lib/Analysis/ScalarEvolutionExpander.cpp

   1 //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===//
   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 contains the implementation of the scalar evolution expander,
  11 // which is used to generate the code corresponding to a given scalar evolution
  12 // expression.
  13 //
  14 //===----------------------------------------------------------------------===//
  15
  16 #include "llvm/Analysis/ScalarEvolutionExpander.h"
  17 #include "llvm/Analysis/LoopInfo.h"
  18 #include "llvm/LLVMContext.h"
  19 #include "llvm/Target/TargetData.h"
  20 #include "llvm/ADT/STLExtras.h"
  21 using namespace llvm;
  22
  23 /// InsertNoopCastOfTo - Insert a cast of V to the specified type,
  24 /// which must be possible with a noop cast, doing what we can to share
  25 /// the casts.
  26 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) {
  27   Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false);
  28   assert((Op == Instruction::BitCast ||
  29           Op == Instruction::PtrToInt ||
  30           Op == Instruction::IntToPtr) &&
  31          "InsertNoopCastOfTo cannot perform non-noop casts!");
  32   assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) &&
  33          "InsertNoopCastOfTo cannot change sizes!");
  34
  35   // Short-circuit unnecessary bitcasts.
  36   if (Op == Instruction::BitCast && V->getType() == Ty)
  37     return V;
  38
  39   // Short-circuit unnecessary inttoptr<->ptrtoint casts.
  40   if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) &&
  41       SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) {
  42     if (CastInst *CI = dyn_cast<CastInst>(V))
  43       if ((CI->getOpcode() == Instruction::PtrToInt ||
  44            CI->getOpcode() == Instruction::IntToPtr) &&
  45           SE.getTypeSizeInBits(CI->getType()) ==
  46           SE.getTypeSizeInBits(CI->getOperand(0)->getType()))
  47         return CI->getOperand(0);
  48     if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
  49       if ((CE->getOpcode() == Instruction::PtrToInt ||
  50            CE->getOpcode() == Instruction::IntToPtr) &&
  51           SE.getTypeSizeInBits(CE->getType()) ==
  52           SE.getTypeSizeInBits(CE->getOperand(0)->getType()))
  53         return CE->getOperand(0);
  54   }
  55
  56   if (Constant *C = dyn_cast<Constant>(V))
  57     return ConstantExpr::getCast(Op, C, Ty);
  58
  59   if (Argument *A = dyn_cast<Argument>(V)) {
  60     // Check to see if there is already a cast!
  61     for (Value::use_iterator UI = A->use_begin(), E = A->use_end();
  62          UI != E; ++UI)
  63       if ((*UI)->getType() == Ty)
  64         if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
  65           if (CI->getOpcode() == Op) {
  66             // If the cast isn't the first instruction of the function, move it.
  67             if (BasicBlock::iterator(CI) !=
  68                 A->getParent()->getEntryBlock().begin()) {
  69               // Recreate the cast at the beginning of the entry block.
  70               // The old cast is left in place in case it is being used
  71               // as an insert point.
  72               Instruction *NewCI =
  73                 CastInst::Create(Op, V, Ty, "",
  74                                  A->getParent()->getEntryBlock().begin());
  75               NewCI->takeName(CI);
  76               CI->replaceAllUsesWith(NewCI);
  77               return NewCI;
  78             }
  79             return CI;
  80           }
  81
  82     Instruction *I = CastInst::Create(Op, V, Ty, V->getName(),
  83                                       A->getParent()->getEntryBlock().begin());
  84     InsertedValues.insert(I);
  85     return I;
  86   }
  87
  88   Instruction *I = cast<Instruction>(V);
  89
  90   // Check to see if there is already a cast.  If there is, use it.
  91   for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
  92        UI != E; ++UI) {
  93     if ((*UI)->getType() == Ty)
  94       if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI)))
  95         if (CI->getOpcode() == Op) {
  96           BasicBlock::iterator It = I; ++It;
  97           if (isa<InvokeInst>(I))
  98             It = cast<InvokeInst>(I)->getNormalDest()->begin();
  99           while (isa<PHINode>(It)) ++It;
 100           if (It != BasicBlock::iterator(CI)) {
 101             // Recreate the cast at the beginning of the entry block.
 102             // The old cast is left in place in case it is being used
 103             // as an insert point.
 104             Instruction *NewCI = CastInst::Create(Op, V, Ty, "", It);
 105             NewCI->takeName(CI);
 106             CI->replaceAllUsesWith(NewCI);
 107             return NewCI;
 108           }
 109           return CI;
 110         }
 111   }
 112   BasicBlock::iterator IP = I; ++IP;
 113   if (InvokeInst *II = dyn_cast<InvokeInst>(I))
 114     IP = II->getNormalDest()->begin();
 115   while (isa<PHINode>(IP)) ++IP;
 116   Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP);
 117   InsertedValues.insert(CI);
 118   return CI;
 119 }
 120
 121 /// InsertBinop - Insert the specified binary operator, doing a small amount
 122 /// of work to avoid inserting an obviously redundant operation.
 123 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode,
 124                                  Value *LHS, Value *RHS) {
 125   // Fold a binop with constant operands.
 126   if (Constant *CLHS = dyn_cast<Constant>(LHS))
 127     if (Constant *CRHS = dyn_cast<Constant>(RHS))
 128       return ConstantExpr::get(Opcode, CLHS, CRHS);
 129
 130   // Do a quick scan to see if we have this binop nearby.  If so, reuse it.
 131   unsigned ScanLimit = 6;
 132   BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
 133   // Scanning starts from the last instruction before the insertion point.
 134   BasicBlock::iterator IP = Builder.GetInsertPoint();
 135   if (IP != BlockBegin) {
 136     --IP;
 137     for (; ScanLimit; --IP, --ScanLimit) {
 138       if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS &&
 139           IP->getOperand(1) == RHS)
 140         return IP;
 141       if (IP == BlockBegin) break;
 142     }
 143   }
 144
 145   // If we haven't found this binop, insert it.
 146   Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp");
 147   InsertedValues.insert(BO);
 148   return BO;
 149 }
 150
 151 /// FactorOutConstant - Test if S is divisible by Factor, using signed
 152 /// division. If so, update S with Factor divided out and return true.
 153 /// S need not be evenly divisble if a reasonable remainder can be
 154 /// computed.
 155 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made
 156 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and
 157 /// check to see if the divide was folded.
 158 static bool FactorOutConstant(const SCEV *&S,
 159                               const SCEV *&Remainder,
 160                               const SCEV *Factor,
 161                               ScalarEvolution &SE,
 162                               const TargetData *TD) {
 163   // Everything is divisible by one.
 164   if (Factor->isOne())
 165     return true;
 166
 167   // x/x == 1.
 168   if (S == Factor) {
 169     S = SE.getIntegerSCEV(1, S->getType());
 170     return true;
 171   }
 172
 173   // For a Constant, check for a multiple of the given factor.
 174   if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) {
 175     // 0/x == 0.
 176     if (C->isZero())
 177       return true;
 178     // Check for divisibility.
 179     if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) {
 180       ConstantInt *CI =
 181         ConstantInt::get(SE.getContext(),
 182                          C->getValue()->getValue().sdiv(
 183                                                    FC->getValue()->getValue()));
 184       // If the quotient is zero and the remainder is non-zero, reject
 185       // the value at this scale. It will be considered for subsequent
 186       // smaller scales.
 187       if (!CI->isZero()) {
 188         const SCEV *Div = SE.getConstant(CI);
 189         S = Div;
 190         Remainder =
 191           SE.getAddExpr(Remainder,
 192                         SE.getConstant(C->getValue()->getValue().srem(
 193                                                   FC->getValue()->getValue())));
 194         return true;
 195       }
 196     }
 197   }
 198
 199   // In a Mul, check if there is a constant operand which is a multiple
 200   // of the given factor.
 201   if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) {
 202     if (TD) {
 203       // With TargetData, the size is known. Check if there is a constant
 204       // operand which is a multiple of the given factor. If so, we can
 205       // factor it.
 206       const SCEVConstant *FC = cast<SCEVConstant>(Factor);
 207       if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0)))
 208         if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) {
 209           const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
 210           SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
 211                                                  MOperands.end());
 212           NewMulOps[0] =
 213             SE.getConstant(C->getValue()->getValue().sdiv(
 214                                                    FC->getValue()->getValue()));
 215           S = SE.getMulExpr(NewMulOps);
 216           return true;
 217         }
 218     } else {
 219       // Without TargetData, check if Factor can be factored out of any of the
 220       // Mul's operands. If so, we can just remove it.
 221       for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) {
 222         const SCEV *SOp = M->getOperand(i);
 223         const SCEV *Remainder = SE.getIntegerSCEV(0, SOp->getType());
 224         if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) &&
 225             Remainder->isZero()) {
 226           const SmallVectorImpl<const SCEV *> &MOperands = M->getOperands();
 227           SmallVector<const SCEV *, 4> NewMulOps(MOperands.begin(),
 228                                                  MOperands.end());
 229           NewMulOps[i] = SOp;
 230           S = SE.getMulExpr(NewMulOps);
 231           return true;
 232         }
 233       }
 234     }
 235   }
 236
 237   // In an AddRec, check if both start and step are divisible.
 238   if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) {
 239     const SCEV *Step = A->getStepRecurrence(SE);
 240     const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType());
 241     if (!FactorOutConstant(Step, StepRem, Factor, SE, TD))
 242       return false;
 243     if (!StepRem->isZero())
 244       return false;
 245     const SCEV *Start = A->getStart();
 246     if (!FactorOutConstant(Start, Remainder, Factor, SE, TD))
 247       return false;
 248     S = SE.getAddRecExpr(Start, Step, A->getLoop());
 249     return true;
 250   }
 251
 252   return false;
 253 }
 254
 255 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs
 256 /// is the number of SCEVAddRecExprs present, which are kept at the end of
 257 /// the list.
 258 ///
 259 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops,
 260                                 const Type *Ty,
 261                                 ScalarEvolution &SE) {
 262   unsigned NumAddRecs = 0;
 263   for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i)
 264     ++NumAddRecs;
 265   // Group Ops into non-addrecs and addrecs.
 266   SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs);
 267   SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end());
 268   // Let ScalarEvolution sort and simplify the non-addrecs list.
 269   const SCEV *Sum = NoAddRecs.empty() ?
 270                     SE.getIntegerSCEV(0, Ty) :
 271                     SE.getAddExpr(NoAddRecs);
 272   // If it returned an add, use the operands. Otherwise it simplified
 273   // the sum into a single value, so just use that.
 274   if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum))
 275     Ops = Add->getOperands();
 276   else {
 277     Ops.clear();
 278     if (!Sum->isZero())
 279       Ops.push_back(Sum);
 280   }
 281   // Then append the addrecs.
 282   Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
 283 }
 284
 285 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values
 286 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}.
 287 /// This helps expose more opportunities for folding parts of the expressions
 288 /// into GEP indices.
 289 ///
 290 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops,
 291                          const Type *Ty,
 292                          ScalarEvolution &SE) {
 293   // Find the addrecs.
 294   SmallVector<const SCEV *, 8> AddRecs;
 295   for (unsigned i = 0, e = Ops.size(); i != e; ++i)
 296     while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) {
 297       const SCEV *Start = A->getStart();
 298       if (Start->isZero()) break;
 299       const SCEV *Zero = SE.getIntegerSCEV(0, Ty);
 300       AddRecs.push_back(SE.getAddRecExpr(Zero,
 301                                          A->getStepRecurrence(SE),
 302                                          A->getLoop()));
 303       if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) {
 304         Ops[i] = Zero;
 305         Ops.insert(Ops.end(), Add->op_begin(), Add->op_end());
 306         e += Add->getNumOperands();
 307       } else {
 308         Ops[i] = Start;
 309       }
 310     }
 311   if (!AddRecs.empty()) {
 312     // Add the addrecs onto the end of the list.
 313     Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end());
 314     // Resort the operand list, moving any constants to the front.
 315     SimplifyAddOperands(Ops, Ty, SE);
 316   }
 317 }
 318
 319 /// expandAddToGEP - Expand an addition expression with a pointer type into
 320 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps
 321 /// BasicAliasAnalysis and other passes analyze the result. See the rules
 322 /// for getelementptr vs. inttoptr in
 323 /// http://llvm.org/docs/LangRef.html#pointeraliasing
 324 /// for details.
 325 ///
 326 /// Design note: The correctness of using getelmeentptr here depends on
 327 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as
 328 /// they may introduce pointer arithmetic which may not be safely converted
 329 /// into getelementptr.
 330 ///
 331 /// Design note: It might seem desirable for this function to be more
 332 /// loop-aware. If some of the indices are loop-invariant while others
 333 /// aren't, it might seem desirable to emit multiple GEPs, keeping the
 334 /// loop-invariant portions of the overall computation outside the loop.
 335 /// However, there are a few reasons this is not done here. Hoisting simple
 336 /// arithmetic is a low-level optimization that often isn't very
 337 /// important until late in the optimization process. In fact, passes
 338 /// like InstructionCombining will combine GEPs, even if it means
 339 /// pushing loop-invariant computation down into loops, so even if the
 340 /// GEPs were split here, the work would quickly be undone. The
 341 /// LoopStrengthReduction pass, which is usually run quite late (and
 342 /// after the last InstructionCombining pass), takes care of hoisting
 343 /// loop-invariant portions of expressions, after considering what
 344 /// can be folded using target addressing modes.
 345 ///
 346 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin,
 347                                     const SCEV *const *op_end,
 348                                     const PointerType *PTy,
 349                                     const Type *Ty,
 350                                     Value *V) {
 351   const Type *ElTy = PTy->getElementType();
 352   SmallVector<Value *, 4> GepIndices;
 353   SmallVector<const SCEV *, 8> Ops(op_begin, op_end);
 354   bool AnyNonZeroIndices = false;
 355
 356   // Split AddRecs up into parts as either of the parts may be usable
 357   // without the other.
 358   SplitAddRecs(Ops, Ty, SE);
 359
 360   // Decend down the pointer's type and attempt to convert the other
 361   // operands into GEP indices, at each level. The first index in a GEP
 362   // indexes into the array implied by the pointer operand; the rest of
 363   // the indices index into the element or field type selected by the
 364   // preceding index.
 365   for (;;) {
 366     const SCEV *ElSize = SE.getAllocSizeExpr(ElTy);
 367     // If the scale size is not 0, attempt to factor out a scale for
 368     // array indexing.
 369     SmallVector<const SCEV *, 8> ScaledOps;
 370     if (ElTy->isSized() && !ElSize->isZero()) {
 371       SmallVector<const SCEV *, 8> NewOps;
 372       for (unsigned i = 0, e = Ops.size(); i != e; ++i) {
 373         const SCEV *Op = Ops[i];
 374         const SCEV *Remainder = SE.getIntegerSCEV(0, Ty);
 375         if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) {
 376           // Op now has ElSize factored out.
 377           ScaledOps.push_back(Op);
 378           if (!Remainder->isZero())
 379             NewOps.push_back(Remainder);
 380           AnyNonZeroIndices = true;
 381         } else {
 382           // The operand was not divisible, so add it to the list of operands
 383           // we'll scan next iteration.
 384           NewOps.push_back(Ops[i]);
 385         }
 386       }
 387       // If we made any changes, update Ops.
 388       if (!ScaledOps.empty()) {
 389         Ops = NewOps;
 390         SimplifyAddOperands(Ops, Ty, SE);
 391       }
 392     }
 393
 394     // Record the scaled array index for this level of the type. If
 395     // we didn't find any operands that could be factored, tentatively
 396     // assume that element zero was selected (since the zero offset
 397     // would obviously be folded away).
 398     Value *Scaled = ScaledOps.empty() ?
 399                     Constant::getNullValue(Ty) :
 400                     expandCodeFor(SE.getAddExpr(ScaledOps), Ty);
 401     GepIndices.push_back(Scaled);
 402
 403     // Collect struct field index operands.
 404     while (const StructType *STy = dyn_cast<StructType>(ElTy)) {
 405       bool FoundFieldNo = false;
 406       // An empty struct has no fields.
 407       if (STy->getNumElements() == 0) break;
 408       if (SE.TD) {
 409         // With TargetData, field offsets are known. See if a constant offset
 410         // falls within any of the struct fields.
 411         if (Ops.empty()) break;
 412         if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0]))
 413           if (SE.getTypeSizeInBits(C->getType()) <= 64) {
 414             const StructLayout &SL = *SE.TD->getStructLayout(STy);
 415             uint64_t FullOffset = C->getValue()->getZExtValue();
 416             if (FullOffset < SL.getSizeInBytes()) {
 417               unsigned ElIdx = SL.getElementContainingOffset(FullOffset);
 418               GepIndices.push_back(
 419                   ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx));
 420               ElTy = STy->getTypeAtIndex(ElIdx);
 421               Ops[0] =
 422                 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx));
 423               AnyNonZeroIndices = true;
 424               FoundFieldNo = true;
 425             }
 426           }
 427       } else {
 428         // Without TargetData, just check for a SCEVFieldOffsetExpr of the
 429         // appropriate struct type.
 430         for (unsigned i = 0, e = Ops.size(); i != e; ++i)
 431           if (const SCEVFieldOffsetExpr *FO =
 432                 dyn_cast<SCEVFieldOffsetExpr>(Ops[i]))
 433             if (FO->getStructType() == STy) {
 434               unsigned FieldNo = FO->getFieldNo();
 435               GepIndices.push_back(
 436                   ConstantInt::get(Type::getInt32Ty(Ty->getContext()),
 437                                    FieldNo));
 438               ElTy = STy->getTypeAtIndex(FieldNo);
 439               Ops[i] = SE.getConstant(Ty, 0);
 440               AnyNonZeroIndices = true;
 441               FoundFieldNo = true;
 442               break;
 443             }
 444       }
 445       // If no struct field offsets were found, tentatively assume that
 446       // field zero was selected (since the zero offset would obviously
 447       // be folded away).
 448       if (!FoundFieldNo) {
 449         ElTy = STy->getTypeAtIndex(0u);
 450         GepIndices.push_back(
 451           Constant::getNullValue(Type::getInt32Ty(Ty->getContext())));
 452       }
 453     }
 454
 455     if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy))
 456       ElTy = ATy->getElementType();
 457     else
 458       break;
 459   }
 460
 461   // If none of the operands were convertable to proper GEP indices, cast
 462   // the base to i8* and do an ugly getelementptr with that. It's still
 463   // better than ptrtoint+arithmetic+inttoptr at least.
 464   if (!AnyNonZeroIndices) {
 465     // Cast the base to i8*.
 466     V = InsertNoopCastOfTo(V,
 467        Type::getInt8Ty(Ty->getContext())->getPointerTo(PTy->getAddressSpace()));
 468
 469     // Expand the operands for a plain byte offset.
 470     Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty);
 471
 472     // Fold a GEP with constant operands.
 473     if (Constant *CLHS = dyn_cast<Constant>(V))
 474       if (Constant *CRHS = dyn_cast<Constant>(Idx))
 475         return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1);
 476
 477     // Do a quick scan to see if we have this GEP nearby.  If so, reuse it.
 478     unsigned ScanLimit = 6;
 479     BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin();
 480     // Scanning starts from the last instruction before the insertion point.
 481     BasicBlock::iterator IP = Builder.GetInsertPoint();
 482     if (IP != BlockBegin) {
 483       --IP;
 484       for (; ScanLimit; --IP, --ScanLimit) {
 485         if (IP->getOpcode() == Instruction::GetElementPtr &&
 486             IP->getOperand(0) == V && IP->getOperand(1) == Idx)
 487           return IP;
 488         if (IP == BlockBegin) break;
 489       }
 490     }
 491
 492     // Emit a GEP.
 493     Value *GEP = Builder.CreateGEP(V, Idx, "uglygep");
 494     InsertedValues.insert(GEP);
 495     return GEP;
 496   }
 497
 498   // Insert a pretty getelementptr. Note that this GEP is not marked inbounds,
 499   // because ScalarEvolution may have changed the address arithmetic to
 500   // compute a value which is beyond the end of the allocated object.
 501   Value *GEP = Builder.CreateGEP(V,
 502                                  GepIndices.begin(),
 503                                  GepIndices.end(),
 504                                  "scevgep");
 505   Ops.push_back(SE.getUnknown(GEP));
 506   InsertedValues.insert(GEP);
 507   return expand(SE.getAddExpr(Ops));
 508 }
 509
 510 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) {
 511   const Type *Ty = SE.getEffectiveSCEVType(S->getType());
 512   Value *V = expand(S->getOperand(S->getNumOperands()-1));
 513
 514   // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
 515   // comments on expandAddToGEP for details.
 516   if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) {
 517     const SmallVectorImpl<const SCEV *> &Ops = S->getOperands();
 518     return expandAddToGEP(&Ops[0], &Ops[Ops.size() - 1], PTy, Ty, V);
 519   }
 520
 521   V = InsertNoopCastOfTo(V, Ty);
 522
 523   // Emit a bunch of add instructions
 524   for (int i = S->getNumOperands()-2; i >= 0; --i) {
 525     Value *W = expandCodeFor(S->getOperand(i), Ty);
 526     V = InsertBinop(Instruction::Add, V, W);
 527   }
 528   return V;
 529 }
 530
 531 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) {
 532   const Type *Ty = SE.getEffectiveSCEVType(S->getType());
 533   int FirstOp = 0;  // Set if we should emit a subtract.
 534   if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0)))
 535     if (SC->getValue()->isAllOnesValue())
 536       FirstOp = 1;
 537
 538   int i = S->getNumOperands()-2;
 539   Value *V = expandCodeFor(S->getOperand(i+1), Ty);
 540
 541   // Emit a bunch of multiply instructions
 542   for (; i >= FirstOp; --i) {
 543     Value *W = expandCodeFor(S->getOperand(i), Ty);
 544     V = InsertBinop(Instruction::Mul, V, W);
 545   }
 546
 547   // -1 * ...  --->  0 - ...
 548   if (FirstOp == 1)
 549     V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V);
 550   return V;
 551 }
 552
 553 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) {
 554   const Type *Ty = SE.getEffectiveSCEVType(S->getType());
 555
 556   Value *LHS = expandCodeFor(S->getLHS(), Ty);
 557   if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) {
 558     const APInt &RHS = SC->getValue()->getValue();
 559     if (RHS.isPowerOf2())
 560       return InsertBinop(Instruction::LShr, LHS,
 561                          ConstantInt::get(Ty, RHS.logBase2()));
 562   }
 563
 564   Value *RHS = expandCodeFor(S->getRHS(), Ty);
 565   return InsertBinop(Instruction::UDiv, LHS, RHS);
 566 }
 567
 568 /// Move parts of Base into Rest to leave Base with the minimal
 569 /// expression that provides a pointer operand suitable for a
 570 /// GEP expansion.
 571 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest,
 572                               ScalarEvolution &SE) {
 573   while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) {
 574     Base = A->getStart();
 575     Rest = SE.getAddExpr(Rest,
 576                          SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()),
 577                                           A->getStepRecurrence(SE),
 578                                           A->getLoop()));
 579   }
 580   if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) {
 581     Base = A->getOperand(A->getNumOperands()-1);
 582     SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end());
 583     NewAddOps.back() = Rest;
 584     Rest = SE.getAddExpr(NewAddOps);
 585     ExposePointerBase(Base, Rest, SE);
 586   }
 587 }
 588
 589 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) {
 590   const Type *Ty = SE.getEffectiveSCEVType(S->getType());
 591   const Loop *L = S->getLoop();
 592
 593   // First check for an existing canonical IV in a suitable type.
 594   PHINode *CanonicalIV = 0;
 595   if (PHINode *PN = L->getCanonicalInductionVariable())
 596     if (SE.isSCEVable(PN->getType()) &&
 597         isa<IntegerType>(SE.getEffectiveSCEVType(PN->getType())) &&
 598         SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty))
 599       CanonicalIV = PN;
 600
 601   // Rewrite an AddRec in terms of the canonical induction variable, if
 602   // its type is more narrow.
 603   if (CanonicalIV &&
 604       SE.getTypeSizeInBits(CanonicalIV->getType()) >
 605       SE.getTypeSizeInBits(Ty)) {
 606     const SCEV *Start = SE.getAnyExtendExpr(S->getStart(),
 607                                             CanonicalIV->getType());
 608     const SCEV *Step = SE.getAnyExtendExpr(S->getStepRecurrence(SE),
 609                                            CanonicalIV->getType());
 610     Value *V = expand(SE.getAddRecExpr(Start, Step, S->getLoop()));
 611     BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
 612     BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
 613     BasicBlock::iterator NewInsertPt =
 614       next(BasicBlock::iterator(cast<Instruction>(V)));
 615     while (isa<PHINode>(NewInsertPt)) ++NewInsertPt;
 616     V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0,
 617                       NewInsertPt);
 618     Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
 619     return V;
 620   }
 621
 622   // {X,+,F} --> X + {0,+,F}
 623   if (!S->getStart()->isZero()) {
 624     const SmallVectorImpl<const SCEV *> &SOperands = S->getOperands();
 625     SmallVector<const SCEV *, 4> NewOps(SOperands.begin(), SOperands.end());
 626     NewOps[0] = SE.getIntegerSCEV(0, Ty);
 627     const SCEV *Rest = SE.getAddRecExpr(NewOps, L);
 628
 629     // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the
 630     // comments on expandAddToGEP for details.
 631     const SCEV *Base = S->getStart();
 632     const SCEV *RestArray[1] = { Rest };
 633     // Dig into the expression to find the pointer base for a GEP.
 634     ExposePointerBase(Base, RestArray[0], SE);
 635     // If we found a pointer, expand the AddRec with a GEP.
 636     if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) {
 637       // Make sure the Base isn't something exotic, such as a multiplied
 638       // or divided pointer value. In those cases, the result type isn't
 639       // actually a pointer type.
 640       if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) {
 641         Value *StartV = expand(Base);
 642         assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!");
 643         return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV);
 644       }
 645     }
 646
 647     // Just do a normal add. Pre-expand the operands to suppress folding.
 648     return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())),
 649                                 SE.getUnknown(expand(Rest))));
 650   }
 651
 652   // {0,+,1} --> Insert a canonical induction variable into the loop!
 653   if (S->isAffine() &&
 654       S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) {
 655     // If there's a canonical IV, just use it.
 656     if (CanonicalIV) {
 657       assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) &&
 658              "IVs with types different from the canonical IV should "
 659              "already have been handled!");
 660       return CanonicalIV;
 661     }
 662
 663     // Create and insert the PHI node for the induction variable in the
 664     // specified loop.
 665     BasicBlock *Header = L->getHeader();
 666     BasicBlock *Preheader = L->getLoopPreheader();
 667     PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin());
 668     InsertedValues.insert(PN);
 669     PN->addIncoming(Constant::getNullValue(Ty), Preheader);
 670
 671     pred_iterator HPI = pred_begin(Header);
 672     assert(HPI != pred_end(Header) && "Loop with zero preds???");
 673     if (!L->contains(*HPI)) ++HPI;
 674     assert(HPI != pred_end(Header) && L->contains(*HPI) &&
 675            "No backedge in loop?");
 676
 677     // Insert a unit add instruction right before the terminator corresponding
 678     // to the back-edge.
 679     Constant *One = ConstantInt::get(Ty, 1);
 680     Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next",
 681                                                  (*HPI)->getTerminator());
 682     InsertedValues.insert(Add);
 683
 684     pred_iterator PI = pred_begin(Header);
 685     if (*PI == Preheader)
 686       ++PI;
 687     PN->addIncoming(Add, *PI);
 688     return PN;
 689   }
 690
 691   // {0,+,F} --> {0,+,1} * F
 692   // Get the canonical induction variable I for this loop.
 693   Value *I = CanonicalIV ?
 694              CanonicalIV :
 695              getOrInsertCanonicalInductionVariable(L, Ty);
 696
 697   // If this is a simple linear addrec, emit it now as a special case.
 698   if (S->isAffine())    // {0,+,F} --> i*F
 699     return
 700       expand(SE.getTruncateOrNoop(
 701         SE.getMulExpr(SE.getUnknown(I),
 702                       SE.getNoopOrAnyExtend(S->getOperand(1),
 703                                             I->getType())),
 704         Ty));
 705
 706   // If this is a chain of recurrences, turn it into a closed form, using the
 707   // folders, then expandCodeFor the closed form.  This allows the folders to
 708   // simplify the expression without having to build a bunch of special code
 709   // into this folder.
 710   const SCEV *IH = SE.getUnknown(I);   // Get I as a "symbolic" SCEV.
 711
 712   // Promote S up to the canonical IV type, if the cast is foldable.
 713   const SCEV *NewS = S;
 714   const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType());
 715   if (isa<SCEVAddRecExpr>(Ext))
 716     NewS = Ext;
 717
 718   const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE);
 719   //cerr << "Evaluated: " << *this << "\n     to: " << *V << "\n";
 720
 721   // Truncate the result down to the original type, if needed.
 722   const SCEV *T = SE.getTruncateOrNoop(V, Ty);
 723   return expand(T);
 724 }
 725
 726 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) {
 727   const Type *Ty = SE.getEffectiveSCEVType(S->getType());
 728   Value *V = expandCodeFor(S->getOperand(),
 729                            SE.getEffectiveSCEVType(S->getOperand()->getType()));
 730   Value *I = Builder.CreateTrunc(V, Ty, "tmp");
 731   InsertedValues.insert(I);
 732   return I;
 733 }
 734
 735 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) {
 736   const Type *Ty = SE.getEffectiveSCEVType(S->getType());
 737   Value *V = expandCodeFor(S->getOperand(),
 738                            SE.getEffectiveSCEVType(S->getOperand()->getType()));
 739   Value *I = Builder.CreateZExt(V, Ty, "tmp");
 740   InsertedValues.insert(I);
 741   return I;
 742 }
 743
 744 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) {
 745   const Type *Ty = SE.getEffectiveSCEVType(S->getType());
 746   Value *V = expandCodeFor(S->getOperand(),
 747                            SE.getEffectiveSCEVType(S->getOperand()->getType()));
 748   Value *I = Builder.CreateSExt(V, Ty, "tmp");
 749   InsertedValues.insert(I);
 750   return I;
 751 }
 752
 753 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) {
 754   Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
 755   const Type *Ty = LHS->getType();
 756   for (int i = S->getNumOperands()-2; i >= 0; --i) {
 757     // In the case of mixed integer and pointer types, do the
 758     // rest of the comparisons as integer.
 759     if (S->getOperand(i)->getType() != Ty) {
 760       Ty = SE.getEffectiveSCEVType(Ty);
 761       LHS = InsertNoopCastOfTo(LHS, Ty);
 762     }
 763     Value *RHS = expandCodeFor(S->getOperand(i), Ty);
 764     Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp");
 765     InsertedValues.insert(ICmp);
 766     Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax");
 767     InsertedValues.insert(Sel);
 768     LHS = Sel;
 769   }
 770   // In the case of mixed integer and pointer types, cast the
 771   // final result back to the pointer type.
 772   if (LHS->getType() != S->getType())
 773     LHS = InsertNoopCastOfTo(LHS, S->getType());
 774   return LHS;
 775 }
 776
 777 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) {
 778   Value *LHS = expand(S->getOperand(S->getNumOperands()-1));
 779   const Type *Ty = LHS->getType();
 780   for (int i = S->getNumOperands()-2; i >= 0; --i) {
 781     // In the case of mixed integer and pointer types, do the
 782     // rest of the comparisons as integer.
 783     if (S->getOperand(i)->getType() != Ty) {
 784       Ty = SE.getEffectiveSCEVType(Ty);
 785       LHS = InsertNoopCastOfTo(LHS, Ty);
 786     }
 787     Value *RHS = expandCodeFor(S->getOperand(i), Ty);
 788     Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp");
 789     InsertedValues.insert(ICmp);
 790     Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax");
 791     InsertedValues.insert(Sel);
 792     LHS = Sel;
 793   }
 794   // In the case of mixed integer and pointer types, cast the
 795   // final result back to the pointer type.
 796   if (LHS->getType() != S->getType())
 797     LHS = InsertNoopCastOfTo(LHS, S->getType());
 798   return LHS;
 799 }
 800
 801 Value *SCEVExpander::visitFieldOffsetExpr(const SCEVFieldOffsetExpr *S) {
 802   return ConstantExpr::getOffsetOf(S->getStructType(), S->getFieldNo());
 803 }
 804
 805 Value *SCEVExpander::visitAllocSizeExpr(const SCEVAllocSizeExpr *S) {
 806   return ConstantExpr::getSizeOf(S->getAllocType());
 807 }
 808
 809 Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) {
 810   // Expand the code for this SCEV.
 811   Value *V = expand(SH);
 812   if (Ty) {
 813     assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) &&
 814            "non-trivial casts should be done with the SCEVs directly!");
 815     V = InsertNoopCastOfTo(V, Ty);
 816   }
 817   return V;
 818 }
 819
 820 Value *SCEVExpander::expand(const SCEV *S) {
 821   // Compute an insertion point for this SCEV object. Hoist the instructions
 822   // as far out in the loop nest as possible.
 823   Instruction *InsertPt = Builder.GetInsertPoint();
 824   for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ;
 825        L = L->getParentLoop())
 826     if (S->isLoopInvariant(L)) {
 827       if (!L) break;
 828       if (BasicBlock *Preheader = L->getLoopPreheader())
 829         InsertPt = Preheader->getTerminator();
 830     } else {
 831       // If the SCEV is computable at this level, insert it into the header
 832       // after the PHIs (and after any other instructions that we've inserted
 833       // there) so that it is guaranteed to dominate any user inside the loop.
 834       if (L && S->hasComputableLoopEvolution(L))
 835         InsertPt = L->getHeader()->getFirstNonPHI();
 836       while (isInsertedInstruction(InsertPt))
 837         InsertPt = next(BasicBlock::iterator(InsertPt));
 838       break;
 839     }
 840
 841   // Check to see if we already expanded this here.
 842   std::map<std::pair<const SCEV *, Instruction *>,
 843            AssertingVH<Value> >::iterator I =
 844     InsertedExpressions.find(std::make_pair(S, InsertPt));
 845   if (I != InsertedExpressions.end())
 846     return I->second;
 847
 848   BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
 849   BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
 850   Builder.SetInsertPoint(InsertPt->getParent(), InsertPt);
 851
 852   // Expand the expression into instructions.
 853   Value *V = visit(S);
 854
 855   // Remember the expanded value for this SCEV at this location.
 856   InsertedExpressions[std::make_pair(S, InsertPt)] = V;
 857
 858   Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
 859   return V;
 860 }
 861
 862 /// getOrInsertCanonicalInductionVariable - This method returns the
 863 /// canonical induction variable of the specified type for the specified
 864 /// loop (inserting one if there is none).  A canonical induction variable
 865 /// starts at zero and steps by one on each iteration.
 866 Value *
 867 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L,
 868                                                     const Type *Ty) {
 869   assert(Ty->isInteger() && "Can only insert integer induction variables!");
 870   const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty),
 871                                    SE.getIntegerSCEV(1, Ty), L);
 872   BasicBlock *SaveInsertBB = Builder.GetInsertBlock();
 873   BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint();
 874   Value *V = expandCodeFor(H, 0, L->getHeader()->begin());
 875   if (SaveInsertBB)
 876     Builder.SetInsertPoint(SaveInsertBB, SaveInsertPt);
 877   return V;
 878 }