1 //===- ConstantMerge.cpp - Merge duplicate global constants ---------------===//
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 // This file defines the interface to a pass that merges duplicate global
11 // constants together into a single constant that is shared.  This is useful
12 // because some passes (ie TraceValues) insert a lot of string constants into
13 // the program, regardless of whether or not an existing string is available.
14 //
15 // Algorithm: ConstantMerge is designed to build up a map of available constants
16 // and eliminate duplicates when it is initialized.
17 //
18 //===----------------------------------------------------------------------===//
19 
20 #include "llvm/Transforms/IPO.h"
21 #include "llvm/ADT/DenseMap.h"
22 #include "llvm/ADT/PointerIntPair.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Module.h"
29 #include "llvm/IR/Operator.h"
30 #include "llvm/Pass.h"
31 using namespace llvm;
32 
33 #define DEBUG_TYPE "constmerge"
34 
35 STATISTIC(NumMerged, "Number of global constants merged");
36 
37 namespace {
38   struct ConstantMerge : public ModulePass {
39     static char ID; // Pass identification, replacement for typeid
ConstantMerge__anonf592c4e10111::ConstantMerge40     ConstantMerge() : ModulePass(ID) {
41       initializeConstantMergePass(*PassRegistry::getPassRegistry());
42     }
43 
44     // For this pass, process all of the globals in the module, eliminating
45     // duplicate constants.
46     bool runOnModule(Module &M) override;
47 
48     // Return true iff we can determine the alignment of this global variable.
49     bool hasKnownAlignment(GlobalVariable *GV) const;
50 
51     // Return the alignment of the global, including converting the default
52     // alignment to a concrete value.
53     unsigned getAlignment(GlobalVariable *GV) const;
54 
55   };
56 }
57 
58 char ConstantMerge::ID = 0;
59 INITIALIZE_PASS(ConstantMerge, "constmerge",
60                 "Merge Duplicate Global Constants", false, false)
61 
createConstantMergePass()62 ModulePass *llvm::createConstantMergePass() { return new ConstantMerge(); }
63 
64 
65 
66 /// Find values that are marked as llvm.used.
FindUsedValues(GlobalVariable * LLVMUsed,SmallPtrSetImpl<const GlobalValue * > & UsedValues)67 static void FindUsedValues(GlobalVariable *LLVMUsed,
68                            SmallPtrSetImpl<const GlobalValue*> &UsedValues) {
69   if (!LLVMUsed) return;
70   ConstantArray *Inits = cast<ConstantArray>(LLVMUsed->getInitializer());
71 
72   for (unsigned i = 0, e = Inits->getNumOperands(); i != e; ++i) {
73     Value *Operand = Inits->getOperand(i)->stripPointerCastsNoFollowAliases();
74     GlobalValue *GV = cast<GlobalValue>(Operand);
75     UsedValues.insert(GV);
76   }
77 }
78 
79 // True if A is better than B.
IsBetterCanonical(const GlobalVariable & A,const GlobalVariable & B)80 static bool IsBetterCanonical(const GlobalVariable &A,
81                               const GlobalVariable &B) {
82   if (!A.hasLocalLinkage() && B.hasLocalLinkage())
83     return true;
84 
85   if (A.hasLocalLinkage() && !B.hasLocalLinkage())
86     return false;
87 
88   return A.hasUnnamedAddr();
89 }
90 
getAlignment(GlobalVariable * GV) const91 unsigned ConstantMerge::getAlignment(GlobalVariable *GV) const {
92   unsigned Align = GV->getAlignment();
93   if (Align)
94     return Align;
95   return GV->getParent()->getDataLayout().getPreferredAlignment(GV);
96 }
97 
runOnModule(Module & M)98 bool ConstantMerge::runOnModule(Module &M) {
99 
100   // Find all the globals that are marked "used".  These cannot be merged.
101   SmallPtrSet<const GlobalValue*, 8> UsedGlobals;
102   FindUsedValues(M.getGlobalVariable("llvm.used"), UsedGlobals);
103   FindUsedValues(M.getGlobalVariable("llvm.compiler.used"), UsedGlobals);
104 
105   // Map unique constants to globals.
106   DenseMap<Constant *, GlobalVariable *> CMap;
107 
108   // Replacements - This vector contains a list of replacements to perform.
109   SmallVector<std::pair<GlobalVariable*, GlobalVariable*>, 32> Replacements;
110 
111   bool MadeChange = false;
112 
113   // Iterate constant merging while we are still making progress.  Merging two
114   // constants together may allow us to merge other constants together if the
115   // second level constants have initializers which point to the globals that
116   // were just merged.
117   while (1) {
118 
119     // First: Find the canonical constants others will be merged with.
120     for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
121          GVI != E; ) {
122       GlobalVariable *GV = &*GVI++;
123 
124       // If this GV is dead, remove it.
125       GV->removeDeadConstantUsers();
126       if (GV->use_empty() && GV->hasLocalLinkage()) {
127         GV->eraseFromParent();
128         continue;
129       }
130 
131       // Only process constants with initializers in the default address space.
132       if (!GV->isConstant() || !GV->hasDefinitiveInitializer() ||
133           GV->getType()->getAddressSpace() != 0 || GV->hasSection() ||
134           // Don't touch values marked with attribute(used).
135           UsedGlobals.count(GV))
136         continue;
137 
138       // This transformation is legal for weak ODR globals in the sense it
139       // doesn't change semantics, but we really don't want to perform it
140       // anyway; it's likely to pessimize code generation, and some tools
141       // (like the Darwin linker in cases involving CFString) don't expect it.
142       if (GV->isWeakForLinker())
143         continue;
144 
145       Constant *Init = GV->getInitializer();
146 
147       // Check to see if the initializer is already known.
148       GlobalVariable *&Slot = CMap[Init];
149 
150       // If this is the first constant we find or if the old one is local,
151       // replace with the current one. If the current is externally visible
152       // it cannot be replace, but can be the canonical constant we merge with.
153       if (!Slot || IsBetterCanonical(*GV, *Slot))
154         Slot = GV;
155     }
156 
157     // Second: identify all globals that can be merged together, filling in
158     // the Replacements vector.  We cannot do the replacement in this pass
159     // because doing so may cause initializers of other globals to be rewritten,
160     // invalidating the Constant* pointers in CMap.
161     for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
162          GVI != E; ) {
163       GlobalVariable *GV = &*GVI++;
164 
165       // Only process constants with initializers in the default address space.
166       if (!GV->isConstant() || !GV->hasDefinitiveInitializer() ||
167           GV->getType()->getAddressSpace() != 0 || GV->hasSection() ||
168           // Don't touch values marked with attribute(used).
169           UsedGlobals.count(GV))
170         continue;
171 
172       // We can only replace constant with local linkage.
173       if (!GV->hasLocalLinkage())
174         continue;
175 
176       Constant *Init = GV->getInitializer();
177 
178       // Check to see if the initializer is already known.
179       GlobalVariable *Slot = CMap[Init];
180 
181       if (!Slot || Slot == GV)
182         continue;
183 
184       if (!Slot->hasUnnamedAddr() && !GV->hasUnnamedAddr())
185         continue;
186 
187       if (!GV->hasUnnamedAddr())
188         Slot->setUnnamedAddr(false);
189 
190       // Make all uses of the duplicate constant use the canonical version.
191       Replacements.push_back(std::make_pair(GV, Slot));
192     }
193 
194     if (Replacements.empty())
195       return MadeChange;
196     CMap.clear();
197 
198     // Now that we have figured out which replacements must be made, do them all
199     // now.  This avoid invalidating the pointers in CMap, which are unneeded
200     // now.
201     for (unsigned i = 0, e = Replacements.size(); i != e; ++i) {
202       // Bump the alignment if necessary.
203       if (Replacements[i].first->getAlignment() ||
204           Replacements[i].second->getAlignment()) {
205         Replacements[i].second->setAlignment(
206             std::max(getAlignment(Replacements[i].first),
207                      getAlignment(Replacements[i].second)));
208       }
209 
210       // Eliminate any uses of the dead global.
211       Replacements[i].first->replaceAllUsesWith(Replacements[i].second);
212 
213       // Delete the global value from the module.
214       assert(Replacements[i].first->hasLocalLinkage() &&
215              "Refusing to delete an externally visible global variable.");
216       Replacements[i].first->eraseFromParent();
217     }
218 
219     NumMerged += Replacements.size();
220     Replacements.clear();
221   }
222 }
223