LLVM  6.0.0svn
AMDGPUAnnotateUniformValues.cpp
Go to the documentation of this file.
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"
20 #include "llvm/Analysis/LoopInfo.h"
22 #include "llvm/IR/IRBuilder.h"
23 #include "llvm/IR/InstVisitor.h"
24 #include "llvm/Support/Debug.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> {
37  LoopInfo *LI;
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 {
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)
70 INITIALIZE_PASS_END(AMDGPUAnnotateUniformValues, DEBUG_TYPE,
72 
73 char AMDGPUAnnotateUniformValues::ID = 0;
74 
76  I->setMetadata("amdgpu.uniform", MDNode::get(I->getContext(), {}));
77 }
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 
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();
153  Type::getInt32Ty(Ptr->getContext()), APInt(64, 0));
154  // Insert GEP at the entry to make it dominate all uses
156  Ptr->getType()->getPointerElementType(), Ptr,
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 
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 *
191  return new AMDGPUAnnotateUniformValues();
192 }
Provides a lazy, caching interface for making common memory aliasing information queries, backed by LLVM&#39;s alias analysis passes.
Base class for instruction visitors.
Definition: InstVisitor.h:81
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
A Module instance is used to store all the information related to an LLVM module. ...
Definition: Module.h:63
static GetElementPtrInst * Create(Type *PointeeType, Value *Ptr, ArrayRef< Value *> IdxList, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Definition: Instructions.h:863
AMDGPUAS getAMDGPUAS(const Module &M)
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:728
F(f)
An instruction for reading from memory.
Definition: Instructions.h:164
Value * getCondition() const
AnalysisUsage & addRequired()
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
Type * getPointerElementType() const
Definition: Type.h:373
Twine - A lightweight data structure for efficiently representing the concatenation of temporary valu...
Definition: Twine.h:81
OpenCL uses address spaces to differentiate between various memory regions on the hardware...
Definition: AMDGPU.h:214
T uniform(GenT &Gen, T Min, T Max)
Return a uniformly distributed random value between Min and Max.
Definition: Random.h:22
BlockT * getHeader() const
Definition: LoopInfo.h:100
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
bool insert(const value_type &X)
Insert a new element into the SetVector.
Definition: SetVector.h:142
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:33
void replaceUsesOfWith(Value *From, Value *To)
Replace uses of one Value with another.
Definition: User.cpp:22
const BasicBlock & getEntryBlock() const
Definition: Function.h:572
static MDTuple * get(LLVMContext &Context, ArrayRef< Metadata *> MDs)
Definition: Metadata.h:1164
static bool runOnFunction(Function &F, bool PostInlining)
#define P(N)
const Instruction * getFirstNonPHI() const
Returns a pointer to the first instruction in this block that is not a PHINode instruction.
Definition: BasicBlock.cpp:171
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
Conditional or Unconditional Branch instruction.
Represent the analysis usage information of a pass.
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
Calling convention for AMDGPU code object kernels.
Definition: CallingConv.h:201
Value * getPointerOperand()
Definition: Instructions.h:270
static void setNoClobberMetadata(Instruction *I)
#define DEBUG_TYPE
A wrapper analysis pass for the legacy pass manager that exposes a MemoryDepnedenceResults instance...
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
Add AMDGPU uniform metadata
void setMetadata(unsigned KindID, MDNode *Node)
Set the metadata of the specified kind to the specified node.
Definition: Metadata.cpp:1214
FunctionPass * createAMDGPUAnnotateUniformValues()
static Constant * getIntegerValue(Type *Ty, const APInt &V)
Return the value for an integer or pointer constant, or a vector thereof, with the given scalar value...
Definition: Constants.cpp:244
Representation for a specific memory location.
Iterator for intrusive lists based on ilist_node.
CallingConv::ID getCallingConv() const
getCallingConv()/setCallingConv(CC) - These method get and set the calling convention of this functio...
Definition: Function.h:194
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:110
INITIALIZE_PASS_BEGIN(AMDGPUAnnotateUniformValues, DEBUG_TYPE, "Add AMDGPU uniform metadata", false, false) INITIALIZE_PASS_END(AMDGPUAnnotateUniformValues
Class for arbitrary precision integers.
Definition: APInt.h:69
Interface for the AMDGPU Implementation of the Intrinsic Info class.
void setPreservesAll()
Set by analyses that do not transform their input at all.
InstListType::iterator iterator
Instruction iterators...
Definition: BasicBlock.h:91
static void DFS(BasicBlock *Root, SetVector< BasicBlock *> &Set)
LoopT * getParentLoop() const
Definition: LoopInfo.h:101
static IntegerType * getInt32Ty(LLVMContext &C)
Definition: Type.cpp:176
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:439
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
#define I(x, y, z)
Definition: MD5.cpp:58
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
block_iterator block_end() const
Definition: LoopInfo.h:155
bool isUnconditional() const
unsigned getPointerAddressSpace() const
Returns the address space of the pointer operand.
Definition: Instructions.h:276
LLVM Value Representation.
Definition: Value.h:73
A vector that has set insertion semantics.
Definition: SetVector.h:41
The legacy pass manager&#39;s analysis pass to compute loop information.
Definition: LoopInfo.h:958
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:49
const TerminatorInst * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:120
block_iterator block_begin() const
Definition: LoopInfo.h:154
static void setUniformMetadata(Instruction *I)
const BasicBlock * getParent() const
Definition: Instruction.h:66