XCoreLowerThreadLocal.cpp revision 296417
1226584Sdim//===-- XCoreLowerThreadLocal - Lower thread local variables --------------===// 2226584Sdim// 3226584Sdim// The LLVM Compiler Infrastructure 4226584Sdim// 5226584Sdim// This file is distributed under the University of Illinois Open Source 6226584Sdim// License. See LICENSE.TXT for details. 7226584Sdim// 8226584Sdim//===----------------------------------------------------------------------===// 9226584Sdim/// 10226584Sdim/// \file 11226584Sdim/// \brief This file contains a pass that lowers thread local variables on the 12226584Sdim/// XCore. 13226584Sdim/// 14226584Sdim//===----------------------------------------------------------------------===// 15226584Sdim 16226584Sdim#include "XCore.h" 17226584Sdim#include "llvm/IR/Constants.h" 18249423Sdim#include "llvm/IR/DerivedTypes.h" 19226584Sdim#include "llvm/IR/GlobalVariable.h" 20226584Sdim#include "llvm/IR/IRBuilder.h" 21249423Sdim#include "llvm/IR/Intrinsics.h" 22234353Sdim#include "llvm/IR/Module.h" 23249423Sdim#include "llvm/IR/NoFolder.h" 24226584Sdim#include "llvm/IR/ValueHandle.h" 25249423Sdim#include "llvm/Pass.h" 26226584Sdim#include "llvm/Support/CommandLine.h" 27234353Sdim#include "llvm/Transforms/Utils/BasicBlockUtils.h" 28226584Sdim 29226584Sdim#define DEBUG_TYPE "xcore-lower-thread-local" 30226584Sdim 31234353Sdimusing namespace llvm; 32263508Sdim 33243830Sdimstatic cl::opt<unsigned> MaxThreads( 34234353Sdim "xcore-max-threads", cl::Optional, 35234982Sdim cl::desc("Maximum number of threads (for emulation thread-local storage)"), 36263508Sdim cl::Hidden, cl::value_desc("number"), cl::init(8)); 37263508Sdim 38263508Sdimnamespace { 39263508Sdim /// Lowers thread local variables on the XCore. Each thread local variable is 40263508Sdim /// expanded to an array of n elements indexed by the thread ID where n is the 41263508Sdim /// fixed number hardware threads supported by the device. 42251662Sdim struct XCoreLowerThreadLocal : public ModulePass { 43243830Sdim static char ID; 44243830Sdim 45263508Sdim XCoreLowerThreadLocal() : ModulePass(ID) { 46251662Sdim initializeXCoreLowerThreadLocalPass(*PassRegistry::getPassRegistry()); 47251662Sdim } 48251662Sdim 49234982Sdim bool lowerGlobal(GlobalVariable *GV); 50234982Sdim 51234982Sdim bool runOnModule(Module &M) override; 52234982Sdim }; 53243830Sdim} 54243830Sdim 55243830Sdimchar XCoreLowerThreadLocal::ID = 0; 56243830Sdim 57263508SdimINITIALIZE_PASS(XCoreLowerThreadLocal, "xcore-lower-thread-local", 58263508Sdim "Lower thread local variables", false, false) 59263508Sdim 60263508SdimModulePass *llvm::createXCoreLowerThreadLocalPass() { 61226584Sdim return new XCoreLowerThreadLocal(); 62226584Sdim} 63226584Sdim 64226584Sdimstatic ArrayType *createLoweredType(Type *OriginalType) { 65226584Sdim return ArrayType::get(OriginalType, MaxThreads); 66226584Sdim} 67226584Sdim 68251662Sdimstatic Constant * 69251662SdimcreateLoweredInitializer(ArrayType *NewType, Constant *OriginalInitializer) { 70243830Sdim SmallVector<Constant *, 8> Elements(MaxThreads); 71263508Sdim for (unsigned i = 0; i != MaxThreads; ++i) { 72263508Sdim Elements[i] = OriginalInitializer; 73226584Sdim } 74226584Sdim return ConstantArray::get(NewType, Elements); 75226584Sdim} 76226584Sdim 77226584Sdimstatic Instruction * 78226584SdimcreateReplacementInstr(ConstantExpr *CE, Instruction *Instr) { 79226584Sdim IRBuilder<true,NoFolder> Builder(Instr); 80226584Sdim unsigned OpCode = CE->getOpcode(); 81226584Sdim switch (OpCode) { 82226584Sdim case Instruction::GetElementPtr: { 83226584Sdim SmallVector<Value *,4> CEOpVec(CE->op_begin(), CE->op_end()); 84226584Sdim ArrayRef<Value *> CEOps(CEOpVec); 85226584Sdim return dyn_cast<Instruction>(Builder.CreateInBoundsGEP( 86226584Sdim cast<GEPOperator>(CE)->getSourceElementType(), CEOps[0], 87226584Sdim CEOps.slice(1))); 88226584Sdim } 89226584Sdim case Instruction::Add: 90226584Sdim case Instruction::Sub: 91226584Sdim case Instruction::Mul: 92226584Sdim case Instruction::UDiv: 93226584Sdim case Instruction::SDiv: 94226584Sdim case Instruction::FDiv: 95226584Sdim case Instruction::URem: 96226584Sdim case Instruction::SRem: 97226584Sdim case Instruction::FRem: 98226584Sdim case Instruction::Shl: 99226584Sdim case Instruction::LShr: 100226584Sdim case Instruction::AShr: 101226584Sdim case Instruction::And: 102226584Sdim case Instruction::Or: 103226584Sdim case Instruction::Xor: 104226584Sdim return dyn_cast<Instruction>( 105226584Sdim Builder.CreateBinOp((Instruction::BinaryOps)OpCode, 106226584Sdim CE->getOperand(0), CE->getOperand(1), 107226584Sdim CE->getName())); 108226584Sdim case Instruction::Trunc: 109226584Sdim case Instruction::ZExt: 110226584Sdim case Instruction::SExt: 111226584Sdim case Instruction::FPToUI: 112226584Sdim case Instruction::FPToSI: 113226584Sdim case Instruction::UIToFP: 114226584Sdim case Instruction::SIToFP: 115226584Sdim case Instruction::FPTrunc: 116226584Sdim case Instruction::FPExt: 117226584Sdim case Instruction::PtrToInt: 118226584Sdim case Instruction::IntToPtr: 119226584Sdim case Instruction::BitCast: 120226584Sdim return dyn_cast<Instruction>( 121226584Sdim Builder.CreateCast((Instruction::CastOps)OpCode, 122226584Sdim CE->getOperand(0), CE->getType(), 123226584Sdim CE->getName())); 124243830Sdim default: 125243830Sdim llvm_unreachable("Unhandled constant expression!\n"); 126243830Sdim } 127243830Sdim} 128226584Sdim 129226584Sdimstatic bool replaceConstantExprOp(ConstantExpr *CE, Pass *P) { 130226584Sdim do { 131226584Sdim SmallVector<WeakVH,8> WUsers(CE->user_begin(), CE->user_end()); 132226584Sdim std::sort(WUsers.begin(), WUsers.end()); 133226584Sdim WUsers.erase(std::unique(WUsers.begin(), WUsers.end()), WUsers.end()); 134226584Sdim while (!WUsers.empty()) 135226584Sdim if (WeakVH WU = WUsers.pop_back_val()) { 136226584Sdim if (PHINode *PN = dyn_cast<PHINode>(WU)) { 137226584Sdim for (int I = 0, E = PN->getNumIncomingValues(); I < E; ++I) 138226584Sdim if (PN->getIncomingValue(I) == CE) { 139243830Sdim BasicBlock *PredBB = PN->getIncomingBlock(I); 140243830Sdim if (PredBB->getTerminator()->getNumSuccessors() > 1) 141243830Sdim PredBB = SplitEdge(PredBB, PN->getParent()); 142243830Sdim Instruction *InsertPos = PredBB->getTerminator(); 143243830Sdim Instruction *NewInst = createReplacementInstr(CE, InsertPos); 144243830Sdim PN->setOperand(I, NewInst); 145243830Sdim } 146243830Sdim } else if (Instruction *Instr = dyn_cast<Instruction>(WU)) { 147234353Sdim Instruction *NewInst = createReplacementInstr(CE, Instr); 148226584Sdim Instr->replaceUsesOfWith(CE, NewInst); 149226584Sdim } else { 150226584Sdim ConstantExpr *CExpr = dyn_cast<ConstantExpr>(WU); 151226584Sdim if (!CExpr || !replaceConstantExprOp(CExpr, P)) 152226584Sdim return false; 153226584Sdim } 154226584Sdim } 155226584Sdim } while (CE->hasNUsesOrMore(1)); // We need to check because a recursive 156226584Sdim // sibling may have used 'CE' when createReplacementInstr was called. 157234353Sdim CE->destroyConstant(); 158234353Sdim return true; 159226584Sdim} 160226584Sdim 161226584Sdimstatic bool rewriteNonInstructionUses(GlobalVariable *GV, Pass *P) { 162226584Sdim SmallVector<WeakVH,8> WUsers; 163226584Sdim for (User *U : GV->users()) 164226584Sdim if (!isa<Instruction>(U)) 165226584Sdim WUsers.push_back(WeakVH(U)); 166226584Sdim while (!WUsers.empty()) 167226584Sdim if (WeakVH WU = WUsers.pop_back_val()) { 168226584Sdim ConstantExpr *CE = dyn_cast<ConstantExpr>(WU); 169226584Sdim if (!CE || !replaceConstantExprOp(CE, P)) 170226584Sdim return false; 171226584Sdim } 172226584Sdim return true; 173226584Sdim} 174226584Sdim 175226584Sdimstatic bool isZeroLengthArray(Type *Ty) { 176226584Sdim ArrayType *AT = dyn_cast<ArrayType>(Ty); 177226584Sdim return AT && (AT->getNumElements() == 0); 178226584Sdim} 179226584Sdim 180226584Sdimbool XCoreLowerThreadLocal::lowerGlobal(GlobalVariable *GV) { 181226584Sdim Module *M = GV->getParent(); 182243830Sdim LLVMContext &Ctx = M->getContext(); 183243830Sdim if (!GV->isThreadLocal()) 184243830Sdim return false; 185243830Sdim 186226584Sdim // Skip globals that we can't lower and leave it for the backend to error. 187226584Sdim if (!rewriteNonInstructionUses(GV, this) || 188226584Sdim !GV->getType()->isSized() || isZeroLengthArray(GV->getType())) 189226584Sdim return false; 190226584Sdim 191226584Sdim // Create replacement global. 192226584Sdim ArrayType *NewType = createLoweredType(GV->getType()->getElementType()); 193226584Sdim Constant *NewInitializer = nullptr; 194226584Sdim if (GV->hasInitializer()) 195226584Sdim NewInitializer = createLoweredInitializer(NewType, 196226584Sdim GV->getInitializer()); 197226584Sdim GlobalVariable *NewGV = 198226584Sdim new GlobalVariable(*M, NewType, GV->isConstant(), GV->getLinkage(), 199226584Sdim NewInitializer, "", nullptr, 200226584Sdim GlobalVariable::NotThreadLocal, 201226584Sdim GV->getType()->getAddressSpace(), 202243830Sdim GV->isExternallyInitialized()); 203263508Sdim 204263508Sdim // Update uses. 205243830Sdim SmallVector<User *, 16> Users(GV->user_begin(), GV->user_end()); 206226584Sdim for (unsigned I = 0, E = Users.size(); I != E; ++I) { 207226584Sdim User *U = Users[I]; 208226584Sdim Instruction *Inst = cast<Instruction>(U); 209226584Sdim IRBuilder<> Builder(Inst); 210226584Sdim Function *GetID = Intrinsic::getDeclaration(GV->getParent(), 211226584Sdim Intrinsic::xcore_getid); 212226584Sdim Value *ThreadID = Builder.CreateCall(GetID, {}); 213226584Sdim SmallVector<Value *, 2> Indices; 214226584Sdim Indices.push_back(Constant::getNullValue(Type::getInt64Ty(Ctx))); 215226584Sdim Indices.push_back(ThreadID); 216226584Sdim Value *Addr = 217226584Sdim Builder.CreateInBoundsGEP(NewGV->getValueType(), NewGV, Indices); 218226584Sdim U->replaceUsesOfWith(GV, Addr); 219226584Sdim } 220226584Sdim 221226584Sdim // Remove old global. 222226584Sdim NewGV->takeName(GV); 223226584Sdim GV->eraseFromParent(); 224263508Sdim return true; 225263508Sdim} 226251662Sdim 227263508Sdimbool XCoreLowerThreadLocal::runOnModule(Module &M) { 228251662Sdim // Find thread local globals. 229251662Sdim bool MadeChange = false; 230234353Sdim SmallVector<GlobalVariable *, 16> ThreadLocalGlobals; 231234353Sdim for (GlobalVariable &GV : M.globals()) 232234982Sdim if (GV.isThreadLocal()) 233263508Sdim ThreadLocalGlobals.push_back(&GV); 234234982Sdim for (unsigned I = 0, E = ThreadLocalGlobals.size(); I != E; ++I) { 235263508Sdim MadeChange |= lowerGlobal(ThreadLocalGlobals[I]); 236249423Sdim } 237249423Sdim return MadeChange; 238249423Sdim} 239249423Sdim