tilelink2 RAMModel: clear Mems on power-up

src/main/scala/uncore/tilelink2/RAMModel.scala

@@ -15,10 +15,11 @@ class TLRAMModel extends LazyModule
       val out = node.bundleOut
     }
 
-    // Pass through all signals unchanged
-    io.out <> io.in
     require (io.out.size == 1) // !!! support multiple clients
 
+    val in = io.in(0)
+    val out = io.out(0)
+
     val edge         = node.edgesIn(0)
     val endAddress   = edge.manager.maxAddress + 1
     val endSourceId  = edge.client.endSourceId
@@ -32,6 +33,27 @@ class TLRAMModel extends LazyModule
     val countBits    = log2Up(endSourceId)
     val sizeBits     = edge.bundle.sizeBits
 
+    // Reset control logic
+    val wipeIndex = RegInit(UInt(0, width = log2Ceil(endAddressHi) + 1))
+    val wipe = !wipeIndex(log2Ceil(endAddressHi))
+    wipeIndex := wipeIndex + wipe.asUInt
+
+    // Block traffic while wiping Mems
+    in.a.ready := out.a.ready && !wipe
+    out.a.valid := in.a.valid && !wipe
+    out.a.bits  := in.a.bits
+    out.d.ready := in.d.ready && !wipe
+    in.d.valid := out.d.valid && !wipe
+    in.d.bits  := out.d.bits
+
+    // BCE unsupported
+    in.b.valid := Bool(false)
+    out.c.valid := Bool(false)
+    out.e.valid := Bool(false)
+    out.b.ready := Bool(true)
+    in.c.ready := Bool(true)
+    in.e.ready := Bool(true)
+
     class ByteMonitor extends Bundle {
       val valid = Bool()
       val value = UInt(width = 8)
@@ -42,33 +64,39 @@ class TLRAMModel extends LazyModule
       val opcode  = UInt(width = 3)
     }
 
-    // !!! Must somehow power-on with these all initialized with 0
     val shadow = Seq.fill(beatBytes) { Mem(endAddressHi, new ByteMonitor) }
     val inc_bytes = Seq.fill(beatBytes) { Mem(endAddressHi, UInt(width = countBits)) }
     val dec_bytes = Seq.fill(beatBytes) { Mem(endAddressHi, UInt(width = countBits)) }
     val inc_trees = Seq.tabulate(decTrees) { i => Mem(endAddressHi >> (i+1), UInt(width = countBits)) }
     val dec_trees = Seq.tabulate(decTrees) { i => Mem(endAddressHi >> (i+1), UInt(width = countBits)) }
 
-    // Don't care on power-up
+    val shadow_wen = Wire(init = Fill(beatBytes, wipe))
+    val inc_bytes_wen = Wire(init = Fill(beatBytes, wipe))
+    val dec_bytes_wen = Wire(init = Fill(beatBytes, wipe))
+    val inc_trees_wen = Wire(init = Fill(decTrees, wipe))
+    val dec_trees_wen = Wire(init = Fill(decTrees, wipe))
+
+    // Don't care on power-up !!! Mem ?
     val flight = Reg(Vec(endSourceId, new FlightMonitor))
 
     // Process A access requests
-    val a = RegNext(io.in(0).a)
+    val a = Reg(next = in.a.bits)
-    val a_beats1 = edge.numBeats1(a.bits)
+    val a_fire = Reg(next = in.a.fire(), init = Bool(false))
-    val a_size = edge.size(a.bits)
+    val a_beats1 = edge.numBeats1(a)
+    val a_size = edge.size(a)
     val a_sizeOH = UIntToOH(a_size)
     val a_counter = RegInit(UInt(0, width = maxLgBeats))
     val a_counter1 = a_counter - UInt(1)
     val a_first = a_counter === UInt(0)
-    val a_addr_hi = a.bits.addr_hi | (a_beats1 & ~a_counter1)
+    val a_addr_hi = a.addr_hi | (a_beats1 & ~a_counter1)
-    val a_base = edge.address(a.bits)
+    val a_base = edge.address(a)
     val a_mask = edge.mask(a_base, a_size)
 
     // What is the request?
     val a_flight = Wire(new FlightMonitor)
     a_flight.base   := a_base
     a_flight.size   := a_size
-    a_flight.opcode := a.bits.opcode
+    a_flight.opcode := a.opcode
 
     // Grab the concurrency state we need
     val a_inc_bytes = inc_bytes.map(_.read(a_addr_hi))
@@ -80,56 +108,72 @@ class TLRAMModel extends LazyModule
     val a_inc = a_inc_bytes.map(_ + a_inc_tree)
     val a_dec = a_dec_bytes.map(_ + a_dec_tree)
 
-    when (a.fire()) {
+    when (a_fire) {
       // Record the request so we can handle it's response
-      flight(a.bits.source) := a_flight
+      flight(a.source) := a_flight
       a_counter := Mux(a_first, a_beats1, a_counter1)
 
       // !!! atomics
-      assert (a.bits.opcode =/= TLMessages.Acquire)
+      assert (a.opcode =/= TLMessages.Acquire)
 
       // Increase the per-byte flight counter for the whole transaction
-      when (a_first && a.bits.opcode =/= TLMessages.Hint) {
+      when (a_first && a.opcode =/= TLMessages.Hint) {
         when (a_size <= UInt(shift)) {
-          for (i <- 0 until beatBytes) {
+          inc_bytes_wen := a_mask
-            when (a_mask(i)) { // not a.bits.mask; the full mask
-              inc_bytes(i).write(a_addr_hi, a_inc_bytes(i) + UInt(1))
-            }
-          }
-        }
-        for (i <- 0 until inc_trees.size) {
-          when (a_sizeOH(i+shift+1)) {
-            inc_trees(i).write(a_addr_hi >> (i+1), a_inc_trees(i) + UInt(1))
-          }
         }
+        inc_trees_wen := a_sizeOH >> (shift+1)
       }
 
-      when (a.bits.opcode === TLMessages.PutFullData || a.bits.opcode === TLMessages.PutPartialData) {
+      when (a.opcode === TLMessages.PutFullData || a.opcode === TLMessages.PutPartialData) {
+        shadow_wen := a.mask
         for (i <- 0 until beatBytes) {
-          val set = Wire(new ByteMonitor)
           val busy = a_inc(i) - a_dec(i) - (!a_first).asUInt
-          set.valid := busy === UInt(0)
+          val byte = a.data(8*(i+1)-1, 8*i)
-          set.value := a.bits.data(8*(i+1)-1, 8*i)
+          when (a.mask(i)) {
-          when (a.bits.mask(i)) {
+            printf("P 0x%x := 0x%x #%d\n", a_addr_hi << shift | UInt(i), byte, busy)
-            shadow(i).write(a_addr_hi, set)
-            printf("P 0x%x := 0x%x #%d\n", a_addr_hi << shift | UInt(i), set.value, busy)
           }
         }
       }
     }
 
+    val a_waddr = Mux(wipe, wipeIndex, a_addr_hi)
+    for (i <- 0 until beatBytes) {
+      val data = Wire(new ByteMonitor)
+      val busy = a_inc(i) - a_dec(i) - (!a_first).asUInt
+      data.valid := Mux(wipe, Bool(false), busy === UInt(0))
+      data.value := a.data(8*(i+1)-1, 8*i)
+      when (shadow_wen(i)) {
+        shadow(i).write(a_waddr, data)
+      }
+    }
+
+    for (i <- 0 until beatBytes) {
+      val data = Mux(wipe, UInt(0), a_inc_bytes(i) + UInt(1))
+      when (inc_bytes_wen(i)) {
+        inc_bytes(i).write(a_waddr, data)
+      }
+    }
+
+    for (i <- 0 until inc_trees.size) {
+      val data = Mux(wipe, UInt(0), a_inc_trees(i) + UInt(1))
+      when (inc_trees_wen(i)) {
+        inc_trees(i).write(a_waddr >> (i+1), data)
+      }
+    }
+
     // Process D access responses
-    val d = RegNext(io.out(0).d)
+    val d = RegNext(out.d.bits)
-    val d_bypass = a.valid && d.bits.source === a.bits.source
+    val d_fire = Reg(next = out.d.fire(), init = Bool(false))
-    val d_flight = Mux(d_bypass, a_flight, flight(d.bits.source))
+    val d_bypass = a_fire && d.source === a.source
-    val d_beats1 = edge.numBeats1(d.bits)
+    val d_flight = Mux(d_bypass, a_flight, flight(d.source))
-    val d_size = edge.size(d.bits)
+    val d_beats1 = edge.numBeats1(d)
+    val d_size = edge.size(d)
     val d_sizeOH = UIntToOH(d_size)
     val d_counter = RegInit(UInt(0, width = maxLgBeats))
     val d_counter1 = d_counter - UInt(1)
     val d_first = d_counter === UInt(0)
     val d_last  = d_counter === UInt(1) || d_beats1 === UInt(0)
-    val d_base = d_flight.base
+    val d_base = d_flight.base // !!! not a register => can't be absorbed
     val d_addr_hi = d_base >> shift | (d_beats1 & ~d_counter1)
     val d_mask = edge.mask(d_base, d_size)
 
@@ -144,44 +188,36 @@ class TLRAMModel extends LazyModule
     val d_dec = d_dec_bytes.map(_ + d_dec_tree)
     val d_shadow = shadow.map(_.read(d_addr_hi))
 
-    when (d.fire()) {
+    when (d_fire) {
       assert (d_size === d_flight.size)
       d_counter := Mux(d_first, d_beats1, d_counter1)
 
       when (d_flight.opcode === TLMessages.Hint) {
-        assert (d.bits.opcode === TLMessages.HintAck)
+        assert (d.opcode === TLMessages.HintAck)
       }
 
       // Decreaes the per-byte flight counter for the whole transaction
       when (d_last && d_flight.opcode =/= TLMessages.Hint) {
         when (d_size <= UInt(shift)) {
-          for (i <- 0 until beatBytes) {
+          dec_bytes_wen := d_mask
-            when (d_mask(i)) {
-              dec_bytes(i).write(d_addr_hi, d_dec_bytes(i) + UInt(1))
-            }
-          }
-        }
-        for (i <- 0 until dec_trees.size) {
-          when (d_sizeOH(i+shift+1)) {
-            dec_trees(i).write(d_addr_hi >> (i+1), d_dec_trees(i) + UInt(1))
-          }
         }
+        dec_trees_wen := d_sizeOH >> (shift+1)
       }
 
       when (d_flight.opcode === TLMessages.PutFullData || d_flight.opcode === TLMessages.PutPartialData) {
-        assert (d.bits.opcode === TLMessages.AccessAck)
+        assert (d.opcode === TLMessages.AccessAck)
       }
 
       // !!! atomics
 
       when (d_flight.opcode === TLMessages.Get) {
-        assert (d.bits.opcode === TLMessages.AccessAckData)
+        assert (d.opcode === TLMessages.AccessAckData)
         for (i <- 0 until beatBytes) {
-          val got = d.bits.data(8*(i+1)-1, 8*i)
+          val got = d.data(8*(i+1)-1, 8*i)
           val shadow = Wire(init = d_shadow(i))
           when (d_mask(i)) {
             when (!shadow.valid) {
-              printf("G 0x%x := undefined due to concurrent accesses\n", d_addr_hi << shift | UInt(i))
+              printf("G 0x%x := undefined\n", d_addr_hi << shift | UInt(i))
             } .otherwise {
               printf("G 0x%x := 0x%x\n", d_addr_hi << shift | UInt(i), shadow.value)
               assert (shadow.value === got)
@@ -190,5 +226,20 @@ class TLRAMModel extends LazyModule
         }
       }
     }
+
+    val d_waddr = Mux(wipe, wipeIndex, d_addr_hi)
+    for (i <- 0 until beatBytes) {
+      val data = Mux(wipe, UInt(0), d_dec_bytes(i) + UInt(1))
+      when (dec_bytes_wen(i)) {
+        dec_bytes(i).write(d_waddr, data)
+      }
+    }
+
+    for (i <- 0 until dec_trees.size) {
+      val data = Mux(wipe, UInt(0), d_dec_trees(i) + UInt(1))
+      when (dec_trees_wen(i)) {
+        dec_trees(i).write(d_waddr >> (i+1), data)
+      }
+    }
   }
 }