feat(npc): implement all riscv32i intructions (untested)

npc/core/src/main/resources/RamDpi.v

@@ -8,7 +8,9 @@ module RamDpi (
   input [31:0] writeData,
   input [3:0] writeMask,
   input reg [31:0] readAddr,
-  output reg [31:0] readData
+  output reg [31:0] readData,
+  input reg [31:0] pc,
+  output reg [31:0] inst
 );
   always @(*) begin
     if (valid) begin // 有读写请求时
@@ -20,5 +22,6 @@ module RamDpi (
     else begin
       readData = 0;
     end
+    inst = pmem_read(pc);
   end
 endmodule

npc/core/src/main/scala/components/ALU.scala

@@ -6,41 +6,50 @@ import shapeless.{HNil, ::}
 
 class ALUControlInterface extends Bundle {
   object OpSelect extends ChiselEnum {
-    val aOpAdd, aOpSub, aOpNot, aOpAnd, aOpOr, aOpXor, aOpSlt, aOpEq, aOpNop = Value
+    val aOpAdd, aOpSub, aOpNot, aOpAnd, aOpOr, aOpXor, aOpSlt, aOpSll, aOpSrl, aOpSra = Value
   }
-  object SrcSelect extends ChiselEnum {
+  object SrcASelect extends ChiselEnum {
-    val aSrcRs2, aSrcImm = Value
+    val aSrcARs1, aSrcAPc, aSrcAZero = Value
+  }
+  object SrcBSelect extends ChiselEnum {
+    val aSrcBRs2, aSrcBImmI, aSrcBImmJ, aSrcBImmB, aSrcBImmS = Value
   }
   val op = Input(OpSelect())
-  val src = Input(SrcSelect())
+  val srcASelect = Input(SrcASelect())
+  val srcBSelect = Input(SrcBSelect())
+  val signExt = Input(Bool())
 
-  type CtrlTypes = OpSelect.Type :: SrcSelect.Type :: HNil
+  def ctrlBindPorts = {
-  def ctrlBindPorts: CtrlTypes = {
+    op :: srcASelect :: srcBSelect :: signExt :: HNil
-    op :: src :: HNil
   }
 }
 
 class ALU[T <: UInt](tpe: T) extends Module {
   val control = IO(new ALUControlInterface)
   val in = IO(new Bundle {
-    val a = Input(Vec(control.SrcSelect.all.length, tpe))
+    val a = Input(Vec(control.SrcASelect.all.length, tpe))
-    val b = Input(tpe)
+    val b = Input(Vec(control.SrcBSelect.all.length, tpe))
   })
   val out = IO(new Bundle {
+    val eq = Output(Bool())
     val result = Output(tpe)
   })
 
-  val a = in.a(control.src.asUInt)
+  val a = in.a(control.srcASelect.asUInt)
+  val b = in.b(control.srcBSelect.asUInt)
 
   // val adder_b = (Fill(tpe.getWidth, io.op(0)) ^ io.b) + io.op(0)  // take (-b) if sub
-  val add = a + in.b
+  val add = a + b
-  val sub = a - in.b
+  val sub = a - b
-  val and = a & in.b
+  val and = a & b
   val not = ~a
-  val or = a | in.b
+  val or = a | b
-  val xor = a ^ in.b
+  val xor = a ^ b
-  val slt = a < in.b
+  val slt = a < b
-  val eq = a === in.b
+  val sll = a << b(log2Ceil(tpe.getWidth), 0)
+  val srl = a >> b(log2Ceil(tpe.getWidth), 0)
+  val sra = a.asSInt >> b(log2Ceil(tpe.getWidth), 0)
+  out.eq := a === b
 
   import control.OpSelect._
 
@@ -52,7 +61,9 @@ class ALU[T <: UInt](tpe: T) extends Module {
     aOpOr  -> or,
     aOpXor -> xor,
     aOpSlt -> slt,
-    aOpEq -> eq
+    aOpSll -> sll,
+    aOpSrl -> srl,
+    aOpSra -> sra.asUInt
   ))
 }
 

npc/core/src/main/scala/components/Mem.scala

@@ -6,19 +6,32 @@ import chisel3.util.HasBlackBoxResource
 import chisel3.util.log2Ceil
 import chisel3.experimental.noPrefix
 import scala.collection.SeqMap
-import javax.swing.plaf.synth.SynthRadioButtonMenuItemUI
+import flow.components
+import shapeless.{HNil, ::}
 
-class RamInterface[T <: Data](tpe: T, addrWidth: Int) extends Bundle {
+class RamControlInterface(addrWidth: Int) extends Bundle {
   val valid = Input(Bool())
+  val writeMask = Input(UInt((addrWidth / 8).W))
   val writeEnable = Input(Bool())
+  def ctrlBindPorts = {
+    valid :: writeMask :: writeEnable :: HNil
+  }
+
+}
+
+/* FIXME: Extends here might not be the best solution.
+ *        We need a way to merge two bundles together
+ */
+class RamInterface[T <: Data](tpe: T, addrWidth: Int) extends RamControlInterface(addrWidth) {
   val writeAddr = Input(UInt(addrWidth.W))
   val writeData = Input(tpe)
-  val writeMask = Input(UInt((addrWidth / 8).W))
   val readAddr = Input(UInt(addrWidth.W))
   val readData = Output(tpe)
+  val pc = Input(UInt(addrWidth.W))
+  val inst = Output(tpe)
 }
 
 class RamDpi extends BlackBox with HasBlackBoxResource {
-  val io = IO(new RamInterface(UInt(32.W), 32))
+  val io = IO((new RamInterface(UInt(32.W), 32)))
   addResource("/RamDpi.v")
 }

npc/core/src/main/scala/components/ProgramCounter.scala

@@ -6,33 +6,42 @@ import shapeless.{HNil, ::}
 
 class PcControlInterface extends Bundle {
   object SrcSelect extends ChiselEnum {
-    val pStaticNpc, pBranchResult = Value
+    val pStaticNpc, pExeOut = Value
   }
 
+  val useImmB = Input(Bool())
   val srcSelect = Input(SrcSelect())
 
-  type CtrlTypes = SrcSelect.Type :: HNil
+  def ctrlBindPorts = {
-  def ctrlBindPorts: CtrlTypes = {
+    useImmB :: srcSelect :: HNil
-    srcSelect :: HNil
   }
 }
 
-class ProgramCounter[T <: Data](tpe: T) extends Module {
+class ProgramCounter[T <: UInt](tpe: T) extends Module {
 
   val control = IO(new PcControlInterface)
   val in = IO(new Bundle {
-    val pcSrcs = Input(Vec(control.SrcSelect.all.length, tpe))
+    val immB = Input(tpe)
+    val exeOut = Input(tpe)
   })
   val out = IO(Output(tpe))
 
-  private val pc = RegInit(0x80000000L.U)
+  private val pc_reg = RegInit(0x80000000L.U)
 
-  pc := in.pcSrcs(control.srcSelect.asUInt)
+//   pc := in.pcSrcs(control.srcSelect.asUInt)
-  out := pc
+  import control.SrcSelect._
+  when( control.useImmB === true.B ) {
+    pc_reg := pc_reg + in.immB
+  }. elsewhen( control.srcSelect === pStaticNpc) {
+    pc_reg := pc_reg + 4.U
+  }. elsewhen( control.srcSelect === pExeOut) {
+    pc_reg := in.exeOut
+  }
+  out := pc_reg
 }
 
 object ProgramCounter {
-  def apply[T <: Data](tpe: T): ProgramCounter[T] = {
+  def apply[T <: UInt](tpe: T): ProgramCounter[T] = {
     val pc = Module(new ProgramCounter(tpe))
     pc
   }

npc/core/src/main/scala/components/RegisterFile.scala

@@ -5,18 +5,17 @@ import chisel3.util.log2Ceil
 import chisel3.util.UIntToOH
 import chisel3.util.MuxLookup
 import chisel3.experimental.Trace._
-import shapeless.{ HNil, :: }
+import shapeless.{ HList, HNil, :: }
 
 class RegControl extends Bundle {
   object WriteSelect extends ChiselEnum {
-    val rAluOut, rMemOut = Value
+    val rAluOut, rMemOut, rNpc = Value
   }
 
   val writeEnable = Input(Bool())
   val writeSelect = Input(WriteSelect())
 
-  type CtrlTypes = Bool :: WriteSelect.Type :: HNil
+  def ctrlBindPorts = {
-  def ctrlBindPorts: CtrlTypes = {
     writeEnable :: writeSelect :: HNil
   }
   traceName(writeEnable)
@@ -40,7 +39,8 @@ class RegisterFile[T <: Data](tpe: T, regCount: Int, numReadPorts: Int) extends
   val writeAddrOH = UIntToOH(in.writeAddr)
 
   for ((reg, i) <- regFile.zipWithIndex.tail) {
-    reg := Mux(writeAddrOH(i.U(log2Ceil(regCount).W)) && control.writeEnable, in.writeData(control.writeSelect.asUInt), reg)
+    reg := Mux(
+      writeAddrOH(i.U(log2Ceil(regCount).W)) && control.writeEnable, in.writeData(control.writeSelect.asUInt), reg)
   }
   regFile(0) := 0.U
 

npc/core/src/main/scala/top/FlowMain.scala

@@ -2,12 +2,8 @@ package flow
 
 import scala.reflect.runtime.universe._
 import chisel3._
-import chisel3.util.{MuxLookup, Fill, Decoupled, Counter, Queue, Reverse}
-import chisel3.util.{SRAM}
 import chisel3.util.experimental.decode.{decoder, TruthTable}
-import chisel3.util.log2Ceil
+import chisel3.util._
-import chisel3.util.BitPat
-import chisel3.util.Enum
 import chisel3.experimental.Trace._
 import shapeless.{HNil, ::}
 import shapeless.HList
@@ -15,68 +11,322 @@ import shapeless.ops.coproduct.Prepend
 import chisel3.util.{ BinaryMemoryFile, HexMemoryFile }
 
 import chisel3.experimental.Trace
+import scala.collection.IndexedSeqView
+import shapeless.Poly1
+import flow.components.RamControlInterface
 
 object RV32Inst {
   private val bp = BitPat
+
+  val lui    = this.bp("b???????_?????_?????_???_?????_01101_11")
+  val auipc  = this.bp("b???????_?????_?????_???_?????_00101_11")
+
+  val jal    = this.bp("b???????_?????_?????_???_?????_11011_11")
+  val jalr   = this.bp("b???????_?????_?????_???_?????_11001_11")
+  val beq    = this.bp("b???????_?????_?????_000_?????_11000_11")
+  val bne    = this.bp("b???????_?????_?????_001_?????_11000_11")
+  val blt    = this.bp("b???????_?????_?????_100_?????_11000_11")
+  val bge    = this.bp("b???????_?????_?????_101_?????_11000_11")
+  val bltu   = this.bp("b???????_?????_?????_110_?????_11000_11")
+  val bgeu   = this.bp("b???????_?????_?????_111_?????_11000_11")
+
+  val lb     = this.bp("b???????_?????_?????_000_?????_00000_11")
+  val lh     = this.bp("b???????_?????_?????_001_?????_00000_11")
+  val lw     = this.bp("b???????_?????_?????_010_?????_00000_11")
+  val lbu    = this.bp("b???????_?????_?????_100_?????_00000_11")
+  val lhu    = this.bp("b???????_?????_?????_101_?????_00000_11")
+  val sb     = this.bp("b???????_?????_?????_000_?????_01000_11")
+  val sh     = this.bp("b???????_?????_?????_001_?????_01000_11")
+  val sw     = this.bp("b???????_?????_?????_010_?????_01000_11")
+
   val addi   = this.bp("b???????_?????_?????_000_?????_00100_11")
+  val slti   = this.bp("b???????_?????_?????_010_?????_00100_11")
+  val sltiu  = this.bp("b???????_?????_?????_011_?????_00100_11")
+  val xori   = this.bp("b???????_?????_?????_100_?????_00100_11")
+  val ori    = this.bp("b???????_?????_?????_110_?????_00100_11")
+  val andi   = this.bp("b???????_?????_?????_111_?????_00100_11")
+  val slli   = this.bp("b0000000_?????_?????_001_?????_00100_11")
+  val srli   = this.bp("b0000000_?????_?????_101_?????_00100_11")
+  val srai   = this.bp("b0100000_?????_?????_101_?????_00100_11")
+  val add    = this.bp("b0000000_?????_?????_000_?????_01100_11")
+  val sub    = this.bp("b0100000_?????_?????_000_?????_01100_11")
+  val sll    = this.bp("b0000000_?????_?????_001_?????_01100_11")
+  val slt    = this.bp("b0000000_?????_?????_010_?????_01100_11")
+  val sltu   = this.bp("b0000000_?????_?????_011_?????_01100_11")
+  val xor    = this.bp("b0000000_?????_?????_100_?????_01100_11")
+  val srl    = this.bp("b0000000_?????_?????_101_?????_01100_11")
+  val sra    = this.bp("b0100000_?????_?????_101_?????_01100_11")
+  val or     = this.bp("b0000000_?????_?????_110_?????_01100_11")
+  val and    = this.bp("b0000000_?????_?????_111_?????_01100_11")
+
+  val ebreak = this.bp("b0000000_00001_00000_000_00000_11100_11")
+
+
+  val mul    = this.bp("b0000001_?????_?????_000_?????_01100_11")
+  val mulh   = this.bp("b0000001_?????_?????_001_?????_01100_11")
+  val mulhsu = this.bp("b0000001_?????_?????_010_?????_01100_11")
+  val mulhu  = this.bp("b0000001_?????_?????_011_?????_01100_11")
+  val div    = this.bp("b0000001_?????_?????_100_?????_01100_11")
+  val divu   = this.bp("b0000001_?????_?????_101_?????_01100_11")
+  val rem    = this.bp("b0000001_?????_?????_110_?????_01100_11")
+  val remu   = this.bp("b0000001_?????_?????_111_?????_01100_11")
+
   val inv    = this.bp("b???????_?????_?????_???_?????_?????_??")
 }
 
-class PcControl(width: Int) extends Bundle {
-  object SrcSelect extends ChiselEnum {
-    val pPC, pExeResult = Value
-  }
-  val srcSelect = Output(SrcSelect())
-}
-
 import flow.components.{RegControl, PcControlInterface, ALUControlInterface}
-class Control(width: Int) extends Module {
+class Control(width: Int) extends RawModule {
+  // Helpers
+  class WrapList[T](vl: T) { type Type = T; val v = vl}
+  object wrap extends Poly1 {
+    implicit def default[A] = at[A](Right(_).withLeft[Int])
+  }
+  def lit(x: Element) = { x.litValue.toInt }
+  def toBits(t: dst.Type): BitPat = {
+    val list = t.toList
+    list.map(e => e match {
+      case Right(x) => BitPat(lit(x).U(x.getWidth.W))
+      case Left(x) => BitPat.dontCare(x)
+    }).reduceLeft(_ ## _)
+  }
+  val r = Right
+  def l[T <: Any](x: T) = x match {
+    case x: ChiselEnum => Left(log2Ceil(x.all.length))
+    case x: Data => Left(x.getWidth)
+    case _ => throw new IllegalArgumentException
+  }
+
   val inst = IO(Input(UInt(width.W)))
 
   val reg = IO(Flipped(new RegControl))
   val pc = IO(Flipped(new PcControlInterface))
   val alu = IO(Flipped(new ALUControlInterface))
+  val ram = IO(Flipped(new RamControlInterface(32)))
 
-  // TODO: Add .ctrlTypes together instead of writing them by hand.
+  val dst = new WrapList((
-  type T =
+    reg.ctrlBindPorts ++
-    Bool :: reg.WriteSelect.Type :: pc.SrcSelect.Type :: alu.OpSelect.Type :: alu.SrcSelect.Type :: HNil
+    pc.ctrlBindPorts ++
-  val dst: T = reg.ctrlBindPorts ++ pc.ctrlBindPorts ++ alu.ctrlBindPorts
+    alu.ctrlBindPorts ++
+    ram.ctrlBindPorts).map(wrap))
 
-  val dstList = dst.toList
+  val dstList = dst.v.toList
-  val reversePrefixSum = dstList.scanLeft(0)(_ + _.getWidth).reverse
+  val controlWidth = dstList.map(_.toOption.get.getWidth).reduce(_ + _)
-  val slices = reversePrefixSum.zip(reversePrefixSum.tail)
+  val reversePrefixSum = dstList.scanLeft(0)(_ + _.toOption.get.getWidth)
+  val sliceIndex = reversePrefixSum.map(controlWidth - _)
+  val slices = sliceIndex.map(_ - 1).zip(sliceIndex.tail)
 
   import reg.WriteSelect._
+  import reg._
   import pc.SrcSelect._
+  import pc._
   import alu.OpSelect._
-  import alu.SrcSelect._
+  import alu.SrcASelect._
+  import alu.SrcBSelect._
+  import pc._
   import RV32Inst._
-  val ControlMapping: Array[(BitPat, T)] = Array(
+  val ControlMapping: Array[(BitPat, dst.Type)] = Array(
-    //     Regs                       :: PC         :: Exe
+    //     Regs | writeEnable :: writeSelect :: HNil
-    //     writeEnable :: writeSelect :: srcSelect  ::
+    //     PC   | useImmB :: srcSelect :: HNil
-    (addi, true.B      :: rAluOut     :: pStaticNpc :: aOpAdd :: aSrcImm :: HNil),
+    //     Exe  | op :: srcASelect :: srcBSelect :: signExt :: HNil
-  )
+    //     Mem  | valid :: writeMask :: writeEnable :: HNil
-  val default = BitPat.dontCare(dstList.map(_.getWidth).reduce(_ + _))
 
-  def toBits(t: T): BitPat = {
+    // ---- Control Transfer Instructions ----
-    val list: List[Data] = t.toList
+    (jal   , (r(true.B)     :: r(rNpc)     ::
-    list.map(x => BitPat(x.litValue.toInt.U(x.getWidth.W))).reduceLeft(_ ## _)
+              r(false.B)    :: r(pExeOut)  ::
-  }
+              r(aOpAdd)     :: r(aSrcAPc)  :: r(aSrcBImmJ) :: r(false.B) ::
+              r(false.B)    :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (jalr  , (r(true.B)     :: r(rNpc)     ::
+              r(false.B)    :: r(pExeOut)    ::
+              r(aOpAdd)     :: r(aSrcARs1) :: r(aSrcBImmJ) :: r(false.B) ::
+              r(false.B)    :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (beq   , (r(false.B)    :: l(WriteSelect) ::
+              r(true.B)     :: r(pExeOut)     ::
+              r(aOpSlt)     :: r(aSrcAPc)  :: r(aSrcBImmB) :: l(Bool())    ::
+              r(false.B)    :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (bne   , (r(false.B)    :: l(WriteSelect) ::
+              r(true.B)     :: r(pExeOut)     ::
+              r(aOpSlt)     :: r(aSrcAPc)  :: r(aSrcBImmB) :: l(Bool())    ::
+              r(false.B)    :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (blt   , (r(false.B)    :: l(WriteSelect) ::
+              r(true.B)     :: r(pExeOut)     ::
+              r(aOpSlt)     :: r(aSrcAPc)  :: r(aSrcBImmB) :: r(true.B)  ::
+              r(false.B)    :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (bge   , (r(false.B)    :: l(WriteSelect) ::
+              r(true.B)     :: r(pExeOut)     ::
+              r(aOpSlt)     :: r(aSrcAPc)  :: r(aSrcBImmB) :: r(true.B)  ::
+              r(false.B)    :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (bltu  , (r(false.B)    :: l(WriteSelect)::
+              r(true.B)     :: r(pExeOut)    ::
+              r(aOpSlt)     :: r(aSrcAPc)    :: r(aSrcBImmB) :: r(false.B) ::
+              r(false.B)    :: l(UInt(4.W))  :: r(false.B)   :: HNil)),
+
+    (bgeu  , (r(false.B)    :: l(WriteSelect)::
+              r(true.B)     :: r(pExeOut)    ::
+              r(aOpSlt)     :: r(aSrcAPc)    :: r(aSrcBImmB) :: r(false.B) ::
+              r(false.B)    :: l(UInt(4.W))  :: r(false.B)   :: HNil)),
+
+    // ---- Memory Access Instructions ----
+
+    (lb    , (r(true.B)     :: r(rMemOut)    ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAdd)     :: r(aSrcARs1)   :: r(aSrcBImmI) :: l(Bool()) ::
+              r(true.B)     :: l(UInt(4.W))  :: r(false.B)   :: HNil)),
+
+    (lh    , (r(true.B)     :: r(rMemOut)    ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAdd)     :: r(aSrcARs1)   :: r(aSrcBImmI) :: l(Bool()) ::
+              r(true.B)     :: l(UInt(4.W))  :: r(false.B)   :: HNil)),
+
+    (lw    , (r(true.B)     :: r(rMemOut)    ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAdd)     :: r(aSrcARs1)   :: r(aSrcBImmI) :: l(Bool()) ::
+              r(true.B)     :: l(UInt(4.W))  :: r(false.B)   :: HNil)),
+
+    (sb    , (r(false.B)    :: l(WriteSelect)::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAdd)     :: r(aSrcARs1)   :: r(aSrcBImmI) :: l(Bool()) ::
+              r(true.B)     :: r(1.U(4.W))   :: r(false.B)   :: HNil)),
+
+    (sh    , (r(false.B)    :: l(WriteSelect)::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAdd)     :: r(aSrcARs1)   :: r(aSrcBImmI) :: l(Bool()) ::
+              r(true.B)     :: r(3.U(4.W))   :: r(false.B)   :: HNil)),
+
+    (sw    , (r(false.B)    :: l(WriteSelect)::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAdd)     :: r(aSrcARs1)   :: r(aSrcBImmI) :: l(Bool()) ::
+              r(true.B)     :: r(15.U(4.W))  :: r(false.B)   :: HNil)),
+
+    // ---- Integer Computational Instructions ---
+
+    (addi  , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAdd)     :: r(aSrcARs1) :: r(aSrcBImmI) :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (slti  , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSlt)     :: r(aSrcARs1) :: r(aSrcBImmI) :: r(true.B)  ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (sltiu , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSlt)     :: r(aSrcARs1) :: r(aSrcBImmI) :: r(false.B) ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (xori  , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpOr)      :: r(aSrcARs1) :: r(aSrcBImmI) :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (ori   , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpXor)     :: r(aSrcARs1) :: r(aSrcBImmI) :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (andi  , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAnd)     :: r(aSrcARs1) :: r(aSrcBImmI) :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (slli  , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSll)     :: r(aSrcARs1) :: r(aSrcBImmI) :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (srli  , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSrl)     :: r(aSrcARs1) :: r(aSrcBImmI) :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (srai  , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSra)     :: r(aSrcARs1) :: r(aSrcBImmI) :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (add   , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAdd)     :: r(aSrcARs1) :: r(aSrcBRs2)  :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (sub   , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSub)     :: r(aSrcARs1) :: r(aSrcBRs2)  :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (sll   , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSll)     :: r(aSrcARs1) :: r(aSrcBRs2)  :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (slt   , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSlt)     :: r(aSrcARs1) :: r(aSrcBRs2)  :: r(true.B)  ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (sltu  , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSlt)     :: r(aSrcARs1) :: r(aSrcBRs2)  :: r(false.B) ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (xor   , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpXor)     :: r(aSrcARs1) :: r(aSrcBRs2)  :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (srl   , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSrl)     :: r(aSrcARs1) :: r(aSrcBRs2)  :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (sra   , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpSra)     :: r(aSrcARs1) :: r(aSrcBRs2)  :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (or    , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpOr)      :: r(aSrcARs1) :: r(aSrcBRs2)  :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+
+    (and   , (r(true.B)     :: r(rAluOut)  ::
+              r(false.B)    :: r(pStaticNpc) ::
+              r(aOpAnd)     :: r(aSrcARs1) :: r(aSrcBRs2)  :: l(Bool())    ::
+              r(false.B)     :: l(UInt(4.W)):: r(false.B)   :: HNil)),
+  )
+
+  val default = BitPat(0.U(controlWidth.W))
+
+//   println(s"ControlMapping = ${ControlMapping.map(it => (it._1 -> toBits(it._2))).foreach(x => println(x._2))}\n")
   val out = decoder(
     inst,
     TruthTable(ControlMapping.map(it => (it._1 -> toBits(it._2))), default))
-  val srcList = slices.map(s => out(s._1 - 1, s._2))
+  val srcList = slices.map(s => out(s._1, s._2))
 
+  assert(out != default)
+  println(s"out = $out, default = $default\n")
+  println(s"dstList = ${dstList}\n")
+  println(s"srcList = ${srcList}\n")
   srcList
-    .zip(dstList.reverse)
+    .zip(dstList)
     .foreach({ case (src, dst) =>
-      dst := src.asTypeOf(dst)
+      dst.toOption.get := src.asTypeOf(dst.toOption.get)
     })
+
+  pc.useImmB := DontCare
 }
 
 import flow.components.{RegisterFile, ProgramCounter, ALU, RamDpi}
 import chisel3.util.experimental.loadMemoryFromFileInline
 class Flow extends Module {
+  def lit(x: Data) = { x.litValue.toInt }
+
   val dataType = UInt(32.W)
   val ram = Module(new RamDpi)
   val control = Module(new Control(32))
@@ -84,42 +334,61 @@ class Flow extends Module {
   val pc = Module(new ProgramCounter(dataType))
   val alu = Module(new ALU(dataType))
 
-  ram.io.readAddr := pc.out
+  // TODO: Switch to Decoupled and Arbiter later
-  val inst = ram.io.readData
+  ram.io.pc := pc.out
+  val inst = ram.io.inst
 
   dontTouch(reg.control.writeEnable)
 
   import control.pc.SrcSelect._
 
-  pc.in.pcSrcs(pStaticNpc.litValue.toInt) := pc.out + 4.U
+  val npc = Wire(dataType)
-  pc.in.pcSrcs(pBranchResult.litValue.toInt) := alu.out.result
+  npc := pc.out + 4.U
+  pc.in.exeOut := alu.out.result
+  pc.in.immB := Cat(Fill(20, inst(31)), inst(7), inst(30, 25), inst(11, 8), inst(0))
 
   control.inst := inst
   reg.control <> control.reg
+  // FIXME: Probably optimizable with bulk connection
   pc.control <> control.pc
   alu.control <> control.alu
+  val branchUseSlt = Wire(Bool())
+  val branchInvertResult = Wire(Bool())
+  branchUseSlt := inst(14)
+  branchInvertResult := inst(12)
+  val _branchResult = Mux(branchUseSlt, alu.out.result(0), alu.out.eq)
+  val branchResult = Mux(branchInvertResult, _branchResult, !_branchResult)
+  pc.control.useImmB := control.pc.useImmB && _branchResult
 
   import control.reg.WriteSelect._
-  reg.in.writeData(rAluOut.litValue.toInt) := alu.out.result
+  reg.in.writeData(lit(rAluOut)) := alu.out.result
-
   // TODO: Read address in load command goes here
-  reg.in.writeData(rMemOut.litValue.toInt) := DontCare
+  reg.in.writeData(lit(rMemOut)) := ram.io.readData
+  reg.in.writeData(lit(rNpc)) := npc
 
   reg.in.writeAddr := inst(11, 7)
   reg.in.rs(0) := inst(19, 15) // rs1
   reg.in.rs(1) := inst(24, 20) // rs2
 
   // TODO: Memory write goes here
-  ram.io.writeAddr := DontCare
+  ram.io.writeAddr := alu.out.result
-  ram.io.writeData := DontCare
+  ram.io.writeData := reg.out.src(1)
-  ram.io.writeMask := DontCare
+  ram.io.writeMask := control.ram.writeMask
-  ram.io.writeEnable := false.B
+  ram.io.writeEnable := control.ram.writeEnable
   ram.io.valid := true.B
+  ram.io.readAddr := alu.out.result
 
-  import control.alu.SrcSelect._
+  import control.alu.SrcASelect._
-  alu.in.a(aSrcRs2.litValue.toInt) := reg.out.src(1)
+  import control.alu.SrcBSelect._
-  alu.in.a(aSrcImm.litValue.toInt) := inst(31, 20)
+  alu.in.a(lit(aSrcARs1)) := reg.out.src(0)
-  alu.in.b := reg.out.src(0)
+  alu.in.a(lit(aSrcAPc)) := pc.out
+  alu.in.a(lit(aSrcAZero)) := 0.U
+
+  alu.in.b(lit(aSrcBRs2)) := reg.out.src(1)
+  alu.in.b(lit(aSrcBImmI)) := inst(31, 20).pad(aSrcBImmI.getWidth)
+  alu.in.b(lit(aSrcBImmJ)) := Cat(Fill(12, inst(31)), inst(19, 12), inst(20), inst(30, 25), inst(24, 21), 0.U(1.W))
+  alu.in.b(lit(aSrcBImmB)) := Cat(Fill(20, inst(31)), inst(7), inst(30, 25), inst(11, 8), inst(0))
+  alu.in.b(lit(aSrcBImmS)) := Cat(inst(31), inst(19, 12), inst(20), inst(30, 25), inst(24, 21), 0.U(1.W)).pad(aSrcBImmS.getWidth)
 
   Trace.traceName(pc.out);
   dontTouch(control.out)

npc/core/src/main/scala/top/Main.scala

@@ -34,7 +34,7 @@ object VerilogMain extends App {
   }
 
   val annos = (new ChiselStage).execute(
-    Array("--target-dir", opt.targetDir.toString, "--target", "systemverilog", "--split-verilog"),
+    Array("--target-dir", opt.targetDir.toString, "--target", "systemverilog", "--split-verilog", "--full-stacktrace"),
     Seq(
 
     ) ++ (if(config.traceConfig.enable) Seq(ChiselGeneratorAnnotation(() => new Flow)) else Seq())