More broadcast hub bugfixes

rocket/src/main/scala/coherence.scala

@@ -276,12 +276,12 @@ class XactTracker(id: Int) extends Component with FourStateCoherence {
     when(req_data.ready && req_data.valid) {
       pop_data := UFix(1) << tile_id 
       mem_cnt  := mem_cnt_next
-    }
       when(mem_cnt_next === UFix(0)) {
         pop_dep := UFix(1) << tile_id
         trigger := Bool(false)
       }
     }
+  }
 
   def doMemReqRead(req_cmd: ioDecoupled[MemReqCmd], trigger: Bool) {
     req_cmd.valid := Bool(true)
@@ -535,10 +535,11 @@ class CoherenceHubBroadcast extends CoherenceHub  with FourStateCoherence{
       rep.bits.t_type := getTransactionReplyType(t_type_arr(ack_idx), sh_count_arr(ack_idx))
       rep.bits.tile_xact_id := tile_xact_id_arr(ack_idx)
       rep.bits.global_xact_id := ack_idx
-      rep.valid := (UFix(j) === init_tile_id_arr(ack_idx)) && send_x_rep_ack_arr.toBits.orR
+      val do_send_ack = (UFix(j) === init_tile_id_arr(ack_idx)) && send_x_rep_ack_arr.toBits.orR
+      rep.valid := do_send_ack
+      sent_x_rep_ack_arr(ack_idx) := do_send_ack
     }
   }
-  sent_x_rep_ack_arr(ack_idx) := !io.mem.resp.valid
   // If there were a ready signal due to e.g. intervening network use:
   //io.mem.resp.ready  := io.tiles(init_tile_id_arr.read(mem_idx)).xact_rep.ready
 

rocket/src/main/scala/util.scala

@@ -194,40 +194,33 @@ class ioLockingArbiter[T <: Data](n: Int)(data: => T) extends Bundle {
 class LockingArbiter[T <: Data](n: Int)(data: => T) extends Component {
   val io = new ioLockingArbiter(n)(data)
   val locked = Vec(n) { Reg(resetVal = Bool(false)) }
-  var dout = io.in(0).bits
-  var vout = Bool(false)
-
-  for (i <- 0 until n) {
-    io.in(i).ready := io.out.ready
-  }
-
   val any_lock_held = (locked.toBits & io.lock.toBits).orR
-  when(any_lock_held) {
-    vout = io.in(0).valid && locked(0)
+  val valid_arr = Vec(n) { Wire() { Bool() } }
+  val bits_arr = Vec(n) { Wire() { data } }
   for(i <- 0 until n) {
-      io.in(i).ready := io.out.ready && locked(i)
-      dout = Mux(locked(i), io.in(i).bits, dout)
-      vout = vout || io.in(i).valid && locked(i)
-    }
-  } .otherwise {
-    io.in(0).ready := io.out.ready
-    locked(0) := io.out.ready && io.lock(0)
-    for (i <- 1 until n) {
-      io.in(i).ready := !io.in(i-1).valid && io.in(i-1).ready
-      locked(i) := !io.in(i-1).valid && io.in(i-1).ready && io.lock(i)
+    valid_arr(i) := io.in(i).valid
+    bits_arr(i) := io.in(i).bits
   }
 
-    dout = io.in(n-1).bits
+  io.in(0).ready := Mux(any_lock_held, io.out.ready && locked(0), io.out.ready)
+  locked(0) := Mux(any_lock_held, locked(0), io.in(0).ready && io.lock(0))
+  for (i <- 1 until n) {
+    io.in(i).ready := Mux(any_lock_held, io.out.ready && locked(i), 
+                          !io.in(i-1).valid && io.in(i-1).ready)
+    locked(i) := Mux(any_lock_held, locked(i), io.in(i).ready)
+  }
+
+  var dout = io.in(n-1).bits
   for (i <- 1 until n)
     dout = Mux(io.in(n-1-i).valid, io.in(n-1-i).bits, dout)
 
-    vout = io.in(0).valid
+  var vout = io.in(0).valid
   for (i <- 1 until n)
     vout = vout || io.in(i).valid
-  }
 
-  vout <> io.out.valid
-  dout <> io.out.bits
+  val lock_idx = PriorityEncoder(locked.toBits)
+  io.out.valid := Mux(any_lock_held, valid_arr(lock_idx), vout)
+  io.out.bits  := Mux(any_lock_held, bits_arr(lock_idx), dout)
 }
 
 object PriorityEncoder