1 //===-- AMDGPUAnnotateUniformValues.cpp - ---------------------------------===//
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
7 //===----------------------------------------------------------------------===//
10 /// This pass adds amdgpu.uniform metadata to IR values so this information
11 /// can be used during instruction selection.
13 //===----------------------------------------------------------------------===//
16 #include "llvm/ADT/SetVector.h"
17 #include "llvm/Analysis/LegacyDivergenceAnalysis.h"
18 #include "llvm/Analysis/LoopInfo.h"
19 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
20 #include "llvm/IR/IRBuilder.h"
21 #include "llvm/IR/InstVisitor.h"
22 #include "llvm/InitializePasses.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/raw_ostream.h"
26 #define DEBUG_TYPE "amdgpu-annotate-uniform"
32 class AMDGPUAnnotateUniformValues : public FunctionPass,
33 public InstVisitor<AMDGPUAnnotateUniformValues> {
34 LegacyDivergenceAnalysis *DA;
35 MemoryDependenceResults *MDR;
37 DenseMap<Value*, GetElementPtrInst*> noClobberClones;
42 AMDGPUAnnotateUniformValues() :
44 bool doInitialization(Module &M) override;
45 bool runOnFunction(Function &F) override;
46 StringRef getPassName() const override {
47 return "AMDGPU Annotate Uniform Values";
49 void getAnalysisUsage(AnalysisUsage &AU) const override {
50 AU.addRequired<LegacyDivergenceAnalysis>();
51 AU.addRequired<MemoryDependenceWrapperPass>();
52 AU.addRequired<LoopInfoWrapperPass>();
56 void visitBranchInst(BranchInst &I);
57 void visitLoadInst(LoadInst &I);
58 bool isClobberedInFunction(LoadInst * Load);
61 } // End anonymous namespace
63 INITIALIZE_PASS_BEGIN(AMDGPUAnnotateUniformValues, DEBUG_TYPE,
64 "Add AMDGPU uniform metadata", false, false)
65 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis)
66 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
67 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
68 INITIALIZE_PASS_END(AMDGPUAnnotateUniformValues, DEBUG_TYPE,
69 "Add AMDGPU uniform metadata", false, false)
71 char AMDGPUAnnotateUniformValues::ID = 0;
73 static void setUniformMetadata(Instruction *I) {
74 I->setMetadata("amdgpu.uniform", MDNode::get(I->getContext(), {}));
76 static void setNoClobberMetadata(Instruction *I) {
77 I->setMetadata("amdgpu.noclobber", MDNode::get(I->getContext(), {}));
80 static void DFS(BasicBlock *Root, SetVector<BasicBlock*> & Set) {
81 for (auto I : predecessors(Root))
86 bool AMDGPUAnnotateUniformValues::isClobberedInFunction(LoadInst * Load) {
87 // 1. get Loop for the Load->getparent();
88 // 2. if it exists, collect all the BBs from the most outer
89 // loop and check for the writes. If NOT - start DFS over all preds.
90 // 3. Start DFS over all preds from the most outer loop header.
91 SetVector<BasicBlock *> Checklist;
92 BasicBlock *Start = Load->getParent();
93 Checklist.insert(Start);
94 const Value *Ptr = Load->getPointerOperand();
95 const Loop *L = LI->getLoopFor(Start);
100 P = P->getParentLoop();
102 Checklist.insert(L->block_begin(), L->block_end());
103 Start = L->getHeader();
106 DFS(Start, Checklist);
107 for (auto &BB : Checklist) {
108 BasicBlock::iterator StartIt = (!L && (BB == Load->getParent())) ?
109 BasicBlock::iterator(Load) : BB->end();
110 auto Q = MDR->getPointerDependencyFrom(MemoryLocation(Ptr), true,
112 if (Q.isClobber() || Q.isUnknown())
118 void AMDGPUAnnotateUniformValues::visitBranchInst(BranchInst &I) {
119 if (DA->isUniform(&I))
120 setUniformMetadata(I.getParent()->getTerminator());
123 void AMDGPUAnnotateUniformValues::visitLoadInst(LoadInst &I) {
124 Value *Ptr = I.getPointerOperand();
125 if (!DA->isUniform(Ptr))
127 auto isGlobalLoad = [&](LoadInst &Load)->bool {
128 return Load.getPointerAddressSpace() == AMDGPUAS::GLOBAL_ADDRESS;
130 // We're tracking up to the Function boundaries
131 // We cannot go beyond because of FunctionPass restrictions
132 // Thus we can ensure that memory not clobbered for memory
133 // operations that live in kernel only.
134 bool NotClobbered = isKernelFunc && !isClobberedInFunction(&I);
135 Instruction *PtrI = dyn_cast<Instruction>(Ptr);
136 if (!PtrI && NotClobbered && isGlobalLoad(I)) {
137 if (isa<Argument>(Ptr) || isa<GlobalValue>(Ptr)) {
138 // Lookup for the existing GEP
139 if (noClobberClones.count(Ptr)) {
140 PtrI = noClobberClones[Ptr];
142 // Create GEP of the Value
143 Function *F = I.getParent()->getParent();
144 Value *Idx = Constant::getIntegerValue(
145 Type::getInt32Ty(Ptr->getContext()), APInt(64, 0));
146 // Insert GEP at the entry to make it dominate all uses
147 PtrI = GetElementPtrInst::Create(
148 Ptr->getType()->getPointerElementType(), Ptr,
149 ArrayRef<Value*>(Idx), Twine(""), F->getEntryBlock().getFirstNonPHI());
151 I.replaceUsesOfWith(Ptr, PtrI);
156 setUniformMetadata(PtrI);
158 setNoClobberMetadata(PtrI);
162 bool AMDGPUAnnotateUniformValues::doInitialization(Module &M) {
166 bool AMDGPUAnnotateUniformValues::runOnFunction(Function &F) {
170 DA = &getAnalysis<LegacyDivergenceAnalysis>();
171 MDR = &getAnalysis<MemoryDependenceWrapperPass>().getMemDep();
172 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
173 isKernelFunc = F.getCallingConv() == CallingConv::AMDGPU_KERNEL;
176 noClobberClones.clear();
181 llvm::createAMDGPUAnnotateUniformValues() {
182 return new AMDGPUAnnotateUniformValues();