lib/Transforms/IPO/MergeFunctions.cpp

   1 //===- MergeFunctions.cpp - Merge identical functions ---------------------===//
   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 pass looks for equivalent functions that are mergable and folds them.
  11 //
  12 // A Function will not be analyzed if:
  13 // * it is overridable at runtime (except for weak linkage), or
  14 // * it is used by anything other than the callee parameter of a call/invoke
  15 //
  16 // A hash is computed from the function, based on its type and number of
  17 // basic blocks.
  18 //
  19 // Once all hashes are computed, we perform an expensive equality comparison
  20 // on each function pair. This takes n^2/2 comparisons per bucket, so it's
  21 // important that the hash function be high quality. The equality comparison
  22 // iterates through each instruction in each basic block.
  23 //
  24 // When a match is found, the functions are folded. We can only fold two
  25 // functions when we know that the definition of one of them is not
  26 // overridable.
  27 // * fold a function marked internal by replacing all of its users.
  28 // * fold extern or weak functions by replacing them with a global alias
  29 //
  30 //===----------------------------------------------------------------------===//
  31 //
  32 // Future work:
  33 //
  34 // * fold vector<T*>::push_back and vector<S*>::push_back.
  35 //
  36 // These two functions have different types, but in a way that doesn't matter
  37 // to us. As long as we never see an S or T itself, using S* and S** is the
  38 // same as using a T* and T**.
  39 //
  40 // * virtual functions.
  41 //
  42 // Many functions have their address taken by the virtual function table for
  43 // the object they belong to. However, as long as it's only used for a lookup
  44 // and call, this is irrelevant, and we'd like to fold such implementations.
  45 //
  46 //===----------------------------------------------------------------------===//
  47
  48 #define DEBUG_TYPE "mergefunc"
  49 #include "llvm/Transforms/IPO.h"
  50 #include "llvm/ADT/DenseMap.h"
  51 #include "llvm/ADT/Statistic.h"
  52 #include "llvm/Constants.h"
  53 #include "llvm/InlineAsm.h"
  54 #include "llvm/Instructions.h"
  55 #include "llvm/Module.h"
  56 #include "llvm/Pass.h"
  57 #include "llvm/Support/CallSite.h"
  58 #include "llvm/Support/Compiler.h"
  59 #include "llvm/Support/Debug.h"
  60 #include <map>
  61 #include <vector>
  62 using namespace llvm;
  63
  64 STATISTIC(NumFunctionsMerged, "Number of functions merged");
  65 STATISTIC(NumMergeFails, "Number of identical function pairings not merged");
  66
  67 namespace {
  68   struct VISIBILITY_HIDDEN MergeFunctions : public ModulePass {
  69     static char ID; // Pass identification, replacement for typeid
  70     MergeFunctions() : ModulePass((intptr_t)&ID) {}
  71
  72     bool runOnModule(Module &M);
  73   };
  74 }
  75
  76 char MergeFunctions::ID = 0;
  77 static RegisterPass<MergeFunctions>
  78 X("mergefunc", "Merge Functions");
  79
  80 ModulePass *llvm::createMergeFunctionsPass() {
  81   return new MergeFunctions();
  82 }
  83
  84 static unsigned long hash(const Function *F) {
  85   return F->size() ^ reinterpret_cast<unsigned long>(F->getType());
  86   //return F->size() ^ F->arg_size() ^ F->getReturnType();
  87 }
  88
  89 static bool compare(const Value *V, const Value *U) {
  90   assert(!isa<BasicBlock>(V) && !isa<BasicBlock>(U) &&
  91          "Must not compare basic blocks.");
  92
  93   assert(V->getType() == U->getType() &&
  94         "Two of the same operation have operands of different type.");
  95
  96   // TODO: If the constant is an expression of F, we should accept that it's
  97   // equal to the same expression in terms of G.
  98   if (isa<Constant>(V))
  99     return V == U;
 100
 101   // The caller has ensured that ValueMap[V] != U. Since Arguments are
 102   // pre-loaded into the ValueMap, and Instructions are added as we go, we know
 103   // that this can only be a mis-match.
 104   if (isa<Instruction>(V) || isa<Argument>(V))
 105     return false;
 106
 107   if (isa<InlineAsm>(V) && isa<InlineAsm>(U)) {
 108     const InlineAsm *IAF = cast<InlineAsm>(V);
 109     const InlineAsm *IAG = cast<InlineAsm>(U);
 110     return IAF->getAsmString() == IAG->getAsmString() &&
 111            IAF->getConstraintString() == IAG->getConstraintString();
 112   }
 113
 114   return false;
 115 }
 116
 117 static bool equals(const BasicBlock *BB1, const BasicBlock *BB2,
 118                    DenseMap<const Value *, const Value *> &ValueMap,
 119                    DenseMap<const Value *, const Value *> &SpeculationMap) {
 120   // Specutively add it anyways. If it's false, we'll notice a difference later, and
 121   // this won't matter.
 122   ValueMap[BB1] = BB2;
 123
 124   BasicBlock::const_iterator FI = BB1->begin(), FE = BB1->end();
 125   BasicBlock::const_iterator GI = BB2->begin(), GE = BB2->end();
 126
 127   do {
 128     if (!FI->isSameOperationAs(const_cast<Instruction *>(&*GI)))
 129       return false;
 130
 131     if (FI->getNumOperands() != GI->getNumOperands())
 132       return false;
 133
 134     if (ValueMap[FI] == GI) {
 135       ++FI, ++GI;
 136       continue;
 137     }
 138
 139     if (ValueMap[FI] != NULL)
 140       return false;
 141
 142     for (unsigned i = 0, e = FI->getNumOperands(); i != e; ++i) {
 143       Value *OpF = FI->getOperand(i);
 144       Value *OpG = GI->getOperand(i);
 145
 146       if (ValueMap[OpF] == OpG)
 147         continue;
 148
 149       if (ValueMap[OpF] != NULL)
 150         return false;
 151
 152       assert(OpF->getType() == OpG->getType() &&
 153              "Two of the same operation has operands of different type.");
 154
 155       if (OpF->getValueID() != OpG->getValueID())
 156         return false;
 157
 158       if (isa<PHINode>(FI)) {
 159         if (SpeculationMap[OpF] == NULL)
 160           SpeculationMap[OpF] = OpG;
 161         else if (SpeculationMap[OpF] != OpG)
 162           return false;
 163         continue;
 164       } else if (isa<BasicBlock>(OpF)) {
 165         assert(isa<TerminatorInst>(FI) &&
 166                "BasicBlock referenced by non-Terminator non-PHI");
 167         // This call changes the ValueMap, hence we can't use
 168         // Value *& = ValueMap[...]
 169         if (!equals(cast<BasicBlock>(OpF), cast<BasicBlock>(OpG), ValueMap,
 170                     SpeculationMap))
 171           return false;
 172       } else {
 173         if (!compare(OpF, OpG))
 174           return false;
 175       }
 176
 177       ValueMap[OpF] = OpG;
 178     }
 179
 180     ValueMap[FI] = GI;
 181     ++FI, ++GI;
 182   } while (FI != FE && GI != GE);
 183
 184   return FI == FE && GI == GE;
 185 }
 186
 187 static bool equals(const Function *F, const Function *G) {
 188   // We need to recheck everything, but check the things that weren't included
 189   // in the hash first.
 190
 191   if (F->getAttributes() != G->getAttributes())
 192     return false;
 193
 194   if (F->hasGC() != G->hasGC())
 195     return false;
 196
 197   if (F->hasGC() && F->getGC() != G->getGC())
 198     return false;
 199
 200   if (F->hasSection() != G->hasSection())
 201     return false;
 202
 203   if (F->hasSection() && F->getSection() != G->getSection())
 204     return false;
 205
 206   // TODO: if it's internal and only used in direct calls, we could handle this
 207   // case too.
 208   if (F->getCallingConv() != G->getCallingConv())
 209     return false;
 210
 211   // TODO: We want to permit cases where two functions take T* and S* but
 212   // only load or store them into T** and S**.
 213   if (F->getType() != G->getType())
 214     return false;
 215
 216   DenseMap<const Value *, const Value *> ValueMap;
 217   DenseMap<const Value *, const Value *> SpeculationMap;
 218   ValueMap[F] = G;
 219
 220   assert(F->arg_size() == G->arg_size() &&
 221          "Identical functions have a different number of args.");
 222
 223   for (Function::const_arg_iterator fi = F->arg_begin(), gi = G->arg_begin(),
 224          fe = F->arg_end(); fi != fe; ++fi, ++gi)
 225     ValueMap[fi] = gi;
 226
 227   if (!equals(&F->getEntryBlock(), &G->getEntryBlock(), ValueMap,
 228               SpeculationMap))
 229     return false;
 230
 231   for (DenseMap<const Value *, const Value *>::iterator
 232          I = SpeculationMap.begin(), E = SpeculationMap.end(); I != E; ++I) {
 233     if (ValueMap[I->first] != I->second)
 234       return false;
 235   }
 236
 237   return true;
 238 }
 239
 240 static bool fold(std::vector<Function *> &FnVec, unsigned i, unsigned j) {
 241   if (FnVec[i]->mayBeOverridden() && !FnVec[j]->mayBeOverridden())
 242     std::swap(FnVec[i], FnVec[j]);
 243
 244   Function *F = FnVec[i];
 245   Function *G = FnVec[j];
 246
 247   if (!F->mayBeOverridden()) {
 248     if (G->hasLocalLinkage()) {
 249       F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
 250       G->replaceAllUsesWith(F);
 251       G->eraseFromParent();
 252       ++NumFunctionsMerged;
 253       return true;
 254     }
 255
 256     if (G->hasExternalLinkage() || G->hasWeakLinkage()) {
 257       GlobalAlias *GA = new GlobalAlias(G->getType(), G->getLinkage(), "",
 258                                         F, G->getParent());
 259       F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
 260       GA->takeName(G);
 261       GA->setVisibility(G->getVisibility());
 262       G->replaceAllUsesWith(GA);
 263       G->eraseFromParent();
 264       ++NumFunctionsMerged;
 265       return true;
 266     }
 267   }
 268
 269   if (F->hasWeakLinkage() && G->hasWeakLinkage()) {
 270     GlobalAlias *GA_F = new GlobalAlias(F->getType(), F->getLinkage(), "",
 271                                         0, F->getParent());
 272     GA_F->takeName(F);
 273     GA_F->setVisibility(F->getVisibility());
 274     F->setAlignment(std::max(F->getAlignment(), G->getAlignment()));
 275     F->replaceAllUsesWith(GA_F);
 276     F->setName("folded." + GA_F->getName());
 277     F->setLinkage(GlobalValue::ExternalLinkage);
 278     GA_F->setAliasee(F);
 279
 280     GlobalAlias *GA_G = new GlobalAlias(G->getType(), G->getLinkage(), "",
 281                                         F, G->getParent());
 282     GA_G->takeName(G);
 283     GA_G->setVisibility(G->getVisibility());
 284     G->replaceAllUsesWith(GA_G);
 285     G->eraseFromParent();
 286
 287     ++NumFunctionsMerged;
 288     return true;
 289   }
 290
 291   DOUT << "Failed on " << F->getName() << " and " << G->getName() << "\n";
 292
 293   ++NumMergeFails;
 294   return false;
 295 }
 296
 297 static bool hasAddressTaken(User *U) {
 298   for (User::use_iterator I = U->use_begin(), E = U->use_end(); I != E; ++I) {
 299     User *Use = *I;
 300
 301     // 'call (bitcast @F to ...)' happens a lot.
 302     while (isa<ConstantExpr>(Use) && Use->hasOneUse()) {
 303       Use = *Use->use_begin();
 304     }
 305
 306     if (isa<ConstantExpr>(Use)) {
 307       if (hasAddressTaken(Use))
 308         return true;
 309     }
 310
 311     if (!isa<CallInst>(Use) && !isa<InvokeInst>(Use))
 312       return true;
 313
 314     // Make sure we aren't passing U as a parameter to call instead of the
 315     // callee.
 316     if (CallSite(cast<Instruction>(Use)).hasArgument(U))
 317       return true;
 318   }
 319
 320   return false;
 321 }
 322
 323 bool MergeFunctions::runOnModule(Module &M) {
 324   bool Changed = false;
 325
 326   std::map<unsigned long, std::vector<Function *> > FnMap;
 327
 328   for (Module::iterator F = M.begin(), E = M.end(); F != E; ++F) {
 329     if (F->isDeclaration() || F->isIntrinsic())
 330       continue;
 331
 332     if (!F->hasLocalLinkage() && !F->hasExternalLinkage() &&
 333         !F->hasWeakLinkage())
 334       continue;
 335
 336     if (hasAddressTaken(F))
 337       continue;
 338
 339     FnMap[hash(F)].push_back(F);
 340   }
 341
 342   // TODO: instead of running in a loop, we could also fold functions in callgraph
 343   // order. Constructing the CFG probably isn't cheaper than just running in a loop.
 344
 345   bool LocalChanged;
 346   do {
 347     LocalChanged = false;
 348     for (std::map<unsigned long, std::vector<Function *> >::iterator
 349          I = FnMap.begin(), E = FnMap.end(); I != E; ++I) {
 350       DOUT << "size: " << FnMap.size() << "\n";
 351       std::vector<Function *> &FnVec = I->second;
 352       DOUT << "hash (" << I->first << "): " << FnVec.size() << "\n";
 353
 354       for (int i = 0, e = FnVec.size(); i != e; ++i) {
 355         for (int j = i + 1; j != e; ++j) {
 356           bool isEqual = equals(FnVec[i], FnVec[j]);
 357
 358           DOUT << "  " << FnVec[i]->getName()
 359                << (isEqual ? " == " : " != ")
 360                << FnVec[j]->getName() << "\n";
 361
 362           if (isEqual) {
 363             if (fold(FnVec, i, j)) {
 364               LocalChanged = true;
 365               FnVec.erase(FnVec.begin() + j);
 366               --j, --e;
 367             }
 368           }
 369         }
 370       }
 371
 372     }
 373     Changed |= LocalChanged;
 374   } while (LocalChanged);
 375
 376   return Changed;
 377 }