DAGCombiner.cpp revision 5bce67a95feb136389ca630cc5dd6a81e97ff1eb
1//===-- DAGCombiner.cpp - Implement a DAG node combiner -------------------===// 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 pass combines dag nodes to form fewer, simpler DAG nodes. It can be run 11// both before and after the DAG is legalized. 12// 13// This pass is not a substitute for the LLVM IR instcombine pass. This pass is 14// primarily intended to handle simplification opportunities that are implicit 15// in the LLVM IR and exposed by the various codegen lowering phases. 16// 17//===----------------------------------------------------------------------===// 18 19#define DEBUG_TYPE "dagcombine" 20#include "llvm/CodeGen/SelectionDAG.h" 21#include "llvm/ADT/SmallPtrSet.h" 22#include "llvm/ADT/Statistic.h" 23#include "llvm/Analysis/AliasAnalysis.h" 24#include "llvm/CodeGen/MachineFrameInfo.h" 25#include "llvm/CodeGen/MachineFunction.h" 26#include "llvm/IR/DataLayout.h" 27#include "llvm/IR/DerivedTypes.h" 28#include "llvm/IR/Function.h" 29#include "llvm/IR/LLVMContext.h" 30#include "llvm/Support/CommandLine.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/ErrorHandling.h" 33#include "llvm/Support/MathExtras.h" 34#include "llvm/Support/raw_ostream.h" 35#include "llvm/Target/TargetLowering.h" 36#include "llvm/Target/TargetMachine.h" 37#include "llvm/Target/TargetOptions.h" 38#include <algorithm> 39using namespace llvm; 40 41STATISTIC(NodesCombined , "Number of dag nodes combined"); 42STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created"); 43STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created"); 44STATISTIC(OpsNarrowed , "Number of load/op/store narrowed"); 45STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int"); 46 47namespace { 48 static cl::opt<bool> 49 CombinerAA("combiner-alias-analysis", cl::Hidden, 50 cl::desc("Turn on alias analysis during testing")); 51 52 static cl::opt<bool> 53 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden, 54 cl::desc("Include global information in alias analysis")); 55 56//------------------------------ DAGCombiner ---------------------------------// 57 58 class DAGCombiner { 59 SelectionDAG &DAG; 60 const TargetLowering &TLI; 61 CombineLevel Level; 62 CodeGenOpt::Level OptLevel; 63 bool LegalOperations; 64 bool LegalTypes; 65 66 // Worklist of all of the nodes that need to be simplified. 67 // 68 // This has the semantics that when adding to the worklist, 69 // the item added must be next to be processed. It should 70 // also only appear once. The naive approach to this takes 71 // linear time. 72 // 73 // To reduce the insert/remove time to logarithmic, we use 74 // a set and a vector to maintain our worklist. 75 // 76 // The set contains the items on the worklist, but does not 77 // maintain the order they should be visited. 78 // 79 // The vector maintains the order nodes should be visited, but may 80 // contain duplicate or removed nodes. When choosing a node to 81 // visit, we pop off the order stack until we find an item that is 82 // also in the contents set. All operations are O(log N). 83 SmallPtrSet<SDNode*, 64> WorkListContents; 84 SmallVector<SDNode*, 64> WorkListOrder; 85 86 // AA - Used for DAG load/store alias analysis. 87 AliasAnalysis &AA; 88 89 /// AddUsersToWorkList - When an instruction is simplified, add all users of 90 /// the instruction to the work lists because they might get more simplified 91 /// now. 92 /// 93 void AddUsersToWorkList(SDNode *N) { 94 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 95 UI != UE; ++UI) 96 AddToWorkList(*UI); 97 } 98 99 /// visit - call the node-specific routine that knows how to fold each 100 /// particular type of node. 101 SDValue visit(SDNode *N); 102 103 public: 104 /// AddToWorkList - Add to the work list making sure its instance is at the 105 /// back (next to be processed.) 106 void AddToWorkList(SDNode *N) { 107 WorkListContents.insert(N); 108 WorkListOrder.push_back(N); 109 } 110 111 /// removeFromWorkList - remove all instances of N from the worklist. 112 /// 113 void removeFromWorkList(SDNode *N) { 114 WorkListContents.erase(N); 115 } 116 117 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, 118 bool AddTo = true); 119 120 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) { 121 return CombineTo(N, &Res, 1, AddTo); 122 } 123 124 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, 125 bool AddTo = true) { 126 SDValue To[] = { Res0, Res1 }; 127 return CombineTo(N, To, 2, AddTo); 128 } 129 130 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO); 131 132 private: 133 134 /// SimplifyDemandedBits - Check the specified integer node value to see if 135 /// it can be simplified or if things it uses can be simplified by bit 136 /// propagation. If so, return true. 137 bool SimplifyDemandedBits(SDValue Op) { 138 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits(); 139 APInt Demanded = APInt::getAllOnesValue(BitWidth); 140 return SimplifyDemandedBits(Op, Demanded); 141 } 142 143 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded); 144 145 bool CombineToPreIndexedLoadStore(SDNode *N); 146 bool CombineToPostIndexedLoadStore(SDNode *N); 147 148 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad); 149 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace); 150 SDValue SExtPromoteOperand(SDValue Op, EVT PVT); 151 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT); 152 SDValue PromoteIntBinOp(SDValue Op); 153 SDValue PromoteIntShiftOp(SDValue Op); 154 SDValue PromoteExtend(SDValue Op); 155 bool PromoteLoad(SDValue Op); 156 157 void ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs, 158 SDValue Trunc, SDValue ExtLoad, SDLoc DL, 159 ISD::NodeType ExtType); 160 161 /// combine - call the node-specific routine that knows how to fold each 162 /// particular type of node. If that doesn't do anything, try the 163 /// target-specific DAG combines. 164 SDValue combine(SDNode *N); 165 166 // Visitation implementation - Implement dag node combining for different 167 // node types. The semantics are as follows: 168 // Return Value: 169 // SDValue.getNode() == 0 - No change was made 170 // SDValue.getNode() == N - N was replaced, is dead and has been handled. 171 // otherwise - N should be replaced by the returned Operand. 172 // 173 SDValue visitTokenFactor(SDNode *N); 174 SDValue visitMERGE_VALUES(SDNode *N); 175 SDValue visitADD(SDNode *N); 176 SDValue visitSUB(SDNode *N); 177 SDValue visitADDC(SDNode *N); 178 SDValue visitSUBC(SDNode *N); 179 SDValue visitADDE(SDNode *N); 180 SDValue visitSUBE(SDNode *N); 181 SDValue visitMUL(SDNode *N); 182 SDValue visitSDIV(SDNode *N); 183 SDValue visitUDIV(SDNode *N); 184 SDValue visitSREM(SDNode *N); 185 SDValue visitUREM(SDNode *N); 186 SDValue visitMULHU(SDNode *N); 187 SDValue visitMULHS(SDNode *N); 188 SDValue visitSMUL_LOHI(SDNode *N); 189 SDValue visitUMUL_LOHI(SDNode *N); 190 SDValue visitSMULO(SDNode *N); 191 SDValue visitUMULO(SDNode *N); 192 SDValue visitSDIVREM(SDNode *N); 193 SDValue visitUDIVREM(SDNode *N); 194 SDValue visitAND(SDNode *N); 195 SDValue visitOR(SDNode *N); 196 SDValue visitXOR(SDNode *N); 197 SDValue SimplifyVBinOp(SDNode *N); 198 SDValue SimplifyVUnaryOp(SDNode *N); 199 SDValue visitSHL(SDNode *N); 200 SDValue visitSRA(SDNode *N); 201 SDValue visitSRL(SDNode *N); 202 SDValue visitCTLZ(SDNode *N); 203 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N); 204 SDValue visitCTTZ(SDNode *N); 205 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N); 206 SDValue visitCTPOP(SDNode *N); 207 SDValue visitSELECT(SDNode *N); 208 SDValue visitVSELECT(SDNode *N); 209 SDValue visitSELECT_CC(SDNode *N); 210 SDValue visitSETCC(SDNode *N); 211 SDValue visitSIGN_EXTEND(SDNode *N); 212 SDValue visitZERO_EXTEND(SDNode *N); 213 SDValue visitANY_EXTEND(SDNode *N); 214 SDValue visitSIGN_EXTEND_INREG(SDNode *N); 215 SDValue visitTRUNCATE(SDNode *N); 216 SDValue visitBITCAST(SDNode *N); 217 SDValue visitBUILD_PAIR(SDNode *N); 218 SDValue visitFADD(SDNode *N); 219 SDValue visitFSUB(SDNode *N); 220 SDValue visitFMUL(SDNode *N); 221 SDValue visitFMA(SDNode *N); 222 SDValue visitFDIV(SDNode *N); 223 SDValue visitFREM(SDNode *N); 224 SDValue visitFCOPYSIGN(SDNode *N); 225 SDValue visitSINT_TO_FP(SDNode *N); 226 SDValue visitUINT_TO_FP(SDNode *N); 227 SDValue visitFP_TO_SINT(SDNode *N); 228 SDValue visitFP_TO_UINT(SDNode *N); 229 SDValue visitFP_ROUND(SDNode *N); 230 SDValue visitFP_ROUND_INREG(SDNode *N); 231 SDValue visitFP_EXTEND(SDNode *N); 232 SDValue visitFNEG(SDNode *N); 233 SDValue visitFABS(SDNode *N); 234 SDValue visitFCEIL(SDNode *N); 235 SDValue visitFTRUNC(SDNode *N); 236 SDValue visitFFLOOR(SDNode *N); 237 SDValue visitBRCOND(SDNode *N); 238 SDValue visitBR_CC(SDNode *N); 239 SDValue visitLOAD(SDNode *N); 240 SDValue visitSTORE(SDNode *N); 241 SDValue visitINSERT_VECTOR_ELT(SDNode *N); 242 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N); 243 SDValue visitBUILD_VECTOR(SDNode *N); 244 SDValue visitCONCAT_VECTORS(SDNode *N); 245 SDValue visitEXTRACT_SUBVECTOR(SDNode *N); 246 SDValue visitVECTOR_SHUFFLE(SDNode *N); 247 248 SDValue XformToShuffleWithZero(SDNode *N); 249 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS); 250 251 SDValue visitShiftByConstant(SDNode *N, unsigned Amt); 252 253 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS); 254 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N); 255 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2); 256 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2, 257 SDValue N3, ISD::CondCode CC, 258 bool NotExtCompare = false); 259 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, 260 SDLoc DL, bool foldBooleans = true); 261 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, 262 unsigned HiOp); 263 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT); 264 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT); 265 SDValue BuildSDIV(SDNode *N); 266 SDValue BuildUDIV(SDNode *N); 267 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, 268 bool DemandHighBits = true); 269 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1); 270 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL); 271 SDValue ReduceLoadWidth(SDNode *N); 272 SDValue ReduceLoadOpStoreWidth(SDNode *N); 273 SDValue TransformFPLoadStorePair(SDNode *N); 274 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N); 275 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N); 276 277 SDValue GetDemandedBits(SDValue V, const APInt &Mask); 278 279 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes, 280 /// looking for aliasing nodes and adding them to the Aliases vector. 281 void GatherAllAliases(SDNode *N, SDValue OriginalChain, 282 SmallVector<SDValue, 8> &Aliases); 283 284 /// isAlias - Return true if there is any possibility that the two addresses 285 /// overlap. 286 bool isAlias(SDValue Ptr1, int64_t Size1, 287 const Value *SrcValue1, int SrcValueOffset1, 288 unsigned SrcValueAlign1, 289 const MDNode *TBAAInfo1, 290 SDValue Ptr2, int64_t Size2, 291 const Value *SrcValue2, int SrcValueOffset2, 292 unsigned SrcValueAlign2, 293 const MDNode *TBAAInfo2) const; 294 295 /// isAlias - Return true if there is any possibility that the two addresses 296 /// overlap. 297 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1); 298 299 /// FindAliasInfo - Extracts the relevant alias information from the memory 300 /// node. Returns true if the operand was a load. 301 bool FindAliasInfo(SDNode *N, 302 SDValue &Ptr, int64_t &Size, 303 const Value *&SrcValue, int &SrcValueOffset, 304 unsigned &SrcValueAlignment, 305 const MDNode *&TBAAInfo) const; 306 307 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, 308 /// looking for a better chain (aliasing node.) 309 SDValue FindBetterChain(SDNode *N, SDValue Chain); 310 311 /// Merge consecutive store operations into a wide store. 312 /// This optimization uses wide integers or vectors when possible. 313 /// \return True if some memory operations were changed. 314 bool MergeConsecutiveStores(StoreSDNode *N); 315 316 public: 317 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL) 318 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes), 319 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {} 320 321 /// Run - runs the dag combiner on all nodes in the work list 322 void Run(CombineLevel AtLevel); 323 324 SelectionDAG &getDAG() const { return DAG; } 325 326 /// getShiftAmountTy - Returns a type large enough to hold any valid 327 /// shift amount - before type legalization these can be huge. 328 EVT getShiftAmountTy(EVT LHSTy) { 329 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 330 if (LHSTy.isVector()) 331 return LHSTy; 332 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy) : TLI.getPointerTy(); 333 } 334 335 /// isTypeLegal - This method returns true if we are running before type 336 /// legalization or if the specified VT is legal. 337 bool isTypeLegal(const EVT &VT) { 338 if (!LegalTypes) return true; 339 return TLI.isTypeLegal(VT); 340 } 341 342 /// getSetCCResultType - Convenience wrapper around 343 /// TargetLowering::getSetCCResultType 344 EVT getSetCCResultType(EVT VT) const { 345 return TLI.getSetCCResultType(*DAG.getContext(), VT); 346 } 347 }; 348} 349 350 351namespace { 352/// WorkListRemover - This class is a DAGUpdateListener that removes any deleted 353/// nodes from the worklist. 354class WorkListRemover : public SelectionDAG::DAGUpdateListener { 355 DAGCombiner &DC; 356public: 357 explicit WorkListRemover(DAGCombiner &dc) 358 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {} 359 360 virtual void NodeDeleted(SDNode *N, SDNode *E) { 361 DC.removeFromWorkList(N); 362 } 363}; 364} 365 366//===----------------------------------------------------------------------===// 367// TargetLowering::DAGCombinerInfo implementation 368//===----------------------------------------------------------------------===// 369 370void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) { 371 ((DAGCombiner*)DC)->AddToWorkList(N); 372} 373 374void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) { 375 ((DAGCombiner*)DC)->removeFromWorkList(N); 376} 377 378SDValue TargetLowering::DAGCombinerInfo:: 379CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) { 380 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo); 381} 382 383SDValue TargetLowering::DAGCombinerInfo:: 384CombineTo(SDNode *N, SDValue Res, bool AddTo) { 385 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo); 386} 387 388 389SDValue TargetLowering::DAGCombinerInfo:: 390CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) { 391 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo); 392} 393 394void TargetLowering::DAGCombinerInfo:: 395CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { 396 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO); 397} 398 399//===----------------------------------------------------------------------===// 400// Helper Functions 401//===----------------------------------------------------------------------===// 402 403/// isNegatibleForFree - Return 1 if we can compute the negated form of the 404/// specified expression for the same cost as the expression itself, or 2 if we 405/// can compute the negated form more cheaply than the expression itself. 406static char isNegatibleForFree(SDValue Op, bool LegalOperations, 407 const TargetLowering &TLI, 408 const TargetOptions *Options, 409 unsigned Depth = 0) { 410 // fneg is removable even if it has multiple uses. 411 if (Op.getOpcode() == ISD::FNEG) return 2; 412 413 // Don't allow anything with multiple uses. 414 if (!Op.hasOneUse()) return 0; 415 416 // Don't recurse exponentially. 417 if (Depth > 6) return 0; 418 419 switch (Op.getOpcode()) { 420 default: return false; 421 case ISD::ConstantFP: 422 // Don't invert constant FP values after legalize. The negated constant 423 // isn't necessarily legal. 424 return LegalOperations ? 0 : 1; 425 case ISD::FADD: 426 // FIXME: determine better conditions for this xform. 427 if (!Options->UnsafeFPMath) return 0; 428 429 // After operation legalization, it might not be legal to create new FSUBs. 430 if (LegalOperations && 431 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) 432 return 0; 433 434 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B) 435 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, 436 Options, Depth + 1)) 437 return V; 438 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A) 439 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options, 440 Depth + 1); 441 case ISD::FSUB: 442 // We can't turn -(A-B) into B-A when we honor signed zeros. 443 if (!Options->UnsafeFPMath) return 0; 444 445 // fold (fneg (fsub A, B)) -> (fsub B, A) 446 return 1; 447 448 case ISD::FMUL: 449 case ISD::FDIV: 450 if (Options->HonorSignDependentRoundingFPMath()) return 0; 451 452 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y)) 453 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, 454 Options, Depth + 1)) 455 return V; 456 457 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options, 458 Depth + 1); 459 460 case ISD::FP_EXTEND: 461 case ISD::FP_ROUND: 462 case ISD::FSIN: 463 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options, 464 Depth + 1); 465 } 466} 467 468/// GetNegatedExpression - If isNegatibleForFree returns true, this function 469/// returns the newly negated expression. 470static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG, 471 bool LegalOperations, unsigned Depth = 0) { 472 // fneg is removable even if it has multiple uses. 473 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0); 474 475 // Don't allow anything with multiple uses. 476 assert(Op.hasOneUse() && "Unknown reuse!"); 477 478 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree"); 479 switch (Op.getOpcode()) { 480 default: llvm_unreachable("Unknown code"); 481 case ISD::ConstantFP: { 482 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF(); 483 V.changeSign(); 484 return DAG.getConstantFP(V, Op.getValueType()); 485 } 486 case ISD::FADD: 487 // FIXME: determine better conditions for this xform. 488 assert(DAG.getTarget().Options.UnsafeFPMath); 489 490 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B) 491 if (isNegatibleForFree(Op.getOperand(0), LegalOperations, 492 DAG.getTargetLoweringInfo(), 493 &DAG.getTarget().Options, Depth+1)) 494 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), 495 GetNegatedExpression(Op.getOperand(0), DAG, 496 LegalOperations, Depth+1), 497 Op.getOperand(1)); 498 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A) 499 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), 500 GetNegatedExpression(Op.getOperand(1), DAG, 501 LegalOperations, Depth+1), 502 Op.getOperand(0)); 503 case ISD::FSUB: 504 // We can't turn -(A-B) into B-A when we honor signed zeros. 505 assert(DAG.getTarget().Options.UnsafeFPMath); 506 507 // fold (fneg (fsub 0, B)) -> B 508 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0))) 509 if (N0CFP->getValueAPF().isZero()) 510 return Op.getOperand(1); 511 512 // fold (fneg (fsub A, B)) -> (fsub B, A) 513 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), 514 Op.getOperand(1), Op.getOperand(0)); 515 516 case ISD::FMUL: 517 case ISD::FDIV: 518 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath()); 519 520 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) 521 if (isNegatibleForFree(Op.getOperand(0), LegalOperations, 522 DAG.getTargetLoweringInfo(), 523 &DAG.getTarget().Options, Depth+1)) 524 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), 525 GetNegatedExpression(Op.getOperand(0), DAG, 526 LegalOperations, Depth+1), 527 Op.getOperand(1)); 528 529 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y)) 530 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), 531 Op.getOperand(0), 532 GetNegatedExpression(Op.getOperand(1), DAG, 533 LegalOperations, Depth+1)); 534 535 case ISD::FP_EXTEND: 536 case ISD::FSIN: 537 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), 538 GetNegatedExpression(Op.getOperand(0), DAG, 539 LegalOperations, Depth+1)); 540 case ISD::FP_ROUND: 541 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(), 542 GetNegatedExpression(Op.getOperand(0), DAG, 543 LegalOperations, Depth+1), 544 Op.getOperand(1)); 545 } 546} 547 548 549// isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc 550// that selects between the values 1 and 0, making it equivalent to a setcc. 551// Also, set the incoming LHS, RHS, and CC references to the appropriate 552// nodes based on the type of node we are checking. This simplifies life a 553// bit for the callers. 554static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, 555 SDValue &CC) { 556 if (N.getOpcode() == ISD::SETCC) { 557 LHS = N.getOperand(0); 558 RHS = N.getOperand(1); 559 CC = N.getOperand(2); 560 return true; 561 } 562 if (N.getOpcode() == ISD::SELECT_CC && 563 N.getOperand(2).getOpcode() == ISD::Constant && 564 N.getOperand(3).getOpcode() == ISD::Constant && 565 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 && 566 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) { 567 LHS = N.getOperand(0); 568 RHS = N.getOperand(1); 569 CC = N.getOperand(4); 570 return true; 571 } 572 return false; 573} 574 575// isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only 576// one use. If this is true, it allows the users to invert the operation for 577// free when it is profitable to do so. 578static bool isOneUseSetCC(SDValue N) { 579 SDValue N0, N1, N2; 580 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse()) 581 return true; 582 return false; 583} 584 585SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL, 586 SDValue N0, SDValue N1) { 587 EVT VT = N0.getValueType(); 588 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) { 589 if (isa<ConstantSDNode>(N1)) { 590 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2)) 591 SDValue OpNode = 592 DAG.FoldConstantArithmetic(Opc, VT, 593 cast<ConstantSDNode>(N0.getOperand(1)), 594 cast<ConstantSDNode>(N1)); 595 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode); 596 } 597 if (N0.hasOneUse()) { 598 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use 599 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, 600 N0.getOperand(0), N1); 601 AddToWorkList(OpNode.getNode()); 602 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1)); 603 } 604 } 605 606 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) { 607 if (isa<ConstantSDNode>(N0)) { 608 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2)) 609 SDValue OpNode = 610 DAG.FoldConstantArithmetic(Opc, VT, 611 cast<ConstantSDNode>(N1.getOperand(1)), 612 cast<ConstantSDNode>(N0)); 613 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode); 614 } 615 if (N1.hasOneUse()) { 616 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use 617 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, 618 N1.getOperand(0), N0); 619 AddToWorkList(OpNode.getNode()); 620 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1)); 621 } 622 } 623 624 return SDValue(); 625} 626 627SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, 628 bool AddTo) { 629 assert(N->getNumValues() == NumTo && "Broken CombineTo call!"); 630 ++NodesCombined; 631 DEBUG(dbgs() << "\nReplacing.1 "; 632 N->dump(&DAG); 633 dbgs() << "\nWith: "; 634 To[0].getNode()->dump(&DAG); 635 dbgs() << " and " << NumTo-1 << " other values\n"; 636 for (unsigned i = 0, e = NumTo; i != e; ++i) 637 assert((!To[i].getNode() || 638 N->getValueType(i) == To[i].getValueType()) && 639 "Cannot combine value to value of different type!")); 640 WorkListRemover DeadNodes(*this); 641 DAG.ReplaceAllUsesWith(N, To); 642 if (AddTo) { 643 // Push the new nodes and any users onto the worklist 644 for (unsigned i = 0, e = NumTo; i != e; ++i) { 645 if (To[i].getNode()) { 646 AddToWorkList(To[i].getNode()); 647 AddUsersToWorkList(To[i].getNode()); 648 } 649 } 650 } 651 652 // Finally, if the node is now dead, remove it from the graph. The node 653 // may not be dead if the replacement process recursively simplified to 654 // something else needing this node. 655 if (N->use_empty()) { 656 // Nodes can be reintroduced into the worklist. Make sure we do not 657 // process a node that has been replaced. 658 removeFromWorkList(N); 659 660 // Finally, since the node is now dead, remove it from the graph. 661 DAG.DeleteNode(N); 662 } 663 return SDValue(N, 0); 664} 665 666void DAGCombiner:: 667CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { 668 // Replace all uses. If any nodes become isomorphic to other nodes and 669 // are deleted, make sure to remove them from our worklist. 670 WorkListRemover DeadNodes(*this); 671 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New); 672 673 // Push the new node and any (possibly new) users onto the worklist. 674 AddToWorkList(TLO.New.getNode()); 675 AddUsersToWorkList(TLO.New.getNode()); 676 677 // Finally, if the node is now dead, remove it from the graph. The node 678 // may not be dead if the replacement process recursively simplified to 679 // something else needing this node. 680 if (TLO.Old.getNode()->use_empty()) { 681 removeFromWorkList(TLO.Old.getNode()); 682 683 // If the operands of this node are only used by the node, they will now 684 // be dead. Make sure to visit them first to delete dead nodes early. 685 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i) 686 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse()) 687 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode()); 688 689 DAG.DeleteNode(TLO.Old.getNode()); 690 } 691} 692 693/// SimplifyDemandedBits - Check the specified integer node value to see if 694/// it can be simplified or if things it uses can be simplified by bit 695/// propagation. If so, return true. 696bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) { 697 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); 698 APInt KnownZero, KnownOne; 699 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO)) 700 return false; 701 702 // Revisit the node. 703 AddToWorkList(Op.getNode()); 704 705 // Replace the old value with the new one. 706 ++NodesCombined; 707 DEBUG(dbgs() << "\nReplacing.2 "; 708 TLO.Old.getNode()->dump(&DAG); 709 dbgs() << "\nWith: "; 710 TLO.New.getNode()->dump(&DAG); 711 dbgs() << '\n'); 712 713 CommitTargetLoweringOpt(TLO); 714 return true; 715} 716 717void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) { 718 SDLoc dl(Load); 719 EVT VT = Load->getValueType(0); 720 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0)); 721 722 DEBUG(dbgs() << "\nReplacing.9 "; 723 Load->dump(&DAG); 724 dbgs() << "\nWith: "; 725 Trunc.getNode()->dump(&DAG); 726 dbgs() << '\n'); 727 WorkListRemover DeadNodes(*this); 728 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc); 729 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1)); 730 removeFromWorkList(Load); 731 DAG.DeleteNode(Load); 732 AddToWorkList(Trunc.getNode()); 733} 734 735SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) { 736 Replace = false; 737 SDLoc dl(Op); 738 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) { 739 EVT MemVT = LD->getMemoryVT(); 740 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) 741 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD 742 : ISD::EXTLOAD) 743 : LD->getExtensionType(); 744 Replace = true; 745 return DAG.getExtLoad(ExtType, dl, PVT, 746 LD->getChain(), LD->getBasePtr(), 747 LD->getPointerInfo(), 748 MemVT, LD->isVolatile(), 749 LD->isNonTemporal(), LD->getAlignment()); 750 } 751 752 unsigned Opc = Op.getOpcode(); 753 switch (Opc) { 754 default: break; 755 case ISD::AssertSext: 756 return DAG.getNode(ISD::AssertSext, dl, PVT, 757 SExtPromoteOperand(Op.getOperand(0), PVT), 758 Op.getOperand(1)); 759 case ISD::AssertZext: 760 return DAG.getNode(ISD::AssertZext, dl, PVT, 761 ZExtPromoteOperand(Op.getOperand(0), PVT), 762 Op.getOperand(1)); 763 case ISD::Constant: { 764 unsigned ExtOpc = 765 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; 766 return DAG.getNode(ExtOpc, dl, PVT, Op); 767 } 768 } 769 770 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT)) 771 return SDValue(); 772 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op); 773} 774 775SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) { 776 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT)) 777 return SDValue(); 778 EVT OldVT = Op.getValueType(); 779 SDLoc dl(Op); 780 bool Replace = false; 781 SDValue NewOp = PromoteOperand(Op, PVT, Replace); 782 if (NewOp.getNode() == 0) 783 return SDValue(); 784 AddToWorkList(NewOp.getNode()); 785 786 if (Replace) 787 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); 788 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp, 789 DAG.getValueType(OldVT)); 790} 791 792SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) { 793 EVT OldVT = Op.getValueType(); 794 SDLoc dl(Op); 795 bool Replace = false; 796 SDValue NewOp = PromoteOperand(Op, PVT, Replace); 797 if (NewOp.getNode() == 0) 798 return SDValue(); 799 AddToWorkList(NewOp.getNode()); 800 801 if (Replace) 802 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); 803 return DAG.getZeroExtendInReg(NewOp, dl, OldVT); 804} 805 806/// PromoteIntBinOp - Promote the specified integer binary operation if the 807/// target indicates it is beneficial. e.g. On x86, it's usually better to 808/// promote i16 operations to i32 since i16 instructions are longer. 809SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) { 810 if (!LegalOperations) 811 return SDValue(); 812 813 EVT VT = Op.getValueType(); 814 if (VT.isVector() || !VT.isInteger()) 815 return SDValue(); 816 817 // If operation type is 'undesirable', e.g. i16 on x86, consider 818 // promoting it. 819 unsigned Opc = Op.getOpcode(); 820 if (TLI.isTypeDesirableForOp(Opc, VT)) 821 return SDValue(); 822 823 EVT PVT = VT; 824 // Consult target whether it is a good idea to promote this operation and 825 // what's the right type to promote it to. 826 if (TLI.IsDesirableToPromoteOp(Op, PVT)) { 827 assert(PVT != VT && "Don't know what type to promote to!"); 828 829 bool Replace0 = false; 830 SDValue N0 = Op.getOperand(0); 831 SDValue NN0 = PromoteOperand(N0, PVT, Replace0); 832 if (NN0.getNode() == 0) 833 return SDValue(); 834 835 bool Replace1 = false; 836 SDValue N1 = Op.getOperand(1); 837 SDValue NN1; 838 if (N0 == N1) 839 NN1 = NN0; 840 else { 841 NN1 = PromoteOperand(N1, PVT, Replace1); 842 if (NN1.getNode() == 0) 843 return SDValue(); 844 } 845 846 AddToWorkList(NN0.getNode()); 847 if (NN1.getNode()) 848 AddToWorkList(NN1.getNode()); 849 850 if (Replace0) 851 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode()); 852 if (Replace1) 853 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode()); 854 855 DEBUG(dbgs() << "\nPromoting "; 856 Op.getNode()->dump(&DAG)); 857 SDLoc dl(Op); 858 return DAG.getNode(ISD::TRUNCATE, dl, VT, 859 DAG.getNode(Opc, dl, PVT, NN0, NN1)); 860 } 861 return SDValue(); 862} 863 864/// PromoteIntShiftOp - Promote the specified integer shift operation if the 865/// target indicates it is beneficial. e.g. On x86, it's usually better to 866/// promote i16 operations to i32 since i16 instructions are longer. 867SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) { 868 if (!LegalOperations) 869 return SDValue(); 870 871 EVT VT = Op.getValueType(); 872 if (VT.isVector() || !VT.isInteger()) 873 return SDValue(); 874 875 // If operation type is 'undesirable', e.g. i16 on x86, consider 876 // promoting it. 877 unsigned Opc = Op.getOpcode(); 878 if (TLI.isTypeDesirableForOp(Opc, VT)) 879 return SDValue(); 880 881 EVT PVT = VT; 882 // Consult target whether it is a good idea to promote this operation and 883 // what's the right type to promote it to. 884 if (TLI.IsDesirableToPromoteOp(Op, PVT)) { 885 assert(PVT != VT && "Don't know what type to promote to!"); 886 887 bool Replace = false; 888 SDValue N0 = Op.getOperand(0); 889 if (Opc == ISD::SRA) 890 N0 = SExtPromoteOperand(Op.getOperand(0), PVT); 891 else if (Opc == ISD::SRL) 892 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT); 893 else 894 N0 = PromoteOperand(N0, PVT, Replace); 895 if (N0.getNode() == 0) 896 return SDValue(); 897 898 AddToWorkList(N0.getNode()); 899 if (Replace) 900 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode()); 901 902 DEBUG(dbgs() << "\nPromoting "; 903 Op.getNode()->dump(&DAG)); 904 SDLoc dl(Op); 905 return DAG.getNode(ISD::TRUNCATE, dl, VT, 906 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1))); 907 } 908 return SDValue(); 909} 910 911SDValue DAGCombiner::PromoteExtend(SDValue Op) { 912 if (!LegalOperations) 913 return SDValue(); 914 915 EVT VT = Op.getValueType(); 916 if (VT.isVector() || !VT.isInteger()) 917 return SDValue(); 918 919 // If operation type is 'undesirable', e.g. i16 on x86, consider 920 // promoting it. 921 unsigned Opc = Op.getOpcode(); 922 if (TLI.isTypeDesirableForOp(Opc, VT)) 923 return SDValue(); 924 925 EVT PVT = VT; 926 // Consult target whether it is a good idea to promote this operation and 927 // what's the right type to promote it to. 928 if (TLI.IsDesirableToPromoteOp(Op, PVT)) { 929 assert(PVT != VT && "Don't know what type to promote to!"); 930 // fold (aext (aext x)) -> (aext x) 931 // fold (aext (zext x)) -> (zext x) 932 // fold (aext (sext x)) -> (sext x) 933 DEBUG(dbgs() << "\nPromoting "; 934 Op.getNode()->dump(&DAG)); 935 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0)); 936 } 937 return SDValue(); 938} 939 940bool DAGCombiner::PromoteLoad(SDValue Op) { 941 if (!LegalOperations) 942 return false; 943 944 EVT VT = Op.getValueType(); 945 if (VT.isVector() || !VT.isInteger()) 946 return false; 947 948 // If operation type is 'undesirable', e.g. i16 on x86, consider 949 // promoting it. 950 unsigned Opc = Op.getOpcode(); 951 if (TLI.isTypeDesirableForOp(Opc, VT)) 952 return false; 953 954 EVT PVT = VT; 955 // Consult target whether it is a good idea to promote this operation and 956 // what's the right type to promote it to. 957 if (TLI.IsDesirableToPromoteOp(Op, PVT)) { 958 assert(PVT != VT && "Don't know what type to promote to!"); 959 960 SDLoc dl(Op); 961 SDNode *N = Op.getNode(); 962 LoadSDNode *LD = cast<LoadSDNode>(N); 963 EVT MemVT = LD->getMemoryVT(); 964 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) 965 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD 966 : ISD::EXTLOAD) 967 : LD->getExtensionType(); 968 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT, 969 LD->getChain(), LD->getBasePtr(), 970 LD->getPointerInfo(), 971 MemVT, LD->isVolatile(), 972 LD->isNonTemporal(), LD->getAlignment()); 973 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD); 974 975 DEBUG(dbgs() << "\nPromoting "; 976 N->dump(&DAG); 977 dbgs() << "\nTo: "; 978 Result.getNode()->dump(&DAG); 979 dbgs() << '\n'); 980 WorkListRemover DeadNodes(*this); 981 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result); 982 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1)); 983 removeFromWorkList(N); 984 DAG.DeleteNode(N); 985 AddToWorkList(Result.getNode()); 986 return true; 987 } 988 return false; 989} 990 991 992//===----------------------------------------------------------------------===// 993// Main DAG Combiner implementation 994//===----------------------------------------------------------------------===// 995 996void DAGCombiner::Run(CombineLevel AtLevel) { 997 // set the instance variables, so that the various visit routines may use it. 998 Level = AtLevel; 999 LegalOperations = Level >= AfterLegalizeVectorOps; 1000 LegalTypes = Level >= AfterLegalizeTypes; 1001 1002 // Add all the dag nodes to the worklist. 1003 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(), 1004 E = DAG.allnodes_end(); I != E; ++I) 1005 AddToWorkList(I); 1006 1007 // Create a dummy node (which is not added to allnodes), that adds a reference 1008 // to the root node, preventing it from being deleted, and tracking any 1009 // changes of the root. 1010 HandleSDNode Dummy(DAG.getRoot()); 1011 1012 // The root of the dag may dangle to deleted nodes until the dag combiner is 1013 // done. Set it to null to avoid confusion. 1014 DAG.setRoot(SDValue()); 1015 1016 // while the worklist isn't empty, find a node and 1017 // try and combine it. 1018 while (!WorkListContents.empty()) { 1019 SDNode *N; 1020 // The WorkListOrder holds the SDNodes in order, but it may contain duplicates. 1021 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the 1022 // worklist *should* contain, and check the node we want to visit is should 1023 // actually be visited. 1024 do { 1025 N = WorkListOrder.pop_back_val(); 1026 } while (!WorkListContents.erase(N)); 1027 1028 // If N has no uses, it is dead. Make sure to revisit all N's operands once 1029 // N is deleted from the DAG, since they too may now be dead or may have a 1030 // reduced number of uses, allowing other xforms. 1031 if (N->use_empty() && N != &Dummy) { 1032 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 1033 AddToWorkList(N->getOperand(i).getNode()); 1034 1035 DAG.DeleteNode(N); 1036 continue; 1037 } 1038 1039 SDValue RV = combine(N); 1040 1041 if (RV.getNode() == 0) 1042 continue; 1043 1044 ++NodesCombined; 1045 1046 // If we get back the same node we passed in, rather than a new node or 1047 // zero, we know that the node must have defined multiple values and 1048 // CombineTo was used. Since CombineTo takes care of the worklist 1049 // mechanics for us, we have no work to do in this case. 1050 if (RV.getNode() == N) 1051 continue; 1052 1053 assert(N->getOpcode() != ISD::DELETED_NODE && 1054 RV.getNode()->getOpcode() != ISD::DELETED_NODE && 1055 "Node was deleted but visit returned new node!"); 1056 1057 DEBUG(dbgs() << "\nReplacing.3 "; 1058 N->dump(&DAG); 1059 dbgs() << "\nWith: "; 1060 RV.getNode()->dump(&DAG); 1061 dbgs() << '\n'); 1062 1063 // Transfer debug value. 1064 DAG.TransferDbgValues(SDValue(N, 0), RV); 1065 WorkListRemover DeadNodes(*this); 1066 if (N->getNumValues() == RV.getNode()->getNumValues()) 1067 DAG.ReplaceAllUsesWith(N, RV.getNode()); 1068 else { 1069 assert(N->getValueType(0) == RV.getValueType() && 1070 N->getNumValues() == 1 && "Type mismatch"); 1071 SDValue OpV = RV; 1072 DAG.ReplaceAllUsesWith(N, &OpV); 1073 } 1074 1075 // Push the new node and any users onto the worklist 1076 AddToWorkList(RV.getNode()); 1077 AddUsersToWorkList(RV.getNode()); 1078 1079 // Add any uses of the old node to the worklist in case this node is the 1080 // last one that uses them. They may become dead after this node is 1081 // deleted. 1082 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 1083 AddToWorkList(N->getOperand(i).getNode()); 1084 1085 // Finally, if the node is now dead, remove it from the graph. The node 1086 // may not be dead if the replacement process recursively simplified to 1087 // something else needing this node. 1088 if (N->use_empty()) { 1089 // Nodes can be reintroduced into the worklist. Make sure we do not 1090 // process a node that has been replaced. 1091 removeFromWorkList(N); 1092 1093 // Finally, since the node is now dead, remove it from the graph. 1094 DAG.DeleteNode(N); 1095 } 1096 } 1097 1098 // If the root changed (e.g. it was a dead load, update the root). 1099 DAG.setRoot(Dummy.getValue()); 1100 DAG.RemoveDeadNodes(); 1101} 1102 1103SDValue DAGCombiner::visit(SDNode *N) { 1104 switch (N->getOpcode()) { 1105 default: break; 1106 case ISD::TokenFactor: return visitTokenFactor(N); 1107 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N); 1108 case ISD::ADD: return visitADD(N); 1109 case ISD::SUB: return visitSUB(N); 1110 case ISD::ADDC: return visitADDC(N); 1111 case ISD::SUBC: return visitSUBC(N); 1112 case ISD::ADDE: return visitADDE(N); 1113 case ISD::SUBE: return visitSUBE(N); 1114 case ISD::MUL: return visitMUL(N); 1115 case ISD::SDIV: return visitSDIV(N); 1116 case ISD::UDIV: return visitUDIV(N); 1117 case ISD::SREM: return visitSREM(N); 1118 case ISD::UREM: return visitUREM(N); 1119 case ISD::MULHU: return visitMULHU(N); 1120 case ISD::MULHS: return visitMULHS(N); 1121 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N); 1122 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N); 1123 case ISD::SMULO: return visitSMULO(N); 1124 case ISD::UMULO: return visitUMULO(N); 1125 case ISD::SDIVREM: return visitSDIVREM(N); 1126 case ISD::UDIVREM: return visitUDIVREM(N); 1127 case ISD::AND: return visitAND(N); 1128 case ISD::OR: return visitOR(N); 1129 case ISD::XOR: return visitXOR(N); 1130 case ISD::SHL: return visitSHL(N); 1131 case ISD::SRA: return visitSRA(N); 1132 case ISD::SRL: return visitSRL(N); 1133 case ISD::CTLZ: return visitCTLZ(N); 1134 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N); 1135 case ISD::CTTZ: return visitCTTZ(N); 1136 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N); 1137 case ISD::CTPOP: return visitCTPOP(N); 1138 case ISD::SELECT: return visitSELECT(N); 1139 case ISD::VSELECT: return visitVSELECT(N); 1140 case ISD::SELECT_CC: return visitSELECT_CC(N); 1141 case ISD::SETCC: return visitSETCC(N); 1142 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N); 1143 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N); 1144 case ISD::ANY_EXTEND: return visitANY_EXTEND(N); 1145 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N); 1146 case ISD::TRUNCATE: return visitTRUNCATE(N); 1147 case ISD::BITCAST: return visitBITCAST(N); 1148 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N); 1149 case ISD::FADD: return visitFADD(N); 1150 case ISD::FSUB: return visitFSUB(N); 1151 case ISD::FMUL: return visitFMUL(N); 1152 case ISD::FMA: return visitFMA(N); 1153 case ISD::FDIV: return visitFDIV(N); 1154 case ISD::FREM: return visitFREM(N); 1155 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N); 1156 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N); 1157 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N); 1158 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N); 1159 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N); 1160 case ISD::FP_ROUND: return visitFP_ROUND(N); 1161 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N); 1162 case ISD::FP_EXTEND: return visitFP_EXTEND(N); 1163 case ISD::FNEG: return visitFNEG(N); 1164 case ISD::FABS: return visitFABS(N); 1165 case ISD::FFLOOR: return visitFFLOOR(N); 1166 case ISD::FCEIL: return visitFCEIL(N); 1167 case ISD::FTRUNC: return visitFTRUNC(N); 1168 case ISD::BRCOND: return visitBRCOND(N); 1169 case ISD::BR_CC: return visitBR_CC(N); 1170 case ISD::LOAD: return visitLOAD(N); 1171 case ISD::STORE: return visitSTORE(N); 1172 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N); 1173 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N); 1174 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N); 1175 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N); 1176 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N); 1177 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N); 1178 } 1179 return SDValue(); 1180} 1181 1182SDValue DAGCombiner::combine(SDNode *N) { 1183 SDValue RV = visit(N); 1184 1185 // If nothing happened, try a target-specific DAG combine. 1186 if (RV.getNode() == 0) { 1187 assert(N->getOpcode() != ISD::DELETED_NODE && 1188 "Node was deleted but visit returned NULL!"); 1189 1190 if (N->getOpcode() >= ISD::BUILTIN_OP_END || 1191 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) { 1192 1193 // Expose the DAG combiner to the target combiner impls. 1194 TargetLowering::DAGCombinerInfo 1195 DagCombineInfo(DAG, Level, false, this); 1196 1197 RV = TLI.PerformDAGCombine(N, DagCombineInfo); 1198 } 1199 } 1200 1201 // If nothing happened still, try promoting the operation. 1202 if (RV.getNode() == 0) { 1203 switch (N->getOpcode()) { 1204 default: break; 1205 case ISD::ADD: 1206 case ISD::SUB: 1207 case ISD::MUL: 1208 case ISD::AND: 1209 case ISD::OR: 1210 case ISD::XOR: 1211 RV = PromoteIntBinOp(SDValue(N, 0)); 1212 break; 1213 case ISD::SHL: 1214 case ISD::SRA: 1215 case ISD::SRL: 1216 RV = PromoteIntShiftOp(SDValue(N, 0)); 1217 break; 1218 case ISD::SIGN_EXTEND: 1219 case ISD::ZERO_EXTEND: 1220 case ISD::ANY_EXTEND: 1221 RV = PromoteExtend(SDValue(N, 0)); 1222 break; 1223 case ISD::LOAD: 1224 if (PromoteLoad(SDValue(N, 0))) 1225 RV = SDValue(N, 0); 1226 break; 1227 } 1228 } 1229 1230 // If N is a commutative binary node, try commuting it to enable more 1231 // sdisel CSE. 1232 if (RV.getNode() == 0 && 1233 SelectionDAG::isCommutativeBinOp(N->getOpcode()) && 1234 N->getNumValues() == 1) { 1235 SDValue N0 = N->getOperand(0); 1236 SDValue N1 = N->getOperand(1); 1237 1238 // Constant operands are canonicalized to RHS. 1239 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) { 1240 SDValue Ops[] = { N1, N0 }; 1241 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), 1242 Ops, 2); 1243 if (CSENode) 1244 return SDValue(CSENode, 0); 1245 } 1246 } 1247 1248 return RV; 1249} 1250 1251/// getInputChainForNode - Given a node, return its input chain if it has one, 1252/// otherwise return a null sd operand. 1253static SDValue getInputChainForNode(SDNode *N) { 1254 if (unsigned NumOps = N->getNumOperands()) { 1255 if (N->getOperand(0).getValueType() == MVT::Other) 1256 return N->getOperand(0); 1257 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other) 1258 return N->getOperand(NumOps-1); 1259 for (unsigned i = 1; i < NumOps-1; ++i) 1260 if (N->getOperand(i).getValueType() == MVT::Other) 1261 return N->getOperand(i); 1262 } 1263 return SDValue(); 1264} 1265 1266SDValue DAGCombiner::visitTokenFactor(SDNode *N) { 1267 // If N has two operands, where one has an input chain equal to the other, 1268 // the 'other' chain is redundant. 1269 if (N->getNumOperands() == 2) { 1270 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1)) 1271 return N->getOperand(0); 1272 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0)) 1273 return N->getOperand(1); 1274 } 1275 1276 SmallVector<SDNode *, 8> TFs; // List of token factors to visit. 1277 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor. 1278 SmallPtrSet<SDNode*, 16> SeenOps; 1279 bool Changed = false; // If we should replace this token factor. 1280 1281 // Start out with this token factor. 1282 TFs.push_back(N); 1283 1284 // Iterate through token factors. The TFs grows when new token factors are 1285 // encountered. 1286 for (unsigned i = 0; i < TFs.size(); ++i) { 1287 SDNode *TF = TFs[i]; 1288 1289 // Check each of the operands. 1290 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) { 1291 SDValue Op = TF->getOperand(i); 1292 1293 switch (Op.getOpcode()) { 1294 case ISD::EntryToken: 1295 // Entry tokens don't need to be added to the list. They are 1296 // rededundant. 1297 Changed = true; 1298 break; 1299 1300 case ISD::TokenFactor: 1301 if (Op.hasOneUse() && 1302 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) { 1303 // Queue up for processing. 1304 TFs.push_back(Op.getNode()); 1305 // Clean up in case the token factor is removed. 1306 AddToWorkList(Op.getNode()); 1307 Changed = true; 1308 break; 1309 } 1310 // Fall thru 1311 1312 default: 1313 // Only add if it isn't already in the list. 1314 if (SeenOps.insert(Op.getNode())) 1315 Ops.push_back(Op); 1316 else 1317 Changed = true; 1318 break; 1319 } 1320 } 1321 } 1322 1323 SDValue Result; 1324 1325 // If we've change things around then replace token factor. 1326 if (Changed) { 1327 if (Ops.empty()) { 1328 // The entry token is the only possible outcome. 1329 Result = DAG.getEntryNode(); 1330 } else { 1331 // New and improved token factor. 1332 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), 1333 MVT::Other, &Ops[0], Ops.size()); 1334 } 1335 1336 // Don't add users to work list. 1337 return CombineTo(N, Result, false); 1338 } 1339 1340 return Result; 1341} 1342 1343/// MERGE_VALUES can always be eliminated. 1344SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) { 1345 WorkListRemover DeadNodes(*this); 1346 // Replacing results may cause a different MERGE_VALUES to suddenly 1347 // be CSE'd with N, and carry its uses with it. Iterate until no 1348 // uses remain, to ensure that the node can be safely deleted. 1349 // First add the users of this node to the work list so that they 1350 // can be tried again once they have new operands. 1351 AddUsersToWorkList(N); 1352 do { 1353 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 1354 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i)); 1355 } while (!N->use_empty()); 1356 removeFromWorkList(N); 1357 DAG.DeleteNode(N); 1358 return SDValue(N, 0); // Return N so it doesn't get rechecked! 1359} 1360 1361static 1362SDValue combineShlAddConstant(SDLoc DL, SDValue N0, SDValue N1, 1363 SelectionDAG &DAG) { 1364 EVT VT = N0.getValueType(); 1365 SDValue N00 = N0.getOperand(0); 1366 SDValue N01 = N0.getOperand(1); 1367 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01); 1368 1369 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() && 1370 isa<ConstantSDNode>(N00.getOperand(1))) { 1371 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), ) 1372 N0 = DAG.getNode(ISD::ADD, SDLoc(N0), VT, 1373 DAG.getNode(ISD::SHL, SDLoc(N00), VT, 1374 N00.getOperand(0), N01), 1375 DAG.getNode(ISD::SHL, SDLoc(N01), VT, 1376 N00.getOperand(1), N01)); 1377 return DAG.getNode(ISD::ADD, DL, VT, N0, N1); 1378 } 1379 1380 return SDValue(); 1381} 1382 1383SDValue DAGCombiner::visitADD(SDNode *N) { 1384 SDValue N0 = N->getOperand(0); 1385 SDValue N1 = N->getOperand(1); 1386 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1387 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1388 EVT VT = N0.getValueType(); 1389 1390 // fold vector ops 1391 if (VT.isVector()) { 1392 SDValue FoldedVOp = SimplifyVBinOp(N); 1393 if (FoldedVOp.getNode()) return FoldedVOp; 1394 1395 // fold (add x, 0) -> x, vector edition 1396 if (ISD::isBuildVectorAllZeros(N1.getNode())) 1397 return N0; 1398 if (ISD::isBuildVectorAllZeros(N0.getNode())) 1399 return N1; 1400 } 1401 1402 // fold (add x, undef) -> undef 1403 if (N0.getOpcode() == ISD::UNDEF) 1404 return N0; 1405 if (N1.getOpcode() == ISD::UNDEF) 1406 return N1; 1407 // fold (add c1, c2) -> c1+c2 1408 if (N0C && N1C) 1409 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C); 1410 // canonicalize constant to RHS 1411 if (N0C && !N1C) 1412 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0); 1413 // fold (add x, 0) -> x 1414 if (N1C && N1C->isNullValue()) 1415 return N0; 1416 // fold (add Sym, c) -> Sym+c 1417 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0)) 1418 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C && 1419 GA->getOpcode() == ISD::GlobalAddress) 1420 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT, 1421 GA->getOffset() + 1422 (uint64_t)N1C->getSExtValue()); 1423 // fold ((c1-A)+c2) -> (c1+c2)-A 1424 if (N1C && N0.getOpcode() == ISD::SUB) 1425 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0))) 1426 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1427 DAG.getConstant(N1C->getAPIntValue()+ 1428 N0C->getAPIntValue(), VT), 1429 N0.getOperand(1)); 1430 // reassociate add 1431 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1); 1432 if (RADD.getNode() != 0) 1433 return RADD; 1434 // fold ((0-A) + B) -> B-A 1435 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) && 1436 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue()) 1437 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1)); 1438 // fold (A + (0-B)) -> A-B 1439 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) && 1440 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue()) 1441 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1)); 1442 // fold (A+(B-A)) -> B 1443 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1)) 1444 return N1.getOperand(0); 1445 // fold ((B-A)+A) -> B 1446 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1)) 1447 return N0.getOperand(0); 1448 // fold (A+(B-(A+C))) to (B-C) 1449 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && 1450 N0 == N1.getOperand(1).getOperand(0)) 1451 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0), 1452 N1.getOperand(1).getOperand(1)); 1453 // fold (A+(B-(C+A))) to (B-C) 1454 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && 1455 N0 == N1.getOperand(1).getOperand(1)) 1456 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0), 1457 N1.getOperand(1).getOperand(0)); 1458 // fold (A+((B-A)+or-C)) to (B+or-C) 1459 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) && 1460 N1.getOperand(0).getOpcode() == ISD::SUB && 1461 N0 == N1.getOperand(0).getOperand(1)) 1462 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT, 1463 N1.getOperand(0).getOperand(0), N1.getOperand(1)); 1464 1465 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant 1466 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) { 1467 SDValue N00 = N0.getOperand(0); 1468 SDValue N01 = N0.getOperand(1); 1469 SDValue N10 = N1.getOperand(0); 1470 SDValue N11 = N1.getOperand(1); 1471 1472 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10)) 1473 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1474 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10), 1475 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11)); 1476 } 1477 1478 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0))) 1479 return SDValue(N, 0); 1480 1481 // fold (a+b) -> (a|b) iff a and b share no bits. 1482 if (VT.isInteger() && !VT.isVector()) { 1483 APInt LHSZero, LHSOne; 1484 APInt RHSZero, RHSOne; 1485 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne); 1486 1487 if (LHSZero.getBoolValue()) { 1488 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne); 1489 1490 // If all possibly-set bits on the LHS are clear on the RHS, return an OR. 1491 // If all possibly-set bits on the RHS are clear on the LHS, return an OR. 1492 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero) 1493 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1); 1494 } 1495 } 1496 1497 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), ) 1498 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) { 1499 SDValue Result = combineShlAddConstant(SDLoc(N), N0, N1, DAG); 1500 if (Result.getNode()) return Result; 1501 } 1502 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) { 1503 SDValue Result = combineShlAddConstant(SDLoc(N), N1, N0, DAG); 1504 if (Result.getNode()) return Result; 1505 } 1506 1507 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n)) 1508 if (N1.getOpcode() == ISD::SHL && 1509 N1.getOperand(0).getOpcode() == ISD::SUB) 1510 if (ConstantSDNode *C = 1511 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0))) 1512 if (C->getAPIntValue() == 0) 1513 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, 1514 DAG.getNode(ISD::SHL, SDLoc(N), VT, 1515 N1.getOperand(0).getOperand(1), 1516 N1.getOperand(1))); 1517 if (N0.getOpcode() == ISD::SHL && 1518 N0.getOperand(0).getOpcode() == ISD::SUB) 1519 if (ConstantSDNode *C = 1520 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0))) 1521 if (C->getAPIntValue() == 0) 1522 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, 1523 DAG.getNode(ISD::SHL, SDLoc(N), VT, 1524 N0.getOperand(0).getOperand(1), 1525 N0.getOperand(1))); 1526 1527 if (N1.getOpcode() == ISD::AND) { 1528 SDValue AndOp0 = N1.getOperand(0); 1529 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1)); 1530 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0); 1531 unsigned DestBits = VT.getScalarType().getSizeInBits(); 1532 1533 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x)) 1534 // and similar xforms where the inner op is either ~0 or 0. 1535 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) { 1536 SDLoc DL(N); 1537 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0); 1538 } 1539 } 1540 1541 // add (sext i1), X -> sub X, (zext i1) 1542 if (N0.getOpcode() == ISD::SIGN_EXTEND && 1543 N0.getOperand(0).getValueType() == MVT::i1 && 1544 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) { 1545 SDLoc DL(N); 1546 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)); 1547 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt); 1548 } 1549 1550 return SDValue(); 1551} 1552 1553SDValue DAGCombiner::visitADDC(SDNode *N) { 1554 SDValue N0 = N->getOperand(0); 1555 SDValue N1 = N->getOperand(1); 1556 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1557 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1558 EVT VT = N0.getValueType(); 1559 1560 // If the flag result is dead, turn this into an ADD. 1561 if (!N->hasAnyUseOfValue(1)) 1562 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1), 1563 DAG.getNode(ISD::CARRY_FALSE, 1564 SDLoc(N), MVT::Glue)); 1565 1566 // canonicalize constant to RHS. 1567 if (N0C && !N1C) 1568 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0); 1569 1570 // fold (addc x, 0) -> x + no carry out 1571 if (N1C && N1C->isNullValue()) 1572 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, 1573 SDLoc(N), MVT::Glue)); 1574 1575 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits. 1576 APInt LHSZero, LHSOne; 1577 APInt RHSZero, RHSOne; 1578 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne); 1579 1580 if (LHSZero.getBoolValue()) { 1581 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne); 1582 1583 // If all possibly-set bits on the LHS are clear on the RHS, return an OR. 1584 // If all possibly-set bits on the RHS are clear on the LHS, return an OR. 1585 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero) 1586 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1), 1587 DAG.getNode(ISD::CARRY_FALSE, 1588 SDLoc(N), MVT::Glue)); 1589 } 1590 1591 return SDValue(); 1592} 1593 1594SDValue DAGCombiner::visitADDE(SDNode *N) { 1595 SDValue N0 = N->getOperand(0); 1596 SDValue N1 = N->getOperand(1); 1597 SDValue CarryIn = N->getOperand(2); 1598 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1599 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1600 1601 // canonicalize constant to RHS 1602 if (N0C && !N1C) 1603 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(), 1604 N1, N0, CarryIn); 1605 1606 // fold (adde x, y, false) -> (addc x, y) 1607 if (CarryIn.getOpcode() == ISD::CARRY_FALSE) 1608 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1); 1609 1610 return SDValue(); 1611} 1612 1613// Since it may not be valid to emit a fold to zero for vector initializers 1614// check if we can before folding. 1615static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT, 1616 SelectionDAG &DAG, bool LegalOperations) { 1617 if (!VT.isVector()) { 1618 return DAG.getConstant(0, VT); 1619 } 1620 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) { 1621 // Produce a vector of zeros. 1622 SDValue El = DAG.getConstant(0, VT.getVectorElementType()); 1623 std::vector<SDValue> Ops(VT.getVectorNumElements(), El); 1624 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, 1625 &Ops[0], Ops.size()); 1626 } 1627 return SDValue(); 1628} 1629 1630SDValue DAGCombiner::visitSUB(SDNode *N) { 1631 SDValue N0 = N->getOperand(0); 1632 SDValue N1 = N->getOperand(1); 1633 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode()); 1634 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 1635 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 : 1636 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode()); 1637 EVT VT = N0.getValueType(); 1638 1639 // fold vector ops 1640 if (VT.isVector()) { 1641 SDValue FoldedVOp = SimplifyVBinOp(N); 1642 if (FoldedVOp.getNode()) return FoldedVOp; 1643 1644 // fold (sub x, 0) -> x, vector edition 1645 if (ISD::isBuildVectorAllZeros(N1.getNode())) 1646 return N0; 1647 } 1648 1649 // fold (sub x, x) -> 0 1650 // FIXME: Refactor this and xor and other similar operations together. 1651 if (N0 == N1) 1652 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations); 1653 // fold (sub c1, c2) -> c1-c2 1654 if (N0C && N1C) 1655 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C); 1656 // fold (sub x, c) -> (add x, -c) 1657 if (N1C) 1658 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, 1659 DAG.getConstant(-N1C->getAPIntValue(), VT)); 1660 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) 1661 if (N0C && N0C->isAllOnesValue()) 1662 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0); 1663 // fold A-(A-B) -> B 1664 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0)) 1665 return N1.getOperand(1); 1666 // fold (A+B)-A -> B 1667 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1) 1668 return N0.getOperand(1); 1669 // fold (A+B)-B -> A 1670 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1) 1671 return N0.getOperand(0); 1672 // fold C2-(A+C1) -> (C2-C1)-A 1673 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) { 1674 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(), 1675 VT); 1676 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC, 1677 N1.getOperand(0)); 1678 } 1679 // fold ((A+(B+or-C))-B) -> A+or-C 1680 if (N0.getOpcode() == ISD::ADD && 1681 (N0.getOperand(1).getOpcode() == ISD::SUB || 1682 N0.getOperand(1).getOpcode() == ISD::ADD) && 1683 N0.getOperand(1).getOperand(0) == N1) 1684 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT, 1685 N0.getOperand(0), N0.getOperand(1).getOperand(1)); 1686 // fold ((A+(C+B))-B) -> A+C 1687 if (N0.getOpcode() == ISD::ADD && 1688 N0.getOperand(1).getOpcode() == ISD::ADD && 1689 N0.getOperand(1).getOperand(1) == N1) 1690 return DAG.getNode(ISD::ADD, SDLoc(N), VT, 1691 N0.getOperand(0), N0.getOperand(1).getOperand(0)); 1692 // fold ((A-(B-C))-C) -> A-B 1693 if (N0.getOpcode() == ISD::SUB && 1694 N0.getOperand(1).getOpcode() == ISD::SUB && 1695 N0.getOperand(1).getOperand(1) == N1) 1696 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1697 N0.getOperand(0), N0.getOperand(1).getOperand(0)); 1698 1699 // If either operand of a sub is undef, the result is undef 1700 if (N0.getOpcode() == ISD::UNDEF) 1701 return N0; 1702 if (N1.getOpcode() == ISD::UNDEF) 1703 return N1; 1704 1705 // If the relocation model supports it, consider symbol offsets. 1706 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0)) 1707 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) { 1708 // fold (sub Sym, c) -> Sym-c 1709 if (N1C && GA->getOpcode() == ISD::GlobalAddress) 1710 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT, 1711 GA->getOffset() - 1712 (uint64_t)N1C->getSExtValue()); 1713 // fold (sub Sym+c1, Sym+c2) -> c1-c2 1714 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1)) 1715 if (GA->getGlobal() == GB->getGlobal()) 1716 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(), 1717 VT); 1718 } 1719 1720 return SDValue(); 1721} 1722 1723SDValue DAGCombiner::visitSUBC(SDNode *N) { 1724 SDValue N0 = N->getOperand(0); 1725 SDValue N1 = N->getOperand(1); 1726 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1727 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1728 EVT VT = N0.getValueType(); 1729 1730 // If the flag result is dead, turn this into an SUB. 1731 if (!N->hasAnyUseOfValue(1)) 1732 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1), 1733 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N), 1734 MVT::Glue)); 1735 1736 // fold (subc x, x) -> 0 + no borrow 1737 if (N0 == N1) 1738 return CombineTo(N, DAG.getConstant(0, VT), 1739 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N), 1740 MVT::Glue)); 1741 1742 // fold (subc x, 0) -> x + no borrow 1743 if (N1C && N1C->isNullValue()) 1744 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N), 1745 MVT::Glue)); 1746 1747 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow 1748 if (N0C && N0C->isAllOnesValue()) 1749 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0), 1750 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N), 1751 MVT::Glue)); 1752 1753 return SDValue(); 1754} 1755 1756SDValue DAGCombiner::visitSUBE(SDNode *N) { 1757 SDValue N0 = N->getOperand(0); 1758 SDValue N1 = N->getOperand(1); 1759 SDValue CarryIn = N->getOperand(2); 1760 1761 // fold (sube x, y, false) -> (subc x, y) 1762 if (CarryIn.getOpcode() == ISD::CARRY_FALSE) 1763 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1); 1764 1765 return SDValue(); 1766} 1767 1768/// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are 1769/// all the same constant or undefined. 1770static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) { 1771 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N); 1772 if (!C) 1773 return false; 1774 1775 APInt SplatUndef; 1776 unsigned SplatBitSize; 1777 bool HasAnyUndefs; 1778 EVT EltVT = N->getValueType(0).getVectorElementType(); 1779 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, 1780 HasAnyUndefs) && 1781 EltVT.getSizeInBits() >= SplatBitSize); 1782} 1783 1784SDValue DAGCombiner::visitMUL(SDNode *N) { 1785 SDValue N0 = N->getOperand(0); 1786 SDValue N1 = N->getOperand(1); 1787 EVT VT = N0.getValueType(); 1788 1789 // fold (mul x, undef) -> 0 1790 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) 1791 return DAG.getConstant(0, VT); 1792 1793 bool N0IsConst = false; 1794 bool N1IsConst = false; 1795 APInt ConstValue0, ConstValue1; 1796 // fold vector ops 1797 if (VT.isVector()) { 1798 SDValue FoldedVOp = SimplifyVBinOp(N); 1799 if (FoldedVOp.getNode()) return FoldedVOp; 1800 1801 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0); 1802 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1); 1803 } else { 1804 N0IsConst = dyn_cast<ConstantSDNode>(N0) != 0; 1805 ConstValue0 = N0IsConst? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue() : APInt(); 1806 N1IsConst = dyn_cast<ConstantSDNode>(N1) != 0; 1807 ConstValue1 = N1IsConst? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue() : APInt(); 1808 } 1809 1810 // fold (mul c1, c2) -> c1*c2 1811 if (N0IsConst && N1IsConst) 1812 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode()); 1813 1814 // canonicalize constant to RHS 1815 if (N0IsConst && !N1IsConst) 1816 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0); 1817 // fold (mul x, 0) -> 0 1818 if (N1IsConst && ConstValue1 == 0) 1819 return N1; 1820 // fold (mul x, 1) -> x 1821 if (N1IsConst && ConstValue1 == 1) 1822 return N0; 1823 // fold (mul x, -1) -> 0-x 1824 if (N1IsConst && ConstValue1.isAllOnesValue()) 1825 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1826 DAG.getConstant(0, VT), N0); 1827 // fold (mul x, (1 << c)) -> x << c 1828 if (N1IsConst && ConstValue1.isPowerOf2()) 1829 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, 1830 DAG.getConstant(ConstValue1.logBase2(), 1831 getShiftAmountTy(N0.getValueType()))); 1832 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c 1833 if (N1IsConst && (-ConstValue1).isPowerOf2()) { 1834 unsigned Log2Val = (-ConstValue1).logBase2(); 1835 // FIXME: If the input is something that is easily negated (e.g. a 1836 // single-use add), we should put the negate there. 1837 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1838 DAG.getConstant(0, VT), 1839 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, 1840 DAG.getConstant(Log2Val, 1841 getShiftAmountTy(N0.getValueType())))); 1842 } 1843 1844 APInt Val; 1845 // (mul (shl X, c1), c2) -> (mul X, c2 << c1) 1846 if (N1IsConst && N0.getOpcode() == ISD::SHL && 1847 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) || 1848 isa<ConstantSDNode>(N0.getOperand(1)))) { 1849 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT, 1850 N1, N0.getOperand(1)); 1851 AddToWorkList(C3.getNode()); 1852 return DAG.getNode(ISD::MUL, SDLoc(N), VT, 1853 N0.getOperand(0), C3); 1854 } 1855 1856 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one 1857 // use. 1858 { 1859 SDValue Sh(0,0), Y(0,0); 1860 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)). 1861 if (N0.getOpcode() == ISD::SHL && 1862 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) || 1863 isa<ConstantSDNode>(N0.getOperand(1))) && 1864 N0.getNode()->hasOneUse()) { 1865 Sh = N0; Y = N1; 1866 } else if (N1.getOpcode() == ISD::SHL && 1867 isa<ConstantSDNode>(N1.getOperand(1)) && 1868 N1.getNode()->hasOneUse()) { 1869 Sh = N1; Y = N0; 1870 } 1871 1872 if (Sh.getNode()) { 1873 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, 1874 Sh.getOperand(0), Y); 1875 return DAG.getNode(ISD::SHL, SDLoc(N), VT, 1876 Mul, Sh.getOperand(1)); 1877 } 1878 } 1879 1880 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2) 1881 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() && 1882 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) || 1883 isa<ConstantSDNode>(N0.getOperand(1)))) 1884 return DAG.getNode(ISD::ADD, SDLoc(N), VT, 1885 DAG.getNode(ISD::MUL, SDLoc(N0), VT, 1886 N0.getOperand(0), N1), 1887 DAG.getNode(ISD::MUL, SDLoc(N1), VT, 1888 N0.getOperand(1), N1)); 1889 1890 // reassociate mul 1891 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1); 1892 if (RMUL.getNode() != 0) 1893 return RMUL; 1894 1895 return SDValue(); 1896} 1897 1898SDValue DAGCombiner::visitSDIV(SDNode *N) { 1899 SDValue N0 = N->getOperand(0); 1900 SDValue N1 = N->getOperand(1); 1901 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode()); 1902 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 1903 EVT VT = N->getValueType(0); 1904 1905 // fold vector ops 1906 if (VT.isVector()) { 1907 SDValue FoldedVOp = SimplifyVBinOp(N); 1908 if (FoldedVOp.getNode()) return FoldedVOp; 1909 } 1910 1911 // fold (sdiv c1, c2) -> c1/c2 1912 if (N0C && N1C && !N1C->isNullValue()) 1913 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C); 1914 // fold (sdiv X, 1) -> X 1915 if (N1C && N1C->getAPIntValue() == 1LL) 1916 return N0; 1917 // fold (sdiv X, -1) -> 0-X 1918 if (N1C && N1C->isAllOnesValue()) 1919 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1920 DAG.getConstant(0, VT), N0); 1921 // If we know the sign bits of both operands are zero, strength reduce to a 1922 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2 1923 if (!VT.isVector()) { 1924 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) 1925 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(), 1926 N0, N1); 1927 } 1928 // fold (sdiv X, pow2) -> simple ops after legalize 1929 if (N1C && !N1C->isNullValue() && 1930 (N1C->getAPIntValue().isPowerOf2() || 1931 (-N1C->getAPIntValue()).isPowerOf2())) { 1932 // If dividing by powers of two is cheap, then don't perform the following 1933 // fold. 1934 if (TLI.isPow2DivCheap()) 1935 return SDValue(); 1936 1937 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros(); 1938 1939 // Splat the sign bit into the register 1940 SDValue SGN = DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, 1941 DAG.getConstant(VT.getSizeInBits()-1, 1942 getShiftAmountTy(N0.getValueType()))); 1943 AddToWorkList(SGN.getNode()); 1944 1945 // Add (N0 < 0) ? abs2 - 1 : 0; 1946 SDValue SRL = DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN, 1947 DAG.getConstant(VT.getSizeInBits() - lg2, 1948 getShiftAmountTy(SGN.getValueType()))); 1949 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL); 1950 AddToWorkList(SRL.getNode()); 1951 AddToWorkList(ADD.getNode()); // Divide by pow2 1952 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD, 1953 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType()))); 1954 1955 // If we're dividing by a positive value, we're done. Otherwise, we must 1956 // negate the result. 1957 if (N1C->getAPIntValue().isNonNegative()) 1958 return SRA; 1959 1960 AddToWorkList(SRA.getNode()); 1961 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1962 DAG.getConstant(0, VT), SRA); 1963 } 1964 1965 // if integer divide is expensive and we satisfy the requirements, emit an 1966 // alternate sequence. 1967 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) { 1968 SDValue Op = BuildSDIV(N); 1969 if (Op.getNode()) return Op; 1970 } 1971 1972 // undef / X -> 0 1973 if (N0.getOpcode() == ISD::UNDEF) 1974 return DAG.getConstant(0, VT); 1975 // X / undef -> undef 1976 if (N1.getOpcode() == ISD::UNDEF) 1977 return N1; 1978 1979 return SDValue(); 1980} 1981 1982SDValue DAGCombiner::visitUDIV(SDNode *N) { 1983 SDValue N0 = N->getOperand(0); 1984 SDValue N1 = N->getOperand(1); 1985 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode()); 1986 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 1987 EVT VT = N->getValueType(0); 1988 1989 // fold vector ops 1990 if (VT.isVector()) { 1991 SDValue FoldedVOp = SimplifyVBinOp(N); 1992 if (FoldedVOp.getNode()) return FoldedVOp; 1993 } 1994 1995 // fold (udiv c1, c2) -> c1/c2 1996 if (N0C && N1C && !N1C->isNullValue()) 1997 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C); 1998 // fold (udiv x, (1 << c)) -> x >>u c 1999 if (N1C && N1C->getAPIntValue().isPowerOf2()) 2000 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, 2001 DAG.getConstant(N1C->getAPIntValue().logBase2(), 2002 getShiftAmountTy(N0.getValueType()))); 2003 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2 2004 if (N1.getOpcode() == ISD::SHL) { 2005 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) { 2006 if (SHC->getAPIntValue().isPowerOf2()) { 2007 EVT ADDVT = N1.getOperand(1).getValueType(); 2008 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT, 2009 N1.getOperand(1), 2010 DAG.getConstant(SHC->getAPIntValue() 2011 .logBase2(), 2012 ADDVT)); 2013 AddToWorkList(Add.getNode()); 2014 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add); 2015 } 2016 } 2017 } 2018 // fold (udiv x, c) -> alternate 2019 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) { 2020 SDValue Op = BuildUDIV(N); 2021 if (Op.getNode()) return Op; 2022 } 2023 2024 // undef / X -> 0 2025 if (N0.getOpcode() == ISD::UNDEF) 2026 return DAG.getConstant(0, VT); 2027 // X / undef -> undef 2028 if (N1.getOpcode() == ISD::UNDEF) 2029 return N1; 2030 2031 return SDValue(); 2032} 2033 2034SDValue DAGCombiner::visitSREM(SDNode *N) { 2035 SDValue N0 = N->getOperand(0); 2036 SDValue N1 = N->getOperand(1); 2037 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2038 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2039 EVT VT = N->getValueType(0); 2040 2041 // fold (srem c1, c2) -> c1%c2 2042 if (N0C && N1C && !N1C->isNullValue()) 2043 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C); 2044 // If we know the sign bits of both operands are zero, strength reduce to a 2045 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15 2046 if (!VT.isVector()) { 2047 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) 2048 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1); 2049 } 2050 2051 // If X/C can be simplified by the division-by-constant logic, lower 2052 // X%C to the equivalent of X-X/C*C. 2053 if (N1C && !N1C->isNullValue()) { 2054 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1); 2055 AddToWorkList(Div.getNode()); 2056 SDValue OptimizedDiv = combine(Div.getNode()); 2057 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) { 2058 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, 2059 OptimizedDiv, N1); 2060 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul); 2061 AddToWorkList(Mul.getNode()); 2062 return Sub; 2063 } 2064 } 2065 2066 // undef % X -> 0 2067 if (N0.getOpcode() == ISD::UNDEF) 2068 return DAG.getConstant(0, VT); 2069 // X % undef -> undef 2070 if (N1.getOpcode() == ISD::UNDEF) 2071 return N1; 2072 2073 return SDValue(); 2074} 2075 2076SDValue DAGCombiner::visitUREM(SDNode *N) { 2077 SDValue N0 = N->getOperand(0); 2078 SDValue N1 = N->getOperand(1); 2079 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2080 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2081 EVT VT = N->getValueType(0); 2082 2083 // fold (urem c1, c2) -> c1%c2 2084 if (N0C && N1C && !N1C->isNullValue()) 2085 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C); 2086 // fold (urem x, pow2) -> (and x, pow2-1) 2087 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2()) 2088 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, 2089 DAG.getConstant(N1C->getAPIntValue()-1,VT)); 2090 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1)) 2091 if (N1.getOpcode() == ISD::SHL) { 2092 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) { 2093 if (SHC->getAPIntValue().isPowerOf2()) { 2094 SDValue Add = 2095 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, 2096 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), 2097 VT)); 2098 AddToWorkList(Add.getNode()); 2099 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add); 2100 } 2101 } 2102 } 2103 2104 // If X/C can be simplified by the division-by-constant logic, lower 2105 // X%C to the equivalent of X-X/C*C. 2106 if (N1C && !N1C->isNullValue()) { 2107 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1); 2108 AddToWorkList(Div.getNode()); 2109 SDValue OptimizedDiv = combine(Div.getNode()); 2110 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) { 2111 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, 2112 OptimizedDiv, N1); 2113 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul); 2114 AddToWorkList(Mul.getNode()); 2115 return Sub; 2116 } 2117 } 2118 2119 // undef % X -> 0 2120 if (N0.getOpcode() == ISD::UNDEF) 2121 return DAG.getConstant(0, VT); 2122 // X % undef -> undef 2123 if (N1.getOpcode() == ISD::UNDEF) 2124 return N1; 2125 2126 return SDValue(); 2127} 2128 2129SDValue DAGCombiner::visitMULHS(SDNode *N) { 2130 SDValue N0 = N->getOperand(0); 2131 SDValue N1 = N->getOperand(1); 2132 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2133 EVT VT = N->getValueType(0); 2134 SDLoc DL(N); 2135 2136 // fold (mulhs x, 0) -> 0 2137 if (N1C && N1C->isNullValue()) 2138 return N1; 2139 // fold (mulhs x, 1) -> (sra x, size(x)-1) 2140 if (N1C && N1C->getAPIntValue() == 1) 2141 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0, 2142 DAG.getConstant(N0.getValueType().getSizeInBits() - 1, 2143 getShiftAmountTy(N0.getValueType()))); 2144 // fold (mulhs x, undef) -> 0 2145 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) 2146 return DAG.getConstant(0, VT); 2147 2148 // If the type twice as wide is legal, transform the mulhs to a wider multiply 2149 // plus a shift. 2150 if (VT.isSimple() && !VT.isVector()) { 2151 MVT Simple = VT.getSimpleVT(); 2152 unsigned SimpleSize = Simple.getSizeInBits(); 2153 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); 2154 if (TLI.isOperationLegal(ISD::MUL, NewVT)) { 2155 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0); 2156 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1); 2157 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); 2158 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, 2159 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType()))); 2160 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); 2161 } 2162 } 2163 2164 return SDValue(); 2165} 2166 2167SDValue DAGCombiner::visitMULHU(SDNode *N) { 2168 SDValue N0 = N->getOperand(0); 2169 SDValue N1 = N->getOperand(1); 2170 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2171 EVT VT = N->getValueType(0); 2172 SDLoc DL(N); 2173 2174 // fold (mulhu x, 0) -> 0 2175 if (N1C && N1C->isNullValue()) 2176 return N1; 2177 // fold (mulhu x, 1) -> 0 2178 if (N1C && N1C->getAPIntValue() == 1) 2179 return DAG.getConstant(0, N0.getValueType()); 2180 // fold (mulhu x, undef) -> 0 2181 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) 2182 return DAG.getConstant(0, VT); 2183 2184 // If the type twice as wide is legal, transform the mulhu to a wider multiply 2185 // plus a shift. 2186 if (VT.isSimple() && !VT.isVector()) { 2187 MVT Simple = VT.getSimpleVT(); 2188 unsigned SimpleSize = Simple.getSizeInBits(); 2189 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); 2190 if (TLI.isOperationLegal(ISD::MUL, NewVT)) { 2191 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0); 2192 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1); 2193 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); 2194 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, 2195 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType()))); 2196 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); 2197 } 2198 } 2199 2200 return SDValue(); 2201} 2202 2203/// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that 2204/// compute two values. LoOp and HiOp give the opcodes for the two computations 2205/// that are being performed. Return true if a simplification was made. 2206/// 2207SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, 2208 unsigned HiOp) { 2209 // If the high half is not needed, just compute the low half. 2210 bool HiExists = N->hasAnyUseOfValue(1); 2211 if (!HiExists && 2212 (!LegalOperations || 2213 TLI.isOperationLegal(LoOp, N->getValueType(0)))) { 2214 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), 2215 N->op_begin(), N->getNumOperands()); 2216 return CombineTo(N, Res, Res); 2217 } 2218 2219 // If the low half is not needed, just compute the high half. 2220 bool LoExists = N->hasAnyUseOfValue(0); 2221 if (!LoExists && 2222 (!LegalOperations || 2223 TLI.isOperationLegal(HiOp, N->getValueType(1)))) { 2224 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), 2225 N->op_begin(), N->getNumOperands()); 2226 return CombineTo(N, Res, Res); 2227 } 2228 2229 // If both halves are used, return as it is. 2230 if (LoExists && HiExists) 2231 return SDValue(); 2232 2233 // If the two computed results can be simplified separately, separate them. 2234 if (LoExists) { 2235 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), 2236 N->op_begin(), N->getNumOperands()); 2237 AddToWorkList(Lo.getNode()); 2238 SDValue LoOpt = combine(Lo.getNode()); 2239 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() && 2240 (!LegalOperations || 2241 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType()))) 2242 return CombineTo(N, LoOpt, LoOpt); 2243 } 2244 2245 if (HiExists) { 2246 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), 2247 N->op_begin(), N->getNumOperands()); 2248 AddToWorkList(Hi.getNode()); 2249 SDValue HiOpt = combine(Hi.getNode()); 2250 if (HiOpt.getNode() && HiOpt != Hi && 2251 (!LegalOperations || 2252 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType()))) 2253 return CombineTo(N, HiOpt, HiOpt); 2254 } 2255 2256 return SDValue(); 2257} 2258 2259SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) { 2260 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS); 2261 if (Res.getNode()) return Res; 2262 2263 EVT VT = N->getValueType(0); 2264 SDLoc DL(N); 2265 2266 // If the type twice as wide is legal, transform the mulhu to a wider multiply 2267 // plus a shift. 2268 if (VT.isSimple() && !VT.isVector()) { 2269 MVT Simple = VT.getSimpleVT(); 2270 unsigned SimpleSize = Simple.getSizeInBits(); 2271 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); 2272 if (TLI.isOperationLegal(ISD::MUL, NewVT)) { 2273 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0)); 2274 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1)); 2275 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); 2276 // Compute the high part as N1. 2277 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, 2278 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType()))); 2279 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); 2280 // Compute the low part as N0. 2281 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); 2282 return CombineTo(N, Lo, Hi); 2283 } 2284 } 2285 2286 return SDValue(); 2287} 2288 2289SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) { 2290 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU); 2291 if (Res.getNode()) return Res; 2292 2293 EVT VT = N->getValueType(0); 2294 SDLoc DL(N); 2295 2296 // If the type twice as wide is legal, transform the mulhu to a wider multiply 2297 // plus a shift. 2298 if (VT.isSimple() && !VT.isVector()) { 2299 MVT Simple = VT.getSimpleVT(); 2300 unsigned SimpleSize = Simple.getSizeInBits(); 2301 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); 2302 if (TLI.isOperationLegal(ISD::MUL, NewVT)) { 2303 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0)); 2304 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1)); 2305 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); 2306 // Compute the high part as N1. 2307 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, 2308 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType()))); 2309 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); 2310 // Compute the low part as N0. 2311 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); 2312 return CombineTo(N, Lo, Hi); 2313 } 2314 } 2315 2316 return SDValue(); 2317} 2318 2319SDValue DAGCombiner::visitSMULO(SDNode *N) { 2320 // (smulo x, 2) -> (saddo x, x) 2321 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1))) 2322 if (C2->getAPIntValue() == 2) 2323 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(), 2324 N->getOperand(0), N->getOperand(0)); 2325 2326 return SDValue(); 2327} 2328 2329SDValue DAGCombiner::visitUMULO(SDNode *N) { 2330 // (umulo x, 2) -> (uaddo x, x) 2331 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1))) 2332 if (C2->getAPIntValue() == 2) 2333 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(), 2334 N->getOperand(0), N->getOperand(0)); 2335 2336 return SDValue(); 2337} 2338 2339SDValue DAGCombiner::visitSDIVREM(SDNode *N) { 2340 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM); 2341 if (Res.getNode()) return Res; 2342 2343 return SDValue(); 2344} 2345 2346SDValue DAGCombiner::visitUDIVREM(SDNode *N) { 2347 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM); 2348 if (Res.getNode()) return Res; 2349 2350 return SDValue(); 2351} 2352 2353/// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with 2354/// two operands of the same opcode, try to simplify it. 2355SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) { 2356 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 2357 EVT VT = N0.getValueType(); 2358 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!"); 2359 2360 // Bail early if none of these transforms apply. 2361 if (N0.getNode()->getNumOperands() == 0) return SDValue(); 2362 2363 // For each of OP in AND/OR/XOR: 2364 // fold (OP (zext x), (zext y)) -> (zext (OP x, y)) 2365 // fold (OP (sext x), (sext y)) -> (sext (OP x, y)) 2366 // fold (OP (aext x), (aext y)) -> (aext (OP x, y)) 2367 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free) 2368 // 2369 // do not sink logical op inside of a vector extend, since it may combine 2370 // into a vsetcc. 2371 EVT Op0VT = N0.getOperand(0).getValueType(); 2372 if ((N0.getOpcode() == ISD::ZERO_EXTEND || 2373 N0.getOpcode() == ISD::SIGN_EXTEND || 2374 // Avoid infinite looping with PromoteIntBinOp. 2375 (N0.getOpcode() == ISD::ANY_EXTEND && 2376 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) || 2377 (N0.getOpcode() == ISD::TRUNCATE && 2378 (!TLI.isZExtFree(VT, Op0VT) || 2379 !TLI.isTruncateFree(Op0VT, VT)) && 2380 TLI.isTypeLegal(Op0VT))) && 2381 !VT.isVector() && 2382 Op0VT == N1.getOperand(0).getValueType() && 2383 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) { 2384 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0), 2385 N0.getOperand(0).getValueType(), 2386 N0.getOperand(0), N1.getOperand(0)); 2387 AddToWorkList(ORNode.getNode()); 2388 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode); 2389 } 2390 2391 // For each of OP in SHL/SRL/SRA/AND... 2392 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z) 2393 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z) 2394 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z) 2395 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL || 2396 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) && 2397 N0.getOperand(1) == N1.getOperand(1)) { 2398 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0), 2399 N0.getOperand(0).getValueType(), 2400 N0.getOperand(0), N1.getOperand(0)); 2401 AddToWorkList(ORNode.getNode()); 2402 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, 2403 ORNode, N0.getOperand(1)); 2404 } 2405 2406 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B)) 2407 // Only perform this optimization after type legalization and before 2408 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by 2409 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and 2410 // we don't want to undo this promotion. 2411 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper 2412 // on scalars. 2413 if ((N0.getOpcode() == ISD::BITCAST || 2414 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) && 2415 Level == AfterLegalizeTypes) { 2416 SDValue In0 = N0.getOperand(0); 2417 SDValue In1 = N1.getOperand(0); 2418 EVT In0Ty = In0.getValueType(); 2419 EVT In1Ty = In1.getValueType(); 2420 SDLoc DL(N); 2421 // If both incoming values are integers, and the original types are the 2422 // same. 2423 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) { 2424 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1); 2425 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op); 2426 AddToWorkList(Op.getNode()); 2427 return BC; 2428 } 2429 } 2430 2431 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value). 2432 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B)) 2433 // If both shuffles use the same mask, and both shuffle within a single 2434 // vector, then it is worthwhile to move the swizzle after the operation. 2435 // The type-legalizer generates this pattern when loading illegal 2436 // vector types from memory. In many cases this allows additional shuffle 2437 // optimizations. 2438 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG && 2439 N0.getOperand(1).getOpcode() == ISD::UNDEF && 2440 N1.getOperand(1).getOpcode() == ISD::UNDEF) { 2441 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0); 2442 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1); 2443 2444 assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() && 2445 "Inputs to shuffles are not the same type"); 2446 2447 unsigned NumElts = VT.getVectorNumElements(); 2448 2449 // Check that both shuffles use the same mask. The masks are known to be of 2450 // the same length because the result vector type is the same. 2451 bool SameMask = true; 2452 for (unsigned i = 0; i != NumElts; ++i) { 2453 int Idx0 = SVN0->getMaskElt(i); 2454 int Idx1 = SVN1->getMaskElt(i); 2455 if (Idx0 != Idx1) { 2456 SameMask = false; 2457 break; 2458 } 2459 } 2460 2461 if (SameMask) { 2462 SDValue Op = DAG.getNode(N->getOpcode(), SDLoc(N), VT, 2463 N0.getOperand(0), N1.getOperand(0)); 2464 AddToWorkList(Op.getNode()); 2465 return DAG.getVectorShuffle(VT, SDLoc(N), Op, 2466 DAG.getUNDEF(VT), &SVN0->getMask()[0]); 2467 } 2468 } 2469 2470 return SDValue(); 2471} 2472 2473SDValue DAGCombiner::visitAND(SDNode *N) { 2474 SDValue N0 = N->getOperand(0); 2475 SDValue N1 = N->getOperand(1); 2476 SDValue LL, LR, RL, RR, CC0, CC1; 2477 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2478 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2479 EVT VT = N1.getValueType(); 2480 unsigned BitWidth = VT.getScalarType().getSizeInBits(); 2481 2482 // fold vector ops 2483 if (VT.isVector()) { 2484 SDValue FoldedVOp = SimplifyVBinOp(N); 2485 if (FoldedVOp.getNode()) return FoldedVOp; 2486 2487 // fold (and x, 0) -> 0, vector edition 2488 if (ISD::isBuildVectorAllZeros(N0.getNode())) 2489 return N0; 2490 if (ISD::isBuildVectorAllZeros(N1.getNode())) 2491 return N1; 2492 2493 // fold (and x, -1) -> x, vector edition 2494 if (ISD::isBuildVectorAllOnes(N0.getNode())) 2495 return N1; 2496 if (ISD::isBuildVectorAllOnes(N1.getNode())) 2497 return N0; 2498 } 2499 2500 // fold (and x, undef) -> 0 2501 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) 2502 return DAG.getConstant(0, VT); 2503 // fold (and c1, c2) -> c1&c2 2504 if (N0C && N1C) 2505 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C); 2506 // canonicalize constant to RHS 2507 if (N0C && !N1C) 2508 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0); 2509 // fold (and x, -1) -> x 2510 if (N1C && N1C->isAllOnesValue()) 2511 return N0; 2512 // if (and x, c) is known to be zero, return 0 2513 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), 2514 APInt::getAllOnesValue(BitWidth))) 2515 return DAG.getConstant(0, VT); 2516 // reassociate and 2517 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1); 2518 if (RAND.getNode() != 0) 2519 return RAND; 2520 // fold (and (or x, C), D) -> D if (C & D) == D 2521 if (N1C && N0.getOpcode() == ISD::OR) 2522 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 2523 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue()) 2524 return N1; 2525 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits. 2526 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { 2527 SDValue N0Op0 = N0.getOperand(0); 2528 APInt Mask = ~N1C->getAPIntValue(); 2529 Mask = Mask.trunc(N0Op0.getValueSizeInBits()); 2530 if (DAG.MaskedValueIsZero(N0Op0, Mask)) { 2531 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), 2532 N0.getValueType(), N0Op0); 2533 2534 // Replace uses of the AND with uses of the Zero extend node. 2535 CombineTo(N, Zext); 2536 2537 // We actually want to replace all uses of the any_extend with the 2538 // zero_extend, to avoid duplicating things. This will later cause this 2539 // AND to be folded. 2540 CombineTo(N0.getNode(), Zext); 2541 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2542 } 2543 } 2544 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) -> 2545 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must 2546 // already be zero by virtue of the width of the base type of the load. 2547 // 2548 // the 'X' node here can either be nothing or an extract_vector_elt to catch 2549 // more cases. 2550 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && 2551 N0.getOperand(0).getOpcode() == ISD::LOAD) || 2552 N0.getOpcode() == ISD::LOAD) { 2553 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ? 2554 N0 : N0.getOperand(0) ); 2555 2556 // Get the constant (if applicable) the zero'th operand is being ANDed with. 2557 // This can be a pure constant or a vector splat, in which case we treat the 2558 // vector as a scalar and use the splat value. 2559 APInt Constant = APInt::getNullValue(1); 2560 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { 2561 Constant = C->getAPIntValue(); 2562 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) { 2563 APInt SplatValue, SplatUndef; 2564 unsigned SplatBitSize; 2565 bool HasAnyUndefs; 2566 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef, 2567 SplatBitSize, HasAnyUndefs); 2568 if (IsSplat) { 2569 // Undef bits can contribute to a possible optimisation if set, so 2570 // set them. 2571 SplatValue |= SplatUndef; 2572 2573 // The splat value may be something like "0x00FFFFFF", which means 0 for 2574 // the first vector value and FF for the rest, repeating. We need a mask 2575 // that will apply equally to all members of the vector, so AND all the 2576 // lanes of the constant together. 2577 EVT VT = Vector->getValueType(0); 2578 unsigned BitWidth = VT.getVectorElementType().getSizeInBits(); 2579 2580 // If the splat value has been compressed to a bitlength lower 2581 // than the size of the vector lane, we need to re-expand it to 2582 // the lane size. 2583 if (BitWidth > SplatBitSize) 2584 for (SplatValue = SplatValue.zextOrTrunc(BitWidth); 2585 SplatBitSize < BitWidth; 2586 SplatBitSize = SplatBitSize * 2) 2587 SplatValue |= SplatValue.shl(SplatBitSize); 2588 2589 Constant = APInt::getAllOnesValue(BitWidth); 2590 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i) 2591 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth); 2592 } 2593 } 2594 2595 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is 2596 // actually legal and isn't going to get expanded, else this is a false 2597 // optimisation. 2598 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD, 2599 Load->getMemoryVT()); 2600 2601 // Resize the constant to the same size as the original memory access before 2602 // extension. If it is still the AllOnesValue then this AND is completely 2603 // unneeded. 2604 Constant = 2605 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits()); 2606 2607 bool B; 2608 switch (Load->getExtensionType()) { 2609 default: B = false; break; 2610 case ISD::EXTLOAD: B = CanZextLoadProfitably; break; 2611 case ISD::ZEXTLOAD: 2612 case ISD::NON_EXTLOAD: B = true; break; 2613 } 2614 2615 if (B && Constant.isAllOnesValue()) { 2616 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to 2617 // preserve semantics once we get rid of the AND. 2618 SDValue NewLoad(Load, 0); 2619 if (Load->getExtensionType() == ISD::EXTLOAD) { 2620 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD, 2621 Load->getValueType(0), SDLoc(Load), 2622 Load->getChain(), Load->getBasePtr(), 2623 Load->getOffset(), Load->getMemoryVT(), 2624 Load->getMemOperand()); 2625 // Replace uses of the EXTLOAD with the new ZEXTLOAD. 2626 if (Load->getNumValues() == 3) { 2627 // PRE/POST_INC loads have 3 values. 2628 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1), 2629 NewLoad.getValue(2) }; 2630 CombineTo(Load, To, 3, true); 2631 } else { 2632 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1)); 2633 } 2634 } 2635 2636 // Fold the AND away, taking care not to fold to the old load node if we 2637 // replaced it. 2638 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0); 2639 2640 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2641 } 2642 } 2643 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y)) 2644 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 2645 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 2646 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 2647 2648 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 2649 LL.getValueType().isInteger()) { 2650 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0) 2651 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) { 2652 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0), 2653 LR.getValueType(), LL, RL); 2654 AddToWorkList(ORNode.getNode()); 2655 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1); 2656 } 2657 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1) 2658 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) { 2659 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0), 2660 LR.getValueType(), LL, RL); 2661 AddToWorkList(ANDNode.getNode()); 2662 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1); 2663 } 2664 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1) 2665 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) { 2666 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0), 2667 LR.getValueType(), LL, RL); 2668 AddToWorkList(ORNode.getNode()); 2669 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1); 2670 } 2671 } 2672 // canonicalize equivalent to ll == rl 2673 if (LL == RR && LR == RL) { 2674 Op1 = ISD::getSetCCSwappedOperands(Op1); 2675 std::swap(RL, RR); 2676 } 2677 if (LL == RL && LR == RR) { 2678 bool isInteger = LL.getValueType().isInteger(); 2679 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger); 2680 if (Result != ISD::SETCC_INVALID && 2681 (!LegalOperations || 2682 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) && 2683 TLI.isOperationLegal(ISD::SETCC, 2684 getSetCCResultType(N0.getSimpleValueType()))))) 2685 return DAG.getSetCC(SDLoc(N), N0.getValueType(), 2686 LL, LR, Result); 2687 } 2688 } 2689 2690 // Simplify: (and (op x...), (op y...)) -> (op (and x, y)) 2691 if (N0.getOpcode() == N1.getOpcode()) { 2692 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N); 2693 if (Tmp.getNode()) return Tmp; 2694 } 2695 2696 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1) 2697 // fold (and (sra)) -> (and (srl)) when possible. 2698 if (!VT.isVector() && 2699 SimplifyDemandedBits(SDValue(N, 0))) 2700 return SDValue(N, 0); 2701 2702 // fold (zext_inreg (extload x)) -> (zextload x) 2703 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) { 2704 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2705 EVT MemVT = LN0->getMemoryVT(); 2706 // If we zero all the possible extended bits, then we can turn this into 2707 // a zextload if we are running before legalize or the operation is legal. 2708 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits(); 2709 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth, 2710 BitWidth - MemVT.getScalarType().getSizeInBits())) && 2711 ((!LegalOperations && !LN0->isVolatile()) || 2712 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) { 2713 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT, 2714 LN0->getChain(), LN0->getBasePtr(), 2715 LN0->getPointerInfo(), MemVT, 2716 LN0->isVolatile(), LN0->isNonTemporal(), 2717 LN0->getAlignment()); 2718 AddToWorkList(N); 2719 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); 2720 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2721 } 2722 } 2723 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use 2724 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && 2725 N0.hasOneUse()) { 2726 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2727 EVT MemVT = LN0->getMemoryVT(); 2728 // If we zero all the possible extended bits, then we can turn this into 2729 // a zextload if we are running before legalize or the operation is legal. 2730 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits(); 2731 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth, 2732 BitWidth - MemVT.getScalarType().getSizeInBits())) && 2733 ((!LegalOperations && !LN0->isVolatile()) || 2734 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) { 2735 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT, 2736 LN0->getChain(), 2737 LN0->getBasePtr(), LN0->getPointerInfo(), 2738 MemVT, 2739 LN0->isVolatile(), LN0->isNonTemporal(), 2740 LN0->getAlignment()); 2741 AddToWorkList(N); 2742 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); 2743 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2744 } 2745 } 2746 2747 // fold (and (load x), 255) -> (zextload x, i8) 2748 // fold (and (extload x, i16), 255) -> (zextload x, i8) 2749 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8) 2750 if (N1C && (N0.getOpcode() == ISD::LOAD || 2751 (N0.getOpcode() == ISD::ANY_EXTEND && 2752 N0.getOperand(0).getOpcode() == ISD::LOAD))) { 2753 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND; 2754 LoadSDNode *LN0 = HasAnyExt 2755 ? cast<LoadSDNode>(N0.getOperand(0)) 2756 : cast<LoadSDNode>(N0); 2757 if (LN0->getExtensionType() != ISD::SEXTLOAD && 2758 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) { 2759 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits(); 2760 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){ 2761 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); 2762 EVT LoadedVT = LN0->getMemoryVT(); 2763 2764 if (ExtVT == LoadedVT && 2765 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) { 2766 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT; 2767 2768 SDValue NewLoad = 2769 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy, 2770 LN0->getChain(), LN0->getBasePtr(), 2771 LN0->getPointerInfo(), 2772 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(), 2773 LN0->getAlignment()); 2774 AddToWorkList(N); 2775 CombineTo(LN0, NewLoad, NewLoad.getValue(1)); 2776 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2777 } 2778 2779 // Do not change the width of a volatile load. 2780 // Do not generate loads of non-round integer types since these can 2781 // be expensive (and would be wrong if the type is not byte sized). 2782 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() && 2783 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) { 2784 EVT PtrType = LN0->getOperand(1).getValueType(); 2785 2786 unsigned Alignment = LN0->getAlignment(); 2787 SDValue NewPtr = LN0->getBasePtr(); 2788 2789 // For big endian targets, we need to add an offset to the pointer 2790 // to load the correct bytes. For little endian systems, we merely 2791 // need to read fewer bytes from the same pointer. 2792 if (TLI.isBigEndian()) { 2793 unsigned LVTStoreBytes = LoadedVT.getStoreSize(); 2794 unsigned EVTStoreBytes = ExtVT.getStoreSize(); 2795 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes; 2796 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType, 2797 NewPtr, DAG.getConstant(PtrOff, PtrType)); 2798 Alignment = MinAlign(Alignment, PtrOff); 2799 } 2800 2801 AddToWorkList(NewPtr.getNode()); 2802 2803 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT; 2804 SDValue Load = 2805 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy, 2806 LN0->getChain(), NewPtr, 2807 LN0->getPointerInfo(), 2808 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(), 2809 Alignment); 2810 AddToWorkList(N); 2811 CombineTo(LN0, Load, Load.getValue(1)); 2812 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2813 } 2814 } 2815 } 2816 } 2817 2818 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL && 2819 VT.getSizeInBits() <= 64) { 2820 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2821 APInt ADDC = ADDI->getAPIntValue(); 2822 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) { 2823 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal 2824 // immediate for an add, but it is legal if its top c2 bits are set, 2825 // transform the ADD so the immediate doesn't need to be materialized 2826 // in a register. 2827 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) { 2828 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), 2829 SRLI->getZExtValue()); 2830 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) { 2831 ADDC |= Mask; 2832 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) { 2833 SDValue NewAdd = 2834 DAG.getNode(ISD::ADD, SDLoc(N0), VT, 2835 N0.getOperand(0), DAG.getConstant(ADDC, VT)); 2836 CombineTo(N0.getNode(), NewAdd); 2837 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2838 } 2839 } 2840 } 2841 } 2842 } 2843 } 2844 2845 return SDValue(); 2846} 2847 2848/// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16 2849/// 2850SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, 2851 bool DemandHighBits) { 2852 if (!LegalOperations) 2853 return SDValue(); 2854 2855 EVT VT = N->getValueType(0); 2856 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16) 2857 return SDValue(); 2858 if (!TLI.isOperationLegal(ISD::BSWAP, VT)) 2859 return SDValue(); 2860 2861 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00) 2862 bool LookPassAnd0 = false; 2863 bool LookPassAnd1 = false; 2864 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL) 2865 std::swap(N0, N1); 2866 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL) 2867 std::swap(N0, N1); 2868 if (N0.getOpcode() == ISD::AND) { 2869 if (!N0.getNode()->hasOneUse()) 2870 return SDValue(); 2871 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 2872 if (!N01C || N01C->getZExtValue() != 0xFF00) 2873 return SDValue(); 2874 N0 = N0.getOperand(0); 2875 LookPassAnd0 = true; 2876 } 2877 2878 if (N1.getOpcode() == ISD::AND) { 2879 if (!N1.getNode()->hasOneUse()) 2880 return SDValue(); 2881 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 2882 if (!N11C || N11C->getZExtValue() != 0xFF) 2883 return SDValue(); 2884 N1 = N1.getOperand(0); 2885 LookPassAnd1 = true; 2886 } 2887 2888 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL) 2889 std::swap(N0, N1); 2890 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL) 2891 return SDValue(); 2892 if (!N0.getNode()->hasOneUse() || 2893 !N1.getNode()->hasOneUse()) 2894 return SDValue(); 2895 2896 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 2897 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 2898 if (!N01C || !N11C) 2899 return SDValue(); 2900 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8) 2901 return SDValue(); 2902 2903 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8) 2904 SDValue N00 = N0->getOperand(0); 2905 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) { 2906 if (!N00.getNode()->hasOneUse()) 2907 return SDValue(); 2908 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1)); 2909 if (!N001C || N001C->getZExtValue() != 0xFF) 2910 return SDValue(); 2911 N00 = N00.getOperand(0); 2912 LookPassAnd0 = true; 2913 } 2914 2915 SDValue N10 = N1->getOperand(0); 2916 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) { 2917 if (!N10.getNode()->hasOneUse()) 2918 return SDValue(); 2919 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1)); 2920 if (!N101C || N101C->getZExtValue() != 0xFF00) 2921 return SDValue(); 2922 N10 = N10.getOperand(0); 2923 LookPassAnd1 = true; 2924 } 2925 2926 if (N00 != N10) 2927 return SDValue(); 2928 2929 // Make sure everything beyond the low halfword is zero since the SRL 16 2930 // will clear the top bits. 2931 unsigned OpSizeInBits = VT.getSizeInBits(); 2932 if (DemandHighBits && OpSizeInBits > 16 && 2933 (!LookPassAnd0 || !LookPassAnd1) && 2934 !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16))) 2935 return SDValue(); 2936 2937 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00); 2938 if (OpSizeInBits > 16) 2939 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res, 2940 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT))); 2941 return Res; 2942} 2943 2944/// isBSwapHWordElement - Return true if the specified node is an element 2945/// that makes up a 32-bit packed halfword byteswap. i.e. 2946/// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8) 2947static bool isBSwapHWordElement(SDValue N, SmallVector<SDNode*,4> &Parts) { 2948 if (!N.getNode()->hasOneUse()) 2949 return false; 2950 2951 unsigned Opc = N.getOpcode(); 2952 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL) 2953 return false; 2954 2955 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1)); 2956 if (!N1C) 2957 return false; 2958 2959 unsigned Num; 2960 switch (N1C->getZExtValue()) { 2961 default: 2962 return false; 2963 case 0xFF: Num = 0; break; 2964 case 0xFF00: Num = 1; break; 2965 case 0xFF0000: Num = 2; break; 2966 case 0xFF000000: Num = 3; break; 2967 } 2968 2969 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00). 2970 SDValue N0 = N.getOperand(0); 2971 if (Opc == ISD::AND) { 2972 if (Num == 0 || Num == 2) { 2973 // (x >> 8) & 0xff 2974 // (x >> 8) & 0xff0000 2975 if (N0.getOpcode() != ISD::SRL) 2976 return false; 2977 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 2978 if (!C || C->getZExtValue() != 8) 2979 return false; 2980 } else { 2981 // (x << 8) & 0xff00 2982 // (x << 8) & 0xff000000 2983 if (N0.getOpcode() != ISD::SHL) 2984 return false; 2985 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 2986 if (!C || C->getZExtValue() != 8) 2987 return false; 2988 } 2989 } else if (Opc == ISD::SHL) { 2990 // (x & 0xff) << 8 2991 // (x & 0xff0000) << 8 2992 if (Num != 0 && Num != 2) 2993 return false; 2994 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1)); 2995 if (!C || C->getZExtValue() != 8) 2996 return false; 2997 } else { // Opc == ISD::SRL 2998 // (x & 0xff00) >> 8 2999 // (x & 0xff000000) >> 8 3000 if (Num != 1 && Num != 3) 3001 return false; 3002 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1)); 3003 if (!C || C->getZExtValue() != 8) 3004 return false; 3005 } 3006 3007 if (Parts[Num]) 3008 return false; 3009 3010 Parts[Num] = N0.getOperand(0).getNode(); 3011 return true; 3012} 3013 3014/// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is 3015/// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8) 3016/// => (rotl (bswap x), 16) 3017SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) { 3018 if (!LegalOperations) 3019 return SDValue(); 3020 3021 EVT VT = N->getValueType(0); 3022 if (VT != MVT::i32) 3023 return SDValue(); 3024 if (!TLI.isOperationLegal(ISD::BSWAP, VT)) 3025 return SDValue(); 3026 3027 SmallVector<SDNode*,4> Parts(4, (SDNode*)0); 3028 // Look for either 3029 // (or (or (and), (and)), (or (and), (and))) 3030 // (or (or (or (and), (and)), (and)), (and)) 3031 if (N0.getOpcode() != ISD::OR) 3032 return SDValue(); 3033 SDValue N00 = N0.getOperand(0); 3034 SDValue N01 = N0.getOperand(1); 3035 3036 if (N1.getOpcode() == ISD::OR && 3037 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) { 3038 // (or (or (and), (and)), (or (and), (and))) 3039 SDValue N000 = N00.getOperand(0); 3040 if (!isBSwapHWordElement(N000, Parts)) 3041 return SDValue(); 3042 3043 SDValue N001 = N00.getOperand(1); 3044 if (!isBSwapHWordElement(N001, Parts)) 3045 return SDValue(); 3046 SDValue N010 = N01.getOperand(0); 3047 if (!isBSwapHWordElement(N010, Parts)) 3048 return SDValue(); 3049 SDValue N011 = N01.getOperand(1); 3050 if (!isBSwapHWordElement(N011, Parts)) 3051 return SDValue(); 3052 } else { 3053 // (or (or (or (and), (and)), (and)), (and)) 3054 if (!isBSwapHWordElement(N1, Parts)) 3055 return SDValue(); 3056 if (!isBSwapHWordElement(N01, Parts)) 3057 return SDValue(); 3058 if (N00.getOpcode() != ISD::OR) 3059 return SDValue(); 3060 SDValue N000 = N00.getOperand(0); 3061 if (!isBSwapHWordElement(N000, Parts)) 3062 return SDValue(); 3063 SDValue N001 = N00.getOperand(1); 3064 if (!isBSwapHWordElement(N001, Parts)) 3065 return SDValue(); 3066 } 3067 3068 // Make sure the parts are all coming from the same node. 3069 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3]) 3070 return SDValue(); 3071 3072 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, 3073 SDValue(Parts[0],0)); 3074 3075 // Result of the bswap should be rotated by 16. If it's not legal, than 3076 // do (x << 16) | (x >> 16). 3077 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT)); 3078 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT)) 3079 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt); 3080 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT)) 3081 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt); 3082 return DAG.getNode(ISD::OR, SDLoc(N), VT, 3083 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt), 3084 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt)); 3085} 3086 3087SDValue DAGCombiner::visitOR(SDNode *N) { 3088 SDValue N0 = N->getOperand(0); 3089 SDValue N1 = N->getOperand(1); 3090 SDValue LL, LR, RL, RR, CC0, CC1; 3091 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3092 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3093 EVT VT = N1.getValueType(); 3094 3095 // fold vector ops 3096 if (VT.isVector()) { 3097 SDValue FoldedVOp = SimplifyVBinOp(N); 3098 if (FoldedVOp.getNode()) return FoldedVOp; 3099 3100 // fold (or x, 0) -> x, vector edition 3101 if (ISD::isBuildVectorAllZeros(N0.getNode())) 3102 return N1; 3103 if (ISD::isBuildVectorAllZeros(N1.getNode())) 3104 return N0; 3105 3106 // fold (or x, -1) -> -1, vector edition 3107 if (ISD::isBuildVectorAllOnes(N0.getNode())) 3108 return N0; 3109 if (ISD::isBuildVectorAllOnes(N1.getNode())) 3110 return N1; 3111 } 3112 3113 // fold (or x, undef) -> -1 3114 if (!LegalOperations && 3115 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) { 3116 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT; 3117 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT); 3118 } 3119 // fold (or c1, c2) -> c1|c2 3120 if (N0C && N1C) 3121 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C); 3122 // canonicalize constant to RHS 3123 if (N0C && !N1C) 3124 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0); 3125 // fold (or x, 0) -> x 3126 if (N1C && N1C->isNullValue()) 3127 return N0; 3128 // fold (or x, -1) -> -1 3129 if (N1C && N1C->isAllOnesValue()) 3130 return N1; 3131 // fold (or x, c) -> c iff (x & ~c) == 0 3132 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue())) 3133 return N1; 3134 3135 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16) 3136 SDValue BSwap = MatchBSwapHWord(N, N0, N1); 3137 if (BSwap.getNode() != 0) 3138 return BSwap; 3139 BSwap = MatchBSwapHWordLow(N, N0, N1); 3140 if (BSwap.getNode() != 0) 3141 return BSwap; 3142 3143 // reassociate or 3144 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1); 3145 if (ROR.getNode() != 0) 3146 return ROR; 3147 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2) 3148 // iff (c1 & c2) == 0. 3149 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() && 3150 isa<ConstantSDNode>(N0.getOperand(1))) { 3151 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1)); 3152 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) 3153 return DAG.getNode(ISD::AND, SDLoc(N), VT, 3154 DAG.getNode(ISD::OR, SDLoc(N0), VT, 3155 N0.getOperand(0), N1), 3156 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1)); 3157 } 3158 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y)) 3159 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 3160 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 3161 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 3162 3163 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 3164 LL.getValueType().isInteger()) { 3165 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0) 3166 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0) 3167 if (cast<ConstantSDNode>(LR)->isNullValue() && 3168 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) { 3169 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR), 3170 LR.getValueType(), LL, RL); 3171 AddToWorkList(ORNode.getNode()); 3172 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1); 3173 } 3174 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1) 3175 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1) 3176 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && 3177 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) { 3178 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR), 3179 LR.getValueType(), LL, RL); 3180 AddToWorkList(ANDNode.getNode()); 3181 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1); 3182 } 3183 } 3184 // canonicalize equivalent to ll == rl 3185 if (LL == RR && LR == RL) { 3186 Op1 = ISD::getSetCCSwappedOperands(Op1); 3187 std::swap(RL, RR); 3188 } 3189 if (LL == RL && LR == RR) { 3190 bool isInteger = LL.getValueType().isInteger(); 3191 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger); 3192 if (Result != ISD::SETCC_INVALID && 3193 (!LegalOperations || 3194 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) && 3195 TLI.isOperationLegal(ISD::SETCC, 3196 getSetCCResultType(N0.getValueType()))))) 3197 return DAG.getSetCC(SDLoc(N), N0.getValueType(), 3198 LL, LR, Result); 3199 } 3200 } 3201 3202 // Simplify: (or (op x...), (op y...)) -> (op (or x, y)) 3203 if (N0.getOpcode() == N1.getOpcode()) { 3204 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N); 3205 if (Tmp.getNode()) return Tmp; 3206 } 3207 3208 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible. 3209 if (N0.getOpcode() == ISD::AND && 3210 N1.getOpcode() == ISD::AND && 3211 N0.getOperand(1).getOpcode() == ISD::Constant && 3212 N1.getOperand(1).getOpcode() == ISD::Constant && 3213 // Don't increase # computations. 3214 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) { 3215 // We can only do this xform if we know that bits from X that are set in C2 3216 // but not in C1 are already zero. Likewise for Y. 3217 const APInt &LHSMask = 3218 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 3219 const APInt &RHSMask = 3220 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue(); 3221 3222 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) && 3223 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) { 3224 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT, 3225 N0.getOperand(0), N1.getOperand(0)); 3226 return DAG.getNode(ISD::AND, SDLoc(N), VT, X, 3227 DAG.getConstant(LHSMask | RHSMask, VT)); 3228 } 3229 } 3230 3231 // See if this is some rotate idiom. 3232 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N))) 3233 return SDValue(Rot, 0); 3234 3235 // Simplify the operands using demanded-bits information. 3236 if (!VT.isVector() && 3237 SimplifyDemandedBits(SDValue(N, 0))) 3238 return SDValue(N, 0); 3239 3240 return SDValue(); 3241} 3242 3243/// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present. 3244static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) { 3245 if (Op.getOpcode() == ISD::AND) { 3246 if (isa<ConstantSDNode>(Op.getOperand(1))) { 3247 Mask = Op.getOperand(1); 3248 Op = Op.getOperand(0); 3249 } else { 3250 return false; 3251 } 3252 } 3253 3254 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) { 3255 Shift = Op; 3256 return true; 3257 } 3258 3259 return false; 3260} 3261 3262// MatchRotate - Handle an 'or' of two operands. If this is one of the many 3263// idioms for rotate, and if the target supports rotation instructions, generate 3264// a rot[lr]. 3265SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) { 3266 // Must be a legal type. Expanded 'n promoted things won't work with rotates. 3267 EVT VT = LHS.getValueType(); 3268 if (!TLI.isTypeLegal(VT)) return 0; 3269 3270 // The target must have at least one rotate flavor. 3271 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT); 3272 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT); 3273 if (!HasROTL && !HasROTR) return 0; 3274 3275 // Match "(X shl/srl V1) & V2" where V2 may not be present. 3276 SDValue LHSShift; // The shift. 3277 SDValue LHSMask; // AND value if any. 3278 if (!MatchRotateHalf(LHS, LHSShift, LHSMask)) 3279 return 0; // Not part of a rotate. 3280 3281 SDValue RHSShift; // The shift. 3282 SDValue RHSMask; // AND value if any. 3283 if (!MatchRotateHalf(RHS, RHSShift, RHSMask)) 3284 return 0; // Not part of a rotate. 3285 3286 if (LHSShift.getOperand(0) != RHSShift.getOperand(0)) 3287 return 0; // Not shifting the same value. 3288 3289 if (LHSShift.getOpcode() == RHSShift.getOpcode()) 3290 return 0; // Shifts must disagree. 3291 3292 // Canonicalize shl to left side in a shl/srl pair. 3293 if (RHSShift.getOpcode() == ISD::SHL) { 3294 std::swap(LHS, RHS); 3295 std::swap(LHSShift, RHSShift); 3296 std::swap(LHSMask , RHSMask ); 3297 } 3298 3299 unsigned OpSizeInBits = VT.getSizeInBits(); 3300 SDValue LHSShiftArg = LHSShift.getOperand(0); 3301 SDValue LHSShiftAmt = LHSShift.getOperand(1); 3302 SDValue RHSShiftAmt = RHSShift.getOperand(1); 3303 3304 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1) 3305 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2) 3306 if (LHSShiftAmt.getOpcode() == ISD::Constant && 3307 RHSShiftAmt.getOpcode() == ISD::Constant) { 3308 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue(); 3309 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue(); 3310 if ((LShVal + RShVal) != OpSizeInBits) 3311 return 0; 3312 3313 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, 3314 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt); 3315 3316 // If there is an AND of either shifted operand, apply it to the result. 3317 if (LHSMask.getNode() || RHSMask.getNode()) { 3318 APInt Mask = APInt::getAllOnesValue(OpSizeInBits); 3319 3320 if (LHSMask.getNode()) { 3321 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal); 3322 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits; 3323 } 3324 if (RHSMask.getNode()) { 3325 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal); 3326 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits; 3327 } 3328 3329 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT)); 3330 } 3331 3332 return Rot.getNode(); 3333 } 3334 3335 // If there is a mask here, and we have a variable shift, we can't be sure 3336 // that we're masking out the right stuff. 3337 if (LHSMask.getNode() || RHSMask.getNode()) 3338 return 0; 3339 3340 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y) 3341 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y)) 3342 if (RHSShiftAmt.getOpcode() == ISD::SUB && 3343 LHSShiftAmt == RHSShiftAmt.getOperand(1)) { 3344 if (ConstantSDNode *SUBC = 3345 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) { 3346 if (SUBC->getAPIntValue() == OpSizeInBits) { 3347 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, LHSShiftArg, 3348 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode(); 3349 } 3350 } 3351 } 3352 3353 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y) 3354 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y)) 3355 if (LHSShiftAmt.getOpcode() == ISD::SUB && 3356 RHSShiftAmt == LHSShiftAmt.getOperand(1)) { 3357 if (ConstantSDNode *SUBC = 3358 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) { 3359 if (SUBC->getAPIntValue() == OpSizeInBits) { 3360 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT, LHSShiftArg, 3361 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode(); 3362 } 3363 } 3364 } 3365 3366 // Look for sign/zext/any-extended or truncate cases: 3367 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND || 3368 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND || 3369 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND || 3370 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) && 3371 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND || 3372 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND || 3373 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND || 3374 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) { 3375 SDValue LExtOp0 = LHSShiftAmt.getOperand(0); 3376 SDValue RExtOp0 = RHSShiftAmt.getOperand(0); 3377 if (RExtOp0.getOpcode() == ISD::SUB && 3378 RExtOp0.getOperand(1) == LExtOp0) { 3379 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) -> 3380 // (rotl x, y) 3381 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) -> 3382 // (rotr x, (sub 32, y)) 3383 if (ConstantSDNode *SUBC = 3384 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) { 3385 if (SUBC->getAPIntValue() == OpSizeInBits) { 3386 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, 3387 LHSShiftArg, 3388 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode(); 3389 } 3390 } 3391 } else if (LExtOp0.getOpcode() == ISD::SUB && 3392 RExtOp0 == LExtOp0.getOperand(1)) { 3393 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) -> 3394 // (rotr x, y) 3395 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) -> 3396 // (rotl x, (sub 32, y)) 3397 if (ConstantSDNode *SUBC = 3398 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) { 3399 if (SUBC->getAPIntValue() == OpSizeInBits) { 3400 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT, 3401 LHSShiftArg, 3402 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode(); 3403 } 3404 } 3405 } 3406 } 3407 3408 return 0; 3409} 3410 3411SDValue DAGCombiner::visitXOR(SDNode *N) { 3412 SDValue N0 = N->getOperand(0); 3413 SDValue N1 = N->getOperand(1); 3414 SDValue LHS, RHS, CC; 3415 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3416 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3417 EVT VT = N0.getValueType(); 3418 3419 // fold vector ops 3420 if (VT.isVector()) { 3421 SDValue FoldedVOp = SimplifyVBinOp(N); 3422 if (FoldedVOp.getNode()) return FoldedVOp; 3423 3424 // fold (xor x, 0) -> x, vector edition 3425 if (ISD::isBuildVectorAllZeros(N0.getNode())) 3426 return N1; 3427 if (ISD::isBuildVectorAllZeros(N1.getNode())) 3428 return N0; 3429 } 3430 3431 // fold (xor undef, undef) -> 0. This is a common idiom (misuse). 3432 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF) 3433 return DAG.getConstant(0, VT); 3434 // fold (xor x, undef) -> undef 3435 if (N0.getOpcode() == ISD::UNDEF) 3436 return N0; 3437 if (N1.getOpcode() == ISD::UNDEF) 3438 return N1; 3439 // fold (xor c1, c2) -> c1^c2 3440 if (N0C && N1C) 3441 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C); 3442 // canonicalize constant to RHS 3443 if (N0C && !N1C) 3444 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0); 3445 // fold (xor x, 0) -> x 3446 if (N1C && N1C->isNullValue()) 3447 return N0; 3448 // reassociate xor 3449 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1); 3450 if (RXOR.getNode() != 0) 3451 return RXOR; 3452 3453 // fold !(x cc y) -> (x !cc y) 3454 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) { 3455 bool isInt = LHS.getValueType().isInteger(); 3456 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), 3457 isInt); 3458 3459 if (!LegalOperations || 3460 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) { 3461 switch (N0.getOpcode()) { 3462 default: 3463 llvm_unreachable("Unhandled SetCC Equivalent!"); 3464 case ISD::SETCC: 3465 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC); 3466 case ISD::SELECT_CC: 3467 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2), 3468 N0.getOperand(3), NotCC); 3469 } 3470 } 3471 } 3472 3473 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y))) 3474 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND && 3475 N0.getNode()->hasOneUse() && 3476 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){ 3477 SDValue V = N0.getOperand(0); 3478 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V, 3479 DAG.getConstant(1, V.getValueType())); 3480 AddToWorkList(V.getNode()); 3481 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V); 3482 } 3483 3484 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc 3485 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 && 3486 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 3487 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); 3488 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) { 3489 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 3490 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS 3491 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS 3492 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode()); 3493 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS); 3494 } 3495 } 3496 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants 3497 if (N1C && N1C->isAllOnesValue() && 3498 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 3499 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); 3500 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) { 3501 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 3502 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS 3503 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS 3504 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode()); 3505 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS); 3506 } 3507 } 3508 // fold (xor (and x, y), y) -> (and (not x), y) 3509 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() && 3510 N0->getOperand(1) == N1) { 3511 SDValue X = N0->getOperand(0); 3512 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT); 3513 AddToWorkList(NotX.getNode()); 3514 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1); 3515 } 3516 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2)) 3517 if (N1C && N0.getOpcode() == ISD::XOR) { 3518 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0)); 3519 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 3520 if (N00C) 3521 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1), 3522 DAG.getConstant(N1C->getAPIntValue() ^ 3523 N00C->getAPIntValue(), VT)); 3524 if (N01C) 3525 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0), 3526 DAG.getConstant(N1C->getAPIntValue() ^ 3527 N01C->getAPIntValue(), VT)); 3528 } 3529 // fold (xor x, x) -> 0 3530 if (N0 == N1) 3531 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations); 3532 3533 // Simplify: xor (op x...), (op y...) -> (op (xor x, y)) 3534 if (N0.getOpcode() == N1.getOpcode()) { 3535 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N); 3536 if (Tmp.getNode()) return Tmp; 3537 } 3538 3539 // Simplify the expression using non-local knowledge. 3540 if (!VT.isVector() && 3541 SimplifyDemandedBits(SDValue(N, 0))) 3542 return SDValue(N, 0); 3543 3544 return SDValue(); 3545} 3546 3547/// visitShiftByConstant - Handle transforms common to the three shifts, when 3548/// the shift amount is a constant. 3549SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) { 3550 SDNode *LHS = N->getOperand(0).getNode(); 3551 if (!LHS->hasOneUse()) return SDValue(); 3552 3553 // We want to pull some binops through shifts, so that we have (and (shift)) 3554 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of 3555 // thing happens with address calculations, so it's important to canonicalize 3556 // it. 3557 bool HighBitSet = false; // Can we transform this if the high bit is set? 3558 3559 switch (LHS->getOpcode()) { 3560 default: return SDValue(); 3561 case ISD::OR: 3562 case ISD::XOR: 3563 HighBitSet = false; // We can only transform sra if the high bit is clear. 3564 break; 3565 case ISD::AND: 3566 HighBitSet = true; // We can only transform sra if the high bit is set. 3567 break; 3568 case ISD::ADD: 3569 if (N->getOpcode() != ISD::SHL) 3570 return SDValue(); // only shl(add) not sr[al](add). 3571 HighBitSet = false; // We can only transform sra if the high bit is clear. 3572 break; 3573 } 3574 3575 // We require the RHS of the binop to be a constant as well. 3576 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1)); 3577 if (!BinOpCst) return SDValue(); 3578 3579 // FIXME: disable this unless the input to the binop is a shift by a constant. 3580 // If it is not a shift, it pessimizes some common cases like: 3581 // 3582 // void foo(int *X, int i) { X[i & 1235] = 1; } 3583 // int bar(int *X, int i) { return X[i & 255]; } 3584 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode(); 3585 if ((BinOpLHSVal->getOpcode() != ISD::SHL && 3586 BinOpLHSVal->getOpcode() != ISD::SRA && 3587 BinOpLHSVal->getOpcode() != ISD::SRL) || 3588 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1))) 3589 return SDValue(); 3590 3591 EVT VT = N->getValueType(0); 3592 3593 // If this is a signed shift right, and the high bit is modified by the 3594 // logical operation, do not perform the transformation. The highBitSet 3595 // boolean indicates the value of the high bit of the constant which would 3596 // cause it to be modified for this operation. 3597 if (N->getOpcode() == ISD::SRA) { 3598 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative(); 3599 if (BinOpRHSSignSet != HighBitSet) 3600 return SDValue(); 3601 } 3602 3603 // Fold the constants, shifting the binop RHS by the shift amount. 3604 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)), 3605 N->getValueType(0), 3606 LHS->getOperand(1), N->getOperand(1)); 3607 3608 // Create the new shift. 3609 SDValue NewShift = DAG.getNode(N->getOpcode(), 3610 SDLoc(LHS->getOperand(0)), 3611 VT, LHS->getOperand(0), N->getOperand(1)); 3612 3613 // Create the new binop. 3614 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS); 3615} 3616 3617SDValue DAGCombiner::visitSHL(SDNode *N) { 3618 SDValue N0 = N->getOperand(0); 3619 SDValue N1 = N->getOperand(1); 3620 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3621 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3622 EVT VT = N0.getValueType(); 3623 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits(); 3624 3625 // fold (shl c1, c2) -> c1<<c2 3626 if (N0C && N1C) 3627 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C); 3628 // fold (shl 0, x) -> 0 3629 if (N0C && N0C->isNullValue()) 3630 return N0; 3631 // fold (shl x, c >= size(x)) -> undef 3632 if (N1C && N1C->getZExtValue() >= OpSizeInBits) 3633 return DAG.getUNDEF(VT); 3634 // fold (shl x, 0) -> x 3635 if (N1C && N1C->isNullValue()) 3636 return N0; 3637 // fold (shl undef, x) -> 0 3638 if (N0.getOpcode() == ISD::UNDEF) 3639 return DAG.getConstant(0, VT); 3640 // if (shl x, c) is known to be zero, return 0 3641 if (DAG.MaskedValueIsZero(SDValue(N, 0), 3642 APInt::getAllOnesValue(OpSizeInBits))) 3643 return DAG.getConstant(0, VT); 3644 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))). 3645 if (N1.getOpcode() == ISD::TRUNCATE && 3646 N1.getOperand(0).getOpcode() == ISD::AND && 3647 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) { 3648 SDValue N101 = N1.getOperand(0).getOperand(1); 3649 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) { 3650 EVT TruncVT = N1.getValueType(); 3651 SDValue N100 = N1.getOperand(0).getOperand(0); 3652 APInt TruncC = N101C->getAPIntValue(); 3653 TruncC = TruncC.trunc(TruncVT.getSizeInBits()); 3654 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, 3655 DAG.getNode(ISD::AND, SDLoc(N), TruncVT, 3656 DAG.getNode(ISD::TRUNCATE, 3657 SDLoc(N), 3658 TruncVT, N100), 3659 DAG.getConstant(TruncC, TruncVT))); 3660 } 3661 } 3662 3663 if (N1C && SimplifyDemandedBits(SDValue(N, 0))) 3664 return SDValue(N, 0); 3665 3666 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2)) 3667 if (N1C && N0.getOpcode() == ISD::SHL && 3668 N0.getOperand(1).getOpcode() == ISD::Constant) { 3669 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 3670 uint64_t c2 = N1C->getZExtValue(); 3671 if (c1 + c2 >= OpSizeInBits) 3672 return DAG.getConstant(0, VT); 3673 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0), 3674 DAG.getConstant(c1 + c2, N1.getValueType())); 3675 } 3676 3677 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2))) 3678 // For this to be valid, the second form must not preserve any of the bits 3679 // that are shifted out by the inner shift in the first form. This means 3680 // the outer shift size must be >= the number of bits added by the ext. 3681 // As a corollary, we don't care what kind of ext it is. 3682 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND || 3683 N0.getOpcode() == ISD::ANY_EXTEND || 3684 N0.getOpcode() == ISD::SIGN_EXTEND) && 3685 N0.getOperand(0).getOpcode() == ISD::SHL && 3686 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) { 3687 uint64_t c1 = 3688 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue(); 3689 uint64_t c2 = N1C->getZExtValue(); 3690 EVT InnerShiftVT = N0.getOperand(0).getValueType(); 3691 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits(); 3692 if (c2 >= OpSizeInBits - InnerShiftSize) { 3693 if (c1 + c2 >= OpSizeInBits) 3694 return DAG.getConstant(0, VT); 3695 return DAG.getNode(ISD::SHL, SDLoc(N0), VT, 3696 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT, 3697 N0.getOperand(0)->getOperand(0)), 3698 DAG.getConstant(c1 + c2, N1.getValueType())); 3699 } 3700 } 3701 3702 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or 3703 // (and (srl x, (sub c1, c2), MASK) 3704 // Only fold this if the inner shift has no other uses -- if it does, folding 3705 // this will increase the total number of instructions. 3706 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() && 3707 N0.getOperand(1).getOpcode() == ISD::Constant) { 3708 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 3709 if (c1 < VT.getSizeInBits()) { 3710 uint64_t c2 = N1C->getZExtValue(); 3711 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), 3712 VT.getSizeInBits() - c1); 3713 SDValue Shift; 3714 if (c2 > c1) { 3715 Mask = Mask.shl(c2-c1); 3716 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0), 3717 DAG.getConstant(c2-c1, N1.getValueType())); 3718 } else { 3719 Mask = Mask.lshr(c1-c2); 3720 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), 3721 DAG.getConstant(c1-c2, N1.getValueType())); 3722 } 3723 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift, 3724 DAG.getConstant(Mask, VT)); 3725 } 3726 } 3727 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1)) 3728 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) { 3729 SDValue HiBitsMask = 3730 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(), 3731 VT.getSizeInBits() - 3732 N1C->getZExtValue()), 3733 VT); 3734 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0), 3735 HiBitsMask); 3736 } 3737 3738 if (N1C) { 3739 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue()); 3740 if (NewSHL.getNode()) 3741 return NewSHL; 3742 } 3743 3744 return SDValue(); 3745} 3746 3747SDValue DAGCombiner::visitSRA(SDNode *N) { 3748 SDValue N0 = N->getOperand(0); 3749 SDValue N1 = N->getOperand(1); 3750 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3751 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3752 EVT VT = N0.getValueType(); 3753 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits(); 3754 3755 // fold (sra c1, c2) -> (sra c1, c2) 3756 if (N0C && N1C) 3757 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C); 3758 // fold (sra 0, x) -> 0 3759 if (N0C && N0C->isNullValue()) 3760 return N0; 3761 // fold (sra -1, x) -> -1 3762 if (N0C && N0C->isAllOnesValue()) 3763 return N0; 3764 // fold (sra x, (setge c, size(x))) -> undef 3765 if (N1C && N1C->getZExtValue() >= OpSizeInBits) 3766 return DAG.getUNDEF(VT); 3767 // fold (sra x, 0) -> x 3768 if (N1C && N1C->isNullValue()) 3769 return N0; 3770 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports 3771 // sext_inreg. 3772 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) { 3773 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue(); 3774 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits); 3775 if (VT.isVector()) 3776 ExtVT = EVT::getVectorVT(*DAG.getContext(), 3777 ExtVT, VT.getVectorNumElements()); 3778 if ((!LegalOperations || 3779 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT))) 3780 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, 3781 N0.getOperand(0), DAG.getValueType(ExtVT)); 3782 } 3783 3784 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2)) 3785 if (N1C && N0.getOpcode() == ISD::SRA) { 3786 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 3787 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue(); 3788 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1; 3789 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0), 3790 DAG.getConstant(Sum, N1C->getValueType(0))); 3791 } 3792 } 3793 3794 // fold (sra (shl X, m), (sub result_size, n)) 3795 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for 3796 // result_size - n != m. 3797 // If truncate is free for the target sext(shl) is likely to result in better 3798 // code. 3799 if (N0.getOpcode() == ISD::SHL) { 3800 // Get the two constanst of the shifts, CN0 = m, CN = n. 3801 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 3802 if (N01C && N1C) { 3803 // Determine what the truncate's result bitsize and type would be. 3804 EVT TruncVT = 3805 EVT::getIntegerVT(*DAG.getContext(), 3806 OpSizeInBits - N1C->getZExtValue()); 3807 // Determine the residual right-shift amount. 3808 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue(); 3809 3810 // If the shift is not a no-op (in which case this should be just a sign 3811 // extend already), the truncated to type is legal, sign_extend is legal 3812 // on that type, and the truncate to that type is both legal and free, 3813 // perform the transform. 3814 if ((ShiftAmt > 0) && 3815 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) && 3816 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) && 3817 TLI.isTruncateFree(VT, TruncVT)) { 3818 3819 SDValue Amt = DAG.getConstant(ShiftAmt, 3820 getShiftAmountTy(N0.getOperand(0).getValueType())); 3821 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT, 3822 N0.getOperand(0), Amt); 3823 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT, 3824 Shift); 3825 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), 3826 N->getValueType(0), Trunc); 3827 } 3828 } 3829 } 3830 3831 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))). 3832 if (N1.getOpcode() == ISD::TRUNCATE && 3833 N1.getOperand(0).getOpcode() == ISD::AND && 3834 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) { 3835 SDValue N101 = N1.getOperand(0).getOperand(1); 3836 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) { 3837 EVT TruncVT = N1.getValueType(); 3838 SDValue N100 = N1.getOperand(0).getOperand(0); 3839 APInt TruncC = N101C->getAPIntValue(); 3840 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits()); 3841 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, 3842 DAG.getNode(ISD::AND, SDLoc(N), 3843 TruncVT, 3844 DAG.getNode(ISD::TRUNCATE, 3845 SDLoc(N), 3846 TruncVT, N100), 3847 DAG.getConstant(TruncC, TruncVT))); 3848 } 3849 } 3850 3851 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2)) 3852 // if c1 is equal to the number of bits the trunc removes 3853 if (N0.getOpcode() == ISD::TRUNCATE && 3854 (N0.getOperand(0).getOpcode() == ISD::SRL || 3855 N0.getOperand(0).getOpcode() == ISD::SRA) && 3856 N0.getOperand(0).hasOneUse() && 3857 N0.getOperand(0).getOperand(1).hasOneUse() && 3858 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) { 3859 EVT LargeVT = N0.getOperand(0).getValueType(); 3860 ConstantSDNode *LargeShiftAmt = 3861 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1)); 3862 3863 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits == 3864 LargeShiftAmt->getZExtValue()) { 3865 SDValue Amt = 3866 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(), 3867 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType())); 3868 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT, 3869 N0.getOperand(0).getOperand(0), Amt); 3870 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA); 3871 } 3872 } 3873 3874 // Simplify, based on bits shifted out of the LHS. 3875 if (N1C && SimplifyDemandedBits(SDValue(N, 0))) 3876 return SDValue(N, 0); 3877 3878 3879 // If the sign bit is known to be zero, switch this to a SRL. 3880 if (DAG.SignBitIsZero(N0)) 3881 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1); 3882 3883 if (N1C) { 3884 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue()); 3885 if (NewSRA.getNode()) 3886 return NewSRA; 3887 } 3888 3889 return SDValue(); 3890} 3891 3892SDValue DAGCombiner::visitSRL(SDNode *N) { 3893 SDValue N0 = N->getOperand(0); 3894 SDValue N1 = N->getOperand(1); 3895 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3896 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3897 EVT VT = N0.getValueType(); 3898 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits(); 3899 3900 // fold (srl c1, c2) -> c1 >>u c2 3901 if (N0C && N1C) 3902 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C); 3903 // fold (srl 0, x) -> 0 3904 if (N0C && N0C->isNullValue()) 3905 return N0; 3906 // fold (srl x, c >= size(x)) -> undef 3907 if (N1C && N1C->getZExtValue() >= OpSizeInBits) 3908 return DAG.getUNDEF(VT); 3909 // fold (srl x, 0) -> x 3910 if (N1C && N1C->isNullValue()) 3911 return N0; 3912 // if (srl x, c) is known to be zero, return 0 3913 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), 3914 APInt::getAllOnesValue(OpSizeInBits))) 3915 return DAG.getConstant(0, VT); 3916 3917 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2)) 3918 if (N1C && N0.getOpcode() == ISD::SRL && 3919 N0.getOperand(1).getOpcode() == ISD::Constant) { 3920 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 3921 uint64_t c2 = N1C->getZExtValue(); 3922 if (c1 + c2 >= OpSizeInBits) 3923 return DAG.getConstant(0, VT); 3924 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), 3925 DAG.getConstant(c1 + c2, N1.getValueType())); 3926 } 3927 3928 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2))) 3929 if (N1C && N0.getOpcode() == ISD::TRUNCATE && 3930 N0.getOperand(0).getOpcode() == ISD::SRL && 3931 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) { 3932 uint64_t c1 = 3933 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue(); 3934 uint64_t c2 = N1C->getZExtValue(); 3935 EVT InnerShiftVT = N0.getOperand(0).getValueType(); 3936 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType(); 3937 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits(); 3938 // This is only valid if the OpSizeInBits + c1 = size of inner shift. 3939 if (c1 + OpSizeInBits == InnerShiftSize) { 3940 if (c1 + c2 >= InnerShiftSize) 3941 return DAG.getConstant(0, VT); 3942 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, 3943 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT, 3944 N0.getOperand(0)->getOperand(0), 3945 DAG.getConstant(c1 + c2, ShiftCountVT))); 3946 } 3947 } 3948 3949 // fold (srl (shl x, c), c) -> (and x, cst2) 3950 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 && 3951 N0.getValueSizeInBits() <= 64) { 3952 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits(); 3953 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0), 3954 DAG.getConstant(~0ULL >> ShAmt, VT)); 3955 } 3956 3957 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask) 3958 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { 3959 // Shifting in all undef bits? 3960 EVT SmallVT = N0.getOperand(0).getValueType(); 3961 if (N1C->getZExtValue() >= SmallVT.getSizeInBits()) 3962 return DAG.getUNDEF(VT); 3963 3964 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) { 3965 uint64_t ShiftAmt = N1C->getZExtValue(); 3966 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT, 3967 N0.getOperand(0), 3968 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT))); 3969 AddToWorkList(SmallShift.getNode()); 3970 APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits()).lshr(ShiftAmt); 3971 return DAG.getNode(ISD::AND, SDLoc(N), VT, 3972 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift), 3973 DAG.getConstant(Mask, VT)); 3974 } 3975 } 3976 3977 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign 3978 // bit, which is unmodified by sra. 3979 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) { 3980 if (N0.getOpcode() == ISD::SRA) 3981 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1); 3982 } 3983 3984 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit). 3985 if (N1C && N0.getOpcode() == ISD::CTLZ && 3986 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) { 3987 APInt KnownZero, KnownOne; 3988 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne); 3989 3990 // If any of the input bits are KnownOne, then the input couldn't be all 3991 // zeros, thus the result of the srl will always be zero. 3992 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT); 3993 3994 // If all of the bits input the to ctlz node are known to be zero, then 3995 // the result of the ctlz is "32" and the result of the shift is one. 3996 APInt UnknownBits = ~KnownZero; 3997 if (UnknownBits == 0) return DAG.getConstant(1, VT); 3998 3999 // Otherwise, check to see if there is exactly one bit input to the ctlz. 4000 if ((UnknownBits & (UnknownBits - 1)) == 0) { 4001 // Okay, we know that only that the single bit specified by UnknownBits 4002 // could be set on input to the CTLZ node. If this bit is set, the SRL 4003 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair 4004 // to an SRL/XOR pair, which is likely to simplify more. 4005 unsigned ShAmt = UnknownBits.countTrailingZeros(); 4006 SDValue Op = N0.getOperand(0); 4007 4008 if (ShAmt) { 4009 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op, 4010 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType()))); 4011 AddToWorkList(Op.getNode()); 4012 } 4013 4014 return DAG.getNode(ISD::XOR, SDLoc(N), VT, 4015 Op, DAG.getConstant(1, VT)); 4016 } 4017 } 4018 4019 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))). 4020 if (N1.getOpcode() == ISD::TRUNCATE && 4021 N1.getOperand(0).getOpcode() == ISD::AND && 4022 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) { 4023 SDValue N101 = N1.getOperand(0).getOperand(1); 4024 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) { 4025 EVT TruncVT = N1.getValueType(); 4026 SDValue N100 = N1.getOperand(0).getOperand(0); 4027 APInt TruncC = N101C->getAPIntValue(); 4028 TruncC = TruncC.trunc(TruncVT.getSizeInBits()); 4029 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, 4030 DAG.getNode(ISD::AND, SDLoc(N), 4031 TruncVT, 4032 DAG.getNode(ISD::TRUNCATE, 4033 SDLoc(N), 4034 TruncVT, N100), 4035 DAG.getConstant(TruncC, TruncVT))); 4036 } 4037 } 4038 4039 // fold operands of srl based on knowledge that the low bits are not 4040 // demanded. 4041 if (N1C && SimplifyDemandedBits(SDValue(N, 0))) 4042 return SDValue(N, 0); 4043 4044 if (N1C) { 4045 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue()); 4046 if (NewSRL.getNode()) 4047 return NewSRL; 4048 } 4049 4050 // Attempt to convert a srl of a load into a narrower zero-extending load. 4051 SDValue NarrowLoad = ReduceLoadWidth(N); 4052 if (NarrowLoad.getNode()) 4053 return NarrowLoad; 4054 4055 // Here is a common situation. We want to optimize: 4056 // 4057 // %a = ... 4058 // %b = and i32 %a, 2 4059 // %c = srl i32 %b, 1 4060 // brcond i32 %c ... 4061 // 4062 // into 4063 // 4064 // %a = ... 4065 // %b = and %a, 2 4066 // %c = setcc eq %b, 0 4067 // brcond %c ... 4068 // 4069 // However when after the source operand of SRL is optimized into AND, the SRL 4070 // itself may not be optimized further. Look for it and add the BRCOND into 4071 // the worklist. 4072 if (N->hasOneUse()) { 4073 SDNode *Use = *N->use_begin(); 4074 if (Use->getOpcode() == ISD::BRCOND) 4075 AddToWorkList(Use); 4076 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) { 4077 // Also look pass the truncate. 4078 Use = *Use->use_begin(); 4079 if (Use->getOpcode() == ISD::BRCOND) 4080 AddToWorkList(Use); 4081 } 4082 } 4083 4084 return SDValue(); 4085} 4086 4087SDValue DAGCombiner::visitCTLZ(SDNode *N) { 4088 SDValue N0 = N->getOperand(0); 4089 EVT VT = N->getValueType(0); 4090 4091 // fold (ctlz c1) -> c2 4092 if (isa<ConstantSDNode>(N0)) 4093 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0); 4094 return SDValue(); 4095} 4096 4097SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) { 4098 SDValue N0 = N->getOperand(0); 4099 EVT VT = N->getValueType(0); 4100 4101 // fold (ctlz_zero_undef c1) -> c2 4102 if (isa<ConstantSDNode>(N0)) 4103 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0); 4104 return SDValue(); 4105} 4106 4107SDValue DAGCombiner::visitCTTZ(SDNode *N) { 4108 SDValue N0 = N->getOperand(0); 4109 EVT VT = N->getValueType(0); 4110 4111 // fold (cttz c1) -> c2 4112 if (isa<ConstantSDNode>(N0)) 4113 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0); 4114 return SDValue(); 4115} 4116 4117SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) { 4118 SDValue N0 = N->getOperand(0); 4119 EVT VT = N->getValueType(0); 4120 4121 // fold (cttz_zero_undef c1) -> c2 4122 if (isa<ConstantSDNode>(N0)) 4123 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0); 4124 return SDValue(); 4125} 4126 4127SDValue DAGCombiner::visitCTPOP(SDNode *N) { 4128 SDValue N0 = N->getOperand(0); 4129 EVT VT = N->getValueType(0); 4130 4131 // fold (ctpop c1) -> c2 4132 if (isa<ConstantSDNode>(N0)) 4133 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0); 4134 return SDValue(); 4135} 4136 4137SDValue DAGCombiner::visitSELECT(SDNode *N) { 4138 SDValue N0 = N->getOperand(0); 4139 SDValue N1 = N->getOperand(1); 4140 SDValue N2 = N->getOperand(2); 4141 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 4142 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 4143 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); 4144 EVT VT = N->getValueType(0); 4145 EVT VT0 = N0.getValueType(); 4146 4147 // fold (select C, X, X) -> X 4148 if (N1 == N2) 4149 return N1; 4150 // fold (select true, X, Y) -> X 4151 if (N0C && !N0C->isNullValue()) 4152 return N1; 4153 // fold (select false, X, Y) -> Y 4154 if (N0C && N0C->isNullValue()) 4155 return N2; 4156 // fold (select C, 1, X) -> (or C, X) 4157 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1) 4158 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2); 4159 // fold (select C, 0, 1) -> (xor C, 1) 4160 if (VT.isInteger() && 4161 (VT0 == MVT::i1 || 4162 (VT0.isInteger() && 4163 TLI.getBooleanContents(false) == 4164 TargetLowering::ZeroOrOneBooleanContent)) && 4165 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) { 4166 SDValue XORNode; 4167 if (VT == VT0) 4168 return DAG.getNode(ISD::XOR, SDLoc(N), VT0, 4169 N0, DAG.getConstant(1, VT0)); 4170 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0, 4171 N0, DAG.getConstant(1, VT0)); 4172 AddToWorkList(XORNode.getNode()); 4173 if (VT.bitsGT(VT0)) 4174 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode); 4175 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode); 4176 } 4177 // fold (select C, 0, X) -> (and (not C), X) 4178 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) { 4179 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT); 4180 AddToWorkList(NOTNode.getNode()); 4181 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2); 4182 } 4183 // fold (select C, X, 1) -> (or (not C), X) 4184 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) { 4185 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT); 4186 AddToWorkList(NOTNode.getNode()); 4187 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1); 4188 } 4189 // fold (select C, X, 0) -> (and C, X) 4190 if (VT == MVT::i1 && N2C && N2C->isNullValue()) 4191 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1); 4192 // fold (select X, X, Y) -> (or X, Y) 4193 // fold (select X, 1, Y) -> (or X, Y) 4194 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1))) 4195 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2); 4196 // fold (select X, Y, X) -> (and X, Y) 4197 // fold (select X, Y, 0) -> (and X, Y) 4198 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0))) 4199 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1); 4200 4201 // If we can fold this based on the true/false value, do so. 4202 if (SimplifySelectOps(N, N1, N2)) 4203 return SDValue(N, 0); // Don't revisit N. 4204 4205 // fold selects based on a setcc into other things, such as min/max/abs 4206 if (N0.getOpcode() == ISD::SETCC) { 4207 // FIXME: 4208 // Check against MVT::Other for SELECT_CC, which is a workaround for targets 4209 // having to say they don't support SELECT_CC on every type the DAG knows 4210 // about, since there is no way to mark an opcode illegal at all value types 4211 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) && 4212 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) 4213 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, 4214 N0.getOperand(0), N0.getOperand(1), 4215 N1, N2, N0.getOperand(2)); 4216 return SimplifySelect(SDLoc(N), N0, N1, N2); 4217 } 4218 4219 return SDValue(); 4220} 4221 4222SDValue DAGCombiner::visitVSELECT(SDNode *N) { 4223 SDValue N0 = N->getOperand(0); 4224 SDValue N1 = N->getOperand(1); 4225 SDValue N2 = N->getOperand(2); 4226 SDLoc DL(N); 4227 4228 // Canonicalize integer abs. 4229 // vselect (setg[te] X, 0), X, -X -> 4230 // vselect (setgt X, -1), X, -X -> 4231 // vselect (setl[te] X, 0), -X, X -> 4232 // Y = sra (X, size(X)-1); xor (add (X, Y), Y) 4233 if (N0.getOpcode() == ISD::SETCC) { 4234 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); 4235 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 4236 bool isAbs = false; 4237 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode()); 4238 4239 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) || 4240 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) && 4241 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1)) 4242 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode()); 4243 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) && 4244 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1)) 4245 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode()); 4246 4247 if (isAbs) { 4248 EVT VT = LHS.getValueType(); 4249 SDValue Shift = DAG.getNode( 4250 ISD::SRA, DL, VT, LHS, 4251 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT)); 4252 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift); 4253 AddToWorkList(Shift.getNode()); 4254 AddToWorkList(Add.getNode()); 4255 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift); 4256 } 4257 } 4258 4259 return SDValue(); 4260} 4261 4262SDValue DAGCombiner::visitSELECT_CC(SDNode *N) { 4263 SDValue N0 = N->getOperand(0); 4264 SDValue N1 = N->getOperand(1); 4265 SDValue N2 = N->getOperand(2); 4266 SDValue N3 = N->getOperand(3); 4267 SDValue N4 = N->getOperand(4); 4268 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get(); 4269 4270 // fold select_cc lhs, rhs, x, x, cc -> x 4271 if (N2 == N3) 4272 return N2; 4273 4274 // Determine if the condition we're dealing with is constant 4275 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()), 4276 N0, N1, CC, SDLoc(N), false); 4277 if (SCC.getNode()) { 4278 AddToWorkList(SCC.getNode()); 4279 4280 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) { 4281 if (!SCCC->isNullValue()) 4282 return N2; // cond always true -> true val 4283 else 4284 return N3; // cond always false -> false val 4285 } 4286 4287 // Fold to a simpler select_cc 4288 if (SCC.getOpcode() == ISD::SETCC) 4289 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(), 4290 SCC.getOperand(0), SCC.getOperand(1), N2, N3, 4291 SCC.getOperand(2)); 4292 } 4293 4294 // If we can fold this based on the true/false value, do so. 4295 if (SimplifySelectOps(N, N2, N3)) 4296 return SDValue(N, 0); // Don't revisit N. 4297 4298 // fold select_cc into other things, such as min/max/abs 4299 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC); 4300} 4301 4302SDValue DAGCombiner::visitSETCC(SDNode *N) { 4303 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1), 4304 cast<CondCodeSDNode>(N->getOperand(2))->get(), 4305 SDLoc(N)); 4306} 4307 4308// ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this: 4309// "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))" 4310// transformation. Returns true if extension are possible and the above 4311// mentioned transformation is profitable. 4312static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0, 4313 unsigned ExtOpc, 4314 SmallVector<SDNode*, 4> &ExtendNodes, 4315 const TargetLowering &TLI) { 4316 bool HasCopyToRegUses = false; 4317 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType()); 4318 for (SDNode::use_iterator UI = N0.getNode()->use_begin(), 4319 UE = N0.getNode()->use_end(); 4320 UI != UE; ++UI) { 4321 SDNode *User = *UI; 4322 if (User == N) 4323 continue; 4324 if (UI.getUse().getResNo() != N0.getResNo()) 4325 continue; 4326 // FIXME: Only extend SETCC N, N and SETCC N, c for now. 4327 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) { 4328 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get(); 4329 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC)) 4330 // Sign bits will be lost after a zext. 4331 return false; 4332 bool Add = false; 4333 for (unsigned i = 0; i != 2; ++i) { 4334 SDValue UseOp = User->getOperand(i); 4335 if (UseOp == N0) 4336 continue; 4337 if (!isa<ConstantSDNode>(UseOp)) 4338 return false; 4339 Add = true; 4340 } 4341 if (Add) 4342 ExtendNodes.push_back(User); 4343 continue; 4344 } 4345 // If truncates aren't free and there are users we can't 4346 // extend, it isn't worthwhile. 4347 if (!isTruncFree) 4348 return false; 4349 // Remember if this value is live-out. 4350 if (User->getOpcode() == ISD::CopyToReg) 4351 HasCopyToRegUses = true; 4352 } 4353 4354 if (HasCopyToRegUses) { 4355 bool BothLiveOut = false; 4356 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 4357 UI != UE; ++UI) { 4358 SDUse &Use = UI.getUse(); 4359 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) { 4360 BothLiveOut = true; 4361 break; 4362 } 4363 } 4364 if (BothLiveOut) 4365 // Both unextended and extended values are live out. There had better be 4366 // a good reason for the transformation. 4367 return ExtendNodes.size(); 4368 } 4369 return true; 4370} 4371 4372void DAGCombiner::ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs, 4373 SDValue Trunc, SDValue ExtLoad, SDLoc DL, 4374 ISD::NodeType ExtType) { 4375 // Extend SetCC uses if necessary. 4376 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) { 4377 SDNode *SetCC = SetCCs[i]; 4378 SmallVector<SDValue, 4> Ops; 4379 4380 for (unsigned j = 0; j != 2; ++j) { 4381 SDValue SOp = SetCC->getOperand(j); 4382 if (SOp == Trunc) 4383 Ops.push_back(ExtLoad); 4384 else 4385 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp)); 4386 } 4387 4388 Ops.push_back(SetCC->getOperand(2)); 4389 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), 4390 &Ops[0], Ops.size())); 4391 } 4392} 4393 4394SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) { 4395 SDValue N0 = N->getOperand(0); 4396 EVT VT = N->getValueType(0); 4397 4398 // fold (sext c1) -> c1 4399 if (isa<ConstantSDNode>(N0)) 4400 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N0); 4401 4402 // fold (sext (sext x)) -> (sext x) 4403 // fold (sext (aext x)) -> (sext x) 4404 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) 4405 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, 4406 N0.getOperand(0)); 4407 4408 if (N0.getOpcode() == ISD::TRUNCATE) { 4409 // fold (sext (truncate (load x))) -> (sext (smaller load x)) 4410 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n))) 4411 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode()); 4412 if (NarrowLoad.getNode()) { 4413 SDNode* oye = N0.getNode()->getOperand(0).getNode(); 4414 if (NarrowLoad.getNode() != N0.getNode()) { 4415 CombineTo(N0.getNode(), NarrowLoad); 4416 // CombineTo deleted the truncate, if needed, but not what's under it. 4417 AddToWorkList(oye); 4418 } 4419 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4420 } 4421 4422 // See if the value being truncated is already sign extended. If so, just 4423 // eliminate the trunc/sext pair. 4424 SDValue Op = N0.getOperand(0); 4425 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits(); 4426 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits(); 4427 unsigned DestBits = VT.getScalarType().getSizeInBits(); 4428 unsigned NumSignBits = DAG.ComputeNumSignBits(Op); 4429 4430 if (OpBits == DestBits) { 4431 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign 4432 // bits, it is already ready. 4433 if (NumSignBits > DestBits-MidBits) 4434 return Op; 4435 } else if (OpBits < DestBits) { 4436 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign 4437 // bits, just sext from i32. 4438 if (NumSignBits > OpBits-MidBits) 4439 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op); 4440 } else { 4441 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign 4442 // bits, just truncate to i32. 4443 if (NumSignBits > OpBits-MidBits) 4444 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op); 4445 } 4446 4447 // fold (sext (truncate x)) -> (sextinreg x). 4448 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, 4449 N0.getValueType())) { 4450 if (OpBits < DestBits) 4451 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op); 4452 else if (OpBits > DestBits) 4453 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op); 4454 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op, 4455 DAG.getValueType(N0.getValueType())); 4456 } 4457 } 4458 4459 // fold (sext (load x)) -> (sext (truncate (sextload x))) 4460 // None of the supported targets knows how to perform load and sign extend 4461 // on vectors in one instruction. We only perform this transformation on 4462 // scalars. 4463 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() && 4464 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 4465 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) { 4466 bool DoXform = true; 4467 SmallVector<SDNode*, 4> SetCCs; 4468 if (!N0.hasOneUse()) 4469 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI); 4470 if (DoXform) { 4471 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 4472 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, 4473 LN0->getChain(), 4474 LN0->getBasePtr(), LN0->getPointerInfo(), 4475 N0.getValueType(), 4476 LN0->isVolatile(), LN0->isNonTemporal(), 4477 LN0->getAlignment()); 4478 CombineTo(N, ExtLoad); 4479 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 4480 N0.getValueType(), ExtLoad); 4481 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1)); 4482 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 4483 ISD::SIGN_EXTEND); 4484 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4485 } 4486 } 4487 4488 // fold (sext (sextload x)) -> (sext (truncate (sextload x))) 4489 // fold (sext ( extload x)) -> (sext (truncate (sextload x))) 4490 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) && 4491 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) { 4492 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 4493 EVT MemVT = LN0->getMemoryVT(); 4494 if ((!LegalOperations && !LN0->isVolatile()) || 4495 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) { 4496 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, 4497 LN0->getChain(), 4498 LN0->getBasePtr(), LN0->getPointerInfo(), 4499 MemVT, 4500 LN0->isVolatile(), LN0->isNonTemporal(), 4501 LN0->getAlignment()); 4502 CombineTo(N, ExtLoad); 4503 CombineTo(N0.getNode(), 4504 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 4505 N0.getValueType(), ExtLoad), 4506 ExtLoad.getValue(1)); 4507 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4508 } 4509 } 4510 4511 // fold (sext (and/or/xor (load x), cst)) -> 4512 // (and/or/xor (sextload x), (sext cst)) 4513 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR || 4514 N0.getOpcode() == ISD::XOR) && 4515 isa<LoadSDNode>(N0.getOperand(0)) && 4516 N0.getOperand(1).getOpcode() == ISD::Constant && 4517 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) && 4518 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) { 4519 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0)); 4520 if (LN0->getExtensionType() != ISD::ZEXTLOAD) { 4521 bool DoXform = true; 4522 SmallVector<SDNode*, 4> SetCCs; 4523 if (!N0.hasOneUse()) 4524 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND, 4525 SetCCs, TLI); 4526 if (DoXform) { 4527 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT, 4528 LN0->getChain(), LN0->getBasePtr(), 4529 LN0->getPointerInfo(), 4530 LN0->getMemoryVT(), 4531 LN0->isVolatile(), 4532 LN0->isNonTemporal(), 4533 LN0->getAlignment()); 4534 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 4535 Mask = Mask.sext(VT.getSizeInBits()); 4536 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT, 4537 ExtLoad, DAG.getConstant(Mask, VT)); 4538 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, 4539 SDLoc(N0.getOperand(0)), 4540 N0.getOperand(0).getValueType(), ExtLoad); 4541 CombineTo(N, And); 4542 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1)); 4543 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 4544 ISD::SIGN_EXTEND); 4545 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4546 } 4547 } 4548 } 4549 4550 if (N0.getOpcode() == ISD::SETCC) { 4551 // sext(setcc) -> sext_in_reg(vsetcc) for vectors. 4552 // Only do this before legalize for now. 4553 if (VT.isVector() && !LegalOperations && 4554 TLI.getBooleanContents(true) == 4555 TargetLowering::ZeroOrNegativeOneBooleanContent) { 4556 EVT N0VT = N0.getOperand(0).getValueType(); 4557 // On some architectures (such as SSE/NEON/etc) the SETCC result type is 4558 // of the same size as the compared operands. Only optimize sext(setcc()) 4559 // if this is the case. 4560 EVT SVT = getSetCCResultType(N0VT); 4561 4562 // We know that the # elements of the results is the same as the 4563 // # elements of the compare (and the # elements of the compare result 4564 // for that matter). Check to see that they are the same size. If so, 4565 // we know that the element size of the sext'd result matches the 4566 // element size of the compare operands. 4567 if (VT.getSizeInBits() == SVT.getSizeInBits()) 4568 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0), 4569 N0.getOperand(1), 4570 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 4571 4572 // If the desired elements are smaller or larger than the source 4573 // elements we can use a matching integer vector type and then 4574 // truncate/sign extend 4575 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger(); 4576 if (SVT == MatchingVectorType) { 4577 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType, 4578 N0.getOperand(0), N0.getOperand(1), 4579 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 4580 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT); 4581 } 4582 } 4583 4584 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc) 4585 unsigned ElementWidth = VT.getScalarType().getSizeInBits(); 4586 SDValue NegOne = 4587 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT); 4588 SDValue SCC = 4589 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1), 4590 NegOne, DAG.getConstant(0, VT), 4591 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true); 4592 if (SCC.getNode()) return SCC; 4593 if (!VT.isVector() && 4594 (!LegalOperations || 4595 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(VT)))) { 4596 return DAG.getSelect(SDLoc(N), VT, 4597 DAG.getSetCC(SDLoc(N), 4598 getSetCCResultType(VT), 4599 N0.getOperand(0), N0.getOperand(1), 4600 cast<CondCodeSDNode>(N0.getOperand(2))->get()), 4601 NegOne, DAG.getConstant(0, VT)); 4602 } 4603 } 4604 4605 // fold (sext x) -> (zext x) if the sign bit is known zero. 4606 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) && 4607 DAG.SignBitIsZero(N0)) 4608 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0); 4609 4610 return SDValue(); 4611} 4612 4613// isTruncateOf - If N is a truncate of some other value, return true, record 4614// the value being truncated in Op and which of Op's bits are zero in KnownZero. 4615// This function computes KnownZero to avoid a duplicated call to 4616// ComputeMaskedBits in the caller. 4617static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op, 4618 APInt &KnownZero) { 4619 APInt KnownOne; 4620 if (N->getOpcode() == ISD::TRUNCATE) { 4621 Op = N->getOperand(0); 4622 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne); 4623 return true; 4624 } 4625 4626 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 || 4627 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE) 4628 return false; 4629 4630 SDValue Op0 = N->getOperand(0); 4631 SDValue Op1 = N->getOperand(1); 4632 assert(Op0.getValueType() == Op1.getValueType()); 4633 4634 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0); 4635 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1); 4636 if (COp0 && COp0->isNullValue()) 4637 Op = Op1; 4638 else if (COp1 && COp1->isNullValue()) 4639 Op = Op0; 4640 else 4641 return false; 4642 4643 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne); 4644 4645 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue()) 4646 return false; 4647 4648 return true; 4649} 4650 4651SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) { 4652 SDValue N0 = N->getOperand(0); 4653 EVT VT = N->getValueType(0); 4654 4655 // fold (zext c1) -> c1 4656 if (isa<ConstantSDNode>(N0)) 4657 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0); 4658 // fold (zext (zext x)) -> (zext x) 4659 // fold (zext (aext x)) -> (zext x) 4660 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) 4661 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, 4662 N0.getOperand(0)); 4663 4664 // fold (zext (truncate x)) -> (zext x) or 4665 // (zext (truncate x)) -> (truncate x) 4666 // This is valid when the truncated bits of x are already zero. 4667 // FIXME: We should extend this to work for vectors too. 4668 SDValue Op; 4669 APInt KnownZero; 4670 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) { 4671 APInt TruncatedBits = 4672 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ? 4673 APInt(Op.getValueSizeInBits(), 0) : 4674 APInt::getBitsSet(Op.getValueSizeInBits(), 4675 N0.getValueSizeInBits(), 4676 std::min(Op.getValueSizeInBits(), 4677 VT.getSizeInBits())); 4678 if (TruncatedBits == (KnownZero & TruncatedBits)) { 4679 if (VT.bitsGT(Op.getValueType())) 4680 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op); 4681 if (VT.bitsLT(Op.getValueType())) 4682 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op); 4683 4684 return Op; 4685 } 4686 } 4687 4688 // fold (zext (truncate (load x))) -> (zext (smaller load x)) 4689 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n))) 4690 if (N0.getOpcode() == ISD::TRUNCATE) { 4691 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode()); 4692 if (NarrowLoad.getNode()) { 4693 SDNode* oye = N0.getNode()->getOperand(0).getNode(); 4694 if (NarrowLoad.getNode() != N0.getNode()) { 4695 CombineTo(N0.getNode(), NarrowLoad); 4696 // CombineTo deleted the truncate, if needed, but not what's under it. 4697 AddToWorkList(oye); 4698 } 4699 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4700 } 4701 } 4702 4703 // fold (zext (truncate x)) -> (and x, mask) 4704 if (N0.getOpcode() == ISD::TRUNCATE && 4705 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) { 4706 4707 // fold (zext (truncate (load x))) -> (zext (smaller load x)) 4708 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n))) 4709 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode()); 4710 if (NarrowLoad.getNode()) { 4711 SDNode* oye = N0.getNode()->getOperand(0).getNode(); 4712 if (NarrowLoad.getNode() != N0.getNode()) { 4713 CombineTo(N0.getNode(), NarrowLoad); 4714 // CombineTo deleted the truncate, if needed, but not what's under it. 4715 AddToWorkList(oye); 4716 } 4717 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4718 } 4719 4720 SDValue Op = N0.getOperand(0); 4721 if (Op.getValueType().bitsLT(VT)) { 4722 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op); 4723 AddToWorkList(Op.getNode()); 4724 } else if (Op.getValueType().bitsGT(VT)) { 4725 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op); 4726 AddToWorkList(Op.getNode()); 4727 } 4728 return DAG.getZeroExtendInReg(Op, SDLoc(N), 4729 N0.getValueType().getScalarType()); 4730 } 4731 4732 // Fold (zext (and (trunc x), cst)) -> (and x, cst), 4733 // if either of the casts is not free. 4734 if (N0.getOpcode() == ISD::AND && 4735 N0.getOperand(0).getOpcode() == ISD::TRUNCATE && 4736 N0.getOperand(1).getOpcode() == ISD::Constant && 4737 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(), 4738 N0.getValueType()) || 4739 !TLI.isZExtFree(N0.getValueType(), VT))) { 4740 SDValue X = N0.getOperand(0).getOperand(0); 4741 if (X.getValueType().bitsLT(VT)) { 4742 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X); 4743 } else if (X.getValueType().bitsGT(VT)) { 4744 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X); 4745 } 4746 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 4747 Mask = Mask.zext(VT.getSizeInBits()); 4748 return DAG.getNode(ISD::AND, SDLoc(N), VT, 4749 X, DAG.getConstant(Mask, VT)); 4750 } 4751 4752 // fold (zext (load x)) -> (zext (truncate (zextload x))) 4753 // None of the supported targets knows how to perform load and vector_zext 4754 // on vectors in one instruction. We only perform this transformation on 4755 // scalars. 4756 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() && 4757 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 4758 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) { 4759 bool DoXform = true; 4760 SmallVector<SDNode*, 4> SetCCs; 4761 if (!N0.hasOneUse()) 4762 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI); 4763 if (DoXform) { 4764 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 4765 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT, 4766 LN0->getChain(), 4767 LN0->getBasePtr(), LN0->getPointerInfo(), 4768 N0.getValueType(), 4769 LN0->isVolatile(), LN0->isNonTemporal(), 4770 LN0->getAlignment()); 4771 CombineTo(N, ExtLoad); 4772 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 4773 N0.getValueType(), ExtLoad); 4774 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1)); 4775 4776 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 4777 ISD::ZERO_EXTEND); 4778 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4779 } 4780 } 4781 4782 // fold (zext (and/or/xor (load x), cst)) -> 4783 // (and/or/xor (zextload x), (zext cst)) 4784 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR || 4785 N0.getOpcode() == ISD::XOR) && 4786 isa<LoadSDNode>(N0.getOperand(0)) && 4787 N0.getOperand(1).getOpcode() == ISD::Constant && 4788 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) && 4789 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) { 4790 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0)); 4791 if (LN0->getExtensionType() != ISD::SEXTLOAD) { 4792 bool DoXform = true; 4793 SmallVector<SDNode*, 4> SetCCs; 4794 if (!N0.hasOneUse()) 4795 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND, 4796 SetCCs, TLI); 4797 if (DoXform) { 4798 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT, 4799 LN0->getChain(), LN0->getBasePtr(), 4800 LN0->getPointerInfo(), 4801 LN0->getMemoryVT(), 4802 LN0->isVolatile(), 4803 LN0->isNonTemporal(), 4804 LN0->getAlignment()); 4805 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 4806 Mask = Mask.zext(VT.getSizeInBits()); 4807 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT, 4808 ExtLoad, DAG.getConstant(Mask, VT)); 4809 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, 4810 SDLoc(N0.getOperand(0)), 4811 N0.getOperand(0).getValueType(), ExtLoad); 4812 CombineTo(N, And); 4813 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1)); 4814 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 4815 ISD::ZERO_EXTEND); 4816 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4817 } 4818 } 4819 } 4820 4821 // fold (zext (zextload x)) -> (zext (truncate (zextload x))) 4822 // fold (zext ( extload x)) -> (zext (truncate (zextload x))) 4823 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) && 4824 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) { 4825 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 4826 EVT MemVT = LN0->getMemoryVT(); 4827 if ((!LegalOperations && !LN0->isVolatile()) || 4828 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) { 4829 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT, 4830 LN0->getChain(), 4831 LN0->getBasePtr(), LN0->getPointerInfo(), 4832 MemVT, 4833 LN0->isVolatile(), LN0->isNonTemporal(), 4834 LN0->getAlignment()); 4835 CombineTo(N, ExtLoad); 4836 CombineTo(N0.getNode(), 4837 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), 4838 ExtLoad), 4839 ExtLoad.getValue(1)); 4840 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4841 } 4842 } 4843 4844 if (N0.getOpcode() == ISD::SETCC) { 4845 if (!LegalOperations && VT.isVector()) { 4846 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors. 4847 // Only do this before legalize for now. 4848 EVT N0VT = N0.getOperand(0).getValueType(); 4849 EVT EltVT = VT.getVectorElementType(); 4850 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(), 4851 DAG.getConstant(1, EltVT)); 4852 if (VT.getSizeInBits() == N0VT.getSizeInBits()) 4853 // We know that the # elements of the results is the same as the 4854 // # elements of the compare (and the # elements of the compare result 4855 // for that matter). Check to see that they are the same size. If so, 4856 // we know that the element size of the sext'd result matches the 4857 // element size of the compare operands. 4858 return DAG.getNode(ISD::AND, SDLoc(N), VT, 4859 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0), 4860 N0.getOperand(1), 4861 cast<CondCodeSDNode>(N0.getOperand(2))->get()), 4862 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, 4863 &OneOps[0], OneOps.size())); 4864 4865 // If the desired elements are smaller or larger than the source 4866 // elements we can use a matching integer vector type and then 4867 // truncate/sign extend 4868 EVT MatchingElementType = 4869 EVT::getIntegerVT(*DAG.getContext(), 4870 N0VT.getScalarType().getSizeInBits()); 4871 EVT MatchingVectorType = 4872 EVT::getVectorVT(*DAG.getContext(), MatchingElementType, 4873 N0VT.getVectorNumElements()); 4874 SDValue VsetCC = 4875 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0), 4876 N0.getOperand(1), 4877 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 4878 return DAG.getNode(ISD::AND, SDLoc(N), VT, 4879 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT), 4880 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, 4881 &OneOps[0], OneOps.size())); 4882 } 4883 4884 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc 4885 SDValue SCC = 4886 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1), 4887 DAG.getConstant(1, VT), DAG.getConstant(0, VT), 4888 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true); 4889 if (SCC.getNode()) return SCC; 4890 } 4891 4892 // (zext (shl (zext x), cst)) -> (shl (zext x), cst) 4893 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) && 4894 isa<ConstantSDNode>(N0.getOperand(1)) && 4895 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND && 4896 N0.hasOneUse()) { 4897 SDValue ShAmt = N0.getOperand(1); 4898 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue(); 4899 if (N0.getOpcode() == ISD::SHL) { 4900 SDValue InnerZExt = N0.getOperand(0); 4901 // If the original shl may be shifting out bits, do not perform this 4902 // transformation. 4903 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() - 4904 InnerZExt.getOperand(0).getValueType().getSizeInBits(); 4905 if (ShAmtVal > KnownZeroBits) 4906 return SDValue(); 4907 } 4908 4909 SDLoc DL(N); 4910 4911 // Ensure that the shift amount is wide enough for the shifted value. 4912 if (VT.getSizeInBits() >= 256) 4913 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt); 4914 4915 return DAG.getNode(N0.getOpcode(), DL, VT, 4916 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)), 4917 ShAmt); 4918 } 4919 4920 return SDValue(); 4921} 4922 4923SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) { 4924 SDValue N0 = N->getOperand(0); 4925 EVT VT = N->getValueType(0); 4926 4927 // fold (aext c1) -> c1 4928 if (isa<ConstantSDNode>(N0)) 4929 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, N0); 4930 // fold (aext (aext x)) -> (aext x) 4931 // fold (aext (zext x)) -> (zext x) 4932 // fold (aext (sext x)) -> (sext x) 4933 if (N0.getOpcode() == ISD::ANY_EXTEND || 4934 N0.getOpcode() == ISD::ZERO_EXTEND || 4935 N0.getOpcode() == ISD::SIGN_EXTEND) 4936 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0)); 4937 4938 // fold (aext (truncate (load x))) -> (aext (smaller load x)) 4939 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n))) 4940 if (N0.getOpcode() == ISD::TRUNCATE) { 4941 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode()); 4942 if (NarrowLoad.getNode()) { 4943 SDNode* oye = N0.getNode()->getOperand(0).getNode(); 4944 if (NarrowLoad.getNode() != N0.getNode()) { 4945 CombineTo(N0.getNode(), NarrowLoad); 4946 // CombineTo deleted the truncate, if needed, but not what's under it. 4947 AddToWorkList(oye); 4948 } 4949 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4950 } 4951 } 4952 4953 // fold (aext (truncate x)) 4954 if (N0.getOpcode() == ISD::TRUNCATE) { 4955 SDValue TruncOp = N0.getOperand(0); 4956 if (TruncOp.getValueType() == VT) 4957 return TruncOp; // x iff x size == zext size. 4958 if (TruncOp.getValueType().bitsGT(VT)) 4959 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp); 4960 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp); 4961 } 4962 4963 // Fold (aext (and (trunc x), cst)) -> (and x, cst) 4964 // if the trunc is not free. 4965 if (N0.getOpcode() == ISD::AND && 4966 N0.getOperand(0).getOpcode() == ISD::TRUNCATE && 4967 N0.getOperand(1).getOpcode() == ISD::Constant && 4968 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(), 4969 N0.getValueType())) { 4970 SDValue X = N0.getOperand(0).getOperand(0); 4971 if (X.getValueType().bitsLT(VT)) { 4972 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X); 4973 } else if (X.getValueType().bitsGT(VT)) { 4974 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X); 4975 } 4976 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 4977 Mask = Mask.zext(VT.getSizeInBits()); 4978 return DAG.getNode(ISD::AND, SDLoc(N), VT, 4979 X, DAG.getConstant(Mask, VT)); 4980 } 4981 4982 // fold (aext (load x)) -> (aext (truncate (extload x))) 4983 // None of the supported targets knows how to perform load and any_ext 4984 // on vectors in one instruction. We only perform this transformation on 4985 // scalars. 4986 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() && 4987 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 4988 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) { 4989 bool DoXform = true; 4990 SmallVector<SDNode*, 4> SetCCs; 4991 if (!N0.hasOneUse()) 4992 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI); 4993 if (DoXform) { 4994 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 4995 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT, 4996 LN0->getChain(), 4997 LN0->getBasePtr(), LN0->getPointerInfo(), 4998 N0.getValueType(), 4999 LN0->isVolatile(), LN0->isNonTemporal(), 5000 LN0->getAlignment()); 5001 CombineTo(N, ExtLoad); 5002 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 5003 N0.getValueType(), ExtLoad); 5004 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1)); 5005 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 5006 ISD::ANY_EXTEND); 5007 return SDValue(N, 0); // Return N so it doesn't get rechecked! 5008 } 5009 } 5010 5011 // fold (aext (zextload x)) -> (aext (truncate (zextload x))) 5012 // fold (aext (sextload x)) -> (aext (truncate (sextload x))) 5013 // fold (aext ( extload x)) -> (aext (truncate (extload x))) 5014 if (N0.getOpcode() == ISD::LOAD && 5015 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && 5016 N0.hasOneUse()) { 5017 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5018 EVT MemVT = LN0->getMemoryVT(); 5019 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(N), 5020 VT, LN0->getChain(), LN0->getBasePtr(), 5021 LN0->getPointerInfo(), MemVT, 5022 LN0->isVolatile(), LN0->isNonTemporal(), 5023 LN0->getAlignment()); 5024 CombineTo(N, ExtLoad); 5025 CombineTo(N0.getNode(), 5026 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 5027 N0.getValueType(), ExtLoad), 5028 ExtLoad.getValue(1)); 5029 return SDValue(N, 0); // Return N so it doesn't get rechecked! 5030 } 5031 5032 if (N0.getOpcode() == ISD::SETCC) { 5033 // aext(setcc) -> sext_in_reg(vsetcc) for vectors. 5034 // Only do this before legalize for now. 5035 if (VT.isVector() && !LegalOperations) { 5036 EVT N0VT = N0.getOperand(0).getValueType(); 5037 // We know that the # elements of the results is the same as the 5038 // # elements of the compare (and the # elements of the compare result 5039 // for that matter). Check to see that they are the same size. If so, 5040 // we know that the element size of the sext'd result matches the 5041 // element size of the compare operands. 5042 if (VT.getSizeInBits() == N0VT.getSizeInBits()) 5043 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0), 5044 N0.getOperand(1), 5045 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 5046 // If the desired elements are smaller or larger than the source 5047 // elements we can use a matching integer vector type and then 5048 // truncate/sign extend 5049 else { 5050 EVT MatchingElementType = 5051 EVT::getIntegerVT(*DAG.getContext(), 5052 N0VT.getScalarType().getSizeInBits()); 5053 EVT MatchingVectorType = 5054 EVT::getVectorVT(*DAG.getContext(), MatchingElementType, 5055 N0VT.getVectorNumElements()); 5056 SDValue VsetCC = 5057 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0), 5058 N0.getOperand(1), 5059 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 5060 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT); 5061 } 5062 } 5063 5064 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc 5065 SDValue SCC = 5066 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1), 5067 DAG.getConstant(1, VT), DAG.getConstant(0, VT), 5068 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true); 5069 if (SCC.getNode()) 5070 return SCC; 5071 } 5072 5073 return SDValue(); 5074} 5075 5076/// GetDemandedBits - See if the specified operand can be simplified with the 5077/// knowledge that only the bits specified by Mask are used. If so, return the 5078/// simpler operand, otherwise return a null SDValue. 5079SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) { 5080 switch (V.getOpcode()) { 5081 default: break; 5082 case ISD::Constant: { 5083 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode()); 5084 assert(CV != 0 && "Const value should be ConstSDNode."); 5085 const APInt &CVal = CV->getAPIntValue(); 5086 APInt NewVal = CVal & Mask; 5087 if (NewVal != CVal) { 5088 return DAG.getConstant(NewVal, V.getValueType()); 5089 } 5090 break; 5091 } 5092 case ISD::OR: 5093 case ISD::XOR: 5094 // If the LHS or RHS don't contribute bits to the or, drop them. 5095 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask)) 5096 return V.getOperand(1); 5097 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask)) 5098 return V.getOperand(0); 5099 break; 5100 case ISD::SRL: 5101 // Only look at single-use SRLs. 5102 if (!V.getNode()->hasOneUse()) 5103 break; 5104 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) { 5105 // See if we can recursively simplify the LHS. 5106 unsigned Amt = RHSC->getZExtValue(); 5107 5108 // Watch out for shift count overflow though. 5109 if (Amt >= Mask.getBitWidth()) break; 5110 APInt NewMask = Mask << Amt; 5111 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask); 5112 if (SimplifyLHS.getNode()) 5113 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(), 5114 SimplifyLHS, V.getOperand(1)); 5115 } 5116 } 5117 return SDValue(); 5118} 5119 5120/// ReduceLoadWidth - If the result of a wider load is shifted to right of N 5121/// bits and then truncated to a narrower type and where N is a multiple 5122/// of number of bits of the narrower type, transform it to a narrower load 5123/// from address + N / num of bits of new type. If the result is to be 5124/// extended, also fold the extension to form a extending load. 5125SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) { 5126 unsigned Opc = N->getOpcode(); 5127 5128 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD; 5129 SDValue N0 = N->getOperand(0); 5130 EVT VT = N->getValueType(0); 5131 EVT ExtVT = VT; 5132 5133 // This transformation isn't valid for vector loads. 5134 if (VT.isVector()) 5135 return SDValue(); 5136 5137 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then 5138 // extended to VT. 5139 if (Opc == ISD::SIGN_EXTEND_INREG) { 5140 ExtType = ISD::SEXTLOAD; 5141 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT(); 5142 } else if (Opc == ISD::SRL) { 5143 // Another special-case: SRL is basically zero-extending a narrower value. 5144 ExtType = ISD::ZEXTLOAD; 5145 N0 = SDValue(N, 0); 5146 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 5147 if (!N01) return SDValue(); 5148 ExtVT = EVT::getIntegerVT(*DAG.getContext(), 5149 VT.getSizeInBits() - N01->getZExtValue()); 5150 } 5151 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT)) 5152 return SDValue(); 5153 5154 unsigned EVTBits = ExtVT.getSizeInBits(); 5155 5156 // Do not generate loads of non-round integer types since these can 5157 // be expensive (and would be wrong if the type is not byte sized). 5158 if (!ExtVT.isRound()) 5159 return SDValue(); 5160 5161 unsigned ShAmt = 0; 5162 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) { 5163 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 5164 ShAmt = N01->getZExtValue(); 5165 // Is the shift amount a multiple of size of VT? 5166 if ((ShAmt & (EVTBits-1)) == 0) { 5167 N0 = N0.getOperand(0); 5168 // Is the load width a multiple of size of VT? 5169 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0) 5170 return SDValue(); 5171 } 5172 5173 // At this point, we must have a load or else we can't do the transform. 5174 if (!isa<LoadSDNode>(N0)) return SDValue(); 5175 5176 // Because a SRL must be assumed to *need* to zero-extend the high bits 5177 // (as opposed to anyext the high bits), we can't combine the zextload 5178 // lowering of SRL and an sextload. 5179 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD) 5180 return SDValue(); 5181 5182 // If the shift amount is larger than the input type then we're not 5183 // accessing any of the loaded bytes. If the load was a zextload/extload 5184 // then the result of the shift+trunc is zero/undef (handled elsewhere). 5185 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits()) 5186 return SDValue(); 5187 } 5188 } 5189 5190 // If the load is shifted left (and the result isn't shifted back right), 5191 // we can fold the truncate through the shift. 5192 unsigned ShLeftAmt = 0; 5193 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() && 5194 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) { 5195 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 5196 ShLeftAmt = N01->getZExtValue(); 5197 N0 = N0.getOperand(0); 5198 } 5199 } 5200 5201 // If we haven't found a load, we can't narrow it. Don't transform one with 5202 // multiple uses, this would require adding a new load. 5203 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse()) 5204 return SDValue(); 5205 5206 // Don't change the width of a volatile load. 5207 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5208 if (LN0->isVolatile()) 5209 return SDValue(); 5210 5211 // Verify that we are actually reducing a load width here. 5212 if (LN0->getMemoryVT().getSizeInBits() < EVTBits) 5213 return SDValue(); 5214 5215 // For the transform to be legal, the load must produce only two values 5216 // (the value loaded and the chain). Don't transform a pre-increment 5217 // load, for example, which produces an extra value. Otherwise the 5218 // transformation is not equivalent, and the downstream logic to replace 5219 // uses gets things wrong. 5220 if (LN0->getNumValues() > 2) 5221 return SDValue(); 5222 5223 EVT PtrType = N0.getOperand(1).getValueType(); 5224 5225 if (PtrType == MVT::Untyped || PtrType.isExtended()) 5226 // It's not possible to generate a constant of extended or untyped type. 5227 return SDValue(); 5228 5229 // For big endian targets, we need to adjust the offset to the pointer to 5230 // load the correct bytes. 5231 if (TLI.isBigEndian()) { 5232 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits(); 5233 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits(); 5234 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt; 5235 } 5236 5237 uint64_t PtrOff = ShAmt / 8; 5238 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff); 5239 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), 5240 PtrType, LN0->getBasePtr(), 5241 DAG.getConstant(PtrOff, PtrType)); 5242 AddToWorkList(NewPtr.getNode()); 5243 5244 SDValue Load; 5245 if (ExtType == ISD::NON_EXTLOAD) 5246 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr, 5247 LN0->getPointerInfo().getWithOffset(PtrOff), 5248 LN0->isVolatile(), LN0->isNonTemporal(), 5249 LN0->isInvariant(), NewAlign); 5250 else 5251 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr, 5252 LN0->getPointerInfo().getWithOffset(PtrOff), 5253 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(), 5254 NewAlign); 5255 5256 // Replace the old load's chain with the new load's chain. 5257 WorkListRemover DeadNodes(*this); 5258 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1)); 5259 5260 // Shift the result left, if we've swallowed a left shift. 5261 SDValue Result = Load; 5262 if (ShLeftAmt != 0) { 5263 EVT ShImmTy = getShiftAmountTy(Result.getValueType()); 5264 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt)) 5265 ShImmTy = VT; 5266 // If the shift amount is as large as the result size (but, presumably, 5267 // no larger than the source) then the useful bits of the result are 5268 // zero; we can't simply return the shortened shift, because the result 5269 // of that operation is undefined. 5270 if (ShLeftAmt >= VT.getSizeInBits()) 5271 Result = DAG.getConstant(0, VT); 5272 else 5273 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT, 5274 Result, DAG.getConstant(ShLeftAmt, ShImmTy)); 5275 } 5276 5277 // Return the new loaded value. 5278 return Result; 5279} 5280 5281SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) { 5282 SDValue N0 = N->getOperand(0); 5283 SDValue N1 = N->getOperand(1); 5284 EVT VT = N->getValueType(0); 5285 EVT EVT = cast<VTSDNode>(N1)->getVT(); 5286 unsigned VTBits = VT.getScalarType().getSizeInBits(); 5287 unsigned EVTBits = EVT.getScalarType().getSizeInBits(); 5288 5289 // fold (sext_in_reg c1) -> c1 5290 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF) 5291 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1); 5292 5293 // If the input is already sign extended, just drop the extension. 5294 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1) 5295 return N0; 5296 5297 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2 5298 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 5299 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) { 5300 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, 5301 N0.getOperand(0), N1); 5302 } 5303 5304 // fold (sext_in_reg (sext x)) -> (sext x) 5305 // fold (sext_in_reg (aext x)) -> (sext x) 5306 // if x is small enough. 5307 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) { 5308 SDValue N00 = N0.getOperand(0); 5309 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits && 5310 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT))) 5311 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1); 5312 } 5313 5314 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero. 5315 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits))) 5316 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT); 5317 5318 // fold operands of sext_in_reg based on knowledge that the top bits are not 5319 // demanded. 5320 if (SimplifyDemandedBits(SDValue(N, 0))) 5321 return SDValue(N, 0); 5322 5323 // fold (sext_in_reg (load x)) -> (smaller sextload x) 5324 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits)) 5325 SDValue NarrowLoad = ReduceLoadWidth(N); 5326 if (NarrowLoad.getNode()) 5327 return NarrowLoad; 5328 5329 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24) 5330 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible. 5331 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above. 5332 if (N0.getOpcode() == ISD::SRL) { 5333 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 5334 if (ShAmt->getZExtValue()+EVTBits <= VTBits) { 5335 // We can turn this into an SRA iff the input to the SRL is already sign 5336 // extended enough. 5337 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0)); 5338 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits) 5339 return DAG.getNode(ISD::SRA, SDLoc(N), VT, 5340 N0.getOperand(0), N0.getOperand(1)); 5341 } 5342 } 5343 5344 // fold (sext_inreg (extload x)) -> (sextload x) 5345 if (ISD::isEXTLoad(N0.getNode()) && 5346 ISD::isUNINDEXEDLoad(N0.getNode()) && 5347 EVT == cast<LoadSDNode>(N0)->getMemoryVT() && 5348 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 5349 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) { 5350 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5351 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, 5352 LN0->getChain(), 5353 LN0->getBasePtr(), LN0->getPointerInfo(), 5354 EVT, 5355 LN0->isVolatile(), LN0->isNonTemporal(), 5356 LN0->getAlignment()); 5357 CombineTo(N, ExtLoad); 5358 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); 5359 AddToWorkList(ExtLoad.getNode()); 5360 return SDValue(N, 0); // Return N so it doesn't get rechecked! 5361 } 5362 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use 5363 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && 5364 N0.hasOneUse() && 5365 EVT == cast<LoadSDNode>(N0)->getMemoryVT() && 5366 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 5367 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) { 5368 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5369 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, 5370 LN0->getChain(), 5371 LN0->getBasePtr(), LN0->getPointerInfo(), 5372 EVT, 5373 LN0->isVolatile(), LN0->isNonTemporal(), 5374 LN0->getAlignment()); 5375 CombineTo(N, ExtLoad); 5376 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); 5377 return SDValue(N, 0); // Return N so it doesn't get rechecked! 5378 } 5379 5380 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16)) 5381 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) { 5382 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), 5383 N0.getOperand(1), false); 5384 if (BSwap.getNode() != 0) 5385 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, 5386 BSwap, N1); 5387 } 5388 5389 return SDValue(); 5390} 5391 5392SDValue DAGCombiner::visitTRUNCATE(SDNode *N) { 5393 SDValue N0 = N->getOperand(0); 5394 EVT VT = N->getValueType(0); 5395 bool isLE = TLI.isLittleEndian(); 5396 5397 // noop truncate 5398 if (N0.getValueType() == N->getValueType(0)) 5399 return N0; 5400 // fold (truncate c1) -> c1 5401 if (isa<ConstantSDNode>(N0)) 5402 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0); 5403 // fold (truncate (truncate x)) -> (truncate x) 5404 if (N0.getOpcode() == ISD::TRUNCATE) 5405 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0)); 5406 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x 5407 if (N0.getOpcode() == ISD::ZERO_EXTEND || 5408 N0.getOpcode() == ISD::SIGN_EXTEND || 5409 N0.getOpcode() == ISD::ANY_EXTEND) { 5410 if (N0.getOperand(0).getValueType().bitsLT(VT)) 5411 // if the source is smaller than the dest, we still need an extend 5412 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, 5413 N0.getOperand(0)); 5414 if (N0.getOperand(0).getValueType().bitsGT(VT)) 5415 // if the source is larger than the dest, than we just need the truncate 5416 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0)); 5417 // if the source and dest are the same type, we can drop both the extend 5418 // and the truncate. 5419 return N0.getOperand(0); 5420 } 5421 5422 // Fold extract-and-trunc into a narrow extract. For example: 5423 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1) 5424 // i32 y = TRUNCATE(i64 x) 5425 // -- becomes -- 5426 // v16i8 b = BITCAST (v2i64 val) 5427 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8) 5428 // 5429 // Note: We only run this optimization after type legalization (which often 5430 // creates this pattern) and before operation legalization after which 5431 // we need to be more careful about the vector instructions that we generate. 5432 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && 5433 LegalTypes && !LegalOperations && N0->hasOneUse()) { 5434 5435 EVT VecTy = N0.getOperand(0).getValueType(); 5436 EVT ExTy = N0.getValueType(); 5437 EVT TrTy = N->getValueType(0); 5438 5439 unsigned NumElem = VecTy.getVectorNumElements(); 5440 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits(); 5441 5442 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem); 5443 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size"); 5444 5445 SDValue EltNo = N0->getOperand(1); 5446 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) { 5447 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); 5448 EVT IndexTy = N0->getOperand(1).getValueType(); 5449 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1)); 5450 5451 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N), 5452 NVT, N0.getOperand(0)); 5453 5454 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, 5455 SDLoc(N), TrTy, V, 5456 DAG.getConstant(Index, IndexTy)); 5457 } 5458 } 5459 5460 // Fold a series of buildvector, bitcast, and truncate if possible. 5461 // For example fold 5462 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to 5463 // (2xi32 (buildvector x, y)). 5464 if (Level == AfterLegalizeVectorOps && VT.isVector() && 5465 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() && 5466 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR && 5467 N0.getOperand(0).hasOneUse()) { 5468 5469 SDValue BuildVect = N0.getOperand(0); 5470 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType(); 5471 EVT TruncVecEltTy = VT.getVectorElementType(); 5472 5473 // Check that the element types match. 5474 if (BuildVectEltTy == TruncVecEltTy) { 5475 // Now we only need to compute the offset of the truncated elements. 5476 unsigned BuildVecNumElts = BuildVect.getNumOperands(); 5477 unsigned TruncVecNumElts = VT.getVectorNumElements(); 5478 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts; 5479 5480 assert((BuildVecNumElts % TruncVecNumElts) == 0 && 5481 "Invalid number of elements"); 5482 5483 SmallVector<SDValue, 8> Opnds; 5484 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset) 5485 Opnds.push_back(BuildVect.getOperand(i)); 5486 5487 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Opnds[0], 5488 Opnds.size()); 5489 } 5490 } 5491 5492 // See if we can simplify the input to this truncate through knowledge that 5493 // only the low bits are being used. 5494 // For example "trunc (or (shl x, 8), y)" // -> trunc y 5495 // Currently we only perform this optimization on scalars because vectors 5496 // may have different active low bits. 5497 if (!VT.isVector()) { 5498 SDValue Shorter = 5499 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(), 5500 VT.getSizeInBits())); 5501 if (Shorter.getNode()) 5502 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter); 5503 } 5504 // fold (truncate (load x)) -> (smaller load x) 5505 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits)) 5506 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) { 5507 SDValue Reduced = ReduceLoadWidth(N); 5508 if (Reduced.getNode()) 5509 return Reduced; 5510 } 5511 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)), 5512 // where ... are all 'undef'. 5513 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) { 5514 SmallVector<EVT, 8> VTs; 5515 SDValue V; 5516 unsigned Idx = 0; 5517 unsigned NumDefs = 0; 5518 5519 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) { 5520 SDValue X = N0.getOperand(i); 5521 if (X.getOpcode() != ISD::UNDEF) { 5522 V = X; 5523 Idx = i; 5524 NumDefs++; 5525 } 5526 // Stop if more than one members are non-undef. 5527 if (NumDefs > 1) 5528 break; 5529 VTs.push_back(EVT::getVectorVT(*DAG.getContext(), 5530 VT.getVectorElementType(), 5531 X.getValueType().getVectorNumElements())); 5532 } 5533 5534 if (NumDefs == 0) 5535 return DAG.getUNDEF(VT); 5536 5537 if (NumDefs == 1) { 5538 assert(V.getNode() && "The single defined operand is empty!"); 5539 SmallVector<SDValue, 8> Opnds; 5540 for (unsigned i = 0, e = VTs.size(); i != e; ++i) { 5541 if (i != Idx) { 5542 Opnds.push_back(DAG.getUNDEF(VTs[i])); 5543 continue; 5544 } 5545 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V); 5546 AddToWorkList(NV.getNode()); 5547 Opnds.push_back(NV); 5548 } 5549 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, 5550 &Opnds[0], Opnds.size()); 5551 } 5552 } 5553 5554 // Simplify the operands using demanded-bits information. 5555 if (!VT.isVector() && 5556 SimplifyDemandedBits(SDValue(N, 0))) 5557 return SDValue(N, 0); 5558 5559 return SDValue(); 5560} 5561 5562static SDNode *getBuildPairElt(SDNode *N, unsigned i) { 5563 SDValue Elt = N->getOperand(i); 5564 if (Elt.getOpcode() != ISD::MERGE_VALUES) 5565 return Elt.getNode(); 5566 return Elt.getOperand(Elt.getResNo()).getNode(); 5567} 5568 5569/// CombineConsecutiveLoads - build_pair (load, load) -> load 5570/// if load locations are consecutive. 5571SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) { 5572 assert(N->getOpcode() == ISD::BUILD_PAIR); 5573 5574 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0)); 5575 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1)); 5576 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() || 5577 LD1->getPointerInfo().getAddrSpace() != 5578 LD2->getPointerInfo().getAddrSpace()) 5579 return SDValue(); 5580 EVT LD1VT = LD1->getValueType(0); 5581 5582 if (ISD::isNON_EXTLoad(LD2) && 5583 LD2->hasOneUse() && 5584 // If both are volatile this would reduce the number of volatile loads. 5585 // If one is volatile it might be ok, but play conservative and bail out. 5586 !LD1->isVolatile() && 5587 !LD2->isVolatile() && 5588 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) { 5589 unsigned Align = LD1->getAlignment(); 5590 unsigned NewAlign = TLI.getDataLayout()-> 5591 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext())); 5592 5593 if (NewAlign <= Align && 5594 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) 5595 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(), 5596 LD1->getBasePtr(), LD1->getPointerInfo(), 5597 false, false, false, Align); 5598 } 5599 5600 return SDValue(); 5601} 5602 5603SDValue DAGCombiner::visitBITCAST(SDNode *N) { 5604 SDValue N0 = N->getOperand(0); 5605 EVT VT = N->getValueType(0); 5606 5607 // If the input is a BUILD_VECTOR with all constant elements, fold this now. 5608 // Only do this before legalize, since afterward the target may be depending 5609 // on the bitconvert. 5610 // First check to see if this is all constant. 5611 if (!LegalTypes && 5612 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() && 5613 VT.isVector()) { 5614 bool isSimple = true; 5615 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) 5616 if (N0.getOperand(i).getOpcode() != ISD::UNDEF && 5617 N0.getOperand(i).getOpcode() != ISD::Constant && 5618 N0.getOperand(i).getOpcode() != ISD::ConstantFP) { 5619 isSimple = false; 5620 break; 5621 } 5622 5623 EVT DestEltVT = N->getValueType(0).getVectorElementType(); 5624 assert(!DestEltVT.isVector() && 5625 "Element type of vector ValueType must not be vector!"); 5626 if (isSimple) 5627 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT); 5628 } 5629 5630 // If the input is a constant, let getNode fold it. 5631 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) { 5632 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0); 5633 if (Res.getNode() != N) { 5634 if (!LegalOperations || 5635 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT)) 5636 return Res; 5637 5638 // Folding it resulted in an illegal node, and it's too late to 5639 // do that. Clean up the old node and forego the transformation. 5640 // Ideally this won't happen very often, because instcombine 5641 // and the earlier dagcombine runs (where illegal nodes are 5642 // permitted) should have folded most of them already. 5643 DAG.DeleteNode(Res.getNode()); 5644 } 5645 } 5646 5647 // (conv (conv x, t1), t2) -> (conv x, t2) 5648 if (N0.getOpcode() == ISD::BITCAST) 5649 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, 5650 N0.getOperand(0)); 5651 5652 // fold (conv (load x)) -> (load (conv*)x) 5653 // If the resultant load doesn't need a higher alignment than the original! 5654 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && 5655 // Do not change the width of a volatile load. 5656 !cast<LoadSDNode>(N0)->isVolatile() && 5657 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) { 5658 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5659 unsigned Align = TLI.getDataLayout()-> 5660 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext())); 5661 unsigned OrigAlign = LN0->getAlignment(); 5662 5663 if (Align <= OrigAlign) { 5664 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), 5665 LN0->getBasePtr(), LN0->getPointerInfo(), 5666 LN0->isVolatile(), LN0->isNonTemporal(), 5667 LN0->isInvariant(), OrigAlign); 5668 AddToWorkList(N); 5669 CombineTo(N0.getNode(), 5670 DAG.getNode(ISD::BITCAST, SDLoc(N0), 5671 N0.getValueType(), Load), 5672 Load.getValue(1)); 5673 return Load; 5674 } 5675 } 5676 5677 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit) 5678 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit)) 5679 // This often reduces constant pool loads. 5680 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(VT)) || 5681 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(VT))) && 5682 N0.getNode()->hasOneUse() && VT.isInteger() && 5683 !VT.isVector() && !N0.getValueType().isVector()) { 5684 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT, 5685 N0.getOperand(0)); 5686 AddToWorkList(NewConv.getNode()); 5687 5688 APInt SignBit = APInt::getSignBit(VT.getSizeInBits()); 5689 if (N0.getOpcode() == ISD::FNEG) 5690 return DAG.getNode(ISD::XOR, SDLoc(N), VT, 5691 NewConv, DAG.getConstant(SignBit, VT)); 5692 assert(N0.getOpcode() == ISD::FABS); 5693 return DAG.getNode(ISD::AND, SDLoc(N), VT, 5694 NewConv, DAG.getConstant(~SignBit, VT)); 5695 } 5696 5697 // fold (bitconvert (fcopysign cst, x)) -> 5698 // (or (and (bitconvert x), sign), (and cst, (not sign))) 5699 // Note that we don't handle (copysign x, cst) because this can always be 5700 // folded to an fneg or fabs. 5701 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() && 5702 isa<ConstantFPSDNode>(N0.getOperand(0)) && 5703 VT.isInteger() && !VT.isVector()) { 5704 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits(); 5705 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth); 5706 if (isTypeLegal(IntXVT)) { 5707 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0), 5708 IntXVT, N0.getOperand(1)); 5709 AddToWorkList(X.getNode()); 5710 5711 // If X has a different width than the result/lhs, sext it or truncate it. 5712 unsigned VTWidth = VT.getSizeInBits(); 5713 if (OrigXWidth < VTWidth) { 5714 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X); 5715 AddToWorkList(X.getNode()); 5716 } else if (OrigXWidth > VTWidth) { 5717 // To get the sign bit in the right place, we have to shift it right 5718 // before truncating. 5719 X = DAG.getNode(ISD::SRL, SDLoc(X), 5720 X.getValueType(), X, 5721 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType())); 5722 AddToWorkList(X.getNode()); 5723 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X); 5724 AddToWorkList(X.getNode()); 5725 } 5726 5727 APInt SignBit = APInt::getSignBit(VT.getSizeInBits()); 5728 X = DAG.getNode(ISD::AND, SDLoc(X), VT, 5729 X, DAG.getConstant(SignBit, VT)); 5730 AddToWorkList(X.getNode()); 5731 5732 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0), 5733 VT, N0.getOperand(0)); 5734 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT, 5735 Cst, DAG.getConstant(~SignBit, VT)); 5736 AddToWorkList(Cst.getNode()); 5737 5738 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst); 5739 } 5740 } 5741 5742 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive. 5743 if (N0.getOpcode() == ISD::BUILD_PAIR) { 5744 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT); 5745 if (CombineLD.getNode()) 5746 return CombineLD; 5747 } 5748 5749 return SDValue(); 5750} 5751 5752SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) { 5753 EVT VT = N->getValueType(0); 5754 return CombineConsecutiveLoads(N, VT); 5755} 5756 5757/// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector 5758/// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the 5759/// destination element value type. 5760SDValue DAGCombiner:: 5761ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) { 5762 EVT SrcEltVT = BV->getValueType(0).getVectorElementType(); 5763 5764 // If this is already the right type, we're done. 5765 if (SrcEltVT == DstEltVT) return SDValue(BV, 0); 5766 5767 unsigned SrcBitSize = SrcEltVT.getSizeInBits(); 5768 unsigned DstBitSize = DstEltVT.getSizeInBits(); 5769 5770 // If this is a conversion of N elements of one type to N elements of another 5771 // type, convert each element. This handles FP<->INT cases. 5772 if (SrcBitSize == DstBitSize) { 5773 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, 5774 BV->getValueType(0).getVectorNumElements()); 5775 5776 // Due to the FP element handling below calling this routine recursively, 5777 // we can end up with a scalar-to-vector node here. 5778 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR) 5779 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT, 5780 DAG.getNode(ISD::BITCAST, SDLoc(BV), 5781 DstEltVT, BV->getOperand(0))); 5782 5783 SmallVector<SDValue, 8> Ops; 5784 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) { 5785 SDValue Op = BV->getOperand(i); 5786 // If the vector element type is not legal, the BUILD_VECTOR operands 5787 // are promoted and implicitly truncated. Make that explicit here. 5788 if (Op.getValueType() != SrcEltVT) 5789 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op); 5790 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV), 5791 DstEltVT, Op)); 5792 AddToWorkList(Ops.back().getNode()); 5793 } 5794 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, 5795 &Ops[0], Ops.size()); 5796 } 5797 5798 // Otherwise, we're growing or shrinking the elements. To avoid having to 5799 // handle annoying details of growing/shrinking FP values, we convert them to 5800 // int first. 5801 if (SrcEltVT.isFloatingPoint()) { 5802 // Convert the input float vector to a int vector where the elements are the 5803 // same sizes. 5804 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!"); 5805 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits()); 5806 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode(); 5807 SrcEltVT = IntVT; 5808 } 5809 5810 // Now we know the input is an integer vector. If the output is a FP type, 5811 // convert to integer first, then to FP of the right size. 5812 if (DstEltVT.isFloatingPoint()) { 5813 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!"); 5814 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits()); 5815 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode(); 5816 5817 // Next, convert to FP elements of the same size. 5818 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT); 5819 } 5820 5821 // Okay, we know the src/dst types are both integers of differing types. 5822 // Handling growing first. 5823 assert(SrcEltVT.isInteger() && DstEltVT.isInteger()); 5824 if (SrcBitSize < DstBitSize) { 5825 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize; 5826 5827 SmallVector<SDValue, 8> Ops; 5828 for (unsigned i = 0, e = BV->getNumOperands(); i != e; 5829 i += NumInputsPerOutput) { 5830 bool isLE = TLI.isLittleEndian(); 5831 APInt NewBits = APInt(DstBitSize, 0); 5832 bool EltIsUndef = true; 5833 for (unsigned j = 0; j != NumInputsPerOutput; ++j) { 5834 // Shift the previously computed bits over. 5835 NewBits <<= SrcBitSize; 5836 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j)); 5837 if (Op.getOpcode() == ISD::UNDEF) continue; 5838 EltIsUndef = false; 5839 5840 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue(). 5841 zextOrTrunc(SrcBitSize).zext(DstBitSize); 5842 } 5843 5844 if (EltIsUndef) 5845 Ops.push_back(DAG.getUNDEF(DstEltVT)); 5846 else 5847 Ops.push_back(DAG.getConstant(NewBits, DstEltVT)); 5848 } 5849 5850 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size()); 5851 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, 5852 &Ops[0], Ops.size()); 5853 } 5854 5855 // Finally, this must be the case where we are shrinking elements: each input 5856 // turns into multiple outputs. 5857 bool isS2V = ISD::isScalarToVector(BV); 5858 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize; 5859 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, 5860 NumOutputsPerInput*BV->getNumOperands()); 5861 SmallVector<SDValue, 8> Ops; 5862 5863 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) { 5864 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) { 5865 for (unsigned j = 0; j != NumOutputsPerInput; ++j) 5866 Ops.push_back(DAG.getUNDEF(DstEltVT)); 5867 continue; 5868 } 5869 5870 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))-> 5871 getAPIntValue().zextOrTrunc(SrcBitSize); 5872 5873 for (unsigned j = 0; j != NumOutputsPerInput; ++j) { 5874 APInt ThisVal = OpVal.trunc(DstBitSize); 5875 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT)); 5876 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal) 5877 // Simply turn this into a SCALAR_TO_VECTOR of the new type. 5878 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT, 5879 Ops[0]); 5880 OpVal = OpVal.lshr(DstBitSize); 5881 } 5882 5883 // For big endian targets, swap the order of the pieces of each element. 5884 if (TLI.isBigEndian()) 5885 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end()); 5886 } 5887 5888 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, 5889 &Ops[0], Ops.size()); 5890} 5891 5892SDValue DAGCombiner::visitFADD(SDNode *N) { 5893 SDValue N0 = N->getOperand(0); 5894 SDValue N1 = N->getOperand(1); 5895 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 5896 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 5897 EVT VT = N->getValueType(0); 5898 5899 // fold vector ops 5900 if (VT.isVector()) { 5901 SDValue FoldedVOp = SimplifyVBinOp(N); 5902 if (FoldedVOp.getNode()) return FoldedVOp; 5903 } 5904 5905 // fold (fadd c1, c2) -> c1 + c2 5906 if (N0CFP && N1CFP) 5907 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1); 5908 // canonicalize constant to RHS 5909 if (N0CFP && !N1CFP) 5910 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0); 5911 // fold (fadd A, 0) -> A 5912 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && 5913 N1CFP->getValueAPF().isZero()) 5914 return N0; 5915 // fold (fadd A, (fneg B)) -> (fsub A, B) 5916 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) && 5917 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2) 5918 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, 5919 GetNegatedExpression(N1, DAG, LegalOperations)); 5920 // fold (fadd (fneg A), B) -> (fsub B, A) 5921 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) && 5922 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2) 5923 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1, 5924 GetNegatedExpression(N0, DAG, LegalOperations)); 5925 5926 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2)) 5927 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && 5928 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() && 5929 isa<ConstantFPSDNode>(N0.getOperand(1))) 5930 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0), 5931 DAG.getNode(ISD::FADD, SDLoc(N), VT, 5932 N0.getOperand(1), N1)); 5933 5934 // No FP constant should be created after legalization as Instruction 5935 // Selection pass has hard time in dealing with FP constant. 5936 // 5937 // We don't need test this condition for transformation like following, as 5938 // the DAG being transformed implies it is legal to take FP constant as 5939 // operand. 5940 // 5941 // (fadd (fmul c, x), x) -> (fmul c+1, x) 5942 // 5943 bool AllowNewFpConst = (Level < AfterLegalizeDAG); 5944 5945 // If allow, fold (fadd (fneg x), x) -> 0.0 5946 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath && 5947 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1) { 5948 return DAG.getConstantFP(0.0, VT); 5949 } 5950 5951 // If allow, fold (fadd x, (fneg x)) -> 0.0 5952 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath && 5953 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0) { 5954 return DAG.getConstantFP(0.0, VT); 5955 } 5956 5957 // In unsafe math mode, we can fold chains of FADD's of the same value 5958 // into multiplications. This transform is not safe in general because 5959 // we are reducing the number of rounding steps. 5960 if (DAG.getTarget().Options.UnsafeFPMath && 5961 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && 5962 !N0CFP && !N1CFP) { 5963 if (N0.getOpcode() == ISD::FMUL) { 5964 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0)); 5965 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1)); 5966 5967 // (fadd (fmul c, x), x) -> (fmul x, c+1) 5968 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) { 5969 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 5970 SDValue(CFP00, 0), 5971 DAG.getConstantFP(1.0, VT)); 5972 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 5973 N1, NewCFP); 5974 } 5975 5976 // (fadd (fmul x, c), x) -> (fmul x, c+1) 5977 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) { 5978 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 5979 SDValue(CFP01, 0), 5980 DAG.getConstantFP(1.0, VT)); 5981 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 5982 N1, NewCFP); 5983 } 5984 5985 // (fadd (fmul c, x), (fadd x, x)) -> (fmul x, c+2) 5986 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD && 5987 N1.getOperand(0) == N1.getOperand(1) && 5988 N0.getOperand(1) == N1.getOperand(0)) { 5989 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 5990 SDValue(CFP00, 0), 5991 DAG.getConstantFP(2.0, VT)); 5992 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 5993 N0.getOperand(1), NewCFP); 5994 } 5995 5996 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2) 5997 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD && 5998 N1.getOperand(0) == N1.getOperand(1) && 5999 N0.getOperand(0) == N1.getOperand(0)) { 6000 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6001 SDValue(CFP01, 0), 6002 DAG.getConstantFP(2.0, VT)); 6003 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6004 N0.getOperand(0), NewCFP); 6005 } 6006 } 6007 6008 if (N1.getOpcode() == ISD::FMUL) { 6009 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0)); 6010 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1)); 6011 6012 // (fadd x, (fmul c, x)) -> (fmul x, c+1) 6013 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) { 6014 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6015 SDValue(CFP10, 0), 6016 DAG.getConstantFP(1.0, VT)); 6017 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6018 N0, NewCFP); 6019 } 6020 6021 // (fadd x, (fmul x, c)) -> (fmul x, c+1) 6022 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) { 6023 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6024 SDValue(CFP11, 0), 6025 DAG.getConstantFP(1.0, VT)); 6026 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6027 N0, NewCFP); 6028 } 6029 6030 6031 // (fadd (fadd x, x), (fmul c, x)) -> (fmul x, c+2) 6032 if (CFP10 && !CFP11 && N0.getOpcode() == ISD::FADD && 6033 N0.getOperand(0) == N0.getOperand(1) && 6034 N1.getOperand(1) == N0.getOperand(0)) { 6035 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6036 SDValue(CFP10, 0), 6037 DAG.getConstantFP(2.0, VT)); 6038 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6039 N1.getOperand(1), NewCFP); 6040 } 6041 6042 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2) 6043 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD && 6044 N0.getOperand(0) == N0.getOperand(1) && 6045 N1.getOperand(0) == N0.getOperand(0)) { 6046 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6047 SDValue(CFP11, 0), 6048 DAG.getConstantFP(2.0, VT)); 6049 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6050 N1.getOperand(0), NewCFP); 6051 } 6052 } 6053 6054 if (N0.getOpcode() == ISD::FADD && AllowNewFpConst) { 6055 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0)); 6056 // (fadd (fadd x, x), x) -> (fmul x, 3.0) 6057 if (!CFP && N0.getOperand(0) == N0.getOperand(1) && 6058 (N0.getOperand(0) == N1)) { 6059 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6060 N1, DAG.getConstantFP(3.0, VT)); 6061 } 6062 } 6063 6064 if (N1.getOpcode() == ISD::FADD && AllowNewFpConst) { 6065 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0)); 6066 // (fadd x, (fadd x, x)) -> (fmul x, 3.0) 6067 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) && 6068 N1.getOperand(0) == N0) { 6069 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6070 N0, DAG.getConstantFP(3.0, VT)); 6071 } 6072 } 6073 6074 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0) 6075 if (AllowNewFpConst && 6076 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD && 6077 N0.getOperand(0) == N0.getOperand(1) && 6078 N1.getOperand(0) == N1.getOperand(1) && 6079 N0.getOperand(0) == N1.getOperand(0)) { 6080 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6081 N0.getOperand(0), 6082 DAG.getConstantFP(4.0, VT)); 6083 } 6084 } 6085 6086 // FADD -> FMA combines: 6087 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast || 6088 DAG.getTarget().Options.UnsafeFPMath) && 6089 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) && 6090 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) { 6091 6092 // fold (fadd (fmul x, y), z) -> (fma x, y, z) 6093 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) { 6094 return DAG.getNode(ISD::FMA, SDLoc(N), VT, 6095 N0.getOperand(0), N0.getOperand(1), N1); 6096 } 6097 6098 // fold (fadd x, (fmul y, z)) -> (fma y, z, x) 6099 // Note: Commutes FADD operands. 6100 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) { 6101 return DAG.getNode(ISD::FMA, SDLoc(N), VT, 6102 N1.getOperand(0), N1.getOperand(1), N0); 6103 } 6104 } 6105 6106 return SDValue(); 6107} 6108 6109SDValue DAGCombiner::visitFSUB(SDNode *N) { 6110 SDValue N0 = N->getOperand(0); 6111 SDValue N1 = N->getOperand(1); 6112 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6113 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6114 EVT VT = N->getValueType(0); 6115 SDLoc dl(N); 6116 6117 // fold vector ops 6118 if (VT.isVector()) { 6119 SDValue FoldedVOp = SimplifyVBinOp(N); 6120 if (FoldedVOp.getNode()) return FoldedVOp; 6121 } 6122 6123 // fold (fsub c1, c2) -> c1-c2 6124 if (N0CFP && N1CFP) 6125 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1); 6126 // fold (fsub A, 0) -> A 6127 if (DAG.getTarget().Options.UnsafeFPMath && 6128 N1CFP && N1CFP->getValueAPF().isZero()) 6129 return N0; 6130 // fold (fsub 0, B) -> -B 6131 if (DAG.getTarget().Options.UnsafeFPMath && 6132 N0CFP && N0CFP->getValueAPF().isZero()) { 6133 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options)) 6134 return GetNegatedExpression(N1, DAG, LegalOperations); 6135 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) 6136 return DAG.getNode(ISD::FNEG, dl, VT, N1); 6137 } 6138 // fold (fsub A, (fneg B)) -> (fadd A, B) 6139 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options)) 6140 return DAG.getNode(ISD::FADD, dl, VT, N0, 6141 GetNegatedExpression(N1, DAG, LegalOperations)); 6142 6143 // If 'unsafe math' is enabled, fold 6144 // (fsub x, x) -> 0.0 & 6145 // (fsub x, (fadd x, y)) -> (fneg y) & 6146 // (fsub x, (fadd y, x)) -> (fneg y) 6147 if (DAG.getTarget().Options.UnsafeFPMath) { 6148 if (N0 == N1) 6149 return DAG.getConstantFP(0.0f, VT); 6150 6151 if (N1.getOpcode() == ISD::FADD) { 6152 SDValue N10 = N1->getOperand(0); 6153 SDValue N11 = N1->getOperand(1); 6154 6155 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, 6156 &DAG.getTarget().Options)) 6157 return GetNegatedExpression(N11, DAG, LegalOperations); 6158 else if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, 6159 &DAG.getTarget().Options)) 6160 return GetNegatedExpression(N10, DAG, LegalOperations); 6161 } 6162 } 6163 6164 // FSUB -> FMA combines: 6165 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast || 6166 DAG.getTarget().Options.UnsafeFPMath) && 6167 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) && 6168 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) { 6169 6170 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z)) 6171 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) { 6172 return DAG.getNode(ISD::FMA, dl, VT, 6173 N0.getOperand(0), N0.getOperand(1), 6174 DAG.getNode(ISD::FNEG, dl, VT, N1)); 6175 } 6176 6177 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x) 6178 // Note: Commutes FSUB operands. 6179 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) { 6180 return DAG.getNode(ISD::FMA, dl, VT, 6181 DAG.getNode(ISD::FNEG, dl, VT, 6182 N1.getOperand(0)), 6183 N1.getOperand(1), N0); 6184 } 6185 6186 // fold (fsub (-(fmul, x, y)), z) -> (fma (fneg x), y, (fneg z)) 6187 if (N0.getOpcode() == ISD::FNEG && 6188 N0.getOperand(0).getOpcode() == ISD::FMUL && 6189 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) { 6190 SDValue N00 = N0.getOperand(0).getOperand(0); 6191 SDValue N01 = N0.getOperand(0).getOperand(1); 6192 return DAG.getNode(ISD::FMA, dl, VT, 6193 DAG.getNode(ISD::FNEG, dl, VT, N00), N01, 6194 DAG.getNode(ISD::FNEG, dl, VT, N1)); 6195 } 6196 } 6197 6198 return SDValue(); 6199} 6200 6201SDValue DAGCombiner::visitFMUL(SDNode *N) { 6202 SDValue N0 = N->getOperand(0); 6203 SDValue N1 = N->getOperand(1); 6204 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6205 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6206 EVT VT = N->getValueType(0); 6207 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 6208 6209 // fold vector ops 6210 if (VT.isVector()) { 6211 SDValue FoldedVOp = SimplifyVBinOp(N); 6212 if (FoldedVOp.getNode()) return FoldedVOp; 6213 } 6214 6215 // fold (fmul c1, c2) -> c1*c2 6216 if (N0CFP && N1CFP) 6217 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1); 6218 // canonicalize constant to RHS 6219 if (N0CFP && !N1CFP) 6220 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0); 6221 // fold (fmul A, 0) -> 0 6222 if (DAG.getTarget().Options.UnsafeFPMath && 6223 N1CFP && N1CFP->getValueAPF().isZero()) 6224 return N1; 6225 // fold (fmul A, 0) -> 0, vector edition. 6226 if (DAG.getTarget().Options.UnsafeFPMath && 6227 ISD::isBuildVectorAllZeros(N1.getNode())) 6228 return N1; 6229 // fold (fmul A, 1.0) -> A 6230 if (N1CFP && N1CFP->isExactlyValue(1.0)) 6231 return N0; 6232 // fold (fmul X, 2.0) -> (fadd X, X) 6233 if (N1CFP && N1CFP->isExactlyValue(+2.0)) 6234 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0); 6235 // fold (fmul X, -1.0) -> (fneg X) 6236 if (N1CFP && N1CFP->isExactlyValue(-1.0)) 6237 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) 6238 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0); 6239 6240 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y) 6241 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, 6242 &DAG.getTarget().Options)) { 6243 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, 6244 &DAG.getTarget().Options)) { 6245 // Both can be negated for free, check to see if at least one is cheaper 6246 // negated. 6247 if (LHSNeg == 2 || RHSNeg == 2) 6248 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6249 GetNegatedExpression(N0, DAG, LegalOperations), 6250 GetNegatedExpression(N1, DAG, LegalOperations)); 6251 } 6252 } 6253 6254 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2)) 6255 if (DAG.getTarget().Options.UnsafeFPMath && 6256 N1CFP && N0.getOpcode() == ISD::FMUL && 6257 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1))) 6258 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), 6259 DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6260 N0.getOperand(1), N1)); 6261 6262 return SDValue(); 6263} 6264 6265SDValue DAGCombiner::visitFMA(SDNode *N) { 6266 SDValue N0 = N->getOperand(0); 6267 SDValue N1 = N->getOperand(1); 6268 SDValue N2 = N->getOperand(2); 6269 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6270 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6271 EVT VT = N->getValueType(0); 6272 SDLoc dl(N); 6273 6274 if (DAG.getTarget().Options.UnsafeFPMath) { 6275 if (N0CFP && N0CFP->isZero()) 6276 return N2; 6277 if (N1CFP && N1CFP->isZero()) 6278 return N2; 6279 } 6280 if (N0CFP && N0CFP->isExactlyValue(1.0)) 6281 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2); 6282 if (N1CFP && N1CFP->isExactlyValue(1.0)) 6283 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2); 6284 6285 // Canonicalize (fma c, x, y) -> (fma x, c, y) 6286 if (N0CFP && !N1CFP) 6287 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2); 6288 6289 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2) 6290 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && 6291 N2.getOpcode() == ISD::FMUL && 6292 N0 == N2.getOperand(0) && 6293 N2.getOperand(1).getOpcode() == ISD::ConstantFP) { 6294 return DAG.getNode(ISD::FMUL, dl, VT, N0, 6295 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1))); 6296 } 6297 6298 6299 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y) 6300 if (DAG.getTarget().Options.UnsafeFPMath && 6301 N0.getOpcode() == ISD::FMUL && N1CFP && 6302 N0.getOperand(1).getOpcode() == ISD::ConstantFP) { 6303 return DAG.getNode(ISD::FMA, dl, VT, 6304 N0.getOperand(0), 6305 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)), 6306 N2); 6307 } 6308 6309 // (fma x, 1, y) -> (fadd x, y) 6310 // (fma x, -1, y) -> (fadd (fneg x), y) 6311 if (N1CFP) { 6312 if (N1CFP->isExactlyValue(1.0)) 6313 return DAG.getNode(ISD::FADD, dl, VT, N0, N2); 6314 6315 if (N1CFP->isExactlyValue(-1.0) && 6316 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) { 6317 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0); 6318 AddToWorkList(RHSNeg.getNode()); 6319 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg); 6320 } 6321 } 6322 6323 // (fma x, c, x) -> (fmul x, (c+1)) 6324 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2) { 6325 return DAG.getNode(ISD::FMUL, dl, VT, 6326 N0, 6327 DAG.getNode(ISD::FADD, dl, VT, 6328 N1, DAG.getConstantFP(1.0, VT))); 6329 } 6330 6331 // (fma x, c, (fneg x)) -> (fmul x, (c-1)) 6332 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && 6333 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0) { 6334 return DAG.getNode(ISD::FMUL, dl, VT, 6335 N0, 6336 DAG.getNode(ISD::FADD, dl, VT, 6337 N1, DAG.getConstantFP(-1.0, VT))); 6338 } 6339 6340 6341 return SDValue(); 6342} 6343 6344SDValue DAGCombiner::visitFDIV(SDNode *N) { 6345 SDValue N0 = N->getOperand(0); 6346 SDValue N1 = N->getOperand(1); 6347 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6348 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6349 EVT VT = N->getValueType(0); 6350 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 6351 6352 // fold vector ops 6353 if (VT.isVector()) { 6354 SDValue FoldedVOp = SimplifyVBinOp(N); 6355 if (FoldedVOp.getNode()) return FoldedVOp; 6356 } 6357 6358 // fold (fdiv c1, c2) -> c1/c2 6359 if (N0CFP && N1CFP) 6360 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1); 6361 6362 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable. 6363 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) { 6364 // Compute the reciprocal 1.0 / c2. 6365 APFloat N1APF = N1CFP->getValueAPF(); 6366 APFloat Recip(N1APF.getSemantics(), 1); // 1.0 6367 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven); 6368 // Only do the transform if the reciprocal is a legal fp immediate that 6369 // isn't too nasty (eg NaN, denormal, ...). 6370 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty 6371 (!LegalOperations || 6372 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM 6373 // backend)... we should handle this gracefully after Legalize. 6374 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) || 6375 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) || 6376 TLI.isFPImmLegal(Recip, VT))) 6377 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, 6378 DAG.getConstantFP(Recip, VT)); 6379 } 6380 6381 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y) 6382 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, 6383 &DAG.getTarget().Options)) { 6384 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, 6385 &DAG.getTarget().Options)) { 6386 // Both can be negated for free, check to see if at least one is cheaper 6387 // negated. 6388 if (LHSNeg == 2 || RHSNeg == 2) 6389 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, 6390 GetNegatedExpression(N0, DAG, LegalOperations), 6391 GetNegatedExpression(N1, DAG, LegalOperations)); 6392 } 6393 } 6394 6395 return SDValue(); 6396} 6397 6398SDValue DAGCombiner::visitFREM(SDNode *N) { 6399 SDValue N0 = N->getOperand(0); 6400 SDValue N1 = N->getOperand(1); 6401 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6402 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6403 EVT VT = N->getValueType(0); 6404 6405 // fold (frem c1, c2) -> fmod(c1,c2) 6406 if (N0CFP && N1CFP) 6407 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1); 6408 6409 return SDValue(); 6410} 6411 6412SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) { 6413 SDValue N0 = N->getOperand(0); 6414 SDValue N1 = N->getOperand(1); 6415 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6416 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6417 EVT VT = N->getValueType(0); 6418 6419 if (N0CFP && N1CFP) // Constant fold 6420 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1); 6421 6422 if (N1CFP) { 6423 const APFloat& V = N1CFP->getValueAPF(); 6424 // copysign(x, c1) -> fabs(x) iff ispos(c1) 6425 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1) 6426 if (!V.isNegative()) { 6427 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT)) 6428 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); 6429 } else { 6430 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) 6431 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, 6432 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0)); 6433 } 6434 } 6435 6436 // copysign(fabs(x), y) -> copysign(x, y) 6437 // copysign(fneg(x), y) -> copysign(x, y) 6438 // copysign(copysign(x,z), y) -> copysign(x, y) 6439 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG || 6440 N0.getOpcode() == ISD::FCOPYSIGN) 6441 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, 6442 N0.getOperand(0), N1); 6443 6444 // copysign(x, abs(y)) -> abs(x) 6445 if (N1.getOpcode() == ISD::FABS) 6446 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); 6447 6448 // copysign(x, copysign(y,z)) -> copysign(x, z) 6449 if (N1.getOpcode() == ISD::FCOPYSIGN) 6450 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, 6451 N0, N1.getOperand(1)); 6452 6453 // copysign(x, fp_extend(y)) -> copysign(x, y) 6454 // copysign(x, fp_round(y)) -> copysign(x, y) 6455 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND) 6456 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, 6457 N0, N1.getOperand(0)); 6458 6459 return SDValue(); 6460} 6461 6462SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) { 6463 SDValue N0 = N->getOperand(0); 6464 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 6465 EVT VT = N->getValueType(0); 6466 EVT OpVT = N0.getValueType(); 6467 6468 // fold (sint_to_fp c1) -> c1fp 6469 if (N0C && 6470 // ...but only if the target supports immediate floating-point values 6471 (!LegalOperations || 6472 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) 6473 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0); 6474 6475 // If the input is a legal type, and SINT_TO_FP is not legal on this target, 6476 // but UINT_TO_FP is legal on this target, try to convert. 6477 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) && 6478 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) { 6479 // If the sign bit is known to be zero, we can change this to UINT_TO_FP. 6480 if (DAG.SignBitIsZero(N0)) 6481 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0); 6482 } 6483 6484 // The next optimizations are desireable only if SELECT_CC can be lowered. 6485 // Check against MVT::Other for SELECT_CC, which is a workaround for targets 6486 // having to say they don't support SELECT_CC on every type the DAG knows 6487 // about, since there is no way to mark an opcode illegal at all value types 6488 // (See also visitSELECT) 6489 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) { 6490 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc) 6491 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 && 6492 !VT.isVector() && 6493 (!LegalOperations || 6494 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) { 6495 SDValue Ops[] = 6496 { N0.getOperand(0), N0.getOperand(1), 6497 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT), 6498 N0.getOperand(2) }; 6499 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5); 6500 } 6501 6502 // fold (sint_to_fp (zext (setcc x, y, cc))) -> 6503 // (select_cc x, y, 1.0, 0.0,, cc) 6504 if (N0.getOpcode() == ISD::ZERO_EXTEND && 6505 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() && 6506 (!LegalOperations || 6507 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) { 6508 SDValue Ops[] = 6509 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1), 6510 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT), 6511 N0.getOperand(0).getOperand(2) }; 6512 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5); 6513 } 6514 } 6515 6516 return SDValue(); 6517} 6518 6519SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) { 6520 SDValue N0 = N->getOperand(0); 6521 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 6522 EVT VT = N->getValueType(0); 6523 EVT OpVT = N0.getValueType(); 6524 6525 // fold (uint_to_fp c1) -> c1fp 6526 if (N0C && 6527 // ...but only if the target supports immediate floating-point values 6528 (!LegalOperations || 6529 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) 6530 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0); 6531 6532 // If the input is a legal type, and UINT_TO_FP is not legal on this target, 6533 // but SINT_TO_FP is legal on this target, try to convert. 6534 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) && 6535 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) { 6536 // If the sign bit is known to be zero, we can change this to SINT_TO_FP. 6537 if (DAG.SignBitIsZero(N0)) 6538 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0); 6539 } 6540 6541 // The next optimizations are desireable only if SELECT_CC can be lowered. 6542 // Check against MVT::Other for SELECT_CC, which is a workaround for targets 6543 // having to say they don't support SELECT_CC on every type the DAG knows 6544 // about, since there is no way to mark an opcode illegal at all value types 6545 // (See also visitSELECT) 6546 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) { 6547 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc) 6548 6549 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() && 6550 (!LegalOperations || 6551 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) { 6552 SDValue Ops[] = 6553 { N0.getOperand(0), N0.getOperand(1), 6554 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT), 6555 N0.getOperand(2) }; 6556 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5); 6557 } 6558 } 6559 6560 return SDValue(); 6561} 6562 6563SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) { 6564 SDValue N0 = N->getOperand(0); 6565 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6566 EVT VT = N->getValueType(0); 6567 6568 // fold (fp_to_sint c1fp) -> c1 6569 if (N0CFP) 6570 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0); 6571 6572 return SDValue(); 6573} 6574 6575SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) { 6576 SDValue N0 = N->getOperand(0); 6577 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6578 EVT VT = N->getValueType(0); 6579 6580 // fold (fp_to_uint c1fp) -> c1 6581 if (N0CFP) 6582 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0); 6583 6584 return SDValue(); 6585} 6586 6587SDValue DAGCombiner::visitFP_ROUND(SDNode *N) { 6588 SDValue N0 = N->getOperand(0); 6589 SDValue N1 = N->getOperand(1); 6590 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6591 EVT VT = N->getValueType(0); 6592 6593 // fold (fp_round c1fp) -> c1fp 6594 if (N0CFP) 6595 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1); 6596 6597 // fold (fp_round (fp_extend x)) -> x 6598 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType()) 6599 return N0.getOperand(0); 6600 6601 // fold (fp_round (fp_round x)) -> (fp_round x) 6602 if (N0.getOpcode() == ISD::FP_ROUND) { 6603 // This is a value preserving truncation if both round's are. 6604 bool IsTrunc = N->getConstantOperandVal(1) == 1 && 6605 N0.getNode()->getConstantOperandVal(1) == 1; 6606 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0), 6607 DAG.getIntPtrConstant(IsTrunc)); 6608 } 6609 6610 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y) 6611 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) { 6612 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT, 6613 N0.getOperand(0), N1); 6614 AddToWorkList(Tmp.getNode()); 6615 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, 6616 Tmp, N0.getOperand(1)); 6617 } 6618 6619 return SDValue(); 6620} 6621 6622SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) { 6623 SDValue N0 = N->getOperand(0); 6624 EVT VT = N->getValueType(0); 6625 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT(); 6626 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6627 6628 // fold (fp_round_inreg c1fp) -> c1fp 6629 if (N0CFP && isTypeLegal(EVT)) { 6630 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT); 6631 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round); 6632 } 6633 6634 return SDValue(); 6635} 6636 6637SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) { 6638 SDValue N0 = N->getOperand(0); 6639 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6640 EVT VT = N->getValueType(0); 6641 6642 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded. 6643 if (N->hasOneUse() && 6644 N->use_begin()->getOpcode() == ISD::FP_ROUND) 6645 return SDValue(); 6646 6647 // fold (fp_extend c1fp) -> c1fp 6648 if (N0CFP) 6649 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0); 6650 6651 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the 6652 // value of X. 6653 if (N0.getOpcode() == ISD::FP_ROUND 6654 && N0.getNode()->getConstantOperandVal(1) == 1) { 6655 SDValue In = N0.getOperand(0); 6656 if (In.getValueType() == VT) return In; 6657 if (VT.bitsLT(In.getValueType())) 6658 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, 6659 In, N0.getOperand(1)); 6660 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In); 6661 } 6662 6663 // fold (fpext (load x)) -> (fpext (fptrunc (extload x))) 6664 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() && 6665 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 6666 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) { 6667 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 6668 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT, 6669 LN0->getChain(), 6670 LN0->getBasePtr(), LN0->getPointerInfo(), 6671 N0.getValueType(), 6672 LN0->isVolatile(), LN0->isNonTemporal(), 6673 LN0->getAlignment()); 6674 CombineTo(N, ExtLoad); 6675 CombineTo(N0.getNode(), 6676 DAG.getNode(ISD::FP_ROUND, SDLoc(N0), 6677 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)), 6678 ExtLoad.getValue(1)); 6679 return SDValue(N, 0); // Return N so it doesn't get rechecked! 6680 } 6681 6682 return SDValue(); 6683} 6684 6685SDValue DAGCombiner::visitFNEG(SDNode *N) { 6686 SDValue N0 = N->getOperand(0); 6687 EVT VT = N->getValueType(0); 6688 6689 if (VT.isVector()) { 6690 SDValue FoldedVOp = SimplifyVUnaryOp(N); 6691 if (FoldedVOp.getNode()) return FoldedVOp; 6692 } 6693 6694 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(), 6695 &DAG.getTarget().Options)) 6696 return GetNegatedExpression(N0, DAG, LegalOperations); 6697 6698 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading 6699 // constant pool values. 6700 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST && 6701 !VT.isVector() && 6702 N0.getNode()->hasOneUse() && 6703 N0.getOperand(0).getValueType().isInteger()) { 6704 SDValue Int = N0.getOperand(0); 6705 EVT IntVT = Int.getValueType(); 6706 if (IntVT.isInteger() && !IntVT.isVector()) { 6707 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int, 6708 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT)); 6709 AddToWorkList(Int.getNode()); 6710 return DAG.getNode(ISD::BITCAST, SDLoc(N), 6711 VT, Int); 6712 } 6713 } 6714 6715 // (fneg (fmul c, x)) -> (fmul -c, x) 6716 if (N0.getOpcode() == ISD::FMUL) { 6717 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1)); 6718 if (CFP1) { 6719 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6720 N0.getOperand(0), 6721 DAG.getNode(ISD::FNEG, SDLoc(N), VT, 6722 N0.getOperand(1))); 6723 } 6724 } 6725 6726 return SDValue(); 6727} 6728 6729SDValue DAGCombiner::visitFCEIL(SDNode *N) { 6730 SDValue N0 = N->getOperand(0); 6731 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6732 EVT VT = N->getValueType(0); 6733 6734 // fold (fceil c1) -> fceil(c1) 6735 if (N0CFP) 6736 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0); 6737 6738 return SDValue(); 6739} 6740 6741SDValue DAGCombiner::visitFTRUNC(SDNode *N) { 6742 SDValue N0 = N->getOperand(0); 6743 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6744 EVT VT = N->getValueType(0); 6745 6746 // fold (ftrunc c1) -> ftrunc(c1) 6747 if (N0CFP) 6748 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0); 6749 6750 return SDValue(); 6751} 6752 6753SDValue DAGCombiner::visitFFLOOR(SDNode *N) { 6754 SDValue N0 = N->getOperand(0); 6755 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6756 EVT VT = N->getValueType(0); 6757 6758 // fold (ffloor c1) -> ffloor(c1) 6759 if (N0CFP) 6760 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0); 6761 6762 return SDValue(); 6763} 6764 6765SDValue DAGCombiner::visitFABS(SDNode *N) { 6766 SDValue N0 = N->getOperand(0); 6767 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6768 EVT VT = N->getValueType(0); 6769 6770 if (VT.isVector()) { 6771 SDValue FoldedVOp = SimplifyVUnaryOp(N); 6772 if (FoldedVOp.getNode()) return FoldedVOp; 6773 } 6774 6775 // fold (fabs c1) -> fabs(c1) 6776 if (N0CFP) 6777 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); 6778 // fold (fabs (fabs x)) -> (fabs x) 6779 if (N0.getOpcode() == ISD::FABS) 6780 return N->getOperand(0); 6781 // fold (fabs (fneg x)) -> (fabs x) 6782 // fold (fabs (fcopysign x, y)) -> (fabs x) 6783 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN) 6784 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0)); 6785 6786 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading 6787 // constant pool values. 6788 if (!TLI.isFAbsFree(VT) && 6789 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() && 6790 N0.getOperand(0).getValueType().isInteger() && 6791 !N0.getOperand(0).getValueType().isVector()) { 6792 SDValue Int = N0.getOperand(0); 6793 EVT IntVT = Int.getValueType(); 6794 if (IntVT.isInteger() && !IntVT.isVector()) { 6795 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int, 6796 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT)); 6797 AddToWorkList(Int.getNode()); 6798 return DAG.getNode(ISD::BITCAST, SDLoc(N), 6799 N->getValueType(0), Int); 6800 } 6801 } 6802 6803 return SDValue(); 6804} 6805 6806SDValue DAGCombiner::visitBRCOND(SDNode *N) { 6807 SDValue Chain = N->getOperand(0); 6808 SDValue N1 = N->getOperand(1); 6809 SDValue N2 = N->getOperand(2); 6810 6811 // If N is a constant we could fold this into a fallthrough or unconditional 6812 // branch. However that doesn't happen very often in normal code, because 6813 // Instcombine/SimplifyCFG should have handled the available opportunities. 6814 // If we did this folding here, it would be necessary to update the 6815 // MachineBasicBlock CFG, which is awkward. 6816 6817 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal 6818 // on the target. 6819 if (N1.getOpcode() == ISD::SETCC && 6820 TLI.isOperationLegalOrCustom(ISD::BR_CC, 6821 N1.getOperand(0).getValueType())) { 6822 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other, 6823 Chain, N1.getOperand(2), 6824 N1.getOperand(0), N1.getOperand(1), N2); 6825 } 6826 6827 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) || 6828 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) && 6829 (N1.getOperand(0).hasOneUse() && 6830 N1.getOperand(0).getOpcode() == ISD::SRL))) { 6831 SDNode *Trunc = 0; 6832 if (N1.getOpcode() == ISD::TRUNCATE) { 6833 // Look pass the truncate. 6834 Trunc = N1.getNode(); 6835 N1 = N1.getOperand(0); 6836 } 6837 6838 // Match this pattern so that we can generate simpler code: 6839 // 6840 // %a = ... 6841 // %b = and i32 %a, 2 6842 // %c = srl i32 %b, 1 6843 // brcond i32 %c ... 6844 // 6845 // into 6846 // 6847 // %a = ... 6848 // %b = and i32 %a, 2 6849 // %c = setcc eq %b, 0 6850 // brcond %c ... 6851 // 6852 // This applies only when the AND constant value has one bit set and the 6853 // SRL constant is equal to the log2 of the AND constant. The back-end is 6854 // smart enough to convert the result into a TEST/JMP sequence. 6855 SDValue Op0 = N1.getOperand(0); 6856 SDValue Op1 = N1.getOperand(1); 6857 6858 if (Op0.getOpcode() == ISD::AND && 6859 Op1.getOpcode() == ISD::Constant) { 6860 SDValue AndOp1 = Op0.getOperand(1); 6861 6862 if (AndOp1.getOpcode() == ISD::Constant) { 6863 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue(); 6864 6865 if (AndConst.isPowerOf2() && 6866 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) { 6867 SDValue SetCC = 6868 DAG.getSetCC(SDLoc(N), 6869 getSetCCResultType(Op0.getValueType()), 6870 Op0, DAG.getConstant(0, Op0.getValueType()), 6871 ISD::SETNE); 6872 6873 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N), 6874 MVT::Other, Chain, SetCC, N2); 6875 // Don't add the new BRCond into the worklist or else SimplifySelectCC 6876 // will convert it back to (X & C1) >> C2. 6877 CombineTo(N, NewBRCond, false); 6878 // Truncate is dead. 6879 if (Trunc) { 6880 removeFromWorkList(Trunc); 6881 DAG.DeleteNode(Trunc); 6882 } 6883 // Replace the uses of SRL with SETCC 6884 WorkListRemover DeadNodes(*this); 6885 DAG.ReplaceAllUsesOfValueWith(N1, SetCC); 6886 removeFromWorkList(N1.getNode()); 6887 DAG.DeleteNode(N1.getNode()); 6888 return SDValue(N, 0); // Return N so it doesn't get rechecked! 6889 } 6890 } 6891 } 6892 6893 if (Trunc) 6894 // Restore N1 if the above transformation doesn't match. 6895 N1 = N->getOperand(1); 6896 } 6897 6898 // Transform br(xor(x, y)) -> br(x != y) 6899 // Transform br(xor(xor(x,y), 1)) -> br (x == y) 6900 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) { 6901 SDNode *TheXor = N1.getNode(); 6902 SDValue Op0 = TheXor->getOperand(0); 6903 SDValue Op1 = TheXor->getOperand(1); 6904 if (Op0.getOpcode() == Op1.getOpcode()) { 6905 // Avoid missing important xor optimizations. 6906 SDValue Tmp = visitXOR(TheXor); 6907 if (Tmp.getNode()) { 6908 if (Tmp.getNode() != TheXor) { 6909 DEBUG(dbgs() << "\nReplacing.8 "; 6910 TheXor->dump(&DAG); 6911 dbgs() << "\nWith: "; 6912 Tmp.getNode()->dump(&DAG); 6913 dbgs() << '\n'); 6914 WorkListRemover DeadNodes(*this); 6915 DAG.ReplaceAllUsesOfValueWith(N1, Tmp); 6916 removeFromWorkList(TheXor); 6917 DAG.DeleteNode(TheXor); 6918 return DAG.getNode(ISD::BRCOND, SDLoc(N), 6919 MVT::Other, Chain, Tmp, N2); 6920 } 6921 6922 // visitXOR has changed XOR's operands or replaced the XOR completely, 6923 // bail out. 6924 return SDValue(N, 0); 6925 } 6926 } 6927 6928 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) { 6929 bool Equal = false; 6930 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0)) 6931 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() && 6932 Op0.getOpcode() == ISD::XOR) { 6933 TheXor = Op0.getNode(); 6934 Equal = true; 6935 } 6936 6937 EVT SetCCVT = N1.getValueType(); 6938 if (LegalTypes) 6939 SetCCVT = getSetCCResultType(SetCCVT); 6940 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor), 6941 SetCCVT, 6942 Op0, Op1, 6943 Equal ? ISD::SETEQ : ISD::SETNE); 6944 // Replace the uses of XOR with SETCC 6945 WorkListRemover DeadNodes(*this); 6946 DAG.ReplaceAllUsesOfValueWith(N1, SetCC); 6947 removeFromWorkList(N1.getNode()); 6948 DAG.DeleteNode(N1.getNode()); 6949 return DAG.getNode(ISD::BRCOND, SDLoc(N), 6950 MVT::Other, Chain, SetCC, N2); 6951 } 6952 } 6953 6954 return SDValue(); 6955} 6956 6957// Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB. 6958// 6959SDValue DAGCombiner::visitBR_CC(SDNode *N) { 6960 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1)); 6961 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3); 6962 6963 // If N is a constant we could fold this into a fallthrough or unconditional 6964 // branch. However that doesn't happen very often in normal code, because 6965 // Instcombine/SimplifyCFG should have handled the available opportunities. 6966 // If we did this folding here, it would be necessary to update the 6967 // MachineBasicBlock CFG, which is awkward. 6968 6969 // Use SimplifySetCC to simplify SETCC's. 6970 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()), 6971 CondLHS, CondRHS, CC->get(), SDLoc(N), 6972 false); 6973 if (Simp.getNode()) AddToWorkList(Simp.getNode()); 6974 6975 // fold to a simpler setcc 6976 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC) 6977 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other, 6978 N->getOperand(0), Simp.getOperand(2), 6979 Simp.getOperand(0), Simp.getOperand(1), 6980 N->getOperand(4)); 6981 6982 return SDValue(); 6983} 6984 6985/// canFoldInAddressingMode - Return true if 'Use' is a load or a store that 6986/// uses N as its base pointer and that N may be folded in the load / store 6987/// addressing mode. 6988static bool canFoldInAddressingMode(SDNode *N, SDNode *Use, 6989 SelectionDAG &DAG, 6990 const TargetLowering &TLI) { 6991 EVT VT; 6992 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) { 6993 if (LD->isIndexed() || LD->getBasePtr().getNode() != N) 6994 return false; 6995 VT = Use->getValueType(0); 6996 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) { 6997 if (ST->isIndexed() || ST->getBasePtr().getNode() != N) 6998 return false; 6999 VT = ST->getValue().getValueType(); 7000 } else 7001 return false; 7002 7003 TargetLowering::AddrMode AM; 7004 if (N->getOpcode() == ISD::ADD) { 7005 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); 7006 if (Offset) 7007 // [reg +/- imm] 7008 AM.BaseOffs = Offset->getSExtValue(); 7009 else 7010 // [reg +/- reg] 7011 AM.Scale = 1; 7012 } else if (N->getOpcode() == ISD::SUB) { 7013 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); 7014 if (Offset) 7015 // [reg +/- imm] 7016 AM.BaseOffs = -Offset->getSExtValue(); 7017 else 7018 // [reg +/- reg] 7019 AM.Scale = 1; 7020 } else 7021 return false; 7022 7023 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext())); 7024} 7025 7026/// CombineToPreIndexedLoadStore - Try turning a load / store into a 7027/// pre-indexed load / store when the base pointer is an add or subtract 7028/// and it has other uses besides the load / store. After the 7029/// transformation, the new indexed load / store has effectively folded 7030/// the add / subtract in and all of its other uses are redirected to the 7031/// new load / store. 7032bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) { 7033 if (Level < AfterLegalizeDAG) 7034 return false; 7035 7036 bool isLoad = true; 7037 SDValue Ptr; 7038 EVT VT; 7039 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 7040 if (LD->isIndexed()) 7041 return false; 7042 VT = LD->getMemoryVT(); 7043 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) && 7044 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT)) 7045 return false; 7046 Ptr = LD->getBasePtr(); 7047 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 7048 if (ST->isIndexed()) 7049 return false; 7050 VT = ST->getMemoryVT(); 7051 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) && 7052 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT)) 7053 return false; 7054 Ptr = ST->getBasePtr(); 7055 isLoad = false; 7056 } else { 7057 return false; 7058 } 7059 7060 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail 7061 // out. There is no reason to make this a preinc/predec. 7062 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) || 7063 Ptr.getNode()->hasOneUse()) 7064 return false; 7065 7066 // Ask the target to do addressing mode selection. 7067 SDValue BasePtr; 7068 SDValue Offset; 7069 ISD::MemIndexedMode AM = ISD::UNINDEXED; 7070 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG)) 7071 return false; 7072 7073 // Backends without true r+i pre-indexed forms may need to pass a 7074 // constant base with a variable offset so that constant coercion 7075 // will work with the patterns in canonical form. 7076 bool Swapped = false; 7077 if (isa<ConstantSDNode>(BasePtr)) { 7078 std::swap(BasePtr, Offset); 7079 Swapped = true; 7080 } 7081 7082 // Don't create a indexed load / store with zero offset. 7083 if (isa<ConstantSDNode>(Offset) && 7084 cast<ConstantSDNode>(Offset)->isNullValue()) 7085 return false; 7086 7087 // Try turning it into a pre-indexed load / store except when: 7088 // 1) The new base ptr is a frame index. 7089 // 2) If N is a store and the new base ptr is either the same as or is a 7090 // predecessor of the value being stored. 7091 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded 7092 // that would create a cycle. 7093 // 4) All uses are load / store ops that use it as old base ptr. 7094 7095 // Check #1. Preinc'ing a frame index would require copying the stack pointer 7096 // (plus the implicit offset) to a register to preinc anyway. 7097 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr)) 7098 return false; 7099 7100 // Check #2. 7101 if (!isLoad) { 7102 SDValue Val = cast<StoreSDNode>(N)->getValue(); 7103 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode())) 7104 return false; 7105 } 7106 7107 // If the offset is a constant, there may be other adds of constants that 7108 // can be folded with this one. We should do this to avoid having to keep 7109 // a copy of the original base pointer. 7110 SmallVector<SDNode *, 16> OtherUses; 7111 if (isa<ConstantSDNode>(Offset)) 7112 for (SDNode::use_iterator I = BasePtr.getNode()->use_begin(), 7113 E = BasePtr.getNode()->use_end(); I != E; ++I) { 7114 SDNode *Use = *I; 7115 if (Use == Ptr.getNode()) 7116 continue; 7117 7118 if (Use->isPredecessorOf(N)) 7119 continue; 7120 7121 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) { 7122 OtherUses.clear(); 7123 break; 7124 } 7125 7126 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1); 7127 if (Op1.getNode() == BasePtr.getNode()) 7128 std::swap(Op0, Op1); 7129 assert(Op0.getNode() == BasePtr.getNode() && 7130 "Use of ADD/SUB but not an operand"); 7131 7132 if (!isa<ConstantSDNode>(Op1)) { 7133 OtherUses.clear(); 7134 break; 7135 } 7136 7137 // FIXME: In some cases, we can be smarter about this. 7138 if (Op1.getValueType() != Offset.getValueType()) { 7139 OtherUses.clear(); 7140 break; 7141 } 7142 7143 OtherUses.push_back(Use); 7144 } 7145 7146 if (Swapped) 7147 std::swap(BasePtr, Offset); 7148 7149 // Now check for #3 and #4. 7150 bool RealUse = false; 7151 7152 // Caches for hasPredecessorHelper 7153 SmallPtrSet<const SDNode *, 32> Visited; 7154 SmallVector<const SDNode *, 16> Worklist; 7155 7156 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(), 7157 E = Ptr.getNode()->use_end(); I != E; ++I) { 7158 SDNode *Use = *I; 7159 if (Use == N) 7160 continue; 7161 if (N->hasPredecessorHelper(Use, Visited, Worklist)) 7162 return false; 7163 7164 // If Ptr may be folded in addressing mode of other use, then it's 7165 // not profitable to do this transformation. 7166 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI)) 7167 RealUse = true; 7168 } 7169 7170 if (!RealUse) 7171 return false; 7172 7173 SDValue Result; 7174 if (isLoad) 7175 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N), 7176 BasePtr, Offset, AM); 7177 else 7178 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N), 7179 BasePtr, Offset, AM); 7180 ++PreIndexedNodes; 7181 ++NodesCombined; 7182 DEBUG(dbgs() << "\nReplacing.4 "; 7183 N->dump(&DAG); 7184 dbgs() << "\nWith: "; 7185 Result.getNode()->dump(&DAG); 7186 dbgs() << '\n'); 7187 WorkListRemover DeadNodes(*this); 7188 if (isLoad) { 7189 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0)); 7190 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2)); 7191 } else { 7192 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1)); 7193 } 7194 7195 // Finally, since the node is now dead, remove it from the graph. 7196 DAG.DeleteNode(N); 7197 7198 if (Swapped) 7199 std::swap(BasePtr, Offset); 7200 7201 // Replace other uses of BasePtr that can be updated to use Ptr 7202 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) { 7203 unsigned OffsetIdx = 1; 7204 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode()) 7205 OffsetIdx = 0; 7206 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() == 7207 BasePtr.getNode() && "Expected BasePtr operand"); 7208 7209 // We need to replace ptr0 in the following expression: 7210 // x0 * offset0 + y0 * ptr0 = t0 7211 // knowing that 7212 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store) 7213 // 7214 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the 7215 // indexed load/store and the expresion that needs to be re-written. 7216 // 7217 // Therefore, we have: 7218 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1 7219 7220 ConstantSDNode *CN = 7221 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx)); 7222 int X0, X1, Y0, Y1; 7223 APInt Offset0 = CN->getAPIntValue(); 7224 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue(); 7225 7226 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1; 7227 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1; 7228 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1; 7229 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1; 7230 7231 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD; 7232 7233 APInt CNV = Offset0; 7234 if (X0 < 0) CNV = -CNV; 7235 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1; 7236 else CNV = CNV - Offset1; 7237 7238 // We can now generate the new expression. 7239 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0)); 7240 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0); 7241 7242 SDValue NewUse = DAG.getNode(Opcode, 7243 SDLoc(OtherUses[i]), 7244 OtherUses[i]->getValueType(0), NewOp1, NewOp2); 7245 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse); 7246 removeFromWorkList(OtherUses[i]); 7247 DAG.DeleteNode(OtherUses[i]); 7248 } 7249 7250 // Replace the uses of Ptr with uses of the updated base value. 7251 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0)); 7252 removeFromWorkList(Ptr.getNode()); 7253 DAG.DeleteNode(Ptr.getNode()); 7254 7255 return true; 7256} 7257 7258/// CombineToPostIndexedLoadStore - Try to combine a load / store with a 7259/// add / sub of the base pointer node into a post-indexed load / store. 7260/// The transformation folded the add / subtract into the new indexed 7261/// load / store effectively and all of its uses are redirected to the 7262/// new load / store. 7263bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) { 7264 if (Level < AfterLegalizeDAG) 7265 return false; 7266 7267 bool isLoad = true; 7268 SDValue Ptr; 7269 EVT VT; 7270 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 7271 if (LD->isIndexed()) 7272 return false; 7273 VT = LD->getMemoryVT(); 7274 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) && 7275 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT)) 7276 return false; 7277 Ptr = LD->getBasePtr(); 7278 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 7279 if (ST->isIndexed()) 7280 return false; 7281 VT = ST->getMemoryVT(); 7282 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) && 7283 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT)) 7284 return false; 7285 Ptr = ST->getBasePtr(); 7286 isLoad = false; 7287 } else { 7288 return false; 7289 } 7290 7291 if (Ptr.getNode()->hasOneUse()) 7292 return false; 7293 7294 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(), 7295 E = Ptr.getNode()->use_end(); I != E; ++I) { 7296 SDNode *Op = *I; 7297 if (Op == N || 7298 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)) 7299 continue; 7300 7301 SDValue BasePtr; 7302 SDValue Offset; 7303 ISD::MemIndexedMode AM = ISD::UNINDEXED; 7304 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) { 7305 // Don't create a indexed load / store with zero offset. 7306 if (isa<ConstantSDNode>(Offset) && 7307 cast<ConstantSDNode>(Offset)->isNullValue()) 7308 continue; 7309 7310 // Try turning it into a post-indexed load / store except when 7311 // 1) All uses are load / store ops that use it as base ptr (and 7312 // it may be folded as addressing mmode). 7313 // 2) Op must be independent of N, i.e. Op is neither a predecessor 7314 // nor a successor of N. Otherwise, if Op is folded that would 7315 // create a cycle. 7316 7317 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr)) 7318 continue; 7319 7320 // Check for #1. 7321 bool TryNext = false; 7322 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(), 7323 EE = BasePtr.getNode()->use_end(); II != EE; ++II) { 7324 SDNode *Use = *II; 7325 if (Use == Ptr.getNode()) 7326 continue; 7327 7328 // If all the uses are load / store addresses, then don't do the 7329 // transformation. 7330 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){ 7331 bool RealUse = false; 7332 for (SDNode::use_iterator III = Use->use_begin(), 7333 EEE = Use->use_end(); III != EEE; ++III) { 7334 SDNode *UseUse = *III; 7335 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI)) 7336 RealUse = true; 7337 } 7338 7339 if (!RealUse) { 7340 TryNext = true; 7341 break; 7342 } 7343 } 7344 } 7345 7346 if (TryNext) 7347 continue; 7348 7349 // Check for #2 7350 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) { 7351 SDValue Result = isLoad 7352 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N), 7353 BasePtr, Offset, AM) 7354 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N), 7355 BasePtr, Offset, AM); 7356 ++PostIndexedNodes; 7357 ++NodesCombined; 7358 DEBUG(dbgs() << "\nReplacing.5 "; 7359 N->dump(&DAG); 7360 dbgs() << "\nWith: "; 7361 Result.getNode()->dump(&DAG); 7362 dbgs() << '\n'); 7363 WorkListRemover DeadNodes(*this); 7364 if (isLoad) { 7365 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0)); 7366 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2)); 7367 } else { 7368 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1)); 7369 } 7370 7371 // Finally, since the node is now dead, remove it from the graph. 7372 DAG.DeleteNode(N); 7373 7374 // Replace the uses of Use with uses of the updated base value. 7375 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0), 7376 Result.getValue(isLoad ? 1 : 0)); 7377 removeFromWorkList(Op); 7378 DAG.DeleteNode(Op); 7379 return true; 7380 } 7381 } 7382 } 7383 7384 return false; 7385} 7386 7387SDValue DAGCombiner::visitLOAD(SDNode *N) { 7388 LoadSDNode *LD = cast<LoadSDNode>(N); 7389 SDValue Chain = LD->getChain(); 7390 SDValue Ptr = LD->getBasePtr(); 7391 7392 // If load is not volatile and there are no uses of the loaded value (and 7393 // the updated indexed value in case of indexed loads), change uses of the 7394 // chain value into uses of the chain input (i.e. delete the dead load). 7395 if (!LD->isVolatile()) { 7396 if (N->getValueType(1) == MVT::Other) { 7397 // Unindexed loads. 7398 if (!N->hasAnyUseOfValue(0)) { 7399 // It's not safe to use the two value CombineTo variant here. e.g. 7400 // v1, chain2 = load chain1, loc 7401 // v2, chain3 = load chain2, loc 7402 // v3 = add v2, c 7403 // Now we replace use of chain2 with chain1. This makes the second load 7404 // isomorphic to the one we are deleting, and thus makes this load live. 7405 DEBUG(dbgs() << "\nReplacing.6 "; 7406 N->dump(&DAG); 7407 dbgs() << "\nWith chain: "; 7408 Chain.getNode()->dump(&DAG); 7409 dbgs() << "\n"); 7410 WorkListRemover DeadNodes(*this); 7411 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain); 7412 7413 if (N->use_empty()) { 7414 removeFromWorkList(N); 7415 DAG.DeleteNode(N); 7416 } 7417 7418 return SDValue(N, 0); // Return N so it doesn't get rechecked! 7419 } 7420 } else { 7421 // Indexed loads. 7422 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?"); 7423 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) { 7424 SDValue Undef = DAG.getUNDEF(N->getValueType(0)); 7425 DEBUG(dbgs() << "\nReplacing.7 "; 7426 N->dump(&DAG); 7427 dbgs() << "\nWith: "; 7428 Undef.getNode()->dump(&DAG); 7429 dbgs() << " and 2 other values\n"); 7430 WorkListRemover DeadNodes(*this); 7431 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef); 7432 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), 7433 DAG.getUNDEF(N->getValueType(1))); 7434 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain); 7435 removeFromWorkList(N); 7436 DAG.DeleteNode(N); 7437 return SDValue(N, 0); // Return N so it doesn't get rechecked! 7438 } 7439 } 7440 } 7441 7442 // If this load is directly stored, replace the load value with the stored 7443 // value. 7444 // TODO: Handle store large -> read small portion. 7445 // TODO: Handle TRUNCSTORE/LOADEXT 7446 if (ISD::isNormalLoad(N) && !LD->isVolatile()) { 7447 if (ISD::isNON_TRUNCStore(Chain.getNode())) { 7448 StoreSDNode *PrevST = cast<StoreSDNode>(Chain); 7449 if (PrevST->getBasePtr() == Ptr && 7450 PrevST->getValue().getValueType() == N->getValueType(0)) 7451 return CombineTo(N, Chain.getOperand(1), Chain); 7452 } 7453 } 7454 7455 // Try to infer better alignment information than the load already has. 7456 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) { 7457 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) { 7458 if (Align > LD->getMemOperand()->getBaseAlignment()) { 7459 SDValue NewLoad = 7460 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N), 7461 LD->getValueType(0), 7462 Chain, Ptr, LD->getPointerInfo(), 7463 LD->getMemoryVT(), 7464 LD->isVolatile(), LD->isNonTemporal(), Align); 7465 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true); 7466 } 7467 } 7468 } 7469 7470 if (CombinerAA) { 7471 // Walk up chain skipping non-aliasing memory nodes. 7472 SDValue BetterChain = FindBetterChain(N, Chain); 7473 7474 // If there is a better chain. 7475 if (Chain != BetterChain) { 7476 SDValue ReplLoad; 7477 7478 // Replace the chain to void dependency. 7479 if (LD->getExtensionType() == ISD::NON_EXTLOAD) { 7480 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD), 7481 BetterChain, Ptr, LD->getPointerInfo(), 7482 LD->isVolatile(), LD->isNonTemporal(), 7483 LD->isInvariant(), LD->getAlignment()); 7484 } else { 7485 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), 7486 LD->getValueType(0), 7487 BetterChain, Ptr, LD->getPointerInfo(), 7488 LD->getMemoryVT(), 7489 LD->isVolatile(), 7490 LD->isNonTemporal(), 7491 LD->getAlignment()); 7492 } 7493 7494 // Create token factor to keep old chain connected. 7495 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N), 7496 MVT::Other, Chain, ReplLoad.getValue(1)); 7497 7498 // Make sure the new and old chains are cleaned up. 7499 AddToWorkList(Token.getNode()); 7500 7501 // Replace uses with load result and token factor. Don't add users 7502 // to work list. 7503 return CombineTo(N, ReplLoad.getValue(0), Token, false); 7504 } 7505 } 7506 7507 // Try transforming N to an indexed load. 7508 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) 7509 return SDValue(N, 0); 7510 7511 return SDValue(); 7512} 7513 7514/// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the 7515/// load is having specific bytes cleared out. If so, return the byte size 7516/// being masked out and the shift amount. 7517static std::pair<unsigned, unsigned> 7518CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) { 7519 std::pair<unsigned, unsigned> Result(0, 0); 7520 7521 // Check for the structure we're looking for. 7522 if (V->getOpcode() != ISD::AND || 7523 !isa<ConstantSDNode>(V->getOperand(1)) || 7524 !ISD::isNormalLoad(V->getOperand(0).getNode())) 7525 return Result; 7526 7527 // Check the chain and pointer. 7528 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0)); 7529 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer. 7530 7531 // The store should be chained directly to the load or be an operand of a 7532 // tokenfactor. 7533 if (LD == Chain.getNode()) 7534 ; // ok. 7535 else if (Chain->getOpcode() != ISD::TokenFactor) 7536 return Result; // Fail. 7537 else { 7538 bool isOk = false; 7539 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i) 7540 if (Chain->getOperand(i).getNode() == LD) { 7541 isOk = true; 7542 break; 7543 } 7544 if (!isOk) return Result; 7545 } 7546 7547 // This only handles simple types. 7548 if (V.getValueType() != MVT::i16 && 7549 V.getValueType() != MVT::i32 && 7550 V.getValueType() != MVT::i64) 7551 return Result; 7552 7553 // Check the constant mask. Invert it so that the bits being masked out are 7554 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits 7555 // follow the sign bit for uniformity. 7556 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue(); 7557 unsigned NotMaskLZ = countLeadingZeros(NotMask); 7558 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte. 7559 unsigned NotMaskTZ = countTrailingZeros(NotMask); 7560 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte. 7561 if (NotMaskLZ == 64) return Result; // All zero mask. 7562 7563 // See if we have a continuous run of bits. If so, we have 0*1+0* 7564 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64) 7565 return Result; 7566 7567 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64. 7568 if (V.getValueType() != MVT::i64 && NotMaskLZ) 7569 NotMaskLZ -= 64-V.getValueSizeInBits(); 7570 7571 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8; 7572 switch (MaskedBytes) { 7573 case 1: 7574 case 2: 7575 case 4: break; 7576 default: return Result; // All one mask, or 5-byte mask. 7577 } 7578 7579 // Verify that the first bit starts at a multiple of mask so that the access 7580 // is aligned the same as the access width. 7581 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result; 7582 7583 Result.first = MaskedBytes; 7584 Result.second = NotMaskTZ/8; 7585 return Result; 7586} 7587 7588 7589/// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that 7590/// provides a value as specified by MaskInfo. If so, replace the specified 7591/// store with a narrower store of truncated IVal. 7592static SDNode * 7593ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo, 7594 SDValue IVal, StoreSDNode *St, 7595 DAGCombiner *DC) { 7596 unsigned NumBytes = MaskInfo.first; 7597 unsigned ByteShift = MaskInfo.second; 7598 SelectionDAG &DAG = DC->getDAG(); 7599 7600 // Check to see if IVal is all zeros in the part being masked in by the 'or' 7601 // that uses this. If not, this is not a replacement. 7602 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(), 7603 ByteShift*8, (ByteShift+NumBytes)*8); 7604 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0; 7605 7606 // Check that it is legal on the target to do this. It is legal if the new 7607 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type 7608 // legalization. 7609 MVT VT = MVT::getIntegerVT(NumBytes*8); 7610 if (!DC->isTypeLegal(VT)) 7611 return 0; 7612 7613 // Okay, we can do this! Replace the 'St' store with a store of IVal that is 7614 // shifted by ByteShift and truncated down to NumBytes. 7615 if (ByteShift) 7616 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal, 7617 DAG.getConstant(ByteShift*8, 7618 DC->getShiftAmountTy(IVal.getValueType()))); 7619 7620 // Figure out the offset for the store and the alignment of the access. 7621 unsigned StOffset; 7622 unsigned NewAlign = St->getAlignment(); 7623 7624 if (DAG.getTargetLoweringInfo().isLittleEndian()) 7625 StOffset = ByteShift; 7626 else 7627 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes; 7628 7629 SDValue Ptr = St->getBasePtr(); 7630 if (StOffset) { 7631 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(), 7632 Ptr, DAG.getConstant(StOffset, Ptr.getValueType())); 7633 NewAlign = MinAlign(NewAlign, StOffset); 7634 } 7635 7636 // Truncate down to the new size. 7637 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal); 7638 7639 ++OpsNarrowed; 7640 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr, 7641 St->getPointerInfo().getWithOffset(StOffset), 7642 false, false, NewAlign).getNode(); 7643} 7644 7645 7646/// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is 7647/// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some 7648/// of the loaded bits, try narrowing the load and store if it would end up 7649/// being a win for performance or code size. 7650SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) { 7651 StoreSDNode *ST = cast<StoreSDNode>(N); 7652 if (ST->isVolatile()) 7653 return SDValue(); 7654 7655 SDValue Chain = ST->getChain(); 7656 SDValue Value = ST->getValue(); 7657 SDValue Ptr = ST->getBasePtr(); 7658 EVT VT = Value.getValueType(); 7659 7660 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse()) 7661 return SDValue(); 7662 7663 unsigned Opc = Value.getOpcode(); 7664 7665 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst 7666 // is a byte mask indicating a consecutive number of bytes, check to see if 7667 // Y is known to provide just those bytes. If so, we try to replace the 7668 // load + replace + store sequence with a single (narrower) store, which makes 7669 // the load dead. 7670 if (Opc == ISD::OR) { 7671 std::pair<unsigned, unsigned> MaskedLoad; 7672 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain); 7673 if (MaskedLoad.first) 7674 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad, 7675 Value.getOperand(1), ST,this)) 7676 return SDValue(NewST, 0); 7677 7678 // Or is commutative, so try swapping X and Y. 7679 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain); 7680 if (MaskedLoad.first) 7681 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad, 7682 Value.getOperand(0), ST,this)) 7683 return SDValue(NewST, 0); 7684 } 7685 7686 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) || 7687 Value.getOperand(1).getOpcode() != ISD::Constant) 7688 return SDValue(); 7689 7690 SDValue N0 = Value.getOperand(0); 7691 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && 7692 Chain == SDValue(N0.getNode(), 1)) { 7693 LoadSDNode *LD = cast<LoadSDNode>(N0); 7694 if (LD->getBasePtr() != Ptr || 7695 LD->getPointerInfo().getAddrSpace() != 7696 ST->getPointerInfo().getAddrSpace()) 7697 return SDValue(); 7698 7699 // Find the type to narrow it the load / op / store to. 7700 SDValue N1 = Value.getOperand(1); 7701 unsigned BitWidth = N1.getValueSizeInBits(); 7702 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue(); 7703 if (Opc == ISD::AND) 7704 Imm ^= APInt::getAllOnesValue(BitWidth); 7705 if (Imm == 0 || Imm.isAllOnesValue()) 7706 return SDValue(); 7707 unsigned ShAmt = Imm.countTrailingZeros(); 7708 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1; 7709 unsigned NewBW = NextPowerOf2(MSB - ShAmt); 7710 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW); 7711 while (NewBW < BitWidth && 7712 !(TLI.isOperationLegalOrCustom(Opc, NewVT) && 7713 TLI.isNarrowingProfitable(VT, NewVT))) { 7714 NewBW = NextPowerOf2(NewBW); 7715 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW); 7716 } 7717 if (NewBW >= BitWidth) 7718 return SDValue(); 7719 7720 // If the lsb changed does not start at the type bitwidth boundary, 7721 // start at the previous one. 7722 if (ShAmt % NewBW) 7723 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW; 7724 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, 7725 std::min(BitWidth, ShAmt + NewBW)); 7726 if ((Imm & Mask) == Imm) { 7727 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW); 7728 if (Opc == ISD::AND) 7729 NewImm ^= APInt::getAllOnesValue(NewBW); 7730 uint64_t PtrOff = ShAmt / 8; 7731 // For big endian targets, we need to adjust the offset to the pointer to 7732 // load the correct bytes. 7733 if (TLI.isBigEndian()) 7734 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff; 7735 7736 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff); 7737 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext()); 7738 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy)) 7739 return SDValue(); 7740 7741 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD), 7742 Ptr.getValueType(), Ptr, 7743 DAG.getConstant(PtrOff, Ptr.getValueType())); 7744 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0), 7745 LD->getChain(), NewPtr, 7746 LD->getPointerInfo().getWithOffset(PtrOff), 7747 LD->isVolatile(), LD->isNonTemporal(), 7748 LD->isInvariant(), NewAlign); 7749 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD, 7750 DAG.getConstant(NewImm, NewVT)); 7751 SDValue NewST = DAG.getStore(Chain, SDLoc(N), 7752 NewVal, NewPtr, 7753 ST->getPointerInfo().getWithOffset(PtrOff), 7754 false, false, NewAlign); 7755 7756 AddToWorkList(NewPtr.getNode()); 7757 AddToWorkList(NewLD.getNode()); 7758 AddToWorkList(NewVal.getNode()); 7759 WorkListRemover DeadNodes(*this); 7760 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1)); 7761 ++OpsNarrowed; 7762 return NewST; 7763 } 7764 } 7765 7766 return SDValue(); 7767} 7768 7769/// TransformFPLoadStorePair - For a given floating point load / store pair, 7770/// if the load value isn't used by any other operations, then consider 7771/// transforming the pair to integer load / store operations if the target 7772/// deems the transformation profitable. 7773SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) { 7774 StoreSDNode *ST = cast<StoreSDNode>(N); 7775 SDValue Chain = ST->getChain(); 7776 SDValue Value = ST->getValue(); 7777 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) && 7778 Value.hasOneUse() && 7779 Chain == SDValue(Value.getNode(), 1)) { 7780 LoadSDNode *LD = cast<LoadSDNode>(Value); 7781 EVT VT = LD->getMemoryVT(); 7782 if (!VT.isFloatingPoint() || 7783 VT != ST->getMemoryVT() || 7784 LD->isNonTemporal() || 7785 ST->isNonTemporal() || 7786 LD->getPointerInfo().getAddrSpace() != 0 || 7787 ST->getPointerInfo().getAddrSpace() != 0) 7788 return SDValue(); 7789 7790 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits()); 7791 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) || 7792 !TLI.isOperationLegal(ISD::STORE, IntVT) || 7793 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) || 7794 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT)) 7795 return SDValue(); 7796 7797 unsigned LDAlign = LD->getAlignment(); 7798 unsigned STAlign = ST->getAlignment(); 7799 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext()); 7800 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy); 7801 if (LDAlign < ABIAlign || STAlign < ABIAlign) 7802 return SDValue(); 7803 7804 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value), 7805 LD->getChain(), LD->getBasePtr(), 7806 LD->getPointerInfo(), 7807 false, false, false, LDAlign); 7808 7809 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N), 7810 NewLD, ST->getBasePtr(), 7811 ST->getPointerInfo(), 7812 false, false, STAlign); 7813 7814 AddToWorkList(NewLD.getNode()); 7815 AddToWorkList(NewST.getNode()); 7816 WorkListRemover DeadNodes(*this); 7817 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1)); 7818 ++LdStFP2Int; 7819 return NewST; 7820 } 7821 7822 return SDValue(); 7823} 7824 7825/// Helper struct to parse and store a memory address as base + index + offset. 7826/// We ignore sign extensions when it is safe to do so. 7827/// The following two expressions are not equivalent. To differentiate we need 7828/// to store whether there was a sign extension involved in the index 7829/// computation. 7830/// (load (i64 add (i64 copyfromreg %c) 7831/// (i64 signextend (add (i8 load %index) 7832/// (i8 1)))) 7833/// vs 7834/// 7835/// (load (i64 add (i64 copyfromreg %c) 7836/// (i64 signextend (i32 add (i32 signextend (i8 load %index)) 7837/// (i32 1))))) 7838struct BaseIndexOffset { 7839 SDValue Base; 7840 SDValue Index; 7841 int64_t Offset; 7842 bool IsIndexSignExt; 7843 7844 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {} 7845 7846 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset, 7847 bool IsIndexSignExt) : 7848 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {} 7849 7850 bool equalBaseIndex(const BaseIndexOffset &Other) { 7851 return Other.Base == Base && Other.Index == Index && 7852 Other.IsIndexSignExt == IsIndexSignExt; 7853 } 7854 7855 /// Parses tree in Ptr for base, index, offset addresses. 7856 static BaseIndexOffset match(SDValue Ptr) { 7857 bool IsIndexSignExt = false; 7858 7859 // Just Base or possibly anything else. 7860 if (Ptr->getOpcode() != ISD::ADD) 7861 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt); 7862 7863 // Base + offset. 7864 if (isa<ConstantSDNode>(Ptr->getOperand(1))) { 7865 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue(); 7866 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset, 7867 IsIndexSignExt); 7868 } 7869 7870 // Look at Base + Index + Offset cases. 7871 SDValue Base = Ptr->getOperand(0); 7872 SDValue IndexOffset = Ptr->getOperand(1); 7873 7874 // Skip signextends. 7875 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) { 7876 IndexOffset = IndexOffset->getOperand(0); 7877 IsIndexSignExt = true; 7878 } 7879 7880 // Either the case of Base + Index (no offset) or something else. 7881 if (IndexOffset->getOpcode() != ISD::ADD) 7882 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt); 7883 7884 // Now we have the case of Base + Index + offset. 7885 SDValue Index = IndexOffset->getOperand(0); 7886 SDValue Offset = IndexOffset->getOperand(1); 7887 7888 if (!isa<ConstantSDNode>(Offset)) 7889 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt); 7890 7891 // Ignore signextends. 7892 if (Index->getOpcode() == ISD::SIGN_EXTEND) { 7893 Index = Index->getOperand(0); 7894 IsIndexSignExt = true; 7895 } else IsIndexSignExt = false; 7896 7897 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue(); 7898 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt); 7899 } 7900}; 7901 7902/// Holds a pointer to an LSBaseSDNode as well as information on where it 7903/// is located in a sequence of memory operations connected by a chain. 7904struct MemOpLink { 7905 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq): 7906 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { } 7907 // Ptr to the mem node. 7908 LSBaseSDNode *MemNode; 7909 // Offset from the base ptr. 7910 int64_t OffsetFromBase; 7911 // What is the sequence number of this mem node. 7912 // Lowest mem operand in the DAG starts at zero. 7913 unsigned SequenceNum; 7914}; 7915 7916/// Sorts store nodes in a link according to their offset from a shared 7917// base ptr. 7918struct ConsecutiveMemoryChainSorter { 7919 bool operator()(MemOpLink LHS, MemOpLink RHS) { 7920 return LHS.OffsetFromBase < RHS.OffsetFromBase; 7921 } 7922}; 7923 7924bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) { 7925 EVT MemVT = St->getMemoryVT(); 7926 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8; 7927 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes(). 7928 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat); 7929 7930 // Don't merge vectors into wider inputs. 7931 if (MemVT.isVector() || !MemVT.isSimple()) 7932 return false; 7933 7934 // Perform an early exit check. Do not bother looking at stored values that 7935 // are not constants or loads. 7936 SDValue StoredVal = St->getValue(); 7937 bool IsLoadSrc = isa<LoadSDNode>(StoredVal); 7938 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) && 7939 !IsLoadSrc) 7940 return false; 7941 7942 // Only look at ends of store sequences. 7943 SDValue Chain = SDValue(St, 1); 7944 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE) 7945 return false; 7946 7947 // This holds the base pointer, index, and the offset in bytes from the base 7948 // pointer. 7949 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr()); 7950 7951 // We must have a base and an offset. 7952 if (!BasePtr.Base.getNode()) 7953 return false; 7954 7955 // Do not handle stores to undef base pointers. 7956 if (BasePtr.Base.getOpcode() == ISD::UNDEF) 7957 return false; 7958 7959 // Save the LoadSDNodes that we find in the chain. 7960 // We need to make sure that these nodes do not interfere with 7961 // any of the store nodes. 7962 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes; 7963 7964 // Save the StoreSDNodes that we find in the chain. 7965 SmallVector<MemOpLink, 8> StoreNodes; 7966 7967 // Walk up the chain and look for nodes with offsets from the same 7968 // base pointer. Stop when reaching an instruction with a different kind 7969 // or instruction which has a different base pointer. 7970 unsigned Seq = 0; 7971 StoreSDNode *Index = St; 7972 while (Index) { 7973 // If the chain has more than one use, then we can't reorder the mem ops. 7974 if (Index != St && !SDValue(Index, 1)->hasOneUse()) 7975 break; 7976 7977 // Find the base pointer and offset for this memory node. 7978 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr()); 7979 7980 // Check that the base pointer is the same as the original one. 7981 if (!Ptr.equalBaseIndex(BasePtr)) 7982 break; 7983 7984 // Check that the alignment is the same. 7985 if (Index->getAlignment() != St->getAlignment()) 7986 break; 7987 7988 // The memory operands must not be volatile. 7989 if (Index->isVolatile() || Index->isIndexed()) 7990 break; 7991 7992 // No truncation. 7993 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index)) 7994 if (St->isTruncatingStore()) 7995 break; 7996 7997 // The stored memory type must be the same. 7998 if (Index->getMemoryVT() != MemVT) 7999 break; 8000 8001 // We do not allow unaligned stores because we want to prevent overriding 8002 // stores. 8003 if (Index->getAlignment()*8 != MemVT.getSizeInBits()) 8004 break; 8005 8006 // We found a potential memory operand to merge. 8007 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++)); 8008 8009 // Find the next memory operand in the chain. If the next operand in the 8010 // chain is a store then move up and continue the scan with the next 8011 // memory operand. If the next operand is a load save it and use alias 8012 // information to check if it interferes with anything. 8013 SDNode *NextInChain = Index->getChain().getNode(); 8014 while (1) { 8015 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) { 8016 // We found a store node. Use it for the next iteration. 8017 Index = STn; 8018 break; 8019 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) { 8020 // Save the load node for later. Continue the scan. 8021 AliasLoadNodes.push_back(Ldn); 8022 NextInChain = Ldn->getChain().getNode(); 8023 continue; 8024 } else { 8025 Index = NULL; 8026 break; 8027 } 8028 } 8029 } 8030 8031 // Check if there is anything to merge. 8032 if (StoreNodes.size() < 2) 8033 return false; 8034 8035 // Sort the memory operands according to their distance from the base pointer. 8036 std::sort(StoreNodes.begin(), StoreNodes.end(), 8037 ConsecutiveMemoryChainSorter()); 8038 8039 // Scan the memory operations on the chain and find the first non-consecutive 8040 // store memory address. 8041 unsigned LastConsecutiveStore = 0; 8042 int64_t StartAddress = StoreNodes[0].OffsetFromBase; 8043 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) { 8044 8045 // Check that the addresses are consecutive starting from the second 8046 // element in the list of stores. 8047 if (i > 0) { 8048 int64_t CurrAddress = StoreNodes[i].OffsetFromBase; 8049 if (CurrAddress - StartAddress != (ElementSizeBytes * i)) 8050 break; 8051 } 8052 8053 bool Alias = false; 8054 // Check if this store interferes with any of the loads that we found. 8055 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld) 8056 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) { 8057 Alias = true; 8058 break; 8059 } 8060 // We found a load that alias with this store. Stop the sequence. 8061 if (Alias) 8062 break; 8063 8064 // Mark this node as useful. 8065 LastConsecutiveStore = i; 8066 } 8067 8068 // The node with the lowest store address. 8069 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; 8070 8071 // Store the constants into memory as one consecutive store. 8072 if (!IsLoadSrc) { 8073 unsigned LastLegalType = 0; 8074 unsigned LastLegalVectorType = 0; 8075 bool NonZero = false; 8076 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) { 8077 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); 8078 SDValue StoredVal = St->getValue(); 8079 8080 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) { 8081 NonZero |= !C->isNullValue(); 8082 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) { 8083 NonZero |= !C->getConstantFPValue()->isNullValue(); 8084 } else { 8085 // Non constant. 8086 break; 8087 } 8088 8089 // Find a legal type for the constant store. 8090 unsigned StoreBW = (i+1) * ElementSizeBytes * 8; 8091 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW); 8092 if (TLI.isTypeLegal(StoreTy)) 8093 LastLegalType = i+1; 8094 // Or check whether a truncstore is legal. 8095 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) == 8096 TargetLowering::TypePromoteInteger) { 8097 EVT LegalizedStoredValueTy = 8098 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType()); 8099 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy)) 8100 LastLegalType = i+1; 8101 } 8102 8103 // Find a legal type for the vector store. 8104 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1); 8105 if (TLI.isTypeLegal(Ty)) 8106 LastLegalVectorType = i + 1; 8107 } 8108 8109 // We only use vectors if the constant is known to be zero and the 8110 // function is not marked with the noimplicitfloat attribute. 8111 if (NonZero || NoVectors) 8112 LastLegalVectorType = 0; 8113 8114 // Check if we found a legal integer type to store. 8115 if (LastLegalType == 0 && LastLegalVectorType == 0) 8116 return false; 8117 8118 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors; 8119 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType; 8120 8121 // Make sure we have something to merge. 8122 if (NumElem < 2) 8123 return false; 8124 8125 unsigned EarliestNodeUsed = 0; 8126 for (unsigned i=0; i < NumElem; ++i) { 8127 // Find a chain for the new wide-store operand. Notice that some 8128 // of the store nodes that we found may not be selected for inclusion 8129 // in the wide store. The chain we use needs to be the chain of the 8130 // earliest store node which is *used* and replaced by the wide store. 8131 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum) 8132 EarliestNodeUsed = i; 8133 } 8134 8135 // The earliest Node in the DAG. 8136 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode; 8137 SDLoc DL(StoreNodes[0].MemNode); 8138 8139 SDValue StoredVal; 8140 if (UseVector) { 8141 // Find a legal type for the vector store. 8142 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem); 8143 assert(TLI.isTypeLegal(Ty) && "Illegal vector store"); 8144 StoredVal = DAG.getConstant(0, Ty); 8145 } else { 8146 unsigned StoreBW = NumElem * ElementSizeBytes * 8; 8147 APInt StoreInt(StoreBW, 0); 8148 8149 // Construct a single integer constant which is made of the smaller 8150 // constant inputs. 8151 bool IsLE = TLI.isLittleEndian(); 8152 for (unsigned i = 0; i < NumElem ; ++i) { 8153 unsigned Idx = IsLE ?(NumElem - 1 - i) : i; 8154 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode); 8155 SDValue Val = St->getValue(); 8156 StoreInt<<=ElementSizeBytes*8; 8157 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) { 8158 StoreInt|=C->getAPIntValue().zext(StoreBW); 8159 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) { 8160 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW); 8161 } else { 8162 assert(false && "Invalid constant element type"); 8163 } 8164 } 8165 8166 // Create the new Load and Store operations. 8167 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW); 8168 StoredVal = DAG.getConstant(StoreInt, StoreTy); 8169 } 8170 8171 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal, 8172 FirstInChain->getBasePtr(), 8173 FirstInChain->getPointerInfo(), 8174 false, false, 8175 FirstInChain->getAlignment()); 8176 8177 // Replace the first store with the new store 8178 CombineTo(EarliestOp, NewStore); 8179 // Erase all other stores. 8180 for (unsigned i = 0; i < NumElem ; ++i) { 8181 if (StoreNodes[i].MemNode == EarliestOp) 8182 continue; 8183 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); 8184 // ReplaceAllUsesWith will replace all uses that existed when it was 8185 // called, but graph optimizations may cause new ones to appear. For 8186 // example, the case in pr14333 looks like 8187 // 8188 // St's chain -> St -> another store -> X 8189 // 8190 // And the only difference from St to the other store is the chain. 8191 // When we change it's chain to be St's chain they become identical, 8192 // get CSEed and the net result is that X is now a use of St. 8193 // Since we know that St is redundant, just iterate. 8194 while (!St->use_empty()) 8195 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain()); 8196 removeFromWorkList(St); 8197 DAG.DeleteNode(St); 8198 } 8199 8200 return true; 8201 } 8202 8203 // Below we handle the case of multiple consecutive stores that 8204 // come from multiple consecutive loads. We merge them into a single 8205 // wide load and a single wide store. 8206 8207 // Look for load nodes which are used by the stored values. 8208 SmallVector<MemOpLink, 8> LoadNodes; 8209 8210 // Find acceptable loads. Loads need to have the same chain (token factor), 8211 // must not be zext, volatile, indexed, and they must be consecutive. 8212 BaseIndexOffset LdBasePtr; 8213 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) { 8214 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); 8215 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue()); 8216 if (!Ld) break; 8217 8218 // Loads must only have one use. 8219 if (!Ld->hasNUsesOfValue(1, 0)) 8220 break; 8221 8222 // Check that the alignment is the same as the stores. 8223 if (Ld->getAlignment() != St->getAlignment()) 8224 break; 8225 8226 // The memory operands must not be volatile. 8227 if (Ld->isVolatile() || Ld->isIndexed()) 8228 break; 8229 8230 // We do not accept ext loads. 8231 if (Ld->getExtensionType() != ISD::NON_EXTLOAD) 8232 break; 8233 8234 // The stored memory type must be the same. 8235 if (Ld->getMemoryVT() != MemVT) 8236 break; 8237 8238 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr()); 8239 // If this is not the first ptr that we check. 8240 if (LdBasePtr.Base.getNode()) { 8241 // The base ptr must be the same. 8242 if (!LdPtr.equalBaseIndex(LdBasePtr)) 8243 break; 8244 } else { 8245 // Check that all other base pointers are the same as this one. 8246 LdBasePtr = LdPtr; 8247 } 8248 8249 // We found a potential memory operand to merge. 8250 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0)); 8251 } 8252 8253 if (LoadNodes.size() < 2) 8254 return false; 8255 8256 // Scan the memory operations on the chain and find the first non-consecutive 8257 // load memory address. These variables hold the index in the store node 8258 // array. 8259 unsigned LastConsecutiveLoad = 0; 8260 // This variable refers to the size and not index in the array. 8261 unsigned LastLegalVectorType = 0; 8262 unsigned LastLegalIntegerType = 0; 8263 StartAddress = LoadNodes[0].OffsetFromBase; 8264 SDValue FirstChain = LoadNodes[0].MemNode->getChain(); 8265 for (unsigned i = 1; i < LoadNodes.size(); ++i) { 8266 // All loads much share the same chain. 8267 if (LoadNodes[i].MemNode->getChain() != FirstChain) 8268 break; 8269 8270 int64_t CurrAddress = LoadNodes[i].OffsetFromBase; 8271 if (CurrAddress - StartAddress != (ElementSizeBytes * i)) 8272 break; 8273 LastConsecutiveLoad = i; 8274 8275 // Find a legal type for the vector store. 8276 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1); 8277 if (TLI.isTypeLegal(StoreTy)) 8278 LastLegalVectorType = i + 1; 8279 8280 // Find a legal type for the integer store. 8281 unsigned StoreBW = (i+1) * ElementSizeBytes * 8; 8282 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW); 8283 if (TLI.isTypeLegal(StoreTy)) 8284 LastLegalIntegerType = i + 1; 8285 // Or check whether a truncstore and extload is legal. 8286 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) == 8287 TargetLowering::TypePromoteInteger) { 8288 EVT LegalizedStoredValueTy = 8289 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy); 8290 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) && 8291 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) && 8292 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) && 8293 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy)) 8294 LastLegalIntegerType = i+1; 8295 } 8296 } 8297 8298 // Only use vector types if the vector type is larger than the integer type. 8299 // If they are the same, use integers. 8300 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors; 8301 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType); 8302 8303 // We add +1 here because the LastXXX variables refer to location while 8304 // the NumElem refers to array/index size. 8305 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1; 8306 NumElem = std::min(LastLegalType, NumElem); 8307 8308 if (NumElem < 2) 8309 return false; 8310 8311 // The earliest Node in the DAG. 8312 unsigned EarliestNodeUsed = 0; 8313 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode; 8314 for (unsigned i=1; i<NumElem; ++i) { 8315 // Find a chain for the new wide-store operand. Notice that some 8316 // of the store nodes that we found may not be selected for inclusion 8317 // in the wide store. The chain we use needs to be the chain of the 8318 // earliest store node which is *used* and replaced by the wide store. 8319 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum) 8320 EarliestNodeUsed = i; 8321 } 8322 8323 // Find if it is better to use vectors or integers to load and store 8324 // to memory. 8325 EVT JointMemOpVT; 8326 if (UseVectorTy) { 8327 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem); 8328 } else { 8329 unsigned StoreBW = NumElem * ElementSizeBytes * 8; 8330 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW); 8331 } 8332 8333 SDLoc LoadDL(LoadNodes[0].MemNode); 8334 SDLoc StoreDL(StoreNodes[0].MemNode); 8335 8336 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode); 8337 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL, 8338 FirstLoad->getChain(), 8339 FirstLoad->getBasePtr(), 8340 FirstLoad->getPointerInfo(), 8341 false, false, false, 8342 FirstLoad->getAlignment()); 8343 8344 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad, 8345 FirstInChain->getBasePtr(), 8346 FirstInChain->getPointerInfo(), false, false, 8347 FirstInChain->getAlignment()); 8348 8349 // Replace one of the loads with the new load. 8350 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode); 8351 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), 8352 SDValue(NewLoad.getNode(), 1)); 8353 8354 // Remove the rest of the load chains. 8355 for (unsigned i = 1; i < NumElem ; ++i) { 8356 // Replace all chain users of the old load nodes with the chain of the new 8357 // load node. 8358 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode); 8359 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain()); 8360 } 8361 8362 // Replace the first store with the new store. 8363 CombineTo(EarliestOp, NewStore); 8364 // Erase all other stores. 8365 for (unsigned i = 0; i < NumElem ; ++i) { 8366 // Remove all Store nodes. 8367 if (StoreNodes[i].MemNode == EarliestOp) 8368 continue; 8369 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); 8370 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain()); 8371 removeFromWorkList(St); 8372 DAG.DeleteNode(St); 8373 } 8374 8375 return true; 8376} 8377 8378SDValue DAGCombiner::visitSTORE(SDNode *N) { 8379 StoreSDNode *ST = cast<StoreSDNode>(N); 8380 SDValue Chain = ST->getChain(); 8381 SDValue Value = ST->getValue(); 8382 SDValue Ptr = ST->getBasePtr(); 8383 8384 // If this is a store of a bit convert, store the input value if the 8385 // resultant store does not need a higher alignment than the original. 8386 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() && 8387 ST->isUnindexed()) { 8388 unsigned OrigAlign = ST->getAlignment(); 8389 EVT SVT = Value.getOperand(0).getValueType(); 8390 unsigned Align = TLI.getDataLayout()-> 8391 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext())); 8392 if (Align <= OrigAlign && 8393 ((!LegalOperations && !ST->isVolatile()) || 8394 TLI.isOperationLegalOrCustom(ISD::STORE, SVT))) 8395 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0), 8396 Ptr, ST->getPointerInfo(), ST->isVolatile(), 8397 ST->isNonTemporal(), OrigAlign); 8398 } 8399 8400 // Turn 'store undef, Ptr' -> nothing. 8401 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed()) 8402 return Chain; 8403 8404 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr' 8405 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) { 8406 // NOTE: If the original store is volatile, this transform must not increase 8407 // the number of stores. For example, on x86-32 an f64 can be stored in one 8408 // processor operation but an i64 (which is not legal) requires two. So the 8409 // transform should not be done in this case. 8410 if (Value.getOpcode() != ISD::TargetConstantFP) { 8411 SDValue Tmp; 8412 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) { 8413 default: llvm_unreachable("Unknown FP type"); 8414 case MVT::f16: // We don't do this for these yet. 8415 case MVT::f80: 8416 case MVT::f128: 8417 case MVT::ppcf128: 8418 break; 8419 case MVT::f32: 8420 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) || 8421 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) { 8422 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF(). 8423 bitcastToAPInt().getZExtValue(), MVT::i32); 8424 return DAG.getStore(Chain, SDLoc(N), Tmp, 8425 Ptr, ST->getPointerInfo(), ST->isVolatile(), 8426 ST->isNonTemporal(), ST->getAlignment()); 8427 } 8428 break; 8429 case MVT::f64: 8430 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations && 8431 !ST->isVolatile()) || 8432 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) { 8433 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt(). 8434 getZExtValue(), MVT::i64); 8435 return DAG.getStore(Chain, SDLoc(N), Tmp, 8436 Ptr, ST->getPointerInfo(), ST->isVolatile(), 8437 ST->isNonTemporal(), ST->getAlignment()); 8438 } 8439 8440 if (!ST->isVolatile() && 8441 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) { 8442 // Many FP stores are not made apparent until after legalize, e.g. for 8443 // argument passing. Since this is so common, custom legalize the 8444 // 64-bit integer store into two 32-bit stores. 8445 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue(); 8446 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32); 8447 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32); 8448 if (TLI.isBigEndian()) std::swap(Lo, Hi); 8449 8450 unsigned Alignment = ST->getAlignment(); 8451 bool isVolatile = ST->isVolatile(); 8452 bool isNonTemporal = ST->isNonTemporal(); 8453 8454 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo, 8455 Ptr, ST->getPointerInfo(), 8456 isVolatile, isNonTemporal, 8457 ST->getAlignment()); 8458 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr, 8459 DAG.getConstant(4, Ptr.getValueType())); 8460 Alignment = MinAlign(Alignment, 4U); 8461 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi, 8462 Ptr, ST->getPointerInfo().getWithOffset(4), 8463 isVolatile, isNonTemporal, 8464 Alignment); 8465 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, 8466 St0, St1); 8467 } 8468 8469 break; 8470 } 8471 } 8472 } 8473 8474 // Try to infer better alignment information than the store already has. 8475 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) { 8476 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) { 8477 if (Align > ST->getAlignment()) 8478 return DAG.getTruncStore(Chain, SDLoc(N), Value, 8479 Ptr, ST->getPointerInfo(), ST->getMemoryVT(), 8480 ST->isVolatile(), ST->isNonTemporal(), Align); 8481 } 8482 } 8483 8484 // Try transforming a pair floating point load / store ops to integer 8485 // load / store ops. 8486 SDValue NewST = TransformFPLoadStorePair(N); 8487 if (NewST.getNode()) 8488 return NewST; 8489 8490 if (CombinerAA) { 8491 // Walk up chain skipping non-aliasing memory nodes. 8492 SDValue BetterChain = FindBetterChain(N, Chain); 8493 8494 // If there is a better chain. 8495 if (Chain != BetterChain) { 8496 SDValue ReplStore; 8497 8498 // Replace the chain to avoid dependency. 8499 if (ST->isTruncatingStore()) { 8500 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr, 8501 ST->getPointerInfo(), 8502 ST->getMemoryVT(), ST->isVolatile(), 8503 ST->isNonTemporal(), ST->getAlignment()); 8504 } else { 8505 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr, 8506 ST->getPointerInfo(), 8507 ST->isVolatile(), ST->isNonTemporal(), 8508 ST->getAlignment()); 8509 } 8510 8511 // Create token to keep both nodes around. 8512 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N), 8513 MVT::Other, Chain, ReplStore); 8514 8515 // Make sure the new and old chains are cleaned up. 8516 AddToWorkList(Token.getNode()); 8517 8518 // Don't add users to work list. 8519 return CombineTo(N, Token, false); 8520 } 8521 } 8522 8523 // Try transforming N to an indexed store. 8524 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) 8525 return SDValue(N, 0); 8526 8527 // FIXME: is there such a thing as a truncating indexed store? 8528 if (ST->isTruncatingStore() && ST->isUnindexed() && 8529 Value.getValueType().isInteger()) { 8530 // See if we can simplify the input to this truncstore with knowledge that 8531 // only the low bits are being used. For example: 8532 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8" 8533 SDValue Shorter = 8534 GetDemandedBits(Value, 8535 APInt::getLowBitsSet( 8536 Value.getValueType().getScalarType().getSizeInBits(), 8537 ST->getMemoryVT().getScalarType().getSizeInBits())); 8538 AddToWorkList(Value.getNode()); 8539 if (Shorter.getNode()) 8540 return DAG.getTruncStore(Chain, SDLoc(N), Shorter, 8541 Ptr, ST->getPointerInfo(), ST->getMemoryVT(), 8542 ST->isVolatile(), ST->isNonTemporal(), 8543 ST->getAlignment()); 8544 8545 // Otherwise, see if we can simplify the operation with 8546 // SimplifyDemandedBits, which only works if the value has a single use. 8547 if (SimplifyDemandedBits(Value, 8548 APInt::getLowBitsSet( 8549 Value.getValueType().getScalarType().getSizeInBits(), 8550 ST->getMemoryVT().getScalarType().getSizeInBits()))) 8551 return SDValue(N, 0); 8552 } 8553 8554 // If this is a load followed by a store to the same location, then the store 8555 // is dead/noop. 8556 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) { 8557 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() && 8558 ST->isUnindexed() && !ST->isVolatile() && 8559 // There can't be any side effects between the load and store, such as 8560 // a call or store. 8561 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) { 8562 // The store is dead, remove it. 8563 return Chain; 8564 } 8565 } 8566 8567 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a 8568 // truncating store. We can do this even if this is already a truncstore. 8569 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE) 8570 && Value.getNode()->hasOneUse() && ST->isUnindexed() && 8571 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(), 8572 ST->getMemoryVT())) { 8573 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0), 8574 Ptr, ST->getPointerInfo(), ST->getMemoryVT(), 8575 ST->isVolatile(), ST->isNonTemporal(), 8576 ST->getAlignment()); 8577 } 8578 8579 // Only perform this optimization before the types are legal, because we 8580 // don't want to perform this optimization on every DAGCombine invocation. 8581 if (!LegalTypes) { 8582 bool EverChanged = false; 8583 8584 do { 8585 // There can be multiple store sequences on the same chain. 8586 // Keep trying to merge store sequences until we are unable to do so 8587 // or until we merge the last store on the chain. 8588 bool Changed = MergeConsecutiveStores(ST); 8589 EverChanged |= Changed; 8590 if (!Changed) break; 8591 } while (ST->getOpcode() != ISD::DELETED_NODE); 8592 8593 if (EverChanged) 8594 return SDValue(N, 0); 8595 } 8596 8597 return ReduceLoadOpStoreWidth(N); 8598} 8599 8600SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) { 8601 SDValue InVec = N->getOperand(0); 8602 SDValue InVal = N->getOperand(1); 8603 SDValue EltNo = N->getOperand(2); 8604 SDLoc dl(N); 8605 8606 // If the inserted element is an UNDEF, just use the input vector. 8607 if (InVal.getOpcode() == ISD::UNDEF) 8608 return InVec; 8609 8610 EVT VT = InVec.getValueType(); 8611 8612 // If we can't generate a legal BUILD_VECTOR, exit 8613 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) 8614 return SDValue(); 8615 8616 // Check that we know which element is being inserted 8617 if (!isa<ConstantSDNode>(EltNo)) 8618 return SDValue(); 8619 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); 8620 8621 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially 8622 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the 8623 // vector elements. 8624 SmallVector<SDValue, 8> Ops; 8625 if (InVec.getOpcode() == ISD::BUILD_VECTOR) { 8626 Ops.append(InVec.getNode()->op_begin(), 8627 InVec.getNode()->op_end()); 8628 } else if (InVec.getOpcode() == ISD::UNDEF) { 8629 unsigned NElts = VT.getVectorNumElements(); 8630 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType())); 8631 } else { 8632 return SDValue(); 8633 } 8634 8635 // Insert the element 8636 if (Elt < Ops.size()) { 8637 // All the operands of BUILD_VECTOR must have the same type; 8638 // we enforce that here. 8639 EVT OpVT = Ops[0].getValueType(); 8640 if (InVal.getValueType() != OpVT) 8641 InVal = OpVT.bitsGT(InVal.getValueType()) ? 8642 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) : 8643 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal); 8644 Ops[Elt] = InVal; 8645 } 8646 8647 // Return the new vector 8648 return DAG.getNode(ISD::BUILD_VECTOR, dl, 8649 VT, &Ops[0], Ops.size()); 8650} 8651 8652SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) { 8653 // (vextract (scalar_to_vector val, 0) -> val 8654 SDValue InVec = N->getOperand(0); 8655 EVT VT = InVec.getValueType(); 8656 EVT NVT = N->getValueType(0); 8657 8658 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) { 8659 // Check if the result type doesn't match the inserted element type. A 8660 // SCALAR_TO_VECTOR may truncate the inserted element and the 8661 // EXTRACT_VECTOR_ELT may widen the extracted vector. 8662 SDValue InOp = InVec.getOperand(0); 8663 if (InOp.getValueType() != NVT) { 8664 assert(InOp.getValueType().isInteger() && NVT.isInteger()); 8665 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT); 8666 } 8667 return InOp; 8668 } 8669 8670 SDValue EltNo = N->getOperand(1); 8671 bool ConstEltNo = isa<ConstantSDNode>(EltNo); 8672 8673 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT. 8674 // We only perform this optimization before the op legalization phase because 8675 // we may introduce new vector instructions which are not backed by TD 8676 // patterns. For example on AVX, extracting elements from a wide vector 8677 // without using extract_subvector. 8678 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE 8679 && ConstEltNo && !LegalOperations) { 8680 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); 8681 int NumElem = VT.getVectorNumElements(); 8682 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec); 8683 // Find the new index to extract from. 8684 int OrigElt = SVOp->getMaskElt(Elt); 8685 8686 // Extracting an undef index is undef. 8687 if (OrigElt == -1) 8688 return DAG.getUNDEF(NVT); 8689 8690 // Select the right vector half to extract from. 8691 if (OrigElt < NumElem) { 8692 InVec = InVec->getOperand(0); 8693 } else { 8694 InVec = InVec->getOperand(1); 8695 OrigElt -= NumElem; 8696 } 8697 8698 EVT IndexTy = N->getOperand(1).getValueType(); 8699 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT, 8700 InVec, DAG.getConstant(OrigElt, IndexTy)); 8701 } 8702 8703 // Perform only after legalization to ensure build_vector / vector_shuffle 8704 // optimizations have already been done. 8705 if (!LegalOperations) return SDValue(); 8706 8707 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size) 8708 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size) 8709 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr) 8710 8711 if (ConstEltNo) { 8712 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); 8713 bool NewLoad = false; 8714 bool BCNumEltsChanged = false; 8715 EVT ExtVT = VT.getVectorElementType(); 8716 EVT LVT = ExtVT; 8717 8718 // If the result of load has to be truncated, then it's not necessarily 8719 // profitable. 8720 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT)) 8721 return SDValue(); 8722 8723 if (InVec.getOpcode() == ISD::BITCAST) { 8724 // Don't duplicate a load with other uses. 8725 if (!InVec.hasOneUse()) 8726 return SDValue(); 8727 8728 EVT BCVT = InVec.getOperand(0).getValueType(); 8729 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType())) 8730 return SDValue(); 8731 if (VT.getVectorNumElements() != BCVT.getVectorNumElements()) 8732 BCNumEltsChanged = true; 8733 InVec = InVec.getOperand(0); 8734 ExtVT = BCVT.getVectorElementType(); 8735 NewLoad = true; 8736 } 8737 8738 LoadSDNode *LN0 = NULL; 8739 const ShuffleVectorSDNode *SVN = NULL; 8740 if (ISD::isNormalLoad(InVec.getNode())) { 8741 LN0 = cast<LoadSDNode>(InVec); 8742 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR && 8743 InVec.getOperand(0).getValueType() == ExtVT && 8744 ISD::isNormalLoad(InVec.getOperand(0).getNode())) { 8745 // Don't duplicate a load with other uses. 8746 if (!InVec.hasOneUse()) 8747 return SDValue(); 8748 8749 LN0 = cast<LoadSDNode>(InVec.getOperand(0)); 8750 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) { 8751 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1) 8752 // => 8753 // (load $addr+1*size) 8754 8755 // Don't duplicate a load with other uses. 8756 if (!InVec.hasOneUse()) 8757 return SDValue(); 8758 8759 // If the bit convert changed the number of elements, it is unsafe 8760 // to examine the mask. 8761 if (BCNumEltsChanged) 8762 return SDValue(); 8763 8764 // Select the input vector, guarding against out of range extract vector. 8765 unsigned NumElems = VT.getVectorNumElements(); 8766 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt); 8767 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1); 8768 8769 if (InVec.getOpcode() == ISD::BITCAST) { 8770 // Don't duplicate a load with other uses. 8771 if (!InVec.hasOneUse()) 8772 return SDValue(); 8773 8774 InVec = InVec.getOperand(0); 8775 } 8776 if (ISD::isNormalLoad(InVec.getNode())) { 8777 LN0 = cast<LoadSDNode>(InVec); 8778 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems; 8779 } 8780 } 8781 8782 // Make sure we found a non-volatile load and the extractelement is 8783 // the only use. 8784 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile()) 8785 return SDValue(); 8786 8787 // If Idx was -1 above, Elt is going to be -1, so just return undef. 8788 if (Elt == -1) 8789 return DAG.getUNDEF(LVT); 8790 8791 unsigned Align = LN0->getAlignment(); 8792 if (NewLoad) { 8793 // Check the resultant load doesn't need a higher alignment than the 8794 // original load. 8795 unsigned NewAlign = 8796 TLI.getDataLayout() 8797 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext())); 8798 8799 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT)) 8800 return SDValue(); 8801 8802 Align = NewAlign; 8803 } 8804 8805 SDValue NewPtr = LN0->getBasePtr(); 8806 unsigned PtrOff = 0; 8807 8808 if (Elt) { 8809 PtrOff = LVT.getSizeInBits() * Elt / 8; 8810 EVT PtrType = NewPtr.getValueType(); 8811 if (TLI.isBigEndian()) 8812 PtrOff = VT.getSizeInBits() / 8 - PtrOff; 8813 NewPtr = DAG.getNode(ISD::ADD, SDLoc(N), PtrType, NewPtr, 8814 DAG.getConstant(PtrOff, PtrType)); 8815 } 8816 8817 // The replacement we need to do here is a little tricky: we need to 8818 // replace an extractelement of a load with a load. 8819 // Use ReplaceAllUsesOfValuesWith to do the replacement. 8820 // Note that this replacement assumes that the extractvalue is the only 8821 // use of the load; that's okay because we don't want to perform this 8822 // transformation in other cases anyway. 8823 SDValue Load; 8824 SDValue Chain; 8825 if (NVT.bitsGT(LVT)) { 8826 // If the result type of vextract is wider than the load, then issue an 8827 // extending load instead. 8828 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT) 8829 ? ISD::ZEXTLOAD : ISD::EXTLOAD; 8830 Load = DAG.getExtLoad(ExtType, SDLoc(N), NVT, LN0->getChain(), 8831 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff), 8832 LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align); 8833 Chain = Load.getValue(1); 8834 } else { 8835 Load = DAG.getLoad(LVT, SDLoc(N), LN0->getChain(), NewPtr, 8836 LN0->getPointerInfo().getWithOffset(PtrOff), 8837 LN0->isVolatile(), LN0->isNonTemporal(), 8838 LN0->isInvariant(), Align); 8839 Chain = Load.getValue(1); 8840 if (NVT.bitsLT(LVT)) 8841 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, Load); 8842 else 8843 Load = DAG.getNode(ISD::BITCAST, SDLoc(N), NVT, Load); 8844 } 8845 WorkListRemover DeadNodes(*this); 8846 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) }; 8847 SDValue To[] = { Load, Chain }; 8848 DAG.ReplaceAllUsesOfValuesWith(From, To, 2); 8849 // Since we're explcitly calling ReplaceAllUses, add the new node to the 8850 // worklist explicitly as well. 8851 AddToWorkList(Load.getNode()); 8852 AddUsersToWorkList(Load.getNode()); // Add users too 8853 // Make sure to revisit this node to clean it up; it will usually be dead. 8854 AddToWorkList(N); 8855 return SDValue(N, 0); 8856 } 8857 8858 return SDValue(); 8859} 8860 8861// Simplify (build_vec (ext )) to (bitcast (build_vec )) 8862SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) { 8863 // We perform this optimization post type-legalization because 8864 // the type-legalizer often scalarizes integer-promoted vectors. 8865 // Performing this optimization before may create bit-casts which 8866 // will be type-legalized to complex code sequences. 8867 // We perform this optimization only before the operation legalizer because we 8868 // may introduce illegal operations. 8869 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes) 8870 return SDValue(); 8871 8872 unsigned NumInScalars = N->getNumOperands(); 8873 SDLoc dl(N); 8874 EVT VT = N->getValueType(0); 8875 8876 // Check to see if this is a BUILD_VECTOR of a bunch of values 8877 // which come from any_extend or zero_extend nodes. If so, we can create 8878 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR 8879 // optimizations. We do not handle sign-extend because we can't fill the sign 8880 // using shuffles. 8881 EVT SourceType = MVT::Other; 8882 bool AllAnyExt = true; 8883 8884 for (unsigned i = 0; i != NumInScalars; ++i) { 8885 SDValue In = N->getOperand(i); 8886 // Ignore undef inputs. 8887 if (In.getOpcode() == ISD::UNDEF) continue; 8888 8889 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND; 8890 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND; 8891 8892 // Abort if the element is not an extension. 8893 if (!ZeroExt && !AnyExt) { 8894 SourceType = MVT::Other; 8895 break; 8896 } 8897 8898 // The input is a ZeroExt or AnyExt. Check the original type. 8899 EVT InTy = In.getOperand(0).getValueType(); 8900 8901 // Check that all of the widened source types are the same. 8902 if (SourceType == MVT::Other) 8903 // First time. 8904 SourceType = InTy; 8905 else if (InTy != SourceType) { 8906 // Multiple income types. Abort. 8907 SourceType = MVT::Other; 8908 break; 8909 } 8910 8911 // Check if all of the extends are ANY_EXTENDs. 8912 AllAnyExt &= AnyExt; 8913 } 8914 8915 // In order to have valid types, all of the inputs must be extended from the 8916 // same source type and all of the inputs must be any or zero extend. 8917 // Scalar sizes must be a power of two. 8918 EVT OutScalarTy = VT.getScalarType(); 8919 bool ValidTypes = SourceType != MVT::Other && 8920 isPowerOf2_32(OutScalarTy.getSizeInBits()) && 8921 isPowerOf2_32(SourceType.getSizeInBits()); 8922 8923 // Create a new simpler BUILD_VECTOR sequence which other optimizations can 8924 // turn into a single shuffle instruction. 8925 if (!ValidTypes) 8926 return SDValue(); 8927 8928 bool isLE = TLI.isLittleEndian(); 8929 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits(); 8930 assert(ElemRatio > 1 && "Invalid element size ratio"); 8931 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType): 8932 DAG.getConstant(0, SourceType); 8933 8934 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements(); 8935 SmallVector<SDValue, 8> Ops(NewBVElems, Filler); 8936 8937 // Populate the new build_vector 8938 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 8939 SDValue Cast = N->getOperand(i); 8940 assert((Cast.getOpcode() == ISD::ANY_EXTEND || 8941 Cast.getOpcode() == ISD::ZERO_EXTEND || 8942 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode"); 8943 SDValue In; 8944 if (Cast.getOpcode() == ISD::UNDEF) 8945 In = DAG.getUNDEF(SourceType); 8946 else 8947 In = Cast->getOperand(0); 8948 unsigned Index = isLE ? (i * ElemRatio) : 8949 (i * ElemRatio + (ElemRatio - 1)); 8950 8951 assert(Index < Ops.size() && "Invalid index"); 8952 Ops[Index] = In; 8953 } 8954 8955 // The type of the new BUILD_VECTOR node. 8956 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems); 8957 assert(VecVT.getSizeInBits() == VT.getSizeInBits() && 8958 "Invalid vector size"); 8959 // Check if the new vector type is legal. 8960 if (!isTypeLegal(VecVT)) return SDValue(); 8961 8962 // Make the new BUILD_VECTOR. 8963 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], Ops.size()); 8964 8965 // The new BUILD_VECTOR node has the potential to be further optimized. 8966 AddToWorkList(BV.getNode()); 8967 // Bitcast to the desired type. 8968 return DAG.getNode(ISD::BITCAST, dl, VT, BV); 8969} 8970 8971SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) { 8972 EVT VT = N->getValueType(0); 8973 8974 unsigned NumInScalars = N->getNumOperands(); 8975 SDLoc dl(N); 8976 8977 EVT SrcVT = MVT::Other; 8978 unsigned Opcode = ISD::DELETED_NODE; 8979 unsigned NumDefs = 0; 8980 8981 for (unsigned i = 0; i != NumInScalars; ++i) { 8982 SDValue In = N->getOperand(i); 8983 unsigned Opc = In.getOpcode(); 8984 8985 if (Opc == ISD::UNDEF) 8986 continue; 8987 8988 // If all scalar values are floats and converted from integers. 8989 if (Opcode == ISD::DELETED_NODE && 8990 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) { 8991 Opcode = Opc; 8992 } 8993 8994 if (Opc != Opcode) 8995 return SDValue(); 8996 8997 EVT InVT = In.getOperand(0).getValueType(); 8998 8999 // If all scalar values are typed differently, bail out. It's chosen to 9000 // simplify BUILD_VECTOR of integer types. 9001 if (SrcVT == MVT::Other) 9002 SrcVT = InVT; 9003 if (SrcVT != InVT) 9004 return SDValue(); 9005 NumDefs++; 9006 } 9007 9008 // If the vector has just one element defined, it's not worth to fold it into 9009 // a vectorized one. 9010 if (NumDefs < 2) 9011 return SDValue(); 9012 9013 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP) 9014 && "Should only handle conversion from integer to float."); 9015 assert(SrcVT != MVT::Other && "Cannot determine source type!"); 9016 9017 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars); 9018 9019 if (!TLI.isOperationLegalOrCustom(Opcode, NVT)) 9020 return SDValue(); 9021 9022 SmallVector<SDValue, 8> Opnds; 9023 for (unsigned i = 0; i != NumInScalars; ++i) { 9024 SDValue In = N->getOperand(i); 9025 9026 if (In.getOpcode() == ISD::UNDEF) 9027 Opnds.push_back(DAG.getUNDEF(SrcVT)); 9028 else 9029 Opnds.push_back(In.getOperand(0)); 9030 } 9031 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, 9032 &Opnds[0], Opnds.size()); 9033 AddToWorkList(BV.getNode()); 9034 9035 return DAG.getNode(Opcode, dl, VT, BV); 9036} 9037 9038SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) { 9039 unsigned NumInScalars = N->getNumOperands(); 9040 SDLoc dl(N); 9041 EVT VT = N->getValueType(0); 9042 9043 // A vector built entirely of undefs is undef. 9044 if (ISD::allOperandsUndef(N)) 9045 return DAG.getUNDEF(VT); 9046 9047 SDValue V = reduceBuildVecExtToExtBuildVec(N); 9048 if (V.getNode()) 9049 return V; 9050 9051 V = reduceBuildVecConvertToConvertBuildVec(N); 9052 if (V.getNode()) 9053 return V; 9054 9055 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT 9056 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from 9057 // at most two distinct vectors, turn this into a shuffle node. 9058 9059 // May only combine to shuffle after legalize if shuffle is legal. 9060 if (LegalOperations && 9061 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT)) 9062 return SDValue(); 9063 9064 SDValue VecIn1, VecIn2; 9065 for (unsigned i = 0; i != NumInScalars; ++i) { 9066 // Ignore undef inputs. 9067 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue; 9068 9069 // If this input is something other than a EXTRACT_VECTOR_ELT with a 9070 // constant index, bail out. 9071 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT || 9072 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) { 9073 VecIn1 = VecIn2 = SDValue(0, 0); 9074 break; 9075 } 9076 9077 // We allow up to two distinct input vectors. 9078 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0); 9079 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2) 9080 continue; 9081 9082 if (VecIn1.getNode() == 0) { 9083 VecIn1 = ExtractedFromVec; 9084 } else if (VecIn2.getNode() == 0) { 9085 VecIn2 = ExtractedFromVec; 9086 } else { 9087 // Too many inputs. 9088 VecIn1 = VecIn2 = SDValue(0, 0); 9089 break; 9090 } 9091 } 9092 9093 // If everything is good, we can make a shuffle operation. 9094 if (VecIn1.getNode()) { 9095 SmallVector<int, 8> Mask; 9096 for (unsigned i = 0; i != NumInScalars; ++i) { 9097 if (N->getOperand(i).getOpcode() == ISD::UNDEF) { 9098 Mask.push_back(-1); 9099 continue; 9100 } 9101 9102 // If extracting from the first vector, just use the index directly. 9103 SDValue Extract = N->getOperand(i); 9104 SDValue ExtVal = Extract.getOperand(1); 9105 if (Extract.getOperand(0) == VecIn1) { 9106 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue(); 9107 if (ExtIndex > VT.getVectorNumElements()) 9108 return SDValue(); 9109 9110 Mask.push_back(ExtIndex); 9111 continue; 9112 } 9113 9114 // Otherwise, use InIdx + VecSize 9115 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue(); 9116 Mask.push_back(Idx+NumInScalars); 9117 } 9118 9119 // We can't generate a shuffle node with mismatched input and output types. 9120 // Attempt to transform a single input vector to the correct type. 9121 if ((VT != VecIn1.getValueType())) { 9122 // We don't support shuffeling between TWO values of different types. 9123 if (VecIn2.getNode() != 0) 9124 return SDValue(); 9125 9126 // We only support widening of vectors which are half the size of the 9127 // output registers. For example XMM->YMM widening on X86 with AVX. 9128 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits()) 9129 return SDValue(); 9130 9131 // If the input vector type has a different base type to the output 9132 // vector type, bail out. 9133 if (VecIn1.getValueType().getVectorElementType() != 9134 VT.getVectorElementType()) 9135 return SDValue(); 9136 9137 // Widen the input vector by adding undef values. 9138 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, 9139 VecIn1, DAG.getUNDEF(VecIn1.getValueType())); 9140 } 9141 9142 // If VecIn2 is unused then change it to undef. 9143 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT); 9144 9145 // Check that we were able to transform all incoming values to the same 9146 // type. 9147 if (VecIn2.getValueType() != VecIn1.getValueType() || 9148 VecIn1.getValueType() != VT) 9149 return SDValue(); 9150 9151 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes. 9152 if (!isTypeLegal(VT)) 9153 return SDValue(); 9154 9155 // Return the new VECTOR_SHUFFLE node. 9156 SDValue Ops[2]; 9157 Ops[0] = VecIn1; 9158 Ops[1] = VecIn2; 9159 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]); 9160 } 9161 9162 return SDValue(); 9163} 9164 9165SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) { 9166 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of 9167 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector 9168 // inputs come from at most two distinct vectors, turn this into a shuffle 9169 // node. 9170 9171 // If we only have one input vector, we don't need to do any concatenation. 9172 if (N->getNumOperands() == 1) 9173 return N->getOperand(0); 9174 9175 // Check if all of the operands are undefs. 9176 if (ISD::allOperandsUndef(N)) 9177 return DAG.getUNDEF(N->getValueType(0)); 9178 9179 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR 9180 // nodes often generate nop CONCAT_VECTOR nodes. 9181 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that 9182 // place the incoming vectors at the exact same location. 9183 SDValue SingleSource = SDValue(); 9184 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements(); 9185 9186 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 9187 SDValue Op = N->getOperand(i); 9188 9189 if (Op.getOpcode() == ISD::UNDEF) 9190 continue; 9191 9192 // Check if this is the identity extract: 9193 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR) 9194 return SDValue(); 9195 9196 // Find the single incoming vector for the extract_subvector. 9197 if (SingleSource.getNode()) { 9198 if (Op.getOperand(0) != SingleSource) 9199 return SDValue(); 9200 } else { 9201 SingleSource = Op.getOperand(0); 9202 9203 // Check the source type is the same as the type of the result. 9204 // If not, this concat may extend the vector, so we can not 9205 // optimize it away. 9206 if (SingleSource.getValueType() != N->getValueType(0)) 9207 return SDValue(); 9208 } 9209 9210 unsigned IdentityIndex = i * PartNumElem; 9211 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 9212 // The extract index must be constant. 9213 if (!CS) 9214 return SDValue(); 9215 9216 // Check that we are reading from the identity index. 9217 if (CS->getZExtValue() != IdentityIndex) 9218 return SDValue(); 9219 } 9220 9221 if (SingleSource.getNode()) 9222 return SingleSource; 9223 9224 return SDValue(); 9225} 9226 9227SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) { 9228 EVT NVT = N->getValueType(0); 9229 SDValue V = N->getOperand(0); 9230 9231 if (V->getOpcode() == ISD::CONCAT_VECTORS) { 9232 // Combine: 9233 // (extract_subvec (concat V1, V2, ...), i) 9234 // Into: 9235 // Vi if possible 9236 // Only operand 0 is checked as 'concat' assumes all inputs of the same type. 9237 if (V->getOperand(0).getValueType() != NVT) 9238 return SDValue(); 9239 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 9240 unsigned NumElems = NVT.getVectorNumElements(); 9241 assert((Idx % NumElems) == 0 && 9242 "IDX in concat is not a multiple of the result vector length."); 9243 return V->getOperand(Idx / NumElems); 9244 } 9245 9246 // Skip bitcasting 9247 if (V->getOpcode() == ISD::BITCAST) 9248 V = V.getOperand(0); 9249 9250 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) { 9251 SDLoc dl(N); 9252 // Handle only simple case where vector being inserted and vector 9253 // being extracted are of same type, and are half size of larger vectors. 9254 EVT BigVT = V->getOperand(0).getValueType(); 9255 EVT SmallVT = V->getOperand(1).getValueType(); 9256 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits()) 9257 return SDValue(); 9258 9259 // Only handle cases where both indexes are constants with the same type. 9260 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1)); 9261 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2)); 9262 9263 if (InsIdx && ExtIdx && 9264 InsIdx->getValueType(0).getSizeInBits() <= 64 && 9265 ExtIdx->getValueType(0).getSizeInBits() <= 64) { 9266 // Combine: 9267 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx) 9268 // Into: 9269 // indices are equal or bit offsets are equal => V1 9270 // otherwise => (extract_subvec V1, ExtIdx) 9271 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() == 9272 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits()) 9273 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1)); 9274 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT, 9275 DAG.getNode(ISD::BITCAST, dl, 9276 N->getOperand(0).getValueType(), 9277 V->getOperand(0)), N->getOperand(1)); 9278 } 9279 } 9280 9281 return SDValue(); 9282} 9283 9284// Tries to turn a shuffle of two CONCAT_VECTORS into a single concat. 9285static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) { 9286 EVT VT = N->getValueType(0); 9287 unsigned NumElts = VT.getVectorNumElements(); 9288 9289 SDValue N0 = N->getOperand(0); 9290 SDValue N1 = N->getOperand(1); 9291 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 9292 9293 SmallVector<SDValue, 4> Ops; 9294 EVT ConcatVT = N0.getOperand(0).getValueType(); 9295 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements(); 9296 unsigned NumConcats = NumElts / NumElemsPerConcat; 9297 9298 // Look at every vector that's inserted. We're looking for exact 9299 // subvector-sized copies from a concatenated vector 9300 for (unsigned I = 0; I != NumConcats; ++I) { 9301 // Make sure we're dealing with a copy. 9302 unsigned Begin = I * NumElemsPerConcat; 9303 bool AllUndef = true, NoUndef = true; 9304 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) { 9305 if (SVN->getMaskElt(J) >= 0) 9306 AllUndef = false; 9307 else 9308 NoUndef = false; 9309 } 9310 9311 if (NoUndef) { 9312 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0) 9313 return SDValue(); 9314 9315 for (unsigned J = 1; J != NumElemsPerConcat; ++J) 9316 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J)) 9317 return SDValue(); 9318 9319 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat; 9320 if (FirstElt < N0.getNumOperands()) 9321 Ops.push_back(N0.getOperand(FirstElt)); 9322 else 9323 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands())); 9324 9325 } else if (AllUndef) { 9326 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType())); 9327 } else { // Mixed with general masks and undefs, can't do optimization. 9328 return SDValue(); 9329 } 9330 } 9331 9332 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops.data(), 9333 Ops.size()); 9334} 9335 9336SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) { 9337 EVT VT = N->getValueType(0); 9338 unsigned NumElts = VT.getVectorNumElements(); 9339 9340 SDValue N0 = N->getOperand(0); 9341 SDValue N1 = N->getOperand(1); 9342 9343 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG"); 9344 9345 // Canonicalize shuffle undef, undef -> undef 9346 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF) 9347 return DAG.getUNDEF(VT); 9348 9349 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 9350 9351 // Canonicalize shuffle v, v -> v, undef 9352 if (N0 == N1) { 9353 SmallVector<int, 8> NewMask; 9354 for (unsigned i = 0; i != NumElts; ++i) { 9355 int Idx = SVN->getMaskElt(i); 9356 if (Idx >= (int)NumElts) Idx -= NumElts; 9357 NewMask.push_back(Idx); 9358 } 9359 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT), 9360 &NewMask[0]); 9361 } 9362 9363 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask. 9364 if (N0.getOpcode() == ISD::UNDEF) { 9365 SmallVector<int, 8> NewMask; 9366 for (unsigned i = 0; i != NumElts; ++i) { 9367 int Idx = SVN->getMaskElt(i); 9368 if (Idx >= 0) { 9369 if (Idx < (int)NumElts) 9370 Idx += NumElts; 9371 else 9372 Idx -= NumElts; 9373 } 9374 NewMask.push_back(Idx); 9375 } 9376 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT), 9377 &NewMask[0]); 9378 } 9379 9380 // Remove references to rhs if it is undef 9381 if (N1.getOpcode() == ISD::UNDEF) { 9382 bool Changed = false; 9383 SmallVector<int, 8> NewMask; 9384 for (unsigned i = 0; i != NumElts; ++i) { 9385 int Idx = SVN->getMaskElt(i); 9386 if (Idx >= (int)NumElts) { 9387 Idx = -1; 9388 Changed = true; 9389 } 9390 NewMask.push_back(Idx); 9391 } 9392 if (Changed) 9393 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]); 9394 } 9395 9396 // If it is a splat, check if the argument vector is another splat or a 9397 // build_vector with all scalar elements the same. 9398 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) { 9399 SDNode *V = N0.getNode(); 9400 9401 // If this is a bit convert that changes the element type of the vector but 9402 // not the number of vector elements, look through it. Be careful not to 9403 // look though conversions that change things like v4f32 to v2f64. 9404 if (V->getOpcode() == ISD::BITCAST) { 9405 SDValue ConvInput = V->getOperand(0); 9406 if (ConvInput.getValueType().isVector() && 9407 ConvInput.getValueType().getVectorNumElements() == NumElts) 9408 V = ConvInput.getNode(); 9409 } 9410 9411 if (V->getOpcode() == ISD::BUILD_VECTOR) { 9412 assert(V->getNumOperands() == NumElts && 9413 "BUILD_VECTOR has wrong number of operands"); 9414 SDValue Base; 9415 bool AllSame = true; 9416 for (unsigned i = 0; i != NumElts; ++i) { 9417 if (V->getOperand(i).getOpcode() != ISD::UNDEF) { 9418 Base = V->getOperand(i); 9419 break; 9420 } 9421 } 9422 // Splat of <u, u, u, u>, return <u, u, u, u> 9423 if (!Base.getNode()) 9424 return N0; 9425 for (unsigned i = 0; i != NumElts; ++i) { 9426 if (V->getOperand(i) != Base) { 9427 AllSame = false; 9428 break; 9429 } 9430 } 9431 // Splat of <x, x, x, x>, return <x, x, x, x> 9432 if (AllSame) 9433 return N0; 9434 } 9435 } 9436 9437 if (N0.getOpcode() == ISD::CONCAT_VECTORS && 9438 Level < AfterLegalizeVectorOps && 9439 (N1.getOpcode() == ISD::UNDEF || 9440 (N1.getOpcode() == ISD::CONCAT_VECTORS && 9441 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) { 9442 SDValue V = partitionShuffleOfConcats(N, DAG); 9443 9444 if (V.getNode()) 9445 return V; 9446 } 9447 9448 // If this shuffle node is simply a swizzle of another shuffle node, 9449 // and it reverses the swizzle of the previous shuffle then we can 9450 // optimize shuffle(shuffle(x, undef), undef) -> x. 9451 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG && 9452 N1.getOpcode() == ISD::UNDEF) { 9453 9454 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0); 9455 9456 // Shuffle nodes can only reverse shuffles with a single non-undef value. 9457 if (N0.getOperand(1).getOpcode() != ISD::UNDEF) 9458 return SDValue(); 9459 9460 // The incoming shuffle must be of the same type as the result of the 9461 // current shuffle. 9462 assert(OtherSV->getOperand(0).getValueType() == VT && 9463 "Shuffle types don't match"); 9464 9465 for (unsigned i = 0; i != NumElts; ++i) { 9466 int Idx = SVN->getMaskElt(i); 9467 assert(Idx < (int)NumElts && "Index references undef operand"); 9468 // Next, this index comes from the first value, which is the incoming 9469 // shuffle. Adopt the incoming index. 9470 if (Idx >= 0) 9471 Idx = OtherSV->getMaskElt(Idx); 9472 9473 // The combined shuffle must map each index to itself. 9474 if (Idx >= 0 && (unsigned)Idx != i) 9475 return SDValue(); 9476 } 9477 9478 return OtherSV->getOperand(0); 9479 } 9480 9481 return SDValue(); 9482} 9483 9484/// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform 9485/// an AND to a vector_shuffle with the destination vector and a zero vector. 9486/// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==> 9487/// vector_shuffle V, Zero, <0, 4, 2, 4> 9488SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) { 9489 EVT VT = N->getValueType(0); 9490 SDLoc dl(N); 9491 SDValue LHS = N->getOperand(0); 9492 SDValue RHS = N->getOperand(1); 9493 if (N->getOpcode() == ISD::AND) { 9494 if (RHS.getOpcode() == ISD::BITCAST) 9495 RHS = RHS.getOperand(0); 9496 if (RHS.getOpcode() == ISD::BUILD_VECTOR) { 9497 SmallVector<int, 8> Indices; 9498 unsigned NumElts = RHS.getNumOperands(); 9499 for (unsigned i = 0; i != NumElts; ++i) { 9500 SDValue Elt = RHS.getOperand(i); 9501 if (!isa<ConstantSDNode>(Elt)) 9502 return SDValue(); 9503 9504 if (cast<ConstantSDNode>(Elt)->isAllOnesValue()) 9505 Indices.push_back(i); 9506 else if (cast<ConstantSDNode>(Elt)->isNullValue()) 9507 Indices.push_back(NumElts); 9508 else 9509 return SDValue(); 9510 } 9511 9512 // Let's see if the target supports this vector_shuffle. 9513 EVT RVT = RHS.getValueType(); 9514 if (!TLI.isVectorClearMaskLegal(Indices, RVT)) 9515 return SDValue(); 9516 9517 // Return the new VECTOR_SHUFFLE node. 9518 EVT EltVT = RVT.getVectorElementType(); 9519 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(), 9520 DAG.getConstant(0, EltVT)); 9521 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), 9522 RVT, &ZeroOps[0], ZeroOps.size()); 9523 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS); 9524 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]); 9525 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf); 9526 } 9527 } 9528 9529 return SDValue(); 9530} 9531 9532/// SimplifyVBinOp - Visit a binary vector operation, like ADD. 9533SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) { 9534 assert(N->getValueType(0).isVector() && 9535 "SimplifyVBinOp only works on vectors!"); 9536 9537 SDValue LHS = N->getOperand(0); 9538 SDValue RHS = N->getOperand(1); 9539 SDValue Shuffle = XformToShuffleWithZero(N); 9540 if (Shuffle.getNode()) return Shuffle; 9541 9542 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold 9543 // this operation. 9544 if (LHS.getOpcode() == ISD::BUILD_VECTOR && 9545 RHS.getOpcode() == ISD::BUILD_VECTOR) { 9546 SmallVector<SDValue, 8> Ops; 9547 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) { 9548 SDValue LHSOp = LHS.getOperand(i); 9549 SDValue RHSOp = RHS.getOperand(i); 9550 // If these two elements can't be folded, bail out. 9551 if ((LHSOp.getOpcode() != ISD::UNDEF && 9552 LHSOp.getOpcode() != ISD::Constant && 9553 LHSOp.getOpcode() != ISD::ConstantFP) || 9554 (RHSOp.getOpcode() != ISD::UNDEF && 9555 RHSOp.getOpcode() != ISD::Constant && 9556 RHSOp.getOpcode() != ISD::ConstantFP)) 9557 break; 9558 9559 // Can't fold divide by zero. 9560 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV || 9561 N->getOpcode() == ISD::FDIV) { 9562 if ((RHSOp.getOpcode() == ISD::Constant && 9563 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) || 9564 (RHSOp.getOpcode() == ISD::ConstantFP && 9565 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero())) 9566 break; 9567 } 9568 9569 EVT VT = LHSOp.getValueType(); 9570 EVT RVT = RHSOp.getValueType(); 9571 if (RVT != VT) { 9572 // Integer BUILD_VECTOR operands may have types larger than the element 9573 // size (e.g., when the element type is not legal). Prior to type 9574 // legalization, the types may not match between the two BUILD_VECTORS. 9575 // Truncate one of the operands to make them match. 9576 if (RVT.getSizeInBits() > VT.getSizeInBits()) { 9577 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp); 9578 } else { 9579 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp); 9580 VT = RVT; 9581 } 9582 } 9583 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT, 9584 LHSOp, RHSOp); 9585 if (FoldOp.getOpcode() != ISD::UNDEF && 9586 FoldOp.getOpcode() != ISD::Constant && 9587 FoldOp.getOpcode() != ISD::ConstantFP) 9588 break; 9589 Ops.push_back(FoldOp); 9590 AddToWorkList(FoldOp.getNode()); 9591 } 9592 9593 if (Ops.size() == LHS.getNumOperands()) 9594 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), 9595 LHS.getValueType(), &Ops[0], Ops.size()); 9596 } 9597 9598 return SDValue(); 9599} 9600 9601/// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG. 9602SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) { 9603 assert(N->getValueType(0).isVector() && 9604 "SimplifyVUnaryOp only works on vectors!"); 9605 9606 SDValue N0 = N->getOperand(0); 9607 9608 if (N0.getOpcode() != ISD::BUILD_VECTOR) 9609 return SDValue(); 9610 9611 // Operand is a BUILD_VECTOR node, see if we can constant fold it. 9612 SmallVector<SDValue, 8> Ops; 9613 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) { 9614 SDValue Op = N0.getOperand(i); 9615 if (Op.getOpcode() != ISD::UNDEF && 9616 Op.getOpcode() != ISD::ConstantFP) 9617 break; 9618 EVT EltVT = Op.getValueType(); 9619 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op); 9620 if (FoldOp.getOpcode() != ISD::UNDEF && 9621 FoldOp.getOpcode() != ISD::ConstantFP) 9622 break; 9623 Ops.push_back(FoldOp); 9624 AddToWorkList(FoldOp.getNode()); 9625 } 9626 9627 if (Ops.size() != N0.getNumOperands()) 9628 return SDValue(); 9629 9630 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), 9631 N0.getValueType(), &Ops[0], Ops.size()); 9632} 9633 9634SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0, 9635 SDValue N1, SDValue N2){ 9636 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!"); 9637 9638 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2, 9639 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 9640 9641 // If we got a simplified select_cc node back from SimplifySelectCC, then 9642 // break it down into a new SETCC node, and a new SELECT node, and then return 9643 // the SELECT node, since we were called with a SELECT node. 9644 if (SCC.getNode()) { 9645 // Check to see if we got a select_cc back (to turn into setcc/select). 9646 // Otherwise, just return whatever node we got back, like fabs. 9647 if (SCC.getOpcode() == ISD::SELECT_CC) { 9648 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0), 9649 N0.getValueType(), 9650 SCC.getOperand(0), SCC.getOperand(1), 9651 SCC.getOperand(4)); 9652 AddToWorkList(SETCC.getNode()); 9653 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), 9654 SCC.getOperand(2), SCC.getOperand(3), SETCC); 9655 } 9656 9657 return SCC; 9658 } 9659 return SDValue(); 9660} 9661 9662/// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS 9663/// are the two values being selected between, see if we can simplify the 9664/// select. Callers of this should assume that TheSelect is deleted if this 9665/// returns true. As such, they should return the appropriate thing (e.g. the 9666/// node) back to the top-level of the DAG combiner loop to avoid it being 9667/// looked at. 9668bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS, 9669 SDValue RHS) { 9670 9671 // Cannot simplify select with vector condition 9672 if (TheSelect->getOperand(0).getValueType().isVector()) return false; 9673 9674 // If this is a select from two identical things, try to pull the operation 9675 // through the select. 9676 if (LHS.getOpcode() != RHS.getOpcode() || 9677 !LHS.hasOneUse() || !RHS.hasOneUse()) 9678 return false; 9679 9680 // If this is a load and the token chain is identical, replace the select 9681 // of two loads with a load through a select of the address to load from. 9682 // This triggers in things like "select bool X, 10.0, 123.0" after the FP 9683 // constants have been dropped into the constant pool. 9684 if (LHS.getOpcode() == ISD::LOAD) { 9685 LoadSDNode *LLD = cast<LoadSDNode>(LHS); 9686 LoadSDNode *RLD = cast<LoadSDNode>(RHS); 9687 9688 // Token chains must be identical. 9689 if (LHS.getOperand(0) != RHS.getOperand(0) || 9690 // Do not let this transformation reduce the number of volatile loads. 9691 LLD->isVolatile() || RLD->isVolatile() || 9692 // If this is an EXTLOAD, the VT's must match. 9693 LLD->getMemoryVT() != RLD->getMemoryVT() || 9694 // If this is an EXTLOAD, the kind of extension must match. 9695 (LLD->getExtensionType() != RLD->getExtensionType() && 9696 // The only exception is if one of the extensions is anyext. 9697 LLD->getExtensionType() != ISD::EXTLOAD && 9698 RLD->getExtensionType() != ISD::EXTLOAD) || 9699 // FIXME: this discards src value information. This is 9700 // over-conservative. It would be beneficial to be able to remember 9701 // both potential memory locations. Since we are discarding 9702 // src value info, don't do the transformation if the memory 9703 // locations are not in the default address space. 9704 LLD->getPointerInfo().getAddrSpace() != 0 || 9705 RLD->getPointerInfo().getAddrSpace() != 0 || 9706 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(), 9707 LLD->getBasePtr().getValueType())) 9708 return false; 9709 9710 // Check that the select condition doesn't reach either load. If so, 9711 // folding this will induce a cycle into the DAG. If not, this is safe to 9712 // xform, so create a select of the addresses. 9713 SDValue Addr; 9714 if (TheSelect->getOpcode() == ISD::SELECT) { 9715 SDNode *CondNode = TheSelect->getOperand(0).getNode(); 9716 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) || 9717 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode))) 9718 return false; 9719 // The loads must not depend on one another. 9720 if (LLD->isPredecessorOf(RLD) || 9721 RLD->isPredecessorOf(LLD)) 9722 return false; 9723 Addr = DAG.getSelect(SDLoc(TheSelect), 9724 LLD->getBasePtr().getValueType(), 9725 TheSelect->getOperand(0), LLD->getBasePtr(), 9726 RLD->getBasePtr()); 9727 } else { // Otherwise SELECT_CC 9728 SDNode *CondLHS = TheSelect->getOperand(0).getNode(); 9729 SDNode *CondRHS = TheSelect->getOperand(1).getNode(); 9730 9731 if ((LLD->hasAnyUseOfValue(1) && 9732 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) || 9733 (RLD->hasAnyUseOfValue(1) && 9734 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS)))) 9735 return false; 9736 9737 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect), 9738 LLD->getBasePtr().getValueType(), 9739 TheSelect->getOperand(0), 9740 TheSelect->getOperand(1), 9741 LLD->getBasePtr(), RLD->getBasePtr(), 9742 TheSelect->getOperand(4)); 9743 } 9744 9745 SDValue Load; 9746 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) { 9747 Load = DAG.getLoad(TheSelect->getValueType(0), 9748 SDLoc(TheSelect), 9749 // FIXME: Discards pointer info. 9750 LLD->getChain(), Addr, MachinePointerInfo(), 9751 LLD->isVolatile(), LLD->isNonTemporal(), 9752 LLD->isInvariant(), LLD->getAlignment()); 9753 } else { 9754 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ? 9755 RLD->getExtensionType() : LLD->getExtensionType(), 9756 SDLoc(TheSelect), 9757 TheSelect->getValueType(0), 9758 // FIXME: Discards pointer info. 9759 LLD->getChain(), Addr, MachinePointerInfo(), 9760 LLD->getMemoryVT(), LLD->isVolatile(), 9761 LLD->isNonTemporal(), LLD->getAlignment()); 9762 } 9763 9764 // Users of the select now use the result of the load. 9765 CombineTo(TheSelect, Load); 9766 9767 // Users of the old loads now use the new load's chain. We know the 9768 // old-load value is dead now. 9769 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1)); 9770 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1)); 9771 return true; 9772 } 9773 9774 return false; 9775} 9776 9777/// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3 9778/// where 'cond' is the comparison specified by CC. 9779SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, 9780 SDValue N2, SDValue N3, 9781 ISD::CondCode CC, bool NotExtCompare) { 9782 // (x ? y : y) -> y. 9783 if (N2 == N3) return N2; 9784 9785 EVT VT = N2.getValueType(); 9786 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 9787 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode()); 9788 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode()); 9789 9790 // Determine if the condition we're dealing with is constant 9791 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()), 9792 N0, N1, CC, DL, false); 9793 if (SCC.getNode()) AddToWorkList(SCC.getNode()); 9794 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode()); 9795 9796 // fold select_cc true, x, y -> x 9797 if (SCCC && !SCCC->isNullValue()) 9798 return N2; 9799 // fold select_cc false, x, y -> y 9800 if (SCCC && SCCC->isNullValue()) 9801 return N3; 9802 9803 // Check to see if we can simplify the select into an fabs node 9804 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) { 9805 // Allow either -0.0 or 0.0 9806 if (CFP->getValueAPF().isZero()) { 9807 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs 9808 if ((CC == ISD::SETGE || CC == ISD::SETGT) && 9809 N0 == N2 && N3.getOpcode() == ISD::FNEG && 9810 N2 == N3.getOperand(0)) 9811 return DAG.getNode(ISD::FABS, DL, VT, N0); 9812 9813 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs 9814 if ((CC == ISD::SETLT || CC == ISD::SETLE) && 9815 N0 == N3 && N2.getOpcode() == ISD::FNEG && 9816 N2.getOperand(0) == N3) 9817 return DAG.getNode(ISD::FABS, DL, VT, N3); 9818 } 9819 } 9820 9821 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)" 9822 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0 9823 // in it. This is a win when the constant is not otherwise available because 9824 // it replaces two constant pool loads with one. We only do this if the FP 9825 // type is known to be legal, because if it isn't, then we are before legalize 9826 // types an we want the other legalization to happen first (e.g. to avoid 9827 // messing with soft float) and if the ConstantFP is not legal, because if 9828 // it is legal, we may not need to store the FP constant in a constant pool. 9829 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2)) 9830 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) { 9831 if (TLI.isTypeLegal(N2.getValueType()) && 9832 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) != 9833 TargetLowering::Legal) && 9834 // If both constants have multiple uses, then we won't need to do an 9835 // extra load, they are likely around in registers for other users. 9836 (TV->hasOneUse() || FV->hasOneUse())) { 9837 Constant *Elts[] = { 9838 const_cast<ConstantFP*>(FV->getConstantFPValue()), 9839 const_cast<ConstantFP*>(TV->getConstantFPValue()) 9840 }; 9841 Type *FPTy = Elts[0]->getType(); 9842 const DataLayout &TD = *TLI.getDataLayout(); 9843 9844 // Create a ConstantArray of the two constants. 9845 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts); 9846 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(), 9847 TD.getPrefTypeAlignment(FPTy)); 9848 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment(); 9849 9850 // Get the offsets to the 0 and 1 element of the array so that we can 9851 // select between them. 9852 SDValue Zero = DAG.getIntPtrConstant(0); 9853 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType()); 9854 SDValue One = DAG.getIntPtrConstant(EltSize); 9855 9856 SDValue Cond = DAG.getSetCC(DL, 9857 getSetCCResultType(N0.getValueType()), 9858 N0, N1, CC); 9859 AddToWorkList(Cond.getNode()); 9860 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(), 9861 Cond, One, Zero); 9862 AddToWorkList(CstOffset.getNode()); 9863 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx, 9864 CstOffset); 9865 AddToWorkList(CPIdx.getNode()); 9866 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx, 9867 MachinePointerInfo::getConstantPool(), false, 9868 false, false, Alignment); 9869 9870 } 9871 } 9872 9873 // Check to see if we can perform the "gzip trick", transforming 9874 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A) 9875 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT && 9876 (N1C->isNullValue() || // (a < 0) ? b : 0 9877 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0 9878 EVT XType = N0.getValueType(); 9879 EVT AType = N2.getValueType(); 9880 if (XType.bitsGE(AType)) { 9881 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a 9882 // single-bit constant. 9883 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) { 9884 unsigned ShCtV = N2C->getAPIntValue().logBase2(); 9885 ShCtV = XType.getSizeInBits()-ShCtV-1; 9886 SDValue ShCt = DAG.getConstant(ShCtV, 9887 getShiftAmountTy(N0.getValueType())); 9888 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), 9889 XType, N0, ShCt); 9890 AddToWorkList(Shift.getNode()); 9891 9892 if (XType.bitsGT(AType)) { 9893 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift); 9894 AddToWorkList(Shift.getNode()); 9895 } 9896 9897 return DAG.getNode(ISD::AND, DL, AType, Shift, N2); 9898 } 9899 9900 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), 9901 XType, N0, 9902 DAG.getConstant(XType.getSizeInBits()-1, 9903 getShiftAmountTy(N0.getValueType()))); 9904 AddToWorkList(Shift.getNode()); 9905 9906 if (XType.bitsGT(AType)) { 9907 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift); 9908 AddToWorkList(Shift.getNode()); 9909 } 9910 9911 return DAG.getNode(ISD::AND, DL, AType, Shift, N2); 9912 } 9913 } 9914 9915 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A) 9916 // where y is has a single bit set. 9917 // A plaintext description would be, we can turn the SELECT_CC into an AND 9918 // when the condition can be materialized as an all-ones register. Any 9919 // single bit-test can be materialized as an all-ones register with 9920 // shift-left and shift-right-arith. 9921 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND && 9922 N0->getValueType(0) == VT && 9923 N1C && N1C->isNullValue() && 9924 N2C && N2C->isNullValue()) { 9925 SDValue AndLHS = N0->getOperand(0); 9926 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1)); 9927 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) { 9928 // Shift the tested bit over the sign bit. 9929 APInt AndMask = ConstAndRHS->getAPIntValue(); 9930 SDValue ShlAmt = 9931 DAG.getConstant(AndMask.countLeadingZeros(), 9932 getShiftAmountTy(AndLHS.getValueType())); 9933 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt); 9934 9935 // Now arithmetic right shift it all the way over, so the result is either 9936 // all-ones, or zero. 9937 SDValue ShrAmt = 9938 DAG.getConstant(AndMask.getBitWidth()-1, 9939 getShiftAmountTy(Shl.getValueType())); 9940 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt); 9941 9942 return DAG.getNode(ISD::AND, DL, VT, Shr, N3); 9943 } 9944 } 9945 9946 // fold select C, 16, 0 -> shl C, 4 9947 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() && 9948 TLI.getBooleanContents(N0.getValueType().isVector()) == 9949 TargetLowering::ZeroOrOneBooleanContent) { 9950 9951 // If the caller doesn't want us to simplify this into a zext of a compare, 9952 // don't do it. 9953 if (NotExtCompare && N2C->getAPIntValue() == 1) 9954 return SDValue(); 9955 9956 // Get a SetCC of the condition 9957 // NOTE: Don't create a SETCC if it's not legal on this target. 9958 if (!LegalOperations || 9959 TLI.isOperationLegal(ISD::SETCC, 9960 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) { 9961 SDValue Temp, SCC; 9962 // cast from setcc result type to select result type 9963 if (LegalTypes) { 9964 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()), 9965 N0, N1, CC); 9966 if (N2.getValueType().bitsLT(SCC.getValueType())) 9967 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2), 9968 N2.getValueType()); 9969 else 9970 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2), 9971 N2.getValueType(), SCC); 9972 } else { 9973 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC); 9974 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2), 9975 N2.getValueType(), SCC); 9976 } 9977 9978 AddToWorkList(SCC.getNode()); 9979 AddToWorkList(Temp.getNode()); 9980 9981 if (N2C->getAPIntValue() == 1) 9982 return Temp; 9983 9984 // shl setcc result by log2 n2c 9985 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp, 9986 DAG.getConstant(N2C->getAPIntValue().logBase2(), 9987 getShiftAmountTy(Temp.getValueType()))); 9988 } 9989 } 9990 9991 // Check to see if this is the equivalent of setcc 9992 // FIXME: Turn all of these into setcc if setcc if setcc is legal 9993 // otherwise, go ahead with the folds. 9994 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) { 9995 EVT XType = N0.getValueType(); 9996 if (!LegalOperations || 9997 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) { 9998 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC); 9999 if (Res.getValueType() != VT) 10000 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res); 10001 return Res; 10002 } 10003 10004 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X)))) 10005 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ && 10006 (!LegalOperations || 10007 TLI.isOperationLegal(ISD::CTLZ, XType))) { 10008 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0); 10009 return DAG.getNode(ISD::SRL, DL, XType, Ctlz, 10010 DAG.getConstant(Log2_32(XType.getSizeInBits()), 10011 getShiftAmountTy(Ctlz.getValueType()))); 10012 } 10013 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1)) 10014 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) { 10015 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0), 10016 XType, DAG.getConstant(0, XType), N0); 10017 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType); 10018 return DAG.getNode(ISD::SRL, DL, XType, 10019 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0), 10020 DAG.getConstant(XType.getSizeInBits()-1, 10021 getShiftAmountTy(XType))); 10022 } 10023 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1)) 10024 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) { 10025 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0, 10026 DAG.getConstant(XType.getSizeInBits()-1, 10027 getShiftAmountTy(N0.getValueType()))); 10028 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType)); 10029 } 10030 } 10031 10032 // Check to see if this is an integer abs. 10033 // select_cc setg[te] X, 0, X, -X -> 10034 // select_cc setgt X, -1, X, -X -> 10035 // select_cc setl[te] X, 0, -X, X -> 10036 // select_cc setlt X, 1, -X, X -> 10037 // Y = sra (X, size(X)-1); xor (add (X, Y), Y) 10038 if (N1C) { 10039 ConstantSDNode *SubC = NULL; 10040 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) || 10041 (N1C->isAllOnesValue() && CC == ISD::SETGT)) && 10042 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1)) 10043 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0)); 10044 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) || 10045 (N1C->isOne() && CC == ISD::SETLT)) && 10046 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) 10047 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0)); 10048 10049 EVT XType = N0.getValueType(); 10050 if (SubC && SubC->isNullValue() && XType.isInteger()) { 10051 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType, 10052 N0, 10053 DAG.getConstant(XType.getSizeInBits()-1, 10054 getShiftAmountTy(N0.getValueType()))); 10055 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0), 10056 XType, N0, Shift); 10057 AddToWorkList(Shift.getNode()); 10058 AddToWorkList(Add.getNode()); 10059 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift); 10060 } 10061 } 10062 10063 return SDValue(); 10064} 10065 10066/// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC. 10067SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0, 10068 SDValue N1, ISD::CondCode Cond, 10069 SDLoc DL, bool foldBooleans) { 10070 TargetLowering::DAGCombinerInfo 10071 DagCombineInfo(DAG, Level, false, this); 10072 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL); 10073} 10074 10075/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 10076/// return a DAG expression to select that will generate the same value by 10077/// multiplying by a magic number. See: 10078/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 10079SDValue DAGCombiner::BuildSDIV(SDNode *N) { 10080 std::vector<SDNode*> Built; 10081 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built); 10082 10083 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end(); 10084 ii != ee; ++ii) 10085 AddToWorkList(*ii); 10086 return S; 10087} 10088 10089/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 10090/// return a DAG expression to select that will generate the same value by 10091/// multiplying by a magic number. See: 10092/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 10093SDValue DAGCombiner::BuildUDIV(SDNode *N) { 10094 std::vector<SDNode*> Built; 10095 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built); 10096 10097 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end(); 10098 ii != ee; ++ii) 10099 AddToWorkList(*ii); 10100 return S; 10101} 10102 10103/// FindBaseOffset - Return true if base is a frame index, which is known not 10104// to alias with anything but itself. Provides base object and offset as 10105// results. 10106static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset, 10107 const GlobalValue *&GV, const void *&CV) { 10108 // Assume it is a primitive operation. 10109 Base = Ptr; Offset = 0; GV = 0; CV = 0; 10110 10111 // If it's an adding a simple constant then integrate the offset. 10112 if (Base.getOpcode() == ISD::ADD) { 10113 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) { 10114 Base = Base.getOperand(0); 10115 Offset += C->getZExtValue(); 10116 } 10117 } 10118 10119 // Return the underlying GlobalValue, and update the Offset. Return false 10120 // for GlobalAddressSDNode since the same GlobalAddress may be represented 10121 // by multiple nodes with different offsets. 10122 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) { 10123 GV = G->getGlobal(); 10124 Offset += G->getOffset(); 10125 return false; 10126 } 10127 10128 // Return the underlying Constant value, and update the Offset. Return false 10129 // for ConstantSDNodes since the same constant pool entry may be represented 10130 // by multiple nodes with different offsets. 10131 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) { 10132 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal() 10133 : (const void *)C->getConstVal(); 10134 Offset += C->getOffset(); 10135 return false; 10136 } 10137 // If it's any of the following then it can't alias with anything but itself. 10138 return isa<FrameIndexSDNode>(Base); 10139} 10140 10141/// isAlias - Return true if there is any possibility that the two addresses 10142/// overlap. 10143bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1, 10144 const Value *SrcValue1, int SrcValueOffset1, 10145 unsigned SrcValueAlign1, 10146 const MDNode *TBAAInfo1, 10147 SDValue Ptr2, int64_t Size2, 10148 const Value *SrcValue2, int SrcValueOffset2, 10149 unsigned SrcValueAlign2, 10150 const MDNode *TBAAInfo2) const { 10151 // If they are the same then they must be aliases. 10152 if (Ptr1 == Ptr2) return true; 10153 10154 // Gather base node and offset information. 10155 SDValue Base1, Base2; 10156 int64_t Offset1, Offset2; 10157 const GlobalValue *GV1, *GV2; 10158 const void *CV1, *CV2; 10159 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1); 10160 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2); 10161 10162 // If they have a same base address then check to see if they overlap. 10163 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2))) 10164 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1); 10165 10166 // It is possible for different frame indices to alias each other, mostly 10167 // when tail call optimization reuses return address slots for arguments. 10168 // To catch this case, look up the actual index of frame indices to compute 10169 // the real alias relationship. 10170 if (isFrameIndex1 && isFrameIndex2) { 10171 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 10172 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex()); 10173 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex()); 10174 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1); 10175 } 10176 10177 // Otherwise, if we know what the bases are, and they aren't identical, then 10178 // we know they cannot alias. 10179 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2)) 10180 return false; 10181 10182 // If we know required SrcValue1 and SrcValue2 have relatively large alignment 10183 // compared to the size and offset of the access, we may be able to prove they 10184 // do not alias. This check is conservative for now to catch cases created by 10185 // splitting vector types. 10186 if ((SrcValueAlign1 == SrcValueAlign2) && 10187 (SrcValueOffset1 != SrcValueOffset2) && 10188 (Size1 == Size2) && (SrcValueAlign1 > Size1)) { 10189 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1; 10190 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1; 10191 10192 // There is no overlap between these relatively aligned accesses of similar 10193 // size, return no alias. 10194 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1) 10195 return false; 10196 } 10197 10198 if (CombinerGlobalAA) { 10199 // Use alias analysis information. 10200 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2); 10201 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset; 10202 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset; 10203 AliasAnalysis::AliasResult AAResult = 10204 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1), 10205 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2)); 10206 if (AAResult == AliasAnalysis::NoAlias) 10207 return false; 10208 } 10209 10210 // Otherwise we have to assume they alias. 10211 return true; 10212} 10213 10214bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) { 10215 SDValue Ptr0, Ptr1; 10216 int64_t Size0, Size1; 10217 const Value *SrcValue0, *SrcValue1; 10218 int SrcValueOffset0, SrcValueOffset1; 10219 unsigned SrcValueAlign0, SrcValueAlign1; 10220 const MDNode *SrcTBAAInfo0, *SrcTBAAInfo1; 10221 FindAliasInfo(Op0, Ptr0, Size0, SrcValue0, SrcValueOffset0, 10222 SrcValueAlign0, SrcTBAAInfo0); 10223 FindAliasInfo(Op1, Ptr1, Size1, SrcValue1, SrcValueOffset1, 10224 SrcValueAlign1, SrcTBAAInfo1); 10225 return isAlias(Ptr0, Size0, SrcValue0, SrcValueOffset0, 10226 SrcValueAlign0, SrcTBAAInfo0, 10227 Ptr1, Size1, SrcValue1, SrcValueOffset1, 10228 SrcValueAlign1, SrcTBAAInfo1); 10229} 10230 10231/// FindAliasInfo - Extracts the relevant alias information from the memory 10232/// node. Returns true if the operand was a load. 10233bool DAGCombiner::FindAliasInfo(SDNode *N, 10234 SDValue &Ptr, int64_t &Size, 10235 const Value *&SrcValue, 10236 int &SrcValueOffset, 10237 unsigned &SrcValueAlign, 10238 const MDNode *&TBAAInfo) const { 10239 LSBaseSDNode *LS = cast<LSBaseSDNode>(N); 10240 10241 Ptr = LS->getBasePtr(); 10242 Size = LS->getMemoryVT().getSizeInBits() >> 3; 10243 SrcValue = LS->getSrcValue(); 10244 SrcValueOffset = LS->getSrcValueOffset(); 10245 SrcValueAlign = LS->getOriginalAlignment(); 10246 TBAAInfo = LS->getTBAAInfo(); 10247 return isa<LoadSDNode>(LS); 10248} 10249 10250/// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes, 10251/// looking for aliasing nodes and adding them to the Aliases vector. 10252void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain, 10253 SmallVector<SDValue, 8> &Aliases) { 10254 SmallVector<SDValue, 8> Chains; // List of chains to visit. 10255 SmallPtrSet<SDNode *, 16> Visited; // Visited node set. 10256 10257 // Get alias information for node. 10258 SDValue Ptr; 10259 int64_t Size; 10260 const Value *SrcValue; 10261 int SrcValueOffset; 10262 unsigned SrcValueAlign; 10263 const MDNode *SrcTBAAInfo; 10264 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset, 10265 SrcValueAlign, SrcTBAAInfo); 10266 10267 // Starting off. 10268 Chains.push_back(OriginalChain); 10269 unsigned Depth = 0; 10270 10271 // Look at each chain and determine if it is an alias. If so, add it to the 10272 // aliases list. If not, then continue up the chain looking for the next 10273 // candidate. 10274 while (!Chains.empty()) { 10275 SDValue Chain = Chains.back(); 10276 Chains.pop_back(); 10277 10278 // For TokenFactor nodes, look at each operand and only continue up the 10279 // chain until we find two aliases. If we've seen two aliases, assume we'll 10280 // find more and revert to original chain since the xform is unlikely to be 10281 // profitable. 10282 // 10283 // FIXME: The depth check could be made to return the last non-aliasing 10284 // chain we found before we hit a tokenfactor rather than the original 10285 // chain. 10286 if (Depth > 6 || Aliases.size() == 2) { 10287 Aliases.clear(); 10288 Aliases.push_back(OriginalChain); 10289 break; 10290 } 10291 10292 // Don't bother if we've been before. 10293 if (!Visited.insert(Chain.getNode())) 10294 continue; 10295 10296 switch (Chain.getOpcode()) { 10297 case ISD::EntryToken: 10298 // Entry token is ideal chain operand, but handled in FindBetterChain. 10299 break; 10300 10301 case ISD::LOAD: 10302 case ISD::STORE: { 10303 // Get alias information for Chain. 10304 SDValue OpPtr; 10305 int64_t OpSize; 10306 const Value *OpSrcValue; 10307 int OpSrcValueOffset; 10308 unsigned OpSrcValueAlign; 10309 const MDNode *OpSrcTBAAInfo; 10310 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize, 10311 OpSrcValue, OpSrcValueOffset, 10312 OpSrcValueAlign, 10313 OpSrcTBAAInfo); 10314 10315 // If chain is alias then stop here. 10316 if (!(IsLoad && IsOpLoad) && 10317 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign, 10318 SrcTBAAInfo, 10319 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset, 10320 OpSrcValueAlign, OpSrcTBAAInfo)) { 10321 Aliases.push_back(Chain); 10322 } else { 10323 // Look further up the chain. 10324 Chains.push_back(Chain.getOperand(0)); 10325 ++Depth; 10326 } 10327 break; 10328 } 10329 10330 case ISD::TokenFactor: 10331 // We have to check each of the operands of the token factor for "small" 10332 // token factors, so we queue them up. Adding the operands to the queue 10333 // (stack) in reverse order maintains the original order and increases the 10334 // likelihood that getNode will find a matching token factor (CSE.) 10335 if (Chain.getNumOperands() > 16) { 10336 Aliases.push_back(Chain); 10337 break; 10338 } 10339 for (unsigned n = Chain.getNumOperands(); n;) 10340 Chains.push_back(Chain.getOperand(--n)); 10341 ++Depth; 10342 break; 10343 10344 default: 10345 // For all other instructions we will just have to take what we can get. 10346 Aliases.push_back(Chain); 10347 break; 10348 } 10349 } 10350} 10351 10352/// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking 10353/// for a better chain (aliasing node.) 10354SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) { 10355 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor. 10356 10357 // Accumulate all the aliases to this node. 10358 GatherAllAliases(N, OldChain, Aliases); 10359 10360 // If no operands then chain to entry token. 10361 if (Aliases.size() == 0) 10362 return DAG.getEntryNode(); 10363 10364 // If a single operand then chain to it. We don't need to revisit it. 10365 if (Aliases.size() == 1) 10366 return Aliases[0]; 10367 10368 // Construct a custom tailored token factor. 10369 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, 10370 &Aliases[0], Aliases.size()); 10371} 10372 10373// SelectionDAG::Combine - This is the entry point for the file. 10374// 10375void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA, 10376 CodeGenOpt::Level OptLevel) { 10377 /// run - This is the main entry point to this class. 10378 /// 10379 DAGCombiner(*this, AA, OptLevel).Run(Level); 10380} 10381