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<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(
     86           cast<GEPOperator>(CE)->getSourceElementType(), CEOps[0],
     87           CEOps.slice(1)));
     88     }
     89     case Instruction::Add:
     90     case Instruction::Sub:
     91     case Instruction::Mul:
     92     case Instruction::UDiv:
     93     case Instruction::SDiv:
     94     case Instruction::FDiv:
     95     case Instruction::URem:
     96     case Instruction::SRem:
     97     case Instruction::FRem:
     98     case Instruction::Shl:
     99     case Instruction::LShr:
    100     case Instruction::AShr:
    101     case Instruction::And:
    102     case Instruction::Or:
    103     case Instruction::Xor:
    104       return dyn_cast<Instruction>(
    105                   Builder.CreateBinOp((Instruction::BinaryOps)OpCode,
    106                                       CE->getOperand(0), CE->getOperand(1),
    107                                       CE->getName()));
    108     case Instruction::Trunc:
    109     case Instruction::ZExt:
    110     case Instruction::SExt:
    111     case Instruction::FPToUI:
    112     case Instruction::FPToSI:
    113     case Instruction::UIToFP:
    114     case Instruction::SIToFP:
    115     case Instruction::FPTrunc:
    116     case Instruction::FPExt:
    117     case Instruction::PtrToInt:
    118     case Instruction::IntToPtr:
    119     case Instruction::BitCast:
    120       return dyn_cast<Instruction>(
    121                   Builder.CreateCast((Instruction::CastOps)OpCode,
    122                                      CE->getOperand(0), CE->getType(),
    123                                      CE->getName()));
    124     default:
    125       llvm_unreachable("Unhandled constant expression!\n");
    126   }
    127 }
    128 
    129 static bool replaceConstantExprOp(ConstantExpr *CE, Pass *P) {
    130   do {
    131     SmallVector<WeakVH,8> WUsers(CE->user_begin(), CE->user_end());
    132     std::sort(WUsers.begin(), WUsers.end());
    133     WUsers.erase(std::unique(WUsers.begin(), WUsers.end()), WUsers.end());
    134     while (!WUsers.empty())
    135       if (WeakVH WU = WUsers.pop_back_val()) {
    136         if (PHINode *PN = dyn_cast<PHINode>(WU)) {
    137           for (int I = 0, E = PN->getNumIncomingValues(); I < E; ++I)
    138             if (PN->getIncomingValue(I) == CE) {
    139               BasicBlock *PredBB = PN->getIncomingBlock(I);
    140               if (PredBB->getTerminator()->getNumSuccessors() > 1)
    141                 PredBB = SplitEdge(PredBB, PN->getParent());
    142               Instruction *InsertPos = PredBB->getTerminator();
    143               Instruction *NewInst = createReplacementInstr(CE, InsertPos);
    144               PN->setOperand(I, NewInst);
    145             }
    146         } else if (Instruction *Instr = dyn_cast<Instruction>(WU)) {
    147           Instruction *NewInst = createReplacementInstr(CE, Instr);
    148           Instr->replaceUsesOfWith(CE, NewInst);
    149         } else {
    150           ConstantExpr *CExpr = dyn_cast<ConstantExpr>(WU);
    151           if (!CExpr || !replaceConstantExprOp(CExpr, P))
    152             return false;
    153         }
    154       }
    155   } while (CE->hasNUsesOrMore(1)); // We need to check because a recursive
    156   // sibling may have used 'CE' when createReplacementInstr was called.
    157   CE->destroyConstant();
    158   return true;
    159 }
    160 
    161 static bool rewriteNonInstructionUses(GlobalVariable *GV, Pass *P) {
    162   SmallVector<WeakVH,8> WUsers;
    163   for (User *U : GV->users())
    164     if (!isa<Instruction>(U))
    165       WUsers.push_back(WeakVH(U));
    166   while (!WUsers.empty())
    167     if (WeakVH WU = WUsers.pop_back_val()) {
    168       ConstantExpr *CE = dyn_cast<ConstantExpr>(WU);
    169       if (!CE || !replaceConstantExprOp(CE, P))
    170         return false;
    171     }
    172   return true;
    173 }
    174 
    175 static bool isZeroLengthArray(Type *Ty) {
    176   ArrayType *AT = dyn_cast<ArrayType>(Ty);
    177   return AT && (AT->getNumElements() == 0);
    178 }
    179 
    180 bool XCoreLowerThreadLocal::lowerGlobal(GlobalVariable *GV) {
    181   Module *M = GV->getParent();
    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->getValueType());
    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     Value *Addr = Builder.CreateInBoundsGEP(NewGV->getValueType(), NewGV,
    213                                             {Builder.getInt64(0), ThreadID});
    214     U->replaceUsesOfWith(GV, Addr);
    215   }
    216 
    217   // Remove old global.
    218   NewGV->takeName(GV);
    219   GV->eraseFromParent();
    220   return true;
    221 }
    222 
    223 bool XCoreLowerThreadLocal::runOnModule(Module &M) {
    224   // Find thread local globals.
    225   bool MadeChange = false;
    226   SmallVector<GlobalVariable *, 16> ThreadLocalGlobals;
    227   for (GlobalVariable &GV : M.globals())
    228     if (GV.isThreadLocal())
    229       ThreadLocalGlobals.push_back(&GV);
    230   for (unsigned I = 0, E = ThreadLocalGlobals.size(); I != E; ++I) {
    231     MadeChange |= lowerGlobal(ThreadLocalGlobals[I]);
    232   }
    233   return MadeChange;
    234 }
    235