gather send optimization

gather send optimization

Author: bcheng0127

visa/G4_Declare.h

@@ -94,6 +94,10 @@ class G4_Declare {
   uint16_t forceSpilled : 1;
   uint16_t exclusiveLoad : 1;
   uint16_t isCmpUseOnly : 1;
+  // indicate if the declare is local referenced only
+  // Especially for the variable with pseodu_kill,
+  // while will be removed in removeLifetimeOps pass.
+  uint16_t isBBLocal : 1;
 
   unsigned declId; // global decl id for this builder
 
@@ -334,6 +338,9 @@ class G4_Declare {
   void setIsCmpUseOnly(bool b) { isCmpUseOnly = b; }
   bool getIsCmpUseOnly() const { return isCmpUseOnly; }
 
+  void setIsBBLocal(bool b) { isBBLocal = b; }
+  bool getIsBBLocal() const { return isBBLocal; }
+
   unsigned getNumRegNeeded() const;
 
   void emit(std::ostream &output) const;

visa/LocalDataflow.cpp

@@ -453,6 +453,47 @@ static void processReadOpnds(G4_BB *BB, G4_INST *Inst, LocalLivenessInfo &LLI) {
   }
 }
 
+static void
+processReadOpndsForPseudoKill(G4_BB *BB, G4_INST *Inst,
+                              std::unordered_set<G4_Declare *> &pseudoKills) {
+  if (Inst->isPseudoKill()) {
+    return;
+  }
+  // (1) Indirect dst operand reads address.
+  G4_DstRegRegion *Dst = Inst->getDst();
+  if (Dst && Dst->isIndirect()) {
+    G4_Declare *dcl = Dst->getTopDcl();
+    if (pseudoKills.find(dcl) != pseudoKills.end()) {
+      dcl->setIsBBLocal(false);
+      pseudoKills.erase(dcl);
+    }
+  }
+
+  // (2) Direct and indirect source operands.
+  for (auto OpNum :
+       {Gen4_Operand_Number::Opnd_src0, Gen4_Operand_Number::Opnd_src1,
+        Gen4_Operand_Number::Opnd_src2, Gen4_Operand_Number::Opnd_src3,
+        Gen4_Operand_Number::Opnd_src4, Gen4_Operand_Number::Opnd_src5,
+        Gen4_Operand_Number::Opnd_src6, Gen4_Operand_Number::Opnd_src7,
+        Gen4_Operand_Number::Opnd_pred, Gen4_Operand_Number::Opnd_implAccSrc}) {
+    G4_Operand *opnd = Inst->getOperand(OpNum);
+    if (opnd == nullptr || opnd->isImm() || opnd->isNullReg() ||
+        opnd->isLabel())
+      continue;
+
+    G4_Declare *dcl = nullptr;
+    if (Inst->isPseudoAddrMovIntrinsic()) {
+      dcl =opnd->asAddrExp()->getRegVar()->getDeclare();
+    } else {
+      dcl = opnd->getTopDcl();
+    }
+    if (pseudoKills.find(dcl) != pseudoKills.end()) {
+      dcl->setIsBBLocal(false);
+      pseudoKills.erase(dcl);
+    }
+  }
+}
+
 // Process writes. If this is a partial definition, then record this partial
 // definition. When all partial definitions together define this live read node,
 // it is killed and du/ud links are added.
@@ -487,6 +528,12 @@ static void processWriteOpnds(G4_BB *BB, G4_INST *Inst,
 }
 
 void FlowGraph::localDataFlowAnalysis() {
+  // For pseudo kill varaible
+  // If there is use exposed in a BB, it's treated as global.
+  // Otherwise, it's treated as local even the same pseudo kill may appear in
+  // multiple BBs
+  std::unordered_set<G4_Declare *> pesudoKilledDcls;
+
   for (auto BB : BBs) {
     LocalLivenessInfo LLI(!BB->isAllLaneActive());
     for (auto I = BB->rbegin(), E = BB->rend(); I != E; ++I) {
@@ -504,7 +551,24 @@ void FlowGraph::localDataFlowAnalysis() {
         continue;
       }
       processWriteOpnds(BB, Inst, LLI);
+
+      if (Inst->isPseudoKill() && Inst->getDst() && !Inst->getDst()->isNullReg()) {
+        G4_Declare *dcl = Inst->getDst()->getTopDcl();
+        pesudoKilledDcls.insert(dcl);
+        // In case the use in anther BB is analyzed before define
+        if (!globalOpndHT.isOpndGlobal(Inst->getDst())) {
+          G4_Declare *dcl = Inst->getDst()->getTopDcl();
+          dcl->setIsBBLocal(true);
+        }
+      }
+
       processReadOpnds(BB, Inst, LLI);
+      if (pesudoKilledDcls
+              .size()) { // Process the operand using variable which
+                         // has psuedo kill. Since the scan is from back to
+                         // front, exposed use will make variable global
+        processReadOpndsForPseudoKill(BB, Inst, pesudoKilledDcls);
+      }
     }
 
     // All left over live nodes are global.

visa/Optimizer.cpp

@@ -834,7 +834,7 @@ void Optimizer::s0SubAfterRA() {
   kernel.fg.resetLocalDataFlowData();
   kernel.fg.localDataFlowAnalysis();
 
-  SRSubPassBeforeRA s0Sub(builder, kernel);
+  SRSubPassAfterRA s0Sub(builder, kernel);
   s0Sub.run();
 }