ScalarEvolutionExpander.cpp revision 263508
1//===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===// 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 file contains the implementation of the scalar evolution expander, 11// which is used to generate the code corresponding to a given scalar evolution 12// expression. 13// 14//===----------------------------------------------------------------------===// 15 16#include "llvm/Analysis/ScalarEvolutionExpander.h" 17#include "llvm/ADT/SmallSet.h" 18#include "llvm/ADT/STLExtras.h" 19#include "llvm/Analysis/LoopInfo.h" 20#include "llvm/Analysis/TargetTransformInfo.h" 21#include "llvm/IR/DataLayout.h" 22#include "llvm/IR/IntrinsicInst.h" 23#include "llvm/IR/LLVMContext.h" 24#include "llvm/Support/Debug.h" 25 26using namespace llvm; 27 28/// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, 29/// reusing an existing cast if a suitable one exists, moving an existing 30/// cast if a suitable one exists but isn't in the right place, or 31/// creating a new one. 32Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, 33 Instruction::CastOps Op, 34 BasicBlock::iterator IP) { 35 // This function must be called with the builder having a valid insertion 36 // point. It doesn't need to be the actual IP where the uses of the returned 37 // cast will be added, but it must dominate such IP. 38 // We use this precondition to produce a cast that will dominate all its 39 // uses. In particular, this is crucial for the case where the builder's 40 // insertion point *is* the point where we were asked to put the cast. 41 // Since we don't know the builder's insertion point is actually 42 // where the uses will be added (only that it dominates it), we are 43 // not allowed to move it. 44 BasicBlock::iterator BIP = Builder.GetInsertPoint(); 45 46 Instruction *Ret = NULL; 47 48 // Check to see if there is already a cast! 49 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); 50 UI != E; ++UI) { 51 User *U = *UI; 52 if (U->getType() == Ty) 53 if (CastInst *CI = dyn_cast<CastInst>(U)) 54 if (CI->getOpcode() == Op) { 55 // If the cast isn't where we want it, create a new cast at IP. 56 // Likewise, do not reuse a cast at BIP because it must dominate 57 // instructions that might be inserted before BIP. 58 if (BasicBlock::iterator(CI) != IP || BIP == IP) { 59 // Create a new cast, and leave the old cast in place in case 60 // it is being used as an insert point. Clear its operand 61 // so that it doesn't hold anything live. 62 Ret = CastInst::Create(Op, V, Ty, "", IP); 63 Ret->takeName(CI); 64 CI->replaceAllUsesWith(Ret); 65 CI->setOperand(0, UndefValue::get(V->getType())); 66 break; 67 } 68 Ret = CI; 69 break; 70 } 71 } 72 73 // Create a new cast. 74 if (!Ret) 75 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP); 76 77 // We assert at the end of the function since IP might point to an 78 // instruction with different dominance properties than a cast 79 // (an invoke for example) and not dominate BIP (but the cast does). 80 assert(SE.DT->dominates(Ret, BIP)); 81 82 rememberInstruction(Ret); 83 return Ret; 84} 85 86/// InsertNoopCastOfTo - Insert a cast of V to the specified type, 87/// which must be possible with a noop cast, doing what we can to share 88/// the casts. 89Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 90 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 91 assert((Op == Instruction::BitCast || 92 Op == Instruction::PtrToInt || 93 Op == Instruction::IntToPtr) && 94 "InsertNoopCastOfTo cannot perform non-noop casts!"); 95 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 96 "InsertNoopCastOfTo cannot change sizes!"); 97 98 // Short-circuit unnecessary bitcasts. 99 if (Op == Instruction::BitCast) { 100 if (V->getType() == Ty) 101 return V; 102 if (CastInst *CI = dyn_cast<CastInst>(V)) { 103 if (CI->getOperand(0)->getType() == Ty) 104 return CI->getOperand(0); 105 } 106 } 107 // Short-circuit unnecessary inttoptr<->ptrtoint casts. 108 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 109 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 110 if (CastInst *CI = dyn_cast<CastInst>(V)) 111 if ((CI->getOpcode() == Instruction::PtrToInt || 112 CI->getOpcode() == Instruction::IntToPtr) && 113 SE.getTypeSizeInBits(CI->getType()) == 114 SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 115 return CI->getOperand(0); 116 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 117 if ((CE->getOpcode() == Instruction::PtrToInt || 118 CE->getOpcode() == Instruction::IntToPtr) && 119 SE.getTypeSizeInBits(CE->getType()) == 120 SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 121 return CE->getOperand(0); 122 } 123 124 // Fold a cast of a constant. 125 if (Constant *C = dyn_cast<Constant>(V)) 126 return ConstantExpr::getCast(Op, C, Ty); 127 128 // Cast the argument at the beginning of the entry block, after 129 // any bitcasts of other arguments. 130 if (Argument *A = dyn_cast<Argument>(V)) { 131 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 132 while ((isa<BitCastInst>(IP) && 133 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 134 cast<BitCastInst>(IP)->getOperand(0) != A) || 135 isa<DbgInfoIntrinsic>(IP) || 136 isa<LandingPadInst>(IP)) 137 ++IP; 138 return ReuseOrCreateCast(A, Ty, Op, IP); 139 } 140 141 // Cast the instruction immediately after the instruction. 142 Instruction *I = cast<Instruction>(V); 143 BasicBlock::iterator IP = I; ++IP; 144 if (InvokeInst *II = dyn_cast<InvokeInst>(I)) 145 IP = II->getNormalDest()->begin(); 146 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP)) 147 ++IP; 148 return ReuseOrCreateCast(I, Ty, Op, IP); 149} 150 151/// InsertBinop - Insert the specified binary operator, doing a small amount 152/// of work to avoid inserting an obviously redundant operation. 153Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 154 Value *LHS, Value *RHS) { 155 // Fold a binop with constant operands. 156 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 157 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 158 return ConstantExpr::get(Opcode, CLHS, CRHS); 159 160 // Do a quick scan to see if we have this binop nearby. If so, reuse it. 161 unsigned ScanLimit = 6; 162 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 163 // Scanning starts from the last instruction before the insertion point. 164 BasicBlock::iterator IP = Builder.GetInsertPoint(); 165 if (IP != BlockBegin) { 166 --IP; 167 for (; ScanLimit; --IP, --ScanLimit) { 168 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 169 // generated code. 170 if (isa<DbgInfoIntrinsic>(IP)) 171 ScanLimit++; 172 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 173 IP->getOperand(1) == RHS) 174 return IP; 175 if (IP == BlockBegin) break; 176 } 177 } 178 179 // Save the original insertion point so we can restore it when we're done. 180 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); 181 BuilderType::InsertPointGuard Guard(Builder); 182 183 // Move the insertion point out of as many loops as we can. 184 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 185 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 186 BasicBlock *Preheader = L->getLoopPreheader(); 187 if (!Preheader) break; 188 189 // Ok, move up a level. 190 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 191 } 192 193 // If we haven't found this binop, insert it. 194 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 195 BO->setDebugLoc(Loc); 196 rememberInstruction(BO); 197 198 return BO; 199} 200 201/// FactorOutConstant - Test if S is divisible by Factor, using signed 202/// division. If so, update S with Factor divided out and return true. 203/// S need not be evenly divisible if a reasonable remainder can be 204/// computed. 205/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 206/// unnecessary; in its place, just signed-divide Ops[i] by the scale and 207/// check to see if the divide was folded. 208static bool FactorOutConstant(const SCEV *&S, 209 const SCEV *&Remainder, 210 const SCEV *Factor, 211 ScalarEvolution &SE, 212 const DataLayout *TD) { 213 // Everything is divisible by one. 214 if (Factor->isOne()) 215 return true; 216 217 // x/x == 1. 218 if (S == Factor) { 219 S = SE.getConstant(S->getType(), 1); 220 return true; 221 } 222 223 // For a Constant, check for a multiple of the given factor. 224 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 225 // 0/x == 0. 226 if (C->isZero()) 227 return true; 228 // Check for divisibility. 229 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 230 ConstantInt *CI = 231 ConstantInt::get(SE.getContext(), 232 C->getValue()->getValue().sdiv( 233 FC->getValue()->getValue())); 234 // If the quotient is zero and the remainder is non-zero, reject 235 // the value at this scale. It will be considered for subsequent 236 // smaller scales. 237 if (!CI->isZero()) { 238 const SCEV *Div = SE.getConstant(CI); 239 S = Div; 240 Remainder = 241 SE.getAddExpr(Remainder, 242 SE.getConstant(C->getValue()->getValue().srem( 243 FC->getValue()->getValue()))); 244 return true; 245 } 246 } 247 } 248 249 // In a Mul, check if there is a constant operand which is a multiple 250 // of the given factor. 251 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 252 if (TD) { 253 // With DataLayout, the size is known. Check if there is a constant 254 // operand which is a multiple of the given factor. If so, we can 255 // factor it. 256 const SCEVConstant *FC = cast<SCEVConstant>(Factor); 257 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 258 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) { 259 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 260 NewMulOps[0] = 261 SE.getConstant(C->getValue()->getValue().sdiv( 262 FC->getValue()->getValue())); 263 S = SE.getMulExpr(NewMulOps); 264 return true; 265 } 266 } else { 267 // Without DataLayout, check if Factor can be factored out of any of the 268 // Mul's operands. If so, we can just remove it. 269 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) { 270 const SCEV *SOp = M->getOperand(i); 271 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0); 272 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) && 273 Remainder->isZero()) { 274 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 275 NewMulOps[i] = SOp; 276 S = SE.getMulExpr(NewMulOps); 277 return true; 278 } 279 } 280 } 281 } 282 283 // In an AddRec, check if both start and step are divisible. 284 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 285 const SCEV *Step = A->getStepRecurrence(SE); 286 const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 287 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD)) 288 return false; 289 if (!StepRem->isZero()) 290 return false; 291 const SCEV *Start = A->getStart(); 292 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD)) 293 return false; 294 S = SE.getAddRecExpr(Start, Step, A->getLoop(), 295 A->getNoWrapFlags(SCEV::FlagNW)); 296 return true; 297 } 298 299 return false; 300} 301 302/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 303/// is the number of SCEVAddRecExprs present, which are kept at the end of 304/// the list. 305/// 306static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 307 Type *Ty, 308 ScalarEvolution &SE) { 309 unsigned NumAddRecs = 0; 310 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 311 ++NumAddRecs; 312 // Group Ops into non-addrecs and addrecs. 313 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 314 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 315 // Let ScalarEvolution sort and simplify the non-addrecs list. 316 const SCEV *Sum = NoAddRecs.empty() ? 317 SE.getConstant(Ty, 0) : 318 SE.getAddExpr(NoAddRecs); 319 // If it returned an add, use the operands. Otherwise it simplified 320 // the sum into a single value, so just use that. 321 Ops.clear(); 322 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 323 Ops.append(Add->op_begin(), Add->op_end()); 324 else if (!Sum->isZero()) 325 Ops.push_back(Sum); 326 // Then append the addrecs. 327 Ops.append(AddRecs.begin(), AddRecs.end()); 328} 329 330/// SplitAddRecs - Flatten a list of add operands, moving addrec start values 331/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 332/// This helps expose more opportunities for folding parts of the expressions 333/// into GEP indices. 334/// 335static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 336 Type *Ty, 337 ScalarEvolution &SE) { 338 // Find the addrecs. 339 SmallVector<const SCEV *, 8> AddRecs; 340 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 341 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 342 const SCEV *Start = A->getStart(); 343 if (Start->isZero()) break; 344 const SCEV *Zero = SE.getConstant(Ty, 0); 345 AddRecs.push_back(SE.getAddRecExpr(Zero, 346 A->getStepRecurrence(SE), 347 A->getLoop(), 348 A->getNoWrapFlags(SCEV::FlagNW))); 349 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 350 Ops[i] = Zero; 351 Ops.append(Add->op_begin(), Add->op_end()); 352 e += Add->getNumOperands(); 353 } else { 354 Ops[i] = Start; 355 } 356 } 357 if (!AddRecs.empty()) { 358 // Add the addrecs onto the end of the list. 359 Ops.append(AddRecs.begin(), AddRecs.end()); 360 // Resort the operand list, moving any constants to the front. 361 SimplifyAddOperands(Ops, Ty, SE); 362 } 363} 364 365/// expandAddToGEP - Expand an addition expression with a pointer type into 366/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 367/// BasicAliasAnalysis and other passes analyze the result. See the rules 368/// for getelementptr vs. inttoptr in 369/// http://llvm.org/docs/LangRef.html#pointeraliasing 370/// for details. 371/// 372/// Design note: The correctness of using getelementptr here depends on 373/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 374/// they may introduce pointer arithmetic which may not be safely converted 375/// into getelementptr. 376/// 377/// Design note: It might seem desirable for this function to be more 378/// loop-aware. If some of the indices are loop-invariant while others 379/// aren't, it might seem desirable to emit multiple GEPs, keeping the 380/// loop-invariant portions of the overall computation outside the loop. 381/// However, there are a few reasons this is not done here. Hoisting simple 382/// arithmetic is a low-level optimization that often isn't very 383/// important until late in the optimization process. In fact, passes 384/// like InstructionCombining will combine GEPs, even if it means 385/// pushing loop-invariant computation down into loops, so even if the 386/// GEPs were split here, the work would quickly be undone. The 387/// LoopStrengthReduction pass, which is usually run quite late (and 388/// after the last InstructionCombining pass), takes care of hoisting 389/// loop-invariant portions of expressions, after considering what 390/// can be folded using target addressing modes. 391/// 392Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 393 const SCEV *const *op_end, 394 PointerType *PTy, 395 Type *Ty, 396 Value *V) { 397 Type *ElTy = PTy->getElementType(); 398 SmallVector<Value *, 4> GepIndices; 399 SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 400 bool AnyNonZeroIndices = false; 401 402 // Split AddRecs up into parts as either of the parts may be usable 403 // without the other. 404 SplitAddRecs(Ops, Ty, SE); 405 406 Type *IntPtrTy = SE.TD 407 ? SE.TD->getIntPtrType(PTy) 408 : Type::getInt64Ty(PTy->getContext()); 409 410 // Descend down the pointer's type and attempt to convert the other 411 // operands into GEP indices, at each level. The first index in a GEP 412 // indexes into the array implied by the pointer operand; the rest of 413 // the indices index into the element or field type selected by the 414 // preceding index. 415 for (;;) { 416 // If the scale size is not 0, attempt to factor out a scale for 417 // array indexing. 418 SmallVector<const SCEV *, 8> ScaledOps; 419 if (ElTy->isSized()) { 420 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy); 421 if (!ElSize->isZero()) { 422 SmallVector<const SCEV *, 8> NewOps; 423 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 424 const SCEV *Op = Ops[i]; 425 const SCEV *Remainder = SE.getConstant(Ty, 0); 426 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) { 427 // Op now has ElSize factored out. 428 ScaledOps.push_back(Op); 429 if (!Remainder->isZero()) 430 NewOps.push_back(Remainder); 431 AnyNonZeroIndices = true; 432 } else { 433 // The operand was not divisible, so add it to the list of operands 434 // we'll scan next iteration. 435 NewOps.push_back(Ops[i]); 436 } 437 } 438 // If we made any changes, update Ops. 439 if (!ScaledOps.empty()) { 440 Ops = NewOps; 441 SimplifyAddOperands(Ops, Ty, SE); 442 } 443 } 444 } 445 446 // Record the scaled array index for this level of the type. If 447 // we didn't find any operands that could be factored, tentatively 448 // assume that element zero was selected (since the zero offset 449 // would obviously be folded away). 450 Value *Scaled = ScaledOps.empty() ? 451 Constant::getNullValue(Ty) : 452 expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 453 GepIndices.push_back(Scaled); 454 455 // Collect struct field index operands. 456 while (StructType *STy = dyn_cast<StructType>(ElTy)) { 457 bool FoundFieldNo = false; 458 // An empty struct has no fields. 459 if (STy->getNumElements() == 0) break; 460 if (SE.TD) { 461 // With DataLayout, field offsets are known. See if a constant offset 462 // falls within any of the struct fields. 463 if (Ops.empty()) break; 464 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 465 if (SE.getTypeSizeInBits(C->getType()) <= 64) { 466 const StructLayout &SL = *SE.TD->getStructLayout(STy); 467 uint64_t FullOffset = C->getValue()->getZExtValue(); 468 if (FullOffset < SL.getSizeInBytes()) { 469 unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 470 GepIndices.push_back( 471 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 472 ElTy = STy->getTypeAtIndex(ElIdx); 473 Ops[0] = 474 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 475 AnyNonZeroIndices = true; 476 FoundFieldNo = true; 477 } 478 } 479 } else { 480 // Without DataLayout, just check for an offsetof expression of the 481 // appropriate struct type. 482 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 483 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) { 484 Type *CTy; 485 Constant *FieldNo; 486 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) { 487 GepIndices.push_back(FieldNo); 488 ElTy = 489 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue()); 490 Ops[i] = SE.getConstant(Ty, 0); 491 AnyNonZeroIndices = true; 492 FoundFieldNo = true; 493 break; 494 } 495 } 496 } 497 // If no struct field offsets were found, tentatively assume that 498 // field zero was selected (since the zero offset would obviously 499 // be folded away). 500 if (!FoundFieldNo) { 501 ElTy = STy->getTypeAtIndex(0u); 502 GepIndices.push_back( 503 Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 504 } 505 } 506 507 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 508 ElTy = ATy->getElementType(); 509 else 510 break; 511 } 512 513 // If none of the operands were convertible to proper GEP indices, cast 514 // the base to i8* and do an ugly getelementptr with that. It's still 515 // better than ptrtoint+arithmetic+inttoptr at least. 516 if (!AnyNonZeroIndices) { 517 // Cast the base to i8*. 518 V = InsertNoopCastOfTo(V, 519 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 520 521 assert(!isa<Instruction>(V) || 522 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint())); 523 524 // Expand the operands for a plain byte offset. 525 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 526 527 // Fold a GEP with constant operands. 528 if (Constant *CLHS = dyn_cast<Constant>(V)) 529 if (Constant *CRHS = dyn_cast<Constant>(Idx)) 530 return ConstantExpr::getGetElementPtr(CLHS, CRHS); 531 532 // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 533 unsigned ScanLimit = 6; 534 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 535 // Scanning starts from the last instruction before the insertion point. 536 BasicBlock::iterator IP = Builder.GetInsertPoint(); 537 if (IP != BlockBegin) { 538 --IP; 539 for (; ScanLimit; --IP, --ScanLimit) { 540 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 541 // generated code. 542 if (isa<DbgInfoIntrinsic>(IP)) 543 ScanLimit++; 544 if (IP->getOpcode() == Instruction::GetElementPtr && 545 IP->getOperand(0) == V && IP->getOperand(1) == Idx) 546 return IP; 547 if (IP == BlockBegin) break; 548 } 549 } 550 551 // Save the original insertion point so we can restore it when we're done. 552 BuilderType::InsertPointGuard Guard(Builder); 553 554 // Move the insertion point out of as many loops as we can. 555 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 556 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 557 BasicBlock *Preheader = L->getLoopPreheader(); 558 if (!Preheader) break; 559 560 // Ok, move up a level. 561 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 562 } 563 564 // Emit a GEP. 565 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep"); 566 rememberInstruction(GEP); 567 568 return GEP; 569 } 570 571 // Save the original insertion point so we can restore it when we're done. 572 BuilderType::InsertPoint SaveInsertPt = Builder.saveIP(); 573 574 // Move the insertion point out of as many loops as we can. 575 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 576 if (!L->isLoopInvariant(V)) break; 577 578 bool AnyIndexNotLoopInvariant = false; 579 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(), 580 E = GepIndices.end(); I != E; ++I) 581 if (!L->isLoopInvariant(*I)) { 582 AnyIndexNotLoopInvariant = true; 583 break; 584 } 585 if (AnyIndexNotLoopInvariant) 586 break; 587 588 BasicBlock *Preheader = L->getLoopPreheader(); 589 if (!Preheader) break; 590 591 // Ok, move up a level. 592 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 593 } 594 595 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 596 // because ScalarEvolution may have changed the address arithmetic to 597 // compute a value which is beyond the end of the allocated object. 598 Value *Casted = V; 599 if (V->getType() != PTy) 600 Casted = InsertNoopCastOfTo(Casted, PTy); 601 Value *GEP = Builder.CreateGEP(Casted, 602 GepIndices, 603 "scevgep"); 604 Ops.push_back(SE.getUnknown(GEP)); 605 rememberInstruction(GEP); 606 607 // Restore the original insert point. 608 Builder.restoreIP(SaveInsertPt); 609 610 return expand(SE.getAddExpr(Ops)); 611} 612 613/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 614/// SCEV expansion. If they are nested, this is the most nested. If they are 615/// neighboring, pick the later. 616static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 617 DominatorTree &DT) { 618 if (!A) return B; 619 if (!B) return A; 620 if (A->contains(B)) return B; 621 if (B->contains(A)) return A; 622 if (DT.dominates(A->getHeader(), B->getHeader())) return B; 623 if (DT.dominates(B->getHeader(), A->getHeader())) return A; 624 return A; // Arbitrarily break the tie. 625} 626 627/// getRelevantLoop - Get the most relevant loop associated with the given 628/// expression, according to PickMostRelevantLoop. 629const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 630 // Test whether we've already computed the most relevant loop for this SCEV. 631 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair = 632 RelevantLoops.insert(std::make_pair(S, static_cast<const Loop *>(0))); 633 if (!Pair.second) 634 return Pair.first->second; 635 636 if (isa<SCEVConstant>(S)) 637 // A constant has no relevant loops. 638 return 0; 639 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 640 if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 641 return Pair.first->second = SE.LI->getLoopFor(I->getParent()); 642 // A non-instruction has no relevant loops. 643 return 0; 644 } 645 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 646 const Loop *L = 0; 647 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 648 L = AR->getLoop(); 649 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end(); 650 I != E; ++I) 651 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT); 652 return RelevantLoops[N] = L; 653 } 654 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 655 const Loop *Result = getRelevantLoop(C->getOperand()); 656 return RelevantLoops[C] = Result; 657 } 658 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 659 const Loop *Result = 660 PickMostRelevantLoop(getRelevantLoop(D->getLHS()), 661 getRelevantLoop(D->getRHS()), 662 *SE.DT); 663 return RelevantLoops[D] = Result; 664 } 665 llvm_unreachable("Unexpected SCEV type!"); 666} 667 668namespace { 669 670/// LoopCompare - Compare loops by PickMostRelevantLoop. 671class LoopCompare { 672 DominatorTree &DT; 673public: 674 explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 675 676 bool operator()(std::pair<const Loop *, const SCEV *> LHS, 677 std::pair<const Loop *, const SCEV *> RHS) const { 678 // Keep pointer operands sorted at the end. 679 if (LHS.second->getType()->isPointerTy() != 680 RHS.second->getType()->isPointerTy()) 681 return LHS.second->getType()->isPointerTy(); 682 683 // Compare loops with PickMostRelevantLoop. 684 if (LHS.first != RHS.first) 685 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 686 687 // If one operand is a non-constant negative and the other is not, 688 // put the non-constant negative on the right so that a sub can 689 // be used instead of a negate and add. 690 if (LHS.second->isNonConstantNegative()) { 691 if (!RHS.second->isNonConstantNegative()) 692 return false; 693 } else if (RHS.second->isNonConstantNegative()) 694 return true; 695 696 // Otherwise they are equivalent according to this comparison. 697 return false; 698 } 699}; 700 701} 702 703Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 704 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 705 706 // Collect all the add operands in a loop, along with their associated loops. 707 // Iterate in reverse so that constants are emitted last, all else equal, and 708 // so that pointer operands are inserted first, which the code below relies on 709 // to form more involved GEPs. 710 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 711 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 712 E(S->op_begin()); I != E; ++I) 713 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 714 715 // Sort by loop. Use a stable sort so that constants follow non-constants and 716 // pointer operands precede non-pointer operands. 717 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 718 719 // Emit instructions to add all the operands. Hoist as much as possible 720 // out of loops, and form meaningful getelementptrs where possible. 721 Value *Sum = 0; 722 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 723 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 724 const Loop *CurLoop = I->first; 725 const SCEV *Op = I->second; 726 if (!Sum) { 727 // This is the first operand. Just expand it. 728 Sum = expand(Op); 729 ++I; 730 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 731 // The running sum expression is a pointer. Try to form a getelementptr 732 // at this level with that as the base. 733 SmallVector<const SCEV *, 4> NewOps; 734 for (; I != E && I->first == CurLoop; ++I) { 735 // If the operand is SCEVUnknown and not instructions, peek through 736 // it, to enable more of it to be folded into the GEP. 737 const SCEV *X = I->second; 738 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 739 if (!isa<Instruction>(U->getValue())) 740 X = SE.getSCEV(U->getValue()); 741 NewOps.push_back(X); 742 } 743 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 744 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 745 // The running sum is an integer, and there's a pointer at this level. 746 // Try to form a getelementptr. If the running sum is instructions, 747 // use a SCEVUnknown to avoid re-analyzing them. 748 SmallVector<const SCEV *, 4> NewOps; 749 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 750 SE.getSCEV(Sum)); 751 for (++I; I != E && I->first == CurLoop; ++I) 752 NewOps.push_back(I->second); 753 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 754 } else if (Op->isNonConstantNegative()) { 755 // Instead of doing a negate and add, just do a subtract. 756 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 757 Sum = InsertNoopCastOfTo(Sum, Ty); 758 Sum = InsertBinop(Instruction::Sub, Sum, W); 759 ++I; 760 } else { 761 // A simple add. 762 Value *W = expandCodeFor(Op, Ty); 763 Sum = InsertNoopCastOfTo(Sum, Ty); 764 // Canonicalize a constant to the RHS. 765 if (isa<Constant>(Sum)) std::swap(Sum, W); 766 Sum = InsertBinop(Instruction::Add, Sum, W); 767 ++I; 768 } 769 } 770 771 return Sum; 772} 773 774Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 775 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 776 777 // Collect all the mul operands in a loop, along with their associated loops. 778 // Iterate in reverse so that constants are emitted last, all else equal. 779 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 780 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 781 E(S->op_begin()); I != E; ++I) 782 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 783 784 // Sort by loop. Use a stable sort so that constants follow non-constants. 785 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 786 787 // Emit instructions to mul all the operands. Hoist as much as possible 788 // out of loops. 789 Value *Prod = 0; 790 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 791 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 792 const SCEV *Op = I->second; 793 if (!Prod) { 794 // This is the first operand. Just expand it. 795 Prod = expand(Op); 796 ++I; 797 } else if (Op->isAllOnesValue()) { 798 // Instead of doing a multiply by negative one, just do a negate. 799 Prod = InsertNoopCastOfTo(Prod, Ty); 800 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 801 ++I; 802 } else { 803 // A simple mul. 804 Value *W = expandCodeFor(Op, Ty); 805 Prod = InsertNoopCastOfTo(Prod, Ty); 806 // Canonicalize a constant to the RHS. 807 if (isa<Constant>(Prod)) std::swap(Prod, W); 808 Prod = InsertBinop(Instruction::Mul, Prod, W); 809 ++I; 810 } 811 } 812 813 return Prod; 814} 815 816Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 817 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 818 819 Value *LHS = expandCodeFor(S->getLHS(), Ty); 820 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 821 const APInt &RHS = SC->getValue()->getValue(); 822 if (RHS.isPowerOf2()) 823 return InsertBinop(Instruction::LShr, LHS, 824 ConstantInt::get(Ty, RHS.logBase2())); 825 } 826 827 Value *RHS = expandCodeFor(S->getRHS(), Ty); 828 return InsertBinop(Instruction::UDiv, LHS, RHS); 829} 830 831/// Move parts of Base into Rest to leave Base with the minimal 832/// expression that provides a pointer operand suitable for a 833/// GEP expansion. 834static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 835 ScalarEvolution &SE) { 836 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 837 Base = A->getStart(); 838 Rest = SE.getAddExpr(Rest, 839 SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 840 A->getStepRecurrence(SE), 841 A->getLoop(), 842 A->getNoWrapFlags(SCEV::FlagNW))); 843 } 844 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 845 Base = A->getOperand(A->getNumOperands()-1); 846 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 847 NewAddOps.back() = Rest; 848 Rest = SE.getAddExpr(NewAddOps); 849 ExposePointerBase(Base, Rest, SE); 850 } 851} 852 853/// Determine if this is a well-behaved chain of instructions leading back to 854/// the PHI. If so, it may be reused by expanded expressions. 855bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 856 const Loop *L) { 857 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 858 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 859 return false; 860 // If any of the operands don't dominate the insert position, bail. 861 // Addrec operands are always loop-invariant, so this can only happen 862 // if there are instructions which haven't been hoisted. 863 if (L == IVIncInsertLoop) { 864 for (User::op_iterator OI = IncV->op_begin()+1, 865 OE = IncV->op_end(); OI != OE; ++OI) 866 if (Instruction *OInst = dyn_cast<Instruction>(OI)) 867 if (!SE.DT->dominates(OInst, IVIncInsertPos)) 868 return false; 869 } 870 // Advance to the next instruction. 871 IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 872 if (!IncV) 873 return false; 874 875 if (IncV->mayHaveSideEffects()) 876 return false; 877 878 if (IncV != PN) 879 return true; 880 881 return isNormalAddRecExprPHI(PN, IncV, L); 882} 883 884/// getIVIncOperand returns an induction variable increment's induction 885/// variable operand. 886/// 887/// If allowScale is set, any type of GEP is allowed as long as the nonIV 888/// operands dominate InsertPos. 889/// 890/// If allowScale is not set, ensure that a GEP increment conforms to one of the 891/// simple patterns generated by getAddRecExprPHILiterally and 892/// expandAddtoGEP. If the pattern isn't recognized, return NULL. 893Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 894 Instruction *InsertPos, 895 bool allowScale) { 896 if (IncV == InsertPos) 897 return NULL; 898 899 switch (IncV->getOpcode()) { 900 default: 901 return NULL; 902 // Check for a simple Add/Sub or GEP of a loop invariant step. 903 case Instruction::Add: 904 case Instruction::Sub: { 905 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 906 if (!OInst || SE.DT->dominates(OInst, InsertPos)) 907 return dyn_cast<Instruction>(IncV->getOperand(0)); 908 return NULL; 909 } 910 case Instruction::BitCast: 911 return dyn_cast<Instruction>(IncV->getOperand(0)); 912 case Instruction::GetElementPtr: 913 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end(); 914 I != E; ++I) { 915 if (isa<Constant>(*I)) 916 continue; 917 if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 918 if (!SE.DT->dominates(OInst, InsertPos)) 919 return NULL; 920 } 921 if (allowScale) { 922 // allow any kind of GEP as long as it can be hoisted. 923 continue; 924 } 925 // This must be a pointer addition of constants (pretty), which is already 926 // handled, or some number of address-size elements (ugly). Ugly geps 927 // have 2 operands. i1* is used by the expander to represent an 928 // address-size element. 929 if (IncV->getNumOperands() != 2) 930 return NULL; 931 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 932 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 933 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 934 return NULL; 935 break; 936 } 937 return dyn_cast<Instruction>(IncV->getOperand(0)); 938 } 939} 940 941/// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 942/// it available to other uses in this loop. Recursively hoist any operands, 943/// until we reach a value that dominates InsertPos. 944bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 945 if (SE.DT->dominates(IncV, InsertPos)) 946 return true; 947 948 // InsertPos must itself dominate IncV so that IncV's new position satisfies 949 // its existing users. 950 if (isa<PHINode>(InsertPos) 951 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent())) 952 return false; 953 954 // Check that the chain of IV operands leading back to Phi can be hoisted. 955 SmallVector<Instruction*, 4> IVIncs; 956 for(;;) { 957 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 958 if (!Oper) 959 return false; 960 // IncV is safe to hoist. 961 IVIncs.push_back(IncV); 962 IncV = Oper; 963 if (SE.DT->dominates(IncV, InsertPos)) 964 break; 965 } 966 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(), 967 E = IVIncs.rend(); I != E; ++I) { 968 (*I)->moveBefore(InsertPos); 969 } 970 return true; 971} 972 973/// Determine if this cyclic phi is in a form that would have been generated by 974/// LSR. We don't care if the phi was actually expanded in this pass, as long 975/// as it is in a low-cost form, for example, no implied multiplication. This 976/// should match any patterns generated by getAddRecExprPHILiterally and 977/// expandAddtoGEP. 978bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 979 const Loop *L) { 980 for(Instruction *IVOper = IncV; 981 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 982 /*allowScale=*/false));) { 983 if (IVOper == PN) 984 return true; 985 } 986 return false; 987} 988 989/// expandIVInc - Expand an IV increment at Builder's current InsertPos. 990/// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 991/// need to materialize IV increments elsewhere to handle difficult situations. 992Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 993 Type *ExpandTy, Type *IntTy, 994 bool useSubtract) { 995 Value *IncV; 996 // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 997 if (ExpandTy->isPointerTy()) { 998 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 999 // If the step isn't constant, don't use an implicitly scaled GEP, because 1000 // that would require a multiply inside the loop. 1001 if (!isa<ConstantInt>(StepV)) 1002 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 1003 GEPPtrTy->getAddressSpace()); 1004 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 1005 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 1006 if (IncV->getType() != PN->getType()) { 1007 IncV = Builder.CreateBitCast(IncV, PN->getType()); 1008 rememberInstruction(IncV); 1009 } 1010 } else { 1011 IncV = useSubtract ? 1012 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 1013 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 1014 rememberInstruction(IncV); 1015 } 1016 return IncV; 1017} 1018 1019/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 1020/// the base addrec, which is the addrec without any non-loop-dominating 1021/// values, and return the PHI. 1022PHINode * 1023SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 1024 const Loop *L, 1025 Type *ExpandTy, 1026 Type *IntTy) { 1027 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 1028 1029 // Reuse a previously-inserted PHI, if present. 1030 BasicBlock *LatchBlock = L->getLoopLatch(); 1031 if (LatchBlock) { 1032 for (BasicBlock::iterator I = L->getHeader()->begin(); 1033 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1034 if (!SE.isSCEVable(PN->getType()) || 1035 (SE.getEffectiveSCEVType(PN->getType()) != 1036 SE.getEffectiveSCEVType(Normalized->getType())) || 1037 SE.getSCEV(PN) != Normalized) 1038 continue; 1039 1040 Instruction *IncV = 1041 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 1042 1043 if (LSRMode) { 1044 if (!isExpandedAddRecExprPHI(PN, IncV, L)) 1045 continue; 1046 if (L == IVIncInsertLoop && !hoistIVInc(IncV, IVIncInsertPos)) 1047 continue; 1048 } 1049 else { 1050 if (!isNormalAddRecExprPHI(PN, IncV, L)) 1051 continue; 1052 if (L == IVIncInsertLoop) 1053 do { 1054 if (SE.DT->dominates(IncV, IVIncInsertPos)) 1055 break; 1056 // Make sure the increment is where we want it. But don't move it 1057 // down past a potential existing post-inc user. 1058 IncV->moveBefore(IVIncInsertPos); 1059 IVIncInsertPos = IncV; 1060 IncV = cast<Instruction>(IncV->getOperand(0)); 1061 } while (IncV != PN); 1062 } 1063 // Ok, the add recurrence looks usable. 1064 // Remember this PHI, even in post-inc mode. 1065 InsertedValues.insert(PN); 1066 // Remember the increment. 1067 rememberInstruction(IncV); 1068 return PN; 1069 } 1070 } 1071 1072 // Save the original insertion point so we can restore it when we're done. 1073 BuilderType::InsertPointGuard Guard(Builder); 1074 1075 // Another AddRec may need to be recursively expanded below. For example, if 1076 // this AddRec is quadratic, the StepV may itself be an AddRec in this 1077 // loop. Remove this loop from the PostIncLoops set before expanding such 1078 // AddRecs. Otherwise, we cannot find a valid position for the step 1079 // (i.e. StepV can never dominate its loop header). Ideally, we could do 1080 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 1081 // so it's not worth implementing SmallPtrSet::swap. 1082 PostIncLoopSet SavedPostIncLoops = PostIncLoops; 1083 PostIncLoops.clear(); 1084 1085 // Expand code for the start value. 1086 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 1087 L->getHeader()->begin()); 1088 1089 // StartV must be hoisted into L's preheader to dominate the new phi. 1090 assert(!isa<Instruction>(StartV) || 1091 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(), 1092 L->getHeader())); 1093 1094 // Expand code for the step value. Do this before creating the PHI so that PHI 1095 // reuse code doesn't see an incomplete PHI. 1096 const SCEV *Step = Normalized->getStepRecurrence(SE); 1097 // If the stride is negative, insert a sub instead of an add for the increment 1098 // (unless it's a constant, because subtracts of constants are canonicalized 1099 // to adds). 1100 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1101 if (useSubtract) 1102 Step = SE.getNegativeSCEV(Step); 1103 // Expand the step somewhere that dominates the loop header. 1104 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1105 1106 // Create the PHI. 1107 BasicBlock *Header = L->getHeader(); 1108 Builder.SetInsertPoint(Header, Header->begin()); 1109 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1110 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 1111 Twine(IVName) + ".iv"); 1112 rememberInstruction(PN); 1113 1114 // Create the step instructions and populate the PHI. 1115 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1116 BasicBlock *Pred = *HPI; 1117 1118 // Add a start value. 1119 if (!L->contains(Pred)) { 1120 PN->addIncoming(StartV, Pred); 1121 continue; 1122 } 1123 1124 // Create a step value and add it to the PHI. 1125 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 1126 // instructions at IVIncInsertPos. 1127 Instruction *InsertPos = L == IVIncInsertLoop ? 1128 IVIncInsertPos : Pred->getTerminator(); 1129 Builder.SetInsertPoint(InsertPos); 1130 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1131 if (isa<OverflowingBinaryOperator>(IncV)) { 1132 if (Normalized->getNoWrapFlags(SCEV::FlagNUW)) 1133 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap(); 1134 if (Normalized->getNoWrapFlags(SCEV::FlagNSW)) 1135 cast<BinaryOperator>(IncV)->setHasNoSignedWrap(); 1136 } 1137 PN->addIncoming(IncV, Pred); 1138 } 1139 1140 // After expanding subexpressions, restore the PostIncLoops set so the caller 1141 // can ensure that IVIncrement dominates the current uses. 1142 PostIncLoops = SavedPostIncLoops; 1143 1144 // Remember this PHI, even in post-inc mode. 1145 InsertedValues.insert(PN); 1146 1147 return PN; 1148} 1149 1150Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 1151 Type *STy = S->getType(); 1152 Type *IntTy = SE.getEffectiveSCEVType(STy); 1153 const Loop *L = S->getLoop(); 1154 1155 // Determine a normalized form of this expression, which is the expression 1156 // before any post-inc adjustment is made. 1157 const SCEVAddRecExpr *Normalized = S; 1158 if (PostIncLoops.count(L)) { 1159 PostIncLoopSet Loops; 1160 Loops.insert(L); 1161 Normalized = 1162 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0, 1163 Loops, SE, *SE.DT)); 1164 } 1165 1166 // Strip off any non-loop-dominating component from the addrec start. 1167 const SCEV *Start = Normalized->getStart(); 1168 const SCEV *PostLoopOffset = 0; 1169 if (!SE.properlyDominates(Start, L->getHeader())) { 1170 PostLoopOffset = Start; 1171 Start = SE.getConstant(Normalized->getType(), 0); 1172 Normalized = cast<SCEVAddRecExpr>( 1173 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 1174 Normalized->getLoop(), 1175 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1176 } 1177 1178 // Strip off any non-loop-dominating component from the addrec step. 1179 const SCEV *Step = Normalized->getStepRecurrence(SE); 1180 const SCEV *PostLoopScale = 0; 1181 if (!SE.dominates(Step, L->getHeader())) { 1182 PostLoopScale = Step; 1183 Step = SE.getConstant(Normalized->getType(), 1); 1184 Normalized = 1185 cast<SCEVAddRecExpr>(SE.getAddRecExpr( 1186 Start, Step, Normalized->getLoop(), 1187 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1188 } 1189 1190 // Expand the core addrec. If we need post-loop scaling, force it to 1191 // expand to an integer type to avoid the need for additional casting. 1192 Type *ExpandTy = PostLoopScale ? IntTy : STy; 1193 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy); 1194 1195 // Accommodate post-inc mode, if necessary. 1196 Value *Result; 1197 if (!PostIncLoops.count(L)) 1198 Result = PN; 1199 else { 1200 // In PostInc mode, use the post-incremented value. 1201 BasicBlock *LatchBlock = L->getLoopLatch(); 1202 assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1203 Result = PN->getIncomingValueForBlock(LatchBlock); 1204 1205 // For an expansion to use the postinc form, the client must call 1206 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 1207 // or dominated by IVIncInsertPos. 1208 if (isa<Instruction>(Result) 1209 && !SE.DT->dominates(cast<Instruction>(Result), 1210 Builder.GetInsertPoint())) { 1211 // The induction variable's postinc expansion does not dominate this use. 1212 // IVUsers tries to prevent this case, so it is rare. However, it can 1213 // happen when an IVUser outside the loop is not dominated by the latch 1214 // block. Adjusting IVIncInsertPos before expansion begins cannot handle 1215 // all cases. Consider a phi outide whose operand is replaced during 1216 // expansion with the value of the postinc user. Without fundamentally 1217 // changing the way postinc users are tracked, the only remedy is 1218 // inserting an extra IV increment. StepV might fold into PostLoopOffset, 1219 // but hopefully expandCodeFor handles that. 1220 bool useSubtract = 1221 !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1222 if (useSubtract) 1223 Step = SE.getNegativeSCEV(Step); 1224 Value *StepV; 1225 { 1226 // Expand the step somewhere that dominates the loop header. 1227 BuilderType::InsertPointGuard Guard(Builder); 1228 StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1229 } 1230 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1231 } 1232 } 1233 1234 // Re-apply any non-loop-dominating scale. 1235 if (PostLoopScale) { 1236 assert(S->isAffine() && "Can't linearly scale non-affine recurrences."); 1237 Result = InsertNoopCastOfTo(Result, IntTy); 1238 Result = Builder.CreateMul(Result, 1239 expandCodeFor(PostLoopScale, IntTy)); 1240 rememberInstruction(Result); 1241 } 1242 1243 // Re-apply any non-loop-dominating offset. 1244 if (PostLoopOffset) { 1245 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1246 const SCEV *const OffsetArray[1] = { PostLoopOffset }; 1247 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 1248 } else { 1249 Result = InsertNoopCastOfTo(Result, IntTy); 1250 Result = Builder.CreateAdd(Result, 1251 expandCodeFor(PostLoopOffset, IntTy)); 1252 rememberInstruction(Result); 1253 } 1254 } 1255 1256 return Result; 1257} 1258 1259Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1260 if (!CanonicalMode) return expandAddRecExprLiterally(S); 1261 1262 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1263 const Loop *L = S->getLoop(); 1264 1265 // First check for an existing canonical IV in a suitable type. 1266 PHINode *CanonicalIV = 0; 1267 if (PHINode *PN = L->getCanonicalInductionVariable()) 1268 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 1269 CanonicalIV = PN; 1270 1271 // Rewrite an AddRec in terms of the canonical induction variable, if 1272 // its type is more narrow. 1273 if (CanonicalIV && 1274 SE.getTypeSizeInBits(CanonicalIV->getType()) > 1275 SE.getTypeSizeInBits(Ty)) { 1276 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1277 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1278 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 1279 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 1280 S->getNoWrapFlags(SCEV::FlagNW))); 1281 BasicBlock::iterator NewInsertPt = 1282 llvm::next(BasicBlock::iterator(cast<Instruction>(V))); 1283 BuilderType::InsertPointGuard Guard(Builder); 1284 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) || 1285 isa<LandingPadInst>(NewInsertPt)) 1286 ++NewInsertPt; 1287 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0, 1288 NewInsertPt); 1289 return V; 1290 } 1291 1292 // {X,+,F} --> X + {0,+,F} 1293 if (!S->getStart()->isZero()) { 1294 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1295 NewOps[0] = SE.getConstant(Ty, 0); 1296 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, 1297 S->getNoWrapFlags(SCEV::FlagNW)); 1298 1299 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1300 // comments on expandAddToGEP for details. 1301 const SCEV *Base = S->getStart(); 1302 const SCEV *RestArray[1] = { Rest }; 1303 // Dig into the expression to find the pointer base for a GEP. 1304 ExposePointerBase(Base, RestArray[0], SE); 1305 // If we found a pointer, expand the AddRec with a GEP. 1306 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1307 // Make sure the Base isn't something exotic, such as a multiplied 1308 // or divided pointer value. In those cases, the result type isn't 1309 // actually a pointer type. 1310 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1311 Value *StartV = expand(Base); 1312 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1313 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 1314 } 1315 } 1316 1317 // Just do a normal add. Pre-expand the operands to suppress folding. 1318 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), 1319 SE.getUnknown(expand(Rest)))); 1320 } 1321 1322 // If we don't yet have a canonical IV, create one. 1323 if (!CanonicalIV) { 1324 // Create and insert the PHI node for the induction variable in the 1325 // specified loop. 1326 BasicBlock *Header = L->getHeader(); 1327 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1328 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 1329 Header->begin()); 1330 rememberInstruction(CanonicalIV); 1331 1332 SmallSet<BasicBlock *, 4> PredSeen; 1333 Constant *One = ConstantInt::get(Ty, 1); 1334 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1335 BasicBlock *HP = *HPI; 1336 if (!PredSeen.insert(HP)) 1337 continue; 1338 1339 if (L->contains(HP)) { 1340 // Insert a unit add instruction right before the terminator 1341 // corresponding to the back-edge. 1342 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 1343 "indvar.next", 1344 HP->getTerminator()); 1345 Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 1346 rememberInstruction(Add); 1347 CanonicalIV->addIncoming(Add, HP); 1348 } else { 1349 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 1350 } 1351 } 1352 } 1353 1354 // {0,+,1} --> Insert a canonical induction variable into the loop! 1355 if (S->isAffine() && S->getOperand(1)->isOne()) { 1356 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 1357 "IVs with types different from the canonical IV should " 1358 "already have been handled!"); 1359 return CanonicalIV; 1360 } 1361 1362 // {0,+,F} --> {0,+,1} * F 1363 1364 // If this is a simple linear addrec, emit it now as a special case. 1365 if (S->isAffine()) // {0,+,F} --> i*F 1366 return 1367 expand(SE.getTruncateOrNoop( 1368 SE.getMulExpr(SE.getUnknown(CanonicalIV), 1369 SE.getNoopOrAnyExtend(S->getOperand(1), 1370 CanonicalIV->getType())), 1371 Ty)); 1372 1373 // If this is a chain of recurrences, turn it into a closed form, using the 1374 // folders, then expandCodeFor the closed form. This allows the folders to 1375 // simplify the expression without having to build a bunch of special code 1376 // into this folder. 1377 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 1378 1379 // Promote S up to the canonical IV type, if the cast is foldable. 1380 const SCEV *NewS = S; 1381 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 1382 if (isa<SCEVAddRecExpr>(Ext)) 1383 NewS = Ext; 1384 1385 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1386 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 1387 1388 // Truncate the result down to the original type, if needed. 1389 const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1390 return expand(T); 1391} 1392 1393Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1394 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1395 Value *V = expandCodeFor(S->getOperand(), 1396 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1397 Value *I = Builder.CreateTrunc(V, Ty); 1398 rememberInstruction(I); 1399 return I; 1400} 1401 1402Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1403 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1404 Value *V = expandCodeFor(S->getOperand(), 1405 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1406 Value *I = Builder.CreateZExt(V, Ty); 1407 rememberInstruction(I); 1408 return I; 1409} 1410 1411Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1412 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1413 Value *V = expandCodeFor(S->getOperand(), 1414 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1415 Value *I = Builder.CreateSExt(V, Ty); 1416 rememberInstruction(I); 1417 return I; 1418} 1419 1420Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 1421 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1422 Type *Ty = LHS->getType(); 1423 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1424 // In the case of mixed integer and pointer types, do the 1425 // rest of the comparisons as integer. 1426 if (S->getOperand(i)->getType() != Ty) { 1427 Ty = SE.getEffectiveSCEVType(Ty); 1428 LHS = InsertNoopCastOfTo(LHS, Ty); 1429 } 1430 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1431 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 1432 rememberInstruction(ICmp); 1433 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1434 rememberInstruction(Sel); 1435 LHS = Sel; 1436 } 1437 // In the case of mixed integer and pointer types, cast the 1438 // final result back to the pointer type. 1439 if (LHS->getType() != S->getType()) 1440 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1441 return LHS; 1442} 1443 1444Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 1445 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1446 Type *Ty = LHS->getType(); 1447 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1448 // In the case of mixed integer and pointer types, do the 1449 // rest of the comparisons as integer. 1450 if (S->getOperand(i)->getType() != Ty) { 1451 Ty = SE.getEffectiveSCEVType(Ty); 1452 LHS = InsertNoopCastOfTo(LHS, Ty); 1453 } 1454 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1455 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 1456 rememberInstruction(ICmp); 1457 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1458 rememberInstruction(Sel); 1459 LHS = Sel; 1460 } 1461 // In the case of mixed integer and pointer types, cast the 1462 // final result back to the pointer type. 1463 if (LHS->getType() != S->getType()) 1464 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1465 return LHS; 1466} 1467 1468Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 1469 Instruction *IP) { 1470 Builder.SetInsertPoint(IP->getParent(), IP); 1471 return expandCodeFor(SH, Ty); 1472} 1473 1474Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 1475 // Expand the code for this SCEV. 1476 Value *V = expand(SH); 1477 if (Ty) { 1478 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 1479 "non-trivial casts should be done with the SCEVs directly!"); 1480 V = InsertNoopCastOfTo(V, Ty); 1481 } 1482 return V; 1483} 1484 1485Value *SCEVExpander::expand(const SCEV *S) { 1486 // Compute an insertion point for this SCEV object. Hoist the instructions 1487 // as far out in the loop nest as possible. 1488 Instruction *InsertPt = Builder.GetInsertPoint(); 1489 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ; 1490 L = L->getParentLoop()) 1491 if (SE.isLoopInvariant(S, L)) { 1492 if (!L) break; 1493 if (BasicBlock *Preheader = L->getLoopPreheader()) 1494 InsertPt = Preheader->getTerminator(); 1495 else { 1496 // LSR sets the insertion point for AddRec start/step values to the 1497 // block start to simplify value reuse, even though it's an invalid 1498 // position. SCEVExpander must correct for this in all cases. 1499 InsertPt = L->getHeader()->getFirstInsertionPt(); 1500 } 1501 } else { 1502 // If the SCEV is computable at this level, insert it into the header 1503 // after the PHIs (and after any other instructions that we've inserted 1504 // there) so that it is guaranteed to dominate any user inside the loop. 1505 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 1506 InsertPt = L->getHeader()->getFirstInsertionPt(); 1507 while (InsertPt != Builder.GetInsertPoint() 1508 && (isInsertedInstruction(InsertPt) 1509 || isa<DbgInfoIntrinsic>(InsertPt))) { 1510 InsertPt = llvm::next(BasicBlock::iterator(InsertPt)); 1511 } 1512 break; 1513 } 1514 1515 // Check to see if we already expanded this here. 1516 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator 1517 I = InsertedExpressions.find(std::make_pair(S, InsertPt)); 1518 if (I != InsertedExpressions.end()) 1519 return I->second; 1520 1521 BuilderType::InsertPointGuard Guard(Builder); 1522 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt); 1523 1524 // Expand the expression into instructions. 1525 Value *V = visit(S); 1526 1527 // Remember the expanded value for this SCEV at this location. 1528 // 1529 // This is independent of PostIncLoops. The mapped value simply materializes 1530 // the expression at this insertion point. If the mapped value happened to be 1531 // a postinc expansion, it could be reused by a non postinc user, but only if 1532 // its insertion point was already at the head of the loop. 1533 InsertedExpressions[std::make_pair(S, InsertPt)] = V; 1534 return V; 1535} 1536 1537void SCEVExpander::rememberInstruction(Value *I) { 1538 if (!PostIncLoops.empty()) 1539 InsertedPostIncValues.insert(I); 1540 else 1541 InsertedValues.insert(I); 1542} 1543 1544/// getOrInsertCanonicalInductionVariable - This method returns the 1545/// canonical induction variable of the specified type for the specified 1546/// loop (inserting one if there is none). A canonical induction variable 1547/// starts at zero and steps by one on each iteration. 1548PHINode * 1549SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1550 Type *Ty) { 1551 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1552 1553 // Build a SCEV for {0,+,1}<L>. 1554 // Conservatively use FlagAnyWrap for now. 1555 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 1556 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 1557 1558 // Emit code for it. 1559 BuilderType::InsertPointGuard Guard(Builder); 1560 PHINode *V = cast<PHINode>(expandCodeFor(H, 0, L->getHeader()->begin())); 1561 1562 return V; 1563} 1564 1565/// Sort values by integer width for replaceCongruentIVs. 1566static bool width_descending(Value *lhs, Value *rhs) { 1567 // Put pointers at the back and make sure pointer < pointer = false. 1568 if (!lhs->getType()->isIntegerTy() || !rhs->getType()->isIntegerTy()) 1569 return rhs->getType()->isIntegerTy() && !lhs->getType()->isIntegerTy(); 1570 return rhs->getType()->getPrimitiveSizeInBits() 1571 < lhs->getType()->getPrimitiveSizeInBits(); 1572} 1573 1574/// replaceCongruentIVs - Check for congruent phis in this loop header and 1575/// replace them with their most canonical representative. Return the number of 1576/// phis eliminated. 1577/// 1578/// This does not depend on any SCEVExpander state but should be used in 1579/// the same context that SCEVExpander is used. 1580unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 1581 SmallVectorImpl<WeakVH> &DeadInsts, 1582 const TargetTransformInfo *TTI) { 1583 // Find integer phis in order of increasing width. 1584 SmallVector<PHINode*, 8> Phis; 1585 for (BasicBlock::iterator I = L->getHeader()->begin(); 1586 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 1587 Phis.push_back(Phi); 1588 } 1589 if (TTI) 1590 std::sort(Phis.begin(), Phis.end(), width_descending); 1591 1592 unsigned NumElim = 0; 1593 DenseMap<const SCEV *, PHINode *> ExprToIVMap; 1594 // Process phis from wide to narrow. Mapping wide phis to the their truncation 1595 // so narrow phis can reuse them. 1596 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(), 1597 PEnd = Phis.end(); PIter != PEnd; ++PIter) { 1598 PHINode *Phi = *PIter; 1599 1600 // Fold constant phis. They may be congruent to other constant phis and 1601 // would confuse the logic below that expects proper IVs. 1602 if (Value *V = Phi->hasConstantValue()) { 1603 Phi->replaceAllUsesWith(V); 1604 DeadInsts.push_back(Phi); 1605 ++NumElim; 1606 DEBUG_WITH_TYPE(DebugType, dbgs() 1607 << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 1608 continue; 1609 } 1610 1611 if (!SE.isSCEVable(Phi->getType())) 1612 continue; 1613 1614 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 1615 if (!OrigPhiRef) { 1616 OrigPhiRef = Phi; 1617 if (Phi->getType()->isIntegerTy() && TTI 1618 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 1619 // This phi can be freely truncated to the narrowest phi type. Map the 1620 // truncated expression to it so it will be reused for narrow types. 1621 const SCEV *TruncExpr = 1622 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 1623 ExprToIVMap[TruncExpr] = Phi; 1624 } 1625 continue; 1626 } 1627 1628 // Replacing a pointer phi with an integer phi or vice-versa doesn't make 1629 // sense. 1630 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 1631 continue; 1632 1633 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1634 Instruction *OrigInc = 1635 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 1636 Instruction *IsomorphicInc = 1637 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 1638 1639 // If this phi has the same width but is more canonical, replace the 1640 // original with it. As part of the "more canonical" determination, 1641 // respect a prior decision to use an IV chain. 1642 if (OrigPhiRef->getType() == Phi->getType() 1643 && !(ChainedPhis.count(Phi) 1644 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) 1645 && (ChainedPhis.count(Phi) 1646 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 1647 std::swap(OrigPhiRef, Phi); 1648 std::swap(OrigInc, IsomorphicInc); 1649 } 1650 // Replacing the congruent phi is sufficient because acyclic redundancy 1651 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves 1652 // that a phi is congruent, it's often the head of an IV user cycle that 1653 // is isomorphic with the original phi. It's worth eagerly cleaning up the 1654 // common case of a single IV increment so that DeleteDeadPHIs can remove 1655 // cycles that had postinc uses. 1656 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc), 1657 IsomorphicInc->getType()); 1658 if (OrigInc != IsomorphicInc 1659 && TruncExpr == SE.getSCEV(IsomorphicInc) 1660 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc)) 1661 || hoistIVInc(OrigInc, IsomorphicInc))) { 1662 DEBUG_WITH_TYPE(DebugType, dbgs() 1663 << "INDVARS: Eliminated congruent iv.inc: " 1664 << *IsomorphicInc << '\n'); 1665 Value *NewInc = OrigInc; 1666 if (OrigInc->getType() != IsomorphicInc->getType()) { 1667 Instruction *IP = isa<PHINode>(OrigInc) 1668 ? (Instruction*)L->getHeader()->getFirstInsertionPt() 1669 : OrigInc->getNextNode(); 1670 IRBuilder<> Builder(IP); 1671 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); 1672 NewInc = Builder. 1673 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName); 1674 } 1675 IsomorphicInc->replaceAllUsesWith(NewInc); 1676 DeadInsts.push_back(IsomorphicInc); 1677 } 1678 } 1679 DEBUG_WITH_TYPE(DebugType, dbgs() 1680 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n'); 1681 ++NumElim; 1682 Value *NewIV = OrigPhiRef; 1683 if (OrigPhiRef->getType() != Phi->getType()) { 1684 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt()); 1685 Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); 1686 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName); 1687 } 1688 Phi->replaceAllUsesWith(NewIV); 1689 DeadInsts.push_back(Phi); 1690 } 1691 return NumElim; 1692} 1693 1694namespace { 1695// Search for a SCEV subexpression that is not safe to expand. Any expression 1696// that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely 1697// UDiv expressions. We don't know if the UDiv is derived from an IR divide 1698// instruction, but the important thing is that we prove the denominator is 1699// nonzero before expansion. 1700// 1701// IVUsers already checks that IV-derived expressions are safe. So this check is 1702// only needed when the expression includes some subexpression that is not IV 1703// derived. 1704// 1705// Currently, we only allow division by a nonzero constant here. If this is 1706// inadequate, we could easily allow division by SCEVUnknown by using 1707// ValueTracking to check isKnownNonZero(). 1708// 1709// We cannot generally expand recurrences unless the step dominates the loop 1710// header. The expander handles the special case of affine recurrences by 1711// scaling the recurrence outside the loop, but this technique isn't generally 1712// applicable. Expanding a nested recurrence outside a loop requires computing 1713// binomial coefficients. This could be done, but the recurrence has to be in a 1714// perfectly reduced form, which can't be guaranteed. 1715struct SCEVFindUnsafe { 1716 ScalarEvolution &SE; 1717 bool IsUnsafe; 1718 1719 SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {} 1720 1721 bool follow(const SCEV *S) { 1722 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 1723 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS()); 1724 if (!SC || SC->getValue()->isZero()) { 1725 IsUnsafe = true; 1726 return false; 1727 } 1728 } 1729 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 1730 const SCEV *Step = AR->getStepRecurrence(SE); 1731 if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) { 1732 IsUnsafe = true; 1733 return false; 1734 } 1735 } 1736 return true; 1737 } 1738 bool isDone() const { return IsUnsafe; } 1739}; 1740} 1741 1742namespace llvm { 1743bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) { 1744 SCEVFindUnsafe Search(SE); 1745 visitAll(S, Search); 1746 return !Search.IsUnsafe; 1747} 1748} 1749