lib/Transforms/InstCombine/InstCombinePHI.cpp

   1 //===- InstCombinePHI.cpp -------------------------------------------------===//
   2 //
   3 //                     The LLVM Compiler Infrastructure
   4 //
   5 // This file is distributed under the University of Illinois Open Source
   6 // License. See LICENSE.TXT for details.
   7 //
   8 //===----------------------------------------------------------------------===//
   9 //
  10 // This file implements the visitPHINode function.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "InstCombine.h"
  15 #include "llvm/Analysis/InstructionSimplify.h"
  16 #include "llvm/Target/TargetData.h"
  17 #include "llvm/ADT/SmallPtrSet.h"
  18 #include "llvm/ADT/STLExtras.h"
  19 using namespace llvm;
  20
  21 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(a,c)]
  22 /// and if a/b/c and the add's all have a single use, turn this into a phi
  23 /// and a single binop.
  24 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
  25   Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
  26   assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
  27   unsigned Opc = FirstInst->getOpcode();
  28   Value *LHSVal = FirstInst->getOperand(0);
  29   Value *RHSVal = FirstInst->getOperand(1);
  30
  31   const Type *LHSType = LHSVal->getType();
  32   const Type *RHSType = RHSVal->getType();
  33
  34   bool isNUW = false, isNSW = false, isExact = false;
  35   if (OverflowingBinaryOperator *BO =
  36         dyn_cast<OverflowingBinaryOperator>(FirstInst)) {
  37     isNUW = BO->hasNoUnsignedWrap();
  38     isNSW = BO->hasNoSignedWrap();
  39   } else if (PossiblyExactOperator *PEO =
  40                dyn_cast<PossiblyExactOperator>(FirstInst))
  41     isExact = PEO->isExact();
  42
  43   // Scan to see if all operands are the same opcode, and all have one use.
  44   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
  45     Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
  46     if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
  47         // Verify type of the LHS matches so we don't fold cmp's of different
  48         // types.
  49         I->getOperand(0)->getType() != LHSType ||
  50         I->getOperand(1)->getType() != RHSType)
  51       return 0;
  52
  53     // If they are CmpInst instructions, check their predicates
  54     if (CmpInst *CI = dyn_cast<CmpInst>(I))
  55       if (CI->getPredicate() != cast<CmpInst>(FirstInst)->getPredicate())
  56         return 0;
  57
  58     if (isNUW)
  59       isNUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
  60     if (isNSW)
  61       isNSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
  62     if (isExact)
  63       isExact = cast<PossiblyExactOperator>(I)->isExact();
  64
  65     // Keep track of which operand needs a phi node.
  66     if (I->getOperand(0) != LHSVal) LHSVal = 0;
  67     if (I->getOperand(1) != RHSVal) RHSVal = 0;
  68   }
  69
  70   // If both LHS and RHS would need a PHI, don't do this transformation,
  71   // because it would increase the number of PHIs entering the block,
  72   // which leads to higher register pressure. This is especially
  73   // bad when the PHIs are in the header of a loop.
  74   if (!LHSVal && !RHSVal)
  75     return 0;
  76
  77   // Otherwise, this is safe to transform!
  78
  79   Value *InLHS = FirstInst->getOperand(0);
  80   Value *InRHS = FirstInst->getOperand(1);
  81   PHINode *NewLHS = 0, *NewRHS = 0;
  82   if (LHSVal == 0) {
  83     NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(),
  84                              FirstInst->getOperand(0)->getName() + ".pn");
  85     NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
  86     InsertNewInstBefore(NewLHS, PN);
  87     LHSVal = NewLHS;
  88   }
  89
  90   if (RHSVal == 0) {
  91     NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(),
  92                              FirstInst->getOperand(1)->getName() + ".pn");
  93     NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
  94     InsertNewInstBefore(NewRHS, PN);
  95     RHSVal = NewRHS;
  96   }
  97
  98   // Add all operands to the new PHIs.
  99   if (NewLHS || NewRHS) {
 100     for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
 101       Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
 102       if (NewLHS) {
 103         Value *NewInLHS = InInst->getOperand(0);
 104         NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
 105       }
 106       if (NewRHS) {
 107         Value *NewInRHS = InInst->getOperand(1);
 108         NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
 109       }
 110     }
 111   }
 112
 113   if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) {
 114     CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
 115                                      LHSVal, RHSVal);
 116     NewCI->setDebugLoc(FirstInst->getDebugLoc());
 117     return NewCI;
 118   }
 119
 120   BinaryOperator *BinOp = cast<BinaryOperator>(FirstInst);
 121   BinaryOperator *NewBinOp =
 122     BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
 123   if (isNUW) NewBinOp->setHasNoUnsignedWrap();
 124   if (isNSW) NewBinOp->setHasNoSignedWrap();
 125   if (isExact) NewBinOp->setIsExact();
 126   NewBinOp->setDebugLoc(FirstInst->getDebugLoc());
 127   return NewBinOp;
 128 }
 129
 130 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
 131   GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
 132
 133   SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
 134                                         FirstInst->op_end());
 135   // This is true if all GEP bases are allocas and if all indices into them are
 136   // constants.
 137   bool AllBasePointersAreAllocas = true;
 138
 139   // We don't want to replace this phi if the replacement would require
 140   // more than one phi, which leads to higher register pressure. This is
 141   // especially bad when the PHIs are in the header of a loop.
 142   bool NeededPhi = false;
 143
 144   bool AllInBounds = true;
 145
 146   // Scan to see if all operands are the same opcode, and all have one use.
 147   for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
 148     GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
 149     if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
 150       GEP->getNumOperands() != FirstInst->getNumOperands())
 151       return 0;
 152
 153     AllInBounds &= GEP->isInBounds();
 154
 155     // Keep track of whether or not all GEPs are of alloca pointers.
 156     if (AllBasePointersAreAllocas &&
 157         (!isa<AllocaInst>(GEP->getOperand(0)) ||
 158          !GEP->hasAllConstantIndices()))
 159       AllBasePointersAreAllocas = false;
 160
 161     // Compare the operand lists.
 162     for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
 163       if (FirstInst->getOperand(op) == GEP->getOperand(op))
 164         continue;
 165
 166       // Don't merge two GEPs when two operands differ (introducing phi nodes)
 167       // if one of the PHIs has a constant for the index.  The index may be
 168       // substantially cheaper to compute for the constants, so making it a
 169       // variable index could pessimize the path.  This also handles the case
 170       // for struct indices, which must always be constant.
 171       if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
 172           isa<ConstantInt>(GEP->getOperand(op)))
 173         return 0;
 174
 175       if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
 176         return 0;
 177
 178       // If we already needed a PHI for an earlier operand, and another operand
 179       // also requires a PHI, we'd be introducing more PHIs than we're
 180       // eliminating, which increases register pressure on entry to the PHI's
 181       // block.
 182       if (NeededPhi)
 183         return 0;
 184
 185       FixedOperands[op] = 0;  // Needs a PHI.
 186       NeededPhi = true;
 187     }
 188   }
 189
 190   // If all of the base pointers of the PHI'd GEPs are from allocas, don't
 191   // bother doing this transformation.  At best, this will just save a bit of
 192   // offset calculation, but all the predecessors will have to materialize the
 193   // stack address into a register anyway.  We'd actually rather *clone* the
 194   // load up into the predecessors so that we have a load of a gep of an alloca,
 195   // which can usually all be folded into the load.
 196   if (AllBasePointersAreAllocas)
 197     return 0;
 198
 199   // Otherwise, this is safe to transform.  Insert PHI nodes for each operand
 200   // that is variable.
 201   SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
 202
 203   bool HasAnyPHIs = false;
 204   for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
 205     if (FixedOperands[i]) continue;  // operand doesn't need a phi.
 206     Value *FirstOp = FirstInst->getOperand(i);
 207     PHINode *NewPN = PHINode::Create(FirstOp->getType(), e,
 208                                      FirstOp->getName()+".pn");
 209     InsertNewInstBefore(NewPN, PN);
 210
 211     NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
 212     OperandPhis[i] = NewPN;
 213     FixedOperands[i] = NewPN;
 214     HasAnyPHIs = true;
 215   }
 216
 217
 218   // Add all operands to the new PHIs.
 219   if (HasAnyPHIs) {
 220     for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
 221       GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
 222       BasicBlock *InBB = PN.getIncomingBlock(i);
 223
 224       for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
 225         if (PHINode *OpPhi = OperandPhis[op])
 226           OpPhi->addIncoming(InGEP->getOperand(op), InBB);
 227     }
 228   }
 229
 230   Value *Base = FixedOperands[0];
 231   GetElementPtrInst *NewGEP =
 232     GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
 233                               FixedOperands.end());
 234   if (AllInBounds) NewGEP->setIsInBounds();
 235   NewGEP->setDebugLoc(FirstInst->getDebugLoc());
 236   return NewGEP;
 237 }
 238
 239
 240 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
 241 /// sink the load out of the block that defines it.  This means that it must be
 242 /// obvious the value of the load is not changed from the point of the load to
 243 /// the end of the block it is in.
 244 ///
 245 /// Finally, it is safe, but not profitable, to sink a load targeting a
 246 /// non-address-taken alloca.  Doing so will cause us to not promote the alloca
 247 /// to a register.
 248 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
 249   BasicBlock::iterator BBI = L, E = L->getParent()->end();
 250
 251   for (++BBI; BBI != E; ++BBI)
 252     if (BBI->mayWriteToMemory())
 253       return false;
 254
 255   // Check for non-address taken alloca.  If not address-taken already, it isn't
 256   // profitable to do this xform.
 257   if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
 258     bool isAddressTaken = false;
 259     for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
 260          UI != E; ++UI) {
 261       User *U = *UI;
 262       if (isa<LoadInst>(U)) continue;
 263       if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
 264         // If storing TO the alloca, then the address isn't taken.
 265         if (SI->getOperand(1) == AI) continue;
 266       }
 267       isAddressTaken = true;
 268       break;
 269     }
 270
 271     if (!isAddressTaken && AI->isStaticAlloca())
 272       return false;
 273   }
 274
 275   // If this load is a load from a GEP with a constant offset from an alloca,
 276   // then we don't want to sink it.  In its present form, it will be
 277   // load [constant stack offset].  Sinking it will cause us to have to
 278   // materialize the stack addresses in each predecessor in a register only to
 279   // do a shared load from register in the successor.
 280   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
 281     if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
 282       if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
 283         return false;
 284
 285   return true;
 286 }
 287
 288 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) {
 289   LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0));
 290
 291   // When processing loads, we need to propagate two bits of information to the
 292   // sunk load: whether it is volatile, and what its alignment is.  We currently
 293   // don't sink loads when some have their alignment specified and some don't.
 294   // visitLoadInst will propagate an alignment onto the load when TD is around,
 295   // and if TD isn't around, we can't handle the mixed case.
 296   bool isVolatile = FirstLI->isVolatile();
 297   unsigned LoadAlignment = FirstLI->getAlignment();
 298   unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace();
 299
 300   // We can't sink the load if the loaded value could be modified between the
 301   // load and the PHI.
 302   if (FirstLI->getParent() != PN.getIncomingBlock(0) ||
 303       !isSafeAndProfitableToSinkLoad(FirstLI))
 304     return 0;
 305
 306   // If the PHI is of volatile loads and the load block has multiple
 307   // successors, sinking it would remove a load of the volatile value from
 308   // the path through the other successor.
 309   if (isVolatile &&
 310       FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1)
 311     return 0;
 312
 313   // Check to see if all arguments are the same operation.
 314   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
 315     LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i));
 316     if (!LI || !LI->hasOneUse())
 317       return 0;
 318
 319     // We can't sink the load if the loaded value could be modified between
 320     // the load and the PHI.
 321     if (LI->isVolatile() != isVolatile ||
 322         LI->getParent() != PN.getIncomingBlock(i) ||
 323         LI->getPointerAddressSpace() != LoadAddrSpace ||
 324         !isSafeAndProfitableToSinkLoad(LI))
 325       return 0;
 326
 327     // If some of the loads have an alignment specified but not all of them,
 328     // we can't do the transformation.
 329     if ((LoadAlignment != 0) != (LI->getAlignment() != 0))
 330       return 0;
 331
 332     LoadAlignment = std::min(LoadAlignment, LI->getAlignment());
 333
 334     // If the PHI is of volatile loads and the load block has multiple
 335     // successors, sinking it would remove a load of the volatile value from
 336     // the path through the other successor.
 337     if (isVolatile &&
 338         LI->getParent()->getTerminator()->getNumSuccessors() != 1)
 339       return 0;
 340   }
 341
 342   // Okay, they are all the same operation.  Create a new PHI node of the
 343   // correct type, and PHI together all of the LHS's of the instructions.
 344   PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(),
 345                                    PN.getNumIncomingValues(),
 346                                    PN.getName()+".in");
 347
 348   Value *InVal = FirstLI->getOperand(0);
 349   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
 350
 351   // Add all operands to the new PHI.
 352   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
 353     Value *NewInVal = cast<LoadInst>(PN.getIncomingValue(i))->getOperand(0);
 354     if (NewInVal != InVal)
 355       InVal = 0;
 356     NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
 357   }
 358
 359   Value *PhiVal;
 360   if (InVal) {
 361     // The new PHI unions all of the same values together.  This is really
 362     // common, so we handle it intelligently here for compile-time speed.
 363     PhiVal = InVal;
 364     delete NewPN;
 365   } else {
 366     InsertNewInstBefore(NewPN, PN);
 367     PhiVal = NewPN;
 368   }
 369
 370   // If this was a volatile load that we are merging, make sure to loop through
 371   // and mark all the input loads as non-volatile.  If we don't do this, we will
 372   // insert a new volatile load and the old ones will not be deletable.
 373   if (isVolatile)
 374     for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
 375       cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
 376
 377   LoadInst *NewLI = new LoadInst(PhiVal, "", isVolatile, LoadAlignment);
 378   NewLI->setDebugLoc(FirstLI->getDebugLoc());
 379   return NewLI;
 380 }
 381
 382
 383
 384 /// FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
 385 /// operator and they all are only used by the PHI, PHI together their
 386 /// inputs, and do the operation once, to the result of the PHI.
 387 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
 388   Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
 389
 390   if (isa<GetElementPtrInst>(FirstInst))
 391     return FoldPHIArgGEPIntoPHI(PN);
 392   if (isa<LoadInst>(FirstInst))
 393     return FoldPHIArgLoadIntoPHI(PN);
 394
 395   // Scan the instruction, looking for input operations that can be folded away.
 396   // If all input operands to the phi are the same instruction (e.g. a cast from
 397   // the same type or "+42") we can pull the operation through the PHI, reducing
 398   // code size and simplifying code.
 399   Constant *ConstantOp = 0;
 400   const Type *CastSrcTy = 0;
 401   bool isNUW = false, isNSW = false, isExact = false;
 402
 403   if (isa<CastInst>(FirstInst)) {
 404     CastSrcTy = FirstInst->getOperand(0)->getType();
 405
 406     // Be careful about transforming integer PHIs.  We don't want to pessimize
 407     // the code by turning an i32 into an i1293.
 408     if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) {
 409       if (!ShouldChangeType(PN.getType(), CastSrcTy))
 410         return 0;
 411     }
 412   } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
 413     // Can fold binop, compare or shift here if the RHS is a constant,
 414     // otherwise call FoldPHIArgBinOpIntoPHI.
 415     ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
 416     if (ConstantOp == 0)
 417       return FoldPHIArgBinOpIntoPHI(PN);
 418
 419     if (OverflowingBinaryOperator *BO =
 420         dyn_cast<OverflowingBinaryOperator>(FirstInst)) {
 421       isNUW = BO->hasNoUnsignedWrap();
 422       isNSW = BO->hasNoSignedWrap();
 423     } else if (PossiblyExactOperator *PEO =
 424                dyn_cast<PossiblyExactOperator>(FirstInst))
 425       isExact = PEO->isExact();
 426   } else {
 427     return 0;  // Cannot fold this operation.
 428   }
 429
 430   // Check to see if all arguments are the same operation.
 431   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
 432     Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
 433     if (I == 0 || !I->hasOneUse() || !I->isSameOperationAs(FirstInst))
 434       return 0;
 435     if (CastSrcTy) {
 436       if (I->getOperand(0)->getType() != CastSrcTy)
 437         return 0;  // Cast operation must match.
 438     } else if (I->getOperand(1) != ConstantOp) {
 439       return 0;
 440     }
 441
 442     if (isNUW)
 443       isNUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap();
 444     if (isNSW)
 445       isNSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
 446     if (isExact)
 447       isExact = cast<PossiblyExactOperator>(I)->isExact();
 448   }
 449
 450   // Okay, they are all the same operation.  Create a new PHI node of the
 451   // correct type, and PHI together all of the LHS's of the instructions.
 452   PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
 453                                    PN.getNumIncomingValues(),
 454                                    PN.getName()+".in");
 455
 456   Value *InVal = FirstInst->getOperand(0);
 457   NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
 458
 459   // Add all operands to the new PHI.
 460   for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
 461     Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
 462     if (NewInVal != InVal)
 463       InVal = 0;
 464     NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
 465   }
 466
 467   Value *PhiVal;
 468   if (InVal) {
 469     // The new PHI unions all of the same values together.  This is really
 470     // common, so we handle it intelligently here for compile-time speed.
 471     PhiVal = InVal;
 472     delete NewPN;
 473   } else {
 474     InsertNewInstBefore(NewPN, PN);
 475     PhiVal = NewPN;
 476   }
 477
 478   // Insert and return the new operation.
 479   if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) {
 480     CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal,
 481                                        PN.getType());
 482     NewCI->setDebugLoc(FirstInst->getDebugLoc());
 483     return NewCI;
 484   }
 485
 486   if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) {
 487     BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
 488     if (isNUW) BinOp->setHasNoUnsignedWrap();
 489     if (isNSW) BinOp->setHasNoSignedWrap();
 490     if (isExact) BinOp->setIsExact();
 491     BinOp->setDebugLoc(FirstInst->getDebugLoc());
 492     return BinOp;
 493   }
 494
 495   CmpInst *CIOp = cast<CmpInst>(FirstInst);
 496   CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
 497                                    PhiVal, ConstantOp);
 498   NewCI->setDebugLoc(FirstInst->getDebugLoc());
 499   return NewCI;
 500 }
 501
 502 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
 503 /// that is dead.
 504 static bool DeadPHICycle(PHINode *PN,
 505                          SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
 506   if (PN->use_empty()) return true;
 507   if (!PN->hasOneUse()) return false;
 508
 509   // Remember this node, and if we find the cycle, return.
 510   if (!PotentiallyDeadPHIs.insert(PN))
 511     return true;
 512
 513   // Don't scan crazily complex things.
 514   if (PotentiallyDeadPHIs.size() == 16)
 515     return false;
 516
 517   if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
 518     return DeadPHICycle(PU, PotentiallyDeadPHIs);
 519
 520   return false;
 521 }
 522
 523 /// PHIsEqualValue - Return true if this phi node is always equal to
 524 /// NonPhiInVal.  This happens with mutually cyclic phi nodes like:
 525 ///   z = some value; x = phi (y, z); y = phi (x, z)
 526 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
 527                            SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
 528   // See if we already saw this PHI node.
 529   if (!ValueEqualPHIs.insert(PN))
 530     return true;
 531
 532   // Don't scan crazily complex things.
 533   if (ValueEqualPHIs.size() == 16)
 534     return false;
 535
 536   // Scan the operands to see if they are either phi nodes or are equal to
 537   // the value.
 538   for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
 539     Value *Op = PN->getIncomingValue(i);
 540     if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
 541       if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
 542         return false;
 543     } else if (Op != NonPhiInVal)
 544       return false;
 545   }
 546
 547   return true;
 548 }
 549
 550
 551 namespace {
 552 struct PHIUsageRecord {
 553   unsigned PHIId;     // The ID # of the PHI (something determinstic to sort on)
 554   unsigned Shift;     // The amount shifted.
 555   Instruction *Inst;  // The trunc instruction.
 556
 557   PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User)
 558     : PHIId(pn), Shift(Sh), Inst(User) {}
 559
 560   bool operator<(const PHIUsageRecord &RHS) const {
 561     if (PHIId < RHS.PHIId) return true;
 562     if (PHIId > RHS.PHIId) return false;
 563     if (Shift < RHS.Shift) return true;
 564     if (Shift > RHS.Shift) return false;
 565     return Inst->getType()->getPrimitiveSizeInBits() <
 566            RHS.Inst->getType()->getPrimitiveSizeInBits();
 567   }
 568 };
 569
 570 struct LoweredPHIRecord {
 571   PHINode *PN;        // The PHI that was lowered.
 572   unsigned Shift;     // The amount shifted.
 573   unsigned Width;     // The width extracted.
 574
 575   LoweredPHIRecord(PHINode *pn, unsigned Sh, const Type *Ty)
 576     : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {}
 577
 578   // Ctor form used by DenseMap.
 579   LoweredPHIRecord(PHINode *pn, unsigned Sh)
 580     : PN(pn), Shift(Sh), Width(0) {}
 581 };
 582 }
 583
 584 namespace llvm {
 585   template<>
 586   struct DenseMapInfo<LoweredPHIRecord> {
 587     static inline LoweredPHIRecord getEmptyKey() {
 588       return LoweredPHIRecord(0, 0);
 589     }
 590     static inline LoweredPHIRecord getTombstoneKey() {
 591       return LoweredPHIRecord(0, 1);
 592     }
 593     static unsigned getHashValue(const LoweredPHIRecord &Val) {
 594       return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^
 595              (Val.Width>>3);
 596     }
 597     static bool isEqual(const LoweredPHIRecord &LHS,
 598                         const LoweredPHIRecord &RHS) {
 599       return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift &&
 600              LHS.Width == RHS.Width;
 601     }
 602   };
 603   template <>
 604   struct isPodLike<LoweredPHIRecord> { static const bool value = true; };
 605 }
 606
 607
 608 /// SliceUpIllegalIntegerPHI - This is an integer PHI and we know that it has an
 609 /// illegal type: see if it is only used by trunc or trunc(lshr) operations.  If
 610 /// so, we split the PHI into the various pieces being extracted.  This sort of
 611 /// thing is introduced when SROA promotes an aggregate to large integer values.
 612 ///
 613 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an
 614 /// inttoptr.  We should produce new PHIs in the right type.
 615 ///
 616 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) {
 617   // PHIUsers - Keep track of all of the truncated values extracted from a set
 618   // of PHIs, along with their offset.  These are the things we want to rewrite.
 619   SmallVector<PHIUsageRecord, 16> PHIUsers;
 620
 621   // PHIs are often mutually cyclic, so we keep track of a whole set of PHI
 622   // nodes which are extracted from. PHIsToSlice is a set we use to avoid
 623   // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to
 624   // check the uses of (to ensure they are all extracts).
 625   SmallVector<PHINode*, 8> PHIsToSlice;
 626   SmallPtrSet<PHINode*, 8> PHIsInspected;
 627
 628   PHIsToSlice.push_back(&FirstPhi);
 629   PHIsInspected.insert(&FirstPhi);
 630
 631   for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) {
 632     PHINode *PN = PHIsToSlice[PHIId];
 633
 634     // Scan the input list of the PHI.  If any input is an invoke, and if the
 635     // input is defined in the predecessor, then we won't be split the critical
 636     // edge which is required to insert a truncate.  Because of this, we have to
 637     // bail out.
 638     for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
 639       InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i));
 640       if (II == 0) continue;
 641       if (II->getParent() != PN->getIncomingBlock(i))
 642         continue;
 643
 644       // If we have a phi, and if it's directly in the predecessor, then we have
 645       // a critical edge where we need to put the truncate.  Since we can't
 646       // split the edge in instcombine, we have to bail out.
 647       return 0;
 648     }
 649
 650
 651     for (Value::use_iterator UI = PN->use_begin(), E = PN->use_end();
 652          UI != E; ++UI) {
 653       Instruction *User = cast<Instruction>(*UI);
 654
 655       // If the user is a PHI, inspect its uses recursively.
 656       if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
 657         if (PHIsInspected.insert(UserPN))
 658           PHIsToSlice.push_back(UserPN);
 659         continue;
 660       }
 661
 662       // Truncates are always ok.
 663       if (isa<TruncInst>(User)) {
 664         PHIUsers.push_back(PHIUsageRecord(PHIId, 0, User));
 665         continue;
 666       }
 667
 668       // Otherwise it must be a lshr which can only be used by one trunc.
 669       if (User->getOpcode() != Instruction::LShr ||
 670           !User->hasOneUse() || !isa<TruncInst>(User->use_back()) ||
 671           !isa<ConstantInt>(User->getOperand(1)))
 672         return 0;
 673
 674       unsigned Shift = cast<ConstantInt>(User->getOperand(1))->getZExtValue();
 675       PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, User->use_back()));
 676     }
 677   }
 678
 679   // If we have no users, they must be all self uses, just nuke the PHI.
 680   if (PHIUsers.empty())
 681     return ReplaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType()));
 682
 683   // If this phi node is transformable, create new PHIs for all the pieces
 684   // extracted out of it.  First, sort the users by their offset and size.
 685   array_pod_sort(PHIUsers.begin(), PHIUsers.end());
 686
 687   DEBUG(errs() << "SLICING UP PHI: " << FirstPhi << '\n';
 688             for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
 689               errs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] <<'\n';
 690         );
 691
 692   // PredValues - This is a temporary used when rewriting PHI nodes.  It is
 693   // hoisted out here to avoid construction/destruction thrashing.
 694   DenseMap<BasicBlock*, Value*> PredValues;
 695
 696   // ExtractedVals - Each new PHI we introduce is saved here so we don't
 697   // introduce redundant PHIs.
 698   DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals;
 699
 700   for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) {
 701     unsigned PHIId = PHIUsers[UserI].PHIId;
 702     PHINode *PN = PHIsToSlice[PHIId];
 703     unsigned Offset = PHIUsers[UserI].Shift;
 704     const Type *Ty = PHIUsers[UserI].Inst->getType();
 705
 706     PHINode *EltPHI;
 707
 708     // If we've already lowered a user like this, reuse the previously lowered
 709     // value.
 710     if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == 0) {
 711
 712       // Otherwise, Create the new PHI node for this user.
 713       EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(),
 714                                PN->getName()+".off"+Twine(Offset), PN);
 715       assert(EltPHI->getType() != PN->getType() &&
 716              "Truncate didn't shrink phi?");
 717
 718       for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
 719         BasicBlock *Pred = PN->getIncomingBlock(i);
 720         Value *&PredVal = PredValues[Pred];
 721
 722         // If we already have a value for this predecessor, reuse it.
 723         if (PredVal) {
 724           EltPHI->addIncoming(PredVal, Pred);
 725           continue;
 726         }
 727
 728         // Handle the PHI self-reuse case.
 729         Value *InVal = PN->getIncomingValue(i);
 730         if (InVal == PN) {
 731           PredVal = EltPHI;
 732           EltPHI->addIncoming(PredVal, Pred);
 733           continue;
 734         }
 735
 736         if (PHINode *InPHI = dyn_cast<PHINode>(PN)) {
 737           // If the incoming value was a PHI, and if it was one of the PHIs we
 738           // already rewrote it, just use the lowered value.
 739           if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) {
 740             PredVal = Res;
 741             EltPHI->addIncoming(PredVal, Pred);
 742             continue;
 743           }
 744         }
 745
 746         // Otherwise, do an extract in the predecessor.
 747         Builder->SetInsertPoint(Pred, Pred->getTerminator());
 748         Value *Res = InVal;
 749         if (Offset)
 750           Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(),
 751                                                           Offset), "extract");
 752         Res = Builder->CreateTrunc(Res, Ty, "extract.t");
 753         PredVal = Res;
 754         EltPHI->addIncoming(Res, Pred);
 755
 756         // If the incoming value was a PHI, and if it was one of the PHIs we are
 757         // rewriting, we will ultimately delete the code we inserted.  This
 758         // means we need to revisit that PHI to make sure we extract out the
 759         // needed piece.
 760         if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i)))
 761           if (PHIsInspected.count(OldInVal)) {
 762             unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(),
 763                                           OldInVal)-PHIsToSlice.begin();
 764             PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset,
 765                                               cast<Instruction>(Res)));
 766             ++UserE;
 767           }
 768       }
 769       PredValues.clear();
 770
 771       DEBUG(errs() << "  Made element PHI for offset " << Offset << ": "
 772                    << *EltPHI << '\n');
 773       ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI;
 774     }
 775
 776     // Replace the use of this piece with the PHI node.
 777     ReplaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI);
 778   }
 779
 780   // Replace all the remaining uses of the PHI nodes (self uses and the lshrs)
 781   // with undefs.
 782   Value *Undef = UndefValue::get(FirstPhi.getType());
 783   for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i)
 784     ReplaceInstUsesWith(*PHIsToSlice[i], Undef);
 785   return ReplaceInstUsesWith(FirstPhi, Undef);
 786 }
 787
 788 // PHINode simplification
 789 //
 790 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
 791   if (Value *V = SimplifyInstruction(&PN, TD))
 792     return ReplaceInstUsesWith(PN, V);
 793
 794   // If all PHI operands are the same operation, pull them through the PHI,
 795   // reducing code size.
 796   if (isa<Instruction>(PN.getIncomingValue(0)) &&
 797       isa<Instruction>(PN.getIncomingValue(1)) &&
 798       cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
 799       cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
 800       // FIXME: The hasOneUse check will fail for PHIs that use the value more
 801       // than themselves more than once.
 802       PN.getIncomingValue(0)->hasOneUse())
 803     if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
 804       return Result;
 805
 806   // If this is a trivial cycle in the PHI node graph, remove it.  Basically, if
 807   // this PHI only has a single use (a PHI), and if that PHI only has one use (a
 808   // PHI)... break the cycle.
 809   if (PN.hasOneUse()) {
 810     Instruction *PHIUser = cast<Instruction>(PN.use_back());
 811     if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
 812       SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
 813       PotentiallyDeadPHIs.insert(&PN);
 814       if (DeadPHICycle(PU, PotentiallyDeadPHIs))
 815         return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
 816     }
 817
 818     // If this phi has a single use, and if that use just computes a value for
 819     // the next iteration of a loop, delete the phi.  This occurs with unused
 820     // induction variables, e.g. "for (int j = 0; ; ++j);".  Detecting this
 821     // common case here is good because the only other things that catch this
 822     // are induction variable analysis (sometimes) and ADCE, which is only run
 823     // late.
 824     if (PHIUser->hasOneUse() &&
 825         (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
 826         PHIUser->use_back() == &PN) {
 827       return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
 828     }
 829   }
 830
 831   // We sometimes end up with phi cycles that non-obviously end up being the
 832   // same value, for example:
 833   //   z = some value; x = phi (y, z); y = phi (x, z)
 834   // where the phi nodes don't necessarily need to be in the same block.  Do a
 835   // quick check to see if the PHI node only contains a single non-phi value, if
 836   // so, scan to see if the phi cycle is actually equal to that value.
 837   {
 838     unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues();
 839     // Scan for the first non-phi operand.
 840     while (InValNo != NumIncomingVals &&
 841            isa<PHINode>(PN.getIncomingValue(InValNo)))
 842       ++InValNo;
 843
 844     if (InValNo != NumIncomingVals) {
 845       Value *NonPhiInVal = PN.getIncomingValue(InValNo);
 846
 847       // Scan the rest of the operands to see if there are any conflicts, if so
 848       // there is no need to recursively scan other phis.
 849       for (++InValNo; InValNo != NumIncomingVals; ++InValNo) {
 850         Value *OpVal = PN.getIncomingValue(InValNo);
 851         if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
 852           break;
 853       }
 854
 855       // If we scanned over all operands, then we have one unique value plus
 856       // phi values.  Scan PHI nodes to see if they all merge in each other or
 857       // the value.
 858       if (InValNo == NumIncomingVals) {
 859         SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
 860         if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
 861           return ReplaceInstUsesWith(PN, NonPhiInVal);
 862       }
 863     }
 864   }
 865
 866   // If there are multiple PHIs, sort their operands so that they all list
 867   // the blocks in the same order. This will help identical PHIs be eliminated
 868   // by other passes. Other passes shouldn't depend on this for correctness
 869   // however.
 870   PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin());
 871   if (&PN != FirstPN)
 872     for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) {
 873       BasicBlock *BBA = PN.getIncomingBlock(i);
 874       BasicBlock *BBB = FirstPN->getIncomingBlock(i);
 875       if (BBA != BBB) {
 876         Value *VA = PN.getIncomingValue(i);
 877         unsigned j = PN.getBasicBlockIndex(BBB);
 878         Value *VB = PN.getIncomingValue(j);
 879         PN.setIncomingBlock(i, BBB);
 880         PN.setIncomingValue(i, VB);
 881         PN.setIncomingBlock(j, BBA);
 882         PN.setIncomingValue(j, VA);
 883         // NOTE: Instcombine normally would want us to "return &PN" if we
 884         // modified any of the operands of an instruction.  However, since we
 885         // aren't adding or removing uses (just rearranging them) we don't do
 886         // this in this case.
 887       }
 888     }
 889
 890   // If this is an integer PHI and we know that it has an illegal type, see if
 891   // it is only used by trunc or trunc(lshr) operations.  If so, we split the
 892   // PHI into the various pieces being extracted.  This sort of thing is
 893   // introduced when SROA promotes an aggregate to a single large integer type.
 894   if (PN.getType()->isIntegerTy() && TD &&
 895       !TD->isLegalInteger(PN.getType()->getPrimitiveSizeInBits()))
 896     if (Instruction *Res = SliceUpIllegalIntegerPHI(PN))
 897       return Res;
 898
 899   return 0;
 900 }