ScalarEvolutionExpander.cpp revision 53b73a283e0a0339f7a273775ee21ebcc220b089
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/Analysis/LoopInfo.h" 18#include "llvm/IntrinsicInst.h" 19#include "llvm/LLVMContext.h" 20#include "llvm/Target/TargetData.h" 21#include "llvm/ADT/STLExtras.h" 22using namespace llvm; 23 24/// InsertNoopCastOfTo - Insert a cast of V to the specified type, 25/// which must be possible with a noop cast, doing what we can to share 26/// the casts. 27Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) { 28 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 29 assert((Op == Instruction::BitCast || 30 Op == Instruction::PtrToInt || 31 Op == Instruction::IntToPtr) && 32 "InsertNoopCastOfTo cannot perform non-noop casts!"); 33 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 34 "InsertNoopCastOfTo cannot change sizes!"); 35 36 // Short-circuit unnecessary bitcasts. 37 if (Op == Instruction::BitCast && V->getType() == Ty) 38 return V; 39 40 // Short-circuit unnecessary inttoptr<->ptrtoint casts. 41 if ((Op == Instruction::PtrToInt || Op == Instruction::IntToPtr) && 42 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 43 if (CastInst *CI = dyn_cast<CastInst>(V)) 44 if ((CI->getOpcode() == Instruction::PtrToInt || 45 CI->getOpcode() == Instruction::IntToPtr) && 46 SE.getTypeSizeInBits(CI->getType()) == 47 SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 48 return CI->getOperand(0); 49 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 50 if ((CE->getOpcode() == Instruction::PtrToInt || 51 CE->getOpcode() == Instruction::IntToPtr) && 52 SE.getTypeSizeInBits(CE->getType()) == 53 SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 54 return CE->getOperand(0); 55 } 56 57 if (Constant *C = dyn_cast<Constant>(V)) 58 return ConstantExpr::getCast(Op, C, Ty); 59 60 if (Argument *A = dyn_cast<Argument>(V)) { 61 // Check to see if there is already a cast! 62 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); 63 UI != E; ++UI) 64 if ((*UI)->getType() == Ty) 65 if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI))) 66 if (CI->getOpcode() == Op) { 67 // If the cast isn't the first instruction of the function, move it. 68 if (BasicBlock::iterator(CI) != 69 A->getParent()->getEntryBlock().begin()) { 70 // Recreate the cast at the beginning of the entry block. 71 // The old cast is left in place in case it is being used 72 // as an insert point. 73 Instruction *NewCI = 74 CastInst::Create(Op, V, Ty, "", 75 A->getParent()->getEntryBlock().begin()); 76 NewCI->takeName(CI); 77 CI->replaceAllUsesWith(NewCI); 78 return NewCI; 79 } 80 return CI; 81 } 82 83 Instruction *I = CastInst::Create(Op, V, Ty, V->getName(), 84 A->getParent()->getEntryBlock().begin()); 85 rememberInstruction(I); 86 return I; 87 } 88 89 Instruction *I = cast<Instruction>(V); 90 91 // Check to see if there is already a cast. If there is, use it. 92 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 93 UI != E; ++UI) { 94 if ((*UI)->getType() == Ty) 95 if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI))) 96 if (CI->getOpcode() == Op) { 97 BasicBlock::iterator It = I; ++It; 98 if (isa<InvokeInst>(I)) 99 It = cast<InvokeInst>(I)->getNormalDest()->begin(); 100 while (isa<PHINode>(It)) ++It; 101 if (It != BasicBlock::iterator(CI)) { 102 // Recreate the cast after the user. 103 // The old cast is left in place in case it is being used 104 // as an insert point. 105 Instruction *NewCI = CastInst::Create(Op, V, Ty, "", It); 106 NewCI->takeName(CI); 107 CI->replaceAllUsesWith(NewCI); 108 rememberInstruction(NewCI); 109 return NewCI; 110 } 111 rememberInstruction(CI); 112 return CI; 113 } 114 } 115 BasicBlock::iterator IP = I; ++IP; 116 if (InvokeInst *II = dyn_cast<InvokeInst>(I)) 117 IP = II->getNormalDest()->begin(); 118 while (isa<PHINode>(IP)) ++IP; 119 Instruction *CI = CastInst::Create(Op, V, Ty, V->getName(), IP); 120 rememberInstruction(CI); 121 return CI; 122} 123 124/// InsertBinop - Insert the specified binary operator, doing a small amount 125/// of work to avoid inserting an obviously redundant operation. 126Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, 127 Value *LHS, Value *RHS) { 128 // Fold a binop with constant operands. 129 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 130 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 131 return ConstantExpr::get(Opcode, CLHS, CRHS); 132 133 // Do a quick scan to see if we have this binop nearby. If so, reuse it. 134 unsigned ScanLimit = 6; 135 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 136 // Scanning starts from the last instruction before the insertion point. 137 BasicBlock::iterator IP = Builder.GetInsertPoint(); 138 if (IP != BlockBegin) { 139 --IP; 140 for (; ScanLimit; --IP, --ScanLimit) { 141 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 142 // generated code. 143 if (isa<DbgInfoIntrinsic>(IP)) 144 ScanLimit++; 145 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 146 IP->getOperand(1) == RHS) 147 return IP; 148 if (IP == BlockBegin) break; 149 } 150 } 151 152 // Save the original insertion point so we can restore it when we're done. 153 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 154 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 155 156 // Move the insertion point out of as many loops as we can. 157 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 158 if (!L->isLoopInvariant(LHS) || !L->isLoopInvariant(RHS)) break; 159 BasicBlock *Preheader = L->getLoopPreheader(); 160 if (!Preheader) break; 161 162 // Ok, move up a level. 163 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 164 } 165 166 // If we haven't found this binop, insert it. 167 Value *BO = Builder.CreateBinOp(Opcode, LHS, RHS, "tmp"); 168 rememberInstruction(BO); 169 170 // Restore the original insert point. 171 if (SaveInsertBB) 172 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 173 174 return BO; 175} 176 177/// FactorOutConstant - Test if S is divisible by Factor, using signed 178/// division. If so, update S with Factor divided out and return true. 179/// S need not be evenly divisible if a reasonable remainder can be 180/// computed. 181/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 182/// unnecessary; in its place, just signed-divide Ops[i] by the scale and 183/// check to see if the divide was folded. 184static bool FactorOutConstant(const SCEV *&S, 185 const SCEV *&Remainder, 186 const SCEV *Factor, 187 ScalarEvolution &SE, 188 const TargetData *TD) { 189 // Everything is divisible by one. 190 if (Factor->isOne()) 191 return true; 192 193 // x/x == 1. 194 if (S == Factor) { 195 S = SE.getIntegerSCEV(1, S->getType()); 196 return true; 197 } 198 199 // For a Constant, check for a multiple of the given factor. 200 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 201 // 0/x == 0. 202 if (C->isZero()) 203 return true; 204 // Check for divisibility. 205 if (const SCEVConstant *FC = dyn_cast<SCEVConstant>(Factor)) { 206 ConstantInt *CI = 207 ConstantInt::get(SE.getContext(), 208 C->getValue()->getValue().sdiv( 209 FC->getValue()->getValue())); 210 // If the quotient is zero and the remainder is non-zero, reject 211 // the value at this scale. It will be considered for subsequent 212 // smaller scales. 213 if (!CI->isZero()) { 214 const SCEV *Div = SE.getConstant(CI); 215 S = Div; 216 Remainder = 217 SE.getAddExpr(Remainder, 218 SE.getConstant(C->getValue()->getValue().srem( 219 FC->getValue()->getValue()))); 220 return true; 221 } 222 } 223 } 224 225 // In a Mul, check if there is a constant operand which is a multiple 226 // of the given factor. 227 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 228 if (TD) { 229 // With TargetData, the size is known. Check if there is a constant 230 // operand which is a multiple of the given factor. If so, we can 231 // factor it. 232 const SCEVConstant *FC = cast<SCEVConstant>(Factor); 233 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 234 if (!C->getValue()->getValue().srem(FC->getValue()->getValue())) { 235 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 236 NewMulOps[0] = 237 SE.getConstant(C->getValue()->getValue().sdiv( 238 FC->getValue()->getValue())); 239 S = SE.getMulExpr(NewMulOps); 240 return true; 241 } 242 } else { 243 // Without TargetData, check if Factor can be factored out of any of the 244 // Mul's operands. If so, we can just remove it. 245 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) { 246 const SCEV *SOp = M->getOperand(i); 247 const SCEV *Remainder = SE.getIntegerSCEV(0, SOp->getType()); 248 if (FactorOutConstant(SOp, Remainder, Factor, SE, TD) && 249 Remainder->isZero()) { 250 SmallVector<const SCEV *, 4> NewMulOps(M->op_begin(), M->op_end()); 251 NewMulOps[i] = SOp; 252 S = SE.getMulExpr(NewMulOps); 253 return true; 254 } 255 } 256 } 257 } 258 259 // In an AddRec, check if both start and step are divisible. 260 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 261 const SCEV *Step = A->getStepRecurrence(SE); 262 const SCEV *StepRem = SE.getIntegerSCEV(0, Step->getType()); 263 if (!FactorOutConstant(Step, StepRem, Factor, SE, TD)) 264 return false; 265 if (!StepRem->isZero()) 266 return false; 267 const SCEV *Start = A->getStart(); 268 if (!FactorOutConstant(Start, Remainder, Factor, SE, TD)) 269 return false; 270 S = SE.getAddRecExpr(Start, Step, A->getLoop()); 271 return true; 272 } 273 274 return false; 275} 276 277/// SimplifyAddOperands - Sort and simplify a list of add operands. NumAddRecs 278/// is the number of SCEVAddRecExprs present, which are kept at the end of 279/// the list. 280/// 281static void SimplifyAddOperands(SmallVectorImpl<const SCEV *> &Ops, 282 const Type *Ty, 283 ScalarEvolution &SE) { 284 unsigned NumAddRecs = 0; 285 for (unsigned i = Ops.size(); i > 0 && isa<SCEVAddRecExpr>(Ops[i-1]); --i) 286 ++NumAddRecs; 287 // Group Ops into non-addrecs and addrecs. 288 SmallVector<const SCEV *, 8> NoAddRecs(Ops.begin(), Ops.end() - NumAddRecs); 289 SmallVector<const SCEV *, 8> AddRecs(Ops.end() - NumAddRecs, Ops.end()); 290 // Let ScalarEvolution sort and simplify the non-addrecs list. 291 const SCEV *Sum = NoAddRecs.empty() ? 292 SE.getIntegerSCEV(0, Ty) : 293 SE.getAddExpr(NoAddRecs); 294 // If it returned an add, use the operands. Otherwise it simplified 295 // the sum into a single value, so just use that. 296 Ops.clear(); 297 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Sum)) 298 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end()); 299 else if (!Sum->isZero()) 300 Ops.push_back(Sum); 301 // Then append the addrecs. 302 Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end()); 303} 304 305/// SplitAddRecs - Flatten a list of add operands, moving addrec start values 306/// out to the top level. For example, convert {a + b,+,c} to a, b, {0,+,d}. 307/// This helps expose more opportunities for folding parts of the expressions 308/// into GEP indices. 309/// 310static void SplitAddRecs(SmallVectorImpl<const SCEV *> &Ops, 311 const Type *Ty, 312 ScalarEvolution &SE) { 313 // Find the addrecs. 314 SmallVector<const SCEV *, 8> AddRecs; 315 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 316 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) { 317 const SCEV *Start = A->getStart(); 318 if (Start->isZero()) break; 319 const SCEV *Zero = SE.getIntegerSCEV(0, Ty); 320 AddRecs.push_back(SE.getAddRecExpr(Zero, 321 A->getStepRecurrence(SE), 322 A->getLoop())); 323 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Start)) { 324 Ops[i] = Zero; 325 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end()); 326 e += Add->getNumOperands(); 327 } else { 328 Ops[i] = Start; 329 } 330 } 331 if (!AddRecs.empty()) { 332 // Add the addrecs onto the end of the list. 333 Ops.insert(Ops.end(), AddRecs.begin(), AddRecs.end()); 334 // Resort the operand list, moving any constants to the front. 335 SimplifyAddOperands(Ops, Ty, SE); 336 } 337} 338 339/// expandAddToGEP - Expand an addition expression with a pointer type into 340/// a GEP instead of using ptrtoint+arithmetic+inttoptr. This helps 341/// BasicAliasAnalysis and other passes analyze the result. See the rules 342/// for getelementptr vs. inttoptr in 343/// http://llvm.org/docs/LangRef.html#pointeraliasing 344/// for details. 345/// 346/// Design note: The correctness of using getelementptr here depends on 347/// ScalarEvolution not recognizing inttoptr and ptrtoint operators, as 348/// they may introduce pointer arithmetic which may not be safely converted 349/// into getelementptr. 350/// 351/// Design note: It might seem desirable for this function to be more 352/// loop-aware. If some of the indices are loop-invariant while others 353/// aren't, it might seem desirable to emit multiple GEPs, keeping the 354/// loop-invariant portions of the overall computation outside the loop. 355/// However, there are a few reasons this is not done here. Hoisting simple 356/// arithmetic is a low-level optimization that often isn't very 357/// important until late in the optimization process. In fact, passes 358/// like InstructionCombining will combine GEPs, even if it means 359/// pushing loop-invariant computation down into loops, so even if the 360/// GEPs were split here, the work would quickly be undone. The 361/// LoopStrengthReduction pass, which is usually run quite late (and 362/// after the last InstructionCombining pass), takes care of hoisting 363/// loop-invariant portions of expressions, after considering what 364/// can be folded using target addressing modes. 365/// 366Value *SCEVExpander::expandAddToGEP(const SCEV *const *op_begin, 367 const SCEV *const *op_end, 368 const PointerType *PTy, 369 const Type *Ty, 370 Value *V) { 371 const Type *ElTy = PTy->getElementType(); 372 SmallVector<Value *, 4> GepIndices; 373 SmallVector<const SCEV *, 8> Ops(op_begin, op_end); 374 bool AnyNonZeroIndices = false; 375 376 // Split AddRecs up into parts as either of the parts may be usable 377 // without the other. 378 SplitAddRecs(Ops, Ty, SE); 379 380 // Descend down the pointer's type and attempt to convert the other 381 // operands into GEP indices, at each level. The first index in a GEP 382 // indexes into the array implied by the pointer operand; the rest of 383 // the indices index into the element or field type selected by the 384 // preceding index. 385 for (;;) { 386 // If the scale size is not 0, attempt to factor out a scale for 387 // array indexing. 388 SmallVector<const SCEV *, 8> ScaledOps; 389 if (ElTy->isSized()) { 390 const SCEV *ElSize = SE.getSizeOfExpr(ElTy); 391 if (!ElSize->isZero()) { 392 SmallVector<const SCEV *, 8> NewOps; 393 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 394 const SCEV *Op = Ops[i]; 395 const SCEV *Remainder = SE.getIntegerSCEV(0, Ty); 396 if (FactorOutConstant(Op, Remainder, ElSize, SE, SE.TD)) { 397 // Op now has ElSize factored out. 398 ScaledOps.push_back(Op); 399 if (!Remainder->isZero()) 400 NewOps.push_back(Remainder); 401 AnyNonZeroIndices = true; 402 } else { 403 // The operand was not divisible, so add it to the list of operands 404 // we'll scan next iteration. 405 NewOps.push_back(Ops[i]); 406 } 407 } 408 // If we made any changes, update Ops. 409 if (!ScaledOps.empty()) { 410 Ops = NewOps; 411 SimplifyAddOperands(Ops, Ty, SE); 412 } 413 } 414 } 415 416 // Record the scaled array index for this level of the type. If 417 // we didn't find any operands that could be factored, tentatively 418 // assume that element zero was selected (since the zero offset 419 // would obviously be folded away). 420 Value *Scaled = ScaledOps.empty() ? 421 Constant::getNullValue(Ty) : 422 expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 423 GepIndices.push_back(Scaled); 424 425 // Collect struct field index operands. 426 while (const StructType *STy = dyn_cast<StructType>(ElTy)) { 427 bool FoundFieldNo = false; 428 // An empty struct has no fields. 429 if (STy->getNumElements() == 0) break; 430 if (SE.TD) { 431 // With TargetData, field offsets are known. See if a constant offset 432 // falls within any of the struct fields. 433 if (Ops.empty()) break; 434 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 435 if (SE.getTypeSizeInBits(C->getType()) <= 64) { 436 const StructLayout &SL = *SE.TD->getStructLayout(STy); 437 uint64_t FullOffset = C->getValue()->getZExtValue(); 438 if (FullOffset < SL.getSizeInBytes()) { 439 unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 440 GepIndices.push_back( 441 ConstantInt::get(Type::getInt32Ty(Ty->getContext()), ElIdx)); 442 ElTy = STy->getTypeAtIndex(ElIdx); 443 Ops[0] = 444 SE.getConstant(Ty, FullOffset - SL.getElementOffset(ElIdx)); 445 AnyNonZeroIndices = true; 446 FoundFieldNo = true; 447 } 448 } 449 } else { 450 // Without TargetData, just check for an offsetof expression of the 451 // appropriate struct type. 452 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 453 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Ops[i])) { 454 const Type *CTy; 455 Constant *FieldNo; 456 if (U->isOffsetOf(CTy, FieldNo) && CTy == STy) { 457 GepIndices.push_back(FieldNo); 458 ElTy = 459 STy->getTypeAtIndex(cast<ConstantInt>(FieldNo)->getZExtValue()); 460 Ops[i] = SE.getConstant(Ty, 0); 461 AnyNonZeroIndices = true; 462 FoundFieldNo = true; 463 break; 464 } 465 } 466 } 467 // If no struct field offsets were found, tentatively assume that 468 // field zero was selected (since the zero offset would obviously 469 // be folded away). 470 if (!FoundFieldNo) { 471 ElTy = STy->getTypeAtIndex(0u); 472 GepIndices.push_back( 473 Constant::getNullValue(Type::getInt32Ty(Ty->getContext()))); 474 } 475 } 476 477 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) 478 ElTy = ATy->getElementType(); 479 else 480 break; 481 } 482 483 // If none of the operands were convertible to proper GEP indices, cast 484 // the base to i8* and do an ugly getelementptr with that. It's still 485 // better than ptrtoint+arithmetic+inttoptr at least. 486 if (!AnyNonZeroIndices) { 487 // Cast the base to i8*. 488 V = InsertNoopCastOfTo(V, 489 Type::getInt8PtrTy(Ty->getContext(), PTy->getAddressSpace())); 490 491 // Expand the operands for a plain byte offset. 492 Value *Idx = expandCodeFor(SE.getAddExpr(Ops), Ty); 493 494 // Fold a GEP with constant operands. 495 if (Constant *CLHS = dyn_cast<Constant>(V)) 496 if (Constant *CRHS = dyn_cast<Constant>(Idx)) 497 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1); 498 499 // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 500 unsigned ScanLimit = 6; 501 BasicBlock::iterator BlockBegin = Builder.GetInsertBlock()->begin(); 502 // Scanning starts from the last instruction before the insertion point. 503 BasicBlock::iterator IP = Builder.GetInsertPoint(); 504 if (IP != BlockBegin) { 505 --IP; 506 for (; ScanLimit; --IP, --ScanLimit) { 507 // Don't count dbg.value against the ScanLimit, to avoid perturbing the 508 // generated code. 509 if (isa<DbgInfoIntrinsic>(IP)) 510 ScanLimit++; 511 if (IP->getOpcode() == Instruction::GetElementPtr && 512 IP->getOperand(0) == V && IP->getOperand(1) == Idx) 513 return IP; 514 if (IP == BlockBegin) break; 515 } 516 } 517 518 // Save the original insertion point so we can restore it when we're done. 519 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 520 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 521 522 // Move the insertion point out of as many loops as we can. 523 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 524 if (!L->isLoopInvariant(V) || !L->isLoopInvariant(Idx)) break; 525 BasicBlock *Preheader = L->getLoopPreheader(); 526 if (!Preheader) break; 527 528 // Ok, move up a level. 529 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 530 } 531 532 // Emit a GEP. 533 Value *GEP = Builder.CreateGEP(V, Idx, "uglygep"); 534 rememberInstruction(GEP); 535 536 // Restore the original insert point. 537 if (SaveInsertBB) 538 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 539 540 return GEP; 541 } 542 543 // Save the original insertion point so we can restore it when we're done. 544 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 545 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 546 547 // Move the insertion point out of as many loops as we can. 548 while (const Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock())) { 549 if (!L->isLoopInvariant(V)) break; 550 551 bool AnyIndexNotLoopInvariant = false; 552 for (SmallVectorImpl<Value *>::const_iterator I = GepIndices.begin(), 553 E = GepIndices.end(); I != E; ++I) 554 if (!L->isLoopInvariant(*I)) { 555 AnyIndexNotLoopInvariant = true; 556 break; 557 } 558 if (AnyIndexNotLoopInvariant) 559 break; 560 561 BasicBlock *Preheader = L->getLoopPreheader(); 562 if (!Preheader) break; 563 564 // Ok, move up a level. 565 Builder.SetInsertPoint(Preheader, Preheader->getTerminator()); 566 } 567 568 // Insert a pretty getelementptr. Note that this GEP is not marked inbounds, 569 // because ScalarEvolution may have changed the address arithmetic to 570 // compute a value which is beyond the end of the allocated object. 571 Value *Casted = V; 572 if (V->getType() != PTy) 573 Casted = InsertNoopCastOfTo(Casted, PTy); 574 Value *GEP = Builder.CreateGEP(Casted, 575 GepIndices.begin(), 576 GepIndices.end(), 577 "scevgep"); 578 Ops.push_back(SE.getUnknown(GEP)); 579 rememberInstruction(GEP); 580 581 // Restore the original insert point. 582 if (SaveInsertBB) 583 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 584 585 return expand(SE.getAddExpr(Ops)); 586} 587 588/// isNonConstantNegative - Return true if the specified scev is negated, but 589/// not a constant. 590static bool isNonConstantNegative(const SCEV *F) { 591 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(F); 592 if (!Mul) return false; 593 594 // If there is a constant factor, it will be first. 595 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0)); 596 if (!SC) return false; 597 598 // Return true if the value is negative, this matches things like (-42 * V). 599 return SC->getValue()->getValue().isNegative(); 600} 601 602/// PickMostRelevantLoop - Given two loops pick the one that's most relevant for 603/// SCEV expansion. If they are nested, this is the most nested. If they are 604/// neighboring, pick the later. 605static const Loop *PickMostRelevantLoop(const Loop *A, const Loop *B, 606 DominatorTree &DT) { 607 if (!A) return B; 608 if (!B) return A; 609 if (A->contains(B)) return B; 610 if (B->contains(A)) return A; 611 if (DT.dominates(A->getHeader(), B->getHeader())) return B; 612 if (DT.dominates(B->getHeader(), A->getHeader())) return A; 613 return A; // Arbitrarily break the tie. 614} 615 616/// GetRelevantLoop - Get the most relevant loop associated with the given 617/// expression, according to PickMostRelevantLoop. 618static const Loop *GetRelevantLoop(const SCEV *S, LoopInfo &LI, 619 DominatorTree &DT) { 620 if (isa<SCEVConstant>(S)) 621 return 0; 622 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 623 if (const Instruction *I = dyn_cast<Instruction>(U->getValue())) 624 return LI.getLoopFor(I->getParent()); 625 return 0; 626 } 627 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) { 628 const Loop *L = 0; 629 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 630 L = AR->getLoop(); 631 for (SCEVNAryExpr::op_iterator I = N->op_begin(), E = N->op_end(); 632 I != E; ++I) 633 L = PickMostRelevantLoop(L, GetRelevantLoop(*I, LI, DT), DT); 634 return L; 635 } 636 if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) 637 return GetRelevantLoop(C->getOperand(), LI, DT); 638 if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) 639 return PickMostRelevantLoop(GetRelevantLoop(D->getLHS(), LI, DT), 640 GetRelevantLoop(D->getRHS(), LI, DT), 641 DT); 642 llvm_unreachable("Unexpected SCEV type!"); 643} 644 645/// LoopCompare - Compare loops by PickMostRelevantLoop. 646class LoopCompare { 647 DominatorTree &DT; 648public: 649 explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 650 651 bool operator()(std::pair<const Loop *, const SCEV *> LHS, 652 std::pair<const Loop *, const SCEV *> RHS) const { 653 // Compare loops with PickMostRelevantLoop. 654 if (LHS.first != RHS.first) 655 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 656 657 // If one operand is a non-constant negative and the other is not, 658 // put the non-constant negative on the right so that a sub can 659 // be used instead of a negate and add. 660 if (isNonConstantNegative(LHS.second)) { 661 if (!isNonConstantNegative(RHS.second)) 662 return false; 663 } else if (isNonConstantNegative(RHS.second)) 664 return true; 665 666 // Otherwise they are equivalent according to this comparison. 667 return false; 668 } 669}; 670 671Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 672 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 673 674 // Collect all the add operands in a loop, along with their associated loops. 675 // Iterate in reverse so that constants are emitted last, all else equal, and 676 // so that pointer operands are inserted first, which the code below relies on 677 // to form more involved GEPs. 678 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 679 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 680 E(S->op_begin()); I != E; ++I) 681 OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT), 682 *I)); 683 684 // Sort by loop. Use a stable sort so that constants follow non-constants and 685 // pointer operands precede non-pointer operands. 686 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 687 688 // Emit instructions to add all the operands. Hoist as much as possible 689 // out of loops, and form meaningful getelementptrs where possible. 690 Value *Sum = 0; 691 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 692 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 693 const Loop *CurLoop = I->first; 694 const SCEV *Op = I->second; 695 if (!Sum) { 696 // This is the first operand. Just expand it. 697 Sum = expand(Op); 698 ++I; 699 } else if (const PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 700 // The running sum expression is a pointer. Try to form a getelementptr 701 // at this level with that as the base. 702 SmallVector<const SCEV *, 4> NewOps; 703 for (; I != E && I->first == CurLoop; ++I) 704 NewOps.push_back(I->second); 705 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 706 } else if (const PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 707 // The running sum is an integer, and there's a pointer at this level. 708 // Try to form a getelementptr. Use a SCEVUnknown so that we don't 709 // re-analyze the instructions that we just emitted. 710 SmallVector<const SCEV *, 4> NewOps; 711 NewOps.push_back(SE.getUnknown(Sum)); 712 for (++I; I != E && I->first == CurLoop; ++I) 713 NewOps.push_back(I->second); 714 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 715 } else if (isNonConstantNegative(Op)) { 716 // Instead of doing a negate and add, just do a subtract. 717 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 718 Sum = InsertNoopCastOfTo(Sum, Ty); 719 Sum = InsertBinop(Instruction::Sub, Sum, W); 720 ++I; 721 } else { 722 // A simple add. 723 Value *W = expandCodeFor(Op, Ty); 724 Sum = InsertNoopCastOfTo(Sum, Ty); 725 // Canonicalize a constant to the RHS. 726 if (isa<Constant>(Sum)) std::swap(Sum, W); 727 Sum = InsertBinop(Instruction::Add, Sum, W); 728 ++I; 729 } 730 } 731 732 return Sum; 733} 734 735Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 736 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 737 738 // Collect all the mul operands in a loop, along with their associated loops. 739 // Iterate in reverse so that constants are emitted last, all else equal. 740 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 741 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 742 E(S->op_begin()); I != E; ++I) 743 OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT), 744 *I)); 745 746 // Sort by loop. Use a stable sort so that constants follow non-constants. 747 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 748 749 // Emit instructions to mul all the operands. Hoist as much as possible 750 // out of loops. 751 Value *Prod = 0; 752 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 753 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 754 const SCEV *Op = I->second; 755 if (!Prod) { 756 // This is the first operand. Just expand it. 757 Prod = expand(Op); 758 ++I; 759 } else if (Op->isAllOnesValue()) { 760 // Instead of doing a multiply by negative one, just do a negate. 761 Prod = InsertNoopCastOfTo(Prod, Ty); 762 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 763 ++I; 764 } else { 765 // A simple mul. 766 Value *W = expandCodeFor(Op, Ty); 767 Prod = InsertNoopCastOfTo(Prod, Ty); 768 // Canonicalize a constant to the RHS. 769 if (isa<Constant>(Prod)) std::swap(Prod, W); 770 Prod = InsertBinop(Instruction::Mul, Prod, W); 771 ++I; 772 } 773 } 774 775 return Prod; 776} 777 778Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 779 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 780 781 Value *LHS = expandCodeFor(S->getLHS(), Ty); 782 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 783 const APInt &RHS = SC->getValue()->getValue(); 784 if (RHS.isPowerOf2()) 785 return InsertBinop(Instruction::LShr, LHS, 786 ConstantInt::get(Ty, RHS.logBase2())); 787 } 788 789 Value *RHS = expandCodeFor(S->getRHS(), Ty); 790 return InsertBinop(Instruction::UDiv, LHS, RHS); 791} 792 793/// Move parts of Base into Rest to leave Base with the minimal 794/// expression that provides a pointer operand suitable for a 795/// GEP expansion. 796static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 797 ScalarEvolution &SE) { 798 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 799 Base = A->getStart(); 800 Rest = SE.getAddExpr(Rest, 801 SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()), 802 A->getStepRecurrence(SE), 803 A->getLoop())); 804 } 805 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 806 Base = A->getOperand(A->getNumOperands()-1); 807 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 808 NewAddOps.back() = Rest; 809 Rest = SE.getAddExpr(NewAddOps); 810 ExposePointerBase(Base, Rest, SE); 811 } 812} 813 814/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 815/// the base addrec, which is the addrec without any non-loop-dominating 816/// values, and return the PHI. 817PHINode * 818SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 819 const Loop *L, 820 const Type *ExpandTy, 821 const Type *IntTy) { 822 // Reuse a previously-inserted PHI, if present. 823 for (BasicBlock::iterator I = L->getHeader()->begin(); 824 PHINode *PN = dyn_cast<PHINode>(I); ++I) 825 if (SE.isSCEVable(PN->getType()) && 826 (SE.getEffectiveSCEVType(PN->getType()) == 827 SE.getEffectiveSCEVType(Normalized->getType())) && 828 SE.getSCEV(PN) == Normalized) 829 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 830 Instruction *IncV = 831 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 832 833 // Determine if this is a well-behaved chain of instructions leading 834 // back to the PHI. It probably will be, if we're scanning an inner 835 // loop already visited by LSR for example, but it wouldn't have 836 // to be. 837 do { 838 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV)) { 839 IncV = 0; 840 break; 841 } 842 // If any of the operands don't dominate the insert position, bail. 843 // Addrec operands are always loop-invariant, so this can only happen 844 // if there are instructions which haven't been hoisted. 845 for (User::op_iterator OI = IncV->op_begin()+1, 846 OE = IncV->op_end(); OI != OE; ++OI) 847 if (Instruction *OInst = dyn_cast<Instruction>(OI)) 848 if (!SE.DT->dominates(OInst, IVIncInsertPos)) { 849 IncV = 0; 850 break; 851 } 852 if (!IncV) 853 break; 854 // Advance to the next instruction. 855 IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 856 if (!IncV) 857 break; 858 if (IncV->mayHaveSideEffects()) { 859 IncV = 0; 860 break; 861 } 862 } while (IncV != PN); 863 864 if (IncV) { 865 // Ok, the add recurrence looks usable. 866 // Remember this PHI, even in post-inc mode. 867 InsertedValues.insert(PN); 868 // Remember the increment. 869 IncV = cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 870 rememberInstruction(IncV); 871 if (L == IVIncInsertLoop) 872 do { 873 if (SE.DT->dominates(IncV, IVIncInsertPos)) 874 break; 875 // Make sure the increment is where we want it. But don't move it 876 // down past a potential existing post-inc user. 877 IncV->moveBefore(IVIncInsertPos); 878 IVIncInsertPos = IncV; 879 IncV = cast<Instruction>(IncV->getOperand(0)); 880 } while (IncV != PN); 881 return PN; 882 } 883 } 884 885 // Save the original insertion point so we can restore it when we're done. 886 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 887 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 888 889 // Expand code for the start value. 890 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 891 L->getHeader()->begin()); 892 893 // Expand code for the step value. Insert instructions right before the 894 // terminator corresponding to the back-edge. Do this before creating the PHI 895 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is 896 // negative, insert a sub instead of an add for the increment (unless it's a 897 // constant, because subtracts of constants are canonicalized to adds). 898 const SCEV *Step = Normalized->getStepRecurrence(SE); 899 bool isPointer = ExpandTy->isPointerTy(); 900 bool isNegative = !isPointer && isNonConstantNegative(Step); 901 if (isNegative) 902 Step = SE.getNegativeSCEV(Step); 903 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 904 905 // Create the PHI. 906 Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin()); 907 PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv"); 908 rememberInstruction(PN); 909 910 // Create the step instructions and populate the PHI. 911 BasicBlock *Header = L->getHeader(); 912 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header); 913 HPI != HPE; ++HPI) { 914 BasicBlock *Pred = *HPI; 915 916 // Add a start value. 917 if (!L->contains(Pred)) { 918 PN->addIncoming(StartV, Pred); 919 continue; 920 } 921 922 // Create a step value and add it to the PHI. If IVIncInsertLoop is 923 // non-null and equal to the addrec's loop, insert the instructions 924 // at IVIncInsertPos. 925 Instruction *InsertPos = L == IVIncInsertLoop ? 926 IVIncInsertPos : Pred->getTerminator(); 927 Builder.SetInsertPoint(InsertPos->getParent(), InsertPos); 928 Value *IncV; 929 // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 930 if (isPointer) { 931 const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 932 // If the step isn't constant, don't use an implicitly scaled GEP, because 933 // that would require a multiply inside the loop. 934 if (!isa<ConstantInt>(StepV)) 935 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 936 GEPPtrTy->getAddressSpace()); 937 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 938 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 939 if (IncV->getType() != PN->getType()) { 940 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp"); 941 rememberInstruction(IncV); 942 } 943 } else { 944 IncV = isNegative ? 945 Builder.CreateSub(PN, StepV, "lsr.iv.next") : 946 Builder.CreateAdd(PN, StepV, "lsr.iv.next"); 947 rememberInstruction(IncV); 948 } 949 PN->addIncoming(IncV, Pred); 950 } 951 952 // Restore the original insert point. 953 if (SaveInsertBB) 954 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 955 956 // Remember this PHI, even in post-inc mode. 957 InsertedValues.insert(PN); 958 959 return PN; 960} 961 962Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 963 const Type *STy = S->getType(); 964 const Type *IntTy = SE.getEffectiveSCEVType(STy); 965 const Loop *L = S->getLoop(); 966 967 // Determine a normalized form of this expression, which is the expression 968 // before any post-inc adjustment is made. 969 const SCEVAddRecExpr *Normalized = S; 970 if (PostIncLoops.count(L)) { 971 PostIncLoopSet Loops; 972 Loops.insert(L); 973 Normalized = 974 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0, 975 Loops, SE, *SE.DT)); 976 } 977 978 // Strip off any non-loop-dominating component from the addrec start. 979 const SCEV *Start = Normalized->getStart(); 980 const SCEV *PostLoopOffset = 0; 981 if (!Start->properlyDominates(L->getHeader(), SE.DT)) { 982 PostLoopOffset = Start; 983 Start = SE.getIntegerSCEV(0, Normalized->getType()); 984 Normalized = 985 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, 986 Normalized->getStepRecurrence(SE), 987 Normalized->getLoop())); 988 } 989 990 // Strip off any non-loop-dominating component from the addrec step. 991 const SCEV *Step = Normalized->getStepRecurrence(SE); 992 const SCEV *PostLoopScale = 0; 993 if (!Step->hasComputableLoopEvolution(L) && 994 !Step->dominates(L->getHeader(), SE.DT)) { 995 PostLoopScale = Step; 996 Step = SE.getIntegerSCEV(1, Normalized->getType()); 997 Normalized = 998 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step, 999 Normalized->getLoop())); 1000 } 1001 1002 // Expand the core addrec. If we need post-loop scaling, force it to 1003 // expand to an integer type to avoid the need for additional casting. 1004 const Type *ExpandTy = PostLoopScale ? IntTy : STy; 1005 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy); 1006 1007 // Accommodate post-inc mode, if necessary. 1008 Value *Result; 1009 if (!PostIncLoops.count(L)) 1010 Result = PN; 1011 else { 1012 // In PostInc mode, use the post-incremented value. 1013 BasicBlock *LatchBlock = L->getLoopLatch(); 1014 assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1015 Result = PN->getIncomingValueForBlock(LatchBlock); 1016 } 1017 1018 // Re-apply any non-loop-dominating scale. 1019 if (PostLoopScale) { 1020 Result = InsertNoopCastOfTo(Result, IntTy); 1021 Result = Builder.CreateMul(Result, 1022 expandCodeFor(PostLoopScale, IntTy)); 1023 rememberInstruction(Result); 1024 } 1025 1026 // Re-apply any non-loop-dominating offset. 1027 if (PostLoopOffset) { 1028 if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1029 const SCEV *const OffsetArray[1] = { PostLoopOffset }; 1030 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 1031 } else { 1032 Result = InsertNoopCastOfTo(Result, IntTy); 1033 Result = Builder.CreateAdd(Result, 1034 expandCodeFor(PostLoopOffset, IntTy)); 1035 rememberInstruction(Result); 1036 } 1037 } 1038 1039 return Result; 1040} 1041 1042Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1043 if (!CanonicalMode) return expandAddRecExprLiterally(S); 1044 1045 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1046 const Loop *L = S->getLoop(); 1047 1048 // First check for an existing canonical IV in a suitable type. 1049 PHINode *CanonicalIV = 0; 1050 if (PHINode *PN = L->getCanonicalInductionVariable()) 1051 if (SE.isSCEVable(PN->getType()) && 1052 SE.getEffectiveSCEVType(PN->getType())->isIntegerTy() && 1053 SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 1054 CanonicalIV = PN; 1055 1056 // Rewrite an AddRec in terms of the canonical induction variable, if 1057 // its type is more narrow. 1058 if (CanonicalIV && 1059 SE.getTypeSizeInBits(CanonicalIV->getType()) > 1060 SE.getTypeSizeInBits(Ty)) { 1061 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1062 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1063 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 1064 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop())); 1065 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1066 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 1067 BasicBlock::iterator NewInsertPt = 1068 llvm::next(BasicBlock::iterator(cast<Instruction>(V))); 1069 while (isa<PHINode>(NewInsertPt)) ++NewInsertPt; 1070 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0, 1071 NewInsertPt); 1072 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 1073 return V; 1074 } 1075 1076 // {X,+,F} --> X + {0,+,F} 1077 if (!S->getStart()->isZero()) { 1078 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1079 NewOps[0] = SE.getIntegerSCEV(0, Ty); 1080 const SCEV *Rest = SE.getAddRecExpr(NewOps, L); 1081 1082 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1083 // comments on expandAddToGEP for details. 1084 const SCEV *Base = S->getStart(); 1085 const SCEV *RestArray[1] = { Rest }; 1086 // Dig into the expression to find the pointer base for a GEP. 1087 ExposePointerBase(Base, RestArray[0], SE); 1088 // If we found a pointer, expand the AddRec with a GEP. 1089 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1090 // Make sure the Base isn't something exotic, such as a multiplied 1091 // or divided pointer value. In those cases, the result type isn't 1092 // actually a pointer type. 1093 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1094 Value *StartV = expand(Base); 1095 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1096 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 1097 } 1098 } 1099 1100 // Just do a normal add. Pre-expand the operands to suppress folding. 1101 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), 1102 SE.getUnknown(expand(Rest)))); 1103 } 1104 1105 // {0,+,1} --> Insert a canonical induction variable into the loop! 1106 if (S->isAffine() && 1107 S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) { 1108 // If there's a canonical IV, just use it. 1109 if (CanonicalIV) { 1110 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 1111 "IVs with types different from the canonical IV should " 1112 "already have been handled!"); 1113 return CanonicalIV; 1114 } 1115 1116 // Create and insert the PHI node for the induction variable in the 1117 // specified loop. 1118 BasicBlock *Header = L->getHeader(); 1119 PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin()); 1120 rememberInstruction(PN); 1121 1122 Constant *One = ConstantInt::get(Ty, 1); 1123 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header); 1124 HPI != HPE; ++HPI) 1125 if (L->contains(*HPI)) { 1126 // Insert a unit add instruction right before the terminator 1127 // corresponding to the back-edge. 1128 Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next", 1129 (*HPI)->getTerminator()); 1130 rememberInstruction(Add); 1131 PN->addIncoming(Add, *HPI); 1132 } else { 1133 PN->addIncoming(Constant::getNullValue(Ty), *HPI); 1134 } 1135 } 1136 1137 // {0,+,F} --> {0,+,1} * F 1138 // Get the canonical induction variable I for this loop. 1139 Value *I = CanonicalIV ? 1140 CanonicalIV : 1141 getOrInsertCanonicalInductionVariable(L, Ty); 1142 1143 // If this is a simple linear addrec, emit it now as a special case. 1144 if (S->isAffine()) // {0,+,F} --> i*F 1145 return 1146 expand(SE.getTruncateOrNoop( 1147 SE.getMulExpr(SE.getUnknown(I), 1148 SE.getNoopOrAnyExtend(S->getOperand(1), 1149 I->getType())), 1150 Ty)); 1151 1152 // If this is a chain of recurrences, turn it into a closed form, using the 1153 // folders, then expandCodeFor the closed form. This allows the folders to 1154 // simplify the expression without having to build a bunch of special code 1155 // into this folder. 1156 const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV. 1157 1158 // Promote S up to the canonical IV type, if the cast is foldable. 1159 const SCEV *NewS = S; 1160 const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType()); 1161 if (isa<SCEVAddRecExpr>(Ext)) 1162 NewS = Ext; 1163 1164 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1165 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 1166 1167 // Truncate the result down to the original type, if needed. 1168 const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1169 return expand(T); 1170} 1171 1172Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1173 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1174 Value *V = expandCodeFor(S->getOperand(), 1175 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1176 Value *I = Builder.CreateTrunc(V, Ty, "tmp"); 1177 rememberInstruction(I); 1178 return I; 1179} 1180 1181Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1182 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1183 Value *V = expandCodeFor(S->getOperand(), 1184 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1185 Value *I = Builder.CreateZExt(V, Ty, "tmp"); 1186 rememberInstruction(I); 1187 return I; 1188} 1189 1190Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1191 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1192 Value *V = expandCodeFor(S->getOperand(), 1193 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1194 Value *I = Builder.CreateSExt(V, Ty, "tmp"); 1195 rememberInstruction(I); 1196 return I; 1197} 1198 1199Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 1200 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1201 const Type *Ty = LHS->getType(); 1202 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1203 // In the case of mixed integer and pointer types, do the 1204 // rest of the comparisons as integer. 1205 if (S->getOperand(i)->getType() != Ty) { 1206 Ty = SE.getEffectiveSCEVType(Ty); 1207 LHS = InsertNoopCastOfTo(LHS, Ty); 1208 } 1209 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1210 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp"); 1211 rememberInstruction(ICmp); 1212 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1213 rememberInstruction(Sel); 1214 LHS = Sel; 1215 } 1216 // In the case of mixed integer and pointer types, cast the 1217 // final result back to the pointer type. 1218 if (LHS->getType() != S->getType()) 1219 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1220 return LHS; 1221} 1222 1223Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 1224 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1225 const Type *Ty = LHS->getType(); 1226 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1227 // In the case of mixed integer and pointer types, do the 1228 // rest of the comparisons as integer. 1229 if (S->getOperand(i)->getType() != Ty) { 1230 Ty = SE.getEffectiveSCEVType(Ty); 1231 LHS = InsertNoopCastOfTo(LHS, Ty); 1232 } 1233 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1234 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp"); 1235 rememberInstruction(ICmp); 1236 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1237 rememberInstruction(Sel); 1238 LHS = Sel; 1239 } 1240 // In the case of mixed integer and pointer types, cast the 1241 // final result back to the pointer type. 1242 if (LHS->getType() != S->getType()) 1243 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1244 return LHS; 1245} 1246 1247Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty, 1248 Instruction *I) { 1249 BasicBlock::iterator IP = I; 1250 while (isInsertedInstruction(IP) || isa<DbgInfoIntrinsic>(IP)) 1251 ++IP; 1252 Builder.SetInsertPoint(IP->getParent(), IP); 1253 return expandCodeFor(SH, Ty); 1254} 1255 1256Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) { 1257 // Expand the code for this SCEV. 1258 Value *V = expand(SH); 1259 if (Ty) { 1260 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 1261 "non-trivial casts should be done with the SCEVs directly!"); 1262 V = InsertNoopCastOfTo(V, Ty); 1263 } 1264 return V; 1265} 1266 1267Value *SCEVExpander::expand(const SCEV *S) { 1268 // Compute an insertion point for this SCEV object. Hoist the instructions 1269 // as far out in the loop nest as possible. 1270 Instruction *InsertPt = Builder.GetInsertPoint(); 1271 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ; 1272 L = L->getParentLoop()) 1273 if (S->isLoopInvariant(L)) { 1274 if (!L) break; 1275 if (BasicBlock *Preheader = L->getLoopPreheader()) 1276 InsertPt = Preheader->getTerminator(); 1277 } else { 1278 // If the SCEV is computable at this level, insert it into the header 1279 // after the PHIs (and after any other instructions that we've inserted 1280 // there) so that it is guaranteed to dominate any user inside the loop. 1281 if (L && S->hasComputableLoopEvolution(L) && !PostIncLoops.count(L)) 1282 InsertPt = L->getHeader()->getFirstNonPHI(); 1283 while (isInsertedInstruction(InsertPt) || isa<DbgInfoIntrinsic>(InsertPt)) 1284 InsertPt = llvm::next(BasicBlock::iterator(InsertPt)); 1285 break; 1286 } 1287 1288 // Check to see if we already expanded this here. 1289 std::map<std::pair<const SCEV *, Instruction *>, 1290 AssertingVH<Value> >::iterator I = 1291 InsertedExpressions.find(std::make_pair(S, InsertPt)); 1292 if (I != InsertedExpressions.end()) 1293 return I->second; 1294 1295 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1296 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 1297 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt); 1298 1299 // Expand the expression into instructions. 1300 Value *V = visit(S); 1301 1302 // Remember the expanded value for this SCEV at this location. 1303 if (PostIncLoops.empty()) 1304 InsertedExpressions[std::make_pair(S, InsertPt)] = V; 1305 1306 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 1307 return V; 1308} 1309 1310void SCEVExpander::rememberInstruction(Value *I) { 1311 if (PostIncLoops.empty()) 1312 InsertedValues.insert(I); 1313 1314 // If we just claimed an existing instruction and that instruction had 1315 // been the insert point, adjust the insert point forward so that 1316 // subsequently inserted code will be dominated. 1317 if (Builder.GetInsertPoint() == I) { 1318 BasicBlock::iterator It = cast<Instruction>(I); 1319 do { ++It; } while (isInsertedInstruction(It) || 1320 isa<DbgInfoIntrinsic>(It)); 1321 Builder.SetInsertPoint(Builder.GetInsertBlock(), It); 1322 } 1323} 1324 1325void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) { 1326 // If we acquired more instructions since the old insert point was saved, 1327 // advance past them. 1328 while (isInsertedInstruction(I) || isa<DbgInfoIntrinsic>(I)) ++I; 1329 1330 Builder.SetInsertPoint(BB, I); 1331} 1332 1333/// getOrInsertCanonicalInductionVariable - This method returns the 1334/// canonical induction variable of the specified type for the specified 1335/// loop (inserting one if there is none). A canonical induction variable 1336/// starts at zero and steps by one on each iteration. 1337Value * 1338SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1339 const Type *Ty) { 1340 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1341 const SCEV *H = SE.getAddRecExpr(SE.getIntegerSCEV(0, Ty), 1342 SE.getIntegerSCEV(1, Ty), L); 1343 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1344 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 1345 Value *V = expandCodeFor(H, 0, L->getHeader()->begin()); 1346 if (SaveInsertBB) 1347 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 1348 return V; 1349} 1350