IndVarSimplify.cpp revision 7137909128e6447c556b2713a60cd8e62f823612
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) || !isSafeToExpand(ExitValue)) 536 continue; 537 538 // Computing the value outside of the loop brings no benefit if : 539 // - it is definitely used inside the loop in a way which can not be 540 // optimized away. 541 // - no use outside of the loop can take advantage of hoisting the 542 // computation out of the loop 543 if (ExitValue->getSCEVType()>=scMulExpr) { 544 unsigned NumHardInternalUses = 0; 545 unsigned NumSoftExternalUses = 0; 546 unsigned NumUses = 0; 547 for (Value::use_iterator IB=Inst->use_begin(), IE=Inst->use_end(); 548 IB!=IE && NumUses<=6 ; ++IB) { 549 Instruction *UseInstr = cast<Instruction>(*IB); 550 unsigned Opc = UseInstr->getOpcode(); 551 NumUses++; 552 if (L->contains(UseInstr)) { 553 if (Opc == Instruction::Call || Opc == Instruction::Ret) 554 NumHardInternalUses++; 555 } else { 556 if (Opc == Instruction::PHI) { 557 // Do not count the Phi as a use. LCSSA may have inserted 558 // plenty of trivial ones. 559 NumUses--; 560 for (Value::use_iterator PB=UseInstr->use_begin(), 561 PE=UseInstr->use_end(); 562 PB!=PE && NumUses<=6 ; ++PB, ++NumUses) { 563 unsigned PhiOpc = cast<Instruction>(*PB)->getOpcode(); 564 if (PhiOpc != Instruction::Call && PhiOpc != Instruction::Ret) 565 NumSoftExternalUses++; 566 } 567 continue; 568 } 569 if (Opc != Instruction::Call && Opc != Instruction::Ret) 570 NumSoftExternalUses++; 571 } 572 } 573 if (NumUses <= 6 && NumHardInternalUses && !NumSoftExternalUses) 574 continue; 575 } 576 577 Value *ExitVal = Rewriter.expandCodeFor(ExitValue, PN->getType(), Inst); 578 579 DEBUG(dbgs() << "INDVARS: RLEV: AfterLoopVal = " << *ExitVal << '\n' 580 << " LoopVal = " << *Inst << "\n"); 581 582 if (!isValidRewrite(Inst, ExitVal)) { 583 DeadInsts.push_back(ExitVal); 584 continue; 585 } 586 Changed = true; 587 ++NumReplaced; 588 589 PN->setIncomingValue(i, ExitVal); 590 591 // If this instruction is dead now, delete it. Don't do it now to avoid 592 // invalidating iterators. 593 if (isInstructionTriviallyDead(Inst, TLI)) 594 DeadInsts.push_back(Inst); 595 596 if (NumPreds == 1) { 597 // Completely replace a single-pred PHI. This is safe, because the 598 // NewVal won't be variant in the loop, so we don't need an LCSSA phi 599 // node anymore. 600 PN->replaceAllUsesWith(ExitVal); 601 PN->eraseFromParent(); 602 } 603 } 604 if (NumPreds != 1) { 605 // Clone the PHI and delete the original one. This lets IVUsers and 606 // any other maps purge the original user from their records. 607 PHINode *NewPN = cast<PHINode>(PN->clone()); 608 NewPN->takeName(PN); 609 NewPN->insertBefore(PN); 610 PN->replaceAllUsesWith(NewPN); 611 PN->eraseFromParent(); 612 } 613 } 614 } 615 616 // The insertion point instruction may have been deleted; clear it out 617 // so that the rewriter doesn't trip over it later. 618 Rewriter.clearInsertPoint(); 619} 620 621//===----------------------------------------------------------------------===// 622// IV Widening - Extend the width of an IV to cover its widest uses. 623//===----------------------------------------------------------------------===// 624 625namespace { 626 // Collect information about induction variables that are used by sign/zero 627 // extend operations. This information is recorded by CollectExtend and 628 // provides the input to WidenIV. 629 struct WideIVInfo { 630 PHINode *NarrowIV; 631 Type *WidestNativeType; // Widest integer type created [sz]ext 632 bool IsSigned; // Was an sext user seen before a zext? 633 634 WideIVInfo() : NarrowIV(0), WidestNativeType(0), IsSigned(false) {} 635 }; 636 637 class WideIVVisitor : public IVVisitor { 638 ScalarEvolution *SE; 639 const DataLayout *TD; 640 641 public: 642 WideIVInfo WI; 643 644 WideIVVisitor(PHINode *NarrowIV, ScalarEvolution *SCEV, 645 const DataLayout *TData) : 646 SE(SCEV), TD(TData) { WI.NarrowIV = NarrowIV; } 647 648 // Implement the interface used by simplifyUsersOfIV. 649 virtual void visitCast(CastInst *Cast); 650 }; 651} 652 653/// visitCast - Update information about the induction variable that is 654/// extended by this sign or zero extend operation. This is used to determine 655/// the final width of the IV before actually widening it. 656void WideIVVisitor::visitCast(CastInst *Cast) { 657 bool IsSigned = Cast->getOpcode() == Instruction::SExt; 658 if (!IsSigned && Cast->getOpcode() != Instruction::ZExt) 659 return; 660 661 Type *Ty = Cast->getType(); 662 uint64_t Width = SE->getTypeSizeInBits(Ty); 663 if (TD && !TD->isLegalInteger(Width)) 664 return; 665 666 if (!WI.WidestNativeType) { 667 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 668 WI.IsSigned = IsSigned; 669 return; 670 } 671 672 // We extend the IV to satisfy the sign of its first user, arbitrarily. 673 if (WI.IsSigned != IsSigned) 674 return; 675 676 if (Width > SE->getTypeSizeInBits(WI.WidestNativeType)) 677 WI.WidestNativeType = SE->getEffectiveSCEVType(Ty); 678} 679 680namespace { 681 682/// NarrowIVDefUse - Record a link in the Narrow IV def-use chain along with the 683/// WideIV that computes the same value as the Narrow IV def. This avoids 684/// caching Use* pointers. 685struct NarrowIVDefUse { 686 Instruction *NarrowDef; 687 Instruction *NarrowUse; 688 Instruction *WideDef; 689 690 NarrowIVDefUse(): NarrowDef(0), NarrowUse(0), WideDef(0) {} 691 692 NarrowIVDefUse(Instruction *ND, Instruction *NU, Instruction *WD): 693 NarrowDef(ND), NarrowUse(NU), WideDef(WD) {} 694}; 695 696/// WidenIV - The goal of this transform is to remove sign and zero extends 697/// without creating any new induction variables. To do this, it creates a new 698/// phi of the wider type and redirects all users, either removing extends or 699/// inserting truncs whenever we stop propagating the type. 700/// 701class WidenIV { 702 // Parameters 703 PHINode *OrigPhi; 704 Type *WideType; 705 bool IsSigned; 706 707 // Context 708 LoopInfo *LI; 709 Loop *L; 710 ScalarEvolution *SE; 711 DominatorTree *DT; 712 713 // Result 714 PHINode *WidePhi; 715 Instruction *WideInc; 716 const SCEV *WideIncExpr; 717 SmallVectorImpl<WeakVH> &DeadInsts; 718 719 SmallPtrSet<Instruction*,16> Widened; 720 SmallVector<NarrowIVDefUse, 8> NarrowIVUsers; 721 722public: 723 WidenIV(const WideIVInfo &WI, LoopInfo *LInfo, 724 ScalarEvolution *SEv, DominatorTree *DTree, 725 SmallVectorImpl<WeakVH> &DI) : 726 OrigPhi(WI.NarrowIV), 727 WideType(WI.WidestNativeType), 728 IsSigned(WI.IsSigned), 729 LI(LInfo), 730 L(LI->getLoopFor(OrigPhi->getParent())), 731 SE(SEv), 732 DT(DTree), 733 WidePhi(0), 734 WideInc(0), 735 WideIncExpr(0), 736 DeadInsts(DI) { 737 assert(L->getHeader() == OrigPhi->getParent() && "Phi must be an IV"); 738 } 739 740 PHINode *CreateWideIV(SCEVExpander &Rewriter); 741 742protected: 743 Value *getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, 744 Instruction *Use); 745 746 Instruction *CloneIVUser(NarrowIVDefUse DU); 747 748 const SCEVAddRecExpr *GetWideRecurrence(Instruction *NarrowUse); 749 750 const SCEVAddRecExpr* GetExtendedOperandRecurrence(NarrowIVDefUse DU); 751 752 Instruction *WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter); 753 754 void pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef); 755}; 756} // anonymous namespace 757 758/// isLoopInvariant - Perform a quick domtree based check for loop invariance 759/// assuming that V is used within the loop. LoopInfo::isLoopInvariant() seems 760/// gratuitous for this purpose. 761static bool isLoopInvariant(Value *V, const Loop *L, const DominatorTree *DT) { 762 Instruction *Inst = dyn_cast<Instruction>(V); 763 if (!Inst) 764 return true; 765 766 return DT->properlyDominates(Inst->getParent(), L->getHeader()); 767} 768 769Value *WidenIV::getExtend(Value *NarrowOper, Type *WideType, bool IsSigned, 770 Instruction *Use) { 771 // Set the debug location and conservative insertion point. 772 IRBuilder<> Builder(Use); 773 // Hoist the insertion point into loop preheaders as far as possible. 774 for (const Loop *L = LI->getLoopFor(Use->getParent()); 775 L && L->getLoopPreheader() && isLoopInvariant(NarrowOper, L, DT); 776 L = L->getParentLoop()) 777 Builder.SetInsertPoint(L->getLoopPreheader()->getTerminator()); 778 779 return IsSigned ? Builder.CreateSExt(NarrowOper, WideType) : 780 Builder.CreateZExt(NarrowOper, WideType); 781} 782 783/// CloneIVUser - Instantiate a wide operation to replace a narrow 784/// operation. This only needs to handle operations that can evaluation to 785/// SCEVAddRec. It can safely return 0 for any operation we decide not to clone. 786Instruction *WidenIV::CloneIVUser(NarrowIVDefUse DU) { 787 unsigned Opcode = DU.NarrowUse->getOpcode(); 788 switch (Opcode) { 789 default: 790 return 0; 791 case Instruction::Add: 792 case Instruction::Mul: 793 case Instruction::UDiv: 794 case Instruction::Sub: 795 case Instruction::And: 796 case Instruction::Or: 797 case Instruction::Xor: 798 case Instruction::Shl: 799 case Instruction::LShr: 800 case Instruction::AShr: 801 DEBUG(dbgs() << "Cloning IVUser: " << *DU.NarrowUse << "\n"); 802 803 // Replace NarrowDef operands with WideDef. Otherwise, we don't know 804 // anything about the narrow operand yet so must insert a [sz]ext. It is 805 // probably loop invariant and will be folded or hoisted. If it actually 806 // comes from a widened IV, it should be removed during a future call to 807 // WidenIVUse. 808 Value *LHS = (DU.NarrowUse->getOperand(0) == DU.NarrowDef) ? DU.WideDef : 809 getExtend(DU.NarrowUse->getOperand(0), WideType, IsSigned, DU.NarrowUse); 810 Value *RHS = (DU.NarrowUse->getOperand(1) == DU.NarrowDef) ? DU.WideDef : 811 getExtend(DU.NarrowUse->getOperand(1), WideType, IsSigned, DU.NarrowUse); 812 813 BinaryOperator *NarrowBO = cast<BinaryOperator>(DU.NarrowUse); 814 BinaryOperator *WideBO = BinaryOperator::Create(NarrowBO->getOpcode(), 815 LHS, RHS, 816 NarrowBO->getName()); 817 IRBuilder<> Builder(DU.NarrowUse); 818 Builder.Insert(WideBO); 819 if (const OverflowingBinaryOperator *OBO = 820 dyn_cast<OverflowingBinaryOperator>(NarrowBO)) { 821 if (OBO->hasNoUnsignedWrap()) WideBO->setHasNoUnsignedWrap(); 822 if (OBO->hasNoSignedWrap()) WideBO->setHasNoSignedWrap(); 823 } 824 return WideBO; 825 } 826} 827 828/// No-wrap operations can transfer sign extension of their result to their 829/// operands. Generate the SCEV value for the widened operation without 830/// actually modifying the IR yet. If the expression after extending the 831/// operands is an AddRec for this loop, return it. 832const SCEVAddRecExpr* WidenIV::GetExtendedOperandRecurrence(NarrowIVDefUse DU) { 833 // Handle the common case of add<nsw/nuw> 834 if (DU.NarrowUse->getOpcode() != Instruction::Add) 835 return 0; 836 837 // One operand (NarrowDef) has already been extended to WideDef. Now determine 838 // if extending the other will lead to a recurrence. 839 unsigned ExtendOperIdx = DU.NarrowUse->getOperand(0) == DU.NarrowDef ? 1 : 0; 840 assert(DU.NarrowUse->getOperand(1-ExtendOperIdx) == DU.NarrowDef && "bad DU"); 841 842 const SCEV *ExtendOperExpr = 0; 843 const OverflowingBinaryOperator *OBO = 844 cast<OverflowingBinaryOperator>(DU.NarrowUse); 845 if (IsSigned && OBO->hasNoSignedWrap()) 846 ExtendOperExpr = SE->getSignExtendExpr( 847 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 848 else if(!IsSigned && OBO->hasNoUnsignedWrap()) 849 ExtendOperExpr = SE->getZeroExtendExpr( 850 SE->getSCEV(DU.NarrowUse->getOperand(ExtendOperIdx)), WideType); 851 else 852 return 0; 853 854 // When creating this AddExpr, don't apply the current operations NSW or NUW 855 // flags. This instruction may be guarded by control flow that the no-wrap 856 // behavior depends on. Non-control-equivalent instructions can be mapped to 857 // the same SCEV expression, and it would be incorrect to transfer NSW/NUW 858 // semantics to those operations. 859 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>( 860 SE->getAddExpr(SE->getSCEV(DU.WideDef), ExtendOperExpr)); 861 862 if (!AddRec || AddRec->getLoop() != L) 863 return 0; 864 return AddRec; 865} 866 867/// GetWideRecurrence - Is this instruction potentially interesting from 868/// IVUsers' perspective after widening it's type? In other words, can the 869/// extend be safely hoisted out of the loop with SCEV reducing the value to a 870/// recurrence on the same loop. If so, return the sign or zero extended 871/// recurrence. Otherwise return NULL. 872const SCEVAddRecExpr *WidenIV::GetWideRecurrence(Instruction *NarrowUse) { 873 if (!SE->isSCEVable(NarrowUse->getType())) 874 return 0; 875 876 const SCEV *NarrowExpr = SE->getSCEV(NarrowUse); 877 if (SE->getTypeSizeInBits(NarrowExpr->getType()) 878 >= SE->getTypeSizeInBits(WideType)) { 879 // NarrowUse implicitly widens its operand. e.g. a gep with a narrow 880 // index. So don't follow this use. 881 return 0; 882 } 883 884 const SCEV *WideExpr = IsSigned ? 885 SE->getSignExtendExpr(NarrowExpr, WideType) : 886 SE->getZeroExtendExpr(NarrowExpr, WideType); 887 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(WideExpr); 888 if (!AddRec || AddRec->getLoop() != L) 889 return 0; 890 return AddRec; 891} 892 893/// WidenIVUse - Determine whether an individual user of the narrow IV can be 894/// widened. If so, return the wide clone of the user. 895Instruction *WidenIV::WidenIVUse(NarrowIVDefUse DU, SCEVExpander &Rewriter) { 896 897 // Stop traversing the def-use chain at inner-loop phis or post-loop phis. 898 if (isa<PHINode>(DU.NarrowUse) && 899 LI->getLoopFor(DU.NarrowUse->getParent()) != L) 900 return 0; 901 902 // Our raison d'etre! Eliminate sign and zero extension. 903 if (IsSigned ? isa<SExtInst>(DU.NarrowUse) : isa<ZExtInst>(DU.NarrowUse)) { 904 Value *NewDef = DU.WideDef; 905 if (DU.NarrowUse->getType() != WideType) { 906 unsigned CastWidth = SE->getTypeSizeInBits(DU.NarrowUse->getType()); 907 unsigned IVWidth = SE->getTypeSizeInBits(WideType); 908 if (CastWidth < IVWidth) { 909 // The cast isn't as wide as the IV, so insert a Trunc. 910 IRBuilder<> Builder(DU.NarrowUse); 911 NewDef = Builder.CreateTrunc(DU.WideDef, DU.NarrowUse->getType()); 912 } 913 else { 914 // A wider extend was hidden behind a narrower one. This may induce 915 // another round of IV widening in which the intermediate IV becomes 916 // dead. It should be very rare. 917 DEBUG(dbgs() << "INDVARS: New IV " << *WidePhi 918 << " not wide enough to subsume " << *DU.NarrowUse << "\n"); 919 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, DU.WideDef); 920 NewDef = DU.NarrowUse; 921 } 922 } 923 if (NewDef != DU.NarrowUse) { 924 DEBUG(dbgs() << "INDVARS: eliminating " << *DU.NarrowUse 925 << " replaced by " << *DU.WideDef << "\n"); 926 ++NumElimExt; 927 DU.NarrowUse->replaceAllUsesWith(NewDef); 928 DeadInsts.push_back(DU.NarrowUse); 929 } 930 // Now that the extend is gone, we want to expose it's uses for potential 931 // further simplification. We don't need to directly inform SimplifyIVUsers 932 // of the new users, because their parent IV will be processed later as a 933 // new loop phi. If we preserved IVUsers analysis, we would also want to 934 // push the uses of WideDef here. 935 936 // No further widening is needed. The deceased [sz]ext had done it for us. 937 return 0; 938 } 939 940 // Does this user itself evaluate to a recurrence after widening? 941 const SCEVAddRecExpr *WideAddRec = GetWideRecurrence(DU.NarrowUse); 942 if (!WideAddRec) { 943 WideAddRec = GetExtendedOperandRecurrence(DU); 944 } 945 if (!WideAddRec) { 946 // This user does not evaluate to a recurence after widening, so don't 947 // follow it. Instead insert a Trunc to kill off the original use, 948 // eventually isolating the original narrow IV so it can be removed. 949 IRBuilder<> Builder(getInsertPointForUses(DU.NarrowUse, DU.NarrowDef, DT)); 950 Value *Trunc = Builder.CreateTrunc(DU.WideDef, DU.NarrowDef->getType()); 951 DU.NarrowUse->replaceUsesOfWith(DU.NarrowDef, Trunc); 952 return 0; 953 } 954 // Assume block terminators cannot evaluate to a recurrence. We can't to 955 // insert a Trunc after a terminator if there happens to be a critical edge. 956 assert(DU.NarrowUse != DU.NarrowUse->getParent()->getTerminator() && 957 "SCEV is not expected to evaluate a block terminator"); 958 959 // Reuse the IV increment that SCEVExpander created as long as it dominates 960 // NarrowUse. 961 Instruction *WideUse = 0; 962 if (WideAddRec == WideIncExpr 963 && Rewriter.hoistIVInc(WideInc, DU.NarrowUse)) 964 WideUse = WideInc; 965 else { 966 WideUse = CloneIVUser(DU); 967 if (!WideUse) 968 return 0; 969 } 970 // Evaluation of WideAddRec ensured that the narrow expression could be 971 // extended outside the loop without overflow. This suggests that the wide use 972 // evaluates to the same expression as the extended narrow use, but doesn't 973 // absolutely guarantee it. Hence the following failsafe check. In rare cases 974 // where it fails, we simply throw away the newly created wide use. 975 if (WideAddRec != SE->getSCEV(WideUse)) { 976 DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse 977 << ": " << *SE->getSCEV(WideUse) << " != " << *WideAddRec << "\n"); 978 DeadInsts.push_back(WideUse); 979 return 0; 980 } 981 982 // Returning WideUse pushes it on the worklist. 983 return WideUse; 984} 985 986/// pushNarrowIVUsers - Add eligible users of NarrowDef to NarrowIVUsers. 987/// 988void WidenIV::pushNarrowIVUsers(Instruction *NarrowDef, Instruction *WideDef) { 989 for (Value::use_iterator UI = NarrowDef->use_begin(), 990 UE = NarrowDef->use_end(); UI != UE; ++UI) { 991 Instruction *NarrowUse = cast<Instruction>(*UI); 992 993 // Handle data flow merges and bizarre phi cycles. 994 if (!Widened.insert(NarrowUse)) 995 continue; 996 997 NarrowIVUsers.push_back(NarrowIVDefUse(NarrowDef, NarrowUse, WideDef)); 998 } 999} 1000 1001/// CreateWideIV - Process a single induction variable. First use the 1002/// SCEVExpander to create a wide induction variable that evaluates to the same 1003/// recurrence as the original narrow IV. Then use a worklist to forward 1004/// traverse the narrow IV's def-use chain. After WidenIVUse has processed all 1005/// interesting IV users, the narrow IV will be isolated for removal by 1006/// DeleteDeadPHIs. 1007/// 1008/// It would be simpler to delete uses as they are processed, but we must avoid 1009/// invalidating SCEV expressions. 1010/// 1011PHINode *WidenIV::CreateWideIV(SCEVExpander &Rewriter) { 1012 // Is this phi an induction variable? 1013 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(OrigPhi)); 1014 if (!AddRec) 1015 return NULL; 1016 1017 // Widen the induction variable expression. 1018 const SCEV *WideIVExpr = IsSigned ? 1019 SE->getSignExtendExpr(AddRec, WideType) : 1020 SE->getZeroExtendExpr(AddRec, WideType); 1021 1022 assert(SE->getEffectiveSCEVType(WideIVExpr->getType()) == WideType && 1023 "Expect the new IV expression to preserve its type"); 1024 1025 // Can the IV be extended outside the loop without overflow? 1026 AddRec = dyn_cast<SCEVAddRecExpr>(WideIVExpr); 1027 if (!AddRec || AddRec->getLoop() != L) 1028 return NULL; 1029 1030 // An AddRec must have loop-invariant operands. Since this AddRec is 1031 // materialized by a loop header phi, the expression cannot have any post-loop 1032 // operands, so they must dominate the loop header. 1033 assert(SE->properlyDominates(AddRec->getStart(), L->getHeader()) && 1034 SE->properlyDominates(AddRec->getStepRecurrence(*SE), L->getHeader()) 1035 && "Loop header phi recurrence inputs do not dominate the loop"); 1036 1037 // The rewriter provides a value for the desired IV expression. This may 1038 // either find an existing phi or materialize a new one. Either way, we 1039 // expect a well-formed cyclic phi-with-increments. i.e. any operand not part 1040 // of the phi-SCC dominates the loop entry. 1041 Instruction *InsertPt = L->getHeader()->begin(); 1042 WidePhi = cast<PHINode>(Rewriter.expandCodeFor(AddRec, WideType, InsertPt)); 1043 1044 // Remembering the WideIV increment generated by SCEVExpander allows 1045 // WidenIVUse to reuse it when widening the narrow IV's increment. We don't 1046 // employ a general reuse mechanism because the call above is the only call to 1047 // SCEVExpander. Henceforth, we produce 1-to-1 narrow to wide uses. 1048 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1049 WideInc = 1050 cast<Instruction>(WidePhi->getIncomingValueForBlock(LatchBlock)); 1051 WideIncExpr = SE->getSCEV(WideInc); 1052 } 1053 1054 DEBUG(dbgs() << "Wide IV: " << *WidePhi << "\n"); 1055 ++NumWidened; 1056 1057 // Traverse the def-use chain using a worklist starting at the original IV. 1058 assert(Widened.empty() && NarrowIVUsers.empty() && "expect initial state" ); 1059 1060 Widened.insert(OrigPhi); 1061 pushNarrowIVUsers(OrigPhi, WidePhi); 1062 1063 while (!NarrowIVUsers.empty()) { 1064 NarrowIVDefUse DU = NarrowIVUsers.pop_back_val(); 1065 1066 // Process a def-use edge. This may replace the use, so don't hold a 1067 // use_iterator across it. 1068 Instruction *WideUse = WidenIVUse(DU, Rewriter); 1069 1070 // Follow all def-use edges from the previous narrow use. 1071 if (WideUse) 1072 pushNarrowIVUsers(DU.NarrowUse, WideUse); 1073 1074 // WidenIVUse may have removed the def-use edge. 1075 if (DU.NarrowDef->use_empty()) 1076 DeadInsts.push_back(DU.NarrowDef); 1077 } 1078 return WidePhi; 1079} 1080 1081//===----------------------------------------------------------------------===// 1082// Simplification of IV users based on SCEV evaluation. 1083//===----------------------------------------------------------------------===// 1084 1085 1086/// SimplifyAndExtend - Iteratively perform simplification on a worklist of IV 1087/// users. Each successive simplification may push more users which may 1088/// themselves be candidates for simplification. 1089/// 1090/// Sign/Zero extend elimination is interleaved with IV simplification. 1091/// 1092void IndVarSimplify::SimplifyAndExtend(Loop *L, 1093 SCEVExpander &Rewriter, 1094 LPPassManager &LPM) { 1095 SmallVector<WideIVInfo, 8> WideIVs; 1096 1097 SmallVector<PHINode*, 8> LoopPhis; 1098 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1099 LoopPhis.push_back(cast<PHINode>(I)); 1100 } 1101 // Each round of simplification iterates through the SimplifyIVUsers worklist 1102 // for all current phis, then determines whether any IVs can be 1103 // widened. Widening adds new phis to LoopPhis, inducing another round of 1104 // simplification on the wide IVs. 1105 while (!LoopPhis.empty()) { 1106 // Evaluate as many IV expressions as possible before widening any IVs. This 1107 // forces SCEV to set no-wrap flags before evaluating sign/zero 1108 // extension. The first time SCEV attempts to normalize sign/zero extension, 1109 // the result becomes final. So for the most predictable results, we delay 1110 // evaluation of sign/zero extend evaluation until needed, and avoid running 1111 // other SCEV based analysis prior to SimplifyAndExtend. 1112 do { 1113 PHINode *CurrIV = LoopPhis.pop_back_val(); 1114 1115 // Information about sign/zero extensions of CurrIV. 1116 WideIVVisitor WIV(CurrIV, SE, TD); 1117 1118 Changed |= simplifyUsersOfIV(CurrIV, SE, &LPM, DeadInsts, &WIV); 1119 1120 if (WIV.WI.WidestNativeType) { 1121 WideIVs.push_back(WIV.WI); 1122 } 1123 } while(!LoopPhis.empty()); 1124 1125 for (; !WideIVs.empty(); WideIVs.pop_back()) { 1126 WidenIV Widener(WideIVs.back(), LI, SE, DT, DeadInsts); 1127 if (PHINode *WidePhi = Widener.CreateWideIV(Rewriter)) { 1128 Changed = true; 1129 LoopPhis.push_back(WidePhi); 1130 } 1131 } 1132 } 1133} 1134 1135//===----------------------------------------------------------------------===// 1136// LinearFunctionTestReplace and its kin. Rewrite the loop exit condition. 1137//===----------------------------------------------------------------------===// 1138 1139/// Check for expressions that ScalarEvolution generates to compute 1140/// BackedgeTakenInfo. If these expressions have not been reduced, then 1141/// expanding them may incur additional cost (albeit in the loop preheader). 1142static bool isHighCostExpansion(const SCEV *S, BranchInst *BI, 1143 SmallPtrSet<const SCEV*, 8> &Processed, 1144 ScalarEvolution *SE) { 1145 if (!Processed.insert(S)) 1146 return false; 1147 1148 // If the backedge-taken count is a UDiv, it's very likely a UDiv that 1149 // ScalarEvolution's HowFarToZero or HowManyLessThans produced to compute a 1150 // precise expression, rather than a UDiv from the user's code. If we can't 1151 // find a UDiv in the code with some simple searching, assume the former and 1152 // forego rewriting the loop. 1153 if (isa<SCEVUDivExpr>(S)) { 1154 ICmpInst *OrigCond = dyn_cast<ICmpInst>(BI->getCondition()); 1155 if (!OrigCond) return true; 1156 const SCEV *R = SE->getSCEV(OrigCond->getOperand(1)); 1157 R = SE->getMinusSCEV(R, SE->getConstant(R->getType(), 1)); 1158 if (R != S) { 1159 const SCEV *L = SE->getSCEV(OrigCond->getOperand(0)); 1160 L = SE->getMinusSCEV(L, SE->getConstant(L->getType(), 1)); 1161 if (L != S) 1162 return true; 1163 } 1164 } 1165 1166 // Recurse past add expressions, which commonly occur in the 1167 // BackedgeTakenCount. They may already exist in program code, and if not, 1168 // they are not too expensive rematerialize. 1169 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 1170 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 1171 I != E; ++I) { 1172 if (isHighCostExpansion(*I, BI, Processed, SE)) 1173 return true; 1174 } 1175 return false; 1176 } 1177 1178 // HowManyLessThans uses a Max expression whenever the loop is not guarded by 1179 // the exit condition. 1180 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S)) 1181 return true; 1182 1183 // If we haven't recognized an expensive SCEV pattern, assume it's an 1184 // expression produced by program code. 1185 return false; 1186} 1187 1188/// canExpandBackedgeTakenCount - Return true if this loop's backedge taken 1189/// count expression can be safely and cheaply expanded into an instruction 1190/// sequence that can be used by LinearFunctionTestReplace. 1191/// 1192/// TODO: This fails for pointer-type loop counters with greater than one byte 1193/// strides, consequently preventing LFTR from running. For the purpose of LFTR 1194/// we could skip this check in the case that the LFTR loop counter (chosen by 1195/// FindLoopCounter) is also pointer type. Instead, we could directly convert 1196/// the loop test to an inequality test by checking the target data's alignment 1197/// of element types (given that the initial pointer value originates from or is 1198/// used by ABI constrained operation, as opposed to inttoptr/ptrtoint). 1199/// However, we don't yet have a strong motivation for converting loop tests 1200/// into inequality tests. 1201static bool canExpandBackedgeTakenCount(Loop *L, ScalarEvolution *SE) { 1202 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1203 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount) || 1204 BackedgeTakenCount->isZero()) 1205 return false; 1206 1207 if (!L->getExitingBlock()) 1208 return false; 1209 1210 // Can't rewrite non-branch yet. 1211 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1212 if (!BI) 1213 return false; 1214 1215 SmallPtrSet<const SCEV*, 8> Processed; 1216 if (isHighCostExpansion(BackedgeTakenCount, BI, Processed, SE)) 1217 return false; 1218 1219 return true; 1220} 1221 1222/// getLoopPhiForCounter - Return the loop header phi IFF IncV adds a loop 1223/// invariant value to the phi. 1224static PHINode *getLoopPhiForCounter(Value *IncV, Loop *L, DominatorTree *DT) { 1225 Instruction *IncI = dyn_cast<Instruction>(IncV); 1226 if (!IncI) 1227 return 0; 1228 1229 switch (IncI->getOpcode()) { 1230 case Instruction::Add: 1231 case Instruction::Sub: 1232 break; 1233 case Instruction::GetElementPtr: 1234 // An IV counter must preserve its type. 1235 if (IncI->getNumOperands() == 2) 1236 break; 1237 default: 1238 return 0; 1239 } 1240 1241 PHINode *Phi = dyn_cast<PHINode>(IncI->getOperand(0)); 1242 if (Phi && Phi->getParent() == L->getHeader()) { 1243 if (isLoopInvariant(IncI->getOperand(1), L, DT)) 1244 return Phi; 1245 return 0; 1246 } 1247 if (IncI->getOpcode() == Instruction::GetElementPtr) 1248 return 0; 1249 1250 // Allow add/sub to be commuted. 1251 Phi = dyn_cast<PHINode>(IncI->getOperand(1)); 1252 if (Phi && Phi->getParent() == L->getHeader()) { 1253 if (isLoopInvariant(IncI->getOperand(0), L, DT)) 1254 return Phi; 1255 } 1256 return 0; 1257} 1258 1259/// Return the compare guarding the loop latch, or NULL for unrecognized tests. 1260static ICmpInst *getLoopTest(Loop *L) { 1261 assert(L->getExitingBlock() && "expected loop exit"); 1262 1263 BasicBlock *LatchBlock = L->getLoopLatch(); 1264 // Don't bother with LFTR if the loop is not properly simplified. 1265 if (!LatchBlock) 1266 return 0; 1267 1268 BranchInst *BI = dyn_cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1269 assert(BI && "expected exit branch"); 1270 1271 return dyn_cast<ICmpInst>(BI->getCondition()); 1272} 1273 1274/// needsLFTR - LinearFunctionTestReplace policy. Return true unless we can show 1275/// that the current exit test is already sufficiently canonical. 1276static bool needsLFTR(Loop *L, DominatorTree *DT) { 1277 // Do LFTR to simplify the exit condition to an ICMP. 1278 ICmpInst *Cond = getLoopTest(L); 1279 if (!Cond) 1280 return true; 1281 1282 // Do LFTR to simplify the exit ICMP to EQ/NE 1283 ICmpInst::Predicate Pred = Cond->getPredicate(); 1284 if (Pred != ICmpInst::ICMP_NE && Pred != ICmpInst::ICMP_EQ) 1285 return true; 1286 1287 // Look for a loop invariant RHS 1288 Value *LHS = Cond->getOperand(0); 1289 Value *RHS = Cond->getOperand(1); 1290 if (!isLoopInvariant(RHS, L, DT)) { 1291 if (!isLoopInvariant(LHS, L, DT)) 1292 return true; 1293 std::swap(LHS, RHS); 1294 } 1295 // Look for a simple IV counter LHS 1296 PHINode *Phi = dyn_cast<PHINode>(LHS); 1297 if (!Phi) 1298 Phi = getLoopPhiForCounter(LHS, L, DT); 1299 1300 if (!Phi) 1301 return true; 1302 1303 // Do LFTR if PHI node is defined in the loop, but is *not* a counter. 1304 int Idx = Phi->getBasicBlockIndex(L->getLoopLatch()); 1305 if (Idx < 0) 1306 return true; 1307 1308 // Do LFTR if the exit condition's IV is *not* a simple counter. 1309 Value *IncV = Phi->getIncomingValue(Idx); 1310 return Phi != getLoopPhiForCounter(IncV, L, DT); 1311} 1312 1313/// Recursive helper for hasConcreteDef(). Unfortunately, this currently boils 1314/// down to checking that all operands are constant and listing instructions 1315/// that may hide undef. 1316static bool hasConcreteDefImpl(Value *V, SmallPtrSet<Value*, 8> &Visited, 1317 unsigned Depth) { 1318 if (isa<Constant>(V)) 1319 return !isa<UndefValue>(V); 1320 1321 if (Depth >= 6) 1322 return false; 1323 1324 // Conservatively handle non-constant non-instructions. For example, Arguments 1325 // may be undef. 1326 Instruction *I = dyn_cast<Instruction>(V); 1327 if (!I) 1328 return false; 1329 1330 // Load and return values may be undef. 1331 if(I->mayReadFromMemory() || isa<CallInst>(I) || isa<InvokeInst>(I)) 1332 return false; 1333 1334 // Optimistically handle other instructions. 1335 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) { 1336 if (!Visited.insert(*OI)) 1337 continue; 1338 if (!hasConcreteDefImpl(*OI, Visited, Depth+1)) 1339 return false; 1340 } 1341 return true; 1342} 1343 1344/// Return true if the given value is concrete. We must prove that undef can 1345/// never reach it. 1346/// 1347/// TODO: If we decide that this is a good approach to checking for undef, we 1348/// may factor it into a common location. 1349static bool hasConcreteDef(Value *V) { 1350 SmallPtrSet<Value*, 8> Visited; 1351 Visited.insert(V); 1352 return hasConcreteDefImpl(V, Visited, 0); 1353} 1354 1355/// AlmostDeadIV - Return true if this IV has any uses other than the (soon to 1356/// be rewritten) loop exit test. 1357static bool AlmostDeadIV(PHINode *Phi, BasicBlock *LatchBlock, Value *Cond) { 1358 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1359 Value *IncV = Phi->getIncomingValue(LatchIdx); 1360 1361 for (Value::use_iterator UI = Phi->use_begin(), UE = Phi->use_end(); 1362 UI != UE; ++UI) { 1363 if (*UI != Cond && *UI != IncV) return false; 1364 } 1365 1366 for (Value::use_iterator UI = IncV->use_begin(), UE = IncV->use_end(); 1367 UI != UE; ++UI) { 1368 if (*UI != Cond && *UI != Phi) return false; 1369 } 1370 return true; 1371} 1372 1373/// FindLoopCounter - Find an affine IV in canonical form. 1374/// 1375/// BECount may be an i8* pointer type. The pointer difference is already 1376/// valid count without scaling the address stride, so it remains a pointer 1377/// expression as far as SCEV is concerned. 1378/// 1379/// Currently only valid for LFTR. See the comments on hasConcreteDef below. 1380/// 1381/// FIXME: Accept -1 stride and set IVLimit = IVInit - BECount 1382/// 1383/// FIXME: Accept non-unit stride as long as SCEV can reduce BECount * Stride. 1384/// This is difficult in general for SCEV because of potential overflow. But we 1385/// could at least handle constant BECounts. 1386static PHINode * 1387FindLoopCounter(Loop *L, const SCEV *BECount, 1388 ScalarEvolution *SE, DominatorTree *DT, const DataLayout *TD) { 1389 uint64_t BCWidth = SE->getTypeSizeInBits(BECount->getType()); 1390 1391 Value *Cond = 1392 cast<BranchInst>(L->getExitingBlock()->getTerminator())->getCondition(); 1393 1394 // Loop over all of the PHI nodes, looking for a simple counter. 1395 PHINode *BestPhi = 0; 1396 const SCEV *BestInit = 0; 1397 BasicBlock *LatchBlock = L->getLoopLatch(); 1398 assert(LatchBlock && "needsLFTR should guarantee a loop latch"); 1399 1400 for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I) { 1401 PHINode *Phi = cast<PHINode>(I); 1402 if (!SE->isSCEVable(Phi->getType())) 1403 continue; 1404 1405 // Avoid comparing an integer IV against a pointer Limit. 1406 if (BECount->getType()->isPointerTy() && !Phi->getType()->isPointerTy()) 1407 continue; 1408 1409 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Phi)); 1410 if (!AR || AR->getLoop() != L || !AR->isAffine()) 1411 continue; 1412 1413 // AR may be a pointer type, while BECount is an integer type. 1414 // AR may be wider than BECount. With eq/ne tests overflow is immaterial. 1415 // AR may not be a narrower type, or we may never exit. 1416 uint64_t PhiWidth = SE->getTypeSizeInBits(AR->getType()); 1417 if (PhiWidth < BCWidth || (TD && !TD->isLegalInteger(PhiWidth))) 1418 continue; 1419 1420 const SCEV *Step = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE)); 1421 if (!Step || !Step->isOne()) 1422 continue; 1423 1424 int LatchIdx = Phi->getBasicBlockIndex(LatchBlock); 1425 Value *IncV = Phi->getIncomingValue(LatchIdx); 1426 if (getLoopPhiForCounter(IncV, L, DT) != Phi) 1427 continue; 1428 1429 // Avoid reusing a potentially undef value to compute other values that may 1430 // have originally had a concrete definition. 1431 if (!hasConcreteDef(Phi)) { 1432 // We explicitly allow unknown phis as long as they are already used by 1433 // the loop test. In this case we assume that performing LFTR could not 1434 // increase the number of undef users. 1435 if (ICmpInst *Cond = getLoopTest(L)) { 1436 if (Phi != getLoopPhiForCounter(Cond->getOperand(0), L, DT) 1437 && Phi != getLoopPhiForCounter(Cond->getOperand(1), L, DT)) { 1438 continue; 1439 } 1440 } 1441 } 1442 const SCEV *Init = AR->getStart(); 1443 1444 if (BestPhi && !AlmostDeadIV(BestPhi, LatchBlock, Cond)) { 1445 // Don't force a live loop counter if another IV can be used. 1446 if (AlmostDeadIV(Phi, LatchBlock, Cond)) 1447 continue; 1448 1449 // Prefer to count-from-zero. This is a more "canonical" counter form. It 1450 // also prefers integer to pointer IVs. 1451 if (BestInit->isZero() != Init->isZero()) { 1452 if (BestInit->isZero()) 1453 continue; 1454 } 1455 // If two IVs both count from zero or both count from nonzero then the 1456 // narrower is likely a dead phi that has been widened. Use the wider phi 1457 // to allow the other to be eliminated. 1458 else if (PhiWidth <= SE->getTypeSizeInBits(BestPhi->getType())) 1459 continue; 1460 } 1461 BestPhi = Phi; 1462 BestInit = Init; 1463 } 1464 return BestPhi; 1465} 1466 1467/// genLoopLimit - Help LinearFunctionTestReplace by generating a value that 1468/// holds the RHS of the new loop test. 1469static Value *genLoopLimit(PHINode *IndVar, const SCEV *IVCount, Loop *L, 1470 SCEVExpander &Rewriter, ScalarEvolution *SE) { 1471 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(IndVar)); 1472 assert(AR && AR->getLoop() == L && AR->isAffine() && "bad loop counter"); 1473 const SCEV *IVInit = AR->getStart(); 1474 1475 // IVInit may be a pointer while IVCount is an integer when FindLoopCounter 1476 // finds a valid pointer IV. Sign extend BECount in order to materialize a 1477 // GEP. Avoid running SCEVExpander on a new pointer value, instead reusing 1478 // the existing GEPs whenever possible. 1479 if (IndVar->getType()->isPointerTy() 1480 && !IVCount->getType()->isPointerTy()) { 1481 1482 Type *OfsTy = SE->getEffectiveSCEVType(IVInit->getType()); 1483 const SCEV *IVOffset = SE->getTruncateOrSignExtend(IVCount, OfsTy); 1484 1485 // Expand the code for the iteration count. 1486 assert(SE->isLoopInvariant(IVOffset, L) && 1487 "Computed iteration count is not loop invariant!"); 1488 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1489 Value *GEPOffset = Rewriter.expandCodeFor(IVOffset, OfsTy, BI); 1490 1491 Value *GEPBase = IndVar->getIncomingValueForBlock(L->getLoopPreheader()); 1492 assert(AR->getStart() == SE->getSCEV(GEPBase) && "bad loop counter"); 1493 // We could handle pointer IVs other than i8*, but we need to compensate for 1494 // gep index scaling. See canExpandBackedgeTakenCount comments. 1495 assert(SE->getSizeOfExpr( 1496 cast<PointerType>(GEPBase->getType())->getElementType())->isOne() 1497 && "unit stride pointer IV must be i8*"); 1498 1499 IRBuilder<> Builder(L->getLoopPreheader()->getTerminator()); 1500 return Builder.CreateGEP(GEPBase, GEPOffset, "lftr.limit"); 1501 } 1502 else { 1503 // In any other case, convert both IVInit and IVCount to integers before 1504 // comparing. This may result in SCEV expension of pointers, but in practice 1505 // SCEV will fold the pointer arithmetic away as such: 1506 // BECount = (IVEnd - IVInit - 1) => IVLimit = IVInit (postinc). 1507 // 1508 // Valid Cases: (1) both integers is most common; (2) both may be pointers 1509 // for simple memset-style loops; (3) IVInit is an integer and IVCount is a 1510 // pointer may occur when enable-iv-rewrite generates a canonical IV on top 1511 // of case #2. 1512 1513 const SCEV *IVLimit = 0; 1514 // For unit stride, IVCount = Start + BECount with 2's complement overflow. 1515 // For non-zero Start, compute IVCount here. 1516 if (AR->getStart()->isZero()) 1517 IVLimit = IVCount; 1518 else { 1519 assert(AR->getStepRecurrence(*SE)->isOne() && "only handles unit stride"); 1520 const SCEV *IVInit = AR->getStart(); 1521 1522 // For integer IVs, truncate the IV before computing IVInit + BECount. 1523 if (SE->getTypeSizeInBits(IVInit->getType()) 1524 > SE->getTypeSizeInBits(IVCount->getType())) 1525 IVInit = SE->getTruncateExpr(IVInit, IVCount->getType()); 1526 1527 IVLimit = SE->getAddExpr(IVInit, IVCount); 1528 } 1529 // Expand the code for the iteration count. 1530 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1531 IRBuilder<> Builder(BI); 1532 assert(SE->isLoopInvariant(IVLimit, L) && 1533 "Computed iteration count is not loop invariant!"); 1534 // Ensure that we generate the same type as IndVar, or a smaller integer 1535 // type. In the presence of null pointer values, we have an integer type 1536 // SCEV expression (IVInit) for a pointer type IV value (IndVar). 1537 Type *LimitTy = IVCount->getType()->isPointerTy() ? 1538 IndVar->getType() : IVCount->getType(); 1539 return Rewriter.expandCodeFor(IVLimit, LimitTy, BI); 1540 } 1541} 1542 1543/// LinearFunctionTestReplace - This method rewrites the exit condition of the 1544/// loop to be a canonical != comparison against the incremented loop induction 1545/// variable. This pass is able to rewrite the exit tests of any loop where the 1546/// SCEV analysis can determine a loop-invariant trip count of the loop, which 1547/// is actually a much broader range than just linear tests. 1548Value *IndVarSimplify:: 1549LinearFunctionTestReplace(Loop *L, 1550 const SCEV *BackedgeTakenCount, 1551 PHINode *IndVar, 1552 SCEVExpander &Rewriter) { 1553 assert(canExpandBackedgeTakenCount(L, SE) && "precondition"); 1554 1555 // LFTR can ignore IV overflow and truncate to the width of 1556 // BECount. This avoids materializing the add(zext(add)) expression. 1557 Type *CntTy = BackedgeTakenCount->getType(); 1558 1559 // If the exiting block is the same as the backedge block, we prefer to 1560 // compare against the post-incremented value, otherwise we must compare 1561 // against the preincremented value. 1562 Value *CmpIndVar; 1563 const SCEV *IVCount; 1564 if (L->getExitingBlock() == L->getLoopLatch()) { 1565 // Add one to the "backedge-taken" count to get the trip count. 1566 // If this addition may overflow, we have to be more pessimistic and 1567 // cast the induction variable before doing the add. 1568 const SCEV *N = 1569 SE->getAddExpr(BackedgeTakenCount, 1570 SE->getConstant(BackedgeTakenCount->getType(), 1)); 1571 if (CntTy == BackedgeTakenCount->getType()) 1572 IVCount = N; 1573 else { 1574 const SCEV *Zero = SE->getConstant(BackedgeTakenCount->getType(), 0); 1575 if ((isa<SCEVConstant>(N) && !N->isZero()) || 1576 SE->isLoopEntryGuardedByCond(L, ICmpInst::ICMP_NE, N, Zero)) { 1577 // No overflow. Cast the sum. 1578 IVCount = SE->getTruncateOrZeroExtend(N, CntTy); 1579 } else { 1580 // Potential overflow. Cast before doing the add. 1581 IVCount = SE->getTruncateOrZeroExtend(BackedgeTakenCount, CntTy); 1582 IVCount = SE->getAddExpr(IVCount, SE->getConstant(CntTy, 1)); 1583 } 1584 } 1585 // The BackedgeTaken expression contains the number of times that the 1586 // backedge branches to the loop header. This is one less than the 1587 // number of times the loop executes, so use the incremented indvar. 1588 CmpIndVar = IndVar->getIncomingValueForBlock(L->getExitingBlock()); 1589 } else { 1590 // We must use the preincremented value... 1591 IVCount = SE->getTruncateOrZeroExtend(BackedgeTakenCount, CntTy); 1592 CmpIndVar = IndVar; 1593 } 1594 1595 Value *ExitCnt = genLoopLimit(IndVar, IVCount, L, Rewriter, SE); 1596 assert(ExitCnt->getType()->isPointerTy() == IndVar->getType()->isPointerTy() 1597 && "genLoopLimit missed a cast"); 1598 1599 // Insert a new icmp_ne or icmp_eq instruction before the branch. 1600 BranchInst *BI = cast<BranchInst>(L->getExitingBlock()->getTerminator()); 1601 ICmpInst::Predicate P; 1602 if (L->contains(BI->getSuccessor(0))) 1603 P = ICmpInst::ICMP_NE; 1604 else 1605 P = ICmpInst::ICMP_EQ; 1606 1607 DEBUG(dbgs() << "INDVARS: Rewriting loop exit condition to:\n" 1608 << " LHS:" << *CmpIndVar << '\n' 1609 << " op:\t" 1610 << (P == ICmpInst::ICMP_NE ? "!=" : "==") << "\n" 1611 << " RHS:\t" << *ExitCnt << "\n" 1612 << " IVCount:\t" << *IVCount << "\n"); 1613 1614 IRBuilder<> Builder(BI); 1615 if (SE->getTypeSizeInBits(CmpIndVar->getType()) 1616 > SE->getTypeSizeInBits(ExitCnt->getType())) { 1617 CmpIndVar = Builder.CreateTrunc(CmpIndVar, ExitCnt->getType(), 1618 "lftr.wideiv"); 1619 } 1620 1621 Value *Cond = Builder.CreateICmp(P, CmpIndVar, ExitCnt, "exitcond"); 1622 Value *OrigCond = BI->getCondition(); 1623 // It's tempting to use replaceAllUsesWith here to fully replace the old 1624 // comparison, but that's not immediately safe, since users of the old 1625 // comparison may not be dominated by the new comparison. Instead, just 1626 // update the branch to use the new comparison; in the common case this 1627 // will make old comparison dead. 1628 BI->setCondition(Cond); 1629 DeadInsts.push_back(OrigCond); 1630 1631 ++NumLFTR; 1632 Changed = true; 1633 return Cond; 1634} 1635 1636//===----------------------------------------------------------------------===// 1637// SinkUnusedInvariants. A late subpass to cleanup loop preheaders. 1638//===----------------------------------------------------------------------===// 1639 1640/// If there's a single exit block, sink any loop-invariant values that 1641/// were defined in the preheader but not used inside the loop into the 1642/// exit block to reduce register pressure in the loop. 1643void IndVarSimplify::SinkUnusedInvariants(Loop *L) { 1644 BasicBlock *ExitBlock = L->getExitBlock(); 1645 if (!ExitBlock) return; 1646 1647 BasicBlock *Preheader = L->getLoopPreheader(); 1648 if (!Preheader) return; 1649 1650 Instruction *InsertPt = ExitBlock->getFirstInsertionPt(); 1651 BasicBlock::iterator I = Preheader->getTerminator(); 1652 while (I != Preheader->begin()) { 1653 --I; 1654 // New instructions were inserted at the end of the preheader. 1655 if (isa<PHINode>(I)) 1656 break; 1657 1658 // Don't move instructions which might have side effects, since the side 1659 // effects need to complete before instructions inside the loop. Also don't 1660 // move instructions which might read memory, since the loop may modify 1661 // memory. Note that it's okay if the instruction might have undefined 1662 // behavior: LoopSimplify guarantees that the preheader dominates the exit 1663 // block. 1664 if (I->mayHaveSideEffects() || I->mayReadFromMemory()) 1665 continue; 1666 1667 // Skip debug info intrinsics. 1668 if (isa<DbgInfoIntrinsic>(I)) 1669 continue; 1670 1671 // Skip landingpad instructions. 1672 if (isa<LandingPadInst>(I)) 1673 continue; 1674 1675 // Don't sink alloca: we never want to sink static alloca's out of the 1676 // entry block, and correctly sinking dynamic alloca's requires 1677 // checks for stacksave/stackrestore intrinsics. 1678 // FIXME: Refactor this check somehow? 1679 if (isa<AllocaInst>(I)) 1680 continue; 1681 1682 // Determine if there is a use in or before the loop (direct or 1683 // otherwise). 1684 bool UsedInLoop = false; 1685 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); 1686 UI != UE; ++UI) { 1687 User *U = *UI; 1688 BasicBlock *UseBB = cast<Instruction>(U)->getParent(); 1689 if (PHINode *P = dyn_cast<PHINode>(U)) { 1690 unsigned i = 1691 PHINode::getIncomingValueNumForOperand(UI.getOperandNo()); 1692 UseBB = P->getIncomingBlock(i); 1693 } 1694 if (UseBB == Preheader || L->contains(UseBB)) { 1695 UsedInLoop = true; 1696 break; 1697 } 1698 } 1699 1700 // If there is, the def must remain in the preheader. 1701 if (UsedInLoop) 1702 continue; 1703 1704 // Otherwise, sink it to the exit block. 1705 Instruction *ToMove = I; 1706 bool Done = false; 1707 1708 if (I != Preheader->begin()) { 1709 // Skip debug info intrinsics. 1710 do { 1711 --I; 1712 } while (isa<DbgInfoIntrinsic>(I) && I != Preheader->begin()); 1713 1714 if (isa<DbgInfoIntrinsic>(I) && I == Preheader->begin()) 1715 Done = true; 1716 } else { 1717 Done = true; 1718 } 1719 1720 ToMove->moveBefore(InsertPt); 1721 if (Done) break; 1722 InsertPt = ToMove; 1723 } 1724} 1725 1726//===----------------------------------------------------------------------===// 1727// IndVarSimplify driver. Manage several subpasses of IV simplification. 1728//===----------------------------------------------------------------------===// 1729 1730bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) { 1731 // If LoopSimplify form is not available, stay out of trouble. Some notes: 1732 // - LSR currently only supports LoopSimplify-form loops. Indvars' 1733 // canonicalization can be a pessimization without LSR to "clean up" 1734 // afterwards. 1735 // - We depend on having a preheader; in particular, 1736 // Loop::getCanonicalInductionVariable only supports loops with preheaders, 1737 // and we're in trouble if we can't find the induction variable even when 1738 // we've manually inserted one. 1739 if (!L->isLoopSimplifyForm()) 1740 return false; 1741 1742 LI = &getAnalysis<LoopInfo>(); 1743 SE = &getAnalysis<ScalarEvolution>(); 1744 DT = &getAnalysis<DominatorTree>(); 1745 TD = getAnalysisIfAvailable<DataLayout>(); 1746 TLI = getAnalysisIfAvailable<TargetLibraryInfo>(); 1747 1748 DeadInsts.clear(); 1749 Changed = false; 1750 1751 // If there are any floating-point recurrences, attempt to 1752 // transform them to use integer recurrences. 1753 RewriteNonIntegerIVs(L); 1754 1755 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 1756 1757 // Create a rewriter object which we'll use to transform the code with. 1758 SCEVExpander Rewriter(*SE, "indvars"); 1759#ifndef NDEBUG 1760 Rewriter.setDebugType(DEBUG_TYPE); 1761#endif 1762 1763 // Eliminate redundant IV users. 1764 // 1765 // Simplification works best when run before other consumers of SCEV. We 1766 // attempt to avoid evaluating SCEVs for sign/zero extend operations until 1767 // other expressions involving loop IVs have been evaluated. This helps SCEV 1768 // set no-wrap flags before normalizing sign/zero extension. 1769 Rewriter.disableCanonicalMode(); 1770 SimplifyAndExtend(L, Rewriter, LPM); 1771 1772 // Check to see if this loop has a computable loop-invariant execution count. 1773 // If so, this means that we can compute the final value of any expressions 1774 // that are recurrent in the loop, and substitute the exit values from the 1775 // loop into any instructions outside of the loop that use the final values of 1776 // the current expressions. 1777 // 1778 if (!isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1779 RewriteLoopExitValues(L, Rewriter); 1780 1781 // Eliminate redundant IV cycles. 1782 NumElimIV += Rewriter.replaceCongruentIVs(L, DT, DeadInsts); 1783 1784 // If we have a trip count expression, rewrite the loop's exit condition 1785 // using it. We can currently only handle loops with a single exit. 1786 if (canExpandBackedgeTakenCount(L, SE) && needsLFTR(L, DT)) { 1787 PHINode *IndVar = FindLoopCounter(L, BackedgeTakenCount, SE, DT, TD); 1788 if (IndVar) { 1789 // Check preconditions for proper SCEVExpander operation. SCEV does not 1790 // express SCEVExpander's dependencies, such as LoopSimplify. Instead any 1791 // pass that uses the SCEVExpander must do it. This does not work well for 1792 // loop passes because SCEVExpander makes assumptions about all loops, while 1793 // LoopPassManager only forces the current loop to be simplified. 1794 // 1795 // FIXME: SCEV expansion has no way to bail out, so the caller must 1796 // explicitly check any assumptions made by SCEV. Brittle. 1797 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(BackedgeTakenCount); 1798 if (!AR || AR->getLoop()->getLoopPreheader()) 1799 (void)LinearFunctionTestReplace(L, BackedgeTakenCount, IndVar, 1800 Rewriter); 1801 } 1802 } 1803 // Clear the rewriter cache, because values that are in the rewriter's cache 1804 // can be deleted in the loop below, causing the AssertingVH in the cache to 1805 // trigger. 1806 Rewriter.clear(); 1807 1808 // Now that we're done iterating through lists, clean up any instructions 1809 // which are now dead. 1810 while (!DeadInsts.empty()) 1811 if (Instruction *Inst = 1812 dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val())) 1813 RecursivelyDeleteTriviallyDeadInstructions(Inst, TLI); 1814 1815 // The Rewriter may not be used from this point on. 1816 1817 // Loop-invariant instructions in the preheader that aren't used in the 1818 // loop may be sunk below the loop to reduce register pressure. 1819 SinkUnusedInvariants(L); 1820 1821 // Clean up dead instructions. 1822 Changed |= DeleteDeadPHIs(L->getHeader(), TLI); 1823 // Check a post-condition. 1824 assert(L->isLCSSAForm(*DT) && 1825 "Indvars did not leave the loop in lcssa form!"); 1826 1827 // Verify that LFTR, and any other change have not interfered with SCEV's 1828 // ability to compute trip count. 1829#ifndef NDEBUG 1830 if (VerifyIndvars && !isa<SCEVCouldNotCompute>(BackedgeTakenCount)) { 1831 SE->forgetLoop(L); 1832 const SCEV *NewBECount = SE->getBackedgeTakenCount(L); 1833 if (SE->getTypeSizeInBits(BackedgeTakenCount->getType()) < 1834 SE->getTypeSizeInBits(NewBECount->getType())) 1835 NewBECount = SE->getTruncateOrNoop(NewBECount, 1836 BackedgeTakenCount->getType()); 1837 else 1838 BackedgeTakenCount = SE->getTruncateOrNoop(BackedgeTakenCount, 1839 NewBECount->getType()); 1840 assert(BackedgeTakenCount == NewBECount && "indvars must preserve SCEV"); 1841 } 1842#endif 1843 1844 return Changed; 1845} 1846