WaitTable: use 2-bit counter and optimize XORFold logic (#681)

src/main/scala/utils/BitUtils.scala

@@ -133,8 +133,10 @@ object GetOddBits {
 }
 
 object XORFold {
-  def apply(input: UInt, reswidth: Int): UInt = {
-    require(input.getWidth*2>=reswidth)
-    input(reswidth-1, 0) ^ input(2*reswidth-1, reswidth)
+  def apply(input: UInt, resWidth: Int): UInt = {
+    require(resWidth > 0)
+    val fold_range = input.getWidth / resWidth
+    val value = ZeroExt(input, fold_range * resWidth)
+    ParallelXOR((0 until fold_range).map(i => value(i*resWidth+resWidth-1, i*resWidth)))
   }
-}
\ No newline at end of file
+}

src/main/scala/xiangshan/backend/decode/DecodeStage.scala

@@ -36,8 +36,8 @@ class DecodeStage extends XSModule {
   }
   waittable.io.csrCtrl <> io.csrCtrl
 
-  val loadWaitBitSet = PopCount(VecInit((0 until DecodeWidth).map(i => waittable.io.rdata(i) && io.out(i).fire())))
-  XSPerf("loadWaitBitSet", loadWaitBitSet, acc = true) // rollback redirect generated
+  val loadWaitBitSet = PopCount(io.out.map(o => o.fire() && o.bits.cf.loadWaitBit))
+  XSPerf("loadWaitBitSet", loadWaitBitSet)
 
   val hasValid = VecInit(io.in.map(_.valid)).asUInt.orR
   XSPerf("utilization", PopCount(io.in.map(_.valid)))

src/main/scala/xiangshan/backend/decode/WaitTable.scala

@@ -21,19 +21,19 @@ class WaitTable extends XSModule with WaitTableParameters {
     val csrCtrl = Input(new CustomCSRCtrlIO)
   })
 
-  val data = Reg(Vec(WaitTableSize, Bool())) // init val false.B
+  val data = RegInit(VecInit(Seq.fill(WaitTableSize)(0.U(2.W))))
   val resetCounter = RegInit(0.U(ResetTimeMax2Pow.W))
   resetCounter := resetCounter + 1.U
 
   // read ports
   for (i <- 0 until DecodeWidth) {
-    io.rdata(i) := (data(io.raddr(i)) || io.csrCtrl.no_spec_load) && !io.csrCtrl.lvpred_disable
+    io.rdata(i) := (data(io.raddr(i))(1) || io.csrCtrl.no_spec_load) && !io.csrCtrl.lvpred_disable
   }
 
   // write ports (with priority)
   (0 until StorePipelineWidth).map(i => {
     when(io.update(i).valid){
-      data(io.update(i).waddr) := io.update(i).wdata
+      data(io.update(i).waddr) := Cat(data(io.update(i).waddr)(0), true.B)
     }
   })
 
@@ -41,7 +41,7 @@ class WaitTable extends XSModule with WaitTableParameters {
   // reset period: ResetTimeMax2Pow
   when(resetCounter(ResetTimeMax2Pow-1, ResetTimeMin2Pow)(RegNext(io.csrCtrl.waittable_timeout))) {
     for (j <- 0 until WaitTableSize) {
-      data(j) := false.B
+      data(j) := 0.U
     }
     resetCounter:= 0.U
   }
@@ -53,5 +53,5 @@ class WaitTable extends XSModule with WaitTableParameters {
     }
   }
 
-  XSPerf("wait_table_bit_set", PopCount(data))
+  XSPerf("wait_table_bit_set", PopCount(data.map(d => d(1))))
 }