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