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