1//===-- XCoreLowerThreadLocal - Lower thread local variables --------------===//
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/// \brief This file contains a pass that lowers thread local variables on the
12///        XCore.
13///
14//===----------------------------------------------------------------------===//
15
16#include "XCore.h"
17#include "llvm/IR/Constants.h"
18#include "llvm/IR/DerivedTypes.h"
19#include "llvm/IR/GlobalVariable.h"
20#include "llvm/IR/Intrinsics.h"
21#include "llvm/IR/IRBuilder.h"
22#include "llvm/IR/Module.h"
23#include "llvm/Pass.h"
24#include "llvm/Support/CommandLine.h"
25#include "llvm/Support/NoFolder.h"
26#include "llvm/Support/ValueHandle.h"
27#include "llvm/Transforms/Utils/BasicBlockUtils.h"
28
29#define DEBUG_TYPE "xcore-lower-thread-local"
30
31using namespace llvm;
32
33static cl::opt<unsigned> MaxThreads(
34  "xcore-max-threads", cl::Optional,
35  cl::desc("Maximum number of threads (for emulation thread-local storage)"),
36  cl::Hidden, cl::value_desc("number"), cl::init(8));
37
38namespace {
39  /// Lowers thread local variables on the XCore. Each thread local variable is
40  /// expanded to an array of n elements indexed by the thread ID where n is the
41  /// fixed number hardware threads supported by the device.
42  struct XCoreLowerThreadLocal : public ModulePass {
43    static char ID;
44
45    XCoreLowerThreadLocal() : ModulePass(ID) {
46      initializeXCoreLowerThreadLocalPass(*PassRegistry::getPassRegistry());
47    }
48
49    bool lowerGlobal(GlobalVariable *GV);
50
51    bool runOnModule(Module &M);
52  };
53}
54
55char XCoreLowerThreadLocal::ID = 0;
56
57INITIALIZE_PASS(XCoreLowerThreadLocal, "xcore-lower-thread-local",
58                "Lower thread local variables", false, false)
59
60ModulePass *llvm::createXCoreLowerThreadLocalPass() {
61  return new XCoreLowerThreadLocal();
62}
63
64static ArrayType *createLoweredType(Type *OriginalType) {
65  return ArrayType::get(OriginalType, MaxThreads);
66}
67
68static Constant *
69createLoweredInitializer(ArrayType *NewType, Constant *OriginalInitializer) {
70  SmallVector<Constant *, 8> Elements(MaxThreads);
71  for (unsigned i = 0; i != MaxThreads; ++i) {
72    Elements[i] = OriginalInitializer;
73  }
74  return ConstantArray::get(NewType, Elements);
75}
76
77static Instruction *
78createReplacementInstr(ConstantExpr *CE, Instruction *Instr) {
79  IRBuilder<true,NoFolder> Builder(Instr);
80  unsigned OpCode = CE->getOpcode();
81  switch (OpCode) {
82    case Instruction::GetElementPtr: {
83      SmallVector<Value *,4> CEOpVec(CE->op_begin(), CE->op_end());
84      ArrayRef<Value *> CEOps(CEOpVec);
85      return dyn_cast<Instruction>(Builder.CreateInBoundsGEP(CEOps[0],
86                                                             CEOps.slice(1)));
87    }
88    case Instruction::Add:
89    case Instruction::Sub:
90    case Instruction::Mul:
91    case Instruction::UDiv:
92    case Instruction::SDiv:
93    case Instruction::FDiv:
94    case Instruction::URem:
95    case Instruction::SRem:
96    case Instruction::FRem:
97    case Instruction::Shl:
98    case Instruction::LShr:
99    case Instruction::AShr:
100    case Instruction::And:
101    case Instruction::Or:
102    case Instruction::Xor:
103      return dyn_cast<Instruction>(
104                  Builder.CreateBinOp((Instruction::BinaryOps)OpCode,
105                                      CE->getOperand(0), CE->getOperand(1),
106                                      CE->getName()));
107    case Instruction::Trunc:
108    case Instruction::ZExt:
109    case Instruction::SExt:
110    case Instruction::FPToUI:
111    case Instruction::FPToSI:
112    case Instruction::UIToFP:
113    case Instruction::SIToFP:
114    case Instruction::FPTrunc:
115    case Instruction::FPExt:
116    case Instruction::PtrToInt:
117    case Instruction::IntToPtr:
118    case Instruction::BitCast:
119      return dyn_cast<Instruction>(
120                  Builder.CreateCast((Instruction::CastOps)OpCode,
121                                     CE->getOperand(0), CE->getType(),
122                                     CE->getName()));
123    default:
124      llvm_unreachable("Unhandled constant expression!\n");
125  }
126}
127
128static bool replaceConstantExprOp(ConstantExpr *CE, Pass *P) {
129  do {
130    SmallVector<WeakVH,8> WUsers;
131    for (Value::use_iterator I = CE->use_begin(), E = CE->use_end();
132         I != E; ++I)
133      WUsers.push_back(WeakVH(*I));
134    std::sort(WUsers.begin(), WUsers.end());
135    WUsers.erase(std::unique(WUsers.begin(), WUsers.end()), WUsers.end());
136    while (!WUsers.empty())
137      if (WeakVH WU = WUsers.pop_back_val()) {
138        if (PHINode *PN = dyn_cast<PHINode>(WU)) {
139          for (int I = 0, E = PN->getNumIncomingValues(); I < E; ++I)
140            if (PN->getIncomingValue(I) == CE) {
141              BasicBlock *PredBB = PN->getIncomingBlock(I);
142              if (PredBB->getTerminator()->getNumSuccessors() > 1)
143                PredBB = SplitEdge(PredBB, PN->getParent(), P);
144              Instruction *InsertPos = PredBB->getTerminator();
145              Instruction *NewInst = createReplacementInstr(CE, InsertPos);
146              PN->setOperand(I, NewInst);
147            }
148        } else if (Instruction *Instr = dyn_cast<Instruction>(WU)) {
149          Instruction *NewInst = createReplacementInstr(CE, Instr);
150          Instr->replaceUsesOfWith(CE, NewInst);
151        } else {
152          ConstantExpr *CExpr = dyn_cast<ConstantExpr>(WU);
153          if (!CExpr || !replaceConstantExprOp(CExpr, P))
154            return false;
155        }
156      }
157  } while (CE->hasNUsesOrMore(1)); // We need to check becasue a recursive
158  // sibbling may have used 'CE' when createReplacementInstr was called.
159  CE->destroyConstant();
160  return true;
161}
162
163static bool rewriteNonInstructionUses(GlobalVariable *GV, Pass *P) {
164  SmallVector<WeakVH,8> WUsers;
165  for (Value::use_iterator I = GV->use_begin(), E = GV->use_end(); I != E; ++I)
166    if (!isa<Instruction>(*I))
167      WUsers.push_back(WeakVH(*I));
168  while (!WUsers.empty())
169    if (WeakVH WU = WUsers.pop_back_val()) {
170      ConstantExpr *CE = dyn_cast<ConstantExpr>(WU);
171      if (!CE || !replaceConstantExprOp(CE, P))
172        return false;
173    }
174  return true;
175}
176
177static bool isZeroLengthArray(Type *Ty) {
178  ArrayType *AT = dyn_cast<ArrayType>(Ty);
179  return AT && (AT->getNumElements() == 0);
180}
181
182bool XCoreLowerThreadLocal::lowerGlobal(GlobalVariable *GV) {
183  Module *M = GV->getParent();
184  LLVMContext &Ctx = M->getContext();
185  if (!GV->isThreadLocal())
186    return false;
187
188  // Skip globals that we can't lower and leave it for the backend to error.
189  if (!rewriteNonInstructionUses(GV, this) ||
190      !GV->getType()->isSized() || isZeroLengthArray(GV->getType()))
191    return false;
192
193  // Create replacement global.
194  ArrayType *NewType = createLoweredType(GV->getType()->getElementType());
195  Constant *NewInitializer = 0;
196  if (GV->hasInitializer())
197    NewInitializer = createLoweredInitializer(NewType,
198                                              GV->getInitializer());
199  GlobalVariable *NewGV =
200    new GlobalVariable(*M, NewType, GV->isConstant(), GV->getLinkage(),
201                       NewInitializer, "", 0, GlobalVariable::NotThreadLocal,
202                       GV->getType()->getAddressSpace(),
203                       GV->isExternallyInitialized());
204
205  // Update uses.
206  SmallVector<User *, 16> Users(GV->use_begin(), GV->use_end());
207  for (unsigned I = 0, E = Users.size(); I != E; ++I) {
208    User *U = Users[I];
209    Instruction *Inst = cast<Instruction>(U);
210    IRBuilder<> Builder(Inst);
211    Function *GetID = Intrinsic::getDeclaration(GV->getParent(),
212                                                Intrinsic::xcore_getid);
213    Value *ThreadID = Builder.CreateCall(GetID);
214    SmallVector<Value *, 2> Indices;
215    Indices.push_back(Constant::getNullValue(Type::getInt64Ty(Ctx)));
216    Indices.push_back(ThreadID);
217    Value *Addr = Builder.CreateInBoundsGEP(NewGV, Indices);
218    U->replaceUsesOfWith(GV, Addr);
219  }
220
221  // Remove old global.
222  NewGV->takeName(GV);
223  GV->eraseFromParent();
224  return true;
225}
226
227bool XCoreLowerThreadLocal::runOnModule(Module &M) {
228  // Find thread local globals.
229  bool MadeChange = false;
230  SmallVector<GlobalVariable *, 16> ThreadLocalGlobals;
231  for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
232       GVI != E; ++GVI) {
233    GlobalVariable *GV = GVI;
234    if (GV->isThreadLocal())
235      ThreadLocalGlobals.push_back(GV);
236  }
237  for (unsigned I = 0, E = ThreadLocalGlobals.size(); I != E; ++I) {
238    MadeChange |= lowerGlobal(ThreadLocalGlobals[I]);
239  }
240  return MadeChange;
241}
242