changes for verilator debug

dev/sim/verilator/tb_litex.cpp

@@ -3,18 +3,30 @@
 
 #define TRACING
 
+// old public access method
+//#define OLD_PUBLIC
+
 #include <cstddef>
 #include <iostream>
 #include <fstream>
 #include <iomanip>
+#include <unordered_map>
 
 #include "verilated.h"
 #include "Va2owb.h"
 
+#ifndef OLD_PUBLIC
 // internal nets
 #include "Va2owb___024root.h"
+#endif
+
 #include "Va2owb_a2owb.h"
 #include "Va2owb_a2l2wb.h"
+#include "Va2owb_c.h"
+#include "Va2owb_iuq.h"
+#include "Va2owb_iuq_cpl_top.h"
+#include "Va2owb_iuq_cpl.h"
+#include "Va2owb_iuq_cpl_ctrl.h"
 
 #ifdef TRACING
 #include "verilated_vcd_c.h"
@@ -28,9 +40,13 @@ unsigned int t = 0;
 */
 
 Va2owb* m;
+#ifdef OLD_PUBLIC
+Va2owb* root;
+#else
+Va2owb___024root* root;
+#endif
 
 vluint64_t main_time = 0;     // in units of timeprecision used in verilog or --timescale-override
-// what is it?  it changed to 941621251 after calling loadmem()
 
 double sc_time_stamp() {      // $time in verilog
    return main_time;
@@ -38,10 +54,12 @@ double sc_time_stamp() {      // $time in verilog
 
 const int resetCycle = 10;
 const int threadRunCycle = resetCycle + 5;
-const int runCycles = 500;
+const int runCycles = 501;
 const int hbCycles = 500;
 const int threads = 1;
-const std::string testFile = "../mem/test1/rom.init";
+const std::string testFile = "../mem/test3/rom.init";
+const unsigned int bootAdr = 0x00000000;
+
 
 // Cythonize this and use it for cocotb too...
 
@@ -139,16 +157,23 @@ int main(int argc, char **argv) {
 
    Verilated::commandArgs(argc, argv);
    m = new Va2owb;
+#ifdef OLD_PUBLIC
+   root = m;
+#else
+   root = m->rootp;
+#endif
 
 #ifdef TRACING
       Verilated::traceEverOn(true);
       t = new VerilatedVcdC;
       m->trace(t, 99);
-      t->open("a2onode.vcd");
+      t->open("a2olitex.vcd");
       cout << "Tracing enabled." << endl;
 #endif
 
+   bool ok = true;
    bool resetDone = false;
+   bool booted = false;
    unsigned int threadStop = 0x3;
 
    unsigned int tick = 0;
@@ -160,20 +185,107 @@ int main(int argc, char **argv) {
    unsigned int countReads = 0;
    bool wbRdPending = false, wbWrPending = false;
 
-   //unsigned int iu0Comp = m->rootp->a2owb->c0->iu_lq_i0_completed;
+   unsigned int iu0Comp, iu1Comp;
-   //unsigned int iu0Comp = m->rootp->a2owb__DOT__c0__DOT__lq0__DOT__lsq__DOT__odq__DOT__iu_lq_i0_completed_itag_int;
+   unsigned long iu0CompIFAR, iu1CompIFAR, cp3NIA;
+/*
+   creditsLdErr = sim.a2o.root.lq0.lsq.arb.ld_cred_err_q
+   creditsStErr = sim.a2o.root.lq0.lsq.arb.st_cred_err_q
+*/
+
+
    /*
    iu0CompIFAR = sim.a2o.root.iuq0.iuq_cpl_top0.iuq_cpl0.cp2_i0_ifar
-   iu1Comp = sim.a2o.root.iu_lq_i1_completed
+   iu1Comp =                                             cp2_i0_completed
    iu1CompIFAR = sim.a2o.root.iuq0.iuq_cpl_top0.iuq_cpl0.cp2_i1_ifar
    iuCompFlushIFAR = sim.a2o.root.cp_t0_flush_ifar
    cp3NIA = sim.a2o.root.iuq0.iuq_cpl_top0.iuq_cpl0.iuq_cpl_ctrl.cp3_nia_q           # nia after last cycle's completions
+      comp = ''
+      #wtf seeing something weird here
+      # there are cases where x's are in some bits of comp ifar's; maybe ok (predict array?) but why is completed indicated?
+
+      if iu0Comp.value == 1:
+         comp = f'0:{sim.safeint(iu0CompIFAR.value.binstr + "00", 2):06X} '
+
+      if iu1Comp.value == 1:
+         comp = f'{comp}1:{sim.safeint(iu1CompIFAR.value.binstr + "00", 2):06X} '
+
+      if comp == '':
+         if sim.a2o.stopOnHang != 0 and sim.cycle - lastCompCycle > sim.a2o.stopOnHang:
+            sim.ok = False
+            sim.fail = f'No completion detected in {sim.a2o.stopOnHang} cycles'
+            assert False, sim.fail
+            break
+      else:
+         comp = f'{comp}{sim.safeint(iuCompFlushIFAR.value.binstr + "00", 2):016X}'
+         sim.msg(f'C0: CP {comp}')
+         lastCompCycle = sim.cycle
+
+   */
+
+/*
+   # GPR pool and arch map
+   gprCompMap = []
+   lastGprCompMap = []
+   #wtf check what 33:36 are!
+   for i in range(36):
+      gprCompMap.append(sim.a2o.root.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.gpr_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastGprCompMap.append(i)
+
+   gpr = []
+   for i in range(144):
+      gpr.append(sim.a2o.root.xu0.gpr.gpr0.loc[i].dat)
+
+   # CR fields pool and arch map
+   crCompMap = []
+   lastCrCompMap = []
+   for i in range(8):
+      crCompMap.append(sim.a2o.root.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.cr_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastCrCompMap.append(i)
+
+   cr = []
+   for i in range(24):
+      cr.append(sim.a2o.root.xu0.cr.entry[i].reg_latch.dout)
+
+   # XER pool and arch map
+   xerCompMap = []
+   lastXerCompMap = []
+   for i in range(1):
+      xerCompMap.append(sim.a2o.root.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.xer_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastXerCompMap.append(i)
+
+   xer = []
+   for i in range(12):
+      xer.append(sim.a2o.root.xu0.xer.entry[i].reg_latch.dout)
+
+   # CTR pool and arch map
+   ctrCompMap = []
+   lastCtrCompMap = []
+   for i in range(1):
+      ctrCompMap.append(sim.a2o.root.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.ctr_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastCtrCompMap.append(i)
+
+   ctr = []
+   for i in range(8):
+      ctr.append(sim.a2o.root.xu0.ctr.entry[i].reg_latch.dout)
+
+   # LR pool and arch map
+   lrCompMap = []
+   lastLrCompMap = []
+   for i in range(1):
+      lrCompMap.append(sim.a2o.root.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.lr_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastLrCompMap.append(i)
+
+   lr = []
+   for i in range(8):
+      lr.append(sim.a2o.root.xu0.lr.entry[i].reg_latch.dout)
+
 */
 
    mem.write(0xFFFFFFFC, 0x48000002);
    mem.loadFile(testFile);
 
    m->rst = 1;
+
    cout << dec << setw(8) << cycle << " Resetting..." << endl;
 
    //m->an_ac_pm_thread_stop = threadStop;
@@ -182,7 +294,7 @@ int main(int argc, char **argv) {
    const int clocks[4] = {0x3, 0x2, 0x1, 0x0};     // 1x, 2x
    const int ticks1x = 4;
 
-   while (!Verilated::gotFinish()) {
+   while (!Verilated::gotFinish() && ok && cycle <= runCycles) {
 
       if (!resetDone && (cycle > resetCycle)) {
          m->rst = 0;
@@ -198,29 +310,50 @@ int main(int argc, char **argv) {
 
       m->clk_1x = clocks[tick % ticks1x] >> 1;
       m->clk_2x = clocks[tick % ticks1x] & 0x1;
-
-      tick++;
       m->eval();
 
-      // bus is 1x clock
+      // 1x clock
       if ((tick % ticks1x) == 0) {
 
+         // core
+         iu0Comp = root->a2owb->c0->iuq0->iuq_cpl_top0->iuq_cpl0->cp2_i0_completed;
+         iu1Comp = root->a2owb->c0->iuq0->iuq_cpl_top0->iuq_cpl0->cp2_i1_completed;
+         iu0CompIFAR = root->a2owb->c0->iuq0->iuq_cpl_top0->iuq_cpl0->cp2_i0_ifar;
+         iu1CompIFAR = root->a2owb->c0->iuq0->iuq_cpl_top0->iuq_cpl0->cp2_i1_ifar;
+
+         if (iu0Comp || iu1Comp) {
+            cout << dec << setw(8) << setfill('0') << uppercase << cycle << " Completed:";
+            if (iu0Comp)
+               cout << " I0:" << iu0Comp << " "  << setw(16) << setfill('0') << hex << (iu0CompIFAR << 2);
+            if (iu1Comp)
+               cout << " I1:" << iu1Comp << " "  << setw(16) << setfill('0') << hex << (iu1CompIFAR << 2);
+            cout << endl;
+         }
+
          // wb
          m->wb_ack = 0;
          if (wbRdPending) {
-            m->wb_datr = mem.read(m->wb_adr);
+            m->wb_datr = mem.read(m->wb_adr & 0xFFFFFFFC);
             m->wb_ack = 1;
+            cout << dec << setw(8) << setfill('0') << uppercase << cycle << " WB RD ACK RA=" << setw(8)  << hex << setfill('0') << (m->wb_adr & 0xFFFFFFFC) <<
+                  " DATA=" <<  m->wb_datr << endl;
             wbRdPending = false;
          } else if (wbWrPending) {
             mem.write(m->wb_adr, m->wb_datw, m->wb_sel);
             m->wb_ack = 1;
             wbWrPending = false;
-         }
+         } else if (m->wb_cyc && m->wb_stb) {
-
-         if (!wbRdPending && !wbWrPending && m->wb_cyc && m->wb_stb) {
             if (!m->wb_we) {
                cout << dec << setw(8) << setfill('0') << uppercase << cycle << " WB RD RA=" << setw(8)  << hex << setfill('0') << m->wb_adr << endl;
                wbRdPending = true;
+               if (m->wb_adr == bootAdr) {
+                  if (booted) {
+                     cout << "*** Fetch to boot address (" << dec << setw(8) << bootAdr << ") after initial boot! ***" << endl;
+                     ok = false;
+                  } else {
+                     booted = true;
+                  }
+               }
             } else {
                cout << dec << setw(8) << setfill('0') << uppercase << cycle << " WB WR RA=" << setw(8)  << hex << setfill('0') << m->wb_adr <<
                   " SEL=" << m->wb_sel << " DATA=" << m->wb_datw << endl;
@@ -230,6 +363,8 @@ int main(int argc, char **argv) {
 
       }
 
+      m->eval(); // NEED THIS!!!!
+
       // finish clock stuff
       if ((tick % ticks1x) == 0) {
          cycle++;
@@ -237,6 +372,9 @@ int main(int argc, char **argv) {
             cout << dec << setw(8) << setfill('0') << cycle << " ...tick..." << endl;
          }
       }
+
+      tick++;
+
       #ifdef TRACING
       t->dump(tick);
       t->flush();
@@ -244,11 +382,6 @@ int main(int argc, char **argv) {
 
       // check for fails
 
-      // hit limit
-      if (cycle > runCycles) {
-         break;
-      }
-
    }
 
 #ifdef TRACING
@@ -256,6 +389,13 @@ int main(int argc, char **argv) {
 #endif
    m->final();
 
+   cout << endl << "Cycles run=" << cycle << endl << endl;
+   if (!ok) {
+      cout << "You are worthless and weak." << endl;
+      exit(EXIT_FAILURE);
+   } else {
+      cout << "You has opulence." << endl;
       exit(EXIT_SUCCESS);
+   }
 
 }