[Autobackout][FuncReg]Revert of change: 9fa74ab

 merge uav writes from branches

Merge complementary UAV writes from if/else branches to avoid partial
writes

Author: fftzeng

IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp

@@ -2143,7 +2143,6 @@ void OptimizeIR(CodeGenContext* const pContext)
             // more efficient sequences of multiplies, shifts, and adds
             mpm.add(createIntDivConstantReductionPass());
         }
-        mpm.add(createMergeMemFromBranchOptPass());
 
         mpm.add(createConstantMergePass());
 

IGC/Compiler/CustomSafeOptPass.cpp

@@ -5686,185 +5686,6 @@ bool CleanPHINode::runOnFunction(Function& F)
     return changed;
 }
 
-namespace {
-    class MergeMemFromBranchOpt : public FunctionPass, public llvm::InstVisitor<MergeMemFromBranchOpt>
-    {
-    public:
-        static char ID;
-        MergeMemFromBranchOpt();
-
-        StringRef getPassName() const override { return "MergeMemFromBranchOpt"; }
-
-        bool runOnFunction(Function& F) override;
-        virtual void getAnalysisUsage(llvm::AnalysisUsage& AU) const override
-        {
-            AU.setPreservesCFG();
-            AU.addRequired<CodeGenContextWrapper>();
-        }
-
-        void visitCallInst(llvm::CallInst& C);
-        void visitTypedWrite(llvm::CallInst* inst);
-    private:
-        CodeGenContext* pContext = nullptr;
-    };
-}
-
-#undef PASS_FLAG
-#undef PASS_DESCRIPTION
-#undef PASS_CFG_ONLY
-#undef PASS_ANALYSIS
-#define PASS_FLAG "igc-MergeMemFromBranchOpt"
-#define PASS_DESCRIPTION "Merge Mem from Branch Opt"
-#define PASS_CFG_ONLY false
-#define PASS_ANALYSIS false
-IGC_INITIALIZE_PASS_BEGIN(MergeMemFromBranchOpt, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
-IGC_INITIALIZE_PASS_END(MergeMemFromBranchOpt, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
-
-/*
-This optimization merge the UAV writes from if/else branch to the common successor to avoid partial writes
-
-ex: Change
-    Label-1337.i:                                     ; preds = %Label-1339.i, %Label-1780.i
-      <skip instructions>
-      %494 = inttoptr i32 %0 to %__2D_DIM_Resource addrspace(2490368)*
-      call void @llvm.genx.GenISA.typedwrite.p2490368__2D_DIM_Resource(%__2D_DIM_Resource addrspace(2490368)* %494, i32 %17, i32 %18, i32 0, i32 0, float %491, float %492, float %493, float %Temp-2360.i163)
-      br label %Output
-
-    Label-1313.i:                                     ; preds = %Label-1577.i
-      %495 = inttoptr i32 %0 to %__2D_DIM_Resource addrspace(2490368)*
-      call void @llvm.genx.GenISA.typedwrite.p2490368__2D_DIM_Resource(%__2D_DIM_Resource addrspace(2490368)* %495, i32 %17, i32 %18, i32 0, i32 0, float %scalar39, float %scalar40, float %scalar41, float %scalar42)
-      br label %Output
-
-    Output:                                           ; preds = %Label-1313.i, %Label-1337.i
-      ret void
-
-                      to
-
-    Output:                                           ; preds = %Label-1313.i, %Label-1337.i
-      %494 = phi float [ %scalar39, %Label-1313.i ], [ %491, %Label-1337.i ]
-      %495 = phi float [ %scalar40, %Label-1313.i ], [ %492, %Label-1337.i ]
-      %496 = phi float [ %scalar41, %Label-1313.i ], [ %493, %Label-1337.i ]
-      %497 = phi float [ %scalar42, %Label-1313.i ], [ %Temp-2360.i163, %Label-1337.i ]
-      %498 = inttoptr i32 %0 to %__2D_DIM_Resource addrspace(2490368)*
-      call void @llvm.genx.GenISA.typedwrite.p2490368__2D_DIM_Resource(%__2D_DIM_Resource addrspace(2490368)* %498, i32 %17, i32 %18, i32 0, i32 0, float %494, float %495, float %496, float %497)
-      ret void
-    }
-*/
-
-char MergeMemFromBranchOpt::ID = 0;
-FunctionPass* IGC::createMergeMemFromBranchOptPass()
-{
-    return new MergeMemFromBranchOpt();
-}
-
-MergeMemFromBranchOpt::MergeMemFromBranchOpt() : FunctionPass(ID)
-{
-    initializeMergeMemFromBranchOptPass(*PassRegistry::getPassRegistry());
-}
-
-void MergeMemFromBranchOpt::visitTypedWrite(llvm::CallInst* inst)
-{
-    std::vector<CallInst*> callSet;
-    // last instruction in the BB
-    if (dyn_cast<BranchInst>(inst->getNextNode()))
-    {
-        BasicBlock* BB = inst->getParent();
-        BasicBlock* BB2 = nullptr;
-        // the basicblock with inst has single successor BB
-        if (BasicBlock* succBB = BB->getSingleSuccessor())
-        {
-            callSet.push_back(inst);
-            // find the other BB with the same successor
-            for (pred_iterator PI = pred_begin(succBB), E = pred_end(succBB); PI != E; ++PI)
-            {
-                BB2 = *PI;
-                if (BB2 != BB && BB2->getInstList().size() > 1)
-                {
-                    BranchInst* br = dyn_cast<BranchInst>(&BB2->getInstList().back());
-                    if (br && br->getPrevNode())
-                    {
-                        if (llvm::GenIntrinsicInst* giInst = llvm::dyn_cast<GenIntrinsicInst>(br->getPrevNode()))
-                        {
-                            if (giInst->getIntrinsicID() == GenISAIntrinsic::GenISA_typedwrite)
-                            {
-                                if (((giInst->getOperand(0))->getType()->getPointerAddressSpace() ==
-                                    (inst->getOperand(0))->getType()->getPointerAddressSpace()) &&
-                                    (giInst->getOperand(1) == inst->getOperand(1)) &&
-                                    (giInst->getOperand(2) == inst->getOperand(2)) &&
-                                    (giInst->getOperand(3) == inst->getOperand(3)) &&
-                                    (giInst->getOperand(4) == inst->getOperand(4)))
-                                {
-                                    callSet.push_back(giInst);
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-
-            // if there are more predecessor than the ones with urb partial write,
-            // skip the optimization as it might increase the number of times it needs to do the urb write.
-            if (succBB->hasNPredecessorsOrMore(callSet.size() + 1))
-                return;
-
-            if (callSet.size() < 2)
-                return;
-
-            // start the conversion
-            PHINode* p8 = PHINode::Create(inst->getOperand(8)->getType(), callSet.size(), "", &(succBB->front()));
-            PHINode* p7 = PHINode::Create(inst->getOperand(7)->getType(), callSet.size(), "", &(succBB->front()));
-            PHINode* p6 = PHINode::Create(inst->getOperand(6)->getType(), callSet.size(), "", &(succBB->front()));
-            PHINode* p5 = PHINode::Create(inst->getOperand(5)->getType(), callSet.size(), "", &(succBB->front()));
-            for (unsigned int i = 0; i < callSet.size(); i++)
-            {
-                p5->addIncoming(callSet[i]->getOperand(5), callSet[i]->getParent());
-                p6->addIncoming(callSet[i]->getOperand(6), callSet[i]->getParent());
-                p7->addIncoming(callSet[i]->getOperand(7), callSet[i]->getParent());
-                p8->addIncoming(callSet[i]->getOperand(8), callSet[i]->getParent());
-            }
-
-            inst->removeFromParent();
-            inst->insertAfter(p8);
-
-            if (inst->getOperand(0) && inst->getOperand(0)->hasOneUse())
-            {
-                if (Instruction* itp = dyn_cast<Instruction>(inst->getOperand(0)))
-                {
-                    itp->removeFromParent();
-                    itp->insertAfter(p8);
-                }
-            }
-
-            inst->setOperand(5, p5);
-            inst->setOperand(6, p6);
-            inst->setOperand(7, p7);
-            inst->setOperand(8, p8);
-
-            for (unsigned int i = 1; i < callSet.size(); i++)
-            {
-                callSet[i]->eraseFromParent();
-            }
-        }
-    }
-}
-
-void MergeMemFromBranchOpt::visitCallInst(CallInst& C)
-{
-    if (llvm::GenIntrinsicInst* inst = llvm::dyn_cast<GenIntrinsicInst>(&C))
-    {
-        if (inst->getIntrinsicID() == GenISAIntrinsic::GenISA_typedwrite)
-        {
-            visitTypedWrite(inst);
-        }
-    }
-}
-
-bool MergeMemFromBranchOpt::runOnFunction(Function& F)
-{
-    visit(F);
-    return false;
-}
-
 namespace {
     class InsertBranchOpt : public FunctionPass
     {

IGC/Compiler/CustomSafeOptPass.hpp

@@ -250,5 +250,5 @@ namespace IGC
     llvm::FunctionPass* createLogicalAndToBranchPass();
     llvm::FunctionPass* createCleanPHINodePass();
     llvm::FunctionPass* createInsertBranchOptPass();
-    llvm::FunctionPass* createMergeMemFromBranchOptPass();
+
 } // namespace IGC