DAGCombiner.cpp revision ceb3b4649086c4aa6aa535440748c5860093f6e6
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 "llvm/Target/TargetSubtargetInfo.h" 39#include <algorithm> 40using namespace llvm; 41 42STATISTIC(NodesCombined , "Number of dag nodes combined"); 43STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created"); 44STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created"); 45STATISTIC(OpsNarrowed , "Number of load/op/store narrowed"); 46STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int"); 47 48namespace { 49 static cl::opt<bool> 50 CombinerAA("combiner-alias-analysis", cl::Hidden, 51 cl::desc("Turn on alias analysis during testing")); 52 53 static cl::opt<bool> 54 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden, 55 cl::desc("Include global information in alias analysis")); 56 57//------------------------------ DAGCombiner ---------------------------------// 58 59 class DAGCombiner { 60 SelectionDAG &DAG; 61 const TargetLowering &TLI; 62 CombineLevel Level; 63 CodeGenOpt::Level OptLevel; 64 bool LegalOperations; 65 bool LegalTypes; 66 67 // Worklist of all of the nodes that need to be simplified. 68 // 69 // This has the semantics that when adding to the worklist, 70 // the item added must be next to be processed. It should 71 // also only appear once. The naive approach to this takes 72 // linear time. 73 // 74 // To reduce the insert/remove time to logarithmic, we use 75 // a set and a vector to maintain our worklist. 76 // 77 // The set contains the items on the worklist, but does not 78 // maintain the order they should be visited. 79 // 80 // The vector maintains the order nodes should be visited, but may 81 // contain duplicate or removed nodes. When choosing a node to 82 // visit, we pop off the order stack until we find an item that is 83 // also in the contents set. All operations are O(log N). 84 SmallPtrSet<SDNode*, 64> WorkListContents; 85 SmallVector<SDNode*, 64> WorkListOrder; 86 87 // AA - Used for DAG load/store alias analysis. 88 AliasAnalysis &AA; 89 90 /// AddUsersToWorkList - When an instruction is simplified, add all users of 91 /// the instruction to the work lists because they might get more simplified 92 /// now. 93 /// 94 void AddUsersToWorkList(SDNode *N) { 95 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 96 UI != UE; ++UI) 97 AddToWorkList(*UI); 98 } 99 100 /// visit - call the node-specific routine that knows how to fold each 101 /// particular type of node. 102 SDValue visit(SDNode *N); 103 104 public: 105 /// AddToWorkList - Add to the work list making sure its instance is at the 106 /// back (next to be processed.) 107 void AddToWorkList(SDNode *N) { 108 WorkListContents.insert(N); 109 WorkListOrder.push_back(N); 110 } 111 112 /// removeFromWorkList - remove all instances of N from the worklist. 113 /// 114 void removeFromWorkList(SDNode *N) { 115 WorkListContents.erase(N); 116 } 117 118 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, 119 bool AddTo = true); 120 121 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) { 122 return CombineTo(N, &Res, 1, AddTo); 123 } 124 125 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, 126 bool AddTo = true) { 127 SDValue To[] = { Res0, Res1 }; 128 return CombineTo(N, To, 2, AddTo); 129 } 130 131 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO); 132 133 private: 134 135 /// SimplifyDemandedBits - Check the specified integer node value to see if 136 /// it can be simplified or if things it uses can be simplified by bit 137 /// propagation. If so, return true. 138 bool SimplifyDemandedBits(SDValue Op) { 139 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits(); 140 APInt Demanded = APInt::getAllOnesValue(BitWidth); 141 return SimplifyDemandedBits(Op, Demanded); 142 } 143 144 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded); 145 146 bool CombineToPreIndexedLoadStore(SDNode *N); 147 bool CombineToPostIndexedLoadStore(SDNode *N); 148 149 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad); 150 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace); 151 SDValue SExtPromoteOperand(SDValue Op, EVT PVT); 152 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT); 153 SDValue PromoteIntBinOp(SDValue Op); 154 SDValue PromoteIntShiftOp(SDValue Op); 155 SDValue PromoteExtend(SDValue Op); 156 bool PromoteLoad(SDValue Op); 157 158 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs, 159 SDValue Trunc, SDValue ExtLoad, SDLoc DL, 160 ISD::NodeType ExtType); 161 162 /// combine - call the node-specific routine that knows how to fold each 163 /// particular type of node. If that doesn't do anything, try the 164 /// target-specific DAG combines. 165 SDValue combine(SDNode *N); 166 167 // Visitation implementation - Implement dag node combining for different 168 // node types. The semantics are as follows: 169 // Return Value: 170 // SDValue.getNode() == 0 - No change was made 171 // SDValue.getNode() == N - N was replaced, is dead and has been handled. 172 // otherwise - N should be replaced by the returned Operand. 173 // 174 SDValue visitTokenFactor(SDNode *N); 175 SDValue visitMERGE_VALUES(SDNode *N); 176 SDValue visitADD(SDNode *N); 177 SDValue visitSUB(SDNode *N); 178 SDValue visitADDC(SDNode *N); 179 SDValue visitSUBC(SDNode *N); 180 SDValue visitADDE(SDNode *N); 181 SDValue visitSUBE(SDNode *N); 182 SDValue visitMUL(SDNode *N); 183 SDValue visitSDIV(SDNode *N); 184 SDValue visitUDIV(SDNode *N); 185 SDValue visitSREM(SDNode *N); 186 SDValue visitUREM(SDNode *N); 187 SDValue visitMULHU(SDNode *N); 188 SDValue visitMULHS(SDNode *N); 189 SDValue visitSMUL_LOHI(SDNode *N); 190 SDValue visitUMUL_LOHI(SDNode *N); 191 SDValue visitSMULO(SDNode *N); 192 SDValue visitUMULO(SDNode *N); 193 SDValue visitSDIVREM(SDNode *N); 194 SDValue visitUDIVREM(SDNode *N); 195 SDValue visitAND(SDNode *N); 196 SDValue visitOR(SDNode *N); 197 SDValue visitXOR(SDNode *N); 198 SDValue SimplifyVBinOp(SDNode *N); 199 SDValue SimplifyVUnaryOp(SDNode *N); 200 SDValue visitSHL(SDNode *N); 201 SDValue visitSRA(SDNode *N); 202 SDValue visitSRL(SDNode *N); 203 SDValue visitCTLZ(SDNode *N); 204 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N); 205 SDValue visitCTTZ(SDNode *N); 206 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N); 207 SDValue visitCTPOP(SDNode *N); 208 SDValue visitSELECT(SDNode *N); 209 SDValue visitVSELECT(SDNode *N); 210 SDValue visitSELECT_CC(SDNode *N); 211 SDValue visitSETCC(SDNode *N); 212 SDValue visitSIGN_EXTEND(SDNode *N); 213 SDValue visitZERO_EXTEND(SDNode *N); 214 SDValue visitANY_EXTEND(SDNode *N); 215 SDValue visitSIGN_EXTEND_INREG(SDNode *N); 216 SDValue visitTRUNCATE(SDNode *N); 217 SDValue visitBITCAST(SDNode *N); 218 SDValue visitBUILD_PAIR(SDNode *N); 219 SDValue visitFADD(SDNode *N); 220 SDValue visitFSUB(SDNode *N); 221 SDValue visitFMUL(SDNode *N); 222 SDValue visitFMA(SDNode *N); 223 SDValue visitFDIV(SDNode *N); 224 SDValue visitFREM(SDNode *N); 225 SDValue visitFCOPYSIGN(SDNode *N); 226 SDValue visitSINT_TO_FP(SDNode *N); 227 SDValue visitUINT_TO_FP(SDNode *N); 228 SDValue visitFP_TO_SINT(SDNode *N); 229 SDValue visitFP_TO_UINT(SDNode *N); 230 SDValue visitFP_ROUND(SDNode *N); 231 SDValue visitFP_ROUND_INREG(SDNode *N); 232 SDValue visitFP_EXTEND(SDNode *N); 233 SDValue visitFNEG(SDNode *N); 234 SDValue visitFABS(SDNode *N); 235 SDValue visitFCEIL(SDNode *N); 236 SDValue visitFTRUNC(SDNode *N); 237 SDValue visitFFLOOR(SDNode *N); 238 SDValue visitBRCOND(SDNode *N); 239 SDValue visitBR_CC(SDNode *N); 240 SDValue visitLOAD(SDNode *N); 241 SDValue visitSTORE(SDNode *N); 242 SDValue visitINSERT_VECTOR_ELT(SDNode *N); 243 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N); 244 SDValue visitBUILD_VECTOR(SDNode *N); 245 SDValue visitCONCAT_VECTORS(SDNode *N); 246 SDValue visitEXTRACT_SUBVECTOR(SDNode *N); 247 SDValue visitVECTOR_SHUFFLE(SDNode *N); 248 249 SDValue XformToShuffleWithZero(SDNode *N); 250 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS); 251 252 SDValue visitShiftByConstant(SDNode *N, unsigned Amt); 253 254 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS); 255 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N); 256 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2); 257 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2, 258 SDValue N3, ISD::CondCode CC, 259 bool NotExtCompare = false); 260 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond, 261 SDLoc DL, bool foldBooleans = true); 262 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, 263 unsigned HiOp); 264 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT); 265 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT); 266 SDValue BuildSDIV(SDNode *N); 267 SDValue BuildUDIV(SDNode *N); 268 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, 269 bool DemandHighBits = true); 270 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1); 271 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL); 272 SDValue ReduceLoadWidth(SDNode *N); 273 SDValue ReduceLoadOpStoreWidth(SDNode *N); 274 SDValue TransformFPLoadStorePair(SDNode *N); 275 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N); 276 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N); 277 278 SDValue GetDemandedBits(SDValue V, const APInt &Mask); 279 280 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes, 281 /// looking for aliasing nodes and adding them to the Aliases vector. 282 void GatherAllAliases(SDNode *N, SDValue OriginalChain, 283 SmallVectorImpl<SDValue> &Aliases); 284 285 /// isAlias - Return true if there is any possibility that the two addresses 286 /// overlap. 287 bool isAlias(SDValue Ptr1, int64_t Size1, 288 const Value *SrcValue1, int SrcValueOffset1, 289 unsigned SrcValueAlign1, 290 const MDNode *TBAAInfo1, 291 SDValue Ptr2, int64_t Size2, 292 const Value *SrcValue2, int SrcValueOffset2, 293 unsigned SrcValueAlign2, 294 const MDNode *TBAAInfo2) const; 295 296 /// isAlias - Return true if there is any possibility that the two addresses 297 /// overlap. 298 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1); 299 300 /// FindAliasInfo - Extracts the relevant alias information from the memory 301 /// node. Returns true if the operand was a load. 302 bool FindAliasInfo(SDNode *N, 303 SDValue &Ptr, int64_t &Size, 304 const Value *&SrcValue, int &SrcValueOffset, 305 unsigned &SrcValueAlignment, 306 const MDNode *&TBAAInfo) const; 307 308 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, 309 /// looking for a better chain (aliasing node.) 310 SDValue FindBetterChain(SDNode *N, SDValue Chain); 311 312 /// Merge consecutive store operations into a wide store. 313 /// This optimization uses wide integers or vectors when possible. 314 /// \return True if some memory operations were changed. 315 bool MergeConsecutiveStores(StoreSDNode *N); 316 317 public: 318 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL) 319 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes), 320 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {} 321 322 /// Run - runs the dag combiner on all nodes in the work list 323 void Run(CombineLevel AtLevel); 324 325 SelectionDAG &getDAG() const { return DAG; } 326 327 /// getShiftAmountTy - Returns a type large enough to hold any valid 328 /// shift amount - before type legalization these can be huge. 329 EVT getShiftAmountTy(EVT LHSTy) { 330 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 331 if (LHSTy.isVector()) 332 return LHSTy; 333 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy) : TLI.getPointerTy(); 334 } 335 336 /// isTypeLegal - This method returns true if we are running before type 337 /// legalization or if the specified VT is legal. 338 bool isTypeLegal(const EVT &VT) { 339 if (!LegalTypes) return true; 340 return TLI.isTypeLegal(VT); 341 } 342 343 /// getSetCCResultType - Convenience wrapper around 344 /// TargetLowering::getSetCCResultType 345 EVT getSetCCResultType(EVT VT) const { 346 return TLI.getSetCCResultType(*DAG.getContext(), VT); 347 } 348 }; 349} 350 351 352namespace { 353/// WorkListRemover - This class is a DAGUpdateListener that removes any deleted 354/// nodes from the worklist. 355class WorkListRemover : public SelectionDAG::DAGUpdateListener { 356 DAGCombiner &DC; 357public: 358 explicit WorkListRemover(DAGCombiner &dc) 359 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {} 360 361 virtual void NodeDeleted(SDNode *N, SDNode *E) { 362 DC.removeFromWorkList(N); 363 } 364}; 365} 366 367//===----------------------------------------------------------------------===// 368// TargetLowering::DAGCombinerInfo implementation 369//===----------------------------------------------------------------------===// 370 371void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) { 372 ((DAGCombiner*)DC)->AddToWorkList(N); 373} 374 375void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) { 376 ((DAGCombiner*)DC)->removeFromWorkList(N); 377} 378 379SDValue TargetLowering::DAGCombinerInfo:: 380CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) { 381 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo); 382} 383 384SDValue TargetLowering::DAGCombinerInfo:: 385CombineTo(SDNode *N, SDValue Res, bool AddTo) { 386 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo); 387} 388 389 390SDValue TargetLowering::DAGCombinerInfo:: 391CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) { 392 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo); 393} 394 395void TargetLowering::DAGCombinerInfo:: 396CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { 397 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO); 398} 399 400//===----------------------------------------------------------------------===// 401// Helper Functions 402//===----------------------------------------------------------------------===// 403 404/// isNegatibleForFree - Return 1 if we can compute the negated form of the 405/// specified expression for the same cost as the expression itself, or 2 if we 406/// can compute the negated form more cheaply than the expression itself. 407static char isNegatibleForFree(SDValue Op, bool LegalOperations, 408 const TargetLowering &TLI, 409 const TargetOptions *Options, 410 unsigned Depth = 0) { 411 // fneg is removable even if it has multiple uses. 412 if (Op.getOpcode() == ISD::FNEG) return 2; 413 414 // Don't allow anything with multiple uses. 415 if (!Op.hasOneUse()) return 0; 416 417 // Don't recurse exponentially. 418 if (Depth > 6) return 0; 419 420 switch (Op.getOpcode()) { 421 default: return false; 422 case ISD::ConstantFP: 423 // Don't invert constant FP values after legalize. The negated constant 424 // isn't necessarily legal. 425 return LegalOperations ? 0 : 1; 426 case ISD::FADD: 427 // FIXME: determine better conditions for this xform. 428 if (!Options->UnsafeFPMath) return 0; 429 430 // After operation legalization, it might not be legal to create new FSUBs. 431 if (LegalOperations && 432 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType())) 433 return 0; 434 435 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B) 436 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, 437 Options, Depth + 1)) 438 return V; 439 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A) 440 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options, 441 Depth + 1); 442 case ISD::FSUB: 443 // We can't turn -(A-B) into B-A when we honor signed zeros. 444 if (!Options->UnsafeFPMath) return 0; 445 446 // fold (fneg (fsub A, B)) -> (fsub B, A) 447 return 1; 448 449 case ISD::FMUL: 450 case ISD::FDIV: 451 if (Options->HonorSignDependentRoundingFPMath()) return 0; 452 453 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y)) 454 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, 455 Options, Depth + 1)) 456 return V; 457 458 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options, 459 Depth + 1); 460 461 case ISD::FP_EXTEND: 462 case ISD::FP_ROUND: 463 case ISD::FSIN: 464 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options, 465 Depth + 1); 466 } 467} 468 469/// GetNegatedExpression - If isNegatibleForFree returns true, this function 470/// returns the newly negated expression. 471static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG, 472 bool LegalOperations, unsigned Depth = 0) { 473 // fneg is removable even if it has multiple uses. 474 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0); 475 476 // Don't allow anything with multiple uses. 477 assert(Op.hasOneUse() && "Unknown reuse!"); 478 479 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree"); 480 switch (Op.getOpcode()) { 481 default: llvm_unreachable("Unknown code"); 482 case ISD::ConstantFP: { 483 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF(); 484 V.changeSign(); 485 return DAG.getConstantFP(V, Op.getValueType()); 486 } 487 case ISD::FADD: 488 // FIXME: determine better conditions for this xform. 489 assert(DAG.getTarget().Options.UnsafeFPMath); 490 491 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B) 492 if (isNegatibleForFree(Op.getOperand(0), LegalOperations, 493 DAG.getTargetLoweringInfo(), 494 &DAG.getTarget().Options, Depth+1)) 495 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), 496 GetNegatedExpression(Op.getOperand(0), DAG, 497 LegalOperations, Depth+1), 498 Op.getOperand(1)); 499 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A) 500 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), 501 GetNegatedExpression(Op.getOperand(1), DAG, 502 LegalOperations, Depth+1), 503 Op.getOperand(0)); 504 case ISD::FSUB: 505 // We can't turn -(A-B) into B-A when we honor signed zeros. 506 assert(DAG.getTarget().Options.UnsafeFPMath); 507 508 // fold (fneg (fsub 0, B)) -> B 509 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0))) 510 if (N0CFP->getValueAPF().isZero()) 511 return Op.getOperand(1); 512 513 // fold (fneg (fsub A, B)) -> (fsub B, A) 514 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(), 515 Op.getOperand(1), Op.getOperand(0)); 516 517 case ISD::FMUL: 518 case ISD::FDIV: 519 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath()); 520 521 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) 522 if (isNegatibleForFree(Op.getOperand(0), LegalOperations, 523 DAG.getTargetLoweringInfo(), 524 &DAG.getTarget().Options, Depth+1)) 525 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), 526 GetNegatedExpression(Op.getOperand(0), DAG, 527 LegalOperations, Depth+1), 528 Op.getOperand(1)); 529 530 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y)) 531 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), 532 Op.getOperand(0), 533 GetNegatedExpression(Op.getOperand(1), DAG, 534 LegalOperations, Depth+1)); 535 536 case ISD::FP_EXTEND: 537 case ISD::FSIN: 538 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), 539 GetNegatedExpression(Op.getOperand(0), DAG, 540 LegalOperations, Depth+1)); 541 case ISD::FP_ROUND: 542 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(), 543 GetNegatedExpression(Op.getOperand(0), DAG, 544 LegalOperations, Depth+1), 545 Op.getOperand(1)); 546 } 547} 548 549 550// isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc 551// that selects between the values 1 and 0, making it equivalent to a setcc. 552// Also, set the incoming LHS, RHS, and CC references to the appropriate 553// nodes based on the type of node we are checking. This simplifies life a 554// bit for the callers. 555static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS, 556 SDValue &CC) { 557 if (N.getOpcode() == ISD::SETCC) { 558 LHS = N.getOperand(0); 559 RHS = N.getOperand(1); 560 CC = N.getOperand(2); 561 return true; 562 } 563 if (N.getOpcode() == ISD::SELECT_CC && 564 N.getOperand(2).getOpcode() == ISD::Constant && 565 N.getOperand(3).getOpcode() == ISD::Constant && 566 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 && 567 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) { 568 LHS = N.getOperand(0); 569 RHS = N.getOperand(1); 570 CC = N.getOperand(4); 571 return true; 572 } 573 return false; 574} 575 576// isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only 577// one use. If this is true, it allows the users to invert the operation for 578// free when it is profitable to do so. 579static bool isOneUseSetCC(SDValue N) { 580 SDValue N0, N1, N2; 581 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse()) 582 return true; 583 return false; 584} 585 586SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL, 587 SDValue N0, SDValue N1) { 588 EVT VT = N0.getValueType(); 589 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) { 590 if (isa<ConstantSDNode>(N1)) { 591 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2)) 592 SDValue OpNode = 593 DAG.FoldConstantArithmetic(Opc, VT, 594 cast<ConstantSDNode>(N0.getOperand(1)), 595 cast<ConstantSDNode>(N1)); 596 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode); 597 } 598 if (N0.hasOneUse()) { 599 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use 600 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, 601 N0.getOperand(0), N1); 602 AddToWorkList(OpNode.getNode()); 603 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1)); 604 } 605 } 606 607 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) { 608 if (isa<ConstantSDNode>(N0)) { 609 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2)) 610 SDValue OpNode = 611 DAG.FoldConstantArithmetic(Opc, VT, 612 cast<ConstantSDNode>(N1.getOperand(1)), 613 cast<ConstantSDNode>(N0)); 614 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode); 615 } 616 if (N1.hasOneUse()) { 617 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use 618 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, 619 N1.getOperand(0), N0); 620 AddToWorkList(OpNode.getNode()); 621 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1)); 622 } 623 } 624 625 return SDValue(); 626} 627 628SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo, 629 bool AddTo) { 630 assert(N->getNumValues() == NumTo && "Broken CombineTo call!"); 631 ++NodesCombined; 632 DEBUG(dbgs() << "\nReplacing.1 "; 633 N->dump(&DAG); 634 dbgs() << "\nWith: "; 635 To[0].getNode()->dump(&DAG); 636 dbgs() << " and " << NumTo-1 << " other values\n"; 637 for (unsigned i = 0, e = NumTo; i != e; ++i) 638 assert((!To[i].getNode() || 639 N->getValueType(i) == To[i].getValueType()) && 640 "Cannot combine value to value of different type!")); 641 WorkListRemover DeadNodes(*this); 642 DAG.ReplaceAllUsesWith(N, To); 643 if (AddTo) { 644 // Push the new nodes and any users onto the worklist 645 for (unsigned i = 0, e = NumTo; i != e; ++i) { 646 if (To[i].getNode()) { 647 AddToWorkList(To[i].getNode()); 648 AddUsersToWorkList(To[i].getNode()); 649 } 650 } 651 } 652 653 // Finally, if the node is now dead, remove it from the graph. The node 654 // may not be dead if the replacement process recursively simplified to 655 // something else needing this node. 656 if (N->use_empty()) { 657 // Nodes can be reintroduced into the worklist. Make sure we do not 658 // process a node that has been replaced. 659 removeFromWorkList(N); 660 661 // Finally, since the node is now dead, remove it from the graph. 662 DAG.DeleteNode(N); 663 } 664 return SDValue(N, 0); 665} 666 667void DAGCombiner:: 668CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) { 669 // Replace all uses. If any nodes become isomorphic to other nodes and 670 // are deleted, make sure to remove them from our worklist. 671 WorkListRemover DeadNodes(*this); 672 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New); 673 674 // Push the new node and any (possibly new) users onto the worklist. 675 AddToWorkList(TLO.New.getNode()); 676 AddUsersToWorkList(TLO.New.getNode()); 677 678 // Finally, if the node is now dead, remove it from the graph. The node 679 // may not be dead if the replacement process recursively simplified to 680 // something else needing this node. 681 if (TLO.Old.getNode()->use_empty()) { 682 removeFromWorkList(TLO.Old.getNode()); 683 684 // If the operands of this node are only used by the node, they will now 685 // be dead. Make sure to visit them first to delete dead nodes early. 686 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i) 687 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse()) 688 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode()); 689 690 DAG.DeleteNode(TLO.Old.getNode()); 691 } 692} 693 694/// SimplifyDemandedBits - Check the specified integer node value to see if 695/// it can be simplified or if things it uses can be simplified by bit 696/// propagation. If so, return true. 697bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) { 698 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations); 699 APInt KnownZero, KnownOne; 700 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO)) 701 return false; 702 703 // Revisit the node. 704 AddToWorkList(Op.getNode()); 705 706 // Replace the old value with the new one. 707 ++NodesCombined; 708 DEBUG(dbgs() << "\nReplacing.2 "; 709 TLO.Old.getNode()->dump(&DAG); 710 dbgs() << "\nWith: "; 711 TLO.New.getNode()->dump(&DAG); 712 dbgs() << '\n'); 713 714 CommitTargetLoweringOpt(TLO); 715 return true; 716} 717 718void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) { 719 SDLoc dl(Load); 720 EVT VT = Load->getValueType(0); 721 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0)); 722 723 DEBUG(dbgs() << "\nReplacing.9 "; 724 Load->dump(&DAG); 725 dbgs() << "\nWith: "; 726 Trunc.getNode()->dump(&DAG); 727 dbgs() << '\n'); 728 WorkListRemover DeadNodes(*this); 729 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc); 730 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1)); 731 removeFromWorkList(Load); 732 DAG.DeleteNode(Load); 733 AddToWorkList(Trunc.getNode()); 734} 735 736SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) { 737 Replace = false; 738 SDLoc dl(Op); 739 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) { 740 EVT MemVT = LD->getMemoryVT(); 741 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) 742 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD 743 : ISD::EXTLOAD) 744 : LD->getExtensionType(); 745 Replace = true; 746 return DAG.getExtLoad(ExtType, dl, PVT, 747 LD->getChain(), LD->getBasePtr(), 748 LD->getPointerInfo(), 749 MemVT, LD->isVolatile(), 750 LD->isNonTemporal(), LD->getAlignment()); 751 } 752 753 unsigned Opc = Op.getOpcode(); 754 switch (Opc) { 755 default: break; 756 case ISD::AssertSext: 757 return DAG.getNode(ISD::AssertSext, dl, PVT, 758 SExtPromoteOperand(Op.getOperand(0), PVT), 759 Op.getOperand(1)); 760 case ISD::AssertZext: 761 return DAG.getNode(ISD::AssertZext, dl, PVT, 762 ZExtPromoteOperand(Op.getOperand(0), PVT), 763 Op.getOperand(1)); 764 case ISD::Constant: { 765 unsigned ExtOpc = 766 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; 767 return DAG.getNode(ExtOpc, dl, PVT, Op); 768 } 769 } 770 771 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT)) 772 return SDValue(); 773 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op); 774} 775 776SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) { 777 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT)) 778 return SDValue(); 779 EVT OldVT = Op.getValueType(); 780 SDLoc dl(Op); 781 bool Replace = false; 782 SDValue NewOp = PromoteOperand(Op, PVT, Replace); 783 if (NewOp.getNode() == 0) 784 return SDValue(); 785 AddToWorkList(NewOp.getNode()); 786 787 if (Replace) 788 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); 789 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp, 790 DAG.getValueType(OldVT)); 791} 792 793SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) { 794 EVT OldVT = Op.getValueType(); 795 SDLoc dl(Op); 796 bool Replace = false; 797 SDValue NewOp = PromoteOperand(Op, PVT, Replace); 798 if (NewOp.getNode() == 0) 799 return SDValue(); 800 AddToWorkList(NewOp.getNode()); 801 802 if (Replace) 803 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode()); 804 return DAG.getZeroExtendInReg(NewOp, dl, OldVT); 805} 806 807/// PromoteIntBinOp - Promote the specified integer binary operation if the 808/// target indicates it is beneficial. e.g. On x86, it's usually better to 809/// promote i16 operations to i32 since i16 instructions are longer. 810SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) { 811 if (!LegalOperations) 812 return SDValue(); 813 814 EVT VT = Op.getValueType(); 815 if (VT.isVector() || !VT.isInteger()) 816 return SDValue(); 817 818 // If operation type is 'undesirable', e.g. i16 on x86, consider 819 // promoting it. 820 unsigned Opc = Op.getOpcode(); 821 if (TLI.isTypeDesirableForOp(Opc, VT)) 822 return SDValue(); 823 824 EVT PVT = VT; 825 // Consult target whether it is a good idea to promote this operation and 826 // what's the right type to promote it to. 827 if (TLI.IsDesirableToPromoteOp(Op, PVT)) { 828 assert(PVT != VT && "Don't know what type to promote to!"); 829 830 bool Replace0 = false; 831 SDValue N0 = Op.getOperand(0); 832 SDValue NN0 = PromoteOperand(N0, PVT, Replace0); 833 if (NN0.getNode() == 0) 834 return SDValue(); 835 836 bool Replace1 = false; 837 SDValue N1 = Op.getOperand(1); 838 SDValue NN1; 839 if (N0 == N1) 840 NN1 = NN0; 841 else { 842 NN1 = PromoteOperand(N1, PVT, Replace1); 843 if (NN1.getNode() == 0) 844 return SDValue(); 845 } 846 847 AddToWorkList(NN0.getNode()); 848 if (NN1.getNode()) 849 AddToWorkList(NN1.getNode()); 850 851 if (Replace0) 852 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode()); 853 if (Replace1) 854 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode()); 855 856 DEBUG(dbgs() << "\nPromoting "; 857 Op.getNode()->dump(&DAG)); 858 SDLoc dl(Op); 859 return DAG.getNode(ISD::TRUNCATE, dl, VT, 860 DAG.getNode(Opc, dl, PVT, NN0, NN1)); 861 } 862 return SDValue(); 863} 864 865/// PromoteIntShiftOp - Promote the specified integer shift operation if the 866/// target indicates it is beneficial. e.g. On x86, it's usually better to 867/// promote i16 operations to i32 since i16 instructions are longer. 868SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) { 869 if (!LegalOperations) 870 return SDValue(); 871 872 EVT VT = Op.getValueType(); 873 if (VT.isVector() || !VT.isInteger()) 874 return SDValue(); 875 876 // If operation type is 'undesirable', e.g. i16 on x86, consider 877 // promoting it. 878 unsigned Opc = Op.getOpcode(); 879 if (TLI.isTypeDesirableForOp(Opc, VT)) 880 return SDValue(); 881 882 EVT PVT = VT; 883 // Consult target whether it is a good idea to promote this operation and 884 // what's the right type to promote it to. 885 if (TLI.IsDesirableToPromoteOp(Op, PVT)) { 886 assert(PVT != VT && "Don't know what type to promote to!"); 887 888 bool Replace = false; 889 SDValue N0 = Op.getOperand(0); 890 if (Opc == ISD::SRA) 891 N0 = SExtPromoteOperand(Op.getOperand(0), PVT); 892 else if (Opc == ISD::SRL) 893 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT); 894 else 895 N0 = PromoteOperand(N0, PVT, Replace); 896 if (N0.getNode() == 0) 897 return SDValue(); 898 899 AddToWorkList(N0.getNode()); 900 if (Replace) 901 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode()); 902 903 DEBUG(dbgs() << "\nPromoting "; 904 Op.getNode()->dump(&DAG)); 905 SDLoc dl(Op); 906 return DAG.getNode(ISD::TRUNCATE, dl, VT, 907 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1))); 908 } 909 return SDValue(); 910} 911 912SDValue DAGCombiner::PromoteExtend(SDValue Op) { 913 if (!LegalOperations) 914 return SDValue(); 915 916 EVT VT = Op.getValueType(); 917 if (VT.isVector() || !VT.isInteger()) 918 return SDValue(); 919 920 // If operation type is 'undesirable', e.g. i16 on x86, consider 921 // promoting it. 922 unsigned Opc = Op.getOpcode(); 923 if (TLI.isTypeDesirableForOp(Opc, VT)) 924 return SDValue(); 925 926 EVT PVT = VT; 927 // Consult target whether it is a good idea to promote this operation and 928 // what's the right type to promote it to. 929 if (TLI.IsDesirableToPromoteOp(Op, PVT)) { 930 assert(PVT != VT && "Don't know what type to promote to!"); 931 // fold (aext (aext x)) -> (aext x) 932 // fold (aext (zext x)) -> (zext x) 933 // fold (aext (sext x)) -> (sext x) 934 DEBUG(dbgs() << "\nPromoting "; 935 Op.getNode()->dump(&DAG)); 936 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0)); 937 } 938 return SDValue(); 939} 940 941bool DAGCombiner::PromoteLoad(SDValue Op) { 942 if (!LegalOperations) 943 return false; 944 945 EVT VT = Op.getValueType(); 946 if (VT.isVector() || !VT.isInteger()) 947 return false; 948 949 // If operation type is 'undesirable', e.g. i16 on x86, consider 950 // promoting it. 951 unsigned Opc = Op.getOpcode(); 952 if (TLI.isTypeDesirableForOp(Opc, VT)) 953 return false; 954 955 EVT PVT = VT; 956 // Consult target whether it is a good idea to promote this operation and 957 // what's the right type to promote it to. 958 if (TLI.IsDesirableToPromoteOp(Op, PVT)) { 959 assert(PVT != VT && "Don't know what type to promote to!"); 960 961 SDLoc dl(Op); 962 SDNode *N = Op.getNode(); 963 LoadSDNode *LD = cast<LoadSDNode>(N); 964 EVT MemVT = LD->getMemoryVT(); 965 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD) 966 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD 967 : ISD::EXTLOAD) 968 : LD->getExtensionType(); 969 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT, 970 LD->getChain(), LD->getBasePtr(), 971 LD->getPointerInfo(), 972 MemVT, LD->isVolatile(), 973 LD->isNonTemporal(), LD->getAlignment()); 974 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD); 975 976 DEBUG(dbgs() << "\nPromoting "; 977 N->dump(&DAG); 978 dbgs() << "\nTo: "; 979 Result.getNode()->dump(&DAG); 980 dbgs() << '\n'); 981 WorkListRemover DeadNodes(*this); 982 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result); 983 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1)); 984 removeFromWorkList(N); 985 DAG.DeleteNode(N); 986 AddToWorkList(Result.getNode()); 987 return true; 988 } 989 return false; 990} 991 992 993//===----------------------------------------------------------------------===// 994// Main DAG Combiner implementation 995//===----------------------------------------------------------------------===// 996 997void DAGCombiner::Run(CombineLevel AtLevel) { 998 // set the instance variables, so that the various visit routines may use it. 999 Level = AtLevel; 1000 LegalOperations = Level >= AfterLegalizeVectorOps; 1001 LegalTypes = Level >= AfterLegalizeTypes; 1002 1003 // Add all the dag nodes to the worklist. 1004 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(), 1005 E = DAG.allnodes_end(); I != E; ++I) 1006 AddToWorkList(I); 1007 1008 // Create a dummy node (which is not added to allnodes), that adds a reference 1009 // to the root node, preventing it from being deleted, and tracking any 1010 // changes of the root. 1011 HandleSDNode Dummy(DAG.getRoot()); 1012 1013 // The root of the dag may dangle to deleted nodes until the dag combiner is 1014 // done. Set it to null to avoid confusion. 1015 DAG.setRoot(SDValue()); 1016 1017 // while the worklist isn't empty, find a node and 1018 // try and combine it. 1019 while (!WorkListContents.empty()) { 1020 SDNode *N; 1021 // The WorkListOrder holds the SDNodes in order, but it may contain duplicates. 1022 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the 1023 // worklist *should* contain, and check the node we want to visit is should 1024 // actually be visited. 1025 do { 1026 N = WorkListOrder.pop_back_val(); 1027 } while (!WorkListContents.erase(N)); 1028 1029 // If N has no uses, it is dead. Make sure to revisit all N's operands once 1030 // N is deleted from the DAG, since they too may now be dead or may have a 1031 // reduced number of uses, allowing other xforms. 1032 if (N->use_empty() && N != &Dummy) { 1033 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 1034 AddToWorkList(N->getOperand(i).getNode()); 1035 1036 DAG.DeleteNode(N); 1037 continue; 1038 } 1039 1040 SDValue RV = combine(N); 1041 1042 if (RV.getNode() == 0) 1043 continue; 1044 1045 ++NodesCombined; 1046 1047 // If we get back the same node we passed in, rather than a new node or 1048 // zero, we know that the node must have defined multiple values and 1049 // CombineTo was used. Since CombineTo takes care of the worklist 1050 // mechanics for us, we have no work to do in this case. 1051 if (RV.getNode() == N) 1052 continue; 1053 1054 assert(N->getOpcode() != ISD::DELETED_NODE && 1055 RV.getNode()->getOpcode() != ISD::DELETED_NODE && 1056 "Node was deleted but visit returned new node!"); 1057 1058 DEBUG(dbgs() << "\nReplacing.3 "; 1059 N->dump(&DAG); 1060 dbgs() << "\nWith: "; 1061 RV.getNode()->dump(&DAG); 1062 dbgs() << '\n'); 1063 1064 // Transfer debug value. 1065 DAG.TransferDbgValues(SDValue(N, 0), RV); 1066 WorkListRemover DeadNodes(*this); 1067 if (N->getNumValues() == RV.getNode()->getNumValues()) 1068 DAG.ReplaceAllUsesWith(N, RV.getNode()); 1069 else { 1070 assert(N->getValueType(0) == RV.getValueType() && 1071 N->getNumValues() == 1 && "Type mismatch"); 1072 SDValue OpV = RV; 1073 DAG.ReplaceAllUsesWith(N, &OpV); 1074 } 1075 1076 // Push the new node and any users onto the worklist 1077 AddToWorkList(RV.getNode()); 1078 AddUsersToWorkList(RV.getNode()); 1079 1080 // Add any uses of the old node to the worklist in case this node is the 1081 // last one that uses them. They may become dead after this node is 1082 // deleted. 1083 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 1084 AddToWorkList(N->getOperand(i).getNode()); 1085 1086 // Finally, if the node is now dead, remove it from the graph. The node 1087 // may not be dead if the replacement process recursively simplified to 1088 // something else needing this node. 1089 if (N->use_empty()) { 1090 // Nodes can be reintroduced into the worklist. Make sure we do not 1091 // process a node that has been replaced. 1092 removeFromWorkList(N); 1093 1094 // Finally, since the node is now dead, remove it from the graph. 1095 DAG.DeleteNode(N); 1096 } 1097 } 1098 1099 // If the root changed (e.g. it was a dead load, update the root). 1100 DAG.setRoot(Dummy.getValue()); 1101 DAG.RemoveDeadNodes(); 1102} 1103 1104SDValue DAGCombiner::visit(SDNode *N) { 1105 switch (N->getOpcode()) { 1106 default: break; 1107 case ISD::TokenFactor: return visitTokenFactor(N); 1108 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N); 1109 case ISD::ADD: return visitADD(N); 1110 case ISD::SUB: return visitSUB(N); 1111 case ISD::ADDC: return visitADDC(N); 1112 case ISD::SUBC: return visitSUBC(N); 1113 case ISD::ADDE: return visitADDE(N); 1114 case ISD::SUBE: return visitSUBE(N); 1115 case ISD::MUL: return visitMUL(N); 1116 case ISD::SDIV: return visitSDIV(N); 1117 case ISD::UDIV: return visitUDIV(N); 1118 case ISD::SREM: return visitSREM(N); 1119 case ISD::UREM: return visitUREM(N); 1120 case ISD::MULHU: return visitMULHU(N); 1121 case ISD::MULHS: return visitMULHS(N); 1122 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N); 1123 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N); 1124 case ISD::SMULO: return visitSMULO(N); 1125 case ISD::UMULO: return visitUMULO(N); 1126 case ISD::SDIVREM: return visitSDIVREM(N); 1127 case ISD::UDIVREM: return visitUDIVREM(N); 1128 case ISD::AND: return visitAND(N); 1129 case ISD::OR: return visitOR(N); 1130 case ISD::XOR: return visitXOR(N); 1131 case ISD::SHL: return visitSHL(N); 1132 case ISD::SRA: return visitSRA(N); 1133 case ISD::SRL: return visitSRL(N); 1134 case ISD::CTLZ: return visitCTLZ(N); 1135 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N); 1136 case ISD::CTTZ: return visitCTTZ(N); 1137 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N); 1138 case ISD::CTPOP: return visitCTPOP(N); 1139 case ISD::SELECT: return visitSELECT(N); 1140 case ISD::VSELECT: return visitVSELECT(N); 1141 case ISD::SELECT_CC: return visitSELECT_CC(N); 1142 case ISD::SETCC: return visitSETCC(N); 1143 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N); 1144 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N); 1145 case ISD::ANY_EXTEND: return visitANY_EXTEND(N); 1146 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N); 1147 case ISD::TRUNCATE: return visitTRUNCATE(N); 1148 case ISD::BITCAST: return visitBITCAST(N); 1149 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N); 1150 case ISD::FADD: return visitFADD(N); 1151 case ISD::FSUB: return visitFSUB(N); 1152 case ISD::FMUL: return visitFMUL(N); 1153 case ISD::FMA: return visitFMA(N); 1154 case ISD::FDIV: return visitFDIV(N); 1155 case ISD::FREM: return visitFREM(N); 1156 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N); 1157 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N); 1158 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N); 1159 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N); 1160 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N); 1161 case ISD::FP_ROUND: return visitFP_ROUND(N); 1162 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N); 1163 case ISD::FP_EXTEND: return visitFP_EXTEND(N); 1164 case ISD::FNEG: return visitFNEG(N); 1165 case ISD::FABS: return visitFABS(N); 1166 case ISD::FFLOOR: return visitFFLOOR(N); 1167 case ISD::FCEIL: return visitFCEIL(N); 1168 case ISD::FTRUNC: return visitFTRUNC(N); 1169 case ISD::BRCOND: return visitBRCOND(N); 1170 case ISD::BR_CC: return visitBR_CC(N); 1171 case ISD::LOAD: return visitLOAD(N); 1172 case ISD::STORE: return visitSTORE(N); 1173 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N); 1174 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N); 1175 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N); 1176 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N); 1177 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N); 1178 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N); 1179 } 1180 return SDValue(); 1181} 1182 1183SDValue DAGCombiner::combine(SDNode *N) { 1184 SDValue RV = visit(N); 1185 1186 // If nothing happened, try a target-specific DAG combine. 1187 if (RV.getNode() == 0) { 1188 assert(N->getOpcode() != ISD::DELETED_NODE && 1189 "Node was deleted but visit returned NULL!"); 1190 1191 if (N->getOpcode() >= ISD::BUILTIN_OP_END || 1192 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) { 1193 1194 // Expose the DAG combiner to the target combiner impls. 1195 TargetLowering::DAGCombinerInfo 1196 DagCombineInfo(DAG, Level, false, this); 1197 1198 RV = TLI.PerformDAGCombine(N, DagCombineInfo); 1199 } 1200 } 1201 1202 // If nothing happened still, try promoting the operation. 1203 if (RV.getNode() == 0) { 1204 switch (N->getOpcode()) { 1205 default: break; 1206 case ISD::ADD: 1207 case ISD::SUB: 1208 case ISD::MUL: 1209 case ISD::AND: 1210 case ISD::OR: 1211 case ISD::XOR: 1212 RV = PromoteIntBinOp(SDValue(N, 0)); 1213 break; 1214 case ISD::SHL: 1215 case ISD::SRA: 1216 case ISD::SRL: 1217 RV = PromoteIntShiftOp(SDValue(N, 0)); 1218 break; 1219 case ISD::SIGN_EXTEND: 1220 case ISD::ZERO_EXTEND: 1221 case ISD::ANY_EXTEND: 1222 RV = PromoteExtend(SDValue(N, 0)); 1223 break; 1224 case ISD::LOAD: 1225 if (PromoteLoad(SDValue(N, 0))) 1226 RV = SDValue(N, 0); 1227 break; 1228 } 1229 } 1230 1231 // If N is a commutative binary node, try commuting it to enable more 1232 // sdisel CSE. 1233 if (RV.getNode() == 0 && 1234 SelectionDAG::isCommutativeBinOp(N->getOpcode()) && 1235 N->getNumValues() == 1) { 1236 SDValue N0 = N->getOperand(0); 1237 SDValue N1 = N->getOperand(1); 1238 1239 // Constant operands are canonicalized to RHS. 1240 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) { 1241 SDValue Ops[] = { N1, N0 }; 1242 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), 1243 Ops, 2); 1244 if (CSENode) 1245 return SDValue(CSENode, 0); 1246 } 1247 } 1248 1249 return RV; 1250} 1251 1252/// getInputChainForNode - Given a node, return its input chain if it has one, 1253/// otherwise return a null sd operand. 1254static SDValue getInputChainForNode(SDNode *N) { 1255 if (unsigned NumOps = N->getNumOperands()) { 1256 if (N->getOperand(0).getValueType() == MVT::Other) 1257 return N->getOperand(0); 1258 if (N->getOperand(NumOps-1).getValueType() == MVT::Other) 1259 return N->getOperand(NumOps-1); 1260 for (unsigned i = 1; i < NumOps-1; ++i) 1261 if (N->getOperand(i).getValueType() == MVT::Other) 1262 return N->getOperand(i); 1263 } 1264 return SDValue(); 1265} 1266 1267SDValue DAGCombiner::visitTokenFactor(SDNode *N) { 1268 // If N has two operands, where one has an input chain equal to the other, 1269 // the 'other' chain is redundant. 1270 if (N->getNumOperands() == 2) { 1271 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1)) 1272 return N->getOperand(0); 1273 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0)) 1274 return N->getOperand(1); 1275 } 1276 1277 SmallVector<SDNode *, 8> TFs; // List of token factors to visit. 1278 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor. 1279 SmallPtrSet<SDNode*, 16> SeenOps; 1280 bool Changed = false; // If we should replace this token factor. 1281 1282 // Start out with this token factor. 1283 TFs.push_back(N); 1284 1285 // Iterate through token factors. The TFs grows when new token factors are 1286 // encountered. 1287 for (unsigned i = 0; i < TFs.size(); ++i) { 1288 SDNode *TF = TFs[i]; 1289 1290 // Check each of the operands. 1291 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) { 1292 SDValue Op = TF->getOperand(i); 1293 1294 switch (Op.getOpcode()) { 1295 case ISD::EntryToken: 1296 // Entry tokens don't need to be added to the list. They are 1297 // rededundant. 1298 Changed = true; 1299 break; 1300 1301 case ISD::TokenFactor: 1302 if (Op.hasOneUse() && 1303 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) { 1304 // Queue up for processing. 1305 TFs.push_back(Op.getNode()); 1306 // Clean up in case the token factor is removed. 1307 AddToWorkList(Op.getNode()); 1308 Changed = true; 1309 break; 1310 } 1311 // Fall thru 1312 1313 default: 1314 // Only add if it isn't already in the list. 1315 if (SeenOps.insert(Op.getNode())) 1316 Ops.push_back(Op); 1317 else 1318 Changed = true; 1319 break; 1320 } 1321 } 1322 } 1323 1324 SDValue Result; 1325 1326 // If we've change things around then replace token factor. 1327 if (Changed) { 1328 if (Ops.empty()) { 1329 // The entry token is the only possible outcome. 1330 Result = DAG.getEntryNode(); 1331 } else { 1332 // New and improved token factor. 1333 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), 1334 MVT::Other, &Ops[0], Ops.size()); 1335 } 1336 1337 // Don't add users to work list. 1338 return CombineTo(N, Result, false); 1339 } 1340 1341 return Result; 1342} 1343 1344/// MERGE_VALUES can always be eliminated. 1345SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) { 1346 WorkListRemover DeadNodes(*this); 1347 // Replacing results may cause a different MERGE_VALUES to suddenly 1348 // be CSE'd with N, and carry its uses with it. Iterate until no 1349 // uses remain, to ensure that the node can be safely deleted. 1350 // First add the users of this node to the work list so that they 1351 // can be tried again once they have new operands. 1352 AddUsersToWorkList(N); 1353 do { 1354 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 1355 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i)); 1356 } while (!N->use_empty()); 1357 removeFromWorkList(N); 1358 DAG.DeleteNode(N); 1359 return SDValue(N, 0); // Return N so it doesn't get rechecked! 1360} 1361 1362static 1363SDValue combineShlAddConstant(SDLoc DL, SDValue N0, SDValue N1, 1364 SelectionDAG &DAG) { 1365 EVT VT = N0.getValueType(); 1366 SDValue N00 = N0.getOperand(0); 1367 SDValue N01 = N0.getOperand(1); 1368 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01); 1369 1370 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() && 1371 isa<ConstantSDNode>(N00.getOperand(1))) { 1372 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), ) 1373 N0 = DAG.getNode(ISD::ADD, SDLoc(N0), VT, 1374 DAG.getNode(ISD::SHL, SDLoc(N00), VT, 1375 N00.getOperand(0), N01), 1376 DAG.getNode(ISD::SHL, SDLoc(N01), VT, 1377 N00.getOperand(1), N01)); 1378 return DAG.getNode(ISD::ADD, DL, VT, N0, N1); 1379 } 1380 1381 return SDValue(); 1382} 1383 1384SDValue DAGCombiner::visitADD(SDNode *N) { 1385 SDValue N0 = N->getOperand(0); 1386 SDValue N1 = N->getOperand(1); 1387 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1388 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1389 EVT VT = N0.getValueType(); 1390 1391 // fold vector ops 1392 if (VT.isVector()) { 1393 SDValue FoldedVOp = SimplifyVBinOp(N); 1394 if (FoldedVOp.getNode()) return FoldedVOp; 1395 1396 // fold (add x, 0) -> x, vector edition 1397 if (ISD::isBuildVectorAllZeros(N1.getNode())) 1398 return N0; 1399 if (ISD::isBuildVectorAllZeros(N0.getNode())) 1400 return N1; 1401 } 1402 1403 // fold (add x, undef) -> undef 1404 if (N0.getOpcode() == ISD::UNDEF) 1405 return N0; 1406 if (N1.getOpcode() == ISD::UNDEF) 1407 return N1; 1408 // fold (add c1, c2) -> c1+c2 1409 if (N0C && N1C) 1410 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C); 1411 // canonicalize constant to RHS 1412 if (N0C && !N1C) 1413 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0); 1414 // fold (add x, 0) -> x 1415 if (N1C && N1C->isNullValue()) 1416 return N0; 1417 // fold (add Sym, c) -> Sym+c 1418 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0)) 1419 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C && 1420 GA->getOpcode() == ISD::GlobalAddress) 1421 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT, 1422 GA->getOffset() + 1423 (uint64_t)N1C->getSExtValue()); 1424 // fold ((c1-A)+c2) -> (c1+c2)-A 1425 if (N1C && N0.getOpcode() == ISD::SUB) 1426 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0))) 1427 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1428 DAG.getConstant(N1C->getAPIntValue()+ 1429 N0C->getAPIntValue(), VT), 1430 N0.getOperand(1)); 1431 // reassociate add 1432 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1); 1433 if (RADD.getNode() != 0) 1434 return RADD; 1435 // fold ((0-A) + B) -> B-A 1436 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) && 1437 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue()) 1438 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1)); 1439 // fold (A + (0-B)) -> A-B 1440 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) && 1441 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue()) 1442 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1)); 1443 // fold (A+(B-A)) -> B 1444 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1)) 1445 return N1.getOperand(0); 1446 // fold ((B-A)+A) -> B 1447 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1)) 1448 return N0.getOperand(0); 1449 // fold (A+(B-(A+C))) to (B-C) 1450 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && 1451 N0 == N1.getOperand(1).getOperand(0)) 1452 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0), 1453 N1.getOperand(1).getOperand(1)); 1454 // fold (A+(B-(C+A))) to (B-C) 1455 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD && 1456 N0 == N1.getOperand(1).getOperand(1)) 1457 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0), 1458 N1.getOperand(1).getOperand(0)); 1459 // fold (A+((B-A)+or-C)) to (B+or-C) 1460 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) && 1461 N1.getOperand(0).getOpcode() == ISD::SUB && 1462 N0 == N1.getOperand(0).getOperand(1)) 1463 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT, 1464 N1.getOperand(0).getOperand(0), N1.getOperand(1)); 1465 1466 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant 1467 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) { 1468 SDValue N00 = N0.getOperand(0); 1469 SDValue N01 = N0.getOperand(1); 1470 SDValue N10 = N1.getOperand(0); 1471 SDValue N11 = N1.getOperand(1); 1472 1473 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10)) 1474 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1475 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10), 1476 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11)); 1477 } 1478 1479 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0))) 1480 return SDValue(N, 0); 1481 1482 // fold (a+b) -> (a|b) iff a and b share no bits. 1483 if (VT.isInteger() && !VT.isVector()) { 1484 APInt LHSZero, LHSOne; 1485 APInt RHSZero, RHSOne; 1486 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne); 1487 1488 if (LHSZero.getBoolValue()) { 1489 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne); 1490 1491 // If all possibly-set bits on the LHS are clear on the RHS, return an OR. 1492 // If all possibly-set bits on the RHS are clear on the LHS, return an OR. 1493 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero) 1494 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1); 1495 } 1496 } 1497 1498 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), ) 1499 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) { 1500 SDValue Result = combineShlAddConstant(SDLoc(N), N0, N1, DAG); 1501 if (Result.getNode()) return Result; 1502 } 1503 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) { 1504 SDValue Result = combineShlAddConstant(SDLoc(N), N1, N0, DAG); 1505 if (Result.getNode()) return Result; 1506 } 1507 1508 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n)) 1509 if (N1.getOpcode() == ISD::SHL && 1510 N1.getOperand(0).getOpcode() == ISD::SUB) 1511 if (ConstantSDNode *C = 1512 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0))) 1513 if (C->getAPIntValue() == 0) 1514 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, 1515 DAG.getNode(ISD::SHL, SDLoc(N), VT, 1516 N1.getOperand(0).getOperand(1), 1517 N1.getOperand(1))); 1518 if (N0.getOpcode() == ISD::SHL && 1519 N0.getOperand(0).getOpcode() == ISD::SUB) 1520 if (ConstantSDNode *C = 1521 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0))) 1522 if (C->getAPIntValue() == 0) 1523 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, 1524 DAG.getNode(ISD::SHL, SDLoc(N), VT, 1525 N0.getOperand(0).getOperand(1), 1526 N0.getOperand(1))); 1527 1528 if (N1.getOpcode() == ISD::AND) { 1529 SDValue AndOp0 = N1.getOperand(0); 1530 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1)); 1531 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0); 1532 unsigned DestBits = VT.getScalarType().getSizeInBits(); 1533 1534 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x)) 1535 // and similar xforms where the inner op is either ~0 or 0. 1536 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) { 1537 SDLoc DL(N); 1538 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0); 1539 } 1540 } 1541 1542 // add (sext i1), X -> sub X, (zext i1) 1543 if (N0.getOpcode() == ISD::SIGN_EXTEND && 1544 N0.getOperand(0).getValueType() == MVT::i1 && 1545 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) { 1546 SDLoc DL(N); 1547 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)); 1548 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt); 1549 } 1550 1551 return SDValue(); 1552} 1553 1554SDValue DAGCombiner::visitADDC(SDNode *N) { 1555 SDValue N0 = N->getOperand(0); 1556 SDValue N1 = N->getOperand(1); 1557 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1558 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1559 EVT VT = N0.getValueType(); 1560 1561 // If the flag result is dead, turn this into an ADD. 1562 if (!N->hasAnyUseOfValue(1)) 1563 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1), 1564 DAG.getNode(ISD::CARRY_FALSE, 1565 SDLoc(N), MVT::Glue)); 1566 1567 // canonicalize constant to RHS. 1568 if (N0C && !N1C) 1569 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0); 1570 1571 // fold (addc x, 0) -> x + no carry out 1572 if (N1C && N1C->isNullValue()) 1573 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, 1574 SDLoc(N), MVT::Glue)); 1575 1576 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits. 1577 APInt LHSZero, LHSOne; 1578 APInt RHSZero, RHSOne; 1579 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne); 1580 1581 if (LHSZero.getBoolValue()) { 1582 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne); 1583 1584 // If all possibly-set bits on the LHS are clear on the RHS, return an OR. 1585 // If all possibly-set bits on the RHS are clear on the LHS, return an OR. 1586 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero) 1587 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1), 1588 DAG.getNode(ISD::CARRY_FALSE, 1589 SDLoc(N), MVT::Glue)); 1590 } 1591 1592 return SDValue(); 1593} 1594 1595SDValue DAGCombiner::visitADDE(SDNode *N) { 1596 SDValue N0 = N->getOperand(0); 1597 SDValue N1 = N->getOperand(1); 1598 SDValue CarryIn = N->getOperand(2); 1599 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1600 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1601 1602 // canonicalize constant to RHS 1603 if (N0C && !N1C) 1604 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(), 1605 N1, N0, CarryIn); 1606 1607 // fold (adde x, y, false) -> (addc x, y) 1608 if (CarryIn.getOpcode() == ISD::CARRY_FALSE) 1609 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1); 1610 1611 return SDValue(); 1612} 1613 1614// Since it may not be valid to emit a fold to zero for vector initializers 1615// check if we can before folding. 1616static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT, 1617 SelectionDAG &DAG, 1618 bool LegalOperations, bool LegalTypes) { 1619 if (!VT.isVector()) 1620 return DAG.getConstant(0, VT); 1621 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) { 1622 // Produce a vector of zeros. 1623 EVT ElemTy = VT.getVectorElementType(); 1624 if (LegalTypes && TLI.getTypeAction(*DAG.getContext(), ElemTy) == 1625 TargetLowering::TypePromoteInteger) 1626 ElemTy = TLI.getTypeToTransformTo(*DAG.getContext(), ElemTy); 1627 assert((!LegalTypes || TLI.isTypeLegal(ElemTy)) && 1628 "Type for zero vector elements is not legal"); 1629 SDValue El = DAG.getConstant(0, ElemTy); 1630 std::vector<SDValue> Ops(VT.getVectorNumElements(), El); 1631 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, 1632 &Ops[0], Ops.size()); 1633 } 1634 return SDValue(); 1635} 1636 1637SDValue DAGCombiner::visitSUB(SDNode *N) { 1638 SDValue N0 = N->getOperand(0); 1639 SDValue N1 = N->getOperand(1); 1640 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode()); 1641 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 1642 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 : 1643 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode()); 1644 EVT VT = N0.getValueType(); 1645 1646 // fold vector ops 1647 if (VT.isVector()) { 1648 SDValue FoldedVOp = SimplifyVBinOp(N); 1649 if (FoldedVOp.getNode()) return FoldedVOp; 1650 1651 // fold (sub x, 0) -> x, vector edition 1652 if (ISD::isBuildVectorAllZeros(N1.getNode())) 1653 return N0; 1654 } 1655 1656 // fold (sub x, x) -> 0 1657 // FIXME: Refactor this and xor and other similar operations together. 1658 if (N0 == N1) 1659 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes); 1660 // fold (sub c1, c2) -> c1-c2 1661 if (N0C && N1C) 1662 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C); 1663 // fold (sub x, c) -> (add x, -c) 1664 if (N1C) 1665 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, 1666 DAG.getConstant(-N1C->getAPIntValue(), VT)); 1667 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) 1668 if (N0C && N0C->isAllOnesValue()) 1669 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0); 1670 // fold A-(A-B) -> B 1671 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0)) 1672 return N1.getOperand(1); 1673 // fold (A+B)-A -> B 1674 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1) 1675 return N0.getOperand(1); 1676 // fold (A+B)-B -> A 1677 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1) 1678 return N0.getOperand(0); 1679 // fold C2-(A+C1) -> (C2-C1)-A 1680 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) { 1681 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(), 1682 VT); 1683 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC, 1684 N1.getOperand(0)); 1685 } 1686 // fold ((A+(B+or-C))-B) -> A+or-C 1687 if (N0.getOpcode() == ISD::ADD && 1688 (N0.getOperand(1).getOpcode() == ISD::SUB || 1689 N0.getOperand(1).getOpcode() == ISD::ADD) && 1690 N0.getOperand(1).getOperand(0) == N1) 1691 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT, 1692 N0.getOperand(0), N0.getOperand(1).getOperand(1)); 1693 // fold ((A+(C+B))-B) -> A+C 1694 if (N0.getOpcode() == ISD::ADD && 1695 N0.getOperand(1).getOpcode() == ISD::ADD && 1696 N0.getOperand(1).getOperand(1) == N1) 1697 return DAG.getNode(ISD::ADD, SDLoc(N), VT, 1698 N0.getOperand(0), N0.getOperand(1).getOperand(0)); 1699 // fold ((A-(B-C))-C) -> A-B 1700 if (N0.getOpcode() == ISD::SUB && 1701 N0.getOperand(1).getOpcode() == ISD::SUB && 1702 N0.getOperand(1).getOperand(1) == N1) 1703 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1704 N0.getOperand(0), N0.getOperand(1).getOperand(0)); 1705 1706 // If either operand of a sub is undef, the result is undef 1707 if (N0.getOpcode() == ISD::UNDEF) 1708 return N0; 1709 if (N1.getOpcode() == ISD::UNDEF) 1710 return N1; 1711 1712 // If the relocation model supports it, consider symbol offsets. 1713 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0)) 1714 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) { 1715 // fold (sub Sym, c) -> Sym-c 1716 if (N1C && GA->getOpcode() == ISD::GlobalAddress) 1717 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT, 1718 GA->getOffset() - 1719 (uint64_t)N1C->getSExtValue()); 1720 // fold (sub Sym+c1, Sym+c2) -> c1-c2 1721 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1)) 1722 if (GA->getGlobal() == GB->getGlobal()) 1723 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(), 1724 VT); 1725 } 1726 1727 return SDValue(); 1728} 1729 1730SDValue DAGCombiner::visitSUBC(SDNode *N) { 1731 SDValue N0 = N->getOperand(0); 1732 SDValue N1 = N->getOperand(1); 1733 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1734 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1735 EVT VT = N0.getValueType(); 1736 1737 // If the flag result is dead, turn this into an SUB. 1738 if (!N->hasAnyUseOfValue(1)) 1739 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1), 1740 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N), 1741 MVT::Glue)); 1742 1743 // fold (subc x, x) -> 0 + no borrow 1744 if (N0 == N1) 1745 return CombineTo(N, DAG.getConstant(0, VT), 1746 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N), 1747 MVT::Glue)); 1748 1749 // fold (subc x, 0) -> x + no borrow 1750 if (N1C && N1C->isNullValue()) 1751 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N), 1752 MVT::Glue)); 1753 1754 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow 1755 if (N0C && N0C->isAllOnesValue()) 1756 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0), 1757 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N), 1758 MVT::Glue)); 1759 1760 return SDValue(); 1761} 1762 1763SDValue DAGCombiner::visitSUBE(SDNode *N) { 1764 SDValue N0 = N->getOperand(0); 1765 SDValue N1 = N->getOperand(1); 1766 SDValue CarryIn = N->getOperand(2); 1767 1768 // fold (sube x, y, false) -> (subc x, y) 1769 if (CarryIn.getOpcode() == ISD::CARRY_FALSE) 1770 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1); 1771 1772 return SDValue(); 1773} 1774 1775/// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are 1776/// all the same constant or undefined. 1777static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) { 1778 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N); 1779 if (!C) 1780 return false; 1781 1782 APInt SplatUndef; 1783 unsigned SplatBitSize; 1784 bool HasAnyUndefs; 1785 EVT EltVT = N->getValueType(0).getVectorElementType(); 1786 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, 1787 HasAnyUndefs) && 1788 EltVT.getSizeInBits() >= SplatBitSize); 1789} 1790 1791SDValue DAGCombiner::visitMUL(SDNode *N) { 1792 SDValue N0 = N->getOperand(0); 1793 SDValue N1 = N->getOperand(1); 1794 EVT VT = N0.getValueType(); 1795 1796 // fold (mul x, undef) -> 0 1797 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) 1798 return DAG.getConstant(0, VT); 1799 1800 bool N0IsConst = false; 1801 bool N1IsConst = false; 1802 APInt ConstValue0, ConstValue1; 1803 // fold vector ops 1804 if (VT.isVector()) { 1805 SDValue FoldedVOp = SimplifyVBinOp(N); 1806 if (FoldedVOp.getNode()) return FoldedVOp; 1807 1808 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0); 1809 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1); 1810 } else { 1811 N0IsConst = dyn_cast<ConstantSDNode>(N0) != 0; 1812 ConstValue0 = N0IsConst? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue() : APInt(); 1813 N1IsConst = dyn_cast<ConstantSDNode>(N1) != 0; 1814 ConstValue1 = N1IsConst? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue() : APInt(); 1815 } 1816 1817 // fold (mul c1, c2) -> c1*c2 1818 if (N0IsConst && N1IsConst) 1819 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode()); 1820 1821 // canonicalize constant to RHS 1822 if (N0IsConst && !N1IsConst) 1823 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0); 1824 // fold (mul x, 0) -> 0 1825 if (N1IsConst && ConstValue1 == 0) 1826 return N1; 1827 // We require a splat of the entire scalar bit width for non-contiguous 1828 // bit patterns. 1829 bool IsFullSplat = 1830 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits(); 1831 // fold (mul x, 1) -> x 1832 if (N1IsConst && ConstValue1 == 1 && IsFullSplat) 1833 return N0; 1834 // fold (mul x, -1) -> 0-x 1835 if (N1IsConst && ConstValue1.isAllOnesValue()) 1836 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1837 DAG.getConstant(0, VT), N0); 1838 // fold (mul x, (1 << c)) -> x << c 1839 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat) 1840 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, 1841 DAG.getConstant(ConstValue1.logBase2(), 1842 getShiftAmountTy(N0.getValueType()))); 1843 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c 1844 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) { 1845 unsigned Log2Val = (-ConstValue1).logBase2(); 1846 // FIXME: If the input is something that is easily negated (e.g. a 1847 // single-use add), we should put the negate there. 1848 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1849 DAG.getConstant(0, VT), 1850 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, 1851 DAG.getConstant(Log2Val, 1852 getShiftAmountTy(N0.getValueType())))); 1853 } 1854 1855 APInt Val; 1856 // (mul (shl X, c1), c2) -> (mul X, c2 << c1) 1857 if (N1IsConst && N0.getOpcode() == ISD::SHL && 1858 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) || 1859 isa<ConstantSDNode>(N0.getOperand(1)))) { 1860 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT, 1861 N1, N0.getOperand(1)); 1862 AddToWorkList(C3.getNode()); 1863 return DAG.getNode(ISD::MUL, SDLoc(N), VT, 1864 N0.getOperand(0), C3); 1865 } 1866 1867 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one 1868 // use. 1869 { 1870 SDValue Sh(0,0), Y(0,0); 1871 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)). 1872 if (N0.getOpcode() == ISD::SHL && 1873 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) || 1874 isa<ConstantSDNode>(N0.getOperand(1))) && 1875 N0.getNode()->hasOneUse()) { 1876 Sh = N0; Y = N1; 1877 } else if (N1.getOpcode() == ISD::SHL && 1878 isa<ConstantSDNode>(N1.getOperand(1)) && 1879 N1.getNode()->hasOneUse()) { 1880 Sh = N1; Y = N0; 1881 } 1882 1883 if (Sh.getNode()) { 1884 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, 1885 Sh.getOperand(0), Y); 1886 return DAG.getNode(ISD::SHL, SDLoc(N), VT, 1887 Mul, Sh.getOperand(1)); 1888 } 1889 } 1890 1891 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2) 1892 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() && 1893 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) || 1894 isa<ConstantSDNode>(N0.getOperand(1)))) 1895 return DAG.getNode(ISD::ADD, SDLoc(N), VT, 1896 DAG.getNode(ISD::MUL, SDLoc(N0), VT, 1897 N0.getOperand(0), N1), 1898 DAG.getNode(ISD::MUL, SDLoc(N1), VT, 1899 N0.getOperand(1), N1)); 1900 1901 // reassociate mul 1902 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1); 1903 if (RMUL.getNode() != 0) 1904 return RMUL; 1905 1906 return SDValue(); 1907} 1908 1909SDValue DAGCombiner::visitSDIV(SDNode *N) { 1910 SDValue N0 = N->getOperand(0); 1911 SDValue N1 = N->getOperand(1); 1912 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode()); 1913 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 1914 EVT VT = N->getValueType(0); 1915 1916 // fold vector ops 1917 if (VT.isVector()) { 1918 SDValue FoldedVOp = SimplifyVBinOp(N); 1919 if (FoldedVOp.getNode()) return FoldedVOp; 1920 } 1921 1922 // fold (sdiv c1, c2) -> c1/c2 1923 if (N0C && N1C && !N1C->isNullValue()) 1924 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C); 1925 // fold (sdiv X, 1) -> X 1926 if (N1C && N1C->getAPIntValue() == 1LL) 1927 return N0; 1928 // fold (sdiv X, -1) -> 0-X 1929 if (N1C && N1C->isAllOnesValue()) 1930 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1931 DAG.getConstant(0, VT), N0); 1932 // If we know the sign bits of both operands are zero, strength reduce to a 1933 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2 1934 if (!VT.isVector()) { 1935 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) 1936 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(), 1937 N0, N1); 1938 } 1939 // fold (sdiv X, pow2) -> simple ops after legalize 1940 if (N1C && !N1C->isNullValue() && 1941 (N1C->getAPIntValue().isPowerOf2() || 1942 (-N1C->getAPIntValue()).isPowerOf2())) { 1943 // If dividing by powers of two is cheap, then don't perform the following 1944 // fold. 1945 if (TLI.isPow2DivCheap()) 1946 return SDValue(); 1947 1948 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros(); 1949 1950 // Splat the sign bit into the register 1951 SDValue SGN = DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, 1952 DAG.getConstant(VT.getSizeInBits()-1, 1953 getShiftAmountTy(N0.getValueType()))); 1954 AddToWorkList(SGN.getNode()); 1955 1956 // Add (N0 < 0) ? abs2 - 1 : 0; 1957 SDValue SRL = DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN, 1958 DAG.getConstant(VT.getSizeInBits() - lg2, 1959 getShiftAmountTy(SGN.getValueType()))); 1960 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL); 1961 AddToWorkList(SRL.getNode()); 1962 AddToWorkList(ADD.getNode()); // Divide by pow2 1963 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD, 1964 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType()))); 1965 1966 // If we're dividing by a positive value, we're done. Otherwise, we must 1967 // negate the result. 1968 if (N1C->getAPIntValue().isNonNegative()) 1969 return SRA; 1970 1971 AddToWorkList(SRA.getNode()); 1972 return DAG.getNode(ISD::SUB, SDLoc(N), VT, 1973 DAG.getConstant(0, VT), SRA); 1974 } 1975 1976 // if integer divide is expensive and we satisfy the requirements, emit an 1977 // alternate sequence. 1978 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) { 1979 SDValue Op = BuildSDIV(N); 1980 if (Op.getNode()) return Op; 1981 } 1982 1983 // undef / X -> 0 1984 if (N0.getOpcode() == ISD::UNDEF) 1985 return DAG.getConstant(0, VT); 1986 // X / undef -> undef 1987 if (N1.getOpcode() == ISD::UNDEF) 1988 return N1; 1989 1990 return SDValue(); 1991} 1992 1993SDValue DAGCombiner::visitUDIV(SDNode *N) { 1994 SDValue N0 = N->getOperand(0); 1995 SDValue N1 = N->getOperand(1); 1996 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode()); 1997 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 1998 EVT VT = N->getValueType(0); 1999 2000 // fold vector ops 2001 if (VT.isVector()) { 2002 SDValue FoldedVOp = SimplifyVBinOp(N); 2003 if (FoldedVOp.getNode()) return FoldedVOp; 2004 } 2005 2006 // fold (udiv c1, c2) -> c1/c2 2007 if (N0C && N1C && !N1C->isNullValue()) 2008 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C); 2009 // fold (udiv x, (1 << c)) -> x >>u c 2010 if (N1C && N1C->getAPIntValue().isPowerOf2()) 2011 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, 2012 DAG.getConstant(N1C->getAPIntValue().logBase2(), 2013 getShiftAmountTy(N0.getValueType()))); 2014 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2 2015 if (N1.getOpcode() == ISD::SHL) { 2016 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) { 2017 if (SHC->getAPIntValue().isPowerOf2()) { 2018 EVT ADDVT = N1.getOperand(1).getValueType(); 2019 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT, 2020 N1.getOperand(1), 2021 DAG.getConstant(SHC->getAPIntValue() 2022 .logBase2(), 2023 ADDVT)); 2024 AddToWorkList(Add.getNode()); 2025 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add); 2026 } 2027 } 2028 } 2029 // fold (udiv x, c) -> alternate 2030 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) { 2031 SDValue Op = BuildUDIV(N); 2032 if (Op.getNode()) return Op; 2033 } 2034 2035 // undef / X -> 0 2036 if (N0.getOpcode() == ISD::UNDEF) 2037 return DAG.getConstant(0, VT); 2038 // X / undef -> undef 2039 if (N1.getOpcode() == ISD::UNDEF) 2040 return N1; 2041 2042 return SDValue(); 2043} 2044 2045SDValue DAGCombiner::visitSREM(SDNode *N) { 2046 SDValue N0 = N->getOperand(0); 2047 SDValue N1 = N->getOperand(1); 2048 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2049 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2050 EVT VT = N->getValueType(0); 2051 2052 // fold (srem c1, c2) -> c1%c2 2053 if (N0C && N1C && !N1C->isNullValue()) 2054 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C); 2055 // If we know the sign bits of both operands are zero, strength reduce to a 2056 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15 2057 if (!VT.isVector()) { 2058 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0)) 2059 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1); 2060 } 2061 2062 // If X/C can be simplified by the division-by-constant logic, lower 2063 // X%C to the equivalent of X-X/C*C. 2064 if (N1C && !N1C->isNullValue()) { 2065 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1); 2066 AddToWorkList(Div.getNode()); 2067 SDValue OptimizedDiv = combine(Div.getNode()); 2068 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) { 2069 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, 2070 OptimizedDiv, N1); 2071 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul); 2072 AddToWorkList(Mul.getNode()); 2073 return Sub; 2074 } 2075 } 2076 2077 // undef % X -> 0 2078 if (N0.getOpcode() == ISD::UNDEF) 2079 return DAG.getConstant(0, VT); 2080 // X % undef -> undef 2081 if (N1.getOpcode() == ISD::UNDEF) 2082 return N1; 2083 2084 return SDValue(); 2085} 2086 2087SDValue DAGCombiner::visitUREM(SDNode *N) { 2088 SDValue N0 = N->getOperand(0); 2089 SDValue N1 = N->getOperand(1); 2090 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2091 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2092 EVT VT = N->getValueType(0); 2093 2094 // fold (urem c1, c2) -> c1%c2 2095 if (N0C && N1C && !N1C->isNullValue()) 2096 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C); 2097 // fold (urem x, pow2) -> (and x, pow2-1) 2098 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2()) 2099 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, 2100 DAG.getConstant(N1C->getAPIntValue()-1,VT)); 2101 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1)) 2102 if (N1.getOpcode() == ISD::SHL) { 2103 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) { 2104 if (SHC->getAPIntValue().isPowerOf2()) { 2105 SDValue Add = 2106 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, 2107 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), 2108 VT)); 2109 AddToWorkList(Add.getNode()); 2110 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add); 2111 } 2112 } 2113 } 2114 2115 // If X/C can be simplified by the division-by-constant logic, lower 2116 // X%C to the equivalent of X-X/C*C. 2117 if (N1C && !N1C->isNullValue()) { 2118 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1); 2119 AddToWorkList(Div.getNode()); 2120 SDValue OptimizedDiv = combine(Div.getNode()); 2121 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) { 2122 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT, 2123 OptimizedDiv, N1); 2124 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul); 2125 AddToWorkList(Mul.getNode()); 2126 return Sub; 2127 } 2128 } 2129 2130 // undef % X -> 0 2131 if (N0.getOpcode() == ISD::UNDEF) 2132 return DAG.getConstant(0, VT); 2133 // X % undef -> undef 2134 if (N1.getOpcode() == ISD::UNDEF) 2135 return N1; 2136 2137 return SDValue(); 2138} 2139 2140SDValue DAGCombiner::visitMULHS(SDNode *N) { 2141 SDValue N0 = N->getOperand(0); 2142 SDValue N1 = N->getOperand(1); 2143 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2144 EVT VT = N->getValueType(0); 2145 SDLoc DL(N); 2146 2147 // fold (mulhs x, 0) -> 0 2148 if (N1C && N1C->isNullValue()) 2149 return N1; 2150 // fold (mulhs x, 1) -> (sra x, size(x)-1) 2151 if (N1C && N1C->getAPIntValue() == 1) 2152 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0, 2153 DAG.getConstant(N0.getValueType().getSizeInBits() - 1, 2154 getShiftAmountTy(N0.getValueType()))); 2155 // fold (mulhs x, undef) -> 0 2156 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) 2157 return DAG.getConstant(0, VT); 2158 2159 // If the type twice as wide is legal, transform the mulhs to a wider multiply 2160 // plus a shift. 2161 if (VT.isSimple() && !VT.isVector()) { 2162 MVT Simple = VT.getSimpleVT(); 2163 unsigned SimpleSize = Simple.getSizeInBits(); 2164 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); 2165 if (TLI.isOperationLegal(ISD::MUL, NewVT)) { 2166 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0); 2167 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1); 2168 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); 2169 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, 2170 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType()))); 2171 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); 2172 } 2173 } 2174 2175 return SDValue(); 2176} 2177 2178SDValue DAGCombiner::visitMULHU(SDNode *N) { 2179 SDValue N0 = N->getOperand(0); 2180 SDValue N1 = N->getOperand(1); 2181 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2182 EVT VT = N->getValueType(0); 2183 SDLoc DL(N); 2184 2185 // fold (mulhu x, 0) -> 0 2186 if (N1C && N1C->isNullValue()) 2187 return N1; 2188 // fold (mulhu x, 1) -> 0 2189 if (N1C && N1C->getAPIntValue() == 1) 2190 return DAG.getConstant(0, N0.getValueType()); 2191 // fold (mulhu x, undef) -> 0 2192 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) 2193 return DAG.getConstant(0, VT); 2194 2195 // If the type twice as wide is legal, transform the mulhu to a wider multiply 2196 // plus a shift. 2197 if (VT.isSimple() && !VT.isVector()) { 2198 MVT Simple = VT.getSimpleVT(); 2199 unsigned SimpleSize = Simple.getSizeInBits(); 2200 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); 2201 if (TLI.isOperationLegal(ISD::MUL, NewVT)) { 2202 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0); 2203 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1); 2204 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1); 2205 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1, 2206 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType()))); 2207 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1); 2208 } 2209 } 2210 2211 return SDValue(); 2212} 2213 2214/// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that 2215/// compute two values. LoOp and HiOp give the opcodes for the two computations 2216/// that are being performed. Return true if a simplification was made. 2217/// 2218SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp, 2219 unsigned HiOp) { 2220 // If the high half is not needed, just compute the low half. 2221 bool HiExists = N->hasAnyUseOfValue(1); 2222 if (!HiExists && 2223 (!LegalOperations || 2224 TLI.isOperationLegal(LoOp, N->getValueType(0)))) { 2225 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), 2226 N->op_begin(), N->getNumOperands()); 2227 return CombineTo(N, Res, Res); 2228 } 2229 2230 // If the low half is not needed, just compute the high half. 2231 bool LoExists = N->hasAnyUseOfValue(0); 2232 if (!LoExists && 2233 (!LegalOperations || 2234 TLI.isOperationLegal(HiOp, N->getValueType(1)))) { 2235 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), 2236 N->op_begin(), N->getNumOperands()); 2237 return CombineTo(N, Res, Res); 2238 } 2239 2240 // If both halves are used, return as it is. 2241 if (LoExists && HiExists) 2242 return SDValue(); 2243 2244 // If the two computed results can be simplified separately, separate them. 2245 if (LoExists) { 2246 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), 2247 N->op_begin(), N->getNumOperands()); 2248 AddToWorkList(Lo.getNode()); 2249 SDValue LoOpt = combine(Lo.getNode()); 2250 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() && 2251 (!LegalOperations || 2252 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType()))) 2253 return CombineTo(N, LoOpt, LoOpt); 2254 } 2255 2256 if (HiExists) { 2257 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), 2258 N->op_begin(), N->getNumOperands()); 2259 AddToWorkList(Hi.getNode()); 2260 SDValue HiOpt = combine(Hi.getNode()); 2261 if (HiOpt.getNode() && HiOpt != Hi && 2262 (!LegalOperations || 2263 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType()))) 2264 return CombineTo(N, HiOpt, HiOpt); 2265 } 2266 2267 return SDValue(); 2268} 2269 2270SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) { 2271 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS); 2272 if (Res.getNode()) return Res; 2273 2274 EVT VT = N->getValueType(0); 2275 SDLoc DL(N); 2276 2277 // If the type twice as wide is legal, transform the mulhu to a wider multiply 2278 // plus a shift. 2279 if (VT.isSimple() && !VT.isVector()) { 2280 MVT Simple = VT.getSimpleVT(); 2281 unsigned SimpleSize = Simple.getSizeInBits(); 2282 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); 2283 if (TLI.isOperationLegal(ISD::MUL, NewVT)) { 2284 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0)); 2285 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1)); 2286 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); 2287 // Compute the high part as N1. 2288 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, 2289 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType()))); 2290 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); 2291 // Compute the low part as N0. 2292 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); 2293 return CombineTo(N, Lo, Hi); 2294 } 2295 } 2296 2297 return SDValue(); 2298} 2299 2300SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) { 2301 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU); 2302 if (Res.getNode()) return Res; 2303 2304 EVT VT = N->getValueType(0); 2305 SDLoc DL(N); 2306 2307 // If the type twice as wide is legal, transform the mulhu to a wider multiply 2308 // plus a shift. 2309 if (VT.isSimple() && !VT.isVector()) { 2310 MVT Simple = VT.getSimpleVT(); 2311 unsigned SimpleSize = Simple.getSizeInBits(); 2312 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2); 2313 if (TLI.isOperationLegal(ISD::MUL, NewVT)) { 2314 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0)); 2315 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1)); 2316 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi); 2317 // Compute the high part as N1. 2318 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo, 2319 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType()))); 2320 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi); 2321 // Compute the low part as N0. 2322 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo); 2323 return CombineTo(N, Lo, Hi); 2324 } 2325 } 2326 2327 return SDValue(); 2328} 2329 2330SDValue DAGCombiner::visitSMULO(SDNode *N) { 2331 // (smulo x, 2) -> (saddo x, x) 2332 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1))) 2333 if (C2->getAPIntValue() == 2) 2334 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(), 2335 N->getOperand(0), N->getOperand(0)); 2336 2337 return SDValue(); 2338} 2339 2340SDValue DAGCombiner::visitUMULO(SDNode *N) { 2341 // (umulo x, 2) -> (uaddo x, x) 2342 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1))) 2343 if (C2->getAPIntValue() == 2) 2344 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(), 2345 N->getOperand(0), N->getOperand(0)); 2346 2347 return SDValue(); 2348} 2349 2350SDValue DAGCombiner::visitSDIVREM(SDNode *N) { 2351 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM); 2352 if (Res.getNode()) return Res; 2353 2354 return SDValue(); 2355} 2356 2357SDValue DAGCombiner::visitUDIVREM(SDNode *N) { 2358 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM); 2359 if (Res.getNode()) return Res; 2360 2361 return SDValue(); 2362} 2363 2364/// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with 2365/// two operands of the same opcode, try to simplify it. 2366SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) { 2367 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 2368 EVT VT = N0.getValueType(); 2369 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!"); 2370 2371 // Bail early if none of these transforms apply. 2372 if (N0.getNode()->getNumOperands() == 0) return SDValue(); 2373 2374 // For each of OP in AND/OR/XOR: 2375 // fold (OP (zext x), (zext y)) -> (zext (OP x, y)) 2376 // fold (OP (sext x), (sext y)) -> (sext (OP x, y)) 2377 // fold (OP (aext x), (aext y)) -> (aext (OP x, y)) 2378 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free) 2379 // 2380 // do not sink logical op inside of a vector extend, since it may combine 2381 // into a vsetcc. 2382 EVT Op0VT = N0.getOperand(0).getValueType(); 2383 if ((N0.getOpcode() == ISD::ZERO_EXTEND || 2384 N0.getOpcode() == ISD::SIGN_EXTEND || 2385 // Avoid infinite looping with PromoteIntBinOp. 2386 (N0.getOpcode() == ISD::ANY_EXTEND && 2387 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) || 2388 (N0.getOpcode() == ISD::TRUNCATE && 2389 (!TLI.isZExtFree(VT, Op0VT) || 2390 !TLI.isTruncateFree(Op0VT, VT)) && 2391 TLI.isTypeLegal(Op0VT))) && 2392 !VT.isVector() && 2393 Op0VT == N1.getOperand(0).getValueType() && 2394 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) { 2395 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0), 2396 N0.getOperand(0).getValueType(), 2397 N0.getOperand(0), N1.getOperand(0)); 2398 AddToWorkList(ORNode.getNode()); 2399 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode); 2400 } 2401 2402 // For each of OP in SHL/SRL/SRA/AND... 2403 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z) 2404 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z) 2405 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z) 2406 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL || 2407 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) && 2408 N0.getOperand(1) == N1.getOperand(1)) { 2409 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0), 2410 N0.getOperand(0).getValueType(), 2411 N0.getOperand(0), N1.getOperand(0)); 2412 AddToWorkList(ORNode.getNode()); 2413 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, 2414 ORNode, N0.getOperand(1)); 2415 } 2416 2417 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B)) 2418 // Only perform this optimization after type legalization and before 2419 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by 2420 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and 2421 // we don't want to undo this promotion. 2422 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper 2423 // on scalars. 2424 if ((N0.getOpcode() == ISD::BITCAST || 2425 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) && 2426 Level == AfterLegalizeTypes) { 2427 SDValue In0 = N0.getOperand(0); 2428 SDValue In1 = N1.getOperand(0); 2429 EVT In0Ty = In0.getValueType(); 2430 EVT In1Ty = In1.getValueType(); 2431 SDLoc DL(N); 2432 // If both incoming values are integers, and the original types are the 2433 // same. 2434 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) { 2435 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1); 2436 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op); 2437 AddToWorkList(Op.getNode()); 2438 return BC; 2439 } 2440 } 2441 2442 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value). 2443 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B)) 2444 // If both shuffles use the same mask, and both shuffle within a single 2445 // vector, then it is worthwhile to move the swizzle after the operation. 2446 // The type-legalizer generates this pattern when loading illegal 2447 // vector types from memory. In many cases this allows additional shuffle 2448 // optimizations. 2449 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG && 2450 N0.getOperand(1).getOpcode() == ISD::UNDEF && 2451 N1.getOperand(1).getOpcode() == ISD::UNDEF) { 2452 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0); 2453 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1); 2454 2455 assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() && 2456 "Inputs to shuffles are not the same type"); 2457 2458 unsigned NumElts = VT.getVectorNumElements(); 2459 2460 // Check that both shuffles use the same mask. The masks are known to be of 2461 // the same length because the result vector type is the same. 2462 bool SameMask = true; 2463 for (unsigned i = 0; i != NumElts; ++i) { 2464 int Idx0 = SVN0->getMaskElt(i); 2465 int Idx1 = SVN1->getMaskElt(i); 2466 if (Idx0 != Idx1) { 2467 SameMask = false; 2468 break; 2469 } 2470 } 2471 2472 if (SameMask) { 2473 SDValue Op = DAG.getNode(N->getOpcode(), SDLoc(N), VT, 2474 N0.getOperand(0), N1.getOperand(0)); 2475 AddToWorkList(Op.getNode()); 2476 return DAG.getVectorShuffle(VT, SDLoc(N), Op, 2477 DAG.getUNDEF(VT), &SVN0->getMask()[0]); 2478 } 2479 } 2480 2481 return SDValue(); 2482} 2483 2484SDValue DAGCombiner::visitAND(SDNode *N) { 2485 SDValue N0 = N->getOperand(0); 2486 SDValue N1 = N->getOperand(1); 2487 SDValue LL, LR, RL, RR, CC0, CC1; 2488 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2489 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2490 EVT VT = N1.getValueType(); 2491 unsigned BitWidth = VT.getScalarType().getSizeInBits(); 2492 2493 // fold vector ops 2494 if (VT.isVector()) { 2495 SDValue FoldedVOp = SimplifyVBinOp(N); 2496 if (FoldedVOp.getNode()) return FoldedVOp; 2497 2498 // fold (and x, 0) -> 0, vector edition 2499 if (ISD::isBuildVectorAllZeros(N0.getNode())) 2500 return N0; 2501 if (ISD::isBuildVectorAllZeros(N1.getNode())) 2502 return N1; 2503 2504 // fold (and x, -1) -> x, vector edition 2505 if (ISD::isBuildVectorAllOnes(N0.getNode())) 2506 return N1; 2507 if (ISD::isBuildVectorAllOnes(N1.getNode())) 2508 return N0; 2509 } 2510 2511 // fold (and x, undef) -> 0 2512 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF) 2513 return DAG.getConstant(0, VT); 2514 // fold (and c1, c2) -> c1&c2 2515 if (N0C && N1C) 2516 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C); 2517 // canonicalize constant to RHS 2518 if (N0C && !N1C) 2519 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0); 2520 // fold (and x, -1) -> x 2521 if (N1C && N1C->isAllOnesValue()) 2522 return N0; 2523 // if (and x, c) is known to be zero, return 0 2524 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), 2525 APInt::getAllOnesValue(BitWidth))) 2526 return DAG.getConstant(0, VT); 2527 // reassociate and 2528 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1); 2529 if (RAND.getNode() != 0) 2530 return RAND; 2531 // fold (and (or x, C), D) -> D if (C & D) == D 2532 if (N1C && N0.getOpcode() == ISD::OR) 2533 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 2534 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue()) 2535 return N1; 2536 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits. 2537 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { 2538 SDValue N0Op0 = N0.getOperand(0); 2539 APInt Mask = ~N1C->getAPIntValue(); 2540 Mask = Mask.trunc(N0Op0.getValueSizeInBits()); 2541 if (DAG.MaskedValueIsZero(N0Op0, Mask)) { 2542 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), 2543 N0.getValueType(), N0Op0); 2544 2545 // Replace uses of the AND with uses of the Zero extend node. 2546 CombineTo(N, Zext); 2547 2548 // We actually want to replace all uses of the any_extend with the 2549 // zero_extend, to avoid duplicating things. This will later cause this 2550 // AND to be folded. 2551 CombineTo(N0.getNode(), Zext); 2552 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2553 } 2554 } 2555 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) -> 2556 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must 2557 // already be zero by virtue of the width of the base type of the load. 2558 // 2559 // the 'X' node here can either be nothing or an extract_vector_elt to catch 2560 // more cases. 2561 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && 2562 N0.getOperand(0).getOpcode() == ISD::LOAD) || 2563 N0.getOpcode() == ISD::LOAD) { 2564 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ? 2565 N0 : N0.getOperand(0) ); 2566 2567 // Get the constant (if applicable) the zero'th operand is being ANDed with. 2568 // This can be a pure constant or a vector splat, in which case we treat the 2569 // vector as a scalar and use the splat value. 2570 APInt Constant = APInt::getNullValue(1); 2571 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { 2572 Constant = C->getAPIntValue(); 2573 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) { 2574 APInt SplatValue, SplatUndef; 2575 unsigned SplatBitSize; 2576 bool HasAnyUndefs; 2577 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef, 2578 SplatBitSize, HasAnyUndefs); 2579 if (IsSplat) { 2580 // Undef bits can contribute to a possible optimisation if set, so 2581 // set them. 2582 SplatValue |= SplatUndef; 2583 2584 // The splat value may be something like "0x00FFFFFF", which means 0 for 2585 // the first vector value and FF for the rest, repeating. We need a mask 2586 // that will apply equally to all members of the vector, so AND all the 2587 // lanes of the constant together. 2588 EVT VT = Vector->getValueType(0); 2589 unsigned BitWidth = VT.getVectorElementType().getSizeInBits(); 2590 2591 // If the splat value has been compressed to a bitlength lower 2592 // than the size of the vector lane, we need to re-expand it to 2593 // the lane size. 2594 if (BitWidth > SplatBitSize) 2595 for (SplatValue = SplatValue.zextOrTrunc(BitWidth); 2596 SplatBitSize < BitWidth; 2597 SplatBitSize = SplatBitSize * 2) 2598 SplatValue |= SplatValue.shl(SplatBitSize); 2599 2600 Constant = APInt::getAllOnesValue(BitWidth); 2601 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i) 2602 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth); 2603 } 2604 } 2605 2606 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is 2607 // actually legal and isn't going to get expanded, else this is a false 2608 // optimisation. 2609 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD, 2610 Load->getMemoryVT()); 2611 2612 // Resize the constant to the same size as the original memory access before 2613 // extension. If it is still the AllOnesValue then this AND is completely 2614 // unneeded. 2615 Constant = 2616 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits()); 2617 2618 bool B; 2619 switch (Load->getExtensionType()) { 2620 default: B = false; break; 2621 case ISD::EXTLOAD: B = CanZextLoadProfitably; break; 2622 case ISD::ZEXTLOAD: 2623 case ISD::NON_EXTLOAD: B = true; break; 2624 } 2625 2626 if (B && Constant.isAllOnesValue()) { 2627 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to 2628 // preserve semantics once we get rid of the AND. 2629 SDValue NewLoad(Load, 0); 2630 if (Load->getExtensionType() == ISD::EXTLOAD) { 2631 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD, 2632 Load->getValueType(0), SDLoc(Load), 2633 Load->getChain(), Load->getBasePtr(), 2634 Load->getOffset(), Load->getMemoryVT(), 2635 Load->getMemOperand()); 2636 // Replace uses of the EXTLOAD with the new ZEXTLOAD. 2637 if (Load->getNumValues() == 3) { 2638 // PRE/POST_INC loads have 3 values. 2639 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1), 2640 NewLoad.getValue(2) }; 2641 CombineTo(Load, To, 3, true); 2642 } else { 2643 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1)); 2644 } 2645 } 2646 2647 // Fold the AND away, taking care not to fold to the old load node if we 2648 // replaced it. 2649 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0); 2650 2651 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2652 } 2653 } 2654 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y)) 2655 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 2656 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 2657 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 2658 2659 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 2660 LL.getValueType().isInteger()) { 2661 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0) 2662 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) { 2663 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0), 2664 LR.getValueType(), LL, RL); 2665 AddToWorkList(ORNode.getNode()); 2666 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1); 2667 } 2668 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1) 2669 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) { 2670 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0), 2671 LR.getValueType(), LL, RL); 2672 AddToWorkList(ANDNode.getNode()); 2673 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1); 2674 } 2675 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1) 2676 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) { 2677 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0), 2678 LR.getValueType(), LL, RL); 2679 AddToWorkList(ORNode.getNode()); 2680 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1); 2681 } 2682 } 2683 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2) 2684 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) && 2685 Op0 == Op1 && LL.getValueType().isInteger() && 2686 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() && 2687 cast<ConstantSDNode>(RR)->isAllOnesValue()) || 2688 (cast<ConstantSDNode>(LR)->isAllOnesValue() && 2689 cast<ConstantSDNode>(RR)->isNullValue()))) { 2690 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(), 2691 LL, DAG.getConstant(1, LL.getValueType())); 2692 AddToWorkList(ADDNode.getNode()); 2693 return DAG.getSetCC(SDLoc(N), VT, ADDNode, 2694 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE); 2695 } 2696 // canonicalize equivalent to ll == rl 2697 if (LL == RR && LR == RL) { 2698 Op1 = ISD::getSetCCSwappedOperands(Op1); 2699 std::swap(RL, RR); 2700 } 2701 if (LL == RL && LR == RR) { 2702 bool isInteger = LL.getValueType().isInteger(); 2703 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger); 2704 if (Result != ISD::SETCC_INVALID && 2705 (!LegalOperations || 2706 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) && 2707 TLI.isOperationLegal(ISD::SETCC, 2708 getSetCCResultType(N0.getSimpleValueType()))))) 2709 return DAG.getSetCC(SDLoc(N), N0.getValueType(), 2710 LL, LR, Result); 2711 } 2712 } 2713 2714 // Simplify: (and (op x...), (op y...)) -> (op (and x, y)) 2715 if (N0.getOpcode() == N1.getOpcode()) { 2716 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N); 2717 if (Tmp.getNode()) return Tmp; 2718 } 2719 2720 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1) 2721 // fold (and (sra)) -> (and (srl)) when possible. 2722 if (!VT.isVector() && 2723 SimplifyDemandedBits(SDValue(N, 0))) 2724 return SDValue(N, 0); 2725 2726 // fold (zext_inreg (extload x)) -> (zextload x) 2727 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) { 2728 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2729 EVT MemVT = LN0->getMemoryVT(); 2730 // If we zero all the possible extended bits, then we can turn this into 2731 // a zextload if we are running before legalize or the operation is legal. 2732 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits(); 2733 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth, 2734 BitWidth - MemVT.getScalarType().getSizeInBits())) && 2735 ((!LegalOperations && !LN0->isVolatile()) || 2736 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) { 2737 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT, 2738 LN0->getChain(), LN0->getBasePtr(), 2739 LN0->getPointerInfo(), MemVT, 2740 LN0->isVolatile(), LN0->isNonTemporal(), 2741 LN0->getAlignment()); 2742 AddToWorkList(N); 2743 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); 2744 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2745 } 2746 } 2747 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use 2748 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && 2749 N0.hasOneUse()) { 2750 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2751 EVT MemVT = LN0->getMemoryVT(); 2752 // If we zero all the possible extended bits, then we can turn this into 2753 // a zextload if we are running before legalize or the operation is legal. 2754 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits(); 2755 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth, 2756 BitWidth - MemVT.getScalarType().getSizeInBits())) && 2757 ((!LegalOperations && !LN0->isVolatile()) || 2758 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) { 2759 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT, 2760 LN0->getChain(), 2761 LN0->getBasePtr(), LN0->getPointerInfo(), 2762 MemVT, 2763 LN0->isVolatile(), LN0->isNonTemporal(), 2764 LN0->getAlignment()); 2765 AddToWorkList(N); 2766 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); 2767 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2768 } 2769 } 2770 2771 // fold (and (load x), 255) -> (zextload x, i8) 2772 // fold (and (extload x, i16), 255) -> (zextload x, i8) 2773 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8) 2774 if (N1C && (N0.getOpcode() == ISD::LOAD || 2775 (N0.getOpcode() == ISD::ANY_EXTEND && 2776 N0.getOperand(0).getOpcode() == ISD::LOAD))) { 2777 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND; 2778 LoadSDNode *LN0 = HasAnyExt 2779 ? cast<LoadSDNode>(N0.getOperand(0)) 2780 : cast<LoadSDNode>(N0); 2781 if (LN0->getExtensionType() != ISD::SEXTLOAD && 2782 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) { 2783 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits(); 2784 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){ 2785 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits); 2786 EVT LoadedVT = LN0->getMemoryVT(); 2787 2788 if (ExtVT == LoadedVT && 2789 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) { 2790 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT; 2791 2792 SDValue NewLoad = 2793 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy, 2794 LN0->getChain(), LN0->getBasePtr(), 2795 LN0->getPointerInfo(), 2796 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(), 2797 LN0->getAlignment()); 2798 AddToWorkList(N); 2799 CombineTo(LN0, NewLoad, NewLoad.getValue(1)); 2800 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2801 } 2802 2803 // Do not change the width of a volatile load. 2804 // Do not generate loads of non-round integer types since these can 2805 // be expensive (and would be wrong if the type is not byte sized). 2806 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() && 2807 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) { 2808 EVT PtrType = LN0->getOperand(1).getValueType(); 2809 2810 unsigned Alignment = LN0->getAlignment(); 2811 SDValue NewPtr = LN0->getBasePtr(); 2812 2813 // For big endian targets, we need to add an offset to the pointer 2814 // to load the correct bytes. For little endian systems, we merely 2815 // need to read fewer bytes from the same pointer. 2816 if (TLI.isBigEndian()) { 2817 unsigned LVTStoreBytes = LoadedVT.getStoreSize(); 2818 unsigned EVTStoreBytes = ExtVT.getStoreSize(); 2819 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes; 2820 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType, 2821 NewPtr, DAG.getConstant(PtrOff, PtrType)); 2822 Alignment = MinAlign(Alignment, PtrOff); 2823 } 2824 2825 AddToWorkList(NewPtr.getNode()); 2826 2827 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT; 2828 SDValue Load = 2829 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy, 2830 LN0->getChain(), NewPtr, 2831 LN0->getPointerInfo(), 2832 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(), 2833 Alignment); 2834 AddToWorkList(N); 2835 CombineTo(LN0, Load, Load.getValue(1)); 2836 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2837 } 2838 } 2839 } 2840 } 2841 2842 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL && 2843 VT.getSizeInBits() <= 64) { 2844 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2845 APInt ADDC = ADDI->getAPIntValue(); 2846 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) { 2847 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal 2848 // immediate for an add, but it is legal if its top c2 bits are set, 2849 // transform the ADD so the immediate doesn't need to be materialized 2850 // in a register. 2851 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) { 2852 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), 2853 SRLI->getZExtValue()); 2854 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) { 2855 ADDC |= Mask; 2856 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) { 2857 SDValue NewAdd = 2858 DAG.getNode(ISD::ADD, SDLoc(N0), VT, 2859 N0.getOperand(0), DAG.getConstant(ADDC, VT)); 2860 CombineTo(N0.getNode(), NewAdd); 2861 return SDValue(N, 0); // Return N so it doesn't get rechecked! 2862 } 2863 } 2864 } 2865 } 2866 } 2867 } 2868 2869 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const) 2870 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) { 2871 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), 2872 N0.getOperand(1), false); 2873 if (BSwap.getNode()) 2874 return BSwap; 2875 } 2876 2877 return SDValue(); 2878} 2879 2880/// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16 2881/// 2882SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1, 2883 bool DemandHighBits) { 2884 if (!LegalOperations) 2885 return SDValue(); 2886 2887 EVT VT = N->getValueType(0); 2888 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16) 2889 return SDValue(); 2890 if (!TLI.isOperationLegal(ISD::BSWAP, VT)) 2891 return SDValue(); 2892 2893 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00) 2894 bool LookPassAnd0 = false; 2895 bool LookPassAnd1 = false; 2896 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL) 2897 std::swap(N0, N1); 2898 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL) 2899 std::swap(N0, N1); 2900 if (N0.getOpcode() == ISD::AND) { 2901 if (!N0.getNode()->hasOneUse()) 2902 return SDValue(); 2903 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 2904 if (!N01C || N01C->getZExtValue() != 0xFF00) 2905 return SDValue(); 2906 N0 = N0.getOperand(0); 2907 LookPassAnd0 = true; 2908 } 2909 2910 if (N1.getOpcode() == ISD::AND) { 2911 if (!N1.getNode()->hasOneUse()) 2912 return SDValue(); 2913 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 2914 if (!N11C || N11C->getZExtValue() != 0xFF) 2915 return SDValue(); 2916 N1 = N1.getOperand(0); 2917 LookPassAnd1 = true; 2918 } 2919 2920 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL) 2921 std::swap(N0, N1); 2922 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL) 2923 return SDValue(); 2924 if (!N0.getNode()->hasOneUse() || 2925 !N1.getNode()->hasOneUse()) 2926 return SDValue(); 2927 2928 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 2929 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1)); 2930 if (!N01C || !N11C) 2931 return SDValue(); 2932 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8) 2933 return SDValue(); 2934 2935 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8) 2936 SDValue N00 = N0->getOperand(0); 2937 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) { 2938 if (!N00.getNode()->hasOneUse()) 2939 return SDValue(); 2940 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1)); 2941 if (!N001C || N001C->getZExtValue() != 0xFF) 2942 return SDValue(); 2943 N00 = N00.getOperand(0); 2944 LookPassAnd0 = true; 2945 } 2946 2947 SDValue N10 = N1->getOperand(0); 2948 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) { 2949 if (!N10.getNode()->hasOneUse()) 2950 return SDValue(); 2951 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1)); 2952 if (!N101C || N101C->getZExtValue() != 0xFF00) 2953 return SDValue(); 2954 N10 = N10.getOperand(0); 2955 LookPassAnd1 = true; 2956 } 2957 2958 if (N00 != N10) 2959 return SDValue(); 2960 2961 // Make sure everything beyond the low halfword gets set to zero since the SRL 2962 // 16 will clear the top bits. 2963 unsigned OpSizeInBits = VT.getSizeInBits(); 2964 if (DemandHighBits && OpSizeInBits > 16) { 2965 // If the left-shift isn't masked out then the only way this is a bswap is 2966 // if all bits beyond the low 8 are 0. In that case the entire pattern 2967 // reduces to a left shift anyway: leave it for other parts of the combiner. 2968 if (!LookPassAnd0) 2969 return SDValue(); 2970 2971 // However, if the right shift isn't masked out then it might be because 2972 // it's not needed. See if we can spot that too. 2973 if (!LookPassAnd1 && 2974 !DAG.MaskedValueIsZero( 2975 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16))) 2976 return SDValue(); 2977 } 2978 2979 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00); 2980 if (OpSizeInBits > 16) 2981 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res, 2982 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT))); 2983 return Res; 2984} 2985 2986/// isBSwapHWordElement - Return true if the specified node is an element 2987/// that makes up a 32-bit packed halfword byteswap. i.e. 2988/// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8) 2989static bool isBSwapHWordElement(SDValue N, SmallVectorImpl<SDNode *> &Parts) { 2990 if (!N.getNode()->hasOneUse()) 2991 return false; 2992 2993 unsigned Opc = N.getOpcode(); 2994 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL) 2995 return false; 2996 2997 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1)); 2998 if (!N1C) 2999 return false; 3000 3001 unsigned Num; 3002 switch (N1C->getZExtValue()) { 3003 default: 3004 return false; 3005 case 0xFF: Num = 0; break; 3006 case 0xFF00: Num = 1; break; 3007 case 0xFF0000: Num = 2; break; 3008 case 0xFF000000: Num = 3; break; 3009 } 3010 3011 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00). 3012 SDValue N0 = N.getOperand(0); 3013 if (Opc == ISD::AND) { 3014 if (Num == 0 || Num == 2) { 3015 // (x >> 8) & 0xff 3016 // (x >> 8) & 0xff0000 3017 if (N0.getOpcode() != ISD::SRL) 3018 return false; 3019 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 3020 if (!C || C->getZExtValue() != 8) 3021 return false; 3022 } else { 3023 // (x << 8) & 0xff00 3024 // (x << 8) & 0xff000000 3025 if (N0.getOpcode() != ISD::SHL) 3026 return false; 3027 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 3028 if (!C || C->getZExtValue() != 8) 3029 return false; 3030 } 3031 } else if (Opc == ISD::SHL) { 3032 // (x & 0xff) << 8 3033 // (x & 0xff0000) << 8 3034 if (Num != 0 && Num != 2) 3035 return false; 3036 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1)); 3037 if (!C || C->getZExtValue() != 8) 3038 return false; 3039 } else { // Opc == ISD::SRL 3040 // (x & 0xff00) >> 8 3041 // (x & 0xff000000) >> 8 3042 if (Num != 1 && Num != 3) 3043 return false; 3044 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1)); 3045 if (!C || C->getZExtValue() != 8) 3046 return false; 3047 } 3048 3049 if (Parts[Num]) 3050 return false; 3051 3052 Parts[Num] = N0.getOperand(0).getNode(); 3053 return true; 3054} 3055 3056/// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is 3057/// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8) 3058/// => (rotl (bswap x), 16) 3059SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) { 3060 if (!LegalOperations) 3061 return SDValue(); 3062 3063 EVT VT = N->getValueType(0); 3064 if (VT != MVT::i32) 3065 return SDValue(); 3066 if (!TLI.isOperationLegal(ISD::BSWAP, VT)) 3067 return SDValue(); 3068 3069 SmallVector<SDNode*,4> Parts(4, (SDNode*)0); 3070 // Look for either 3071 // (or (or (and), (and)), (or (and), (and))) 3072 // (or (or (or (and), (and)), (and)), (and)) 3073 if (N0.getOpcode() != ISD::OR) 3074 return SDValue(); 3075 SDValue N00 = N0.getOperand(0); 3076 SDValue N01 = N0.getOperand(1); 3077 3078 if (N1.getOpcode() == ISD::OR && 3079 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) { 3080 // (or (or (and), (and)), (or (and), (and))) 3081 SDValue N000 = N00.getOperand(0); 3082 if (!isBSwapHWordElement(N000, Parts)) 3083 return SDValue(); 3084 3085 SDValue N001 = N00.getOperand(1); 3086 if (!isBSwapHWordElement(N001, Parts)) 3087 return SDValue(); 3088 SDValue N010 = N01.getOperand(0); 3089 if (!isBSwapHWordElement(N010, Parts)) 3090 return SDValue(); 3091 SDValue N011 = N01.getOperand(1); 3092 if (!isBSwapHWordElement(N011, Parts)) 3093 return SDValue(); 3094 } else { 3095 // (or (or (or (and), (and)), (and)), (and)) 3096 if (!isBSwapHWordElement(N1, Parts)) 3097 return SDValue(); 3098 if (!isBSwapHWordElement(N01, Parts)) 3099 return SDValue(); 3100 if (N00.getOpcode() != ISD::OR) 3101 return SDValue(); 3102 SDValue N000 = N00.getOperand(0); 3103 if (!isBSwapHWordElement(N000, Parts)) 3104 return SDValue(); 3105 SDValue N001 = N00.getOperand(1); 3106 if (!isBSwapHWordElement(N001, Parts)) 3107 return SDValue(); 3108 } 3109 3110 // Make sure the parts are all coming from the same node. 3111 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3]) 3112 return SDValue(); 3113 3114 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, 3115 SDValue(Parts[0],0)); 3116 3117 // Result of the bswap should be rotated by 16. If it's not legal, than 3118 // do (x << 16) | (x >> 16). 3119 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT)); 3120 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT)) 3121 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt); 3122 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT)) 3123 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt); 3124 return DAG.getNode(ISD::OR, SDLoc(N), VT, 3125 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt), 3126 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt)); 3127} 3128 3129SDValue DAGCombiner::visitOR(SDNode *N) { 3130 SDValue N0 = N->getOperand(0); 3131 SDValue N1 = N->getOperand(1); 3132 SDValue LL, LR, RL, RR, CC0, CC1; 3133 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3134 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3135 EVT VT = N1.getValueType(); 3136 3137 // fold vector ops 3138 if (VT.isVector()) { 3139 SDValue FoldedVOp = SimplifyVBinOp(N); 3140 if (FoldedVOp.getNode()) return FoldedVOp; 3141 3142 // fold (or x, 0) -> x, vector edition 3143 if (ISD::isBuildVectorAllZeros(N0.getNode())) 3144 return N1; 3145 if (ISD::isBuildVectorAllZeros(N1.getNode())) 3146 return N0; 3147 3148 // fold (or x, -1) -> -1, vector edition 3149 if (ISD::isBuildVectorAllOnes(N0.getNode())) 3150 return N0; 3151 if (ISD::isBuildVectorAllOnes(N1.getNode())) 3152 return N1; 3153 } 3154 3155 // fold (or x, undef) -> -1 3156 if (!LegalOperations && 3157 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) { 3158 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT; 3159 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT); 3160 } 3161 // fold (or c1, c2) -> c1|c2 3162 if (N0C && N1C) 3163 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C); 3164 // canonicalize constant to RHS 3165 if (N0C && !N1C) 3166 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0); 3167 // fold (or x, 0) -> x 3168 if (N1C && N1C->isNullValue()) 3169 return N0; 3170 // fold (or x, -1) -> -1 3171 if (N1C && N1C->isAllOnesValue()) 3172 return N1; 3173 // fold (or x, c) -> c iff (x & ~c) == 0 3174 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue())) 3175 return N1; 3176 3177 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16) 3178 SDValue BSwap = MatchBSwapHWord(N, N0, N1); 3179 if (BSwap.getNode() != 0) 3180 return BSwap; 3181 BSwap = MatchBSwapHWordLow(N, N0, N1); 3182 if (BSwap.getNode() != 0) 3183 return BSwap; 3184 3185 // reassociate or 3186 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1); 3187 if (ROR.getNode() != 0) 3188 return ROR; 3189 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2) 3190 // iff (c1 & c2) == 0. 3191 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() && 3192 isa<ConstantSDNode>(N0.getOperand(1))) { 3193 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1)); 3194 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) 3195 return DAG.getNode(ISD::AND, SDLoc(N), VT, 3196 DAG.getNode(ISD::OR, SDLoc(N0), VT, 3197 N0.getOperand(0), N1), 3198 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1)); 3199 } 3200 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y)) 3201 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 3202 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 3203 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 3204 3205 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 3206 LL.getValueType().isInteger()) { 3207 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0) 3208 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0) 3209 if (cast<ConstantSDNode>(LR)->isNullValue() && 3210 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) { 3211 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR), 3212 LR.getValueType(), LL, RL); 3213 AddToWorkList(ORNode.getNode()); 3214 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1); 3215 } 3216 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1) 3217 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1) 3218 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && 3219 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) { 3220 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR), 3221 LR.getValueType(), LL, RL); 3222 AddToWorkList(ANDNode.getNode()); 3223 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1); 3224 } 3225 } 3226 // canonicalize equivalent to ll == rl 3227 if (LL == RR && LR == RL) { 3228 Op1 = ISD::getSetCCSwappedOperands(Op1); 3229 std::swap(RL, RR); 3230 } 3231 if (LL == RL && LR == RR) { 3232 bool isInteger = LL.getValueType().isInteger(); 3233 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger); 3234 if (Result != ISD::SETCC_INVALID && 3235 (!LegalOperations || 3236 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) && 3237 TLI.isOperationLegal(ISD::SETCC, 3238 getSetCCResultType(N0.getValueType()))))) 3239 return DAG.getSetCC(SDLoc(N), N0.getValueType(), 3240 LL, LR, Result); 3241 } 3242 } 3243 3244 // Simplify: (or (op x...), (op y...)) -> (op (or x, y)) 3245 if (N0.getOpcode() == N1.getOpcode()) { 3246 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N); 3247 if (Tmp.getNode()) return Tmp; 3248 } 3249 3250 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible. 3251 if (N0.getOpcode() == ISD::AND && 3252 N1.getOpcode() == ISD::AND && 3253 N0.getOperand(1).getOpcode() == ISD::Constant && 3254 N1.getOperand(1).getOpcode() == ISD::Constant && 3255 // Don't increase # computations. 3256 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) { 3257 // We can only do this xform if we know that bits from X that are set in C2 3258 // but not in C1 are already zero. Likewise for Y. 3259 const APInt &LHSMask = 3260 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 3261 const APInt &RHSMask = 3262 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue(); 3263 3264 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) && 3265 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) { 3266 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT, 3267 N0.getOperand(0), N1.getOperand(0)); 3268 return DAG.getNode(ISD::AND, SDLoc(N), VT, X, 3269 DAG.getConstant(LHSMask | RHSMask, VT)); 3270 } 3271 } 3272 3273 // See if this is some rotate idiom. 3274 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N))) 3275 return SDValue(Rot, 0); 3276 3277 // Simplify the operands using demanded-bits information. 3278 if (!VT.isVector() && 3279 SimplifyDemandedBits(SDValue(N, 0))) 3280 return SDValue(N, 0); 3281 3282 return SDValue(); 3283} 3284 3285/// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present. 3286static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) { 3287 if (Op.getOpcode() == ISD::AND) { 3288 if (isa<ConstantSDNode>(Op.getOperand(1))) { 3289 Mask = Op.getOperand(1); 3290 Op = Op.getOperand(0); 3291 } else { 3292 return false; 3293 } 3294 } 3295 3296 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) { 3297 Shift = Op; 3298 return true; 3299 } 3300 3301 return false; 3302} 3303 3304// MatchRotate - Handle an 'or' of two operands. If this is one of the many 3305// idioms for rotate, and if the target supports rotation instructions, generate 3306// a rot[lr]. 3307SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) { 3308 // Must be a legal type. Expanded 'n promoted things won't work with rotates. 3309 EVT VT = LHS.getValueType(); 3310 if (!TLI.isTypeLegal(VT)) return 0; 3311 3312 // The target must have at least one rotate flavor. 3313 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT); 3314 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT); 3315 if (!HasROTL && !HasROTR) return 0; 3316 3317 // Match "(X shl/srl V1) & V2" where V2 may not be present. 3318 SDValue LHSShift; // The shift. 3319 SDValue LHSMask; // AND value if any. 3320 if (!MatchRotateHalf(LHS, LHSShift, LHSMask)) 3321 return 0; // Not part of a rotate. 3322 3323 SDValue RHSShift; // The shift. 3324 SDValue RHSMask; // AND value if any. 3325 if (!MatchRotateHalf(RHS, RHSShift, RHSMask)) 3326 return 0; // Not part of a rotate. 3327 3328 if (LHSShift.getOperand(0) != RHSShift.getOperand(0)) 3329 return 0; // Not shifting the same value. 3330 3331 if (LHSShift.getOpcode() == RHSShift.getOpcode()) 3332 return 0; // Shifts must disagree. 3333 3334 // Canonicalize shl to left side in a shl/srl pair. 3335 if (RHSShift.getOpcode() == ISD::SHL) { 3336 std::swap(LHS, RHS); 3337 std::swap(LHSShift, RHSShift); 3338 std::swap(LHSMask , RHSMask ); 3339 } 3340 3341 unsigned OpSizeInBits = VT.getSizeInBits(); 3342 SDValue LHSShiftArg = LHSShift.getOperand(0); 3343 SDValue LHSShiftAmt = LHSShift.getOperand(1); 3344 SDValue RHSShiftArg = RHSShift.getOperand(0); 3345 SDValue RHSShiftAmt = RHSShift.getOperand(1); 3346 3347 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1) 3348 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2) 3349 if (LHSShiftAmt.getOpcode() == ISD::Constant && 3350 RHSShiftAmt.getOpcode() == ISD::Constant) { 3351 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue(); 3352 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue(); 3353 if ((LShVal + RShVal) != OpSizeInBits) 3354 return 0; 3355 3356 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, 3357 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt); 3358 3359 // If there is an AND of either shifted operand, apply it to the result. 3360 if (LHSMask.getNode() || RHSMask.getNode()) { 3361 APInt Mask = APInt::getAllOnesValue(OpSizeInBits); 3362 3363 if (LHSMask.getNode()) { 3364 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal); 3365 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits; 3366 } 3367 if (RHSMask.getNode()) { 3368 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal); 3369 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits; 3370 } 3371 3372 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT)); 3373 } 3374 3375 return Rot.getNode(); 3376 } 3377 3378 // If there is a mask here, and we have a variable shift, we can't be sure 3379 // that we're masking out the right stuff. 3380 if (LHSMask.getNode() || RHSMask.getNode()) 3381 return 0; 3382 3383 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y) 3384 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y)) 3385 if (RHSShiftAmt.getOpcode() == ISD::SUB && 3386 LHSShiftAmt == RHSShiftAmt.getOperand(1)) { 3387 if (ConstantSDNode *SUBC = 3388 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) { 3389 if (SUBC->getAPIntValue() == OpSizeInBits) 3390 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, LHSShiftArg, 3391 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode(); 3392 } 3393 } 3394 3395 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y) 3396 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y)) 3397 if (LHSShiftAmt.getOpcode() == ISD::SUB && 3398 RHSShiftAmt == LHSShiftAmt.getOperand(1)) 3399 if (ConstantSDNode *SUBC = 3400 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) 3401 if (SUBC->getAPIntValue() == OpSizeInBits) 3402 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT, LHSShiftArg, 3403 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode(); 3404 3405 // Look for sign/zext/any-extended or truncate cases: 3406 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND || 3407 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND || 3408 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND || 3409 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) && 3410 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND || 3411 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND || 3412 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND || 3413 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) { 3414 SDValue LExtOp0 = LHSShiftAmt.getOperand(0); 3415 SDValue RExtOp0 = RHSShiftAmt.getOperand(0); 3416 if (RExtOp0.getOpcode() == ISD::SUB && 3417 RExtOp0.getOperand(1) == LExtOp0) { 3418 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) -> 3419 // (rotl x, y) 3420 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) -> 3421 // (rotr x, (sub 32, y)) 3422 if (ConstantSDNode *SUBC = 3423 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) 3424 if (SUBC->getAPIntValue() == OpSizeInBits) { 3425 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, 3426 LHSShiftArg, 3427 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode(); 3428 } else if (LHSShiftArg.getOpcode() == ISD::ZERO_EXTEND || 3429 LHSShiftArg.getOpcode() == ISD::ANY_EXTEND) { 3430 // fold (or (shl (*ext x), (*ext y)), 3431 // (srl (*ext x), (*ext (sub 32, y)))) -> 3432 // (*ext (rotl x, y)) 3433 // fold (or (shl (*ext x), (*ext y)), 3434 // (srl (*ext x), (*ext (sub 32, y)))) -> 3435 // (*ext (rotr x, (sub 32, y))) 3436 SDValue LArgExtOp0 = LHSShiftArg.getOperand(0); 3437 EVT LArgVT = LArgExtOp0.getValueType(); 3438 if (LArgVT.getSizeInBits() == SUBC->getAPIntValue()) { 3439 SDValue V = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, LArgVT, 3440 LArgExtOp0, 3441 HasROTL ? LHSShiftAmt : RHSShiftAmt); 3442 return DAG.getNode(LHSShiftArg.getOpcode(), DL, VT, V).getNode(); 3443 } 3444 } 3445 } else if (LExtOp0.getOpcode() == ISD::SUB && 3446 RExtOp0 == LExtOp0.getOperand(1)) { 3447 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) -> 3448 // (rotr x, y) 3449 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) -> 3450 // (rotl x, (sub 32, y)) 3451 if (ConstantSDNode *SUBC = 3452 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) 3453 if (SUBC->getAPIntValue() == OpSizeInBits) { 3454 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT, 3455 LHSShiftArg, 3456 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode(); 3457 } else if (RHSShiftArg.getOpcode() == ISD::ZERO_EXTEND || 3458 RHSShiftArg.getOpcode() == ISD::ANY_EXTEND) { 3459 // fold (or (shl (*ext x), (*ext (sub 32, y))), 3460 // (srl (*ext x), (*ext y))) -> 3461 // (*ext (rotl x, y)) 3462 // fold (or (shl (*ext x), (*ext (sub 32, y))), 3463 // (srl (*ext x), (*ext y))) -> 3464 // (*ext (rotr x, (sub 32, y))) 3465 SDValue RArgExtOp0 = RHSShiftArg.getOperand(0); 3466 EVT RArgVT = RArgExtOp0.getValueType(); 3467 if (RArgVT.getSizeInBits() == SUBC->getAPIntValue()) { 3468 SDValue V = DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, RArgVT, 3469 RArgExtOp0, 3470 HasROTR ? RHSShiftAmt : LHSShiftAmt); 3471 return DAG.getNode(RHSShiftArg.getOpcode(), DL, VT, V).getNode(); 3472 } 3473 } 3474 } 3475 } 3476 3477 return 0; 3478} 3479 3480SDValue DAGCombiner::visitXOR(SDNode *N) { 3481 SDValue N0 = N->getOperand(0); 3482 SDValue N1 = N->getOperand(1); 3483 SDValue LHS, RHS, CC; 3484 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3485 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3486 EVT VT = N0.getValueType(); 3487 3488 // fold vector ops 3489 if (VT.isVector()) { 3490 SDValue FoldedVOp = SimplifyVBinOp(N); 3491 if (FoldedVOp.getNode()) return FoldedVOp; 3492 3493 // fold (xor x, 0) -> x, vector edition 3494 if (ISD::isBuildVectorAllZeros(N0.getNode())) 3495 return N1; 3496 if (ISD::isBuildVectorAllZeros(N1.getNode())) 3497 return N0; 3498 } 3499 3500 // fold (xor undef, undef) -> 0. This is a common idiom (misuse). 3501 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF) 3502 return DAG.getConstant(0, VT); 3503 // fold (xor x, undef) -> undef 3504 if (N0.getOpcode() == ISD::UNDEF) 3505 return N0; 3506 if (N1.getOpcode() == ISD::UNDEF) 3507 return N1; 3508 // fold (xor c1, c2) -> c1^c2 3509 if (N0C && N1C) 3510 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C); 3511 // canonicalize constant to RHS 3512 if (N0C && !N1C) 3513 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0); 3514 // fold (xor x, 0) -> x 3515 if (N1C && N1C->isNullValue()) 3516 return N0; 3517 // reassociate xor 3518 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1); 3519 if (RXOR.getNode() != 0) 3520 return RXOR; 3521 3522 // fold !(x cc y) -> (x !cc y) 3523 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) { 3524 bool isInt = LHS.getValueType().isInteger(); 3525 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), 3526 isInt); 3527 3528 if (!LegalOperations || 3529 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) { 3530 switch (N0.getOpcode()) { 3531 default: 3532 llvm_unreachable("Unhandled SetCC Equivalent!"); 3533 case ISD::SETCC: 3534 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC); 3535 case ISD::SELECT_CC: 3536 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2), 3537 N0.getOperand(3), NotCC); 3538 } 3539 } 3540 } 3541 3542 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y))) 3543 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND && 3544 N0.getNode()->hasOneUse() && 3545 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){ 3546 SDValue V = N0.getOperand(0); 3547 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V, 3548 DAG.getConstant(1, V.getValueType())); 3549 AddToWorkList(V.getNode()); 3550 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V); 3551 } 3552 3553 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc 3554 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 && 3555 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 3556 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); 3557 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) { 3558 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 3559 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS 3560 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS 3561 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode()); 3562 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS); 3563 } 3564 } 3565 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants 3566 if (N1C && N1C->isAllOnesValue() && 3567 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 3568 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); 3569 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) { 3570 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 3571 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS 3572 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS 3573 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode()); 3574 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS); 3575 } 3576 } 3577 // fold (xor (and x, y), y) -> (and (not x), y) 3578 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() && 3579 N0->getOperand(1) == N1) { 3580 SDValue X = N0->getOperand(0); 3581 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT); 3582 AddToWorkList(NotX.getNode()); 3583 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1); 3584 } 3585 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2)) 3586 if (N1C && N0.getOpcode() == ISD::XOR) { 3587 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0)); 3588 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 3589 if (N00C) 3590 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1), 3591 DAG.getConstant(N1C->getAPIntValue() ^ 3592 N00C->getAPIntValue(), VT)); 3593 if (N01C) 3594 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0), 3595 DAG.getConstant(N1C->getAPIntValue() ^ 3596 N01C->getAPIntValue(), VT)); 3597 } 3598 // fold (xor x, x) -> 0 3599 if (N0 == N1) 3600 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes); 3601 3602 // Simplify: xor (op x...), (op y...) -> (op (xor x, y)) 3603 if (N0.getOpcode() == N1.getOpcode()) { 3604 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N); 3605 if (Tmp.getNode()) return Tmp; 3606 } 3607 3608 // Simplify the expression using non-local knowledge. 3609 if (!VT.isVector() && 3610 SimplifyDemandedBits(SDValue(N, 0))) 3611 return SDValue(N, 0); 3612 3613 return SDValue(); 3614} 3615 3616/// visitShiftByConstant - Handle transforms common to the three shifts, when 3617/// the shift amount is a constant. 3618SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) { 3619 SDNode *LHS = N->getOperand(0).getNode(); 3620 if (!LHS->hasOneUse()) return SDValue(); 3621 3622 // We want to pull some binops through shifts, so that we have (and (shift)) 3623 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of 3624 // thing happens with address calculations, so it's important to canonicalize 3625 // it. 3626 bool HighBitSet = false; // Can we transform this if the high bit is set? 3627 3628 switch (LHS->getOpcode()) { 3629 default: return SDValue(); 3630 case ISD::OR: 3631 case ISD::XOR: 3632 HighBitSet = false; // We can only transform sra if the high bit is clear. 3633 break; 3634 case ISD::AND: 3635 HighBitSet = true; // We can only transform sra if the high bit is set. 3636 break; 3637 case ISD::ADD: 3638 if (N->getOpcode() != ISD::SHL) 3639 return SDValue(); // only shl(add) not sr[al](add). 3640 HighBitSet = false; // We can only transform sra if the high bit is clear. 3641 break; 3642 } 3643 3644 // We require the RHS of the binop to be a constant as well. 3645 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1)); 3646 if (!BinOpCst) return SDValue(); 3647 3648 // FIXME: disable this unless the input to the binop is a shift by a constant. 3649 // If it is not a shift, it pessimizes some common cases like: 3650 // 3651 // void foo(int *X, int i) { X[i & 1235] = 1; } 3652 // int bar(int *X, int i) { return X[i & 255]; } 3653 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode(); 3654 if ((BinOpLHSVal->getOpcode() != ISD::SHL && 3655 BinOpLHSVal->getOpcode() != ISD::SRA && 3656 BinOpLHSVal->getOpcode() != ISD::SRL) || 3657 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1))) 3658 return SDValue(); 3659 3660 EVT VT = N->getValueType(0); 3661 3662 // If this is a signed shift right, and the high bit is modified by the 3663 // logical operation, do not perform the transformation. The highBitSet 3664 // boolean indicates the value of the high bit of the constant which would 3665 // cause it to be modified for this operation. 3666 if (N->getOpcode() == ISD::SRA) { 3667 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative(); 3668 if (BinOpRHSSignSet != HighBitSet) 3669 return SDValue(); 3670 } 3671 3672 // Fold the constants, shifting the binop RHS by the shift amount. 3673 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)), 3674 N->getValueType(0), 3675 LHS->getOperand(1), N->getOperand(1)); 3676 3677 // Create the new shift. 3678 SDValue NewShift = DAG.getNode(N->getOpcode(), 3679 SDLoc(LHS->getOperand(0)), 3680 VT, LHS->getOperand(0), N->getOperand(1)); 3681 3682 // Create the new binop. 3683 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS); 3684} 3685 3686SDValue DAGCombiner::visitSHL(SDNode *N) { 3687 SDValue N0 = N->getOperand(0); 3688 SDValue N1 = N->getOperand(1); 3689 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3690 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3691 EVT VT = N0.getValueType(); 3692 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits(); 3693 3694 // fold (shl c1, c2) -> c1<<c2 3695 if (N0C && N1C) 3696 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C); 3697 // fold (shl 0, x) -> 0 3698 if (N0C && N0C->isNullValue()) 3699 return N0; 3700 // fold (shl x, c >= size(x)) -> undef 3701 if (N1C && N1C->getZExtValue() >= OpSizeInBits) 3702 return DAG.getUNDEF(VT); 3703 // fold (shl x, 0) -> x 3704 if (N1C && N1C->isNullValue()) 3705 return N0; 3706 // fold (shl undef, x) -> 0 3707 if (N0.getOpcode() == ISD::UNDEF) 3708 return DAG.getConstant(0, VT); 3709 // if (shl x, c) is known to be zero, return 0 3710 if (DAG.MaskedValueIsZero(SDValue(N, 0), 3711 APInt::getAllOnesValue(OpSizeInBits))) 3712 return DAG.getConstant(0, VT); 3713 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))). 3714 if (N1.getOpcode() == ISD::TRUNCATE && 3715 N1.getOperand(0).getOpcode() == ISD::AND && 3716 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) { 3717 SDValue N101 = N1.getOperand(0).getOperand(1); 3718 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) { 3719 EVT TruncVT = N1.getValueType(); 3720 SDValue N100 = N1.getOperand(0).getOperand(0); 3721 APInt TruncC = N101C->getAPIntValue(); 3722 TruncC = TruncC.trunc(TruncVT.getSizeInBits()); 3723 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, 3724 DAG.getNode(ISD::AND, SDLoc(N), TruncVT, 3725 DAG.getNode(ISD::TRUNCATE, 3726 SDLoc(N), 3727 TruncVT, N100), 3728 DAG.getConstant(TruncC, TruncVT))); 3729 } 3730 } 3731 3732 if (N1C && SimplifyDemandedBits(SDValue(N, 0))) 3733 return SDValue(N, 0); 3734 3735 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2)) 3736 if (N1C && N0.getOpcode() == ISD::SHL && 3737 N0.getOperand(1).getOpcode() == ISD::Constant) { 3738 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 3739 uint64_t c2 = N1C->getZExtValue(); 3740 if (c1 + c2 >= OpSizeInBits) 3741 return DAG.getConstant(0, VT); 3742 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0), 3743 DAG.getConstant(c1 + c2, N1.getValueType())); 3744 } 3745 3746 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2))) 3747 // For this to be valid, the second form must not preserve any of the bits 3748 // that are shifted out by the inner shift in the first form. This means 3749 // the outer shift size must be >= the number of bits added by the ext. 3750 // As a corollary, we don't care what kind of ext it is. 3751 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND || 3752 N0.getOpcode() == ISD::ANY_EXTEND || 3753 N0.getOpcode() == ISD::SIGN_EXTEND) && 3754 N0.getOperand(0).getOpcode() == ISD::SHL && 3755 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) { 3756 uint64_t c1 = 3757 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue(); 3758 uint64_t c2 = N1C->getZExtValue(); 3759 EVT InnerShiftVT = N0.getOperand(0).getValueType(); 3760 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits(); 3761 if (c2 >= OpSizeInBits - InnerShiftSize) { 3762 if (c1 + c2 >= OpSizeInBits) 3763 return DAG.getConstant(0, VT); 3764 return DAG.getNode(ISD::SHL, SDLoc(N0), VT, 3765 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT, 3766 N0.getOperand(0)->getOperand(0)), 3767 DAG.getConstant(c1 + c2, N1.getValueType())); 3768 } 3769 } 3770 3771 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or 3772 // (and (srl x, (sub c1, c2), MASK) 3773 // Only fold this if the inner shift has no other uses -- if it does, folding 3774 // this will increase the total number of instructions. 3775 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() && 3776 N0.getOperand(1).getOpcode() == ISD::Constant) { 3777 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 3778 if (c1 < VT.getSizeInBits()) { 3779 uint64_t c2 = N1C->getZExtValue(); 3780 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(), 3781 VT.getSizeInBits() - c1); 3782 SDValue Shift; 3783 if (c2 > c1) { 3784 Mask = Mask.shl(c2-c1); 3785 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0), 3786 DAG.getConstant(c2-c1, N1.getValueType())); 3787 } else { 3788 Mask = Mask.lshr(c1-c2); 3789 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), 3790 DAG.getConstant(c1-c2, N1.getValueType())); 3791 } 3792 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift, 3793 DAG.getConstant(Mask, VT)); 3794 } 3795 } 3796 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1)) 3797 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) { 3798 SDValue HiBitsMask = 3799 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(), 3800 VT.getSizeInBits() - 3801 N1C->getZExtValue()), 3802 VT); 3803 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0), 3804 HiBitsMask); 3805 } 3806 3807 if (N1C) { 3808 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue()); 3809 if (NewSHL.getNode()) 3810 return NewSHL; 3811 } 3812 3813 return SDValue(); 3814} 3815 3816SDValue DAGCombiner::visitSRA(SDNode *N) { 3817 SDValue N0 = N->getOperand(0); 3818 SDValue N1 = N->getOperand(1); 3819 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3820 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3821 EVT VT = N0.getValueType(); 3822 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits(); 3823 3824 // fold (sra c1, c2) -> (sra c1, c2) 3825 if (N0C && N1C) 3826 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C); 3827 // fold (sra 0, x) -> 0 3828 if (N0C && N0C->isNullValue()) 3829 return N0; 3830 // fold (sra -1, x) -> -1 3831 if (N0C && N0C->isAllOnesValue()) 3832 return N0; 3833 // fold (sra x, (setge c, size(x))) -> undef 3834 if (N1C && N1C->getZExtValue() >= OpSizeInBits) 3835 return DAG.getUNDEF(VT); 3836 // fold (sra x, 0) -> x 3837 if (N1C && N1C->isNullValue()) 3838 return N0; 3839 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports 3840 // sext_inreg. 3841 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) { 3842 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue(); 3843 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits); 3844 if (VT.isVector()) 3845 ExtVT = EVT::getVectorVT(*DAG.getContext(), 3846 ExtVT, VT.getVectorNumElements()); 3847 if ((!LegalOperations || 3848 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT))) 3849 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, 3850 N0.getOperand(0), DAG.getValueType(ExtVT)); 3851 } 3852 3853 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2)) 3854 if (N1C && N0.getOpcode() == ISD::SRA) { 3855 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 3856 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue(); 3857 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1; 3858 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0), 3859 DAG.getConstant(Sum, N1C->getValueType(0))); 3860 } 3861 } 3862 3863 // fold (sra (shl X, m), (sub result_size, n)) 3864 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for 3865 // result_size - n != m. 3866 // If truncate is free for the target sext(shl) is likely to result in better 3867 // code. 3868 if (N0.getOpcode() == ISD::SHL) { 3869 // Get the two constanst of the shifts, CN0 = m, CN = n. 3870 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 3871 if (N01C && N1C) { 3872 // Determine what the truncate's result bitsize and type would be. 3873 EVT TruncVT = 3874 EVT::getIntegerVT(*DAG.getContext(), 3875 OpSizeInBits - N1C->getZExtValue()); 3876 // Determine the residual right-shift amount. 3877 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue(); 3878 3879 // If the shift is not a no-op (in which case this should be just a sign 3880 // extend already), the truncated to type is legal, sign_extend is legal 3881 // on that type, and the truncate to that type is both legal and free, 3882 // perform the transform. 3883 if ((ShiftAmt > 0) && 3884 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) && 3885 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) && 3886 TLI.isTruncateFree(VT, TruncVT)) { 3887 3888 SDValue Amt = DAG.getConstant(ShiftAmt, 3889 getShiftAmountTy(N0.getOperand(0).getValueType())); 3890 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT, 3891 N0.getOperand(0), Amt); 3892 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT, 3893 Shift); 3894 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), 3895 N->getValueType(0), Trunc); 3896 } 3897 } 3898 } 3899 3900 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))). 3901 if (N1.getOpcode() == ISD::TRUNCATE && 3902 N1.getOperand(0).getOpcode() == ISD::AND && 3903 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) { 3904 SDValue N101 = N1.getOperand(0).getOperand(1); 3905 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) { 3906 EVT TruncVT = N1.getValueType(); 3907 SDValue N100 = N1.getOperand(0).getOperand(0); 3908 APInt TruncC = N101C->getAPIntValue(); 3909 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits()); 3910 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, 3911 DAG.getNode(ISD::AND, SDLoc(N), 3912 TruncVT, 3913 DAG.getNode(ISD::TRUNCATE, 3914 SDLoc(N), 3915 TruncVT, N100), 3916 DAG.getConstant(TruncC, TruncVT))); 3917 } 3918 } 3919 3920 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2)) 3921 // if c1 is equal to the number of bits the trunc removes 3922 if (N0.getOpcode() == ISD::TRUNCATE && 3923 (N0.getOperand(0).getOpcode() == ISD::SRL || 3924 N0.getOperand(0).getOpcode() == ISD::SRA) && 3925 N0.getOperand(0).hasOneUse() && 3926 N0.getOperand(0).getOperand(1).hasOneUse() && 3927 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) { 3928 EVT LargeVT = N0.getOperand(0).getValueType(); 3929 ConstantSDNode *LargeShiftAmt = 3930 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1)); 3931 3932 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits == 3933 LargeShiftAmt->getZExtValue()) { 3934 SDValue Amt = 3935 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(), 3936 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType())); 3937 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT, 3938 N0.getOperand(0).getOperand(0), Amt); 3939 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA); 3940 } 3941 } 3942 3943 // Simplify, based on bits shifted out of the LHS. 3944 if (N1C && SimplifyDemandedBits(SDValue(N, 0))) 3945 return SDValue(N, 0); 3946 3947 3948 // If the sign bit is known to be zero, switch this to a SRL. 3949 if (DAG.SignBitIsZero(N0)) 3950 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1); 3951 3952 if (N1C) { 3953 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue()); 3954 if (NewSRA.getNode()) 3955 return NewSRA; 3956 } 3957 3958 return SDValue(); 3959} 3960 3961SDValue DAGCombiner::visitSRL(SDNode *N) { 3962 SDValue N0 = N->getOperand(0); 3963 SDValue N1 = N->getOperand(1); 3964 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 3965 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 3966 EVT VT = N0.getValueType(); 3967 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits(); 3968 3969 // fold (srl c1, c2) -> c1 >>u c2 3970 if (N0C && N1C) 3971 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C); 3972 // fold (srl 0, x) -> 0 3973 if (N0C && N0C->isNullValue()) 3974 return N0; 3975 // fold (srl x, c >= size(x)) -> undef 3976 if (N1C && N1C->getZExtValue() >= OpSizeInBits) 3977 return DAG.getUNDEF(VT); 3978 // fold (srl x, 0) -> x 3979 if (N1C && N1C->isNullValue()) 3980 return N0; 3981 // if (srl x, c) is known to be zero, return 0 3982 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0), 3983 APInt::getAllOnesValue(OpSizeInBits))) 3984 return DAG.getConstant(0, VT); 3985 3986 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2)) 3987 if (N1C && N0.getOpcode() == ISD::SRL && 3988 N0.getOperand(1).getOpcode() == ISD::Constant) { 3989 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 3990 uint64_t c2 = N1C->getZExtValue(); 3991 if (c1 + c2 >= OpSizeInBits) 3992 return DAG.getConstant(0, VT); 3993 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), 3994 DAG.getConstant(c1 + c2, N1.getValueType())); 3995 } 3996 3997 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2))) 3998 if (N1C && N0.getOpcode() == ISD::TRUNCATE && 3999 N0.getOperand(0).getOpcode() == ISD::SRL && 4000 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) { 4001 uint64_t c1 = 4002 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue(); 4003 uint64_t c2 = N1C->getZExtValue(); 4004 EVT InnerShiftVT = N0.getOperand(0).getValueType(); 4005 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType(); 4006 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits(); 4007 // This is only valid if the OpSizeInBits + c1 = size of inner shift. 4008 if (c1 + OpSizeInBits == InnerShiftSize) { 4009 if (c1 + c2 >= InnerShiftSize) 4010 return DAG.getConstant(0, VT); 4011 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, 4012 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT, 4013 N0.getOperand(0)->getOperand(0), 4014 DAG.getConstant(c1 + c2, ShiftCountVT))); 4015 } 4016 } 4017 4018 // fold (srl (shl x, c), c) -> (and x, cst2) 4019 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 && 4020 N0.getValueSizeInBits() <= 64) { 4021 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits(); 4022 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0), 4023 DAG.getConstant(~0ULL >> ShAmt, VT)); 4024 } 4025 4026 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask) 4027 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { 4028 // Shifting in all undef bits? 4029 EVT SmallVT = N0.getOperand(0).getValueType(); 4030 if (N1C->getZExtValue() >= SmallVT.getSizeInBits()) 4031 return DAG.getUNDEF(VT); 4032 4033 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) { 4034 uint64_t ShiftAmt = N1C->getZExtValue(); 4035 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT, 4036 N0.getOperand(0), 4037 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT))); 4038 AddToWorkList(SmallShift.getNode()); 4039 APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits()).lshr(ShiftAmt); 4040 return DAG.getNode(ISD::AND, SDLoc(N), VT, 4041 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift), 4042 DAG.getConstant(Mask, VT)); 4043 } 4044 } 4045 4046 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign 4047 // bit, which is unmodified by sra. 4048 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) { 4049 if (N0.getOpcode() == ISD::SRA) 4050 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1); 4051 } 4052 4053 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit). 4054 if (N1C && N0.getOpcode() == ISD::CTLZ && 4055 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) { 4056 APInt KnownZero, KnownOne; 4057 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne); 4058 4059 // If any of the input bits are KnownOne, then the input couldn't be all 4060 // zeros, thus the result of the srl will always be zero. 4061 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT); 4062 4063 // If all of the bits input the to ctlz node are known to be zero, then 4064 // the result of the ctlz is "32" and the result of the shift is one. 4065 APInt UnknownBits = ~KnownZero; 4066 if (UnknownBits == 0) return DAG.getConstant(1, VT); 4067 4068 // Otherwise, check to see if there is exactly one bit input to the ctlz. 4069 if ((UnknownBits & (UnknownBits - 1)) == 0) { 4070 // Okay, we know that only that the single bit specified by UnknownBits 4071 // could be set on input to the CTLZ node. If this bit is set, the SRL 4072 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair 4073 // to an SRL/XOR pair, which is likely to simplify more. 4074 unsigned ShAmt = UnknownBits.countTrailingZeros(); 4075 SDValue Op = N0.getOperand(0); 4076 4077 if (ShAmt) { 4078 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op, 4079 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType()))); 4080 AddToWorkList(Op.getNode()); 4081 } 4082 4083 return DAG.getNode(ISD::XOR, SDLoc(N), VT, 4084 Op, DAG.getConstant(1, VT)); 4085 } 4086 } 4087 4088 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))). 4089 if (N1.getOpcode() == ISD::TRUNCATE && 4090 N1.getOperand(0).getOpcode() == ISD::AND && 4091 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) { 4092 SDValue N101 = N1.getOperand(0).getOperand(1); 4093 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) { 4094 EVT TruncVT = N1.getValueType(); 4095 SDValue N100 = N1.getOperand(0).getOperand(0); 4096 APInt TruncC = N101C->getAPIntValue(); 4097 TruncC = TruncC.trunc(TruncVT.getSizeInBits()); 4098 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, 4099 DAG.getNode(ISD::AND, SDLoc(N), 4100 TruncVT, 4101 DAG.getNode(ISD::TRUNCATE, 4102 SDLoc(N), 4103 TruncVT, N100), 4104 DAG.getConstant(TruncC, TruncVT))); 4105 } 4106 } 4107 4108 // fold operands of srl based on knowledge that the low bits are not 4109 // demanded. 4110 if (N1C && SimplifyDemandedBits(SDValue(N, 0))) 4111 return SDValue(N, 0); 4112 4113 if (N1C) { 4114 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue()); 4115 if (NewSRL.getNode()) 4116 return NewSRL; 4117 } 4118 4119 // Attempt to convert a srl of a load into a narrower zero-extending load. 4120 SDValue NarrowLoad = ReduceLoadWidth(N); 4121 if (NarrowLoad.getNode()) 4122 return NarrowLoad; 4123 4124 // Here is a common situation. We want to optimize: 4125 // 4126 // %a = ... 4127 // %b = and i32 %a, 2 4128 // %c = srl i32 %b, 1 4129 // brcond i32 %c ... 4130 // 4131 // into 4132 // 4133 // %a = ... 4134 // %b = and %a, 2 4135 // %c = setcc eq %b, 0 4136 // brcond %c ... 4137 // 4138 // However when after the source operand of SRL is optimized into AND, the SRL 4139 // itself may not be optimized further. Look for it and add the BRCOND into 4140 // the worklist. 4141 if (N->hasOneUse()) { 4142 SDNode *Use = *N->use_begin(); 4143 if (Use->getOpcode() == ISD::BRCOND) 4144 AddToWorkList(Use); 4145 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) { 4146 // Also look pass the truncate. 4147 Use = *Use->use_begin(); 4148 if (Use->getOpcode() == ISD::BRCOND) 4149 AddToWorkList(Use); 4150 } 4151 } 4152 4153 return SDValue(); 4154} 4155 4156SDValue DAGCombiner::visitCTLZ(SDNode *N) { 4157 SDValue N0 = N->getOperand(0); 4158 EVT VT = N->getValueType(0); 4159 4160 // fold (ctlz c1) -> c2 4161 if (isa<ConstantSDNode>(N0)) 4162 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0); 4163 return SDValue(); 4164} 4165 4166SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) { 4167 SDValue N0 = N->getOperand(0); 4168 EVT VT = N->getValueType(0); 4169 4170 // fold (ctlz_zero_undef c1) -> c2 4171 if (isa<ConstantSDNode>(N0)) 4172 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0); 4173 return SDValue(); 4174} 4175 4176SDValue DAGCombiner::visitCTTZ(SDNode *N) { 4177 SDValue N0 = N->getOperand(0); 4178 EVT VT = N->getValueType(0); 4179 4180 // fold (cttz c1) -> c2 4181 if (isa<ConstantSDNode>(N0)) 4182 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0); 4183 return SDValue(); 4184} 4185 4186SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) { 4187 SDValue N0 = N->getOperand(0); 4188 EVT VT = N->getValueType(0); 4189 4190 // fold (cttz_zero_undef c1) -> c2 4191 if (isa<ConstantSDNode>(N0)) 4192 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0); 4193 return SDValue(); 4194} 4195 4196SDValue DAGCombiner::visitCTPOP(SDNode *N) { 4197 SDValue N0 = N->getOperand(0); 4198 EVT VT = N->getValueType(0); 4199 4200 // fold (ctpop c1) -> c2 4201 if (isa<ConstantSDNode>(N0)) 4202 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0); 4203 return SDValue(); 4204} 4205 4206SDValue DAGCombiner::visitSELECT(SDNode *N) { 4207 SDValue N0 = N->getOperand(0); 4208 SDValue N1 = N->getOperand(1); 4209 SDValue N2 = N->getOperand(2); 4210 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 4211 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 4212 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); 4213 EVT VT = N->getValueType(0); 4214 EVT VT0 = N0.getValueType(); 4215 4216 // fold (select C, X, X) -> X 4217 if (N1 == N2) 4218 return N1; 4219 // fold (select true, X, Y) -> X 4220 if (N0C && !N0C->isNullValue()) 4221 return N1; 4222 // fold (select false, X, Y) -> Y 4223 if (N0C && N0C->isNullValue()) 4224 return N2; 4225 // fold (select C, 1, X) -> (or C, X) 4226 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1) 4227 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2); 4228 // fold (select C, 0, 1) -> (xor C, 1) 4229 if (VT.isInteger() && 4230 (VT0 == MVT::i1 || 4231 (VT0.isInteger() && 4232 TLI.getBooleanContents(false) == 4233 TargetLowering::ZeroOrOneBooleanContent)) && 4234 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) { 4235 SDValue XORNode; 4236 if (VT == VT0) 4237 return DAG.getNode(ISD::XOR, SDLoc(N), VT0, 4238 N0, DAG.getConstant(1, VT0)); 4239 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0, 4240 N0, DAG.getConstant(1, VT0)); 4241 AddToWorkList(XORNode.getNode()); 4242 if (VT.bitsGT(VT0)) 4243 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode); 4244 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode); 4245 } 4246 // fold (select C, 0, X) -> (and (not C), X) 4247 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) { 4248 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT); 4249 AddToWorkList(NOTNode.getNode()); 4250 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2); 4251 } 4252 // fold (select C, X, 1) -> (or (not C), X) 4253 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) { 4254 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT); 4255 AddToWorkList(NOTNode.getNode()); 4256 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1); 4257 } 4258 // fold (select C, X, 0) -> (and C, X) 4259 if (VT == MVT::i1 && N2C && N2C->isNullValue()) 4260 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1); 4261 // fold (select X, X, Y) -> (or X, Y) 4262 // fold (select X, 1, Y) -> (or X, Y) 4263 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1))) 4264 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2); 4265 // fold (select X, Y, X) -> (and X, Y) 4266 // fold (select X, Y, 0) -> (and X, Y) 4267 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0))) 4268 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1); 4269 4270 // If we can fold this based on the true/false value, do so. 4271 if (SimplifySelectOps(N, N1, N2)) 4272 return SDValue(N, 0); // Don't revisit N. 4273 4274 // fold selects based on a setcc into other things, such as min/max/abs 4275 if (N0.getOpcode() == ISD::SETCC) { 4276 // FIXME: 4277 // Check against MVT::Other for SELECT_CC, which is a workaround for targets 4278 // having to say they don't support SELECT_CC on every type the DAG knows 4279 // about, since there is no way to mark an opcode illegal at all value types 4280 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) && 4281 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) 4282 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, 4283 N0.getOperand(0), N0.getOperand(1), 4284 N1, N2, N0.getOperand(2)); 4285 return SimplifySelect(SDLoc(N), N0, N1, N2); 4286 } 4287 4288 return SDValue(); 4289} 4290 4291SDValue DAGCombiner::visitVSELECT(SDNode *N) { 4292 SDValue N0 = N->getOperand(0); 4293 SDValue N1 = N->getOperand(1); 4294 SDValue N2 = N->getOperand(2); 4295 SDLoc DL(N); 4296 4297 // Canonicalize integer abs. 4298 // vselect (setg[te] X, 0), X, -X -> 4299 // vselect (setgt X, -1), X, -X -> 4300 // vselect (setl[te] X, 0), -X, X -> 4301 // Y = sra (X, size(X)-1); xor (add (X, Y), Y) 4302 if (N0.getOpcode() == ISD::SETCC) { 4303 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1); 4304 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 4305 bool isAbs = false; 4306 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode()); 4307 4308 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) || 4309 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) && 4310 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1)) 4311 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode()); 4312 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) && 4313 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1)) 4314 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode()); 4315 4316 if (isAbs) { 4317 EVT VT = LHS.getValueType(); 4318 SDValue Shift = DAG.getNode( 4319 ISD::SRA, DL, VT, LHS, 4320 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT)); 4321 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift); 4322 AddToWorkList(Shift.getNode()); 4323 AddToWorkList(Add.getNode()); 4324 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift); 4325 } 4326 } 4327 4328 return SDValue(); 4329} 4330 4331SDValue DAGCombiner::visitSELECT_CC(SDNode *N) { 4332 SDValue N0 = N->getOperand(0); 4333 SDValue N1 = N->getOperand(1); 4334 SDValue N2 = N->getOperand(2); 4335 SDValue N3 = N->getOperand(3); 4336 SDValue N4 = N->getOperand(4); 4337 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get(); 4338 4339 // fold select_cc lhs, rhs, x, x, cc -> x 4340 if (N2 == N3) 4341 return N2; 4342 4343 // Determine if the condition we're dealing with is constant 4344 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()), 4345 N0, N1, CC, SDLoc(N), false); 4346 if (SCC.getNode()) { 4347 AddToWorkList(SCC.getNode()); 4348 4349 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) { 4350 if (!SCCC->isNullValue()) 4351 return N2; // cond always true -> true val 4352 else 4353 return N3; // cond always false -> false val 4354 } 4355 4356 // Fold to a simpler select_cc 4357 if (SCC.getOpcode() == ISD::SETCC) 4358 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(), 4359 SCC.getOperand(0), SCC.getOperand(1), N2, N3, 4360 SCC.getOperand(2)); 4361 } 4362 4363 // If we can fold this based on the true/false value, do so. 4364 if (SimplifySelectOps(N, N2, N3)) 4365 return SDValue(N, 0); // Don't revisit N. 4366 4367 // fold select_cc into other things, such as min/max/abs 4368 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC); 4369} 4370 4371SDValue DAGCombiner::visitSETCC(SDNode *N) { 4372 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1), 4373 cast<CondCodeSDNode>(N->getOperand(2))->get(), 4374 SDLoc(N)); 4375} 4376 4377// ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this: 4378// "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))" 4379// transformation. Returns true if extension are possible and the above 4380// mentioned transformation is profitable. 4381static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0, 4382 unsigned ExtOpc, 4383 SmallVectorImpl<SDNode *> &ExtendNodes, 4384 const TargetLowering &TLI) { 4385 bool HasCopyToRegUses = false; 4386 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType()); 4387 for (SDNode::use_iterator UI = N0.getNode()->use_begin(), 4388 UE = N0.getNode()->use_end(); 4389 UI != UE; ++UI) { 4390 SDNode *User = *UI; 4391 if (User == N) 4392 continue; 4393 if (UI.getUse().getResNo() != N0.getResNo()) 4394 continue; 4395 // FIXME: Only extend SETCC N, N and SETCC N, c for now. 4396 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) { 4397 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get(); 4398 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC)) 4399 // Sign bits will be lost after a zext. 4400 return false; 4401 bool Add = false; 4402 for (unsigned i = 0; i != 2; ++i) { 4403 SDValue UseOp = User->getOperand(i); 4404 if (UseOp == N0) 4405 continue; 4406 if (!isa<ConstantSDNode>(UseOp)) 4407 return false; 4408 Add = true; 4409 } 4410 if (Add) 4411 ExtendNodes.push_back(User); 4412 continue; 4413 } 4414 // If truncates aren't free and there are users we can't 4415 // extend, it isn't worthwhile. 4416 if (!isTruncFree) 4417 return false; 4418 // Remember if this value is live-out. 4419 if (User->getOpcode() == ISD::CopyToReg) 4420 HasCopyToRegUses = true; 4421 } 4422 4423 if (HasCopyToRegUses) { 4424 bool BothLiveOut = false; 4425 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 4426 UI != UE; ++UI) { 4427 SDUse &Use = UI.getUse(); 4428 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) { 4429 BothLiveOut = true; 4430 break; 4431 } 4432 } 4433 if (BothLiveOut) 4434 // Both unextended and extended values are live out. There had better be 4435 // a good reason for the transformation. 4436 return ExtendNodes.size(); 4437 } 4438 return true; 4439} 4440 4441void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs, 4442 SDValue Trunc, SDValue ExtLoad, SDLoc DL, 4443 ISD::NodeType ExtType) { 4444 // Extend SetCC uses if necessary. 4445 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) { 4446 SDNode *SetCC = SetCCs[i]; 4447 SmallVector<SDValue, 4> Ops; 4448 4449 for (unsigned j = 0; j != 2; ++j) { 4450 SDValue SOp = SetCC->getOperand(j); 4451 if (SOp == Trunc) 4452 Ops.push_back(ExtLoad); 4453 else 4454 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp)); 4455 } 4456 4457 Ops.push_back(SetCC->getOperand(2)); 4458 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), 4459 &Ops[0], Ops.size())); 4460 } 4461} 4462 4463SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) { 4464 SDValue N0 = N->getOperand(0); 4465 EVT VT = N->getValueType(0); 4466 4467 // fold (sext c1) -> c1 4468 if (isa<ConstantSDNode>(N0)) 4469 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N0); 4470 4471 // fold (sext (sext x)) -> (sext x) 4472 // fold (sext (aext x)) -> (sext x) 4473 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) 4474 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, 4475 N0.getOperand(0)); 4476 4477 if (N0.getOpcode() == ISD::TRUNCATE) { 4478 // fold (sext (truncate (load x))) -> (sext (smaller load x)) 4479 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n))) 4480 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode()); 4481 if (NarrowLoad.getNode()) { 4482 SDNode* oye = N0.getNode()->getOperand(0).getNode(); 4483 if (NarrowLoad.getNode() != N0.getNode()) { 4484 CombineTo(N0.getNode(), NarrowLoad); 4485 // CombineTo deleted the truncate, if needed, but not what's under it. 4486 AddToWorkList(oye); 4487 } 4488 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4489 } 4490 4491 // See if the value being truncated is already sign extended. If so, just 4492 // eliminate the trunc/sext pair. 4493 SDValue Op = N0.getOperand(0); 4494 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits(); 4495 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits(); 4496 unsigned DestBits = VT.getScalarType().getSizeInBits(); 4497 unsigned NumSignBits = DAG.ComputeNumSignBits(Op); 4498 4499 if (OpBits == DestBits) { 4500 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign 4501 // bits, it is already ready. 4502 if (NumSignBits > DestBits-MidBits) 4503 return Op; 4504 } else if (OpBits < DestBits) { 4505 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign 4506 // bits, just sext from i32. 4507 if (NumSignBits > OpBits-MidBits) 4508 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op); 4509 } else { 4510 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign 4511 // bits, just truncate to i32. 4512 if (NumSignBits > OpBits-MidBits) 4513 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op); 4514 } 4515 4516 // fold (sext (truncate x)) -> (sextinreg x). 4517 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, 4518 N0.getValueType())) { 4519 if (OpBits < DestBits) 4520 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op); 4521 else if (OpBits > DestBits) 4522 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op); 4523 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op, 4524 DAG.getValueType(N0.getValueType())); 4525 } 4526 } 4527 4528 // fold (sext (load x)) -> (sext (truncate (sextload x))) 4529 // None of the supported targets knows how to perform load and sign extend 4530 // on vectors in one instruction. We only perform this transformation on 4531 // scalars. 4532 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() && 4533 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 4534 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) { 4535 bool DoXform = true; 4536 SmallVector<SDNode*, 4> SetCCs; 4537 if (!N0.hasOneUse()) 4538 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI); 4539 if (DoXform) { 4540 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 4541 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, 4542 LN0->getChain(), 4543 LN0->getBasePtr(), LN0->getPointerInfo(), 4544 N0.getValueType(), 4545 LN0->isVolatile(), LN0->isNonTemporal(), 4546 LN0->getAlignment()); 4547 CombineTo(N, ExtLoad); 4548 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 4549 N0.getValueType(), ExtLoad); 4550 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1)); 4551 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 4552 ISD::SIGN_EXTEND); 4553 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4554 } 4555 } 4556 4557 // fold (sext (sextload x)) -> (sext (truncate (sextload x))) 4558 // fold (sext ( extload x)) -> (sext (truncate (sextload x))) 4559 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) && 4560 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) { 4561 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 4562 EVT MemVT = LN0->getMemoryVT(); 4563 if ((!LegalOperations && !LN0->isVolatile()) || 4564 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) { 4565 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, 4566 LN0->getChain(), 4567 LN0->getBasePtr(), LN0->getPointerInfo(), 4568 MemVT, 4569 LN0->isVolatile(), LN0->isNonTemporal(), 4570 LN0->getAlignment()); 4571 CombineTo(N, ExtLoad); 4572 CombineTo(N0.getNode(), 4573 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 4574 N0.getValueType(), ExtLoad), 4575 ExtLoad.getValue(1)); 4576 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4577 } 4578 } 4579 4580 // fold (sext (and/or/xor (load x), cst)) -> 4581 // (and/or/xor (sextload x), (sext cst)) 4582 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR || 4583 N0.getOpcode() == ISD::XOR) && 4584 isa<LoadSDNode>(N0.getOperand(0)) && 4585 N0.getOperand(1).getOpcode() == ISD::Constant && 4586 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) && 4587 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) { 4588 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0)); 4589 if (LN0->getExtensionType() != ISD::ZEXTLOAD) { 4590 bool DoXform = true; 4591 SmallVector<SDNode*, 4> SetCCs; 4592 if (!N0.hasOneUse()) 4593 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND, 4594 SetCCs, TLI); 4595 if (DoXform) { 4596 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT, 4597 LN0->getChain(), LN0->getBasePtr(), 4598 LN0->getPointerInfo(), 4599 LN0->getMemoryVT(), 4600 LN0->isVolatile(), 4601 LN0->isNonTemporal(), 4602 LN0->getAlignment()); 4603 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 4604 Mask = Mask.sext(VT.getSizeInBits()); 4605 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT, 4606 ExtLoad, DAG.getConstant(Mask, VT)); 4607 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, 4608 SDLoc(N0.getOperand(0)), 4609 N0.getOperand(0).getValueType(), ExtLoad); 4610 CombineTo(N, And); 4611 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1)); 4612 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 4613 ISD::SIGN_EXTEND); 4614 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4615 } 4616 } 4617 } 4618 4619 if (N0.getOpcode() == ISD::SETCC) { 4620 // sext(setcc) -> sext_in_reg(vsetcc) for vectors. 4621 // Only do this before legalize for now. 4622 if (VT.isVector() && !LegalOperations && 4623 TLI.getBooleanContents(true) == 4624 TargetLowering::ZeroOrNegativeOneBooleanContent) { 4625 EVT N0VT = N0.getOperand(0).getValueType(); 4626 // On some architectures (such as SSE/NEON/etc) the SETCC result type is 4627 // of the same size as the compared operands. Only optimize sext(setcc()) 4628 // if this is the case. 4629 EVT SVT = getSetCCResultType(N0VT); 4630 4631 // We know that the # elements of the results is the same as the 4632 // # elements of the compare (and the # elements of the compare result 4633 // for that matter). Check to see that they are the same size. If so, 4634 // we know that the element size of the sext'd result matches the 4635 // element size of the compare operands. 4636 if (VT.getSizeInBits() == SVT.getSizeInBits()) 4637 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0), 4638 N0.getOperand(1), 4639 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 4640 4641 // If the desired elements are smaller or larger than the source 4642 // elements we can use a matching integer vector type and then 4643 // truncate/sign extend 4644 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger(); 4645 if (SVT == MatchingVectorType) { 4646 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType, 4647 N0.getOperand(0), N0.getOperand(1), 4648 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 4649 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT); 4650 } 4651 } 4652 4653 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc) 4654 unsigned ElementWidth = VT.getScalarType().getSizeInBits(); 4655 SDValue NegOne = 4656 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT); 4657 SDValue SCC = 4658 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1), 4659 NegOne, DAG.getConstant(0, VT), 4660 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true); 4661 if (SCC.getNode()) return SCC; 4662 if (!VT.isVector() && 4663 (!LegalOperations || 4664 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(VT)))) { 4665 return DAG.getSelect(SDLoc(N), VT, 4666 DAG.getSetCC(SDLoc(N), 4667 getSetCCResultType(VT), 4668 N0.getOperand(0), N0.getOperand(1), 4669 cast<CondCodeSDNode>(N0.getOperand(2))->get()), 4670 NegOne, DAG.getConstant(0, VT)); 4671 } 4672 } 4673 4674 // fold (sext x) -> (zext x) if the sign bit is known zero. 4675 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) && 4676 DAG.SignBitIsZero(N0)) 4677 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0); 4678 4679 return SDValue(); 4680} 4681 4682// isTruncateOf - If N is a truncate of some other value, return true, record 4683// the value being truncated in Op and which of Op's bits are zero in KnownZero. 4684// This function computes KnownZero to avoid a duplicated call to 4685// ComputeMaskedBits in the caller. 4686static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op, 4687 APInt &KnownZero) { 4688 APInt KnownOne; 4689 if (N->getOpcode() == ISD::TRUNCATE) { 4690 Op = N->getOperand(0); 4691 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne); 4692 return true; 4693 } 4694 4695 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 || 4696 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE) 4697 return false; 4698 4699 SDValue Op0 = N->getOperand(0); 4700 SDValue Op1 = N->getOperand(1); 4701 assert(Op0.getValueType() == Op1.getValueType()); 4702 4703 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0); 4704 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1); 4705 if (COp0 && COp0->isNullValue()) 4706 Op = Op1; 4707 else if (COp1 && COp1->isNullValue()) 4708 Op = Op0; 4709 else 4710 return false; 4711 4712 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne); 4713 4714 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue()) 4715 return false; 4716 4717 return true; 4718} 4719 4720SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) { 4721 SDValue N0 = N->getOperand(0); 4722 EVT VT = N->getValueType(0); 4723 4724 // fold (zext c1) -> c1 4725 if (isa<ConstantSDNode>(N0)) 4726 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0); 4727 // fold (zext (zext x)) -> (zext x) 4728 // fold (zext (aext x)) -> (zext x) 4729 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) 4730 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, 4731 N0.getOperand(0)); 4732 4733 // fold (zext (truncate x)) -> (zext x) or 4734 // (zext (truncate x)) -> (truncate x) 4735 // This is valid when the truncated bits of x are already zero. 4736 // FIXME: We should extend this to work for vectors too. 4737 SDValue Op; 4738 APInt KnownZero; 4739 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) { 4740 APInt TruncatedBits = 4741 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ? 4742 APInt(Op.getValueSizeInBits(), 0) : 4743 APInt::getBitsSet(Op.getValueSizeInBits(), 4744 N0.getValueSizeInBits(), 4745 std::min(Op.getValueSizeInBits(), 4746 VT.getSizeInBits())); 4747 if (TruncatedBits == (KnownZero & TruncatedBits)) { 4748 if (VT.bitsGT(Op.getValueType())) 4749 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op); 4750 if (VT.bitsLT(Op.getValueType())) 4751 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op); 4752 4753 return Op; 4754 } 4755 } 4756 4757 // fold (zext (truncate (load x))) -> (zext (smaller load x)) 4758 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n))) 4759 if (N0.getOpcode() == ISD::TRUNCATE) { 4760 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode()); 4761 if (NarrowLoad.getNode()) { 4762 SDNode* oye = N0.getNode()->getOperand(0).getNode(); 4763 if (NarrowLoad.getNode() != N0.getNode()) { 4764 CombineTo(N0.getNode(), NarrowLoad); 4765 // CombineTo deleted the truncate, if needed, but not what's under it. 4766 AddToWorkList(oye); 4767 } 4768 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4769 } 4770 } 4771 4772 // fold (zext (truncate x)) -> (and x, mask) 4773 if (N0.getOpcode() == ISD::TRUNCATE && 4774 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) { 4775 4776 // fold (zext (truncate (load x))) -> (zext (smaller load x)) 4777 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n))) 4778 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode()); 4779 if (NarrowLoad.getNode()) { 4780 SDNode* oye = N0.getNode()->getOperand(0).getNode(); 4781 if (NarrowLoad.getNode() != N0.getNode()) { 4782 CombineTo(N0.getNode(), NarrowLoad); 4783 // CombineTo deleted the truncate, if needed, but not what's under it. 4784 AddToWorkList(oye); 4785 } 4786 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4787 } 4788 4789 SDValue Op = N0.getOperand(0); 4790 if (Op.getValueType().bitsLT(VT)) { 4791 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op); 4792 AddToWorkList(Op.getNode()); 4793 } else if (Op.getValueType().bitsGT(VT)) { 4794 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op); 4795 AddToWorkList(Op.getNode()); 4796 } 4797 return DAG.getZeroExtendInReg(Op, SDLoc(N), 4798 N0.getValueType().getScalarType()); 4799 } 4800 4801 // Fold (zext (and (trunc x), cst)) -> (and x, cst), 4802 // if either of the casts is not free. 4803 if (N0.getOpcode() == ISD::AND && 4804 N0.getOperand(0).getOpcode() == ISD::TRUNCATE && 4805 N0.getOperand(1).getOpcode() == ISD::Constant && 4806 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(), 4807 N0.getValueType()) || 4808 !TLI.isZExtFree(N0.getValueType(), VT))) { 4809 SDValue X = N0.getOperand(0).getOperand(0); 4810 if (X.getValueType().bitsLT(VT)) { 4811 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X); 4812 } else if (X.getValueType().bitsGT(VT)) { 4813 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X); 4814 } 4815 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 4816 Mask = Mask.zext(VT.getSizeInBits()); 4817 return DAG.getNode(ISD::AND, SDLoc(N), VT, 4818 X, DAG.getConstant(Mask, VT)); 4819 } 4820 4821 // fold (zext (load x)) -> (zext (truncate (zextload x))) 4822 // None of the supported targets knows how to perform load and vector_zext 4823 // on vectors in one instruction. We only perform this transformation on 4824 // scalars. 4825 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() && 4826 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 4827 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) { 4828 bool DoXform = true; 4829 SmallVector<SDNode*, 4> SetCCs; 4830 if (!N0.hasOneUse()) 4831 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI); 4832 if (DoXform) { 4833 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 4834 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT, 4835 LN0->getChain(), 4836 LN0->getBasePtr(), LN0->getPointerInfo(), 4837 N0.getValueType(), 4838 LN0->isVolatile(), LN0->isNonTemporal(), 4839 LN0->getAlignment()); 4840 CombineTo(N, ExtLoad); 4841 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 4842 N0.getValueType(), ExtLoad); 4843 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1)); 4844 4845 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 4846 ISD::ZERO_EXTEND); 4847 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4848 } 4849 } 4850 4851 // fold (zext (and/or/xor (load x), cst)) -> 4852 // (and/or/xor (zextload x), (zext cst)) 4853 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR || 4854 N0.getOpcode() == ISD::XOR) && 4855 isa<LoadSDNode>(N0.getOperand(0)) && 4856 N0.getOperand(1).getOpcode() == ISD::Constant && 4857 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) && 4858 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) { 4859 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0)); 4860 if (LN0->getExtensionType() != ISD::SEXTLOAD) { 4861 bool DoXform = true; 4862 SmallVector<SDNode*, 4> SetCCs; 4863 if (!N0.hasOneUse()) 4864 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND, 4865 SetCCs, TLI); 4866 if (DoXform) { 4867 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT, 4868 LN0->getChain(), LN0->getBasePtr(), 4869 LN0->getPointerInfo(), 4870 LN0->getMemoryVT(), 4871 LN0->isVolatile(), 4872 LN0->isNonTemporal(), 4873 LN0->getAlignment()); 4874 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 4875 Mask = Mask.zext(VT.getSizeInBits()); 4876 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT, 4877 ExtLoad, DAG.getConstant(Mask, VT)); 4878 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, 4879 SDLoc(N0.getOperand(0)), 4880 N0.getOperand(0).getValueType(), ExtLoad); 4881 CombineTo(N, And); 4882 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1)); 4883 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 4884 ISD::ZERO_EXTEND); 4885 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4886 } 4887 } 4888 } 4889 4890 // fold (zext (zextload x)) -> (zext (truncate (zextload x))) 4891 // fold (zext ( extload x)) -> (zext (truncate (zextload x))) 4892 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) && 4893 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) { 4894 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 4895 EVT MemVT = LN0->getMemoryVT(); 4896 if ((!LegalOperations && !LN0->isVolatile()) || 4897 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) { 4898 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT, 4899 LN0->getChain(), 4900 LN0->getBasePtr(), LN0->getPointerInfo(), 4901 MemVT, 4902 LN0->isVolatile(), LN0->isNonTemporal(), 4903 LN0->getAlignment()); 4904 CombineTo(N, ExtLoad); 4905 CombineTo(N0.getNode(), 4906 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), 4907 ExtLoad), 4908 ExtLoad.getValue(1)); 4909 return SDValue(N, 0); // Return N so it doesn't get rechecked! 4910 } 4911 } 4912 4913 if (N0.getOpcode() == ISD::SETCC) { 4914 if (!LegalOperations && VT.isVector()) { 4915 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors. 4916 // Only do this before legalize for now. 4917 EVT N0VT = N0.getOperand(0).getValueType(); 4918 EVT EltVT = VT.getVectorElementType(); 4919 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(), 4920 DAG.getConstant(1, EltVT)); 4921 if (VT.getSizeInBits() == N0VT.getSizeInBits()) 4922 // We know that the # elements of the results is the same as the 4923 // # elements of the compare (and the # elements of the compare result 4924 // for that matter). Check to see that they are the same size. If so, 4925 // we know that the element size of the sext'd result matches the 4926 // element size of the compare operands. 4927 return DAG.getNode(ISD::AND, SDLoc(N), VT, 4928 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0), 4929 N0.getOperand(1), 4930 cast<CondCodeSDNode>(N0.getOperand(2))->get()), 4931 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, 4932 &OneOps[0], OneOps.size())); 4933 4934 // If the desired elements are smaller or larger than the source 4935 // elements we can use a matching integer vector type and then 4936 // truncate/sign extend 4937 EVT MatchingElementType = 4938 EVT::getIntegerVT(*DAG.getContext(), 4939 N0VT.getScalarType().getSizeInBits()); 4940 EVT MatchingVectorType = 4941 EVT::getVectorVT(*DAG.getContext(), MatchingElementType, 4942 N0VT.getVectorNumElements()); 4943 SDValue VsetCC = 4944 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0), 4945 N0.getOperand(1), 4946 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 4947 return DAG.getNode(ISD::AND, SDLoc(N), VT, 4948 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT), 4949 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, 4950 &OneOps[0], OneOps.size())); 4951 } 4952 4953 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc 4954 SDValue SCC = 4955 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1), 4956 DAG.getConstant(1, VT), DAG.getConstant(0, VT), 4957 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true); 4958 if (SCC.getNode()) return SCC; 4959 } 4960 4961 // (zext (shl (zext x), cst)) -> (shl (zext x), cst) 4962 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) && 4963 isa<ConstantSDNode>(N0.getOperand(1)) && 4964 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND && 4965 N0.hasOneUse()) { 4966 SDValue ShAmt = N0.getOperand(1); 4967 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue(); 4968 if (N0.getOpcode() == ISD::SHL) { 4969 SDValue InnerZExt = N0.getOperand(0); 4970 // If the original shl may be shifting out bits, do not perform this 4971 // transformation. 4972 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() - 4973 InnerZExt.getOperand(0).getValueType().getSizeInBits(); 4974 if (ShAmtVal > KnownZeroBits) 4975 return SDValue(); 4976 } 4977 4978 SDLoc DL(N); 4979 4980 // Ensure that the shift amount is wide enough for the shifted value. 4981 if (VT.getSizeInBits() >= 256) 4982 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt); 4983 4984 return DAG.getNode(N0.getOpcode(), DL, VT, 4985 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)), 4986 ShAmt); 4987 } 4988 4989 return SDValue(); 4990} 4991 4992SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) { 4993 SDValue N0 = N->getOperand(0); 4994 EVT VT = N->getValueType(0); 4995 4996 // fold (aext c1) -> c1 4997 if (isa<ConstantSDNode>(N0)) 4998 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, N0); 4999 // fold (aext (aext x)) -> (aext x) 5000 // fold (aext (zext x)) -> (zext x) 5001 // fold (aext (sext x)) -> (sext x) 5002 if (N0.getOpcode() == ISD::ANY_EXTEND || 5003 N0.getOpcode() == ISD::ZERO_EXTEND || 5004 N0.getOpcode() == ISD::SIGN_EXTEND) 5005 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0)); 5006 5007 // fold (aext (truncate (load x))) -> (aext (smaller load x)) 5008 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n))) 5009 if (N0.getOpcode() == ISD::TRUNCATE) { 5010 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode()); 5011 if (NarrowLoad.getNode()) { 5012 SDNode* oye = N0.getNode()->getOperand(0).getNode(); 5013 if (NarrowLoad.getNode() != N0.getNode()) { 5014 CombineTo(N0.getNode(), NarrowLoad); 5015 // CombineTo deleted the truncate, if needed, but not what's under it. 5016 AddToWorkList(oye); 5017 } 5018 return SDValue(N, 0); // Return N so it doesn't get rechecked! 5019 } 5020 } 5021 5022 // fold (aext (truncate x)) 5023 if (N0.getOpcode() == ISD::TRUNCATE) { 5024 SDValue TruncOp = N0.getOperand(0); 5025 if (TruncOp.getValueType() == VT) 5026 return TruncOp; // x iff x size == zext size. 5027 if (TruncOp.getValueType().bitsGT(VT)) 5028 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp); 5029 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp); 5030 } 5031 5032 // Fold (aext (and (trunc x), cst)) -> (and x, cst) 5033 // if the trunc is not free. 5034 if (N0.getOpcode() == ISD::AND && 5035 N0.getOperand(0).getOpcode() == ISD::TRUNCATE && 5036 N0.getOperand(1).getOpcode() == ISD::Constant && 5037 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(), 5038 N0.getValueType())) { 5039 SDValue X = N0.getOperand(0).getOperand(0); 5040 if (X.getValueType().bitsLT(VT)) { 5041 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X); 5042 } else if (X.getValueType().bitsGT(VT)) { 5043 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X); 5044 } 5045 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 5046 Mask = Mask.zext(VT.getSizeInBits()); 5047 return DAG.getNode(ISD::AND, SDLoc(N), VT, 5048 X, DAG.getConstant(Mask, VT)); 5049 } 5050 5051 // fold (aext (load x)) -> (aext (truncate (extload x))) 5052 // None of the supported targets knows how to perform load and any_ext 5053 // on vectors in one instruction. We only perform this transformation on 5054 // scalars. 5055 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() && 5056 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 5057 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) { 5058 bool DoXform = true; 5059 SmallVector<SDNode*, 4> SetCCs; 5060 if (!N0.hasOneUse()) 5061 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI); 5062 if (DoXform) { 5063 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5064 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT, 5065 LN0->getChain(), 5066 LN0->getBasePtr(), LN0->getPointerInfo(), 5067 N0.getValueType(), 5068 LN0->isVolatile(), LN0->isNonTemporal(), 5069 LN0->getAlignment()); 5070 CombineTo(N, ExtLoad); 5071 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 5072 N0.getValueType(), ExtLoad); 5073 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1)); 5074 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N), 5075 ISD::ANY_EXTEND); 5076 return SDValue(N, 0); // Return N so it doesn't get rechecked! 5077 } 5078 } 5079 5080 // fold (aext (zextload x)) -> (aext (truncate (zextload x))) 5081 // fold (aext (sextload x)) -> (aext (truncate (sextload x))) 5082 // fold (aext ( extload x)) -> (aext (truncate (extload x))) 5083 if (N0.getOpcode() == ISD::LOAD && 5084 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && 5085 N0.hasOneUse()) { 5086 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5087 EVT MemVT = LN0->getMemoryVT(); 5088 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(N), 5089 VT, LN0->getChain(), LN0->getBasePtr(), 5090 LN0->getPointerInfo(), MemVT, 5091 LN0->isVolatile(), LN0->isNonTemporal(), 5092 LN0->getAlignment()); 5093 CombineTo(N, ExtLoad); 5094 CombineTo(N0.getNode(), 5095 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), 5096 N0.getValueType(), ExtLoad), 5097 ExtLoad.getValue(1)); 5098 return SDValue(N, 0); // Return N so it doesn't get rechecked! 5099 } 5100 5101 if (N0.getOpcode() == ISD::SETCC) { 5102 // aext(setcc) -> sext_in_reg(vsetcc) for vectors. 5103 // Only do this before legalize for now. 5104 if (VT.isVector() && !LegalOperations) { 5105 EVT N0VT = N0.getOperand(0).getValueType(); 5106 // We know that the # elements of the results is the same as the 5107 // # elements of the compare (and the # elements of the compare result 5108 // for that matter). Check to see that they are the same size. If so, 5109 // we know that the element size of the sext'd result matches the 5110 // element size of the compare operands. 5111 if (VT.getSizeInBits() == N0VT.getSizeInBits()) 5112 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0), 5113 N0.getOperand(1), 5114 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 5115 // If the desired elements are smaller or larger than the source 5116 // elements we can use a matching integer vector type and then 5117 // truncate/sign extend 5118 else { 5119 EVT MatchingElementType = 5120 EVT::getIntegerVT(*DAG.getContext(), 5121 N0VT.getScalarType().getSizeInBits()); 5122 EVT MatchingVectorType = 5123 EVT::getVectorVT(*DAG.getContext(), MatchingElementType, 5124 N0VT.getVectorNumElements()); 5125 SDValue VsetCC = 5126 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0), 5127 N0.getOperand(1), 5128 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 5129 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT); 5130 } 5131 } 5132 5133 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc 5134 SDValue SCC = 5135 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1), 5136 DAG.getConstant(1, VT), DAG.getConstant(0, VT), 5137 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true); 5138 if (SCC.getNode()) 5139 return SCC; 5140 } 5141 5142 return SDValue(); 5143} 5144 5145/// GetDemandedBits - See if the specified operand can be simplified with the 5146/// knowledge that only the bits specified by Mask are used. If so, return the 5147/// simpler operand, otherwise return a null SDValue. 5148SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) { 5149 switch (V.getOpcode()) { 5150 default: break; 5151 case ISD::Constant: { 5152 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode()); 5153 assert(CV != 0 && "Const value should be ConstSDNode."); 5154 const APInt &CVal = CV->getAPIntValue(); 5155 APInt NewVal = CVal & Mask; 5156 if (NewVal != CVal) 5157 return DAG.getConstant(NewVal, V.getValueType()); 5158 break; 5159 } 5160 case ISD::OR: 5161 case ISD::XOR: 5162 // If the LHS or RHS don't contribute bits to the or, drop them. 5163 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask)) 5164 return V.getOperand(1); 5165 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask)) 5166 return V.getOperand(0); 5167 break; 5168 case ISD::SRL: 5169 // Only look at single-use SRLs. 5170 if (!V.getNode()->hasOneUse()) 5171 break; 5172 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) { 5173 // See if we can recursively simplify the LHS. 5174 unsigned Amt = RHSC->getZExtValue(); 5175 5176 // Watch out for shift count overflow though. 5177 if (Amt >= Mask.getBitWidth()) break; 5178 APInt NewMask = Mask << Amt; 5179 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask); 5180 if (SimplifyLHS.getNode()) 5181 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(), 5182 SimplifyLHS, V.getOperand(1)); 5183 } 5184 } 5185 return SDValue(); 5186} 5187 5188/// ReduceLoadWidth - If the result of a wider load is shifted to right of N 5189/// bits and then truncated to a narrower type and where N is a multiple 5190/// of number of bits of the narrower type, transform it to a narrower load 5191/// from address + N / num of bits of new type. If the result is to be 5192/// extended, also fold the extension to form a extending load. 5193SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) { 5194 unsigned Opc = N->getOpcode(); 5195 5196 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD; 5197 SDValue N0 = N->getOperand(0); 5198 EVT VT = N->getValueType(0); 5199 EVT ExtVT = VT; 5200 5201 // This transformation isn't valid for vector loads. 5202 if (VT.isVector()) 5203 return SDValue(); 5204 5205 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then 5206 // extended to VT. 5207 if (Opc == ISD::SIGN_EXTEND_INREG) { 5208 ExtType = ISD::SEXTLOAD; 5209 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT(); 5210 } else if (Opc == ISD::SRL) { 5211 // Another special-case: SRL is basically zero-extending a narrower value. 5212 ExtType = ISD::ZEXTLOAD; 5213 N0 = SDValue(N, 0); 5214 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 5215 if (!N01) return SDValue(); 5216 ExtVT = EVT::getIntegerVT(*DAG.getContext(), 5217 VT.getSizeInBits() - N01->getZExtValue()); 5218 } 5219 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT)) 5220 return SDValue(); 5221 5222 unsigned EVTBits = ExtVT.getSizeInBits(); 5223 5224 // Do not generate loads of non-round integer types since these can 5225 // be expensive (and would be wrong if the type is not byte sized). 5226 if (!ExtVT.isRound()) 5227 return SDValue(); 5228 5229 unsigned ShAmt = 0; 5230 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) { 5231 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 5232 ShAmt = N01->getZExtValue(); 5233 // Is the shift amount a multiple of size of VT? 5234 if ((ShAmt & (EVTBits-1)) == 0) { 5235 N0 = N0.getOperand(0); 5236 // Is the load width a multiple of size of VT? 5237 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0) 5238 return SDValue(); 5239 } 5240 5241 // At this point, we must have a load or else we can't do the transform. 5242 if (!isa<LoadSDNode>(N0)) return SDValue(); 5243 5244 // Because a SRL must be assumed to *need* to zero-extend the high bits 5245 // (as opposed to anyext the high bits), we can't combine the zextload 5246 // lowering of SRL and an sextload. 5247 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD) 5248 return SDValue(); 5249 5250 // If the shift amount is larger than the input type then we're not 5251 // accessing any of the loaded bytes. If the load was a zextload/extload 5252 // then the result of the shift+trunc is zero/undef (handled elsewhere). 5253 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits()) 5254 return SDValue(); 5255 } 5256 } 5257 5258 // If the load is shifted left (and the result isn't shifted back right), 5259 // we can fold the truncate through the shift. 5260 unsigned ShLeftAmt = 0; 5261 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() && 5262 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) { 5263 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 5264 ShLeftAmt = N01->getZExtValue(); 5265 N0 = N0.getOperand(0); 5266 } 5267 } 5268 5269 // If we haven't found a load, we can't narrow it. Don't transform one with 5270 // multiple uses, this would require adding a new load. 5271 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse()) 5272 return SDValue(); 5273 5274 // Don't change the width of a volatile load. 5275 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5276 if (LN0->isVolatile()) 5277 return SDValue(); 5278 5279 // Verify that we are actually reducing a load width here. 5280 if (LN0->getMemoryVT().getSizeInBits() < EVTBits) 5281 return SDValue(); 5282 5283 // For the transform to be legal, the load must produce only two values 5284 // (the value loaded and the chain). Don't transform a pre-increment 5285 // load, for example, which produces an extra value. Otherwise the 5286 // transformation is not equivalent, and the downstream logic to replace 5287 // uses gets things wrong. 5288 if (LN0->getNumValues() > 2) 5289 return SDValue(); 5290 5291 // If the load that we're shrinking is an extload and we're not just 5292 // discarding the extension we can't simply shrink the load. Bail. 5293 // TODO: It would be possible to merge the extensions in some cases. 5294 if (LN0->getExtensionType() != ISD::NON_EXTLOAD && 5295 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt) 5296 return SDValue(); 5297 5298 EVT PtrType = N0.getOperand(1).getValueType(); 5299 5300 if (PtrType == MVT::Untyped || PtrType.isExtended()) 5301 // It's not possible to generate a constant of extended or untyped type. 5302 return SDValue(); 5303 5304 // For big endian targets, we need to adjust the offset to the pointer to 5305 // load the correct bytes. 5306 if (TLI.isBigEndian()) { 5307 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits(); 5308 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits(); 5309 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt; 5310 } 5311 5312 uint64_t PtrOff = ShAmt / 8; 5313 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff); 5314 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), 5315 PtrType, LN0->getBasePtr(), 5316 DAG.getConstant(PtrOff, PtrType)); 5317 AddToWorkList(NewPtr.getNode()); 5318 5319 SDValue Load; 5320 if (ExtType == ISD::NON_EXTLOAD) 5321 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr, 5322 LN0->getPointerInfo().getWithOffset(PtrOff), 5323 LN0->isVolatile(), LN0->isNonTemporal(), 5324 LN0->isInvariant(), NewAlign); 5325 else 5326 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr, 5327 LN0->getPointerInfo().getWithOffset(PtrOff), 5328 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(), 5329 NewAlign); 5330 5331 // Replace the old load's chain with the new load's chain. 5332 WorkListRemover DeadNodes(*this); 5333 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1)); 5334 5335 // Shift the result left, if we've swallowed a left shift. 5336 SDValue Result = Load; 5337 if (ShLeftAmt != 0) { 5338 EVT ShImmTy = getShiftAmountTy(Result.getValueType()); 5339 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt)) 5340 ShImmTy = VT; 5341 // If the shift amount is as large as the result size (but, presumably, 5342 // no larger than the source) then the useful bits of the result are 5343 // zero; we can't simply return the shortened shift, because the result 5344 // of that operation is undefined. 5345 if (ShLeftAmt >= VT.getSizeInBits()) 5346 Result = DAG.getConstant(0, VT); 5347 else 5348 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT, 5349 Result, DAG.getConstant(ShLeftAmt, ShImmTy)); 5350 } 5351 5352 // Return the new loaded value. 5353 return Result; 5354} 5355 5356SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) { 5357 SDValue N0 = N->getOperand(0); 5358 SDValue N1 = N->getOperand(1); 5359 EVT VT = N->getValueType(0); 5360 EVT EVT = cast<VTSDNode>(N1)->getVT(); 5361 unsigned VTBits = VT.getScalarType().getSizeInBits(); 5362 unsigned EVTBits = EVT.getScalarType().getSizeInBits(); 5363 5364 // fold (sext_in_reg c1) -> c1 5365 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF) 5366 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1); 5367 5368 // If the input is already sign extended, just drop the extension. 5369 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1) 5370 return N0; 5371 5372 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2 5373 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 5374 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) 5375 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, 5376 N0.getOperand(0), N1); 5377 5378 // fold (sext_in_reg (sext x)) -> (sext x) 5379 // fold (sext_in_reg (aext x)) -> (sext x) 5380 // if x is small enough. 5381 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) { 5382 SDValue N00 = N0.getOperand(0); 5383 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits && 5384 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT))) 5385 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1); 5386 } 5387 5388 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero. 5389 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits))) 5390 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT); 5391 5392 // fold operands of sext_in_reg based on knowledge that the top bits are not 5393 // demanded. 5394 if (SimplifyDemandedBits(SDValue(N, 0))) 5395 return SDValue(N, 0); 5396 5397 // fold (sext_in_reg (load x)) -> (smaller sextload x) 5398 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits)) 5399 SDValue NarrowLoad = ReduceLoadWidth(N); 5400 if (NarrowLoad.getNode()) 5401 return NarrowLoad; 5402 5403 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24) 5404 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible. 5405 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above. 5406 if (N0.getOpcode() == ISD::SRL) { 5407 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 5408 if (ShAmt->getZExtValue()+EVTBits <= VTBits) { 5409 // We can turn this into an SRA iff the input to the SRL is already sign 5410 // extended enough. 5411 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0)); 5412 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits) 5413 return DAG.getNode(ISD::SRA, SDLoc(N), VT, 5414 N0.getOperand(0), N0.getOperand(1)); 5415 } 5416 } 5417 5418 // fold (sext_inreg (extload x)) -> (sextload x) 5419 if (ISD::isEXTLoad(N0.getNode()) && 5420 ISD::isUNINDEXEDLoad(N0.getNode()) && 5421 EVT == cast<LoadSDNode>(N0)->getMemoryVT() && 5422 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 5423 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) { 5424 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5425 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, 5426 LN0->getChain(), 5427 LN0->getBasePtr(), LN0->getPointerInfo(), 5428 EVT, 5429 LN0->isVolatile(), LN0->isNonTemporal(), 5430 LN0->getAlignment()); 5431 CombineTo(N, ExtLoad); 5432 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); 5433 AddToWorkList(ExtLoad.getNode()); 5434 return SDValue(N, 0); // Return N so it doesn't get rechecked! 5435 } 5436 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use 5437 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) && 5438 N0.hasOneUse() && 5439 EVT == cast<LoadSDNode>(N0)->getMemoryVT() && 5440 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 5441 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) { 5442 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5443 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT, 5444 LN0->getChain(), 5445 LN0->getBasePtr(), LN0->getPointerInfo(), 5446 EVT, 5447 LN0->isVolatile(), LN0->isNonTemporal(), 5448 LN0->getAlignment()); 5449 CombineTo(N, ExtLoad); 5450 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); 5451 return SDValue(N, 0); // Return N so it doesn't get rechecked! 5452 } 5453 5454 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16)) 5455 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) { 5456 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), 5457 N0.getOperand(1), false); 5458 if (BSwap.getNode() != 0) 5459 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, 5460 BSwap, N1); 5461 } 5462 5463 return SDValue(); 5464} 5465 5466SDValue DAGCombiner::visitTRUNCATE(SDNode *N) { 5467 SDValue N0 = N->getOperand(0); 5468 EVT VT = N->getValueType(0); 5469 bool isLE = TLI.isLittleEndian(); 5470 5471 // noop truncate 5472 if (N0.getValueType() == N->getValueType(0)) 5473 return N0; 5474 // fold (truncate c1) -> c1 5475 if (isa<ConstantSDNode>(N0)) 5476 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0); 5477 // fold (truncate (truncate x)) -> (truncate x) 5478 if (N0.getOpcode() == ISD::TRUNCATE) 5479 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0)); 5480 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x 5481 if (N0.getOpcode() == ISD::ZERO_EXTEND || 5482 N0.getOpcode() == ISD::SIGN_EXTEND || 5483 N0.getOpcode() == ISD::ANY_EXTEND) { 5484 if (N0.getOperand(0).getValueType().bitsLT(VT)) 5485 // if the source is smaller than the dest, we still need an extend 5486 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, 5487 N0.getOperand(0)); 5488 if (N0.getOperand(0).getValueType().bitsGT(VT)) 5489 // if the source is larger than the dest, than we just need the truncate 5490 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0)); 5491 // if the source and dest are the same type, we can drop both the extend 5492 // and the truncate. 5493 return N0.getOperand(0); 5494 } 5495 5496 // Fold extract-and-trunc into a narrow extract. For example: 5497 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1) 5498 // i32 y = TRUNCATE(i64 x) 5499 // -- becomes -- 5500 // v16i8 b = BITCAST (v2i64 val) 5501 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8) 5502 // 5503 // Note: We only run this optimization after type legalization (which often 5504 // creates this pattern) and before operation legalization after which 5505 // we need to be more careful about the vector instructions that we generate. 5506 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT && 5507 LegalTypes && !LegalOperations && N0->hasOneUse()) { 5508 5509 EVT VecTy = N0.getOperand(0).getValueType(); 5510 EVT ExTy = N0.getValueType(); 5511 EVT TrTy = N->getValueType(0); 5512 5513 unsigned NumElem = VecTy.getVectorNumElements(); 5514 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits(); 5515 5516 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem); 5517 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size"); 5518 5519 SDValue EltNo = N0->getOperand(1); 5520 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) { 5521 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); 5522 EVT IndexTy = TLI.getVectorIdxTy(); 5523 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1)); 5524 5525 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N), 5526 NVT, N0.getOperand(0)); 5527 5528 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, 5529 SDLoc(N), TrTy, V, 5530 DAG.getConstant(Index, IndexTy)); 5531 } 5532 } 5533 5534 // Fold a series of buildvector, bitcast, and truncate if possible. 5535 // For example fold 5536 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to 5537 // (2xi32 (buildvector x, y)). 5538 if (Level == AfterLegalizeVectorOps && VT.isVector() && 5539 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() && 5540 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR && 5541 N0.getOperand(0).hasOneUse()) { 5542 5543 SDValue BuildVect = N0.getOperand(0); 5544 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType(); 5545 EVT TruncVecEltTy = VT.getVectorElementType(); 5546 5547 // Check that the element types match. 5548 if (BuildVectEltTy == TruncVecEltTy) { 5549 // Now we only need to compute the offset of the truncated elements. 5550 unsigned BuildVecNumElts = BuildVect.getNumOperands(); 5551 unsigned TruncVecNumElts = VT.getVectorNumElements(); 5552 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts; 5553 5554 assert((BuildVecNumElts % TruncVecNumElts) == 0 && 5555 "Invalid number of elements"); 5556 5557 SmallVector<SDValue, 8> Opnds; 5558 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset) 5559 Opnds.push_back(BuildVect.getOperand(i)); 5560 5561 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Opnds[0], 5562 Opnds.size()); 5563 } 5564 } 5565 5566 // See if we can simplify the input to this truncate through knowledge that 5567 // only the low bits are being used. 5568 // For example "trunc (or (shl x, 8), y)" // -> trunc y 5569 // Currently we only perform this optimization on scalars because vectors 5570 // may have different active low bits. 5571 if (!VT.isVector()) { 5572 SDValue Shorter = 5573 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(), 5574 VT.getSizeInBits())); 5575 if (Shorter.getNode()) 5576 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter); 5577 } 5578 // fold (truncate (load x)) -> (smaller load x) 5579 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits)) 5580 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) { 5581 SDValue Reduced = ReduceLoadWidth(N); 5582 if (Reduced.getNode()) 5583 return Reduced; 5584 } 5585 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)), 5586 // where ... are all 'undef'. 5587 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) { 5588 SmallVector<EVT, 8> VTs; 5589 SDValue V; 5590 unsigned Idx = 0; 5591 unsigned NumDefs = 0; 5592 5593 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) { 5594 SDValue X = N0.getOperand(i); 5595 if (X.getOpcode() != ISD::UNDEF) { 5596 V = X; 5597 Idx = i; 5598 NumDefs++; 5599 } 5600 // Stop if more than one members are non-undef. 5601 if (NumDefs > 1) 5602 break; 5603 VTs.push_back(EVT::getVectorVT(*DAG.getContext(), 5604 VT.getVectorElementType(), 5605 X.getValueType().getVectorNumElements())); 5606 } 5607 5608 if (NumDefs == 0) 5609 return DAG.getUNDEF(VT); 5610 5611 if (NumDefs == 1) { 5612 assert(V.getNode() && "The single defined operand is empty!"); 5613 SmallVector<SDValue, 8> Opnds; 5614 for (unsigned i = 0, e = VTs.size(); i != e; ++i) { 5615 if (i != Idx) { 5616 Opnds.push_back(DAG.getUNDEF(VTs[i])); 5617 continue; 5618 } 5619 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V); 5620 AddToWorkList(NV.getNode()); 5621 Opnds.push_back(NV); 5622 } 5623 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, 5624 &Opnds[0], Opnds.size()); 5625 } 5626 } 5627 5628 // Simplify the operands using demanded-bits information. 5629 if (!VT.isVector() && 5630 SimplifyDemandedBits(SDValue(N, 0))) 5631 return SDValue(N, 0); 5632 5633 return SDValue(); 5634} 5635 5636static SDNode *getBuildPairElt(SDNode *N, unsigned i) { 5637 SDValue Elt = N->getOperand(i); 5638 if (Elt.getOpcode() != ISD::MERGE_VALUES) 5639 return Elt.getNode(); 5640 return Elt.getOperand(Elt.getResNo()).getNode(); 5641} 5642 5643/// CombineConsecutiveLoads - build_pair (load, load) -> load 5644/// if load locations are consecutive. 5645SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) { 5646 assert(N->getOpcode() == ISD::BUILD_PAIR); 5647 5648 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0)); 5649 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1)); 5650 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() || 5651 LD1->getPointerInfo().getAddrSpace() != 5652 LD2->getPointerInfo().getAddrSpace()) 5653 return SDValue(); 5654 EVT LD1VT = LD1->getValueType(0); 5655 5656 if (ISD::isNON_EXTLoad(LD2) && 5657 LD2->hasOneUse() && 5658 // If both are volatile this would reduce the number of volatile loads. 5659 // If one is volatile it might be ok, but play conservative and bail out. 5660 !LD1->isVolatile() && 5661 !LD2->isVolatile() && 5662 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) { 5663 unsigned Align = LD1->getAlignment(); 5664 unsigned NewAlign = TLI.getDataLayout()-> 5665 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext())); 5666 5667 if (NewAlign <= Align && 5668 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) 5669 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(), 5670 LD1->getBasePtr(), LD1->getPointerInfo(), 5671 false, false, false, Align); 5672 } 5673 5674 return SDValue(); 5675} 5676 5677SDValue DAGCombiner::visitBITCAST(SDNode *N) { 5678 SDValue N0 = N->getOperand(0); 5679 EVT VT = N->getValueType(0); 5680 5681 // If the input is a BUILD_VECTOR with all constant elements, fold this now. 5682 // Only do this before legalize, since afterward the target may be depending 5683 // on the bitconvert. 5684 // First check to see if this is all constant. 5685 if (!LegalTypes && 5686 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() && 5687 VT.isVector()) { 5688 bool isSimple = true; 5689 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) 5690 if (N0.getOperand(i).getOpcode() != ISD::UNDEF && 5691 N0.getOperand(i).getOpcode() != ISD::Constant && 5692 N0.getOperand(i).getOpcode() != ISD::ConstantFP) { 5693 isSimple = false; 5694 break; 5695 } 5696 5697 EVT DestEltVT = N->getValueType(0).getVectorElementType(); 5698 assert(!DestEltVT.isVector() && 5699 "Element type of vector ValueType must not be vector!"); 5700 if (isSimple) 5701 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT); 5702 } 5703 5704 // If the input is a constant, let getNode fold it. 5705 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) { 5706 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0); 5707 if (Res.getNode() != N) { 5708 if (!LegalOperations || 5709 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT)) 5710 return Res; 5711 5712 // Folding it resulted in an illegal node, and it's too late to 5713 // do that. Clean up the old node and forego the transformation. 5714 // Ideally this won't happen very often, because instcombine 5715 // and the earlier dagcombine runs (where illegal nodes are 5716 // permitted) should have folded most of them already. 5717 DAG.DeleteNode(Res.getNode()); 5718 } 5719 } 5720 5721 // (conv (conv x, t1), t2) -> (conv x, t2) 5722 if (N0.getOpcode() == ISD::BITCAST) 5723 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, 5724 N0.getOperand(0)); 5725 5726 // fold (conv (load x)) -> (load (conv*)x) 5727 // If the resultant load doesn't need a higher alignment than the original! 5728 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && 5729 // Do not change the width of a volatile load. 5730 !cast<LoadSDNode>(N0)->isVolatile() && 5731 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) { 5732 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 5733 unsigned Align = TLI.getDataLayout()-> 5734 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext())); 5735 unsigned OrigAlign = LN0->getAlignment(); 5736 5737 if (Align <= OrigAlign) { 5738 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), 5739 LN0->getBasePtr(), LN0->getPointerInfo(), 5740 LN0->isVolatile(), LN0->isNonTemporal(), 5741 LN0->isInvariant(), OrigAlign); 5742 AddToWorkList(N); 5743 CombineTo(N0.getNode(), 5744 DAG.getNode(ISD::BITCAST, SDLoc(N0), 5745 N0.getValueType(), Load), 5746 Load.getValue(1)); 5747 return Load; 5748 } 5749 } 5750 5751 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit) 5752 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit)) 5753 // This often reduces constant pool loads. 5754 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) || 5755 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) && 5756 N0.getNode()->hasOneUse() && VT.isInteger() && 5757 !VT.isVector() && !N0.getValueType().isVector()) { 5758 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT, 5759 N0.getOperand(0)); 5760 AddToWorkList(NewConv.getNode()); 5761 5762 APInt SignBit = APInt::getSignBit(VT.getSizeInBits()); 5763 if (N0.getOpcode() == ISD::FNEG) 5764 return DAG.getNode(ISD::XOR, SDLoc(N), VT, 5765 NewConv, DAG.getConstant(SignBit, VT)); 5766 assert(N0.getOpcode() == ISD::FABS); 5767 return DAG.getNode(ISD::AND, SDLoc(N), VT, 5768 NewConv, DAG.getConstant(~SignBit, VT)); 5769 } 5770 5771 // fold (bitconvert (fcopysign cst, x)) -> 5772 // (or (and (bitconvert x), sign), (and cst, (not sign))) 5773 // Note that we don't handle (copysign x, cst) because this can always be 5774 // folded to an fneg or fabs. 5775 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() && 5776 isa<ConstantFPSDNode>(N0.getOperand(0)) && 5777 VT.isInteger() && !VT.isVector()) { 5778 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits(); 5779 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth); 5780 if (isTypeLegal(IntXVT)) { 5781 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0), 5782 IntXVT, N0.getOperand(1)); 5783 AddToWorkList(X.getNode()); 5784 5785 // If X has a different width than the result/lhs, sext it or truncate it. 5786 unsigned VTWidth = VT.getSizeInBits(); 5787 if (OrigXWidth < VTWidth) { 5788 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X); 5789 AddToWorkList(X.getNode()); 5790 } else if (OrigXWidth > VTWidth) { 5791 // To get the sign bit in the right place, we have to shift it right 5792 // before truncating. 5793 X = DAG.getNode(ISD::SRL, SDLoc(X), 5794 X.getValueType(), X, 5795 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType())); 5796 AddToWorkList(X.getNode()); 5797 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X); 5798 AddToWorkList(X.getNode()); 5799 } 5800 5801 APInt SignBit = APInt::getSignBit(VT.getSizeInBits()); 5802 X = DAG.getNode(ISD::AND, SDLoc(X), VT, 5803 X, DAG.getConstant(SignBit, VT)); 5804 AddToWorkList(X.getNode()); 5805 5806 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0), 5807 VT, N0.getOperand(0)); 5808 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT, 5809 Cst, DAG.getConstant(~SignBit, VT)); 5810 AddToWorkList(Cst.getNode()); 5811 5812 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst); 5813 } 5814 } 5815 5816 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive. 5817 if (N0.getOpcode() == ISD::BUILD_PAIR) { 5818 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT); 5819 if (CombineLD.getNode()) 5820 return CombineLD; 5821 } 5822 5823 return SDValue(); 5824} 5825 5826SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) { 5827 EVT VT = N->getValueType(0); 5828 return CombineConsecutiveLoads(N, VT); 5829} 5830 5831/// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector 5832/// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the 5833/// destination element value type. 5834SDValue DAGCombiner:: 5835ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) { 5836 EVT SrcEltVT = BV->getValueType(0).getVectorElementType(); 5837 5838 // If this is already the right type, we're done. 5839 if (SrcEltVT == DstEltVT) return SDValue(BV, 0); 5840 5841 unsigned SrcBitSize = SrcEltVT.getSizeInBits(); 5842 unsigned DstBitSize = DstEltVT.getSizeInBits(); 5843 5844 // If this is a conversion of N elements of one type to N elements of another 5845 // type, convert each element. This handles FP<->INT cases. 5846 if (SrcBitSize == DstBitSize) { 5847 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, 5848 BV->getValueType(0).getVectorNumElements()); 5849 5850 // Due to the FP element handling below calling this routine recursively, 5851 // we can end up with a scalar-to-vector node here. 5852 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR) 5853 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT, 5854 DAG.getNode(ISD::BITCAST, SDLoc(BV), 5855 DstEltVT, BV->getOperand(0))); 5856 5857 SmallVector<SDValue, 8> Ops; 5858 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) { 5859 SDValue Op = BV->getOperand(i); 5860 // If the vector element type is not legal, the BUILD_VECTOR operands 5861 // are promoted and implicitly truncated. Make that explicit here. 5862 if (Op.getValueType() != SrcEltVT) 5863 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op); 5864 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV), 5865 DstEltVT, Op)); 5866 AddToWorkList(Ops.back().getNode()); 5867 } 5868 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, 5869 &Ops[0], Ops.size()); 5870 } 5871 5872 // Otherwise, we're growing or shrinking the elements. To avoid having to 5873 // handle annoying details of growing/shrinking FP values, we convert them to 5874 // int first. 5875 if (SrcEltVT.isFloatingPoint()) { 5876 // Convert the input float vector to a int vector where the elements are the 5877 // same sizes. 5878 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!"); 5879 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits()); 5880 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode(); 5881 SrcEltVT = IntVT; 5882 } 5883 5884 // Now we know the input is an integer vector. If the output is a FP type, 5885 // convert to integer first, then to FP of the right size. 5886 if (DstEltVT.isFloatingPoint()) { 5887 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!"); 5888 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits()); 5889 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode(); 5890 5891 // Next, convert to FP elements of the same size. 5892 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT); 5893 } 5894 5895 // Okay, we know the src/dst types are both integers of differing types. 5896 // Handling growing first. 5897 assert(SrcEltVT.isInteger() && DstEltVT.isInteger()); 5898 if (SrcBitSize < DstBitSize) { 5899 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize; 5900 5901 SmallVector<SDValue, 8> Ops; 5902 for (unsigned i = 0, e = BV->getNumOperands(); i != e; 5903 i += NumInputsPerOutput) { 5904 bool isLE = TLI.isLittleEndian(); 5905 APInt NewBits = APInt(DstBitSize, 0); 5906 bool EltIsUndef = true; 5907 for (unsigned j = 0; j != NumInputsPerOutput; ++j) { 5908 // Shift the previously computed bits over. 5909 NewBits <<= SrcBitSize; 5910 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j)); 5911 if (Op.getOpcode() == ISD::UNDEF) continue; 5912 EltIsUndef = false; 5913 5914 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue(). 5915 zextOrTrunc(SrcBitSize).zext(DstBitSize); 5916 } 5917 5918 if (EltIsUndef) 5919 Ops.push_back(DAG.getUNDEF(DstEltVT)); 5920 else 5921 Ops.push_back(DAG.getConstant(NewBits, DstEltVT)); 5922 } 5923 5924 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size()); 5925 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, 5926 &Ops[0], Ops.size()); 5927 } 5928 5929 // Finally, this must be the case where we are shrinking elements: each input 5930 // turns into multiple outputs. 5931 bool isS2V = ISD::isScalarToVector(BV); 5932 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize; 5933 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, 5934 NumOutputsPerInput*BV->getNumOperands()); 5935 SmallVector<SDValue, 8> Ops; 5936 5937 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) { 5938 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) { 5939 for (unsigned j = 0; j != NumOutputsPerInput; ++j) 5940 Ops.push_back(DAG.getUNDEF(DstEltVT)); 5941 continue; 5942 } 5943 5944 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))-> 5945 getAPIntValue().zextOrTrunc(SrcBitSize); 5946 5947 for (unsigned j = 0; j != NumOutputsPerInput; ++j) { 5948 APInt ThisVal = OpVal.trunc(DstBitSize); 5949 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT)); 5950 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal) 5951 // Simply turn this into a SCALAR_TO_VECTOR of the new type. 5952 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT, 5953 Ops[0]); 5954 OpVal = OpVal.lshr(DstBitSize); 5955 } 5956 5957 // For big endian targets, swap the order of the pieces of each element. 5958 if (TLI.isBigEndian()) 5959 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end()); 5960 } 5961 5962 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, 5963 &Ops[0], Ops.size()); 5964} 5965 5966SDValue DAGCombiner::visitFADD(SDNode *N) { 5967 SDValue N0 = N->getOperand(0); 5968 SDValue N1 = N->getOperand(1); 5969 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 5970 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 5971 EVT VT = N->getValueType(0); 5972 5973 // fold vector ops 5974 if (VT.isVector()) { 5975 SDValue FoldedVOp = SimplifyVBinOp(N); 5976 if (FoldedVOp.getNode()) return FoldedVOp; 5977 } 5978 5979 // fold (fadd c1, c2) -> c1 + c2 5980 if (N0CFP && N1CFP) 5981 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1); 5982 // canonicalize constant to RHS 5983 if (N0CFP && !N1CFP) 5984 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0); 5985 // fold (fadd A, 0) -> A 5986 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && 5987 N1CFP->getValueAPF().isZero()) 5988 return N0; 5989 // fold (fadd A, (fneg B)) -> (fsub A, B) 5990 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) && 5991 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2) 5992 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, 5993 GetNegatedExpression(N1, DAG, LegalOperations)); 5994 // fold (fadd (fneg A), B) -> (fsub B, A) 5995 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) && 5996 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2) 5997 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1, 5998 GetNegatedExpression(N0, DAG, LegalOperations)); 5999 6000 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2)) 6001 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && 6002 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() && 6003 isa<ConstantFPSDNode>(N0.getOperand(1))) 6004 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0), 6005 DAG.getNode(ISD::FADD, SDLoc(N), VT, 6006 N0.getOperand(1), N1)); 6007 6008 // No FP constant should be created after legalization as Instruction 6009 // Selection pass has hard time in dealing with FP constant. 6010 // 6011 // We don't need test this condition for transformation like following, as 6012 // the DAG being transformed implies it is legal to take FP constant as 6013 // operand. 6014 // 6015 // (fadd (fmul c, x), x) -> (fmul c+1, x) 6016 // 6017 bool AllowNewFpConst = (Level < AfterLegalizeDAG); 6018 6019 // If allow, fold (fadd (fneg x), x) -> 0.0 6020 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath && 6021 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1) 6022 return DAG.getConstantFP(0.0, VT); 6023 6024 // If allow, fold (fadd x, (fneg x)) -> 0.0 6025 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath && 6026 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0) 6027 return DAG.getConstantFP(0.0, VT); 6028 6029 // In unsafe math mode, we can fold chains of FADD's of the same value 6030 // into multiplications. This transform is not safe in general because 6031 // we are reducing the number of rounding steps. 6032 if (DAG.getTarget().Options.UnsafeFPMath && 6033 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && 6034 !N0CFP && !N1CFP) { 6035 if (N0.getOpcode() == ISD::FMUL) { 6036 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0)); 6037 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1)); 6038 6039 // (fadd (fmul c, x), x) -> (fmul x, c+1) 6040 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) { 6041 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6042 SDValue(CFP00, 0), 6043 DAG.getConstantFP(1.0, VT)); 6044 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6045 N1, NewCFP); 6046 } 6047 6048 // (fadd (fmul x, c), x) -> (fmul x, c+1) 6049 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) { 6050 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6051 SDValue(CFP01, 0), 6052 DAG.getConstantFP(1.0, VT)); 6053 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6054 N1, NewCFP); 6055 } 6056 6057 // (fadd (fmul c, x), (fadd x, x)) -> (fmul x, c+2) 6058 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD && 6059 N1.getOperand(0) == N1.getOperand(1) && 6060 N0.getOperand(1) == N1.getOperand(0)) { 6061 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6062 SDValue(CFP00, 0), 6063 DAG.getConstantFP(2.0, VT)); 6064 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6065 N0.getOperand(1), NewCFP); 6066 } 6067 6068 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2) 6069 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD && 6070 N1.getOperand(0) == N1.getOperand(1) && 6071 N0.getOperand(0) == N1.getOperand(0)) { 6072 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6073 SDValue(CFP01, 0), 6074 DAG.getConstantFP(2.0, VT)); 6075 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6076 N0.getOperand(0), NewCFP); 6077 } 6078 } 6079 6080 if (N1.getOpcode() == ISD::FMUL) { 6081 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0)); 6082 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1)); 6083 6084 // (fadd x, (fmul c, x)) -> (fmul x, c+1) 6085 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) { 6086 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6087 SDValue(CFP10, 0), 6088 DAG.getConstantFP(1.0, VT)); 6089 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6090 N0, NewCFP); 6091 } 6092 6093 // (fadd x, (fmul x, c)) -> (fmul x, c+1) 6094 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) { 6095 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6096 SDValue(CFP11, 0), 6097 DAG.getConstantFP(1.0, VT)); 6098 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6099 N0, NewCFP); 6100 } 6101 6102 6103 // (fadd (fadd x, x), (fmul c, x)) -> (fmul x, c+2) 6104 if (CFP10 && !CFP11 && N0.getOpcode() == ISD::FADD && 6105 N0.getOperand(0) == N0.getOperand(1) && 6106 N1.getOperand(1) == N0.getOperand(0)) { 6107 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6108 SDValue(CFP10, 0), 6109 DAG.getConstantFP(2.0, VT)); 6110 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6111 N1.getOperand(1), NewCFP); 6112 } 6113 6114 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2) 6115 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD && 6116 N0.getOperand(0) == N0.getOperand(1) && 6117 N1.getOperand(0) == N0.getOperand(0)) { 6118 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT, 6119 SDValue(CFP11, 0), 6120 DAG.getConstantFP(2.0, VT)); 6121 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6122 N1.getOperand(0), NewCFP); 6123 } 6124 } 6125 6126 if (N0.getOpcode() == ISD::FADD && AllowNewFpConst) { 6127 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0)); 6128 // (fadd (fadd x, x), x) -> (fmul x, 3.0) 6129 if (!CFP && N0.getOperand(0) == N0.getOperand(1) && 6130 (N0.getOperand(0) == N1)) 6131 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6132 N1, DAG.getConstantFP(3.0, VT)); 6133 } 6134 6135 if (N1.getOpcode() == ISD::FADD && AllowNewFpConst) { 6136 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0)); 6137 // (fadd x, (fadd x, x)) -> (fmul x, 3.0) 6138 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) && 6139 N1.getOperand(0) == N0) 6140 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6141 N0, DAG.getConstantFP(3.0, VT)); 6142 } 6143 6144 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0) 6145 if (AllowNewFpConst && 6146 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD && 6147 N0.getOperand(0) == N0.getOperand(1) && 6148 N1.getOperand(0) == N1.getOperand(1) && 6149 N0.getOperand(0) == N1.getOperand(0)) 6150 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6151 N0.getOperand(0), 6152 DAG.getConstantFP(4.0, VT)); 6153 } 6154 6155 // FADD -> FMA combines: 6156 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast || 6157 DAG.getTarget().Options.UnsafeFPMath) && 6158 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) && 6159 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) { 6160 6161 // fold (fadd (fmul x, y), z) -> (fma x, y, z) 6162 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) 6163 return DAG.getNode(ISD::FMA, SDLoc(N), VT, 6164 N0.getOperand(0), N0.getOperand(1), N1); 6165 6166 // fold (fadd x, (fmul y, z)) -> (fma y, z, x) 6167 // Note: Commutes FADD operands. 6168 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) 6169 return DAG.getNode(ISD::FMA, SDLoc(N), VT, 6170 N1.getOperand(0), N1.getOperand(1), N0); 6171 } 6172 6173 return SDValue(); 6174} 6175 6176SDValue DAGCombiner::visitFSUB(SDNode *N) { 6177 SDValue N0 = N->getOperand(0); 6178 SDValue N1 = N->getOperand(1); 6179 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6180 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6181 EVT VT = N->getValueType(0); 6182 SDLoc dl(N); 6183 6184 // fold vector ops 6185 if (VT.isVector()) { 6186 SDValue FoldedVOp = SimplifyVBinOp(N); 6187 if (FoldedVOp.getNode()) return FoldedVOp; 6188 } 6189 6190 // fold (fsub c1, c2) -> c1-c2 6191 if (N0CFP && N1CFP) 6192 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1); 6193 // fold (fsub A, 0) -> A 6194 if (DAG.getTarget().Options.UnsafeFPMath && 6195 N1CFP && N1CFP->getValueAPF().isZero()) 6196 return N0; 6197 // fold (fsub 0, B) -> -B 6198 if (DAG.getTarget().Options.UnsafeFPMath && 6199 N0CFP && N0CFP->getValueAPF().isZero()) { 6200 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options)) 6201 return GetNegatedExpression(N1, DAG, LegalOperations); 6202 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) 6203 return DAG.getNode(ISD::FNEG, dl, VT, N1); 6204 } 6205 // fold (fsub A, (fneg B)) -> (fadd A, B) 6206 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options)) 6207 return DAG.getNode(ISD::FADD, dl, VT, N0, 6208 GetNegatedExpression(N1, DAG, LegalOperations)); 6209 6210 // If 'unsafe math' is enabled, fold 6211 // (fsub x, x) -> 0.0 & 6212 // (fsub x, (fadd x, y)) -> (fneg y) & 6213 // (fsub x, (fadd y, x)) -> (fneg y) 6214 if (DAG.getTarget().Options.UnsafeFPMath) { 6215 if (N0 == N1) 6216 return DAG.getConstantFP(0.0f, VT); 6217 6218 if (N1.getOpcode() == ISD::FADD) { 6219 SDValue N10 = N1->getOperand(0); 6220 SDValue N11 = N1->getOperand(1); 6221 6222 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, 6223 &DAG.getTarget().Options)) 6224 return GetNegatedExpression(N11, DAG, LegalOperations); 6225 6226 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, 6227 &DAG.getTarget().Options)) 6228 return GetNegatedExpression(N10, DAG, LegalOperations); 6229 } 6230 } 6231 6232 // FSUB -> FMA combines: 6233 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast || 6234 DAG.getTarget().Options.UnsafeFPMath) && 6235 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) && 6236 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) { 6237 6238 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z)) 6239 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) 6240 return DAG.getNode(ISD::FMA, dl, VT, 6241 N0.getOperand(0), N0.getOperand(1), 6242 DAG.getNode(ISD::FNEG, dl, VT, N1)); 6243 6244 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x) 6245 // Note: Commutes FSUB operands. 6246 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) 6247 return DAG.getNode(ISD::FMA, dl, VT, 6248 DAG.getNode(ISD::FNEG, dl, VT, 6249 N1.getOperand(0)), 6250 N1.getOperand(1), N0); 6251 6252 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z)) 6253 if (N0.getOpcode() == ISD::FNEG && 6254 N0.getOperand(0).getOpcode() == ISD::FMUL && 6255 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) { 6256 SDValue N00 = N0.getOperand(0).getOperand(0); 6257 SDValue N01 = N0.getOperand(0).getOperand(1); 6258 return DAG.getNode(ISD::FMA, dl, VT, 6259 DAG.getNode(ISD::FNEG, dl, VT, N00), N01, 6260 DAG.getNode(ISD::FNEG, dl, VT, N1)); 6261 } 6262 } 6263 6264 return SDValue(); 6265} 6266 6267SDValue DAGCombiner::visitFMUL(SDNode *N) { 6268 SDValue N0 = N->getOperand(0); 6269 SDValue N1 = N->getOperand(1); 6270 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6271 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6272 EVT VT = N->getValueType(0); 6273 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 6274 6275 // fold vector ops 6276 if (VT.isVector()) { 6277 SDValue FoldedVOp = SimplifyVBinOp(N); 6278 if (FoldedVOp.getNode()) return FoldedVOp; 6279 } 6280 6281 // fold (fmul c1, c2) -> c1*c2 6282 if (N0CFP && N1CFP) 6283 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1); 6284 // canonicalize constant to RHS 6285 if (N0CFP && !N1CFP) 6286 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0); 6287 // fold (fmul A, 0) -> 0 6288 if (DAG.getTarget().Options.UnsafeFPMath && 6289 N1CFP && N1CFP->getValueAPF().isZero()) 6290 return N1; 6291 // fold (fmul A, 0) -> 0, vector edition. 6292 if (DAG.getTarget().Options.UnsafeFPMath && 6293 ISD::isBuildVectorAllZeros(N1.getNode())) 6294 return N1; 6295 // fold (fmul A, 1.0) -> A 6296 if (N1CFP && N1CFP->isExactlyValue(1.0)) 6297 return N0; 6298 // fold (fmul X, 2.0) -> (fadd X, X) 6299 if (N1CFP && N1CFP->isExactlyValue(+2.0)) 6300 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0); 6301 // fold (fmul X, -1.0) -> (fneg X) 6302 if (N1CFP && N1CFP->isExactlyValue(-1.0)) 6303 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) 6304 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0); 6305 6306 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y) 6307 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, 6308 &DAG.getTarget().Options)) { 6309 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, 6310 &DAG.getTarget().Options)) { 6311 // Both can be negated for free, check to see if at least one is cheaper 6312 // negated. 6313 if (LHSNeg == 2 || RHSNeg == 2) 6314 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6315 GetNegatedExpression(N0, DAG, LegalOperations), 6316 GetNegatedExpression(N1, DAG, LegalOperations)); 6317 } 6318 } 6319 6320 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2)) 6321 if (DAG.getTarget().Options.UnsafeFPMath && 6322 N1CFP && N0.getOpcode() == ISD::FMUL && 6323 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1))) 6324 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), 6325 DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6326 N0.getOperand(1), N1)); 6327 6328 return SDValue(); 6329} 6330 6331SDValue DAGCombiner::visitFMA(SDNode *N) { 6332 SDValue N0 = N->getOperand(0); 6333 SDValue N1 = N->getOperand(1); 6334 SDValue N2 = N->getOperand(2); 6335 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6336 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6337 EVT VT = N->getValueType(0); 6338 SDLoc dl(N); 6339 6340 if (DAG.getTarget().Options.UnsafeFPMath) { 6341 if (N0CFP && N0CFP->isZero()) 6342 return N2; 6343 if (N1CFP && N1CFP->isZero()) 6344 return N2; 6345 } 6346 if (N0CFP && N0CFP->isExactlyValue(1.0)) 6347 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2); 6348 if (N1CFP && N1CFP->isExactlyValue(1.0)) 6349 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2); 6350 6351 // Canonicalize (fma c, x, y) -> (fma x, c, y) 6352 if (N0CFP && !N1CFP) 6353 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2); 6354 6355 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2) 6356 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && 6357 N2.getOpcode() == ISD::FMUL && 6358 N0 == N2.getOperand(0) && 6359 N2.getOperand(1).getOpcode() == ISD::ConstantFP) { 6360 return DAG.getNode(ISD::FMUL, dl, VT, N0, 6361 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1))); 6362 } 6363 6364 6365 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y) 6366 if (DAG.getTarget().Options.UnsafeFPMath && 6367 N0.getOpcode() == ISD::FMUL && N1CFP && 6368 N0.getOperand(1).getOpcode() == ISD::ConstantFP) { 6369 return DAG.getNode(ISD::FMA, dl, VT, 6370 N0.getOperand(0), 6371 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)), 6372 N2); 6373 } 6374 6375 // (fma x, 1, y) -> (fadd x, y) 6376 // (fma x, -1, y) -> (fadd (fneg x), y) 6377 if (N1CFP) { 6378 if (N1CFP->isExactlyValue(1.0)) 6379 return DAG.getNode(ISD::FADD, dl, VT, N0, N2); 6380 6381 if (N1CFP->isExactlyValue(-1.0) && 6382 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) { 6383 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0); 6384 AddToWorkList(RHSNeg.getNode()); 6385 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg); 6386 } 6387 } 6388 6389 // (fma x, c, x) -> (fmul x, (c+1)) 6390 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2) 6391 return DAG.getNode(ISD::FMUL, dl, VT, N0, 6392 DAG.getNode(ISD::FADD, dl, VT, 6393 N1, DAG.getConstantFP(1.0, VT))); 6394 6395 // (fma x, c, (fneg x)) -> (fmul x, (c-1)) 6396 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && 6397 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0) 6398 return DAG.getNode(ISD::FMUL, dl, VT, N0, 6399 DAG.getNode(ISD::FADD, dl, VT, 6400 N1, DAG.getConstantFP(-1.0, VT))); 6401 6402 6403 return SDValue(); 6404} 6405 6406SDValue DAGCombiner::visitFDIV(SDNode *N) { 6407 SDValue N0 = N->getOperand(0); 6408 SDValue N1 = N->getOperand(1); 6409 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6410 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6411 EVT VT = N->getValueType(0); 6412 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 6413 6414 // fold vector ops 6415 if (VT.isVector()) { 6416 SDValue FoldedVOp = SimplifyVBinOp(N); 6417 if (FoldedVOp.getNode()) return FoldedVOp; 6418 } 6419 6420 // fold (fdiv c1, c2) -> c1/c2 6421 if (N0CFP && N1CFP) 6422 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1); 6423 6424 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable. 6425 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) { 6426 // Compute the reciprocal 1.0 / c2. 6427 APFloat N1APF = N1CFP->getValueAPF(); 6428 APFloat Recip(N1APF.getSemantics(), 1); // 1.0 6429 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven); 6430 // Only do the transform if the reciprocal is a legal fp immediate that 6431 // isn't too nasty (eg NaN, denormal, ...). 6432 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty 6433 (!LegalOperations || 6434 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM 6435 // backend)... we should handle this gracefully after Legalize. 6436 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) || 6437 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) || 6438 TLI.isFPImmLegal(Recip, VT))) 6439 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, 6440 DAG.getConstantFP(Recip, VT)); 6441 } 6442 6443 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y) 6444 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, 6445 &DAG.getTarget().Options)) { 6446 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, 6447 &DAG.getTarget().Options)) { 6448 // Both can be negated for free, check to see if at least one is cheaper 6449 // negated. 6450 if (LHSNeg == 2 || RHSNeg == 2) 6451 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, 6452 GetNegatedExpression(N0, DAG, LegalOperations), 6453 GetNegatedExpression(N1, DAG, LegalOperations)); 6454 } 6455 } 6456 6457 return SDValue(); 6458} 6459 6460SDValue DAGCombiner::visitFREM(SDNode *N) { 6461 SDValue N0 = N->getOperand(0); 6462 SDValue N1 = N->getOperand(1); 6463 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6464 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6465 EVT VT = N->getValueType(0); 6466 6467 // fold (frem c1, c2) -> fmod(c1,c2) 6468 if (N0CFP && N1CFP) 6469 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1); 6470 6471 return SDValue(); 6472} 6473 6474SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) { 6475 SDValue N0 = N->getOperand(0); 6476 SDValue N1 = N->getOperand(1); 6477 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6478 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 6479 EVT VT = N->getValueType(0); 6480 6481 if (N0CFP && N1CFP) // Constant fold 6482 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1); 6483 6484 if (N1CFP) { 6485 const APFloat& V = N1CFP->getValueAPF(); 6486 // copysign(x, c1) -> fabs(x) iff ispos(c1) 6487 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1) 6488 if (!V.isNegative()) { 6489 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT)) 6490 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); 6491 } else { 6492 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT)) 6493 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, 6494 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0)); 6495 } 6496 } 6497 6498 // copysign(fabs(x), y) -> copysign(x, y) 6499 // copysign(fneg(x), y) -> copysign(x, y) 6500 // copysign(copysign(x,z), y) -> copysign(x, y) 6501 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG || 6502 N0.getOpcode() == ISD::FCOPYSIGN) 6503 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, 6504 N0.getOperand(0), N1); 6505 6506 // copysign(x, abs(y)) -> abs(x) 6507 if (N1.getOpcode() == ISD::FABS) 6508 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); 6509 6510 // copysign(x, copysign(y,z)) -> copysign(x, z) 6511 if (N1.getOpcode() == ISD::FCOPYSIGN) 6512 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, 6513 N0, N1.getOperand(1)); 6514 6515 // copysign(x, fp_extend(y)) -> copysign(x, y) 6516 // copysign(x, fp_round(y)) -> copysign(x, y) 6517 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND) 6518 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, 6519 N0, N1.getOperand(0)); 6520 6521 return SDValue(); 6522} 6523 6524SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) { 6525 SDValue N0 = N->getOperand(0); 6526 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 6527 EVT VT = N->getValueType(0); 6528 EVT OpVT = N0.getValueType(); 6529 6530 // fold (sint_to_fp c1) -> c1fp 6531 if (N0C && 6532 // ...but only if the target supports immediate floating-point values 6533 (!LegalOperations || 6534 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) 6535 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0); 6536 6537 // If the input is a legal type, and SINT_TO_FP is not legal on this target, 6538 // but UINT_TO_FP is legal on this target, try to convert. 6539 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) && 6540 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) { 6541 // If the sign bit is known to be zero, we can change this to UINT_TO_FP. 6542 if (DAG.SignBitIsZero(N0)) 6543 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0); 6544 } 6545 6546 // The next optimizations are desireable only if SELECT_CC can be lowered. 6547 // Check against MVT::Other for SELECT_CC, which is a workaround for targets 6548 // having to say they don't support SELECT_CC on every type the DAG knows 6549 // about, since there is no way to mark an opcode illegal at all value types 6550 // (See also visitSELECT) 6551 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) { 6552 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc) 6553 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 && 6554 !VT.isVector() && 6555 (!LegalOperations || 6556 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) { 6557 SDValue Ops[] = 6558 { N0.getOperand(0), N0.getOperand(1), 6559 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT), 6560 N0.getOperand(2) }; 6561 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5); 6562 } 6563 6564 // fold (sint_to_fp (zext (setcc x, y, cc))) -> 6565 // (select_cc x, y, 1.0, 0.0,, cc) 6566 if (N0.getOpcode() == ISD::ZERO_EXTEND && 6567 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() && 6568 (!LegalOperations || 6569 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) { 6570 SDValue Ops[] = 6571 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1), 6572 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT), 6573 N0.getOperand(0).getOperand(2) }; 6574 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5); 6575 } 6576 } 6577 6578 return SDValue(); 6579} 6580 6581SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) { 6582 SDValue N0 = N->getOperand(0); 6583 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 6584 EVT VT = N->getValueType(0); 6585 EVT OpVT = N0.getValueType(); 6586 6587 // fold (uint_to_fp c1) -> c1fp 6588 if (N0C && 6589 // ...but only if the target supports immediate floating-point values 6590 (!LegalOperations || 6591 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) 6592 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0); 6593 6594 // If the input is a legal type, and UINT_TO_FP is not legal on this target, 6595 // but SINT_TO_FP is legal on this target, try to convert. 6596 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) && 6597 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) { 6598 // If the sign bit is known to be zero, we can change this to SINT_TO_FP. 6599 if (DAG.SignBitIsZero(N0)) 6600 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0); 6601 } 6602 6603 // The next optimizations are desireable only if SELECT_CC can be lowered. 6604 // Check against MVT::Other for SELECT_CC, which is a workaround for targets 6605 // having to say they don't support SELECT_CC on every type the DAG knows 6606 // about, since there is no way to mark an opcode illegal at all value types 6607 // (See also visitSELECT) 6608 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) { 6609 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc) 6610 6611 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() && 6612 (!LegalOperations || 6613 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) { 6614 SDValue Ops[] = 6615 { N0.getOperand(0), N0.getOperand(1), 6616 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT), 6617 N0.getOperand(2) }; 6618 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5); 6619 } 6620 } 6621 6622 return SDValue(); 6623} 6624 6625SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) { 6626 SDValue N0 = N->getOperand(0); 6627 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6628 EVT VT = N->getValueType(0); 6629 6630 // fold (fp_to_sint c1fp) -> c1 6631 if (N0CFP) 6632 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0); 6633 6634 return SDValue(); 6635} 6636 6637SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) { 6638 SDValue N0 = N->getOperand(0); 6639 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6640 EVT VT = N->getValueType(0); 6641 6642 // fold (fp_to_uint c1fp) -> c1 6643 if (N0CFP) 6644 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0); 6645 6646 return SDValue(); 6647} 6648 6649SDValue DAGCombiner::visitFP_ROUND(SDNode *N) { 6650 SDValue N0 = N->getOperand(0); 6651 SDValue N1 = N->getOperand(1); 6652 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6653 EVT VT = N->getValueType(0); 6654 6655 // fold (fp_round c1fp) -> c1fp 6656 if (N0CFP) 6657 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1); 6658 6659 // fold (fp_round (fp_extend x)) -> x 6660 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType()) 6661 return N0.getOperand(0); 6662 6663 // fold (fp_round (fp_round x)) -> (fp_round x) 6664 if (N0.getOpcode() == ISD::FP_ROUND) { 6665 // This is a value preserving truncation if both round's are. 6666 bool IsTrunc = N->getConstantOperandVal(1) == 1 && 6667 N0.getNode()->getConstantOperandVal(1) == 1; 6668 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0), 6669 DAG.getIntPtrConstant(IsTrunc)); 6670 } 6671 6672 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y) 6673 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) { 6674 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT, 6675 N0.getOperand(0), N1); 6676 AddToWorkList(Tmp.getNode()); 6677 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, 6678 Tmp, N0.getOperand(1)); 6679 } 6680 6681 return SDValue(); 6682} 6683 6684SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) { 6685 SDValue N0 = N->getOperand(0); 6686 EVT VT = N->getValueType(0); 6687 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT(); 6688 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6689 6690 // fold (fp_round_inreg c1fp) -> c1fp 6691 if (N0CFP && isTypeLegal(EVT)) { 6692 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT); 6693 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round); 6694 } 6695 6696 return SDValue(); 6697} 6698 6699SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) { 6700 SDValue N0 = N->getOperand(0); 6701 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6702 EVT VT = N->getValueType(0); 6703 6704 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded. 6705 if (N->hasOneUse() && 6706 N->use_begin()->getOpcode() == ISD::FP_ROUND) 6707 return SDValue(); 6708 6709 // fold (fp_extend c1fp) -> c1fp 6710 if (N0CFP) 6711 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0); 6712 6713 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the 6714 // value of X. 6715 if (N0.getOpcode() == ISD::FP_ROUND 6716 && N0.getNode()->getConstantOperandVal(1) == 1) { 6717 SDValue In = N0.getOperand(0); 6718 if (In.getValueType() == VT) return In; 6719 if (VT.bitsLT(In.getValueType())) 6720 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, 6721 In, N0.getOperand(1)); 6722 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In); 6723 } 6724 6725 // fold (fpext (load x)) -> (fpext (fptrunc (extload x))) 6726 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() && 6727 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) || 6728 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) { 6729 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 6730 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT, 6731 LN0->getChain(), 6732 LN0->getBasePtr(), LN0->getPointerInfo(), 6733 N0.getValueType(), 6734 LN0->isVolatile(), LN0->isNonTemporal(), 6735 LN0->getAlignment()); 6736 CombineTo(N, ExtLoad); 6737 CombineTo(N0.getNode(), 6738 DAG.getNode(ISD::FP_ROUND, SDLoc(N0), 6739 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)), 6740 ExtLoad.getValue(1)); 6741 return SDValue(N, 0); // Return N so it doesn't get rechecked! 6742 } 6743 6744 return SDValue(); 6745} 6746 6747SDValue DAGCombiner::visitFNEG(SDNode *N) { 6748 SDValue N0 = N->getOperand(0); 6749 EVT VT = N->getValueType(0); 6750 6751 if (VT.isVector()) { 6752 SDValue FoldedVOp = SimplifyVUnaryOp(N); 6753 if (FoldedVOp.getNode()) return FoldedVOp; 6754 } 6755 6756 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(), 6757 &DAG.getTarget().Options)) 6758 return GetNegatedExpression(N0, DAG, LegalOperations); 6759 6760 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading 6761 // constant pool values. 6762 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST && 6763 !VT.isVector() && 6764 N0.getNode()->hasOneUse() && 6765 N0.getOperand(0).getValueType().isInteger()) { 6766 SDValue Int = N0.getOperand(0); 6767 EVT IntVT = Int.getValueType(); 6768 if (IntVT.isInteger() && !IntVT.isVector()) { 6769 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int, 6770 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT)); 6771 AddToWorkList(Int.getNode()); 6772 return DAG.getNode(ISD::BITCAST, SDLoc(N), 6773 VT, Int); 6774 } 6775 } 6776 6777 // (fneg (fmul c, x)) -> (fmul -c, x) 6778 if (N0.getOpcode() == ISD::FMUL) { 6779 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1)); 6780 if (CFP1) 6781 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, 6782 N0.getOperand(0), 6783 DAG.getNode(ISD::FNEG, SDLoc(N), VT, 6784 N0.getOperand(1))); 6785 } 6786 6787 return SDValue(); 6788} 6789 6790SDValue DAGCombiner::visitFCEIL(SDNode *N) { 6791 SDValue N0 = N->getOperand(0); 6792 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6793 EVT VT = N->getValueType(0); 6794 6795 // fold (fceil c1) -> fceil(c1) 6796 if (N0CFP) 6797 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0); 6798 6799 return SDValue(); 6800} 6801 6802SDValue DAGCombiner::visitFTRUNC(SDNode *N) { 6803 SDValue N0 = N->getOperand(0); 6804 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6805 EVT VT = N->getValueType(0); 6806 6807 // fold (ftrunc c1) -> ftrunc(c1) 6808 if (N0CFP) 6809 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0); 6810 6811 return SDValue(); 6812} 6813 6814SDValue DAGCombiner::visitFFLOOR(SDNode *N) { 6815 SDValue N0 = N->getOperand(0); 6816 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6817 EVT VT = N->getValueType(0); 6818 6819 // fold (ffloor c1) -> ffloor(c1) 6820 if (N0CFP) 6821 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0); 6822 6823 return SDValue(); 6824} 6825 6826SDValue DAGCombiner::visitFABS(SDNode *N) { 6827 SDValue N0 = N->getOperand(0); 6828 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 6829 EVT VT = N->getValueType(0); 6830 6831 if (VT.isVector()) { 6832 SDValue FoldedVOp = SimplifyVUnaryOp(N); 6833 if (FoldedVOp.getNode()) return FoldedVOp; 6834 } 6835 6836 // fold (fabs c1) -> fabs(c1) 6837 if (N0CFP) 6838 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0); 6839 // fold (fabs (fabs x)) -> (fabs x) 6840 if (N0.getOpcode() == ISD::FABS) 6841 return N->getOperand(0); 6842 // fold (fabs (fneg x)) -> (fabs x) 6843 // fold (fabs (fcopysign x, y)) -> (fabs x) 6844 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN) 6845 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0)); 6846 6847 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading 6848 // constant pool values. 6849 if (!TLI.isFAbsFree(VT) && 6850 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() && 6851 N0.getOperand(0).getValueType().isInteger() && 6852 !N0.getOperand(0).getValueType().isVector()) { 6853 SDValue Int = N0.getOperand(0); 6854 EVT IntVT = Int.getValueType(); 6855 if (IntVT.isInteger() && !IntVT.isVector()) { 6856 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int, 6857 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT)); 6858 AddToWorkList(Int.getNode()); 6859 return DAG.getNode(ISD::BITCAST, SDLoc(N), 6860 N->getValueType(0), Int); 6861 } 6862 } 6863 6864 return SDValue(); 6865} 6866 6867SDValue DAGCombiner::visitBRCOND(SDNode *N) { 6868 SDValue Chain = N->getOperand(0); 6869 SDValue N1 = N->getOperand(1); 6870 SDValue N2 = N->getOperand(2); 6871 6872 // If N is a constant we could fold this into a fallthrough or unconditional 6873 // branch. However that doesn't happen very often in normal code, because 6874 // Instcombine/SimplifyCFG should have handled the available opportunities. 6875 // If we did this folding here, it would be necessary to update the 6876 // MachineBasicBlock CFG, which is awkward. 6877 6878 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal 6879 // on the target. 6880 if (N1.getOpcode() == ISD::SETCC && 6881 TLI.isOperationLegalOrCustom(ISD::BR_CC, 6882 N1.getOperand(0).getValueType())) { 6883 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other, 6884 Chain, N1.getOperand(2), 6885 N1.getOperand(0), N1.getOperand(1), N2); 6886 } 6887 6888 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) || 6889 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) && 6890 (N1.getOperand(0).hasOneUse() && 6891 N1.getOperand(0).getOpcode() == ISD::SRL))) { 6892 SDNode *Trunc = 0; 6893 if (N1.getOpcode() == ISD::TRUNCATE) { 6894 // Look pass the truncate. 6895 Trunc = N1.getNode(); 6896 N1 = N1.getOperand(0); 6897 } 6898 6899 // Match this pattern so that we can generate simpler code: 6900 // 6901 // %a = ... 6902 // %b = and i32 %a, 2 6903 // %c = srl i32 %b, 1 6904 // brcond i32 %c ... 6905 // 6906 // into 6907 // 6908 // %a = ... 6909 // %b = and i32 %a, 2 6910 // %c = setcc eq %b, 0 6911 // brcond %c ... 6912 // 6913 // This applies only when the AND constant value has one bit set and the 6914 // SRL constant is equal to the log2 of the AND constant. The back-end is 6915 // smart enough to convert the result into a TEST/JMP sequence. 6916 SDValue Op0 = N1.getOperand(0); 6917 SDValue Op1 = N1.getOperand(1); 6918 6919 if (Op0.getOpcode() == ISD::AND && 6920 Op1.getOpcode() == ISD::Constant) { 6921 SDValue AndOp1 = Op0.getOperand(1); 6922 6923 if (AndOp1.getOpcode() == ISD::Constant) { 6924 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue(); 6925 6926 if (AndConst.isPowerOf2() && 6927 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) { 6928 SDValue SetCC = 6929 DAG.getSetCC(SDLoc(N), 6930 getSetCCResultType(Op0.getValueType()), 6931 Op0, DAG.getConstant(0, Op0.getValueType()), 6932 ISD::SETNE); 6933 6934 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N), 6935 MVT::Other, Chain, SetCC, N2); 6936 // Don't add the new BRCond into the worklist or else SimplifySelectCC 6937 // will convert it back to (X & C1) >> C2. 6938 CombineTo(N, NewBRCond, false); 6939 // Truncate is dead. 6940 if (Trunc) { 6941 removeFromWorkList(Trunc); 6942 DAG.DeleteNode(Trunc); 6943 } 6944 // Replace the uses of SRL with SETCC 6945 WorkListRemover DeadNodes(*this); 6946 DAG.ReplaceAllUsesOfValueWith(N1, SetCC); 6947 removeFromWorkList(N1.getNode()); 6948 DAG.DeleteNode(N1.getNode()); 6949 return SDValue(N, 0); // Return N so it doesn't get rechecked! 6950 } 6951 } 6952 } 6953 6954 if (Trunc) 6955 // Restore N1 if the above transformation doesn't match. 6956 N1 = N->getOperand(1); 6957 } 6958 6959 // Transform br(xor(x, y)) -> br(x != y) 6960 // Transform br(xor(xor(x,y), 1)) -> br (x == y) 6961 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) { 6962 SDNode *TheXor = N1.getNode(); 6963 SDValue Op0 = TheXor->getOperand(0); 6964 SDValue Op1 = TheXor->getOperand(1); 6965 if (Op0.getOpcode() == Op1.getOpcode()) { 6966 // Avoid missing important xor optimizations. 6967 SDValue Tmp = visitXOR(TheXor); 6968 if (Tmp.getNode()) { 6969 if (Tmp.getNode() != TheXor) { 6970 DEBUG(dbgs() << "\nReplacing.8 "; 6971 TheXor->dump(&DAG); 6972 dbgs() << "\nWith: "; 6973 Tmp.getNode()->dump(&DAG); 6974 dbgs() << '\n'); 6975 WorkListRemover DeadNodes(*this); 6976 DAG.ReplaceAllUsesOfValueWith(N1, Tmp); 6977 removeFromWorkList(TheXor); 6978 DAG.DeleteNode(TheXor); 6979 return DAG.getNode(ISD::BRCOND, SDLoc(N), 6980 MVT::Other, Chain, Tmp, N2); 6981 } 6982 6983 // visitXOR has changed XOR's operands or replaced the XOR completely, 6984 // bail out. 6985 return SDValue(N, 0); 6986 } 6987 } 6988 6989 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) { 6990 bool Equal = false; 6991 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0)) 6992 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() && 6993 Op0.getOpcode() == ISD::XOR) { 6994 TheXor = Op0.getNode(); 6995 Equal = true; 6996 } 6997 6998 EVT SetCCVT = N1.getValueType(); 6999 if (LegalTypes) 7000 SetCCVT = getSetCCResultType(SetCCVT); 7001 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor), 7002 SetCCVT, 7003 Op0, Op1, 7004 Equal ? ISD::SETEQ : ISD::SETNE); 7005 // Replace the uses of XOR with SETCC 7006 WorkListRemover DeadNodes(*this); 7007 DAG.ReplaceAllUsesOfValueWith(N1, SetCC); 7008 removeFromWorkList(N1.getNode()); 7009 DAG.DeleteNode(N1.getNode()); 7010 return DAG.getNode(ISD::BRCOND, SDLoc(N), 7011 MVT::Other, Chain, SetCC, N2); 7012 } 7013 } 7014 7015 return SDValue(); 7016} 7017 7018// Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB. 7019// 7020SDValue DAGCombiner::visitBR_CC(SDNode *N) { 7021 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1)); 7022 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3); 7023 7024 // If N is a constant we could fold this into a fallthrough or unconditional 7025 // branch. However that doesn't happen very often in normal code, because 7026 // Instcombine/SimplifyCFG should have handled the available opportunities. 7027 // If we did this folding here, it would be necessary to update the 7028 // MachineBasicBlock CFG, which is awkward. 7029 7030 // Use SimplifySetCC to simplify SETCC's. 7031 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()), 7032 CondLHS, CondRHS, CC->get(), SDLoc(N), 7033 false); 7034 if (Simp.getNode()) AddToWorkList(Simp.getNode()); 7035 7036 // fold to a simpler setcc 7037 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC) 7038 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other, 7039 N->getOperand(0), Simp.getOperand(2), 7040 Simp.getOperand(0), Simp.getOperand(1), 7041 N->getOperand(4)); 7042 7043 return SDValue(); 7044} 7045 7046/// canFoldInAddressingMode - Return true if 'Use' is a load or a store that 7047/// uses N as its base pointer and that N may be folded in the load / store 7048/// addressing mode. 7049static bool canFoldInAddressingMode(SDNode *N, SDNode *Use, 7050 SelectionDAG &DAG, 7051 const TargetLowering &TLI) { 7052 EVT VT; 7053 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) { 7054 if (LD->isIndexed() || LD->getBasePtr().getNode() != N) 7055 return false; 7056 VT = Use->getValueType(0); 7057 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) { 7058 if (ST->isIndexed() || ST->getBasePtr().getNode() != N) 7059 return false; 7060 VT = ST->getValue().getValueType(); 7061 } else 7062 return false; 7063 7064 TargetLowering::AddrMode AM; 7065 if (N->getOpcode() == ISD::ADD) { 7066 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); 7067 if (Offset) 7068 // [reg +/- imm] 7069 AM.BaseOffs = Offset->getSExtValue(); 7070 else 7071 // [reg +/- reg] 7072 AM.Scale = 1; 7073 } else if (N->getOpcode() == ISD::SUB) { 7074 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1)); 7075 if (Offset) 7076 // [reg +/- imm] 7077 AM.BaseOffs = -Offset->getSExtValue(); 7078 else 7079 // [reg +/- reg] 7080 AM.Scale = 1; 7081 } else 7082 return false; 7083 7084 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext())); 7085} 7086 7087/// CombineToPreIndexedLoadStore - Try turning a load / store into a 7088/// pre-indexed load / store when the base pointer is an add or subtract 7089/// and it has other uses besides the load / store. After the 7090/// transformation, the new indexed load / store has effectively folded 7091/// the add / subtract in and all of its other uses are redirected to the 7092/// new load / store. 7093bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) { 7094 if (Level < AfterLegalizeDAG) 7095 return false; 7096 7097 bool isLoad = true; 7098 SDValue Ptr; 7099 EVT VT; 7100 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 7101 if (LD->isIndexed()) 7102 return false; 7103 VT = LD->getMemoryVT(); 7104 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) && 7105 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT)) 7106 return false; 7107 Ptr = LD->getBasePtr(); 7108 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 7109 if (ST->isIndexed()) 7110 return false; 7111 VT = ST->getMemoryVT(); 7112 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) && 7113 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT)) 7114 return false; 7115 Ptr = ST->getBasePtr(); 7116 isLoad = false; 7117 } else { 7118 return false; 7119 } 7120 7121 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail 7122 // out. There is no reason to make this a preinc/predec. 7123 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) || 7124 Ptr.getNode()->hasOneUse()) 7125 return false; 7126 7127 // Ask the target to do addressing mode selection. 7128 SDValue BasePtr; 7129 SDValue Offset; 7130 ISD::MemIndexedMode AM = ISD::UNINDEXED; 7131 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG)) 7132 return false; 7133 7134 // Backends without true r+i pre-indexed forms may need to pass a 7135 // constant base with a variable offset so that constant coercion 7136 // will work with the patterns in canonical form. 7137 bool Swapped = false; 7138 if (isa<ConstantSDNode>(BasePtr)) { 7139 std::swap(BasePtr, Offset); 7140 Swapped = true; 7141 } 7142 7143 // Don't create a indexed load / store with zero offset. 7144 if (isa<ConstantSDNode>(Offset) && 7145 cast<ConstantSDNode>(Offset)->isNullValue()) 7146 return false; 7147 7148 // Try turning it into a pre-indexed load / store except when: 7149 // 1) The new base ptr is a frame index. 7150 // 2) If N is a store and the new base ptr is either the same as or is a 7151 // predecessor of the value being stored. 7152 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded 7153 // that would create a cycle. 7154 // 4) All uses are load / store ops that use it as old base ptr. 7155 7156 // Check #1. Preinc'ing a frame index would require copying the stack pointer 7157 // (plus the implicit offset) to a register to preinc anyway. 7158 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr)) 7159 return false; 7160 7161 // Check #2. 7162 if (!isLoad) { 7163 SDValue Val = cast<StoreSDNode>(N)->getValue(); 7164 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode())) 7165 return false; 7166 } 7167 7168 // If the offset is a constant, there may be other adds of constants that 7169 // can be folded with this one. We should do this to avoid having to keep 7170 // a copy of the original base pointer. 7171 SmallVector<SDNode *, 16> OtherUses; 7172 if (isa<ConstantSDNode>(Offset)) 7173 for (SDNode::use_iterator I = BasePtr.getNode()->use_begin(), 7174 E = BasePtr.getNode()->use_end(); I != E; ++I) { 7175 SDNode *Use = *I; 7176 if (Use == Ptr.getNode()) 7177 continue; 7178 7179 if (Use->isPredecessorOf(N)) 7180 continue; 7181 7182 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) { 7183 OtherUses.clear(); 7184 break; 7185 } 7186 7187 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1); 7188 if (Op1.getNode() == BasePtr.getNode()) 7189 std::swap(Op0, Op1); 7190 assert(Op0.getNode() == BasePtr.getNode() && 7191 "Use of ADD/SUB but not an operand"); 7192 7193 if (!isa<ConstantSDNode>(Op1)) { 7194 OtherUses.clear(); 7195 break; 7196 } 7197 7198 // FIXME: In some cases, we can be smarter about this. 7199 if (Op1.getValueType() != Offset.getValueType()) { 7200 OtherUses.clear(); 7201 break; 7202 } 7203 7204 OtherUses.push_back(Use); 7205 } 7206 7207 if (Swapped) 7208 std::swap(BasePtr, Offset); 7209 7210 // Now check for #3 and #4. 7211 bool RealUse = false; 7212 7213 // Caches for hasPredecessorHelper 7214 SmallPtrSet<const SDNode *, 32> Visited; 7215 SmallVector<const SDNode *, 16> Worklist; 7216 7217 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(), 7218 E = Ptr.getNode()->use_end(); I != E; ++I) { 7219 SDNode *Use = *I; 7220 if (Use == N) 7221 continue; 7222 if (N->hasPredecessorHelper(Use, Visited, Worklist)) 7223 return false; 7224 7225 // If Ptr may be folded in addressing mode of other use, then it's 7226 // not profitable to do this transformation. 7227 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI)) 7228 RealUse = true; 7229 } 7230 7231 if (!RealUse) 7232 return false; 7233 7234 SDValue Result; 7235 if (isLoad) 7236 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N), 7237 BasePtr, Offset, AM); 7238 else 7239 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N), 7240 BasePtr, Offset, AM); 7241 ++PreIndexedNodes; 7242 ++NodesCombined; 7243 DEBUG(dbgs() << "\nReplacing.4 "; 7244 N->dump(&DAG); 7245 dbgs() << "\nWith: "; 7246 Result.getNode()->dump(&DAG); 7247 dbgs() << '\n'); 7248 WorkListRemover DeadNodes(*this); 7249 if (isLoad) { 7250 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0)); 7251 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2)); 7252 } else { 7253 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1)); 7254 } 7255 7256 // Finally, since the node is now dead, remove it from the graph. 7257 DAG.DeleteNode(N); 7258 7259 if (Swapped) 7260 std::swap(BasePtr, Offset); 7261 7262 // Replace other uses of BasePtr that can be updated to use Ptr 7263 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) { 7264 unsigned OffsetIdx = 1; 7265 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode()) 7266 OffsetIdx = 0; 7267 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() == 7268 BasePtr.getNode() && "Expected BasePtr operand"); 7269 7270 // We need to replace ptr0 in the following expression: 7271 // x0 * offset0 + y0 * ptr0 = t0 7272 // knowing that 7273 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store) 7274 // 7275 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the 7276 // indexed load/store and the expresion that needs to be re-written. 7277 // 7278 // Therefore, we have: 7279 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1 7280 7281 ConstantSDNode *CN = 7282 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx)); 7283 int X0, X1, Y0, Y1; 7284 APInt Offset0 = CN->getAPIntValue(); 7285 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue(); 7286 7287 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1; 7288 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1; 7289 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1; 7290 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1; 7291 7292 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD; 7293 7294 APInt CNV = Offset0; 7295 if (X0 < 0) CNV = -CNV; 7296 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1; 7297 else CNV = CNV - Offset1; 7298 7299 // We can now generate the new expression. 7300 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0)); 7301 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0); 7302 7303 SDValue NewUse = DAG.getNode(Opcode, 7304 SDLoc(OtherUses[i]), 7305 OtherUses[i]->getValueType(0), NewOp1, NewOp2); 7306 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse); 7307 removeFromWorkList(OtherUses[i]); 7308 DAG.DeleteNode(OtherUses[i]); 7309 } 7310 7311 // Replace the uses of Ptr with uses of the updated base value. 7312 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0)); 7313 removeFromWorkList(Ptr.getNode()); 7314 DAG.DeleteNode(Ptr.getNode()); 7315 7316 return true; 7317} 7318 7319/// CombineToPostIndexedLoadStore - Try to combine a load / store with a 7320/// add / sub of the base pointer node into a post-indexed load / store. 7321/// The transformation folded the add / subtract into the new indexed 7322/// load / store effectively and all of its uses are redirected to the 7323/// new load / store. 7324bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) { 7325 if (Level < AfterLegalizeDAG) 7326 return false; 7327 7328 bool isLoad = true; 7329 SDValue Ptr; 7330 EVT VT; 7331 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 7332 if (LD->isIndexed()) 7333 return false; 7334 VT = LD->getMemoryVT(); 7335 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) && 7336 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT)) 7337 return false; 7338 Ptr = LD->getBasePtr(); 7339 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 7340 if (ST->isIndexed()) 7341 return false; 7342 VT = ST->getMemoryVT(); 7343 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) && 7344 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT)) 7345 return false; 7346 Ptr = ST->getBasePtr(); 7347 isLoad = false; 7348 } else { 7349 return false; 7350 } 7351 7352 if (Ptr.getNode()->hasOneUse()) 7353 return false; 7354 7355 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(), 7356 E = Ptr.getNode()->use_end(); I != E; ++I) { 7357 SDNode *Op = *I; 7358 if (Op == N || 7359 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)) 7360 continue; 7361 7362 SDValue BasePtr; 7363 SDValue Offset; 7364 ISD::MemIndexedMode AM = ISD::UNINDEXED; 7365 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) { 7366 // Don't create a indexed load / store with zero offset. 7367 if (isa<ConstantSDNode>(Offset) && 7368 cast<ConstantSDNode>(Offset)->isNullValue()) 7369 continue; 7370 7371 // Try turning it into a post-indexed load / store except when 7372 // 1) All uses are load / store ops that use it as base ptr (and 7373 // it may be folded as addressing mmode). 7374 // 2) Op must be independent of N, i.e. Op is neither a predecessor 7375 // nor a successor of N. Otherwise, if Op is folded that would 7376 // create a cycle. 7377 7378 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr)) 7379 continue; 7380 7381 // Check for #1. 7382 bool TryNext = false; 7383 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(), 7384 EE = BasePtr.getNode()->use_end(); II != EE; ++II) { 7385 SDNode *Use = *II; 7386 if (Use == Ptr.getNode()) 7387 continue; 7388 7389 // If all the uses are load / store addresses, then don't do the 7390 // transformation. 7391 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){ 7392 bool RealUse = false; 7393 for (SDNode::use_iterator III = Use->use_begin(), 7394 EEE = Use->use_end(); III != EEE; ++III) { 7395 SDNode *UseUse = *III; 7396 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI)) 7397 RealUse = true; 7398 } 7399 7400 if (!RealUse) { 7401 TryNext = true; 7402 break; 7403 } 7404 } 7405 } 7406 7407 if (TryNext) 7408 continue; 7409 7410 // Check for #2 7411 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) { 7412 SDValue Result = isLoad 7413 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N), 7414 BasePtr, Offset, AM) 7415 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N), 7416 BasePtr, Offset, AM); 7417 ++PostIndexedNodes; 7418 ++NodesCombined; 7419 DEBUG(dbgs() << "\nReplacing.5 "; 7420 N->dump(&DAG); 7421 dbgs() << "\nWith: "; 7422 Result.getNode()->dump(&DAG); 7423 dbgs() << '\n'); 7424 WorkListRemover DeadNodes(*this); 7425 if (isLoad) { 7426 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0)); 7427 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2)); 7428 } else { 7429 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1)); 7430 } 7431 7432 // Finally, since the node is now dead, remove it from the graph. 7433 DAG.DeleteNode(N); 7434 7435 // Replace the uses of Use with uses of the updated base value. 7436 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0), 7437 Result.getValue(isLoad ? 1 : 0)); 7438 removeFromWorkList(Op); 7439 DAG.DeleteNode(Op); 7440 return true; 7441 } 7442 } 7443 } 7444 7445 return false; 7446} 7447 7448SDValue DAGCombiner::visitLOAD(SDNode *N) { 7449 LoadSDNode *LD = cast<LoadSDNode>(N); 7450 SDValue Chain = LD->getChain(); 7451 SDValue Ptr = LD->getBasePtr(); 7452 7453 // If load is not volatile and there are no uses of the loaded value (and 7454 // the updated indexed value in case of indexed loads), change uses of the 7455 // chain value into uses of the chain input (i.e. delete the dead load). 7456 if (!LD->isVolatile()) { 7457 if (N->getValueType(1) == MVT::Other) { 7458 // Unindexed loads. 7459 if (!N->hasAnyUseOfValue(0)) { 7460 // It's not safe to use the two value CombineTo variant here. e.g. 7461 // v1, chain2 = load chain1, loc 7462 // v2, chain3 = load chain2, loc 7463 // v3 = add v2, c 7464 // Now we replace use of chain2 with chain1. This makes the second load 7465 // isomorphic to the one we are deleting, and thus makes this load live. 7466 DEBUG(dbgs() << "\nReplacing.6 "; 7467 N->dump(&DAG); 7468 dbgs() << "\nWith chain: "; 7469 Chain.getNode()->dump(&DAG); 7470 dbgs() << "\n"); 7471 WorkListRemover DeadNodes(*this); 7472 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain); 7473 7474 if (N->use_empty()) { 7475 removeFromWorkList(N); 7476 DAG.DeleteNode(N); 7477 } 7478 7479 return SDValue(N, 0); // Return N so it doesn't get rechecked! 7480 } 7481 } else { 7482 // Indexed loads. 7483 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?"); 7484 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) { 7485 SDValue Undef = DAG.getUNDEF(N->getValueType(0)); 7486 DEBUG(dbgs() << "\nReplacing.7 "; 7487 N->dump(&DAG); 7488 dbgs() << "\nWith: "; 7489 Undef.getNode()->dump(&DAG); 7490 dbgs() << " and 2 other values\n"); 7491 WorkListRemover DeadNodes(*this); 7492 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef); 7493 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), 7494 DAG.getUNDEF(N->getValueType(1))); 7495 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain); 7496 removeFromWorkList(N); 7497 DAG.DeleteNode(N); 7498 return SDValue(N, 0); // Return N so it doesn't get rechecked! 7499 } 7500 } 7501 } 7502 7503 // If this load is directly stored, replace the load value with the stored 7504 // value. 7505 // TODO: Handle store large -> read small portion. 7506 // TODO: Handle TRUNCSTORE/LOADEXT 7507 if (ISD::isNormalLoad(N) && !LD->isVolatile()) { 7508 if (ISD::isNON_TRUNCStore(Chain.getNode())) { 7509 StoreSDNode *PrevST = cast<StoreSDNode>(Chain); 7510 if (PrevST->getBasePtr() == Ptr && 7511 PrevST->getValue().getValueType() == N->getValueType(0)) 7512 return CombineTo(N, Chain.getOperand(1), Chain); 7513 } 7514 } 7515 7516 // Try to infer better alignment information than the load already has. 7517 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) { 7518 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) { 7519 if (Align > LD->getMemOperand()->getBaseAlignment()) { 7520 SDValue NewLoad = 7521 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N), 7522 LD->getValueType(0), 7523 Chain, Ptr, LD->getPointerInfo(), 7524 LD->getMemoryVT(), 7525 LD->isVolatile(), LD->isNonTemporal(), Align); 7526 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true); 7527 } 7528 } 7529 } 7530 7531 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA : 7532 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA(); 7533 if (UseAA) { 7534 // Walk up chain skipping non-aliasing memory nodes. 7535 SDValue BetterChain = FindBetterChain(N, Chain); 7536 7537 // If there is a better chain. 7538 if (Chain != BetterChain) { 7539 SDValue ReplLoad; 7540 7541 // Replace the chain to void dependency. 7542 if (LD->getExtensionType() == ISD::NON_EXTLOAD) { 7543 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD), 7544 BetterChain, Ptr, LD->getPointerInfo(), 7545 LD->isVolatile(), LD->isNonTemporal(), 7546 LD->isInvariant(), LD->getAlignment()); 7547 } else { 7548 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD), 7549 LD->getValueType(0), 7550 BetterChain, Ptr, LD->getPointerInfo(), 7551 LD->getMemoryVT(), 7552 LD->isVolatile(), 7553 LD->isNonTemporal(), 7554 LD->getAlignment()); 7555 } 7556 7557 // Create token factor to keep old chain connected. 7558 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N), 7559 MVT::Other, Chain, ReplLoad.getValue(1)); 7560 7561 // Make sure the new and old chains are cleaned up. 7562 AddToWorkList(Token.getNode()); 7563 7564 // Replace uses with load result and token factor. Don't add users 7565 // to work list. 7566 return CombineTo(N, ReplLoad.getValue(0), Token, false); 7567 } 7568 } 7569 7570 // Try transforming N to an indexed load. 7571 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) 7572 return SDValue(N, 0); 7573 7574 return SDValue(); 7575} 7576 7577/// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the 7578/// load is having specific bytes cleared out. If so, return the byte size 7579/// being masked out and the shift amount. 7580static std::pair<unsigned, unsigned> 7581CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) { 7582 std::pair<unsigned, unsigned> Result(0, 0); 7583 7584 // Check for the structure we're looking for. 7585 if (V->getOpcode() != ISD::AND || 7586 !isa<ConstantSDNode>(V->getOperand(1)) || 7587 !ISD::isNormalLoad(V->getOperand(0).getNode())) 7588 return Result; 7589 7590 // Check the chain and pointer. 7591 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0)); 7592 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer. 7593 7594 // The store should be chained directly to the load or be an operand of a 7595 // tokenfactor. 7596 if (LD == Chain.getNode()) 7597 ; // ok. 7598 else if (Chain->getOpcode() != ISD::TokenFactor) 7599 return Result; // Fail. 7600 else { 7601 bool isOk = false; 7602 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i) 7603 if (Chain->getOperand(i).getNode() == LD) { 7604 isOk = true; 7605 break; 7606 } 7607 if (!isOk) return Result; 7608 } 7609 7610 // This only handles simple types. 7611 if (V.getValueType() != MVT::i16 && 7612 V.getValueType() != MVT::i32 && 7613 V.getValueType() != MVT::i64) 7614 return Result; 7615 7616 // Check the constant mask. Invert it so that the bits being masked out are 7617 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits 7618 // follow the sign bit for uniformity. 7619 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue(); 7620 unsigned NotMaskLZ = countLeadingZeros(NotMask); 7621 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte. 7622 unsigned NotMaskTZ = countTrailingZeros(NotMask); 7623 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte. 7624 if (NotMaskLZ == 64) return Result; // All zero mask. 7625 7626 // See if we have a continuous run of bits. If so, we have 0*1+0* 7627 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64) 7628 return Result; 7629 7630 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64. 7631 if (V.getValueType() != MVT::i64 && NotMaskLZ) 7632 NotMaskLZ -= 64-V.getValueSizeInBits(); 7633 7634 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8; 7635 switch (MaskedBytes) { 7636 case 1: 7637 case 2: 7638 case 4: break; 7639 default: return Result; // All one mask, or 5-byte mask. 7640 } 7641 7642 // Verify that the first bit starts at a multiple of mask so that the access 7643 // is aligned the same as the access width. 7644 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result; 7645 7646 Result.first = MaskedBytes; 7647 Result.second = NotMaskTZ/8; 7648 return Result; 7649} 7650 7651 7652/// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that 7653/// provides a value as specified by MaskInfo. If so, replace the specified 7654/// store with a narrower store of truncated IVal. 7655static SDNode * 7656ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo, 7657 SDValue IVal, StoreSDNode *St, 7658 DAGCombiner *DC) { 7659 unsigned NumBytes = MaskInfo.first; 7660 unsigned ByteShift = MaskInfo.second; 7661 SelectionDAG &DAG = DC->getDAG(); 7662 7663 // Check to see if IVal is all zeros in the part being masked in by the 'or' 7664 // that uses this. If not, this is not a replacement. 7665 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(), 7666 ByteShift*8, (ByteShift+NumBytes)*8); 7667 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0; 7668 7669 // Check that it is legal on the target to do this. It is legal if the new 7670 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type 7671 // legalization. 7672 MVT VT = MVT::getIntegerVT(NumBytes*8); 7673 if (!DC->isTypeLegal(VT)) 7674 return 0; 7675 7676 // Okay, we can do this! Replace the 'St' store with a store of IVal that is 7677 // shifted by ByteShift and truncated down to NumBytes. 7678 if (ByteShift) 7679 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal, 7680 DAG.getConstant(ByteShift*8, 7681 DC->getShiftAmountTy(IVal.getValueType()))); 7682 7683 // Figure out the offset for the store and the alignment of the access. 7684 unsigned StOffset; 7685 unsigned NewAlign = St->getAlignment(); 7686 7687 if (DAG.getTargetLoweringInfo().isLittleEndian()) 7688 StOffset = ByteShift; 7689 else 7690 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes; 7691 7692 SDValue Ptr = St->getBasePtr(); 7693 if (StOffset) { 7694 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(), 7695 Ptr, DAG.getConstant(StOffset, Ptr.getValueType())); 7696 NewAlign = MinAlign(NewAlign, StOffset); 7697 } 7698 7699 // Truncate down to the new size. 7700 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal); 7701 7702 ++OpsNarrowed; 7703 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr, 7704 St->getPointerInfo().getWithOffset(StOffset), 7705 false, false, NewAlign).getNode(); 7706} 7707 7708 7709/// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is 7710/// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some 7711/// of the loaded bits, try narrowing the load and store if it would end up 7712/// being a win for performance or code size. 7713SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) { 7714 StoreSDNode *ST = cast<StoreSDNode>(N); 7715 if (ST->isVolatile()) 7716 return SDValue(); 7717 7718 SDValue Chain = ST->getChain(); 7719 SDValue Value = ST->getValue(); 7720 SDValue Ptr = ST->getBasePtr(); 7721 EVT VT = Value.getValueType(); 7722 7723 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse()) 7724 return SDValue(); 7725 7726 unsigned Opc = Value.getOpcode(); 7727 7728 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst 7729 // is a byte mask indicating a consecutive number of bytes, check to see if 7730 // Y is known to provide just those bytes. If so, we try to replace the 7731 // load + replace + store sequence with a single (narrower) store, which makes 7732 // the load dead. 7733 if (Opc == ISD::OR) { 7734 std::pair<unsigned, unsigned> MaskedLoad; 7735 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain); 7736 if (MaskedLoad.first) 7737 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad, 7738 Value.getOperand(1), ST,this)) 7739 return SDValue(NewST, 0); 7740 7741 // Or is commutative, so try swapping X and Y. 7742 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain); 7743 if (MaskedLoad.first) 7744 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad, 7745 Value.getOperand(0), ST,this)) 7746 return SDValue(NewST, 0); 7747 } 7748 7749 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) || 7750 Value.getOperand(1).getOpcode() != ISD::Constant) 7751 return SDValue(); 7752 7753 SDValue N0 = Value.getOperand(0); 7754 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() && 7755 Chain == SDValue(N0.getNode(), 1)) { 7756 LoadSDNode *LD = cast<LoadSDNode>(N0); 7757 if (LD->getBasePtr() != Ptr || 7758 LD->getPointerInfo().getAddrSpace() != 7759 ST->getPointerInfo().getAddrSpace()) 7760 return SDValue(); 7761 7762 // Find the type to narrow it the load / op / store to. 7763 SDValue N1 = Value.getOperand(1); 7764 unsigned BitWidth = N1.getValueSizeInBits(); 7765 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue(); 7766 if (Opc == ISD::AND) 7767 Imm ^= APInt::getAllOnesValue(BitWidth); 7768 if (Imm == 0 || Imm.isAllOnesValue()) 7769 return SDValue(); 7770 unsigned ShAmt = Imm.countTrailingZeros(); 7771 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1; 7772 unsigned NewBW = NextPowerOf2(MSB - ShAmt); 7773 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW); 7774 while (NewBW < BitWidth && 7775 !(TLI.isOperationLegalOrCustom(Opc, NewVT) && 7776 TLI.isNarrowingProfitable(VT, NewVT))) { 7777 NewBW = NextPowerOf2(NewBW); 7778 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW); 7779 } 7780 if (NewBW >= BitWidth) 7781 return SDValue(); 7782 7783 // If the lsb changed does not start at the type bitwidth boundary, 7784 // start at the previous one. 7785 if (ShAmt % NewBW) 7786 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW; 7787 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, 7788 std::min(BitWidth, ShAmt + NewBW)); 7789 if ((Imm & Mask) == Imm) { 7790 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW); 7791 if (Opc == ISD::AND) 7792 NewImm ^= APInt::getAllOnesValue(NewBW); 7793 uint64_t PtrOff = ShAmt / 8; 7794 // For big endian targets, we need to adjust the offset to the pointer to 7795 // load the correct bytes. 7796 if (TLI.isBigEndian()) 7797 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff; 7798 7799 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff); 7800 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext()); 7801 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy)) 7802 return SDValue(); 7803 7804 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD), 7805 Ptr.getValueType(), Ptr, 7806 DAG.getConstant(PtrOff, Ptr.getValueType())); 7807 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0), 7808 LD->getChain(), NewPtr, 7809 LD->getPointerInfo().getWithOffset(PtrOff), 7810 LD->isVolatile(), LD->isNonTemporal(), 7811 LD->isInvariant(), NewAlign); 7812 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD, 7813 DAG.getConstant(NewImm, NewVT)); 7814 SDValue NewST = DAG.getStore(Chain, SDLoc(N), 7815 NewVal, NewPtr, 7816 ST->getPointerInfo().getWithOffset(PtrOff), 7817 false, false, NewAlign); 7818 7819 AddToWorkList(NewPtr.getNode()); 7820 AddToWorkList(NewLD.getNode()); 7821 AddToWorkList(NewVal.getNode()); 7822 WorkListRemover DeadNodes(*this); 7823 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1)); 7824 ++OpsNarrowed; 7825 return NewST; 7826 } 7827 } 7828 7829 return SDValue(); 7830} 7831 7832/// TransformFPLoadStorePair - For a given floating point load / store pair, 7833/// if the load value isn't used by any other operations, then consider 7834/// transforming the pair to integer load / store operations if the target 7835/// deems the transformation profitable. 7836SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) { 7837 StoreSDNode *ST = cast<StoreSDNode>(N); 7838 SDValue Chain = ST->getChain(); 7839 SDValue Value = ST->getValue(); 7840 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) && 7841 Value.hasOneUse() && 7842 Chain == SDValue(Value.getNode(), 1)) { 7843 LoadSDNode *LD = cast<LoadSDNode>(Value); 7844 EVT VT = LD->getMemoryVT(); 7845 if (!VT.isFloatingPoint() || 7846 VT != ST->getMemoryVT() || 7847 LD->isNonTemporal() || 7848 ST->isNonTemporal() || 7849 LD->getPointerInfo().getAddrSpace() != 0 || 7850 ST->getPointerInfo().getAddrSpace() != 0) 7851 return SDValue(); 7852 7853 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits()); 7854 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) || 7855 !TLI.isOperationLegal(ISD::STORE, IntVT) || 7856 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) || 7857 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT)) 7858 return SDValue(); 7859 7860 unsigned LDAlign = LD->getAlignment(); 7861 unsigned STAlign = ST->getAlignment(); 7862 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext()); 7863 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy); 7864 if (LDAlign < ABIAlign || STAlign < ABIAlign) 7865 return SDValue(); 7866 7867 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value), 7868 LD->getChain(), LD->getBasePtr(), 7869 LD->getPointerInfo(), 7870 false, false, false, LDAlign); 7871 7872 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N), 7873 NewLD, ST->getBasePtr(), 7874 ST->getPointerInfo(), 7875 false, false, STAlign); 7876 7877 AddToWorkList(NewLD.getNode()); 7878 AddToWorkList(NewST.getNode()); 7879 WorkListRemover DeadNodes(*this); 7880 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1)); 7881 ++LdStFP2Int; 7882 return NewST; 7883 } 7884 7885 return SDValue(); 7886} 7887 7888/// Helper struct to parse and store a memory address as base + index + offset. 7889/// We ignore sign extensions when it is safe to do so. 7890/// The following two expressions are not equivalent. To differentiate we need 7891/// to store whether there was a sign extension involved in the index 7892/// computation. 7893/// (load (i64 add (i64 copyfromreg %c) 7894/// (i64 signextend (add (i8 load %index) 7895/// (i8 1)))) 7896/// vs 7897/// 7898/// (load (i64 add (i64 copyfromreg %c) 7899/// (i64 signextend (i32 add (i32 signextend (i8 load %index)) 7900/// (i32 1))))) 7901struct BaseIndexOffset { 7902 SDValue Base; 7903 SDValue Index; 7904 int64_t Offset; 7905 bool IsIndexSignExt; 7906 7907 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {} 7908 7909 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset, 7910 bool IsIndexSignExt) : 7911 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {} 7912 7913 bool equalBaseIndex(const BaseIndexOffset &Other) { 7914 return Other.Base == Base && Other.Index == Index && 7915 Other.IsIndexSignExt == IsIndexSignExt; 7916 } 7917 7918 /// Parses tree in Ptr for base, index, offset addresses. 7919 static BaseIndexOffset match(SDValue Ptr) { 7920 bool IsIndexSignExt = false; 7921 7922 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD 7923 // instruction, then it could be just the BASE or everything else we don't 7924 // know how to handle. Just use Ptr as BASE and give up. 7925 if (Ptr->getOpcode() != ISD::ADD) 7926 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt); 7927 7928 // We know that we have at least an ADD instruction. Try to pattern match 7929 // the simple case of BASE + OFFSET. 7930 if (isa<ConstantSDNode>(Ptr->getOperand(1))) { 7931 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue(); 7932 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset, 7933 IsIndexSignExt); 7934 } 7935 7936 // Inside a loop the current BASE pointer is calculated using an ADD and a 7937 // MUL instruction. In this case Ptr is the actual BASE pointer. 7938 // (i64 add (i64 %array_ptr) 7939 // (i64 mul (i64 %induction_var) 7940 // (i64 %element_size))) 7941 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL) 7942 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt); 7943 7944 // Look at Base + Index + Offset cases. 7945 SDValue Base = Ptr->getOperand(0); 7946 SDValue IndexOffset = Ptr->getOperand(1); 7947 7948 // Skip signextends. 7949 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) { 7950 IndexOffset = IndexOffset->getOperand(0); 7951 IsIndexSignExt = true; 7952 } 7953 7954 // Either the case of Base + Index (no offset) or something else. 7955 if (IndexOffset->getOpcode() != ISD::ADD) 7956 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt); 7957 7958 // Now we have the case of Base + Index + offset. 7959 SDValue Index = IndexOffset->getOperand(0); 7960 SDValue Offset = IndexOffset->getOperand(1); 7961 7962 if (!isa<ConstantSDNode>(Offset)) 7963 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt); 7964 7965 // Ignore signextends. 7966 if (Index->getOpcode() == ISD::SIGN_EXTEND) { 7967 Index = Index->getOperand(0); 7968 IsIndexSignExt = true; 7969 } else IsIndexSignExt = false; 7970 7971 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue(); 7972 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt); 7973 } 7974}; 7975 7976/// Holds a pointer to an LSBaseSDNode as well as information on where it 7977/// is located in a sequence of memory operations connected by a chain. 7978struct MemOpLink { 7979 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq): 7980 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { } 7981 // Ptr to the mem node. 7982 LSBaseSDNode *MemNode; 7983 // Offset from the base ptr. 7984 int64_t OffsetFromBase; 7985 // What is the sequence number of this mem node. 7986 // Lowest mem operand in the DAG starts at zero. 7987 unsigned SequenceNum; 7988}; 7989 7990/// Sorts store nodes in a link according to their offset from a shared 7991// base ptr. 7992struct ConsecutiveMemoryChainSorter { 7993 bool operator()(MemOpLink LHS, MemOpLink RHS) { 7994 return LHS.OffsetFromBase < RHS.OffsetFromBase; 7995 } 7996}; 7997 7998bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) { 7999 EVT MemVT = St->getMemoryVT(); 8000 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8; 8001 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes(). 8002 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat); 8003 8004 // Don't merge vectors into wider inputs. 8005 if (MemVT.isVector() || !MemVT.isSimple()) 8006 return false; 8007 8008 // Perform an early exit check. Do not bother looking at stored values that 8009 // are not constants or loads. 8010 SDValue StoredVal = St->getValue(); 8011 bool IsLoadSrc = isa<LoadSDNode>(StoredVal); 8012 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) && 8013 !IsLoadSrc) 8014 return false; 8015 8016 // Only look at ends of store sequences. 8017 SDValue Chain = SDValue(St, 1); 8018 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE) 8019 return false; 8020 8021 // This holds the base pointer, index, and the offset in bytes from the base 8022 // pointer. 8023 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr()); 8024 8025 // We must have a base and an offset. 8026 if (!BasePtr.Base.getNode()) 8027 return false; 8028 8029 // Do not handle stores to undef base pointers. 8030 if (BasePtr.Base.getOpcode() == ISD::UNDEF) 8031 return false; 8032 8033 // Save the LoadSDNodes that we find in the chain. 8034 // We need to make sure that these nodes do not interfere with 8035 // any of the store nodes. 8036 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes; 8037 8038 // Save the StoreSDNodes that we find in the chain. 8039 SmallVector<MemOpLink, 8> StoreNodes; 8040 8041 // Walk up the chain and look for nodes with offsets from the same 8042 // base pointer. Stop when reaching an instruction with a different kind 8043 // or instruction which has a different base pointer. 8044 unsigned Seq = 0; 8045 StoreSDNode *Index = St; 8046 while (Index) { 8047 // If the chain has more than one use, then we can't reorder the mem ops. 8048 if (Index != St && !SDValue(Index, 1)->hasOneUse()) 8049 break; 8050 8051 // Find the base pointer and offset for this memory node. 8052 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr()); 8053 8054 // Check that the base pointer is the same as the original one. 8055 if (!Ptr.equalBaseIndex(BasePtr)) 8056 break; 8057 8058 // Check that the alignment is the same. 8059 if (Index->getAlignment() != St->getAlignment()) 8060 break; 8061 8062 // The memory operands must not be volatile. 8063 if (Index->isVolatile() || Index->isIndexed()) 8064 break; 8065 8066 // No truncation. 8067 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index)) 8068 if (St->isTruncatingStore()) 8069 break; 8070 8071 // The stored memory type must be the same. 8072 if (Index->getMemoryVT() != MemVT) 8073 break; 8074 8075 // We do not allow unaligned stores because we want to prevent overriding 8076 // stores. 8077 if (Index->getAlignment()*8 != MemVT.getSizeInBits()) 8078 break; 8079 8080 // We found a potential memory operand to merge. 8081 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++)); 8082 8083 // Find the next memory operand in the chain. If the next operand in the 8084 // chain is a store then move up and continue the scan with the next 8085 // memory operand. If the next operand is a load save it and use alias 8086 // information to check if it interferes with anything. 8087 SDNode *NextInChain = Index->getChain().getNode(); 8088 while (1) { 8089 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) { 8090 // We found a store node. Use it for the next iteration. 8091 Index = STn; 8092 break; 8093 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) { 8094 // Save the load node for later. Continue the scan. 8095 AliasLoadNodes.push_back(Ldn); 8096 NextInChain = Ldn->getChain().getNode(); 8097 continue; 8098 } else { 8099 Index = NULL; 8100 break; 8101 } 8102 } 8103 } 8104 8105 // Check if there is anything to merge. 8106 if (StoreNodes.size() < 2) 8107 return false; 8108 8109 // Sort the memory operands according to their distance from the base pointer. 8110 std::sort(StoreNodes.begin(), StoreNodes.end(), 8111 ConsecutiveMemoryChainSorter()); 8112 8113 // Scan the memory operations on the chain and find the first non-consecutive 8114 // store memory address. 8115 unsigned LastConsecutiveStore = 0; 8116 int64_t StartAddress = StoreNodes[0].OffsetFromBase; 8117 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) { 8118 8119 // Check that the addresses are consecutive starting from the second 8120 // element in the list of stores. 8121 if (i > 0) { 8122 int64_t CurrAddress = StoreNodes[i].OffsetFromBase; 8123 if (CurrAddress - StartAddress != (ElementSizeBytes * i)) 8124 break; 8125 } 8126 8127 bool Alias = false; 8128 // Check if this store interferes with any of the loads that we found. 8129 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld) 8130 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) { 8131 Alias = true; 8132 break; 8133 } 8134 // We found a load that alias with this store. Stop the sequence. 8135 if (Alias) 8136 break; 8137 8138 // Mark this node as useful. 8139 LastConsecutiveStore = i; 8140 } 8141 8142 // The node with the lowest store address. 8143 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode; 8144 8145 // Store the constants into memory as one consecutive store. 8146 if (!IsLoadSrc) { 8147 unsigned LastLegalType = 0; 8148 unsigned LastLegalVectorType = 0; 8149 bool NonZero = false; 8150 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) { 8151 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); 8152 SDValue StoredVal = St->getValue(); 8153 8154 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) { 8155 NonZero |= !C->isNullValue(); 8156 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) { 8157 NonZero |= !C->getConstantFPValue()->isNullValue(); 8158 } else { 8159 // Non constant. 8160 break; 8161 } 8162 8163 // Find a legal type for the constant store. 8164 unsigned StoreBW = (i+1) * ElementSizeBytes * 8; 8165 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW); 8166 if (TLI.isTypeLegal(StoreTy)) 8167 LastLegalType = i+1; 8168 // Or check whether a truncstore is legal. 8169 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) == 8170 TargetLowering::TypePromoteInteger) { 8171 EVT LegalizedStoredValueTy = 8172 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType()); 8173 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy)) 8174 LastLegalType = i+1; 8175 } 8176 8177 // Find a legal type for the vector store. 8178 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1); 8179 if (TLI.isTypeLegal(Ty)) 8180 LastLegalVectorType = i + 1; 8181 } 8182 8183 // We only use vectors if the constant is known to be zero and the 8184 // function is not marked with the noimplicitfloat attribute. 8185 if (NonZero || NoVectors) 8186 LastLegalVectorType = 0; 8187 8188 // Check if we found a legal integer type to store. 8189 if (LastLegalType == 0 && LastLegalVectorType == 0) 8190 return false; 8191 8192 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors; 8193 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType; 8194 8195 // Make sure we have something to merge. 8196 if (NumElem < 2) 8197 return false; 8198 8199 unsigned EarliestNodeUsed = 0; 8200 for (unsigned i=0; i < NumElem; ++i) { 8201 // Find a chain for the new wide-store operand. Notice that some 8202 // of the store nodes that we found may not be selected for inclusion 8203 // in the wide store. The chain we use needs to be the chain of the 8204 // earliest store node which is *used* and replaced by the wide store. 8205 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum) 8206 EarliestNodeUsed = i; 8207 } 8208 8209 // The earliest Node in the DAG. 8210 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode; 8211 SDLoc DL(StoreNodes[0].MemNode); 8212 8213 SDValue StoredVal; 8214 if (UseVector) { 8215 // Find a legal type for the vector store. 8216 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem); 8217 assert(TLI.isTypeLegal(Ty) && "Illegal vector store"); 8218 StoredVal = DAG.getConstant(0, Ty); 8219 } else { 8220 unsigned StoreBW = NumElem * ElementSizeBytes * 8; 8221 APInt StoreInt(StoreBW, 0); 8222 8223 // Construct a single integer constant which is made of the smaller 8224 // constant inputs. 8225 bool IsLE = TLI.isLittleEndian(); 8226 for (unsigned i = 0; i < NumElem ; ++i) { 8227 unsigned Idx = IsLE ?(NumElem - 1 - i) : i; 8228 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode); 8229 SDValue Val = St->getValue(); 8230 StoreInt<<=ElementSizeBytes*8; 8231 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) { 8232 StoreInt|=C->getAPIntValue().zext(StoreBW); 8233 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) { 8234 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW); 8235 } else { 8236 assert(false && "Invalid constant element type"); 8237 } 8238 } 8239 8240 // Create the new Load and Store operations. 8241 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW); 8242 StoredVal = DAG.getConstant(StoreInt, StoreTy); 8243 } 8244 8245 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal, 8246 FirstInChain->getBasePtr(), 8247 FirstInChain->getPointerInfo(), 8248 false, false, 8249 FirstInChain->getAlignment()); 8250 8251 // Replace the first store with the new store 8252 CombineTo(EarliestOp, NewStore); 8253 // Erase all other stores. 8254 for (unsigned i = 0; i < NumElem ; ++i) { 8255 if (StoreNodes[i].MemNode == EarliestOp) 8256 continue; 8257 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); 8258 // ReplaceAllUsesWith will replace all uses that existed when it was 8259 // called, but graph optimizations may cause new ones to appear. For 8260 // example, the case in pr14333 looks like 8261 // 8262 // St's chain -> St -> another store -> X 8263 // 8264 // And the only difference from St to the other store is the chain. 8265 // When we change it's chain to be St's chain they become identical, 8266 // get CSEed and the net result is that X is now a use of St. 8267 // Since we know that St is redundant, just iterate. 8268 while (!St->use_empty()) 8269 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain()); 8270 removeFromWorkList(St); 8271 DAG.DeleteNode(St); 8272 } 8273 8274 return true; 8275 } 8276 8277 // Below we handle the case of multiple consecutive stores that 8278 // come from multiple consecutive loads. We merge them into a single 8279 // wide load and a single wide store. 8280 8281 // Look for load nodes which are used by the stored values. 8282 SmallVector<MemOpLink, 8> LoadNodes; 8283 8284 // Find acceptable loads. Loads need to have the same chain (token factor), 8285 // must not be zext, volatile, indexed, and they must be consecutive. 8286 BaseIndexOffset LdBasePtr; 8287 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) { 8288 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); 8289 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue()); 8290 if (!Ld) break; 8291 8292 // Loads must only have one use. 8293 if (!Ld->hasNUsesOfValue(1, 0)) 8294 break; 8295 8296 // Check that the alignment is the same as the stores. 8297 if (Ld->getAlignment() != St->getAlignment()) 8298 break; 8299 8300 // The memory operands must not be volatile. 8301 if (Ld->isVolatile() || Ld->isIndexed()) 8302 break; 8303 8304 // We do not accept ext loads. 8305 if (Ld->getExtensionType() != ISD::NON_EXTLOAD) 8306 break; 8307 8308 // The stored memory type must be the same. 8309 if (Ld->getMemoryVT() != MemVT) 8310 break; 8311 8312 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr()); 8313 // If this is not the first ptr that we check. 8314 if (LdBasePtr.Base.getNode()) { 8315 // The base ptr must be the same. 8316 if (!LdPtr.equalBaseIndex(LdBasePtr)) 8317 break; 8318 } else { 8319 // Check that all other base pointers are the same as this one. 8320 LdBasePtr = LdPtr; 8321 } 8322 8323 // We found a potential memory operand to merge. 8324 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0)); 8325 } 8326 8327 if (LoadNodes.size() < 2) 8328 return false; 8329 8330 // Scan the memory operations on the chain and find the first non-consecutive 8331 // load memory address. These variables hold the index in the store node 8332 // array. 8333 unsigned LastConsecutiveLoad = 0; 8334 // This variable refers to the size and not index in the array. 8335 unsigned LastLegalVectorType = 0; 8336 unsigned LastLegalIntegerType = 0; 8337 StartAddress = LoadNodes[0].OffsetFromBase; 8338 SDValue FirstChain = LoadNodes[0].MemNode->getChain(); 8339 for (unsigned i = 1; i < LoadNodes.size(); ++i) { 8340 // All loads much share the same chain. 8341 if (LoadNodes[i].MemNode->getChain() != FirstChain) 8342 break; 8343 8344 int64_t CurrAddress = LoadNodes[i].OffsetFromBase; 8345 if (CurrAddress - StartAddress != (ElementSizeBytes * i)) 8346 break; 8347 LastConsecutiveLoad = i; 8348 8349 // Find a legal type for the vector store. 8350 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1); 8351 if (TLI.isTypeLegal(StoreTy)) 8352 LastLegalVectorType = i + 1; 8353 8354 // Find a legal type for the integer store. 8355 unsigned StoreBW = (i+1) * ElementSizeBytes * 8; 8356 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW); 8357 if (TLI.isTypeLegal(StoreTy)) 8358 LastLegalIntegerType = i + 1; 8359 // Or check whether a truncstore and extload is legal. 8360 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) == 8361 TargetLowering::TypePromoteInteger) { 8362 EVT LegalizedStoredValueTy = 8363 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy); 8364 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) && 8365 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) && 8366 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) && 8367 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy)) 8368 LastLegalIntegerType = i+1; 8369 } 8370 } 8371 8372 // Only use vector types if the vector type is larger than the integer type. 8373 // If they are the same, use integers. 8374 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors; 8375 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType); 8376 8377 // We add +1 here because the LastXXX variables refer to location while 8378 // the NumElem refers to array/index size. 8379 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1; 8380 NumElem = std::min(LastLegalType, NumElem); 8381 8382 if (NumElem < 2) 8383 return false; 8384 8385 // The earliest Node in the DAG. 8386 unsigned EarliestNodeUsed = 0; 8387 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode; 8388 for (unsigned i=1; i<NumElem; ++i) { 8389 // Find a chain for the new wide-store operand. Notice that some 8390 // of the store nodes that we found may not be selected for inclusion 8391 // in the wide store. The chain we use needs to be the chain of the 8392 // earliest store node which is *used* and replaced by the wide store. 8393 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum) 8394 EarliestNodeUsed = i; 8395 } 8396 8397 // Find if it is better to use vectors or integers to load and store 8398 // to memory. 8399 EVT JointMemOpVT; 8400 if (UseVectorTy) { 8401 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem); 8402 } else { 8403 unsigned StoreBW = NumElem * ElementSizeBytes * 8; 8404 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW); 8405 } 8406 8407 SDLoc LoadDL(LoadNodes[0].MemNode); 8408 SDLoc StoreDL(StoreNodes[0].MemNode); 8409 8410 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode); 8411 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL, 8412 FirstLoad->getChain(), 8413 FirstLoad->getBasePtr(), 8414 FirstLoad->getPointerInfo(), 8415 false, false, false, 8416 FirstLoad->getAlignment()); 8417 8418 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad, 8419 FirstInChain->getBasePtr(), 8420 FirstInChain->getPointerInfo(), false, false, 8421 FirstInChain->getAlignment()); 8422 8423 // Replace one of the loads with the new load. 8424 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode); 8425 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), 8426 SDValue(NewLoad.getNode(), 1)); 8427 8428 // Remove the rest of the load chains. 8429 for (unsigned i = 1; i < NumElem ; ++i) { 8430 // Replace all chain users of the old load nodes with the chain of the new 8431 // load node. 8432 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode); 8433 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain()); 8434 } 8435 8436 // Replace the first store with the new store. 8437 CombineTo(EarliestOp, NewStore); 8438 // Erase all other stores. 8439 for (unsigned i = 0; i < NumElem ; ++i) { 8440 // Remove all Store nodes. 8441 if (StoreNodes[i].MemNode == EarliestOp) 8442 continue; 8443 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode); 8444 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain()); 8445 removeFromWorkList(St); 8446 DAG.DeleteNode(St); 8447 } 8448 8449 return true; 8450} 8451 8452SDValue DAGCombiner::visitSTORE(SDNode *N) { 8453 StoreSDNode *ST = cast<StoreSDNode>(N); 8454 SDValue Chain = ST->getChain(); 8455 SDValue Value = ST->getValue(); 8456 SDValue Ptr = ST->getBasePtr(); 8457 8458 // If this is a store of a bit convert, store the input value if the 8459 // resultant store does not need a higher alignment than the original. 8460 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() && 8461 ST->isUnindexed()) { 8462 unsigned OrigAlign = ST->getAlignment(); 8463 EVT SVT = Value.getOperand(0).getValueType(); 8464 unsigned Align = TLI.getDataLayout()-> 8465 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext())); 8466 if (Align <= OrigAlign && 8467 ((!LegalOperations && !ST->isVolatile()) || 8468 TLI.isOperationLegalOrCustom(ISD::STORE, SVT))) 8469 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0), 8470 Ptr, ST->getPointerInfo(), ST->isVolatile(), 8471 ST->isNonTemporal(), OrigAlign); 8472 } 8473 8474 // Turn 'store undef, Ptr' -> nothing. 8475 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed()) 8476 return Chain; 8477 8478 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr' 8479 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) { 8480 // NOTE: If the original store is volatile, this transform must not increase 8481 // the number of stores. For example, on x86-32 an f64 can be stored in one 8482 // processor operation but an i64 (which is not legal) requires two. So the 8483 // transform should not be done in this case. 8484 if (Value.getOpcode() != ISD::TargetConstantFP) { 8485 SDValue Tmp; 8486 switch (CFP->getSimpleValueType(0).SimpleTy) { 8487 default: llvm_unreachable("Unknown FP type"); 8488 case MVT::f16: // We don't do this for these yet. 8489 case MVT::f80: 8490 case MVT::f128: 8491 case MVT::ppcf128: 8492 break; 8493 case MVT::f32: 8494 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) || 8495 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) { 8496 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF(). 8497 bitcastToAPInt().getZExtValue(), MVT::i32); 8498 return DAG.getStore(Chain, SDLoc(N), Tmp, 8499 Ptr, ST->getPointerInfo(), ST->isVolatile(), 8500 ST->isNonTemporal(), ST->getAlignment()); 8501 } 8502 break; 8503 case MVT::f64: 8504 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations && 8505 !ST->isVolatile()) || 8506 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) { 8507 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt(). 8508 getZExtValue(), MVT::i64); 8509 return DAG.getStore(Chain, SDLoc(N), Tmp, 8510 Ptr, ST->getPointerInfo(), ST->isVolatile(), 8511 ST->isNonTemporal(), ST->getAlignment()); 8512 } 8513 8514 if (!ST->isVolatile() && 8515 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) { 8516 // Many FP stores are not made apparent until after legalize, e.g. for 8517 // argument passing. Since this is so common, custom legalize the 8518 // 64-bit integer store into two 32-bit stores. 8519 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue(); 8520 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32); 8521 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32); 8522 if (TLI.isBigEndian()) std::swap(Lo, Hi); 8523 8524 unsigned Alignment = ST->getAlignment(); 8525 bool isVolatile = ST->isVolatile(); 8526 bool isNonTemporal = ST->isNonTemporal(); 8527 8528 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo, 8529 Ptr, ST->getPointerInfo(), 8530 isVolatile, isNonTemporal, 8531 ST->getAlignment()); 8532 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr, 8533 DAG.getConstant(4, Ptr.getValueType())); 8534 Alignment = MinAlign(Alignment, 4U); 8535 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi, 8536 Ptr, ST->getPointerInfo().getWithOffset(4), 8537 isVolatile, isNonTemporal, 8538 Alignment); 8539 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, 8540 St0, St1); 8541 } 8542 8543 break; 8544 } 8545 } 8546 } 8547 8548 // Try to infer better alignment information than the store already has. 8549 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) { 8550 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) { 8551 if (Align > ST->getAlignment()) 8552 return DAG.getTruncStore(Chain, SDLoc(N), Value, 8553 Ptr, ST->getPointerInfo(), ST->getMemoryVT(), 8554 ST->isVolatile(), ST->isNonTemporal(), Align); 8555 } 8556 } 8557 8558 // Try transforming a pair floating point load / store ops to integer 8559 // load / store ops. 8560 SDValue NewST = TransformFPLoadStorePair(N); 8561 if (NewST.getNode()) 8562 return NewST; 8563 8564 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA : 8565 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA(); 8566 if (UseAA) { 8567 // Walk up chain skipping non-aliasing memory nodes. 8568 SDValue BetterChain = FindBetterChain(N, Chain); 8569 8570 // If there is a better chain. 8571 if (Chain != BetterChain) { 8572 SDValue ReplStore; 8573 8574 // Replace the chain to avoid dependency. 8575 if (ST->isTruncatingStore()) { 8576 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr, 8577 ST->getPointerInfo(), 8578 ST->getMemoryVT(), ST->isVolatile(), 8579 ST->isNonTemporal(), ST->getAlignment()); 8580 } else { 8581 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr, 8582 ST->getPointerInfo(), 8583 ST->isVolatile(), ST->isNonTemporal(), 8584 ST->getAlignment()); 8585 } 8586 8587 // Create token to keep both nodes around. 8588 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N), 8589 MVT::Other, Chain, ReplStore); 8590 8591 // Make sure the new and old chains are cleaned up. 8592 AddToWorkList(Token.getNode()); 8593 8594 // Don't add users to work list. 8595 return CombineTo(N, Token, false); 8596 } 8597 } 8598 8599 // Try transforming N to an indexed store. 8600 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) 8601 return SDValue(N, 0); 8602 8603 // FIXME: is there such a thing as a truncating indexed store? 8604 if (ST->isTruncatingStore() && ST->isUnindexed() && 8605 Value.getValueType().isInteger()) { 8606 // See if we can simplify the input to this truncstore with knowledge that 8607 // only the low bits are being used. For example: 8608 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8" 8609 SDValue Shorter = 8610 GetDemandedBits(Value, 8611 APInt::getLowBitsSet( 8612 Value.getValueType().getScalarType().getSizeInBits(), 8613 ST->getMemoryVT().getScalarType().getSizeInBits())); 8614 AddToWorkList(Value.getNode()); 8615 if (Shorter.getNode()) 8616 return DAG.getTruncStore(Chain, SDLoc(N), Shorter, 8617 Ptr, ST->getPointerInfo(), ST->getMemoryVT(), 8618 ST->isVolatile(), ST->isNonTemporal(), 8619 ST->getAlignment()); 8620 8621 // Otherwise, see if we can simplify the operation with 8622 // SimplifyDemandedBits, which only works if the value has a single use. 8623 if (SimplifyDemandedBits(Value, 8624 APInt::getLowBitsSet( 8625 Value.getValueType().getScalarType().getSizeInBits(), 8626 ST->getMemoryVT().getScalarType().getSizeInBits()))) 8627 return SDValue(N, 0); 8628 } 8629 8630 // If this is a load followed by a store to the same location, then the store 8631 // is dead/noop. 8632 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) { 8633 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() && 8634 ST->isUnindexed() && !ST->isVolatile() && 8635 // There can't be any side effects between the load and store, such as 8636 // a call or store. 8637 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) { 8638 // The store is dead, remove it. 8639 return Chain; 8640 } 8641 } 8642 8643 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a 8644 // truncating store. We can do this even if this is already a truncstore. 8645 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE) 8646 && Value.getNode()->hasOneUse() && ST->isUnindexed() && 8647 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(), 8648 ST->getMemoryVT())) { 8649 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0), 8650 Ptr, ST->getPointerInfo(), ST->getMemoryVT(), 8651 ST->isVolatile(), ST->isNonTemporal(), 8652 ST->getAlignment()); 8653 } 8654 8655 // Only perform this optimization before the types are legal, because we 8656 // don't want to perform this optimization on every DAGCombine invocation. 8657 if (!LegalTypes) { 8658 bool EverChanged = false; 8659 8660 do { 8661 // There can be multiple store sequences on the same chain. 8662 // Keep trying to merge store sequences until we are unable to do so 8663 // or until we merge the last store on the chain. 8664 bool Changed = MergeConsecutiveStores(ST); 8665 EverChanged |= Changed; 8666 if (!Changed) break; 8667 } while (ST->getOpcode() != ISD::DELETED_NODE); 8668 8669 if (EverChanged) 8670 return SDValue(N, 0); 8671 } 8672 8673 return ReduceLoadOpStoreWidth(N); 8674} 8675 8676SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) { 8677 SDValue InVec = N->getOperand(0); 8678 SDValue InVal = N->getOperand(1); 8679 SDValue EltNo = N->getOperand(2); 8680 SDLoc dl(N); 8681 8682 // If the inserted element is an UNDEF, just use the input vector. 8683 if (InVal.getOpcode() == ISD::UNDEF) 8684 return InVec; 8685 8686 EVT VT = InVec.getValueType(); 8687 8688 // If we can't generate a legal BUILD_VECTOR, exit 8689 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) 8690 return SDValue(); 8691 8692 // Check that we know which element is being inserted 8693 if (!isa<ConstantSDNode>(EltNo)) 8694 return SDValue(); 8695 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); 8696 8697 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially 8698 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the 8699 // vector elements. 8700 SmallVector<SDValue, 8> Ops; 8701 // Do not combine these two vectors if the output vector will not replace 8702 // the input vector. 8703 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) { 8704 Ops.append(InVec.getNode()->op_begin(), 8705 InVec.getNode()->op_end()); 8706 } else if (InVec.getOpcode() == ISD::UNDEF) { 8707 unsigned NElts = VT.getVectorNumElements(); 8708 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType())); 8709 } else { 8710 return SDValue(); 8711 } 8712 8713 // Insert the element 8714 if (Elt < Ops.size()) { 8715 // All the operands of BUILD_VECTOR must have the same type; 8716 // we enforce that here. 8717 EVT OpVT = Ops[0].getValueType(); 8718 if (InVal.getValueType() != OpVT) 8719 InVal = OpVT.bitsGT(InVal.getValueType()) ? 8720 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) : 8721 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal); 8722 Ops[Elt] = InVal; 8723 } 8724 8725 // Return the new vector 8726 return DAG.getNode(ISD::BUILD_VECTOR, dl, 8727 VT, &Ops[0], Ops.size()); 8728} 8729 8730SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) { 8731 // (vextract (scalar_to_vector val, 0) -> val 8732 SDValue InVec = N->getOperand(0); 8733 EVT VT = InVec.getValueType(); 8734 EVT NVT = N->getValueType(0); 8735 8736 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) { 8737 // Check if the result type doesn't match the inserted element type. A 8738 // SCALAR_TO_VECTOR may truncate the inserted element and the 8739 // EXTRACT_VECTOR_ELT may widen the extracted vector. 8740 SDValue InOp = InVec.getOperand(0); 8741 if (InOp.getValueType() != NVT) { 8742 assert(InOp.getValueType().isInteger() && NVT.isInteger()); 8743 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT); 8744 } 8745 return InOp; 8746 } 8747 8748 SDValue EltNo = N->getOperand(1); 8749 bool ConstEltNo = isa<ConstantSDNode>(EltNo); 8750 8751 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT. 8752 // We only perform this optimization before the op legalization phase because 8753 // we may introduce new vector instructions which are not backed by TD 8754 // patterns. For example on AVX, extracting elements from a wide vector 8755 // without using extract_subvector. 8756 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE 8757 && ConstEltNo && !LegalOperations) { 8758 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); 8759 int NumElem = VT.getVectorNumElements(); 8760 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec); 8761 // Find the new index to extract from. 8762 int OrigElt = SVOp->getMaskElt(Elt); 8763 8764 // Extracting an undef index is undef. 8765 if (OrigElt == -1) 8766 return DAG.getUNDEF(NVT); 8767 8768 // Select the right vector half to extract from. 8769 if (OrigElt < NumElem) { 8770 InVec = InVec->getOperand(0); 8771 } else { 8772 InVec = InVec->getOperand(1); 8773 OrigElt -= NumElem; 8774 } 8775 8776 EVT IndexTy = TLI.getVectorIdxTy(); 8777 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT, 8778 InVec, DAG.getConstant(OrigElt, IndexTy)); 8779 } 8780 8781 // Perform only after legalization to ensure build_vector / vector_shuffle 8782 // optimizations have already been done. 8783 if (!LegalOperations) return SDValue(); 8784 8785 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size) 8786 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size) 8787 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr) 8788 8789 if (ConstEltNo) { 8790 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue(); 8791 bool NewLoad = false; 8792 bool BCNumEltsChanged = false; 8793 EVT ExtVT = VT.getVectorElementType(); 8794 EVT LVT = ExtVT; 8795 8796 // If the result of load has to be truncated, then it's not necessarily 8797 // profitable. 8798 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT)) 8799 return SDValue(); 8800 8801 if (InVec.getOpcode() == ISD::BITCAST) { 8802 // Don't duplicate a load with other uses. 8803 if (!InVec.hasOneUse()) 8804 return SDValue(); 8805 8806 EVT BCVT = InVec.getOperand(0).getValueType(); 8807 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType())) 8808 return SDValue(); 8809 if (VT.getVectorNumElements() != BCVT.getVectorNumElements()) 8810 BCNumEltsChanged = true; 8811 InVec = InVec.getOperand(0); 8812 ExtVT = BCVT.getVectorElementType(); 8813 NewLoad = true; 8814 } 8815 8816 LoadSDNode *LN0 = NULL; 8817 const ShuffleVectorSDNode *SVN = NULL; 8818 if (ISD::isNormalLoad(InVec.getNode())) { 8819 LN0 = cast<LoadSDNode>(InVec); 8820 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR && 8821 InVec.getOperand(0).getValueType() == ExtVT && 8822 ISD::isNormalLoad(InVec.getOperand(0).getNode())) { 8823 // Don't duplicate a load with other uses. 8824 if (!InVec.hasOneUse()) 8825 return SDValue(); 8826 8827 LN0 = cast<LoadSDNode>(InVec.getOperand(0)); 8828 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) { 8829 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1) 8830 // => 8831 // (load $addr+1*size) 8832 8833 // Don't duplicate a load with other uses. 8834 if (!InVec.hasOneUse()) 8835 return SDValue(); 8836 8837 // If the bit convert changed the number of elements, it is unsafe 8838 // to examine the mask. 8839 if (BCNumEltsChanged) 8840 return SDValue(); 8841 8842 // Select the input vector, guarding against out of range extract vector. 8843 unsigned NumElems = VT.getVectorNumElements(); 8844 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt); 8845 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1); 8846 8847 if (InVec.getOpcode() == ISD::BITCAST) { 8848 // Don't duplicate a load with other uses. 8849 if (!InVec.hasOneUse()) 8850 return SDValue(); 8851 8852 InVec = InVec.getOperand(0); 8853 } 8854 if (ISD::isNormalLoad(InVec.getNode())) { 8855 LN0 = cast<LoadSDNode>(InVec); 8856 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems; 8857 } 8858 } 8859 8860 // Make sure we found a non-volatile load and the extractelement is 8861 // the only use. 8862 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile()) 8863 return SDValue(); 8864 8865 // If Idx was -1 above, Elt is going to be -1, so just return undef. 8866 if (Elt == -1) 8867 return DAG.getUNDEF(LVT); 8868 8869 unsigned Align = LN0->getAlignment(); 8870 if (NewLoad) { 8871 // Check the resultant load doesn't need a higher alignment than the 8872 // original load. 8873 unsigned NewAlign = 8874 TLI.getDataLayout() 8875 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext())); 8876 8877 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT)) 8878 return SDValue(); 8879 8880 Align = NewAlign; 8881 } 8882 8883 SDValue NewPtr = LN0->getBasePtr(); 8884 unsigned PtrOff = 0; 8885 8886 if (Elt) { 8887 PtrOff = LVT.getSizeInBits() * Elt / 8; 8888 EVT PtrType = NewPtr.getValueType(); 8889 if (TLI.isBigEndian()) 8890 PtrOff = VT.getSizeInBits() / 8 - PtrOff; 8891 NewPtr = DAG.getNode(ISD::ADD, SDLoc(N), PtrType, NewPtr, 8892 DAG.getConstant(PtrOff, PtrType)); 8893 } 8894 8895 // The replacement we need to do here is a little tricky: we need to 8896 // replace an extractelement of a load with a load. 8897 // Use ReplaceAllUsesOfValuesWith to do the replacement. 8898 // Note that this replacement assumes that the extractvalue is the only 8899 // use of the load; that's okay because we don't want to perform this 8900 // transformation in other cases anyway. 8901 SDValue Load; 8902 SDValue Chain; 8903 if (NVT.bitsGT(LVT)) { 8904 // If the result type of vextract is wider than the load, then issue an 8905 // extending load instead. 8906 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT) 8907 ? ISD::ZEXTLOAD : ISD::EXTLOAD; 8908 Load = DAG.getExtLoad(ExtType, SDLoc(N), NVT, LN0->getChain(), 8909 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff), 8910 LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align); 8911 Chain = Load.getValue(1); 8912 } else { 8913 Load = DAG.getLoad(LVT, SDLoc(N), LN0->getChain(), NewPtr, 8914 LN0->getPointerInfo().getWithOffset(PtrOff), 8915 LN0->isVolatile(), LN0->isNonTemporal(), 8916 LN0->isInvariant(), Align); 8917 Chain = Load.getValue(1); 8918 if (NVT.bitsLT(LVT)) 8919 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, Load); 8920 else 8921 Load = DAG.getNode(ISD::BITCAST, SDLoc(N), NVT, Load); 8922 } 8923 WorkListRemover DeadNodes(*this); 8924 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) }; 8925 SDValue To[] = { Load, Chain }; 8926 DAG.ReplaceAllUsesOfValuesWith(From, To, 2); 8927 // Since we're explcitly calling ReplaceAllUses, add the new node to the 8928 // worklist explicitly as well. 8929 AddToWorkList(Load.getNode()); 8930 AddUsersToWorkList(Load.getNode()); // Add users too 8931 // Make sure to revisit this node to clean it up; it will usually be dead. 8932 AddToWorkList(N); 8933 return SDValue(N, 0); 8934 } 8935 8936 return SDValue(); 8937} 8938 8939// Simplify (build_vec (ext )) to (bitcast (build_vec )) 8940SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) { 8941 // We perform this optimization post type-legalization because 8942 // the type-legalizer often scalarizes integer-promoted vectors. 8943 // Performing this optimization before may create bit-casts which 8944 // will be type-legalized to complex code sequences. 8945 // We perform this optimization only before the operation legalizer because we 8946 // may introduce illegal operations. 8947 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes) 8948 return SDValue(); 8949 8950 unsigned NumInScalars = N->getNumOperands(); 8951 SDLoc dl(N); 8952 EVT VT = N->getValueType(0); 8953 8954 // Check to see if this is a BUILD_VECTOR of a bunch of values 8955 // which come from any_extend or zero_extend nodes. If so, we can create 8956 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR 8957 // optimizations. We do not handle sign-extend because we can't fill the sign 8958 // using shuffles. 8959 EVT SourceType = MVT::Other; 8960 bool AllAnyExt = true; 8961 8962 for (unsigned i = 0; i != NumInScalars; ++i) { 8963 SDValue In = N->getOperand(i); 8964 // Ignore undef inputs. 8965 if (In.getOpcode() == ISD::UNDEF) continue; 8966 8967 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND; 8968 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND; 8969 8970 // Abort if the element is not an extension. 8971 if (!ZeroExt && !AnyExt) { 8972 SourceType = MVT::Other; 8973 break; 8974 } 8975 8976 // The input is a ZeroExt or AnyExt. Check the original type. 8977 EVT InTy = In.getOperand(0).getValueType(); 8978 8979 // Check that all of the widened source types are the same. 8980 if (SourceType == MVT::Other) 8981 // First time. 8982 SourceType = InTy; 8983 else if (InTy != SourceType) { 8984 // Multiple income types. Abort. 8985 SourceType = MVT::Other; 8986 break; 8987 } 8988 8989 // Check if all of the extends are ANY_EXTENDs. 8990 AllAnyExt &= AnyExt; 8991 } 8992 8993 // In order to have valid types, all of the inputs must be extended from the 8994 // same source type and all of the inputs must be any or zero extend. 8995 // Scalar sizes must be a power of two. 8996 EVT OutScalarTy = VT.getScalarType(); 8997 bool ValidTypes = SourceType != MVT::Other && 8998 isPowerOf2_32(OutScalarTy.getSizeInBits()) && 8999 isPowerOf2_32(SourceType.getSizeInBits()); 9000 9001 // Create a new simpler BUILD_VECTOR sequence which other optimizations can 9002 // turn into a single shuffle instruction. 9003 if (!ValidTypes) 9004 return SDValue(); 9005 9006 bool isLE = TLI.isLittleEndian(); 9007 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits(); 9008 assert(ElemRatio > 1 && "Invalid element size ratio"); 9009 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType): 9010 DAG.getConstant(0, SourceType); 9011 9012 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements(); 9013 SmallVector<SDValue, 8> Ops(NewBVElems, Filler); 9014 9015 // Populate the new build_vector 9016 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 9017 SDValue Cast = N->getOperand(i); 9018 assert((Cast.getOpcode() == ISD::ANY_EXTEND || 9019 Cast.getOpcode() == ISD::ZERO_EXTEND || 9020 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode"); 9021 SDValue In; 9022 if (Cast.getOpcode() == ISD::UNDEF) 9023 In = DAG.getUNDEF(SourceType); 9024 else 9025 In = Cast->getOperand(0); 9026 unsigned Index = isLE ? (i * ElemRatio) : 9027 (i * ElemRatio + (ElemRatio - 1)); 9028 9029 assert(Index < Ops.size() && "Invalid index"); 9030 Ops[Index] = In; 9031 } 9032 9033 // The type of the new BUILD_VECTOR node. 9034 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems); 9035 assert(VecVT.getSizeInBits() == VT.getSizeInBits() && 9036 "Invalid vector size"); 9037 // Check if the new vector type is legal. 9038 if (!isTypeLegal(VecVT)) return SDValue(); 9039 9040 // Make the new BUILD_VECTOR. 9041 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], Ops.size()); 9042 9043 // The new BUILD_VECTOR node has the potential to be further optimized. 9044 AddToWorkList(BV.getNode()); 9045 // Bitcast to the desired type. 9046 return DAG.getNode(ISD::BITCAST, dl, VT, BV); 9047} 9048 9049SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) { 9050 EVT VT = N->getValueType(0); 9051 9052 unsigned NumInScalars = N->getNumOperands(); 9053 SDLoc dl(N); 9054 9055 EVT SrcVT = MVT::Other; 9056 unsigned Opcode = ISD::DELETED_NODE; 9057 unsigned NumDefs = 0; 9058 9059 for (unsigned i = 0; i != NumInScalars; ++i) { 9060 SDValue In = N->getOperand(i); 9061 unsigned Opc = In.getOpcode(); 9062 9063 if (Opc == ISD::UNDEF) 9064 continue; 9065 9066 // If all scalar values are floats and converted from integers. 9067 if (Opcode == ISD::DELETED_NODE && 9068 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) { 9069 Opcode = Opc; 9070 } 9071 9072 if (Opc != Opcode) 9073 return SDValue(); 9074 9075 EVT InVT = In.getOperand(0).getValueType(); 9076 9077 // If all scalar values are typed differently, bail out. It's chosen to 9078 // simplify BUILD_VECTOR of integer types. 9079 if (SrcVT == MVT::Other) 9080 SrcVT = InVT; 9081 if (SrcVT != InVT) 9082 return SDValue(); 9083 NumDefs++; 9084 } 9085 9086 // If the vector has just one element defined, it's not worth to fold it into 9087 // a vectorized one. 9088 if (NumDefs < 2) 9089 return SDValue(); 9090 9091 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP) 9092 && "Should only handle conversion from integer to float."); 9093 assert(SrcVT != MVT::Other && "Cannot determine source type!"); 9094 9095 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars); 9096 9097 if (!TLI.isOperationLegalOrCustom(Opcode, NVT)) 9098 return SDValue(); 9099 9100 SmallVector<SDValue, 8> Opnds; 9101 for (unsigned i = 0; i != NumInScalars; ++i) { 9102 SDValue In = N->getOperand(i); 9103 9104 if (In.getOpcode() == ISD::UNDEF) 9105 Opnds.push_back(DAG.getUNDEF(SrcVT)); 9106 else 9107 Opnds.push_back(In.getOperand(0)); 9108 } 9109 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, 9110 &Opnds[0], Opnds.size()); 9111 AddToWorkList(BV.getNode()); 9112 9113 return DAG.getNode(Opcode, dl, VT, BV); 9114} 9115 9116SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) { 9117 unsigned NumInScalars = N->getNumOperands(); 9118 SDLoc dl(N); 9119 EVT VT = N->getValueType(0); 9120 9121 // A vector built entirely of undefs is undef. 9122 if (ISD::allOperandsUndef(N)) 9123 return DAG.getUNDEF(VT); 9124 9125 SDValue V = reduceBuildVecExtToExtBuildVec(N); 9126 if (V.getNode()) 9127 return V; 9128 9129 V = reduceBuildVecConvertToConvertBuildVec(N); 9130 if (V.getNode()) 9131 return V; 9132 9133 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT 9134 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from 9135 // at most two distinct vectors, turn this into a shuffle node. 9136 9137 // May only combine to shuffle after legalize if shuffle is legal. 9138 if (LegalOperations && 9139 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT)) 9140 return SDValue(); 9141 9142 SDValue VecIn1, VecIn2; 9143 for (unsigned i = 0; i != NumInScalars; ++i) { 9144 // Ignore undef inputs. 9145 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue; 9146 9147 // If this input is something other than a EXTRACT_VECTOR_ELT with a 9148 // constant index, bail out. 9149 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT || 9150 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) { 9151 VecIn1 = VecIn2 = SDValue(0, 0); 9152 break; 9153 } 9154 9155 // We allow up to two distinct input vectors. 9156 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0); 9157 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2) 9158 continue; 9159 9160 if (VecIn1.getNode() == 0) { 9161 VecIn1 = ExtractedFromVec; 9162 } else if (VecIn2.getNode() == 0) { 9163 VecIn2 = ExtractedFromVec; 9164 } else { 9165 // Too many inputs. 9166 VecIn1 = VecIn2 = SDValue(0, 0); 9167 break; 9168 } 9169 } 9170 9171 // If everything is good, we can make a shuffle operation. 9172 if (VecIn1.getNode()) { 9173 SmallVector<int, 8> Mask; 9174 for (unsigned i = 0; i != NumInScalars; ++i) { 9175 if (N->getOperand(i).getOpcode() == ISD::UNDEF) { 9176 Mask.push_back(-1); 9177 continue; 9178 } 9179 9180 // If extracting from the first vector, just use the index directly. 9181 SDValue Extract = N->getOperand(i); 9182 SDValue ExtVal = Extract.getOperand(1); 9183 if (Extract.getOperand(0) == VecIn1) { 9184 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue(); 9185 if (ExtIndex > VT.getVectorNumElements()) 9186 return SDValue(); 9187 9188 Mask.push_back(ExtIndex); 9189 continue; 9190 } 9191 9192 // Otherwise, use InIdx + VecSize 9193 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue(); 9194 Mask.push_back(Idx+NumInScalars); 9195 } 9196 9197 // We can't generate a shuffle node with mismatched input and output types. 9198 // Attempt to transform a single input vector to the correct type. 9199 if ((VT != VecIn1.getValueType())) { 9200 // We don't support shuffeling between TWO values of different types. 9201 if (VecIn2.getNode() != 0) 9202 return SDValue(); 9203 9204 // We only support widening of vectors which are half the size of the 9205 // output registers. For example XMM->YMM widening on X86 with AVX. 9206 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits()) 9207 return SDValue(); 9208 9209 // If the input vector type has a different base type to the output 9210 // vector type, bail out. 9211 if (VecIn1.getValueType().getVectorElementType() != 9212 VT.getVectorElementType()) 9213 return SDValue(); 9214 9215 // Widen the input vector by adding undef values. 9216 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, 9217 VecIn1, DAG.getUNDEF(VecIn1.getValueType())); 9218 } 9219 9220 // If VecIn2 is unused then change it to undef. 9221 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT); 9222 9223 // Check that we were able to transform all incoming values to the same 9224 // type. 9225 if (VecIn2.getValueType() != VecIn1.getValueType() || 9226 VecIn1.getValueType() != VT) 9227 return SDValue(); 9228 9229 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes. 9230 if (!isTypeLegal(VT)) 9231 return SDValue(); 9232 9233 // Return the new VECTOR_SHUFFLE node. 9234 SDValue Ops[2]; 9235 Ops[0] = VecIn1; 9236 Ops[1] = VecIn2; 9237 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]); 9238 } 9239 9240 return SDValue(); 9241} 9242 9243SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) { 9244 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of 9245 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector 9246 // inputs come from at most two distinct vectors, turn this into a shuffle 9247 // node. 9248 9249 // If we only have one input vector, we don't need to do any concatenation. 9250 if (N->getNumOperands() == 1) 9251 return N->getOperand(0); 9252 9253 // Check if all of the operands are undefs. 9254 if (ISD::allOperandsUndef(N)) 9255 return DAG.getUNDEF(N->getValueType(0)); 9256 9257 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR 9258 // nodes often generate nop CONCAT_VECTOR nodes. 9259 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that 9260 // place the incoming vectors at the exact same location. 9261 SDValue SingleSource = SDValue(); 9262 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements(); 9263 9264 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 9265 SDValue Op = N->getOperand(i); 9266 9267 if (Op.getOpcode() == ISD::UNDEF) 9268 continue; 9269 9270 // Check if this is the identity extract: 9271 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR) 9272 return SDValue(); 9273 9274 // Find the single incoming vector for the extract_subvector. 9275 if (SingleSource.getNode()) { 9276 if (Op.getOperand(0) != SingleSource) 9277 return SDValue(); 9278 } else { 9279 SingleSource = Op.getOperand(0); 9280 9281 // Check the source type is the same as the type of the result. 9282 // If not, this concat may extend the vector, so we can not 9283 // optimize it away. 9284 if (SingleSource.getValueType() != N->getValueType(0)) 9285 return SDValue(); 9286 } 9287 9288 unsigned IdentityIndex = i * PartNumElem; 9289 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 9290 // The extract index must be constant. 9291 if (!CS) 9292 return SDValue(); 9293 9294 // Check that we are reading from the identity index. 9295 if (CS->getZExtValue() != IdentityIndex) 9296 return SDValue(); 9297 } 9298 9299 if (SingleSource.getNode()) 9300 return SingleSource; 9301 9302 return SDValue(); 9303} 9304 9305SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) { 9306 EVT NVT = N->getValueType(0); 9307 SDValue V = N->getOperand(0); 9308 9309 if (V->getOpcode() == ISD::CONCAT_VECTORS) { 9310 // Combine: 9311 // (extract_subvec (concat V1, V2, ...), i) 9312 // Into: 9313 // Vi if possible 9314 // Only operand 0 is checked as 'concat' assumes all inputs of the same type. 9315 if (V->getOperand(0).getValueType() != NVT) 9316 return SDValue(); 9317 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue(); 9318 unsigned NumElems = NVT.getVectorNumElements(); 9319 assert((Idx % NumElems) == 0 && 9320 "IDX in concat is not a multiple of the result vector length."); 9321 return V->getOperand(Idx / NumElems); 9322 } 9323 9324 // Skip bitcasting 9325 if (V->getOpcode() == ISD::BITCAST) 9326 V = V.getOperand(0); 9327 9328 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) { 9329 SDLoc dl(N); 9330 // Handle only simple case where vector being inserted and vector 9331 // being extracted are of same type, and are half size of larger vectors. 9332 EVT BigVT = V->getOperand(0).getValueType(); 9333 EVT SmallVT = V->getOperand(1).getValueType(); 9334 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits()) 9335 return SDValue(); 9336 9337 // Only handle cases where both indexes are constants with the same type. 9338 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1)); 9339 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2)); 9340 9341 if (InsIdx && ExtIdx && 9342 InsIdx->getValueType(0).getSizeInBits() <= 64 && 9343 ExtIdx->getValueType(0).getSizeInBits() <= 64) { 9344 // Combine: 9345 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx) 9346 // Into: 9347 // indices are equal or bit offsets are equal => V1 9348 // otherwise => (extract_subvec V1, ExtIdx) 9349 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() == 9350 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits()) 9351 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1)); 9352 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT, 9353 DAG.getNode(ISD::BITCAST, dl, 9354 N->getOperand(0).getValueType(), 9355 V->getOperand(0)), N->getOperand(1)); 9356 } 9357 } 9358 9359 return SDValue(); 9360} 9361 9362// Tries to turn a shuffle of two CONCAT_VECTORS into a single concat. 9363static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) { 9364 EVT VT = N->getValueType(0); 9365 unsigned NumElts = VT.getVectorNumElements(); 9366 9367 SDValue N0 = N->getOperand(0); 9368 SDValue N1 = N->getOperand(1); 9369 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 9370 9371 SmallVector<SDValue, 4> Ops; 9372 EVT ConcatVT = N0.getOperand(0).getValueType(); 9373 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements(); 9374 unsigned NumConcats = NumElts / NumElemsPerConcat; 9375 9376 // Look at every vector that's inserted. We're looking for exact 9377 // subvector-sized copies from a concatenated vector 9378 for (unsigned I = 0; I != NumConcats; ++I) { 9379 // Make sure we're dealing with a copy. 9380 unsigned Begin = I * NumElemsPerConcat; 9381 bool AllUndef = true, NoUndef = true; 9382 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) { 9383 if (SVN->getMaskElt(J) >= 0) 9384 AllUndef = false; 9385 else 9386 NoUndef = false; 9387 } 9388 9389 if (NoUndef) { 9390 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0) 9391 return SDValue(); 9392 9393 for (unsigned J = 1; J != NumElemsPerConcat; ++J) 9394 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J)) 9395 return SDValue(); 9396 9397 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat; 9398 if (FirstElt < N0.getNumOperands()) 9399 Ops.push_back(N0.getOperand(FirstElt)); 9400 else 9401 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands())); 9402 9403 } else if (AllUndef) { 9404 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType())); 9405 } else { // Mixed with general masks and undefs, can't do optimization. 9406 return SDValue(); 9407 } 9408 } 9409 9410 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops.data(), 9411 Ops.size()); 9412} 9413 9414SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) { 9415 EVT VT = N->getValueType(0); 9416 unsigned NumElts = VT.getVectorNumElements(); 9417 9418 SDValue N0 = N->getOperand(0); 9419 SDValue N1 = N->getOperand(1); 9420 9421 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG"); 9422 9423 // Canonicalize shuffle undef, undef -> undef 9424 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF) 9425 return DAG.getUNDEF(VT); 9426 9427 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 9428 9429 // Canonicalize shuffle v, v -> v, undef 9430 if (N0 == N1) { 9431 SmallVector<int, 8> NewMask; 9432 for (unsigned i = 0; i != NumElts; ++i) { 9433 int Idx = SVN->getMaskElt(i); 9434 if (Idx >= (int)NumElts) Idx -= NumElts; 9435 NewMask.push_back(Idx); 9436 } 9437 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT), 9438 &NewMask[0]); 9439 } 9440 9441 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask. 9442 if (N0.getOpcode() == ISD::UNDEF) { 9443 SmallVector<int, 8> NewMask; 9444 for (unsigned i = 0; i != NumElts; ++i) { 9445 int Idx = SVN->getMaskElt(i); 9446 if (Idx >= 0) { 9447 if (Idx >= (int)NumElts) 9448 Idx -= NumElts; 9449 else 9450 Idx = -1; // remove reference to lhs 9451 } 9452 NewMask.push_back(Idx); 9453 } 9454 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT), 9455 &NewMask[0]); 9456 } 9457 9458 // Remove references to rhs if it is undef 9459 if (N1.getOpcode() == ISD::UNDEF) { 9460 bool Changed = false; 9461 SmallVector<int, 8> NewMask; 9462 for (unsigned i = 0; i != NumElts; ++i) { 9463 int Idx = SVN->getMaskElt(i); 9464 if (Idx >= (int)NumElts) { 9465 Idx = -1; 9466 Changed = true; 9467 } 9468 NewMask.push_back(Idx); 9469 } 9470 if (Changed) 9471 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]); 9472 } 9473 9474 // If it is a splat, check if the argument vector is another splat or a 9475 // build_vector with all scalar elements the same. 9476 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) { 9477 SDNode *V = N0.getNode(); 9478 9479 // If this is a bit convert that changes the element type of the vector but 9480 // not the number of vector elements, look through it. Be careful not to 9481 // look though conversions that change things like v4f32 to v2f64. 9482 if (V->getOpcode() == ISD::BITCAST) { 9483 SDValue ConvInput = V->getOperand(0); 9484 if (ConvInput.getValueType().isVector() && 9485 ConvInput.getValueType().getVectorNumElements() == NumElts) 9486 V = ConvInput.getNode(); 9487 } 9488 9489 if (V->getOpcode() == ISD::BUILD_VECTOR) { 9490 assert(V->getNumOperands() == NumElts && 9491 "BUILD_VECTOR has wrong number of operands"); 9492 SDValue Base; 9493 bool AllSame = true; 9494 for (unsigned i = 0; i != NumElts; ++i) { 9495 if (V->getOperand(i).getOpcode() != ISD::UNDEF) { 9496 Base = V->getOperand(i); 9497 break; 9498 } 9499 } 9500 // Splat of <u, u, u, u>, return <u, u, u, u> 9501 if (!Base.getNode()) 9502 return N0; 9503 for (unsigned i = 0; i != NumElts; ++i) { 9504 if (V->getOperand(i) != Base) { 9505 AllSame = false; 9506 break; 9507 } 9508 } 9509 // Splat of <x, x, x, x>, return <x, x, x, x> 9510 if (AllSame) 9511 return N0; 9512 } 9513 } 9514 9515 if (N0.getOpcode() == ISD::CONCAT_VECTORS && 9516 Level < AfterLegalizeVectorOps && 9517 (N1.getOpcode() == ISD::UNDEF || 9518 (N1.getOpcode() == ISD::CONCAT_VECTORS && 9519 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) { 9520 SDValue V = partitionShuffleOfConcats(N, DAG); 9521 9522 if (V.getNode()) 9523 return V; 9524 } 9525 9526 // If this shuffle node is simply a swizzle of another shuffle node, 9527 // and it reverses the swizzle of the previous shuffle then we can 9528 // optimize shuffle(shuffle(x, undef), undef) -> x. 9529 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG && 9530 N1.getOpcode() == ISD::UNDEF) { 9531 9532 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0); 9533 9534 // Shuffle nodes can only reverse shuffles with a single non-undef value. 9535 if (N0.getOperand(1).getOpcode() != ISD::UNDEF) 9536 return SDValue(); 9537 9538 // The incoming shuffle must be of the same type as the result of the 9539 // current shuffle. 9540 assert(OtherSV->getOperand(0).getValueType() == VT && 9541 "Shuffle types don't match"); 9542 9543 for (unsigned i = 0; i != NumElts; ++i) { 9544 int Idx = SVN->getMaskElt(i); 9545 assert(Idx < (int)NumElts && "Index references undef operand"); 9546 // Next, this index comes from the first value, which is the incoming 9547 // shuffle. Adopt the incoming index. 9548 if (Idx >= 0) 9549 Idx = OtherSV->getMaskElt(Idx); 9550 9551 // The combined shuffle must map each index to itself. 9552 if (Idx >= 0 && (unsigned)Idx != i) 9553 return SDValue(); 9554 } 9555 9556 return OtherSV->getOperand(0); 9557 } 9558 9559 return SDValue(); 9560} 9561 9562/// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform 9563/// an AND to a vector_shuffle with the destination vector and a zero vector. 9564/// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==> 9565/// vector_shuffle V, Zero, <0, 4, 2, 4> 9566SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) { 9567 EVT VT = N->getValueType(0); 9568 SDLoc dl(N); 9569 SDValue LHS = N->getOperand(0); 9570 SDValue RHS = N->getOperand(1); 9571 if (N->getOpcode() == ISD::AND) { 9572 if (RHS.getOpcode() == ISD::BITCAST) 9573 RHS = RHS.getOperand(0); 9574 if (RHS.getOpcode() == ISD::BUILD_VECTOR) { 9575 SmallVector<int, 8> Indices; 9576 unsigned NumElts = RHS.getNumOperands(); 9577 for (unsigned i = 0; i != NumElts; ++i) { 9578 SDValue Elt = RHS.getOperand(i); 9579 if (!isa<ConstantSDNode>(Elt)) 9580 return SDValue(); 9581 9582 if (cast<ConstantSDNode>(Elt)->isAllOnesValue()) 9583 Indices.push_back(i); 9584 else if (cast<ConstantSDNode>(Elt)->isNullValue()) 9585 Indices.push_back(NumElts); 9586 else 9587 return SDValue(); 9588 } 9589 9590 // Let's see if the target supports this vector_shuffle. 9591 EVT RVT = RHS.getValueType(); 9592 if (!TLI.isVectorClearMaskLegal(Indices, RVT)) 9593 return SDValue(); 9594 9595 // Return the new VECTOR_SHUFFLE node. 9596 EVT EltVT = RVT.getVectorElementType(); 9597 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(), 9598 DAG.getConstant(0, EltVT)); 9599 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), 9600 RVT, &ZeroOps[0], ZeroOps.size()); 9601 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS); 9602 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]); 9603 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf); 9604 } 9605 } 9606 9607 return SDValue(); 9608} 9609 9610/// SimplifyVBinOp - Visit a binary vector operation, like ADD. 9611SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) { 9612 assert(N->getValueType(0).isVector() && 9613 "SimplifyVBinOp only works on vectors!"); 9614 9615 SDValue LHS = N->getOperand(0); 9616 SDValue RHS = N->getOperand(1); 9617 SDValue Shuffle = XformToShuffleWithZero(N); 9618 if (Shuffle.getNode()) return Shuffle; 9619 9620 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold 9621 // this operation. 9622 if (LHS.getOpcode() == ISD::BUILD_VECTOR && 9623 RHS.getOpcode() == ISD::BUILD_VECTOR) { 9624 SmallVector<SDValue, 8> Ops; 9625 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) { 9626 SDValue LHSOp = LHS.getOperand(i); 9627 SDValue RHSOp = RHS.getOperand(i); 9628 // If these two elements can't be folded, bail out. 9629 if ((LHSOp.getOpcode() != ISD::UNDEF && 9630 LHSOp.getOpcode() != ISD::Constant && 9631 LHSOp.getOpcode() != ISD::ConstantFP) || 9632 (RHSOp.getOpcode() != ISD::UNDEF && 9633 RHSOp.getOpcode() != ISD::Constant && 9634 RHSOp.getOpcode() != ISD::ConstantFP)) 9635 break; 9636 9637 // Can't fold divide by zero. 9638 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV || 9639 N->getOpcode() == ISD::FDIV) { 9640 if ((RHSOp.getOpcode() == ISD::Constant && 9641 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) || 9642 (RHSOp.getOpcode() == ISD::ConstantFP && 9643 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero())) 9644 break; 9645 } 9646 9647 EVT VT = LHSOp.getValueType(); 9648 EVT RVT = RHSOp.getValueType(); 9649 if (RVT != VT) { 9650 // Integer BUILD_VECTOR operands may have types larger than the element 9651 // size (e.g., when the element type is not legal). Prior to type 9652 // legalization, the types may not match between the two BUILD_VECTORS. 9653 // Truncate one of the operands to make them match. 9654 if (RVT.getSizeInBits() > VT.getSizeInBits()) { 9655 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp); 9656 } else { 9657 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp); 9658 VT = RVT; 9659 } 9660 } 9661 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT, 9662 LHSOp, RHSOp); 9663 if (FoldOp.getOpcode() != ISD::UNDEF && 9664 FoldOp.getOpcode() != ISD::Constant && 9665 FoldOp.getOpcode() != ISD::ConstantFP) 9666 break; 9667 Ops.push_back(FoldOp); 9668 AddToWorkList(FoldOp.getNode()); 9669 } 9670 9671 if (Ops.size() == LHS.getNumOperands()) 9672 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), 9673 LHS.getValueType(), &Ops[0], Ops.size()); 9674 } 9675 9676 return SDValue(); 9677} 9678 9679/// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG. 9680SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) { 9681 assert(N->getValueType(0).isVector() && 9682 "SimplifyVUnaryOp only works on vectors!"); 9683 9684 SDValue N0 = N->getOperand(0); 9685 9686 if (N0.getOpcode() != ISD::BUILD_VECTOR) 9687 return SDValue(); 9688 9689 // Operand is a BUILD_VECTOR node, see if we can constant fold it. 9690 SmallVector<SDValue, 8> Ops; 9691 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) { 9692 SDValue Op = N0.getOperand(i); 9693 if (Op.getOpcode() != ISD::UNDEF && 9694 Op.getOpcode() != ISD::ConstantFP) 9695 break; 9696 EVT EltVT = Op.getValueType(); 9697 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op); 9698 if (FoldOp.getOpcode() != ISD::UNDEF && 9699 FoldOp.getOpcode() != ISD::ConstantFP) 9700 break; 9701 Ops.push_back(FoldOp); 9702 AddToWorkList(FoldOp.getNode()); 9703 } 9704 9705 if (Ops.size() != N0.getNumOperands()) 9706 return SDValue(); 9707 9708 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), 9709 N0.getValueType(), &Ops[0], Ops.size()); 9710} 9711 9712SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0, 9713 SDValue N1, SDValue N2){ 9714 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!"); 9715 9716 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2, 9717 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 9718 9719 // If we got a simplified select_cc node back from SimplifySelectCC, then 9720 // break it down into a new SETCC node, and a new SELECT node, and then return 9721 // the SELECT node, since we were called with a SELECT node. 9722 if (SCC.getNode()) { 9723 // Check to see if we got a select_cc back (to turn into setcc/select). 9724 // Otherwise, just return whatever node we got back, like fabs. 9725 if (SCC.getOpcode() == ISD::SELECT_CC) { 9726 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0), 9727 N0.getValueType(), 9728 SCC.getOperand(0), SCC.getOperand(1), 9729 SCC.getOperand(4)); 9730 AddToWorkList(SETCC.getNode()); 9731 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), 9732 SCC.getOperand(2), SCC.getOperand(3), SETCC); 9733 } 9734 9735 return SCC; 9736 } 9737 return SDValue(); 9738} 9739 9740/// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS 9741/// are the two values being selected between, see if we can simplify the 9742/// select. Callers of this should assume that TheSelect is deleted if this 9743/// returns true. As such, they should return the appropriate thing (e.g. the 9744/// node) back to the top-level of the DAG combiner loop to avoid it being 9745/// looked at. 9746bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS, 9747 SDValue RHS) { 9748 9749 // Cannot simplify select with vector condition 9750 if (TheSelect->getOperand(0).getValueType().isVector()) return false; 9751 9752 // If this is a select from two identical things, try to pull the operation 9753 // through the select. 9754 if (LHS.getOpcode() != RHS.getOpcode() || 9755 !LHS.hasOneUse() || !RHS.hasOneUse()) 9756 return false; 9757 9758 // If this is a load and the token chain is identical, replace the select 9759 // of two loads with a load through a select of the address to load from. 9760 // This triggers in things like "select bool X, 10.0, 123.0" after the FP 9761 // constants have been dropped into the constant pool. 9762 if (LHS.getOpcode() == ISD::LOAD) { 9763 LoadSDNode *LLD = cast<LoadSDNode>(LHS); 9764 LoadSDNode *RLD = cast<LoadSDNode>(RHS); 9765 9766 // Token chains must be identical. 9767 if (LHS.getOperand(0) != RHS.getOperand(0) || 9768 // Do not let this transformation reduce the number of volatile loads. 9769 LLD->isVolatile() || RLD->isVolatile() || 9770 // If this is an EXTLOAD, the VT's must match. 9771 LLD->getMemoryVT() != RLD->getMemoryVT() || 9772 // If this is an EXTLOAD, the kind of extension must match. 9773 (LLD->getExtensionType() != RLD->getExtensionType() && 9774 // The only exception is if one of the extensions is anyext. 9775 LLD->getExtensionType() != ISD::EXTLOAD && 9776 RLD->getExtensionType() != ISD::EXTLOAD) || 9777 // FIXME: this discards src value information. This is 9778 // over-conservative. It would be beneficial to be able to remember 9779 // both potential memory locations. Since we are discarding 9780 // src value info, don't do the transformation if the memory 9781 // locations are not in the default address space. 9782 LLD->getPointerInfo().getAddrSpace() != 0 || 9783 RLD->getPointerInfo().getAddrSpace() != 0 || 9784 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(), 9785 LLD->getBasePtr().getValueType())) 9786 return false; 9787 9788 // Check that the select condition doesn't reach either load. If so, 9789 // folding this will induce a cycle into the DAG. If not, this is safe to 9790 // xform, so create a select of the addresses. 9791 SDValue Addr; 9792 if (TheSelect->getOpcode() == ISD::SELECT) { 9793 SDNode *CondNode = TheSelect->getOperand(0).getNode(); 9794 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) || 9795 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode))) 9796 return false; 9797 // The loads must not depend on one another. 9798 if (LLD->isPredecessorOf(RLD) || 9799 RLD->isPredecessorOf(LLD)) 9800 return false; 9801 Addr = DAG.getSelect(SDLoc(TheSelect), 9802 LLD->getBasePtr().getValueType(), 9803 TheSelect->getOperand(0), LLD->getBasePtr(), 9804 RLD->getBasePtr()); 9805 } else { // Otherwise SELECT_CC 9806 SDNode *CondLHS = TheSelect->getOperand(0).getNode(); 9807 SDNode *CondRHS = TheSelect->getOperand(1).getNode(); 9808 9809 if ((LLD->hasAnyUseOfValue(1) && 9810 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) || 9811 (RLD->hasAnyUseOfValue(1) && 9812 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS)))) 9813 return false; 9814 9815 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect), 9816 LLD->getBasePtr().getValueType(), 9817 TheSelect->getOperand(0), 9818 TheSelect->getOperand(1), 9819 LLD->getBasePtr(), RLD->getBasePtr(), 9820 TheSelect->getOperand(4)); 9821 } 9822 9823 SDValue Load; 9824 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) { 9825 Load = DAG.getLoad(TheSelect->getValueType(0), 9826 SDLoc(TheSelect), 9827 // FIXME: Discards pointer info. 9828 LLD->getChain(), Addr, MachinePointerInfo(), 9829 LLD->isVolatile(), LLD->isNonTemporal(), 9830 LLD->isInvariant(), LLD->getAlignment()); 9831 } else { 9832 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ? 9833 RLD->getExtensionType() : LLD->getExtensionType(), 9834 SDLoc(TheSelect), 9835 TheSelect->getValueType(0), 9836 // FIXME: Discards pointer info. 9837 LLD->getChain(), Addr, MachinePointerInfo(), 9838 LLD->getMemoryVT(), LLD->isVolatile(), 9839 LLD->isNonTemporal(), LLD->getAlignment()); 9840 } 9841 9842 // Users of the select now use the result of the load. 9843 CombineTo(TheSelect, Load); 9844 9845 // Users of the old loads now use the new load's chain. We know the 9846 // old-load value is dead now. 9847 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1)); 9848 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1)); 9849 return true; 9850 } 9851 9852 return false; 9853} 9854 9855/// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3 9856/// where 'cond' is the comparison specified by CC. 9857SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, 9858 SDValue N2, SDValue N3, 9859 ISD::CondCode CC, bool NotExtCompare) { 9860 // (x ? y : y) -> y. 9861 if (N2 == N3) return N2; 9862 9863 EVT VT = N2.getValueType(); 9864 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 9865 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode()); 9866 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode()); 9867 9868 // Determine if the condition we're dealing with is constant 9869 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()), 9870 N0, N1, CC, DL, false); 9871 if (SCC.getNode()) AddToWorkList(SCC.getNode()); 9872 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode()); 9873 9874 // fold select_cc true, x, y -> x 9875 if (SCCC && !SCCC->isNullValue()) 9876 return N2; 9877 // fold select_cc false, x, y -> y 9878 if (SCCC && SCCC->isNullValue()) 9879 return N3; 9880 9881 // Check to see if we can simplify the select into an fabs node 9882 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) { 9883 // Allow either -0.0 or 0.0 9884 if (CFP->getValueAPF().isZero()) { 9885 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs 9886 if ((CC == ISD::SETGE || CC == ISD::SETGT) && 9887 N0 == N2 && N3.getOpcode() == ISD::FNEG && 9888 N2 == N3.getOperand(0)) 9889 return DAG.getNode(ISD::FABS, DL, VT, N0); 9890 9891 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs 9892 if ((CC == ISD::SETLT || CC == ISD::SETLE) && 9893 N0 == N3 && N2.getOpcode() == ISD::FNEG && 9894 N2.getOperand(0) == N3) 9895 return DAG.getNode(ISD::FABS, DL, VT, N3); 9896 } 9897 } 9898 9899 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)" 9900 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0 9901 // in it. This is a win when the constant is not otherwise available because 9902 // it replaces two constant pool loads with one. We only do this if the FP 9903 // type is known to be legal, because if it isn't, then we are before legalize 9904 // types an we want the other legalization to happen first (e.g. to avoid 9905 // messing with soft float) and if the ConstantFP is not legal, because if 9906 // it is legal, we may not need to store the FP constant in a constant pool. 9907 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2)) 9908 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) { 9909 if (TLI.isTypeLegal(N2.getValueType()) && 9910 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) != 9911 TargetLowering::Legal) && 9912 // If both constants have multiple uses, then we won't need to do an 9913 // extra load, they are likely around in registers for other users. 9914 (TV->hasOneUse() || FV->hasOneUse())) { 9915 Constant *Elts[] = { 9916 const_cast<ConstantFP*>(FV->getConstantFPValue()), 9917 const_cast<ConstantFP*>(TV->getConstantFPValue()) 9918 }; 9919 Type *FPTy = Elts[0]->getType(); 9920 const DataLayout &TD = *TLI.getDataLayout(); 9921 9922 // Create a ConstantArray of the two constants. 9923 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts); 9924 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(), 9925 TD.getPrefTypeAlignment(FPTy)); 9926 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment(); 9927 9928 // Get the offsets to the 0 and 1 element of the array so that we can 9929 // select between them. 9930 SDValue Zero = DAG.getIntPtrConstant(0); 9931 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType()); 9932 SDValue One = DAG.getIntPtrConstant(EltSize); 9933 9934 SDValue Cond = DAG.getSetCC(DL, 9935 getSetCCResultType(N0.getValueType()), 9936 N0, N1, CC); 9937 AddToWorkList(Cond.getNode()); 9938 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(), 9939 Cond, One, Zero); 9940 AddToWorkList(CstOffset.getNode()); 9941 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx, 9942 CstOffset); 9943 AddToWorkList(CPIdx.getNode()); 9944 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx, 9945 MachinePointerInfo::getConstantPool(), false, 9946 false, false, Alignment); 9947 9948 } 9949 } 9950 9951 // Check to see if we can perform the "gzip trick", transforming 9952 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A) 9953 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT && 9954 (N1C->isNullValue() || // (a < 0) ? b : 0 9955 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0 9956 EVT XType = N0.getValueType(); 9957 EVT AType = N2.getValueType(); 9958 if (XType.bitsGE(AType)) { 9959 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a 9960 // single-bit constant. 9961 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) { 9962 unsigned ShCtV = N2C->getAPIntValue().logBase2(); 9963 ShCtV = XType.getSizeInBits()-ShCtV-1; 9964 SDValue ShCt = DAG.getConstant(ShCtV, 9965 getShiftAmountTy(N0.getValueType())); 9966 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), 9967 XType, N0, ShCt); 9968 AddToWorkList(Shift.getNode()); 9969 9970 if (XType.bitsGT(AType)) { 9971 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift); 9972 AddToWorkList(Shift.getNode()); 9973 } 9974 9975 return DAG.getNode(ISD::AND, DL, AType, Shift, N2); 9976 } 9977 9978 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), 9979 XType, N0, 9980 DAG.getConstant(XType.getSizeInBits()-1, 9981 getShiftAmountTy(N0.getValueType()))); 9982 AddToWorkList(Shift.getNode()); 9983 9984 if (XType.bitsGT(AType)) { 9985 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift); 9986 AddToWorkList(Shift.getNode()); 9987 } 9988 9989 return DAG.getNode(ISD::AND, DL, AType, Shift, N2); 9990 } 9991 } 9992 9993 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A) 9994 // where y is has a single bit set. 9995 // A plaintext description would be, we can turn the SELECT_CC into an AND 9996 // when the condition can be materialized as an all-ones register. Any 9997 // single bit-test can be materialized as an all-ones register with 9998 // shift-left and shift-right-arith. 9999 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND && 10000 N0->getValueType(0) == VT && 10001 N1C && N1C->isNullValue() && 10002 N2C && N2C->isNullValue()) { 10003 SDValue AndLHS = N0->getOperand(0); 10004 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1)); 10005 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) { 10006 // Shift the tested bit over the sign bit. 10007 APInt AndMask = ConstAndRHS->getAPIntValue(); 10008 SDValue ShlAmt = 10009 DAG.getConstant(AndMask.countLeadingZeros(), 10010 getShiftAmountTy(AndLHS.getValueType())); 10011 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt); 10012 10013 // Now arithmetic right shift it all the way over, so the result is either 10014 // all-ones, or zero. 10015 SDValue ShrAmt = 10016 DAG.getConstant(AndMask.getBitWidth()-1, 10017 getShiftAmountTy(Shl.getValueType())); 10018 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt); 10019 10020 return DAG.getNode(ISD::AND, DL, VT, Shr, N3); 10021 } 10022 } 10023 10024 // fold select C, 16, 0 -> shl C, 4 10025 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() && 10026 TLI.getBooleanContents(N0.getValueType().isVector()) == 10027 TargetLowering::ZeroOrOneBooleanContent) { 10028 10029 // If the caller doesn't want us to simplify this into a zext of a compare, 10030 // don't do it. 10031 if (NotExtCompare && N2C->getAPIntValue() == 1) 10032 return SDValue(); 10033 10034 // Get a SetCC of the condition 10035 // NOTE: Don't create a SETCC if it's not legal on this target. 10036 if (!LegalOperations || 10037 TLI.isOperationLegal(ISD::SETCC, 10038 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) { 10039 SDValue Temp, SCC; 10040 // cast from setcc result type to select result type 10041 if (LegalTypes) { 10042 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()), 10043 N0, N1, CC); 10044 if (N2.getValueType().bitsLT(SCC.getValueType())) 10045 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2), 10046 N2.getValueType()); 10047 else 10048 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2), 10049 N2.getValueType(), SCC); 10050 } else { 10051 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC); 10052 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2), 10053 N2.getValueType(), SCC); 10054 } 10055 10056 AddToWorkList(SCC.getNode()); 10057 AddToWorkList(Temp.getNode()); 10058 10059 if (N2C->getAPIntValue() == 1) 10060 return Temp; 10061 10062 // shl setcc result by log2 n2c 10063 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp, 10064 DAG.getConstant(N2C->getAPIntValue().logBase2(), 10065 getShiftAmountTy(Temp.getValueType()))); 10066 } 10067 } 10068 10069 // Check to see if this is the equivalent of setcc 10070 // FIXME: Turn all of these into setcc if setcc if setcc is legal 10071 // otherwise, go ahead with the folds. 10072 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) { 10073 EVT XType = N0.getValueType(); 10074 if (!LegalOperations || 10075 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) { 10076 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC); 10077 if (Res.getValueType() != VT) 10078 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res); 10079 return Res; 10080 } 10081 10082 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X)))) 10083 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ && 10084 (!LegalOperations || 10085 TLI.isOperationLegal(ISD::CTLZ, XType))) { 10086 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0); 10087 return DAG.getNode(ISD::SRL, DL, XType, Ctlz, 10088 DAG.getConstant(Log2_32(XType.getSizeInBits()), 10089 getShiftAmountTy(Ctlz.getValueType()))); 10090 } 10091 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1)) 10092 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) { 10093 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0), 10094 XType, DAG.getConstant(0, XType), N0); 10095 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType); 10096 return DAG.getNode(ISD::SRL, DL, XType, 10097 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0), 10098 DAG.getConstant(XType.getSizeInBits()-1, 10099 getShiftAmountTy(XType))); 10100 } 10101 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1)) 10102 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) { 10103 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0, 10104 DAG.getConstant(XType.getSizeInBits()-1, 10105 getShiftAmountTy(N0.getValueType()))); 10106 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType)); 10107 } 10108 } 10109 10110 // Check to see if this is an integer abs. 10111 // select_cc setg[te] X, 0, X, -X -> 10112 // select_cc setgt X, -1, X, -X -> 10113 // select_cc setl[te] X, 0, -X, X -> 10114 // select_cc setlt X, 1, -X, X -> 10115 // Y = sra (X, size(X)-1); xor (add (X, Y), Y) 10116 if (N1C) { 10117 ConstantSDNode *SubC = NULL; 10118 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) || 10119 (N1C->isAllOnesValue() && CC == ISD::SETGT)) && 10120 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1)) 10121 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0)); 10122 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) || 10123 (N1C->isOne() && CC == ISD::SETLT)) && 10124 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) 10125 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0)); 10126 10127 EVT XType = N0.getValueType(); 10128 if (SubC && SubC->isNullValue() && XType.isInteger()) { 10129 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType, 10130 N0, 10131 DAG.getConstant(XType.getSizeInBits()-1, 10132 getShiftAmountTy(N0.getValueType()))); 10133 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0), 10134 XType, N0, Shift); 10135 AddToWorkList(Shift.getNode()); 10136 AddToWorkList(Add.getNode()); 10137 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift); 10138 } 10139 } 10140 10141 return SDValue(); 10142} 10143 10144/// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC. 10145SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0, 10146 SDValue N1, ISD::CondCode Cond, 10147 SDLoc DL, bool foldBooleans) { 10148 TargetLowering::DAGCombinerInfo 10149 DagCombineInfo(DAG, Level, false, this); 10150 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL); 10151} 10152 10153/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 10154/// return a DAG expression to select that will generate the same value by 10155/// multiplying by a magic number. See: 10156/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 10157SDValue DAGCombiner::BuildSDIV(SDNode *N) { 10158 std::vector<SDNode*> Built; 10159 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built); 10160 10161 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end(); 10162 ii != ee; ++ii) 10163 AddToWorkList(*ii); 10164 return S; 10165} 10166 10167/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 10168/// return a DAG expression to select that will generate the same value by 10169/// multiplying by a magic number. See: 10170/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 10171SDValue DAGCombiner::BuildUDIV(SDNode *N) { 10172 std::vector<SDNode*> Built; 10173 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built); 10174 10175 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end(); 10176 ii != ee; ++ii) 10177 AddToWorkList(*ii); 10178 return S; 10179} 10180 10181/// FindBaseOffset - Return true if base is a frame index, which is known not 10182// to alias with anything but itself. Provides base object and offset as 10183// results. 10184static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset, 10185 const GlobalValue *&GV, const void *&CV) { 10186 // Assume it is a primitive operation. 10187 Base = Ptr; Offset = 0; GV = 0; CV = 0; 10188 10189 // If it's an adding a simple constant then integrate the offset. 10190 if (Base.getOpcode() == ISD::ADD) { 10191 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) { 10192 Base = Base.getOperand(0); 10193 Offset += C->getZExtValue(); 10194 } 10195 } 10196 10197 // Return the underlying GlobalValue, and update the Offset. Return false 10198 // for GlobalAddressSDNode since the same GlobalAddress may be represented 10199 // by multiple nodes with different offsets. 10200 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) { 10201 GV = G->getGlobal(); 10202 Offset += G->getOffset(); 10203 return false; 10204 } 10205 10206 // Return the underlying Constant value, and update the Offset. Return false 10207 // for ConstantSDNodes since the same constant pool entry may be represented 10208 // by multiple nodes with different offsets. 10209 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) { 10210 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal() 10211 : (const void *)C->getConstVal(); 10212 Offset += C->getOffset(); 10213 return false; 10214 } 10215 // If it's any of the following then it can't alias with anything but itself. 10216 return isa<FrameIndexSDNode>(Base); 10217} 10218 10219/// isAlias - Return true if there is any possibility that the two addresses 10220/// overlap. 10221bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1, 10222 const Value *SrcValue1, int SrcValueOffset1, 10223 unsigned SrcValueAlign1, 10224 const MDNode *TBAAInfo1, 10225 SDValue Ptr2, int64_t Size2, 10226 const Value *SrcValue2, int SrcValueOffset2, 10227 unsigned SrcValueAlign2, 10228 const MDNode *TBAAInfo2) const { 10229 // If they are the same then they must be aliases. 10230 if (Ptr1 == Ptr2) return true; 10231 10232 // Gather base node and offset information. 10233 SDValue Base1, Base2; 10234 int64_t Offset1, Offset2; 10235 const GlobalValue *GV1, *GV2; 10236 const void *CV1, *CV2; 10237 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1); 10238 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2); 10239 10240 // If they have a same base address then check to see if they overlap. 10241 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2))) 10242 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1); 10243 10244 // It is possible for different frame indices to alias each other, mostly 10245 // when tail call optimization reuses return address slots for arguments. 10246 // To catch this case, look up the actual index of frame indices to compute 10247 // the real alias relationship. 10248 if (isFrameIndex1 && isFrameIndex2) { 10249 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 10250 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex()); 10251 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex()); 10252 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1); 10253 } 10254 10255 // Otherwise, if we know what the bases are, and they aren't identical, then 10256 // we know they cannot alias. 10257 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2)) 10258 return false; 10259 10260 // If we know required SrcValue1 and SrcValue2 have relatively large alignment 10261 // compared to the size and offset of the access, we may be able to prove they 10262 // do not alias. This check is conservative for now to catch cases created by 10263 // splitting vector types. 10264 if ((SrcValueAlign1 == SrcValueAlign2) && 10265 (SrcValueOffset1 != SrcValueOffset2) && 10266 (Size1 == Size2) && (SrcValueAlign1 > Size1)) { 10267 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1; 10268 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1; 10269 10270 // There is no overlap between these relatively aligned accesses of similar 10271 // size, return no alias. 10272 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1) 10273 return false; 10274 } 10275 10276 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA : 10277 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA(); 10278 if (UseAA && SrcValue1 && SrcValue2) { 10279 // Use alias analysis information. 10280 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2); 10281 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset; 10282 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset; 10283 AliasAnalysis::AliasResult AAResult = 10284 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1), 10285 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2)); 10286 if (AAResult == AliasAnalysis::NoAlias) 10287 return false; 10288 } 10289 10290 // Otherwise we have to assume they alias. 10291 return true; 10292} 10293 10294bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) { 10295 SDValue Ptr0, Ptr1; 10296 int64_t Size0, Size1; 10297 const Value *SrcValue0, *SrcValue1; 10298 int SrcValueOffset0, SrcValueOffset1; 10299 unsigned SrcValueAlign0, SrcValueAlign1; 10300 const MDNode *SrcTBAAInfo0, *SrcTBAAInfo1; 10301 FindAliasInfo(Op0, Ptr0, Size0, SrcValue0, SrcValueOffset0, 10302 SrcValueAlign0, SrcTBAAInfo0); 10303 FindAliasInfo(Op1, Ptr1, Size1, SrcValue1, SrcValueOffset1, 10304 SrcValueAlign1, SrcTBAAInfo1); 10305 return isAlias(Ptr0, Size0, SrcValue0, SrcValueOffset0, 10306 SrcValueAlign0, SrcTBAAInfo0, 10307 Ptr1, Size1, SrcValue1, SrcValueOffset1, 10308 SrcValueAlign1, SrcTBAAInfo1); 10309} 10310 10311/// FindAliasInfo - Extracts the relevant alias information from the memory 10312/// node. Returns true if the operand was a load. 10313bool DAGCombiner::FindAliasInfo(SDNode *N, 10314 SDValue &Ptr, int64_t &Size, 10315 const Value *&SrcValue, 10316 int &SrcValueOffset, 10317 unsigned &SrcValueAlign, 10318 const MDNode *&TBAAInfo) const { 10319 LSBaseSDNode *LS = cast<LSBaseSDNode>(N); 10320 10321 Ptr = LS->getBasePtr(); 10322 Size = LS->getMemoryVT().getSizeInBits() >> 3; 10323 SrcValue = LS->getSrcValue(); 10324 SrcValueOffset = LS->getSrcValueOffset(); 10325 SrcValueAlign = LS->getOriginalAlignment(); 10326 TBAAInfo = LS->getTBAAInfo(); 10327 return isa<LoadSDNode>(LS); 10328} 10329 10330/// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes, 10331/// looking for aliasing nodes and adding them to the Aliases vector. 10332void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain, 10333 SmallVectorImpl<SDValue> &Aliases) { 10334 SmallVector<SDValue, 8> Chains; // List of chains to visit. 10335 SmallPtrSet<SDNode *, 16> Visited; // Visited node set. 10336 10337 // Get alias information for node. 10338 SDValue Ptr; 10339 int64_t Size; 10340 const Value *SrcValue; 10341 int SrcValueOffset; 10342 unsigned SrcValueAlign; 10343 const MDNode *SrcTBAAInfo; 10344 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset, 10345 SrcValueAlign, SrcTBAAInfo); 10346 10347 // Starting off. 10348 Chains.push_back(OriginalChain); 10349 unsigned Depth = 0; 10350 10351 // Look at each chain and determine if it is an alias. If so, add it to the 10352 // aliases list. If not, then continue up the chain looking for the next 10353 // candidate. 10354 while (!Chains.empty()) { 10355 SDValue Chain = Chains.back(); 10356 Chains.pop_back(); 10357 10358 // For TokenFactor nodes, look at each operand and only continue up the 10359 // chain until we find two aliases. If we've seen two aliases, assume we'll 10360 // find more and revert to original chain since the xform is unlikely to be 10361 // profitable. 10362 // 10363 // FIXME: The depth check could be made to return the last non-aliasing 10364 // chain we found before we hit a tokenfactor rather than the original 10365 // chain. 10366 if (Depth > 6 || Aliases.size() == 2) { 10367 Aliases.clear(); 10368 Aliases.push_back(OriginalChain); 10369 break; 10370 } 10371 10372 // Don't bother if we've been before. 10373 if (!Visited.insert(Chain.getNode())) 10374 continue; 10375 10376 switch (Chain.getOpcode()) { 10377 case ISD::EntryToken: 10378 // Entry token is ideal chain operand, but handled in FindBetterChain. 10379 break; 10380 10381 case ISD::LOAD: 10382 case ISD::STORE: { 10383 // Get alias information for Chain. 10384 SDValue OpPtr; 10385 int64_t OpSize; 10386 const Value *OpSrcValue; 10387 int OpSrcValueOffset; 10388 unsigned OpSrcValueAlign; 10389 const MDNode *OpSrcTBAAInfo; 10390 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize, 10391 OpSrcValue, OpSrcValueOffset, 10392 OpSrcValueAlign, 10393 OpSrcTBAAInfo); 10394 10395 // If chain is alias then stop here. 10396 if (!(IsLoad && IsOpLoad) && 10397 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign, 10398 SrcTBAAInfo, 10399 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset, 10400 OpSrcValueAlign, OpSrcTBAAInfo)) { 10401 Aliases.push_back(Chain); 10402 } else { 10403 // Look further up the chain. 10404 Chains.push_back(Chain.getOperand(0)); 10405 ++Depth; 10406 } 10407 break; 10408 } 10409 10410 case ISD::TokenFactor: 10411 // We have to check each of the operands of the token factor for "small" 10412 // token factors, so we queue them up. Adding the operands to the queue 10413 // (stack) in reverse order maintains the original order and increases the 10414 // likelihood that getNode will find a matching token factor (CSE.) 10415 if (Chain.getNumOperands() > 16) { 10416 Aliases.push_back(Chain); 10417 break; 10418 } 10419 for (unsigned n = Chain.getNumOperands(); n;) 10420 Chains.push_back(Chain.getOperand(--n)); 10421 ++Depth; 10422 break; 10423 10424 default: 10425 // For all other instructions we will just have to take what we can get. 10426 Aliases.push_back(Chain); 10427 break; 10428 } 10429 } 10430} 10431 10432/// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking 10433/// for a better chain (aliasing node.) 10434SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) { 10435 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor. 10436 10437 // Accumulate all the aliases to this node. 10438 GatherAllAliases(N, OldChain, Aliases); 10439 10440 // If no operands then chain to entry token. 10441 if (Aliases.size() == 0) 10442 return DAG.getEntryNode(); 10443 10444 // If a single operand then chain to it. We don't need to revisit it. 10445 if (Aliases.size() == 1) 10446 return Aliases[0]; 10447 10448 // Construct a custom tailored token factor. 10449 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, 10450 &Aliases[0], Aliases.size()); 10451} 10452 10453// SelectionDAG::Combine - This is the entry point for the file. 10454// 10455void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA, 10456 CodeGenOpt::Level OptLevel) { 10457 /// run - This is the main entry point to this class. 10458 /// 10459 DAGCombiner(*this, AA, OptLevel).Run(Level); 10460} 10461