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/Analysis/ValueTracking.h" 19 #include "llvm/IR/PatternMatch.h" 20 #include "llvm/Transforms/Utils/Local.h" 21 using namespace llvm; 22 using namespace llvm::PatternMatch; 23 24 #define DEBUG_TYPE "instcombine" 25 26 /// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the 27 /// adds all have a single use, turn this into a phi and a single binop. 28 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) { 29 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0)); 30 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)); 31 unsigned Opc = FirstInst->getOpcode(); 32 Value *LHSVal = FirstInst->getOperand(0); 33 Value *RHSVal = FirstInst->getOperand(1); 34 35 Type *LHSType = LHSVal->getType(); 36 Type *RHSType = RHSVal->getType(); 37 38 // Scan to see if all operands are the same opcode, and all have one use. 39 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) { 40 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i)); 41 if (!I || I->getOpcode() != Opc || !I->hasOneUse() || 42 // Verify type of the LHS matches so we don't fold cmp's of different 43 // types. 44 I->getOperand(0)->getType() != LHSType || 45 I->getOperand(1)->getType() != RHSType) 46 return nullptr; 47 48 // If they are CmpInst instructions, check their predicates 49 if (CmpInst *CI = dyn_cast<CmpInst>(I)) 50 if (CI->getPredicate() != cast<CmpInst>(FirstInst)->getPredicate()) 51 return nullptr; 52 53 // Keep track of which operand needs a phi node. 54 if (I->getOperand(0) != LHSVal) LHSVal = nullptr; 55 if (I->getOperand(1) != RHSVal) RHSVal = nullptr; 56 } 57 58 // If both LHS and RHS would need a PHI, don't do this transformation, 59 // because it would increase the number of PHIs entering the block, 60 // which leads to higher register pressure. This is especially 61 // bad when the PHIs are in the header of a loop. 62 if (!LHSVal && !RHSVal) 63 return nullptr; 64 65 // Otherwise, this is safe to transform! 66 67 Value *InLHS = FirstInst->getOperand(0); 68 Value *InRHS = FirstInst->getOperand(1); 69 PHINode *NewLHS = nullptr, *NewRHS = nullptr; 70 if (!LHSVal) { 71 NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(), 72 FirstInst->getOperand(0)->getName() + ".pn"); 73 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0)); 74 InsertNewInstBefore(NewLHS, PN); 75 LHSVal = NewLHS; 76 } 77 78 if (!RHSVal) { 79 NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(), 80 FirstInst->getOperand(1)->getName() + ".pn"); 81 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0)); 82 InsertNewInstBefore(NewRHS, PN); 83 RHSVal = NewRHS; 84 } 85 86 // Add all operands to the new PHIs. 87 if (NewLHS || NewRHS) { 88 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 89 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i)); 90 if (NewLHS) { 91 Value *NewInLHS = InInst->getOperand(0); 92 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i)); 93 } 94 if (NewRHS) { 95 Value *NewInRHS = InInst->getOperand(1); 96 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i)); 97 } 98 } 99 } 100 101 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) { 102 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), 103 LHSVal, RHSVal); 104 NewCI->setDebugLoc(FirstInst->getDebugLoc()); 105 return NewCI; 106 } 107 108 BinaryOperator *BinOp = cast<BinaryOperator>(FirstInst); 109 BinaryOperator *NewBinOp = 110 BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal); 111 112 NewBinOp->copyIRFlags(PN.getIncomingValue(0)); 113 114 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) 115 NewBinOp->andIRFlags(PN.getIncomingValue(i)); 116 117 NewBinOp->setDebugLoc(FirstInst->getDebugLoc()); 118 return NewBinOp; 119 } 120 121 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) { 122 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0)); 123 124 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(), 125 FirstInst->op_end()); 126 // This is true if all GEP bases are allocas and if all indices into them are 127 // constants. 128 bool AllBasePointersAreAllocas = true; 129 130 // We don't want to replace this phi if the replacement would require 131 // more than one phi, which leads to higher register pressure. This is 132 // especially bad when the PHIs are in the header of a loop. 133 bool NeededPhi = false; 134 135 bool AllInBounds = true; 136 137 // Scan to see if all operands are the same opcode, and all have one use. 138 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) { 139 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i)); 140 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() || 141 GEP->getNumOperands() != FirstInst->getNumOperands()) 142 return nullptr; 143 144 AllInBounds &= GEP->isInBounds(); 145 146 // Keep track of whether or not all GEPs are of alloca pointers. 147 if (AllBasePointersAreAllocas && 148 (!isa<AllocaInst>(GEP->getOperand(0)) || 149 !GEP->hasAllConstantIndices())) 150 AllBasePointersAreAllocas = false; 151 152 // Compare the operand lists. 153 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) { 154 if (FirstInst->getOperand(op) == GEP->getOperand(op)) 155 continue; 156 157 // Don't merge two GEPs when two operands differ (introducing phi nodes) 158 // if one of the PHIs has a constant for the index. The index may be 159 // substantially cheaper to compute for the constants, so making it a 160 // variable index could pessimize the path. This also handles the case 161 // for struct indices, which must always be constant. 162 if (isa<ConstantInt>(FirstInst->getOperand(op)) || 163 isa<ConstantInt>(GEP->getOperand(op))) 164 return nullptr; 165 166 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType()) 167 return nullptr; 168 169 // If we already needed a PHI for an earlier operand, and another operand 170 // also requires a PHI, we'd be introducing more PHIs than we're 171 // eliminating, which increases register pressure on entry to the PHI's 172 // block. 173 if (NeededPhi) 174 return nullptr; 175 176 FixedOperands[op] = nullptr; // Needs a PHI. 177 NeededPhi = true; 178 } 179 } 180 181 // If all of the base pointers of the PHI'd GEPs are from allocas, don't 182 // bother doing this transformation. At best, this will just save a bit of 183 // offset calculation, but all the predecessors will have to materialize the 184 // stack address into a register anyway. We'd actually rather *clone* the 185 // load up into the predecessors so that we have a load of a gep of an alloca, 186 // which can usually all be folded into the load. 187 if (AllBasePointersAreAllocas) 188 return nullptr; 189 190 // Otherwise, this is safe to transform. Insert PHI nodes for each operand 191 // that is variable. 192 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size()); 193 194 bool HasAnyPHIs = false; 195 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) { 196 if (FixedOperands[i]) continue; // operand doesn't need a phi. 197 Value *FirstOp = FirstInst->getOperand(i); 198 PHINode *NewPN = PHINode::Create(FirstOp->getType(), e, 199 FirstOp->getName()+".pn"); 200 InsertNewInstBefore(NewPN, PN); 201 202 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0)); 203 OperandPhis[i] = NewPN; 204 FixedOperands[i] = NewPN; 205 HasAnyPHIs = true; 206 } 207 208 209 // Add all operands to the new PHIs. 210 if (HasAnyPHIs) { 211 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 212 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i)); 213 BasicBlock *InBB = PN.getIncomingBlock(i); 214 215 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op) 216 if (PHINode *OpPhi = OperandPhis[op]) 217 OpPhi->addIncoming(InGEP->getOperand(op), InBB); 218 } 219 } 220 221 Value *Base = FixedOperands[0]; 222 GetElementPtrInst *NewGEP = 223 GetElementPtrInst::Create(FirstInst->getSourceElementType(), Base, 224 makeArrayRef(FixedOperands).slice(1)); 225 if (AllInBounds) NewGEP->setIsInBounds(); 226 NewGEP->setDebugLoc(FirstInst->getDebugLoc()); 227 return NewGEP; 228 } 229 230 231 /// Return true if we know that it is safe to sink the load out of the block 232 /// that defines it. This means that it must be obvious the value of the load is 233 /// not changed from the point of the load to the end of the block it is in. 234 /// 235 /// Finally, it is safe, but not profitable, to sink a load targeting a 236 /// non-address-taken alloca. Doing so will cause us to not promote the alloca 237 /// to a register. 238 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) { 239 BasicBlock::iterator BBI = L->getIterator(), E = L->getParent()->end(); 240 241 for (++BBI; BBI != E; ++BBI) 242 if (BBI->mayWriteToMemory()) 243 return false; 244 245 // Check for non-address taken alloca. If not address-taken already, it isn't 246 // profitable to do this xform. 247 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) { 248 bool isAddressTaken = false; 249 for (User *U : AI->users()) { 250 if (isa<LoadInst>(U)) continue; 251 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 252 // If storing TO the alloca, then the address isn't taken. 253 if (SI->getOperand(1) == AI) continue; 254 } 255 isAddressTaken = true; 256 break; 257 } 258 259 if (!isAddressTaken && AI->isStaticAlloca()) 260 return false; 261 } 262 263 // If this load is a load from a GEP with a constant offset from an alloca, 264 // then we don't want to sink it. In its present form, it will be 265 // load [constant stack offset]. Sinking it will cause us to have to 266 // materialize the stack addresses in each predecessor in a register only to 267 // do a shared load from register in the successor. 268 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0))) 269 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0))) 270 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices()) 271 return false; 272 273 return true; 274 } 275 276 Instruction *InstCombiner::FoldPHIArgLoadIntoPHI(PHINode &PN) { 277 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0)); 278 279 // FIXME: This is overconservative; this transform is allowed in some cases 280 // for atomic operations. 281 if (FirstLI->isAtomic()) 282 return nullptr; 283 284 // When processing loads, we need to propagate two bits of information to the 285 // sunk load: whether it is volatile, and what its alignment is. We currently 286 // don't sink loads when some have their alignment specified and some don't. 287 // visitLoadInst will propagate an alignment onto the load when TD is around, 288 // and if TD isn't around, we can't handle the mixed case. 289 bool isVolatile = FirstLI->isVolatile(); 290 unsigned LoadAlignment = FirstLI->getAlignment(); 291 unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace(); 292 293 // We can't sink the load if the loaded value could be modified between the 294 // load and the PHI. 295 if (FirstLI->getParent() != PN.getIncomingBlock(0) || 296 !isSafeAndProfitableToSinkLoad(FirstLI)) 297 return nullptr; 298 299 // If the PHI is of volatile loads and the load block has multiple 300 // successors, sinking it would remove a load of the volatile value from 301 // the path through the other successor. 302 if (isVolatile && 303 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1) 304 return nullptr; 305 306 // Check to see if all arguments are the same operation. 307 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 308 LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i)); 309 if (!LI || !LI->hasOneUse()) 310 return nullptr; 311 312 // We can't sink the load if the loaded value could be modified between 313 // the load and the PHI. 314 if (LI->isVolatile() != isVolatile || 315 LI->getParent() != PN.getIncomingBlock(i) || 316 LI->getPointerAddressSpace() != LoadAddrSpace || 317 !isSafeAndProfitableToSinkLoad(LI)) 318 return nullptr; 319 320 // If some of the loads have an alignment specified but not all of them, 321 // we can't do the transformation. 322 if ((LoadAlignment != 0) != (LI->getAlignment() != 0)) 323 return nullptr; 324 325 LoadAlignment = std::min(LoadAlignment, LI->getAlignment()); 326 327 // If the PHI is of volatile loads and the load block has multiple 328 // successors, sinking it would remove a load of the volatile value from 329 // the path through the other successor. 330 if (isVolatile && 331 LI->getParent()->getTerminator()->getNumSuccessors() != 1) 332 return nullptr; 333 } 334 335 // Okay, they are all the same operation. Create a new PHI node of the 336 // correct type, and PHI together all of the LHS's of the instructions. 337 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(), 338 PN.getNumIncomingValues(), 339 PN.getName()+".in"); 340 341 Value *InVal = FirstLI->getOperand(0); 342 NewPN->addIncoming(InVal, PN.getIncomingBlock(0)); 343 LoadInst *NewLI = new LoadInst(NewPN, "", isVolatile, LoadAlignment); 344 345 unsigned KnownIDs[] = { 346 LLVMContext::MD_tbaa, 347 LLVMContext::MD_range, 348 LLVMContext::MD_invariant_load, 349 LLVMContext::MD_alias_scope, 350 LLVMContext::MD_noalias, 351 LLVMContext::MD_nonnull, 352 LLVMContext::MD_align, 353 LLVMContext::MD_dereferenceable, 354 LLVMContext::MD_dereferenceable_or_null, 355 }; 356 357 for (unsigned ID : KnownIDs) 358 NewLI->setMetadata(ID, FirstLI->getMetadata(ID)); 359 360 // Add all operands to the new PHI and combine TBAA metadata. 361 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 362 LoadInst *LI = cast<LoadInst>(PN.getIncomingValue(i)); 363 combineMetadata(NewLI, LI, KnownIDs); 364 Value *NewInVal = LI->getOperand(0); 365 if (NewInVal != InVal) 366 InVal = nullptr; 367 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i)); 368 } 369 370 if (InVal) { 371 // The new PHI unions all of the same values together. This is really 372 // common, so we handle it intelligently here for compile-time speed. 373 NewLI->setOperand(0, InVal); 374 delete NewPN; 375 } else { 376 InsertNewInstBefore(NewPN, PN); 377 } 378 379 // If this was a volatile load that we are merging, make sure to loop through 380 // and mark all the input loads as non-volatile. If we don't do this, we will 381 // insert a new volatile load and the old ones will not be deletable. 382 if (isVolatile) 383 for (Value *IncValue : PN.incoming_values()) 384 cast<LoadInst>(IncValue)->setVolatile(false); 385 386 NewLI->setDebugLoc(FirstLI->getDebugLoc()); 387 return NewLI; 388 } 389 390 /// TODO: This function could handle other cast types, but then it might 391 /// require special-casing a cast from the 'i1' type. See the comment in 392 /// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types. 393 Instruction *InstCombiner::FoldPHIArgZextsIntoPHI(PHINode &Phi) { 394 // We cannot create a new instruction after the PHI if the terminator is an 395 // EHPad because there is no valid insertion point. 396 if (TerminatorInst *TI = Phi.getParent()->getTerminator()) 397 if (TI->isEHPad()) 398 return nullptr; 399 400 // Early exit for the common case of a phi with two operands. These are 401 // handled elsewhere. See the comment below where we check the count of zexts 402 // and constants for more details. 403 unsigned NumIncomingValues = Phi.getNumIncomingValues(); 404 if (NumIncomingValues < 3) 405 return nullptr; 406 407 // Find the narrower type specified by the first zext. 408 Type *NarrowType = nullptr; 409 for (Value *V : Phi.incoming_values()) { 410 if (auto *Zext = dyn_cast<ZExtInst>(V)) { 411 NarrowType = Zext->getSrcTy(); 412 break; 413 } 414 } 415 if (!NarrowType) 416 return nullptr; 417 418 // Walk the phi operands checking that we only have zexts or constants that 419 // we can shrink for free. Store the new operands for the new phi. 420 SmallVector<Value *, 4> NewIncoming; 421 unsigned NumZexts = 0; 422 unsigned NumConsts = 0; 423 for (Value *V : Phi.incoming_values()) { 424 if (auto *Zext = dyn_cast<ZExtInst>(V)) { 425 // All zexts must be identical and have one use. 426 if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUse()) 427 return nullptr; 428 NewIncoming.push_back(Zext->getOperand(0)); 429 NumZexts++; 430 } else if (auto *C = dyn_cast<Constant>(V)) { 431 // Make sure that constants can fit in the new type. 432 Constant *Trunc = ConstantExpr::getTrunc(C, NarrowType); 433 if (ConstantExpr::getZExt(Trunc, C->getType()) != C) 434 return nullptr; 435 NewIncoming.push_back(Trunc); 436 NumConsts++; 437 } else { 438 // If it's not a cast or a constant, bail out. 439 return nullptr; 440 } 441 } 442 443 // The more common cases of a phi with no constant operands or just one 444 // variable operand are handled by FoldPHIArgOpIntoPHI() and FoldOpIntoPhi() 445 // respectively. FoldOpIntoPhi() wants to do the opposite transform that is 446 // performed here. It tries to replicate a cast in the phi operand's basic 447 // block to expose other folding opportunities. Thus, InstCombine will 448 // infinite loop without this check. 449 if (NumConsts == 0 || NumZexts < 2) 450 return nullptr; 451 452 // All incoming values are zexts or constants that are safe to truncate. 453 // Create a new phi node of the narrow type, phi together all of the new 454 // operands, and zext the result back to the original type. 455 PHINode *NewPhi = PHINode::Create(NarrowType, NumIncomingValues, 456 Phi.getName() + ".shrunk"); 457 for (unsigned i = 0; i != NumIncomingValues; ++i) 458 NewPhi->addIncoming(NewIncoming[i], Phi.getIncomingBlock(i)); 459 460 InsertNewInstBefore(NewPhi, Phi); 461 return CastInst::CreateZExtOrBitCast(NewPhi, Phi.getType()); 462 } 463 464 /// If all operands to a PHI node are the same "unary" operator and they all are 465 /// only used by the PHI, PHI together their inputs, and do the operation once, 466 /// to the result of the PHI. 467 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) { 468 // We cannot create a new instruction after the PHI if the terminator is an 469 // EHPad because there is no valid insertion point. 470 if (TerminatorInst *TI = PN.getParent()->getTerminator()) 471 if (TI->isEHPad()) 472 return nullptr; 473 474 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0)); 475 476 if (isa<GetElementPtrInst>(FirstInst)) 477 return FoldPHIArgGEPIntoPHI(PN); 478 if (isa<LoadInst>(FirstInst)) 479 return FoldPHIArgLoadIntoPHI(PN); 480 481 // Scan the instruction, looking for input operations that can be folded away. 482 // If all input operands to the phi are the same instruction (e.g. a cast from 483 // the same type or "+42") we can pull the operation through the PHI, reducing 484 // code size and simplifying code. 485 Constant *ConstantOp = nullptr; 486 Type *CastSrcTy = nullptr; 487 488 if (isa<CastInst>(FirstInst)) { 489 CastSrcTy = FirstInst->getOperand(0)->getType(); 490 491 // Be careful about transforming integer PHIs. We don't want to pessimize 492 // the code by turning an i32 into an i1293. 493 if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) { 494 if (!ShouldChangeType(PN.getType(), CastSrcTy)) 495 return nullptr; 496 } 497 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) { 498 // Can fold binop, compare or shift here if the RHS is a constant, 499 // otherwise call FoldPHIArgBinOpIntoPHI. 500 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1)); 501 if (!ConstantOp) 502 return FoldPHIArgBinOpIntoPHI(PN); 503 } else { 504 return nullptr; // Cannot fold this operation. 505 } 506 507 // Check to see if all arguments are the same operation. 508 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 509 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i)); 510 if (!I || !I->hasOneUse() || !I->isSameOperationAs(FirstInst)) 511 return nullptr; 512 if (CastSrcTy) { 513 if (I->getOperand(0)->getType() != CastSrcTy) 514 return nullptr; // Cast operation must match. 515 } else if (I->getOperand(1) != ConstantOp) { 516 return nullptr; 517 } 518 } 519 520 // Okay, they are all the same operation. Create a new PHI node of the 521 // correct type, and PHI together all of the LHS's of the instructions. 522 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(), 523 PN.getNumIncomingValues(), 524 PN.getName()+".in"); 525 526 Value *InVal = FirstInst->getOperand(0); 527 NewPN->addIncoming(InVal, PN.getIncomingBlock(0)); 528 529 // Add all operands to the new PHI. 530 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 531 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0); 532 if (NewInVal != InVal) 533 InVal = nullptr; 534 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i)); 535 } 536 537 Value *PhiVal; 538 if (InVal) { 539 // The new PHI unions all of the same values together. This is really 540 // common, so we handle it intelligently here for compile-time speed. 541 PhiVal = InVal; 542 delete NewPN; 543 } else { 544 InsertNewInstBefore(NewPN, PN); 545 PhiVal = NewPN; 546 } 547 548 // Insert and return the new operation. 549 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) { 550 CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal, 551 PN.getType()); 552 NewCI->setDebugLoc(FirstInst->getDebugLoc()); 553 return NewCI; 554 } 555 556 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) { 557 BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp); 558 BinOp->copyIRFlags(PN.getIncomingValue(0)); 559 560 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) 561 BinOp->andIRFlags(PN.getIncomingValue(i)); 562 563 BinOp->setDebugLoc(FirstInst->getDebugLoc()); 564 return BinOp; 565 } 566 567 CmpInst *CIOp = cast<CmpInst>(FirstInst); 568 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), 569 PhiVal, ConstantOp); 570 NewCI->setDebugLoc(FirstInst->getDebugLoc()); 571 return NewCI; 572 } 573 574 /// Return true if this PHI node is only used by a PHI node cycle that is dead. 575 static bool DeadPHICycle(PHINode *PN, 576 SmallPtrSetImpl<PHINode*> &PotentiallyDeadPHIs) { 577 if (PN->use_empty()) return true; 578 if (!PN->hasOneUse()) return false; 579 580 // Remember this node, and if we find the cycle, return. 581 if (!PotentiallyDeadPHIs.insert(PN).second) 582 return true; 583 584 // Don't scan crazily complex things. 585 if (PotentiallyDeadPHIs.size() == 16) 586 return false; 587 588 if (PHINode *PU = dyn_cast<PHINode>(PN->user_back())) 589 return DeadPHICycle(PU, PotentiallyDeadPHIs); 590 591 return false; 592 } 593 594 /// Return true if this phi node is always equal to NonPhiInVal. 595 /// This happens with mutually cyclic phi nodes like: 596 /// z = some value; x = phi (y, z); y = phi (x, z) 597 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal, 598 SmallPtrSetImpl<PHINode*> &ValueEqualPHIs) { 599 // See if we already saw this PHI node. 600 if (!ValueEqualPHIs.insert(PN).second) 601 return true; 602 603 // Don't scan crazily complex things. 604 if (ValueEqualPHIs.size() == 16) 605 return false; 606 607 // Scan the operands to see if they are either phi nodes or are equal to 608 // the value. 609 for (Value *Op : PN->incoming_values()) { 610 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) { 611 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs)) 612 return false; 613 } else if (Op != NonPhiInVal) 614 return false; 615 } 616 617 return true; 618 } 619 620 /// Return an existing non-zero constant if this phi node has one, otherwise 621 /// return constant 1. 622 static ConstantInt *GetAnyNonZeroConstInt(PHINode &PN) { 623 assert(isa<IntegerType>(PN.getType()) && "Expect only intger type phi"); 624 for (Value *V : PN.operands()) 625 if (auto *ConstVA = dyn_cast<ConstantInt>(V)) 626 if (!ConstVA->isZeroValue()) 627 return ConstVA; 628 return ConstantInt::get(cast<IntegerType>(PN.getType()), 1); 629 } 630 631 namespace { 632 struct PHIUsageRecord { 633 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on) 634 unsigned Shift; // The amount shifted. 635 Instruction *Inst; // The trunc instruction. 636 637 PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User) 638 : PHIId(pn), Shift(Sh), Inst(User) {} 639 640 bool operator<(const PHIUsageRecord &RHS) const { 641 if (PHIId < RHS.PHIId) return true; 642 if (PHIId > RHS.PHIId) return false; 643 if (Shift < RHS.Shift) return true; 644 if (Shift > RHS.Shift) return false; 645 return Inst->getType()->getPrimitiveSizeInBits() < 646 RHS.Inst->getType()->getPrimitiveSizeInBits(); 647 } 648 }; 649 650 struct LoweredPHIRecord { 651 PHINode *PN; // The PHI that was lowered. 652 unsigned Shift; // The amount shifted. 653 unsigned Width; // The width extracted. 654 655 LoweredPHIRecord(PHINode *pn, unsigned Sh, Type *Ty) 656 : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {} 657 658 // Ctor form used by DenseMap. 659 LoweredPHIRecord(PHINode *pn, unsigned Sh) 660 : PN(pn), Shift(Sh), Width(0) {} 661 }; 662 } 663 664 namespace llvm { 665 template<> 666 struct DenseMapInfo<LoweredPHIRecord> { 667 static inline LoweredPHIRecord getEmptyKey() { 668 return LoweredPHIRecord(nullptr, 0); 669 } 670 static inline LoweredPHIRecord getTombstoneKey() { 671 return LoweredPHIRecord(nullptr, 1); 672 } 673 static unsigned getHashValue(const LoweredPHIRecord &Val) { 674 return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^ 675 (Val.Width>>3); 676 } 677 static bool isEqual(const LoweredPHIRecord &LHS, 678 const LoweredPHIRecord &RHS) { 679 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift && 680 LHS.Width == RHS.Width; 681 } 682 }; 683 } 684 685 686 /// This is an integer PHI and we know that it has an illegal type: see if it is 687 /// only used by trunc or trunc(lshr) operations. If so, we split the PHI into 688 /// the various pieces being extracted. This sort of thing is introduced when 689 /// SROA promotes an aggregate to large integer values. 690 /// 691 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an 692 /// inttoptr. We should produce new PHIs in the right type. 693 /// 694 Instruction *InstCombiner::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) { 695 // PHIUsers - Keep track of all of the truncated values extracted from a set 696 // of PHIs, along with their offset. These are the things we want to rewrite. 697 SmallVector<PHIUsageRecord, 16> PHIUsers; 698 699 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI 700 // nodes which are extracted from. PHIsToSlice is a set we use to avoid 701 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to 702 // check the uses of (to ensure they are all extracts). 703 SmallVector<PHINode*, 8> PHIsToSlice; 704 SmallPtrSet<PHINode*, 8> PHIsInspected; 705 706 PHIsToSlice.push_back(&FirstPhi); 707 PHIsInspected.insert(&FirstPhi); 708 709 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) { 710 PHINode *PN = PHIsToSlice[PHIId]; 711 712 // Scan the input list of the PHI. If any input is an invoke, and if the 713 // input is defined in the predecessor, then we won't be split the critical 714 // edge which is required to insert a truncate. Because of this, we have to 715 // bail out. 716 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 717 InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i)); 718 if (!II) continue; 719 if (II->getParent() != PN->getIncomingBlock(i)) 720 continue; 721 722 // If we have a phi, and if it's directly in the predecessor, then we have 723 // a critical edge where we need to put the truncate. Since we can't 724 // split the edge in instcombine, we have to bail out. 725 return nullptr; 726 } 727 728 for (User *U : PN->users()) { 729 Instruction *UserI = cast<Instruction>(U); 730 731 // If the user is a PHI, inspect its uses recursively. 732 if (PHINode *UserPN = dyn_cast<PHINode>(UserI)) { 733 if (PHIsInspected.insert(UserPN).second) 734 PHIsToSlice.push_back(UserPN); 735 continue; 736 } 737 738 // Truncates are always ok. 739 if (isa<TruncInst>(UserI)) { 740 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, UserI)); 741 continue; 742 } 743 744 // Otherwise it must be a lshr which can only be used by one trunc. 745 if (UserI->getOpcode() != Instruction::LShr || 746 !UserI->hasOneUse() || !isa<TruncInst>(UserI->user_back()) || 747 !isa<ConstantInt>(UserI->getOperand(1))) 748 return nullptr; 749 750 unsigned Shift = cast<ConstantInt>(UserI->getOperand(1))->getZExtValue(); 751 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, UserI->user_back())); 752 } 753 } 754 755 // If we have no users, they must be all self uses, just nuke the PHI. 756 if (PHIUsers.empty()) 757 return replaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType())); 758 759 // If this phi node is transformable, create new PHIs for all the pieces 760 // extracted out of it. First, sort the users by their offset and size. 761 array_pod_sort(PHIUsers.begin(), PHIUsers.end()); 762 763 DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n'; 764 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i) 765 dbgs() << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] << '\n'; 766 ); 767 768 // PredValues - This is a temporary used when rewriting PHI nodes. It is 769 // hoisted out here to avoid construction/destruction thrashing. 770 DenseMap<BasicBlock*, Value*> PredValues; 771 772 // ExtractedVals - Each new PHI we introduce is saved here so we don't 773 // introduce redundant PHIs. 774 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals; 775 776 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) { 777 unsigned PHIId = PHIUsers[UserI].PHIId; 778 PHINode *PN = PHIsToSlice[PHIId]; 779 unsigned Offset = PHIUsers[UserI].Shift; 780 Type *Ty = PHIUsers[UserI].Inst->getType(); 781 782 PHINode *EltPHI; 783 784 // If we've already lowered a user like this, reuse the previously lowered 785 // value. 786 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) { 787 788 // Otherwise, Create the new PHI node for this user. 789 EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(), 790 PN->getName()+".off"+Twine(Offset), PN); 791 assert(EltPHI->getType() != PN->getType() && 792 "Truncate didn't shrink phi?"); 793 794 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 795 BasicBlock *Pred = PN->getIncomingBlock(i); 796 Value *&PredVal = PredValues[Pred]; 797 798 // If we already have a value for this predecessor, reuse it. 799 if (PredVal) { 800 EltPHI->addIncoming(PredVal, Pred); 801 continue; 802 } 803 804 // Handle the PHI self-reuse case. 805 Value *InVal = PN->getIncomingValue(i); 806 if (InVal == PN) { 807 PredVal = EltPHI; 808 EltPHI->addIncoming(PredVal, Pred); 809 continue; 810 } 811 812 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) { 813 // If the incoming value was a PHI, and if it was one of the PHIs we 814 // already rewrote it, just use the lowered value. 815 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) { 816 PredVal = Res; 817 EltPHI->addIncoming(PredVal, Pred); 818 continue; 819 } 820 } 821 822 // Otherwise, do an extract in the predecessor. 823 Builder->SetInsertPoint(Pred->getTerminator()); 824 Value *Res = InVal; 825 if (Offset) 826 Res = Builder->CreateLShr(Res, ConstantInt::get(InVal->getType(), 827 Offset), "extract"); 828 Res = Builder->CreateTrunc(Res, Ty, "extract.t"); 829 PredVal = Res; 830 EltPHI->addIncoming(Res, Pred); 831 832 // If the incoming value was a PHI, and if it was one of the PHIs we are 833 // rewriting, we will ultimately delete the code we inserted. This 834 // means we need to revisit that PHI to make sure we extract out the 835 // needed piece. 836 if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i))) 837 if (PHIsInspected.count(OldInVal)) { 838 unsigned RefPHIId = std::find(PHIsToSlice.begin(),PHIsToSlice.end(), 839 OldInVal)-PHIsToSlice.begin(); 840 PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset, 841 cast<Instruction>(Res))); 842 ++UserE; 843 } 844 } 845 PredValues.clear(); 846 847 DEBUG(dbgs() << " Made element PHI for offset " << Offset << ": " 848 << *EltPHI << '\n'); 849 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI; 850 } 851 852 // Replace the use of this piece with the PHI node. 853 replaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI); 854 } 855 856 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs) 857 // with undefs. 858 Value *Undef = UndefValue::get(FirstPhi.getType()); 859 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i) 860 replaceInstUsesWith(*PHIsToSlice[i], Undef); 861 return replaceInstUsesWith(FirstPhi, Undef); 862 } 863 864 // PHINode simplification 865 // 866 Instruction *InstCombiner::visitPHINode(PHINode &PN) { 867 if (Value *V = SimplifyInstruction(&PN, DL, TLI, DT, AC)) 868 return replaceInstUsesWith(PN, V); 869 870 if (Instruction *Result = FoldPHIArgZextsIntoPHI(PN)) 871 return Result; 872 873 // If all PHI operands are the same operation, pull them through the PHI, 874 // reducing code size. 875 if (isa<Instruction>(PN.getIncomingValue(0)) && 876 isa<Instruction>(PN.getIncomingValue(1)) && 877 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() == 878 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() && 879 // FIXME: The hasOneUse check will fail for PHIs that use the value more 880 // than themselves more than once. 881 PN.getIncomingValue(0)->hasOneUse()) 882 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN)) 883 return Result; 884 885 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if 886 // this PHI only has a single use (a PHI), and if that PHI only has one use (a 887 // PHI)... break the cycle. 888 if (PN.hasOneUse()) { 889 Instruction *PHIUser = cast<Instruction>(PN.user_back()); 890 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) { 891 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs; 892 PotentiallyDeadPHIs.insert(&PN); 893 if (DeadPHICycle(PU, PotentiallyDeadPHIs)) 894 return replaceInstUsesWith(PN, UndefValue::get(PN.getType())); 895 } 896 897 // If this phi has a single use, and if that use just computes a value for 898 // the next iteration of a loop, delete the phi. This occurs with unused 899 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this 900 // common case here is good because the only other things that catch this 901 // are induction variable analysis (sometimes) and ADCE, which is only run 902 // late. 903 if (PHIUser->hasOneUse() && 904 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) && 905 PHIUser->user_back() == &PN) { 906 return replaceInstUsesWith(PN, UndefValue::get(PN.getType())); 907 } 908 // When a PHI is used only to be compared with zero, it is safe to replace 909 // an incoming value proved as known nonzero with any non-zero constant. 910 // For example, in the code below, the incoming value %v can be replaced 911 // with any non-zero constant based on the fact that the PHI is only used to 912 // be compared with zero and %v is a known non-zero value: 913 // %v = select %cond, 1, 2 914 // %p = phi [%v, BB] ... 915 // icmp eq, %p, 0 916 auto *CmpInst = dyn_cast<ICmpInst>(PHIUser); 917 // FIXME: To be simple, handle only integer type for now. 918 if (CmpInst && isa<IntegerType>(PN.getType()) && CmpInst->isEquality() && 919 match(CmpInst->getOperand(1), m_Zero())) { 920 ConstantInt *NonZeroConst = nullptr; 921 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 922 Instruction *CtxI = PN.getIncomingBlock(i)->getTerminator(); 923 Value *VA = PN.getIncomingValue(i); 924 if (isKnownNonZero(VA, DL, 0, AC, CtxI, DT)) { 925 if (!NonZeroConst) 926 NonZeroConst = GetAnyNonZeroConstInt(PN); 927 PN.setIncomingValue(i, NonZeroConst); 928 } 929 } 930 } 931 } 932 933 // We sometimes end up with phi cycles that non-obviously end up being the 934 // same value, for example: 935 // z = some value; x = phi (y, z); y = phi (x, z) 936 // where the phi nodes don't necessarily need to be in the same block. Do a 937 // quick check to see if the PHI node only contains a single non-phi value, if 938 // so, scan to see if the phi cycle is actually equal to that value. 939 { 940 unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues(); 941 // Scan for the first non-phi operand. 942 while (InValNo != NumIncomingVals && 943 isa<PHINode>(PN.getIncomingValue(InValNo))) 944 ++InValNo; 945 946 if (InValNo != NumIncomingVals) { 947 Value *NonPhiInVal = PN.getIncomingValue(InValNo); 948 949 // Scan the rest of the operands to see if there are any conflicts, if so 950 // there is no need to recursively scan other phis. 951 for (++InValNo; InValNo != NumIncomingVals; ++InValNo) { 952 Value *OpVal = PN.getIncomingValue(InValNo); 953 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal)) 954 break; 955 } 956 957 // If we scanned over all operands, then we have one unique value plus 958 // phi values. Scan PHI nodes to see if they all merge in each other or 959 // the value. 960 if (InValNo == NumIncomingVals) { 961 SmallPtrSet<PHINode*, 16> ValueEqualPHIs; 962 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs)) 963 return replaceInstUsesWith(PN, NonPhiInVal); 964 } 965 } 966 } 967 968 // If there are multiple PHIs, sort their operands so that they all list 969 // the blocks in the same order. This will help identical PHIs be eliminated 970 // by other passes. Other passes shouldn't depend on this for correctness 971 // however. 972 PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin()); 973 if (&PN != FirstPN) 974 for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) { 975 BasicBlock *BBA = PN.getIncomingBlock(i); 976 BasicBlock *BBB = FirstPN->getIncomingBlock(i); 977 if (BBA != BBB) { 978 Value *VA = PN.getIncomingValue(i); 979 unsigned j = PN.getBasicBlockIndex(BBB); 980 Value *VB = PN.getIncomingValue(j); 981 PN.setIncomingBlock(i, BBB); 982 PN.setIncomingValue(i, VB); 983 PN.setIncomingBlock(j, BBA); 984 PN.setIncomingValue(j, VA); 985 // NOTE: Instcombine normally would want us to "return &PN" if we 986 // modified any of the operands of an instruction. However, since we 987 // aren't adding or removing uses (just rearranging them) we don't do 988 // this in this case. 989 } 990 } 991 992 // If this is an integer PHI and we know that it has an illegal type, see if 993 // it is only used by trunc or trunc(lshr) operations. If so, we split the 994 // PHI into the various pieces being extracted. This sort of thing is 995 // introduced when SROA promotes an aggregate to a single large integer type. 996 if (PN.getType()->isIntegerTy() && 997 !DL.isLegalInteger(PN.getType()->getPrimitiveSizeInBits())) 998 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN)) 999 return Res; 1000 1001 return nullptr; 1002 } 1003