MSHR: fix bug when SnpOnce hits on a TRUNK block (#187)

When a SnpOnce hits on a TRUNK block, CoupledL2 will probe upstream
cache to acquire the most up-to-date block. When DCache in XiangShan
receives Probe toT on a dirty cache line, DCache will echo ProbeAckData
TtoT to L2, and erase the dirty flag. Therefore either DCache reserves
the dirty flag or L2 marks the block dirty on ProbeAckData TtoT to make
sure the dirty data will not be lost. This commit adopts the latter
solution.

src/main/scala/coupledL2/tl2chi/MSHR.scala

@@ -491,7 +491,8 @@ class MSHR(implicit p: Parameters) extends TL2CHIL2Module {
         * 3. If the snoop is SnpCleanShared
         * Otherwise, the dirty bit should stay the same as before.
         */
-      dirty = !snpToN && !snpToB && req_chiOpcode =/= SnpCleanShared && meta.dirty,
+      dirty = !snpToN && !snpToB && req_chiOpcode =/= SnpCleanShared && meta.dirty ||
+        isSnpOnceX(req_chiOpcode) && probeDirty,
       state = Mux(
         snpToN,
         INVALID,