1 //===- EarlyCSE.cpp - Simple and fast CSE pass ----------------------------===// 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 pass performs a simple dominator tree walk that eliminates trivially 10 // redundant instructions. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Transforms/Scalar/EarlyCSE.h" 15 #include "llvm/ADT/DenseMapInfo.h" 16 #include "llvm/ADT/Hashing.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/ScopedHashTable.h" 19 #include "llvm/ADT/SetVector.h" 20 #include "llvm/ADT/SmallVector.h" 21 #include "llvm/ADT/Statistic.h" 22 #include "llvm/Analysis/AssumptionCache.h" 23 #include "llvm/Analysis/GlobalsModRef.h" 24 #include "llvm/Analysis/GuardUtils.h" 25 #include "llvm/Analysis/InstructionSimplify.h" 26 #include "llvm/Analysis/MemorySSA.h" 27 #include "llvm/Analysis/MemorySSAUpdater.h" 28 #include "llvm/Analysis/TargetLibraryInfo.h" 29 #include "llvm/Analysis/TargetTransformInfo.h" 30 #include "llvm/Analysis/ValueTracking.h" 31 #include "llvm/IR/BasicBlock.h" 32 #include "llvm/IR/Constants.h" 33 #include "llvm/IR/DataLayout.h" 34 #include "llvm/IR/Dominators.h" 35 #include "llvm/IR/Function.h" 36 #include "llvm/IR/InstrTypes.h" 37 #include "llvm/IR/Instruction.h" 38 #include "llvm/IR/Instructions.h" 39 #include "llvm/IR/IntrinsicInst.h" 40 #include "llvm/IR/Intrinsics.h" 41 #include "llvm/IR/LLVMContext.h" 42 #include "llvm/IR/PassManager.h" 43 #include "llvm/IR/PatternMatch.h" 44 #include "llvm/IR/Type.h" 45 #include "llvm/IR/Use.h" 46 #include "llvm/IR/Value.h" 47 #include "llvm/InitializePasses.h" 48 #include "llvm/Pass.h" 49 #include "llvm/Support/Allocator.h" 50 #include "llvm/Support/AtomicOrdering.h" 51 #include "llvm/Support/Casting.h" 52 #include "llvm/Support/Debug.h" 53 #include "llvm/Support/DebugCounter.h" 54 #include "llvm/Support/RecyclingAllocator.h" 55 #include "llvm/Support/raw_ostream.h" 56 #include "llvm/Transforms/Scalar.h" 57 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" 58 #include "llvm/Transforms/Utils/GuardUtils.h" 59 #include "llvm/Transforms/Utils/Local.h" 60 #include <cassert> 61 #include <deque> 62 #include <memory> 63 #include <utility> 64 65 using namespace llvm; 66 using namespace llvm::PatternMatch; 67 68 #define DEBUG_TYPE "early-cse" 69 70 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd"); 71 STATISTIC(NumCSE, "Number of instructions CSE'd"); 72 STATISTIC(NumCSECVP, "Number of compare instructions CVP'd"); 73 STATISTIC(NumCSELoad, "Number of load instructions CSE'd"); 74 STATISTIC(NumCSECall, "Number of call instructions CSE'd"); 75 STATISTIC(NumDSE, "Number of trivial dead stores removed"); 76 77 DEBUG_COUNTER(CSECounter, "early-cse", 78 "Controls which instructions are removed"); 79 80 static cl::opt<unsigned> EarlyCSEMssaOptCap( 81 "earlycse-mssa-optimization-cap", cl::init(500), cl::Hidden, 82 cl::desc("Enable imprecision in EarlyCSE in pathological cases, in exchange " 83 "for faster compile. Caps the MemorySSA clobbering calls.")); 84 85 static cl::opt<bool> EarlyCSEDebugHash( 86 "earlycse-debug-hash", cl::init(false), cl::Hidden, 87 cl::desc("Perform extra assertion checking to verify that SimpleValue's hash " 88 "function is well-behaved w.r.t. its isEqual predicate")); 89 90 //===----------------------------------------------------------------------===// 91 // SimpleValue 92 //===----------------------------------------------------------------------===// 93 94 namespace { 95 96 /// Struct representing the available values in the scoped hash table. 97 struct SimpleValue { 98 Instruction *Inst; 99 100 SimpleValue(Instruction *I) : Inst(I) { 101 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!"); 102 } 103 104 bool isSentinel() const { 105 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() || 106 Inst == DenseMapInfo<Instruction *>::getTombstoneKey(); 107 } 108 109 static bool canHandle(Instruction *Inst) { 110 // This can only handle non-void readnone functions. 111 // Also handled are constrained intrinsic that look like the types 112 // of instruction handled below (UnaryOperator, etc.). 113 if (CallInst *CI = dyn_cast<CallInst>(Inst)) { 114 if (Function *F = CI->getCalledFunction()) { 115 switch ((Intrinsic::ID)F->getIntrinsicID()) { 116 case Intrinsic::experimental_constrained_fadd: 117 case Intrinsic::experimental_constrained_fsub: 118 case Intrinsic::experimental_constrained_fmul: 119 case Intrinsic::experimental_constrained_fdiv: 120 case Intrinsic::experimental_constrained_frem: 121 case Intrinsic::experimental_constrained_fptosi: 122 case Intrinsic::experimental_constrained_sitofp: 123 case Intrinsic::experimental_constrained_fptoui: 124 case Intrinsic::experimental_constrained_uitofp: 125 case Intrinsic::experimental_constrained_fcmp: 126 case Intrinsic::experimental_constrained_fcmps: { 127 auto *CFP = cast<ConstrainedFPIntrinsic>(CI); 128 return CFP->isDefaultFPEnvironment(); 129 } 130 } 131 } 132 return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy(); 133 } 134 return isa<CastInst>(Inst) || isa<UnaryOperator>(Inst) || 135 isa<BinaryOperator>(Inst) || isa<GetElementPtrInst>(Inst) || 136 isa<CmpInst>(Inst) || isa<SelectInst>(Inst) || 137 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) || 138 isa<ShuffleVectorInst>(Inst) || isa<ExtractValueInst>(Inst) || 139 isa<InsertValueInst>(Inst) || isa<FreezeInst>(Inst); 140 } 141 }; 142 143 } // end anonymous namespace 144 145 namespace llvm { 146 147 template <> struct DenseMapInfo<SimpleValue> { 148 static inline SimpleValue getEmptyKey() { 149 return DenseMapInfo<Instruction *>::getEmptyKey(); 150 } 151 152 static inline SimpleValue getTombstoneKey() { 153 return DenseMapInfo<Instruction *>::getTombstoneKey(); 154 } 155 156 static unsigned getHashValue(SimpleValue Val); 157 static bool isEqual(SimpleValue LHS, SimpleValue RHS); 158 }; 159 160 } // end namespace llvm 161 162 /// Match a 'select' including an optional 'not's of the condition. 163 static bool matchSelectWithOptionalNotCond(Value *V, Value *&Cond, Value *&A, 164 Value *&B, 165 SelectPatternFlavor &Flavor) { 166 // Return false if V is not even a select. 167 if (!match(V, m_Select(m_Value(Cond), m_Value(A), m_Value(B)))) 168 return false; 169 170 // Look through a 'not' of the condition operand by swapping A/B. 171 Value *CondNot; 172 if (match(Cond, m_Not(m_Value(CondNot)))) { 173 Cond = CondNot; 174 std::swap(A, B); 175 } 176 177 // Match canonical forms of min/max. We are not using ValueTracking's 178 // more powerful matchSelectPattern() because it may rely on instruction flags 179 // such as "nsw". That would be incompatible with the current hashing 180 // mechanism that may remove flags to increase the likelihood of CSE. 181 182 Flavor = SPF_UNKNOWN; 183 CmpInst::Predicate Pred; 184 185 if (!match(Cond, m_ICmp(Pred, m_Specific(A), m_Specific(B)))) { 186 // Check for commuted variants of min/max by swapping predicate. 187 // If we do not match the standard or commuted patterns, this is not a 188 // recognized form of min/max, but it is still a select, so return true. 189 if (!match(Cond, m_ICmp(Pred, m_Specific(B), m_Specific(A)))) 190 return true; 191 Pred = ICmpInst::getSwappedPredicate(Pred); 192 } 193 194 switch (Pred) { 195 case CmpInst::ICMP_UGT: Flavor = SPF_UMAX; break; 196 case CmpInst::ICMP_ULT: Flavor = SPF_UMIN; break; 197 case CmpInst::ICMP_SGT: Flavor = SPF_SMAX; break; 198 case CmpInst::ICMP_SLT: Flavor = SPF_SMIN; break; 199 // Non-strict inequalities. 200 case CmpInst::ICMP_ULE: Flavor = SPF_UMIN; break; 201 case CmpInst::ICMP_UGE: Flavor = SPF_UMAX; break; 202 case CmpInst::ICMP_SLE: Flavor = SPF_SMIN; break; 203 case CmpInst::ICMP_SGE: Flavor = SPF_SMAX; break; 204 default: break; 205 } 206 207 return true; 208 } 209 210 static unsigned getHashValueImpl(SimpleValue Val) { 211 Instruction *Inst = Val.Inst; 212 // Hash in all of the operands as pointers. 213 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) { 214 Value *LHS = BinOp->getOperand(0); 215 Value *RHS = BinOp->getOperand(1); 216 if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1)) 217 std::swap(LHS, RHS); 218 219 return hash_combine(BinOp->getOpcode(), LHS, RHS); 220 } 221 222 if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) { 223 // Compares can be commuted by swapping the comparands and 224 // updating the predicate. Choose the form that has the 225 // comparands in sorted order, or in the case of a tie, the 226 // one with the lower predicate. 227 Value *LHS = CI->getOperand(0); 228 Value *RHS = CI->getOperand(1); 229 CmpInst::Predicate Pred = CI->getPredicate(); 230 CmpInst::Predicate SwappedPred = CI->getSwappedPredicate(); 231 if (std::tie(LHS, Pred) > std::tie(RHS, SwappedPred)) { 232 std::swap(LHS, RHS); 233 Pred = SwappedPred; 234 } 235 return hash_combine(Inst->getOpcode(), Pred, LHS, RHS); 236 } 237 238 // Hash general selects to allow matching commuted true/false operands. 239 SelectPatternFlavor SPF; 240 Value *Cond, *A, *B; 241 if (matchSelectWithOptionalNotCond(Inst, Cond, A, B, SPF)) { 242 // Hash min/max (cmp + select) to allow for commuted operands. 243 // Min/max may also have non-canonical compare predicate (eg, the compare for 244 // smin may use 'sgt' rather than 'slt'), and non-canonical operands in the 245 // compare. 246 // TODO: We should also detect FP min/max. 247 if (SPF == SPF_SMIN || SPF == SPF_SMAX || 248 SPF == SPF_UMIN || SPF == SPF_UMAX) { 249 if (A > B) 250 std::swap(A, B); 251 return hash_combine(Inst->getOpcode(), SPF, A, B); 252 } 253 254 // Hash general selects to allow matching commuted true/false operands. 255 256 // If we do not have a compare as the condition, just hash in the condition. 257 CmpInst::Predicate Pred; 258 Value *X, *Y; 259 if (!match(Cond, m_Cmp(Pred, m_Value(X), m_Value(Y)))) 260 return hash_combine(Inst->getOpcode(), Cond, A, B); 261 262 // Similar to cmp normalization (above) - canonicalize the predicate value: 263 // select (icmp Pred, X, Y), A, B --> select (icmp InvPred, X, Y), B, A 264 if (CmpInst::getInversePredicate(Pred) < Pred) { 265 Pred = CmpInst::getInversePredicate(Pred); 266 std::swap(A, B); 267 } 268 return hash_combine(Inst->getOpcode(), Pred, X, Y, A, B); 269 } 270 271 if (CastInst *CI = dyn_cast<CastInst>(Inst)) 272 return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0)); 273 274 if (FreezeInst *FI = dyn_cast<FreezeInst>(Inst)) 275 return hash_combine(FI->getOpcode(), FI->getOperand(0)); 276 277 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst)) 278 return hash_combine(EVI->getOpcode(), EVI->getOperand(0), 279 hash_combine_range(EVI->idx_begin(), EVI->idx_end())); 280 281 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst)) 282 return hash_combine(IVI->getOpcode(), IVI->getOperand(0), 283 IVI->getOperand(1), 284 hash_combine_range(IVI->idx_begin(), IVI->idx_end())); 285 286 assert((isa<CallInst>(Inst) || isa<GetElementPtrInst>(Inst) || 287 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) || 288 isa<ShuffleVectorInst>(Inst) || isa<UnaryOperator>(Inst) || 289 isa<FreezeInst>(Inst)) && 290 "Invalid/unknown instruction"); 291 292 // Handle intrinsics with commutative operands. 293 // TODO: Extend this to handle intrinsics with >2 operands where the 1st 294 // 2 operands are commutative. 295 auto *II = dyn_cast<IntrinsicInst>(Inst); 296 if (II && II->isCommutative() && II->getNumArgOperands() == 2) { 297 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 298 if (LHS > RHS) 299 std::swap(LHS, RHS); 300 return hash_combine(II->getOpcode(), LHS, RHS); 301 } 302 303 // gc.relocate is 'special' call: its second and third operands are 304 // not real values, but indices into statepoint's argument list. 305 // Get values they point to. 306 if (const GCRelocateInst *GCR = dyn_cast<GCRelocateInst>(Inst)) 307 return hash_combine(GCR->getOpcode(), GCR->getOperand(0), 308 GCR->getBasePtr(), GCR->getDerivedPtr()); 309 310 // Mix in the opcode. 311 return hash_combine( 312 Inst->getOpcode(), 313 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end())); 314 } 315 316 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) { 317 #ifndef NDEBUG 318 // If -earlycse-debug-hash was specified, return a constant -- this 319 // will force all hashing to collide, so we'll exhaustively search 320 // the table for a match, and the assertion in isEqual will fire if 321 // there's a bug causing equal keys to hash differently. 322 if (EarlyCSEDebugHash) 323 return 0; 324 #endif 325 return getHashValueImpl(Val); 326 } 327 328 static bool isEqualImpl(SimpleValue LHS, SimpleValue RHS) { 329 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst; 330 331 if (LHS.isSentinel() || RHS.isSentinel()) 332 return LHSI == RHSI; 333 334 if (LHSI->getOpcode() != RHSI->getOpcode()) 335 return false; 336 if (LHSI->isIdenticalToWhenDefined(RHSI)) 337 return true; 338 339 // If we're not strictly identical, we still might be a commutable instruction 340 if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) { 341 if (!LHSBinOp->isCommutative()) 342 return false; 343 344 assert(isa<BinaryOperator>(RHSI) && 345 "same opcode, but different instruction type?"); 346 BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI); 347 348 // Commuted equality 349 return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) && 350 LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0); 351 } 352 if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) { 353 assert(isa<CmpInst>(RHSI) && 354 "same opcode, but different instruction type?"); 355 CmpInst *RHSCmp = cast<CmpInst>(RHSI); 356 // Commuted equality 357 return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) && 358 LHSCmp->getOperand(1) == RHSCmp->getOperand(0) && 359 LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate(); 360 } 361 362 // TODO: Extend this for >2 args by matching the trailing N-2 args. 363 auto *LII = dyn_cast<IntrinsicInst>(LHSI); 364 auto *RII = dyn_cast<IntrinsicInst>(RHSI); 365 if (LII && RII && LII->getIntrinsicID() == RII->getIntrinsicID() && 366 LII->isCommutative() && LII->getNumArgOperands() == 2) { 367 return LII->getArgOperand(0) == RII->getArgOperand(1) && 368 LII->getArgOperand(1) == RII->getArgOperand(0); 369 } 370 371 // See comment above in `getHashValue()`. 372 if (const GCRelocateInst *GCR1 = dyn_cast<GCRelocateInst>(LHSI)) 373 if (const GCRelocateInst *GCR2 = dyn_cast<GCRelocateInst>(RHSI)) 374 return GCR1->getOperand(0) == GCR2->getOperand(0) && 375 GCR1->getBasePtr() == GCR2->getBasePtr() && 376 GCR1->getDerivedPtr() == GCR2->getDerivedPtr(); 377 378 // Min/max can occur with commuted operands, non-canonical predicates, 379 // and/or non-canonical operands. 380 // Selects can be non-trivially equivalent via inverted conditions and swaps. 381 SelectPatternFlavor LSPF, RSPF; 382 Value *CondL, *CondR, *LHSA, *RHSA, *LHSB, *RHSB; 383 if (matchSelectWithOptionalNotCond(LHSI, CondL, LHSA, LHSB, LSPF) && 384 matchSelectWithOptionalNotCond(RHSI, CondR, RHSA, RHSB, RSPF)) { 385 if (LSPF == RSPF) { 386 // TODO: We should also detect FP min/max. 387 if (LSPF == SPF_SMIN || LSPF == SPF_SMAX || 388 LSPF == SPF_UMIN || LSPF == SPF_UMAX) 389 return ((LHSA == RHSA && LHSB == RHSB) || 390 (LHSA == RHSB && LHSB == RHSA)); 391 392 // select Cond, A, B <--> select not(Cond), B, A 393 if (CondL == CondR && LHSA == RHSA && LHSB == RHSB) 394 return true; 395 } 396 397 // If the true/false operands are swapped and the conditions are compares 398 // with inverted predicates, the selects are equal: 399 // select (icmp Pred, X, Y), A, B <--> select (icmp InvPred, X, Y), B, A 400 // 401 // This also handles patterns with a double-negation in the sense of not + 402 // inverse, because we looked through a 'not' in the matching function and 403 // swapped A/B: 404 // select (cmp Pred, X, Y), A, B <--> select (not (cmp InvPred, X, Y)), B, A 405 // 406 // This intentionally does NOT handle patterns with a double-negation in 407 // the sense of not + not, because doing so could result in values 408 // comparing 409 // as equal that hash differently in the min/max cases like: 410 // select (cmp slt, X, Y), X, Y <--> select (not (not (cmp slt, X, Y))), X, Y 411 // ^ hashes as min ^ would not hash as min 412 // In the context of the EarlyCSE pass, however, such cases never reach 413 // this code, as we simplify the double-negation before hashing the second 414 // select (and so still succeed at CSEing them). 415 if (LHSA == RHSB && LHSB == RHSA) { 416 CmpInst::Predicate PredL, PredR; 417 Value *X, *Y; 418 if (match(CondL, m_Cmp(PredL, m_Value(X), m_Value(Y))) && 419 match(CondR, m_Cmp(PredR, m_Specific(X), m_Specific(Y))) && 420 CmpInst::getInversePredicate(PredL) == PredR) 421 return true; 422 } 423 } 424 425 return false; 426 } 427 428 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) { 429 // These comparisons are nontrivial, so assert that equality implies 430 // hash equality (DenseMap demands this as an invariant). 431 bool Result = isEqualImpl(LHS, RHS); 432 assert(!Result || (LHS.isSentinel() && LHS.Inst == RHS.Inst) || 433 getHashValueImpl(LHS) == getHashValueImpl(RHS)); 434 return Result; 435 } 436 437 //===----------------------------------------------------------------------===// 438 // CallValue 439 //===----------------------------------------------------------------------===// 440 441 namespace { 442 443 /// Struct representing the available call values in the scoped hash 444 /// table. 445 struct CallValue { 446 Instruction *Inst; 447 448 CallValue(Instruction *I) : Inst(I) { 449 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!"); 450 } 451 452 bool isSentinel() const { 453 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() || 454 Inst == DenseMapInfo<Instruction *>::getTombstoneKey(); 455 } 456 457 static bool canHandle(Instruction *Inst) { 458 // Don't value number anything that returns void. 459 if (Inst->getType()->isVoidTy()) 460 return false; 461 462 CallInst *CI = dyn_cast<CallInst>(Inst); 463 if (!CI || !CI->onlyReadsMemory()) 464 return false; 465 return true; 466 } 467 }; 468 469 } // end anonymous namespace 470 471 namespace llvm { 472 473 template <> struct DenseMapInfo<CallValue> { 474 static inline CallValue getEmptyKey() { 475 return DenseMapInfo<Instruction *>::getEmptyKey(); 476 } 477 478 static inline CallValue getTombstoneKey() { 479 return DenseMapInfo<Instruction *>::getTombstoneKey(); 480 } 481 482 static unsigned getHashValue(CallValue Val); 483 static bool isEqual(CallValue LHS, CallValue RHS); 484 }; 485 486 } // end namespace llvm 487 488 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) { 489 Instruction *Inst = Val.Inst; 490 491 // Hash all of the operands as pointers and mix in the opcode. 492 return hash_combine( 493 Inst->getOpcode(), 494 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end())); 495 } 496 497 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) { 498 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst; 499 if (LHS.isSentinel() || RHS.isSentinel()) 500 return LHSI == RHSI; 501 502 return LHSI->isIdenticalTo(RHSI); 503 } 504 505 //===----------------------------------------------------------------------===// 506 // EarlyCSE implementation 507 //===----------------------------------------------------------------------===// 508 509 namespace { 510 511 /// A simple and fast domtree-based CSE pass. 512 /// 513 /// This pass does a simple depth-first walk over the dominator tree, 514 /// eliminating trivially redundant instructions and using instsimplify to 515 /// canonicalize things as it goes. It is intended to be fast and catch obvious 516 /// cases so that instcombine and other passes are more effective. It is 517 /// expected that a later pass of GVN will catch the interesting/hard cases. 518 class EarlyCSE { 519 public: 520 const TargetLibraryInfo &TLI; 521 const TargetTransformInfo &TTI; 522 DominatorTree &DT; 523 AssumptionCache &AC; 524 const SimplifyQuery SQ; 525 MemorySSA *MSSA; 526 std::unique_ptr<MemorySSAUpdater> MSSAUpdater; 527 528 using AllocatorTy = 529 RecyclingAllocator<BumpPtrAllocator, 530 ScopedHashTableVal<SimpleValue, Value *>>; 531 using ScopedHTType = 532 ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>, 533 AllocatorTy>; 534 535 /// A scoped hash table of the current values of all of our simple 536 /// scalar expressions. 537 /// 538 /// As we walk down the domtree, we look to see if instructions are in this: 539 /// if so, we replace them with what we find, otherwise we insert them so 540 /// that dominated values can succeed in their lookup. 541 ScopedHTType AvailableValues; 542 543 /// A scoped hash table of the current values of previously encountered 544 /// memory locations. 545 /// 546 /// This allows us to get efficient access to dominating loads or stores when 547 /// we have a fully redundant load. In addition to the most recent load, we 548 /// keep track of a generation count of the read, which is compared against 549 /// the current generation count. The current generation count is incremented 550 /// after every possibly writing memory operation, which ensures that we only 551 /// CSE loads with other loads that have no intervening store. Ordering 552 /// events (such as fences or atomic instructions) increment the generation 553 /// count as well; essentially, we model these as writes to all possible 554 /// locations. Note that atomic and/or volatile loads and stores can be 555 /// present the table; it is the responsibility of the consumer to inspect 556 /// the atomicity/volatility if needed. 557 struct LoadValue { 558 Instruction *DefInst = nullptr; 559 unsigned Generation = 0; 560 int MatchingId = -1; 561 bool IsAtomic = false; 562 563 LoadValue() = default; 564 LoadValue(Instruction *Inst, unsigned Generation, unsigned MatchingId, 565 bool IsAtomic) 566 : DefInst(Inst), Generation(Generation), MatchingId(MatchingId), 567 IsAtomic(IsAtomic) {} 568 }; 569 570 using LoadMapAllocator = 571 RecyclingAllocator<BumpPtrAllocator, 572 ScopedHashTableVal<Value *, LoadValue>>; 573 using LoadHTType = 574 ScopedHashTable<Value *, LoadValue, DenseMapInfo<Value *>, 575 LoadMapAllocator>; 576 577 LoadHTType AvailableLoads; 578 579 // A scoped hash table mapping memory locations (represented as typed 580 // addresses) to generation numbers at which that memory location became 581 // (henceforth indefinitely) invariant. 582 using InvariantMapAllocator = 583 RecyclingAllocator<BumpPtrAllocator, 584 ScopedHashTableVal<MemoryLocation, unsigned>>; 585 using InvariantHTType = 586 ScopedHashTable<MemoryLocation, unsigned, DenseMapInfo<MemoryLocation>, 587 InvariantMapAllocator>; 588 InvariantHTType AvailableInvariants; 589 590 /// A scoped hash table of the current values of read-only call 591 /// values. 592 /// 593 /// It uses the same generation count as loads. 594 using CallHTType = 595 ScopedHashTable<CallValue, std::pair<Instruction *, unsigned>>; 596 CallHTType AvailableCalls; 597 598 /// This is the current generation of the memory value. 599 unsigned CurrentGeneration = 0; 600 601 /// Set up the EarlyCSE runner for a particular function. 602 EarlyCSE(const DataLayout &DL, const TargetLibraryInfo &TLI, 603 const TargetTransformInfo &TTI, DominatorTree &DT, 604 AssumptionCache &AC, MemorySSA *MSSA) 605 : TLI(TLI), TTI(TTI), DT(DT), AC(AC), SQ(DL, &TLI, &DT, &AC), MSSA(MSSA), 606 MSSAUpdater(std::make_unique<MemorySSAUpdater>(MSSA)) {} 607 608 bool run(); 609 610 private: 611 unsigned ClobberCounter = 0; 612 // Almost a POD, but needs to call the constructors for the scoped hash 613 // tables so that a new scope gets pushed on. These are RAII so that the 614 // scope gets popped when the NodeScope is destroyed. 615 class NodeScope { 616 public: 617 NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads, 618 InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls) 619 : Scope(AvailableValues), LoadScope(AvailableLoads), 620 InvariantScope(AvailableInvariants), CallScope(AvailableCalls) {} 621 NodeScope(const NodeScope &) = delete; 622 NodeScope &operator=(const NodeScope &) = delete; 623 624 private: 625 ScopedHTType::ScopeTy Scope; 626 LoadHTType::ScopeTy LoadScope; 627 InvariantHTType::ScopeTy InvariantScope; 628 CallHTType::ScopeTy CallScope; 629 }; 630 631 // Contains all the needed information to create a stack for doing a depth 632 // first traversal of the tree. This includes scopes for values, loads, and 633 // calls as well as the generation. There is a child iterator so that the 634 // children do not need to be store separately. 635 class StackNode { 636 public: 637 StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads, 638 InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls, 639 unsigned cg, DomTreeNode *n, DomTreeNode::const_iterator child, 640 DomTreeNode::const_iterator end) 641 : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child), 642 EndIter(end), 643 Scopes(AvailableValues, AvailableLoads, AvailableInvariants, 644 AvailableCalls) 645 {} 646 StackNode(const StackNode &) = delete; 647 StackNode &operator=(const StackNode &) = delete; 648 649 // Accessors. 650 unsigned currentGeneration() const { return CurrentGeneration; } 651 unsigned childGeneration() const { return ChildGeneration; } 652 void childGeneration(unsigned generation) { ChildGeneration = generation; } 653 DomTreeNode *node() { return Node; } 654 DomTreeNode::const_iterator childIter() const { return ChildIter; } 655 656 DomTreeNode *nextChild() { 657 DomTreeNode *child = *ChildIter; 658 ++ChildIter; 659 return child; 660 } 661 662 DomTreeNode::const_iterator end() const { return EndIter; } 663 bool isProcessed() const { return Processed; } 664 void process() { Processed = true; } 665 666 private: 667 unsigned CurrentGeneration; 668 unsigned ChildGeneration; 669 DomTreeNode *Node; 670 DomTreeNode::const_iterator ChildIter; 671 DomTreeNode::const_iterator EndIter; 672 NodeScope Scopes; 673 bool Processed = false; 674 }; 675 676 /// Wrapper class to handle memory instructions, including loads, 677 /// stores and intrinsic loads and stores defined by the target. 678 class ParseMemoryInst { 679 public: 680 ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI) 681 : Inst(Inst) { 682 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 683 IntrID = II->getIntrinsicID(); 684 if (TTI.getTgtMemIntrinsic(II, Info)) 685 return; 686 if (isHandledNonTargetIntrinsic(IntrID)) { 687 switch (IntrID) { 688 case Intrinsic::masked_load: 689 Info.PtrVal = Inst->getOperand(0); 690 Info.MatchingId = Intrinsic::masked_load; 691 Info.ReadMem = true; 692 Info.WriteMem = false; 693 Info.IsVolatile = false; 694 break; 695 case Intrinsic::masked_store: 696 Info.PtrVal = Inst->getOperand(1); 697 // Use the ID of masked load as the "matching id". This will 698 // prevent matching non-masked loads/stores with masked ones 699 // (which could be done), but at the moment, the code here 700 // does not support matching intrinsics with non-intrinsics, 701 // so keep the MatchingIds specific to masked instructions 702 // for now (TODO). 703 Info.MatchingId = Intrinsic::masked_load; 704 Info.ReadMem = false; 705 Info.WriteMem = true; 706 Info.IsVolatile = false; 707 break; 708 } 709 } 710 } 711 } 712 713 Instruction *get() { return Inst; } 714 const Instruction *get() const { return Inst; } 715 716 bool isLoad() const { 717 if (IntrID != 0) 718 return Info.ReadMem; 719 return isa<LoadInst>(Inst); 720 } 721 722 bool isStore() const { 723 if (IntrID != 0) 724 return Info.WriteMem; 725 return isa<StoreInst>(Inst); 726 } 727 728 bool isAtomic() const { 729 if (IntrID != 0) 730 return Info.Ordering != AtomicOrdering::NotAtomic; 731 return Inst->isAtomic(); 732 } 733 734 bool isUnordered() const { 735 if (IntrID != 0) 736 return Info.isUnordered(); 737 738 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 739 return LI->isUnordered(); 740 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 741 return SI->isUnordered(); 742 } 743 // Conservative answer 744 return !Inst->isAtomic(); 745 } 746 747 bool isVolatile() const { 748 if (IntrID != 0) 749 return Info.IsVolatile; 750 751 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 752 return LI->isVolatile(); 753 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 754 return SI->isVolatile(); 755 } 756 // Conservative answer 757 return true; 758 } 759 760 bool isInvariantLoad() const { 761 if (auto *LI = dyn_cast<LoadInst>(Inst)) 762 return LI->hasMetadata(LLVMContext::MD_invariant_load); 763 return false; 764 } 765 766 bool isValid() const { return getPointerOperand() != nullptr; } 767 768 // For regular (non-intrinsic) loads/stores, this is set to -1. For 769 // intrinsic loads/stores, the id is retrieved from the corresponding 770 // field in the MemIntrinsicInfo structure. That field contains 771 // non-negative values only. 772 int getMatchingId() const { 773 if (IntrID != 0) 774 return Info.MatchingId; 775 return -1; 776 } 777 778 Value *getPointerOperand() const { 779 if (IntrID != 0) 780 return Info.PtrVal; 781 return getLoadStorePointerOperand(Inst); 782 } 783 784 bool mayReadFromMemory() const { 785 if (IntrID != 0) 786 return Info.ReadMem; 787 return Inst->mayReadFromMemory(); 788 } 789 790 bool mayWriteToMemory() const { 791 if (IntrID != 0) 792 return Info.WriteMem; 793 return Inst->mayWriteToMemory(); 794 } 795 796 private: 797 Intrinsic::ID IntrID = 0; 798 MemIntrinsicInfo Info; 799 Instruction *Inst; 800 }; 801 802 // This function is to prevent accidentally passing a non-target 803 // intrinsic ID to TargetTransformInfo. 804 static bool isHandledNonTargetIntrinsic(Intrinsic::ID ID) { 805 switch (ID) { 806 case Intrinsic::masked_load: 807 case Intrinsic::masked_store: 808 return true; 809 } 810 return false; 811 } 812 static bool isHandledNonTargetIntrinsic(const Value *V) { 813 if (auto *II = dyn_cast<IntrinsicInst>(V)) 814 return isHandledNonTargetIntrinsic(II->getIntrinsicID()); 815 return false; 816 } 817 818 bool processNode(DomTreeNode *Node); 819 820 bool handleBranchCondition(Instruction *CondInst, const BranchInst *BI, 821 const BasicBlock *BB, const BasicBlock *Pred); 822 823 Value *getMatchingValue(LoadValue &InVal, ParseMemoryInst &MemInst, 824 unsigned CurrentGeneration); 825 826 bool overridingStores(const ParseMemoryInst &Earlier, 827 const ParseMemoryInst &Later); 828 829 Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const { 830 if (auto *LI = dyn_cast<LoadInst>(Inst)) 831 return LI; 832 if (auto *SI = dyn_cast<StoreInst>(Inst)) 833 return SI->getValueOperand(); 834 assert(isa<IntrinsicInst>(Inst) && "Instruction not supported"); 835 auto *II = cast<IntrinsicInst>(Inst); 836 if (isHandledNonTargetIntrinsic(II->getIntrinsicID())) 837 return getOrCreateResultNonTargetMemIntrinsic(II, ExpectedType); 838 return TTI.getOrCreateResultFromMemIntrinsic(II, ExpectedType); 839 } 840 841 Value *getOrCreateResultNonTargetMemIntrinsic(IntrinsicInst *II, 842 Type *ExpectedType) const { 843 switch (II->getIntrinsicID()) { 844 case Intrinsic::masked_load: 845 return II; 846 case Intrinsic::masked_store: 847 return II->getOperand(0); 848 } 849 return nullptr; 850 } 851 852 /// Return true if the instruction is known to only operate on memory 853 /// provably invariant in the given "generation". 854 bool isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt); 855 856 bool isSameMemGeneration(unsigned EarlierGeneration, unsigned LaterGeneration, 857 Instruction *EarlierInst, Instruction *LaterInst); 858 859 bool isNonTargetIntrinsicMatch(const IntrinsicInst *Earlier, 860 const IntrinsicInst *Later) { 861 auto IsSubmask = [](const Value *Mask0, const Value *Mask1) { 862 // Is Mask0 a submask of Mask1? 863 if (Mask0 == Mask1) 864 return true; 865 if (isa<UndefValue>(Mask0) || isa<UndefValue>(Mask1)) 866 return false; 867 auto *Vec0 = dyn_cast<ConstantVector>(Mask0); 868 auto *Vec1 = dyn_cast<ConstantVector>(Mask1); 869 if (!Vec0 || !Vec1) 870 return false; 871 assert(Vec0->getType() == Vec1->getType() && 872 "Masks should have the same type"); 873 for (int i = 0, e = Vec0->getNumOperands(); i != e; ++i) { 874 Constant *Elem0 = Vec0->getOperand(i); 875 Constant *Elem1 = Vec1->getOperand(i); 876 auto *Int0 = dyn_cast<ConstantInt>(Elem0); 877 if (Int0 && Int0->isZero()) 878 continue; 879 auto *Int1 = dyn_cast<ConstantInt>(Elem1); 880 if (Int1 && !Int1->isZero()) 881 continue; 882 if (isa<UndefValue>(Elem0) || isa<UndefValue>(Elem1)) 883 return false; 884 if (Elem0 == Elem1) 885 continue; 886 return false; 887 } 888 return true; 889 }; 890 auto PtrOp = [](const IntrinsicInst *II) { 891 if (II->getIntrinsicID() == Intrinsic::masked_load) 892 return II->getOperand(0); 893 if (II->getIntrinsicID() == Intrinsic::masked_store) 894 return II->getOperand(1); 895 llvm_unreachable("Unexpected IntrinsicInst"); 896 }; 897 auto MaskOp = [](const IntrinsicInst *II) { 898 if (II->getIntrinsicID() == Intrinsic::masked_load) 899 return II->getOperand(2); 900 if (II->getIntrinsicID() == Intrinsic::masked_store) 901 return II->getOperand(3); 902 llvm_unreachable("Unexpected IntrinsicInst"); 903 }; 904 auto ThruOp = [](const IntrinsicInst *II) { 905 if (II->getIntrinsicID() == Intrinsic::masked_load) 906 return II->getOperand(3); 907 llvm_unreachable("Unexpected IntrinsicInst"); 908 }; 909 910 if (PtrOp(Earlier) != PtrOp(Later)) 911 return false; 912 913 Intrinsic::ID IDE = Earlier->getIntrinsicID(); 914 Intrinsic::ID IDL = Later->getIntrinsicID(); 915 // We could really use specific intrinsic classes for masked loads 916 // and stores in IntrinsicInst.h. 917 if (IDE == Intrinsic::masked_load && IDL == Intrinsic::masked_load) { 918 // Trying to replace later masked load with the earlier one. 919 // Check that the pointers are the same, and 920 // - masks and pass-throughs are the same, or 921 // - replacee's pass-through is "undef" and replacer's mask is a 922 // super-set of the replacee's mask. 923 if (MaskOp(Earlier) == MaskOp(Later) && ThruOp(Earlier) == ThruOp(Later)) 924 return true; 925 if (!isa<UndefValue>(ThruOp(Later))) 926 return false; 927 return IsSubmask(MaskOp(Later), MaskOp(Earlier)); 928 } 929 if (IDE == Intrinsic::masked_store && IDL == Intrinsic::masked_load) { 930 // Trying to replace a load of a stored value with the store's value. 931 // Check that the pointers are the same, and 932 // - load's mask is a subset of store's mask, and 933 // - load's pass-through is "undef". 934 if (!IsSubmask(MaskOp(Later), MaskOp(Earlier))) 935 return false; 936 return isa<UndefValue>(ThruOp(Later)); 937 } 938 if (IDE == Intrinsic::masked_load && IDL == Intrinsic::masked_store) { 939 // Trying to remove a store of the loaded value. 940 // Check that the pointers are the same, and 941 // - store's mask is a subset of the load's mask. 942 return IsSubmask(MaskOp(Later), MaskOp(Earlier)); 943 } 944 if (IDE == Intrinsic::masked_store && IDL == Intrinsic::masked_store) { 945 // Trying to remove a dead store (earlier). 946 // Check that the pointers are the same, 947 // - the to-be-removed store's mask is a subset of the other store's 948 // mask. 949 return IsSubmask(MaskOp(Earlier), MaskOp(Later)); 950 } 951 return false; 952 } 953 954 void removeMSSA(Instruction &Inst) { 955 if (!MSSA) 956 return; 957 if (VerifyMemorySSA) 958 MSSA->verifyMemorySSA(); 959 // Removing a store here can leave MemorySSA in an unoptimized state by 960 // creating MemoryPhis that have identical arguments and by creating 961 // MemoryUses whose defining access is not an actual clobber. The phi case 962 // is handled by MemorySSA when passing OptimizePhis = true to 963 // removeMemoryAccess. The non-optimized MemoryUse case is lazily updated 964 // by MemorySSA's getClobberingMemoryAccess. 965 MSSAUpdater->removeMemoryAccess(&Inst, true); 966 } 967 }; 968 969 } // end anonymous namespace 970 971 /// Determine if the memory referenced by LaterInst is from the same heap 972 /// version as EarlierInst. 973 /// This is currently called in two scenarios: 974 /// 975 /// load p 976 /// ... 977 /// load p 978 /// 979 /// and 980 /// 981 /// x = load p 982 /// ... 983 /// store x, p 984 /// 985 /// in both cases we want to verify that there are no possible writes to the 986 /// memory referenced by p between the earlier and later instruction. 987 bool EarlyCSE::isSameMemGeneration(unsigned EarlierGeneration, 988 unsigned LaterGeneration, 989 Instruction *EarlierInst, 990 Instruction *LaterInst) { 991 // Check the simple memory generation tracking first. 992 if (EarlierGeneration == LaterGeneration) 993 return true; 994 995 if (!MSSA) 996 return false; 997 998 // If MemorySSA has determined that one of EarlierInst or LaterInst does not 999 // read/write memory, then we can safely return true here. 1000 // FIXME: We could be more aggressive when checking doesNotAccessMemory(), 1001 // onlyReadsMemory(), mayReadFromMemory(), and mayWriteToMemory() in this pass 1002 // by also checking the MemorySSA MemoryAccess on the instruction. Initial 1003 // experiments suggest this isn't worthwhile, at least for C/C++ code compiled 1004 // with the default optimization pipeline. 1005 auto *EarlierMA = MSSA->getMemoryAccess(EarlierInst); 1006 if (!EarlierMA) 1007 return true; 1008 auto *LaterMA = MSSA->getMemoryAccess(LaterInst); 1009 if (!LaterMA) 1010 return true; 1011 1012 // Since we know LaterDef dominates LaterInst and EarlierInst dominates 1013 // LaterInst, if LaterDef dominates EarlierInst then it can't occur between 1014 // EarlierInst and LaterInst and neither can any other write that potentially 1015 // clobbers LaterInst. 1016 MemoryAccess *LaterDef; 1017 if (ClobberCounter < EarlyCSEMssaOptCap) { 1018 LaterDef = MSSA->getWalker()->getClobberingMemoryAccess(LaterInst); 1019 ClobberCounter++; 1020 } else 1021 LaterDef = LaterMA->getDefiningAccess(); 1022 1023 return MSSA->dominates(LaterDef, EarlierMA); 1024 } 1025 1026 bool EarlyCSE::isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt) { 1027 // A location loaded from with an invariant_load is assumed to *never* change 1028 // within the visible scope of the compilation. 1029 if (auto *LI = dyn_cast<LoadInst>(I)) 1030 if (LI->hasMetadata(LLVMContext::MD_invariant_load)) 1031 return true; 1032 1033 auto MemLocOpt = MemoryLocation::getOrNone(I); 1034 if (!MemLocOpt) 1035 // "target" intrinsic forms of loads aren't currently known to 1036 // MemoryLocation::get. TODO 1037 return false; 1038 MemoryLocation MemLoc = *MemLocOpt; 1039 if (!AvailableInvariants.count(MemLoc)) 1040 return false; 1041 1042 // Is the generation at which this became invariant older than the 1043 // current one? 1044 return AvailableInvariants.lookup(MemLoc) <= GenAt; 1045 } 1046 1047 bool EarlyCSE::handleBranchCondition(Instruction *CondInst, 1048 const BranchInst *BI, const BasicBlock *BB, 1049 const BasicBlock *Pred) { 1050 assert(BI->isConditional() && "Should be a conditional branch!"); 1051 assert(BI->getCondition() == CondInst && "Wrong condition?"); 1052 assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB); 1053 auto *TorF = (BI->getSuccessor(0) == BB) 1054 ? ConstantInt::getTrue(BB->getContext()) 1055 : ConstantInt::getFalse(BB->getContext()); 1056 auto MatchBinOp = [](Instruction *I, unsigned Opcode, Value *&LHS, 1057 Value *&RHS) { 1058 if (Opcode == Instruction::And && 1059 match(I, m_LogicalAnd(m_Value(LHS), m_Value(RHS)))) 1060 return true; 1061 else if (Opcode == Instruction::Or && 1062 match(I, m_LogicalOr(m_Value(LHS), m_Value(RHS)))) 1063 return true; 1064 return false; 1065 }; 1066 // If the condition is AND operation, we can propagate its operands into the 1067 // true branch. If it is OR operation, we can propagate them into the false 1068 // branch. 1069 unsigned PropagateOpcode = 1070 (BI->getSuccessor(0) == BB) ? Instruction::And : Instruction::Or; 1071 1072 bool MadeChanges = false; 1073 SmallVector<Instruction *, 4> WorkList; 1074 SmallPtrSet<Instruction *, 4> Visited; 1075 WorkList.push_back(CondInst); 1076 while (!WorkList.empty()) { 1077 Instruction *Curr = WorkList.pop_back_val(); 1078 1079 AvailableValues.insert(Curr, TorF); 1080 LLVM_DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '" 1081 << Curr->getName() << "' as " << *TorF << " in " 1082 << BB->getName() << "\n"); 1083 if (!DebugCounter::shouldExecute(CSECounter)) { 1084 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1085 } else { 1086 // Replace all dominated uses with the known value. 1087 if (unsigned Count = replaceDominatedUsesWith(Curr, TorF, DT, 1088 BasicBlockEdge(Pred, BB))) { 1089 NumCSECVP += Count; 1090 MadeChanges = true; 1091 } 1092 } 1093 1094 Value *LHS, *RHS; 1095 if (MatchBinOp(Curr, PropagateOpcode, LHS, RHS)) 1096 for (auto &Op : { LHS, RHS }) 1097 if (Instruction *OPI = dyn_cast<Instruction>(Op)) 1098 if (SimpleValue::canHandle(OPI) && Visited.insert(OPI).second) 1099 WorkList.push_back(OPI); 1100 } 1101 1102 return MadeChanges; 1103 } 1104 1105 Value *EarlyCSE::getMatchingValue(LoadValue &InVal, ParseMemoryInst &MemInst, 1106 unsigned CurrentGeneration) { 1107 if (InVal.DefInst == nullptr) 1108 return nullptr; 1109 if (InVal.MatchingId != MemInst.getMatchingId()) 1110 return nullptr; 1111 // We don't yet handle removing loads with ordering of any kind. 1112 if (MemInst.isVolatile() || !MemInst.isUnordered()) 1113 return nullptr; 1114 // We can't replace an atomic load with one which isn't also atomic. 1115 if (MemInst.isLoad() && !InVal.IsAtomic && MemInst.isAtomic()) 1116 return nullptr; 1117 // The value V returned from this function is used differently depending 1118 // on whether MemInst is a load or a store. If it's a load, we will replace 1119 // MemInst with V, if it's a store, we will check if V is the same as the 1120 // available value. 1121 bool MemInstMatching = !MemInst.isLoad(); 1122 Instruction *Matching = MemInstMatching ? MemInst.get() : InVal.DefInst; 1123 Instruction *Other = MemInstMatching ? InVal.DefInst : MemInst.get(); 1124 1125 // For stores check the result values before checking memory generation 1126 // (otherwise isSameMemGeneration may crash). 1127 Value *Result = MemInst.isStore() 1128 ? getOrCreateResult(Matching, Other->getType()) 1129 : nullptr; 1130 if (MemInst.isStore() && InVal.DefInst != Result) 1131 return nullptr; 1132 1133 // Deal with non-target memory intrinsics. 1134 bool MatchingNTI = isHandledNonTargetIntrinsic(Matching); 1135 bool OtherNTI = isHandledNonTargetIntrinsic(Other); 1136 if (OtherNTI != MatchingNTI) 1137 return nullptr; 1138 if (OtherNTI && MatchingNTI) { 1139 if (!isNonTargetIntrinsicMatch(cast<IntrinsicInst>(InVal.DefInst), 1140 cast<IntrinsicInst>(MemInst.get()))) 1141 return nullptr; 1142 } 1143 1144 if (!isOperatingOnInvariantMemAt(MemInst.get(), InVal.Generation) && 1145 !isSameMemGeneration(InVal.Generation, CurrentGeneration, InVal.DefInst, 1146 MemInst.get())) 1147 return nullptr; 1148 1149 if (!Result) 1150 Result = getOrCreateResult(Matching, Other->getType()); 1151 return Result; 1152 } 1153 1154 bool EarlyCSE::overridingStores(const ParseMemoryInst &Earlier, 1155 const ParseMemoryInst &Later) { 1156 // Can we remove Earlier store because of Later store? 1157 1158 assert(Earlier.isUnordered() && !Earlier.isVolatile() && 1159 "Violated invariant"); 1160 if (Earlier.getPointerOperand() != Later.getPointerOperand()) 1161 return false; 1162 if (Earlier.getMatchingId() != Later.getMatchingId()) 1163 return false; 1164 // At the moment, we don't remove ordered stores, but do remove 1165 // unordered atomic stores. There's no special requirement (for 1166 // unordered atomics) about removing atomic stores only in favor of 1167 // other atomic stores since we were going to execute the non-atomic 1168 // one anyway and the atomic one might never have become visible. 1169 if (!Earlier.isUnordered() || !Later.isUnordered()) 1170 return false; 1171 1172 // Deal with non-target memory intrinsics. 1173 bool ENTI = isHandledNonTargetIntrinsic(Earlier.get()); 1174 bool LNTI = isHandledNonTargetIntrinsic(Later.get()); 1175 if (ENTI && LNTI) 1176 return isNonTargetIntrinsicMatch(cast<IntrinsicInst>(Earlier.get()), 1177 cast<IntrinsicInst>(Later.get())); 1178 1179 // Because of the check above, at least one of them is false. 1180 // For now disallow matching intrinsics with non-intrinsics, 1181 // so assume that the stores match if neither is an intrinsic. 1182 return ENTI == LNTI; 1183 } 1184 1185 bool EarlyCSE::processNode(DomTreeNode *Node) { 1186 bool Changed = false; 1187 BasicBlock *BB = Node->getBlock(); 1188 1189 // If this block has a single predecessor, then the predecessor is the parent 1190 // of the domtree node and all of the live out memory values are still current 1191 // in this block. If this block has multiple predecessors, then they could 1192 // have invalidated the live-out memory values of our parent value. For now, 1193 // just be conservative and invalidate memory if this block has multiple 1194 // predecessors. 1195 if (!BB->getSinglePredecessor()) 1196 ++CurrentGeneration; 1197 1198 // If this node has a single predecessor which ends in a conditional branch, 1199 // we can infer the value of the branch condition given that we took this 1200 // path. We need the single predecessor to ensure there's not another path 1201 // which reaches this block where the condition might hold a different 1202 // value. Since we're adding this to the scoped hash table (like any other 1203 // def), it will have been popped if we encounter a future merge block. 1204 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 1205 auto *BI = dyn_cast<BranchInst>(Pred->getTerminator()); 1206 if (BI && BI->isConditional()) { 1207 auto *CondInst = dyn_cast<Instruction>(BI->getCondition()); 1208 if (CondInst && SimpleValue::canHandle(CondInst)) 1209 Changed |= handleBranchCondition(CondInst, BI, BB, Pred); 1210 } 1211 } 1212 1213 /// LastStore - Keep track of the last non-volatile store that we saw... for 1214 /// as long as there in no instruction that reads memory. If we see a store 1215 /// to the same location, we delete the dead store. This zaps trivial dead 1216 /// stores which can occur in bitfield code among other things. 1217 Instruction *LastStore = nullptr; 1218 1219 // See if any instructions in the block can be eliminated. If so, do it. If 1220 // not, add them to AvailableValues. 1221 for (Instruction &Inst : make_early_inc_range(BB->getInstList())) { 1222 // Dead instructions should just be removed. 1223 if (isInstructionTriviallyDead(&Inst, &TLI)) { 1224 LLVM_DEBUG(dbgs() << "EarlyCSE DCE: " << Inst << '\n'); 1225 if (!DebugCounter::shouldExecute(CSECounter)) { 1226 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1227 continue; 1228 } 1229 1230 salvageKnowledge(&Inst, &AC); 1231 salvageDebugInfo(Inst); 1232 removeMSSA(Inst); 1233 Inst.eraseFromParent(); 1234 Changed = true; 1235 ++NumSimplify; 1236 continue; 1237 } 1238 1239 // Skip assume intrinsics, they don't really have side effects (although 1240 // they're marked as such to ensure preservation of control dependencies), 1241 // and this pass will not bother with its removal. However, we should mark 1242 // its condition as true for all dominated blocks. 1243 if (auto *Assume = dyn_cast<AssumeInst>(&Inst)) { 1244 auto *CondI = dyn_cast<Instruction>(Assume->getArgOperand(0)); 1245 if (CondI && SimpleValue::canHandle(CondI)) { 1246 LLVM_DEBUG(dbgs() << "EarlyCSE considering assumption: " << Inst 1247 << '\n'); 1248 AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext())); 1249 } else 1250 LLVM_DEBUG(dbgs() << "EarlyCSE skipping assumption: " << Inst << '\n'); 1251 continue; 1252 } 1253 1254 // Likewise, noalias intrinsics don't actually write. 1255 if (match(&Inst, 1256 m_Intrinsic<Intrinsic::experimental_noalias_scope_decl>())) { 1257 LLVM_DEBUG(dbgs() << "EarlyCSE skipping noalias intrinsic: " << Inst 1258 << '\n'); 1259 continue; 1260 } 1261 1262 // Skip sideeffect intrinsics, for the same reason as assume intrinsics. 1263 if (match(&Inst, m_Intrinsic<Intrinsic::sideeffect>())) { 1264 LLVM_DEBUG(dbgs() << "EarlyCSE skipping sideeffect: " << Inst << '\n'); 1265 continue; 1266 } 1267 1268 // We can skip all invariant.start intrinsics since they only read memory, 1269 // and we can forward values across it. For invariant starts without 1270 // invariant ends, we can use the fact that the invariantness never ends to 1271 // start a scope in the current generaton which is true for all future 1272 // generations. Also, we dont need to consume the last store since the 1273 // semantics of invariant.start allow us to perform DSE of the last 1274 // store, if there was a store following invariant.start. Consider: 1275 // 1276 // store 30, i8* p 1277 // invariant.start(p) 1278 // store 40, i8* p 1279 // We can DSE the store to 30, since the store 40 to invariant location p 1280 // causes undefined behaviour. 1281 if (match(&Inst, m_Intrinsic<Intrinsic::invariant_start>())) { 1282 // If there are any uses, the scope might end. 1283 if (!Inst.use_empty()) 1284 continue; 1285 MemoryLocation MemLoc = 1286 MemoryLocation::getForArgument(&cast<CallInst>(Inst), 1, TLI); 1287 // Don't start a scope if we already have a better one pushed 1288 if (!AvailableInvariants.count(MemLoc)) 1289 AvailableInvariants.insert(MemLoc, CurrentGeneration); 1290 continue; 1291 } 1292 1293 if (isGuard(&Inst)) { 1294 if (auto *CondI = 1295 dyn_cast<Instruction>(cast<CallInst>(Inst).getArgOperand(0))) { 1296 if (SimpleValue::canHandle(CondI)) { 1297 // Do we already know the actual value of this condition? 1298 if (auto *KnownCond = AvailableValues.lookup(CondI)) { 1299 // Is the condition known to be true? 1300 if (isa<ConstantInt>(KnownCond) && 1301 cast<ConstantInt>(KnownCond)->isOne()) { 1302 LLVM_DEBUG(dbgs() 1303 << "EarlyCSE removing guard: " << Inst << '\n'); 1304 salvageKnowledge(&Inst, &AC); 1305 removeMSSA(Inst); 1306 Inst.eraseFromParent(); 1307 Changed = true; 1308 continue; 1309 } else 1310 // Use the known value if it wasn't true. 1311 cast<CallInst>(Inst).setArgOperand(0, KnownCond); 1312 } 1313 // The condition we're on guarding here is true for all dominated 1314 // locations. 1315 AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext())); 1316 } 1317 } 1318 1319 // Guard intrinsics read all memory, but don't write any memory. 1320 // Accordingly, don't update the generation but consume the last store (to 1321 // avoid an incorrect DSE). 1322 LastStore = nullptr; 1323 continue; 1324 } 1325 1326 // If the instruction can be simplified (e.g. X+0 = X) then replace it with 1327 // its simpler value. 1328 if (Value *V = SimplifyInstruction(&Inst, SQ)) { 1329 LLVM_DEBUG(dbgs() << "EarlyCSE Simplify: " << Inst << " to: " << *V 1330 << '\n'); 1331 if (!DebugCounter::shouldExecute(CSECounter)) { 1332 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1333 } else { 1334 bool Killed = false; 1335 if (!Inst.use_empty()) { 1336 Inst.replaceAllUsesWith(V); 1337 Changed = true; 1338 } 1339 if (isInstructionTriviallyDead(&Inst, &TLI)) { 1340 salvageKnowledge(&Inst, &AC); 1341 removeMSSA(Inst); 1342 Inst.eraseFromParent(); 1343 Changed = true; 1344 Killed = true; 1345 } 1346 if (Changed) 1347 ++NumSimplify; 1348 if (Killed) 1349 continue; 1350 } 1351 } 1352 1353 // If this is a simple instruction that we can value number, process it. 1354 if (SimpleValue::canHandle(&Inst)) { 1355 // See if the instruction has an available value. If so, use it. 1356 if (Value *V = AvailableValues.lookup(&Inst)) { 1357 LLVM_DEBUG(dbgs() << "EarlyCSE CSE: " << Inst << " to: " << *V 1358 << '\n'); 1359 if (!DebugCounter::shouldExecute(CSECounter)) { 1360 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1361 continue; 1362 } 1363 if (auto *I = dyn_cast<Instruction>(V)) 1364 I->andIRFlags(&Inst); 1365 Inst.replaceAllUsesWith(V); 1366 salvageKnowledge(&Inst, &AC); 1367 removeMSSA(Inst); 1368 Inst.eraseFromParent(); 1369 Changed = true; 1370 ++NumCSE; 1371 continue; 1372 } 1373 1374 // Otherwise, just remember that this value is available. 1375 AvailableValues.insert(&Inst, &Inst); 1376 continue; 1377 } 1378 1379 ParseMemoryInst MemInst(&Inst, TTI); 1380 // If this is a non-volatile load, process it. 1381 if (MemInst.isValid() && MemInst.isLoad()) { 1382 // (conservatively) we can't peak past the ordering implied by this 1383 // operation, but we can add this load to our set of available values 1384 if (MemInst.isVolatile() || !MemInst.isUnordered()) { 1385 LastStore = nullptr; 1386 ++CurrentGeneration; 1387 } 1388 1389 if (MemInst.isInvariantLoad()) { 1390 // If we pass an invariant load, we know that memory location is 1391 // indefinitely constant from the moment of first dereferenceability. 1392 // We conservatively treat the invariant_load as that moment. If we 1393 // pass a invariant load after already establishing a scope, don't 1394 // restart it since we want to preserve the earliest point seen. 1395 auto MemLoc = MemoryLocation::get(&Inst); 1396 if (!AvailableInvariants.count(MemLoc)) 1397 AvailableInvariants.insert(MemLoc, CurrentGeneration); 1398 } 1399 1400 // If we have an available version of this load, and if it is the right 1401 // generation or the load is known to be from an invariant location, 1402 // replace this instruction. 1403 // 1404 // If either the dominating load or the current load are invariant, then 1405 // we can assume the current load loads the same value as the dominating 1406 // load. 1407 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand()); 1408 if (Value *Op = getMatchingValue(InVal, MemInst, CurrentGeneration)) { 1409 LLVM_DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << Inst 1410 << " to: " << *InVal.DefInst << '\n'); 1411 if (!DebugCounter::shouldExecute(CSECounter)) { 1412 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1413 continue; 1414 } 1415 if (!Inst.use_empty()) 1416 Inst.replaceAllUsesWith(Op); 1417 salvageKnowledge(&Inst, &AC); 1418 removeMSSA(Inst); 1419 Inst.eraseFromParent(); 1420 Changed = true; 1421 ++NumCSELoad; 1422 continue; 1423 } 1424 1425 // Otherwise, remember that we have this instruction. 1426 AvailableLoads.insert(MemInst.getPointerOperand(), 1427 LoadValue(&Inst, CurrentGeneration, 1428 MemInst.getMatchingId(), 1429 MemInst.isAtomic())); 1430 LastStore = nullptr; 1431 continue; 1432 } 1433 1434 // If this instruction may read from memory or throw (and potentially read 1435 // from memory in the exception handler), forget LastStore. Load/store 1436 // intrinsics will indicate both a read and a write to memory. The target 1437 // may override this (e.g. so that a store intrinsic does not read from 1438 // memory, and thus will be treated the same as a regular store for 1439 // commoning purposes). 1440 if ((Inst.mayReadFromMemory() || Inst.mayThrow()) && 1441 !(MemInst.isValid() && !MemInst.mayReadFromMemory())) 1442 LastStore = nullptr; 1443 1444 // If this is a read-only call, process it. 1445 if (CallValue::canHandle(&Inst)) { 1446 // If we have an available version of this call, and if it is the right 1447 // generation, replace this instruction. 1448 std::pair<Instruction *, unsigned> InVal = AvailableCalls.lookup(&Inst); 1449 if (InVal.first != nullptr && 1450 isSameMemGeneration(InVal.second, CurrentGeneration, InVal.first, 1451 &Inst)) { 1452 LLVM_DEBUG(dbgs() << "EarlyCSE CSE CALL: " << Inst 1453 << " to: " << *InVal.first << '\n'); 1454 if (!DebugCounter::shouldExecute(CSECounter)) { 1455 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1456 continue; 1457 } 1458 if (!Inst.use_empty()) 1459 Inst.replaceAllUsesWith(InVal.first); 1460 salvageKnowledge(&Inst, &AC); 1461 removeMSSA(Inst); 1462 Inst.eraseFromParent(); 1463 Changed = true; 1464 ++NumCSECall; 1465 continue; 1466 } 1467 1468 // Otherwise, remember that we have this instruction. 1469 AvailableCalls.insert(&Inst, std::make_pair(&Inst, CurrentGeneration)); 1470 continue; 1471 } 1472 1473 // A release fence requires that all stores complete before it, but does 1474 // not prevent the reordering of following loads 'before' the fence. As a 1475 // result, we don't need to consider it as writing to memory and don't need 1476 // to advance the generation. We do need to prevent DSE across the fence, 1477 // but that's handled above. 1478 if (auto *FI = dyn_cast<FenceInst>(&Inst)) 1479 if (FI->getOrdering() == AtomicOrdering::Release) { 1480 assert(Inst.mayReadFromMemory() && "relied on to prevent DSE above"); 1481 continue; 1482 } 1483 1484 // write back DSE - If we write back the same value we just loaded from 1485 // the same location and haven't passed any intervening writes or ordering 1486 // operations, we can remove the write. The primary benefit is in allowing 1487 // the available load table to remain valid and value forward past where 1488 // the store originally was. 1489 if (MemInst.isValid() && MemInst.isStore()) { 1490 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand()); 1491 if (InVal.DefInst && 1492 InVal.DefInst == getMatchingValue(InVal, MemInst, CurrentGeneration)) { 1493 // It is okay to have a LastStore to a different pointer here if MemorySSA 1494 // tells us that the load and store are from the same memory generation. 1495 // In that case, LastStore should keep its present value since we're 1496 // removing the current store. 1497 assert((!LastStore || 1498 ParseMemoryInst(LastStore, TTI).getPointerOperand() == 1499 MemInst.getPointerOperand() || 1500 MSSA) && 1501 "can't have an intervening store if not using MemorySSA!"); 1502 LLVM_DEBUG(dbgs() << "EarlyCSE DSE (writeback): " << Inst << '\n'); 1503 if (!DebugCounter::shouldExecute(CSECounter)) { 1504 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1505 continue; 1506 } 1507 salvageKnowledge(&Inst, &AC); 1508 removeMSSA(Inst); 1509 Inst.eraseFromParent(); 1510 Changed = true; 1511 ++NumDSE; 1512 // We can avoid incrementing the generation count since we were able 1513 // to eliminate this store. 1514 continue; 1515 } 1516 } 1517 1518 // Okay, this isn't something we can CSE at all. Check to see if it is 1519 // something that could modify memory. If so, our available memory values 1520 // cannot be used so bump the generation count. 1521 if (Inst.mayWriteToMemory()) { 1522 ++CurrentGeneration; 1523 1524 if (MemInst.isValid() && MemInst.isStore()) { 1525 // We do a trivial form of DSE if there are two stores to the same 1526 // location with no intervening loads. Delete the earlier store. 1527 if (LastStore) { 1528 if (overridingStores(ParseMemoryInst(LastStore, TTI), MemInst)) { 1529 LLVM_DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore 1530 << " due to: " << Inst << '\n'); 1531 if (!DebugCounter::shouldExecute(CSECounter)) { 1532 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1533 } else { 1534 salvageKnowledge(&Inst, &AC); 1535 removeMSSA(*LastStore); 1536 LastStore->eraseFromParent(); 1537 Changed = true; 1538 ++NumDSE; 1539 LastStore = nullptr; 1540 } 1541 } 1542 // fallthrough - we can exploit information about this store 1543 } 1544 1545 // Okay, we just invalidated anything we knew about loaded values. Try 1546 // to salvage *something* by remembering that the stored value is a live 1547 // version of the pointer. It is safe to forward from volatile stores 1548 // to non-volatile loads, so we don't have to check for volatility of 1549 // the store. 1550 AvailableLoads.insert(MemInst.getPointerOperand(), 1551 LoadValue(&Inst, CurrentGeneration, 1552 MemInst.getMatchingId(), 1553 MemInst.isAtomic())); 1554 1555 // Remember that this was the last unordered store we saw for DSE. We 1556 // don't yet handle DSE on ordered or volatile stores since we don't 1557 // have a good way to model the ordering requirement for following 1558 // passes once the store is removed. We could insert a fence, but 1559 // since fences are slightly stronger than stores in their ordering, 1560 // it's not clear this is a profitable transform. Another option would 1561 // be to merge the ordering with that of the post dominating store. 1562 if (MemInst.isUnordered() && !MemInst.isVolatile()) 1563 LastStore = &Inst; 1564 else 1565 LastStore = nullptr; 1566 } 1567 } 1568 } 1569 1570 return Changed; 1571 } 1572 1573 bool EarlyCSE::run() { 1574 // Note, deque is being used here because there is significant performance 1575 // gains over vector when the container becomes very large due to the 1576 // specific access patterns. For more information see the mailing list 1577 // discussion on this: 1578 // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html 1579 std::deque<StackNode *> nodesToProcess; 1580 1581 bool Changed = false; 1582 1583 // Process the root node. 1584 nodesToProcess.push_back(new StackNode( 1585 AvailableValues, AvailableLoads, AvailableInvariants, AvailableCalls, 1586 CurrentGeneration, DT.getRootNode(), 1587 DT.getRootNode()->begin(), DT.getRootNode()->end())); 1588 1589 assert(!CurrentGeneration && "Create a new EarlyCSE instance to rerun it."); 1590 1591 // Process the stack. 1592 while (!nodesToProcess.empty()) { 1593 // Grab the first item off the stack. Set the current generation, remove 1594 // the node from the stack, and process it. 1595 StackNode *NodeToProcess = nodesToProcess.back(); 1596 1597 // Initialize class members. 1598 CurrentGeneration = NodeToProcess->currentGeneration(); 1599 1600 // Check if the node needs to be processed. 1601 if (!NodeToProcess->isProcessed()) { 1602 // Process the node. 1603 Changed |= processNode(NodeToProcess->node()); 1604 NodeToProcess->childGeneration(CurrentGeneration); 1605 NodeToProcess->process(); 1606 } else if (NodeToProcess->childIter() != NodeToProcess->end()) { 1607 // Push the next child onto the stack. 1608 DomTreeNode *child = NodeToProcess->nextChild(); 1609 nodesToProcess.push_back( 1610 new StackNode(AvailableValues, AvailableLoads, AvailableInvariants, 1611 AvailableCalls, NodeToProcess->childGeneration(), 1612 child, child->begin(), child->end())); 1613 } else { 1614 // It has been processed, and there are no more children to process, 1615 // so delete it and pop it off the stack. 1616 delete NodeToProcess; 1617 nodesToProcess.pop_back(); 1618 } 1619 } // while (!nodes...) 1620 1621 return Changed; 1622 } 1623 1624 PreservedAnalyses EarlyCSEPass::run(Function &F, 1625 FunctionAnalysisManager &AM) { 1626 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1627 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 1628 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 1629 auto &AC = AM.getResult<AssumptionAnalysis>(F); 1630 auto *MSSA = 1631 UseMemorySSA ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA() : nullptr; 1632 1633 EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA); 1634 1635 if (!CSE.run()) 1636 return PreservedAnalyses::all(); 1637 1638 PreservedAnalyses PA; 1639 PA.preserveSet<CFGAnalyses>(); 1640 if (UseMemorySSA) 1641 PA.preserve<MemorySSAAnalysis>(); 1642 return PA; 1643 } 1644 1645 namespace { 1646 1647 /// A simple and fast domtree-based CSE pass. 1648 /// 1649 /// This pass does a simple depth-first walk over the dominator tree, 1650 /// eliminating trivially redundant instructions and using instsimplify to 1651 /// canonicalize things as it goes. It is intended to be fast and catch obvious 1652 /// cases so that instcombine and other passes are more effective. It is 1653 /// expected that a later pass of GVN will catch the interesting/hard cases. 1654 template<bool UseMemorySSA> 1655 class EarlyCSELegacyCommonPass : public FunctionPass { 1656 public: 1657 static char ID; 1658 1659 EarlyCSELegacyCommonPass() : FunctionPass(ID) { 1660 if (UseMemorySSA) 1661 initializeEarlyCSEMemSSALegacyPassPass(*PassRegistry::getPassRegistry()); 1662 else 1663 initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry()); 1664 } 1665 1666 bool runOnFunction(Function &F) override { 1667 if (skipFunction(F)) 1668 return false; 1669 1670 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1671 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 1672 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1673 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 1674 auto *MSSA = 1675 UseMemorySSA ? &getAnalysis<MemorySSAWrapperPass>().getMSSA() : nullptr; 1676 1677 EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA); 1678 1679 return CSE.run(); 1680 } 1681 1682 void getAnalysisUsage(AnalysisUsage &AU) const override { 1683 AU.addRequired<AssumptionCacheTracker>(); 1684 AU.addRequired<DominatorTreeWrapperPass>(); 1685 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1686 AU.addRequired<TargetTransformInfoWrapperPass>(); 1687 if (UseMemorySSA) { 1688 AU.addRequired<AAResultsWrapperPass>(); 1689 AU.addRequired<MemorySSAWrapperPass>(); 1690 AU.addPreserved<MemorySSAWrapperPass>(); 1691 } 1692 AU.addPreserved<GlobalsAAWrapperPass>(); 1693 AU.addPreserved<AAResultsWrapperPass>(); 1694 AU.setPreservesCFG(); 1695 } 1696 }; 1697 1698 } // end anonymous namespace 1699 1700 using EarlyCSELegacyPass = EarlyCSELegacyCommonPass</*UseMemorySSA=*/false>; 1701 1702 template<> 1703 char EarlyCSELegacyPass::ID = 0; 1704 1705 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false, 1706 false) 1707 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 1708 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 1709 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 1710 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1711 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false) 1712 1713 using EarlyCSEMemSSALegacyPass = 1714 EarlyCSELegacyCommonPass</*UseMemorySSA=*/true>; 1715 1716 template<> 1717 char EarlyCSEMemSSALegacyPass::ID = 0; 1718 1719 FunctionPass *llvm::createEarlyCSEPass(bool UseMemorySSA) { 1720 if (UseMemorySSA) 1721 return new EarlyCSEMemSSALegacyPass(); 1722 else 1723 return new EarlyCSELegacyPass(); 1724 } 1725 1726 INITIALIZE_PASS_BEGIN(EarlyCSEMemSSALegacyPass, "early-cse-memssa", 1727 "Early CSE w/ MemorySSA", false, false) 1728 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 1729 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 1730 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 1731 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 1732 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1733 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) 1734 INITIALIZE_PASS_END(EarlyCSEMemSSALegacyPass, "early-cse-memssa", 1735 "Early CSE w/ MemorySSA", false, false) 1736