lib/Transforms/Scalar/IndVarSimplify.cpp

   1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===//
   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 transformation analyzes and transforms the induction variables (and
  11 // computations derived from them) into simpler forms suitable for subsequent
  12 // analysis and transformation.
  13 //
  14 // This transformation makes the following changes to each loop with an
  15 // identifiable induction variable:
  16 //   1. All loops are transformed to have a SINGLE canonical induction variable
  17 //      which starts at zero and steps by one.
  18 //   2. The canonical induction variable is guaranteed to be the first PHI node
  19 //      in the loop header block.
  20 //   3. Any pointer arithmetic recurrences are raised to use array subscripts.
  21 //
  22 // If the trip count of a loop is computable, this pass also makes the following
  23 // changes:
  24 //   1. The exit condition for the loop is canonicalized to compare the
  25 //      induction value against the exit value.  This turns loops like:
  26 //        'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)'
  27 //   2. Any use outside of the loop of an expression derived from the indvar
  28 //      is changed to compute the derived value outside of the loop, eliminating
  29 //      the dependence on the exit value of the induction variable.  If the only
  30 //      purpose of the loop is to compute the exit value of some derived
  31 //      expression, this transformation will make the loop dead.
  32 //
  33 // This transformation should be followed by strength reduction after all of the
  34 // desired loop transformations have been performed.  Additionally, on targets
  35 // where it is profitable, the loop could be transformed to count down to zero
  36 // (the "do loop" optimization).
  37 //
  38 //===----------------------------------------------------------------------===//
  39
  40 #define DEBUG_TYPE "indvars"
  41 #include "llvm/Transforms/Scalar.h"
  42 #include "llvm/BasicBlock.h"
  43 #include "llvm/Constants.h"
  44 #include "llvm/Instructions.h"
  45 #include "llvm/Type.h"
  46 #include "llvm/Analysis/ScalarEvolutionExpander.h"
  47 #include "llvm/Analysis/LoopInfo.h"
  48 #include "llvm/Analysis/LoopPass.h"
  49 #include "llvm/Support/CFG.h"
  50 #include "llvm/Support/Compiler.h"
  51 #include "llvm/Support/Debug.h"
  52 #include "llvm/Support/GetElementPtrTypeIterator.h"
  53 #include "llvm/Transforms/Utils/Local.h"
  54 #include "llvm/Support/CommandLine.h"
  55 #include "llvm/ADT/SmallVector.h"
  56 #include "llvm/ADT/SetVector.h"
  57 #include "llvm/ADT/SmallPtrSet.h"
  58 #include "llvm/ADT/Statistic.h"
  59 using namespace llvm;
  60
  61 STATISTIC(NumRemoved , "Number of aux indvars removed");
  62 STATISTIC(NumInserted, "Number of canonical indvars added");
  63 STATISTIC(NumReplaced, "Number of exit values replaced");
  64 STATISTIC(NumLFTR    , "Number of loop exit tests replaced");
  65
  66 namespace {
  67   class VISIBILITY_HIDDEN IndVarSimplify : public LoopPass {
  68     LoopInfo        *LI;
  69     ScalarEvolution *SE;
  70     bool Changed;
  71   public:
  72
  73    static char ID; // Pass identification, replacement for typeid
  74    IndVarSimplify() : LoopPass(&ID) {}
  75
  76    virtual bool runOnLoop(Loop *L, LPPassManager &LPM);
  77
  78    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
  79      AU.addRequired<ScalarEvolution>();
  80      AU.addRequiredID(LCSSAID);
  81      AU.addRequiredID(LoopSimplifyID);
  82      AU.addRequired<LoopInfo>();
  83      AU.addPreserved<ScalarEvolution>();
  84      AU.addPreservedID(LoopSimplifyID);
  85      AU.addPreservedID(LCSSAID);
  86      AU.setPreservesCFG();
  87    }
  88
  89   private:
  90
  91     void RewriteNonIntegerIVs(Loop *L);
  92
  93     void LinearFunctionTestReplace(Loop *L, SCEVHandle BackedgeTakenCount,
  94                                    Value *IndVar,
  95                                    BasicBlock *ExitingBlock,
  96                                    BranchInst *BI,
  97                                    SCEVExpander &Rewriter);
  98     void RewriteLoopExitValues(Loop *L, const SCEV *BackedgeTakenCount);
  99
 100     void DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*, 16> &Insts);
 101
 102     void HandleFloatingPointIV(Loop *L, PHINode *PH,
 103                                SmallPtrSet<Instruction*, 16> &DeadInsts);
 104   };
 105 }
 106
 107 char IndVarSimplify::ID = 0;
 108 static RegisterPass<IndVarSimplify>
 109 X("indvars", "Canonicalize Induction Variables");
 110
 111 Pass *llvm::createIndVarSimplifyPass() {
 112   return new IndVarSimplify();
 113 }
 114
 115 /// DeleteTriviallyDeadInstructions - If any of the instructions is the
 116 /// specified set are trivially dead, delete them and see if this makes any of
 117 /// their operands subsequently dead.
 118 void IndVarSimplify::
 119 DeleteTriviallyDeadInstructions(SmallPtrSet<Instruction*, 16> &Insts) {
 120   while (!Insts.empty()) {
 121     Instruction *I = *Insts.begin();
 122     Insts.erase(I);
 123     if (isInstructionTriviallyDead(I)) {
 124       for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
 125         if (Instruction *U = dyn_cast<Instruction>(I->getOperand(i)))
 126           Insts.insert(U);
 127       DOUT << "INDVARS: Deleting: " << *I;
 128       I->eraseFromParent();
 129       Changed = true;
 130     }
 131   }
 132 }
 133
 134 /// LinearFunctionTestReplace - This method rewrites the exit condition of the
 135 /// loop to be a canonical != comparison against the incremented loop induction
 136 /// variable.  This pass is able to rewrite the exit tests of any loop where the
 137 /// SCEV analysis can determine a loop-invariant trip count of the loop, which
 138 /// is actually a much broader range than just linear tests.
 139 void IndVarSimplify::LinearFunctionTestReplace(Loop *L,
 140                                    SCEVHandle BackedgeTakenCount,
 141                                    Value *IndVar,
 142                                    BasicBlock *ExitingBlock,
 143                                    BranchInst *BI,
 144                                    SCEVExpander &Rewriter) {
 145   // If the exiting block is not the same as the backedge block, we must compare
 146   // against the preincremented value, otherwise we prefer to compare against
 147   // the post-incremented value.
 148   Value *CmpIndVar;
 149   SCEVHandle RHS = BackedgeTakenCount;
 150   if (ExitingBlock == L->getLoopLatch()) {
 151     // Add one to the "backedge-taken" count to get the trip count.
 152     // If this addition may overflow, we have to be more pessimistic and
 153     // cast the induction variable before doing the add.
 154     SCEVHandle Zero = SE->getIntegerSCEV(0, BackedgeTakenCount->getType());
 155     SCEVHandle N =
 156       SE->getAddExpr(BackedgeTakenCount,
 157                      SE->getIntegerSCEV(1, BackedgeTakenCount->getType()));
 158     if ((isa<SCEVConstant>(N) && !N->isZero()) ||
 159         SE->isLoopGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) {
 160       // No overflow. Cast the sum.
 161       RHS = SE->getTruncateOrZeroExtend(N, IndVar->getType());
 162     } else {
 163       // Potential overflow. Cast before doing the add.
 164       RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
 165                                         IndVar->getType());
 166       RHS = SE->getAddExpr(RHS,
 167                            SE->getIntegerSCEV(1, IndVar->getType()));
 168     }
 169
 170     // The BackedgeTaken expression contains the number of times that the
 171     // backedge branches to the loop header.  This is one less than the
 172     // number of times the loop executes, so use the incremented indvar.
 173     CmpIndVar = L->getCanonicalInductionVariableIncrement();
 174   } else {
 175     // We have to use the preincremented value...
 176     RHS = SE->getTruncateOrZeroExtend(BackedgeTakenCount,
 177                                       IndVar->getType());
 178     CmpIndVar = IndVar;
 179   }
 180
 181   // Expand the code for the iteration count into the preheader of the loop.
 182   BasicBlock *Preheader = L->getLoopPreheader();
 183   Value *ExitCnt = Rewriter.expandCodeFor(RHS, IndVar->getType(),
 184                                           Preheader->getTerminator());
 185
 186   // Insert a new icmp_ne or icmp_eq instruction before the branch.
 187   ICmpInst::Predicate Opcode;
 188   if (L->contains(BI->getSuccessor(0)))
 189     Opcode = ICmpInst::ICMP_NE;
 190   else
 191     Opcode = ICmpInst::ICMP_EQ;
 192
 193   DOUT << "INDVARS: Rewriting loop exit condition to:\n"
 194        << "      LHS:" << *CmpIndVar // includes a newline
 195        << "       op:\t"
 196        << (Opcode == ICmpInst::ICMP_NE ? "!=" : "==") << "\n"
 197        << "      RHS:\t" << *RHS << "\n";
 198
 199   Value *Cond = new ICmpInst(Opcode, CmpIndVar, ExitCnt, "exitcond", BI);
 200   BI->setCondition(Cond);
 201   ++NumLFTR;
 202   Changed = true;
 203 }
 204
 205 /// RewriteLoopExitValues - Check to see if this loop has a computable
 206 /// loop-invariant execution count.  If so, this means that we can compute the
 207 /// final value of any expressions that are recurrent in the loop, and
 208 /// substitute the exit values from the loop into any instructions outside of
 209 /// the loop that use the final values of the current expressions.
 210 void IndVarSimplify::RewriteLoopExitValues(Loop *L,
 211                                            const SCEV *BackedgeTakenCount) {
 212   BasicBlock *Preheader = L->getLoopPreheader();
 213
 214   // Scan all of the instructions in the loop, looking at those that have
 215   // extra-loop users and which are recurrences.
 216   SCEVExpander Rewriter(*SE, *LI);
 217
 218   // We insert the code into the preheader of the loop if the loop contains
 219   // multiple exit blocks, or in the exit block if there is exactly one.
 220   BasicBlock *BlockToInsertInto;
 221   SmallVector<BasicBlock*, 8> ExitBlocks;
 222   L->getUniqueExitBlocks(ExitBlocks);
 223   if (ExitBlocks.size() == 1)
 224     BlockToInsertInto = ExitBlocks[0];
 225   else
 226     BlockToInsertInto = Preheader;
 227   BasicBlock::iterator InsertPt = BlockToInsertInto->getFirstNonPHI();
 228
 229   bool HasConstantItCount = isa<SCEVConstant>(BackedgeTakenCount);
 230
 231   SmallPtrSet<Instruction*, 16> InstructionsToDelete;
 232   std::map<Instruction*, Value*> ExitValues;
 233
 234   // Find all values that are computed inside the loop, but used outside of it.
 235   // Because of LCSSA, these values will only occur in LCSSA PHI Nodes.  Scan
 236   // the exit blocks of the loop to find them.
 237   for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
 238     BasicBlock *ExitBB = ExitBlocks[i];
 239
 240     // If there are no PHI nodes in this exit block, then no values defined
 241     // inside the loop are used on this path, skip it.
 242     PHINode *PN = dyn_cast<PHINode>(ExitBB->begin());
 243     if (!PN) continue;
 244
 245     unsigned NumPreds = PN->getNumIncomingValues();
 246
 247     // Iterate over all of the PHI nodes.
 248     BasicBlock::iterator BBI = ExitBB->begin();
 249     while ((PN = dyn_cast<PHINode>(BBI++))) {
 250
 251       // Iterate over all of the values in all the PHI nodes.
 252       for (unsigned i = 0; i != NumPreds; ++i) {
 253         // If the value being merged in is not integer or is not defined
 254         // in the loop, skip it.
 255         Value *InVal = PN->getIncomingValue(i);
 256         if (!isa<Instruction>(InVal) ||
 257             // SCEV only supports integer expressions for now.
 258             (!isa<IntegerType>(InVal->getType()) &&
 259              !isa<PointerType>(InVal->getType())))
 260           continue;
 261
 262         // If this pred is for a subloop, not L itself, skip it.
 263         if (LI->getLoopFor(PN->getIncomingBlock(i)) != L)
 264           continue; // The Block is in a subloop, skip it.
 265
 266         // Check that InVal is defined in the loop.
 267         Instruction *Inst = cast<Instruction>(InVal);
 268         if (!L->contains(Inst->getParent()))
 269           continue;
 270
 271         // We require that this value either have a computable evolution or that
 272         // the loop have a constant iteration count.  In the case where the loop
 273         // has a constant iteration count, we can sometimes force evaluation of
 274         // the exit value through brute force.
 275         SCEVHandle SH = SE->getSCEV(Inst);
 276         if (!SH->hasComputableLoopEvolution(L) && !HasConstantItCount)
 277           continue;          // Cannot get exit evolution for the loop value.
 278
 279         // Okay, this instruction has a user outside of the current loop
 280         // and varies predictably *inside* the loop.  Evaluate the value it
 281         // contains when the loop exits, if possible.
 282         SCEVHandle ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop());
 283         if (isa<SCEVCouldNotCompute>(ExitValue) ||
 284             !ExitValue->isLoopInvariant(L))
 285           continue;
 286
 287         Changed = true;
 288         ++NumReplaced;
 289
 290         // See if we already computed the exit value for the instruction, if so,
 291         // just reuse it.
 292         Value *&ExitVal = ExitValues[Inst];
 293         if (!ExitVal)
 294           ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), InsertPt);
 295
 296         DOUT << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal
 297              << "  LoopVal = " << *Inst << "\n";
 298
 299         PN->setIncomingValue(i, ExitVal);
 300
 301         // If this instruction is dead now, schedule it to be removed.
 302         if (Inst->use_empty())
 303           InstructionsToDelete.insert(Inst);
 304
 305         // See if this is a single-entry LCSSA PHI node.  If so, we can (and
 306         // have to) remove
 307         // the PHI entirely.  This is safe, because the NewVal won't be variant
 308         // in the loop, so we don't need an LCSSA phi node anymore.
 309         if (NumPreds == 1) {
 310           PN->replaceAllUsesWith(ExitVal);
 311           PN->eraseFromParent();
 312           break;
 313         }
 314       }
 315     }
 316   }
 317
 318   DeleteTriviallyDeadInstructions(InstructionsToDelete);
 319 }
 320
 321 void IndVarSimplify::RewriteNonIntegerIVs(Loop *L) {
 322   // First step.  Check to see if there are any floating-point recurrences.
 323   // If there are, change them into integer recurrences, permitting analysis by
 324   // the SCEV routines.
 325   //
 326   BasicBlock *Header    = L->getHeader();
 327
 328   SmallPtrSet<Instruction*, 16> DeadInsts;
 329   for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
 330     PHINode *PN = cast<PHINode>(I);
 331     HandleFloatingPointIV(L, PN, DeadInsts);
 332   }
 333
 334   // If the loop previously had floating-point IV, ScalarEvolution
 335   // may not have been able to compute a trip count. Now that we've done some
 336   // re-writing, the trip count may be computable.
 337   if (Changed)
 338     SE->forgetLoopBackedgeTakenCount(L);
 339
 340   if (!DeadInsts.empty())
 341     DeleteTriviallyDeadInstructions(DeadInsts);
 342 }
 343
 344 /// getEffectiveIndvarType - Determine the widest type that the
 345 /// induction-variable PHINode Phi is cast to.
 346 ///
 347 static const Type *getEffectiveIndvarType(const PHINode *Phi,
 348                                           const ScalarEvolution *SE) {
 349   const Type *Ty = Phi->getType();
 350
 351   for (Value::use_const_iterator UI = Phi->use_begin(), UE = Phi->use_end();
 352        UI != UE; ++UI) {
 353     const Type *CandidateType = NULL;
 354     if (const ZExtInst *ZI = dyn_cast<ZExtInst>(UI))
 355       CandidateType = ZI->getDestTy();
 356     else if (const SExtInst *SI = dyn_cast<SExtInst>(UI))
 357       CandidateType = SI->getDestTy();
 358     else if (const IntToPtrInst *IP = dyn_cast<IntToPtrInst>(UI))
 359       CandidateType = IP->getDestTy();
 360     else if (const PtrToIntInst *PI = dyn_cast<PtrToIntInst>(UI))
 361       CandidateType = PI->getDestTy();
 362     if (CandidateType &&
 363         SE->isSCEVable(CandidateType) &&
 364         SE->getTypeSizeInBits(CandidateType) > SE->getTypeSizeInBits(Ty))
 365       Ty = CandidateType;
 366   }
 367
 368   return Ty;
 369 }
 370
 371 /// TestOrigIVForWrap - Analyze the original induction variable that
 372 /// controls the loop's iteration to determine whether it would ever
 373 /// undergo signed or unsigned overflow.
 374 ///
 375 /// In addition to setting the NoSignedWrap and NoUnsignedWrap
 376 /// variables to true when appropriate (they are not set to false here),
 377 /// return the PHI for this induction variable.  Also record the initial
 378 /// and final values and the increment; these are not meaningful unless
 379 /// either NoSignedWrap or NoUnsignedWrap is true, and are always meaningful
 380 /// in that case, although the final value may be 0 indicating a nonconstant.
 381 ///
 382 /// TODO: This duplicates a fair amount of ScalarEvolution logic.
 383 /// Perhaps this can be merged with
 384 /// ScalarEvolution::getBackedgeTakenCount
 385 /// and/or ScalarEvolution::get{Sign,Zero}ExtendExpr.
 386 ///
 387 static const PHINode *TestOrigIVForWrap(const Loop *L,
 388                                         const BranchInst *BI,
 389                                         const Instruction *OrigCond,
 390                                         const ScalarEvolution &SE,
 391                                         bool &NoSignedWrap,
 392                                         bool &NoUnsignedWrap,
 393                                         const ConstantInt* &InitialVal,
 394                                         const ConstantInt* &IncrVal,
 395                                         const ConstantInt* &LimitVal) {
 396   // Verify that the loop is sane and find the exit condition.
 397   const ICmpInst *Cmp = dyn_cast<ICmpInst>(OrigCond);
 398   if (!Cmp) return 0;
 399
 400   const Value *CmpLHS = Cmp->getOperand(0);
 401   const Value *CmpRHS = Cmp->getOperand(1);
 402   const BasicBlock *TrueBB = BI->getSuccessor(0);
 403   const BasicBlock *FalseBB = BI->getSuccessor(1);
 404   ICmpInst::Predicate Pred = Cmp->getPredicate();
 405
 406   // Canonicalize a constant to the RHS.
 407   if (isa<ConstantInt>(CmpLHS)) {
 408     Pred = ICmpInst::getSwappedPredicate(Pred);
 409     std::swap(CmpLHS, CmpRHS);
 410   }
 411   // Canonicalize SLE to SLT.
 412   if (Pred == ICmpInst::ICMP_SLE)
 413     if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
 414       if (!CI->getValue().isMaxSignedValue()) {
 415         CmpRHS = ConstantInt::get(CI->getValue() + 1);
 416         Pred = ICmpInst::ICMP_SLT;
 417       }
 418   // Canonicalize SGT to SGE.
 419   if (Pred == ICmpInst::ICMP_SGT)
 420     if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
 421       if (!CI->getValue().isMaxSignedValue()) {
 422         CmpRHS = ConstantInt::get(CI->getValue() + 1);
 423         Pred = ICmpInst::ICMP_SGE;
 424       }
 425   // Canonicalize SGE to SLT.
 426   if (Pred == ICmpInst::ICMP_SGE) {
 427     std::swap(TrueBB, FalseBB);
 428     Pred = ICmpInst::ICMP_SLT;
 429   }
 430   // Canonicalize ULE to ULT.
 431   if (Pred == ICmpInst::ICMP_ULE)
 432     if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
 433       if (!CI->getValue().isMaxValue()) {
 434         CmpRHS = ConstantInt::get(CI->getValue() + 1);
 435         Pred = ICmpInst::ICMP_ULT;
 436       }
 437   // Canonicalize UGT to UGE.
 438   if (Pred == ICmpInst::ICMP_UGT)
 439     if (const ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS))
 440       if (!CI->getValue().isMaxValue()) {
 441         CmpRHS = ConstantInt::get(CI->getValue() + 1);
 442         Pred = ICmpInst::ICMP_UGE;
 443       }
 444   // Canonicalize UGE to ULT.
 445   if (Pred == ICmpInst::ICMP_UGE) {
 446     std::swap(TrueBB, FalseBB);
 447     Pred = ICmpInst::ICMP_ULT;
 448   }
 449   // For now, analyze only LT loops for signed overflow.
 450   if (Pred != ICmpInst::ICMP_SLT && Pred != ICmpInst::ICMP_ULT)
 451     return 0;
 452
 453   bool isSigned = Pred == ICmpInst::ICMP_SLT;
 454
 455   // Get the increment instruction. Look past casts if we will
 456   // be able to prove that the original induction variable doesn't
 457   // undergo signed or unsigned overflow, respectively.
 458   const Value *IncrInst = CmpLHS;
 459   if (isSigned) {
 460     if (const SExtInst *SI = dyn_cast<SExtInst>(CmpLHS)) {
 461       if (!isa<ConstantInt>(CmpRHS) ||
 462           !cast<ConstantInt>(CmpRHS)->getValue()
 463             .isSignedIntN(SE.getTypeSizeInBits(IncrInst->getType())))
 464         return 0;
 465       IncrInst = SI->getOperand(0);
 466     }
 467   } else {
 468     if (const ZExtInst *ZI = dyn_cast<ZExtInst>(CmpLHS)) {
 469       if (!isa<ConstantInt>(CmpRHS) ||
 470           !cast<ConstantInt>(CmpRHS)->getValue()
 471             .isIntN(SE.getTypeSizeInBits(IncrInst->getType())))
 472         return 0;
 473       IncrInst = ZI->getOperand(0);
 474     }
 475   }
 476
 477   // For now, only analyze induction variables that have simple increments.
 478   const BinaryOperator *IncrOp = dyn_cast<BinaryOperator>(IncrInst);
 479   if (!IncrOp || IncrOp->getOpcode() != Instruction::Add)
 480     return 0;
 481   IncrVal = dyn_cast<ConstantInt>(IncrOp->getOperand(1));
 482   if (!IncrVal)
 483     return 0;
 484
 485   // Make sure the PHI looks like a normal IV.
 486   const PHINode *PN = dyn_cast<PHINode>(IncrOp->getOperand(0));
 487   if (!PN || PN->getNumIncomingValues() != 2)
 488     return 0;
 489   unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0));
 490   unsigned BackEdge = !IncomingEdge;
 491   if (!L->contains(PN->getIncomingBlock(BackEdge)) ||
 492       PN->getIncomingValue(BackEdge) != IncrOp)
 493     return 0;
 494   if (!L->contains(TrueBB))
 495     return 0;
 496
 497   // For now, only analyze loops with a constant start value, so that
 498   // we can easily determine if the start value is not a maximum value
 499   // which would wrap on the first iteration.
 500   InitialVal = dyn_cast<ConstantInt>(PN->getIncomingValue(IncomingEdge));
 501   if (!InitialVal)
 502     return 0;
 503
 504   // The upper limit need not be a constant; we'll check later.
 505   LimitVal = dyn_cast<ConstantInt>(CmpRHS);
 506
 507   // We detect the impossibility of wrapping in two cases, both of
 508   // which require starting with a non-max value:
 509   // - The IV counts up by one, and the loop iterates only while it remains
 510   // less than a limiting value (any) in the same type.
 511   // - The IV counts up by a positive increment other than 1, and the
 512   // constant limiting value + the increment is less than the max value
 513   // (computed as max-increment to avoid overflow)
 514   if (isSigned && !InitialVal->getValue().isMaxSignedValue()) {
 515     if (IncrVal->equalsInt(1))
 516       NoSignedWrap = true;    // LimitVal need not be constant
 517     else if (LimitVal) {
 518       uint64_t numBits = LimitVal->getValue().getBitWidth();
 519       if (IncrVal->getValue().sgt(APInt::getNullValue(numBits)) &&
 520           (APInt::getSignedMaxValue(numBits) - IncrVal->getValue())
 521             .sgt(LimitVal->getValue()))
 522         NoSignedWrap = true;
 523     }
 524   } else if (!isSigned && !InitialVal->getValue().isMaxValue()) {
 525     if (IncrVal->equalsInt(1))
 526       NoUnsignedWrap = true;  // LimitVal need not be constant
 527     else if (LimitVal) {
 528       uint64_t numBits = LimitVal->getValue().getBitWidth();
 529       if (IncrVal->getValue().ugt(APInt::getNullValue(numBits)) &&
 530           (APInt::getMaxValue(numBits) - IncrVal->getValue())
 531             .ugt(LimitVal->getValue()))
 532         NoUnsignedWrap = true;
 533     }
 534   }
 535   return PN;
 536 }
 537
 538 static Value *getSignExtendedTruncVar(const SCEVAddRecExpr *AR,
 539                                       ScalarEvolution *SE,
 540                                       const Type *LargestType, Loop *L,
 541                                       const Type *myType,
 542                                       SCEVExpander &Rewriter) {
 543   SCEVHandle ExtendedStart =
 544     SE->getSignExtendExpr(AR->getStart(), LargestType);
 545   SCEVHandle ExtendedStep =
 546     SE->getSignExtendExpr(AR->getStepRecurrence(*SE), LargestType);
 547   SCEVHandle ExtendedAddRec =
 548     SE->getAddRecExpr(ExtendedStart, ExtendedStep, L);
 549   if (LargestType != myType)
 550     ExtendedAddRec = SE->getTruncateExpr(ExtendedAddRec, myType);
 551   return Rewriter.expandCodeFor(ExtendedAddRec, myType);
 552 }
 553
 554 static Value *getZeroExtendedTruncVar(const SCEVAddRecExpr *AR,
 555                                       ScalarEvolution *SE,
 556                                       const Type *LargestType, Loop *L,
 557                                       const Type *myType,
 558                                       SCEVExpander &Rewriter) {
 559   SCEVHandle ExtendedStart =
 560     SE->getZeroExtendExpr(AR->getStart(), LargestType);
 561   SCEVHandle ExtendedStep =
 562     SE->getZeroExtendExpr(AR->getStepRecurrence(*SE), LargestType);
 563   SCEVHandle ExtendedAddRec =
 564     SE->getAddRecExpr(ExtendedStart, ExtendedStep, L);
 565   if (LargestType != myType)
 566     ExtendedAddRec = SE->getTruncateExpr(ExtendedAddRec, myType);
 567   return Rewriter.expandCodeFor(ExtendedAddRec, myType);
 568 }
 569
 570 /// allUsesAreSameTyped - See whether all Uses of I are instructions
 571 /// with the same Opcode and the same type.
 572 static bool allUsesAreSameTyped(unsigned int Opcode, Instruction *I) {
 573   const Type* firstType = NULL;
 574   for (Value::use_iterator UI = I->use_begin(), UE = I->use_end();
 575        UI != UE; ++UI) {
 576     Instruction *II = dyn_cast<Instruction>(*UI);
 577     if (!II || II->getOpcode() != Opcode)
 578       return false;
 579     if (!firstType)
 580       firstType = II->getType();
 581     else if (firstType != II->getType())
 582       return false;
 583   }
 584   return true;
 585 }
 586
 587 bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
 588   LI = &getAnalysis<LoopInfo>();
 589   SE = &getAnalysis<ScalarEvolution>();
 590   Changed = false;
 591
 592   // If there are any floating-point recurrences, attempt to
 593   // transform them to use integer recurrences.
 594   RewriteNonIntegerIVs(L);
 595
 596   BasicBlock *Header       = L->getHeader();
 597   BasicBlock *ExitingBlock = L->getExitingBlock();
 598   SmallPtrSet<Instruction*, 16> DeadInsts;
 599
 600   // Verify the input to the pass in already in LCSSA form.
 601   assert(L->isLCSSAForm());
 602
 603   // Check to see if this loop has a computable loop-invariant execution count.
 604   // If so, this means that we can compute the final value of any expressions
 605   // that are recurrent in the loop, and substitute the exit values from the
 606   // loop into any instructions outside of the loop that use the final values of
 607   // the current expressions.
 608   //
 609   SCEVHandle BackedgeTakenCount = SE->getBackedgeTakenCount(L);
 610   if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount))
 611     RewriteLoopExitValues(L, BackedgeTakenCount);
 612
 613   // Next, analyze all of the induction variables in the loop, canonicalizing
 614   // auxillary induction variables.
 615   std::vector<std::pair<PHINode*, SCEVHandle> > IndVars;
 616
 617   for (BasicBlock::iterator I = Header->begin(); isa<PHINode>(I); ++I) {
 618     PHINode *PN = cast<PHINode>(I);
 619     if (SE->isSCEVable(PN->getType())) {
 620       SCEVHandle SCEV = SE->getSCEV(PN);
 621       // FIXME: It is an extremely bad idea to indvar substitute anything more
 622       // complex than affine induction variables.  Doing so will put expensive
 623       // polynomial evaluations inside of the loop, and the str reduction pass
 624       // currently can only reduce affine polynomials.  For now just disable
 625       // indvar subst on anything more complex than an affine addrec.
 626       if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SCEV))
 627         if (AR->getLoop() == L && AR->isAffine())
 628           IndVars.push_back(std::make_pair(PN, SCEV));
 629     }
 630   }
 631
 632   // Compute the type of the largest recurrence expression, and collect
 633   // the set of the types of the other recurrence expressions.
 634   const Type *LargestType = 0;
 635   SmallSetVector<const Type *, 4> SizesToInsert;
 636   if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) {
 637     LargestType = BackedgeTakenCount->getType();
 638     LargestType = SE->getEffectiveSCEVType(LargestType);
 639     SizesToInsert.insert(LargestType);
 640   }
 641   for (unsigned i = 0, e = IndVars.size(); i != e; ++i) {
 642     const PHINode *PN = IndVars[i].first;
 643     const Type *PNTy = PN->getType();
 644     PNTy = SE->getEffectiveSCEVType(PNTy);
 645     SizesToInsert.insert(PNTy);
 646     const Type *EffTy = getEffectiveIndvarType(PN, SE);
 647     EffTy = SE->getEffectiveSCEVType(EffTy);
 648     SizesToInsert.insert(EffTy);
 649     if (!LargestType ||
 650         SE->getTypeSizeInBits(EffTy) >
 651           SE->getTypeSizeInBits(LargestType))
 652       LargestType = EffTy;
 653   }
 654
 655   // Create a rewriter object which we'll use to transform the code with.
 656   SCEVExpander Rewriter(*SE, *LI);
 657
 658   // Now that we know the largest of of the induction variables in this loop,
 659   // insert a canonical induction variable of the largest size.
 660   Value *IndVar = 0;
 661   if (!SizesToInsert.empty()) {
 662     IndVar = Rewriter.getOrInsertCanonicalInductionVariable(L,LargestType);
 663     ++NumInserted;
 664     Changed = true;
 665     DOUT << "INDVARS: New CanIV: " << *IndVar;
 666   }
 667
 668   // If we have a trip count expression, rewrite the loop's exit condition
 669   // using it.  We can currently only handle loops with a single exit.
 670   bool NoSignedWrap = false;
 671   bool NoUnsignedWrap = false;
 672   const ConstantInt* InitialVal, * IncrVal, * LimitVal;
 673   const PHINode *OrigControllingPHI = 0;
 674   if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount) && ExitingBlock)
 675     // Can't rewrite non-branch yet.
 676     if (BranchInst *BI = dyn_cast<BranchInst>(ExitingBlock->getTerminator())) {
 677       if (Instruction *OrigCond = dyn_cast<Instruction>(BI->getCondition())) {
 678         // Determine if the OrigIV will ever undergo overflow.
 679         OrigControllingPHI =
 680           TestOrigIVForWrap(L, BI, OrigCond, *SE,
 681                             NoSignedWrap, NoUnsignedWrap,
 682                             InitialVal, IncrVal, LimitVal);
 683
 684         // We'll be replacing the original condition, so it'll be dead.
 685         DeadInsts.insert(OrigCond);
 686       }
 687
 688       LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar,
 689                                 ExitingBlock, BI, Rewriter);
 690     }
 691
 692   // Now that we have a canonical induction variable, we can rewrite any
 693   // recurrences in terms of the induction variable.  Start with the auxillary
 694   // induction variables, and recursively rewrite any of their uses.
 695   BasicBlock::iterator InsertPt = Header->getFirstNonPHI();
 696   Rewriter.setInsertionPoint(InsertPt);
 697
 698   // If there were induction variables of other sizes, cast the primary
 699   // induction variable to the right size for them, avoiding the need for the
 700   // code evaluation methods to insert induction variables of different sizes.
 701   for (unsigned i = 0, e = SizesToInsert.size(); i != e; ++i) {
 702     const Type *Ty = SizesToInsert[i];
 703     if (Ty != LargestType) {
 704       Instruction *New = new TruncInst(IndVar, Ty, "indvar", InsertPt);
 705       Rewriter.addInsertedValue(New, SE->getSCEV(New));
 706       DOUT << "INDVARS: Made trunc IV for type " << *Ty << ": "
 707            << *New << "\n";
 708     }
 709   }
 710
 711   // Rewrite all induction variables in terms of the canonical induction
 712   // variable.
 713   while (!IndVars.empty()) {
 714     PHINode *PN = IndVars.back().first;
 715     const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(IndVars.back().second);
 716     Value *NewVal = Rewriter.expandCodeFor(AR, PN->getType());
 717     DOUT << "INDVARS: Rewrote IV '" << *AR << "' " << *PN
 718          << "   into = " << *NewVal << "\n";
 719     NewVal->takeName(PN);
 720
 721     /// If the new canonical induction variable is wider than the original,
 722     /// and the original has uses that are casts to wider types, see if the
 723     /// truncate and extend can be omitted.
 724     if (PN == OrigControllingPHI && PN->getType() != LargestType)
 725       for (Value::use_iterator UI = PN->use_begin(), UE = PN->use_end();
 726            UI != UE; ++UI) {
 727         Instruction *UInst = dyn_cast<Instruction>(*UI);
 728         if (UInst && isa<SExtInst>(UInst) && NoSignedWrap) {
 729           Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType, L,
 730                                          UInst->getType(), Rewriter);
 731           UInst->replaceAllUsesWith(TruncIndVar);
 732           DeadInsts.insert(UInst);
 733         }
 734         // See if we can figure out sext(i+constant) doesn't wrap, so we can
 735         // use a larger add.  This is common in subscripting.
 736         if (UInst && UInst->getOpcode()==Instruction::Add &&
 737             !UInst->use_empty() &&
 738             allUsesAreSameTyped(Instruction::SExt, UInst) &&
 739             isa<ConstantInt>(UInst->getOperand(1)) &&
 740             NoSignedWrap && LimitVal) {
 741           uint64_t oldBitSize = LimitVal->getValue().getBitWidth();
 742           uint64_t newBitSize = LargestType->getPrimitiveSizeInBits();
 743           ConstantInt* AddRHS = dyn_cast<ConstantInt>(UInst->getOperand(1));
 744           if (((APInt::getSignedMaxValue(oldBitSize) - IncrVal->getValue()) -
 745                 AddRHS->getValue()).sgt(LimitVal->getValue())) {
 746             // We've determined this is (i+constant) and it won't overflow.
 747             if (isa<SExtInst>(UInst->use_begin())) {
 748               SExtInst* oldSext = dyn_cast<SExtInst>(UInst->use_begin());
 749               uint64_t truncSize = oldSext->getType()->getPrimitiveSizeInBits();
 750               Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType,
 751                                                 L, oldSext->getType(), Rewriter);
 752               APInt APnewAddRHS = APInt(AddRHS->getValue()).sext(newBitSize);
 753               if (newBitSize > truncSize)
 754                 APnewAddRHS = APnewAddRHS.trunc(truncSize);
 755               ConstantInt* newAddRHS =ConstantInt::get(APnewAddRHS);
 756               Value *NewAdd =
 757                     BinaryOperator::CreateAdd(TruncIndVar, newAddRHS,
 758                                               UInst->getName()+".nosex", UInst);
 759               for (Value::use_iterator UI2 = UInst->use_begin(),
 760                     UE2 = UInst->use_end(); UI2 != UE2; ++UI2) {
 761                 Instruction *II = dyn_cast<Instruction>(UI2);
 762                 II->replaceAllUsesWith(NewAdd);
 763                 DeadInsts.insert(II);
 764               }
 765               DeadInsts.insert(UInst);
 766             }
 767           }
 768         }
 769         // Try for sext(i | constant).  This is safe as long as the
 770         // high bit of the constant is not set.
 771         if (UInst && UInst->getOpcode()==Instruction::Or &&
 772             !UInst->use_empty() &&
 773             allUsesAreSameTyped(Instruction::SExt, UInst) && NoSignedWrap &&
 774             isa<ConstantInt>(UInst->getOperand(1))) {
 775           ConstantInt* RHS = dyn_cast<ConstantInt>(UInst->getOperand(1));
 776           if (!RHS->getValue().isNegative()) {
 777             uint64_t newBitSize = LargestType->getPrimitiveSizeInBits();
 778             SExtInst* oldSext = dyn_cast<SExtInst>(UInst->use_begin());
 779             uint64_t truncSize = oldSext->getType()->getPrimitiveSizeInBits();
 780             Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType,
 781                                               L, oldSext->getType(), Rewriter);
 782             APInt APnewOrRHS = APInt(RHS->getValue()).sext(newBitSize);
 783             if (newBitSize > truncSize)
 784               APnewOrRHS = APnewOrRHS.trunc(truncSize);
 785             ConstantInt* newOrRHS =ConstantInt::get(APnewOrRHS);
 786             Value *NewOr =
 787                   BinaryOperator::CreateOr(TruncIndVar, newOrRHS,
 788                                             UInst->getName()+".nosex", UInst);
 789             for (Value::use_iterator UI2 = UInst->use_begin(),
 790                   UE2 = UInst->use_end(); UI2 != UE2; ++UI2) {
 791               Instruction *II = dyn_cast<Instruction>(UI2);
 792               II->replaceAllUsesWith(NewOr);
 793               DeadInsts.insert(II);
 794             }
 795             DeadInsts.insert(UInst);
 796           }
 797         }
 798         // A zext of a signed variable known not to overflow is still safe.
 799         if (UInst && isa<ZExtInst>(UInst) && (NoUnsignedWrap || NoSignedWrap)) {
 800           Value *TruncIndVar = getZeroExtendedTruncVar(AR, SE, LargestType, L,
 801                                          UInst->getType(), Rewriter);
 802           UInst->replaceAllUsesWith(TruncIndVar);
 803           DeadInsts.insert(UInst);
 804         }
 805         // If we have zext(i&constant), it's always safe to use the larger
 806         // variable.  This is not common but is a bottleneck in Openssl.
 807         // (RHS doesn't have to be constant.  There should be a better approach
 808         // than bottom-up pattern matching for this...)
 809         if (UInst && UInst->getOpcode()==Instruction::And &&
 810             !UInst->use_empty() &&
 811             allUsesAreSameTyped(Instruction::ZExt, UInst) &&
 812             isa<ConstantInt>(UInst->getOperand(1))) {
 813           uint64_t newBitSize = LargestType->getPrimitiveSizeInBits();
 814           ConstantInt* AndRHS = dyn_cast<ConstantInt>(UInst->getOperand(1));
 815           ZExtInst* oldZext = dyn_cast<ZExtInst>(UInst->use_begin());
 816           uint64_t truncSize = oldZext->getType()->getPrimitiveSizeInBits();
 817           Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType,
 818                                   L, oldZext->getType(), Rewriter);
 819           APInt APnewAndRHS = APInt(AndRHS->getValue()).zext(newBitSize);
 820           if (newBitSize > truncSize)
 821             APnewAndRHS = APnewAndRHS.trunc(truncSize);
 822           ConstantInt* newAndRHS = ConstantInt::get(APnewAndRHS);
 823           Value *NewAnd =
 824                 BinaryOperator::CreateAnd(TruncIndVar, newAndRHS,
 825                                           UInst->getName()+".nozex", UInst);
 826           for (Value::use_iterator UI2 = UInst->use_begin(),
 827                 UE2 = UInst->use_end(); UI2 != UE2; ++UI2) {
 828             Instruction *II = dyn_cast<Instruction>(UI2);
 829             II->replaceAllUsesWith(NewAnd);
 830             DeadInsts.insert(II);
 831           }
 832           DeadInsts.insert(UInst);
 833         }
 834         // If we have zext((i+constant)&constant), we can use the larger
 835         // variable even if the add does overflow.  This works whenever the
 836         // constant being ANDed is the same size as i, which it presumably is.
 837         // We don't need to restrict the expression being and'ed to i+const,
 838         // but we have to promote everything in it, so it's convenient.
 839         // zext((i | constant)&constant) is also valid and accepted here.
 840         if (UInst && (UInst->getOpcode()==Instruction::Add ||
 841                       UInst->getOpcode()==Instruction::Or) &&
 842             UInst->hasOneUse() &&
 843             isa<ConstantInt>(UInst->getOperand(1))) {
 844           uint64_t newBitSize = LargestType->getPrimitiveSizeInBits();
 845           ConstantInt* AddRHS = dyn_cast<ConstantInt>(UInst->getOperand(1));
 846           Instruction *UInst2 = dyn_cast<Instruction>(UInst->use_begin());
 847           if (UInst2 && UInst2->getOpcode() == Instruction::And &&
 848               !UInst2->use_empty() &&
 849               allUsesAreSameTyped(Instruction::ZExt, UInst2) &&
 850               isa<ConstantInt>(UInst2->getOperand(1))) {
 851             ZExtInst* oldZext = dyn_cast<ZExtInst>(UInst2->use_begin());
 852             uint64_t truncSize = oldZext->getType()->getPrimitiveSizeInBits();
 853             Value *TruncIndVar = getSignExtendedTruncVar(AR, SE, LargestType,
 854                                     L, oldZext->getType(), Rewriter);
 855             ConstantInt* AndRHS = dyn_cast<ConstantInt>(UInst2->getOperand(1));
 856             APInt APnewAddRHS = APInt(AddRHS->getValue()).zext(newBitSize);
 857             if (newBitSize > truncSize)
 858               APnewAddRHS = APnewAddRHS.trunc(truncSize);
 859             ConstantInt* newAddRHS = ConstantInt::get(APnewAddRHS);
 860             Value *NewAdd = ((UInst->getOpcode()==Instruction::Add) ?
 861                   BinaryOperator::CreateAdd(TruncIndVar, newAddRHS,
 862                                             UInst->getName()+".nozex", UInst2) :
 863                   BinaryOperator::CreateOr(TruncIndVar, newAddRHS,
 864                                             UInst->getName()+".nozex", UInst2));
 865             APInt APcopy2 = APInt(AndRHS->getValue());
 866             ConstantInt* newAndRHS = ConstantInt::get(APcopy2.zext(newBitSize));
 867             Value *NewAnd =
 868                   BinaryOperator::CreateAnd(NewAdd, newAndRHS,
 869                                             UInst->getName()+".nozex", UInst2);
 870             for (Value::use_iterator UI2 = UInst2->use_begin(),
 871                   UE2 = UInst2->use_end(); UI2 != UE2; ++UI2) {
 872               Instruction *II = dyn_cast<Instruction>(UI2);
 873               II->replaceAllUsesWith(NewAnd);
 874               DeadInsts.insert(II);
 875             }
 876             DeadInsts.insert(UInst);
 877             DeadInsts.insert(UInst2);
 878           }
 879         }
 880       }
 881
 882     // Replace the old PHI Node with the inserted computation.
 883     PN->replaceAllUsesWith(NewVal);
 884     DeadInsts.insert(PN);
 885     IndVars.pop_back();
 886     ++NumRemoved;
 887     Changed = true;
 888   }
 889
 890   DeleteTriviallyDeadInstructions(DeadInsts);
 891   assert(L->isLCSSAForm());
 892   return Changed;
 893 }
 894
 895 /// Return true if it is OK to use SIToFPInst for an inducation variable
 896 /// with given inital and exit values.
 897 static bool useSIToFPInst(ConstantFP &InitV, ConstantFP &ExitV,
 898                           uint64_t intIV, uint64_t intEV) {
 899
 900   if (InitV.getValueAPF().isNegative() || ExitV.getValueAPF().isNegative())
 901     return true;
 902
 903   // If the iteration range can be handled by SIToFPInst then use it.
 904   APInt Max = APInt::getSignedMaxValue(32);
 905   if (Max.getZExtValue() > static_cast<uint64_t>(abs(intEV - intIV)))
 906     return true;
 907
 908   return false;
 909 }
 910
 911 /// convertToInt - Convert APF to an integer, if possible.
 912 static bool convertToInt(const APFloat &APF, uint64_t *intVal) {
 913
 914   bool isExact = false;
 915   if (&APF.getSemantics() == &APFloat::PPCDoubleDouble)
 916     return false;
 917   if (APF.convertToInteger(intVal, 32, APF.isNegative(),
 918                            APFloat::rmTowardZero, &isExact)
 919       != APFloat::opOK)
 920     return false;
 921   if (!isExact)
 922     return false;
 923   return true;
 924
 925 }
 926
 927 /// HandleFloatingPointIV - If the loop has floating induction variable
 928 /// then insert corresponding integer induction variable if possible.
 929 /// For example,
 930 /// for(double i = 0; i < 10000; ++i)
 931 ///   bar(i)
 932 /// is converted into
 933 /// for(int i = 0; i < 10000; ++i)
 934 ///   bar((double)i);
 935 ///
 936 void IndVarSimplify::HandleFloatingPointIV(Loop *L, PHINode *PH,
 937                                    SmallPtrSet<Instruction*, 16> &DeadInsts) {
 938
 939   unsigned IncomingEdge = L->contains(PH->getIncomingBlock(0));
 940   unsigned BackEdge     = IncomingEdge^1;
 941
 942   // Check incoming value.
 943   ConstantFP *InitValue = dyn_cast<ConstantFP>(PH->getIncomingValue(IncomingEdge));
 944   if (!InitValue) return;
 945   uint64_t newInitValue = Type::Int32Ty->getPrimitiveSizeInBits();
 946   if (!convertToInt(InitValue->getValueAPF(), &newInitValue))
 947     return;
 948
 949   // Check IV increment. Reject this PH if increement operation is not
 950   // an add or increment value can not be represented by an integer.
 951   BinaryOperator *Incr =
 952     dyn_cast<BinaryOperator>(PH->getIncomingValue(BackEdge));
 953   if (!Incr) return;
 954   if (Incr->getOpcode() != Instruction::Add) return;
 955   ConstantFP *IncrValue = NULL;
 956   unsigned IncrVIndex = 1;
 957   if (Incr->getOperand(1) == PH)
 958     IncrVIndex = 0;
 959   IncrValue = dyn_cast<ConstantFP>(Incr->getOperand(IncrVIndex));
 960   if (!IncrValue) return;
 961   uint64_t newIncrValue = Type::Int32Ty->getPrimitiveSizeInBits();
 962   if (!convertToInt(IncrValue->getValueAPF(), &newIncrValue))
 963     return;
 964
 965   // Check Incr uses. One user is PH and the other users is exit condition used
 966   // by the conditional terminator.
 967   Value::use_iterator IncrUse = Incr->use_begin();
 968   Instruction *U1 = cast<Instruction>(IncrUse++);
 969   if (IncrUse == Incr->use_end()) return;
 970   Instruction *U2 = cast<Instruction>(IncrUse++);
 971   if (IncrUse != Incr->use_end()) return;
 972
 973   // Find exit condition.
 974   FCmpInst *EC = dyn_cast<FCmpInst>(U1);
 975   if (!EC)
 976     EC = dyn_cast<FCmpInst>(U2);
 977   if (!EC) return;
 978
 979   if (BranchInst *BI = dyn_cast<BranchInst>(EC->getParent()->getTerminator())) {
 980     if (!BI->isConditional()) return;
 981     if (BI->getCondition() != EC) return;
 982   }
 983
 984   // Find exit value. If exit value can not be represented as an interger then
 985   // do not handle this floating point PH.
 986   ConstantFP *EV = NULL;
 987   unsigned EVIndex = 1;
 988   if (EC->getOperand(1) == Incr)
 989     EVIndex = 0;
 990   EV = dyn_cast<ConstantFP>(EC->getOperand(EVIndex));
 991   if (!EV) return;
 992   uint64_t intEV = Type::Int32Ty->getPrimitiveSizeInBits();
 993   if (!convertToInt(EV->getValueAPF(), &intEV))
 994     return;
 995
 996   // Find new predicate for integer comparison.
 997   CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE;
 998   switch (EC->getPredicate()) {
 999   case CmpInst::FCMP_OEQ:
1000   case CmpInst::FCMP_UEQ:
1001     NewPred = CmpInst::ICMP_EQ;
1002     break;
1003   case CmpInst::FCMP_OGT:
1004   case CmpInst::FCMP_UGT:
1005     NewPred = CmpInst::ICMP_UGT;
1006     break;
1007   case CmpInst::FCMP_OGE:
1008   case CmpInst::FCMP_UGE:
1009     NewPred = CmpInst::ICMP_UGE;
1010     break;
1011   case CmpInst::FCMP_OLT:
1012   case CmpInst::FCMP_ULT:
1013     NewPred = CmpInst::ICMP_ULT;
1014     break;
1015   case CmpInst::FCMP_OLE:
1016   case CmpInst::FCMP_ULE:
1017     NewPred = CmpInst::ICMP_ULE;
1018     break;
1019   default:
1020     break;
1021   }
1022   if (NewPred == CmpInst::BAD_ICMP_PREDICATE) return;
1023
1024   // Insert new integer induction variable.
1025   PHINode *NewPHI = PHINode::Create(Type::Int32Ty,
1026                                     PH->getName()+".int", PH);
1027   NewPHI->addIncoming(ConstantInt::get(Type::Int32Ty, newInitValue),
1028                       PH->getIncomingBlock(IncomingEdge));
1029
1030   Value *NewAdd = BinaryOperator::CreateAdd(NewPHI,
1031                                             ConstantInt::get(Type::Int32Ty,
1032                                                              newIncrValue),
1033                                             Incr->getName()+".int", Incr);
1034   NewPHI->addIncoming(NewAdd, PH->getIncomingBlock(BackEdge));
1035
1036   // The back edge is edge 1 of newPHI, whatever it may have been in the
1037   // original PHI.
1038   ConstantInt *NewEV = ConstantInt::get(Type::Int32Ty, intEV);
1039   Value *LHS = (EVIndex == 1 ? NewPHI->getIncomingValue(1) : NewEV);
1040   Value *RHS = (EVIndex == 1 ? NewEV : NewPHI->getIncomingValue(1));
1041   ICmpInst *NewEC = new ICmpInst(NewPred, LHS, RHS, EC->getNameStart(),
1042                                  EC->getParent()->getTerminator());
1043
1044   // Delete old, floating point, exit comparision instruction.
1045   EC->replaceAllUsesWith(NewEC);
1046   DeadInsts.insert(EC);
1047
1048   // Delete old, floating point, increment instruction.
1049   Incr->replaceAllUsesWith(UndefValue::get(Incr->getType()));
1050   DeadInsts.insert(Incr);
1051
1052   // Replace floating induction variable. Give SIToFPInst preference over
1053   // UIToFPInst because it is faster on platforms that are widely used.
1054   if (useSIToFPInst(*InitValue, *EV, newInitValue, intEV)) {
1055     SIToFPInst *Conv = new SIToFPInst(NewPHI, PH->getType(), "indvar.conv",
1056                                       PH->getParent()->getFirstNonPHI());
1057     PH->replaceAllUsesWith(Conv);
1058   } else {
1059     UIToFPInst *Conv = new UIToFPInst(NewPHI, PH->getType(), "indvar.conv",
1060                                       PH->getParent()->getFirstNonPHI());
1061     PH->replaceAllUsesWith(Conv);
1062   }
1063   DeadInsts.insert(PH);
1064 }
1065