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