diff --git a/rocket/src/main/scala/consts.scala b/rocket/src/main/scala/consts.scala
index 030753d0..4d7cb68d 100644
--- a/rocket/src/main/scala/consts.scala
+++ b/rocket/src/main/scala/consts.scala
@@ -21,7 +21,7 @@ trait ScalarOpConstants {
   val PC_EX  = UInt(0, 2)
   val PC_MEM = UInt(1, 2)
   val PC_WB  = UInt(2, 2)
-  val PC_PCR = UInt(3, 2)
+  val PC_CSR = UInt(3, 2)
 
   val A1_X    = Bits("b??", 2)
   val A1_ZERO = UInt(0, 2)
diff --git a/rocket/src/main/scala/csr.scala b/rocket/src/main/scala/csr.scala
index 8e0c875a..fa263d19 100644
--- a/rocket/src/main/scala/csr.scala
+++ b/rocket/src/main/scala/csr.scala
@@ -93,7 +93,6 @@ class CSRFileIO extends CoreBundle {
   val uarch_counters = Vec.fill(16)(UInt(INPUT, log2Up(1+retireWidth)))
   val custom_mrw_csrs = Vec.fill(params(NCustomMRWCSRs))(UInt(INPUT, xLen))
   val cause = UInt(INPUT, xLen)
-  val mbadaddr_wen = Bool(INPUT)
   val pc = SInt(INPUT, vaddrBits+1)
   val fatc = Bool(OUTPUT)
   val time = UInt(OUTPUT, xLen)
diff --git a/rocket/src/main/scala/ctrl.scala b/rocket/src/main/scala/ctrl.scala
index b083371c..3e0e3a8c 100644
--- a/rocket/src/main/scala/ctrl.scala
+++ b/rocket/src/main/scala/ctrl.scala
@@ -553,7 +553,7 @@ class Control extends CoreModule
   take_pc_wb := replay_wb || wb_xcpt || io.dpath.eret
 
   io.dpath.sel_pc :=
-    Mux(wb_xcpt || io.dpath.eret,     PC_PCR, // exception or [m|s]ret
+    Mux(wb_xcpt || io.dpath.eret,     PC_CSR, // exception or [m|s]ret
     Mux(replay_wb,                    PC_WB,  // replay
                                       PC_MEM))
 
@@ -589,7 +589,7 @@ class Control extends CoreModule
     io.dpath.bypass_src(i) := PriorityEncoder(doBypass(i))
   }
 
-  // stall for RAW/WAW hazards on PCRs, loads, AMOs, and mul/div in execute stage.
+  // stall for RAW/WAW hazards on CSRs, loads, AMOs, and mul/div in execute stage.
   val id_renx1_not0 = id_ctrl.rxs1 && id_raddr1 != UInt(0)
   val id_renx2_not0 = id_ctrl.rxs2 && id_raddr2 != UInt(0)
   val id_wen_not0 = id_ctrl.wxd && id_waddr != UInt(0)
@@ -605,7 +605,7 @@ class Control extends CoreModule
      io.fpu.dec.wen  && id_waddr  === io.dpath.ex_waddr)
   val id_ex_hazard = ex_reg_valid && (data_hazard_ex && ex_cannot_bypass || fp_data_hazard_ex)
 
-  // stall for RAW/WAW hazards on PCRs, LB/LH, and mul/div in memory stage.
+  // stall for RAW/WAW hazards on CSRs, LB/LH, and mul/div in memory stage.
   val mem_mem_cmd_bh =
     if (params(FastLoadWord)) Bool(!params(FastLoadByte)) && mem_reg_slow_bypass
     else Bool(true)
@@ -674,7 +674,7 @@ class Control extends CoreModule
   io.dmem.req.bits.cmd  := ex_ctrl.mem_cmd
   io.dmem.req.bits.typ  := ex_ctrl.mem_type
   io.dmem.req.bits.phys := Bool(false)
-  io.dmem.sret := wb_xcpt // obviously not an sret, but sufficient
+  io.dmem.invalidate_lr := wb_xcpt
 
   io.rocc.cmd.valid := wb_rocc_val
   io.rocc.exception := wb_xcpt && io.dpath.status.xs.orR
diff --git a/rocket/src/main/scala/dpath.scala b/rocket/src/main/scala/dpath.scala
index bc25a1bd..a60ff72f 100644
--- a/rocket/src/main/scala/dpath.scala
+++ b/rocket/src/main/scala/dpath.scala
@@ -166,17 +166,17 @@ class Datapath extends CoreModule
   require(params(CoreDCacheReqTagBits) >= 6)
 
   // processor control regfile read
-  val pcr = Module(new CSRFile)
-  pcr.io.host <> io.host
-  pcr.io <> io.ctrl
-  pcr.io <> io.fpu
-  pcr.io.rocc <> io.rocc
-  pcr.io.pc := wb_reg_pc
-  pcr.io.uarch_counters.foreach(_ := Bool(false))
+  val csr = Module(new CSRFile)
+  csr.io.host <> io.host
+  csr.io <> io.ctrl
+  csr.io <> io.fpu
+  csr.io.rocc <> io.rocc
+  csr.io.pc := wb_reg_pc
+  csr.io.uarch_counters.foreach(_ := Bool(false))
 
-  io.ptw.ptbr := pcr.io.ptbr
-  io.ptw.invalidate := pcr.io.fatc
-  io.ptw.status := pcr.io.status
+  io.ptw.ptbr := csr.io.ptbr
+  io.ptw.invalidate := csr.io.fatc
+  io.ptw.status := csr.io.status
 
   // memory stage
   mem_reg_kill := ex_reg_kill
@@ -246,7 +246,7 @@ class Datapath extends CoreModule
   }
   wb_wdata := Mux(dmem_resp_valid && dmem_resp_xpu, io.dmem.resp.bits.data_subword,
               Mux(io.ctrl.ll_wen, ll_wdata,
-              Mux(io.ctrl.csr_cmd != CSR.N, pcr.io.rw.rdata,
+              Mux(io.ctrl.csr_cmd != CSR.N, csr.io.rw.rdata,
               wb_reg_wdata)))
 
   val wb_wen = io.ctrl.ll_wen || io.ctrl.wb_wen
@@ -258,9 +258,9 @@ class Datapath extends CoreModule
   io.ctrl.fp_sboard_clra := dmem_resp_waddr
 
   // processor control regfile write
-  pcr.io.rw.addr := wb_reg_inst(31,20)
-  pcr.io.rw.cmd := io.ctrl.csr_cmd
-  pcr.io.rw.wdata := wb_reg_wdata
+  csr.io.rw.addr := wb_reg_inst(31,20)
+  csr.io.rw.cmd := io.ctrl.csr_cmd
+  csr.io.rw.wdata := wb_reg_wdata
 
   io.rocc.cmd.bits.inst := new RoCCInstruction().fromBits(wb_reg_inst)
   io.rocc.cmd.bits.rs1 := wb_reg_wdata
@@ -269,7 +269,7 @@ class Datapath extends CoreModule
   // hook up I$
   io.imem.req.bits.pc :=
     Mux(io.ctrl.sel_pc === PC_MEM, mem_npc,
-    Mux(io.ctrl.sel_pc === PC_PCR, pcr.io.evec,
+    Mux(io.ctrl.sel_pc === PC_CSR, csr.io.evec,
         wb_reg_pc)).toUInt // PC_WB
   io.imem.btb_update.bits.pc := mem_reg_pc
   io.imem.btb_update.bits.target := io.imem.req.bits.pc
@@ -283,7 +283,7 @@ class Datapath extends CoreModule
   io.ctrl.wb_waddr := wb_reg_inst(11,7)
 
   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.host.id, pcr.io.time(32,0), io.ctrl.retire, wb_reg_pc,
+         io.host.id, csr.io.time(32,0), io.ctrl.retire, wb_reg_pc,
          Mux(wb_wen, wb_waddr, UInt(0)), wb_wdata, wb_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))),
diff --git a/rocket/src/main/scala/icache.scala b/rocket/src/main/scala/icache.scala
index d7a820db..1103f22d 100644
--- a/rocket/src/main/scala/icache.scala
+++ b/rocket/src/main/scala/icache.scala
@@ -114,8 +114,11 @@ class Frontend(btb_updates_out_of_order: Boolean = false) extends FrontendModule
   io.cpu.resp.valid := s2_valid && (s2_xcpt_if || icache.io.resp.valid)
   io.cpu.resp.bits.pc := s2_pc
 
+  require(coreFetchWidth * coreInstBytes <= rowBytes)
+  val fetch_data =
+    if (coreFetchWidth * coreInstBytes == rowBytes) icache.io.resp.bits.datablock
+    else icache.io.resp.bits.datablock >> (s2_pc(log2Up(rowBytes)-1,log2Up(coreFetchWidth*coreInstBytes)) << log2Up(coreFetchWidth*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)
   }
diff --git a/rocket/src/main/scala/nbdcache.scala b/rocket/src/main/scala/nbdcache.scala
index 3f24ba4c..3b55ec8c 100644
--- a/rocket/src/main/scala/nbdcache.scala
+++ b/rocket/src/main/scala/nbdcache.scala
@@ -83,7 +83,7 @@ class HellaCacheIO extends CoreBundle {
   val resp = Valid(new HellaCacheResp).flip
   val replay_next = Valid(Bits(width = coreDCacheReqTagBits)).flip
   val xcpt = (new HellaCacheExceptions).asInput
-  val sret = Bool(OUTPUT)
+  val invalidate_lr = Bool(OUTPUT)
   val ordered = Bool(INPUT)
 }
 
@@ -686,6 +686,10 @@ class HellaCache extends L1HellaCacheModule {
   io.cpu.xcpt.pf.ld := s1_read && dtlb.io.resp.xcpt_ld
   io.cpu.xcpt.pf.st := s1_write && dtlb.io.resp.xcpt_st
 
+  assert (!(Reg(next=
+    (io.cpu.xcpt.ma.ld || io.cpu.xcpt.ma.st || io.cpu.xcpt.pf.ld || io.cpu.xcpt.pf.st)) &&
+    io.cpu.resp.valid), "DCache exception occurred - cache response not killed.")
+
   // tags
   def onReset = L1Metadata(UInt(0), ClientMetadata.onReset)
   val meta = Module(new MetadataArray(onReset _))
@@ -752,7 +756,7 @@ class HellaCache extends L1HellaCacheModule {
       lrsc_count := 0
     }
   }
-  when (io.cpu.sret) { lrsc_count := 0 }
+  when (io.cpu.invalidate_lr) { lrsc_count := 0 }
 
   val s2_data = Vec.fill(nWays){Bits(width = encRowBits)}
   for (w <- 0 until nWays) {
diff --git a/rocket/src/main/scala/rocc.scala b/rocket/src/main/scala/rocc.scala
index bf49aca5..da37785d 100644
--- a/rocket/src/main/scala/rocc.scala
+++ b/rocket/src/main/scala/rocc.scala
@@ -125,7 +125,7 @@ class AccumulatorExample extends RoCC
   io.mem.req.bits.cmd := M_XRD // perform a load (M_XWR for stores)
   io.mem.req.bits.typ := MT_D // D = 8 bytes, W = 4, H = 2, B = 1
   io.mem.req.bits.data := Bits(0) // we're not performing any stores...
-  io.mem.sret := false
+  io.mem.invalidate_lr := false
 
   io.imem.acquire.valid := false
   io.imem.grant.ready := false
diff --git a/rocket/src/main/scala/tile.scala b/rocket/src/main/scala/tile.scala
index def17452..edb5566c 100644
--- a/rocket/src/main/scala/tile.scala
+++ b/rocket/src/main/scala/tile.scala
@@ -25,7 +25,7 @@ class RocketTile(resetSignal: Bool = null) extends Tile(resetSignal) {
   val ptw = Module(new PTW(params(NPTWPorts)))
   val core = Module(new Core, { case CoreName => "Rocket" })
 
-  dcache.io.cpu.sret := core.io.dmem.sret // Bypass sret to dcache
+  dcache.io.cpu.invalidate_lr := core.io.dmem.invalidate_lr // Bypass signal to dcache
   val dcArb = Module(new HellaCacheArbiter(params(NDCachePorts)))
   dcArb.io.requestor(0) <> ptw.io.mem
   dcArb.io.requestor(1) <> core.io.dmem