diff --git a/rocket/src/main/scala/dcache.scala b/rocket/src/main/scala/dcache.scala
index bb11c55f..484a39a9 100644
--- a/rocket/src/main/scala/dcache.scala
+++ b/rocket/src/main/scala/dcache.scala
@@ -51,7 +51,7 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   val replacer = p(Replacer)()
   def onReset = L1Metadata(UInt(0), ClientMetadata.onReset)
   val meta = Module(new MetadataArray(onReset _))
-  val metaReadArb = Module(new Arbiter(new MetaReadReq, 2))
+  val metaReadArb = Module(new Arbiter(new MetaReadReq, 3))
   val metaWriteArb = Module(new Arbiter(new L1MetaWriteReq, 3))
   meta.io.read <> metaReadArb.io.out
   meta.io.write <> metaWriteArb.io.out
@@ -68,12 +68,19 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   val s1_nack = Wire(init=Bool(false))
   val s1_valid_masked = s1_valid && !io.cpu.s1_kill
   val s1_valid_not_nacked = s1_valid_masked && !s1_nack
-  val s1_req = RegEnable(io.cpu.req.bits, io.cpu.req.valid)
+  val s1_req = Reg(io.cpu.req.bits)
+  when (metaReadArb.io.out.valid) {
+    s1_req := io.cpu.req.bits
+    s1_req.addr := Cat(io.cpu.req.bits.addr >> pgIdxBits, metaReadArb.io.out.bits.idx, io.cpu.req.bits.addr(blockOffBits-1,0))
+  }
   val s1_read = isRead(s1_req.cmd)
   val s1_write = isWrite(s1_req.cmd)
+  val s1_readwrite = s1_read || s1_write
+  val s1_flush_valid = Reg(Bool())
 
   val s_ready :: s_grant_wait :: s_voluntary_writeback :: s_probe_rep_dirty :: s_probe_rep_clean :: s_probe_rep_miss :: s_voluntary_write_meta :: s_probe_write_meta :: Nil = Enum(UInt(), 8)
   val grant_wait = Reg(init=Bool(false))
+  val release_ack_wait = Reg(init=Bool(false))
   val release_state = Reg(init=s_ready)
   val pstore1_valid = Wire(Bool())
   val pstore2_valid = Reg(Bool())
@@ -87,22 +94,22 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   dataArb.io.in(3).bits.addr := io.cpu.req.bits.addr
   dataArb.io.in(3).bits.way_en := ~UInt(0, nWays)
   when (!dataArb.io.in(3).ready && isRead(io.cpu.req.bits.cmd)) { io.cpu.req.ready := false }
-  metaReadArb.io.in(1).valid := io.cpu.req.valid
-  metaReadArb.io.in(1).bits.idx := io.cpu.req.bits.addr(idxMSB, idxLSB)
-  metaReadArb.io.in(1).bits.way_en := ~UInt(0, nWays)
-  when (!metaReadArb.io.in(1).ready) { io.cpu.req.ready := false }
+  metaReadArb.io.in(2).valid := io.cpu.req.valid
+  metaReadArb.io.in(2).bits.idx := io.cpu.req.bits.addr(idxMSB, idxLSB)
+  metaReadArb.io.in(2).bits.way_en := ~UInt(0, nWays)
+  when (!metaReadArb.io.in(2).ready) { io.cpu.req.ready := false }
 
   // address translation
   val tlb = Module(new TLB)
   io.ptw <> tlb.io.ptw
-  tlb.io.req.valid := s1_valid_masked
+  tlb.io.req.valid := s1_valid_masked && s1_readwrite
   tlb.io.req.bits.passthrough := s1_req.phys
   tlb.io.req.bits.asid := 0
   tlb.io.req.bits.vpn := s1_req.addr >> pgIdxBits
   tlb.io.req.bits.instruction := false
   tlb.io.req.bits.store := s1_write
   when (!tlb.io.req.ready && !io.cpu.req.bits.phys) { io.cpu.req.ready := false }
-  when (s1_valid && tlb.io.resp.miss) { s1_nack := true }
+  when (s1_valid && s1_readwrite && tlb.io.resp.miss) { s1_nack := true }
 
   val s1_paddr = Cat(tlb.io.resp.ppn, s1_req.addr(pgIdxBits-1,0))
   val s1_tag = Mux(s1_probe, probe_bits.addr_block >> idxBits, s1_paddr(paddrBits-1, untagBits))
@@ -112,30 +119,37 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
     .reduce (_|_))
   val s1_data_way = Mux(inWriteback, releaseWay, s1_hit_way)
   val s1_data = Mux1H(s1_data_way, data.io.resp) // retime into s2 if critical
+  val s1_victim_way = Wire(init = replacer.way)
 
   val s2_valid = Reg(next=s1_valid_masked, init=Bool(false))
   val s2_probe = Reg(next=s1_probe, init=Bool(false))
   val releaseInFlight = s1_probe || s2_probe || release_state =/= s_ready
   val s2_valid_masked = s2_valid && Reg(next = !s1_nack)
   val s2_req = Reg(io.cpu.req.bits)
-  when (s1_valid_not_nacked) {
+  when (s1_valid_not_nacked || s1_flush_valid) {
     s2_req := s1_req
     s2_req.addr := s1_paddr
   }
+  val s2_read = isRead(s2_req.cmd)
+  val s2_write = isWrite(s2_req.cmd)
+  val s2_readwrite = s2_read || s2_write
+  val s2_flush_valid = RegNext(s1_flush_valid)
   val s2_data = RegEnable(s1_data, s1_valid || inWriteback)
   val s2_probe_way = RegEnable(s1_hit_way, s1_probe)
   val s2_probe_state = RegEnable(s1_hit_state, s1_probe)
   val s2_hit_way = RegEnable(s1_hit_way, s1_valid_not_nacked)
   val s2_hit_state = RegEnable(s1_hit_state, s1_valid_not_nacked)
   val s2_hit = s2_hit_state.isHit(s2_req.cmd)
-  val s2_valid_hit = s2_valid_masked && s2_hit
-  val s2_valid_miss = s2_valid_masked && !s2_hit && !(pstore1_valid || pstore2_valid)
+  val s2_valid_hit = s2_valid_masked && s2_readwrite && s2_hit
+  val s2_valid_miss = s2_valid_masked && s2_readwrite && !s2_hit && !(pstore1_valid || pstore2_valid) && !release_ack_wait
   val s2_uncached = !addrMap.isCacheable(s2_req.addr)
   val s2_valid_cached_miss = s2_valid_miss && !s2_uncached
+  val s2_victimize = s2_valid_cached_miss || s2_flush_valid
   val s2_valid_uncached = s2_valid_miss && s2_uncached
-  val s2_victim_way = Mux(s2_hit_state.isValid(), s2_hit_way, UIntToOH(replacer.way))
-  val s2_victim_tag = RegEnable(meta.io.resp(replacer.way).tag, s1_valid_not_nacked)
-  val s2_victim_state = Mux(s2_hit_state.isValid(), s2_hit_state, RegEnable(meta.io.resp(replacer.way).coh, s1_valid_not_nacked))
+  val s2_victim_way = Mux(s2_hit_state.isValid() && !s2_flush_valid, s2_hit_way, UIntToOH(RegEnable(s1_victim_way, s1_valid_not_nacked || s1_flush_valid)))
+  val s2_victim_tag = RegEnable(meta.io.resp(s1_victim_way).tag, s1_valid_not_nacked || s1_flush_valid)
+  val s2_victim_state = Mux(s2_hit_state.isValid() && !s2_flush_valid, s2_hit_state, RegEnable(meta.io.resp(s1_victim_way).coh, s1_valid_not_nacked || s1_flush_valid))
+  val s2_victim_valid = s2_victim_state.isValid()
   val s2_victim_dirty = s2_victim_state.requiresVoluntaryWriteback()
   io.cpu.s2_nack := s2_valid && !s2_valid_hit && !(s2_valid_uncached && io.mem.acquire.ready)
   when (s2_valid && !s2_valid_hit) { s1_nack := true }
@@ -148,7 +162,7 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   io.cpu.xcpt.pf.st := s1_write && tlb.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),
+    s2_valid_masked),
       "DCache exception occurred - cache response not killed.")
 
   // load reservations
@@ -181,7 +195,7 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
     Bool(usingAtomics) && pstore_drain_structural ||
     (((pstore1_valid && !pstore1_amo) || pstore2_valid) && (pstore_drain_opportunistic || pstore_drain_on_miss))
   pstore1_valid := {
-    val s2_store_valid = s2_valid_hit && isWrite(s2_req.cmd) && !s2_sc_fail
+    val s2_store_valid = s2_valid_hit && s2_write && !s2_sc_fail
     val pstore1_held = Reg(Bool())
     assert(!s2_store_valid || !pstore1_held)
     pstore1_held := (s2_store_valid || pstore1_held) && pstore2_valid && !pstore_drain
@@ -211,7 +225,7 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   when (s1_valid && s1_raw_hazard) { s1_nack := true }
 
   val s2_new_hit_state = s2_hit_state.onHit(s2_req.cmd)
-  metaWriteArb.io.in(0).valid := (s2_valid_hit && s2_hit_state =/= s2_new_hit_state) || (s2_valid_cached_miss && !s2_victim_dirty)
+  metaWriteArb.io.in(0).valid := (s2_valid_hit && s2_hit_state =/= s2_new_hit_state) || (s2_victimize && !s2_victim_dirty)
   metaWriteArb.io.in(0).bits.way_en := s2_victim_way
   metaWriteArb.io.in(0).bits.idx := s2_req.addr(idxMSB, idxLSB)
   metaWriteArb.io.in(0).bits.data.coh := Mux(s2_hit, s2_new_hit_state, ClientMetadata.onReset)
@@ -252,7 +266,7 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   when (s2_uncached) {
     assert(!s2_valid_masked || !s2_hit_state.isValid(), "cache hit on uncached access")
     io.mem.acquire.bits := uncachedGetMessage
-    when (isWrite(s2_req.cmd)) {
+    when (s2_write) {
       io.mem.acquire.bits := uncachedPutMessage
       when (pstore1_amo) {
         io.mem.acquire.bits := uncachedPutAtomicMessage
@@ -266,8 +280,9 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   val grantIsVoluntary = io.mem.grant.bits.isVoluntary()
   val grantIsUncached = !grantIsRefill && !grantIsVoluntary
   when (io.mem.grant.valid) {
-    assert(grant_wait || grantIsVoluntary, "unexpected grant")
+    assert(grant_wait || grantIsVoluntary && release_ack_wait, "unexpected grant")
     when (grantIsUncached) { s2_data := io.mem.grant.bits.data }
+    when (grantIsVoluntary) { release_ack_wait := false }
   }
   val (refillCount, refillDone) = Counter(io.mem.grant.fire() && grantIsRefill, refillCycles)
   val grantDone = refillDone || grantIsUncached
@@ -299,10 +314,10 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
 
   // probe
   val block_probe = releaseInFlight || lrscValid || (s2_valid_hit && s2_lr)
-  metaReadArb.io.in(0).valid := io.mem.probe.valid && !block_probe
-  io.mem.probe.ready := metaReadArb.io.in(0).ready && !block_probe && !s1_valid && (!s2_valid || s2_valid_hit)
-  metaReadArb.io.in(0).bits.idx := io.mem.probe.bits.addr_block
-  metaReadArb.io.in(0).bits.way_en := ~UInt(0, nWays)
+  metaReadArb.io.in(1).valid := io.mem.probe.valid && !block_probe
+  io.mem.probe.ready := metaReadArb.io.in(1).ready && !block_probe && !s1_valid && (!s2_valid || s2_valid_hit)
+  metaReadArb.io.in(1).bits.idx := io.mem.probe.bits.addr_block
+  metaReadArb.io.in(1).bits.way_en := ~UInt(0, nWays)
 
   // release
   val (writebackCount, writebackDone) = Counter(io.mem.release.fire() && inWriteback, refillCycles)
@@ -319,7 +334,7 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   val probeNewCoh = s2_probe_state.onProbe(probe_bits)
   val newCoh = Wire(init = probeNewCoh)
   releaseWay := s2_probe_way
-  when (s2_valid_cached_miss && s2_victim_dirty) {
+  when (s2_victimize && s2_victim_dirty) {
     assert(!s2_hit_state.isValid())
     release_state := s_voluntary_writeback
     probe_bits.addr_block := Cat(s2_victim_tag, s2_req.addr(idxMSB, idxLSB))
@@ -344,7 +359,10 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
     io.mem.release.bits := voluntaryReleaseMessage
     newCoh := voluntaryNewCoh
     releaseWay := s2_victim_way
-    when (releaseDone) { release_state := s_voluntary_write_meta }
+    when (releaseDone) {
+      release_state := s_voluntary_write_meta
+      release_ack_wait := true
+    }
   }
   when (s2_probe && !io.mem.release.fire()) { s1_nack := true }
   io.mem.release.bits.addr_block := probe_bits.addr_block
@@ -366,7 +384,7 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
   // cached response
   io.cpu.resp.valid := s2_valid_hit
   io.cpu.resp.bits := s2_req
-  io.cpu.resp.bits.has_data := isRead(s2_req.cmd)
+  io.cpu.resp.bits.has_data := s2_read
   io.cpu.resp.bits.replay := false
   io.cpu.ordered := !(s1_valid || s2_valid || grant_wait)
 
@@ -399,6 +417,33 @@ class DCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
     amoalu.io.rhs := pstore1_data
     pstore1_storegen_data := amoalu.io.out
   } else {
-    assert(!(s1_valid_masked && isRead(s1_req.cmd) && isWrite(s1_req.cmd)), "unsupported D$ operation")
+    assert(!(s1_valid_masked && s1_read && s1_write), "unsupported D$ operation")
+  }
+
+  // flushes
+  val flushed = Reg(init=Bool(true))
+  val flushing = Reg(init=Bool(false))
+  val flushCounter = Counter(nSets * nWays)
+  when (io.mem.acquire.fire()) { flushed := false }
+  when (s2_valid_masked && s2_req.cmd === M_FLUSH_ALL) {
+    io.cpu.s2_nack := !flushed
+    when (!flushed) {
+      flushing := !release_ack_wait
+    }
+  }
+  s1_flush_valid := metaReadArb.io.in(0).fire() && !s1_flush_valid && !s2_flush_valid && release_state === s_ready && !release_ack_wait
+  metaReadArb.io.in(0).valid := flushing
+  metaReadArb.io.in(0).bits.idx := flushCounter.value
+  metaReadArb.io.in(0).bits.way_en := ~UInt(0, nWays)
+  when (flushing) {
+    s1_victim_way := flushCounter.value >> log2Up(nSets)
+    when (s2_flush_valid) {
+      when (flushCounter.inc()) {
+        flushed := true
+      }
+    }
+    when (flushed && release_state === s_ready && !release_ack_wait) {
+      flushing := false
+    }
   }
 }
diff --git a/rocket/src/main/scala/idecode.scala b/rocket/src/main/scala/idecode.scala
index 6b02baa1..33319e93 100644
--- a/rocket/src/main/scala/idecode.scala
+++ b/rocket/src/main/scala/idecode.scala
@@ -43,15 +43,15 @@ class IntCtrlSigs extends Bundle {
   val amo = Bool()
 
   def default: List[BitPat] =
-                //           jal                                                               renf1             fence.i
-                //   val     | jalr                                                            | renf2           |
-                //   | fp_val| | renx2                                                         | | renf3         |
-                //   | | rocc| | | renx1     s_alu1                          mem_val           | | | wfd         | 
-                //   | | | br| | | | s_alu2  |       imm    dw     alu       | mem_cmd mem_type| | | | div       | 
-                //   | | | | | | | | |       |       |      |      |         | |         |     | | | | | wxd     | fence
-                //   | | | | | | | | |       |       |      |      |         | |         |     | | | | | | csr   | | amo
-                //   | | | | | | | | |       |       |      |      |         | |         |     | | | | | | |     | | |
-                List(N,X,X,X,X,X,X,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, X,X,X,X,X,X,CSR.X,X,X,X)
+                //           jal                                                                 renf1             fence.i
+                //   val     | jalr                                                              | renf2           |
+                //   | fp_val| | renx2                                                           | | renf3         |
+                //   | | rocc| | | renx1     s_alu1                          mem_val             | | | wfd         | 
+                //   | | | br| | | | s_alu2  |       imm    dw     alu       | mem_cmd   mem_type| | | | div       | 
+                //   | | | | | | | | |       |       |      |      |         | |           |     | | | | | wxd     | fence
+                //   | | | | | | | | |       |       |      |      |         | |           |     | | | | | | csr   | | amo
+                //   | | | | | | | | |       |       |      |      |         | |           |     | | | | | | |     | | |
+                List(N,X,X,X,X,X,X,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, X,X,X,X,X,X,CSR.X,X,X,X)
 
   def decode(inst: UInt, table: Iterable[(BitPat, List[BitPat])]) = {
     val decoder = DecodeLogic(inst, default, table)
@@ -66,243 +66,243 @@ class IntCtrlSigs extends Bundle {
 class IDecode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    BNE->       List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SNE,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    BEQ->       List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SEQ,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    BLT->       List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SLT,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    BLTU->      List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SLTU,  N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    BGE->       List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SGE,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    BGEU->      List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SGEU,  N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    BNE->       List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SNE,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    BEQ->       List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SEQ,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    BLT->       List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SLT,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    BLTU->      List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SLTU,  N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    BGE->       List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SGE,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    BGEU->      List(Y,N,N,Y,N,N,Y,Y,A2_RS2, A1_RS1, IMM_SB,DW_X,  FN_SGEU,  N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
 
-    JAL->       List(Y,N,N,N,Y,N,N,N,A2_FOUR,A1_PC,  IMM_UJ,DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    JALR->      List(Y,N,N,N,N,Y,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    AUIPC->     List(Y,N,N,N,N,N,N,N,A2_IMM, A1_PC,  IMM_U, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    JAL->       List(Y,N,N,N,Y,N,N,N,A2_FOUR,A1_PC,  IMM_UJ,DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    JALR->      List(Y,N,N,N,N,Y,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    AUIPC->     List(Y,N,N,N,N,N,N,N,A2_IMM, A1_PC,  IMM_U, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
 
-    LB->        List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,    MT_B, N,N,N,N,N,Y,CSR.N,N,N,N),
-    LH->        List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,    MT_H, N,N,N,N,N,Y,CSR.N,N,N,N),
-    LW->        List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,    MT_W, N,N,N,N,N,Y,CSR.N,N,N,N),
-    LBU->       List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,    MT_BU,N,N,N,N,N,Y,CSR.N,N,N,N),
-    LHU->       List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,    MT_HU,N,N,N,N,N,Y,CSR.N,N,N,N),
-    SB->        List(Y,N,N,N,N,N,Y,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,    MT_B, N,N,N,N,N,N,CSR.N,N,N,N),
-    SH->        List(Y,N,N,N,N,N,Y,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,    MT_H, N,N,N,N,N,N,CSR.N,N,N,N),
-    SW->        List(Y,N,N,N,N,N,Y,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,    MT_W, N,N,N,N,N,N,CSR.N,N,N,N),
+    LB->        List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,      MT_B, N,N,N,N,N,Y,CSR.N,N,N,N),
+    LH->        List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,      MT_H, N,N,N,N,N,Y,CSR.N,N,N,N),
+    LW->        List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,      MT_W, N,N,N,N,N,Y,CSR.N,N,N,N),
+    LBU->       List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,      MT_BU,N,N,N,N,N,Y,CSR.N,N,N,N),
+    LHU->       List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,      MT_HU,N,N,N,N,N,Y,CSR.N,N,N,N),
+    SB->        List(Y,N,N,N,N,N,Y,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,      MT_B, N,N,N,N,N,N,CSR.N,N,N,N),
+    SH->        List(Y,N,N,N,N,N,Y,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,      MT_H, N,N,N,N,N,N,CSR.N,N,N,N),
+    SW->        List(Y,N,N,N,N,N,Y,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,      MT_W, N,N,N,N,N,N,CSR.N,N,N,N),
 
-    LUI->       List(Y,N,N,N,N,N,N,N,A2_IMM, A1_ZERO,IMM_U, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    ADDI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SLTI ->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SLT,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SLTIU->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SLTU,  N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    ANDI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_AND,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    ORI->       List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_OR,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    XORI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_XOR,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SLLI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SL,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SRLI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SR,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SRAI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SRA,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    ADD->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SUB->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SUB,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SLT->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SLT,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SLTU->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SLTU,  N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    AND->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_AND,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    OR->        List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_OR,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    XOR->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_XOR,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SLL->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SL,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SRL->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SR,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SRA->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SRA,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    LUI->       List(Y,N,N,N,N,N,N,N,A2_IMM, A1_ZERO,IMM_U, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    ADDI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SLTI ->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SLT,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SLTIU->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SLTU,  N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    ANDI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_AND,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    ORI->       List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_OR,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    XORI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_XOR,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SLLI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SL,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SRLI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SR,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SRAI->      List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_SRA,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    ADD->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SUB->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SUB,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SLT->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SLT,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SLTU->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SLTU,  N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    AND->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_AND,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    OR->        List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_OR,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    XOR->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_XOR,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SLL->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SL,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SRL->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SR,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SRA->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_SRA,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
 
-    FENCE->     List(Y,N,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,Y,N),
-    FENCE_I->   List(Y,N,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,Y,N,N),
+    FENCE->     List(Y,N,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,Y,N),
+    FENCE_I->   List(Y,N,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     Y,M_FLUSH_ALL,MT_X, N,N,N,N,N,N,CSR.N,Y,N,N),
 
-    SFENCE_VM-> List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
-    SCALL->     List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
-    SBREAK->    List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
-    SRET->      List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
-    MRET->      List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
-    WFI->       List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
-    CSRRW->     List(Y,N,N,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.W,N,N,N),
-    CSRRS->     List(Y,N,N,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.S,N,N,N),
-    CSRRC->     List(Y,N,N,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.C,N,N,N),
-    CSRRWI->    List(Y,N,N,N,N,N,N,N,A2_IMM, A1_ZERO,IMM_Z, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.W,N,N,N),
-    CSRRSI->    List(Y,N,N,N,N,N,N,N,A2_IMM, A1_ZERO,IMM_Z, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.S,N,N,N),
-    CSRRCI->    List(Y,N,N,N,N,N,N,N,A2_IMM, A1_ZERO,IMM_Z, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.C,N,N,N))
+    SFENCE_VM-> List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
+    SCALL->     List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
+    SBREAK->    List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
+    SRET->      List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
+    MRET->      List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
+    WFI->       List(Y,N,N,N,N,N,N,X,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,N,N,N,CSR.I,N,N,N),
+    CSRRW->     List(Y,N,N,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.W,N,N,N),
+    CSRRS->     List(Y,N,N,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.S,N,N,N),
+    CSRRC->     List(Y,N,N,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.C,N,N,N),
+    CSRRWI->    List(Y,N,N,N,N,N,N,N,A2_IMM, A1_ZERO,IMM_Z, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.W,N,N,N),
+    CSRRSI->    List(Y,N,N,N,N,N,N,N,A2_IMM, A1_ZERO,IMM_Z, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.S,N,N,N),
+    CSRRCI->    List(Y,N,N,N,N,N,N,N,A2_IMM, A1_ZERO,IMM_Z, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.C,N,N,N))
 }
 
 class I64Decode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    LD->        List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,    MT_D, N,N,N,N,N,Y,CSR.N,N,N,N),
-    LWU->       List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,    MT_WU,N,N,N,N,N,Y,CSR.N,N,N,N),
-    SD->        List(Y,N,N,N,N,N,Y,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,    MT_D, N,N,N,N,N,N,CSR.N,N,N,N),
+    LD->        List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,      MT_D, N,N,N,N,N,Y,CSR.N,N,N,N),
+    LWU->       List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,      MT_WU,N,N,N,N,N,Y,CSR.N,N,N,N),
+    SD->        List(Y,N,N,N,N,N,Y,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,      MT_D, N,N,N,N,N,N,CSR.N,N,N,N),
 
-    ADDIW->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_32,FN_ADD,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SLLIW->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_32,FN_SL,     N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SRLIW->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_32,FN_SR,     N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SRAIW->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_32,FN_SRA,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    ADDW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_ADD,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SUBW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_SUB,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SLLW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_SL,     N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SRLW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_SR,     N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    SRAW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_SRA,    N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N))
+    ADDIW->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_32,FN_ADD,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SLLIW->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_32,FN_SL,     N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SRLIW->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_32,FN_SR,     N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SRAIW->     List(Y,N,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_32,FN_SRA,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    ADDW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_ADD,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SUBW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_SUB,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SLLW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_SL,     N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SRLW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_SR,     N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    SRAW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32,FN_SRA,    N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N))
 }
 
 class MDecode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    MUL->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_MUL,   N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
-    MULH->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_MULH,  N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
-    MULHU->     List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_MULHU, N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
-    MULHSU->    List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_MULHSU,N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    MUL->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_MUL,   N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    MULH->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_MULH,  N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    MULHU->     List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_MULHU, N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    MULHSU->    List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_MULHSU,N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
 
-    DIV->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_DIV,   N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
-    DIVU->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_DIVU,  N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
-    REM->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_REM,   N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
-    REMU->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_REMU,  N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N))
+    DIV->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_DIV,   N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    DIVU->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_DIVU,  N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    REM->       List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_REM,   N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    REMU->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_XPR,FN_REMU,  N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N))
 }
 
 class M64Decode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    MULW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_MUL,   N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    MULW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_MUL,   N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
 
-    DIVW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_DIV,   N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
-    DIVUW->     List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_DIVU,  N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
-    REMW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_REM,   N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
-    REMUW->     List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_REMU,  N,M_X,      MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N))
+    DIVW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_DIV,   N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    DIVUW->     List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_DIVU,  N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    REMW->      List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_REM,   N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N),
+    REMUW->     List(Y,N,N,N,N,N,Y,Y,A2_RS2, A1_RS1, IMM_X, DW_32, FN_REMU,  N,M_X,        MT_X, N,N,N,N,Y,Y,CSR.N,N,N,N))
 }
 
 class ADecode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    AMOADD_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_ADD, MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOXOR_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_XOR, MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOSWAP_W-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_SWAP,MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOAND_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_AND, MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOOR_W->   List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_OR,  MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOMIN_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MIN, MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOMINU_W-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MINU,MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOMAX_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MAX, MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOMAXU_W-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MAXU,MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOADD_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_ADD,   MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOXOR_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_XOR,   MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOSWAP_W-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_SWAP,  MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOAND_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_AND,   MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOOR_W->   List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_OR,    MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOMIN_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MIN,   MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOMINU_W-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MINU,  MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOMAX_W->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MAX,   MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOMAXU_W-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MAXU,  MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
 
-    LR_W->      List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XLR,    MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    SC_W->      List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XSC,    MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y))
+    LR_W->      List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XLR,      MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    SC_W->      List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XSC,      MT_W, N,N,N,N,N,Y,CSR.N,N,N,Y))
 }
 
 class A64Decode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    AMOADD_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_ADD, MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOSWAP_D-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_SWAP,MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOXOR_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_XOR, MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOAND_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_AND, MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOOR_D->   List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_OR,  MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOMIN_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MIN, MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOMINU_D-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MINU,MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOMAX_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MAX, MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    AMOMAXU_D-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MAXU,MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOADD_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_ADD,   MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOSWAP_D-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_SWAP,  MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOXOR_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_XOR,   MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOAND_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_AND,   MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOOR_D->   List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_OR,    MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOMIN_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MIN,   MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOMINU_D-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MINU,  MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOMAX_D->  List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MAX,   MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    AMOMAXU_D-> List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XA_MAXU,  MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
 
-    LR_D->      List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XLR,    MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
-    SC_D->      List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XSC,    MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y))
+    LR_D->      List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XLR,      MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y),
+    SC_D->      List(Y,N,N,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   Y,M_XSC,      MT_D, N,N,N,N,N,Y,CSR.N,N,N,Y))
 }
 
 class FDecode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    FCVT_S_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,Y,N,N,CSR.N,N,N,N),
-    FCVT_D_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,Y,N,N,CSR.N,N,N,N),
-    FSGNJ_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FSGNJ_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FSGNJX_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FSGNJX_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FSGNJN_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FSGNJN_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FMIN_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FMIN_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FMAX_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FMAX_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FADD_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FADD_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FSUB_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FSUB_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FMUL_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FMUL_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FMADD_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
-    FMADD_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
-    FMSUB_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
-    FMSUB_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
-    FNMADD_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
-    FNMADD_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
-    FNMSUB_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
-    FNMSUB_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
-    FCLASS_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FCLASS_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FMV_X_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FCVT_W_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FCVT_W_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FCVT_WU_S-> List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FCVT_WU_D-> List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FEQ_S->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
-    FEQ_D->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
-    FLT_S->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
-    FLT_D->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
-    FLE_S->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
-    FLE_D->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
-    FMV_S_X->   List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FCVT_S_W->  List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FCVT_D_W->  List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FCVT_S_WU-> List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FCVT_D_WU-> List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FLW->       List(Y,Y,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,    MT_W, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FLD->       List(Y,Y,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,    MT_D, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FSW->       List(Y,Y,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,    MT_W, N,Y,N,N,N,N,CSR.N,N,N,N),
-    FSD->       List(Y,Y,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,    MT_D, N,Y,N,N,N,N,CSR.N,N,N,N))
+    FCVT_S_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,Y,N,N,CSR.N,N,N,N),
+    FCVT_D_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,Y,N,N,CSR.N,N,N,N),
+    FSGNJ_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FSGNJ_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FSGNJX_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FSGNJX_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FSGNJN_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FSGNJN_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FMIN_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FMIN_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FMAX_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FMAX_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FADD_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FADD_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FSUB_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FSUB_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FMUL_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FMUL_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FMADD_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
+    FMADD_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
+    FMSUB_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
+    FMSUB_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
+    FNMADD_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
+    FNMADD_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
+    FNMSUB_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
+    FNMSUB_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,Y,Y,N,N,CSR.N,N,N,N),
+    FCLASS_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FCLASS_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FMV_X_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FCVT_W_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FCVT_W_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FCVT_WU_S-> List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FCVT_WU_D-> List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FEQ_S->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
+    FEQ_D->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
+    FLT_S->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
+    FLT_D->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
+    FLE_S->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
+    FLE_D->     List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,N,N,Y,CSR.N,N,N,N),
+    FMV_S_X->   List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FCVT_S_W->  List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FCVT_D_W->  List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FCVT_S_WU-> List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FCVT_D_WU-> List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FLW->       List(Y,Y,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,      MT_W, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FLD->       List(Y,Y,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_I, DW_XPR,FN_ADD,   Y,M_XRD,      MT_D, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FSW->       List(Y,Y,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,      MT_W, N,Y,N,N,N,N,CSR.N,N,N,N),
+    FSD->       List(Y,Y,N,N,N,N,N,Y,A2_IMM, A1_RS1, IMM_S, DW_XPR,FN_ADD,   Y,M_XWR,      MT_D, N,Y,N,N,N,N,CSR.N,N,N,N))
 }
 
 class F64Decode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    FMV_X_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FCVT_L_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FCVT_L_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FCVT_LU_S-> List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FCVT_LU_D-> List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
-    FMV_D_X->   List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FCVT_S_L->  List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FCVT_D_L->  List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FCVT_S_LU-> List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
-    FCVT_D_LU-> List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, N,N,N,Y,N,N,CSR.N,N,N,N))
+    FMV_X_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FCVT_L_S->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FCVT_L_D->  List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FCVT_LU_S-> List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FCVT_LU_D-> List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,N,N,N,N,Y,CSR.N,N,N,N),
+    FMV_D_X->   List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FCVT_S_L->  List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FCVT_D_L->  List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FCVT_S_LU-> List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N),
+    FCVT_D_LU-> List(Y,Y,N,N,N,N,N,Y,A2_X,   A1_RS1, IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, N,N,N,Y,N,N,CSR.N,N,N,N))
 }
 
 class FDivSqrtDecode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    FDIV_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FDIV_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FSQRT_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
-    FSQRT_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,      MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N))
+    FDIV_S->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FDIV_D->    List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FSQRT_S->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N),
+    FSQRT_D->   List(Y,Y,N,N,N,N,N,N,A2_X,   A1_X,   IMM_X, DW_X,  FN_X,     N,M_X,        MT_X, Y,Y,N,Y,N,N,CSR.N,N,N,N))
 }
 
 class RoCCDecode(implicit val p: Parameters) extends DecodeConstants
 {
   val table: Array[(BitPat, List[BitPat])] = Array(
-    CUSTOM0->           List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM0_RS1->       List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM0_RS1_RS2->   List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM0_RD->        List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM0_RD_RS1->    List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM0_RD_RS1_RS2->List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM1->           List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM1_RS1->       List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM1_RS1_RS2->   List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM1_RD->        List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM1_RD_RS1->    List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM1_RD_RS1_RS2->List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM2->           List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM2_RS1->       List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM2_RS1_RS2->   List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM2_RD->        List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM2_RD_RS1->    List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM2_RD_RS1_RS2->List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM3->           List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM3_RS1->       List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM3_RS1_RS2->   List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
-    CUSTOM3_RD->        List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM3_RD_RS1->    List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
-    CUSTOM3_RD_RS1_RS2->List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,      MT_X, N,N,N,N,N,Y,CSR.N,N,N,N))
+    CUSTOM0->           List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM0_RS1->       List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM0_RS1_RS2->   List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM0_RD->        List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM0_RD_RS1->    List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM0_RD_RS1_RS2->List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM1->           List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM1_RS1->       List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM1_RS1_RS2->   List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM1_RD->        List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM1_RD_RS1->    List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM1_RD_RS1_RS2->List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM2->           List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM2_RS1->       List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM2_RS1_RS2->   List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM2_RD->        List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM2_RD_RS1->    List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM2_RD_RS1_RS2->List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM3->           List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM3_RS1->       List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM3_RS1_RS2->   List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,N,CSR.N,N,N,N),
+    CUSTOM3_RD->        List(Y,N,Y,N,N,N,N,N,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM3_RD_RS1->    List(Y,N,Y,N,N,N,N,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N),
+    CUSTOM3_RD_RS1_RS2->List(Y,N,Y,N,N,N,Y,Y,A2_ZERO,A1_RS1, IMM_X, DW_XPR,FN_ADD,   N,M_X,        MT_X, N,N,N,N,N,Y,CSR.N,N,N,N))
 }
diff --git a/rocket/src/main/scala/nbdcache.scala b/rocket/src/main/scala/nbdcache.scala
index c776596c..87a8179a 100644
--- a/rocket/src/main/scala/nbdcache.scala
+++ b/rocket/src/main/scala/nbdcache.scala
@@ -410,7 +410,7 @@ class MSHR(id: Int)(implicit p: Parameters) extends L1HellaCacheModule()(p) {
 
   when (!io.meta_read.ready) {
     rpq.io.deq.ready := Bool(false)
-    io.replay.bits.cmd := M_NOP
+    io.replay.bits.cmd := M_FLUSH_ALL /* nop */
   }
 }
 
@@ -800,7 +800,7 @@ class HellaCache(implicit p: Parameters) extends L1HellaCacheModule()(p) {
 
   val s2_valid = Reg(next=s1_valid_masked, init=Bool(false))
   val s2_req = Reg(io.cpu.req.bits)
-  val s2_replay = Reg(next=s1_replay, init=Bool(false)) && s2_req.cmd =/= M_NOP
+  val s2_replay = Reg(next=s1_replay, init=Bool(false)) && s2_req.cmd =/= M_FLUSH_ALL
   val s2_recycle = Wire(Bool())
   val s2_valid_masked = Wire(Bool())
 
diff --git a/rocket/src/main/scala/rocket.scala b/rocket/src/main/scala/rocket.scala
index 823d0c09..e365b217 100644
--- a/rocket/src/main/scala/rocket.scala
+++ b/rocket/src/main/scala/rocket.scala
@@ -509,7 +509,7 @@ class Rocket(implicit p: Parameters) extends CoreModule()(p) {
     Mux(wb_xcpt || csr.io.eret, csr.io.evec,     // exception or [m|s]ret
     Mux(replay_wb,              wb_reg_pc,       // replay
                                 mem_npc)).toUInt // mispredicted branch
-  io.imem.flush_icache := wb_reg_valid && wb_ctrl.fence_i
+  io.imem.flush_icache := wb_reg_valid && wb_ctrl.fence_i && !io.dmem.s2_nack
   io.imem.flush_tlb := csr.io.fatc
   io.imem.resp.ready := !ctrl_stalld || csr.io.interrupt || take_pc_mem