Merge pull request #147 from ngc7331/fix-btbhit

fix(BPU): btbHit is incorrectly invalidated causing the correct PHT result to be invalidated

src/main/scala/nutcore/frontend/BPU.scala

@@ -102,7 +102,7 @@ class BPU_ooo extends NutCoreModule {
   // we should latch the input pc for one cycle
   val pcLatch = RegEnable(io.in.pc.bits, io.in.pc.valid)
   val btbHit = Wire(Vec(4, Bool()))
-  (0 to 3).map(i => btbHit(i) := btbRead(i).valid && btbRead(i).tag === btbAddr.getTag(pcLatch) && !flush && RegNext(btb(i).io.r.req.fire, init = false.B))
+  (0 to 3).map(i => btbHit(i) := btbRead(i).valid && btbRead(i).tag === btbAddr.getTag(pcLatch) && !flush && RegNext(btb(i).io.r.req.ready, init = false.B))
   // btbHit will ignore pc(2,0). pc(2,0) is used to build brIdx
   val crosslineJump = btbRead(3).crosslineJump && btbHit(3) && !io.brIdx(0) && !io.brIdx(1) && !io.brIdx(2)
   io.crosslineJump := crosslineJump
@@ -317,7 +317,10 @@ class BPU_inorder extends NutCoreModule {
   // since there is one cycle latency to read SyncReadMem,
   // we should latch the input pc for one cycle
   val pcLatch = RegEnable(io.in.pc.bits, io.in.pc.valid)
-  val btbHit = btbRead.valid && btbRead.tag === btbAddr.getTag(pcLatch) && !flush && RegNext(btb.io.r.req.fire, init = false.B) && !(pcLatch(1) && btbRead.brIdx(0))
+  val btbHit = btbRead.valid && btbRead.tag === btbAddr.getTag(pcLatch) && !flush && RegNext(btb.io.r.req.ready, init = false.B) && !(pcLatch(1) && btbRead.brIdx(0))
+  // btb.io.r.req.ready is used to indicate whether BTB SRAM is doing reset (if is true.B, btbRead.valid can be fake)
+  // we don't use btb.io.r.req.fire because of btb.io.r.req.valid only last for 1 cycle, further causes fake BTB miss
+  // See https://github.com/OSCPU/NutShell/pull/147
   // btbHit will ignore pc(1,0). pc(1,0) is used to build brIdx
   // !(pcLatch(1) && btbRead.brIdx(0)) is used to deal with the following case:
   // -------------------------------------------------