diff --git a/rocket/src/main/scala/btb.scala b/rocket/src/main/scala/btb.scala
index 02ca111e..e3762bfd 100644
--- a/rocket/src/main/scala/btb.scala
+++ b/rocket/src/main/scala/btb.scala
@@ -62,6 +62,14 @@ class BHT(nbht: Int) {
   val history = Reg(UInt(width = nbhtbits))
 }
 
+// BTB update occurs during branch resolution (i.e., PC redirection if a mispredict).
+//  - "pc" is what future fetch PCs will tag match against.
+//  - "br_pc" is the PC of the branch instruction.
+//  - "resp.mask" provides a mask of valid instructions (instructions are
+//      masked off by the predicted, taken branch).
+// Assumption: superscalar commits are batched together into a single
+// "taken" update ("history compression"), and correspond to the
+// superscalar fetch 1:1.
 class BTBUpdate extends Bundle with BTBParameters {
   val prediction = Valid(new BTBResp)
   val pc = UInt(width = vaddrBits)
@@ -71,11 +79,13 @@ class BTBUpdate extends Bundle with BTBParameters {
   val isJump = Bool()
   val isCall = Bool()
   val isReturn = Bool()
+  val br_pc = UInt(width = vaddrBits)
   val incorrectTarget = Bool()
 }
 
 class BTBResp extends Bundle with BTBParameters {
   val taken = Bool()
+  val mask = Bits(width = log2Up(params(FetchWidth)))
   val target = UInt(width = vaddrBits)
   val entry = UInt(width = opaqueBits)
   val bht = new BHTResp
@@ -102,6 +112,7 @@ class BTB extends Module with BTBParameters {
 
   val useRAS = Reg(UInt(width = entries))
   val isJump = Reg(UInt(width = entries))
+  val brIdx  = Mem(UInt(width=log2Up(params(FetchWidth))), entries)
 
   private def page(addr: UInt) = addr >> matchBits
   private def pageMatch(addr: UInt) = {
@@ -167,6 +178,7 @@ class BTB extends Module with BTBParameters {
       tgtPages(waddr) := tgtPageUpdate
       useRAS(waddr) := update.bits.isReturn
       isJump(waddr) := update.bits.isJump
+      brIdx(waddr)  := update.bits.br_pc
     }
 
     require(nPages % 2 == 0)
@@ -193,6 +205,7 @@ class BTB extends Module with BTBParameters {
   io.resp.bits.taken := io.resp.valid
   io.resp.bits.target := Cat(Mux1H(Mux1H(hits, tgtPagesOH), pages), Mux1H(hits, tgts))
   io.resp.bits.entry := OHToUInt(hits)
+  io.resp.bits.mask := Cat((UInt(1) << brIdx(io.resp.bits.entry))-1, UInt(1)) 
 
   if (nBHT > 0) {
     val bht = new BHT(nBHT)
diff --git a/rocket/src/main/scala/core.scala b/rocket/src/main/scala/core.scala
index f02f1424..43409b46 100644
--- a/rocket/src/main/scala/core.scala
+++ b/rocket/src/main/scala/core.scala
@@ -9,6 +9,7 @@ import uncore._
 case object BuildFPU extends Field[Option[() => FPU]]
 case object XprLen extends Field[Int]
 case object NMultXpr extends Field[Int]
+case object FetchWidth extends Field[Int]
 case object RetireWidth extends Field[Int]
 case object UseVM extends Field[Boolean]
 case object FastLoadWord extends Field[Boolean]
@@ -20,6 +21,7 @@ case object CoreDCacheReqTagBits extends Field[Int]
 
 abstract trait CoreParameters extends UsesParameters {
   val xprLen = params(XprLen)
+  val coreFetchWidth = params(FetchWidth)
   val coreInstBits = params(CoreInstBits)
   val coreInstBytes = coreInstBits/8
   val coreDataBits = xprLen
@@ -28,6 +30,7 @@ abstract trait CoreParameters extends UsesParameters {
   val coreMaxAddrBits = math.max(params(PPNBits),params(VPNBits)+1) + params(PgIdxBits)
 
   if(params(FastLoadByte)) require(params(FastLoadWord))
+  require(params(FetchWidth) == 1)  // for now...
   require(params(RetireWidth) == 1) // for now...
 }
 abstract class CoreBundle extends Bundle with CoreParameters
diff --git a/rocket/src/main/scala/dpath.scala b/rocket/src/main/scala/dpath.scala
index 7df1a8d3..ef4f353c 100644
--- a/rocket/src/main/scala/dpath.scala
+++ b/rocket/src/main/scala/dpath.scala
@@ -47,9 +47,9 @@ class Datapath extends Module
   val wb_reg_rs2 = Reg(Bits())
 
   // instruction decode stage
-  val id_inst = io.imem.resp.bits.data
+  val id_inst = io.imem.resp.bits.data(0).toBits; require(params(FetchWidth) == 1)
   val id_pc = io.imem.resp.bits.pc
- 
+
   class RegFile {
     private val rf = Mem(UInt(width = 64), 31)
     private val reads = collection.mutable.ArrayBuffer[(UInt,UInt)]()
diff --git a/rocket/src/main/scala/icache.scala b/rocket/src/main/scala/icache.scala
index fb2dbb0c..05c36902 100644
--- a/rocket/src/main/scala/icache.scala
+++ b/rocket/src/main/scala/icache.scala
@@ -1,5 +1,3 @@
-// See LICENSE for license details.
-
 package rocket
 
 import Chisel._
@@ -24,7 +22,8 @@ class FrontendReq extends CoreBundle {
 
 class FrontendResp extends CoreBundle {
   val pc = UInt(width = params(VAddrBits)+1)  // ID stage PC
-  val data = Bits(width = coreInstBits)
+  val data = Vec.fill(coreFetchWidth) (Bits(width = coreInstBits))
+  val mask = Bits(width = coreFetchWidth)
   val xcpt_ma = Bool()
   val xcpt_if = Bool()
 }
@@ -60,12 +59,12 @@ class Frontend extends FrontendModule
 
   val msb = vaddrBits-1
   val btbTarget = Cat(btb.io.resp.bits.target(msb), btb.io.resp.bits.target)
-  val pcp4_0 = s1_pc + UInt(coreInstBytes)
-  val pcp4 = Cat(s1_pc(msb) & pcp4_0(msb), pcp4_0(msb,0))
+  val ntpc_0 = s1_pc + UInt(coreInstBytes)
+  val ntpc = Cat(s1_pc(msb) & ntpc_0(msb), ntpc_0(msb,0))
   val icmiss = s2_valid && !icache.io.resp.valid
-  val predicted_npc = Mux(btb.io.resp.bits.taken, btbTarget, pcp4)
+  val predicted_npc = Mux(btb.io.resp.bits.taken, btbTarget, ntpc)
   val npc = Mux(icmiss, s2_pc, predicted_npc).toUInt
-  val s0_same_block = !icmiss && !io.cpu.req.valid && !btb.io.resp.bits.taken && ((pcp4 & rowBytes) === (s1_pc & rowBytes))
+  val s0_same_block = !icmiss && !io.cpu.req.valid && !btb.io.resp.bits.taken && ((ntpc & rowBytes) === (s1_pc & rowBytes))
 
   val stall = io.cpu.resp.valid && !io.cpu.resp.ready
   when (!stall) {
@@ -106,7 +105,17 @@ class Frontend extends FrontendModule
 
   io.cpu.resp.valid := s2_valid && (s2_xcpt_if || icache.io.resp.valid)
   io.cpu.resp.bits.pc := s2_pc & SInt(-coreInstBytes) // discard PC LSBs
-  io.cpu.resp.bits.data := icache.io.resp.bits.datablock >> (s2_pc(log2Up(rowBytes)-1,log2Up(coreInstBytes)) << log2Up(coreInstBits))
+
+
+  val fetch_data = icache.io.resp.bits.datablock >> (s2_pc(log2Up(rowBytes)-1,log2Up(coreFetchWidth*coreInstBytes)) << log2Up(coreFetchWidth*coreInstBits))
+  for (i <- 0 until coreFetchWidth) {
+    io.cpu.resp.bits.data(i) := fetch_data(i*coreInstBits+coreInstBits-1, i*coreInstBits)
+  }
+
+  val all_ones = UInt((1 << coreFetchWidth)-1)
+  val msk_pc = all_ones << s2_pc(log2Up(coreFetchWidth)-1+2,2)
+  io.cpu.resp.bits.mask := msk_pc & btb.io.resp.bits.mask
+
   io.cpu.resp.bits.xcpt_ma := s2_pc(log2Up(coreInstBytes)-1,0) != UInt(0)
   io.cpu.resp.bits.xcpt_if := s2_xcpt_if