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/IntrinsicsXCore.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 
78 static bool replaceConstantExprOp(ConstantExpr *CE, Pass *P) {
79   do {
80     SmallVector<WeakTrackingVH, 8> WUsers(CE->users());
81     llvm::sort(WUsers);
82     WUsers.erase(std::unique(WUsers.begin(), WUsers.end()), WUsers.end());
83     while (!WUsers.empty())
84       if (WeakTrackingVH WU = WUsers.pop_back_val()) {
85         if (PHINode *PN = dyn_cast<PHINode>(WU)) {
86           for (int I = 0, E = PN->getNumIncomingValues(); I < E; ++I)
87             if (PN->getIncomingValue(I) == CE) {
88               BasicBlock *PredBB = PN->getIncomingBlock(I);
89               if (PredBB->getTerminator()->getNumSuccessors() > 1)
90                 PredBB = SplitEdge(PredBB, PN->getParent());
91               Instruction *InsertPos = PredBB->getTerminator();
92               Instruction *NewInst = CE->getAsInstruction(InsertPos);
93               PN->setOperand(I, NewInst);
94             }
95         } else if (Instruction *Instr = dyn_cast<Instruction>(WU)) {
96           Instruction *NewInst = CE->getAsInstruction(Instr);
97           Instr->replaceUsesOfWith(CE, NewInst);
98         } else {
99           ConstantExpr *CExpr = dyn_cast<ConstantExpr>(WU);
100           if (!CExpr || !replaceConstantExprOp(CExpr, P))
101             return false;
102         }
103       }
104   } while (CE->hasNUsesOrMore(1)); // We need to check because a recursive
105   // sibling may have used 'CE' when getAsInstruction was called.
106   CE->destroyConstant();
107   return true;
108 }
109 
110 static bool rewriteNonInstructionUses(GlobalVariable *GV, Pass *P) {
111   SmallVector<WeakTrackingVH, 8> WUsers;
112   for (User *U : GV->users())
113     if (!isa<Instruction>(U))
114       WUsers.push_back(WeakTrackingVH(U));
115   while (!WUsers.empty())
116     if (WeakTrackingVH WU = WUsers.pop_back_val()) {
117       ConstantExpr *CE = dyn_cast<ConstantExpr>(WU);
118       if (!CE || !replaceConstantExprOp(CE, P))
119         return false;
120     }
121   return true;
122 }
123 
124 static bool isZeroLengthArray(Type *Ty) {
125   ArrayType *AT = dyn_cast<ArrayType>(Ty);
126   return AT && (AT->getNumElements() == 0);
127 }
128 
129 bool XCoreLowerThreadLocal::lowerGlobal(GlobalVariable *GV) {
130   Module *M = GV->getParent();
131   if (!GV->isThreadLocal())
132     return false;
133 
134   // Skip globals that we can't lower and leave it for the backend to error.
135   if (!rewriteNonInstructionUses(GV, this) ||
136       !GV->getType()->isSized() || isZeroLengthArray(GV->getType()))
137     return false;
138 
139   // Create replacement global.
140   ArrayType *NewType = createLoweredType(GV->getValueType());
141   Constant *NewInitializer = nullptr;
142   if (GV->hasInitializer())
143     NewInitializer = createLoweredInitializer(NewType,
144                                               GV->getInitializer());
145   GlobalVariable *NewGV =
146     new GlobalVariable(*M, NewType, GV->isConstant(), GV->getLinkage(),
147                        NewInitializer, "", nullptr,
148                        GlobalVariable::NotThreadLocal,
149                        GV->getType()->getAddressSpace(),
150                        GV->isExternallyInitialized());
151 
152   // Update uses.
153   SmallVector<User *, 16> Users(GV->users());
154   for (unsigned I = 0, E = Users.size(); I != E; ++I) {
155     User *U = Users[I];
156     Instruction *Inst = cast<Instruction>(U);
157     IRBuilder<> Builder(Inst);
158     Function *GetID = Intrinsic::getDeclaration(GV->getParent(),
159                                                 Intrinsic::xcore_getid);
160     Value *ThreadID = Builder.CreateCall(GetID, {});
161     Value *Addr = Builder.CreateInBoundsGEP(NewGV->getValueType(), NewGV,
162                                             {Builder.getInt64(0), ThreadID});
163     U->replaceUsesOfWith(GV, Addr);
164   }
165 
166   // Remove old global.
167   NewGV->takeName(GV);
168   GV->eraseFromParent();
169   return true;
170 }
171 
172 bool XCoreLowerThreadLocal::runOnModule(Module &M) {
173   // Find thread local globals.
174   bool MadeChange = false;
175   SmallVector<GlobalVariable *, 16> ThreadLocalGlobals;
176   for (GlobalVariable &GV : M.globals())
177     if (GV.isThreadLocal())
178       ThreadLocalGlobals.push_back(&GV);
179   for (unsigned I = 0, E = ThreadLocalGlobals.size(); I != E; ++I) {
180     MadeChange |= lowerGlobal(ThreadLocalGlobals[I]);
181   }
182   return MadeChange;
183 }
184