dispatch1: compute lsqNeedAlloc in rename for better timing

src/main/scala/xiangshan/backend/CtrlBlock.scala

@@ -292,6 +292,7 @@ class CtrlBlock extends XSModule with HasCircularQueuePtrHelper {
   rename.io.roqCommits <> roq.io.commits
   rename.io.out <> dispatch.io.fromRename
   rename.io.renameBypass <> dispatch.io.renameBypass
+  rename.io.dispatchInfo <> dispatch.io.preDpInfo
 
   dispatch.io.redirect <> backendRedirect
   dispatch.io.flush := flushReg

src/main/scala/xiangshan/backend/dispatch/Dispatch.scala

@@ -28,6 +28,7 @@ class Dispatch extends XSModule {
     // from rename
     val fromRename = Vec(RenameWidth, Flipped(DecoupledIO(new MicroOp)))
     val renameBypass = Input(new RenameBypassInfo)
+    val preDpInfo = Input(new PreDispatchInfo)
     // to busytable: set pdest to busy (not ready) when they are dispatched
     val allocPregs = Vec(RenameWidth, Output(new ReplayPregReq))
     // enq Roq
@@ -66,6 +67,7 @@ class Dispatch extends XSModule {
   // dispatch 1: accept uops from rename and dispatch them to the three dispatch queues
   // dispatch1.io.redirect <> io.redirect
   dispatch1.io.renameBypass := RegEnable(io.renameBypass, io.fromRename(0).valid && dispatch1.io.fromRename(0).ready)
+  dispatch1.io.preDpInfo := RegEnable(io.preDpInfo, io.fromRename(0).valid && dispatch1.io.fromRename(0).ready)
   dispatch1.io.enqRoq <> io.enqRoq
   dispatch1.io.enqLsq <> io.enqLsq
   dispatch1.io.toIntDq <> intDq.io.enq

src/main/scala/xiangshan/backend/dispatch/Dispatch1.scala

@@ -10,12 +10,18 @@ import xiangshan.backend.rename.RenameBypassInfo
 import xiangshan.mem.LsqEnqIO
 import xiangshan.backend.fu.HasExceptionNO
 
+
+class PreDispatchInfo extends XSBundle {
+  val lsqNeedAlloc = Vec(RenameWidth, UInt(2.W))
+}
+
 // read rob and enqueue
 class Dispatch1 extends XSModule with HasExceptionNO {
   val io = IO(new Bundle() {
     // from rename
     val fromRename = Vec(RenameWidth, Flipped(DecoupledIO(new MicroOp)))
     val renameBypass = Input(new RenameBypassInfo)
+    val preDpInfo = Input(new PreDispatchInfo)
     val recv = Output(Vec(RenameWidth, Bool()))
     // enq Roq
     val enqRoq = Flipped(new RoqEnqIO)
@@ -147,7 +153,7 @@ class Dispatch1 extends XSModule with HasExceptionNO {
     io.enqRoq.req(i).bits := updatedUop(i)
     XSDebug(io.enqRoq.req(i).valid, p"pc 0x${Hexadecimal(io.fromRename(i).bits.cf.pc)} receives nroq ${io.enqRoq.resp(i)}\n")
 
-    io.enqLsq.needAlloc(i) := io.fromRename(i).valid && isLs(i)
+    io.enqLsq.needAlloc(i) := Mux(io.fromRename(i).valid, io.preDpInfo.lsqNeedAlloc(i), 0.U)
     io.enqLsq.req(i).valid := io.fromRename(i).valid && isLs(i) && thisCanActualOut(i) && io.enqRoq.canAccept && io.toIntDq.canAccept && io.toFpDq.canAccept && io.toLsDq.canAccept
     io.enqLsq.req(i).bits := updatedUop(i)
     io.enqLsq.req(i).bits.roqIdx := io.enqRoq.resp(i)

src/main/scala/xiangshan/backend/rename/Rename.scala

@@ -5,6 +5,7 @@ import chisel3.util._
 import xiangshan._
 import utils._
 import xiangshan.backend.roq.RoqPtr
+import xiangshan.backend.dispatch.PreDispatchInfo
 
 class RenameBypassInfo extends XSBundle {
   val lsrc1_bypass = MixedVec(List.tabulate(RenameWidth-1)(i => UInt((i+1).W)))
@@ -23,6 +24,7 @@ class Rename extends XSModule with HasCircularQueuePtrHelper {
     // to dispatch1
     val out = Vec(RenameWidth, DecoupledIO(new MicroOp))
     val renameBypass = Output(new RenameBypassInfo)
+    val dispatchInfo = Output(new PreDispatchInfo)
   })
 
   def printRenameInfo(in: DecoupledIO[CfCtrl], out: DecoupledIO[MicroOp]) = {
@@ -202,6 +204,12 @@ class Rename extends XSModule with HasCircularQueuePtrHelper {
     }).reverse)
   }
 
+  val isLs    = VecInit(uops.map(uop => FuType.isLoadStore(uop.ctrl.fuType)))
+  val isStore = VecInit(uops.map(uop => FuType.isStoreExu(uop.ctrl.fuType)))
+  val isAMO   = VecInit(uops.map(uop => FuType.isAMO(uop.ctrl.fuType)))
+  io.dispatchInfo.lsqNeedAlloc := VecInit((0 until RenameWidth).map(i =>
+    Mux(isLs(i), Mux(isStore(i) && !isAMO(i), 2.U, 1.U), 0.U)))
+
   /**
     * Instructions commit: update freelist and rename table
     */

src/main/scala/xiangshan/mem/lsqueue/LSQWrapper.scala

@@ -29,7 +29,7 @@ class InflightBlockInfo extends XSBundle {
 
 class LsqEnqIO extends XSBundle {
   val canAccept = Output(Bool())
-  val needAlloc = Vec(RenameWidth, Input(Bool()))
+  val needAlloc = Vec(RenameWidth, Input(UInt(2.W)))
   val req = Vec(RenameWidth, Flipped(ValidIO(new MicroOp)))
   val resp = Vec(RenameWidth, Output(new LSIdx))
 }
@@ -75,15 +75,13 @@ class LsqWrappper extends XSModule with HasDCacheParameters {
   loadQueue.io.enq.sqCanAccept := storeQueue.io.enq.canAccept
   storeQueue.io.enq.lqCanAccept := loadQueue.io.enq.canAccept
   for (i <- 0 until RenameWidth) {
-    val isStore = CommitType.lsInstIsStore(io.enq.req(i).bits.ctrl.commitType)
-
-    loadQueue.io.enq.needAlloc(i) := io.enq.needAlloc(i) && !isStore
-    loadQueue.io.enq.req(i).valid  := !isStore && io.enq.req(i).valid
+    loadQueue.io.enq.needAlloc(i) := io.enq.needAlloc(i)(0)
+    loadQueue.io.enq.req(i).valid := io.enq.needAlloc(i)(0) && io.enq.req(i).valid
     loadQueue.io.enq.req(i).bits  := io.enq.req(i).bits
 
-    storeQueue.io.enq.needAlloc(i) := io.enq.needAlloc(i) && isStore
-    storeQueue.io.enq.req(i).valid :=  isStore && io.enq.req(i).valid
-    storeQueue.io.enq.req(i).bits := io.enq.req(i).bits
+    storeQueue.io.enq.needAlloc(i) := io.enq.needAlloc(i)(1)
+    storeQueue.io.enq.req(i).valid := io.enq.needAlloc(i)(1) && io.enq.req(i).valid
+    storeQueue.io.enq.req(i).bits  := io.enq.req(i).bits
 
     io.enq.resp(i).lqIdx := loadQueue.io.enq.resp(i)
     io.enq.resp(i).sqIdx := storeQueue.io.enq.resp(i)