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"
31using namespace llvm;
32
33#define DEBUG_TYPE "constmerge"
34
35STATISTIC(NumMerged, "Number of global constants merged");
36
37namespace {
38  struct ConstantMerge : public ModulePass {
39    static char ID; // Pass identification, replacement for typeid
40    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    const DataLayout *DL;
56  };
57}
58
59char ConstantMerge::ID = 0;
60INITIALIZE_PASS(ConstantMerge, "constmerge",
61                "Merge Duplicate Global Constants", false, false)
62
63ModulePass *llvm::createConstantMergePass() { return new ConstantMerge(); }
64
65
66
67/// Find values that are marked as llvm.used.
68static void FindUsedValues(GlobalVariable *LLVMUsed,
69                           SmallPtrSet<const GlobalValue*, 8> &UsedValues) {
70  if (!LLVMUsed) return;
71  ConstantArray *Inits = cast<ConstantArray>(LLVMUsed->getInitializer());
72
73  for (unsigned i = 0, e = Inits->getNumOperands(); i != e; ++i) {
74    Value *Operand = Inits->getOperand(i)->stripPointerCastsNoFollowAliases();
75    GlobalValue *GV = cast<GlobalValue>(Operand);
76    UsedValues.insert(GV);
77  }
78}
79
80// True if A is better than B.
81static bool IsBetterCanonical(const GlobalVariable &A,
82                              const GlobalVariable &B) {
83  if (!A.hasLocalLinkage() && B.hasLocalLinkage())
84    return true;
85
86  if (A.hasLocalLinkage() && !B.hasLocalLinkage())
87    return false;
88
89  return A.hasUnnamedAddr();
90}
91
92bool ConstantMerge::hasKnownAlignment(GlobalVariable *GV) const {
93  return DL || GV->getAlignment() != 0;
94}
95
96unsigned ConstantMerge::getAlignment(GlobalVariable *GV) const {
97  unsigned Align = GV->getAlignment();
98  if (Align)
99    return Align;
100  if (DL)
101    return DL->getPreferredAlignment(GV);
102  return 0;
103}
104
105bool ConstantMerge::runOnModule(Module &M) {
106  DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
107  DL = DLP ? &DLP->getDataLayout() : nullptr;
108
109  // Find all the globals that are marked "used".  These cannot be merged.
110  SmallPtrSet<const GlobalValue*, 8> UsedGlobals;
111  FindUsedValues(M.getGlobalVariable("llvm.used"), UsedGlobals);
112  FindUsedValues(M.getGlobalVariable("llvm.compiler.used"), UsedGlobals);
113
114  // Map unique <constants, has-unknown-alignment> pairs to globals.  We don't
115  // want to merge globals of unknown alignment with those of explicit
116  // alignment.  If we have DataLayout, we always know the alignment.
117  DenseMap<PointerIntPair<Constant*, 1, bool>, GlobalVariable*> CMap;
118
119  // Replacements - This vector contains a list of replacements to perform.
120  SmallVector<std::pair<GlobalVariable*, GlobalVariable*>, 32> Replacements;
121
122  bool MadeChange = false;
123
124  // Iterate constant merging while we are still making progress.  Merging two
125  // constants together may allow us to merge other constants together if the
126  // second level constants have initializers which point to the globals that
127  // were just merged.
128  while (1) {
129
130    // First: Find the canonical constants others will be merged with.
131    for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
132         GVI != E; ) {
133      GlobalVariable *GV = GVI++;
134
135      // If this GV is dead, remove it.
136      GV->removeDeadConstantUsers();
137      if (GV->use_empty() && GV->hasLocalLinkage()) {
138        GV->eraseFromParent();
139        continue;
140      }
141
142      // Only process constants with initializers in the default address space.
143      if (!GV->isConstant() || !GV->hasDefinitiveInitializer() ||
144          GV->getType()->getAddressSpace() != 0 || GV->hasSection() ||
145          // Don't touch values marked with attribute(used).
146          UsedGlobals.count(GV))
147        continue;
148
149      // This transformation is legal for weak ODR globals in the sense it
150      // doesn't change semantics, but we really don't want to perform it
151      // anyway; it's likely to pessimize code generation, and some tools
152      // (like the Darwin linker in cases involving CFString) don't expect it.
153      if (GV->isWeakForLinker())
154        continue;
155
156      Constant *Init = GV->getInitializer();
157
158      // Check to see if the initializer is already known.
159      PointerIntPair<Constant*, 1, bool> Pair(Init, hasKnownAlignment(GV));
160      GlobalVariable *&Slot = CMap[Pair];
161
162      // If this is the first constant we find or if the old one is local,
163      // replace with the current one. If the current is externally visible
164      // it cannot be replace, but can be the canonical constant we merge with.
165      if (!Slot || IsBetterCanonical(*GV, *Slot))
166        Slot = GV;
167    }
168
169    // Second: identify all globals that can be merged together, filling in
170    // the Replacements vector.  We cannot do the replacement in this pass
171    // because doing so may cause initializers of other globals to be rewritten,
172    // invalidating the Constant* pointers in CMap.
173    for (Module::global_iterator GVI = M.global_begin(), E = M.global_end();
174         GVI != E; ) {
175      GlobalVariable *GV = GVI++;
176
177      // Only process constants with initializers in the default address space.
178      if (!GV->isConstant() || !GV->hasDefinitiveInitializer() ||
179          GV->getType()->getAddressSpace() != 0 || GV->hasSection() ||
180          // Don't touch values marked with attribute(used).
181          UsedGlobals.count(GV))
182        continue;
183
184      // We can only replace constant with local linkage.
185      if (!GV->hasLocalLinkage())
186        continue;
187
188      Constant *Init = GV->getInitializer();
189
190      // Check to see if the initializer is already known.
191      PointerIntPair<Constant*, 1, bool> Pair(Init, hasKnownAlignment(GV));
192      GlobalVariable *Slot = CMap[Pair];
193
194      if (!Slot || Slot == GV)
195        continue;
196
197      if (!Slot->hasUnnamedAddr() && !GV->hasUnnamedAddr())
198        continue;
199
200      if (!GV->hasUnnamedAddr())
201        Slot->setUnnamedAddr(false);
202
203      // Make all uses of the duplicate constant use the canonical version.
204      Replacements.push_back(std::make_pair(GV, Slot));
205    }
206
207    if (Replacements.empty())
208      return MadeChange;
209    CMap.clear();
210
211    // Now that we have figured out which replacements must be made, do them all
212    // now.  This avoid invalidating the pointers in CMap, which are unneeded
213    // now.
214    for (unsigned i = 0, e = Replacements.size(); i != e; ++i) {
215      // Bump the alignment if necessary.
216      if (Replacements[i].first->getAlignment() ||
217          Replacements[i].second->getAlignment()) {
218        Replacements[i].second->setAlignment(
219            std::max(getAlignment(Replacements[i].first),
220                     getAlignment(Replacements[i].second)));
221      }
222
223      // Eliminate any uses of the dead global.
224      Replacements[i].first->replaceAllUsesWith(Replacements[i].second);
225
226      // Delete the global value from the module.
227      assert(Replacements[i].first->hasLocalLinkage() &&
228             "Refusing to delete an externally visible global variable.");
229      Replacements[i].first->eraseFromParent();
230    }
231
232    NumMerged += Replacements.size();
233    Replacements.clear();
234  }
235}
236