ScalarEvolutionExpander.cpp revision deff621abdd48bd70434bd4d7ef30f08ddba1cd8
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.getConstant(S->getType(), 1); 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.getConstant(SOp->getType(), 0); 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.getConstant(Step->getType(), 0); 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.getConstant(Ty, 0) : 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.getConstant(Ty, 0); 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.getConstant(Ty, 0); 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 645namespace { 646 647/// LoopCompare - Compare loops by PickMostRelevantLoop. 648class LoopCompare { 649 DominatorTree &DT; 650public: 651 explicit LoopCompare(DominatorTree &dt) : DT(dt) {} 652 653 bool operator()(std::pair<const Loop *, const SCEV *> LHS, 654 std::pair<const Loop *, const SCEV *> RHS) const { 655 // Compare loops with PickMostRelevantLoop. 656 if (LHS.first != RHS.first) 657 return PickMostRelevantLoop(LHS.first, RHS.first, DT) != LHS.first; 658 659 // If one operand is a non-constant negative and the other is not, 660 // put the non-constant negative on the right so that a sub can 661 // be used instead of a negate and add. 662 if (isNonConstantNegative(LHS.second)) { 663 if (!isNonConstantNegative(RHS.second)) 664 return false; 665 } else if (isNonConstantNegative(RHS.second)) 666 return true; 667 668 // Otherwise they are equivalent according to this comparison. 669 return false; 670 } 671}; 672 673} 674 675Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 676 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 677 678 // Collect all the add operands in a loop, along with their associated loops. 679 // Iterate in reverse so that constants are emitted last, all else equal, and 680 // so that pointer operands are inserted first, which the code below relies on 681 // to form more involved GEPs. 682 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 683 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(S->op_end()), 684 E(S->op_begin()); I != E; ++I) 685 OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT), 686 *I)); 687 688 // Sort by loop. Use a stable sort so that constants follow non-constants and 689 // pointer operands precede non-pointer operands. 690 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 691 692 // Emit instructions to add all the operands. Hoist as much as possible 693 // out of loops, and form meaningful getelementptrs where possible. 694 Value *Sum = 0; 695 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 696 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 697 const Loop *CurLoop = I->first; 698 const SCEV *Op = I->second; 699 if (!Sum) { 700 // This is the first operand. Just expand it. 701 Sum = expand(Op); 702 ++I; 703 } else if (const PointerType *PTy = dyn_cast<PointerType>(Sum->getType())) { 704 // The running sum expression is a pointer. Try to form a getelementptr 705 // at this level with that as the base. 706 SmallVector<const SCEV *, 4> NewOps; 707 for (; I != E && I->first == CurLoop; ++I) 708 NewOps.push_back(I->second); 709 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, Sum); 710 } else if (const PointerType *PTy = dyn_cast<PointerType>(Op->getType())) { 711 // The running sum is an integer, and there's a pointer at this level. 712 // Try to form a getelementptr. If the running sum is instructions, 713 // use a SCEVUnknown to avoid re-analyzing them. 714 SmallVector<const SCEV *, 4> NewOps; 715 NewOps.push_back(isa<Instruction>(Sum) ? SE.getUnknown(Sum) : 716 SE.getSCEV(Sum)); 717 for (++I; I != E && I->first == CurLoop; ++I) 718 NewOps.push_back(I->second); 719 Sum = expandAddToGEP(NewOps.begin(), NewOps.end(), PTy, Ty, expand(Op)); 720 } else if (isNonConstantNegative(Op)) { 721 // Instead of doing a negate and add, just do a subtract. 722 Value *W = expandCodeFor(SE.getNegativeSCEV(Op), Ty); 723 Sum = InsertNoopCastOfTo(Sum, Ty); 724 Sum = InsertBinop(Instruction::Sub, Sum, W); 725 ++I; 726 } else { 727 // A simple add. 728 Value *W = expandCodeFor(Op, Ty); 729 Sum = InsertNoopCastOfTo(Sum, Ty); 730 // Canonicalize a constant to the RHS. 731 if (isa<Constant>(Sum)) std::swap(Sum, W); 732 Sum = InsertBinop(Instruction::Add, Sum, W); 733 ++I; 734 } 735 } 736 737 return Sum; 738} 739 740Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 741 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 742 743 // Collect all the mul operands in a loop, along with their associated loops. 744 // Iterate in reverse so that constants are emitted last, all else equal. 745 SmallVector<std::pair<const Loop *, const SCEV *>, 8> OpsAndLoops; 746 for (std::reverse_iterator<SCEVMulExpr::op_iterator> I(S->op_end()), 747 E(S->op_begin()); I != E; ++I) 748 OpsAndLoops.push_back(std::make_pair(GetRelevantLoop(*I, *SE.LI, *SE.DT), 749 *I)); 750 751 // Sort by loop. Use a stable sort so that constants follow non-constants. 752 std::stable_sort(OpsAndLoops.begin(), OpsAndLoops.end(), LoopCompare(*SE.DT)); 753 754 // Emit instructions to mul all the operands. Hoist as much as possible 755 // out of loops. 756 Value *Prod = 0; 757 for (SmallVectorImpl<std::pair<const Loop *, const SCEV *> >::iterator 758 I = OpsAndLoops.begin(), E = OpsAndLoops.end(); I != E; ) { 759 const SCEV *Op = I->second; 760 if (!Prod) { 761 // This is the first operand. Just expand it. 762 Prod = expand(Op); 763 ++I; 764 } else if (Op->isAllOnesValue()) { 765 // Instead of doing a multiply by negative one, just do a negate. 766 Prod = InsertNoopCastOfTo(Prod, Ty); 767 Prod = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), Prod); 768 ++I; 769 } else { 770 // A simple mul. 771 Value *W = expandCodeFor(Op, Ty); 772 Prod = InsertNoopCastOfTo(Prod, Ty); 773 // Canonicalize a constant to the RHS. 774 if (isa<Constant>(Prod)) std::swap(Prod, W); 775 Prod = InsertBinop(Instruction::Mul, Prod, W); 776 ++I; 777 } 778 } 779 780 return Prod; 781} 782 783Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 784 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 785 786 Value *LHS = expandCodeFor(S->getLHS(), Ty); 787 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 788 const APInt &RHS = SC->getValue()->getValue(); 789 if (RHS.isPowerOf2()) 790 return InsertBinop(Instruction::LShr, LHS, 791 ConstantInt::get(Ty, RHS.logBase2())); 792 } 793 794 Value *RHS = expandCodeFor(S->getRHS(), Ty); 795 return InsertBinop(Instruction::UDiv, LHS, RHS); 796} 797 798/// Move parts of Base into Rest to leave Base with the minimal 799/// expression that provides a pointer operand suitable for a 800/// GEP expansion. 801static void ExposePointerBase(const SCEV *&Base, const SCEV *&Rest, 802 ScalarEvolution &SE) { 803 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 804 Base = A->getStart(); 805 Rest = SE.getAddExpr(Rest, 806 SE.getAddRecExpr(SE.getConstant(A->getType(), 0), 807 A->getStepRecurrence(SE), 808 A->getLoop())); 809 } 810 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 811 Base = A->getOperand(A->getNumOperands()-1); 812 SmallVector<const SCEV *, 8> NewAddOps(A->op_begin(), A->op_end()); 813 NewAddOps.back() = Rest; 814 Rest = SE.getAddExpr(NewAddOps); 815 ExposePointerBase(Base, Rest, SE); 816 } 817} 818 819/// getAddRecExprPHILiterally - Helper for expandAddRecExprLiterally. Expand 820/// the base addrec, which is the addrec without any non-loop-dominating 821/// values, and return the PHI. 822PHINode * 823SCEVExpander::getAddRecExprPHILiterally(const SCEVAddRecExpr *Normalized, 824 const Loop *L, 825 const Type *ExpandTy, 826 const Type *IntTy) { 827 // Reuse a previously-inserted PHI, if present. 828 for (BasicBlock::iterator I = L->getHeader()->begin(); 829 PHINode *PN = dyn_cast<PHINode>(I); ++I) 830 if (SE.isSCEVable(PN->getType()) && 831 (SE.getEffectiveSCEVType(PN->getType()) == 832 SE.getEffectiveSCEVType(Normalized->getType())) && 833 SE.getSCEV(PN) == Normalized) 834 if (BasicBlock *LatchBlock = L->getLoopLatch()) { 835 Instruction *IncV = 836 cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 837 838 // Determine if this is a well-behaved chain of instructions leading 839 // back to the PHI. It probably will be, if we're scanning an inner 840 // loop already visited by LSR for example, but it wouldn't have 841 // to be. 842 do { 843 if (IncV->getNumOperands() == 0 || isa<PHINode>(IncV)) { 844 IncV = 0; 845 break; 846 } 847 // If any of the operands don't dominate the insert position, bail. 848 // Addrec operands are always loop-invariant, so this can only happen 849 // if there are instructions which haven't been hoisted. 850 for (User::op_iterator OI = IncV->op_begin()+1, 851 OE = IncV->op_end(); OI != OE; ++OI) 852 if (Instruction *OInst = dyn_cast<Instruction>(OI)) 853 if (!SE.DT->dominates(OInst, IVIncInsertPos)) { 854 IncV = 0; 855 break; 856 } 857 if (!IncV) 858 break; 859 // Advance to the next instruction. 860 IncV = dyn_cast<Instruction>(IncV->getOperand(0)); 861 if (!IncV) 862 break; 863 if (IncV->mayHaveSideEffects()) { 864 IncV = 0; 865 break; 866 } 867 } while (IncV != PN); 868 869 if (IncV) { 870 // Ok, the add recurrence looks usable. 871 // Remember this PHI, even in post-inc mode. 872 InsertedValues.insert(PN); 873 // Remember the increment. 874 IncV = cast<Instruction>(PN->getIncomingValueForBlock(LatchBlock)); 875 rememberInstruction(IncV); 876 if (L == IVIncInsertLoop) 877 do { 878 if (SE.DT->dominates(IncV, IVIncInsertPos)) 879 break; 880 // Make sure the increment is where we want it. But don't move it 881 // down past a potential existing post-inc user. 882 IncV->moveBefore(IVIncInsertPos); 883 IVIncInsertPos = IncV; 884 IncV = cast<Instruction>(IncV->getOperand(0)); 885 } while (IncV != PN); 886 return PN; 887 } 888 } 889 890 // Save the original insertion point so we can restore it when we're done. 891 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 892 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 893 894 // Expand code for the start value. 895 Value *StartV = expandCodeFor(Normalized->getStart(), ExpandTy, 896 L->getHeader()->begin()); 897 898 // Expand code for the step value. Insert instructions right before the 899 // terminator corresponding to the back-edge. Do this before creating the PHI 900 // so that PHI reuse code doesn't see an incomplete PHI. If the stride is 901 // negative, insert a sub instead of an add for the increment (unless it's a 902 // constant, because subtracts of constants are canonicalized to adds). 903 const SCEV *Step = Normalized->getStepRecurrence(SE); 904 bool isPointer = ExpandTy->isPointerTy(); 905 bool isNegative = !isPointer && isNonConstantNegative(Step); 906 if (isNegative) 907 Step = SE.getNegativeSCEV(Step); 908 Value *StepV = expandCodeFor(Step, IntTy, L->getHeader()->begin()); 909 910 // Create the PHI. 911 Builder.SetInsertPoint(L->getHeader(), L->getHeader()->begin()); 912 PHINode *PN = Builder.CreatePHI(ExpandTy, "lsr.iv"); 913 rememberInstruction(PN); 914 915 // Create the step instructions and populate the PHI. 916 BasicBlock *Header = L->getHeader(); 917 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header); 918 HPI != HPE; ++HPI) { 919 BasicBlock *Pred = *HPI; 920 921 // Add a start value. 922 if (!L->contains(Pred)) { 923 PN->addIncoming(StartV, Pred); 924 continue; 925 } 926 927 // Create a step value and add it to the PHI. If IVIncInsertLoop is 928 // non-null and equal to the addrec's loop, insert the instructions 929 // at IVIncInsertPos. 930 Instruction *InsertPos = L == IVIncInsertLoop ? 931 IVIncInsertPos : Pred->getTerminator(); 932 Builder.SetInsertPoint(InsertPos->getParent(), InsertPos); 933 Value *IncV; 934 // If the PHI is a pointer, use a GEP, otherwise use an add or sub. 935 if (isPointer) { 936 const PointerType *GEPPtrTy = cast<PointerType>(ExpandTy); 937 // If the step isn't constant, don't use an implicitly scaled GEP, because 938 // that would require a multiply inside the loop. 939 if (!isa<ConstantInt>(StepV)) 940 GEPPtrTy = PointerType::get(Type::getInt1Ty(SE.getContext()), 941 GEPPtrTy->getAddressSpace()); 942 const SCEV *const StepArray[1] = { SE.getSCEV(StepV) }; 943 IncV = expandAddToGEP(StepArray, StepArray+1, GEPPtrTy, IntTy, PN); 944 if (IncV->getType() != PN->getType()) { 945 IncV = Builder.CreateBitCast(IncV, PN->getType(), "tmp"); 946 rememberInstruction(IncV); 947 } 948 } else { 949 IncV = isNegative ? 950 Builder.CreateSub(PN, StepV, "lsr.iv.next") : 951 Builder.CreateAdd(PN, StepV, "lsr.iv.next"); 952 rememberInstruction(IncV); 953 } 954 PN->addIncoming(IncV, Pred); 955 } 956 957 // Restore the original insert point. 958 if (SaveInsertBB) 959 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 960 961 // Remember this PHI, even in post-inc mode. 962 InsertedValues.insert(PN); 963 964 return PN; 965} 966 967Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) { 968 const Type *STy = S->getType(); 969 const Type *IntTy = SE.getEffectiveSCEVType(STy); 970 const Loop *L = S->getLoop(); 971 972 // Determine a normalized form of this expression, which is the expression 973 // before any post-inc adjustment is made. 974 const SCEVAddRecExpr *Normalized = S; 975 if (PostIncLoops.count(L)) { 976 PostIncLoopSet Loops; 977 Loops.insert(L); 978 Normalized = 979 cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0, 980 Loops, SE, *SE.DT)); 981 } 982 983 // Strip off any non-loop-dominating component from the addrec start. 984 const SCEV *Start = Normalized->getStart(); 985 const SCEV *PostLoopOffset = 0; 986 if (!Start->properlyDominates(L->getHeader(), SE.DT)) { 987 PostLoopOffset = Start; 988 Start = SE.getConstant(Normalized->getType(), 0); 989 Normalized = 990 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, 991 Normalized->getStepRecurrence(SE), 992 Normalized->getLoop())); 993 } 994 995 // Strip off any non-loop-dominating component from the addrec step. 996 const SCEV *Step = Normalized->getStepRecurrence(SE); 997 const SCEV *PostLoopScale = 0; 998 if (!Step->dominates(L->getHeader(), SE.DT)) { 999 PostLoopScale = Step; 1000 Step = SE.getConstant(Normalized->getType(), 1); 1001 Normalized = 1002 cast<SCEVAddRecExpr>(SE.getAddRecExpr(Start, Step, 1003 Normalized->getLoop())); 1004 } 1005 1006 // Expand the core addrec. If we need post-loop scaling, force it to 1007 // expand to an integer type to avoid the need for additional casting. 1008 const Type *ExpandTy = PostLoopScale ? IntTy : STy; 1009 PHINode *PN = getAddRecExprPHILiterally(Normalized, L, ExpandTy, IntTy); 1010 1011 // Accommodate post-inc mode, if necessary. 1012 Value *Result; 1013 if (!PostIncLoops.count(L)) 1014 Result = PN; 1015 else { 1016 // In PostInc mode, use the post-incremented value. 1017 BasicBlock *LatchBlock = L->getLoopLatch(); 1018 assert(LatchBlock && "PostInc mode requires a unique loop latch!"); 1019 Result = PN->getIncomingValueForBlock(LatchBlock); 1020 } 1021 1022 // Re-apply any non-loop-dominating scale. 1023 if (PostLoopScale) { 1024 Result = InsertNoopCastOfTo(Result, IntTy); 1025 Result = Builder.CreateMul(Result, 1026 expandCodeFor(PostLoopScale, IntTy)); 1027 rememberInstruction(Result); 1028 } 1029 1030 // Re-apply any non-loop-dominating offset. 1031 if (PostLoopOffset) { 1032 if (const PointerType *PTy = dyn_cast<PointerType>(ExpandTy)) { 1033 const SCEV *const OffsetArray[1] = { PostLoopOffset }; 1034 Result = expandAddToGEP(OffsetArray, OffsetArray+1, PTy, IntTy, Result); 1035 } else { 1036 Result = InsertNoopCastOfTo(Result, IntTy); 1037 Result = Builder.CreateAdd(Result, 1038 expandCodeFor(PostLoopOffset, IntTy)); 1039 rememberInstruction(Result); 1040 } 1041 } 1042 1043 return Result; 1044} 1045 1046Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 1047 if (!CanonicalMode) return expandAddRecExprLiterally(S); 1048 1049 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1050 const Loop *L = S->getLoop(); 1051 1052 // First check for an existing canonical IV in a suitable type. 1053 PHINode *CanonicalIV = 0; 1054 if (PHINode *PN = L->getCanonicalInductionVariable()) 1055 if (SE.isSCEVable(PN->getType()) && 1056 SE.getEffectiveSCEVType(PN->getType())->isIntegerTy() && 1057 SE.getTypeSizeInBits(PN->getType()) >= SE.getTypeSizeInBits(Ty)) 1058 CanonicalIV = PN; 1059 1060 // Rewrite an AddRec in terms of the canonical induction variable, if 1061 // its type is more narrow. 1062 if (CanonicalIV && 1063 SE.getTypeSizeInBits(CanonicalIV->getType()) > 1064 SE.getTypeSizeInBits(Ty)) { 1065 SmallVector<const SCEV *, 4> NewOps(S->getNumOperands()); 1066 for (unsigned i = 0, e = S->getNumOperands(); i != e; ++i) 1067 NewOps[i] = SE.getAnyExtendExpr(S->op_begin()[i], CanonicalIV->getType()); 1068 Value *V = expand(SE.getAddRecExpr(NewOps, S->getLoop())); 1069 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1070 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 1071 BasicBlock::iterator NewInsertPt = 1072 llvm::next(BasicBlock::iterator(cast<Instruction>(V))); 1073 while (isa<PHINode>(NewInsertPt)) ++NewInsertPt; 1074 V = expandCodeFor(SE.getTruncateExpr(SE.getUnknown(V), Ty), 0, 1075 NewInsertPt); 1076 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 1077 return V; 1078 } 1079 1080 // {X,+,F} --> X + {0,+,F} 1081 if (!S->getStart()->isZero()) { 1082 SmallVector<const SCEV *, 4> NewOps(S->op_begin(), S->op_end()); 1083 NewOps[0] = SE.getConstant(Ty, 0); 1084 const SCEV *Rest = SE.getAddRecExpr(NewOps, L); 1085 1086 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 1087 // comments on expandAddToGEP for details. 1088 const SCEV *Base = S->getStart(); 1089 const SCEV *RestArray[1] = { Rest }; 1090 // Dig into the expression to find the pointer base for a GEP. 1091 ExposePointerBase(Base, RestArray[0], SE); 1092 // If we found a pointer, expand the AddRec with a GEP. 1093 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 1094 // Make sure the Base isn't something exotic, such as a multiplied 1095 // or divided pointer value. In those cases, the result type isn't 1096 // actually a pointer type. 1097 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 1098 Value *StartV = expand(Base); 1099 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 1100 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 1101 } 1102 } 1103 1104 // Just do a normal add. Pre-expand the operands to suppress folding. 1105 return expand(SE.getAddExpr(SE.getUnknown(expand(S->getStart())), 1106 SE.getUnknown(expand(Rest)))); 1107 } 1108 1109 // {0,+,1} --> Insert a canonical induction variable into the loop! 1110 if (S->isAffine() && 1111 S->getOperand(1) == SE.getConstant(Ty, 1)) { 1112 // If there's a canonical IV, just use it. 1113 if (CanonicalIV) { 1114 assert(Ty == SE.getEffectiveSCEVType(CanonicalIV->getType()) && 1115 "IVs with types different from the canonical IV should " 1116 "already have been handled!"); 1117 return CanonicalIV; 1118 } 1119 1120 // Create and insert the PHI node for the induction variable in the 1121 // specified loop. 1122 BasicBlock *Header = L->getHeader(); 1123 PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin()); 1124 rememberInstruction(PN); 1125 1126 Constant *One = ConstantInt::get(Ty, 1); 1127 for (pred_iterator HPI = pred_begin(Header), HPE = pred_end(Header); 1128 HPI != HPE; ++HPI) 1129 if (L->contains(*HPI)) { 1130 // Insert a unit add instruction right before the terminator 1131 // corresponding to the back-edge. 1132 Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next", 1133 (*HPI)->getTerminator()); 1134 rememberInstruction(Add); 1135 PN->addIncoming(Add, *HPI); 1136 } else { 1137 PN->addIncoming(Constant::getNullValue(Ty), *HPI); 1138 } 1139 } 1140 1141 // {0,+,F} --> {0,+,1} * F 1142 // Get the canonical induction variable I for this loop. 1143 Value *I = CanonicalIV ? 1144 CanonicalIV : 1145 getOrInsertCanonicalInductionVariable(L, Ty); 1146 1147 // If this is a simple linear addrec, emit it now as a special case. 1148 if (S->isAffine()) // {0,+,F} --> i*F 1149 return 1150 expand(SE.getTruncateOrNoop( 1151 SE.getMulExpr(SE.getUnknown(I), 1152 SE.getNoopOrAnyExtend(S->getOperand(1), 1153 I->getType())), 1154 Ty)); 1155 1156 // If this is a chain of recurrences, turn it into a closed form, using the 1157 // folders, then expandCodeFor the closed form. This allows the folders to 1158 // simplify the expression without having to build a bunch of special code 1159 // into this folder. 1160 const SCEV *IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV. 1161 1162 // Promote S up to the canonical IV type, if the cast is foldable. 1163 const SCEV *NewS = S; 1164 const SCEV *Ext = SE.getNoopOrAnyExtend(S, I->getType()); 1165 if (isa<SCEVAddRecExpr>(Ext)) 1166 NewS = Ext; 1167 1168 const SCEV *V = cast<SCEVAddRecExpr>(NewS)->evaluateAtIteration(IH, SE); 1169 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 1170 1171 // Truncate the result down to the original type, if needed. 1172 const SCEV *T = SE.getTruncateOrNoop(V, Ty); 1173 return expand(T); 1174} 1175 1176Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 1177 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1178 Value *V = expandCodeFor(S->getOperand(), 1179 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1180 Value *I = Builder.CreateTrunc(V, Ty, "tmp"); 1181 rememberInstruction(I); 1182 return I; 1183} 1184 1185Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 1186 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1187 Value *V = expandCodeFor(S->getOperand(), 1188 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1189 Value *I = Builder.CreateZExt(V, Ty, "tmp"); 1190 rememberInstruction(I); 1191 return I; 1192} 1193 1194Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 1195 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 1196 Value *V = expandCodeFor(S->getOperand(), 1197 SE.getEffectiveSCEVType(S->getOperand()->getType())); 1198 Value *I = Builder.CreateSExt(V, Ty, "tmp"); 1199 rememberInstruction(I); 1200 return I; 1201} 1202 1203Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 1204 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1205 const Type *Ty = LHS->getType(); 1206 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1207 // In the case of mixed integer and pointer types, do the 1208 // rest of the comparisons as integer. 1209 if (S->getOperand(i)->getType() != Ty) { 1210 Ty = SE.getEffectiveSCEVType(Ty); 1211 LHS = InsertNoopCastOfTo(LHS, Ty); 1212 } 1213 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1214 Value *ICmp = Builder.CreateICmpSGT(LHS, RHS, "tmp"); 1215 rememberInstruction(ICmp); 1216 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "smax"); 1217 rememberInstruction(Sel); 1218 LHS = Sel; 1219 } 1220 // In the case of mixed integer and pointer types, cast the 1221 // final result back to the pointer type. 1222 if (LHS->getType() != S->getType()) 1223 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1224 return LHS; 1225} 1226 1227Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 1228 Value *LHS = expand(S->getOperand(S->getNumOperands()-1)); 1229 const Type *Ty = LHS->getType(); 1230 for (int i = S->getNumOperands()-2; i >= 0; --i) { 1231 // In the case of mixed integer and pointer types, do the 1232 // rest of the comparisons as integer. 1233 if (S->getOperand(i)->getType() != Ty) { 1234 Ty = SE.getEffectiveSCEVType(Ty); 1235 LHS = InsertNoopCastOfTo(LHS, Ty); 1236 } 1237 Value *RHS = expandCodeFor(S->getOperand(i), Ty); 1238 Value *ICmp = Builder.CreateICmpUGT(LHS, RHS, "tmp"); 1239 rememberInstruction(ICmp); 1240 Value *Sel = Builder.CreateSelect(ICmp, LHS, RHS, "umax"); 1241 rememberInstruction(Sel); 1242 LHS = Sel; 1243 } 1244 // In the case of mixed integer and pointer types, cast the 1245 // final result back to the pointer type. 1246 if (LHS->getType() != S->getType()) 1247 LHS = InsertNoopCastOfTo(LHS, S->getType()); 1248 return LHS; 1249} 1250 1251Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty, 1252 Instruction *I) { 1253 BasicBlock::iterator IP = I; 1254 while (isInsertedInstruction(IP) || isa<DbgInfoIntrinsic>(IP)) 1255 ++IP; 1256 Builder.SetInsertPoint(IP->getParent(), IP); 1257 return expandCodeFor(SH, Ty); 1258} 1259 1260Value *SCEVExpander::expandCodeFor(const SCEV *SH, const Type *Ty) { 1261 // Expand the code for this SCEV. 1262 Value *V = expand(SH); 1263 if (Ty) { 1264 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 1265 "non-trivial casts should be done with the SCEVs directly!"); 1266 V = InsertNoopCastOfTo(V, Ty); 1267 } 1268 return V; 1269} 1270 1271Value *SCEVExpander::expand(const SCEV *S) { 1272 // Compute an insertion point for this SCEV object. Hoist the instructions 1273 // as far out in the loop nest as possible. 1274 Instruction *InsertPt = Builder.GetInsertPoint(); 1275 for (Loop *L = SE.LI->getLoopFor(Builder.GetInsertBlock()); ; 1276 L = L->getParentLoop()) 1277 if (S->isLoopInvariant(L)) { 1278 if (!L) break; 1279 if (BasicBlock *Preheader = L->getLoopPreheader()) 1280 InsertPt = Preheader->getTerminator(); 1281 } else { 1282 // If the SCEV is computable at this level, insert it into the header 1283 // after the PHIs (and after any other instructions that we've inserted 1284 // there) so that it is guaranteed to dominate any user inside the loop. 1285 if (L && S->hasComputableLoopEvolution(L) && !PostIncLoops.count(L)) 1286 InsertPt = L->getHeader()->getFirstNonPHI(); 1287 while (isInsertedInstruction(InsertPt) || isa<DbgInfoIntrinsic>(InsertPt)) 1288 InsertPt = llvm::next(BasicBlock::iterator(InsertPt)); 1289 break; 1290 } 1291 1292 // Check to see if we already expanded this here. 1293 std::map<std::pair<const SCEV *, Instruction *>, 1294 AssertingVH<Value> >::iterator I = 1295 InsertedExpressions.find(std::make_pair(S, InsertPt)); 1296 if (I != InsertedExpressions.end()) 1297 return I->second; 1298 1299 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1300 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 1301 Builder.SetInsertPoint(InsertPt->getParent(), InsertPt); 1302 1303 // Expand the expression into instructions. 1304 Value *V = visit(S); 1305 1306 // Remember the expanded value for this SCEV at this location. 1307 if (PostIncLoops.empty()) 1308 InsertedExpressions[std::make_pair(S, InsertPt)] = V; 1309 1310 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 1311 return V; 1312} 1313 1314void SCEVExpander::rememberInstruction(Value *I) { 1315 if (PostIncLoops.empty()) 1316 InsertedValues.insert(I); 1317 1318 // If we just claimed an existing instruction and that instruction had 1319 // been the insert point, adjust the insert point forward so that 1320 // subsequently inserted code will be dominated. 1321 if (Builder.GetInsertPoint() == I) { 1322 BasicBlock::iterator It = cast<Instruction>(I); 1323 do { ++It; } while (isInsertedInstruction(It) || 1324 isa<DbgInfoIntrinsic>(It)); 1325 Builder.SetInsertPoint(Builder.GetInsertBlock(), It); 1326 } 1327} 1328 1329void SCEVExpander::restoreInsertPoint(BasicBlock *BB, BasicBlock::iterator I) { 1330 // If we acquired more instructions since the old insert point was saved, 1331 // advance past them. 1332 while (isInsertedInstruction(I) || isa<DbgInfoIntrinsic>(I)) ++I; 1333 1334 Builder.SetInsertPoint(BB, I); 1335} 1336 1337/// getOrInsertCanonicalInductionVariable - This method returns the 1338/// canonical induction variable of the specified type for the specified 1339/// loop (inserting one if there is none). A canonical induction variable 1340/// starts at zero and steps by one on each iteration. 1341Value * 1342SCEVExpander::getOrInsertCanonicalInductionVariable(const Loop *L, 1343 const Type *Ty) { 1344 assert(Ty->isIntegerTy() && "Can only insert integer induction variables!"); 1345 const SCEV *H = SE.getAddRecExpr(SE.getConstant(Ty, 0), 1346 SE.getConstant(Ty, 1), L); 1347 BasicBlock *SaveInsertBB = Builder.GetInsertBlock(); 1348 BasicBlock::iterator SaveInsertPt = Builder.GetInsertPoint(); 1349 Value *V = expandCodeFor(H, 0, L->getHeader()->begin()); 1350 if (SaveInsertBB) 1351 restoreInsertPoint(SaveInsertBB, SaveInsertPt); 1352 return V; 1353} 1354