diff --git a/el2_lsu_bus_buffer.anno.json b/el2_lsu_bus_buffer.anno.json
index 33ba6fdd..11c11b9c 100644
--- a/el2_lsu_bus_buffer.anno.json
+++ b/el2_lsu_bus_buffer.anno.json
@@ -1,4 +1,14 @@
 [
+  {
+    "class":"firrtl.transforms.CombinationalPath",
+    "sink":"~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_lsu_bus_buffer_full_any",
+    "sources":[
+      "~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_ldst_dual_d",
+      "~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_dec_lsu_valid_raw_d",
+      "~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_lsu_busreq_m",
+      "~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_lsu_busreq_r"
+    ]
+  },
   {
     "class":"firrtl.transforms.CombinationalPath",
     "sink":"~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_wdata_in",
@@ -36,14 +46,6 @@
       "~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_end_addr_m"
     ]
   },
-  {
-    "class":"firrtl.transforms.CombinationalPath",
-    "sink":"~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_lsu_bus_buffer_full_any",
-    "sources":[
-      "~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_ldst_dual_d",
-      "~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_dec_lsu_valid_raw_d"
-    ]
-  },
   {
     "class":"firrtl.transforms.CombinationalPath",
     "sink":"~el2_lsu_bus_buffer|el2_lsu_bus_buffer>io_lsu_nonblock_load_valid_m",
diff --git a/el2_lsu_bus_buffer.fir b/el2_lsu_bus_buffer.fir
index 1adc1027..55d02b8c 100644
--- a/el2_lsu_bus_buffer.fir
+++ b/el2_lsu_bus_buffer.fir
@@ -5863,673 +5863,672 @@ circuit el2_lsu_bus_buffer :
     node _T_4431 = cat(_T_4430, buf_data_en[1]) @[Cat.scala 29:58]
     node _T_4432 = cat(_T_4431, buf_data_en[0]) @[Cat.scala 29:58]
     io.data_en <= _T_4432 @[el2_lsu_bus_buffer.scala 578:14]
-    node _T_4433 = neq(buf_state[0], UInt<3>("h00")) @[el2_lsu_bus_buffer.scala 580:60]
-    node _T_4434 = neq(buf_state[1], UInt<3>("h00")) @[el2_lsu_bus_buffer.scala 580:60]
-    node _T_4435 = neq(buf_state[2], UInt<3>("h00")) @[el2_lsu_bus_buffer.scala 580:60]
-    node _T_4436 = neq(buf_state[3], UInt<3>("h00")) @[el2_lsu_bus_buffer.scala 580:60]
-    node _T_4437 = add(_T_4436, _T_4435) @[el2_lsu_bus_buffer.scala 580:96]
-    node _T_4438 = add(_T_4437, _T_4434) @[el2_lsu_bus_buffer.scala 580:96]
-    node buf_numvld_any = add(_T_4438, _T_4433) @[el2_lsu_bus_buffer.scala 580:96]
-    node _T_4439 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 581:60]
-    node _T_4440 = eq(buf_state[0], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 581:79]
-    node _T_4441 = and(_T_4439, _T_4440) @[el2_lsu_bus_buffer.scala 581:64]
-    node _T_4442 = eq(buf_cmd_state_bus_en[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 581:91]
-    node _T_4443 = and(_T_4441, _T_4442) @[el2_lsu_bus_buffer.scala 581:89]
-    node _T_4444 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 581:60]
-    node _T_4445 = eq(buf_state[1], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 581:79]
-    node _T_4446 = and(_T_4444, _T_4445) @[el2_lsu_bus_buffer.scala 581:64]
-    node _T_4447 = eq(buf_cmd_state_bus_en[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 581:91]
-    node _T_4448 = and(_T_4446, _T_4447) @[el2_lsu_bus_buffer.scala 581:89]
-    node _T_4449 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 581:60]
-    node _T_4450 = eq(buf_state[2], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 581:79]
-    node _T_4451 = and(_T_4449, _T_4450) @[el2_lsu_bus_buffer.scala 581:64]
-    node _T_4452 = eq(buf_cmd_state_bus_en[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 581:91]
-    node _T_4453 = and(_T_4451, _T_4452) @[el2_lsu_bus_buffer.scala 581:89]
-    node _T_4454 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 581:60]
-    node _T_4455 = eq(buf_state[3], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 581:79]
-    node _T_4456 = and(_T_4454, _T_4455) @[el2_lsu_bus_buffer.scala 581:64]
-    node _T_4457 = eq(buf_cmd_state_bus_en[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 581:91]
-    node _T_4458 = and(_T_4456, _T_4457) @[el2_lsu_bus_buffer.scala 581:89]
-    node _T_4459 = add(_T_4458, _T_4453) @[el2_lsu_bus_buffer.scala 581:142]
-    node _T_4460 = add(_T_4459, _T_4448) @[el2_lsu_bus_buffer.scala 581:142]
-    node _T_4461 = add(_T_4460, _T_4443) @[el2_lsu_bus_buffer.scala 581:142]
-    buf_numvld_wrcmd_any <= _T_4461 @[el2_lsu_bus_buffer.scala 581:24]
-    node _T_4462 = eq(buf_state[0], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 582:63]
-    node _T_4463 = eq(buf_cmd_state_bus_en[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 582:75]
-    node _T_4464 = and(_T_4462, _T_4463) @[el2_lsu_bus_buffer.scala 582:73]
-    node _T_4465 = eq(buf_state[1], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 582:63]
-    node _T_4466 = eq(buf_cmd_state_bus_en[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 582:75]
-    node _T_4467 = and(_T_4465, _T_4466) @[el2_lsu_bus_buffer.scala 582:73]
-    node _T_4468 = eq(buf_state[2], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 582:63]
-    node _T_4469 = eq(buf_cmd_state_bus_en[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 582:75]
-    node _T_4470 = and(_T_4468, _T_4469) @[el2_lsu_bus_buffer.scala 582:73]
-    node _T_4471 = eq(buf_state[3], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 582:63]
-    node _T_4472 = eq(buf_cmd_state_bus_en[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 582:75]
-    node _T_4473 = and(_T_4471, _T_4472) @[el2_lsu_bus_buffer.scala 582:73]
-    node _T_4474 = add(_T_4473, _T_4470) @[el2_lsu_bus_buffer.scala 582:126]
-    node _T_4475 = add(_T_4474, _T_4467) @[el2_lsu_bus_buffer.scala 582:126]
-    node _T_4476 = add(_T_4475, _T_4464) @[el2_lsu_bus_buffer.scala 582:126]
-    buf_numvld_cmd_any <= _T_4476 @[el2_lsu_bus_buffer.scala 582:22]
-    node _T_4477 = eq(buf_state[0], UInt<3>("h01")) @[el2_lsu_bus_buffer.scala 583:63]
-    node _T_4478 = eq(buf_state[0], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 583:90]
-    node _T_4479 = eq(buf_cmd_state_bus_en[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 583:102]
-    node _T_4480 = and(_T_4478, _T_4479) @[el2_lsu_bus_buffer.scala 583:100]
-    node _T_4481 = or(_T_4477, _T_4480) @[el2_lsu_bus_buffer.scala 583:74]
-    node _T_4482 = eq(buf_state[1], UInt<3>("h01")) @[el2_lsu_bus_buffer.scala 583:63]
-    node _T_4483 = eq(buf_state[1], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 583:90]
-    node _T_4484 = eq(buf_cmd_state_bus_en[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 583:102]
-    node _T_4485 = and(_T_4483, _T_4484) @[el2_lsu_bus_buffer.scala 583:100]
-    node _T_4486 = or(_T_4482, _T_4485) @[el2_lsu_bus_buffer.scala 583:74]
-    node _T_4487 = eq(buf_state[2], UInt<3>("h01")) @[el2_lsu_bus_buffer.scala 583:63]
-    node _T_4488 = eq(buf_state[2], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 583:90]
-    node _T_4489 = eq(buf_cmd_state_bus_en[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 583:102]
-    node _T_4490 = and(_T_4488, _T_4489) @[el2_lsu_bus_buffer.scala 583:100]
-    node _T_4491 = or(_T_4487, _T_4490) @[el2_lsu_bus_buffer.scala 583:74]
-    node _T_4492 = eq(buf_state[3], UInt<3>("h01")) @[el2_lsu_bus_buffer.scala 583:63]
-    node _T_4493 = eq(buf_state[3], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 583:90]
-    node _T_4494 = eq(buf_cmd_state_bus_en[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 583:102]
-    node _T_4495 = and(_T_4493, _T_4494) @[el2_lsu_bus_buffer.scala 583:100]
-    node _T_4496 = or(_T_4492, _T_4495) @[el2_lsu_bus_buffer.scala 583:74]
-    node _T_4497 = add(_T_4496, _T_4491) @[el2_lsu_bus_buffer.scala 583:154]
-    node _T_4498 = add(_T_4497, _T_4486) @[el2_lsu_bus_buffer.scala 583:154]
-    node _T_4499 = add(_T_4498, _T_4481) @[el2_lsu_bus_buffer.scala 583:154]
-    buf_numvld_pend_any <= _T_4499 @[el2_lsu_bus_buffer.scala 583:23]
-    node _T_4500 = eq(buf_state[0], UInt<3>("h05")) @[el2_lsu_bus_buffer.scala 584:61]
-    node _T_4501 = eq(buf_state[1], UInt<3>("h05")) @[el2_lsu_bus_buffer.scala 584:61]
-    node _T_4502 = eq(buf_state[2], UInt<3>("h05")) @[el2_lsu_bus_buffer.scala 584:61]
-    node _T_4503 = eq(buf_state[3], UInt<3>("h05")) @[el2_lsu_bus_buffer.scala 584:61]
-    node _T_4504 = or(_T_4503, _T_4502) @[el2_lsu_bus_buffer.scala 584:93]
-    node _T_4505 = or(_T_4504, _T_4501) @[el2_lsu_bus_buffer.scala 584:93]
-    node _T_4506 = or(_T_4505, _T_4500) @[el2_lsu_bus_buffer.scala 584:93]
-    any_done_wait_state <= _T_4506 @[el2_lsu_bus_buffer.scala 584:23]
-    node _T_4507 = orr(buf_numvld_pend_any) @[el2_lsu_bus_buffer.scala 585:53]
-    io.lsu_bus_buffer_pend_any <= _T_4507 @[el2_lsu_bus_buffer.scala 585:30]
-    node _T_4508 = and(io.ldst_dual_d, io.dec_lsu_valid_raw_d) @[el2_lsu_bus_buffer.scala 586:52]
-    node _T_4509 = geq(buf_numvld_any, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 586:92]
-    node _T_4510 = eq(buf_numvld_any, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 586:119]
-    node _T_4511 = mux(_T_4508, _T_4509, _T_4510) @[el2_lsu_bus_buffer.scala 586:36]
-    io.lsu_bus_buffer_full_any <= _T_4511 @[el2_lsu_bus_buffer.scala 586:30]
-    node _T_4512 = orr(buf_state[0]) @[el2_lsu_bus_buffer.scala 587:52]
-    node _T_4513 = orr(buf_state[1]) @[el2_lsu_bus_buffer.scala 587:52]
-    node _T_4514 = orr(buf_state[2]) @[el2_lsu_bus_buffer.scala 587:52]
-    node _T_4515 = orr(buf_state[3]) @[el2_lsu_bus_buffer.scala 587:52]
-    node _T_4516 = or(_T_4512, _T_4513) @[el2_lsu_bus_buffer.scala 587:65]
+    node _T_4433 = cat(io.lsu_busreq_m, UInt<1>("h00")) @[Cat.scala 29:58]
+    node _T_4434 = mux(io.lsu_busreq_m, _T_4433, io.lsu_busreq_m) @[el2_lsu_bus_buffer.scala 580:27]
+    node _T_4435 = cat(io.lsu_busreq_r, UInt<1>("h00")) @[Cat.scala 29:58]
+    node _T_4436 = mux(io.lsu_busreq_r, _T_4435, io.lsu_busreq_r) @[el2_lsu_bus_buffer.scala 580:94]
+    node _T_4437 = add(_T_4434, _T_4436) @[el2_lsu_bus_buffer.scala 580:88]
+    node buf_numvld_any = add(_T_4437, ibuf_valid) @[el2_lsu_bus_buffer.scala 580:155]
+    node _T_4438 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 581:60]
+    node _T_4439 = eq(buf_state[0], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 581:79]
+    node _T_4440 = and(_T_4438, _T_4439) @[el2_lsu_bus_buffer.scala 581:64]
+    node _T_4441 = eq(buf_cmd_state_bus_en[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 581:91]
+    node _T_4442 = and(_T_4440, _T_4441) @[el2_lsu_bus_buffer.scala 581:89]
+    node _T_4443 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 581:60]
+    node _T_4444 = eq(buf_state[1], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 581:79]
+    node _T_4445 = and(_T_4443, _T_4444) @[el2_lsu_bus_buffer.scala 581:64]
+    node _T_4446 = eq(buf_cmd_state_bus_en[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 581:91]
+    node _T_4447 = and(_T_4445, _T_4446) @[el2_lsu_bus_buffer.scala 581:89]
+    node _T_4448 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 581:60]
+    node _T_4449 = eq(buf_state[2], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 581:79]
+    node _T_4450 = and(_T_4448, _T_4449) @[el2_lsu_bus_buffer.scala 581:64]
+    node _T_4451 = eq(buf_cmd_state_bus_en[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 581:91]
+    node _T_4452 = and(_T_4450, _T_4451) @[el2_lsu_bus_buffer.scala 581:89]
+    node _T_4453 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 581:60]
+    node _T_4454 = eq(buf_state[3], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 581:79]
+    node _T_4455 = and(_T_4453, _T_4454) @[el2_lsu_bus_buffer.scala 581:64]
+    node _T_4456 = eq(buf_cmd_state_bus_en[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 581:91]
+    node _T_4457 = and(_T_4455, _T_4456) @[el2_lsu_bus_buffer.scala 581:89]
+    node _T_4458 = add(_T_4457, _T_4452) @[el2_lsu_bus_buffer.scala 581:142]
+    node _T_4459 = add(_T_4458, _T_4447) @[el2_lsu_bus_buffer.scala 581:142]
+    node _T_4460 = add(_T_4459, _T_4442) @[el2_lsu_bus_buffer.scala 581:142]
+    buf_numvld_wrcmd_any <= _T_4460 @[el2_lsu_bus_buffer.scala 581:24]
+    node _T_4461 = eq(buf_state[0], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 582:63]
+    node _T_4462 = eq(buf_cmd_state_bus_en[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 582:75]
+    node _T_4463 = and(_T_4461, _T_4462) @[el2_lsu_bus_buffer.scala 582:73]
+    node _T_4464 = eq(buf_state[1], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 582:63]
+    node _T_4465 = eq(buf_cmd_state_bus_en[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 582:75]
+    node _T_4466 = and(_T_4464, _T_4465) @[el2_lsu_bus_buffer.scala 582:73]
+    node _T_4467 = eq(buf_state[2], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 582:63]
+    node _T_4468 = eq(buf_cmd_state_bus_en[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 582:75]
+    node _T_4469 = and(_T_4467, _T_4468) @[el2_lsu_bus_buffer.scala 582:73]
+    node _T_4470 = eq(buf_state[3], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 582:63]
+    node _T_4471 = eq(buf_cmd_state_bus_en[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 582:75]
+    node _T_4472 = and(_T_4470, _T_4471) @[el2_lsu_bus_buffer.scala 582:73]
+    node _T_4473 = add(_T_4472, _T_4469) @[el2_lsu_bus_buffer.scala 582:126]
+    node _T_4474 = add(_T_4473, _T_4466) @[el2_lsu_bus_buffer.scala 582:126]
+    node _T_4475 = add(_T_4474, _T_4463) @[el2_lsu_bus_buffer.scala 582:126]
+    buf_numvld_cmd_any <= _T_4475 @[el2_lsu_bus_buffer.scala 582:22]
+    node _T_4476 = eq(buf_state[0], UInt<3>("h01")) @[el2_lsu_bus_buffer.scala 583:63]
+    node _T_4477 = eq(buf_state[0], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 583:90]
+    node _T_4478 = eq(buf_cmd_state_bus_en[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 583:102]
+    node _T_4479 = and(_T_4477, _T_4478) @[el2_lsu_bus_buffer.scala 583:100]
+    node _T_4480 = or(_T_4476, _T_4479) @[el2_lsu_bus_buffer.scala 583:74]
+    node _T_4481 = eq(buf_state[1], UInt<3>("h01")) @[el2_lsu_bus_buffer.scala 583:63]
+    node _T_4482 = eq(buf_state[1], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 583:90]
+    node _T_4483 = eq(buf_cmd_state_bus_en[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 583:102]
+    node _T_4484 = and(_T_4482, _T_4483) @[el2_lsu_bus_buffer.scala 583:100]
+    node _T_4485 = or(_T_4481, _T_4484) @[el2_lsu_bus_buffer.scala 583:74]
+    node _T_4486 = eq(buf_state[2], UInt<3>("h01")) @[el2_lsu_bus_buffer.scala 583:63]
+    node _T_4487 = eq(buf_state[2], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 583:90]
+    node _T_4488 = eq(buf_cmd_state_bus_en[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 583:102]
+    node _T_4489 = and(_T_4487, _T_4488) @[el2_lsu_bus_buffer.scala 583:100]
+    node _T_4490 = or(_T_4486, _T_4489) @[el2_lsu_bus_buffer.scala 583:74]
+    node _T_4491 = eq(buf_state[3], UInt<3>("h01")) @[el2_lsu_bus_buffer.scala 583:63]
+    node _T_4492 = eq(buf_state[3], UInt<3>("h02")) @[el2_lsu_bus_buffer.scala 583:90]
+    node _T_4493 = eq(buf_cmd_state_bus_en[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 583:102]
+    node _T_4494 = and(_T_4492, _T_4493) @[el2_lsu_bus_buffer.scala 583:100]
+    node _T_4495 = or(_T_4491, _T_4494) @[el2_lsu_bus_buffer.scala 583:74]
+    node _T_4496 = add(_T_4495, _T_4490) @[el2_lsu_bus_buffer.scala 583:154]
+    node _T_4497 = add(_T_4496, _T_4485) @[el2_lsu_bus_buffer.scala 583:154]
+    node _T_4498 = add(_T_4497, _T_4480) @[el2_lsu_bus_buffer.scala 583:154]
+    buf_numvld_pend_any <= _T_4498 @[el2_lsu_bus_buffer.scala 583:23]
+    node _T_4499 = eq(buf_state[0], UInt<3>("h05")) @[el2_lsu_bus_buffer.scala 584:61]
+    node _T_4500 = eq(buf_state[1], UInt<3>("h05")) @[el2_lsu_bus_buffer.scala 584:61]
+    node _T_4501 = eq(buf_state[2], UInt<3>("h05")) @[el2_lsu_bus_buffer.scala 584:61]
+    node _T_4502 = eq(buf_state[3], UInt<3>("h05")) @[el2_lsu_bus_buffer.scala 584:61]
+    node _T_4503 = or(_T_4502, _T_4501) @[el2_lsu_bus_buffer.scala 584:93]
+    node _T_4504 = or(_T_4503, _T_4500) @[el2_lsu_bus_buffer.scala 584:93]
+    node _T_4505 = or(_T_4504, _T_4499) @[el2_lsu_bus_buffer.scala 584:93]
+    any_done_wait_state <= _T_4505 @[el2_lsu_bus_buffer.scala 584:23]
+    node _T_4506 = orr(buf_numvld_pend_any) @[el2_lsu_bus_buffer.scala 585:53]
+    io.lsu_bus_buffer_pend_any <= _T_4506 @[el2_lsu_bus_buffer.scala 585:30]
+    node _T_4507 = and(io.ldst_dual_d, io.dec_lsu_valid_raw_d) @[el2_lsu_bus_buffer.scala 586:52]
+    node _T_4508 = geq(buf_numvld_any, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 586:92]
+    node _T_4509 = eq(buf_numvld_any, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 586:119]
+    node _T_4510 = mux(_T_4507, _T_4508, _T_4509) @[el2_lsu_bus_buffer.scala 586:36]
+    io.lsu_bus_buffer_full_any <= _T_4510 @[el2_lsu_bus_buffer.scala 586:30]
+    node _T_4511 = orr(buf_state[0]) @[el2_lsu_bus_buffer.scala 587:52]
+    node _T_4512 = orr(buf_state[1]) @[el2_lsu_bus_buffer.scala 587:52]
+    node _T_4513 = orr(buf_state[2]) @[el2_lsu_bus_buffer.scala 587:52]
+    node _T_4514 = orr(buf_state[3]) @[el2_lsu_bus_buffer.scala 587:52]
+    node _T_4515 = or(_T_4511, _T_4512) @[el2_lsu_bus_buffer.scala 587:65]
+    node _T_4516 = or(_T_4515, _T_4513) @[el2_lsu_bus_buffer.scala 587:65]
     node _T_4517 = or(_T_4516, _T_4514) @[el2_lsu_bus_buffer.scala 587:65]
-    node _T_4518 = or(_T_4517, _T_4515) @[el2_lsu_bus_buffer.scala 587:65]
-    node _T_4519 = eq(_T_4518, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 587:34]
-    node _T_4520 = eq(ibuf_valid, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 587:72]
-    node _T_4521 = and(_T_4519, _T_4520) @[el2_lsu_bus_buffer.scala 587:70]
-    node _T_4522 = eq(obuf_valid, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 587:86]
-    node _T_4523 = and(_T_4521, _T_4522) @[el2_lsu_bus_buffer.scala 587:84]
-    io.lsu_bus_buffer_empty_any <= _T_4523 @[el2_lsu_bus_buffer.scala 587:31]
-    node _T_4524 = and(io.lsu_busreq_m, io.lsu_pkt_m.valid) @[el2_lsu_bus_buffer.scala 589:51]
-    node _T_4525 = and(_T_4524, io.lsu_pkt_m.load) @[el2_lsu_bus_buffer.scala 589:72]
-    node _T_4526 = eq(io.flush_m_up, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 589:94]
-    node _T_4527 = and(_T_4525, _T_4526) @[el2_lsu_bus_buffer.scala 589:92]
-    node _T_4528 = eq(io.ld_full_hit_m, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 589:111]
-    node _T_4529 = and(_T_4527, _T_4528) @[el2_lsu_bus_buffer.scala 589:109]
-    io.lsu_nonblock_load_valid_m <= _T_4529 @[el2_lsu_bus_buffer.scala 589:32]
+    node _T_4518 = eq(_T_4517, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 587:34]
+    node _T_4519 = eq(ibuf_valid, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 587:72]
+    node _T_4520 = and(_T_4518, _T_4519) @[el2_lsu_bus_buffer.scala 587:70]
+    node _T_4521 = eq(obuf_valid, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 587:86]
+    node _T_4522 = and(_T_4520, _T_4521) @[el2_lsu_bus_buffer.scala 587:84]
+    io.lsu_bus_buffer_empty_any <= _T_4522 @[el2_lsu_bus_buffer.scala 587:31]
+    node _T_4523 = and(io.lsu_busreq_m, io.lsu_pkt_m.valid) @[el2_lsu_bus_buffer.scala 589:51]
+    node _T_4524 = and(_T_4523, io.lsu_pkt_m.load) @[el2_lsu_bus_buffer.scala 589:72]
+    node _T_4525 = eq(io.flush_m_up, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 589:94]
+    node _T_4526 = and(_T_4524, _T_4525) @[el2_lsu_bus_buffer.scala 589:92]
+    node _T_4527 = eq(io.ld_full_hit_m, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 589:111]
+    node _T_4528 = and(_T_4526, _T_4527) @[el2_lsu_bus_buffer.scala 589:109]
+    io.lsu_nonblock_load_valid_m <= _T_4528 @[el2_lsu_bus_buffer.scala 589:32]
     io.lsu_nonblock_load_tag_m <= WrPtr0_m @[el2_lsu_bus_buffer.scala 590:30]
     wire lsu_nonblock_load_valid_r : UInt<1>
     lsu_nonblock_load_valid_r <= UInt<1>("h00")
-    node _T_4530 = eq(io.lsu_commit_r, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 592:61]
-    node _T_4531 = and(lsu_nonblock_load_valid_r, _T_4530) @[el2_lsu_bus_buffer.scala 592:59]
-    io.lsu_nonblock_load_inv_r <= _T_4531 @[el2_lsu_bus_buffer.scala 592:30]
+    node _T_4529 = eq(io.lsu_commit_r, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 592:61]
+    node _T_4530 = and(lsu_nonblock_load_valid_r, _T_4529) @[el2_lsu_bus_buffer.scala 592:59]
+    io.lsu_nonblock_load_inv_r <= _T_4530 @[el2_lsu_bus_buffer.scala 592:30]
     io.lsu_nonblock_load_inv_tag_r <= WrPtr0_r @[el2_lsu_bus_buffer.scala 593:34]
-    node _T_4532 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 594:80]
-    node _T_4533 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 594:127]
-    node _T_4534 = and(UInt<1>("h01"), _T_4533) @[el2_lsu_bus_buffer.scala 594:116]
-    node _T_4535 = eq(_T_4534, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 594:95]
-    node _T_4536 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 594:80]
-    node _T_4537 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 594:127]
-    node _T_4538 = and(UInt<1>("h01"), _T_4537) @[el2_lsu_bus_buffer.scala 594:116]
-    node _T_4539 = eq(_T_4538, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 594:95]
-    node _T_4540 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 594:80]
-    node _T_4541 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 594:127]
-    node _T_4542 = and(UInt<1>("h01"), _T_4541) @[el2_lsu_bus_buffer.scala 594:116]
-    node _T_4543 = eq(_T_4542, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 594:95]
-    node _T_4544 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 594:80]
-    node _T_4545 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 594:127]
-    node _T_4546 = and(UInt<1>("h01"), _T_4545) @[el2_lsu_bus_buffer.scala 594:116]
-    node _T_4547 = eq(_T_4546, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 594:95]
-    node _T_4548 = mux(_T_4532, _T_4535, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4549 = mux(_T_4536, _T_4539, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4550 = mux(_T_4540, _T_4543, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4551 = mux(_T_4544, _T_4547, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4552 = or(_T_4548, _T_4549) @[Mux.scala 27:72]
+    node _T_4531 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 594:80]
+    node _T_4532 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 594:127]
+    node _T_4533 = and(UInt<1>("h01"), _T_4532) @[el2_lsu_bus_buffer.scala 594:116]
+    node _T_4534 = eq(_T_4533, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 594:95]
+    node _T_4535 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 594:80]
+    node _T_4536 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 594:127]
+    node _T_4537 = and(UInt<1>("h01"), _T_4536) @[el2_lsu_bus_buffer.scala 594:116]
+    node _T_4538 = eq(_T_4537, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 594:95]
+    node _T_4539 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 594:80]
+    node _T_4540 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 594:127]
+    node _T_4541 = and(UInt<1>("h01"), _T_4540) @[el2_lsu_bus_buffer.scala 594:116]
+    node _T_4542 = eq(_T_4541, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 594:95]
+    node _T_4543 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 594:80]
+    node _T_4544 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 594:127]
+    node _T_4545 = and(UInt<1>("h01"), _T_4544) @[el2_lsu_bus_buffer.scala 594:116]
+    node _T_4546 = eq(_T_4545, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 594:95]
+    node _T_4547 = mux(_T_4531, _T_4534, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4548 = mux(_T_4535, _T_4538, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4549 = mux(_T_4539, _T_4542, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4550 = mux(_T_4543, _T_4546, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4551 = or(_T_4547, _T_4548) @[Mux.scala 27:72]
+    node _T_4552 = or(_T_4551, _T_4549) @[Mux.scala 27:72]
     node _T_4553 = or(_T_4552, _T_4550) @[Mux.scala 27:72]
-    node _T_4554 = or(_T_4553, _T_4551) @[Mux.scala 27:72]
     wire lsu_nonblock_load_data_ready : UInt<1> @[Mux.scala 27:72]
-    lsu_nonblock_load_data_ready <= _T_4554 @[Mux.scala 27:72]
-    node _T_4555 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 595:80]
-    node _T_4556 = bits(buf_error, 0, 0) @[el2_lsu_bus_buffer.scala 595:104]
-    node _T_4557 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 595:120]
-    node _T_4558 = eq(_T_4557, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 595:110]
-    node _T_4559 = and(_T_4556, _T_4558) @[el2_lsu_bus_buffer.scala 595:108]
-    node _T_4560 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 595:80]
-    node _T_4561 = bits(buf_error, 1, 1) @[el2_lsu_bus_buffer.scala 595:104]
-    node _T_4562 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 595:120]
-    node _T_4563 = eq(_T_4562, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 595:110]
-    node _T_4564 = and(_T_4561, _T_4563) @[el2_lsu_bus_buffer.scala 595:108]
-    node _T_4565 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 595:80]
-    node _T_4566 = bits(buf_error, 2, 2) @[el2_lsu_bus_buffer.scala 595:104]
-    node _T_4567 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 595:120]
-    node _T_4568 = eq(_T_4567, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 595:110]
-    node _T_4569 = and(_T_4566, _T_4568) @[el2_lsu_bus_buffer.scala 595:108]
-    node _T_4570 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 595:80]
-    node _T_4571 = bits(buf_error, 3, 3) @[el2_lsu_bus_buffer.scala 595:104]
-    node _T_4572 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 595:120]
-    node _T_4573 = eq(_T_4572, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 595:110]
-    node _T_4574 = and(_T_4571, _T_4573) @[el2_lsu_bus_buffer.scala 595:108]
-    node _T_4575 = mux(_T_4555, _T_4559, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4576 = mux(_T_4560, _T_4564, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4577 = mux(_T_4565, _T_4569, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4578 = mux(_T_4570, _T_4574, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4579 = or(_T_4575, _T_4576) @[Mux.scala 27:72]
+    lsu_nonblock_load_data_ready <= _T_4553 @[Mux.scala 27:72]
+    node _T_4554 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 595:80]
+    node _T_4555 = bits(buf_error, 0, 0) @[el2_lsu_bus_buffer.scala 595:104]
+    node _T_4556 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 595:120]
+    node _T_4557 = eq(_T_4556, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 595:110]
+    node _T_4558 = and(_T_4555, _T_4557) @[el2_lsu_bus_buffer.scala 595:108]
+    node _T_4559 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 595:80]
+    node _T_4560 = bits(buf_error, 1, 1) @[el2_lsu_bus_buffer.scala 595:104]
+    node _T_4561 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 595:120]
+    node _T_4562 = eq(_T_4561, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 595:110]
+    node _T_4563 = and(_T_4560, _T_4562) @[el2_lsu_bus_buffer.scala 595:108]
+    node _T_4564 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 595:80]
+    node _T_4565 = bits(buf_error, 2, 2) @[el2_lsu_bus_buffer.scala 595:104]
+    node _T_4566 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 595:120]
+    node _T_4567 = eq(_T_4566, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 595:110]
+    node _T_4568 = and(_T_4565, _T_4567) @[el2_lsu_bus_buffer.scala 595:108]
+    node _T_4569 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 595:80]
+    node _T_4570 = bits(buf_error, 3, 3) @[el2_lsu_bus_buffer.scala 595:104]
+    node _T_4571 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 595:120]
+    node _T_4572 = eq(_T_4571, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 595:110]
+    node _T_4573 = and(_T_4570, _T_4572) @[el2_lsu_bus_buffer.scala 595:108]
+    node _T_4574 = mux(_T_4554, _T_4558, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4575 = mux(_T_4559, _T_4563, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4576 = mux(_T_4564, _T_4568, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4577 = mux(_T_4569, _T_4573, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4578 = or(_T_4574, _T_4575) @[Mux.scala 27:72]
+    node _T_4579 = or(_T_4578, _T_4576) @[Mux.scala 27:72]
     node _T_4580 = or(_T_4579, _T_4577) @[Mux.scala 27:72]
-    node _T_4581 = or(_T_4580, _T_4578) @[Mux.scala 27:72]
-    wire _T_4582 : UInt<1> @[Mux.scala 27:72]
-    _T_4582 <= _T_4581 @[Mux.scala 27:72]
-    io.lsu_nonblock_load_data_error <= _T_4582 @[el2_lsu_bus_buffer.scala 595:35]
-    node _T_4583 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 596:79]
-    node _T_4584 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 596:102]
-    node _T_4585 = eq(_T_4584, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:92]
-    node _T_4586 = and(_T_4583, _T_4585) @[el2_lsu_bus_buffer.scala 596:90]
-    node _T_4587 = eq(buf_dual[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:109]
-    node _T_4588 = eq(buf_dualhi[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:124]
-    node _T_4589 = or(_T_4587, _T_4588) @[el2_lsu_bus_buffer.scala 596:122]
-    node _T_4590 = and(_T_4586, _T_4589) @[el2_lsu_bus_buffer.scala 596:106]
-    node _T_4591 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 596:79]
-    node _T_4592 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 596:102]
-    node _T_4593 = eq(_T_4592, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:92]
-    node _T_4594 = and(_T_4591, _T_4593) @[el2_lsu_bus_buffer.scala 596:90]
-    node _T_4595 = eq(buf_dual[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:109]
-    node _T_4596 = eq(buf_dualhi[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:124]
-    node _T_4597 = or(_T_4595, _T_4596) @[el2_lsu_bus_buffer.scala 596:122]
-    node _T_4598 = and(_T_4594, _T_4597) @[el2_lsu_bus_buffer.scala 596:106]
-    node _T_4599 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 596:79]
-    node _T_4600 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 596:102]
-    node _T_4601 = eq(_T_4600, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:92]
-    node _T_4602 = and(_T_4599, _T_4601) @[el2_lsu_bus_buffer.scala 596:90]
-    node _T_4603 = eq(buf_dual[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:109]
-    node _T_4604 = eq(buf_dualhi[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:124]
-    node _T_4605 = or(_T_4603, _T_4604) @[el2_lsu_bus_buffer.scala 596:122]
-    node _T_4606 = and(_T_4602, _T_4605) @[el2_lsu_bus_buffer.scala 596:106]
-    node _T_4607 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 596:79]
-    node _T_4608 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 596:102]
-    node _T_4609 = eq(_T_4608, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:92]
-    node _T_4610 = and(_T_4607, _T_4609) @[el2_lsu_bus_buffer.scala 596:90]
-    node _T_4611 = eq(buf_dual[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:109]
-    node _T_4612 = eq(buf_dualhi[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:124]
-    node _T_4613 = or(_T_4611, _T_4612) @[el2_lsu_bus_buffer.scala 596:122]
-    node _T_4614 = and(_T_4610, _T_4613) @[el2_lsu_bus_buffer.scala 596:106]
-    node _T_4615 = mux(_T_4590, UInt<1>("h00"), UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4616 = mux(_T_4598, UInt<1>("h01"), UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4617 = mux(_T_4606, UInt<2>("h02"), UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4618 = mux(_T_4614, UInt<2>("h03"), UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4619 = or(_T_4615, _T_4616) @[Mux.scala 27:72]
+    wire _T_4581 : UInt<1> @[Mux.scala 27:72]
+    _T_4581 <= _T_4580 @[Mux.scala 27:72]
+    io.lsu_nonblock_load_data_error <= _T_4581 @[el2_lsu_bus_buffer.scala 595:35]
+    node _T_4582 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 596:79]
+    node _T_4583 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 596:102]
+    node _T_4584 = eq(_T_4583, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:92]
+    node _T_4585 = and(_T_4582, _T_4584) @[el2_lsu_bus_buffer.scala 596:90]
+    node _T_4586 = eq(buf_dual[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:109]
+    node _T_4587 = eq(buf_dualhi[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:124]
+    node _T_4588 = or(_T_4586, _T_4587) @[el2_lsu_bus_buffer.scala 596:122]
+    node _T_4589 = and(_T_4585, _T_4588) @[el2_lsu_bus_buffer.scala 596:106]
+    node _T_4590 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 596:79]
+    node _T_4591 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 596:102]
+    node _T_4592 = eq(_T_4591, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:92]
+    node _T_4593 = and(_T_4590, _T_4592) @[el2_lsu_bus_buffer.scala 596:90]
+    node _T_4594 = eq(buf_dual[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:109]
+    node _T_4595 = eq(buf_dualhi[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:124]
+    node _T_4596 = or(_T_4594, _T_4595) @[el2_lsu_bus_buffer.scala 596:122]
+    node _T_4597 = and(_T_4593, _T_4596) @[el2_lsu_bus_buffer.scala 596:106]
+    node _T_4598 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 596:79]
+    node _T_4599 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 596:102]
+    node _T_4600 = eq(_T_4599, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:92]
+    node _T_4601 = and(_T_4598, _T_4600) @[el2_lsu_bus_buffer.scala 596:90]
+    node _T_4602 = eq(buf_dual[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:109]
+    node _T_4603 = eq(buf_dualhi[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:124]
+    node _T_4604 = or(_T_4602, _T_4603) @[el2_lsu_bus_buffer.scala 596:122]
+    node _T_4605 = and(_T_4601, _T_4604) @[el2_lsu_bus_buffer.scala 596:106]
+    node _T_4606 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 596:79]
+    node _T_4607 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 596:102]
+    node _T_4608 = eq(_T_4607, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:92]
+    node _T_4609 = and(_T_4606, _T_4608) @[el2_lsu_bus_buffer.scala 596:90]
+    node _T_4610 = eq(buf_dual[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:109]
+    node _T_4611 = eq(buf_dualhi[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 596:124]
+    node _T_4612 = or(_T_4610, _T_4611) @[el2_lsu_bus_buffer.scala 596:122]
+    node _T_4613 = and(_T_4609, _T_4612) @[el2_lsu_bus_buffer.scala 596:106]
+    node _T_4614 = mux(_T_4589, UInt<1>("h00"), UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4615 = mux(_T_4597, UInt<1>("h01"), UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4616 = mux(_T_4605, UInt<2>("h02"), UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4617 = mux(_T_4613, UInt<2>("h03"), UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4618 = or(_T_4614, _T_4615) @[Mux.scala 27:72]
+    node _T_4619 = or(_T_4618, _T_4616) @[Mux.scala 27:72]
     node _T_4620 = or(_T_4619, _T_4617) @[Mux.scala 27:72]
-    node _T_4621 = or(_T_4620, _T_4618) @[Mux.scala 27:72]
-    wire _T_4622 : UInt<2> @[Mux.scala 27:72]
-    _T_4622 <= _T_4621 @[Mux.scala 27:72]
-    io.lsu_nonblock_load_data_tag <= _T_4622 @[el2_lsu_bus_buffer.scala 596:33]
-    node _T_4623 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 597:78]
-    node _T_4624 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 597:101]
-    node _T_4625 = eq(_T_4624, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:91]
-    node _T_4626 = and(_T_4623, _T_4625) @[el2_lsu_bus_buffer.scala 597:89]
-    node _T_4627 = eq(buf_dual[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:108]
-    node _T_4628 = eq(buf_dualhi[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:123]
-    node _T_4629 = or(_T_4627, _T_4628) @[el2_lsu_bus_buffer.scala 597:121]
-    node _T_4630 = and(_T_4626, _T_4629) @[el2_lsu_bus_buffer.scala 597:105]
-    node _T_4631 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 597:78]
-    node _T_4632 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 597:101]
-    node _T_4633 = eq(_T_4632, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:91]
-    node _T_4634 = and(_T_4631, _T_4633) @[el2_lsu_bus_buffer.scala 597:89]
-    node _T_4635 = eq(buf_dual[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:108]
-    node _T_4636 = eq(buf_dualhi[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:123]
-    node _T_4637 = or(_T_4635, _T_4636) @[el2_lsu_bus_buffer.scala 597:121]
-    node _T_4638 = and(_T_4634, _T_4637) @[el2_lsu_bus_buffer.scala 597:105]
-    node _T_4639 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 597:78]
-    node _T_4640 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 597:101]
-    node _T_4641 = eq(_T_4640, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:91]
-    node _T_4642 = and(_T_4639, _T_4641) @[el2_lsu_bus_buffer.scala 597:89]
-    node _T_4643 = eq(buf_dual[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:108]
-    node _T_4644 = eq(buf_dualhi[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:123]
-    node _T_4645 = or(_T_4643, _T_4644) @[el2_lsu_bus_buffer.scala 597:121]
-    node _T_4646 = and(_T_4642, _T_4645) @[el2_lsu_bus_buffer.scala 597:105]
-    node _T_4647 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 597:78]
-    node _T_4648 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 597:101]
-    node _T_4649 = eq(_T_4648, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:91]
-    node _T_4650 = and(_T_4647, _T_4649) @[el2_lsu_bus_buffer.scala 597:89]
-    node _T_4651 = eq(buf_dual[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:108]
-    node _T_4652 = eq(buf_dualhi[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:123]
-    node _T_4653 = or(_T_4651, _T_4652) @[el2_lsu_bus_buffer.scala 597:121]
-    node _T_4654 = and(_T_4650, _T_4653) @[el2_lsu_bus_buffer.scala 597:105]
-    node _T_4655 = mux(_T_4630, buf_data[0], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4656 = mux(_T_4638, buf_data[1], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4657 = mux(_T_4646, buf_data[2], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4658 = mux(_T_4654, buf_data[3], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4659 = or(_T_4655, _T_4656) @[Mux.scala 27:72]
+    wire _T_4621 : UInt<2> @[Mux.scala 27:72]
+    _T_4621 <= _T_4620 @[Mux.scala 27:72]
+    io.lsu_nonblock_load_data_tag <= _T_4621 @[el2_lsu_bus_buffer.scala 596:33]
+    node _T_4622 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 597:78]
+    node _T_4623 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 597:101]
+    node _T_4624 = eq(_T_4623, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:91]
+    node _T_4625 = and(_T_4622, _T_4624) @[el2_lsu_bus_buffer.scala 597:89]
+    node _T_4626 = eq(buf_dual[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:108]
+    node _T_4627 = eq(buf_dualhi[0], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:123]
+    node _T_4628 = or(_T_4626, _T_4627) @[el2_lsu_bus_buffer.scala 597:121]
+    node _T_4629 = and(_T_4625, _T_4628) @[el2_lsu_bus_buffer.scala 597:105]
+    node _T_4630 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 597:78]
+    node _T_4631 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 597:101]
+    node _T_4632 = eq(_T_4631, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:91]
+    node _T_4633 = and(_T_4630, _T_4632) @[el2_lsu_bus_buffer.scala 597:89]
+    node _T_4634 = eq(buf_dual[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:108]
+    node _T_4635 = eq(buf_dualhi[1], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:123]
+    node _T_4636 = or(_T_4634, _T_4635) @[el2_lsu_bus_buffer.scala 597:121]
+    node _T_4637 = and(_T_4633, _T_4636) @[el2_lsu_bus_buffer.scala 597:105]
+    node _T_4638 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 597:78]
+    node _T_4639 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 597:101]
+    node _T_4640 = eq(_T_4639, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:91]
+    node _T_4641 = and(_T_4638, _T_4640) @[el2_lsu_bus_buffer.scala 597:89]
+    node _T_4642 = eq(buf_dual[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:108]
+    node _T_4643 = eq(buf_dualhi[2], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:123]
+    node _T_4644 = or(_T_4642, _T_4643) @[el2_lsu_bus_buffer.scala 597:121]
+    node _T_4645 = and(_T_4641, _T_4644) @[el2_lsu_bus_buffer.scala 597:105]
+    node _T_4646 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 597:78]
+    node _T_4647 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 597:101]
+    node _T_4648 = eq(_T_4647, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:91]
+    node _T_4649 = and(_T_4646, _T_4648) @[el2_lsu_bus_buffer.scala 597:89]
+    node _T_4650 = eq(buf_dual[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:108]
+    node _T_4651 = eq(buf_dualhi[3], UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 597:123]
+    node _T_4652 = or(_T_4650, _T_4651) @[el2_lsu_bus_buffer.scala 597:121]
+    node _T_4653 = and(_T_4649, _T_4652) @[el2_lsu_bus_buffer.scala 597:105]
+    node _T_4654 = mux(_T_4629, buf_data[0], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4655 = mux(_T_4637, buf_data[1], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4656 = mux(_T_4645, buf_data[2], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4657 = mux(_T_4653, buf_data[3], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4658 = or(_T_4654, _T_4655) @[Mux.scala 27:72]
+    node _T_4659 = or(_T_4658, _T_4656) @[Mux.scala 27:72]
     node _T_4660 = or(_T_4659, _T_4657) @[Mux.scala 27:72]
-    node _T_4661 = or(_T_4660, _T_4658) @[Mux.scala 27:72]
     wire lsu_nonblock_load_data_lo : UInt<32> @[Mux.scala 27:72]
-    lsu_nonblock_load_data_lo <= _T_4661 @[Mux.scala 27:72]
-    node _T_4662 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 598:78]
-    node _T_4663 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 598:101]
-    node _T_4664 = eq(_T_4663, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 598:91]
-    node _T_4665 = and(_T_4662, _T_4664) @[el2_lsu_bus_buffer.scala 598:89]
-    node _T_4666 = or(buf_dual[0], buf_dualhi[0]) @[el2_lsu_bus_buffer.scala 598:120]
-    node _T_4667 = and(_T_4665, _T_4666) @[el2_lsu_bus_buffer.scala 598:105]
-    node _T_4668 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 598:78]
-    node _T_4669 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 598:101]
-    node _T_4670 = eq(_T_4669, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 598:91]
-    node _T_4671 = and(_T_4668, _T_4670) @[el2_lsu_bus_buffer.scala 598:89]
-    node _T_4672 = or(buf_dual[1], buf_dualhi[1]) @[el2_lsu_bus_buffer.scala 598:120]
-    node _T_4673 = and(_T_4671, _T_4672) @[el2_lsu_bus_buffer.scala 598:105]
-    node _T_4674 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 598:78]
-    node _T_4675 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 598:101]
-    node _T_4676 = eq(_T_4675, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 598:91]
-    node _T_4677 = and(_T_4674, _T_4676) @[el2_lsu_bus_buffer.scala 598:89]
-    node _T_4678 = or(buf_dual[2], buf_dualhi[2]) @[el2_lsu_bus_buffer.scala 598:120]
-    node _T_4679 = and(_T_4677, _T_4678) @[el2_lsu_bus_buffer.scala 598:105]
-    node _T_4680 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 598:78]
-    node _T_4681 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 598:101]
-    node _T_4682 = eq(_T_4681, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 598:91]
-    node _T_4683 = and(_T_4680, _T_4682) @[el2_lsu_bus_buffer.scala 598:89]
-    node _T_4684 = or(buf_dual[3], buf_dualhi[3]) @[el2_lsu_bus_buffer.scala 598:120]
-    node _T_4685 = and(_T_4683, _T_4684) @[el2_lsu_bus_buffer.scala 598:105]
-    node _T_4686 = mux(_T_4667, buf_data[0], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4687 = mux(_T_4673, buf_data[1], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4688 = mux(_T_4679, buf_data[2], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4689 = mux(_T_4685, buf_data[3], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4690 = or(_T_4686, _T_4687) @[Mux.scala 27:72]
+    lsu_nonblock_load_data_lo <= _T_4660 @[Mux.scala 27:72]
+    node _T_4661 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 598:78]
+    node _T_4662 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 598:101]
+    node _T_4663 = eq(_T_4662, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 598:91]
+    node _T_4664 = and(_T_4661, _T_4663) @[el2_lsu_bus_buffer.scala 598:89]
+    node _T_4665 = or(buf_dual[0], buf_dualhi[0]) @[el2_lsu_bus_buffer.scala 598:120]
+    node _T_4666 = and(_T_4664, _T_4665) @[el2_lsu_bus_buffer.scala 598:105]
+    node _T_4667 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 598:78]
+    node _T_4668 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 598:101]
+    node _T_4669 = eq(_T_4668, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 598:91]
+    node _T_4670 = and(_T_4667, _T_4669) @[el2_lsu_bus_buffer.scala 598:89]
+    node _T_4671 = or(buf_dual[1], buf_dualhi[1]) @[el2_lsu_bus_buffer.scala 598:120]
+    node _T_4672 = and(_T_4670, _T_4671) @[el2_lsu_bus_buffer.scala 598:105]
+    node _T_4673 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 598:78]
+    node _T_4674 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 598:101]
+    node _T_4675 = eq(_T_4674, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 598:91]
+    node _T_4676 = and(_T_4673, _T_4675) @[el2_lsu_bus_buffer.scala 598:89]
+    node _T_4677 = or(buf_dual[2], buf_dualhi[2]) @[el2_lsu_bus_buffer.scala 598:120]
+    node _T_4678 = and(_T_4676, _T_4677) @[el2_lsu_bus_buffer.scala 598:105]
+    node _T_4679 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 598:78]
+    node _T_4680 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 598:101]
+    node _T_4681 = eq(_T_4680, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 598:91]
+    node _T_4682 = and(_T_4679, _T_4681) @[el2_lsu_bus_buffer.scala 598:89]
+    node _T_4683 = or(buf_dual[3], buf_dualhi[3]) @[el2_lsu_bus_buffer.scala 598:120]
+    node _T_4684 = and(_T_4682, _T_4683) @[el2_lsu_bus_buffer.scala 598:105]
+    node _T_4685 = mux(_T_4666, buf_data[0], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4686 = mux(_T_4672, buf_data[1], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4687 = mux(_T_4678, buf_data[2], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4688 = mux(_T_4684, buf_data[3], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4689 = or(_T_4685, _T_4686) @[Mux.scala 27:72]
+    node _T_4690 = or(_T_4689, _T_4687) @[Mux.scala 27:72]
     node _T_4691 = or(_T_4690, _T_4688) @[Mux.scala 27:72]
-    node _T_4692 = or(_T_4691, _T_4689) @[Mux.scala 27:72]
     wire lsu_nonblock_load_data_hi : UInt<32> @[Mux.scala 27:72]
-    lsu_nonblock_load_data_hi <= _T_4692 @[Mux.scala 27:72]
-    node _T_4693 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4694 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4695 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4696 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4697 = mux(_T_4693, buf_addr[0], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4698 = mux(_T_4694, buf_addr[1], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4699 = mux(_T_4695, buf_addr[2], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4700 = mux(_T_4696, buf_addr[3], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4701 = or(_T_4697, _T_4698) @[Mux.scala 27:72]
+    lsu_nonblock_load_data_hi <= _T_4691 @[Mux.scala 27:72]
+    node _T_4692 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4693 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4694 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4695 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4696 = mux(_T_4692, buf_addr[0], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4697 = mux(_T_4693, buf_addr[1], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4698 = mux(_T_4694, buf_addr[2], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4699 = mux(_T_4695, buf_addr[3], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4700 = or(_T_4696, _T_4697) @[Mux.scala 27:72]
+    node _T_4701 = or(_T_4700, _T_4698) @[Mux.scala 27:72]
     node _T_4702 = or(_T_4701, _T_4699) @[Mux.scala 27:72]
-    node _T_4703 = or(_T_4702, _T_4700) @[Mux.scala 27:72]
-    wire _T_4704 : UInt<32> @[Mux.scala 27:72]
-    _T_4704 <= _T_4703 @[Mux.scala 27:72]
-    node lsu_nonblock_addr_offset = bits(_T_4704, 1, 0) @[el2_lsu_bus_buffer.scala 599:83]
-    node _T_4705 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4706 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4707 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4708 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4709 = mux(_T_4705, buf_sz[0], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4710 = mux(_T_4706, buf_sz[1], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4711 = mux(_T_4707, buf_sz[2], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4712 = mux(_T_4708, buf_sz[3], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4713 = or(_T_4709, _T_4710) @[Mux.scala 27:72]
+    wire _T_4703 : UInt<32> @[Mux.scala 27:72]
+    _T_4703 <= _T_4702 @[Mux.scala 27:72]
+    node lsu_nonblock_addr_offset = bits(_T_4703, 1, 0) @[el2_lsu_bus_buffer.scala 599:83]
+    node _T_4704 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4705 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4706 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4707 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4708 = mux(_T_4704, buf_sz[0], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4709 = mux(_T_4705, buf_sz[1], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4710 = mux(_T_4706, buf_sz[2], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4711 = mux(_T_4707, buf_sz[3], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4712 = or(_T_4708, _T_4709) @[Mux.scala 27:72]
+    node _T_4713 = or(_T_4712, _T_4710) @[Mux.scala 27:72]
     node _T_4714 = or(_T_4713, _T_4711) @[Mux.scala 27:72]
-    node _T_4715 = or(_T_4714, _T_4712) @[Mux.scala 27:72]
     wire lsu_nonblock_sz : UInt<2> @[Mux.scala 27:72]
-    lsu_nonblock_sz <= _T_4715 @[Mux.scala 27:72]
-    node _T_4716 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 119:118]
-    node _T_4717 = bits(buf_unsign, 0, 0) @[el2_lsu_bus_buffer.scala 119:129]
-    node _T_4718 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 119:118]
-    node _T_4719 = bits(buf_unsign, 1, 1) @[el2_lsu_bus_buffer.scala 119:129]
-    node _T_4720 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 119:118]
-    node _T_4721 = bits(buf_unsign, 2, 2) @[el2_lsu_bus_buffer.scala 119:129]
-    node _T_4722 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 119:118]
-    node _T_4723 = bits(buf_unsign, 3, 3) @[el2_lsu_bus_buffer.scala 119:129]
-    node _T_4724 = mux(_T_4716, _T_4717, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4725 = mux(_T_4718, _T_4719, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4726 = mux(_T_4720, _T_4721, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4727 = mux(_T_4722, _T_4723, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4728 = or(_T_4724, _T_4725) @[Mux.scala 27:72]
+    lsu_nonblock_sz <= _T_4714 @[Mux.scala 27:72]
+    node _T_4715 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 119:118]
+    node _T_4716 = bits(buf_unsign, 0, 0) @[el2_lsu_bus_buffer.scala 119:129]
+    node _T_4717 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 119:118]
+    node _T_4718 = bits(buf_unsign, 1, 1) @[el2_lsu_bus_buffer.scala 119:129]
+    node _T_4719 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 119:118]
+    node _T_4720 = bits(buf_unsign, 2, 2) @[el2_lsu_bus_buffer.scala 119:129]
+    node _T_4721 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 119:118]
+    node _T_4722 = bits(buf_unsign, 3, 3) @[el2_lsu_bus_buffer.scala 119:129]
+    node _T_4723 = mux(_T_4715, _T_4716, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4724 = mux(_T_4717, _T_4718, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4725 = mux(_T_4719, _T_4720, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4726 = mux(_T_4721, _T_4722, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4727 = or(_T_4723, _T_4724) @[Mux.scala 27:72]
+    node _T_4728 = or(_T_4727, _T_4725) @[Mux.scala 27:72]
     node _T_4729 = or(_T_4728, _T_4726) @[Mux.scala 27:72]
-    node _T_4730 = or(_T_4729, _T_4727) @[Mux.scala 27:72]
     wire lsu_nonblock_unsign : UInt<1> @[Mux.scala 27:72]
-    lsu_nonblock_unsign <= _T_4730 @[Mux.scala 27:72]
-    node _T_4731 = cat(buf_dual[3], buf_dual[2]) @[Cat.scala 29:58]
-    node _T_4732 = cat(_T_4731, buf_dual[1]) @[Cat.scala 29:58]
-    node _T_4733 = cat(_T_4732, buf_dual[0]) @[Cat.scala 29:58]
-    node _T_4734 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 119:118]
-    node _T_4735 = bits(_T_4733, 0, 0) @[el2_lsu_bus_buffer.scala 119:129]
-    node _T_4736 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 119:118]
-    node _T_4737 = bits(_T_4733, 1, 1) @[el2_lsu_bus_buffer.scala 119:129]
-    node _T_4738 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 119:118]
-    node _T_4739 = bits(_T_4733, 2, 2) @[el2_lsu_bus_buffer.scala 119:129]
-    node _T_4740 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 119:118]
-    node _T_4741 = bits(_T_4733, 3, 3) @[el2_lsu_bus_buffer.scala 119:129]
-    node _T_4742 = mux(_T_4734, _T_4735, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4743 = mux(_T_4736, _T_4737, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4744 = mux(_T_4738, _T_4739, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4745 = mux(_T_4740, _T_4741, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4746 = or(_T_4742, _T_4743) @[Mux.scala 27:72]
+    lsu_nonblock_unsign <= _T_4729 @[Mux.scala 27:72]
+    node _T_4730 = cat(buf_dual[3], buf_dual[2]) @[Cat.scala 29:58]
+    node _T_4731 = cat(_T_4730, buf_dual[1]) @[Cat.scala 29:58]
+    node _T_4732 = cat(_T_4731, buf_dual[0]) @[Cat.scala 29:58]
+    node _T_4733 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 119:118]
+    node _T_4734 = bits(_T_4732, 0, 0) @[el2_lsu_bus_buffer.scala 119:129]
+    node _T_4735 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 119:118]
+    node _T_4736 = bits(_T_4732, 1, 1) @[el2_lsu_bus_buffer.scala 119:129]
+    node _T_4737 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 119:118]
+    node _T_4738 = bits(_T_4732, 2, 2) @[el2_lsu_bus_buffer.scala 119:129]
+    node _T_4739 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 119:118]
+    node _T_4740 = bits(_T_4732, 3, 3) @[el2_lsu_bus_buffer.scala 119:129]
+    node _T_4741 = mux(_T_4733, _T_4734, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4742 = mux(_T_4735, _T_4736, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4743 = mux(_T_4737, _T_4738, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4744 = mux(_T_4739, _T_4740, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4745 = or(_T_4741, _T_4742) @[Mux.scala 27:72]
+    node _T_4746 = or(_T_4745, _T_4743) @[Mux.scala 27:72]
     node _T_4747 = or(_T_4746, _T_4744) @[Mux.scala 27:72]
-    node _T_4748 = or(_T_4747, _T_4745) @[Mux.scala 27:72]
     wire lsu_nonblock_dual : UInt<1> @[Mux.scala 27:72]
-    lsu_nonblock_dual <= _T_4748 @[Mux.scala 27:72]
-    node _T_4749 = cat(lsu_nonblock_load_data_hi, lsu_nonblock_load_data_lo) @[Cat.scala 29:58]
-    node _T_4750 = mul(lsu_nonblock_addr_offset, UInt<4>("h08")) @[el2_lsu_bus_buffer.scala 603:121]
-    node lsu_nonblock_data_unalgn = dshr(_T_4749, _T_4750) @[el2_lsu_bus_buffer.scala 603:92]
+    lsu_nonblock_dual <= _T_4747 @[Mux.scala 27:72]
+    node _T_4748 = cat(lsu_nonblock_load_data_hi, lsu_nonblock_load_data_lo) @[Cat.scala 29:58]
+    node _T_4749 = mul(lsu_nonblock_addr_offset, UInt<4>("h08")) @[el2_lsu_bus_buffer.scala 603:121]
+    node lsu_nonblock_data_unalgn = dshr(_T_4748, _T_4749) @[el2_lsu_bus_buffer.scala 603:92]
     io.data_hi <= lsu_nonblock_load_data_hi @[el2_lsu_bus_buffer.scala 604:14]
     io.data_lo <= lsu_nonblock_load_data_lo @[el2_lsu_bus_buffer.scala 605:14]
-    node _T_4751 = eq(io.lsu_nonblock_load_data_error, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 606:69]
-    node _T_4752 = and(lsu_nonblock_load_data_ready, _T_4751) @[el2_lsu_bus_buffer.scala 606:67]
-    io.lsu_nonblock_load_data_valid <= _T_4752 @[el2_lsu_bus_buffer.scala 606:35]
-    node _T_4753 = eq(lsu_nonblock_sz, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 607:81]
-    node _T_4754 = and(lsu_nonblock_unsign, _T_4753) @[el2_lsu_bus_buffer.scala 607:63]
-    node _T_4755 = bits(lsu_nonblock_data_unalgn, 7, 0) @[el2_lsu_bus_buffer.scala 607:131]
-    node _T_4756 = cat(UInt<24>("h00"), _T_4755) @[Cat.scala 29:58]
-    node _T_4757 = eq(lsu_nonblock_sz, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 608:45]
-    node _T_4758 = and(lsu_nonblock_unsign, _T_4757) @[el2_lsu_bus_buffer.scala 608:26]
-    node _T_4759 = bits(lsu_nonblock_data_unalgn, 15, 0) @[el2_lsu_bus_buffer.scala 608:95]
-    node _T_4760 = cat(UInt<16>("h00"), _T_4759) @[Cat.scala 29:58]
-    node _T_4761 = eq(lsu_nonblock_unsign, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 609:6]
-    node _T_4762 = eq(lsu_nonblock_sz, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 609:45]
-    node _T_4763 = and(_T_4761, _T_4762) @[el2_lsu_bus_buffer.scala 609:27]
-    node _T_4764 = bits(lsu_nonblock_data_unalgn, 7, 7) @[el2_lsu_bus_buffer.scala 609:93]
-    node _T_4765 = bits(_T_4764, 0, 0) @[Bitwise.scala 72:15]
-    node _T_4766 = mux(_T_4765, UInt<24>("h0ffffff"), UInt<24>("h00")) @[Bitwise.scala 72:12]
-    node _T_4767 = bits(lsu_nonblock_data_unalgn, 7, 0) @[el2_lsu_bus_buffer.scala 609:123]
-    node _T_4768 = cat(_T_4766, _T_4767) @[Cat.scala 29:58]
-    node _T_4769 = eq(lsu_nonblock_unsign, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 610:6]
-    node _T_4770 = eq(lsu_nonblock_sz, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 610:45]
-    node _T_4771 = and(_T_4769, _T_4770) @[el2_lsu_bus_buffer.scala 610:27]
-    node _T_4772 = bits(lsu_nonblock_data_unalgn, 15, 15) @[el2_lsu_bus_buffer.scala 610:93]
-    node _T_4773 = bits(_T_4772, 0, 0) @[Bitwise.scala 72:15]
-    node _T_4774 = mux(_T_4773, UInt<16>("h0ffff"), UInt<16>("h00")) @[Bitwise.scala 72:12]
-    node _T_4775 = bits(lsu_nonblock_data_unalgn, 15, 0) @[el2_lsu_bus_buffer.scala 610:124]
-    node _T_4776 = cat(_T_4774, _T_4775) @[Cat.scala 29:58]
-    node _T_4777 = eq(lsu_nonblock_sz, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 611:21]
-    node _T_4778 = mux(_T_4754, _T_4756, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4779 = mux(_T_4758, _T_4760, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4780 = mux(_T_4763, _T_4768, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4781 = mux(_T_4771, _T_4776, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4782 = mux(_T_4777, lsu_nonblock_data_unalgn, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4783 = or(_T_4778, _T_4779) @[Mux.scala 27:72]
+    node _T_4750 = eq(io.lsu_nonblock_load_data_error, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 606:69]
+    node _T_4751 = and(lsu_nonblock_load_data_ready, _T_4750) @[el2_lsu_bus_buffer.scala 606:67]
+    io.lsu_nonblock_load_data_valid <= _T_4751 @[el2_lsu_bus_buffer.scala 606:35]
+    node _T_4752 = eq(lsu_nonblock_sz, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 607:81]
+    node _T_4753 = and(lsu_nonblock_unsign, _T_4752) @[el2_lsu_bus_buffer.scala 607:63]
+    node _T_4754 = bits(lsu_nonblock_data_unalgn, 7, 0) @[el2_lsu_bus_buffer.scala 607:131]
+    node _T_4755 = cat(UInt<24>("h00"), _T_4754) @[Cat.scala 29:58]
+    node _T_4756 = eq(lsu_nonblock_sz, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 608:45]
+    node _T_4757 = and(lsu_nonblock_unsign, _T_4756) @[el2_lsu_bus_buffer.scala 608:26]
+    node _T_4758 = bits(lsu_nonblock_data_unalgn, 15, 0) @[el2_lsu_bus_buffer.scala 608:95]
+    node _T_4759 = cat(UInt<16>("h00"), _T_4758) @[Cat.scala 29:58]
+    node _T_4760 = eq(lsu_nonblock_unsign, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 609:6]
+    node _T_4761 = eq(lsu_nonblock_sz, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 609:45]
+    node _T_4762 = and(_T_4760, _T_4761) @[el2_lsu_bus_buffer.scala 609:27]
+    node _T_4763 = bits(lsu_nonblock_data_unalgn, 7, 7) @[el2_lsu_bus_buffer.scala 609:93]
+    node _T_4764 = bits(_T_4763, 0, 0) @[Bitwise.scala 72:15]
+    node _T_4765 = mux(_T_4764, UInt<24>("h0ffffff"), UInt<24>("h00")) @[Bitwise.scala 72:12]
+    node _T_4766 = bits(lsu_nonblock_data_unalgn, 7, 0) @[el2_lsu_bus_buffer.scala 609:123]
+    node _T_4767 = cat(_T_4765, _T_4766) @[Cat.scala 29:58]
+    node _T_4768 = eq(lsu_nonblock_unsign, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 610:6]
+    node _T_4769 = eq(lsu_nonblock_sz, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 610:45]
+    node _T_4770 = and(_T_4768, _T_4769) @[el2_lsu_bus_buffer.scala 610:27]
+    node _T_4771 = bits(lsu_nonblock_data_unalgn, 15, 15) @[el2_lsu_bus_buffer.scala 610:93]
+    node _T_4772 = bits(_T_4771, 0, 0) @[Bitwise.scala 72:15]
+    node _T_4773 = mux(_T_4772, UInt<16>("h0ffff"), UInt<16>("h00")) @[Bitwise.scala 72:12]
+    node _T_4774 = bits(lsu_nonblock_data_unalgn, 15, 0) @[el2_lsu_bus_buffer.scala 610:124]
+    node _T_4775 = cat(_T_4773, _T_4774) @[Cat.scala 29:58]
+    node _T_4776 = eq(lsu_nonblock_sz, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 611:21]
+    node _T_4777 = mux(_T_4753, _T_4755, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4778 = mux(_T_4757, _T_4759, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4779 = mux(_T_4762, _T_4767, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4780 = mux(_T_4770, _T_4775, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4781 = mux(_T_4776, lsu_nonblock_data_unalgn, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4782 = or(_T_4777, _T_4778) @[Mux.scala 27:72]
+    node _T_4783 = or(_T_4782, _T_4779) @[Mux.scala 27:72]
     node _T_4784 = or(_T_4783, _T_4780) @[Mux.scala 27:72]
     node _T_4785 = or(_T_4784, _T_4781) @[Mux.scala 27:72]
-    node _T_4786 = or(_T_4785, _T_4782) @[Mux.scala 27:72]
-    wire _T_4787 : UInt<64> @[Mux.scala 27:72]
-    _T_4787 <= _T_4786 @[Mux.scala 27:72]
-    io.lsu_nonblock_load_data <= _T_4787 @[el2_lsu_bus_buffer.scala 607:29]
-    node _T_4788 = eq(buf_state[0], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 612:62]
-    node _T_4789 = bits(buf_sideeffect, 0, 0) @[el2_lsu_bus_buffer.scala 612:89]
-    node _T_4790 = and(_T_4788, _T_4789) @[el2_lsu_bus_buffer.scala 612:73]
-    node _T_4791 = and(_T_4790, io.dec_tlu_sideeffect_posted_disable) @[el2_lsu_bus_buffer.scala 612:93]
-    node _T_4792 = eq(buf_state[1], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 612:62]
-    node _T_4793 = bits(buf_sideeffect, 1, 1) @[el2_lsu_bus_buffer.scala 612:89]
-    node _T_4794 = and(_T_4792, _T_4793) @[el2_lsu_bus_buffer.scala 612:73]
-    node _T_4795 = and(_T_4794, io.dec_tlu_sideeffect_posted_disable) @[el2_lsu_bus_buffer.scala 612:93]
-    node _T_4796 = eq(buf_state[2], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 612:62]
-    node _T_4797 = bits(buf_sideeffect, 2, 2) @[el2_lsu_bus_buffer.scala 612:89]
-    node _T_4798 = and(_T_4796, _T_4797) @[el2_lsu_bus_buffer.scala 612:73]
-    node _T_4799 = and(_T_4798, io.dec_tlu_sideeffect_posted_disable) @[el2_lsu_bus_buffer.scala 612:93]
-    node _T_4800 = eq(buf_state[3], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 612:62]
-    node _T_4801 = bits(buf_sideeffect, 3, 3) @[el2_lsu_bus_buffer.scala 612:89]
-    node _T_4802 = and(_T_4800, _T_4801) @[el2_lsu_bus_buffer.scala 612:73]
-    node _T_4803 = and(_T_4802, io.dec_tlu_sideeffect_posted_disable) @[el2_lsu_bus_buffer.scala 612:93]
-    node _T_4804 = or(_T_4791, _T_4795) @[el2_lsu_bus_buffer.scala 612:141]
-    node _T_4805 = or(_T_4804, _T_4799) @[el2_lsu_bus_buffer.scala 612:141]
-    node _T_4806 = or(_T_4805, _T_4803) @[el2_lsu_bus_buffer.scala 612:141]
-    bus_sideeffect_pend <= _T_4806 @[el2_lsu_bus_buffer.scala 612:23]
-    node _T_4807 = eq(buf_state[0], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 613:71]
-    node _T_4808 = and(UInt<1>("h01"), obuf_valid) @[el2_lsu_bus_buffer.scala 614:25]
-    node _T_4809 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 614:50]
-    node _T_4810 = bits(buf_addr[0], 31, 3) @[el2_lsu_bus_buffer.scala 614:70]
-    node _T_4811 = eq(_T_4809, _T_4810) @[el2_lsu_bus_buffer.scala 614:56]
-    node _T_4812 = and(_T_4808, _T_4811) @[el2_lsu_bus_buffer.scala 614:38]
-    node _T_4813 = eq(obuf_tag0, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:92]
-    node _T_4814 = eq(obuf_tag1, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:126]
-    node _T_4815 = and(obuf_merge, _T_4814) @[el2_lsu_bus_buffer.scala 614:114]
-    node _T_4816 = or(_T_4813, _T_4815) @[el2_lsu_bus_buffer.scala 614:100]
-    node _T_4817 = eq(_T_4816, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:80]
-    node _T_4818 = and(_T_4812, _T_4817) @[el2_lsu_bus_buffer.scala 614:78]
-    node _T_4819 = eq(buf_state[1], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 613:71]
-    node _T_4820 = and(UInt<1>("h01"), obuf_valid) @[el2_lsu_bus_buffer.scala 614:25]
-    node _T_4821 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 614:50]
-    node _T_4822 = bits(buf_addr[1], 31, 3) @[el2_lsu_bus_buffer.scala 614:70]
-    node _T_4823 = eq(_T_4821, _T_4822) @[el2_lsu_bus_buffer.scala 614:56]
-    node _T_4824 = and(_T_4820, _T_4823) @[el2_lsu_bus_buffer.scala 614:38]
-    node _T_4825 = eq(obuf_tag0, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 614:92]
-    node _T_4826 = eq(obuf_tag1, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 614:126]
-    node _T_4827 = and(obuf_merge, _T_4826) @[el2_lsu_bus_buffer.scala 614:114]
-    node _T_4828 = or(_T_4825, _T_4827) @[el2_lsu_bus_buffer.scala 614:100]
-    node _T_4829 = eq(_T_4828, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:80]
-    node _T_4830 = and(_T_4824, _T_4829) @[el2_lsu_bus_buffer.scala 614:78]
-    node _T_4831 = eq(buf_state[2], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 613:71]
-    node _T_4832 = and(UInt<1>("h01"), obuf_valid) @[el2_lsu_bus_buffer.scala 614:25]
-    node _T_4833 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 614:50]
-    node _T_4834 = bits(buf_addr[2], 31, 3) @[el2_lsu_bus_buffer.scala 614:70]
-    node _T_4835 = eq(_T_4833, _T_4834) @[el2_lsu_bus_buffer.scala 614:56]
-    node _T_4836 = and(_T_4832, _T_4835) @[el2_lsu_bus_buffer.scala 614:38]
-    node _T_4837 = eq(obuf_tag0, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 614:92]
-    node _T_4838 = eq(obuf_tag1, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 614:126]
-    node _T_4839 = and(obuf_merge, _T_4838) @[el2_lsu_bus_buffer.scala 614:114]
-    node _T_4840 = or(_T_4837, _T_4839) @[el2_lsu_bus_buffer.scala 614:100]
-    node _T_4841 = eq(_T_4840, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:80]
-    node _T_4842 = and(_T_4836, _T_4841) @[el2_lsu_bus_buffer.scala 614:78]
-    node _T_4843 = eq(buf_state[3], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 613:71]
-    node _T_4844 = and(UInt<1>("h01"), obuf_valid) @[el2_lsu_bus_buffer.scala 614:25]
-    node _T_4845 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 614:50]
-    node _T_4846 = bits(buf_addr[3], 31, 3) @[el2_lsu_bus_buffer.scala 614:70]
-    node _T_4847 = eq(_T_4845, _T_4846) @[el2_lsu_bus_buffer.scala 614:56]
-    node _T_4848 = and(_T_4844, _T_4847) @[el2_lsu_bus_buffer.scala 614:38]
-    node _T_4849 = eq(obuf_tag0, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 614:92]
-    node _T_4850 = eq(obuf_tag1, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 614:126]
-    node _T_4851 = and(obuf_merge, _T_4850) @[el2_lsu_bus_buffer.scala 614:114]
-    node _T_4852 = or(_T_4849, _T_4851) @[el2_lsu_bus_buffer.scala 614:100]
-    node _T_4853 = eq(_T_4852, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:80]
-    node _T_4854 = and(_T_4848, _T_4853) @[el2_lsu_bus_buffer.scala 614:78]
-    node _T_4855 = mux(_T_4807, _T_4818, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4856 = mux(_T_4819, _T_4830, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4857 = mux(_T_4831, _T_4842, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4858 = mux(_T_4843, _T_4854, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4859 = or(_T_4855, _T_4856) @[Mux.scala 27:72]
+    wire _T_4786 : UInt<64> @[Mux.scala 27:72]
+    _T_4786 <= _T_4785 @[Mux.scala 27:72]
+    io.lsu_nonblock_load_data <= _T_4786 @[el2_lsu_bus_buffer.scala 607:29]
+    node _T_4787 = eq(buf_state[0], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 612:62]
+    node _T_4788 = bits(buf_sideeffect, 0, 0) @[el2_lsu_bus_buffer.scala 612:89]
+    node _T_4789 = and(_T_4787, _T_4788) @[el2_lsu_bus_buffer.scala 612:73]
+    node _T_4790 = and(_T_4789, io.dec_tlu_sideeffect_posted_disable) @[el2_lsu_bus_buffer.scala 612:93]
+    node _T_4791 = eq(buf_state[1], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 612:62]
+    node _T_4792 = bits(buf_sideeffect, 1, 1) @[el2_lsu_bus_buffer.scala 612:89]
+    node _T_4793 = and(_T_4791, _T_4792) @[el2_lsu_bus_buffer.scala 612:73]
+    node _T_4794 = and(_T_4793, io.dec_tlu_sideeffect_posted_disable) @[el2_lsu_bus_buffer.scala 612:93]
+    node _T_4795 = eq(buf_state[2], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 612:62]
+    node _T_4796 = bits(buf_sideeffect, 2, 2) @[el2_lsu_bus_buffer.scala 612:89]
+    node _T_4797 = and(_T_4795, _T_4796) @[el2_lsu_bus_buffer.scala 612:73]
+    node _T_4798 = and(_T_4797, io.dec_tlu_sideeffect_posted_disable) @[el2_lsu_bus_buffer.scala 612:93]
+    node _T_4799 = eq(buf_state[3], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 612:62]
+    node _T_4800 = bits(buf_sideeffect, 3, 3) @[el2_lsu_bus_buffer.scala 612:89]
+    node _T_4801 = and(_T_4799, _T_4800) @[el2_lsu_bus_buffer.scala 612:73]
+    node _T_4802 = and(_T_4801, io.dec_tlu_sideeffect_posted_disable) @[el2_lsu_bus_buffer.scala 612:93]
+    node _T_4803 = or(_T_4790, _T_4794) @[el2_lsu_bus_buffer.scala 612:141]
+    node _T_4804 = or(_T_4803, _T_4798) @[el2_lsu_bus_buffer.scala 612:141]
+    node _T_4805 = or(_T_4804, _T_4802) @[el2_lsu_bus_buffer.scala 612:141]
+    bus_sideeffect_pend <= _T_4805 @[el2_lsu_bus_buffer.scala 612:23]
+    node _T_4806 = eq(buf_state[0], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 613:71]
+    node _T_4807 = and(UInt<1>("h01"), obuf_valid) @[el2_lsu_bus_buffer.scala 614:25]
+    node _T_4808 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 614:50]
+    node _T_4809 = bits(buf_addr[0], 31, 3) @[el2_lsu_bus_buffer.scala 614:70]
+    node _T_4810 = eq(_T_4808, _T_4809) @[el2_lsu_bus_buffer.scala 614:56]
+    node _T_4811 = and(_T_4807, _T_4810) @[el2_lsu_bus_buffer.scala 614:38]
+    node _T_4812 = eq(obuf_tag0, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:92]
+    node _T_4813 = eq(obuf_tag1, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:126]
+    node _T_4814 = and(obuf_merge, _T_4813) @[el2_lsu_bus_buffer.scala 614:114]
+    node _T_4815 = or(_T_4812, _T_4814) @[el2_lsu_bus_buffer.scala 614:100]
+    node _T_4816 = eq(_T_4815, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:80]
+    node _T_4817 = and(_T_4811, _T_4816) @[el2_lsu_bus_buffer.scala 614:78]
+    node _T_4818 = eq(buf_state[1], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 613:71]
+    node _T_4819 = and(UInt<1>("h01"), obuf_valid) @[el2_lsu_bus_buffer.scala 614:25]
+    node _T_4820 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 614:50]
+    node _T_4821 = bits(buf_addr[1], 31, 3) @[el2_lsu_bus_buffer.scala 614:70]
+    node _T_4822 = eq(_T_4820, _T_4821) @[el2_lsu_bus_buffer.scala 614:56]
+    node _T_4823 = and(_T_4819, _T_4822) @[el2_lsu_bus_buffer.scala 614:38]
+    node _T_4824 = eq(obuf_tag0, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 614:92]
+    node _T_4825 = eq(obuf_tag1, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 614:126]
+    node _T_4826 = and(obuf_merge, _T_4825) @[el2_lsu_bus_buffer.scala 614:114]
+    node _T_4827 = or(_T_4824, _T_4826) @[el2_lsu_bus_buffer.scala 614:100]
+    node _T_4828 = eq(_T_4827, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:80]
+    node _T_4829 = and(_T_4823, _T_4828) @[el2_lsu_bus_buffer.scala 614:78]
+    node _T_4830 = eq(buf_state[2], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 613:71]
+    node _T_4831 = and(UInt<1>("h01"), obuf_valid) @[el2_lsu_bus_buffer.scala 614:25]
+    node _T_4832 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 614:50]
+    node _T_4833 = bits(buf_addr[2], 31, 3) @[el2_lsu_bus_buffer.scala 614:70]
+    node _T_4834 = eq(_T_4832, _T_4833) @[el2_lsu_bus_buffer.scala 614:56]
+    node _T_4835 = and(_T_4831, _T_4834) @[el2_lsu_bus_buffer.scala 614:38]
+    node _T_4836 = eq(obuf_tag0, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 614:92]
+    node _T_4837 = eq(obuf_tag1, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 614:126]
+    node _T_4838 = and(obuf_merge, _T_4837) @[el2_lsu_bus_buffer.scala 614:114]
+    node _T_4839 = or(_T_4836, _T_4838) @[el2_lsu_bus_buffer.scala 614:100]
+    node _T_4840 = eq(_T_4839, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:80]
+    node _T_4841 = and(_T_4835, _T_4840) @[el2_lsu_bus_buffer.scala 614:78]
+    node _T_4842 = eq(buf_state[3], UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 613:71]
+    node _T_4843 = and(UInt<1>("h01"), obuf_valid) @[el2_lsu_bus_buffer.scala 614:25]
+    node _T_4844 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 614:50]
+    node _T_4845 = bits(buf_addr[3], 31, 3) @[el2_lsu_bus_buffer.scala 614:70]
+    node _T_4846 = eq(_T_4844, _T_4845) @[el2_lsu_bus_buffer.scala 614:56]
+    node _T_4847 = and(_T_4843, _T_4846) @[el2_lsu_bus_buffer.scala 614:38]
+    node _T_4848 = eq(obuf_tag0, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 614:92]
+    node _T_4849 = eq(obuf_tag1, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 614:126]
+    node _T_4850 = and(obuf_merge, _T_4849) @[el2_lsu_bus_buffer.scala 614:114]
+    node _T_4851 = or(_T_4848, _T_4850) @[el2_lsu_bus_buffer.scala 614:100]
+    node _T_4852 = eq(_T_4851, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 614:80]
+    node _T_4853 = and(_T_4847, _T_4852) @[el2_lsu_bus_buffer.scala 614:78]
+    node _T_4854 = mux(_T_4806, _T_4817, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4855 = mux(_T_4818, _T_4829, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4856 = mux(_T_4830, _T_4841, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4857 = mux(_T_4842, _T_4853, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4858 = or(_T_4854, _T_4855) @[Mux.scala 27:72]
+    node _T_4859 = or(_T_4858, _T_4856) @[Mux.scala 27:72]
     node _T_4860 = or(_T_4859, _T_4857) @[Mux.scala 27:72]
-    node _T_4861 = or(_T_4860, _T_4858) @[Mux.scala 27:72]
-    wire _T_4862 : UInt<1> @[Mux.scala 27:72]
-    _T_4862 <= _T_4861 @[Mux.scala 27:72]
-    bus_addr_match_pending <= _T_4862 @[el2_lsu_bus_buffer.scala 613:26]
-    node _T_4863 = or(obuf_cmd_done, obuf_data_done) @[el2_lsu_bus_buffer.scala 616:54]
-    node _T_4864 = mux(obuf_cmd_done, io.lsu_axi_wready, io.lsu_axi_awready) @[el2_lsu_bus_buffer.scala 616:75]
-    node _T_4865 = and(io.lsu_axi_awready, io.lsu_axi_awready) @[el2_lsu_bus_buffer.scala 616:150]
-    node _T_4866 = mux(_T_4863, _T_4864, _T_4865) @[el2_lsu_bus_buffer.scala 616:39]
-    node _T_4867 = mux(obuf_write, _T_4866, io.lsu_axi_arready) @[el2_lsu_bus_buffer.scala 616:23]
-    bus_cmd_ready <= _T_4867 @[el2_lsu_bus_buffer.scala 616:17]
-    node _T_4868 = and(io.lsu_axi_awvalid, io.lsu_axi_awready) @[el2_lsu_bus_buffer.scala 617:39]
-    bus_wcmd_sent <= _T_4868 @[el2_lsu_bus_buffer.scala 617:17]
-    node _T_4869 = and(io.lsu_axi_wvalid, io.lsu_axi_wready) @[el2_lsu_bus_buffer.scala 618:39]
-    bus_wdata_sent <= _T_4869 @[el2_lsu_bus_buffer.scala 618:18]
-    node _T_4870 = or(obuf_cmd_done, bus_wcmd_sent) @[el2_lsu_bus_buffer.scala 619:35]
-    node _T_4871 = or(obuf_data_done, bus_wdata_sent) @[el2_lsu_bus_buffer.scala 619:70]
-    node _T_4872 = and(_T_4870, _T_4871) @[el2_lsu_bus_buffer.scala 619:52]
-    node _T_4873 = and(io.lsu_axi_arvalid, io.lsu_axi_arready) @[el2_lsu_bus_buffer.scala 619:111]
-    node _T_4874 = or(_T_4872, _T_4873) @[el2_lsu_bus_buffer.scala 619:89]
-    bus_cmd_sent <= _T_4874 @[el2_lsu_bus_buffer.scala 619:16]
-    node _T_4875 = and(io.lsu_axi_rvalid, io.lsu_axi_rready) @[el2_lsu_bus_buffer.scala 620:37]
-    bus_rsp_read <= _T_4875 @[el2_lsu_bus_buffer.scala 620:16]
-    node _T_4876 = and(io.lsu_axi_bvalid, io.lsu_axi_bready) @[el2_lsu_bus_buffer.scala 621:38]
-    bus_rsp_write <= _T_4876 @[el2_lsu_bus_buffer.scala 621:17]
+    wire _T_4861 : UInt<1> @[Mux.scala 27:72]
+    _T_4861 <= _T_4860 @[Mux.scala 27:72]
+    bus_addr_match_pending <= _T_4861 @[el2_lsu_bus_buffer.scala 613:26]
+    node _T_4862 = or(obuf_cmd_done, obuf_data_done) @[el2_lsu_bus_buffer.scala 616:54]
+    node _T_4863 = mux(obuf_cmd_done, io.lsu_axi_wready, io.lsu_axi_awready) @[el2_lsu_bus_buffer.scala 616:75]
+    node _T_4864 = and(io.lsu_axi_awready, io.lsu_axi_awready) @[el2_lsu_bus_buffer.scala 616:150]
+    node _T_4865 = mux(_T_4862, _T_4863, _T_4864) @[el2_lsu_bus_buffer.scala 616:39]
+    node _T_4866 = mux(obuf_write, _T_4865, io.lsu_axi_arready) @[el2_lsu_bus_buffer.scala 616:23]
+    bus_cmd_ready <= _T_4866 @[el2_lsu_bus_buffer.scala 616:17]
+    node _T_4867 = and(io.lsu_axi_awvalid, io.lsu_axi_awready) @[el2_lsu_bus_buffer.scala 617:39]
+    bus_wcmd_sent <= _T_4867 @[el2_lsu_bus_buffer.scala 617:17]
+    node _T_4868 = and(io.lsu_axi_wvalid, io.lsu_axi_wready) @[el2_lsu_bus_buffer.scala 618:39]
+    bus_wdata_sent <= _T_4868 @[el2_lsu_bus_buffer.scala 618:18]
+    node _T_4869 = or(obuf_cmd_done, bus_wcmd_sent) @[el2_lsu_bus_buffer.scala 619:35]
+    node _T_4870 = or(obuf_data_done, bus_wdata_sent) @[el2_lsu_bus_buffer.scala 619:70]
+    node _T_4871 = and(_T_4869, _T_4870) @[el2_lsu_bus_buffer.scala 619:52]
+    node _T_4872 = and(io.lsu_axi_arvalid, io.lsu_axi_arready) @[el2_lsu_bus_buffer.scala 619:111]
+    node _T_4873 = or(_T_4871, _T_4872) @[el2_lsu_bus_buffer.scala 619:89]
+    bus_cmd_sent <= _T_4873 @[el2_lsu_bus_buffer.scala 619:16]
+    node _T_4874 = and(io.lsu_axi_rvalid, io.lsu_axi_rready) @[el2_lsu_bus_buffer.scala 620:37]
+    bus_rsp_read <= _T_4874 @[el2_lsu_bus_buffer.scala 620:16]
+    node _T_4875 = and(io.lsu_axi_bvalid, io.lsu_axi_bready) @[el2_lsu_bus_buffer.scala 621:38]
+    bus_rsp_write <= _T_4875 @[el2_lsu_bus_buffer.scala 621:17]
     bus_rsp_read_tag <= io.lsu_axi_rid @[el2_lsu_bus_buffer.scala 622:20]
     bus_rsp_write_tag <= io.lsu_axi_bid @[el2_lsu_bus_buffer.scala 623:21]
-    node _T_4877 = neq(io.lsu_axi_bresp, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 624:60]
-    node _T_4878 = and(bus_rsp_write, _T_4877) @[el2_lsu_bus_buffer.scala 624:40]
-    bus_rsp_write_error <= _T_4878 @[el2_lsu_bus_buffer.scala 624:23]
-    node _T_4879 = neq(io.lsu_axi_bresp, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 625:58]
-    node _T_4880 = and(bus_rsp_read, _T_4879) @[el2_lsu_bus_buffer.scala 625:38]
-    bus_rsp_read_error <= _T_4880 @[el2_lsu_bus_buffer.scala 625:22]
+    node _T_4876 = neq(io.lsu_axi_bresp, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 624:60]
+    node _T_4877 = and(bus_rsp_write, _T_4876) @[el2_lsu_bus_buffer.scala 624:40]
+    bus_rsp_write_error <= _T_4877 @[el2_lsu_bus_buffer.scala 624:23]
+    node _T_4878 = neq(io.lsu_axi_bresp, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 625:58]
+    node _T_4879 = and(bus_rsp_read, _T_4878) @[el2_lsu_bus_buffer.scala 625:38]
+    bus_rsp_read_error <= _T_4879 @[el2_lsu_bus_buffer.scala 625:22]
     bus_rsp_rdata <= io.lsu_axi_rdata @[el2_lsu_bus_buffer.scala 626:17]
-    node _T_4881 = and(obuf_valid, obuf_write) @[el2_lsu_bus_buffer.scala 629:36]
-    node _T_4882 = eq(obuf_cmd_done, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 629:51]
-    node _T_4883 = and(_T_4881, _T_4882) @[el2_lsu_bus_buffer.scala 629:49]
-    node _T_4884 = eq(bus_addr_match_pending, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 629:68]
-    node _T_4885 = and(_T_4883, _T_4884) @[el2_lsu_bus_buffer.scala 629:66]
-    io.lsu_axi_awvalid <= _T_4885 @[el2_lsu_bus_buffer.scala 629:22]
+    node _T_4880 = and(obuf_valid, obuf_write) @[el2_lsu_bus_buffer.scala 629:36]
+    node _T_4881 = eq(obuf_cmd_done, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 629:51]
+    node _T_4882 = and(_T_4880, _T_4881) @[el2_lsu_bus_buffer.scala 629:49]
+    node _T_4883 = eq(bus_addr_match_pending, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 629:68]
+    node _T_4884 = and(_T_4882, _T_4883) @[el2_lsu_bus_buffer.scala 629:66]
+    io.lsu_axi_awvalid <= _T_4884 @[el2_lsu_bus_buffer.scala 629:22]
     io.lsu_axi_awid <= obuf_tag0 @[el2_lsu_bus_buffer.scala 630:19]
-    node _T_4886 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 631:69]
-    node _T_4887 = cat(_T_4886, UInt<3>("h00")) @[Cat.scala 29:58]
-    node _T_4888 = mux(obuf_sideeffect, obuf_addr, _T_4887) @[el2_lsu_bus_buffer.scala 631:27]
-    io.lsu_axi_awaddr <= _T_4888 @[el2_lsu_bus_buffer.scala 631:21]
-    node _T_4889 = cat(UInt<1>("h00"), obuf_sz) @[Cat.scala 29:58]
-    node _T_4890 = mux(obuf_sideeffect, _T_4889, UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 632:27]
-    io.lsu_axi_awsize <= _T_4890 @[el2_lsu_bus_buffer.scala 632:21]
+    node _T_4885 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 631:69]
+    node _T_4886 = cat(_T_4885, UInt<3>("h00")) @[Cat.scala 29:58]
+    node _T_4887 = mux(obuf_sideeffect, obuf_addr, _T_4886) @[el2_lsu_bus_buffer.scala 631:27]
+    io.lsu_axi_awaddr <= _T_4887 @[el2_lsu_bus_buffer.scala 631:21]
+    node _T_4888 = cat(UInt<1>("h00"), obuf_sz) @[Cat.scala 29:58]
+    node _T_4889 = mux(obuf_sideeffect, _T_4888, UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 632:27]
+    io.lsu_axi_awsize <= _T_4889 @[el2_lsu_bus_buffer.scala 632:21]
     io.lsu_axi_awprot <= UInt<1>("h00") @[el2_lsu_bus_buffer.scala 633:21]
-    node _T_4891 = mux(obuf_sideeffect, UInt<1>("h00"), UInt<4>("h0f")) @[el2_lsu_bus_buffer.scala 634:28]
-    io.lsu_axi_awcache <= _T_4891 @[el2_lsu_bus_buffer.scala 634:22]
-    node _T_4892 = bits(obuf_addr, 31, 28) @[el2_lsu_bus_buffer.scala 635:35]
-    io.lsu_axi_awregion <= _T_4892 @[el2_lsu_bus_buffer.scala 635:23]
+    node _T_4890 = mux(obuf_sideeffect, UInt<1>("h00"), UInt<4>("h0f")) @[el2_lsu_bus_buffer.scala 634:28]
+    io.lsu_axi_awcache <= _T_4890 @[el2_lsu_bus_buffer.scala 634:22]
+    node _T_4891 = bits(obuf_addr, 31, 28) @[el2_lsu_bus_buffer.scala 635:35]
+    io.lsu_axi_awregion <= _T_4891 @[el2_lsu_bus_buffer.scala 635:23]
     io.lsu_axi_awlen <= UInt<1>("h00") @[el2_lsu_bus_buffer.scala 636:20]
     io.lsu_axi_awburst <= UInt<2>("h01") @[el2_lsu_bus_buffer.scala 637:22]
     io.lsu_axi_awqos <= UInt<1>("h00") @[el2_lsu_bus_buffer.scala 638:20]
     io.lsu_axi_awlock <= UInt<1>("h00") @[el2_lsu_bus_buffer.scala 639:21]
-    node _T_4893 = and(obuf_valid, obuf_write) @[el2_lsu_bus_buffer.scala 641:35]
-    node _T_4894 = eq(obuf_data_done, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 641:50]
-    node _T_4895 = and(_T_4893, _T_4894) @[el2_lsu_bus_buffer.scala 641:48]
-    node _T_4896 = eq(bus_addr_match_pending, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 641:68]
-    node _T_4897 = and(_T_4895, _T_4896) @[el2_lsu_bus_buffer.scala 641:66]
-    io.lsu_axi_wvalid <= _T_4897 @[el2_lsu_bus_buffer.scala 641:21]
-    node _T_4898 = bits(obuf_write, 0, 0) @[Bitwise.scala 72:15]
-    node _T_4899 = mux(_T_4898, UInt<8>("h0ff"), UInt<8>("h00")) @[Bitwise.scala 72:12]
-    node _T_4900 = and(obuf_byteen, _T_4899) @[el2_lsu_bus_buffer.scala 642:35]
-    io.lsu_axi_wstrb <= _T_4900 @[el2_lsu_bus_buffer.scala 642:20]
+    node _T_4892 = and(obuf_valid, obuf_write) @[el2_lsu_bus_buffer.scala 641:35]
+    node _T_4893 = eq(obuf_data_done, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 641:50]
+    node _T_4894 = and(_T_4892, _T_4893) @[el2_lsu_bus_buffer.scala 641:48]
+    node _T_4895 = eq(bus_addr_match_pending, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 641:68]
+    node _T_4896 = and(_T_4894, _T_4895) @[el2_lsu_bus_buffer.scala 641:66]
+    io.lsu_axi_wvalid <= _T_4896 @[el2_lsu_bus_buffer.scala 641:21]
+    node _T_4897 = bits(obuf_write, 0, 0) @[Bitwise.scala 72:15]
+    node _T_4898 = mux(_T_4897, UInt<8>("h0ff"), UInt<8>("h00")) @[Bitwise.scala 72:12]
+    node _T_4899 = and(obuf_byteen, _T_4898) @[el2_lsu_bus_buffer.scala 642:35]
+    io.lsu_axi_wstrb <= _T_4899 @[el2_lsu_bus_buffer.scala 642:20]
     io.lsu_axi_wdata <= obuf_data @[el2_lsu_bus_buffer.scala 643:20]
     io.lsu_axi_wlast <= UInt<1>("h01") @[el2_lsu_bus_buffer.scala 644:20]
-    node _T_4901 = eq(obuf_write, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 646:38]
-    node _T_4902 = and(obuf_valid, _T_4901) @[el2_lsu_bus_buffer.scala 646:36]
-    node _T_4903 = eq(obuf_nosend, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 646:52]
-    node _T_4904 = and(_T_4902, _T_4903) @[el2_lsu_bus_buffer.scala 646:50]
-    node _T_4905 = eq(bus_addr_match_pending, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 646:67]
-    node _T_4906 = and(_T_4904, _T_4905) @[el2_lsu_bus_buffer.scala 646:65]
-    io.lsu_axi_arvalid <= _T_4906 @[el2_lsu_bus_buffer.scala 646:22]
+    node _T_4900 = eq(obuf_write, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 646:38]
+    node _T_4901 = and(obuf_valid, _T_4900) @[el2_lsu_bus_buffer.scala 646:36]
+    node _T_4902 = eq(obuf_nosend, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 646:52]
+    node _T_4903 = and(_T_4901, _T_4902) @[el2_lsu_bus_buffer.scala 646:50]
+    node _T_4904 = eq(bus_addr_match_pending, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 646:67]
+    node _T_4905 = and(_T_4903, _T_4904) @[el2_lsu_bus_buffer.scala 646:65]
+    io.lsu_axi_arvalid <= _T_4905 @[el2_lsu_bus_buffer.scala 646:22]
     io.lsu_axi_arid <= obuf_tag0 @[el2_lsu_bus_buffer.scala 647:19]
-    node _T_4907 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 648:69]
-    node _T_4908 = cat(_T_4907, UInt<3>("h00")) @[Cat.scala 29:58]
-    node _T_4909 = mux(obuf_sideeffect, obuf_addr, _T_4908) @[el2_lsu_bus_buffer.scala 648:27]
-    io.lsu_axi_araddr <= _T_4909 @[el2_lsu_bus_buffer.scala 648:21]
-    node _T_4910 = cat(UInt<1>("h00"), obuf_sz) @[Cat.scala 29:58]
-    node _T_4911 = mux(obuf_sideeffect, _T_4910, UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 649:27]
-    io.lsu_axi_arsize <= _T_4911 @[el2_lsu_bus_buffer.scala 649:21]
+    node _T_4906 = bits(obuf_addr, 31, 3) @[el2_lsu_bus_buffer.scala 648:69]
+    node _T_4907 = cat(_T_4906, UInt<3>("h00")) @[Cat.scala 29:58]
+    node _T_4908 = mux(obuf_sideeffect, obuf_addr, _T_4907) @[el2_lsu_bus_buffer.scala 648:27]
+    io.lsu_axi_araddr <= _T_4908 @[el2_lsu_bus_buffer.scala 648:21]
+    node _T_4909 = cat(UInt<1>("h00"), obuf_sz) @[Cat.scala 29:58]
+    node _T_4910 = mux(obuf_sideeffect, _T_4909, UInt<3>("h03")) @[el2_lsu_bus_buffer.scala 649:27]
+    io.lsu_axi_arsize <= _T_4910 @[el2_lsu_bus_buffer.scala 649:21]
     io.lsu_axi_arprot <= UInt<1>("h00") @[el2_lsu_bus_buffer.scala 650:21]
-    node _T_4912 = mux(obuf_sideeffect, UInt<4>("h00"), UInt<4>("h0f")) @[el2_lsu_bus_buffer.scala 651:28]
-    io.lsu_axi_arcache <= _T_4912 @[el2_lsu_bus_buffer.scala 651:22]
-    node _T_4913 = bits(obuf_addr, 31, 28) @[el2_lsu_bus_buffer.scala 652:35]
-    io.lsu_axi_arregion <= _T_4913 @[el2_lsu_bus_buffer.scala 652:23]
+    node _T_4911 = mux(obuf_sideeffect, UInt<4>("h00"), UInt<4>("h0f")) @[el2_lsu_bus_buffer.scala 651:28]
+    io.lsu_axi_arcache <= _T_4911 @[el2_lsu_bus_buffer.scala 651:22]
+    node _T_4912 = bits(obuf_addr, 31, 28) @[el2_lsu_bus_buffer.scala 652:35]
+    io.lsu_axi_arregion <= _T_4912 @[el2_lsu_bus_buffer.scala 652:23]
     io.lsu_axi_arlen <= UInt<1>("h00") @[el2_lsu_bus_buffer.scala 653:20]
     io.lsu_axi_arburst <= UInt<2>("h01") @[el2_lsu_bus_buffer.scala 654:22]
     io.lsu_axi_arqos <= UInt<1>("h00") @[el2_lsu_bus_buffer.scala 655:20]
     io.lsu_axi_arlock <= UInt<1>("h00") @[el2_lsu_bus_buffer.scala 656:21]
     io.lsu_axi_bready <= UInt<1>("h01") @[el2_lsu_bus_buffer.scala 657:21]
     io.lsu_axi_rready <= UInt<1>("h01") @[el2_lsu_bus_buffer.scala 658:21]
-    node _T_4914 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 659:81]
-    node _T_4915 = bits(buf_error, 0, 0) @[el2_lsu_bus_buffer.scala 659:125]
-    node _T_4916 = and(io.lsu_bus_clk_en_q, _T_4915) @[el2_lsu_bus_buffer.scala 659:114]
-    node _T_4917 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 659:140]
-    node _T_4918 = and(_T_4916, _T_4917) @[el2_lsu_bus_buffer.scala 659:129]
-    node _T_4919 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 659:81]
-    node _T_4920 = bits(buf_error, 1, 1) @[el2_lsu_bus_buffer.scala 659:125]
-    node _T_4921 = and(io.lsu_bus_clk_en_q, _T_4920) @[el2_lsu_bus_buffer.scala 659:114]
-    node _T_4922 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 659:140]
-    node _T_4923 = and(_T_4921, _T_4922) @[el2_lsu_bus_buffer.scala 659:129]
-    node _T_4924 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 659:81]
-    node _T_4925 = bits(buf_error, 2, 2) @[el2_lsu_bus_buffer.scala 659:125]
-    node _T_4926 = and(io.lsu_bus_clk_en_q, _T_4925) @[el2_lsu_bus_buffer.scala 659:114]
-    node _T_4927 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 659:140]
-    node _T_4928 = and(_T_4926, _T_4927) @[el2_lsu_bus_buffer.scala 659:129]
-    node _T_4929 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 659:81]
-    node _T_4930 = bits(buf_error, 3, 3) @[el2_lsu_bus_buffer.scala 659:125]
-    node _T_4931 = and(io.lsu_bus_clk_en_q, _T_4930) @[el2_lsu_bus_buffer.scala 659:114]
-    node _T_4932 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 659:140]
-    node _T_4933 = and(_T_4931, _T_4932) @[el2_lsu_bus_buffer.scala 659:129]
-    node _T_4934 = mux(_T_4914, _T_4918, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4935 = mux(_T_4919, _T_4923, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4936 = mux(_T_4924, _T_4928, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4937 = mux(_T_4929, _T_4933, UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4938 = or(_T_4934, _T_4935) @[Mux.scala 27:72]
+    node _T_4913 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 659:81]
+    node _T_4914 = bits(buf_error, 0, 0) @[el2_lsu_bus_buffer.scala 659:125]
+    node _T_4915 = and(io.lsu_bus_clk_en_q, _T_4914) @[el2_lsu_bus_buffer.scala 659:114]
+    node _T_4916 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 659:140]
+    node _T_4917 = and(_T_4915, _T_4916) @[el2_lsu_bus_buffer.scala 659:129]
+    node _T_4918 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 659:81]
+    node _T_4919 = bits(buf_error, 1, 1) @[el2_lsu_bus_buffer.scala 659:125]
+    node _T_4920 = and(io.lsu_bus_clk_en_q, _T_4919) @[el2_lsu_bus_buffer.scala 659:114]
+    node _T_4921 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 659:140]
+    node _T_4922 = and(_T_4920, _T_4921) @[el2_lsu_bus_buffer.scala 659:129]
+    node _T_4923 = eq(buf_state[2], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 659:81]
+    node _T_4924 = bits(buf_error, 2, 2) @[el2_lsu_bus_buffer.scala 659:125]
+    node _T_4925 = and(io.lsu_bus_clk_en_q, _T_4924) @[el2_lsu_bus_buffer.scala 659:114]
+    node _T_4926 = bits(buf_write, 2, 2) @[el2_lsu_bus_buffer.scala 659:140]
+    node _T_4927 = and(_T_4925, _T_4926) @[el2_lsu_bus_buffer.scala 659:129]
+    node _T_4928 = eq(buf_state[3], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 659:81]
+    node _T_4929 = bits(buf_error, 3, 3) @[el2_lsu_bus_buffer.scala 659:125]
+    node _T_4930 = and(io.lsu_bus_clk_en_q, _T_4929) @[el2_lsu_bus_buffer.scala 659:114]
+    node _T_4931 = bits(buf_write, 3, 3) @[el2_lsu_bus_buffer.scala 659:140]
+    node _T_4932 = and(_T_4930, _T_4931) @[el2_lsu_bus_buffer.scala 659:129]
+    node _T_4933 = mux(_T_4913, _T_4917, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4934 = mux(_T_4918, _T_4922, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4935 = mux(_T_4923, _T_4927, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4936 = mux(_T_4928, _T_4932, UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4937 = or(_T_4933, _T_4934) @[Mux.scala 27:72]
+    node _T_4938 = or(_T_4937, _T_4935) @[Mux.scala 27:72]
     node _T_4939 = or(_T_4938, _T_4936) @[Mux.scala 27:72]
-    node _T_4940 = or(_T_4939, _T_4937) @[Mux.scala 27:72]
-    wire _T_4941 : UInt<1> @[Mux.scala 27:72]
-    _T_4941 <= _T_4940 @[Mux.scala 27:72]
-    io.lsu_imprecise_error_store_any <= _T_4941 @[el2_lsu_bus_buffer.scala 659:36]
-    node _T_4942 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 660:87]
-    node _T_4943 = bits(buf_error, 0, 0) @[el2_lsu_bus_buffer.scala 660:109]
-    node _T_4944 = and(_T_4942, _T_4943) @[el2_lsu_bus_buffer.scala 660:98]
-    node _T_4945 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 660:124]
-    node _T_4946 = and(_T_4944, _T_4945) @[el2_lsu_bus_buffer.scala 660:113]
-    node _T_4947 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 660:87]
-    node _T_4948 = bits(buf_error, 1, 1) @[el2_lsu_bus_buffer.scala 660:109]
-    node _T_4949 = and(_T_4947, _T_4948) @[el2_lsu_bus_buffer.scala 660:98]
-    node _T_4950 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 660:124]
-    node _T_4951 = and(_T_4949, _T_4950) @[el2_lsu_bus_buffer.scala 660:113]
-    node _T_4952 = mux(_T_4946, UInt<1>("h00"), UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4953 = mux(_T_4951, UInt<1>("h01"), UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4954 = or(_T_4952, _T_4953) @[Mux.scala 27:72]
+    wire _T_4940 : UInt<1> @[Mux.scala 27:72]
+    _T_4940 <= _T_4939 @[Mux.scala 27:72]
+    io.lsu_imprecise_error_store_any <= _T_4940 @[el2_lsu_bus_buffer.scala 659:36]
+    node _T_4941 = eq(buf_state[0], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 660:87]
+    node _T_4942 = bits(buf_error, 0, 0) @[el2_lsu_bus_buffer.scala 660:109]
+    node _T_4943 = and(_T_4941, _T_4942) @[el2_lsu_bus_buffer.scala 660:98]
+    node _T_4944 = bits(buf_write, 0, 0) @[el2_lsu_bus_buffer.scala 660:124]
+    node _T_4945 = and(_T_4943, _T_4944) @[el2_lsu_bus_buffer.scala 660:113]
+    node _T_4946 = eq(buf_state[1], UInt<3>("h06")) @[el2_lsu_bus_buffer.scala 660:87]
+    node _T_4947 = bits(buf_error, 1, 1) @[el2_lsu_bus_buffer.scala 660:109]
+    node _T_4948 = and(_T_4946, _T_4947) @[el2_lsu_bus_buffer.scala 660:98]
+    node _T_4949 = bits(buf_write, 1, 1) @[el2_lsu_bus_buffer.scala 660:124]
+    node _T_4950 = and(_T_4948, _T_4949) @[el2_lsu_bus_buffer.scala 660:113]
+    node _T_4951 = mux(_T_4945, UInt<1>("h00"), UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4952 = mux(_T_4950, UInt<1>("h01"), UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4953 = or(_T_4951, _T_4952) @[Mux.scala 27:72]
     wire lsu_imprecise_error_store_tag : UInt<1> @[Mux.scala 27:72]
-    lsu_imprecise_error_store_tag <= _T_4954 @[Mux.scala 27:72]
-    node _T_4955 = eq(io.lsu_imprecise_error_store_any, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 662:72]
-    node _T_4956 = and(io.lsu_nonblock_load_data_error, _T_4955) @[el2_lsu_bus_buffer.scala 662:70]
-    io.lsu_imprecise_error_load_any <= _T_4956 @[el2_lsu_bus_buffer.scala 662:35]
-    node _T_4957 = eq(lsu_imprecise_error_store_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4958 = eq(lsu_imprecise_error_store_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4959 = mux(_T_4957, buf_addr[0], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4960 = mux(_T_4958, buf_addr[1], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4961 = or(_T_4959, _T_4960) @[Mux.scala 27:72]
-    wire _T_4962 : UInt<32> @[Mux.scala 27:72]
-    _T_4962 <= _T_4961 @[Mux.scala 27:72]
-    node _T_4963 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4964 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4965 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4966 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 120:123]
-    node _T_4967 = mux(_T_4963, buf_addr[0], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4968 = mux(_T_4964, buf_addr[1], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4969 = mux(_T_4965, buf_addr[2], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4970 = mux(_T_4966, buf_addr[3], UInt<1>("h00")) @[Mux.scala 27:72]
-    node _T_4971 = or(_T_4967, _T_4968) @[Mux.scala 27:72]
+    lsu_imprecise_error_store_tag <= _T_4953 @[Mux.scala 27:72]
+    node _T_4954 = eq(io.lsu_imprecise_error_store_any, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 662:72]
+    node _T_4955 = and(io.lsu_nonblock_load_data_error, _T_4954) @[el2_lsu_bus_buffer.scala 662:70]
+    io.lsu_imprecise_error_load_any <= _T_4955 @[el2_lsu_bus_buffer.scala 662:35]
+    node _T_4956 = eq(lsu_imprecise_error_store_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4957 = eq(lsu_imprecise_error_store_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4958 = mux(_T_4956, buf_addr[0], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4959 = mux(_T_4957, buf_addr[1], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4960 = or(_T_4958, _T_4959) @[Mux.scala 27:72]
+    wire _T_4961 : UInt<32> @[Mux.scala 27:72]
+    _T_4961 <= _T_4960 @[Mux.scala 27:72]
+    node _T_4962 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4963 = eq(io.lsu_nonblock_load_data_tag, UInt<1>("h01")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4964 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h02")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4965 = eq(io.lsu_nonblock_load_data_tag, UInt<2>("h03")) @[el2_lsu_bus_buffer.scala 120:123]
+    node _T_4966 = mux(_T_4962, buf_addr[0], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4967 = mux(_T_4963, buf_addr[1], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4968 = mux(_T_4964, buf_addr[2], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4969 = mux(_T_4965, buf_addr[3], UInt<1>("h00")) @[Mux.scala 27:72]
+    node _T_4970 = or(_T_4966, _T_4967) @[Mux.scala 27:72]
+    node _T_4971 = or(_T_4970, _T_4968) @[Mux.scala 27:72]
     node _T_4972 = or(_T_4971, _T_4969) @[Mux.scala 27:72]
-    node _T_4973 = or(_T_4972, _T_4970) @[Mux.scala 27:72]
-    wire _T_4974 : UInt<32> @[Mux.scala 27:72]
-    _T_4974 <= _T_4973 @[Mux.scala 27:72]
-    node _T_4975 = mux(io.lsu_imprecise_error_store_any, _T_4962, _T_4974) @[el2_lsu_bus_buffer.scala 663:41]
-    io.lsu_imprecise_error_addr_any <= _T_4975 @[el2_lsu_bus_buffer.scala 663:35]
+    wire _T_4973 : UInt<32> @[Mux.scala 27:72]
+    _T_4973 <= _T_4972 @[Mux.scala 27:72]
+    node _T_4974 = mux(io.lsu_imprecise_error_store_any, _T_4961, _T_4973) @[el2_lsu_bus_buffer.scala 663:41]
+    io.lsu_imprecise_error_addr_any <= _T_4974 @[el2_lsu_bus_buffer.scala 663:35]
     lsu_bus_cntr_overflow <= UInt<1>("h00") @[el2_lsu_bus_buffer.scala 664:25]
     io.lsu_bus_idle_any <= UInt<1>("h01") @[el2_lsu_bus_buffer.scala 666:23]
-    node _T_4976 = and(io.lsu_axi_awvalid, io.lsu_axi_awready) @[el2_lsu_bus_buffer.scala 669:46]
-    node _T_4977 = and(io.lsu_axi_wvalid, io.lsu_axi_wready) @[el2_lsu_bus_buffer.scala 669:89]
-    node _T_4978 = or(_T_4976, _T_4977) @[el2_lsu_bus_buffer.scala 669:68]
-    node _T_4979 = and(io.lsu_axi_arvalid, io.lsu_axi_arready) @[el2_lsu_bus_buffer.scala 669:132]
-    node _T_4980 = or(_T_4978, _T_4979) @[el2_lsu_bus_buffer.scala 669:110]
-    io.lsu_pmu_bus_trxn <= _T_4980 @[el2_lsu_bus_buffer.scala 669:23]
-    node _T_4981 = and(io.lsu_busreq_r, io.ldst_dual_r) @[el2_lsu_bus_buffer.scala 670:48]
-    node _T_4982 = and(_T_4981, io.lsu_commit_r) @[el2_lsu_bus_buffer.scala 670:65]
-    io.lsu_pmu_bus_misaligned <= _T_4982 @[el2_lsu_bus_buffer.scala 670:29]
-    node _T_4983 = or(io.lsu_imprecise_error_load_any, io.lsu_imprecise_error_store_any) @[el2_lsu_bus_buffer.scala 671:59]
-    io.lsu_pmu_bus_error <= _T_4983 @[el2_lsu_bus_buffer.scala 671:24]
-    node _T_4984 = eq(io.lsu_axi_awready, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 673:48]
-    node _T_4985 = and(io.lsu_axi_awvalid, _T_4984) @[el2_lsu_bus_buffer.scala 673:46]
-    node _T_4986 = eq(io.lsu_axi_wready, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 673:92]
-    node _T_4987 = and(io.lsu_axi_wvalid, _T_4986) @[el2_lsu_bus_buffer.scala 673:90]
-    node _T_4988 = or(_T_4985, _T_4987) @[el2_lsu_bus_buffer.scala 673:69]
-    node _T_4989 = eq(io.lsu_axi_arready, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 673:136]
-    node _T_4990 = and(io.lsu_axi_arvalid, _T_4989) @[el2_lsu_bus_buffer.scala 673:134]
-    node _T_4991 = or(_T_4988, _T_4990) @[el2_lsu_bus_buffer.scala 673:112]
-    io.lsu_pmu_bus_busy <= _T_4991 @[el2_lsu_bus_buffer.scala 673:23]
-    reg _T_4992 : UInt, io.lsu_c2_r_clk with : (reset => (reset, UInt<1>("h00"))) @[el2_lsu_bus_buffer.scala 675:49]
-    _T_4992 <= WrPtr0_m @[el2_lsu_bus_buffer.scala 675:49]
-    WrPtr0_r <= _T_4992 @[el2_lsu_bus_buffer.scala 675:12]
-    reg _T_4993 : UInt, io.lsu_c2_r_clk with : (reset => (reset, UInt<1>("h00"))) @[el2_lsu_bus_buffer.scala 676:49]
-    _T_4993 <= WrPtr1_m @[el2_lsu_bus_buffer.scala 676:49]
-    WrPtr1_r <= _T_4993 @[el2_lsu_bus_buffer.scala 676:12]
-    node _T_4994 = eq(io.flush_r, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 677:75]
-    node _T_4995 = and(io.lsu_busreq_m, _T_4994) @[el2_lsu_bus_buffer.scala 677:73]
-    node _T_4996 = eq(io.ld_full_hit_m, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 677:89]
-    node _T_4997 = and(_T_4995, _T_4996) @[el2_lsu_bus_buffer.scala 677:87]
-    reg _T_4998 : UInt<1>, io.lsu_c2_r_clk with : (reset => (reset, UInt<1>("h00"))) @[el2_lsu_bus_buffer.scala 677:56]
-    _T_4998 <= _T_4997 @[el2_lsu_bus_buffer.scala 677:56]
-    io.lsu_busreq_r <= _T_4998 @[el2_lsu_bus_buffer.scala 677:19]
-    reg _T_4999 : UInt<1>, io.lsu_c2_r_clk with : (reset => (reset, UInt<1>("h00"))) @[el2_lsu_bus_buffer.scala 678:66]
-    _T_4999 <= io.lsu_nonblock_load_valid_m @[el2_lsu_bus_buffer.scala 678:66]
-    lsu_nonblock_load_valid_r <= _T_4999 @[el2_lsu_bus_buffer.scala 678:29]
+    node _T_4975 = and(io.lsu_axi_awvalid, io.lsu_axi_awready) @[el2_lsu_bus_buffer.scala 669:46]
+    node _T_4976 = and(io.lsu_axi_wvalid, io.lsu_axi_wready) @[el2_lsu_bus_buffer.scala 669:89]
+    node _T_4977 = or(_T_4975, _T_4976) @[el2_lsu_bus_buffer.scala 669:68]
+    node _T_4978 = and(io.lsu_axi_arvalid, io.lsu_axi_arready) @[el2_lsu_bus_buffer.scala 669:132]
+    node _T_4979 = or(_T_4977, _T_4978) @[el2_lsu_bus_buffer.scala 669:110]
+    io.lsu_pmu_bus_trxn <= _T_4979 @[el2_lsu_bus_buffer.scala 669:23]
+    node _T_4980 = and(io.lsu_busreq_r, io.ldst_dual_r) @[el2_lsu_bus_buffer.scala 670:48]
+    node _T_4981 = and(_T_4980, io.lsu_commit_r) @[el2_lsu_bus_buffer.scala 670:65]
+    io.lsu_pmu_bus_misaligned <= _T_4981 @[el2_lsu_bus_buffer.scala 670:29]
+    node _T_4982 = or(io.lsu_imprecise_error_load_any, io.lsu_imprecise_error_store_any) @[el2_lsu_bus_buffer.scala 671:59]
+    io.lsu_pmu_bus_error <= _T_4982 @[el2_lsu_bus_buffer.scala 671:24]
+    node _T_4983 = eq(io.lsu_axi_awready, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 673:48]
+    node _T_4984 = and(io.lsu_axi_awvalid, _T_4983) @[el2_lsu_bus_buffer.scala 673:46]
+    node _T_4985 = eq(io.lsu_axi_wready, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 673:92]
+    node _T_4986 = and(io.lsu_axi_wvalid, _T_4985) @[el2_lsu_bus_buffer.scala 673:90]
+    node _T_4987 = or(_T_4984, _T_4986) @[el2_lsu_bus_buffer.scala 673:69]
+    node _T_4988 = eq(io.lsu_axi_arready, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 673:136]
+    node _T_4989 = and(io.lsu_axi_arvalid, _T_4988) @[el2_lsu_bus_buffer.scala 673:134]
+    node _T_4990 = or(_T_4987, _T_4989) @[el2_lsu_bus_buffer.scala 673:112]
+    io.lsu_pmu_bus_busy <= _T_4990 @[el2_lsu_bus_buffer.scala 673:23]
+    reg _T_4991 : UInt, io.lsu_c2_r_clk with : (reset => (reset, UInt<1>("h00"))) @[el2_lsu_bus_buffer.scala 675:49]
+    _T_4991 <= WrPtr0_m @[el2_lsu_bus_buffer.scala 675:49]
+    WrPtr0_r <= _T_4991 @[el2_lsu_bus_buffer.scala 675:12]
+    reg _T_4992 : UInt, io.lsu_c2_r_clk with : (reset => (reset, UInt<1>("h00"))) @[el2_lsu_bus_buffer.scala 676:49]
+    _T_4992 <= WrPtr1_m @[el2_lsu_bus_buffer.scala 676:49]
+    WrPtr1_r <= _T_4992 @[el2_lsu_bus_buffer.scala 676:12]
+    node _T_4993 = eq(io.flush_r, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 677:75]
+    node _T_4994 = and(io.lsu_busreq_m, _T_4993) @[el2_lsu_bus_buffer.scala 677:73]
+    node _T_4995 = eq(io.ld_full_hit_m, UInt<1>("h00")) @[el2_lsu_bus_buffer.scala 677:89]
+    node _T_4996 = and(_T_4994, _T_4995) @[el2_lsu_bus_buffer.scala 677:87]
+    reg _T_4997 : UInt<1>, io.lsu_c2_r_clk with : (reset => (reset, UInt<1>("h00"))) @[el2_lsu_bus_buffer.scala 677:56]
+    _T_4997 <= _T_4996 @[el2_lsu_bus_buffer.scala 677:56]
+    io.lsu_busreq_r <= _T_4997 @[el2_lsu_bus_buffer.scala 677:19]
+    reg _T_4998 : UInt<1>, io.lsu_c2_r_clk with : (reset => (reset, UInt<1>("h00"))) @[el2_lsu_bus_buffer.scala 678:66]
+    _T_4998 <= io.lsu_nonblock_load_valid_m @[el2_lsu_bus_buffer.scala 678:66]
+    lsu_nonblock_load_valid_r <= _T_4998 @[el2_lsu_bus_buffer.scala 678:29]
     
diff --git a/el2_lsu_bus_buffer.v b/el2_lsu_bus_buffer.v
index 6b6a4262..fa389bad 100644
--- a/el2_lsu_bus_buffer.v
+++ b/el2_lsu_bus_buffer.v
@@ -1006,33 +1006,33 @@ module el2_lsu_bus_buffer(
   reg  ibuf_nomerge; // @[Reg.scala 27:20]
   reg  ibuf_unsign; // @[Reg.scala 27:20]
   reg [1:0] ibuf_sz; // @[Reg.scala 27:20]
-  wire  _T_4456 = buf_write[3] & _T_2636; // @[el2_lsu_bus_buffer.scala 581:64]
-  wire  _T_4457 = ~buf_cmd_state_bus_en_3; // @[el2_lsu_bus_buffer.scala 581:91]
-  wire  _T_4458 = _T_4456 & _T_4457; // @[el2_lsu_bus_buffer.scala 581:89]
-  wire  _T_4451 = buf_write[2] & _T_2631; // @[el2_lsu_bus_buffer.scala 581:64]
-  wire  _T_4452 = ~buf_cmd_state_bus_en_2; // @[el2_lsu_bus_buffer.scala 581:91]
-  wire  _T_4453 = _T_4451 & _T_4452; // @[el2_lsu_bus_buffer.scala 581:89]
-  wire [1:0] _T_4459 = _T_4458 + _T_4453; // @[el2_lsu_bus_buffer.scala 581:142]
-  wire  _T_4446 = buf_write[1] & _T_2626; // @[el2_lsu_bus_buffer.scala 581:64]
-  wire  _T_4447 = ~buf_cmd_state_bus_en_1; // @[el2_lsu_bus_buffer.scala 581:91]
-  wire  _T_4448 = _T_4446 & _T_4447; // @[el2_lsu_bus_buffer.scala 581:89]
-  wire [1:0] _GEN_362 = {{1'd0}, _T_4448}; // @[el2_lsu_bus_buffer.scala 581:142]
-  wire [2:0] _T_4460 = _T_4459 + _GEN_362; // @[el2_lsu_bus_buffer.scala 581:142]
-  wire  _T_4441 = buf_write[0] & _T_2621; // @[el2_lsu_bus_buffer.scala 581:64]
-  wire  _T_4442 = ~buf_cmd_state_bus_en_0; // @[el2_lsu_bus_buffer.scala 581:91]
-  wire  _T_4443 = _T_4441 & _T_4442; // @[el2_lsu_bus_buffer.scala 581:89]
-  wire [2:0] _GEN_363 = {{2'd0}, _T_4443}; // @[el2_lsu_bus_buffer.scala 581:142]
-  wire [3:0] buf_numvld_wrcmd_any = _T_4460 + _GEN_363; // @[el2_lsu_bus_buffer.scala 581:142]
+  wire  _T_4455 = buf_write[3] & _T_2636; // @[el2_lsu_bus_buffer.scala 581:64]
+  wire  _T_4456 = ~buf_cmd_state_bus_en_3; // @[el2_lsu_bus_buffer.scala 581:91]
+  wire  _T_4457 = _T_4455 & _T_4456; // @[el2_lsu_bus_buffer.scala 581:89]
+  wire  _T_4450 = buf_write[2] & _T_2631; // @[el2_lsu_bus_buffer.scala 581:64]
+  wire  _T_4451 = ~buf_cmd_state_bus_en_2; // @[el2_lsu_bus_buffer.scala 581:91]
+  wire  _T_4452 = _T_4450 & _T_4451; // @[el2_lsu_bus_buffer.scala 581:89]
+  wire [1:0] _T_4458 = _T_4457 + _T_4452; // @[el2_lsu_bus_buffer.scala 581:142]
+  wire  _T_4445 = buf_write[1] & _T_2626; // @[el2_lsu_bus_buffer.scala 581:64]
+  wire  _T_4446 = ~buf_cmd_state_bus_en_1; // @[el2_lsu_bus_buffer.scala 581:91]
+  wire  _T_4447 = _T_4445 & _T_4446; // @[el2_lsu_bus_buffer.scala 581:89]
+  wire [1:0] _GEN_362 = {{1'd0}, _T_4447}; // @[el2_lsu_bus_buffer.scala 581:142]
+  wire [2:0] _T_4459 = _T_4458 + _GEN_362; // @[el2_lsu_bus_buffer.scala 581:142]
+  wire  _T_4440 = buf_write[0] & _T_2621; // @[el2_lsu_bus_buffer.scala 581:64]
+  wire  _T_4441 = ~buf_cmd_state_bus_en_0; // @[el2_lsu_bus_buffer.scala 581:91]
+  wire  _T_4442 = _T_4440 & _T_4441; // @[el2_lsu_bus_buffer.scala 581:89]
+  wire [2:0] _GEN_363 = {{2'd0}, _T_4442}; // @[el2_lsu_bus_buffer.scala 581:142]
+  wire [3:0] buf_numvld_wrcmd_any = _T_4459 + _GEN_363; // @[el2_lsu_bus_buffer.scala 581:142]
   wire  _T_1029 = buf_numvld_wrcmd_any == 4'h1; // @[el2_lsu_bus_buffer.scala 320:43]
-  wire  _T_4473 = _T_2636 & _T_4457; // @[el2_lsu_bus_buffer.scala 582:73]
-  wire  _T_4470 = _T_2631 & _T_4452; // @[el2_lsu_bus_buffer.scala 582:73]
-  wire [1:0] _T_4474 = _T_4473 + _T_4470; // @[el2_lsu_bus_buffer.scala 582:126]
-  wire  _T_4467 = _T_2626 & _T_4447; // @[el2_lsu_bus_buffer.scala 582:73]
-  wire [1:0] _GEN_364 = {{1'd0}, _T_4467}; // @[el2_lsu_bus_buffer.scala 582:126]
-  wire [2:0] _T_4475 = _T_4474 + _GEN_364; // @[el2_lsu_bus_buffer.scala 582:126]
-  wire  _T_4464 = _T_2621 & _T_4442; // @[el2_lsu_bus_buffer.scala 582:73]
-  wire [2:0] _GEN_365 = {{2'd0}, _T_4464}; // @[el2_lsu_bus_buffer.scala 582:126]
-  wire [3:0] buf_numvld_cmd_any = _T_4475 + _GEN_365; // @[el2_lsu_bus_buffer.scala 582:126]
+  wire  _T_4472 = _T_2636 & _T_4456; // @[el2_lsu_bus_buffer.scala 582:73]
+  wire  _T_4469 = _T_2631 & _T_4451; // @[el2_lsu_bus_buffer.scala 582:73]
+  wire [1:0] _T_4473 = _T_4472 + _T_4469; // @[el2_lsu_bus_buffer.scala 582:126]
+  wire  _T_4466 = _T_2626 & _T_4446; // @[el2_lsu_bus_buffer.scala 582:73]
+  wire [1:0] _GEN_364 = {{1'd0}, _T_4466}; // @[el2_lsu_bus_buffer.scala 582:126]
+  wire [2:0] _T_4474 = _T_4473 + _GEN_364; // @[el2_lsu_bus_buffer.scala 582:126]
+  wire  _T_4463 = _T_2621 & _T_4441; // @[el2_lsu_bus_buffer.scala 582:73]
+  wire [2:0] _GEN_365 = {{2'd0}, _T_4463}; // @[el2_lsu_bus_buffer.scala 582:126]
+  wire [3:0] buf_numvld_cmd_any = _T_4474 + _GEN_365; // @[el2_lsu_bus_buffer.scala 582:126]
   wire  _T_1030 = buf_numvld_cmd_any == 4'h1; // @[el2_lsu_bus_buffer.scala 320:72]
   wire  _T_1031 = _T_1029 & _T_1030; // @[el2_lsu_bus_buffer.scala 320:51]
   reg [2:0] obuf_wr_timer; // @[el2_lsu_bus_buffer.scala 419:54]
@@ -1042,19 +1042,19 @@ module el2_lsu_bus_buffer(
   wire  _T_1994 = |buf_age_3; // @[el2_lsu_bus_buffer.scala 436:58]
   wire  _T_1995 = ~_T_1994; // @[el2_lsu_bus_buffer.scala 436:45]
   wire  _T_1997 = _T_1995 & _T_2636; // @[el2_lsu_bus_buffer.scala 436:63]
-  wire  _T_1999 = _T_1997 & _T_4457; // @[el2_lsu_bus_buffer.scala 436:88]
+  wire  _T_1999 = _T_1997 & _T_4456; // @[el2_lsu_bus_buffer.scala 436:88]
   wire  _T_1988 = |buf_age_2; // @[el2_lsu_bus_buffer.scala 436:58]
   wire  _T_1989 = ~_T_1988; // @[el2_lsu_bus_buffer.scala 436:45]
   wire  _T_1991 = _T_1989 & _T_2631; // @[el2_lsu_bus_buffer.scala 436:63]
-  wire  _T_1993 = _T_1991 & _T_4452; // @[el2_lsu_bus_buffer.scala 436:88]
+  wire  _T_1993 = _T_1991 & _T_4451; // @[el2_lsu_bus_buffer.scala 436:88]
   wire  _T_1982 = |buf_age_1; // @[el2_lsu_bus_buffer.scala 436:58]
   wire  _T_1983 = ~_T_1982; // @[el2_lsu_bus_buffer.scala 436:45]
   wire  _T_1985 = _T_1983 & _T_2626; // @[el2_lsu_bus_buffer.scala 436:63]
-  wire  _T_1987 = _T_1985 & _T_4447; // @[el2_lsu_bus_buffer.scala 436:88]
+  wire  _T_1987 = _T_1985 & _T_4446; // @[el2_lsu_bus_buffer.scala 436:88]
   wire  _T_1976 = |buf_age_0; // @[el2_lsu_bus_buffer.scala 436:58]
   wire  _T_1977 = ~_T_1976; // @[el2_lsu_bus_buffer.scala 436:45]
   wire  _T_1979 = _T_1977 & _T_2621; // @[el2_lsu_bus_buffer.scala 436:63]
-  wire  _T_1981 = _T_1979 & _T_4442; // @[el2_lsu_bus_buffer.scala 436:88]
+  wire  _T_1981 = _T_1979 & _T_4441; // @[el2_lsu_bus_buffer.scala 436:88]
   wire [3:0] CmdPtr0Dec = {_T_1999,_T_1993,_T_1987,_T_1981}; // @[Cat.scala 29:58]
   wire [7:0] _T_2069 = {4'h0,_T_1999,_T_1993,_T_1987,_T_1981}; // @[Cat.scala 29:58]
   wire  _T_2072 = _T_2069[4] | _T_2069[5]; // @[el2_lsu_bus_buffer.scala 444:42]
@@ -1116,40 +1116,40 @@ module el2_lsu_bus_buffer(
   wire  _T_1070 = obuf_wr_timer < 3'h7; // @[el2_lsu_bus_buffer.scala 323:95]
   wire  _T_1071 = _T_1069 & _T_1070; // @[el2_lsu_bus_buffer.scala 323:79]
   wire [2:0] _T_1073 = obuf_wr_timer + 3'h1; // @[el2_lsu_bus_buffer.scala 323:121]
-  wire  _T_4492 = buf_state_3 == 3'h1; // @[el2_lsu_bus_buffer.scala 583:63]
-  wire  _T_4496 = _T_4492 | _T_4473; // @[el2_lsu_bus_buffer.scala 583:74]
-  wire  _T_4487 = buf_state_2 == 3'h1; // @[el2_lsu_bus_buffer.scala 583:63]
-  wire  _T_4491 = _T_4487 | _T_4470; // @[el2_lsu_bus_buffer.scala 583:74]
-  wire [1:0] _T_4497 = _T_4496 + _T_4491; // @[el2_lsu_bus_buffer.scala 583:154]
-  wire  _T_4482 = buf_state_1 == 3'h1; // @[el2_lsu_bus_buffer.scala 583:63]
-  wire  _T_4486 = _T_4482 | _T_4467; // @[el2_lsu_bus_buffer.scala 583:74]
-  wire [1:0] _GEN_366 = {{1'd0}, _T_4486}; // @[el2_lsu_bus_buffer.scala 583:154]
-  wire [2:0] _T_4498 = _T_4497 + _GEN_366; // @[el2_lsu_bus_buffer.scala 583:154]
-  wire  _T_4477 = buf_state_0 == 3'h1; // @[el2_lsu_bus_buffer.scala 583:63]
-  wire  _T_4481 = _T_4477 | _T_4464; // @[el2_lsu_bus_buffer.scala 583:74]
-  wire [2:0] _GEN_367 = {{2'd0}, _T_4481}; // @[el2_lsu_bus_buffer.scala 583:154]
-  wire [3:0] buf_numvld_pend_any = _T_4498 + _GEN_367; // @[el2_lsu_bus_buffer.scala 583:154]
+  wire  _T_4491 = buf_state_3 == 3'h1; // @[el2_lsu_bus_buffer.scala 583:63]
+  wire  _T_4495 = _T_4491 | _T_4472; // @[el2_lsu_bus_buffer.scala 583:74]
+  wire  _T_4486 = buf_state_2 == 3'h1; // @[el2_lsu_bus_buffer.scala 583:63]
+  wire  _T_4490 = _T_4486 | _T_4469; // @[el2_lsu_bus_buffer.scala 583:74]
+  wire [1:0] _T_4496 = _T_4495 + _T_4490; // @[el2_lsu_bus_buffer.scala 583:154]
+  wire  _T_4481 = buf_state_1 == 3'h1; // @[el2_lsu_bus_buffer.scala 583:63]
+  wire  _T_4485 = _T_4481 | _T_4466; // @[el2_lsu_bus_buffer.scala 583:74]
+  wire [1:0] _GEN_366 = {{1'd0}, _T_4485}; // @[el2_lsu_bus_buffer.scala 583:154]
+  wire [2:0] _T_4497 = _T_4496 + _GEN_366; // @[el2_lsu_bus_buffer.scala 583:154]
+  wire  _T_4476 = buf_state_0 == 3'h1; // @[el2_lsu_bus_buffer.scala 583:63]
+  wire  _T_4480 = _T_4476 | _T_4463; // @[el2_lsu_bus_buffer.scala 583:74]
+  wire [2:0] _GEN_367 = {{2'd0}, _T_4480}; // @[el2_lsu_bus_buffer.scala 583:154]
+  wire [3:0] buf_numvld_pend_any = _T_4497 + _GEN_367; // @[el2_lsu_bus_buffer.scala 583:154]
   wire  _T_1100 = buf_numvld_pend_any == 4'h0; // @[el2_lsu_bus_buffer.scala 326:53]
   wire  _T_1101 = ibuf_byp & _T_1100; // @[el2_lsu_bus_buffer.scala 326:31]
   wire  _T_1102 = ~io_lsu_pkt_r_store; // @[el2_lsu_bus_buffer.scala 326:64]
   wire  _T_1103 = _T_1102 | io_no_dword_merge_r; // @[el2_lsu_bus_buffer.scala 326:84]
   wire  ibuf_buf_byp = _T_1101 & _T_1103; // @[el2_lsu_bus_buffer.scala 326:61]
   wire  _T_1104 = ibuf_buf_byp & io_lsu_commit_r; // @[el2_lsu_bus_buffer.scala 341:32]
-  wire  _T_4788 = buf_state_0 == 3'h3; // @[el2_lsu_bus_buffer.scala 612:62]
-  wire  _T_4790 = _T_4788 & buf_sideeffect[0]; // @[el2_lsu_bus_buffer.scala 612:73]
-  wire  _T_4791 = _T_4790 & io_dec_tlu_sideeffect_posted_disable; // @[el2_lsu_bus_buffer.scala 612:93]
-  wire  _T_4792 = buf_state_1 == 3'h3; // @[el2_lsu_bus_buffer.scala 612:62]
-  wire  _T_4794 = _T_4792 & buf_sideeffect[1]; // @[el2_lsu_bus_buffer.scala 612:73]
-  wire  _T_4795 = _T_4794 & io_dec_tlu_sideeffect_posted_disable; // @[el2_lsu_bus_buffer.scala 612:93]
-  wire  _T_4804 = _T_4791 | _T_4795; // @[el2_lsu_bus_buffer.scala 612:141]
-  wire  _T_4796 = buf_state_2 == 3'h3; // @[el2_lsu_bus_buffer.scala 612:62]
-  wire  _T_4798 = _T_4796 & buf_sideeffect[2]; // @[el2_lsu_bus_buffer.scala 612:73]
-  wire  _T_4799 = _T_4798 & io_dec_tlu_sideeffect_posted_disable; // @[el2_lsu_bus_buffer.scala 612:93]
-  wire  _T_4805 = _T_4804 | _T_4799; // @[el2_lsu_bus_buffer.scala 612:141]
-  wire  _T_4800 = buf_state_3 == 3'h3; // @[el2_lsu_bus_buffer.scala 612:62]
-  wire  _T_4802 = _T_4800 & buf_sideeffect[3]; // @[el2_lsu_bus_buffer.scala 612:73]
-  wire  _T_4803 = _T_4802 & io_dec_tlu_sideeffect_posted_disable; // @[el2_lsu_bus_buffer.scala 612:93]
-  wire  bus_sideeffect_pend = _T_4805 | _T_4803; // @[el2_lsu_bus_buffer.scala 612:141]
+  wire  _T_4787 = buf_state_0 == 3'h3; // @[el2_lsu_bus_buffer.scala 612:62]
+  wire  _T_4789 = _T_4787 & buf_sideeffect[0]; // @[el2_lsu_bus_buffer.scala 612:73]
+  wire  _T_4790 = _T_4789 & io_dec_tlu_sideeffect_posted_disable; // @[el2_lsu_bus_buffer.scala 612:93]
+  wire  _T_4791 = buf_state_1 == 3'h3; // @[el2_lsu_bus_buffer.scala 612:62]
+  wire  _T_4793 = _T_4791 & buf_sideeffect[1]; // @[el2_lsu_bus_buffer.scala 612:73]
+  wire  _T_4794 = _T_4793 & io_dec_tlu_sideeffect_posted_disable; // @[el2_lsu_bus_buffer.scala 612:93]
+  wire  _T_4803 = _T_4790 | _T_4794; // @[el2_lsu_bus_buffer.scala 612:141]
+  wire  _T_4795 = buf_state_2 == 3'h3; // @[el2_lsu_bus_buffer.scala 612:62]
+  wire  _T_4797 = _T_4795 & buf_sideeffect[2]; // @[el2_lsu_bus_buffer.scala 612:73]
+  wire  _T_4798 = _T_4797 & io_dec_tlu_sideeffect_posted_disable; // @[el2_lsu_bus_buffer.scala 612:93]
+  wire  _T_4804 = _T_4803 | _T_4798; // @[el2_lsu_bus_buffer.scala 612:141]
+  wire  _T_4799 = buf_state_3 == 3'h3; // @[el2_lsu_bus_buffer.scala 612:62]
+  wire  _T_4801 = _T_4799 & buf_sideeffect[3]; // @[el2_lsu_bus_buffer.scala 612:73]
+  wire  _T_4802 = _T_4801 & io_dec_tlu_sideeffect_posted_disable; // @[el2_lsu_bus_buffer.scala 612:93]
+  wire  bus_sideeffect_pend = _T_4804 | _T_4802; // @[el2_lsu_bus_buffer.scala 612:141]
   wire  _T_1105 = io_is_sideeffects_r & bus_sideeffect_pend; // @[el2_lsu_bus_buffer.scala 341:74]
   wire  _T_1106 = ~_T_1105; // @[el2_lsu_bus_buffer.scala 341:52]
   wire  _T_1107 = _T_1104 & _T_1106; // @[el2_lsu_bus_buffer.scala 341:50]
@@ -1218,28 +1218,28 @@ module el2_lsu_bus_buffer(
   wire  _T_2040 = ~CmdPtr0Dec[3]; // @[el2_lsu_bus_buffer.scala 437:83]
   wire  _T_2041 = _T_2038 & _T_2040; // @[el2_lsu_bus_buffer.scala 437:81]
   wire  _T_2043 = _T_2041 & _T_2636; // @[el2_lsu_bus_buffer.scala 437:98]
-  wire  _T_2045 = _T_2043 & _T_4457; // @[el2_lsu_bus_buffer.scala 437:123]
+  wire  _T_2045 = _T_2043 & _T_4456; // @[el2_lsu_bus_buffer.scala 437:123]
   wire [3:0] _T_2025 = buf_age_2 & _T_2035; // @[el2_lsu_bus_buffer.scala 437:59]
   wire  _T_2026 = |_T_2025; // @[el2_lsu_bus_buffer.scala 437:76]
   wire  _T_2027 = ~_T_2026; // @[el2_lsu_bus_buffer.scala 437:45]
   wire  _T_2029 = ~CmdPtr0Dec[2]; // @[el2_lsu_bus_buffer.scala 437:83]
   wire  _T_2030 = _T_2027 & _T_2029; // @[el2_lsu_bus_buffer.scala 437:81]
   wire  _T_2032 = _T_2030 & _T_2631; // @[el2_lsu_bus_buffer.scala 437:98]
-  wire  _T_2034 = _T_2032 & _T_4452; // @[el2_lsu_bus_buffer.scala 437:123]
+  wire  _T_2034 = _T_2032 & _T_4451; // @[el2_lsu_bus_buffer.scala 437:123]
   wire [3:0] _T_2014 = buf_age_1 & _T_2035; // @[el2_lsu_bus_buffer.scala 437:59]
   wire  _T_2015 = |_T_2014; // @[el2_lsu_bus_buffer.scala 437:76]
   wire  _T_2016 = ~_T_2015; // @[el2_lsu_bus_buffer.scala 437:45]
   wire  _T_2018 = ~CmdPtr0Dec[1]; // @[el2_lsu_bus_buffer.scala 437:83]
   wire  _T_2019 = _T_2016 & _T_2018; // @[el2_lsu_bus_buffer.scala 437:81]
   wire  _T_2021 = _T_2019 & _T_2626; // @[el2_lsu_bus_buffer.scala 437:98]
-  wire  _T_2023 = _T_2021 & _T_4447; // @[el2_lsu_bus_buffer.scala 437:123]
+  wire  _T_2023 = _T_2021 & _T_4446; // @[el2_lsu_bus_buffer.scala 437:123]
   wire [3:0] _T_2003 = buf_age_0 & _T_2035; // @[el2_lsu_bus_buffer.scala 437:59]
   wire  _T_2004 = |_T_2003; // @[el2_lsu_bus_buffer.scala 437:76]
   wire  _T_2005 = ~_T_2004; // @[el2_lsu_bus_buffer.scala 437:45]
   wire  _T_2007 = ~CmdPtr0Dec[0]; // @[el2_lsu_bus_buffer.scala 437:83]
   wire  _T_2008 = _T_2005 & _T_2007; // @[el2_lsu_bus_buffer.scala 437:81]
   wire  _T_2010 = _T_2008 & _T_2621; // @[el2_lsu_bus_buffer.scala 437:98]
-  wire  _T_2012 = _T_2010 & _T_4442; // @[el2_lsu_bus_buffer.scala 437:123]
+  wire  _T_2012 = _T_2010 & _T_4441; // @[el2_lsu_bus_buffer.scala 437:123]
   wire [3:0] CmdPtr1Dec = {_T_2045,_T_2034,_T_2023,_T_2012}; // @[Cat.scala 29:58]
   wire  found_cmdptr1 = |CmdPtr1Dec; // @[el2_lsu_bus_buffer.scala 442:31]
   wire  _T_1220 = _T_1219 | found_cmdptr1; // @[el2_lsu_bus_buffer.scala 344:181]
@@ -1258,10 +1258,10 @@ module el2_lsu_bus_buffer(
   reg  obuf_write; // @[Reg.scala 27:20]
   reg  obuf_cmd_done; // @[el2_lsu_bus_buffer.scala 406:54]
   reg  obuf_data_done; // @[el2_lsu_bus_buffer.scala 407:55]
-  wire  _T_4863 = obuf_cmd_done | obuf_data_done; // @[el2_lsu_bus_buffer.scala 616:54]
-  wire  _T_4864 = obuf_cmd_done ? io_lsu_axi_wready : io_lsu_axi_awready; // @[el2_lsu_bus_buffer.scala 616:75]
-  wire  _T_4866 = _T_4863 ? _T_4864 : io_lsu_axi_awready; // @[el2_lsu_bus_buffer.scala 616:39]
-  wire  bus_cmd_ready = obuf_write ? _T_4866 : io_lsu_axi_arready; // @[el2_lsu_bus_buffer.scala 616:23]
+  wire  _T_4862 = obuf_cmd_done | obuf_data_done; // @[el2_lsu_bus_buffer.scala 616:54]
+  wire  _T_4863 = obuf_cmd_done ? io_lsu_axi_wready : io_lsu_axi_awready; // @[el2_lsu_bus_buffer.scala 616:75]
+  wire  _T_4865 = _T_4862 ? _T_4863 : io_lsu_axi_awready; // @[el2_lsu_bus_buffer.scala 616:39]
+  wire  bus_cmd_ready = obuf_write ? _T_4865 : io_lsu_axi_arready; // @[el2_lsu_bus_buffer.scala 616:23]
   wire  _T_1244 = ~obuf_valid; // @[el2_lsu_bus_buffer.scala 345:48]
   wire  _T_1245 = bus_cmd_ready | _T_1244; // @[el2_lsu_bus_buffer.scala 345:46]
   reg  obuf_nosend; // @[Reg.scala 27:20]
@@ -1270,52 +1270,52 @@ module el2_lsu_bus_buffer(
   wire  _T_1248 = ~obuf_wr_wait; // @[el2_lsu_bus_buffer.scala 345:77]
   wire  _T_1249 = _T_1247 & _T_1248; // @[el2_lsu_bus_buffer.scala 345:75]
   reg [31:0] obuf_addr; // @[el2_lib.scala 491:16]
-  wire  _T_4811 = obuf_addr[31:3] == buf_addr_0[31:3]; // @[el2_lsu_bus_buffer.scala 614:56]
-  wire  _T_4812 = obuf_valid & _T_4811; // @[el2_lsu_bus_buffer.scala 614:38]
-  wire  _T_4814 = obuf_tag1 == 2'h0; // @[el2_lsu_bus_buffer.scala 614:126]
-  wire  _T_4815 = obuf_merge & _T_4814; // @[el2_lsu_bus_buffer.scala 614:114]
-  wire  _T_4816 = _T_3577 | _T_4815; // @[el2_lsu_bus_buffer.scala 614:100]
-  wire  _T_4817 = ~_T_4816; // @[el2_lsu_bus_buffer.scala 614:80]
-  wire  _T_4818 = _T_4812 & _T_4817; // @[el2_lsu_bus_buffer.scala 614:78]
-  wire  _T_4855 = _T_4788 & _T_4818; // @[Mux.scala 27:72]
-  wire  _T_4823 = obuf_addr[31:3] == buf_addr_1[31:3]; // @[el2_lsu_bus_buffer.scala 614:56]
-  wire  _T_4824 = obuf_valid & _T_4823; // @[el2_lsu_bus_buffer.scala 614:38]
-  wire  _T_4826 = obuf_tag1 == 2'h1; // @[el2_lsu_bus_buffer.scala 614:126]
-  wire  _T_4827 = obuf_merge & _T_4826; // @[el2_lsu_bus_buffer.scala 614:114]
-  wire  _T_4828 = _T_3770 | _T_4827; // @[el2_lsu_bus_buffer.scala 614:100]
-  wire  _T_4829 = ~_T_4828; // @[el2_lsu_bus_buffer.scala 614:80]
-  wire  _T_4830 = _T_4824 & _T_4829; // @[el2_lsu_bus_buffer.scala 614:78]
-  wire  _T_4856 = _T_4792 & _T_4830; // @[Mux.scala 27:72]
-  wire  _T_4859 = _T_4855 | _T_4856; // @[Mux.scala 27:72]
-  wire  _T_4835 = obuf_addr[31:3] == buf_addr_2[31:3]; // @[el2_lsu_bus_buffer.scala 614:56]
-  wire  _T_4836 = obuf_valid & _T_4835; // @[el2_lsu_bus_buffer.scala 614:38]
-  wire  _T_4838 = obuf_tag1 == 2'h2; // @[el2_lsu_bus_buffer.scala 614:126]
-  wire  _T_4839 = obuf_merge & _T_4838; // @[el2_lsu_bus_buffer.scala 614:114]
-  wire  _T_4840 = _T_3963 | _T_4839; // @[el2_lsu_bus_buffer.scala 614:100]
-  wire  _T_4841 = ~_T_4840; // @[el2_lsu_bus_buffer.scala 614:80]
-  wire  _T_4842 = _T_4836 & _T_4841; // @[el2_lsu_bus_buffer.scala 614:78]
-  wire  _T_4857 = _T_4796 & _T_4842; // @[Mux.scala 27:72]
-  wire  _T_4860 = _T_4859 | _T_4857; // @[Mux.scala 27:72]
-  wire  _T_4847 = obuf_addr[31:3] == buf_addr_3[31:3]; // @[el2_lsu_bus_buffer.scala 614:56]
-  wire  _T_4848 = obuf_valid & _T_4847; // @[el2_lsu_bus_buffer.scala 614:38]
-  wire  _T_4850 = obuf_tag1 == 2'h3; // @[el2_lsu_bus_buffer.scala 614:126]
-  wire  _T_4851 = obuf_merge & _T_4850; // @[el2_lsu_bus_buffer.scala 614:114]
-  wire  _T_4852 = _T_4156 | _T_4851; // @[el2_lsu_bus_buffer.scala 614:100]
-  wire  _T_4853 = ~_T_4852; // @[el2_lsu_bus_buffer.scala 614:80]
-  wire  _T_4854 = _T_4848 & _T_4853; // @[el2_lsu_bus_buffer.scala 614:78]
-  wire  _T_4858 = _T_4800 & _T_4854; // @[Mux.scala 27:72]
-  wire  bus_addr_match_pending = _T_4860 | _T_4858; // @[Mux.scala 27:72]
+  wire  _T_4810 = obuf_addr[31:3] == buf_addr_0[31:3]; // @[el2_lsu_bus_buffer.scala 614:56]
+  wire  _T_4811 = obuf_valid & _T_4810; // @[el2_lsu_bus_buffer.scala 614:38]
+  wire  _T_4813 = obuf_tag1 == 2'h0; // @[el2_lsu_bus_buffer.scala 614:126]
+  wire  _T_4814 = obuf_merge & _T_4813; // @[el2_lsu_bus_buffer.scala 614:114]
+  wire  _T_4815 = _T_3577 | _T_4814; // @[el2_lsu_bus_buffer.scala 614:100]
+  wire  _T_4816 = ~_T_4815; // @[el2_lsu_bus_buffer.scala 614:80]
+  wire  _T_4817 = _T_4811 & _T_4816; // @[el2_lsu_bus_buffer.scala 614:78]
+  wire  _T_4854 = _T_4787 & _T_4817; // @[Mux.scala 27:72]
+  wire  _T_4822 = obuf_addr[31:3] == buf_addr_1[31:3]; // @[el2_lsu_bus_buffer.scala 614:56]
+  wire  _T_4823 = obuf_valid & _T_4822; // @[el2_lsu_bus_buffer.scala 614:38]
+  wire  _T_4825 = obuf_tag1 == 2'h1; // @[el2_lsu_bus_buffer.scala 614:126]
+  wire  _T_4826 = obuf_merge & _T_4825; // @[el2_lsu_bus_buffer.scala 614:114]
+  wire  _T_4827 = _T_3770 | _T_4826; // @[el2_lsu_bus_buffer.scala 614:100]
+  wire  _T_4828 = ~_T_4827; // @[el2_lsu_bus_buffer.scala 614:80]
+  wire  _T_4829 = _T_4823 & _T_4828; // @[el2_lsu_bus_buffer.scala 614:78]
+  wire  _T_4855 = _T_4791 & _T_4829; // @[Mux.scala 27:72]
+  wire  _T_4858 = _T_4854 | _T_4855; // @[Mux.scala 27:72]
+  wire  _T_4834 = obuf_addr[31:3] == buf_addr_2[31:3]; // @[el2_lsu_bus_buffer.scala 614:56]
+  wire  _T_4835 = obuf_valid & _T_4834; // @[el2_lsu_bus_buffer.scala 614:38]
+  wire  _T_4837 = obuf_tag1 == 2'h2; // @[el2_lsu_bus_buffer.scala 614:126]
+  wire  _T_4838 = obuf_merge & _T_4837; // @[el2_lsu_bus_buffer.scala 614:114]
+  wire  _T_4839 = _T_3963 | _T_4838; // @[el2_lsu_bus_buffer.scala 614:100]
+  wire  _T_4840 = ~_T_4839; // @[el2_lsu_bus_buffer.scala 614:80]
+  wire  _T_4841 = _T_4835 & _T_4840; // @[el2_lsu_bus_buffer.scala 614:78]
+  wire  _T_4856 = _T_4795 & _T_4841; // @[Mux.scala 27:72]
+  wire  _T_4859 = _T_4858 | _T_4856; // @[Mux.scala 27:72]
+  wire  _T_4846 = obuf_addr[31:3] == buf_addr_3[31:3]; // @[el2_lsu_bus_buffer.scala 614:56]
+  wire  _T_4847 = obuf_valid & _T_4846; // @[el2_lsu_bus_buffer.scala 614:38]
+  wire  _T_4849 = obuf_tag1 == 2'h3; // @[el2_lsu_bus_buffer.scala 614:126]
+  wire  _T_4850 = obuf_merge & _T_4849; // @[el2_lsu_bus_buffer.scala 614:114]
+  wire  _T_4851 = _T_4156 | _T_4850; // @[el2_lsu_bus_buffer.scala 614:100]
+  wire  _T_4852 = ~_T_4851; // @[el2_lsu_bus_buffer.scala 614:80]
+  wire  _T_4853 = _T_4847 & _T_4852; // @[el2_lsu_bus_buffer.scala 614:78]
+  wire  _T_4857 = _T_4799 & _T_4853; // @[Mux.scala 27:72]
+  wire  bus_addr_match_pending = _T_4859 | _T_4857; // @[Mux.scala 27:72]
   wire  _T_1252 = ~bus_addr_match_pending; // @[el2_lsu_bus_buffer.scala 345:118]
   wire  _T_1253 = _T_1249 & _T_1252; // @[el2_lsu_bus_buffer.scala 345:116]
   wire  obuf_wr_en = _T_1253 & io_lsu_bus_clk_en; // @[el2_lsu_bus_buffer.scala 345:142]
   wire  _T_1255 = obuf_valid & obuf_nosend; // @[el2_lsu_bus_buffer.scala 347:47]
   wire  bus_wcmd_sent = io_lsu_axi_awvalid & io_lsu_axi_awready; // @[el2_lsu_bus_buffer.scala 617:39]
-  wire  _T_4870 = obuf_cmd_done | bus_wcmd_sent; // @[el2_lsu_bus_buffer.scala 619:35]
+  wire  _T_4869 = obuf_cmd_done | bus_wcmd_sent; // @[el2_lsu_bus_buffer.scala 619:35]
   wire  bus_wdata_sent = io_lsu_axi_wvalid & io_lsu_axi_wready; // @[el2_lsu_bus_buffer.scala 618:39]
-  wire  _T_4871 = obuf_data_done | bus_wdata_sent; // @[el2_lsu_bus_buffer.scala 619:70]
-  wire  _T_4872 = _T_4870 & _T_4871; // @[el2_lsu_bus_buffer.scala 619:52]
-  wire  _T_4873 = io_lsu_axi_arvalid & io_lsu_axi_arready; // @[el2_lsu_bus_buffer.scala 619:111]
-  wire  bus_cmd_sent = _T_4872 | _T_4873; // @[el2_lsu_bus_buffer.scala 619:89]
+  wire  _T_4870 = obuf_data_done | bus_wdata_sent; // @[el2_lsu_bus_buffer.scala 619:70]
+  wire  _T_4871 = _T_4869 & _T_4870; // @[el2_lsu_bus_buffer.scala 619:52]
+  wire  _T_4872 = io_lsu_axi_arvalid & io_lsu_axi_arready; // @[el2_lsu_bus_buffer.scala 619:111]
+  wire  bus_cmd_sent = _T_4871 | _T_4872; // @[el2_lsu_bus_buffer.scala 619:89]
   wire  _T_1256 = bus_cmd_sent | _T_1255; // @[el2_lsu_bus_buffer.scala 347:33]
   wire  _T_1257 = ~obuf_wr_en; // @[el2_lsu_bus_buffer.scala 347:65]
   wire  _T_1258 = _T_1256 & _T_1257; // @[el2_lsu_bus_buffer.scala 347:63]
@@ -1694,7 +1694,7 @@ module el2_lsu_bus_buffer(
   wire  _T_2152 = _T_2150 & _T_3547; // @[el2_lsu_bus_buffer.scala 467:41]
   wire  _T_2154 = _T_2152 & _T_1869; // @[el2_lsu_bus_buffer.scala 467:71]
   wire  _T_2156 = _T_2154 & _T_1867; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2157 = _T_4481 | _T_2156; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2157 = _T_4480 | _T_2156; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2158 = ibuf_byp & io_lsu_busreq_r; // @[el2_lsu_bus_buffer.scala 468:17]
   wire  _T_2159 = _T_2158 & io_ldst_dual_r; // @[el2_lsu_bus_buffer.scala 468:35]
   wire  _T_2161 = _T_2159 & _T_1870; // @[el2_lsu_bus_buffer.scala 468:52]
@@ -1703,19 +1703,19 @@ module el2_lsu_bus_buffer(
   wire  _T_2165 = _T_2144 & _T_2164; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2167 = _T_2165 | buf_age_0[0]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2181 = _T_2154 & _T_1878; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2182 = _T_4486 | _T_2181; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2182 = _T_4485 | _T_2181; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2188 = _T_2161 & _T_1880; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2189 = _T_2182 | _T_2188; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2190 = _T_2144 & _T_2189; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2192 = _T_2190 | buf_age_0[1]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2206 = _T_2154 & _T_1889; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2207 = _T_4491 | _T_2206; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2207 = _T_4490 | _T_2206; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2213 = _T_2161 & _T_1891; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2214 = _T_2207 | _T_2213; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2215 = _T_2144 & _T_2214; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2217 = _T_2215 | buf_age_0[2]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2231 = _T_2154 & _T_1900; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2232 = _T_4496 | _T_2231; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2232 = _T_4495 | _T_2231; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2238 = _T_2161 & _T_1902; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2239 = _T_2232 | _T_2238; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2240 = _T_2144 & _T_2239; // @[el2_lsu_bus_buffer.scala 465:113]
@@ -1786,26 +1786,26 @@ module el2_lsu_bus_buffer(
   wire  _T_2246 = _T_1877 & buf_state_en_1; // @[el2_lsu_bus_buffer.scala 465:94]
   wire  _T_2256 = _T_2152 & _T_1880; // @[el2_lsu_bus_buffer.scala 467:71]
   wire  _T_2258 = _T_2256 & _T_1867; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2259 = _T_4481 | _T_2258; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2259 = _T_4480 | _T_2258; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2263 = _T_2159 & _T_1881; // @[el2_lsu_bus_buffer.scala 468:52]
   wire  _T_2265 = _T_2263 & _T_1869; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2266 = _T_2259 | _T_2265; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2267 = _T_2246 & _T_2266; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2269 = _T_2267 | buf_age_1[0]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2283 = _T_2256 & _T_1878; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2284 = _T_4486 | _T_2283; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2284 = _T_4485 | _T_2283; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2290 = _T_2263 & _T_1880; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2291 = _T_2284 | _T_2290; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2292 = _T_2246 & _T_2291; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2294 = _T_2292 | buf_age_1[1]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2308 = _T_2256 & _T_1889; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2309 = _T_4491 | _T_2308; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2309 = _T_4490 | _T_2308; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2315 = _T_2263 & _T_1891; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2316 = _T_2309 | _T_2315; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2317 = _T_2246 & _T_2316; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2319 = _T_2317 | buf_age_1[2]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2333 = _T_2256 & _T_1900; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2334 = _T_4496 | _T_2333; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2334 = _T_4495 | _T_2333; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2340 = _T_2263 & _T_1902; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2341 = _T_2334 | _T_2340; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2342 = _T_2246 & _T_2341; // @[el2_lsu_bus_buffer.scala 465:113]
@@ -1876,26 +1876,26 @@ module el2_lsu_bus_buffer(
   wire  _T_2348 = _T_1888 & buf_state_en_2; // @[el2_lsu_bus_buffer.scala 465:94]
   wire  _T_2358 = _T_2152 & _T_1891; // @[el2_lsu_bus_buffer.scala 467:71]
   wire  _T_2360 = _T_2358 & _T_1867; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2361 = _T_4481 | _T_2360; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2361 = _T_4480 | _T_2360; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2365 = _T_2159 & _T_1892; // @[el2_lsu_bus_buffer.scala 468:52]
   wire  _T_2367 = _T_2365 & _T_1869; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2368 = _T_2361 | _T_2367; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2369 = _T_2348 & _T_2368; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2371 = _T_2369 | buf_age_2[0]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2385 = _T_2358 & _T_1878; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2386 = _T_4486 | _T_2385; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2386 = _T_4485 | _T_2385; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2392 = _T_2365 & _T_1880; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2393 = _T_2386 | _T_2392; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2394 = _T_2348 & _T_2393; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2396 = _T_2394 | buf_age_2[1]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2410 = _T_2358 & _T_1889; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2411 = _T_4491 | _T_2410; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2411 = _T_4490 | _T_2410; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2417 = _T_2365 & _T_1891; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2418 = _T_2411 | _T_2417; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2419 = _T_2348 & _T_2418; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2421 = _T_2419 | buf_age_2[2]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2435 = _T_2358 & _T_1900; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2436 = _T_4496 | _T_2435; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2436 = _T_4495 | _T_2435; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2442 = _T_2365 & _T_1902; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2443 = _T_2436 | _T_2442; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2444 = _T_2348 & _T_2443; // @[el2_lsu_bus_buffer.scala 465:113]
@@ -1966,26 +1966,26 @@ module el2_lsu_bus_buffer(
   wire  _T_2450 = _T_1899 & buf_state_en_3; // @[el2_lsu_bus_buffer.scala 465:94]
   wire  _T_2460 = _T_2152 & _T_1902; // @[el2_lsu_bus_buffer.scala 467:71]
   wire  _T_2462 = _T_2460 & _T_1867; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2463 = _T_4481 | _T_2462; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2463 = _T_4480 | _T_2462; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2467 = _T_2159 & _T_1903; // @[el2_lsu_bus_buffer.scala 468:52]
   wire  _T_2469 = _T_2467 & _T_1869; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2470 = _T_2463 | _T_2469; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2471 = _T_2450 & _T_2470; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2473 = _T_2471 | buf_age_3[0]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2487 = _T_2460 & _T_1878; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2488 = _T_4486 | _T_2487; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2488 = _T_4485 | _T_2487; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2494 = _T_2467 & _T_1880; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2495 = _T_2488 | _T_2494; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2496 = _T_2450 & _T_2495; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2498 = _T_2496 | buf_age_3[1]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2512 = _T_2460 & _T_1889; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2513 = _T_4491 | _T_2512; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2513 = _T_4490 | _T_2512; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2519 = _T_2467 & _T_1891; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2520 = _T_2513 | _T_2519; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2521 = _T_2450 & _T_2520; // @[el2_lsu_bus_buffer.scala 465:113]
   wire  _T_2523 = _T_2521 | buf_age_3[2]; // @[el2_lsu_bus_buffer.scala 468:97]
   wire  _T_2537 = _T_2460 & _T_1900; // @[el2_lsu_bus_buffer.scala 467:92]
-  wire  _T_2538 = _T_4496 | _T_2537; // @[el2_lsu_bus_buffer.scala 466:86]
+  wire  _T_2538 = _T_4495 | _T_2537; // @[el2_lsu_bus_buffer.scala 466:86]
   wire  _T_2544 = _T_2467 & _T_1902; // @[el2_lsu_bus_buffer.scala 468:73]
   wire  _T_2545 = _T_2538 | _T_2544; // @[el2_lsu_bus_buffer.scala 467:114]
   wire  _T_2546 = _T_2450 & _T_2545; // @[el2_lsu_bus_buffer.scala 465:113]
@@ -2155,14 +2155,14 @@ module el2_lsu_bus_buffer(
   wire  _T_3590 = _T_3588 & _T_1346; // @[el2_lsu_bus_buffer.scala 518:74]
   wire  _T_3593 = _T_3583 & obuf_nosend; // @[el2_lsu_bus_buffer.scala 520:67]
   wire  _T_3594 = _T_3593 & bus_rsp_read; // @[el2_lsu_bus_buffer.scala 520:81]
-  wire  _T_4879 = io_lsu_axi_bresp != 2'h0; // @[el2_lsu_bus_buffer.scala 625:58]
-  wire  bus_rsp_read_error = bus_rsp_read & _T_4879; // @[el2_lsu_bus_buffer.scala 625:38]
+  wire  _T_4878 = io_lsu_axi_bresp != 2'h0; // @[el2_lsu_bus_buffer.scala 625:58]
+  wire  bus_rsp_read_error = bus_rsp_read & _T_4878; // @[el2_lsu_bus_buffer.scala 625:38]
   wire  _T_3597 = _T_3593 & bus_rsp_read_error; // @[el2_lsu_bus_buffer.scala 521:82]
   wire  _T_3672 = bus_rsp_read_error & _T_3651; // @[el2_lsu_bus_buffer.scala 535:91]
   wire  _T_3674 = bus_rsp_read_error & buf_ldfwd[0]; // @[el2_lsu_bus_buffer.scala 536:31]
   wire  _T_3676 = _T_3674 & _T_3653; // @[el2_lsu_bus_buffer.scala 536:46]
   wire  _T_3677 = _T_3672 | _T_3676; // @[el2_lsu_bus_buffer.scala 535:143]
-  wire  bus_rsp_write_error = bus_rsp_write & _T_4879; // @[el2_lsu_bus_buffer.scala 624:40]
+  wire  bus_rsp_write_error = bus_rsp_write & _T_4878; // @[el2_lsu_bus_buffer.scala 624:40]
   wire  _T_3680 = bus_rsp_write_error & _T_3649; // @[el2_lsu_bus_buffer.scala 537:53]
   wire  _T_3681 = _T_3677 | _T_3680; // @[el2_lsu_bus_buffer.scala 536:88]
   wire  _T_3682 = _T_3583 & _T_3681; // @[el2_lsu_bus_buffer.scala 535:68]
@@ -2181,9 +2181,9 @@ module el2_lsu_bus_buffer(
   wire [2:0] _GEN_21 = 2'h3 == buf_dualtag_0 ? buf_state_3 : _GEN_20; // @[el2_lsu_bus_buffer.scala 526:90]
   wire  _T_3616 = _GEN_21 != 3'h4; // @[el2_lsu_bus_buffer.scala 526:90]
   wire  _T_3617 = _T_3615 & _T_3616; // @[el2_lsu_bus_buffer.scala 526:61]
-  wire  _T_4504 = _T_2761 | _T_2758; // @[el2_lsu_bus_buffer.scala 584:93]
-  wire  _T_4505 = _T_4504 | _T_2755; // @[el2_lsu_bus_buffer.scala 584:93]
-  wire  any_done_wait_state = _T_4505 | _T_2752; // @[el2_lsu_bus_buffer.scala 584:93]
+  wire  _T_4503 = _T_2761 | _T_2758; // @[el2_lsu_bus_buffer.scala 584:93]
+  wire  _T_4504 = _T_4503 | _T_2755; // @[el2_lsu_bus_buffer.scala 584:93]
+  wire  any_done_wait_state = _T_4504 | _T_2752; // @[el2_lsu_bus_buffer.scala 584:93]
   wire  _T_3619 = buf_ldfwd[0] | any_done_wait_state; // @[el2_lsu_bus_buffer.scala 527:31]
   wire  _T_3625 = buf_dualtag_0 == 2'h0; // @[el2_lsu_bus_buffer.scala 119:118]
   wire  _T_3627 = buf_dualtag_0 == 2'h1; // @[el2_lsu_bus_buffer.scala 119:118]
@@ -2448,187 +2448,189 @@ module el2_lsu_bus_buffer(
   wire  _T_4423 = buf_error_en_3 | buf_error[3]; // @[el2_lsu_bus_buffer.scala 577:86]
   wire  _T_4424 = ~buf_rst_3; // @[el2_lsu_bus_buffer.scala 577:128]
   wire [2:0] _T_4431 = {buf_data_en_3,buf_data_en_2,buf_data_en_1}; // @[Cat.scala 29:58]
-  wire [1:0] _T_4437 = _T_26 + _T_19; // @[el2_lsu_bus_buffer.scala 580:96]
-  wire [1:0] _GEN_388 = {{1'd0}, _T_12}; // @[el2_lsu_bus_buffer.scala 580:96]
-  wire [2:0] _T_4438 = _T_4437 + _GEN_388; // @[el2_lsu_bus_buffer.scala 580:96]
-  wire [2:0] _GEN_389 = {{2'd0}, _T_5}; // @[el2_lsu_bus_buffer.scala 580:96]
-  wire [3:0] buf_numvld_any = _T_4438 + _GEN_389; // @[el2_lsu_bus_buffer.scala 580:96]
-  wire  _T_4508 = io_ldst_dual_d & io_dec_lsu_valid_raw_d; // @[el2_lsu_bus_buffer.scala 586:52]
-  wire  _T_4509 = buf_numvld_any >= 4'h3; // @[el2_lsu_bus_buffer.scala 586:92]
-  wire  _T_4510 = buf_numvld_any == 4'h3; // @[el2_lsu_bus_buffer.scala 586:119]
-  wire  _T_4512 = |buf_state_0; // @[el2_lsu_bus_buffer.scala 587:52]
-  wire  _T_4513 = |buf_state_1; // @[el2_lsu_bus_buffer.scala 587:52]
-  wire  _T_4514 = |buf_state_2; // @[el2_lsu_bus_buffer.scala 587:52]
-  wire  _T_4515 = |buf_state_3; // @[el2_lsu_bus_buffer.scala 587:52]
-  wire  _T_4516 = _T_4512 | _T_4513; // @[el2_lsu_bus_buffer.scala 587:65]
+  wire [1:0] _T_4433 = {io_lsu_busreq_m,1'h0}; // @[Cat.scala 29:58]
+  wire [1:0] _T_4434 = io_lsu_busreq_m ? _T_4433 : {{1'd0}, io_lsu_busreq_m}; // @[el2_lsu_bus_buffer.scala 580:27]
+  wire [1:0] _T_4435 = {io_lsu_busreq_r,1'h0}; // @[Cat.scala 29:58]
+  wire [1:0] _T_4436 = io_lsu_busreq_r ? _T_4435 : {{1'd0}, io_lsu_busreq_r}; // @[el2_lsu_bus_buffer.scala 580:94]
+  wire [2:0] _T_4437 = _T_4434 + _T_4436; // @[el2_lsu_bus_buffer.scala 580:88]
+  wire [2:0] _GEN_388 = {{2'd0}, ibuf_valid}; // @[el2_lsu_bus_buffer.scala 580:155]
+  wire [3:0] buf_numvld_any = _T_4437 + _GEN_388; // @[el2_lsu_bus_buffer.scala 580:155]
+  wire  _T_4507 = io_ldst_dual_d & io_dec_lsu_valid_raw_d; // @[el2_lsu_bus_buffer.scala 586:52]
+  wire  _T_4508 = buf_numvld_any >= 4'h3; // @[el2_lsu_bus_buffer.scala 586:92]
+  wire  _T_4509 = buf_numvld_any == 4'h3; // @[el2_lsu_bus_buffer.scala 586:119]
+  wire  _T_4511 = |buf_state_0; // @[el2_lsu_bus_buffer.scala 587:52]
+  wire  _T_4512 = |buf_state_1; // @[el2_lsu_bus_buffer.scala 587:52]
+  wire  _T_4513 = |buf_state_2; // @[el2_lsu_bus_buffer.scala 587:52]
+  wire  _T_4514 = |buf_state_3; // @[el2_lsu_bus_buffer.scala 587:52]
+  wire  _T_4515 = _T_4511 | _T_4512; // @[el2_lsu_bus_buffer.scala 587:65]
+  wire  _T_4516 = _T_4515 | _T_4513; // @[el2_lsu_bus_buffer.scala 587:65]
   wire  _T_4517 = _T_4516 | _T_4514; // @[el2_lsu_bus_buffer.scala 587:65]
-  wire  _T_4518 = _T_4517 | _T_4515; // @[el2_lsu_bus_buffer.scala 587:65]
-  wire  _T_4519 = ~_T_4518; // @[el2_lsu_bus_buffer.scala 587:34]
-  wire  _T_4521 = _T_4519 & _T_844; // @[el2_lsu_bus_buffer.scala 587:70]
-  wire  _T_4524 = io_lsu_busreq_m & io_lsu_pkt_m_valid; // @[el2_lsu_bus_buffer.scala 589:51]
-  wire  _T_4525 = _T_4524 & io_lsu_pkt_m_load; // @[el2_lsu_bus_buffer.scala 589:72]
-  wire  _T_4526 = ~io_flush_m_up; // @[el2_lsu_bus_buffer.scala 589:94]
-  wire  _T_4527 = _T_4525 & _T_4526; // @[el2_lsu_bus_buffer.scala 589:92]
-  wire  _T_4528 = ~io_ld_full_hit_m; // @[el2_lsu_bus_buffer.scala 589:111]
-  wire  _T_4530 = ~io_lsu_commit_r; // @[el2_lsu_bus_buffer.scala 592:61]
+  wire  _T_4518 = ~_T_4517; // @[el2_lsu_bus_buffer.scala 587:34]
+  wire  _T_4520 = _T_4518 & _T_844; // @[el2_lsu_bus_buffer.scala 587:70]
+  wire  _T_4523 = io_lsu_busreq_m & io_lsu_pkt_m_valid; // @[el2_lsu_bus_buffer.scala 589:51]
+  wire  _T_4524 = _T_4523 & io_lsu_pkt_m_load; // @[el2_lsu_bus_buffer.scala 589:72]
+  wire  _T_4525 = ~io_flush_m_up; // @[el2_lsu_bus_buffer.scala 589:94]
+  wire  _T_4526 = _T_4524 & _T_4525; // @[el2_lsu_bus_buffer.scala 589:92]
+  wire  _T_4527 = ~io_ld_full_hit_m; // @[el2_lsu_bus_buffer.scala 589:111]
+  wire  _T_4529 = ~io_lsu_commit_r; // @[el2_lsu_bus_buffer.scala 592:61]
   reg  lsu_nonblock_load_valid_r; // @[el2_lsu_bus_buffer.scala 678:66]
-  wire  _T_4548 = _T_2814 & _T_3658; // @[Mux.scala 27:72]
-  wire  _T_4549 = _T_2836 & _T_3851; // @[Mux.scala 27:72]
-  wire  _T_4550 = _T_2858 & _T_4044; // @[Mux.scala 27:72]
-  wire  _T_4551 = _T_2880 & _T_4237; // @[Mux.scala 27:72]
-  wire  _T_4552 = _T_4548 | _T_4549; // @[Mux.scala 27:72]
-  wire  _T_4553 = _T_4552 | _T_4550; // @[Mux.scala 27:72]
-  wire  lsu_nonblock_load_data_ready = _T_4553 | _T_4551; // @[Mux.scala 27:72]
-  wire  _T_4559 = buf_error[0] & _T_3658; // @[el2_lsu_bus_buffer.scala 595:108]
-  wire  _T_4564 = buf_error[1] & _T_3851; // @[el2_lsu_bus_buffer.scala 595:108]
-  wire  _T_4569 = buf_error[2] & _T_4044; // @[el2_lsu_bus_buffer.scala 595:108]
-  wire  _T_4574 = buf_error[3] & _T_4237; // @[el2_lsu_bus_buffer.scala 595:108]
-  wire  _T_4575 = _T_2814 & _T_4559; // @[Mux.scala 27:72]
-  wire  _T_4576 = _T_2836 & _T_4564; // @[Mux.scala 27:72]
-  wire  _T_4577 = _T_2858 & _T_4569; // @[Mux.scala 27:72]
-  wire  _T_4578 = _T_2880 & _T_4574; // @[Mux.scala 27:72]
-  wire  _T_4579 = _T_4575 | _T_4576; // @[Mux.scala 27:72]
-  wire  _T_4580 = _T_4579 | _T_4577; // @[Mux.scala 27:72]
-  wire  _T_4587 = ~buf_dual_0; // @[el2_lsu_bus_buffer.scala 596:109]
-  wire  _T_4588 = ~buf_dualhi_0; // @[el2_lsu_bus_buffer.scala 596:124]
-  wire  _T_4589 = _T_4587 | _T_4588; // @[el2_lsu_bus_buffer.scala 596:122]
-  wire  _T_4590 = _T_4548 & _T_4589; // @[el2_lsu_bus_buffer.scala 596:106]
-  wire  _T_4595 = ~buf_dual_1; // @[el2_lsu_bus_buffer.scala 596:109]
-  wire  _T_4596 = ~buf_dualhi_1; // @[el2_lsu_bus_buffer.scala 596:124]
-  wire  _T_4597 = _T_4595 | _T_4596; // @[el2_lsu_bus_buffer.scala 596:122]
-  wire  _T_4598 = _T_4549 & _T_4597; // @[el2_lsu_bus_buffer.scala 596:106]
-  wire  _T_4603 = ~buf_dual_2; // @[el2_lsu_bus_buffer.scala 596:109]
-  wire  _T_4604 = ~buf_dualhi_2; // @[el2_lsu_bus_buffer.scala 596:124]
-  wire  _T_4605 = _T_4603 | _T_4604; // @[el2_lsu_bus_buffer.scala 596:122]
-  wire  _T_4606 = _T_4550 & _T_4605; // @[el2_lsu_bus_buffer.scala 596:106]
-  wire  _T_4611 = ~buf_dual_3; // @[el2_lsu_bus_buffer.scala 596:109]
-  wire  _T_4612 = ~buf_dualhi_3; // @[el2_lsu_bus_buffer.scala 596:124]
-  wire  _T_4613 = _T_4611 | _T_4612; // @[el2_lsu_bus_buffer.scala 596:122]
-  wire  _T_4614 = _T_4551 & _T_4613; // @[el2_lsu_bus_buffer.scala 596:106]
-  wire [1:0] _T_4617 = _T_4606 ? 2'h2 : 2'h0; // @[Mux.scala 27:72]
-  wire [1:0] _T_4618 = _T_4614 ? 2'h3 : 2'h0; // @[Mux.scala 27:72]
-  wire [1:0] _GEN_390 = {{1'd0}, _T_4598}; // @[Mux.scala 27:72]
-  wire [1:0] _T_4620 = _GEN_390 | _T_4617; // @[Mux.scala 27:72]
-  wire [31:0] _T_4655 = _T_4590 ? buf_data_0 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4656 = _T_4598 ? buf_data_1 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4657 = _T_4606 ? buf_data_2 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4658 = _T_4614 ? buf_data_3 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4659 = _T_4655 | _T_4656; // @[Mux.scala 27:72]
-  wire [31:0] _T_4660 = _T_4659 | _T_4657; // @[Mux.scala 27:72]
-  wire [31:0] lsu_nonblock_load_data_lo = _T_4660 | _T_4658; // @[Mux.scala 27:72]
-  wire  _T_4666 = buf_dual_0 | buf_dualhi_0; // @[el2_lsu_bus_buffer.scala 598:120]
-  wire  _T_4667 = _T_4548 & _T_4666; // @[el2_lsu_bus_buffer.scala 598:105]
-  wire  _T_4672 = buf_dual_1 | buf_dualhi_1; // @[el2_lsu_bus_buffer.scala 598:120]
-  wire  _T_4673 = _T_4549 & _T_4672; // @[el2_lsu_bus_buffer.scala 598:105]
-  wire  _T_4678 = buf_dual_2 | buf_dualhi_2; // @[el2_lsu_bus_buffer.scala 598:120]
-  wire  _T_4679 = _T_4550 & _T_4678; // @[el2_lsu_bus_buffer.scala 598:105]
-  wire  _T_4684 = buf_dual_3 | buf_dualhi_3; // @[el2_lsu_bus_buffer.scala 598:120]
-  wire  _T_4685 = _T_4551 & _T_4684; // @[el2_lsu_bus_buffer.scala 598:105]
-  wire [31:0] _T_4686 = _T_4667 ? buf_data_0 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4687 = _T_4673 ? buf_data_1 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4688 = _T_4679 ? buf_data_2 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4689 = _T_4685 ? buf_data_3 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4690 = _T_4686 | _T_4687; // @[Mux.scala 27:72]
-  wire [31:0] _T_4691 = _T_4690 | _T_4688; // @[Mux.scala 27:72]
-  wire [31:0] lsu_nonblock_load_data_hi = _T_4691 | _T_4689; // @[Mux.scala 27:72]
-  wire  _T_4693 = io_lsu_nonblock_load_data_tag == 2'h0; // @[el2_lsu_bus_buffer.scala 120:123]
-  wire  _T_4694 = io_lsu_nonblock_load_data_tag == 2'h1; // @[el2_lsu_bus_buffer.scala 120:123]
-  wire  _T_4695 = io_lsu_nonblock_load_data_tag == 2'h2; // @[el2_lsu_bus_buffer.scala 120:123]
-  wire  _T_4696 = io_lsu_nonblock_load_data_tag == 2'h3; // @[el2_lsu_bus_buffer.scala 120:123]
-  wire [31:0] _T_4697 = _T_4693 ? buf_addr_0 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4698 = _T_4694 ? buf_addr_1 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4699 = _T_4695 ? buf_addr_2 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4700 = _T_4696 ? buf_addr_3 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4701 = _T_4697 | _T_4698; // @[Mux.scala 27:72]
+  wire  _T_4547 = _T_2814 & _T_3658; // @[Mux.scala 27:72]
+  wire  _T_4548 = _T_2836 & _T_3851; // @[Mux.scala 27:72]
+  wire  _T_4549 = _T_2858 & _T_4044; // @[Mux.scala 27:72]
+  wire  _T_4550 = _T_2880 & _T_4237; // @[Mux.scala 27:72]
+  wire  _T_4551 = _T_4547 | _T_4548; // @[Mux.scala 27:72]
+  wire  _T_4552 = _T_4551 | _T_4549; // @[Mux.scala 27:72]
+  wire  lsu_nonblock_load_data_ready = _T_4552 | _T_4550; // @[Mux.scala 27:72]
+  wire  _T_4558 = buf_error[0] & _T_3658; // @[el2_lsu_bus_buffer.scala 595:108]
+  wire  _T_4563 = buf_error[1] & _T_3851; // @[el2_lsu_bus_buffer.scala 595:108]
+  wire  _T_4568 = buf_error[2] & _T_4044; // @[el2_lsu_bus_buffer.scala 595:108]
+  wire  _T_4573 = buf_error[3] & _T_4237; // @[el2_lsu_bus_buffer.scala 595:108]
+  wire  _T_4574 = _T_2814 & _T_4558; // @[Mux.scala 27:72]
+  wire  _T_4575 = _T_2836 & _T_4563; // @[Mux.scala 27:72]
+  wire  _T_4576 = _T_2858 & _T_4568; // @[Mux.scala 27:72]
+  wire  _T_4577 = _T_2880 & _T_4573; // @[Mux.scala 27:72]
+  wire  _T_4578 = _T_4574 | _T_4575; // @[Mux.scala 27:72]
+  wire  _T_4579 = _T_4578 | _T_4576; // @[Mux.scala 27:72]
+  wire  _T_4586 = ~buf_dual_0; // @[el2_lsu_bus_buffer.scala 596:109]
+  wire  _T_4587 = ~buf_dualhi_0; // @[el2_lsu_bus_buffer.scala 596:124]
+  wire  _T_4588 = _T_4586 | _T_4587; // @[el2_lsu_bus_buffer.scala 596:122]
+  wire  _T_4589 = _T_4547 & _T_4588; // @[el2_lsu_bus_buffer.scala 596:106]
+  wire  _T_4594 = ~buf_dual_1; // @[el2_lsu_bus_buffer.scala 596:109]
+  wire  _T_4595 = ~buf_dualhi_1; // @[el2_lsu_bus_buffer.scala 596:124]
+  wire  _T_4596 = _T_4594 | _T_4595; // @[el2_lsu_bus_buffer.scala 596:122]
+  wire  _T_4597 = _T_4548 & _T_4596; // @[el2_lsu_bus_buffer.scala 596:106]
+  wire  _T_4602 = ~buf_dual_2; // @[el2_lsu_bus_buffer.scala 596:109]
+  wire  _T_4603 = ~buf_dualhi_2; // @[el2_lsu_bus_buffer.scala 596:124]
+  wire  _T_4604 = _T_4602 | _T_4603; // @[el2_lsu_bus_buffer.scala 596:122]
+  wire  _T_4605 = _T_4549 & _T_4604; // @[el2_lsu_bus_buffer.scala 596:106]
+  wire  _T_4610 = ~buf_dual_3; // @[el2_lsu_bus_buffer.scala 596:109]
+  wire  _T_4611 = ~buf_dualhi_3; // @[el2_lsu_bus_buffer.scala 596:124]
+  wire  _T_4612 = _T_4610 | _T_4611; // @[el2_lsu_bus_buffer.scala 596:122]
+  wire  _T_4613 = _T_4550 & _T_4612; // @[el2_lsu_bus_buffer.scala 596:106]
+  wire [1:0] _T_4616 = _T_4605 ? 2'h2 : 2'h0; // @[Mux.scala 27:72]
+  wire [1:0] _T_4617 = _T_4613 ? 2'h3 : 2'h0; // @[Mux.scala 27:72]
+  wire [1:0] _GEN_389 = {{1'd0}, _T_4597}; // @[Mux.scala 27:72]
+  wire [1:0] _T_4619 = _GEN_389 | _T_4616; // @[Mux.scala 27:72]
+  wire [31:0] _T_4654 = _T_4589 ? buf_data_0 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4655 = _T_4597 ? buf_data_1 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4656 = _T_4605 ? buf_data_2 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4657 = _T_4613 ? buf_data_3 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4658 = _T_4654 | _T_4655; // @[Mux.scala 27:72]
+  wire [31:0] _T_4659 = _T_4658 | _T_4656; // @[Mux.scala 27:72]
+  wire [31:0] lsu_nonblock_load_data_lo = _T_4659 | _T_4657; // @[Mux.scala 27:72]
+  wire  _T_4665 = buf_dual_0 | buf_dualhi_0; // @[el2_lsu_bus_buffer.scala 598:120]
+  wire  _T_4666 = _T_4547 & _T_4665; // @[el2_lsu_bus_buffer.scala 598:105]
+  wire  _T_4671 = buf_dual_1 | buf_dualhi_1; // @[el2_lsu_bus_buffer.scala 598:120]
+  wire  _T_4672 = _T_4548 & _T_4671; // @[el2_lsu_bus_buffer.scala 598:105]
+  wire  _T_4677 = buf_dual_2 | buf_dualhi_2; // @[el2_lsu_bus_buffer.scala 598:120]
+  wire  _T_4678 = _T_4549 & _T_4677; // @[el2_lsu_bus_buffer.scala 598:105]
+  wire  _T_4683 = buf_dual_3 | buf_dualhi_3; // @[el2_lsu_bus_buffer.scala 598:120]
+  wire  _T_4684 = _T_4550 & _T_4683; // @[el2_lsu_bus_buffer.scala 598:105]
+  wire [31:0] _T_4685 = _T_4666 ? buf_data_0 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4686 = _T_4672 ? buf_data_1 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4687 = _T_4678 ? buf_data_2 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4688 = _T_4684 ? buf_data_3 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4689 = _T_4685 | _T_4686; // @[Mux.scala 27:72]
+  wire [31:0] _T_4690 = _T_4689 | _T_4687; // @[Mux.scala 27:72]
+  wire [31:0] lsu_nonblock_load_data_hi = _T_4690 | _T_4688; // @[Mux.scala 27:72]
+  wire  _T_4692 = io_lsu_nonblock_load_data_tag == 2'h0; // @[el2_lsu_bus_buffer.scala 120:123]
+  wire  _T_4693 = io_lsu_nonblock_load_data_tag == 2'h1; // @[el2_lsu_bus_buffer.scala 120:123]
+  wire  _T_4694 = io_lsu_nonblock_load_data_tag == 2'h2; // @[el2_lsu_bus_buffer.scala 120:123]
+  wire  _T_4695 = io_lsu_nonblock_load_data_tag == 2'h3; // @[el2_lsu_bus_buffer.scala 120:123]
+  wire [31:0] _T_4696 = _T_4692 ? buf_addr_0 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4697 = _T_4693 ? buf_addr_1 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4698 = _T_4694 ? buf_addr_2 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4699 = _T_4695 ? buf_addr_3 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4700 = _T_4696 | _T_4697; // @[Mux.scala 27:72]
+  wire [31:0] _T_4701 = _T_4700 | _T_4698; // @[Mux.scala 27:72]
   wire [31:0] _T_4702 = _T_4701 | _T_4699; // @[Mux.scala 27:72]
-  wire [31:0] _T_4703 = _T_4702 | _T_4700; // @[Mux.scala 27:72]
-  wire [1:0] lsu_nonblock_addr_offset = _T_4703[1:0]; // @[el2_lsu_bus_buffer.scala 599:83]
-  wire [1:0] _T_4709 = _T_4693 ? buf_sz_0 : 2'h0; // @[Mux.scala 27:72]
-  wire [1:0] _T_4710 = _T_4694 ? buf_sz_1 : 2'h0; // @[Mux.scala 27:72]
-  wire [1:0] _T_4711 = _T_4695 ? buf_sz_2 : 2'h0; // @[Mux.scala 27:72]
-  wire [1:0] _T_4712 = _T_4696 ? buf_sz_3 : 2'h0; // @[Mux.scala 27:72]
-  wire [1:0] _T_4713 = _T_4709 | _T_4710; // @[Mux.scala 27:72]
-  wire [1:0] _T_4714 = _T_4713 | _T_4711; // @[Mux.scala 27:72]
-  wire [1:0] lsu_nonblock_sz = _T_4714 | _T_4712; // @[Mux.scala 27:72]
-  wire  _T_4724 = _T_4693 & buf_unsign[0]; // @[Mux.scala 27:72]
-  wire  _T_4725 = _T_4694 & buf_unsign[1]; // @[Mux.scala 27:72]
-  wire  _T_4726 = _T_4695 & buf_unsign[2]; // @[Mux.scala 27:72]
-  wire  _T_4727 = _T_4696 & buf_unsign[3]; // @[Mux.scala 27:72]
-  wire  _T_4728 = _T_4724 | _T_4725; // @[Mux.scala 27:72]
-  wire  _T_4729 = _T_4728 | _T_4726; // @[Mux.scala 27:72]
-  wire  lsu_nonblock_unsign = _T_4729 | _T_4727; // @[Mux.scala 27:72]
-  wire [63:0] _T_4749 = {lsu_nonblock_load_data_hi,lsu_nonblock_load_data_lo}; // @[Cat.scala 29:58]
-  wire [3:0] _GEN_391 = {{2'd0}, lsu_nonblock_addr_offset}; // @[el2_lsu_bus_buffer.scala 603:121]
-  wire [5:0] _T_4750 = _GEN_391 * 4'h8; // @[el2_lsu_bus_buffer.scala 603:121]
-  wire [63:0] lsu_nonblock_data_unalgn = _T_4749 >> _T_4750; // @[el2_lsu_bus_buffer.scala 603:92]
-  wire  _T_4751 = ~io_lsu_nonblock_load_data_error; // @[el2_lsu_bus_buffer.scala 606:69]
-  wire  _T_4753 = lsu_nonblock_sz == 2'h0; // @[el2_lsu_bus_buffer.scala 607:81]
-  wire  _T_4754 = lsu_nonblock_unsign & _T_4753; // @[el2_lsu_bus_buffer.scala 607:63]
-  wire [31:0] _T_4756 = {24'h0,lsu_nonblock_data_unalgn[7:0]}; // @[Cat.scala 29:58]
-  wire  _T_4757 = lsu_nonblock_sz == 2'h1; // @[el2_lsu_bus_buffer.scala 608:45]
-  wire  _T_4758 = lsu_nonblock_unsign & _T_4757; // @[el2_lsu_bus_buffer.scala 608:26]
-  wire [31:0] _T_4760 = {16'h0,lsu_nonblock_data_unalgn[15:0]}; // @[Cat.scala 29:58]
-  wire  _T_4761 = ~lsu_nonblock_unsign; // @[el2_lsu_bus_buffer.scala 609:6]
-  wire  _T_4763 = _T_4761 & _T_4753; // @[el2_lsu_bus_buffer.scala 609:27]
-  wire [23:0] _T_4766 = lsu_nonblock_data_unalgn[7] ? 24'hffffff : 24'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_4768 = {_T_4766,lsu_nonblock_data_unalgn[7:0]}; // @[Cat.scala 29:58]
-  wire  _T_4771 = _T_4761 & _T_4757; // @[el2_lsu_bus_buffer.scala 610:27]
-  wire [15:0] _T_4774 = lsu_nonblock_data_unalgn[15] ? 16'hffff : 16'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_4776 = {_T_4774,lsu_nonblock_data_unalgn[15:0]}; // @[Cat.scala 29:58]
-  wire  _T_4777 = lsu_nonblock_sz == 2'h2; // @[el2_lsu_bus_buffer.scala 611:21]
-  wire [31:0] _T_4778 = _T_4754 ? _T_4756 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4779 = _T_4758 ? _T_4760 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4780 = _T_4763 ? _T_4768 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4781 = _T_4771 ? _T_4776 : 32'h0; // @[Mux.scala 27:72]
-  wire [63:0] _T_4782 = _T_4777 ? lsu_nonblock_data_unalgn : 64'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4783 = _T_4778 | _T_4779; // @[Mux.scala 27:72]
+  wire [1:0] lsu_nonblock_addr_offset = _T_4702[1:0]; // @[el2_lsu_bus_buffer.scala 599:83]
+  wire [1:0] _T_4708 = _T_4692 ? buf_sz_0 : 2'h0; // @[Mux.scala 27:72]
+  wire [1:0] _T_4709 = _T_4693 ? buf_sz_1 : 2'h0; // @[Mux.scala 27:72]
+  wire [1:0] _T_4710 = _T_4694 ? buf_sz_2 : 2'h0; // @[Mux.scala 27:72]
+  wire [1:0] _T_4711 = _T_4695 ? buf_sz_3 : 2'h0; // @[Mux.scala 27:72]
+  wire [1:0] _T_4712 = _T_4708 | _T_4709; // @[Mux.scala 27:72]
+  wire [1:0] _T_4713 = _T_4712 | _T_4710; // @[Mux.scala 27:72]
+  wire [1:0] lsu_nonblock_sz = _T_4713 | _T_4711; // @[Mux.scala 27:72]
+  wire  _T_4723 = _T_4692 & buf_unsign[0]; // @[Mux.scala 27:72]
+  wire  _T_4724 = _T_4693 & buf_unsign[1]; // @[Mux.scala 27:72]
+  wire  _T_4725 = _T_4694 & buf_unsign[2]; // @[Mux.scala 27:72]
+  wire  _T_4726 = _T_4695 & buf_unsign[3]; // @[Mux.scala 27:72]
+  wire  _T_4727 = _T_4723 | _T_4724; // @[Mux.scala 27:72]
+  wire  _T_4728 = _T_4727 | _T_4725; // @[Mux.scala 27:72]
+  wire  lsu_nonblock_unsign = _T_4728 | _T_4726; // @[Mux.scala 27:72]
+  wire [63:0] _T_4748 = {lsu_nonblock_load_data_hi,lsu_nonblock_load_data_lo}; // @[Cat.scala 29:58]
+  wire [3:0] _GEN_390 = {{2'd0}, lsu_nonblock_addr_offset}; // @[el2_lsu_bus_buffer.scala 603:121]
+  wire [5:0] _T_4749 = _GEN_390 * 4'h8; // @[el2_lsu_bus_buffer.scala 603:121]
+  wire [63:0] lsu_nonblock_data_unalgn = _T_4748 >> _T_4749; // @[el2_lsu_bus_buffer.scala 603:92]
+  wire  _T_4750 = ~io_lsu_nonblock_load_data_error; // @[el2_lsu_bus_buffer.scala 606:69]
+  wire  _T_4752 = lsu_nonblock_sz == 2'h0; // @[el2_lsu_bus_buffer.scala 607:81]
+  wire  _T_4753 = lsu_nonblock_unsign & _T_4752; // @[el2_lsu_bus_buffer.scala 607:63]
+  wire [31:0] _T_4755 = {24'h0,lsu_nonblock_data_unalgn[7:0]}; // @[Cat.scala 29:58]
+  wire  _T_4756 = lsu_nonblock_sz == 2'h1; // @[el2_lsu_bus_buffer.scala 608:45]
+  wire  _T_4757 = lsu_nonblock_unsign & _T_4756; // @[el2_lsu_bus_buffer.scala 608:26]
+  wire [31:0] _T_4759 = {16'h0,lsu_nonblock_data_unalgn[15:0]}; // @[Cat.scala 29:58]
+  wire  _T_4760 = ~lsu_nonblock_unsign; // @[el2_lsu_bus_buffer.scala 609:6]
+  wire  _T_4762 = _T_4760 & _T_4752; // @[el2_lsu_bus_buffer.scala 609:27]
+  wire [23:0] _T_4765 = lsu_nonblock_data_unalgn[7] ? 24'hffffff : 24'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_4767 = {_T_4765,lsu_nonblock_data_unalgn[7:0]}; // @[Cat.scala 29:58]
+  wire  _T_4770 = _T_4760 & _T_4756; // @[el2_lsu_bus_buffer.scala 610:27]
+  wire [15:0] _T_4773 = lsu_nonblock_data_unalgn[15] ? 16'hffff : 16'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_4775 = {_T_4773,lsu_nonblock_data_unalgn[15:0]}; // @[Cat.scala 29:58]
+  wire  _T_4776 = lsu_nonblock_sz == 2'h2; // @[el2_lsu_bus_buffer.scala 611:21]
+  wire [31:0] _T_4777 = _T_4753 ? _T_4755 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4778 = _T_4757 ? _T_4759 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4779 = _T_4762 ? _T_4767 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4780 = _T_4770 ? _T_4775 : 32'h0; // @[Mux.scala 27:72]
+  wire [63:0] _T_4781 = _T_4776 ? lsu_nonblock_data_unalgn : 64'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4782 = _T_4777 | _T_4778; // @[Mux.scala 27:72]
+  wire [31:0] _T_4783 = _T_4782 | _T_4779; // @[Mux.scala 27:72]
   wire [31:0] _T_4784 = _T_4783 | _T_4780; // @[Mux.scala 27:72]
-  wire [31:0] _T_4785 = _T_4784 | _T_4781; // @[Mux.scala 27:72]
-  wire [63:0] _GEN_392 = {{32'd0}, _T_4785}; // @[Mux.scala 27:72]
-  wire [63:0] _T_4786 = _GEN_392 | _T_4782; // @[Mux.scala 27:72]
-  wire  _T_4881 = obuf_valid & obuf_write; // @[el2_lsu_bus_buffer.scala 629:36]
-  wire  _T_4882 = ~obuf_cmd_done; // @[el2_lsu_bus_buffer.scala 629:51]
-  wire  _T_4883 = _T_4881 & _T_4882; // @[el2_lsu_bus_buffer.scala 629:49]
-  wire [31:0] _T_4887 = {obuf_addr[31:3],3'h0}; // @[Cat.scala 29:58]
-  wire [2:0] _T_4889 = {1'h0,obuf_sz}; // @[Cat.scala 29:58]
-  wire  _T_4894 = ~obuf_data_done; // @[el2_lsu_bus_buffer.scala 641:50]
-  wire  _T_4895 = _T_4881 & _T_4894; // @[el2_lsu_bus_buffer.scala 641:48]
-  wire [7:0] _T_4899 = obuf_write ? 8'hff : 8'h0; // @[Bitwise.scala 72:12]
-  wire  _T_4902 = obuf_valid & _T_1356; // @[el2_lsu_bus_buffer.scala 646:36]
-  wire  _T_4904 = _T_4902 & _T_1362; // @[el2_lsu_bus_buffer.scala 646:50]
-  wire  _T_4916 = io_lsu_bus_clk_en_q & buf_error[0]; // @[el2_lsu_bus_buffer.scala 659:114]
-  wire  _T_4918 = _T_4916 & buf_write[0]; // @[el2_lsu_bus_buffer.scala 659:129]
-  wire  _T_4921 = io_lsu_bus_clk_en_q & buf_error[1]; // @[el2_lsu_bus_buffer.scala 659:114]
-  wire  _T_4923 = _T_4921 & buf_write[1]; // @[el2_lsu_bus_buffer.scala 659:129]
-  wire  _T_4926 = io_lsu_bus_clk_en_q & buf_error[2]; // @[el2_lsu_bus_buffer.scala 659:114]
-  wire  _T_4928 = _T_4926 & buf_write[2]; // @[el2_lsu_bus_buffer.scala 659:129]
-  wire  _T_4931 = io_lsu_bus_clk_en_q & buf_error[3]; // @[el2_lsu_bus_buffer.scala 659:114]
-  wire  _T_4933 = _T_4931 & buf_write[3]; // @[el2_lsu_bus_buffer.scala 659:129]
-  wire  _T_4934 = _T_2814 & _T_4918; // @[Mux.scala 27:72]
-  wire  _T_4935 = _T_2836 & _T_4923; // @[Mux.scala 27:72]
-  wire  _T_4936 = _T_2858 & _T_4928; // @[Mux.scala 27:72]
-  wire  _T_4937 = _T_2880 & _T_4933; // @[Mux.scala 27:72]
-  wire  _T_4938 = _T_4934 | _T_4935; // @[Mux.scala 27:72]
-  wire  _T_4939 = _T_4938 | _T_4936; // @[Mux.scala 27:72]
-  wire  _T_4949 = _T_2836 & buf_error[1]; // @[el2_lsu_bus_buffer.scala 660:98]
-  wire  lsu_imprecise_error_store_tag = _T_4949 & buf_write[1]; // @[el2_lsu_bus_buffer.scala 660:113]
-  wire  _T_4955 = ~io_lsu_imprecise_error_store_any; // @[el2_lsu_bus_buffer.scala 662:72]
-  wire  _T_4957 = ~lsu_imprecise_error_store_tag; // @[el2_lsu_bus_buffer.scala 120:123]
-  wire [31:0] _T_4959 = _T_4957 ? buf_addr_0 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4960 = lsu_imprecise_error_store_tag ? buf_addr_1 : 32'h0; // @[Mux.scala 27:72]
-  wire [31:0] _T_4961 = _T_4959 | _T_4960; // @[Mux.scala 27:72]
-  wire  _T_4978 = bus_wcmd_sent | bus_wdata_sent; // @[el2_lsu_bus_buffer.scala 669:68]
-  wire  _T_4981 = io_lsu_busreq_r & io_ldst_dual_r; // @[el2_lsu_bus_buffer.scala 670:48]
-  wire  _T_4984 = ~io_lsu_axi_awready; // @[el2_lsu_bus_buffer.scala 673:48]
-  wire  _T_4985 = io_lsu_axi_awvalid & _T_4984; // @[el2_lsu_bus_buffer.scala 673:46]
-  wire  _T_4986 = ~io_lsu_axi_wready; // @[el2_lsu_bus_buffer.scala 673:92]
-  wire  _T_4987 = io_lsu_axi_wvalid & _T_4986; // @[el2_lsu_bus_buffer.scala 673:90]
-  wire  _T_4988 = _T_4985 | _T_4987; // @[el2_lsu_bus_buffer.scala 673:69]
-  wire  _T_4989 = ~io_lsu_axi_arready; // @[el2_lsu_bus_buffer.scala 673:136]
-  wire  _T_4990 = io_lsu_axi_arvalid & _T_4989; // @[el2_lsu_bus_buffer.scala 673:134]
-  wire  _T_4994 = ~io_flush_r; // @[el2_lsu_bus_buffer.scala 677:75]
-  wire  _T_4995 = io_lsu_busreq_m & _T_4994; // @[el2_lsu_bus_buffer.scala 677:73]
-  reg  _T_4998; // @[el2_lsu_bus_buffer.scala 677:56]
+  wire [63:0] _GEN_391 = {{32'd0}, _T_4784}; // @[Mux.scala 27:72]
+  wire [63:0] _T_4785 = _GEN_391 | _T_4781; // @[Mux.scala 27:72]
+  wire  _T_4880 = obuf_valid & obuf_write; // @[el2_lsu_bus_buffer.scala 629:36]
+  wire  _T_4881 = ~obuf_cmd_done; // @[el2_lsu_bus_buffer.scala 629:51]
+  wire  _T_4882 = _T_4880 & _T_4881; // @[el2_lsu_bus_buffer.scala 629:49]
+  wire [31:0] _T_4886 = {obuf_addr[31:3],3'h0}; // @[Cat.scala 29:58]
+  wire [2:0] _T_4888 = {1'h0,obuf_sz}; // @[Cat.scala 29:58]
+  wire  _T_4893 = ~obuf_data_done; // @[el2_lsu_bus_buffer.scala 641:50]
+  wire  _T_4894 = _T_4880 & _T_4893; // @[el2_lsu_bus_buffer.scala 641:48]
+  wire [7:0] _T_4898 = obuf_write ? 8'hff : 8'h0; // @[Bitwise.scala 72:12]
+  wire  _T_4901 = obuf_valid & _T_1356; // @[el2_lsu_bus_buffer.scala 646:36]
+  wire  _T_4903 = _T_4901 & _T_1362; // @[el2_lsu_bus_buffer.scala 646:50]
+  wire  _T_4915 = io_lsu_bus_clk_en_q & buf_error[0]; // @[el2_lsu_bus_buffer.scala 659:114]
+  wire  _T_4917 = _T_4915 & buf_write[0]; // @[el2_lsu_bus_buffer.scala 659:129]
+  wire  _T_4920 = io_lsu_bus_clk_en_q & buf_error[1]; // @[el2_lsu_bus_buffer.scala 659:114]
+  wire  _T_4922 = _T_4920 & buf_write[1]; // @[el2_lsu_bus_buffer.scala 659:129]
+  wire  _T_4925 = io_lsu_bus_clk_en_q & buf_error[2]; // @[el2_lsu_bus_buffer.scala 659:114]
+  wire  _T_4927 = _T_4925 & buf_write[2]; // @[el2_lsu_bus_buffer.scala 659:129]
+  wire  _T_4930 = io_lsu_bus_clk_en_q & buf_error[3]; // @[el2_lsu_bus_buffer.scala 659:114]
+  wire  _T_4932 = _T_4930 & buf_write[3]; // @[el2_lsu_bus_buffer.scala 659:129]
+  wire  _T_4933 = _T_2814 & _T_4917; // @[Mux.scala 27:72]
+  wire  _T_4934 = _T_2836 & _T_4922; // @[Mux.scala 27:72]
+  wire  _T_4935 = _T_2858 & _T_4927; // @[Mux.scala 27:72]
+  wire  _T_4936 = _T_2880 & _T_4932; // @[Mux.scala 27:72]
+  wire  _T_4937 = _T_4933 | _T_4934; // @[Mux.scala 27:72]
+  wire  _T_4938 = _T_4937 | _T_4935; // @[Mux.scala 27:72]
+  wire  _T_4948 = _T_2836 & buf_error[1]; // @[el2_lsu_bus_buffer.scala 660:98]
+  wire  lsu_imprecise_error_store_tag = _T_4948 & buf_write[1]; // @[el2_lsu_bus_buffer.scala 660:113]
+  wire  _T_4954 = ~io_lsu_imprecise_error_store_any; // @[el2_lsu_bus_buffer.scala 662:72]
+  wire  _T_4956 = ~lsu_imprecise_error_store_tag; // @[el2_lsu_bus_buffer.scala 120:123]
+  wire [31:0] _T_4958 = _T_4956 ? buf_addr_0 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4959 = lsu_imprecise_error_store_tag ? buf_addr_1 : 32'h0; // @[Mux.scala 27:72]
+  wire [31:0] _T_4960 = _T_4958 | _T_4959; // @[Mux.scala 27:72]
+  wire  _T_4977 = bus_wcmd_sent | bus_wdata_sent; // @[el2_lsu_bus_buffer.scala 669:68]
+  wire  _T_4980 = io_lsu_busreq_r & io_ldst_dual_r; // @[el2_lsu_bus_buffer.scala 670:48]
+  wire  _T_4983 = ~io_lsu_axi_awready; // @[el2_lsu_bus_buffer.scala 673:48]
+  wire  _T_4984 = io_lsu_axi_awvalid & _T_4983; // @[el2_lsu_bus_buffer.scala 673:46]
+  wire  _T_4985 = ~io_lsu_axi_wready; // @[el2_lsu_bus_buffer.scala 673:92]
+  wire  _T_4986 = io_lsu_axi_wvalid & _T_4985; // @[el2_lsu_bus_buffer.scala 673:90]
+  wire  _T_4987 = _T_4984 | _T_4986; // @[el2_lsu_bus_buffer.scala 673:69]
+  wire  _T_4988 = ~io_lsu_axi_arready; // @[el2_lsu_bus_buffer.scala 673:136]
+  wire  _T_4989 = io_lsu_axi_arvalid & _T_4988; // @[el2_lsu_bus_buffer.scala 673:134]
+  wire  _T_4993 = ~io_flush_r; // @[el2_lsu_bus_buffer.scala 677:75]
+  wire  _T_4994 = io_lsu_busreq_m & _T_4993; // @[el2_lsu_bus_buffer.scala 677:73]
+  reg  _T_4997; // @[el2_lsu_bus_buffer.scala 677:56]
   rvclkhdr rvclkhdr ( // @[el2_lib.scala 485:23]
     .io_l1clk(rvclkhdr_io_l1clk),
     .io_clk(rvclkhdr_io_clk),
@@ -2701,52 +2703,52 @@ module el2_lsu_bus_buffer(
     .io_en(rvclkhdr_11_io_en),
     .io_scan_mode(rvclkhdr_11_io_scan_mode)
   );
-  assign io_lsu_busreq_r = _T_4998; // @[el2_lsu_bus_buffer.scala 677:19]
+  assign io_lsu_busreq_r = _T_4997; // @[el2_lsu_bus_buffer.scala 677:19]
   assign io_lsu_bus_buffer_pend_any = |buf_numvld_pend_any; // @[el2_lsu_bus_buffer.scala 585:30]
-  assign io_lsu_bus_buffer_full_any = _T_4508 ? _T_4509 : _T_4510; // @[el2_lsu_bus_buffer.scala 586:30]
-  assign io_lsu_bus_buffer_empty_any = _T_4521 & _T_1244; // @[el2_lsu_bus_buffer.scala 587:31]
+  assign io_lsu_bus_buffer_full_any = _T_4507 ? _T_4508 : _T_4509; // @[el2_lsu_bus_buffer.scala 586:30]
+  assign io_lsu_bus_buffer_empty_any = _T_4520 & _T_1244; // @[el2_lsu_bus_buffer.scala 587:31]
   assign io_lsu_bus_idle_any = 1'h1; // @[el2_lsu_bus_buffer.scala 666:23]
   assign io_ld_byte_hit_buf_lo = {_T_69,_T_58}; // @[el2_lsu_bus_buffer.scala 198:25]
   assign io_ld_byte_hit_buf_hi = {_T_84,_T_73}; // @[el2_lsu_bus_buffer.scala 199:25]
   assign io_ld_fwddata_buf_lo = _T_646[31:0]; // @[el2_lsu_bus_buffer.scala 224:24]
   assign io_ld_fwddata_buf_hi = _T_741[31:0]; // @[el2_lsu_bus_buffer.scala 229:24]
-  assign io_lsu_imprecise_error_load_any = io_lsu_nonblock_load_data_error & _T_4955; // @[el2_lsu_bus_buffer.scala 662:35]
-  assign io_lsu_imprecise_error_store_any = _T_4939 | _T_4937; // @[el2_lsu_bus_buffer.scala 659:36]
-  assign io_lsu_imprecise_error_addr_any = io_lsu_imprecise_error_store_any ? _T_4961 : _T_4703; // @[el2_lsu_bus_buffer.scala 663:35]
-  assign io_lsu_nonblock_load_valid_m = _T_4527 & _T_4528; // @[el2_lsu_bus_buffer.scala 589:32]
+  assign io_lsu_imprecise_error_load_any = io_lsu_nonblock_load_data_error & _T_4954; // @[el2_lsu_bus_buffer.scala 662:35]
+  assign io_lsu_imprecise_error_store_any = _T_4938 | _T_4936; // @[el2_lsu_bus_buffer.scala 659:36]
+  assign io_lsu_imprecise_error_addr_any = io_lsu_imprecise_error_store_any ? _T_4960 : _T_4702; // @[el2_lsu_bus_buffer.scala 663:35]
+  assign io_lsu_nonblock_load_valid_m = _T_4526 & _T_4527; // @[el2_lsu_bus_buffer.scala 589:32]
   assign io_lsu_nonblock_load_tag_m = _T_1876 ? 2'h0 : _T_1912; // @[el2_lsu_bus_buffer.scala 590:30]
-  assign io_lsu_nonblock_load_inv_r = lsu_nonblock_load_valid_r & _T_4530; // @[el2_lsu_bus_buffer.scala 592:30]
+  assign io_lsu_nonblock_load_inv_r = lsu_nonblock_load_valid_r & _T_4529; // @[el2_lsu_bus_buffer.scala 592:30]
   assign io_lsu_nonblock_load_inv_tag_r = WrPtr0_r; // @[el2_lsu_bus_buffer.scala 593:34]
-  assign io_lsu_nonblock_load_data_valid = lsu_nonblock_load_data_ready & _T_4751; // @[el2_lsu_bus_buffer.scala 606:35]
-  assign io_lsu_nonblock_load_data_error = _T_4580 | _T_4578; // @[el2_lsu_bus_buffer.scala 595:35]
-  assign io_lsu_nonblock_load_data_tag = _T_4620 | _T_4618; // @[el2_lsu_bus_buffer.scala 596:33]
-  assign io_lsu_nonblock_load_data = _T_4786[31:0]; // @[el2_lsu_bus_buffer.scala 607:29]
-  assign io_lsu_pmu_bus_trxn = _T_4978 | _T_4873; // @[el2_lsu_bus_buffer.scala 669:23]
-  assign io_lsu_pmu_bus_misaligned = _T_4981 & io_lsu_commit_r; // @[el2_lsu_bus_buffer.scala 670:29]
+  assign io_lsu_nonblock_load_data_valid = lsu_nonblock_load_data_ready & _T_4750; // @[el2_lsu_bus_buffer.scala 606:35]
+  assign io_lsu_nonblock_load_data_error = _T_4579 | _T_4577; // @[el2_lsu_bus_buffer.scala 595:35]
+  assign io_lsu_nonblock_load_data_tag = _T_4619 | _T_4617; // @[el2_lsu_bus_buffer.scala 596:33]
+  assign io_lsu_nonblock_load_data = _T_4785[31:0]; // @[el2_lsu_bus_buffer.scala 607:29]
+  assign io_lsu_pmu_bus_trxn = _T_4977 | _T_4872; // @[el2_lsu_bus_buffer.scala 669:23]
+  assign io_lsu_pmu_bus_misaligned = _T_4980 & io_lsu_commit_r; // @[el2_lsu_bus_buffer.scala 670:29]
   assign io_lsu_pmu_bus_error = io_lsu_imprecise_error_load_any | io_lsu_imprecise_error_store_any; // @[el2_lsu_bus_buffer.scala 671:24]
-  assign io_lsu_pmu_bus_busy = _T_4988 | _T_4990; // @[el2_lsu_bus_buffer.scala 673:23]
-  assign io_lsu_axi_awvalid = _T_4883 & _T_1252; // @[el2_lsu_bus_buffer.scala 629:22]
+  assign io_lsu_pmu_bus_busy = _T_4987 | _T_4989; // @[el2_lsu_bus_buffer.scala 673:23]
+  assign io_lsu_axi_awvalid = _T_4882 & _T_1252; // @[el2_lsu_bus_buffer.scala 629:22]
   assign io_lsu_axi_awid = {{1'd0}, _T_1861}; // @[el2_lsu_bus_buffer.scala 630:19]
-  assign io_lsu_axi_awaddr = obuf_sideeffect ? obuf_addr : _T_4887; // @[el2_lsu_bus_buffer.scala 631:21]
+  assign io_lsu_axi_awaddr = obuf_sideeffect ? obuf_addr : _T_4886; // @[el2_lsu_bus_buffer.scala 631:21]
   assign io_lsu_axi_awregion = obuf_addr[31:28]; // @[el2_lsu_bus_buffer.scala 635:23]
   assign io_lsu_axi_awlen = 8'h0; // @[el2_lsu_bus_buffer.scala 636:20]
-  assign io_lsu_axi_awsize = obuf_sideeffect ? _T_4889 : 3'h3; // @[el2_lsu_bus_buffer.scala 632:21]
+  assign io_lsu_axi_awsize = obuf_sideeffect ? _T_4888 : 3'h3; // @[el2_lsu_bus_buffer.scala 632:21]
   assign io_lsu_axi_awburst = 2'h1; // @[el2_lsu_bus_buffer.scala 637:22]
   assign io_lsu_axi_awlock = 1'h0; // @[el2_lsu_bus_buffer.scala 639:21]
   assign io_lsu_axi_awcache = obuf_sideeffect ? 4'h0 : 4'hf; // @[el2_lsu_bus_buffer.scala 634:22]
   assign io_lsu_axi_awprot = 3'h0; // @[el2_lsu_bus_buffer.scala 633:21]
   assign io_lsu_axi_awqos = 4'h0; // @[el2_lsu_bus_buffer.scala 638:20]
-  assign io_lsu_axi_wvalid = _T_4895 & _T_1252; // @[el2_lsu_bus_buffer.scala 641:21]
+  assign io_lsu_axi_wvalid = _T_4894 & _T_1252; // @[el2_lsu_bus_buffer.scala 641:21]
   assign io_lsu_axi_wdata = obuf_data; // @[el2_lsu_bus_buffer.scala 643:20]
-  assign io_lsu_axi_wstrb = obuf_byteen & _T_4899; // @[el2_lsu_bus_buffer.scala 642:20]
+  assign io_lsu_axi_wstrb = obuf_byteen & _T_4898; // @[el2_lsu_bus_buffer.scala 642:20]
   assign io_lsu_axi_wlast = 1'h1; // @[el2_lsu_bus_buffer.scala 644:20]
   assign io_lsu_axi_bready = 1'h1; // @[el2_lsu_bus_buffer.scala 657:21]
-  assign io_lsu_axi_arvalid = _T_4904 & _T_1252; // @[el2_lsu_bus_buffer.scala 646:22]
+  assign io_lsu_axi_arvalid = _T_4903 & _T_1252; // @[el2_lsu_bus_buffer.scala 646:22]
   assign io_lsu_axi_arid = {{1'd0}, _T_1861}; // @[el2_lsu_bus_buffer.scala 647:19]
-  assign io_lsu_axi_araddr = obuf_sideeffect ? obuf_addr : _T_4887; // @[el2_lsu_bus_buffer.scala 648:21]
+  assign io_lsu_axi_araddr = obuf_sideeffect ? obuf_addr : _T_4886; // @[el2_lsu_bus_buffer.scala 648:21]
   assign io_lsu_axi_arregion = obuf_addr[31:28]; // @[el2_lsu_bus_buffer.scala 652:23]
   assign io_lsu_axi_arlen = 8'h0; // @[el2_lsu_bus_buffer.scala 653:20]
-  assign io_lsu_axi_arsize = obuf_sideeffect ? _T_4889 : 3'h3; // @[el2_lsu_bus_buffer.scala 649:21]
+  assign io_lsu_axi_arsize = obuf_sideeffect ? _T_4888 : 3'h3; // @[el2_lsu_bus_buffer.scala 649:21]
   assign io_lsu_axi_arburst = 2'h1; // @[el2_lsu_bus_buffer.scala 654:22]
   assign io_lsu_axi_arlock = 1'h0; // @[el2_lsu_bus_buffer.scala 656:21]
   assign io_lsu_axi_arcache = obuf_sideeffect ? 4'h0 : 4'hf; // @[el2_lsu_bus_buffer.scala 651:22]
@@ -2754,8 +2756,8 @@ module el2_lsu_bus_buffer(
   assign io_lsu_axi_arqos = 4'h0; // @[el2_lsu_bus_buffer.scala 655:20]
   assign io_lsu_axi_rready = 1'h1; // @[el2_lsu_bus_buffer.scala 658:21]
   assign io_test = _T_745 | _T_744; // @[el2_lsu_bus_buffer.scala 258:11]
-  assign io_data_hi = _T_4691 | _T_4689; // @[el2_lsu_bus_buffer.scala 604:14]
-  assign io_data_lo = _T_4660 | _T_4658; // @[el2_lsu_bus_buffer.scala 605:14]
+  assign io_data_hi = _T_4690 | _T_4688; // @[el2_lsu_bus_buffer.scala 604:14]
+  assign io_data_lo = _T_4659 | _T_4657; // @[el2_lsu_bus_buffer.scala 605:14]
   assign io_data_en = {_T_4431,buf_data_en_0}; // @[el2_lsu_bus_buffer.scala 578:14]
   assign io_Cmdptr0 = _T_2092[1:0]; // @[el2_lsu_bus_buffer.scala 447:14]
   assign io_Cmdptr1 = 2'h0; // @[el2_lsu_bus_buffer.scala 358:14]
@@ -3047,7 +3049,7 @@ initial begin
   _RAND_105 = {1{`RANDOM}};
   lsu_nonblock_load_valid_r = _RAND_105[0:0];
   _RAND_106 = {1{`RANDOM}};
-  _T_4998 = _RAND_106[0:0];
+  _T_4997 = _RAND_106[0:0];
 `endif // RANDOMIZE_REG_INIT
   if (reset) begin
     buf_addr_0 = 32'h0;
@@ -3368,7 +3370,7 @@ initial begin
     lsu_nonblock_load_valid_r = 1'h0;
   end
   if (reset) begin
-    _T_4998 = 1'h0;
+    _T_4997 = 1'h0;
   end
   `endif // RANDOMIZE
 end // initial
@@ -4195,14 +4197,14 @@ end // initial
     if (reset) begin
       obuf_cmd_done <= 1'h0;
     end else begin
-      obuf_cmd_done <= _T_1318 & _T_4870;
+      obuf_cmd_done <= _T_1318 & _T_4869;
     end
   end
   always @(posedge io_lsu_busm_clk or posedge reset) begin
     if (reset) begin
       obuf_data_done <= 1'h0;
     end else begin
-      obuf_data_done <= _T_1318 & _T_4871;
+      obuf_data_done <= _T_1318 & _T_4870;
     end
   end
   always @(posedge io_lsu_free_c2_clk or posedge reset) begin
@@ -4600,9 +4602,9 @@ end // initial
   end
   always @(posedge io_lsu_c2_r_clk or posedge reset) begin
     if (reset) begin
-      _T_4998 <= 1'h0;
+      _T_4997 <= 1'h0;
     end else begin
-      _T_4998 <= _T_4995 & _T_4528;
+      _T_4997 <= _T_4994 & _T_4527;
     end
   end
 endmodule
diff --git a/src/main/scala/lsu/el2_lsu_bus_buffer.scala b/src/main/scala/lsu/el2_lsu_bus_buffer.scala
index dc50ed48..183042eb 100644
--- a/src/main/scala/lsu/el2_lsu_bus_buffer.scala
+++ b/src/main/scala/lsu/el2_lsu_bus_buffer.scala
@@ -577,7 +577,7 @@ class  el2_lsu_bus_buffer extends Module with RequireAsyncReset with el2_lib {
     buf_error := (0 until DEPTH).map(i=>(withClock(io.lsu_bus_buf_c1_clk){RegNext(Mux(buf_error_en(i), true.B, buf_error(i)) & !buf_rst(i), false.B)}).asUInt()).reverse.reduce(Cat(_,_))
   io.data_en := (0 until DEPTH).map(i=>buf_data_en(i).asUInt()).reverse.reduce(Cat(_,_))
 
-  val buf_numvld_any = (0 until DEPTH).map(i=>(buf_state(i)=/=idle_C).asUInt).reverse.reduce(_ +& _)
+  val buf_numvld_any = Mux(io.lsu_busreq_m, Cat(io.lsu_busreq_m, 0.U),io.lsu_busreq_m) +& Mux(io.lsu_busreq_r, Cat(io.lsu_busreq_r, 0.U),io.lsu_busreq_r) +& ibuf_valid
   buf_numvld_wrcmd_any := (0 until DEPTH).map(i=>(buf_write(i) & (buf_state(i)===cmd_C) & !buf_cmd_state_bus_en(i)).asUInt).reverse.reduce(_ +& _)
   buf_numvld_cmd_any :=  (0 until DEPTH).map(i=>((buf_state(i)===cmd_C) & !buf_cmd_state_bus_en(i)).asUInt).reverse.reduce(_ +& _)
   buf_numvld_pend_any := (0 until DEPTH).map(i=>((buf_state(i)===wait_C) | ((buf_state(i)===cmd_C) & !buf_cmd_state_bus_en(i))).asUInt).reverse.reduce(_ +& _)
diff --git a/target/scala-2.12/classes/lsu/el2_lsu_bus_buffer.class b/target/scala-2.12/classes/lsu/el2_lsu_bus_buffer.class
index fb9ec507..75f6c02f 100644
Binary files a/target/scala-2.12/classes/lsu/el2_lsu_bus_buffer.class and b/target/scala-2.12/classes/lsu/el2_lsu_bus_buffer.class differ