replace tile memory interface with ioTileLink

work in progress towards coherent HTIF. for now, requests
are incoherently passed through a null coherence hub.

rocket/src/main/scala/arbiter.scala

@@ -4,75 +4,64 @@ import Chisel._;
 import Node._;
 import Constants._;
 
-class ioMem() extends Bundle
-{
-  val req_val = Bool(OUTPUT);
-  val req_rdy = Bool(INPUT);
-  val req_rw  = Bool(OUTPUT);
-  val req_addr = UFix(PADDR_BITS - OFFSET_BITS, OUTPUT);
-  val req_tag = Bits(MEM_TAG_BITS, OUTPUT);
-
-  val req_data_val = Bool(OUTPUT);
-  val req_data_rdy = Bool(INPUT);
-  val req_data_bits = Bits(MEM_DATA_BITS, OUTPUT);
-  
-  val resp_val = Bool(INPUT);
-  val resp_tag = Bits(MEM_TAG_BITS, INPUT);
-  val resp_data = Bits(MEM_DATA_BITS, INPUT);
-}
-
-class ioMemArbiter(n: Int) extends Bundle() {
-  val mem = new ioMem();
-  val requestor = Vec(n) { new ioMem().flip() }
-}
-
 class rocketMemArbiter(n: Int) extends Component {
-  val io = new ioMemArbiter(n);
+  val io = new Bundle {
+    val mem = new ioTileLink
+    val requestor = Vec(n) { new ioTileLink().flip }
+  }
 
   var req_val = Bool(false)
-  var req_rdy = io.mem.req_rdy
+  var req_rdy = io.mem.xact_init.ready
   for (i <- 0 until n)
   {
-    io.requestor(i).req_rdy := req_rdy
+    io.requestor(i).xact_init.ready := req_rdy
-    req_val = req_val || io.requestor(i).req_val
+    req_val = req_val || io.requestor(i).xact_init.valid
-    req_rdy = req_rdy && !io.requestor(i).req_val
+    req_rdy = req_rdy && !io.requestor(i).xact_init.valid
   }
 
+  // if more than one requestor at a time can write back, the data
+  // arbiter needs to be made stateful: one xact's write data must
+  // be sent to the memory system contiguously.
   var req_data_val = Bool(false)
-  var req_data_rdy = io.mem.req_data_rdy
+  var req_data_rdy = io.mem.xact_init_data.ready
   for (i <- 0 until n)
   {
-    io.requestor(i).req_data_rdy := req_data_rdy
+    io.requestor(i).xact_init_data.ready := req_data_rdy
-    req_data_val = req_data_val || io.requestor(i).req_data_val
+    req_data_val = req_data_val || io.requestor(i).xact_init_data.valid
-    req_data_rdy = req_data_rdy && !io.requestor(i).req_data_val
+    req_data_rdy = req_data_rdy && !io.requestor(i).xact_init_data.valid
   }
 
-  var req_rw = io.requestor(n-1).req_rw
+  var req_bits = Wire() { new TransactionInit }
-  var req_addr = io.requestor(n-1).req_addr
+  req_bits := io.requestor(n-1).xact_init.bits
-  var req_tag = Cat(io.requestor(n-1).req_tag, UFix(n-1, log2up(n)))
+  req_bits.tile_xact_id := Cat(io.requestor(n-1).xact_init.bits.tile_xact_id, UFix(n-1, log2up(n)))
-  for (i <- n-1 to 0 by -1)
+  for (i <- n-2 to 0 by -1)
   {
-    req_rw = Mux(io.requestor(i).req_val, io.requestor(i).req_rw, req_rw)
+    var my_req_bits = Wire() { new TransactionInit }
-    req_addr = Mux(io.requestor(i).req_val, io.requestor(i).req_addr, req_addr)
+    my_req_bits := io.requestor(i).xact_init.bits
-    req_tag = Mux(io.requestor(i).req_val, Cat(io.requestor(i).req_tag, UFix(i, log2up(n))), req_tag)
+    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)
   }
 
-  var req_data_bits = io.requestor(n-1).req_data_bits
+  var req_data_bits = io.requestor(n-1).xact_init_data.bits
-  for (i <- n-1 to 0 by -1)
+  for (i <- n-2 to 0 by -1)
-    req_data_bits = Mux(io.requestor(i).req_data_val, io.requestor(i).req_data_bits, req_data_bits)
+    req_data_bits = Mux(io.requestor(i).xact_init_data.valid, io.requestor(i).xact_init_data.bits, req_data_bits)
 
-  io.mem.req_val := req_val
+  io.mem.xact_init.valid := req_val
-  io.mem.req_rw := req_rw
+  io.mem.xact_init.bits := req_bits
-  io.mem.req_addr := req_addr
-  io.mem.req_tag := req_tag
 
-  io.mem.req_data_val := req_data_val
+  io.mem.xact_init_data.valid := req_data_val
-  io.mem.req_data_bits := req_data_bits
+  io.mem.xact_init_data.bits := req_data_bits
 
   for (i <- 0 until n)
   {
-    io.requestor(i).resp_val := io.mem.resp_val && io.mem.resp_tag(log2up(n)-1,0) === UFix(i)
+    val tag = io.mem.xact_rep.bits.tile_xact_id
-    io.requestor(i).resp_data := io.mem.resp_data
+    io.requestor(i).xact_rep.valid := io.mem.xact_rep.valid && tag(log2up(n)-1,0) === UFix(i)
-    io.requestor(i).resp_tag := io.mem.resp_tag >> UFix(log2up(n))
+    io.requestor(i).xact_rep.bits.data := io.mem.xact_rep.bits.data
+    io.requestor(i).xact_rep.bits.t_type := io.mem.xact_rep.bits.t_type
+    io.requestor(i).xact_rep.bits.has_data := io.mem.xact_rep.bits.has_data
+    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
   }
+  io.mem.xact_rep.ready := Bool(true) // XXX we shouldn't have xact_rep.ready
 }

rocket/src/main/scala/coherence.scala

@@ -10,22 +10,21 @@ class MemData extends Bundle {
 class MemReqCmd() extends Bundle 
 {
   val rw = Bool()
-  val addr = UFix(PADDR_BITS - OFFSET_BITS)
+  val addr = UFix(width = PADDR_BITS - OFFSET_BITS)
-  val tag = Bits(MEM_TAG_BITS)
+  val tag = Bits(width = MEM_TAG_BITS)
 }
 
 class MemResp () extends Bundle 
 {
-  val tag = Bits(MEM_TAG_BITS)
+  val tag = Bits(width = MEM_TAG_BITS)
   val data = Bits(width = MEM_DATA_BITS)
-  val valid = Bool()
 }
 
-class ioMemHub() extends Bundle
+class ioMem() extends Bundle
 {
   val req_cmd  = (new ioDecoupled) { new MemReqCmd() }.flip
   val req_data = (new ioDecoupled) { new MemData() }.flip
-  val resp     = new MemResp() 
+  val resp     = (new ioValid) { new MemResp() }
 }
 
 class HubMemReq extends Bundle {
@@ -49,7 +48,7 @@ class TransactionInit extends Bundle {
   val t_type = Bits(width = TTYPE_BITS)
   val has_data = Bool()
   val tile_xact_id = Bits(width = TILE_XACT_ID_BITS)
-  val address = Bits(width = PADDR_BITS)
+  val address = UFix(width = PADDR_BITS)
 }
 
 class TransactionInitData extends MemData
@@ -348,8 +347,8 @@ abstract class CoherenceHub extends Component with CoherencePolicy
 
 class CoherenceHubNull extends Component {
   val io = new Bundle {
-    val tile = new ioTileLink() 
+    val tile = new ioTileLink().flip
-    val mem = new ioMemHub()
+    val mem = new ioMem
   }
 
   val x_init = io.tile.xact_init
@@ -362,11 +361,11 @@ class CoherenceHubNull extends Component {
   io.mem.req_data <> io.tile.xact_init_data
 
   val x_rep = io.tile.xact_rep
-  x_rep.bits.t_type := Mux(is_write, X_WRITE_UNCACHED, X_READ_EXCLUSIVE)
+  x_rep.bits.t_type := Mux(io.mem.resp.valid, X_READ_EXCLUSIVE, X_WRITE_UNCACHED)
-  x_rep.bits.tile_xact_id := Mux(is_write, x_init.bits.tile_xact_id, io.mem.resp.tag)
+  x_rep.bits.tile_xact_id := Mux(io.mem.resp.valid, io.mem.resp.bits.tag, x_init.bits.tile_xact_id)
   x_rep.bits.global_xact_id := UFix(0) // don't care
-  x_rep.bits.data := io.mem.resp.data
+  x_rep.bits.data := io.mem.resp.bits.data
-  x_rep.valid := io.mem.resp.valid || is_write
+  x_rep.valid := io.mem.resp.valid || x_init.valid && is_write
 }
 
 
@@ -388,7 +387,7 @@ class CoherenceHubNoDir extends CoherenceHub {
 
   val io = new Bundle {
     val tiles = Vec(NTILES) { new ioTileLink() }
-    val mem = new ioMemHub
+    val mem = new ioMem
   }
   
   val trackerList = (0 until NGLOBAL_XACTS).map(new XactTracker(_))
@@ -427,12 +426,12 @@ class CoherenceHubNoDir extends CoherenceHub {
 
   // Reply to initial requestor
   // Forward memory responses from mem to tile
-  val idx = io.mem.resp.tag
+  val idx = io.mem.resp.bits.tag
   for( j <- 0 until NTILES ) {
     io.tiles(j).xact_rep.bits.t_type := getTransactionReplyType(t_type_arr.read(idx), sh_count_arr.read(idx))
     io.tiles(j).xact_rep.bits.tile_xact_id := tile_xact_id_arr.read(idx)
     io.tiles(j).xact_rep.bits.global_xact_id := idx
-    io.tiles(j).xact_rep.bits.data := io.mem.resp.data
+    io.tiles(j).xact_rep.bits.data := io.mem.resp.bits.data
     io.tiles(j).xact_rep.valid := (UFix(j) === init_tile_id_arr.read(idx)) && (io.mem.resp.valid || send_x_rep_ack_arr.read(idx))
   }
   // If there were a ready signal due to e.g. intervening network use:

rocket/src/main/scala/consts.scala

@@ -182,7 +182,7 @@ object Constants
   val NTILES = 1
   val COHERENCE_DATA_BITS = (1 << OFFSET_BITS)*8 
   val TILE_ID_BITS = 1
-  val TILE_XACT_ID_BITS = 1 // log2(NMSHR)
+  val TILE_XACT_ID_BITS = log2up(NMSHR)+2
   val GLOBAL_XACT_ID_BITS = 4
   val NGLOBAL_XACTS = 1 << GLOBAL_XACT_ID_BITS
 
@@ -201,7 +201,7 @@ object Constants
   val MEM_TAG_BITS = 4
   val MEM_DATA_BITS = 128
   val REFILL_CYCLES = (1 << OFFSET_BITS)*8/MEM_DATA_BITS
-  require(MEM_TAG_BITS >= max(log2up(NMSHR)+1, GLOBAL_XACT_ID_BITS))
+  require(MEM_TAG_BITS >= max(TILE_XACT_ID_BITS, GLOBAL_XACT_ID_BITS))
   
   val DTLB_ENTRIES = 8;
   val ITLB_ENTRIES = 8;

rocket/src/main/scala/htif.scala

@@ -26,7 +26,7 @@ class rocketHTIF(w: Int, ncores: Int) extends Component
   val io = new Bundle {
     val host = new ioHost(w)
     val cpu = Vec(ncores) { new ioHTIF().flip() }
-    val mem = new ioMem
+    val mem = new ioTileLink
   }
 
   val short_request_bits = 64
@@ -93,11 +93,11 @@ class rocketHTIF(w: Int, ncores: Int) extends Component
   }
 
   val mem_cnt = Reg(resetVal = UFix(0, log2up(REFILL_CYCLES)))
-  when (state === state_mem_req && io.mem.req_rdy) {
+  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.req_data_rdy || 
+  when (state === state_mem_wdata && io.mem.xact_init_data.ready || 
-        state === state_mem_rdata && io.mem.resp_val) {
+        state === state_mem_rdata && io.mem.xact_rep.valid) {
     when (mem_cnt.andR)  {
       state := state_tx
     }
@@ -112,16 +112,17 @@ class rocketHTIF(w: Int, ncores: Int) extends Component
   var mem_req_data: Bits = null
   for (i <- 0 until MEM_DATA_BITS/short_request_bits) {
     val idx = Cat(mem_cnt, UFix(i, log2up(MEM_DATA_BITS/short_request_bits)))
-    packet_ram.write(idx, io.mem.resp_data((i+1)*short_request_bits-1, i*short_request_bits),
+    packet_ram.write(idx, io.mem.xact_rep.bits.data((i+1)*short_request_bits-1, i*short_request_bits),
-                     state === state_mem_rdata && io.mem.resp_val)
+                     state === state_mem_rdata && io.mem.xact_rep.valid)
     mem_req_data = Cat(packet_ram.read(idx), mem_req_data)
   }
-  io.mem.req_val := state === state_mem_req
+  io.mem.xact_init.valid := state === state_mem_req
-  io.mem.req_rw := cmd === cmd_writemem
+  io.mem.xact_init.bits.t_type := Mux(cmd === cmd_writemem, X_WRITE_UNCACHED, X_READ_UNCACHED)
-  io.mem.req_addr := addr >> UFix(OFFSET_BITS-3)
+  io.mem.xact_init.bits.has_data := cmd === cmd_writemem
+  io.mem.xact_init.bits.address := addr >> UFix(OFFSET_BITS-3)
 
-  io.mem.req_data_val := state === state_mem_wdata
+  io.mem.xact_init_data.valid:= state === state_mem_wdata
-  io.mem.req_data_bits := mem_req_data
+  io.mem.xact_init_data.bits.data := mem_req_data
 
   pcr_done := Bool(false)
   val pcr_mux = (new Mux1H(ncores)) { Bits(width = 64) }

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 ioMem
+  val mem = new ioTileLink
 }
 
 // basic direct mapped instruction cache
@@ -75,7 +75,7 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
 
   // refill counter
   val refill_count = Reg(resetVal = UFix(0, rf_cnt_bits));
-  when (io.mem.resp_val) {
+  when (io.mem.xact_rep.valid) {
     refill_count := refill_count + UFix(1);
   }
 
@@ -84,7 +84,7 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
   val tag_addr = 
     Mux((state === s_refill_wait), r_cpu_req_idx(indexmsb,indexlsb),
       io.cpu.req_idx(indexmsb,indexlsb)).toUFix;
-  val tag_we = (state === s_refill_wait) && io.mem.resp_val;
+  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;
@@ -112,10 +112,10 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
     val hit = valid && (tag_rdata === r_cpu_hit_addr(tagmsb,taglsb))
     
     // data array
-    val data_array = Mem(sets*REFILL_CYCLES){ io.mem.resp_data }
+    val data_array = Mem(sets*REFILL_CYCLES){ io.mem.xact_rep.bits.data }
     data_array.setReadLatency(1);
     data_array.setTarget('inst);
-    val data_out = data_array.rw(data_addr, io.mem.resp_data, io.mem.resp_val && repl_me)
+    val data_out = data_array.rw(data_addr, io.mem.xact_rep.bits.data, io.mem.xact_rep.valid && repl_me)
 
     data_mux.io.sel(i) := hit
     data_mux.io.in(i) := (data_out >> word_shift)(databits-1,0);
@@ -128,10 +128,11 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
   io.cpu.resp_val := !io.cpu.itlb_miss && (state === s_ready) && r_cpu_req_val && tag_hit;
   rdy := !io.cpu.itlb_miss && (state === s_ready) && (!r_cpu_req_val || tag_hit);
   io.cpu.resp_data := data_mux.io.out
-  io.mem.req_val := (state === s_request);
+  io.mem.xact_init.valid := (state === s_request)
-  io.mem.req_rw := Bool(false)
+  io.mem.xact_init.bits.t_type := X_READ_UNCACHED
-  io.mem.req_addr := r_cpu_miss_addr(tagmsb,indexlsb).toUFix
+  io.mem.xact_init.bits.has_data := Bool(false)
-  io.mem.req_data_val := Bool(false)
+  io.mem.xact_init.bits.address := r_cpu_miss_addr(tagmsb,indexlsb).toUFix
+  io.mem.xact_init_data.valid := Bool(false)
 
   // control state machine
   switch (state) {
@@ -148,17 +149,17 @@ class rocketICache(sets: Int, assoc: Int) extends Component {
     }
     is (s_request)
     {
-      when (io.mem.req_rdy) {
+      when (io.mem.xact_init.ready) {
         state := s_refill_wait;
       }
     }
     is (s_refill_wait) {
-      when (io.mem.resp_val) {
+      when (io.mem.xact_rep.valid) {
         state := s_refill;
       }
     }
     is (s_refill) {
-      when (io.mem.resp_val && (~refill_count === UFix(0))) {
+      when (io.mem.xact_rep.valid && refill_count.andR) {
         state := s_ready;
       }
     }

rocket/src/main/scala/icache_prefetch.scala

@@ -6,8 +6,8 @@ import Constants._;
 import scala.math._;
 
 class ioIPrefetcher extends Bundle() {
-  val icache = new ioMem().flip
+  val icache = new ioTileLink().flip
-  val mem = new ioMem
+  val mem = new ioTileLink
   val invalidate = Bool(INPUT)
 }
 
@@ -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 demand_miss = io.icache.req_val & io.icache.req_rdy;
+  val demand_miss = io.icache.xact_init.valid && io.icache.xact_init.ready
-  val prefetch_addr = Reg() { UFix(width = io.icache.req_addr.width) };
+  val prefetch_addr = Reg() { UFix(width = io.icache.xact_init.bits.address.width) };
-  when (demand_miss) { prefetch_addr := io.icache.req_addr + UFix(1); }
+  val addr_match = (prefetch_addr === io.icache.xact_init.bits.address);
-  
-  val addr_match = (prefetch_addr === io.icache.req_addr);
   val hit = (state != s_invalid) & (state != s_req_wait) & addr_match;
+  val prefetch_miss = io.icache.xact_init.valid && !hit
+  when (demand_miss) { prefetch_addr := io.icache.xact_init.bits.address + UFix(1); }
   
-  io.icache.req_rdy := io.mem.req_rdy;
+  io.icache.xact_init.ready := io.mem.xact_init.ready
-  val ip_mem_req_rdy  = io.mem.req_rdy & ~(io.icache.req_val & ~hit);
+  val ip_mem_req_rdy  = io.mem.xact_init.ready && !prefetch_miss
-  val ip_mem_resp_val = io.mem.resp_val && io.mem.resp_tag(0).toBool; 
+  val ip_mem_resp_val = io.mem.xact_rep.valid && io.mem.xact_rep.bits.tile_xact_id(0)
   
-  io.mem.req_val  := io.icache.req_val & ~hit | (state === s_req_wait);
+  io.mem.xact_init.valid  := prefetch_miss || (state === s_req_wait)
-  io.mem.req_rw   := Bool(false)
+  io.mem.xact_init.bits.t_type := X_READ_UNCACHED
-  io.mem.req_tag  := Mux(io.icache.req_val && !hit, UFix(0), UFix(1))
+  io.mem.xact_init.bits.has_data   := Bool(false)
-  io.mem.req_addr := Mux(io.mem.req_tag(0).toBool, prefetch_addr, io.icache.req_addr);
+  io.mem.xact_init.bits.tile_xact_id  := Mux(prefetch_miss, UFix(0), UFix(1))
-  io.mem.req_data_val := Bool(false)
+  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); }
@@ -45,11 +46,11 @@ class rocketIPrefetcher extends Component() {
   val forward_done = (~forward_cnt === UFix(0)) & pdq.io.deq.valid;
   forward := (demand_miss & hit | forward & ~forward_done);  
 
-  io.icache.resp_val  := (io.mem.resp_val && !io.mem.resp_tag(0).toBool) || (forward && pdq.io.deq.valid);
+  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.resp_data := Mux(forward, pdq.io.deq.bits, io.mem.resp_data);
+  io.icache.xact_rep.bits.data := Mux(forward, pdq.io.deq.bits, io.mem.xact_rep.bits.data)
   
   pdq.io.flush := Reg(demand_miss && !hit || (state === s_bad_resp_wait), resetVal = Bool(false))
-  pdq.io.enq.bits := io.mem.resp_data;
+  pdq.io.enq.bits := io.mem.xact_rep.bits.data
   pdq.io.enq.valid  := ip_mem_resp_val.toBool;
   pdq.io.deq.ready  := forward;
   

rocket/src/main/scala/nbdcache.scala

@@ -133,12 +133,6 @@ class DataArrayArrayReq extends Bundle {
   val way_en = Bits(width = NWAYS)
 }
 
-class MemReq extends Bundle {
-  val rw   = Bool()
-  val addr = UFix(width = PADDR_BITS-OFFSET_BITS)
-  val tag  = Bits(width = MEM_TAG_BITS)
-}
-
 class WritebackReq extends Bundle {
   val ppn = Bits(width = TAG_BITS)
   val idx = Bits(width = IDX_BITS)
@@ -182,7 +176,7 @@ class MSHR(id: Int) extends Component with ThreeStateIncoherence {
     val way_oh         = Bits(NWAYS, OUTPUT)
 
     val mem_resp_val = Bool(INPUT)
-    val mem_req  = (new ioDecoupled) { new MemReq()  }.flip
+    val mem_req  = (new ioDecoupled) { new TransactionInit }.flip
     val meta_req = (new ioDecoupled) { new MetaArrayArrayReq() }.flip
     val replay   = (new ioDecoupled) { new Replay()   }.flip
   }
@@ -257,10 +251,10 @@ class MSHR(id: Int) extends Component with ThreeStateIncoherence {
   io.meta_req.bits.way_en := way_oh_
 
   io.mem_req.valid := valid && !requested
-  //io.mem_req.bits.itm := next_dirty
+  io.mem_req.bits.t_type := Mux(needsWriteback(next_state), X_READ_EXCLUSIVE, X_READ_SHARED)
-  io.mem_req.bits.rw := Bool(false)
+  io.mem_req.bits.has_data := Bool(false)
-  io.mem_req.bits.addr := Cat(ppn, idx_).toUFix
+  io.mem_req.bits.address := Cat(ppn, idx_).toUFix
-  io.mem_req.bits.tag := Bits(id)
+  io.mem_req.bits.tile_xact_id := Bits(id)
 
   io.replay.valid := rpq.io.deq.valid && refilled
   io.replay.bits.idx := idx_
@@ -287,7 +281,7 @@ class MSHRFile extends Component {
 
     val fence_rdy = Bool(OUTPUT)
 
-    val mem_req  = (new ioDecoupled) { new MemReq()  }.flip()
+    val mem_req  = (new ioDecoupled) { new TransactionInit }.flip()
     val meta_req = (new ioDecoupled) { new MetaArrayArrayReq() }.flip()
     val replay   = (new ioDecoupled) { new Replay()   }.flip()
   }
@@ -296,7 +290,7 @@ class MSHRFile extends Component {
   val mem_resp_idx_mux = (new Mux1H(NMSHR)){ Bits(width = IDX_BITS) }
   val mem_resp_way_oh_mux = (new Mux1H(NMSHR)){ Bits(width =  NWAYS) }
   val meta_req_arb = (new Arbiter(NMSHR)) { new MetaArrayArrayReq() }
-  val mem_req_arb = (new Arbiter(NMSHR)) { new MemReq() }
+  val mem_req_arb = (new Arbiter(NMSHR)) { new TransactionInit }
   val replay_arb = (new Arbiter(NMSHR)) { new Replay() }
 
   val alloc_arb = (new Arbiter(NMSHR)) { Bool() }
@@ -421,9 +415,9 @@ class WritebackUnit extends Component {
     val req    = (new ioDecoupled) { new WritebackReq() }
     val data_req  = (new ioDecoupled) { new DataArrayArrayReq() }.flip()
     val data_resp = Bits(MEM_DATA_BITS, INPUT)
-    val refill_req = (new ioDecoupled) { new MemReq() }
+    val refill_req = (new ioDecoupled) { new TransactionInit }
-    val mem_req   = (new ioDecoupled) { new MemReq() }.flip()
+    val mem_req   = (new ioDecoupled) { new TransactionInit }.flip
-    val mem_req_data = (new ioDecoupled) { Bits(width = MEM_DATA_BITS) }.flip()
+    val mem_req_data = (new ioDecoupled) { new TransactionInitData }.flip
   }
 
   val valid = Reg(resetVal = Bool(false))
@@ -449,12 +443,13 @@ class WritebackUnit extends Component {
   val wb_req_val = io.req.valid && !valid
   io.refill_req.ready := io.mem_req.ready && !wb_req_val
   io.mem_req.valid := io.refill_req.valid || wb_req_val
-  io.mem_req.bits.rw := wb_req_val
+  io.mem_req.bits.t_type := Mux(wb_req_val, X_WRITE_UNCACHED, io.refill_req.bits.t_type)
-  io.mem_req.bits.addr := Mux(wb_req_val, Cat(io.req.bits.ppn, io.req.bits.idx).toUFix, io.refill_req.bits.addr)
+  io.mem_req.bits.has_data := wb_req_val
-  io.mem_req.bits.tag := io.refill_req.bits.tag
+  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.tile_xact_id := Mux(wb_req_val, Bits(NMSHR), io.refill_req.bits.tile_xact_id)
 
   io.mem_req_data.valid := data_req_fired
-  io.mem_req_data.bits := io.data_resp
+  io.mem_req_data.bits.data := io.data_resp
 }
 
 class FlushUnit(lines: Int) extends Component with ThreeStateIncoherence{
@@ -680,7 +675,7 @@ abstract class HellaCache extends Component {
 class HellaCacheUniproc extends HellaCache with ThreeStateIncoherence {
   val io = new Bundle {
     val cpu = new ioDmem()
-    val mem = new ioMem
+    val mem = new ioTileLink
   }
  
   val lines       = 1 << IDX_BITS
@@ -749,9 +744,11 @@ class HellaCacheUniproc extends HellaCache with ThreeStateIncoherence {
   val cpu_req_data = Mux(r_replay_amo, r_amo_replay_data, io.cpu.req_data)
 
   // refill counter
+  val mem_resp_type = io.mem.xact_rep.bits.t_type
+  val refill_val = io.mem.xact_rep.valid && (mem_resp_type === X_READ_SHARED || mem_resp_type === X_READ_EXCLUSIVE)
   val rr_count = Reg(resetVal = UFix(0, log2up(REFILL_CYCLES)))
   val rr_count_next = rr_count + UFix(1)
-  when (io.mem.resp_val) { rr_count := rr_count_next }
+  when (refill_val) { rr_count := rr_count_next }
 
   val misaligned =
     (((r_cpu_req_type === MT_H) || (r_cpu_req_type === MT_HU)) && (r_cpu_req_idx(0) != Bits(0))) ||
@@ -806,19 +803,19 @@ class HellaCacheUniproc extends HellaCache with ThreeStateIncoherence {
   val needs_writeback = needsWriteback(meta_wb_mux.state)
 
   // refill response
-  val block_during_refill = !io.mem.resp_val && (rr_count != UFix(0))
+  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.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.resp_data
+  data_arb.io.in(0).bits.inner_req.data := io.mem.xact_rep.bits.data
-  data_arb.io.in(0).valid := io.mem.resp_val || block_during_refill
+  data_arb.io.in(0).valid := refill_val || block_during_refill
 
   // load hits
   data_arb.io.in(4).bits.inner_req.offset := io.cpu.req_idx(offsetmsb,ramindexlsb)
   data_arb.io.in(4).bits.inner_req.idx := io.cpu.req_idx(indexmsb,indexlsb)
   data_arb.io.in(4).bits.inner_req.rw := Bool(false)
   data_arb.io.in(4).bits.inner_req.wmask := UFix(0) // don't care
-  data_arb.io.in(4).bits.inner_req.data := io.mem.resp_data // don't care
+  data_arb.io.in(4).bits.inner_req.data := io.mem.xact_rep.bits.data // don't care
   data_arb.io.in(4).valid := io.cpu.req_val && req_read
   data_arb.io.in(4).bits.way_en := ~UFix(0, NWAYS) // intiate load on all ways, mux after tag check
   val early_load_nack = req_read && !data_arb.io.in(4).ready
@@ -884,8 +881,8 @@ class HellaCacheUniproc extends HellaCache with ThreeStateIncoherence {
   mshr.io.req_type := r_cpu_req_type
   mshr.io.req_sdq_id := replayer.io.sdq_id
   mshr.io.req_way_oh := replaced_way_oh
-  mshr.io.mem_resp_val := io.mem.resp_val && (~rr_count === UFix(0))
+  mshr.io.mem_resp_val := refill_val && (~rr_count === UFix(0))
-  mshr.io.mem_resp_tag := io.mem.resp_tag
+  mshr.io.mem_resp_tag := io.mem.xact_rep.bits.tile_xact_id
   mshr.io.mem_req <> wb.io.refill_req
   mshr.io.meta_req <> meta_arb.io.in(1)
   mshr.io.replay <> replayer.io.replay
@@ -969,13 +966,6 @@ class HellaCacheUniproc extends HellaCache with ThreeStateIncoherence {
   io.cpu.resp_data := loadgen.io.dout
   io.cpu.resp_data_subword := loadgen.io.r_dout_subword
   
-  wb.io.mem_req.ready := io.mem.req_rdy
+  io.mem.xact_init <> wb.io.mem_req
-  io.mem.req_val   := wb.io.mem_req.valid
+  io.mem.xact_init_data <> wb.io.mem_req_data
-  io.mem.req_rw    := wb.io.mem_req.bits.rw
-  io.mem.req_tag   := wb.io.mem_req.bits.tag.toUFix
-  io.mem.req_addr  := wb.io.mem_req.bits.addr
-
-  io.mem.req_data_val := wb.io.mem_req_data.valid
-  wb.io.mem_req_data.ready := io.mem.req_data_rdy
-  io.mem.req_data_bits := wb.io.mem_req_data.bits
 }

rocket/src/main/scala/queues.scala

@@ -10,7 +10,7 @@ class ioQueue[T <: Data](flushable: Boolean)(data: => T) extends Bundle
   val deq   = new ioDecoupled()(data).flip
 }
 
-class queue[T <: Data](entries: Int, flushable: Boolean = false)(data: => T) extends Component
+class queue[T <: Data](entries: Int, pipe: Boolean = false, flushable: Boolean = false)(data: => T) extends Component
 {
   val io = new ioQueue(flushable)(data)
 
@@ -50,6 +50,6 @@ class queue[T <: Data](entries: Int, flushable: Boolean = false)(data: => T) ext
   }
 
   io.deq.valid :=  maybe_full || enq_ptr != deq_ptr
-  io.enq.ready := !maybe_full || enq_ptr != deq_ptr
+  io.enq.ready := !maybe_full || enq_ptr != deq_ptr || (if (pipe) io.deq.ready else Bool(false))
   io.deq.bits <> Mem(entries, do_enq, enq_ptr, io.enq.bits).read(deq_ptr)
 }

rocket/src/main/scala/top.scala

@@ -7,7 +7,7 @@ import Constants._;
 class ioTop(htif_width: Int) extends Bundle  {
   val debug   = new ioDebug();
   val host    = new ioHost(htif_width);
-  val mem     = new ioMem();
+  val mem     = new ioMem
 }
 
 class Top() extends Component {
@@ -21,23 +21,39 @@ class Top() extends Component {
   val icache_pf = new rocketIPrefetcher();
   val dcache    = new HellaCacheUniproc();
 
-  val arbiter   = new rocketMemArbiter(4);
+  val arbiter   = new rocketMemArbiter(3 + (if (HAVE_VEC) 1 else 0));
   arbiter.io.requestor(0) <> dcache.io.mem
   arbiter.io.requestor(1) <> icache_pf.io.mem
-  arbiter.io.requestor(3) <> htif.io.mem
+  arbiter.io.requestor(2) <> htif.io.mem
-  arbiter.io.mem <> io.mem
+
+  val hub = new CoherenceHubNull
+  // connect tile to hub (figure out how to do this more compactly)
+  val xact_init_q = (new queue(2)) { new TransactionInit }
+  xact_init_q.io.enq <> arbiter.io.mem.xact_init
+  xact_init_q.io.deq <> hub.io.tile.xact_init
+  val xact_init_data_q = (new queue(2)) { new TransactionInitData }
+  xact_init_data_q.io.enq <> arbiter.io.mem.xact_init_data
+  xact_init_data_q.io.deq <> hub.io.tile.xact_init_data
+  val xact_rep_q = (new queue(1, pipe = true)) { new TransactionReply }
+  xact_rep_q.io.enq <> hub.io.tile.xact_rep
+  xact_rep_q.io.deq <> arbiter.io.mem.xact_rep
+  // connect hub to memory
+  val mem_req_q = (new queue(2)) { new MemReqCmd }
+  mem_req_q.io.enq <> hub.io.mem.req_cmd
+  mem_req_q.io.deq <> io.mem.req_cmd
+  val mem_req_data_q = (new queue(2)) { new MemData }
+  mem_req_data_q.io.enq <> hub.io.mem.req_data
+  mem_req_data_q.io.deq <> io.mem.req_data
+  hub.io.mem.resp.valid := Reg(io.mem.resp.valid, resetVal = Bool(false))
+  hub.io.mem.resp.bits := Reg(io.mem.resp.bits)
+
 
   if (HAVE_VEC)
   {
     val vicache = new rocketICache(128, 2); // 128 sets x 2 ways
-    arbiter.io.requestor(2) <> vicache.io.mem
+    arbiter.io.requestor(3) <> vicache.io.mem
     cpu.io.vimem <> vicache.io.cpu;
   }
-  else
-  {
-    arbiter.io.requestor(2).req_val := Bool(false)
-    arbiter.io.requestor(2).req_data_val := Bool(false)
-  }
 
   htif.io.host <> io.host
   cpu.io.host       <> htif.io.cpu(0);

rocket/src/main/scala/util.scala

@@ -180,6 +180,12 @@ class ioDecoupled[T <: Data]()(data: => T) extends Bundle
   val bits  = data.asInput
 }
 
+class ioValid[T <: Data]()(data: => T) extends Bundle
+{
+  val valid = Bool(INPUT)
+  val bits = data.asInput
+}
+
 class ioArbiter[T <: Data](n: Int)(data: => T) extends Bundle {
   val in  = Vec(n) { (new ioDecoupled()) { data } }
   val out = (new ioDecoupled()) { data }.flip()