1 //===- InstCombinePHI.cpp -------------------------------------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file implements the visitPHINode function. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "InstCombineInternal.h" 14 #include "llvm/ADT/STLExtras.h" 15 #include "llvm/ADT/SmallPtrSet.h" 16 #include "llvm/ADT/Statistic.h" 17 #include "llvm/Analysis/InstructionSimplify.h" 18 #include "llvm/Analysis/ValueTracking.h" 19 #include "llvm/IR/PatternMatch.h" 20 #include "llvm/Support/CommandLine.h" 21 #include "llvm/Transforms/InstCombine/InstCombiner.h" 22 #include "llvm/Transforms/Utils/Local.h" 23 24 using namespace llvm; 25 using namespace llvm::PatternMatch; 26 27 #define DEBUG_TYPE "instcombine" 28 29 static cl::opt<unsigned> 30 MaxNumPhis("instcombine-max-num-phis", cl::init(512), 31 cl::desc("Maximum number phis to handle in intptr/ptrint folding")); 32 33 STATISTIC(NumPHIsOfInsertValues, 34 "Number of phi-of-insertvalue turned into insertvalue-of-phis"); 35 STATISTIC(NumPHIsOfExtractValues, 36 "Number of phi-of-extractvalue turned into extractvalue-of-phi"); 37 STATISTIC(NumPHICSEs, "Number of PHI's that got CSE'd"); 38 39 /// The PHI arguments will be folded into a single operation with a PHI node 40 /// as input. The debug location of the single operation will be the merged 41 /// locations of the original PHI node arguments. 42 void InstCombinerImpl::PHIArgMergedDebugLoc(Instruction *Inst, PHINode &PN) { 43 auto *FirstInst = cast<Instruction>(PN.getIncomingValue(0)); 44 Inst->setDebugLoc(FirstInst->getDebugLoc()); 45 // We do not expect a CallInst here, otherwise, N-way merging of DebugLoc 46 // will be inefficient. 47 assert(!isa<CallInst>(Inst)); 48 49 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) { 50 auto *I = cast<Instruction>(PN.getIncomingValue(i)); 51 Inst->applyMergedLocation(Inst->getDebugLoc(), I->getDebugLoc()); 52 } 53 } 54 55 // Replace Integer typed PHI PN if the PHI's value is used as a pointer value. 56 // If there is an existing pointer typed PHI that produces the same value as PN, 57 // replace PN and the IntToPtr operation with it. Otherwise, synthesize a new 58 // PHI node: 59 // 60 // Case-1: 61 // bb1: 62 // int_init = PtrToInt(ptr_init) 63 // br label %bb2 64 // bb2: 65 // int_val = PHI([int_init, %bb1], [int_val_inc, %bb2] 66 // ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2] 67 // ptr_val2 = IntToPtr(int_val) 68 // ... 69 // use(ptr_val2) 70 // ptr_val_inc = ... 71 // inc_val_inc = PtrToInt(ptr_val_inc) 72 // 73 // ==> 74 // bb1: 75 // br label %bb2 76 // bb2: 77 // ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2] 78 // ... 79 // use(ptr_val) 80 // ptr_val_inc = ... 81 // 82 // Case-2: 83 // bb1: 84 // int_ptr = BitCast(ptr_ptr) 85 // int_init = Load(int_ptr) 86 // br label %bb2 87 // bb2: 88 // int_val = PHI([int_init, %bb1], [int_val_inc, %bb2] 89 // ptr_val2 = IntToPtr(int_val) 90 // ... 91 // use(ptr_val2) 92 // ptr_val_inc = ... 93 // inc_val_inc = PtrToInt(ptr_val_inc) 94 // ==> 95 // bb1: 96 // ptr_init = Load(ptr_ptr) 97 // br label %bb2 98 // bb2: 99 // ptr_val = PHI([ptr_init, %bb1], [ptr_val_inc, %bb2] 100 // ... 101 // use(ptr_val) 102 // ptr_val_inc = ... 103 // ... 104 // 105 Instruction *InstCombinerImpl::foldIntegerTypedPHI(PHINode &PN) { 106 if (!PN.getType()->isIntegerTy()) 107 return nullptr; 108 if (!PN.hasOneUse()) 109 return nullptr; 110 111 auto *IntToPtr = dyn_cast<IntToPtrInst>(PN.user_back()); 112 if (!IntToPtr) 113 return nullptr; 114 115 // Check if the pointer is actually used as pointer: 116 auto HasPointerUse = [](Instruction *IIP) { 117 for (User *U : IIP->users()) { 118 Value *Ptr = nullptr; 119 if (LoadInst *LoadI = dyn_cast<LoadInst>(U)) { 120 Ptr = LoadI->getPointerOperand(); 121 } else if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 122 Ptr = SI->getPointerOperand(); 123 } else if (GetElementPtrInst *GI = dyn_cast<GetElementPtrInst>(U)) { 124 Ptr = GI->getPointerOperand(); 125 } 126 127 if (Ptr && Ptr == IIP) 128 return true; 129 } 130 return false; 131 }; 132 133 if (!HasPointerUse(IntToPtr)) 134 return nullptr; 135 136 if (DL.getPointerSizeInBits(IntToPtr->getAddressSpace()) != 137 DL.getTypeSizeInBits(IntToPtr->getOperand(0)->getType())) 138 return nullptr; 139 140 SmallVector<Value *, 4> AvailablePtrVals; 141 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) { 142 Value *Arg = PN.getIncomingValue(i); 143 144 // First look backward: 145 if (auto *PI = dyn_cast<PtrToIntInst>(Arg)) { 146 AvailablePtrVals.emplace_back(PI->getOperand(0)); 147 continue; 148 } 149 150 // Next look forward: 151 Value *ArgIntToPtr = nullptr; 152 for (User *U : Arg->users()) { 153 if (isa<IntToPtrInst>(U) && U->getType() == IntToPtr->getType() && 154 (DT.dominates(cast<Instruction>(U), PN.getIncomingBlock(i)) || 155 cast<Instruction>(U)->getParent() == PN.getIncomingBlock(i))) { 156 ArgIntToPtr = U; 157 break; 158 } 159 } 160 161 if (ArgIntToPtr) { 162 AvailablePtrVals.emplace_back(ArgIntToPtr); 163 continue; 164 } 165 166 // If Arg is defined by a PHI, allow it. This will also create 167 // more opportunities iteratively. 168 if (isa<PHINode>(Arg)) { 169 AvailablePtrVals.emplace_back(Arg); 170 continue; 171 } 172 173 // For a single use integer load: 174 auto *LoadI = dyn_cast<LoadInst>(Arg); 175 if (!LoadI) 176 return nullptr; 177 178 if (!LoadI->hasOneUse()) 179 return nullptr; 180 181 // Push the integer typed Load instruction into the available 182 // value set, and fix it up later when the pointer typed PHI 183 // is synthesized. 184 AvailablePtrVals.emplace_back(LoadI); 185 } 186 187 // Now search for a matching PHI 188 auto *BB = PN.getParent(); 189 assert(AvailablePtrVals.size() == PN.getNumIncomingValues() && 190 "Not enough available ptr typed incoming values"); 191 PHINode *MatchingPtrPHI = nullptr; 192 unsigned NumPhis = 0; 193 for (auto II = BB->begin(); II != BB->end(); II++, NumPhis++) { 194 // FIXME: consider handling this in AggressiveInstCombine 195 PHINode *PtrPHI = dyn_cast<PHINode>(II); 196 if (!PtrPHI) 197 break; 198 if (NumPhis > MaxNumPhis) 199 return nullptr; 200 if (PtrPHI == &PN || PtrPHI->getType() != IntToPtr->getType()) 201 continue; 202 MatchingPtrPHI = PtrPHI; 203 for (unsigned i = 0; i != PtrPHI->getNumIncomingValues(); ++i) { 204 if (AvailablePtrVals[i] != 205 PtrPHI->getIncomingValueForBlock(PN.getIncomingBlock(i))) { 206 MatchingPtrPHI = nullptr; 207 break; 208 } 209 } 210 211 if (MatchingPtrPHI) 212 break; 213 } 214 215 if (MatchingPtrPHI) { 216 assert(MatchingPtrPHI->getType() == IntToPtr->getType() && 217 "Phi's Type does not match with IntToPtr"); 218 // The PtrToCast + IntToPtr will be simplified later 219 return CastInst::CreateBitOrPointerCast(MatchingPtrPHI, 220 IntToPtr->getOperand(0)->getType()); 221 } 222 223 // If it requires a conversion for every PHI operand, do not do it. 224 if (all_of(AvailablePtrVals, [&](Value *V) { 225 return (V->getType() != IntToPtr->getType()) || isa<IntToPtrInst>(V); 226 })) 227 return nullptr; 228 229 // If any of the operand that requires casting is a terminator 230 // instruction, do not do it. Similarly, do not do the transform if the value 231 // is PHI in a block with no insertion point, for example, a catchswitch 232 // block, since we will not be able to insert a cast after the PHI. 233 if (any_of(AvailablePtrVals, [&](Value *V) { 234 if (V->getType() == IntToPtr->getType()) 235 return false; 236 auto *Inst = dyn_cast<Instruction>(V); 237 if (!Inst) 238 return false; 239 if (Inst->isTerminator()) 240 return true; 241 auto *BB = Inst->getParent(); 242 if (isa<PHINode>(Inst) && BB->getFirstInsertionPt() == BB->end()) 243 return true; 244 return false; 245 })) 246 return nullptr; 247 248 PHINode *NewPtrPHI = PHINode::Create( 249 IntToPtr->getType(), PN.getNumIncomingValues(), PN.getName() + ".ptr"); 250 251 InsertNewInstBefore(NewPtrPHI, PN); 252 SmallDenseMap<Value *, Instruction *> Casts; 253 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) { 254 auto *IncomingBB = PN.getIncomingBlock(i); 255 auto *IncomingVal = AvailablePtrVals[i]; 256 257 if (IncomingVal->getType() == IntToPtr->getType()) { 258 NewPtrPHI->addIncoming(IncomingVal, IncomingBB); 259 continue; 260 } 261 262 #ifndef NDEBUG 263 LoadInst *LoadI = dyn_cast<LoadInst>(IncomingVal); 264 assert((isa<PHINode>(IncomingVal) || 265 IncomingVal->getType()->isPointerTy() || 266 (LoadI && LoadI->hasOneUse())) && 267 "Can not replace LoadInst with multiple uses"); 268 #endif 269 // Need to insert a BitCast. 270 // For an integer Load instruction with a single use, the load + IntToPtr 271 // cast will be simplified into a pointer load: 272 // %v = load i64, i64* %a.ip, align 8 273 // %v.cast = inttoptr i64 %v to float ** 274 // ==> 275 // %v.ptrp = bitcast i64 * %a.ip to float ** 276 // %v.cast = load float *, float ** %v.ptrp, align 8 277 Instruction *&CI = Casts[IncomingVal]; 278 if (!CI) { 279 CI = CastInst::CreateBitOrPointerCast(IncomingVal, IntToPtr->getType(), 280 IncomingVal->getName() + ".ptr"); 281 if (auto *IncomingI = dyn_cast<Instruction>(IncomingVal)) { 282 BasicBlock::iterator InsertPos(IncomingI); 283 InsertPos++; 284 BasicBlock *BB = IncomingI->getParent(); 285 if (isa<PHINode>(IncomingI)) 286 InsertPos = BB->getFirstInsertionPt(); 287 assert(InsertPos != BB->end() && "should have checked above"); 288 InsertNewInstBefore(CI, *InsertPos); 289 } else { 290 auto *InsertBB = &IncomingBB->getParent()->getEntryBlock(); 291 InsertNewInstBefore(CI, *InsertBB->getFirstInsertionPt()); 292 } 293 } 294 NewPtrPHI->addIncoming(CI, IncomingBB); 295 } 296 297 // The PtrToCast + IntToPtr will be simplified later 298 return CastInst::CreateBitOrPointerCast(NewPtrPHI, 299 IntToPtr->getOperand(0)->getType()); 300 } 301 302 /// If we have something like phi [insertvalue(a,b,0), insertvalue(c,d,0)], 303 /// turn this into a phi[a,c] and phi[b,d] and a single insertvalue. 304 Instruction * 305 InstCombinerImpl::foldPHIArgInsertValueInstructionIntoPHI(PHINode &PN) { 306 auto *FirstIVI = cast<InsertValueInst>(PN.getIncomingValue(0)); 307 308 // Scan to see if all operands are `insertvalue`'s with the same indicies, 309 // and all have a single use. 310 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) { 311 auto *I = dyn_cast<InsertValueInst>(PN.getIncomingValue(i)); 312 if (!I || !I->hasOneUser() || I->getIndices() != FirstIVI->getIndices()) 313 return nullptr; 314 } 315 316 // For each operand of an `insertvalue` 317 std::array<PHINode *, 2> NewOperands; 318 for (int OpIdx : {0, 1}) { 319 auto *&NewOperand = NewOperands[OpIdx]; 320 // Create a new PHI node to receive the values the operand has in each 321 // incoming basic block. 322 NewOperand = PHINode::Create( 323 FirstIVI->getOperand(OpIdx)->getType(), PN.getNumIncomingValues(), 324 FirstIVI->getOperand(OpIdx)->getName() + ".pn"); 325 // And populate each operand's PHI with said values. 326 for (auto Incoming : zip(PN.blocks(), PN.incoming_values())) 327 NewOperand->addIncoming( 328 cast<InsertValueInst>(std::get<1>(Incoming))->getOperand(OpIdx), 329 std::get<0>(Incoming)); 330 InsertNewInstBefore(NewOperand, PN); 331 } 332 333 // And finally, create `insertvalue` over the newly-formed PHI nodes. 334 auto *NewIVI = InsertValueInst::Create(NewOperands[0], NewOperands[1], 335 FirstIVI->getIndices(), PN.getName()); 336 337 PHIArgMergedDebugLoc(NewIVI, PN); 338 ++NumPHIsOfInsertValues; 339 return NewIVI; 340 } 341 342 /// If we have something like phi [extractvalue(a,0), extractvalue(b,0)], 343 /// turn this into a phi[a,b] and a single extractvalue. 344 Instruction * 345 InstCombinerImpl::foldPHIArgExtractValueInstructionIntoPHI(PHINode &PN) { 346 auto *FirstEVI = cast<ExtractValueInst>(PN.getIncomingValue(0)); 347 348 // Scan to see if all operands are `extractvalue`'s with the same indicies, 349 // and all have a single use. 350 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) { 351 auto *I = dyn_cast<ExtractValueInst>(PN.getIncomingValue(i)); 352 if (!I || !I->hasOneUser() || I->getIndices() != FirstEVI->getIndices() || 353 I->getAggregateOperand()->getType() != 354 FirstEVI->getAggregateOperand()->getType()) 355 return nullptr; 356 } 357 358 // Create a new PHI node to receive the values the aggregate operand has 359 // in each incoming basic block. 360 auto *NewAggregateOperand = PHINode::Create( 361 FirstEVI->getAggregateOperand()->getType(), PN.getNumIncomingValues(), 362 FirstEVI->getAggregateOperand()->getName() + ".pn"); 363 // And populate the PHI with said values. 364 for (auto Incoming : zip(PN.blocks(), PN.incoming_values())) 365 NewAggregateOperand->addIncoming( 366 cast<ExtractValueInst>(std::get<1>(Incoming))->getAggregateOperand(), 367 std::get<0>(Incoming)); 368 InsertNewInstBefore(NewAggregateOperand, PN); 369 370 // And finally, create `extractvalue` over the newly-formed PHI nodes. 371 auto *NewEVI = ExtractValueInst::Create(NewAggregateOperand, 372 FirstEVI->getIndices(), PN.getName()); 373 374 PHIArgMergedDebugLoc(NewEVI, PN); 375 ++NumPHIsOfExtractValues; 376 return NewEVI; 377 } 378 379 /// If we have something like phi [add (a,b), add(a,c)] and if a/b/c and the 380 /// adds all have a single user, turn this into a phi and a single binop. 381 Instruction *InstCombinerImpl::foldPHIArgBinOpIntoPHI(PHINode &PN) { 382 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0)); 383 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)); 384 unsigned Opc = FirstInst->getOpcode(); 385 Value *LHSVal = FirstInst->getOperand(0); 386 Value *RHSVal = FirstInst->getOperand(1); 387 388 Type *LHSType = LHSVal->getType(); 389 Type *RHSType = RHSVal->getType(); 390 391 // Scan to see if all operands are the same opcode, and all have one user. 392 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) { 393 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i)); 394 if (!I || I->getOpcode() != Opc || !I->hasOneUser() || 395 // Verify type of the LHS matches so we don't fold cmp's of different 396 // types. 397 I->getOperand(0)->getType() != LHSType || 398 I->getOperand(1)->getType() != RHSType) 399 return nullptr; 400 401 // If they are CmpInst instructions, check their predicates 402 if (CmpInst *CI = dyn_cast<CmpInst>(I)) 403 if (CI->getPredicate() != cast<CmpInst>(FirstInst)->getPredicate()) 404 return nullptr; 405 406 // Keep track of which operand needs a phi node. 407 if (I->getOperand(0) != LHSVal) LHSVal = nullptr; 408 if (I->getOperand(1) != RHSVal) RHSVal = nullptr; 409 } 410 411 // If both LHS and RHS would need a PHI, don't do this transformation, 412 // because it would increase the number of PHIs entering the block, 413 // which leads to higher register pressure. This is especially 414 // bad when the PHIs are in the header of a loop. 415 if (!LHSVal && !RHSVal) 416 return nullptr; 417 418 // Otherwise, this is safe to transform! 419 420 Value *InLHS = FirstInst->getOperand(0); 421 Value *InRHS = FirstInst->getOperand(1); 422 PHINode *NewLHS = nullptr, *NewRHS = nullptr; 423 if (!LHSVal) { 424 NewLHS = PHINode::Create(LHSType, PN.getNumIncomingValues(), 425 FirstInst->getOperand(0)->getName() + ".pn"); 426 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0)); 427 InsertNewInstBefore(NewLHS, PN); 428 LHSVal = NewLHS; 429 } 430 431 if (!RHSVal) { 432 NewRHS = PHINode::Create(RHSType, PN.getNumIncomingValues(), 433 FirstInst->getOperand(1)->getName() + ".pn"); 434 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0)); 435 InsertNewInstBefore(NewRHS, PN); 436 RHSVal = NewRHS; 437 } 438 439 // Add all operands to the new PHIs. 440 if (NewLHS || NewRHS) { 441 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 442 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i)); 443 if (NewLHS) { 444 Value *NewInLHS = InInst->getOperand(0); 445 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i)); 446 } 447 if (NewRHS) { 448 Value *NewInRHS = InInst->getOperand(1); 449 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i)); 450 } 451 } 452 } 453 454 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst)) { 455 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), 456 LHSVal, RHSVal); 457 PHIArgMergedDebugLoc(NewCI, PN); 458 return NewCI; 459 } 460 461 BinaryOperator *BinOp = cast<BinaryOperator>(FirstInst); 462 BinaryOperator *NewBinOp = 463 BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal); 464 465 NewBinOp->copyIRFlags(PN.getIncomingValue(0)); 466 467 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) 468 NewBinOp->andIRFlags(PN.getIncomingValue(i)); 469 470 PHIArgMergedDebugLoc(NewBinOp, PN); 471 return NewBinOp; 472 } 473 474 Instruction *InstCombinerImpl::foldPHIArgGEPIntoPHI(PHINode &PN) { 475 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0)); 476 477 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(), 478 FirstInst->op_end()); 479 // This is true if all GEP bases are allocas and if all indices into them are 480 // constants. 481 bool AllBasePointersAreAllocas = true; 482 483 // We don't want to replace this phi if the replacement would require 484 // more than one phi, which leads to higher register pressure. This is 485 // especially bad when the PHIs are in the header of a loop. 486 bool NeededPhi = false; 487 488 bool AllInBounds = true; 489 490 // Scan to see if all operands are the same opcode, and all have one user. 491 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) { 492 GetElementPtrInst *GEP = 493 dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i)); 494 if (!GEP || !GEP->hasOneUser() || GEP->getType() != FirstInst->getType() || 495 GEP->getNumOperands() != FirstInst->getNumOperands()) 496 return nullptr; 497 498 AllInBounds &= GEP->isInBounds(); 499 500 // Keep track of whether or not all GEPs are of alloca pointers. 501 if (AllBasePointersAreAllocas && 502 (!isa<AllocaInst>(GEP->getOperand(0)) || 503 !GEP->hasAllConstantIndices())) 504 AllBasePointersAreAllocas = false; 505 506 // Compare the operand lists. 507 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) { 508 if (FirstInst->getOperand(op) == GEP->getOperand(op)) 509 continue; 510 511 // Don't merge two GEPs when two operands differ (introducing phi nodes) 512 // if one of the PHIs has a constant for the index. The index may be 513 // substantially cheaper to compute for the constants, so making it a 514 // variable index could pessimize the path. This also handles the case 515 // for struct indices, which must always be constant. 516 if (isa<ConstantInt>(FirstInst->getOperand(op)) || 517 isa<ConstantInt>(GEP->getOperand(op))) 518 return nullptr; 519 520 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType()) 521 return nullptr; 522 523 // If we already needed a PHI for an earlier operand, and another operand 524 // also requires a PHI, we'd be introducing more PHIs than we're 525 // eliminating, which increases register pressure on entry to the PHI's 526 // block. 527 if (NeededPhi) 528 return nullptr; 529 530 FixedOperands[op] = nullptr; // Needs a PHI. 531 NeededPhi = true; 532 } 533 } 534 535 // If all of the base pointers of the PHI'd GEPs are from allocas, don't 536 // bother doing this transformation. At best, this will just save a bit of 537 // offset calculation, but all the predecessors will have to materialize the 538 // stack address into a register anyway. We'd actually rather *clone* the 539 // load up into the predecessors so that we have a load of a gep of an alloca, 540 // which can usually all be folded into the load. 541 if (AllBasePointersAreAllocas) 542 return nullptr; 543 544 // Otherwise, this is safe to transform. Insert PHI nodes for each operand 545 // that is variable. 546 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size()); 547 548 bool HasAnyPHIs = false; 549 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) { 550 if (FixedOperands[i]) continue; // operand doesn't need a phi. 551 Value *FirstOp = FirstInst->getOperand(i); 552 PHINode *NewPN = PHINode::Create(FirstOp->getType(), e, 553 FirstOp->getName()+".pn"); 554 InsertNewInstBefore(NewPN, PN); 555 556 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0)); 557 OperandPhis[i] = NewPN; 558 FixedOperands[i] = NewPN; 559 HasAnyPHIs = true; 560 } 561 562 563 // Add all operands to the new PHIs. 564 if (HasAnyPHIs) { 565 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 566 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i)); 567 BasicBlock *InBB = PN.getIncomingBlock(i); 568 569 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op) 570 if (PHINode *OpPhi = OperandPhis[op]) 571 OpPhi->addIncoming(InGEP->getOperand(op), InBB); 572 } 573 } 574 575 Value *Base = FixedOperands[0]; 576 GetElementPtrInst *NewGEP = 577 GetElementPtrInst::Create(FirstInst->getSourceElementType(), Base, 578 makeArrayRef(FixedOperands).slice(1)); 579 if (AllInBounds) NewGEP->setIsInBounds(); 580 PHIArgMergedDebugLoc(NewGEP, PN); 581 return NewGEP; 582 } 583 584 /// Return true if we know that it is safe to sink the load out of the block 585 /// that defines it. This means that it must be obvious the value of the load is 586 /// not changed from the point of the load to the end of the block it is in. 587 /// 588 /// Finally, it is safe, but not profitable, to sink a load targeting a 589 /// non-address-taken alloca. Doing so will cause us to not promote the alloca 590 /// to a register. 591 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) { 592 BasicBlock::iterator BBI = L->getIterator(), E = L->getParent()->end(); 593 594 for (++BBI; BBI != E; ++BBI) 595 if (BBI->mayWriteToMemory()) { 596 // Calls that only access inaccessible memory do not block sinking the 597 // load. 598 if (auto *CB = dyn_cast<CallBase>(BBI)) 599 if (CB->onlyAccessesInaccessibleMemory()) 600 continue; 601 return false; 602 } 603 604 // Check for non-address taken alloca. If not address-taken already, it isn't 605 // profitable to do this xform. 606 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) { 607 bool isAddressTaken = false; 608 for (User *U : AI->users()) { 609 if (isa<LoadInst>(U)) continue; 610 if (StoreInst *SI = dyn_cast<StoreInst>(U)) { 611 // If storing TO the alloca, then the address isn't taken. 612 if (SI->getOperand(1) == AI) continue; 613 } 614 isAddressTaken = true; 615 break; 616 } 617 618 if (!isAddressTaken && AI->isStaticAlloca()) 619 return false; 620 } 621 622 // If this load is a load from a GEP with a constant offset from an alloca, 623 // then we don't want to sink it. In its present form, it will be 624 // load [constant stack offset]. Sinking it will cause us to have to 625 // materialize the stack addresses in each predecessor in a register only to 626 // do a shared load from register in the successor. 627 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0))) 628 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0))) 629 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices()) 630 return false; 631 632 return true; 633 } 634 635 Instruction *InstCombinerImpl::foldPHIArgLoadIntoPHI(PHINode &PN) { 636 LoadInst *FirstLI = cast<LoadInst>(PN.getIncomingValue(0)); 637 638 // FIXME: This is overconservative; this transform is allowed in some cases 639 // for atomic operations. 640 if (FirstLI->isAtomic()) 641 return nullptr; 642 643 // When processing loads, we need to propagate two bits of information to the 644 // sunk load: whether it is volatile, and what its alignment is. We currently 645 // don't sink loads when some have their alignment specified and some don't. 646 // visitLoadInst will propagate an alignment onto the load when TD is around, 647 // and if TD isn't around, we can't handle the mixed case. 648 bool isVolatile = FirstLI->isVolatile(); 649 Align LoadAlignment = FirstLI->getAlign(); 650 unsigned LoadAddrSpace = FirstLI->getPointerAddressSpace(); 651 652 // We can't sink the load if the loaded value could be modified between the 653 // load and the PHI. 654 if (FirstLI->getParent() != PN.getIncomingBlock(0) || 655 !isSafeAndProfitableToSinkLoad(FirstLI)) 656 return nullptr; 657 658 // If the PHI is of volatile loads and the load block has multiple 659 // successors, sinking it would remove a load of the volatile value from 660 // the path through the other successor. 661 if (isVolatile && 662 FirstLI->getParent()->getTerminator()->getNumSuccessors() != 1) 663 return nullptr; 664 665 // Check to see if all arguments are the same operation. 666 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 667 LoadInst *LI = dyn_cast<LoadInst>(PN.getIncomingValue(i)); 668 if (!LI || !LI->hasOneUser()) 669 return nullptr; 670 671 // We can't sink the load if the loaded value could be modified between 672 // the load and the PHI. 673 if (LI->isVolatile() != isVolatile || 674 LI->getParent() != PN.getIncomingBlock(i) || 675 LI->getPointerAddressSpace() != LoadAddrSpace || 676 !isSafeAndProfitableToSinkLoad(LI)) 677 return nullptr; 678 679 LoadAlignment = std::min(LoadAlignment, Align(LI->getAlign())); 680 681 // If the PHI is of volatile loads and the load block has multiple 682 // successors, sinking it would remove a load of the volatile value from 683 // the path through the other successor. 684 if (isVolatile && 685 LI->getParent()->getTerminator()->getNumSuccessors() != 1) 686 return nullptr; 687 } 688 689 // Okay, they are all the same operation. Create a new PHI node of the 690 // correct type, and PHI together all of the LHS's of the instructions. 691 PHINode *NewPN = PHINode::Create(FirstLI->getOperand(0)->getType(), 692 PN.getNumIncomingValues(), 693 PN.getName()+".in"); 694 695 Value *InVal = FirstLI->getOperand(0); 696 NewPN->addIncoming(InVal, PN.getIncomingBlock(0)); 697 LoadInst *NewLI = 698 new LoadInst(FirstLI->getType(), NewPN, "", isVolatile, LoadAlignment); 699 700 unsigned KnownIDs[] = { 701 LLVMContext::MD_tbaa, 702 LLVMContext::MD_range, 703 LLVMContext::MD_invariant_load, 704 LLVMContext::MD_alias_scope, 705 LLVMContext::MD_noalias, 706 LLVMContext::MD_nonnull, 707 LLVMContext::MD_align, 708 LLVMContext::MD_dereferenceable, 709 LLVMContext::MD_dereferenceable_or_null, 710 LLVMContext::MD_access_group, 711 }; 712 713 for (unsigned ID : KnownIDs) 714 NewLI->setMetadata(ID, FirstLI->getMetadata(ID)); 715 716 // Add all operands to the new PHI and combine TBAA metadata. 717 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 718 LoadInst *LI = cast<LoadInst>(PN.getIncomingValue(i)); 719 combineMetadata(NewLI, LI, KnownIDs, true); 720 Value *NewInVal = LI->getOperand(0); 721 if (NewInVal != InVal) 722 InVal = nullptr; 723 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i)); 724 } 725 726 if (InVal) { 727 // The new PHI unions all of the same values together. This is really 728 // common, so we handle it intelligently here for compile-time speed. 729 NewLI->setOperand(0, InVal); 730 delete NewPN; 731 } else { 732 InsertNewInstBefore(NewPN, PN); 733 } 734 735 // If this was a volatile load that we are merging, make sure to loop through 736 // and mark all the input loads as non-volatile. If we don't do this, we will 737 // insert a new volatile load and the old ones will not be deletable. 738 if (isVolatile) 739 for (Value *IncValue : PN.incoming_values()) 740 cast<LoadInst>(IncValue)->setVolatile(false); 741 742 PHIArgMergedDebugLoc(NewLI, PN); 743 return NewLI; 744 } 745 746 /// TODO: This function could handle other cast types, but then it might 747 /// require special-casing a cast from the 'i1' type. See the comment in 748 /// FoldPHIArgOpIntoPHI() about pessimizing illegal integer types. 749 Instruction *InstCombinerImpl::foldPHIArgZextsIntoPHI(PHINode &Phi) { 750 // We cannot create a new instruction after the PHI if the terminator is an 751 // EHPad because there is no valid insertion point. 752 if (Instruction *TI = Phi.getParent()->getTerminator()) 753 if (TI->isEHPad()) 754 return nullptr; 755 756 // Early exit for the common case of a phi with two operands. These are 757 // handled elsewhere. See the comment below where we check the count of zexts 758 // and constants for more details. 759 unsigned NumIncomingValues = Phi.getNumIncomingValues(); 760 if (NumIncomingValues < 3) 761 return nullptr; 762 763 // Find the narrower type specified by the first zext. 764 Type *NarrowType = nullptr; 765 for (Value *V : Phi.incoming_values()) { 766 if (auto *Zext = dyn_cast<ZExtInst>(V)) { 767 NarrowType = Zext->getSrcTy(); 768 break; 769 } 770 } 771 if (!NarrowType) 772 return nullptr; 773 774 // Walk the phi operands checking that we only have zexts or constants that 775 // we can shrink for free. Store the new operands for the new phi. 776 SmallVector<Value *, 4> NewIncoming; 777 unsigned NumZexts = 0; 778 unsigned NumConsts = 0; 779 for (Value *V : Phi.incoming_values()) { 780 if (auto *Zext = dyn_cast<ZExtInst>(V)) { 781 // All zexts must be identical and have one user. 782 if (Zext->getSrcTy() != NarrowType || !Zext->hasOneUser()) 783 return nullptr; 784 NewIncoming.push_back(Zext->getOperand(0)); 785 NumZexts++; 786 } else if (auto *C = dyn_cast<Constant>(V)) { 787 // Make sure that constants can fit in the new type. 788 Constant *Trunc = ConstantExpr::getTrunc(C, NarrowType); 789 if (ConstantExpr::getZExt(Trunc, C->getType()) != C) 790 return nullptr; 791 NewIncoming.push_back(Trunc); 792 NumConsts++; 793 } else { 794 // If it's not a cast or a constant, bail out. 795 return nullptr; 796 } 797 } 798 799 // The more common cases of a phi with no constant operands or just one 800 // variable operand are handled by FoldPHIArgOpIntoPHI() and foldOpIntoPhi() 801 // respectively. foldOpIntoPhi() wants to do the opposite transform that is 802 // performed here. It tries to replicate a cast in the phi operand's basic 803 // block to expose other folding opportunities. Thus, InstCombine will 804 // infinite loop without this check. 805 if (NumConsts == 0 || NumZexts < 2) 806 return nullptr; 807 808 // All incoming values are zexts or constants that are safe to truncate. 809 // Create a new phi node of the narrow type, phi together all of the new 810 // operands, and zext the result back to the original type. 811 PHINode *NewPhi = PHINode::Create(NarrowType, NumIncomingValues, 812 Phi.getName() + ".shrunk"); 813 for (unsigned i = 0; i != NumIncomingValues; ++i) 814 NewPhi->addIncoming(NewIncoming[i], Phi.getIncomingBlock(i)); 815 816 InsertNewInstBefore(NewPhi, Phi); 817 return CastInst::CreateZExtOrBitCast(NewPhi, Phi.getType()); 818 } 819 820 /// If all operands to a PHI node are the same "unary" operator and they all are 821 /// only used by the PHI, PHI together their inputs, and do the operation once, 822 /// to the result of the PHI. 823 Instruction *InstCombinerImpl::foldPHIArgOpIntoPHI(PHINode &PN) { 824 // We cannot create a new instruction after the PHI if the terminator is an 825 // EHPad because there is no valid insertion point. 826 if (Instruction *TI = PN.getParent()->getTerminator()) 827 if (TI->isEHPad()) 828 return nullptr; 829 830 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0)); 831 832 if (isa<GetElementPtrInst>(FirstInst)) 833 return foldPHIArgGEPIntoPHI(PN); 834 if (isa<LoadInst>(FirstInst)) 835 return foldPHIArgLoadIntoPHI(PN); 836 if (isa<InsertValueInst>(FirstInst)) 837 return foldPHIArgInsertValueInstructionIntoPHI(PN); 838 if (isa<ExtractValueInst>(FirstInst)) 839 return foldPHIArgExtractValueInstructionIntoPHI(PN); 840 841 // Scan the instruction, looking for input operations that can be folded away. 842 // If all input operands to the phi are the same instruction (e.g. a cast from 843 // the same type or "+42") we can pull the operation through the PHI, reducing 844 // code size and simplifying code. 845 Constant *ConstantOp = nullptr; 846 Type *CastSrcTy = nullptr; 847 848 if (isa<CastInst>(FirstInst)) { 849 CastSrcTy = FirstInst->getOperand(0)->getType(); 850 851 // Be careful about transforming integer PHIs. We don't want to pessimize 852 // the code by turning an i32 into an i1293. 853 if (PN.getType()->isIntegerTy() && CastSrcTy->isIntegerTy()) { 854 if (!shouldChangeType(PN.getType(), CastSrcTy)) 855 return nullptr; 856 } 857 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) { 858 // Can fold binop, compare or shift here if the RHS is a constant, 859 // otherwise call FoldPHIArgBinOpIntoPHI. 860 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1)); 861 if (!ConstantOp) 862 return foldPHIArgBinOpIntoPHI(PN); 863 } else { 864 return nullptr; // Cannot fold this operation. 865 } 866 867 // Check to see if all arguments are the same operation. 868 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 869 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i)); 870 if (!I || !I->hasOneUser() || !I->isSameOperationAs(FirstInst)) 871 return nullptr; 872 if (CastSrcTy) { 873 if (I->getOperand(0)->getType() != CastSrcTy) 874 return nullptr; // Cast operation must match. 875 } else if (I->getOperand(1) != ConstantOp) { 876 return nullptr; 877 } 878 } 879 880 // Okay, they are all the same operation. Create a new PHI node of the 881 // correct type, and PHI together all of the LHS's of the instructions. 882 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(), 883 PN.getNumIncomingValues(), 884 PN.getName()+".in"); 885 886 Value *InVal = FirstInst->getOperand(0); 887 NewPN->addIncoming(InVal, PN.getIncomingBlock(0)); 888 889 // Add all operands to the new PHI. 890 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) { 891 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0); 892 if (NewInVal != InVal) 893 InVal = nullptr; 894 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i)); 895 } 896 897 Value *PhiVal; 898 if (InVal) { 899 // The new PHI unions all of the same values together. This is really 900 // common, so we handle it intelligently here for compile-time speed. 901 PhiVal = InVal; 902 delete NewPN; 903 } else { 904 InsertNewInstBefore(NewPN, PN); 905 PhiVal = NewPN; 906 } 907 908 // Insert and return the new operation. 909 if (CastInst *FirstCI = dyn_cast<CastInst>(FirstInst)) { 910 CastInst *NewCI = CastInst::Create(FirstCI->getOpcode(), PhiVal, 911 PN.getType()); 912 PHIArgMergedDebugLoc(NewCI, PN); 913 return NewCI; 914 } 915 916 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst)) { 917 BinOp = BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp); 918 BinOp->copyIRFlags(PN.getIncomingValue(0)); 919 920 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) 921 BinOp->andIRFlags(PN.getIncomingValue(i)); 922 923 PHIArgMergedDebugLoc(BinOp, PN); 924 return BinOp; 925 } 926 927 CmpInst *CIOp = cast<CmpInst>(FirstInst); 928 CmpInst *NewCI = CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(), 929 PhiVal, ConstantOp); 930 PHIArgMergedDebugLoc(NewCI, PN); 931 return NewCI; 932 } 933 934 /// Return true if this PHI node is only used by a PHI node cycle that is dead. 935 static bool DeadPHICycle(PHINode *PN, 936 SmallPtrSetImpl<PHINode*> &PotentiallyDeadPHIs) { 937 if (PN->use_empty()) return true; 938 if (!PN->hasOneUse()) return false; 939 940 // Remember this node, and if we find the cycle, return. 941 if (!PotentiallyDeadPHIs.insert(PN).second) 942 return true; 943 944 // Don't scan crazily complex things. 945 if (PotentiallyDeadPHIs.size() == 16) 946 return false; 947 948 if (PHINode *PU = dyn_cast<PHINode>(PN->user_back())) 949 return DeadPHICycle(PU, PotentiallyDeadPHIs); 950 951 return false; 952 } 953 954 /// Return true if this phi node is always equal to NonPhiInVal. 955 /// This happens with mutually cyclic phi nodes like: 956 /// z = some value; x = phi (y, z); y = phi (x, z) 957 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal, 958 SmallPtrSetImpl<PHINode*> &ValueEqualPHIs) { 959 // See if we already saw this PHI node. 960 if (!ValueEqualPHIs.insert(PN).second) 961 return true; 962 963 // Don't scan crazily complex things. 964 if (ValueEqualPHIs.size() == 16) 965 return false; 966 967 // Scan the operands to see if they are either phi nodes or are equal to 968 // the value. 969 for (Value *Op : PN->incoming_values()) { 970 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) { 971 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs)) 972 return false; 973 } else if (Op != NonPhiInVal) 974 return false; 975 } 976 977 return true; 978 } 979 980 /// Return an existing non-zero constant if this phi node has one, otherwise 981 /// return constant 1. 982 static ConstantInt *GetAnyNonZeroConstInt(PHINode &PN) { 983 assert(isa<IntegerType>(PN.getType()) && "Expect only integer type phi"); 984 for (Value *V : PN.operands()) 985 if (auto *ConstVA = dyn_cast<ConstantInt>(V)) 986 if (!ConstVA->isZero()) 987 return ConstVA; 988 return ConstantInt::get(cast<IntegerType>(PN.getType()), 1); 989 } 990 991 namespace { 992 struct PHIUsageRecord { 993 unsigned PHIId; // The ID # of the PHI (something determinstic to sort on) 994 unsigned Shift; // The amount shifted. 995 Instruction *Inst; // The trunc instruction. 996 997 PHIUsageRecord(unsigned pn, unsigned Sh, Instruction *User) 998 : PHIId(pn), Shift(Sh), Inst(User) {} 999 1000 bool operator<(const PHIUsageRecord &RHS) const { 1001 if (PHIId < RHS.PHIId) return true; 1002 if (PHIId > RHS.PHIId) return false; 1003 if (Shift < RHS.Shift) return true; 1004 if (Shift > RHS.Shift) return false; 1005 return Inst->getType()->getPrimitiveSizeInBits() < 1006 RHS.Inst->getType()->getPrimitiveSizeInBits(); 1007 } 1008 }; 1009 1010 struct LoweredPHIRecord { 1011 PHINode *PN; // The PHI that was lowered. 1012 unsigned Shift; // The amount shifted. 1013 unsigned Width; // The width extracted. 1014 1015 LoweredPHIRecord(PHINode *pn, unsigned Sh, Type *Ty) 1016 : PN(pn), Shift(Sh), Width(Ty->getPrimitiveSizeInBits()) {} 1017 1018 // Ctor form used by DenseMap. 1019 LoweredPHIRecord(PHINode *pn, unsigned Sh) 1020 : PN(pn), Shift(Sh), Width(0) {} 1021 }; 1022 } // namespace 1023 1024 namespace llvm { 1025 template<> 1026 struct DenseMapInfo<LoweredPHIRecord> { 1027 static inline LoweredPHIRecord getEmptyKey() { 1028 return LoweredPHIRecord(nullptr, 0); 1029 } 1030 static inline LoweredPHIRecord getTombstoneKey() { 1031 return LoweredPHIRecord(nullptr, 1); 1032 } 1033 static unsigned getHashValue(const LoweredPHIRecord &Val) { 1034 return DenseMapInfo<PHINode*>::getHashValue(Val.PN) ^ (Val.Shift>>3) ^ 1035 (Val.Width>>3); 1036 } 1037 static bool isEqual(const LoweredPHIRecord &LHS, 1038 const LoweredPHIRecord &RHS) { 1039 return LHS.PN == RHS.PN && LHS.Shift == RHS.Shift && 1040 LHS.Width == RHS.Width; 1041 } 1042 }; 1043 } // namespace llvm 1044 1045 1046 /// This is an integer PHI and we know that it has an illegal type: see if it is 1047 /// only used by trunc or trunc(lshr) operations. If so, we split the PHI into 1048 /// the various pieces being extracted. This sort of thing is introduced when 1049 /// SROA promotes an aggregate to large integer values. 1050 /// 1051 /// TODO: The user of the trunc may be an bitcast to float/double/vector or an 1052 /// inttoptr. We should produce new PHIs in the right type. 1053 /// 1054 Instruction *InstCombinerImpl::SliceUpIllegalIntegerPHI(PHINode &FirstPhi) { 1055 // PHIUsers - Keep track of all of the truncated values extracted from a set 1056 // of PHIs, along with their offset. These are the things we want to rewrite. 1057 SmallVector<PHIUsageRecord, 16> PHIUsers; 1058 1059 // PHIs are often mutually cyclic, so we keep track of a whole set of PHI 1060 // nodes which are extracted from. PHIsToSlice is a set we use to avoid 1061 // revisiting PHIs, PHIsInspected is a ordered list of PHIs that we need to 1062 // check the uses of (to ensure they are all extracts). 1063 SmallVector<PHINode*, 8> PHIsToSlice; 1064 SmallPtrSet<PHINode*, 8> PHIsInspected; 1065 1066 PHIsToSlice.push_back(&FirstPhi); 1067 PHIsInspected.insert(&FirstPhi); 1068 1069 for (unsigned PHIId = 0; PHIId != PHIsToSlice.size(); ++PHIId) { 1070 PHINode *PN = PHIsToSlice[PHIId]; 1071 1072 // Scan the input list of the PHI. If any input is an invoke, and if the 1073 // input is defined in the predecessor, then we won't be split the critical 1074 // edge which is required to insert a truncate. Because of this, we have to 1075 // bail out. 1076 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1077 InvokeInst *II = dyn_cast<InvokeInst>(PN->getIncomingValue(i)); 1078 if (!II) continue; 1079 if (II->getParent() != PN->getIncomingBlock(i)) 1080 continue; 1081 1082 // If we have a phi, and if it's directly in the predecessor, then we have 1083 // a critical edge where we need to put the truncate. Since we can't 1084 // split the edge in instcombine, we have to bail out. 1085 return nullptr; 1086 } 1087 1088 for (User *U : PN->users()) { 1089 Instruction *UserI = cast<Instruction>(U); 1090 1091 // If the user is a PHI, inspect its uses recursively. 1092 if (PHINode *UserPN = dyn_cast<PHINode>(UserI)) { 1093 if (PHIsInspected.insert(UserPN).second) 1094 PHIsToSlice.push_back(UserPN); 1095 continue; 1096 } 1097 1098 // Truncates are always ok. 1099 if (isa<TruncInst>(UserI)) { 1100 PHIUsers.push_back(PHIUsageRecord(PHIId, 0, UserI)); 1101 continue; 1102 } 1103 1104 // Otherwise it must be a lshr which can only be used by one trunc. 1105 if (UserI->getOpcode() != Instruction::LShr || 1106 !UserI->hasOneUse() || !isa<TruncInst>(UserI->user_back()) || 1107 !isa<ConstantInt>(UserI->getOperand(1))) 1108 return nullptr; 1109 1110 // Bail on out of range shifts. 1111 unsigned SizeInBits = UserI->getType()->getScalarSizeInBits(); 1112 if (cast<ConstantInt>(UserI->getOperand(1))->getValue().uge(SizeInBits)) 1113 return nullptr; 1114 1115 unsigned Shift = cast<ConstantInt>(UserI->getOperand(1))->getZExtValue(); 1116 PHIUsers.push_back(PHIUsageRecord(PHIId, Shift, UserI->user_back())); 1117 } 1118 } 1119 1120 // If we have no users, they must be all self uses, just nuke the PHI. 1121 if (PHIUsers.empty()) 1122 return replaceInstUsesWith(FirstPhi, UndefValue::get(FirstPhi.getType())); 1123 1124 // If this phi node is transformable, create new PHIs for all the pieces 1125 // extracted out of it. First, sort the users by their offset and size. 1126 array_pod_sort(PHIUsers.begin(), PHIUsers.end()); 1127 1128 LLVM_DEBUG(dbgs() << "SLICING UP PHI: " << FirstPhi << '\n'; 1129 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i) dbgs() 1130 << "AND USER PHI #" << i << ": " << *PHIsToSlice[i] << '\n';); 1131 1132 // PredValues - This is a temporary used when rewriting PHI nodes. It is 1133 // hoisted out here to avoid construction/destruction thrashing. 1134 DenseMap<BasicBlock*, Value*> PredValues; 1135 1136 // ExtractedVals - Each new PHI we introduce is saved here so we don't 1137 // introduce redundant PHIs. 1138 DenseMap<LoweredPHIRecord, PHINode*> ExtractedVals; 1139 1140 for (unsigned UserI = 0, UserE = PHIUsers.size(); UserI != UserE; ++UserI) { 1141 unsigned PHIId = PHIUsers[UserI].PHIId; 1142 PHINode *PN = PHIsToSlice[PHIId]; 1143 unsigned Offset = PHIUsers[UserI].Shift; 1144 Type *Ty = PHIUsers[UserI].Inst->getType(); 1145 1146 PHINode *EltPHI; 1147 1148 // If we've already lowered a user like this, reuse the previously lowered 1149 // value. 1150 if ((EltPHI = ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)]) == nullptr) { 1151 1152 // Otherwise, Create the new PHI node for this user. 1153 EltPHI = PHINode::Create(Ty, PN->getNumIncomingValues(), 1154 PN->getName()+".off"+Twine(Offset), PN); 1155 assert(EltPHI->getType() != PN->getType() && 1156 "Truncate didn't shrink phi?"); 1157 1158 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1159 BasicBlock *Pred = PN->getIncomingBlock(i); 1160 Value *&PredVal = PredValues[Pred]; 1161 1162 // If we already have a value for this predecessor, reuse it. 1163 if (PredVal) { 1164 EltPHI->addIncoming(PredVal, Pred); 1165 continue; 1166 } 1167 1168 // Handle the PHI self-reuse case. 1169 Value *InVal = PN->getIncomingValue(i); 1170 if (InVal == PN) { 1171 PredVal = EltPHI; 1172 EltPHI->addIncoming(PredVal, Pred); 1173 continue; 1174 } 1175 1176 if (PHINode *InPHI = dyn_cast<PHINode>(PN)) { 1177 // If the incoming value was a PHI, and if it was one of the PHIs we 1178 // already rewrote it, just use the lowered value. 1179 if (Value *Res = ExtractedVals[LoweredPHIRecord(InPHI, Offset, Ty)]) { 1180 PredVal = Res; 1181 EltPHI->addIncoming(PredVal, Pred); 1182 continue; 1183 } 1184 } 1185 1186 // Otherwise, do an extract in the predecessor. 1187 Builder.SetInsertPoint(Pred->getTerminator()); 1188 Value *Res = InVal; 1189 if (Offset) 1190 Res = Builder.CreateLShr(Res, ConstantInt::get(InVal->getType(), 1191 Offset), "extract"); 1192 Res = Builder.CreateTrunc(Res, Ty, "extract.t"); 1193 PredVal = Res; 1194 EltPHI->addIncoming(Res, Pred); 1195 1196 // If the incoming value was a PHI, and if it was one of the PHIs we are 1197 // rewriting, we will ultimately delete the code we inserted. This 1198 // means we need to revisit that PHI to make sure we extract out the 1199 // needed piece. 1200 if (PHINode *OldInVal = dyn_cast<PHINode>(PN->getIncomingValue(i))) 1201 if (PHIsInspected.count(OldInVal)) { 1202 unsigned RefPHIId = 1203 find(PHIsToSlice, OldInVal) - PHIsToSlice.begin(); 1204 PHIUsers.push_back(PHIUsageRecord(RefPHIId, Offset, 1205 cast<Instruction>(Res))); 1206 ++UserE; 1207 } 1208 } 1209 PredValues.clear(); 1210 1211 LLVM_DEBUG(dbgs() << " Made element PHI for offset " << Offset << ": " 1212 << *EltPHI << '\n'); 1213 ExtractedVals[LoweredPHIRecord(PN, Offset, Ty)] = EltPHI; 1214 } 1215 1216 // Replace the use of this piece with the PHI node. 1217 replaceInstUsesWith(*PHIUsers[UserI].Inst, EltPHI); 1218 } 1219 1220 // Replace all the remaining uses of the PHI nodes (self uses and the lshrs) 1221 // with undefs. 1222 Value *Undef = UndefValue::get(FirstPhi.getType()); 1223 for (unsigned i = 1, e = PHIsToSlice.size(); i != e; ++i) 1224 replaceInstUsesWith(*PHIsToSlice[i], Undef); 1225 return replaceInstUsesWith(FirstPhi, Undef); 1226 } 1227 1228 static Value *SimplifyUsingControlFlow(InstCombiner &Self, PHINode &PN, 1229 const DominatorTree &DT) { 1230 // Simplify the following patterns: 1231 // if (cond) 1232 // / \ 1233 // ... ... 1234 // \ / 1235 // phi [true] [false] 1236 if (!PN.getType()->isIntegerTy(1)) 1237 return nullptr; 1238 1239 if (PN.getNumOperands() != 2) 1240 return nullptr; 1241 1242 // Make sure all inputs are constants. 1243 if (!all_of(PN.operands(), [](Value *V) { return isa<ConstantInt>(V); })) 1244 return nullptr; 1245 1246 BasicBlock *BB = PN.getParent(); 1247 // Do not bother with unreachable instructions. 1248 if (!DT.isReachableFromEntry(BB)) 1249 return nullptr; 1250 1251 // Same inputs. 1252 if (PN.getOperand(0) == PN.getOperand(1)) 1253 return PN.getOperand(0); 1254 1255 BasicBlock *TruePred = nullptr, *FalsePred = nullptr; 1256 for (auto *Pred : predecessors(BB)) { 1257 auto *Input = cast<ConstantInt>(PN.getIncomingValueForBlock(Pred)); 1258 if (Input->isAllOnesValue()) 1259 TruePred = Pred; 1260 else 1261 FalsePred = Pred; 1262 } 1263 assert(TruePred && FalsePred && "Must be!"); 1264 1265 // Check which edge of the dominator dominates the true input. If it is the 1266 // false edge, we should invert the condition. 1267 auto *IDom = DT.getNode(BB)->getIDom()->getBlock(); 1268 auto *BI = dyn_cast<BranchInst>(IDom->getTerminator()); 1269 if (!BI || BI->isUnconditional()) 1270 return nullptr; 1271 1272 // Check that edges outgoing from the idom's terminators dominate respective 1273 // inputs of the Phi. 1274 BasicBlockEdge TrueOutEdge(IDom, BI->getSuccessor(0)); 1275 BasicBlockEdge FalseOutEdge(IDom, BI->getSuccessor(1)); 1276 1277 BasicBlockEdge TrueIncEdge(TruePred, BB); 1278 BasicBlockEdge FalseIncEdge(FalsePred, BB); 1279 1280 auto *Cond = BI->getCondition(); 1281 if (DT.dominates(TrueOutEdge, TrueIncEdge) && 1282 DT.dominates(FalseOutEdge, FalseIncEdge)) 1283 // This Phi is actually equivalent to branching condition of IDom. 1284 return Cond; 1285 else if (DT.dominates(TrueOutEdge, FalseIncEdge) && 1286 DT.dominates(FalseOutEdge, TrueIncEdge)) { 1287 // This Phi is actually opposite to branching condition of IDom. We invert 1288 // the condition that will potentially open up some opportunities for 1289 // sinking. 1290 auto InsertPt = BB->getFirstInsertionPt(); 1291 if (InsertPt != BB->end()) { 1292 Self.Builder.SetInsertPoint(&*InsertPt); 1293 return Self.Builder.CreateNot(Cond); 1294 } 1295 } 1296 1297 return nullptr; 1298 } 1299 1300 // PHINode simplification 1301 // 1302 Instruction *InstCombinerImpl::visitPHINode(PHINode &PN) { 1303 if (Value *V = SimplifyInstruction(&PN, SQ.getWithInstruction(&PN))) 1304 return replaceInstUsesWith(PN, V); 1305 1306 if (Instruction *Result = foldPHIArgZextsIntoPHI(PN)) 1307 return Result; 1308 1309 // If all PHI operands are the same operation, pull them through the PHI, 1310 // reducing code size. 1311 if (isa<Instruction>(PN.getIncomingValue(0)) && 1312 isa<Instruction>(PN.getIncomingValue(1)) && 1313 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() == 1314 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() && 1315 PN.getIncomingValue(0)->hasOneUser()) 1316 if (Instruction *Result = foldPHIArgOpIntoPHI(PN)) 1317 return Result; 1318 1319 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if 1320 // this PHI only has a single use (a PHI), and if that PHI only has one use (a 1321 // PHI)... break the cycle. 1322 if (PN.hasOneUse()) { 1323 if (Instruction *Result = foldIntegerTypedPHI(PN)) 1324 return Result; 1325 1326 Instruction *PHIUser = cast<Instruction>(PN.user_back()); 1327 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) { 1328 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs; 1329 PotentiallyDeadPHIs.insert(&PN); 1330 if (DeadPHICycle(PU, PotentiallyDeadPHIs)) 1331 return replaceInstUsesWith(PN, UndefValue::get(PN.getType())); 1332 } 1333 1334 // If this phi has a single use, and if that use just computes a value for 1335 // the next iteration of a loop, delete the phi. This occurs with unused 1336 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this 1337 // common case here is good because the only other things that catch this 1338 // are induction variable analysis (sometimes) and ADCE, which is only run 1339 // late. 1340 if (PHIUser->hasOneUse() && 1341 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) && 1342 PHIUser->user_back() == &PN) { 1343 return replaceInstUsesWith(PN, UndefValue::get(PN.getType())); 1344 } 1345 // When a PHI is used only to be compared with zero, it is safe to replace 1346 // an incoming value proved as known nonzero with any non-zero constant. 1347 // For example, in the code below, the incoming value %v can be replaced 1348 // with any non-zero constant based on the fact that the PHI is only used to 1349 // be compared with zero and %v is a known non-zero value: 1350 // %v = select %cond, 1, 2 1351 // %p = phi [%v, BB] ... 1352 // icmp eq, %p, 0 1353 auto *CmpInst = dyn_cast<ICmpInst>(PHIUser); 1354 // FIXME: To be simple, handle only integer type for now. 1355 if (CmpInst && isa<IntegerType>(PN.getType()) && CmpInst->isEquality() && 1356 match(CmpInst->getOperand(1), m_Zero())) { 1357 ConstantInt *NonZeroConst = nullptr; 1358 bool MadeChange = false; 1359 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) { 1360 Instruction *CtxI = PN.getIncomingBlock(i)->getTerminator(); 1361 Value *VA = PN.getIncomingValue(i); 1362 if (isKnownNonZero(VA, DL, 0, &AC, CtxI, &DT)) { 1363 if (!NonZeroConst) 1364 NonZeroConst = GetAnyNonZeroConstInt(PN); 1365 1366 if (NonZeroConst != VA) { 1367 replaceOperand(PN, i, NonZeroConst); 1368 MadeChange = true; 1369 } 1370 } 1371 } 1372 if (MadeChange) 1373 return &PN; 1374 } 1375 } 1376 1377 // We sometimes end up with phi cycles that non-obviously end up being the 1378 // same value, for example: 1379 // z = some value; x = phi (y, z); y = phi (x, z) 1380 // where the phi nodes don't necessarily need to be in the same block. Do a 1381 // quick check to see if the PHI node only contains a single non-phi value, if 1382 // so, scan to see if the phi cycle is actually equal to that value. 1383 { 1384 unsigned InValNo = 0, NumIncomingVals = PN.getNumIncomingValues(); 1385 // Scan for the first non-phi operand. 1386 while (InValNo != NumIncomingVals && 1387 isa<PHINode>(PN.getIncomingValue(InValNo))) 1388 ++InValNo; 1389 1390 if (InValNo != NumIncomingVals) { 1391 Value *NonPhiInVal = PN.getIncomingValue(InValNo); 1392 1393 // Scan the rest of the operands to see if there are any conflicts, if so 1394 // there is no need to recursively scan other phis. 1395 for (++InValNo; InValNo != NumIncomingVals; ++InValNo) { 1396 Value *OpVal = PN.getIncomingValue(InValNo); 1397 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal)) 1398 break; 1399 } 1400 1401 // If we scanned over all operands, then we have one unique value plus 1402 // phi values. Scan PHI nodes to see if they all merge in each other or 1403 // the value. 1404 if (InValNo == NumIncomingVals) { 1405 SmallPtrSet<PHINode*, 16> ValueEqualPHIs; 1406 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs)) 1407 return replaceInstUsesWith(PN, NonPhiInVal); 1408 } 1409 } 1410 } 1411 1412 // If there are multiple PHIs, sort their operands so that they all list 1413 // the blocks in the same order. This will help identical PHIs be eliminated 1414 // by other passes. Other passes shouldn't depend on this for correctness 1415 // however. 1416 PHINode *FirstPN = cast<PHINode>(PN.getParent()->begin()); 1417 if (&PN != FirstPN) 1418 for (unsigned i = 0, e = FirstPN->getNumIncomingValues(); i != e; ++i) { 1419 BasicBlock *BBA = PN.getIncomingBlock(i); 1420 BasicBlock *BBB = FirstPN->getIncomingBlock(i); 1421 if (BBA != BBB) { 1422 Value *VA = PN.getIncomingValue(i); 1423 unsigned j = PN.getBasicBlockIndex(BBB); 1424 Value *VB = PN.getIncomingValue(j); 1425 PN.setIncomingBlock(i, BBB); 1426 PN.setIncomingValue(i, VB); 1427 PN.setIncomingBlock(j, BBA); 1428 PN.setIncomingValue(j, VA); 1429 // NOTE: Instcombine normally would want us to "return &PN" if we 1430 // modified any of the operands of an instruction. However, since we 1431 // aren't adding or removing uses (just rearranging them) we don't do 1432 // this in this case. 1433 } 1434 } 1435 1436 // Is there an identical PHI node in this basic block? 1437 for (PHINode &IdenticalPN : PN.getParent()->phis()) { 1438 // Ignore the PHI node itself. 1439 if (&IdenticalPN == &PN) 1440 continue; 1441 // Note that even though we've just canonicalized this PHI, due to the 1442 // worklist visitation order, there are no guarantess that *every* PHI 1443 // has been canonicalized, so we can't just compare operands ranges. 1444 if (!PN.isIdenticalToWhenDefined(&IdenticalPN)) 1445 continue; 1446 // Just use that PHI instead then. 1447 ++NumPHICSEs; 1448 return replaceInstUsesWith(PN, &IdenticalPN); 1449 } 1450 1451 // If this is an integer PHI and we know that it has an illegal type, see if 1452 // it is only used by trunc or trunc(lshr) operations. If so, we split the 1453 // PHI into the various pieces being extracted. This sort of thing is 1454 // introduced when SROA promotes an aggregate to a single large integer type. 1455 if (PN.getType()->isIntegerTy() && 1456 !DL.isLegalInteger(PN.getType()->getPrimitiveSizeInBits())) 1457 if (Instruction *Res = SliceUpIllegalIntegerPHI(PN)) 1458 return Res; 1459 1460 // Ultimately, try to replace this Phi with a dominating condition. 1461 if (auto *V = SimplifyUsingControlFlow(*this, PN, DT)) 1462 return replaceInstUsesWith(PN, V); 1463 1464 return nullptr; 1465 } 1466