Use cross block load vectorization for new inline raytracing by default

Use cross block load vectorization for new inline raytracing by default

Author: jaladreips

IGC/AdaptorCommon/LivenessUtils/AllocationLivenessAnalyzer.cpp

@@ -33,7 +33,7 @@ using namespace llvm;
 using namespace IGC;
 
 AllocationLivenessAnalyzer::LivenessData
-AllocationLivenessAnalyzer::ProcessInstruction(Instruction *I, DominatorTree &DT, LoopInfo &LI) {
+AllocationLivenessAnalyzer::ProcessInstruction(Instruction *I, DominatorTree &DT, LoopInfo &LI, bool includeOrigin) {
   // static allocas are usually going to be in the entry block
   // that's a practice, but we only care about the last block that dominates all uses
   BasicBlock *commonDominator = nullptr;
@@ -93,14 +93,9 @@ AllocationLivenessAnalyzer::ProcessInstruction(Instruction *I, DominatorTree &DT
         }
       }
     } break;
-    case Instruction::Call: {
-      auto *callI = cast<CallInst>(II);
-      if (!callI->doesNotCapture(use->getOperandNo()))
-        lifetimeLeakingUsers.insert(II);
-
-      if (II->getType()->isPointerTy())
-        addUsesFn(II->uses());
-    } break;
+    case Instruction::Call:
+      implementCallSpecificBehavior(cast<CallInst>(II), use, worklist, allUsers, lifetimeLeakingUsers);
+      break;
     case Instruction::Load:
       if (II->getType()->isPointerTy())
         addUsesFn(II->uses());
@@ -111,6 +106,9 @@ AllocationLivenessAnalyzer::ProcessInstruction(Instruction *I, DominatorTree &DT
     }
   }
 
+  if (includeOrigin)
+    allUsers.insert(I);
+
   return LivenessData(I, std::move(allUsers), LI, DT, commonDominator, std::move(lifetimeLeakingUsers));
 }
 
@@ -120,6 +118,20 @@ void AllocationLivenessAnalyzer::getAnalysisUsage(llvm::AnalysisUsage &AU) const
   getAdditionalAnalysisUsage(AU);
 }
 
+void AllocationLivenessAnalyzer::implementCallSpecificBehavior(CallInst *callI, Use* use, SmallVector<Use *> &worklist,
+                                                               SetVector<Instruction *> &allUsers,
+                                                               SetVector<Instruction *> &lifetimeLeakingUsers) {
+
+  if (!callI->doesNotCapture(use->getOperandNo()))
+    lifetimeLeakingUsers.insert(callI);
+
+  if (callI->getType()->isPointerTy()) {
+
+    for (auto &use : callI->uses())
+      worklist.push_back(&use);
+  }
+}
+
 template <typename range>
 static inline void doWorkLoop(SmallVector<BasicBlock *> &worklist, DenseSet<BasicBlock *> &bbSet1,
                               DenseSet<BasicBlock *> &bbSet2, std::function<range(BasicBlock *)> iterate,

IGC/AdaptorCommon/LivenessUtils/AllocationLivenessAnalyzer.h

@@ -15,9 +15,11 @@ SPDX-License-Identifier: MIT
 
 namespace llvm {
 class BasicBlock;
+class CallInst;
 class DominatorTree;
 class Instruction;
 class LoopInfo;
+class Use;
 } // namespace llvm
 
 namespace IGC {
@@ -49,10 +51,15 @@ class AllocationLivenessAnalyzer : public llvm::FunctionPass {
   AllocationLivenessAnalyzer(char &pid) : llvm::FunctionPass(pid) {}
 
 protected:
-  LivenessData ProcessInstruction(llvm::Instruction *I, llvm::DominatorTree &DT, llvm::LoopInfo &LI);
+  LivenessData ProcessInstruction(llvm::Instruction *I, llvm::DominatorTree &DT, llvm::LoopInfo &LI,
+                                  bool includeOrigin = false);
 
   void getAnalysisUsage(llvm::AnalysisUsage &AU) const override;
   virtual void getAdditionalAnalysisUsage(llvm::AnalysisUsage &AU) const = 0;
+  virtual void implementCallSpecificBehavior(llvm::CallInst *I, llvm::Use *use,
+                                             llvm::SmallVector<llvm::Use *> &worklist,
+                                             llvm::SetVector<llvm::Instruction *> &allUsers,
+                                             llvm::SetVector<llvm::Instruction *> &lifetimeLeakingUsers);
 };
 
 namespace Provenance {

IGC/common/igc_flags.h

@@ -1844,7 +1844,7 @@ DECLARE_IGC_REGKEY_BITMASK(UseNewInlineRaytracing, 4, "Use the new rayquery impl
                            NEW_INLINE_RAYTRACING_MASK, true)
 DECLARE_IGC_REGKEY(DWORD, AddDummySlotsForNewInlineRaytracing, 0,
                    "Add dummy rayquery slots when doing new inline raytracing", true)
-DECLARE_IGC_REGKEY(bool, UseCrossBlockLoadVectorizationForInlineRaytracing, false,
+DECLARE_IGC_REGKEY(bool, UseCrossBlockLoadVectorizationForInlineRaytracing, true,
                    "If enabled, will try to vectorize loads that are not adjacent to each other. May increase GRF pressure", true)
 DECLARE_IGC_REGKEY(bool, OverrideRayQueryThrottling, false,
                    "Force rayquery throttling (dynamic ray management) to be enabled or disabled. Default value of "