Home | History | Annotate | Download | only in XCore 
      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/IRBuilder.h"
     21 #include "llvm/IR/Intrinsics.h"
     22 #include "llvm/IR/Module.h"
     23 #include "llvm/IR/NoFolder.h"
     24 #include "llvm/IR/ValueHandle.h"
     25 #include "llvm/Pass.h"
     26 #include "llvm/Support/CommandLine.h"
     27 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
     28 
     29 #define DEBUG_TYPE "xcore-lower-thread-local"
     30 
     31 using namespace llvm;
     32 
     33 static 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 
     38 namespace {
     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) override;
     52   };
     53 }
     54 
     55 char XCoreLowerThreadLocal::ID = 0;
     56 
     57 INITIALIZE_PASS(XCoreLowerThreadLocal, "xcore-lower-thread-local",
     58                 "Lower thread local variables", false, false)
     59 
     60 ModulePass *llvm::createXCoreLowerThreadLocalPass() {
     61   return new XCoreLowerThreadLocal();
     62 }
     63 
     64 static ArrayType *createLoweredType(Type *OriginalType) {
     65   return ArrayType::get(OriginalType, MaxThreads);
     66 }
     67 
     68 static Constant *
     69 createLoweredInitializer(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 
     77 static Instruction *
     78 createReplacementInstr(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 
    128 static bool replaceConstantExprOp(ConstantExpr *CE, Pass *P) {
    129   do {
    130     SmallVector<WeakVH,8> WUsers(CE->user_begin(), CE->user_end());
    131     std::sort(WUsers.begin(), WUsers.end());
    132     WUsers.erase(std::unique(WUsers.begin(), WUsers.end()), WUsers.end());
    133     while (!WUsers.empty())
    134       if (WeakVH WU = WUsers.pop_back_val()) {
    135         if (PHINode *PN = dyn_cast<PHINode>(WU)) {
    136           for (int I = 0, E = PN->getNumIncomingValues(); I < E; ++I)
    137             if (PN->getIncomingValue(I) == CE) {
    138               BasicBlock *PredBB = PN->getIncomingBlock(I);
    139               if (PredBB->getTerminator()->getNumSuccessors() > 1)
    140                 PredBB = SplitEdge(PredBB, PN->getParent(), P);
    141               Instruction *InsertPos = PredBB->getTerminator();
    142               Instruction *NewInst = createReplacementInstr(CE, InsertPos);
    143               PN->setOperand(I, NewInst);
    144             }
    145         } else if (Instruction *Instr = dyn_cast<Instruction>(WU)) {
    146           Instruction *NewInst = createReplacementInstr(CE, Instr);
    147           Instr->replaceUsesOfWith(CE, NewInst);
    148         } else {
    149           ConstantExpr *CExpr = dyn_cast<ConstantExpr>(WU);
    150           if (!CExpr || !replaceConstantExprOp(CExpr, P))
    151             return false;
    152         }
    153       }
    154   } while (CE->hasNUsesOrMore(1)); // We need to check because a recursive
    155   // sibling may have used 'CE' when createReplacementInstr was called.
    156   CE->destroyConstant();
    157   return true;
    158 }
    159 
    160 static bool rewriteNonInstructionUses(GlobalVariable *GV, Pass *P) {
    161   SmallVector<WeakVH,8> WUsers;
    162   for (User *U : GV->users())
    163     if (!isa<Instruction>(U))
    164       WUsers.push_back(WeakVH(U));
    165   while (!WUsers.empty())
    166     if (WeakVH WU = WUsers.pop_back_val()) {
    167       ConstantExpr *CE = dyn_cast<ConstantExpr>(WU);
    168       if (!CE || !replaceConstantExprOp(CE, P))
    169         return false;
    170     }
    171   return true;
    172 }
    173 
    174 static bool isZeroLengthArray(Type *Ty) {
    175   ArrayType *AT = dyn_cast<ArrayType>(Ty);
    176   return AT && (AT->getNumElements() == 0);
    177 }
    178 
    179 bool XCoreLowerThreadLocal::lowerGlobal(GlobalVariable *GV) {
    180   Module *M = GV->getParent();
    181   LLVMContext &Ctx = M->getContext();
    182   if (!GV->isThreadLocal())
    183     return false;
    184 
    185   // Skip globals that we can't lower and leave it for the backend to error.
    186   if (!rewriteNonInstructionUses(GV, this) ||
    187       !GV->getType()->isSized() || isZeroLengthArray(GV->getType()))
    188     return false;
    189 
    190   // Create replacement global.
    191   ArrayType *NewType = createLoweredType(GV->getType()->getElementType());
    192   Constant *NewInitializer = nullptr;
    193   if (GV->hasInitializer())
    194     NewInitializer = createLoweredInitializer(NewType,
    195                                               GV->getInitializer());
    196   GlobalVariable *NewGV =
    197     new GlobalVariable(*M, NewType, GV->isConstant(), GV->getLinkage(),
    198                        NewInitializer, "", nullptr,
    199                        GlobalVariable::NotThreadLocal,
    200                        GV->getType()->getAddressSpace(),
    201                        GV->isExternallyInitialized());
    202 
    203   // Update uses.
    204   SmallVector<User *, 16> Users(GV->user_begin(), GV->user_end());
    205   for (unsigned I = 0, E = Users.size(); I != E; ++I) {
    206     User *U = Users[I];
    207     Instruction *Inst = cast<Instruction>(U);
    208     IRBuilder<> Builder(Inst);
    209     Function *GetID = Intrinsic::getDeclaration(GV->getParent(),
    210                                                 Intrinsic::xcore_getid);
    211     Value *ThreadID = Builder.CreateCall(GetID);
    212     SmallVector<Value *, 2> Indices;
    213     Indices.push_back(Constant::getNullValue(Type::getInt64Ty(Ctx)));
    214     Indices.push_back(ThreadID);
    215     Value *Addr = Builder.CreateInBoundsGEP(NewGV, Indices);
    216     U->replaceUsesOfWith(GV, Addr);
    217   }
    218 
    219   // Remove old global.
    220   NewGV->takeName(GV);
    221   GV->eraseFromParent();
    222   return true;
    223 }
    224 
    225 bool XCoreLowerThreadLocal::runOnModule(Module &M) {
    226   // Find thread local globals.
    227   bool MadeChange = false;
    228   SmallVector<GlobalVariable *, 16> ThreadLocalGlobals;
    229   for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
    230        GVI != E; ++GVI) {
    231     GlobalVariable *GV = GVI;
    232     if (GV->isThreadLocal())
    233       ThreadLocalGlobals.push_back(GV);
    234   }
    235   for (unsigned I = 0, E = ThreadLocalGlobals.size(); I != E; ++I) {
    236     MadeChange |= lowerGlobal(ThreadLocalGlobals[I]);
    237   }
    238   return MadeChange;
    239 }
    240