Add performance counter facility

src/main/scala/rocket/Arbiter.scala

@@ -56,6 +56,8 @@ class HellaCacheArbiter(n: Int)(implicit p: Parameters) extends Module
       resp.valid := io.mem.resp.valid && tag_hit
       io.requestor(i).xcpt := io.mem.xcpt
       io.requestor(i).ordered := io.mem.ordered
+      io.requestor(i).acquire := io.mem.acquire
+      io.requestor(i).release := io.mem.release
       io.requestor(i).s2_nack := io.mem.s2_nack && s2_id === UInt(i)
       resp.bits := io.mem.resp.bits
       resp.bits.tag := io.mem.resp.bits.tag >> log2Up(n)

src/main/scala/rocket/CSR.scala

@@ -124,14 +124,23 @@ object CSR
   }
 
   val firstCtr = CSRs.cycle
+  val firstCtrH = CSRs.cycleh
   val firstHPC = CSRs.hpmcounter3
+  val firstHPCH = CSRs.hpmcounter3h
   val firstHPE = CSRs.mhpmevent3
   val firstMHPC = CSRs.mhpmcounter3
+  val firstMHPCH = CSRs.mhpmcounter3h
   val firstHPM = 3
   val nCtr = 32
   val nHPM = nCtr - firstHPM
 }
 
+class PerfCounterIO(implicit p: Parameters) extends CoreBundle
+    with HasRocketCoreParameters {
+  val eventSel = UInt(OUTPUT, xLen)
+  val inc = UInt(INPUT, log2Ceil(1+retireWidth))
+}
+
 class CSRFileIO(implicit p: Parameters) extends CoreBundle
     with HasRocketCoreParameters {
   val interrupts = new TileInterrupts().asInput
@@ -174,10 +183,10 @@ class CSRFileIO(implicit p: Parameters) extends CoreBundle
   val interrupt = Bool(OUTPUT)
   val interrupt_cause = UInt(OUTPUT, xLen)
   val bp = Vec(nBreakpoints, new BP).asOutput
-  val events = Vec(nPerfEvents, Bool()).asInput
+  val counters = Vec(nPerfCounters, new PerfCounterIO)
 }
 
-class CSRFile(implicit p: Parameters) extends CoreModule()(p)
+class CSRFile(perfEventSets: EventSets = new EventSets(Seq()))(implicit p: Parameters) extends CoreModule()(p)
     with HasRocketCoreParameters {
   val io = new CSRFileIO
 
@@ -258,8 +267,9 @@ class CSRFile(implicit p: Parameters) extends CoreModule()(p)
 
   val reg_instret = WideCounter(64, io.retire)
   val reg_cycle = if (enableCommitLog) reg_instret else WideCounter(64)
-  val reg_hpmevent = Seq.fill(nPerfCounters)(if (nPerfEvents > 1) Reg(UInt(width = log2Ceil(nPerfEvents))) else UInt(0))
+  val reg_hpmevent = io.counters.map(c => Reg(init = UInt(0, xLen)))
-  val reg_hpmcounter = reg_hpmevent.map(e => WideCounter(64, ((UInt(0) +: io.events): Seq[UInt])(e)))
+  (io.counters zip reg_hpmevent) foreach { case (c, e) => c.eventSel := e }
+  val reg_hpmcounter = io.counters.map(c => WideCounter(40, c.inc, reset = false))
   val hpm_mask = reg_mcounteren & Mux((!usingVM).B || reg_mstatus.prv === PRV.S, delegable_counters.U, reg_scounteren)
 
   val mip = Wire(init=reg_mip)
@@ -339,6 +349,10 @@ class CSRFile(implicit p: Parameters) extends CoreModule()(p)
     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 (usingVM) {
@@ -407,7 +421,7 @@ class CSRFile(implicit p: Parameters) extends CoreModule()(p)
   io.decode.read_illegal := effective_prv < io.decode.csr(9,8) ||
     !read_mapping.keys.map(io.decode.csr === _).reduce(_||_) ||
     io.decode.csr === CSRs.sptbr && !allow_sfence_vma ||
-    io.decode.csr >= CSR.firstCtr && io.decode.csr < CSR.firstCtr + CSR.nCtr && effective_prv <= PRV.S && hpm_mask(io.decode.csr(log2Ceil(CSR.firstCtr)-1,0)) ||
+    (io.decode.csr.inRange(CSR.firstCtr, CSR.firstCtr + CSR.nCtr) || io.decode.csr.inRange(CSR.firstCtrH, CSR.firstCtrH + CSR.nCtr)) && effective_prv <= PRV.S && hpm_mask(io.decode.csr(log2Ceil(CSR.firstCtr)-1,0)) ||
     Bool(usingDebug) && !reg_debug && debug_csrs.keys.map(io.decode.csr === _).reduce(_||_) ||
     Bool(usingFPU) && fp_csrs.keys.map(io.decode.csr === _).reduce(_||_) && io.decode.fp_illegal
   io.decode.write_illegal := io.decode.csr(11,10).andR
@@ -561,8 +575,7 @@ class CSRFile(implicit p: Parameters) extends CoreModule()(p)
 
     for (((e, c), i) <- (reg_hpmevent zip reg_hpmcounter) zipWithIndex) {
       writeCounter(i + CSR.firstMHPC, c, wdata)
-      if (nPerfEvents > 1)
+      when (decoded_addr(i + CSR.firstHPE)) { e := perfEventSets.maskEventSelector(wdata) }
-        when (decoded_addr(i + CSR.firstHPE)) { e := wdata }
     }
     writeCounter(CSRs.mcycle, reg_cycle, wdata)
     writeCounter(CSRs.minstret, reg_instret, wdata)
@@ -688,10 +701,10 @@ class CSRFile(implicit p: Parameters) extends CoreModule()(p)
   def writeCounter(lo: Int, ctr: WideCounter, wdata: UInt) = {
     if (xLen == 32) {
       val hi = lo + CSRs.mcycleh - CSRs.mcycle
-      when (decoded_addr(lo)) { ctr := Cat(ctr(63, 32), wdata) }
+      when (decoded_addr(lo)) { ctr := Cat(ctr(ctr.getWidth-1, 32), wdata) }
-      when (decoded_addr(hi)) { ctr := Cat(wdata, ctr(31, 0)) }
+      when (decoded_addr(hi)) { ctr := Cat(wdata(ctr.getWidth-33, 0), ctr(31, 0)) }
     } else {
-      when (decoded_addr(lo)) { ctr := wdata }
+      when (decoded_addr(lo)) { ctr := wdata(ctr.getWidth-1, 0) }
     }
   }
   def formEPC(x: UInt) = ~(~x | Cat(!reg_misa('c'-'a'), UInt(1)))

src/main/scala/rocket/DCache.scala

@@ -503,4 +503,8 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
       flushing := false
     }
   }
+
+  // performance events
+  io.cpu.acquire := tl_out.a.fire()
+  io.cpu.release := tl_out.c.fire()
 }

src/main/scala/rocket/Events.scala

@@ -0,0 +1,35 @@
+// See LICENSE.Berkeley for license details.
+// See LICENSE.SiFive for license details.
+
+package rocket
+
+import util._
+import Chisel._
+
+class EventSet(gate: (UInt, UInt) => Bool, events: Seq[(String, () => Bool)]) {
+  def size = events.size
+  def hits = events.map(_._2()).asUInt
+  def check(mask: UInt) = gate(mask, hits)
+}
+
+class EventSets(eventSets: Seq[EventSet]) {
+  def maskEventSelector(eventSel: UInt): UInt = {
+    // allow full associativity between counters and event sets (for now?)
+    val setMask = (BigInt(1) << log2Ceil(eventSets.size)) - 1
+    val maskMask = ((BigInt(1) << eventSets.map(_.size).max) - 1) << eventSetIdBits
+    eventSel & (setMask | maskMask).U
+  }
+
+  private def decode(counter: UInt): (UInt, UInt) = {
+    require(eventSets.size <= (1 << eventSetIdBits))
+    (counter(log2Ceil(eventSets.size)-1, 0), counter >> eventSetIdBits)
+  }
+
+  def evaluate(eventSel: UInt): Bool = {
+    val (set, mask) = decode(eventSel)
+    val sets = eventSets map (_ check mask)
+    sets(set)
+  }
+
+  private def eventSetIdBits = 8
+}

src/main/scala/rocket/Frontend.scala

@@ -35,6 +35,9 @@ class FrontendIO(implicit p: Parameters) extends CoreBundle()(p) {
   val flush_icache = Bool(OUTPUT)
   val flush_tlb = Bool(OUTPUT)
   val npc = UInt(INPUT, width = vaddrBitsExtended)
+
+  // performance events
+  val acquire = Bool(INPUT)
 }
 
 class Frontend(implicit p: Parameters) extends LazyModule {
@@ -150,6 +153,9 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   io.cpu.resp.bits.replay := icache.io.s2_kill && !icache.io.resp.valid && !s2_xcpt_if
   io.cpu.resp.bits.btb.valid := s2_btb_resp_valid
   io.cpu.resp.bits.btb.bits := s2_btb_resp_bits
+
+  // performance events
+  io.cpu.acquire := icache.io.mem(0).a.fire()
 }
 
 /** Mix-ins for constructing tiles that have an ICache-based pipeline frontend */

src/main/scala/rocket/HellaCache.scala

@@ -121,6 +121,10 @@ class HellaCacheIO(implicit p: Parameters) extends CoreBundle()(p) {
   val s1_data = Bits(OUTPUT, coreDataBits) // data for previous cycle's req
   val s2_nack = Bool(INPUT) // req from two cycles ago is rejected
 
+  // performance events
+  val acquire = Bool(INPUT)
+  val release = Bool(INPUT)
+
   val resp = Valid(new HellaCacheResp).flip
   val replay_next = Bool(INPUT)
   val xcpt = (new HellaCacheExceptions).asInput

src/main/scala/rocket/NBDcache.scala

@@ -973,4 +973,8 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   io.cpu.resp.bits.data_word_bypass := loadgen.wordData
   io.cpu.ordered := mshrs.io.fence_rdy && !s1_valid && !s2_valid
   io.cpu.replay_next := (s1_replay && s1_read) || mshrs.io.replay_next
+
+  // performance events
+  io.cpu.acquire := tl_out.a.fire()
+  io.cpu.release := tl_out.c.fire()
 }

src/main/scala/rocket/Rocket.scala

@@ -19,7 +19,6 @@ case class RocketCoreParams(
   useCompressed: Boolean = true,
   nBreakpoints: Int = 1,
   nPerfCounters: Int = 0,
-  nPerfEvents: Int = 0,
   nCustomMRWCSRs: Int = 0,
   mtvecInit: Option[BigInt] = Some(BigInt(0)),
   mtvecWritable: Boolean = true,
@@ -44,7 +43,6 @@ trait HasRocketCoreParameters extends HasCoreParameters {
   val fastJAL = rocketParams.fastJAL
   val nBreakpoints = rocketParams.nBreakpoints
   val nPerfCounters = rocketParams.nPerfCounters
-  val nPerfEvents = rocketParams.nPerfEvents
   val nCustomMrwCsrs = rocketParams.nCustomMRWCSRs
   val mtvecInit = rocketParams.mtvecInit
   val mtvecWritable = rocketParams.mtvecWritable
@@ -58,6 +56,50 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
     with HasRocketCoreParameters
     with HasCoreIO {
 
+  // performance counters
+  def pipelineIDToWB[T <: Data](x: T): T =
+    RegEnable(RegEnable(RegEnable(x, !ctrl_killd), ex_pc_valid), mem_pc_valid)
+  val perfEvents = new EventSets(Seq(
+    new EventSet((mask, hits) => Mux(mask(0), wb_xcpt, wb_valid && pipelineIDToWB((mask & hits).orR)), Seq(
+      ("exception", () => false.B),
+      ("load", () => id_ctrl.mem && id_ctrl.mem_cmd === M_XRD && !id_ctrl.fp),
+      ("store", () => id_ctrl.mem && id_ctrl.mem_cmd === M_XWR && !id_ctrl.fp),
+      ("amo", () => Bool(usingAtomics) && id_ctrl.mem && (isAMO(id_ctrl.mem_cmd) || id_ctrl.mem_cmd.isOneOf(M_XLR, M_XSC))),
+      ("system", () => id_ctrl.csr =/= CSR.N),
+      ("arith", () => id_ctrl.wxd && !(id_ctrl.jal || id_ctrl.jalr || id_ctrl.mem || id_ctrl.fp || id_ctrl.div || id_ctrl.csr =/= CSR.N)),
+      ("branch", () => id_ctrl.branch),
+      ("jal", () => id_ctrl.jal),
+      ("jalr", () => id_ctrl.jalr))
+      ++ (if (!usingMulDiv) Seq() else Seq(
+        ("mul", () => id_ctrl.div && (id_ctrl.alu_fn & ALU.FN_DIV) =/= ALU.FN_DIV),
+        ("div", () => id_ctrl.div && (id_ctrl.alu_fn & ALU.FN_DIV) === ALU.FN_DIV)))
+      ++ (if (!usingFPU) Seq() else Seq(
+        ("fp load", () => id_ctrl.fp && io.fpu.dec.ldst && io.fpu.dec.wen),
+        ("fp store", () => id_ctrl.fp && io.fpu.dec.ldst && !io.fpu.dec.wen),
+        ("fp add", () => id_ctrl.fp && io.fpu.dec.fma && io.fpu.dec.swap23),
+        ("fp mul", () => id_ctrl.fp && io.fpu.dec.fma && !io.fpu.dec.swap23 && !io.fpu.dec.ren3),
+        ("fp mul-add", () => id_ctrl.fp && io.fpu.dec.fma && io.fpu.dec.ren3),
+        ("fp div/sqrt", () => id_ctrl.fp && (io.fpu.dec.div || io.fpu.dec.sqrt)),
+        ("fp other", () => id_ctrl.fp && !(io.fpu.dec.ldst || io.fpu.dec.fma || io.fpu.dec.div || io.fpu.dec.sqrt))))),
+    new EventSet((mask, hits) => (mask & hits).orR, Seq(
+      ("load-use interlock", () => id_ex_hazard && ex_ctrl.mem || id_mem_hazard && mem_ctrl.mem || id_wb_hazard && wb_ctrl.mem),
+      ("long-latency interlock", () => id_sboard_hazard),
+      ("csr interlock", () => id_ex_hazard && ex_ctrl.csr =/= CSR.N || id_mem_hazard && mem_ctrl.csr =/= CSR.N || id_wb_hazard && wb_ctrl.csr =/= CSR.N),
+      ("I$ blocked", () => !(ibuf.io.inst(0).valid || Reg(next = take_pc))),
+      ("D$ blocked", () => id_ctrl.mem && dcache_blocked),
+      ("branch misprediction", () => take_pc_mem && mem_direction_misprediction),
+      ("control-flow target misprediction", () => take_pc_mem && mem_misprediction && !mem_direction_misprediction),
+      ("flush", () => take_pc_mem && mem_reg_flush_pipe),
+      ("replay", () => replay_wb))
+      ++ (if (!usingMulDiv) Seq() else Seq(
+        ("mul/div interlock", () => id_ex_hazard && ex_ctrl.div || id_mem_hazard && mem_ctrl.div || id_wb_hazard && wb_ctrl.div)))
+      ++ (if (!usingFPU) Seq() else Seq(
+        ("fp interlock", () => id_ex_hazard && ex_ctrl.fp || id_mem_hazard && mem_ctrl.fp || id_wb_hazard && wb_ctrl.fp || id_ctrl.fp && id_stall_fpu)))),
+    new EventSet((mask, hits) => (mask & hits).orR, Seq(
+      ("I$ miss", () => io.imem.acquire),
+      ("D$ miss", () => io.dmem.acquire),
+      ("D$ release", () => io.dmem.release)))))
+
   val decode_table = {
     (if (usingMulDiv) new MDecode +: (xLen > 32).option(new M64Decode).toSeq else Nil) ++:
     (if (usingAtomics) new ADecode +: (xLen > 32).option(new A64Decode).toSeq else Nil) ++:
@@ -142,7 +184,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   val id_npc = (ibuf.io.pc.asSInt + ImmGen(IMM_UJ, id_inst(0))).asUInt
   take_pc_id := Bool(fastJAL) && !ctrl_killd && id_ctrl.jal
 
-  val csr = Module(new CSRFile)
+  val csr = Module(new CSRFile(perfEvents))
   val id_csr_en = id_ctrl.csr.isOneOf(CSR.S, CSR.C, CSR.W)
   val id_system_insn = id_ctrl.csr >= CSR.I
   val id_csr_ren = id_ctrl.csr.isOneOf(CSR.S, CSR.C) && id_raddr1 === UInt(0)
@@ -292,6 +334,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
     (ex_reg_xcpt_interrupt || ex_reg_xcpt, ex_reg_cause)))
 
   // memory stage
+  val mem_pc_valid = mem_reg_valid || mem_reg_replay || mem_reg_xcpt_interrupt
   val mem_br_taken = mem_reg_wdata(0)
   val mem_br_target = mem_reg_pc.asSInt +
     Mux(mem_ctrl.branch && mem_br_taken, ImmGen(IMM_SB, mem_reg_inst),
@@ -303,6 +346,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   val mem_int_wdata = Mux(!mem_reg_xcpt && (mem_ctrl.jalr ^ mem_npc_misaligned), mem_br_target, mem_reg_wdata.asSInt).asUInt
   val mem_cfi = mem_ctrl.branch || mem_ctrl.jalr || mem_ctrl.jal
   val mem_cfi_taken = (mem_ctrl.branch && mem_br_taken) || mem_ctrl.jalr || (Bool(!fastJAL) && mem_ctrl.jal)
+  val mem_direction_misprediction = mem_reg_btb_hit && mem_ctrl.branch && mem_br_taken =/= mem_reg_btb_resp.taken
   val mem_misprediction = if (usingBTB) mem_wrong_npc else mem_cfi_taken
   take_pc_mem := mem_reg_valid && (mem_misprediction || mem_reg_flush_pipe)
 
@@ -357,7 +401,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   wb_reg_replay := replay_mem && !take_pc_wb
   wb_reg_xcpt := mem_xcpt && !take_pc_wb
   when (mem_xcpt) { wb_reg_cause := mem_cause }
-  when (mem_reg_valid || mem_reg_replay || mem_reg_xcpt_interrupt) {
+  when (mem_pc_valid) {
     wb_ctrl := mem_ctrl
     wb_reg_wdata := Mux(!mem_reg_xcpt && mem_ctrl.fp && mem_ctrl.wxd, io.fpu.toint_data, mem_int_wdata)
     when (mem_ctrl.rocc) {
@@ -556,6 +600,9 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p)
   io.rocc.cmd.bits.rs1 := wb_reg_wdata
   io.rocc.cmd.bits.rs2 := wb_reg_rs2
 
+  // evaluate performance counters
+  csr.io.counters foreach { c => c.inc := RegNext(perfEvents.evaluate(c.eventSel)) }
+
   if (enableCommitLog) {
     val pc = Wire(SInt(width=xLen))
     pc := wb_reg_pc

src/main/scala/util/Counters.scala

@@ -50,7 +50,7 @@ case class WideCounter(width: Int, inc: UInt = UInt(1), reset: Boolean = true)
 
   private val large = if (isWide) {
     val r = if (reset) Reg(init=UInt(0, width - smallWidth)) else Reg(UInt(width = width - smallWidth))
-    when (nextSmall(smallWidth)) { r := r +& UInt(1) }
+    when (nextSmall(smallWidth)) { r := r + UInt(1) }
     r
   } else null
 

src/main/scala/util/Package.scala

@@ -43,6 +43,8 @@ package object util {
       if (hi == lo-1) UInt(0)
       else x(hi, lo)
     }
+
+    def inRange(base: UInt, bounds: UInt) = x >= base && x < bounds
   }
 
   implicit class BooleanToAugmentedBoolean(val x: Boolean) extends AnyVal {