Don't double-count release traffic in perfctrs

src/main/scala/rocket/DCache.scala

@@ -505,6 +505,6 @@ class DCacheModule(outer: DCache) extends HellaCacheModule(outer) {
   }
 
   // performance events
-  io.cpu.acquire := tl_out.a.fire()
+  io.cpu.acquire := edge.last(tl_out.a)
-  io.cpu.release := tl_out.c.fire()
+  io.cpu.release := edge.last(tl_out.c)
 }

src/main/scala/rocket/Frontend.scala

@@ -155,7 +155,7 @@ class FrontendModule(outer: Frontend) extends LazyModuleImp(outer)
   io.cpu.resp.bits.btb.bits := s2_btb_resp_bits
 
   // performance events
-  io.cpu.acquire := icache.io.mem(0).a.fire()
+  io.cpu.acquire := edge.last(icache.io.mem(0).a)
 }
 
 /** Mix-ins for constructing tiles that have an ICache-based pipeline frontend */

src/main/scala/rocket/NBDcache.scala

@@ -975,6 +975,6 @@ class NonBlockingDCacheModule(outer: NonBlockingDCache) extends HellaCacheModule
   io.cpu.replay_next := (s1_replay && s1_read) || mshrs.io.replay_next
 
   // performance events
-  io.cpu.acquire := tl_out.a.fire()
+  io.cpu.acquire := edge.last(tl_out.a)
-  io.cpu.release := tl_out.c.fire()
+  io.cpu.release := edge.last(tl_out.c)
 }