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 "InstCombineInternal.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/SmallPtrSet.h" 17 #include "llvm/Analysis/InstructionSimplify.h" 18 #include "llvm/Transforms/Utils/Local.h" 19 using namespace llvm; 20 21 #define DEBUG_TYPE "instcombine" 22 23 /// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the 24 /// adds all have a single use, turn this into a phi and a single binop. 25 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) { 26 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0)); 27 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)); 28 unsigned Opc = FirstInst->getOpcode(); 29 Value *LHSVal = FirstInst->getOperand(0); 30 Value *RHSVal = FirstInst->getOperand(1); 31 32 Type *LHSType = LHSVal->getType(); 33 Type *RHSType = RHSVal->getType(); 34 35 bool isNUW = false, isNSW = false, isExact = false; 36 if (OverflowingBinaryOperator *BO = 37 dyn_cast<OverflowingBinaryOperator>(FirstInst)) { 38 isNUW = BO->hasNoUnsignedWrap(); 39 isNSW = BO->hasNoSignedWrap(); 40 } else if (PossiblyExactOperator *PEO = 41 dyn_cast<PossiblyExactOperator>(FirstInst)) 42 isExact = PEO->isExact(); 43 44 // Scan to see if all operands are the same opcode, and all have one use. 45 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) { 46 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i)); 47 if (!I || I->getOpcode() != Opc || !I->hasOneUse() || 48 // Verify type of the LHS matches so we don't fold cmp's of different 49 // types. 50 I->getOperand(0)->getType() != LHSType || 51 I->getOperand(1)->getType() != RHSType) 52 return nullptr; 53 54 // If they are CmpInst instructions, check their predicates 55 if (CmpInst *CI = dyn_cast<CmpInst>(I)) 56 if (CI->getPredicate() != cast<CmpInst>(FirstInst)->getPredicate()) 57 return nullptr; 58 59 if (isNUW) 60 isNUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap(); 61 if (isNSW) 62 isNSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap(); 63 if (isExact) 64 isExact = cast<PossiblyExactOperator>(I)->isExact(); 65 66 // Keep track of which operand needs a phi node. 67 if (I->getOperand(0) != LHSVal) LHSVal = nullptr; 68 if (I->getOperand(1) != RHSVal) RHSVal = nullptr; 69 } 70 71 // If both LHS and RHS would need a PHI, don't do this transformation, 72 // because it would increase the number of PHIs entering the block, 73 // which leads to higher register pressure. This is especially 74 // bad when the PHIs are in the header of a loop. 75 if (!LHSVal && !RHSVal) 76 return nullptr; 77 78 // Otherwise, this is safe to transform! 79 80 Value *InLHS = FirstInst->getOperand(0); 81 Value *InRHS = FirstInst->getOperand(1); 82 PHINode *NewLHS = nullptr, *NewRHS = nullptr; 83 if (!LHSVal) { 84 NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(), 85 FirstInst->getOperand(0)->getName() + ".pn"); 86 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0)); 87 InsertNewInstBefore(NewLHS, PN); 88 LHSVal = NewLHS; 89 } 90 91 if (!RHSVal) { 92 NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(), 93 FirstInst->getOperand(1)->getName() + ".pn"); 94 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0)); 95 InsertNewInstBefore(NewRHS, PN); 96 RHSVal = NewRHS; 97 } 98 99 // Add all operands to the new PHIs. 100 if (NewLHS || NewRHS) { 101 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 102 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i)); 103 if (NewLHS) { 104 Value *NewInLHS = InInst->getOperand(0); 105 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i)); 106 } 107 if (NewRHS) { 108 Value *NewInRHS = InInst->getOperand(1); 109 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i)); 110 } 111 } 112 } 113 114 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) { 115 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), 116 LHSVal, RHSVal); 117 NewCI->setDebugLoc(FirstInst->getDebugLoc()); 118 return NewCI; 119 } 120 121 BinaryOperator *BinOp = cast<BinaryOperator>(FirstInst); 122 BinaryOperator *NewBinOp = 123 BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal); 124 if (isNUW) NewBinOp->setHasNoUnsignedWrap(); 125 if (isNSW) NewBinOp->setHasNoSignedWrap(); 126 if (isExact) NewBinOp->setIsExact(); 127 NewBinOp->setDebugLoc(FirstInst->getDebugLoc()); 128 return NewBinOp; 129 } 130 131 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) { 132 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0)); 133 134 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(), 135 FirstInst->op_end()); 136 // This is true if all GEP bases are allocas and if all indices into them are 137 // constants. 138 bool AllBasePointersAreAllocas = true; 139 140 // We don't want to replace this phi if the replacement would require 141 // more than one phi, which leads to higher register pressure. This is 142 // especially bad when the PHIs are in the header of a loop. 143 bool NeededPhi = false; 144 145 bool AllInBounds = true; 146 147 // Scan to see if all operands are the same opcode, and all have one use. 148 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) { 149 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i)); 150 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() || 151 GEP->getNumOperands() != FirstInst->getNumOperands()) 152 return nullptr; 153 154 AllInBounds &= GEP->isInBounds(); 155 156 // Keep track of whether or not all GEPs are of alloca pointers. 157 if (AllBasePointersAreAllocas && 158 (!isa<AllocaInst>(GEP->getOperand(0)) || 159 !GEP->hasAllConstantIndices())) 160 AllBasePointersAreAllocas = false; 161 162 // Compare the operand lists. 163 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) { 164 if (FirstInst->getOperand(op) == GEP->getOperand(op)) 165 continue; 166 167 // Don't merge two GEPs when two operands differ (introducing phi nodes) 168 // if one of the PHIs has a constant for the index. The index may be 169 // substantially cheaper to compute for the constants, so making it a 170 // variable index could pessimize the path. This also handles the case 171 // for struct indices, which must always be constant. 172 if (isa<ConstantInt>(FirstInst->getOperand(op)) || 173 isa<ConstantInt>(GEP->getOperand(op))) 174 return nullptr; 175 176 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType()) 177 return nullptr; 178 179 // If we already needed a PHI for an earlier operand, and another operand 180 // also requires a PHI, we'd be introducing more PHIs than we're 181 // eliminating, which increases register pressure on entry to the PHI's 182 // block. 183 if (NeededPhi) 184 return nullptr; 185 186 FixedOperands[op] = nullptr; // Needs a PHI. 187 NeededPhi = true; 188 } 189 } 190 191 // If all of the base pointers of the PHI'd GEPs are from allocas, don't 192 // bother doing this transformation. At best, this will just save a bit of 193 // offset calculation, but all the predecessors will have to materialize the 194 // stack address into a register anyway. We'd actually rather *clone* the 195 // load up into the predecessors so that we have a load of a gep of an alloca, 196 // which can usually all be folded into the load. 197 if (AllBasePointersAreAllocas) 198 return nullptr; 199 200 // Otherwise, this is safe to transform. Insert PHI nodes for each operand 201 // that is variable. 202 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size()); 203 204 bool HasAnyPHIs = false; 205 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) { 206 if (FixedOperands[i]) continue; // operand doesn't need a phi. 207 Value *FirstOp = FirstInst->getOperand(i); 208 PHINode *NewPN = PHINode::Create(FirstOp->getType(), e, 209 FirstOp->getName()+".pn"); 210 InsertNewInstBefore(NewPN, PN); 211 212 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0)); 213 OperandPhis[i] = NewPN; 214 FixedOperands[i] = NewPN; 215 HasAnyPHIs = true; 216 } 217 218 219 // Add all operands to the new PHIs. 220 if (HasAnyPHIs) { 221 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 222 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i)); 223 BasicBlock *InBB = PN.getIncomingBlock(i); 224 225 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op) 226 if (PHINode *OpPhi = OperandPhis[op]) 227 OpPhi->addIncoming(InGEP->getOperand(op), InBB); 228 } 229 } 230 231 Value *Base = FixedOperands[0]; 232 GetElementPtrInst *NewGEP = 233 GetElementPtrInst::Create(FirstInst->getSourceElementType(), Base, 234 makeArrayRef(FixedOperands).slice(1)); 235 if (AllInBounds) NewGEP->setIsInBounds(); 236 NewGEP->setDebugLoc(FirstInst->getDebugLoc()); 237 return NewGEP; 238 } 239 240 241 /// Return true if we know that it is safe to sink the load out of the block 242 /// that defines it. This means that it must be obvious the value of the load is 243 /// not changed from the point of the load to 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->getIterator(), 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 (User *U : AI->users()) { 260 if (isa<LoadInst>(U)) continue; 261 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 262 // If storing TO the alloca, then the address isn't taken. 263 if (SI->getOperand(1) == AI) continue; 264 } 265 isAddressTaken = true; 266 break; 267 } 268 269 if (!isAddressTaken && AI->isStaticAlloca()) 270 return false; 271 } 272 273 // If this load is a load from a GEP with a constant offset from an alloca, 274 // then we don't want to sink it. In its present form, it will be 275 // load [constant stack offset]. Sinking it will cause us to have to 276 // materialize the stack addresses in each predecessor in a register only to 277 // do a shared load from register in the successor. 278 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0))) 279 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0))) 280 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices()) 281 return false; 282 283 return true; 284 } 285 286 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) { 287 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0)); 288 289 // FIXME: This is overconservative; this transform is allowed in some cases 290 // for atomic operations. 291 if (FirstLI->isAtomic()) 292 return nullptr; 293 294 // When processing loads, we need to propagate two bits of information to the 295 // sunk load: whether it is volatile, and what its alignment is. We currently 296 // don't sink loads when some have their alignment specified and some don't. 297 // visitLoadInst will propagate an alignment onto the load when TD is around, 298 // and if TD isn't around, we can't handle the mixed case. 299 bool isVolatile = FirstLI->isVolatile(); 300 unsigned LoadAlignment = FirstLI->getAlignment(); 301 unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace(); 302 303 // We can't sink the load if the loaded value could be modified between the 304 // load and the PHI. 305 if (FirstLI->getParent() != PN.getIncomingBlock(0) || 306 !isSafeAndProfitableToSinkLoad(FirstLI)) 307 return nullptr; 308 309 // If the PHI is of volatile loads and the load block has multiple 310 // successors, sinking it would remove a load of the volatile value from 311 // the path through the other successor. 312 if (isVolatile && 313 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1) 314 return nullptr; 315 316 // Check to see if all arguments are the same operation. 317 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 318 LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i)); 319 if (!LI || !LI->hasOneUse()) 320 return nullptr; 321 322 // We can't sink the load if the loaded value could be modified between 323 // the load and the PHI. 324 if (LI->isVolatile() != isVolatile || 325 LI->getParent() != PN.getIncomingBlock(i) || 326 LI->getPointerAddressSpace() != LoadAddrSpace || 327 !isSafeAndProfitableToSinkLoad(LI)) 328 return nullptr; 329 330 // If some of the loads have an alignment specified but not all of them, 331 // we can't do the transformation. 332 if ((LoadAlignment != 0) != (LI->getAlignment() != 0)) 333 return nullptr; 334 335 LoadAlignment = std::min(LoadAlignment, LI->getAlignment()); 336 337 // If the PHI is of volatile loads and the load block has multiple 338 // successors, sinking it would remove a load of the volatile value from 339 // the path through the other successor. 340 if (isVolatile && 341 LI->getParent()->getTerminator()->getNumSuccessors() != 1) 342 return nullptr; 343 } 344 345 // Okay, they are all the same operation. Create a new PHI node of the 346 // correct type, and PHI together all of the LHS's of the instructions. 347 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(), 348 PN.getNumIncomingValues(), 349 PN.getName()+".in"); 350 351 Value *InVal = FirstLI->getOperand(0); 352 NewPN->addIncoming(InVal, PN.getIncomingBlock(0)); 353 LoadInst *NewLI = new LoadInst(NewPN, "", isVolatile, LoadAlignment); 354 355 unsigned KnownIDs[] = { 356 LLVMContext::MD_tbaa, 357 LLVMContext::MD_range, 358 LLVMContext::MD_invariant_load, 359 LLVMContext::MD_alias_scope, 360 LLVMContext::MD_noalias, 361 LLVMContext::MD_nonnull, 362 LLVMContext::MD_align, 363 LLVMContext::MD_dereferenceable, 364 LLVMContext::MD_dereferenceable_or_null, 365 }; 366 367 for (unsigned ID : KnownIDs) 368 NewLI->setMetadata(ID, FirstLI->getMetadata(ID)); 369 370 // Add all operands to the new PHI and combine TBAA metadata. 371 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 372 LoadInst *LI = cast<LoadInst>(PN.getIncomingValue(i)); 373 combineMetadata(NewLI, LI, KnownIDs); 374 Value *NewInVal = LI->getOperand(0); 375 if (NewInVal != InVal) 376 InVal = nullptr; 377 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i)); 378 } 379 380 if (InVal) { 381 // The new PHI unions all of the same values together. This is really 382 // common, so we handle it intelligently here for compile-time speed. 383 NewLI->setOperand(0, InVal); 384 delete NewPN; 385 } else { 386 InsertNewInstBefore(NewPN, PN); 387 } 388 389 // If this was a volatile load that we are merging, make sure to loop through 390 // and mark all the input loads as non-volatile. If we don't do this, we will 391 // insert a new volatile load and the old ones will not be deletable. 392 if (isVolatile) 393 for (Value *IncValue : PN.incoming_values()) 394 cast<LoadInst>(IncValue)->setVolatile(false); 395 396 NewLI->setDebugLoc(FirstLI->getDebugLoc()); 397 return NewLI; 398 } 399 400 /// TODO: This function could handle other cast types, but then it might 401 /// require special-casing a cast from the 'i1' type. See the comment in 402 /// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types. 403 Instruction *InstCombiner::FoldPHIArgZextsIntoPHI(PHINode &Phi) { 404 // We cannot create a new instruction after the PHI if the terminator is an 405 // EHPad because there is no valid insertion point. 406 if (TerminatorInst *TI = Phi.getParent()->getTerminator()) 407 if (TI->isEHPad()) 408 return nullptr; 409 410 // Early exit for the common case of a phi with two operands. These are 411 // handled elsewhere. See the comment below where we check the count of zexts 412 // and constants for more details. 413 unsigned NumIncomingValues = Phi.getNumIncomingValues(); 414 if (NumIncomingValues < 3) 415 return nullptr; 416 417 // Find the narrower type specified by the first zext. 418 Type *NarrowType = nullptr; 419 for (Value *V : Phi.incoming_values()) { 420 if (auto *Zext = dyn_cast<ZExtInst>(V)) { 421 NarrowType = Zext->getSrcTy(); 422 break; 423 } 424 } 425 if (!NarrowType) 426 return nullptr; 427 428 // Walk the phi operands checking that we only have zexts or constants that 429 // we can shrink for free. Store the new operands for the new phi. 430 SmallVector<Value *, 4> NewIncoming; 431 unsigned NumZexts = 0; 432 unsigned NumConsts = 0; 433 for (Value *V : Phi.incoming_values()) { 434 if (auto *Zext = dyn_cast<ZExtInst>(V)) { 435 // All zexts must be identical and have one use. 436 if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUse()) 437 return nullptr; 438 NewIncoming.push_back(Zext->getOperand(0)); 439 NumZexts++; 440 } else if (auto *C = dyn_cast<Constant>(V)) { 441 // Make sure that constants can fit in the new type. 442 Constant *Trunc = ConstantExpr::getTrunc(C, NarrowType); 443 if (ConstantExpr::getZExt(Trunc, C->getType()) != C) 444 return nullptr; 445 NewIncoming.push_back(Trunc); 446 NumConsts++; 447 } else { 448 // If it's not a cast or a constant, bail out. 449 return nullptr; 450 } 451 } 452 453 // The more common cases of a phi with no constant operands or just one 454 // variable operand are handled by FoldPHIArgOpIntoPHI() and FoldOpIntoPhi() 455 // respectively. FoldOpIntoPhi() wants to do the opposite transform that is 456 // performed here. It tries to replicate a cast in the phi operand's basic 457 // block to expose other folding opportunities. Thus, InstCombine will 458 // infinite loop without this check. 459 if (NumConsts == 0 || NumZexts < 2) 460 return nullptr; 461 462 // All incoming values are zexts or constants that are safe to truncate. 463 // Create a new phi node of the narrow type, phi together all of the new 464 // operands, and zext the result back to the original type. 465 PHINode *NewPhi = PHINode::Create(NarrowType, NumIncomingValues, 466 Phi.getName() + ".shrunk"); 467 for (unsigned i = 0; i != NumIncomingValues; ++i) 468 NewPhi->addIncoming(NewIncoming[i], Phi.getIncomingBlock(i)); 469 470 InsertNewInstBefore(NewPhi, Phi); 471 return CastInst::CreateZExtOrBitCast(NewPhi, Phi.getType()); 472 } 473 474 /// If all operands to a PHI node are the same "unary" operator and they all are 475 /// only used by the PHI, PHI together their inputs, and do the operation once, 476 /// to the result of the PHI. 477 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) { 478 // We cannot create a new instruction after the PHI if the terminator is an 479 // EHPad because there is no valid insertion point. 480 if (TerminatorInst *TI = PN.getParent()->getTerminator()) 481 if (TI->isEHPad()) 482 return nullptr; 483 484 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0)); 485 486 if (isa<GetElementPtrInst>(FirstInst)) 487 return FoldPHIArgGEPIntoPHI(PN); 488 if (isa<LoadInst>(FirstInst)) 489 return FoldPHIArgLoadIntoPHI(PN); 490 491 // Scan the instruction, looking for input operations that can be folded away. 492 // If all input operands to the phi are the same instruction (e.g. a cast from 493 // the same type or "+42") we can pull the operation through the PHI, reducing 494 // code size and simplifying code. 495 Constant *ConstantOp = nullptr; 496 Type *CastSrcTy = nullptr; 497 bool isNUW = false, isNSW = false, isExact = false; 498 499 if (isa<CastInst>(FirstInst)) { 500 CastSrcTy = FirstInst->getOperand(0)->getType(); 501 502 // Be careful about transforming integer PHIs. We don't want to pessimize 503 // the code by turning an i32 into an i1293. 504 if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) { 505 if (!ShouldChangeType(PN.getType(), CastSrcTy)) 506 return nullptr; 507 } 508 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) { 509 // Can fold binop, compare or shift here if the RHS is a constant, 510 // otherwise call FoldPHIArgBinOpIntoPHI. 511 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1)); 512 if (!ConstantOp) 513 return FoldPHIArgBinOpIntoPHI(PN); 514 515 if (OverflowingBinaryOperator *BO = 516 dyn_cast<OverflowingBinaryOperator>(FirstInst)) { 517 isNUW = BO->hasNoUnsignedWrap(); 518 isNSW = BO->hasNoSignedWrap(); 519 } else if (PossiblyExactOperator *PEO = 520 dyn_cast<PossiblyExactOperator>(FirstInst)) 521 isExact = PEO->isExact(); 522 } else { 523 return nullptr; // Cannot fold this operation. 524 } 525 526 // Check to see if all arguments are the same operation. 527 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 528 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i)); 529 if (!I || !I->hasOneUse() || !I->isSameOperationAs(FirstInst)) 530 return nullptr; 531 if (CastSrcTy) { 532 if (I->getOperand(0)->getType() != CastSrcTy) 533 return nullptr; // Cast operation must match. 534 } else if (I->getOperand(1) != ConstantOp) { 535 return nullptr; 536 } 537 538 if (isNUW) 539 isNUW = cast<OverflowingBinaryOperator>(I)->hasNoUnsignedWrap(); 540 if (isNSW) 541 isNSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap(); 542 if (isExact) 543 isExact = cast<PossiblyExactOperator>(I)->isExact(); 544 } 545 546 // Okay, they are all the same operation. Create a new PHI node of the 547 // correct type, and PHI together all of the LHS's of the instructions. 548 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(), 549 PN.getNumIncomingValues(), 550 PN.getName()+".in"); 551 552 Value *InVal = FirstInst->getOperand(0); 553 NewPN->addIncoming(InVal, PN.getIncomingBlock(0)); 554 555 // Add all operands to the new PHI. 556 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 557 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0); 558 if (NewInVal != InVal) 559 InVal = nullptr; 560 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i)); 561 } 562 563 Value *PhiVal; 564 if (InVal) { 565 // The new PHI unions all of the same values together. This is really 566 // common, so we handle it intelligently here for compile-time speed. 567 PhiVal = InVal; 568 delete NewPN; 569 } else { 570 InsertNewInstBefore(NewPN, PN); 571 PhiVal = NewPN; 572 } 573 574 // Insert and return the new operation. 575 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) { 576 CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal, 577 PN.getType()); 578 NewCI->setDebugLoc(FirstInst->getDebugLoc()); 579 return NewCI; 580 } 581 582 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) { 583 BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp); 584 if (isNUW) BinOp->setHasNoUnsignedWrap(); 585 if (isNSW) BinOp->setHasNoSignedWrap(); 586 if (isExact) BinOp->setIsExact(); 587 BinOp->setDebugLoc(FirstInst->getDebugLoc()); 588 return BinOp; 589 } 590 591 CmpInst *CIOp = cast<CmpInst>(FirstInst); 592 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), 593 PhiVal, ConstantOp); 594 NewCI->setDebugLoc(FirstInst->getDebugLoc()); 595 return NewCI; 596 } 597 598 /// Return true if this PHI node is only used by a PHI node cycle that is dead. 599 static bool DeadPHICycle(PHINode *PN, 600 SmallPtrSetImpl<PHINode*> &PotentiallyDeadPHIs) { 601 if (PN->use_empty()) return true; 602 if (!PN->hasOneUse()) return false; 603 604 // Remember this node, and if we find the cycle, return. 605 if (!PotentiallyDeadPHIs.insert(PN).second) 606 return true; 607 608 // Don't scan crazily complex things. 609 if (PotentiallyDeadPHIs.size() == 16) 610 return false; 611 612 if (PHINode *PU = dyn_cast<PHINode>(PN->user_back())) 613 return DeadPHICycle(PU, PotentiallyDeadPHIs); 614 615 return false; 616 } 617 618 /// Return true if this phi node is always equal to NonPhiInVal. 619 /// This happens with mutually cyclic phi nodes like: 620 /// z = some value; x = phi (y, z); y = phi (x, z) 621 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal, 622 SmallPtrSetImpl<PHINode*> &ValueEqualPHIs) { 623 // See if we already saw this PHI node. 624 if (!ValueEqualPHIs.insert(PN).second) 625 return true; 626 627 // Don't scan crazily complex things. 628 if (ValueEqualPHIs.size() == 16) 629 return false; 630 631 // Scan the operands to see if they are either phi nodes or are equal to 632 // the value. 633 for (Value *Op : PN->incoming_values()) { 634 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) { 635 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs)) 636 return false; 637 } else if (Op != NonPhiInVal) 638 return false; 639 } 640 641 return true; 642 } 643 644 645 namespace { 646 struct PHIUsageRecord { 647 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on) 648 unsigned Shift; // The amount shifted. 649 Instruction *Inst; // The trunc instruction. 650 651 PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User) 652 : PHIId(pn), Shift(Sh), Inst(User) {} 653 654 bool operator<(const PHIUsageRecord &RHS) const { 655 if (PHIId < RHS.PHIId) return true; 656 if (PHIId > RHS.PHIId) return false; 657 if (Shift < RHS.Shift) return true; 658 if (Shift > RHS.Shift) return false; 659 return Inst->getType()->getPrimitiveSizeInBits() < 660 RHS.Inst->getType()->getPrimitiveSizeInBits(); 661 } 662 }; 663 664 struct LoweredPHIRecord { 665 PHINode *PN; // The PHI that was lowered. 666 unsigned Shift; // The amount shifted. 667 unsigned Width; // The width extracted. 668 669 LoweredPHIRecord(PHINode *pn, unsigned Sh, Type *Ty) 670 : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {} 671 672 // Ctor form used by DenseMap. 673 LoweredPHIRecord(PHINode *pn, unsigned Sh) 674 : PN(pn), Shift(Sh), Width(0) {} 675 }; 676 } 677 678 namespace llvm { 679 template<> 680 struct DenseMapInfo<LoweredPHIRecord> { 681 static inline LoweredPHIRecord getEmptyKey() { 682 return LoweredPHIRecord(nullptr, 0); 683 } 684 static inline LoweredPHIRecord getTombstoneKey() { 685 return LoweredPHIRecord(nullptr, 1); 686 } 687 static unsigned getHashValue(const LoweredPHIRecord &Val) { 688 return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^ 689 (Val.Width>>3); 690 } 691 static bool isEqual(const LoweredPHIRecord &LHS, 692 const LoweredPHIRecord &RHS) { 693 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift && 694 LHS.Width == RHS.Width; 695 } 696 }; 697 } 698 699 700 /// This is an integer PHI and we know that it has an illegal type: see if it is 701 /// only used by trunc or trunc(lshr) operations. If so, we split the PHI into 702 /// the various pieces being extracted. This sort of thing is introduced when 703 /// SROA promotes an aggregate to large integer values. 704 /// 705 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an 706 /// inttoptr. We should produce new PHIs in the right type. 707 /// 708 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) { 709 // PHIUsers - Keep track of all of the truncated values extracted from a set 710 // of PHIs, along with their offset. These are the things we want to rewrite. 711 SmallVector<PHIUsageRecord, 16> PHIUsers; 712 713 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI 714 // nodes which are extracted from. PHIsToSlice is a set we use to avoid 715 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to 716 // check the uses of (to ensure they are all extracts). 717 SmallVector<PHINode*, 8> PHIsToSlice; 718 SmallPtrSet<PHINode*, 8> PHIsInspected; 719 720 PHIsToSlice.push_back(&FirstPhi); 721 PHIsInspected.insert(&FirstPhi); 722 723 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) { 724 PHINode *PN = PHIsToSlice[PHIId]; 725 726 // Scan the input list of the PHI. If any input is an invoke, and if the 727 // input is defined in the predecessor, then we won't be split the critical 728 // edge which is required to insert a truncate. Because of this, we have to 729 // bail out. 730 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 731 InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i)); 732 if (!II) continue; 733 if (II->getParent() != PN->getIncomingBlock(i)) 734 continue; 735 736 // If we have a phi, and if it's directly in the predecessor, then we have 737 // a critical edge where we need to put the truncate. Since we can't 738 // split the edge in instcombine, we have to bail out. 739 return nullptr; 740 } 741 742 for (User *U : PN->users()) { 743 Instruction *UserI = cast<Instruction>(U); 744 745 // If the user is a PHI, inspect its uses recursively. 746 if (PHINode *UserPN = dyn_cast<PHINode>(UserI)) { 747 if (PHIsInspected.insert(UserPN).second) 748 PHIsToSlice.push_back(UserPN); 749 continue; 750 } 751 752 // Truncates are always ok. 753 if (isa<TruncInst>(UserI)) { 754 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, UserI)); 755 continue; 756 } 757 758 // Otherwise it must be a lshr which can only be used by one trunc. 759 if (UserI->getOpcode() != Instruction::LShr || 760 !UserI->hasOneUse() || !isa<TruncInst>(UserI->user_back()) || 761 !isa<ConstantInt>(UserI->getOperand(1))) 762 return nullptr; 763 764 unsigned Shift = cast<ConstantInt>(UserI->getOperand(1))->getZExtValue(); 765 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, UserI->user_back())); 766 } 767 } 768 769 // If we have no users, they must be all self uses, just nuke the PHI. 770 if (PHIUsers.empty()) 771 return ReplaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType())); 772 773 // If this phi node is transformable, create new PHIs for all the pieces 774 // extracted out of it. First, sort the users by their offset and size. 775 array_pod_sort(PHIUsers.begin(), PHIUsers.end()); 776 777 DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n'; 778 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i) 779 dbgs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] << '\n'; 780 ); 781 782 // PredValues - This is a temporary used when rewriting PHI nodes. It is 783 // hoisted out here to avoid construction/destruction thrashing. 784 DenseMap<BasicBlock*, Value*> PredValues; 785 786 // ExtractedVals - Each new PHI we introduce is saved here so we don't 787 // introduce redundant PHIs. 788 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals; 789 790 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) { 791 unsigned PHIId = PHIUsers[UserI].PHIId; 792 PHINode *PN = PHIsToSlice[PHIId]; 793 unsigned Offset = PHIUsers[UserI].Shift; 794 Type *Ty = PHIUsers[UserI].Inst->getType(); 795 796 PHINode *EltPHI; 797 798 // If we've already lowered a user like this, reuse the previously lowered 799 // value. 800 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) { 801 802 // Otherwise, Create the new PHI node for this user. 803 EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(), 804 PN->getName()+".off"+Twine(Offset), PN); 805 assert(EltPHI->getType() != PN->getType() && 806 "Truncate didn't shrink phi?"); 807 808 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 809 BasicBlock *Pred = PN->getIncomingBlock(i); 810 Value *&PredVal = PredValues[Pred]; 811 812 // If we already have a value for this predecessor, reuse it. 813 if (PredVal) { 814 EltPHI->addIncoming(PredVal, Pred); 815 continue; 816 } 817 818 // Handle the PHI self-reuse case. 819 Value *InVal = PN->getIncomingValue(i); 820 if (InVal == PN) { 821 PredVal = EltPHI; 822 EltPHI->addIncoming(PredVal, Pred); 823 continue; 824 } 825 826 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) { 827 // If the incoming value was a PHI, and if it was one of the PHIs we 828 // already rewrote it, just use the lowered value. 829 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) { 830 PredVal = Res; 831 EltPHI->addIncoming(PredVal, Pred); 832 continue; 833 } 834 } 835 836 // Otherwise, do an extract in the predecessor. 837 Builder->SetInsertPoint(Pred->getTerminator()); 838 Value *Res = InVal; 839 if (Offset) 840 Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(), 841 Offset), "extract"); 842 Res = Builder->CreateTrunc(Res, Ty, "extract.t"); 843 PredVal = Res; 844 EltPHI->addIncoming(Res, Pred); 845 846 // If the incoming value was a PHI, and if it was one of the PHIs we are 847 // rewriting, we will ultimately delete the code we inserted. This 848 // means we need to revisit that PHI to make sure we extract out the 849 // needed piece. 850 if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i))) 851 if (PHIsInspected.count(OldInVal)) { 852 unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(), 853 OldInVal)-PHIsToSlice.begin(); 854 PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset, 855 cast<Instruction>(Res))); 856 ++UserE; 857 } 858 } 859 PredValues.clear(); 860 861 DEBUG(dbgs() << " Made element PHI for offset " << Offset << ": " 862 << *EltPHI << '\n'); 863 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI; 864 } 865 866 // Replace the use of this piece with the PHI node. 867 ReplaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI); 868 } 869 870 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs) 871 // with undefs. 872 Value *Undef = UndefValue::get(FirstPhi.getType()); 873 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i) 874 ReplaceInstUsesWith(*PHIsToSlice[i], Undef); 875 return ReplaceInstUsesWith(FirstPhi, Undef); 876 } 877 878 // PHINode simplification 879 // 880 Instruction *InstCombiner::visitPHINode(PHINode &PN) { 881 if (Value *V = SimplifyInstruction(&PN, DL, TLI, DT, AC)) 882 return ReplaceInstUsesWith(PN, V); 883 884 if (Instruction *Result = FoldPHIArgZextsIntoPHI(PN)) 885 return Result; 886 887 // If all PHI operands are the same operation, pull them through the PHI, 888 // reducing code size. 889 if (isa<Instruction>(PN.getIncomingValue(0)) && 890 isa<Instruction>(PN.getIncomingValue(1)) && 891 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() == 892 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() && 893 // FIXME: The hasOneUse check will fail for PHIs that use the value more 894 // than themselves more than once. 895 PN.getIncomingValue(0)->hasOneUse()) 896 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN)) 897 return Result; 898 899 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if 900 // this PHI only has a single use (a PHI), and if that PHI only has one use (a 901 // PHI)... break the cycle. 902 if (PN.hasOneUse()) { 903 Instruction *PHIUser = cast<Instruction>(PN.user_back()); 904 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) { 905 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs; 906 PotentiallyDeadPHIs.insert(&PN); 907 if (DeadPHICycle(PU, PotentiallyDeadPHIs)) 908 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType())); 909 } 910 911 // If this phi has a single use, and if that use just computes a value for 912 // the next iteration of a loop, delete the phi. This occurs with unused 913 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this 914 // common case here is good because the only other things that catch this 915 // are induction variable analysis (sometimes) and ADCE, which is only run 916 // late. 917 if (PHIUser->hasOneUse() && 918 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) && 919 PHIUser->user_back() == &PN) { 920 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType())); 921 } 922 } 923 924 // We sometimes end up with phi cycles that non-obviously end up being the 925 // same value, for example: 926 // z = some value; x = phi (y, z); y = phi (x, z) 927 // where the phi nodes don't necessarily need to be in the same block. Do a 928 // quick check to see if the PHI node only contains a single non-phi value, if 929 // so, scan to see if the phi cycle is actually equal to that value. 930 { 931 unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues(); 932 // Scan for the first non-phi operand. 933 while (InValNo != NumIncomingVals && 934 isa<PHINode>(PN.getIncomingValue(InValNo))) 935 ++InValNo; 936 937 if (InValNo != NumIncomingVals) { 938 Value *NonPhiInVal = PN.getIncomingValue(InValNo); 939 940 // Scan the rest of the operands to see if there are any conflicts, if so 941 // there is no need to recursively scan other phis. 942 for (++InValNo; InValNo != NumIncomingVals; ++InValNo) { 943 Value *OpVal = PN.getIncomingValue(InValNo); 944 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal)) 945 break; 946 } 947 948 // If we scanned over all operands, then we have one unique value plus 949 // phi values. Scan PHI nodes to see if they all merge in each other or 950 // the value. 951 if (InValNo == NumIncomingVals) { 952 SmallPtrSet<PHINode*, 16> ValueEqualPHIs; 953 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs)) 954 return ReplaceInstUsesWith(PN, NonPhiInVal); 955 } 956 } 957 } 958 959 // If there are multiple PHIs, sort their operands so that they all list 960 // the blocks in the same order. This will help identical PHIs be eliminated 961 // by other passes. Other passes shouldn't depend on this for correctness 962 // however. 963 PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin()); 964 if (&PN != FirstPN) 965 for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) { 966 BasicBlock *BBA = PN.getIncomingBlock(i); 967 BasicBlock *BBB = FirstPN->getIncomingBlock(i); 968 if (BBA != BBB) { 969 Value *VA = PN.getIncomingValue(i); 970 unsigned j = PN.getBasicBlockIndex(BBB); 971 Value *VB = PN.getIncomingValue(j); 972 PN.setIncomingBlock(i, BBB); 973 PN.setIncomingValue(i, VB); 974 PN.setIncomingBlock(j, BBA); 975 PN.setIncomingValue(j, VA); 976 // NOTE: Instcombine normally would want us to "return &PN" if we 977 // modified any of the operands of an instruction. However, since we 978 // aren't adding or removing uses (just rearranging them) we don't do 979 // this in this case. 980 } 981 } 982 983 // If this is an integer PHI and we know that it has an illegal type, see if 984 // it is only used by trunc or trunc(lshr) operations. If so, we split the 985 // PHI into the various pieces being extracted. This sort of thing is 986 // introduced when SROA promotes an aggregate to a single large integer type. 987 if (PN.getType()->isIntegerTy() && 988 !DL.isLegalInteger(PN.getType()->getPrimitiveSizeInBits())) 989 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN)) 990 return Res; 991 992 return nullptr; 993 } 994