LSU Top updated

lsu.anno.json

@@ -180,6 +180,7 @@
     "sink":"~lsu|lsu>io_lsu_store_stall_any",
     "sources":[
       "~lsu|lsu>io_dec_lsu_valid_raw_d",
+      "~lsu|lsu>io_dec_tlu_flush_lower_r",
       "~lsu|lsu>io_lsu_exu_exu_lsu_rs1_d",
       "~lsu|lsu>io_lsu_dma_dma_lsc_ctl_dma_mem_addr",
       "~lsu|lsu>io_dec_lsu_offset_d",
@@ -188,7 +189,6 @@
       "~lsu|lsu>io_lsu_p_bits_half",
       "~lsu|lsu>io_lsu_p_bits_word",
       "~lsu|lsu>io_lsu_dma_dma_lsc_ctl_dma_mem_sz",
-      "~lsu|lsu>io_dec_tlu_flush_lower_r",
       "~lsu|lsu>io_dec_tlu_force_halt",
       "~lsu|lsu>io_lsu_pic_picm_rd_data",
       "~lsu|lsu>io_dccm_rd_data_hi",
@@ -380,6 +380,7 @@
     "sink":"~lsu|lsu>io_lsu_load_stall_any",
     "sources":[
       "~lsu|lsu>io_dec_lsu_valid_raw_d",
+      "~lsu|lsu>io_dec_tlu_flush_lower_r",
       "~lsu|lsu>io_lsu_exu_exu_lsu_rs1_d",
       "~lsu|lsu>io_lsu_dma_dma_lsc_ctl_dma_mem_addr",
       "~lsu|lsu>io_dec_lsu_offset_d",
@@ -388,7 +389,6 @@
       "~lsu|lsu>io_lsu_p_bits_half",
       "~lsu|lsu>io_lsu_p_bits_word",
       "~lsu|lsu>io_lsu_dma_dma_lsc_ctl_dma_mem_sz",
-      "~lsu|lsu>io_dec_tlu_flush_lower_r",
       "~lsu|lsu>io_dec_tlu_force_halt",
       "~lsu|lsu>io_lsu_pic_picm_rd_data",
       "~lsu|lsu>io_dccm_rd_data_hi",

lsu_lsc_ctl.anno.json

@@ -1,11 +1,4 @@
 [
-  {
-    "class":"firrtl.transforms.CombinationalPath",
-    "sink":"~lsu_lsc_ctl|lsu_lsc_ctl>io_end_addr_r",
-    "sources":[
-      "~lsu_lsc_ctl|lsu_lsc_ctl>io_lsu_addr_r"
-    ]
-  },
   {
     "class":"firrtl.transforms.CombinationalPath",
     "sink":"~lsu_lsc_ctl|lsu_lsc_ctl>io_lsu_pkt_d_bits_word",
@@ -31,6 +24,14 @@
       "~lsu_lsc_ctl|lsu_lsc_ctl>io_dec_lsu_valid_raw_d"
     ]
   },
+  {
+    "class":"firrtl.transforms.CombinationalPath",
+    "sink":"~lsu_lsc_ctl|lsu_lsc_ctl>io_end_addr_r",
+    "sources":[
+      "~lsu_lsc_ctl|lsu_lsc_ctl>io_ldst_dual_r",
+      "~lsu_lsc_ctl|lsu_lsc_ctl>io_lsu_addr_r"
+    ]
+  },
   {
     "class":"firrtl.transforms.CombinationalPath",
     "sink":"~lsu_lsc_ctl|lsu_lsc_ctl>io_lsu_pkt_d_bits_store",
@@ -56,6 +57,14 @@
       "~lsu_lsc_ctl|lsu_lsc_ctl>io_lsu_ld_data_corr_r"
     ]
   },
+  {
+    "class":"firrtl.transforms.CombinationalPath",
+    "sink":"~lsu_lsc_ctl|lsu_lsc_ctl>io_end_addr_m",
+    "sources":[
+      "~lsu_lsc_ctl|lsu_lsc_ctl>io_ldst_dual_m",
+      "~lsu_lsc_ctl|lsu_lsc_ctl>io_lsu_addr_m"
+    ]
+  },
   {
     "class":"firrtl.transforms.CombinationalPath",
     "sink":"~lsu_lsc_ctl|lsu_lsc_ctl>io_store_data_m",
@@ -109,13 +118,6 @@
       "~lsu_lsc_ctl|lsu_lsc_ctl>io_dec_lsu_valid_raw_d"
     ]
   },
-  {
-    "class":"firrtl.transforms.CombinationalPath",
-    "sink":"~lsu_lsc_ctl|lsu_lsc_ctl>io_end_addr_m",
-    "sources":[
-      "~lsu_lsc_ctl|lsu_lsc_ctl>io_lsu_addr_m"
-    ]
-  },
   {
     "class":"firrtl.transforms.CombinationalPath",
     "sink":"~lsu_lsc_ctl|lsu_lsc_ctl>io_lsu_pkt_d_bits_stack",

lsu_lsc_ctl.fir

@@ -783,205 +783,195 @@ circuit lsu_lsc_ctl :
     node _T_151 = bits(io.lsu_exu.lsu_result_m, 31, 0) @[lsu_lsc_ctl.scala 223:103]
     node _T_152 = bits(store_data_d, 31, 0) @[lsu_lsc_ctl.scala 223:122]
     node store_data_m_in = mux(_T_150, _T_151, _T_152) @[lsu_lsc_ctl.scala 223:34]
-    node _T_153 = bits(io.lsu_addr_d, 2, 2) @[lsu_lsc_ctl.scala 224:62]
+    reg store_data_pre_m : UInt, io.lsu_store_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 225:72]
-    reg _T_154 : UInt<1>, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 224:48]
+    store_data_pre_m <= store_data_m_in @[lsu_lsc_ctl.scala 225:72]
-    _T_154 <= _T_153 @[lsu_lsc_ctl.scala 224:48]
+    reg _T_153 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 226:62]
-    node _T_155 = bits(io.end_addr_d, 2, 2) @[lsu_lsc_ctl.scala 224:124]
+    _T_153 <= io.lsu_addr_d @[lsu_lsc_ctl.scala 226:62]
-    reg _T_156 : UInt<1>, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 224:110]
+    io.lsu_addr_m <= _T_153 @[lsu_lsc_ctl.scala 226:24]
-    _T_156 <= _T_155 @[lsu_lsc_ctl.scala 224:110]
+    reg _T_154 : UInt, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 227:62]
-    node int = neq(_T_154, _T_156) @[lsu_lsc_ctl.scala 224:72]
+    _T_154 <= io.lsu_addr_m @[lsu_lsc_ctl.scala 227:62]
-    node _T_157 = bits(io.lsu_addr_m, 2, 2) @[lsu_lsc_ctl.scala 225:62]
+    io.lsu_addr_r <= _T_154 @[lsu_lsc_ctl.scala 227:24]
-    reg _T_158 : UInt<1>, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 225:48]
+    node _T_155 = bits(io.ldst_dual_m, 0, 0) @[lib.scala 8:44]
-    _T_158 <= _T_157 @[lsu_lsc_ctl.scala 225:48]
+    node _T_156 = bits(io.lsu_addr_m, 31, 3) @[lsu_lsc_ctl.scala 228:71]
-    node _T_159 = bits(io.end_addr_m, 2, 2) @[lsu_lsc_ctl.scala 225:124]
+    node _T_157 = mux(_T_155, end_addr_pre_m, _T_156) @[lsu_lsc_ctl.scala 228:27]
-    reg _T_160 : UInt<1>, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 225:110]
+    node _T_158 = bits(io.end_addr_d, 2, 0) @[lsu_lsc_ctl.scala 228:128]
-    _T_160 <= _T_159 @[lsu_lsc_ctl.scala 225:110]
+    reg _T_159 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 228:114]
-    node int1 = neq(_T_158, _T_160) @[lsu_lsc_ctl.scala 225:72]
+    _T_159 <= _T_158 @[lsu_lsc_ctl.scala 228:114]
-    reg store_data_pre_m : UInt, io.lsu_store_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 226:72]
+    node _T_160 = cat(_T_157, _T_159) @[Cat.scala 29:58]
-    store_data_pre_m <= store_data_m_in @[lsu_lsc_ctl.scala 226:72]
+    io.end_addr_m <= _T_160 @[lsu_lsc_ctl.scala 228:17]
-    reg _T_161 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 227:62]
+    node _T_161 = bits(io.ldst_dual_r, 0, 0) @[lib.scala 8:44]
-    _T_161 <= io.lsu_addr_d @[lsu_lsc_ctl.scala 227:62]
+    node _T_162 = bits(io.lsu_addr_r, 31, 3) @[lsu_lsc_ctl.scala 229:71]
-    io.lsu_addr_m <= _T_161 @[lsu_lsc_ctl.scala 227:24]
+    node _T_163 = mux(_T_161, end_addr_pre_r, _T_162) @[lsu_lsc_ctl.scala 229:27]
-    reg _T_162 : UInt, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 228:62]
+    node _T_164 = bits(io.end_addr_m, 2, 0) @[lsu_lsc_ctl.scala 229:128]
-    _T_162 <= io.lsu_addr_m @[lsu_lsc_ctl.scala 228:62]
+    reg _T_165 : UInt, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 229:114]
-    io.lsu_addr_r <= _T_162 @[lsu_lsc_ctl.scala 228:24]
+    _T_165 <= _T_164 @[lsu_lsc_ctl.scala 229:114]
-    node _T_163 = bits(io.lsu_addr_m, 31, 3) @[lsu_lsc_ctl.scala 229:60]
+    node _T_166 = cat(_T_163, _T_165) @[Cat.scala 29:58]
-    node _T_164 = mux(int, end_addr_pre_m, _T_163) @[lsu_lsc_ctl.scala 229:27]
+    io.end_addr_r <= _T_166 @[lsu_lsc_ctl.scala 229:17]
-    node _T_165 = bits(io.end_addr_d, 2, 0) @[lsu_lsc_ctl.scala 229:117]
+    node _T_167 = bits(io.end_addr_d, 31, 3) @[lsu_lsc_ctl.scala 230:41]
-    reg _T_166 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 229:103]
+    node _T_168 = and(io.lsu_pkt_d.valid, io.ldst_dual_d) @[lsu_lsc_ctl.scala 230:69]
-    _T_166 <= _T_165 @[lsu_lsc_ctl.scala 229:103]
+    node _T_169 = or(_T_168, io.clk_override) @[lsu_lsc_ctl.scala 230:87]
-    node _T_167 = cat(_T_164, _T_166) @[Cat.scala 29:58]
+    node _T_170 = bits(_T_169, 0, 0) @[lib.scala 8:44]
-    io.end_addr_m <= _T_167 @[lsu_lsc_ctl.scala 229:17]
+    node _T_171 = bits(io.scan_mode, 0, 0) @[lib.scala 8:44]
-    node _T_168 = bits(io.lsu_addr_r, 31, 3) @[lsu_lsc_ctl.scala 230:61]
-    node _T_169 = mux(int1, end_addr_pre_r, _T_168) @[lsu_lsc_ctl.scala 230:27]
-    node _T_170 = bits(io.end_addr_m, 2, 0) @[lsu_lsc_ctl.scala 230:118]
-    reg _T_171 : UInt, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 230:104]
-    _T_171 <= _T_170 @[lsu_lsc_ctl.scala 230:104]
-    node _T_172 = cat(_T_169, _T_171) @[Cat.scala 29:58]
-    io.end_addr_r <= _T_172 @[lsu_lsc_ctl.scala 230:17]
-    node _T_173 = bits(io.end_addr_d, 31, 3) @[lsu_lsc_ctl.scala 231:41]
-    node _T_174 = and(io.lsu_pkt_d.valid, io.ldst_dual_d) @[lsu_lsc_ctl.scala 231:69]
-    node _T_175 = or(_T_174, io.clk_override) @[lsu_lsc_ctl.scala 231:87]
-    node _T_176 = bits(io.scan_mode, 0, 0) @[lib.scala 8:44]
     inst rvclkhdr_1 of rvclkhdr_1 @[lib.scala 404:23]
     rvclkhdr_1.clock <= clock
     rvclkhdr_1.reset <= reset
     rvclkhdr_1.io.clk <= clock @[lib.scala 406:18]
-    rvclkhdr_1.io.en <= _T_175 @[lib.scala 407:17]
+    rvclkhdr_1.io.en <= _T_170 @[lib.scala 407:17]
     rvclkhdr_1.io.scan_mode <= UInt<1>("h00") @[lib.scala 408:24]
-    reg _T_177 : UInt, clock with : (reset => (reset, UInt<1>("h00"))) @[Reg.scala 27:20]
+    reg _T_172 : UInt, clock with : (reset => (reset, UInt<1>("h00"))) @[Reg.scala 27:20]
-    when _T_175 : @[Reg.scala 28:19]
+    when _T_170 : @[Reg.scala 28:19]
-      _T_177 <= _T_173 @[Reg.scala 28:23]
+      _T_172 <= _T_167 @[Reg.scala 28:23]
       skip @[Reg.scala 28:19]
-    end_addr_pre_m <= _T_177 @[lsu_lsc_ctl.scala 231:18]
+    end_addr_pre_m <= _T_172 @[lsu_lsc_ctl.scala 230:18]
-    node _T_178 = bits(io.end_addr_m, 31, 3) @[lsu_lsc_ctl.scala 232:41]
+    node _T_173 = bits(io.end_addr_m, 31, 3) @[lsu_lsc_ctl.scala 231:41]
-    node _T_179 = and(io.lsu_pkt_m.valid, int) @[lsu_lsc_ctl.scala 232:69]
+    node _T_174 = and(io.lsu_pkt_m.valid, io.ldst_dual_m) @[lsu_lsc_ctl.scala 231:69]
-    node _T_180 = or(_T_179, io.clk_override) @[lsu_lsc_ctl.scala 232:76]
+    node _T_175 = or(_T_174, io.clk_override) @[lsu_lsc_ctl.scala 231:87]
-    node _T_181 = bits(io.scan_mode, 0, 0) @[lib.scala 8:44]
+    node _T_176 = bits(_T_175, 0, 0) @[lib.scala 8:44]
+    node _T_177 = bits(io.scan_mode, 0, 0) @[lib.scala 8:44]
     inst rvclkhdr_2 of rvclkhdr_2 @[lib.scala 404:23]
     rvclkhdr_2.clock <= clock
     rvclkhdr_2.reset <= reset
     rvclkhdr_2.io.clk <= clock @[lib.scala 406:18]
-    rvclkhdr_2.io.en <= _T_180 @[lib.scala 407:17]
+    rvclkhdr_2.io.en <= _T_176 @[lib.scala 407:17]
     rvclkhdr_2.io.scan_mode <= UInt<1>("h00") @[lib.scala 408:24]
-    reg _T_182 : UInt, clock with : (reset => (reset, UInt<1>("h00"))) @[Reg.scala 27:20]
+    reg _T_178 : UInt, clock with : (reset => (reset, UInt<1>("h00"))) @[Reg.scala 27:20]
-    when _T_180 : @[Reg.scala 28:19]
+    when _T_176 : @[Reg.scala 28:19]
-      _T_182 <= _T_178 @[Reg.scala 28:23]
+      _T_178 <= _T_173 @[Reg.scala 28:23]
       skip @[Reg.scala 28:19]
-    end_addr_pre_r <= _T_182 @[lsu_lsc_ctl.scala 232:18]
+    end_addr_pre_r <= _T_178 @[lsu_lsc_ctl.scala 231:18]
-    reg _T_183 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 233:62]
+    reg _T_179 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 232:62]
-    _T_183 <= io.addr_in_dccm_d @[lsu_lsc_ctl.scala 233:62]
+    _T_179 <= io.addr_in_dccm_d @[lsu_lsc_ctl.scala 232:62]
-    io.addr_in_dccm_m <= _T_183 @[lsu_lsc_ctl.scala 233:24]
+    io.addr_in_dccm_m <= _T_179 @[lsu_lsc_ctl.scala 232:24]
-    reg _T_184 : UInt, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 234:62]
+    reg _T_180 : UInt, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 233:62]
-    _T_184 <= io.addr_in_dccm_m @[lsu_lsc_ctl.scala 234:62]
+    _T_180 <= io.addr_in_dccm_m @[lsu_lsc_ctl.scala 233:62]
-    io.addr_in_dccm_r <= _T_184 @[lsu_lsc_ctl.scala 234:24]
+    io.addr_in_dccm_r <= _T_180 @[lsu_lsc_ctl.scala 233:24]
-    reg _T_185 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 235:62]
+    reg _T_181 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 234:62]
-    _T_185 <= io.addr_in_pic_d @[lsu_lsc_ctl.scala 235:62]
+    _T_181 <= io.addr_in_pic_d @[lsu_lsc_ctl.scala 234:62]
-    io.addr_in_pic_m <= _T_185 @[lsu_lsc_ctl.scala 235:24]
+    io.addr_in_pic_m <= _T_181 @[lsu_lsc_ctl.scala 234:24]
-    reg _T_186 : UInt, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 236:62]
+    reg _T_182 : UInt, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 235:62]
-    _T_186 <= io.addr_in_pic_m @[lsu_lsc_ctl.scala 236:62]
+    _T_182 <= io.addr_in_pic_m @[lsu_lsc_ctl.scala 235:62]
-    io.addr_in_pic_r <= _T_186 @[lsu_lsc_ctl.scala 236:24]
+    io.addr_in_pic_r <= _T_182 @[lsu_lsc_ctl.scala 235:24]
-    reg _T_187 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 237:62]
+    reg _T_183 : UInt, io.lsu_c1_m_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 236:62]
-    _T_187 <= addrcheck.io.addr_external_d @[lsu_lsc_ctl.scala 237:62]
+    _T_183 <= addrcheck.io.addr_external_d @[lsu_lsc_ctl.scala 236:62]
-    io.addr_external_m <= _T_187 @[lsu_lsc_ctl.scala 237:24]
+    io.addr_external_m <= _T_183 @[lsu_lsc_ctl.scala 236:24]
-    reg addr_external_r : UInt<1>, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 238:66]
+    reg addr_external_r : UInt<1>, io.lsu_c1_r_clk with : (reset => (reset, UInt<1>("h00"))) @[lsu_lsc_ctl.scala 237:66]
-    addr_external_r <= io.addr_external_m @[lsu_lsc_ctl.scala 238:66]
+    addr_external_r <= io.addr_external_m @[lsu_lsc_ctl.scala 237:66]
-    node _T_188 = or(io.addr_external_m, io.clk_override) @[lsu_lsc_ctl.scala 239:77]
+    node _T_184 = or(io.addr_external_m, io.clk_override) @[lsu_lsc_ctl.scala 238:77]
-    node _T_189 = bits(io.scan_mode, 0, 0) @[lib.scala 8:44]
+    node _T_185 = bits(io.scan_mode, 0, 0) @[lib.scala 8:44]
     inst rvclkhdr_3 of rvclkhdr_3 @[lib.scala 404:23]
     rvclkhdr_3.clock <= clock
     rvclkhdr_3.reset <= reset
     rvclkhdr_3.io.clk <= clock @[lib.scala 406:18]
-    rvclkhdr_3.io.en <= _T_188 @[lib.scala 407:17]
+    rvclkhdr_3.io.en <= _T_184 @[lib.scala 407:17]
     rvclkhdr_3.io.scan_mode <= UInt<1>("h00") @[lib.scala 408:24]
     reg bus_read_data_r : UInt, clock with : (reset => (reset, UInt<1>("h00"))) @[Reg.scala 27:20]
-    when _T_188 : @[Reg.scala 28:19]
+    when _T_184 : @[Reg.scala 28:19]
       bus_read_data_r <= io.bus_read_data_m @[Reg.scala 28:23]
       skip @[Reg.scala 28:19]
-    node _T_190 = bits(io.lsu_ld_data_corr_r, 31, 1) @[lsu_lsc_ctl.scala 242:52]
+    node _T_186 = bits(io.lsu_ld_data_corr_r, 31, 1) @[lsu_lsc_ctl.scala 241:52]
-    io.lsu_fir_addr <= _T_190 @[lsu_lsc_ctl.scala 242:28]
+    io.lsu_fir_addr <= _T_186 @[lsu_lsc_ctl.scala 241:28]
-    io.lsu_addr_d <= full_addr_d @[lsu_lsc_ctl.scala 244:28]
+    io.lsu_addr_d <= full_addr_d @[lsu_lsc_ctl.scala 243:28]
-    node _T_191 = or(io.lsu_pkt_r.bits.store, io.lsu_pkt_r.bits.load) @[lsu_lsc_ctl.scala 246:68]
+    node _T_187 = or(io.lsu_pkt_r.bits.store, io.lsu_pkt_r.bits.load) @[lsu_lsc_ctl.scala 245:68]
-    node _T_192 = and(io.lsu_pkt_r.valid, _T_191) @[lsu_lsc_ctl.scala 246:41]
+    node _T_188 = and(io.lsu_pkt_r.valid, _T_187) @[lsu_lsc_ctl.scala 245:41]
-    node _T_193 = eq(io.flush_r, UInt<1>("h00")) @[lsu_lsc_ctl.scala 246:96]
+    node _T_189 = eq(io.flush_r, UInt<1>("h00")) @[lsu_lsc_ctl.scala 245:96]
-    node _T_194 = and(_T_192, _T_193) @[lsu_lsc_ctl.scala 246:94]
+    node _T_190 = and(_T_188, _T_189) @[lsu_lsc_ctl.scala 245:94]
-    node _T_195 = eq(io.lsu_pkt_r.bits.dma, UInt<1>("h00")) @[lsu_lsc_ctl.scala 246:110]
+    node _T_191 = eq(io.lsu_pkt_r.bits.dma, UInt<1>("h00")) @[lsu_lsc_ctl.scala 245:110]
-    node _T_196 = and(_T_194, _T_195) @[lsu_lsc_ctl.scala 246:108]
+    node _T_192 = and(_T_190, _T_191) @[lsu_lsc_ctl.scala 245:108]
-    io.lsu_commit_r <= _T_196 @[lsu_lsc_ctl.scala 246:19]
+    io.lsu_commit_r <= _T_192 @[lsu_lsc_ctl.scala 245:19]
-    node _T_197 = bits(io.picm_mask_data_m, 31, 0) @[lsu_lsc_ctl.scala 247:52]
+    node _T_193 = bits(io.picm_mask_data_m, 31, 0) @[lsu_lsc_ctl.scala 246:52]
-    node _T_198 = eq(io.addr_in_pic_m, UInt<1>("h00")) @[lsu_lsc_ctl.scala 247:69]
+    node _T_194 = eq(io.addr_in_pic_m, UInt<1>("h00")) @[lsu_lsc_ctl.scala 246:69]
-    node _T_199 = bits(_T_198, 0, 0) @[Bitwise.scala 72:15]
+    node _T_195 = bits(_T_194, 0, 0) @[Bitwise.scala 72:15]
-    node _T_200 = mux(_T_199, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_196 = mux(_T_195, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_201 = or(_T_197, _T_200) @[lsu_lsc_ctl.scala 247:59]
+    node _T_197 = or(_T_193, _T_196) @[lsu_lsc_ctl.scala 246:59]
-    node _T_202 = bits(io.lsu_pkt_m.bits.store_data_bypass_m, 0, 0) @[lsu_lsc_ctl.scala 247:133]
+    node _T_198 = bits(io.lsu_pkt_m.bits.store_data_bypass_m, 0, 0) @[lsu_lsc_ctl.scala 246:133]
-    node _T_203 = mux(_T_202, io.lsu_exu.lsu_result_m, store_data_pre_m) @[lsu_lsc_ctl.scala 247:94]
+    node _T_199 = mux(_T_198, io.lsu_exu.lsu_result_m, store_data_pre_m) @[lsu_lsc_ctl.scala 246:94]
-    node _T_204 = and(_T_201, _T_203) @[lsu_lsc_ctl.scala 247:89]
+    node _T_200 = and(_T_197, _T_199) @[lsu_lsc_ctl.scala 246:89]
-    io.store_data_m <= _T_204 @[lsu_lsc_ctl.scala 247:29]
+    io.store_data_m <= _T_200 @[lsu_lsc_ctl.scala 246:29]
-    node _T_205 = mux(io.addr_external_m, io.bus_read_data_m, io.lsu_ld_data_m) @[lsu_lsc_ctl.scala 268:33]
+    node _T_201 = mux(io.addr_external_m, io.bus_read_data_m, io.lsu_ld_data_m) @[lsu_lsc_ctl.scala 267:33]
-    lsu_ld_datafn_m <= _T_205 @[lsu_lsc_ctl.scala 268:27]
+    lsu_ld_datafn_m <= _T_201 @[lsu_lsc_ctl.scala 267:27]
-    node _T_206 = eq(addr_external_r, UInt<1>("h01")) @[lsu_lsc_ctl.scala 269:49]
+    node _T_202 = eq(addr_external_r, UInt<1>("h01")) @[lsu_lsc_ctl.scala 268:49]
-    node _T_207 = mux(_T_206, bus_read_data_r, io.lsu_ld_data_corr_r) @[lsu_lsc_ctl.scala 269:33]
+    node _T_203 = mux(_T_202, bus_read_data_r, io.lsu_ld_data_corr_r) @[lsu_lsc_ctl.scala 268:33]
-    lsu_ld_datafn_corr_r <= _T_207 @[lsu_lsc_ctl.scala 269:27]
+    lsu_ld_datafn_corr_r <= _T_203 @[lsu_lsc_ctl.scala 268:27]
-    node _T_208 = and(io.lsu_pkt_m.bits.unsign, io.lsu_pkt_m.bits.by) @[lsu_lsc_ctl.scala 270:74]
+    node _T_204 = and(io.lsu_pkt_m.bits.unsign, io.lsu_pkt_m.bits.by) @[lsu_lsc_ctl.scala 269:74]
-    node _T_209 = bits(_T_208, 0, 0) @[Bitwise.scala 72:15]
+    node _T_205 = bits(_T_204, 0, 0) @[Bitwise.scala 72:15]
-    node _T_210 = mux(_T_209, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_206 = mux(_T_205, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_211 = bits(lsu_ld_datafn_m, 7, 0) @[lsu_lsc_ctl.scala 270:133]
+    node _T_207 = bits(lsu_ld_datafn_m, 7, 0) @[lsu_lsc_ctl.scala 269:133]
-    node _T_212 = cat(UInt<24>("h00"), _T_211) @[Cat.scala 29:58]
+    node _T_208 = cat(UInt<24>("h00"), _T_207) @[Cat.scala 29:58]
-    node _T_213 = and(_T_210, _T_212) @[lsu_lsc_ctl.scala 270:102]
+    node _T_209 = and(_T_206, _T_208) @[lsu_lsc_ctl.scala 269:102]
-    node _T_214 = and(io.lsu_pkt_m.bits.unsign, io.lsu_pkt_m.bits.half) @[lsu_lsc_ctl.scala 271:43]
+    node _T_210 = and(io.lsu_pkt_m.bits.unsign, io.lsu_pkt_m.bits.half) @[lsu_lsc_ctl.scala 270:43]
-    node _T_215 = bits(_T_214, 0, 0) @[Bitwise.scala 72:15]
+    node _T_211 = bits(_T_210, 0, 0) @[Bitwise.scala 72:15]
-    node _T_216 = mux(_T_215, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_212 = mux(_T_211, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_217 = bits(lsu_ld_datafn_m, 15, 0) @[lsu_lsc_ctl.scala 271:102]
+    node _T_213 = bits(lsu_ld_datafn_m, 15, 0) @[lsu_lsc_ctl.scala 270:102]
-    node _T_218 = cat(UInt<16>("h00"), _T_217) @[Cat.scala 29:58]
+    node _T_214 = cat(UInt<16>("h00"), _T_213) @[Cat.scala 29:58]
-    node _T_219 = and(_T_216, _T_218) @[lsu_lsc_ctl.scala 271:71]
+    node _T_215 = and(_T_212, _T_214) @[lsu_lsc_ctl.scala 270:71]
-    node _T_220 = or(_T_213, _T_219) @[lsu_lsc_ctl.scala 270:141]
+    node _T_216 = or(_T_209, _T_215) @[lsu_lsc_ctl.scala 269:141]
-    node _T_221 = eq(io.lsu_pkt_m.bits.unsign, UInt<1>("h00")) @[lsu_lsc_ctl.scala 272:17]
+    node _T_217 = eq(io.lsu_pkt_m.bits.unsign, UInt<1>("h00")) @[lsu_lsc_ctl.scala 271:17]
-    node _T_222 = and(_T_221, io.lsu_pkt_m.bits.by) @[lsu_lsc_ctl.scala 272:43]
+    node _T_218 = and(_T_217, io.lsu_pkt_m.bits.by) @[lsu_lsc_ctl.scala 271:43]
-    node _T_223 = bits(_T_222, 0, 0) @[Bitwise.scala 72:15]
+    node _T_219 = bits(_T_218, 0, 0) @[Bitwise.scala 72:15]
-    node _T_224 = mux(_T_223, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_220 = mux(_T_219, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_225 = bits(lsu_ld_datafn_m, 7, 7) @[lsu_lsc_ctl.scala 272:102]
+    node _T_221 = bits(lsu_ld_datafn_m, 7, 7) @[lsu_lsc_ctl.scala 271:102]
-    node _T_226 = bits(_T_225, 0, 0) @[Bitwise.scala 72:15]
+    node _T_222 = bits(_T_221, 0, 0) @[Bitwise.scala 72:15]
-    node _T_227 = mux(_T_226, UInt<24>("h0ffffff"), UInt<24>("h00")) @[Bitwise.scala 72:12]
+    node _T_223 = mux(_T_222, UInt<24>("h0ffffff"), UInt<24>("h00")) @[Bitwise.scala 72:12]
-    node _T_228 = bits(lsu_ld_datafn_m, 7, 0) @[lsu_lsc_ctl.scala 272:125]
+    node _T_224 = bits(lsu_ld_datafn_m, 7, 0) @[lsu_lsc_ctl.scala 271:125]
-    node _T_229 = cat(_T_227, _T_228) @[Cat.scala 29:58]
+    node _T_225 = cat(_T_223, _T_224) @[Cat.scala 29:58]
-    node _T_230 = and(_T_224, _T_229) @[lsu_lsc_ctl.scala 272:71]
+    node _T_226 = and(_T_220, _T_225) @[lsu_lsc_ctl.scala 271:71]
-    node _T_231 = or(_T_220, _T_230) @[lsu_lsc_ctl.scala 271:114]
+    node _T_227 = or(_T_216, _T_226) @[lsu_lsc_ctl.scala 270:114]
-    node _T_232 = eq(io.lsu_pkt_m.bits.unsign, UInt<1>("h00")) @[lsu_lsc_ctl.scala 273:17]
+    node _T_228 = eq(io.lsu_pkt_m.bits.unsign, UInt<1>("h00")) @[lsu_lsc_ctl.scala 272:17]
-    node _T_233 = and(_T_232, io.lsu_pkt_m.bits.half) @[lsu_lsc_ctl.scala 273:43]
+    node _T_229 = and(_T_228, io.lsu_pkt_m.bits.half) @[lsu_lsc_ctl.scala 272:43]
-    node _T_234 = bits(_T_233, 0, 0) @[Bitwise.scala 72:15]
+    node _T_230 = bits(_T_229, 0, 0) @[Bitwise.scala 72:15]
-    node _T_235 = mux(_T_234, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_231 = mux(_T_230, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_236 = bits(lsu_ld_datafn_m, 15, 15) @[lsu_lsc_ctl.scala 273:101]
+    node _T_232 = bits(lsu_ld_datafn_m, 15, 15) @[lsu_lsc_ctl.scala 272:101]
-    node _T_237 = bits(_T_236, 0, 0) @[Bitwise.scala 72:15]
+    node _T_233 = bits(_T_232, 0, 0) @[Bitwise.scala 72:15]
-    node _T_238 = mux(_T_237, UInt<16>("h0ffff"), UInt<16>("h00")) @[Bitwise.scala 72:12]
+    node _T_234 = mux(_T_233, UInt<16>("h0ffff"), UInt<16>("h00")) @[Bitwise.scala 72:12]
-    node _T_239 = bits(lsu_ld_datafn_m, 15, 0) @[lsu_lsc_ctl.scala 273:125]
+    node _T_235 = bits(lsu_ld_datafn_m, 15, 0) @[lsu_lsc_ctl.scala 272:125]
-    node _T_240 = cat(_T_238, _T_239) @[Cat.scala 29:58]
+    node _T_236 = cat(_T_234, _T_235) @[Cat.scala 29:58]
-    node _T_241 = and(_T_235, _T_240) @[lsu_lsc_ctl.scala 273:71]
+    node _T_237 = and(_T_231, _T_236) @[lsu_lsc_ctl.scala 272:71]
-    node _T_242 = or(_T_231, _T_241) @[lsu_lsc_ctl.scala 272:134]
+    node _T_238 = or(_T_227, _T_237) @[lsu_lsc_ctl.scala 271:134]
-    node _T_243 = bits(io.lsu_pkt_m.bits.word, 0, 0) @[Bitwise.scala 72:15]
+    node _T_239 = bits(io.lsu_pkt_m.bits.word, 0, 0) @[Bitwise.scala 72:15]
-    node _T_244 = mux(_T_243, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_240 = mux(_T_239, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_245 = bits(lsu_ld_datafn_m, 31, 0) @[lsu_lsc_ctl.scala 274:60]
+    node _T_241 = bits(lsu_ld_datafn_m, 31, 0) @[lsu_lsc_ctl.scala 273:60]
-    node _T_246 = and(_T_244, _T_245) @[lsu_lsc_ctl.scala 274:43]
+    node _T_242 = and(_T_240, _T_241) @[lsu_lsc_ctl.scala 273:43]
-    node _T_247 = or(_T_242, _T_246) @[lsu_lsc_ctl.scala 273:134]
+    node _T_243 = or(_T_238, _T_242) @[lsu_lsc_ctl.scala 272:134]
-    io.lsu_exu.lsu_result_m <= _T_247 @[lsu_lsc_ctl.scala 270:35]
+    io.lsu_exu.lsu_result_m <= _T_243 @[lsu_lsc_ctl.scala 269:35]
-    node _T_248 = and(io.lsu_pkt_r.bits.unsign, io.lsu_pkt_r.bits.by) @[lsu_lsc_ctl.scala 275:66]
+    node _T_244 = and(io.lsu_pkt_r.bits.unsign, io.lsu_pkt_r.bits.by) @[lsu_lsc_ctl.scala 274:66]
-    node _T_249 = bits(_T_248, 0, 0) @[Bitwise.scala 72:15]
+    node _T_245 = bits(_T_244, 0, 0) @[Bitwise.scala 72:15]
-    node _T_250 = mux(_T_249, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_246 = mux(_T_245, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_251 = bits(lsu_ld_datafn_corr_r, 7, 0) @[lsu_lsc_ctl.scala 275:130]
+    node _T_247 = bits(lsu_ld_datafn_corr_r, 7, 0) @[lsu_lsc_ctl.scala 274:130]
-    node _T_252 = cat(UInt<24>("h00"), _T_251) @[Cat.scala 29:58]
+    node _T_248 = cat(UInt<24>("h00"), _T_247) @[Cat.scala 29:58]
-    node _T_253 = and(_T_250, _T_252) @[lsu_lsc_ctl.scala 275:94]
+    node _T_249 = and(_T_246, _T_248) @[lsu_lsc_ctl.scala 274:94]
-    node _T_254 = and(io.lsu_pkt_r.bits.unsign, io.lsu_pkt_r.bits.half) @[lsu_lsc_ctl.scala 276:43]
+    node _T_250 = and(io.lsu_pkt_r.bits.unsign, io.lsu_pkt_r.bits.half) @[lsu_lsc_ctl.scala 275:43]
-    node _T_255 = bits(_T_254, 0, 0) @[Bitwise.scala 72:15]
+    node _T_251 = bits(_T_250, 0, 0) @[Bitwise.scala 72:15]
-    node _T_256 = mux(_T_255, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_252 = mux(_T_251, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_257 = bits(lsu_ld_datafn_corr_r, 15, 0) @[lsu_lsc_ctl.scala 276:107]
+    node _T_253 = bits(lsu_ld_datafn_corr_r, 15, 0) @[lsu_lsc_ctl.scala 275:107]
-    node _T_258 = cat(UInt<16>("h00"), _T_257) @[Cat.scala 29:58]
+    node _T_254 = cat(UInt<16>("h00"), _T_253) @[Cat.scala 29:58]
-    node _T_259 = and(_T_256, _T_258) @[lsu_lsc_ctl.scala 276:71]
+    node _T_255 = and(_T_252, _T_254) @[lsu_lsc_ctl.scala 275:71]
-    node _T_260 = or(_T_253, _T_259) @[lsu_lsc_ctl.scala 275:138]
+    node _T_256 = or(_T_249, _T_255) @[lsu_lsc_ctl.scala 274:138]
-    node _T_261 = eq(io.lsu_pkt_r.bits.unsign, UInt<1>("h00")) @[lsu_lsc_ctl.scala 277:17]
+    node _T_257 = eq(io.lsu_pkt_r.bits.unsign, UInt<1>("h00")) @[lsu_lsc_ctl.scala 276:17]
-    node _T_262 = and(_T_261, io.lsu_pkt_r.bits.by) @[lsu_lsc_ctl.scala 277:43]
+    node _T_258 = and(_T_257, io.lsu_pkt_r.bits.by) @[lsu_lsc_ctl.scala 276:43]
-    node _T_263 = bits(_T_262, 0, 0) @[Bitwise.scala 72:15]
+    node _T_259 = bits(_T_258, 0, 0) @[Bitwise.scala 72:15]
-    node _T_264 = mux(_T_263, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_260 = mux(_T_259, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_265 = bits(lsu_ld_datafn_corr_r, 7, 7) @[lsu_lsc_ctl.scala 277:107]
+    node _T_261 = bits(lsu_ld_datafn_corr_r, 7, 7) @[lsu_lsc_ctl.scala 276:107]
-    node _T_266 = bits(_T_265, 0, 0) @[Bitwise.scala 72:15]
+    node _T_262 = bits(_T_261, 0, 0) @[Bitwise.scala 72:15]
-    node _T_267 = mux(_T_266, UInt<24>("h0ffffff"), UInt<24>("h00")) @[Bitwise.scala 72:12]
+    node _T_263 = mux(_T_262, UInt<24>("h0ffffff"), UInt<24>("h00")) @[Bitwise.scala 72:12]
-    node _T_268 = bits(lsu_ld_datafn_corr_r, 7, 0) @[lsu_lsc_ctl.scala 277:135]
+    node _T_264 = bits(lsu_ld_datafn_corr_r, 7, 0) @[lsu_lsc_ctl.scala 276:135]
-    node _T_269 = cat(_T_267, _T_268) @[Cat.scala 29:58]
+    node _T_265 = cat(_T_263, _T_264) @[Cat.scala 29:58]
-    node _T_270 = and(_T_264, _T_269) @[lsu_lsc_ctl.scala 277:71]
+    node _T_266 = and(_T_260, _T_265) @[lsu_lsc_ctl.scala 276:71]
-    node _T_271 = or(_T_260, _T_270) @[lsu_lsc_ctl.scala 276:119]
+    node _T_267 = or(_T_256, _T_266) @[lsu_lsc_ctl.scala 275:119]
-    node _T_272 = eq(io.lsu_pkt_r.bits.unsign, UInt<1>("h00")) @[lsu_lsc_ctl.scala 278:17]
+    node _T_268 = eq(io.lsu_pkt_r.bits.unsign, UInt<1>("h00")) @[lsu_lsc_ctl.scala 277:17]
-    node _T_273 = and(_T_272, io.lsu_pkt_r.bits.half) @[lsu_lsc_ctl.scala 278:43]
+    node _T_269 = and(_T_268, io.lsu_pkt_r.bits.half) @[lsu_lsc_ctl.scala 277:43]
-    node _T_274 = bits(_T_273, 0, 0) @[Bitwise.scala 72:15]
+    node _T_270 = bits(_T_269, 0, 0) @[Bitwise.scala 72:15]
-    node _T_275 = mux(_T_274, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_271 = mux(_T_270, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_276 = bits(lsu_ld_datafn_corr_r, 15, 15) @[lsu_lsc_ctl.scala 278:106]
+    node _T_272 = bits(lsu_ld_datafn_corr_r, 15, 15) @[lsu_lsc_ctl.scala 277:106]
-    node _T_277 = bits(_T_276, 0, 0) @[Bitwise.scala 72:15]
+    node _T_273 = bits(_T_272, 0, 0) @[Bitwise.scala 72:15]
-    node _T_278 = mux(_T_277, UInt<16>("h0ffff"), UInt<16>("h00")) @[Bitwise.scala 72:12]
+    node _T_274 = mux(_T_273, UInt<16>("h0ffff"), UInt<16>("h00")) @[Bitwise.scala 72:12]
-    node _T_279 = bits(lsu_ld_datafn_corr_r, 15, 0) @[lsu_lsc_ctl.scala 278:135]
+    node _T_275 = bits(lsu_ld_datafn_corr_r, 15, 0) @[lsu_lsc_ctl.scala 277:135]
-    node _T_280 = cat(_T_278, _T_279) @[Cat.scala 29:58]
+    node _T_276 = cat(_T_274, _T_275) @[Cat.scala 29:58]
-    node _T_281 = and(_T_275, _T_280) @[lsu_lsc_ctl.scala 278:71]
+    node _T_277 = and(_T_271, _T_276) @[lsu_lsc_ctl.scala 277:71]
-    node _T_282 = or(_T_271, _T_281) @[lsu_lsc_ctl.scala 277:144]
+    node _T_278 = or(_T_267, _T_277) @[lsu_lsc_ctl.scala 276:144]
-    node _T_283 = bits(io.lsu_pkt_r.bits.word, 0, 0) @[Bitwise.scala 72:15]
+    node _T_279 = bits(io.lsu_pkt_r.bits.word, 0, 0) @[Bitwise.scala 72:15]
-    node _T_284 = mux(_T_283, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
+    node _T_280 = mux(_T_279, UInt<32>("h0ffffffff"), UInt<32>("h00")) @[Bitwise.scala 72:12]
-    node _T_285 = bits(lsu_ld_datafn_corr_r, 31, 0) @[lsu_lsc_ctl.scala 279:65]
+    node _T_281 = bits(lsu_ld_datafn_corr_r, 31, 0) @[lsu_lsc_ctl.scala 278:65]
-    node _T_286 = and(_T_284, _T_285) @[lsu_lsc_ctl.scala 279:43]
+    node _T_282 = and(_T_280, _T_281) @[lsu_lsc_ctl.scala 278:43]
-    node _T_287 = or(_T_282, _T_286) @[lsu_lsc_ctl.scala 278:144]
+    node _T_283 = or(_T_278, _T_282) @[lsu_lsc_ctl.scala 277:144]
-    io.lsu_result_corr_r <= _T_287 @[lsu_lsc_ctl.scala 275:27]
+    io.lsu_result_corr_r <= _T_283 @[lsu_lsc_ctl.scala 274:27]
     

lsu_lsc_ctl.v

@@ -375,10 +375,6 @@ module lsu_lsc_ctl(
   reg [31:0] _RAND_51;
   reg [31:0] _RAND_52;
   reg [31:0] _RAND_53;
-  reg [31:0] _RAND_54;
-  reg [31:0] _RAND_55;
-  reg [31:0] _RAND_56;
-  reg [31:0] _RAND_57;
 `endif // RANDOMIZE_REG_INIT
   wire  addrcheck_reset; // @[lsu_lsc_ctl.scala 117:25]
   wire  addrcheck_io_lsu_c2_m_clk; // @[lsu_lsc_ctl.scala 117:25]
@@ -526,92 +522,86 @@ module lsu_lsc_ctl(
   reg  _T_143; // @[lsu_lsc_ctl.scala 219:65]
   wire [5:0] _T_146 = {io_dma_lsc_ctl_dma_mem_addr[2:0],3'h0}; // @[Cat.scala 29:58]
   wire [63:0] dma_mem_wdata_shifted = io_dma_lsc_ctl_dma_mem_wdata >> _T_146; // @[lsu_lsc_ctl.scala 221:66]
-  reg  _T_154; // @[lsu_lsc_ctl.scala 224:48]
+  reg [31:0] store_data_pre_m; // @[lsu_lsc_ctl.scala 225:72]
-  reg  _T_156; // @[lsu_lsc_ctl.scala 224:110]
+  reg [31:0] _T_153; // @[lsu_lsc_ctl.scala 226:62]
-  wire  int_ = _T_154 != _T_156; // @[lsu_lsc_ctl.scala 224:72]
+  reg [31:0] _T_154; // @[lsu_lsc_ctl.scala 227:62]
-  reg  _T_158; // @[lsu_lsc_ctl.scala 225:48]
-  reg  _T_160; // @[lsu_lsc_ctl.scala 225:110]
-  wire  int1 = _T_158 != _T_160; // @[lsu_lsc_ctl.scala 225:72]
-  reg [31:0] store_data_pre_m; // @[lsu_lsc_ctl.scala 226:72]
-  reg [31:0] _T_161; // @[lsu_lsc_ctl.scala 227:62]
-  reg [31:0] _T_162; // @[lsu_lsc_ctl.scala 228:62]
   reg [28:0] end_addr_pre_m; // @[Reg.scala 27:20]
-  wire [28:0] _T_164 = int_ ? end_addr_pre_m : io_lsu_addr_m[31:3]; // @[lsu_lsc_ctl.scala 229:27]
+  wire [28:0] _T_157 = io_ldst_dual_m ? end_addr_pre_m : io_lsu_addr_m[31:3]; // @[lsu_lsc_ctl.scala 228:27]
-  reg [2:0] _T_166; // @[lsu_lsc_ctl.scala 229:103]
+  reg [2:0] _T_159; // @[lsu_lsc_ctl.scala 228:114]
   reg [28:0] end_addr_pre_r; // @[Reg.scala 27:20]
-  wire [28:0] _T_169 = int1 ? end_addr_pre_r : io_lsu_addr_r[31:3]; // @[lsu_lsc_ctl.scala 230:27]
+  wire [28:0] _T_163 = io_ldst_dual_r ? end_addr_pre_r : io_lsu_addr_r[31:3]; // @[lsu_lsc_ctl.scala 229:27]
-  reg [2:0] _T_171; // @[lsu_lsc_ctl.scala 230:104]
+  reg [2:0] _T_165; // @[lsu_lsc_ctl.scala 229:114]
-  wire  _T_174 = io_lsu_pkt_d_valid & io_ldst_dual_d; // @[lsu_lsc_ctl.scala 231:69]
+  wire  _T_168 = io_lsu_pkt_d_valid & io_ldst_dual_d; // @[lsu_lsc_ctl.scala 230:69]
+  wire  _T_169 = _T_168 | io_clk_override; // @[lsu_lsc_ctl.scala 230:87]
+  wire  _T_174 = io_lsu_pkt_m_valid & io_ldst_dual_m; // @[lsu_lsc_ctl.scala 231:69]
   wire  _T_175 = _T_174 | io_clk_override; // @[lsu_lsc_ctl.scala 231:87]
-  wire  _T_179 = io_lsu_pkt_m_valid & int_; // @[lsu_lsc_ctl.scala 232:69]
+  reg  _T_179; // @[lsu_lsc_ctl.scala 232:62]
-  wire  _T_180 = _T_179 | io_clk_override; // @[lsu_lsc_ctl.scala 232:76]
+  reg  _T_180; // @[lsu_lsc_ctl.scala 233:62]
-  reg  _T_183; // @[lsu_lsc_ctl.scala 233:62]
+  reg  _T_181; // @[lsu_lsc_ctl.scala 234:62]
-  reg  _T_184; // @[lsu_lsc_ctl.scala 234:62]
+  reg  _T_182; // @[lsu_lsc_ctl.scala 235:62]
-  reg  _T_185; // @[lsu_lsc_ctl.scala 235:62]
+  reg  _T_183; // @[lsu_lsc_ctl.scala 236:62]
-  reg  _T_186; // @[lsu_lsc_ctl.scala 236:62]
+  reg  addr_external_r; // @[lsu_lsc_ctl.scala 237:66]
-  reg  _T_187; // @[lsu_lsc_ctl.scala 237:62]
+  wire  _T_184 = io_addr_external_m | io_clk_override; // @[lsu_lsc_ctl.scala 238:77]
-  reg  addr_external_r; // @[lsu_lsc_ctl.scala 238:66]
-  wire  _T_188 = io_addr_external_m | io_clk_override; // @[lsu_lsc_ctl.scala 239:77]
   reg [31:0] bus_read_data_r; // @[Reg.scala 27:20]
-  wire  _T_191 = io_lsu_pkt_r_bits_store | io_lsu_pkt_r_bits_load; // @[lsu_lsc_ctl.scala 246:68]
+  wire  _T_187 = io_lsu_pkt_r_bits_store | io_lsu_pkt_r_bits_load; // @[lsu_lsc_ctl.scala 245:68]
-  wire  _T_192 = io_lsu_pkt_r_valid & _T_191; // @[lsu_lsc_ctl.scala 246:41]
+  wire  _T_188 = io_lsu_pkt_r_valid & _T_187; // @[lsu_lsc_ctl.scala 245:41]
-  wire  _T_193 = ~io_flush_r; // @[lsu_lsc_ctl.scala 246:96]
+  wire  _T_189 = ~io_flush_r; // @[lsu_lsc_ctl.scala 245:96]
-  wire  _T_194 = _T_192 & _T_193; // @[lsu_lsc_ctl.scala 246:94]
+  wire  _T_190 = _T_188 & _T_189; // @[lsu_lsc_ctl.scala 245:94]
-  wire  _T_195 = ~io_lsu_pkt_r_bits_dma; // @[lsu_lsc_ctl.scala 246:110]
+  wire  _T_191 = ~io_lsu_pkt_r_bits_dma; // @[lsu_lsc_ctl.scala 245:110]
-  wire  _T_198 = ~io_addr_in_pic_m; // @[lsu_lsc_ctl.scala 247:69]
+  wire  _T_194 = ~io_addr_in_pic_m; // @[lsu_lsc_ctl.scala 246:69]
-  wire [31:0] _T_200 = _T_198 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_196 = _T_194 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_201 = io_picm_mask_data_m | _T_200; // @[lsu_lsc_ctl.scala 247:59]
+  wire [31:0] _T_197 = io_picm_mask_data_m | _T_196; // @[lsu_lsc_ctl.scala 246:59]
-  wire [31:0] _T_203 = io_lsu_pkt_m_bits_store_data_bypass_m ? io_lsu_exu_lsu_result_m : store_data_pre_m; // @[lsu_lsc_ctl.scala 247:94]
+  wire [31:0] _T_199 = io_lsu_pkt_m_bits_store_data_bypass_m ? io_lsu_exu_lsu_result_m : store_data_pre_m; // @[lsu_lsc_ctl.scala 246:94]
-  wire [31:0] lsu_ld_datafn_m = io_addr_external_m ? io_bus_read_data_m : io_lsu_ld_data_m; // @[lsu_lsc_ctl.scala 268:33]
+  wire [31:0] lsu_ld_datafn_m = io_addr_external_m ? io_bus_read_data_m : io_lsu_ld_data_m; // @[lsu_lsc_ctl.scala 267:33]
-  wire [31:0] lsu_ld_datafn_corr_r = addr_external_r ? bus_read_data_r : io_lsu_ld_data_corr_r; // @[lsu_lsc_ctl.scala 269:33]
+  wire [31:0] lsu_ld_datafn_corr_r = addr_external_r ? bus_read_data_r : io_lsu_ld_data_corr_r; // @[lsu_lsc_ctl.scala 268:33]
-  wire  _T_208 = io_lsu_pkt_m_bits_unsign & io_lsu_pkt_m_bits_by; // @[lsu_lsc_ctl.scala 270:74]
+  wire  _T_204 = io_lsu_pkt_m_bits_unsign & io_lsu_pkt_m_bits_by; // @[lsu_lsc_ctl.scala 269:74]
-  wire [31:0] _T_210 = _T_208 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_206 = _T_204 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_212 = {24'h0,lsu_ld_datafn_m[7:0]}; // @[Cat.scala 29:58]
+  wire [31:0] _T_208 = {24'h0,lsu_ld_datafn_m[7:0]}; // @[Cat.scala 29:58]
-  wire [31:0] _T_213 = _T_210 & _T_212; // @[lsu_lsc_ctl.scala 270:102]
+  wire [31:0] _T_209 = _T_206 & _T_208; // @[lsu_lsc_ctl.scala 269:102]
-  wire  _T_214 = io_lsu_pkt_m_bits_unsign & io_lsu_pkt_m_bits_half; // @[lsu_lsc_ctl.scala 271:43]
+  wire  _T_210 = io_lsu_pkt_m_bits_unsign & io_lsu_pkt_m_bits_half; // @[lsu_lsc_ctl.scala 270:43]
-  wire [31:0] _T_216 = _T_214 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_212 = _T_210 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_218 = {16'h0,lsu_ld_datafn_m[15:0]}; // @[Cat.scala 29:58]
+  wire [31:0] _T_214 = {16'h0,lsu_ld_datafn_m[15:0]}; // @[Cat.scala 29:58]
-  wire [31:0] _T_219 = _T_216 & _T_218; // @[lsu_lsc_ctl.scala 271:71]
+  wire [31:0] _T_215 = _T_212 & _T_214; // @[lsu_lsc_ctl.scala 270:71]
-  wire [31:0] _T_220 = _T_213 | _T_219; // @[lsu_lsc_ctl.scala 270:141]
+  wire [31:0] _T_216 = _T_209 | _T_215; // @[lsu_lsc_ctl.scala 269:141]
-  wire  _T_221 = ~io_lsu_pkt_m_bits_unsign; // @[lsu_lsc_ctl.scala 272:17]
+  wire  _T_217 = ~io_lsu_pkt_m_bits_unsign; // @[lsu_lsc_ctl.scala 271:17]
-  wire  _T_222 = _T_221 & io_lsu_pkt_m_bits_by; // @[lsu_lsc_ctl.scala 272:43]
+  wire  _T_218 = _T_217 & io_lsu_pkt_m_bits_by; // @[lsu_lsc_ctl.scala 271:43]
-  wire [31:0] _T_224 = _T_222 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_220 = _T_218 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [23:0] _T_227 = lsu_ld_datafn_m[7] ? 24'hffffff : 24'h0; // @[Bitwise.scala 72:12]
+  wire [23:0] _T_223 = lsu_ld_datafn_m[7] ? 24'hffffff : 24'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_229 = {_T_227,lsu_ld_datafn_m[7:0]}; // @[Cat.scala 29:58]
+  wire [31:0] _T_225 = {_T_223,lsu_ld_datafn_m[7:0]}; // @[Cat.scala 29:58]
-  wire [31:0] _T_230 = _T_224 & _T_229; // @[lsu_lsc_ctl.scala 272:71]
+  wire [31:0] _T_226 = _T_220 & _T_225; // @[lsu_lsc_ctl.scala 271:71]
-  wire [31:0] _T_231 = _T_220 | _T_230; // @[lsu_lsc_ctl.scala 271:114]
+  wire [31:0] _T_227 = _T_216 | _T_226; // @[lsu_lsc_ctl.scala 270:114]
-  wire  _T_233 = _T_221 & io_lsu_pkt_m_bits_half; // @[lsu_lsc_ctl.scala 273:43]
+  wire  _T_229 = _T_217 & io_lsu_pkt_m_bits_half; // @[lsu_lsc_ctl.scala 272:43]
-  wire [31:0] _T_235 = _T_233 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_231 = _T_229 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [15:0] _T_238 = lsu_ld_datafn_m[15] ? 16'hffff : 16'h0; // @[Bitwise.scala 72:12]
+  wire [15:0] _T_234 = lsu_ld_datafn_m[15] ? 16'hffff : 16'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_240 = {_T_238,lsu_ld_datafn_m[15:0]}; // @[Cat.scala 29:58]
+  wire [31:0] _T_236 = {_T_234,lsu_ld_datafn_m[15:0]}; // @[Cat.scala 29:58]
-  wire [31:0] _T_241 = _T_235 & _T_240; // @[lsu_lsc_ctl.scala 273:71]
+  wire [31:0] _T_237 = _T_231 & _T_236; // @[lsu_lsc_ctl.scala 272:71]
-  wire [31:0] _T_242 = _T_231 | _T_241; // @[lsu_lsc_ctl.scala 272:134]
+  wire [31:0] _T_238 = _T_227 | _T_237; // @[lsu_lsc_ctl.scala 271:134]
-  wire [31:0] _T_244 = io_lsu_pkt_m_bits_word ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_240 = io_lsu_pkt_m_bits_word ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_246 = _T_244 & lsu_ld_datafn_m; // @[lsu_lsc_ctl.scala 274:43]
+  wire [31:0] _T_242 = _T_240 & lsu_ld_datafn_m; // @[lsu_lsc_ctl.scala 273:43]
-  wire  _T_248 = io_lsu_pkt_r_bits_unsign & io_lsu_pkt_r_bits_by; // @[lsu_lsc_ctl.scala 275:66]
+  wire  _T_244 = io_lsu_pkt_r_bits_unsign & io_lsu_pkt_r_bits_by; // @[lsu_lsc_ctl.scala 274:66]
-  wire [31:0] _T_250 = _T_248 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_246 = _T_244 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_252 = {24'h0,lsu_ld_datafn_corr_r[7:0]}; // @[Cat.scala 29:58]
+  wire [31:0] _T_248 = {24'h0,lsu_ld_datafn_corr_r[7:0]}; // @[Cat.scala 29:58]
-  wire [31:0] _T_253 = _T_250 & _T_252; // @[lsu_lsc_ctl.scala 275:94]
+  wire [31:0] _T_249 = _T_246 & _T_248; // @[lsu_lsc_ctl.scala 274:94]
-  wire  _T_254 = io_lsu_pkt_r_bits_unsign & io_lsu_pkt_r_bits_half; // @[lsu_lsc_ctl.scala 276:43]
+  wire  _T_250 = io_lsu_pkt_r_bits_unsign & io_lsu_pkt_r_bits_half; // @[lsu_lsc_ctl.scala 275:43]
-  wire [31:0] _T_256 = _T_254 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_252 = _T_250 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_258 = {16'h0,lsu_ld_datafn_corr_r[15:0]}; // @[Cat.scala 29:58]
+  wire [31:0] _T_254 = {16'h0,lsu_ld_datafn_corr_r[15:0]}; // @[Cat.scala 29:58]
-  wire [31:0] _T_259 = _T_256 & _T_258; // @[lsu_lsc_ctl.scala 276:71]
+  wire [31:0] _T_255 = _T_252 & _T_254; // @[lsu_lsc_ctl.scala 275:71]
-  wire [31:0] _T_260 = _T_253 | _T_259; // @[lsu_lsc_ctl.scala 275:138]
+  wire [31:0] _T_256 = _T_249 | _T_255; // @[lsu_lsc_ctl.scala 274:138]
-  wire  _T_261 = ~io_lsu_pkt_r_bits_unsign; // @[lsu_lsc_ctl.scala 277:17]
+  wire  _T_257 = ~io_lsu_pkt_r_bits_unsign; // @[lsu_lsc_ctl.scala 276:17]
-  wire  _T_262 = _T_261 & io_lsu_pkt_r_bits_by; // @[lsu_lsc_ctl.scala 277:43]
+  wire  _T_258 = _T_257 & io_lsu_pkt_r_bits_by; // @[lsu_lsc_ctl.scala 276:43]
-  wire [31:0] _T_264 = _T_262 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_260 = _T_258 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [23:0] _T_267 = lsu_ld_datafn_corr_r[7] ? 24'hffffff : 24'h0; // @[Bitwise.scala 72:12]
+  wire [23:0] _T_263 = lsu_ld_datafn_corr_r[7] ? 24'hffffff : 24'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_269 = {_T_267,lsu_ld_datafn_corr_r[7:0]}; // @[Cat.scala 29:58]
+  wire [31:0] _T_265 = {_T_263,lsu_ld_datafn_corr_r[7:0]}; // @[Cat.scala 29:58]
-  wire [31:0] _T_270 = _T_264 & _T_269; // @[lsu_lsc_ctl.scala 277:71]
+  wire [31:0] _T_266 = _T_260 & _T_265; // @[lsu_lsc_ctl.scala 276:71]
-  wire [31:0] _T_271 = _T_260 | _T_270; // @[lsu_lsc_ctl.scala 276:119]
+  wire [31:0] _T_267 = _T_256 | _T_266; // @[lsu_lsc_ctl.scala 275:119]
-  wire  _T_273 = _T_261 & io_lsu_pkt_r_bits_half; // @[lsu_lsc_ctl.scala 278:43]
+  wire  _T_269 = _T_257 & io_lsu_pkt_r_bits_half; // @[lsu_lsc_ctl.scala 277:43]
-  wire [31:0] _T_275 = _T_273 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_271 = _T_269 ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [15:0] _T_278 = lsu_ld_datafn_corr_r[15] ? 16'hffff : 16'h0; // @[Bitwise.scala 72:12]
+  wire [15:0] _T_274 = lsu_ld_datafn_corr_r[15] ? 16'hffff : 16'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_280 = {_T_278,lsu_ld_datafn_corr_r[15:0]}; // @[Cat.scala 29:58]
+  wire [31:0] _T_276 = {_T_274,lsu_ld_datafn_corr_r[15:0]}; // @[Cat.scala 29:58]
-  wire [31:0] _T_281 = _T_275 & _T_280; // @[lsu_lsc_ctl.scala 278:71]
+  wire [31:0] _T_277 = _T_271 & _T_276; // @[lsu_lsc_ctl.scala 277:71]
-  wire [31:0] _T_282 = _T_271 | _T_281; // @[lsu_lsc_ctl.scala 277:144]
+  wire [31:0] _T_278 = _T_267 | _T_277; // @[lsu_lsc_ctl.scala 276:144]
-  wire [31:0] _T_284 = io_lsu_pkt_r_bits_word ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
+  wire [31:0] _T_280 = io_lsu_pkt_r_bits_word ? 32'hffffffff : 32'h0; // @[Bitwise.scala 72:12]
-  wire [31:0] _T_286 = _T_284 & lsu_ld_datafn_corr_r; // @[lsu_lsc_ctl.scala 279:43]
+  wire [31:0] _T_282 = _T_280 & lsu_ld_datafn_corr_r; // @[lsu_lsc_ctl.scala 278:43]
   lsu_addrcheck addrcheck ( // @[lsu_lsc_ctl.scala 117:25]
     .reset(addrcheck_reset),
     .io_lsu_c2_m_clk(addrcheck_io_lsu_c2_m_clk),
@@ -653,18 +643,18 @@ module lsu_lsc_ctl(
     .io_clk(rvclkhdr_3_io_clk),
     .io_en(rvclkhdr_3_io_en)
   );
-  assign io_lsu_exu_lsu_result_m = _T_242 | _T_246; // @[lsu_lsc_ctl.scala 270:35]
+  assign io_lsu_exu_lsu_result_m = _T_238 | _T_242; // @[lsu_lsc_ctl.scala 269:35]
-  assign io_lsu_result_corr_r = _T_282 | _T_286; // @[lsu_lsc_ctl.scala 275:27]
+  assign io_lsu_result_corr_r = _T_278 | _T_282; // @[lsu_lsc_ctl.scala 274:27]
-  assign io_lsu_addr_d = {_T_41,_T_11[11:0]}; // @[lsu_lsc_ctl.scala 244:28]
+  assign io_lsu_addr_d = {_T_41,_T_11[11:0]}; // @[lsu_lsc_ctl.scala 243:28]
-  assign io_lsu_addr_m = _T_161; // @[lsu_lsc_ctl.scala 227:24]
+  assign io_lsu_addr_m = _T_153; // @[lsu_lsc_ctl.scala 226:24]
-  assign io_lsu_addr_r = _T_162; // @[lsu_lsc_ctl.scala 228:24]
+  assign io_lsu_addr_r = _T_154; // @[lsu_lsc_ctl.scala 227:24]
   assign io_end_addr_d = rs1_d + _T_65; // @[lsu_lsc_ctl.scala 114:24]
-  assign io_end_addr_m = {_T_164,_T_166}; // @[lsu_lsc_ctl.scala 229:17]
+  assign io_end_addr_m = {_T_157,_T_159}; // @[lsu_lsc_ctl.scala 228:17]
-  assign io_end_addr_r = {_T_169,_T_171}; // @[lsu_lsc_ctl.scala 230:17]
+  assign io_end_addr_r = {_T_163,_T_165}; // @[lsu_lsc_ctl.scala 229:17]
-  assign io_store_data_m = _T_201 & _T_203; // @[lsu_lsc_ctl.scala 247:29]
+  assign io_store_data_m = _T_197 & _T_199; // @[lsu_lsc_ctl.scala 246:29]
   assign io_lsu_exc_m = access_fault_m | misaligned_fault_m; // @[lsu_lsc_ctl.scala 154:16]
   assign io_is_sideeffects_m = addrcheck_io_is_sideeffects_m; // @[lsu_lsc_ctl.scala 127:42]
-  assign io_lsu_commit_r = _T_194 & _T_195; // @[lsu_lsc_ctl.scala 246:19]
+  assign io_lsu_commit_r = _T_190 & _T_191; // @[lsu_lsc_ctl.scala 245:19]
   assign io_lsu_single_ecc_error_incr = _T_74 & io_lsu_pkt_r_valid; // @[lsu_lsc_ctl.scala 155:32]
   assign io_lsu_error_pkt_r_valid = _T_112; // @[lsu_lsc_ctl.scala 184:24 lsu_lsc_ctl.scala 186:30]
   assign io_lsu_error_pkt_r_bits_single_ecc_error = _T_111; // @[lsu_lsc_ctl.scala 184:24 lsu_lsc_ctl.scala 185:46]
@@ -672,15 +662,15 @@ module lsu_lsc_ctl(
   assign io_lsu_error_pkt_r_bits_exc_type = _T_110_bits_exc_type; // @[lsu_lsc_ctl.scala 184:24]
   assign io_lsu_error_pkt_r_bits_mscause = _T_110_bits_mscause; // @[lsu_lsc_ctl.scala 184:24]
   assign io_lsu_error_pkt_r_bits_addr = _T_110_bits_addr; // @[lsu_lsc_ctl.scala 184:24]
-  assign io_lsu_fir_addr = io_lsu_ld_data_corr_r[31:1]; // @[lsu_lsc_ctl.scala 242:28]
+  assign io_lsu_fir_addr = io_lsu_ld_data_corr_r[31:1]; // @[lsu_lsc_ctl.scala 241:28]
   assign io_lsu_fir_error = _T_113; // @[lsu_lsc_ctl.scala 187:38]
   assign io_addr_in_dccm_d = addrcheck_io_addr_in_dccm_d; // @[lsu_lsc_ctl.scala 128:42]
-  assign io_addr_in_dccm_m = _T_183; // @[lsu_lsc_ctl.scala 233:24]
+  assign io_addr_in_dccm_m = _T_179; // @[lsu_lsc_ctl.scala 232:24]
-  assign io_addr_in_dccm_r = _T_184; // @[lsu_lsc_ctl.scala 234:24]
+  assign io_addr_in_dccm_r = _T_180; // @[lsu_lsc_ctl.scala 233:24]
   assign io_addr_in_pic_d = addrcheck_io_addr_in_pic_d; // @[lsu_lsc_ctl.scala 129:42]
-  assign io_addr_in_pic_m = _T_185; // @[lsu_lsc_ctl.scala 235:24]
+  assign io_addr_in_pic_m = _T_181; // @[lsu_lsc_ctl.scala 234:24]
-  assign io_addr_in_pic_r = _T_186; // @[lsu_lsc_ctl.scala 236:24]
+  assign io_addr_in_pic_r = _T_182; // @[lsu_lsc_ctl.scala 235:24]
-  assign io_addr_external_m = _T_187; // @[lsu_lsc_ctl.scala 237:24]
+  assign io_addr_external_m = _T_183; // @[lsu_lsc_ctl.scala 236:24]
   assign io_lsu_pkt_d_valid = _T_128 | io_dma_lsc_ctl_dma_dccm_req; // @[lsu_lsc_ctl.scala 208:20 lsu_lsc_ctl.scala 212:24]
   assign io_lsu_pkt_d_bits_fast_int = io_dec_lsu_valid_raw_d & io_lsu_p_bits_fast_int; // @[lsu_lsc_ctl.scala 208:20]
   assign io_lsu_pkt_d_bits_stack = io_dec_lsu_valid_raw_d & io_lsu_p_bits_stack; // @[lsu_lsc_ctl.scala 208:20]
@@ -740,9 +730,9 @@ module lsu_lsc_ctl(
   assign rvclkhdr_io_clk = clock; // @[lib.scala 419:18]
   assign rvclkhdr_io_en = _T_105 | io_clk_override; // @[lib.scala 420:17]
   assign rvclkhdr_1_io_clk = clock; // @[lib.scala 406:18]
-  assign rvclkhdr_1_io_en = _T_174 | io_clk_override; // @[lib.scala 407:17]
+  assign rvclkhdr_1_io_en = _T_168 | io_clk_override; // @[lib.scala 407:17]
   assign rvclkhdr_2_io_clk = clock; // @[lib.scala 406:18]
-  assign rvclkhdr_2_io_en = _T_179 | io_clk_override; // @[lib.scala 407:17]
+  assign rvclkhdr_2_io_en = _T_174 | io_clk_override; // @[lib.scala 407:17]
   assign rvclkhdr_3_io_clk = clock; // @[lib.scala 406:18]
   assign rvclkhdr_3_io_en = io_addr_external_m | io_clk_override; // @[lib.scala 407:17]
 `ifdef RANDOMIZE_GARBAGE_ASSIGN
@@ -861,41 +851,33 @@ initial begin
   _RAND_39 = {1{`RANDOM}};
   _T_143 = _RAND_39[0:0];
   _RAND_40 = {1{`RANDOM}};
-  _T_154 = _RAND_40[0:0];
+  store_data_pre_m = _RAND_40[31:0];
   _RAND_41 = {1{`RANDOM}};
-  _T_156 = _RAND_41[0:0];
+  _T_153 = _RAND_41[31:0];
   _RAND_42 = {1{`RANDOM}};
-  _T_158 = _RAND_42[0:0];
+  _T_154 = _RAND_42[31:0];
   _RAND_43 = {1{`RANDOM}};
-  _T_160 = _RAND_43[0:0];
+  end_addr_pre_m = _RAND_43[28:0];
   _RAND_44 = {1{`RANDOM}};
-  store_data_pre_m = _RAND_44[31:0];
+  _T_159 = _RAND_44[2:0];
   _RAND_45 = {1{`RANDOM}};
-  _T_161 = _RAND_45[31:0];
+  end_addr_pre_r = _RAND_45[28:0];
   _RAND_46 = {1{`RANDOM}};
-  _T_162 = _RAND_46[31:0];
+  _T_165 = _RAND_46[2:0];
   _RAND_47 = {1{`RANDOM}};
-  end_addr_pre_m = _RAND_47[28:0];
+  _T_179 = _RAND_47[0:0];
   _RAND_48 = {1{`RANDOM}};
-  _T_166 = _RAND_48[2:0];
+  _T_180 = _RAND_48[0:0];
   _RAND_49 = {1{`RANDOM}};
-  end_addr_pre_r = _RAND_49[28:0];
+  _T_181 = _RAND_49[0:0];
   _RAND_50 = {1{`RANDOM}};
-  _T_171 = _RAND_50[2:0];
+  _T_182 = _RAND_50[0:0];
   _RAND_51 = {1{`RANDOM}};
   _T_183 = _RAND_51[0:0];
   _RAND_52 = {1{`RANDOM}};
-  _T_184 = _RAND_52[0:0];
+  addr_external_r = _RAND_52[0:0];
   _RAND_53 = {1{`RANDOM}};
-  _T_185 = _RAND_53[0:0];
+  bus_read_data_r = _RAND_53[31:0];
-  _RAND_54 = {1{`RANDOM}};
-  _T_186 = _RAND_54[0:0];
-  _RAND_55 = {1{`RANDOM}};
-  _T_187 = _RAND_55[0:0];
-  _RAND_56 = {1{`RANDOM}};
-  addr_external_r = _RAND_56[0:0];
-  _RAND_57 = {1{`RANDOM}};
-  bus_read_data_r = _RAND_57[31:0];
 `endif // RANDOMIZE_REG_INIT
   if (reset) begin
     access_fault_m = 1'h0;
@@ -1017,54 +999,42 @@ initial begin
   if (reset) begin
     _T_143 = 1'h0;
   end
-  if (reset) begin
-    _T_154 = 1'h0;
-  end
-  if (reset) begin
-    _T_156 = 1'h0;
-  end
-  if (reset) begin
-    _T_158 = 1'h0;
-  end
-  if (reset) begin
-    _T_160 = 1'h0;
-  end
   if (reset) begin
     store_data_pre_m = 32'h0;
   end
   if (reset) begin
-    _T_161 = 32'h0;
+    _T_153 = 32'h0;
   end
   if (reset) begin
-    _T_162 = 32'h0;
+    _T_154 = 32'h0;
   end
   if (reset) begin
     end_addr_pre_m = 29'h0;
   end
   if (reset) begin
-    _T_166 = 3'h0;
+    _T_159 = 3'h0;
   end
   if (reset) begin
     end_addr_pre_r = 29'h0;
   end
   if (reset) begin
-    _T_171 = 3'h0;
+    _T_165 = 3'h0;
+  end
+  if (reset) begin
+    _T_179 = 1'h0;
+  end
+  if (reset) begin
+    _T_180 = 1'h0;
+  end
+  if (reset) begin
+    _T_181 = 1'h0;
+  end
+  if (reset) begin
+    _T_182 = 1'h0;
   end
   if (reset) begin
     _T_183 = 1'h0;
   end
-  if (reset) begin
-    _T_184 = 1'h0;
-  end
-  if (reset) begin
-    _T_185 = 1'h0;
-  end
-  if (reset) begin
-    _T_186 = 1'h0;
-  end
-  if (reset) begin
-    _T_187 = 1'h0;
-  end
   if (reset) begin
     addr_external_r = 1'h0;
   end
@@ -1367,34 +1337,6 @@ end // initial
       _T_143 <= io_lsu_pkt_m_valid & _T_136;
     end
   end
-  always @(posedge io_lsu_c1_m_clk or posedge reset) begin
-    if (reset) begin
-      _T_154 <= 1'h0;
-    end else begin
-      _T_154 <= io_lsu_addr_d[2];
-    end
-  end
-  always @(posedge io_lsu_c1_m_clk or posedge reset) begin
-    if (reset) begin
-      _T_156 <= 1'h0;
-    end else begin
-      _T_156 <= io_end_addr_d[2];
-    end
-  end
-  always @(posedge io_lsu_c1_r_clk or posedge reset) begin
-    if (reset) begin
-      _T_158 <= 1'h0;
-    end else begin
-      _T_158 <= io_lsu_addr_m[2];
-    end
-  end
-  always @(posedge io_lsu_c1_r_clk or posedge reset) begin
-    if (reset) begin
-      _T_160 <= 1'h0;
-    end else begin
-      _T_160 <= io_end_addr_m[2];
-    end
-  end
   always @(posedge io_lsu_store_c1_m_clk or posedge reset) begin
     if (reset) begin
       store_data_pre_m <= 32'h0;
@@ -1408,79 +1350,79 @@ end // initial
   end
   always @(posedge io_lsu_c1_m_clk or posedge reset) begin
     if (reset) begin
-      _T_161 <= 32'h0;
+      _T_153 <= 32'h0;
     end else begin
-      _T_161 <= io_lsu_addr_d;
+      _T_153 <= io_lsu_addr_d;
     end
   end
   always @(posedge io_lsu_c1_r_clk or posedge reset) begin
     if (reset) begin
-      _T_162 <= 32'h0;
+      _T_154 <= 32'h0;
     end else begin
-      _T_162 <= io_lsu_addr_m;
+      _T_154 <= io_lsu_addr_m;
     end
   end
   always @(posedge clock or posedge reset) begin
     if (reset) begin
       end_addr_pre_m <= 29'h0;
-    end else if (_T_175) begin
+    end else if (_T_169) begin
       end_addr_pre_m <= io_end_addr_d[31:3];
     end
   end
   always @(posedge io_lsu_c1_m_clk or posedge reset) begin
     if (reset) begin
-      _T_166 <= 3'h0;
+      _T_159 <= 3'h0;
     end else begin
-      _T_166 <= io_end_addr_d[2:0];
+      _T_159 <= io_end_addr_d[2:0];
     end
   end
   always @(posedge clock or posedge reset) begin
     if (reset) begin
       end_addr_pre_r <= 29'h0;
-    end else if (_T_180) begin
+    end else if (_T_175) begin
       end_addr_pre_r <= io_end_addr_m[31:3];
     end
   end
   always @(posedge io_lsu_c1_r_clk or posedge reset) begin
     if (reset) begin
-      _T_171 <= 3'h0;
+      _T_165 <= 3'h0;
     end else begin
-      _T_171 <= io_end_addr_m[2:0];
+      _T_165 <= io_end_addr_m[2:0];
+    end
+  end
+  always @(posedge io_lsu_c1_m_clk or posedge reset) begin
+    if (reset) begin
+      _T_179 <= 1'h0;
+    end else begin
+      _T_179 <= io_addr_in_dccm_d;
+    end
+  end
+  always @(posedge io_lsu_c1_r_clk or posedge reset) begin
+    if (reset) begin
+      _T_180 <= 1'h0;
+    end else begin
+      _T_180 <= io_addr_in_dccm_m;
+    end
+  end
+  always @(posedge io_lsu_c1_m_clk or posedge reset) begin
+    if (reset) begin
+      _T_181 <= 1'h0;
+    end else begin
+      _T_181 <= io_addr_in_pic_d;
+    end
+  end
+  always @(posedge io_lsu_c1_r_clk or posedge reset) begin
+    if (reset) begin
+      _T_182 <= 1'h0;
+    end else begin
+      _T_182 <= io_addr_in_pic_m;
     end
   end
   always @(posedge io_lsu_c1_m_clk or posedge reset) begin
     if (reset) begin
       _T_183 <= 1'h0;
     end else begin
-      _T_183 <= io_addr_in_dccm_d;
+      _T_183 <= addrcheck_io_addr_external_d;
-    end
-  end
-  always @(posedge io_lsu_c1_r_clk or posedge reset) begin
-    if (reset) begin
-      _T_184 <= 1'h0;
-    end else begin
-      _T_184 <= io_addr_in_dccm_m;
-    end
-  end
-  always @(posedge io_lsu_c1_m_clk or posedge reset) begin
-    if (reset) begin
-      _T_185 <= 1'h0;
-    end else begin
-      _T_185 <= io_addr_in_pic_d;
-    end
-  end
-  always @(posedge io_lsu_c1_r_clk or posedge reset) begin
-    if (reset) begin
-      _T_186 <= 1'h0;
-    end else begin
-      _T_186 <= io_addr_in_pic_m;
-    end
-  end
-  always @(posedge io_lsu_c1_m_clk or posedge reset) begin
-    if (reset) begin
-      _T_187 <= 1'h0;
-    end else begin
-      _T_187 <= addrcheck_io_addr_external_d;
     end
   end
   always @(posedge io_lsu_c1_r_clk or posedge reset) begin
@@ -1493,7 +1435,7 @@ end // initial
   always @(posedge clock or posedge reset) begin
     if (reset) begin
       bus_read_data_r <= 32'h0;
-    end else if (_T_188) begin
+    end else if (_T_184) begin
       bus_read_data_r <= io_bus_read_data_m;
     end
   end

src/main/scala/lsu/lsu.scala

@@ -1,7 +1,7 @@
 package lsu
 
 import lib._
-import chisel3._
+import chisel3.{withClock, _}
 import chisel3.util._
 import include._
 import mem._
@@ -65,9 +65,9 @@ class lsu extends Module with RequireAsyncReset with param with lib {
 //  val end_addr_m                   = WireInit(0.U(32.W))
 //  val end_addr_r                   = WireInit(0.U(32.W))
   val lsu_busreq_r                 = WireInit(Bool(),false.B)
-//  val ldst_dual_d = WireInit(Bool(),false.B)
+  val ldst_dual_d = WireInit(Bool(),false.B)
-//  val ldst_dual_m = WireInit(Bool(),false.B)
+  val ldst_dual_m = WireInit(Bool(),false.B)
-//  val ldst_dual_r = WireInit(Bool(),false.B)
+  val ldst_dual_r = WireInit(Bool(),false.B)
 
   val lsu_lsc_ctl    = Module(new lsu_lsc_ctl())
  // io.lsu_exu.lsu_result_m := lsu_lsc_ctl.io.lsu_result_m
@@ -138,9 +138,9 @@ class lsu extends Module with RequireAsyncReset with param with lib {
   lsu_lsc_ctl.io.lsu_double_ecc_error_m      := ecc.io.lsu_double_ecc_error_m
   lsu_lsc_ctl.io.flush_m_up                  := flush_m_up
   lsu_lsc_ctl.io.flush_r                     := flush_r
-  lsu_lsc_ctl.io.ldst_dual_d                 := lsu_lsc_ctl.io.lsu_addr_d(2) =/= lsu_lsc_ctl.io.end_addr_d(2)
+  lsu_lsc_ctl.io.ldst_dual_d                 := ldst_dual_d
-  lsu_lsc_ctl.io.ldst_dual_m                 := lsu_lsc_ctl.io.lsu_addr_m(2) =/= lsu_lsc_ctl.io.end_addr_m(2)
+  lsu_lsc_ctl.io.ldst_dual_m                 := ldst_dual_m
-  lsu_lsc_ctl.io.ldst_dual_r                 := lsu_lsc_ctl.io.lsu_addr_r(2) =/= lsu_lsc_ctl.io.end_addr_r(2)
+  lsu_lsc_ctl.io.ldst_dual_r                 := ldst_dual_r
   lsu_lsc_ctl.io.lsu_exu                     <> io.lsu_exu
   lsu_lsc_ctl.io.lsu_p                       <> io.lsu_p
   lsu_lsc_ctl.io.dec_lsu_valid_raw_d         := io.dec_lsu_valid_raw_d
@@ -153,9 +153,9 @@ class lsu extends Module with RequireAsyncReset with param with lib {
   //Outputs
 
 
-//  ldst_dual_d  := lsu_lsc_ctl.io.lsu_addr_d(2) =/= lsu_lsc_ctl.io.end_addr_d(2)
+  ldst_dual_d  := lsu_lsc_ctl.io.lsu_addr_d(2) =/= lsu_lsc_ctl.io.end_addr_d(2)
-//  ldst_dual_m  := lsu_lsc_ctl.io.lsu_addr_m(2) =/= lsu_lsc_ctl.io.end_addr_m(2)
+  ldst_dual_m  := lsu_lsc_ctl.io.lsu_addr_m(2) =/= withClock(clkdomain.io.lsu_c1_m_clk){RegNext(lsu_lsc_ctl.io.end_addr_d(2),0.U)}//=/= lsu_lsc_ctl.io.end_addr_m(2)
-//  ldst_dual_r  := lsu_lsc_ctl.io.lsu_addr_r(2) =/= lsu_lsc_ctl.io.end_addr_r(2)
+  ldst_dual_r  := lsu_lsc_ctl.io.lsu_addr_r(2) =/= withClock(clkdomain.io.lsu_c1_r_clk){RegNext(lsu_lsc_ctl.io.end_addr_m(2),0.U)}//=/= lsu_lsc_ctl.io.end_addr_r(2)
 
   io.lsu_single_ecc_error_incr                  := lsu_lsc_ctl.io.lsu_single_ecc_error_incr
   io.lsu_error_pkt_r                            <> lsu_lsc_ctl.io.lsu_error_pkt_r
@@ -164,8 +164,8 @@ class lsu extends Module with RequireAsyncReset with param with lib {
   // DCCM Control
   //Inputs
   dccm_ctl.io.clk_override                   := io.clk_override
-  dccm_ctl.io.ldst_dual_m                    := lsu_lsc_ctl.io.lsu_addr_m(2) =/= lsu_lsc_ctl.io.end_addr_m(2)
+  dccm_ctl.io.ldst_dual_m                    := ldst_dual_m
-  dccm_ctl.io.ldst_dual_r                    := lsu_lsc_ctl.io.lsu_addr_r(2) =/= lsu_lsc_ctl.io.end_addr_r(2)
+  dccm_ctl.io.ldst_dual_r                    := ldst_dual_r
   dccm_ctl.io.lsu_c2_m_clk                   := clkdomain.io.lsu_c2_m_clk
   dccm_ctl.io.lsu_c2_r_clk                   := clkdomain.io.lsu_c2_r_clk
   dccm_ctl.io.lsu_free_c2_clk                := clkdomain.io.lsu_free_c2_clk
@@ -224,9 +224,9 @@ class lsu extends Module with RequireAsyncReset with param with lib {
   io.lsu_pic <> dccm_ctl.io.lsu_pic
   //Store Buffer
   //Inputs
-  stbuf.io.ldst_dual_d                           := lsu_lsc_ctl.io.lsu_addr_d(2) =/= lsu_lsc_ctl.io.end_addr_d(2)
+  stbuf.io.ldst_dual_d                           := ldst_dual_d
-  stbuf.io.ldst_dual_m                           := lsu_lsc_ctl.io.lsu_addr_m(2) =/= lsu_lsc_ctl.io.end_addr_m(2)
+  stbuf.io.ldst_dual_m                           := ldst_dual_m
-  stbuf.io.ldst_dual_r                           := lsu_lsc_ctl.io.lsu_addr_r(2) =/= lsu_lsc_ctl.io.end_addr_r(2)
+  stbuf.io.ldst_dual_r                           := ldst_dual_r
   stbuf.io.lsu_stbuf_c1_clk       	             := clkdomain.io.lsu_stbuf_c1_clk
   stbuf.io.lsu_free_c2_clk        	             := clkdomain.io.lsu_free_c2_clk
   stbuf.io.lsu_pkt_m           	                 <> lsu_lsc_ctl.io.lsu_pkt_m
@@ -327,9 +327,9 @@ class lsu extends Module with RequireAsyncReset with param with lib {
   bus_intf.io.lsu_busm_clk                      := clkdomain.io.lsu_busm_clk
   bus_intf.io.dec_lsu_valid_raw_d               := io.dec_lsu_valid_raw_d
   bus_intf.io.lsu_busreq_m                      := lsu_busreq_m
-  bus_intf.io.ldst_dual_d                       := lsu_lsc_ctl.io.lsu_addr_d(2) =/= lsu_lsc_ctl.io.end_addr_d(2)
+  bus_intf.io.ldst_dual_d                       := ldst_dual_d
-  bus_intf.io.ldst_dual_m                       := lsu_lsc_ctl.io.lsu_addr_m(2) =/= lsu_lsc_ctl.io.end_addr_m(2)
+  bus_intf.io.ldst_dual_m                       := ldst_dual_m
-  bus_intf.io.ldst_dual_r                       := lsu_lsc_ctl.io.lsu_addr_r(2) =/= lsu_lsc_ctl.io.end_addr_r(2)
+  bus_intf.io.ldst_dual_r                       := ldst_dual_r
   bus_intf.io.lsu_addr_m                        := lsu_lsc_ctl.io.lsu_addr_m & Fill(32,lsu_lsc_ctl.io.addr_external_m & lsu_lsc_ctl.io.lsu_pkt_m.valid)
   bus_intf.io.lsu_addr_r                        := lsu_lsc_ctl.io.lsu_addr_r & Fill(32,lsu_busreq_r)
   bus_intf.io.end_addr_m                        := lsu_lsc_ctl.io.end_addr_m & Fill(32,lsu_lsc_ctl.io.addr_external_m & lsu_lsc_ctl.io.lsu_pkt_m.valid)

src/main/scala/lsu/lsu_bus_buffer.scala

@@ -166,13 +166,15 @@ class lsu_bus_buffer extends Module with RequireAsyncReset with lib {
   val fwd_data = WireInit(UInt(32.W), 0.U)
   val ld_fwddata_buf_lo_initial = (0 until 4).map(i=>Fill(8, ld_byte_ibuf_hit_lo(i))).reverse.reduce(Cat(_,_))
   val ld_fwddata_buf_hi_initial = (0 until 4).map(i=>Fill(8, ld_byte_ibuf_hit_hi(i))).reverse.reduce(Cat(_,_))
-  io.ld_fwddata_buf_lo := Cat((0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_lo(3)(i)) & buf_data(i)(31, 24)).reduce(_ | _),
+  io.ld_fwddata_buf_lo := Cat(
+    (0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_lo(3)(i)) & buf_data(i)(31, 24)).reduce(_ | _),
     (0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_lo(2)(i)) & buf_data(i)(23, 16)).reduce(_ | _),
     (0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_lo(1)(i)) & buf_data(i)(15,  8)).reduce(_ | _),
     (0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_lo(0)(i)) & buf_data(i)(7 ,  0)).reduce(_ | _)) |
     (ld_fwddata_buf_lo_initial & ibuf_data)
 
-  io.ld_fwddata_buf_hi := Cat((0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_hi(3)(i)) & buf_data(i)(31, 24)).reduce(_ | _),
+  io.ld_fwddata_buf_hi := Cat(
+    (0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_hi(3)(i)) & buf_data(i)(31, 24)).reduce(_ | _),
     (0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_hi(2)(i)) & buf_data(i)(23, 16)).reduce(_ | _),
     (0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_hi(1)(i)) & buf_data(i)(15,  8)).reduce(_ | _),
     (0 until DEPTH).map(i => Fill(8, ld_byte_hitvecfn_hi(0)(i)) & buf_data(i)(7 ,  0)).reduce(_ | _)) |

src/main/scala/lsu/lsu_lsc_ctl.scala

@@ -26,9 +26,9 @@ class  lsu_lsc_ctl extends Module with RequireAsyncReset with lib
 
     val flush_m_up = Input(UInt(1.W))
     val flush_r    = Input(UInt(1.W))
-    val ldst_dual_d    = Input(Bool())
+    val ldst_dual_d    = Input(UInt(1.W))
-    val ldst_dual_m    = Input(Bool())
+    val ldst_dual_m    = Input(UInt(1.W))
-    val ldst_dual_r    = Input(Bool())
+    val ldst_dual_r    = Input(UInt(1.W))
 
     val lsu_exu = new lsu_exu()
 
@@ -221,15 +221,14 @@ class  lsu_lsc_ctl extends Module with RequireAsyncReset with lib
   val dma_mem_wdata_shifted = io.dma_lsc_ctl.dma_mem_wdata(63,0) >> Cat(io.dma_lsc_ctl.dma_mem_addr(2,0), 0.U(3.W))   // Shift the dma data to lower bits to make it consistent to lsu stores
   val store_data_d          = Mux(io.dma_lsc_ctl.dma_dccm_req.asBool,dma_mem_wdata_shifted(31,0),io.lsu_exu.exu_lsu_rs2_d(31,0))  // Write to PIC still happens in r stage
   val store_data_m_in       = Mux(io.lsu_pkt_d.bits.store_data_bypass_d.asBool,io.lsu_exu.lsu_result_m(31,0),store_data_d(31,0))
-  val int  = withClock(io.lsu_c1_m_clk){RegNext(io.lsu_addr_d(2),0.U)} =/= withClock(io.lsu_c1_m_clk){RegNext(io.end_addr_d(2),0.U)}
+
-  val int1 = withClock(io.lsu_c1_r_clk){RegNext(io.lsu_addr_m(2),0.U)} =/= withClock(io.lsu_c1_r_clk){RegNext(io.end_addr_m(2),0.U)}
   val store_data_pre_m      =  withClock(io.lsu_store_c1_m_clk){RegNext(store_data_m_in,0.U)}
   io.lsu_addr_m        :=  withClock(io.lsu_c1_m_clk){RegNext(io.lsu_addr_d,0.U)}
   io.lsu_addr_r        :=  withClock(io.lsu_c1_r_clk){RegNext(io.lsu_addr_m,0.U)}
-  io.end_addr_m := Cat(Mux(int,end_addr_pre_m,io.lsu_addr_m(31,3)), withClock(io.lsu_c1_m_clk){RegNext(io.end_addr_d(2,0),0.U)})
+  io.end_addr_m := Cat(Mux(io.ldst_dual_m,end_addr_pre_m,io.lsu_addr_m(31,3)), withClock(io.lsu_c1_m_clk){RegNext(io.end_addr_d(2,0),0.U)})
-  io.end_addr_r := Cat(Mux(int1,end_addr_pre_r,io.lsu_addr_r(31,3)), withClock(io.lsu_c1_r_clk){RegNext(io.end_addr_m(2,0),0.U)})
+  io.end_addr_r := Cat(Mux(io.ldst_dual_r,end_addr_pre_r,io.lsu_addr_r(31,3)), withClock(io.lsu_c1_r_clk){RegNext(io.end_addr_m(2,0),0.U)})
   end_addr_pre_m := rvdffe(io.end_addr_d(31,3),((io.lsu_pkt_d.valid & io.ldst_dual_d) | io.clk_override),clock,io.scan_mode)
-  end_addr_pre_r := rvdffe(io.end_addr_m(31,3),((io.lsu_pkt_m.valid & int) | io.clk_override),clock,io.scan_mode)
+  end_addr_pre_r := rvdffe(io.end_addr_m(31,3),((io.lsu_pkt_m.valid & io.ldst_dual_m) | io.clk_override),clock,io.scan_mode)
   io.addr_in_dccm_m    :=  withClock(io.lsu_c1_m_clk){RegNext(io.addr_in_dccm_d,0.U)}
   io.addr_in_dccm_r    :=  withClock(io.lsu_c1_r_clk){RegNext(io.addr_in_dccm_m,0.U)}
   io.addr_in_pic_m     :=  withClock(io.lsu_c1_m_clk){RegNext(io.addr_in_pic_d,0.U)}