llvm/lib/Transforms/Scalar/LoopFlatten.cpp

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