lib/Transforms/Utils/FunctionComparator.cpp

   1 //===- FunctionComparator.h - Function Comparator -------------------------===//
   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 implements the FunctionComparator and GlobalNumberState classes
  11 // which are used by the MergeFunctions pass for comparing functions.
  12 //
  13 //===----------------------------------------------------------------------===//
  14
  15 #include "llvm/Transforms/Utils/FunctionComparator.h"
  16 #include "llvm/ADT/SmallSet.h"
  17 #include "llvm/IR/CallSite.h"
  18 #include "llvm/IR/InlineAsm.h"
  19 #include "llvm/IR/Instructions.h"
  20 #include "llvm/IR/Module.h"
  21 #include "llvm/Support/Debug.h"
  22 #include "llvm/Support/raw_ostream.h"
  23
  24 using namespace llvm;
  25
  26 #define DEBUG_TYPE "functioncomparator"
  27
  28 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
  29   if (L < R) return -1;
  30   if (L > R) return 1;
  31   return 0;
  32 }
  33
  34 int FunctionComparator::cmpOrderings(AtomicOrdering L, AtomicOrdering R) const {
  35   if ((int)L < (int)R) return -1;
  36   if ((int)L > (int)R) return 1;
  37   return 0;
  38 }
  39
  40 int FunctionComparator::cmpAPInts(const APInt &L, const APInt &R) const {
  41   if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
  42     return Res;
  43   if (L.ugt(R)) return 1;
  44   if (R.ugt(L)) return -1;
  45   return 0;
  46 }
  47
  48 int FunctionComparator::cmpAPFloats(const APFloat &L, const APFloat &R) const {
  49   // Floats are ordered first by semantics (i.e. float, double, half, etc.),
  50   // then by value interpreted as a bitstring (aka APInt).
  51   const fltSemantics &SL = L.getSemantics(), &SR = R.getSemantics();
  52   if (int Res = cmpNumbers(APFloat::semanticsPrecision(SL),
  53                            APFloat::semanticsPrecision(SR)))
  54     return Res;
  55   if (int Res = cmpNumbers(APFloat::semanticsMaxExponent(SL),
  56                            APFloat::semanticsMaxExponent(SR)))
  57     return Res;
  58   if (int Res = cmpNumbers(APFloat::semanticsMinExponent(SL),
  59                            APFloat::semanticsMinExponent(SR)))
  60     return Res;
  61   if (int Res = cmpNumbers(APFloat::semanticsSizeInBits(SL),
  62                            APFloat::semanticsSizeInBits(SR)))
  63     return Res;
  64   return cmpAPInts(L.bitcastToAPInt(), R.bitcastToAPInt());
  65 }
  66
  67 int FunctionComparator::cmpMem(StringRef L, StringRef R) const {
  68   // Prevent heavy comparison, compare sizes first.
  69   if (int Res = cmpNumbers(L.size(), R.size()))
  70     return Res;
  71
  72   // Compare strings lexicographically only when it is necessary: only when
  73   // strings are equal in size.
  74   return L.compare(R);
  75 }
  76
  77 int FunctionComparator::cmpAttrs(const AttributeList L,
  78                                  const AttributeList R) const {
  79   if (int Res = cmpNumbers(L.getNumAttrSets(), R.getNumAttrSets()))
  80     return Res;
  81
  82   for (unsigned i = L.index_begin(), e = L.index_end(); i != e; ++i) {
  83     AttributeSet LAS = L.getAttributes(i);
  84     AttributeSet RAS = R.getAttributes(i);
  85     AttributeSet::iterator LI = LAS.begin(), LE = LAS.end();
  86     AttributeSet::iterator RI = RAS.begin(), RE = RAS.end();
  87     for (; LI != LE && RI != RE; ++LI, ++RI) {
  88       Attribute LA = *LI;
  89       Attribute RA = *RI;
  90       if (LA < RA)
  91         return -1;
  92       if (RA < LA)
  93         return 1;
  94     }
  95     if (LI != LE)
  96       return 1;
  97     if (RI != RE)
  98       return -1;
  99   }
 100   return 0;
 101 }
 102
 103 int FunctionComparator::cmpRangeMetadata(const MDNode *L,
 104                                          const MDNode *R) const {
 105   if (L == R)
 106     return 0;
 107   if (!L)
 108     return -1;
 109   if (!R)
 110     return 1;
 111   // Range metadata is a sequence of numbers. Make sure they are the same
 112   // sequence.
 113   // TODO: Note that as this is metadata, it is possible to drop and/or merge
 114   // this data when considering functions to merge. Thus this comparison would
 115   // return 0 (i.e. equivalent), but merging would become more complicated
 116   // because the ranges would need to be unioned. It is not likely that
 117   // functions differ ONLY in this metadata if they are actually the same
 118   // function semantically.
 119   if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
 120     return Res;
 121   for (size_t I = 0; I < L->getNumOperands(); ++I) {
 122     ConstantInt *LLow = mdconst::extract<ConstantInt>(L->getOperand(I));
 123     ConstantInt *RLow = mdconst::extract<ConstantInt>(R->getOperand(I));
 124     if (int Res = cmpAPInts(LLow->getValue(), RLow->getValue()))
 125       return Res;
 126   }
 127   return 0;
 128 }
 129
 130 int FunctionComparator::cmpOperandBundlesSchema(const Instruction *L,
 131                                                 const Instruction *R) const {
 132   ImmutableCallSite LCS(L);
 133   ImmutableCallSite RCS(R);
 134
 135   assert(LCS && RCS && "Must be calls or invokes!");
 136   assert(LCS.isCall() == RCS.isCall() && "Can't compare otherwise!");
 137
 138   if (int Res =
 139           cmpNumbers(LCS.getNumOperandBundles(), RCS.getNumOperandBundles()))
 140     return Res;
 141
 142   for (unsigned i = 0, e = LCS.getNumOperandBundles(); i != e; ++i) {
 143     auto OBL = LCS.getOperandBundleAt(i);
 144     auto OBR = RCS.getOperandBundleAt(i);
 145
 146     if (int Res = OBL.getTagName().compare(OBR.getTagName()))
 147       return Res;
 148
 149     if (int Res = cmpNumbers(OBL.Inputs.size(), OBR.Inputs.size()))
 150       return Res;
 151   }
 152
 153   return 0;
 154 }
 155
 156 /// Constants comparison:
 157 /// 1. Check whether type of L constant could be losslessly bitcasted to R
 158 /// type.
 159 /// 2. Compare constant contents.
 160 /// For more details see declaration comments.
 161 int FunctionComparator::cmpConstants(const Constant *L,
 162                                      const Constant *R) const {
 163
 164   Type *TyL = L->getType();
 165   Type *TyR = R->getType();
 166
 167   // Check whether types are bitcastable. This part is just re-factored
 168   // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
 169   // we also pack into result which type is "less" for us.
 170   int TypesRes = cmpTypes(TyL, TyR);
 171   if (TypesRes != 0) {
 172     // Types are different, but check whether we can bitcast them.
 173     if (!TyL->isFirstClassType()) {
 174       if (TyR->isFirstClassType())
 175         return -1;
 176       // Neither TyL nor TyR are values of first class type. Return the result
 177       // of comparing the types
 178       return TypesRes;
 179     }
 180     if (!TyR->isFirstClassType()) {
 181       if (TyL->isFirstClassType())
 182         return 1;
 183       return TypesRes;
 184     }
 185
 186     // Vector -> Vector conversions are always lossless if the two vector types
 187     // have the same size, otherwise not.
 188     unsigned TyLWidth = 0;
 189     unsigned TyRWidth = 0;
 190
 191     if (auto *VecTyL = dyn_cast<VectorType>(TyL))
 192       TyLWidth = VecTyL->getBitWidth();
 193     if (auto *VecTyR = dyn_cast<VectorType>(TyR))
 194       TyRWidth = VecTyR->getBitWidth();
 195
 196     if (TyLWidth != TyRWidth)
 197       return cmpNumbers(TyLWidth, TyRWidth);
 198
 199     // Zero bit-width means neither TyL nor TyR are vectors.
 200     if (!TyLWidth) {
 201       PointerType *PTyL = dyn_cast<PointerType>(TyL);
 202       PointerType *PTyR = dyn_cast<PointerType>(TyR);
 203       if (PTyL && PTyR) {
 204         unsigned AddrSpaceL = PTyL->getAddressSpace();
 205         unsigned AddrSpaceR = PTyR->getAddressSpace();
 206         if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
 207           return Res;
 208       }
 209       if (PTyL)
 210         return 1;
 211       if (PTyR)
 212         return -1;
 213
 214       // TyL and TyR aren't vectors, nor pointers. We don't know how to
 215       // bitcast them.
 216       return TypesRes;
 217     }
 218   }
 219
 220   // OK, types are bitcastable, now check constant contents.
 221
 222   if (L->isNullValue() && R->isNullValue())
 223     return TypesRes;
 224   if (L->isNullValue() && !R->isNullValue())
 225     return 1;
 226   if (!L->isNullValue() && R->isNullValue())
 227     return -1;
 228
 229   auto GlobalValueL = const_cast<GlobalValue*>(dyn_cast<GlobalValue>(L));
 230   auto GlobalValueR = const_cast<GlobalValue*>(dyn_cast<GlobalValue>(R));
 231   if (GlobalValueL && GlobalValueR) {
 232     return cmpGlobalValues(GlobalValueL, GlobalValueR);
 233   }
 234
 235   if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
 236     return Res;
 237
 238   if (const auto *SeqL = dyn_cast<ConstantDataSequential>(L)) {
 239     const auto *SeqR = cast<ConstantDataSequential>(R);
 240     // This handles ConstantDataArray and ConstantDataVector. Note that we
 241     // compare the two raw data arrays, which might differ depending on the host
 242     // endianness. This isn't a problem though, because the endiness of a module
 243     // will affect the order of the constants, but this order is the same
 244     // for a given input module and host platform.
 245     return cmpMem(SeqL->getRawDataValues(), SeqR->getRawDataValues());
 246   }
 247
 248   switch (L->getValueID()) {
 249   case Value::UndefValueVal:
 250   case Value::ConstantTokenNoneVal:
 251     return TypesRes;
 252   case Value::ConstantIntVal: {
 253     const APInt &LInt = cast<ConstantInt>(L)->getValue();
 254     const APInt &RInt = cast<ConstantInt>(R)->getValue();
 255     return cmpAPInts(LInt, RInt);
 256   }
 257   case Value::ConstantFPVal: {
 258     const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
 259     const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
 260     return cmpAPFloats(LAPF, RAPF);
 261   }
 262   case Value::ConstantArrayVal: {
 263     const ConstantArray *LA = cast<ConstantArray>(L);
 264     const ConstantArray *RA = cast<ConstantArray>(R);
 265     uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
 266     uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
 267     if (int Res = cmpNumbers(NumElementsL, NumElementsR))
 268       return Res;
 269     for (uint64_t i = 0; i < NumElementsL; ++i) {
 270       if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
 271                                  cast<Constant>(RA->getOperand(i))))
 272         return Res;
 273     }
 274     return 0;
 275   }
 276   case Value::ConstantStructVal: {
 277     const ConstantStruct *LS = cast<ConstantStruct>(L);
 278     const ConstantStruct *RS = cast<ConstantStruct>(R);
 279     unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
 280     unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
 281     if (int Res = cmpNumbers(NumElementsL, NumElementsR))
 282       return Res;
 283     for (unsigned i = 0; i != NumElementsL; ++i) {
 284       if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
 285                                  cast<Constant>(RS->getOperand(i))))
 286         return Res;
 287     }
 288     return 0;
 289   }
 290   case Value::ConstantVectorVal: {
 291     const ConstantVector *LV = cast<ConstantVector>(L);
 292     const ConstantVector *RV = cast<ConstantVector>(R);
 293     unsigned NumElementsL = cast<VectorType>(TyL)->getNumElements();
 294     unsigned NumElementsR = cast<VectorType>(TyR)->getNumElements();
 295     if (int Res = cmpNumbers(NumElementsL, NumElementsR))
 296       return Res;
 297     for (uint64_t i = 0; i < NumElementsL; ++i) {
 298       if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
 299                                  cast<Constant>(RV->getOperand(i))))
 300         return Res;
 301     }
 302     return 0;
 303   }
 304   case Value::ConstantExprVal: {
 305     const ConstantExpr *LE = cast<ConstantExpr>(L);
 306     const ConstantExpr *RE = cast<ConstantExpr>(R);
 307     unsigned NumOperandsL = LE->getNumOperands();
 308     unsigned NumOperandsR = RE->getNumOperands();
 309     if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
 310       return Res;
 311     for (unsigned i = 0; i < NumOperandsL; ++i) {
 312       if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
 313                                  cast<Constant>(RE->getOperand(i))))
 314         return Res;
 315     }
 316     return 0;
 317   }
 318   case Value::BlockAddressVal: {
 319     const BlockAddress *LBA = cast<BlockAddress>(L);
 320     const BlockAddress *RBA = cast<BlockAddress>(R);
 321     if (int Res = cmpValues(LBA->getFunction(), RBA->getFunction()))
 322       return Res;
 323     if (LBA->getFunction() == RBA->getFunction()) {
 324       // They are BBs in the same function. Order by which comes first in the
 325       // BB order of the function. This order is deterministic.
 326       Function* F = LBA->getFunction();
 327       BasicBlock *LBB = LBA->getBasicBlock();
 328       BasicBlock *RBB = RBA->getBasicBlock();
 329       if (LBB == RBB)
 330         return 0;
 331       for(BasicBlock &BB : F->getBasicBlockList()) {
 332         if (&BB == LBB) {
 333           assert(&BB != RBB);
 334           return -1;
 335         }
 336         if (&BB == RBB)
 337           return 1;
 338       }
 339       llvm_unreachable("Basic Block Address does not point to a basic block in "
 340                        "its function.");
 341       return -1;
 342     } else {
 343       // cmpValues said the functions are the same. So because they aren't
 344       // literally the same pointer, they must respectively be the left and
 345       // right functions.
 346       assert(LBA->getFunction() == FnL && RBA->getFunction() == FnR);
 347       // cmpValues will tell us if these are equivalent BasicBlocks, in the
 348       // context of their respective functions.
 349       return cmpValues(LBA->getBasicBlock(), RBA->getBasicBlock());
 350     }
 351   }
 352   default: // Unknown constant, abort.
 353     DEBUG(dbgs() << "Looking at valueID " << L->getValueID() << "\n");
 354     llvm_unreachable("Constant ValueID not recognized.");
 355     return -1;
 356   }
 357 }
 358
 359 int FunctionComparator::cmpGlobalValues(GlobalValue *L, GlobalValue *R) const {
 360   uint64_t LNumber = GlobalNumbers->getNumber(L);
 361   uint64_t RNumber = GlobalNumbers->getNumber(R);
 362   return cmpNumbers(LNumber, RNumber);
 363 }
 364
 365 /// cmpType - compares two types,
 366 /// defines total ordering among the types set.
 367 /// See method declaration comments for more details.
 368 int FunctionComparator::cmpTypes(Type *TyL, Type *TyR) const {
 369   PointerType *PTyL = dyn_cast<PointerType>(TyL);
 370   PointerType *PTyR = dyn_cast<PointerType>(TyR);
 371
 372   const DataLayout &DL = FnL->getParent()->getDataLayout();
 373   if (PTyL && PTyL->getAddressSpace() == 0)
 374     TyL = DL.getIntPtrType(TyL);
 375   if (PTyR && PTyR->getAddressSpace() == 0)
 376     TyR = DL.getIntPtrType(TyR);
 377
 378   if (TyL == TyR)
 379     return 0;
 380
 381   if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
 382     return Res;
 383
 384   switch (TyL->getTypeID()) {
 385   default:
 386     llvm_unreachable("Unknown type!");
 387     // Fall through in Release mode.
 388     LLVM_FALLTHROUGH;
 389   case Type::IntegerTyID:
 390     return cmpNumbers(cast<IntegerType>(TyL)->getBitWidth(),
 391                       cast<IntegerType>(TyR)->getBitWidth());
 392   // TyL == TyR would have returned true earlier, because types are uniqued.
 393   case Type::VoidTyID:
 394   case Type::FloatTyID:
 395   case Type::DoubleTyID:
 396   case Type::X86_FP80TyID:
 397   case Type::FP128TyID:
 398   case Type::PPC_FP128TyID:
 399   case Type::LabelTyID:
 400   case Type::MetadataTyID:
 401   case Type::TokenTyID:
 402     return 0;
 403
 404   case Type::PointerTyID: {
 405     assert(PTyL && PTyR && "Both types must be pointers here.");
 406     return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
 407   }
 408
 409   case Type::StructTyID: {
 410     StructType *STyL = cast<StructType>(TyL);
 411     StructType *STyR = cast<StructType>(TyR);
 412     if (STyL->getNumElements() != STyR->getNumElements())
 413       return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
 414
 415     if (STyL->isPacked() != STyR->isPacked())
 416       return cmpNumbers(STyL->isPacked(), STyR->isPacked());
 417
 418     for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
 419       if (int Res = cmpTypes(STyL->getElementType(i), STyR->getElementType(i)))
 420         return Res;
 421     }
 422     return 0;
 423   }
 424
 425   case Type::FunctionTyID: {
 426     FunctionType *FTyL = cast<FunctionType>(TyL);
 427     FunctionType *FTyR = cast<FunctionType>(TyR);
 428     if (FTyL->getNumParams() != FTyR->getNumParams())
 429       return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
 430
 431     if (FTyL->isVarArg() != FTyR->isVarArg())
 432       return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
 433
 434     if (int Res = cmpTypes(FTyL->getReturnType(), FTyR->getReturnType()))
 435       return Res;
 436
 437     for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
 438       if (int Res = cmpTypes(FTyL->getParamType(i), FTyR->getParamType(i)))
 439         return Res;
 440     }
 441     return 0;
 442   }
 443
 444   case Type::ArrayTyID:
 445   case Type::VectorTyID: {
 446     auto *STyL = cast<SequentialType>(TyL);
 447     auto *STyR = cast<SequentialType>(TyR);
 448     if (STyL->getNumElements() != STyR->getNumElements())
 449       return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
 450     return cmpTypes(STyL->getElementType(), STyR->getElementType());
 451   }
 452   }
 453 }
 454
 455 // Determine whether the two operations are the same except that pointer-to-A
 456 // and pointer-to-B are equivalent. This should be kept in sync with
 457 // Instruction::isSameOperationAs.
 458 // Read method declaration comments for more details.
 459 int FunctionComparator::cmpOperations(const Instruction *L,
 460                                       const Instruction *R,
 461                                       bool &needToCmpOperands) const {
 462   needToCmpOperands = true;
 463   if (int Res = cmpValues(L, R))
 464     return Res;
 465
 466   // Differences from Instruction::isSameOperationAs:
 467   //  * replace type comparison with calls to cmpTypes.
 468   //  * we test for I->getRawSubclassOptionalData (nuw/nsw/tail) at the top.
 469   //  * because of the above, we don't test for the tail bit on calls later on.
 470   if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
 471     return Res;
 472
 473   if (const GetElementPtrInst *GEPL = dyn_cast<GetElementPtrInst>(L)) {
 474     needToCmpOperands = false;
 475     const GetElementPtrInst *GEPR = cast<GetElementPtrInst>(R);
 476     if (int Res =
 477             cmpValues(GEPL->getPointerOperand(), GEPR->getPointerOperand()))
 478       return Res;
 479     return cmpGEPs(GEPL, GEPR);
 480   }
 481
 482   if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
 483     return Res;
 484
 485   if (int Res = cmpTypes(L->getType(), R->getType()))
 486     return Res;
 487
 488   if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
 489                            R->getRawSubclassOptionalData()))
 490     return Res;
 491
 492   // We have two instructions of identical opcode and #operands.  Check to see
 493   // if all operands are the same type
 494   for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
 495     if (int Res =
 496             cmpTypes(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
 497       return Res;
 498   }
 499
 500   // Check special state that is a part of some instructions.
 501   if (const AllocaInst *AI = dyn_cast<AllocaInst>(L)) {
 502     if (int Res = cmpTypes(AI->getAllocatedType(),
 503                            cast<AllocaInst>(R)->getAllocatedType()))
 504       return Res;
 505     return cmpNumbers(AI->getAlignment(), cast<AllocaInst>(R)->getAlignment());
 506   }
 507   if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
 508     if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
 509       return Res;
 510     if (int Res =
 511             cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
 512       return Res;
 513     if (int Res =
 514             cmpOrderings(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
 515       return Res;
 516     if (int Res = cmpNumbers(LI->getSyncScopeID(),
 517                              cast<LoadInst>(R)->getSyncScopeID()))
 518       return Res;
 519     return cmpRangeMetadata(LI->getMetadata(LLVMContext::MD_range),
 520         cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
 521   }
 522   if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
 523     if (int Res =
 524             cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
 525       return Res;
 526     if (int Res =
 527             cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
 528       return Res;
 529     if (int Res =
 530             cmpOrderings(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
 531       return Res;
 532     return cmpNumbers(SI->getSyncScopeID(),
 533                       cast<StoreInst>(R)->getSyncScopeID());
 534   }
 535   if (const CmpInst *CI = dyn_cast<CmpInst>(L))
 536     return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
 537   if (const CallInst *CI = dyn_cast<CallInst>(L)) {
 538     if (int Res = cmpNumbers(CI->getCallingConv(),
 539                              cast<CallInst>(R)->getCallingConv()))
 540       return Res;
 541     if (int Res =
 542             cmpAttrs(CI->getAttributes(), cast<CallInst>(R)->getAttributes()))
 543       return Res;
 544     if (int Res = cmpOperandBundlesSchema(CI, R))
 545       return Res;
 546     return cmpRangeMetadata(
 547         CI->getMetadata(LLVMContext::MD_range),
 548         cast<CallInst>(R)->getMetadata(LLVMContext::MD_range));
 549   }
 550   if (const InvokeInst *II = dyn_cast<InvokeInst>(L)) {
 551     if (int Res = cmpNumbers(II->getCallingConv(),
 552                              cast<InvokeInst>(R)->getCallingConv()))
 553       return Res;
 554     if (int Res =
 555             cmpAttrs(II->getAttributes(), cast<InvokeInst>(R)->getAttributes()))
 556       return Res;
 557     if (int Res = cmpOperandBundlesSchema(II, R))
 558       return Res;
 559     return cmpRangeMetadata(
 560         II->getMetadata(LLVMContext::MD_range),
 561         cast<InvokeInst>(R)->getMetadata(LLVMContext::MD_range));
 562   }
 563   if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
 564     ArrayRef<unsigned> LIndices = IVI->getIndices();
 565     ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
 566     if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
 567       return Res;
 568     for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
 569       if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
 570         return Res;
 571     }
 572     return 0;
 573   }
 574   if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
 575     ArrayRef<unsigned> LIndices = EVI->getIndices();
 576     ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
 577     if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
 578       return Res;
 579     for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
 580       if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
 581         return Res;
 582     }
 583   }
 584   if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
 585     if (int Res =
 586             cmpOrderings(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
 587       return Res;
 588     return cmpNumbers(FI->getSyncScopeID(),
 589                       cast<FenceInst>(R)->getSyncScopeID());
 590   }
 591   if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
 592     if (int Res = cmpNumbers(CXI->isVolatile(),
 593                              cast<AtomicCmpXchgInst>(R)->isVolatile()))
 594       return Res;
 595     if (int Res = cmpNumbers(CXI->isWeak(),
 596                              cast<AtomicCmpXchgInst>(R)->isWeak()))
 597       return Res;
 598     if (int Res =
 599             cmpOrderings(CXI->getSuccessOrdering(),
 600                          cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
 601       return Res;
 602     if (int Res =
 603             cmpOrderings(CXI->getFailureOrdering(),
 604                          cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
 605       return Res;
 606     return cmpNumbers(CXI->getSyncScopeID(),
 607                       cast<AtomicCmpXchgInst>(R)->getSyncScopeID());
 608   }
 609   if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
 610     if (int Res = cmpNumbers(RMWI->getOperation(),
 611                              cast<AtomicRMWInst>(R)->getOperation()))
 612       return Res;
 613     if (int Res = cmpNumbers(RMWI->isVolatile(),
 614                              cast<AtomicRMWInst>(R)->isVolatile()))
 615       return Res;
 616     if (int Res = cmpOrderings(RMWI->getOrdering(),
 617                              cast<AtomicRMWInst>(R)->getOrdering()))
 618       return Res;
 619     return cmpNumbers(RMWI->getSyncScopeID(),
 620                       cast<AtomicRMWInst>(R)->getSyncScopeID());
 621   }
 622   if (const PHINode *PNL = dyn_cast<PHINode>(L)) {
 623     const PHINode *PNR = cast<PHINode>(R);
 624     // Ensure that in addition to the incoming values being identical
 625     // (checked by the caller of this function), the incoming blocks
 626     // are also identical.
 627     for (unsigned i = 0, e = PNL->getNumIncomingValues(); i != e; ++i) {
 628       if (int Res =
 629               cmpValues(PNL->getIncomingBlock(i), PNR->getIncomingBlock(i)))
 630         return Res;
 631     }
 632   }
 633   return 0;
 634 }
 635
 636 // Determine whether two GEP operations perform the same underlying arithmetic.
 637 // Read method declaration comments for more details.
 638 int FunctionComparator::cmpGEPs(const GEPOperator *GEPL,
 639                                 const GEPOperator *GEPR) const {
 640
 641   unsigned int ASL = GEPL->getPointerAddressSpace();
 642   unsigned int ASR = GEPR->getPointerAddressSpace();
 643
 644   if (int Res = cmpNumbers(ASL, ASR))
 645     return Res;
 646
 647   // When we have target data, we can reduce the GEP down to the value in bytes
 648   // added to the address.
 649   const DataLayout &DL = FnL->getParent()->getDataLayout();
 650   unsigned BitWidth = DL.getPointerSizeInBits(ASL);
 651   APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
 652   if (GEPL->accumulateConstantOffset(DL, OffsetL) &&
 653       GEPR->accumulateConstantOffset(DL, OffsetR))
 654     return cmpAPInts(OffsetL, OffsetR);
 655   if (int Res = cmpTypes(GEPL->getSourceElementType(),
 656                          GEPR->getSourceElementType()))
 657     return Res;
 658
 659   if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
 660     return Res;
 661
 662   for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
 663     if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
 664       return Res;
 665   }
 666
 667   return 0;
 668 }
 669
 670 int FunctionComparator::cmpInlineAsm(const InlineAsm *L,
 671                                      const InlineAsm *R) const {
 672   // InlineAsm's are uniqued. If they are the same pointer, obviously they are
 673   // the same, otherwise compare the fields.
 674   if (L == R)
 675     return 0;
 676   if (int Res = cmpTypes(L->getFunctionType(), R->getFunctionType()))
 677     return Res;
 678   if (int Res = cmpMem(L->getAsmString(), R->getAsmString()))
 679     return Res;
 680   if (int Res = cmpMem(L->getConstraintString(), R->getConstraintString()))
 681     return Res;
 682   if (int Res = cmpNumbers(L->hasSideEffects(), R->hasSideEffects()))
 683     return Res;
 684   if (int Res = cmpNumbers(L->isAlignStack(), R->isAlignStack()))
 685     return Res;
 686   if (int Res = cmpNumbers(L->getDialect(), R->getDialect()))
 687     return Res;
 688   llvm_unreachable("InlineAsm blocks were not uniqued.");
 689   return 0;
 690 }
 691
 692 /// Compare two values used by the two functions under pair-wise comparison. If
 693 /// this is the first time the values are seen, they're added to the mapping so
 694 /// that we will detect mismatches on next use.
 695 /// See comments in declaration for more details.
 696 int FunctionComparator::cmpValues(const Value *L, const Value *R) const {
 697   // Catch self-reference case.
 698   if (L == FnL) {
 699     if (R == FnR)
 700       return 0;
 701     return -1;
 702   }
 703   if (R == FnR) {
 704     if (L == FnL)
 705       return 0;
 706     return 1;
 707   }
 708
 709   const Constant *ConstL = dyn_cast<Constant>(L);
 710   const Constant *ConstR = dyn_cast<Constant>(R);
 711   if (ConstL && ConstR) {
 712     if (L == R)
 713       return 0;
 714     return cmpConstants(ConstL, ConstR);
 715   }
 716
 717   if (ConstL)
 718     return 1;
 719   if (ConstR)
 720     return -1;
 721
 722   const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
 723   const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
 724
 725   if (InlineAsmL && InlineAsmR)
 726     return cmpInlineAsm(InlineAsmL, InlineAsmR);
 727   if (InlineAsmL)
 728     return 1;
 729   if (InlineAsmR)
 730     return -1;
 731
 732   auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
 733        RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
 734
 735   return cmpNumbers(LeftSN.first->second, RightSN.first->second);
 736 }
 737
 738 // Test whether two basic blocks have equivalent behaviour.
 739 int FunctionComparator::cmpBasicBlocks(const BasicBlock *BBL,
 740                                        const BasicBlock *BBR) const {
 741   BasicBlock::const_iterator InstL = BBL->begin(), InstLE = BBL->end();
 742   BasicBlock::const_iterator InstR = BBR->begin(), InstRE = BBR->end();
 743
 744   do {
 745     bool needToCmpOperands = true;
 746     if (int Res = cmpOperations(&*InstL, &*InstR, needToCmpOperands))
 747       return Res;
 748     if (needToCmpOperands) {
 749       assert(InstL->getNumOperands() == InstR->getNumOperands());
 750
 751       for (unsigned i = 0, e = InstL->getNumOperands(); i != e; ++i) {
 752         Value *OpL = InstL->getOperand(i);
 753         Value *OpR = InstR->getOperand(i);
 754         if (int Res = cmpValues(OpL, OpR))
 755           return Res;
 756         // cmpValues should ensure this is true.
 757         assert(cmpTypes(OpL->getType(), OpR->getType()) == 0);
 758       }
 759     }
 760
 761     ++InstL;
 762     ++InstR;
 763   } while (InstL != InstLE && InstR != InstRE);
 764
 765   if (InstL != InstLE && InstR == InstRE)
 766     return 1;
 767   if (InstL == InstLE && InstR != InstRE)
 768     return -1;
 769   return 0;
 770 }
 771
 772 int FunctionComparator::compareSignature() const {
 773   if (int Res = cmpAttrs(FnL->getAttributes(), FnR->getAttributes()))
 774     return Res;
 775
 776   if (int Res = cmpNumbers(FnL->hasGC(), FnR->hasGC()))
 777     return Res;
 778
 779   if (FnL->hasGC()) {
 780     if (int Res = cmpMem(FnL->getGC(), FnR->getGC()))
 781       return Res;
 782   }
 783
 784   if (int Res = cmpNumbers(FnL->hasSection(), FnR->hasSection()))
 785     return Res;
 786
 787   if (FnL->hasSection()) {
 788     if (int Res = cmpMem(FnL->getSection(), FnR->getSection()))
 789       return Res;
 790   }
 791
 792   if (int Res = cmpNumbers(FnL->isVarArg(), FnR->isVarArg()))
 793     return Res;
 794
 795   // TODO: if it's internal and only used in direct calls, we could handle this
 796   // case too.
 797   if (int Res = cmpNumbers(FnL->getCallingConv(), FnR->getCallingConv()))
 798     return Res;
 799
 800   if (int Res = cmpTypes(FnL->getFunctionType(), FnR->getFunctionType()))
 801     return Res;
 802
 803   assert(FnL->arg_size() == FnR->arg_size() &&
 804          "Identically typed functions have different numbers of args!");
 805
 806   // Visit the arguments so that they get enumerated in the order they're
 807   // passed in.
 808   for (Function::const_arg_iterator ArgLI = FnL->arg_begin(),
 809        ArgRI = FnR->arg_begin(),
 810        ArgLE = FnL->arg_end();
 811        ArgLI != ArgLE; ++ArgLI, ++ArgRI) {
 812     if (cmpValues(&*ArgLI, &*ArgRI) != 0)
 813       llvm_unreachable("Arguments repeat!");
 814   }
 815   return 0;
 816 }
 817
 818 // Test whether the two functions have equivalent behaviour.
 819 int FunctionComparator::compare() {
 820   beginCompare();
 821
 822   if (int Res = compareSignature())
 823     return Res;
 824
 825   // We do a CFG-ordered walk since the actual ordering of the blocks in the
 826   // linked list is immaterial. Our walk starts at the entry block for both
 827   // functions, then takes each block from each terminator in order. As an
 828   // artifact, this also means that unreachable blocks are ignored.
 829   SmallVector<const BasicBlock *, 8> FnLBBs, FnRBBs;
 830   SmallPtrSet<const BasicBlock *, 32> VisitedBBs; // in terms of F1.
 831
 832   FnLBBs.push_back(&FnL->getEntryBlock());
 833   FnRBBs.push_back(&FnR->getEntryBlock());
 834
 835   VisitedBBs.insert(FnLBBs[0]);
 836   while (!FnLBBs.empty()) {
 837     const BasicBlock *BBL = FnLBBs.pop_back_val();
 838     const BasicBlock *BBR = FnRBBs.pop_back_val();
 839
 840     if (int Res = cmpValues(BBL, BBR))
 841       return Res;
 842
 843     if (int Res = cmpBasicBlocks(BBL, BBR))
 844       return Res;
 845
 846     const TerminatorInst *TermL = BBL->getTerminator();
 847     const TerminatorInst *TermR = BBR->getTerminator();
 848
 849     assert(TermL->getNumSuccessors() == TermR->getNumSuccessors());
 850     for (unsigned i = 0, e = TermL->getNumSuccessors(); i != e; ++i) {
 851       if (!VisitedBBs.insert(TermL->getSuccessor(i)).second)
 852         continue;
 853
 854       FnLBBs.push_back(TermL->getSuccessor(i));
 855       FnRBBs.push_back(TermR->getSuccessor(i));
 856     }
 857   }
 858   return 0;
 859 }
 860
 861 namespace {
 862
 863 // Accumulate the hash of a sequence of 64-bit integers. This is similar to a
 864 // hash of a sequence of 64bit ints, but the entire input does not need to be
 865 // available at once. This interface is necessary for functionHash because it
 866 // needs to accumulate the hash as the structure of the function is traversed
 867 // without saving these values to an intermediate buffer. This form of hashing
 868 // is not often needed, as usually the object to hash is just read from a
 869 // buffer.
 870 class HashAccumulator64 {
 871   uint64_t Hash;
 872 public:
 873   // Initialize to random constant, so the state isn't zero.
 874   HashAccumulator64() { Hash = 0x6acaa36bef8325c5ULL; }
 875   void add(uint64_t V) {
 876      Hash = llvm::hashing::detail::hash_16_bytes(Hash, V);
 877   }
 878   // No finishing is required, because the entire hash value is used.
 879   uint64_t getHash() { return Hash; }
 880 };
 881 } // end anonymous namespace
 882
 883 // A function hash is calculated by considering only the number of arguments and
 884 // whether a function is varargs, the order of basic blocks (given by the
 885 // successors of each basic block in depth first order), and the order of
 886 // opcodes of each instruction within each of these basic blocks. This mirrors
 887 // the strategy compare() uses to compare functions by walking the BBs in depth
 888 // first order and comparing each instruction in sequence. Because this hash
 889 // does not look at the operands, it is insensitive to things such as the
 890 // target of calls and the constants used in the function, which makes it useful
 891 // when possibly merging functions which are the same modulo constants and call
 892 // targets.
 893 FunctionComparator::FunctionHash FunctionComparator::functionHash(Function &F) {
 894   HashAccumulator64 H;
 895   H.add(F.isVarArg());
 896   H.add(F.arg_size());
 897
 898   SmallVector<const BasicBlock *, 8> BBs;
 899   SmallSet<const BasicBlock *, 16> VisitedBBs;
 900
 901   // Walk the blocks in the same order as FunctionComparator::cmpBasicBlocks(),
 902   // accumulating the hash of the function "structure." (BB and opcode sequence)
 903   BBs.push_back(&F.getEntryBlock());
 904   VisitedBBs.insert(BBs[0]);
 905   while (!BBs.empty()) {
 906     const BasicBlock *BB = BBs.pop_back_val();
 907     // This random value acts as a block header, as otherwise the partition of
 908     // opcodes into BBs wouldn't affect the hash, only the order of the opcodes
 909     H.add(45798);
 910     for (auto &Inst : *BB) {
 911       H.add(Inst.getOpcode());
 912     }
 913     const TerminatorInst *Term = BB->getTerminator();
 914     for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
 915       if (!VisitedBBs.insert(Term->getSuccessor(i)).second)
 916         continue;
 917       BBs.push_back(Term->getSuccessor(i));
 918     }
 919   }
 920   return H.getHash();
 921 }
 922
 923