ScalarEvolutionExpander.cpp revision 4a4f767235dac50965fc266fb822d9a2e37d5212
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/Target/TargetData.h" 19using namespace llvm; 20 21/// InsertCastOfTo - Insert a cast of V to the specified type, doing what 22/// we can to share the casts. 23Value *SCEVExpander::InsertCastOfTo(Instruction::CastOps opcode, Value *V, 24 const Type *Ty) { 25 // Short-circuit unnecessary bitcasts. 26 if (opcode == Instruction::BitCast && V->getType() == Ty) 27 return V; 28 29 // Short-circuit unnecessary inttoptr<->ptrtoint casts. 30 if ((opcode == Instruction::PtrToInt || opcode == Instruction::IntToPtr) && 31 SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(V->getType())) { 32 if (CastInst *CI = dyn_cast<CastInst>(V)) 33 if ((CI->getOpcode() == Instruction::PtrToInt || 34 CI->getOpcode() == Instruction::IntToPtr) && 35 SE.getTypeSizeInBits(CI->getType()) == 36 SE.getTypeSizeInBits(CI->getOperand(0)->getType())) 37 return CI->getOperand(0); 38 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 39 if ((CE->getOpcode() == Instruction::PtrToInt || 40 CE->getOpcode() == Instruction::IntToPtr) && 41 SE.getTypeSizeInBits(CE->getType()) == 42 SE.getTypeSizeInBits(CE->getOperand(0)->getType())) 43 return CE->getOperand(0); 44 } 45 46 // FIXME: keep track of the cast instruction. 47 if (Constant *C = dyn_cast<Constant>(V)) 48 return ConstantExpr::getCast(opcode, C, Ty); 49 50 if (Argument *A = dyn_cast<Argument>(V)) { 51 // Check to see if there is already a cast! 52 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); 53 UI != E; ++UI) { 54 if ((*UI)->getType() == Ty) 55 if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI))) 56 if (CI->getOpcode() == opcode) { 57 // If the cast isn't the first instruction of the function, move it. 58 if (BasicBlock::iterator(CI) != 59 A->getParent()->getEntryBlock().begin()) { 60 // If the CastInst is the insert point, change the insert point. 61 if (CI == InsertPt) ++InsertPt; 62 // Splice the cast at the beginning of the entry block. 63 CI->moveBefore(A->getParent()->getEntryBlock().begin()); 64 } 65 return CI; 66 } 67 } 68 Instruction *I = CastInst::Create(opcode, V, Ty, V->getName(), 69 A->getParent()->getEntryBlock().begin()); 70 InsertedValues.insert(I); 71 return I; 72 } 73 74 Instruction *I = cast<Instruction>(V); 75 76 // Check to see if there is already a cast. If there is, use it. 77 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 78 UI != E; ++UI) { 79 if ((*UI)->getType() == Ty) 80 if (CastInst *CI = dyn_cast<CastInst>(cast<Instruction>(*UI))) 81 if (CI->getOpcode() == opcode) { 82 BasicBlock::iterator It = I; ++It; 83 if (isa<InvokeInst>(I)) 84 It = cast<InvokeInst>(I)->getNormalDest()->begin(); 85 while (isa<PHINode>(It)) ++It; 86 if (It != BasicBlock::iterator(CI)) { 87 // If the CastInst is the insert point, change the insert point. 88 if (CI == InsertPt) ++InsertPt; 89 // Splice the cast immediately after the operand in question. 90 CI->moveBefore(It); 91 } 92 return CI; 93 } 94 } 95 BasicBlock::iterator IP = I; ++IP; 96 if (InvokeInst *II = dyn_cast<InvokeInst>(I)) 97 IP = II->getNormalDest()->begin(); 98 while (isa<PHINode>(IP)) ++IP; 99 Instruction *CI = CastInst::Create(opcode, V, Ty, V->getName(), IP); 100 InsertedValues.insert(CI); 101 return CI; 102} 103 104/// InsertNoopCastOfTo - Insert a cast of V to the specified type, 105/// which must be possible with a noop cast. 106Value *SCEVExpander::InsertNoopCastOfTo(Value *V, const Type *Ty) { 107 Instruction::CastOps Op = CastInst::getCastOpcode(V, false, Ty, false); 108 assert((Op == Instruction::BitCast || 109 Op == Instruction::PtrToInt || 110 Op == Instruction::IntToPtr) && 111 "InsertNoopCastOfTo cannot perform non-noop casts!"); 112 assert(SE.getTypeSizeInBits(V->getType()) == SE.getTypeSizeInBits(Ty) && 113 "InsertNoopCastOfTo cannot change sizes!"); 114 return InsertCastOfTo(Op, V, Ty); 115} 116 117/// InsertBinop - Insert the specified binary operator, doing a small amount 118/// of work to avoid inserting an obviously redundant operation. 119Value *SCEVExpander::InsertBinop(Instruction::BinaryOps Opcode, Value *LHS, 120 Value *RHS, BasicBlock::iterator InsertPt) { 121 // Fold a binop with constant operands. 122 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 123 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 124 return ConstantExpr::get(Opcode, CLHS, CRHS); 125 126 // Do a quick scan to see if we have this binop nearby. If so, reuse it. 127 unsigned ScanLimit = 6; 128 BasicBlock::iterator BlockBegin = InsertPt->getParent()->begin(); 129 if (InsertPt != BlockBegin) { 130 // Scanning starts from the last instruction before InsertPt. 131 BasicBlock::iterator IP = InsertPt; 132 --IP; 133 for (; ScanLimit; --IP, --ScanLimit) { 134 if (IP->getOpcode() == (unsigned)Opcode && IP->getOperand(0) == LHS && 135 IP->getOperand(1) == RHS) 136 return IP; 137 if (IP == BlockBegin) break; 138 } 139 } 140 141 // If we haven't found this binop, insert it. 142 Instruction *BO = BinaryOperator::Create(Opcode, LHS, RHS, "tmp", InsertPt); 143 InsertedValues.insert(BO); 144 return BO; 145} 146 147/// FactorOutConstant - Test if S is divisible by Factor, using signed 148/// division. If so, update S with Factor divided out and return true. 149/// S need not be evenly divisble if a reasonable remainder can be 150/// computed. 151/// TODO: When ScalarEvolution gets a SCEVSDivExpr, this can be made 152/// unnecessary; in its place, just signed-divide Ops[i] by the scale and 153/// check to see if the divide was folded. 154static bool FactorOutConstant(SCEVHandle &S, 155 SCEVHandle &Remainder, 156 const APInt &Factor, 157 ScalarEvolution &SE) { 158 // Everything is divisible by one. 159 if (Factor == 1) 160 return true; 161 162 // For a Constant, check for a multiple of the given factor. 163 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 164 ConstantInt *CI = 165 ConstantInt::get(C->getValue()->getValue().sdiv(Factor)); 166 // If the quotient is zero and the remainder is non-zero, reject 167 // the value at this scale. It will be considered for subsequent 168 // smaller scales. 169 if (C->isZero() || !CI->isZero()) { 170 SCEVHandle Div = SE.getConstant(CI); 171 S = Div; 172 Remainder = 173 SE.getAddExpr(Remainder, 174 SE.getConstant(C->getValue()->getValue().srem(Factor))); 175 return true; 176 } 177 } 178 179 // In a Mul, check if there is a constant operand which is a multiple 180 // of the given factor. 181 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) 182 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(M->getOperand(0))) 183 if (!C->getValue()->getValue().srem(Factor)) { 184 std::vector<SCEVHandle> NewMulOps(M->getOperands()); 185 NewMulOps[0] = 186 SE.getConstant(C->getValue()->getValue().sdiv(Factor)); 187 S = SE.getMulExpr(NewMulOps); 188 return true; 189 } 190 191 // In an AddRec, check if both start and step are divisible. 192 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 193 SCEVHandle Step = A->getStepRecurrence(SE); 194 SCEVHandle StepRem = SE.getIntegerSCEV(0, Step->getType()); 195 if (!FactorOutConstant(Step, StepRem, Factor, SE)) 196 return false; 197 if (!StepRem->isZero()) 198 return false; 199 SCEVHandle Start = A->getStart(); 200 if (!FactorOutConstant(Start, Remainder, Factor, SE)) 201 return false; 202 S = SE.getAddRecExpr(Start, Step, A->getLoop()); 203 return true; 204 } 205 206 return false; 207} 208 209/// expandAddToGEP - Expand a SCEVAddExpr with a pointer type into a GEP 210/// instead of using ptrtoint+arithmetic+inttoptr. This helps 211/// BasicAliasAnalysis analyze the result. However, it suffers from the 212/// underlying bug described in PR2831. Addition in LLVM currently always 213/// has two's complement wrapping guaranteed. However, the semantics for 214/// getelementptr overflow are ambiguous. In the common case though, this 215/// expansion gets used when a GEP in the original code has been converted 216/// into integer arithmetic, in which case the resulting code will be no 217/// more undefined than it was originally. 218/// 219/// Design note: It might seem desirable for this function to be more 220/// loop-aware. If some of the indices are loop-invariant while others 221/// aren't, it might seem desirable to emit multiple GEPs, keeping the 222/// loop-invariant portions of the overall computation outside the loop. 223/// However, there are a few reasons this is not done here. Hoisting simple 224/// arithmetic is a low-level optimization that often isn't very 225/// important until late in the optimization process. In fact, passes 226/// like InstructionCombining will combine GEPs, even if it means 227/// pushing loop-invariant computation down into loops, so even if the 228/// GEPs were split here, the work would quickly be undone. The 229/// LoopStrengthReduction pass, which is usually run quite late (and 230/// after the last InstructionCombining pass), takes care of hoisting 231/// loop-invariant portions of expressions, after considering what 232/// can be folded using target addressing modes. 233/// 234Value *SCEVExpander::expandAddToGEP(const SCEVHandle *op_begin, 235 const SCEVHandle *op_end, 236 const PointerType *PTy, 237 const Type *Ty, 238 Value *V) { 239 const Type *ElTy = PTy->getElementType(); 240 SmallVector<Value *, 4> GepIndices; 241 std::vector<SCEVHandle> Ops(op_begin, op_end); 242 bool AnyNonZeroIndices = false; 243 244 // Decend down the pointer's type and attempt to convert the other 245 // operands into GEP indices, at each level. The first index in a GEP 246 // indexes into the array implied by the pointer operand; the rest of 247 // the indices index into the element or field type selected by the 248 // preceding index. 249 for (;;) { 250 APInt ElSize = APInt(SE.getTypeSizeInBits(Ty), 251 ElTy->isSized() ? SE.TD->getTypeAllocSize(ElTy) : 0); 252 std::vector<SCEVHandle> NewOps; 253 std::vector<SCEVHandle> ScaledOps; 254 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 255 // Split AddRecs up into parts as either of the parts may be usable 256 // without the other. 257 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Ops[i])) 258 if (!A->getStart()->isZero()) { 259 SCEVHandle Start = A->getStart(); 260 Ops.push_back(SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()), 261 A->getStepRecurrence(SE), 262 A->getLoop())); 263 Ops[i] = Start; 264 ++e; 265 } 266 // If the scale size is not 0, attempt to factor out a scale. 267 if (ElSize != 0) { 268 SCEVHandle Op = Ops[i]; 269 SCEVHandle Remainder = SE.getIntegerSCEV(0, Op->getType()); 270 if (FactorOutConstant(Op, Remainder, ElSize, SE)) { 271 ScaledOps.push_back(Op); // Op now has ElSize factored out. 272 NewOps.push_back(Remainder); 273 continue; 274 } 275 } 276 // If the operand was not divisible, add it to the list of operands 277 // we'll scan next iteration. 278 NewOps.push_back(Ops[i]); 279 } 280 Ops = NewOps; 281 AnyNonZeroIndices |= !ScaledOps.empty(); 282 Value *Scaled = ScaledOps.empty() ? 283 Constant::getNullValue(Ty) : 284 expandCodeFor(SE.getAddExpr(ScaledOps), Ty); 285 GepIndices.push_back(Scaled); 286 287 // Collect struct field index operands. 288 if (!Ops.empty()) 289 while (const StructType *STy = dyn_cast<StructType>(ElTy)) { 290 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[0])) 291 if (SE.getTypeSizeInBits(C->getType()) <= 64) { 292 const StructLayout &SL = *SE.TD->getStructLayout(STy); 293 uint64_t FullOffset = C->getValue()->getZExtValue(); 294 if (FullOffset < SL.getSizeInBytes()) { 295 unsigned ElIdx = SL.getElementContainingOffset(FullOffset); 296 GepIndices.push_back(ConstantInt::get(Type::Int32Ty, ElIdx)); 297 ElTy = STy->getTypeAtIndex(ElIdx); 298 Ops[0] = 299 SE.getConstant(ConstantInt::get(Ty, 300 FullOffset - 301 SL.getElementOffset(ElIdx))); 302 AnyNonZeroIndices = true; 303 continue; 304 } 305 } 306 break; 307 } 308 309 if (const ArrayType *ATy = dyn_cast<ArrayType>(ElTy)) { 310 ElTy = ATy->getElementType(); 311 continue; 312 } 313 break; 314 } 315 316 // If none of the operands were convertable to proper GEP indices, cast 317 // the base to i8* and do an ugly getelementptr with that. It's still 318 // better than ptrtoint+arithmetic+inttoptr at least. 319 if (!AnyNonZeroIndices) { 320 V = InsertNoopCastOfTo(V, 321 Type::Int8Ty->getPointerTo(PTy->getAddressSpace())); 322 Value *Idx = expand(SE.getAddExpr(Ops)); 323 Idx = InsertNoopCastOfTo(Idx, Ty); 324 325 // Fold a GEP with constant operands. 326 if (Constant *CLHS = dyn_cast<Constant>(V)) 327 if (Constant *CRHS = dyn_cast<Constant>(Idx)) 328 return ConstantExpr::getGetElementPtr(CLHS, &CRHS, 1); 329 330 // Do a quick scan to see if we have this GEP nearby. If so, reuse it. 331 unsigned ScanLimit = 6; 332 BasicBlock::iterator BlockBegin = InsertPt->getParent()->begin(); 333 if (InsertPt != BlockBegin) { 334 // Scanning starts from the last instruction before InsertPt. 335 BasicBlock::iterator IP = InsertPt; 336 --IP; 337 for (; ScanLimit; --IP, --ScanLimit) { 338 if (IP->getOpcode() == Instruction::GetElementPtr && 339 IP->getOperand(0) == V && IP->getOperand(1) == Idx) 340 return IP; 341 if (IP == BlockBegin) break; 342 } 343 } 344 345 Value *GEP = GetElementPtrInst::Create(V, Idx, "scevgep", InsertPt); 346 InsertedValues.insert(GEP); 347 return GEP; 348 } 349 350 // Insert a pretty getelementptr. 351 Value *GEP = GetElementPtrInst::Create(V, 352 GepIndices.begin(), 353 GepIndices.end(), 354 "scevgep", InsertPt); 355 Ops.push_back(SE.getUnknown(GEP)); 356 InsertedValues.insert(GEP); 357 return expand(SE.getAddExpr(Ops)); 358} 359 360Value *SCEVExpander::visitAddExpr(const SCEVAddExpr *S) { 361 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 362 Value *V = expand(S->getOperand(S->getNumOperands()-1)); 363 364 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 365 // comments on expandAddToGEP for details. 366 if (SE.TD) 367 if (const PointerType *PTy = dyn_cast<PointerType>(V->getType())) { 368 const std::vector<SCEVHandle> &Ops = S->getOperands(); 369 return expandAddToGEP(&Ops[0], &Ops[Ops.size() - 1], 370 PTy, Ty, V); 371 } 372 373 V = InsertNoopCastOfTo(V, Ty); 374 375 // Emit a bunch of add instructions 376 for (int i = S->getNumOperands()-2; i >= 0; --i) { 377 Value *W = expand(S->getOperand(i)); 378 W = InsertNoopCastOfTo(W, Ty); 379 V = InsertBinop(Instruction::Add, V, W, InsertPt); 380 } 381 return V; 382} 383 384Value *SCEVExpander::visitMulExpr(const SCEVMulExpr *S) { 385 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 386 int FirstOp = 0; // Set if we should emit a subtract. 387 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getOperand(0))) 388 if (SC->getValue()->isAllOnesValue()) 389 FirstOp = 1; 390 391 int i = S->getNumOperands()-2; 392 Value *V = expand(S->getOperand(i+1)); 393 V = InsertNoopCastOfTo(V, Ty); 394 395 // Emit a bunch of multiply instructions 396 for (; i >= FirstOp; --i) { 397 Value *W = expand(S->getOperand(i)); 398 W = InsertNoopCastOfTo(W, Ty); 399 V = InsertBinop(Instruction::Mul, V, W, InsertPt); 400 } 401 402 // -1 * ... ---> 0 - ... 403 if (FirstOp == 1) 404 V = InsertBinop(Instruction::Sub, Constant::getNullValue(Ty), V, InsertPt); 405 return V; 406} 407 408Value *SCEVExpander::visitUDivExpr(const SCEVUDivExpr *S) { 409 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 410 411 Value *LHS = expand(S->getLHS()); 412 LHS = InsertNoopCastOfTo(LHS, Ty); 413 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(S->getRHS())) { 414 const APInt &RHS = SC->getValue()->getValue(); 415 if (RHS.isPowerOf2()) 416 return InsertBinop(Instruction::LShr, LHS, 417 ConstantInt::get(Ty, RHS.logBase2()), 418 InsertPt); 419 } 420 421 Value *RHS = expand(S->getRHS()); 422 RHS = InsertNoopCastOfTo(RHS, Ty); 423 return InsertBinop(Instruction::UDiv, LHS, RHS, InsertPt); 424} 425 426/// Move parts of Base into Rest to leave Base with the minimal 427/// expression that provides a pointer operand suitable for a 428/// GEP expansion. 429static void ExposePointerBase(SCEVHandle &Base, SCEVHandle &Rest, 430 ScalarEvolution &SE) { 431 while (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(Base)) { 432 Base = A->getStart(); 433 Rest = SE.getAddExpr(Rest, 434 SE.getAddRecExpr(SE.getIntegerSCEV(0, A->getType()), 435 A->getStepRecurrence(SE), 436 A->getLoop())); 437 } 438 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(Base)) { 439 Base = A->getOperand(A->getNumOperands()-1); 440 std::vector<SCEVHandle> NewAddOps(A->op_begin(), A->op_end()); 441 NewAddOps.back() = Rest; 442 Rest = SE.getAddExpr(NewAddOps); 443 ExposePointerBase(Base, Rest, SE); 444 } 445} 446 447Value *SCEVExpander::visitAddRecExpr(const SCEVAddRecExpr *S) { 448 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 449 const Loop *L = S->getLoop(); 450 451 // {X,+,F} --> X + {0,+,F} 452 if (!S->getStart()->isZero()) { 453 std::vector<SCEVHandle> NewOps(S->getOperands()); 454 NewOps[0] = SE.getIntegerSCEV(0, Ty); 455 SCEVHandle Rest = SE.getAddRecExpr(NewOps, L); 456 457 // Turn things like ptrtoint+arithmetic+inttoptr into GEP. See the 458 // comments on expandAddToGEP for details. 459 if (SE.TD) { 460 SCEVHandle Base = S->getStart(); 461 SCEVHandle RestArray[1] = Rest; 462 // Dig into the expression to find the pointer base for a GEP. 463 ExposePointerBase(Base, RestArray[0], SE); 464 // If we found a pointer, expand the AddRec with a GEP. 465 if (const PointerType *PTy = dyn_cast<PointerType>(Base->getType())) { 466 // Make sure the Base isn't something exotic, such as a multiplied 467 // or divided pointer value. In those cases, the result type isn't 468 // actually a pointer type. 469 if (!isa<SCEVMulExpr>(Base) && !isa<SCEVUDivExpr>(Base)) { 470 Value *StartV = expand(Base); 471 assert(StartV->getType() == PTy && "Pointer type mismatch for GEP!"); 472 return expandAddToGEP(RestArray, RestArray+1, PTy, Ty, StartV); 473 } 474 } 475 } 476 477 Value *RestV = expand(Rest); 478 return expand(SE.getAddExpr(S->getStart(), SE.getUnknown(RestV))); 479 } 480 481 // {0,+,1} --> Insert a canonical induction variable into the loop! 482 if (S->isAffine() && 483 S->getOperand(1) == SE.getIntegerSCEV(1, Ty)) { 484 // Create and insert the PHI node for the induction variable in the 485 // specified loop. 486 BasicBlock *Header = L->getHeader(); 487 PHINode *PN = PHINode::Create(Ty, "indvar", Header->begin()); 488 InsertedValues.insert(PN); 489 PN->addIncoming(Constant::getNullValue(Ty), L->getLoopPreheader()); 490 491 pred_iterator HPI = pred_begin(Header); 492 assert(HPI != pred_end(Header) && "Loop with zero preds???"); 493 if (!L->contains(*HPI)) ++HPI; 494 assert(HPI != pred_end(Header) && L->contains(*HPI) && 495 "No backedge in loop?"); 496 497 // Insert a unit add instruction right before the terminator corresponding 498 // to the back-edge. 499 Constant *One = ConstantInt::get(Ty, 1); 500 Instruction *Add = BinaryOperator::CreateAdd(PN, One, "indvar.next", 501 (*HPI)->getTerminator()); 502 InsertedValues.insert(Add); 503 504 pred_iterator PI = pred_begin(Header); 505 if (*PI == L->getLoopPreheader()) 506 ++PI; 507 PN->addIncoming(Add, *PI); 508 return PN; 509 } 510 511 // Get the canonical induction variable I for this loop. 512 Value *I = getOrInsertCanonicalInductionVariable(L, Ty); 513 514 // If this is a simple linear addrec, emit it now as a special case. 515 if (S->isAffine()) { // {0,+,F} --> i*F 516 Value *F = expand(S->getOperand(1)); 517 F = InsertNoopCastOfTo(F, Ty); 518 519 // IF the step is by one, just return the inserted IV. 520 if (ConstantInt *CI = dyn_cast<ConstantInt>(F)) 521 if (CI->getValue() == 1) 522 return I; 523 524 // If the insert point is directly inside of the loop, emit the multiply at 525 // the insert point. Otherwise, L is a loop that is a parent of the insert 526 // point loop. If we can, move the multiply to the outer most loop that it 527 // is safe to be in. 528 BasicBlock::iterator MulInsertPt = getInsertionPoint(); 529 Loop *InsertPtLoop = SE.LI->getLoopFor(MulInsertPt->getParent()); 530 if (InsertPtLoop != L && InsertPtLoop && 531 L->contains(InsertPtLoop->getHeader())) { 532 do { 533 // If we cannot hoist the multiply out of this loop, don't. 534 if (!InsertPtLoop->isLoopInvariant(F)) break; 535 536 BasicBlock *InsertPtLoopPH = InsertPtLoop->getLoopPreheader(); 537 538 // If this loop hasn't got a preheader, we aren't able to hoist the 539 // multiply. 540 if (!InsertPtLoopPH) 541 break; 542 543 // Otherwise, move the insert point to the preheader. 544 MulInsertPt = InsertPtLoopPH->getTerminator(); 545 InsertPtLoop = InsertPtLoop->getParentLoop(); 546 } while (InsertPtLoop != L); 547 } 548 549 return InsertBinop(Instruction::Mul, I, F, MulInsertPt); 550 } 551 552 // If this is a chain of recurrences, turn it into a closed form, using the 553 // folders, then expandCodeFor the closed form. This allows the folders to 554 // simplify the expression without having to build a bunch of special code 555 // into this folder. 556 SCEVHandle IH = SE.getUnknown(I); // Get I as a "symbolic" SCEV. 557 558 SCEVHandle V = S->evaluateAtIteration(IH, SE); 559 //cerr << "Evaluated: " << *this << "\n to: " << *V << "\n"; 560 561 return expand(V); 562} 563 564Value *SCEVExpander::visitTruncateExpr(const SCEVTruncateExpr *S) { 565 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 566 Value *V = expand(S->getOperand()); 567 V = InsertNoopCastOfTo(V, SE.getEffectiveSCEVType(V->getType())); 568 Instruction *I = new TruncInst(V, Ty, "tmp.", InsertPt); 569 InsertedValues.insert(I); 570 return I; 571} 572 573Value *SCEVExpander::visitZeroExtendExpr(const SCEVZeroExtendExpr *S) { 574 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 575 Value *V = expand(S->getOperand()); 576 V = InsertNoopCastOfTo(V, SE.getEffectiveSCEVType(V->getType())); 577 Instruction *I = new ZExtInst(V, Ty, "tmp.", InsertPt); 578 InsertedValues.insert(I); 579 return I; 580} 581 582Value *SCEVExpander::visitSignExtendExpr(const SCEVSignExtendExpr *S) { 583 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 584 Value *V = expand(S->getOperand()); 585 V = InsertNoopCastOfTo(V, SE.getEffectiveSCEVType(V->getType())); 586 Instruction *I = new SExtInst(V, Ty, "tmp.", InsertPt); 587 InsertedValues.insert(I); 588 return I; 589} 590 591Value *SCEVExpander::visitSMaxExpr(const SCEVSMaxExpr *S) { 592 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 593 Value *LHS = expand(S->getOperand(0)); 594 LHS = InsertNoopCastOfTo(LHS, Ty); 595 for (unsigned i = 1; i < S->getNumOperands(); ++i) { 596 Value *RHS = expand(S->getOperand(i)); 597 RHS = InsertNoopCastOfTo(RHS, Ty); 598 Instruction *ICmp = 599 new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS, "tmp", InsertPt); 600 InsertedValues.insert(ICmp); 601 Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "smax", InsertPt); 602 InsertedValues.insert(Sel); 603 LHS = Sel; 604 } 605 return LHS; 606} 607 608Value *SCEVExpander::visitUMaxExpr(const SCEVUMaxExpr *S) { 609 const Type *Ty = SE.getEffectiveSCEVType(S->getType()); 610 Value *LHS = expand(S->getOperand(0)); 611 LHS = InsertNoopCastOfTo(LHS, Ty); 612 for (unsigned i = 1; i < S->getNumOperands(); ++i) { 613 Value *RHS = expand(S->getOperand(i)); 614 RHS = InsertNoopCastOfTo(RHS, Ty); 615 Instruction *ICmp = 616 new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS, "tmp", InsertPt); 617 InsertedValues.insert(ICmp); 618 Instruction *Sel = SelectInst::Create(ICmp, LHS, RHS, "umax", InsertPt); 619 InsertedValues.insert(Sel); 620 LHS = Sel; 621 } 622 return LHS; 623} 624 625Value *SCEVExpander::expandCodeFor(SCEVHandle SH, const Type *Ty) { 626 // Expand the code for this SCEV. 627 Value *V = expand(SH); 628 if (Ty) { 629 assert(SE.getTypeSizeInBits(Ty) == SE.getTypeSizeInBits(SH->getType()) && 630 "non-trivial casts should be done with the SCEVs directly!"); 631 V = InsertNoopCastOfTo(V, Ty); 632 } 633 return V; 634} 635 636Value *SCEVExpander::expand(const SCEV *S) { 637 // Check to see if we already expanded this. 638 std::map<SCEVHandle, AssertingVH<Value> >::iterator I = 639 InsertedExpressions.find(S); 640 if (I != InsertedExpressions.end()) 641 return I->second; 642 643 Value *V = visit(S); 644 InsertedExpressions[S] = V; 645 return V; 646} 647