1//===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis ------------===// 2// 3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4// See https://llvm.org/LICENSE.txt for license information. 5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6// 7//===----------------------------------------------------------------------===// 8// 9// This file contains the implementation of the scalar evolution expander, 10// which is used to generate the code corresponding to a given scalar evolution 11// expression. 12// 13//===----------------------------------------------------------------------===// 14 15#include "llvm/Analysis/ScalarEvolutionExpander.h" 16#include "llvm/ADT/STLExtras.h" 17#include "llvm/ADT/SmallSet.h" 18#include "llvm/Analysis/InstructionSimplify.h" 19#include "llvm/Analysis/LoopInfo.h" 20#include "llvm/Analysis/TargetTransformInfo.h" 21#include "llvm/IR/DataLayout.h" 22#include "llvm/IR/Dominators.h" 23#include "llvm/IR/IntrinsicInst.h" 24#include "llvm/IR/LLVMContext.h" 25#include "llvm/IR/Module.h" 26#include "llvm/IR/PatternMatch.h" 27#include "llvm/Support/Debug.h" 28#include "llvm/Support/raw_ostream.h" 29 30using namespace llvm; 31using namespace PatternMatch; 32 33/// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, 34/// reusing an existing cast if a suitable one exists, moving an existing 35/// cast if a suitable one exists but isn't in the right place, or 36/// creating a new one. 37Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, 38 Instruction::CastOps Op, 39 BasicBlock::iterator IP) { 40 // This function must be called with the builder having a valid insertion 41 // point. It doesn't need to be the actual IP where the uses of the returned 42 // cast will be added, but it must dominate such IP. 43 // We use this precondition to produce a cast that will dominate all its 44 // uses. In particular, this is crucial for the case where the builder's 45 // insertion point *is* the point where we were asked to put the cast. 46 // Since we don't know the builder's insertion point is actually 47 // where the uses will be added (only that it dominates it), we are 48 // not allowed to move it. 49 BasicBlock::iterator BIP = Builder.GetInsertPoint(); 50 51 Instruction *Ret = nullptr; 52 53 // Check to see if there is already a cast! 54 for (User *U : V->users()) 55 if (U->getType() == Ty) 56 if (CastInst *CI = dyn_cast<CastInst>(U)) 57 if (CI->getOpcode() == Op) { 58 // If the cast isn't where we want it, create a new cast at IP. 59 // Likewise, do not reuse a cast at BIP because it must dominate 60 // instructions that might be inserted before BIP. 61 if (BasicBlock::iterator(CI) != IP || BIP == IP) { 62 // Create a new cast, and leave the old cast in place in case 63 // it is being used as an insert point. 64 Ret = CastInst::Create(Op, V, Ty, "", &*IP); 65 Ret->takeName(CI); 66 CI->replaceAllUsesWith(Ret); 67 break; 68 } 69 Ret = CI; 70 break; 71 } 72 73 // Create a new cast. 74 if (!Ret) 75 Ret = CastInst::Create(Op, V, Ty, V->getName(), &*IP); 76 77 // We assert at the end of the function since IP might point to an 78 // instruction with different dominance properties than a cast 79 // (an invoke for example) and not dominate BIP (but the cast does). 80 assert(SE.DT.dominates(Ret, &*BIP)); 81 82 rememberInstruction(Ret); 83 return Ret; 84} 85 86static BasicBlock::iterator findInsertPointAfter(Instruction *I, 87 BasicBlock *MustDominate) { 88 BasicBlock::iterator IP = ++I->getIterator(); 89 if (auto *II = dyn_cast<InvokeInst>(I)) 90 IP = II->getNormalDest()->begin(); 91 92 while (isa<PHINode>(IP)) 93 ++IP; 94 95 if (isa<FuncletPadInst>(IP) || isa<LandingPadInst>(IP)) { 96 ++IP; 97 } else if (isa<CatchSwitchInst>(IP)) { 98 IP = MustDominate->getFirstInsertionPt(); 99 } else { 100 assert(!IP->isEHPad() && "unexpected eh pad!"); 101 } 102 103 return IP; 104} 105 106/// InsertNoopCastOfTo - Insert a cast of V to the specified type, 107/// which must be possible with a noop cast, doing what we can to share 108/// the casts. 109Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 110 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 111 assert((Op == Instruction::BitCast || 112 Op == Instruction::PtrToInt || 113 Op == Instruction::IntToPtr) && 114 "InsertNoopCastOfTo cannot perform non-noop casts!"); 115 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 116 "InsertNoopCastOfTo cannot change sizes!"); 117 118 // Short-circuit unnecessary bitcasts. 119 if (Op == Instruction::BitCast) { 120 if (V->getType() == Ty) 121 return V; 122 if (CastInst *CI = dyn_cast<CastInst>(V)) { 123 if (CI->getOperand(0)->getType() == Ty) 124 return CI->getOperand(0); 125 } 126 } 127 // Short-circuit unnecessary inttoptr<->ptrtoint casts. 128 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 129 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 130 if (CastInst *CI = dyn_cast<CastInst>(V)) 131 if ((CI->getOpcode() == Instruction::PtrToInt || 132 CI->getOpcode() == Instruction::IntToPtr) && 133 SE.getTypeSizeInBits(CI->getType()) == 134 SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 135 return CI->getOperand(0); 136 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 137 if ((CE->getOpcode() == Instruction::PtrToInt || 138 CE->getOpcode() == Instruction::IntToPtr) && 139 SE.getTypeSizeInBits(CE->getType()) == 140 SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 141 return CE->getOperand(0); 142 } 143 144 // Fold a cast of a constant. 145 if (Constant *C = dyn_cast<Constant>(V)) 146 return ConstantExpr::getCast(Op, C, Ty); 147 148 // Cast the argument at the beginning of the entry block, after 149 // any bitcasts of other arguments. 150 if (Argument *A = dyn_cast<Argument>(V)) { 151 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 152 while ((isa<BitCastInst>(IP) && 153 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 154 cast<BitCastInst>(IP)->getOperand(0) != A) || 155 isa<DbgInfoIntrinsic>(IP)) 156 ++IP; 157 return ReuseOrCreateCast(A, Ty, Op, IP); 158 } 159 160 // Cast the instruction immediately after the instruction. 161 Instruction *I = cast<Instruction>(V); 162 BasicBlock::iterator IP = findInsertPointAfter(I, Builder.GetInsertBlock()); 163 return ReuseOrCreateCast(I, Ty, Op, IP); 164} 165 166/// InsertBinop - Insert the specified binary operator, doing a small amount 167/// of work to avoid inserting an obviously redundant operation, and hoisting 168/// to an outer loop when the opportunity is there and it is safe. 169Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 170 Value *LHS, Value *RHS, 171 SCEV::NoWrapFlags Flags, bool IsSafeToHoist) { 172 // Fold a binop with constant operands. 173 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 174 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 175 return ConstantExpr::get(Opcode, CLHS, CRHS); 176 177 // Do a quick scan to see if we have this binop nearby. If so, reuse it. 178 unsigned ScanLimit = 6; 179 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 180 // Scanning starts from the last instruction before the insertion point. 181 BasicBlock::iterator IP = Builder.GetInsertPoint(); 182 if (IP != BlockBegin) { 183 --IP; 184 for (; ScanLimit; --IP, --ScanLimit) { 185 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 186 // generated code. 187 if (isa<DbgInfoIntrinsic>(IP)) 188 ScanLimit++; 189 190 auto canGenerateIncompatiblePoison = [&Flags](Instruction *I) { 191 // Ensure that no-wrap flags match. 192 if (isa<OverflowingBinaryOperator>(I)) { 193 if (I->hasNoSignedWrap() != (Flags & SCEV::FlagNSW)) 194 return true; 195 if (I->hasNoUnsignedWrap() != (Flags & SCEV::FlagNUW)) 196 return true; 197 } 198 // Conservatively, do not use any instruction which has any of exact 199 // flags installed. 200 if (isa<PossiblyExactOperator>(I) && I->isExact()) 201 return true; 202 return false; 203 }; 204 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 205 IP->getOperand(1) == RHS && !canGenerateIncompatiblePoison(&*IP)) 206 return &*IP; 207 if (IP == BlockBegin) break; 208 } 209 } 210 211 // Save the original insertion point so we can restore it when we're done. 212 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); 213 SCEVInsertPointGuard Guard(Builder, this); 214 215 if (IsSafeToHoist) { 216 // Move the insertion point out of as many loops as we can. 217 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 218 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 219 BasicBlock *Preheader = L->getLoopPreheader(); 220 if (!Preheader) break; 221 222 // Ok, move up a level. 223 Builder.SetInsertPoint(Preheader->getTerminator()); 224 } 225 } 226 227 // If we haven't found this binop, insert it. 228 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 229 BO->setDebugLoc(Loc); 230 if (Flags & SCEV::FlagNUW) 231 BO->setHasNoUnsignedWrap(); 232 if (Flags & SCEV::FlagNSW) 233 BO->setHasNoSignedWrap(); 234 rememberInstruction(BO); 235 236 return BO; 237} 238 239/// FactorOutConstant - Test if S is divisible by Factor, using signed 240/// division. If so, update S with Factor divided out and return true. 241/// S need not be evenly divisible if a reasonable remainder can be 242/// computed. 243static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder, 244 const SCEV *Factor, ScalarEvolution &SE, 245 const DataLayout &DL) { 246 // Everything is divisible by one. 247 if (Factor->isOne()) 248 return true; 249 250 // x/x == 1. 251 if (S == Factor) { 252 S = SE.getConstant(S->getType(), 1); 253 return true; 254 } 255 256 // For a Constant, check for a multiple of the given factor. 257 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 258 // 0/x == 0. 259 if (C->isZero()) 260 return true; 261 // Check for divisibility. 262 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 263 ConstantInt *CI = 264 ConstantInt::get(SE.getContext(), C->getAPInt().sdiv(FC->getAPInt())); 265 // If the quotient is zero and the remainder is non-zero, reject 266 // the value at this scale. It will be considered for subsequent 267 // smaller scales. 268 if (!CI->isZero()) { 269 const SCEV *Div = SE.getConstant(CI); 270 S = Div; 271 Remainder = SE.getAddExpr( 272 Remainder, SE.getConstant(C->getAPInt().srem(FC->getAPInt()))); 273 return true; 274 } 275 } 276 } 277 278 // In a Mul, check if there is a constant operand which is a multiple 279 // of the given factor. 280 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 281 // Size is known, check if there is a constant operand which is a multiple 282 // of the given factor. If so, we can factor it. 283 const SCEVConstant *FC = cast<SCEVConstant>(Factor); 284 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 285 if (!C->getAPInt().srem(FC->getAPInt())) { 286 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 287 NewMulOps[0] = SE.getConstant(C->getAPInt().sdiv(FC->getAPInt())); 288 S = SE.getMulExpr(NewMulOps); 289 return true; 290 } 291 } 292 293 // In an AddRec, check if both start and step are divisible. 294 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 295 const SCEV *Step = A->getStepRecurrence(SE); 296 const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 297 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL)) 298 return false; 299 if (!StepRem->isZero()) 300 return false; 301 const SCEV *Start = A->getStart(); 302 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL)) 303 return false; 304 S = SE.getAddRecExpr(Start, Step, A->getLoop(), 305 A->getNoWrapFlags(SCEV::FlagNW)); 306 return true; 307 } 308 309 return false; 310} 311 312/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 313/// is the number of SCEVAddRecExprs present, which are kept at the end of 314/// the list. 315/// 316static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 317 Type *Ty, 318 ScalarEvolution &SE) { 319 unsigned NumAddRecs = 0; 320 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 321 ++NumAddRecs; 322 // Group Ops into non-addrecs and addrecs. 323 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 324 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 325 // Let ScalarEvolution sort and simplify the non-addrecs list. 326 const SCEV *Sum = NoAddRecs.empty() ? 327 SE.getConstant(Ty, 0) : 328 SE.getAddExpr(NoAddRecs); 329 // If it returned an add, use the operands. Otherwise it simplified 330 // the sum into a single value, so just use that. 331 Ops.clear(); 332 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 333 Ops.append(Add->op_begin(), Add->op_end()); 334 else if (!Sum->isZero()) 335 Ops.push_back(Sum); 336 // Then append the addrecs. 337 Ops.append(AddRecs.begin(), AddRecs.end()); 338} 339 340/// SplitAddRecs - Flatten a list of add operands, moving addrec start values 341/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 342/// This helps expose more opportunities for folding parts of the expressions 343/// into GEP indices. 344/// 345static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 346 Type *Ty, 347 ScalarEvolution &SE) { 348 // Find the addrecs. 349 SmallVector<const SCEV *, 8> AddRecs; 350 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 351 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 352 const SCEV *Start = A->getStart(); 353 if (Start->isZero()) break; 354 const SCEV *Zero = SE.getConstant(Ty, 0); 355 AddRecs.push_back(SE.getAddRecExpr(Zero, 356 A->getStepRecurrence(SE), 357 A->getLoop(), 358 A->getNoWrapFlags(SCEV::FlagNW))); 359 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 360 Ops[i] = Zero; 361 Ops.append(Add->op_begin(), Add->op_end()); 362 e += Add->getNumOperands(); 363 } else { 364 Ops[i] = Start; 365 } 366 } 367 if (!AddRecs.empty()) { 368 // Add the addrecs onto the end of the list. 369 Ops.append(AddRecs.begin(), AddRecs.end()); 370 // Resort the operand list, moving any constants to the front. 371 SimplifyAddOperands(Ops, Ty, SE); 372 } 373} 374 375/// expandAddToGEP - Expand an addition expression with a pointer type into 376/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 377/// BasicAliasAnalysis and other passes analyze the result. See the rules 378/// for getelementptr vs. inttoptr in 379/// http://llvm.org/docs/LangRef.html#pointeraliasing 380/// for details. 381/// 382/// Design note: The correctness of using getelementptr here depends on 383/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 384/// they may introduce pointer arithmetic which may not be safely converted 385/// into getelementptr. 386/// 387/// Design note: It might seem desirable for this function to be more 388/// loop-aware. If some of the indices are loop-invariant while others 389/// aren't, it might seem desirable to emit multiple GEPs, keeping the 390/// loop-invariant portions of the overall computation outside the loop. 391/// However, there are a few reasons this is not done here. Hoisting simple 392/// arithmetic is a low-level optimization that often isn't very 393/// important until late in the optimization process. In fact, passes 394/// like InstructionCombining will combine GEPs, even if it means 395/// pushing loop-invariant computation down into loops, so even if the 396/// GEPs were split here, the work would quickly be undone. The 397/// LoopStrengthReduction pass, which is usually run quite late (and 398/// after the last InstructionCombining pass), takes care of hoisting 399/// loop-invariant portions of expressions, after considering what 400/// can be folded using target addressing modes. 401/// 402Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 403 const SCEV *const *op_end, 404 PointerType *PTy, 405 Type *Ty, 406 Value *V) { 407 Type *OriginalElTy = PTy->getElementType(); 408 Type *ElTy = OriginalElTy; 409 SmallVector<Value *, 4> GepIndices; 410 SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 411 bool AnyNonZeroIndices = false; 412 413 // Split AddRecs up into parts as either of the parts may be usable 414 // without the other. 415 SplitAddRecs(Ops, Ty, SE); 416 417 Type *IntIdxTy = DL.getIndexType(PTy); 418 419 // Descend down the pointer's type and attempt to convert the other 420 // operands into GEP indices, at each level. The first index in a GEP 421 // indexes into the array implied by the pointer operand; the rest of 422 // the indices index into the element or field type selected by the 423 // preceding index. 424 for (;;) { 425 // If the scale size is not 0, attempt to factor out a scale for 426 // array indexing. 427 SmallVector<const SCEV *, 8> ScaledOps; 428 if (ElTy->isSized()) { 429 const SCEV *ElSize = SE.getSizeOfExpr(IntIdxTy, ElTy); 430 if (!ElSize->isZero()) { 431 SmallVector<const SCEV *, 8> NewOps; 432 for (const SCEV *Op : Ops) { 433 const SCEV *Remainder = SE.getConstant(Ty, 0); 434 if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) { 435 // Op now has ElSize factored out. 436 ScaledOps.push_back(Op); 437 if (!Remainder->isZero()) 438 NewOps.push_back(Remainder); 439 AnyNonZeroIndices = true; 440 } else { 441 // The operand was not divisible, so add it to the list of operands 442 // we'll scan next iteration. 443 NewOps.push_back(Op); 444 } 445 } 446 // If we made any changes, update Ops. 447 if (!ScaledOps.empty()) { 448 Ops = NewOps; 449 SimplifyAddOperands(Ops, Ty, SE); 450 } 451 } 452 } 453 454 // Record the scaled array index for this level of the type. If 455 // we didn't find any operands that could be factored, tentatively 456 // assume that element zero was selected (since the zero offset 457 // would obviously be folded away). 458 Value *Scaled = ScaledOps.empty() ? 459 Constant::getNullValue(Ty) : 460 expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 461 GepIndices.push_back(Scaled); 462 463 // Collect struct field index operands. 464 while (StructType *STy = dyn_cast<StructType>(ElTy)) { 465 bool FoundFieldNo = false; 466 // An empty struct has no fields. 467 if (STy->getNumElements() == 0) break; 468 // Field offsets are known. See if a constant offset falls within any of 469 // the struct fields. 470 if (Ops.empty()) 471 break; 472 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 473 if (SE.getTypeSizeInBits(C->getType()) <= 64) { 474 const StructLayout &SL = *DL.getStructLayout(STy); 475 uint64_t FullOffset = C->getValue()->getZExtValue(); 476 if (FullOffset < SL.getSizeInBytes()) { 477 unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 478 GepIndices.push_back( 479 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 480 ElTy = STy->getTypeAtIndex(ElIdx); 481 Ops[0] = 482 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 483 AnyNonZeroIndices = true; 484 FoundFieldNo = true; 485 } 486 } 487 // If no struct field offsets were found, tentatively assume that 488 // field zero was selected (since the zero offset would obviously 489 // be folded away). 490 if (!FoundFieldNo) { 491 ElTy = STy->getTypeAtIndex(0u); 492 GepIndices.push_back( 493 Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 494 } 495 } 496 497 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 498 ElTy = ATy->getElementType(); 499 else 500 break; 501 } 502 503 // If none of the operands were convertible to proper GEP indices, cast 504 // the base to i8* and do an ugly getelementptr with that. It's still 505 // better than ptrtoint+arithmetic+inttoptr at least. 506 if (!AnyNonZeroIndices) { 507 // Cast the base to i8*. 508 V = InsertNoopCastOfTo(V, 509 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 510 511 assert(!isa<Instruction>(V) || 512 SE.DT.dominates(cast<Instruction>(V), &*Builder.GetInsertPoint())); 513 514 // Expand the operands for a plain byte offset. 515 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 516 517 // Fold a GEP with constant operands. 518 if (Constant *CLHS = dyn_cast<Constant>(V)) 519 if (Constant *CRHS = dyn_cast<Constant>(Idx)) 520 return ConstantExpr::getGetElementPtr(Type::getInt8Ty(Ty->getContext()), 521 CLHS, CRHS); 522 523 // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 524 unsigned ScanLimit = 6; 525 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 526 // Scanning starts from the last instruction before the insertion point. 527 BasicBlock::iterator IP = Builder.GetInsertPoint(); 528 if (IP != BlockBegin) { 529 --IP; 530 for (; ScanLimit; --IP, --ScanLimit) { 531 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 532 // generated code. 533 if (isa<DbgInfoIntrinsic>(IP)) 534 ScanLimit++; 535 if (IP->getOpcode() == Instruction::GetElementPtr && 536 IP->getOperand(0) == V && IP->getOperand(1) == Idx) 537 return &*IP; 538 if (IP == BlockBegin) break; 539 } 540 } 541 542 // Save the original insertion point so we can restore it when we're done. 543 SCEVInsertPointGuard Guard(Builder, this); 544 545 // Move the insertion point out of as many loops as we can. 546 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 547 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 548 BasicBlock *Preheader = L->getLoopPreheader(); 549 if (!Preheader) break; 550 551 // Ok, move up a level. 552 Builder.SetInsertPoint(Preheader->getTerminator()); 553 } 554 555 // Emit a GEP. 556 Value *GEP = Builder.CreateGEP(Builder.getInt8Ty(), V, Idx, "uglygep"); 557 rememberInstruction(GEP); 558 559 return GEP; 560 } 561 562 { 563 SCEVInsertPointGuard Guard(Builder, this); 564 565 // Move the insertion point out of as many loops as we can. 566 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 567 if (!L->isLoopInvariant(V)) break; 568 569 bool AnyIndexNotLoopInvariant = any_of( 570 GepIndices, [L](Value *Op) { return !L->isLoopInvariant(Op); }); 571 572 if (AnyIndexNotLoopInvariant) 573 break; 574 575 BasicBlock *Preheader = L->getLoopPreheader(); 576 if (!Preheader) break; 577 578 // Ok, move up a level. 579 Builder.SetInsertPoint(Preheader->getTerminator()); 580 } 581 582 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 583 // because ScalarEvolution may have changed the address arithmetic to 584 // compute a value which is beyond the end of the allocated object. 585 Value *Casted = V; 586 if (V->getType() != PTy) 587 Casted = InsertNoopCastOfTo(Casted, PTy); 588 Value *GEP = Builder.CreateGEP(OriginalElTy, Casted, GepIndices, "scevgep"); 589 Ops.push_back(SE.getUnknown(GEP)); 590 rememberInstruction(GEP); 591 } 592 593 return expand(SE.getAddExpr(Ops)); 594} 595 596Value *SCEVExpander::expandAddToGEP(const SCEV *Op, PointerType *PTy, Type *Ty, 597 Value *V) { 598 const SCEV *const Ops[1] = {Op}; 599 return expandAddToGEP(Ops, Ops + 1, PTy, Ty, V); 600} 601 602/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 603/// SCEV expansion. If they are nested, this is the most nested. If they are 604/// neighboring, pick the later. 605static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 606 DominatorTree &DT) { 607 if (!A) return B; 608 if (!B) return A; 609 if (A->contains(B)) return B; 610 if (B->contains(A)) return A; 611 if (DT.dominates(A->getHeader(), B->getHeader())) return B; 612 if (DT.dominates(B->getHeader(), A->getHeader())) return A; 613 return A; // Arbitrarily break the tie. 614} 615 616/// getRelevantLoop - Get the most relevant loop associated with the given 617/// expression, according to PickMostRelevantLoop. 618const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 619 // Test whether we've already computed the most relevant loop for this SCEV. 620 auto Pair = RelevantLoops.insert(std::make_pair(S, nullptr)); 621 if (!Pair.second) 622 return Pair.first->second; 623 624 if (isa<SCEVConstant>(S)) 625 // A constant has no relevant loops. 626 return nullptr; 627 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 628 if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 629 return Pair.first->second = SE.LI.getLoopFor(I->getParent()); 630 // A non-instruction has no relevant loops. 631 return nullptr; 632 } 633 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 634 const Loop *L = nullptr; 635 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 636 L = AR->getLoop(); 637 for (const SCEV *Op : N->operands()) 638 L = PickMostRelevantLoop(L, getRelevantLoop(Op), SE.DT); 639 return RelevantLoops[N] = L; 640 } 641 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 642 const Loop *Result = getRelevantLoop(C->getOperand()); 643 return RelevantLoops[C] = Result; 644 } 645 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 646 const Loop *Result = PickMostRelevantLoop( 647 getRelevantLoop(D->getLHS()), getRelevantLoop(D->getRHS()), SE.DT); 648 return RelevantLoops[D] = Result; 649 } 650 llvm_unreachable("Unexpected SCEV type!"); 651} 652 653namespace { 654 655/// LoopCompare - Compare loops by PickMostRelevantLoop. 656class LoopCompare { 657 DominatorTree &DT; 658public: 659 explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 660 661 bool operator()(std::pair<const Loop *, const SCEV *> LHS, 662 std::pair<const Loop *, const SCEV *> RHS) const { 663 // Keep pointer operands sorted at the end. 664 if (LHS.second->getType()->isPointerTy() != 665 RHS.second->getType()->isPointerTy()) 666 return LHS.second->getType()->isPointerTy(); 667 668 // Compare loops with PickMostRelevantLoop. 669 if (LHS.first != RHS.first) 670 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 671 672 // If one operand is a non-constant negative and the other is not, 673 // put the non-constant negative on the right so that a sub can 674 // be used instead of a negate and add. 675 if (LHS.second->isNonConstantNegative()) { 676 if (!RHS.second->isNonConstantNegative()) 677 return false; 678 } else if (RHS.second->isNonConstantNegative()) 679 return true; 680 681 // Otherwise they are equivalent according to this comparison. 682 return false; 683 } 684}; 685 686} 687 688Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 689 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 690 691 // Collect all the add operands in a loop, along with their associated loops. 692 // Iterate in reverse so that constants are emitted last, all else equal, and 693 // so that pointer operands are inserted first, which the code below relies on 694 // to form more involved GEPs. 695 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 696 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 697 E(S->op_begin()); I != E; ++I) 698 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 699 700 // Sort by loop. Use a stable sort so that constants follow non-constants and 701 // pointer operands precede non-pointer operands. 702 llvm::stable_sort(OpsAndLoops, LoopCompare(SE.DT)); 703 704 // Emit instructions to add all the operands. Hoist as much as possible 705 // out of loops, and form meaningful getelementptrs where possible. 706 Value *Sum = nullptr; 707 for (auto I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E;) { 708 const Loop *CurLoop = I->first; 709 const SCEV *Op = I->second; 710 if (!Sum) { 711 // This is the first operand. Just expand it. 712 Sum = expand(Op); 713 ++I; 714 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 715 // The running sum expression is a pointer. Try to form a getelementptr 716 // at this level with that as the base. 717 SmallVector<const SCEV *, 4> NewOps; 718 for (; I != E && I->first == CurLoop; ++I) { 719 // If the operand is SCEVUnknown and not instructions, peek through 720 // it, to enable more of it to be folded into the GEP. 721 const SCEV *X = I->second; 722 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 723 if (!isa<Instruction>(U->getValue())) 724 X = SE.getSCEV(U->getValue()); 725 NewOps.push_back(X); 726 } 727 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 728 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 729 // The running sum is an integer, and there's a pointer at this level. 730 // Try to form a getelementptr. If the running sum is instructions, 731 // use a SCEVUnknown to avoid re-analyzing them. 732 SmallVector<const SCEV *, 4> NewOps; 733 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 734 SE.getSCEV(Sum)); 735 for (++I; I != E && I->first == CurLoop; ++I) 736 NewOps.push_back(I->second); 737 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 738 } else if (Op->isNonConstantNegative()) { 739 // Instead of doing a negate and add, just do a subtract. 740 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 741 Sum = InsertNoopCastOfTo(Sum, Ty); 742 Sum = InsertBinop(Instruction::Sub, Sum, W, SCEV::FlagAnyWrap, 743 /*IsSafeToHoist*/ true); 744 ++I; 745 } else { 746 // A simple add. 747 Value *W = expandCodeFor(Op, Ty); 748 Sum = InsertNoopCastOfTo(Sum, Ty); 749 // Canonicalize a constant to the RHS. 750 if (isa<Constant>(Sum)) std::swap(Sum, W); 751 Sum = InsertBinop(Instruction::Add, Sum, W, S->getNoWrapFlags(), 752 /*IsSafeToHoist*/ true); 753 ++I; 754 } 755 } 756 757 return Sum; 758} 759 760Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 761 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 762 763 // Collect all the mul operands in a loop, along with their associated loops. 764 // Iterate in reverse so that constants are emitted last, all else equal. 765 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 766 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 767 E(S->op_begin()); I != E; ++I) 768 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 769 770 // Sort by loop. Use a stable sort so that constants follow non-constants. 771 llvm::stable_sort(OpsAndLoops, LoopCompare(SE.DT)); 772 773 // Emit instructions to mul all the operands. Hoist as much as possible 774 // out of loops. 775 Value *Prod = nullptr; 776 auto I = OpsAndLoops.begin(); 777 778 // Expand the calculation of X pow N in the following manner: 779 // Let N = P1 + P2 + ... + PK, where all P are powers of 2. Then: 780 // X pow N = (X pow P1) * (X pow P2) * ... * (X pow PK). 781 const auto ExpandOpBinPowN = [this, &I, &OpsAndLoops, &Ty]() { 782 auto E = I; 783 // Calculate how many times the same operand from the same loop is included 784 // into this power. 785 uint64_t Exponent = 0; 786 const uint64_t MaxExponent = UINT64_MAX >> 1; 787 // No one sane will ever try to calculate such huge exponents, but if we 788 // need this, we stop on UINT64_MAX / 2 because we need to exit the loop 789 // below when the power of 2 exceeds our Exponent, and we want it to be 790 // 1u << 31 at most to not deal with unsigned overflow. 791 while (E != OpsAndLoops.end() && *I == *E && Exponent != MaxExponent) { 792 ++Exponent; 793 ++E; 794 } 795 assert(Exponent > 0 && "Trying to calculate a zeroth exponent of operand?"); 796 797 // Calculate powers with exponents 1, 2, 4, 8 etc. and include those of them 798 // that are needed into the result. 799 Value *P = expandCodeFor(I->second, Ty); 800 Value *Result = nullptr; 801 if (Exponent & 1) 802 Result = P; 803 for (uint64_t BinExp = 2; BinExp <= Exponent; BinExp <<= 1) { 804 P = InsertBinop(Instruction::Mul, P, P, SCEV::FlagAnyWrap, 805 /*IsSafeToHoist*/ true); 806 if (Exponent & BinExp) 807 Result = Result ? InsertBinop(Instruction::Mul, Result, P, 808 SCEV::FlagAnyWrap, 809 /*IsSafeToHoist*/ true) 810 : P; 811 } 812 813 I = E; 814 assert(Result && "Nothing was expanded?"); 815 return Result; 816 }; 817 818 while (I != OpsAndLoops.end()) { 819 if (!Prod) { 820 // This is the first operand. Just expand it. 821 Prod = ExpandOpBinPowN(); 822 } else if (I->second->isAllOnesValue()) { 823 // Instead of doing a multiply by negative one, just do a negate. 824 Prod = InsertNoopCastOfTo(Prod, Ty); 825 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod, 826 SCEV::FlagAnyWrap, /*IsSafeToHoist*/ true); 827 ++I; 828 } else { 829 // A simple mul. 830 Value *W = ExpandOpBinPowN(); 831 Prod = InsertNoopCastOfTo(Prod, Ty); 832 // Canonicalize a constant to the RHS. 833 if (isa<Constant>(Prod)) std::swap(Prod, W); 834 const APInt *RHS; 835 if (match(W, m_Power2(RHS))) { 836 // Canonicalize Prod*(1<<C) to Prod<<C. 837 assert(!Ty->isVectorTy() && "vector types are not SCEVable"); 838 auto NWFlags = S->getNoWrapFlags(); 839 // clear nsw flag if shl will produce poison value. 840 if (RHS->logBase2() == RHS->getBitWidth() - 1) 841 NWFlags = ScalarEvolution::clearFlags(NWFlags, SCEV::FlagNSW); 842 Prod = InsertBinop(Instruction::Shl, Prod, 843 ConstantInt::get(Ty, RHS->logBase2()), NWFlags, 844 /*IsSafeToHoist*/ true); 845 } else { 846 Prod = InsertBinop(Instruction::Mul, Prod, W, S->getNoWrapFlags(), 847 /*IsSafeToHoist*/ true); 848 } 849 } 850 } 851 852 return Prod; 853} 854 855Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 856 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 857 858 Value *LHS = expandCodeFor(S->getLHS(), Ty); 859 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 860 const APInt &RHS = SC->getAPInt(); 861 if (RHS.isPowerOf2()) 862 return InsertBinop(Instruction::LShr, LHS, 863 ConstantInt::get(Ty, RHS.logBase2()), 864 SCEV::FlagAnyWrap, /*IsSafeToHoist*/ true); 865 } 866 867 Value *RHS = expandCodeFor(S->getRHS(), Ty); 868 return InsertBinop(Instruction::UDiv, LHS, RHS, SCEV::FlagAnyWrap, 869 /*IsSafeToHoist*/ SE.isKnownNonZero(S->getRHS())); 870} 871 872/// Move parts of Base into Rest to leave Base with the minimal 873/// expression that provides a pointer operand suitable for a 874/// GEP expansion. 875static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 876 ScalarEvolution &SE) { 877 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 878 Base = A->getStart(); 879 Rest = SE.getAddExpr(Rest, 880 SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 881 A->getStepRecurrence(SE), 882 A->getLoop(), 883 A->getNoWrapFlags(SCEV::FlagNW))); 884 } 885 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 886 Base = A->getOperand(A->getNumOperands()-1); 887 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 888 NewAddOps.back() = Rest; 889 Rest = SE.getAddExpr(NewAddOps); 890 ExposePointerBase(Base, Rest, SE); 891 } 892} 893 894/// Determine if this is a well-behaved chain of instructions leading back to 895/// the PHI. If so, it may be reused by expanded expressions. 896bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 897 const Loop *L) { 898 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 899 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 900 return false; 901 // If any of the operands don't dominate the insert position, bail. 902 // Addrec operands are always loop-invariant, so this can only happen 903 // if there are instructions which haven't been hoisted. 904 if (L == IVIncInsertLoop) { 905 for (User::op_iterator OI = IncV->op_begin()+1, 906 OE = IncV->op_end(); OI != OE; ++OI) 907 if (Instruction *OInst = dyn_cast<Instruction>(OI)) 908 if (!SE.DT.dominates(OInst, IVIncInsertPos)) 909 return false; 910 } 911 // Advance to the next instruction. 912 IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 913 if (!IncV) 914 return false; 915 916 if (IncV->mayHaveSideEffects()) 917 return false; 918 919 if (IncV == PN) 920 return true; 921 922 return isNormalAddRecExprPHI(PN, IncV, L); 923} 924 925/// getIVIncOperand returns an induction variable increment's induction 926/// variable operand. 927/// 928/// If allowScale is set, any type of GEP is allowed as long as the nonIV 929/// operands dominate InsertPos. 930/// 931/// If allowScale is not set, ensure that a GEP increment conforms to one of the 932/// simple patterns generated by getAddRecExprPHILiterally and 933/// expandAddtoGEP. If the pattern isn't recognized, return NULL. 934Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 935 Instruction *InsertPos, 936 bool allowScale) { 937 if (IncV == InsertPos) 938 return nullptr; 939 940 switch (IncV->getOpcode()) { 941 default: 942 return nullptr; 943 // Check for a simple Add/Sub or GEP of a loop invariant step. 944 case Instruction::Add: 945 case Instruction::Sub: { 946 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 947 if (!OInst || SE.DT.dominates(OInst, InsertPos)) 948 return dyn_cast<Instruction>(IncV->getOperand(0)); 949 return nullptr; 950 } 951 case Instruction::BitCast: 952 return dyn_cast<Instruction>(IncV->getOperand(0)); 953 case Instruction::GetElementPtr: 954 for (auto I = IncV->op_begin() + 1, E = IncV->op_end(); I != E; ++I) { 955 if (isa<Constant>(*I)) 956 continue; 957 if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 958 if (!SE.DT.dominates(OInst, InsertPos)) 959 return nullptr; 960 } 961 if (allowScale) { 962 // allow any kind of GEP as long as it can be hoisted. 963 continue; 964 } 965 // This must be a pointer addition of constants (pretty), which is already 966 // handled, or some number of address-size elements (ugly). Ugly geps 967 // have 2 operands. i1* is used by the expander to represent an 968 // address-size element. 969 if (IncV->getNumOperands() != 2) 970 return nullptr; 971 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 972 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 973 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 974 return nullptr; 975 break; 976 } 977 return dyn_cast<Instruction>(IncV->getOperand(0)); 978 } 979} 980 981/// If the insert point of the current builder or any of the builders on the 982/// stack of saved builders has 'I' as its insert point, update it to point to 983/// the instruction after 'I'. This is intended to be used when the instruction 984/// 'I' is being moved. If this fixup is not done and 'I' is moved to a 985/// different block, the inconsistent insert point (with a mismatched 986/// Instruction and Block) can lead to an instruction being inserted in a block 987/// other than its parent. 988void SCEVExpander::fixupInsertPoints(Instruction *I) { 989 BasicBlock::iterator It(*I); 990 BasicBlock::iterator NewInsertPt = std::next(It); 991 if (Builder.GetInsertPoint() == It) 992 Builder.SetInsertPoint(&*NewInsertPt); 993 for (auto *InsertPtGuard : InsertPointGuards) 994 if (InsertPtGuard->GetInsertPoint() == It) 995 InsertPtGuard->SetInsertPoint(NewInsertPt); 996} 997 998/// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 999/// it available to other uses in this loop. Recursively hoist any operands, 1000/// until we reach a value that dominates InsertPos. 1001bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 1002 if (SE.DT.dominates(IncV, InsertPos)) 1003 return true; 1004 1005 // InsertPos must itself dominate IncV so that IncV's new position satisfies 1006 // its existing users. 1007 if (isa<PHINode>(InsertPos) || 1008 !SE.DT.dominates(InsertPos->getParent(), IncV->getParent())) 1009 return false; 1010 1011 if (!SE.LI.movementPreservesLCSSAForm(IncV, InsertPos)) 1012 return false; 1013 1014 // Check that the chain of IV operands leading back to Phi can be hoisted. 1015 SmallVector<Instruction*, 4> IVIncs; 1016 for(;;) { 1017 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 1018 if (!Oper) 1019 return false; 1020 // IncV is safe to hoist. 1021 IVIncs.push_back(IncV); 1022 IncV = Oper; 1023 if (SE.DT.dominates(IncV, InsertPos)) 1024 break; 1025 } 1026 for (auto I = IVIncs.rbegin(), E = IVIncs.rend(); I != E; ++I) { 1027 fixupInsertPoints(*I); 1028 (*I)->moveBefore(InsertPos); 1029 } 1030 return true; 1031} 1032 1033/// Determine if this cyclic phi is in a form that would have been generated by 1034/// LSR. We don't care if the phi was actually expanded in this pass, as long 1035/// as it is in a low-cost form, for example, no implied multiplication. This 1036/// should match any patterns generated by getAddRecExprPHILiterally and 1037/// expandAddtoGEP. 1038bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 1039 const Loop *L) { 1040 for(Instruction *IVOper = IncV; 1041 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 1042 /*allowScale=*/false));) { 1043 if (IVOper == PN) 1044 return true; 1045 } 1046 return false; 1047} 1048 1049/// expandIVInc - Expand an IV increment at Builder's current InsertPos. 1050/// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 1051/// need to materialize IV increments elsewhere to handle difficult situations. 1052Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 1053 Type *ExpandTy, Type *IntTy, 1054 bool useSubtract) { 1055 Value *IncV; 1056 // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 1057 if (ExpandTy->isPointerTy()) { 1058 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 1059 // If the step isn't constant, don't use an implicitly scaled GEP, because 1060 // that would require a multiply inside the loop. 1061 if (!isa<ConstantInt>(StepV)) 1062 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 1063 GEPPtrTy->getAddressSpace()); 1064 IncV = expandAddToGEP(SE.getSCEV(StepV), GEPPtrTy, IntTy, PN); 1065 if (IncV->getType() != PN->getType()) { 1066 IncV = Builder.CreateBitCast(IncV, PN->getType()); 1067 rememberInstruction(IncV); 1068 } 1069 } else { 1070 IncV = useSubtract ? 1071 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 1072 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 1073 rememberInstruction(IncV); 1074 } 1075 return IncV; 1076} 1077 1078/// Hoist the addrec instruction chain rooted in the loop phi above the 1079/// position. This routine assumes that this is possible (has been checked). 1080void SCEVExpander::hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist, 1081 Instruction *Pos, PHINode *LoopPhi) { 1082 do { 1083 if (DT->dominates(InstToHoist, Pos)) 1084 break; 1085 // Make sure the increment is where we want it. But don't move it 1086 // down past a potential existing post-inc user. 1087 fixupInsertPoints(InstToHoist); 1088 InstToHoist->moveBefore(Pos); 1089 Pos = InstToHoist; 1090 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0)); 1091 } while (InstToHoist != LoopPhi); 1092} 1093 1094/// Check whether we can cheaply express the requested SCEV in terms of 1095/// the available PHI SCEV by truncation and/or inversion of the step. 1096static bool canBeCheaplyTransformed(ScalarEvolution &SE, 1097 const SCEVAddRecExpr *Phi, 1098 const SCEVAddRecExpr *Requested, 1099 bool &InvertStep) { 1100 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType()); 1101 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType()); 1102 1103 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth()) 1104 return false; 1105 1106 // Try truncate it if necessary. 1107 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy)); 1108 if (!Phi) 1109 return false; 1110 1111 // Check whether truncation will help. 1112 if (Phi == Requested) { 1113 InvertStep = false; 1114 return true; 1115 } 1116 1117 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}. 1118 if (SE.getAddExpr(Requested->getStart(), 1119 SE.getNegativeSCEV(Requested)) == Phi) { 1120 InvertStep = true; 1121 return true; 1122 } 1123 1124 return false; 1125} 1126 1127static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1128 if (!isa<IntegerType>(AR->getType())) 1129 return false; 1130 1131 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1132 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1133 const SCEV *Step = AR->getStepRecurrence(SE); 1134 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy), 1135 SE.getSignExtendExpr(AR, WideTy)); 1136 const SCEV *ExtendAfterOp = 1137 SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1138 return ExtendAfterOp == OpAfterExtend; 1139} 1140 1141static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1142 if (!isa<IntegerType>(AR->getType())) 1143 return false; 1144 1145 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1146 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1147 const SCEV *Step = AR->getStepRecurrence(SE); 1148 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy), 1149 SE.getZeroExtendExpr(AR, WideTy)); 1150 const SCEV *ExtendAfterOp = 1151 SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1152 return ExtendAfterOp == OpAfterExtend; 1153} 1154 1155/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 1156/// the base addrec, which is the addrec without any non-loop-dominating 1157/// values, and return the PHI. 1158PHINode * 1159SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 1160 const Loop *L, 1161 Type *ExpandTy, 1162 Type *IntTy, 1163 Type *&TruncTy, 1164 bool &InvertStep) { 1165 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 1166 1167 // Reuse a previously-inserted PHI, if present. 1168 BasicBlock *LatchBlock = L->getLoopLatch(); 1169 if (LatchBlock) { 1170 PHINode *AddRecPhiMatch = nullptr; 1171 Instruction *IncV = nullptr; 1172 TruncTy = nullptr; 1173 InvertStep = false; 1174 1175 // Only try partially matching scevs that need truncation and/or 1176 // step-inversion if we know this loop is outside the current loop. 1177 bool TryNonMatchingSCEV = 1178 IVIncInsertLoop && 1179 SE.DT.properlyDominates(LatchBlock, IVIncInsertLoop->getHeader()); 1180 1181 for (PHINode &PN : L->getHeader()->phis()) { 1182 if (!SE.isSCEVable(PN.getType())) 1183 continue; 1184 1185 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(&PN)); 1186 if (!PhiSCEV) 1187 continue; 1188 1189 bool IsMatchingSCEV = PhiSCEV == Normalized; 1190 // We only handle truncation and inversion of phi recurrences for the 1191 // expanded expression if the expanded expression's loop dominates the 1192 // loop we insert to. Check now, so we can bail out early. 1193 if (!IsMatchingSCEV && !TryNonMatchingSCEV) 1194 continue; 1195 1196 // TODO: this possibly can be reworked to avoid this cast at all. 1197 Instruction *TempIncV = 1198 dyn_cast<Instruction>(PN.getIncomingValueForBlock(LatchBlock)); 1199 if (!TempIncV) 1200 continue; 1201 1202 // Check whether we can reuse this PHI node. 1203 if (LSRMode) { 1204 if (!isExpandedAddRecExprPHI(&PN, TempIncV, L)) 1205 continue; 1206 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos)) 1207 continue; 1208 } else { 1209 if (!isNormalAddRecExprPHI(&PN, TempIncV, L)) 1210 continue; 1211 } 1212 1213 // Stop if we have found an exact match SCEV. 1214 if (IsMatchingSCEV) { 1215 IncV = TempIncV; 1216 TruncTy = nullptr; 1217 InvertStep = false; 1218 AddRecPhiMatch = &PN; 1219 break; 1220 } 1221 1222 // Try whether the phi can be translated into the requested form 1223 // (truncated and/or offset by a constant). 1224 if ((!TruncTy || InvertStep) && 1225 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) { 1226 // Record the phi node. But don't stop we might find an exact match 1227 // later. 1228 AddRecPhiMatch = &PN; 1229 IncV = TempIncV; 1230 TruncTy = SE.getEffectiveSCEVType(Normalized->getType()); 1231 } 1232 } 1233 1234 if (AddRecPhiMatch) { 1235 // Potentially, move the increment. We have made sure in 1236 // isExpandedAddRecExprPHI or hoistIVInc that this is possible. 1237 if (L == IVIncInsertLoop) 1238 hoistBeforePos(&SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch); 1239 1240 // Ok, the add recurrence looks usable. 1241 // Remember this PHI, even in post-inc mode. 1242 InsertedValues.insert(AddRecPhiMatch); 1243 // Remember the increment. 1244 rememberInstruction(IncV); 1245 return AddRecPhiMatch; 1246 } 1247 } 1248 1249 // Save the original insertion point so we can restore it when we're done. 1250 SCEVInsertPointGuard Guard(Builder, this); 1251 1252 // Another AddRec may need to be recursively expanded below. For example, if 1253 // this AddRec is quadratic, the StepV may itself be an AddRec in this 1254 // loop. Remove this loop from the PostIncLoops set before expanding such 1255 // AddRecs. Otherwise, we cannot find a valid position for the step 1256 // (i.e. StepV can never dominate its loop header). Ideally, we could do 1257 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 1258 // so it's not worth implementing SmallPtrSet::swap. 1259 PostIncLoopSet SavedPostIncLoops = PostIncLoops; 1260 PostIncLoops.clear(); 1261 1262 // Expand code for the start value into the loop preheader. 1263 assert(L->getLoopPreheader() && 1264 "Can't expand add recurrences without a loop preheader!"); 1265 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 1266 L->getLoopPreheader()->getTerminator()); 1267 1268 // StartV must have been be inserted into L's preheader to dominate the new 1269 // phi. 1270 assert(!isa<Instruction>(StartV) || 1271 SE.DT.properlyDominates(cast<Instruction>(StartV)->getParent(), 1272 L->getHeader())); 1273 1274 // Expand code for the step value. Do this before creating the PHI so that PHI 1275 // reuse code doesn't see an incomplete PHI. 1276 const SCEV *Step = Normalized->getStepRecurrence(SE); 1277 // If the stride is negative, insert a sub instead of an add for the increment 1278 // (unless it's a constant, because subtracts of constants are canonicalized 1279 // to adds). 1280 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1281 if (useSubtract) 1282 Step = SE.getNegativeSCEV(Step); 1283 // Expand the step somewhere that dominates the loop header. 1284 Value *StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front()); 1285 1286 // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if 1287 // we actually do emit an addition. It does not apply if we emit a 1288 // subtraction. 1289 bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized); 1290 bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized); 1291 1292 // Create the PHI. 1293 BasicBlock *Header = L->getHeader(); 1294 Builder.SetInsertPoint(Header, Header->begin()); 1295 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1296 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 1297 Twine(IVName) + ".iv"); 1298 rememberInstruction(PN); 1299 1300 // Create the step instructions and populate the PHI. 1301 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1302 BasicBlock *Pred = *HPI; 1303 1304 // Add a start value. 1305 if (!L->contains(Pred)) { 1306 PN->addIncoming(StartV, Pred); 1307 continue; 1308 } 1309 1310 // Create a step value and add it to the PHI. 1311 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 1312 // instructions at IVIncInsertPos. 1313 Instruction *InsertPos = L == IVIncInsertLoop ? 1314 IVIncInsertPos : Pred->getTerminator(); 1315 Builder.SetInsertPoint(InsertPos); 1316 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1317 1318 if (isa<OverflowingBinaryOperator>(IncV)) { 1319 if (IncrementIsNUW) 1320 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap(); 1321 if (IncrementIsNSW) 1322 cast<BinaryOperator>(IncV)->setHasNoSignedWrap(); 1323 } 1324 PN->addIncoming(IncV, Pred); 1325 } 1326 1327 // After expanding subexpressions, restore the PostIncLoops set so the caller 1328 // can ensure that IVIncrement dominates the current uses. 1329 PostIncLoops = SavedPostIncLoops; 1330 1331 // Remember this PHI, even in post-inc mode. 1332 InsertedValues.insert(PN); 1333 1334 return PN; 1335} 1336 1337Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 1338 Type *STy = S->getType(); 1339 Type *IntTy = SE.getEffectiveSCEVType(STy); 1340 const Loop *L = S->getLoop(); 1341 1342 // Determine a normalized form of this expression, which is the expression 1343 // before any post-inc adjustment is made. 1344 const SCEVAddRecExpr *Normalized = S; 1345 if (PostIncLoops.count(L)) { 1346 PostIncLoopSet Loops; 1347 Loops.insert(L); 1348 Normalized = cast<SCEVAddRecExpr>(normalizeForPostIncUse(S, Loops, SE)); 1349 } 1350 1351 // Strip off any non-loop-dominating component from the addrec start. 1352 const SCEV *Start = Normalized->getStart(); 1353 const SCEV *PostLoopOffset = nullptr; 1354 if (!SE.properlyDominates(Start, L->getHeader())) { 1355 PostLoopOffset = Start; 1356 Start = SE.getConstant(Normalized->getType(), 0); 1357 Normalized = cast<SCEVAddRecExpr>( 1358 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 1359 Normalized->getLoop(), 1360 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1361 } 1362 1363 // Strip off any non-loop-dominating component from the addrec step. 1364 const SCEV *Step = Normalized->getStepRecurrence(SE); 1365 const SCEV *PostLoopScale = nullptr; 1366 if (!SE.dominates(Step, L->getHeader())) { 1367 PostLoopScale = Step; 1368 Step = SE.getConstant(Normalized->getType(), 1); 1369 if (!Start->isZero()) { 1370 // The normalization below assumes that Start is constant zero, so if 1371 // it isn't re-associate Start to PostLoopOffset. 1372 assert(!PostLoopOffset && "Start not-null but PostLoopOffset set?"); 1373 PostLoopOffset = Start; 1374 Start = SE.getConstant(Normalized->getType(), 0); 1375 } 1376 Normalized = 1377 cast<SCEVAddRecExpr>(SE.getAddRecExpr( 1378 Start, Step, Normalized->getLoop(), 1379 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1380 } 1381 1382 // Expand the core addrec. If we need post-loop scaling, force it to 1383 // expand to an integer type to avoid the need for additional casting. 1384 Type *ExpandTy = PostLoopScale ? IntTy : STy; 1385 // We can't use a pointer type for the addrec if the pointer type is 1386 // non-integral. 1387 Type *AddRecPHIExpandTy = 1388 DL.isNonIntegralPointerType(STy) ? Normalized->getType() : ExpandTy; 1389 1390 // In some cases, we decide to reuse an existing phi node but need to truncate 1391 // it and/or invert the step. 1392 Type *TruncTy = nullptr; 1393 bool InvertStep = false; 1394 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, AddRecPHIExpandTy, 1395 IntTy, TruncTy, InvertStep); 1396 1397 // Accommodate post-inc mode, if necessary. 1398 Value *Result; 1399 if (!PostIncLoops.count(L)) 1400 Result = PN; 1401 else { 1402 // In PostInc mode, use the post-incremented value. 1403 BasicBlock *LatchBlock = L->getLoopLatch(); 1404 assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1405 Result = PN->getIncomingValueForBlock(LatchBlock); 1406 1407 // For an expansion to use the postinc form, the client must call 1408 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 1409 // or dominated by IVIncInsertPos. 1410 if (isa<Instruction>(Result) && 1411 !SE.DT.dominates(cast<Instruction>(Result), 1412 &*Builder.GetInsertPoint())) { 1413 // The induction variable's postinc expansion does not dominate this use. 1414 // IVUsers tries to prevent this case, so it is rare. However, it can 1415 // happen when an IVUser outside the loop is not dominated by the latch 1416 // block. Adjusting IVIncInsertPos before expansion begins cannot handle 1417 // all cases. Consider a phi outside whose operand is replaced during 1418 // expansion with the value of the postinc user. Without fundamentally 1419 // changing the way postinc users are tracked, the only remedy is 1420 // inserting an extra IV increment. StepV might fold into PostLoopOffset, 1421 // but hopefully expandCodeFor handles that. 1422 bool useSubtract = 1423 !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1424 if (useSubtract) 1425 Step = SE.getNegativeSCEV(Step); 1426 Value *StepV; 1427 { 1428 // Expand the step somewhere that dominates the loop header. 1429 SCEVInsertPointGuard Guard(Builder, this); 1430 StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front()); 1431 } 1432 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1433 } 1434 } 1435 1436 // We have decided to reuse an induction variable of a dominating loop. Apply 1437 // truncation and/or inversion of the step. 1438 if (TruncTy) { 1439 Type *ResTy = Result->getType(); 1440 // Normalize the result type. 1441 if (ResTy != SE.getEffectiveSCEVType(ResTy)) 1442 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy)); 1443 // Truncate the result. 1444 if (TruncTy != Result->getType()) { 1445 Result = Builder.CreateTrunc(Result, TruncTy); 1446 rememberInstruction(Result); 1447 } 1448 // Invert the result. 1449 if (InvertStep) { 1450 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy), 1451 Result); 1452 rememberInstruction(Result); 1453 } 1454 } 1455 1456 // Re-apply any non-loop-dominating scale. 1457 if (PostLoopScale) { 1458 assert(S->isAffine() && "Can't linearly scale non-affine recurrences."); 1459 Result = InsertNoopCastOfTo(Result, IntTy); 1460 Result = Builder.CreateMul(Result, 1461 expandCodeFor(PostLoopScale, IntTy)); 1462 rememberInstruction(Result); 1463 } 1464 1465 // Re-apply any non-loop-dominating offset. 1466 if (PostLoopOffset) { 1467 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1468 if (Result->getType()->isIntegerTy()) { 1469 Value *Base = expandCodeFor(PostLoopOffset, ExpandTy); 1470 Result = expandAddToGEP(SE.getUnknown(Result), PTy, IntTy, Base); 1471 } else { 1472 Result = expandAddToGEP(PostLoopOffset, PTy, IntTy, Result); 1473 } 1474 } else { 1475 Result = InsertNoopCastOfTo(Result, IntTy); 1476 Result = Builder.CreateAdd(Result, 1477 expandCodeFor(PostLoopOffset, IntTy)); 1478 rememberInstruction(Result); 1479 } 1480 } 1481 1482 return Result; 1483} 1484 1485Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1486 // In canonical mode we compute the addrec as an expression of a canonical IV 1487 // using evaluateAtIteration and expand the resulting SCEV expression. This 1488 // way we avoid introducing new IVs to carry on the comutation of the addrec 1489 // throughout the loop. 1490 // 1491 // For nested addrecs evaluateAtIteration might need a canonical IV of a 1492 // type wider than the addrec itself. Emitting a canonical IV of the 1493 // proper type might produce non-legal types, for example expanding an i64 1494 // {0,+,2,+,1} addrec would need an i65 canonical IV. To avoid this just fall 1495 // back to non-canonical mode for nested addrecs. 1496 if (!CanonicalMode || (S->getNumOperands() > 2)) 1497 return expandAddRecExprLiterally(S); 1498 1499 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1500 const Loop *L = S->getLoop(); 1501 1502 // First check for an existing canonical IV in a suitable type. 1503 PHINode *CanonicalIV = nullptr; 1504 if (PHINode *PN = L->getCanonicalInductionVariable()) 1505 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 1506 CanonicalIV = PN; 1507 1508 // Rewrite an AddRec in terms of the canonical induction variable, if 1509 // its type is more narrow. 1510 if (CanonicalIV && 1511 SE.getTypeSizeInBits(CanonicalIV->getType()) > 1512 SE.getTypeSizeInBits(Ty)) { 1513 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1514 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1515 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 1516 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 1517 S->getNoWrapFlags(SCEV::FlagNW))); 1518 BasicBlock::iterator NewInsertPt = 1519 findInsertPointAfter(cast<Instruction>(V), Builder.GetInsertBlock()); 1520 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr, 1521 &*NewInsertPt); 1522 return V; 1523 } 1524 1525 // {X,+,F} --> X + {0,+,F} 1526 if (!S->getStart()->isZero()) { 1527 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1528 NewOps[0] = SE.getConstant(Ty, 0); 1529 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, 1530 S->getNoWrapFlags(SCEV::FlagNW)); 1531 1532 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1533 // comments on expandAddToGEP for details. 1534 const SCEV *Base = S->getStart(); 1535 // Dig into the expression to find the pointer base for a GEP. 1536 const SCEV *ExposedRest = Rest; 1537 ExposePointerBase(Base, ExposedRest, SE); 1538 // If we found a pointer, expand the AddRec with a GEP. 1539 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1540 // Make sure the Base isn't something exotic, such as a multiplied 1541 // or divided pointer value. In those cases, the result type isn't 1542 // actually a pointer type. 1543 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1544 Value *StartV = expand(Base); 1545 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1546 return expandAddToGEP(ExposedRest, PTy, Ty, StartV); 1547 } 1548 } 1549 1550 // Just do a normal add. Pre-expand the operands to suppress folding. 1551 // 1552 // The LHS and RHS values are factored out of the expand call to make the 1553 // output independent of the argument evaluation order. 1554 const SCEV *AddExprLHS = SE.getUnknown(expand(S->getStart())); 1555 const SCEV *AddExprRHS = SE.getUnknown(expand(Rest)); 1556 return expand(SE.getAddExpr(AddExprLHS, AddExprRHS)); 1557 } 1558 1559 // If we don't yet have a canonical IV, create one. 1560 if (!CanonicalIV) { 1561 // Create and insert the PHI node for the induction variable in the 1562 // specified loop. 1563 BasicBlock *Header = L->getHeader(); 1564 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1565 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 1566 &Header->front()); 1567 rememberInstruction(CanonicalIV); 1568 1569 SmallSet<BasicBlock *, 4> PredSeen; 1570 Constant *One = ConstantInt::get(Ty, 1); 1571 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1572 BasicBlock *HP = *HPI; 1573 if (!PredSeen.insert(HP).second) { 1574 // There must be an incoming value for each predecessor, even the 1575 // duplicates! 1576 CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP); 1577 continue; 1578 } 1579 1580 if (L->contains(HP)) { 1581 // Insert a unit add instruction right before the terminator 1582 // corresponding to the back-edge. 1583 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 1584 "indvar.next", 1585 HP->getTerminator()); 1586 Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 1587 rememberInstruction(Add); 1588 CanonicalIV->addIncoming(Add, HP); 1589 } else { 1590 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 1591 } 1592 } 1593 } 1594 1595 // {0,+,1} --> Insert a canonical induction variable into the loop! 1596 if (S->isAffine() && S->getOperand(1)->isOne()) { 1597 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 1598 "IVs with types different from the canonical IV should " 1599 "already have been handled!"); 1600 return CanonicalIV; 1601 } 1602 1603 // {0,+,F} --> {0,+,1} * F 1604 1605 // If this is a simple linear addrec, emit it now as a special case. 1606 if (S->isAffine()) // {0,+,F} --> i*F 1607 return 1608 expand(SE.getTruncateOrNoop( 1609 SE.getMulExpr(SE.getUnknown(CanonicalIV), 1610 SE.getNoopOrAnyExtend(S->getOperand(1), 1611 CanonicalIV->getType())), 1612 Ty)); 1613 1614 // If this is a chain of recurrences, turn it into a closed form, using the 1615 // folders, then expandCodeFor the closed form. This allows the folders to 1616 // simplify the expression without having to build a bunch of special code 1617 // into this folder. 1618 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 1619 1620 // Promote S up to the canonical IV type, if the cast is foldable. 1621 const SCEV *NewS = S; 1622 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 1623 if (isa<SCEVAddRecExpr>(Ext)) 1624 NewS = Ext; 1625 1626 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1627 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 1628 1629 // Truncate the result down to the original type, if needed. 1630 const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1631 return expand(T); 1632} 1633 1634Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1635 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1636 Value *V = expandCodeFor(S->getOperand(), 1637 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1638 Value *I = Builder.CreateTrunc(V, Ty); 1639 rememberInstruction(I); 1640 return I; 1641} 1642 1643Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1644 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1645 Value *V = expandCodeFor(S->getOperand(), 1646 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1647 Value *I = Builder.CreateZExt(V, Ty); 1648 rememberInstruction(I); 1649 return I; 1650} 1651 1652Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1653 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1654 Value *V = expandCodeFor(S->getOperand(), 1655 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1656 Value *I = Builder.CreateSExt(V, Ty); 1657 rememberInstruction(I); 1658 return I; 1659} 1660 1661Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 1662 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1663 Type *Ty = LHS->getType(); 1664 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1665 // In the case of mixed integer and pointer types, do the 1666 // rest of the comparisons as integer. 1667 Type *OpTy = S->getOperand(i)->getType(); 1668 if (OpTy->isIntegerTy() != Ty->isIntegerTy()) { 1669 Ty = SE.getEffectiveSCEVType(Ty); 1670 LHS = InsertNoopCastOfTo(LHS, Ty); 1671 } 1672 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1673 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 1674 rememberInstruction(ICmp); 1675 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1676 rememberInstruction(Sel); 1677 LHS = Sel; 1678 } 1679 // In the case of mixed integer and pointer types, cast the 1680 // final result back to the pointer type. 1681 if (LHS->getType() != S->getType()) 1682 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1683 return LHS; 1684} 1685 1686Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 1687 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1688 Type *Ty = LHS->getType(); 1689 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1690 // In the case of mixed integer and pointer types, do the 1691 // rest of the comparisons as integer. 1692 Type *OpTy = S->getOperand(i)->getType(); 1693 if (OpTy->isIntegerTy() != Ty->isIntegerTy()) { 1694 Ty = SE.getEffectiveSCEVType(Ty); 1695 LHS = InsertNoopCastOfTo(LHS, Ty); 1696 } 1697 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1698 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 1699 rememberInstruction(ICmp); 1700 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1701 rememberInstruction(Sel); 1702 LHS = Sel; 1703 } 1704 // In the case of mixed integer and pointer types, cast the 1705 // final result back to the pointer type. 1706 if (LHS->getType() != S->getType()) 1707 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1708 return LHS; 1709} 1710 1711Value *SCEVExpander::visitSMinExpr(const SCEVSMinExpr *S) { 1712 Value *LHS = expand(S->getOperand(S->getNumOperands() - 1)); 1713 Type *Ty = LHS->getType(); 1714 for (int i = S->getNumOperands() - 2; i >= 0; --i) { 1715 // In the case of mixed integer and pointer types, do the 1716 // rest of the comparisons as integer. 1717 Type *OpTy = S->getOperand(i)->getType(); 1718 if (OpTy->isIntegerTy() != Ty->isIntegerTy()) { 1719 Ty = SE.getEffectiveSCEVType(Ty); 1720 LHS = InsertNoopCastOfTo(LHS, Ty); 1721 } 1722 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1723 Value *ICmp = Builder.CreateICmpSLT(LHS, RHS); 1724 rememberInstruction(ICmp); 1725 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smin"); 1726 rememberInstruction(Sel); 1727 LHS = Sel; 1728 } 1729 // In the case of mixed integer and pointer types, cast the 1730 // final result back to the pointer type. 1731 if (LHS->getType() != S->getType()) 1732 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1733 return LHS; 1734} 1735 1736Value *SCEVExpander::visitUMinExpr(const SCEVUMinExpr *S) { 1737 Value *LHS = expand(S->getOperand(S->getNumOperands() - 1)); 1738 Type *Ty = LHS->getType(); 1739 for (int i = S->getNumOperands() - 2; i >= 0; --i) { 1740 // In the case of mixed integer and pointer types, do the 1741 // rest of the comparisons as integer. 1742 Type *OpTy = S->getOperand(i)->getType(); 1743 if (OpTy->isIntegerTy() != Ty->isIntegerTy()) { 1744 Ty = SE.getEffectiveSCEVType(Ty); 1745 LHS = InsertNoopCastOfTo(LHS, Ty); 1746 } 1747 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1748 Value *ICmp = Builder.CreateICmpULT(LHS, RHS); 1749 rememberInstruction(ICmp); 1750 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umin"); 1751 rememberInstruction(Sel); 1752 LHS = Sel; 1753 } 1754 // In the case of mixed integer and pointer types, cast the 1755 // final result back to the pointer type. 1756 if (LHS->getType() != S->getType()) 1757 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1758 return LHS; 1759} 1760 1761Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 1762 Instruction *IP) { 1763 setInsertPoint(IP); 1764 return expandCodeFor(SH, Ty); 1765} 1766 1767Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 1768 // Expand the code for this SCEV. 1769 Value *V = expand(SH); 1770 if (Ty) { 1771 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 1772 "non-trivial casts should be done with the SCEVs directly!"); 1773 V = InsertNoopCastOfTo(V, Ty); 1774 } 1775 return V; 1776} 1777 1778ScalarEvolution::ValueOffsetPair 1779SCEVExpander::FindValueInExprValueMap(const SCEV *S, 1780 const Instruction *InsertPt) { 1781 SetVector<ScalarEvolution::ValueOffsetPair> *Set = SE.getSCEVValues(S); 1782 // If the expansion is not in CanonicalMode, and the SCEV contains any 1783 // sub scAddRecExpr type SCEV, it is required to expand the SCEV literally. 1784 if (CanonicalMode || !SE.containsAddRecurrence(S)) { 1785 // If S is scConstant, it may be worse to reuse an existing Value. 1786 if (S->getSCEVType() != scConstant && Set) { 1787 // Choose a Value from the set which dominates the insertPt. 1788 // insertPt should be inside the Value's parent loop so as not to break 1789 // the LCSSA form. 1790 for (auto const &VOPair : *Set) { 1791 Value *V = VOPair.first; 1792 ConstantInt *Offset = VOPair.second; 1793 Instruction *EntInst = nullptr; 1794 if (V && isa<Instruction>(V) && (EntInst = cast<Instruction>(V)) && 1795 S->getType() == V->getType() && 1796 EntInst->getFunction() == InsertPt->getFunction() && 1797 SE.DT.dominates(EntInst, InsertPt) && 1798 (SE.LI.getLoopFor(EntInst->getParent()) == nullptr || 1799 SE.LI.getLoopFor(EntInst->getParent())->contains(InsertPt))) 1800 return {V, Offset}; 1801 } 1802 } 1803 } 1804 return {nullptr, nullptr}; 1805} 1806 1807// The expansion of SCEV will either reuse a previous Value in ExprValueMap, 1808// or expand the SCEV literally. Specifically, if the expansion is in LSRMode, 1809// and the SCEV contains any sub scAddRecExpr type SCEV, it will be expanded 1810// literally, to prevent LSR's transformed SCEV from being reverted. Otherwise, 1811// the expansion will try to reuse Value from ExprValueMap, and only when it 1812// fails, expand the SCEV literally. 1813Value *SCEVExpander::expand(const SCEV *S) { 1814 // Compute an insertion point for this SCEV object. Hoist the instructions 1815 // as far out in the loop nest as possible. 1816 Instruction *InsertPt = &*Builder.GetInsertPoint(); 1817 1818 // We can move insertion point only if there is no div or rem operations 1819 // otherwise we are risky to move it over the check for zero denominator. 1820 auto SafeToHoist = [](const SCEV *S) { 1821 return !SCEVExprContains(S, [](const SCEV *S) { 1822 if (const auto *D = dyn_cast<SCEVUDivExpr>(S)) { 1823 if (const auto *SC = dyn_cast<SCEVConstant>(D->getRHS())) 1824 // Division by non-zero constants can be hoisted. 1825 return SC->getValue()->isZero(); 1826 // All other divisions should not be moved as they may be 1827 // divisions by zero and should be kept within the 1828 // conditions of the surrounding loops that guard their 1829 // execution (see PR35406). 1830 return true; 1831 } 1832 return false; 1833 }); 1834 }; 1835 if (SafeToHoist(S)) { 1836 for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());; 1837 L = L->getParentLoop()) { 1838 if (SE.isLoopInvariant(S, L)) { 1839 if (!L) break; 1840 if (BasicBlock *Preheader = L->getLoopPreheader()) 1841 InsertPt = Preheader->getTerminator(); 1842 else 1843 // LSR sets the insertion point for AddRec start/step values to the 1844 // block start to simplify value reuse, even though it's an invalid 1845 // position. SCEVExpander must correct for this in all cases. 1846 InsertPt = &*L->getHeader()->getFirstInsertionPt(); 1847 } else { 1848 // If the SCEV is computable at this level, insert it into the header 1849 // after the PHIs (and after any other instructions that we've inserted 1850 // there) so that it is guaranteed to dominate any user inside the loop. 1851 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 1852 InsertPt = &*L->getHeader()->getFirstInsertionPt(); 1853 while (InsertPt->getIterator() != Builder.GetInsertPoint() && 1854 (isInsertedInstruction(InsertPt) || 1855 isa<DbgInfoIntrinsic>(InsertPt))) 1856 InsertPt = &*std::next(InsertPt->getIterator()); 1857 break; 1858 } 1859 } 1860 } 1861 1862 // IndVarSimplify sometimes sets the insertion point at the block start, even 1863 // when there are PHIs at that point. We must correct for this. 1864 if (isa<PHINode>(*InsertPt)) 1865 InsertPt = &*InsertPt->getParent()->getFirstInsertionPt(); 1866 1867 // Check to see if we already expanded this here. 1868 auto I = InsertedExpressions.find(std::make_pair(S, InsertPt)); 1869 if (I != InsertedExpressions.end()) 1870 return I->second; 1871 1872 SCEVInsertPointGuard Guard(Builder, this); 1873 Builder.SetInsertPoint(InsertPt); 1874 1875 // Expand the expression into instructions. 1876 ScalarEvolution::ValueOffsetPair VO = FindValueInExprValueMap(S, InsertPt); 1877 Value *V = VO.first; 1878 1879 if (!V) 1880 V = visit(S); 1881 else if (VO.second) { 1882 if (PointerType *Vty = dyn_cast<PointerType>(V->getType())) { 1883 Type *Ety = Vty->getPointerElementType(); 1884 int64_t Offset = VO.second->getSExtValue(); 1885 int64_t ESize = SE.getTypeSizeInBits(Ety); 1886 if ((Offset * 8) % ESize == 0) { 1887 ConstantInt *Idx = 1888 ConstantInt::getSigned(VO.second->getType(), -(Offset * 8) / ESize); 1889 V = Builder.CreateGEP(Ety, V, Idx, "scevgep"); 1890 } else { 1891 ConstantInt *Idx = 1892 ConstantInt::getSigned(VO.second->getType(), -Offset); 1893 unsigned AS = Vty->getAddressSpace(); 1894 V = Builder.CreateBitCast(V, Type::getInt8PtrTy(SE.getContext(), AS)); 1895 V = Builder.CreateGEP(Type::getInt8Ty(SE.getContext()), V, Idx, 1896 "uglygep"); 1897 V = Builder.CreateBitCast(V, Vty); 1898 } 1899 } else { 1900 V = Builder.CreateSub(V, VO.second); 1901 } 1902 } 1903 // Remember the expanded value for this SCEV at this location. 1904 // 1905 // This is independent of PostIncLoops. The mapped value simply materializes 1906 // the expression at this insertion point. If the mapped value happened to be 1907 // a postinc expansion, it could be reused by a non-postinc user, but only if 1908 // its insertion point was already at the head of the loop. 1909 InsertedExpressions[std::make_pair(S, InsertPt)] = V; 1910 return V; 1911} 1912 1913void SCEVExpander::rememberInstruction(Value *I) { 1914 if (!PostIncLoops.empty()) 1915 InsertedPostIncValues.insert(I); 1916 else 1917 InsertedValues.insert(I); 1918} 1919 1920/// getOrInsertCanonicalInductionVariable - This method returns the 1921/// canonical induction variable of the specified type for the specified 1922/// loop (inserting one if there is none). A canonical induction variable 1923/// starts at zero and steps by one on each iteration. 1924PHINode * 1925SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1926 Type *Ty) { 1927 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1928 1929 // Build a SCEV for {0,+,1}<L>. 1930 // Conservatively use FlagAnyWrap for now. 1931 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 1932 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 1933 1934 // Emit code for it. 1935 SCEVInsertPointGuard Guard(Builder, this); 1936 PHINode *V = 1937 cast<PHINode>(expandCodeFor(H, nullptr, &L->getHeader()->front())); 1938 1939 return V; 1940} 1941 1942/// replaceCongruentIVs - Check for congruent phis in this loop header and 1943/// replace them with their most canonical representative. Return the number of 1944/// phis eliminated. 1945/// 1946/// This does not depend on any SCEVExpander state but should be used in 1947/// the same context that SCEVExpander is used. 1948unsigned 1949SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 1950 SmallVectorImpl<WeakTrackingVH> &DeadInsts, 1951 const TargetTransformInfo *TTI) { 1952 // Find integer phis in order of increasing width. 1953 SmallVector<PHINode*, 8> Phis; 1954 for (PHINode &PN : L->getHeader()->phis()) 1955 Phis.push_back(&PN); 1956 1957 if (TTI) 1958 llvm::sort(Phis, [](Value *LHS, Value *RHS) { 1959 // Put pointers at the back and make sure pointer < pointer = false. 1960 if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy()) 1961 return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy(); 1962 return RHS->getType()->getPrimitiveSizeInBits() < 1963 LHS->getType()->getPrimitiveSizeInBits(); 1964 }); 1965 1966 unsigned NumElim = 0; 1967 DenseMap<const SCEV *, PHINode *> ExprToIVMap; 1968 // Process phis from wide to narrow. Map wide phis to their truncation 1969 // so narrow phis can reuse them. 1970 for (PHINode *Phi : Phis) { 1971 auto SimplifyPHINode = [&](PHINode *PN) -> Value * { 1972 if (Value *V = SimplifyInstruction(PN, {DL, &SE.TLI, &SE.DT, &SE.AC})) 1973 return V; 1974 if (!SE.isSCEVable(PN->getType())) 1975 return nullptr; 1976 auto *Const = dyn_cast<SCEVConstant>(SE.getSCEV(PN)); 1977 if (!Const) 1978 return nullptr; 1979 return Const->getValue(); 1980 }; 1981 1982 // Fold constant phis. They may be congruent to other constant phis and 1983 // would confuse the logic below that expects proper IVs. 1984 if (Value *V = SimplifyPHINode(Phi)) { 1985 if (V->getType() != Phi->getType()) 1986 continue; 1987 Phi->replaceAllUsesWith(V); 1988 DeadInsts.emplace_back(Phi); 1989 ++NumElim; 1990 DEBUG_WITH_TYPE(DebugType, dbgs() 1991 << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 1992 continue; 1993 } 1994 1995 if (!SE.isSCEVable(Phi->getType())) 1996 continue; 1997 1998 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 1999 if (!OrigPhiRef) { 2000 OrigPhiRef = Phi; 2001 if (Phi->getType()->isIntegerTy() && TTI && 2002 TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 2003 // This phi can be freely truncated to the narrowest phi type. Map the 2004 // truncated expression to it so it will be reused for narrow types. 2005 const SCEV *TruncExpr = 2006 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 2007 ExprToIVMap[TruncExpr] = Phi; 2008 } 2009 continue; 2010 } 2011 2012 // Replacing a pointer phi with an integer phi or vice-versa doesn't make 2013 // sense. 2014 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 2015 continue; 2016 2017 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 2018 Instruction *OrigInc = dyn_cast<Instruction>( 2019 OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 2020 Instruction *IsomorphicInc = 2021 dyn_cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 2022 2023 if (OrigInc && IsomorphicInc) { 2024 // If this phi has the same width but is more canonical, replace the 2025 // original with it. As part of the "more canonical" determination, 2026 // respect a prior decision to use an IV chain. 2027 if (OrigPhiRef->getType() == Phi->getType() && 2028 !(ChainedPhis.count(Phi) || 2029 isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) && 2030 (ChainedPhis.count(Phi) || 2031 isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 2032 std::swap(OrigPhiRef, Phi); 2033 std::swap(OrigInc, IsomorphicInc); 2034 } 2035 // Replacing the congruent phi is sufficient because acyclic 2036 // redundancy elimination, CSE/GVN, should handle the 2037 // rest. However, once SCEV proves that a phi is congruent, 2038 // it's often the head of an IV user cycle that is isomorphic 2039 // with the original phi. It's worth eagerly cleaning up the 2040 // common case of a single IV increment so that DeleteDeadPHIs 2041 // can remove cycles that had postinc uses. 2042 const SCEV *TruncExpr = 2043 SE.getTruncateOrNoop(SE.getSCEV(OrigInc), IsomorphicInc->getType()); 2044 if (OrigInc != IsomorphicInc && 2045 TruncExpr == SE.getSCEV(IsomorphicInc) && 2046 SE.LI.replacementPreservesLCSSAForm(IsomorphicInc, OrigInc) && 2047 hoistIVInc(OrigInc, IsomorphicInc)) { 2048 DEBUG_WITH_TYPE(DebugType, 2049 dbgs() << "INDVARS: Eliminated congruent iv.inc: " 2050 << *IsomorphicInc << '\n'); 2051 Value *NewInc = OrigInc; 2052 if (OrigInc->getType() != IsomorphicInc->getType()) { 2053 Instruction *IP = nullptr; 2054 if (PHINode *PN = dyn_cast<PHINode>(OrigInc)) 2055 IP = &*PN->getParent()->getFirstInsertionPt(); 2056 else 2057 IP = OrigInc->getNextNode(); 2058 2059 IRBuilder<> Builder(IP); 2060 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); 2061 NewInc = Builder.CreateTruncOrBitCast( 2062 OrigInc, IsomorphicInc->getType(), IVName); 2063 } 2064 IsomorphicInc->replaceAllUsesWith(NewInc); 2065 DeadInsts.emplace_back(IsomorphicInc); 2066 } 2067 } 2068 } 2069 DEBUG_WITH_TYPE(DebugType, dbgs() << "INDVARS: Eliminated congruent iv: " 2070 << *Phi << '\n'); 2071 ++NumElim; 2072 Value *NewIV = OrigPhiRef; 2073 if (OrigPhiRef->getType() != Phi->getType()) { 2074 IRBuilder<> Builder(&*L->getHeader()->getFirstInsertionPt()); 2075 Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); 2076 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName); 2077 } 2078 Phi->replaceAllUsesWith(NewIV); 2079 DeadInsts.emplace_back(Phi); 2080 } 2081 return NumElim; 2082} 2083 2084Value *SCEVExpander::getExactExistingExpansion(const SCEV *S, 2085 const Instruction *At, Loop *L) { 2086 Optional<ScalarEvolution::ValueOffsetPair> VO = 2087 getRelatedExistingExpansion(S, At, L); 2088 if (VO && VO.getValue().second == nullptr) 2089 return VO.getValue().first; 2090 return nullptr; 2091} 2092 2093Optional<ScalarEvolution::ValueOffsetPair> 2094SCEVExpander::getRelatedExistingExpansion(const SCEV *S, const Instruction *At, 2095 Loop *L) { 2096 using namespace llvm::PatternMatch; 2097 2098 SmallVector<BasicBlock *, 4> ExitingBlocks; 2099 L->getExitingBlocks(ExitingBlocks); 2100 2101 // Look for suitable value in simple conditions at the loop exits. 2102 for (BasicBlock *BB : ExitingBlocks) { 2103 ICmpInst::Predicate Pred; 2104 Instruction *LHS, *RHS; 2105 2106 if (!match(BB->getTerminator(), 2107 m_Br(m_ICmp(Pred, m_Instruction(LHS), m_Instruction(RHS)), 2108 m_BasicBlock(), m_BasicBlock()))) 2109 continue; 2110 2111 if (SE.getSCEV(LHS) == S && SE.DT.dominates(LHS, At)) 2112 return ScalarEvolution::ValueOffsetPair(LHS, nullptr); 2113 2114 if (SE.getSCEV(RHS) == S && SE.DT.dominates(RHS, At)) 2115 return ScalarEvolution::ValueOffsetPair(RHS, nullptr); 2116 } 2117 2118 // Use expand's logic which is used for reusing a previous Value in 2119 // ExprValueMap. 2120 ScalarEvolution::ValueOffsetPair VO = FindValueInExprValueMap(S, At); 2121 if (VO.first) 2122 return VO; 2123 2124 // There is potential to make this significantly smarter, but this simple 2125 // heuristic already gets some interesting cases. 2126 2127 // Can not find suitable value. 2128 return None; 2129} 2130 2131bool SCEVExpander::isHighCostExpansionHelper( 2132 const SCEV *S, Loop *L, const Instruction *At, 2133 SmallPtrSetImpl<const SCEV *> &Processed) { 2134 2135 // If we can find an existing value for this scev available at the point "At" 2136 // then consider the expression cheap. 2137 if (At && getRelatedExistingExpansion(S, At, L)) 2138 return false; 2139 2140 // Zero/One operand expressions 2141 switch (S->getSCEVType()) { 2142 case scUnknown: 2143 case scConstant: 2144 return false; 2145 case scTruncate: 2146 return isHighCostExpansionHelper(cast<SCEVTruncateExpr>(S)->getOperand(), 2147 L, At, Processed); 2148 case scZeroExtend: 2149 return isHighCostExpansionHelper(cast<SCEVZeroExtendExpr>(S)->getOperand(), 2150 L, At, Processed); 2151 case scSignExtend: 2152 return isHighCostExpansionHelper(cast<SCEVSignExtendExpr>(S)->getOperand(), 2153 L, At, Processed); 2154 } 2155 2156 if (!Processed.insert(S).second) 2157 return false; 2158 2159 if (auto *UDivExpr = dyn_cast<SCEVUDivExpr>(S)) { 2160 // If the divisor is a power of two and the SCEV type fits in a native 2161 // integer (and the LHS not expensive), consider the division cheap 2162 // irrespective of whether it occurs in the user code since it can be 2163 // lowered into a right shift. 2164 if (auto *SC = dyn_cast<SCEVConstant>(UDivExpr->getRHS())) 2165 if (SC->getAPInt().isPowerOf2()) { 2166 if (isHighCostExpansionHelper(UDivExpr->getLHS(), L, At, Processed)) 2167 return true; 2168 const DataLayout &DL = 2169 L->getHeader()->getParent()->getParent()->getDataLayout(); 2170 unsigned Width = cast<IntegerType>(UDivExpr->getType())->getBitWidth(); 2171 return DL.isIllegalInteger(Width); 2172 } 2173 2174 // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or 2175 // HowManyLessThans produced to compute a precise expression, rather than a 2176 // UDiv from the user's code. If we can't find a UDiv in the code with some 2177 // simple searching, assume the former consider UDivExpr expensive to 2178 // compute. 2179 BasicBlock *ExitingBB = L->getExitingBlock(); 2180 if (!ExitingBB) 2181 return true; 2182 2183 // At the beginning of this function we already tried to find existing value 2184 // for plain 'S'. Now try to lookup 'S + 1' since it is common pattern 2185 // involving division. This is just a simple search heuristic. 2186 if (!At) 2187 At = &ExitingBB->back(); 2188 if (!getRelatedExistingExpansion( 2189 SE.getAddExpr(S, SE.getConstant(S->getType(), 1)), At, L)) 2190 return true; 2191 } 2192 2193 // HowManyLessThans uses a Max expression whenever the loop is not guarded by 2194 // the exit condition. 2195 if (isa<SCEVMinMaxExpr>(S)) 2196 return true; 2197 2198 // Recurse past nary expressions, which commonly occur in the 2199 // BackedgeTakenCount. They may already exist in program code, and if not, 2200 // they are not too expensive rematerialize. 2201 if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(S)) { 2202 for (auto *Op : NAry->operands()) 2203 if (isHighCostExpansionHelper(Op, L, At, Processed)) 2204 return true; 2205 } 2206 2207 // If we haven't recognized an expensive SCEV pattern, assume it's an 2208 // expression produced by program code. 2209 return false; 2210} 2211 2212Value *SCEVExpander::expandCodeForPredicate(const SCEVPredicate *Pred, 2213 Instruction *IP) { 2214 assert(IP); 2215 switch (Pred->getKind()) { 2216 case SCEVPredicate::P_Union: 2217 return expandUnionPredicate(cast<SCEVUnionPredicate>(Pred), IP); 2218 case SCEVPredicate::P_Equal: 2219 return expandEqualPredicate(cast<SCEVEqualPredicate>(Pred), IP); 2220 case SCEVPredicate::P_Wrap: { 2221 auto *AddRecPred = cast<SCEVWrapPredicate>(Pred); 2222 return expandWrapPredicate(AddRecPred, IP); 2223 } 2224 } 2225 llvm_unreachable("Unknown SCEV predicate type"); 2226} 2227 2228Value *SCEVExpander::expandEqualPredicate(const SCEVEqualPredicate *Pred, 2229 Instruction *IP) { 2230 Value *Expr0 = expandCodeFor(Pred->getLHS(), Pred->getLHS()->getType(), IP); 2231 Value *Expr1 = expandCodeFor(Pred->getRHS(), Pred->getRHS()->getType(), IP); 2232 2233 Builder.SetInsertPoint(IP); 2234 auto *I = Builder.CreateICmpNE(Expr0, Expr1, "ident.check"); 2235 return I; 2236} 2237 2238Value *SCEVExpander::generateOverflowCheck(const SCEVAddRecExpr *AR, 2239 Instruction *Loc, bool Signed) { 2240 assert(AR->isAffine() && "Cannot generate RT check for " 2241 "non-affine expression"); 2242 2243 SCEVUnionPredicate Pred; 2244 const SCEV *ExitCount = 2245 SE.getPredicatedBackedgeTakenCount(AR->getLoop(), Pred); 2246 2247 assert(ExitCount != SE.getCouldNotCompute() && "Invalid loop count"); 2248 2249 const SCEV *Step = AR->getStepRecurrence(SE); 2250 const SCEV *Start = AR->getStart(); 2251 2252 Type *ARTy = AR->getType(); 2253 unsigned SrcBits = SE.getTypeSizeInBits(ExitCount->getType()); 2254 unsigned DstBits = SE.getTypeSizeInBits(ARTy); 2255 2256 // The expression {Start,+,Step} has nusw/nssw if 2257 // Step < 0, Start - |Step| * Backedge <= Start 2258 // Step >= 0, Start + |Step| * Backedge > Start 2259 // and |Step| * Backedge doesn't unsigned overflow. 2260 2261 IntegerType *CountTy = IntegerType::get(Loc->getContext(), SrcBits); 2262 Builder.SetInsertPoint(Loc); 2263 Value *TripCountVal = expandCodeFor(ExitCount, CountTy, Loc); 2264 2265 IntegerType *Ty = 2266 IntegerType::get(Loc->getContext(), SE.getTypeSizeInBits(ARTy)); 2267 Type *ARExpandTy = DL.isNonIntegralPointerType(ARTy) ? ARTy : Ty; 2268 2269 Value *StepValue = expandCodeFor(Step, Ty, Loc); 2270 Value *NegStepValue = expandCodeFor(SE.getNegativeSCEV(Step), Ty, Loc); 2271 Value *StartValue = expandCodeFor(Start, ARExpandTy, Loc); 2272 2273 ConstantInt *Zero = 2274 ConstantInt::get(Loc->getContext(), APInt::getNullValue(DstBits)); 2275 2276 Builder.SetInsertPoint(Loc); 2277 // Compute |Step| 2278 Value *StepCompare = Builder.CreateICmp(ICmpInst::ICMP_SLT, StepValue, Zero); 2279 Value *AbsStep = Builder.CreateSelect(StepCompare, NegStepValue, StepValue); 2280 2281 // Get the backedge taken count and truncate or extended to the AR type. 2282 Value *TruncTripCount = Builder.CreateZExtOrTrunc(TripCountVal, Ty); 2283 auto *MulF = Intrinsic::getDeclaration(Loc->getModule(), 2284 Intrinsic::umul_with_overflow, Ty); 2285 2286 // Compute |Step| * Backedge 2287 CallInst *Mul = Builder.CreateCall(MulF, {AbsStep, TruncTripCount}, "mul"); 2288 Value *MulV = Builder.CreateExtractValue(Mul, 0, "mul.result"); 2289 Value *OfMul = Builder.CreateExtractValue(Mul, 1, "mul.overflow"); 2290 2291 // Compute: 2292 // Start + |Step| * Backedge < Start 2293 // Start - |Step| * Backedge > Start 2294 Value *Add = nullptr, *Sub = nullptr; 2295 if (PointerType *ARPtrTy = dyn_cast<PointerType>(ARExpandTy)) { 2296 const SCEV *MulS = SE.getSCEV(MulV); 2297 const SCEV *NegMulS = SE.getNegativeSCEV(MulS); 2298 Add = Builder.CreateBitCast(expandAddToGEP(MulS, ARPtrTy, Ty, StartValue), 2299 ARPtrTy); 2300 Sub = Builder.CreateBitCast( 2301 expandAddToGEP(NegMulS, ARPtrTy, Ty, StartValue), ARPtrTy); 2302 } else { 2303 Add = Builder.CreateAdd(StartValue, MulV); 2304 Sub = Builder.CreateSub(StartValue, MulV); 2305 } 2306 2307 Value *EndCompareGT = Builder.CreateICmp( 2308 Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT, Sub, StartValue); 2309 2310 Value *EndCompareLT = Builder.CreateICmp( 2311 Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, Add, StartValue); 2312 2313 // Select the answer based on the sign of Step. 2314 Value *EndCheck = 2315 Builder.CreateSelect(StepCompare, EndCompareGT, EndCompareLT); 2316 2317 // If the backedge taken count type is larger than the AR type, 2318 // check that we don't drop any bits by truncating it. If we are 2319 // dropping bits, then we have overflow (unless the step is zero). 2320 if (SE.getTypeSizeInBits(CountTy) > SE.getTypeSizeInBits(Ty)) { 2321 auto MaxVal = APInt::getMaxValue(DstBits).zext(SrcBits); 2322 auto *BackedgeCheck = 2323 Builder.CreateICmp(ICmpInst::ICMP_UGT, TripCountVal, 2324 ConstantInt::get(Loc->getContext(), MaxVal)); 2325 BackedgeCheck = Builder.CreateAnd( 2326 BackedgeCheck, Builder.CreateICmp(ICmpInst::ICMP_NE, StepValue, Zero)); 2327 2328 EndCheck = Builder.CreateOr(EndCheck, BackedgeCheck); 2329 } 2330 2331 EndCheck = Builder.CreateOr(EndCheck, OfMul); 2332 return EndCheck; 2333} 2334 2335Value *SCEVExpander::expandWrapPredicate(const SCEVWrapPredicate *Pred, 2336 Instruction *IP) { 2337 const auto *A = cast<SCEVAddRecExpr>(Pred->getExpr()); 2338 Value *NSSWCheck = nullptr, *NUSWCheck = nullptr; 2339 2340 // Add a check for NUSW 2341 if (Pred->getFlags() & SCEVWrapPredicate::IncrementNUSW) 2342 NUSWCheck = generateOverflowCheck(A, IP, false); 2343 2344 // Add a check for NSSW 2345 if (Pred->getFlags() & SCEVWrapPredicate::IncrementNSSW) 2346 NSSWCheck = generateOverflowCheck(A, IP, true); 2347 2348 if (NUSWCheck && NSSWCheck) 2349 return Builder.CreateOr(NUSWCheck, NSSWCheck); 2350 2351 if (NUSWCheck) 2352 return NUSWCheck; 2353 2354 if (NSSWCheck) 2355 return NSSWCheck; 2356 2357 return ConstantInt::getFalse(IP->getContext()); 2358} 2359 2360Value *SCEVExpander::expandUnionPredicate(const SCEVUnionPredicate *Union, 2361 Instruction *IP) { 2362 auto *BoolType = IntegerType::get(IP->getContext(), 1); 2363 Value *Check = ConstantInt::getNullValue(BoolType); 2364 2365 // Loop over all checks in this set. 2366 for (auto Pred : Union->getPredicates()) { 2367 auto *NextCheck = expandCodeForPredicate(Pred, IP); 2368 Builder.SetInsertPoint(IP); 2369 Check = Builder.CreateOr(Check, NextCheck); 2370 } 2371 2372 return Check; 2373} 2374 2375namespace { 2376// Search for a SCEV subexpression that is not safe to expand. Any expression 2377// that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely 2378// UDiv expressions. We don't know if the UDiv is derived from an IR divide 2379// instruction, but the important thing is that we prove the denominator is 2380// nonzero before expansion. 2381// 2382// IVUsers already checks that IV-derived expressions are safe. So this check is 2383// only needed when the expression includes some subexpression that is not IV 2384// derived. 2385// 2386// Currently, we only allow division by a nonzero constant here. If this is 2387// inadequate, we could easily allow division by SCEVUnknown by using 2388// ValueTracking to check isKnownNonZero(). 2389// 2390// We cannot generally expand recurrences unless the step dominates the loop 2391// header. The expander handles the special case of affine recurrences by 2392// scaling the recurrence outside the loop, but this technique isn't generally 2393// applicable. Expanding a nested recurrence outside a loop requires computing 2394// binomial coefficients. This could be done, but the recurrence has to be in a 2395// perfectly reduced form, which can't be guaranteed. 2396struct SCEVFindUnsafe { 2397 ScalarEvolution &SE; 2398 bool IsUnsafe; 2399 2400 SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {} 2401 2402 bool follow(const SCEV *S) { 2403 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 2404 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS()); 2405 if (!SC || SC->getValue()->isZero()) { 2406 IsUnsafe = true; 2407 return false; 2408 } 2409 } 2410 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2411 const SCEV *Step = AR->getStepRecurrence(SE); 2412 if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) { 2413 IsUnsafe = true; 2414 return false; 2415 } 2416 } 2417 return true; 2418 } 2419 bool isDone() const { return IsUnsafe; } 2420}; 2421} 2422 2423namespace llvm { 2424bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) { 2425 SCEVFindUnsafe Search(SE); 2426 visitAll(S, Search); 2427 return !Search.IsUnsafe; 2428} 2429 2430bool isSafeToExpandAt(const SCEV *S, const Instruction *InsertionPoint, 2431 ScalarEvolution &SE) { 2432 if (!isSafeToExpand(S, SE)) 2433 return false; 2434 // We have to prove that the expanded site of S dominates InsertionPoint. 2435 // This is easy when not in the same block, but hard when S is an instruction 2436 // to be expanded somewhere inside the same block as our insertion point. 2437 // What we really need here is something analogous to an OrderedBasicBlock, 2438 // but for the moment, we paper over the problem by handling two common and 2439 // cheap to check cases. 2440 if (SE.properlyDominates(S, InsertionPoint->getParent())) 2441 return true; 2442 if (SE.dominates(S, InsertionPoint->getParent())) { 2443 if (InsertionPoint->getParent()->getTerminator() == InsertionPoint) 2444 return true; 2445 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) 2446 for (const Value *V : InsertionPoint->operand_values()) 2447 if (V == U->getValue()) 2448 return true; 2449 } 2450 return false; 2451} 2452} 2453