Reduce compilation time with subroutine call

Sync all the token dependences before callsites and return, so that
there is no dependence will cross them.

visa/LocalScheduler/SWSB_G4IR.cpp

@@ -1465,6 +1465,35 @@ void SWSB::handleFuncCall() {
  }
 }
 
+void SWSB::SWSBInitializeGlobalSends(
+ SparseBitVector &allDstGlobalIDs, // node global ID bit vector in buckets
+ // touched by dst operands
+ SparseBitVector &allSrcGlobalIDs, // node global ID bit vector in buckets
+ // touched by src operands
+ LiveGRFBuckets &globalSendsLB) { // live buckets of global sends
+
+
+ // Scan the global send LB to set the node bit in dst or src bit set of each
+ // bucket that node touches.
+ for (unsigned curBucket = 0;
+ curBucket <
+ kernel.getNumRegTotal() + fg.builder->getNumScalarRegisters();
+ curBucket++) {
+
+ for (LiveGRFBuckets::BN_iterator bn_it = globalSendsLB.begin(curBucket);
+ bn_it != globalSendsLB.end(curBucket); ++bn_it) {
+ SBBucketNode *liveBN = (*bn_it);
+ SBNode *curLiveNode = liveBN->node;
+
+ if (liveBN->opndNum == Opnd_dst) {
+ allDstGlobalIDs.set(curLiveNode->globalID);
+ } else {
+ allSrcGlobalIDs.set(curLiveNode->globalID);
+ }
+ }
+ }
+}
+
 //
 // Set the global ID bit vector of each bucket touched by corresponding
 // operands
@@ -1533,6 +1562,12 @@ void SWSB::SWSBGlobalTokenGenerator(PointsToAnalysis &p, LiveGRFBuckets &LB,
 
  std::vector<SparseBitVector> dstGlobalIDs;
  std::vector<SparseBitVector> srcGlobalIDs;
+ SparseBitVector allDstGlobalIDs;
+ SparseBitVector allSrcGlobalIDs;
+
+ //Get all the global sends bitvector
+ SWSBInitializeGlobalSends(allDstGlobalIDs, allSrcGlobalIDs, globalSendsLB);
+
  if (!fg.builder->hasReadSuppressionOrSharedLocalMemoryWAs()) {
  // Initialilze for setSendGlobalIDMayKilledByCurrentBB only
  SWSBInitializeGlobalNodesInBuckets(dstGlobalIDs, srcGlobalIDs,
@@ -1579,7 +1614,9 @@ void SWSB::SWSBGlobalTokenGenerator(PointsToAnalysis &p, LiveGRFBuckets &LB,
  bb->setSendOpndMayKilled(&globalSendsLB, SBNodes, p);
  }
  } else {
- bb->setSendOpndMayKilled(&globalSendsLB, SBNodes, p);
+ if (!bb->handleCallForMayKilled(&allDstGlobalIDs, &allSrcGlobalIDs)) {
+ bb->setSendOpndMayKilled(&globalSendsLB, SBNodes, p);
+ }
  }
 
 #ifdef DEBUG_VERBOSE_ON
@@ -4408,6 +4445,23 @@ void SWSB::insertTokenSync() {
  continue;
  }
 
+ if (inst->isCall() || inst->isFCall() || inst->isReturn() ||
+ inst->isFReturn()) {
+ G4_INST *synAllInst = insertSyncAllWRInstruction(bb, 0, inst_it);
+ // In case previous inst need force A@1 in following inst
+ synAllInst->setDistance(1);
+ if (kernel.fg.builder->hasThreeALUPipes() ||
+ kernel.fg.builder->hasFourALUPipes()) {
+ synAllInst->setDistanceTypeXe(G4_INST::DistanceType::DISTALL);
+ }
+ synAllInst->setLexicalId(newInstID);
+ synAllInst = insertSyncAllRDInstruction(bb, 0, inst_it);
+ synAllInst->setLexicalId(newInstID);
+ newInstID++;
+ node_it++;
+ continue;
+ }
+
  auto jitInfo = kernel.fg.builder->getJitInfo();
  if (inst->getDistance() == 1) {
  if (kernel.fg.builder->hasThreeALUPipes() ||
@@ -5385,6 +5439,26 @@ void G4_BB_SB::getLiveOutToken(unsigned allSendNum,
  }
 }
 
+// All global dependencies will be resolved in function call
+bool G4_BB_SB::handleCallForMayKilled(SparseBitVector *allDstGlobalIDs,
+ SparseBitVector *allSrcGlobalIDs) {
+ if (!bb->size()) {
+ return false;
+ }
+
+ G4_INST *lastInst = bb->back();
+ if (!lastInst->isReturn() && !lastInst->isCall() && !lastInst->isFReturn() &&
+ !lastInst->isFCall()) {
+ return false;
+ }
+
+ send_may_kill.dst = *allDstGlobalIDs;
+ send_may_kill.src = *allSrcGlobalIDs;
+ send_WAW_may_kill = *allDstGlobalIDs;
+
+ return true;
+}
+
 //
 // Scan to check which global send operand for sends will be killed by current
 // BB. Note that there is no guarantee the send operand will in the live in set

visa/LocalScheduler/SWSB_G4IR.h

@@ -784,6 +784,9 @@ class G4_BB_SB {
  bool &sameDstSrc);
 
  // Global SBID dependence analysis
+ bool handleCallForMayKilled(SparseBitVector *allDstGlobalIDs,
+ SparseBitVector *allSrcGlobalIDs);
+
  void setSendOpndMayKilled(LiveGRFBuckets *globalSendsLB, SBNODE_VECT &SBNodes,
  PointsToAnalysis &p);
  void
@@ -1079,6 +1082,10 @@ class SWSB {
  INST_LIST_ITER inst_it, int newInstID, BitSet *dstTokens,
  BitSet *srcTokens, bool &keepDst, bool removeAllToken);
 
+ void SWSBInitializeGlobalSends(SparseBitVector &allDstGlobalIDs,
+ SparseBitVector &allSrcGlobalIDs,
+ LiveGRFBuckets &globalSendsLB);
+
  void
  SWSBInitializeGlobalNodesInBuckets(std::vector<SparseBitVector> &dstGlobalIDs,
  std::vector<SparseBitVector> &srcGlobalIDs,