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/STLExtras.h" 18#include "llvm/ADT/SmallSet.h" 19#include "llvm/Analysis/InstructionSimplify.h" 20#include "llvm/Analysis/LoopInfo.h" 21#include "llvm/Analysis/TargetTransformInfo.h" 22#include "llvm/IR/DataLayout.h" 23#include "llvm/IR/Dominators.h" 24#include "llvm/IR/IntrinsicInst.h" 25#include "llvm/IR/LLVMContext.h" 26#include "llvm/Support/Debug.h" 27 28using namespace llvm; 29 30/// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, 31/// reusing an existing cast if a suitable one exists, moving an existing 32/// cast if a suitable one exists but isn't in the right place, or 33/// creating a new one. 34Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, 35 Instruction::CastOps Op, 36 BasicBlock::iterator IP) { 37 // This function must be called with the builder having a valid insertion 38 // point. It doesn't need to be the actual IP where the uses of the returned 39 // cast will be added, but it must dominate such IP. 40 // We use this precondition to produce a cast that will dominate all its 41 // uses. In particular, this is crucial for the case where the builder's 42 // insertion point *is* the point where we were asked to put the cast. 43 // Since we don't know the builder's insertion point is actually 44 // where the uses will be added (only that it dominates it), we are 45 // not allowed to move it. 46 BasicBlock::iterator BIP = Builder.GetInsertPoint(); 47 48 Instruction *Ret = nullptr; 49 50 // Check to see if there is already a cast! 51 for (User *U : V->users()) 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 // Create a new cast. 73 if (!Ret) 74 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP); 75 76 // We assert at the end of the function since IP might point to an 77 // instruction with different dominance properties than a cast 78 // (an invoke for example) and not dominate BIP (but the cast does). 79 assert(SE.DT->dominates(Ret, BIP)); 80 81 rememberInstruction(Ret); 82 return Ret; 83} 84 85/// InsertNoopCastOfTo - Insert a cast of V to the specified type, 86/// which must be possible with a noop cast, doing what we can to share 87/// the casts. 88Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 89 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 90 assert((Op == Instruction::BitCast || 91 Op == Instruction::PtrToInt || 92 Op == Instruction::IntToPtr) && 93 "InsertNoopCastOfTo cannot perform non-noop casts!"); 94 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 95 "InsertNoopCastOfTo cannot change sizes!"); 96 97 // Short-circuit unnecessary bitcasts. 98 if (Op == Instruction::BitCast) { 99 if (V->getType() == Ty) 100 return V; 101 if (CastInst *CI = dyn_cast<CastInst>(V)) { 102 if (CI->getOperand(0)->getType() == Ty) 103 return CI->getOperand(0); 104 } 105 } 106 // Short-circuit unnecessary inttoptr<->ptrtoint casts. 107 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 108 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 109 if (CastInst *CI = dyn_cast<CastInst>(V)) 110 if ((CI->getOpcode() == Instruction::PtrToInt || 111 CI->getOpcode() == Instruction::IntToPtr) && 112 SE.getTypeSizeInBits(CI->getType()) == 113 SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 114 return CI->getOperand(0); 115 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 116 if ((CE->getOpcode() == Instruction::PtrToInt || 117 CE->getOpcode() == Instruction::IntToPtr) && 118 SE.getTypeSizeInBits(CE->getType()) == 119 SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 120 return CE->getOperand(0); 121 } 122 123 // Fold a cast of a constant. 124 if (Constant *C = dyn_cast<Constant>(V)) 125 return ConstantExpr::getCast(Op, C, Ty); 126 127 // Cast the argument at the beginning of the entry block, after 128 // any bitcasts of other arguments. 129 if (Argument *A = dyn_cast<Argument>(V)) { 130 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 131 while ((isa<BitCastInst>(IP) && 132 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 133 cast<BitCastInst>(IP)->getOperand(0) != A) || 134 isa<DbgInfoIntrinsic>(IP) || 135 isa<LandingPadInst>(IP)) 136 ++IP; 137 return ReuseOrCreateCast(A, Ty, Op, IP); 138 } 139 140 // Cast the instruction immediately after the instruction. 141 Instruction *I = cast<Instruction>(V); 142 BasicBlock::iterator IP = I; ++IP; 143 if (InvokeInst *II = dyn_cast<InvokeInst>(I)) 144 IP = II->getNormalDest()->begin(); 145 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP)) 146 ++IP; 147 return ReuseOrCreateCast(I, Ty, Op, IP); 148} 149 150/// InsertBinop - Insert the specified binary operator, doing a small amount 151/// of work to avoid inserting an obviously redundant operation. 152Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 153 Value *LHS, Value *RHS) { 154 // Fold a binop with constant operands. 155 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 156 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 157 return ConstantExpr::get(Opcode, CLHS, CRHS); 158 159 // Do a quick scan to see if we have this binop nearby. If so, reuse it. 160 unsigned ScanLimit = 6; 161 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 162 // Scanning starts from the last instruction before the insertion point. 163 BasicBlock::iterator IP = Builder.GetInsertPoint(); 164 if (IP != BlockBegin) { 165 --IP; 166 for (; ScanLimit; --IP, --ScanLimit) { 167 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 168 // generated code. 169 if (isa<DbgInfoIntrinsic>(IP)) 170 ScanLimit++; 171 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 172 IP->getOperand(1) == RHS) 173 return IP; 174 if (IP == BlockBegin) break; 175 } 176 } 177 178 // Save the original insertion point so we can restore it when we're done. 179 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); 180 BuilderType::InsertPointGuard Guard(Builder); 181 182 // Move the insertion point out of as many loops as we can. 183 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 184 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 185 BasicBlock *Preheader = L->getLoopPreheader(); 186 if (!Preheader) break; 187 188 // Ok, move up a level. 189 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 190 } 191 192 // If we haven't found this binop, insert it. 193 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 194 BO->setDebugLoc(Loc); 195 rememberInstruction(BO); 196 197 return BO; 198} 199 200/// FactorOutConstant - Test if S is divisible by Factor, using signed 201/// division. If so, update S with Factor divided out and return true. 202/// S need not be evenly divisible if a reasonable remainder can be 203/// computed. 204/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 205/// unnecessary; in its place, just signed-divide Ops[i] by the scale and 206/// check to see if the divide was folded. 207static bool FactorOutConstant(const SCEV *&S, 208 const SCEV *&Remainder, 209 const SCEV *Factor, 210 ScalarEvolution &SE, 211 const DataLayout *DL) { 212 // Everything is divisible by one. 213 if (Factor->isOne()) 214 return true; 215 216 // x/x == 1. 217 if (S == Factor) { 218 S = SE.getConstant(S->getType(), 1); 219 return true; 220 } 221 222 // For a Constant, check for a multiple of the given factor. 223 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 224 // 0/x == 0. 225 if (C->isZero()) 226 return true; 227 // Check for divisibility. 228 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 229 ConstantInt *CI = 230 ConstantInt::get(SE.getContext(), 231 C->getValue()->getValue().sdiv( 232 FC->getValue()->getValue())); 233 // If the quotient is zero and the remainder is non-zero, reject 234 // the value at this scale. It will be considered for subsequent 235 // smaller scales. 236 if (!CI->isZero()) { 237 const SCEV *Div = SE.getConstant(CI); 238 S = Div; 239 Remainder = 240 SE.getAddExpr(Remainder, 241 SE.getConstant(C->getValue()->getValue().srem( 242 FC->getValue()->getValue()))); 243 return true; 244 } 245 } 246 } 247 248 // In a Mul, check if there is a constant operand which is a multiple 249 // of the given factor. 250 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 251 if (DL) { 252 // With DataLayout, the size is known. Check if there is a constant 253 // operand which is a multiple of the given factor. If so, we can 254 // factor it. 255 const SCEVConstant *FC = cast<SCEVConstant>(Factor); 256 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 257 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) { 258 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 259 NewMulOps[0] = 260 SE.getConstant(C->getValue()->getValue().sdiv( 261 FC->getValue()->getValue())); 262 S = SE.getMulExpr(NewMulOps); 263 return true; 264 } 265 } else { 266 // Without DataLayout, check if Factor can be factored out of any of the 267 // Mul's operands. If so, we can just remove it. 268 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) { 269 const SCEV *SOp = M->getOperand(i); 270 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0); 271 if (FactorOutConstant(SOp, Remainder, Factor, SE, DL) && 272 Remainder->isZero()) { 273 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 274 NewMulOps[i] = SOp; 275 S = SE.getMulExpr(NewMulOps); 276 return true; 277 } 278 } 279 } 280 } 281 282 // In an AddRec, check if both start and step are divisible. 283 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 284 const SCEV *Step = A->getStepRecurrence(SE); 285 const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 286 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL)) 287 return false; 288 if (!StepRem->isZero()) 289 return false; 290 const SCEV *Start = A->getStart(); 291 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL)) 292 return false; 293 S = SE.getAddRecExpr(Start, Step, A->getLoop(), 294 A->getNoWrapFlags(SCEV::FlagNW)); 295 return true; 296 } 297 298 return false; 299} 300 301/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 302/// is the number of SCEVAddRecExprs present, which are kept at the end of 303/// the list. 304/// 305static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 306 Type *Ty, 307 ScalarEvolution &SE) { 308 unsigned NumAddRecs = 0; 309 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 310 ++NumAddRecs; 311 // Group Ops into non-addrecs and addrecs. 312 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 313 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 314 // Let ScalarEvolution sort and simplify the non-addrecs list. 315 const SCEV *Sum = NoAddRecs.empty() ? 316 SE.getConstant(Ty, 0) : 317 SE.getAddExpr(NoAddRecs); 318 // If it returned an add, use the operands. Otherwise it simplified 319 // the sum into a single value, so just use that. 320 Ops.clear(); 321 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 322 Ops.append(Add->op_begin(), Add->op_end()); 323 else if (!Sum->isZero()) 324 Ops.push_back(Sum); 325 // Then append the addrecs. 326 Ops.append(AddRecs.begin(), AddRecs.end()); 327} 328 329/// SplitAddRecs - Flatten a list of add operands, moving addrec start values 330/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 331/// This helps expose more opportunities for folding parts of the expressions 332/// into GEP indices. 333/// 334static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 335 Type *Ty, 336 ScalarEvolution &SE) { 337 // Find the addrecs. 338 SmallVector<const SCEV *, 8> AddRecs; 339 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 340 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 341 const SCEV *Start = A->getStart(); 342 if (Start->isZero()) break; 343 const SCEV *Zero = SE.getConstant(Ty, 0); 344 AddRecs.push_back(SE.getAddRecExpr(Zero, 345 A->getStepRecurrence(SE), 346 A->getLoop(), 347 A->getNoWrapFlags(SCEV::FlagNW))); 348 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 349 Ops[i] = Zero; 350 Ops.append(Add->op_begin(), Add->op_end()); 351 e += Add->getNumOperands(); 352 } else { 353 Ops[i] = Start; 354 } 355 } 356 if (!AddRecs.empty()) { 357 // Add the addrecs onto the end of the list. 358 Ops.append(AddRecs.begin(), AddRecs.end()); 359 // Resort the operand list, moving any constants to the front. 360 SimplifyAddOperands(Ops, Ty, SE); 361 } 362} 363 364/// expandAddToGEP - Expand an addition expression with a pointer type into 365/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 366/// BasicAliasAnalysis and other passes analyze the result. See the rules 367/// for getelementptr vs. inttoptr in 368/// http://llvm.org/docs/LangRef.html#pointeraliasing 369/// for details. 370/// 371/// Design note: The correctness of using getelementptr here depends on 372/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 373/// they may introduce pointer arithmetic which may not be safely converted 374/// into getelementptr. 375/// 376/// Design note: It might seem desirable for this function to be more 377/// loop-aware. If some of the indices are loop-invariant while others 378/// aren't, it might seem desirable to emit multiple GEPs, keeping the 379/// loop-invariant portions of the overall computation outside the loop. 380/// However, there are a few reasons this is not done here. Hoisting simple 381/// arithmetic is a low-level optimization that often isn't very 382/// important until late in the optimization process. In fact, passes 383/// like InstructionCombining will combine GEPs, even if it means 384/// pushing loop-invariant computation down into loops, so even if the 385/// GEPs were split here, the work would quickly be undone. The 386/// LoopStrengthReduction pass, which is usually run quite late (and 387/// after the last InstructionCombining pass), takes care of hoisting 388/// loop-invariant portions of expressions, after considering what 389/// can be folded using target addressing modes. 390/// 391Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 392 const SCEV *const *op_end, 393 PointerType *PTy, 394 Type *Ty, 395 Value *V) { 396 Type *ElTy = PTy->getElementType(); 397 SmallVector<Value *, 4> GepIndices; 398 SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 399 bool AnyNonZeroIndices = false; 400 401 // Split AddRecs up into parts as either of the parts may be usable 402 // without the other. 403 SplitAddRecs(Ops, Ty, SE); 404 405 Type *IntPtrTy = SE.DL 406 ? SE.DL->getIntPtrType(PTy) 407 : Type::getInt64Ty(PTy->getContext()); 408 409 // Descend down the pointer's type and attempt to convert the other 410 // operands into GEP indices, at each level. The first index in a GEP 411 // indexes into the array implied by the pointer operand; the rest of 412 // the indices index into the element or field type selected by the 413 // preceding index. 414 for (;;) { 415 // If the scale size is not 0, attempt to factor out a scale for 416 // array indexing. 417 SmallVector<const SCEV *, 8> ScaledOps; 418 if (ElTy->isSized()) { 419 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy); 420 if (!ElSize->isZero()) { 421 SmallVector<const SCEV *, 8> NewOps; 422 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 423 const SCEV *Op = Ops[i]; 424 const SCEV *Remainder = SE.getConstant(Ty, 0); 425 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.DL)) { 426 // Op now has ElSize factored out. 427 ScaledOps.push_back(Op); 428 if (!Remainder->isZero()) 429 NewOps.push_back(Remainder); 430 AnyNonZeroIndices = true; 431 } else { 432 // The operand was not divisible, so add it to the list of operands 433 // we'll scan next iteration. 434 NewOps.push_back(Ops[i]); 435 } 436 } 437 // If we made any changes, update Ops. 438 if (!ScaledOps.empty()) { 439 Ops = NewOps; 440 SimplifyAddOperands(Ops, Ty, SE); 441 } 442 } 443 } 444 445 // Record the scaled array index for this level of the type. If 446 // we didn't find any operands that could be factored, tentatively 447 // assume that element zero was selected (since the zero offset 448 // would obviously be folded away). 449 Value *Scaled = ScaledOps.empty() ? 450 Constant::getNullValue(Ty) : 451 expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 452 GepIndices.push_back(Scaled); 453 454 // Collect struct field index operands. 455 while (StructType *STy = dyn_cast<StructType>(ElTy)) { 456 bool FoundFieldNo = false; 457 // An empty struct has no fields. 458 if (STy->getNumElements() == 0) break; 459 if (SE.DL) { 460 // With DataLayout, field offsets are known. See if a constant offset 461 // falls within any of the struct fields. 462 if (Ops.empty()) break; 463 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 464 if (SE.getTypeSizeInBits(C->getType()) <= 64) { 465 const StructLayout &SL = *SE.DL->getStructLayout(STy); 466 uint64_t FullOffset = C->getValue()->getZExtValue(); 467 if (FullOffset < SL.getSizeInBytes()) { 468 unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 469 GepIndices.push_back( 470 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 471 ElTy = STy->getTypeAtIndex(ElIdx); 472 Ops[0] = 473 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 474 AnyNonZeroIndices = true; 475 FoundFieldNo = true; 476 } 477 } 478 } else { 479 // Without DataLayout, just check for an offsetof expression of the 480 // appropriate struct type. 481 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 482 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) { 483 Type *CTy; 484 Constant *FieldNo; 485 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) { 486 GepIndices.push_back(FieldNo); 487 ElTy = 488 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue()); 489 Ops[i] = SE.getConstant(Ty, 0); 490 AnyNonZeroIndices = true; 491 FoundFieldNo = true; 492 break; 493 } 494 } 495 } 496 // If no struct field offsets were found, tentatively assume that 497 // field zero was selected (since the zero offset would obviously 498 // be folded away). 499 if (!FoundFieldNo) { 500 ElTy = STy->getTypeAtIndex(0u); 501 GepIndices.push_back( 502 Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 503 } 504 } 505 506 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 507 ElTy = ATy->getElementType(); 508 else 509 break; 510 } 511 512 // If none of the operands were convertible to proper GEP indices, cast 513 // the base to i8* and do an ugly getelementptr with that. It's still 514 // better than ptrtoint+arithmetic+inttoptr at least. 515 if (!AnyNonZeroIndices) { 516 // Cast the base to i8*. 517 V = InsertNoopCastOfTo(V, 518 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 519 520 assert(!isa<Instruction>(V) || 521 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint())); 522 523 // Expand the operands for a plain byte offset. 524 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 525 526 // Fold a GEP with constant operands. 527 if (Constant *CLHS = dyn_cast<Constant>(V)) 528 if (Constant *CRHS = dyn_cast<Constant>(Idx)) 529 return ConstantExpr::getGetElementPtr(CLHS, CRHS); 530 531 // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 532 unsigned ScanLimit = 6; 533 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 534 // Scanning starts from the last instruction before the insertion point. 535 BasicBlock::iterator IP = Builder.GetInsertPoint(); 536 if (IP != BlockBegin) { 537 --IP; 538 for (; ScanLimit; --IP, --ScanLimit) { 539 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 540 // generated code. 541 if (isa<DbgInfoIntrinsic>(IP)) 542 ScanLimit++; 543 if (IP->getOpcode() == Instruction::GetElementPtr && 544 IP->getOperand(0) == V && IP->getOperand(1) == Idx) 545 return IP; 546 if (IP == BlockBegin) break; 547 } 548 } 549 550 // Save the original insertion point so we can restore it when we're done. 551 BuilderType::InsertPointGuard Guard(Builder); 552 553 // Move the insertion point out of as many loops as we can. 554 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 555 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 556 BasicBlock *Preheader = L->getLoopPreheader(); 557 if (!Preheader) break; 558 559 // Ok, move up a level. 560 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 561 } 562 563 // Emit a GEP. 564 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep"); 565 rememberInstruction(GEP); 566 567 return GEP; 568 } 569 570 // Save the original insertion point so we can restore it when we're done. 571 BuilderType::InsertPoint SaveInsertPt = Builder.saveIP(); 572 573 // Move the insertion point out of as many loops as we can. 574 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 575 if (!L->isLoopInvariant(V)) break; 576 577 bool AnyIndexNotLoopInvariant = false; 578 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(), 579 E = GepIndices.end(); I != E; ++I) 580 if (!L->isLoopInvariant(*I)) { 581 AnyIndexNotLoopInvariant = true; 582 break; 583 } 584 if (AnyIndexNotLoopInvariant) 585 break; 586 587 BasicBlock *Preheader = L->getLoopPreheader(); 588 if (!Preheader) break; 589 590 // Ok, move up a level. 591 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 592 } 593 594 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 595 // because ScalarEvolution may have changed the address arithmetic to 596 // compute a value which is beyond the end of the allocated object. 597 Value *Casted = V; 598 if (V->getType() != PTy) 599 Casted = InsertNoopCastOfTo(Casted, PTy); 600 Value *GEP = Builder.CreateGEP(Casted, 601 GepIndices, 602 "scevgep"); 603 Ops.push_back(SE.getUnknown(GEP)); 604 rememberInstruction(GEP); 605 606 // Restore the original insert point. 607 Builder.restoreIP(SaveInsertPt); 608 609 return expand(SE.getAddExpr(Ops)); 610} 611 612/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 613/// SCEV expansion. If they are nested, this is the most nested. If they are 614/// neighboring, pick the later. 615static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 616 DominatorTree &DT) { 617 if (!A) return B; 618 if (!B) return A; 619 if (A->contains(B)) return B; 620 if (B->contains(A)) return A; 621 if (DT.dominates(A->getHeader(), B->getHeader())) return B; 622 if (DT.dominates(B->getHeader(), A->getHeader())) return A; 623 return A; // Arbitrarily break the tie. 624} 625 626/// getRelevantLoop - Get the most relevant loop associated with the given 627/// expression, according to PickMostRelevantLoop. 628const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 629 // Test whether we've already computed the most relevant loop for this SCEV. 630 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair = 631 RelevantLoops.insert(std::make_pair(S, nullptr)); 632 if (!Pair.second) 633 return Pair.first->second; 634 635 if (isa<SCEVConstant>(S)) 636 // A constant has no relevant loops. 637 return nullptr; 638 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 639 if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 640 return Pair.first->second = SE.LI->getLoopFor(I->getParent()); 641 // A non-instruction has no relevant loops. 642 return nullptr; 643 } 644 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 645 const Loop *L = nullptr; 646 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 647 L = AR->getLoop(); 648 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end(); 649 I != E; ++I) 650 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT); 651 return RelevantLoops[N] = L; 652 } 653 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 654 const Loop *Result = getRelevantLoop(C->getOperand()); 655 return RelevantLoops[C] = Result; 656 } 657 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 658 const Loop *Result = 659 PickMostRelevantLoop(getRelevantLoop(D->getLHS()), 660 getRelevantLoop(D->getRHS()), 661 *SE.DT); 662 return RelevantLoops[D] = Result; 663 } 664 llvm_unreachable("Unexpected SCEV type!"); 665} 666 667namespace { 668 669/// LoopCompare - Compare loops by PickMostRelevantLoop. 670class LoopCompare { 671 DominatorTree &DT; 672public: 673 explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 674 675 bool operator()(std::pair<const Loop *, const SCEV *> LHS, 676 std::pair<const Loop *, const SCEV *> RHS) const { 677 // Keep pointer operands sorted at the end. 678 if (LHS.second->getType()->isPointerTy() != 679 RHS.second->getType()->isPointerTy()) 680 return LHS.second->getType()->isPointerTy(); 681 682 // Compare loops with PickMostRelevantLoop. 683 if (LHS.first != RHS.first) 684 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 685 686 // If one operand is a non-constant negative and the other is not, 687 // put the non-constant negative on the right so that a sub can 688 // be used instead of a negate and add. 689 if (LHS.second->isNonConstantNegative()) { 690 if (!RHS.second->isNonConstantNegative()) 691 return false; 692 } else if (RHS.second->isNonConstantNegative()) 693 return true; 694 695 // Otherwise they are equivalent according to this comparison. 696 return false; 697 } 698}; 699 700} 701 702Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 703 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 704 705 // Collect all the add operands in a loop, along with their associated loops. 706 // Iterate in reverse so that constants are emitted last, all else equal, and 707 // so that pointer operands are inserted first, which the code below relies on 708 // to form more involved GEPs. 709 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 710 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 711 E(S->op_begin()); I != E; ++I) 712 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 713 714 // Sort by loop. Use a stable sort so that constants follow non-constants and 715 // pointer operands precede non-pointer operands. 716 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 717 718 // Emit instructions to add all the operands. Hoist as much as possible 719 // out of loops, and form meaningful getelementptrs where possible. 720 Value *Sum = nullptr; 721 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 722 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 723 const Loop *CurLoop = I->first; 724 const SCEV *Op = I->second; 725 if (!Sum) { 726 // This is the first operand. Just expand it. 727 Sum = expand(Op); 728 ++I; 729 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 730 // The running sum expression is a pointer. Try to form a getelementptr 731 // at this level with that as the base. 732 SmallVector<const SCEV *, 4> NewOps; 733 for (; I != E && I->first == CurLoop; ++I) { 734 // If the operand is SCEVUnknown and not instructions, peek through 735 // it, to enable more of it to be folded into the GEP. 736 const SCEV *X = I->second; 737 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 738 if (!isa<Instruction>(U->getValue())) 739 X = SE.getSCEV(U->getValue()); 740 NewOps.push_back(X); 741 } 742 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 743 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 744 // The running sum is an integer, and there's a pointer at this level. 745 // Try to form a getelementptr. If the running sum is instructions, 746 // use a SCEVUnknown to avoid re-analyzing them. 747 SmallVector<const SCEV *, 4> NewOps; 748 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 749 SE.getSCEV(Sum)); 750 for (++I; I != E && I->first == CurLoop; ++I) 751 NewOps.push_back(I->second); 752 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 753 } else if (Op->isNonConstantNegative()) { 754 // Instead of doing a negate and add, just do a subtract. 755 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 756 Sum = InsertNoopCastOfTo(Sum, Ty); 757 Sum = InsertBinop(Instruction::Sub, Sum, W); 758 ++I; 759 } else { 760 // A simple add. 761 Value *W = expandCodeFor(Op, Ty); 762 Sum = InsertNoopCastOfTo(Sum, Ty); 763 // Canonicalize a constant to the RHS. 764 if (isa<Constant>(Sum)) std::swap(Sum, W); 765 Sum = InsertBinop(Instruction::Add, Sum, W); 766 ++I; 767 } 768 } 769 770 return Sum; 771} 772 773Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 774 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 775 776 // Collect all the mul operands in a loop, along with their associated loops. 777 // Iterate in reverse so that constants are emitted last, all else equal. 778 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 779 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 780 E(S->op_begin()); I != E; ++I) 781 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 782 783 // Sort by loop. Use a stable sort so that constants follow non-constants. 784 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 785 786 // Emit instructions to mul all the operands. Hoist as much as possible 787 // out of loops. 788 Value *Prod = nullptr; 789 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 790 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 791 const SCEV *Op = I->second; 792 if (!Prod) { 793 // This is the first operand. Just expand it. 794 Prod = expand(Op); 795 ++I; 796 } else if (Op->isAllOnesValue()) { 797 // Instead of doing a multiply by negative one, just do a negate. 798 Prod = InsertNoopCastOfTo(Prod, Ty); 799 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 800 ++I; 801 } else { 802 // A simple mul. 803 Value *W = expandCodeFor(Op, Ty); 804 Prod = InsertNoopCastOfTo(Prod, Ty); 805 // Canonicalize a constant to the RHS. 806 if (isa<Constant>(Prod)) std::swap(Prod, W); 807 Prod = InsertBinop(Instruction::Mul, Prod, W); 808 ++I; 809 } 810 } 811 812 return Prod; 813} 814 815Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 816 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 817 818 Value *LHS = expandCodeFor(S->getLHS(), Ty); 819 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 820 const APInt &RHS = SC->getValue()->getValue(); 821 if (RHS.isPowerOf2()) 822 return InsertBinop(Instruction::LShr, LHS, 823 ConstantInt::get(Ty, RHS.logBase2())); 824 } 825 826 Value *RHS = expandCodeFor(S->getRHS(), Ty); 827 return InsertBinop(Instruction::UDiv, LHS, RHS); 828} 829 830/// Move parts of Base into Rest to leave Base with the minimal 831/// expression that provides a pointer operand suitable for a 832/// GEP expansion. 833static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 834 ScalarEvolution &SE) { 835 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 836 Base = A->getStart(); 837 Rest = SE.getAddExpr(Rest, 838 SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 839 A->getStepRecurrence(SE), 840 A->getLoop(), 841 A->getNoWrapFlags(SCEV::FlagNW))); 842 } 843 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 844 Base = A->getOperand(A->getNumOperands()-1); 845 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 846 NewAddOps.back() = Rest; 847 Rest = SE.getAddExpr(NewAddOps); 848 ExposePointerBase(Base, Rest, SE); 849 } 850} 851 852/// Determine if this is a well-behaved chain of instructions leading back to 853/// the PHI. If so, it may be reused by expanded expressions. 854bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 855 const Loop *L) { 856 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 857 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 858 return false; 859 // If any of the operands don't dominate the insert position, bail. 860 // Addrec operands are always loop-invariant, so this can only happen 861 // if there are instructions which haven't been hoisted. 862 if (L == IVIncInsertLoop) { 863 for (User::op_iterator OI = IncV->op_begin()+1, 864 OE = IncV->op_end(); OI != OE; ++OI) 865 if (Instruction *OInst = dyn_cast<Instruction>(OI)) 866 if (!SE.DT->dominates(OInst, IVIncInsertPos)) 867 return false; 868 } 869 // Advance to the next instruction. 870 IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 871 if (!IncV) 872 return false; 873 874 if (IncV->mayHaveSideEffects()) 875 return false; 876 877 if (IncV != PN) 878 return true; 879 880 return isNormalAddRecExprPHI(PN, IncV, L); 881} 882 883/// getIVIncOperand returns an induction variable increment's induction 884/// variable operand. 885/// 886/// If allowScale is set, any type of GEP is allowed as long as the nonIV 887/// operands dominate InsertPos. 888/// 889/// If allowScale is not set, ensure that a GEP increment conforms to one of the 890/// simple patterns generated by getAddRecExprPHILiterally and 891/// expandAddtoGEP. If the pattern isn't recognized, return NULL. 892Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 893 Instruction *InsertPos, 894 bool allowScale) { 895 if (IncV == InsertPos) 896 return nullptr; 897 898 switch (IncV->getOpcode()) { 899 default: 900 return nullptr; 901 // Check for a simple Add/Sub or GEP of a loop invariant step. 902 case Instruction::Add: 903 case Instruction::Sub: { 904 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 905 if (!OInst || SE.DT->dominates(OInst, InsertPos)) 906 return dyn_cast<Instruction>(IncV->getOperand(0)); 907 return nullptr; 908 } 909 case Instruction::BitCast: 910 return dyn_cast<Instruction>(IncV->getOperand(0)); 911 case Instruction::GetElementPtr: 912 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end(); 913 I != E; ++I) { 914 if (isa<Constant>(*I)) 915 continue; 916 if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 917 if (!SE.DT->dominates(OInst, InsertPos)) 918 return nullptr; 919 } 920 if (allowScale) { 921 // allow any kind of GEP as long as it can be hoisted. 922 continue; 923 } 924 // This must be a pointer addition of constants (pretty), which is already 925 // handled, or some number of address-size elements (ugly). Ugly geps 926 // have 2 operands. i1* is used by the expander to represent an 927 // address-size element. 928 if (IncV->getNumOperands() != 2) 929 return nullptr; 930 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 931 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 932 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 933 return nullptr; 934 break; 935 } 936 return dyn_cast<Instruction>(IncV->getOperand(0)); 937 } 938} 939 940/// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 941/// it available to other uses in this loop. Recursively hoist any operands, 942/// until we reach a value that dominates InsertPos. 943bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 944 if (SE.DT->dominates(IncV, InsertPos)) 945 return true; 946 947 // InsertPos must itself dominate IncV so that IncV's new position satisfies 948 // its existing users. 949 if (isa<PHINode>(InsertPos) 950 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent())) 951 return false; 952 953 // Check that the chain of IV operands leading back to Phi can be hoisted. 954 SmallVector<Instruction*, 4> IVIncs; 955 for(;;) { 956 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 957 if (!Oper) 958 return false; 959 // IncV is safe to hoist. 960 IVIncs.push_back(IncV); 961 IncV = Oper; 962 if (SE.DT->dominates(IncV, InsertPos)) 963 break; 964 } 965 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(), 966 E = IVIncs.rend(); I != E; ++I) { 967 (*I)->moveBefore(InsertPos); 968 } 969 return true; 970} 971 972/// Determine if this cyclic phi is in a form that would have been generated by 973/// LSR. We don't care if the phi was actually expanded in this pass, as long 974/// as it is in a low-cost form, for example, no implied multiplication. This 975/// should match any patterns generated by getAddRecExprPHILiterally and 976/// expandAddtoGEP. 977bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 978 const Loop *L) { 979 for(Instruction *IVOper = IncV; 980 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 981 /*allowScale=*/false));) { 982 if (IVOper == PN) 983 return true; 984 } 985 return false; 986} 987 988/// expandIVInc - Expand an IV increment at Builder's current InsertPos. 989/// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 990/// need to materialize IV increments elsewhere to handle difficult situations. 991Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 992 Type *ExpandTy, Type *IntTy, 993 bool useSubtract) { 994 Value *IncV; 995 // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 996 if (ExpandTy->isPointerTy()) { 997 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 998 // If the step isn't constant, don't use an implicitly scaled GEP, because 999 // that would require a multiply inside the loop. 1000 if (!isa<ConstantInt>(StepV)) 1001 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 1002 GEPPtrTy->getAddressSpace()); 1003 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 1004 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 1005 if (IncV->getType() != PN->getType()) { 1006 IncV = Builder.CreateBitCast(IncV, PN->getType()); 1007 rememberInstruction(IncV); 1008 } 1009 } else { 1010 IncV = useSubtract ? 1011 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 1012 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 1013 rememberInstruction(IncV); 1014 } 1015 return IncV; 1016} 1017 1018/// \brief Hoist the addrec instruction chain rooted in the loop phi above the 1019/// position. This routine assumes that this is possible (has been checked). 1020static void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist, 1021 Instruction *Pos, PHINode *LoopPhi) { 1022 do { 1023 if (DT->dominates(InstToHoist, Pos)) 1024 break; 1025 // Make sure the increment is where we want it. But don't move it 1026 // down past a potential existing post-inc user. 1027 InstToHoist->moveBefore(Pos); 1028 Pos = InstToHoist; 1029 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0)); 1030 } while (InstToHoist != LoopPhi); 1031} 1032 1033/// \brief Check whether we can cheaply express the requested SCEV in terms of 1034/// the available PHI SCEV by truncation and/or invertion of the step. 1035static bool canBeCheaplyTransformed(ScalarEvolution &SE, 1036 const SCEVAddRecExpr *Phi, 1037 const SCEVAddRecExpr *Requested, 1038 bool &InvertStep) { 1039 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType()); 1040 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType()); 1041 1042 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth()) 1043 return false; 1044 1045 // Try truncate it if necessary. 1046 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy)); 1047 if (!Phi) 1048 return false; 1049 1050 // Check whether truncation will help. 1051 if (Phi == Requested) { 1052 InvertStep = false; 1053 return true; 1054 } 1055 1056 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}. 1057 if (SE.getAddExpr(Requested->getStart(), 1058 SE.getNegativeSCEV(Requested)) == Phi) { 1059 InvertStep = true; 1060 return true; 1061 } 1062 1063 return false; 1064} 1065 1066/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 1067/// the base addrec, which is the addrec without any non-loop-dominating 1068/// values, and return the PHI. 1069PHINode * 1070SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 1071 const Loop *L, 1072 Type *ExpandTy, 1073 Type *IntTy, 1074 Type *&TruncTy, 1075 bool &InvertStep) { 1076 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 1077 1078 // Reuse a previously-inserted PHI, if present. 1079 BasicBlock *LatchBlock = L->getLoopLatch(); 1080 if (LatchBlock) { 1081 PHINode *AddRecPhiMatch = nullptr; 1082 Instruction *IncV = nullptr; 1083 TruncTy = nullptr; 1084 InvertStep = false; 1085 1086 // Only try partially matching scevs that need truncation and/or 1087 // step-inversion if we know this loop is outside the current loop. 1088 bool TryNonMatchingSCEV = IVIncInsertLoop && 1089 SE.DT->properlyDominates(LatchBlock, IVIncInsertLoop->getHeader()); 1090 1091 for (BasicBlock::iterator I = L->getHeader()->begin(); 1092 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1093 if (!SE.isSCEVable(PN->getType())) 1094 continue; 1095 1096 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN)); 1097 if (!PhiSCEV) 1098 continue; 1099 1100 bool IsMatchingSCEV = PhiSCEV == Normalized; 1101 // We only handle truncation and inversion of phi recurrences for the 1102 // expanded expression if the expanded expression's loop dominates the 1103 // loop we insert to. Check now, so we can bail out early. 1104 if (!IsMatchingSCEV && !TryNonMatchingSCEV) 1105 continue; 1106 1107 Instruction *TempIncV = 1108 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 1109 1110 // Check whether we can reuse this PHI node. 1111 if (LSRMode) { 1112 if (!isExpandedAddRecExprPHI(PN, TempIncV, L)) 1113 continue; 1114 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos)) 1115 continue; 1116 } else { 1117 if (!isNormalAddRecExprPHI(PN, TempIncV, L)) 1118 continue; 1119 } 1120 1121 // Stop if we have found an exact match SCEV. 1122 if (IsMatchingSCEV) { 1123 IncV = TempIncV; 1124 TruncTy = nullptr; 1125 InvertStep = false; 1126 AddRecPhiMatch = PN; 1127 break; 1128 } 1129 1130 // Try whether the phi can be translated into the requested form 1131 // (truncated and/or offset by a constant). 1132 if ((!TruncTy || InvertStep) && 1133 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) { 1134 // Record the phi node. But don't stop we might find an exact match 1135 // later. 1136 AddRecPhiMatch = PN; 1137 IncV = TempIncV; 1138 TruncTy = SE.getEffectiveSCEVType(Normalized->getType()); 1139 } 1140 } 1141 1142 if (AddRecPhiMatch) { 1143 // Potentially, move the increment. We have made sure in 1144 // isExpandedAddRecExprPHI or hoistIVInc that this is possible. 1145 if (L == IVIncInsertLoop) 1146 hoistBeforePos(SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch); 1147 1148 // Ok, the add recurrence looks usable. 1149 // Remember this PHI, even in post-inc mode. 1150 InsertedValues.insert(AddRecPhiMatch); 1151 // Remember the increment. 1152 rememberInstruction(IncV); 1153 return AddRecPhiMatch; 1154 } 1155 } 1156 1157 // Save the original insertion point so we can restore it when we're done. 1158 BuilderType::InsertPointGuard Guard(Builder); 1159 1160 // Another AddRec may need to be recursively expanded below. For example, if 1161 // this AddRec is quadratic, the StepV may itself be an AddRec in this 1162 // loop. Remove this loop from the PostIncLoops set before expanding such 1163 // AddRecs. Otherwise, we cannot find a valid position for the step 1164 // (i.e. StepV can never dominate its loop header). Ideally, we could do 1165 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 1166 // so it's not worth implementing SmallPtrSet::swap. 1167 PostIncLoopSet SavedPostIncLoops = PostIncLoops; 1168 PostIncLoops.clear(); 1169 1170 // Expand code for the start value. 1171 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 1172 L->getHeader()->begin()); 1173 1174 // StartV must be hoisted into L's preheader to dominate the new phi. 1175 assert(!isa<Instruction>(StartV) || 1176 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(), 1177 L->getHeader())); 1178 1179 // Expand code for the step value. Do this before creating the PHI so that PHI 1180 // reuse code doesn't see an incomplete PHI. 1181 const SCEV *Step = Normalized->getStepRecurrence(SE); 1182 // If the stride is negative, insert a sub instead of an add for the increment 1183 // (unless it's a constant, because subtracts of constants are canonicalized 1184 // to adds). 1185 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1186 if (useSubtract) 1187 Step = SE.getNegativeSCEV(Step); 1188 // Expand the step somewhere that dominates the loop header. 1189 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1190 1191 // Create the PHI. 1192 BasicBlock *Header = L->getHeader(); 1193 Builder.SetInsertPoint(Header, Header->begin()); 1194 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1195 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 1196 Twine(IVName) + ".iv"); 1197 rememberInstruction(PN); 1198 1199 // Create the step instructions and populate the PHI. 1200 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1201 BasicBlock *Pred = *HPI; 1202 1203 // Add a start value. 1204 if (!L->contains(Pred)) { 1205 PN->addIncoming(StartV, Pred); 1206 continue; 1207 } 1208 1209 // Create a step value and add it to the PHI. 1210 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 1211 // instructions at IVIncInsertPos. 1212 Instruction *InsertPos = L == IVIncInsertLoop ? 1213 IVIncInsertPos : Pred->getTerminator(); 1214 Builder.SetInsertPoint(InsertPos); 1215 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1216 if (isa<OverflowingBinaryOperator>(IncV)) { 1217 if (Normalized->getNoWrapFlags(SCEV::FlagNUW)) 1218 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap(); 1219 if (Normalized->getNoWrapFlags(SCEV::FlagNSW)) 1220 cast<BinaryOperator>(IncV)->setHasNoSignedWrap(); 1221 } 1222 PN->addIncoming(IncV, Pred); 1223 } 1224 1225 // After expanding subexpressions, restore the PostIncLoops set so the caller 1226 // can ensure that IVIncrement dominates the current uses. 1227 PostIncLoops = SavedPostIncLoops; 1228 1229 // Remember this PHI, even in post-inc mode. 1230 InsertedValues.insert(PN); 1231 1232 return PN; 1233} 1234 1235Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 1236 Type *STy = S->getType(); 1237 Type *IntTy = SE.getEffectiveSCEVType(STy); 1238 const Loop *L = S->getLoop(); 1239 1240 // Determine a normalized form of this expression, which is the expression 1241 // before any post-inc adjustment is made. 1242 const SCEVAddRecExpr *Normalized = S; 1243 if (PostIncLoops.count(L)) { 1244 PostIncLoopSet Loops; 1245 Loops.insert(L); 1246 Normalized = 1247 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, nullptr, 1248 nullptr, Loops, SE, *SE.DT)); 1249 } 1250 1251 // Strip off any non-loop-dominating component from the addrec start. 1252 const SCEV *Start = Normalized->getStart(); 1253 const SCEV *PostLoopOffset = nullptr; 1254 if (!SE.properlyDominates(Start, L->getHeader())) { 1255 PostLoopOffset = Start; 1256 Start = SE.getConstant(Normalized->getType(), 0); 1257 Normalized = cast<SCEVAddRecExpr>( 1258 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 1259 Normalized->getLoop(), 1260 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1261 } 1262 1263 // Strip off any non-loop-dominating component from the addrec step. 1264 const SCEV *Step = Normalized->getStepRecurrence(SE); 1265 const SCEV *PostLoopScale = nullptr; 1266 if (!SE.dominates(Step, L->getHeader())) { 1267 PostLoopScale = Step; 1268 Step = SE.getConstant(Normalized->getType(), 1); 1269 Normalized = 1270 cast<SCEVAddRecExpr>(SE.getAddRecExpr( 1271 Start, Step, Normalized->getLoop(), 1272 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1273 } 1274 1275 // Expand the core addrec. If we need post-loop scaling, force it to 1276 // expand to an integer type to avoid the need for additional casting. 1277 Type *ExpandTy = PostLoopScale ? IntTy : STy; 1278 // In some cases, we decide to reuse an existing phi node but need to truncate 1279 // it and/or invert the step. 1280 Type *TruncTy = nullptr; 1281 bool InvertStep = false; 1282 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy, 1283 TruncTy, InvertStep); 1284 1285 // Accommodate post-inc mode, if necessary. 1286 Value *Result; 1287 if (!PostIncLoops.count(L)) 1288 Result = PN; 1289 else { 1290 // In PostInc mode, use the post-incremented value. 1291 BasicBlock *LatchBlock = L->getLoopLatch(); 1292 assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1293 Result = PN->getIncomingValueForBlock(LatchBlock); 1294 1295 // For an expansion to use the postinc form, the client must call 1296 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 1297 // or dominated by IVIncInsertPos. 1298 if (isa<Instruction>(Result) 1299 && !SE.DT->dominates(cast<Instruction>(Result), 1300 Builder.GetInsertPoint())) { 1301 // The induction variable's postinc expansion does not dominate this use. 1302 // IVUsers tries to prevent this case, so it is rare. However, it can 1303 // happen when an IVUser outside the loop is not dominated by the latch 1304 // block. Adjusting IVIncInsertPos before expansion begins cannot handle 1305 // all cases. Consider a phi outide whose operand is replaced during 1306 // expansion with the value of the postinc user. Without fundamentally 1307 // changing the way postinc users are tracked, the only remedy is 1308 // inserting an extra IV increment. StepV might fold into PostLoopOffset, 1309 // but hopefully expandCodeFor handles that. 1310 bool useSubtract = 1311 !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1312 if (useSubtract) 1313 Step = SE.getNegativeSCEV(Step); 1314 Value *StepV; 1315 { 1316 // Expand the step somewhere that dominates the loop header. 1317 BuilderType::InsertPointGuard Guard(Builder); 1318 StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1319 } 1320 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1321 } 1322 } 1323 1324 // We have decided to reuse an induction variable of a dominating loop. Apply 1325 // truncation and/or invertion of the step. 1326 if (TruncTy) { 1327 Type *ResTy = Result->getType(); 1328 // Normalize the result type. 1329 if (ResTy != SE.getEffectiveSCEVType(ResTy)) 1330 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy)); 1331 // Truncate the result. 1332 if (TruncTy != Result->getType()) { 1333 Result = Builder.CreateTrunc(Result, TruncTy); 1334 rememberInstruction(Result); 1335 } 1336 // Invert the result. 1337 if (InvertStep) { 1338 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy), 1339 Result); 1340 rememberInstruction(Result); 1341 } 1342 } 1343 1344 // Re-apply any non-loop-dominating scale. 1345 if (PostLoopScale) { 1346 assert(S->isAffine() && "Can't linearly scale non-affine recurrences."); 1347 Result = InsertNoopCastOfTo(Result, IntTy); 1348 Result = Builder.CreateMul(Result, 1349 expandCodeFor(PostLoopScale, IntTy)); 1350 rememberInstruction(Result); 1351 } 1352 1353 // Re-apply any non-loop-dominating offset. 1354 if (PostLoopOffset) { 1355 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1356 const SCEV *const OffsetArray[1] = { PostLoopOffset }; 1357 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 1358 } else { 1359 Result = InsertNoopCastOfTo(Result, IntTy); 1360 Result = Builder.CreateAdd(Result, 1361 expandCodeFor(PostLoopOffset, IntTy)); 1362 rememberInstruction(Result); 1363 } 1364 } 1365 1366 return Result; 1367} 1368 1369Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1370 if (!CanonicalMode) return expandAddRecExprLiterally(S); 1371 1372 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1373 const Loop *L = S->getLoop(); 1374 1375 // First check for an existing canonical IV in a suitable type. 1376 PHINode *CanonicalIV = nullptr; 1377 if (PHINode *PN = L->getCanonicalInductionVariable()) 1378 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 1379 CanonicalIV = PN; 1380 1381 // Rewrite an AddRec in terms of the canonical induction variable, if 1382 // its type is more narrow. 1383 if (CanonicalIV && 1384 SE.getTypeSizeInBits(CanonicalIV->getType()) > 1385 SE.getTypeSizeInBits(Ty)) { 1386 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1387 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1388 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 1389 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 1390 S->getNoWrapFlags(SCEV::FlagNW))); 1391 BasicBlock::iterator NewInsertPt = 1392 std::next(BasicBlock::iterator(cast<Instruction>(V))); 1393 BuilderType::InsertPointGuard Guard(Builder); 1394 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) || 1395 isa<LandingPadInst>(NewInsertPt)) 1396 ++NewInsertPt; 1397 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr, 1398 NewInsertPt); 1399 return V; 1400 } 1401 1402 // {X,+,F} --> X + {0,+,F} 1403 if (!S->getStart()->isZero()) { 1404 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1405 NewOps[0] = SE.getConstant(Ty, 0); 1406 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, 1407 S->getNoWrapFlags(SCEV::FlagNW)); 1408 1409 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1410 // comments on expandAddToGEP for details. 1411 const SCEV *Base = S->getStart(); 1412 const SCEV *RestArray[1] = { Rest }; 1413 // Dig into the expression to find the pointer base for a GEP. 1414 ExposePointerBase(Base, RestArray[0], SE); 1415 // If we found a pointer, expand the AddRec with a GEP. 1416 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1417 // Make sure the Base isn't something exotic, such as a multiplied 1418 // or divided pointer value. In those cases, the result type isn't 1419 // actually a pointer type. 1420 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1421 Value *StartV = expand(Base); 1422 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1423 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 1424 } 1425 } 1426 1427 // Just do a normal add. Pre-expand the operands to suppress folding. 1428 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), 1429 SE.getUnknown(expand(Rest)))); 1430 } 1431 1432 // If we don't yet have a canonical IV, create one. 1433 if (!CanonicalIV) { 1434 // Create and insert the PHI node for the induction variable in the 1435 // specified loop. 1436 BasicBlock *Header = L->getHeader(); 1437 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1438 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 1439 Header->begin()); 1440 rememberInstruction(CanonicalIV); 1441 1442 SmallSet<BasicBlock *, 4> PredSeen; 1443 Constant *One = ConstantInt::get(Ty, 1); 1444 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1445 BasicBlock *HP = *HPI; 1446 if (!PredSeen.insert(HP).second) { 1447 // There must be an incoming value for each predecessor, even the 1448 // duplicates! 1449 CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP); 1450 continue; 1451 } 1452 1453 if (L->contains(HP)) { 1454 // Insert a unit add instruction right before the terminator 1455 // corresponding to the back-edge. 1456 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 1457 "indvar.next", 1458 HP->getTerminator()); 1459 Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 1460 rememberInstruction(Add); 1461 CanonicalIV->addIncoming(Add, HP); 1462 } else { 1463 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 1464 } 1465 } 1466 } 1467 1468 // {0,+,1} --> Insert a canonical induction variable into the loop! 1469 if (S->isAffine() && S->getOperand(1)->isOne()) { 1470 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 1471 "IVs with types different from the canonical IV should " 1472 "already have been handled!"); 1473 return CanonicalIV; 1474 } 1475 1476 // {0,+,F} --> {0,+,1} * F 1477 1478 // If this is a simple linear addrec, emit it now as a special case. 1479 if (S->isAffine()) // {0,+,F} --> i*F 1480 return 1481 expand(SE.getTruncateOrNoop( 1482 SE.getMulExpr(SE.getUnknown(CanonicalIV), 1483 SE.getNoopOrAnyExtend(S->getOperand(1), 1484 CanonicalIV->getType())), 1485 Ty)); 1486 1487 // If this is a chain of recurrences, turn it into a closed form, using the 1488 // folders, then expandCodeFor the closed form. This allows the folders to 1489 // simplify the expression without having to build a bunch of special code 1490 // into this folder. 1491 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 1492 1493 // Promote S up to the canonical IV type, if the cast is foldable. 1494 const SCEV *NewS = S; 1495 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 1496 if (isa<SCEVAddRecExpr>(Ext)) 1497 NewS = Ext; 1498 1499 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1500 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 1501 1502 // Truncate the result down to the original type, if needed. 1503 const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1504 return expand(T); 1505} 1506 1507Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1508 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1509 Value *V = expandCodeFor(S->getOperand(), 1510 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1511 Value *I = Builder.CreateTrunc(V, Ty); 1512 rememberInstruction(I); 1513 return I; 1514} 1515 1516Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1517 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1518 Value *V = expandCodeFor(S->getOperand(), 1519 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1520 Value *I = Builder.CreateZExt(V, Ty); 1521 rememberInstruction(I); 1522 return I; 1523} 1524 1525Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1526 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1527 Value *V = expandCodeFor(S->getOperand(), 1528 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1529 Value *I = Builder.CreateSExt(V, Ty); 1530 rememberInstruction(I); 1531 return I; 1532} 1533 1534Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 1535 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1536 Type *Ty = LHS->getType(); 1537 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1538 // In the case of mixed integer and pointer types, do the 1539 // rest of the comparisons as integer. 1540 if (S->getOperand(i)->getType() != Ty) { 1541 Ty = SE.getEffectiveSCEVType(Ty); 1542 LHS = InsertNoopCastOfTo(LHS, Ty); 1543 } 1544 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1545 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 1546 rememberInstruction(ICmp); 1547 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1548 rememberInstruction(Sel); 1549 LHS = Sel; 1550 } 1551 // In the case of mixed integer and pointer types, cast the 1552 // final result back to the pointer type. 1553 if (LHS->getType() != S->getType()) 1554 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1555 return LHS; 1556} 1557 1558Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 1559 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1560 Type *Ty = LHS->getType(); 1561 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1562 // In the case of mixed integer and pointer types, do the 1563 // rest of the comparisons as integer. 1564 if (S->getOperand(i)->getType() != Ty) { 1565 Ty = SE.getEffectiveSCEVType(Ty); 1566 LHS = InsertNoopCastOfTo(LHS, Ty); 1567 } 1568 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1569 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 1570 rememberInstruction(ICmp); 1571 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1572 rememberInstruction(Sel); 1573 LHS = Sel; 1574 } 1575 // In the case of mixed integer and pointer types, cast the 1576 // final result back to the pointer type. 1577 if (LHS->getType() != S->getType()) 1578 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1579 return LHS; 1580} 1581 1582Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 1583 Instruction *IP) { 1584 Builder.SetInsertPoint(IP->getParent(), IP); 1585 return expandCodeFor(SH, Ty); 1586} 1587 1588Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 1589 // Expand the code for this SCEV. 1590 Value *V = expand(SH); 1591 if (Ty) { 1592 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 1593 "non-trivial casts should be done with the SCEVs directly!"); 1594 V = InsertNoopCastOfTo(V, Ty); 1595 } 1596 return V; 1597} 1598 1599Value *SCEVExpander::expand(const SCEV *S) { 1600 // Compute an insertion point for this SCEV object. Hoist the instructions 1601 // as far out in the loop nest as possible. 1602 Instruction *InsertPt = Builder.GetInsertPoint(); 1603 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ; 1604 L = L->getParentLoop()) 1605 if (SE.isLoopInvariant(S, L)) { 1606 if (!L) break; 1607 if (BasicBlock *Preheader = L->getLoopPreheader()) 1608 InsertPt = Preheader->getTerminator(); 1609 else { 1610 // LSR sets the insertion point for AddRec start/step values to the 1611 // block start to simplify value reuse, even though it's an invalid 1612 // position. SCEVExpander must correct for this in all cases. 1613 InsertPt = L->getHeader()->getFirstInsertionPt(); 1614 } 1615 } else { 1616 // If the SCEV is computable at this level, insert it into the header 1617 // after the PHIs (and after any other instructions that we've inserted 1618 // there) so that it is guaranteed to dominate any user inside the loop. 1619 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 1620 InsertPt = L->getHeader()->getFirstInsertionPt(); 1621 while (InsertPt != Builder.GetInsertPoint() 1622 && (isInsertedInstruction(InsertPt) 1623 || isa<DbgInfoIntrinsic>(InsertPt))) { 1624 InsertPt = std::next(BasicBlock::iterator(InsertPt)); 1625 } 1626 break; 1627 } 1628 1629 // Check to see if we already expanded this here. 1630 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator 1631 I = InsertedExpressions.find(std::make_pair(S, InsertPt)); 1632 if (I != InsertedExpressions.end()) 1633 return I->second; 1634 1635 BuilderType::InsertPointGuard Guard(Builder); 1636 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt); 1637 1638 // Expand the expression into instructions. 1639 Value *V = visit(S); 1640 1641 // Remember the expanded value for this SCEV at this location. 1642 // 1643 // This is independent of PostIncLoops. The mapped value simply materializes 1644 // the expression at this insertion point. If the mapped value happened to be 1645 // a postinc expansion, it could be reused by a non-postinc user, but only if 1646 // its insertion point was already at the head of the loop. 1647 InsertedExpressions[std::make_pair(S, InsertPt)] = V; 1648 return V; 1649} 1650 1651void SCEVExpander::rememberInstruction(Value *I) { 1652 if (!PostIncLoops.empty()) 1653 InsertedPostIncValues.insert(I); 1654 else 1655 InsertedValues.insert(I); 1656} 1657 1658/// getOrInsertCanonicalInductionVariable - This method returns the 1659/// canonical induction variable of the specified type for the specified 1660/// loop (inserting one if there is none). A canonical induction variable 1661/// starts at zero and steps by one on each iteration. 1662PHINode * 1663SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1664 Type *Ty) { 1665 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1666 1667 // Build a SCEV for {0,+,1}<L>. 1668 // Conservatively use FlagAnyWrap for now. 1669 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 1670 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 1671 1672 // Emit code for it. 1673 BuilderType::InsertPointGuard Guard(Builder); 1674 PHINode *V = cast<PHINode>(expandCodeFor(H, nullptr, 1675 L->getHeader()->begin())); 1676 1677 return V; 1678} 1679 1680/// replaceCongruentIVs - Check for congruent phis in this loop header and 1681/// replace them with their most canonical representative. Return the number of 1682/// phis eliminated. 1683/// 1684/// This does not depend on any SCEVExpander state but should be used in 1685/// the same context that SCEVExpander is used. 1686unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 1687 SmallVectorImpl<WeakVH> &DeadInsts, 1688 const TargetTransformInfo *TTI) { 1689 // Find integer phis in order of increasing width. 1690 SmallVector<PHINode*, 8> Phis; 1691 for (BasicBlock::iterator I = L->getHeader()->begin(); 1692 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 1693 Phis.push_back(Phi); 1694 } 1695 if (TTI) 1696 std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) { 1697 // Put pointers at the back and make sure pointer < pointer = false. 1698 if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy()) 1699 return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy(); 1700 return RHS->getType()->getPrimitiveSizeInBits() < 1701 LHS->getType()->getPrimitiveSizeInBits(); 1702 }); 1703 1704 unsigned NumElim = 0; 1705 DenseMap<const SCEV *, PHINode *> ExprToIVMap; 1706 // Process phis from wide to narrow. Mapping wide phis to the their truncation 1707 // so narrow phis can reuse them. 1708 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(), 1709 PEnd = Phis.end(); PIter != PEnd; ++PIter) { 1710 PHINode *Phi = *PIter; 1711 1712 // Fold constant phis. They may be congruent to other constant phis and 1713 // would confuse the logic below that expects proper IVs. 1714 if (Value *V = SimplifyInstruction(Phi, SE.DL, SE.TLI, SE.DT, SE.AC)) { 1715 Phi->replaceAllUsesWith(V); 1716 DeadInsts.push_back(Phi); 1717 ++NumElim; 1718 DEBUG_WITH_TYPE(DebugType, dbgs() 1719 << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 1720 continue; 1721 } 1722 1723 if (!SE.isSCEVable(Phi->getType())) 1724 continue; 1725 1726 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 1727 if (!OrigPhiRef) { 1728 OrigPhiRef = Phi; 1729 if (Phi->getType()->isIntegerTy() && TTI 1730 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 1731 // This phi can be freely truncated to the narrowest phi type. Map the 1732 // truncated expression to it so it will be reused for narrow types. 1733 const SCEV *TruncExpr = 1734 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 1735 ExprToIVMap[TruncExpr] = Phi; 1736 } 1737 continue; 1738 } 1739 1740 // Replacing a pointer phi with an integer phi or vice-versa doesn't make 1741 // sense. 1742 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 1743 continue; 1744 1745 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1746 Instruction *OrigInc = 1747 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 1748 Instruction *IsomorphicInc = 1749 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 1750 1751 // If this phi has the same width but is more canonical, replace the 1752 // original with it. As part of the "more canonical" determination, 1753 // respect a prior decision to use an IV chain. 1754 if (OrigPhiRef->getType() == Phi->getType() 1755 && !(ChainedPhis.count(Phi) 1756 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) 1757 && (ChainedPhis.count(Phi) 1758 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 1759 std::swap(OrigPhiRef, Phi); 1760 std::swap(OrigInc, IsomorphicInc); 1761 } 1762 // Replacing the congruent phi is sufficient because acyclic redundancy 1763 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves 1764 // that a phi is congruent, it's often the head of an IV user cycle that 1765 // is isomorphic with the original phi. It's worth eagerly cleaning up the 1766 // common case of a single IV increment so that DeleteDeadPHIs can remove 1767 // cycles that had postinc uses. 1768 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc), 1769 IsomorphicInc->getType()); 1770 if (OrigInc != IsomorphicInc 1771 && TruncExpr == SE.getSCEV(IsomorphicInc) 1772 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc)) 1773 || hoistIVInc(OrigInc, IsomorphicInc))) { 1774 DEBUG_WITH_TYPE(DebugType, dbgs() 1775 << "INDVARS: Eliminated congruent iv.inc: " 1776 << *IsomorphicInc << '\n'); 1777 Value *NewInc = OrigInc; 1778 if (OrigInc->getType() != IsomorphicInc->getType()) {
| 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/STLExtras.h" 18#include "llvm/ADT/SmallSet.h" 19#include "llvm/Analysis/InstructionSimplify.h" 20#include "llvm/Analysis/LoopInfo.h" 21#include "llvm/Analysis/TargetTransformInfo.h" 22#include "llvm/IR/DataLayout.h" 23#include "llvm/IR/Dominators.h" 24#include "llvm/IR/IntrinsicInst.h" 25#include "llvm/IR/LLVMContext.h" 26#include "llvm/Support/Debug.h" 27 28using namespace llvm; 29 30/// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, 31/// reusing an existing cast if a suitable one exists, moving an existing 32/// cast if a suitable one exists but isn't in the right place, or 33/// creating a new one. 34Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, 35 Instruction::CastOps Op, 36 BasicBlock::iterator IP) { 37 // This function must be called with the builder having a valid insertion 38 // point. It doesn't need to be the actual IP where the uses of the returned 39 // cast will be added, but it must dominate such IP. 40 // We use this precondition to produce a cast that will dominate all its 41 // uses. In particular, this is crucial for the case where the builder's 42 // insertion point *is* the point where we were asked to put the cast. 43 // Since we don't know the builder's insertion point is actually 44 // where the uses will be added (only that it dominates it), we are 45 // not allowed to move it. 46 BasicBlock::iterator BIP = Builder.GetInsertPoint(); 47 48 Instruction *Ret = nullptr; 49 50 // Check to see if there is already a cast! 51 for (User *U : V->users()) 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 // Create a new cast. 73 if (!Ret) 74 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP); 75 76 // We assert at the end of the function since IP might point to an 77 // instruction with different dominance properties than a cast 78 // (an invoke for example) and not dominate BIP (but the cast does). 79 assert(SE.DT->dominates(Ret, BIP)); 80 81 rememberInstruction(Ret); 82 return Ret; 83} 84 85/// InsertNoopCastOfTo - Insert a cast of V to the specified type, 86/// which must be possible with a noop cast, doing what we can to share 87/// the casts. 88Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 89 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 90 assert((Op == Instruction::BitCast || 91 Op == Instruction::PtrToInt || 92 Op == Instruction::IntToPtr) && 93 "InsertNoopCastOfTo cannot perform non-noop casts!"); 94 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 95 "InsertNoopCastOfTo cannot change sizes!"); 96 97 // Short-circuit unnecessary bitcasts. 98 if (Op == Instruction::BitCast) { 99 if (V->getType() == Ty) 100 return V; 101 if (CastInst *CI = dyn_cast<CastInst>(V)) { 102 if (CI->getOperand(0)->getType() == Ty) 103 return CI->getOperand(0); 104 } 105 } 106 // Short-circuit unnecessary inttoptr<->ptrtoint casts. 107 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 108 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 109 if (CastInst *CI = dyn_cast<CastInst>(V)) 110 if ((CI->getOpcode() == Instruction::PtrToInt || 111 CI->getOpcode() == Instruction::IntToPtr) && 112 SE.getTypeSizeInBits(CI->getType()) == 113 SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 114 return CI->getOperand(0); 115 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 116 if ((CE->getOpcode() == Instruction::PtrToInt || 117 CE->getOpcode() == Instruction::IntToPtr) && 118 SE.getTypeSizeInBits(CE->getType()) == 119 SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 120 return CE->getOperand(0); 121 } 122 123 // Fold a cast of a constant. 124 if (Constant *C = dyn_cast<Constant>(V)) 125 return ConstantExpr::getCast(Op, C, Ty); 126 127 // Cast the argument at the beginning of the entry block, after 128 // any bitcasts of other arguments. 129 if (Argument *A = dyn_cast<Argument>(V)) { 130 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 131 while ((isa<BitCastInst>(IP) && 132 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 133 cast<BitCastInst>(IP)->getOperand(0) != A) || 134 isa<DbgInfoIntrinsic>(IP) || 135 isa<LandingPadInst>(IP)) 136 ++IP; 137 return ReuseOrCreateCast(A, Ty, Op, IP); 138 } 139 140 // Cast the instruction immediately after the instruction. 141 Instruction *I = cast<Instruction>(V); 142 BasicBlock::iterator IP = I; ++IP; 143 if (InvokeInst *II = dyn_cast<InvokeInst>(I)) 144 IP = II->getNormalDest()->begin(); 145 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP)) 146 ++IP; 147 return ReuseOrCreateCast(I, Ty, Op, IP); 148} 149 150/// InsertBinop - Insert the specified binary operator, doing a small amount 151/// of work to avoid inserting an obviously redundant operation. 152Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 153 Value *LHS, Value *RHS) { 154 // Fold a binop with constant operands. 155 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 156 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 157 return ConstantExpr::get(Opcode, CLHS, CRHS); 158 159 // Do a quick scan to see if we have this binop nearby. If so, reuse it. 160 unsigned ScanLimit = 6; 161 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 162 // Scanning starts from the last instruction before the insertion point. 163 BasicBlock::iterator IP = Builder.GetInsertPoint(); 164 if (IP != BlockBegin) { 165 --IP; 166 for (; ScanLimit; --IP, --ScanLimit) { 167 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 168 // generated code. 169 if (isa<DbgInfoIntrinsic>(IP)) 170 ScanLimit++; 171 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 172 IP->getOperand(1) == RHS) 173 return IP; 174 if (IP == BlockBegin) break; 175 } 176 } 177 178 // Save the original insertion point so we can restore it when we're done. 179 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); 180 BuilderType::InsertPointGuard Guard(Builder); 181 182 // Move the insertion point out of as many loops as we can. 183 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 184 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 185 BasicBlock *Preheader = L->getLoopPreheader(); 186 if (!Preheader) break; 187 188 // Ok, move up a level. 189 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 190 } 191 192 // If we haven't found this binop, insert it. 193 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 194 BO->setDebugLoc(Loc); 195 rememberInstruction(BO); 196 197 return BO; 198} 199 200/// FactorOutConstant - Test if S is divisible by Factor, using signed 201/// division. If so, update S with Factor divided out and return true. 202/// S need not be evenly divisible if a reasonable remainder can be 203/// computed. 204/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 205/// unnecessary; in its place, just signed-divide Ops[i] by the scale and 206/// check to see if the divide was folded. 207static bool FactorOutConstant(const SCEV *&S, 208 const SCEV *&Remainder, 209 const SCEV *Factor, 210 ScalarEvolution &SE, 211 const DataLayout *DL) { 212 // Everything is divisible by one. 213 if (Factor->isOne()) 214 return true; 215 216 // x/x == 1. 217 if (S == Factor) { 218 S = SE.getConstant(S->getType(), 1); 219 return true; 220 } 221 222 // For a Constant, check for a multiple of the given factor. 223 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 224 // 0/x == 0. 225 if (C->isZero()) 226 return true; 227 // Check for divisibility. 228 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 229 ConstantInt *CI = 230 ConstantInt::get(SE.getContext(), 231 C->getValue()->getValue().sdiv( 232 FC->getValue()->getValue())); 233 // If the quotient is zero and the remainder is non-zero, reject 234 // the value at this scale. It will be considered for subsequent 235 // smaller scales. 236 if (!CI->isZero()) { 237 const SCEV *Div = SE.getConstant(CI); 238 S = Div; 239 Remainder = 240 SE.getAddExpr(Remainder, 241 SE.getConstant(C->getValue()->getValue().srem( 242 FC->getValue()->getValue()))); 243 return true; 244 } 245 } 246 } 247 248 // In a Mul, check if there is a constant operand which is a multiple 249 // of the given factor. 250 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 251 if (DL) { 252 // With DataLayout, the size is known. Check if there is a constant 253 // operand which is a multiple of the given factor. If so, we can 254 // factor it. 255 const SCEVConstant *FC = cast<SCEVConstant>(Factor); 256 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 257 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) { 258 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 259 NewMulOps[0] = 260 SE.getConstant(C->getValue()->getValue().sdiv( 261 FC->getValue()->getValue())); 262 S = SE.getMulExpr(NewMulOps); 263 return true; 264 } 265 } else { 266 // Without DataLayout, check if Factor can be factored out of any of the 267 // Mul's operands. If so, we can just remove it. 268 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) { 269 const SCEV *SOp = M->getOperand(i); 270 const SCEV *Remainder = SE.getConstant(SOp->getType(), 0); 271 if (FactorOutConstant(SOp, Remainder, Factor, SE, DL) && 272 Remainder->isZero()) { 273 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 274 NewMulOps[i] = SOp; 275 S = SE.getMulExpr(NewMulOps); 276 return true; 277 } 278 } 279 } 280 } 281 282 // In an AddRec, check if both start and step are divisible. 283 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 284 const SCEV *Step = A->getStepRecurrence(SE); 285 const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 286 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL)) 287 return false; 288 if (!StepRem->isZero()) 289 return false; 290 const SCEV *Start = A->getStart(); 291 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL)) 292 return false; 293 S = SE.getAddRecExpr(Start, Step, A->getLoop(), 294 A->getNoWrapFlags(SCEV::FlagNW)); 295 return true; 296 } 297 298 return false; 299} 300 301/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 302/// is the number of SCEVAddRecExprs present, which are kept at the end of 303/// the list. 304/// 305static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 306 Type *Ty, 307 ScalarEvolution &SE) { 308 unsigned NumAddRecs = 0; 309 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 310 ++NumAddRecs; 311 // Group Ops into non-addrecs and addrecs. 312 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 313 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 314 // Let ScalarEvolution sort and simplify the non-addrecs list. 315 const SCEV *Sum = NoAddRecs.empty() ? 316 SE.getConstant(Ty, 0) : 317 SE.getAddExpr(NoAddRecs); 318 // If it returned an add, use the operands. Otherwise it simplified 319 // the sum into a single value, so just use that. 320 Ops.clear(); 321 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 322 Ops.append(Add->op_begin(), Add->op_end()); 323 else if (!Sum->isZero()) 324 Ops.push_back(Sum); 325 // Then append the addrecs. 326 Ops.append(AddRecs.begin(), AddRecs.end()); 327} 328 329/// SplitAddRecs - Flatten a list of add operands, moving addrec start values 330/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 331/// This helps expose more opportunities for folding parts of the expressions 332/// into GEP indices. 333/// 334static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 335 Type *Ty, 336 ScalarEvolution &SE) { 337 // Find the addrecs. 338 SmallVector<const SCEV *, 8> AddRecs; 339 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 340 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 341 const SCEV *Start = A->getStart(); 342 if (Start->isZero()) break; 343 const SCEV *Zero = SE.getConstant(Ty, 0); 344 AddRecs.push_back(SE.getAddRecExpr(Zero, 345 A->getStepRecurrence(SE), 346 A->getLoop(), 347 A->getNoWrapFlags(SCEV::FlagNW))); 348 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 349 Ops[i] = Zero; 350 Ops.append(Add->op_begin(), Add->op_end()); 351 e += Add->getNumOperands(); 352 } else { 353 Ops[i] = Start; 354 } 355 } 356 if (!AddRecs.empty()) { 357 // Add the addrecs onto the end of the list. 358 Ops.append(AddRecs.begin(), AddRecs.end()); 359 // Resort the operand list, moving any constants to the front. 360 SimplifyAddOperands(Ops, Ty, SE); 361 } 362} 363 364/// expandAddToGEP - Expand an addition expression with a pointer type into 365/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 366/// BasicAliasAnalysis and other passes analyze the result. See the rules 367/// for getelementptr vs. inttoptr in 368/// http://llvm.org/docs/LangRef.html#pointeraliasing 369/// for details. 370/// 371/// Design note: The correctness of using getelementptr here depends on 372/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 373/// they may introduce pointer arithmetic which may not be safely converted 374/// into getelementptr. 375/// 376/// Design note: It might seem desirable for this function to be more 377/// loop-aware. If some of the indices are loop-invariant while others 378/// aren't, it might seem desirable to emit multiple GEPs, keeping the 379/// loop-invariant portions of the overall computation outside the loop. 380/// However, there are a few reasons this is not done here. Hoisting simple 381/// arithmetic is a low-level optimization that often isn't very 382/// important until late in the optimization process. In fact, passes 383/// like InstructionCombining will combine GEPs, even if it means 384/// pushing loop-invariant computation down into loops, so even if the 385/// GEPs were split here, the work would quickly be undone. The 386/// LoopStrengthReduction pass, which is usually run quite late (and 387/// after the last InstructionCombining pass), takes care of hoisting 388/// loop-invariant portions of expressions, after considering what 389/// can be folded using target addressing modes. 390/// 391Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 392 const SCEV *const *op_end, 393 PointerType *PTy, 394 Type *Ty, 395 Value *V) { 396 Type *ElTy = PTy->getElementType(); 397 SmallVector<Value *, 4> GepIndices; 398 SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 399 bool AnyNonZeroIndices = false; 400 401 // Split AddRecs up into parts as either of the parts may be usable 402 // without the other. 403 SplitAddRecs(Ops, Ty, SE); 404 405 Type *IntPtrTy = SE.DL 406 ? SE.DL->getIntPtrType(PTy) 407 : Type::getInt64Ty(PTy->getContext()); 408 409 // Descend down the pointer's type and attempt to convert the other 410 // operands into GEP indices, at each level. The first index in a GEP 411 // indexes into the array implied by the pointer operand; the rest of 412 // the indices index into the element or field type selected by the 413 // preceding index. 414 for (;;) { 415 // If the scale size is not 0, attempt to factor out a scale for 416 // array indexing. 417 SmallVector<const SCEV *, 8> ScaledOps; 418 if (ElTy->isSized()) { 419 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy); 420 if (!ElSize->isZero()) { 421 SmallVector<const SCEV *, 8> NewOps; 422 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 423 const SCEV *Op = Ops[i]; 424 const SCEV *Remainder = SE.getConstant(Ty, 0); 425 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.DL)) { 426 // Op now has ElSize factored out. 427 ScaledOps.push_back(Op); 428 if (!Remainder->isZero()) 429 NewOps.push_back(Remainder); 430 AnyNonZeroIndices = true; 431 } else { 432 // The operand was not divisible, so add it to the list of operands 433 // we'll scan next iteration. 434 NewOps.push_back(Ops[i]); 435 } 436 } 437 // If we made any changes, update Ops. 438 if (!ScaledOps.empty()) { 439 Ops = NewOps; 440 SimplifyAddOperands(Ops, Ty, SE); 441 } 442 } 443 } 444 445 // Record the scaled array index for this level of the type. If 446 // we didn't find any operands that could be factored, tentatively 447 // assume that element zero was selected (since the zero offset 448 // would obviously be folded away). 449 Value *Scaled = ScaledOps.empty() ? 450 Constant::getNullValue(Ty) : 451 expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 452 GepIndices.push_back(Scaled); 453 454 // Collect struct field index operands. 455 while (StructType *STy = dyn_cast<StructType>(ElTy)) { 456 bool FoundFieldNo = false; 457 // An empty struct has no fields. 458 if (STy->getNumElements() == 0) break; 459 if (SE.DL) { 460 // With DataLayout, field offsets are known. See if a constant offset 461 // falls within any of the struct fields. 462 if (Ops.empty()) break; 463 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 464 if (SE.getTypeSizeInBits(C->getType()) <= 64) { 465 const StructLayout &SL = *SE.DL->getStructLayout(STy); 466 uint64_t FullOffset = C->getValue()->getZExtValue(); 467 if (FullOffset < SL.getSizeInBytes()) { 468 unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 469 GepIndices.push_back( 470 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 471 ElTy = STy->getTypeAtIndex(ElIdx); 472 Ops[0] = 473 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 474 AnyNonZeroIndices = true; 475 FoundFieldNo = true; 476 } 477 } 478 } else { 479 // Without DataLayout, just check for an offsetof expression of the 480 // appropriate struct type. 481 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 482 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) { 483 Type *CTy; 484 Constant *FieldNo; 485 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) { 486 GepIndices.push_back(FieldNo); 487 ElTy = 488 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue()); 489 Ops[i] = SE.getConstant(Ty, 0); 490 AnyNonZeroIndices = true; 491 FoundFieldNo = true; 492 break; 493 } 494 } 495 } 496 // If no struct field offsets were found, tentatively assume that 497 // field zero was selected (since the zero offset would obviously 498 // be folded away). 499 if (!FoundFieldNo) { 500 ElTy = STy->getTypeAtIndex(0u); 501 GepIndices.push_back( 502 Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 503 } 504 } 505 506 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 507 ElTy = ATy->getElementType(); 508 else 509 break; 510 } 511 512 // If none of the operands were convertible to proper GEP indices, cast 513 // the base to i8* and do an ugly getelementptr with that. It's still 514 // better than ptrtoint+arithmetic+inttoptr at least. 515 if (!AnyNonZeroIndices) { 516 // Cast the base to i8*. 517 V = InsertNoopCastOfTo(V, 518 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 519 520 assert(!isa<Instruction>(V) || 521 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint())); 522 523 // Expand the operands for a plain byte offset. 524 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 525 526 // Fold a GEP with constant operands. 527 if (Constant *CLHS = dyn_cast<Constant>(V)) 528 if (Constant *CRHS = dyn_cast<Constant>(Idx)) 529 return ConstantExpr::getGetElementPtr(CLHS, CRHS); 530 531 // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 532 unsigned ScanLimit = 6; 533 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 534 // Scanning starts from the last instruction before the insertion point. 535 BasicBlock::iterator IP = Builder.GetInsertPoint(); 536 if (IP != BlockBegin) { 537 --IP; 538 for (; ScanLimit; --IP, --ScanLimit) { 539 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 540 // generated code. 541 if (isa<DbgInfoIntrinsic>(IP)) 542 ScanLimit++; 543 if (IP->getOpcode() == Instruction::GetElementPtr && 544 IP->getOperand(0) == V && IP->getOperand(1) == Idx) 545 return IP; 546 if (IP == BlockBegin) break; 547 } 548 } 549 550 // Save the original insertion point so we can restore it when we're done. 551 BuilderType::InsertPointGuard Guard(Builder); 552 553 // Move the insertion point out of as many loops as we can. 554 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 555 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 556 BasicBlock *Preheader = L->getLoopPreheader(); 557 if (!Preheader) break; 558 559 // Ok, move up a level. 560 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 561 } 562 563 // Emit a GEP. 564 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep"); 565 rememberInstruction(GEP); 566 567 return GEP; 568 } 569 570 // Save the original insertion point so we can restore it when we're done. 571 BuilderType::InsertPoint SaveInsertPt = Builder.saveIP(); 572 573 // Move the insertion point out of as many loops as we can. 574 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 575 if (!L->isLoopInvariant(V)) break; 576 577 bool AnyIndexNotLoopInvariant = false; 578 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(), 579 E = GepIndices.end(); I != E; ++I) 580 if (!L->isLoopInvariant(*I)) { 581 AnyIndexNotLoopInvariant = true; 582 break; 583 } 584 if (AnyIndexNotLoopInvariant) 585 break; 586 587 BasicBlock *Preheader = L->getLoopPreheader(); 588 if (!Preheader) break; 589 590 // Ok, move up a level. 591 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 592 } 593 594 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 595 // because ScalarEvolution may have changed the address arithmetic to 596 // compute a value which is beyond the end of the allocated object. 597 Value *Casted = V; 598 if (V->getType() != PTy) 599 Casted = InsertNoopCastOfTo(Casted, PTy); 600 Value *GEP = Builder.CreateGEP(Casted, 601 GepIndices, 602 "scevgep"); 603 Ops.push_back(SE.getUnknown(GEP)); 604 rememberInstruction(GEP); 605 606 // Restore the original insert point. 607 Builder.restoreIP(SaveInsertPt); 608 609 return expand(SE.getAddExpr(Ops)); 610} 611 612/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 613/// SCEV expansion. If they are nested, this is the most nested. If they are 614/// neighboring, pick the later. 615static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 616 DominatorTree &DT) { 617 if (!A) return B; 618 if (!B) return A; 619 if (A->contains(B)) return B; 620 if (B->contains(A)) return A; 621 if (DT.dominates(A->getHeader(), B->getHeader())) return B; 622 if (DT.dominates(B->getHeader(), A->getHeader())) return A; 623 return A; // Arbitrarily break the tie. 624} 625 626/// getRelevantLoop - Get the most relevant loop associated with the given 627/// expression, according to PickMostRelevantLoop. 628const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 629 // Test whether we've already computed the most relevant loop for this SCEV. 630 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair = 631 RelevantLoops.insert(std::make_pair(S, nullptr)); 632 if (!Pair.second) 633 return Pair.first->second; 634 635 if (isa<SCEVConstant>(S)) 636 // A constant has no relevant loops. 637 return nullptr; 638 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 639 if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 640 return Pair.first->second = SE.LI->getLoopFor(I->getParent()); 641 // A non-instruction has no relevant loops. 642 return nullptr; 643 } 644 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 645 const Loop *L = nullptr; 646 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 647 L = AR->getLoop(); 648 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end(); 649 I != E; ++I) 650 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT); 651 return RelevantLoops[N] = L; 652 } 653 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 654 const Loop *Result = getRelevantLoop(C->getOperand()); 655 return RelevantLoops[C] = Result; 656 } 657 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 658 const Loop *Result = 659 PickMostRelevantLoop(getRelevantLoop(D->getLHS()), 660 getRelevantLoop(D->getRHS()), 661 *SE.DT); 662 return RelevantLoops[D] = Result; 663 } 664 llvm_unreachable("Unexpected SCEV type!"); 665} 666 667namespace { 668 669/// LoopCompare - Compare loops by PickMostRelevantLoop. 670class LoopCompare { 671 DominatorTree &DT; 672public: 673 explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 674 675 bool operator()(std::pair<const Loop *, const SCEV *> LHS, 676 std::pair<const Loop *, const SCEV *> RHS) const { 677 // Keep pointer operands sorted at the end. 678 if (LHS.second->getType()->isPointerTy() != 679 RHS.second->getType()->isPointerTy()) 680 return LHS.second->getType()->isPointerTy(); 681 682 // Compare loops with PickMostRelevantLoop. 683 if (LHS.first != RHS.first) 684 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 685 686 // If one operand is a non-constant negative and the other is not, 687 // put the non-constant negative on the right so that a sub can 688 // be used instead of a negate and add. 689 if (LHS.second->isNonConstantNegative()) { 690 if (!RHS.second->isNonConstantNegative()) 691 return false; 692 } else if (RHS.second->isNonConstantNegative()) 693 return true; 694 695 // Otherwise they are equivalent according to this comparison. 696 return false; 697 } 698}; 699 700} 701 702Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 703 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 704 705 // Collect all the add operands in a loop, along with their associated loops. 706 // Iterate in reverse so that constants are emitted last, all else equal, and 707 // so that pointer operands are inserted first, which the code below relies on 708 // to form more involved GEPs. 709 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 710 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 711 E(S->op_begin()); I != E; ++I) 712 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 713 714 // Sort by loop. Use a stable sort so that constants follow non-constants and 715 // pointer operands precede non-pointer operands. 716 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 717 718 // Emit instructions to add all the operands. Hoist as much as possible 719 // out of loops, and form meaningful getelementptrs where possible. 720 Value *Sum = nullptr; 721 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 722 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 723 const Loop *CurLoop = I->first; 724 const SCEV *Op = I->second; 725 if (!Sum) { 726 // This is the first operand. Just expand it. 727 Sum = expand(Op); 728 ++I; 729 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 730 // The running sum expression is a pointer. Try to form a getelementptr 731 // at this level with that as the base. 732 SmallVector<const SCEV *, 4> NewOps; 733 for (; I != E && I->first == CurLoop; ++I) { 734 // If the operand is SCEVUnknown and not instructions, peek through 735 // it, to enable more of it to be folded into the GEP. 736 const SCEV *X = I->second; 737 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 738 if (!isa<Instruction>(U->getValue())) 739 X = SE.getSCEV(U->getValue()); 740 NewOps.push_back(X); 741 } 742 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 743 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 744 // The running sum is an integer, and there's a pointer at this level. 745 // Try to form a getelementptr. If the running sum is instructions, 746 // use a SCEVUnknown to avoid re-analyzing them. 747 SmallVector<const SCEV *, 4> NewOps; 748 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 749 SE.getSCEV(Sum)); 750 for (++I; I != E && I->first == CurLoop; ++I) 751 NewOps.push_back(I->second); 752 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 753 } else if (Op->isNonConstantNegative()) { 754 // Instead of doing a negate and add, just do a subtract. 755 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 756 Sum = InsertNoopCastOfTo(Sum, Ty); 757 Sum = InsertBinop(Instruction::Sub, Sum, W); 758 ++I; 759 } else { 760 // A simple add. 761 Value *W = expandCodeFor(Op, Ty); 762 Sum = InsertNoopCastOfTo(Sum, Ty); 763 // Canonicalize a constant to the RHS. 764 if (isa<Constant>(Sum)) std::swap(Sum, W); 765 Sum = InsertBinop(Instruction::Add, Sum, W); 766 ++I; 767 } 768 } 769 770 return Sum; 771} 772 773Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 774 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 775 776 // Collect all the mul operands in a loop, along with their associated loops. 777 // Iterate in reverse so that constants are emitted last, all else equal. 778 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 779 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 780 E(S->op_begin()); I != E; ++I) 781 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 782 783 // Sort by loop. Use a stable sort so that constants follow non-constants. 784 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 785 786 // Emit instructions to mul all the operands. Hoist as much as possible 787 // out of loops. 788 Value *Prod = nullptr; 789 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 790 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 791 const SCEV *Op = I->second; 792 if (!Prod) { 793 // This is the first operand. Just expand it. 794 Prod = expand(Op); 795 ++I; 796 } else if (Op->isAllOnesValue()) { 797 // Instead of doing a multiply by negative one, just do a negate. 798 Prod = InsertNoopCastOfTo(Prod, Ty); 799 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 800 ++I; 801 } else { 802 // A simple mul. 803 Value *W = expandCodeFor(Op, Ty); 804 Prod = InsertNoopCastOfTo(Prod, Ty); 805 // Canonicalize a constant to the RHS. 806 if (isa<Constant>(Prod)) std::swap(Prod, W); 807 Prod = InsertBinop(Instruction::Mul, Prod, W); 808 ++I; 809 } 810 } 811 812 return Prod; 813} 814 815Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 816 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 817 818 Value *LHS = expandCodeFor(S->getLHS(), Ty); 819 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 820 const APInt &RHS = SC->getValue()->getValue(); 821 if (RHS.isPowerOf2()) 822 return InsertBinop(Instruction::LShr, LHS, 823 ConstantInt::get(Ty, RHS.logBase2())); 824 } 825 826 Value *RHS = expandCodeFor(S->getRHS(), Ty); 827 return InsertBinop(Instruction::UDiv, LHS, RHS); 828} 829 830/// Move parts of Base into Rest to leave Base with the minimal 831/// expression that provides a pointer operand suitable for a 832/// GEP expansion. 833static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 834 ScalarEvolution &SE) { 835 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 836 Base = A->getStart(); 837 Rest = SE.getAddExpr(Rest, 838 SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 839 A->getStepRecurrence(SE), 840 A->getLoop(), 841 A->getNoWrapFlags(SCEV::FlagNW))); 842 } 843 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 844 Base = A->getOperand(A->getNumOperands()-1); 845 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 846 NewAddOps.back() = Rest; 847 Rest = SE.getAddExpr(NewAddOps); 848 ExposePointerBase(Base, Rest, SE); 849 } 850} 851 852/// Determine if this is a well-behaved chain of instructions leading back to 853/// the PHI. If so, it may be reused by expanded expressions. 854bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 855 const Loop *L) { 856 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 857 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 858 return false; 859 // If any of the operands don't dominate the insert position, bail. 860 // Addrec operands are always loop-invariant, so this can only happen 861 // if there are instructions which haven't been hoisted. 862 if (L == IVIncInsertLoop) { 863 for (User::op_iterator OI = IncV->op_begin()+1, 864 OE = IncV->op_end(); OI != OE; ++OI) 865 if (Instruction *OInst = dyn_cast<Instruction>(OI)) 866 if (!SE.DT->dominates(OInst, IVIncInsertPos)) 867 return false; 868 } 869 // Advance to the next instruction. 870 IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 871 if (!IncV) 872 return false; 873 874 if (IncV->mayHaveSideEffects()) 875 return false; 876 877 if (IncV != PN) 878 return true; 879 880 return isNormalAddRecExprPHI(PN, IncV, L); 881} 882 883/// getIVIncOperand returns an induction variable increment's induction 884/// variable operand. 885/// 886/// If allowScale is set, any type of GEP is allowed as long as the nonIV 887/// operands dominate InsertPos. 888/// 889/// If allowScale is not set, ensure that a GEP increment conforms to one of the 890/// simple patterns generated by getAddRecExprPHILiterally and 891/// expandAddtoGEP. If the pattern isn't recognized, return NULL. 892Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 893 Instruction *InsertPos, 894 bool allowScale) { 895 if (IncV == InsertPos) 896 return nullptr; 897 898 switch (IncV->getOpcode()) { 899 default: 900 return nullptr; 901 // Check for a simple Add/Sub or GEP of a loop invariant step. 902 case Instruction::Add: 903 case Instruction::Sub: { 904 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 905 if (!OInst || SE.DT->dominates(OInst, InsertPos)) 906 return dyn_cast<Instruction>(IncV->getOperand(0)); 907 return nullptr; 908 } 909 case Instruction::BitCast: 910 return dyn_cast<Instruction>(IncV->getOperand(0)); 911 case Instruction::GetElementPtr: 912 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end(); 913 I != E; ++I) { 914 if (isa<Constant>(*I)) 915 continue; 916 if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 917 if (!SE.DT->dominates(OInst, InsertPos)) 918 return nullptr; 919 } 920 if (allowScale) { 921 // allow any kind of GEP as long as it can be hoisted. 922 continue; 923 } 924 // This must be a pointer addition of constants (pretty), which is already 925 // handled, or some number of address-size elements (ugly). Ugly geps 926 // have 2 operands. i1* is used by the expander to represent an 927 // address-size element. 928 if (IncV->getNumOperands() != 2) 929 return nullptr; 930 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 931 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 932 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 933 return nullptr; 934 break; 935 } 936 return dyn_cast<Instruction>(IncV->getOperand(0)); 937 } 938} 939 940/// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 941/// it available to other uses in this loop. Recursively hoist any operands, 942/// until we reach a value that dominates InsertPos. 943bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 944 if (SE.DT->dominates(IncV, InsertPos)) 945 return true; 946 947 // InsertPos must itself dominate IncV so that IncV's new position satisfies 948 // its existing users. 949 if (isa<PHINode>(InsertPos) 950 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent())) 951 return false; 952 953 // Check that the chain of IV operands leading back to Phi can be hoisted. 954 SmallVector<Instruction*, 4> IVIncs; 955 for(;;) { 956 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 957 if (!Oper) 958 return false; 959 // IncV is safe to hoist. 960 IVIncs.push_back(IncV); 961 IncV = Oper; 962 if (SE.DT->dominates(IncV, InsertPos)) 963 break; 964 } 965 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(), 966 E = IVIncs.rend(); I != E; ++I) { 967 (*I)->moveBefore(InsertPos); 968 } 969 return true; 970} 971 972/// Determine if this cyclic phi is in a form that would have been generated by 973/// LSR. We don't care if the phi was actually expanded in this pass, as long 974/// as it is in a low-cost form, for example, no implied multiplication. This 975/// should match any patterns generated by getAddRecExprPHILiterally and 976/// expandAddtoGEP. 977bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 978 const Loop *L) { 979 for(Instruction *IVOper = IncV; 980 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 981 /*allowScale=*/false));) { 982 if (IVOper == PN) 983 return true; 984 } 985 return false; 986} 987 988/// expandIVInc - Expand an IV increment at Builder's current InsertPos. 989/// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 990/// need to materialize IV increments elsewhere to handle difficult situations. 991Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 992 Type *ExpandTy, Type *IntTy, 993 bool useSubtract) { 994 Value *IncV; 995 // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 996 if (ExpandTy->isPointerTy()) { 997 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 998 // If the step isn't constant, don't use an implicitly scaled GEP, because 999 // that would require a multiply inside the loop. 1000 if (!isa<ConstantInt>(StepV)) 1001 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 1002 GEPPtrTy->getAddressSpace()); 1003 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 1004 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 1005 if (IncV->getType() != PN->getType()) { 1006 IncV = Builder.CreateBitCast(IncV, PN->getType()); 1007 rememberInstruction(IncV); 1008 } 1009 } else { 1010 IncV = useSubtract ? 1011 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 1012 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 1013 rememberInstruction(IncV); 1014 } 1015 return IncV; 1016} 1017 1018/// \brief Hoist the addrec instruction chain rooted in the loop phi above the 1019/// position. This routine assumes that this is possible (has been checked). 1020static void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist, 1021 Instruction *Pos, PHINode *LoopPhi) { 1022 do { 1023 if (DT->dominates(InstToHoist, Pos)) 1024 break; 1025 // Make sure the increment is where we want it. But don't move it 1026 // down past a potential existing post-inc user. 1027 InstToHoist->moveBefore(Pos); 1028 Pos = InstToHoist; 1029 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0)); 1030 } while (InstToHoist != LoopPhi); 1031} 1032 1033/// \brief Check whether we can cheaply express the requested SCEV in terms of 1034/// the available PHI SCEV by truncation and/or invertion of the step. 1035static bool canBeCheaplyTransformed(ScalarEvolution &SE, 1036 const SCEVAddRecExpr *Phi, 1037 const SCEVAddRecExpr *Requested, 1038 bool &InvertStep) { 1039 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType()); 1040 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType()); 1041 1042 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth()) 1043 return false; 1044 1045 // Try truncate it if necessary. 1046 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy)); 1047 if (!Phi) 1048 return false; 1049 1050 // Check whether truncation will help. 1051 if (Phi == Requested) { 1052 InvertStep = false; 1053 return true; 1054 } 1055 1056 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}. 1057 if (SE.getAddExpr(Requested->getStart(), 1058 SE.getNegativeSCEV(Requested)) == Phi) { 1059 InvertStep = true; 1060 return true; 1061 } 1062 1063 return false; 1064} 1065 1066/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 1067/// the base addrec, which is the addrec without any non-loop-dominating 1068/// values, and return the PHI. 1069PHINode * 1070SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 1071 const Loop *L, 1072 Type *ExpandTy, 1073 Type *IntTy, 1074 Type *&TruncTy, 1075 bool &InvertStep) { 1076 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 1077 1078 // Reuse a previously-inserted PHI, if present. 1079 BasicBlock *LatchBlock = L->getLoopLatch(); 1080 if (LatchBlock) { 1081 PHINode *AddRecPhiMatch = nullptr; 1082 Instruction *IncV = nullptr; 1083 TruncTy = nullptr; 1084 InvertStep = false; 1085 1086 // Only try partially matching scevs that need truncation and/or 1087 // step-inversion if we know this loop is outside the current loop. 1088 bool TryNonMatchingSCEV = IVIncInsertLoop && 1089 SE.DT->properlyDominates(LatchBlock, IVIncInsertLoop->getHeader()); 1090 1091 for (BasicBlock::iterator I = L->getHeader()->begin(); 1092 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1093 if (!SE.isSCEVable(PN->getType())) 1094 continue; 1095 1096 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN)); 1097 if (!PhiSCEV) 1098 continue; 1099 1100 bool IsMatchingSCEV = PhiSCEV == Normalized; 1101 // We only handle truncation and inversion of phi recurrences for the 1102 // expanded expression if the expanded expression's loop dominates the 1103 // loop we insert to. Check now, so we can bail out early. 1104 if (!IsMatchingSCEV && !TryNonMatchingSCEV) 1105 continue; 1106 1107 Instruction *TempIncV = 1108 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 1109 1110 // Check whether we can reuse this PHI node. 1111 if (LSRMode) { 1112 if (!isExpandedAddRecExprPHI(PN, TempIncV, L)) 1113 continue; 1114 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos)) 1115 continue; 1116 } else { 1117 if (!isNormalAddRecExprPHI(PN, TempIncV, L)) 1118 continue; 1119 } 1120 1121 // Stop if we have found an exact match SCEV. 1122 if (IsMatchingSCEV) { 1123 IncV = TempIncV; 1124 TruncTy = nullptr; 1125 InvertStep = false; 1126 AddRecPhiMatch = PN; 1127 break; 1128 } 1129 1130 // Try whether the phi can be translated into the requested form 1131 // (truncated and/or offset by a constant). 1132 if ((!TruncTy || InvertStep) && 1133 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) { 1134 // Record the phi node. But don't stop we might find an exact match 1135 // later. 1136 AddRecPhiMatch = PN; 1137 IncV = TempIncV; 1138 TruncTy = SE.getEffectiveSCEVType(Normalized->getType()); 1139 } 1140 } 1141 1142 if (AddRecPhiMatch) { 1143 // Potentially, move the increment. We have made sure in 1144 // isExpandedAddRecExprPHI or hoistIVInc that this is possible. 1145 if (L == IVIncInsertLoop) 1146 hoistBeforePos(SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch); 1147 1148 // Ok, the add recurrence looks usable. 1149 // Remember this PHI, even in post-inc mode. 1150 InsertedValues.insert(AddRecPhiMatch); 1151 // Remember the increment. 1152 rememberInstruction(IncV); 1153 return AddRecPhiMatch; 1154 } 1155 } 1156 1157 // Save the original insertion point so we can restore it when we're done. 1158 BuilderType::InsertPointGuard Guard(Builder); 1159 1160 // Another AddRec may need to be recursively expanded below. For example, if 1161 // this AddRec is quadratic, the StepV may itself be an AddRec in this 1162 // loop. Remove this loop from the PostIncLoops set before expanding such 1163 // AddRecs. Otherwise, we cannot find a valid position for the step 1164 // (i.e. StepV can never dominate its loop header). Ideally, we could do 1165 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 1166 // so it's not worth implementing SmallPtrSet::swap. 1167 PostIncLoopSet SavedPostIncLoops = PostIncLoops; 1168 PostIncLoops.clear(); 1169 1170 // Expand code for the start value. 1171 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 1172 L->getHeader()->begin()); 1173 1174 // StartV must be hoisted into L's preheader to dominate the new phi. 1175 assert(!isa<Instruction>(StartV) || 1176 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(), 1177 L->getHeader())); 1178 1179 // Expand code for the step value. Do this before creating the PHI so that PHI 1180 // reuse code doesn't see an incomplete PHI. 1181 const SCEV *Step = Normalized->getStepRecurrence(SE); 1182 // If the stride is negative, insert a sub instead of an add for the increment 1183 // (unless it's a constant, because subtracts of constants are canonicalized 1184 // to adds). 1185 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1186 if (useSubtract) 1187 Step = SE.getNegativeSCEV(Step); 1188 // Expand the step somewhere that dominates the loop header. 1189 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1190 1191 // Create the PHI. 1192 BasicBlock *Header = L->getHeader(); 1193 Builder.SetInsertPoint(Header, Header->begin()); 1194 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1195 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 1196 Twine(IVName) + ".iv"); 1197 rememberInstruction(PN); 1198 1199 // Create the step instructions and populate the PHI. 1200 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1201 BasicBlock *Pred = *HPI; 1202 1203 // Add a start value. 1204 if (!L->contains(Pred)) { 1205 PN->addIncoming(StartV, Pred); 1206 continue; 1207 } 1208 1209 // Create a step value and add it to the PHI. 1210 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 1211 // instructions at IVIncInsertPos. 1212 Instruction *InsertPos = L == IVIncInsertLoop ? 1213 IVIncInsertPos : Pred->getTerminator(); 1214 Builder.SetInsertPoint(InsertPos); 1215 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1216 if (isa<OverflowingBinaryOperator>(IncV)) { 1217 if (Normalized->getNoWrapFlags(SCEV::FlagNUW)) 1218 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap(); 1219 if (Normalized->getNoWrapFlags(SCEV::FlagNSW)) 1220 cast<BinaryOperator>(IncV)->setHasNoSignedWrap(); 1221 } 1222 PN->addIncoming(IncV, Pred); 1223 } 1224 1225 // After expanding subexpressions, restore the PostIncLoops set so the caller 1226 // can ensure that IVIncrement dominates the current uses. 1227 PostIncLoops = SavedPostIncLoops; 1228 1229 // Remember this PHI, even in post-inc mode. 1230 InsertedValues.insert(PN); 1231 1232 return PN; 1233} 1234 1235Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 1236 Type *STy = S->getType(); 1237 Type *IntTy = SE.getEffectiveSCEVType(STy); 1238 const Loop *L = S->getLoop(); 1239 1240 // Determine a normalized form of this expression, which is the expression 1241 // before any post-inc adjustment is made. 1242 const SCEVAddRecExpr *Normalized = S; 1243 if (PostIncLoops.count(L)) { 1244 PostIncLoopSet Loops; 1245 Loops.insert(L); 1246 Normalized = 1247 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, nullptr, 1248 nullptr, Loops, SE, *SE.DT)); 1249 } 1250 1251 // Strip off any non-loop-dominating component from the addrec start. 1252 const SCEV *Start = Normalized->getStart(); 1253 const SCEV *PostLoopOffset = nullptr; 1254 if (!SE.properlyDominates(Start, L->getHeader())) { 1255 PostLoopOffset = Start; 1256 Start = SE.getConstant(Normalized->getType(), 0); 1257 Normalized = cast<SCEVAddRecExpr>( 1258 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 1259 Normalized->getLoop(), 1260 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1261 } 1262 1263 // Strip off any non-loop-dominating component from the addrec step. 1264 const SCEV *Step = Normalized->getStepRecurrence(SE); 1265 const SCEV *PostLoopScale = nullptr; 1266 if (!SE.dominates(Step, L->getHeader())) { 1267 PostLoopScale = Step; 1268 Step = SE.getConstant(Normalized->getType(), 1); 1269 Normalized = 1270 cast<SCEVAddRecExpr>(SE.getAddRecExpr( 1271 Start, Step, Normalized->getLoop(), 1272 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1273 } 1274 1275 // Expand the core addrec. If we need post-loop scaling, force it to 1276 // expand to an integer type to avoid the need for additional casting. 1277 Type *ExpandTy = PostLoopScale ? IntTy : STy; 1278 // In some cases, we decide to reuse an existing phi node but need to truncate 1279 // it and/or invert the step. 1280 Type *TruncTy = nullptr; 1281 bool InvertStep = false; 1282 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy, 1283 TruncTy, InvertStep); 1284 1285 // Accommodate post-inc mode, if necessary. 1286 Value *Result; 1287 if (!PostIncLoops.count(L)) 1288 Result = PN; 1289 else { 1290 // In PostInc mode, use the post-incremented value. 1291 BasicBlock *LatchBlock = L->getLoopLatch(); 1292 assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1293 Result = PN->getIncomingValueForBlock(LatchBlock); 1294 1295 // For an expansion to use the postinc form, the client must call 1296 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 1297 // or dominated by IVIncInsertPos. 1298 if (isa<Instruction>(Result) 1299 && !SE.DT->dominates(cast<Instruction>(Result), 1300 Builder.GetInsertPoint())) { 1301 // The induction variable's postinc expansion does not dominate this use. 1302 // IVUsers tries to prevent this case, so it is rare. However, it can 1303 // happen when an IVUser outside the loop is not dominated by the latch 1304 // block. Adjusting IVIncInsertPos before expansion begins cannot handle 1305 // all cases. Consider a phi outide whose operand is replaced during 1306 // expansion with the value of the postinc user. Without fundamentally 1307 // changing the way postinc users are tracked, the only remedy is 1308 // inserting an extra IV increment. StepV might fold into PostLoopOffset, 1309 // but hopefully expandCodeFor handles that. 1310 bool useSubtract = 1311 !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1312 if (useSubtract) 1313 Step = SE.getNegativeSCEV(Step); 1314 Value *StepV; 1315 { 1316 // Expand the step somewhere that dominates the loop header. 1317 BuilderType::InsertPointGuard Guard(Builder); 1318 StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1319 } 1320 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1321 } 1322 } 1323 1324 // We have decided to reuse an induction variable of a dominating loop. Apply 1325 // truncation and/or invertion of the step. 1326 if (TruncTy) { 1327 Type *ResTy = Result->getType(); 1328 // Normalize the result type. 1329 if (ResTy != SE.getEffectiveSCEVType(ResTy)) 1330 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy)); 1331 // Truncate the result. 1332 if (TruncTy != Result->getType()) { 1333 Result = Builder.CreateTrunc(Result, TruncTy); 1334 rememberInstruction(Result); 1335 } 1336 // Invert the result. 1337 if (InvertStep) { 1338 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy), 1339 Result); 1340 rememberInstruction(Result); 1341 } 1342 } 1343 1344 // Re-apply any non-loop-dominating scale. 1345 if (PostLoopScale) { 1346 assert(S->isAffine() && "Can't linearly scale non-affine recurrences."); 1347 Result = InsertNoopCastOfTo(Result, IntTy); 1348 Result = Builder.CreateMul(Result, 1349 expandCodeFor(PostLoopScale, IntTy)); 1350 rememberInstruction(Result); 1351 } 1352 1353 // Re-apply any non-loop-dominating offset. 1354 if (PostLoopOffset) { 1355 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1356 const SCEV *const OffsetArray[1] = { PostLoopOffset }; 1357 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 1358 } else { 1359 Result = InsertNoopCastOfTo(Result, IntTy); 1360 Result = Builder.CreateAdd(Result, 1361 expandCodeFor(PostLoopOffset, IntTy)); 1362 rememberInstruction(Result); 1363 } 1364 } 1365 1366 return Result; 1367} 1368 1369Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1370 if (!CanonicalMode) return expandAddRecExprLiterally(S); 1371 1372 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1373 const Loop *L = S->getLoop(); 1374 1375 // First check for an existing canonical IV in a suitable type. 1376 PHINode *CanonicalIV = nullptr; 1377 if (PHINode *PN = L->getCanonicalInductionVariable()) 1378 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 1379 CanonicalIV = PN; 1380 1381 // Rewrite an AddRec in terms of the canonical induction variable, if 1382 // its type is more narrow. 1383 if (CanonicalIV && 1384 SE.getTypeSizeInBits(CanonicalIV->getType()) > 1385 SE.getTypeSizeInBits(Ty)) { 1386 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1387 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1388 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 1389 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 1390 S->getNoWrapFlags(SCEV::FlagNW))); 1391 BasicBlock::iterator NewInsertPt = 1392 std::next(BasicBlock::iterator(cast<Instruction>(V))); 1393 BuilderType::InsertPointGuard Guard(Builder); 1394 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) || 1395 isa<LandingPadInst>(NewInsertPt)) 1396 ++NewInsertPt; 1397 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr, 1398 NewInsertPt); 1399 return V; 1400 } 1401 1402 // {X,+,F} --> X + {0,+,F} 1403 if (!S->getStart()->isZero()) { 1404 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1405 NewOps[0] = SE.getConstant(Ty, 0); 1406 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, 1407 S->getNoWrapFlags(SCEV::FlagNW)); 1408 1409 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1410 // comments on expandAddToGEP for details. 1411 const SCEV *Base = S->getStart(); 1412 const SCEV *RestArray[1] = { Rest }; 1413 // Dig into the expression to find the pointer base for a GEP. 1414 ExposePointerBase(Base, RestArray[0], SE); 1415 // If we found a pointer, expand the AddRec with a GEP. 1416 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1417 // Make sure the Base isn't something exotic, such as a multiplied 1418 // or divided pointer value. In those cases, the result type isn't 1419 // actually a pointer type. 1420 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1421 Value *StartV = expand(Base); 1422 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1423 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 1424 } 1425 } 1426 1427 // Just do a normal add. Pre-expand the operands to suppress folding. 1428 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), 1429 SE.getUnknown(expand(Rest)))); 1430 } 1431 1432 // If we don't yet have a canonical IV, create one. 1433 if (!CanonicalIV) { 1434 // Create and insert the PHI node for the induction variable in the 1435 // specified loop. 1436 BasicBlock *Header = L->getHeader(); 1437 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1438 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 1439 Header->begin()); 1440 rememberInstruction(CanonicalIV); 1441 1442 SmallSet<BasicBlock *, 4> PredSeen; 1443 Constant *One = ConstantInt::get(Ty, 1); 1444 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1445 BasicBlock *HP = *HPI; 1446 if (!PredSeen.insert(HP).second) { 1447 // There must be an incoming value for each predecessor, even the 1448 // duplicates! 1449 CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP); 1450 continue; 1451 } 1452 1453 if (L->contains(HP)) { 1454 // Insert a unit add instruction right before the terminator 1455 // corresponding to the back-edge. 1456 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 1457 "indvar.next", 1458 HP->getTerminator()); 1459 Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 1460 rememberInstruction(Add); 1461 CanonicalIV->addIncoming(Add, HP); 1462 } else { 1463 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 1464 } 1465 } 1466 } 1467 1468 // {0,+,1} --> Insert a canonical induction variable into the loop! 1469 if (S->isAffine() && S->getOperand(1)->isOne()) { 1470 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 1471 "IVs with types different from the canonical IV should " 1472 "already have been handled!"); 1473 return CanonicalIV; 1474 } 1475 1476 // {0,+,F} --> {0,+,1} * F 1477 1478 // If this is a simple linear addrec, emit it now as a special case. 1479 if (S->isAffine()) // {0,+,F} --> i*F 1480 return 1481 expand(SE.getTruncateOrNoop( 1482 SE.getMulExpr(SE.getUnknown(CanonicalIV), 1483 SE.getNoopOrAnyExtend(S->getOperand(1), 1484 CanonicalIV->getType())), 1485 Ty)); 1486 1487 // If this is a chain of recurrences, turn it into a closed form, using the 1488 // folders, then expandCodeFor the closed form. This allows the folders to 1489 // simplify the expression without having to build a bunch of special code 1490 // into this folder. 1491 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 1492 1493 // Promote S up to the canonical IV type, if the cast is foldable. 1494 const SCEV *NewS = S; 1495 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 1496 if (isa<SCEVAddRecExpr>(Ext)) 1497 NewS = Ext; 1498 1499 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1500 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 1501 1502 // Truncate the result down to the original type, if needed. 1503 const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1504 return expand(T); 1505} 1506 1507Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1508 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1509 Value *V = expandCodeFor(S->getOperand(), 1510 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1511 Value *I = Builder.CreateTrunc(V, Ty); 1512 rememberInstruction(I); 1513 return I; 1514} 1515 1516Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1517 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1518 Value *V = expandCodeFor(S->getOperand(), 1519 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1520 Value *I = Builder.CreateZExt(V, Ty); 1521 rememberInstruction(I); 1522 return I; 1523} 1524 1525Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1526 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1527 Value *V = expandCodeFor(S->getOperand(), 1528 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1529 Value *I = Builder.CreateSExt(V, Ty); 1530 rememberInstruction(I); 1531 return I; 1532} 1533 1534Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 1535 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1536 Type *Ty = LHS->getType(); 1537 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1538 // In the case of mixed integer and pointer types, do the 1539 // rest of the comparisons as integer. 1540 if (S->getOperand(i)->getType() != Ty) { 1541 Ty = SE.getEffectiveSCEVType(Ty); 1542 LHS = InsertNoopCastOfTo(LHS, Ty); 1543 } 1544 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1545 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 1546 rememberInstruction(ICmp); 1547 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1548 rememberInstruction(Sel); 1549 LHS = Sel; 1550 } 1551 // In the case of mixed integer and pointer types, cast the 1552 // final result back to the pointer type. 1553 if (LHS->getType() != S->getType()) 1554 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1555 return LHS; 1556} 1557 1558Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 1559 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1560 Type *Ty = LHS->getType(); 1561 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1562 // In the case of mixed integer and pointer types, do the 1563 // rest of the comparisons as integer. 1564 if (S->getOperand(i)->getType() != Ty) { 1565 Ty = SE.getEffectiveSCEVType(Ty); 1566 LHS = InsertNoopCastOfTo(LHS, Ty); 1567 } 1568 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1569 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 1570 rememberInstruction(ICmp); 1571 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1572 rememberInstruction(Sel); 1573 LHS = Sel; 1574 } 1575 // In the case of mixed integer and pointer types, cast the 1576 // final result back to the pointer type. 1577 if (LHS->getType() != S->getType()) 1578 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1579 return LHS; 1580} 1581 1582Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 1583 Instruction *IP) { 1584 Builder.SetInsertPoint(IP->getParent(), IP); 1585 return expandCodeFor(SH, Ty); 1586} 1587 1588Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 1589 // Expand the code for this SCEV. 1590 Value *V = expand(SH); 1591 if (Ty) { 1592 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 1593 "non-trivial casts should be done with the SCEVs directly!"); 1594 V = InsertNoopCastOfTo(V, Ty); 1595 } 1596 return V; 1597} 1598 1599Value *SCEVExpander::expand(const SCEV *S) { 1600 // Compute an insertion point for this SCEV object. Hoist the instructions 1601 // as far out in the loop nest as possible. 1602 Instruction *InsertPt = Builder.GetInsertPoint(); 1603 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ; 1604 L = L->getParentLoop()) 1605 if (SE.isLoopInvariant(S, L)) { 1606 if (!L) break; 1607 if (BasicBlock *Preheader = L->getLoopPreheader()) 1608 InsertPt = Preheader->getTerminator(); 1609 else { 1610 // LSR sets the insertion point for AddRec start/step values to the 1611 // block start to simplify value reuse, even though it's an invalid 1612 // position. SCEVExpander must correct for this in all cases. 1613 InsertPt = L->getHeader()->getFirstInsertionPt(); 1614 } 1615 } else { 1616 // If the SCEV is computable at this level, insert it into the header 1617 // after the PHIs (and after any other instructions that we've inserted 1618 // there) so that it is guaranteed to dominate any user inside the loop. 1619 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 1620 InsertPt = L->getHeader()->getFirstInsertionPt(); 1621 while (InsertPt != Builder.GetInsertPoint() 1622 && (isInsertedInstruction(InsertPt) 1623 || isa<DbgInfoIntrinsic>(InsertPt))) { 1624 InsertPt = std::next(BasicBlock::iterator(InsertPt)); 1625 } 1626 break; 1627 } 1628 1629 // Check to see if we already expanded this here. 1630 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator 1631 I = InsertedExpressions.find(std::make_pair(S, InsertPt)); 1632 if (I != InsertedExpressions.end()) 1633 return I->second; 1634 1635 BuilderType::InsertPointGuard Guard(Builder); 1636 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt); 1637 1638 // Expand the expression into instructions. 1639 Value *V = visit(S); 1640 1641 // Remember the expanded value for this SCEV at this location. 1642 // 1643 // This is independent of PostIncLoops. The mapped value simply materializes 1644 // the expression at this insertion point. If the mapped value happened to be 1645 // a postinc expansion, it could be reused by a non-postinc user, but only if 1646 // its insertion point was already at the head of the loop. 1647 InsertedExpressions[std::make_pair(S, InsertPt)] = V; 1648 return V; 1649} 1650 1651void SCEVExpander::rememberInstruction(Value *I) { 1652 if (!PostIncLoops.empty()) 1653 InsertedPostIncValues.insert(I); 1654 else 1655 InsertedValues.insert(I); 1656} 1657 1658/// getOrInsertCanonicalInductionVariable - This method returns the 1659/// canonical induction variable of the specified type for the specified 1660/// loop (inserting one if there is none). A canonical induction variable 1661/// starts at zero and steps by one on each iteration. 1662PHINode * 1663SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1664 Type *Ty) { 1665 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1666 1667 // Build a SCEV for {0,+,1}<L>. 1668 // Conservatively use FlagAnyWrap for now. 1669 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 1670 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 1671 1672 // Emit code for it. 1673 BuilderType::InsertPointGuard Guard(Builder); 1674 PHINode *V = cast<PHINode>(expandCodeFor(H, nullptr, 1675 L->getHeader()->begin())); 1676 1677 return V; 1678} 1679 1680/// replaceCongruentIVs - Check for congruent phis in this loop header and 1681/// replace them with their most canonical representative. Return the number of 1682/// phis eliminated. 1683/// 1684/// This does not depend on any SCEVExpander state but should be used in 1685/// the same context that SCEVExpander is used. 1686unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 1687 SmallVectorImpl<WeakVH> &DeadInsts, 1688 const TargetTransformInfo *TTI) { 1689 // Find integer phis in order of increasing width. 1690 SmallVector<PHINode*, 8> Phis; 1691 for (BasicBlock::iterator I = L->getHeader()->begin(); 1692 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 1693 Phis.push_back(Phi); 1694 } 1695 if (TTI) 1696 std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) { 1697 // Put pointers at the back and make sure pointer < pointer = false. 1698 if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy()) 1699 return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy(); 1700 return RHS->getType()->getPrimitiveSizeInBits() < 1701 LHS->getType()->getPrimitiveSizeInBits(); 1702 }); 1703 1704 unsigned NumElim = 0; 1705 DenseMap<const SCEV *, PHINode *> ExprToIVMap; 1706 // Process phis from wide to narrow. Mapping wide phis to the their truncation 1707 // so narrow phis can reuse them. 1708 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(), 1709 PEnd = Phis.end(); PIter != PEnd; ++PIter) { 1710 PHINode *Phi = *PIter; 1711 1712 // Fold constant phis. They may be congruent to other constant phis and 1713 // would confuse the logic below that expects proper IVs. 1714 if (Value *V = SimplifyInstruction(Phi, SE.DL, SE.TLI, SE.DT, SE.AC)) { 1715 Phi->replaceAllUsesWith(V); 1716 DeadInsts.push_back(Phi); 1717 ++NumElim; 1718 DEBUG_WITH_TYPE(DebugType, dbgs() 1719 << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 1720 continue; 1721 } 1722 1723 if (!SE.isSCEVable(Phi->getType())) 1724 continue; 1725 1726 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 1727 if (!OrigPhiRef) { 1728 OrigPhiRef = Phi; 1729 if (Phi->getType()->isIntegerTy() && TTI 1730 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 1731 // This phi can be freely truncated to the narrowest phi type. Map the 1732 // truncated expression to it so it will be reused for narrow types. 1733 const SCEV *TruncExpr = 1734 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 1735 ExprToIVMap[TruncExpr] = Phi; 1736 } 1737 continue; 1738 } 1739 1740 // Replacing a pointer phi with an integer phi or vice-versa doesn't make 1741 // sense. 1742 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 1743 continue; 1744 1745 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1746 Instruction *OrigInc = 1747 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 1748 Instruction *IsomorphicInc = 1749 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 1750 1751 // If this phi has the same width but is more canonical, replace the 1752 // original with it. As part of the "more canonical" determination, 1753 // respect a prior decision to use an IV chain. 1754 if (OrigPhiRef->getType() == Phi->getType() 1755 && !(ChainedPhis.count(Phi) 1756 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) 1757 && (ChainedPhis.count(Phi) 1758 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 1759 std::swap(OrigPhiRef, Phi); 1760 std::swap(OrigInc, IsomorphicInc); 1761 } 1762 // Replacing the congruent phi is sufficient because acyclic redundancy 1763 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves 1764 // that a phi is congruent, it's often the head of an IV user cycle that 1765 // is isomorphic with the original phi. It's worth eagerly cleaning up the 1766 // common case of a single IV increment so that DeleteDeadPHIs can remove 1767 // cycles that had postinc uses. 1768 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc), 1769 IsomorphicInc->getType()); 1770 if (OrigInc != IsomorphicInc 1771 && TruncExpr == SE.getSCEV(IsomorphicInc) 1772 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc)) 1773 || hoistIVInc(OrigInc, IsomorphicInc))) { 1774 DEBUG_WITH_TYPE(DebugType, dbgs() 1775 << "INDVARS: Eliminated congruent iv.inc: " 1776 << *IsomorphicInc << '\n'); 1777 Value *NewInc = OrigInc; 1778 if (OrigInc->getType() != IsomorphicInc->getType()) {
|