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/Transforms/Utils/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/CommandLine.h" 28 #include "llvm/Support/Debug.h" 29 #include "llvm/Support/raw_ostream.h" 30 31 using namespace llvm; 32 33 cl::opt<unsigned> llvm::SCEVCheapExpansionBudget( 34 "scev-cheap-expansion-budget", cl::Hidden, cl::init(4), 35 cl::desc("When performing SCEV expansion only if it is cheap to do, this " 36 "controls the budget that is considered cheap (default = 4)")); 37 38 using namespace PatternMatch; 39 40 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, 41 /// reusing an existing cast if a suitable one exists, moving an existing 42 /// cast if a suitable one exists but isn't in the right place, or 43 /// creating a new one. 44 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, 45 Instruction::CastOps Op, 46 BasicBlock::iterator IP) { 47 // This function must be called with the builder having a valid insertion 48 // point. It doesn't need to be the actual IP where the uses of the returned 49 // cast will be added, but it must dominate such IP. 50 // We use this precondition to produce a cast that will dominate all its 51 // uses. In particular, this is crucial for the case where the builder's 52 // insertion point *is* the point where we were asked to put the cast. 53 // Since we don't know the builder's insertion point is actually 54 // where the uses will be added (only that it dominates it), we are 55 // not allowed to move it. 56 BasicBlock::iterator BIP = Builder.GetInsertPoint(); 57 58 Instruction *Ret = nullptr; 59 60 // Check to see if there is already a cast! 61 for (User *U : V->users()) 62 if (U->getType() == Ty) 63 if (CastInst *CI = dyn_cast<CastInst>(U)) 64 if (CI->getOpcode() == Op) { 65 // If the cast isn't where we want it, create a new cast at IP. 66 // Likewise, do not reuse a cast at BIP because it must dominate 67 // instructions that might be inserted before BIP. 68 if (BasicBlock::iterator(CI) != IP || BIP == IP) { 69 // Create a new cast, and leave the old cast in place in case 70 // it is being used as an insert point. 71 Ret = CastInst::Create(Op, V, Ty, "", &*IP); 72 Ret->takeName(CI); 73 CI->replaceAllUsesWith(Ret); 74 break; 75 } 76 Ret = CI; 77 break; 78 } 79 80 // Create a new cast. 81 if (!Ret) 82 Ret = CastInst::Create(Op, V, Ty, V->getName(), &*IP); 83 84 // We assert at the end of the function since IP might point to an 85 // instruction with different dominance properties than a cast 86 // (an invoke for example) and not dominate BIP (but the cast does). 87 assert(SE.DT.dominates(Ret, &*BIP)); 88 89 rememberInstruction(Ret); 90 return Ret; 91 } 92 93 static BasicBlock::iterator findInsertPointAfter(Instruction *I, 94 BasicBlock *MustDominate) { 95 BasicBlock::iterator IP = ++I->getIterator(); 96 if (auto *II = dyn_cast<InvokeInst>(I)) 97 IP = II->getNormalDest()->begin(); 98 99 while (isa<PHINode>(IP)) 100 ++IP; 101 102 if (isa<FuncletPadInst>(IP) || isa<LandingPadInst>(IP)) { 103 ++IP; 104 } else if (isa<CatchSwitchInst>(IP)) { 105 IP = MustDominate->getFirstInsertionPt(); 106 } else { 107 assert(!IP->isEHPad() && "unexpected eh pad!"); 108 } 109 110 return IP; 111 } 112 113 /// InsertNoopCastOfTo - Insert a cast of V to the specified type, 114 /// which must be possible with a noop cast, doing what we can to share 115 /// the casts. 116 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 117 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 118 assert((Op == Instruction::BitCast || 119 Op == Instruction::PtrToInt || 120 Op == Instruction::IntToPtr) && 121 "InsertNoopCastOfTo cannot perform non-noop casts!"); 122 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 123 "InsertNoopCastOfTo cannot change sizes!"); 124 125 // Short-circuit unnecessary bitcasts. 126 if (Op == Instruction::BitCast) { 127 if (V->getType() == Ty) 128 return V; 129 if (CastInst *CI = dyn_cast<CastInst>(V)) { 130 if (CI->getOperand(0)->getType() == Ty) 131 return CI->getOperand(0); 132 } 133 } 134 // Short-circuit unnecessary inttoptr<->ptrtoint casts. 135 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 136 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 137 if (CastInst *CI = dyn_cast<CastInst>(V)) 138 if ((CI->getOpcode() == Instruction::PtrToInt || 139 CI->getOpcode() == Instruction::IntToPtr) && 140 SE.getTypeSizeInBits(CI->getType()) == 141 SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 142 return CI->getOperand(0); 143 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 144 if ((CE->getOpcode() == Instruction::PtrToInt || 145 CE->getOpcode() == Instruction::IntToPtr) && 146 SE.getTypeSizeInBits(CE->getType()) == 147 SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 148 return CE->getOperand(0); 149 } 150 151 // Fold a cast of a constant. 152 if (Constant *C = dyn_cast<Constant>(V)) 153 return ConstantExpr::getCast(Op, C, Ty); 154 155 // Cast the argument at the beginning of the entry block, after 156 // any bitcasts of other arguments. 157 if (Argument *A = dyn_cast<Argument>(V)) { 158 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 159 while ((isa<BitCastInst>(IP) && 160 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 161 cast<BitCastInst>(IP)->getOperand(0) != A) || 162 isa<DbgInfoIntrinsic>(IP)) 163 ++IP; 164 return ReuseOrCreateCast(A, Ty, Op, IP); 165 } 166 167 // Cast the instruction immediately after the instruction. 168 Instruction *I = cast<Instruction>(V); 169 BasicBlock::iterator IP = findInsertPointAfter(I, Builder.GetInsertBlock()); 170 return ReuseOrCreateCast(I, Ty, Op, IP); 171 } 172 173 /// InsertBinop - Insert the specified binary operator, doing a small amount 174 /// of work to avoid inserting an obviously redundant operation, and hoisting 175 /// to an outer loop when the opportunity is there and it is safe. 176 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 177 Value *LHS, Value *RHS, 178 SCEV::NoWrapFlags Flags, bool IsSafeToHoist) { 179 // Fold a binop with constant operands. 180 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 181 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 182 return ConstantExpr::get(Opcode, CLHS, CRHS); 183 184 // Do a quick scan to see if we have this binop nearby. If so, reuse it. 185 unsigned ScanLimit = 6; 186 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 187 // Scanning starts from the last instruction before the insertion point. 188 BasicBlock::iterator IP = Builder.GetInsertPoint(); 189 if (IP != BlockBegin) { 190 --IP; 191 for (; ScanLimit; --IP, --ScanLimit) { 192 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 193 // generated code. 194 if (isa<DbgInfoIntrinsic>(IP)) 195 ScanLimit++; 196 197 auto canGenerateIncompatiblePoison = [&Flags](Instruction *I) { 198 // Ensure that no-wrap flags match. 199 if (isa<OverflowingBinaryOperator>(I)) { 200 if (I->hasNoSignedWrap() != (Flags & SCEV::FlagNSW)) 201 return true; 202 if (I->hasNoUnsignedWrap() != (Flags & SCEV::FlagNUW)) 203 return true; 204 } 205 // Conservatively, do not use any instruction which has any of exact 206 // flags installed. 207 if (isa<PossiblyExactOperator>(I) && I->isExact()) 208 return true; 209 return false; 210 }; 211 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 212 IP->getOperand(1) == RHS && !canGenerateIncompatiblePoison(&*IP)) 213 return &*IP; 214 if (IP == BlockBegin) break; 215 } 216 } 217 218 // Save the original insertion point so we can restore it when we're done. 219 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); 220 SCEVInsertPointGuard Guard(Builder, this); 221 222 if (IsSafeToHoist) { 223 // Move the insertion point out of as many loops as we can. 224 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 225 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 226 BasicBlock *Preheader = L->getLoopPreheader(); 227 if (!Preheader) break; 228 229 // Ok, move up a level. 230 Builder.SetInsertPoint(Preheader->getTerminator()); 231 } 232 } 233 234 // If we haven't found this binop, insert it. 235 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 236 BO->setDebugLoc(Loc); 237 if (Flags & SCEV::FlagNUW) 238 BO->setHasNoUnsignedWrap(); 239 if (Flags & SCEV::FlagNSW) 240 BO->setHasNoSignedWrap(); 241 rememberInstruction(BO); 242 243 return BO; 244 } 245 246 /// FactorOutConstant - Test if S is divisible by Factor, using signed 247 /// division. If so, update S with Factor divided out and return true. 248 /// S need not be evenly divisible if a reasonable remainder can be 249 /// computed. 250 static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder, 251 const SCEV *Factor, ScalarEvolution &SE, 252 const DataLayout &DL) { 253 // Everything is divisible by one. 254 if (Factor->isOne()) 255 return true; 256 257 // x/x == 1. 258 if (S == Factor) { 259 S = SE.getConstant(S->getType(), 1); 260 return true; 261 } 262 263 // For a Constant, check for a multiple of the given factor. 264 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 265 // 0/x == 0. 266 if (C->isZero()) 267 return true; 268 // Check for divisibility. 269 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 270 ConstantInt *CI = 271 ConstantInt::get(SE.getContext(), C->getAPInt().sdiv(FC->getAPInt())); 272 // If the quotient is zero and the remainder is non-zero, reject 273 // the value at this scale. It will be considered for subsequent 274 // smaller scales. 275 if (!CI->isZero()) { 276 const SCEV *Div = SE.getConstant(CI); 277 S = Div; 278 Remainder = SE.getAddExpr( 279 Remainder, SE.getConstant(C->getAPInt().srem(FC->getAPInt()))); 280 return true; 281 } 282 } 283 } 284 285 // In a Mul, check if there is a constant operand which is a multiple 286 // of the given factor. 287 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 288 // Size is known, check if there is a constant operand which is a multiple 289 // of the given factor. If so, we can factor it. 290 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) 291 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 292 if (!C->getAPInt().srem(FC->getAPInt())) { 293 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 294 NewMulOps[0] = SE.getConstant(C->getAPInt().sdiv(FC->getAPInt())); 295 S = SE.getMulExpr(NewMulOps); 296 return true; 297 } 298 } 299 300 // In an AddRec, check if both start and step are divisible. 301 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 302 const SCEV *Step = A->getStepRecurrence(SE); 303 const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 304 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL)) 305 return false; 306 if (!StepRem->isZero()) 307 return false; 308 const SCEV *Start = A->getStart(); 309 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL)) 310 return false; 311 S = SE.getAddRecExpr(Start, Step, A->getLoop(), 312 A->getNoWrapFlags(SCEV::FlagNW)); 313 return true; 314 } 315 316 return false; 317 } 318 319 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 320 /// is the number of SCEVAddRecExprs present, which are kept at the end of 321 /// the list. 322 /// 323 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 324 Type *Ty, 325 ScalarEvolution &SE) { 326 unsigned NumAddRecs = 0; 327 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 328 ++NumAddRecs; 329 // Group Ops into non-addrecs and addrecs. 330 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 331 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 332 // Let ScalarEvolution sort and simplify the non-addrecs list. 333 const SCEV *Sum = NoAddRecs.empty() ? 334 SE.getConstant(Ty, 0) : 335 SE.getAddExpr(NoAddRecs); 336 // If it returned an add, use the operands. Otherwise it simplified 337 // the sum into a single value, so just use that. 338 Ops.clear(); 339 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 340 Ops.append(Add->op_begin(), Add->op_end()); 341 else if (!Sum->isZero()) 342 Ops.push_back(Sum); 343 // Then append the addrecs. 344 Ops.append(AddRecs.begin(), AddRecs.end()); 345 } 346 347 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values 348 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 349 /// This helps expose more opportunities for folding parts of the expressions 350 /// into GEP indices. 351 /// 352 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 353 Type *Ty, 354 ScalarEvolution &SE) { 355 // Find the addrecs. 356 SmallVector<const SCEV *, 8> AddRecs; 357 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 358 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 359 const SCEV *Start = A->getStart(); 360 if (Start->isZero()) break; 361 const SCEV *Zero = SE.getConstant(Ty, 0); 362 AddRecs.push_back(SE.getAddRecExpr(Zero, 363 A->getStepRecurrence(SE), 364 A->getLoop(), 365 A->getNoWrapFlags(SCEV::FlagNW))); 366 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 367 Ops[i] = Zero; 368 Ops.append(Add->op_begin(), Add->op_end()); 369 e += Add->getNumOperands(); 370 } else { 371 Ops[i] = Start; 372 } 373 } 374 if (!AddRecs.empty()) { 375 // Add the addrecs onto the end of the list. 376 Ops.append(AddRecs.begin(), AddRecs.end()); 377 // Resort the operand list, moving any constants to the front. 378 SimplifyAddOperands(Ops, Ty, SE); 379 } 380 } 381 382 /// expandAddToGEP - Expand an addition expression with a pointer type into 383 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 384 /// BasicAliasAnalysis and other passes analyze the result. See the rules 385 /// for getelementptr vs. inttoptr in 386 /// http://llvm.org/docs/LangRef.html#pointeraliasing 387 /// for details. 388 /// 389 /// Design note: The correctness of using getelementptr here depends on 390 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 391 /// they may introduce pointer arithmetic which may not be safely converted 392 /// into getelementptr. 393 /// 394 /// Design note: It might seem desirable for this function to be more 395 /// loop-aware. If some of the indices are loop-invariant while others 396 /// aren't, it might seem desirable to emit multiple GEPs, keeping the 397 /// loop-invariant portions of the overall computation outside the loop. 398 /// However, there are a few reasons this is not done here. Hoisting simple 399 /// arithmetic is a low-level optimization that often isn't very 400 /// important until late in the optimization process. In fact, passes 401 /// like InstructionCombining will combine GEPs, even if it means 402 /// pushing loop-invariant computation down into loops, so even if the 403 /// GEPs were split here, the work would quickly be undone. The 404 /// LoopStrengthReduction pass, which is usually run quite late (and 405 /// after the last InstructionCombining pass), takes care of hoisting 406 /// loop-invariant portions of expressions, after considering what 407 /// can be folded using target addressing modes. 408 /// 409 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 410 const SCEV *const *op_end, 411 PointerType *PTy, 412 Type *Ty, 413 Value *V) { 414 Type *OriginalElTy = PTy->getElementType(); 415 Type *ElTy = OriginalElTy; 416 SmallVector<Value *, 4> GepIndices; 417 SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 418 bool AnyNonZeroIndices = false; 419 420 // Split AddRecs up into parts as either of the parts may be usable 421 // without the other. 422 SplitAddRecs(Ops, Ty, SE); 423 424 Type *IntIdxTy = DL.getIndexType(PTy); 425 426 // Descend down the pointer's type and attempt to convert the other 427 // operands into GEP indices, at each level. The first index in a GEP 428 // indexes into the array implied by the pointer operand; the rest of 429 // the indices index into the element or field type selected by the 430 // preceding index. 431 for (;;) { 432 // If the scale size is not 0, attempt to factor out a scale for 433 // array indexing. 434 SmallVector<const SCEV *, 8> ScaledOps; 435 if (ElTy->isSized()) { 436 const SCEV *ElSize = SE.getSizeOfExpr(IntIdxTy, ElTy); 437 if (!ElSize->isZero()) { 438 SmallVector<const SCEV *, 8> NewOps; 439 for (const SCEV *Op : Ops) { 440 const SCEV *Remainder = SE.getConstant(Ty, 0); 441 if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) { 442 // Op now has ElSize factored out. 443 ScaledOps.push_back(Op); 444 if (!Remainder->isZero()) 445 NewOps.push_back(Remainder); 446 AnyNonZeroIndices = true; 447 } else { 448 // The operand was not divisible, so add it to the list of operands 449 // we'll scan next iteration. 450 NewOps.push_back(Op); 451 } 452 } 453 // If we made any changes, update Ops. 454 if (!ScaledOps.empty()) { 455 Ops = NewOps; 456 SimplifyAddOperands(Ops, Ty, SE); 457 } 458 } 459 } 460 461 // Record the scaled array index for this level of the type. If 462 // we didn't find any operands that could be factored, tentatively 463 // assume that element zero was selected (since the zero offset 464 // would obviously be folded away). 465 Value *Scaled = ScaledOps.empty() ? 466 Constant::getNullValue(Ty) : 467 expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 468 GepIndices.push_back(Scaled); 469 470 // Collect struct field index operands. 471 while (StructType *STy = dyn_cast<StructType>(ElTy)) { 472 bool FoundFieldNo = false; 473 // An empty struct has no fields. 474 if (STy->getNumElements() == 0) break; 475 // Field offsets are known. See if a constant offset falls within any of 476 // the struct fields. 477 if (Ops.empty()) 478 break; 479 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 480 if (SE.getTypeSizeInBits(C->getType()) <= 64) { 481 const StructLayout &SL = *DL.getStructLayout(STy); 482 uint64_t FullOffset = C->getValue()->getZExtValue(); 483 if (FullOffset < SL.getSizeInBytes()) { 484 unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 485 GepIndices.push_back( 486 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 487 ElTy = STy->getTypeAtIndex(ElIdx); 488 Ops[0] = 489 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 490 AnyNonZeroIndices = true; 491 FoundFieldNo = true; 492 } 493 } 494 // If no struct field offsets were found, tentatively assume that 495 // field zero was selected (since the zero offset would obviously 496 // be folded away). 497 if (!FoundFieldNo) { 498 ElTy = STy->getTypeAtIndex(0u); 499 GepIndices.push_back( 500 Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 501 } 502 } 503 504 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 505 ElTy = ATy->getElementType(); 506 else 507 // FIXME: Handle VectorType. 508 // E.g., If ElTy is scalable vector, then ElSize is not a compile-time 509 // constant, therefore can not be factored out. The generated IR is less 510 // ideal with base 'V' cast to i8* and do ugly getelementptr over that. 511 break; 512 } 513 514 // If none of the operands were convertible to proper GEP indices, cast 515 // the base to i8* and do an ugly getelementptr with that. It's still 516 // better than ptrtoint+arithmetic+inttoptr at least. 517 if (!AnyNonZeroIndices) { 518 // Cast the base to i8*. 519 V = InsertNoopCastOfTo(V, 520 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 521 522 assert(!isa<Instruction>(V) || 523 SE.DT.dominates(cast<Instruction>(V), &*Builder.GetInsertPoint())); 524 525 // Expand the operands for a plain byte offset. 526 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 527 528 // Fold a GEP with constant operands. 529 if (Constant *CLHS = dyn_cast<Constant>(V)) 530 if (Constant *CRHS = dyn_cast<Constant>(Idx)) 531 return ConstantExpr::getGetElementPtr(Type::getInt8Ty(Ty->getContext()), 532 CLHS, CRHS); 533 534 // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 535 unsigned ScanLimit = 6; 536 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 537 // Scanning starts from the last instruction before the insertion point. 538 BasicBlock::iterator IP = Builder.GetInsertPoint(); 539 if (IP != BlockBegin) { 540 --IP; 541 for (; ScanLimit; --IP, --ScanLimit) { 542 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 543 // generated code. 544 if (isa<DbgInfoIntrinsic>(IP)) 545 ScanLimit++; 546 if (IP->getOpcode() == Instruction::GetElementPtr && 547 IP->getOperand(0) == V && IP->getOperand(1) == Idx) 548 return &*IP; 549 if (IP == BlockBegin) break; 550 } 551 } 552 553 // Save the original insertion point so we can restore it when we're done. 554 SCEVInsertPointGuard Guard(Builder, this); 555 556 // Move the insertion point out of as many loops as we can. 557 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 558 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 559 BasicBlock *Preheader = L->getLoopPreheader(); 560 if (!Preheader) break; 561 562 // Ok, move up a level. 563 Builder.SetInsertPoint(Preheader->getTerminator()); 564 } 565 566 // Emit a GEP. 567 Value *GEP = Builder.CreateGEP(Builder.getInt8Ty(), V, Idx, "uglygep"); 568 rememberInstruction(GEP); 569 570 return GEP; 571 } 572 573 { 574 SCEVInsertPointGuard Guard(Builder, this); 575 576 // Move the insertion point out of as many loops as we can. 577 while (const Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock())) { 578 if (!L->isLoopInvariant(V)) break; 579 580 bool AnyIndexNotLoopInvariant = any_of( 581 GepIndices, [L](Value *Op) { return !L->isLoopInvariant(Op); }); 582 583 if (AnyIndexNotLoopInvariant) 584 break; 585 586 BasicBlock *Preheader = L->getLoopPreheader(); 587 if (!Preheader) break; 588 589 // Ok, move up a level. 590 Builder.SetInsertPoint(Preheader->getTerminator()); 591 } 592 593 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 594 // because ScalarEvolution may have changed the address arithmetic to 595 // compute a value which is beyond the end of the allocated object. 596 Value *Casted = V; 597 if (V->getType() != PTy) 598 Casted = InsertNoopCastOfTo(Casted, PTy); 599 Value *GEP = Builder.CreateGEP(OriginalElTy, Casted, GepIndices, "scevgep"); 600 Ops.push_back(SE.getUnknown(GEP)); 601 rememberInstruction(GEP); 602 } 603 604 return expand(SE.getAddExpr(Ops)); 605 } 606 607 Value *SCEVExpander::expandAddToGEP(const SCEV *Op, PointerType *PTy, Type *Ty, 608 Value *V) { 609 const SCEV *const Ops[1] = {Op}; 610 return expandAddToGEP(Ops, Ops + 1, PTy, Ty, V); 611 } 612 613 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 614 /// SCEV expansion. If they are nested, this is the most nested. If they are 615 /// neighboring, pick the later. 616 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 617 DominatorTree &DT) { 618 if (!A) return B; 619 if (!B) return A; 620 if (A->contains(B)) return B; 621 if (B->contains(A)) return A; 622 if (DT.dominates(A->getHeader(), B->getHeader())) return B; 623 if (DT.dominates(B->getHeader(), A->getHeader())) return A; 624 return A; // Arbitrarily break the tie. 625 } 626 627 /// getRelevantLoop - Get the most relevant loop associated with the given 628 /// expression, according to PickMostRelevantLoop. 629 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 630 // Test whether we've already computed the most relevant loop for this SCEV. 631 auto Pair = RelevantLoops.insert(std::make_pair(S, nullptr)); 632 if (!Pair.second) 633 return Pair.first->second; 634 635 if (isa<SCEVConstant>(S)) 636 // A constant has no relevant loops. 637 return nullptr; 638 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 639 if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 640 return Pair.first->second = SE.LI.getLoopFor(I->getParent()); 641 // A non-instruction has no relevant loops. 642 return nullptr; 643 } 644 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 645 const Loop *L = nullptr; 646 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 647 L = AR->getLoop(); 648 for (const SCEV *Op : N->operands()) 649 L = PickMostRelevantLoop(L, getRelevantLoop(Op), SE.DT); 650 return RelevantLoops[N] = L; 651 } 652 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 653 const Loop *Result = getRelevantLoop(C->getOperand()); 654 return RelevantLoops[C] = Result; 655 } 656 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 657 const Loop *Result = PickMostRelevantLoop( 658 getRelevantLoop(D->getLHS()), getRelevantLoop(D->getRHS()), SE.DT); 659 return RelevantLoops[D] = Result; 660 } 661 llvm_unreachable("Unexpected SCEV type!"); 662 } 663 664 namespace { 665 666 /// LoopCompare - Compare loops by PickMostRelevantLoop. 667 class LoopCompare { 668 DominatorTree &DT; 669 public: 670 explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 671 672 bool operator()(std::pair<const Loop *, const SCEV *> LHS, 673 std::pair<const Loop *, const SCEV *> RHS) const { 674 // Keep pointer operands sorted at the end. 675 if (LHS.second->getType()->isPointerTy() != 676 RHS.second->getType()->isPointerTy()) 677 return LHS.second->getType()->isPointerTy(); 678 679 // Compare loops with PickMostRelevantLoop. 680 if (LHS.first != RHS.first) 681 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 682 683 // If one operand is a non-constant negative and the other is not, 684 // put the non-constant negative on the right so that a sub can 685 // be used instead of a negate and add. 686 if (LHS.second->isNonConstantNegative()) { 687 if (!RHS.second->isNonConstantNegative()) 688 return false; 689 } else if (RHS.second->isNonConstantNegative()) 690 return true; 691 692 // Otherwise they are equivalent according to this comparison. 693 return false; 694 } 695 }; 696 697 } 698 699 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 700 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 701 702 // Collect all the add operands in a loop, along with their associated loops. 703 // Iterate in reverse so that constants are emitted last, all else equal, and 704 // so that pointer operands are inserted first, which the code below relies on 705 // to form more involved GEPs. 706 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 707 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 708 E(S->op_begin()); I != E; ++I) 709 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 710 711 // Sort by loop. Use a stable sort so that constants follow non-constants and 712 // pointer operands precede non-pointer operands. 713 llvm::stable_sort(OpsAndLoops, LoopCompare(SE.DT)); 714 715 // Emit instructions to add all the operands. Hoist as much as possible 716 // out of loops, and form meaningful getelementptrs where possible. 717 Value *Sum = nullptr; 718 for (auto I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E;) { 719 const Loop *CurLoop = I->first; 720 const SCEV *Op = I->second; 721 if (!Sum) { 722 // This is the first operand. Just expand it. 723 Sum = expand(Op); 724 ++I; 725 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 726 // The running sum expression is a pointer. Try to form a getelementptr 727 // at this level with that as the base. 728 SmallVector<const SCEV *, 4> NewOps; 729 for (; I != E && I->first == CurLoop; ++I) { 730 // If the operand is SCEVUnknown and not instructions, peek through 731 // it, to enable more of it to be folded into the GEP. 732 const SCEV *X = I->second; 733 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 734 if (!isa<Instruction>(U->getValue())) 735 X = SE.getSCEV(U->getValue()); 736 NewOps.push_back(X); 737 } 738 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 739 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 740 // The running sum is an integer, and there's a pointer at this level. 741 // Try to form a getelementptr. If the running sum is instructions, 742 // use a SCEVUnknown to avoid re-analyzing them. 743 SmallVector<const SCEV *, 4> NewOps; 744 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 745 SE.getSCEV(Sum)); 746 for (++I; I != E && I->first == CurLoop; ++I) 747 NewOps.push_back(I->second); 748 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 749 } else if (Op->isNonConstantNegative()) { 750 // Instead of doing a negate and add, just do a subtract. 751 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 752 Sum = InsertNoopCastOfTo(Sum, Ty); 753 Sum = InsertBinop(Instruction::Sub, Sum, W, SCEV::FlagAnyWrap, 754 /*IsSafeToHoist*/ true); 755 ++I; 756 } else { 757 // A simple add. 758 Value *W = expandCodeFor(Op, Ty); 759 Sum = InsertNoopCastOfTo(Sum, Ty); 760 // Canonicalize a constant to the RHS. 761 if (isa<Constant>(Sum)) std::swap(Sum, W); 762 Sum = InsertBinop(Instruction::Add, Sum, W, S->getNoWrapFlags(), 763 /*IsSafeToHoist*/ true); 764 ++I; 765 } 766 } 767 768 return Sum; 769 } 770 771 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 772 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 773 774 // Collect all the mul operands in a loop, along with their associated loops. 775 // Iterate in reverse so that constants are emitted last, all else equal. 776 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 777 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 778 E(S->op_begin()); I != E; ++I) 779 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 780 781 // Sort by loop. Use a stable sort so that constants follow non-constants. 782 llvm::stable_sort(OpsAndLoops, LoopCompare(SE.DT)); 783 784 // Emit instructions to mul all the operands. Hoist as much as possible 785 // out of loops. 786 Value *Prod = nullptr; 787 auto I = OpsAndLoops.begin(); 788 789 // Expand the calculation of X pow N in the following manner: 790 // Let N = P1 + P2 + ... + PK, where all P are powers of 2. Then: 791 // X pow N = (X pow P1) * (X pow P2) * ... * (X pow PK). 792 const auto ExpandOpBinPowN = [this, &I, &OpsAndLoops, &Ty]() { 793 auto E = I; 794 // Calculate how many times the same operand from the same loop is included 795 // into this power. 796 uint64_t Exponent = 0; 797 const uint64_t MaxExponent = UINT64_MAX >> 1; 798 // No one sane will ever try to calculate such huge exponents, but if we 799 // need this, we stop on UINT64_MAX / 2 because we need to exit the loop 800 // below when the power of 2 exceeds our Exponent, and we want it to be 801 // 1u << 31 at most to not deal with unsigned overflow. 802 while (E != OpsAndLoops.end() && *I == *E && Exponent != MaxExponent) { 803 ++Exponent; 804 ++E; 805 } 806 assert(Exponent > 0 && "Trying to calculate a zeroth exponent of operand?"); 807 808 // Calculate powers with exponents 1, 2, 4, 8 etc. and include those of them 809 // that are needed into the result. 810 Value *P = expandCodeFor(I->second, Ty); 811 Value *Result = nullptr; 812 if (Exponent & 1) 813 Result = P; 814 for (uint64_t BinExp = 2; BinExp <= Exponent; BinExp <<= 1) { 815 P = InsertBinop(Instruction::Mul, P, P, SCEV::FlagAnyWrap, 816 /*IsSafeToHoist*/ true); 817 if (Exponent & BinExp) 818 Result = Result ? InsertBinop(Instruction::Mul, Result, P, 819 SCEV::FlagAnyWrap, 820 /*IsSafeToHoist*/ true) 821 : P; 822 } 823 824 I = E; 825 assert(Result && "Nothing was expanded?"); 826 return Result; 827 }; 828 829 while (I != OpsAndLoops.end()) { 830 if (!Prod) { 831 // This is the first operand. Just expand it. 832 Prod = ExpandOpBinPowN(); 833 } else if (I->second->isAllOnesValue()) { 834 // Instead of doing a multiply by negative one, just do a negate. 835 Prod = InsertNoopCastOfTo(Prod, Ty); 836 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod, 837 SCEV::FlagAnyWrap, /*IsSafeToHoist*/ true); 838 ++I; 839 } else { 840 // A simple mul. 841 Value *W = ExpandOpBinPowN(); 842 Prod = InsertNoopCastOfTo(Prod, Ty); 843 // Canonicalize a constant to the RHS. 844 if (isa<Constant>(Prod)) std::swap(Prod, W); 845 const APInt *RHS; 846 if (match(W, m_Power2(RHS))) { 847 // Canonicalize Prod*(1<<C) to Prod<<C. 848 assert(!Ty->isVectorTy() && "vector types are not SCEVable"); 849 auto NWFlags = S->getNoWrapFlags(); 850 // clear nsw flag if shl will produce poison value. 851 if (RHS->logBase2() == RHS->getBitWidth() - 1) 852 NWFlags = ScalarEvolution::clearFlags(NWFlags, SCEV::FlagNSW); 853 Prod = InsertBinop(Instruction::Shl, Prod, 854 ConstantInt::get(Ty, RHS->logBase2()), NWFlags, 855 /*IsSafeToHoist*/ true); 856 } else { 857 Prod = InsertBinop(Instruction::Mul, Prod, W, S->getNoWrapFlags(), 858 /*IsSafeToHoist*/ true); 859 } 860 } 861 } 862 863 return Prod; 864 } 865 866 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 867 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 868 869 Value *LHS = expandCodeFor(S->getLHS(), Ty); 870 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 871 const APInt &RHS = SC->getAPInt(); 872 if (RHS.isPowerOf2()) 873 return InsertBinop(Instruction::LShr, LHS, 874 ConstantInt::get(Ty, RHS.logBase2()), 875 SCEV::FlagAnyWrap, /*IsSafeToHoist*/ true); 876 } 877 878 Value *RHS = expandCodeFor(S->getRHS(), Ty); 879 return InsertBinop(Instruction::UDiv, LHS, RHS, SCEV::FlagAnyWrap, 880 /*IsSafeToHoist*/ SE.isKnownNonZero(S->getRHS())); 881 } 882 883 /// Move parts of Base into Rest to leave Base with the minimal 884 /// expression that provides a pointer operand suitable for a 885 /// GEP expansion. 886 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 887 ScalarEvolution &SE) { 888 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 889 Base = A->getStart(); 890 Rest = SE.getAddExpr(Rest, 891 SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 892 A->getStepRecurrence(SE), 893 A->getLoop(), 894 A->getNoWrapFlags(SCEV::FlagNW))); 895 } 896 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 897 Base = A->getOperand(A->getNumOperands()-1); 898 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 899 NewAddOps.back() = Rest; 900 Rest = SE.getAddExpr(NewAddOps); 901 ExposePointerBase(Base, Rest, SE); 902 } 903 } 904 905 /// Determine if this is a well-behaved chain of instructions leading back to 906 /// the PHI. If so, it may be reused by expanded expressions. 907 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 908 const Loop *L) { 909 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 910 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 911 return false; 912 // If any of the operands don't dominate the insert position, bail. 913 // Addrec operands are always loop-invariant, so this can only happen 914 // if there are instructions which haven't been hoisted. 915 if (L == IVIncInsertLoop) { 916 for (User::op_iterator OI = IncV->op_begin()+1, 917 OE = IncV->op_end(); OI != OE; ++OI) 918 if (Instruction *OInst = dyn_cast<Instruction>(OI)) 919 if (!SE.DT.dominates(OInst, IVIncInsertPos)) 920 return false; 921 } 922 // Advance to the next instruction. 923 IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 924 if (!IncV) 925 return false; 926 927 if (IncV->mayHaveSideEffects()) 928 return false; 929 930 if (IncV == PN) 931 return true; 932 933 return isNormalAddRecExprPHI(PN, IncV, L); 934 } 935 936 /// getIVIncOperand returns an induction variable increment's induction 937 /// variable operand. 938 /// 939 /// If allowScale is set, any type of GEP is allowed as long as the nonIV 940 /// operands dominate InsertPos. 941 /// 942 /// If allowScale is not set, ensure that a GEP increment conforms to one of the 943 /// simple patterns generated by getAddRecExprPHILiterally and 944 /// expandAddtoGEP. If the pattern isn't recognized, return NULL. 945 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 946 Instruction *InsertPos, 947 bool allowScale) { 948 if (IncV == InsertPos) 949 return nullptr; 950 951 switch (IncV->getOpcode()) { 952 default: 953 return nullptr; 954 // Check for a simple Add/Sub or GEP of a loop invariant step. 955 case Instruction::Add: 956 case Instruction::Sub: { 957 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 958 if (!OInst || SE.DT.dominates(OInst, InsertPos)) 959 return dyn_cast<Instruction>(IncV->getOperand(0)); 960 return nullptr; 961 } 962 case Instruction::BitCast: 963 return dyn_cast<Instruction>(IncV->getOperand(0)); 964 case Instruction::GetElementPtr: 965 for (auto I = IncV->op_begin() + 1, E = IncV->op_end(); I != E; ++I) { 966 if (isa<Constant>(*I)) 967 continue; 968 if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 969 if (!SE.DT.dominates(OInst, InsertPos)) 970 return nullptr; 971 } 972 if (allowScale) { 973 // allow any kind of GEP as long as it can be hoisted. 974 continue; 975 } 976 // This must be a pointer addition of constants (pretty), which is already 977 // handled, or some number of address-size elements (ugly). Ugly geps 978 // have 2 operands. i1* is used by the expander to represent an 979 // address-size element. 980 if (IncV->getNumOperands() != 2) 981 return nullptr; 982 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 983 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 984 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 985 return nullptr; 986 break; 987 } 988 return dyn_cast<Instruction>(IncV->getOperand(0)); 989 } 990 } 991 992 /// If the insert point of the current builder or any of the builders on the 993 /// stack of saved builders has 'I' as its insert point, update it to point to 994 /// the instruction after 'I'. This is intended to be used when the instruction 995 /// 'I' is being moved. If this fixup is not done and 'I' is moved to a 996 /// different block, the inconsistent insert point (with a mismatched 997 /// Instruction and Block) can lead to an instruction being inserted in a block 998 /// other than its parent. 999 void SCEVExpander::fixupInsertPoints(Instruction *I) { 1000 BasicBlock::iterator It(*I); 1001 BasicBlock::iterator NewInsertPt = std::next(It); 1002 if (Builder.GetInsertPoint() == It) 1003 Builder.SetInsertPoint(&*NewInsertPt); 1004 for (auto *InsertPtGuard : InsertPointGuards) 1005 if (InsertPtGuard->GetInsertPoint() == It) 1006 InsertPtGuard->SetInsertPoint(NewInsertPt); 1007 } 1008 1009 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 1010 /// it available to other uses in this loop. Recursively hoist any operands, 1011 /// until we reach a value that dominates InsertPos. 1012 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 1013 if (SE.DT.dominates(IncV, InsertPos)) 1014 return true; 1015 1016 // InsertPos must itself dominate IncV so that IncV's new position satisfies 1017 // its existing users. 1018 if (isa<PHINode>(InsertPos) || 1019 !SE.DT.dominates(InsertPos->getParent(), IncV->getParent())) 1020 return false; 1021 1022 if (!SE.LI.movementPreservesLCSSAForm(IncV, InsertPos)) 1023 return false; 1024 1025 // Check that the chain of IV operands leading back to Phi can be hoisted. 1026 SmallVector<Instruction*, 4> IVIncs; 1027 for(;;) { 1028 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 1029 if (!Oper) 1030 return false; 1031 // IncV is safe to hoist. 1032 IVIncs.push_back(IncV); 1033 IncV = Oper; 1034 if (SE.DT.dominates(IncV, InsertPos)) 1035 break; 1036 } 1037 for (auto I = IVIncs.rbegin(), E = IVIncs.rend(); I != E; ++I) { 1038 fixupInsertPoints(*I); 1039 (*I)->moveBefore(InsertPos); 1040 } 1041 return true; 1042 } 1043 1044 /// Determine if this cyclic phi is in a form that would have been generated by 1045 /// LSR. We don't care if the phi was actually expanded in this pass, as long 1046 /// as it is in a low-cost form, for example, no implied multiplication. This 1047 /// should match any patterns generated by getAddRecExprPHILiterally and 1048 /// expandAddtoGEP. 1049 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 1050 const Loop *L) { 1051 for(Instruction *IVOper = IncV; 1052 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 1053 /*allowScale=*/false));) { 1054 if (IVOper == PN) 1055 return true; 1056 } 1057 return false; 1058 } 1059 1060 /// expandIVInc - Expand an IV increment at Builder's current InsertPos. 1061 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 1062 /// need to materialize IV increments elsewhere to handle difficult situations. 1063 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 1064 Type *ExpandTy, Type *IntTy, 1065 bool useSubtract) { 1066 Value *IncV; 1067 // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 1068 if (ExpandTy->isPointerTy()) { 1069 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 1070 // If the step isn't constant, don't use an implicitly scaled GEP, because 1071 // that would require a multiply inside the loop. 1072 if (!isa<ConstantInt>(StepV)) 1073 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 1074 GEPPtrTy->getAddressSpace()); 1075 IncV = expandAddToGEP(SE.getSCEV(StepV), GEPPtrTy, IntTy, PN); 1076 if (IncV->getType() != PN->getType()) { 1077 IncV = Builder.CreateBitCast(IncV, PN->getType()); 1078 rememberInstruction(IncV); 1079 } 1080 } else { 1081 IncV = useSubtract ? 1082 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 1083 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 1084 rememberInstruction(IncV); 1085 } 1086 return IncV; 1087 } 1088 1089 /// Hoist the addrec instruction chain rooted in the loop phi above the 1090 /// position. This routine assumes that this is possible (has been checked). 1091 void SCEVExpander::hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist, 1092 Instruction *Pos, PHINode *LoopPhi) { 1093 do { 1094 if (DT->dominates(InstToHoist, Pos)) 1095 break; 1096 // Make sure the increment is where we want it. But don't move it 1097 // down past a potential existing post-inc user. 1098 fixupInsertPoints(InstToHoist); 1099 InstToHoist->moveBefore(Pos); 1100 Pos = InstToHoist; 1101 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0)); 1102 } while (InstToHoist != LoopPhi); 1103 } 1104 1105 /// Check whether we can cheaply express the requested SCEV in terms of 1106 /// the available PHI SCEV by truncation and/or inversion of the step. 1107 static bool canBeCheaplyTransformed(ScalarEvolution &SE, 1108 const SCEVAddRecExpr *Phi, 1109 const SCEVAddRecExpr *Requested, 1110 bool &InvertStep) { 1111 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType()); 1112 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType()); 1113 1114 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth()) 1115 return false; 1116 1117 // Try truncate it if necessary. 1118 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy)); 1119 if (!Phi) 1120 return false; 1121 1122 // Check whether truncation will help. 1123 if (Phi == Requested) { 1124 InvertStep = false; 1125 return true; 1126 } 1127 1128 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}. 1129 if (SE.getAddExpr(Requested->getStart(), 1130 SE.getNegativeSCEV(Requested)) == Phi) { 1131 InvertStep = true; 1132 return true; 1133 } 1134 1135 return false; 1136 } 1137 1138 static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1139 if (!isa<IntegerType>(AR->getType())) 1140 return false; 1141 1142 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1143 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1144 const SCEV *Step = AR->getStepRecurrence(SE); 1145 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy), 1146 SE.getSignExtendExpr(AR, WideTy)); 1147 const SCEV *ExtendAfterOp = 1148 SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1149 return ExtendAfterOp == OpAfterExtend; 1150 } 1151 1152 static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1153 if (!isa<IntegerType>(AR->getType())) 1154 return false; 1155 1156 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1157 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1158 const SCEV *Step = AR->getStepRecurrence(SE); 1159 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy), 1160 SE.getZeroExtendExpr(AR, WideTy)); 1161 const SCEV *ExtendAfterOp = 1162 SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1163 return ExtendAfterOp == OpAfterExtend; 1164 } 1165 1166 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 1167 /// the base addrec, which is the addrec without any non-loop-dominating 1168 /// values, and return the PHI. 1169 PHINode * 1170 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 1171 const Loop *L, 1172 Type *ExpandTy, 1173 Type *IntTy, 1174 Type *&TruncTy, 1175 bool &InvertStep) { 1176 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 1177 1178 // Reuse a previously-inserted PHI, if present. 1179 BasicBlock *LatchBlock = L->getLoopLatch(); 1180 if (LatchBlock) { 1181 PHINode *AddRecPhiMatch = nullptr; 1182 Instruction *IncV = nullptr; 1183 TruncTy = nullptr; 1184 InvertStep = false; 1185 1186 // Only try partially matching scevs that need truncation and/or 1187 // step-inversion if we know this loop is outside the current loop. 1188 bool TryNonMatchingSCEV = 1189 IVIncInsertLoop && 1190 SE.DT.properlyDominates(LatchBlock, IVIncInsertLoop->getHeader()); 1191 1192 for (PHINode &PN : L->getHeader()->phis()) { 1193 if (!SE.isSCEVable(PN.getType())) 1194 continue; 1195 1196 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(&PN)); 1197 if (!PhiSCEV) 1198 continue; 1199 1200 bool IsMatchingSCEV = PhiSCEV == Normalized; 1201 // We only handle truncation and inversion of phi recurrences for the 1202 // expanded expression if the expanded expression's loop dominates the 1203 // loop we insert to. Check now, so we can bail out early. 1204 if (!IsMatchingSCEV && !TryNonMatchingSCEV) 1205 continue; 1206 1207 // TODO: this possibly can be reworked to avoid this cast at all. 1208 Instruction *TempIncV = 1209 dyn_cast<Instruction>(PN.getIncomingValueForBlock(LatchBlock)); 1210 if (!TempIncV) 1211 continue; 1212 1213 // Check whether we can reuse this PHI node. 1214 if (LSRMode) { 1215 if (!isExpandedAddRecExprPHI(&PN, TempIncV, L)) 1216 continue; 1217 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos)) 1218 continue; 1219 } else { 1220 if (!isNormalAddRecExprPHI(&PN, TempIncV, L)) 1221 continue; 1222 } 1223 1224 // Stop if we have found an exact match SCEV. 1225 if (IsMatchingSCEV) { 1226 IncV = TempIncV; 1227 TruncTy = nullptr; 1228 InvertStep = false; 1229 AddRecPhiMatch = &PN; 1230 break; 1231 } 1232 1233 // Try whether the phi can be translated into the requested form 1234 // (truncated and/or offset by a constant). 1235 if ((!TruncTy || InvertStep) && 1236 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) { 1237 // Record the phi node. But don't stop we might find an exact match 1238 // later. 1239 AddRecPhiMatch = &PN; 1240 IncV = TempIncV; 1241 TruncTy = SE.getEffectiveSCEVType(Normalized->getType()); 1242 } 1243 } 1244 1245 if (AddRecPhiMatch) { 1246 // Potentially, move the increment. We have made sure in 1247 // isExpandedAddRecExprPHI or hoistIVInc that this is possible. 1248 if (L == IVIncInsertLoop) 1249 hoistBeforePos(&SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch); 1250 1251 // Ok, the add recurrence looks usable. 1252 // Remember this PHI, even in post-inc mode. 1253 InsertedValues.insert(AddRecPhiMatch); 1254 // Remember the increment. 1255 rememberInstruction(IncV); 1256 return AddRecPhiMatch; 1257 } 1258 } 1259 1260 // Save the original insertion point so we can restore it when we're done. 1261 SCEVInsertPointGuard Guard(Builder, this); 1262 1263 // Another AddRec may need to be recursively expanded below. For example, if 1264 // this AddRec is quadratic, the StepV may itself be an AddRec in this 1265 // loop. Remove this loop from the PostIncLoops set before expanding such 1266 // AddRecs. Otherwise, we cannot find a valid position for the step 1267 // (i.e. StepV can never dominate its loop header). Ideally, we could do 1268 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 1269 // so it's not worth implementing SmallPtrSet::swap. 1270 PostIncLoopSet SavedPostIncLoops = PostIncLoops; 1271 PostIncLoops.clear(); 1272 1273 // Expand code for the start value into the loop preheader. 1274 assert(L->getLoopPreheader() && 1275 "Can't expand add recurrences without a loop preheader!"); 1276 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 1277 L->getLoopPreheader()->getTerminator()); 1278 1279 // StartV must have been be inserted into L's preheader to dominate the new 1280 // phi. 1281 assert(!isa<Instruction>(StartV) || 1282 SE.DT.properlyDominates(cast<Instruction>(StartV)->getParent(), 1283 L->getHeader())); 1284 1285 // Expand code for the step value. Do this before creating the PHI so that PHI 1286 // reuse code doesn't see an incomplete PHI. 1287 const SCEV *Step = Normalized->getStepRecurrence(SE); 1288 // If the stride is negative, insert a sub instead of an add for the increment 1289 // (unless it's a constant, because subtracts of constants are canonicalized 1290 // to adds). 1291 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1292 if (useSubtract) 1293 Step = SE.getNegativeSCEV(Step); 1294 // Expand the step somewhere that dominates the loop header. 1295 Value *StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front()); 1296 1297 // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if 1298 // we actually do emit an addition. It does not apply if we emit a 1299 // subtraction. 1300 bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized); 1301 bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized); 1302 1303 // Create the PHI. 1304 BasicBlock *Header = L->getHeader(); 1305 Builder.SetInsertPoint(Header, Header->begin()); 1306 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1307 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 1308 Twine(IVName) + ".iv"); 1309 rememberInstruction(PN); 1310 1311 // Create the step instructions and populate the PHI. 1312 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1313 BasicBlock *Pred = *HPI; 1314 1315 // Add a start value. 1316 if (!L->contains(Pred)) { 1317 PN->addIncoming(StartV, Pred); 1318 continue; 1319 } 1320 1321 // Create a step value and add it to the PHI. 1322 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 1323 // instructions at IVIncInsertPos. 1324 Instruction *InsertPos = L == IVIncInsertLoop ? 1325 IVIncInsertPos : Pred->getTerminator(); 1326 Builder.SetInsertPoint(InsertPos); 1327 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1328 1329 if (isa<OverflowingBinaryOperator>(IncV)) { 1330 if (IncrementIsNUW) 1331 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap(); 1332 if (IncrementIsNSW) 1333 cast<BinaryOperator>(IncV)->setHasNoSignedWrap(); 1334 } 1335 PN->addIncoming(IncV, Pred); 1336 } 1337 1338 // After expanding subexpressions, restore the PostIncLoops set so the caller 1339 // can ensure that IVIncrement dominates the current uses. 1340 PostIncLoops = SavedPostIncLoops; 1341 1342 // Remember this PHI, even in post-inc mode. 1343 InsertedValues.insert(PN); 1344 1345 return PN; 1346 } 1347 1348 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 1349 Type *STy = S->getType(); 1350 Type *IntTy = SE.getEffectiveSCEVType(STy); 1351 const Loop *L = S->getLoop(); 1352 1353 // Determine a normalized form of this expression, which is the expression 1354 // before any post-inc adjustment is made. 1355 const SCEVAddRecExpr *Normalized = S; 1356 if (PostIncLoops.count(L)) { 1357 PostIncLoopSet Loops; 1358 Loops.insert(L); 1359 Normalized = cast<SCEVAddRecExpr>(normalizeForPostIncUse(S, Loops, SE)); 1360 } 1361 1362 // Strip off any non-loop-dominating component from the addrec start. 1363 const SCEV *Start = Normalized->getStart(); 1364 const SCEV *PostLoopOffset = nullptr; 1365 if (!SE.properlyDominates(Start, L->getHeader())) { 1366 PostLoopOffset = Start; 1367 Start = SE.getConstant(Normalized->getType(), 0); 1368 Normalized = cast<SCEVAddRecExpr>( 1369 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 1370 Normalized->getLoop(), 1371 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1372 } 1373 1374 // Strip off any non-loop-dominating component from the addrec step. 1375 const SCEV *Step = Normalized->getStepRecurrence(SE); 1376 const SCEV *PostLoopScale = nullptr; 1377 if (!SE.dominates(Step, L->getHeader())) { 1378 PostLoopScale = Step; 1379 Step = SE.getConstant(Normalized->getType(), 1); 1380 if (!Start->isZero()) { 1381 // The normalization below assumes that Start is constant zero, so if 1382 // it isn't re-associate Start to PostLoopOffset. 1383 assert(!PostLoopOffset && "Start not-null but PostLoopOffset set?"); 1384 PostLoopOffset = Start; 1385 Start = SE.getConstant(Normalized->getType(), 0); 1386 } 1387 Normalized = 1388 cast<SCEVAddRecExpr>(SE.getAddRecExpr( 1389 Start, Step, Normalized->getLoop(), 1390 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1391 } 1392 1393 // Expand the core addrec. If we need post-loop scaling, force it to 1394 // expand to an integer type to avoid the need for additional casting. 1395 Type *ExpandTy = PostLoopScale ? IntTy : STy; 1396 // We can't use a pointer type for the addrec if the pointer type is 1397 // non-integral. 1398 Type *AddRecPHIExpandTy = 1399 DL.isNonIntegralPointerType(STy) ? Normalized->getType() : ExpandTy; 1400 1401 // In some cases, we decide to reuse an existing phi node but need to truncate 1402 // it and/or invert the step. 1403 Type *TruncTy = nullptr; 1404 bool InvertStep = false; 1405 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, AddRecPHIExpandTy, 1406 IntTy, TruncTy, InvertStep); 1407 1408 // Accommodate post-inc mode, if necessary. 1409 Value *Result; 1410 if (!PostIncLoops.count(L)) 1411 Result = PN; 1412 else { 1413 // In PostInc mode, use the post-incremented value. 1414 BasicBlock *LatchBlock = L->getLoopLatch(); 1415 assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1416 Result = PN->getIncomingValueForBlock(LatchBlock); 1417 1418 // For an expansion to use the postinc form, the client must call 1419 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 1420 // or dominated by IVIncInsertPos. 1421 if (isa<Instruction>(Result) && 1422 !SE.DT.dominates(cast<Instruction>(Result), 1423 &*Builder.GetInsertPoint())) { 1424 // The induction variable's postinc expansion does not dominate this use. 1425 // IVUsers tries to prevent this case, so it is rare. However, it can 1426 // happen when an IVUser outside the loop is not dominated by the latch 1427 // block. Adjusting IVIncInsertPos before expansion begins cannot handle 1428 // all cases. Consider a phi outside whose operand is replaced during 1429 // expansion with the value of the postinc user. Without fundamentally 1430 // changing the way postinc users are tracked, the only remedy is 1431 // inserting an extra IV increment. StepV might fold into PostLoopOffset, 1432 // but hopefully expandCodeFor handles that. 1433 bool useSubtract = 1434 !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1435 if (useSubtract) 1436 Step = SE.getNegativeSCEV(Step); 1437 Value *StepV; 1438 { 1439 // Expand the step somewhere that dominates the loop header. 1440 SCEVInsertPointGuard Guard(Builder, this); 1441 StepV = expandCodeFor(Step, IntTy, &L->getHeader()->front()); 1442 } 1443 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1444 } 1445 } 1446 1447 // We have decided to reuse an induction variable of a dominating loop. Apply 1448 // truncation and/or inversion of the step. 1449 if (TruncTy) { 1450 Type *ResTy = Result->getType(); 1451 // Normalize the result type. 1452 if (ResTy != SE.getEffectiveSCEVType(ResTy)) 1453 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy)); 1454 // Truncate the result. 1455 if (TruncTy != Result->getType()) { 1456 Result = Builder.CreateTrunc(Result, TruncTy); 1457 rememberInstruction(Result); 1458 } 1459 // Invert the result. 1460 if (InvertStep) { 1461 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy), 1462 Result); 1463 rememberInstruction(Result); 1464 } 1465 } 1466 1467 // Re-apply any non-loop-dominating scale. 1468 if (PostLoopScale) { 1469 assert(S->isAffine() && "Can't linearly scale non-affine recurrences."); 1470 Result = InsertNoopCastOfTo(Result, IntTy); 1471 Result = Builder.CreateMul(Result, 1472 expandCodeFor(PostLoopScale, IntTy)); 1473 rememberInstruction(Result); 1474 } 1475 1476 // Re-apply any non-loop-dominating offset. 1477 if (PostLoopOffset) { 1478 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1479 if (Result->getType()->isIntegerTy()) { 1480 Value *Base = expandCodeFor(PostLoopOffset, ExpandTy); 1481 Result = expandAddToGEP(SE.getUnknown(Result), PTy, IntTy, Base); 1482 } else { 1483 Result = expandAddToGEP(PostLoopOffset, PTy, IntTy, Result); 1484 } 1485 } else { 1486 Result = InsertNoopCastOfTo(Result, IntTy); 1487 Result = Builder.CreateAdd(Result, 1488 expandCodeFor(PostLoopOffset, IntTy)); 1489 rememberInstruction(Result); 1490 } 1491 } 1492 1493 return Result; 1494 } 1495 1496 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1497 // In canonical mode we compute the addrec as an expression of a canonical IV 1498 // using evaluateAtIteration and expand the resulting SCEV expression. This 1499 // way we avoid introducing new IVs to carry on the comutation of the addrec 1500 // throughout the loop. 1501 // 1502 // For nested addrecs evaluateAtIteration might need a canonical IV of a 1503 // type wider than the addrec itself. Emitting a canonical IV of the 1504 // proper type might produce non-legal types, for example expanding an i64 1505 // {0,+,2,+,1} addrec would need an i65 canonical IV. To avoid this just fall 1506 // back to non-canonical mode for nested addrecs. 1507 if (!CanonicalMode || (S->getNumOperands() > 2)) 1508 return expandAddRecExprLiterally(S); 1509 1510 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1511 const Loop *L = S->getLoop(); 1512 1513 // First check for an existing canonical IV in a suitable type. 1514 PHINode *CanonicalIV = nullptr; 1515 if (PHINode *PN = L->getCanonicalInductionVariable()) 1516 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 1517 CanonicalIV = PN; 1518 1519 // Rewrite an AddRec in terms of the canonical induction variable, if 1520 // its type is more narrow. 1521 if (CanonicalIV && 1522 SE.getTypeSizeInBits(CanonicalIV->getType()) > 1523 SE.getTypeSizeInBits(Ty)) { 1524 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1525 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1526 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 1527 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 1528 S->getNoWrapFlags(SCEV::FlagNW))); 1529 BasicBlock::iterator NewInsertPt = 1530 findInsertPointAfter(cast<Instruction>(V), Builder.GetInsertBlock()); 1531 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr, 1532 &*NewInsertPt); 1533 return V; 1534 } 1535 1536 // {X,+,F} --> X + {0,+,F} 1537 if (!S->getStart()->isZero()) { 1538 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1539 NewOps[0] = SE.getConstant(Ty, 0); 1540 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, 1541 S->getNoWrapFlags(SCEV::FlagNW)); 1542 1543 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1544 // comments on expandAddToGEP for details. 1545 const SCEV *Base = S->getStart(); 1546 // Dig into the expression to find the pointer base for a GEP. 1547 const SCEV *ExposedRest = Rest; 1548 ExposePointerBase(Base, ExposedRest, SE); 1549 // If we found a pointer, expand the AddRec with a GEP. 1550 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1551 // Make sure the Base isn't something exotic, such as a multiplied 1552 // or divided pointer value. In those cases, the result type isn't 1553 // actually a pointer type. 1554 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1555 Value *StartV = expand(Base); 1556 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1557 return expandAddToGEP(ExposedRest, PTy, Ty, StartV); 1558 } 1559 } 1560 1561 // Just do a normal add. Pre-expand the operands to suppress folding. 1562 // 1563 // The LHS and RHS values are factored out of the expand call to make the 1564 // output independent of the argument evaluation order. 1565 const SCEV *AddExprLHS = SE.getUnknown(expand(S->getStart())); 1566 const SCEV *AddExprRHS = SE.getUnknown(expand(Rest)); 1567 return expand(SE.getAddExpr(AddExprLHS, AddExprRHS)); 1568 } 1569 1570 // If we don't yet have a canonical IV, create one. 1571 if (!CanonicalIV) { 1572 // Create and insert the PHI node for the induction variable in the 1573 // specified loop. 1574 BasicBlock *Header = L->getHeader(); 1575 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1576 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 1577 &Header->front()); 1578 rememberInstruction(CanonicalIV); 1579 1580 SmallSet<BasicBlock *, 4> PredSeen; 1581 Constant *One = ConstantInt::get(Ty, 1); 1582 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1583 BasicBlock *HP = *HPI; 1584 if (!PredSeen.insert(HP).second) { 1585 // There must be an incoming value for each predecessor, even the 1586 // duplicates! 1587 CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP); 1588 continue; 1589 } 1590 1591 if (L->contains(HP)) { 1592 // Insert a unit add instruction right before the terminator 1593 // corresponding to the back-edge. 1594 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 1595 "indvar.next", 1596 HP->getTerminator()); 1597 Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 1598 rememberInstruction(Add); 1599 CanonicalIV->addIncoming(Add, HP); 1600 } else { 1601 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 1602 } 1603 } 1604 } 1605 1606 // {0,+,1} --> Insert a canonical induction variable into the loop! 1607 if (S->isAffine() && S->getOperand(1)->isOne()) { 1608 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 1609 "IVs with types different from the canonical IV should " 1610 "already have been handled!"); 1611 return CanonicalIV; 1612 } 1613 1614 // {0,+,F} --> {0,+,1} * F 1615 1616 // If this is a simple linear addrec, emit it now as a special case. 1617 if (S->isAffine()) // {0,+,F} --> i*F 1618 return 1619 expand(SE.getTruncateOrNoop( 1620 SE.getMulExpr(SE.getUnknown(CanonicalIV), 1621 SE.getNoopOrAnyExtend(S->getOperand(1), 1622 CanonicalIV->getType())), 1623 Ty)); 1624 1625 // If this is a chain of recurrences, turn it into a closed form, using the 1626 // folders, then expandCodeFor the closed form. This allows the folders to 1627 // simplify the expression without having to build a bunch of special code 1628 // into this folder. 1629 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 1630 1631 // Promote S up to the canonical IV type, if the cast is foldable. 1632 const SCEV *NewS = S; 1633 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 1634 if (isa<SCEVAddRecExpr>(Ext)) 1635 NewS = Ext; 1636 1637 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1638 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 1639 1640 // Truncate the result down to the original type, if needed. 1641 const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1642 return expand(T); 1643 } 1644 1645 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1646 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1647 Value *V = expandCodeFor(S->getOperand(), 1648 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1649 Value *I = Builder.CreateTrunc(V, Ty); 1650 rememberInstruction(I); 1651 return I; 1652 } 1653 1654 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1655 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1656 Value *V = expandCodeFor(S->getOperand(), 1657 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1658 Value *I = Builder.CreateZExt(V, Ty); 1659 rememberInstruction(I); 1660 return I; 1661 } 1662 1663 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1664 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1665 Value *V = expandCodeFor(S->getOperand(), 1666 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1667 Value *I = Builder.CreateSExt(V, Ty); 1668 rememberInstruction(I); 1669 return I; 1670 } 1671 1672 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 1673 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1674 Type *Ty = LHS->getType(); 1675 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1676 // In the case of mixed integer and pointer types, do the 1677 // rest of the comparisons as integer. 1678 Type *OpTy = S->getOperand(i)->getType(); 1679 if (OpTy->isIntegerTy() != Ty->isIntegerTy()) { 1680 Ty = SE.getEffectiveSCEVType(Ty); 1681 LHS = InsertNoopCastOfTo(LHS, Ty); 1682 } 1683 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1684 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 1685 rememberInstruction(ICmp); 1686 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1687 rememberInstruction(Sel); 1688 LHS = Sel; 1689 } 1690 // In the case of mixed integer and pointer types, cast the 1691 // final result back to the pointer type. 1692 if (LHS->getType() != S->getType()) 1693 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1694 return LHS; 1695 } 1696 1697 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 1698 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1699 Type *Ty = LHS->getType(); 1700 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1701 // In the case of mixed integer and pointer types, do the 1702 // rest of the comparisons as integer. 1703 Type *OpTy = S->getOperand(i)->getType(); 1704 if (OpTy->isIntegerTy() != Ty->isIntegerTy()) { 1705 Ty = SE.getEffectiveSCEVType(Ty); 1706 LHS = InsertNoopCastOfTo(LHS, Ty); 1707 } 1708 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1709 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 1710 rememberInstruction(ICmp); 1711 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1712 rememberInstruction(Sel); 1713 LHS = Sel; 1714 } 1715 // In the case of mixed integer and pointer types, cast the 1716 // final result back to the pointer type. 1717 if (LHS->getType() != S->getType()) 1718 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1719 return LHS; 1720 } 1721 1722 Value *SCEVExpander::visitSMinExpr(const SCEVSMinExpr *S) { 1723 Value *LHS = expand(S->getOperand(S->getNumOperands() - 1)); 1724 Type *Ty = LHS->getType(); 1725 for (int i = S->getNumOperands() - 2; i >= 0; --i) { 1726 // In the case of mixed integer and pointer types, do the 1727 // rest of the comparisons as integer. 1728 Type *OpTy = S->getOperand(i)->getType(); 1729 if (OpTy->isIntegerTy() != Ty->isIntegerTy()) { 1730 Ty = SE.getEffectiveSCEVType(Ty); 1731 LHS = InsertNoopCastOfTo(LHS, Ty); 1732 } 1733 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1734 Value *ICmp = Builder.CreateICmpSLT(LHS, RHS); 1735 rememberInstruction(ICmp); 1736 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smin"); 1737 rememberInstruction(Sel); 1738 LHS = Sel; 1739 } 1740 // In the case of mixed integer and pointer types, cast the 1741 // final result back to the pointer type. 1742 if (LHS->getType() != S->getType()) 1743 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1744 return LHS; 1745 } 1746 1747 Value *SCEVExpander::visitUMinExpr(const SCEVUMinExpr *S) { 1748 Value *LHS = expand(S->getOperand(S->getNumOperands() - 1)); 1749 Type *Ty = LHS->getType(); 1750 for (int i = S->getNumOperands() - 2; i >= 0; --i) { 1751 // In the case of mixed integer and pointer types, do the 1752 // rest of the comparisons as integer. 1753 Type *OpTy = S->getOperand(i)->getType(); 1754 if (OpTy->isIntegerTy() != Ty->isIntegerTy()) { 1755 Ty = SE.getEffectiveSCEVType(Ty); 1756 LHS = InsertNoopCastOfTo(LHS, Ty); 1757 } 1758 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1759 Value *ICmp = Builder.CreateICmpULT(LHS, RHS); 1760 rememberInstruction(ICmp); 1761 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umin"); 1762 rememberInstruction(Sel); 1763 LHS = Sel; 1764 } 1765 // In the case of mixed integer and pointer types, cast the 1766 // final result back to the pointer type. 1767 if (LHS->getType() != S->getType()) 1768 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1769 return LHS; 1770 } 1771 1772 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 1773 Instruction *IP) { 1774 setInsertPoint(IP); 1775 return expandCodeFor(SH, Ty); 1776 } 1777 1778 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 1779 // Expand the code for this SCEV. 1780 Value *V = expand(SH); 1781 if (Ty) { 1782 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 1783 "non-trivial casts should be done with the SCEVs directly!"); 1784 V = InsertNoopCastOfTo(V, Ty); 1785 } 1786 return V; 1787 } 1788 1789 ScalarEvolution::ValueOffsetPair 1790 SCEVExpander::FindValueInExprValueMap(const SCEV *S, 1791 const Instruction *InsertPt) { 1792 SetVector<ScalarEvolution::ValueOffsetPair> *Set = SE.getSCEVValues(S); 1793 // If the expansion is not in CanonicalMode, and the SCEV contains any 1794 // sub scAddRecExpr type SCEV, it is required to expand the SCEV literally. 1795 if (CanonicalMode || !SE.containsAddRecurrence(S)) { 1796 // If S is scConstant, it may be worse to reuse an existing Value. 1797 if (S->getSCEVType() != scConstant && Set) { 1798 // Choose a Value from the set which dominates the insertPt. 1799 // insertPt should be inside the Value's parent loop so as not to break 1800 // the LCSSA form. 1801 for (auto const &VOPair : *Set) { 1802 Value *V = VOPair.first; 1803 ConstantInt *Offset = VOPair.second; 1804 Instruction *EntInst = nullptr; 1805 if (V && isa<Instruction>(V) && (EntInst = cast<Instruction>(V)) && 1806 S->getType() == V->getType() && 1807 EntInst->getFunction() == InsertPt->getFunction() && 1808 SE.DT.dominates(EntInst, InsertPt) && 1809 (SE.LI.getLoopFor(EntInst->getParent()) == nullptr || 1810 SE.LI.getLoopFor(EntInst->getParent())->contains(InsertPt))) 1811 return {V, Offset}; 1812 } 1813 } 1814 } 1815 return {nullptr, nullptr}; 1816 } 1817 1818 // The expansion of SCEV will either reuse a previous Value in ExprValueMap, 1819 // or expand the SCEV literally. Specifically, if the expansion is in LSRMode, 1820 // and the SCEV contains any sub scAddRecExpr type SCEV, it will be expanded 1821 // literally, to prevent LSR's transformed SCEV from being reverted. Otherwise, 1822 // the expansion will try to reuse Value from ExprValueMap, and only when it 1823 // fails, expand the SCEV literally. 1824 Value *SCEVExpander::expand(const SCEV *S) { 1825 // Compute an insertion point for this SCEV object. Hoist the instructions 1826 // as far out in the loop nest as possible. 1827 Instruction *InsertPt = &*Builder.GetInsertPoint(); 1828 1829 // We can move insertion point only if there is no div or rem operations 1830 // otherwise we are risky to move it over the check for zero denominator. 1831 auto SafeToHoist = [](const SCEV *S) { 1832 return !SCEVExprContains(S, [](const SCEV *S) { 1833 if (const auto *D = dyn_cast<SCEVUDivExpr>(S)) { 1834 if (const auto *SC = dyn_cast<SCEVConstant>(D->getRHS())) 1835 // Division by non-zero constants can be hoisted. 1836 return SC->getValue()->isZero(); 1837 // All other divisions should not be moved as they may be 1838 // divisions by zero and should be kept within the 1839 // conditions of the surrounding loops that guard their 1840 // execution (see PR35406). 1841 return true; 1842 } 1843 return false; 1844 }); 1845 }; 1846 if (SafeToHoist(S)) { 1847 for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());; 1848 L = L->getParentLoop()) { 1849 if (SE.isLoopInvariant(S, L)) { 1850 if (!L) break; 1851 if (BasicBlock *Preheader = L->getLoopPreheader()) 1852 InsertPt = Preheader->getTerminator(); 1853 else 1854 // LSR sets the insertion point for AddRec start/step values to the 1855 // block start to simplify value reuse, even though it's an invalid 1856 // position. SCEVExpander must correct for this in all cases. 1857 InsertPt = &*L->getHeader()->getFirstInsertionPt(); 1858 } else { 1859 // If the SCEV is computable at this level, insert it into the header 1860 // after the PHIs (and after any other instructions that we've inserted 1861 // there) so that it is guaranteed to dominate any user inside the loop. 1862 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 1863 InsertPt = &*L->getHeader()->getFirstInsertionPt(); 1864 while (InsertPt->getIterator() != Builder.GetInsertPoint() && 1865 (isInsertedInstruction(InsertPt) || 1866 isa<DbgInfoIntrinsic>(InsertPt))) 1867 InsertPt = &*std::next(InsertPt->getIterator()); 1868 break; 1869 } 1870 } 1871 } 1872 1873 // IndVarSimplify sometimes sets the insertion point at the block start, even 1874 // when there are PHIs at that point. We must correct for this. 1875 if (isa<PHINode>(*InsertPt)) 1876 InsertPt = &*InsertPt->getParent()->getFirstInsertionPt(); 1877 1878 // Check to see if we already expanded this here. 1879 auto I = InsertedExpressions.find(std::make_pair(S, InsertPt)); 1880 if (I != InsertedExpressions.end()) 1881 return I->second; 1882 1883 SCEVInsertPointGuard Guard(Builder, this); 1884 Builder.SetInsertPoint(InsertPt); 1885 1886 // Expand the expression into instructions. 1887 ScalarEvolution::ValueOffsetPair VO = FindValueInExprValueMap(S, InsertPt); 1888 Value *V = VO.first; 1889 1890 if (!V) 1891 V = visit(S); 1892 else if (VO.second) { 1893 if (PointerType *Vty = dyn_cast<PointerType>(V->getType())) { 1894 Type *Ety = Vty->getPointerElementType(); 1895 int64_t Offset = VO.second->getSExtValue(); 1896 int64_t ESize = SE.getTypeSizeInBits(Ety); 1897 if ((Offset * 8) % ESize == 0) { 1898 ConstantInt *Idx = 1899 ConstantInt::getSigned(VO.second->getType(), -(Offset * 8) / ESize); 1900 V = Builder.CreateGEP(Ety, V, Idx, "scevgep"); 1901 } else { 1902 ConstantInt *Idx = 1903 ConstantInt::getSigned(VO.second->getType(), -Offset); 1904 unsigned AS = Vty->getAddressSpace(); 1905 V = Builder.CreateBitCast(V, Type::getInt8PtrTy(SE.getContext(), AS)); 1906 V = Builder.CreateGEP(Type::getInt8Ty(SE.getContext()), V, Idx, 1907 "uglygep"); 1908 V = Builder.CreateBitCast(V, Vty); 1909 } 1910 } else { 1911 V = Builder.CreateSub(V, VO.second); 1912 } 1913 } 1914 // Remember the expanded value for this SCEV at this location. 1915 // 1916 // This is independent of PostIncLoops. The mapped value simply materializes 1917 // the expression at this insertion point. If the mapped value happened to be 1918 // a postinc expansion, it could be reused by a non-postinc user, but only if 1919 // its insertion point was already at the head of the loop. 1920 InsertedExpressions[std::make_pair(S, InsertPt)] = V; 1921 return V; 1922 } 1923 1924 void SCEVExpander::rememberInstruction(Value *I) { 1925 if (!PostIncLoops.empty()) 1926 InsertedPostIncValues.insert(I); 1927 else 1928 InsertedValues.insert(I); 1929 } 1930 1931 /// getOrInsertCanonicalInductionVariable - This method returns the 1932 /// canonical induction variable of the specified type for the specified 1933 /// loop (inserting one if there is none). A canonical induction variable 1934 /// starts at zero and steps by one on each iteration. 1935 PHINode * 1936 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1937 Type *Ty) { 1938 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1939 1940 // Build a SCEV for {0,+,1}<L>. 1941 // Conservatively use FlagAnyWrap for now. 1942 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 1943 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 1944 1945 // Emit code for it. 1946 SCEVInsertPointGuard Guard(Builder, this); 1947 PHINode *V = 1948 cast<PHINode>(expandCodeFor(H, nullptr, &L->getHeader()->front())); 1949 1950 return V; 1951 } 1952 1953 /// replaceCongruentIVs - Check for congruent phis in this loop header and 1954 /// replace them with their most canonical representative. Return the number of 1955 /// phis eliminated. 1956 /// 1957 /// This does not depend on any SCEVExpander state but should be used in 1958 /// the same context that SCEVExpander is used. 1959 unsigned 1960 SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 1961 SmallVectorImpl<WeakTrackingVH> &DeadInsts, 1962 const TargetTransformInfo *TTI) { 1963 // Find integer phis in order of increasing width. 1964 SmallVector<PHINode*, 8> Phis; 1965 for (PHINode &PN : L->getHeader()->phis()) 1966 Phis.push_back(&PN); 1967 1968 if (TTI) 1969 llvm::sort(Phis, [](Value *LHS, Value *RHS) { 1970 // Put pointers at the back and make sure pointer < pointer = false. 1971 if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy()) 1972 return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy(); 1973 return RHS->getType()->getPrimitiveSizeInBits() < 1974 LHS->getType()->getPrimitiveSizeInBits(); 1975 }); 1976 1977 unsigned NumElim = 0; 1978 DenseMap<const SCEV *, PHINode *> ExprToIVMap; 1979 // Process phis from wide to narrow. Map wide phis to their truncation 1980 // so narrow phis can reuse them. 1981 for (PHINode *Phi : Phis) { 1982 auto SimplifyPHINode = [&](PHINode *PN) -> Value * { 1983 if (Value *V = SimplifyInstruction(PN, {DL, &SE.TLI, &SE.DT, &SE.AC})) 1984 return V; 1985 if (!SE.isSCEVable(PN->getType())) 1986 return nullptr; 1987 auto *Const = dyn_cast<SCEVConstant>(SE.getSCEV(PN)); 1988 if (!Const) 1989 return nullptr; 1990 return Const->getValue(); 1991 }; 1992 1993 // Fold constant phis. They may be congruent to other constant phis and 1994 // would confuse the logic below that expects proper IVs. 1995 if (Value *V = SimplifyPHINode(Phi)) { 1996 if (V->getType() != Phi->getType()) 1997 continue; 1998 Phi->replaceAllUsesWith(V); 1999 DeadInsts.emplace_back(Phi); 2000 ++NumElim; 2001 DEBUG_WITH_TYPE(DebugType, dbgs() 2002 << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 2003 continue; 2004 } 2005 2006 if (!SE.isSCEVable(Phi->getType())) 2007 continue; 2008 2009 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 2010 if (!OrigPhiRef) { 2011 OrigPhiRef = Phi; 2012 if (Phi->getType()->isIntegerTy() && TTI && 2013 TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 2014 // This phi can be freely truncated to the narrowest phi type. Map the 2015 // truncated expression to it so it will be reused for narrow types. 2016 const SCEV *TruncExpr = 2017 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 2018 ExprToIVMap[TruncExpr] = Phi; 2019 } 2020 continue; 2021 } 2022 2023 // Replacing a pointer phi with an integer phi or vice-versa doesn't make 2024 // sense. 2025 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 2026 continue; 2027 2028 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 2029 Instruction *OrigInc = dyn_cast<Instruction>( 2030 OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 2031 Instruction *IsomorphicInc = 2032 dyn_cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 2033 2034 if (OrigInc && IsomorphicInc) { 2035 // If this phi has the same width but is more canonical, replace the 2036 // original with it. As part of the "more canonical" determination, 2037 // respect a prior decision to use an IV chain. 2038 if (OrigPhiRef->getType() == Phi->getType() && 2039 !(ChainedPhis.count(Phi) || 2040 isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) && 2041 (ChainedPhis.count(Phi) || 2042 isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 2043 std::swap(OrigPhiRef, Phi); 2044 std::swap(OrigInc, IsomorphicInc); 2045 } 2046 // Replacing the congruent phi is sufficient because acyclic 2047 // redundancy elimination, CSE/GVN, should handle the 2048 // rest. However, once SCEV proves that a phi is congruent, 2049 // it's often the head of an IV user cycle that is isomorphic 2050 // with the original phi. It's worth eagerly cleaning up the 2051 // common case of a single IV increment so that DeleteDeadPHIs 2052 // can remove cycles that had postinc uses. 2053 const SCEV *TruncExpr = 2054 SE.getTruncateOrNoop(SE.getSCEV(OrigInc), IsomorphicInc->getType()); 2055 if (OrigInc != IsomorphicInc && 2056 TruncExpr == SE.getSCEV(IsomorphicInc) && 2057 SE.LI.replacementPreservesLCSSAForm(IsomorphicInc, OrigInc) && 2058 hoistIVInc(OrigInc, IsomorphicInc)) { 2059 DEBUG_WITH_TYPE(DebugType, 2060 dbgs() << "INDVARS: Eliminated congruent iv.inc: " 2061 << *IsomorphicInc << '\n'); 2062 Value *NewInc = OrigInc; 2063 if (OrigInc->getType() != IsomorphicInc->getType()) { 2064 Instruction *IP = nullptr; 2065 if (PHINode *PN = dyn_cast<PHINode>(OrigInc)) 2066 IP = &*PN->getParent()->getFirstInsertionPt(); 2067 else 2068 IP = OrigInc->getNextNode(); 2069 2070 IRBuilder<> Builder(IP); 2071 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); 2072 NewInc = Builder.CreateTruncOrBitCast( 2073 OrigInc, IsomorphicInc->getType(), IVName); 2074 } 2075 IsomorphicInc->replaceAllUsesWith(NewInc); 2076 DeadInsts.emplace_back(IsomorphicInc); 2077 } 2078 } 2079 } 2080 DEBUG_WITH_TYPE(DebugType, dbgs() << "INDVARS: Eliminated congruent iv: " 2081 << *Phi << '\n'); 2082 ++NumElim; 2083 Value *NewIV = OrigPhiRef; 2084 if (OrigPhiRef->getType() != Phi->getType()) { 2085 IRBuilder<> Builder(&*L->getHeader()->getFirstInsertionPt()); 2086 Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); 2087 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName); 2088 } 2089 Phi->replaceAllUsesWith(NewIV); 2090 DeadInsts.emplace_back(Phi); 2091 } 2092 return NumElim; 2093 } 2094 2095 Value *SCEVExpander::getExactExistingExpansion(const SCEV *S, 2096 const Instruction *At, Loop *L) { 2097 Optional<ScalarEvolution::ValueOffsetPair> VO = 2098 getRelatedExistingExpansion(S, At, L); 2099 if (VO && VO.getValue().second == nullptr) 2100 return VO.getValue().first; 2101 return nullptr; 2102 } 2103 2104 Optional<ScalarEvolution::ValueOffsetPair> 2105 SCEVExpander::getRelatedExistingExpansion(const SCEV *S, const Instruction *At, 2106 Loop *L) { 2107 using namespace llvm::PatternMatch; 2108 2109 SmallVector<BasicBlock *, 4> ExitingBlocks; 2110 L->getExitingBlocks(ExitingBlocks); 2111 2112 // Look for suitable value in simple conditions at the loop exits. 2113 for (BasicBlock *BB : ExitingBlocks) { 2114 ICmpInst::Predicate Pred; 2115 Instruction *LHS, *RHS; 2116 2117 if (!match(BB->getTerminator(), 2118 m_Br(m_ICmp(Pred, m_Instruction(LHS), m_Instruction(RHS)), 2119 m_BasicBlock(), m_BasicBlock()))) 2120 continue; 2121 2122 if (SE.getSCEV(LHS) == S && SE.DT.dominates(LHS, At)) 2123 return ScalarEvolution::ValueOffsetPair(LHS, nullptr); 2124 2125 if (SE.getSCEV(RHS) == S && SE.DT.dominates(RHS, At)) 2126 return ScalarEvolution::ValueOffsetPair(RHS, nullptr); 2127 } 2128 2129 // Use expand's logic which is used for reusing a previous Value in 2130 // ExprValueMap. 2131 ScalarEvolution::ValueOffsetPair VO = FindValueInExprValueMap(S, At); 2132 if (VO.first) 2133 return VO; 2134 2135 // There is potential to make this significantly smarter, but this simple 2136 // heuristic already gets some interesting cases. 2137 2138 // Can not find suitable value. 2139 return None; 2140 } 2141 2142 bool SCEVExpander::isHighCostExpansionHelper( 2143 const SCEV *S, Loop *L, const Instruction &At, int &BudgetRemaining, 2144 const TargetTransformInfo &TTI, SmallPtrSetImpl<const SCEV *> &Processed, 2145 SmallVectorImpl<const SCEV *> &Worklist) { 2146 if (BudgetRemaining < 0) 2147 return true; // Already run out of budget, give up. 2148 2149 // Was the cost of expansion of this expression already accounted for? 2150 if (!Processed.insert(S).second) 2151 return false; // We have already accounted for this expression. 2152 2153 // If we can find an existing value for this scev available at the point "At" 2154 // then consider the expression cheap. 2155 if (getRelatedExistingExpansion(S, &At, L)) 2156 return false; // Consider the expression to be free. 2157 2158 switch (S->getSCEVType()) { 2159 case scUnknown: 2160 case scConstant: 2161 return false; // Assume to be zero-cost. 2162 } 2163 2164 TargetTransformInfo::TargetCostKind CostKind = 2165 TargetTransformInfo::TCK_RecipThroughput; 2166 2167 if (auto *CastExpr = dyn_cast<SCEVCastExpr>(S)) { 2168 unsigned Opcode; 2169 switch (S->getSCEVType()) { 2170 case scTruncate: 2171 Opcode = Instruction::Trunc; 2172 break; 2173 case scZeroExtend: 2174 Opcode = Instruction::ZExt; 2175 break; 2176 case scSignExtend: 2177 Opcode = Instruction::SExt; 2178 break; 2179 default: 2180 llvm_unreachable("There are no other cast types."); 2181 } 2182 const SCEV *Op = CastExpr->getOperand(); 2183 BudgetRemaining -= TTI.getCastInstrCost(Opcode, /*Dst=*/S->getType(), 2184 /*Src=*/Op->getType(), CostKind); 2185 Worklist.emplace_back(Op); 2186 return false; // Will answer upon next entry into this function. 2187 } 2188 2189 if (auto *UDivExpr = dyn_cast<SCEVUDivExpr>(S)) { 2190 // If the divisor is a power of two count this as a logical right-shift. 2191 if (auto *SC = dyn_cast<SCEVConstant>(UDivExpr->getRHS())) { 2192 if (SC->getAPInt().isPowerOf2()) { 2193 BudgetRemaining -= 2194 TTI.getArithmeticInstrCost(Instruction::LShr, S->getType(), 2195 CostKind); 2196 // Note that we don't count the cost of RHS, because it is a constant, 2197 // and we consider those to be free. But if that changes, we would need 2198 // to log2() it first before calling isHighCostExpansionHelper(). 2199 Worklist.emplace_back(UDivExpr->getLHS()); 2200 return false; // Will answer upon next entry into this function. 2201 } 2202 } 2203 2204 // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or 2205 // HowManyLessThans produced to compute a precise expression, rather than a 2206 // UDiv from the user's code. If we can't find a UDiv in the code with some 2207 // simple searching, we need to account for it's cost. 2208 2209 // At the beginning of this function we already tried to find existing 2210 // value for plain 'S'. Now try to lookup 'S + 1' since it is common 2211 // pattern involving division. This is just a simple search heuristic. 2212 if (getRelatedExistingExpansion( 2213 SE.getAddExpr(S, SE.getConstant(S->getType(), 1)), &At, L)) 2214 return false; // Consider it to be free. 2215 2216 // Need to count the cost of this UDiv. 2217 BudgetRemaining -= 2218 TTI.getArithmeticInstrCost(Instruction::UDiv, S->getType(), 2219 CostKind); 2220 Worklist.insert(Worklist.end(), {UDivExpr->getLHS(), UDivExpr->getRHS()}); 2221 return false; // Will answer upon next entry into this function. 2222 } 2223 2224 if (const auto *NAry = dyn_cast<SCEVAddRecExpr>(S)) { 2225 Type *OpType = NAry->getType(); 2226 2227 assert(NAry->getNumOperands() >= 2 && 2228 "Polynomial should be at least linear"); 2229 2230 int AddCost = 2231 TTI.getArithmeticInstrCost(Instruction::Add, OpType, CostKind); 2232 int MulCost = 2233 TTI.getArithmeticInstrCost(Instruction::Mul, OpType, CostKind); 2234 2235 // In this polynominal, we may have some zero operands, and we shouldn't 2236 // really charge for those. So how many non-zero coeffients are there? 2237 int NumTerms = llvm::count_if(NAry->operands(), 2238 [](const SCEV *S) { return !S->isZero(); }); 2239 assert(NumTerms >= 1 && "Polynominal should have at least one term."); 2240 assert(!(*std::prev(NAry->operands().end()))->isZero() && 2241 "Last operand should not be zero"); 2242 2243 // Much like with normal add expr, the polynominal will require 2244 // one less addition than the number of it's terms. 2245 BudgetRemaining -= AddCost * (NumTerms - 1); 2246 if (BudgetRemaining < 0) 2247 return true; 2248 2249 // Ignoring constant term (operand 0), how many of the coeffients are u> 1? 2250 int NumNonZeroDegreeNonOneTerms = 2251 llvm::count_if(make_range(std::next(NAry->op_begin()), NAry->op_end()), 2252 [](const SCEV *S) { 2253 auto *SConst = dyn_cast<SCEVConstant>(S); 2254 return !SConst || SConst->getAPInt().ugt(1); 2255 }); 2256 // Here, *each* one of those will require a multiplication. 2257 BudgetRemaining -= MulCost * NumNonZeroDegreeNonOneTerms; 2258 if (BudgetRemaining < 0) 2259 return true; 2260 2261 // What is the degree of this polynominal? 2262 int PolyDegree = NAry->getNumOperands() - 1; 2263 assert(PolyDegree >= 1 && "Should be at least affine."); 2264 2265 // The final term will be: 2266 // Op_{PolyDegree} * x ^ {PolyDegree} 2267 // Where x ^ {PolyDegree} will again require PolyDegree-1 mul operations. 2268 // Note that x ^ {PolyDegree} = x * x ^ {PolyDegree-1} so charging for 2269 // x ^ {PolyDegree} will give us x ^ {2} .. x ^ {PolyDegree-1} for free. 2270 // FIXME: this is conservatively correct, but might be overly pessimistic. 2271 BudgetRemaining -= MulCost * (PolyDegree - 1); 2272 if (BudgetRemaining < 0) 2273 return true; 2274 2275 // And finally, the operands themselves should fit within the budget. 2276 Worklist.insert(Worklist.end(), NAry->operands().begin(), 2277 NAry->operands().end()); 2278 return false; // So far so good, though ops may be too costly? 2279 } 2280 2281 if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(S)) { 2282 Type *OpType = NAry->getType(); 2283 2284 int PairCost; 2285 switch (S->getSCEVType()) { 2286 case scAddExpr: 2287 PairCost = 2288 TTI.getArithmeticInstrCost(Instruction::Add, OpType, CostKind); 2289 break; 2290 case scMulExpr: 2291 // TODO: this is a very pessimistic cost modelling for Mul, 2292 // because of Bin Pow algorithm actually used by the expander, 2293 // see SCEVExpander::visitMulExpr(), ExpandOpBinPowN(). 2294 PairCost = 2295 TTI.getArithmeticInstrCost(Instruction::Mul, OpType, CostKind); 2296 break; 2297 case scSMaxExpr: 2298 case scUMaxExpr: 2299 case scSMinExpr: 2300 case scUMinExpr: 2301 PairCost = TTI.getCmpSelInstrCost(Instruction::ICmp, OpType, 2302 CmpInst::makeCmpResultType(OpType), 2303 CostKind) + 2304 TTI.getCmpSelInstrCost(Instruction::Select, OpType, 2305 CmpInst::makeCmpResultType(OpType), 2306 CostKind); 2307 break; 2308 default: 2309 llvm_unreachable("There are no other variants here."); 2310 } 2311 2312 assert(NAry->getNumOperands() > 1 && 2313 "Nary expr should have more than 1 operand."); 2314 // The simple nary expr will require one less op (or pair of ops) 2315 // than the number of it's terms. 2316 BudgetRemaining -= PairCost * (NAry->getNumOperands() - 1); 2317 if (BudgetRemaining < 0) 2318 return true; 2319 2320 // And finally, the operands themselves should fit within the budget. 2321 Worklist.insert(Worklist.end(), NAry->operands().begin(), 2322 NAry->operands().end()); 2323 return false; // So far so good, though ops may be too costly? 2324 } 2325 2326 llvm_unreachable("No other scev expressions possible."); 2327 } 2328 2329 Value *SCEVExpander::expandCodeForPredicate(const SCEVPredicate *Pred, 2330 Instruction *IP) { 2331 assert(IP); 2332 switch (Pred->getKind()) { 2333 case SCEVPredicate::P_Union: 2334 return expandUnionPredicate(cast<SCEVUnionPredicate>(Pred), IP); 2335 case SCEVPredicate::P_Equal: 2336 return expandEqualPredicate(cast<SCEVEqualPredicate>(Pred), IP); 2337 case SCEVPredicate::P_Wrap: { 2338 auto *AddRecPred = cast<SCEVWrapPredicate>(Pred); 2339 return expandWrapPredicate(AddRecPred, IP); 2340 } 2341 } 2342 llvm_unreachable("Unknown SCEV predicate type"); 2343 } 2344 2345 Value *SCEVExpander::expandEqualPredicate(const SCEVEqualPredicate *Pred, 2346 Instruction *IP) { 2347 Value *Expr0 = expandCodeFor(Pred->getLHS(), Pred->getLHS()->getType(), IP); 2348 Value *Expr1 = expandCodeFor(Pred->getRHS(), Pred->getRHS()->getType(), IP); 2349 2350 Builder.SetInsertPoint(IP); 2351 auto *I = Builder.CreateICmpNE(Expr0, Expr1, "ident.check"); 2352 return I; 2353 } 2354 2355 Value *SCEVExpander::generateOverflowCheck(const SCEVAddRecExpr *AR, 2356 Instruction *Loc, bool Signed) { 2357 assert(AR->isAffine() && "Cannot generate RT check for " 2358 "non-affine expression"); 2359 2360 SCEVUnionPredicate Pred; 2361 const SCEV *ExitCount = 2362 SE.getPredicatedBackedgeTakenCount(AR->getLoop(), Pred); 2363 2364 assert(ExitCount != SE.getCouldNotCompute() && "Invalid loop count"); 2365 2366 const SCEV *Step = AR->getStepRecurrence(SE); 2367 const SCEV *Start = AR->getStart(); 2368 2369 Type *ARTy = AR->getType(); 2370 unsigned SrcBits = SE.getTypeSizeInBits(ExitCount->getType()); 2371 unsigned DstBits = SE.getTypeSizeInBits(ARTy); 2372 2373 // The expression {Start,+,Step} has nusw/nssw if 2374 // Step < 0, Start - |Step| * Backedge <= Start 2375 // Step >= 0, Start + |Step| * Backedge > Start 2376 // and |Step| * Backedge doesn't unsigned overflow. 2377 2378 IntegerType *CountTy = IntegerType::get(Loc->getContext(), SrcBits); 2379 Builder.SetInsertPoint(Loc); 2380 Value *TripCountVal = expandCodeFor(ExitCount, CountTy, Loc); 2381 2382 IntegerType *Ty = 2383 IntegerType::get(Loc->getContext(), SE.getTypeSizeInBits(ARTy)); 2384 Type *ARExpandTy = DL.isNonIntegralPointerType(ARTy) ? ARTy : Ty; 2385 2386 Value *StepValue = expandCodeFor(Step, Ty, Loc); 2387 Value *NegStepValue = expandCodeFor(SE.getNegativeSCEV(Step), Ty, Loc); 2388 Value *StartValue = expandCodeFor(Start, ARExpandTy, Loc); 2389 2390 ConstantInt *Zero = 2391 ConstantInt::get(Loc->getContext(), APInt::getNullValue(DstBits)); 2392 2393 Builder.SetInsertPoint(Loc); 2394 // Compute |Step| 2395 Value *StepCompare = Builder.CreateICmp(ICmpInst::ICMP_SLT, StepValue, Zero); 2396 Value *AbsStep = Builder.CreateSelect(StepCompare, NegStepValue, StepValue); 2397 2398 // Get the backedge taken count and truncate or extended to the AR type. 2399 Value *TruncTripCount = Builder.CreateZExtOrTrunc(TripCountVal, Ty); 2400 auto *MulF = Intrinsic::getDeclaration(Loc->getModule(), 2401 Intrinsic::umul_with_overflow, Ty); 2402 2403 // Compute |Step| * Backedge 2404 CallInst *Mul = Builder.CreateCall(MulF, {AbsStep, TruncTripCount}, "mul"); 2405 Value *MulV = Builder.CreateExtractValue(Mul, 0, "mul.result"); 2406 Value *OfMul = Builder.CreateExtractValue(Mul, 1, "mul.overflow"); 2407 2408 // Compute: 2409 // Start + |Step| * Backedge < Start 2410 // Start - |Step| * Backedge > Start 2411 Value *Add = nullptr, *Sub = nullptr; 2412 if (PointerType *ARPtrTy = dyn_cast<PointerType>(ARExpandTy)) { 2413 const SCEV *MulS = SE.getSCEV(MulV); 2414 const SCEV *NegMulS = SE.getNegativeSCEV(MulS); 2415 Add = Builder.CreateBitCast(expandAddToGEP(MulS, ARPtrTy, Ty, StartValue), 2416 ARPtrTy); 2417 Sub = Builder.CreateBitCast( 2418 expandAddToGEP(NegMulS, ARPtrTy, Ty, StartValue), ARPtrTy); 2419 } else { 2420 Add = Builder.CreateAdd(StartValue, MulV); 2421 Sub = Builder.CreateSub(StartValue, MulV); 2422 } 2423 2424 Value *EndCompareGT = Builder.CreateICmp( 2425 Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT, Sub, StartValue); 2426 2427 Value *EndCompareLT = Builder.CreateICmp( 2428 Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT, Add, StartValue); 2429 2430 // Select the answer based on the sign of Step. 2431 Value *EndCheck = 2432 Builder.CreateSelect(StepCompare, EndCompareGT, EndCompareLT); 2433 2434 // If the backedge taken count type is larger than the AR type, 2435 // check that we don't drop any bits by truncating it. If we are 2436 // dropping bits, then we have overflow (unless the step is zero). 2437 if (SE.getTypeSizeInBits(CountTy) > SE.getTypeSizeInBits(Ty)) { 2438 auto MaxVal = APInt::getMaxValue(DstBits).zext(SrcBits); 2439 auto *BackedgeCheck = 2440 Builder.CreateICmp(ICmpInst::ICMP_UGT, TripCountVal, 2441 ConstantInt::get(Loc->getContext(), MaxVal)); 2442 BackedgeCheck = Builder.CreateAnd( 2443 BackedgeCheck, Builder.CreateICmp(ICmpInst::ICMP_NE, StepValue, Zero)); 2444 2445 EndCheck = Builder.CreateOr(EndCheck, BackedgeCheck); 2446 } 2447 2448 EndCheck = Builder.CreateOr(EndCheck, OfMul); 2449 return EndCheck; 2450 } 2451 2452 Value *SCEVExpander::expandWrapPredicate(const SCEVWrapPredicate *Pred, 2453 Instruction *IP) { 2454 const auto *A = cast<SCEVAddRecExpr>(Pred->getExpr()); 2455 Value *NSSWCheck = nullptr, *NUSWCheck = nullptr; 2456 2457 // Add a check for NUSW 2458 if (Pred->getFlags() & SCEVWrapPredicate::IncrementNUSW) 2459 NUSWCheck = generateOverflowCheck(A, IP, false); 2460 2461 // Add a check for NSSW 2462 if (Pred->getFlags() & SCEVWrapPredicate::IncrementNSSW) 2463 NSSWCheck = generateOverflowCheck(A, IP, true); 2464 2465 if (NUSWCheck && NSSWCheck) 2466 return Builder.CreateOr(NUSWCheck, NSSWCheck); 2467 2468 if (NUSWCheck) 2469 return NUSWCheck; 2470 2471 if (NSSWCheck) 2472 return NSSWCheck; 2473 2474 return ConstantInt::getFalse(IP->getContext()); 2475 } 2476 2477 Value *SCEVExpander::expandUnionPredicate(const SCEVUnionPredicate *Union, 2478 Instruction *IP) { 2479 auto *BoolType = IntegerType::get(IP->getContext(), 1); 2480 Value *Check = ConstantInt::getNullValue(BoolType); 2481 2482 // Loop over all checks in this set. 2483 for (auto Pred : Union->getPredicates()) { 2484 auto *NextCheck = expandCodeForPredicate(Pred, IP); 2485 Builder.SetInsertPoint(IP); 2486 Check = Builder.CreateOr(Check, NextCheck); 2487 } 2488 2489 return Check; 2490 } 2491 2492 namespace { 2493 // Search for a SCEV subexpression that is not safe to expand. Any expression 2494 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely 2495 // UDiv expressions. We don't know if the UDiv is derived from an IR divide 2496 // instruction, but the important thing is that we prove the denominator is 2497 // nonzero before expansion. 2498 // 2499 // IVUsers already checks that IV-derived expressions are safe. So this check is 2500 // only needed when the expression includes some subexpression that is not IV 2501 // derived. 2502 // 2503 // Currently, we only allow division by a nonzero constant here. If this is 2504 // inadequate, we could easily allow division by SCEVUnknown by using 2505 // ValueTracking to check isKnownNonZero(). 2506 // 2507 // We cannot generally expand recurrences unless the step dominates the loop 2508 // header. The expander handles the special case of affine recurrences by 2509 // scaling the recurrence outside the loop, but this technique isn't generally 2510 // applicable. Expanding a nested recurrence outside a loop requires computing 2511 // binomial coefficients. This could be done, but the recurrence has to be in a 2512 // perfectly reduced form, which can't be guaranteed. 2513 struct SCEVFindUnsafe { 2514 ScalarEvolution &SE; 2515 bool IsUnsafe; 2516 2517 SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {} 2518 2519 bool follow(const SCEV *S) { 2520 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 2521 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS()); 2522 if (!SC || SC->getValue()->isZero()) { 2523 IsUnsafe = true; 2524 return false; 2525 } 2526 } 2527 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2528 const SCEV *Step = AR->getStepRecurrence(SE); 2529 if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) { 2530 IsUnsafe = true; 2531 return false; 2532 } 2533 } 2534 return true; 2535 } 2536 bool isDone() const { return IsUnsafe; } 2537 }; 2538 } 2539 2540 namespace llvm { 2541 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) { 2542 SCEVFindUnsafe Search(SE); 2543 visitAll(S, Search); 2544 return !Search.IsUnsafe; 2545 } 2546 2547 bool isSafeToExpandAt(const SCEV *S, const Instruction *InsertionPoint, 2548 ScalarEvolution &SE) { 2549 if (!isSafeToExpand(S, SE)) 2550 return false; 2551 // We have to prove that the expanded site of S dominates InsertionPoint. 2552 // This is easy when not in the same block, but hard when S is an instruction 2553 // to be expanded somewhere inside the same block as our insertion point. 2554 // What we really need here is something analogous to an OrderedBasicBlock, 2555 // but for the moment, we paper over the problem by handling two common and 2556 // cheap to check cases. 2557 if (SE.properlyDominates(S, InsertionPoint->getParent())) 2558 return true; 2559 if (SE.dominates(S, InsertionPoint->getParent())) { 2560 if (InsertionPoint->getParent()->getTerminator() == InsertionPoint) 2561 return true; 2562 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) 2563 for (const Value *V : InsertionPoint->operand_values()) 2564 if (V == U->getValue()) 2565 return true; 2566 } 2567 return false; 2568 } 2569 } 2570