Fixed assertion in call wa code.

call wa needs to keep certain BBs without deleting/merging, etc. visa optimization ifcvt
merged them, therefore, causing assertion in the later phase of callwa.

This submit does:
  1. fixed assertion by setting BB type to G4_BB_KEEP_TYPE, so that ifcvt can
     avoid if-conversion for BBs involving this type.
  2. call wa inserts join and set this join's jip to the next join instruction.
     Although this is correct functionally, it breaks if-conversion as if-conversion
     only checks control-flow and ignore any jump due to jip. Igorning jip caused
     SWSB assertion later.

     As the new join inserted by call wa will never jump, it is safe to set its JIP
     to null. Doing so will not break ifcvt. Thus, SWSB assertion will not happen.

Author: jgu222

visa/G4_BB.cpp

@@ -1296,6 +1296,8 @@ std::string G4_BB::getBBTypeStr() const {
     addTyString("FCALL");
   if ((bbTy & G4_BB_NM_WA_TYPE) != 0)
     addTyString("NoMaskWA");
+  if ((bbTy & G4_BB_KEEP_TYPE) != 0)
+    addTyString("KEEP");
   return ss.str();
 }
 

visa/G4_BB.hpp

@@ -32,7 +32,11 @@ enum G4_BB_TYPE {
   G4_BB_INIT_TYPE = 0x04,
   G4_BB_EXIT_TYPE = 0x08,
   G4_BB_NM_WA_TYPE = 0x10, // For NoMaskWA
-  G4_BB_FCALL_TYPE = 0x20  // For NoMaskWA
+  G4_BB_FCALL_TYPE = 0x20, // For NoMaskWA
+
+  // Keep BB without deleting. Its instructions stay and are not moved out.
+  // Moving instructions into BB are okay.
+  G4_BB_KEEP_TYPE = 0x40
 };
 
 class FuncInfo;

visa/Optimizer.cpp

@@ -12655,20 +12655,6 @@ void Optimizer::applyNoMaskWA() {
 // running the inst always.)
 //
 void Optimizer::applyFusedCallWA() {
-  auto getNextJoinLabel = [&](BB_LIST_ITER ITER) -> G4_Label * {
-    for (auto II = ITER, IE = fg.end(); II != IE; ++II) {
-      G4_BB *B = (*II);
-      if (G4_INST *Inst = B->getFirstInst()) {
-        if (Inst->opcode() == G4_join || Inst->opcode() == G4_endif ||
-            Inst->opcode() == G4_while) {
-          G4_INST *labelInst = B->front();
-          return labelInst->getLabel();
-        }
-      }
-    }
-    return nullptr;
-  };
-
   auto updateSubroutineTableIfNeeded = [&](G4_BB *aLeadBB, G4_BB *aB0,
                                            G4_BB *aB1, G4_BB *aS0, G4_BB *aS1,
                                            G4_BB *aEndB_or_null) {
@@ -12775,10 +12761,11 @@ void Optimizer::applyFusedCallWA() {
     bigB1->push_back(gotoEnd);
 
     // Need to insert a join in nextBB
+    // This join will never jump, thus set its JIP to nullptr.
     G4_INST *tjoin = nextBB->getFirstInst();
     if (tjoin == nullptr || tjoin->opcode() != G4_join) {
       G4_INST *finalJoin = builder.createCFInst(nullptr, G4_join, simdsz,
-                                                getNextJoinLabel(nextBI),
+                                                nullptr,
                                                 nullptr, InstOpt_NoOpt, false);
       if (tjoin == nullptr) {
         nextBB->insertBefore(nextBB->end(), finalJoin);
@@ -12943,6 +12930,12 @@ void Optimizer::applyFusedCallWA() {
       kernel.m_indirectCallWAInfo.emplace(
           BB, IndirectCallWAInfo(bigB0, smallB0, nullptr, nullptr, nullptr,
                                  nullptr, nullptr, callI, nCallI));
+      // BB, bigB0, smallB0 should not be deleted and its instructions shall
+      // stay inside. Set BB type to G4_BB_KEEP_TYPE so the other optim passes
+      // will not delete them.
+      BB->setBBType(G4_BB_KEEP_TYPE);
+      bigB0->setBBType(G4_BB_KEEP_TYPE);
+      smallB0->setBBType(G4_BB_KEEP_TYPE);
     } else {
       // relative target:  need to patch offset after SWSB in
       // finishFusedCallWA()

visa/ifcvt.cpp

@@ -451,6 +451,13 @@ void IfConverter::analyze(std::vector<IfConvertible> &list) {
       continue;
     }
 
+    // Conservatively skip if BB is set with G4_BB_KEEP_TYPE
+    if ((t && (t->getBBType() & G4_BB_KEEP_TYPE)) ||
+      (s0 && (s0->getBBType() & G4_BB_KEEP_TYPE)) ||
+      (s1 && (s1->getBBType() & G4_BB_KEEP_TYPE))) {
+      continue;
+    }
+
     G4_Predicate *pred = ifInst->getPredicate();
 
     unsigned n0 = getPredictableInsts(s0, ifInst);