Merge pull request #333 from ucb-bar/tl2-api-changes

Tl2 api changes

src/main/scala/rocketchip/BaseTop.scala

@@ -49,7 +49,7 @@ abstract class BaseTop(q: Parameters) extends LazyModule {
 
   val legacy = LazyModule(new TLLegacy()(p.alterPartial({ case TLId => "L2toMMIO" })))
 
-  peripheryBus.node := TLWidthWidget(TLBuffer(TLAtomicAutomata()(TLHintHandler(legacy.node))), legacy.tlDataBytes)
+  peripheryBus.node := TLWidthWidget(legacy.tlDataBytes)(TLBuffer()(TLAtomicAutomata()(TLHintHandler()(legacy.node))))
 }
 
 abstract class BaseTopBundle(val p: Parameters) extends Bundle {

src/main/scala/rocketchip/Periphery.scala

@@ -288,7 +288,7 @@ trait PeripheryCoreplexLocalInterrupter extends LazyModule with HasPeripheryPara
   val clintConfig = CoreplexLocalInterrupterConfig(beatBytes)
   val clint = LazyModule(new CoreplexLocalInterrupter(clintConfig)(innerMMIOParams))
   // The periphery bus is 32-bit, so we may need to adapt its width to XLen
-  clint.node := TLFragmenter(TLWidthWidget(peripheryBus.node, 4), beatBytes, 256)
+  clint.node := TLFragmenter(beatBytes, 256)(TLWidthWidget(4)(peripheryBus.node))
 }
 
 trait PeripheryCoreplexLocalInterrupterBundle {
@@ -314,7 +314,7 @@ trait PeripheryBootROM extends LazyModule {
   val address = 0x1000
   val size = 0x1000
   val rom = LazyModule(new TLROM(address, size, GenerateBootROM(p, address)) { override def name = "bootrom" })
-  rom.node := TLFragmenter(peripheryBus.node, 4, 256)
+  rom.node := TLFragmenter(4, 256)(peripheryBus.node)
 }
 
 trait PeripheryBootROMBundle {
@@ -337,7 +337,7 @@ trait PeripheryTestRAM extends LazyModule {
   val ramSize = 0x1000
 
   val sram = LazyModule(new TLRAM(AddressSet(ramBase, ramSize-1)) { override def name = "testram" })
-  sram.node := TLFragmenter(peripheryBus.node, 4, 256)
+  sram.node := TLFragmenter(4, 256)(peripheryBus.node)
 }
 
 trait PeripheryTestRAMBundle {

src/main/scala/uncore/devices/Prci.scala

@@ -74,14 +74,11 @@ trait CoreplexLocalInterrupterModule extends Module with HasRegMap with MixCorep
    * bff8 mtime lo
    * bffc mtime hi
    */
-  val ipi_base     = 0
-  val timecmp_base = c.timecmpOffset(0) / c.beatBytes
-  val time_base    = c.timeOffset / c.beatBytes
 
-  regmap((
+  regmap(
-    RegField.split(makeRegFields(ipi),             ipi_base,     c.beatBytes) ++
+    0                  -> makeRegFields(ipi),
-    RegField.split(makeRegFields(timecmp.flatten), timecmp_base, c.beatBytes) ++
+    c.timecmpOffset(0) -> makeRegFields(timecmp.flatten),
-    RegField.split(makeRegFields(time),            time_base,    c.beatBytes)):_*)
+    c.timeOffset       -> makeRegFields(time))
 
   def makeRegFields(s: Seq[UInt]) = s.map(r => RegField(regWidth, r))
 }

src/main/scala/uncore/tilelink2/Buffer.scala

@@ -48,12 +48,12 @@ class TLBuffer(a: Int = 2, b: Int = 2, c: Int = 2, d: Int = 2, e: Int = 2, pipe:
 object TLBuffer
 {
   // applied to the TL source node; y.node := TLBuffer(x.node)
-  def apply(x: TLBaseNode)                                  (implicit sourceInfo: SourceInfo): TLBaseNode = apply(x, 2)
+  def apply()                                (x: TLBaseNode)(implicit sourceInfo: SourceInfo): TLBaseNode = apply(2)(x)
-  def apply(x: TLBaseNode, entries: Int)                    (implicit sourceInfo: SourceInfo): TLBaseNode = apply(x, entries, true)
+  def apply(entries: Int)                    (x: TLBaseNode)(implicit sourceInfo: SourceInfo): TLBaseNode = apply(entries, true)(x)
-  def apply(x: TLBaseNode, entries: Int, pipe: Boolean)     (implicit sourceInfo: SourceInfo): TLBaseNode = apply(x, entries, entries, pipe)
+  def apply(entries: Int, pipe: Boolean)     (x: TLBaseNode)(implicit sourceInfo: SourceInfo): TLBaseNode = apply(entries, entries, pipe)(x)
-  def apply(x: TLBaseNode, ace: Int, bd: Int)               (implicit sourceInfo: SourceInfo): TLBaseNode = apply(x, ace, bd, true)
+  def apply(ace: Int, bd: Int)               (x: TLBaseNode)(implicit sourceInfo: SourceInfo): TLBaseNode = apply(ace, bd, true)(x)
-  def apply(x: TLBaseNode, ace: Int, bd: Int, pipe: Boolean)(implicit sourceInfo: SourceInfo): TLBaseNode = apply(x, ace, bd, ace, bd, ace, pipe)
+  def apply(ace: Int, bd: Int, pipe: Boolean)(x: TLBaseNode)(implicit sourceInfo: SourceInfo): TLBaseNode = apply(ace, bd, ace, bd, ace, pipe)(x)
-  def apply(x: TLBaseNode, a: Int, b: Int, c: Int, d: Int, e: Int, pipe: Boolean = true)(implicit sourceInfo: SourceInfo): TLBaseNode = {
+  def apply(a: Int, b: Int, c: Int, d: Int, e: Int, pipe: Boolean = true)(x: TLBaseNode)(implicit sourceInfo: SourceInfo): TLBaseNode = {
     val buffer = LazyModule(new TLBuffer(a, b, c, d, e, pipe))
     buffer.node := x
     buffer.node

src/main/scala/uncore/tilelink2/Example.scala

@@ -27,7 +27,7 @@ trait ExampleModule extends HasRegMap
   regmap(
     0 -> Seq(
       RegField(params.num, state)),
-    1 -> Seq(
+    4 -> Seq(
       RegField.w1ToClear(4, pending, state)))
 }
 

src/main/scala/uncore/tilelink2/Fragmenter.scala

@@ -244,7 +244,7 @@ class TLFragmenter(minSize: Int, maxSize: Int, alwaysMin: Boolean = false) exten
 object TLFragmenter
 {
   // applied to the TL source node; y.node := TLFragmenter(x.node, 256, 4)
-  def apply(x: TLBaseNode, minSize: Int, maxSize: Int, alwaysMin: Boolean = false)(implicit sourceInfo: SourceInfo): TLBaseNode = {
+  def apply(minSize: Int, maxSize: Int, alwaysMin: Boolean = false)(x: TLBaseNode)(implicit sourceInfo: SourceInfo): TLBaseNode = {
     val fragmenter = LazyModule(new TLFragmenter(minSize, maxSize, alwaysMin))
     fragmenter.node := x
     fragmenter.node

src/main/scala/uncore/tilelink2/Fuzzer.scala

@@ -222,11 +222,11 @@ class TLFuzzRAM extends LazyModule
 
   model.node := fuzz.node
   xbar2.node := TLAtomicAutomata()(model.node)
-  ram2.node := TLFragmenter(xbar2.node, 16, 256)
+  ram2.node := TLFragmenter(16, 256)(xbar2.node)
-  xbar.node := TLWidthWidget(TLHintHandler(xbar2.node), 16)
+  xbar.node := TLWidthWidget(16)(TLHintHandler()(xbar2.node))
-  cross.node := TLFragmenter(TLBuffer(xbar.node), 4, 256)
+  cross.node := TLFragmenter(4, 256)(TLBuffer()(xbar.node))
   val monitor = (ram.node := cross.node)
-  gpio.node := TLFragmenter(TLBuffer(xbar.node), 4, 32)
+  gpio.node := TLFragmenter(4, 32)(TLBuffer()(xbar.node))
 
   lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
     io.finished := fuzz.module.io.finished

src/main/scala/uncore/tilelink2/HintHandler.scala

@@ -135,7 +135,7 @@ class TLHintHandler(supportManagers: Boolean = true, supportClients: Boolean = f
 object TLHintHandler
 {
   // applied to the TL source node; y.node := TLHintHandler(x.node)
-  def apply(x: TLBaseNode, supportManagers: Boolean = true, supportClients: Boolean = false, passthrough: Boolean = true)(implicit sourceInfo: SourceInfo): TLBaseNode = {
+  def apply(supportManagers: Boolean = true, supportClients: Boolean = false, passthrough: Boolean = true)(x: TLBaseNode)(implicit sourceInfo: SourceInfo): TLBaseNode = {
     val hints = LazyModule(new TLHintHandler(supportManagers, supportClients, passthrough))
     hints.node := x
     hints.node

src/main/scala/uncore/tilelink2/RegField.scala

@@ -77,7 +77,7 @@ object RegWriteFn
   implicit def apply(x: Unit): RegWriteFn = RegWriteFn((valid, data) => { Bool(true) })
 }
 
-case class RegField(width: Int, read: RegReadFn, write: RegWriteFn)
+case class RegField(width: Int, read: RegReadFn, write: RegWriteFn, name: String, description: String)
 {
   require (width > 0)
   def pipelined = !read.combinational || !write.combinational
@@ -85,11 +85,15 @@ case class RegField(width: Int, read: RegReadFn, write: RegWriteFn)
 
 object RegField
 {
+  // Byte address => sequence of bitfields, lowest index => lowest address
   type Map = (Int, Seq[RegField])
-  def apply(n: Int)            : RegField = apply(n, (), ())
+
-  def apply(n: Int, rw: UInt)  : RegField = apply(n, rw, rw)
+  def apply(n: Int)                                                         : RegField = apply(n, (), (), "", "")
-  def r(n: Int, r: RegReadFn)  : RegField = apply(n, r, ())
+  def apply(n: Int, r: RegReadFn, w: RegWriteFn)                            : RegField = apply(n, r, w,   "", "")
-  def w(n: Int, w: RegWriteFn) : RegField = apply(n, (), w)
+  def apply(n: Int, rw: UInt)                                               : RegField = apply(n, rw, rw, "", "")
+  def apply(n: Int, rw: UInt, name: String, description: String)            : RegField = apply(n, rw, rw, name, description)
+  def r(n: Int, r: RegReadFn,  name: String = "", description: String = "") : RegField = apply(n, r,  (), name, description)
+  def w(n: Int, w: RegWriteFn, name: String = "", description: String = "") : RegField = apply(n, (), w,  name, description)
 
   // This RegField allows 'set' to set bits in 'reg'.
   // and to clear bits when the bus writes bits of value 1.
@@ -105,28 +109,6 @@ object RegField
       bb.d := data
       Bool(true)
     }))
-
-  // Split a large register into a sequence of byte fields
-  // The bytes can be individually written, as they are one byte per field
-  def bytes(x: UInt): Seq[RegField] = {
-    require (x.getWidth % 8 == 0)
-    val bytes = Seq.tabulate(x.getWidth/8) { i => x(8*(i+1)-1, 8*i) }
-    val wires = bytes.map { b => Wire(init = b) }
-    x := Cat(wires.reverse)
-    Seq.tabulate(x.getWidth/8) { i =>
-      RegField(8, bytes(i), RegWriteFn { (valid, data) =>
-        when (valid) { wires(i) := data }
-        Bool(true)
-      })
-    }
-  }
-
-  // Divide a long sequence of RegFields into a maximum sized registers
-  // Your input RegFields may not cross a beatBytes boundary!
-  def split(fields: Seq[RegField], base: Int, beatBytes: Int = 4): Seq[RegField.Map] = {
-    val offsets = fields.map(_.width).scanLeft(0)(_ + _).init
-    (offsets zip fields).groupBy(_._1 / (beatBytes*8)).toList.map(r => (r._1 + base, r._2.map(_._2)))
-  }
 }
 
 trait HasRegMap

src/main/scala/uncore/tilelink2/RegMapper.scala

@@ -29,26 +29,35 @@ object RegMapper
 {
   // Create a generic register-based device
   def apply(bytes: Int, concurrency: Int, undefZero: Boolean, in: DecoupledIO[RegMapperInput], mapping: RegField.Map*) = {
-    val regmap = mapping.toList.filter(!_._2.isEmpty)
+    val bytemap = mapping.toList
-    require (!regmap.isEmpty)
-
-    // Ensure no register appears twice
-    regmap.combinations(2).foreach { case Seq((reg1, _), (reg2, _)) =>
-      require (reg1 != reg2)
-    }
     // Don't be an asshole...
-    regmap.foreach { reg => require (reg._1 >= 0) }
+    bytemap.foreach { byte => require (byte._1 >= 0) }
+
+    // Transform all fields into bit offsets Seq[(bit, field)]
+    val bitmap = bytemap.map { case (byte, fields) =>
+      val bits = fields.scanLeft(byte * 8)(_ + _.width).init
+      bits zip fields
+    }.flatten.sortBy(_._1)
+
+    // Detect overlaps
+    (bitmap.init zip bitmap.tail) foreach { case ((lbit, lfield), (rbit, rfield)) =>
+      require (lbit + lfield.width <= rbit, s"Register map overlaps at bit ${rbit}.")
+    }
+
+    // Group those fields into bus words Map[word, List[(bit, field)]]
+    val wordmap = bitmap.groupBy(_._1 / (8*bytes))
+
     // Make sure registers fit
     val inParams = in.bits.params
     val inBits = inParams.indexBits
-    assert (regmap.map(_._1).max < (1 << inBits))
+    assert (wordmap.keySet.max < (1 << inBits), "Register map does not fit in device")
 
     val out = Wire(Irrevocable(new RegMapperOutput(inParams)))
     val front = Wire(Irrevocable(new RegMapperInput(inParams)))
     front.bits := in.bits
 
     // Must this device pipeline the control channel?
-    val pipelined = regmap.map(_._2.map(_.pipelined)).flatten.reduce(_ || _)
+    val pipelined = wordmap.values.map(_.map(_._2.pipelined)).flatten.reduce(_ || _)
     val depth = concurrency
     require (depth >= 0)
     require (!pipelined || depth > 0, "Register-based device with request/response handshaking needs concurrency > 0")
@@ -60,7 +69,7 @@ object RegMapper
     def ofBits(bits: List[Boolean]) = bits.foldRight(0){ case (x,y) => (if (x) 1 else 0) | y << 1 }
 
     // Find the minimal mask that can decide the register map
-    val mask = AddressDecoder(regmap.map(_._1))
+    val mask = AddressDecoder(wordmap.keySet.toList)
     val maskMatch = ~UInt(mask, width = inBits)
     val maskFilter = toBits(mask)
     val maskBits = maskFilter.filter(x => x).size
@@ -75,17 +84,22 @@ object RegMapper
     val iRightReg = Array.fill(regSize) { Bool(true) }
     val oRightReg = Array.fill(regSize) { Bool(true) }
 
-    // Flatten the regmap into (RegIndex:Int, Offset:Int, field:RegField)
+    // Transform the wordmap into minimal decoded indexes, Seq[(index, bit, field)]
-    val flat = regmap.map { case (reg, fields) =>
+    val flat = wordmap.toList.map { case (word, fields) =>
-      val offsets = fields.scanLeft(0)(_ + _.width).init
+      val index = regIndexI(word)
-      val index = regIndexI(reg)
+      val uint = UInt(word, width = inBits)
-      val uint = UInt(reg, width = inBits)
       if (undefZero) {
         iRightReg(index) = ((front.bits.index ^ uint) & maskMatch) === UInt(0)
         oRightReg(index) = ((back .bits.index ^ uint) & maskMatch) === UInt(0)
       }
-      // println("mapping 0x%x -> 0x%x for 0x%x/%d".format(reg, index, mask, maskBits))
+      // Confirm that no field spans a word boundary
-      (offsets zip fields) map { case (o, f) => (index, o, f) }
+      fields foreach { case (bit, field) =>
+        val off = bit - 8*bytes*word
+        // println(s"Reg ${word}: [${off}, ${off+field.width})")
+        require (off + field.width <= bytes * 8, s"Field at word ${word}*(${bytes}B) has bits [${off}, ${off+field.width}), which exceeds word limit.")
+      }
+      // println("mapping 0x%x -> 0x%x for 0x%x/%d".format(word, index, mask, maskBits))
+      fields.map { case (bit, field) => (index, bit-8*bytes*word, field) }
     }.flatten
 
     // Forward declaration of all flow control signals

src/main/scala/uncore/tilelink2/RegisterRouterTest.scala

@@ -183,14 +183,14 @@ object RRTest0Map
 
   // All fields must respect byte alignment, or else it won't behave like an SRAM
   val map = Seq(
-    0 -> Seq(aa(8), ar(8), ad(8), ae(8)),
+    0  -> Seq(aa(8), ar(8), ad(8), ae(8)),
-    1 -> Seq(ra(8), rr(8), rd(8), re(8)),
+    4  -> Seq(ra(8), rr(8), rd(8), re(8)),
-    2 -> Seq(da(8), dr(8), dd(8), de(8)),
+    8  -> Seq(da(8), dr(8), dd(8), de(8)),
-    3 -> Seq(ea(8), er(8), ed(8), ee(8)),
+    12 -> Seq(ea(8), er(8), ed(8), ee(8)),
-    4 -> Seq(aa(3), ar(5), ad(1), ae(7), ra(2), rr(6), rd(4), re(4)),
+    16 -> Seq(aa(3), ar(5), ad(1), ae(7), ra(2), rr(6), rd(4), re(4)),
-    5 -> Seq(da(3), dr(5), dd(1), de(7), ea(2), er(6), ed(4), ee(4)),
+    20 -> Seq(da(3), dr(5), dd(1), de(7), ea(2), er(6), ed(4), ee(4)),
-    6 -> Seq(aa(8), rr(8), dd(8), ee(8)),
+    24 -> Seq(aa(8), rr(8), dd(8), ee(8)),
-    7 -> Seq(ar(8), rd(8), de(8), ea(8)))
+    28 -> Seq(ar(8), rd(8), de(8), ea(8)))
 }
 
 object RRTest1Map
@@ -203,8 +203,8 @@ object RRTest1Map
   def bb(bits: Int) = request(bits, busy, busy)
 
   val map = RRTest0Map.map.take(6) ++ Seq(
-    6 -> Seq(pp(8), pb(8), bp(8), bb(8)),
+    24 -> Seq(pp(8), pb(8), bp(8), bb(8)),
-    7 -> Seq(pp(3), pb(5), bp(1), bb(7), pb(5), bp(3), pp(4), bb(4)))
+    28 -> Seq(pp(3), pb(5), bp(1), bb(7), pb(5), bp(3), pp(4), bb(4)))
 }
 
 trait RRTest0Bundle

src/main/scala/uncore/tilelink2/WidthWidget.scala

@@ -174,7 +174,7 @@ class TLWidthWidget(innerBeatBytes: Int) extends LazyModule
 object TLWidthWidget
 {
   // applied to the TL source node; y.node := WidthWidget(x.node, 16)
-  def apply(x: TLBaseNode, innerBeatBytes: Int)(implicit sourceInfo: SourceInfo): TLBaseNode = {
+  def apply(innerBeatBytes: Int)(x: TLBaseNode)(implicit sourceInfo: SourceInfo): TLBaseNode = {
     val widget = LazyModule(new TLWidthWidget(innerBeatBytes))
     widget.node := x
     widget.node