lib/Transforms/Utils/FunctionComparator.cpp

   1 //===- FunctionComparator.h - Function Comparator -------------------------===//
   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 // This file implements the FunctionComparator and GlobalNumberState classes
  10 // which are used by the MergeFunctions pass for comparing functions.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "llvm/Transforms/Utils/FunctionComparator.h"
  15 #include "llvm/ADT/APFloat.h"
  16 #include "llvm/ADT/APInt.h"
  17 #include "llvm/ADT/ArrayRef.h"
  18 #include "llvm/ADT/Hashing.h"
  19 #include "llvm/ADT/SmallPtrSet.h"
  20 #include "llvm/ADT/SmallVector.h"
  21 #include "llvm/IR/Attributes.h"
  22 #include "llvm/IR/BasicBlock.h"
  23 #include "llvm/IR/CallSite.h"
  24 #include "llvm/IR/Constant.h"
  25 #include "llvm/IR/Constants.h"
  26 #include "llvm/IR/DataLayout.h"
  27 #include "llvm/IR/DerivedTypes.h"
  28 #include "llvm/IR/Function.h"
  29 #include "llvm/IR/GlobalValue.h"
  30 #include "llvm/IR/InlineAsm.h"
  31 #include "llvm/IR/InstrTypes.h"
  32 #include "llvm/IR/Instruction.h"
  33 #include "llvm/IR/Instructions.h"
  34 #include "llvm/IR/LLVMContext.h"
  35 #include "llvm/IR/Metadata.h"
  36 #include "llvm/IR/Module.h"
  37 #include "llvm/IR/Operator.h"
  38 #include "llvm/IR/Type.h"
  39 #include "llvm/IR/Value.h"
  40 #include "llvm/Support/Casting.h"
  41 #include "llvm/Support/Compiler.h"
  42 #include "llvm/Support/Debug.h"
  43 #include "llvm/Support/ErrorHandling.h"
  44 #include "llvm/Support/raw_ostream.h"
  45 #include <cassert>
  46 #include <cstddef>
  47 #include <cstdint>
  48 #include <utility>
  49
  50 using namespace llvm;
  51
  52 #define DEBUG_TYPE "functioncomparator"
  53
  54 int FunctionComparator::cmpNumbers(uint64_t L, uint64_t R) const {
  55   if (L < R) return -1;
  56   if (L > R) return 1;
  57   return 0;
  58 }
  59
  60 int FunctionComparator::cmpOrderings(AtomicOrdering L, AtomicOrdering R) const {
  61   if ((int)L < (int)R) return -1;
  62   if ((int)L > (int)R) return 1;
  63   return 0;
  64 }
  65
  66 int FunctionComparator::cmpAPInts(const APInt &L, const APInt &R) const {
  67   if (int Res = cmpNumbers(L.getBitWidth(), R.getBitWidth()))
  68     return Res;
  69   if (L.ugt(R)) return 1;
  70   if (R.ugt(L)) return -1;
  71   return 0;
  72 }
  73
  74 int FunctionComparator::cmpAPFloats(const APFloat &L, const APFloat &R) const {
  75   // Floats are ordered first by semantics (i.e. float, double, half, etc.),
  76   // then by value interpreted as a bitstring (aka APInt).
  77   const fltSemantics &SL = L.getSemantics(), &SR = R.getSemantics();
  78   if (int Res = cmpNumbers(APFloat::semanticsPrecision(SL),
  79                            APFloat::semanticsPrecision(SR)))
  80     return Res;
  81   if (int Res = cmpNumbers(APFloat::semanticsMaxExponent(SL),
  82                            APFloat::semanticsMaxExponent(SR)))
  83     return Res;
  84   if (int Res = cmpNumbers(APFloat::semanticsMinExponent(SL),
  85                            APFloat::semanticsMinExponent(SR)))
  86     return Res;
  87   if (int Res = cmpNumbers(APFloat::semanticsSizeInBits(SL),
  88                            APFloat::semanticsSizeInBits(SR)))
  89     return Res;
  90   return cmpAPInts(L.bitcastToAPInt(), R.bitcastToAPInt());
  91 }
  92
  93 int FunctionComparator::cmpMem(StringRef L, StringRef R) const {
  94   // Prevent heavy comparison, compare sizes first.
  95   if (int Res = cmpNumbers(L.size(), R.size()))
  96     return Res;
  97
  98   // Compare strings lexicographically only when it is necessary: only when
  99   // strings are equal in size.
 100   return L.compare(R);
 101 }
 102
 103 int FunctionComparator::cmpAttrs(const AttributeList L,
 104                                  const AttributeList R) const {
 105   if (int Res = cmpNumbers(L.getNumAttrSets(), R.getNumAttrSets()))
 106     return Res;
 107
 108   for (unsigned i = L.index_begin(), e = L.index_end(); i != e; ++i) {
 109     AttributeSet LAS = L.getAttributes(i);
 110     AttributeSet RAS = R.getAttributes(i);
 111     AttributeSet::iterator LI = LAS.begin(), LE = LAS.end();
 112     AttributeSet::iterator RI = RAS.begin(), RE = RAS.end();
 113     for (; LI != LE && RI != RE; ++LI, ++RI) {
 114       Attribute LA = *LI;
 115       Attribute RA = *RI;
 116       if (LA.isTypeAttribute() && RA.isTypeAttribute()) {
 117         if (LA.getKindAsEnum() != RA.getKindAsEnum())
 118           return cmpNumbers(LA.getKindAsEnum(), RA.getKindAsEnum());
 119
 120         Type *TyL = LA.getValueAsType();
 121         Type *TyR = RA.getValueAsType();
 122         if (TyL && TyR)
 123           return cmpTypes(TyL, TyR);
 124
 125         // Two pointers, at least one null, so the comparison result is
 126         // independent of the value of a real pointer.
 127         return cmpNumbers((uint64_t)TyL, (uint64_t)TyR);
 128       }
 129       if (LA < RA)
 130         return -1;
 131       if (RA < LA)
 132         return 1;
 133     }
 134     if (LI != LE)
 135       return 1;
 136     if (RI != RE)
 137       return -1;
 138   }
 139   return 0;
 140 }
 141
 142 int FunctionComparator::cmpRangeMetadata(const MDNode *L,
 143                                          const MDNode *R) const {
 144   if (L == R)
 145     return 0;
 146   if (!L)
 147     return -1;
 148   if (!R)
 149     return 1;
 150   // Range metadata is a sequence of numbers. Make sure they are the same
 151   // sequence.
 152   // TODO: Note that as this is metadata, it is possible to drop and/or merge
 153   // this data when considering functions to merge. Thus this comparison would
 154   // return 0 (i.e. equivalent), but merging would become more complicated
 155   // because the ranges would need to be unioned. It is not likely that
 156   // functions differ ONLY in this metadata if they are actually the same
 157   // function semantically.
 158   if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
 159     return Res;
 160   for (size_t I = 0; I < L->getNumOperands(); ++I) {
 161     ConstantInt *LLow = mdconst::extract<ConstantInt>(L->getOperand(I));
 162     ConstantInt *RLow = mdconst::extract<ConstantInt>(R->getOperand(I));
 163     if (int Res = cmpAPInts(LLow->getValue(), RLow->getValue()))
 164       return Res;
 165   }
 166   return 0;
 167 }
 168
 169 int FunctionComparator::cmpOperandBundlesSchema(const Instruction *L,
 170                                                 const Instruction *R) const {
 171   ImmutableCallSite LCS(L);
 172   ImmutableCallSite RCS(R);
 173
 174   assert(LCS && RCS && "Must be calls or invokes!");
 175   assert(LCS.isCall() == RCS.isCall() && "Can't compare otherwise!");
 176
 177   if (int Res =
 178           cmpNumbers(LCS.getNumOperandBundles(), RCS.getNumOperandBundles()))
 179     return Res;
 180
 181   for (unsigned i = 0, e = LCS.getNumOperandBundles(); i != e; ++i) {
 182     auto OBL = LCS.getOperandBundleAt(i);
 183     auto OBR = RCS.getOperandBundleAt(i);
 184
 185     if (int Res = OBL.getTagName().compare(OBR.getTagName()))
 186       return Res;
 187
 188     if (int Res = cmpNumbers(OBL.Inputs.size(), OBR.Inputs.size()))
 189       return Res;
 190   }
 191
 192   return 0;
 193 }
 194
 195 /// Constants comparison:
 196 /// 1. Check whether type of L constant could be losslessly bitcasted to R
 197 /// type.
 198 /// 2. Compare constant contents.
 199 /// For more details see declaration comments.
 200 int FunctionComparator::cmpConstants(const Constant *L,
 201                                      const Constant *R) const {
 202   Type *TyL = L->getType();
 203   Type *TyR = R->getType();
 204
 205   // Check whether types are bitcastable. This part is just re-factored
 206   // Type::canLosslesslyBitCastTo method, but instead of returning true/false,
 207   // we also pack into result which type is "less" for us.
 208   int TypesRes = cmpTypes(TyL, TyR);
 209   if (TypesRes != 0) {
 210     // Types are different, but check whether we can bitcast them.
 211     if (!TyL->isFirstClassType()) {
 212       if (TyR->isFirstClassType())
 213         return -1;
 214       // Neither TyL nor TyR are values of first class type. Return the result
 215       // of comparing the types
 216       return TypesRes;
 217     }
 218     if (!TyR->isFirstClassType()) {
 219       if (TyL->isFirstClassType())
 220         return 1;
 221       return TypesRes;
 222     }
 223
 224     // Vector -> Vector conversions are always lossless if the two vector types
 225     // have the same size, otherwise not.
 226     unsigned TyLWidth = 0;
 227     unsigned TyRWidth = 0;
 228
 229     if (auto *VecTyL = dyn_cast<VectorType>(TyL))
 230       TyLWidth = VecTyL->getBitWidth();
 231     if (auto *VecTyR = dyn_cast<VectorType>(TyR))
 232       TyRWidth = VecTyR->getBitWidth();
 233
 234     if (TyLWidth != TyRWidth)
 235       return cmpNumbers(TyLWidth, TyRWidth);
 236
 237     // Zero bit-width means neither TyL nor TyR are vectors.
 238     if (!TyLWidth) {
 239       PointerType *PTyL = dyn_cast<PointerType>(TyL);
 240       PointerType *PTyR = dyn_cast<PointerType>(TyR);
 241       if (PTyL && PTyR) {
 242         unsigned AddrSpaceL = PTyL->getAddressSpace();
 243         unsigned AddrSpaceR = PTyR->getAddressSpace();
 244         if (int Res = cmpNumbers(AddrSpaceL, AddrSpaceR))
 245           return Res;
 246       }
 247       if (PTyL)
 248         return 1;
 249       if (PTyR)
 250         return -1;
 251
 252       // TyL and TyR aren't vectors, nor pointers. We don't know how to
 253       // bitcast them.
 254       return TypesRes;
 255     }
 256   }
 257
 258   // OK, types are bitcastable, now check constant contents.
 259
 260   if (L->isNullValue() && R->isNullValue())
 261     return TypesRes;
 262   if (L->isNullValue() && !R->isNullValue())
 263     return 1;
 264   if (!L->isNullValue() && R->isNullValue())
 265     return -1;
 266
 267   auto GlobalValueL = const_cast<GlobalValue *>(dyn_cast<GlobalValue>(L));
 268   auto GlobalValueR = const_cast<GlobalValue *>(dyn_cast<GlobalValue>(R));
 269   if (GlobalValueL && GlobalValueR) {
 270     return cmpGlobalValues(GlobalValueL, GlobalValueR);
 271   }
 272
 273   if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))
 274     return Res;
 275
 276   if (const auto *SeqL = dyn_cast<ConstantDataSequential>(L)) {
 277     const auto *SeqR = cast<ConstantDataSequential>(R);
 278     // This handles ConstantDataArray and ConstantDataVector. Note that we
 279     // compare the two raw data arrays, which might differ depending on the host
 280     // endianness. This isn't a problem though, because the endiness of a module
 281     // will affect the order of the constants, but this order is the same
 282     // for a given input module and host platform.
 283     return cmpMem(SeqL->getRawDataValues(), SeqR->getRawDataValues());
 284   }
 285
 286   switch (L->getValueID()) {
 287   case Value::UndefValueVal:
 288   case Value::ConstantTokenNoneVal:
 289     return TypesRes;
 290   case Value::ConstantIntVal: {
 291     const APInt &LInt = cast<ConstantInt>(L)->getValue();
 292     const APInt &RInt = cast<ConstantInt>(R)->getValue();
 293     return cmpAPInts(LInt, RInt);
 294   }
 295   case Value::ConstantFPVal: {
 296     const APFloat &LAPF = cast<ConstantFP>(L)->getValueAPF();
 297     const APFloat &RAPF = cast<ConstantFP>(R)->getValueAPF();
 298     return cmpAPFloats(LAPF, RAPF);
 299   }
 300   case Value::ConstantArrayVal: {
 301     const ConstantArray *LA = cast<ConstantArray>(L);
 302     const ConstantArray *RA = cast<ConstantArray>(R);
 303     uint64_t NumElementsL = cast<ArrayType>(TyL)->getNumElements();
 304     uint64_t NumElementsR = cast<ArrayType>(TyR)->getNumElements();
 305     if (int Res = cmpNumbers(NumElementsL, NumElementsR))
 306       return Res;
 307     for (uint64_t i = 0; i < NumElementsL; ++i) {
 308       if (int Res = cmpConstants(cast<Constant>(LA->getOperand(i)),
 309                                  cast<Constant>(RA->getOperand(i))))
 310         return Res;
 311     }
 312     return 0;
 313   }
 314   case Value::ConstantStructVal: {
 315     const ConstantStruct *LS = cast<ConstantStruct>(L);
 316     const ConstantStruct *RS = cast<ConstantStruct>(R);
 317     unsigned NumElementsL = cast<StructType>(TyL)->getNumElements();
 318     unsigned NumElementsR = cast<StructType>(TyR)->getNumElements();
 319     if (int Res = cmpNumbers(NumElementsL, NumElementsR))
 320       return Res;
 321     for (unsigned i = 0; i != NumElementsL; ++i) {
 322       if (int Res = cmpConstants(cast<Constant>(LS->getOperand(i)),
 323                                  cast<Constant>(RS->getOperand(i))))
 324         return Res;
 325     }
 326     return 0;
 327   }
 328   case Value::ConstantVectorVal: {
 329     const ConstantVector *LV = cast<ConstantVector>(L);
 330     const ConstantVector *RV = cast<ConstantVector>(R);
 331     unsigned NumElementsL = cast<VectorType>(TyL)->getNumElements();
 332     unsigned NumElementsR = cast<VectorType>(TyR)->getNumElements();
 333     if (int Res = cmpNumbers(NumElementsL, NumElementsR))
 334       return Res;
 335     for (uint64_t i = 0; i < NumElementsL; ++i) {
 336       if (int Res = cmpConstants(cast<Constant>(LV->getOperand(i)),
 337                                  cast<Constant>(RV->getOperand(i))))
 338         return Res;
 339     }
 340     return 0;
 341   }
 342   case Value::ConstantExprVal: {
 343     const ConstantExpr *LE = cast<ConstantExpr>(L);
 344     const ConstantExpr *RE = cast<ConstantExpr>(R);
 345     unsigned NumOperandsL = LE->getNumOperands();
 346     unsigned NumOperandsR = RE->getNumOperands();
 347     if (int Res = cmpNumbers(NumOperandsL, NumOperandsR))
 348       return Res;
 349     for (unsigned i = 0; i < NumOperandsL; ++i) {
 350       if (int Res = cmpConstants(cast<Constant>(LE->getOperand(i)),
 351                                  cast<Constant>(RE->getOperand(i))))
 352         return Res;
 353     }
 354     return 0;
 355   }
 356   case Value::BlockAddressVal: {
 357     const BlockAddress *LBA = cast<BlockAddress>(L);
 358     const BlockAddress *RBA = cast<BlockAddress>(R);
 359     if (int Res = cmpValues(LBA->getFunction(), RBA->getFunction()))
 360       return Res;
 361     if (LBA->getFunction() == RBA->getFunction()) {
 362       // They are BBs in the same function. Order by which comes first in the
 363       // BB order of the function. This order is deterministic.
 364       Function* F = LBA->getFunction();
 365       BasicBlock *LBB = LBA->getBasicBlock();
 366       BasicBlock *RBB = RBA->getBasicBlock();
 367       if (LBB == RBB)
 368         return 0;
 369       for(BasicBlock &BB : F->getBasicBlockList()) {
 370         if (&BB == LBB) {
 371           assert(&BB != RBB);
 372           return -1;
 373         }
 374         if (&BB == RBB)
 375           return 1;
 376       }
 377       llvm_unreachable("Basic Block Address does not point to a basic block in "
 378                        "its function.");
 379       return -1;
 380     } else {
 381       // cmpValues said the functions are the same. So because they aren't
 382       // literally the same pointer, they must respectively be the left and
 383       // right functions.
 384       assert(LBA->getFunction() == FnL && RBA->getFunction() == FnR);
 385       // cmpValues will tell us if these are equivalent BasicBlocks, in the
 386       // context of their respective functions.
 387       return cmpValues(LBA->getBasicBlock(), RBA->getBasicBlock());
 388     }
 389   }
 390   default: // Unknown constant, abort.
 391     LLVM_DEBUG(dbgs() << "Looking at valueID " << L->getValueID() << "\n");
 392     llvm_unreachable("Constant ValueID not recognized.");
 393     return -1;
 394   }
 395 }
 396
 397 int FunctionComparator::cmpGlobalValues(GlobalValue *L, GlobalValue *R) const {
 398   uint64_t LNumber = GlobalNumbers->getNumber(L);
 399   uint64_t RNumber = GlobalNumbers->getNumber(R);
 400   return cmpNumbers(LNumber, RNumber);
 401 }
 402
 403 /// cmpType - compares two types,
 404 /// defines total ordering among the types set.
 405 /// See method declaration comments for more details.
 406 int FunctionComparator::cmpTypes(Type *TyL, Type *TyR) const {
 407   PointerType *PTyL = dyn_cast<PointerType>(TyL);
 408   PointerType *PTyR = dyn_cast<PointerType>(TyR);
 409
 410   const DataLayout &DL = FnL->getParent()->getDataLayout();
 411   if (PTyL && PTyL->getAddressSpace() == 0)
 412     TyL = DL.getIntPtrType(TyL);
 413   if (PTyR && PTyR->getAddressSpace() == 0)
 414     TyR = DL.getIntPtrType(TyR);
 415
 416   if (TyL == TyR)
 417     return 0;
 418
 419   if (int Res = cmpNumbers(TyL->getTypeID(), TyR->getTypeID()))
 420     return Res;
 421
 422   switch (TyL->getTypeID()) {
 423   default:
 424     llvm_unreachable("Unknown type!");
 425   case Type::IntegerTyID:
 426     return cmpNumbers(cast<IntegerType>(TyL)->getBitWidth(),
 427                       cast<IntegerType>(TyR)->getBitWidth());
 428   // TyL == TyR would have returned true earlier, because types are uniqued.
 429   case Type::VoidTyID:
 430   case Type::FloatTyID:
 431   case Type::DoubleTyID:
 432   case Type::X86_FP80TyID:
 433   case Type::FP128TyID:
 434   case Type::PPC_FP128TyID:
 435   case Type::LabelTyID:
 436   case Type::MetadataTyID:
 437   case Type::TokenTyID:
 438     return 0;
 439
 440   case Type::PointerTyID:
 441     assert(PTyL && PTyR && "Both types must be pointers here.");
 442     return cmpNumbers(PTyL->getAddressSpace(), PTyR->getAddressSpace());
 443
 444   case Type::StructTyID: {
 445     StructType *STyL = cast<StructType>(TyL);
 446     StructType *STyR = cast<StructType>(TyR);
 447     if (STyL->getNumElements() != STyR->getNumElements())
 448       return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
 449
 450     if (STyL->isPacked() != STyR->isPacked())
 451       return cmpNumbers(STyL->isPacked(), STyR->isPacked());
 452
 453     for (unsigned i = 0, e = STyL->getNumElements(); i != e; ++i) {
 454       if (int Res = cmpTypes(STyL->getElementType(i), STyR->getElementType(i)))
 455         return Res;
 456     }
 457     return 0;
 458   }
 459
 460   case Type::FunctionTyID: {
 461     FunctionType *FTyL = cast<FunctionType>(TyL);
 462     FunctionType *FTyR = cast<FunctionType>(TyR);
 463     if (FTyL->getNumParams() != FTyR->getNumParams())
 464       return cmpNumbers(FTyL->getNumParams(), FTyR->getNumParams());
 465
 466     if (FTyL->isVarArg() != FTyR->isVarArg())
 467       return cmpNumbers(FTyL->isVarArg(), FTyR->isVarArg());
 468
 469     if (int Res = cmpTypes(FTyL->getReturnType(), FTyR->getReturnType()))
 470       return Res;
 471
 472     for (unsigned i = 0, e = FTyL->getNumParams(); i != e; ++i) {
 473       if (int Res = cmpTypes(FTyL->getParamType(i), FTyR->getParamType(i)))
 474         return Res;
 475     }
 476     return 0;
 477   }
 478
 479   case Type::ArrayTyID:
 480   case Type::VectorTyID: {
 481     auto *STyL = cast<SequentialType>(TyL);
 482     auto *STyR = cast<SequentialType>(TyR);
 483     if (STyL->getNumElements() != STyR->getNumElements())
 484       return cmpNumbers(STyL->getNumElements(), STyR->getNumElements());
 485     return cmpTypes(STyL->getElementType(), STyR->getElementType());
 486   }
 487   }
 488 }
 489
 490 // Determine whether the two operations are the same except that pointer-to-A
 491 // and pointer-to-B are equivalent. This should be kept in sync with
 492 // Instruction::isSameOperationAs.
 493 // Read method declaration comments for more details.
 494 int FunctionComparator::cmpOperations(const Instruction *L,
 495                                       const Instruction *R,
 496                                       bool &needToCmpOperands) const {
 497   needToCmpOperands = true;
 498   if (int Res = cmpValues(L, R))
 499     return Res;
 500
 501   // Differences from Instruction::isSameOperationAs:
 502   //  * replace type comparison with calls to cmpTypes.
 503   //  * we test for I->getRawSubclassOptionalData (nuw/nsw/tail) at the top.
 504   //  * because of the above, we don't test for the tail bit on calls later on.
 505   if (int Res = cmpNumbers(L->getOpcode(), R->getOpcode()))
 506     return Res;
 507
 508   if (const GetElementPtrInst *GEPL = dyn_cast<GetElementPtrInst>(L)) {
 509     needToCmpOperands = false;
 510     const GetElementPtrInst *GEPR = cast<GetElementPtrInst>(R);
 511     if (int Res =
 512             cmpValues(GEPL->getPointerOperand(), GEPR->getPointerOperand()))
 513       return Res;
 514     return cmpGEPs(GEPL, GEPR);
 515   }
 516
 517   if (int Res = cmpNumbers(L->getNumOperands(), R->getNumOperands()))
 518     return Res;
 519
 520   if (int Res = cmpTypes(L->getType(), R->getType()))
 521     return Res;
 522
 523   if (int Res = cmpNumbers(L->getRawSubclassOptionalData(),
 524                            R->getRawSubclassOptionalData()))
 525     return Res;
 526
 527   // We have two instructions of identical opcode and #operands.  Check to see
 528   // if all operands are the same type
 529   for (unsigned i = 0, e = L->getNumOperands(); i != e; ++i) {
 530     if (int Res =
 531             cmpTypes(L->getOperand(i)->getType(), R->getOperand(i)->getType()))
 532       return Res;
 533   }
 534
 535   // Check special state that is a part of some instructions.
 536   if (const AllocaInst *AI = dyn_cast<AllocaInst>(L)) {
 537     if (int Res = cmpTypes(AI->getAllocatedType(),
 538                            cast<AllocaInst>(R)->getAllocatedType()))
 539       return Res;
 540     return cmpNumbers(AI->getAlignment(), cast<AllocaInst>(R)->getAlignment());
 541   }
 542   if (const LoadInst *LI = dyn_cast<LoadInst>(L)) {
 543     if (int Res = cmpNumbers(LI->isVolatile(), cast<LoadInst>(R)->isVolatile()))
 544       return Res;
 545     if (int Res =
 546             cmpNumbers(LI->getAlignment(), cast<LoadInst>(R)->getAlignment()))
 547       return Res;
 548     if (int Res =
 549             cmpOrderings(LI->getOrdering(), cast<LoadInst>(R)->getOrdering()))
 550       return Res;
 551     if (int Res = cmpNumbers(LI->getSyncScopeID(),
 552                              cast<LoadInst>(R)->getSyncScopeID()))
 553       return Res;
 554     return cmpRangeMetadata(LI->getMetadata(LLVMContext::MD_range),
 555         cast<LoadInst>(R)->getMetadata(LLVMContext::MD_range));
 556   }
 557   if (const StoreInst *SI = dyn_cast<StoreInst>(L)) {
 558     if (int Res =
 559             cmpNumbers(SI->isVolatile(), cast<StoreInst>(R)->isVolatile()))
 560       return Res;
 561     if (int Res =
 562             cmpNumbers(SI->getAlignment(), cast<StoreInst>(R)->getAlignment()))
 563       return Res;
 564     if (int Res =
 565             cmpOrderings(SI->getOrdering(), cast<StoreInst>(R)->getOrdering()))
 566       return Res;
 567     return cmpNumbers(SI->getSyncScopeID(),
 568                       cast<StoreInst>(R)->getSyncScopeID());
 569   }
 570   if (const CmpInst *CI = dyn_cast<CmpInst>(L))
 571     return cmpNumbers(CI->getPredicate(), cast<CmpInst>(R)->getPredicate());
 572   if (auto CSL = CallSite(const_cast<Instruction *>(L))) {
 573     auto CSR = CallSite(const_cast<Instruction *>(R));
 574     if (int Res = cmpNumbers(CSL.getCallingConv(), CSR.getCallingConv()))
 575       return Res;
 576     if (int Res = cmpAttrs(CSL.getAttributes(), CSR.getAttributes()))
 577       return Res;
 578     if (int Res = cmpOperandBundlesSchema(L, R))
 579       return Res;
 580     if (const CallInst *CI = dyn_cast<CallInst>(L))
 581       if (int Res = cmpNumbers(CI->getTailCallKind(),
 582                                cast<CallInst>(R)->getTailCallKind()))
 583         return Res;
 584     return cmpRangeMetadata(L->getMetadata(LLVMContext::MD_range),
 585                             R->getMetadata(LLVMContext::MD_range));
 586   }
 587   if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(L)) {
 588     ArrayRef<unsigned> LIndices = IVI->getIndices();
 589     ArrayRef<unsigned> RIndices = cast<InsertValueInst>(R)->getIndices();
 590     if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
 591       return Res;
 592     for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
 593       if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
 594         return Res;
 595     }
 596     return 0;
 597   }
 598   if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(L)) {
 599     ArrayRef<unsigned> LIndices = EVI->getIndices();
 600     ArrayRef<unsigned> RIndices = cast<ExtractValueInst>(R)->getIndices();
 601     if (int Res = cmpNumbers(LIndices.size(), RIndices.size()))
 602       return Res;
 603     for (size_t i = 0, e = LIndices.size(); i != e; ++i) {
 604       if (int Res = cmpNumbers(LIndices[i], RIndices[i]))
 605         return Res;
 606     }
 607   }
 608   if (const FenceInst *FI = dyn_cast<FenceInst>(L)) {
 609     if (int Res =
 610             cmpOrderings(FI->getOrdering(), cast<FenceInst>(R)->getOrdering()))
 611       return Res;
 612     return cmpNumbers(FI->getSyncScopeID(),
 613                       cast<FenceInst>(R)->getSyncScopeID());
 614   }
 615   if (const AtomicCmpXchgInst *CXI = dyn_cast<AtomicCmpXchgInst>(L)) {
 616     if (int Res = cmpNumbers(CXI->isVolatile(),
 617                              cast<AtomicCmpXchgInst>(R)->isVolatile()))
 618       return Res;
 619     if (int Res = cmpNumbers(CXI->isWeak(),
 620                              cast<AtomicCmpXchgInst>(R)->isWeak()))
 621       return Res;
 622     if (int Res =
 623             cmpOrderings(CXI->getSuccessOrdering(),
 624                          cast<AtomicCmpXchgInst>(R)->getSuccessOrdering()))
 625       return Res;
 626     if (int Res =
 627             cmpOrderings(CXI->getFailureOrdering(),
 628                          cast<AtomicCmpXchgInst>(R)->getFailureOrdering()))
 629       return Res;
 630     return cmpNumbers(CXI->getSyncScopeID(),
 631                       cast<AtomicCmpXchgInst>(R)->getSyncScopeID());
 632   }
 633   if (const AtomicRMWInst *RMWI = dyn_cast<AtomicRMWInst>(L)) {
 634     if (int Res = cmpNumbers(RMWI->getOperation(),
 635                              cast<AtomicRMWInst>(R)->getOperation()))
 636       return Res;
 637     if (int Res = cmpNumbers(RMWI->isVolatile(),
 638                              cast<AtomicRMWInst>(R)->isVolatile()))
 639       return Res;
 640     if (int Res = cmpOrderings(RMWI->getOrdering(),
 641                              cast<AtomicRMWInst>(R)->getOrdering()))
 642       return Res;
 643     return cmpNumbers(RMWI->getSyncScopeID(),
 644                       cast<AtomicRMWInst>(R)->getSyncScopeID());
 645   }
 646   if (const PHINode *PNL = dyn_cast<PHINode>(L)) {
 647     const PHINode *PNR = cast<PHINode>(R);
 648     // Ensure that in addition to the incoming values being identical
 649     // (checked by the caller of this function), the incoming blocks
 650     // are also identical.
 651     for (unsigned i = 0, e = PNL->getNumIncomingValues(); i != e; ++i) {
 652       if (int Res =
 653               cmpValues(PNL->getIncomingBlock(i), PNR->getIncomingBlock(i)))
 654         return Res;
 655     }
 656   }
 657   return 0;
 658 }
 659
 660 // Determine whether two GEP operations perform the same underlying arithmetic.
 661 // Read method declaration comments for more details.
 662 int FunctionComparator::cmpGEPs(const GEPOperator *GEPL,
 663                                 const GEPOperator *GEPR) const {
 664   unsigned int ASL = GEPL->getPointerAddressSpace();
 665   unsigned int ASR = GEPR->getPointerAddressSpace();
 666
 667   if (int Res = cmpNumbers(ASL, ASR))
 668     return Res;
 669
 670   // When we have target data, we can reduce the GEP down to the value in bytes
 671   // added to the address.
 672   const DataLayout &DL = FnL->getParent()->getDataLayout();
 673   unsigned BitWidth = DL.getPointerSizeInBits(ASL);
 674   APInt OffsetL(BitWidth, 0), OffsetR(BitWidth, 0);
 675   if (GEPL->accumulateConstantOffset(DL, OffsetL) &&
 676       GEPR->accumulateConstantOffset(DL, OffsetR))
 677     return cmpAPInts(OffsetL, OffsetR);
 678   if (int Res = cmpTypes(GEPL->getSourceElementType(),
 679                          GEPR->getSourceElementType()))
 680     return Res;
 681
 682   if (int Res = cmpNumbers(GEPL->getNumOperands(), GEPR->getNumOperands()))
 683     return Res;
 684
 685   for (unsigned i = 0, e = GEPL->getNumOperands(); i != e; ++i) {
 686     if (int Res = cmpValues(GEPL->getOperand(i), GEPR->getOperand(i)))
 687       return Res;
 688   }
 689
 690   return 0;
 691 }
 692
 693 int FunctionComparator::cmpInlineAsm(const InlineAsm *L,
 694                                      const InlineAsm *R) const {
 695   // InlineAsm's are uniqued. If they are the same pointer, obviously they are
 696   // the same, otherwise compare the fields.
 697   if (L == R)
 698     return 0;
 699   if (int Res = cmpTypes(L->getFunctionType(), R->getFunctionType()))
 700     return Res;
 701   if (int Res = cmpMem(L->getAsmString(), R->getAsmString()))
 702     return Res;
 703   if (int Res = cmpMem(L->getConstraintString(), R->getConstraintString()))
 704     return Res;
 705   if (int Res = cmpNumbers(L->hasSideEffects(), R->hasSideEffects()))
 706     return Res;
 707   if (int Res = cmpNumbers(L->isAlignStack(), R->isAlignStack()))
 708     return Res;
 709   if (int Res = cmpNumbers(L->getDialect(), R->getDialect()))
 710     return Res;
 711   assert(L->getFunctionType() != R->getFunctionType());
 712   return 0;
 713 }
 714
 715 /// Compare two values used by the two functions under pair-wise comparison. If
 716 /// this is the first time the values are seen, they're added to the mapping so
 717 /// that we will detect mismatches on next use.
 718 /// See comments in declaration for more details.
 719 int FunctionComparator::cmpValues(const Value *L, const Value *R) const {
 720   // Catch self-reference case.
 721   if (L == FnL) {
 722     if (R == FnR)
 723       return 0;
 724     return -1;
 725   }
 726   if (R == FnR) {
 727     if (L == FnL)
 728       return 0;
 729     return 1;
 730   }
 731
 732   const Constant *ConstL = dyn_cast<Constant>(L);
 733   const Constant *ConstR = dyn_cast<Constant>(R);
 734   if (ConstL && ConstR) {
 735     if (L == R)
 736       return 0;
 737     return cmpConstants(ConstL, ConstR);
 738   }
 739
 740   if (ConstL)
 741     return 1;
 742   if (ConstR)
 743     return -1;
 744
 745   const InlineAsm *InlineAsmL = dyn_cast<InlineAsm>(L);
 746   const InlineAsm *InlineAsmR = dyn_cast<InlineAsm>(R);
 747
 748   if (InlineAsmL && InlineAsmR)
 749     return cmpInlineAsm(InlineAsmL, InlineAsmR);
 750   if (InlineAsmL)
 751     return 1;
 752   if (InlineAsmR)
 753     return -1;
 754
 755   auto LeftSN = sn_mapL.insert(std::make_pair(L, sn_mapL.size())),
 756        RightSN = sn_mapR.insert(std::make_pair(R, sn_mapR.size()));
 757
 758   return cmpNumbers(LeftSN.first->second, RightSN.first->second);
 759 }
 760
 761 // Test whether two basic blocks have equivalent behaviour.
 762 int FunctionComparator::cmpBasicBlocks(const BasicBlock *BBL,
 763                                        const BasicBlock *BBR) const {
 764   BasicBlock::const_iterator InstL = BBL->begin(), InstLE = BBL->end();
 765   BasicBlock::const_iterator InstR = BBR->begin(), InstRE = BBR->end();
 766
 767   do {
 768     bool needToCmpOperands = true;
 769     if (int Res = cmpOperations(&*InstL, &*InstR, needToCmpOperands))
 770       return Res;
 771     if (needToCmpOperands) {
 772       assert(InstL->getNumOperands() == InstR->getNumOperands());
 773
 774       for (unsigned i = 0, e = InstL->getNumOperands(); i != e; ++i) {
 775         Value *OpL = InstL->getOperand(i);
 776         Value *OpR = InstR->getOperand(i);
 777         if (int Res = cmpValues(OpL, OpR))
 778           return Res;
 779         // cmpValues should ensure this is true.
 780         assert(cmpTypes(OpL->getType(), OpR->getType()) == 0);
 781       }
 782     }
 783
 784     ++InstL;
 785     ++InstR;
 786   } while (InstL != InstLE && InstR != InstRE);
 787
 788   if (InstL != InstLE && InstR == InstRE)
 789     return 1;
 790   if (InstL == InstLE && InstR != InstRE)
 791     return -1;
 792   return 0;
 793 }
 794
 795 int FunctionComparator::compareSignature() const {
 796   if (int Res = cmpAttrs(FnL->getAttributes(), FnR->getAttributes()))
 797     return Res;
 798
 799   if (int Res = cmpNumbers(FnL->hasGC(), FnR->hasGC()))
 800     return Res;
 801
 802   if (FnL->hasGC()) {
 803     if (int Res = cmpMem(FnL->getGC(), FnR->getGC()))
 804       return Res;
 805   }
 806
 807   if (int Res = cmpNumbers(FnL->hasSection(), FnR->hasSection()))
 808     return Res;
 809
 810   if (FnL->hasSection()) {
 811     if (int Res = cmpMem(FnL->getSection(), FnR->getSection()))
 812       return Res;
 813   }
 814
 815   if (int Res = cmpNumbers(FnL->isVarArg(), FnR->isVarArg()))
 816     return Res;
 817
 818   // TODO: if it's internal and only used in direct calls, we could handle this
 819   // case too.
 820   if (int Res = cmpNumbers(FnL->getCallingConv(), FnR->getCallingConv()))
 821     return Res;
 822
 823   if (int Res = cmpTypes(FnL->getFunctionType(), FnR->getFunctionType()))
 824     return Res;
 825
 826   assert(FnL->arg_size() == FnR->arg_size() &&
 827          "Identically typed functions have different numbers of args!");
 828
 829   // Visit the arguments so that they get enumerated in the order they're
 830   // passed in.
 831   for (Function::const_arg_iterator ArgLI = FnL->arg_begin(),
 832        ArgRI = FnR->arg_begin(),
 833        ArgLE = FnL->arg_end();
 834        ArgLI != ArgLE; ++ArgLI, ++ArgRI) {
 835     if (cmpValues(&*ArgLI, &*ArgRI) != 0)
 836       llvm_unreachable("Arguments repeat!");
 837   }
 838   return 0;
 839 }
 840
 841 // Test whether the two functions have equivalent behaviour.
 842 int FunctionComparator::compare() {
 843   beginCompare();
 844
 845   if (int Res = compareSignature())
 846     return Res;
 847
 848   // We do a CFG-ordered walk since the actual ordering of the blocks in the
 849   // linked list is immaterial. Our walk starts at the entry block for both
 850   // functions, then takes each block from each terminator in order. As an
 851   // artifact, this also means that unreachable blocks are ignored.
 852   SmallVector<const BasicBlock *, 8> FnLBBs, FnRBBs;
 853   SmallPtrSet<const BasicBlock *, 32> VisitedBBs; // in terms of F1.
 854
 855   FnLBBs.push_back(&FnL->getEntryBlock());
 856   FnRBBs.push_back(&FnR->getEntryBlock());
 857
 858   VisitedBBs.insert(FnLBBs[0]);
 859   while (!FnLBBs.empty()) {
 860     const BasicBlock *BBL = FnLBBs.pop_back_val();
 861     const BasicBlock *BBR = FnRBBs.pop_back_val();
 862
 863     if (int Res = cmpValues(BBL, BBR))
 864       return Res;
 865
 866     if (int Res = cmpBasicBlocks(BBL, BBR))
 867       return Res;
 868
 869     const Instruction *TermL = BBL->getTerminator();
 870     const Instruction *TermR = BBR->getTerminator();
 871
 872     assert(TermL->getNumSuccessors() == TermR->getNumSuccessors());
 873     for (unsigned i = 0, e = TermL->getNumSuccessors(); i != e; ++i) {
 874       if (!VisitedBBs.insert(TermL->getSuccessor(i)).second)
 875         continue;
 876
 877       FnLBBs.push_back(TermL->getSuccessor(i));
 878       FnRBBs.push_back(TermR->getSuccessor(i));
 879     }
 880   }
 881   return 0;
 882 }
 883
 884 namespace {
 885
 886 // Accumulate the hash of a sequence of 64-bit integers. This is similar to a
 887 // hash of a sequence of 64bit ints, but the entire input does not need to be
 888 // available at once. This interface is necessary for functionHash because it
 889 // needs to accumulate the hash as the structure of the function is traversed
 890 // without saving these values to an intermediate buffer. This form of hashing
 891 // is not often needed, as usually the object to hash is just read from a
 892 // buffer.
 893 class HashAccumulator64 {
 894   uint64_t Hash;
 895
 896 public:
 897   // Initialize to random constant, so the state isn't zero.
 898   HashAccumulator64() { Hash = 0x6acaa36bef8325c5ULL; }
 899
 900   void add(uint64_t V) {
 901      Hash = hashing::detail::hash_16_bytes(Hash, V);
 902   }
 903
 904   // No finishing is required, because the entire hash value is used.
 905   uint64_t getHash() { return Hash; }
 906 };
 907
 908 } // end anonymous namespace
 909
 910 // A function hash is calculated by considering only the number of arguments and
 911 // whether a function is varargs, the order of basic blocks (given by the
 912 // successors of each basic block in depth first order), and the order of
 913 // opcodes of each instruction within each of these basic blocks. This mirrors
 914 // the strategy compare() uses to compare functions by walking the BBs in depth
 915 // first order and comparing each instruction in sequence. Because this hash
 916 // does not look at the operands, it is insensitive to things such as the
 917 // target of calls and the constants used in the function, which makes it useful
 918 // when possibly merging functions which are the same modulo constants and call
 919 // targets.
 920 FunctionComparator::FunctionHash FunctionComparator::functionHash(Function &F) {
 921   HashAccumulator64 H;
 922   H.add(F.isVarArg());
 923   H.add(F.arg_size());
 924
 925   SmallVector<const BasicBlock *, 8> BBs;
 926   SmallPtrSet<const BasicBlock *, 16> VisitedBBs;
 927
 928   // Walk the blocks in the same order as FunctionComparator::cmpBasicBlocks(),
 929   // accumulating the hash of the function "structure." (BB and opcode sequence)
 930   BBs.push_back(&F.getEntryBlock());
 931   VisitedBBs.insert(BBs[0]);
 932   while (!BBs.empty()) {
 933     const BasicBlock *BB = BBs.pop_back_val();
 934     // This random value acts as a block header, as otherwise the partition of
 935     // opcodes into BBs wouldn't affect the hash, only the order of the opcodes
 936     H.add(45798);
 937     for (auto &Inst : *BB) {
 938       H.add(Inst.getOpcode());
 939     }
 940     const Instruction *Term = BB->getTerminator();
 941     for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) {
 942       if (!VisitedBBs.insert(Term->getSuccessor(i)).second)
 943         continue;
 944       BBs.push_back(Term->getSuccessor(i));
 945     }
 946   }
 947   return H.getHash();
 948 }