1 //===- IndVarSimplify.cpp - Induction Variable Elimination ----------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This transformation analyzes and transforms the induction variables (and 11 // computations derived from them) into simpler forms suitable for subsequent 12 // analysis and transformation. 13 // 14 // If the trip count of a loop is computable, this pass also makes the following 15 // changes: 16 // 1. The exit condition for the loop is canonicalized to compare the 17 // induction value against the exit value. This turns loops like: 18 // 'for (i = 7; i*i < 1000; ++i)' into 'for (i = 0; i != 25; ++i)' 19 // 2. Any use outside of the loop of an expression derived from the indvar 20 // is changed to compute the derived value outside of the loop, eliminating 21 // the dependence on the exit value of the induction variable. If the only 22 // purpose of the loop is to compute the exit value of some derived 23 // expression, this transformation will make the loop dead. 24 // 25 //===----------------------------------------------------------------------===// 26 27 #include "llvm/Transforms/Scalar/IndVarSimplify.h" 28 #include "llvm/Transforms/Scalar.h" 29 #include "llvm/ADT/SmallVector.h" 30 #include "llvm/ADT/Statistic.h" 31 #include "llvm/Analysis/GlobalsModRef.h" 32 #include "llvm/Analysis/LoopInfo.h" 33 #include "llvm/Analysis/LoopPass.h" 34 #include "llvm/Analysis/LoopPassManager.h" 35 #include "llvm/Analysis/ScalarEvolutionExpander.h" 36 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h" 37 #include "llvm/Analysis/TargetLibraryInfo.h" 38 #include "llvm/Analysis/TargetTransformInfo.h" 39 #include "llvm/IR/BasicBlock.h" 40 #include "llvm/IR/CFG.h" 41 #include "llvm/IR/Constants.h" 42 #include "llvm/IR/DataLayout.h" 43 #include "llvm/IR/Dominators.h" 44 #include "llvm/IR/Instructions.h" 45 #include "llvm/IR/IntrinsicInst.h" 46 #include "llvm/IR/LLVMContext.h" 47 #include "llvm/IR/PatternMatch.h" 48 #include "llvm/IR/Type.h" 49 #include "llvm/Support/CommandLine.h" 50 #include "llvm/Support/Debug.h" 51 #include "llvm/Support/raw_ostream.h" 52 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 53 #include "llvm/Transforms/Utils/Local.h" 54 #include "llvm/Transforms/Utils/LoopUtils.h" 55 #include "llvm/Transforms/Utils/SimplifyIndVar.h" 56 using namespace llvm; 57 58 #define DEBUG_TYPE "indvars" 59 60 STATISTIC(NumWidened , "Number of indvars widened"); 61 STATISTIC(NumReplaced , "Number of exit values replaced"); 62 STATISTIC(NumLFTR , "Number of loop exit tests replaced"); 63 STATISTIC(NumElimExt , "Number of IV sign/zero extends eliminated"); 64 STATISTIC(NumElimIV , "Number of congruent IVs eliminated"); 65 66 // Trip count verification can be enabled by default under NDEBUG if we 67 // implement a strong expression equivalence checker in SCEV. Until then, we 68 // use the verify-indvars flag, which may assert in some cases. 69 static cl::opt<bool> VerifyIndvars( 70 "verify-indvars", cl::Hidden, 71 cl::desc("Verify the ScalarEvolution result after running indvars")); 72 73 enum ReplaceExitVal { NeverRepl, OnlyCheapRepl, AlwaysRepl }; 74 75 static cl::opt<ReplaceExitVal> ReplaceExitValue( 76 "replexitval", cl::Hidden, cl::init(OnlyCheapRepl), 77 cl::desc("Choose the strategy to replace exit value in IndVarSimplify"), 78 cl::values(clEnumValN(NeverRepl, "never", "never replace exit value"), 79 clEnumValN(OnlyCheapRepl, "cheap", 80 "only replace exit value when the cost is cheap"), 81 clEnumValN(AlwaysRepl, "always", 82 "always replace exit value whenever possible"), 83 clEnumValEnd)); 84 85 namespace { 86 struct RewritePhi; 87 88 class IndVarSimplify { 89 LoopInfo *LI; 90 ScalarEvolution *SE; 91 DominatorTree *DT; 92 const DataLayout &DL; 93 TargetLibraryInfo *TLI; 94 const TargetTransformInfo *TTI; 95 96 SmallVector<WeakVH, 16> DeadInsts; 97 bool Changed = false; 98 99 bool isValidRewrite(Value *FromVal, Value *ToVal); 100 101 void handleFloatingPointIV(Loop *L, PHINode *PH); 102 void rewriteNonIntegerIVs(Loop *L); 103 104 void simplifyAndExtend(Loop *L, SCEVExpander &Rewriter, LoopInfo *LI); 105 106 bool canLoopBeDeleted(Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet); 107 void rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter); 108 void rewriteFirstIterationLoopExitValues(Loop *L); 109 110 Value *linearFunctionTestReplace(Loop *L, const SCEV *BackedgeTakenCount, 111 PHINode *IndVar, SCEVExpander &Rewriter); 112 113 void sinkUnusedInvariants(Loop *L); 114 115 Value *expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, Loop *L, 116 Instruction *InsertPt, Type *Ty); 117 118 public: 119 IndVarSimplify(LoopInfo *LI, ScalarEvolution *SE, DominatorTree *DT, 120 const DataLayout &DL, TargetLibraryInfo *TLI, 121 TargetTransformInfo *TTI) 122 : LI(LI), SE(SE), DT(DT), DL(DL), TLI(TLI), TTI(TTI) {} 123 124 bool run(Loop *L); 125 }; 126 } 127 128 /// Return true if the SCEV expansion generated by the rewriter can replace the 129 /// original value. SCEV guarantees that it produces the same value, but the way 130 /// it is produced may be illegal IR. Ideally, this function will only be 131 /// called for verification. 132 bool IndVarSimplify::isValidRewrite(Value *FromVal, Value *ToVal) { 133 // If an SCEV expression subsumed multiple pointers, its expansion could 134 // reassociate the GEP changing the base pointer. This is illegal because the 135 // final address produced by a GEP chain must be inbounds relative to its 136 // underlying object. Otherwise basic alias analysis, among other things, 137 // could fail in a dangerous way. Ultimately, SCEV will be improved to avoid 138 // producing an expression involving multiple pointers. Until then, we must 139 // bail out here. 140 // 141 // Retrieve the pointer operand of the GEP. Don't use GetUnderlyingObject 142 // because it understands lcssa phis while SCEV does not. 143 Value *FromPtr = FromVal; 144 Value *ToPtr = ToVal; 145 if (auto *GEP = dyn_cast<GEPOperator>(FromVal)) { 146 FromPtr = GEP->getPointerOperand(); 147 } 148 if (auto *GEP = dyn_cast<GEPOperator>(ToVal)) { 149 ToPtr = GEP->getPointerOperand(); 150 } 151 if (FromPtr != FromVal || ToPtr != ToVal) { 152 // Quickly check the common case 153 if (FromPtr == ToPtr) 154 return true; 155 156 // SCEV may have rewritten an expression that produces the GEP's pointer 157 // operand. That's ok as long as the pointer operand has the same base 158 // pointer. Unlike GetUnderlyingObject(), getPointerBase() will find the 159 // base of a recurrence. This handles the case in which SCEV expansion 160 // converts a pointer type recurrence into a nonrecurrent pointer base 161 // indexed by an integer recurrence. 162 163 // If the GEP base pointer is a vector of pointers, abort. 164 if (!FromPtr->getType()->isPointerTy() || !ToPtr->getType()->isPointerTy()) 165 return false; 166 167 const SCEV *FromBase = SE->getPointerBase(SE->getSCEV(FromPtr)); 168 const SCEV *ToBase = SE->getPointerBase(SE->getSCEV(ToPtr)); 169 if (FromBase == ToBase) 170 return true; 171 172 DEBUG(dbgs() << "INDVARS: GEP rewrite bail out " 173 << *FromBase << " != " << *ToBase << "\n"); 174 175 return false; 176 } 177 return true; 178 } 179 180 /// Determine the insertion point for this user. By default, insert immediately 181 /// before the user. SCEVExpander or LICM will hoist loop invariants out of the 182 /// loop. For PHI nodes, there may be multiple uses, so compute the nearest 183 /// common dominator for the incoming blocks. 184 static Instruction *getInsertPointForUses(Instruction *User, Value *Def, 185 DominatorTree *DT, LoopInfo *LI) { 186 PHINode *PHI = dyn_cast<PHINode>(User); 187 if (!PHI) 188 return User; 189 190 Instruction *InsertPt = nullptr; 191 for (unsigned i = 0, e = PHI->getNumIncomingValues(); i != e; ++i) { 192 if (PHI->getIncomingValue(i) != Def) 193 continue; 194 195 BasicBlock *InsertBB = PHI->getIncomingBlock(i); 196 if (!InsertPt) { 197 InsertPt = InsertBB->getTerminator(); 198 continue; 199 } 200 InsertBB = DT->findNearestCommonDominator(InsertPt->getParent(), InsertBB); 201 InsertPt = InsertBB->getTerminator(); 202 } 203 assert(InsertPt && "Missing phi operand"); 204 205 auto *DefI = dyn_cast<Instruction>(Def); 206 if (!DefI) 207 return InsertPt; 208 209 assert(DT->dominates(DefI, InsertPt) && "def does not dominate all uses"); 210 211 auto *L = LI->getLoopFor(DefI->getParent()); 212 assert(!L || L->contains(LI->getLoopFor(InsertPt->getParent()))); 213 214 for (auto *DTN = (*DT)[InsertPt->getParent()]; DTN; DTN = DTN->getIDom()) 215 if (LI->getLoopFor(DTN->getBlock()) == L) 216 return DTN->getBlock()->getTerminator(); 217 218 llvm_unreachable("DefI dominates InsertPt!"); 219 } 220 221 //===----------------------------------------------------------------------===// 222 // rewriteNonIntegerIVs and helpers. Prefer integer IVs. 223 //===----------------------------------------------------------------------===// 224 225 /// Convert APF to an integer, if possible. 226 static bool ConvertToSInt(const APFloat &APF, int64_t &IntVal) { 227 bool isExact = false; 228 // See if we can convert this to an int64_t 229 uint64_t UIntVal; 230 if (APF.convertToInteger(&UIntVal, 64, true, APFloat::rmTowardZero, 231 &isExact) != APFloat::opOK || !isExact) 232 return false; 233 IntVal = UIntVal; 234 return true; 235 } 236 237 /// If the loop has floating induction variable then insert corresponding 238 /// integer induction variable if possible. 239 /// For example, 240 /// for(double i = 0; i < 10000; ++i) 241 /// bar(i) 242 /// is converted into 243 /// for(int i = 0; i < 10000; ++i) 244 /// bar((double)i); 245 /// 246 void IndVarSimplify::handleFloatingPointIV(Loop *L, PHINode *PN) { 247 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0)); 248 unsigned BackEdge = IncomingEdge^1; 249 250 // Check incoming value. 251 auto *InitValueVal = dyn_cast<ConstantFP>(PN->getIncomingValue(IncomingEdge)); 252 253 int64_t InitValue; 254 if (!InitValueVal || !ConvertToSInt(InitValueVal->getValueAPF(), InitValue)) 255 return; 256 257 // Check IV increment. Reject this PN if increment operation is not 258 // an add or increment value can not be represented by an integer. 259 auto *Incr = dyn_cast<BinaryOperator>(PN->getIncomingValue(BackEdge)); 260 if (Incr == nullptr || Incr->getOpcode() != Instruction::FAdd) return; 261 262 // If this is not an add of the PHI with a constantfp, or if the constant fp 263 // is not an integer, bail out. 264 ConstantFP *IncValueVal = dyn_cast<ConstantFP>(Incr->getOperand(1)); 265 int64_t IncValue; 266 if (IncValueVal == nullptr || Incr->getOperand(0) != PN || 267 !ConvertToSInt(IncValueVal->getValueAPF(), IncValue)) 268 return; 269 270 // Check Incr uses. One user is PN and the other user is an exit condition 271 // used by the conditional terminator. 272 Value::user_iterator IncrUse = Incr->user_begin(); 273 Instruction *U1 = cast<Instruction>(*IncrUse++); 274 if (IncrUse == Incr->user_end()) return; 275 Instruction *U2 = cast<Instruction>(*IncrUse++); 276 if (IncrUse != Incr->user_end()) return; 277 278 // Find exit condition, which is an fcmp. If it doesn't exist, or if it isn't 279 // only used by a branch, we can't transform it. 280 FCmpInst *Compare = dyn_cast<FCmpInst>(U1); 281 if (!Compare) 282 Compare = dyn_cast<FCmpInst>(U2); 283 if (!Compare || !Compare->hasOneUse() || 284 !isa<BranchInst>(Compare->user_back())) 285 return; 286 287 BranchInst *TheBr = cast<BranchInst>(Compare->user_back()); 288 289 // We need to verify that the branch actually controls the iteration count 290 // of the loop. If not, the new IV can overflow and no one will notice. 291 // The branch block must be in the loop and one of the successors must be out 292 // of the loop. 293 assert(TheBr->isConditional() && "Can't use fcmp if not conditional"); 294 if (!L->contains(TheBr->getParent()) || 295 (L->contains(TheBr->getSuccessor(0)) && 296 L->contains(TheBr->getSuccessor(1)))) 297 return; 298 299 300 // If it isn't a comparison with an integer-as-fp (the exit value), we can't 301 // transform it. 302 ConstantFP *ExitValueVal = dyn_cast<ConstantFP>(Compare->getOperand(1)); 303 int64_t ExitValue; 304 if (ExitValueVal == nullptr || 305 !ConvertToSInt(ExitValueVal->getValueAPF(), ExitValue)) 306 return; 307 308 // Find new predicate for integer comparison. 309 CmpInst::Predicate NewPred = CmpInst::BAD_ICMP_PREDICATE; 310 switch (Compare->getPredicate()) { 311 default: return; // Unknown comparison. 312 case CmpInst::FCMP_OEQ: 313 case CmpInst::FCMP_UEQ: NewPred = CmpInst::ICMP_EQ; break; 314 case CmpInst::FCMP_ONE: 315 case CmpInst::FCMP_UNE: NewPred = CmpInst::ICMP_NE; break; 316 case CmpInst::FCMP_OGT: 317 case CmpInst::FCMP_UGT: NewPred = CmpInst::ICMP_SGT; break; 318 case CmpInst::FCMP_OGE: 319 case CmpInst::FCMP_UGE: NewPred = CmpInst::ICMP_SGE; break; 320 case CmpInst::FCMP_OLT: 321 case CmpInst::FCMP_ULT: NewPred = CmpInst::ICMP_SLT; break; 322 case CmpInst::FCMP_OLE: 323 case CmpInst::FCMP_ULE: NewPred = CmpInst::ICMP_SLE; break; 324 } 325 326 // We convert the floating point induction variable to a signed i32 value if 327 // we can. This is only safe if the comparison will not overflow in a way 328 // that won't be trapped by the integer equivalent operations. Check for this 329 // now. 330 // TODO: We could use i64 if it is native and the range requires it. 331 332 // The start/stride/exit values must all fit in signed i32. 333 if (!isInt<32>(InitValue) || !isInt<32>(IncValue) || !isInt<32>(ExitValue)) 334 return; 335 336 // If not actually striding (add x, 0.0), avoid touching the code. 337 if (IncValue == 0) 338 return; 339 340 // Positive and negative strides have different safety conditions. 341 if (IncValue > 0) { 342 // If we have a positive stride, we require the init to be less than the 343 // exit value. 344 if (InitValue >= ExitValue) 345 return; 346 347 uint32_t Range = uint32_t(ExitValue-InitValue); 348 // Check for infinite loop, either: 349 // while (i <= Exit) or until (i > Exit) 350 if (NewPred == CmpInst::ICMP_SLE || NewPred == CmpInst::ICMP_SGT) { 351 if (++Range == 0) return; // Range overflows. 352 } 353 354 unsigned Leftover = Range % uint32_t(IncValue); 355 356 // If this is an equality comparison, we require that the strided value 357 // exactly land on the exit value, otherwise the IV condition will wrap 358 // around and do things the fp IV wouldn't. 359 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && 360 Leftover != 0) 361 return; 362 363 // If the stride would wrap around the i32 before exiting, we can't 364 // transform the IV. 365 if (Leftover != 0 && int32_t(ExitValue+IncValue) < ExitValue) 366 return; 367 368 } else { 369 // If we have a negative stride, we require the init to be greater than the 370 // exit value. 371 if (InitValue <= ExitValue) 372 return; 373 374 uint32_t Range = uint32_t(InitValue-ExitValue); 375 // Check for infinite loop, either: 376 // while (i >= Exit) or until (i < Exit) 377 if (NewPred == CmpInst::ICMP_SGE || NewPred == CmpInst::ICMP_SLT) { 378 if (++Range == 0) return; // Range overflows. 379 } 380 381 unsigned Leftover = Range % uint32_t(-IncValue); 382 383 // If this is an equality comparison, we require that the strided value 384 // exactly land on the exit value, otherwise the IV condition will wrap 385 // around and do things the fp IV wouldn't. 386 if ((NewPred == CmpInst::ICMP_EQ || NewPred == CmpInst::ICMP_NE) && 387 Leftover != 0) 388 return; 389 390 // If the stride would wrap around the i32 before exiting, we can't 391 // transform the IV. 392 if (Leftover != 0 && int32_t(ExitValue+IncValue) > ExitValue) 393 return; 394 } 395 396 IntegerType *Int32Ty = Type::getInt32Ty(PN->getContext()); 397 398 // Insert new integer induction variable. 399 PHINode *NewPHI = PHINode::Create(Int32Ty, 2, PN->getName()+".int", PN); 400 NewPHI->addIncoming(ConstantInt::get(Int32Ty, InitValue), 401 PN->getIncomingBlock(IncomingEdge)); 402 403 Value *NewAdd = 404 BinaryOperator::CreateAdd(NewPHI, ConstantInt::get(Int32Ty, IncValue), 405 Incr->getName()+".int", Incr); 406 NewPHI->addIncoming(NewAdd, PN->getIncomingBlock(BackEdge)); 407 408 ICmpInst *NewCompare = new ICmpInst(TheBr, NewPred, NewAdd, 409 ConstantInt::get(Int32Ty, ExitValue), 410 Compare->getName()); 411 412 // In the following deletions, PN may become dead and may be deleted. 413 // Use a WeakVH to observe whether this happens. 414 WeakVH WeakPH = PN; 415 416 // Delete the old floating point exit comparison. The branch starts using the 417 // new comparison. 418 NewCompare->takeName(Compare); 419 Compare->replaceAllUsesWith(NewCompare); 420 RecursivelyDeleteTriviallyDeadInstructions(Compare, TLI); 421 422 // Delete the old floating point increment. 423 Incr->replaceAllUsesWith(UndefValue::get(Incr->getType())); 424 RecursivelyDeleteTriviallyDeadInstructions(Incr, TLI); 425 426 // If the FP induction variable still has uses, this is because something else 427 // in the loop uses its value. In order to canonicalize the induction 428 // variable, we chose to eliminate the IV and rewrite it in terms of an 429 // int->fp cast. 430 // 431 // We give preference to sitofp over uitofp because it is faster on most 432 // platforms. 433 if (WeakPH) { 434 Value *Conv = new SIToFPInst(NewPHI, PN->getType(), "indvar.conv", 435 &*PN->getParent()->getFirstInsertionPt()); 436 PN->replaceAllUsesWith(Conv); 437 RecursivelyDeleteTriviallyDeadInstructions(PN, TLI); 438 } 439 Changed = true; 440 } 441 442 void IndVarSimplify::rewriteNonIntegerIVs(Loop *L) { 443 // First step. Check to see if there are any floating-point recurrences. 444 // If there are, change them into integer recurrences, permitting analysis by 445 // the SCEV routines. 446 // 447 BasicBlock *Header = L->getHeader(); 448 449 SmallVector<WeakVH, 8> PHIs; 450 for (BasicBlock::iterator I = Header->begin(); 451 PHINode *PN = dyn_cast<PHINode>(I); ++I) 452 PHIs.push_back(PN); 453 454 for (unsigned i = 0, e = PHIs.size(); i != e; ++i) 455 if (PHINode *PN = dyn_cast_or_null<PHINode>(&*PHIs[i])) 456 handleFloatingPointIV(L, PN); 457 458 // If the loop previously had floating-point IV, ScalarEvolution 459 // may not have been able to compute a trip count. Now that we've done some 460 // re-writing, the trip count may be computable. 461 if (Changed) 462 SE->forgetLoop(L); 463 } 464 465 namespace { 466 // Collect information about PHI nodes which can be transformed in 467 // rewriteLoopExitValues. 468 struct RewritePhi { 469 PHINode *PN; 470 unsigned Ith; // Ith incoming value. 471 Value *Val; // Exit value after expansion. 472 bool HighCost; // High Cost when expansion. 473 474 RewritePhi(PHINode *P, unsigned I, Value *V, bool H) 475 : PN(P), Ith(I), Val(V), HighCost(H) {} 476 }; 477 } 478 479 Value *IndVarSimplify::expandSCEVIfNeeded(SCEVExpander &Rewriter, const SCEV *S, 480 Loop *L, Instruction *InsertPt, 481 Type *ResultTy) { 482 // Before expanding S into an expensive LLVM expression, see if we can use an 483 // already existing value as the expansion for S. 484 if (Value *ExistingValue = Rewriter.findExistingExpansion(S, InsertPt, L)) 485 if (ExistingValue->getType() == ResultTy) 486 return ExistingValue; 487 488 // We didn't find anything, fall back to using SCEVExpander. 489 return Rewriter.expandCodeFor(S, ResultTy, InsertPt); 490 } 491 492 //===----------------------------------------------------------------------===// 493 // rewriteLoopExitValues - Optimize IV users outside the loop. 494 // As a side effect, reduces the amount of IV processing within the loop. 495 //===----------------------------------------------------------------------===// 496 497 /// Check to see if this loop has a computable loop-invariant execution count. 498 /// If so, this means that we can compute the final value of any expressions 499 /// that are recurrent in the loop, and substitute the exit values from the loop 500 /// into any instructions outside of the loop that use the final values of the 501 /// current expressions. 502 /// 503 /// This is mostly redundant with the regular IndVarSimplify activities that 504 /// happen later, except that it's more powerful in some cases, because it's 505 /// able to brute-force evaluate arbitrary instructions as long as they have 506 /// constant operands at the beginning of the loop. 507 void IndVarSimplify::rewriteLoopExitValues(Loop *L, SCEVExpander &Rewriter) { 508 // Check a pre-condition. 509 assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!"); 510 511 SmallVector<BasicBlock*, 8> ExitBlocks; 512 L->getUniqueExitBlocks(ExitBlocks); 513 514 SmallVector<RewritePhi, 8> RewritePhiSet; 515 // Find all values that are computed inside the loop, but used outside of it. 516 // Because of LCSSA, these values will only occur in LCSSA PHI Nodes. Scan 517 // the exit blocks of the loop to find them. 518 for (BasicBlock *ExitBB : ExitBlocks) { 519 // If there are no PHI nodes in this exit block, then no values defined 520 // inside the loop are used on this path, skip it. 521 PHINode *PN = dyn_cast<PHINode>(ExitBB->begin()); 522 if (!PN) continue; 523 524 unsigned NumPreds = PN->getNumIncomingValues(); 525 526 // Iterate over all of the PHI nodes. 527 BasicBlock::iterator BBI = ExitBB->begin(); 528 while ((PN = dyn_cast<PHINode>(BBI++))) { 529 if (PN->use_empty()) 530 continue; // dead use, don't replace it 531 532 if (!SE->isSCEVable(PN->getType())) 533 continue; 534 535 // It's necessary to tell ScalarEvolution about this explicitly so that 536 // it can walk the def-use list and forget all SCEVs, as it may not be 537 // watching the PHI itself. Once the new exit value is in place, there 538 // may not be a def-use connection between the loop and every instruction 539 // which got a SCEVAddRecExpr for that loop. 540 SE->forgetValue(PN); 541 542 // Iterate over all of the values in all the PHI nodes. 543 for (unsigned i = 0; i != NumPreds; ++i) { 544 // If the value being merged in is not integer or is not defined 545 // in the loop, skip it. 546 Value *InVal = PN->getIncomingValue(i); 547 if (!isa<Instruction>(InVal)) 548 continue; 549 550 // If this pred is for a subloop, not L itself, skip it. 551 if (LI->getLoopFor(PN->getIncomingBlock(i)) != L) 552 continue; // The Block is in a subloop, skip it. 553 554 // Check that InVal is defined in the loop. 555 Instruction *Inst = cast<Instruction>(InVal); 556 if (!L->contains(Inst)) 557 continue; 558 559 // Okay, this instruction has a user outside of the current loop 560 // and varies predictably *inside* the loop. Evaluate the value it 561 // contains when the loop exits, if possible. 562 const SCEV *ExitValue = SE->getSCEVAtScope(Inst, L->getParentLoop()); 563 if (!SE->isLoopInvariant(ExitValue, L) || 564 !isSafeToExpand(ExitValue, *SE)) 565 continue; 566 567 // Computing the value outside of the loop brings no benefit if : 568 // - it is definitely used inside the loop in a way which can not be 569 // optimized away. 570 // - no use outside of the loop can take advantage of hoisting the 571 // computation out of the loop 572 if (ExitValue->getSCEVType()>=scMulExpr) { 573 unsigned NumHardInternalUses = 0; 574 unsigned NumSoftExternalUses = 0; 575 unsigned NumUses = 0; 576 for (auto IB = Inst->user_begin(), IE = Inst->user_end(); 577 IB != IE && NumUses <= 6; ++IB) { 578 Instruction *UseInstr = cast<Instruction>(*IB); 579 unsigned Opc = UseInstr->getOpcode(); 580 NumUses++; 581 if (L->contains(UseInstr)) { 582 if (Opc == Instruction::Call || Opc == Instruction::Ret) 583 NumHardInternalUses++; 584 } else { 585 if (Opc == Instruction::PHI) { 586 // Do not count the Phi as a use. LCSSA may have inserted 587 // plenty of trivial ones. 588 NumUses--; 589 for (auto PB = UseInstr->user_begin(), 590 PE = UseInstr->user_end(); 591 PB != PE && NumUses <= 6; ++PB, ++NumUses) { 592 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode(); 593 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret) 594 NumSoftExternalUses++; 595 } 596 continue; 597 } 598 if (Opc != Instruction::Call && Opc != Instruction::Ret) 599 NumSoftExternalUses++; 600 } 601 } 602 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses) 603 continue; 604 } 605 606 bool HighCost = Rewriter.isHighCostExpansion(ExitValue, L, Inst); 607 Value *ExitVal = 608 expandSCEVIfNeeded(Rewriter, ExitValue, L, Inst, PN->getType()); 609 610 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n' 611 << " LoopVal = " << *Inst << "\n"); 612 613 if (!isValidRewrite(Inst, ExitVal)) { 614 DeadInsts.push_back(ExitVal); 615 continue; 616 } 617 618 // Collect all the candidate PHINodes to be rewritten. 619 RewritePhiSet.emplace_back(PN, i, ExitVal, HighCost); 620 } 621 } 622 } 623 624 bool LoopCanBeDel = canLoopBeDeleted(L, RewritePhiSet); 625 626 // Transformation. 627 for (const RewritePhi &Phi : RewritePhiSet) { 628 PHINode *PN = Phi.PN; 629 Value *ExitVal = Phi.Val; 630 631 // Only do the rewrite when the ExitValue can be expanded cheaply. 632 // If LoopCanBeDel is true, rewrite exit value aggressively. 633 if (ReplaceExitValue == OnlyCheapRepl && !LoopCanBeDel && Phi.HighCost) { 634 DeadInsts.push_back(ExitVal); 635 continue; 636 } 637 638 Changed = true; 639 ++NumReplaced; 640 Instruction *Inst = cast<Instruction>(PN->getIncomingValue(Phi.Ith)); 641 PN->setIncomingValue(Phi.Ith, ExitVal); 642 643 // If this instruction is dead now, delete it. Don't do it now to avoid 644 // invalidating iterators. 645 if (isInstructionTriviallyDead(Inst, TLI)) 646 DeadInsts.push_back(Inst); 647 648 // Replace PN with ExitVal if that is legal and does not break LCSSA. 649 if (PN->getNumIncomingValues() == 1 && 650 LI->replacementPreservesLCSSAForm(PN, ExitVal)) { 651 PN->replaceAllUsesWith(ExitVal); 652 PN->eraseFromParent(); 653 } 654 } 655 656 // The insertion point instruction may have been deleted; clear it out 657 // so that the rewriter doesn't trip over it later. 658 Rewriter.clearInsertPoint(); 659 } 660 661 //===---------------------------------------------------------------------===// 662 // rewriteFirstIterationLoopExitValues: Rewrite loop exit values if we know 663 // they will exit at the first iteration. 664 //===---------------------------------------------------------------------===// 665 666 /// Check to see if this loop has loop invariant conditions which lead to loop 667 /// exits. If so, we know that if the exit path is taken, it is at the first 668 /// loop iteration. This lets us predict exit values of PHI nodes that live in 669 /// loop header. 670 void IndVarSimplify::rewriteFirstIterationLoopExitValues(Loop *L) { 671 // Verify the input to the pass is already in LCSSA form. 672 assert(L->isLCSSAForm(*DT)); 673 674 SmallVector<BasicBlock *, 8> ExitBlocks; 675 L->getUniqueExitBlocks(ExitBlocks); 676 auto *LoopHeader = L->getHeader(); 677 assert(LoopHeader && "Invalid loop"); 678 679 for (auto *ExitBB : ExitBlocks) { 680 BasicBlock::iterator BBI = ExitBB->begin(); 681 // If there are no more PHI nodes in this exit block, then no more 682 // values defined inside the loop are used on this path. 683 while (auto *PN = dyn_cast<PHINode>(BBI++)) { 684 for (unsigned IncomingValIdx = 0, E = PN->getNumIncomingValues(); 685 IncomingValIdx != E; ++IncomingValIdx) { 686 auto *IncomingBB = PN->getIncomingBlock(IncomingValIdx); 687 688 // We currently only support loop exits from loop header. If the 689 // incoming block is not loop header, we need to recursively check 690 // all conditions starting from loop header are loop invariants. 691 // Additional support might be added in the future. 692 if (IncomingBB != LoopHeader) 693 continue; 694 695 // Get condition that leads to the exit path. 696 auto *TermInst = IncomingBB->getTerminator(); 697 698 Value *Cond = nullptr; 699 if (auto *BI = dyn_cast<BranchInst>(TermInst)) { 700 // Must be a conditional branch, otherwise the block 701 // should not be in the loop. 702 Cond = BI->getCondition(); 703 } else if (auto *SI = dyn_cast<SwitchInst>(TermInst)) 704 Cond = SI->getCondition(); 705 else 706 continue; 707 708 if (!L->isLoopInvariant(Cond)) 709 continue; 710 711 auto *ExitVal = 712 dyn_cast<PHINode>(PN->getIncomingValue(IncomingValIdx)); 713 714 // Only deal with PHIs. 715 if (!ExitVal) 716 continue; 717 718 // If ExitVal is a PHI on the loop header, then we know its 719 // value along this exit because the exit can only be taken 720 // on the first iteration. 721 auto *LoopPreheader = L->getLoopPreheader(); 722 assert(LoopPreheader && "Invalid loop"); 723 int PreheaderIdx = ExitVal->getBasicBlockIndex(LoopPreheader); 724 if (PreheaderIdx != -1) { 725 assert(ExitVal->getParent() == LoopHeader && 726 "ExitVal must be in loop header"); 727 PN->setIncomingValue(IncomingValIdx, 728 ExitVal->getIncomingValue(PreheaderIdx)); 729 } 730 } 731 } 732 } 733 } 734 735 /// Check whether it is possible to delete the loop after rewriting exit 736 /// value. If it is possible, ignore ReplaceExitValue and do rewriting 737 /// aggressively. 738 bool IndVarSimplify::canLoopBeDeleted( 739 Loop *L, SmallVector<RewritePhi, 8> &RewritePhiSet) { 740 741 BasicBlock *Preheader = L->getLoopPreheader(); 742 // If there is no preheader, the loop will not be deleted. 743 if (!Preheader) 744 return false; 745 746 // In LoopDeletion pass Loop can be deleted when ExitingBlocks.size() > 1. 747 // We obviate multiple ExitingBlocks case for simplicity. 748 // TODO: If we see testcase with multiple ExitingBlocks can be deleted 749 // after exit value rewriting, we can enhance the logic here. 750 SmallVector<BasicBlock *, 4> ExitingBlocks; 751 L->getExitingBlocks(ExitingBlocks); 752 SmallVector<BasicBlock *, 8> ExitBlocks; 753 L->getUniqueExitBlocks(ExitBlocks); 754 if (ExitBlocks.size() > 1 || ExitingBlocks.size() > 1) 755 return false; 756 757 BasicBlock *ExitBlock = ExitBlocks[0]; 758 BasicBlock::iterator BI = ExitBlock->begin(); 759 while (PHINode *P = dyn_cast<PHINode>(BI)) { 760 Value *Incoming = P->getIncomingValueForBlock(ExitingBlocks[0]); 761 762 // If the Incoming value of P is found in RewritePhiSet, we know it 763 // could be rewritten to use a loop invariant value in transformation 764 // phase later. Skip it in the loop invariant check below. 765 bool found = false; 766 for (const RewritePhi &Phi : RewritePhiSet) { 767 unsigned i = Phi.Ith; 768 if (Phi.PN == P && (Phi.PN)->getIncomingValue(i) == Incoming) { 769 found = true; 770 break; 771 } 772 } 773 774 Instruction *I; 775 if (!found && (I = dyn_cast<Instruction>(Incoming))) 776 if (!L->hasLoopInvariantOperands(I)) 777 return false; 778 779 ++BI; 780 } 781 782 for (auto *BB : L->blocks()) 783 if (any_of(*BB, [](Instruction &I) { return I.mayHaveSideEffects(); })) 784 return false; 785 786 return true; 787 } 788 789 //===----------------------------------------------------------------------===// 790 // IV Widening - Extend the width of an IV to cover its widest uses. 791 //===----------------------------------------------------------------------===// 792 793 namespace { 794 // Collect information about induction variables that are used by sign/zero 795 // extend operations. This information is recorded by CollectExtend and provides 796 // the input to WidenIV. 797 struct WideIVInfo { 798 PHINode *NarrowIV = nullptr; 799 Type *WidestNativeType = nullptr; // Widest integer type created [sz]ext 800 bool IsSigned = false; // Was a sext user seen before a zext? 801 }; 802 } 803 804 /// Update information about the induction variable that is extended by this 805 /// sign or zero extend operation. This is used to determine the final width of 806 /// the IV before actually widening it. 807 static void visitIVCast(CastInst *Cast, WideIVInfo &WI, ScalarEvolution *SE, 808 const TargetTransformInfo *TTI) { 809 bool IsSigned = Cast->getOpcode() == Instruction::SExt; 810 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt) 811 return; 812 813 Type *Ty = Cast->getType(); 814 uint64_t Width = SE->getTypeSizeInBits(Ty); 815 if (!Cast->getModule()->getDataLayout().isLegalInteger(Width)) 816 return; 817 818 // Cast is either an sext or zext up to this point. 819 // We should not widen an indvar if arithmetics on the wider indvar are more 820 // expensive than those on the narrower indvar. We check only the cost of ADD 821 // because at least an ADD is required to increment the induction variable. We 822 // could compute more comprehensively the cost of all instructions on the 823 // induction variable when necessary. 824 if (TTI && 825 TTI->getArithmeticInstrCost(Instruction::Add, Ty) > 826 TTI->getArithmeticInstrCost(Instruction::Add, 827 Cast->getOperand(0)->getType())) { 828 return; 829 } 830 831 if (!WI.WidestNativeType) { 832 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 833 WI.IsSigned = IsSigned; 834 return; 835 } 836 837 // We extend the IV to satisfy the sign of its first user, arbitrarily. 838 if (WI.IsSigned != IsSigned) 839 return; 840 841 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType)) 842 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 843 } 844 845 namespace { 846 847 /// Record a link in the Narrow IV def-use chain along with the WideIV that 848 /// computes the same value as the Narrow IV def. This avoids caching Use* 849 /// pointers. 850 struct NarrowIVDefUse { 851 Instruction *NarrowDef = nullptr; 852 Instruction *NarrowUse = nullptr; 853 Instruction *WideDef = nullptr; 854 855 // True if the narrow def is never negative. Tracking this information lets 856 // us use a sign extension instead of a zero extension or vice versa, when 857 // profitable and legal. 858 bool NeverNegative = false; 859 860 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD, 861 bool NeverNegative) 862 : NarrowDef(ND), NarrowUse(NU), WideDef(WD), 863 NeverNegative(NeverNegative) {} 864 }; 865 866 /// The goal of this transform is to remove sign and zero extends without 867 /// creating any new induction variables. To do this, it creates a new phi of 868 /// the wider type and redirects all users, either removing extends or inserting 869 /// truncs whenever we stop propagating the type. 870 /// 871 class WidenIV { 872 // Parameters 873 PHINode *OrigPhi; 874 Type *WideType; 875 bool IsSigned; 876 877 // Context 878 LoopInfo *LI; 879 Loop *L; 880 ScalarEvolution *SE; 881 DominatorTree *DT; 882 883 // Result 884 PHINode *WidePhi; 885 Instruction *WideInc; 886 const SCEV *WideIncExpr; 887 SmallVectorImpl<WeakVH> &DeadInsts; 888 889 SmallPtrSet<Instruction*,16> Widened; 890 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers; 891 892 public: 893 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, 894 ScalarEvolution *SEv, DominatorTree *DTree, 895 SmallVectorImpl<WeakVH> &DI) : 896 OrigPhi(WI.NarrowIV), 897 WideType(WI.WidestNativeType), 898 IsSigned(WI.IsSigned), 899 LI(LInfo), 900 L(LI->getLoopFor(OrigPhi->getParent())), 901 SE(SEv), 902 DT(DTree), 903 WidePhi(nullptr), 904 WideInc(nullptr), 905 WideIncExpr(nullptr), 906 DeadInsts(DI) { 907 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV"); 908 } 909 910 PHINode *createWideIV(SCEVExpander &Rewriter); 911 912 protected: 913 Value *createExtendInst(Value *NarrowOper, Type *WideType, bool IsSigned, 914 Instruction *Use); 915 916 Instruction *cloneIVUser(NarrowIVDefUse DU, const SCEVAddRecExpr *WideAR); 917 Instruction *cloneArithmeticIVUser(NarrowIVDefUse DU, 918 const SCEVAddRecExpr *WideAR); 919 Instruction *cloneBitwiseIVUser(NarrowIVDefUse DU); 920 921 const SCEVAddRecExpr *getWideRecurrence(Instruction *NarrowUse); 922 923 const SCEVAddRecExpr* getExtendedOperandRecurrence(NarrowIVDefUse DU); 924 925 const SCEV *getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS, 926 unsigned OpCode) const; 927 928 Instruction *widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter); 929 930 bool widenLoopCompare(NarrowIVDefUse DU); 931 932 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef); 933 }; 934 } // anonymous namespace 935 936 /// Perform a quick domtree based check for loop invariance assuming that V is 937 /// used within the loop. LoopInfo::isLoopInvariant() seems gratuitous for this 938 /// purpose. 939 static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) { 940 Instruction *Inst = dyn_cast<Instruction>(V); 941 if (!Inst) 942 return true; 943 944 return DT->properlyDominates(Inst->getParent(), L->getHeader()); 945 } 946 947 Value *WidenIV::createExtendInst(Value *NarrowOper, Type *WideType, 948 bool IsSigned, Instruction *Use) { 949 // Set the debug location and conservative insertion point. 950 IRBuilder<> Builder(Use); 951 // Hoist the insertion point into loop preheaders as far as possible. 952 for (const Loop *L = LI->getLoopFor(Use->getParent()); 953 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT); 954 L = L->getParentLoop()) 955 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator()); 956 957 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) : 958 Builder.CreateZExt(NarrowOper, WideType); 959 } 960 961 /// Instantiate a wide operation to replace a narrow operation. This only needs 962 /// to handle operations that can evaluation to SCEVAddRec. It can safely return 963 /// 0 for any operation we decide not to clone. 964 Instruction *WidenIV::cloneIVUser(NarrowIVDefUse DU, 965 const SCEVAddRecExpr *WideAR) { 966 unsigned Opcode = DU.NarrowUse->getOpcode(); 967 switch (Opcode) { 968 default: 969 return nullptr; 970 case Instruction::Add: 971 case Instruction::Mul: 972 case Instruction::UDiv: 973 case Instruction::Sub: 974 return cloneArithmeticIVUser(DU, WideAR); 975 976 case Instruction::And: 977 case Instruction::Or: 978 case Instruction::Xor: 979 case Instruction::Shl: 980 case Instruction::LShr: 981 case Instruction::AShr: 982 return cloneBitwiseIVUser(DU); 983 } 984 } 985 986 Instruction *WidenIV::cloneBitwiseIVUser(NarrowIVDefUse DU) { 987 Instruction *NarrowUse = DU.NarrowUse; 988 Instruction *NarrowDef = DU.NarrowDef; 989 Instruction *WideDef = DU.WideDef; 990 991 DEBUG(dbgs() << "Cloning bitwise IVUser: " << *NarrowUse << "\n"); 992 993 // Replace NarrowDef operands with WideDef. Otherwise, we don't know anything 994 // about the narrow operand yet so must insert a [sz]ext. It is probably loop 995 // invariant and will be folded or hoisted. If it actually comes from a 996 // widened IV, it should be removed during a future call to widenIVUse. 997 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) 998 ? WideDef 999 : createExtendInst(NarrowUse->getOperand(0), WideType, 1000 IsSigned, NarrowUse); 1001 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) 1002 ? WideDef 1003 : createExtendInst(NarrowUse->getOperand(1), WideType, 1004 IsSigned, NarrowUse); 1005 1006 auto *NarrowBO = cast<BinaryOperator>(NarrowUse); 1007 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, 1008 NarrowBO->getName()); 1009 IRBuilder<> Builder(NarrowUse); 1010 Builder.Insert(WideBO); 1011 WideBO->copyIRFlags(NarrowBO); 1012 return WideBO; 1013 } 1014 1015 Instruction *WidenIV::cloneArithmeticIVUser(NarrowIVDefUse DU, 1016 const SCEVAddRecExpr *WideAR) { 1017 Instruction *NarrowUse = DU.NarrowUse; 1018 Instruction *NarrowDef = DU.NarrowDef; 1019 Instruction *WideDef = DU.WideDef; 1020 1021 DEBUG(dbgs() << "Cloning arithmetic IVUser: " << *NarrowUse << "\n"); 1022 1023 unsigned IVOpIdx = (NarrowUse->getOperand(0) == NarrowDef) ? 0 : 1; 1024 1025 // We're trying to find X such that 1026 // 1027 // Widen(NarrowDef `op` NonIVNarrowDef) == WideAR == WideDef `op.wide` X 1028 // 1029 // We guess two solutions to X, sext(NonIVNarrowDef) and zext(NonIVNarrowDef), 1030 // and check using SCEV if any of them are correct. 1031 1032 // Returns true if extending NonIVNarrowDef according to `SignExt` is a 1033 // correct solution to X. 1034 auto GuessNonIVOperand = [&](bool SignExt) { 1035 const SCEV *WideLHS; 1036 const SCEV *WideRHS; 1037 1038 auto GetExtend = [this, SignExt](const SCEV *S, Type *Ty) { 1039 if (SignExt) 1040 return SE->getSignExtendExpr(S, Ty); 1041 return SE->getZeroExtendExpr(S, Ty); 1042 }; 1043 1044 if (IVOpIdx == 0) { 1045 WideLHS = SE->getSCEV(WideDef); 1046 const SCEV *NarrowRHS = SE->getSCEV(NarrowUse->getOperand(1)); 1047 WideRHS = GetExtend(NarrowRHS, WideType); 1048 } else { 1049 const SCEV *NarrowLHS = SE->getSCEV(NarrowUse->getOperand(0)); 1050 WideLHS = GetExtend(NarrowLHS, WideType); 1051 WideRHS = SE->getSCEV(WideDef); 1052 } 1053 1054 // WideUse is "WideDef `op.wide` X" as described in the comment. 1055 const SCEV *WideUse = nullptr; 1056 1057 switch (NarrowUse->getOpcode()) { 1058 default: 1059 llvm_unreachable("No other possibility!"); 1060 1061 case Instruction::Add: 1062 WideUse = SE->getAddExpr(WideLHS, WideRHS); 1063 break; 1064 1065 case Instruction::Mul: 1066 WideUse = SE->getMulExpr(WideLHS, WideRHS); 1067 break; 1068 1069 case Instruction::UDiv: 1070 WideUse = SE->getUDivExpr(WideLHS, WideRHS); 1071 break; 1072 1073 case Instruction::Sub: 1074 WideUse = SE->getMinusSCEV(WideLHS, WideRHS); 1075 break; 1076 } 1077 1078 return WideUse == WideAR; 1079 }; 1080 1081 bool SignExtend = IsSigned; 1082 if (!GuessNonIVOperand(SignExtend)) { 1083 SignExtend = !SignExtend; 1084 if (!GuessNonIVOperand(SignExtend)) 1085 return nullptr; 1086 } 1087 1088 Value *LHS = (NarrowUse->getOperand(0) == NarrowDef) 1089 ? WideDef 1090 : createExtendInst(NarrowUse->getOperand(0), WideType, 1091 SignExtend, NarrowUse); 1092 Value *RHS = (NarrowUse->getOperand(1) == NarrowDef) 1093 ? WideDef 1094 : createExtendInst(NarrowUse->getOperand(1), WideType, 1095 SignExtend, NarrowUse); 1096 1097 auto *NarrowBO = cast<BinaryOperator>(NarrowUse); 1098 auto *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), LHS, RHS, 1099 NarrowBO->getName()); 1100 1101 IRBuilder<> Builder(NarrowUse); 1102 Builder.Insert(WideBO); 1103 WideBO->copyIRFlags(NarrowBO); 1104 return WideBO; 1105 } 1106 1107 const SCEV *WidenIV::getSCEVByOpCode(const SCEV *LHS, const SCEV *RHS, 1108 unsigned OpCode) const { 1109 if (OpCode == Instruction::Add) 1110 return SE->getAddExpr(LHS, RHS); 1111 if (OpCode == Instruction::Sub) 1112 return SE->getMinusSCEV(LHS, RHS); 1113 if (OpCode == Instruction::Mul) 1114 return SE->getMulExpr(LHS, RHS); 1115 1116 llvm_unreachable("Unsupported opcode."); 1117 } 1118 1119 /// No-wrap operations can transfer sign extension of their result to their 1120 /// operands. Generate the SCEV value for the widened operation without 1121 /// actually modifying the IR yet. If the expression after extending the 1122 /// operands is an AddRec for this loop, return it. 1123 const SCEVAddRecExpr* WidenIV::getExtendedOperandRecurrence(NarrowIVDefUse DU) { 1124 1125 // Handle the common case of add<nsw/nuw> 1126 const unsigned OpCode = DU.NarrowUse->getOpcode(); 1127 // Only Add/Sub/Mul instructions supported yet. 1128 if (OpCode != Instruction::Add && OpCode != Instruction::Sub && 1129 OpCode != Instruction::Mul) 1130 return nullptr; 1131 1132 // One operand (NarrowDef) has already been extended to WideDef. Now determine 1133 // if extending the other will lead to a recurrence. 1134 const unsigned ExtendOperIdx = 1135 DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0; 1136 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU"); 1137 1138 const SCEV *ExtendOperExpr = nullptr; 1139 const OverflowingBinaryOperator *OBO = 1140 cast<OverflowingBinaryOperator>(DU.NarrowUse); 1141 if (IsSigned && OBO->hasNoSignedWrap()) 1142 ExtendOperExpr = SE->getSignExtendExpr( 1143 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 1144 else if(!IsSigned && OBO->hasNoUnsignedWrap()) 1145 ExtendOperExpr = SE->getZeroExtendExpr( 1146 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 1147 else 1148 return nullptr; 1149 1150 // When creating this SCEV expr, don't apply the current operations NSW or NUW 1151 // flags. This instruction may be guarded by control flow that the no-wrap 1152 // behavior depends on. Non-control-equivalent instructions can be mapped to 1153 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW 1154 // semantics to those operations. 1155 const SCEV *lhs = SE->getSCEV(DU.WideDef); 1156 const SCEV *rhs = ExtendOperExpr; 1157 1158 // Let's swap operands to the initial order for the case of non-commutative 1159 // operations, like SUB. See PR21014. 1160 if (ExtendOperIdx == 0) 1161 std::swap(lhs, rhs); 1162 const SCEVAddRecExpr *AddRec = 1163 dyn_cast<SCEVAddRecExpr>(getSCEVByOpCode(lhs, rhs, OpCode)); 1164 1165 if (!AddRec || AddRec->getLoop() != L) 1166 return nullptr; 1167 return AddRec; 1168 } 1169 1170 /// Is this instruction potentially interesting for further simplification after 1171 /// widening it's type? In other words, can the extend be safely hoisted out of 1172 /// the loop with SCEV reducing the value to a recurrence on the same loop. If 1173 /// so, return the sign or zero extended recurrence. Otherwise return NULL. 1174 const SCEVAddRecExpr *WidenIV::getWideRecurrence(Instruction *NarrowUse) { 1175 if (!SE->isSCEVable(NarrowUse->getType())) 1176 return nullptr; 1177 1178 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse); 1179 if (SE->getTypeSizeInBits(NarrowExpr->getType()) 1180 >= SE->getTypeSizeInBits(WideType)) { 1181 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow 1182 // index. So don't follow this use. 1183 return nullptr; 1184 } 1185 1186 const SCEV *WideExpr = IsSigned ? 1187 SE->getSignExtendExpr(NarrowExpr, WideType) : 1188 SE->getZeroExtendExpr(NarrowExpr, WideType); 1189 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr); 1190 if (!AddRec || AddRec->getLoop() != L) 1191 return nullptr; 1192 return AddRec; 1193 } 1194 1195 /// This IV user cannot be widen. Replace this use of the original narrow IV 1196 /// with a truncation of the new wide IV to isolate and eliminate the narrow IV. 1197 static void truncateIVUse(NarrowIVDefUse DU, DominatorTree *DT, LoopInfo *LI) { 1198 DEBUG(dbgs() << "INDVARS: Truncate IV " << *DU.WideDef 1199 << " for user " << *DU.NarrowUse << "\n"); 1200 IRBuilder<> Builder( 1201 getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI)); 1202 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType()); 1203 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc); 1204 } 1205 1206 /// If the narrow use is a compare instruction, then widen the compare 1207 // (and possibly the other operand). The extend operation is hoisted into the 1208 // loop preheader as far as possible. 1209 bool WidenIV::widenLoopCompare(NarrowIVDefUse DU) { 1210 ICmpInst *Cmp = dyn_cast<ICmpInst>(DU.NarrowUse); 1211 if (!Cmp) 1212 return false; 1213 1214 // We can legally widen the comparison in the following two cases: 1215 // 1216 // - The signedness of the IV extension and comparison match 1217 // 1218 // - The narrow IV is always positive (and thus its sign extension is equal 1219 // to its zero extension). For instance, let's say we're zero extending 1220 // %narrow for the following use 1221 // 1222 // icmp slt i32 %narrow, %val ... (A) 1223 // 1224 // and %narrow is always positive. Then 1225 // 1226 // (A) == icmp slt i32 sext(%narrow), sext(%val) 1227 // == icmp slt i32 zext(%narrow), sext(%val) 1228 1229 if (!(DU.NeverNegative || IsSigned == Cmp->isSigned())) 1230 return false; 1231 1232 Value *Op = Cmp->getOperand(Cmp->getOperand(0) == DU.NarrowDef ? 1 : 0); 1233 unsigned CastWidth = SE->getTypeSizeInBits(Op->getType()); 1234 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 1235 assert (CastWidth <= IVWidth && "Unexpected width while widening compare."); 1236 1237 // Widen the compare instruction. 1238 IRBuilder<> Builder( 1239 getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT, LI)); 1240 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 1241 1242 // Widen the other operand of the compare, if necessary. 1243 if (CastWidth < IVWidth) { 1244 Value *ExtOp = createExtendInst(Op, WideType, Cmp->isSigned(), Cmp); 1245 DU.NarrowUse->replaceUsesOfWith(Op, ExtOp); 1246 } 1247 return true; 1248 } 1249 1250 /// Determine whether an individual user of the narrow IV can be widened. If so, 1251 /// return the wide clone of the user. 1252 Instruction *WidenIV::widenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) { 1253 1254 // Stop traversing the def-use chain at inner-loop phis or post-loop phis. 1255 if (PHINode *UsePhi = dyn_cast<PHINode>(DU.NarrowUse)) { 1256 if (LI->getLoopFor(UsePhi->getParent()) != L) { 1257 // For LCSSA phis, sink the truncate outside the loop. 1258 // After SimplifyCFG most loop exit targets have a single predecessor. 1259 // Otherwise fall back to a truncate within the loop. 1260 if (UsePhi->getNumOperands() != 1) 1261 truncateIVUse(DU, DT, LI); 1262 else { 1263 // Widening the PHI requires us to insert a trunc. The logical place 1264 // for this trunc is in the same BB as the PHI. This is not possible if 1265 // the BB is terminated by a catchswitch. 1266 if (isa<CatchSwitchInst>(UsePhi->getParent()->getTerminator())) 1267 return nullptr; 1268 1269 PHINode *WidePhi = 1270 PHINode::Create(DU.WideDef->getType(), 1, UsePhi->getName() + ".wide", 1271 UsePhi); 1272 WidePhi->addIncoming(DU.WideDef, UsePhi->getIncomingBlock(0)); 1273 IRBuilder<> Builder(&*WidePhi->getParent()->getFirstInsertionPt()); 1274 Value *Trunc = Builder.CreateTrunc(WidePhi, DU.NarrowDef->getType()); 1275 UsePhi->replaceAllUsesWith(Trunc); 1276 DeadInsts.emplace_back(UsePhi); 1277 DEBUG(dbgs() << "INDVARS: Widen lcssa phi " << *UsePhi 1278 << " to " << *WidePhi << "\n"); 1279 } 1280 return nullptr; 1281 } 1282 } 1283 // Our raison d'etre! Eliminate sign and zero extension. 1284 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) { 1285 Value *NewDef = DU.WideDef; 1286 if (DU.NarrowUse->getType() != WideType) { 1287 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType()); 1288 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 1289 if (CastWidth < IVWidth) { 1290 // The cast isn't as wide as the IV, so insert a Trunc. 1291 IRBuilder<> Builder(DU.NarrowUse); 1292 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType()); 1293 } 1294 else { 1295 // A wider extend was hidden behind a narrower one. This may induce 1296 // another round of IV widening in which the intermediate IV becomes 1297 // dead. It should be very rare. 1298 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi 1299 << " not wide enough to subsume " << *DU.NarrowUse << "\n"); 1300 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 1301 NewDef = DU.NarrowUse; 1302 } 1303 } 1304 if (NewDef != DU.NarrowUse) { 1305 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse 1306 << " replaced by " << *DU.WideDef << "\n"); 1307 ++NumElimExt; 1308 DU.NarrowUse->replaceAllUsesWith(NewDef); 1309 DeadInsts.emplace_back(DU.NarrowUse); 1310 } 1311 // Now that the extend is gone, we want to expose it's uses for potential 1312 // further simplification. We don't need to directly inform SimplifyIVUsers 1313 // of the new users, because their parent IV will be processed later as a 1314 // new loop phi. If we preserved IVUsers analysis, we would also want to 1315 // push the uses of WideDef here. 1316 1317 // No further widening is needed. The deceased [sz]ext had done it for us. 1318 return nullptr; 1319 } 1320 1321 // Does this user itself evaluate to a recurrence after widening? 1322 const SCEVAddRecExpr *WideAddRec = getWideRecurrence(DU.NarrowUse); 1323 if (!WideAddRec) 1324 WideAddRec = getExtendedOperandRecurrence(DU); 1325 1326 if (!WideAddRec) { 1327 // If use is a loop condition, try to promote the condition instead of 1328 // truncating the IV first. 1329 if (widenLoopCompare(DU)) 1330 return nullptr; 1331 1332 // This user does not evaluate to a recurence after widening, so don't 1333 // follow it. Instead insert a Trunc to kill off the original use, 1334 // eventually isolating the original narrow IV so it can be removed. 1335 truncateIVUse(DU, DT, LI); 1336 return nullptr; 1337 } 1338 // Assume block terminators cannot evaluate to a recurrence. We can't to 1339 // insert a Trunc after a terminator if there happens to be a critical edge. 1340 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() && 1341 "SCEV is not expected to evaluate a block terminator"); 1342 1343 // Reuse the IV increment that SCEVExpander created as long as it dominates 1344 // NarrowUse. 1345 Instruction *WideUse = nullptr; 1346 if (WideAddRec == WideIncExpr && Rewriter.hoistIVInc(WideInc, DU.NarrowUse)) 1347 WideUse = WideInc; 1348 else { 1349 WideUse = cloneIVUser(DU, WideAddRec); 1350 if (!WideUse) 1351 return nullptr; 1352 } 1353 // Evaluation of WideAddRec ensured that the narrow expression could be 1354 // extended outside the loop without overflow. This suggests that the wide use 1355 // evaluates to the same expression as the extended narrow use, but doesn't 1356 // absolutely guarantee it. Hence the following failsafe check. In rare cases 1357 // where it fails, we simply throw away the newly created wide use. 1358 if (WideAddRec != SE->getSCEV(WideUse)) { 1359 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse 1360 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n"); 1361 DeadInsts.emplace_back(WideUse); 1362 return nullptr; 1363 } 1364 1365 // Returning WideUse pushes it on the worklist. 1366 return WideUse; 1367 } 1368 1369 /// Add eligible users of NarrowDef to NarrowIVUsers. 1370 /// 1371 void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) { 1372 const SCEV *NarrowSCEV = SE->getSCEV(NarrowDef); 1373 bool NeverNegative = 1374 SE->isKnownPredicate(ICmpInst::ICMP_SGE, NarrowSCEV, 1375 SE->getConstant(NarrowSCEV->getType(), 0)); 1376 for (User *U : NarrowDef->users()) { 1377 Instruction *NarrowUser = cast<Instruction>(U); 1378 1379 // Handle data flow merges and bizarre phi cycles. 1380 if (!Widened.insert(NarrowUser).second) 1381 continue; 1382 1383 NarrowIVUsers.emplace_back(NarrowDef, NarrowUser, WideDef, NeverNegative); 1384 } 1385 } 1386 1387 /// Process a single induction variable. First use the SCEVExpander to create a 1388 /// wide induction variable that evaluates to the same recurrence as the 1389 /// original narrow IV. Then use a worklist to forward traverse the narrow IV's 1390 /// def-use chain. After widenIVUse has processed all interesting IV users, the 1391 /// narrow IV will be isolated for removal by DeleteDeadPHIs. 1392 /// 1393 /// It would be simpler to delete uses as they are processed, but we must avoid 1394 /// invalidating SCEV expressions. 1395 /// 1396 PHINode *WidenIV::createWideIV(SCEVExpander &Rewriter) { 1397 // Is this phi an induction variable? 1398 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi)); 1399 if (!AddRec) 1400 return nullptr; 1401 1402 // Widen the induction variable expression. 1403 const SCEV *WideIVExpr = IsSigned ? 1404 SE->getSignExtendExpr(AddRec, WideType) : 1405 SE->getZeroExtendExpr(AddRec, WideType); 1406 1407 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType && 1408 "Expect the new IV expression to preserve its type"); 1409 1410 // Can the IV be extended outside the loop without overflow? 1411 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr); 1412 if (!AddRec || AddRec->getLoop() != L) 1413 return nullptr; 1414 1415 // An AddRec must have loop-invariant operands. Since this AddRec is 1416 // materialized by a loop header phi, the expression cannot have any post-loop 1417 // operands, so they must dominate the loop header. 1418 assert( 1419 SE->properlyDominates(AddRec->getStart(), L->getHeader()) && 1420 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) && 1421 "Loop header phi recurrence inputs do not dominate the loop"); 1422 1423 // The rewriter provides a value for the desired IV expression. This may 1424 // either find an existing phi or materialize a new one. Either way, we 1425 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part 1426 // of the phi-SCC dominates the loop entry. 1427 Instruction *InsertPt = &L->getHeader()->front(); 1428 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt)); 1429 1430 // Remembering the WideIV increment generated by SCEVExpander allows 1431 // widenIVUse to reuse it when widening the narrow IV's increment. We don't 1432 // employ a general reuse mechanism because the call above is the only call to 1433 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses. 1434 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1435 WideInc = 1436 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock)); 1437 WideIncExpr = SE->getSCEV(WideInc); 1438 } 1439 1440 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n"); 1441 ++NumWidened; 1442 1443 // Traverse the def-use chain using a worklist starting at the original IV. 1444 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" ); 1445 1446 Widened.insert(OrigPhi); 1447 pushNarrowIVUsers(OrigPhi, WidePhi); 1448 1449 while (!NarrowIVUsers.empty()) { 1450 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val(); 1451 1452 // Process a def-use edge. This may replace the use, so don't hold a 1453 // use_iterator across it. 1454 Instruction *WideUse = widenIVUse(DU, Rewriter); 1455 1456 // Follow all def-use edges from the previous narrow use. 1457 if (WideUse) 1458 pushNarrowIVUsers(DU.NarrowUse, WideUse); 1459 1460 // widenIVUse may have removed the def-use edge. 1461 if (DU.NarrowDef->use_empty()) 1462 DeadInsts.emplace_back(DU.NarrowDef); 1463 } 1464 return WidePhi; 1465 } 1466 1467 //===----------------------------------------------------------------------===// 1468 // Live IV Reduction - Minimize IVs live across the loop. 1469 //===----------------------------------------------------------------------===// 1470 1471 1472 //===----------------------------------------------------------------------===// 1473 // Simplification of IV users based on SCEV evaluation. 1474 //===----------------------------------------------------------------------===// 1475 1476 namespace { 1477 class IndVarSimplifyVisitor : public IVVisitor { 1478 ScalarEvolution *SE; 1479 const TargetTransformInfo *TTI; 1480 PHINode *IVPhi; 1481 1482 public: 1483 WideIVInfo WI; 1484 1485 IndVarSimplifyVisitor(PHINode *IV, ScalarEvolution *SCEV, 1486 const TargetTransformInfo *TTI, 1487 const DominatorTree *DTree) 1488 : SE(SCEV), TTI(TTI), IVPhi(IV) { 1489 DT = DTree; 1490 WI.NarrowIV = IVPhi; 1491 } 1492 1493 // Implement the interface used by simplifyUsersOfIV. 1494 void visitCast(CastInst *Cast) override { visitIVCast(Cast, WI, SE, TTI); } 1495 }; 1496 } 1497 1498 /// Iteratively perform simplification on a worklist of IV users. Each 1499 /// successive simplification may push more users which may themselves be 1500 /// candidates for simplification. 1501 /// 1502 /// Sign/Zero extend elimination is interleaved with IV simplification. 1503 /// 1504 void IndVarSimplify::simplifyAndExtend(Loop *L, 1505 SCEVExpander &Rewriter, 1506 LoopInfo *LI) { 1507 SmallVector<WideIVInfo, 8> WideIVs; 1508 1509 SmallVector<PHINode*, 8> LoopPhis; 1510 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1511 LoopPhis.push_back(cast<PHINode>(I)); 1512 } 1513 // Each round of simplification iterates through the SimplifyIVUsers worklist 1514 // for all current phis, then determines whether any IVs can be 1515 // widened. Widening adds new phis to LoopPhis, inducing another round of 1516 // simplification on the wide IVs. 1517 while (!LoopPhis.empty()) { 1518 // Evaluate as many IV expressions as possible before widening any IVs. This 1519 // forces SCEV to set no-wrap flags before evaluating sign/zero 1520 // extension. The first time SCEV attempts to normalize sign/zero extension, 1521 // the result becomes final. So for the most predictable results, we delay 1522 // evaluation of sign/zero extend evaluation until needed, and avoid running 1523 // other SCEV based analysis prior to simplifyAndExtend. 1524 do { 1525 PHINode *CurrIV = LoopPhis.pop_back_val(); 1526 1527 // Information about sign/zero extensions of CurrIV. 1528 IndVarSimplifyVisitor Visitor(CurrIV, SE, TTI, DT); 1529 1530 Changed |= simplifyUsersOfIV(CurrIV, SE, DT, LI, DeadInsts, &Visitor); 1531 1532 if (Visitor.WI.WidestNativeType) { 1533 WideIVs.push_back(Visitor.WI); 1534 } 1535 } while(!LoopPhis.empty()); 1536 1537 for (; !WideIVs.empty(); WideIVs.pop_back()) { 1538 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts); 1539 if (PHINode *WidePhi = Widener.createWideIV(Rewriter)) { 1540 Changed = true; 1541 LoopPhis.push_back(WidePhi); 1542 } 1543 } 1544 } 1545 } 1546 1547 //===----------------------------------------------------------------------===// 1548 // linearFunctionTestReplace and its kin. Rewrite the loop exit condition. 1549 //===----------------------------------------------------------------------===// 1550 1551 /// Return true if this loop's backedge taken count expression can be safely and 1552 /// cheaply expanded into an instruction sequence that can be used by 1553 /// linearFunctionTestReplace. 1554 /// 1555 /// TODO: This fails for pointer-type loop counters with greater than one byte 1556 /// strides, consequently preventing LFTR from running. For the purpose of LFTR 1557 /// we could skip this check in the case that the LFTR loop counter (chosen by 1558 /// FindLoopCounter) is also pointer type. Instead, we could directly convert 1559 /// the loop test to an inequality test by checking the target data's alignment 1560 /// of element types (given that the initial pointer value originates from or is 1561 /// used by ABI constrained operation, as opposed to inttoptr/ptrtoint). 1562 /// However, we don't yet have a strong motivation for converting loop tests 1563 /// into inequality tests. 1564 static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE, 1565 SCEVExpander &Rewriter) { 1566 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1567 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) || 1568 BackedgeTakenCount->isZero()) 1569 return false; 1570 1571 if (!L->getExitingBlock()) 1572 return false; 1573 1574 // Can't rewrite non-branch yet. 1575 if (!isa<BranchInst>(L->getExitingBlock()->getTerminator())) 1576 return false; 1577 1578 if (Rewriter.isHighCostExpansion(BackedgeTakenCount, L)) 1579 return false; 1580 1581 return true; 1582 } 1583 1584 /// Return the loop header phi IFF IncV adds a loop invariant value to the phi. 1585 static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) { 1586 Instruction *IncI = dyn_cast<Instruction>(IncV); 1587 if (!IncI) 1588 return nullptr; 1589 1590 switch (IncI->getOpcode()) { 1591 case Instruction::Add: 1592 case Instruction::Sub: 1593 break; 1594 case Instruction::GetElementPtr: 1595 // An IV counter must preserve its type. 1596 if (IncI->getNumOperands() == 2) 1597 break; 1598 default: 1599 return nullptr; 1600 } 1601 1602 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0)); 1603 if (Phi && Phi->getParent() == L->getHeader()) { 1604 if (isLoopInvariant(IncI->getOperand(1), L, DT)) 1605 return Phi; 1606 return nullptr; 1607 } 1608 if (IncI->getOpcode() == Instruction::GetElementPtr) 1609 return nullptr; 1610 1611 // Allow add/sub to be commuted. 1612 Phi = dyn_cast<PHINode>(IncI->getOperand(1)); 1613 if (Phi && Phi->getParent() == L->getHeader()) { 1614 if (isLoopInvariant(IncI->getOperand(0), L, DT)) 1615 return Phi; 1616 } 1617 return nullptr; 1618 } 1619 1620 /// Return the compare guarding the loop latch, or NULL for unrecognized tests. 1621 static ICmpInst *getLoopTest(Loop *L) { 1622 assert(L->getExitingBlock() && "expected loop exit"); 1623 1624 BasicBlock *LatchBlock = L->getLoopLatch(); 1625 // Don't bother with LFTR if the loop is not properly simplified. 1626 if (!LatchBlock) 1627 return nullptr; 1628 1629 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1630 assert(BI && "expected exit branch"); 1631 1632 return dyn_cast<ICmpInst>(BI->getCondition()); 1633 } 1634 1635 /// linearFunctionTestReplace policy. Return true unless we can show that the 1636 /// current exit test is already sufficiently canonical. 1637 static bool needsLFTR(Loop *L, DominatorTree *DT) { 1638 // Do LFTR to simplify the exit condition to an ICMP. 1639 ICmpInst *Cond = getLoopTest(L); 1640 if (!Cond) 1641 return true; 1642 1643 // Do LFTR to simplify the exit ICMP to EQ/NE 1644 ICmpInst::Predicate Pred = Cond->getPredicate(); 1645 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ) 1646 return true; 1647 1648 // Look for a loop invariant RHS 1649 Value *LHS = Cond->getOperand(0); 1650 Value *RHS = Cond->getOperand(1); 1651 if (!isLoopInvariant(RHS, L, DT)) { 1652 if (!isLoopInvariant(LHS, L, DT)) 1653 return true; 1654 std::swap(LHS, RHS); 1655 } 1656 // Look for a simple IV counter LHS 1657 PHINode *Phi = dyn_cast<PHINode>(LHS); 1658 if (!Phi) 1659 Phi = getLoopPhiForCounter(LHS, L, DT); 1660 1661 if (!Phi) 1662 return true; 1663 1664 // Do LFTR if PHI node is defined in the loop, but is *not* a counter. 1665 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch()); 1666 if (Idx < 0) 1667 return true; 1668 1669 // Do LFTR if the exit condition's IV is *not* a simple counter. 1670 Value *IncV = Phi->getIncomingValue(Idx); 1671 return Phi != getLoopPhiForCounter(IncV, L, DT); 1672 } 1673 1674 /// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils 1675 /// down to checking that all operands are constant and listing instructions 1676 /// that may hide undef. 1677 static bool hasConcreteDefImpl(Value *V, SmallPtrSetImpl<Value*> &Visited, 1678 unsigned Depth) { 1679 if (isa<Constant>(V)) 1680 return !isa<UndefValue>(V); 1681 1682 if (Depth >= 6) 1683 return false; 1684 1685 // Conservatively handle non-constant non-instructions. For example, Arguments 1686 // may be undef. 1687 Instruction *I = dyn_cast<Instruction>(V); 1688 if (!I) 1689 return false; 1690 1691 // Load and return values may be undef. 1692 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I)) 1693 return false; 1694 1695 // Optimistically handle other instructions. 1696 for (Value *Op : I->operands()) { 1697 if (!Visited.insert(Op).second) 1698 continue; 1699 if (!hasConcreteDefImpl(Op, Visited, Depth+1)) 1700 return false; 1701 } 1702 return true; 1703 } 1704 1705 /// Return true if the given value is concrete. We must prove that undef can 1706 /// never reach it. 1707 /// 1708 /// TODO: If we decide that this is a good approach to checking for undef, we 1709 /// may factor it into a common location. 1710 static bool hasConcreteDef(Value *V) { 1711 SmallPtrSet<Value*, 8> Visited; 1712 Visited.insert(V); 1713 return hasConcreteDefImpl(V, Visited, 0); 1714 } 1715 1716 /// Return true if this IV has any uses other than the (soon to be rewritten) 1717 /// loop exit test. 1718 static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) { 1719 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1720 Value *IncV = Phi->getIncomingValue(LatchIdx); 1721 1722 for (User *U : Phi->users()) 1723 if (U != Cond && U != IncV) return false; 1724 1725 for (User *U : IncV->users()) 1726 if (U != Cond && U != Phi) return false; 1727 return true; 1728 } 1729 1730 /// Find an affine IV in canonical form. 1731 /// 1732 /// BECount may be an i8* pointer type. The pointer difference is already 1733 /// valid count without scaling the address stride, so it remains a pointer 1734 /// expression as far as SCEV is concerned. 1735 /// 1736 /// Currently only valid for LFTR. See the comments on hasConcreteDef below. 1737 /// 1738 /// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount 1739 /// 1740 /// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride. 1741 /// This is difficult in general for SCEV because of potential overflow. But we 1742 /// could at least handle constant BECounts. 1743 static PHINode *FindLoopCounter(Loop *L, const SCEV *BECount, 1744 ScalarEvolution *SE, DominatorTree *DT) { 1745 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType()); 1746 1747 Value *Cond = 1748 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition(); 1749 1750 // Loop over all of the PHI nodes, looking for a simple counter. 1751 PHINode *BestPhi = nullptr; 1752 const SCEV *BestInit = nullptr; 1753 BasicBlock *LatchBlock = L->getLoopLatch(); 1754 assert(LatchBlock && "needsLFTR should guarantee a loop latch"); 1755 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 1756 1757 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1758 PHINode *Phi = cast<PHINode>(I); 1759 if (!SE->isSCEVable(Phi->getType())) 1760 continue; 1761 1762 // Avoid comparing an integer IV against a pointer Limit. 1763 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy()) 1764 continue; 1765 1766 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi)); 1767 if (!AR || AR->getLoop() != L || !AR->isAffine()) 1768 continue; 1769 1770 // AR may be a pointer type, while BECount is an integer type. 1771 // AR may be wider than BECount. With eq/ne tests overflow is immaterial. 1772 // AR may not be a narrower type, or we may never exit. 1773 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType()); 1774 if (PhiWidth < BCWidth || !DL.isLegalInteger(PhiWidth)) 1775 continue; 1776 1777 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)); 1778 if (!Step || !Step->isOne()) 1779 continue; 1780 1781 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1782 Value *IncV = Phi->getIncomingValue(LatchIdx); 1783 if (getLoopPhiForCounter(IncV, L, DT) != Phi) 1784 continue; 1785 1786 // Avoid reusing a potentially undef value to compute other values that may 1787 // have originally had a concrete definition. 1788 if (!hasConcreteDef(Phi)) { 1789 // We explicitly allow unknown phis as long as they are already used by 1790 // the loop test. In this case we assume that performing LFTR could not 1791 // increase the number of undef users. 1792 if (ICmpInst *Cond = getLoopTest(L)) { 1793 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT) && 1794 Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) { 1795 continue; 1796 } 1797 } 1798 } 1799 const SCEV *Init = AR->getStart(); 1800 1801 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) { 1802 // Don't force a live loop counter if another IV can be used. 1803 if (AlmostDeadIV(Phi, LatchBlock, Cond)) 1804 continue; 1805 1806 // Prefer to count-from-zero. This is a more "canonical" counter form. It 1807 // also prefers integer to pointer IVs. 1808 if (BestInit->isZero() != Init->isZero()) { 1809 if (BestInit->isZero()) 1810 continue; 1811 } 1812 // If two IVs both count from zero or both count from nonzero then the 1813 // narrower is likely a dead phi that has been widened. Use the wider phi 1814 // to allow the other to be eliminated. 1815 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType())) 1816 continue; 1817 } 1818 BestPhi = Phi; 1819 BestInit = Init; 1820 } 1821 return BestPhi; 1822 } 1823 1824 /// Help linearFunctionTestReplace by generating a value that holds the RHS of 1825 /// the new loop test. 1826 static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L, 1827 SCEVExpander &Rewriter, ScalarEvolution *SE) { 1828 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1829 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter"); 1830 const SCEV *IVInit = AR->getStart(); 1831 1832 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter 1833 // finds a valid pointer IV. Sign extend BECount in order to materialize a 1834 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing 1835 // the existing GEPs whenever possible. 1836 if (IndVar->getType()->isPointerTy() && !IVCount->getType()->isPointerTy()) { 1837 // IVOffset will be the new GEP offset that is interpreted by GEP as a 1838 // signed value. IVCount on the other hand represents the loop trip count, 1839 // which is an unsigned value. FindLoopCounter only allows induction 1840 // variables that have a positive unit stride of one. This means we don't 1841 // have to handle the case of negative offsets (yet) and just need to zero 1842 // extend IVCount. 1843 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType()); 1844 const SCEV *IVOffset = SE->getTruncateOrZeroExtend(IVCount, OfsTy); 1845 1846 // Expand the code for the iteration count. 1847 assert(SE->isLoopInvariant(IVOffset, L) && 1848 "Computed iteration count is not loop invariant!"); 1849 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1850 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI); 1851 1852 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader()); 1853 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter"); 1854 // We could handle pointer IVs other than i8*, but we need to compensate for 1855 // gep index scaling. See canExpandBackedgeTakenCount comments. 1856 assert(SE->getSizeOfExpr(IntegerType::getInt64Ty(IndVar->getContext()), 1857 cast<PointerType>(GEPBase->getType()) 1858 ->getElementType())->isOne() && 1859 "unit stride pointer IV must be i8*"); 1860 1861 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); 1862 return Builder.CreateGEP(nullptr, GEPBase, GEPOffset, "lftr.limit"); 1863 } else { 1864 // In any other case, convert both IVInit and IVCount to integers before 1865 // comparing. This may result in SCEV expension of pointers, but in practice 1866 // SCEV will fold the pointer arithmetic away as such: 1867 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc). 1868 // 1869 // Valid Cases: (1) both integers is most common; (2) both may be pointers 1870 // for simple memset-style loops. 1871 // 1872 // IVInit integer and IVCount pointer would only occur if a canonical IV 1873 // were generated on top of case #2, which is not expected. 1874 1875 const SCEV *IVLimit = nullptr; 1876 // For unit stride, IVCount = Start + BECount with 2's complement overflow. 1877 // For non-zero Start, compute IVCount here. 1878 if (AR->getStart()->isZero()) 1879 IVLimit = IVCount; 1880 else { 1881 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride"); 1882 const SCEV *IVInit = AR->getStart(); 1883 1884 // For integer IVs, truncate the IV before computing IVInit + BECount. 1885 if (SE->getTypeSizeInBits(IVInit->getType()) 1886 > SE->getTypeSizeInBits(IVCount->getType())) 1887 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType()); 1888 1889 IVLimit = SE->getAddExpr(IVInit, IVCount); 1890 } 1891 // Expand the code for the iteration count. 1892 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1893 IRBuilder<> Builder(BI); 1894 assert(SE->isLoopInvariant(IVLimit, L) && 1895 "Computed iteration count is not loop invariant!"); 1896 // Ensure that we generate the same type as IndVar, or a smaller integer 1897 // type. In the presence of null pointer values, we have an integer type 1898 // SCEV expression (IVInit) for a pointer type IV value (IndVar). 1899 Type *LimitTy = IVCount->getType()->isPointerTy() ? 1900 IndVar->getType() : IVCount->getType(); 1901 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI); 1902 } 1903 } 1904 1905 /// This method rewrites the exit condition of the loop to be a canonical != 1906 /// comparison against the incremented loop induction variable. This pass is 1907 /// able to rewrite the exit tests of any loop where the SCEV analysis can 1908 /// determine a loop-invariant trip count of the loop, which is actually a much 1909 /// broader range than just linear tests. 1910 Value *IndVarSimplify:: 1911 linearFunctionTestReplace(Loop *L, 1912 const SCEV *BackedgeTakenCount, 1913 PHINode *IndVar, 1914 SCEVExpander &Rewriter) { 1915 assert(canExpandBackedgeTakenCount(L, SE, Rewriter) && "precondition"); 1916 1917 // Initialize CmpIndVar and IVCount to their preincremented values. 1918 Value *CmpIndVar = IndVar; 1919 const SCEV *IVCount = BackedgeTakenCount; 1920 1921 // If the exiting block is the same as the backedge block, we prefer to 1922 // compare against the post-incremented value, otherwise we must compare 1923 // against the preincremented value. 1924 if (L->getExitingBlock() == L->getLoopLatch()) { 1925 // Add one to the "backedge-taken" count to get the trip count. 1926 // This addition may overflow, which is valid as long as the comparison is 1927 // truncated to BackedgeTakenCount->getType(). 1928 IVCount = SE->getAddExpr(BackedgeTakenCount, 1929 SE->getOne(BackedgeTakenCount->getType())); 1930 // The BackedgeTaken expression contains the number of times that the 1931 // backedge branches to the loop header. This is one less than the 1932 // number of times the loop executes, so use the incremented indvar. 1933 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock()); 1934 } 1935 1936 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE); 1937 assert(ExitCnt->getType()->isPointerTy() == 1938 IndVar->getType()->isPointerTy() && 1939 "genLoopLimit missed a cast"); 1940 1941 // Insert a new icmp_ne or icmp_eq instruction before the branch. 1942 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1943 ICmpInst::Predicate P; 1944 if (L->contains(BI->getSuccessor(0))) 1945 P = ICmpInst::ICMP_NE; 1946 else 1947 P = ICmpInst::ICMP_EQ; 1948 1949 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n" 1950 << " LHS:" << *CmpIndVar << '\n' 1951 << " op:\t" 1952 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n" 1953 << " RHS:\t" << *ExitCnt << "\n" 1954 << " IVCount:\t" << *IVCount << "\n"); 1955 1956 IRBuilder<> Builder(BI); 1957 1958 // LFTR can ignore IV overflow and truncate to the width of 1959 // BECount. This avoids materializing the add(zext(add)) expression. 1960 unsigned CmpIndVarSize = SE->getTypeSizeInBits(CmpIndVar->getType()); 1961 unsigned ExitCntSize = SE->getTypeSizeInBits(ExitCnt->getType()); 1962 if (CmpIndVarSize > ExitCntSize) { 1963 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1964 const SCEV *ARStart = AR->getStart(); 1965 const SCEV *ARStep = AR->getStepRecurrence(*SE); 1966 // For constant IVCount, avoid truncation. 1967 if (isa<SCEVConstant>(ARStart) && isa<SCEVConstant>(IVCount)) { 1968 const APInt &Start = cast<SCEVConstant>(ARStart)->getAPInt(); 1969 APInt Count = cast<SCEVConstant>(IVCount)->getAPInt(); 1970 // Note that the post-inc value of BackedgeTakenCount may have overflowed 1971 // above such that IVCount is now zero. 1972 if (IVCount != BackedgeTakenCount && Count == 0) { 1973 Count = APInt::getMaxValue(Count.getBitWidth()).zext(CmpIndVarSize); 1974 ++Count; 1975 } 1976 else 1977 Count = Count.zext(CmpIndVarSize); 1978 APInt NewLimit; 1979 if (cast<SCEVConstant>(ARStep)->getValue()->isNegative()) 1980 NewLimit = Start - Count; 1981 else 1982 NewLimit = Start + Count; 1983 ExitCnt = ConstantInt::get(CmpIndVar->getType(), NewLimit); 1984 1985 DEBUG(dbgs() << " Widen RHS:\t" << *ExitCnt << "\n"); 1986 } else { 1987 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(), 1988 "lftr.wideiv"); 1989 } 1990 } 1991 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond"); 1992 Value *OrigCond = BI->getCondition(); 1993 // It's tempting to use replaceAllUsesWith here to fully replace the old 1994 // comparison, but that's not immediately safe, since users of the old 1995 // comparison may not be dominated by the new comparison. Instead, just 1996 // update the branch to use the new comparison; in the common case this 1997 // will make old comparison dead. 1998 BI->setCondition(Cond); 1999 DeadInsts.push_back(OrigCond); 2000 2001 ++NumLFTR; 2002 Changed = true; 2003 return Cond; 2004 } 2005 2006 //===----------------------------------------------------------------------===// 2007 // sinkUnusedInvariants. A late subpass to cleanup loop preheaders. 2008 //===----------------------------------------------------------------------===// 2009 2010 /// If there's a single exit block, sink any loop-invariant values that 2011 /// were defined in the preheader but not used inside the loop into the 2012 /// exit block to reduce register pressure in the loop. 2013 void IndVarSimplify::sinkUnusedInvariants(Loop *L) { 2014 BasicBlock *ExitBlock = L->getExitBlock(); 2015 if (!ExitBlock) return; 2016 2017 BasicBlock *Preheader = L->getLoopPreheader(); 2018 if (!Preheader) return; 2019 2020 Instruction *InsertPt = &*ExitBlock->getFirstInsertionPt(); 2021 BasicBlock::iterator I(Preheader->getTerminator()); 2022 while (I != Preheader->begin()) { 2023 --I; 2024 // New instructions were inserted at the end of the preheader. 2025 if (isa<PHINode>(I)) 2026 break; 2027 2028 // Don't move instructions which might have side effects, since the side 2029 // effects need to complete before instructions inside the loop. Also don't 2030 // move instructions which might read memory, since the loop may modify 2031 // memory. Note that it's okay if the instruction might have undefined 2032 // behavior: LoopSimplify guarantees that the preheader dominates the exit 2033 // block. 2034 if (I->mayHaveSideEffects() || I->mayReadFromMemory()) 2035 continue; 2036 2037 // Skip debug info intrinsics. 2038 if (isa<DbgInfoIntrinsic>(I)) 2039 continue; 2040 2041 // Skip eh pad instructions. 2042 if (I->isEHPad()) 2043 continue; 2044 2045 // Don't sink alloca: we never want to sink static alloca's out of the 2046 // entry block, and correctly sinking dynamic alloca's requires 2047 // checks for stacksave/stackrestore intrinsics. 2048 // FIXME: Refactor this check somehow? 2049 if (isa<AllocaInst>(I)) 2050 continue; 2051 2052 // Determine if there is a use in or before the loop (direct or 2053 // otherwise). 2054 bool UsedInLoop = false; 2055 for (Use &U : I->uses()) { 2056 Instruction *User = cast<Instruction>(U.getUser()); 2057 BasicBlock *UseBB = User->getParent(); 2058 if (PHINode *P = dyn_cast<PHINode>(User)) { 2059 unsigned i = 2060 PHINode::getIncomingValueNumForOperand(U.getOperandNo()); 2061 UseBB = P->getIncomingBlock(i); 2062 } 2063 if (UseBB == Preheader || L->contains(UseBB)) { 2064 UsedInLoop = true; 2065 break; 2066 } 2067 } 2068 2069 // If there is, the def must remain in the preheader. 2070 if (UsedInLoop) 2071 continue; 2072 2073 // Otherwise, sink it to the exit block. 2074 Instruction *ToMove = &*I; 2075 bool Done = false; 2076 2077 if (I != Preheader->begin()) { 2078 // Skip debug info intrinsics. 2079 do { 2080 --I; 2081 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin()); 2082 2083 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin()) 2084 Done = true; 2085 } else { 2086 Done = true; 2087 } 2088 2089 ToMove->moveBefore(InsertPt); 2090 if (Done) break; 2091 InsertPt = ToMove; 2092 } 2093 } 2094 2095 //===----------------------------------------------------------------------===// 2096 // IndVarSimplify driver. Manage several subpasses of IV simplification. 2097 //===----------------------------------------------------------------------===// 2098 2099 bool IndVarSimplify::run(Loop *L) { 2100 // We need (and expect!) the incoming loop to be in LCSSA. 2101 assert(L->isRecursivelyLCSSAForm(*DT) && "LCSSA required to run indvars!"); 2102 2103 // If LoopSimplify form is not available, stay out of trouble. Some notes: 2104 // - LSR currently only supports LoopSimplify-form loops. Indvars' 2105 // canonicalization can be a pessimization without LSR to "clean up" 2106 // afterwards. 2107 // - We depend on having a preheader; in particular, 2108 // Loop::getCanonicalInductionVariable only supports loops with preheaders, 2109 // and we're in trouble if we can't find the induction variable even when 2110 // we've manually inserted one. 2111 if (!L->isLoopSimplifyForm()) 2112 return false; 2113 2114 // If there are any floating-point recurrences, attempt to 2115 // transform them to use integer recurrences. 2116 rewriteNonIntegerIVs(L); 2117 2118 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 2119 2120 // Create a rewriter object which we'll use to transform the code with. 2121 SCEVExpander Rewriter(*SE, DL, "indvars"); 2122 #ifndef NDEBUG 2123 Rewriter.setDebugType(DEBUG_TYPE); 2124 #endif 2125 2126 // Eliminate redundant IV users. 2127 // 2128 // Simplification works best when run before other consumers of SCEV. We 2129 // attempt to avoid evaluating SCEVs for sign/zero extend operations until 2130 // other expressions involving loop IVs have been evaluated. This helps SCEV 2131 // set no-wrap flags before normalizing sign/zero extension. 2132 Rewriter.disableCanonicalMode(); 2133 simplifyAndExtend(L, Rewriter, LI); 2134 2135 // Check to see if this loop has a computable loop-invariant execution count. 2136 // If so, this means that we can compute the final value of any expressions 2137 // that are recurrent in the loop, and substitute the exit values from the 2138 // loop into any instructions outside of the loop that use the final values of 2139 // the current expressions. 2140 // 2141 if (ReplaceExitValue != NeverRepl && 2142 !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 2143 rewriteLoopExitValues(L, Rewriter); 2144 2145 // Eliminate redundant IV cycles. 2146 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts); 2147 2148 // If we have a trip count expression, rewrite the loop's exit condition 2149 // using it. We can currently only handle loops with a single exit. 2150 if (canExpandBackedgeTakenCount(L, SE, Rewriter) && needsLFTR(L, DT)) { 2151 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT); 2152 if (IndVar) { 2153 // Check preconditions for proper SCEVExpander operation. SCEV does not 2154 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any 2155 // pass that uses the SCEVExpander must do it. This does not work well for 2156 // loop passes because SCEVExpander makes assumptions about all loops, 2157 // while LoopPassManager only forces the current loop to be simplified. 2158 // 2159 // FIXME: SCEV expansion has no way to bail out, so the caller must 2160 // explicitly check any assumptions made by SCEV. Brittle. 2161 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount); 2162 if (!AR || AR->getLoop()->getLoopPreheader()) 2163 (void)linearFunctionTestReplace(L, BackedgeTakenCount, IndVar, 2164 Rewriter); 2165 } 2166 } 2167 // Clear the rewriter cache, because values that are in the rewriter's cache 2168 // can be deleted in the loop below, causing the AssertingVH in the cache to 2169 // trigger. 2170 Rewriter.clear(); 2171 2172 // Now that we're done iterating through lists, clean up any instructions 2173 // which are now dead. 2174 while (!DeadInsts.empty()) 2175 if (Instruction *Inst = 2176 dyn_cast_or_null<Instruction>(DeadInsts.pop_back_val())) 2177 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI); 2178 2179 // The Rewriter may not be used from this point on. 2180 2181 // Loop-invariant instructions in the preheader that aren't used in the 2182 // loop may be sunk below the loop to reduce register pressure. 2183 sinkUnusedInvariants(L); 2184 2185 // rewriteFirstIterationLoopExitValues does not rely on the computation of 2186 // trip count and therefore can further simplify exit values in addition to 2187 // rewriteLoopExitValues. 2188 rewriteFirstIterationLoopExitValues(L); 2189 2190 // Clean up dead instructions. 2191 Changed |= DeleteDeadPHIs(L->getHeader(), TLI); 2192 2193 // Check a post-condition. 2194 assert(L->isRecursivelyLCSSAForm(*DT) && "Indvars did not preserve LCSSA!"); 2195 2196 // Verify that LFTR, and any other change have not interfered with SCEV's 2197 // ability to compute trip count. 2198 #ifndef NDEBUG 2199 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { 2200 SE->forgetLoop(L); 2201 const SCEV *NewBECount = SE->getBackedgeTakenCount(L); 2202 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) < 2203 SE->getTypeSizeInBits(NewBECount->getType())) 2204 NewBECount = SE->getTruncateOrNoop(NewBECount, 2205 BackedgeTakenCount->getType()); 2206 else 2207 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, 2208 NewBECount->getType()); 2209 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV"); 2210 } 2211 #endif 2212 2213 return Changed; 2214 } 2215 2216 PreservedAnalyses IndVarSimplifyPass::run(Loop &L, AnalysisManager<Loop> &AM) { 2217 auto &FAM = AM.getResult<FunctionAnalysisManagerLoopProxy>(L).getManager(); 2218 Function *F = L.getHeader()->getParent(); 2219 const DataLayout &DL = F->getParent()->getDataLayout(); 2220 2221 auto *LI = FAM.getCachedResult<LoopAnalysis>(*F); 2222 auto *SE = FAM.getCachedResult<ScalarEvolutionAnalysis>(*F); 2223 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(*F); 2224 2225 assert((LI && SE && DT) && 2226 "Analyses required for indvarsimplify not available!"); 2227 2228 // Optional analyses. 2229 auto *TTI = FAM.getCachedResult<TargetIRAnalysis>(*F); 2230 auto *TLI = FAM.getCachedResult<TargetLibraryAnalysis>(*F); 2231 2232 IndVarSimplify IVS(LI, SE, DT, DL, TLI, TTI); 2233 if (!IVS.run(&L)) 2234 return PreservedAnalyses::all(); 2235 2236 // FIXME: This should also 'preserve the CFG'. 2237 return getLoopPassPreservedAnalyses(); 2238 } 2239 2240 namespace { 2241 struct IndVarSimplifyLegacyPass : public LoopPass { 2242 static char ID; // Pass identification, replacement for typeid 2243 IndVarSimplifyLegacyPass() : LoopPass(ID) { 2244 initializeIndVarSimplifyLegacyPassPass(*PassRegistry::getPassRegistry()); 2245 } 2246 2247 bool runOnLoop(Loop *L, LPPassManager &LPM) override { 2248 if (skipLoop(L)) 2249 return false; 2250 2251 auto *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 2252 auto *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE(); 2253 auto *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 2254 auto *TLIP = getAnalysisIfAvailable<TargetLibraryInfoWrapperPass>(); 2255 auto *TLI = TLIP ? &TLIP->getTLI() : nullptr; 2256 auto *TTIP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>(); 2257 auto *TTI = TTIP ? &TTIP->getTTI(*L->getHeader()->getParent()) : nullptr; 2258 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 2259 2260 IndVarSimplify IVS(LI, SE, DT, DL, TLI, TTI); 2261 return IVS.run(L); 2262 } 2263 2264 void getAnalysisUsage(AnalysisUsage &AU) const override { 2265 AU.setPreservesCFG(); 2266 getLoopAnalysisUsage(AU); 2267 } 2268 }; 2269 } 2270 2271 char IndVarSimplifyLegacyPass::ID = 0; 2272 INITIALIZE_PASS_BEGIN(IndVarSimplifyLegacyPass, "indvars", 2273 "Induction Variable Simplification", false, false) 2274 INITIALIZE_PASS_DEPENDENCY(LoopPass) 2275 INITIALIZE_PASS_END(IndVarSimplifyLegacyPass, "indvars", 2276 "Induction Variable Simplification", false, false) 2277 2278 Pass *llvm::createIndVarSimplifyPass() { 2279 return new IndVarSimplifyLegacyPass(); 2280 } 2281