support memory transaction aborts

rocket/src/main/scala/arbiter.scala

@@ -4,62 +4,74 @@ import Chisel._;
 import Node._;
 import Constants._;
 
+class ioUncachedRequestor extends Bundle {
+  val xact_init      = (new ioDecoupled) { new TransactionInit() }
+  val xact_abort     = (new ioDecoupled) { new TransactionAbort() }.flip
+  val xact_rep       = (new ioPipe)      { new TransactionReply() }.flip
+  val xact_finish    = (new ioDecoupled) { new TransactionFinish() }
+}
+
 class rocketMemArbiter(n: Int) extends Component {
   val io = new Bundle {
-    val mem = new ioTileLink
+    val mem = new ioUncachedRequestor
-    val requestor = Vec(n) { new ioTileLink().flip }
+    val requestor = Vec(n) { new ioUncachedRequestor().flip }
   }
 
-  var req_val = Bool(false)
+  var xi_val = Bool(false)
-  var req_rdy = io.mem.xact_init.ready
+  var xi_rdy = io.mem.xact_init.ready
   for (i <- 0 until n)
   {
-    io.requestor(i).xact_init.ready := req_rdy
+    io.requestor(i).xact_init.ready := xi_rdy
-    req_val = req_val || io.requestor(i).xact_init.valid
+    xi_val = xi_val || io.requestor(i).xact_init.valid
-    req_rdy = req_rdy && !io.requestor(i).xact_init.valid
+    xi_rdy = xi_rdy && !io.requestor(i).xact_init.valid
   }
 
-  // if more than one requestor at a time can write back, the data
+  var xi_bits = Wire() { new TransactionInit }
-  // arbiter needs to be made stateful: one xact's write data must
+  xi_bits := io.requestor(n-1).xact_init.bits
-  // be sent to the memory system contiguously.
+  xi_bits.tile_xact_id := Cat(io.requestor(n-1).xact_init.bits.tile_xact_id, UFix(n-1, log2up(n)))
-  var req_data_val = Bool(false)
+  for (i <- n-2 to 0 by -1)
-  var req_data_rdy = io.mem.xact_init_data.ready
+  {
+    var my_xi_bits = Wire() { new TransactionInit }
+    my_xi_bits := io.requestor(i).xact_init.bits
+    my_xi_bits.tile_xact_id := Cat(io.requestor(i).xact_init.bits.tile_xact_id, UFix(i, log2up(n)))
+
+    xi_bits = Mux(io.requestor(i).xact_init.valid, my_xi_bits, xi_bits)
+  }
+
+  io.mem.xact_init.valid := xi_val
+  io.mem.xact_init.bits := xi_bits
+
+  var xf_val = Bool(false)
+  var xf_rdy = io.mem.xact_finish.ready
   for (i <- 0 until n)
   {
-    io.requestor(i).xact_init_data.ready := req_data_rdy
+    io.requestor(i).xact_finish.ready := xf_rdy
-    req_data_val = req_data_val || io.requestor(i).xact_init_data.valid
+    xf_val = xf_val || io.requestor(i).xact_finish.valid
-    req_data_rdy = req_data_rdy && !io.requestor(i).xact_init_data.valid
+    xf_rdy = xf_rdy && !io.requestor(i).xact_finish.valid
   }
 
-  var req_bits = Wire() { new TransactionInit }
+  var xf_bits = io.requestor(n-1).xact_finish.bits
-  req_bits := io.requestor(n-1).xact_init.bits
-  req_bits.tile_xact_id := Cat(io.requestor(n-1).xact_init.bits.tile_xact_id, UFix(n-1, log2up(n)))
   for (i <- n-2 to 0 by -1)
-  {
+    xf_bits = Mux(io.requestor(i).xact_finish.valid, io.requestor(i).xact_finish.bits, xf_bits)
-    var my_req_bits = Wire() { new TransactionInit }
-    my_req_bits := io.requestor(i).xact_init.bits
-    my_req_bits.tile_xact_id := Cat(io.requestor(i).xact_init.bits.tile_xact_id, UFix(i, log2up(n)))
 
-    req_bits = Mux(io.requestor(i).xact_init.valid, my_req_bits, req_bits)
+  io.mem.xact_finish.valid := xf_val
-  }
+  io.mem.xact_finish.bits := xf_bits
-
-  var req_data_bits = io.requestor(n-1).xact_init_data.bits
-  for (i <- n-2 to 0 by -1)
-    req_data_bits = Mux(io.requestor(i).xact_init_data.valid, io.requestor(i).xact_init_data.bits, req_data_bits)
-
-  io.mem.xact_init.valid := req_val
-  io.mem.xact_init.bits := req_bits
-
-  io.mem.xact_init_data.valid := req_data_val
-  io.mem.xact_init_data.bits := req_data_bits
 
   for (i <- 0 until n)
   {
     val tag = io.mem.xact_rep.bits.tile_xact_id
     io.requestor(i).xact_rep.valid := io.mem.xact_rep.valid && tag(log2up(n)-1,0) === UFix(i)
-    io.requestor(i).xact_rep.bits.data := io.mem.xact_rep.bits.data
+    io.requestor(i).xact_rep.bits := io.mem.xact_rep.bits
-    io.requestor(i).xact_rep.bits.t_type := io.mem.xact_rep.bits.t_type
     io.requestor(i).xact_rep.bits.tile_xact_id := tag >> UFix(log2up(n))
-    io.requestor(i).xact_rep.bits.global_xact_id := io.mem.xact_rep.bits.global_xact_id
   }
+
+  for (i <- 0 until n)
+  {
+    val tag = io.mem.xact_abort.bits.tile_xact_id
+    io.requestor(i).xact_abort.valid := io.mem.xact_abort.valid && tag(log2up(n)-1,0) === UFix(i)
+    io.requestor(i).xact_abort.bits := io.mem.xact_abort.bits
+    io.requestor(i).xact_abort.bits.tile_xact_id := tag >> UFix(log2up(n))
+  }
+
+  io.mem.xact_abort.ready := Bool(true)
 }

rocket/src/main/scala/coherence.scala

@@ -395,6 +395,8 @@ class CoherenceHubNull extends CoherenceHub {
   x_rep.bits.global_xact_id := UFix(0) // don't care
   x_rep.bits.data := io.mem.resp.bits.data
   x_rep.valid := io.mem.resp.valid || x_init.valid && is_write
+
+  io.tiles(0).xact_abort.valid := Bool(false)
 }
 
 

rocket/src/main/scala/consts.scala

@@ -186,7 +186,7 @@ object Constants
   val NTILES = 1
   val COHERENCE_DATA_BITS = (1 << OFFSET_BITS)*8 
   val TILE_ID_BITS = 1
-  val TILE_XACT_ID_BITS = log2up(NMSHR)+2
+  val TILE_XACT_ID_BITS = log2up(NMSHR)+3
   val GLOBAL_XACT_ID_BITS = 4
   val NGLOBAL_XACTS = 1 << GLOBAL_XACT_ID_BITS
 

rocket/src/main/scala/htif.scala

@@ -78,7 +78,12 @@ class rocketHTIF(w: Int, ncores: Int) extends Component
                 Mux(!nack && cmd === cmd_readcr, UFix(1), UFix(0)))
   val tx_done = packet_ram_raddr - UFix(1) === tx_size
 
-  val state_rx :: state_pcr :: state_mem_req :: state_mem_wdata :: state_mem_rdata :: state_tx :: Nil = Enum(6) { UFix() }
+  val mem_acked = Reg(resetVal = Bool(false))
+  val mem_nacked = Reg(resetVal = Bool(false))
+  when (io.mem.xact_rep.valid) { mem_acked := Bool(true) }
+  when (io.mem.xact_abort.valid) { mem_nacked := Bool(true) }
+
+  val state_rx :: state_pcr :: state_mem_req :: state_mem_wdata :: state_mem_wresp :: state_mem_rdata :: state_tx :: Nil = Enum(7) { UFix() }
   val state = Reg(resetVal = state_rx)
 
   when (state === state_rx && rx_done) {
@@ -96,13 +101,35 @@ class rocketHTIF(w: Int, ncores: Int) extends Component
   when (state === state_mem_req && io.mem.xact_init.ready) {
     state := Mux(cmd === cmd_writemem, state_mem_wdata, state_mem_rdata)
   }
-  when (state === state_mem_wdata && io.mem.xact_init_data.ready || 
+  when (state === state_mem_wdata && io.mem.xact_init_data.ready) {
-        state === state_mem_rdata && io.mem.xact_rep.valid) {
+    when (mem_cnt.andR)  {
+      state := state_mem_wresp
+    }
+    mem_cnt := mem_cnt + UFix(1)
+  }
+  when (state === state_mem_wresp) {
+    when (mem_nacked) {
+      state := state_mem_req
+      mem_nacked := Bool(false)
+    }
+    when (mem_acked) {
+      state := state_tx
+      mem_acked := Bool(false)
+    }
+  }
+  when (state === state_mem_rdata) {
+    when (mem_nacked) {
+      state := state_mem_req
+      mem_nacked := Bool(false)
+    }
+    when (io.mem.xact_rep.valid) {
       when (mem_cnt.andR)  {
         state := state_tx
       }
       mem_cnt := mem_cnt + UFix(1)
     }
+    mem_acked := Bool(false)
+  }
   when (state === state_tx && tx_done) {
     rx_count := UFix(0)
     tx_count := UFix(0)

rocket/src/main/scala/icache.scala

@@ -20,7 +20,7 @@ class ioImem(view: List[String] = null) extends Bundle (view)
 class ioRocketICache extends Bundle()
 {
   val cpu = new ioImem();
-  val mem = new ioTileLink
+  val mem = new ioUncachedRequestor
 }
 
 // basic direct mapped instruction cache
@@ -49,6 +49,7 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
   
   val s_reset :: s_ready :: s_request :: s_refill_wait :: s_refill :: Nil = Enum(5) { UFix() };
   val state = Reg(resetVal = s_reset);
+  val invalidated = Reg() { Bool() }
   
   val r_cpu_req_idx    = Reg { Bits() }
   val r_cpu_req_ppn    = Reg { Bits() }
@@ -78,13 +79,14 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
   when (io.mem.xact_rep.valid) {
     refill_count := refill_count + UFix(1);
   }
+  val refill_done = io.mem.xact_rep.valid && refill_count.andR
 
   val repl_way = LFSR16(state === s_ready && r_cpu_req_val && !io.cpu.itlb_miss && !tag_hit)(log2up(assoc)-1,0)
   val word_shift = Cat(r_cpu_req_idx(offsetmsb-rf_cnt_bits,offsetlsb), UFix(0, log2up(databits))).toUFix
+  val tag_we = (state === s_refill) && refill_done
   val tag_addr = 
-    Mux((state === s_refill_wait), r_cpu_req_idx(indexmsb,indexlsb),
+    Mux((state === s_refill), r_cpu_req_idx(indexmsb,indexlsb),
       io.cpu.req_idx(indexmsb,indexlsb)).toUFix;
-  val tag_we = (state === s_refill_wait) && io.mem.xact_rep.valid;
   val data_addr = 
     Mux((state === s_refill_wait) || (state === s_refill),  Cat(r_cpu_req_idx(indexmsb,offsetbits), refill_count),
       io.cpu.req_idx(indexmsb, offsetbits-rf_cnt_bits)).toUFix;
@@ -102,10 +104,10 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
     // valid bit array
     val vb_array = Reg(resetVal = Bits(0, sets));
     when (io.cpu.invalidate) {
-      vb_array := Bits(0,sets);
+      vb_array := Bits(0)
     }
     .elsewhen (tag_we && repl_me) {
-      vb_array := vb_array.bitSet(r_cpu_req_idx(indexmsb,indexlsb).toUFix, UFix(1,1));
+      vb_array := vb_array.bitSet(r_cpu_req_idx(indexmsb,indexlsb).toUFix, !invalidated)
     }
 
     val valid = vb_array(r_cpu_req_idx(indexmsb,indexlsb)).toBool;
@@ -131,20 +133,20 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
   io.mem.xact_init.valid := (state === s_request)
   io.mem.xact_init.bits.t_type := X_INIT_READ_UNCACHED
   io.mem.xact_init.bits.address := r_cpu_miss_addr(tagmsb,indexlsb).toUFix
-  io.mem.xact_init_data.valid := Bool(false)
 
   // control state machine
+  when (io.cpu.invalidate) {
+    invalidated := Bool(true)
+  }
   switch (state) {
     is (s_reset) {
       state := s_ready;
     }
     is (s_ready) {
-      when (io.cpu.itlb_miss) {
+      when (r_cpu_req_val && !tag_hit && !io.cpu.itlb_miss) {
-        state := s_ready;
-      }
-      .elsewhen (r_cpu_req_val && !tag_hit) {
         state := s_request;
       }
+      invalidated := Bool(false)
     }
     is (s_request)
     {
@@ -153,12 +155,15 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
       }
     }
     is (s_refill_wait) {
+      when (io.mem.xact_abort.valid) {
+        state := s_request
+      }
       when (io.mem.xact_rep.valid) {
         state := s_refill;
       }
     }
     is (s_refill) {
-      when (io.mem.xact_rep.valid && refill_count.andR) {
+      when (refill_done) {
         state := s_ready;
       }
     }

rocket/src/main/scala/icache_prefetch.scala

@@ -18,22 +18,23 @@ class rocketIPrefetcher extends Component() {
   val s_invalid :: s_valid :: s_refilling :: s_req_wait :: s_resp_wait :: s_bad_resp_wait :: Nil = Enum(6) { UFix() };
   val state = Reg(resetVal = s_invalid);
 
+  val ip_mem_resp_abort = io.mem.xact_abort.valid && io.mem.xact_abort.bits.tile_xact_id(0)
   val demand_miss = io.icache.xact_init.valid && io.icache.xact_init.ready
   val prefetch_addr = Reg() { UFix(width = io.icache.xact_init.bits.address.width) };
   val addr_match = (prefetch_addr === io.icache.xact_init.bits.address);
-  val hit = (state != s_invalid) & (state != s_req_wait) & addr_match;
+  val hit = (state != s_invalid) && (state != s_req_wait) && addr_match && !ip_mem_resp_abort
   val prefetch_miss = io.icache.xact_init.valid && !hit
   when (demand_miss) { prefetch_addr := io.icache.xact_init.bits.address + UFix(1); }
 
   io.icache.xact_init.ready := io.mem.xact_init.ready
-  val ip_mem_req_rdy  = io.mem.xact_init.ready && !prefetch_miss
   val ip_mem_resp_val = io.mem.xact_rep.valid && io.mem.xact_rep.bits.tile_xact_id(0)
+  val ip_mem_req_rdy  = io.mem.xact_init.ready && !prefetch_miss
  
+  io.mem.xact_abort.ready := Bool(true)
   io.mem.xact_init.valid  := prefetch_miss || (state === s_req_wait)
   io.mem.xact_init.bits.t_type := X_INIT_READ_UNCACHED
   io.mem.xact_init.bits.tile_xact_id  := Mux(prefetch_miss, UFix(0), UFix(1))
   io.mem.xact_init.bits.address := Mux(prefetch_miss, io.icache.xact_init.bits.address, prefetch_addr);
-  io.mem.xact_init_data.valid := Bool(false)
   
   val fill_cnt = Reg(resetVal = UFix(0, ceil(log(REFILL_CYCLES)/log(2)).toInt));
   when (ip_mem_resp_val.toBool) { fill_cnt := fill_cnt + UFix(1); }
@@ -43,8 +44,10 @@ class rocketIPrefetcher extends Component() {
   val forward_cnt = Reg(resetVal = UFix(0, ceil(log(REFILL_CYCLES)/log(2)).toInt));
   when (forward & pdq.io.deq.valid) { forward_cnt := forward_cnt + UFix(1); }
   val forward_done = (~forward_cnt === UFix(0)) & pdq.io.deq.valid;
-  forward := (demand_miss & hit | forward & ~forward_done);  
+  forward := demand_miss && hit || forward && !forward_done
 
+  io.icache.xact_abort.valid := io.mem.xact_abort.valid && !io.mem.xact_abort.bits.tile_xact_id(0) ||
+                                forward && ip_mem_resp_abort
   io.icache.xact_rep.valid  := io.mem.xact_rep.valid && !io.mem.xact_rep.bits.tile_xact_id(0) || (forward && pdq.io.deq.valid)
   io.icache.xact_rep.bits.data := Mux(forward, pdq.io.deq.bits, io.mem.xact_rep.bits.data)
   
@@ -69,11 +72,14 @@ class rocketIPrefetcher extends Component() {
       when (ip_mem_req_rdy) { state := s_resp_wait; }
     }
     is (s_resp_wait) {
-      when (demand_miss & ~addr_match) { state := s_bad_resp_wait; }
+      when (ip_mem_resp_abort) {
-      .elsewhen (ip_mem_resp_val.toBool) { state := s_refilling; }
+        state := s_invalid
+      }
+      .elsewhen (demand_miss && !addr_match) { state := s_bad_resp_wait }
+      .elsewhen (ip_mem_resp_val) { state := s_refilling }
     }
     is (s_bad_resp_wait) {
-      when (fill_done.toBool & ip_mem_resp_val.toBool) { state := s_req_wait; }
+      when (fill_done || ip_mem_resp_abort) { state := s_req_wait }
     }
   }
 

rocket/src/main/scala/nbdcache.scala

@@ -178,6 +178,7 @@ class MSHR(id: Int) extends Component with ThreeStateIncoherence {
     val way_oh         = Bits(NWAYS, OUTPUT)
 
     val mem_resp_val = Bool(INPUT)
+    val mem_abort_val = Bool(INPUT)
     val mem_req  = (new ioDecoupled) { new TransactionInit }
     val meta_req = (new ioDecoupled) { new MetaArrayArrayReq() }
     val replay   = (new ioDecoupled) { new Replay()   }
@@ -216,6 +217,9 @@ class MSHR(id: Int) extends Component with ThreeStateIncoherence {
     when (io.mem_req.valid && io.mem_req.ready) {
       requested := Bool(true)
     }
+    when (io.mem_abort_val) {
+      requested := Bool(false)
+    }
     when (io.mem_resp_val) {
       refilled := Bool(true)
     }
@@ -264,6 +268,7 @@ class MSHRFile extends Component {
     val mem_req  = (new ioDecoupled) { new TransactionInit }
     val meta_req = (new ioDecoupled) { new MetaArrayArrayReq() }
     val data_req   = (new ioDecoupled) { new DataReq() }
+    val mem_abort = (new ioPipe) { new TransactionAbort }.flip
 
     val cpu_resp_val = Bool(OUTPUT)
     val cpu_resp_tag = Bits(DCACHE_TAG_BITS, OUTPUT)
@@ -310,8 +315,8 @@ class MSHRFile extends Component {
     mshr.io.mem_req <> mem_req_arb.io.in(i)
     mshr.io.replay <> replay_arb.io.in(i)
 
-    val mem_resp_val = io.mem_resp_val && (UFix(i) === io.mem_resp_tag)
+    mshr.io.mem_resp_val := io.mem_resp_val && (UFix(i) === io.mem_resp_tag)
-    mshr.io.mem_resp_val := mem_resp_val
+    mshr.io.mem_abort_val := io.mem_abort.valid && (UFix(i) === io.mem_abort.bits.tile_xact_id)
     mem_resp_idx_mux.io.sel(i) := (UFix(i) === io.mem_resp_tag)
     mem_resp_idx_mux.io.in(i) := mshr.io.idx
     mem_resp_way_oh_mux.io.sel(i) := (UFix(i) === io.mem_resp_tag)
@@ -354,20 +359,53 @@ class WritebackUnit extends Component {
     val refill_req = (new ioDecoupled) { new TransactionInit }.flip
     val mem_req = (new ioDecoupled) { new TransactionInit }
     val mem_req_data = (new ioDecoupled) { new TransactionInitData }
+    val mem_abort = (new ioPipe) { new TransactionAbort }.flip
+    val mem_rep = (new ioPipe) { new TransactionReply }.flip
   }
 
   val valid = Reg(resetVal = Bool(false))
   val data_req_fired = Reg(resetVal = Bool(false))
+  val cmd_sent = Reg() { Bool() }
   val cnt = Reg() { UFix(width = log2up(REFILL_CYCLES+1)) }
   val addr = Reg() { new WritebackReq() }
 
-  data_req_fired := Bool(false)
+  val acked = Reg() { Bool() }
-  when (io.data_req.valid && io.data_req.ready) { data_req_fired := Bool(true); cnt := cnt + UFix(1) }
+  val nacked = Reg() { Bool() }
-  when (data_req_fired && !io.mem_req_data.ready) { data_req_fired := Bool(false); cnt := cnt - UFix(1) }
+  when (io.mem_rep.valid && io.mem_rep.bits.tile_xact_id === UFix(NMSHR)) { acked := Bool(true) }
-  when ((cnt === UFix(REFILL_CYCLES)) && io.mem_req_data.ready) { valid := Bool(false) }
+  when (io.mem_abort.valid && io.mem_abort.bits.tile_xact_id === UFix(NMSHR)) { nacked := Bool(true) }
-  when (io.req.valid && io.req.ready) { valid := Bool(true); cnt := UFix(0); addr := io.req.bits }
 
-  io.req.ready := !valid && io.mem_req.ready
+  data_req_fired := Bool(false)
+  when (valid && io.mem_req.ready) {
+    cmd_sent := Bool(true)
+  }
+  when (io.data_req.valid && io.data_req.ready) {
+    data_req_fired := Bool(true)
+    cnt := cnt + UFix(1)
+  }
+  when (data_req_fired && !io.mem_req_data.ready) {
+    data_req_fired := Bool(false)
+    cnt := cnt - UFix(1)
+  }
+  when ((cnt === UFix(REFILL_CYCLES)) && (!data_req_fired || io.mem_req_data.ready)) {
+    when (acked) {
+      valid := Bool(false)
+    }
+    when (nacked) {
+      cmd_sent := Bool(false)
+      nacked := Bool(false)
+      cnt := UFix(0)
+    }
+  }
+  when (io.req.valid && io.req.ready) {
+    valid := Bool(true)
+    acked := Bool(false)
+    nacked := Bool(false)
+    cmd_sent := Bool(false)
+    cnt := UFix(0)
+    addr := io.req.bits
+  }
+
+  io.req.ready := !valid
   io.data_req.valid := valid && (cnt < UFix(REFILL_CYCLES))
   io.data_req.bits.way_en := addr.way_oh
   io.data_req.bits.inner_req.idx := addr.idx
@@ -376,11 +414,12 @@ class WritebackUnit extends Component {
   io.data_req.bits.inner_req.wmask := Bits(0)
   io.data_req.bits.inner_req.data := Bits(0)
 
-  val wb_req_val = io.req.valid && !valid
+  val wb_req_val = valid && !cmd_sent
-  io.refill_req.ready := io.mem_req.ready && !wb_req_val
+  io.refill_req.ready := io.mem_req.ready && !(valid && !acked)
-  io.mem_req.valid := io.refill_req.valid || wb_req_val
+  io.mem_req.valid := io.refill_req.valid && !(valid && !acked) || wb_req_val
   io.mem_req.bits.t_type := Mux(wb_req_val, X_INIT_WRITE_UNCACHED, io.refill_req.bits.t_type)
-  io.mem_req.bits.address := Mux(wb_req_val, Cat(io.req.bits.ppn, io.req.bits.idx).toUFix, io.refill_req.bits.address)
+  io.mem_req.bits.has_data := wb_req_val
+  io.mem_req.bits.address := Mux(wb_req_val, Cat(addr.ppn, addr.idx).toUFix, io.refill_req.bits.address)
   io.mem_req.bits.tile_xact_id := Mux(wb_req_val, Bits(NMSHR), io.refill_req.bits.tile_xact_id)
 
   io.mem_req_data.valid := data_req_fired
@@ -676,8 +715,12 @@ class HellaCacheUniproc extends HellaCache with ThreeStateIncoherence {
   val cpu_req_data = Mux(r_replay_amo, r_amo_replay_data, io.cpu.req_data)
 
   // refill counter
+<<<<<<< HEAD
   val mem_resp_type = io.mem.xact_rep.bits.t_type
   val refill_val = io.mem.xact_rep.valid && (mem_resp_type === X_REP_READ_SHARED || mem_resp_type === X_REP_READ_EXCLUSIVE)
+=======
+  val refill_val = io.mem.xact_rep.valid && io.mem.xact_rep.bits.tile_xact_id < UFix(NMSHR)
+>>>>>>> support memory transaction aborts
   val rr_count = Reg(resetVal = UFix(0, log2up(REFILL_CYCLES)))
   val rr_count_next = rr_count + UFix(1)
   when (refill_val) { rr_count := rr_count_next }
@@ -725,6 +768,9 @@ class HellaCacheUniproc extends HellaCache with ThreeStateIncoherence {
   wb_arb.io.out <> wb.io.req
   wb.io.data_req <> data_arb.io.in(3)
   wb.io.data_resp <> data_resp_mux
+  wb.io.mem_rep <> io.mem.xact_rep
+  wb.io.mem_abort.valid := io.mem.xact_abort.valid
+  wb.io.mem_abort.bits := io.mem.xact_abort.bits
 
   // replacement policy
   val replacer = new RandomReplacementWayGen()
@@ -737,9 +783,11 @@ class HellaCacheUniproc extends HellaCache with ThreeStateIncoherence {
   // refill response
   val block_during_refill = !refill_val && (rr_count != UFix(0))
   data_arb.io.in(0).bits.inner_req.offset := rr_count
+  data_arb.io.in(0).bits.inner_req.idx := mshr.io.mem_resp_idx
   data_arb.io.in(0).bits.inner_req.rw := !block_during_refill
   data_arb.io.in(0).bits.inner_req.wmask := ~UFix(0, MEM_DATA_BITS/8)
   data_arb.io.in(0).bits.inner_req.data := io.mem.xact_rep.bits.data
+  data_arb.io.in(0).bits.way_en := mshr.io.mem_resp_way_oh
   data_arb.io.in(0).valid := refill_val || block_during_refill
 
   // load hits
@@ -815,10 +863,10 @@ class HellaCacheUniproc extends HellaCache with ThreeStateIncoherence {
 
   mshr.io.mem_resp_val := refill_val && (~rr_count === UFix(0))
   mshr.io.mem_resp_tag := io.mem.xact_rep.bits.tile_xact_id
+  mshr.io.mem_abort.valid := io.mem.xact_abort.valid
+  mshr.io.mem_abort.bits := io.mem.xact_abort.bits
   mshr.io.mem_req <> wb.io.refill_req
   mshr.io.meta_req <> meta_arb.io.in(1)
-  data_arb.io.in(0).bits.inner_req.idx := mshr.io.mem_resp_idx
-  data_arb.io.in(0).bits.way_en := mshr.io.mem_resp_way_oh
   replacer.io.pick_new_way := mshr.io.req.valid && mshr.io.req.ready
 
   // replays

rocket/src/main/scala/top.scala

@@ -29,13 +29,18 @@ class Top() extends Component {
   val hub = new CoherenceHubNull
   // connect tile to hub
   hub.io.tiles(0).xact_init <> Queue(arbiter.io.mem.xact_init)
-  hub.io.tiles(0).xact_init_data <> Queue(arbiter.io.mem.xact_init_data)
+  arbiter.io.mem.xact_abort <> Queue(hub.io.tiles(0).xact_abort)
   arbiter.io.mem.xact_rep <> Pipe(hub.io.tiles(0).xact_rep)
   // connect hub to memory
   io.mem.req_cmd <> Queue(hub.io.mem.req_cmd)
   io.mem.req_data <> Queue(hub.io.mem.req_data)
   hub.io.mem.resp <> Pipe(io.mem.resp)
 
+  // temporary HTIF data connection
+  val data_arb = (new Arbiter(2)) { new TransactionInitData }
+  data_arb.io.in(0) <> Queue(dcache.io.mem.xact_init_data)
+  data_arb.io.in(1) <> Queue(htif.io.mem.xact_init_data)
+  hub.io.tiles(0).xact_init_data <> data_arb.io.out
 
   if (HAVE_VEC)
   {