LoopStrengthReduce.cpp revision e50ed30282bb5b4a9ed952580523f2dda16215ac
1//===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===// 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 forms suitable for efficient execution 12// on the target. 13// 14// This pass performs a strength reduction on array references inside loops that 15// have as one or more of their components the loop induction variable, it 16// rewrites expressions to take advantage of scaled-index addressing modes 17// available on the target, and it performs a variety of other optimizations 18// related to loop induction variables. 19// 20//===----------------------------------------------------------------------===// 21 22#define DEBUG_TYPE "loop-reduce" 23#include "llvm/Transforms/Scalar.h" 24#include "llvm/Constants.h" 25#include "llvm/Instructions.h" 26#include "llvm/IntrinsicInst.h" 27#include "llvm/LLVMContext.h" 28#include "llvm/Type.h" 29#include "llvm/DerivedTypes.h" 30#include "llvm/Analysis/Dominators.h" 31#include "llvm/Analysis/IVUsers.h" 32#include "llvm/Analysis/LoopInfo.h" 33#include "llvm/Analysis/LoopPass.h" 34#include "llvm/Analysis/ScalarEvolutionExpander.h" 35#include "llvm/Transforms/Utils/AddrModeMatcher.h" 36#include "llvm/Transforms/Utils/BasicBlockUtils.h" 37#include "llvm/Transforms/Utils/Local.h" 38#include "llvm/ADT/Statistic.h" 39#include "llvm/Support/CFG.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/Compiler.h" 42#include "llvm/Support/CommandLine.h" 43#include "llvm/Support/ValueHandle.h" 44#include "llvm/Support/raw_ostream.h" 45#include "llvm/Target/TargetLowering.h" 46#include <algorithm> 47using namespace llvm; 48 49STATISTIC(NumReduced , "Number of IV uses strength reduced"); 50STATISTIC(NumInserted, "Number of PHIs inserted"); 51STATISTIC(NumVariable, "Number of PHIs with variable strides"); 52STATISTIC(NumEliminated, "Number of strides eliminated"); 53STATISTIC(NumShadow, "Number of Shadow IVs optimized"); 54STATISTIC(NumImmSunk, "Number of common expr immediates sunk into uses"); 55STATISTIC(NumLoopCond, "Number of loop terminating conds optimized"); 56 57static cl::opt<bool> EnableFullLSRMode("enable-full-lsr", 58 cl::init(false), 59 cl::Hidden); 60 61namespace { 62 63 struct BasedUser; 64 65 /// IVInfo - This structure keeps track of one IV expression inserted during 66 /// StrengthReduceStridedIVUsers. It contains the stride, the common base, as 67 /// well as the PHI node and increment value created for rewrite. 68 struct VISIBILITY_HIDDEN IVExpr { 69 const SCEV *Stride; 70 const SCEV *Base; 71 PHINode *PHI; 72 73 IVExpr(const SCEV *const stride, const SCEV *const base, PHINode *phi) 74 : Stride(stride), Base(base), PHI(phi) {} 75 }; 76 77 /// IVsOfOneStride - This structure keeps track of all IV expression inserted 78 /// during StrengthReduceStridedIVUsers for a particular stride of the IV. 79 struct VISIBILITY_HIDDEN IVsOfOneStride { 80 std::vector<IVExpr> IVs; 81 82 void addIV(const SCEV *const Stride, const SCEV *const Base, PHINode *PHI) { 83 IVs.push_back(IVExpr(Stride, Base, PHI)); 84 } 85 }; 86 87 class VISIBILITY_HIDDEN LoopStrengthReduce : public LoopPass { 88 IVUsers *IU; 89 LoopInfo *LI; 90 DominatorTree *DT; 91 ScalarEvolution *SE; 92 bool Changed; 93 94 /// IVsByStride - Keep track of all IVs that have been inserted for a 95 /// particular stride. 96 std::map<const SCEV *, IVsOfOneStride> IVsByStride; 97 98 /// StrideNoReuse - Keep track of all the strides whose ivs cannot be 99 /// reused (nor should they be rewritten to reuse other strides). 100 SmallSet<const SCEV *, 4> StrideNoReuse; 101 102 /// DeadInsts - Keep track of instructions we may have made dead, so that 103 /// we can remove them after we are done working. 104 SmallVector<WeakVH, 16> DeadInsts; 105 106 /// TLI - Keep a pointer of a TargetLowering to consult for determining 107 /// transformation profitability. 108 const TargetLowering *TLI; 109 110 public: 111 static char ID; // Pass ID, replacement for typeid 112 explicit LoopStrengthReduce(const TargetLowering *tli = NULL) : 113 LoopPass(&ID), TLI(tli) { 114 } 115 116 bool runOnLoop(Loop *L, LPPassManager &LPM); 117 118 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 119 // We split critical edges, so we change the CFG. However, we do update 120 // many analyses if they are around. 121 AU.addPreservedID(LoopSimplifyID); 122 AU.addPreserved<LoopInfo>(); 123 AU.addPreserved<DominanceFrontier>(); 124 AU.addPreserved<DominatorTree>(); 125 126 AU.addRequiredID(LoopSimplifyID); 127 AU.addRequired<LoopInfo>(); 128 AU.addRequired<DominatorTree>(); 129 AU.addRequired<ScalarEvolution>(); 130 AU.addPreserved<ScalarEvolution>(); 131 AU.addRequired<IVUsers>(); 132 AU.addPreserved<IVUsers>(); 133 } 134 135 private: 136 ICmpInst *ChangeCompareStride(Loop *L, ICmpInst *Cond, 137 IVStrideUse* &CondUse, 138 const SCEV *const * &CondStride); 139 140 void OptimizeIndvars(Loop *L); 141 void OptimizeLoopCountIV(Loop *L); 142 void OptimizeLoopTermCond(Loop *L); 143 144 /// OptimizeShadowIV - If IV is used in a int-to-float cast 145 /// inside the loop then try to eliminate the cast opeation. 146 void OptimizeShadowIV(Loop *L); 147 148 /// OptimizeMax - Rewrite the loop's terminating condition 149 /// if it uses a max computation. 150 ICmpInst *OptimizeMax(Loop *L, ICmpInst *Cond, 151 IVStrideUse* &CondUse); 152 153 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse, 154 const SCEV *const * &CondStride); 155 bool RequiresTypeConversion(const Type *Ty, const Type *NewTy); 156 const SCEV *CheckForIVReuse(bool, bool, bool, const SCEV *const&, 157 IVExpr&, const Type*, 158 const std::vector<BasedUser>& UsersToProcess); 159 bool ValidScale(bool, int64_t, 160 const std::vector<BasedUser>& UsersToProcess); 161 bool ValidOffset(bool, int64_t, int64_t, 162 const std::vector<BasedUser>& UsersToProcess); 163 const SCEV *CollectIVUsers(const SCEV *const &Stride, 164 IVUsersOfOneStride &Uses, 165 Loop *L, 166 bool &AllUsesAreAddresses, 167 bool &AllUsesAreOutsideLoop, 168 std::vector<BasedUser> &UsersToProcess); 169 bool ShouldUseFullStrengthReductionMode( 170 const std::vector<BasedUser> &UsersToProcess, 171 const Loop *L, 172 bool AllUsesAreAddresses, 173 const SCEV *Stride); 174 void PrepareToStrengthReduceFully( 175 std::vector<BasedUser> &UsersToProcess, 176 const SCEV *Stride, 177 const SCEV *CommonExprs, 178 const Loop *L, 179 SCEVExpander &PreheaderRewriter); 180 void PrepareToStrengthReduceFromSmallerStride( 181 std::vector<BasedUser> &UsersToProcess, 182 Value *CommonBaseV, 183 const IVExpr &ReuseIV, 184 Instruction *PreInsertPt); 185 void PrepareToStrengthReduceWithNewPhi( 186 std::vector<BasedUser> &UsersToProcess, 187 const SCEV *Stride, 188 const SCEV *CommonExprs, 189 Value *CommonBaseV, 190 Instruction *IVIncInsertPt, 191 const Loop *L, 192 SCEVExpander &PreheaderRewriter); 193 void StrengthReduceStridedIVUsers(const SCEV *const &Stride, 194 IVUsersOfOneStride &Uses, 195 Loop *L); 196 void DeleteTriviallyDeadInstructions(); 197 }; 198} 199 200char LoopStrengthReduce::ID = 0; 201static RegisterPass<LoopStrengthReduce> 202X("loop-reduce", "Loop Strength Reduction"); 203 204Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) { 205 return new LoopStrengthReduce(TLI); 206} 207 208/// DeleteTriviallyDeadInstructions - If any of the instructions is the 209/// specified set are trivially dead, delete them and see if this makes any of 210/// their operands subsequently dead. 211void LoopStrengthReduce::DeleteTriviallyDeadInstructions() { 212 if (DeadInsts.empty()) return; 213 214 while (!DeadInsts.empty()) { 215 Instruction *I = dyn_cast_or_null<Instruction>(DeadInsts.back()); 216 DeadInsts.pop_back(); 217 218 if (I == 0 || !isInstructionTriviallyDead(I)) 219 continue; 220 221 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) { 222 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 223 *OI = 0; 224 if (U->use_empty()) 225 DeadInsts.push_back(U); 226 } 227 } 228 229 I->eraseFromParent(); 230 Changed = true; 231 } 232} 233 234/// containsAddRecFromDifferentLoop - Determine whether expression S involves a 235/// subexpression that is an AddRec from a loop other than L. An outer loop 236/// of L is OK, but not an inner loop nor a disjoint loop. 237static bool containsAddRecFromDifferentLoop(const SCEV *S, Loop *L) { 238 // This is very common, put it first. 239 if (isa<SCEVConstant>(S)) 240 return false; 241 if (const SCEVCommutativeExpr *AE = dyn_cast<SCEVCommutativeExpr>(S)) { 242 for (unsigned int i=0; i< AE->getNumOperands(); i++) 243 if (containsAddRecFromDifferentLoop(AE->getOperand(i), L)) 244 return true; 245 return false; 246 } 247 if (const SCEVAddRecExpr *AE = dyn_cast<SCEVAddRecExpr>(S)) { 248 if (const Loop *newLoop = AE->getLoop()) { 249 if (newLoop == L) 250 return false; 251 // if newLoop is an outer loop of L, this is OK. 252 if (!LoopInfo::isNotAlreadyContainedIn(L, newLoop)) 253 return false; 254 } 255 return true; 256 } 257 if (const SCEVUDivExpr *DE = dyn_cast<SCEVUDivExpr>(S)) 258 return containsAddRecFromDifferentLoop(DE->getLHS(), L) || 259 containsAddRecFromDifferentLoop(DE->getRHS(), L); 260#if 0 261 // SCEVSDivExpr has been backed out temporarily, but will be back; we'll 262 // need this when it is. 263 if (const SCEVSDivExpr *DE = dyn_cast<SCEVSDivExpr>(S)) 264 return containsAddRecFromDifferentLoop(DE->getLHS(), L) || 265 containsAddRecFromDifferentLoop(DE->getRHS(), L); 266#endif 267 if (const SCEVCastExpr *CE = dyn_cast<SCEVCastExpr>(S)) 268 return containsAddRecFromDifferentLoop(CE->getOperand(), L); 269 return false; 270} 271 272/// isAddressUse - Returns true if the specified instruction is using the 273/// specified value as an address. 274static bool isAddressUse(Instruction *Inst, Value *OperandVal) { 275 bool isAddress = isa<LoadInst>(Inst); 276 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 277 if (SI->getOperand(1) == OperandVal) 278 isAddress = true; 279 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 280 // Addressing modes can also be folded into prefetches and a variety 281 // of intrinsics. 282 switch (II->getIntrinsicID()) { 283 default: break; 284 case Intrinsic::prefetch: 285 case Intrinsic::x86_sse2_loadu_dq: 286 case Intrinsic::x86_sse2_loadu_pd: 287 case Intrinsic::x86_sse_loadu_ps: 288 case Intrinsic::x86_sse_storeu_ps: 289 case Intrinsic::x86_sse2_storeu_pd: 290 case Intrinsic::x86_sse2_storeu_dq: 291 case Intrinsic::x86_sse2_storel_dq: 292 if (II->getOperand(1) == OperandVal) 293 isAddress = true; 294 break; 295 } 296 } 297 return isAddress; 298} 299 300/// getAccessType - Return the type of the memory being accessed. 301static const Type *getAccessType(const Instruction *Inst) { 302 const Type *AccessTy = Inst->getType(); 303 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) 304 AccessTy = SI->getOperand(0)->getType(); 305 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 306 // Addressing modes can also be folded into prefetches and a variety 307 // of intrinsics. 308 switch (II->getIntrinsicID()) { 309 default: break; 310 case Intrinsic::x86_sse_storeu_ps: 311 case Intrinsic::x86_sse2_storeu_pd: 312 case Intrinsic::x86_sse2_storeu_dq: 313 case Intrinsic::x86_sse2_storel_dq: 314 AccessTy = II->getOperand(1)->getType(); 315 break; 316 } 317 } 318 return AccessTy; 319} 320 321namespace { 322 /// BasedUser - For a particular base value, keep information about how we've 323 /// partitioned the expression so far. 324 struct BasedUser { 325 /// SE - The current ScalarEvolution object. 326 ScalarEvolution *SE; 327 328 /// Base - The Base value for the PHI node that needs to be inserted for 329 /// this use. As the use is processed, information gets moved from this 330 /// field to the Imm field (below). BasedUser values are sorted by this 331 /// field. 332 const SCEV *Base; 333 334 /// Inst - The instruction using the induction variable. 335 Instruction *Inst; 336 337 /// OperandValToReplace - The operand value of Inst to replace with the 338 /// EmittedBase. 339 Value *OperandValToReplace; 340 341 /// Imm - The immediate value that should be added to the base immediately 342 /// before Inst, because it will be folded into the imm field of the 343 /// instruction. This is also sometimes used for loop-variant values that 344 /// must be added inside the loop. 345 const SCEV *Imm; 346 347 /// Phi - The induction variable that performs the striding that 348 /// should be used for this user. 349 PHINode *Phi; 350 351 // isUseOfPostIncrementedValue - True if this should use the 352 // post-incremented version of this IV, not the preincremented version. 353 // This can only be set in special cases, such as the terminating setcc 354 // instruction for a loop and uses outside the loop that are dominated by 355 // the loop. 356 bool isUseOfPostIncrementedValue; 357 358 BasedUser(IVStrideUse &IVSU, ScalarEvolution *se) 359 : SE(se), Base(IVSU.getOffset()), Inst(IVSU.getUser()), 360 OperandValToReplace(IVSU.getOperandValToReplace()), 361 Imm(SE->getIntegerSCEV(0, Base->getType())), 362 isUseOfPostIncrementedValue(IVSU.isUseOfPostIncrementedValue()) {} 363 364 // Once we rewrite the code to insert the new IVs we want, update the 365 // operands of Inst to use the new expression 'NewBase', with 'Imm' added 366 // to it. 367 void RewriteInstructionToUseNewBase(const SCEV *const &NewBase, 368 Instruction *InsertPt, 369 SCEVExpander &Rewriter, Loop *L, Pass *P, 370 LoopInfo &LI, 371 SmallVectorImpl<WeakVH> &DeadInsts); 372 373 Value *InsertCodeForBaseAtPosition(const SCEV *const &NewBase, 374 const Type *Ty, 375 SCEVExpander &Rewriter, 376 Instruction *IP, Loop *L, 377 LoopInfo &LI); 378 void dump() const; 379 }; 380} 381 382void BasedUser::dump() const { 383 cerr << " Base=" << *Base; 384 cerr << " Imm=" << *Imm; 385 cerr << " Inst: " << *Inst; 386} 387 388Value *BasedUser::InsertCodeForBaseAtPosition(const SCEV *const &NewBase, 389 const Type *Ty, 390 SCEVExpander &Rewriter, 391 Instruction *IP, Loop *L, 392 LoopInfo &LI) { 393 // Figure out where we *really* want to insert this code. In particular, if 394 // the user is inside of a loop that is nested inside of L, we really don't 395 // want to insert this expression before the user, we'd rather pull it out as 396 // many loops as possible. 397 Instruction *BaseInsertPt = IP; 398 399 // Figure out the most-nested loop that IP is in. 400 Loop *InsertLoop = LI.getLoopFor(IP->getParent()); 401 402 // If InsertLoop is not L, and InsertLoop is nested inside of L, figure out 403 // the preheader of the outer-most loop where NewBase is not loop invariant. 404 if (L->contains(IP->getParent())) 405 while (InsertLoop && NewBase->isLoopInvariant(InsertLoop)) { 406 BaseInsertPt = InsertLoop->getLoopPreheader()->getTerminator(); 407 InsertLoop = InsertLoop->getParentLoop(); 408 } 409 410 Value *Base = Rewriter.expandCodeFor(NewBase, 0, BaseInsertPt); 411 412 const SCEV *NewValSCEV = SE->getUnknown(Base); 413 414 // Always emit the immediate into the same block as the user. 415 NewValSCEV = SE->getAddExpr(NewValSCEV, Imm); 416 417 return Rewriter.expandCodeFor(NewValSCEV, Ty, IP); 418} 419 420 421// Once we rewrite the code to insert the new IVs we want, update the 422// operands of Inst to use the new expression 'NewBase', with 'Imm' added 423// to it. NewBasePt is the last instruction which contributes to the 424// value of NewBase in the case that it's a diffferent instruction from 425// the PHI that NewBase is computed from, or null otherwise. 426// 427void BasedUser::RewriteInstructionToUseNewBase(const SCEV *const &NewBase, 428 Instruction *NewBasePt, 429 SCEVExpander &Rewriter, Loop *L, Pass *P, 430 LoopInfo &LI, 431 SmallVectorImpl<WeakVH> &DeadInsts) { 432 if (!isa<PHINode>(Inst)) { 433 // By default, insert code at the user instruction. 434 BasicBlock::iterator InsertPt = Inst; 435 436 // However, if the Operand is itself an instruction, the (potentially 437 // complex) inserted code may be shared by many users. Because of this, we 438 // want to emit code for the computation of the operand right before its old 439 // computation. This is usually safe, because we obviously used to use the 440 // computation when it was computed in its current block. However, in some 441 // cases (e.g. use of a post-incremented induction variable) the NewBase 442 // value will be pinned to live somewhere after the original computation. 443 // In this case, we have to back off. 444 // 445 // If this is a use outside the loop (which means after, since it is based 446 // on a loop indvar) we use the post-incremented value, so that we don't 447 // artificially make the preinc value live out the bottom of the loop. 448 if (!isUseOfPostIncrementedValue && L->contains(Inst->getParent())) { 449 if (NewBasePt && isa<PHINode>(OperandValToReplace)) { 450 InsertPt = NewBasePt; 451 ++InsertPt; 452 } else if (Instruction *OpInst 453 = dyn_cast<Instruction>(OperandValToReplace)) { 454 InsertPt = OpInst; 455 while (isa<PHINode>(InsertPt)) ++InsertPt; 456 } 457 } 458 Value *NewVal = InsertCodeForBaseAtPosition(NewBase, 459 OperandValToReplace->getType(), 460 Rewriter, InsertPt, L, LI); 461 // Replace the use of the operand Value with the new Phi we just created. 462 Inst->replaceUsesOfWith(OperandValToReplace, NewVal); 463 464 DOUT << " Replacing with "; 465 DEBUG(WriteAsOperand(*DOUT, NewVal, /*PrintType=*/false)); 466 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n"; 467 return; 468 } 469 470 // PHI nodes are more complex. We have to insert one copy of the NewBase+Imm 471 // expression into each operand block that uses it. Note that PHI nodes can 472 // have multiple entries for the same predecessor. We use a map to make sure 473 // that a PHI node only has a single Value* for each predecessor (which also 474 // prevents us from inserting duplicate code in some blocks). 475 DenseMap<BasicBlock*, Value*> InsertedCode; 476 PHINode *PN = cast<PHINode>(Inst); 477 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 478 if (PN->getIncomingValue(i) == OperandValToReplace) { 479 // If the original expression is outside the loop, put the replacement 480 // code in the same place as the original expression, 481 // which need not be an immediate predecessor of this PHI. This way we 482 // need only one copy of it even if it is referenced multiple times in 483 // the PHI. We don't do this when the original expression is inside the 484 // loop because multiple copies sometimes do useful sinking of code in 485 // that case(?). 486 Instruction *OldLoc = dyn_cast<Instruction>(OperandValToReplace); 487 if (L->contains(OldLoc->getParent())) { 488 // If this is a critical edge, split the edge so that we do not insert 489 // the code on all predecessor/successor paths. We do this unless this 490 // is the canonical backedge for this loop, as this can make some 491 // inserted code be in an illegal position. 492 BasicBlock *PHIPred = PN->getIncomingBlock(i); 493 if (e != 1 && PHIPred->getTerminator()->getNumSuccessors() > 1 && 494 (PN->getParent() != L->getHeader() || !L->contains(PHIPred))) { 495 496 // First step, split the critical edge. 497 SplitCriticalEdge(PHIPred, PN->getParent(), P, false); 498 499 // Next step: move the basic block. In particular, if the PHI node 500 // is outside of the loop, and PredTI is in the loop, we want to 501 // move the block to be immediately before the PHI block, not 502 // immediately after PredTI. 503 if (L->contains(PHIPred) && !L->contains(PN->getParent())) { 504 BasicBlock *NewBB = PN->getIncomingBlock(i); 505 NewBB->moveBefore(PN->getParent()); 506 } 507 508 // Splitting the edge can reduce the number of PHI entries we have. 509 e = PN->getNumIncomingValues(); 510 } 511 } 512 Value *&Code = InsertedCode[PN->getIncomingBlock(i)]; 513 if (!Code) { 514 // Insert the code into the end of the predecessor block. 515 Instruction *InsertPt = (L->contains(OldLoc->getParent())) ? 516 PN->getIncomingBlock(i)->getTerminator() : 517 OldLoc->getParent()->getTerminator(); 518 Code = InsertCodeForBaseAtPosition(NewBase, PN->getType(), 519 Rewriter, InsertPt, L, LI); 520 521 DOUT << " Changing PHI use to "; 522 DEBUG(WriteAsOperand(*DOUT, Code, /*PrintType=*/false)); 523 DOUT << ", which has value " << *NewBase << " plus IMM " << *Imm << "\n"; 524 } 525 526 // Replace the use of the operand Value with the new Phi we just created. 527 PN->setIncomingValue(i, Code); 528 Rewriter.clear(); 529 } 530 } 531 532 // PHI node might have become a constant value after SplitCriticalEdge. 533 DeadInsts.push_back(Inst); 534} 535 536 537/// fitsInAddressMode - Return true if V can be subsumed within an addressing 538/// mode, and does not need to be put in a register first. 539static bool fitsInAddressMode(const SCEV *const &V, const Type *AccessTy, 540 const TargetLowering *TLI, bool HasBaseReg) { 541 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) { 542 int64_t VC = SC->getValue()->getSExtValue(); 543 if (TLI) { 544 TargetLowering::AddrMode AM; 545 AM.BaseOffs = VC; 546 AM.HasBaseReg = HasBaseReg; 547 return TLI->isLegalAddressingMode(AM, AccessTy); 548 } else { 549 // Defaults to PPC. PPC allows a sign-extended 16-bit immediate field. 550 return (VC > -(1 << 16) && VC < (1 << 16)-1); 551 } 552 } 553 554 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) 555 if (GlobalValue *GV = dyn_cast<GlobalValue>(SU->getValue())) { 556 if (TLI) { 557 TargetLowering::AddrMode AM; 558 AM.BaseGV = GV; 559 AM.HasBaseReg = HasBaseReg; 560 return TLI->isLegalAddressingMode(AM, AccessTy); 561 } else { 562 // Default: assume global addresses are not legal. 563 } 564 } 565 566 return false; 567} 568 569/// MoveLoopVariantsToImmediateField - Move any subexpressions from Val that are 570/// loop varying to the Imm operand. 571static void MoveLoopVariantsToImmediateField(const SCEV *&Val, const SCEV *&Imm, 572 Loop *L, ScalarEvolution *SE) { 573 if (Val->isLoopInvariant(L)) return; // Nothing to do. 574 575 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) { 576 SmallVector<const SCEV *, 4> NewOps; 577 NewOps.reserve(SAE->getNumOperands()); 578 579 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) 580 if (!SAE->getOperand(i)->isLoopInvariant(L)) { 581 // If this is a loop-variant expression, it must stay in the immediate 582 // field of the expression. 583 Imm = SE->getAddExpr(Imm, SAE->getOperand(i)); 584 } else { 585 NewOps.push_back(SAE->getOperand(i)); 586 } 587 588 if (NewOps.empty()) 589 Val = SE->getIntegerSCEV(0, Val->getType()); 590 else 591 Val = SE->getAddExpr(NewOps); 592 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) { 593 // Try to pull immediates out of the start value of nested addrec's. 594 const SCEV *Start = SARE->getStart(); 595 MoveLoopVariantsToImmediateField(Start, Imm, L, SE); 596 597 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end()); 598 Ops[0] = Start; 599 Val = SE->getAddRecExpr(Ops, SARE->getLoop()); 600 } else { 601 // Otherwise, all of Val is variant, move the whole thing over. 602 Imm = SE->getAddExpr(Imm, Val); 603 Val = SE->getIntegerSCEV(0, Val->getType()); 604 } 605} 606 607 608/// MoveImmediateValues - Look at Val, and pull out any additions of constants 609/// that can fit into the immediate field of instructions in the target. 610/// Accumulate these immediate values into the Imm value. 611static void MoveImmediateValues(const TargetLowering *TLI, 612 const Type *AccessTy, 613 const SCEV *&Val, const SCEV *&Imm, 614 bool isAddress, Loop *L, 615 ScalarEvolution *SE) { 616 if (const SCEVAddExpr *SAE = dyn_cast<SCEVAddExpr>(Val)) { 617 SmallVector<const SCEV *, 4> NewOps; 618 NewOps.reserve(SAE->getNumOperands()); 619 620 for (unsigned i = 0; i != SAE->getNumOperands(); ++i) { 621 const SCEV *NewOp = SAE->getOperand(i); 622 MoveImmediateValues(TLI, AccessTy, NewOp, Imm, isAddress, L, SE); 623 624 if (!NewOp->isLoopInvariant(L)) { 625 // If this is a loop-variant expression, it must stay in the immediate 626 // field of the expression. 627 Imm = SE->getAddExpr(Imm, NewOp); 628 } else { 629 NewOps.push_back(NewOp); 630 } 631 } 632 633 if (NewOps.empty()) 634 Val = SE->getIntegerSCEV(0, Val->getType()); 635 else 636 Val = SE->getAddExpr(NewOps); 637 return; 638 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Val)) { 639 // Try to pull immediates out of the start value of nested addrec's. 640 const SCEV *Start = SARE->getStart(); 641 MoveImmediateValues(TLI, AccessTy, Start, Imm, isAddress, L, SE); 642 643 if (Start != SARE->getStart()) { 644 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end()); 645 Ops[0] = Start; 646 Val = SE->getAddRecExpr(Ops, SARE->getLoop()); 647 } 648 return; 649 } else if (const SCEVMulExpr *SME = dyn_cast<SCEVMulExpr>(Val)) { 650 // Transform "8 * (4 + v)" -> "32 + 8*V" if "32" fits in the immed field. 651 if (isAddress && 652 fitsInAddressMode(SME->getOperand(0), AccessTy, TLI, false) && 653 SME->getNumOperands() == 2 && SME->isLoopInvariant(L)) { 654 655 const SCEV *SubImm = SE->getIntegerSCEV(0, Val->getType()); 656 const SCEV *NewOp = SME->getOperand(1); 657 MoveImmediateValues(TLI, AccessTy, NewOp, SubImm, isAddress, L, SE); 658 659 // If we extracted something out of the subexpressions, see if we can 660 // simplify this! 661 if (NewOp != SME->getOperand(1)) { 662 // Scale SubImm up by "8". If the result is a target constant, we are 663 // good. 664 SubImm = SE->getMulExpr(SubImm, SME->getOperand(0)); 665 if (fitsInAddressMode(SubImm, AccessTy, TLI, false)) { 666 // Accumulate the immediate. 667 Imm = SE->getAddExpr(Imm, SubImm); 668 669 // Update what is left of 'Val'. 670 Val = SE->getMulExpr(SME->getOperand(0), NewOp); 671 return; 672 } 673 } 674 } 675 } 676 677 // Loop-variant expressions must stay in the immediate field of the 678 // expression. 679 if ((isAddress && fitsInAddressMode(Val, AccessTy, TLI, false)) || 680 !Val->isLoopInvariant(L)) { 681 Imm = SE->getAddExpr(Imm, Val); 682 Val = SE->getIntegerSCEV(0, Val->getType()); 683 return; 684 } 685 686 // Otherwise, no immediates to move. 687} 688 689static void MoveImmediateValues(const TargetLowering *TLI, 690 Instruction *User, 691 const SCEV *&Val, const SCEV *&Imm, 692 bool isAddress, Loop *L, 693 ScalarEvolution *SE) { 694 const Type *AccessTy = getAccessType(User); 695 MoveImmediateValues(TLI, AccessTy, Val, Imm, isAddress, L, SE); 696} 697 698/// SeparateSubExprs - Decompose Expr into all of the subexpressions that are 699/// added together. This is used to reassociate common addition subexprs 700/// together for maximal sharing when rewriting bases. 701static void SeparateSubExprs(SmallVector<const SCEV *, 16> &SubExprs, 702 const SCEV *Expr, 703 ScalarEvolution *SE) { 704 if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(Expr)) { 705 for (unsigned j = 0, e = AE->getNumOperands(); j != e; ++j) 706 SeparateSubExprs(SubExprs, AE->getOperand(j), SE); 707 } else if (const SCEVAddRecExpr *SARE = dyn_cast<SCEVAddRecExpr>(Expr)) { 708 const SCEV *Zero = SE->getIntegerSCEV(0, Expr->getType()); 709 if (SARE->getOperand(0) == Zero) { 710 SubExprs.push_back(Expr); 711 } else { 712 // Compute the addrec with zero as its base. 713 SmallVector<const SCEV *, 4> Ops(SARE->op_begin(), SARE->op_end()); 714 Ops[0] = Zero; // Start with zero base. 715 SubExprs.push_back(SE->getAddRecExpr(Ops, SARE->getLoop())); 716 717 718 SeparateSubExprs(SubExprs, SARE->getOperand(0), SE); 719 } 720 } else if (!Expr->isZero()) { 721 // Do not add zero. 722 SubExprs.push_back(Expr); 723 } 724} 725 726// This is logically local to the following function, but C++ says we have 727// to make it file scope. 728struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; }; 729 730/// RemoveCommonExpressionsFromUseBases - Look through all of the Bases of all 731/// the Uses, removing any common subexpressions, except that if all such 732/// subexpressions can be folded into an addressing mode for all uses inside 733/// the loop (this case is referred to as "free" in comments herein) we do 734/// not remove anything. This looks for things like (a+b+c) and 735/// (a+c+d) and computes the common (a+c) subexpression. The common expression 736/// is *removed* from the Bases and returned. 737static const SCEV * 738RemoveCommonExpressionsFromUseBases(std::vector<BasedUser> &Uses, 739 ScalarEvolution *SE, Loop *L, 740 const TargetLowering *TLI) { 741 unsigned NumUses = Uses.size(); 742 743 // Only one use? This is a very common case, so we handle it specially and 744 // cheaply. 745 const SCEV *Zero = SE->getIntegerSCEV(0, Uses[0].Base->getType()); 746 const SCEV *Result = Zero; 747 const SCEV *FreeResult = Zero; 748 if (NumUses == 1) { 749 // If the use is inside the loop, use its base, regardless of what it is: 750 // it is clearly shared across all the IV's. If the use is outside the loop 751 // (which means after it) we don't want to factor anything *into* the loop, 752 // so just use 0 as the base. 753 if (L->contains(Uses[0].Inst->getParent())) 754 std::swap(Result, Uses[0].Base); 755 return Result; 756 } 757 758 // To find common subexpressions, count how many of Uses use each expression. 759 // If any subexpressions are used Uses.size() times, they are common. 760 // Also track whether all uses of each expression can be moved into an 761 // an addressing mode "for free"; such expressions are left within the loop. 762 // struct SubExprUseData { unsigned Count; bool notAllUsesAreFree; }; 763 std::map<const SCEV *, SubExprUseData> SubExpressionUseData; 764 765 // UniqueSubExprs - Keep track of all of the subexpressions we see in the 766 // order we see them. 767 SmallVector<const SCEV *, 16> UniqueSubExprs; 768 769 SmallVector<const SCEV *, 16> SubExprs; 770 unsigned NumUsesInsideLoop = 0; 771 for (unsigned i = 0; i != NumUses; ++i) { 772 // If the user is outside the loop, just ignore it for base computation. 773 // Since the user is outside the loop, it must be *after* the loop (if it 774 // were before, it could not be based on the loop IV). We don't want users 775 // after the loop to affect base computation of values *inside* the loop, 776 // because we can always add their offsets to the result IV after the loop 777 // is done, ensuring we get good code inside the loop. 778 if (!L->contains(Uses[i].Inst->getParent())) 779 continue; 780 NumUsesInsideLoop++; 781 782 // If the base is zero (which is common), return zero now, there are no 783 // CSEs we can find. 784 if (Uses[i].Base == Zero) return Zero; 785 786 // If this use is as an address we may be able to put CSEs in the addressing 787 // mode rather than hoisting them. 788 bool isAddrUse = isAddressUse(Uses[i].Inst, Uses[i].OperandValToReplace); 789 // We may need the AccessTy below, but only when isAddrUse, so compute it 790 // only in that case. 791 const Type *AccessTy = 0; 792 if (isAddrUse) 793 AccessTy = getAccessType(Uses[i].Inst); 794 795 // Split the expression into subexprs. 796 SeparateSubExprs(SubExprs, Uses[i].Base, SE); 797 // Add one to SubExpressionUseData.Count for each subexpr present, and 798 // if the subexpr is not a valid immediate within an addressing mode use, 799 // set SubExpressionUseData.notAllUsesAreFree. We definitely want to 800 // hoist these out of the loop (if they are common to all uses). 801 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) { 802 if (++SubExpressionUseData[SubExprs[j]].Count == 1) 803 UniqueSubExprs.push_back(SubExprs[j]); 804 if (!isAddrUse || !fitsInAddressMode(SubExprs[j], AccessTy, TLI, false)) 805 SubExpressionUseData[SubExprs[j]].notAllUsesAreFree = true; 806 } 807 SubExprs.clear(); 808 } 809 810 // Now that we know how many times each is used, build Result. Iterate over 811 // UniqueSubexprs so that we have a stable ordering. 812 for (unsigned i = 0, e = UniqueSubExprs.size(); i != e; ++i) { 813 std::map<const SCEV *, SubExprUseData>::iterator I = 814 SubExpressionUseData.find(UniqueSubExprs[i]); 815 assert(I != SubExpressionUseData.end() && "Entry not found?"); 816 if (I->second.Count == NumUsesInsideLoop) { // Found CSE! 817 if (I->second.notAllUsesAreFree) 818 Result = SE->getAddExpr(Result, I->first); 819 else 820 FreeResult = SE->getAddExpr(FreeResult, I->first); 821 } else 822 // Remove non-cse's from SubExpressionUseData. 823 SubExpressionUseData.erase(I); 824 } 825 826 if (FreeResult != Zero) { 827 // We have some subexpressions that can be subsumed into addressing 828 // modes in every use inside the loop. However, it's possible that 829 // there are so many of them that the combined FreeResult cannot 830 // be subsumed, or that the target cannot handle both a FreeResult 831 // and a Result in the same instruction (for example because it would 832 // require too many registers). Check this. 833 for (unsigned i=0; i<NumUses; ++i) { 834 if (!L->contains(Uses[i].Inst->getParent())) 835 continue; 836 // We know this is an addressing mode use; if there are any uses that 837 // are not, FreeResult would be Zero. 838 const Type *AccessTy = getAccessType(Uses[i].Inst); 839 if (!fitsInAddressMode(FreeResult, AccessTy, TLI, Result!=Zero)) { 840 // FIXME: could split up FreeResult into pieces here, some hoisted 841 // and some not. There is no obvious advantage to this. 842 Result = SE->getAddExpr(Result, FreeResult); 843 FreeResult = Zero; 844 break; 845 } 846 } 847 } 848 849 // If we found no CSE's, return now. 850 if (Result == Zero) return Result; 851 852 // If we still have a FreeResult, remove its subexpressions from 853 // SubExpressionUseData. This means they will remain in the use Bases. 854 if (FreeResult != Zero) { 855 SeparateSubExprs(SubExprs, FreeResult, SE); 856 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) { 857 std::map<const SCEV *, SubExprUseData>::iterator I = 858 SubExpressionUseData.find(SubExprs[j]); 859 SubExpressionUseData.erase(I); 860 } 861 SubExprs.clear(); 862 } 863 864 // Otherwise, remove all of the CSE's we found from each of the base values. 865 for (unsigned i = 0; i != NumUses; ++i) { 866 // Uses outside the loop don't necessarily include the common base, but 867 // the final IV value coming into those uses does. Instead of trying to 868 // remove the pieces of the common base, which might not be there, 869 // subtract off the base to compensate for this. 870 if (!L->contains(Uses[i].Inst->getParent())) { 871 Uses[i].Base = SE->getMinusSCEV(Uses[i].Base, Result); 872 continue; 873 } 874 875 // Split the expression into subexprs. 876 SeparateSubExprs(SubExprs, Uses[i].Base, SE); 877 878 // Remove any common subexpressions. 879 for (unsigned j = 0, e = SubExprs.size(); j != e; ++j) 880 if (SubExpressionUseData.count(SubExprs[j])) { 881 SubExprs.erase(SubExprs.begin()+j); 882 --j; --e; 883 } 884 885 // Finally, add the non-shared expressions together. 886 if (SubExprs.empty()) 887 Uses[i].Base = Zero; 888 else 889 Uses[i].Base = SE->getAddExpr(SubExprs); 890 SubExprs.clear(); 891 } 892 893 return Result; 894} 895 896/// ValidScale - Check whether the given Scale is valid for all loads and 897/// stores in UsersToProcess. 898/// 899bool LoopStrengthReduce::ValidScale(bool HasBaseReg, int64_t Scale, 900 const std::vector<BasedUser>& UsersToProcess) { 901 if (!TLI) 902 return true; 903 904 for (unsigned i = 0, e = UsersToProcess.size(); i!=e; ++i) { 905 // If this is a load or other access, pass the type of the access in. 906 const Type *AccessTy = Type::VoidTy; 907 if (isAddressUse(UsersToProcess[i].Inst, 908 UsersToProcess[i].OperandValToReplace)) 909 AccessTy = getAccessType(UsersToProcess[i].Inst); 910 else if (isa<PHINode>(UsersToProcess[i].Inst)) 911 continue; 912 913 TargetLowering::AddrMode AM; 914 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm)) 915 AM.BaseOffs = SC->getValue()->getSExtValue(); 916 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero(); 917 AM.Scale = Scale; 918 919 // If load[imm+r*scale] is illegal, bail out. 920 if (!TLI->isLegalAddressingMode(AM, AccessTy)) 921 return false; 922 } 923 return true; 924} 925 926/// ValidOffset - Check whether the given Offset is valid for all loads and 927/// stores in UsersToProcess. 928/// 929bool LoopStrengthReduce::ValidOffset(bool HasBaseReg, 930 int64_t Offset, 931 int64_t Scale, 932 const std::vector<BasedUser>& UsersToProcess) { 933 if (!TLI) 934 return true; 935 936 for (unsigned i=0, e = UsersToProcess.size(); i!=e; ++i) { 937 // If this is a load or other access, pass the type of the access in. 938 const Type *AccessTy = Type::VoidTy; 939 if (isAddressUse(UsersToProcess[i].Inst, 940 UsersToProcess[i].OperandValToReplace)) 941 AccessTy = getAccessType(UsersToProcess[i].Inst); 942 else if (isa<PHINode>(UsersToProcess[i].Inst)) 943 continue; 944 945 TargetLowering::AddrMode AM; 946 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(UsersToProcess[i].Imm)) 947 AM.BaseOffs = SC->getValue()->getSExtValue(); 948 AM.BaseOffs = (uint64_t)AM.BaseOffs + (uint64_t)Offset; 949 AM.HasBaseReg = HasBaseReg || !UsersToProcess[i].Base->isZero(); 950 AM.Scale = Scale; 951 952 // If load[imm+r*scale] is illegal, bail out. 953 if (!TLI->isLegalAddressingMode(AM, AccessTy)) 954 return false; 955 } 956 return true; 957} 958 959/// RequiresTypeConversion - Returns true if converting Ty1 to Ty2 is not 960/// a nop. 961bool LoopStrengthReduce::RequiresTypeConversion(const Type *Ty1, 962 const Type *Ty2) { 963 if (Ty1 == Ty2) 964 return false; 965 Ty1 = SE->getEffectiveSCEVType(Ty1); 966 Ty2 = SE->getEffectiveSCEVType(Ty2); 967 if (Ty1 == Ty2) 968 return false; 969 if (Ty1->canLosslesslyBitCastTo(Ty2)) 970 return false; 971 if (TLI && TLI->isTruncateFree(Ty1, Ty2)) 972 return false; 973 return true; 974} 975 976/// CheckForIVReuse - Returns the multiple if the stride is the multiple 977/// of a previous stride and it is a legal value for the target addressing 978/// mode scale component and optional base reg. This allows the users of 979/// this stride to be rewritten as prev iv * factor. It returns 0 if no 980/// reuse is possible. Factors can be negative on same targets, e.g. ARM. 981/// 982/// If all uses are outside the loop, we don't require that all multiplies 983/// be folded into the addressing mode, nor even that the factor be constant; 984/// a multiply (executed once) outside the loop is better than another IV 985/// within. Well, usually. 986const SCEV *LoopStrengthReduce::CheckForIVReuse(bool HasBaseReg, 987 bool AllUsesAreAddresses, 988 bool AllUsesAreOutsideLoop, 989 const SCEV *const &Stride, 990 IVExpr &IV, const Type *Ty, 991 const std::vector<BasedUser>& UsersToProcess) { 992 if (StrideNoReuse.count(Stride)) 993 return SE->getIntegerSCEV(0, Stride->getType()); 994 995 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Stride)) { 996 int64_t SInt = SC->getValue()->getSExtValue(); 997 for (unsigned NewStride = 0, e = IU->StrideOrder.size(); 998 NewStride != e; ++NewStride) { 999 std::map<const SCEV *, IVsOfOneStride>::iterator SI = 1000 IVsByStride.find(IU->StrideOrder[NewStride]); 1001 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first) || 1002 StrideNoReuse.count(SI->first)) 1003 continue; 1004 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue(); 1005 if (SI->first != Stride && 1006 (unsigned(abs64(SInt)) < SSInt || (SInt % SSInt) != 0)) 1007 continue; 1008 int64_t Scale = SInt / SSInt; 1009 // Check that this stride is valid for all the types used for loads and 1010 // stores; if it can be used for some and not others, we might as well use 1011 // the original stride everywhere, since we have to create the IV for it 1012 // anyway. If the scale is 1, then we don't need to worry about folding 1013 // multiplications. 1014 if (Scale == 1 || 1015 (AllUsesAreAddresses && 1016 ValidScale(HasBaseReg, Scale, UsersToProcess))) { 1017 // Prefer to reuse an IV with a base of zero. 1018 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(), 1019 IE = SI->second.IVs.end(); II != IE; ++II) 1020 // Only reuse previous IV if it would not require a type conversion 1021 // and if the base difference can be folded. 1022 if (II->Base->isZero() && 1023 !RequiresTypeConversion(II->Base->getType(), Ty)) { 1024 IV = *II; 1025 return SE->getIntegerSCEV(Scale, Stride->getType()); 1026 } 1027 // Otherwise, settle for an IV with a foldable base. 1028 if (AllUsesAreAddresses) 1029 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(), 1030 IE = SI->second.IVs.end(); II != IE; ++II) 1031 // Only reuse previous IV if it would not require a type conversion 1032 // and if the base difference can be folded. 1033 if (SE->getEffectiveSCEVType(II->Base->getType()) == 1034 SE->getEffectiveSCEVType(Ty) && 1035 isa<SCEVConstant>(II->Base)) { 1036 int64_t Base = 1037 cast<SCEVConstant>(II->Base)->getValue()->getSExtValue(); 1038 if (Base > INT32_MIN && Base <= INT32_MAX && 1039 ValidOffset(HasBaseReg, -Base * Scale, 1040 Scale, UsersToProcess)) { 1041 IV = *II; 1042 return SE->getIntegerSCEV(Scale, Stride->getType()); 1043 } 1044 } 1045 } 1046 } 1047 } else if (AllUsesAreOutsideLoop) { 1048 // Accept nonconstant strides here; it is really really right to substitute 1049 // an existing IV if we can. 1050 for (unsigned NewStride = 0, e = IU->StrideOrder.size(); 1051 NewStride != e; ++NewStride) { 1052 std::map<const SCEV *, IVsOfOneStride>::iterator SI = 1053 IVsByStride.find(IU->StrideOrder[NewStride]); 1054 if (SI == IVsByStride.end() || !isa<SCEVConstant>(SI->first)) 1055 continue; 1056 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue(); 1057 if (SI->first != Stride && SSInt != 1) 1058 continue; 1059 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(), 1060 IE = SI->second.IVs.end(); II != IE; ++II) 1061 // Accept nonzero base here. 1062 // Only reuse previous IV if it would not require a type conversion. 1063 if (!RequiresTypeConversion(II->Base->getType(), Ty)) { 1064 IV = *II; 1065 return Stride; 1066 } 1067 } 1068 // Special case, old IV is -1*x and this one is x. Can treat this one as 1069 // -1*old. 1070 for (unsigned NewStride = 0, e = IU->StrideOrder.size(); 1071 NewStride != e; ++NewStride) { 1072 std::map<const SCEV *, IVsOfOneStride>::iterator SI = 1073 IVsByStride.find(IU->StrideOrder[NewStride]); 1074 if (SI == IVsByStride.end()) 1075 continue; 1076 if (const SCEVMulExpr *ME = dyn_cast<SCEVMulExpr>(SI->first)) 1077 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(ME->getOperand(0))) 1078 if (Stride == ME->getOperand(1) && 1079 SC->getValue()->getSExtValue() == -1LL) 1080 for (std::vector<IVExpr>::iterator II = SI->second.IVs.begin(), 1081 IE = SI->second.IVs.end(); II != IE; ++II) 1082 // Accept nonzero base here. 1083 // Only reuse previous IV if it would not require type conversion. 1084 if (!RequiresTypeConversion(II->Base->getType(), Ty)) { 1085 IV = *II; 1086 return SE->getIntegerSCEV(-1LL, Stride->getType()); 1087 } 1088 } 1089 } 1090 return SE->getIntegerSCEV(0, Stride->getType()); 1091} 1092 1093/// PartitionByIsUseOfPostIncrementedValue - Simple boolean predicate that 1094/// returns true if Val's isUseOfPostIncrementedValue is true. 1095static bool PartitionByIsUseOfPostIncrementedValue(const BasedUser &Val) { 1096 return Val.isUseOfPostIncrementedValue; 1097} 1098 1099/// isNonConstantNegative - Return true if the specified scev is negated, but 1100/// not a constant. 1101static bool isNonConstantNegative(const SCEV *const &Expr) { 1102 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Expr); 1103 if (!Mul) return false; 1104 1105 // If there is a constant factor, it will be first. 1106 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0)); 1107 if (!SC) return false; 1108 1109 // Return true if the value is negative, this matches things like (-42 * V). 1110 return SC->getValue()->getValue().isNegative(); 1111} 1112 1113/// CollectIVUsers - Transform our list of users and offsets to a bit more 1114/// complex table. In this new vector, each 'BasedUser' contains 'Base', the base 1115/// of the strided accesses, as well as the old information from Uses. We 1116/// progressively move information from the Base field to the Imm field, until 1117/// we eventually have the full access expression to rewrite the use. 1118const SCEV *LoopStrengthReduce::CollectIVUsers(const SCEV *const &Stride, 1119 IVUsersOfOneStride &Uses, 1120 Loop *L, 1121 bool &AllUsesAreAddresses, 1122 bool &AllUsesAreOutsideLoop, 1123 std::vector<BasedUser> &UsersToProcess) { 1124 // FIXME: Generalize to non-affine IV's. 1125 if (!Stride->isLoopInvariant(L)) 1126 return SE->getIntegerSCEV(0, Stride->getType()); 1127 1128 UsersToProcess.reserve(Uses.Users.size()); 1129 for (ilist<IVStrideUse>::iterator I = Uses.Users.begin(), 1130 E = Uses.Users.end(); I != E; ++I) { 1131 UsersToProcess.push_back(BasedUser(*I, SE)); 1132 1133 // Move any loop variant operands from the offset field to the immediate 1134 // field of the use, so that we don't try to use something before it is 1135 // computed. 1136 MoveLoopVariantsToImmediateField(UsersToProcess.back().Base, 1137 UsersToProcess.back().Imm, L, SE); 1138 assert(UsersToProcess.back().Base->isLoopInvariant(L) && 1139 "Base value is not loop invariant!"); 1140 } 1141 1142 // We now have a whole bunch of uses of like-strided induction variables, but 1143 // they might all have different bases. We want to emit one PHI node for this 1144 // stride which we fold as many common expressions (between the IVs) into as 1145 // possible. Start by identifying the common expressions in the base values 1146 // for the strides (e.g. if we have "A+C+B" and "A+B+D" as our bases, find 1147 // "A+B"), emit it to the preheader, then remove the expression from the 1148 // UsersToProcess base values. 1149 const SCEV *CommonExprs = 1150 RemoveCommonExpressionsFromUseBases(UsersToProcess, SE, L, TLI); 1151 1152 // Next, figure out what we can represent in the immediate fields of 1153 // instructions. If we can represent anything there, move it to the imm 1154 // fields of the BasedUsers. We do this so that it increases the commonality 1155 // of the remaining uses. 1156 unsigned NumPHI = 0; 1157 bool HasAddress = false; 1158 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) { 1159 // If the user is not in the current loop, this means it is using the exit 1160 // value of the IV. Do not put anything in the base, make sure it's all in 1161 // the immediate field to allow as much factoring as possible. 1162 if (!L->contains(UsersToProcess[i].Inst->getParent())) { 1163 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, 1164 UsersToProcess[i].Base); 1165 UsersToProcess[i].Base = 1166 SE->getIntegerSCEV(0, UsersToProcess[i].Base->getType()); 1167 } else { 1168 // Not all uses are outside the loop. 1169 AllUsesAreOutsideLoop = false; 1170 1171 // Addressing modes can be folded into loads and stores. Be careful that 1172 // the store is through the expression, not of the expression though. 1173 bool isPHI = false; 1174 bool isAddress = isAddressUse(UsersToProcess[i].Inst, 1175 UsersToProcess[i].OperandValToReplace); 1176 if (isa<PHINode>(UsersToProcess[i].Inst)) { 1177 isPHI = true; 1178 ++NumPHI; 1179 } 1180 1181 if (isAddress) 1182 HasAddress = true; 1183 1184 // If this use isn't an address, then not all uses are addresses. 1185 if (!isAddress && !isPHI) 1186 AllUsesAreAddresses = false; 1187 1188 MoveImmediateValues(TLI, UsersToProcess[i].Inst, UsersToProcess[i].Base, 1189 UsersToProcess[i].Imm, isAddress, L, SE); 1190 } 1191 } 1192 1193 // If one of the use is a PHI node and all other uses are addresses, still 1194 // allow iv reuse. Essentially we are trading one constant multiplication 1195 // for one fewer iv. 1196 if (NumPHI > 1) 1197 AllUsesAreAddresses = false; 1198 1199 // There are no in-loop address uses. 1200 if (AllUsesAreAddresses && (!HasAddress && !AllUsesAreOutsideLoop)) 1201 AllUsesAreAddresses = false; 1202 1203 return CommonExprs; 1204} 1205 1206/// ShouldUseFullStrengthReductionMode - Test whether full strength-reduction 1207/// is valid and profitable for the given set of users of a stride. In 1208/// full strength-reduction mode, all addresses at the current stride are 1209/// strength-reduced all the way down to pointer arithmetic. 1210/// 1211bool LoopStrengthReduce::ShouldUseFullStrengthReductionMode( 1212 const std::vector<BasedUser> &UsersToProcess, 1213 const Loop *L, 1214 bool AllUsesAreAddresses, 1215 const SCEV *Stride) { 1216 if (!EnableFullLSRMode) 1217 return false; 1218 1219 // The heuristics below aim to avoid increasing register pressure, but 1220 // fully strength-reducing all the addresses increases the number of 1221 // add instructions, so don't do this when optimizing for size. 1222 // TODO: If the loop is large, the savings due to simpler addresses 1223 // may oughtweight the costs of the extra increment instructions. 1224 if (L->getHeader()->getParent()->hasFnAttr(Attribute::OptimizeForSize)) 1225 return false; 1226 1227 // TODO: For now, don't do full strength reduction if there could 1228 // potentially be greater-stride multiples of the current stride 1229 // which could reuse the current stride IV. 1230 if (IU->StrideOrder.back() != Stride) 1231 return false; 1232 1233 // Iterate through the uses to find conditions that automatically rule out 1234 // full-lsr mode. 1235 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) { 1236 const SCEV *Base = UsersToProcess[i].Base; 1237 const SCEV *Imm = UsersToProcess[i].Imm; 1238 // If any users have a loop-variant component, they can't be fully 1239 // strength-reduced. 1240 if (Imm && !Imm->isLoopInvariant(L)) 1241 return false; 1242 // If there are to users with the same base and the difference between 1243 // the two Imm values can't be folded into the address, full 1244 // strength reduction would increase register pressure. 1245 do { 1246 const SCEV *CurImm = UsersToProcess[i].Imm; 1247 if ((CurImm || Imm) && CurImm != Imm) { 1248 if (!CurImm) CurImm = SE->getIntegerSCEV(0, Stride->getType()); 1249 if (!Imm) Imm = SE->getIntegerSCEV(0, Stride->getType()); 1250 const Instruction *Inst = UsersToProcess[i].Inst; 1251 const Type *AccessTy = getAccessType(Inst); 1252 const SCEV *Diff = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm); 1253 if (!Diff->isZero() && 1254 (!AllUsesAreAddresses || 1255 !fitsInAddressMode(Diff, AccessTy, TLI, /*HasBaseReg=*/true))) 1256 return false; 1257 } 1258 } while (++i != e && Base == UsersToProcess[i].Base); 1259 } 1260 1261 // If there's exactly one user in this stride, fully strength-reducing it 1262 // won't increase register pressure. If it's starting from a non-zero base, 1263 // it'll be simpler this way. 1264 if (UsersToProcess.size() == 1 && !UsersToProcess[0].Base->isZero()) 1265 return true; 1266 1267 // Otherwise, if there are any users in this stride that don't require 1268 // a register for their base, full strength-reduction will increase 1269 // register pressure. 1270 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) 1271 if (UsersToProcess[i].Base->isZero()) 1272 return false; 1273 1274 // Otherwise, go for it. 1275 return true; 1276} 1277 1278/// InsertAffinePhi Create and insert a PHI node for an induction variable 1279/// with the specified start and step values in the specified loop. 1280/// 1281/// If NegateStride is true, the stride should be negated by using a 1282/// subtract instead of an add. 1283/// 1284/// Return the created phi node. 1285/// 1286static PHINode *InsertAffinePhi(const SCEV *Start, const SCEV *Step, 1287 Instruction *IVIncInsertPt, 1288 const Loop *L, 1289 SCEVExpander &Rewriter) { 1290 assert(Start->isLoopInvariant(L) && "New PHI start is not loop invariant!"); 1291 assert(Step->isLoopInvariant(L) && "New PHI stride is not loop invariant!"); 1292 1293 BasicBlock *Header = L->getHeader(); 1294 BasicBlock *Preheader = L->getLoopPreheader(); 1295 BasicBlock *LatchBlock = L->getLoopLatch(); 1296 const Type *Ty = Start->getType(); 1297 Ty = Rewriter.SE.getEffectiveSCEVType(Ty); 1298 1299 PHINode *PN = PHINode::Create(Ty, "lsr.iv", Header->begin()); 1300 PN->addIncoming(Rewriter.expandCodeFor(Start, Ty, Preheader->getTerminator()), 1301 Preheader); 1302 1303 // If the stride is negative, insert a sub instead of an add for the 1304 // increment. 1305 bool isNegative = isNonConstantNegative(Step); 1306 const SCEV *IncAmount = Step; 1307 if (isNegative) 1308 IncAmount = Rewriter.SE.getNegativeSCEV(Step); 1309 1310 // Insert an add instruction right before the terminator corresponding 1311 // to the back-edge or just before the only use. The location is determined 1312 // by the caller and passed in as IVIncInsertPt. 1313 Value *StepV = Rewriter.expandCodeFor(IncAmount, Ty, 1314 Preheader->getTerminator()); 1315 Instruction *IncV; 1316 if (isNegative) { 1317 IncV = BinaryOperator::CreateSub(PN, StepV, "lsr.iv.next", 1318 IVIncInsertPt); 1319 } else { 1320 IncV = BinaryOperator::CreateAdd(PN, StepV, "lsr.iv.next", 1321 IVIncInsertPt); 1322 } 1323 if (!isa<ConstantInt>(StepV)) ++NumVariable; 1324 1325 PN->addIncoming(IncV, LatchBlock); 1326 1327 ++NumInserted; 1328 return PN; 1329} 1330 1331static void SortUsersToProcess(std::vector<BasedUser> &UsersToProcess) { 1332 // We want to emit code for users inside the loop first. To do this, we 1333 // rearrange BasedUser so that the entries at the end have 1334 // isUseOfPostIncrementedValue = false, because we pop off the end of the 1335 // vector (so we handle them first). 1336 std::partition(UsersToProcess.begin(), UsersToProcess.end(), 1337 PartitionByIsUseOfPostIncrementedValue); 1338 1339 // Sort this by base, so that things with the same base are handled 1340 // together. By partitioning first and stable-sorting later, we are 1341 // guaranteed that within each base we will pop off users from within the 1342 // loop before users outside of the loop with a particular base. 1343 // 1344 // We would like to use stable_sort here, but we can't. The problem is that 1345 // const SCEV *'s don't have a deterministic ordering w.r.t to each other, so 1346 // we don't have anything to do a '<' comparison on. Because we think the 1347 // number of uses is small, do a horrible bubble sort which just relies on 1348 // ==. 1349 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) { 1350 // Get a base value. 1351 const SCEV *Base = UsersToProcess[i].Base; 1352 1353 // Compact everything with this base to be consecutive with this one. 1354 for (unsigned j = i+1; j != e; ++j) { 1355 if (UsersToProcess[j].Base == Base) { 1356 std::swap(UsersToProcess[i+1], UsersToProcess[j]); 1357 ++i; 1358 } 1359 } 1360 } 1361} 1362 1363/// PrepareToStrengthReduceFully - Prepare to fully strength-reduce 1364/// UsersToProcess, meaning lowering addresses all the way down to direct 1365/// pointer arithmetic. 1366/// 1367void 1368LoopStrengthReduce::PrepareToStrengthReduceFully( 1369 std::vector<BasedUser> &UsersToProcess, 1370 const SCEV *Stride, 1371 const SCEV *CommonExprs, 1372 const Loop *L, 1373 SCEVExpander &PreheaderRewriter) { 1374 DOUT << " Fully reducing all users\n"; 1375 1376 // Rewrite the UsersToProcess records, creating a separate PHI for each 1377 // unique Base value. 1378 Instruction *IVIncInsertPt = L->getLoopLatch()->getTerminator(); 1379 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ) { 1380 // TODO: The uses are grouped by base, but not sorted. We arbitrarily 1381 // pick the first Imm value here to start with, and adjust it for the 1382 // other uses. 1383 const SCEV *Imm = UsersToProcess[i].Imm; 1384 const SCEV *Base = UsersToProcess[i].Base; 1385 const SCEV *Start = SE->getAddExpr(CommonExprs, Base, Imm); 1386 PHINode *Phi = InsertAffinePhi(Start, Stride, IVIncInsertPt, L, 1387 PreheaderRewriter); 1388 // Loop over all the users with the same base. 1389 do { 1390 UsersToProcess[i].Base = SE->getIntegerSCEV(0, Stride->getType()); 1391 UsersToProcess[i].Imm = SE->getMinusSCEV(UsersToProcess[i].Imm, Imm); 1392 UsersToProcess[i].Phi = Phi; 1393 assert(UsersToProcess[i].Imm->isLoopInvariant(L) && 1394 "ShouldUseFullStrengthReductionMode should reject this!"); 1395 } while (++i != e && Base == UsersToProcess[i].Base); 1396 } 1397} 1398 1399/// FindIVIncInsertPt - Return the location to insert the increment instruction. 1400/// If the only use if a use of postinc value, (must be the loop termination 1401/// condition), then insert it just before the use. 1402static Instruction *FindIVIncInsertPt(std::vector<BasedUser> &UsersToProcess, 1403 const Loop *L) { 1404 if (UsersToProcess.size() == 1 && 1405 UsersToProcess[0].isUseOfPostIncrementedValue && 1406 L->contains(UsersToProcess[0].Inst->getParent())) 1407 return UsersToProcess[0].Inst; 1408 return L->getLoopLatch()->getTerminator(); 1409} 1410 1411/// PrepareToStrengthReduceWithNewPhi - Insert a new induction variable for the 1412/// given users to share. 1413/// 1414void 1415LoopStrengthReduce::PrepareToStrengthReduceWithNewPhi( 1416 std::vector<BasedUser> &UsersToProcess, 1417 const SCEV *Stride, 1418 const SCEV *CommonExprs, 1419 Value *CommonBaseV, 1420 Instruction *IVIncInsertPt, 1421 const Loop *L, 1422 SCEVExpander &PreheaderRewriter) { 1423 DOUT << " Inserting new PHI:\n"; 1424 1425 PHINode *Phi = InsertAffinePhi(SE->getUnknown(CommonBaseV), 1426 Stride, IVIncInsertPt, L, 1427 PreheaderRewriter); 1428 1429 // Remember this in case a later stride is multiple of this. 1430 IVsByStride[Stride].addIV(Stride, CommonExprs, Phi); 1431 1432 // All the users will share this new IV. 1433 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) 1434 UsersToProcess[i].Phi = Phi; 1435 1436 DOUT << " IV="; 1437 DEBUG(WriteAsOperand(*DOUT, Phi, /*PrintType=*/false)); 1438 DOUT << "\n"; 1439} 1440 1441/// PrepareToStrengthReduceFromSmallerStride - Prepare for the given users to 1442/// reuse an induction variable with a stride that is a factor of the current 1443/// induction variable. 1444/// 1445void 1446LoopStrengthReduce::PrepareToStrengthReduceFromSmallerStride( 1447 std::vector<BasedUser> &UsersToProcess, 1448 Value *CommonBaseV, 1449 const IVExpr &ReuseIV, 1450 Instruction *PreInsertPt) { 1451 DOUT << " Rewriting in terms of existing IV of STRIDE " << *ReuseIV.Stride 1452 << " and BASE " << *ReuseIV.Base << "\n"; 1453 1454 // All the users will share the reused IV. 1455 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) 1456 UsersToProcess[i].Phi = ReuseIV.PHI; 1457 1458 Constant *C = dyn_cast<Constant>(CommonBaseV); 1459 if (C && 1460 (!C->isNullValue() && 1461 !fitsInAddressMode(SE->getUnknown(CommonBaseV), CommonBaseV->getType(), 1462 TLI, false))) 1463 // We want the common base emitted into the preheader! This is just 1464 // using cast as a copy so BitCast (no-op cast) is appropriate 1465 CommonBaseV = new BitCastInst(CommonBaseV, CommonBaseV->getType(), 1466 "commonbase", PreInsertPt); 1467} 1468 1469static bool IsImmFoldedIntoAddrMode(GlobalValue *GV, int64_t Offset, 1470 const Type *AccessTy, 1471 std::vector<BasedUser> &UsersToProcess, 1472 const TargetLowering *TLI) { 1473 SmallVector<Instruction*, 16> AddrModeInsts; 1474 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) { 1475 if (UsersToProcess[i].isUseOfPostIncrementedValue) 1476 continue; 1477 ExtAddrMode AddrMode = 1478 AddressingModeMatcher::Match(UsersToProcess[i].OperandValToReplace, 1479 AccessTy, UsersToProcess[i].Inst, 1480 AddrModeInsts, *TLI); 1481 if (GV && GV != AddrMode.BaseGV) 1482 return false; 1483 if (Offset && !AddrMode.BaseOffs) 1484 // FIXME: How to accurate check it's immediate offset is folded. 1485 return false; 1486 AddrModeInsts.clear(); 1487 } 1488 return true; 1489} 1490 1491/// StrengthReduceStridedIVUsers - Strength reduce all of the users of a single 1492/// stride of IV. All of the users may have different starting values, and this 1493/// may not be the only stride. 1494void LoopStrengthReduce::StrengthReduceStridedIVUsers(const SCEV *const &Stride, 1495 IVUsersOfOneStride &Uses, 1496 Loop *L) { 1497 // If all the users are moved to another stride, then there is nothing to do. 1498 if (Uses.Users.empty()) 1499 return; 1500 1501 // Keep track if every use in UsersToProcess is an address. If they all are, 1502 // we may be able to rewrite the entire collection of them in terms of a 1503 // smaller-stride IV. 1504 bool AllUsesAreAddresses = true; 1505 1506 // Keep track if every use of a single stride is outside the loop. If so, 1507 // we want to be more aggressive about reusing a smaller-stride IV; a 1508 // multiply outside the loop is better than another IV inside. Well, usually. 1509 bool AllUsesAreOutsideLoop = true; 1510 1511 // Transform our list of users and offsets to a bit more complex table. In 1512 // this new vector, each 'BasedUser' contains 'Base' the base of the 1513 // strided accessas well as the old information from Uses. We progressively 1514 // move information from the Base field to the Imm field, until we eventually 1515 // have the full access expression to rewrite the use. 1516 std::vector<BasedUser> UsersToProcess; 1517 const SCEV *CommonExprs = CollectIVUsers(Stride, Uses, L, AllUsesAreAddresses, 1518 AllUsesAreOutsideLoop, 1519 UsersToProcess); 1520 1521 // Sort the UsersToProcess array so that users with common bases are 1522 // next to each other. 1523 SortUsersToProcess(UsersToProcess); 1524 1525 // If we managed to find some expressions in common, we'll need to carry 1526 // their value in a register and add it in for each use. This will take up 1527 // a register operand, which potentially restricts what stride values are 1528 // valid. 1529 bool HaveCommonExprs = !CommonExprs->isZero(); 1530 const Type *ReplacedTy = CommonExprs->getType(); 1531 1532 // If all uses are addresses, consider sinking the immediate part of the 1533 // common expression back into uses if they can fit in the immediate fields. 1534 if (TLI && HaveCommonExprs && AllUsesAreAddresses) { 1535 const SCEV *NewCommon = CommonExprs; 1536 const SCEV *Imm = SE->getIntegerSCEV(0, ReplacedTy); 1537 MoveImmediateValues(TLI, Type::VoidTy, NewCommon, Imm, true, L, SE); 1538 if (!Imm->isZero()) { 1539 bool DoSink = true; 1540 1541 // If the immediate part of the common expression is a GV, check if it's 1542 // possible to fold it into the target addressing mode. 1543 GlobalValue *GV = 0; 1544 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(Imm)) 1545 GV = dyn_cast<GlobalValue>(SU->getValue()); 1546 int64_t Offset = 0; 1547 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Imm)) 1548 Offset = SC->getValue()->getSExtValue(); 1549 if (GV || Offset) 1550 // Pass VoidTy as the AccessTy to be conservative, because 1551 // there could be multiple access types among all the uses. 1552 DoSink = IsImmFoldedIntoAddrMode(GV, Offset, Type::VoidTy, 1553 UsersToProcess, TLI); 1554 1555 if (DoSink) { 1556 DOUT << " Sinking " << *Imm << " back down into uses\n"; 1557 for (unsigned i = 0, e = UsersToProcess.size(); i != e; ++i) 1558 UsersToProcess[i].Imm = SE->getAddExpr(UsersToProcess[i].Imm, Imm); 1559 CommonExprs = NewCommon; 1560 HaveCommonExprs = !CommonExprs->isZero(); 1561 ++NumImmSunk; 1562 } 1563 } 1564 } 1565 1566 // Now that we know what we need to do, insert the PHI node itself. 1567 // 1568 DOUT << "LSR: Examining IVs of TYPE " << *ReplacedTy << " of STRIDE " 1569 << *Stride << ":\n" 1570 << " Common base: " << *CommonExprs << "\n"; 1571 1572 SCEVExpander Rewriter(*SE); 1573 SCEVExpander PreheaderRewriter(*SE); 1574 1575 BasicBlock *Preheader = L->getLoopPreheader(); 1576 Instruction *PreInsertPt = Preheader->getTerminator(); 1577 BasicBlock *LatchBlock = L->getLoopLatch(); 1578 Instruction *IVIncInsertPt = LatchBlock->getTerminator(); 1579 1580 Value *CommonBaseV = Constant::getNullValue(ReplacedTy); 1581 1582 const SCEV *RewriteFactor = SE->getIntegerSCEV(0, ReplacedTy); 1583 IVExpr ReuseIV(SE->getIntegerSCEV(0, Type::Int32Ty), 1584 SE->getIntegerSCEV(0, Type::Int32Ty), 1585 0); 1586 1587 /// Choose a strength-reduction strategy and prepare for it by creating 1588 /// the necessary PHIs and adjusting the bookkeeping. 1589 if (ShouldUseFullStrengthReductionMode(UsersToProcess, L, 1590 AllUsesAreAddresses, Stride)) { 1591 PrepareToStrengthReduceFully(UsersToProcess, Stride, CommonExprs, L, 1592 PreheaderRewriter); 1593 } else { 1594 // Emit the initial base value into the loop preheader. 1595 CommonBaseV = PreheaderRewriter.expandCodeFor(CommonExprs, ReplacedTy, 1596 PreInsertPt); 1597 1598 // If all uses are addresses, check if it is possible to reuse an IV. The 1599 // new IV must have a stride that is a multiple of the old stride; the 1600 // multiple must be a number that can be encoded in the scale field of the 1601 // target addressing mode; and we must have a valid instruction after this 1602 // substitution, including the immediate field, if any. 1603 RewriteFactor = CheckForIVReuse(HaveCommonExprs, AllUsesAreAddresses, 1604 AllUsesAreOutsideLoop, 1605 Stride, ReuseIV, ReplacedTy, 1606 UsersToProcess); 1607 if (!RewriteFactor->isZero()) 1608 PrepareToStrengthReduceFromSmallerStride(UsersToProcess, CommonBaseV, 1609 ReuseIV, PreInsertPt); 1610 else { 1611 IVIncInsertPt = FindIVIncInsertPt(UsersToProcess, L); 1612 PrepareToStrengthReduceWithNewPhi(UsersToProcess, Stride, CommonExprs, 1613 CommonBaseV, IVIncInsertPt, 1614 L, PreheaderRewriter); 1615 } 1616 } 1617 1618 // Process all the users now, replacing their strided uses with 1619 // strength-reduced forms. This outer loop handles all bases, the inner 1620 // loop handles all users of a particular base. 1621 while (!UsersToProcess.empty()) { 1622 const SCEV *Base = UsersToProcess.back().Base; 1623 Instruction *Inst = UsersToProcess.back().Inst; 1624 1625 // Emit the code for Base into the preheader. 1626 Value *BaseV = 0; 1627 if (!Base->isZero()) { 1628 BaseV = PreheaderRewriter.expandCodeFor(Base, 0, PreInsertPt); 1629 1630 DOUT << " INSERTING code for BASE = " << *Base << ":"; 1631 if (BaseV->hasName()) 1632 DOUT << " Result value name = %" << BaseV->getNameStr(); 1633 DOUT << "\n"; 1634 1635 // If BaseV is a non-zero constant, make sure that it gets inserted into 1636 // the preheader, instead of being forward substituted into the uses. We 1637 // do this by forcing a BitCast (noop cast) to be inserted into the 1638 // preheader in this case. 1639 if (!fitsInAddressMode(Base, getAccessType(Inst), TLI, false) && 1640 isa<Constant>(BaseV)) { 1641 // We want this constant emitted into the preheader! This is just 1642 // using cast as a copy so BitCast (no-op cast) is appropriate 1643 BaseV = new BitCastInst(BaseV, BaseV->getType(), "preheaderinsert", 1644 PreInsertPt); 1645 } 1646 } 1647 1648 // Emit the code to add the immediate offset to the Phi value, just before 1649 // the instructions that we identified as using this stride and base. 1650 do { 1651 // FIXME: Use emitted users to emit other users. 1652 BasedUser &User = UsersToProcess.back(); 1653 1654 DOUT << " Examining "; 1655 if (User.isUseOfPostIncrementedValue) 1656 DOUT << "postinc"; 1657 else 1658 DOUT << "preinc"; 1659 DOUT << " use "; 1660 DEBUG(WriteAsOperand(*DOUT, UsersToProcess.back().OperandValToReplace, 1661 /*PrintType=*/false)); 1662 DOUT << " in Inst: " << *(User.Inst); 1663 1664 // If this instruction wants to use the post-incremented value, move it 1665 // after the post-inc and use its value instead of the PHI. 1666 Value *RewriteOp = User.Phi; 1667 if (User.isUseOfPostIncrementedValue) { 1668 RewriteOp = User.Phi->getIncomingValueForBlock(LatchBlock); 1669 // If this user is in the loop, make sure it is the last thing in the 1670 // loop to ensure it is dominated by the increment. In case it's the 1671 // only use of the iv, the increment instruction is already before the 1672 // use. 1673 if (L->contains(User.Inst->getParent()) && User.Inst != IVIncInsertPt) 1674 User.Inst->moveBefore(IVIncInsertPt); 1675 } 1676 1677 const SCEV *RewriteExpr = SE->getUnknown(RewriteOp); 1678 1679 if (SE->getEffectiveSCEVType(RewriteOp->getType()) != 1680 SE->getEffectiveSCEVType(ReplacedTy)) { 1681 assert(SE->getTypeSizeInBits(RewriteOp->getType()) > 1682 SE->getTypeSizeInBits(ReplacedTy) && 1683 "Unexpected widening cast!"); 1684 RewriteExpr = SE->getTruncateExpr(RewriteExpr, ReplacedTy); 1685 } 1686 1687 // If we had to insert new instructions for RewriteOp, we have to 1688 // consider that they may not have been able to end up immediately 1689 // next to RewriteOp, because non-PHI instructions may never precede 1690 // PHI instructions in a block. In this case, remember where the last 1691 // instruction was inserted so that if we're replacing a different 1692 // PHI node, we can use the later point to expand the final 1693 // RewriteExpr. 1694 Instruction *NewBasePt = dyn_cast<Instruction>(RewriteOp); 1695 if (RewriteOp == User.Phi) NewBasePt = 0; 1696 1697 // Clear the SCEVExpander's expression map so that we are guaranteed 1698 // to have the code emitted where we expect it. 1699 Rewriter.clear(); 1700 1701 // If we are reusing the iv, then it must be multiplied by a constant 1702 // factor to take advantage of the addressing mode scale component. 1703 if (!RewriteFactor->isZero()) { 1704 // If we're reusing an IV with a nonzero base (currently this happens 1705 // only when all reuses are outside the loop) subtract that base here. 1706 // The base has been used to initialize the PHI node but we don't want 1707 // it here. 1708 if (!ReuseIV.Base->isZero()) { 1709 const SCEV *typedBase = ReuseIV.Base; 1710 if (SE->getEffectiveSCEVType(RewriteExpr->getType()) != 1711 SE->getEffectiveSCEVType(ReuseIV.Base->getType())) { 1712 // It's possible the original IV is a larger type than the new IV, 1713 // in which case we have to truncate the Base. We checked in 1714 // RequiresTypeConversion that this is valid. 1715 assert(SE->getTypeSizeInBits(RewriteExpr->getType()) < 1716 SE->getTypeSizeInBits(ReuseIV.Base->getType()) && 1717 "Unexpected lengthening conversion!"); 1718 typedBase = SE->getTruncateExpr(ReuseIV.Base, 1719 RewriteExpr->getType()); 1720 } 1721 RewriteExpr = SE->getMinusSCEV(RewriteExpr, typedBase); 1722 } 1723 1724 // Multiply old variable, with base removed, by new scale factor. 1725 RewriteExpr = SE->getMulExpr(RewriteFactor, 1726 RewriteExpr); 1727 1728 // The common base is emitted in the loop preheader. But since we 1729 // are reusing an IV, it has not been used to initialize the PHI node. 1730 // Add it to the expression used to rewrite the uses. 1731 // When this use is outside the loop, we earlier subtracted the 1732 // common base, and are adding it back here. Use the same expression 1733 // as before, rather than CommonBaseV, so DAGCombiner will zap it. 1734 if (!CommonExprs->isZero()) { 1735 if (L->contains(User.Inst->getParent())) 1736 RewriteExpr = SE->getAddExpr(RewriteExpr, 1737 SE->getUnknown(CommonBaseV)); 1738 else 1739 RewriteExpr = SE->getAddExpr(RewriteExpr, CommonExprs); 1740 } 1741 } 1742 1743 // Now that we know what we need to do, insert code before User for the 1744 // immediate and any loop-variant expressions. 1745 if (BaseV) 1746 // Add BaseV to the PHI value if needed. 1747 RewriteExpr = SE->getAddExpr(RewriteExpr, SE->getUnknown(BaseV)); 1748 1749 User.RewriteInstructionToUseNewBase(RewriteExpr, NewBasePt, 1750 Rewriter, L, this, *LI, 1751 DeadInsts); 1752 1753 // Mark old value we replaced as possibly dead, so that it is eliminated 1754 // if we just replaced the last use of that value. 1755 DeadInsts.push_back(User.OperandValToReplace); 1756 1757 UsersToProcess.pop_back(); 1758 ++NumReduced; 1759 1760 // If there are any more users to process with the same base, process them 1761 // now. We sorted by base above, so we just have to check the last elt. 1762 } while (!UsersToProcess.empty() && UsersToProcess.back().Base == Base); 1763 // TODO: Next, find out which base index is the most common, pull it out. 1764 } 1765 1766 // IMPORTANT TODO: Figure out how to partition the IV's with this stride, but 1767 // different starting values, into different PHIs. 1768} 1769 1770/// FindIVUserForCond - If Cond has an operand that is an expression of an IV, 1771/// set the IV user and stride information and return true, otherwise return 1772/// false. 1773bool LoopStrengthReduce::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse, 1774 const SCEV *const * &CondStride) { 1775 for (unsigned Stride = 0, e = IU->StrideOrder.size(); 1776 Stride != e && !CondUse; ++Stride) { 1777 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = 1778 IU->IVUsesByStride.find(IU->StrideOrder[Stride]); 1779 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!"); 1780 1781 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(), 1782 E = SI->second->Users.end(); UI != E; ++UI) 1783 if (UI->getUser() == Cond) { 1784 // NOTE: we could handle setcc instructions with multiple uses here, but 1785 // InstCombine does it as well for simple uses, it's not clear that it 1786 // occurs enough in real life to handle. 1787 CondUse = UI; 1788 CondStride = &SI->first; 1789 return true; 1790 } 1791 } 1792 return false; 1793} 1794 1795namespace { 1796 // Constant strides come first which in turns are sorted by their absolute 1797 // values. If absolute values are the same, then positive strides comes first. 1798 // e.g. 1799 // 4, -1, X, 1, 2 ==> 1, -1, 2, 4, X 1800 struct StrideCompare { 1801 const ScalarEvolution *SE; 1802 explicit StrideCompare(const ScalarEvolution *se) : SE(se) {} 1803 1804 bool operator()(const SCEV *const &LHS, const SCEV *const &RHS) { 1805 const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS); 1806 const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS); 1807 if (LHSC && RHSC) { 1808 int64_t LV = LHSC->getValue()->getSExtValue(); 1809 int64_t RV = RHSC->getValue()->getSExtValue(); 1810 uint64_t ALV = (LV < 0) ? -LV : LV; 1811 uint64_t ARV = (RV < 0) ? -RV : RV; 1812 if (ALV == ARV) { 1813 if (LV != RV) 1814 return LV > RV; 1815 } else { 1816 return ALV < ARV; 1817 } 1818 1819 // If it's the same value but different type, sort by bit width so 1820 // that we emit larger induction variables before smaller 1821 // ones, letting the smaller be re-written in terms of larger ones. 1822 return SE->getTypeSizeInBits(RHS->getType()) < 1823 SE->getTypeSizeInBits(LHS->getType()); 1824 } 1825 return LHSC && !RHSC; 1826 } 1827 }; 1828} 1829 1830/// ChangeCompareStride - If a loop termination compare instruction is the 1831/// only use of its stride, and the compaison is against a constant value, 1832/// try eliminate the stride by moving the compare instruction to another 1833/// stride and change its constant operand accordingly. e.g. 1834/// 1835/// loop: 1836/// ... 1837/// v1 = v1 + 3 1838/// v2 = v2 + 1 1839/// if (v2 < 10) goto loop 1840/// => 1841/// loop: 1842/// ... 1843/// v1 = v1 + 3 1844/// if (v1 < 30) goto loop 1845ICmpInst *LoopStrengthReduce::ChangeCompareStride(Loop *L, ICmpInst *Cond, 1846 IVStrideUse* &CondUse, 1847 const SCEV *const* &CondStride) { 1848 // If there's only one stride in the loop, there's nothing to do here. 1849 if (IU->StrideOrder.size() < 2) 1850 return Cond; 1851 // If there are other users of the condition's stride, don't bother 1852 // trying to change the condition because the stride will still 1853 // remain. 1854 std::map<const SCEV *, IVUsersOfOneStride *>::iterator I = 1855 IU->IVUsesByStride.find(*CondStride); 1856 if (I == IU->IVUsesByStride.end() || 1857 I->second->Users.size() != 1) 1858 return Cond; 1859 // Only handle constant strides for now. 1860 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride); 1861 if (!SC) return Cond; 1862 1863 ICmpInst::Predicate Predicate = Cond->getPredicate(); 1864 int64_t CmpSSInt = SC->getValue()->getSExtValue(); 1865 unsigned BitWidth = SE->getTypeSizeInBits((*CondStride)->getType()); 1866 uint64_t SignBit = 1ULL << (BitWidth-1); 1867 const Type *CmpTy = Cond->getOperand(0)->getType(); 1868 const Type *NewCmpTy = NULL; 1869 unsigned TyBits = SE->getTypeSizeInBits(CmpTy); 1870 unsigned NewTyBits = 0; 1871 const SCEV **NewStride = NULL; 1872 Value *NewCmpLHS = NULL; 1873 Value *NewCmpRHS = NULL; 1874 int64_t Scale = 1; 1875 const SCEV *NewOffset = SE->getIntegerSCEV(0, CmpTy); 1876 1877 if (ConstantInt *C = dyn_cast<ConstantInt>(Cond->getOperand(1))) { 1878 int64_t CmpVal = C->getValue().getSExtValue(); 1879 1880 // Check stride constant and the comparision constant signs to detect 1881 // overflow. 1882 if ((CmpVal & SignBit) != (CmpSSInt & SignBit)) 1883 return Cond; 1884 1885 // Look for a suitable stride / iv as replacement. 1886 for (unsigned i = 0, e = IU->StrideOrder.size(); i != e; ++i) { 1887 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = 1888 IU->IVUsesByStride.find(IU->StrideOrder[i]); 1889 if (!isa<SCEVConstant>(SI->first)) 1890 continue; 1891 int64_t SSInt = cast<SCEVConstant>(SI->first)->getValue()->getSExtValue(); 1892 if (SSInt == CmpSSInt || 1893 abs64(SSInt) < abs64(CmpSSInt) || 1894 (SSInt % CmpSSInt) != 0) 1895 continue; 1896 1897 Scale = SSInt / CmpSSInt; 1898 int64_t NewCmpVal = CmpVal * Scale; 1899 APInt Mul = APInt(BitWidth*2, CmpVal, true); 1900 Mul = Mul * APInt(BitWidth*2, Scale, true); 1901 // Check for overflow. 1902 if (!Mul.isSignedIntN(BitWidth)) 1903 continue; 1904 // Check for overflow in the stride's type too. 1905 if (!Mul.isSignedIntN(SE->getTypeSizeInBits(SI->first->getType()))) 1906 continue; 1907 1908 // Watch out for overflow. 1909 if (ICmpInst::isSignedPredicate(Predicate) && 1910 (CmpVal & SignBit) != (NewCmpVal & SignBit)) 1911 continue; 1912 1913 if (NewCmpVal == CmpVal) 1914 continue; 1915 // Pick the best iv to use trying to avoid a cast. 1916 NewCmpLHS = NULL; 1917 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(), 1918 E = SI->second->Users.end(); UI != E; ++UI) { 1919 Value *Op = UI->getOperandValToReplace(); 1920 1921 // If the IVStrideUse implies a cast, check for an actual cast which 1922 // can be used to find the original IV expression. 1923 if (SE->getEffectiveSCEVType(Op->getType()) != 1924 SE->getEffectiveSCEVType(SI->first->getType())) { 1925 CastInst *CI = dyn_cast<CastInst>(Op); 1926 // If it's not a simple cast, it's complicated. 1927 if (!CI) 1928 continue; 1929 // If it's a cast from a type other than the stride type, 1930 // it's complicated. 1931 if (CI->getOperand(0)->getType() != SI->first->getType()) 1932 continue; 1933 // Ok, we found the IV expression in the stride's type. 1934 Op = CI->getOperand(0); 1935 } 1936 1937 NewCmpLHS = Op; 1938 if (NewCmpLHS->getType() == CmpTy) 1939 break; 1940 } 1941 if (!NewCmpLHS) 1942 continue; 1943 1944 NewCmpTy = NewCmpLHS->getType(); 1945 NewTyBits = SE->getTypeSizeInBits(NewCmpTy); 1946 const Type *NewCmpIntTy = IntegerType::get(NewTyBits); 1947 if (RequiresTypeConversion(NewCmpTy, CmpTy)) { 1948 // Check if it is possible to rewrite it using 1949 // an iv / stride of a smaller integer type. 1950 unsigned Bits = NewTyBits; 1951 if (ICmpInst::isSignedPredicate(Predicate)) 1952 --Bits; 1953 uint64_t Mask = (1ULL << Bits) - 1; 1954 if (((uint64_t)NewCmpVal & Mask) != (uint64_t)NewCmpVal) 1955 continue; 1956 } 1957 1958 // Don't rewrite if use offset is non-constant and the new type is 1959 // of a different type. 1960 // FIXME: too conservative? 1961 if (NewTyBits != TyBits && !isa<SCEVConstant>(CondUse->getOffset())) 1962 continue; 1963 1964 bool AllUsesAreAddresses = true; 1965 bool AllUsesAreOutsideLoop = true; 1966 std::vector<BasedUser> UsersToProcess; 1967 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L, 1968 AllUsesAreAddresses, 1969 AllUsesAreOutsideLoop, 1970 UsersToProcess); 1971 // Avoid rewriting the compare instruction with an iv of new stride 1972 // if it's likely the new stride uses will be rewritten using the 1973 // stride of the compare instruction. 1974 if (AllUsesAreAddresses && 1975 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) 1976 continue; 1977 1978 // Avoid rewriting the compare instruction with an iv which has 1979 // implicit extension or truncation built into it. 1980 // TODO: This is over-conservative. 1981 if (SE->getTypeSizeInBits(CondUse->getOffset()->getType()) != TyBits) 1982 continue; 1983 1984 // If scale is negative, use swapped predicate unless it's testing 1985 // for equality. 1986 if (Scale < 0 && !Cond->isEquality()) 1987 Predicate = ICmpInst::getSwappedPredicate(Predicate); 1988 1989 NewStride = &IU->StrideOrder[i]; 1990 if (!isa<PointerType>(NewCmpTy)) 1991 NewCmpRHS = ConstantInt::get(NewCmpTy, NewCmpVal); 1992 else { 1993 Constant *CI = ConstantInt::get(NewCmpIntTy, NewCmpVal); 1994 NewCmpRHS = ConstantExpr::getIntToPtr(CI, NewCmpTy); 1995 } 1996 NewOffset = TyBits == NewTyBits 1997 ? SE->getMulExpr(CondUse->getOffset(), 1998 SE->getConstant(CmpTy, Scale)) 1999 : SE->getConstant(NewCmpIntTy, 2000 cast<SCEVConstant>(CondUse->getOffset())->getValue() 2001 ->getSExtValue()*Scale); 2002 break; 2003 } 2004 } 2005 2006 // Forgo this transformation if it the increment happens to be 2007 // unfortunately positioned after the condition, and the condition 2008 // has multiple uses which prevent it from being moved immediately 2009 // before the branch. See 2010 // test/Transforms/LoopStrengthReduce/change-compare-stride-trickiness-*.ll 2011 // for an example of this situation. 2012 if (!Cond->hasOneUse()) { 2013 for (BasicBlock::iterator I = Cond, E = Cond->getParent()->end(); 2014 I != E; ++I) 2015 if (I == NewCmpLHS) 2016 return Cond; 2017 } 2018 2019 if (NewCmpRHS) { 2020 // Create a new compare instruction using new stride / iv. 2021 ICmpInst *OldCond = Cond; 2022 // Insert new compare instruction. 2023 Cond = new ICmpInst(OldCond, Predicate, NewCmpLHS, NewCmpRHS, 2024 L->getHeader()->getName() + ".termcond"); 2025 2026 // Remove the old compare instruction. The old indvar is probably dead too. 2027 DeadInsts.push_back(CondUse->getOperandValToReplace()); 2028 OldCond->replaceAllUsesWith(Cond); 2029 OldCond->eraseFromParent(); 2030 2031 IU->IVUsesByStride[*NewStride]->addUser(NewOffset, Cond, NewCmpLHS); 2032 CondUse = &IU->IVUsesByStride[*NewStride]->Users.back(); 2033 CondStride = NewStride; 2034 ++NumEliminated; 2035 Changed = true; 2036 } 2037 2038 return Cond; 2039} 2040 2041/// OptimizeMax - Rewrite the loop's terminating condition if it uses 2042/// a max computation. 2043/// 2044/// This is a narrow solution to a specific, but acute, problem. For loops 2045/// like this: 2046/// 2047/// i = 0; 2048/// do { 2049/// p[i] = 0.0; 2050/// } while (++i < n); 2051/// 2052/// the trip count isn't just 'n', because 'n' might not be positive. And 2053/// unfortunately this can come up even for loops where the user didn't use 2054/// a C do-while loop. For example, seemingly well-behaved top-test loops 2055/// will commonly be lowered like this: 2056// 2057/// if (n > 0) { 2058/// i = 0; 2059/// do { 2060/// p[i] = 0.0; 2061/// } while (++i < n); 2062/// } 2063/// 2064/// and then it's possible for subsequent optimization to obscure the if 2065/// test in such a way that indvars can't find it. 2066/// 2067/// When indvars can't find the if test in loops like this, it creates a 2068/// max expression, which allows it to give the loop a canonical 2069/// induction variable: 2070/// 2071/// i = 0; 2072/// max = n < 1 ? 1 : n; 2073/// do { 2074/// p[i] = 0.0; 2075/// } while (++i != max); 2076/// 2077/// Canonical induction variables are necessary because the loop passes 2078/// are designed around them. The most obvious example of this is the 2079/// LoopInfo analysis, which doesn't remember trip count values. It 2080/// expects to be able to rediscover the trip count each time it is 2081/// needed, and it does this using a simple analyis that only succeeds if 2082/// the loop has a canonical induction variable. 2083/// 2084/// However, when it comes time to generate code, the maximum operation 2085/// can be quite costly, especially if it's inside of an outer loop. 2086/// 2087/// This function solves this problem by detecting this type of loop and 2088/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting 2089/// the instructions for the maximum computation. 2090/// 2091ICmpInst *LoopStrengthReduce::OptimizeMax(Loop *L, ICmpInst *Cond, 2092 IVStrideUse* &CondUse) { 2093 // Check that the loop matches the pattern we're looking for. 2094 if (Cond->getPredicate() != CmpInst::ICMP_EQ && 2095 Cond->getPredicate() != CmpInst::ICMP_NE) 2096 return Cond; 2097 2098 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1)); 2099 if (!Sel || !Sel->hasOneUse()) return Cond; 2100 2101 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 2102 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 2103 return Cond; 2104 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType()); 2105 2106 // Add one to the backedge-taken count to get the trip count. 2107 const SCEV *IterationCount = SE->getAddExpr(BackedgeTakenCount, One); 2108 2109 // Check for a max calculation that matches the pattern. 2110 if (!isa<SCEVSMaxExpr>(IterationCount) && !isa<SCEVUMaxExpr>(IterationCount)) 2111 return Cond; 2112 const SCEVNAryExpr *Max = cast<SCEVNAryExpr>(IterationCount); 2113 if (Max != SE->getSCEV(Sel)) return Cond; 2114 2115 // To handle a max with more than two operands, this optimization would 2116 // require additional checking and setup. 2117 if (Max->getNumOperands() != 2) 2118 return Cond; 2119 2120 const SCEV *MaxLHS = Max->getOperand(0); 2121 const SCEV *MaxRHS = Max->getOperand(1); 2122 if (!MaxLHS || MaxLHS != One) return Cond; 2123 2124 // Check the relevant induction variable for conformance to 2125 // the pattern. 2126 const SCEV *IV = SE->getSCEV(Cond->getOperand(0)); 2127 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); 2128 if (!AR || !AR->isAffine() || 2129 AR->getStart() != One || 2130 AR->getStepRecurrence(*SE) != One) 2131 return Cond; 2132 2133 assert(AR->getLoop() == L && 2134 "Loop condition operand is an addrec in a different loop!"); 2135 2136 // Check the right operand of the select, and remember it, as it will 2137 // be used in the new comparison instruction. 2138 Value *NewRHS = 0; 2139 if (SE->getSCEV(Sel->getOperand(1)) == MaxRHS) 2140 NewRHS = Sel->getOperand(1); 2141 else if (SE->getSCEV(Sel->getOperand(2)) == MaxRHS) 2142 NewRHS = Sel->getOperand(2); 2143 if (!NewRHS) return Cond; 2144 2145 // Determine the new comparison opcode. It may be signed or unsigned, 2146 // and the original comparison may be either equality or inequality. 2147 CmpInst::Predicate Pred = 2148 isa<SCEVSMaxExpr>(Max) ? CmpInst::ICMP_SLT : CmpInst::ICMP_ULT; 2149 if (Cond->getPredicate() == CmpInst::ICMP_EQ) 2150 Pred = CmpInst::getInversePredicate(Pred); 2151 2152 // Ok, everything looks ok to change the condition into an SLT or SGE and 2153 // delete the max calculation. 2154 ICmpInst *NewCond = 2155 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp"); 2156 2157 // Delete the max calculation instructions. 2158 Cond->replaceAllUsesWith(NewCond); 2159 CondUse->setUser(NewCond); 2160 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0)); 2161 Cond->eraseFromParent(); 2162 Sel->eraseFromParent(); 2163 if (Cmp->use_empty()) 2164 Cmp->eraseFromParent(); 2165 return NewCond; 2166} 2167 2168/// OptimizeShadowIV - If IV is used in a int-to-float cast 2169/// inside the loop then try to eliminate the cast opeation. 2170void LoopStrengthReduce::OptimizeShadowIV(Loop *L) { 2171 2172 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 2173 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 2174 return; 2175 2176 for (unsigned Stride = 0, e = IU->StrideOrder.size(); Stride != e; 2177 ++Stride) { 2178 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = 2179 IU->IVUsesByStride.find(IU->StrideOrder[Stride]); 2180 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!"); 2181 if (!isa<SCEVConstant>(SI->first)) 2182 continue; 2183 2184 for (ilist<IVStrideUse>::iterator UI = SI->second->Users.begin(), 2185 E = SI->second->Users.end(); UI != E; /* empty */) { 2186 ilist<IVStrideUse>::iterator CandidateUI = UI; 2187 ++UI; 2188 Instruction *ShadowUse = CandidateUI->getUser(); 2189 const Type *DestTy = NULL; 2190 2191 /* If shadow use is a int->float cast then insert a second IV 2192 to eliminate this cast. 2193 2194 for (unsigned i = 0; i < n; ++i) 2195 foo((double)i); 2196 2197 is transformed into 2198 2199 double d = 0.0; 2200 for (unsigned i = 0; i < n; ++i, ++d) 2201 foo(d); 2202 */ 2203 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) 2204 DestTy = UCast->getDestTy(); 2205 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) 2206 DestTy = SCast->getDestTy(); 2207 if (!DestTy) continue; 2208 2209 if (TLI) { 2210 // If target does not support DestTy natively then do not apply 2211 // this transformation. 2212 EVT DVT = TLI->getValueType(DestTy); 2213 if (!TLI->isTypeLegal(DVT)) continue; 2214 } 2215 2216 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); 2217 if (!PH) continue; 2218 if (PH->getNumIncomingValues() != 2) continue; 2219 2220 const Type *SrcTy = PH->getType(); 2221 int Mantissa = DestTy->getFPMantissaWidth(); 2222 if (Mantissa == -1) continue; 2223 if ((int)SE->getTypeSizeInBits(SrcTy) > Mantissa) 2224 continue; 2225 2226 unsigned Entry, Latch; 2227 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { 2228 Entry = 0; 2229 Latch = 1; 2230 } else { 2231 Entry = 1; 2232 Latch = 0; 2233 } 2234 2235 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); 2236 if (!Init) continue; 2237 Constant *NewInit = ConstantFP::get(DestTy, Init->getZExtValue()); 2238 2239 BinaryOperator *Incr = 2240 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); 2241 if (!Incr) continue; 2242 if (Incr->getOpcode() != Instruction::Add 2243 && Incr->getOpcode() != Instruction::Sub) 2244 continue; 2245 2246 /* Initialize new IV, double d = 0.0 in above example. */ 2247 ConstantInt *C = NULL; 2248 if (Incr->getOperand(0) == PH) 2249 C = dyn_cast<ConstantInt>(Incr->getOperand(1)); 2250 else if (Incr->getOperand(1) == PH) 2251 C = dyn_cast<ConstantInt>(Incr->getOperand(0)); 2252 else 2253 continue; 2254 2255 if (!C) continue; 2256 2257 /* Add new PHINode. */ 2258 PHINode *NewPH = PHINode::Create(DestTy, "IV.S.", PH); 2259 2260 /* create new increment. '++d' in above example. */ 2261 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); 2262 BinaryOperator *NewIncr = 2263 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? 2264 Instruction::FAdd : Instruction::FSub, 2265 NewPH, CFP, "IV.S.next.", Incr); 2266 2267 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); 2268 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); 2269 2270 /* Remove cast operation */ 2271 ShadowUse->replaceAllUsesWith(NewPH); 2272 ShadowUse->eraseFromParent(); 2273 NumShadow++; 2274 break; 2275 } 2276 } 2277} 2278 2279/// OptimizeIndvars - Now that IVUsesByStride is set up with all of the indvar 2280/// uses in the loop, look to see if we can eliminate some, in favor of using 2281/// common indvars for the different uses. 2282void LoopStrengthReduce::OptimizeIndvars(Loop *L) { 2283 // TODO: implement optzns here. 2284 2285 OptimizeShadowIV(L); 2286} 2287 2288/// OptimizeLoopTermCond - Change loop terminating condition to use the 2289/// postinc iv when possible. 2290void LoopStrengthReduce::OptimizeLoopTermCond(Loop *L) { 2291 // Finally, get the terminating condition for the loop if possible. If we 2292 // can, we want to change it to use a post-incremented version of its 2293 // induction variable, to allow coalescing the live ranges for the IV into 2294 // one register value. 2295 BasicBlock *LatchBlock = L->getLoopLatch(); 2296 BasicBlock *ExitingBlock = L->getExitingBlock(); 2297 LLVMContext &Context = LatchBlock->getContext(); 2298 2299 if (!ExitingBlock) 2300 // Multiple exits, just look at the exit in the latch block if there is one. 2301 ExitingBlock = LatchBlock; 2302 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 2303 if (!TermBr) 2304 return; 2305 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) 2306 return; 2307 2308 // Search IVUsesByStride to find Cond's IVUse if there is one. 2309 IVStrideUse *CondUse = 0; 2310 const SCEV *const *CondStride = 0; 2311 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); 2312 if (!FindIVUserForCond(Cond, CondUse, CondStride)) 2313 return; // setcc doesn't use the IV. 2314 2315 if (ExitingBlock != LatchBlock) { 2316 if (!Cond->hasOneUse()) 2317 // See below, we don't want the condition to be cloned. 2318 return; 2319 2320 // If exiting block is the latch block, we know it's safe and profitable to 2321 // transform the icmp to use post-inc iv. Otherwise do so only if it would 2322 // not reuse another iv and its iv would be reused by other uses. We are 2323 // optimizing for the case where the icmp is the only use of the iv. 2324 IVUsersOfOneStride &StrideUses = *IU->IVUsesByStride[*CondStride]; 2325 for (ilist<IVStrideUse>::iterator I = StrideUses.Users.begin(), 2326 E = StrideUses.Users.end(); I != E; ++I) { 2327 if (I->getUser() == Cond) 2328 continue; 2329 if (!I->isUseOfPostIncrementedValue()) 2330 return; 2331 } 2332 2333 // FIXME: This is expensive, and worse still ChangeCompareStride does a 2334 // similar check. Can we perform all the icmp related transformations after 2335 // StrengthReduceStridedIVUsers? 2336 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(*CondStride)) { 2337 int64_t SInt = SC->getValue()->getSExtValue(); 2338 for (unsigned NewStride = 0, ee = IU->StrideOrder.size(); NewStride != ee; 2339 ++NewStride) { 2340 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = 2341 IU->IVUsesByStride.find(IU->StrideOrder[NewStride]); 2342 if (!isa<SCEVConstant>(SI->first) || SI->first == *CondStride) 2343 continue; 2344 int64_t SSInt = 2345 cast<SCEVConstant>(SI->first)->getValue()->getSExtValue(); 2346 if (SSInt == SInt) 2347 return; // This can definitely be reused. 2348 if (unsigned(abs64(SSInt)) < SInt || (SSInt % SInt) != 0) 2349 continue; 2350 int64_t Scale = SSInt / SInt; 2351 bool AllUsesAreAddresses = true; 2352 bool AllUsesAreOutsideLoop = true; 2353 std::vector<BasedUser> UsersToProcess; 2354 const SCEV *CommonExprs = CollectIVUsers(SI->first, *SI->second, L, 2355 AllUsesAreAddresses, 2356 AllUsesAreOutsideLoop, 2357 UsersToProcess); 2358 // Avoid rewriting the compare instruction with an iv of new stride 2359 // if it's likely the new stride uses will be rewritten using the 2360 // stride of the compare instruction. 2361 if (AllUsesAreAddresses && 2362 ValidScale(!CommonExprs->isZero(), Scale, UsersToProcess)) 2363 return; 2364 } 2365 } 2366 2367 StrideNoReuse.insert(*CondStride); 2368 } 2369 2370 // If the trip count is computed in terms of a max (due to ScalarEvolution 2371 // being unable to find a sufficient guard, for example), change the loop 2372 // comparison to use SLT or ULT instead of NE. 2373 Cond = OptimizeMax(L, Cond, CondUse); 2374 2375 // If possible, change stride and operands of the compare instruction to 2376 // eliminate one stride. 2377 if (ExitingBlock == LatchBlock) 2378 Cond = ChangeCompareStride(L, Cond, CondUse, CondStride); 2379 2380 // It's possible for the setcc instruction to be anywhere in the loop, and 2381 // possible for it to have multiple users. If it is not immediately before 2382 // the latch block branch, move it. 2383 if (&*++BasicBlock::iterator(Cond) != (Instruction*)TermBr) { 2384 if (Cond->hasOneUse()) { // Condition has a single use, just move it. 2385 Cond->moveBefore(TermBr); 2386 } else { 2387 // Otherwise, clone the terminating condition and insert into the loopend. 2388 Cond = cast<ICmpInst>(Cond->clone(Context)); 2389 Cond->setName(L->getHeader()->getName() + ".termcond"); 2390 LatchBlock->getInstList().insert(TermBr, Cond); 2391 2392 // Clone the IVUse, as the old use still exists! 2393 IU->IVUsesByStride[*CondStride]->addUser(CondUse->getOffset(), Cond, 2394 CondUse->getOperandValToReplace()); 2395 CondUse = &IU->IVUsesByStride[*CondStride]->Users.back(); 2396 } 2397 } 2398 2399 // If we get to here, we know that we can transform the setcc instruction to 2400 // use the post-incremented version of the IV, allowing us to coalesce the 2401 // live ranges for the IV correctly. 2402 CondUse->setOffset(SE->getMinusSCEV(CondUse->getOffset(), *CondStride)); 2403 CondUse->setIsUseOfPostIncrementedValue(true); 2404 Changed = true; 2405 2406 ++NumLoopCond; 2407} 2408 2409/// OptimizeLoopCountIV - If, after all sharing of IVs, the IV used for deciding 2410/// when to exit the loop is used only for that purpose, try to rearrange things 2411/// so it counts down to a test against zero. 2412void LoopStrengthReduce::OptimizeLoopCountIV(Loop *L) { 2413 2414 // If the number of times the loop is executed isn't computable, give up. 2415 const SCEV *BackedgeTakenCount = SE->getBackedgeTakenCount(L); 2416 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 2417 return; 2418 2419 // Get the terminating condition for the loop if possible (this isn't 2420 // necessarily in the latch, or a block that's a predecessor of the header). 2421 if (!L->getExitBlock()) 2422 return; // More than one loop exit blocks. 2423 2424 // Okay, there is one exit block. Try to find the condition that causes the 2425 // loop to be exited. 2426 BasicBlock *ExitingBlock = L->getExitingBlock(); 2427 if (!ExitingBlock) 2428 return; // More than one block exiting! 2429 2430 // Okay, we've computed the exiting block. See what condition causes us to 2431 // exit. 2432 // 2433 // FIXME: we should be able to handle switch instructions (with a single exit) 2434 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 2435 if (TermBr == 0) return; 2436 assert(TermBr->isConditional() && "If unconditional, it can't be in loop!"); 2437 if (!isa<ICmpInst>(TermBr->getCondition())) 2438 return; 2439 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); 2440 2441 // Handle only tests for equality for the moment, and only stride 1. 2442 if (Cond->getPredicate() != CmpInst::ICMP_EQ) 2443 return; 2444 const SCEV *IV = SE->getSCEV(Cond->getOperand(0)); 2445 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); 2446 const SCEV *One = SE->getIntegerSCEV(1, BackedgeTakenCount->getType()); 2447 if (!AR || !AR->isAffine() || AR->getStepRecurrence(*SE) != One) 2448 return; 2449 // If the RHS of the comparison is defined inside the loop, the rewrite 2450 // cannot be done. 2451 if (Instruction *CR = dyn_cast<Instruction>(Cond->getOperand(1))) 2452 if (L->contains(CR->getParent())) 2453 return; 2454 2455 // Make sure the IV is only used for counting. Value may be preinc or 2456 // postinc; 2 uses in either case. 2457 if (!Cond->getOperand(0)->hasNUses(2)) 2458 return; 2459 PHINode *phi = dyn_cast<PHINode>(Cond->getOperand(0)); 2460 Instruction *incr; 2461 if (phi && phi->getParent()==L->getHeader()) { 2462 // value tested is preinc. Find the increment. 2463 // A CmpInst is not a BinaryOperator; we depend on this. 2464 Instruction::use_iterator UI = phi->use_begin(); 2465 incr = dyn_cast<BinaryOperator>(UI); 2466 if (!incr) 2467 incr = dyn_cast<BinaryOperator>(++UI); 2468 // 1 use for postinc value, the phi. Unnecessarily conservative? 2469 if (!incr || !incr->hasOneUse() || incr->getOpcode()!=Instruction::Add) 2470 return; 2471 } else { 2472 // Value tested is postinc. Find the phi node. 2473 incr = dyn_cast<BinaryOperator>(Cond->getOperand(0)); 2474 if (!incr || incr->getOpcode()!=Instruction::Add) 2475 return; 2476 2477 Instruction::use_iterator UI = Cond->getOperand(0)->use_begin(); 2478 phi = dyn_cast<PHINode>(UI); 2479 if (!phi) 2480 phi = dyn_cast<PHINode>(++UI); 2481 // 1 use for preinc value, the increment. 2482 if (!phi || phi->getParent()!=L->getHeader() || !phi->hasOneUse()) 2483 return; 2484 } 2485 2486 // Replace the increment with a decrement. 2487 BinaryOperator *decr = 2488 BinaryOperator::Create(Instruction::Sub, incr->getOperand(0), 2489 incr->getOperand(1), "tmp", incr); 2490 incr->replaceAllUsesWith(decr); 2491 incr->eraseFromParent(); 2492 2493 // Substitute endval-startval for the original startval, and 0 for the 2494 // original endval. Since we're only testing for equality this is OK even 2495 // if the computation wraps around. 2496 BasicBlock *Preheader = L->getLoopPreheader(); 2497 Instruction *PreInsertPt = Preheader->getTerminator(); 2498 int inBlock = L->contains(phi->getIncomingBlock(0)) ? 1 : 0; 2499 Value *startVal = phi->getIncomingValue(inBlock); 2500 Value *endVal = Cond->getOperand(1); 2501 // FIXME check for case where both are constant 2502 Constant* Zero = ConstantInt::get(Cond->getOperand(1)->getType(), 0); 2503 BinaryOperator *NewStartVal = 2504 BinaryOperator::Create(Instruction::Sub, endVal, startVal, 2505 "tmp", PreInsertPt); 2506 phi->setIncomingValue(inBlock, NewStartVal); 2507 Cond->setOperand(1, Zero); 2508 2509 Changed = true; 2510} 2511 2512bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager &LPM) { 2513 2514 IU = &getAnalysis<IVUsers>(); 2515 LI = &getAnalysis<LoopInfo>(); 2516 DT = &getAnalysis<DominatorTree>(); 2517 SE = &getAnalysis<ScalarEvolution>(); 2518 Changed = false; 2519 2520 if (!IU->IVUsesByStride.empty()) { 2521 DEBUG(errs() << "\nLSR on \"" << L->getHeader()->getParent()->getName() 2522 << "\" "; 2523 L->dump()); 2524 2525 // Sort the StrideOrder so we process larger strides first. 2526 std::stable_sort(IU->StrideOrder.begin(), IU->StrideOrder.end(), 2527 StrideCompare(SE)); 2528 2529 // Optimize induction variables. Some indvar uses can be transformed to use 2530 // strides that will be needed for other purposes. A common example of this 2531 // is the exit test for the loop, which can often be rewritten to use the 2532 // computation of some other indvar to decide when to terminate the loop. 2533 OptimizeIndvars(L); 2534 2535 // Change loop terminating condition to use the postinc iv when possible 2536 // and optimize loop terminating compare. FIXME: Move this after 2537 // StrengthReduceStridedIVUsers? 2538 OptimizeLoopTermCond(L); 2539 2540 // FIXME: We can shrink overlarge IV's here. e.g. if the code has 2541 // computation in i64 values and the target doesn't support i64, demote 2542 // the computation to 32-bit if safe. 2543 2544 // FIXME: Attempt to reuse values across multiple IV's. In particular, we 2545 // could have something like "for(i) { foo(i*8); bar(i*16) }", which should 2546 // be codegened as "for (j = 0;; j+=8) { foo(j); bar(j+j); }" on X86/PPC. 2547 // Need to be careful that IV's are all the same type. Only works for 2548 // intptr_t indvars. 2549 2550 // IVsByStride keeps IVs for one particular loop. 2551 assert(IVsByStride.empty() && "Stale entries in IVsByStride?"); 2552 2553 // Note: this processes each stride/type pair individually. All users 2554 // passed into StrengthReduceStridedIVUsers have the same type AND stride. 2555 // Also, note that we iterate over IVUsesByStride indirectly by using 2556 // StrideOrder. This extra layer of indirection makes the ordering of 2557 // strides deterministic - not dependent on map order. 2558 for (unsigned Stride = 0, e = IU->StrideOrder.size(); 2559 Stride != e; ++Stride) { 2560 std::map<const SCEV *, IVUsersOfOneStride *>::iterator SI = 2561 IU->IVUsesByStride.find(IU->StrideOrder[Stride]); 2562 assert(SI != IU->IVUsesByStride.end() && "Stride doesn't exist!"); 2563 // FIXME: Generalize to non-affine IV's. 2564 if (!SI->first->isLoopInvariant(L)) 2565 continue; 2566 StrengthReduceStridedIVUsers(SI->first, *SI->second, L); 2567 } 2568 } 2569 2570 // After all sharing is done, see if we can adjust the loop to test against 2571 // zero instead of counting up to a maximum. This is usually faster. 2572 OptimizeLoopCountIV(L); 2573 2574 // We're done analyzing this loop; release all the state we built up for it. 2575 IVsByStride.clear(); 2576 StrideNoReuse.clear(); 2577 2578 // Clean up after ourselves 2579 if (!DeadInsts.empty()) 2580 DeleteTriviallyDeadInstructions(); 2581 2582 // At this point, it is worth checking to see if any recurrence PHIs are also 2583 // dead, so that we can remove them as well. 2584 DeleteDeadPHIs(L->getHeader()); 2585 2586 return Changed; 2587} 2588