lib/Target/PowerPC/PPCLoopPreIncPrep.cpp

   1 //===------ PPCLoopPreIncPrep.cpp - Loop Pre-Inc. AM Prep. 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 file implements a pass to prepare loops for pre-increment addressing
  10 // modes. Additional PHIs are created for loop induction variables used by
  11 // load/store instructions so that the pre-increment forms can be used.
  12 // Generically, this means transforming loops like this:
  13 //   for (int i = 0; i < n; ++i)
  14 //     array[i] = c;
  15 // to look like this:
  16 //   T *p = array[-1];
  17 //   for (int i = 0; i < n; ++i)
  18 //     *++p = c;
  19 //===----------------------------------------------------------------------===//
  20
  21 #define DEBUG_TYPE "ppc-loop-preinc-prep"
  22
  23 #include "PPC.h"
  24 #include "PPCSubtarget.h"
  25 #include "PPCTargetMachine.h"
  26 #include "llvm/ADT/DepthFirstIterator.h"
  27 #include "llvm/ADT/SmallPtrSet.h"
  28 #include "llvm/ADT/SmallSet.h"
  29 #include "llvm/ADT/SmallVector.h"
  30 #include "llvm/ADT/Statistic.h"
  31 #include "llvm/Analysis/LoopInfo.h"
  32 #include "llvm/Analysis/ScalarEvolution.h"
  33 #include "llvm/Analysis/ScalarEvolutionExpander.h"
  34 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
  35 #include "llvm/Transforms/Utils/Local.h"
  36 #include "llvm/IR/BasicBlock.h"
  37 #include "llvm/IR/CFG.h"
  38 #include "llvm/IR/Dominators.h"
  39 #include "llvm/IR/Instruction.h"
  40 #include "llvm/IR/Instructions.h"
  41 #include "llvm/IR/IntrinsicInst.h"
  42 #include "llvm/IR/Module.h"
  43 #include "llvm/IR/Type.h"
  44 #include "llvm/IR/Value.h"
  45 #include "llvm/Pass.h"
  46 #include "llvm/Support/Casting.h"
  47 #include "llvm/Support/CommandLine.h"
  48 #include "llvm/Support/Debug.h"
  49 #include "llvm/Transforms/Scalar.h"
  50 #include "llvm/Transforms/Utils.h"
  51 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
  52 #include "llvm/Transforms/Utils/LoopUtils.h"
  53 #include <cassert>
  54 #include <iterator>
  55 #include <utility>
  56
  57 using namespace llvm;
  58
  59 // By default, we limit this to creating 16 PHIs (which is a little over half
  60 // of the allocatable register set).
  61 static cl::opt<unsigned> MaxVars("ppc-preinc-prep-max-vars",
  62                                  cl::Hidden, cl::init(16),
  63   cl::desc("Potential PHI threshold for PPC preinc loop prep"));
  64
  65 STATISTIC(PHINodeAlreadyExists, "PHI node already in pre-increment form");
  66
  67 namespace {
  68
  69   class PPCLoopPreIncPrep : public FunctionPass {
  70   public:
  71     static char ID; // Pass ID, replacement for typeid
  72
  73     PPCLoopPreIncPrep() : FunctionPass(ID) {
  74       initializePPCLoopPreIncPrepPass(*PassRegistry::getPassRegistry());
  75     }
  76
  77     PPCLoopPreIncPrep(PPCTargetMachine &TM) : FunctionPass(ID), TM(&TM) {
  78       initializePPCLoopPreIncPrepPass(*PassRegistry::getPassRegistry());
  79     }
  80
  81     void getAnalysisUsage(AnalysisUsage &AU) const override {
  82       AU.addPreserved<DominatorTreeWrapperPass>();
  83       AU.addRequired<LoopInfoWrapperPass>();
  84       AU.addPreserved<LoopInfoWrapperPass>();
  85       AU.addRequired<ScalarEvolutionWrapperPass>();
  86     }
  87
  88     bool alreadyPrepared(Loop *L, Instruction* MemI,
  89                          const SCEV *BasePtrStartSCEV,
  90                          const SCEVConstant *BasePtrIncSCEV);
  91     bool runOnFunction(Function &F) override;
  92
  93     bool runOnLoop(Loop *L);
  94     void simplifyLoopLatch(Loop *L);
  95     bool rotateLoop(Loop *L);
  96
  97   private:
  98     PPCTargetMachine *TM = nullptr;
  99     DominatorTree *DT;
 100     LoopInfo *LI;
 101     ScalarEvolution *SE;
 102     bool PreserveLCSSA;
 103   };
 104
 105 } // end anonymous namespace
 106
 107 char PPCLoopPreIncPrep::ID = 0;
 108 static const char *name = "Prepare loop for pre-inc. addressing modes";
 109 INITIALIZE_PASS_BEGIN(PPCLoopPreIncPrep, DEBUG_TYPE, name, false, false)
 110 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
 111 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
 112 INITIALIZE_PASS_END(PPCLoopPreIncPrep, DEBUG_TYPE, name, false, false)
 113
 114 FunctionPass *llvm::createPPCLoopPreIncPrepPass(PPCTargetMachine &TM) {
 115   return new PPCLoopPreIncPrep(TM);
 116 }
 117
 118 namespace {
 119
 120   struct BucketElement {
 121     BucketElement(const SCEVConstant *O, Instruction *I) : Offset(O), Instr(I) {}
 122     BucketElement(Instruction *I) : Offset(nullptr), Instr(I) {}
 123
 124     const SCEVConstant *Offset;
 125     Instruction *Instr;
 126   };
 127
 128   struct Bucket {
 129     Bucket(const SCEV *B, Instruction *I) : BaseSCEV(B),
 130                                             Elements(1, BucketElement(I)) {}
 131
 132     const SCEV *BaseSCEV;
 133     SmallVector<BucketElement, 16> Elements;
 134   };
 135
 136 } // end anonymous namespace
 137
 138 static bool IsPtrInBounds(Value *BasePtr) {
 139   Value *StrippedBasePtr = BasePtr;
 140   while (BitCastInst *BC = dyn_cast<BitCastInst>(StrippedBasePtr))
 141     StrippedBasePtr = BC->getOperand(0);
 142   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(StrippedBasePtr))
 143     return GEP->isInBounds();
 144
 145   return false;
 146 }
 147
 148 static Value *GetPointerOperand(Value *MemI) {
 149   if (LoadInst *LMemI = dyn_cast<LoadInst>(MemI)) {
 150     return LMemI->getPointerOperand();
 151   } else if (StoreInst *SMemI = dyn_cast<StoreInst>(MemI)) {
 152     return SMemI->getPointerOperand();
 153   } else if (IntrinsicInst *IMemI = dyn_cast<IntrinsicInst>(MemI)) {
 154     if (IMemI->getIntrinsicID() == Intrinsic::prefetch)
 155       return IMemI->getArgOperand(0);
 156   }
 157
 158   return nullptr;
 159 }
 160
 161 bool PPCLoopPreIncPrep::runOnFunction(Function &F) {
 162   if (skipFunction(F))
 163     return false;
 164
 165   LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
 166   SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
 167   auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
 168   DT = DTWP ? &DTWP->getDomTree() : nullptr;
 169   PreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
 170
 171   bool MadeChange = false;
 172
 173   for (auto I = LI->begin(), IE = LI->end(); I != IE; ++I)
 174     for (auto L = df_begin(*I), LE = df_end(*I); L != LE; ++L)
 175       MadeChange |= runOnLoop(*L);
 176
 177   return MadeChange;
 178 }
 179
 180 // In order to prepare for the pre-increment a PHI is added.
 181 // This function will check to see if that PHI already exists and will return
 182 //  true if it found an existing PHI with the same start and increment as the
 183 //  one we wanted to create.
 184 bool PPCLoopPreIncPrep::alreadyPrepared(Loop *L, Instruction* MemI,
 185                                         const SCEV *BasePtrStartSCEV,
 186                                         const SCEVConstant *BasePtrIncSCEV) {
 187   BasicBlock *BB = MemI->getParent();
 188   if (!BB)
 189     return false;
 190
 191   BasicBlock *PredBB = L->getLoopPredecessor();
 192   BasicBlock *LatchBB = L->getLoopLatch();
 193
 194   if (!PredBB || !LatchBB)
 195     return false;
 196
 197   // Run through the PHIs and see if we have some that looks like a preparation
 198   iterator_range<BasicBlock::phi_iterator> PHIIter = BB->phis();
 199   for (auto & CurrentPHI : PHIIter) {
 200     PHINode *CurrentPHINode = dyn_cast<PHINode>(&CurrentPHI);
 201     if (!CurrentPHINode)
 202       continue;
 203
 204     if (!SE->isSCEVable(CurrentPHINode->getType()))
 205       continue;
 206
 207     const SCEV *PHISCEV = SE->getSCEVAtScope(CurrentPHINode, L);
 208
 209     const SCEVAddRecExpr *PHIBasePtrSCEV = dyn_cast<SCEVAddRecExpr>(PHISCEV);
 210     if (!PHIBasePtrSCEV)
 211       continue;
 212
 213     const SCEVConstant *PHIBasePtrIncSCEV =
 214       dyn_cast<SCEVConstant>(PHIBasePtrSCEV->getStepRecurrence(*SE));
 215     if (!PHIBasePtrIncSCEV)
 216       continue;
 217
 218     if (CurrentPHINode->getNumIncomingValues() == 2) {
 219       if ( (CurrentPHINode->getIncomingBlock(0) == LatchBB &&
 220             CurrentPHINode->getIncomingBlock(1) == PredBB) ||
 221             (CurrentPHINode->getIncomingBlock(1) == LatchBB &&
 222             CurrentPHINode->getIncomingBlock(0) == PredBB) ) {
 223         if (PHIBasePtrSCEV->getStart() == BasePtrStartSCEV &&
 224             PHIBasePtrIncSCEV == BasePtrIncSCEV) {
 225           // The existing PHI (CurrentPHINode) has the same start and increment
 226           //  as the PHI that we wanted to create.
 227           ++PHINodeAlreadyExists;
 228           return true;
 229         }
 230       }
 231     }
 232   }
 233   return false;
 234 }
 235
 236 bool PPCLoopPreIncPrep::runOnLoop(Loop *L) {
 237   bool MadeChange = false;
 238
 239   // Only prep. the inner-most loop
 240   if (!L->empty())
 241     return MadeChange;
 242
 243   LLVM_DEBUG(dbgs() << "PIP: Examining: " << *L << "\n");
 244
 245   BasicBlock *Header = L->getHeader();
 246
 247   const PPCSubtarget *ST =
 248     TM ? TM->getSubtargetImpl(*Header->getParent()) : nullptr;
 249
 250   unsigned HeaderLoopPredCount = pred_size(Header);
 251
 252   // Collect buckets of comparable addresses used by loads and stores.
 253   SmallVector<Bucket, 16> Buckets;
 254   for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
 255        I != IE; ++I) {
 256     for (BasicBlock::iterator J = (*I)->begin(), JE = (*I)->end();
 257         J != JE; ++J) {
 258       Value *PtrValue;
 259       Instruction *MemI;
 260
 261       if (LoadInst *LMemI = dyn_cast<LoadInst>(J)) {
 262         MemI = LMemI;
 263         PtrValue = LMemI->getPointerOperand();
 264       } else if (StoreInst *SMemI = dyn_cast<StoreInst>(J)) {
 265         MemI = SMemI;
 266         PtrValue = SMemI->getPointerOperand();
 267       } else if (IntrinsicInst *IMemI = dyn_cast<IntrinsicInst>(J)) {
 268         if (IMemI->getIntrinsicID() == Intrinsic::prefetch) {
 269           MemI = IMemI;
 270           PtrValue = IMemI->getArgOperand(0);
 271         } else continue;
 272       } else continue;
 273
 274       unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace();
 275       if (PtrAddrSpace)
 276         continue;
 277
 278       // There are no update forms for Altivec vector load/stores.
 279       if (ST && ST->hasAltivec() &&
 280           PtrValue->getType()->getPointerElementType()->isVectorTy())
 281         continue;
 282
 283       if (L->isLoopInvariant(PtrValue))
 284         continue;
 285
 286       const SCEV *LSCEV = SE->getSCEVAtScope(PtrValue, L);
 287       if (const SCEVAddRecExpr *LARSCEV = dyn_cast<SCEVAddRecExpr>(LSCEV)) {
 288         if (LARSCEV->getLoop() != L)
 289           continue;
 290         // See getPreIndexedAddressParts, the displacement for LDU/STDU has to
 291         // be 4's multiple (DS-form). For i64 loads/stores when the displacement
 292         // fits in a 16-bit signed field but isn't a multiple of 4, it will be
 293         // useless and possible to break some original well-form addressing mode
 294         // to make this pre-inc prep for it.
 295         if (PtrValue->getType()->getPointerElementType()->isIntegerTy(64)) {
 296           if (const SCEVConstant *StepConst =
 297                   dyn_cast<SCEVConstant>(LARSCEV->getStepRecurrence(*SE))) {
 298             const APInt &ConstInt = StepConst->getValue()->getValue();
 299             if (ConstInt.isSignedIntN(16) && ConstInt.srem(4) != 0)
 300               continue;
 301           }
 302         }
 303       } else {
 304         continue;
 305       }
 306
 307       bool FoundBucket = false;
 308       for (auto &B : Buckets) {
 309         const SCEV *Diff = SE->getMinusSCEV(LSCEV, B.BaseSCEV);
 310         if (const auto *CDiff = dyn_cast<SCEVConstant>(Diff)) {
 311           B.Elements.push_back(BucketElement(CDiff, MemI));
 312           FoundBucket = true;
 313           break;
 314         }
 315       }
 316
 317       if (!FoundBucket) {
 318         if (Buckets.size() == MaxVars)
 319           return MadeChange;
 320         Buckets.push_back(Bucket(LSCEV, MemI));
 321       }
 322     }
 323   }
 324
 325   if (Buckets.empty())
 326     return MadeChange;
 327
 328   BasicBlock *LoopPredecessor = L->getLoopPredecessor();
 329   // If there is no loop predecessor, or the loop predecessor's terminator
 330   // returns a value (which might contribute to determining the loop's
 331   // iteration space), insert a new preheader for the loop.
 332   if (!LoopPredecessor ||
 333       !LoopPredecessor->getTerminator()->getType()->isVoidTy()) {
 334     LoopPredecessor = InsertPreheaderForLoop(L, DT, LI, nullptr, PreserveLCSSA);
 335     if (LoopPredecessor)
 336       MadeChange = true;
 337   }
 338   if (!LoopPredecessor)
 339     return MadeChange;
 340
 341   LLVM_DEBUG(dbgs() << "PIP: Found " << Buckets.size() << " buckets\n");
 342
 343   SmallSet<BasicBlock *, 16> BBChanged;
 344   for (unsigned i = 0, e = Buckets.size(); i != e; ++i) {
 345     // The base address of each bucket is transformed into a phi and the others
 346     // are rewritten as offsets of that variable.
 347
 348     // We have a choice now of which instruction's memory operand we use as the
 349     // base for the generated PHI. Always picking the first instruction in each
 350     // bucket does not work well, specifically because that instruction might
 351     // be a prefetch (and there are no pre-increment dcbt variants). Otherwise,
 352     // the choice is somewhat arbitrary, because the backend will happily
 353     // generate direct offsets from both the pre-incremented and
 354     // post-incremented pointer values. Thus, we'll pick the first non-prefetch
 355     // instruction in each bucket, and adjust the recurrence and other offsets
 356     // accordingly.
 357     for (int j = 0, je = Buckets[i].Elements.size(); j != je; ++j) {
 358       if (auto *II = dyn_cast<IntrinsicInst>(Buckets[i].Elements[j].Instr))
 359         if (II->getIntrinsicID() == Intrinsic::prefetch)
 360           continue;
 361
 362       // If we'd otherwise pick the first element anyway, there's nothing to do.
 363       if (j == 0)
 364         break;
 365
 366       // If our chosen element has no offset from the base pointer, there's
 367       // nothing to do.
 368       if (!Buckets[i].Elements[j].Offset ||
 369           Buckets[i].Elements[j].Offset->isZero())
 370         break;
 371
 372       const SCEV *Offset = Buckets[i].Elements[j].Offset;
 373       Buckets[i].BaseSCEV = SE->getAddExpr(Buckets[i].BaseSCEV, Offset);
 374       for (auto &E : Buckets[i].Elements) {
 375         if (E.Offset)
 376           E.Offset = cast<SCEVConstant>(SE->getMinusSCEV(E.Offset, Offset));
 377         else
 378           E.Offset = cast<SCEVConstant>(SE->getNegativeSCEV(Offset));
 379       }
 380
 381       std::swap(Buckets[i].Elements[j], Buckets[i].Elements[0]);
 382       break;
 383     }
 384
 385     const SCEVAddRecExpr *BasePtrSCEV =
 386       cast<SCEVAddRecExpr>(Buckets[i].BaseSCEV);
 387     if (!BasePtrSCEV->isAffine())
 388       continue;
 389
 390     LLVM_DEBUG(dbgs() << "PIP: Transforming: " << *BasePtrSCEV << "\n");
 391     assert(BasePtrSCEV->getLoop() == L &&
 392            "AddRec for the wrong loop?");
 393
 394     // The instruction corresponding to the Bucket's BaseSCEV must be the first
 395     // in the vector of elements.
 396     Instruction *MemI = Buckets[i].Elements.begin()->Instr;
 397     Value *BasePtr = GetPointerOperand(MemI);
 398     assert(BasePtr && "No pointer operand");
 399
 400     Type *I8Ty = Type::getInt8Ty(MemI->getParent()->getContext());
 401     Type *I8PtrTy = Type::getInt8PtrTy(MemI->getParent()->getContext(),
 402       BasePtr->getType()->getPointerAddressSpace());
 403
 404     const SCEV *BasePtrStartSCEV = BasePtrSCEV->getStart();
 405     if (!SE->isLoopInvariant(BasePtrStartSCEV, L))
 406       continue;
 407
 408     const SCEVConstant *BasePtrIncSCEV =
 409       dyn_cast<SCEVConstant>(BasePtrSCEV->getStepRecurrence(*SE));
 410     if (!BasePtrIncSCEV)
 411       continue;
 412     BasePtrStartSCEV = SE->getMinusSCEV(BasePtrStartSCEV, BasePtrIncSCEV);
 413     if (!isSafeToExpand(BasePtrStartSCEV, *SE))
 414       continue;
 415
 416     LLVM_DEBUG(dbgs() << "PIP: New start is: " << *BasePtrStartSCEV << "\n");
 417
 418     if (alreadyPrepared(L, MemI, BasePtrStartSCEV, BasePtrIncSCEV))
 419       continue;
 420
 421     PHINode *NewPHI = PHINode::Create(I8PtrTy, HeaderLoopPredCount,
 422       MemI->hasName() ? MemI->getName() + ".phi" : "",
 423       Header->getFirstNonPHI());
 424
 425     SCEVExpander SCEVE(*SE, Header->getModule()->getDataLayout(), "pistart");
 426     Value *BasePtrStart = SCEVE.expandCodeFor(BasePtrStartSCEV, I8PtrTy,
 427       LoopPredecessor->getTerminator());
 428
 429     // Note that LoopPredecessor might occur in the predecessor list multiple
 430     // times, and we need to add it the right number of times.
 431     for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
 432          PI != PE; ++PI) {
 433       if (*PI != LoopPredecessor)
 434         continue;
 435
 436       NewPHI->addIncoming(BasePtrStart, LoopPredecessor);
 437     }
 438
 439     Instruction *InsPoint = &*Header->getFirstInsertionPt();
 440     GetElementPtrInst *PtrInc = GetElementPtrInst::Create(
 441         I8Ty, NewPHI, BasePtrIncSCEV->getValue(),
 442         MemI->hasName() ? MemI->getName() + ".inc" : "", InsPoint);
 443     PtrInc->setIsInBounds(IsPtrInBounds(BasePtr));
 444     for (pred_iterator PI = pred_begin(Header), PE = pred_end(Header);
 445          PI != PE; ++PI) {
 446       if (*PI == LoopPredecessor)
 447         continue;
 448
 449       NewPHI->addIncoming(PtrInc, *PI);
 450     }
 451
 452     Instruction *NewBasePtr;
 453     if (PtrInc->getType() != BasePtr->getType())
 454       NewBasePtr = new BitCastInst(PtrInc, BasePtr->getType(),
 455         PtrInc->hasName() ? PtrInc->getName() + ".cast" : "", InsPoint);
 456     else
 457       NewBasePtr = PtrInc;
 458
 459     if (Instruction *IDel = dyn_cast<Instruction>(BasePtr))
 460       BBChanged.insert(IDel->getParent());
 461     BasePtr->replaceAllUsesWith(NewBasePtr);
 462     RecursivelyDeleteTriviallyDeadInstructions(BasePtr);
 463
 464     // Keep track of the replacement pointer values we've inserted so that we
 465     // don't generate more pointer values than necessary.
 466     SmallPtrSet<Value *, 16> NewPtrs;
 467     NewPtrs.insert( NewBasePtr);
 468
 469     for (auto I = std::next(Buckets[i].Elements.begin()),
 470          IE = Buckets[i].Elements.end(); I != IE; ++I) {
 471       Value *Ptr = GetPointerOperand(I->Instr);
 472       assert(Ptr && "No pointer operand");
 473       if (NewPtrs.count(Ptr))
 474         continue;
 475
 476       Instruction *RealNewPtr;
 477       if (!I->Offset || I->Offset->getValue()->isZero()) {
 478         RealNewPtr = NewBasePtr;
 479       } else {
 480         Instruction *PtrIP = dyn_cast<Instruction>(Ptr);
 481         if (PtrIP && isa<Instruction>(NewBasePtr) &&
 482             cast<Instruction>(NewBasePtr)->getParent() == PtrIP->getParent())
 483           PtrIP = nullptr;
 484         else if (isa<PHINode>(PtrIP))
 485           PtrIP = &*PtrIP->getParent()->getFirstInsertionPt();
 486         else if (!PtrIP)
 487           PtrIP = I->Instr;
 488
 489         GetElementPtrInst *NewPtr = GetElementPtrInst::Create(
 490             I8Ty, PtrInc, I->Offset->getValue(),
 491             I->Instr->hasName() ? I->Instr->getName() + ".off" : "", PtrIP);
 492         if (!PtrIP)
 493           NewPtr->insertAfter(cast<Instruction>(PtrInc));
 494         NewPtr->setIsInBounds(IsPtrInBounds(Ptr));
 495         RealNewPtr = NewPtr;
 496       }
 497
 498       if (Instruction *IDel = dyn_cast<Instruction>(Ptr))
 499         BBChanged.insert(IDel->getParent());
 500
 501       Instruction *ReplNewPtr;
 502       if (Ptr->getType() != RealNewPtr->getType()) {
 503         ReplNewPtr = new BitCastInst(RealNewPtr, Ptr->getType(),
 504           Ptr->hasName() ? Ptr->getName() + ".cast" : "");
 505         ReplNewPtr->insertAfter(RealNewPtr);
 506       } else
 507         ReplNewPtr = RealNewPtr;
 508
 509       Ptr->replaceAllUsesWith(ReplNewPtr);
 510       RecursivelyDeleteTriviallyDeadInstructions(Ptr);
 511
 512       NewPtrs.insert(RealNewPtr);
 513     }
 514
 515     MadeChange = true;
 516   }
 517
 518   for (Loop::block_iterator I = L->block_begin(), IE = L->block_end();
 519        I != IE; ++I) {
 520     if (BBChanged.count(*I))
 521       DeleteDeadPHIs(*I);
 522   }
 523
 524   return MadeChange;
 525 }