1 //===-- XCoreLowerThreadLocal - Lower thread local variables --------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 ///
9 /// \file
10 /// This file contains a pass that lowers thread local variables on the
11 ///        XCore.
12 ///
13 //===----------------------------------------------------------------------===//
14 
15 #include "XCore.h"
16 #include "llvm/IR/Constants.h"
17 #include "llvm/IR/DerivedTypes.h"
18 #include "llvm/IR/GlobalVariable.h"
19 #include "llvm/IR/IRBuilder.h"
20 #include "llvm/IR/Intrinsics.h"
21 #include "llvm/IR/Module.h"
22 #include "llvm/IR/NoFolder.h"
23 #include "llvm/IR/ValueHandle.h"
24 #include "llvm/Pass.h"
25 #include "llvm/Support/CommandLine.h"
26 #include "llvm/Transforms/Utils/BasicBlockUtils.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) override;
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<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(
85           cast<GEPOperator>(CE)->getSourceElementType(), 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<WeakTrackingVH, 8> WUsers(CE->user_begin(), CE->user_end());
131     llvm::sort(WUsers);
132     WUsers.erase(std::unique(WUsers.begin(), WUsers.end()), WUsers.end());
133     while (!WUsers.empty())
134       if (WeakTrackingVH 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());
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<WeakTrackingVH, 8> WUsers;
162   for (User *U : GV->users())
163     if (!isa<Instruction>(U))
164       WUsers.push_back(WeakTrackingVH(U));
165   while (!WUsers.empty())
166     if (WeakTrackingVH 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   if (!GV->isThreadLocal())
182     return false;
183 
184   // Skip globals that we can't lower and leave it for the backend to error.
185   if (!rewriteNonInstructionUses(GV, this) ||
186       !GV->getType()->isSized() || isZeroLengthArray(GV->getType()))
187     return false;
188 
189   // Create replacement global.
190   ArrayType *NewType = createLoweredType(GV->getValueType());
191   Constant *NewInitializer = nullptr;
192   if (GV->hasInitializer())
193     NewInitializer = createLoweredInitializer(NewType,
194                                               GV->getInitializer());
195   GlobalVariable *NewGV =
196     new GlobalVariable(*M, NewType, GV->isConstant(), GV->getLinkage(),
197                        NewInitializer, "", nullptr,
198                        GlobalVariable::NotThreadLocal,
199                        GV->getType()->getAddressSpace(),
200                        GV->isExternallyInitialized());
201 
202   // Update uses.
203   SmallVector<User *, 16> Users(GV->user_begin(), GV->user_end());
204   for (unsigned I = 0, E = Users.size(); I != E; ++I) {
205     User *U = Users[I];
206     Instruction *Inst = cast<Instruction>(U);
207     IRBuilder<> Builder(Inst);
208     Function *GetID = Intrinsic::getDeclaration(GV->getParent(),
209                                                 Intrinsic::xcore_getid);
210     Value *ThreadID = Builder.CreateCall(GetID, {});
211     Value *Addr = Builder.CreateInBoundsGEP(NewGV->getValueType(), NewGV,
212                                             {Builder.getInt64(0), ThreadID});
213     U->replaceUsesOfWith(GV, Addr);
214   }
215 
216   // Remove old global.
217   NewGV->takeName(GV);
218   GV->eraseFromParent();
219   return true;
220 }
221 
222 bool XCoreLowerThreadLocal::runOnModule(Module &M) {
223   // Find thread local globals.
224   bool MadeChange = false;
225   SmallVector<GlobalVariable *, 16> ThreadLocalGlobals;
226   for (GlobalVariable &GV : M.globals())
227     if (GV.isThreadLocal())
228       ThreadLocalGlobals.push_back(&GV);
229   for (unsigned I = 0, E = ThreadLocalGlobals.size(); I != E; ++I) {
230     MadeChange |= lowerGlobal(ThreadLocalGlobals[I]);
231   }
232   return MadeChange;
233 }
234