1 //===-- AMDGPUAnnotateUniformValues.cpp - ---------------------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// This pass adds amdgpu.uniform metadata to IR values so this information
12 /// can be used during instruction selection.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "AMDGPU.h"
17 #include "AMDGPUIntrinsicInfo.h"
18 #include "llvm/ADT/SetVector.h"
19 #include "llvm/Analysis/DivergenceAnalysis.h"
20 #include "llvm/Analysis/LoopInfo.h"
21 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
22 #include "llvm/IR/IRBuilder.h"
23 #include "llvm/IR/InstVisitor.h"
24 #include "llvm/Support/Debug.h"
25 #include "llvm/Support/raw_ostream.h"
26 
27 #define DEBUG_TYPE "amdgpu-annotate-uniform"
28 
29 using namespace llvm;
30 
31 namespace {
32 
33 class AMDGPUAnnotateUniformValues : public FunctionPass,
34                        public InstVisitor<AMDGPUAnnotateUniformValues> {
35   DivergenceAnalysis *DA;
36   MemoryDependenceResults *MDR;
37   LoopInfo *LI;
38   DenseMap<Value*, GetElementPtrInst*> noClobberClones;
39   bool isKernelFunc;
40   AMDGPUAS AMDGPUASI;
41 
42 public:
43   static char ID;
44   AMDGPUAnnotateUniformValues() :
45     FunctionPass(ID) { }
46   bool doInitialization(Module &M) override;
47   bool runOnFunction(Function &F) override;
48   StringRef getPassName() const override {
49     return "AMDGPU Annotate Uniform Values";
50   }
51   void getAnalysisUsage(AnalysisUsage &AU) const override {
52     AU.addRequired<DivergenceAnalysis>();
53     AU.addRequired<MemoryDependenceWrapperPass>();
54     AU.addRequired<LoopInfoWrapperPass>();
55     AU.setPreservesAll();
56  }
57 
58   void visitBranchInst(BranchInst &I);
59   void visitLoadInst(LoadInst &I);
60   bool isClobberedInFunction(LoadInst * Load);
61 };
62 
63 } // End anonymous namespace
64 
65 INITIALIZE_PASS_BEGIN(AMDGPUAnnotateUniformValues, DEBUG_TYPE,
66                       "Add AMDGPU uniform metadata", false, false)
67 INITIALIZE_PASS_DEPENDENCY(DivergenceAnalysis)
68 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
69 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
70 INITIALIZE_PASS_END(AMDGPUAnnotateUniformValues, DEBUG_TYPE,
71                     "Add AMDGPU uniform metadata", false, false)
72 
73 char AMDGPUAnnotateUniformValues::ID = 0;
74 
75 static void setUniformMetadata(Instruction *I) {
76   I->setMetadata("amdgpu.uniform", MDNode::get(I->getContext(), {}));
77 }
78 static void setNoClobberMetadata(Instruction *I) {
79   I->setMetadata("amdgpu.noclobber", MDNode::get(I->getContext(), {}));
80 }
81 
82 static void DFS(BasicBlock *Root, SetVector<BasicBlock*> & Set) {
83   for (auto I : predecessors(Root))
84     if (Set.insert(I))
85       DFS(I, Set);
86 }
87 
88 bool AMDGPUAnnotateUniformValues::isClobberedInFunction(LoadInst * Load) {
89   // 1. get Loop for the Load->getparent();
90   // 2. if it exists, collect all the BBs from the most outer
91   // loop and check for the writes. If NOT - start DFS over all preds.
92   // 3. Start DFS over all preds from the most outer loop header.
93   SetVector<BasicBlock *> Checklist;
94   BasicBlock *Start = Load->getParent();
95   Checklist.insert(Start);
96   const Value *Ptr = Load->getPointerOperand();
97   const Loop *L = LI->getLoopFor(Start);
98   if (L) {
99     const Loop *P = L;
100     do {
101       L = P;
102       P = P->getParentLoop();
103     } while (P);
104     Checklist.insert(L->block_begin(), L->block_end());
105     Start = L->getHeader();
106   }
107 
108   DFS(Start, Checklist);
109   for (auto &BB : Checklist) {
110     BasicBlock::iterator StartIt = (!L && (BB == Load->getParent())) ?
111       BasicBlock::iterator(Load) : BB->end();
112     auto Q = MDR->getPointerDependencyFrom(MemoryLocation(Ptr), true,
113                                            StartIt, BB, Load);
114     if (Q.isClobber() || Q.isUnknown())
115       return true;
116   }
117   return false;
118 }
119 
120 void AMDGPUAnnotateUniformValues::visitBranchInst(BranchInst &I) {
121   if (I.isUnconditional())
122     return;
123 
124   Value *Cond = I.getCondition();
125   if (!DA->isUniform(Cond))
126     return;
127 
128   setUniformMetadata(I.getParent()->getTerminator());
129 }
130 
131 void AMDGPUAnnotateUniformValues::visitLoadInst(LoadInst &I) {
132   Value *Ptr = I.getPointerOperand();
133   if (!DA->isUniform(Ptr))
134     return;
135   auto isGlobalLoad = [&](LoadInst &Load)->bool {
136     return Load.getPointerAddressSpace() == AMDGPUASI.GLOBAL_ADDRESS;
137   };
138   // We're tracking up to the Function boundaries
139   // We cannot go beyond because of FunctionPass restrictions
140   // Thus we can ensure that memory not clobbered for memory
141   // operations that live in kernel only.
142   bool NotClobbered = isKernelFunc &&   !isClobberedInFunction(&I);
143   Instruction *PtrI = dyn_cast<Instruction>(Ptr);
144   if (!PtrI && NotClobbered && isGlobalLoad(I)) {
145     if (isa<Argument>(Ptr) || isa<GlobalValue>(Ptr)) {
146       // Lookup for the existing GEP
147       if (noClobberClones.count(Ptr)) {
148         PtrI = noClobberClones[Ptr];
149       } else {
150         // Create GEP of the Value
151         Function *F = I.getParent()->getParent();
152         Value *Idx = Constant::getIntegerValue(
153           Type::getInt32Ty(Ptr->getContext()), APInt(64, 0));
154         // Insert GEP at the entry to make it dominate all uses
155         PtrI = GetElementPtrInst::Create(
156           Ptr->getType()->getPointerElementType(), Ptr,
157           ArrayRef<Value*>(Idx), Twine(""), F->getEntryBlock().getFirstNonPHI());
158       }
159       I.replaceUsesOfWith(Ptr, PtrI);
160     }
161   }
162 
163   if (PtrI) {
164     setUniformMetadata(PtrI);
165     if (NotClobbered)
166       setNoClobberMetadata(PtrI);
167   }
168 }
169 
170 bool AMDGPUAnnotateUniformValues::doInitialization(Module &M) {
171   AMDGPUASI = AMDGPU::getAMDGPUAS(M);
172   return false;
173 }
174 
175 bool AMDGPUAnnotateUniformValues::runOnFunction(Function &F) {
176   if (skipFunction(F))
177     return false;
178 
179   DA  = &getAnalysis<DivergenceAnalysis>();
180   MDR = &getAnalysis<MemoryDependenceWrapperPass>().getMemDep();
181   LI  = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
182   isKernelFunc = F.getCallingConv() == CallingConv::AMDGPU_KERNEL;
183 
184   visit(F);
185   noClobberClones.clear();
186   return true;
187 }
188 
189 FunctionPass *
190 llvm::createAMDGPUAnnotateUniformValues() {
191   return new AMDGPUAnnotateUniformValues();
192 }
193