1 //===-- AMDGPUAnnotateUniformValues.cpp - ---------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 /// \file 10 /// This pass adds amdgpu.uniform metadata to IR values so this information 11 /// can be used during instruction selection. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "AMDGPU.h" 16 #include "Utils/AMDGPUBaseInfo.h" 17 #include "llvm/ADT/SmallSet.h" 18 #include "llvm/Analysis/LegacyDivergenceAnalysis.h" 19 #include "llvm/Analysis/MemorySSA.h" 20 #include "llvm/IR/InstVisitor.h" 21 #include "llvm/IR/IntrinsicsAMDGPU.h" 22 #include "llvm/InitializePasses.h" 23 24 #define DEBUG_TYPE "amdgpu-annotate-uniform" 25 26 using namespace llvm; 27 28 namespace { 29 30 class AMDGPUAnnotateUniformValues : public FunctionPass, 31 public InstVisitor<AMDGPUAnnotateUniformValues> { 32 LegacyDivergenceAnalysis *DA; 33 MemorySSA *MSSA; 34 DenseMap<Value*, GetElementPtrInst*> noClobberClones; 35 bool isEntryFunc; 36 37 public: 38 static char ID; 39 AMDGPUAnnotateUniformValues() : 40 FunctionPass(ID) { } 41 bool doInitialization(Module &M) override; 42 bool runOnFunction(Function &F) override; 43 StringRef getPassName() const override { 44 return "AMDGPU Annotate Uniform Values"; 45 } 46 void getAnalysisUsage(AnalysisUsage &AU) const override { 47 AU.addRequired<LegacyDivergenceAnalysis>(); 48 AU.addRequired<MemorySSAWrapperPass>(); 49 AU.setPreservesAll(); 50 } 51 52 void visitBranchInst(BranchInst &I); 53 void visitLoadInst(LoadInst &I); 54 bool isClobberedInFunction(LoadInst * Load); 55 }; 56 57 } // End anonymous namespace 58 59 INITIALIZE_PASS_BEGIN(AMDGPUAnnotateUniformValues, DEBUG_TYPE, 60 "Add AMDGPU uniform metadata", false, false) 61 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis) 62 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) 63 INITIALIZE_PASS_END(AMDGPUAnnotateUniformValues, DEBUG_TYPE, 64 "Add AMDGPU uniform metadata", false, false) 65 66 char AMDGPUAnnotateUniformValues::ID = 0; 67 68 static void setUniformMetadata(Instruction *I) { 69 I->setMetadata("amdgpu.uniform", MDNode::get(I->getContext(), {})); 70 } 71 static void setNoClobberMetadata(Instruction *I) { 72 I->setMetadata("amdgpu.noclobber", MDNode::get(I->getContext(), {})); 73 } 74 75 bool AMDGPUAnnotateUniformValues::isClobberedInFunction(LoadInst *Load) { 76 MemorySSAWalker *Walker = MSSA->getWalker(); 77 SmallVector<MemoryAccess *> WorkList{Walker->getClobberingMemoryAccess(Load)}; 78 SmallSet<MemoryAccess *, 8> Visited; 79 MemoryLocation Loc(MemoryLocation::get(Load)); 80 81 const auto isReallyAClobber = [](MemoryDef *Def) -> bool { 82 Instruction *DefInst = Def->getMemoryInst(); 83 LLVM_DEBUG(dbgs() << " Def: " << *DefInst << '\n'); 84 85 if (isa<FenceInst>(DefInst)) 86 return false; 87 88 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(DefInst)) { 89 switch (II->getIntrinsicID()) { 90 case Intrinsic::amdgcn_s_barrier: 91 case Intrinsic::amdgcn_wave_barrier: 92 return false; 93 default: 94 break; 95 } 96 } 97 98 return true; 99 }; 100 101 LLVM_DEBUG(dbgs() << "Checking clobbering of: " << *Load << '\n'); 102 103 // Start with a nearest dominating clobbering access, it will be either 104 // live on entry (nothing to do, load is not clobbered), MemoryDef, or 105 // MemoryPhi if several MemoryDefs can define this memory state. In that 106 // case add all Defs to WorkList and continue going up and checking all 107 // the definitions of this memory location until the root. When all the 108 // defs are exhausted and came to the entry state we have no clobber. 109 // Along the scan ignore barriers and fences which are considered clobbers 110 // by the MemorySSA, but not really writing anything into the memory. 111 while (!WorkList.empty()) { 112 MemoryAccess *MA = WorkList.pop_back_val(); 113 if (!Visited.insert(MA).second) 114 continue; 115 116 if (MSSA->isLiveOnEntryDef(MA)) 117 continue; 118 119 if (MemoryDef *Def = dyn_cast<MemoryDef>(MA)) { 120 if (isReallyAClobber(Def)) { 121 LLVM_DEBUG(dbgs() << " -> load is clobbered\n"); 122 return true; 123 } 124 125 WorkList.push_back( 126 Walker->getClobberingMemoryAccess(Def->getDefiningAccess(), Loc)); 127 continue; 128 } 129 130 const MemoryPhi *Phi = cast<MemoryPhi>(MA); 131 for (auto &Use : Phi->incoming_values()) 132 WorkList.push_back(cast<MemoryAccess>(&Use)); 133 } 134 135 LLVM_DEBUG(dbgs() << " -> no clobber\n"); 136 return false; 137 } 138 139 void AMDGPUAnnotateUniformValues::visitBranchInst(BranchInst &I) { 140 if (DA->isUniform(&I)) 141 setUniformMetadata(&I); 142 } 143 144 void AMDGPUAnnotateUniformValues::visitLoadInst(LoadInst &I) { 145 Value *Ptr = I.getPointerOperand(); 146 if (!DA->isUniform(Ptr)) 147 return; 148 // We're tracking up to the Function boundaries, and cannot go beyond because 149 // of FunctionPass restrictions. We can ensure that is memory not clobbered 150 // for memory operations that are live in to entry points only. 151 Instruction *PtrI = dyn_cast<Instruction>(Ptr); 152 153 if (!isEntryFunc) { 154 if (PtrI) 155 setUniformMetadata(PtrI); 156 return; 157 } 158 159 bool NotClobbered = false; 160 bool GlobalLoad = I.getPointerAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS; 161 if (PtrI) 162 NotClobbered = GlobalLoad && !isClobberedInFunction(&I); 163 else if (isa<Argument>(Ptr) || isa<GlobalValue>(Ptr)) { 164 if (GlobalLoad && !isClobberedInFunction(&I)) { 165 NotClobbered = true; 166 // Lookup for the existing GEP 167 if (noClobberClones.count(Ptr)) { 168 PtrI = noClobberClones[Ptr]; 169 } else { 170 // Create GEP of the Value 171 Function *F = I.getParent()->getParent(); 172 Value *Idx = Constant::getIntegerValue( 173 Type::getInt32Ty(Ptr->getContext()), APInt(64, 0)); 174 // Insert GEP at the entry to make it dominate all uses 175 PtrI = GetElementPtrInst::Create(I.getType(), Ptr, 176 ArrayRef<Value *>(Idx), Twine(""), 177 F->getEntryBlock().getFirstNonPHI()); 178 } 179 I.replaceUsesOfWith(Ptr, PtrI); 180 } 181 } 182 183 if (PtrI) { 184 setUniformMetadata(PtrI); 185 if (NotClobbered) 186 setNoClobberMetadata(PtrI); 187 } 188 } 189 190 bool AMDGPUAnnotateUniformValues::doInitialization(Module &M) { 191 return false; 192 } 193 194 bool AMDGPUAnnotateUniformValues::runOnFunction(Function &F) { 195 if (skipFunction(F)) 196 return false; 197 198 DA = &getAnalysis<LegacyDivergenceAnalysis>(); 199 MSSA = &getAnalysis<MemorySSAWrapperPass>().getMSSA(); 200 isEntryFunc = AMDGPU::isEntryFunctionCC(F.getCallingConv()); 201 202 visit(F); 203 noClobberClones.clear(); 204 return true; 205 } 206 207 FunctionPass * 208 llvm::createAMDGPUAnnotateUniformValues() { 209 return new AMDGPUAnnotateUniformValues(); 210 } 211