Merge branch 'master' into ma-fetch

src/main/scala/coreplex/RocketCoreplex.scala

@@ -84,6 +84,8 @@ trait HasRocketTiles extends HasSystemBus
     lip.foreach { coreIntXbar.intnode := _ }                // lip
     wrapper.coreIntNode   := coreIntXbar.intnode
 
+    wrapper.intOutputNode.foreach { plic.intnode := _ }
+
     wrapper
   }
 }

src/main/scala/groundtest/TraceGen.scala

@@ -65,6 +65,7 @@ case class TraceGenParams(
     memStart: BigInt, //p(ExtMem).base
     numGens: Int) extends GroundTestTileParams {
   def build(i: Int, p: Parameters): GroundTestTile = new TraceGenTile(i, this)(p)
+  val trace = false
 }
 
 trait HasTraceGenParams {

src/main/scala/rocket/BusErrorUnit.scala

@@ -0,0 +1,76 @@
+// See LICENSE.SiFive for license details.
+
+package freechips.rocketchip.rocket
+
+import Chisel._
+import Chisel.ImplicitConversions._
+import chisel3.util.Valid
+import freechips.rocketchip.config.Parameters
+import freechips.rocketchip.util._
+import freechips.rocketchip.tile._
+import freechips.rocketchip.diplomacy._
+import freechips.rocketchip.regmapper._
+import freechips.rocketchip.tilelink._
+
+trait BusErrors extends Bundle {
+  def toErrorList: List[Option[Valid[UInt]]]
+}
+
+class L1BusErrors(implicit p: Parameters) extends CoreBundle()(p) with BusErrors {
+  val icache = new ICacheErrors
+  val dcache = new DCacheErrors
+
+  def toErrorList = 
+    List(None, None, icache.correctable, icache.uncorrectable,
+         None, Some(dcache.bus), dcache.correctable, dcache.uncorrectable)
+}
+
+case class BusErrorUnitParams(addr: BigInt, size: Int = 4096)
+
+class BusErrorUnit[T <: BusErrors](t: => T, params: BusErrorUnitParams)(implicit p: Parameters) extends LazyModule {
+  val regWidth = 64
+  val device = new SimpleDevice("bus-error-unit", Seq("sifive,buserror0"))
+  val intNode = IntSourceNode(IntSourcePortSimple(resources = device.int))
+  val node = TLRegisterNode(
+    address   = Seq(AddressSet(params.addr, params.size-1)),
+    device    = device,
+    beatBytes = p(XLen)/8)
+
+  lazy val module = new LazyModuleImp(this) {
+    val io = new Bundle {
+      val tl = node.bundleIn
+      val interrupt = intNode.bundleOut
+      val errors = t.flip
+    }
+
+    val sources = io.errors.toErrorList
+    val mask = sources.map(_.nonEmpty.B).asUInt
+    val cause = Reg(init = UInt(0, log2Ceil(sources.lastIndexWhere(_.nonEmpty) + 1)))
+    val value = Reg(UInt(width = sources.flatten.map(_.bits.getWidth).max))
+    require(value.getWidth <= regWidth)
+    val enable = Reg(init = mask)
+    val interrupt = Reg(init = UInt(0, sources.size))
+    val accrued = Reg(init = UInt(0, sources.size))
+
+    accrued := accrued | sources.map(_.map(_.valid).getOrElse(false.B)).asUInt
+
+    for ((s, i) <- sources.zipWithIndex; if s.nonEmpty) {
+      when (s.get.valid && enable(i) && cause === 0) {
+        cause := i
+        value := s.get.bits
+      }
+    }
+
+    io.interrupt.head(0) := (accrued & interrupt).orR
+
+    def reg(r: UInt) = RegField(regWidth, r)
+    def maskedReg(r: UInt, m: UInt) = RegField(regWidth, r, RegWriteFn((v, d) => { when (v) { r := d & m }; true }))
+
+    node.regmap(
+      0 -> Seq(reg(cause),
+               reg(value),
+               maskedReg(enable, mask),
+               maskedReg(interrupt, mask),
+               maskedReg(accrued, mask)))
+  }
+}

src/main/scala/rocket/CSR.scala

@@ -149,6 +149,16 @@ class PerfCounterIO(implicit p: Parameters) extends CoreBundle
   val inc = UInt(INPUT, log2Ceil(1+retireWidth))
 }
 
+class TracedInstruction(implicit p: Parameters) extends CoreBundle {
+  val valid = Bool()
+  val addr = UInt(width = coreMaxAddrBits)
+  val insn = UInt(width = iLen)
+  val priv = UInt(width = 3)
+  val exception = Bool()
+  val cause = UInt(width = 1 + log2Ceil(xLen))
+  val tval = UInt(width = coreMaxAddrBits max iLen)
+}
+
 class CSRFileIO(implicit p: Parameters) extends CoreBundle
     with HasRocketCoreParameters {
   val interrupts = new TileInterrupts().asInput
@@ -192,6 +202,8 @@ class CSRFileIO(implicit p: Parameters) extends CoreBundle
   val bp = Vec(nBreakpoints, new BP).asOutput
   val pmp = Vec(nPMPs, new PMP).asOutput
   val counters = Vec(nPerfCounters, new PerfCounterIO)
+  val inst = Vec(retireWidth, UInt(width = iLen)).asInput
+  val trace = Vec(retireWidth, new TracedInstruction).asOutput
 }
 
 class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Parameters) extends CoreModule()(p)
@@ -506,15 +518,14 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
   assert(!io.singleStep || io.retire <= UInt(1))
   assert(!reg_singleStepped || io.retire === UInt(0))
 
+  val epc = ~(~io.pc | (coreInstBytes-1))
+  val write_badaddr = cause isOneOf (Causes.illegal_instruction, Causes.breakpoint,
+    Causes.misaligned_load, Causes.misaligned_store,
+    Causes.load_access, Causes.store_access, Causes.fetch_access,
+    Causes.load_page_fault, Causes.store_page_fault, Causes.fetch_page_fault)
+  val badaddr_value = Mux(write_badaddr, io.badaddr, 0.U)
+
   when (exception) {
-    val epc = ~(~io.pc | (coreInstBytes-1))
-
-    val write_badaddr = cause isOneOf (Causes.illegal_instruction, Causes.breakpoint,
-      Causes.misaligned_load, Causes.misaligned_store,
-      Causes.load_access, Causes.store_access, Causes.fetch_access,
-      Causes.load_page_fault, Causes.store_page_fault, Causes.fetch_page_fault)
-    val badaddr_value = Mux(write_badaddr, io.badaddr, 0.U)
-
     when (trapToDebug) {
       when (!reg_debug) {
         reg_debug := true
@@ -756,6 +767,16 @@ class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Param
     }
   }
 
+  for (((t, insn), i) <- (io.trace zip io.inst).zipWithIndex) {
+    t.exception := io.retire >= i && exception
+    t.valid := io.retire > i || t.exception
+    t.insn := insn
+    t.addr := io.pc
+    t.priv := Cat(reg_debug, reg_mstatus.prv)
+    t.cause := Cat(cause(xLen-1), cause(log2Ceil(xLen)-1, 0))
+    t.tval := badaddr_value
+  }
+
   def chooseInterrupt(masks: Seq[UInt]): (Bool, UInt) = {
     val nonstandard = supported_interrupts.getWidth-1 to 12 by -1
     // MEI, MSI, MTI, SEI, SSI, STI, UEI, USI, UTI

src/main/scala/rocket/DCache.scala

@@ -11,6 +11,12 @@ import freechips.rocketchip.tilelink._
 import freechips.rocketchip.util._
 import TLMessages._
 
+class DCacheErrors(implicit p: Parameters) extends L1HellaCacheBundle()(p) {
+  val correctable = (cacheParams.tagECC.canCorrect || cacheParams.dataECC.canCorrect).option(Valid(UInt(width = paddrBits)))
+  val uncorrectable = (cacheParams.tagECC.canDetect || cacheParams.dataECC.canDetect).option(Valid(UInt(width = paddrBits)))
+  val bus = Valid(UInt(width = paddrBits))
+}
+
 class DCacheDataReq(implicit p: Parameters) extends L1HellaCacheBundle()(p) {
   val eccBytes = cacheParams.dataECCBytes
   val addr = Bits(width = untagBits)
@@ -216,9 +222,11 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   val s2_flush_valid_pre_tag_ecc = RegNext(s1_flush_valid)
   val s1_meta_decoded = s1_meta.map(tECC.decode(_))
   val s1_meta_clk_en = s1_valid_not_nacked || s1_flush_valid || s1_probe
-  val s2_meta_errors = s1_meta_decoded.map(m => RegEnable(m.error, s1_meta_clk_en)).asUInt
+  val s2_meta_correctable_errors = s1_meta_decoded.map(m => RegEnable(m.correctable, s1_meta_clk_en)).asUInt
+  val s2_meta_uncorrectable_errors = s1_meta_decoded.map(m => RegEnable(m.uncorrectable, s1_meta_clk_en)).asUInt
+  val s2_meta_error_uncorrectable = s2_meta_uncorrectable_errors.orR
   val s2_meta_corrected = s1_meta_decoded.map(m => RegEnable(m.corrected, s1_meta_clk_en).asTypeOf(new L1Metadata))
-  val s2_meta_error = s2_meta_errors.orR
+  val s2_meta_error = (s2_meta_uncorrectable_errors | s2_meta_correctable_errors).orR
   val s2_flush_valid = s2_flush_valid_pre_tag_ecc && !s2_meta_error
   val s2_data = {
     val en = s1_valid || inWriteback || tl_out.d.fire()
@@ -242,6 +250,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   val s2_word_idx = s2_req.addr.extract(log2Up(rowBits/8)-1, log2Up(wordBytes))
   val s2_did_read = RegEnable(s1_did_read, s1_valid_not_nacked)
   val s2_data_error = s2_did_read && (s2_data_decoded.map(_.error).grouped(wordBits/eccBits).map(_.reduce(_||_)).toSeq)(s2_word_idx)
+  val s2_data_error_uncorrectable = (s2_data_decoded.map(_.uncorrectable).grouped(wordBits/eccBits).map(_.reduce(_||_)).toSeq)(s2_word_idx)
   val s2_data_corrected = (s2_data_decoded.map(_.corrected): Seq[UInt]).asUInt
   val s2_data_uncorrected = (s2_data_decoded.map(_.uncorrected): Seq[UInt]).asUInt
   val s2_valid_hit_pre_data_ecc = s2_valid_masked && s2_readwrite && !s2_meta_error && s2_hit
@@ -264,9 +273,10 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   // tag updates on ECC errors
   metaArb.io.in(1).valid := s2_meta_error && (s2_valid_masked || s2_flush_valid_pre_tag_ecc || s2_probe)
   metaArb.io.in(1).bits.write := true
-  metaArb.io.in(1).bits.way_en := PriorityEncoderOH(s2_meta_errors)
+  metaArb.io.in(1).bits.way_en := s2_meta_uncorrectable_errors | Mux(s2_meta_error_uncorrectable, 0.U, PriorityEncoderOH(s2_meta_correctable_errors))
   metaArb.io.in(1).bits.addr := Cat(io.cpu.req.bits.addr >> untagBits, Mux(s2_probe, probe_bits.address, s2_req.addr)(idxMSB, 0))
-  metaArb.io.in(1).bits.data := PriorityMux(s2_meta_errors, s2_meta_corrected)
+  metaArb.io.in(1).bits.data := PriorityMux(s2_meta_correctable_errors, s2_meta_corrected)
+  when (s2_meta_error_uncorrectable) { metaArb.io.in(1).bits.data.coh := ClientMetadata.onReset }
 
   // tag updates on hit/miss
   metaArb.io.in(2).valid := (s2_valid_hit && s2_update_meta) || (s2_victimize && !s2_victim_dirty)
@@ -703,6 +713,30 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   io.cpu.perf.release := edge.done(tl_out_c)
   io.cpu.perf.tlbMiss := io.ptw.req.fire()
 
+  // report errors
+  {
+    val (data_error, data_error_uncorrectable, data_error_addr) =
+      if (usingDataScratchpad) (s2_valid_data_error, s2_data_error_uncorrectable, s2_req.addr) else {
+        (tl_out_c.valid && edge.hasData(tl_out_c.bits) && s2_data_decoded.map(_.error).reduce(_||_),
+         s2_data_decoded.map(_.uncorrectable).reduce(_||_),
+         tl_out_c.bits.address)
+      }
+    val error_addr =
+      Mux(metaArb.io.in(1).valid, Cat(metaArb.io.in(1).bits.data.tag, metaArb.io.in(1).bits.addr(untagBits-1, idxLSB)),
+          data_error_addr >> idxLSB) << idxLSB
+    io.errors.uncorrectable.foreach { u =>
+      u.valid := metaArb.io.in(1).valid && s2_meta_error_uncorrectable || data_error && data_error_uncorrectable
+      u.bits := error_addr
+    }
+    io.errors.correctable.foreach { c =>
+      c.valid := metaArb.io.in(1).valid || data_error
+      c.bits := error_addr
+      io.errors.uncorrectable.foreach { u => when (u.valid) { c.valid := false } }
+    }
+    io.errors.bus.valid := tl_out.d.fire() && tl_out.d.bits.error
+    io.errors.bus.bits := Mux(grantIsCached, s2_req.addr >> idxLSB << idxLSB, 0.U)
+  }
+
   def encodeData(x: UInt) = x.grouped(eccBits).map(dECC.encode(_)).asUInt
   def dummyEncodeData(x: UInt) = x.grouped(eccBits).map(dECC.swizzle(_)).asUInt
   def decodeData(x: UInt) = x.grouped(dECC.width(eccBits)).map(dECC.decode(_))

src/main/scala/rocket/Frontend.scala

@@ -70,6 +70,7 @@ class FrontendBundle(outer: Frontend) extends CoreBundle()(outer.p)
   val ptw = new TLBPTWIO()
   val tl_out = outer.masterNode.bundleOut
   val tl_in = outer.slaveNode.bundleIn
+  val errors = new ICacheErrors
 }
 
 class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
@@ -286,6 +287,7 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   // performance events
   io.cpu.perf := icache.io.perf
   io.cpu.perf.tlbMiss := io.ptw.req.fire()
+  io.errors := icache.io.errors
 
   def alignPC(pc: UInt) = ~(~pc | (coreInstBytes - 1))
 }

src/main/scala/rocket/HellaCache.scala

@@ -177,6 +177,7 @@ class HellaCacheBundle(outer: HellaCache)(implicit p: Parameters) extends CoreBu
   val cpu = (new HellaCacheIO).flip
   val ptw = new TLBPTWIO()
   val mem = outer.node.bundleOut
+  val errors = new DCacheErrors
 }
 
 class HellaCacheModule(outer: HellaCache) extends LazyModuleImp(outer)

src/main/scala/rocket/IBuf.scala

@@ -16,7 +16,6 @@ class Instruction(implicit val p: Parameters) extends ParameterizedBundle with H
   val rvc = Bool()
   val inst = new ExpandedInstruction
   val raw = UInt(width = 32)
-  val cinst = UInt(width = 32)
   require(coreInstBits == (if (usingCompressed) 16 else 32))
 }
 
@@ -96,7 +95,6 @@ class IBuf(implicit p: Parameters) extends CoreModule {
     exp.io.in := curInst
     io.inst(i).bits.inst := exp.io.out
     io.inst(i).bits.raw := curInst
-    io.inst(i).bits.cinst := Mux(exp.io.rvc, curInst & 0xFFFF, curInst)
 
     if (usingCompressed) {
       val replay = ic_replay(j) || (!exp.io.rvc && (btbHitMask(j) || ic_replay(j+1)))

src/main/scala/rocket/ICache.scala

@@ -36,6 +36,11 @@ class ICacheReq(implicit p: Parameters) extends CoreBundle()(p) with HasL1ICache
   val addr = UInt(width = vaddrBits)
 }
 
+class ICacheErrors(implicit p: Parameters) extends CoreBundle()(p) with HasL1ICacheParameters {
+  val correctable = (cacheParams.tagECC.canDetect || cacheParams.dataECC.canDetect).option(Valid(UInt(width = paddrBits)))
+  val uncorrectable = (cacheParams.itimAddr.nonEmpty && cacheParams.dataECC.canDetect).option(Valid(UInt(width = paddrBits)))
+}
+
 class ICache(val icacheParams: ICacheParams, val hartid: Int)(implicit p: Parameters) extends LazyModule {
   lazy val module = new ICacheModule(this)
   val masterNode = TLClientNode(TLClientParameters(
@@ -87,6 +92,7 @@ class ICacheBundle(outer: ICache) extends CoreBundle()(outer.p) {
   val tl_out = outer.masterNode.bundleOut
   val tl_in = outer.slaveNode.map(_.bundleIn)
 
+  val errors = new ICacheErrors
   val perf = new ICachePerfEvents().asOutput
 }
 
@@ -116,10 +122,10 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
   val scratchpadOn = RegInit(false.B)
   val scratchpadMax = tl_in.map(tl => Reg(UInt(width = log2Ceil(nSets * (nWays - 1)))))
   def lineInScratchpad(line: UInt) = scratchpadMax.map(scratchpadOn && line <= _).getOrElse(false.B)
-  def addrMaybeInScratchpad(addr: UInt) = if (outer.icacheParams.itimAddr.isEmpty) false.B else {
+  val scratchpadBase = outer.icacheParams.itimAddr.map { dummy =>
-    val base = GetPropertyByHartId(p(RocketTilesKey), _.icache.flatMap(_.itimAddr.map(_.U)), io.hartid)
+    GetPropertyByHartId(p(RocketTilesKey), _.icache.flatMap(_.itimAddr.map(_.U)), io.hartid)
-    addr >= base && addr < base + outer.size
   }
+  def addrMaybeInScratchpad(addr: UInt) = scratchpadBase.map(base => addr >= base && addr < base + outer.size).getOrElse(false.B)
   def addrInScratchpad(addr: UInt) = addrMaybeInScratchpad(addr) && lineInScratchpad(addr(untagBits+log2Ceil(nWays)-1, blockOffBits))
   def scratchpadWay(addr: UInt) = addr.extract(untagBits+log2Ceil(nWays)-1, untagBits)
   def scratchpadWayValid(way: UInt) = way < nWays - 1
@@ -248,12 +254,15 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
       io.resp.valid := s1_valid && s1_hit
 
     case 2 =>
-      val s2_tag_hit = RegEnable(s1_tag_hit, s1_valid || s1_slaveValid)
+      val s1_clk_en = s1_valid || s1_slaveValid
-      val s2_dout = RegEnable(s1_dout, s1_valid || s1_slaveValid)
+      val s2_tag_hit = RegEnable(s1_tag_hit, s1_clk_en)
+      val s2_hit_way = OHToUInt(s2_tag_hit)
+      val s2_scratchpad_word_addr = Cat(s2_hit_way, io.s2_vaddr(untagBits-1, log2Ceil(wordBits/8)), UInt(0, log2Ceil(wordBits/8)))
+      val s2_dout = RegEnable(s1_dout, s1_clk_en)
       val s2_way_mux = Mux1H(s2_tag_hit, s2_dout)
 
-      val s2_tag_disparity = RegEnable(s1_tag_disparity, s1_valid || s1_slaveValid).asUInt.orR
+      val s2_tag_disparity = RegEnable(s1_tag_disparity, s1_clk_en).asUInt.orR
-      val s2_tl_error = RegEnable(s1_tl_error.asUInt.orR, s1_valid || s1_slaveValid)
+      val s2_tl_error = RegEnable(s1_tl_error.asUInt.orR, s1_clk_en)
       val s2_data_decoded = dECC.decode(s2_way_mux)
       val s2_disparity = s2_tag_disparity || s2_data_decoded.error
       when (s2_valid && s2_disparity) { invalidate := true }
@@ -263,6 +272,17 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
       io.resp.bits.replay := s2_disparity
       io.resp.valid := s2_valid && s2_hit
 
+      val s1_scratchpad_hit = Mux(s1_slaveValid, lineInScratchpad(scratchpadLine(s1s3_slaveAddr)), addrInScratchpad(io.s1_paddr))
+      val s2_scratchpad_hit = RegEnable(s1_scratchpad_hit, s1_clk_en)
+      io.errors.correctable.foreach { c =>
+        c.valid := s2_valid && Mux(s2_scratchpad_hit, s2_data_decoded.correctable, s2_disparity)
+        c.bits := 0.U
+      }
+      io.errors.uncorrectable.foreach { u =>
+        u.valid := s2_valid && s2_scratchpad_hit && s2_data_decoded.uncorrectable
+        u.bits := scratchpadBase.get + s2_scratchpad_word_addr
+      }
+
       tl_in.map { tl =>
         val respValid = RegInit(false.B)
         tl.a.ready := !(tl_out.d.valid || s1_slaveValid || s2_slaveValid || s3_slaveValid || respValid)
@@ -290,7 +310,7 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
           // a structural hazard on s1s3_slaveData/s1s3_slaveAddress).
           s3_slaveValid := true
           s1s3_slaveData := s2_data_decoded.corrected
-          s1s3_slaveAddr := Cat(OHToUInt(s2_tag_hit), io.s2_vaddr(untagBits-1, log2Ceil(wordBits/8)), s1s3_slaveAddr(log2Ceil(wordBits/8)-1, 0))
+          s1s3_slaveAddr := s2_scratchpad_word_addr | s1s3_slaveAddr(log2Ceil(wordBits/8)-1, 0)
         }
 
         respValid := s2_slaveValid || (respValid && !tl.d.ready)

src/main/scala/rocket/RocketCore.scala

@@ -30,6 +30,7 @@ case class RocketCoreParams(
   fastLoadWord: Boolean = true,
   fastLoadByte: Boolean = false,
   jumpInFrontend: Boolean = true,
+  tileControlAddr: Option[BigInt] = None,
   mulDiv: Option[MulDivParams] = Some(MulDivParams()),
   fpu: Option[FPUParams] = Some(FPUParams())
 ) extends CoreParams {
@@ -136,7 +137,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   val ex_reg_replay = Reg(Bool())
   val ex_reg_pc = Reg(UInt())
   val ex_reg_inst = Reg(Bits())
-  val ex_reg_cinst = Reg(Bits())
+  val ex_reg_raw_inst = Reg(UInt())
 
   val mem_reg_xcpt_interrupt  = Reg(Bool())
   val mem_reg_valid           = Reg(Bool())
@@ -153,7 +154,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   val mem_reg_sfence = Reg(Bool())
   val mem_reg_pc = Reg(UInt())
   val mem_reg_inst = Reg(Bits())
-  val mem_reg_cinst = Reg(Bits())
+  val mem_reg_raw_inst = Reg(UInt())
   val mem_reg_wdata = Reg(Bits())
   val mem_reg_rs2 = Reg(Bits())
   val take_pc_mem = Wire(Bool())
@@ -167,7 +168,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   val wb_reg_sfence = Reg(Bool())
   val wb_reg_pc = Reg(UInt())
   val wb_reg_inst = Reg(Bits())
-  val wb_reg_cinst = Reg(Bits())
+  val wb_reg_raw_inst = Reg(UInt())
   val wb_reg_wdata = Reg(Bits())
   val wb_reg_rs2 = Reg(Bits())
   val take_pc_wb = Wire(Bool())
@@ -178,7 +179,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   // decode stage
   val ibuf = Module(new IBuf)
   val id_expanded_inst = ibuf.io.inst.map(_.bits.inst)
-  val id_nonexpanded_inst = ibuf.io.inst.map(_.bits.cinst)
+  val id_raw_inst = ibuf.io.inst.map(_.bits.raw)
   val id_inst = id_expanded_inst.map(_.bits)
   ibuf.io.imem <> io.imem.resp
   ibuf.io.kill := take_pc
@@ -335,7 +336,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
       }
     }
     when (id_illegal_insn) {
-      val inst = Mux(ibuf.io.inst(0).bits.rvc, ibuf.io.inst(0).bits.raw(15, 0), ibuf.io.inst(0).bits.raw)
+      val inst = Mux(ibuf.io.inst(0).bits.rvc, id_raw_inst(0)(15, 0), id_raw_inst(0))
       ex_reg_rs_bypass(0) := false
       ex_reg_rs_lsb(0) := inst(log2Ceil(bypass_sources.size)-1, 0)
       ex_reg_rs_msb(0) := inst >> log2Ceil(bypass_sources.size)
@@ -344,7 +345,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   when (!ctrl_killd || csr.io.interrupt || ibuf.io.inst(0).bits.replay) {
     ex_reg_cause := id_cause
     ex_reg_inst := id_inst(0)
-    ex_reg_cinst := id_nonexpanded_inst(0)
+    ex_reg_raw_inst := id_raw_inst(0)
     ex_reg_pc := ibuf.io.pc
     ex_reg_btb_resp := ibuf.io.btb_resp
   }
@@ -405,7 +406,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
 
     mem_reg_cause := ex_cause
     mem_reg_inst := ex_reg_inst
-    mem_reg_cinst := ex_reg_cinst
+    mem_reg_raw_inst := ex_reg_raw_inst
     mem_reg_pc := ex_reg_pc
     mem_reg_wdata := alu.io.out
 
@@ -453,7 +454,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
     }
     wb_reg_cause := mem_cause
     wb_reg_inst := mem_reg_inst
-    wb_reg_cinst := mem_reg_cinst
+    wb_reg_raw_inst := mem_reg_raw_inst
     wb_reg_pc := mem_reg_pc
   }
 
@@ -513,10 +514,11 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   when (rf_wen) { rf.write(rf_waddr, rf_wdata) }
 
   // hook up control/status regfile
-  csr.io.decode.csr := ibuf.io.inst(0).bits.raw(31,20)
+  csr.io.decode.csr := id_raw_inst(0)(31,20)
   csr.io.exception := wb_xcpt
   csr.io.cause := wb_cause
   csr.io.retire := wb_valid
+  csr.io.inst(0) := Cat(Mux(wb_reg_rvc, 0.U, wb_reg_inst >> 16), wb_reg_raw_inst(15, 0))
   csr.io.interrupts := io.interrupts
   csr.io.hartid := io.hartid
   io.fpu.fcsr_rm := csr.io.fcsr_rm
@@ -530,6 +532,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   csr.io.rw.addr := wb_reg_inst(31,20)
   csr.io.rw.cmd := Mux(wb_reg_valid, wb_ctrl.csr, CSR.N)
   csr.io.rw.wdata := wb_reg_wdata
+  io.trace := csr.io.trace
 
   val hazard_targets = Seq((id_ctrl.rxs1 && id_raddr1 =/= UInt(0), id_raddr1),
                            (id_ctrl.rxs2 && id_raddr2 =/= UInt(0), id_raddr2),
@@ -664,28 +667,24 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   csr.io.counters foreach { c => c.inc := RegNext(perfEvents.evaluate(c.eventSel)) }
 
   if (enableCommitLog) {
-    val pc = Wire(SInt(width=xLen))
+    val t = csr.io.trace(0)
-    pc := wb_reg_pc.asSInt
+    val rd = wb_waddr
-    val inst = wb_reg_inst
-    val cinst = wb_reg_cinst
-    val rd = RegNext(RegNext(RegNext(id_waddr)))
     val wfd = wb_ctrl.wfd
     val wxd = wb_ctrl.wxd
     val has_data = wb_wen && !wb_set_sboard
-    val priv = csr.io.status.prv
 
-    when (wb_valid) {
+    when (t.valid) {
       when (wfd) {
-        printf ("%d 0x%x (0x%x) f%d p%d 0xXXXXXXXXXXXXXXXX\n", priv, pc, cinst, rd, rd+UInt(32))
+        printf ("%d 0x%x (0x%x) f%d p%d 0xXXXXXXXXXXXXXXXX\n", t.priv, t.addr, t.insn, rd, rd+UInt(32))
       }
       .elsewhen (wxd && rd =/= UInt(0) && has_data) {
-        printf ("%d 0x%x (0x%x) x%d 0x%x\n", priv, pc, cinst, rd, rf_wdata)
+        printf ("%d 0x%x (0x%x) x%d 0x%x\n", t.priv, t.addr, t.insn, rd, rf_wdata)
       }
       .elsewhen (wxd && rd =/= UInt(0) && !has_data) {
-        printf ("%d 0x%x (0x%x) x%d p%d 0xXXXXXXXXXXXXXXXX\n", priv, pc, cinst, rd, rd)
+        printf ("%d 0x%x (0x%x) x%d p%d 0xXXXXXXXXXXXXXXXX\n", t.priv, t.addr, t.insn, rd, rd)
       }
       .otherwise {
-        printf ("%d 0x%x (0x%x)\n", priv, pc, cinst)
+        printf ("%d 0x%x (0x%x)\n", t.priv, t.addr, t.insn)
       }
     }
 
@@ -695,11 +694,11 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   }
   else {
     printf("C%d: %d [%d] pc=[%x] W[r%d=%x][%d] R[r%d=%x] R[r%d=%x] inst=[%x] DASM(%x)\n",
-         io.hartid, csr.io.time(31,0), wb_valid, wb_reg_pc,
+         io.hartid, csr.io.time(31,0), csr.io.trace(0).valid, csr.io.trace(0).addr(vaddrBitsExtended-1, 0),
          Mux(rf_wen && !(wb_set_sboard && wb_wen), rf_waddr, UInt(0)), rf_wdata, rf_wen,
          wb_reg_inst(19,15), Reg(next=Reg(next=ex_rs(0))),
          wb_reg_inst(24,20), Reg(next=Reg(next=ex_rs(1))),
-         wb_reg_cinst, wb_reg_cinst)
+         csr.io.trace(0).insn, csr.io.trace(0).insn)
   }
 
   val max_core_cycles = PlusArg("max-core-cycles",

src/main/scala/rocket/ScratchpadSlavePort.scala

@@ -100,21 +100,32 @@ trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend {
   val module: CanHaveScratchpadModule
   val cacheBlockBytes = p(CacheBlockBytes)
 
-  val slaveNode = TLInputNode() // Up to two uses for this input node:
-
-  // 1) Frontend always exists, but may or may not have a scratchpad node
-  // 2) ScratchpadSlavePort always has a node, but only exists when the HellaCache has a scratchpad
-  val fg = LazyModule(new TLFragmenter(tileParams.core.fetchBytes, cacheBlockBytes, earlyAck=true))
-  val ww = LazyModule(new TLWidthWidget(xBytes))
   val scratch = tileParams.dcache.flatMap { d => d.scratch.map(s =>
     LazyModule(new ScratchpadSlavePort(AddressSet(s, d.dataScratchpadBytes-1), xBytes, tileParams.core.useAtomics)))
   }
 
+  val intOutputNode = tileParams.core.tileControlAddr.map(dummy => IntOutputNode())
+  val busErrorUnit = tileParams.core.tileControlAddr map { a =>
+    val beu = LazyModule(new BusErrorUnit(new L1BusErrors, BusErrorUnitParams(a)))
+    intOutputNode.get := beu.intNode
+    beu
+  }
+
+  // connect any combination of ITIM, DTIM, and BusErrorUnit
+  val slaveNode = TLInputNode()
   DisableMonitors { implicit p =>
-    frontend.slaveNode :*= fg.node
+    val xbarPorts =
-    fg.node :*= ww.node
+      scratch.map(lm => (lm.node, xBytes)) ++
-    ww.node :*= slaveNode
+      busErrorUnit.map(lm => (lm.node, xBytes)) ++
-    scratch foreach { lm => lm.node := TLFragmenter(xBytes, cacheBlockBytes, earlyAck=true)(slaveNode) }
+      tileParams.icache.flatMap(icache => icache.itimAddr.map(a => (frontend.slaveNode, tileParams.core.fetchBytes)))
+
+    if (xbarPorts.nonEmpty) {
+      val xbar = LazyModule(new TLXbar)
+      xbar.node := TLFIFOFixer()(TLFragmenter(xBytes, cacheBlockBytes, earlyAck=true)(slaveNode))
+      xbarPorts.foreach { case (port, bytes) =>
+        port := (if (bytes == xBytes) xbar.node else TLFragmenter(bytes, xBytes, earlyAck=true)(TLWidthWidget(xBytes)(xbar.node)))
+      }
+    }
   }
 
   def findScratchpadFromICache: Option[AddressSet] = scratch.map { s =>
@@ -130,6 +141,7 @@ trait CanHaveScratchpad extends HasHellaCache with HasICacheFrontend {
 trait CanHaveScratchpadBundle extends HasHellaCacheBundle with HasICacheFrontendBundle {
   val outer: CanHaveScratchpad
   val slave = outer.slaveNode.bundleIn
+  val intOutput = outer.intOutputNode.map(_.bundleOut)
 }
 
 trait CanHaveScratchpadModule extends HasHellaCacheModule with HasICacheFrontendModule {
@@ -137,4 +149,8 @@ trait CanHaveScratchpadModule extends HasHellaCacheModule with HasICacheFrontend
   val io: CanHaveScratchpadBundle
 
   outer.scratch.foreach { lm => dcachePorts += lm.module.io.dmem }
+  outer.busErrorUnit.foreach { lm =>
+    lm.module.io.errors.dcache := outer.dcache.module.io.errors
+    lm.module.io.errors.icache := outer.frontend.module.io.errors
+  }
 }

src/main/scala/tile/BaseTile.scala

@@ -21,6 +21,7 @@ trait TileParams {
   val dcache: Option[DCacheParams]
   val rocc: Seq[RoCCParams]
   val btb: Option[BTBParams]
+  val trace: Boolean
 }
 
 trait HasTileParameters {
@@ -37,6 +38,7 @@ trait HasTileParameters {
 
   def xLen: Int = p(XLen)
   def xBytes: Int = xLen / 8
+  def iLen: Int = 32
   def pgIdxBits: Int = 12
   def pgLevelBits: Int = 10 - log2Ceil(xLen / 32)
   def vaddrBits: Int = pgIdxBits + pgLevels * pgLevelBits
@@ -92,6 +94,10 @@ trait HasExternallyDrivenTileConstants extends Bundle with HasTileParameters {
   val reset_vector = UInt(INPUT, resetVectorLen)
 }
 
+trait CanHaveInstructionTracePort extends Bundle with HasTileParameters {
+  val trace = tileParams.trace.option(Vec(tileParams.core.retireWidth, new TracedInstruction).asOutput)
+}
+
 /** Base class for all Tiles that use TileLink */
 abstract class BaseTile(tileParams: TileParams)(implicit p: Parameters) extends BareTile
     with HasTileParameters
@@ -102,6 +108,7 @@ abstract class BaseTile(tileParams: TileParams)(implicit p: Parameters) extends
 class BaseTileBundle[+L <: BaseTile](_outer: L) extends BareTileBundle(_outer)
     with HasTileLinkMasterPortBundle
     with HasExternallyDrivenTileConstants
+    with CanHaveInstructionTracePort
 
 class BaseTileModule[+L <: BaseTile, +B <: BaseTileBundle[L]](_outer: L, _io: () => B) extends BareTileModule(_outer, _io)
     with HasTileParameters

src/main/scala/tile/Core.scala

@@ -27,6 +27,7 @@ trait CoreParams {
   val nLocalInterrupts: Int
   val nL2TLBEntries: Int
   val jumpInFrontend: Boolean
+  val tileControlAddr: Option[BigInt]
 
   def instBytes: Int = instBits / 8
   def fetchBytes: Int = fetchWidth * instBytes
@@ -75,5 +76,6 @@ trait HasCoreIO extends HasTileParameters {
     val ptw = new DatapathPTWIO().flip
     val fpu = new FPUCoreIO().flip
     val rocc = new RoCCCoreIO().flip
+    val trace = Vec(coreParams.retireWidth, new TracedInstruction).asOutput
   }
 }

src/main/scala/tile/RocketTile.scala

@@ -19,6 +19,7 @@ case class RocketTileParams(
     btb: Option[BTBParams] = Some(BTBParams()),
     dataScratchpadBytes: Int = 0,
     boundaryBuffers: Boolean = false,
+    trace: Boolean = false,
     name: Option[String] = Some("tile"),
     externalMasterBuffers: Int = 0,
     externalSlaveBuffers: Int = 0) extends TileParams {
@@ -139,6 +140,7 @@ class RocketTileModule(outer: RocketTile) extends BaseTileModule(outer, () => ne
   val core = Module(p(BuildCore)(outer.p))
   decodeCoreInterrupts(core.io.interrupts) // Decode the interrupt vector
   core.io.hartid := io.hartid // Pass through the hartid
+  io.trace.foreach { _ := core.io.trace }
   outer.frontend.module.io.cpu <> core.io.imem
   outer.frontend.module.io.reset_vector := io.reset_vector
   outer.frontend.module.io.hartid := io.hartid
@@ -168,6 +170,7 @@ abstract class RocketTileWrapper(rtp: RocketTileParams, hartid: Int)(implicit p:
   val rocket = LazyModule(new RocketTile(rtp, hartid))
   val masterNode: OutputNode[_,_,_,_,_]
   val slaveNode: InputNode[_,_,_,_,_]
+  val intOutputNode = rocket.intOutputNode.map(dummy => IntOutputNode())
   val asyncIntNode   = IntInputNode()
   val periphIntNode  = IntInputNode()
   val coreIntNode    = IntInputNode()
@@ -195,10 +198,19 @@ abstract class RocketTileWrapper(rtp: RocketTileParams, hartid: Int)(implicit p:
     }
   }
 
+  def outputInterruptXingLatency: Int
+
+  rocket.intOutputNode.foreach { rocketIntOutputNode =>
+    val outXing = LazyModule(new IntXing(outputInterruptXingLatency))
+    intOutputNode.get := outXing.intnode
+    outXing.intnode := rocketIntOutputNode
+  }
+
   lazy val module = new LazyModuleImp(this) {
-    val io = new CoreBundle with HasExternallyDrivenTileConstants {
+    val io = new CoreBundle with HasExternallyDrivenTileConstants with CanHaveInstructionTracePort {
       val master = masterNode.bundleOut
       val slave = slaveNode.bundleIn
+      val outputInterrupts = intOutputNode.map(_.bundleOut)
       val asyncInterrupts  = asyncIntNode.bundleIn
       val periphInterrupts = periphIntNode.bundleIn
       val coreInterrupts   = coreIntNode.bundleIn
@@ -206,6 +218,7 @@ abstract class RocketTileWrapper(rtp: RocketTileParams, hartid: Int)(implicit p:
     // signals that do not change based on crossing type:
     rocket.module.io.hartid := io.hartid
     rocket.module.io.reset_vector := io.reset_vector
+    io.trace.foreach { _ := rocket.module.io.trace.get }
   }
 }
 
@@ -224,6 +237,8 @@ class SyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters)
   intXbar.intnode  := xing.intnode
   intXbar.intnode  := periphIntNode
   intXbar.intnode  := coreIntNode
+
+  def outputInterruptXingLatency = 0
 }
 
 class AsyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends RocketTileWrapper(rtp, hartid) {
@@ -251,6 +266,8 @@ class AsyncRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters
   intXbar.intnode  := asyncXing.intnode
   intXbar.intnode  := periphXing.intnode
   intXbar.intnode  := coreIntNode
+
+  def outputInterruptXingLatency = 3
 }
 
 class RationalRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Parameters) extends RocketTileWrapper(rtp, hartid) {
@@ -279,4 +296,6 @@ class RationalRocketTile(rtp: RocketTileParams, hartid: Int)(implicit p: Paramet
   intXbar.intnode  := asyncXing.intnode
   intXbar.intnode  := periphXing.intnode
   intXbar.intnode  := coreIntNode
+
+  def outputInterruptXingLatency = 1
 }

src/main/scala/util/ECC.scala

@@ -15,6 +15,9 @@ abstract class Decoding
 
 abstract class Code
 {
+  def canDetect: Boolean
+  def canCorrect: Boolean
+
   def width(w0: Int): Int
   def encode(x: UInt): UInt
   def decode(x: UInt): Decoding
@@ -29,6 +32,9 @@ abstract class Code
 
 class IdentityCode extends Code
 {
+  def canDetect = false
+  def canCorrect = false
+
   def width(w0: Int) = w0
   def encode(x: UInt) = x
   def swizzle(x: UInt) = x
@@ -42,6 +48,9 @@ class IdentityCode extends Code
 
 class ParityCode extends Code
 {
+  def canDetect = true
+  def canCorrect = false
+
   def width(w0: Int) = w0+1
   def encode(x: UInt) = Cat(x.xorR, x)
   def swizzle(x: UInt) = Cat(false.B, x)
@@ -55,6 +64,9 @@ class ParityCode extends Code
 
 class SECCode extends Code
 {
+  def canDetect = true
+  def canCorrect = true
+
   def width(k: Int) = {
     val m = log2Floor(k) + 1
     k + m + (if((1 << m) < m+k+1) 1 else 0)
@@ -101,6 +113,9 @@ class SECCode extends Code
 
 class SECDEDCode extends Code
 {
+  def canDetect = true
+  def canCorrect = true
+
   private val sec = new SECCode
   private val par = new ParityCode
 

src/main/scala/util/ShiftReg.scala

@@ -125,8 +125,10 @@ class SynchronizerShiftReg(w: Int = 1, sync: Int = 3) extends AbstractPipelineRe
 
 
 object SynchronizerShiftReg {
-  def apply [T <: Chisel.Data](in: T, sync: Int = 3, name: Option[String] = None): T =
+  def apply [T <: Chisel.Data](in: T, sync: Int = 3, name: Option[String] = None): T = {
-    AbstractPipelineReg(new SynchronizerShiftReg(in.getWidth, sync), in, name)
+    if (sync == 0) in
+    else AbstractPipelineReg(new SynchronizerShiftReg(in.getWidth, sync), in, name)
+  }
 }
 
 class SyncResetSynchronizerShiftReg(w: Int = 1, sync: Int = 3, init: Int = 0) extends AbstractPipelineReg(w) {