1 //===- LoopFlatten.cpp - Loop flattening 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 flattens pairs nested loops into a single loop. 10 // 11 // The intention is to optimise loop nests like this, which together access an 12 // array linearly: 13 // 14 // for (int i = 0; i < N; ++i) 15 // for (int j = 0; j < M; ++j) 16 // f(A[i*M+j]); 17 // 18 // into one loop: 19 // 20 // for (int i = 0; i < (N*M); ++i) 21 // f(A[i]); 22 // 23 // It can also flatten loops where the induction variables are not used in the 24 // loop. This is only worth doing if the induction variables are only used in an 25 // expression like i*M+j. If they had any other uses, we would have to insert a 26 // div/mod to reconstruct the original values, so this wouldn't be profitable. 27 // 28 // We also need to prove that N*M will not overflow. The preferred solution is 29 // to widen the IV, which avoids overflow checks, so that is tried first. If 30 // the IV cannot be widened, then we try to determine that this new tripcount 31 // expression won't overflow. 32 // 33 // Q: Does LoopFlatten use SCEV? 34 // Short answer: Yes and no. 35 // 36 // Long answer: 37 // For this transformation to be valid, we require all uses of the induction 38 // variables to be linear expressions of the form i*M+j. The different Loop 39 // APIs are used to get some loop components like the induction variable, 40 // compare statement, etc. In addition, we do some pattern matching to find the 41 // linear expressions and other loop components like the loop increment. The 42 // latter are examples of expressions that do use the induction variable, but 43 // are safe to ignore when we check all uses to be of the form i*M+j. We keep 44 // track of all of this in bookkeeping struct FlattenInfo. 45 // We assume the loops to be canonical, i.e. starting at 0 and increment with 46 // 1. This makes RHS of the compare the loop tripcount (with the right 47 // predicate). We use SCEV to then sanity check that this tripcount matches 48 // with the tripcount as computed by SCEV. 49 // 50 //===----------------------------------------------------------------------===// 51 52 #include "llvm/Transforms/Scalar/LoopFlatten.h" 53 54 #include "llvm/ADT/Statistic.h" 55 #include "llvm/Analysis/AssumptionCache.h" 56 #include "llvm/Analysis/LoopInfo.h" 57 #include "llvm/Analysis/LoopNestAnalysis.h" 58 #include "llvm/Analysis/MemorySSAUpdater.h" 59 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 60 #include "llvm/Analysis/ScalarEvolution.h" 61 #include "llvm/Analysis/TargetTransformInfo.h" 62 #include "llvm/Analysis/ValueTracking.h" 63 #include "llvm/IR/Dominators.h" 64 #include "llvm/IR/Function.h" 65 #include "llvm/IR/IRBuilder.h" 66 #include "llvm/IR/Module.h" 67 #include "llvm/IR/PatternMatch.h" 68 #include "llvm/InitializePasses.h" 69 #include "llvm/Pass.h" 70 #include "llvm/Support/Debug.h" 71 #include "llvm/Support/raw_ostream.h" 72 #include "llvm/Transforms/Scalar.h" 73 #include "llvm/Transforms/Scalar/LoopPassManager.h" 74 #include "llvm/Transforms/Utils/Local.h" 75 #include "llvm/Transforms/Utils/LoopUtils.h" 76 #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" 77 #include "llvm/Transforms/Utils/SimplifyIndVar.h" 78 79 using namespace llvm; 80 using namespace llvm::PatternMatch; 81 82 #define DEBUG_TYPE "loop-flatten" 83 84 STATISTIC(NumFlattened, "Number of loops flattened"); 85 86 static cl::opt<unsigned> RepeatedInstructionThreshold( 87 "loop-flatten-cost-threshold", cl::Hidden, cl::init(2), 88 cl::desc("Limit on the cost of instructions that can be repeated due to " 89 "loop flattening")); 90 91 static cl::opt<bool> 92 AssumeNoOverflow("loop-flatten-assume-no-overflow", cl::Hidden, 93 cl::init(false), 94 cl::desc("Assume that the product of the two iteration " 95 "trip counts will never overflow")); 96 97 static cl::opt<bool> 98 WidenIV("loop-flatten-widen-iv", cl::Hidden, cl::init(true), 99 cl::desc("Widen the loop induction variables, if possible, so " 100 "overflow checks won't reject flattening")); 101 102 // We require all uses of both induction variables to match this pattern: 103 // 104 // (OuterPHI * InnerTripCount) + InnerPHI 105 // 106 // I.e., it needs to be a linear expression of the induction variables and the 107 // inner loop trip count. We keep track of all different expressions on which 108 // checks will be performed in this bookkeeping struct. 109 // 110 struct FlattenInfo { 111 Loop *OuterLoop = nullptr; // The loop pair to be flattened. 112 Loop *InnerLoop = nullptr; 113 114 PHINode *InnerInductionPHI = nullptr; // These PHINodes correspond to loop 115 PHINode *OuterInductionPHI = nullptr; // induction variables, which are 116 // expected to start at zero and 117 // increment by one on each loop. 118 119 Value *InnerTripCount = nullptr; // The product of these two tripcounts 120 Value *OuterTripCount = nullptr; // will be the new flattened loop 121 // tripcount. Also used to recognise a 122 // linear expression that will be replaced. 123 124 SmallPtrSet<Value *, 4> LinearIVUses; // Contains the linear expressions 125 // of the form i*M+j that will be 126 // replaced. 127 128 BinaryOperator *InnerIncrement = nullptr; // Uses of induction variables in 129 BinaryOperator *OuterIncrement = nullptr; // loop control statements that 130 BranchInst *InnerBranch = nullptr; // are safe to ignore. 131 132 BranchInst *OuterBranch = nullptr; // The instruction that needs to be 133 // updated with new tripcount. 134 135 SmallPtrSet<PHINode *, 4> InnerPHIsToTransform; 136 137 bool Widened = false; // Whether this holds the flatten info before or after 138 // widening. 139 140 PHINode *NarrowInnerInductionPHI = nullptr; // Holds the old/narrow induction 141 PHINode *NarrowOuterInductionPHI = nullptr; // phis, i.e. the Phis before IV 142 // has been apllied. Used to skip 143 // checks on phi nodes. 144 145 FlattenInfo(Loop *OL, Loop *IL) : OuterLoop(OL), InnerLoop(IL){}; 146 147 bool isNarrowInductionPhi(PHINode *Phi) { 148 // This can't be the narrow phi if we haven't widened the IV first. 149 if (!Widened) 150 return false; 151 return NarrowInnerInductionPHI == Phi || NarrowOuterInductionPHI == Phi; 152 } 153 bool isInnerLoopIncrement(User *U) { 154 return InnerIncrement == U; 155 } 156 bool isOuterLoopIncrement(User *U) { 157 return OuterIncrement == U; 158 } 159 bool isInnerLoopTest(User *U) { 160 return InnerBranch->getCondition() == U; 161 } 162 163 bool checkOuterInductionPhiUsers(SmallPtrSet<Value *, 4> &ValidOuterPHIUses) { 164 for (User *U : OuterInductionPHI->users()) { 165 if (isOuterLoopIncrement(U)) 166 continue; 167 168 auto IsValidOuterPHIUses = [&] (User *U) -> bool { 169 LLVM_DEBUG(dbgs() << "Found use of outer induction variable: "; U->dump()); 170 if (!ValidOuterPHIUses.count(U)) { 171 LLVM_DEBUG(dbgs() << "Did not match expected pattern, bailing\n"); 172 return false; 173 } 174 LLVM_DEBUG(dbgs() << "Use is optimisable\n"); 175 return true; 176 }; 177 178 if (auto *V = dyn_cast<TruncInst>(U)) { 179 for (auto *K : V->users()) { 180 if (!IsValidOuterPHIUses(K)) 181 return false; 182 } 183 continue; 184 } 185 186 if (!IsValidOuterPHIUses(U)) 187 return false; 188 } 189 return true; 190 } 191 192 bool matchLinearIVUser(User *U, Value *InnerTripCount, 193 SmallPtrSet<Value *, 4> &ValidOuterPHIUses) { 194 LLVM_DEBUG(dbgs() << "Found use of inner induction variable: "; U->dump()); 195 Value *MatchedMul = nullptr; 196 Value *MatchedItCount = nullptr; 197 198 bool IsAdd = match(U, m_c_Add(m_Specific(InnerInductionPHI), 199 m_Value(MatchedMul))) && 200 match(MatchedMul, m_c_Mul(m_Specific(OuterInductionPHI), 201 m_Value(MatchedItCount))); 202 203 // Matches the same pattern as above, except it also looks for truncs 204 // on the phi, which can be the result of widening the induction variables. 205 bool IsAddTrunc = 206 match(U, m_c_Add(m_Trunc(m_Specific(InnerInductionPHI)), 207 m_Value(MatchedMul))) && 208 match(MatchedMul, m_c_Mul(m_Trunc(m_Specific(OuterInductionPHI)), 209 m_Value(MatchedItCount))); 210 211 if (!MatchedItCount) 212 return false; 213 214 // Look through extends if the IV has been widened. Don't look through 215 // extends if we already looked through a trunc. 216 if (Widened && IsAdd && 217 (isa<SExtInst>(MatchedItCount) || isa<ZExtInst>(MatchedItCount))) { 218 assert(MatchedItCount->getType() == InnerInductionPHI->getType() && 219 "Unexpected type mismatch in types after widening"); 220 MatchedItCount = isa<SExtInst>(MatchedItCount) 221 ? dyn_cast<SExtInst>(MatchedItCount)->getOperand(0) 222 : dyn_cast<ZExtInst>(MatchedItCount)->getOperand(0); 223 } 224 225 if ((IsAdd || IsAddTrunc) && MatchedItCount == InnerTripCount) { 226 LLVM_DEBUG(dbgs() << "Use is optimisable\n"); 227 ValidOuterPHIUses.insert(MatchedMul); 228 LinearIVUses.insert(U); 229 return true; 230 } 231 232 LLVM_DEBUG(dbgs() << "Did not match expected pattern, bailing\n"); 233 return false; 234 } 235 236 bool checkInnerInductionPhiUsers(SmallPtrSet<Value *, 4> &ValidOuterPHIUses) { 237 Value *SExtInnerTripCount = InnerTripCount; 238 if (Widened && 239 (isa<SExtInst>(InnerTripCount) || isa<ZExtInst>(InnerTripCount))) 240 SExtInnerTripCount = cast<Instruction>(InnerTripCount)->getOperand(0); 241 242 for (User *U : InnerInductionPHI->users()) { 243 if (isInnerLoopIncrement(U)) 244 continue; 245 246 // After widening the IVs, a trunc instruction might have been introduced, 247 // so look through truncs. 248 if (isa<TruncInst>(U)) { 249 if (!U->hasOneUse()) 250 return false; 251 U = *U->user_begin(); 252 } 253 254 // If the use is in the compare (which is also the condition of the inner 255 // branch) then the compare has been altered by another transformation e.g 256 // icmp ult %inc, tripcount -> icmp ult %j, tripcount-1, where tripcount is 257 // a constant. Ignore this use as the compare gets removed later anyway. 258 if (isInnerLoopTest(U)) 259 continue; 260 261 if (!matchLinearIVUser(U, SExtInnerTripCount, ValidOuterPHIUses)) 262 return false; 263 } 264 return true; 265 } 266 }; 267 268 static bool 269 setLoopComponents(Value *&TC, Value *&TripCount, BinaryOperator *&Increment, 270 SmallPtrSetImpl<Instruction *> &IterationInstructions) { 271 TripCount = TC; 272 IterationInstructions.insert(Increment); 273 LLVM_DEBUG(dbgs() << "Found Increment: "; Increment->dump()); 274 LLVM_DEBUG(dbgs() << "Found trip count: "; TripCount->dump()); 275 LLVM_DEBUG(dbgs() << "Successfully found all loop components\n"); 276 return true; 277 } 278 279 // Given the RHS of the loop latch compare instruction, verify with SCEV 280 // that this is indeed the loop tripcount. 281 // TODO: This used to be a straightforward check but has grown to be quite 282 // complicated now. It is therefore worth revisiting what the additional 283 // benefits are of this (compared to relying on canonical loops and pattern 284 // matching). 285 static bool verifyTripCount(Value *RHS, Loop *L, 286 SmallPtrSetImpl<Instruction *> &IterationInstructions, 287 PHINode *&InductionPHI, Value *&TripCount, BinaryOperator *&Increment, 288 BranchInst *&BackBranch, ScalarEvolution *SE, bool IsWidened) { 289 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 290 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { 291 LLVM_DEBUG(dbgs() << "Backedge-taken count is not predictable\n"); 292 return false; 293 } 294 295 // The Extend=false flag is used for getTripCountFromExitCount as we want 296 // to verify and match it with the pattern matched tripcount. Please note 297 // that overflow checks are performed in checkOverflow, but are first tried 298 // to avoid by widening the IV. 299 const SCEV *SCEVTripCount = 300 SE->getTripCountFromExitCount(BackedgeTakenCount, /*Extend=*/false); 301 302 const SCEV *SCEVRHS = SE->getSCEV(RHS); 303 if (SCEVRHS == SCEVTripCount) 304 return setLoopComponents(RHS, TripCount, Increment, IterationInstructions); 305 ConstantInt *ConstantRHS = dyn_cast<ConstantInt>(RHS); 306 if (ConstantRHS) { 307 const SCEV *BackedgeTCExt = nullptr; 308 if (IsWidened) { 309 const SCEV *SCEVTripCountExt; 310 // Find the extended backedge taken count and extended trip count using 311 // SCEV. One of these should now match the RHS of the compare. 312 BackedgeTCExt = SE->getZeroExtendExpr(BackedgeTakenCount, RHS->getType()); 313 SCEVTripCountExt = SE->getTripCountFromExitCount(BackedgeTCExt, false); 314 if (SCEVRHS != BackedgeTCExt && SCEVRHS != SCEVTripCountExt) { 315 LLVM_DEBUG(dbgs() << "Could not find valid trip count\n"); 316 return false; 317 } 318 } 319 // If the RHS of the compare is equal to the backedge taken count we need 320 // to add one to get the trip count. 321 if (SCEVRHS == BackedgeTCExt || SCEVRHS == BackedgeTakenCount) { 322 ConstantInt *One = ConstantInt::get(ConstantRHS->getType(), 1); 323 Value *NewRHS = ConstantInt::get( 324 ConstantRHS->getContext(), ConstantRHS->getValue() + One->getValue()); 325 return setLoopComponents(NewRHS, TripCount, Increment, 326 IterationInstructions); 327 } 328 return setLoopComponents(RHS, TripCount, Increment, IterationInstructions); 329 } 330 // If the RHS isn't a constant then check that the reason it doesn't match 331 // the SCEV trip count is because the RHS is a ZExt or SExt instruction 332 // (and take the trip count to be the RHS). 333 if (!IsWidened) { 334 LLVM_DEBUG(dbgs() << "Could not find valid trip count\n"); 335 return false; 336 } 337 auto *TripCountInst = dyn_cast<Instruction>(RHS); 338 if (!TripCountInst) { 339 LLVM_DEBUG(dbgs() << "Could not find valid trip count\n"); 340 return false; 341 } 342 if ((!isa<ZExtInst>(TripCountInst) && !isa<SExtInst>(TripCountInst)) || 343 SE->getSCEV(TripCountInst->getOperand(0)) != SCEVTripCount) { 344 LLVM_DEBUG(dbgs() << "Could not find valid extended trip count\n"); 345 return false; 346 } 347 return setLoopComponents(RHS, TripCount, Increment, IterationInstructions); 348 } 349 350 // Finds the induction variable, increment and trip count for a simple loop that 351 // we can flatten. 352 static bool findLoopComponents( 353 Loop *L, SmallPtrSetImpl<Instruction *> &IterationInstructions, 354 PHINode *&InductionPHI, Value *&TripCount, BinaryOperator *&Increment, 355 BranchInst *&BackBranch, ScalarEvolution *SE, bool IsWidened) { 356 LLVM_DEBUG(dbgs() << "Finding components of loop: " << L->getName() << "\n"); 357 358 if (!L->isLoopSimplifyForm()) { 359 LLVM_DEBUG(dbgs() << "Loop is not in normal form\n"); 360 return false; 361 } 362 363 // Currently, to simplify the implementation, the Loop induction variable must 364 // start at zero and increment with a step size of one. 365 if (!L->isCanonical(*SE)) { 366 LLVM_DEBUG(dbgs() << "Loop is not canonical\n"); 367 return false; 368 } 369 370 // There must be exactly one exiting block, and it must be the same at the 371 // latch. 372 BasicBlock *Latch = L->getLoopLatch(); 373 if (L->getExitingBlock() != Latch) { 374 LLVM_DEBUG(dbgs() << "Exiting and latch block are different\n"); 375 return false; 376 } 377 378 // Find the induction PHI. If there is no induction PHI, we can't do the 379 // transformation. TODO: could other variables trigger this? Do we have to 380 // search for the best one? 381 InductionPHI = L->getInductionVariable(*SE); 382 if (!InductionPHI) { 383 LLVM_DEBUG(dbgs() << "Could not find induction PHI\n"); 384 return false; 385 } 386 LLVM_DEBUG(dbgs() << "Found induction PHI: "; InductionPHI->dump()); 387 388 bool ContinueOnTrue = L->contains(Latch->getTerminator()->getSuccessor(0)); 389 auto IsValidPredicate = [&](ICmpInst::Predicate Pred) { 390 if (ContinueOnTrue) 391 return Pred == CmpInst::ICMP_NE || Pred == CmpInst::ICMP_ULT; 392 else 393 return Pred == CmpInst::ICMP_EQ; 394 }; 395 396 // Find Compare and make sure it is valid. getLatchCmpInst checks that the 397 // back branch of the latch is conditional. 398 ICmpInst *Compare = L->getLatchCmpInst(); 399 if (!Compare || !IsValidPredicate(Compare->getUnsignedPredicate()) || 400 Compare->hasNUsesOrMore(2)) { 401 LLVM_DEBUG(dbgs() << "Could not find valid comparison\n"); 402 return false; 403 } 404 BackBranch = cast<BranchInst>(Latch->getTerminator()); 405 IterationInstructions.insert(BackBranch); 406 LLVM_DEBUG(dbgs() << "Found back branch: "; BackBranch->dump()); 407 IterationInstructions.insert(Compare); 408 LLVM_DEBUG(dbgs() << "Found comparison: "; Compare->dump()); 409 410 // Find increment and trip count. 411 // There are exactly 2 incoming values to the induction phi; one from the 412 // pre-header and one from the latch. The incoming latch value is the 413 // increment variable. 414 Increment = 415 cast<BinaryOperator>(InductionPHI->getIncomingValueForBlock(Latch)); 416 if (Increment->hasNUsesOrMore(3)) { 417 LLVM_DEBUG(dbgs() << "Could not find valid increment\n"); 418 return false; 419 } 420 // The trip count is the RHS of the compare. If this doesn't match the trip 421 // count computed by SCEV then this is because the trip count variable 422 // has been widened so the types don't match, or because it is a constant and 423 // another transformation has changed the compare (e.g. icmp ult %inc, 424 // tripcount -> icmp ult %j, tripcount-1), or both. 425 Value *RHS = Compare->getOperand(1); 426 427 return verifyTripCount(RHS, L, IterationInstructions, InductionPHI, TripCount, 428 Increment, BackBranch, SE, IsWidened); 429 } 430 431 static bool checkPHIs(FlattenInfo &FI, const TargetTransformInfo *TTI) { 432 // All PHIs in the inner and outer headers must either be: 433 // - The induction PHI, which we are going to rewrite as one induction in 434 // the new loop. This is already checked by findLoopComponents. 435 // - An outer header PHI with all incoming values from outside the loop. 436 // LoopSimplify guarantees we have a pre-header, so we don't need to 437 // worry about that here. 438 // - Pairs of PHIs in the inner and outer headers, which implement a 439 // loop-carried dependency that will still be valid in the new loop. To 440 // be valid, this variable must be modified only in the inner loop. 441 442 // The set of PHI nodes in the outer loop header that we know will still be 443 // valid after the transformation. These will not need to be modified (with 444 // the exception of the induction variable), but we do need to check that 445 // there are no unsafe PHI nodes. 446 SmallPtrSet<PHINode *, 4> SafeOuterPHIs; 447 SafeOuterPHIs.insert(FI.OuterInductionPHI); 448 449 // Check that all PHI nodes in the inner loop header match one of the valid 450 // patterns. 451 for (PHINode &InnerPHI : FI.InnerLoop->getHeader()->phis()) { 452 // The induction PHIs break these rules, and that's OK because we treat 453 // them specially when doing the transformation. 454 if (&InnerPHI == FI.InnerInductionPHI) 455 continue; 456 if (FI.isNarrowInductionPhi(&InnerPHI)) 457 continue; 458 459 // Each inner loop PHI node must have two incoming values/blocks - one 460 // from the pre-header, and one from the latch. 461 assert(InnerPHI.getNumIncomingValues() == 2); 462 Value *PreHeaderValue = 463 InnerPHI.getIncomingValueForBlock(FI.InnerLoop->getLoopPreheader()); 464 Value *LatchValue = 465 InnerPHI.getIncomingValueForBlock(FI.InnerLoop->getLoopLatch()); 466 467 // The incoming value from the outer loop must be the PHI node in the 468 // outer loop header, with no modifications made in the top of the outer 469 // loop. 470 PHINode *OuterPHI = dyn_cast<PHINode>(PreHeaderValue); 471 if (!OuterPHI || OuterPHI->getParent() != FI.OuterLoop->getHeader()) { 472 LLVM_DEBUG(dbgs() << "value modified in top of outer loop\n"); 473 return false; 474 } 475 476 // The other incoming value must come from the inner loop, without any 477 // modifications in the tail end of the outer loop. We are in LCSSA form, 478 // so this will actually be a PHI in the inner loop's exit block, which 479 // only uses values from inside the inner loop. 480 PHINode *LCSSAPHI = dyn_cast<PHINode>( 481 OuterPHI->getIncomingValueForBlock(FI.OuterLoop->getLoopLatch())); 482 if (!LCSSAPHI) { 483 LLVM_DEBUG(dbgs() << "could not find LCSSA PHI\n"); 484 return false; 485 } 486 487 // The value used by the LCSSA PHI must be the same one that the inner 488 // loop's PHI uses. 489 if (LCSSAPHI->hasConstantValue() != LatchValue) { 490 LLVM_DEBUG( 491 dbgs() << "LCSSA PHI incoming value does not match latch value\n"); 492 return false; 493 } 494 495 LLVM_DEBUG(dbgs() << "PHI pair is safe:\n"); 496 LLVM_DEBUG(dbgs() << " Inner: "; InnerPHI.dump()); 497 LLVM_DEBUG(dbgs() << " Outer: "; OuterPHI->dump()); 498 SafeOuterPHIs.insert(OuterPHI); 499 FI.InnerPHIsToTransform.insert(&InnerPHI); 500 } 501 502 for (PHINode &OuterPHI : FI.OuterLoop->getHeader()->phis()) { 503 if (FI.isNarrowInductionPhi(&OuterPHI)) 504 continue; 505 if (!SafeOuterPHIs.count(&OuterPHI)) { 506 LLVM_DEBUG(dbgs() << "found unsafe PHI in outer loop: "; OuterPHI.dump()); 507 return false; 508 } 509 } 510 511 LLVM_DEBUG(dbgs() << "checkPHIs: OK\n"); 512 return true; 513 } 514 515 static bool 516 checkOuterLoopInsts(FlattenInfo &FI, 517 SmallPtrSetImpl<Instruction *> &IterationInstructions, 518 const TargetTransformInfo *TTI) { 519 // Check for instructions in the outer but not inner loop. If any of these 520 // have side-effects then this transformation is not legal, and if there is 521 // a significant amount of code here which can't be optimised out that it's 522 // not profitable (as these instructions would get executed for each 523 // iteration of the inner loop). 524 InstructionCost RepeatedInstrCost = 0; 525 for (auto *B : FI.OuterLoop->getBlocks()) { 526 if (FI.InnerLoop->contains(B)) 527 continue; 528 529 for (auto &I : *B) { 530 if (!isa<PHINode>(&I) && !I.isTerminator() && 531 !isSafeToSpeculativelyExecute(&I)) { 532 LLVM_DEBUG(dbgs() << "Cannot flatten because instruction may have " 533 "side effects: "; 534 I.dump()); 535 return false; 536 } 537 // The execution count of the outer loop's iteration instructions 538 // (increment, compare and branch) will be increased, but the 539 // equivalent instructions will be removed from the inner loop, so 540 // they make a net difference of zero. 541 if (IterationInstructions.count(&I)) 542 continue; 543 // The uncoditional branch to the inner loop's header will turn into 544 // a fall-through, so adds no cost. 545 BranchInst *Br = dyn_cast<BranchInst>(&I); 546 if (Br && Br->isUnconditional() && 547 Br->getSuccessor(0) == FI.InnerLoop->getHeader()) 548 continue; 549 // Multiplies of the outer iteration variable and inner iteration 550 // count will be optimised out. 551 if (match(&I, m_c_Mul(m_Specific(FI.OuterInductionPHI), 552 m_Specific(FI.InnerTripCount)))) 553 continue; 554 InstructionCost Cost = 555 TTI->getUserCost(&I, TargetTransformInfo::TCK_SizeAndLatency); 556 LLVM_DEBUG(dbgs() << "Cost " << Cost << ": "; I.dump()); 557 RepeatedInstrCost += Cost; 558 } 559 } 560 561 LLVM_DEBUG(dbgs() << "Cost of instructions that will be repeated: " 562 << RepeatedInstrCost << "\n"); 563 // Bail out if flattening the loops would cause instructions in the outer 564 // loop but not in the inner loop to be executed extra times. 565 if (RepeatedInstrCost > RepeatedInstructionThreshold) { 566 LLVM_DEBUG(dbgs() << "checkOuterLoopInsts: not profitable, bailing.\n"); 567 return false; 568 } 569 570 LLVM_DEBUG(dbgs() << "checkOuterLoopInsts: OK\n"); 571 return true; 572 } 573 574 575 576 // We require all uses of both induction variables to match this pattern: 577 // 578 // (OuterPHI * InnerTripCount) + InnerPHI 579 // 580 // Any uses of the induction variables not matching that pattern would 581 // require a div/mod to reconstruct in the flattened loop, so the 582 // transformation wouldn't be profitable. 583 static bool checkIVUsers(FlattenInfo &FI) { 584 // Check that all uses of the inner loop's induction variable match the 585 // expected pattern, recording the uses of the outer IV. 586 SmallPtrSet<Value *, 4> ValidOuterPHIUses; 587 if (!FI.checkInnerInductionPhiUsers(ValidOuterPHIUses)) 588 return false; 589 590 // Check that there are no uses of the outer IV other than the ones found 591 // as part of the pattern above. 592 if (!FI.checkOuterInductionPhiUsers(ValidOuterPHIUses)) 593 return false; 594 595 LLVM_DEBUG(dbgs() << "checkIVUsers: OK\n"; 596 dbgs() << "Found " << FI.LinearIVUses.size() 597 << " value(s) that can be replaced:\n"; 598 for (Value *V : FI.LinearIVUses) { 599 dbgs() << " "; 600 V->dump(); 601 }); 602 return true; 603 } 604 605 // Return an OverflowResult dependant on if overflow of the multiplication of 606 // InnerTripCount and OuterTripCount can be assumed not to happen. 607 static OverflowResult checkOverflow(FlattenInfo &FI, DominatorTree *DT, 608 AssumptionCache *AC) { 609 Function *F = FI.OuterLoop->getHeader()->getParent(); 610 const DataLayout &DL = F->getParent()->getDataLayout(); 611 612 // For debugging/testing. 613 if (AssumeNoOverflow) 614 return OverflowResult::NeverOverflows; 615 616 // Check if the multiply could not overflow due to known ranges of the 617 // input values. 618 OverflowResult OR = computeOverflowForUnsignedMul( 619 FI.InnerTripCount, FI.OuterTripCount, DL, AC, 620 FI.OuterLoop->getLoopPreheader()->getTerminator(), DT); 621 if (OR != OverflowResult::MayOverflow) 622 return OR; 623 624 for (Value *V : FI.LinearIVUses) { 625 for (Value *U : V->users()) { 626 if (auto *GEP = dyn_cast<GetElementPtrInst>(U)) { 627 for (Value *GEPUser : U->users()) { 628 auto *GEPUserInst = cast<Instruction>(GEPUser); 629 if (!isa<LoadInst>(GEPUserInst) && 630 !(isa<StoreInst>(GEPUserInst) && 631 GEP == GEPUserInst->getOperand(1))) 632 continue; 633 if (!isGuaranteedToExecuteForEveryIteration(GEPUserInst, 634 FI.InnerLoop)) 635 continue; 636 // The IV is used as the operand of a GEP which dominates the loop 637 // latch, and the IV is at least as wide as the address space of the 638 // GEP. In this case, the GEP would wrap around the address space 639 // before the IV increment wraps, which would be UB. 640 if (GEP->isInBounds() && 641 V->getType()->getIntegerBitWidth() >= 642 DL.getPointerTypeSizeInBits(GEP->getType())) { 643 LLVM_DEBUG( 644 dbgs() << "use of linear IV would be UB if overflow occurred: "; 645 GEP->dump()); 646 return OverflowResult::NeverOverflows; 647 } 648 } 649 } 650 } 651 } 652 653 return OverflowResult::MayOverflow; 654 } 655 656 static bool CanFlattenLoopPair(FlattenInfo &FI, DominatorTree *DT, LoopInfo *LI, 657 ScalarEvolution *SE, AssumptionCache *AC, 658 const TargetTransformInfo *TTI) { 659 SmallPtrSet<Instruction *, 8> IterationInstructions; 660 if (!findLoopComponents(FI.InnerLoop, IterationInstructions, 661 FI.InnerInductionPHI, FI.InnerTripCount, 662 FI.InnerIncrement, FI.InnerBranch, SE, FI.Widened)) 663 return false; 664 if (!findLoopComponents(FI.OuterLoop, IterationInstructions, 665 FI.OuterInductionPHI, FI.OuterTripCount, 666 FI.OuterIncrement, FI.OuterBranch, SE, FI.Widened)) 667 return false; 668 669 // Both of the loop trip count values must be invariant in the outer loop 670 // (non-instructions are all inherently invariant). 671 if (!FI.OuterLoop->isLoopInvariant(FI.InnerTripCount)) { 672 LLVM_DEBUG(dbgs() << "inner loop trip count not invariant\n"); 673 return false; 674 } 675 if (!FI.OuterLoop->isLoopInvariant(FI.OuterTripCount)) { 676 LLVM_DEBUG(dbgs() << "outer loop trip count not invariant\n"); 677 return false; 678 } 679 680 if (!checkPHIs(FI, TTI)) 681 return false; 682 683 // FIXME: it should be possible to handle different types correctly. 684 if (FI.InnerInductionPHI->getType() != FI.OuterInductionPHI->getType()) 685 return false; 686 687 if (!checkOuterLoopInsts(FI, IterationInstructions, TTI)) 688 return false; 689 690 // Find the values in the loop that can be replaced with the linearized 691 // induction variable, and check that there are no other uses of the inner 692 // or outer induction variable. If there were, we could still do this 693 // transformation, but we'd have to insert a div/mod to calculate the 694 // original IVs, so it wouldn't be profitable. 695 if (!checkIVUsers(FI)) 696 return false; 697 698 LLVM_DEBUG(dbgs() << "CanFlattenLoopPair: OK\n"); 699 return true; 700 } 701 702 static bool DoFlattenLoopPair(FlattenInfo &FI, DominatorTree *DT, LoopInfo *LI, 703 ScalarEvolution *SE, AssumptionCache *AC, 704 const TargetTransformInfo *TTI, LPMUpdater *U, 705 MemorySSAUpdater *MSSAU) { 706 Function *F = FI.OuterLoop->getHeader()->getParent(); 707 LLVM_DEBUG(dbgs() << "Checks all passed, doing the transformation\n"); 708 { 709 using namespace ore; 710 OptimizationRemark Remark(DEBUG_TYPE, "Flattened", FI.InnerLoop->getStartLoc(), 711 FI.InnerLoop->getHeader()); 712 OptimizationRemarkEmitter ORE(F); 713 Remark << "Flattened into outer loop"; 714 ORE.emit(Remark); 715 } 716 717 Value *NewTripCount = BinaryOperator::CreateMul( 718 FI.InnerTripCount, FI.OuterTripCount, "flatten.tripcount", 719 FI.OuterLoop->getLoopPreheader()->getTerminator()); 720 LLVM_DEBUG(dbgs() << "Created new trip count in preheader: "; 721 NewTripCount->dump()); 722 723 // Fix up PHI nodes that take values from the inner loop back-edge, which 724 // we are about to remove. 725 FI.InnerInductionPHI->removeIncomingValue(FI.InnerLoop->getLoopLatch()); 726 727 // The old Phi will be optimised away later, but for now we can't leave 728 // leave it in an invalid state, so are updating them too. 729 for (PHINode *PHI : FI.InnerPHIsToTransform) 730 PHI->removeIncomingValue(FI.InnerLoop->getLoopLatch()); 731 732 // Modify the trip count of the outer loop to be the product of the two 733 // trip counts. 734 cast<User>(FI.OuterBranch->getCondition())->setOperand(1, NewTripCount); 735 736 // Replace the inner loop backedge with an unconditional branch to the exit. 737 BasicBlock *InnerExitBlock = FI.InnerLoop->getExitBlock(); 738 BasicBlock *InnerExitingBlock = FI.InnerLoop->getExitingBlock(); 739 InnerExitingBlock->getTerminator()->eraseFromParent(); 740 BranchInst::Create(InnerExitBlock, InnerExitingBlock); 741 742 // Update the DomTree and MemorySSA. 743 DT->deleteEdge(InnerExitingBlock, FI.InnerLoop->getHeader()); 744 if (MSSAU) 745 MSSAU->removeEdge(InnerExitingBlock, FI.InnerLoop->getHeader()); 746 747 // Replace all uses of the polynomial calculated from the two induction 748 // variables with the one new one. 749 IRBuilder<> Builder(FI.OuterInductionPHI->getParent()->getTerminator()); 750 for (Value *V : FI.LinearIVUses) { 751 Value *OuterValue = FI.OuterInductionPHI; 752 if (FI.Widened) 753 OuterValue = Builder.CreateTrunc(FI.OuterInductionPHI, V->getType(), 754 "flatten.trunciv"); 755 756 LLVM_DEBUG(dbgs() << "Replacing: "; V->dump(); dbgs() << "with: "; 757 OuterValue->dump()); 758 V->replaceAllUsesWith(OuterValue); 759 } 760 761 // Tell LoopInfo, SCEV and the pass manager that the inner loop has been 762 // deleted, and any information that have about the outer loop invalidated. 763 SE->forgetLoop(FI.OuterLoop); 764 SE->forgetLoop(FI.InnerLoop); 765 if (U) 766 U->markLoopAsDeleted(*FI.InnerLoop, FI.InnerLoop->getName()); 767 LI->erase(FI.InnerLoop); 768 769 // Increment statistic value. 770 NumFlattened++; 771 772 return true; 773 } 774 775 static bool CanWidenIV(FlattenInfo &FI, DominatorTree *DT, LoopInfo *LI, 776 ScalarEvolution *SE, AssumptionCache *AC, 777 const TargetTransformInfo *TTI) { 778 if (!WidenIV) { 779 LLVM_DEBUG(dbgs() << "Widening the IVs is disabled\n"); 780 return false; 781 } 782 783 LLVM_DEBUG(dbgs() << "Try widening the IVs\n"); 784 Module *M = FI.InnerLoop->getHeader()->getParent()->getParent(); 785 auto &DL = M->getDataLayout(); 786 auto *InnerType = FI.InnerInductionPHI->getType(); 787 auto *OuterType = FI.OuterInductionPHI->getType(); 788 unsigned MaxLegalSize = DL.getLargestLegalIntTypeSizeInBits(); 789 auto *MaxLegalType = DL.getLargestLegalIntType(M->getContext()); 790 791 // If both induction types are less than the maximum legal integer width, 792 // promote both to the widest type available so we know calculating 793 // (OuterTripCount * InnerTripCount) as the new trip count is safe. 794 if (InnerType != OuterType || 795 InnerType->getScalarSizeInBits() >= MaxLegalSize || 796 MaxLegalType->getScalarSizeInBits() < 797 InnerType->getScalarSizeInBits() * 2) { 798 LLVM_DEBUG(dbgs() << "Can't widen the IV\n"); 799 return false; 800 } 801 802 SCEVExpander Rewriter(*SE, DL, "loopflatten"); 803 SmallVector<WeakTrackingVH, 4> DeadInsts; 804 unsigned ElimExt = 0; 805 unsigned Widened = 0; 806 807 auto CreateWideIV = [&](WideIVInfo WideIV, bool &Deleted) -> bool { 808 PHINode *WidePhi = 809 createWideIV(WideIV, LI, SE, Rewriter, DT, DeadInsts, ElimExt, Widened, 810 true /* HasGuards */, true /* UsePostIncrementRanges */); 811 if (!WidePhi) 812 return false; 813 LLVM_DEBUG(dbgs() << "Created wide phi: "; WidePhi->dump()); 814 LLVM_DEBUG(dbgs() << "Deleting old phi: "; WideIV.NarrowIV->dump()); 815 Deleted = RecursivelyDeleteDeadPHINode(WideIV.NarrowIV); 816 return true; 817 }; 818 819 bool Deleted; 820 if (!CreateWideIV({FI.InnerInductionPHI, MaxLegalType, false}, Deleted)) 821 return false; 822 // Add the narrow phi to list, so that it will be adjusted later when the 823 // the transformation is performed. 824 if (!Deleted) 825 FI.InnerPHIsToTransform.insert(FI.InnerInductionPHI); 826 827 if (!CreateWideIV({FI.OuterInductionPHI, MaxLegalType, false}, Deleted)) 828 return false; 829 830 assert(Widened && "Widened IV expected"); 831 FI.Widened = true; 832 833 // Save the old/narrow induction phis, which we need to ignore in CheckPHIs. 834 FI.NarrowInnerInductionPHI = FI.InnerInductionPHI; 835 FI.NarrowOuterInductionPHI = FI.OuterInductionPHI; 836 837 // After widening, rediscover all the loop components. 838 return CanFlattenLoopPair(FI, DT, LI, SE, AC, TTI); 839 } 840 841 static bool FlattenLoopPair(FlattenInfo &FI, DominatorTree *DT, LoopInfo *LI, 842 ScalarEvolution *SE, AssumptionCache *AC, 843 const TargetTransformInfo *TTI, LPMUpdater *U, 844 MemorySSAUpdater *MSSAU) { 845 LLVM_DEBUG( 846 dbgs() << "Loop flattening running on outer loop " 847 << FI.OuterLoop->getHeader()->getName() << " and inner loop " 848 << FI.InnerLoop->getHeader()->getName() << " in " 849 << FI.OuterLoop->getHeader()->getParent()->getName() << "\n"); 850 851 if (!CanFlattenLoopPair(FI, DT, LI, SE, AC, TTI)) 852 return false; 853 854 // Check if we can widen the induction variables to avoid overflow checks. 855 bool CanFlatten = CanWidenIV(FI, DT, LI, SE, AC, TTI); 856 857 // It can happen that after widening of the IV, flattening may not be 858 // possible/happening, e.g. when it is deemed unprofitable. So bail here if 859 // that is the case. 860 // TODO: IV widening without performing the actual flattening transformation 861 // is not ideal. While this codegen change should not matter much, it is an 862 // unnecessary change which is better to avoid. It's unlikely this happens 863 // often, because if it's unprofitibale after widening, it should be 864 // unprofitabe before widening as checked in the first round of checks. But 865 // 'RepeatedInstructionThreshold' is set to only 2, which can probably be 866 // relaxed. Because this is making a code change (the IV widening, but not 867 // the flattening), we return true here. 868 if (FI.Widened && !CanFlatten) 869 return true; 870 871 // If we have widened and can perform the transformation, do that here. 872 if (CanFlatten) 873 return DoFlattenLoopPair(FI, DT, LI, SE, AC, TTI, U, MSSAU); 874 875 // Otherwise, if we haven't widened the IV, check if the new iteration 876 // variable might overflow. In this case, we need to version the loop, and 877 // select the original version at runtime if the iteration space is too 878 // large. 879 // TODO: We currently don't version the loop. 880 OverflowResult OR = checkOverflow(FI, DT, AC); 881 if (OR == OverflowResult::AlwaysOverflowsHigh || 882 OR == OverflowResult::AlwaysOverflowsLow) { 883 LLVM_DEBUG(dbgs() << "Multiply would always overflow, so not profitable\n"); 884 return false; 885 } else if (OR == OverflowResult::MayOverflow) { 886 LLVM_DEBUG(dbgs() << "Multiply might overflow, not flattening\n"); 887 return false; 888 } 889 890 LLVM_DEBUG(dbgs() << "Multiply cannot overflow, modifying loop in-place\n"); 891 return DoFlattenLoopPair(FI, DT, LI, SE, AC, TTI, U, MSSAU); 892 } 893 894 bool Flatten(LoopNest &LN, DominatorTree *DT, LoopInfo *LI, ScalarEvolution *SE, 895 AssumptionCache *AC, TargetTransformInfo *TTI, LPMUpdater *U, 896 MemorySSAUpdater *MSSAU) { 897 bool Changed = false; 898 for (Loop *InnerLoop : LN.getLoops()) { 899 auto *OuterLoop = InnerLoop->getParentLoop(); 900 if (!OuterLoop) 901 continue; 902 FlattenInfo FI(OuterLoop, InnerLoop); 903 Changed |= FlattenLoopPair(FI, DT, LI, SE, AC, TTI, U, MSSAU); 904 } 905 return Changed; 906 } 907 908 PreservedAnalyses LoopFlattenPass::run(LoopNest &LN, LoopAnalysisManager &LAM, 909 LoopStandardAnalysisResults &AR, 910 LPMUpdater &U) { 911 912 bool Changed = false; 913 914 Optional<MemorySSAUpdater> MSSAU; 915 if (AR.MSSA) { 916 MSSAU = MemorySSAUpdater(AR.MSSA); 917 if (VerifyMemorySSA) 918 AR.MSSA->verifyMemorySSA(); 919 } 920 921 // The loop flattening pass requires loops to be 922 // in simplified form, and also needs LCSSA. Running 923 // this pass will simplify all loops that contain inner loops, 924 // regardless of whether anything ends up being flattened. 925 Changed |= Flatten(LN, &AR.DT, &AR.LI, &AR.SE, &AR.AC, &AR.TTI, &U, 926 MSSAU ? MSSAU.getPointer() : nullptr); 927 928 if (!Changed) 929 return PreservedAnalyses::all(); 930 931 if (AR.MSSA && VerifyMemorySSA) 932 AR.MSSA->verifyMemorySSA(); 933 934 auto PA = getLoopPassPreservedAnalyses(); 935 if (AR.MSSA) 936 PA.preserve<MemorySSAAnalysis>(); 937 return PA; 938 } 939 940 namespace { 941 class LoopFlattenLegacyPass : public FunctionPass { 942 public: 943 static char ID; // Pass ID, replacement for typeid 944 LoopFlattenLegacyPass() : FunctionPass(ID) { 945 initializeLoopFlattenLegacyPassPass(*PassRegistry::getPassRegistry()); 946 } 947 948 // Possibly flatten loop L into its child. 949 bool runOnFunction(Function &F) override; 950 951 void getAnalysisUsage(AnalysisUsage &AU) const override { 952 getLoopAnalysisUsage(AU); 953 AU.addRequired<TargetTransformInfoWrapperPass>(); 954 AU.addPreserved<TargetTransformInfoWrapperPass>(); 955 AU.addRequired<AssumptionCacheTracker>(); 956 AU.addPreserved<AssumptionCacheTracker>(); 957 AU.addPreserved<MemorySSAWrapperPass>(); 958 } 959 }; 960 } // namespace 961 962 char LoopFlattenLegacyPass::ID = 0; 963 INITIALIZE_PASS_BEGIN(LoopFlattenLegacyPass, "loop-flatten", "Flattens loops", 964 false, false) 965 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 966 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 967 INITIALIZE_PASS_END(LoopFlattenLegacyPass, "loop-flatten", "Flattens loops", 968 false, false) 969 970 FunctionPass *llvm::createLoopFlattenPass() { 971 return new LoopFlattenLegacyPass(); 972 } 973 974 bool LoopFlattenLegacyPass::runOnFunction(Function &F) { 975 ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 976 LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 977 auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>(); 978 DominatorTree *DT = DTWP ? &DTWP->getDomTree() : nullptr; 979 auto &TTIP = getAnalysis<TargetTransformInfoWrapperPass>(); 980 auto *TTI = &TTIP.getTTI(F); 981 auto *AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F); 982 auto *MSSA = getAnalysisIfAvailable<MemorySSAWrapperPass>(); 983 984 Optional<MemorySSAUpdater> MSSAU; 985 if (MSSA) 986 MSSAU = MemorySSAUpdater(&MSSA->getMSSA()); 987 988 bool Changed = false; 989 for (Loop *L : *LI) { 990 auto LN = LoopNest::getLoopNest(*L, *SE); 991 Changed |= Flatten(*LN, DT, LI, SE, AC, TTI, nullptr, 992 MSSAU ? MSSAU.getPointer() : nullptr); 993 } 994 return Changed; 995 } 996