/src/main/scala/rocket/RocketCore.scala b/src/main/scala/rocket/RocketCore.scala
index bbed2edd..f5c4b78f 100644
--- a/src/main/scala/rocket/RocketCore.scala
+++ b/src/main/scala/rocket/RocketCore.scala
@@ -23,7 +23,8 @@ case class RocketCoreParams(
   nBreakpoints: Int = 1,
   nPMPs: Int = 8,
   nPerfCounters: Int = 0,
-  nCustomMRWCSRs: Int = 0,
+  haveBasicCounters: Boolean = true,
+  misaWritable: Boolean = true,
   nL2TLBEntries: Int = 0,
   mtvecInit: Option[BigInt] = Some(BigInt(0)),
   mtvecWritable: Boolean = true,
@@ -46,12 +47,6 @@ trait HasRocketCoreParameters extends HasCoreParameters {
 
   val fastLoadWord = rocketParams.fastLoadWord
   val fastLoadByte = rocketParams.fastLoadByte
-  val nBreakpoints = rocketParams.nBreakpoints
-  val nPMPs = rocketParams.nPMPs
-  val nPerfCounters = rocketParams.nPerfCounters
-  val nCustomMrwCsrs = rocketParams.nCustomMRWCSRs
-  val mtvecInit = rocketParams.mtvecInit
-  val mtvecWritable = rocketParams.mtvecWritable
 
   val mulDivParams = rocketParams.mulDiv.getOrElse(MulDivParams()) // TODO ask andrew about this
 
diff --git a/src/main/scala/system/Configs.scala b/src/main/scala/system/Configs.scala
index 187e127f..1d5fe758 100644
--- a/src/main/scala/system/Configs.scala
+++ b/src/main/scala/system/Configs.scala
@@ -63,9 +63,6 @@ class EightChannelConfig extends Config(new WithNMemoryChannels(8) ++ new BaseCo
 class DualCoreConfig extends Config(
   new WithNBigCores(2) ++ new BaseConfig)
 
-class HeterogeneousDualCoreConfig extends Config(
-  new WithNSmallCores(1) ++ new WithNBigCores(1) ++ new BaseConfig)
-
 class TinyConfig extends Config(
   new WithNMemoryChannels(0) ++
   new WithStatelessBridge ++
diff --git a/src/main/scala/tile/Core.scala b/src/main/scala/tile/Core.scala
index da400df9..26e20647 100644
--- a/src/main/scala/tile/Core.scala
+++ b/src/main/scala/tile/Core.scala
@@ -25,7 +25,14 @@ trait CoreParams {
   val retireWidth: Int
   val instBits: Int
   val nLocalInterrupts: Int
+  val nPMPs: Int
+  val nBreakpoints: Int
+  val nPerfCounters: Int
+  val haveBasicCounters: Boolean
+  val misaWritable: Boolean
   val nL2TLBEntries: Int
+  val mtvecInit: Option[BigInt]
+  val mtvecWritable: Boolean
   val jumpInFrontend: Boolean
   val tileControlAddr: Option[BigInt]
 
@@ -47,12 +54,19 @@ trait HasCoreParameters extends HasTileParameters {
   val fetchWidth = coreParams.fetchWidth
   val decodeWidth = coreParams.decodeWidth
 
+  val fetchBytes = coreParams.fetchBytes
   val coreInstBits = coreParams.instBits
   val coreInstBytes = coreInstBits/8
   val coreDataBits = xLen max fLen
   val coreDataBytes = coreDataBits/8
   val coreMaxAddrBits = paddrBits max vaddrBitsExtended
 
+  val nBreakpoints = coreParams.nBreakpoints
+  val nPMPs = coreParams.nPMPs
+  val nPerfCounters = coreParams.nPerfCounters
+  val mtvecInit = coreParams.mtvecInit
+  val mtvecWritable = coreParams.mtvecWritable
+
   val coreDCacheReqTagBits = 6
   val dcacheReqTagBits = coreDCacheReqTagBits + log2Ceil(dcacheArbPorts)
 
diff --git a/src/main/scala/tile/FPU.scala b/src/main/scala/tile/FPU.scala
index 8e89a7f3..f64fbd2c 100644
--- a/src/main/scala/tile/FPU.scala
+++ b/src/main/scala/tile/FPU.scala
@@ -542,7 +542,72 @@ class FPToFP(val latency: Int)(implicit p: Parameters) extends FPUModule()(p) {
   io.out <> Pipe(in.valid, mux, latency-1)
 }
 
+class MulAddRecFNPipe(latency: Int, expWidth: Int, sigWidth: Int) extends Module
+{
+    require(latency<=2) 
+
+    val io = new Bundle {
+        val validin = Bool(INPUT)
+        val op = Bits(INPUT, 2)
+        val a = Bits(INPUT, expWidth + sigWidth + 1)
+        val b = Bits(INPUT, expWidth + sigWidth + 1)
+        val c = Bits(INPUT, expWidth + sigWidth + 1)
+        val roundingMode   = UInt(INPUT, 3)
+        val detectTininess = UInt(INPUT, 1)
+        val out = Bits(OUTPUT, expWidth + sigWidth + 1)
+        val exceptionFlags = Bits(OUTPUT, 5)
+        val validout = Bool(OUTPUT)
+    }
+
+    //------------------------------------------------------------------------
+    //------------------------------------------------------------------------
+    val mulAddRecFNToRaw_preMul =
+        Module(new hardfloat.MulAddRecFNToRaw_preMul(expWidth, sigWidth))
+    val mulAddRecFNToRaw_postMul =
+        Module(new hardfloat.MulAddRecFNToRaw_postMul(expWidth, sigWidth))
+
+    mulAddRecFNToRaw_preMul.io.op := io.op
+    mulAddRecFNToRaw_preMul.io.a  := io.a
+    mulAddRecFNToRaw_preMul.io.b  := io.b
+    mulAddRecFNToRaw_preMul.io.c  := io.c
+
+    val mulAddResult =
+        (mulAddRecFNToRaw_preMul.io.mulAddA *
+             mulAddRecFNToRaw_preMul.io.mulAddB) +&
+            mulAddRecFNToRaw_preMul.io.mulAddC
+
+    val valid_stage0 = Wire(Bool())
+    val roundingMode_stage0 = Wire(UInt(width=3))
+    val detectTininess_stage0 = Wire(UInt(width=1))
+  
+    val postmul_regs = if(latency>0) 1 else 0
+    mulAddRecFNToRaw_postMul.io.fromPreMul   := Pipe(io.validin, mulAddRecFNToRaw_preMul.io.toPostMul, postmul_regs).bits
+    mulAddRecFNToRaw_postMul.io.mulAddResult := Pipe(io.validin, mulAddResult, postmul_regs).bits
+    mulAddRecFNToRaw_postMul.io.roundingMode := Pipe(io.validin, io.roundingMode, postmul_regs).bits
+    roundingMode_stage0                      := Pipe(io.validin, io.roundingMode, postmul_regs).bits
+    detectTininess_stage0                    := Pipe(io.validin, io.detectTininess, postmul_regs).bits
+    valid_stage0                             := Pipe(io.validin, false.B, postmul_regs).valid
+    
+    //------------------------------------------------------------------------
+    //------------------------------------------------------------------------
+    val roundRawFNToRecFN = Module(new hardfloat.RoundRawFNToRecFN(expWidth, sigWidth, 0))
+
+    val round_regs = if(latency==2) 1 else 0
+    roundRawFNToRecFN.io.invalidExc         := Pipe(valid_stage0, mulAddRecFNToRaw_postMul.io.invalidExc, round_regs).bits
+    roundRawFNToRecFN.io.in                 := Pipe(valid_stage0, mulAddRecFNToRaw_postMul.io.rawOut, round_regs).bits
+    roundRawFNToRecFN.io.roundingMode       := Pipe(valid_stage0, roundingMode_stage0, round_regs).bits
+    roundRawFNToRecFN.io.detectTininess     := Pipe(valid_stage0, detectTininess_stage0, round_regs).bits
+    io.validout                             := Pipe(valid_stage0, false.B, round_regs).valid
+
+    roundRawFNToRecFN.io.infiniteExc := Bool(false)
+
+    io.out            := roundRawFNToRecFN.io.out
+    io.exceptionFlags := roundRawFNToRecFN.io.exceptionFlags
+}
+
 class FPUFMAPipe(val latency: Int, val t: FType)(implicit p: Parameters) extends FPUModule()(p) {
+  require(latency>0)
+
   val io = new Bundle {
     val in = Valid(new FPInput).flip
     val out = Valid(new FPResult)
@@ -560,7 +625,8 @@ class FPUFMAPipe(val latency: Int, val t: FType)(implicit p: Parameters) extends
     when (!(cmd_fma || cmd_addsub)) { in.in3 := zero }
   }
 
-  val fma = Module(new hardfloat.MulAddRecFN(t.exp, t.sig))
+  val fma = Module(new MulAddRecFNPipe((latency-1) min 2, t.exp, t.sig))
+  fma.io.validin := valid
   fma.io.op := in.fmaCmd
   fma.io.roundingMode := in.rm
   fma.io.detectTininess := hardfloat.consts.tininess_afterRounding
@@ -571,7 +637,8 @@ class FPUFMAPipe(val latency: Int, val t: FType)(implicit p: Parameters) extends
   val res = Wire(new FPResult)
   res.data := sanitizeNaN(fma.io.out, t)
   res.exc := fma.io.exceptionFlags
-  io.out := Pipe(valid, res, latency-1)
+
+  io.out := Pipe(fma.io.validout, res, (latency-3) max 0)
 }
 
 class FPU(cfg: FPUParams)(implicit p: Parameters) extends FPUModule()(p) {
diff --git a/src/main/scala/tile/RocketTile.scala b/src/main/scala/tile/RocketTile.scala
index b9aaa542..9d9b7b94 100644
--- a/src/main/scala/tile/RocketTile.scala
+++ b/src/main/scala/tile/RocketTile.scala
@@ -22,13 +22,14 @@ case class RocketTileParams(
     trace: Boolean = false,
     hcfOnUncorrectable: Boolean = false,
     name: Option[String] = Some("tile"),
+    hartid: Int = 0,
     externalMasterBuffers: Int = 0,
     externalSlaveBuffers: Int = 0) extends TileParams {
   require(icache.isDefined)
   require(dcache.isDefined)
 }
   
-class RocketTile(val rocketParams: RocketTileParams, val hartid: Int)(implicit p: Parameters) extends BaseTile(rocketParams)(p)
+class RocketTile(val rocketParams: RocketTileParams)(implicit p: Parameters) extends BaseTile(rocketParams)(p)
     with HasExternalInterrupts
     with HasLazyRoCC  // implies CanHaveSharedFPU with CanHavePTW with HasHellaCache
     with CanHaveScratchpad { // implies CanHavePTW with HasHellaCache with HasICacheFrontend
@@ -39,6 +40,7 @@ class RocketTile(val rocketParams: RocketTileParams, val hartid: Int)(implicit p
   private def ofStr(x: String) = Seq(ResourceString(x))
   private def ofRef(x: Device) = Seq(ResourceReference(x.label))
 
+  val hartid = rocketParams.hartid
   val cpuDevice = new Device {
     def describe(resources: ResourceBindings): Description = {
       val block =  p(CacheBlockBytes)
@@ -179,8 +181,8 @@ class RocketTileModule(outer: RocketTile) extends BaseTileModule(outer, () => ne
   ptw.io.requestor <> ptwPorts
 }
 
-abstract class RocketTileWrapper(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends LazyModule {
-  val rocket = LazyModule(new RocketTile(rtp, hartid))
+abstract class RocketTileWrapper(rtp: RocketTileParams)(implicit p: Parameters) extends LazyModule {
+  val rocket = LazyModule(new RocketTile(rtp))
   val asyncIntNode   : IntInwardNode
   val periphIntNode  : IntInwardNode
   val coreIntNode    : IntInwardNode
@@ -226,7 +228,7 @@ abstract class RocketTileWrapper(rtp: RocketTileParams, hartid: Int)(implicit p:
   }
 }
 
-class SyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends RocketTileWrapper(rtp, hartid) {
+class SyncRocketTile(rtp: RocketTileParams)(implicit p: Parameters) extends RocketTileWrapper(rtp) {
   val masterNode = optionalMasterBuffer(rocket.masterNode)
   val slaveNode = optionalSlaveBuffer(rocket.slaveNode)
 
@@ -246,7 +248,7 @@ class SyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters)
   def outputInterruptXingLatency = 0
 }
 
-class AsyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends RocketTileWrapper(rtp, hartid) {
+class AsyncRocketTile(rtp: RocketTileParams)(implicit p: Parameters) extends RocketTileWrapper(rtp) {
   val source = LazyModule(new TLAsyncCrossingSource)
   source.node :=* rocket.masterNode
   val masterNode = source.node
@@ -272,7 +274,7 @@ class AsyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters
   def outputInterruptXingLatency = 3
 }
 
-class RationalRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends RocketTileWrapper(rtp, hartid) {
+class RationalRocketTile(rtp: RocketTileParams)(implicit p: Parameters) extends RocketTileWrapper(rtp) {
   val source = LazyModule(new TLRationalCrossingSource)
   source.node :=* optionalMasterBuffer(rocket.masterNode)
   val masterNode = source.node
diff --git a/src/main/scala/tilelink/Atomics.scala b/src/main/scala/tilelink/Atomics.scala
index 5f412cc6..370817a3 100644
--- a/src/main/scala/tilelink/Atomics.scala
+++ b/src/main/scala/tilelink/Atomics.scala
@@ -51,8 +51,8 @@ class Atomics(params: TLBundleParameters) extends Module
     UInt(3),   // LogicalData
     UInt(0),   // Get
     UInt(0),   // Hint
-    UInt(0),   // Acquire
-    UInt(0)))( // Overwrite
+    UInt(0),   // AcquireBlock
+    UInt(0)))( // AcquirePerm
     io.a.opcode))
 
   // Only the masked bytes can be modified
diff --git a/src/main/scala/tilelink/Broadcast.scala b/src/main/scala/tilelink/Broadcast.scala
index 3e203a13..4681b8f3 100644
--- a/src/main/scala/tilelink/Broadcast.scala
+++ b/src/main/scala/tilelink/Broadcast.scala
@@ -176,6 +176,11 @@ class TLBroadcast(lineBytes: Int, numTrackers: Int = 4, bufferless: Boolean = fa
         t.probe := (if (caches.size == 0) UInt(0) else Mux(a_cache.orR(), UInt(caches.size-1), UInt(caches.size)))
       }
 
+      val acq_perms = MuxLookup(in.a.bits.param, Wire(UInt(width = 2)), Array(
+        TLPermissions.NtoB -> TLPermissions.toB,
+        TLPermissions.NtoT -> TLPermissions.toN,
+        TLPermissions.BtoT -> TLPermissions.toN))
+
       when (in.a.fire() && a_first) {
         probe_todo  := ~a_cache // probe all but the cache who poked us
         probe_line  := in.a.bits.address >> lineShift
@@ -188,10 +193,8 @@ class TLBroadcast(lineBytes: Int, numTrackers: Int = 4, bufferless: Boolean = fa
           TLMessages.Hint           -> MuxLookup(in.a.bits.param, Wire(UInt(width = 2)), Array(
             TLHints.PREFETCH_READ   -> TLPermissions.toB,
             TLHints.PREFETCH_WRITE  -> TLPermissions.toN)),
-          TLMessages.Acquire        -> MuxLookup(in.a.bits.param, Wire(UInt(width = 2)), Array(
-            TLPermissions.NtoB      -> TLPermissions.toB,
-            TLPermissions.NtoT      -> TLPermissions.toN,
-            TLPermissions.BtoT      -> TLPermissions.toN))))
+          TLMessages.AcquireBlock   -> acq_perms,
+          TLMessages.AcquirePerm    -> acq_perms))
       }
 
       // The outer TL connections may not be cached
@@ -236,7 +239,7 @@ class TLBroadcastTracker(id: Int, lineBytes: Int, probeCountBits: Int, bufferles
   when (io.in_a.fire() && io.in_a_first) {
     assert (idle)
     sent_d  := Bool(false)
-    got_e   := io.in_a.bits.opcode =/= TLMessages.Acquire
+    got_e   := io.in_a.bits.opcode =/= TLMessages.AcquireBlock && io.in_a.bits.opcode =/= TLMessages.AcquirePerm
     opcode  := io.in_a.bits.opcode
     param   := io.in_a.bits.param
     size    := io.in_a.bits.size
@@ -271,7 +274,7 @@ class TLBroadcastTracker(id: Int, lineBytes: Int, probeCountBits: Int, bufferles
   i_data.bits.data := io.in_a.bits.data
 
   val probe_done = count === UInt(0)
-  val acquire = opcode === TLMessages.Acquire
+  val acquire = opcode === TLMessages.AcquireBlock || opcode === TLMessages.AcquirePerm
 
   val transform = MuxLookup(param, Wire(UInt(width = 2)), Array(
     TLPermissions.NtoB -> TRANSFORM_B,
diff --git a/src/main/scala/tilelink/Bundles.scala b/src/main/scala/tilelink/Bundles.scala
index d58bc907..118b01e5 100644
--- a/src/main/scala/tilelink/Bundles.scala
+++ b/src/main/scala/tilelink/Bundles.scala
@@ -22,7 +22,8 @@ object TLMessages
   def LogicalData    = UInt(3) //     .    .                   => AccessAckData
   def Get            = UInt(4) //     .    .                   => AccessAckData
   def Hint           = UInt(5) //     .    .                   => HintAck
-  def Acquire        = UInt(6) //     .                        => Grant[Data]
+  def AcquireBlock   = UInt(6) //     .                        => Grant[Data]
+  def AcquirePerm    = UInt(7) //     .                        => Grant[Data]
   def Probe          = UInt(6) //          .                   => ProbeAck[Data]
   def AccessAck      = UInt(0) //               .    .
   def AccessAckData  = UInt(1) //               .    .
@@ -36,7 +37,7 @@ object TLMessages
   def ReleaseAck     = UInt(6) //                    .
   def GrantAck       = UInt(0) //                         .
  
-  def isA(x: UInt) = x <= Acquire
+  def isA(x: UInt) = x <= AcquirePerm
   def isB(x: UInt) = x <= Probe
   def isC(x: UInt) = x <= ReleaseData
   def isD(x: UInt) = x <= ReleaseAck
diff --git a/src/main/scala/tilelink/CacheCork.scala b/src/main/scala/tilelink/CacheCork.scala
index 9503e97d..063c58a4 100644
--- a/src/main/scala/tilelink/CacheCork.scala
+++ b/src/main/scala/tilelink/CacheCork.scala
@@ -47,7 +47,8 @@ class TLCacheCork(unsafe: Boolean = false)(implicit p: Parameters) extends LazyM
       val a_a = Wire(out.a)
       val a_d = Wire(in.d)
       val isPut = in.a.bits.opcode === PutFullData || in.a.bits.opcode === PutPartialData
-      val toD = in.a.bits.opcode === Acquire && in.a.bits.param === TLPermissions.BtoT
+      val toD = (in.a.bits.opcode === AcquireBlock && in.a.bits.param === TLPermissions.BtoT) ||
+                (in.a.bits.opcode === AcquirePerm)
       in.a.ready := Mux(toD, a_d.ready, a_a.ready)
 
       a_a.valid := in.a.valid && !toD
@@ -55,7 +56,7 @@ class TLCacheCork(unsafe: Boolean = false)(implicit p: Parameters) extends LazyM
       a_a.bits.source := in.a.bits.source << 1 | Mux(isPut, UInt(1), UInt(0))
 
       // Transform Acquire into Get
-      when (in.a.bits.opcode === Acquire) {
+      when (in.a.bits.opcode === AcquireBlock || in.a.bits.opcode === AcquirePerm) {
         a_a.bits.opcode := Get
         a_a.bits.param  := UInt(0)
         a_a.bits.source := in.a.bits.source << 1 | UInt(1)
@@ -100,10 +101,7 @@ class TLCacheCork(unsafe: Boolean = false)(implicit p: Parameters) extends LazyM
 
       when (out.d.bits.opcode === AccessAckData && out.d.bits.source(0)) {
         d_d.bits.opcode := GrantData
-        // On Grant error, you do NOT get the permissions you asked for.
-        // We only enter this case from NtoT or NtoB, so that means use toN.
-        // (the BtoT case was handled by a_d)
-        d_d.bits.param  := Mux(out.d.bits.error, TLPermissions.toN, TLPermissions.toT)
+        d_d.bits.param := TLPermissions.toT
       }
       when (out.d.bits.opcode === AccessAck && !out.d.bits.source(0)) {
         d_d.bits.opcode := ReleaseAck
diff --git a/src/main/scala/tilelink/Edges.scala b/src/main/scala/tilelink/Edges.scala
index 55c21711..4f59c9a3 100644
--- a/src/main/scala/tilelink/Edges.scala
+++ b/src/main/scala/tilelink/Edges.scala
@@ -271,11 +271,25 @@ class TLEdgeOut(
   extends TLEdge(client, manager, params, sourceInfo)
 {
   // Transfers
-  def Acquire(fromSource: UInt, toAddress: UInt, lgSize: UInt, growPermissions: UInt) = {
+  def AcquireBlock(fromSource: UInt, toAddress: UInt, lgSize: UInt, growPermissions: UInt) = {
     require (manager.anySupportAcquireB)
     val legal = manager.supportsAcquireBFast(toAddress, lgSize)
     val a = Wire(new TLBundleA(bundle))
-    a.opcode  := TLMessages.Acquire
+    a.opcode  := TLMessages.AcquireBlock
+    a.param   := growPermissions
+    a.size    := lgSize
+    a.source  := fromSource
+    a.address := toAddress
+    a.mask    := mask(toAddress, lgSize)
+    a.data    := UInt(0)
+    (legal, a)
+  }
+
+  def AcquirePerm(fromSource: UInt, toAddress: UInt, lgSize: UInt, growPermissions: UInt) = {
+    require (manager.anySupportAcquireB)
+    val legal = manager.supportsAcquireBFast(toAddress, lgSize)
+    val a = Wire(new TLBundleA(bundle))
+    a.opcode  := TLMessages.AcquirePerm
     a.param   := growPermissions
     a.size    := lgSize
     a.source  := fromSource
diff --git a/src/main/scala/tilelink/Monitor.scala b/src/main/scala/tilelink/Monitor.scala
index 11ddd0ec..9b258684 100644
--- a/src/main/scala/tilelink/Monitor.scala
+++ b/src/main/scala/tilelink/Monitor.scala
@@ -37,14 +37,25 @@ class TLMonitor(args: TLMonitorArgs) extends TLMonitorBase(args)
     val is_aligned = edge.isAligned(bundle.address, bundle.size)
     val mask = edge.full_mask(bundle)
 
-    when (bundle.opcode === TLMessages.Acquire) {
-      assert (edge.manager.supportsAcquireBSafe(edge.address(bundle), bundle.size), "'A' channel carries Acquire type unsupported by manager" + extra)
-      assert (edge.client.supportsProbe(edge.source(bundle), bundle.size), "'A' channel carries Acquire from a client which does not support Probe" + extra)
-      assert (source_ok, "'A' channel Acquire carries invalid source ID" + extra)
-      assert (bundle.size >= UInt(log2Ceil(edge.manager.beatBytes)), "'A' channel Acquire smaller than a beat" + extra)
-      assert (is_aligned, "'A' channel Acquire address not aligned to size" + extra)
-      assert (TLPermissions.isGrow(bundle.param), "'A' channel Acquire carries invalid grow param" + extra)
-      assert (~bundle.mask === UInt(0), "'A' channel Acquire contains invalid mask" + extra)
+    when (bundle.opcode === TLMessages.AcquireBlock) {
+      assert (edge.manager.supportsAcquireBSafe(edge.address(bundle), bundle.size), "'A' channel carries AcquireBlock type unsupported by manager" + extra)
+      assert (edge.client.supportsProbe(edge.source(bundle), bundle.size), "'A' channel carries AcquireBlock from a client which does not support Probe" + extra)
+      assert (source_ok, "'A' channel AcquireBlock carries invalid source ID" + extra)
+      assert (bundle.size >= UInt(log2Ceil(edge.manager.beatBytes)), "'A' channel AcquireBlock smaller than a beat" + extra)
+      assert (is_aligned, "'A' channel AcquireBlock address not aligned to size" + extra)
+      assert (TLPermissions.isGrow(bundle.param), "'A' channel AcquireBlock carries invalid grow param" + extra)
+      assert (~bundle.mask === UInt(0), "'A' channel AcquireBlock contains invalid mask" + extra)
+    }
+
+    when (bundle.opcode === TLMessages.AcquirePerm) {
+      assert (edge.manager.supportsAcquireBSafe(edge.address(bundle), bundle.size), "'A' channel carries AcquirePerm type unsupported by manager" + extra)
+      assert (edge.client.supportsProbe(edge.source(bundle), bundle.size), "'A' channel carries AcquirePerm from a client which does not support Probe" + extra)
+      assert (source_ok, "'A' channel AcquirePerm carries invalid source ID" + extra)
+      assert (bundle.size >= UInt(log2Ceil(edge.manager.beatBytes)), "'A' channel AcquirePerm smaller than a beat" + extra)
+      assert (is_aligned, "'A' channel AcquirePerm address not aligned to size" + extra)
+      assert (TLPermissions.isGrow(bundle.param), "'A' channel AcquirePerm carries invalid grow param" + extra)
+      assert (bundle.param =/= TLPermissions.NtoB, "'A' channel AcquirePerm requests NtoB" + extra)
+      assert (~bundle.mask === UInt(0), "'A' channel AcquirePerm contains invalid mask" + extra)
     }
 
     when (bundle.opcode === TLMessages.Get) {
diff --git a/src/main/scala/tilelink/RAMModel.scala b/src/main/scala/tilelink/RAMModel.scala
index c121c002..21c5cf14 100644
--- a/src/main/scala/tilelink/RAMModel.scala
+++ b/src/main/scala/tilelink/RAMModel.scala
@@ -116,7 +116,7 @@ class TLRAMModel(log: String = "")(implicit p: Parameters) extends LazyModule
 
       when (a_fire) {
         // Record the request so we can handle it's response
-        assert (a.opcode =/= TLMessages.Acquire)
+        assert (a.opcode =/= TLMessages.AcquireBlock && a.opcode =/= TLMessages.AcquirePerm)
 
         // Mark the operation as valid
         valid(a.source) := Bool(true)