Merge pull request #849 from freechipsproject/l2-tlb

L1 memory system improvements

src/main/scala/rocket/DCache.scala

@@ -146,6 +146,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   tlb.io.req.bits.sfence.bits.rs1 := s1_req.typ(0)
   tlb.io.req.bits.sfence.bits.rs2 := s1_req.typ(1)
   tlb.io.req.bits.sfence.bits.asid := io.cpu.s1_data.data
+  tlb.io.req.bits.sfence.bits.addr := s1_req.addr
   tlb.io.req.bits.passthrough := s1_req.phys
   tlb.io.req.bits.vaddr := s1_req.addr
   tlb.io.req.bits.instruction := false
@@ -473,7 +474,7 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
     probe_bits.address := Cat(s2_victim_tag, s2_req.addr(idxMSB, idxLSB)) << idxLSB
   }
   when (s2_probe) {
-    s1_nack := true
+    val probeNack = Wire(init = true.B)
     when (s2_meta_error) {
       release_state := s_probe_retry
     }.elsewhen (s2_prb_ack_data) {
@@ -484,9 +485,10 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
       release_state := Mux(releaseDone, s_probe_write_meta, s_probe_rep_clean)
     }.otherwise {
       tl_out.c.valid := true
-      s1_nack := !releaseDone
+      probeNack := !releaseDone
       release_state := Mux(releaseDone, s_ready, s_probe_rep_miss)
     }
+    when (probeNack) { s1_nack := true }
   }
   when (release_state === s_probe_retry) {
     metaArb.io.in(6).valid := true

src/main/scala/rocket/Frontend.scala

@@ -168,12 +168,13 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   fq.io.enq.bits.pc := s2_pc
   io.cpu.npc := ~(~Mux(io.cpu.req.valid, io.cpu.req.bits.pc, npc) | (coreInstBytes-1)) // discard LSB(s)
 
-  fq.io.enq.bits.data := icache.io.resp.bits
+  fq.io.enq.bits.data := icache.io.resp.bits.data
   fq.io.enq.bits.mask := UInt((1 << fetchWidth)-1) << s2_pc.extract(log2Ceil(fetchWidth)+log2Ceil(coreInstBytes)-1, log2Ceil(coreInstBytes))
-  fq.io.enq.bits.xcpt := s2_tlb_resp
   fq.io.enq.bits.replay := icache.io.s2_kill && !icache.io.resp.valid && !s2_xcpt
   fq.io.enq.bits.btb.valid := s2_btb_resp_valid
   fq.io.enq.bits.btb.bits := s2_btb_resp_bits
+  fq.io.enq.bits.xcpt := s2_tlb_resp
+  when (icache.io.resp.valid && icache.io.resp.bits.ae) { fq.io.enq.bits.xcpt.ae.inst := true }
 
   io.cpu.resp <> fq.io.deq
 

src/main/scala/rocket/ICache.scala

@@ -58,6 +58,13 @@ class ICache(val icacheParams: ICacheParams, val hartid: Int)(implicit p: Parame
   }
 }
 
+class ICacheResp(outer: ICache) extends Bundle {
+  val data = UInt(width = outer.icacheParams.fetchBytes*8)
+  val ae = Bool()
+
+  override def cloneType = new ICacheResp(outer).asInstanceOf[this.type]
+}
+
 class ICacheBundle(outer: ICache) extends CoreBundle()(outer.p) {
   val hartid = UInt(INPUT, hartIdLen)
   val req = Decoupled(new ICacheReq).flip
@@ -66,7 +73,7 @@ class ICacheBundle(outer: ICache) extends CoreBundle()(outer.p) {
   val s1_kill = Bool(INPUT) // delayed one cycle w.r.t. req
   val s2_kill = Bool(INPUT) // delayed two cycles; prevents I$ miss emission
 
-  val resp = Valid(UInt(width = outer.icacheParams.fetchBytes*8))
+  val resp = Valid(new ICacheResp(outer))
   val invalidate = Bool(INPUT)
   val tl_out = outer.masterNode.bundleOut
   val tl_in = outer.slaveNode.map(_.bundleIn)
@@ -145,15 +152,18 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
     v
   }
 
-  val tag_array = SeqMem(nSets, Vec(nWays, Bits(width = tECC.width(tagBits))))
+  val tag_array = SeqMem(nSets, Vec(nWays, UInt(width = tECC.width(1 + tagBits))))
   val tag_rdata = tag_array.read(s0_vaddr(untagBits-1,blockOffBits), !refill_done && s0_valid)
+  val accruedRefillError = Reg(Bool())
+  val refillError = tl_out.d.bits.error || (refill_cnt > 0 && accruedRefillError)
   when (refill_done) {
-    val tag = tECC.encode(refill_tag)
+    val enc_tag = tECC.encode(Cat(refillError, refill_tag))
-    tag_array.write(refill_idx, Vec.fill(nWays)(tag), Vec.tabulate(nWays)(repl_way === _))
+    tag_array.write(refill_idx, Vec.fill(nWays)(enc_tag), Seq.tabulate(nWays)(repl_way === _))
   }
 
   val vb_array = Reg(init=Bits(0, nSets*nWays))
   when (tl_out.d.fire()) {
+    accruedRefillError := refillError
     // clear bit when refill starts so hit-under-miss doesn't fetch bad data
     vb_array := vb_array.bitSet(Cat(repl_way, refill_idx), refill_done && !invalidated)
   }
@@ -164,6 +174,7 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
   }
 
   val s1_tag_disparity = Wire(Vec(nWays, Bool()))
+  val s1_tl_error = Wire(Vec(nWays, Bool()))
   val wordBits = outer.icacheParams.fetchBytes*8
   val s1_dout = Wire(Vec(nWays, UInt(width = dECC.width(wordBits))))
 
@@ -179,8 +190,12 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
         lineInScratchpad(scratchpadLine(s1s3_slaveAddr)) && scratchpadWay(s1s3_slaveAddr) === i,
         addrInScratchpad(io.s1_paddr) && scratchpadWay(io.s1_paddr) === i)
     val s1_vb = vb_array(Cat(UInt(i), s1_idx)) && !s1_slaveValid
-    s1_tag_disparity(i) := s1_vb && tECC.decode(tag_rdata(i)).error
+    val enc_tag = tECC.decode(tag_rdata(i))
-    s1_tag_hit(i) := scratchpadHit || (s1_vb && tECC.decode(tag_rdata(i)).uncorrected === s1_tag)
+    val (tl_error, tag) = Split(enc_tag.uncorrected, tagBits)
+    val tagMatch = s1_vb && tag === s1_tag
+    s1_tag_disparity(i) := s1_vb && enc_tag.error
+    s1_tl_error(i) := tagMatch && tl_error.toBool
+    s1_tag_hit(i) := tagMatch || scratchpadHit
   }
   assert(!(s1_valid || s1_slaveValid) || PopCount(s1_tag_hit zip s1_tag_disparity map { case (h, d) => h && !d }) <= 1)
 
@@ -212,7 +227,8 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
       require(tECC.isInstanceOf[uncore.util.IdentityCode])
       require(dECC.isInstanceOf[uncore.util.IdentityCode])
       require(outer.icacheParams.itimAddr.isEmpty)
-      io.resp.bits := Mux1H(s1_tag_hit, s1_dout)
+      io.resp.bits.data := Mux1H(s1_tag_hit, s1_dout)
+      io.resp.bits.ae := s1_tl_error.asUInt.orR
       io.resp.valid := s1_valid && s1_hit
 
     case 2 =>
@@ -221,11 +237,13 @@ class ICacheModule(outer: ICache) extends LazyModuleImp(outer)
       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_tl_error = RegEnable(s1_tl_error.asUInt.orR, s1_valid || s1_slaveValid)
       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 }
 
-      io.resp.bits := s2_data_decoded.uncorrected
+      io.resp.bits.data := s2_data_decoded.uncorrected
+      io.resp.bits.ae := s2_tl_error
       io.resp.valid := s2_valid && s2_hit && !s2_disparity
 
       tl_in.map { tl =>

src/main/scala/rocket/NBDcache.scala

@@ -708,6 +708,7 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   dtlb.io.req.bits.sfence.valid := s1_sfence
   dtlb.io.req.bits.sfence.bits.rs1 := s1_req.typ(0)
   dtlb.io.req.bits.sfence.bits.rs2 := s1_req.typ(1)
+  dtlb.io.req.bits.sfence.bits.addr := s1_req.addr
   dtlb.io.req.bits.sfence.bits.asid := io.cpu.s1_data.data
   dtlb.io.req.bits.passthrough := s1_req.phys
   dtlb.io.req.bits.vaddr := s1_req.addr

src/main/scala/rocket/PTW.scala

@@ -11,6 +11,7 @@ import coreplex.CacheBlockBytes
 import uncore.constants._
 import uncore.tilelink2._
 import util._
+import uncore.util.ParityCode
 
 import scala.collection.mutable.ListBuffer
 
@@ -37,7 +38,7 @@ class TLBPTWIO(implicit p: Parameters) extends CoreBundle()(p)
 class DatapathPTWIO(implicit p: Parameters) extends CoreBundle()(p)
     with HasRocketCoreParameters {
   val ptbr = new PTBR().asInput
-  val invalidate = Bool(INPUT)
+  val sfence = Valid(new SFenceReq).flip
   val status = new MStatus().asInput
   val pmp = Vec(nPMPs, new PMP).asInput
 }
@@ -125,17 +126,69 @@ class PTW(n: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(
       data(r) := pte.ppn
     }
     when (hit && state === s_req) { plru.access(OHToUInt(hits)) }
-    when (io.dpath.invalidate) { valid := 0 }
+    when (io.dpath.sfence.valid && !io.dpath.sfence.bits.rs1) { valid := 0 }
 
     (hit && count < pgLevels-1, Mux1H(hits, data))
   }
 
-  io.mem.req.valid     := state === s_req
+  val l2_refill = RegNext(false.B)
+  val (l2_hit, l2_pte) = if (coreParams.nL2TLBEntries == 0) (false.B, Wire(new PTE)) else {
+    class Entry extends Bundle {
+      val ppn = UInt(width = ppnBits)
+      val d = Bool()
+      val a = Bool()
+      val u = Bool()
+      val x = Bool()
+      val w = Bool()
+      val r = Bool()
+    }
+
+    val code = new ParityCode
+    require(isPow2(coreParams.nL2TLBEntries))
+    val idxBits = log2Ceil(coreParams.nL2TLBEntries)
+    val tagBits = vpnBits - idxBits
+    val ram = SeqMem(coreParams.nL2TLBEntries, UInt(width = code.width(new Entry().getWidth + tagBits)))
+    val g = Reg(UInt(width = coreParams.nL2TLBEntries))
+    val valid = RegInit(UInt(0, coreParams.nL2TLBEntries))
+    val (r_tag, r_idx) = Split(r_req.addr, idxBits)
+    when (l2_refill) {
+      val entry = Wire(new Entry)
+      entry := r_pte
+      ram.write(r_idx, code.encode(Cat(entry.asUInt, r_tag)))
+
+      val mask = UIntToOH(r_idx)
+      valid := valid | mask
+      g := Mux(r_pte.g, g | mask, g & ~mask)
+    }
+    when (io.dpath.sfence.valid) {
+      valid :=
+        Mux(io.dpath.sfence.bits.rs1, valid & ~UIntToOH(io.dpath.sfence.bits.addr(idxBits+pgIdxBits-1, pgIdxBits)),
+        Mux(io.dpath.sfence.bits.rs2, valid & g, 0.U))
+    }
+
+    val s0_valid = !l2_refill && arb.io.out.fire()
+    val s1_valid = RegNext(s0_valid)
+    val s2_valid = RegNext(s1_valid && valid(r_idx))
+    val s1_rdata = ram.read(arb.io.out.bits.addr(idxBits-1, 0), s0_valid)
+    val s2_rdata = code.decode(RegEnable(s1_rdata, s1_valid))
+    when (s2_valid && s2_rdata.error) { valid := 0.U }
+
+    val (s2_entry, s2_tag) = Split(s2_rdata.uncorrected, tagBits)
+    val s2_hit = s2_valid && !s2_rdata.error && r_tag === s2_tag
+    val s2_pte = Wire(new PTE)
+    s2_pte := s2_entry.asTypeOf(new Entry)
+    s2_pte.g := g(r_idx)
+    s2_pte.v := true
+
+    (s2_hit, s2_pte)
+  }
+  
+  io.mem.req.valid := state === s_req && !l2_hit
   io.mem.req.bits.phys := Bool(true)
   io.mem.req.bits.cmd  := M_XRD
   io.mem.req.bits.typ  := log2Ceil(xLen/8)
   io.mem.req.bits.addr := pte_addr
-  io.mem.s1_kill := s1_kill
+  io.mem.s1_kill := s1_kill || l2_hit
   io.mem.invalidate_lr := Bool(false)
   
   val pmaPgLevelHomogeneous = (0 until pgLevels) map { i =>
@@ -159,7 +212,7 @@ class PTW(n: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(
   // control state machine
   switch (state) {
     is (s_ready) {
-      when (arb.io.out.valid) {
+      when (arb.io.out.fire()) {
         state := s_req
       }
       count := UInt(0)
@@ -186,6 +239,7 @@ class PTW(n: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(
           state := s_req
           count := count + 1
         }.otherwise {
+          l2_refill := pte.v && !invalid_paddr && count === pgLevels-1
           resp_ae := pte.v && invalid_paddr
           state := s_ready
           resp_valid(r_req_dest) := true
@@ -198,6 +252,12 @@ class PTW(n: Int)(implicit edge: TLEdgeOut, p: Parameters) extends CoreModule()(
       }
     }
   }
+  when (l2_hit) {
+    state := s_ready
+    resp_valid(r_req_dest) := true
+    resp_ae := false
+    r_pte := l2_pte
+  }
 }
 
 /** Mix-ins for constructing tiles that might have a PTW */

src/main/scala/rocket/Rocket.scala

@@ -25,6 +25,7 @@ case class RocketCoreParams(
   nPMPs: Int = 8,
   nPerfCounters: Int = 0,
   nCustomMRWCSRs: Int = 0,
+  nL2TLBEntries: Int = 0,
   mtvecInit: Option[BigInt] = Some(BigInt(0)),
   mtvecWritable: Boolean = true,
   fastLoadWord: Boolean = true,
@@ -588,8 +589,9 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   io.imem.sfence.valid := wb_reg_valid && wb_reg_sfence
   io.imem.sfence.bits.rs1 := wb_ctrl.mem_type(0)
   io.imem.sfence.bits.rs2 := wb_ctrl.mem_type(1)
+  io.imem.sfence.bits.addr := wb_reg_wdata
   io.imem.sfence.bits.asid := wb_reg_rs2
-  io.ptw.invalidate := io.imem.sfence.valid && !io.imem.sfence.bits.rs1
+  io.ptw.sfence := io.imem.sfence
 
   ibuf.io.inst(0).ready := !ctrl_stalld || csr.io.interrupt
 

src/main/scala/rocket/TLB.scala

@@ -20,6 +20,7 @@ case object ASIdBits extends Field[Int]
 class SFenceReq(implicit p: Parameters) extends CoreBundle()(p) {
   val rs1 = Bool()
   val rs2 = Bool()
+  val addr = UInt(width = vaddrBits)
   val asid = UInt(width = asIdBits max 1) // TODO zero-width
 }
 
@@ -252,6 +253,7 @@ class TLB(lgMaxSize: Int, nEntries: Int)(implicit edge: TLEdgeOut, p: Parameters
     }
 
     when (sfence) {
+      assert((io.req.bits.sfence.bits.addr >> pgIdxBits) === vpn(vpnBits-1,0))
       valid := Mux(io.req.bits.sfence.bits.rs1, valid & ~hits(totalEntries-1, 0),
                Mux(io.req.bits.sfence.bits.rs2, valid & entries.map(_.g).asUInt, 0))
     }

src/main/scala/tile/Core.scala

@@ -24,6 +24,7 @@ trait CoreParams {
   val retireWidth: Int
   val instBits: Int
   val nLocalInterrupts: Int
+  val nL2TLBEntries: Int
 }
 
 trait HasCoreParameters extends HasTileParameters {