DAGCombiner.cpp revision fa1eb27b76ab1e0f78574bf52a432c84a4c1a520
1//===-- DAGCombiner.cpp - Implement a DAG node combiner -------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by Nate Begeman and is distributed under the 6// University of Illinois Open Source 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// FIXME: Missing folds 14// sdiv, udiv, srem, urem (X, const) where X is an integer can be expanded into 15// a sequence of multiplies, shifts, and adds. This should be controlled by 16// some kind of hint from the target that int div is expensive. 17// various folds of mulh[s,u] by constants such as -1, powers of 2, etc. 18// 19// FIXME: select C, pow2, pow2 -> something smart 20// FIXME: trunc(select X, Y, Z) -> select X, trunc(Y), trunc(Z) 21// FIXME: Dead stores -> nuke 22// FIXME: shr X, (and Y,31) -> shr X, Y (TRICKY!) 23// FIXME: mul (x, const) -> shifts + adds 24// FIXME: undef values 25// FIXME: divide by zero is currently left unfolded. do we want to turn this 26// into an undef? 27// FIXME: select ne (select cc, 1, 0), 0, true, false -> select cc, true, false 28// 29//===----------------------------------------------------------------------===// 30 31#define DEBUG_TYPE "dagcombine" 32#include "llvm/ADT/Statistic.h" 33#include "llvm/Analysis/AliasAnalysis.h" 34#include "llvm/CodeGen/SelectionDAG.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/MathExtras.h" 37#include "llvm/Target/TargetLowering.h" 38#include "llvm/Target/TargetOptions.h" 39#include "llvm/Support/Compiler.h" 40#include "llvm/Support/CommandLine.h" 41#include <algorithm> 42using namespace llvm; 43 44STATISTIC(NodesCombined , "Number of dag nodes combined"); 45STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created"); 46STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created"); 47 48namespace { 49#ifndef NDEBUG 50 static cl::opt<bool> 51 ViewDAGCombine1("view-dag-combine1-dags", cl::Hidden, 52 cl::desc("Pop up a window to show dags before the first " 53 "dag combine pass")); 54 static cl::opt<bool> 55 ViewDAGCombine2("view-dag-combine2-dags", cl::Hidden, 56 cl::desc("Pop up a window to show dags before the second " 57 "dag combine pass")); 58#else 59 static const bool ViewDAGCombine1 = false; 60 static const bool ViewDAGCombine2 = false; 61#endif 62 63 static cl::opt<bool> 64 CombinerAA("combiner-alias-analysis", cl::Hidden, 65 cl::desc("Turn on alias analysis during testing")); 66 67 static cl::opt<bool> 68 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden, 69 cl::desc("Include global information in alias analysis")); 70 71//------------------------------ DAGCombiner ---------------------------------// 72 73 class VISIBILITY_HIDDEN DAGCombiner { 74 SelectionDAG &DAG; 75 TargetLowering &TLI; 76 bool AfterLegalize; 77 78 // Worklist of all of the nodes that need to be simplified. 79 std::vector<SDNode*> WorkList; 80 81 // AA - Used for DAG load/store alias analysis. 82 AliasAnalysis &AA; 83 84 /// AddUsersToWorkList - When an instruction is simplified, add all users of 85 /// the instruction to the work lists because they might get more simplified 86 /// now. 87 /// 88 void AddUsersToWorkList(SDNode *N) { 89 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 90 UI != UE; ++UI) 91 AddToWorkList(*UI); 92 } 93 94 /// removeFromWorkList - remove all instances of N from the worklist. 95 /// 96 void removeFromWorkList(SDNode *N) { 97 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), N), 98 WorkList.end()); 99 } 100 101 public: 102 /// AddToWorkList - Add to the work list making sure it's instance is at the 103 /// the back (next to be processed.) 104 void AddToWorkList(SDNode *N) { 105 removeFromWorkList(N); 106 WorkList.push_back(N); 107 } 108 109 SDOperand CombineTo(SDNode *N, const SDOperand *To, unsigned NumTo, 110 bool AddTo = true) { 111 assert(N->getNumValues() == NumTo && "Broken CombineTo call!"); 112 ++NodesCombined; 113 DOUT << "\nReplacing.1 "; DEBUG(N->dump()); 114 DOUT << "\nWith: "; DEBUG(To[0].Val->dump(&DAG)); 115 DOUT << " and " << NumTo-1 << " other values\n"; 116 std::vector<SDNode*> NowDead; 117 DAG.ReplaceAllUsesWith(N, To, &NowDead); 118 119 if (AddTo) { 120 // Push the new nodes and any users onto the worklist 121 for (unsigned i = 0, e = NumTo; i != e; ++i) { 122 AddToWorkList(To[i].Val); 123 AddUsersToWorkList(To[i].Val); 124 } 125 } 126 127 // Nodes can be reintroduced into the worklist. Make sure we do not 128 // process a node that has been replaced. 129 removeFromWorkList(N); 130 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 131 removeFromWorkList(NowDead[i]); 132 133 // Finally, since the node is now dead, remove it from the graph. 134 DAG.DeleteNode(N); 135 return SDOperand(N, 0); 136 } 137 138 SDOperand CombineTo(SDNode *N, SDOperand Res, bool AddTo = true) { 139 return CombineTo(N, &Res, 1, AddTo); 140 } 141 142 SDOperand CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1, 143 bool AddTo = true) { 144 SDOperand To[] = { Res0, Res1 }; 145 return CombineTo(N, To, 2, AddTo); 146 } 147 private: 148 149 /// SimplifyDemandedBits - Check the specified integer node value to see if 150 /// it can be simplified or if things it uses can be simplified by bit 151 /// propagation. If so, return true. 152 bool SimplifyDemandedBits(SDOperand Op) { 153 TargetLowering::TargetLoweringOpt TLO(DAG); 154 uint64_t KnownZero, KnownOne; 155 uint64_t Demanded = MVT::getIntVTBitMask(Op.getValueType()); 156 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO)) 157 return false; 158 159 // Revisit the node. 160 AddToWorkList(Op.Val); 161 162 // Replace the old value with the new one. 163 ++NodesCombined; 164 DOUT << "\nReplacing.2 "; DEBUG(TLO.Old.Val->dump()); 165 DOUT << "\nWith: "; DEBUG(TLO.New.Val->dump(&DAG)); 166 DOUT << '\n'; 167 168 std::vector<SDNode*> NowDead; 169 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, NowDead); 170 171 // Push the new node and any (possibly new) users onto the worklist. 172 AddToWorkList(TLO.New.Val); 173 AddUsersToWorkList(TLO.New.Val); 174 175 // Nodes can end up on the worklist more than once. Make sure we do 176 // not process a node that has been replaced. 177 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 178 removeFromWorkList(NowDead[i]); 179 180 // Finally, if the node is now dead, remove it from the graph. The node 181 // may not be dead if the replacement process recursively simplified to 182 // something else needing this node. 183 if (TLO.Old.Val->use_empty()) { 184 removeFromWorkList(TLO.Old.Val); 185 DAG.DeleteNode(TLO.Old.Val); 186 } 187 return true; 188 } 189 190 bool CombineToPreIndexedLoadStore(SDNode *N); 191 bool CombineToPostIndexedLoadStore(SDNode *N); 192 193 194 /// visit - call the node-specific routine that knows how to fold each 195 /// particular type of node. 196 SDOperand visit(SDNode *N); 197 198 // Visitation implementation - Implement dag node combining for different 199 // node types. The semantics are as follows: 200 // Return Value: 201 // SDOperand.Val == 0 - No change was made 202 // SDOperand.Val == N - N was replaced, is dead, and is already handled. 203 // otherwise - N should be replaced by the returned Operand. 204 // 205 SDOperand visitTokenFactor(SDNode *N); 206 SDOperand visitADD(SDNode *N); 207 SDOperand visitSUB(SDNode *N); 208 SDOperand visitMUL(SDNode *N); 209 SDOperand visitSDIV(SDNode *N); 210 SDOperand visitUDIV(SDNode *N); 211 SDOperand visitSREM(SDNode *N); 212 SDOperand visitUREM(SDNode *N); 213 SDOperand visitMULHU(SDNode *N); 214 SDOperand visitMULHS(SDNode *N); 215 SDOperand visitAND(SDNode *N); 216 SDOperand visitOR(SDNode *N); 217 SDOperand visitXOR(SDNode *N); 218 SDOperand visitVBinOp(SDNode *N, ISD::NodeType IntOp, ISD::NodeType FPOp); 219 SDOperand visitSHL(SDNode *N); 220 SDOperand visitSRA(SDNode *N); 221 SDOperand visitSRL(SDNode *N); 222 SDOperand visitCTLZ(SDNode *N); 223 SDOperand visitCTTZ(SDNode *N); 224 SDOperand visitCTPOP(SDNode *N); 225 SDOperand visitSELECT(SDNode *N); 226 SDOperand visitSELECT_CC(SDNode *N); 227 SDOperand visitSETCC(SDNode *N); 228 SDOperand visitSIGN_EXTEND(SDNode *N); 229 SDOperand visitZERO_EXTEND(SDNode *N); 230 SDOperand visitANY_EXTEND(SDNode *N); 231 SDOperand visitSIGN_EXTEND_INREG(SDNode *N); 232 SDOperand visitTRUNCATE(SDNode *N); 233 SDOperand visitBIT_CONVERT(SDNode *N); 234 SDOperand visitVBIT_CONVERT(SDNode *N); 235 SDOperand visitFADD(SDNode *N); 236 SDOperand visitFSUB(SDNode *N); 237 SDOperand visitFMUL(SDNode *N); 238 SDOperand visitFDIV(SDNode *N); 239 SDOperand visitFREM(SDNode *N); 240 SDOperand visitFCOPYSIGN(SDNode *N); 241 SDOperand visitSINT_TO_FP(SDNode *N); 242 SDOperand visitUINT_TO_FP(SDNode *N); 243 SDOperand visitFP_TO_SINT(SDNode *N); 244 SDOperand visitFP_TO_UINT(SDNode *N); 245 SDOperand visitFP_ROUND(SDNode *N); 246 SDOperand visitFP_ROUND_INREG(SDNode *N); 247 SDOperand visitFP_EXTEND(SDNode *N); 248 SDOperand visitFNEG(SDNode *N); 249 SDOperand visitFABS(SDNode *N); 250 SDOperand visitBRCOND(SDNode *N); 251 SDOperand visitBR_CC(SDNode *N); 252 SDOperand visitLOAD(SDNode *N); 253 SDOperand visitSTORE(SDNode *N); 254 SDOperand visitINSERT_VECTOR_ELT(SDNode *N); 255 SDOperand visitVINSERT_VECTOR_ELT(SDNode *N); 256 SDOperand visitVBUILD_VECTOR(SDNode *N); 257 SDOperand visitVECTOR_SHUFFLE(SDNode *N); 258 SDOperand visitVVECTOR_SHUFFLE(SDNode *N); 259 260 SDOperand XformToShuffleWithZero(SDNode *N); 261 SDOperand ReassociateOps(unsigned Opc, SDOperand LHS, SDOperand RHS); 262 263 bool SimplifySelectOps(SDNode *SELECT, SDOperand LHS, SDOperand RHS); 264 SDOperand SimplifyBinOpWithSameOpcodeHands(SDNode *N); 265 SDOperand SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2); 266 SDOperand SimplifySelectCC(SDOperand N0, SDOperand N1, SDOperand N2, 267 SDOperand N3, ISD::CondCode CC); 268 SDOperand SimplifySetCC(MVT::ValueType VT, SDOperand N0, SDOperand N1, 269 ISD::CondCode Cond, bool foldBooleans = true); 270 SDOperand ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *, MVT::ValueType); 271 SDOperand BuildSDIV(SDNode *N); 272 SDOperand BuildUDIV(SDNode *N); 273 SDNode *MatchRotate(SDOperand LHS, SDOperand RHS); 274 275 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes, 276 /// looking for aliasing nodes and adding them to the Aliases vector. 277 void GatherAllAliases(SDNode *N, SDOperand OriginalChain, 278 SmallVector<SDOperand, 8> &Aliases); 279 280 /// isAlias - Return true if there is any possibility that the two addresses 281 /// overlap. 282 bool isAlias(SDOperand Ptr1, int64_t Size1, 283 const Value *SrcValue1, int SrcValueOffset1, 284 SDOperand Ptr2, int64_t Size2, 285 const Value *SrcValue2, int SrcValueOffset2); 286 287 /// FindAliasInfo - Extracts the relevant alias information from the memory 288 /// node. Returns true if the operand was a load. 289 bool FindAliasInfo(SDNode *N, 290 SDOperand &Ptr, int64_t &Size, 291 const Value *&SrcValue, int &SrcValueOffset); 292 293 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, 294 /// looking for a better chain (aliasing node.) 295 SDOperand FindBetterChain(SDNode *N, SDOperand Chain); 296 297public: 298 DAGCombiner(SelectionDAG &D, AliasAnalysis &A) 299 : DAG(D), 300 TLI(D.getTargetLoweringInfo()), 301 AfterLegalize(false), 302 AA(A) {} 303 304 /// Run - runs the dag combiner on all nodes in the work list 305 void Run(bool RunningAfterLegalize); 306 }; 307} 308 309//===----------------------------------------------------------------------===// 310// TargetLowering::DAGCombinerInfo implementation 311//===----------------------------------------------------------------------===// 312 313void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) { 314 ((DAGCombiner*)DC)->AddToWorkList(N); 315} 316 317SDOperand TargetLowering::DAGCombinerInfo:: 318CombineTo(SDNode *N, const std::vector<SDOperand> &To) { 319 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size()); 320} 321 322SDOperand TargetLowering::DAGCombinerInfo:: 323CombineTo(SDNode *N, SDOperand Res) { 324 return ((DAGCombiner*)DC)->CombineTo(N, Res); 325} 326 327 328SDOperand TargetLowering::DAGCombinerInfo:: 329CombineTo(SDNode *N, SDOperand Res0, SDOperand Res1) { 330 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1); 331} 332 333 334 335 336//===----------------------------------------------------------------------===// 337 338 339// isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc 340// that selects between the values 1 and 0, making it equivalent to a setcc. 341// Also, set the incoming LHS, RHS, and CC references to the appropriate 342// nodes based on the type of node we are checking. This simplifies life a 343// bit for the callers. 344static bool isSetCCEquivalent(SDOperand N, SDOperand &LHS, SDOperand &RHS, 345 SDOperand &CC) { 346 if (N.getOpcode() == ISD::SETCC) { 347 LHS = N.getOperand(0); 348 RHS = N.getOperand(1); 349 CC = N.getOperand(2); 350 return true; 351 } 352 if (N.getOpcode() == ISD::SELECT_CC && 353 N.getOperand(2).getOpcode() == ISD::Constant && 354 N.getOperand(3).getOpcode() == ISD::Constant && 355 cast<ConstantSDNode>(N.getOperand(2))->getValue() == 1 && 356 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) { 357 LHS = N.getOperand(0); 358 RHS = N.getOperand(1); 359 CC = N.getOperand(4); 360 return true; 361 } 362 return false; 363} 364 365// isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only 366// one use. If this is true, it allows the users to invert the operation for 367// free when it is profitable to do so. 368static bool isOneUseSetCC(SDOperand N) { 369 SDOperand N0, N1, N2; 370 if (isSetCCEquivalent(N, N0, N1, N2) && N.Val->hasOneUse()) 371 return true; 372 return false; 373} 374 375SDOperand DAGCombiner::ReassociateOps(unsigned Opc, SDOperand N0, SDOperand N1){ 376 MVT::ValueType VT = N0.getValueType(); 377 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use 378 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2)) 379 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) { 380 if (isa<ConstantSDNode>(N1)) { 381 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(1), N1); 382 AddToWorkList(OpNode.Val); 383 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(0)); 384 } else if (N0.hasOneUse()) { 385 SDOperand OpNode = DAG.getNode(Opc, VT, N0.getOperand(0), N1); 386 AddToWorkList(OpNode.Val); 387 return DAG.getNode(Opc, VT, OpNode, N0.getOperand(1)); 388 } 389 } 390 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use 391 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2)) 392 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) { 393 if (isa<ConstantSDNode>(N0)) { 394 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(1), N0); 395 AddToWorkList(OpNode.Val); 396 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(0)); 397 } else if (N1.hasOneUse()) { 398 SDOperand OpNode = DAG.getNode(Opc, VT, N1.getOperand(0), N0); 399 AddToWorkList(OpNode.Val); 400 return DAG.getNode(Opc, VT, OpNode, N1.getOperand(1)); 401 } 402 } 403 return SDOperand(); 404} 405 406void DAGCombiner::Run(bool RunningAfterLegalize) { 407 // set the instance variable, so that the various visit routines may use it. 408 AfterLegalize = RunningAfterLegalize; 409 410 // Add all the dag nodes to the worklist. 411 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(), 412 E = DAG.allnodes_end(); I != E; ++I) 413 WorkList.push_back(I); 414 415 // Create a dummy node (which is not added to allnodes), that adds a reference 416 // to the root node, preventing it from being deleted, and tracking any 417 // changes of the root. 418 HandleSDNode Dummy(DAG.getRoot()); 419 420 // The root of the dag may dangle to deleted nodes until the dag combiner is 421 // done. Set it to null to avoid confusion. 422 DAG.setRoot(SDOperand()); 423 424 /// DagCombineInfo - Expose the DAG combiner to the target combiner impls. 425 TargetLowering::DAGCombinerInfo 426 DagCombineInfo(DAG, !RunningAfterLegalize, false, this); 427 428 // while the worklist isn't empty, inspect the node on the end of it and 429 // try and combine it. 430 while (!WorkList.empty()) { 431 SDNode *N = WorkList.back(); 432 WorkList.pop_back(); 433 434 // If N has no uses, it is dead. Make sure to revisit all N's operands once 435 // N is deleted from the DAG, since they too may now be dead or may have a 436 // reduced number of uses, allowing other xforms. 437 if (N->use_empty() && N != &Dummy) { 438 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 439 AddToWorkList(N->getOperand(i).Val); 440 441 DAG.DeleteNode(N); 442 continue; 443 } 444 445 SDOperand RV = visit(N); 446 447 // If nothing happened, try a target-specific DAG combine. 448 if (RV.Val == 0) { 449 assert(N->getOpcode() != ISD::DELETED_NODE && 450 "Node was deleted but visit returned NULL!"); 451 if (N->getOpcode() >= ISD::BUILTIN_OP_END || 452 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) 453 RV = TLI.PerformDAGCombine(N, DagCombineInfo); 454 } 455 456 if (RV.Val) { 457 ++NodesCombined; 458 // If we get back the same node we passed in, rather than a new node or 459 // zero, we know that the node must have defined multiple values and 460 // CombineTo was used. Since CombineTo takes care of the worklist 461 // mechanics for us, we have no work to do in this case. 462 if (RV.Val != N) { 463 assert(N->getOpcode() != ISD::DELETED_NODE && 464 RV.Val->getOpcode() != ISD::DELETED_NODE && 465 "Node was deleted but visit returned new node!"); 466 467 DOUT << "\nReplacing.3 "; DEBUG(N->dump()); 468 DOUT << "\nWith: "; DEBUG(RV.Val->dump(&DAG)); 469 DOUT << '\n'; 470 std::vector<SDNode*> NowDead; 471 if (N->getNumValues() == RV.Val->getNumValues()) 472 DAG.ReplaceAllUsesWith(N, RV.Val, &NowDead); 473 else { 474 assert(N->getValueType(0) == RV.getValueType() && "Type mismatch"); 475 SDOperand OpV = RV; 476 DAG.ReplaceAllUsesWith(N, &OpV, &NowDead); 477 } 478 479 // Push the new node and any users onto the worklist 480 AddToWorkList(RV.Val); 481 AddUsersToWorkList(RV.Val); 482 483 // Nodes can be reintroduced into the worklist. Make sure we do not 484 // process a node that has been replaced. 485 removeFromWorkList(N); 486 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 487 removeFromWorkList(NowDead[i]); 488 489 // Finally, since the node is now dead, remove it from the graph. 490 DAG.DeleteNode(N); 491 } 492 } 493 } 494 495 // If the root changed (e.g. it was a dead load, update the root). 496 DAG.setRoot(Dummy.getValue()); 497} 498 499SDOperand DAGCombiner::visit(SDNode *N) { 500 switch(N->getOpcode()) { 501 default: break; 502 case ISD::TokenFactor: return visitTokenFactor(N); 503 case ISD::ADD: return visitADD(N); 504 case ISD::SUB: return visitSUB(N); 505 case ISD::MUL: return visitMUL(N); 506 case ISD::SDIV: return visitSDIV(N); 507 case ISD::UDIV: return visitUDIV(N); 508 case ISD::SREM: return visitSREM(N); 509 case ISD::UREM: return visitUREM(N); 510 case ISD::MULHU: return visitMULHU(N); 511 case ISD::MULHS: return visitMULHS(N); 512 case ISD::AND: return visitAND(N); 513 case ISD::OR: return visitOR(N); 514 case ISD::XOR: return visitXOR(N); 515 case ISD::SHL: return visitSHL(N); 516 case ISD::SRA: return visitSRA(N); 517 case ISD::SRL: return visitSRL(N); 518 case ISD::CTLZ: return visitCTLZ(N); 519 case ISD::CTTZ: return visitCTTZ(N); 520 case ISD::CTPOP: return visitCTPOP(N); 521 case ISD::SELECT: return visitSELECT(N); 522 case ISD::SELECT_CC: return visitSELECT_CC(N); 523 case ISD::SETCC: return visitSETCC(N); 524 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N); 525 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N); 526 case ISD::ANY_EXTEND: return visitANY_EXTEND(N); 527 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N); 528 case ISD::TRUNCATE: return visitTRUNCATE(N); 529 case ISD::BIT_CONVERT: return visitBIT_CONVERT(N); 530 case ISD::VBIT_CONVERT: return visitVBIT_CONVERT(N); 531 case ISD::FADD: return visitFADD(N); 532 case ISD::FSUB: return visitFSUB(N); 533 case ISD::FMUL: return visitFMUL(N); 534 case ISD::FDIV: return visitFDIV(N); 535 case ISD::FREM: return visitFREM(N); 536 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N); 537 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N); 538 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N); 539 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N); 540 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N); 541 case ISD::FP_ROUND: return visitFP_ROUND(N); 542 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N); 543 case ISD::FP_EXTEND: return visitFP_EXTEND(N); 544 case ISD::FNEG: return visitFNEG(N); 545 case ISD::FABS: return visitFABS(N); 546 case ISD::BRCOND: return visitBRCOND(N); 547 case ISD::BR_CC: return visitBR_CC(N); 548 case ISD::LOAD: return visitLOAD(N); 549 case ISD::STORE: return visitSTORE(N); 550 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N); 551 case ISD::VINSERT_VECTOR_ELT: return visitVINSERT_VECTOR_ELT(N); 552 case ISD::VBUILD_VECTOR: return visitVBUILD_VECTOR(N); 553 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N); 554 case ISD::VVECTOR_SHUFFLE: return visitVVECTOR_SHUFFLE(N); 555 case ISD::VADD: return visitVBinOp(N, ISD::ADD , ISD::FADD); 556 case ISD::VSUB: return visitVBinOp(N, ISD::SUB , ISD::FSUB); 557 case ISD::VMUL: return visitVBinOp(N, ISD::MUL , ISD::FMUL); 558 case ISD::VSDIV: return visitVBinOp(N, ISD::SDIV, ISD::FDIV); 559 case ISD::VUDIV: return visitVBinOp(N, ISD::UDIV, ISD::UDIV); 560 case ISD::VAND: return visitVBinOp(N, ISD::AND , ISD::AND); 561 case ISD::VOR: return visitVBinOp(N, ISD::OR , ISD::OR); 562 case ISD::VXOR: return visitVBinOp(N, ISD::XOR , ISD::XOR); 563 } 564 return SDOperand(); 565} 566 567/// getInputChainForNode - Given a node, return its input chain if it has one, 568/// otherwise return a null sd operand. 569static SDOperand getInputChainForNode(SDNode *N) { 570 if (unsigned NumOps = N->getNumOperands()) { 571 if (N->getOperand(0).getValueType() == MVT::Other) 572 return N->getOperand(0); 573 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other) 574 return N->getOperand(NumOps-1); 575 for (unsigned i = 1; i < NumOps-1; ++i) 576 if (N->getOperand(i).getValueType() == MVT::Other) 577 return N->getOperand(i); 578 } 579 return SDOperand(0, 0); 580} 581 582SDOperand DAGCombiner::visitTokenFactor(SDNode *N) { 583 // If N has two operands, where one has an input chain equal to the other, 584 // the 'other' chain is redundant. 585 if (N->getNumOperands() == 2) { 586 if (getInputChainForNode(N->getOperand(0).Val) == N->getOperand(1)) 587 return N->getOperand(0); 588 if (getInputChainForNode(N->getOperand(1).Val) == N->getOperand(0)) 589 return N->getOperand(1); 590 } 591 592 593 SmallVector<SDNode *, 8> TFs; // List of token factors to visit. 594 SmallVector<SDOperand, 8> Ops; // Ops for replacing token factor. 595 bool Changed = false; // If we should replace this token factor. 596 597 // Start out with this token factor. 598 TFs.push_back(N); 599 600 // Iterate through token factors. The TFs grows when new token factors are 601 // encountered. 602 for (unsigned i = 0; i < TFs.size(); ++i) { 603 SDNode *TF = TFs[i]; 604 605 // Check each of the operands. 606 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) { 607 SDOperand Op = TF->getOperand(i); 608 609 switch (Op.getOpcode()) { 610 case ISD::EntryToken: 611 // Entry tokens don't need to be added to the list. They are 612 // rededundant. 613 Changed = true; 614 break; 615 616 case ISD::TokenFactor: 617 if ((CombinerAA || Op.hasOneUse()) && 618 std::find(TFs.begin(), TFs.end(), Op.Val) == TFs.end()) { 619 // Queue up for processing. 620 TFs.push_back(Op.Val); 621 // Clean up in case the token factor is removed. 622 AddToWorkList(Op.Val); 623 Changed = true; 624 break; 625 } 626 // Fall thru 627 628 default: 629 // Only add if not there prior. 630 if (std::find(Ops.begin(), Ops.end(), Op) == Ops.end()) 631 Ops.push_back(Op); 632 break; 633 } 634 } 635 } 636 637 SDOperand Result; 638 639 // If we've change things around then replace token factor. 640 if (Changed) { 641 if (Ops.size() == 0) { 642 // The entry token is the only possible outcome. 643 Result = DAG.getEntryNode(); 644 } else { 645 // New and improved token factor. 646 Result = DAG.getNode(ISD::TokenFactor, MVT::Other, &Ops[0], Ops.size()); 647 } 648 649 // Don't add users to work list. 650 return CombineTo(N, Result, false); 651 } 652 653 return Result; 654} 655 656static 657SDOperand combineShlAddConstant(SDOperand N0, SDOperand N1, SelectionDAG &DAG) { 658 MVT::ValueType VT = N0.getValueType(); 659 SDOperand N00 = N0.getOperand(0); 660 SDOperand N01 = N0.getOperand(1); 661 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01); 662 if (N01C && N00.getOpcode() == ISD::ADD && N00.Val->hasOneUse() && 663 isa<ConstantSDNode>(N00.getOperand(1))) { 664 N0 = DAG.getNode(ISD::ADD, VT, 665 DAG.getNode(ISD::SHL, VT, N00.getOperand(0), N01), 666 DAG.getNode(ISD::SHL, VT, N00.getOperand(1), N01)); 667 return DAG.getNode(ISD::ADD, VT, N0, N1); 668 } 669 return SDOperand(); 670} 671 672SDOperand DAGCombiner::visitADD(SDNode *N) { 673 SDOperand N0 = N->getOperand(0); 674 SDOperand N1 = N->getOperand(1); 675 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 676 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 677 MVT::ValueType VT = N0.getValueType(); 678 679 // fold (add c1, c2) -> c1+c2 680 if (N0C && N1C) 681 return DAG.getNode(ISD::ADD, VT, N0, N1); 682 // canonicalize constant to RHS 683 if (N0C && !N1C) 684 return DAG.getNode(ISD::ADD, VT, N1, N0); 685 // fold (add x, 0) -> x 686 if (N1C && N1C->isNullValue()) 687 return N0; 688 // fold ((c1-A)+c2) -> (c1+c2)-A 689 if (N1C && N0.getOpcode() == ISD::SUB) 690 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0))) 691 return DAG.getNode(ISD::SUB, VT, 692 DAG.getConstant(N1C->getValue()+N0C->getValue(), VT), 693 N0.getOperand(1)); 694 // reassociate add 695 SDOperand RADD = ReassociateOps(ISD::ADD, N0, N1); 696 if (RADD.Val != 0) 697 return RADD; 698 // fold ((0-A) + B) -> B-A 699 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) && 700 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue()) 701 return DAG.getNode(ISD::SUB, VT, N1, N0.getOperand(1)); 702 // fold (A + (0-B)) -> A-B 703 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) && 704 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue()) 705 return DAG.getNode(ISD::SUB, VT, N0, N1.getOperand(1)); 706 // fold (A+(B-A)) -> B 707 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1)) 708 return N1.getOperand(0); 709 710 if (!MVT::isVector(VT) && SimplifyDemandedBits(SDOperand(N, 0))) 711 return SDOperand(N, 0); 712 713 // fold (a+b) -> (a|b) iff a and b share no bits. 714 if (MVT::isInteger(VT) && !MVT::isVector(VT)) { 715 uint64_t LHSZero, LHSOne; 716 uint64_t RHSZero, RHSOne; 717 uint64_t Mask = MVT::getIntVTBitMask(VT); 718 TLI.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne); 719 if (LHSZero) { 720 TLI.ComputeMaskedBits(N1, Mask, RHSZero, RHSOne); 721 722 // If all possibly-set bits on the LHS are clear on the RHS, return an OR. 723 // If all possibly-set bits on the RHS are clear on the LHS, return an OR. 724 if ((RHSZero & (~LHSZero & Mask)) == (~LHSZero & Mask) || 725 (LHSZero & (~RHSZero & Mask)) == (~RHSZero & Mask)) 726 return DAG.getNode(ISD::OR, VT, N0, N1); 727 } 728 } 729 730 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), ) 731 if (N0.getOpcode() == ISD::SHL && N0.Val->hasOneUse()) { 732 SDOperand Result = combineShlAddConstant(N0, N1, DAG); 733 if (Result.Val) return Result; 734 } 735 if (N1.getOpcode() == ISD::SHL && N1.Val->hasOneUse()) { 736 SDOperand Result = combineShlAddConstant(N1, N0, DAG); 737 if (Result.Val) return Result; 738 } 739 740 return SDOperand(); 741} 742 743SDOperand DAGCombiner::visitSUB(SDNode *N) { 744 SDOperand N0 = N->getOperand(0); 745 SDOperand N1 = N->getOperand(1); 746 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 747 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 748 MVT::ValueType VT = N0.getValueType(); 749 750 // fold (sub x, x) -> 0 751 if (N0 == N1) 752 return DAG.getConstant(0, N->getValueType(0)); 753 // fold (sub c1, c2) -> c1-c2 754 if (N0C && N1C) 755 return DAG.getNode(ISD::SUB, VT, N0, N1); 756 // fold (sub x, c) -> (add x, -c) 757 if (N1C) 758 return DAG.getNode(ISD::ADD, VT, N0, DAG.getConstant(-N1C->getValue(), VT)); 759 // fold (A+B)-A -> B 760 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1) 761 return N0.getOperand(1); 762 // fold (A+B)-B -> A 763 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1) 764 return N0.getOperand(0); 765 return SDOperand(); 766} 767 768SDOperand DAGCombiner::visitMUL(SDNode *N) { 769 SDOperand N0 = N->getOperand(0); 770 SDOperand N1 = N->getOperand(1); 771 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 772 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 773 MVT::ValueType VT = N0.getValueType(); 774 775 // fold (mul c1, c2) -> c1*c2 776 if (N0C && N1C) 777 return DAG.getNode(ISD::MUL, VT, N0, N1); 778 // canonicalize constant to RHS 779 if (N0C && !N1C) 780 return DAG.getNode(ISD::MUL, VT, N1, N0); 781 // fold (mul x, 0) -> 0 782 if (N1C && N1C->isNullValue()) 783 return N1; 784 // fold (mul x, -1) -> 0-x 785 if (N1C && N1C->isAllOnesValue()) 786 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0); 787 // fold (mul x, (1 << c)) -> x << c 788 if (N1C && isPowerOf2_64(N1C->getValue())) 789 return DAG.getNode(ISD::SHL, VT, N0, 790 DAG.getConstant(Log2_64(N1C->getValue()), 791 TLI.getShiftAmountTy())); 792 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c 793 if (N1C && isPowerOf2_64(-N1C->getSignExtended())) { 794 // FIXME: If the input is something that is easily negated (e.g. a 795 // single-use add), we should put the negate there. 796 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), 797 DAG.getNode(ISD::SHL, VT, N0, 798 DAG.getConstant(Log2_64(-N1C->getSignExtended()), 799 TLI.getShiftAmountTy()))); 800 } 801 802 // (mul (shl X, c1), c2) -> (mul X, c2 << c1) 803 if (N1C && N0.getOpcode() == ISD::SHL && 804 isa<ConstantSDNode>(N0.getOperand(1))) { 805 SDOperand C3 = DAG.getNode(ISD::SHL, VT, N1, N0.getOperand(1)); 806 AddToWorkList(C3.Val); 807 return DAG.getNode(ISD::MUL, VT, N0.getOperand(0), C3); 808 } 809 810 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one 811 // use. 812 { 813 SDOperand Sh(0,0), Y(0,0); 814 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)). 815 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) && 816 N0.Val->hasOneUse()) { 817 Sh = N0; Y = N1; 818 } else if (N1.getOpcode() == ISD::SHL && 819 isa<ConstantSDNode>(N1.getOperand(1)) && N1.Val->hasOneUse()) { 820 Sh = N1; Y = N0; 821 } 822 if (Sh.Val) { 823 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Sh.getOperand(0), Y); 824 return DAG.getNode(ISD::SHL, VT, Mul, Sh.getOperand(1)); 825 } 826 } 827 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2) 828 if (N1C && N0.getOpcode() == ISD::ADD && N0.Val->hasOneUse() && 829 isa<ConstantSDNode>(N0.getOperand(1))) { 830 return DAG.getNode(ISD::ADD, VT, 831 DAG.getNode(ISD::MUL, VT, N0.getOperand(0), N1), 832 DAG.getNode(ISD::MUL, VT, N0.getOperand(1), N1)); 833 } 834 835 // reassociate mul 836 SDOperand RMUL = ReassociateOps(ISD::MUL, N0, N1); 837 if (RMUL.Val != 0) 838 return RMUL; 839 return SDOperand(); 840} 841 842SDOperand DAGCombiner::visitSDIV(SDNode *N) { 843 SDOperand N0 = N->getOperand(0); 844 SDOperand N1 = N->getOperand(1); 845 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 846 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 847 MVT::ValueType VT = N->getValueType(0); 848 849 // fold (sdiv c1, c2) -> c1/c2 850 if (N0C && N1C && !N1C->isNullValue()) 851 return DAG.getNode(ISD::SDIV, VT, N0, N1); 852 // fold (sdiv X, 1) -> X 853 if (N1C && N1C->getSignExtended() == 1LL) 854 return N0; 855 // fold (sdiv X, -1) -> 0-X 856 if (N1C && N1C->isAllOnesValue()) 857 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), N0); 858 // If we know the sign bits of both operands are zero, strength reduce to a 859 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2 860 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1); 861 if (TLI.MaskedValueIsZero(N1, SignBit) && 862 TLI.MaskedValueIsZero(N0, SignBit)) 863 return DAG.getNode(ISD::UDIV, N1.getValueType(), N0, N1); 864 // fold (sdiv X, pow2) -> simple ops after legalize 865 if (N1C && N1C->getValue() && !TLI.isIntDivCheap() && 866 (isPowerOf2_64(N1C->getSignExtended()) || 867 isPowerOf2_64(-N1C->getSignExtended()))) { 868 // If dividing by powers of two is cheap, then don't perform the following 869 // fold. 870 if (TLI.isPow2DivCheap()) 871 return SDOperand(); 872 int64_t pow2 = N1C->getSignExtended(); 873 int64_t abs2 = pow2 > 0 ? pow2 : -pow2; 874 unsigned lg2 = Log2_64(abs2); 875 // Splat the sign bit into the register 876 SDOperand SGN = DAG.getNode(ISD::SRA, VT, N0, 877 DAG.getConstant(MVT::getSizeInBits(VT)-1, 878 TLI.getShiftAmountTy())); 879 AddToWorkList(SGN.Val); 880 // Add (N0 < 0) ? abs2 - 1 : 0; 881 SDOperand SRL = DAG.getNode(ISD::SRL, VT, SGN, 882 DAG.getConstant(MVT::getSizeInBits(VT)-lg2, 883 TLI.getShiftAmountTy())); 884 SDOperand ADD = DAG.getNode(ISD::ADD, VT, N0, SRL); 885 AddToWorkList(SRL.Val); 886 AddToWorkList(ADD.Val); // Divide by pow2 887 SDOperand SRA = DAG.getNode(ISD::SRA, VT, ADD, 888 DAG.getConstant(lg2, TLI.getShiftAmountTy())); 889 // If we're dividing by a positive value, we're done. Otherwise, we must 890 // negate the result. 891 if (pow2 > 0) 892 return SRA; 893 AddToWorkList(SRA.Val); 894 return DAG.getNode(ISD::SUB, VT, DAG.getConstant(0, VT), SRA); 895 } 896 // if integer divide is expensive and we satisfy the requirements, emit an 897 // alternate sequence. 898 if (N1C && (N1C->getSignExtended() < -1 || N1C->getSignExtended() > 1) && 899 !TLI.isIntDivCheap()) { 900 SDOperand Op = BuildSDIV(N); 901 if (Op.Val) return Op; 902 } 903 return SDOperand(); 904} 905 906SDOperand DAGCombiner::visitUDIV(SDNode *N) { 907 SDOperand N0 = N->getOperand(0); 908 SDOperand N1 = N->getOperand(1); 909 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.Val); 910 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 911 MVT::ValueType VT = N->getValueType(0); 912 913 // fold (udiv c1, c2) -> c1/c2 914 if (N0C && N1C && !N1C->isNullValue()) 915 return DAG.getNode(ISD::UDIV, VT, N0, N1); 916 // fold (udiv x, (1 << c)) -> x >>u c 917 if (N1C && isPowerOf2_64(N1C->getValue())) 918 return DAG.getNode(ISD::SRL, VT, N0, 919 DAG.getConstant(Log2_64(N1C->getValue()), 920 TLI.getShiftAmountTy())); 921 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2 922 if (N1.getOpcode() == ISD::SHL) { 923 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) { 924 if (isPowerOf2_64(SHC->getValue())) { 925 MVT::ValueType ADDVT = N1.getOperand(1).getValueType(); 926 SDOperand Add = DAG.getNode(ISD::ADD, ADDVT, N1.getOperand(1), 927 DAG.getConstant(Log2_64(SHC->getValue()), 928 ADDVT)); 929 AddToWorkList(Add.Val); 930 return DAG.getNode(ISD::SRL, VT, N0, Add); 931 } 932 } 933 } 934 // fold (udiv x, c) -> alternate 935 if (N1C && N1C->getValue() && !TLI.isIntDivCheap()) { 936 SDOperand Op = BuildUDIV(N); 937 if (Op.Val) return Op; 938 } 939 return SDOperand(); 940} 941 942SDOperand DAGCombiner::visitSREM(SDNode *N) { 943 SDOperand N0 = N->getOperand(0); 944 SDOperand N1 = N->getOperand(1); 945 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 946 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 947 MVT::ValueType VT = N->getValueType(0); 948 949 // fold (srem c1, c2) -> c1%c2 950 if (N0C && N1C && !N1C->isNullValue()) 951 return DAG.getNode(ISD::SREM, VT, N0, N1); 952 // If we know the sign bits of both operands are zero, strength reduce to a 953 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15 954 uint64_t SignBit = 1ULL << (MVT::getSizeInBits(VT)-1); 955 if (TLI.MaskedValueIsZero(N1, SignBit) && 956 TLI.MaskedValueIsZero(N0, SignBit)) 957 return DAG.getNode(ISD::UREM, VT, N0, N1); 958 959 // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on 960 // the remainder operation. 961 if (N1C && !N1C->isNullValue()) { 962 SDOperand Div = DAG.getNode(ISD::SDIV, VT, N0, N1); 963 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1); 964 SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul); 965 AddToWorkList(Div.Val); 966 AddToWorkList(Mul.Val); 967 return Sub; 968 } 969 970 return SDOperand(); 971} 972 973SDOperand DAGCombiner::visitUREM(SDNode *N) { 974 SDOperand N0 = N->getOperand(0); 975 SDOperand N1 = N->getOperand(1); 976 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 977 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 978 MVT::ValueType VT = N->getValueType(0); 979 980 // fold (urem c1, c2) -> c1%c2 981 if (N0C && N1C && !N1C->isNullValue()) 982 return DAG.getNode(ISD::UREM, VT, N0, N1); 983 // fold (urem x, pow2) -> (and x, pow2-1) 984 if (N1C && !N1C->isNullValue() && isPowerOf2_64(N1C->getValue())) 985 return DAG.getNode(ISD::AND, VT, N0, DAG.getConstant(N1C->getValue()-1,VT)); 986 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1)) 987 if (N1.getOpcode() == ISD::SHL) { 988 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) { 989 if (isPowerOf2_64(SHC->getValue())) { 990 SDOperand Add = DAG.getNode(ISD::ADD, VT, N1,DAG.getConstant(~0ULL,VT)); 991 AddToWorkList(Add.Val); 992 return DAG.getNode(ISD::AND, VT, N0, Add); 993 } 994 } 995 } 996 997 // Unconditionally lower X%C -> X-X/C*C. This allows the X/C logic to hack on 998 // the remainder operation. 999 if (N1C && !N1C->isNullValue()) { 1000 SDOperand Div = DAG.getNode(ISD::UDIV, VT, N0, N1); 1001 SDOperand Mul = DAG.getNode(ISD::MUL, VT, Div, N1); 1002 SDOperand Sub = DAG.getNode(ISD::SUB, VT, N0, Mul); 1003 AddToWorkList(Div.Val); 1004 AddToWorkList(Mul.Val); 1005 return Sub; 1006 } 1007 1008 return SDOperand(); 1009} 1010 1011SDOperand DAGCombiner::visitMULHS(SDNode *N) { 1012 SDOperand N0 = N->getOperand(0); 1013 SDOperand N1 = N->getOperand(1); 1014 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1015 1016 // fold (mulhs x, 0) -> 0 1017 if (N1C && N1C->isNullValue()) 1018 return N1; 1019 // fold (mulhs x, 1) -> (sra x, size(x)-1) 1020 if (N1C && N1C->getValue() == 1) 1021 return DAG.getNode(ISD::SRA, N0.getValueType(), N0, 1022 DAG.getConstant(MVT::getSizeInBits(N0.getValueType())-1, 1023 TLI.getShiftAmountTy())); 1024 return SDOperand(); 1025} 1026 1027SDOperand DAGCombiner::visitMULHU(SDNode *N) { 1028 SDOperand N0 = N->getOperand(0); 1029 SDOperand N1 = N->getOperand(1); 1030 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1031 1032 // fold (mulhu x, 0) -> 0 1033 if (N1C && N1C->isNullValue()) 1034 return N1; 1035 // fold (mulhu x, 1) -> 0 1036 if (N1C && N1C->getValue() == 1) 1037 return DAG.getConstant(0, N0.getValueType()); 1038 return SDOperand(); 1039} 1040 1041/// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with 1042/// two operands of the same opcode, try to simplify it. 1043SDOperand DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) { 1044 SDOperand N0 = N->getOperand(0), N1 = N->getOperand(1); 1045 MVT::ValueType VT = N0.getValueType(); 1046 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!"); 1047 1048 // For each of OP in AND/OR/XOR: 1049 // fold (OP (zext x), (zext y)) -> (zext (OP x, y)) 1050 // fold (OP (sext x), (sext y)) -> (sext (OP x, y)) 1051 // fold (OP (aext x), (aext y)) -> (aext (OP x, y)) 1052 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) 1053 if ((N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND|| 1054 N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::TRUNCATE) && 1055 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()) { 1056 SDOperand ORNode = DAG.getNode(N->getOpcode(), 1057 N0.getOperand(0).getValueType(), 1058 N0.getOperand(0), N1.getOperand(0)); 1059 AddToWorkList(ORNode.Val); 1060 return DAG.getNode(N0.getOpcode(), VT, ORNode); 1061 } 1062 1063 // For each of OP in SHL/SRL/SRA/AND... 1064 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z) 1065 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z) 1066 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z) 1067 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL || 1068 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) && 1069 N0.getOperand(1) == N1.getOperand(1)) { 1070 SDOperand ORNode = DAG.getNode(N->getOpcode(), 1071 N0.getOperand(0).getValueType(), 1072 N0.getOperand(0), N1.getOperand(0)); 1073 AddToWorkList(ORNode.Val); 1074 return DAG.getNode(N0.getOpcode(), VT, ORNode, N0.getOperand(1)); 1075 } 1076 1077 return SDOperand(); 1078} 1079 1080SDOperand DAGCombiner::visitAND(SDNode *N) { 1081 SDOperand N0 = N->getOperand(0); 1082 SDOperand N1 = N->getOperand(1); 1083 SDOperand LL, LR, RL, RR, CC0, CC1; 1084 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1085 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1086 MVT::ValueType VT = N1.getValueType(); 1087 1088 // fold (and c1, c2) -> c1&c2 1089 if (N0C && N1C) 1090 return DAG.getNode(ISD::AND, VT, N0, N1); 1091 // canonicalize constant to RHS 1092 if (N0C && !N1C) 1093 return DAG.getNode(ISD::AND, VT, N1, N0); 1094 // fold (and x, -1) -> x 1095 if (N1C && N1C->isAllOnesValue()) 1096 return N0; 1097 // if (and x, c) is known to be zero, return 0 1098 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT))) 1099 return DAG.getConstant(0, VT); 1100 // reassociate and 1101 SDOperand RAND = ReassociateOps(ISD::AND, N0, N1); 1102 if (RAND.Val != 0) 1103 return RAND; 1104 // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF 1105 if (N1C && N0.getOpcode() == ISD::OR) 1106 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 1107 if ((ORI->getValue() & N1C->getValue()) == N1C->getValue()) 1108 return N1; 1109 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits. 1110 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { 1111 unsigned InMask = MVT::getIntVTBitMask(N0.getOperand(0).getValueType()); 1112 if (TLI.MaskedValueIsZero(N0.getOperand(0), 1113 ~N1C->getValue() & InMask)) { 1114 SDOperand Zext = DAG.getNode(ISD::ZERO_EXTEND, N0.getValueType(), 1115 N0.getOperand(0)); 1116 1117 // Replace uses of the AND with uses of the Zero extend node. 1118 CombineTo(N, Zext); 1119 1120 // We actually want to replace all uses of the any_extend with the 1121 // zero_extend, to avoid duplicating things. This will later cause this 1122 // AND to be folded. 1123 CombineTo(N0.Val, Zext); 1124 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1125 } 1126 } 1127 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y)) 1128 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 1129 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 1130 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 1131 1132 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 1133 MVT::isInteger(LL.getValueType())) { 1134 // fold (X == 0) & (Y == 0) -> (X|Y == 0) 1135 if (cast<ConstantSDNode>(LR)->getValue() == 0 && Op1 == ISD::SETEQ) { 1136 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL); 1137 AddToWorkList(ORNode.Val); 1138 return DAG.getSetCC(VT, ORNode, LR, Op1); 1139 } 1140 // fold (X == -1) & (Y == -1) -> (X&Y == -1) 1141 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) { 1142 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL); 1143 AddToWorkList(ANDNode.Val); 1144 return DAG.getSetCC(VT, ANDNode, LR, Op1); 1145 } 1146 // fold (X > -1) & (Y > -1) -> (X|Y > -1) 1147 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) { 1148 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL); 1149 AddToWorkList(ORNode.Val); 1150 return DAG.getSetCC(VT, ORNode, LR, Op1); 1151 } 1152 } 1153 // canonicalize equivalent to ll == rl 1154 if (LL == RR && LR == RL) { 1155 Op1 = ISD::getSetCCSwappedOperands(Op1); 1156 std::swap(RL, RR); 1157 } 1158 if (LL == RL && LR == RR) { 1159 bool isInteger = MVT::isInteger(LL.getValueType()); 1160 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger); 1161 if (Result != ISD::SETCC_INVALID) 1162 return DAG.getSetCC(N0.getValueType(), LL, LR, Result); 1163 } 1164 } 1165 1166 // Simplify: and (op x...), (op y...) -> (op (and x, y)) 1167 if (N0.getOpcode() == N1.getOpcode()) { 1168 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N); 1169 if (Tmp.Val) return Tmp; 1170 } 1171 1172 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1) 1173 // fold (and (sra)) -> (and (srl)) when possible. 1174 if (!MVT::isVector(VT) && 1175 SimplifyDemandedBits(SDOperand(N, 0))) 1176 return SDOperand(N, 0); 1177 // fold (zext_inreg (extload x)) -> (zextload x) 1178 if (ISD::isEXTLoad(N0.Val)) { 1179 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 1180 MVT::ValueType EVT = LN0->getLoadedVT(); 1181 // If we zero all the possible extended bits, then we can turn this into 1182 // a zextload if we are running before legalize or the operation is legal. 1183 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) && 1184 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) { 1185 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(), 1186 LN0->getBasePtr(), LN0->getSrcValue(), 1187 LN0->getSrcValueOffset(), EVT); 1188 AddToWorkList(N); 1189 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1190 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1191 } 1192 } 1193 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use 1194 if (ISD::isSEXTLoad(N0.Val) && N0.hasOneUse()) { 1195 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 1196 MVT::ValueType EVT = LN0->getLoadedVT(); 1197 // If we zero all the possible extended bits, then we can turn this into 1198 // a zextload if we are running before legalize or the operation is legal. 1199 if (TLI.MaskedValueIsZero(N1, ~0ULL << MVT::getSizeInBits(EVT)) && 1200 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) { 1201 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(), 1202 LN0->getBasePtr(), LN0->getSrcValue(), 1203 LN0->getSrcValueOffset(), EVT); 1204 AddToWorkList(N); 1205 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 1206 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1207 } 1208 } 1209 1210 // fold (and (load x), 255) -> (zextload x, i8) 1211 // fold (and (extload x, i16), 255) -> (zextload x, i8) 1212 if (N1C && N0.getOpcode() == ISD::LOAD) { 1213 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 1214 if (LN0->getExtensionType() != ISD::SEXTLOAD && 1215 N0.hasOneUse()) { 1216 MVT::ValueType EVT, LoadedVT; 1217 if (N1C->getValue() == 255) 1218 EVT = MVT::i8; 1219 else if (N1C->getValue() == 65535) 1220 EVT = MVT::i16; 1221 else if (N1C->getValue() == ~0U) 1222 EVT = MVT::i32; 1223 else 1224 EVT = MVT::Other; 1225 1226 LoadedVT = LN0->getLoadedVT(); 1227 if (EVT != MVT::Other && LoadedVT > EVT && 1228 (!AfterLegalize || TLI.isLoadXLegal(ISD::ZEXTLOAD, EVT))) { 1229 MVT::ValueType PtrType = N0.getOperand(1).getValueType(); 1230 // For big endian targets, we need to add an offset to the pointer to 1231 // load the correct bytes. For little endian systems, we merely need to 1232 // read fewer bytes from the same pointer. 1233 unsigned PtrOff = 1234 (MVT::getSizeInBits(LoadedVT) - MVT::getSizeInBits(EVT)) / 8; 1235 SDOperand NewPtr = LN0->getBasePtr(); 1236 if (!TLI.isLittleEndian()) 1237 NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr, 1238 DAG.getConstant(PtrOff, PtrType)); 1239 AddToWorkList(NewPtr.Val); 1240 SDOperand Load = 1241 DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(), NewPtr, 1242 LN0->getSrcValue(), LN0->getSrcValueOffset(), EVT); 1243 AddToWorkList(N); 1244 CombineTo(N0.Val, Load, Load.getValue(1)); 1245 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1246 } 1247 } 1248 } 1249 1250 return SDOperand(); 1251} 1252 1253SDOperand DAGCombiner::visitOR(SDNode *N) { 1254 SDOperand N0 = N->getOperand(0); 1255 SDOperand N1 = N->getOperand(1); 1256 SDOperand LL, LR, RL, RR, CC0, CC1; 1257 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1258 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1259 MVT::ValueType VT = N1.getValueType(); 1260 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1261 1262 // fold (or c1, c2) -> c1|c2 1263 if (N0C && N1C) 1264 return DAG.getNode(ISD::OR, VT, N0, N1); 1265 // canonicalize constant to RHS 1266 if (N0C && !N1C) 1267 return DAG.getNode(ISD::OR, VT, N1, N0); 1268 // fold (or x, 0) -> x 1269 if (N1C && N1C->isNullValue()) 1270 return N0; 1271 // fold (or x, -1) -> -1 1272 if (N1C && N1C->isAllOnesValue()) 1273 return N1; 1274 // fold (or x, c) -> c iff (x & ~c) == 0 1275 if (N1C && 1276 TLI.MaskedValueIsZero(N0,~N1C->getValue() & (~0ULL>>(64-OpSizeInBits)))) 1277 return N1; 1278 // reassociate or 1279 SDOperand ROR = ReassociateOps(ISD::OR, N0, N1); 1280 if (ROR.Val != 0) 1281 return ROR; 1282 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2) 1283 if (N1C && N0.getOpcode() == ISD::AND && N0.Val->hasOneUse() && 1284 isa<ConstantSDNode>(N0.getOperand(1))) { 1285 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1)); 1286 return DAG.getNode(ISD::AND, VT, DAG.getNode(ISD::OR, VT, N0.getOperand(0), 1287 N1), 1288 DAG.getConstant(N1C->getValue() | C1->getValue(), VT)); 1289 } 1290 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y)) 1291 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ 1292 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get(); 1293 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get(); 1294 1295 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 && 1296 MVT::isInteger(LL.getValueType())) { 1297 // fold (X != 0) | (Y != 0) -> (X|Y != 0) 1298 // fold (X < 0) | (Y < 0) -> (X|Y < 0) 1299 if (cast<ConstantSDNode>(LR)->getValue() == 0 && 1300 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) { 1301 SDOperand ORNode = DAG.getNode(ISD::OR, LR.getValueType(), LL, RL); 1302 AddToWorkList(ORNode.Val); 1303 return DAG.getSetCC(VT, ORNode, LR, Op1); 1304 } 1305 // fold (X != -1) | (Y != -1) -> (X&Y != -1) 1306 // fold (X > -1) | (Y > -1) -> (X&Y > -1) 1307 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && 1308 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) { 1309 SDOperand ANDNode = DAG.getNode(ISD::AND, LR.getValueType(), LL, RL); 1310 AddToWorkList(ANDNode.Val); 1311 return DAG.getSetCC(VT, ANDNode, LR, Op1); 1312 } 1313 } 1314 // canonicalize equivalent to ll == rl 1315 if (LL == RR && LR == RL) { 1316 Op1 = ISD::getSetCCSwappedOperands(Op1); 1317 std::swap(RL, RR); 1318 } 1319 if (LL == RL && LR == RR) { 1320 bool isInteger = MVT::isInteger(LL.getValueType()); 1321 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger); 1322 if (Result != ISD::SETCC_INVALID) 1323 return DAG.getSetCC(N0.getValueType(), LL, LR, Result); 1324 } 1325 } 1326 1327 // Simplify: or (op x...), (op y...) -> (op (or x, y)) 1328 if (N0.getOpcode() == N1.getOpcode()) { 1329 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N); 1330 if (Tmp.Val) return Tmp; 1331 } 1332 1333 // (X & C1) | (Y & C2) -> (X|Y) & C3 if possible. 1334 if (N0.getOpcode() == ISD::AND && 1335 N1.getOpcode() == ISD::AND && 1336 N0.getOperand(1).getOpcode() == ISD::Constant && 1337 N1.getOperand(1).getOpcode() == ISD::Constant && 1338 // Don't increase # computations. 1339 (N0.Val->hasOneUse() || N1.Val->hasOneUse())) { 1340 // We can only do this xform if we know that bits from X that are set in C2 1341 // but not in C1 are already zero. Likewise for Y. 1342 uint64_t LHSMask = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1343 uint64_t RHSMask = cast<ConstantSDNode>(N1.getOperand(1))->getValue(); 1344 1345 if (TLI.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) && 1346 TLI.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) { 1347 SDOperand X =DAG.getNode(ISD::OR, VT, N0.getOperand(0), N1.getOperand(0)); 1348 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(LHSMask|RHSMask, VT)); 1349 } 1350 } 1351 1352 1353 // See if this is some rotate idiom. 1354 if (SDNode *Rot = MatchRotate(N0, N1)) 1355 return SDOperand(Rot, 0); 1356 1357 return SDOperand(); 1358} 1359 1360 1361/// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present. 1362static bool MatchRotateHalf(SDOperand Op, SDOperand &Shift, SDOperand &Mask) { 1363 if (Op.getOpcode() == ISD::AND) { 1364 if (isa<ConstantSDNode>(Op.getOperand(1))) { 1365 Mask = Op.getOperand(1); 1366 Op = Op.getOperand(0); 1367 } else { 1368 return false; 1369 } 1370 } 1371 1372 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) { 1373 Shift = Op; 1374 return true; 1375 } 1376 return false; 1377} 1378 1379 1380// MatchRotate - Handle an 'or' of two operands. If this is one of the many 1381// idioms for rotate, and if the target supports rotation instructions, generate 1382// a rot[lr]. 1383SDNode *DAGCombiner::MatchRotate(SDOperand LHS, SDOperand RHS) { 1384 // Must be a legal type. Expanded an promoted things won't work with rotates. 1385 MVT::ValueType VT = LHS.getValueType(); 1386 if (!TLI.isTypeLegal(VT)) return 0; 1387 1388 // The target must have at least one rotate flavor. 1389 bool HasROTL = TLI.isOperationLegal(ISD::ROTL, VT); 1390 bool HasROTR = TLI.isOperationLegal(ISD::ROTR, VT); 1391 if (!HasROTL && !HasROTR) return 0; 1392 1393 // Match "(X shl/srl V1) & V2" where V2 may not be present. 1394 SDOperand LHSShift; // The shift. 1395 SDOperand LHSMask; // AND value if any. 1396 if (!MatchRotateHalf(LHS, LHSShift, LHSMask)) 1397 return 0; // Not part of a rotate. 1398 1399 SDOperand RHSShift; // The shift. 1400 SDOperand RHSMask; // AND value if any. 1401 if (!MatchRotateHalf(RHS, RHSShift, RHSMask)) 1402 return 0; // Not part of a rotate. 1403 1404 if (LHSShift.getOperand(0) != RHSShift.getOperand(0)) 1405 return 0; // Not shifting the same value. 1406 1407 if (LHSShift.getOpcode() == RHSShift.getOpcode()) 1408 return 0; // Shifts must disagree. 1409 1410 // Canonicalize shl to left side in a shl/srl pair. 1411 if (RHSShift.getOpcode() == ISD::SHL) { 1412 std::swap(LHS, RHS); 1413 std::swap(LHSShift, RHSShift); 1414 std::swap(LHSMask , RHSMask ); 1415 } 1416 1417 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1418 1419 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1) 1420 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2) 1421 if (LHSShift.getOperand(1).getOpcode() == ISD::Constant && 1422 RHSShift.getOperand(1).getOpcode() == ISD::Constant) { 1423 uint64_t LShVal = cast<ConstantSDNode>(LHSShift.getOperand(1))->getValue(); 1424 uint64_t RShVal = cast<ConstantSDNode>(RHSShift.getOperand(1))->getValue(); 1425 if ((LShVal + RShVal) != OpSizeInBits) 1426 return 0; 1427 1428 SDOperand Rot; 1429 if (HasROTL) 1430 Rot = DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0), 1431 LHSShift.getOperand(1)); 1432 else 1433 Rot = DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0), 1434 RHSShift.getOperand(1)); 1435 1436 // If there is an AND of either shifted operand, apply it to the result. 1437 if (LHSMask.Val || RHSMask.Val) { 1438 uint64_t Mask = MVT::getIntVTBitMask(VT); 1439 1440 if (LHSMask.Val) { 1441 uint64_t RHSBits = (1ULL << LShVal)-1; 1442 Mask &= cast<ConstantSDNode>(LHSMask)->getValue() | RHSBits; 1443 } 1444 if (RHSMask.Val) { 1445 uint64_t LHSBits = ~((1ULL << (OpSizeInBits-RShVal))-1); 1446 Mask &= cast<ConstantSDNode>(RHSMask)->getValue() | LHSBits; 1447 } 1448 1449 Rot = DAG.getNode(ISD::AND, VT, Rot, DAG.getConstant(Mask, VT)); 1450 } 1451 1452 return Rot.Val; 1453 } 1454 1455 // If there is a mask here, and we have a variable shift, we can't be sure 1456 // that we're masking out the right stuff. 1457 if (LHSMask.Val || RHSMask.Val) 1458 return 0; 1459 1460 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y) 1461 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y)) 1462 if (RHSShift.getOperand(1).getOpcode() == ISD::SUB && 1463 LHSShift.getOperand(1) == RHSShift.getOperand(1).getOperand(1)) { 1464 if (ConstantSDNode *SUBC = 1465 dyn_cast<ConstantSDNode>(RHSShift.getOperand(1).getOperand(0))) { 1466 if (SUBC->getValue() == OpSizeInBits) 1467 if (HasROTL) 1468 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0), 1469 LHSShift.getOperand(1)).Val; 1470 else 1471 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0), 1472 LHSShift.getOperand(1)).Val; 1473 } 1474 } 1475 1476 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y) 1477 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y)) 1478 if (LHSShift.getOperand(1).getOpcode() == ISD::SUB && 1479 RHSShift.getOperand(1) == LHSShift.getOperand(1).getOperand(1)) { 1480 if (ConstantSDNode *SUBC = 1481 dyn_cast<ConstantSDNode>(LHSShift.getOperand(1).getOperand(0))) { 1482 if (SUBC->getValue() == OpSizeInBits) 1483 if (HasROTL) 1484 return DAG.getNode(ISD::ROTL, VT, LHSShift.getOperand(0), 1485 LHSShift.getOperand(1)).Val; 1486 else 1487 return DAG.getNode(ISD::ROTR, VT, LHSShift.getOperand(0), 1488 RHSShift.getOperand(1)).Val; 1489 } 1490 } 1491 1492 return 0; 1493} 1494 1495 1496SDOperand DAGCombiner::visitXOR(SDNode *N) { 1497 SDOperand N0 = N->getOperand(0); 1498 SDOperand N1 = N->getOperand(1); 1499 SDOperand LHS, RHS, CC; 1500 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1501 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1502 MVT::ValueType VT = N0.getValueType(); 1503 1504 // fold (xor c1, c2) -> c1^c2 1505 if (N0C && N1C) 1506 return DAG.getNode(ISD::XOR, VT, N0, N1); 1507 // canonicalize constant to RHS 1508 if (N0C && !N1C) 1509 return DAG.getNode(ISD::XOR, VT, N1, N0); 1510 // fold (xor x, 0) -> x 1511 if (N1C && N1C->isNullValue()) 1512 return N0; 1513 // reassociate xor 1514 SDOperand RXOR = ReassociateOps(ISD::XOR, N0, N1); 1515 if (RXOR.Val != 0) 1516 return RXOR; 1517 // fold !(x cc y) -> (x !cc y) 1518 if (N1C && N1C->getValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) { 1519 bool isInt = MVT::isInteger(LHS.getValueType()); 1520 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(), 1521 isInt); 1522 if (N0.getOpcode() == ISD::SETCC) 1523 return DAG.getSetCC(VT, LHS, RHS, NotCC); 1524 if (N0.getOpcode() == ISD::SELECT_CC) 1525 return DAG.getSelectCC(LHS, RHS, N0.getOperand(2),N0.getOperand(3),NotCC); 1526 assert(0 && "Unhandled SetCC Equivalent!"); 1527 abort(); 1528 } 1529 // fold !(x or y) -> (!x and !y) iff x or y are setcc 1530 if (N1C && N1C->getValue() == 1 && VT == MVT::i1 && 1531 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 1532 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1); 1533 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) { 1534 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 1535 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS 1536 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS 1537 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val); 1538 return DAG.getNode(NewOpcode, VT, LHS, RHS); 1539 } 1540 } 1541 // fold !(x or y) -> (!x and !y) iff x or y are constants 1542 if (N1C && N1C->isAllOnesValue() && 1543 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) { 1544 SDOperand LHS = N0.getOperand(0), RHS = N0.getOperand(1); 1545 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) { 1546 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND; 1547 LHS = DAG.getNode(ISD::XOR, VT, LHS, N1); // RHS = ~LHS 1548 RHS = DAG.getNode(ISD::XOR, VT, RHS, N1); // RHS = ~RHS 1549 AddToWorkList(LHS.Val); AddToWorkList(RHS.Val); 1550 return DAG.getNode(NewOpcode, VT, LHS, RHS); 1551 } 1552 } 1553 // fold (xor (xor x, c1), c2) -> (xor x, c1^c2) 1554 if (N1C && N0.getOpcode() == ISD::XOR) { 1555 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0)); 1556 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1)); 1557 if (N00C) 1558 return DAG.getNode(ISD::XOR, VT, N0.getOperand(1), 1559 DAG.getConstant(N1C->getValue()^N00C->getValue(), VT)); 1560 if (N01C) 1561 return DAG.getNode(ISD::XOR, VT, N0.getOperand(0), 1562 DAG.getConstant(N1C->getValue()^N01C->getValue(), VT)); 1563 } 1564 // fold (xor x, x) -> 0 1565 if (N0 == N1) { 1566 if (!MVT::isVector(VT)) { 1567 return DAG.getConstant(0, VT); 1568 } else if (!AfterLegalize || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) { 1569 // Produce a vector of zeros. 1570 SDOperand El = DAG.getConstant(0, MVT::getVectorBaseType(VT)); 1571 std::vector<SDOperand> Ops(MVT::getVectorNumElements(VT), El); 1572 return DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size()); 1573 } 1574 } 1575 1576 // Simplify: xor (op x...), (op y...) -> (op (xor x, y)) 1577 if (N0.getOpcode() == N1.getOpcode()) { 1578 SDOperand Tmp = SimplifyBinOpWithSameOpcodeHands(N); 1579 if (Tmp.Val) return Tmp; 1580 } 1581 1582 // Simplify the expression using non-local knowledge. 1583 if (!MVT::isVector(VT) && 1584 SimplifyDemandedBits(SDOperand(N, 0))) 1585 return SDOperand(N, 0); 1586 1587 return SDOperand(); 1588} 1589 1590SDOperand DAGCombiner::visitSHL(SDNode *N) { 1591 SDOperand N0 = N->getOperand(0); 1592 SDOperand N1 = N->getOperand(1); 1593 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1594 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1595 MVT::ValueType VT = N0.getValueType(); 1596 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1597 1598 // fold (shl c1, c2) -> c1<<c2 1599 if (N0C && N1C) 1600 return DAG.getNode(ISD::SHL, VT, N0, N1); 1601 // fold (shl 0, x) -> 0 1602 if (N0C && N0C->isNullValue()) 1603 return N0; 1604 // fold (shl x, c >= size(x)) -> undef 1605 if (N1C && N1C->getValue() >= OpSizeInBits) 1606 return DAG.getNode(ISD::UNDEF, VT); 1607 // fold (shl x, 0) -> x 1608 if (N1C && N1C->isNullValue()) 1609 return N0; 1610 // if (shl x, c) is known to be zero, return 0 1611 if (TLI.MaskedValueIsZero(SDOperand(N, 0), MVT::getIntVTBitMask(VT))) 1612 return DAG.getConstant(0, VT); 1613 if (SimplifyDemandedBits(SDOperand(N, 0))) 1614 return SDOperand(N, 0); 1615 // fold (shl (shl x, c1), c2) -> 0 or (shl x, c1+c2) 1616 if (N1C && N0.getOpcode() == ISD::SHL && 1617 N0.getOperand(1).getOpcode() == ISD::Constant) { 1618 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1619 uint64_t c2 = N1C->getValue(); 1620 if (c1 + c2 > OpSizeInBits) 1621 return DAG.getConstant(0, VT); 1622 return DAG.getNode(ISD::SHL, VT, N0.getOperand(0), 1623 DAG.getConstant(c1 + c2, N1.getValueType())); 1624 } 1625 // fold (shl (srl x, c1), c2) -> (shl (and x, -1 << c1), c2-c1) or 1626 // (srl (and x, -1 << c1), c1-c2) 1627 if (N1C && N0.getOpcode() == ISD::SRL && 1628 N0.getOperand(1).getOpcode() == ISD::Constant) { 1629 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1630 uint64_t c2 = N1C->getValue(); 1631 SDOperand Mask = DAG.getNode(ISD::AND, VT, N0.getOperand(0), 1632 DAG.getConstant(~0ULL << c1, VT)); 1633 if (c2 > c1) 1634 return DAG.getNode(ISD::SHL, VT, Mask, 1635 DAG.getConstant(c2-c1, N1.getValueType())); 1636 else 1637 return DAG.getNode(ISD::SRL, VT, Mask, 1638 DAG.getConstant(c1-c2, N1.getValueType())); 1639 } 1640 // fold (shl (sra x, c1), c1) -> (and x, -1 << c1) 1641 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) 1642 return DAG.getNode(ISD::AND, VT, N0.getOperand(0), 1643 DAG.getConstant(~0ULL << N1C->getValue(), VT)); 1644 return SDOperand(); 1645} 1646 1647SDOperand DAGCombiner::visitSRA(SDNode *N) { 1648 SDOperand N0 = N->getOperand(0); 1649 SDOperand N1 = N->getOperand(1); 1650 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1651 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1652 MVT::ValueType VT = N0.getValueType(); 1653 1654 // fold (sra c1, c2) -> c1>>c2 1655 if (N0C && N1C) 1656 return DAG.getNode(ISD::SRA, VT, N0, N1); 1657 // fold (sra 0, x) -> 0 1658 if (N0C && N0C->isNullValue()) 1659 return N0; 1660 // fold (sra -1, x) -> -1 1661 if (N0C && N0C->isAllOnesValue()) 1662 return N0; 1663 // fold (sra x, c >= size(x)) -> undef 1664 if (N1C && N1C->getValue() >= MVT::getSizeInBits(VT)) 1665 return DAG.getNode(ISD::UNDEF, VT); 1666 // fold (sra x, 0) -> x 1667 if (N1C && N1C->isNullValue()) 1668 return N0; 1669 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports 1670 // sext_inreg. 1671 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) { 1672 unsigned LowBits = MVT::getSizeInBits(VT) - (unsigned)N1C->getValue(); 1673 MVT::ValueType EVT; 1674 switch (LowBits) { 1675 default: EVT = MVT::Other; break; 1676 case 1: EVT = MVT::i1; break; 1677 case 8: EVT = MVT::i8; break; 1678 case 16: EVT = MVT::i16; break; 1679 case 32: EVT = MVT::i32; break; 1680 } 1681 if (EVT > MVT::Other && TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, EVT)) 1682 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), 1683 DAG.getValueType(EVT)); 1684 } 1685 1686 // fold (sra (sra x, c1), c2) -> (sra x, c1+c2) 1687 if (N1C && N0.getOpcode() == ISD::SRA) { 1688 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1689 unsigned Sum = N1C->getValue() + C1->getValue(); 1690 if (Sum >= MVT::getSizeInBits(VT)) Sum = MVT::getSizeInBits(VT)-1; 1691 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0), 1692 DAG.getConstant(Sum, N1C->getValueType(0))); 1693 } 1694 } 1695 1696 // Simplify, based on bits shifted out of the LHS. 1697 if (N1C && SimplifyDemandedBits(SDOperand(N, 0))) 1698 return SDOperand(N, 0); 1699 1700 1701 // If the sign bit is known to be zero, switch this to a SRL. 1702 if (TLI.MaskedValueIsZero(N0, MVT::getIntVTSignBit(VT))) 1703 return DAG.getNode(ISD::SRL, VT, N0, N1); 1704 return SDOperand(); 1705} 1706 1707SDOperand DAGCombiner::visitSRL(SDNode *N) { 1708 SDOperand N0 = N->getOperand(0); 1709 SDOperand N1 = N->getOperand(1); 1710 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1711 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1712 MVT::ValueType VT = N0.getValueType(); 1713 unsigned OpSizeInBits = MVT::getSizeInBits(VT); 1714 1715 // fold (srl c1, c2) -> c1 >>u c2 1716 if (N0C && N1C) 1717 return DAG.getNode(ISD::SRL, VT, N0, N1); 1718 // fold (srl 0, x) -> 0 1719 if (N0C && N0C->isNullValue()) 1720 return N0; 1721 // fold (srl x, c >= size(x)) -> undef 1722 if (N1C && N1C->getValue() >= OpSizeInBits) 1723 return DAG.getNode(ISD::UNDEF, VT); 1724 // fold (srl x, 0) -> x 1725 if (N1C && N1C->isNullValue()) 1726 return N0; 1727 // if (srl x, c) is known to be zero, return 0 1728 if (N1C && TLI.MaskedValueIsZero(SDOperand(N, 0), ~0ULL >> (64-OpSizeInBits))) 1729 return DAG.getConstant(0, VT); 1730 // fold (srl (srl x, c1), c2) -> 0 or (srl x, c1+c2) 1731 if (N1C && N0.getOpcode() == ISD::SRL && 1732 N0.getOperand(1).getOpcode() == ISD::Constant) { 1733 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 1734 uint64_t c2 = N1C->getValue(); 1735 if (c1 + c2 > OpSizeInBits) 1736 return DAG.getConstant(0, VT); 1737 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0), 1738 DAG.getConstant(c1 + c2, N1.getValueType())); 1739 } 1740 1741 // fold (srl (anyextend x), c) -> (anyextend (srl x, c)) 1742 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) { 1743 // Shifting in all undef bits? 1744 MVT::ValueType SmallVT = N0.getOperand(0).getValueType(); 1745 if (N1C->getValue() >= MVT::getSizeInBits(SmallVT)) 1746 return DAG.getNode(ISD::UNDEF, VT); 1747 1748 SDOperand SmallShift = DAG.getNode(ISD::SRL, SmallVT, N0.getOperand(0), N1); 1749 AddToWorkList(SmallShift.Val); 1750 return DAG.getNode(ISD::ANY_EXTEND, VT, SmallShift); 1751 } 1752 1753 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign 1754 // bit, which is unmodified by sra. 1755 if (N1C && N1C->getValue()+1 == MVT::getSizeInBits(VT)) { 1756 if (N0.getOpcode() == ISD::SRA) 1757 return DAG.getNode(ISD::SRL, VT, N0.getOperand(0), N1); 1758 } 1759 1760 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit). 1761 if (N1C && N0.getOpcode() == ISD::CTLZ && 1762 N1C->getValue() == Log2_32(MVT::getSizeInBits(VT))) { 1763 uint64_t KnownZero, KnownOne, Mask = MVT::getIntVTBitMask(VT); 1764 TLI.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne); 1765 1766 // If any of the input bits are KnownOne, then the input couldn't be all 1767 // zeros, thus the result of the srl will always be zero. 1768 if (KnownOne) return DAG.getConstant(0, VT); 1769 1770 // If all of the bits input the to ctlz node are known to be zero, then 1771 // the result of the ctlz is "32" and the result of the shift is one. 1772 uint64_t UnknownBits = ~KnownZero & Mask; 1773 if (UnknownBits == 0) return DAG.getConstant(1, VT); 1774 1775 // Otherwise, check to see if there is exactly one bit input to the ctlz. 1776 if ((UnknownBits & (UnknownBits-1)) == 0) { 1777 // Okay, we know that only that the single bit specified by UnknownBits 1778 // could be set on input to the CTLZ node. If this bit is set, the SRL 1779 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair 1780 // to an SRL,XOR pair, which is likely to simplify more. 1781 unsigned ShAmt = CountTrailingZeros_64(UnknownBits); 1782 SDOperand Op = N0.getOperand(0); 1783 if (ShAmt) { 1784 Op = DAG.getNode(ISD::SRL, VT, Op, 1785 DAG.getConstant(ShAmt, TLI.getShiftAmountTy())); 1786 AddToWorkList(Op.Val); 1787 } 1788 return DAG.getNode(ISD::XOR, VT, Op, DAG.getConstant(1, VT)); 1789 } 1790 } 1791 1792 return SDOperand(); 1793} 1794 1795SDOperand DAGCombiner::visitCTLZ(SDNode *N) { 1796 SDOperand N0 = N->getOperand(0); 1797 MVT::ValueType VT = N->getValueType(0); 1798 1799 // fold (ctlz c1) -> c2 1800 if (isa<ConstantSDNode>(N0)) 1801 return DAG.getNode(ISD::CTLZ, VT, N0); 1802 return SDOperand(); 1803} 1804 1805SDOperand DAGCombiner::visitCTTZ(SDNode *N) { 1806 SDOperand N0 = N->getOperand(0); 1807 MVT::ValueType VT = N->getValueType(0); 1808 1809 // fold (cttz c1) -> c2 1810 if (isa<ConstantSDNode>(N0)) 1811 return DAG.getNode(ISD::CTTZ, VT, N0); 1812 return SDOperand(); 1813} 1814 1815SDOperand DAGCombiner::visitCTPOP(SDNode *N) { 1816 SDOperand N0 = N->getOperand(0); 1817 MVT::ValueType VT = N->getValueType(0); 1818 1819 // fold (ctpop c1) -> c2 1820 if (isa<ConstantSDNode>(N0)) 1821 return DAG.getNode(ISD::CTPOP, VT, N0); 1822 return SDOperand(); 1823} 1824 1825SDOperand DAGCombiner::visitSELECT(SDNode *N) { 1826 SDOperand N0 = N->getOperand(0); 1827 SDOperand N1 = N->getOperand(1); 1828 SDOperand N2 = N->getOperand(2); 1829 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 1830 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 1831 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2); 1832 MVT::ValueType VT = N->getValueType(0); 1833 1834 // fold select C, X, X -> X 1835 if (N1 == N2) 1836 return N1; 1837 // fold select true, X, Y -> X 1838 if (N0C && !N0C->isNullValue()) 1839 return N1; 1840 // fold select false, X, Y -> Y 1841 if (N0C && N0C->isNullValue()) 1842 return N2; 1843 // fold select C, 1, X -> C | X 1844 if (MVT::i1 == VT && N1C && N1C->getValue() == 1) 1845 return DAG.getNode(ISD::OR, VT, N0, N2); 1846 // fold select C, 0, X -> ~C & X 1847 // FIXME: this should check for C type == X type, not i1? 1848 if (MVT::i1 == VT && N1C && N1C->isNullValue()) { 1849 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT)); 1850 AddToWorkList(XORNode.Val); 1851 return DAG.getNode(ISD::AND, VT, XORNode, N2); 1852 } 1853 // fold select C, X, 1 -> ~C | X 1854 if (MVT::i1 == VT && N2C && N2C->getValue() == 1) { 1855 SDOperand XORNode = DAG.getNode(ISD::XOR, VT, N0, DAG.getConstant(1, VT)); 1856 AddToWorkList(XORNode.Val); 1857 return DAG.getNode(ISD::OR, VT, XORNode, N1); 1858 } 1859 // fold select C, X, 0 -> C & X 1860 // FIXME: this should check for C type == X type, not i1? 1861 if (MVT::i1 == VT && N2C && N2C->isNullValue()) 1862 return DAG.getNode(ISD::AND, VT, N0, N1); 1863 // fold X ? X : Y --> X ? 1 : Y --> X | Y 1864 if (MVT::i1 == VT && N0 == N1) 1865 return DAG.getNode(ISD::OR, VT, N0, N2); 1866 // fold X ? Y : X --> X ? Y : 0 --> X & Y 1867 if (MVT::i1 == VT && N0 == N2) 1868 return DAG.getNode(ISD::AND, VT, N0, N1); 1869 1870 // If we can fold this based on the true/false value, do so. 1871 if (SimplifySelectOps(N, N1, N2)) 1872 return SDOperand(N, 0); // Don't revisit N. 1873 1874 // fold selects based on a setcc into other things, such as min/max/abs 1875 if (N0.getOpcode() == ISD::SETCC) 1876 // FIXME: 1877 // Check against MVT::Other for SELECT_CC, which is a workaround for targets 1878 // having to say they don't support SELECT_CC on every type the DAG knows 1879 // about, since there is no way to mark an opcode illegal at all value types 1880 if (TLI.isOperationLegal(ISD::SELECT_CC, MVT::Other)) 1881 return DAG.getNode(ISD::SELECT_CC, VT, N0.getOperand(0), N0.getOperand(1), 1882 N1, N2, N0.getOperand(2)); 1883 else 1884 return SimplifySelect(N0, N1, N2); 1885 return SDOperand(); 1886} 1887 1888SDOperand DAGCombiner::visitSELECT_CC(SDNode *N) { 1889 SDOperand N0 = N->getOperand(0); 1890 SDOperand N1 = N->getOperand(1); 1891 SDOperand N2 = N->getOperand(2); 1892 SDOperand N3 = N->getOperand(3); 1893 SDOperand N4 = N->getOperand(4); 1894 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get(); 1895 1896 // fold select_cc lhs, rhs, x, x, cc -> x 1897 if (N2 == N3) 1898 return N2; 1899 1900 // Determine if the condition we're dealing with is constant 1901 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false); 1902 if (SCC.Val) AddToWorkList(SCC.Val); 1903 1904 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val)) { 1905 if (SCCC->getValue()) 1906 return N2; // cond always true -> true val 1907 else 1908 return N3; // cond always false -> false val 1909 } 1910 1911 // Fold to a simpler select_cc 1912 if (SCC.Val && SCC.getOpcode() == ISD::SETCC) 1913 return DAG.getNode(ISD::SELECT_CC, N2.getValueType(), 1914 SCC.getOperand(0), SCC.getOperand(1), N2, N3, 1915 SCC.getOperand(2)); 1916 1917 // If we can fold this based on the true/false value, do so. 1918 if (SimplifySelectOps(N, N2, N3)) 1919 return SDOperand(N, 0); // Don't revisit N. 1920 1921 // fold select_cc into other things, such as min/max/abs 1922 return SimplifySelectCC(N0, N1, N2, N3, CC); 1923} 1924 1925SDOperand DAGCombiner::visitSETCC(SDNode *N) { 1926 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1), 1927 cast<CondCodeSDNode>(N->getOperand(2))->get()); 1928} 1929 1930SDOperand DAGCombiner::visitSIGN_EXTEND(SDNode *N) { 1931 SDOperand N0 = N->getOperand(0); 1932 MVT::ValueType VT = N->getValueType(0); 1933 1934 // fold (sext c1) -> c1 1935 if (isa<ConstantSDNode>(N0)) 1936 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0); 1937 1938 // fold (sext (sext x)) -> (sext x) 1939 // fold (sext (aext x)) -> (sext x) 1940 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) 1941 return DAG.getNode(ISD::SIGN_EXTEND, VT, N0.getOperand(0)); 1942 1943 // fold (sext (truncate x)) -> (sextinreg x). 1944 if (N0.getOpcode() == ISD::TRUNCATE && 1945 (!AfterLegalize || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, 1946 N0.getValueType()))) { 1947 SDOperand Op = N0.getOperand(0); 1948 if (Op.getValueType() < VT) { 1949 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op); 1950 } else if (Op.getValueType() > VT) { 1951 Op = DAG.getNode(ISD::TRUNCATE, VT, Op); 1952 } 1953 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, Op, 1954 DAG.getValueType(N0.getValueType())); 1955 } 1956 1957 // fold (sext (load x)) -> (sext (truncate (sextload x))) 1958 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() && 1959 (!AfterLegalize||TLI.isLoadXLegal(ISD::SEXTLOAD, N0.getValueType()))){ 1960 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 1961 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(), 1962 LN0->getBasePtr(), LN0->getSrcValue(), 1963 LN0->getSrcValueOffset(), 1964 N0.getValueType()); 1965 CombineTo(N, ExtLoad); 1966 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1967 ExtLoad.getValue(1)); 1968 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1969 } 1970 1971 // fold (sext (sextload x)) -> (sext (truncate (sextload x))) 1972 // fold (sext ( extload x)) -> (sext (truncate (sextload x))) 1973 if ((ISD::isSEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) && N0.hasOneUse()) { 1974 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 1975 MVT::ValueType EVT = LN0->getLoadedVT(); 1976 if (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT)) { 1977 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(), 1978 LN0->getBasePtr(), LN0->getSrcValue(), 1979 LN0->getSrcValueOffset(), EVT); 1980 CombineTo(N, ExtLoad); 1981 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 1982 ExtLoad.getValue(1)); 1983 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 1984 } 1985 } 1986 1987 return SDOperand(); 1988} 1989 1990SDOperand DAGCombiner::visitZERO_EXTEND(SDNode *N) { 1991 SDOperand N0 = N->getOperand(0); 1992 MVT::ValueType VT = N->getValueType(0); 1993 1994 // fold (zext c1) -> c1 1995 if (isa<ConstantSDNode>(N0)) 1996 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0); 1997 // fold (zext (zext x)) -> (zext x) 1998 // fold (zext (aext x)) -> (zext x) 1999 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) 2000 return DAG.getNode(ISD::ZERO_EXTEND, VT, N0.getOperand(0)); 2001 2002 // fold (zext (truncate x)) -> (and x, mask) 2003 if (N0.getOpcode() == ISD::TRUNCATE && 2004 (!AfterLegalize || TLI.isOperationLegal(ISD::AND, VT))) { 2005 SDOperand Op = N0.getOperand(0); 2006 if (Op.getValueType() < VT) { 2007 Op = DAG.getNode(ISD::ANY_EXTEND, VT, Op); 2008 } else if (Op.getValueType() > VT) { 2009 Op = DAG.getNode(ISD::TRUNCATE, VT, Op); 2010 } 2011 return DAG.getZeroExtendInReg(Op, N0.getValueType()); 2012 } 2013 2014 // fold (zext (and (trunc x), cst)) -> (and x, cst). 2015 if (N0.getOpcode() == ISD::AND && 2016 N0.getOperand(0).getOpcode() == ISD::TRUNCATE && 2017 N0.getOperand(1).getOpcode() == ISD::Constant) { 2018 SDOperand X = N0.getOperand(0).getOperand(0); 2019 if (X.getValueType() < VT) { 2020 X = DAG.getNode(ISD::ANY_EXTEND, VT, X); 2021 } else if (X.getValueType() > VT) { 2022 X = DAG.getNode(ISD::TRUNCATE, VT, X); 2023 } 2024 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 2025 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT)); 2026 } 2027 2028 // fold (zext (load x)) -> (zext (truncate (zextload x))) 2029 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() && 2030 (!AfterLegalize||TLI.isLoadXLegal(ISD::ZEXTLOAD, N0.getValueType()))) { 2031 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2032 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(), 2033 LN0->getBasePtr(), LN0->getSrcValue(), 2034 LN0->getSrcValueOffset(), 2035 N0.getValueType()); 2036 CombineTo(N, ExtLoad); 2037 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 2038 ExtLoad.getValue(1)); 2039 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2040 } 2041 2042 // fold (zext (zextload x)) -> (zext (truncate (zextload x))) 2043 // fold (zext ( extload x)) -> (zext (truncate (zextload x))) 2044 if ((ISD::isZEXTLoad(N0.Val) || ISD::isEXTLoad(N0.Val)) && N0.hasOneUse()) { 2045 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2046 MVT::ValueType EVT = LN0->getLoadedVT(); 2047 SDOperand ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, VT, LN0->getChain(), 2048 LN0->getBasePtr(), LN0->getSrcValue(), 2049 LN0->getSrcValueOffset(), EVT); 2050 CombineTo(N, ExtLoad); 2051 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 2052 ExtLoad.getValue(1)); 2053 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2054 } 2055 return SDOperand(); 2056} 2057 2058SDOperand DAGCombiner::visitANY_EXTEND(SDNode *N) { 2059 SDOperand N0 = N->getOperand(0); 2060 MVT::ValueType VT = N->getValueType(0); 2061 2062 // fold (aext c1) -> c1 2063 if (isa<ConstantSDNode>(N0)) 2064 return DAG.getNode(ISD::ANY_EXTEND, VT, N0); 2065 // fold (aext (aext x)) -> (aext x) 2066 // fold (aext (zext x)) -> (zext x) 2067 // fold (aext (sext x)) -> (sext x) 2068 if (N0.getOpcode() == ISD::ANY_EXTEND || 2069 N0.getOpcode() == ISD::ZERO_EXTEND || 2070 N0.getOpcode() == ISD::SIGN_EXTEND) 2071 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0)); 2072 2073 // fold (aext (truncate x)) 2074 if (N0.getOpcode() == ISD::TRUNCATE) { 2075 SDOperand TruncOp = N0.getOperand(0); 2076 if (TruncOp.getValueType() == VT) 2077 return TruncOp; // x iff x size == zext size. 2078 if (TruncOp.getValueType() > VT) 2079 return DAG.getNode(ISD::TRUNCATE, VT, TruncOp); 2080 return DAG.getNode(ISD::ANY_EXTEND, VT, TruncOp); 2081 } 2082 2083 // fold (aext (and (trunc x), cst)) -> (and x, cst). 2084 if (N0.getOpcode() == ISD::AND && 2085 N0.getOperand(0).getOpcode() == ISD::TRUNCATE && 2086 N0.getOperand(1).getOpcode() == ISD::Constant) { 2087 SDOperand X = N0.getOperand(0).getOperand(0); 2088 if (X.getValueType() < VT) { 2089 X = DAG.getNode(ISD::ANY_EXTEND, VT, X); 2090 } else if (X.getValueType() > VT) { 2091 X = DAG.getNode(ISD::TRUNCATE, VT, X); 2092 } 2093 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getValue(); 2094 return DAG.getNode(ISD::AND, VT, X, DAG.getConstant(Mask, VT)); 2095 } 2096 2097 // fold (aext (load x)) -> (aext (truncate (extload x))) 2098 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() && 2099 (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) { 2100 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2101 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(), 2102 LN0->getBasePtr(), LN0->getSrcValue(), 2103 LN0->getSrcValueOffset(), 2104 N0.getValueType()); 2105 CombineTo(N, ExtLoad); 2106 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 2107 ExtLoad.getValue(1)); 2108 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2109 } 2110 2111 // fold (aext (zextload x)) -> (aext (truncate (zextload x))) 2112 // fold (aext (sextload x)) -> (aext (truncate (sextload x))) 2113 // fold (aext ( extload x)) -> (aext (truncate (extload x))) 2114 if (N0.getOpcode() == ISD::LOAD && !ISD::isNON_EXTLoad(N0.Val) && 2115 N0.hasOneUse()) { 2116 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2117 MVT::ValueType EVT = LN0->getLoadedVT(); 2118 SDOperand ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), VT, 2119 LN0->getChain(), LN0->getBasePtr(), 2120 LN0->getSrcValue(), 2121 LN0->getSrcValueOffset(), EVT); 2122 CombineTo(N, ExtLoad); 2123 CombineTo(N0.Val, DAG.getNode(ISD::TRUNCATE, N0.getValueType(), ExtLoad), 2124 ExtLoad.getValue(1)); 2125 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2126 } 2127 return SDOperand(); 2128} 2129 2130 2131SDOperand DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) { 2132 SDOperand N0 = N->getOperand(0); 2133 SDOperand N1 = N->getOperand(1); 2134 MVT::ValueType VT = N->getValueType(0); 2135 MVT::ValueType EVT = cast<VTSDNode>(N1)->getVT(); 2136 unsigned EVTBits = MVT::getSizeInBits(EVT); 2137 2138 // fold (sext_in_reg c1) -> c1 2139 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF) 2140 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0, N1); 2141 2142 // If the input is already sign extended, just drop the extension. 2143 if (TLI.ComputeNumSignBits(N0) >= MVT::getSizeInBits(VT)-EVTBits+1) 2144 return N0; 2145 2146 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2 2147 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 2148 EVT < cast<VTSDNode>(N0.getOperand(1))->getVT()) { 2149 return DAG.getNode(ISD::SIGN_EXTEND_INREG, VT, N0.getOperand(0), N1); 2150 } 2151 2152 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is zero 2153 if (TLI.MaskedValueIsZero(N0, 1ULL << (EVTBits-1))) 2154 return DAG.getZeroExtendInReg(N0, EVT); 2155 2156 // fold (sext_in_reg (srl X, 24), i8) -> sra X, 24 2157 // fold (sext_in_reg (srl X, 23), i8) -> sra X, 23 iff possible. 2158 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above. 2159 if (N0.getOpcode() == ISD::SRL) { 2160 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1))) 2161 if (ShAmt->getValue()+EVTBits <= MVT::getSizeInBits(VT)) { 2162 // We can turn this into an SRA iff the input to the SRL is already sign 2163 // extended enough. 2164 unsigned InSignBits = TLI.ComputeNumSignBits(N0.getOperand(0)); 2165 if (MVT::getSizeInBits(VT)-(ShAmt->getValue()+EVTBits) < InSignBits) 2166 return DAG.getNode(ISD::SRA, VT, N0.getOperand(0), N0.getOperand(1)); 2167 } 2168 } 2169 2170 // fold (sext_inreg (extload x)) -> (sextload x) 2171 if (ISD::isEXTLoad(N0.Val) && 2172 EVT == cast<LoadSDNode>(N0)->getLoadedVT() && 2173 (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) { 2174 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2175 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(), 2176 LN0->getBasePtr(), LN0->getSrcValue(), 2177 LN0->getSrcValueOffset(), EVT); 2178 CombineTo(N, ExtLoad); 2179 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 2180 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2181 } 2182 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use 2183 if (ISD::isZEXTLoad(N0.Val) && N0.hasOneUse() && 2184 EVT == cast<LoadSDNode>(N0)->getLoadedVT() && 2185 (!AfterLegalize || TLI.isLoadXLegal(ISD::SEXTLOAD, EVT))) { 2186 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2187 SDOperand ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, VT, LN0->getChain(), 2188 LN0->getBasePtr(), LN0->getSrcValue(), 2189 LN0->getSrcValueOffset(), EVT); 2190 CombineTo(N, ExtLoad); 2191 CombineTo(N0.Val, ExtLoad, ExtLoad.getValue(1)); 2192 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2193 } 2194 return SDOperand(); 2195} 2196 2197SDOperand DAGCombiner::visitTRUNCATE(SDNode *N) { 2198 SDOperand N0 = N->getOperand(0); 2199 MVT::ValueType VT = N->getValueType(0); 2200 2201 // noop truncate 2202 if (N0.getValueType() == N->getValueType(0)) 2203 return N0; 2204 // fold (truncate c1) -> c1 2205 if (isa<ConstantSDNode>(N0)) 2206 return DAG.getNode(ISD::TRUNCATE, VT, N0); 2207 // fold (truncate (truncate x)) -> (truncate x) 2208 if (N0.getOpcode() == ISD::TRUNCATE) 2209 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0)); 2210 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x 2211 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::SIGN_EXTEND|| 2212 N0.getOpcode() == ISD::ANY_EXTEND) { 2213 if (N0.getOperand(0).getValueType() < VT) 2214 // if the source is smaller than the dest, we still need an extend 2215 return DAG.getNode(N0.getOpcode(), VT, N0.getOperand(0)); 2216 else if (N0.getOperand(0).getValueType() > VT) 2217 // if the source is larger than the dest, than we just need the truncate 2218 return DAG.getNode(ISD::TRUNCATE, VT, N0.getOperand(0)); 2219 else 2220 // if the source and dest are the same type, we can drop both the extend 2221 // and the truncate 2222 return N0.getOperand(0); 2223 } 2224 // fold (truncate (load x)) -> (smaller load x) 2225 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() && 2226 // Do not allow folding to i1 here. i1 is implicitly stored in memory in 2227 // zero extended form: by shrinking the load, we lose track of the fact 2228 // that it is already zero extended. 2229 // FIXME: This should be reevaluated. 2230 VT != MVT::i1) { 2231 assert(MVT::getSizeInBits(N0.getValueType()) > MVT::getSizeInBits(VT) && 2232 "Cannot truncate to larger type!"); 2233 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2234 MVT::ValueType PtrType = N0.getOperand(1).getValueType(); 2235 // For big endian targets, we need to add an offset to the pointer to load 2236 // the correct bytes. For little endian systems, we merely need to read 2237 // fewer bytes from the same pointer. 2238 uint64_t PtrOff = 2239 (MVT::getSizeInBits(N0.getValueType()) - MVT::getSizeInBits(VT)) / 8; 2240 SDOperand NewPtr = TLI.isLittleEndian() ? LN0->getBasePtr() : 2241 DAG.getNode(ISD::ADD, PtrType, LN0->getBasePtr(), 2242 DAG.getConstant(PtrOff, PtrType)); 2243 AddToWorkList(NewPtr.Val); 2244 SDOperand Load = DAG.getLoad(VT, LN0->getChain(), NewPtr, 2245 LN0->getSrcValue(), LN0->getSrcValueOffset()); 2246 AddToWorkList(N); 2247 CombineTo(N0.Val, Load, Load.getValue(1)); 2248 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2249 } 2250 return SDOperand(); 2251} 2252 2253SDOperand DAGCombiner::visitBIT_CONVERT(SDNode *N) { 2254 SDOperand N0 = N->getOperand(0); 2255 MVT::ValueType VT = N->getValueType(0); 2256 2257 // If the input is a constant, let getNode() fold it. 2258 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) { 2259 SDOperand Res = DAG.getNode(ISD::BIT_CONVERT, VT, N0); 2260 if (Res.Val != N) return Res; 2261 } 2262 2263 if (N0.getOpcode() == ISD::BIT_CONVERT) // conv(conv(x,t1),t2) -> conv(x,t2) 2264 return DAG.getNode(ISD::BIT_CONVERT, VT, N0.getOperand(0)); 2265 2266 // fold (conv (load x)) -> (load (conv*)x) 2267 // FIXME: These xforms need to know that the resultant load doesn't need a 2268 // higher alignment than the original! 2269 if (0 && ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse()) { 2270 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2271 SDOperand Load = DAG.getLoad(VT, LN0->getChain(), LN0->getBasePtr(), 2272 LN0->getSrcValue(), LN0->getSrcValueOffset()); 2273 AddToWorkList(N); 2274 CombineTo(N0.Val, DAG.getNode(ISD::BIT_CONVERT, N0.getValueType(), Load), 2275 Load.getValue(1)); 2276 return Load; 2277 } 2278 2279 return SDOperand(); 2280} 2281 2282SDOperand DAGCombiner::visitVBIT_CONVERT(SDNode *N) { 2283 SDOperand N0 = N->getOperand(0); 2284 MVT::ValueType VT = N->getValueType(0); 2285 2286 // If the input is a VBUILD_VECTOR with all constant elements, fold this now. 2287 // First check to see if this is all constant. 2288 if (N0.getOpcode() == ISD::VBUILD_VECTOR && N0.Val->hasOneUse() && 2289 VT == MVT::Vector) { 2290 bool isSimple = true; 2291 for (unsigned i = 0, e = N0.getNumOperands()-2; i != e; ++i) 2292 if (N0.getOperand(i).getOpcode() != ISD::UNDEF && 2293 N0.getOperand(i).getOpcode() != ISD::Constant && 2294 N0.getOperand(i).getOpcode() != ISD::ConstantFP) { 2295 isSimple = false; 2296 break; 2297 } 2298 2299 MVT::ValueType DestEltVT = cast<VTSDNode>(N->getOperand(2))->getVT(); 2300 if (isSimple && !MVT::isVector(DestEltVT)) { 2301 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(N0.Val, DestEltVT); 2302 } 2303 } 2304 2305 return SDOperand(); 2306} 2307 2308/// ConstantFoldVBIT_CONVERTofVBUILD_VECTOR - We know that BV is a vbuild_vector 2309/// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the 2310/// destination element value type. 2311SDOperand DAGCombiner:: 2312ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(SDNode *BV, MVT::ValueType DstEltVT) { 2313 MVT::ValueType SrcEltVT = BV->getOperand(0).getValueType(); 2314 2315 // If this is already the right type, we're done. 2316 if (SrcEltVT == DstEltVT) return SDOperand(BV, 0); 2317 2318 unsigned SrcBitSize = MVT::getSizeInBits(SrcEltVT); 2319 unsigned DstBitSize = MVT::getSizeInBits(DstEltVT); 2320 2321 // If this is a conversion of N elements of one type to N elements of another 2322 // type, convert each element. This handles FP<->INT cases. 2323 if (SrcBitSize == DstBitSize) { 2324 SmallVector<SDOperand, 8> Ops; 2325 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) { 2326 Ops.push_back(DAG.getNode(ISD::BIT_CONVERT, DstEltVT, BV->getOperand(i))); 2327 AddToWorkList(Ops.back().Val); 2328 } 2329 Ops.push_back(*(BV->op_end()-2)); // Add num elements. 2330 Ops.push_back(DAG.getValueType(DstEltVT)); 2331 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size()); 2332 } 2333 2334 // Otherwise, we're growing or shrinking the elements. To avoid having to 2335 // handle annoying details of growing/shrinking FP values, we convert them to 2336 // int first. 2337 if (MVT::isFloatingPoint(SrcEltVT)) { 2338 // Convert the input float vector to a int vector where the elements are the 2339 // same sizes. 2340 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!"); 2341 MVT::ValueType IntVT = SrcEltVT == MVT::f32 ? MVT::i32 : MVT::i64; 2342 BV = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, IntVT).Val; 2343 SrcEltVT = IntVT; 2344 } 2345 2346 // Now we know the input is an integer vector. If the output is a FP type, 2347 // convert to integer first, then to FP of the right size. 2348 if (MVT::isFloatingPoint(DstEltVT)) { 2349 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!"); 2350 MVT::ValueType TmpVT = DstEltVT == MVT::f32 ? MVT::i32 : MVT::i64; 2351 SDNode *Tmp = ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(BV, TmpVT).Val; 2352 2353 // Next, convert to FP elements of the same size. 2354 return ConstantFoldVBIT_CONVERTofVBUILD_VECTOR(Tmp, DstEltVT); 2355 } 2356 2357 // Okay, we know the src/dst types are both integers of differing types. 2358 // Handling growing first. 2359 assert(MVT::isInteger(SrcEltVT) && MVT::isInteger(DstEltVT)); 2360 if (SrcBitSize < DstBitSize) { 2361 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize; 2362 2363 SmallVector<SDOperand, 8> Ops; 2364 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; 2365 i += NumInputsPerOutput) { 2366 bool isLE = TLI.isLittleEndian(); 2367 uint64_t NewBits = 0; 2368 bool EltIsUndef = true; 2369 for (unsigned j = 0; j != NumInputsPerOutput; ++j) { 2370 // Shift the previously computed bits over. 2371 NewBits <<= SrcBitSize; 2372 SDOperand Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j)); 2373 if (Op.getOpcode() == ISD::UNDEF) continue; 2374 EltIsUndef = false; 2375 2376 NewBits |= cast<ConstantSDNode>(Op)->getValue(); 2377 } 2378 2379 if (EltIsUndef) 2380 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT)); 2381 else 2382 Ops.push_back(DAG.getConstant(NewBits, DstEltVT)); 2383 } 2384 2385 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements. 2386 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size. 2387 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size()); 2388 } 2389 2390 // Finally, this must be the case where we are shrinking elements: each input 2391 // turns into multiple outputs. 2392 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize; 2393 SmallVector<SDOperand, 8> Ops; 2394 for (unsigned i = 0, e = BV->getNumOperands()-2; i != e; ++i) { 2395 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) { 2396 for (unsigned j = 0; j != NumOutputsPerInput; ++j) 2397 Ops.push_back(DAG.getNode(ISD::UNDEF, DstEltVT)); 2398 continue; 2399 } 2400 uint64_t OpVal = cast<ConstantSDNode>(BV->getOperand(i))->getValue(); 2401 2402 for (unsigned j = 0; j != NumOutputsPerInput; ++j) { 2403 unsigned ThisVal = OpVal & ((1ULL << DstBitSize)-1); 2404 OpVal >>= DstBitSize; 2405 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT)); 2406 } 2407 2408 // For big endian targets, swap the order of the pieces of each element. 2409 if (!TLI.isLittleEndian()) 2410 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end()); 2411 } 2412 Ops.push_back(DAG.getConstant(Ops.size(), MVT::i32)); // Add num elements. 2413 Ops.push_back(DAG.getValueType(DstEltVT)); // Add element size. 2414 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size()); 2415} 2416 2417 2418 2419SDOperand DAGCombiner::visitFADD(SDNode *N) { 2420 SDOperand N0 = N->getOperand(0); 2421 SDOperand N1 = N->getOperand(1); 2422 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2423 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2424 MVT::ValueType VT = N->getValueType(0); 2425 2426 // fold (fadd c1, c2) -> c1+c2 2427 if (N0CFP && N1CFP) 2428 return DAG.getNode(ISD::FADD, VT, N0, N1); 2429 // canonicalize constant to RHS 2430 if (N0CFP && !N1CFP) 2431 return DAG.getNode(ISD::FADD, VT, N1, N0); 2432 // fold (A + (-B)) -> A-B 2433 if (N1.getOpcode() == ISD::FNEG) 2434 return DAG.getNode(ISD::FSUB, VT, N0, N1.getOperand(0)); 2435 // fold ((-A) + B) -> B-A 2436 if (N0.getOpcode() == ISD::FNEG) 2437 return DAG.getNode(ISD::FSUB, VT, N1, N0.getOperand(0)); 2438 2439 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2)) 2440 if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FADD && 2441 N0.Val->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1))) 2442 return DAG.getNode(ISD::FADD, VT, N0.getOperand(0), 2443 DAG.getNode(ISD::FADD, VT, N0.getOperand(1), N1)); 2444 2445 return SDOperand(); 2446} 2447 2448SDOperand DAGCombiner::visitFSUB(SDNode *N) { 2449 SDOperand N0 = N->getOperand(0); 2450 SDOperand N1 = N->getOperand(1); 2451 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2452 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2453 MVT::ValueType VT = N->getValueType(0); 2454 2455 // fold (fsub c1, c2) -> c1-c2 2456 if (N0CFP && N1CFP) 2457 return DAG.getNode(ISD::FSUB, VT, N0, N1); 2458 // fold (A-(-B)) -> A+B 2459 if (N1.getOpcode() == ISD::FNEG) 2460 return DAG.getNode(ISD::FADD, VT, N0, N1.getOperand(0)); 2461 return SDOperand(); 2462} 2463 2464SDOperand DAGCombiner::visitFMUL(SDNode *N) { 2465 SDOperand N0 = N->getOperand(0); 2466 SDOperand N1 = N->getOperand(1); 2467 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2468 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2469 MVT::ValueType VT = N->getValueType(0); 2470 2471 // fold (fmul c1, c2) -> c1*c2 2472 if (N0CFP && N1CFP) 2473 return DAG.getNode(ISD::FMUL, VT, N0, N1); 2474 // canonicalize constant to RHS 2475 if (N0CFP && !N1CFP) 2476 return DAG.getNode(ISD::FMUL, VT, N1, N0); 2477 // fold (fmul X, 2.0) -> (fadd X, X) 2478 if (N1CFP && N1CFP->isExactlyValue(+2.0)) 2479 return DAG.getNode(ISD::FADD, VT, N0, N0); 2480 2481 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2)) 2482 if (UnsafeFPMath && N1CFP && N0.getOpcode() == ISD::FMUL && 2483 N0.Val->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1))) 2484 return DAG.getNode(ISD::FMUL, VT, N0.getOperand(0), 2485 DAG.getNode(ISD::FMUL, VT, N0.getOperand(1), N1)); 2486 2487 return SDOperand(); 2488} 2489 2490SDOperand DAGCombiner::visitFDIV(SDNode *N) { 2491 SDOperand N0 = N->getOperand(0); 2492 SDOperand N1 = N->getOperand(1); 2493 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2494 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2495 MVT::ValueType VT = N->getValueType(0); 2496 2497 // fold (fdiv c1, c2) -> c1/c2 2498 if (N0CFP && N1CFP) 2499 return DAG.getNode(ISD::FDIV, VT, N0, N1); 2500 return SDOperand(); 2501} 2502 2503SDOperand DAGCombiner::visitFREM(SDNode *N) { 2504 SDOperand N0 = N->getOperand(0); 2505 SDOperand N1 = N->getOperand(1); 2506 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2507 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2508 MVT::ValueType VT = N->getValueType(0); 2509 2510 // fold (frem c1, c2) -> fmod(c1,c2) 2511 if (N0CFP && N1CFP) 2512 return DAG.getNode(ISD::FREM, VT, N0, N1); 2513 return SDOperand(); 2514} 2515 2516SDOperand DAGCombiner::visitFCOPYSIGN(SDNode *N) { 2517 SDOperand N0 = N->getOperand(0); 2518 SDOperand N1 = N->getOperand(1); 2519 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2520 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1); 2521 MVT::ValueType VT = N->getValueType(0); 2522 2523 if (N0CFP && N1CFP) // Constant fold 2524 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1); 2525 2526 if (N1CFP) { 2527 // copysign(x, c1) -> fabs(x) iff ispos(c1) 2528 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1) 2529 union { 2530 double d; 2531 int64_t i; 2532 } u; 2533 u.d = N1CFP->getValue(); 2534 if (u.i >= 0) 2535 return DAG.getNode(ISD::FABS, VT, N0); 2536 else 2537 return DAG.getNode(ISD::FNEG, VT, DAG.getNode(ISD::FABS, VT, N0)); 2538 } 2539 2540 // copysign(fabs(x), y) -> copysign(x, y) 2541 // copysign(fneg(x), y) -> copysign(x, y) 2542 // copysign(copysign(x,z), y) -> copysign(x, y) 2543 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG || 2544 N0.getOpcode() == ISD::FCOPYSIGN) 2545 return DAG.getNode(ISD::FCOPYSIGN, VT, N0.getOperand(0), N1); 2546 2547 // copysign(x, abs(y)) -> abs(x) 2548 if (N1.getOpcode() == ISD::FABS) 2549 return DAG.getNode(ISD::FABS, VT, N0); 2550 2551 // copysign(x, copysign(y,z)) -> copysign(x, z) 2552 if (N1.getOpcode() == ISD::FCOPYSIGN) 2553 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(1)); 2554 2555 // copysign(x, fp_extend(y)) -> copysign(x, y) 2556 // copysign(x, fp_round(y)) -> copysign(x, y) 2557 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND) 2558 return DAG.getNode(ISD::FCOPYSIGN, VT, N0, N1.getOperand(0)); 2559 2560 return SDOperand(); 2561} 2562 2563 2564 2565SDOperand DAGCombiner::visitSINT_TO_FP(SDNode *N) { 2566 SDOperand N0 = N->getOperand(0); 2567 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2568 MVT::ValueType VT = N->getValueType(0); 2569 2570 // fold (sint_to_fp c1) -> c1fp 2571 if (N0C) 2572 return DAG.getNode(ISD::SINT_TO_FP, VT, N0); 2573 return SDOperand(); 2574} 2575 2576SDOperand DAGCombiner::visitUINT_TO_FP(SDNode *N) { 2577 SDOperand N0 = N->getOperand(0); 2578 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); 2579 MVT::ValueType VT = N->getValueType(0); 2580 2581 // fold (uint_to_fp c1) -> c1fp 2582 if (N0C) 2583 return DAG.getNode(ISD::UINT_TO_FP, VT, N0); 2584 return SDOperand(); 2585} 2586 2587SDOperand DAGCombiner::visitFP_TO_SINT(SDNode *N) { 2588 SDOperand N0 = N->getOperand(0); 2589 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2590 MVT::ValueType VT = N->getValueType(0); 2591 2592 // fold (fp_to_sint c1fp) -> c1 2593 if (N0CFP) 2594 return DAG.getNode(ISD::FP_TO_SINT, VT, N0); 2595 return SDOperand(); 2596} 2597 2598SDOperand DAGCombiner::visitFP_TO_UINT(SDNode *N) { 2599 SDOperand N0 = N->getOperand(0); 2600 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2601 MVT::ValueType VT = N->getValueType(0); 2602 2603 // fold (fp_to_uint c1fp) -> c1 2604 if (N0CFP) 2605 return DAG.getNode(ISD::FP_TO_UINT, VT, N0); 2606 return SDOperand(); 2607} 2608 2609SDOperand DAGCombiner::visitFP_ROUND(SDNode *N) { 2610 SDOperand N0 = N->getOperand(0); 2611 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2612 MVT::ValueType VT = N->getValueType(0); 2613 2614 // fold (fp_round c1fp) -> c1fp 2615 if (N0CFP) 2616 return DAG.getNode(ISD::FP_ROUND, VT, N0); 2617 2618 // fold (fp_round (fp_extend x)) -> x 2619 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType()) 2620 return N0.getOperand(0); 2621 2622 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y) 2623 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.Val->hasOneUse()) { 2624 SDOperand Tmp = DAG.getNode(ISD::FP_ROUND, VT, N0.getOperand(0)); 2625 AddToWorkList(Tmp.Val); 2626 return DAG.getNode(ISD::FCOPYSIGN, VT, Tmp, N0.getOperand(1)); 2627 } 2628 2629 return SDOperand(); 2630} 2631 2632SDOperand DAGCombiner::visitFP_ROUND_INREG(SDNode *N) { 2633 SDOperand N0 = N->getOperand(0); 2634 MVT::ValueType VT = N->getValueType(0); 2635 MVT::ValueType EVT = cast<VTSDNode>(N->getOperand(1))->getVT(); 2636 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2637 2638 // fold (fp_round_inreg c1fp) -> c1fp 2639 if (N0CFP) { 2640 SDOperand Round = DAG.getConstantFP(N0CFP->getValue(), EVT); 2641 return DAG.getNode(ISD::FP_EXTEND, VT, Round); 2642 } 2643 return SDOperand(); 2644} 2645 2646SDOperand DAGCombiner::visitFP_EXTEND(SDNode *N) { 2647 SDOperand N0 = N->getOperand(0); 2648 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2649 MVT::ValueType VT = N->getValueType(0); 2650 2651 // fold (fp_extend c1fp) -> c1fp 2652 if (N0CFP) 2653 return DAG.getNode(ISD::FP_EXTEND, VT, N0); 2654 2655 // fold (fpext (load x)) -> (fpext (fpround (extload x))) 2656 if (ISD::isNON_EXTLoad(N0.Val) && N0.hasOneUse() && 2657 (!AfterLegalize||TLI.isLoadXLegal(ISD::EXTLOAD, N0.getValueType()))) { 2658 LoadSDNode *LN0 = cast<LoadSDNode>(N0); 2659 SDOperand ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, VT, LN0->getChain(), 2660 LN0->getBasePtr(), LN0->getSrcValue(), 2661 LN0->getSrcValueOffset(), 2662 N0.getValueType()); 2663 CombineTo(N, ExtLoad); 2664 CombineTo(N0.Val, DAG.getNode(ISD::FP_ROUND, N0.getValueType(), ExtLoad), 2665 ExtLoad.getValue(1)); 2666 return SDOperand(N, 0); // Return N so it doesn't get rechecked! 2667 } 2668 2669 2670 return SDOperand(); 2671} 2672 2673SDOperand DAGCombiner::visitFNEG(SDNode *N) { 2674 SDOperand N0 = N->getOperand(0); 2675 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2676 MVT::ValueType VT = N->getValueType(0); 2677 2678 // fold (fneg c1) -> -c1 2679 if (N0CFP) 2680 return DAG.getNode(ISD::FNEG, VT, N0); 2681 // fold (fneg (sub x, y)) -> (sub y, x) 2682 if (N0.getOpcode() == ISD::SUB) 2683 return DAG.getNode(ISD::SUB, VT, N0.getOperand(1), N0.getOperand(0)); 2684 // fold (fneg (fneg x)) -> x 2685 if (N0.getOpcode() == ISD::FNEG) 2686 return N0.getOperand(0); 2687 return SDOperand(); 2688} 2689 2690SDOperand DAGCombiner::visitFABS(SDNode *N) { 2691 SDOperand N0 = N->getOperand(0); 2692 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0); 2693 MVT::ValueType VT = N->getValueType(0); 2694 2695 // fold (fabs c1) -> fabs(c1) 2696 if (N0CFP) 2697 return DAG.getNode(ISD::FABS, VT, N0); 2698 // fold (fabs (fabs x)) -> (fabs x) 2699 if (N0.getOpcode() == ISD::FABS) 2700 return N->getOperand(0); 2701 // fold (fabs (fneg x)) -> (fabs x) 2702 // fold (fabs (fcopysign x, y)) -> (fabs x) 2703 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN) 2704 return DAG.getNode(ISD::FABS, VT, N0.getOperand(0)); 2705 2706 return SDOperand(); 2707} 2708 2709SDOperand DAGCombiner::visitBRCOND(SDNode *N) { 2710 SDOperand Chain = N->getOperand(0); 2711 SDOperand N1 = N->getOperand(1); 2712 SDOperand N2 = N->getOperand(2); 2713 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); 2714 2715 // never taken branch, fold to chain 2716 if (N1C && N1C->isNullValue()) 2717 return Chain; 2718 // unconditional branch 2719 if (N1C && N1C->getValue() == 1) 2720 return DAG.getNode(ISD::BR, MVT::Other, Chain, N2); 2721 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal 2722 // on the target. 2723 if (N1.getOpcode() == ISD::SETCC && 2724 TLI.isOperationLegal(ISD::BR_CC, MVT::Other)) { 2725 return DAG.getNode(ISD::BR_CC, MVT::Other, Chain, N1.getOperand(2), 2726 N1.getOperand(0), N1.getOperand(1), N2); 2727 } 2728 return SDOperand(); 2729} 2730 2731// Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB. 2732// 2733SDOperand DAGCombiner::visitBR_CC(SDNode *N) { 2734 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1)); 2735 SDOperand CondLHS = N->getOperand(2), CondRHS = N->getOperand(3); 2736 2737 // Use SimplifySetCC to simplify SETCC's. 2738 SDOperand Simp = SimplifySetCC(MVT::i1, CondLHS, CondRHS, CC->get(), false); 2739 if (Simp.Val) AddToWorkList(Simp.Val); 2740 2741 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(Simp.Val); 2742 2743 // fold br_cc true, dest -> br dest (unconditional branch) 2744 if (SCCC && SCCC->getValue()) 2745 return DAG.getNode(ISD::BR, MVT::Other, N->getOperand(0), 2746 N->getOperand(4)); 2747 // fold br_cc false, dest -> unconditional fall through 2748 if (SCCC && SCCC->isNullValue()) 2749 return N->getOperand(0); 2750 2751 // fold to a simpler setcc 2752 if (Simp.Val && Simp.getOpcode() == ISD::SETCC) 2753 return DAG.getNode(ISD::BR_CC, MVT::Other, N->getOperand(0), 2754 Simp.getOperand(2), Simp.getOperand(0), 2755 Simp.getOperand(1), N->getOperand(4)); 2756 return SDOperand(); 2757} 2758 2759 2760/// CombineToPreIndexedLoadStore - Try turning a load / store and a 2761/// pre-indexed load / store when the base pointer is a add or subtract 2762/// and it has other uses besides the load / store. After the 2763/// transformation, the new indexed load / store has effectively folded 2764/// the add / subtract in and all of its other uses are redirected to the 2765/// new load / store. 2766bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) { 2767 if (!AfterLegalize) 2768 return false; 2769 2770 bool isLoad = true; 2771 SDOperand Ptr; 2772 MVT::ValueType VT; 2773 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 2774 if (LD->getAddressingMode() != ISD::UNINDEXED) 2775 return false; 2776 VT = LD->getLoadedVT(); 2777 if (LD->getAddressingMode() != ISD::UNINDEXED && 2778 !TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) && 2779 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT)) 2780 return false; 2781 Ptr = LD->getBasePtr(); 2782 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 2783 if (ST->getAddressingMode() != ISD::UNINDEXED) 2784 return false; 2785 VT = ST->getStoredVT(); 2786 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) && 2787 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT)) 2788 return false; 2789 Ptr = ST->getBasePtr(); 2790 isLoad = false; 2791 } else 2792 return false; 2793 2794 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail 2795 // out. There is no reason to make this a preinc/predec. 2796 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) || 2797 Ptr.Val->hasOneUse()) 2798 return false; 2799 2800 // Ask the target to do addressing mode selection. 2801 SDOperand BasePtr; 2802 SDOperand Offset; 2803 ISD::MemIndexedMode AM = ISD::UNINDEXED; 2804 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG)) 2805 return false; 2806 2807 // Try turning it into a pre-indexed load / store except when: 2808 // 1) The base is a frame index. 2809 // 2) If N is a store and the ptr is either the same as or is a 2810 // predecessor of the value being stored. 2811 // 3) Another use of base ptr is a predecessor of N. If ptr is folded 2812 // that would create a cycle. 2813 // 4) All uses are load / store ops that use it as base ptr. 2814 2815 // Check #1. Preinc'ing a frame index would require copying the stack pointer 2816 // (plus the implicit offset) to a register to preinc anyway. 2817 if (isa<FrameIndexSDNode>(BasePtr)) 2818 return false; 2819 2820 // Check #2. 2821 if (!isLoad) { 2822 SDOperand Val = cast<StoreSDNode>(N)->getValue(); 2823 if (Val == Ptr || Ptr.Val->isPredecessor(Val.Val)) 2824 return false; 2825 } 2826 2827 // Now check for #2 and #3. 2828 bool RealUse = false; 2829 for (SDNode::use_iterator I = Ptr.Val->use_begin(), 2830 E = Ptr.Val->use_end(); I != E; ++I) { 2831 SDNode *Use = *I; 2832 if (Use == N) 2833 continue; 2834 if (Use->isPredecessor(N)) 2835 return false; 2836 2837 if (!((Use->getOpcode() == ISD::LOAD && 2838 cast<LoadSDNode>(Use)->getBasePtr() == Ptr) || 2839 (Use->getOpcode() == ISD::STORE) && 2840 cast<StoreSDNode>(Use)->getBasePtr() == Ptr)) 2841 RealUse = true; 2842 } 2843 if (!RealUse) 2844 return false; 2845 2846 SDOperand Result; 2847 if (isLoad) 2848 Result = DAG.getIndexedLoad(SDOperand(N,0), BasePtr, Offset, AM); 2849 else 2850 Result = DAG.getIndexedStore(SDOperand(N,0), BasePtr, Offset, AM); 2851 ++PreIndexedNodes; 2852 ++NodesCombined; 2853 DOUT << "\nReplacing.4 "; DEBUG(N->dump()); 2854 DOUT << "\nWith: "; DEBUG(Result.Val->dump(&DAG)); 2855 DOUT << '\n'; 2856 std::vector<SDNode*> NowDead; 2857 if (isLoad) { 2858 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(0), 2859 NowDead); 2860 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Result.getValue(2), 2861 NowDead); 2862 } else { 2863 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(1), 2864 NowDead); 2865 } 2866 2867 // Nodes can end up on the worklist more than once. Make sure we do 2868 // not process a node that has been replaced. 2869 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 2870 removeFromWorkList(NowDead[i]); 2871 // Finally, since the node is now dead, remove it from the graph. 2872 DAG.DeleteNode(N); 2873 2874 // Replace the uses of Ptr with uses of the updated base value. 2875 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0), 2876 NowDead); 2877 removeFromWorkList(Ptr.Val); 2878 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 2879 removeFromWorkList(NowDead[i]); 2880 DAG.DeleteNode(Ptr.Val); 2881 2882 return true; 2883} 2884 2885/// CombineToPostIndexedLoadStore - Try combine a load / store with a 2886/// add / sub of the base pointer node into a post-indexed load / store. 2887/// The transformation folded the add / subtract into the new indexed 2888/// load / store effectively and all of its uses are redirected to the 2889/// new load / store. 2890bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) { 2891 if (!AfterLegalize) 2892 return false; 2893 2894 bool isLoad = true; 2895 SDOperand Ptr; 2896 MVT::ValueType VT; 2897 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 2898 if (LD->getAddressingMode() != ISD::UNINDEXED) 2899 return false; 2900 VT = LD->getLoadedVT(); 2901 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) && 2902 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT)) 2903 return false; 2904 Ptr = LD->getBasePtr(); 2905 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 2906 if (ST->getAddressingMode() != ISD::UNINDEXED) 2907 return false; 2908 VT = ST->getStoredVT(); 2909 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) && 2910 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT)) 2911 return false; 2912 Ptr = ST->getBasePtr(); 2913 isLoad = false; 2914 } else 2915 return false; 2916 2917 if (Ptr.Val->hasOneUse()) 2918 return false; 2919 2920 for (SDNode::use_iterator I = Ptr.Val->use_begin(), 2921 E = Ptr.Val->use_end(); I != E; ++I) { 2922 SDNode *Op = *I; 2923 if (Op == N || 2924 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)) 2925 continue; 2926 2927 SDOperand BasePtr; 2928 SDOperand Offset; 2929 ISD::MemIndexedMode AM = ISD::UNINDEXED; 2930 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) { 2931 if (Ptr == Offset) 2932 std::swap(BasePtr, Offset); 2933 if (Ptr != BasePtr) 2934 continue; 2935 2936 // Try turning it into a post-indexed load / store except when 2937 // 1) All uses are load / store ops that use it as base ptr. 2938 // 2) Op must be independent of N, i.e. Op is neither a predecessor 2939 // nor a successor of N. Otherwise, if Op is folded that would 2940 // create a cycle. 2941 2942 // Check for #1. 2943 bool TryNext = false; 2944 for (SDNode::use_iterator II = BasePtr.Val->use_begin(), 2945 EE = BasePtr.Val->use_end(); II != EE; ++II) { 2946 SDNode *Use = *II; 2947 if (Use == Ptr.Val) 2948 continue; 2949 2950 // If all the uses are load / store addresses, then don't do the 2951 // transformation. 2952 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){ 2953 bool RealUse = false; 2954 for (SDNode::use_iterator III = Use->use_begin(), 2955 EEE = Use->use_end(); III != EEE; ++III) { 2956 SDNode *UseUse = *III; 2957 if (!((UseUse->getOpcode() == ISD::LOAD && 2958 cast<LoadSDNode>(UseUse)->getBasePtr().Val == Use) || 2959 (UseUse->getOpcode() == ISD::STORE) && 2960 cast<StoreSDNode>(UseUse)->getBasePtr().Val == Use)) 2961 RealUse = true; 2962 } 2963 2964 if (!RealUse) { 2965 TryNext = true; 2966 break; 2967 } 2968 } 2969 } 2970 if (TryNext) 2971 continue; 2972 2973 // Check for #2 2974 if (!Op->isPredecessor(N) && !N->isPredecessor(Op)) { 2975 SDOperand Result = isLoad 2976 ? DAG.getIndexedLoad(SDOperand(N,0), BasePtr, Offset, AM) 2977 : DAG.getIndexedStore(SDOperand(N,0), BasePtr, Offset, AM); 2978 ++PostIndexedNodes; 2979 ++NodesCombined; 2980 DOUT << "\nReplacing.5 "; DEBUG(N->dump()); 2981 DOUT << "\nWith: "; DEBUG(Result.Val->dump(&DAG)); 2982 DOUT << '\n'; 2983 std::vector<SDNode*> NowDead; 2984 if (isLoad) { 2985 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(0), 2986 NowDead); 2987 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 1), Result.getValue(2), 2988 NowDead); 2989 } else { 2990 DAG.ReplaceAllUsesOfValueWith(SDOperand(N, 0), Result.getValue(1), 2991 NowDead); 2992 } 2993 2994 // Nodes can end up on the worklist more than once. Make sure we do 2995 // not process a node that has been replaced. 2996 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 2997 removeFromWorkList(NowDead[i]); 2998 // Finally, since the node is now dead, remove it from the graph. 2999 DAG.DeleteNode(N); 3000 3001 // Replace the uses of Use with uses of the updated base value. 3002 DAG.ReplaceAllUsesOfValueWith(SDOperand(Op, 0), 3003 Result.getValue(isLoad ? 1 : 0), 3004 NowDead); 3005 removeFromWorkList(Op); 3006 for (unsigned i = 0, e = NowDead.size(); i != e; ++i) 3007 removeFromWorkList(NowDead[i]); 3008 DAG.DeleteNode(Op); 3009 3010 return true; 3011 } 3012 } 3013 } 3014 return false; 3015} 3016 3017 3018SDOperand DAGCombiner::visitLOAD(SDNode *N) { 3019 LoadSDNode *LD = cast<LoadSDNode>(N); 3020 SDOperand Chain = LD->getChain(); 3021 SDOperand Ptr = LD->getBasePtr(); 3022 3023 // If there are no uses of the loaded value, change uses of the chain value 3024 // into uses of the chain input (i.e. delete the dead load). 3025 if (N->hasNUsesOfValue(0, 0)) 3026 return CombineTo(N, DAG.getNode(ISD::UNDEF, N->getValueType(0)), Chain); 3027 3028 // If this load is directly stored, replace the load value with the stored 3029 // value. 3030 // TODO: Handle store large -> read small portion. 3031 // TODO: Handle TRUNCSTORE/LOADEXT 3032 if (LD->getExtensionType() == ISD::NON_EXTLOAD) { 3033 if (ISD::isNON_TRUNCStore(Chain.Val)) { 3034 StoreSDNode *PrevST = cast<StoreSDNode>(Chain); 3035 if (PrevST->getBasePtr() == Ptr && 3036 PrevST->getValue().getValueType() == N->getValueType(0)) 3037 return CombineTo(N, Chain.getOperand(1), Chain); 3038 } 3039 } 3040 3041 if (CombinerAA) { 3042 // Walk up chain skipping non-aliasing memory nodes. 3043 SDOperand BetterChain = FindBetterChain(N, Chain); 3044 3045 // If there is a better chain. 3046 if (Chain != BetterChain) { 3047 SDOperand ReplLoad; 3048 3049 // Replace the chain to void dependency. 3050 if (LD->getExtensionType() == ISD::NON_EXTLOAD) { 3051 ReplLoad = DAG.getLoad(N->getValueType(0), BetterChain, Ptr, 3052 LD->getSrcValue(), LD->getSrcValueOffset()); 3053 } else { 3054 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), 3055 LD->getValueType(0), 3056 BetterChain, Ptr, LD->getSrcValue(), 3057 LD->getSrcValueOffset(), 3058 LD->getLoadedVT()); 3059 } 3060 3061 // Create token factor to keep old chain connected. 3062 SDOperand Token = DAG.getNode(ISD::TokenFactor, MVT::Other, 3063 Chain, ReplLoad.getValue(1)); 3064 3065 // Replace uses with load result and token factor. Don't add users 3066 // to work list. 3067 return CombineTo(N, ReplLoad.getValue(0), Token, false); 3068 } 3069 } 3070 3071 // Try transforming N to an indexed load. 3072 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) 3073 return SDOperand(N, 0); 3074 3075 return SDOperand(); 3076} 3077 3078SDOperand DAGCombiner::visitSTORE(SDNode *N) { 3079 StoreSDNode *ST = cast<StoreSDNode>(N); 3080 SDOperand Chain = ST->getChain(); 3081 SDOperand Value = ST->getValue(); 3082 SDOperand Ptr = ST->getBasePtr(); 3083 3084 // If this is a store of a bit convert, store the input value. 3085 // FIXME: This needs to know that the resultant store does not need a 3086 // higher alignment than the original. 3087 if (0 && Value.getOpcode() == ISD::BIT_CONVERT) { 3088 return DAG.getStore(Chain, Value.getOperand(0), Ptr, ST->getSrcValue(), 3089 ST->getSrcValueOffset()); 3090 } 3091 3092 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr' 3093 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) { 3094 if (Value.getOpcode() != ISD::TargetConstantFP) { 3095 SDOperand Tmp; 3096 switch (CFP->getValueType(0)) { 3097 default: assert(0 && "Unknown FP type"); 3098 case MVT::f32: 3099 if (!AfterLegalize || TLI.isTypeLegal(MVT::i32)) { 3100 Tmp = DAG.getConstant(FloatToBits(CFP->getValue()), MVT::i32); 3101 return DAG.getStore(Chain, Tmp, Ptr, ST->getSrcValue(), 3102 ST->getSrcValueOffset()); 3103 } 3104 break; 3105 case MVT::f64: 3106 if (!AfterLegalize || TLI.isTypeLegal(MVT::i64)) { 3107 Tmp = DAG.getConstant(DoubleToBits(CFP->getValue()), MVT::i64); 3108 return DAG.getStore(Chain, Tmp, Ptr, ST->getSrcValue(), 3109 ST->getSrcValueOffset()); 3110 } else if (TLI.isTypeLegal(MVT::i32)) { 3111 // Many FP stores are not make apparent until after legalize, e.g. for 3112 // argument passing. Since this is so common, custom legalize the 3113 // 64-bit integer store into two 32-bit stores. 3114 uint64_t Val = DoubleToBits(CFP->getValue()); 3115 SDOperand Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32); 3116 SDOperand Hi = DAG.getConstant(Val >> 32, MVT::i32); 3117 if (!TLI.isLittleEndian()) std::swap(Lo, Hi); 3118 3119 SDOperand St0 = DAG.getStore(Chain, Lo, Ptr, ST->getSrcValue(), 3120 ST->getSrcValueOffset()); 3121 Ptr = DAG.getNode(ISD::ADD, Ptr.getValueType(), Ptr, 3122 DAG.getConstant(4, Ptr.getValueType())); 3123 SDOperand St1 = DAG.getStore(Chain, Hi, Ptr, ST->getSrcValue(), 3124 ST->getSrcValueOffset()+4); 3125 return DAG.getNode(ISD::TokenFactor, MVT::Other, St0, St1); 3126 } 3127 break; 3128 } 3129 } 3130 } 3131 3132 if (CombinerAA) { 3133 // Walk up chain skipping non-aliasing memory nodes. 3134 SDOperand BetterChain = FindBetterChain(N, Chain); 3135 3136 // If there is a better chain. 3137 if (Chain != BetterChain) { 3138 // Replace the chain to avoid dependency. 3139 SDOperand ReplStore; 3140 if (ST->isTruncatingStore()) { 3141 ReplStore = DAG.getTruncStore(BetterChain, Value, Ptr, 3142 ST->getSrcValue(),ST->getSrcValueOffset(), ST->getStoredVT()); 3143 } else { 3144 ReplStore = DAG.getStore(BetterChain, Value, Ptr, 3145 ST->getSrcValue(), ST->getSrcValueOffset()); 3146 } 3147 3148 // Create token to keep both nodes around. 3149 SDOperand Token = 3150 DAG.getNode(ISD::TokenFactor, MVT::Other, Chain, ReplStore); 3151 3152 // Don't add users to work list. 3153 return CombineTo(N, Token, false); 3154 } 3155 } 3156 3157 // Try transforming N to an indexed store. 3158 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N)) 3159 return SDOperand(N, 0); 3160 3161 return SDOperand(); 3162} 3163 3164SDOperand DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) { 3165 SDOperand InVec = N->getOperand(0); 3166 SDOperand InVal = N->getOperand(1); 3167 SDOperand EltNo = N->getOperand(2); 3168 3169 // If the invec is a BUILD_VECTOR and if EltNo is a constant, build a new 3170 // vector with the inserted element. 3171 if (InVec.getOpcode() == ISD::BUILD_VECTOR && isa<ConstantSDNode>(EltNo)) { 3172 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue(); 3173 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end()); 3174 if (Elt < Ops.size()) 3175 Ops[Elt] = InVal; 3176 return DAG.getNode(ISD::BUILD_VECTOR, InVec.getValueType(), 3177 &Ops[0], Ops.size()); 3178 } 3179 3180 return SDOperand(); 3181} 3182 3183SDOperand DAGCombiner::visitVINSERT_VECTOR_ELT(SDNode *N) { 3184 SDOperand InVec = N->getOperand(0); 3185 SDOperand InVal = N->getOperand(1); 3186 SDOperand EltNo = N->getOperand(2); 3187 SDOperand NumElts = N->getOperand(3); 3188 SDOperand EltType = N->getOperand(4); 3189 3190 // If the invec is a VBUILD_VECTOR and if EltNo is a constant, build a new 3191 // vector with the inserted element. 3192 if (InVec.getOpcode() == ISD::VBUILD_VECTOR && isa<ConstantSDNode>(EltNo)) { 3193 unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue(); 3194 SmallVector<SDOperand, 8> Ops(InVec.Val->op_begin(), InVec.Val->op_end()); 3195 if (Elt < Ops.size()-2) 3196 Ops[Elt] = InVal; 3197 return DAG.getNode(ISD::VBUILD_VECTOR, InVec.getValueType(), 3198 &Ops[0], Ops.size()); 3199 } 3200 3201 return SDOperand(); 3202} 3203 3204SDOperand DAGCombiner::visitVBUILD_VECTOR(SDNode *N) { 3205 unsigned NumInScalars = N->getNumOperands()-2; 3206 SDOperand NumElts = N->getOperand(NumInScalars); 3207 SDOperand EltType = N->getOperand(NumInScalars+1); 3208 3209 // Check to see if this is a VBUILD_VECTOR of a bunch of VEXTRACT_VECTOR_ELT 3210 // operations. If so, and if the EXTRACT_ELT vector inputs come from at most 3211 // two distinct vectors, turn this into a shuffle node. 3212 SDOperand VecIn1, VecIn2; 3213 for (unsigned i = 0; i != NumInScalars; ++i) { 3214 // Ignore undef inputs. 3215 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue; 3216 3217 // If this input is something other than a VEXTRACT_VECTOR_ELT with a 3218 // constant index, bail out. 3219 if (N->getOperand(i).getOpcode() != ISD::VEXTRACT_VECTOR_ELT || 3220 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) { 3221 VecIn1 = VecIn2 = SDOperand(0, 0); 3222 break; 3223 } 3224 3225 // If the input vector type disagrees with the result of the vbuild_vector, 3226 // we can't make a shuffle. 3227 SDOperand ExtractedFromVec = N->getOperand(i).getOperand(0); 3228 if (*(ExtractedFromVec.Val->op_end()-2) != NumElts || 3229 *(ExtractedFromVec.Val->op_end()-1) != EltType) { 3230 VecIn1 = VecIn2 = SDOperand(0, 0); 3231 break; 3232 } 3233 3234 // Otherwise, remember this. We allow up to two distinct input vectors. 3235 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2) 3236 continue; 3237 3238 if (VecIn1.Val == 0) { 3239 VecIn1 = ExtractedFromVec; 3240 } else if (VecIn2.Val == 0) { 3241 VecIn2 = ExtractedFromVec; 3242 } else { 3243 // Too many inputs. 3244 VecIn1 = VecIn2 = SDOperand(0, 0); 3245 break; 3246 } 3247 } 3248 3249 // If everything is good, we can make a shuffle operation. 3250 if (VecIn1.Val) { 3251 SmallVector<SDOperand, 8> BuildVecIndices; 3252 for (unsigned i = 0; i != NumInScalars; ++i) { 3253 if (N->getOperand(i).getOpcode() == ISD::UNDEF) { 3254 BuildVecIndices.push_back(DAG.getNode(ISD::UNDEF, TLI.getPointerTy())); 3255 continue; 3256 } 3257 3258 SDOperand Extract = N->getOperand(i); 3259 3260 // If extracting from the first vector, just use the index directly. 3261 if (Extract.getOperand(0) == VecIn1) { 3262 BuildVecIndices.push_back(Extract.getOperand(1)); 3263 continue; 3264 } 3265 3266 // Otherwise, use InIdx + VecSize 3267 unsigned Idx = cast<ConstantSDNode>(Extract.getOperand(1))->getValue(); 3268 BuildVecIndices.push_back(DAG.getConstant(Idx+NumInScalars, 3269 TLI.getPointerTy())); 3270 } 3271 3272 // Add count and size info. 3273 BuildVecIndices.push_back(NumElts); 3274 BuildVecIndices.push_back(DAG.getValueType(TLI.getPointerTy())); 3275 3276 // Return the new VVECTOR_SHUFFLE node. 3277 SDOperand Ops[5]; 3278 Ops[0] = VecIn1; 3279 if (VecIn2.Val) { 3280 Ops[1] = VecIn2; 3281 } else { 3282 // Use an undef vbuild_vector as input for the second operand. 3283 std::vector<SDOperand> UnOps(NumInScalars, 3284 DAG.getNode(ISD::UNDEF, 3285 cast<VTSDNode>(EltType)->getVT())); 3286 UnOps.push_back(NumElts); 3287 UnOps.push_back(EltType); 3288 Ops[1] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, 3289 &UnOps[0], UnOps.size()); 3290 AddToWorkList(Ops[1].Val); 3291 } 3292 Ops[2] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, 3293 &BuildVecIndices[0], BuildVecIndices.size()); 3294 Ops[3] = NumElts; 3295 Ops[4] = EltType; 3296 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, Ops, 5); 3297 } 3298 3299 return SDOperand(); 3300} 3301 3302SDOperand DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) { 3303 SDOperand ShufMask = N->getOperand(2); 3304 unsigned NumElts = ShufMask.getNumOperands(); 3305 3306 // If the shuffle mask is an identity operation on the LHS, return the LHS. 3307 bool isIdentity = true; 3308 for (unsigned i = 0; i != NumElts; ++i) { 3309 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF && 3310 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) { 3311 isIdentity = false; 3312 break; 3313 } 3314 } 3315 if (isIdentity) return N->getOperand(0); 3316 3317 // If the shuffle mask is an identity operation on the RHS, return the RHS. 3318 isIdentity = true; 3319 for (unsigned i = 0; i != NumElts; ++i) { 3320 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF && 3321 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) { 3322 isIdentity = false; 3323 break; 3324 } 3325 } 3326 if (isIdentity) return N->getOperand(1); 3327 3328 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not 3329 // needed at all. 3330 bool isUnary = true; 3331 bool isSplat = true; 3332 int VecNum = -1; 3333 unsigned BaseIdx = 0; 3334 for (unsigned i = 0; i != NumElts; ++i) 3335 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) { 3336 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue(); 3337 int V = (Idx < NumElts) ? 0 : 1; 3338 if (VecNum == -1) { 3339 VecNum = V; 3340 BaseIdx = Idx; 3341 } else { 3342 if (BaseIdx != Idx) 3343 isSplat = false; 3344 if (VecNum != V) { 3345 isUnary = false; 3346 break; 3347 } 3348 } 3349 } 3350 3351 SDOperand N0 = N->getOperand(0); 3352 SDOperand N1 = N->getOperand(1); 3353 // Normalize unary shuffle so the RHS is undef. 3354 if (isUnary && VecNum == 1) 3355 std::swap(N0, N1); 3356 3357 // If it is a splat, check if the argument vector is a build_vector with 3358 // all scalar elements the same. 3359 if (isSplat) { 3360 SDNode *V = N0.Val; 3361 if (V->getOpcode() == ISD::BIT_CONVERT) 3362 V = V->getOperand(0).Val; 3363 if (V->getOpcode() == ISD::BUILD_VECTOR) { 3364 unsigned NumElems = V->getNumOperands()-2; 3365 if (NumElems > BaseIdx) { 3366 SDOperand Base; 3367 bool AllSame = true; 3368 for (unsigned i = 0; i != NumElems; ++i) { 3369 if (V->getOperand(i).getOpcode() != ISD::UNDEF) { 3370 Base = V->getOperand(i); 3371 break; 3372 } 3373 } 3374 // Splat of <u, u, u, u>, return <u, u, u, u> 3375 if (!Base.Val) 3376 return N0; 3377 for (unsigned i = 0; i != NumElems; ++i) { 3378 if (V->getOperand(i).getOpcode() != ISD::UNDEF && 3379 V->getOperand(i) != Base) { 3380 AllSame = false; 3381 break; 3382 } 3383 } 3384 // Splat of <x, x, x, x>, return <x, x, x, x> 3385 if (AllSame) 3386 return N0; 3387 } 3388 } 3389 } 3390 3391 // If it is a unary or the LHS and the RHS are the same node, turn the RHS 3392 // into an undef. 3393 if (isUnary || N0 == N1) { 3394 if (N0.getOpcode() == ISD::UNDEF) 3395 return DAG.getNode(ISD::UNDEF, N->getValueType(0)); 3396 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the 3397 // first operand. 3398 SmallVector<SDOperand, 8> MappedOps; 3399 for (unsigned i = 0, e = ShufMask.getNumOperands(); i != e; ++i) { 3400 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF || 3401 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) { 3402 MappedOps.push_back(ShufMask.getOperand(i)); 3403 } else { 3404 unsigned NewIdx = 3405 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts; 3406 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32)); 3407 } 3408 } 3409 ShufMask = DAG.getNode(ISD::BUILD_VECTOR, ShufMask.getValueType(), 3410 &MappedOps[0], MappedOps.size()); 3411 AddToWorkList(ShufMask.Val); 3412 return DAG.getNode(ISD::VECTOR_SHUFFLE, N->getValueType(0), 3413 N0, 3414 DAG.getNode(ISD::UNDEF, N->getValueType(0)), 3415 ShufMask); 3416 } 3417 3418 return SDOperand(); 3419} 3420 3421SDOperand DAGCombiner::visitVVECTOR_SHUFFLE(SDNode *N) { 3422 SDOperand ShufMask = N->getOperand(2); 3423 unsigned NumElts = ShufMask.getNumOperands()-2; 3424 3425 // If the shuffle mask is an identity operation on the LHS, return the LHS. 3426 bool isIdentity = true; 3427 for (unsigned i = 0; i != NumElts; ++i) { 3428 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF && 3429 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i) { 3430 isIdentity = false; 3431 break; 3432 } 3433 } 3434 if (isIdentity) return N->getOperand(0); 3435 3436 // If the shuffle mask is an identity operation on the RHS, return the RHS. 3437 isIdentity = true; 3438 for (unsigned i = 0; i != NumElts; ++i) { 3439 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF && 3440 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() != i+NumElts) { 3441 isIdentity = false; 3442 break; 3443 } 3444 } 3445 if (isIdentity) return N->getOperand(1); 3446 3447 // Check if the shuffle is a unary shuffle, i.e. one of the vectors is not 3448 // needed at all. 3449 bool isUnary = true; 3450 bool isSplat = true; 3451 int VecNum = -1; 3452 unsigned BaseIdx = 0; 3453 for (unsigned i = 0; i != NumElts; ++i) 3454 if (ShufMask.getOperand(i).getOpcode() != ISD::UNDEF) { 3455 unsigned Idx = cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue(); 3456 int V = (Idx < NumElts) ? 0 : 1; 3457 if (VecNum == -1) { 3458 VecNum = V; 3459 BaseIdx = Idx; 3460 } else { 3461 if (BaseIdx != Idx) 3462 isSplat = false; 3463 if (VecNum != V) { 3464 isUnary = false; 3465 break; 3466 } 3467 } 3468 } 3469 3470 SDOperand N0 = N->getOperand(0); 3471 SDOperand N1 = N->getOperand(1); 3472 // Normalize unary shuffle so the RHS is undef. 3473 if (isUnary && VecNum == 1) 3474 std::swap(N0, N1); 3475 3476 // If it is a splat, check if the argument vector is a build_vector with 3477 // all scalar elements the same. 3478 if (isSplat) { 3479 SDNode *V = N0.Val; 3480 3481 // If this is a vbit convert that changes the element type of the vector but 3482 // not the number of vector elements, look through it. Be careful not to 3483 // look though conversions that change things like v4f32 to v2f64. 3484 if (V->getOpcode() == ISD::VBIT_CONVERT) { 3485 SDOperand ConvInput = V->getOperand(0); 3486 if (ConvInput.getValueType() == MVT::Vector && 3487 NumElts == 3488 ConvInput.getConstantOperandVal(ConvInput.getNumOperands()-2)) 3489 V = ConvInput.Val; 3490 } 3491 3492 if (V->getOpcode() == ISD::VBUILD_VECTOR) { 3493 unsigned NumElems = V->getNumOperands()-2; 3494 if (NumElems > BaseIdx) { 3495 SDOperand Base; 3496 bool AllSame = true; 3497 for (unsigned i = 0; i != NumElems; ++i) { 3498 if (V->getOperand(i).getOpcode() != ISD::UNDEF) { 3499 Base = V->getOperand(i); 3500 break; 3501 } 3502 } 3503 // Splat of <u, u, u, u>, return <u, u, u, u> 3504 if (!Base.Val) 3505 return N0; 3506 for (unsigned i = 0; i != NumElems; ++i) { 3507 if (V->getOperand(i).getOpcode() != ISD::UNDEF && 3508 V->getOperand(i) != Base) { 3509 AllSame = false; 3510 break; 3511 } 3512 } 3513 // Splat of <x, x, x, x>, return <x, x, x, x> 3514 if (AllSame) 3515 return N0; 3516 } 3517 } 3518 } 3519 3520 // If it is a unary or the LHS and the RHS are the same node, turn the RHS 3521 // into an undef. 3522 if (isUnary || N0 == N1) { 3523 // Check the SHUFFLE mask, mapping any inputs from the 2nd operand into the 3524 // first operand. 3525 SmallVector<SDOperand, 8> MappedOps; 3526 for (unsigned i = 0; i != NumElts; ++i) { 3527 if (ShufMask.getOperand(i).getOpcode() == ISD::UNDEF || 3528 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() < NumElts) { 3529 MappedOps.push_back(ShufMask.getOperand(i)); 3530 } else { 3531 unsigned NewIdx = 3532 cast<ConstantSDNode>(ShufMask.getOperand(i))->getValue() - NumElts; 3533 MappedOps.push_back(DAG.getConstant(NewIdx, MVT::i32)); 3534 } 3535 } 3536 // Add the type/#elts values. 3537 MappedOps.push_back(ShufMask.getOperand(NumElts)); 3538 MappedOps.push_back(ShufMask.getOperand(NumElts+1)); 3539 3540 ShufMask = DAG.getNode(ISD::VBUILD_VECTOR, ShufMask.getValueType(), 3541 &MappedOps[0], MappedOps.size()); 3542 AddToWorkList(ShufMask.Val); 3543 3544 // Build the undef vector. 3545 SDOperand UDVal = DAG.getNode(ISD::UNDEF, MappedOps[0].getValueType()); 3546 for (unsigned i = 0; i != NumElts; ++i) 3547 MappedOps[i] = UDVal; 3548 MappedOps[NumElts ] = *(N0.Val->op_end()-2); 3549 MappedOps[NumElts+1] = *(N0.Val->op_end()-1); 3550 UDVal = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, 3551 &MappedOps[0], MappedOps.size()); 3552 3553 return DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, 3554 N0, UDVal, ShufMask, 3555 MappedOps[NumElts], MappedOps[NumElts+1]); 3556 } 3557 3558 return SDOperand(); 3559} 3560 3561/// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform 3562/// a VAND to a vector_shuffle with the destination vector and a zero vector. 3563/// e.g. VAND V, <0xffffffff, 0, 0xffffffff, 0>. ==> 3564/// vector_shuffle V, Zero, <0, 4, 2, 4> 3565SDOperand DAGCombiner::XformToShuffleWithZero(SDNode *N) { 3566 SDOperand LHS = N->getOperand(0); 3567 SDOperand RHS = N->getOperand(1); 3568 if (N->getOpcode() == ISD::VAND) { 3569 SDOperand DstVecSize = *(LHS.Val->op_end()-2); 3570 SDOperand DstVecEVT = *(LHS.Val->op_end()-1); 3571 if (RHS.getOpcode() == ISD::VBIT_CONVERT) 3572 RHS = RHS.getOperand(0); 3573 if (RHS.getOpcode() == ISD::VBUILD_VECTOR) { 3574 std::vector<SDOperand> IdxOps; 3575 unsigned NumOps = RHS.getNumOperands(); 3576 unsigned NumElts = NumOps-2; 3577 MVT::ValueType EVT = cast<VTSDNode>(RHS.getOperand(NumOps-1))->getVT(); 3578 for (unsigned i = 0; i != NumElts; ++i) { 3579 SDOperand Elt = RHS.getOperand(i); 3580 if (!isa<ConstantSDNode>(Elt)) 3581 return SDOperand(); 3582 else if (cast<ConstantSDNode>(Elt)->isAllOnesValue()) 3583 IdxOps.push_back(DAG.getConstant(i, EVT)); 3584 else if (cast<ConstantSDNode>(Elt)->isNullValue()) 3585 IdxOps.push_back(DAG.getConstant(NumElts, EVT)); 3586 else 3587 return SDOperand(); 3588 } 3589 3590 // Let's see if the target supports this vector_shuffle. 3591 if (!TLI.isVectorClearMaskLegal(IdxOps, EVT, DAG)) 3592 return SDOperand(); 3593 3594 // Return the new VVECTOR_SHUFFLE node. 3595 SDOperand NumEltsNode = DAG.getConstant(NumElts, MVT::i32); 3596 SDOperand EVTNode = DAG.getValueType(EVT); 3597 std::vector<SDOperand> Ops; 3598 LHS = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, LHS, NumEltsNode, 3599 EVTNode); 3600 Ops.push_back(LHS); 3601 AddToWorkList(LHS.Val); 3602 std::vector<SDOperand> ZeroOps(NumElts, DAG.getConstant(0, EVT)); 3603 ZeroOps.push_back(NumEltsNode); 3604 ZeroOps.push_back(EVTNode); 3605 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, 3606 &ZeroOps[0], ZeroOps.size())); 3607 IdxOps.push_back(NumEltsNode); 3608 IdxOps.push_back(EVTNode); 3609 Ops.push_back(DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, 3610 &IdxOps[0], IdxOps.size())); 3611 Ops.push_back(NumEltsNode); 3612 Ops.push_back(EVTNode); 3613 SDOperand Result = DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, 3614 &Ops[0], Ops.size()); 3615 if (NumEltsNode != DstVecSize || EVTNode != DstVecEVT) { 3616 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result, 3617 DstVecSize, DstVecEVT); 3618 } 3619 return Result; 3620 } 3621 } 3622 return SDOperand(); 3623} 3624 3625/// visitVBinOp - Visit a binary vector operation, like VADD. IntOp indicates 3626/// the scalar operation of the vop if it is operating on an integer vector 3627/// (e.g. ADD) and FPOp indicates the FP version (e.g. FADD). 3628SDOperand DAGCombiner::visitVBinOp(SDNode *N, ISD::NodeType IntOp, 3629 ISD::NodeType FPOp) { 3630 MVT::ValueType EltType = cast<VTSDNode>(*(N->op_end()-1))->getVT(); 3631 ISD::NodeType ScalarOp = MVT::isInteger(EltType) ? IntOp : FPOp; 3632 SDOperand LHS = N->getOperand(0); 3633 SDOperand RHS = N->getOperand(1); 3634 SDOperand Shuffle = XformToShuffleWithZero(N); 3635 if (Shuffle.Val) return Shuffle; 3636 3637 // If the LHS and RHS are VBUILD_VECTOR nodes, see if we can constant fold 3638 // this operation. 3639 if (LHS.getOpcode() == ISD::VBUILD_VECTOR && 3640 RHS.getOpcode() == ISD::VBUILD_VECTOR) { 3641 SmallVector<SDOperand, 8> Ops; 3642 for (unsigned i = 0, e = LHS.getNumOperands()-2; i != e; ++i) { 3643 SDOperand LHSOp = LHS.getOperand(i); 3644 SDOperand RHSOp = RHS.getOperand(i); 3645 // If these two elements can't be folded, bail out. 3646 if ((LHSOp.getOpcode() != ISD::UNDEF && 3647 LHSOp.getOpcode() != ISD::Constant && 3648 LHSOp.getOpcode() != ISD::ConstantFP) || 3649 (RHSOp.getOpcode() != ISD::UNDEF && 3650 RHSOp.getOpcode() != ISD::Constant && 3651 RHSOp.getOpcode() != ISD::ConstantFP)) 3652 break; 3653 // Can't fold divide by zero. 3654 if (N->getOpcode() == ISD::VSDIV || N->getOpcode() == ISD::VUDIV) { 3655 if ((RHSOp.getOpcode() == ISD::Constant && 3656 cast<ConstantSDNode>(RHSOp.Val)->isNullValue()) || 3657 (RHSOp.getOpcode() == ISD::ConstantFP && 3658 !cast<ConstantFPSDNode>(RHSOp.Val)->getValue())) 3659 break; 3660 } 3661 Ops.push_back(DAG.getNode(ScalarOp, EltType, LHSOp, RHSOp)); 3662 AddToWorkList(Ops.back().Val); 3663 assert((Ops.back().getOpcode() == ISD::UNDEF || 3664 Ops.back().getOpcode() == ISD::Constant || 3665 Ops.back().getOpcode() == ISD::ConstantFP) && 3666 "Scalar binop didn't fold!"); 3667 } 3668 3669 if (Ops.size() == LHS.getNumOperands()-2) { 3670 Ops.push_back(*(LHS.Val->op_end()-2)); 3671 Ops.push_back(*(LHS.Val->op_end()-1)); 3672 return DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size()); 3673 } 3674 } 3675 3676 return SDOperand(); 3677} 3678 3679SDOperand DAGCombiner::SimplifySelect(SDOperand N0, SDOperand N1, SDOperand N2){ 3680 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!"); 3681 3682 SDOperand SCC = SimplifySelectCC(N0.getOperand(0), N0.getOperand(1), N1, N2, 3683 cast<CondCodeSDNode>(N0.getOperand(2))->get()); 3684 // If we got a simplified select_cc node back from SimplifySelectCC, then 3685 // break it down into a new SETCC node, and a new SELECT node, and then return 3686 // the SELECT node, since we were called with a SELECT node. 3687 if (SCC.Val) { 3688 // Check to see if we got a select_cc back (to turn into setcc/select). 3689 // Otherwise, just return whatever node we got back, like fabs. 3690 if (SCC.getOpcode() == ISD::SELECT_CC) { 3691 SDOperand SETCC = DAG.getNode(ISD::SETCC, N0.getValueType(), 3692 SCC.getOperand(0), SCC.getOperand(1), 3693 SCC.getOperand(4)); 3694 AddToWorkList(SETCC.Val); 3695 return DAG.getNode(ISD::SELECT, SCC.getValueType(), SCC.getOperand(2), 3696 SCC.getOperand(3), SETCC); 3697 } 3698 return SCC; 3699 } 3700 return SDOperand(); 3701} 3702 3703/// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS 3704/// are the two values being selected between, see if we can simplify the 3705/// select. Callers of this should assume that TheSelect is deleted if this 3706/// returns true. As such, they should return the appropriate thing (e.g. the 3707/// node) back to the top-level of the DAG combiner loop to avoid it being 3708/// looked at. 3709/// 3710bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDOperand LHS, 3711 SDOperand RHS) { 3712 3713 // If this is a select from two identical things, try to pull the operation 3714 // through the select. 3715 if (LHS.getOpcode() == RHS.getOpcode() && LHS.hasOneUse() && RHS.hasOneUse()){ 3716 // If this is a load and the token chain is identical, replace the select 3717 // of two loads with a load through a select of the address to load from. 3718 // This triggers in things like "select bool X, 10.0, 123.0" after the FP 3719 // constants have been dropped into the constant pool. 3720 if (LHS.getOpcode() == ISD::LOAD && 3721 // Token chains must be identical. 3722 LHS.getOperand(0) == RHS.getOperand(0)) { 3723 LoadSDNode *LLD = cast<LoadSDNode>(LHS); 3724 LoadSDNode *RLD = cast<LoadSDNode>(RHS); 3725 3726 // If this is an EXTLOAD, the VT's must match. 3727 if (LLD->getLoadedVT() == RLD->getLoadedVT()) { 3728 // FIXME: this conflates two src values, discarding one. This is not 3729 // the right thing to do, but nothing uses srcvalues now. When they do, 3730 // turn SrcValue into a list of locations. 3731 SDOperand Addr; 3732 if (TheSelect->getOpcode() == ISD::SELECT) { 3733 // Check that the condition doesn't reach either load. If so, folding 3734 // this will induce a cycle into the DAG. 3735 if (!LLD->isPredecessor(TheSelect->getOperand(0).Val) && 3736 !RLD->isPredecessor(TheSelect->getOperand(0).Val)) { 3737 Addr = DAG.getNode(ISD::SELECT, LLD->getBasePtr().getValueType(), 3738 TheSelect->getOperand(0), LLD->getBasePtr(), 3739 RLD->getBasePtr()); 3740 } 3741 } else { 3742 // Check that the condition doesn't reach either load. If so, folding 3743 // this will induce a cycle into the DAG. 3744 if (!LLD->isPredecessor(TheSelect->getOperand(0).Val) && 3745 !RLD->isPredecessor(TheSelect->getOperand(0).Val) && 3746 !LLD->isPredecessor(TheSelect->getOperand(1).Val) && 3747 !RLD->isPredecessor(TheSelect->getOperand(1).Val)) { 3748 Addr = DAG.getNode(ISD::SELECT_CC, LLD->getBasePtr().getValueType(), 3749 TheSelect->getOperand(0), 3750 TheSelect->getOperand(1), 3751 LLD->getBasePtr(), RLD->getBasePtr(), 3752 TheSelect->getOperand(4)); 3753 } 3754 } 3755 3756 if (Addr.Val) { 3757 SDOperand Load; 3758 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) 3759 Load = DAG.getLoad(TheSelect->getValueType(0), LLD->getChain(), 3760 Addr,LLD->getSrcValue(), 3761 LLD->getSrcValueOffset()); 3762 else { 3763 Load = DAG.getExtLoad(LLD->getExtensionType(), 3764 TheSelect->getValueType(0), 3765 LLD->getChain(), Addr, LLD->getSrcValue(), 3766 LLD->getSrcValueOffset(), 3767 LLD->getLoadedVT()); 3768 } 3769 // Users of the select now use the result of the load. 3770 CombineTo(TheSelect, Load); 3771 3772 // Users of the old loads now use the new load's chain. We know the 3773 // old-load value is dead now. 3774 CombineTo(LHS.Val, Load.getValue(0), Load.getValue(1)); 3775 CombineTo(RHS.Val, Load.getValue(0), Load.getValue(1)); 3776 return true; 3777 } 3778 } 3779 } 3780 } 3781 3782 return false; 3783} 3784 3785SDOperand DAGCombiner::SimplifySelectCC(SDOperand N0, SDOperand N1, 3786 SDOperand N2, SDOperand N3, 3787 ISD::CondCode CC) { 3788 3789 MVT::ValueType VT = N2.getValueType(); 3790 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.Val); 3791 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.Val); 3792 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.Val); 3793 3794 // Determine if the condition we're dealing with is constant 3795 SDOperand SCC = SimplifySetCC(TLI.getSetCCResultTy(), N0, N1, CC, false); 3796 if (SCC.Val) AddToWorkList(SCC.Val); 3797 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.Val); 3798 3799 // fold select_cc true, x, y -> x 3800 if (SCCC && SCCC->getValue()) 3801 return N2; 3802 // fold select_cc false, x, y -> y 3803 if (SCCC && SCCC->getValue() == 0) 3804 return N3; 3805 3806 // Check to see if we can simplify the select into an fabs node 3807 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) { 3808 // Allow either -0.0 or 0.0 3809 if (CFP->getValue() == 0.0) { 3810 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs 3811 if ((CC == ISD::SETGE || CC == ISD::SETGT) && 3812 N0 == N2 && N3.getOpcode() == ISD::FNEG && 3813 N2 == N3.getOperand(0)) 3814 return DAG.getNode(ISD::FABS, VT, N0); 3815 3816 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs 3817 if ((CC == ISD::SETLT || CC == ISD::SETLE) && 3818 N0 == N3 && N2.getOpcode() == ISD::FNEG && 3819 N2.getOperand(0) == N3) 3820 return DAG.getNode(ISD::FABS, VT, N3); 3821 } 3822 } 3823 3824 // Check to see if we can perform the "gzip trick", transforming 3825 // select_cc setlt X, 0, A, 0 -> and (sra X, size(X)-1), A 3826 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT && 3827 MVT::isInteger(N0.getValueType()) && 3828 MVT::isInteger(N2.getValueType()) && 3829 (N1C->isNullValue() || // (a < 0) ? b : 0 3830 (N1C->getValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0 3831 MVT::ValueType XType = N0.getValueType(); 3832 MVT::ValueType AType = N2.getValueType(); 3833 if (XType >= AType) { 3834 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a 3835 // single-bit constant. 3836 if (N2C && ((N2C->getValue() & (N2C->getValue()-1)) == 0)) { 3837 unsigned ShCtV = Log2_64(N2C->getValue()); 3838 ShCtV = MVT::getSizeInBits(XType)-ShCtV-1; 3839 SDOperand ShCt = DAG.getConstant(ShCtV, TLI.getShiftAmountTy()); 3840 SDOperand Shift = DAG.getNode(ISD::SRL, XType, N0, ShCt); 3841 AddToWorkList(Shift.Val); 3842 if (XType > AType) { 3843 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift); 3844 AddToWorkList(Shift.Val); 3845 } 3846 return DAG.getNode(ISD::AND, AType, Shift, N2); 3847 } 3848 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0, 3849 DAG.getConstant(MVT::getSizeInBits(XType)-1, 3850 TLI.getShiftAmountTy())); 3851 AddToWorkList(Shift.Val); 3852 if (XType > AType) { 3853 Shift = DAG.getNode(ISD::TRUNCATE, AType, Shift); 3854 AddToWorkList(Shift.Val); 3855 } 3856 return DAG.getNode(ISD::AND, AType, Shift, N2); 3857 } 3858 } 3859 3860 // fold select C, 16, 0 -> shl C, 4 3861 if (N2C && N3C && N3C->isNullValue() && isPowerOf2_64(N2C->getValue()) && 3862 TLI.getSetCCResultContents() == TargetLowering::ZeroOrOneSetCCResult) { 3863 // Get a SetCC of the condition 3864 // FIXME: Should probably make sure that setcc is legal if we ever have a 3865 // target where it isn't. 3866 SDOperand Temp, SCC; 3867 // cast from setcc result type to select result type 3868 if (AfterLegalize) { 3869 SCC = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC); 3870 if (N2.getValueType() < SCC.getValueType()) 3871 Temp = DAG.getZeroExtendInReg(SCC, N2.getValueType()); 3872 else 3873 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC); 3874 } else { 3875 SCC = DAG.getSetCC(MVT::i1, N0, N1, CC); 3876 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getValueType(), SCC); 3877 } 3878 AddToWorkList(SCC.Val); 3879 AddToWorkList(Temp.Val); 3880 // shl setcc result by log2 n2c 3881 return DAG.getNode(ISD::SHL, N2.getValueType(), Temp, 3882 DAG.getConstant(Log2_64(N2C->getValue()), 3883 TLI.getShiftAmountTy())); 3884 } 3885 3886 // Check to see if this is the equivalent of setcc 3887 // FIXME: Turn all of these into setcc if setcc if setcc is legal 3888 // otherwise, go ahead with the folds. 3889 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getValue() == 1ULL)) { 3890 MVT::ValueType XType = N0.getValueType(); 3891 if (TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultTy())) { 3892 SDOperand Res = DAG.getSetCC(TLI.getSetCCResultTy(), N0, N1, CC); 3893 if (Res.getValueType() != VT) 3894 Res = DAG.getNode(ISD::ZERO_EXTEND, VT, Res); 3895 return Res; 3896 } 3897 3898 // seteq X, 0 -> srl (ctlz X, log2(size(X))) 3899 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ && 3900 TLI.isOperationLegal(ISD::CTLZ, XType)) { 3901 SDOperand Ctlz = DAG.getNode(ISD::CTLZ, XType, N0); 3902 return DAG.getNode(ISD::SRL, XType, Ctlz, 3903 DAG.getConstant(Log2_32(MVT::getSizeInBits(XType)), 3904 TLI.getShiftAmountTy())); 3905 } 3906 // setgt X, 0 -> srl (and (-X, ~X), size(X)-1) 3907 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) { 3908 SDOperand NegN0 = DAG.getNode(ISD::SUB, XType, DAG.getConstant(0, XType), 3909 N0); 3910 SDOperand NotN0 = DAG.getNode(ISD::XOR, XType, N0, 3911 DAG.getConstant(~0ULL, XType)); 3912 return DAG.getNode(ISD::SRL, XType, 3913 DAG.getNode(ISD::AND, XType, NegN0, NotN0), 3914 DAG.getConstant(MVT::getSizeInBits(XType)-1, 3915 TLI.getShiftAmountTy())); 3916 } 3917 // setgt X, -1 -> xor (srl (X, size(X)-1), 1) 3918 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) { 3919 SDOperand Sign = DAG.getNode(ISD::SRL, XType, N0, 3920 DAG.getConstant(MVT::getSizeInBits(XType)-1, 3921 TLI.getShiftAmountTy())); 3922 return DAG.getNode(ISD::XOR, XType, Sign, DAG.getConstant(1, XType)); 3923 } 3924 } 3925 3926 // Check to see if this is an integer abs. select_cc setl[te] X, 0, -X, X -> 3927 // Y = sra (X, size(X)-1); xor (add (X, Y), Y) 3928 if (N1C && N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE) && 3929 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1)) { 3930 if (ConstantSDNode *SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0))) { 3931 MVT::ValueType XType = N0.getValueType(); 3932 if (SubC->isNullValue() && MVT::isInteger(XType)) { 3933 SDOperand Shift = DAG.getNode(ISD::SRA, XType, N0, 3934 DAG.getConstant(MVT::getSizeInBits(XType)-1, 3935 TLI.getShiftAmountTy())); 3936 SDOperand Add = DAG.getNode(ISD::ADD, XType, N0, Shift); 3937 AddToWorkList(Shift.Val); 3938 AddToWorkList(Add.Val); 3939 return DAG.getNode(ISD::XOR, XType, Add, Shift); 3940 } 3941 } 3942 } 3943 3944 return SDOperand(); 3945} 3946 3947/// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC. 3948SDOperand DAGCombiner::SimplifySetCC(MVT::ValueType VT, SDOperand N0, 3949 SDOperand N1, ISD::CondCode Cond, 3950 bool foldBooleans) { 3951 TargetLowering::DAGCombinerInfo 3952 DagCombineInfo(DAG, !AfterLegalize, false, this); 3953 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo); 3954} 3955 3956/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 3957/// return a DAG expression to select that will generate the same value by 3958/// multiplying by a magic number. See: 3959/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3960SDOperand DAGCombiner::BuildSDIV(SDNode *N) { 3961 std::vector<SDNode*> Built; 3962 SDOperand S = TLI.BuildSDIV(N, DAG, &Built); 3963 3964 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end(); 3965 ii != ee; ++ii) 3966 AddToWorkList(*ii); 3967 return S; 3968} 3969 3970/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 3971/// return a DAG expression to select that will generate the same value by 3972/// multiplying by a magic number. See: 3973/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3974SDOperand DAGCombiner::BuildUDIV(SDNode *N) { 3975 std::vector<SDNode*> Built; 3976 SDOperand S = TLI.BuildUDIV(N, DAG, &Built); 3977 3978 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end(); 3979 ii != ee; ++ii) 3980 AddToWorkList(*ii); 3981 return S; 3982} 3983 3984/// FindBaseOffset - Return true if base is known not to alias with anything 3985/// but itself. Provides base object and offset as results. 3986static bool FindBaseOffset(SDOperand Ptr, SDOperand &Base, int64_t &Offset) { 3987 // Assume it is a primitive operation. 3988 Base = Ptr; Offset = 0; 3989 3990 // If it's an adding a simple constant then integrate the offset. 3991 if (Base.getOpcode() == ISD::ADD) { 3992 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) { 3993 Base = Base.getOperand(0); 3994 Offset += C->getValue(); 3995 } 3996 } 3997 3998 // If it's any of the following then it can't alias with anything but itself. 3999 return isa<FrameIndexSDNode>(Base) || 4000 isa<ConstantPoolSDNode>(Base) || 4001 isa<GlobalAddressSDNode>(Base); 4002} 4003 4004/// isAlias - Return true if there is any possibility that the two addresses 4005/// overlap. 4006bool DAGCombiner::isAlias(SDOperand Ptr1, int64_t Size1, 4007 const Value *SrcValue1, int SrcValueOffset1, 4008 SDOperand Ptr2, int64_t Size2, 4009 const Value *SrcValue2, int SrcValueOffset2) 4010{ 4011 // If they are the same then they must be aliases. 4012 if (Ptr1 == Ptr2) return true; 4013 4014 // Gather base node and offset information. 4015 SDOperand Base1, Base2; 4016 int64_t Offset1, Offset2; 4017 bool KnownBase1 = FindBaseOffset(Ptr1, Base1, Offset1); 4018 bool KnownBase2 = FindBaseOffset(Ptr2, Base2, Offset2); 4019 4020 // If they have a same base address then... 4021 if (Base1 == Base2) { 4022 // Check to see if the addresses overlap. 4023 return!((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1); 4024 } 4025 4026 // If we know both bases then they can't alias. 4027 if (KnownBase1 && KnownBase2) return false; 4028 4029 if (CombinerGlobalAA) { 4030 // Use alias analysis information. 4031 int Overlap1 = Size1 + SrcValueOffset1 + Offset1; 4032 int Overlap2 = Size2 + SrcValueOffset2 + Offset2; 4033 AliasAnalysis::AliasResult AAResult = 4034 AA.alias(SrcValue1, Overlap1, SrcValue2, Overlap2); 4035 if (AAResult == AliasAnalysis::NoAlias) 4036 return false; 4037 } 4038 4039 // Otherwise we have to assume they alias. 4040 return true; 4041} 4042 4043/// FindAliasInfo - Extracts the relevant alias information from the memory 4044/// node. Returns true if the operand was a load. 4045bool DAGCombiner::FindAliasInfo(SDNode *N, 4046 SDOperand &Ptr, int64_t &Size, 4047 const Value *&SrcValue, int &SrcValueOffset) { 4048 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 4049 Ptr = LD->getBasePtr(); 4050 Size = MVT::getSizeInBits(LD->getLoadedVT()) >> 3; 4051 SrcValue = LD->getSrcValue(); 4052 SrcValueOffset = LD->getSrcValueOffset(); 4053 return true; 4054 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 4055 Ptr = ST->getBasePtr(); 4056 Size = MVT::getSizeInBits(ST->getStoredVT()) >> 3; 4057 SrcValue = ST->getSrcValue(); 4058 SrcValueOffset = ST->getSrcValueOffset(); 4059 } else { 4060 assert(0 && "FindAliasInfo expected a memory operand"); 4061 } 4062 4063 return false; 4064} 4065 4066/// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes, 4067/// looking for aliasing nodes and adding them to the Aliases vector. 4068void DAGCombiner::GatherAllAliases(SDNode *N, SDOperand OriginalChain, 4069 SmallVector<SDOperand, 8> &Aliases) { 4070 SmallVector<SDOperand, 8> Chains; // List of chains to visit. 4071 std::set<SDNode *> Visited; // Visited node set. 4072 4073 // Get alias information for node. 4074 SDOperand Ptr; 4075 int64_t Size; 4076 const Value *SrcValue; 4077 int SrcValueOffset; 4078 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset); 4079 4080 // Starting off. 4081 Chains.push_back(OriginalChain); 4082 4083 // Look at each chain and determine if it is an alias. If so, add it to the 4084 // aliases list. If not, then continue up the chain looking for the next 4085 // candidate. 4086 while (!Chains.empty()) { 4087 SDOperand Chain = Chains.back(); 4088 Chains.pop_back(); 4089 4090 // Don't bother if we've been before. 4091 if (Visited.find(Chain.Val) != Visited.end()) continue; 4092 Visited.insert(Chain.Val); 4093 4094 switch (Chain.getOpcode()) { 4095 case ISD::EntryToken: 4096 // Entry token is ideal chain operand, but handled in FindBetterChain. 4097 break; 4098 4099 case ISD::LOAD: 4100 case ISD::STORE: { 4101 // Get alias information for Chain. 4102 SDOperand OpPtr; 4103 int64_t OpSize; 4104 const Value *OpSrcValue; 4105 int OpSrcValueOffset; 4106 bool IsOpLoad = FindAliasInfo(Chain.Val, OpPtr, OpSize, 4107 OpSrcValue, OpSrcValueOffset); 4108 4109 // If chain is alias then stop here. 4110 if (!(IsLoad && IsOpLoad) && 4111 isAlias(Ptr, Size, SrcValue, SrcValueOffset, 4112 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset)) { 4113 Aliases.push_back(Chain); 4114 } else { 4115 // Look further up the chain. 4116 Chains.push_back(Chain.getOperand(0)); 4117 // Clean up old chain. 4118 AddToWorkList(Chain.Val); 4119 } 4120 break; 4121 } 4122 4123 case ISD::TokenFactor: 4124 // We have to check each of the operands of the token factor, so we queue 4125 // then up. Adding the operands to the queue (stack) in reverse order 4126 // maintains the original order and increases the likelihood that getNode 4127 // will find a matching token factor (CSE.) 4128 for (unsigned n = Chain.getNumOperands(); n;) 4129 Chains.push_back(Chain.getOperand(--n)); 4130 // Eliminate the token factor if we can. 4131 AddToWorkList(Chain.Val); 4132 break; 4133 4134 default: 4135 // For all other instructions we will just have to take what we can get. 4136 Aliases.push_back(Chain); 4137 break; 4138 } 4139 } 4140} 4141 4142/// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking 4143/// for a better chain (aliasing node.) 4144SDOperand DAGCombiner::FindBetterChain(SDNode *N, SDOperand OldChain) { 4145 SmallVector<SDOperand, 8> Aliases; // Ops for replacing token factor. 4146 4147 // Accumulate all the aliases to this node. 4148 GatherAllAliases(N, OldChain, Aliases); 4149 4150 if (Aliases.size() == 0) { 4151 // If no operands then chain to entry token. 4152 return DAG.getEntryNode(); 4153 } else if (Aliases.size() == 1) { 4154 // If a single operand then chain to it. We don't need to revisit it. 4155 return Aliases[0]; 4156 } 4157 4158 // Construct a custom tailored token factor. 4159 SDOperand NewChain = DAG.getNode(ISD::TokenFactor, MVT::Other, 4160 &Aliases[0], Aliases.size()); 4161 4162 // Make sure the old chain gets cleaned up. 4163 if (NewChain != OldChain) AddToWorkList(OldChain.Val); 4164 4165 return NewChain; 4166} 4167 4168// SelectionDAG::Combine - This is the entry point for the file. 4169// 4170void SelectionDAG::Combine(bool RunningAfterLegalize, AliasAnalysis &AA) { 4171 if (!RunningAfterLegalize && ViewDAGCombine1) 4172 viewGraph(); 4173 if (RunningAfterLegalize && ViewDAGCombine2) 4174 viewGraph(); 4175 /// run - This is the main entry point to this class. 4176 /// 4177 DAGCombiner(*this, AA).Run(RunningAfterLegalize); 4178} 4179