1 //===- ScalarEvolutionExpander.cpp - Scalar Evolution Analysis --*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains the implementation of the scalar evolution expander, 11 // which is used to generate the code corresponding to a given scalar evolution 12 // expression. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #include "llvm/Analysis/ScalarEvolutionExpander.h" 17 #include "llvm/ADT/STLExtras.h" 18 #include "llvm/ADT/SmallSet.h" 19 #include "llvm/Analysis/InstructionSimplify.h" 20 #include "llvm/Analysis/LoopInfo.h" 21 #include "llvm/Analysis/TargetTransformInfo.h" 22 #include "llvm/IR/DataLayout.h" 23 #include "llvm/IR/Dominators.h" 24 #include "llvm/IR/IntrinsicInst.h" 25 #include "llvm/IR/LLVMContext.h" 26 #include "llvm/IR/Module.h" 27 #include "llvm/Support/Debug.h" 28 #include "llvm/Support/raw_ostream.h" 29 30 using namespace llvm; 31 32 /// ReuseOrCreateCast - Arrange for there to be a cast of V to Ty at IP, 33 /// reusing an existing cast if a suitable one exists, moving an existing 34 /// cast if a suitable one exists but isn't in the right place, or 35 /// creating a new one. 36 Value *SCEVExpander::ReuseOrCreateCast(Value *V, Type *Ty, 37 Instruction::CastOps Op, 38 BasicBlock::iterator IP) { 39 // This function must be called with the builder having a valid insertion 40 // point. It doesn't need to be the actual IP where the uses of the returned 41 // cast will be added, but it must dominate such IP. 42 // We use this precondition to produce a cast that will dominate all its 43 // uses. In particular, this is crucial for the case where the builder's 44 // insertion point *is* the point where we were asked to put the cast. 45 // Since we don't know the builder's insertion point is actually 46 // where the uses will be added (only that it dominates it), we are 47 // not allowed to move it. 48 BasicBlock::iterator BIP = Builder.GetInsertPoint(); 49 50 Instruction *Ret = nullptr; 51 52 // Check to see if there is already a cast! 53 for (User *U : V->users()) 54 if (U->getType() == Ty) 55 if (CastInst *CI = dyn_cast<CastInst>(U)) 56 if (CI->getOpcode() == Op) { 57 // If the cast isn't where we want it, create a new cast at IP. 58 // Likewise, do not reuse a cast at BIP because it must dominate 59 // instructions that might be inserted before BIP. 60 if (BasicBlock::iterator(CI) != IP || BIP == IP) { 61 // Create a new cast, and leave the old cast in place in case 62 // it is being used as an insert point. Clear its operand 63 // so that it doesn't hold anything live. 64 Ret = CastInst::Create(Op, V, Ty, "", IP); 65 Ret->takeName(CI); 66 CI->replaceAllUsesWith(Ret); 67 CI->setOperand(0, UndefValue::get(V->getType())); 68 break; 69 } 70 Ret = CI; 71 break; 72 } 73 74 // Create a new cast. 75 if (!Ret) 76 Ret = CastInst::Create(Op, V, Ty, V->getName(), IP); 77 78 // We assert at the end of the function since IP might point to an 79 // instruction with different dominance properties than a cast 80 // (an invoke for example) and not dominate BIP (but the cast does). 81 assert(SE.DT->dominates(Ret, BIP)); 82 83 rememberInstruction(Ret); 84 return Ret; 85 } 86 87 /// InsertNoopCastOfTo - Insert a cast of V to the specified type, 88 /// which must be possible with a noop cast, doing what we can to share 89 /// the casts. 90 Value *SCEVExpander::InsertNoopCastOfTo(Value *V, Type *Ty) { 91 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 92 assert((Op == Instruction::BitCast || 93 Op == Instruction::PtrToInt || 94 Op == Instruction::IntToPtr) && 95 "InsertNoopCastOfTo cannot perform non-noop casts!"); 96 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 97 "InsertNoopCastOfTo cannot change sizes!"); 98 99 // Short-circuit unnecessary bitcasts. 100 if (Op == Instruction::BitCast) { 101 if (V->getType() == Ty) 102 return V; 103 if (CastInst *CI = dyn_cast<CastInst>(V)) { 104 if (CI->getOperand(0)->getType() == Ty) 105 return CI->getOperand(0); 106 } 107 } 108 // Short-circuit unnecessary inttoptr<->ptrtoint casts. 109 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 110 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 111 if (CastInst *CI = dyn_cast<CastInst>(V)) 112 if ((CI->getOpcode() == Instruction::PtrToInt || 113 CI->getOpcode() == Instruction::IntToPtr) && 114 SE.getTypeSizeInBits(CI->getType()) == 115 SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 116 return CI->getOperand(0); 117 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 118 if ((CE->getOpcode() == Instruction::PtrToInt || 119 CE->getOpcode() == Instruction::IntToPtr) && 120 SE.getTypeSizeInBits(CE->getType()) == 121 SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 122 return CE->getOperand(0); 123 } 124 125 // Fold a cast of a constant. 126 if (Constant *C = dyn_cast<Constant>(V)) 127 return ConstantExpr::getCast(Op, C, Ty); 128 129 // Cast the argument at the beginning of the entry block, after 130 // any bitcasts of other arguments. 131 if (Argument *A = dyn_cast<Argument>(V)) { 132 BasicBlock::iterator IP = A->getParent()->getEntryBlock().begin(); 133 while ((isa<BitCastInst>(IP) && 134 isa<Argument>(cast<BitCastInst>(IP)->getOperand(0)) && 135 cast<BitCastInst>(IP)->getOperand(0) != A) || 136 isa<DbgInfoIntrinsic>(IP) || 137 isa<LandingPadInst>(IP)) 138 ++IP; 139 return ReuseOrCreateCast(A, Ty, Op, IP); 140 } 141 142 // Cast the instruction immediately after the instruction. 143 Instruction *I = cast<Instruction>(V); 144 BasicBlock::iterator IP = I; ++IP; 145 if (InvokeInst *II = dyn_cast<InvokeInst>(I)) 146 IP = II->getNormalDest()->begin(); 147 while (isa<PHINode>(IP) || isa<LandingPadInst>(IP)) 148 ++IP; 149 return ReuseOrCreateCast(I, Ty, Op, IP); 150 } 151 152 /// InsertBinop - Insert the specified binary operator, doing a small amount 153 /// of work to avoid inserting an obviously redundant operation. 154 Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 155 Value *LHS, Value *RHS) { 156 // Fold a binop with constant operands. 157 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 158 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 159 return ConstantExpr::get(Opcode, CLHS, CRHS); 160 161 // Do a quick scan to see if we have this binop nearby. If so, reuse it. 162 unsigned ScanLimit = 6; 163 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 164 // Scanning starts from the last instruction before the insertion point. 165 BasicBlock::iterator IP = Builder.GetInsertPoint(); 166 if (IP != BlockBegin) { 167 --IP; 168 for (; ScanLimit; --IP, --ScanLimit) { 169 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 170 // generated code. 171 if (isa<DbgInfoIntrinsic>(IP)) 172 ScanLimit++; 173 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 174 IP->getOperand(1) == RHS) 175 return IP; 176 if (IP == BlockBegin) break; 177 } 178 } 179 180 // Save the original insertion point so we can restore it when we're done. 181 DebugLoc Loc = Builder.GetInsertPoint()->getDebugLoc(); 182 BuilderType::InsertPointGuard Guard(Builder); 183 184 // Move the insertion point out of as many loops as we can. 185 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 186 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 187 BasicBlock *Preheader = L->getLoopPreheader(); 188 if (!Preheader) break; 189 190 // Ok, move up a level. 191 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 192 } 193 194 // If we haven't found this binop, insert it. 195 Instruction *BO = cast<Instruction>(Builder.CreateBinOp(Opcode, LHS, RHS)); 196 BO->setDebugLoc(Loc); 197 rememberInstruction(BO); 198 199 return BO; 200 } 201 202 /// FactorOutConstant - Test if S is divisible by Factor, using signed 203 /// division. If so, update S with Factor divided out and return true. 204 /// S need not be evenly divisible if a reasonable remainder can be 205 /// computed. 206 /// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 207 /// unnecessary; in its place, just signed-divide Ops[i] by the scale and 208 /// check to see if the divide was folded. 209 static bool FactorOutConstant(const SCEV *&S, const SCEV *&Remainder, 210 const SCEV *Factor, ScalarEvolution &SE, 211 const DataLayout &DL) { 212 // Everything is divisible by one. 213 if (Factor->isOne()) 214 return true; 215 216 // x/x == 1. 217 if (S == Factor) { 218 S = SE.getConstant(S->getType(), 1); 219 return true; 220 } 221 222 // For a Constant, check for a multiple of the given factor. 223 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 224 // 0/x == 0. 225 if (C->isZero()) 226 return true; 227 // Check for divisibility. 228 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 229 ConstantInt *CI = 230 ConstantInt::get(SE.getContext(), 231 C->getValue()->getValue().sdiv( 232 FC->getValue()->getValue())); 233 // If the quotient is zero and the remainder is non-zero, reject 234 // the value at this scale. It will be considered for subsequent 235 // smaller scales. 236 if (!CI->isZero()) { 237 const SCEV *Div = SE.getConstant(CI); 238 S = Div; 239 Remainder = 240 SE.getAddExpr(Remainder, 241 SE.getConstant(C->getValue()->getValue().srem( 242 FC->getValue()->getValue()))); 243 return true; 244 } 245 } 246 } 247 248 // In a Mul, check if there is a constant operand which is a multiple 249 // of the given factor. 250 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 251 // Size is known, check if there is a constant operand which is a multiple 252 // of the given factor. If so, we can factor it. 253 const SCEVConstant *FC = cast<SCEVConstant>(Factor); 254 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 255 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) { 256 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 257 NewMulOps[0] = SE.getConstant( 258 C->getValue()->getValue().sdiv(FC->getValue()->getValue())); 259 S = SE.getMulExpr(NewMulOps); 260 return true; 261 } 262 } 263 264 // In an AddRec, check if both start and step are divisible. 265 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 266 const SCEV *Step = A->getStepRecurrence(SE); 267 const SCEV *StepRem = SE.getConstant(Step->getType(), 0); 268 if (!FactorOutConstant(Step, StepRem, Factor, SE, DL)) 269 return false; 270 if (!StepRem->isZero()) 271 return false; 272 const SCEV *Start = A->getStart(); 273 if (!FactorOutConstant(Start, Remainder, Factor, SE, DL)) 274 return false; 275 S = SE.getAddRecExpr(Start, Step, A->getLoop(), 276 A->getNoWrapFlags(SCEV::FlagNW)); 277 return true; 278 } 279 280 return false; 281 } 282 283 /// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 284 /// is the number of SCEVAddRecExprs present, which are kept at the end of 285 /// the list. 286 /// 287 static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 288 Type *Ty, 289 ScalarEvolution &SE) { 290 unsigned NumAddRecs = 0; 291 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 292 ++NumAddRecs; 293 // Group Ops into non-addrecs and addrecs. 294 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 295 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 296 // Let ScalarEvolution sort and simplify the non-addrecs list. 297 const SCEV *Sum = NoAddRecs.empty() ? 298 SE.getConstant(Ty, 0) : 299 SE.getAddExpr(NoAddRecs); 300 // If it returned an add, use the operands. Otherwise it simplified 301 // the sum into a single value, so just use that. 302 Ops.clear(); 303 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 304 Ops.append(Add->op_begin(), Add->op_end()); 305 else if (!Sum->isZero()) 306 Ops.push_back(Sum); 307 // Then append the addrecs. 308 Ops.append(AddRecs.begin(), AddRecs.end()); 309 } 310 311 /// SplitAddRecs - Flatten a list of add operands, moving addrec start values 312 /// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 313 /// This helps expose more opportunities for folding parts of the expressions 314 /// into GEP indices. 315 /// 316 static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 317 Type *Ty, 318 ScalarEvolution &SE) { 319 // Find the addrecs. 320 SmallVector<const SCEV *, 8> AddRecs; 321 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 322 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 323 const SCEV *Start = A->getStart(); 324 if (Start->isZero()) break; 325 const SCEV *Zero = SE.getConstant(Ty, 0); 326 AddRecs.push_back(SE.getAddRecExpr(Zero, 327 A->getStepRecurrence(SE), 328 A->getLoop(), 329 A->getNoWrapFlags(SCEV::FlagNW))); 330 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 331 Ops[i] = Zero; 332 Ops.append(Add->op_begin(), Add->op_end()); 333 e += Add->getNumOperands(); 334 } else { 335 Ops[i] = Start; 336 } 337 } 338 if (!AddRecs.empty()) { 339 // Add the addrecs onto the end of the list. 340 Ops.append(AddRecs.begin(), AddRecs.end()); 341 // Resort the operand list, moving any constants to the front. 342 SimplifyAddOperands(Ops, Ty, SE); 343 } 344 } 345 346 /// expandAddToGEP - Expand an addition expression with a pointer type into 347 /// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 348 /// BasicAliasAnalysis and other passes analyze the result. See the rules 349 /// for getelementptr vs. inttoptr in 350 /// http://llvm.org/docs/LangRef.html#pointeraliasing 351 /// for details. 352 /// 353 /// Design note: The correctness of using getelementptr here depends on 354 /// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 355 /// they may introduce pointer arithmetic which may not be safely converted 356 /// into getelementptr. 357 /// 358 /// Design note: It might seem desirable for this function to be more 359 /// loop-aware. If some of the indices are loop-invariant while others 360 /// aren't, it might seem desirable to emit multiple GEPs, keeping the 361 /// loop-invariant portions of the overall computation outside the loop. 362 /// However, there are a few reasons this is not done here. Hoisting simple 363 /// arithmetic is a low-level optimization that often isn't very 364 /// important until late in the optimization process. In fact, passes 365 /// like InstructionCombining will combine GEPs, even if it means 366 /// pushing loop-invariant computation down into loops, so even if the 367 /// GEPs were split here, the work would quickly be undone. The 368 /// LoopStrengthReduction pass, which is usually run quite late (and 369 /// after the last InstructionCombining pass), takes care of hoisting 370 /// loop-invariant portions of expressions, after considering what 371 /// can be folded using target addressing modes. 372 /// 373 Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 374 const SCEV *const *op_end, 375 PointerType *PTy, 376 Type *Ty, 377 Value *V) { 378 Type *OriginalElTy = PTy->getElementType(); 379 Type *ElTy = OriginalElTy; 380 SmallVector<Value *, 4> GepIndices; 381 SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 382 bool AnyNonZeroIndices = false; 383 384 // Split AddRecs up into parts as either of the parts may be usable 385 // without the other. 386 SplitAddRecs(Ops, Ty, SE); 387 388 Type *IntPtrTy = DL.getIntPtrType(PTy); 389 390 // Descend down the pointer's type and attempt to convert the other 391 // operands into GEP indices, at each level. The first index in a GEP 392 // indexes into the array implied by the pointer operand; the rest of 393 // the indices index into the element or field type selected by the 394 // preceding index. 395 for (;;) { 396 // If the scale size is not 0, attempt to factor out a scale for 397 // array indexing. 398 SmallVector<const SCEV *, 8> ScaledOps; 399 if (ElTy->isSized()) { 400 const SCEV *ElSize = SE.getSizeOfExpr(IntPtrTy, ElTy); 401 if (!ElSize->isZero()) { 402 SmallVector<const SCEV *, 8> NewOps; 403 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 404 const SCEV *Op = Ops[i]; 405 const SCEV *Remainder = SE.getConstant(Ty, 0); 406 if (FactorOutConstant(Op, Remainder, ElSize, SE, DL)) { 407 // Op now has ElSize factored out. 408 ScaledOps.push_back(Op); 409 if (!Remainder->isZero()) 410 NewOps.push_back(Remainder); 411 AnyNonZeroIndices = true; 412 } else { 413 // The operand was not divisible, so add it to the list of operands 414 // we'll scan next iteration. 415 NewOps.push_back(Ops[i]); 416 } 417 } 418 // If we made any changes, update Ops. 419 if (!ScaledOps.empty()) { 420 Ops = NewOps; 421 SimplifyAddOperands(Ops, Ty, SE); 422 } 423 } 424 } 425 426 // Record the scaled array index for this level of the type. If 427 // we didn't find any operands that could be factored, tentatively 428 // assume that element zero was selected (since the zero offset 429 // would obviously be folded away). 430 Value *Scaled = ScaledOps.empty() ? 431 Constant::getNullValue(Ty) : 432 expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 433 GepIndices.push_back(Scaled); 434 435 // Collect struct field index operands. 436 while (StructType *STy = dyn_cast<StructType>(ElTy)) { 437 bool FoundFieldNo = false; 438 // An empty struct has no fields. 439 if (STy->getNumElements() == 0) break; 440 // Field offsets are known. See if a constant offset falls within any of 441 // the struct fields. 442 if (Ops.empty()) 443 break; 444 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 445 if (SE.getTypeSizeInBits(C->getType()) <= 64) { 446 const StructLayout &SL = *DL.getStructLayout(STy); 447 uint64_t FullOffset = C->getValue()->getZExtValue(); 448 if (FullOffset < SL.getSizeInBytes()) { 449 unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 450 GepIndices.push_back( 451 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 452 ElTy = STy->getTypeAtIndex(ElIdx); 453 Ops[0] = 454 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 455 AnyNonZeroIndices = true; 456 FoundFieldNo = true; 457 } 458 } 459 // If no struct field offsets were found, tentatively assume that 460 // field zero was selected (since the zero offset would obviously 461 // be folded away). 462 if (!FoundFieldNo) { 463 ElTy = STy->getTypeAtIndex(0u); 464 GepIndices.push_back( 465 Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 466 } 467 } 468 469 if (ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 470 ElTy = ATy->getElementType(); 471 else 472 break; 473 } 474 475 // If none of the operands were convertible to proper GEP indices, cast 476 // the base to i8* and do an ugly getelementptr with that. It's still 477 // better than ptrtoint+arithmetic+inttoptr at least. 478 if (!AnyNonZeroIndices) { 479 // Cast the base to i8*. 480 V = InsertNoopCastOfTo(V, 481 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 482 483 assert(!isa<Instruction>(V) || 484 SE.DT->dominates(cast<Instruction>(V), Builder.GetInsertPoint())); 485 486 // Expand the operands for a plain byte offset. 487 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 488 489 // Fold a GEP with constant operands. 490 if (Constant *CLHS = dyn_cast<Constant>(V)) 491 if (Constant *CRHS = dyn_cast<Constant>(Idx)) 492 return ConstantExpr::getGetElementPtr(Type::getInt8Ty(Ty->getContext()), 493 CLHS, CRHS); 494 495 // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 496 unsigned ScanLimit = 6; 497 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 498 // Scanning starts from the last instruction before the insertion point. 499 BasicBlock::iterator IP = Builder.GetInsertPoint(); 500 if (IP != BlockBegin) { 501 --IP; 502 for (; ScanLimit; --IP, --ScanLimit) { 503 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 504 // generated code. 505 if (isa<DbgInfoIntrinsic>(IP)) 506 ScanLimit++; 507 if (IP->getOpcode() == Instruction::GetElementPtr && 508 IP->getOperand(0) == V && IP->getOperand(1) == Idx) 509 return IP; 510 if (IP == BlockBegin) break; 511 } 512 } 513 514 // Save the original insertion point so we can restore it when we're done. 515 BuilderType::InsertPointGuard Guard(Builder); 516 517 // Move the insertion point out of as many loops as we can. 518 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 519 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 520 BasicBlock *Preheader = L->getLoopPreheader(); 521 if (!Preheader) break; 522 523 // Ok, move up a level. 524 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 525 } 526 527 // Emit a GEP. 528 Value *GEP = Builder.CreateGEP(Builder.getInt8Ty(), V, Idx, "uglygep"); 529 rememberInstruction(GEP); 530 531 return GEP; 532 } 533 534 // Save the original insertion point so we can restore it when we're done. 535 BuilderType::InsertPoint SaveInsertPt = Builder.saveIP(); 536 537 // Move the insertion point out of as many loops as we can. 538 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 539 if (!L->isLoopInvariant(V)) break; 540 541 bool AnyIndexNotLoopInvariant = false; 542 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(), 543 E = GepIndices.end(); I != E; ++I) 544 if (!L->isLoopInvariant(*I)) { 545 AnyIndexNotLoopInvariant = true; 546 break; 547 } 548 if (AnyIndexNotLoopInvariant) 549 break; 550 551 BasicBlock *Preheader = L->getLoopPreheader(); 552 if (!Preheader) break; 553 554 // Ok, move up a level. 555 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 556 } 557 558 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 559 // because ScalarEvolution may have changed the address arithmetic to 560 // compute a value which is beyond the end of the allocated object. 561 Value *Casted = V; 562 if (V->getType() != PTy) 563 Casted = InsertNoopCastOfTo(Casted, PTy); 564 Value *GEP = Builder.CreateGEP(OriginalElTy, Casted, 565 GepIndices, 566 "scevgep"); 567 Ops.push_back(SE.getUnknown(GEP)); 568 rememberInstruction(GEP); 569 570 // Restore the original insert point. 571 Builder.restoreIP(SaveInsertPt); 572 573 return expand(SE.getAddExpr(Ops)); 574 } 575 576 /// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 577 /// SCEV expansion. If they are nested, this is the most nested. If they are 578 /// neighboring, pick the later. 579 static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 580 DominatorTree &DT) { 581 if (!A) return B; 582 if (!B) return A; 583 if (A->contains(B)) return B; 584 if (B->contains(A)) return A; 585 if (DT.dominates(A->getHeader(), B->getHeader())) return B; 586 if (DT.dominates(B->getHeader(), A->getHeader())) return A; 587 return A; // Arbitrarily break the tie. 588 } 589 590 /// getRelevantLoop - Get the most relevant loop associated with the given 591 /// expression, according to PickMostRelevantLoop. 592 const Loop *SCEVExpander::getRelevantLoop(const SCEV *S) { 593 // Test whether we've already computed the most relevant loop for this SCEV. 594 std::pair<DenseMap<const SCEV *, const Loop *>::iterator, bool> Pair = 595 RelevantLoops.insert(std::make_pair(S, nullptr)); 596 if (!Pair.second) 597 return Pair.first->second; 598 599 if (isa<SCEVConstant>(S)) 600 // A constant has no relevant loops. 601 return nullptr; 602 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 603 if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 604 return Pair.first->second = SE.LI->getLoopFor(I->getParent()); 605 // A non-instruction has no relevant loops. 606 return nullptr; 607 } 608 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 609 const Loop *L = nullptr; 610 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 611 L = AR->getLoop(); 612 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end(); 613 I != E; ++I) 614 L = PickMostRelevantLoop(L, getRelevantLoop(*I), *SE.DT); 615 return RelevantLoops[N] = L; 616 } 617 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) { 618 const Loop *Result = getRelevantLoop(C->getOperand()); 619 return RelevantLoops[C] = Result; 620 } 621 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 622 const Loop *Result = 623 PickMostRelevantLoop(getRelevantLoop(D->getLHS()), 624 getRelevantLoop(D->getRHS()), 625 *SE.DT); 626 return RelevantLoops[D] = Result; 627 } 628 llvm_unreachable("Unexpected SCEV type!"); 629 } 630 631 namespace { 632 633 /// LoopCompare - Compare loops by PickMostRelevantLoop. 634 class LoopCompare { 635 DominatorTree &DT; 636 public: 637 explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 638 639 bool operator()(std::pair<const Loop *, const SCEV *> LHS, 640 std::pair<const Loop *, const SCEV *> RHS) const { 641 // Keep pointer operands sorted at the end. 642 if (LHS.second->getType()->isPointerTy() != 643 RHS.second->getType()->isPointerTy()) 644 return LHS.second->getType()->isPointerTy(); 645 646 // Compare loops with PickMostRelevantLoop. 647 if (LHS.first != RHS.first) 648 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 649 650 // If one operand is a non-constant negative and the other is not, 651 // put the non-constant negative on the right so that a sub can 652 // be used instead of a negate and add. 653 if (LHS.second->isNonConstantNegative()) { 654 if (!RHS.second->isNonConstantNegative()) 655 return false; 656 } else if (RHS.second->isNonConstantNegative()) 657 return true; 658 659 // Otherwise they are equivalent according to this comparison. 660 return false; 661 } 662 }; 663 664 } 665 666 Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 667 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 668 669 // Collect all the add operands in a loop, along with their associated loops. 670 // Iterate in reverse so that constants are emitted last, all else equal, and 671 // so that pointer operands are inserted first, which the code below relies on 672 // to form more involved GEPs. 673 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 674 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 675 E(S->op_begin()); I != E; ++I) 676 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 677 678 // Sort by loop. Use a stable sort so that constants follow non-constants and 679 // pointer operands precede non-pointer operands. 680 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 681 682 // Emit instructions to add all the operands. Hoist as much as possible 683 // out of loops, and form meaningful getelementptrs where possible. 684 Value *Sum = nullptr; 685 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 686 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 687 const Loop *CurLoop = I->first; 688 const SCEV *Op = I->second; 689 if (!Sum) { 690 // This is the first operand. Just expand it. 691 Sum = expand(Op); 692 ++I; 693 } else if (PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 694 // The running sum expression is a pointer. Try to form a getelementptr 695 // at this level with that as the base. 696 SmallVector<const SCEV *, 4> NewOps; 697 for (; I != E && I->first == CurLoop; ++I) { 698 // If the operand is SCEVUnknown and not instructions, peek through 699 // it, to enable more of it to be folded into the GEP. 700 const SCEV *X = I->second; 701 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(X)) 702 if (!isa<Instruction>(U->getValue())) 703 X = SE.getSCEV(U->getValue()); 704 NewOps.push_back(X); 705 } 706 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 707 } else if (PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 708 // The running sum is an integer, and there's a pointer at this level. 709 // Try to form a getelementptr. If the running sum is instructions, 710 // use a SCEVUnknown to avoid re-analyzing them. 711 SmallVector<const SCEV *, 4> NewOps; 712 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 713 SE.getSCEV(Sum)); 714 for (++I; I != E && I->first == CurLoop; ++I) 715 NewOps.push_back(I->second); 716 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 717 } else if (Op->isNonConstantNegative()) { 718 // Instead of doing a negate and add, just do a subtract. 719 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 720 Sum = InsertNoopCastOfTo(Sum, Ty); 721 Sum = InsertBinop(Instruction::Sub, Sum, W); 722 ++I; 723 } else { 724 // A simple add. 725 Value *W = expandCodeFor(Op, Ty); 726 Sum = InsertNoopCastOfTo(Sum, Ty); 727 // Canonicalize a constant to the RHS. 728 if (isa<Constant>(Sum)) std::swap(Sum, W); 729 Sum = InsertBinop(Instruction::Add, Sum, W); 730 ++I; 731 } 732 } 733 734 return Sum; 735 } 736 737 Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 738 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 739 740 // Collect all the mul operands in a loop, along with their associated loops. 741 // Iterate in reverse so that constants are emitted last, all else equal. 742 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 743 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 744 E(S->op_begin()); I != E; ++I) 745 OpsAndLoops.push_back(std::make_pair(getRelevantLoop(*I), *I)); 746 747 // Sort by loop. Use a stable sort so that constants follow non-constants. 748 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 749 750 // Emit instructions to mul all the operands. Hoist as much as possible 751 // out of loops. 752 Value *Prod = nullptr; 753 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 754 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 755 const SCEV *Op = I->second; 756 if (!Prod) { 757 // This is the first operand. Just expand it. 758 Prod = expand(Op); 759 ++I; 760 } else if (Op->isAllOnesValue()) { 761 // Instead of doing a multiply by negative one, just do a negate. 762 Prod = InsertNoopCastOfTo(Prod, Ty); 763 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 764 ++I; 765 } else { 766 // A simple mul. 767 Value *W = expandCodeFor(Op, Ty); 768 Prod = InsertNoopCastOfTo(Prod, Ty); 769 // Canonicalize a constant to the RHS. 770 if (isa<Constant>(Prod)) std::swap(Prod, W); 771 Prod = InsertBinop(Instruction::Mul, Prod, W); 772 ++I; 773 } 774 } 775 776 return Prod; 777 } 778 779 Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 780 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 781 782 Value *LHS = expandCodeFor(S->getLHS(), Ty); 783 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 784 const APInt &RHS = SC->getValue()->getValue(); 785 if (RHS.isPowerOf2()) 786 return InsertBinop(Instruction::LShr, LHS, 787 ConstantInt::get(Ty, RHS.logBase2())); 788 } 789 790 Value *RHS = expandCodeFor(S->getRHS(), Ty); 791 return InsertBinop(Instruction::UDiv, LHS, RHS); 792 } 793 794 /// Move parts of Base into Rest to leave Base with the minimal 795 /// expression that provides a pointer operand suitable for a 796 /// GEP expansion. 797 static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 798 ScalarEvolution &SE) { 799 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 800 Base = A->getStart(); 801 Rest = SE.getAddExpr(Rest, 802 SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 803 A->getStepRecurrence(SE), 804 A->getLoop(), 805 A->getNoWrapFlags(SCEV::FlagNW))); 806 } 807 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 808 Base = A->getOperand(A->getNumOperands()-1); 809 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 810 NewAddOps.back() = Rest; 811 Rest = SE.getAddExpr(NewAddOps); 812 ExposePointerBase(Base, Rest, SE); 813 } 814 } 815 816 /// Determine if this is a well-behaved chain of instructions leading back to 817 /// the PHI. If so, it may be reused by expanded expressions. 818 bool SCEVExpander::isNormalAddRecExprPHI(PHINode *PN, Instruction *IncV, 819 const Loop *L) { 820 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV) || 821 (isa<CastInst>(IncV) && !isa<BitCastInst>(IncV))) 822 return false; 823 // If any of the operands don't dominate the insert position, bail. 824 // Addrec operands are always loop-invariant, so this can only happen 825 // if there are instructions which haven't been hoisted. 826 if (L == IVIncInsertLoop) { 827 for (User::op_iterator OI = IncV->op_begin()+1, 828 OE = IncV->op_end(); OI != OE; ++OI) 829 if (Instruction *OInst = dyn_cast<Instruction>(OI)) 830 if (!SE.DT->dominates(OInst, IVIncInsertPos)) 831 return false; 832 } 833 // Advance to the next instruction. 834 IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 835 if (!IncV) 836 return false; 837 838 if (IncV->mayHaveSideEffects()) 839 return false; 840 841 if (IncV != PN) 842 return true; 843 844 return isNormalAddRecExprPHI(PN, IncV, L); 845 } 846 847 /// getIVIncOperand returns an induction variable increment's induction 848 /// variable operand. 849 /// 850 /// If allowScale is set, any type of GEP is allowed as long as the nonIV 851 /// operands dominate InsertPos. 852 /// 853 /// If allowScale is not set, ensure that a GEP increment conforms to one of the 854 /// simple patterns generated by getAddRecExprPHILiterally and 855 /// expandAddtoGEP. If the pattern isn't recognized, return NULL. 856 Instruction *SCEVExpander::getIVIncOperand(Instruction *IncV, 857 Instruction *InsertPos, 858 bool allowScale) { 859 if (IncV == InsertPos) 860 return nullptr; 861 862 switch (IncV->getOpcode()) { 863 default: 864 return nullptr; 865 // Check for a simple Add/Sub or GEP of a loop invariant step. 866 case Instruction::Add: 867 case Instruction::Sub: { 868 Instruction *OInst = dyn_cast<Instruction>(IncV->getOperand(1)); 869 if (!OInst || SE.DT->dominates(OInst, InsertPos)) 870 return dyn_cast<Instruction>(IncV->getOperand(0)); 871 return nullptr; 872 } 873 case Instruction::BitCast: 874 return dyn_cast<Instruction>(IncV->getOperand(0)); 875 case Instruction::GetElementPtr: 876 for (Instruction::op_iterator I = IncV->op_begin()+1, E = IncV->op_end(); 877 I != E; ++I) { 878 if (isa<Constant>(*I)) 879 continue; 880 if (Instruction *OInst = dyn_cast<Instruction>(*I)) { 881 if (!SE.DT->dominates(OInst, InsertPos)) 882 return nullptr; 883 } 884 if (allowScale) { 885 // allow any kind of GEP as long as it can be hoisted. 886 continue; 887 } 888 // This must be a pointer addition of constants (pretty), which is already 889 // handled, or some number of address-size elements (ugly). Ugly geps 890 // have 2 operands. i1* is used by the expander to represent an 891 // address-size element. 892 if (IncV->getNumOperands() != 2) 893 return nullptr; 894 unsigned AS = cast<PointerType>(IncV->getType())->getAddressSpace(); 895 if (IncV->getType() != Type::getInt1PtrTy(SE.getContext(), AS) 896 && IncV->getType() != Type::getInt8PtrTy(SE.getContext(), AS)) 897 return nullptr; 898 break; 899 } 900 return dyn_cast<Instruction>(IncV->getOperand(0)); 901 } 902 } 903 904 /// hoistStep - Attempt to hoist a simple IV increment above InsertPos to make 905 /// it available to other uses in this loop. Recursively hoist any operands, 906 /// until we reach a value that dominates InsertPos. 907 bool SCEVExpander::hoistIVInc(Instruction *IncV, Instruction *InsertPos) { 908 if (SE.DT->dominates(IncV, InsertPos)) 909 return true; 910 911 // InsertPos must itself dominate IncV so that IncV's new position satisfies 912 // its existing users. 913 if (isa<PHINode>(InsertPos) 914 || !SE.DT->dominates(InsertPos->getParent(), IncV->getParent())) 915 return false; 916 917 // Check that the chain of IV operands leading back to Phi can be hoisted. 918 SmallVector<Instruction*, 4> IVIncs; 919 for(;;) { 920 Instruction *Oper = getIVIncOperand(IncV, InsertPos, /*allowScale*/true); 921 if (!Oper) 922 return false; 923 // IncV is safe to hoist. 924 IVIncs.push_back(IncV); 925 IncV = Oper; 926 if (SE.DT->dominates(IncV, InsertPos)) 927 break; 928 } 929 for (SmallVectorImpl<Instruction*>::reverse_iterator I = IVIncs.rbegin(), 930 E = IVIncs.rend(); I != E; ++I) { 931 (*I)->moveBefore(InsertPos); 932 } 933 return true; 934 } 935 936 /// Determine if this cyclic phi is in a form that would have been generated by 937 /// LSR. We don't care if the phi was actually expanded in this pass, as long 938 /// as it is in a low-cost form, for example, no implied multiplication. This 939 /// should match any patterns generated by getAddRecExprPHILiterally and 940 /// expandAddtoGEP. 941 bool SCEVExpander::isExpandedAddRecExprPHI(PHINode *PN, Instruction *IncV, 942 const Loop *L) { 943 for(Instruction *IVOper = IncV; 944 (IVOper = getIVIncOperand(IVOper, L->getLoopPreheader()->getTerminator(), 945 /*allowScale=*/false));) { 946 if (IVOper == PN) 947 return true; 948 } 949 return false; 950 } 951 952 /// expandIVInc - Expand an IV increment at Builder's current InsertPos. 953 /// Typically this is the LatchBlock terminator or IVIncInsertPos, but we may 954 /// need to materialize IV increments elsewhere to handle difficult situations. 955 Value *SCEVExpander::expandIVInc(PHINode *PN, Value *StepV, const Loop *L, 956 Type *ExpandTy, Type *IntTy, 957 bool useSubtract) { 958 Value *IncV; 959 // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 960 if (ExpandTy->isPointerTy()) { 961 PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 962 // If the step isn't constant, don't use an implicitly scaled GEP, because 963 // that would require a multiply inside the loop. 964 if (!isa<ConstantInt>(StepV)) 965 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 966 GEPPtrTy->getAddressSpace()); 967 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 968 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 969 if (IncV->getType() != PN->getType()) { 970 IncV = Builder.CreateBitCast(IncV, PN->getType()); 971 rememberInstruction(IncV); 972 } 973 } else { 974 IncV = useSubtract ? 975 Builder.CreateSub(PN, StepV, Twine(IVName) + ".iv.next") : 976 Builder.CreateAdd(PN, StepV, Twine(IVName) + ".iv.next"); 977 rememberInstruction(IncV); 978 } 979 return IncV; 980 } 981 982 /// \brief Hoist the addrec instruction chain rooted in the loop phi above the 983 /// position. This routine assumes that this is possible (has been checked). 984 static void hoistBeforePos(DominatorTree *DT, Instruction *InstToHoist, 985 Instruction *Pos, PHINode *LoopPhi) { 986 do { 987 if (DT->dominates(InstToHoist, Pos)) 988 break; 989 // Make sure the increment is where we want it. But don't move it 990 // down past a potential existing post-inc user. 991 InstToHoist->moveBefore(Pos); 992 Pos = InstToHoist; 993 InstToHoist = cast<Instruction>(InstToHoist->getOperand(0)); 994 } while (InstToHoist != LoopPhi); 995 } 996 997 /// \brief Check whether we can cheaply express the requested SCEV in terms of 998 /// the available PHI SCEV by truncation and/or invertion of the step. 999 static bool canBeCheaplyTransformed(ScalarEvolution &SE, 1000 const SCEVAddRecExpr *Phi, 1001 const SCEVAddRecExpr *Requested, 1002 bool &InvertStep) { 1003 Type *PhiTy = SE.getEffectiveSCEVType(Phi->getType()); 1004 Type *RequestedTy = SE.getEffectiveSCEVType(Requested->getType()); 1005 1006 if (RequestedTy->getIntegerBitWidth() > PhiTy->getIntegerBitWidth()) 1007 return false; 1008 1009 // Try truncate it if necessary. 1010 Phi = dyn_cast<SCEVAddRecExpr>(SE.getTruncateOrNoop(Phi, RequestedTy)); 1011 if (!Phi) 1012 return false; 1013 1014 // Check whether truncation will help. 1015 if (Phi == Requested) { 1016 InvertStep = false; 1017 return true; 1018 } 1019 1020 // Check whether inverting will help: {R,+,-1} == R - {0,+,1}. 1021 if (SE.getAddExpr(Requested->getStart(), 1022 SE.getNegativeSCEV(Requested)) == Phi) { 1023 InvertStep = true; 1024 return true; 1025 } 1026 1027 return false; 1028 } 1029 1030 static bool IsIncrementNSW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1031 if (!isa<IntegerType>(AR->getType())) 1032 return false; 1033 1034 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1035 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1036 const SCEV *Step = AR->getStepRecurrence(SE); 1037 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getSignExtendExpr(Step, WideTy), 1038 SE.getSignExtendExpr(AR, WideTy)); 1039 const SCEV *ExtendAfterOp = 1040 SE.getSignExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1041 return ExtendAfterOp == OpAfterExtend; 1042 } 1043 1044 static bool IsIncrementNUW(ScalarEvolution &SE, const SCEVAddRecExpr *AR) { 1045 if (!isa<IntegerType>(AR->getType())) 1046 return false; 1047 1048 unsigned BitWidth = cast<IntegerType>(AR->getType())->getBitWidth(); 1049 Type *WideTy = IntegerType::get(AR->getType()->getContext(), BitWidth * 2); 1050 const SCEV *Step = AR->getStepRecurrence(SE); 1051 const SCEV *OpAfterExtend = SE.getAddExpr(SE.getZeroExtendExpr(Step, WideTy), 1052 SE.getZeroExtendExpr(AR, WideTy)); 1053 const SCEV *ExtendAfterOp = 1054 SE.getZeroExtendExpr(SE.getAddExpr(AR, Step), WideTy); 1055 return ExtendAfterOp == OpAfterExtend; 1056 } 1057 1058 /// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 1059 /// the base addrec, which is the addrec without any non-loop-dominating 1060 /// values, and return the PHI. 1061 PHINode * 1062 SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 1063 const Loop *L, 1064 Type *ExpandTy, 1065 Type *IntTy, 1066 Type *&TruncTy, 1067 bool &InvertStep) { 1068 assert((!IVIncInsertLoop||IVIncInsertPos) && "Uninitialized insert position"); 1069 1070 // Reuse a previously-inserted PHI, if present. 1071 BasicBlock *LatchBlock = L->getLoopLatch(); 1072 if (LatchBlock) { 1073 PHINode *AddRecPhiMatch = nullptr; 1074 Instruction *IncV = nullptr; 1075 TruncTy = nullptr; 1076 InvertStep = false; 1077 1078 // Only try partially matching scevs that need truncation and/or 1079 // step-inversion if we know this loop is outside the current loop. 1080 bool TryNonMatchingSCEV = IVIncInsertLoop && 1081 SE.DT->properlyDominates(LatchBlock, IVIncInsertLoop->getHeader()); 1082 1083 for (BasicBlock::iterator I = L->getHeader()->begin(); 1084 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 1085 if (!SE.isSCEVable(PN->getType())) 1086 continue; 1087 1088 const SCEVAddRecExpr *PhiSCEV = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(PN)); 1089 if (!PhiSCEV) 1090 continue; 1091 1092 bool IsMatchingSCEV = PhiSCEV == Normalized; 1093 // We only handle truncation and inversion of phi recurrences for the 1094 // expanded expression if the expanded expression's loop dominates the 1095 // loop we insert to. Check now, so we can bail out early. 1096 if (!IsMatchingSCEV && !TryNonMatchingSCEV) 1097 continue; 1098 1099 Instruction *TempIncV = 1100 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 1101 1102 // Check whether we can reuse this PHI node. 1103 if (LSRMode) { 1104 if (!isExpandedAddRecExprPHI(PN, TempIncV, L)) 1105 continue; 1106 if (L == IVIncInsertLoop && !hoistIVInc(TempIncV, IVIncInsertPos)) 1107 continue; 1108 } else { 1109 if (!isNormalAddRecExprPHI(PN, TempIncV, L)) 1110 continue; 1111 } 1112 1113 // Stop if we have found an exact match SCEV. 1114 if (IsMatchingSCEV) { 1115 IncV = TempIncV; 1116 TruncTy = nullptr; 1117 InvertStep = false; 1118 AddRecPhiMatch = PN; 1119 break; 1120 } 1121 1122 // Try whether the phi can be translated into the requested form 1123 // (truncated and/or offset by a constant). 1124 if ((!TruncTy || InvertStep) && 1125 canBeCheaplyTransformed(SE, PhiSCEV, Normalized, InvertStep)) { 1126 // Record the phi node. But don't stop we might find an exact match 1127 // later. 1128 AddRecPhiMatch = PN; 1129 IncV = TempIncV; 1130 TruncTy = SE.getEffectiveSCEVType(Normalized->getType()); 1131 } 1132 } 1133 1134 if (AddRecPhiMatch) { 1135 // Potentially, move the increment. We have made sure in 1136 // isExpandedAddRecExprPHI or hoistIVInc that this is possible. 1137 if (L == IVIncInsertLoop) 1138 hoistBeforePos(SE.DT, IncV, IVIncInsertPos, AddRecPhiMatch); 1139 1140 // Ok, the add recurrence looks usable. 1141 // Remember this PHI, even in post-inc mode. 1142 InsertedValues.insert(AddRecPhiMatch); 1143 // Remember the increment. 1144 rememberInstruction(IncV); 1145 return AddRecPhiMatch; 1146 } 1147 } 1148 1149 // Save the original insertion point so we can restore it when we're done. 1150 BuilderType::InsertPointGuard Guard(Builder); 1151 1152 // Another AddRec may need to be recursively expanded below. For example, if 1153 // this AddRec is quadratic, the StepV may itself be an AddRec in this 1154 // loop. Remove this loop from the PostIncLoops set before expanding such 1155 // AddRecs. Otherwise, we cannot find a valid position for the step 1156 // (i.e. StepV can never dominate its loop header). Ideally, we could do 1157 // SavedIncLoops.swap(PostIncLoops), but we generally have a single element, 1158 // so it's not worth implementing SmallPtrSet::swap. 1159 PostIncLoopSet SavedPostIncLoops = PostIncLoops; 1160 PostIncLoops.clear(); 1161 1162 // Expand code for the start value. 1163 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 1164 L->getHeader()->begin()); 1165 1166 // StartV must be hoisted into L's preheader to dominate the new phi. 1167 assert(!isa<Instruction>(StartV) || 1168 SE.DT->properlyDominates(cast<Instruction>(StartV)->getParent(), 1169 L->getHeader())); 1170 1171 // Expand code for the step value. Do this before creating the PHI so that PHI 1172 // reuse code doesn't see an incomplete PHI. 1173 const SCEV *Step = Normalized->getStepRecurrence(SE); 1174 // If the stride is negative, insert a sub instead of an add for the increment 1175 // (unless it's a constant, because subtracts of constants are canonicalized 1176 // to adds). 1177 bool useSubtract = !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1178 if (useSubtract) 1179 Step = SE.getNegativeSCEV(Step); 1180 // Expand the step somewhere that dominates the loop header. 1181 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1182 1183 // The no-wrap behavior proved by IsIncrement(NUW|NSW) is only applicable if 1184 // we actually do emit an addition. It does not apply if we emit a 1185 // subtraction. 1186 bool IncrementIsNUW = !useSubtract && IsIncrementNUW(SE, Normalized); 1187 bool IncrementIsNSW = !useSubtract && IsIncrementNSW(SE, Normalized); 1188 1189 // Create the PHI. 1190 BasicBlock *Header = L->getHeader(); 1191 Builder.SetInsertPoint(Header, Header->begin()); 1192 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1193 PHINode *PN = Builder.CreatePHI(ExpandTy, std::distance(HPB, HPE), 1194 Twine(IVName) + ".iv"); 1195 rememberInstruction(PN); 1196 1197 // Create the step instructions and populate the PHI. 1198 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1199 BasicBlock *Pred = *HPI; 1200 1201 // Add a start value. 1202 if (!L->contains(Pred)) { 1203 PN->addIncoming(StartV, Pred); 1204 continue; 1205 } 1206 1207 // Create a step value and add it to the PHI. 1208 // If IVIncInsertLoop is non-null and equal to the addrec's loop, insert the 1209 // instructions at IVIncInsertPos. 1210 Instruction *InsertPos = L == IVIncInsertLoop ? 1211 IVIncInsertPos : Pred->getTerminator(); 1212 Builder.SetInsertPoint(InsertPos); 1213 Value *IncV = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1214 1215 if (isa<OverflowingBinaryOperator>(IncV)) { 1216 if (IncrementIsNUW) 1217 cast<BinaryOperator>(IncV)->setHasNoUnsignedWrap(); 1218 if (IncrementIsNSW) 1219 cast<BinaryOperator>(IncV)->setHasNoSignedWrap(); 1220 } 1221 PN->addIncoming(IncV, Pred); 1222 } 1223 1224 // After expanding subexpressions, restore the PostIncLoops set so the caller 1225 // can ensure that IVIncrement dominates the current uses. 1226 PostIncLoops = SavedPostIncLoops; 1227 1228 // Remember this PHI, even in post-inc mode. 1229 InsertedValues.insert(PN); 1230 1231 return PN; 1232 } 1233 1234 Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 1235 Type *STy = S->getType(); 1236 Type *IntTy = SE.getEffectiveSCEVType(STy); 1237 const Loop *L = S->getLoop(); 1238 1239 // Determine a normalized form of this expression, which is the expression 1240 // before any post-inc adjustment is made. 1241 const SCEVAddRecExpr *Normalized = S; 1242 if (PostIncLoops.count(L)) { 1243 PostIncLoopSet Loops; 1244 Loops.insert(L); 1245 Normalized = 1246 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, nullptr, 1247 nullptr, Loops, SE, *SE.DT)); 1248 } 1249 1250 // Strip off any non-loop-dominating component from the addrec start. 1251 const SCEV *Start = Normalized->getStart(); 1252 const SCEV *PostLoopOffset = nullptr; 1253 if (!SE.properlyDominates(Start, L->getHeader())) { 1254 PostLoopOffset = Start; 1255 Start = SE.getConstant(Normalized->getType(), 0); 1256 Normalized = cast<SCEVAddRecExpr>( 1257 SE.getAddRecExpr(Start, Normalized->getStepRecurrence(SE), 1258 Normalized->getLoop(), 1259 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1260 } 1261 1262 // Strip off any non-loop-dominating component from the addrec step. 1263 const SCEV *Step = Normalized->getStepRecurrence(SE); 1264 const SCEV *PostLoopScale = nullptr; 1265 if (!SE.dominates(Step, L->getHeader())) { 1266 PostLoopScale = Step; 1267 Step = SE.getConstant(Normalized->getType(), 1); 1268 Normalized = 1269 cast<SCEVAddRecExpr>(SE.getAddRecExpr( 1270 Start, Step, Normalized->getLoop(), 1271 Normalized->getNoWrapFlags(SCEV::FlagNW))); 1272 } 1273 1274 // Expand the core addrec. If we need post-loop scaling, force it to 1275 // expand to an integer type to avoid the need for additional casting. 1276 Type *ExpandTy = PostLoopScale ? IntTy : STy; 1277 // In some cases, we decide to reuse an existing phi node but need to truncate 1278 // it and/or invert the step. 1279 Type *TruncTy = nullptr; 1280 bool InvertStep = false; 1281 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy, 1282 TruncTy, InvertStep); 1283 1284 // Accommodate post-inc mode, if necessary. 1285 Value *Result; 1286 if (!PostIncLoops.count(L)) 1287 Result = PN; 1288 else { 1289 // In PostInc mode, use the post-incremented value. 1290 BasicBlock *LatchBlock = L->getLoopLatch(); 1291 assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1292 Result = PN->getIncomingValueForBlock(LatchBlock); 1293 1294 // For an expansion to use the postinc form, the client must call 1295 // expandCodeFor with an InsertPoint that is either outside the PostIncLoop 1296 // or dominated by IVIncInsertPos. 1297 if (isa<Instruction>(Result) 1298 && !SE.DT->dominates(cast<Instruction>(Result), 1299 Builder.GetInsertPoint())) { 1300 // The induction variable's postinc expansion does not dominate this use. 1301 // IVUsers tries to prevent this case, so it is rare. However, it can 1302 // happen when an IVUser outside the loop is not dominated by the latch 1303 // block. Adjusting IVIncInsertPos before expansion begins cannot handle 1304 // all cases. Consider a phi outide whose operand is replaced during 1305 // expansion with the value of the postinc user. Without fundamentally 1306 // changing the way postinc users are tracked, the only remedy is 1307 // inserting an extra IV increment. StepV might fold into PostLoopOffset, 1308 // but hopefully expandCodeFor handles that. 1309 bool useSubtract = 1310 !ExpandTy->isPointerTy() && Step->isNonConstantNegative(); 1311 if (useSubtract) 1312 Step = SE.getNegativeSCEV(Step); 1313 Value *StepV; 1314 { 1315 // Expand the step somewhere that dominates the loop header. 1316 BuilderType::InsertPointGuard Guard(Builder); 1317 StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 1318 } 1319 Result = expandIVInc(PN, StepV, L, ExpandTy, IntTy, useSubtract); 1320 } 1321 } 1322 1323 // We have decided to reuse an induction variable of a dominating loop. Apply 1324 // truncation and/or invertion of the step. 1325 if (TruncTy) { 1326 Type *ResTy = Result->getType(); 1327 // Normalize the result type. 1328 if (ResTy != SE.getEffectiveSCEVType(ResTy)) 1329 Result = InsertNoopCastOfTo(Result, SE.getEffectiveSCEVType(ResTy)); 1330 // Truncate the result. 1331 if (TruncTy != Result->getType()) { 1332 Result = Builder.CreateTrunc(Result, TruncTy); 1333 rememberInstruction(Result); 1334 } 1335 // Invert the result. 1336 if (InvertStep) { 1337 Result = Builder.CreateSub(expandCodeFor(Normalized->getStart(), TruncTy), 1338 Result); 1339 rememberInstruction(Result); 1340 } 1341 } 1342 1343 // Re-apply any non-loop-dominating scale. 1344 if (PostLoopScale) { 1345 assert(S->isAffine() && "Can't linearly scale non-affine recurrences."); 1346 Result = InsertNoopCastOfTo(Result, IntTy); 1347 Result = Builder.CreateMul(Result, 1348 expandCodeFor(PostLoopScale, IntTy)); 1349 rememberInstruction(Result); 1350 } 1351 1352 // Re-apply any non-loop-dominating offset. 1353 if (PostLoopOffset) { 1354 if (PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1355 const SCEV *const OffsetArray[1] = { PostLoopOffset }; 1356 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 1357 } else { 1358 Result = InsertNoopCastOfTo(Result, IntTy); 1359 Result = Builder.CreateAdd(Result, 1360 expandCodeFor(PostLoopOffset, IntTy)); 1361 rememberInstruction(Result); 1362 } 1363 } 1364 1365 return Result; 1366 } 1367 1368 Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1369 if (!CanonicalMode) return expandAddRecExprLiterally(S); 1370 1371 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1372 const Loop *L = S->getLoop(); 1373 1374 // First check for an existing canonical IV in a suitable type. 1375 PHINode *CanonicalIV = nullptr; 1376 if (PHINode *PN = L->getCanonicalInductionVariable()) 1377 if (SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 1378 CanonicalIV = PN; 1379 1380 // Rewrite an AddRec in terms of the canonical induction variable, if 1381 // its type is more narrow. 1382 if (CanonicalIV && 1383 SE.getTypeSizeInBits(CanonicalIV->getType()) > 1384 SE.getTypeSizeInBits(Ty)) { 1385 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1386 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1387 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 1388 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop(), 1389 S->getNoWrapFlags(SCEV::FlagNW))); 1390 BasicBlock::iterator NewInsertPt = 1391 std::next(BasicBlock::iterator(cast<Instruction>(V))); 1392 BuilderType::InsertPointGuard Guard(Builder); 1393 while (isa<PHINode>(NewInsertPt) || isa<DbgInfoIntrinsic>(NewInsertPt) || 1394 isa<LandingPadInst>(NewInsertPt)) 1395 ++NewInsertPt; 1396 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), nullptr, 1397 NewInsertPt); 1398 return V; 1399 } 1400 1401 // {X,+,F} --> X + {0,+,F} 1402 if (!S->getStart()->isZero()) { 1403 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1404 NewOps[0] = SE.getConstant(Ty, 0); 1405 const SCEV *Rest = SE.getAddRecExpr(NewOps, L, 1406 S->getNoWrapFlags(SCEV::FlagNW)); 1407 1408 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1409 // comments on expandAddToGEP for details. 1410 const SCEV *Base = S->getStart(); 1411 const SCEV *RestArray[1] = { Rest }; 1412 // Dig into the expression to find the pointer base for a GEP. 1413 ExposePointerBase(Base, RestArray[0], SE); 1414 // If we found a pointer, expand the AddRec with a GEP. 1415 if (PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1416 // Make sure the Base isn't something exotic, such as a multiplied 1417 // or divided pointer value. In those cases, the result type isn't 1418 // actually a pointer type. 1419 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1420 Value *StartV = expand(Base); 1421 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1422 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 1423 } 1424 } 1425 1426 // Just do a normal add. Pre-expand the operands to suppress folding. 1427 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), 1428 SE.getUnknown(expand(Rest)))); 1429 } 1430 1431 // If we don't yet have a canonical IV, create one. 1432 if (!CanonicalIV) { 1433 // Create and insert the PHI node for the induction variable in the 1434 // specified loop. 1435 BasicBlock *Header = L->getHeader(); 1436 pred_iterator HPB = pred_begin(Header), HPE = pred_end(Header); 1437 CanonicalIV = PHINode::Create(Ty, std::distance(HPB, HPE), "indvar", 1438 Header->begin()); 1439 rememberInstruction(CanonicalIV); 1440 1441 SmallSet<BasicBlock *, 4> PredSeen; 1442 Constant *One = ConstantInt::get(Ty, 1); 1443 for (pred_iterator HPI = HPB; HPI != HPE; ++HPI) { 1444 BasicBlock *HP = *HPI; 1445 if (!PredSeen.insert(HP).second) { 1446 // There must be an incoming value for each predecessor, even the 1447 // duplicates! 1448 CanonicalIV->addIncoming(CanonicalIV->getIncomingValueForBlock(HP), HP); 1449 continue; 1450 } 1451 1452 if (L->contains(HP)) { 1453 // Insert a unit add instruction right before the terminator 1454 // corresponding to the back-edge. 1455 Instruction *Add = BinaryOperator::CreateAdd(CanonicalIV, One, 1456 "indvar.next", 1457 HP->getTerminator()); 1458 Add->setDebugLoc(HP->getTerminator()->getDebugLoc()); 1459 rememberInstruction(Add); 1460 CanonicalIV->addIncoming(Add, HP); 1461 } else { 1462 CanonicalIV->addIncoming(Constant::getNullValue(Ty), HP); 1463 } 1464 } 1465 } 1466 1467 // {0,+,1} --> Insert a canonical induction variable into the loop! 1468 if (S->isAffine() && S->getOperand(1)->isOne()) { 1469 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 1470 "IVs with types different from the canonical IV should " 1471 "already have been handled!"); 1472 return CanonicalIV; 1473 } 1474 1475 // {0,+,F} --> {0,+,1} * F 1476 1477 // If this is a simple linear addrec, emit it now as a special case. 1478 if (S->isAffine()) // {0,+,F} --> i*F 1479 return 1480 expand(SE.getTruncateOrNoop( 1481 SE.getMulExpr(SE.getUnknown(CanonicalIV), 1482 SE.getNoopOrAnyExtend(S->getOperand(1), 1483 CanonicalIV->getType())), 1484 Ty)); 1485 1486 // If this is a chain of recurrences, turn it into a closed form, using the 1487 // folders, then expandCodeFor the closed form. This allows the folders to 1488 // simplify the expression without having to build a bunch of special code 1489 // into this folder. 1490 const SCEV *IH = SE.getUnknown(CanonicalIV); // Get I as a "symbolic" SCEV. 1491 1492 // Promote S up to the canonical IV type, if the cast is foldable. 1493 const SCEV *NewS = S; 1494 const SCEV *Ext = SE.getNoopOrAnyExtend(S, CanonicalIV->getType()); 1495 if (isa<SCEVAddRecExpr>(Ext)) 1496 NewS = Ext; 1497 1498 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1499 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 1500 1501 // Truncate the result down to the original type, if needed. 1502 const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1503 return expand(T); 1504 } 1505 1506 Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1507 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1508 Value *V = expandCodeFor(S->getOperand(), 1509 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1510 Value *I = Builder.CreateTrunc(V, Ty); 1511 rememberInstruction(I); 1512 return I; 1513 } 1514 1515 Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1516 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1517 Value *V = expandCodeFor(S->getOperand(), 1518 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1519 Value *I = Builder.CreateZExt(V, Ty); 1520 rememberInstruction(I); 1521 return I; 1522 } 1523 1524 Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1525 Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1526 Value *V = expandCodeFor(S->getOperand(), 1527 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1528 Value *I = Builder.CreateSExt(V, Ty); 1529 rememberInstruction(I); 1530 return I; 1531 } 1532 1533 Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 1534 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1535 Type *Ty = LHS->getType(); 1536 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1537 // In the case of mixed integer and pointer types, do the 1538 // rest of the comparisons as integer. 1539 if (S->getOperand(i)->getType() != Ty) { 1540 Ty = SE.getEffectiveSCEVType(Ty); 1541 LHS = InsertNoopCastOfTo(LHS, Ty); 1542 } 1543 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1544 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS); 1545 rememberInstruction(ICmp); 1546 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1547 rememberInstruction(Sel); 1548 LHS = Sel; 1549 } 1550 // In the case of mixed integer and pointer types, cast the 1551 // final result back to the pointer type. 1552 if (LHS->getType() != S->getType()) 1553 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1554 return LHS; 1555 } 1556 1557 Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 1558 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1559 Type *Ty = LHS->getType(); 1560 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1561 // In the case of mixed integer and pointer types, do the 1562 // rest of the comparisons as integer. 1563 if (S->getOperand(i)->getType() != Ty) { 1564 Ty = SE.getEffectiveSCEVType(Ty); 1565 LHS = InsertNoopCastOfTo(LHS, Ty); 1566 } 1567 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1568 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS); 1569 rememberInstruction(ICmp); 1570 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1571 rememberInstruction(Sel); 1572 LHS = Sel; 1573 } 1574 // In the case of mixed integer and pointer types, cast the 1575 // final result back to the pointer type. 1576 if (LHS->getType() != S->getType()) 1577 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1578 return LHS; 1579 } 1580 1581 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty, 1582 Instruction *IP) { 1583 Builder.SetInsertPoint(IP->getParent(), IP); 1584 return expandCodeFor(SH, Ty); 1585 } 1586 1587 Value *SCEVExpander::expandCodeFor(const SCEV *SH, Type *Ty) { 1588 // Expand the code for this SCEV. 1589 Value *V = expand(SH); 1590 if (Ty) { 1591 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 1592 "non-trivial casts should be done with the SCEVs directly!"); 1593 V = InsertNoopCastOfTo(V, Ty); 1594 } 1595 return V; 1596 } 1597 1598 Value *SCEVExpander::expand(const SCEV *S) { 1599 // Compute an insertion point for this SCEV object. Hoist the instructions 1600 // as far out in the loop nest as possible. 1601 Instruction *InsertPt = Builder.GetInsertPoint(); 1602 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ; 1603 L = L->getParentLoop()) 1604 if (SE.isLoopInvariant(S, L)) { 1605 if (!L) break; 1606 if (BasicBlock *Preheader = L->getLoopPreheader()) 1607 InsertPt = Preheader->getTerminator(); 1608 else { 1609 // LSR sets the insertion point for AddRec start/step values to the 1610 // block start to simplify value reuse, even though it's an invalid 1611 // position. SCEVExpander must correct for this in all cases. 1612 InsertPt = L->getHeader()->getFirstInsertionPt(); 1613 } 1614 } else { 1615 // If the SCEV is computable at this level, insert it into the header 1616 // after the PHIs (and after any other instructions that we've inserted 1617 // there) so that it is guaranteed to dominate any user inside the loop. 1618 if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L)) 1619 InsertPt = L->getHeader()->getFirstInsertionPt(); 1620 while (InsertPt != Builder.GetInsertPoint() 1621 && (isInsertedInstruction(InsertPt) 1622 || isa<DbgInfoIntrinsic>(InsertPt))) { 1623 InsertPt = std::next(BasicBlock::iterator(InsertPt)); 1624 } 1625 break; 1626 } 1627 1628 // Check to see if we already expanded this here. 1629 std::map<std::pair<const SCEV *, Instruction *>, TrackingVH<Value> >::iterator 1630 I = InsertedExpressions.find(std::make_pair(S, InsertPt)); 1631 if (I != InsertedExpressions.end()) 1632 return I->second; 1633 1634 BuilderType::InsertPointGuard Guard(Builder); 1635 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt); 1636 1637 // Expand the expression into instructions. 1638 Value *V = visit(S); 1639 1640 // Remember the expanded value for this SCEV at this location. 1641 // 1642 // This is independent of PostIncLoops. The mapped value simply materializes 1643 // the expression at this insertion point. If the mapped value happened to be 1644 // a postinc expansion, it could be reused by a non-postinc user, but only if 1645 // its insertion point was already at the head of the loop. 1646 InsertedExpressions[std::make_pair(S, InsertPt)] = V; 1647 return V; 1648 } 1649 1650 void SCEVExpander::rememberInstruction(Value *I) { 1651 if (!PostIncLoops.empty()) 1652 InsertedPostIncValues.insert(I); 1653 else 1654 InsertedValues.insert(I); 1655 } 1656 1657 /// getOrInsertCanonicalInductionVariable - This method returns the 1658 /// canonical induction variable of the specified type for the specified 1659 /// loop (inserting one if there is none). A canonical induction variable 1660 /// starts at zero and steps by one on each iteration. 1661 PHINode * 1662 SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1663 Type *Ty) { 1664 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1665 1666 // Build a SCEV for {0,+,1}<L>. 1667 // Conservatively use FlagAnyWrap for now. 1668 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 1669 SE.getConstant(Ty, 1), L, SCEV::FlagAnyWrap); 1670 1671 // Emit code for it. 1672 BuilderType::InsertPointGuard Guard(Builder); 1673 PHINode *V = cast<PHINode>(expandCodeFor(H, nullptr, 1674 L->getHeader()->begin())); 1675 1676 return V; 1677 } 1678 1679 /// replaceCongruentIVs - Check for congruent phis in this loop header and 1680 /// replace them with their most canonical representative. Return the number of 1681 /// phis eliminated. 1682 /// 1683 /// This does not depend on any SCEVExpander state but should be used in 1684 /// the same context that SCEVExpander is used. 1685 unsigned SCEVExpander::replaceCongruentIVs(Loop *L, const DominatorTree *DT, 1686 SmallVectorImpl<WeakVH> &DeadInsts, 1687 const TargetTransformInfo *TTI) { 1688 // Find integer phis in order of increasing width. 1689 SmallVector<PHINode*, 8> Phis; 1690 for (BasicBlock::iterator I = L->getHeader()->begin(); 1691 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 1692 Phis.push_back(Phi); 1693 } 1694 if (TTI) 1695 std::sort(Phis.begin(), Phis.end(), [](Value *LHS, Value *RHS) { 1696 // Put pointers at the back and make sure pointer < pointer = false. 1697 if (!LHS->getType()->isIntegerTy() || !RHS->getType()->isIntegerTy()) 1698 return RHS->getType()->isIntegerTy() && !LHS->getType()->isIntegerTy(); 1699 return RHS->getType()->getPrimitiveSizeInBits() < 1700 LHS->getType()->getPrimitiveSizeInBits(); 1701 }); 1702 1703 unsigned NumElim = 0; 1704 DenseMap<const SCEV *, PHINode *> ExprToIVMap; 1705 // Process phis from wide to narrow. Mapping wide phis to the their truncation 1706 // so narrow phis can reuse them. 1707 for (SmallVectorImpl<PHINode*>::const_iterator PIter = Phis.begin(), 1708 PEnd = Phis.end(); PIter != PEnd; ++PIter) { 1709 PHINode *Phi = *PIter; 1710 1711 // Fold constant phis. They may be congruent to other constant phis and 1712 // would confuse the logic below that expects proper IVs. 1713 if (Value *V = SimplifyInstruction(Phi, DL, SE.TLI, SE.DT, SE.AC)) { 1714 Phi->replaceAllUsesWith(V); 1715 DeadInsts.push_back(Phi); 1716 ++NumElim; 1717 DEBUG_WITH_TYPE(DebugType, dbgs() 1718 << "INDVARS: Eliminated constant iv: " << *Phi << '\n'); 1719 continue; 1720 } 1721 1722 if (!SE.isSCEVable(Phi->getType())) 1723 continue; 1724 1725 PHINode *&OrigPhiRef = ExprToIVMap[SE.getSCEV(Phi)]; 1726 if (!OrigPhiRef) { 1727 OrigPhiRef = Phi; 1728 if (Phi->getType()->isIntegerTy() && TTI 1729 && TTI->isTruncateFree(Phi->getType(), Phis.back()->getType())) { 1730 // This phi can be freely truncated to the narrowest phi type. Map the 1731 // truncated expression to it so it will be reused for narrow types. 1732 const SCEV *TruncExpr = 1733 SE.getTruncateExpr(SE.getSCEV(Phi), Phis.back()->getType()); 1734 ExprToIVMap[TruncExpr] = Phi; 1735 } 1736 continue; 1737 } 1738 1739 // Replacing a pointer phi with an integer phi or vice-versa doesn't make 1740 // sense. 1741 if (OrigPhiRef->getType()->isPointerTy() != Phi->getType()->isPointerTy()) 1742 continue; 1743 1744 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 1745 Instruction *OrigInc = 1746 cast<Instruction>(OrigPhiRef->getIncomingValueForBlock(LatchBlock)); 1747 Instruction *IsomorphicInc = 1748 cast<Instruction>(Phi->getIncomingValueForBlock(LatchBlock)); 1749 1750 // If this phi has the same width but is more canonical, replace the 1751 // original with it. As part of the "more canonical" determination, 1752 // respect a prior decision to use an IV chain. 1753 if (OrigPhiRef->getType() == Phi->getType() 1754 && !(ChainedPhis.count(Phi) 1755 || isExpandedAddRecExprPHI(OrigPhiRef, OrigInc, L)) 1756 && (ChainedPhis.count(Phi) 1757 || isExpandedAddRecExprPHI(Phi, IsomorphicInc, L))) { 1758 std::swap(OrigPhiRef, Phi); 1759 std::swap(OrigInc, IsomorphicInc); 1760 } 1761 // Replacing the congruent phi is sufficient because acyclic redundancy 1762 // elimination, CSE/GVN, should handle the rest. However, once SCEV proves 1763 // that a phi is congruent, it's often the head of an IV user cycle that 1764 // is isomorphic with the original phi. It's worth eagerly cleaning up the 1765 // common case of a single IV increment so that DeleteDeadPHIs can remove 1766 // cycles that had postinc uses. 1767 const SCEV *TruncExpr = SE.getTruncateOrNoop(SE.getSCEV(OrigInc), 1768 IsomorphicInc->getType()); 1769 if (OrigInc != IsomorphicInc 1770 && TruncExpr == SE.getSCEV(IsomorphicInc) 1771 && ((isa<PHINode>(OrigInc) && isa<PHINode>(IsomorphicInc)) 1772 || hoistIVInc(OrigInc, IsomorphicInc))) { 1773 DEBUG_WITH_TYPE(DebugType, dbgs() 1774 << "INDVARS: Eliminated congruent iv.inc: " 1775 << *IsomorphicInc << '\n'); 1776 Value *NewInc = OrigInc; 1777 if (OrigInc->getType() != IsomorphicInc->getType()) { 1778 Instruction *IP = nullptr; 1779 if (PHINode *PN = dyn_cast<PHINode>(OrigInc)) 1780 IP = PN->getParent()->getFirstInsertionPt(); 1781 else 1782 IP = OrigInc->getNextNode(); 1783 1784 IRBuilder<> Builder(IP); 1785 Builder.SetCurrentDebugLocation(IsomorphicInc->getDebugLoc()); 1786 NewInc = Builder. 1787 CreateTruncOrBitCast(OrigInc, IsomorphicInc->getType(), IVName); 1788 } 1789 IsomorphicInc->replaceAllUsesWith(NewInc); 1790 DeadInsts.push_back(IsomorphicInc); 1791 } 1792 } 1793 DEBUG_WITH_TYPE(DebugType, dbgs() 1794 << "INDVARS: Eliminated congruent iv: " << *Phi << '\n'); 1795 ++NumElim; 1796 Value *NewIV = OrigPhiRef; 1797 if (OrigPhiRef->getType() != Phi->getType()) { 1798 IRBuilder<> Builder(L->getHeader()->getFirstInsertionPt()); 1799 Builder.SetCurrentDebugLocation(Phi->getDebugLoc()); 1800 NewIV = Builder.CreateTruncOrBitCast(OrigPhiRef, Phi->getType(), IVName); 1801 } 1802 Phi->replaceAllUsesWith(NewIV); 1803 DeadInsts.push_back(Phi); 1804 } 1805 return NumElim; 1806 } 1807 1808 bool SCEVExpander::isHighCostExpansionHelper( 1809 const SCEV *S, Loop *L, SmallPtrSetImpl<const SCEV *> &Processed) { 1810 if (!Processed.insert(S).second) 1811 return false; 1812 1813 if (auto *UDivExpr = dyn_cast<SCEVUDivExpr>(S)) { 1814 // If the divisor is a power of two and the SCEV type fits in a native 1815 // integer, consider the divison cheap irrespective of whether it occurs in 1816 // the user code since it can be lowered into a right shift. 1817 if (auto *SC = dyn_cast<SCEVConstant>(UDivExpr->getRHS())) 1818 if (SC->getValue()->getValue().isPowerOf2()) { 1819 const DataLayout &DL = 1820 L->getHeader()->getParent()->getParent()->getDataLayout(); 1821 unsigned Width = cast<IntegerType>(UDivExpr->getType())->getBitWidth(); 1822 return DL.isIllegalInteger(Width); 1823 } 1824 1825 // UDivExpr is very likely a UDiv that ScalarEvolution's HowFarToZero or 1826 // HowManyLessThans produced to compute a precise expression, rather than a 1827 // UDiv from the user's code. If we can't find a UDiv in the code with some 1828 // simple searching, assume the former consider UDivExpr expensive to 1829 // compute. 1830 BasicBlock *ExitingBB = L->getExitingBlock(); 1831 if (!ExitingBB) 1832 return true; 1833 1834 BranchInst *ExitingBI = dyn_cast<BranchInst>(ExitingBB->getTerminator()); 1835 if (!ExitingBI || !ExitingBI->isConditional()) 1836 return true; 1837 1838 ICmpInst *OrigCond = dyn_cast<ICmpInst>(ExitingBI->getCondition()); 1839 if (!OrigCond) 1840 return true; 1841 1842 const SCEV *RHS = SE.getSCEV(OrigCond->getOperand(1)); 1843 RHS = SE.getMinusSCEV(RHS, SE.getConstant(RHS->getType(), 1)); 1844 if (RHS != S) { 1845 const SCEV *LHS = SE.getSCEV(OrigCond->getOperand(0)); 1846 LHS = SE.getMinusSCEV(LHS, SE.getConstant(LHS->getType(), 1)); 1847 if (LHS != S) 1848 return true; 1849 } 1850 } 1851 1852 // Recurse past add expressions, which commonly occur in the 1853 // BackedgeTakenCount. They may already exist in program code, and if not, 1854 // they are not too expensive rematerialize. 1855 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 1856 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 1857 I != E; ++I) { 1858 if (isHighCostExpansionHelper(*I, L, Processed)) 1859 return true; 1860 } 1861 return false; 1862 } 1863 1864 // HowManyLessThans uses a Max expression whenever the loop is not guarded by 1865 // the exit condition. 1866 if (isa<SCEVSMaxExpr>(S) || isa<SCEVUMaxExpr>(S)) 1867 return true; 1868 1869 // If we haven't recognized an expensive SCEV pattern, assume it's an 1870 // expression produced by program code. 1871 return false; 1872 } 1873 1874 namespace { 1875 // Search for a SCEV subexpression that is not safe to expand. Any expression 1876 // that may expand to a !isSafeToSpeculativelyExecute value is unsafe, namely 1877 // UDiv expressions. We don't know if the UDiv is derived from an IR divide 1878 // instruction, but the important thing is that we prove the denominator is 1879 // nonzero before expansion. 1880 // 1881 // IVUsers already checks that IV-derived expressions are safe. So this check is 1882 // only needed when the expression includes some subexpression that is not IV 1883 // derived. 1884 // 1885 // Currently, we only allow division by a nonzero constant here. If this is 1886 // inadequate, we could easily allow division by SCEVUnknown by using 1887 // ValueTracking to check isKnownNonZero(). 1888 // 1889 // We cannot generally expand recurrences unless the step dominates the loop 1890 // header. The expander handles the special case of affine recurrences by 1891 // scaling the recurrence outside the loop, but this technique isn't generally 1892 // applicable. Expanding a nested recurrence outside a loop requires computing 1893 // binomial coefficients. This could be done, but the recurrence has to be in a 1894 // perfectly reduced form, which can't be guaranteed. 1895 struct SCEVFindUnsafe { 1896 ScalarEvolution &SE; 1897 bool IsUnsafe; 1898 1899 SCEVFindUnsafe(ScalarEvolution &se): SE(se), IsUnsafe(false) {} 1900 1901 bool follow(const SCEV *S) { 1902 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 1903 const SCEVConstant *SC = dyn_cast<SCEVConstant>(D->getRHS()); 1904 if (!SC || SC->getValue()->isZero()) { 1905 IsUnsafe = true; 1906 return false; 1907 } 1908 } 1909 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 1910 const SCEV *Step = AR->getStepRecurrence(SE); 1911 if (!AR->isAffine() && !SE.dominates(Step, AR->getLoop()->getHeader())) { 1912 IsUnsafe = true; 1913 return false; 1914 } 1915 } 1916 return true; 1917 } 1918 bool isDone() const { return IsUnsafe; } 1919 }; 1920 } 1921 1922 namespace llvm { 1923 bool isSafeToExpand(const SCEV *S, ScalarEvolution &SE) { 1924 SCEVFindUnsafe Search(SE); 1925 visitAll(S, Search); 1926 return !Search.IsUnsafe; 1927 } 1928 } 1929