Reland the WA that emits a NOP after branch instruction to support single stepping in debugger.

The WA is enabled only when compiling with -O0 currently.

Author: petechou

visa/HWCaps.inc

@@ -801,4 +801,13 @@ bool WARLocalization() const {
 bool supportPureBF() const {
   return false;
 }
+
+// WA to support debugger stepping. Currently only enabled when compiling with
+// -debug.
+bool needNopAfterCFInstWA() const {
+  const TARGET_PLATFORM p = getPlatform();
+  if (p == Xe_DG2 || p == Xe_PVC || p == Xe_PVCXT)
+    return getOption(vISA_Debug);
+  return false;
+}
 // end HW capabilities

visa/Optimizer.h

@@ -261,6 +261,7 @@ class Optimizer {
   void applyNamedBarrierWA(INST_LIST_ITER it, G4_BB *bb);
   void insertIEEEExceptionTrap();
   void expandIEEEExceptionTrap(INST_LIST_ITER it, G4_BB *bb);
+  void insertNopAfterCFInst(BB_LIST_ITER ib, INST_LIST_ITER it);
 
   typedef std::vector<vISA::G4_INST *> InstListType;
   // create instruction sequence to calculate call offset from ip

visa/SWWA.cpp

@@ -3337,6 +3337,36 @@ void Optimizer::expandIEEEExceptionTrap(INST_LIST_ITER it, G4_BB *bb) {
   *it = restoreFlag;
 }
 
+void Optimizer::insertNopAfterCFInst(BB_LIST_ITER ib, INST_LIST_ITER ii) {
+  // If the CF inst is not the last inst of current BB, simply insert
+  // a nop after it. Otherwise, insert a dummy BB to hold a nop between
+  // the current BB and the fall-through BB.
+  G4_BB *bb = *ib;
+  G4_INST *inst = *ii;
+  if (inst != bb->back()) {
+    bb->insertAfter(ii, builder.createNop(InstOpt_NoOpt));
+  } else {
+    G4_BB *dummyBB = fg.createNewBBWithLabel("CFInstWA_BB");
+    dummyBB->push_back(builder.createNop(InstOpt_NoOpt));
+    BB_LIST_ITER nextBI = std::next(ib);
+    fg.insert(nextBI, dummyBB);
+    // Update the pred/succ edges accordingly.
+    bb->setPhysicalSucc(dummyBB);
+    dummyBB->setPhysicalPred(bb);
+    if (nextBI != fg.end()) {
+      G4_BB *nextBB = *nextBI;
+      dummyBB->setPhysicalSucc(nextBB);
+      nextBB->setPhysicalPred(dummyBB);
+      if (std::any_of(bb->Succs.begin(), bb->Succs.end(),
+                      [=](G4_BB *succ) { return succ == nextBB; } )) {
+        fg.removePredSuccEdges(bb, nextBB);
+        fg.addPredSuccEdges(bb, dummyBB);
+        fg.addPredSuccEdges(dummyBB, nextBB);
+      }
+    }
+  }
+}
+
 // For a subroutine, insert a dummy move with {Switch} option immediately
 // before the first non-label instruction in BB. Otherwie, for a following
 // basic block, insert a dummy move before *any* instruction to ensure that
@@ -3767,6 +3797,14 @@ void Optimizer::HWWorkaround() {
       if (inst->isIEEEExceptionTrap())
         expandIEEEExceptionTrap(ii, bb);
 
+      // Currently skip adding Nop for the predicated backward goto as IP + 1
+      // of the goto might need to be the join instruction.
+      const bool isPredicatedBwdGoto = inst->opcode() == G4_goto &&
+          inst->getPredicate() && inst->asCFInst()->isBackward();
+      if (inst->isCFInst() && !isPredicatedBwdGoto &&
+          builder.needNopAfterCFInstWA())
+        insertNopAfterCFInst(ib, ii);
+
       ii++;
     }
   }