fix WAW hazard handling

rocket/src/main/scala/cpu.scala

@@ -103,6 +103,7 @@ class rocketProc extends Component
   ctrl.io.dmem.resp_miss  := arb.io.cpu.resp_miss;
   ctrl.io.dmem.resp_nack  := arb.io.cpu.resp_nack;
   dpath.io.dmem.resp_val  := arb.io.cpu.resp_val;
+  dpath.io.dmem.resp_replay := io.dmem.resp_replay;
   dpath.io.dmem.resp_tag  := arb.io.cpu.resp_tag;
   dpath.io.dmem.resp_data := arb.io.cpu.resp_data;  
   dpath.io.dmem.resp_data_subword := io.dmem.resp_data_subword;

rocket/src/main/scala/ctrl.scala

@@ -302,10 +302,7 @@ class rocketCtrl extends Component
 
   val id_raddr2 = io.dpath.inst(21,17);
   val id_raddr1 = io.dpath.inst(26,22);
-  val id_waddr  = io.dpath.inst(31,27);
+  val id_waddr  = Mux(id_sel_wa === WA_RA, RA, io.dpath.inst(31,27));
-
-  val id_ren2 = id_renx2.toBool;
-  val id_ren1 = id_renx1.toBool;
 
   val id_console_out_val  = id_wen_pcr.toBool && (id_raddr2 === PCR_CONSOLE);
 
@@ -329,7 +326,6 @@ class rocketCtrl extends Component
   val id_stall_raddr2 = sboard.io.stalla;
   val id_stall_raddr1 = sboard.io.stallb;
   val id_stall_waddr  = sboard.io.stallc;
-  val id_stall_ra     = sboard.io.stallra;
 
   val id_reg_btb_hit      = Reg(resetVal = Bool(false));
   val id_reg_xcpt_itlb    = Reg(resetVal = Bool(false));
@@ -581,62 +577,50 @@ class rocketCtrl extends Component
       io.dpath.stalld
     );
 
-  // check for loads and amos in execute and mem stages to detect load/use hazards
+  // stall for RAW/WAW hazards on loads, AMOs, and mul/div in execute stage.
   val ex_mem_cmd_load = 
     ex_reg_mem_val && ((ex_reg_mem_cmd === M_XRD) || ex_reg_mem_cmd(3).toBool);
-  
+  val data_hazard_ex =
-  val lu_stall_ex = 
+    (ex_mem_cmd_load || ex_reg_div_mul_val) &&
-    ex_mem_cmd_load &&
+    ((id_renx1.toBool && (id_raddr1 === io.dpath.ex_waddr)) ||
-    ((id_ren1 && (id_raddr1 === io.dpath.ex_waddr)) ||
+     (id_renx2.toBool && (id_raddr2 === io.dpath.ex_waddr)) ||
-     (id_ren2 && (id_raddr2 === io.dpath.ex_waddr)));
+     (id_wen.toBool   && (id_waddr  === io.dpath.ex_waddr)));
     
+  // stall for RAW/WAW hazards on LB/LH and mul/div in memory stage.
+  // stall for WAW-but-not-RAW hazards on LW/LD/AMO.
+  val mem_mem_cmd_load =
+    mem_reg_mem_val && ((mem_reg_mem_cmd === M_XRD) || mem_reg_mem_cmd(3).toBool);
   val mem_mem_cmd_load_bh = 
-    mem_reg_mem_val &&
+    mem_mem_cmd_load &&
-    (mem_reg_mem_cmd === M_XRD)   &&
     ((mem_reg_mem_type === MT_B)  ||
      (mem_reg_mem_type === MT_BU) ||
      (mem_reg_mem_type === MT_H)  || 
      (mem_reg_mem_type === MT_HU));
-     
+  val raw_hazard_mem =
-  val lu_stall_mem = 
+    (id_renx1.toBool && (id_raddr1 === io.dpath.mem_waddr)) ||
-    mem_mem_cmd_load_bh &&
+    (id_renx2.toBool && (id_raddr2 === io.dpath.mem_waddr));
-    ((id_ren1 && (id_raddr1 === io.dpath.mem_waddr)) ||
+  val waw_hazard_mem =
-     (id_ren2 && (id_raddr2 === io.dpath.mem_waddr)));
+    (id_wen.toBool   && (id_waddr  === io.dpath.mem_waddr));
+  val data_hazard_mem =
+    (mem_mem_cmd_load_bh || mem_reg_div_mul_val) && (raw_hazard_mem || waw_hazard_mem) ||
+    mem_mem_cmd_load && (!raw_hazard_mem && waw_hazard_mem)
 
-  val lu_stall_wb =
+  // stall for RAW/WAW hazards on load/AMO misses and mul/div in writeback.
-    dcache_miss &&
+  val data_hazard_wb =
-    ((id_ren1 && (id_raddr1 === io.dpath.wb_waddr)) ||
+    (dcache_miss || wb_reg_div_mul_val) &&
-     (id_ren2 && (id_raddr2 === io.dpath.wb_waddr)));
+    ((id_renx1.toBool && (id_raddr1 === io.dpath.wb_waddr)) ||
+     (id_renx2.toBool && (id_raddr2 === io.dpath.wb_waddr)) ||
+     (id_wen.toBool   && (id_waddr  === io.dpath.wb_waddr)));
 
-  val lu_stall = lu_stall_ex || lu_stall_mem || lu_stall_wb;
+  val data_hazard = data_hazard_ex || data_hazard_mem || data_hazard_wb;
-  
-  // check for divide and multiply instructions in ex,mem,wb stages
-  val dm_stall_ex = 
-    ex_reg_div_mul_val &&
-    ((id_ren1 && (id_raddr1 === io.dpath.ex_waddr)) ||
-     (id_ren2 && (id_raddr2 === io.dpath.ex_waddr)));
-
-  val dm_stall_mem = 
-    mem_reg_div_mul_val &&
-    ((id_ren1 && (id_raddr1 === io.dpath.mem_waddr)) ||
-     (id_ren2 && (id_raddr2 === io.dpath.mem_waddr)));
-     
-  val dm_stall_wb = 
-    wb_reg_div_mul_val &&
-    ((id_ren1 && (id_raddr1 === io.dpath.wb_waddr)) ||
-     (id_ren2 && (id_raddr2 === io.dpath.wb_waddr)));
-     
-  val dm_stall = dm_stall_ex || dm_stall_mem || dm_stall_wb;
 
   val ctrl_stalld =
     !take_pc &&
     (
-      dm_stall ||
+      data_hazard ||
-      lu_stall || 
+      id_renx2.toBool && id_stall_raddr2 ||
-      id_ren2 && id_stall_raddr2 ||
+      id_renx1.toBool && id_stall_raddr1 ||
-      id_ren1 && id_stall_raddr1 ||
+      id_wen.toBool   && id_stall_waddr  ||
-      (id_sel_wa === WA_RD) && id_stall_waddr ||
-      (id_sel_wa === WA_RA) && id_stall_ra ||
       id_mem_val.toBool && !(io.dmem.req_rdy && io.dtlb_rdy) ||
       ((id_sync === SYNC_D) || (id_sync === SYNC_I)) && !io.dmem.req_rdy ||
       id_console_out_val && !io.console.rdy ||
@@ -662,8 +646,8 @@ class rocketCtrl extends Component
   io.dpath.killm    := kill_mem;
 
   io.dpath.mem_load := mem_reg_mem_val && ((mem_reg_mem_cmd === M_XRD) || mem_reg_mem_cmd(3).toBool);
-  io.dpath.ren2     := id_ren2;
+  io.dpath.ren2     := id_renx2.toBool;
-  io.dpath.ren1     := id_ren1;
+  io.dpath.ren1     := id_renx1.toBool;
   io.dpath.sel_alu2 := id_sel_alu2;
   io.dpath.sel_alu1 := id_sel_alu1.toBool;
   io.dpath.fn_dw    := id_fn_dw.toBool;

rocket/src/main/scala/ctrl_util.scala

@@ -18,7 +18,6 @@ class ioCtrlSboard extends Bundle()
   val stalla  = Bool('output);
   val stallb  = Bool('output);
   val stallc  = Bool('output);
-  val stallra = Bool('output);
 }
 
 class rocketCtrlSboard extends Component
@@ -34,7 +33,6 @@ class rocketCtrlSboard extends Component
   io.stalla  := reg_busy(io.raddra).toBool;
   io.stallb  := reg_busy(io.raddrb).toBool;
   io.stallc  := reg_busy(io.raddrc).toBool;
-  io.stallra := reg_busy(RA).toBool;
 }
 
 class ioCtrlCnt extends Bundle()

rocket/src/main/scala/dcache.scala

@@ -21,6 +21,7 @@ class ioDmem(view: List[String] = null) extends Bundle(view) {
   val resp_miss = Bool('output);
   val resp_nack = Bool('output);
   val resp_val  = Bool('output);
+  val resp_replay = Bool('output);
   val resp_data = Bits(64, 'output);
   val resp_data_subword = Bits(64, 'output);
   val resp_tag  = Bits(DCACHE_TAG_BITS, 'output);

rocket/src/main/scala/dpath.scala

@@ -11,6 +11,7 @@ class ioDpathDmem extends Bundle()
   val req_tag   = UFix(CPU_TAG_BITS, 'output);
   val req_data  = Bits(64, 'output);
   val resp_val  = Bool('input);
+  val resp_replay = Bool('input);
   val resp_tag  = Bits(CPU_TAG_BITS, 'input);
   val resp_data = Bits(64, 'input);
   val resp_data_subword = Bits(64, 'input);

rocket/src/main/scala/nbdcache.scala

@@ -770,6 +770,7 @@ class HellaCache(lines: Int) extends Component {
   io.cpu.req_rdy   := flusher.io.req.ready && !(r_cpu_req_val_ && r_req_flush) && !pending_fence
   io.cpu.resp_nack := r_cpu_req_val_ && !io.cpu.req_kill && nack
   io.cpu.resp_val  := (tag_hit && !nack_hit && r_req_read) || replayer.io.cpu_resp_val
+  io.cpu.resp_replay := replayer.io.cpu_resp_val
   io.cpu.resp_miss := tag_miss && !nack_miss && r_req_read
   io.cpu.resp_tag  := Mux(replayer.io.cpu_resp_val, replayer.io.cpu_resp_tag, r_cpu_req_tag)
   io.cpu.resp_data := loadgen.io.dout