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/SmallVector.h" 20 #include "llvm/ADT/Statistic.h" 21 #include "llvm/Analysis/AssumptionCache.h" 22 #include "llvm/Analysis/GlobalsModRef.h" 23 #include "llvm/Analysis/GuardUtils.h" 24 #include "llvm/Analysis/InstructionSimplify.h" 25 #include "llvm/Analysis/MemorySSA.h" 26 #include "llvm/Analysis/MemorySSAUpdater.h" 27 #include "llvm/Analysis/TargetLibraryInfo.h" 28 #include "llvm/Analysis/TargetTransformInfo.h" 29 #include "llvm/Analysis/ValueTracking.h" 30 #include "llvm/IR/BasicBlock.h" 31 #include "llvm/IR/Constants.h" 32 #include "llvm/IR/Dominators.h" 33 #include "llvm/IR/Function.h" 34 #include "llvm/IR/InstrTypes.h" 35 #include "llvm/IR/Instruction.h" 36 #include "llvm/IR/Instructions.h" 37 #include "llvm/IR/IntrinsicInst.h" 38 #include "llvm/IR/LLVMContext.h" 39 #include "llvm/IR/PassManager.h" 40 #include "llvm/IR/PatternMatch.h" 41 #include "llvm/IR/Type.h" 42 #include "llvm/IR/Value.h" 43 #include "llvm/InitializePasses.h" 44 #include "llvm/Pass.h" 45 #include "llvm/Support/Allocator.h" 46 #include "llvm/Support/AtomicOrdering.h" 47 #include "llvm/Support/Casting.h" 48 #include "llvm/Support/Debug.h" 49 #include "llvm/Support/DebugCounter.h" 50 #include "llvm/Support/RecyclingAllocator.h" 51 #include "llvm/Support/raw_ostream.h" 52 #include "llvm/Transforms/Scalar.h" 53 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h" 54 #include "llvm/Transforms/Utils/Local.h" 55 #include <cassert> 56 #include <deque> 57 #include <memory> 58 #include <utility> 59 60 using namespace llvm; 61 using namespace llvm::PatternMatch; 62 63 #define DEBUG_TYPE "early-cse" 64 65 STATISTIC(NumSimplify, "Number of instructions simplified or DCE'd"); 66 STATISTIC(NumCSE, "Number of instructions CSE'd"); 67 STATISTIC(NumCSECVP, "Number of compare instructions CVP'd"); 68 STATISTIC(NumCSELoad, "Number of load instructions CSE'd"); 69 STATISTIC(NumCSECall, "Number of call instructions CSE'd"); 70 STATISTIC(NumDSE, "Number of trivial dead stores removed"); 71 72 DEBUG_COUNTER(CSECounter, "early-cse", 73 "Controls which instructions are removed"); 74 75 static cl::opt<unsigned> EarlyCSEMssaOptCap( 76 "earlycse-mssa-optimization-cap", cl::init(500), cl::Hidden, 77 cl::desc("Enable imprecision in EarlyCSE in pathological cases, in exchange " 78 "for faster compile. Caps the MemorySSA clobbering calls.")); 79 80 static cl::opt<bool> EarlyCSEDebugHash( 81 "earlycse-debug-hash", cl::init(false), cl::Hidden, 82 cl::desc("Perform extra assertion checking to verify that SimpleValue's hash " 83 "function is well-behaved w.r.t. its isEqual predicate")); 84 85 //===----------------------------------------------------------------------===// 86 // SimpleValue 87 //===----------------------------------------------------------------------===// 88 89 namespace { 90 91 /// Struct representing the available values in the scoped hash table. 92 struct SimpleValue { 93 Instruction *Inst; 94 95 SimpleValue(Instruction *I) : Inst(I) { 96 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!"); 97 } 98 99 bool isSentinel() const { 100 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() || 101 Inst == DenseMapInfo<Instruction *>::getTombstoneKey(); 102 } 103 104 static bool canHandle(Instruction *Inst) { 105 // This can only handle non-void readnone functions. 106 // Also handled are constrained intrinsic that look like the types 107 // of instruction handled below (UnaryOperator, etc.). 108 if (CallInst *CI = dyn_cast<CallInst>(Inst)) { 109 if (Function *F = CI->getCalledFunction()) { 110 switch ((Intrinsic::ID)F->getIntrinsicID()) { 111 case Intrinsic::experimental_constrained_fadd: 112 case Intrinsic::experimental_constrained_fsub: 113 case Intrinsic::experimental_constrained_fmul: 114 case Intrinsic::experimental_constrained_fdiv: 115 case Intrinsic::experimental_constrained_frem: 116 case Intrinsic::experimental_constrained_fptosi: 117 case Intrinsic::experimental_constrained_sitofp: 118 case Intrinsic::experimental_constrained_fptoui: 119 case Intrinsic::experimental_constrained_uitofp: 120 case Intrinsic::experimental_constrained_fcmp: 121 case Intrinsic::experimental_constrained_fcmps: { 122 auto *CFP = cast<ConstrainedFPIntrinsic>(CI); 123 return CFP->isDefaultFPEnvironment(); 124 } 125 } 126 } 127 return CI->doesNotAccessMemory() && !CI->getType()->isVoidTy(); 128 } 129 return isa<CastInst>(Inst) || isa<UnaryOperator>(Inst) || 130 isa<BinaryOperator>(Inst) || isa<GetElementPtrInst>(Inst) || 131 isa<CmpInst>(Inst) || isa<SelectInst>(Inst) || 132 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) || 133 isa<ShuffleVectorInst>(Inst) || isa<ExtractValueInst>(Inst) || 134 isa<InsertValueInst>(Inst) || isa<FreezeInst>(Inst); 135 } 136 }; 137 138 } // end anonymous namespace 139 140 namespace llvm { 141 142 template <> struct DenseMapInfo<SimpleValue> { 143 static inline SimpleValue getEmptyKey() { 144 return DenseMapInfo<Instruction *>::getEmptyKey(); 145 } 146 147 static inline SimpleValue getTombstoneKey() { 148 return DenseMapInfo<Instruction *>::getTombstoneKey(); 149 } 150 151 static unsigned getHashValue(SimpleValue Val); 152 static bool isEqual(SimpleValue LHS, SimpleValue RHS); 153 }; 154 155 } // end namespace llvm 156 157 /// Match a 'select' including an optional 'not's of the condition. 158 static bool matchSelectWithOptionalNotCond(Value *V, Value *&Cond, Value *&A, 159 Value *&B, 160 SelectPatternFlavor &Flavor) { 161 // Return false if V is not even a select. 162 if (!match(V, m_Select(m_Value(Cond), m_Value(A), m_Value(B)))) 163 return false; 164 165 // Look through a 'not' of the condition operand by swapping A/B. 166 Value *CondNot; 167 if (match(Cond, m_Not(m_Value(CondNot)))) { 168 Cond = CondNot; 169 std::swap(A, B); 170 } 171 172 // Match canonical forms of min/max. We are not using ValueTracking's 173 // more powerful matchSelectPattern() because it may rely on instruction flags 174 // such as "nsw". That would be incompatible with the current hashing 175 // mechanism that may remove flags to increase the likelihood of CSE. 176 177 Flavor = SPF_UNKNOWN; 178 CmpInst::Predicate Pred; 179 180 if (!match(Cond, m_ICmp(Pred, m_Specific(A), m_Specific(B)))) { 181 // Check for commuted variants of min/max by swapping predicate. 182 // If we do not match the standard or commuted patterns, this is not a 183 // recognized form of min/max, but it is still a select, so return true. 184 if (!match(Cond, m_ICmp(Pred, m_Specific(B), m_Specific(A)))) 185 return true; 186 Pred = ICmpInst::getSwappedPredicate(Pred); 187 } 188 189 switch (Pred) { 190 case CmpInst::ICMP_UGT: Flavor = SPF_UMAX; break; 191 case CmpInst::ICMP_ULT: Flavor = SPF_UMIN; break; 192 case CmpInst::ICMP_SGT: Flavor = SPF_SMAX; break; 193 case CmpInst::ICMP_SLT: Flavor = SPF_SMIN; break; 194 // Non-strict inequalities. 195 case CmpInst::ICMP_ULE: Flavor = SPF_UMIN; break; 196 case CmpInst::ICMP_UGE: Flavor = SPF_UMAX; break; 197 case CmpInst::ICMP_SLE: Flavor = SPF_SMIN; break; 198 case CmpInst::ICMP_SGE: Flavor = SPF_SMAX; break; 199 default: break; 200 } 201 202 return true; 203 } 204 205 static unsigned getHashValueImpl(SimpleValue Val) { 206 Instruction *Inst = Val.Inst; 207 // Hash in all of the operands as pointers. 208 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(Inst)) { 209 Value *LHS = BinOp->getOperand(0); 210 Value *RHS = BinOp->getOperand(1); 211 if (BinOp->isCommutative() && BinOp->getOperand(0) > BinOp->getOperand(1)) 212 std::swap(LHS, RHS); 213 214 return hash_combine(BinOp->getOpcode(), LHS, RHS); 215 } 216 217 if (CmpInst *CI = dyn_cast<CmpInst>(Inst)) { 218 // Compares can be commuted by swapping the comparands and 219 // updating the predicate. Choose the form that has the 220 // comparands in sorted order, or in the case of a tie, the 221 // one with the lower predicate. 222 Value *LHS = CI->getOperand(0); 223 Value *RHS = CI->getOperand(1); 224 CmpInst::Predicate Pred = CI->getPredicate(); 225 CmpInst::Predicate SwappedPred = CI->getSwappedPredicate(); 226 if (std::tie(LHS, Pred) > std::tie(RHS, SwappedPred)) { 227 std::swap(LHS, RHS); 228 Pred = SwappedPred; 229 } 230 return hash_combine(Inst->getOpcode(), Pred, LHS, RHS); 231 } 232 233 // Hash general selects to allow matching commuted true/false operands. 234 SelectPatternFlavor SPF; 235 Value *Cond, *A, *B; 236 if (matchSelectWithOptionalNotCond(Inst, Cond, A, B, SPF)) { 237 // Hash min/max (cmp + select) to allow for commuted operands. 238 // Min/max may also have non-canonical compare predicate (eg, the compare for 239 // smin may use 'sgt' rather than 'slt'), and non-canonical operands in the 240 // compare. 241 // TODO: We should also detect FP min/max. 242 if (SPF == SPF_SMIN || SPF == SPF_SMAX || 243 SPF == SPF_UMIN || SPF == SPF_UMAX) { 244 if (A > B) 245 std::swap(A, B); 246 return hash_combine(Inst->getOpcode(), SPF, A, B); 247 } 248 249 // Hash general selects to allow matching commuted true/false operands. 250 251 // If we do not have a compare as the condition, just hash in the condition. 252 CmpInst::Predicate Pred; 253 Value *X, *Y; 254 if (!match(Cond, m_Cmp(Pred, m_Value(X), m_Value(Y)))) 255 return hash_combine(Inst->getOpcode(), Cond, A, B); 256 257 // Similar to cmp normalization (above) - canonicalize the predicate value: 258 // select (icmp Pred, X, Y), A, B --> select (icmp InvPred, X, Y), B, A 259 if (CmpInst::getInversePredicate(Pred) < Pred) { 260 Pred = CmpInst::getInversePredicate(Pred); 261 std::swap(A, B); 262 } 263 return hash_combine(Inst->getOpcode(), Pred, X, Y, A, B); 264 } 265 266 if (CastInst *CI = dyn_cast<CastInst>(Inst)) 267 return hash_combine(CI->getOpcode(), CI->getType(), CI->getOperand(0)); 268 269 if (FreezeInst *FI = dyn_cast<FreezeInst>(Inst)) 270 return hash_combine(FI->getOpcode(), FI->getOperand(0)); 271 272 if (const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(Inst)) 273 return hash_combine(EVI->getOpcode(), EVI->getOperand(0), 274 hash_combine_range(EVI->idx_begin(), EVI->idx_end())); 275 276 if (const InsertValueInst *IVI = dyn_cast<InsertValueInst>(Inst)) 277 return hash_combine(IVI->getOpcode(), IVI->getOperand(0), 278 IVI->getOperand(1), 279 hash_combine_range(IVI->idx_begin(), IVI->idx_end())); 280 281 assert((isa<CallInst>(Inst) || isa<GetElementPtrInst>(Inst) || 282 isa<ExtractElementInst>(Inst) || isa<InsertElementInst>(Inst) || 283 isa<ShuffleVectorInst>(Inst) || isa<UnaryOperator>(Inst) || 284 isa<FreezeInst>(Inst)) && 285 "Invalid/unknown instruction"); 286 287 // Handle intrinsics with commutative operands. 288 // TODO: Extend this to handle intrinsics with >2 operands where the 1st 289 // 2 operands are commutative. 290 auto *II = dyn_cast<IntrinsicInst>(Inst); 291 if (II && II->isCommutative() && II->arg_size() == 2) { 292 Value *LHS = II->getArgOperand(0), *RHS = II->getArgOperand(1); 293 if (LHS > RHS) 294 std::swap(LHS, RHS); 295 return hash_combine(II->getOpcode(), LHS, RHS); 296 } 297 298 // gc.relocate is 'special' call: its second and third operands are 299 // not real values, but indices into statepoint's argument list. 300 // Get values they point to. 301 if (const GCRelocateInst *GCR = dyn_cast<GCRelocateInst>(Inst)) 302 return hash_combine(GCR->getOpcode(), GCR->getOperand(0), 303 GCR->getBasePtr(), GCR->getDerivedPtr()); 304 305 // Mix in the opcode. 306 return hash_combine( 307 Inst->getOpcode(), 308 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end())); 309 } 310 311 unsigned DenseMapInfo<SimpleValue>::getHashValue(SimpleValue Val) { 312 #ifndef NDEBUG 313 // If -earlycse-debug-hash was specified, return a constant -- this 314 // will force all hashing to collide, so we'll exhaustively search 315 // the table for a match, and the assertion in isEqual will fire if 316 // there's a bug causing equal keys to hash differently. 317 if (EarlyCSEDebugHash) 318 return 0; 319 #endif 320 return getHashValueImpl(Val); 321 } 322 323 static bool isEqualImpl(SimpleValue LHS, SimpleValue RHS) { 324 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst; 325 326 if (LHS.isSentinel() || RHS.isSentinel()) 327 return LHSI == RHSI; 328 329 if (LHSI->getOpcode() != RHSI->getOpcode()) 330 return false; 331 if (LHSI->isIdenticalToWhenDefined(RHSI)) 332 return true; 333 334 // If we're not strictly identical, we still might be a commutable instruction 335 if (BinaryOperator *LHSBinOp = dyn_cast<BinaryOperator>(LHSI)) { 336 if (!LHSBinOp->isCommutative()) 337 return false; 338 339 assert(isa<BinaryOperator>(RHSI) && 340 "same opcode, but different instruction type?"); 341 BinaryOperator *RHSBinOp = cast<BinaryOperator>(RHSI); 342 343 // Commuted equality 344 return LHSBinOp->getOperand(0) == RHSBinOp->getOperand(1) && 345 LHSBinOp->getOperand(1) == RHSBinOp->getOperand(0); 346 } 347 if (CmpInst *LHSCmp = dyn_cast<CmpInst>(LHSI)) { 348 assert(isa<CmpInst>(RHSI) && 349 "same opcode, but different instruction type?"); 350 CmpInst *RHSCmp = cast<CmpInst>(RHSI); 351 // Commuted equality 352 return LHSCmp->getOperand(0) == RHSCmp->getOperand(1) && 353 LHSCmp->getOperand(1) == RHSCmp->getOperand(0) && 354 LHSCmp->getSwappedPredicate() == RHSCmp->getPredicate(); 355 } 356 357 // TODO: Extend this for >2 args by matching the trailing N-2 args. 358 auto *LII = dyn_cast<IntrinsicInst>(LHSI); 359 auto *RII = dyn_cast<IntrinsicInst>(RHSI); 360 if (LII && RII && LII->getIntrinsicID() == RII->getIntrinsicID() && 361 LII->isCommutative() && LII->arg_size() == 2) { 362 return LII->getArgOperand(0) == RII->getArgOperand(1) && 363 LII->getArgOperand(1) == RII->getArgOperand(0); 364 } 365 366 // See comment above in `getHashValue()`. 367 if (const GCRelocateInst *GCR1 = dyn_cast<GCRelocateInst>(LHSI)) 368 if (const GCRelocateInst *GCR2 = dyn_cast<GCRelocateInst>(RHSI)) 369 return GCR1->getOperand(0) == GCR2->getOperand(0) && 370 GCR1->getBasePtr() == GCR2->getBasePtr() && 371 GCR1->getDerivedPtr() == GCR2->getDerivedPtr(); 372 373 // Min/max can occur with commuted operands, non-canonical predicates, 374 // and/or non-canonical operands. 375 // Selects can be non-trivially equivalent via inverted conditions and swaps. 376 SelectPatternFlavor LSPF, RSPF; 377 Value *CondL, *CondR, *LHSA, *RHSA, *LHSB, *RHSB; 378 if (matchSelectWithOptionalNotCond(LHSI, CondL, LHSA, LHSB, LSPF) && 379 matchSelectWithOptionalNotCond(RHSI, CondR, RHSA, RHSB, RSPF)) { 380 if (LSPF == RSPF) { 381 // TODO: We should also detect FP min/max. 382 if (LSPF == SPF_SMIN || LSPF == SPF_SMAX || 383 LSPF == SPF_UMIN || LSPF == SPF_UMAX) 384 return ((LHSA == RHSA && LHSB == RHSB) || 385 (LHSA == RHSB && LHSB == RHSA)); 386 387 // select Cond, A, B <--> select not(Cond), B, A 388 if (CondL == CondR && LHSA == RHSA && LHSB == RHSB) 389 return true; 390 } 391 392 // If the true/false operands are swapped and the conditions are compares 393 // with inverted predicates, the selects are equal: 394 // select (icmp Pred, X, Y), A, B <--> select (icmp InvPred, X, Y), B, A 395 // 396 // This also handles patterns with a double-negation in the sense of not + 397 // inverse, because we looked through a 'not' in the matching function and 398 // swapped A/B: 399 // select (cmp Pred, X, Y), A, B <--> select (not (cmp InvPred, X, Y)), B, A 400 // 401 // This intentionally does NOT handle patterns with a double-negation in 402 // the sense of not + not, because doing so could result in values 403 // comparing 404 // as equal that hash differently in the min/max cases like: 405 // select (cmp slt, X, Y), X, Y <--> select (not (not (cmp slt, X, Y))), X, Y 406 // ^ hashes as min ^ would not hash as min 407 // In the context of the EarlyCSE pass, however, such cases never reach 408 // this code, as we simplify the double-negation before hashing the second 409 // select (and so still succeed at CSEing them). 410 if (LHSA == RHSB && LHSB == RHSA) { 411 CmpInst::Predicate PredL, PredR; 412 Value *X, *Y; 413 if (match(CondL, m_Cmp(PredL, m_Value(X), m_Value(Y))) && 414 match(CondR, m_Cmp(PredR, m_Specific(X), m_Specific(Y))) && 415 CmpInst::getInversePredicate(PredL) == PredR) 416 return true; 417 } 418 } 419 420 return false; 421 } 422 423 bool DenseMapInfo<SimpleValue>::isEqual(SimpleValue LHS, SimpleValue RHS) { 424 // These comparisons are nontrivial, so assert that equality implies 425 // hash equality (DenseMap demands this as an invariant). 426 bool Result = isEqualImpl(LHS, RHS); 427 assert(!Result || (LHS.isSentinel() && LHS.Inst == RHS.Inst) || 428 getHashValueImpl(LHS) == getHashValueImpl(RHS)); 429 return Result; 430 } 431 432 //===----------------------------------------------------------------------===// 433 // CallValue 434 //===----------------------------------------------------------------------===// 435 436 namespace { 437 438 /// Struct representing the available call values in the scoped hash 439 /// table. 440 struct CallValue { 441 Instruction *Inst; 442 443 CallValue(Instruction *I) : Inst(I) { 444 assert((isSentinel() || canHandle(I)) && "Inst can't be handled!"); 445 } 446 447 bool isSentinel() const { 448 return Inst == DenseMapInfo<Instruction *>::getEmptyKey() || 449 Inst == DenseMapInfo<Instruction *>::getTombstoneKey(); 450 } 451 452 static bool canHandle(Instruction *Inst) { 453 // Don't value number anything that returns void. 454 if (Inst->getType()->isVoidTy()) 455 return false; 456 457 CallInst *CI = dyn_cast<CallInst>(Inst); 458 if (!CI || !CI->onlyReadsMemory()) 459 return false; 460 return true; 461 } 462 }; 463 464 } // end anonymous namespace 465 466 namespace llvm { 467 468 template <> struct DenseMapInfo<CallValue> { 469 static inline CallValue getEmptyKey() { 470 return DenseMapInfo<Instruction *>::getEmptyKey(); 471 } 472 473 static inline CallValue getTombstoneKey() { 474 return DenseMapInfo<Instruction *>::getTombstoneKey(); 475 } 476 477 static unsigned getHashValue(CallValue Val); 478 static bool isEqual(CallValue LHS, CallValue RHS); 479 }; 480 481 } // end namespace llvm 482 483 unsigned DenseMapInfo<CallValue>::getHashValue(CallValue Val) { 484 Instruction *Inst = Val.Inst; 485 486 // Hash all of the operands as pointers and mix in the opcode. 487 return hash_combine( 488 Inst->getOpcode(), 489 hash_combine_range(Inst->value_op_begin(), Inst->value_op_end())); 490 } 491 492 bool DenseMapInfo<CallValue>::isEqual(CallValue LHS, CallValue RHS) { 493 Instruction *LHSI = LHS.Inst, *RHSI = RHS.Inst; 494 if (LHS.isSentinel() || RHS.isSentinel()) 495 return LHSI == RHSI; 496 497 return LHSI->isIdenticalTo(RHSI); 498 } 499 500 //===----------------------------------------------------------------------===// 501 // EarlyCSE implementation 502 //===----------------------------------------------------------------------===// 503 504 namespace { 505 506 /// A simple and fast domtree-based CSE pass. 507 /// 508 /// This pass does a simple depth-first walk over the dominator tree, 509 /// eliminating trivially redundant instructions and using instsimplify to 510 /// canonicalize things as it goes. It is intended to be fast and catch obvious 511 /// cases so that instcombine and other passes are more effective. It is 512 /// expected that a later pass of GVN will catch the interesting/hard cases. 513 class EarlyCSE { 514 public: 515 const TargetLibraryInfo &TLI; 516 const TargetTransformInfo &TTI; 517 DominatorTree &DT; 518 AssumptionCache &AC; 519 const SimplifyQuery SQ; 520 MemorySSA *MSSA; 521 std::unique_ptr<MemorySSAUpdater> MSSAUpdater; 522 523 using AllocatorTy = 524 RecyclingAllocator<BumpPtrAllocator, 525 ScopedHashTableVal<SimpleValue, Value *>>; 526 using ScopedHTType = 527 ScopedHashTable<SimpleValue, Value *, DenseMapInfo<SimpleValue>, 528 AllocatorTy>; 529 530 /// A scoped hash table of the current values of all of our simple 531 /// scalar expressions. 532 /// 533 /// As we walk down the domtree, we look to see if instructions are in this: 534 /// if so, we replace them with what we find, otherwise we insert them so 535 /// that dominated values can succeed in their lookup. 536 ScopedHTType AvailableValues; 537 538 /// A scoped hash table of the current values of previously encountered 539 /// memory locations. 540 /// 541 /// This allows us to get efficient access to dominating loads or stores when 542 /// we have a fully redundant load. In addition to the most recent load, we 543 /// keep track of a generation count of the read, which is compared against 544 /// the current generation count. The current generation count is incremented 545 /// after every possibly writing memory operation, which ensures that we only 546 /// CSE loads with other loads that have no intervening store. Ordering 547 /// events (such as fences or atomic instructions) increment the generation 548 /// count as well; essentially, we model these as writes to all possible 549 /// locations. Note that atomic and/or volatile loads and stores can be 550 /// present the table; it is the responsibility of the consumer to inspect 551 /// the atomicity/volatility if needed. 552 struct LoadValue { 553 Instruction *DefInst = nullptr; 554 unsigned Generation = 0; 555 int MatchingId = -1; 556 bool IsAtomic = false; 557 558 LoadValue() = default; 559 LoadValue(Instruction *Inst, unsigned Generation, unsigned MatchingId, 560 bool IsAtomic) 561 : DefInst(Inst), Generation(Generation), MatchingId(MatchingId), 562 IsAtomic(IsAtomic) {} 563 }; 564 565 using LoadMapAllocator = 566 RecyclingAllocator<BumpPtrAllocator, 567 ScopedHashTableVal<Value *, LoadValue>>; 568 using LoadHTType = 569 ScopedHashTable<Value *, LoadValue, DenseMapInfo<Value *>, 570 LoadMapAllocator>; 571 572 LoadHTType AvailableLoads; 573 574 // A scoped hash table mapping memory locations (represented as typed 575 // addresses) to generation numbers at which that memory location became 576 // (henceforth indefinitely) invariant. 577 using InvariantMapAllocator = 578 RecyclingAllocator<BumpPtrAllocator, 579 ScopedHashTableVal<MemoryLocation, unsigned>>; 580 using InvariantHTType = 581 ScopedHashTable<MemoryLocation, unsigned, DenseMapInfo<MemoryLocation>, 582 InvariantMapAllocator>; 583 InvariantHTType AvailableInvariants; 584 585 /// A scoped hash table of the current values of read-only call 586 /// values. 587 /// 588 /// It uses the same generation count as loads. 589 using CallHTType = 590 ScopedHashTable<CallValue, std::pair<Instruction *, unsigned>>; 591 CallHTType AvailableCalls; 592 593 /// This is the current generation of the memory value. 594 unsigned CurrentGeneration = 0; 595 596 /// Set up the EarlyCSE runner for a particular function. 597 EarlyCSE(const DataLayout &DL, const TargetLibraryInfo &TLI, 598 const TargetTransformInfo &TTI, DominatorTree &DT, 599 AssumptionCache &AC, MemorySSA *MSSA) 600 : TLI(TLI), TTI(TTI), DT(DT), AC(AC), SQ(DL, &TLI, &DT, &AC), MSSA(MSSA), 601 MSSAUpdater(std::make_unique<MemorySSAUpdater>(MSSA)) {} 602 603 bool run(); 604 605 private: 606 unsigned ClobberCounter = 0; 607 // Almost a POD, but needs to call the constructors for the scoped hash 608 // tables so that a new scope gets pushed on. These are RAII so that the 609 // scope gets popped when the NodeScope is destroyed. 610 class NodeScope { 611 public: 612 NodeScope(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads, 613 InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls) 614 : Scope(AvailableValues), LoadScope(AvailableLoads), 615 InvariantScope(AvailableInvariants), CallScope(AvailableCalls) {} 616 NodeScope(const NodeScope &) = delete; 617 NodeScope &operator=(const NodeScope &) = delete; 618 619 private: 620 ScopedHTType::ScopeTy Scope; 621 LoadHTType::ScopeTy LoadScope; 622 InvariantHTType::ScopeTy InvariantScope; 623 CallHTType::ScopeTy CallScope; 624 }; 625 626 // Contains all the needed information to create a stack for doing a depth 627 // first traversal of the tree. This includes scopes for values, loads, and 628 // calls as well as the generation. There is a child iterator so that the 629 // children do not need to be store separately. 630 class StackNode { 631 public: 632 StackNode(ScopedHTType &AvailableValues, LoadHTType &AvailableLoads, 633 InvariantHTType &AvailableInvariants, CallHTType &AvailableCalls, 634 unsigned cg, DomTreeNode *n, DomTreeNode::const_iterator child, 635 DomTreeNode::const_iterator end) 636 : CurrentGeneration(cg), ChildGeneration(cg), Node(n), ChildIter(child), 637 EndIter(end), 638 Scopes(AvailableValues, AvailableLoads, AvailableInvariants, 639 AvailableCalls) 640 {} 641 StackNode(const StackNode &) = delete; 642 StackNode &operator=(const StackNode &) = delete; 643 644 // Accessors. 645 unsigned currentGeneration() const { return CurrentGeneration; } 646 unsigned childGeneration() const { return ChildGeneration; } 647 void childGeneration(unsigned generation) { ChildGeneration = generation; } 648 DomTreeNode *node() { return Node; } 649 DomTreeNode::const_iterator childIter() const { return ChildIter; } 650 651 DomTreeNode *nextChild() { 652 DomTreeNode *child = *ChildIter; 653 ++ChildIter; 654 return child; 655 } 656 657 DomTreeNode::const_iterator end() const { return EndIter; } 658 bool isProcessed() const { return Processed; } 659 void process() { Processed = true; } 660 661 private: 662 unsigned CurrentGeneration; 663 unsigned ChildGeneration; 664 DomTreeNode *Node; 665 DomTreeNode::const_iterator ChildIter; 666 DomTreeNode::const_iterator EndIter; 667 NodeScope Scopes; 668 bool Processed = false; 669 }; 670 671 /// Wrapper class to handle memory instructions, including loads, 672 /// stores and intrinsic loads and stores defined by the target. 673 class ParseMemoryInst { 674 public: 675 ParseMemoryInst(Instruction *Inst, const TargetTransformInfo &TTI) 676 : Inst(Inst) { 677 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 678 IntrID = II->getIntrinsicID(); 679 if (TTI.getTgtMemIntrinsic(II, Info)) 680 return; 681 if (isHandledNonTargetIntrinsic(IntrID)) { 682 switch (IntrID) { 683 case Intrinsic::masked_load: 684 Info.PtrVal = Inst->getOperand(0); 685 Info.MatchingId = Intrinsic::masked_load; 686 Info.ReadMem = true; 687 Info.WriteMem = false; 688 Info.IsVolatile = false; 689 break; 690 case Intrinsic::masked_store: 691 Info.PtrVal = Inst->getOperand(1); 692 // Use the ID of masked load as the "matching id". This will 693 // prevent matching non-masked loads/stores with masked ones 694 // (which could be done), but at the moment, the code here 695 // does not support matching intrinsics with non-intrinsics, 696 // so keep the MatchingIds specific to masked instructions 697 // for now (TODO). 698 Info.MatchingId = Intrinsic::masked_load; 699 Info.ReadMem = false; 700 Info.WriteMem = true; 701 Info.IsVolatile = false; 702 break; 703 } 704 } 705 } 706 } 707 708 Instruction *get() { return Inst; } 709 const Instruction *get() const { return Inst; } 710 711 bool isLoad() const { 712 if (IntrID != 0) 713 return Info.ReadMem; 714 return isa<LoadInst>(Inst); 715 } 716 717 bool isStore() const { 718 if (IntrID != 0) 719 return Info.WriteMem; 720 return isa<StoreInst>(Inst); 721 } 722 723 bool isAtomic() const { 724 if (IntrID != 0) 725 return Info.Ordering != AtomicOrdering::NotAtomic; 726 return Inst->isAtomic(); 727 } 728 729 bool isUnordered() const { 730 if (IntrID != 0) 731 return Info.isUnordered(); 732 733 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 734 return LI->isUnordered(); 735 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 736 return SI->isUnordered(); 737 } 738 // Conservative answer 739 return !Inst->isAtomic(); 740 } 741 742 bool isVolatile() const { 743 if (IntrID != 0) 744 return Info.IsVolatile; 745 746 if (LoadInst *LI = dyn_cast<LoadInst>(Inst)) { 747 return LI->isVolatile(); 748 } else if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 749 return SI->isVolatile(); 750 } 751 // Conservative answer 752 return true; 753 } 754 755 bool isInvariantLoad() const { 756 if (auto *LI = dyn_cast<LoadInst>(Inst)) 757 return LI->hasMetadata(LLVMContext::MD_invariant_load); 758 return false; 759 } 760 761 bool isValid() const { return getPointerOperand() != nullptr; } 762 763 // For regular (non-intrinsic) loads/stores, this is set to -1. For 764 // intrinsic loads/stores, the id is retrieved from the corresponding 765 // field in the MemIntrinsicInfo structure. That field contains 766 // non-negative values only. 767 int getMatchingId() const { 768 if (IntrID != 0) 769 return Info.MatchingId; 770 return -1; 771 } 772 773 Value *getPointerOperand() const { 774 if (IntrID != 0) 775 return Info.PtrVal; 776 return getLoadStorePointerOperand(Inst); 777 } 778 779 Type *getValueType() const { 780 // TODO: handle target-specific intrinsics. 781 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 782 switch (II->getIntrinsicID()) { 783 case Intrinsic::masked_load: 784 return II->getType(); 785 case Intrinsic::masked_store: 786 return II->getArgOperand(0)->getType(); 787 default: 788 return nullptr; 789 } 790 } 791 return getLoadStoreType(Inst); 792 } 793 794 bool mayReadFromMemory() const { 795 if (IntrID != 0) 796 return Info.ReadMem; 797 return Inst->mayReadFromMemory(); 798 } 799 800 bool mayWriteToMemory() const { 801 if (IntrID != 0) 802 return Info.WriteMem; 803 return Inst->mayWriteToMemory(); 804 } 805 806 private: 807 Intrinsic::ID IntrID = 0; 808 MemIntrinsicInfo Info; 809 Instruction *Inst; 810 }; 811 812 // This function is to prevent accidentally passing a non-target 813 // intrinsic ID to TargetTransformInfo. 814 static bool isHandledNonTargetIntrinsic(Intrinsic::ID ID) { 815 switch (ID) { 816 case Intrinsic::masked_load: 817 case Intrinsic::masked_store: 818 return true; 819 } 820 return false; 821 } 822 static bool isHandledNonTargetIntrinsic(const Value *V) { 823 if (auto *II = dyn_cast<IntrinsicInst>(V)) 824 return isHandledNonTargetIntrinsic(II->getIntrinsicID()); 825 return false; 826 } 827 828 bool processNode(DomTreeNode *Node); 829 830 bool handleBranchCondition(Instruction *CondInst, const BranchInst *BI, 831 const BasicBlock *BB, const BasicBlock *Pred); 832 833 Value *getMatchingValue(LoadValue &InVal, ParseMemoryInst &MemInst, 834 unsigned CurrentGeneration); 835 836 bool overridingStores(const ParseMemoryInst &Earlier, 837 const ParseMemoryInst &Later); 838 839 Value *getOrCreateResult(Value *Inst, Type *ExpectedType) const { 840 // TODO: We could insert relevant casts on type mismatch here. 841 if (auto *LI = dyn_cast<LoadInst>(Inst)) 842 return LI->getType() == ExpectedType ? LI : nullptr; 843 else if (auto *SI = dyn_cast<StoreInst>(Inst)) { 844 Value *V = SI->getValueOperand(); 845 return V->getType() == ExpectedType ? V : nullptr; 846 } 847 assert(isa<IntrinsicInst>(Inst) && "Instruction not supported"); 848 auto *II = cast<IntrinsicInst>(Inst); 849 if (isHandledNonTargetIntrinsic(II->getIntrinsicID())) 850 return getOrCreateResultNonTargetMemIntrinsic(II, ExpectedType); 851 return TTI.getOrCreateResultFromMemIntrinsic(II, ExpectedType); 852 } 853 854 Value *getOrCreateResultNonTargetMemIntrinsic(IntrinsicInst *II, 855 Type *ExpectedType) const { 856 switch (II->getIntrinsicID()) { 857 case Intrinsic::masked_load: 858 return II; 859 case Intrinsic::masked_store: 860 return II->getOperand(0); 861 } 862 return nullptr; 863 } 864 865 /// Return true if the instruction is known to only operate on memory 866 /// provably invariant in the given "generation". 867 bool isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt); 868 869 bool isSameMemGeneration(unsigned EarlierGeneration, unsigned LaterGeneration, 870 Instruction *EarlierInst, Instruction *LaterInst); 871 872 bool isNonTargetIntrinsicMatch(const IntrinsicInst *Earlier, 873 const IntrinsicInst *Later) { 874 auto IsSubmask = [](const Value *Mask0, const Value *Mask1) { 875 // Is Mask0 a submask of Mask1? 876 if (Mask0 == Mask1) 877 return true; 878 if (isa<UndefValue>(Mask0) || isa<UndefValue>(Mask1)) 879 return false; 880 auto *Vec0 = dyn_cast<ConstantVector>(Mask0); 881 auto *Vec1 = dyn_cast<ConstantVector>(Mask1); 882 if (!Vec0 || !Vec1) 883 return false; 884 assert(Vec0->getType() == Vec1->getType() && 885 "Masks should have the same type"); 886 for (int i = 0, e = Vec0->getNumOperands(); i != e; ++i) { 887 Constant *Elem0 = Vec0->getOperand(i); 888 Constant *Elem1 = Vec1->getOperand(i); 889 auto *Int0 = dyn_cast<ConstantInt>(Elem0); 890 if (Int0 && Int0->isZero()) 891 continue; 892 auto *Int1 = dyn_cast<ConstantInt>(Elem1); 893 if (Int1 && !Int1->isZero()) 894 continue; 895 if (isa<UndefValue>(Elem0) || isa<UndefValue>(Elem1)) 896 return false; 897 if (Elem0 == Elem1) 898 continue; 899 return false; 900 } 901 return true; 902 }; 903 auto PtrOp = [](const IntrinsicInst *II) { 904 if (II->getIntrinsicID() == Intrinsic::masked_load) 905 return II->getOperand(0); 906 if (II->getIntrinsicID() == Intrinsic::masked_store) 907 return II->getOperand(1); 908 llvm_unreachable("Unexpected IntrinsicInst"); 909 }; 910 auto MaskOp = [](const IntrinsicInst *II) { 911 if (II->getIntrinsicID() == Intrinsic::masked_load) 912 return II->getOperand(2); 913 if (II->getIntrinsicID() == Intrinsic::masked_store) 914 return II->getOperand(3); 915 llvm_unreachable("Unexpected IntrinsicInst"); 916 }; 917 auto ThruOp = [](const IntrinsicInst *II) { 918 if (II->getIntrinsicID() == Intrinsic::masked_load) 919 return II->getOperand(3); 920 llvm_unreachable("Unexpected IntrinsicInst"); 921 }; 922 923 if (PtrOp(Earlier) != PtrOp(Later)) 924 return false; 925 926 Intrinsic::ID IDE = Earlier->getIntrinsicID(); 927 Intrinsic::ID IDL = Later->getIntrinsicID(); 928 // We could really use specific intrinsic classes for masked loads 929 // and stores in IntrinsicInst.h. 930 if (IDE == Intrinsic::masked_load && IDL == Intrinsic::masked_load) { 931 // Trying to replace later masked load with the earlier one. 932 // Check that the pointers are the same, and 933 // - masks and pass-throughs are the same, or 934 // - replacee's pass-through is "undef" and replacer's mask is a 935 // super-set of the replacee's mask. 936 if (MaskOp(Earlier) == MaskOp(Later) && ThruOp(Earlier) == ThruOp(Later)) 937 return true; 938 if (!isa<UndefValue>(ThruOp(Later))) 939 return false; 940 return IsSubmask(MaskOp(Later), MaskOp(Earlier)); 941 } 942 if (IDE == Intrinsic::masked_store && IDL == Intrinsic::masked_load) { 943 // Trying to replace a load of a stored value with the store's value. 944 // Check that the pointers are the same, and 945 // - load's mask is a subset of store's mask, and 946 // - load's pass-through is "undef". 947 if (!IsSubmask(MaskOp(Later), MaskOp(Earlier))) 948 return false; 949 return isa<UndefValue>(ThruOp(Later)); 950 } 951 if (IDE == Intrinsic::masked_load && IDL == Intrinsic::masked_store) { 952 // Trying to remove a store of the loaded value. 953 // Check that the pointers are the same, and 954 // - store's mask is a subset of the load's mask. 955 return IsSubmask(MaskOp(Later), MaskOp(Earlier)); 956 } 957 if (IDE == Intrinsic::masked_store && IDL == Intrinsic::masked_store) { 958 // Trying to remove a dead store (earlier). 959 // Check that the pointers are the same, 960 // - the to-be-removed store's mask is a subset of the other store's 961 // mask. 962 return IsSubmask(MaskOp(Earlier), MaskOp(Later)); 963 } 964 return false; 965 } 966 967 void removeMSSA(Instruction &Inst) { 968 if (!MSSA) 969 return; 970 if (VerifyMemorySSA) 971 MSSA->verifyMemorySSA(); 972 // Removing a store here can leave MemorySSA in an unoptimized state by 973 // creating MemoryPhis that have identical arguments and by creating 974 // MemoryUses whose defining access is not an actual clobber. The phi case 975 // is handled by MemorySSA when passing OptimizePhis = true to 976 // removeMemoryAccess. The non-optimized MemoryUse case is lazily updated 977 // by MemorySSA's getClobberingMemoryAccess. 978 MSSAUpdater->removeMemoryAccess(&Inst, true); 979 } 980 }; 981 982 } // end anonymous namespace 983 984 /// Determine if the memory referenced by LaterInst is from the same heap 985 /// version as EarlierInst. 986 /// This is currently called in two scenarios: 987 /// 988 /// load p 989 /// ... 990 /// load p 991 /// 992 /// and 993 /// 994 /// x = load p 995 /// ... 996 /// store x, p 997 /// 998 /// in both cases we want to verify that there are no possible writes to the 999 /// memory referenced by p between the earlier and later instruction. 1000 bool EarlyCSE::isSameMemGeneration(unsigned EarlierGeneration, 1001 unsigned LaterGeneration, 1002 Instruction *EarlierInst, 1003 Instruction *LaterInst) { 1004 // Check the simple memory generation tracking first. 1005 if (EarlierGeneration == LaterGeneration) 1006 return true; 1007 1008 if (!MSSA) 1009 return false; 1010 1011 // If MemorySSA has determined that one of EarlierInst or LaterInst does not 1012 // read/write memory, then we can safely return true here. 1013 // FIXME: We could be more aggressive when checking doesNotAccessMemory(), 1014 // onlyReadsMemory(), mayReadFromMemory(), and mayWriteToMemory() in this pass 1015 // by also checking the MemorySSA MemoryAccess on the instruction. Initial 1016 // experiments suggest this isn't worthwhile, at least for C/C++ code compiled 1017 // with the default optimization pipeline. 1018 auto *EarlierMA = MSSA->getMemoryAccess(EarlierInst); 1019 if (!EarlierMA) 1020 return true; 1021 auto *LaterMA = MSSA->getMemoryAccess(LaterInst); 1022 if (!LaterMA) 1023 return true; 1024 1025 // Since we know LaterDef dominates LaterInst and EarlierInst dominates 1026 // LaterInst, if LaterDef dominates EarlierInst then it can't occur between 1027 // EarlierInst and LaterInst and neither can any other write that potentially 1028 // clobbers LaterInst. 1029 MemoryAccess *LaterDef; 1030 if (ClobberCounter < EarlyCSEMssaOptCap) { 1031 LaterDef = MSSA->getWalker()->getClobberingMemoryAccess(LaterInst); 1032 ClobberCounter++; 1033 } else 1034 LaterDef = LaterMA->getDefiningAccess(); 1035 1036 return MSSA->dominates(LaterDef, EarlierMA); 1037 } 1038 1039 bool EarlyCSE::isOperatingOnInvariantMemAt(Instruction *I, unsigned GenAt) { 1040 // A location loaded from with an invariant_load is assumed to *never* change 1041 // within the visible scope of the compilation. 1042 if (auto *LI = dyn_cast<LoadInst>(I)) 1043 if (LI->hasMetadata(LLVMContext::MD_invariant_load)) 1044 return true; 1045 1046 auto MemLocOpt = MemoryLocation::getOrNone(I); 1047 if (!MemLocOpt) 1048 // "target" intrinsic forms of loads aren't currently known to 1049 // MemoryLocation::get. TODO 1050 return false; 1051 MemoryLocation MemLoc = *MemLocOpt; 1052 if (!AvailableInvariants.count(MemLoc)) 1053 return false; 1054 1055 // Is the generation at which this became invariant older than the 1056 // current one? 1057 return AvailableInvariants.lookup(MemLoc) <= GenAt; 1058 } 1059 1060 bool EarlyCSE::handleBranchCondition(Instruction *CondInst, 1061 const BranchInst *BI, const BasicBlock *BB, 1062 const BasicBlock *Pred) { 1063 assert(BI->isConditional() && "Should be a conditional branch!"); 1064 assert(BI->getCondition() == CondInst && "Wrong condition?"); 1065 assert(BI->getSuccessor(0) == BB || BI->getSuccessor(1) == BB); 1066 auto *TorF = (BI->getSuccessor(0) == BB) 1067 ? ConstantInt::getTrue(BB->getContext()) 1068 : ConstantInt::getFalse(BB->getContext()); 1069 auto MatchBinOp = [](Instruction *I, unsigned Opcode, Value *&LHS, 1070 Value *&RHS) { 1071 if (Opcode == Instruction::And && 1072 match(I, m_LogicalAnd(m_Value(LHS), m_Value(RHS)))) 1073 return true; 1074 else if (Opcode == Instruction::Or && 1075 match(I, m_LogicalOr(m_Value(LHS), m_Value(RHS)))) 1076 return true; 1077 return false; 1078 }; 1079 // If the condition is AND operation, we can propagate its operands into the 1080 // true branch. If it is OR operation, we can propagate them into the false 1081 // branch. 1082 unsigned PropagateOpcode = 1083 (BI->getSuccessor(0) == BB) ? Instruction::And : Instruction::Or; 1084 1085 bool MadeChanges = false; 1086 SmallVector<Instruction *, 4> WorkList; 1087 SmallPtrSet<Instruction *, 4> Visited; 1088 WorkList.push_back(CondInst); 1089 while (!WorkList.empty()) { 1090 Instruction *Curr = WorkList.pop_back_val(); 1091 1092 AvailableValues.insert(Curr, TorF); 1093 LLVM_DEBUG(dbgs() << "EarlyCSE CVP: Add conditional value for '" 1094 << Curr->getName() << "' as " << *TorF << " in " 1095 << BB->getName() << "\n"); 1096 if (!DebugCounter::shouldExecute(CSECounter)) { 1097 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1098 } else { 1099 // Replace all dominated uses with the known value. 1100 if (unsigned Count = replaceDominatedUsesWith(Curr, TorF, DT, 1101 BasicBlockEdge(Pred, BB))) { 1102 NumCSECVP += Count; 1103 MadeChanges = true; 1104 } 1105 } 1106 1107 Value *LHS, *RHS; 1108 if (MatchBinOp(Curr, PropagateOpcode, LHS, RHS)) 1109 for (auto &Op : { LHS, RHS }) 1110 if (Instruction *OPI = dyn_cast<Instruction>(Op)) 1111 if (SimpleValue::canHandle(OPI) && Visited.insert(OPI).second) 1112 WorkList.push_back(OPI); 1113 } 1114 1115 return MadeChanges; 1116 } 1117 1118 Value *EarlyCSE::getMatchingValue(LoadValue &InVal, ParseMemoryInst &MemInst, 1119 unsigned CurrentGeneration) { 1120 if (InVal.DefInst == nullptr) 1121 return nullptr; 1122 if (InVal.MatchingId != MemInst.getMatchingId()) 1123 return nullptr; 1124 // We don't yet handle removing loads with ordering of any kind. 1125 if (MemInst.isVolatile() || !MemInst.isUnordered()) 1126 return nullptr; 1127 // We can't replace an atomic load with one which isn't also atomic. 1128 if (MemInst.isLoad() && !InVal.IsAtomic && MemInst.isAtomic()) 1129 return nullptr; 1130 // The value V returned from this function is used differently depending 1131 // on whether MemInst is a load or a store. If it's a load, we will replace 1132 // MemInst with V, if it's a store, we will check if V is the same as the 1133 // available value. 1134 bool MemInstMatching = !MemInst.isLoad(); 1135 Instruction *Matching = MemInstMatching ? MemInst.get() : InVal.DefInst; 1136 Instruction *Other = MemInstMatching ? InVal.DefInst : MemInst.get(); 1137 1138 // For stores check the result values before checking memory generation 1139 // (otherwise isSameMemGeneration may crash). 1140 Value *Result = MemInst.isStore() 1141 ? getOrCreateResult(Matching, Other->getType()) 1142 : nullptr; 1143 if (MemInst.isStore() && InVal.DefInst != Result) 1144 return nullptr; 1145 1146 // Deal with non-target memory intrinsics. 1147 bool MatchingNTI = isHandledNonTargetIntrinsic(Matching); 1148 bool OtherNTI = isHandledNonTargetIntrinsic(Other); 1149 if (OtherNTI != MatchingNTI) 1150 return nullptr; 1151 if (OtherNTI && MatchingNTI) { 1152 if (!isNonTargetIntrinsicMatch(cast<IntrinsicInst>(InVal.DefInst), 1153 cast<IntrinsicInst>(MemInst.get()))) 1154 return nullptr; 1155 } 1156 1157 if (!isOperatingOnInvariantMemAt(MemInst.get(), InVal.Generation) && 1158 !isSameMemGeneration(InVal.Generation, CurrentGeneration, InVal.DefInst, 1159 MemInst.get())) 1160 return nullptr; 1161 1162 if (!Result) 1163 Result = getOrCreateResult(Matching, Other->getType()); 1164 return Result; 1165 } 1166 1167 bool EarlyCSE::overridingStores(const ParseMemoryInst &Earlier, 1168 const ParseMemoryInst &Later) { 1169 // Can we remove Earlier store because of Later store? 1170 1171 assert(Earlier.isUnordered() && !Earlier.isVolatile() && 1172 "Violated invariant"); 1173 if (Earlier.getPointerOperand() != Later.getPointerOperand()) 1174 return false; 1175 if (!Earlier.getValueType() || !Later.getValueType() || 1176 Earlier.getValueType() != Later.getValueType()) 1177 return false; 1178 if (Earlier.getMatchingId() != Later.getMatchingId()) 1179 return false; 1180 // At the moment, we don't remove ordered stores, but do remove 1181 // unordered atomic stores. There's no special requirement (for 1182 // unordered atomics) about removing atomic stores only in favor of 1183 // other atomic stores since we were going to execute the non-atomic 1184 // one anyway and the atomic one might never have become visible. 1185 if (!Earlier.isUnordered() || !Later.isUnordered()) 1186 return false; 1187 1188 // Deal with non-target memory intrinsics. 1189 bool ENTI = isHandledNonTargetIntrinsic(Earlier.get()); 1190 bool LNTI = isHandledNonTargetIntrinsic(Later.get()); 1191 if (ENTI && LNTI) 1192 return isNonTargetIntrinsicMatch(cast<IntrinsicInst>(Earlier.get()), 1193 cast<IntrinsicInst>(Later.get())); 1194 1195 // Because of the check above, at least one of them is false. 1196 // For now disallow matching intrinsics with non-intrinsics, 1197 // so assume that the stores match if neither is an intrinsic. 1198 return ENTI == LNTI; 1199 } 1200 1201 bool EarlyCSE::processNode(DomTreeNode *Node) { 1202 bool Changed = false; 1203 BasicBlock *BB = Node->getBlock(); 1204 1205 // If this block has a single predecessor, then the predecessor is the parent 1206 // of the domtree node and all of the live out memory values are still current 1207 // in this block. If this block has multiple predecessors, then they could 1208 // have invalidated the live-out memory values of our parent value. For now, 1209 // just be conservative and invalidate memory if this block has multiple 1210 // predecessors. 1211 if (!BB->getSinglePredecessor()) 1212 ++CurrentGeneration; 1213 1214 // If this node has a single predecessor which ends in a conditional branch, 1215 // we can infer the value of the branch condition given that we took this 1216 // path. We need the single predecessor to ensure there's not another path 1217 // which reaches this block where the condition might hold a different 1218 // value. Since we're adding this to the scoped hash table (like any other 1219 // def), it will have been popped if we encounter a future merge block. 1220 if (BasicBlock *Pred = BB->getSinglePredecessor()) { 1221 auto *BI = dyn_cast<BranchInst>(Pred->getTerminator()); 1222 if (BI && BI->isConditional()) { 1223 auto *CondInst = dyn_cast<Instruction>(BI->getCondition()); 1224 if (CondInst && SimpleValue::canHandle(CondInst)) 1225 Changed |= handleBranchCondition(CondInst, BI, BB, Pred); 1226 } 1227 } 1228 1229 /// LastStore - Keep track of the last non-volatile store that we saw... for 1230 /// as long as there in no instruction that reads memory. If we see a store 1231 /// to the same location, we delete the dead store. This zaps trivial dead 1232 /// stores which can occur in bitfield code among other things. 1233 Instruction *LastStore = nullptr; 1234 1235 // See if any instructions in the block can be eliminated. If so, do it. If 1236 // not, add them to AvailableValues. 1237 for (Instruction &Inst : make_early_inc_range(BB->getInstList())) { 1238 // Dead instructions should just be removed. 1239 if (isInstructionTriviallyDead(&Inst, &TLI)) { 1240 LLVM_DEBUG(dbgs() << "EarlyCSE DCE: " << Inst << '\n'); 1241 if (!DebugCounter::shouldExecute(CSECounter)) { 1242 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1243 continue; 1244 } 1245 1246 salvageKnowledge(&Inst, &AC); 1247 salvageDebugInfo(Inst); 1248 removeMSSA(Inst); 1249 Inst.eraseFromParent(); 1250 Changed = true; 1251 ++NumSimplify; 1252 continue; 1253 } 1254 1255 // Skip assume intrinsics, they don't really have side effects (although 1256 // they're marked as such to ensure preservation of control dependencies), 1257 // and this pass will not bother with its removal. However, we should mark 1258 // its condition as true for all dominated blocks. 1259 if (auto *Assume = dyn_cast<AssumeInst>(&Inst)) { 1260 auto *CondI = dyn_cast<Instruction>(Assume->getArgOperand(0)); 1261 if (CondI && SimpleValue::canHandle(CondI)) { 1262 LLVM_DEBUG(dbgs() << "EarlyCSE considering assumption: " << Inst 1263 << '\n'); 1264 AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext())); 1265 } else 1266 LLVM_DEBUG(dbgs() << "EarlyCSE skipping assumption: " << Inst << '\n'); 1267 continue; 1268 } 1269 1270 // Likewise, noalias intrinsics don't actually write. 1271 if (match(&Inst, 1272 m_Intrinsic<Intrinsic::experimental_noalias_scope_decl>())) { 1273 LLVM_DEBUG(dbgs() << "EarlyCSE skipping noalias intrinsic: " << Inst 1274 << '\n'); 1275 continue; 1276 } 1277 1278 // Skip sideeffect intrinsics, for the same reason as assume intrinsics. 1279 if (match(&Inst, m_Intrinsic<Intrinsic::sideeffect>())) { 1280 LLVM_DEBUG(dbgs() << "EarlyCSE skipping sideeffect: " << Inst << '\n'); 1281 continue; 1282 } 1283 1284 // Skip pseudoprobe intrinsics, for the same reason as assume intrinsics. 1285 if (match(&Inst, m_Intrinsic<Intrinsic::pseudoprobe>())) { 1286 LLVM_DEBUG(dbgs() << "EarlyCSE skipping pseudoprobe: " << Inst << '\n'); 1287 continue; 1288 } 1289 1290 // We can skip all invariant.start intrinsics since they only read memory, 1291 // and we can forward values across it. For invariant starts without 1292 // invariant ends, we can use the fact that the invariantness never ends to 1293 // start a scope in the current generaton which is true for all future 1294 // generations. Also, we dont need to consume the last store since the 1295 // semantics of invariant.start allow us to perform DSE of the last 1296 // store, if there was a store following invariant.start. Consider: 1297 // 1298 // store 30, i8* p 1299 // invariant.start(p) 1300 // store 40, i8* p 1301 // We can DSE the store to 30, since the store 40 to invariant location p 1302 // causes undefined behaviour. 1303 if (match(&Inst, m_Intrinsic<Intrinsic::invariant_start>())) { 1304 // If there are any uses, the scope might end. 1305 if (!Inst.use_empty()) 1306 continue; 1307 MemoryLocation MemLoc = 1308 MemoryLocation::getForArgument(&cast<CallInst>(Inst), 1, TLI); 1309 // Don't start a scope if we already have a better one pushed 1310 if (!AvailableInvariants.count(MemLoc)) 1311 AvailableInvariants.insert(MemLoc, CurrentGeneration); 1312 continue; 1313 } 1314 1315 if (isGuard(&Inst)) { 1316 if (auto *CondI = 1317 dyn_cast<Instruction>(cast<CallInst>(Inst).getArgOperand(0))) { 1318 if (SimpleValue::canHandle(CondI)) { 1319 // Do we already know the actual value of this condition? 1320 if (auto *KnownCond = AvailableValues.lookup(CondI)) { 1321 // Is the condition known to be true? 1322 if (isa<ConstantInt>(KnownCond) && 1323 cast<ConstantInt>(KnownCond)->isOne()) { 1324 LLVM_DEBUG(dbgs() 1325 << "EarlyCSE removing guard: " << Inst << '\n'); 1326 salvageKnowledge(&Inst, &AC); 1327 removeMSSA(Inst); 1328 Inst.eraseFromParent(); 1329 Changed = true; 1330 continue; 1331 } else 1332 // Use the known value if it wasn't true. 1333 cast<CallInst>(Inst).setArgOperand(0, KnownCond); 1334 } 1335 // The condition we're on guarding here is true for all dominated 1336 // locations. 1337 AvailableValues.insert(CondI, ConstantInt::getTrue(BB->getContext())); 1338 } 1339 } 1340 1341 // Guard intrinsics read all memory, but don't write any memory. 1342 // Accordingly, don't update the generation but consume the last store (to 1343 // avoid an incorrect DSE). 1344 LastStore = nullptr; 1345 continue; 1346 } 1347 1348 // If the instruction can be simplified (e.g. X+0 = X) then replace it with 1349 // its simpler value. 1350 if (Value *V = simplifyInstruction(&Inst, SQ)) { 1351 LLVM_DEBUG(dbgs() << "EarlyCSE Simplify: " << Inst << " to: " << *V 1352 << '\n'); 1353 if (!DebugCounter::shouldExecute(CSECounter)) { 1354 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1355 } else { 1356 bool Killed = false; 1357 if (!Inst.use_empty()) { 1358 Inst.replaceAllUsesWith(V); 1359 Changed = true; 1360 } 1361 if (isInstructionTriviallyDead(&Inst, &TLI)) { 1362 salvageKnowledge(&Inst, &AC); 1363 removeMSSA(Inst); 1364 Inst.eraseFromParent(); 1365 Changed = true; 1366 Killed = true; 1367 } 1368 if (Changed) 1369 ++NumSimplify; 1370 if (Killed) 1371 continue; 1372 } 1373 } 1374 1375 // If this is a simple instruction that we can value number, process it. 1376 if (SimpleValue::canHandle(&Inst)) { 1377 // See if the instruction has an available value. If so, use it. 1378 if (Value *V = AvailableValues.lookup(&Inst)) { 1379 LLVM_DEBUG(dbgs() << "EarlyCSE CSE: " << Inst << " to: " << *V 1380 << '\n'); 1381 if (!DebugCounter::shouldExecute(CSECounter)) { 1382 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1383 continue; 1384 } 1385 if (auto *I = dyn_cast<Instruction>(V)) { 1386 // If I being poison triggers UB, there is no need to drop those 1387 // flags. Otherwise, only retain flags present on both I and Inst. 1388 // TODO: Currently some fast-math flags are not treated as 1389 // poison-generating even though they should. Until this is fixed, 1390 // always retain flags present on both I and Inst for floating point 1391 // instructions. 1392 if (isa<FPMathOperator>(I) || (I->hasPoisonGeneratingFlags() && !programUndefinedIfPoison(I))) 1393 I->andIRFlags(&Inst); 1394 } 1395 Inst.replaceAllUsesWith(V); 1396 salvageKnowledge(&Inst, &AC); 1397 removeMSSA(Inst); 1398 Inst.eraseFromParent(); 1399 Changed = true; 1400 ++NumCSE; 1401 continue; 1402 } 1403 1404 // Otherwise, just remember that this value is available. 1405 AvailableValues.insert(&Inst, &Inst); 1406 continue; 1407 } 1408 1409 ParseMemoryInst MemInst(&Inst, TTI); 1410 // If this is a non-volatile load, process it. 1411 if (MemInst.isValid() && MemInst.isLoad()) { 1412 // (conservatively) we can't peak past the ordering implied by this 1413 // operation, but we can add this load to our set of available values 1414 if (MemInst.isVolatile() || !MemInst.isUnordered()) { 1415 LastStore = nullptr; 1416 ++CurrentGeneration; 1417 } 1418 1419 if (MemInst.isInvariantLoad()) { 1420 // If we pass an invariant load, we know that memory location is 1421 // indefinitely constant from the moment of first dereferenceability. 1422 // We conservatively treat the invariant_load as that moment. If we 1423 // pass a invariant load after already establishing a scope, don't 1424 // restart it since we want to preserve the earliest point seen. 1425 auto MemLoc = MemoryLocation::get(&Inst); 1426 if (!AvailableInvariants.count(MemLoc)) 1427 AvailableInvariants.insert(MemLoc, CurrentGeneration); 1428 } 1429 1430 // If we have an available version of this load, and if it is the right 1431 // generation or the load is known to be from an invariant location, 1432 // replace this instruction. 1433 // 1434 // If either the dominating load or the current load are invariant, then 1435 // we can assume the current load loads the same value as the dominating 1436 // load. 1437 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand()); 1438 if (Value *Op = getMatchingValue(InVal, MemInst, CurrentGeneration)) { 1439 LLVM_DEBUG(dbgs() << "EarlyCSE CSE LOAD: " << Inst 1440 << " to: " << *InVal.DefInst << '\n'); 1441 if (!DebugCounter::shouldExecute(CSECounter)) { 1442 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1443 continue; 1444 } 1445 if (!Inst.use_empty()) 1446 Inst.replaceAllUsesWith(Op); 1447 salvageKnowledge(&Inst, &AC); 1448 removeMSSA(Inst); 1449 Inst.eraseFromParent(); 1450 Changed = true; 1451 ++NumCSELoad; 1452 continue; 1453 } 1454 1455 // Otherwise, remember that we have this instruction. 1456 AvailableLoads.insert(MemInst.getPointerOperand(), 1457 LoadValue(&Inst, CurrentGeneration, 1458 MemInst.getMatchingId(), 1459 MemInst.isAtomic())); 1460 LastStore = nullptr; 1461 continue; 1462 } 1463 1464 // If this instruction may read from memory or throw (and potentially read 1465 // from memory in the exception handler), forget LastStore. Load/store 1466 // intrinsics will indicate both a read and a write to memory. The target 1467 // may override this (e.g. so that a store intrinsic does not read from 1468 // memory, and thus will be treated the same as a regular store for 1469 // commoning purposes). 1470 if ((Inst.mayReadFromMemory() || Inst.mayThrow()) && 1471 !(MemInst.isValid() && !MemInst.mayReadFromMemory())) 1472 LastStore = nullptr; 1473 1474 // If this is a read-only call, process it. 1475 if (CallValue::canHandle(&Inst)) { 1476 // If we have an available version of this call, and if it is the right 1477 // generation, replace this instruction. 1478 std::pair<Instruction *, unsigned> InVal = AvailableCalls.lookup(&Inst); 1479 if (InVal.first != nullptr && 1480 isSameMemGeneration(InVal.second, CurrentGeneration, InVal.first, 1481 &Inst)) { 1482 LLVM_DEBUG(dbgs() << "EarlyCSE CSE CALL: " << Inst 1483 << " to: " << *InVal.first << '\n'); 1484 if (!DebugCounter::shouldExecute(CSECounter)) { 1485 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1486 continue; 1487 } 1488 if (!Inst.use_empty()) 1489 Inst.replaceAllUsesWith(InVal.first); 1490 salvageKnowledge(&Inst, &AC); 1491 removeMSSA(Inst); 1492 Inst.eraseFromParent(); 1493 Changed = true; 1494 ++NumCSECall; 1495 continue; 1496 } 1497 1498 // Otherwise, remember that we have this instruction. 1499 AvailableCalls.insert(&Inst, std::make_pair(&Inst, CurrentGeneration)); 1500 continue; 1501 } 1502 1503 // A release fence requires that all stores complete before it, but does 1504 // not prevent the reordering of following loads 'before' the fence. As a 1505 // result, we don't need to consider it as writing to memory and don't need 1506 // to advance the generation. We do need to prevent DSE across the fence, 1507 // but that's handled above. 1508 if (auto *FI = dyn_cast<FenceInst>(&Inst)) 1509 if (FI->getOrdering() == AtomicOrdering::Release) { 1510 assert(Inst.mayReadFromMemory() && "relied on to prevent DSE above"); 1511 continue; 1512 } 1513 1514 // write back DSE - If we write back the same value we just loaded from 1515 // the same location and haven't passed any intervening writes or ordering 1516 // operations, we can remove the write. The primary benefit is in allowing 1517 // the available load table to remain valid and value forward past where 1518 // the store originally was. 1519 if (MemInst.isValid() && MemInst.isStore()) { 1520 LoadValue InVal = AvailableLoads.lookup(MemInst.getPointerOperand()); 1521 if (InVal.DefInst && 1522 InVal.DefInst == getMatchingValue(InVal, MemInst, CurrentGeneration)) { 1523 // It is okay to have a LastStore to a different pointer here if MemorySSA 1524 // tells us that the load and store are from the same memory generation. 1525 // In that case, LastStore should keep its present value since we're 1526 // removing the current store. 1527 assert((!LastStore || 1528 ParseMemoryInst(LastStore, TTI).getPointerOperand() == 1529 MemInst.getPointerOperand() || 1530 MSSA) && 1531 "can't have an intervening store if not using MemorySSA!"); 1532 LLVM_DEBUG(dbgs() << "EarlyCSE DSE (writeback): " << Inst << '\n'); 1533 if (!DebugCounter::shouldExecute(CSECounter)) { 1534 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1535 continue; 1536 } 1537 salvageKnowledge(&Inst, &AC); 1538 removeMSSA(Inst); 1539 Inst.eraseFromParent(); 1540 Changed = true; 1541 ++NumDSE; 1542 // We can avoid incrementing the generation count since we were able 1543 // to eliminate this store. 1544 continue; 1545 } 1546 } 1547 1548 // Okay, this isn't something we can CSE at all. Check to see if it is 1549 // something that could modify memory. If so, our available memory values 1550 // cannot be used so bump the generation count. 1551 if (Inst.mayWriteToMemory()) { 1552 ++CurrentGeneration; 1553 1554 if (MemInst.isValid() && MemInst.isStore()) { 1555 // We do a trivial form of DSE if there are two stores to the same 1556 // location with no intervening loads. Delete the earlier store. 1557 if (LastStore) { 1558 if (overridingStores(ParseMemoryInst(LastStore, TTI), MemInst)) { 1559 LLVM_DEBUG(dbgs() << "EarlyCSE DEAD STORE: " << *LastStore 1560 << " due to: " << Inst << '\n'); 1561 if (!DebugCounter::shouldExecute(CSECounter)) { 1562 LLVM_DEBUG(dbgs() << "Skipping due to debug counter\n"); 1563 } else { 1564 salvageKnowledge(&Inst, &AC); 1565 removeMSSA(*LastStore); 1566 LastStore->eraseFromParent(); 1567 Changed = true; 1568 ++NumDSE; 1569 LastStore = nullptr; 1570 } 1571 } 1572 // fallthrough - we can exploit information about this store 1573 } 1574 1575 // Okay, we just invalidated anything we knew about loaded values. Try 1576 // to salvage *something* by remembering that the stored value is a live 1577 // version of the pointer. It is safe to forward from volatile stores 1578 // to non-volatile loads, so we don't have to check for volatility of 1579 // the store. 1580 AvailableLoads.insert(MemInst.getPointerOperand(), 1581 LoadValue(&Inst, CurrentGeneration, 1582 MemInst.getMatchingId(), 1583 MemInst.isAtomic())); 1584 1585 // Remember that this was the last unordered store we saw for DSE. We 1586 // don't yet handle DSE on ordered or volatile stores since we don't 1587 // have a good way to model the ordering requirement for following 1588 // passes once the store is removed. We could insert a fence, but 1589 // since fences are slightly stronger than stores in their ordering, 1590 // it's not clear this is a profitable transform. Another option would 1591 // be to merge the ordering with that of the post dominating store. 1592 if (MemInst.isUnordered() && !MemInst.isVolatile()) 1593 LastStore = &Inst; 1594 else 1595 LastStore = nullptr; 1596 } 1597 } 1598 } 1599 1600 return Changed; 1601 } 1602 1603 bool EarlyCSE::run() { 1604 // Note, deque is being used here because there is significant performance 1605 // gains over vector when the container becomes very large due to the 1606 // specific access patterns. For more information see the mailing list 1607 // discussion on this: 1608 // http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20120116/135228.html 1609 std::deque<StackNode *> nodesToProcess; 1610 1611 bool Changed = false; 1612 1613 // Process the root node. 1614 nodesToProcess.push_back(new StackNode( 1615 AvailableValues, AvailableLoads, AvailableInvariants, AvailableCalls, 1616 CurrentGeneration, DT.getRootNode(), 1617 DT.getRootNode()->begin(), DT.getRootNode()->end())); 1618 1619 assert(!CurrentGeneration && "Create a new EarlyCSE instance to rerun it."); 1620 1621 // Process the stack. 1622 while (!nodesToProcess.empty()) { 1623 // Grab the first item off the stack. Set the current generation, remove 1624 // the node from the stack, and process it. 1625 StackNode *NodeToProcess = nodesToProcess.back(); 1626 1627 // Initialize class members. 1628 CurrentGeneration = NodeToProcess->currentGeneration(); 1629 1630 // Check if the node needs to be processed. 1631 if (!NodeToProcess->isProcessed()) { 1632 // Process the node. 1633 Changed |= processNode(NodeToProcess->node()); 1634 NodeToProcess->childGeneration(CurrentGeneration); 1635 NodeToProcess->process(); 1636 } else if (NodeToProcess->childIter() != NodeToProcess->end()) { 1637 // Push the next child onto the stack. 1638 DomTreeNode *child = NodeToProcess->nextChild(); 1639 nodesToProcess.push_back( 1640 new StackNode(AvailableValues, AvailableLoads, AvailableInvariants, 1641 AvailableCalls, NodeToProcess->childGeneration(), 1642 child, child->begin(), child->end())); 1643 } else { 1644 // It has been processed, and there are no more children to process, 1645 // so delete it and pop it off the stack. 1646 delete NodeToProcess; 1647 nodesToProcess.pop_back(); 1648 } 1649 } // while (!nodes...) 1650 1651 return Changed; 1652 } 1653 1654 PreservedAnalyses EarlyCSEPass::run(Function &F, 1655 FunctionAnalysisManager &AM) { 1656 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F); 1657 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 1658 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 1659 auto &AC = AM.getResult<AssumptionAnalysis>(F); 1660 auto *MSSA = 1661 UseMemorySSA ? &AM.getResult<MemorySSAAnalysis>(F).getMSSA() : nullptr; 1662 1663 EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA); 1664 1665 if (!CSE.run()) 1666 return PreservedAnalyses::all(); 1667 1668 PreservedAnalyses PA; 1669 PA.preserveSet<CFGAnalyses>(); 1670 if (UseMemorySSA) 1671 PA.preserve<MemorySSAAnalysis>(); 1672 return PA; 1673 } 1674 1675 void EarlyCSEPass::printPipeline( 1676 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) { 1677 static_cast<PassInfoMixin<EarlyCSEPass> *>(this)->printPipeline( 1678 OS, MapClassName2PassName); 1679 OS << "<"; 1680 if (UseMemorySSA) 1681 OS << "memssa"; 1682 OS << ">"; 1683 } 1684 1685 namespace { 1686 1687 /// A simple and fast domtree-based CSE pass. 1688 /// 1689 /// This pass does a simple depth-first walk over the dominator tree, 1690 /// eliminating trivially redundant instructions and using instsimplify to 1691 /// canonicalize things as it goes. It is intended to be fast and catch obvious 1692 /// cases so that instcombine and other passes are more effective. It is 1693 /// expected that a later pass of GVN will catch the interesting/hard cases. 1694 template<bool UseMemorySSA> 1695 class EarlyCSELegacyCommonPass : public FunctionPass { 1696 public: 1697 static char ID; 1698 1699 EarlyCSELegacyCommonPass() : FunctionPass(ID) { 1700 if (UseMemorySSA) 1701 initializeEarlyCSEMemSSALegacyPassPass(*PassRegistry::getPassRegistry()); 1702 else 1703 initializeEarlyCSELegacyPassPass(*PassRegistry::getPassRegistry()); 1704 } 1705 1706 bool runOnFunction(Function &F) override { 1707 if (skipFunction(F)) 1708 return false; 1709 1710 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F); 1711 auto &TTI = getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F); 1712 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 1713 auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 1714 auto *MSSA = 1715 UseMemorySSA ? &getAnalysis<MemorySSAWrapperPass>().getMSSA() : nullptr; 1716 1717 EarlyCSE CSE(F.getParent()->getDataLayout(), TLI, TTI, DT, AC, MSSA); 1718 1719 return CSE.run(); 1720 } 1721 1722 void getAnalysisUsage(AnalysisUsage &AU) const override { 1723 AU.addRequired<AssumptionCacheTracker>(); 1724 AU.addRequired<DominatorTreeWrapperPass>(); 1725 AU.addRequired<TargetLibraryInfoWrapperPass>(); 1726 AU.addRequired<TargetTransformInfoWrapperPass>(); 1727 if (UseMemorySSA) { 1728 AU.addRequired<AAResultsWrapperPass>(); 1729 AU.addRequired<MemorySSAWrapperPass>(); 1730 AU.addPreserved<MemorySSAWrapperPass>(); 1731 } 1732 AU.addPreserved<GlobalsAAWrapperPass>(); 1733 AU.addPreserved<AAResultsWrapperPass>(); 1734 AU.setPreservesCFG(); 1735 } 1736 }; 1737 1738 } // end anonymous namespace 1739 1740 using EarlyCSELegacyPass = EarlyCSELegacyCommonPass</*UseMemorySSA=*/false>; 1741 1742 template<> 1743 char EarlyCSELegacyPass::ID = 0; 1744 1745 INITIALIZE_PASS_BEGIN(EarlyCSELegacyPass, "early-cse", "Early CSE", false, 1746 false) 1747 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 1748 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 1749 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 1750 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1751 INITIALIZE_PASS_END(EarlyCSELegacyPass, "early-cse", "Early CSE", false, false) 1752 1753 using EarlyCSEMemSSALegacyPass = 1754 EarlyCSELegacyCommonPass</*UseMemorySSA=*/true>; 1755 1756 template<> 1757 char EarlyCSEMemSSALegacyPass::ID = 0; 1758 1759 FunctionPass *llvm::createEarlyCSEPass(bool UseMemorySSA) { 1760 if (UseMemorySSA) 1761 return new EarlyCSEMemSSALegacyPass(); 1762 else 1763 return new EarlyCSELegacyPass(); 1764 } 1765 1766 INITIALIZE_PASS_BEGIN(EarlyCSEMemSSALegacyPass, "early-cse-memssa", 1767 "Early CSE w/ MemorySSA", false, false) 1768 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 1769 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 1770 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass) 1771 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 1772 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1773 INITIALIZE_PASS_DEPENDENCY(MemorySSAWrapperPass) 1774 INITIALIZE_PASS_END(EarlyCSEMemSSALegacyPass, "early-cse-memssa", 1775 "Early CSE w/ MemorySSA", false, false) 1776