Make BTB more complexity-effective

BTB entries reference a small number of unique pages, so we separate the
storage of pages from indices.  This makes much larger BTBs feasible.  It's
easy to exacerbate cycle time this way, so one-hot encoding is used as needed.

rocket/src/main/scala/btb.scala

@@ -0,0 +1,121 @@
+package rocket
+
+import Chisel._
+import Util._
+import Node._
+import uncore.constants.AddressConstants._
+
+case class BTBConfig(entries: Int) {
+  val matchBits = PGIDX_BITS
+  val pages0 = 1 + log2Up(entries) // is this sensible? what about matchBits?
+  val pages = (pages0+1)/2*2 // control logic assumes 2 divides pages
+}
+
+// fully-associative branch target buffer
+class BTB(conf: BTBConfig) extends Module {
+  val io = new Bundle {
+    val current_pc     = UInt(INPUT, VADDR_BITS)
+    val hit            = Bool(OUTPUT)
+    val target         = UInt(OUTPUT, VADDR_BITS)
+    val wen            = Bool(INPUT)
+    val taken          = Bool(INPUT)
+    val invalidate     = Bool(INPUT)
+    val correct_pc     = UInt(INPUT, VADDR_BITS)
+    val correct_target = UInt(INPUT, VADDR_BITS)
+  }
+
+  val idxValid = Vec.fill(conf.entries){Reg(init=Bool(false))}
+  val idxs = Vec.fill(conf.entries){Reg(UInt(width=conf.matchBits))}
+  val idxPages = Vec.fill(conf.entries){Reg(UInt(width=log2Up(conf.pages)))}
+  val idxPagesOH = idxPages.map(UIntToOH(_)(conf.pages-1,0))
+  val tgts = Vec.fill(conf.entries){Reg(UInt(width=conf.matchBits))}
+  val tgtPages = Vec.fill(conf.entries){Reg(UInt(width=log2Up(conf.pages)))}
+  val tgtPagesOH = tgtPages.map(UIntToOH(_)(conf.pages-1,0))
+  val pages = Vec.fill(conf.pages){Reg(UInt(width=VADDR_BITS-conf.matchBits))}
+  val pageValid = Vec.fill(conf.pages){Reg(init=Bool(false))}
+
+  private def page(addr: UInt) = addr >> conf.matchBits
+  private def pageMatch(addr: UInt) = {
+    val p = page(addr)
+    Vec(pages.map(_ === p)).toBits & pageValid.toBits
+  }
+  private def tagMatch(addr: UInt): UInt = tagMatch(addr, pageMatch(addr))
+  private def tagMatch(addr: UInt, pgMatch: UInt): UInt = {
+    val idx = addr(conf.matchBits-1,0)
+    val idxMatch = idxs.map(_ === idx).toBits
+    val idxPageMatch = idxPagesOH.map(_ & pgMatch).map(_.orR).toBits
+    idxValid.toBits & idxMatch & idxPageMatch
+  }
+
+  val hits = tagMatch(io.current_pc)
+  val idxPageMatch = pageMatch(io.correct_pc)
+  val tgtPageMatch = pageMatch(io.correct_target)
+  val updates = tagMatch(io.correct_pc, idxPageMatch)
+  val anyUpdates = updates.orR
+
+  private var lfsr = LFSR16(io.wen)
+  def rand(width: Int) = {
+    lfsr = lfsr(lfsr.getWidth-1,1)
+    Random.oneHot(width, lfsr)
+  }
+  def randOrInvalid(valid: UInt) =
+    Mux(!valid.andR, PriorityEncoderOH(~valid), rand(valid.getWidth))
+
+  val idxRepl = randOrInvalid(idxValid.toBits)
+  val idxWen = updates.toBits | idxRepl & ~anyUpdates.toSInt
+
+  val useIdxPageMatch = idxPageMatch.orR
+  val doIdxPageRepl = !useIdxPageMatch && io.taken
+  val idxPageRepl = rand(conf.pages)
+  val idxPageUpdate = Mux(useIdxPageMatch, idxPageMatch, idxPageRepl)
+  val idxPageReplEn = Mux(doIdxPageRepl, idxPageRepl, UInt(0))
+
+  val samePage = page(io.correct_pc) === page(io.correct_target)
+  val useTgtPageMatch = (tgtPageMatch & ~idxPageReplEn).orR
+  val doTgtPageRepl = !useTgtPageMatch && io.taken && !samePage
+  val tgtPageRepl = Mux(samePage, idxPageUpdate, idxPageUpdate(conf.pages-2,0) << 1 | idxPageUpdate(conf.pages-1))
+  val tgtPageUpdate = Mux(useTgtPageMatch, tgtPageMatch, tgtPageRepl)
+  val tgtPageReplEn = Mux(doTgtPageRepl, tgtPageRepl, UInt(0))
+
+  val pageReplEn = idxPageReplEn | tgtPageReplEn
+
+  when (io.wen) {
+    for (i <- 0 until conf.entries) {
+      when (idxWen(i)) {
+        idxValid(i) := io.taken
+        when (io.taken) {
+          idxs(i) := io.correct_pc
+          idxPages(i) := OHToUInt(idxPageUpdate)
+          tgts(i) := io.correct_target
+          tgtPages(i) := OHToUInt(tgtPageUpdate)
+        }
+      }.elsewhen ((pageReplEn & (idxPagesOH(i) | tgtPagesOH(i))).orR) {
+        idxValid(i) := false
+      }
+    }
+
+    require(conf.pages % 2 == 0)
+    val idxWritesEven = (idxPageUpdate & Fill(conf.pages/2, UInt(1,2))).orR
+
+    def writeBank(i: Int, mod: Int, en: Bool, data: UInt) = {
+      for (i <- i until conf.pages by mod) {
+        when (en && pageReplEn(i)) {
+          pages(i) := data
+          pageValid(i) := true
+        }
+      }
+    }
+    writeBank(0, 2, Mux(idxWritesEven, doIdxPageRepl, doTgtPageRepl),
+      Mux(idxWritesEven, page(io.correct_pc), page(io.correct_target)))
+    writeBank(1, 2, Mux(idxWritesEven, doTgtPageRepl, doIdxPageRepl),
+      Mux(idxWritesEven, page(io.correct_target), page(io.correct_pc)))
+  }
+
+  when (io.invalidate) {
+    idxValid.foreach(_ := false)
+    pageValid.foreach(_ := false)
+  }
+
+  io.hit    := hits.toBits.orR
+  io.target := Cat(Mux1H(Mux1H(hits, tgtPagesOH), pages), Mux1H(hits, tgts))
+}

rocket/src/main/scala/csr.scala

@@ -7,56 +7,6 @@ import uncore.HTIFIO
 import uncore.constants.AddressConstants._
 import scala.math._
 
-class DpathBTBIO extends Bundle
-{
-  val current_pc     = UInt(INPUT, VADDR_BITS)
-  val hit            = Bool(OUTPUT)
-  val target         = UInt(OUTPUT, VADDR_BITS)
-  val wen            = Bool(INPUT)
-  val clr            = Bool(INPUT)
-  val invalidate     = Bool(INPUT)
-  val correct_pc     = UInt(INPUT, VADDR_BITS)
-  val correct_target = UInt(INPUT, VADDR_BITS)
-}
-
-// fully-associative branch target buffer
-class rocketDpathBTB(entries: Int) extends Module
-{
-  val io = new DpathBTBIO
-
-  var hit_reduction = Bool(false)
-  val hit = Bool()
-  val update = Bool()
-  var update_reduction = Bool(false)
-  val valid = Vec.fill(entries){Reg(init=Bool(false))}
-  val hits = Vec.fill(entries){Bool()}
-  val updates = Vec.fill(entries){Bool()}
-  val targets = Vec.fill(entries){Reg(UInt())}
-  val anyUpdate = updates.toBits.orR
-
-  val random_way = Random(entries, io.wen)
-  val invalid_way = valid.indexWhere((x: Bool) => !x)
-  val repl_way = Mux(valid.contains(Bool(false)), invalid_way, random_way)
-
-  for (i <- 0 until entries) {
-    val tag = Reg(UInt())
-    hits(i) := valid(i) && tag === io.current_pc
-    updates(i) := valid(i) && tag === io.correct_pc
-
-    when (io.wen && (updates(i) || !anyUpdate && UInt(i) === repl_way)) {
-      valid(i) := Bool(false)
-      when (!io.clr) {
-        valid(i) := Bool(true)
-        tag := io.correct_pc
-        targets(i) := io.correct_target
-      }
-    }
-  }
-
-  io.hit    := hits.toBits.orR
-  io.target := Mux1H(hits, targets)
-}
-
 class Status extends Bundle {
   val ip = Bits(width = 8)
   val im = Bits(width = 8)
@@ -294,16 +244,3 @@ class CSRFile(implicit conf: RocketConfiguration) extends Module
     reg_status.ip := 0
   }
 }
-
-class ioReadPort(d: Int, w: Int) extends Bundle
-{
-  override def clone = new ioReadPort(d, w).asInstanceOf[this.type]
-}
-
-class ioWritePort(d: Int, w: Int) extends Bundle
-{
-  val addr = UInt(INPUT, log2Up(d))
-  val en   = Bool(INPUT)
-  val data = Bits(INPUT, w)
-  override def clone = new ioWritePort(d, w).asInstanceOf[this.type]
-}

rocket/src/main/scala/icache.scala

@@ -6,7 +6,7 @@ import Util._
 
 case class ICacheConfig(sets: Int, assoc: Int,
                         ibytes: Int = 4,
-                        ntlb: Int = 8, nbtb: Int = 8,
+                        ntlb: Int = 8, btb: BTBConfig = BTBConfig(8),
                         code: Code = new IdentityCode)
 {
   val w = 1
@@ -55,7 +55,7 @@ class Frontend(implicit c: ICacheConfig, tl: TileLinkConfiguration) extends Modu
     val mem = new UncachedTileLinkIO
   }
   
-  val btb = Module(new rocketDpathBTB(c.nbtb))
+  val btb = Module(new BTB(c.btb))
   val icache = Module(new ICache)
   val tlb = Module(new TLB(c.ntlb))
 
@@ -94,7 +94,7 @@ class Frontend(implicit c: ICacheConfig, tl: TileLinkConfiguration) extends Modu
 
   btb.io.current_pc := s1_pc
   btb.io.wen := io.cpu.req.bits.mispredict
-  btb.io.clr := !io.cpu.req.bits.taken
+  btb.io.taken := io.cpu.req.bits.taken
   btb.io.correct_pc := io.cpu.req.bits.currentpc
   btb.io.correct_target := io.cpu.req.bits.pc
   btb.io.invalidate := io.cpu.invalidate || io.cpu.ptw.invalidate

rocket/src/main/scala/util.scala

@@ -11,12 +11,15 @@ object Util {
   implicit def intToUInt(x: Int): UInt = UInt(x)
   implicit def booleanToBool(x: Boolean): Bits = Bool(x)
   implicit def intSeqToUIntSeq(x: Iterable[Int]): Iterable[UInt] = x.map(UInt(_))
+  implicit def seqToVec[T <: Data](x: Iterable[T]): Vec[T] = Vec(x)
   implicit def wcToUInt(c: WideCounter): UInt = c.value
 
   implicit def booleanToInt(x: Boolean): Int = if (x) 1 else 0
   implicit def intToBooleanToInt(x: Int): BooleanToInt = new BooleanToInt(x)
 }
 
+import Util._
+
 object AVec
 {
   def apply[T <: Data](elts: Seq[T]): Vec[T] = Vec(elts)
@@ -141,18 +144,20 @@ case class WideCounter(width: Int, inc: UInt = UInt(1))
 
 object Random
 {
-  def apply(mod: Int, inc: Bool = Bool(true)): UInt = {
+  def apply(mod: Int, random: UInt): UInt = {
-    if (isPow2(mod)) {
+    if (isPow2(mod)) random(log2Up(mod)-1,0)
-      require(mod <= 65536)
+    else PriorityEncoder(partition(apply(1 << log2Up(mod*8), random), mod))
-      LFSR16(inc)(log2Up(mod)-1,0).toUInt
-    } else {
-      val max = 1 << log2Up(mod*8)
-      val rand_pow2 = apply(max, inc)
-
-      var res = UInt(mod-1)
-      for (i <- mod-1 to 1 by -1)
-        res = Mux(rand_pow2 < UInt(i*max/mod), UInt(i-1), res)
-      res
-    }
   }
+  def apply(mod: Int): UInt = apply(mod, randomizer)
+  def oneHot(mod: Int, random: UInt): UInt = {
+    if (isPow2(mod)) UIntToOH(random(log2Up(mod)-1,0))
+    else PriorityEncoderOH(partition(apply(1 << log2Up(mod*8), random), mod)).toBits
+  }
+  def oneHot(mod: Int): UInt = oneHot(mod, randomizer)
+
+  private def randomizer = LFSR16()
+  private def round(x: Double): Int =
+    if (x.toInt.toDouble == x) x.toInt else (x.toInt + 1) & -2
+  private def partition(value: UInt, slices: Int) =
+    Vec.tabulate(slices)(i => value < round((i << value.getWidth).toDouble / slices))
 }