SelectionDAG.cpp revision 7fb5c2dbec26b7de959269ade8906a7a1b6c7539
1//===-- SelectionDAG.cpp - Implement the SelectionDAG data structures -----===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This implements the SelectionDAG class. 11// 12//===----------------------------------------------------------------------===// 13 14#include "llvm/CodeGen/SelectionDAG.h" 15#include "SDNodeOrdering.h" 16#include "SDNodeDbgValue.h" 17#include "llvm/Constants.h" 18#include "llvm/Analysis/ValueTracking.h" 19#include "llvm/Function.h" 20#include "llvm/GlobalAlias.h" 21#include "llvm/GlobalVariable.h" 22#include "llvm/Intrinsics.h" 23#include "llvm/DerivedTypes.h" 24#include "llvm/Assembly/Writer.h" 25#include "llvm/CallingConv.h" 26#include "llvm/CodeGen/MachineBasicBlock.h" 27#include "llvm/CodeGen/MachineConstantPool.h" 28#include "llvm/CodeGen/MachineFrameInfo.h" 29#include "llvm/CodeGen/MachineModuleInfo.h" 30#include "llvm/CodeGen/PseudoSourceValue.h" 31#include "llvm/Target/TargetRegisterInfo.h" 32#include "llvm/Target/TargetData.h" 33#include "llvm/Target/TargetFrameInfo.h" 34#include "llvm/Target/TargetLowering.h" 35#include "llvm/Target/TargetOptions.h" 36#include "llvm/Target/TargetInstrInfo.h" 37#include "llvm/Target/TargetIntrinsicInfo.h" 38#include "llvm/Target/TargetMachine.h" 39#include "llvm/Support/CommandLine.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/ErrorHandling.h" 42#include "llvm/Support/ManagedStatic.h" 43#include "llvm/Support/MathExtras.h" 44#include "llvm/Support/raw_ostream.h" 45#include "llvm/System/Mutex.h" 46#include "llvm/ADT/SetVector.h" 47#include "llvm/ADT/SmallPtrSet.h" 48#include "llvm/ADT/SmallSet.h" 49#include "llvm/ADT/SmallVector.h" 50#include "llvm/ADT/StringExtras.h" 51#include <algorithm> 52#include <cmath> 53using namespace llvm; 54 55/// makeVTList - Return an instance of the SDVTList struct initialized with the 56/// specified members. 57static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) { 58 SDVTList Res = {VTs, NumVTs}; 59 return Res; 60} 61 62static const fltSemantics *EVTToAPFloatSemantics(EVT VT) { 63 switch (VT.getSimpleVT().SimpleTy) { 64 default: llvm_unreachable("Unknown FP format"); 65 case MVT::f32: return &APFloat::IEEEsingle; 66 case MVT::f64: return &APFloat::IEEEdouble; 67 case MVT::f80: return &APFloat::x87DoubleExtended; 68 case MVT::f128: return &APFloat::IEEEquad; 69 case MVT::ppcf128: return &APFloat::PPCDoubleDouble; 70 } 71} 72 73SelectionDAG::DAGUpdateListener::~DAGUpdateListener() {} 74 75//===----------------------------------------------------------------------===// 76// ConstantFPSDNode Class 77//===----------------------------------------------------------------------===// 78 79/// isExactlyValue - We don't rely on operator== working on double values, as 80/// it returns true for things that are clearly not equal, like -0.0 and 0.0. 81/// As such, this method can be used to do an exact bit-for-bit comparison of 82/// two floating point values. 83bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const { 84 return getValueAPF().bitwiseIsEqual(V); 85} 86 87bool ConstantFPSDNode::isValueValidForType(EVT VT, 88 const APFloat& Val) { 89 assert(VT.isFloatingPoint() && "Can only convert between FP types"); 90 91 // PPC long double cannot be converted to any other type. 92 if (VT == MVT::ppcf128 || 93 &Val.getSemantics() == &APFloat::PPCDoubleDouble) 94 return false; 95 96 // convert modifies in place, so make a copy. 97 APFloat Val2 = APFloat(Val); 98 bool losesInfo; 99 (void) Val2.convert(*EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven, 100 &losesInfo); 101 return !losesInfo; 102} 103 104//===----------------------------------------------------------------------===// 105// ISD Namespace 106//===----------------------------------------------------------------------===// 107 108/// isBuildVectorAllOnes - Return true if the specified node is a 109/// BUILD_VECTOR where all of the elements are ~0 or undef. 110bool ISD::isBuildVectorAllOnes(const SDNode *N) { 111 // Look through a bit convert. 112 if (N->getOpcode() == ISD::BIT_CONVERT) 113 N = N->getOperand(0).getNode(); 114 115 if (N->getOpcode() != ISD::BUILD_VECTOR) return false; 116 117 unsigned i = 0, e = N->getNumOperands(); 118 119 // Skip over all of the undef values. 120 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF) 121 ++i; 122 123 // Do not accept an all-undef vector. 124 if (i == e) return false; 125 126 // Do not accept build_vectors that aren't all constants or which have non-~0 127 // elements. 128 SDValue NotZero = N->getOperand(i); 129 if (isa<ConstantSDNode>(NotZero)) { 130 if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue()) 131 return false; 132 } else if (isa<ConstantFPSDNode>(NotZero)) { 133 if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF(). 134 bitcastToAPInt().isAllOnesValue()) 135 return false; 136 } else 137 return false; 138 139 // Okay, we have at least one ~0 value, check to see if the rest match or are 140 // undefs. 141 for (++i; i != e; ++i) 142 if (N->getOperand(i) != NotZero && 143 N->getOperand(i).getOpcode() != ISD::UNDEF) 144 return false; 145 return true; 146} 147 148 149/// isBuildVectorAllZeros - Return true if the specified node is a 150/// BUILD_VECTOR where all of the elements are 0 or undef. 151bool ISD::isBuildVectorAllZeros(const SDNode *N) { 152 // Look through a bit convert. 153 if (N->getOpcode() == ISD::BIT_CONVERT) 154 N = N->getOperand(0).getNode(); 155 156 if (N->getOpcode() != ISD::BUILD_VECTOR) return false; 157 158 unsigned i = 0, e = N->getNumOperands(); 159 160 // Skip over all of the undef values. 161 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF) 162 ++i; 163 164 // Do not accept an all-undef vector. 165 if (i == e) return false; 166 167 // Do not accept build_vectors that aren't all constants or which have non-0 168 // elements. 169 SDValue Zero = N->getOperand(i); 170 if (isa<ConstantSDNode>(Zero)) { 171 if (!cast<ConstantSDNode>(Zero)->isNullValue()) 172 return false; 173 } else if (isa<ConstantFPSDNode>(Zero)) { 174 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero()) 175 return false; 176 } else 177 return false; 178 179 // Okay, we have at least one 0 value, check to see if the rest match or are 180 // undefs. 181 for (++i; i != e; ++i) 182 if (N->getOperand(i) != Zero && 183 N->getOperand(i).getOpcode() != ISD::UNDEF) 184 return false; 185 return true; 186} 187 188/// isScalarToVector - Return true if the specified node is a 189/// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low 190/// element is not an undef. 191bool ISD::isScalarToVector(const SDNode *N) { 192 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) 193 return true; 194 195 if (N->getOpcode() != ISD::BUILD_VECTOR) 196 return false; 197 if (N->getOperand(0).getOpcode() == ISD::UNDEF) 198 return false; 199 unsigned NumElems = N->getNumOperands(); 200 for (unsigned i = 1; i < NumElems; ++i) { 201 SDValue V = N->getOperand(i); 202 if (V.getOpcode() != ISD::UNDEF) 203 return false; 204 } 205 return true; 206} 207 208/// getSetCCSwappedOperands - Return the operation corresponding to (Y op X) 209/// when given the operation for (X op Y). 210ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) { 211 // To perform this operation, we just need to swap the L and G bits of the 212 // operation. 213 unsigned OldL = (Operation >> 2) & 1; 214 unsigned OldG = (Operation >> 1) & 1; 215 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits 216 (OldL << 1) | // New G bit 217 (OldG << 2)); // New L bit. 218} 219 220/// getSetCCInverse - Return the operation corresponding to !(X op Y), where 221/// 'op' is a valid SetCC operation. 222ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) { 223 unsigned Operation = Op; 224 if (isInteger) 225 Operation ^= 7; // Flip L, G, E bits, but not U. 226 else 227 Operation ^= 15; // Flip all of the condition bits. 228 229 if (Operation > ISD::SETTRUE2) 230 Operation &= ~8; // Don't let N and U bits get set. 231 232 return ISD::CondCode(Operation); 233} 234 235 236/// isSignedOp - For an integer comparison, return 1 if the comparison is a 237/// signed operation and 2 if the result is an unsigned comparison. Return zero 238/// if the operation does not depend on the sign of the input (setne and seteq). 239static int isSignedOp(ISD::CondCode Opcode) { 240 switch (Opcode) { 241 default: llvm_unreachable("Illegal integer setcc operation!"); 242 case ISD::SETEQ: 243 case ISD::SETNE: return 0; 244 case ISD::SETLT: 245 case ISD::SETLE: 246 case ISD::SETGT: 247 case ISD::SETGE: return 1; 248 case ISD::SETULT: 249 case ISD::SETULE: 250 case ISD::SETUGT: 251 case ISD::SETUGE: return 2; 252 } 253} 254 255/// getSetCCOrOperation - Return the result of a logical OR between different 256/// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function 257/// returns SETCC_INVALID if it is not possible to represent the resultant 258/// comparison. 259ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2, 260 bool isInteger) { 261 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3) 262 // Cannot fold a signed integer setcc with an unsigned integer setcc. 263 return ISD::SETCC_INVALID; 264 265 unsigned Op = Op1 | Op2; // Combine all of the condition bits. 266 267 // If the N and U bits get set then the resultant comparison DOES suddenly 268 // care about orderedness, and is true when ordered. 269 if (Op > ISD::SETTRUE2) 270 Op &= ~16; // Clear the U bit if the N bit is set. 271 272 // Canonicalize illegal integer setcc's. 273 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT 274 Op = ISD::SETNE; 275 276 return ISD::CondCode(Op); 277} 278 279/// getSetCCAndOperation - Return the result of a logical AND between different 280/// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This 281/// function returns zero if it is not possible to represent the resultant 282/// comparison. 283ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2, 284 bool isInteger) { 285 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3) 286 // Cannot fold a signed setcc with an unsigned setcc. 287 return ISD::SETCC_INVALID; 288 289 // Combine all of the condition bits. 290 ISD::CondCode Result = ISD::CondCode(Op1 & Op2); 291 292 // Canonicalize illegal integer setcc's. 293 if (isInteger) { 294 switch (Result) { 295 default: break; 296 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT 297 case ISD::SETOEQ: // SETEQ & SETU[LG]E 298 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE 299 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE 300 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE 301 } 302 } 303 304 return Result; 305} 306 307const TargetMachine &SelectionDAG::getTarget() const { 308 return MF->getTarget(); 309} 310 311//===----------------------------------------------------------------------===// 312// SDNode Profile Support 313//===----------------------------------------------------------------------===// 314 315/// AddNodeIDOpcode - Add the node opcode to the NodeID data. 316/// 317static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) { 318 ID.AddInteger(OpC); 319} 320 321/// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them 322/// solely with their pointer. 323static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) { 324 ID.AddPointer(VTList.VTs); 325} 326 327/// AddNodeIDOperands - Various routines for adding operands to the NodeID data. 328/// 329static void AddNodeIDOperands(FoldingSetNodeID &ID, 330 const SDValue *Ops, unsigned NumOps) { 331 for (; NumOps; --NumOps, ++Ops) { 332 ID.AddPointer(Ops->getNode()); 333 ID.AddInteger(Ops->getResNo()); 334 } 335} 336 337/// AddNodeIDOperands - Various routines for adding operands to the NodeID data. 338/// 339static void AddNodeIDOperands(FoldingSetNodeID &ID, 340 const SDUse *Ops, unsigned NumOps) { 341 for (; NumOps; --NumOps, ++Ops) { 342 ID.AddPointer(Ops->getNode()); 343 ID.AddInteger(Ops->getResNo()); 344 } 345} 346 347static void AddNodeIDNode(FoldingSetNodeID &ID, 348 unsigned short OpC, SDVTList VTList, 349 const SDValue *OpList, unsigned N) { 350 AddNodeIDOpcode(ID, OpC); 351 AddNodeIDValueTypes(ID, VTList); 352 AddNodeIDOperands(ID, OpList, N); 353} 354 355/// AddNodeIDCustom - If this is an SDNode with special info, add this info to 356/// the NodeID data. 357static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) { 358 switch (N->getOpcode()) { 359 case ISD::TargetExternalSymbol: 360 case ISD::ExternalSymbol: 361 llvm_unreachable("Should only be used on nodes with operands"); 362 default: break; // Normal nodes don't need extra info. 363 case ISD::TargetConstant: 364 case ISD::Constant: 365 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue()); 366 break; 367 case ISD::TargetConstantFP: 368 case ISD::ConstantFP: { 369 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue()); 370 break; 371 } 372 case ISD::TargetGlobalAddress: 373 case ISD::GlobalAddress: 374 case ISD::TargetGlobalTLSAddress: 375 case ISD::GlobalTLSAddress: { 376 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N); 377 ID.AddPointer(GA->getGlobal()); 378 ID.AddInteger(GA->getOffset()); 379 ID.AddInteger(GA->getTargetFlags()); 380 break; 381 } 382 case ISD::BasicBlock: 383 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock()); 384 break; 385 case ISD::Register: 386 ID.AddInteger(cast<RegisterSDNode>(N)->getReg()); 387 break; 388 389 case ISD::SRCVALUE: 390 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue()); 391 break; 392 case ISD::FrameIndex: 393 case ISD::TargetFrameIndex: 394 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex()); 395 break; 396 case ISD::JumpTable: 397 case ISD::TargetJumpTable: 398 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex()); 399 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags()); 400 break; 401 case ISD::ConstantPool: 402 case ISD::TargetConstantPool: { 403 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N); 404 ID.AddInteger(CP->getAlignment()); 405 ID.AddInteger(CP->getOffset()); 406 if (CP->isMachineConstantPoolEntry()) 407 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID); 408 else 409 ID.AddPointer(CP->getConstVal()); 410 ID.AddInteger(CP->getTargetFlags()); 411 break; 412 } 413 case ISD::LOAD: { 414 const LoadSDNode *LD = cast<LoadSDNode>(N); 415 ID.AddInteger(LD->getMemoryVT().getRawBits()); 416 ID.AddInteger(LD->getRawSubclassData()); 417 break; 418 } 419 case ISD::STORE: { 420 const StoreSDNode *ST = cast<StoreSDNode>(N); 421 ID.AddInteger(ST->getMemoryVT().getRawBits()); 422 ID.AddInteger(ST->getRawSubclassData()); 423 break; 424 } 425 case ISD::ATOMIC_CMP_SWAP: 426 case ISD::ATOMIC_SWAP: 427 case ISD::ATOMIC_LOAD_ADD: 428 case ISD::ATOMIC_LOAD_SUB: 429 case ISD::ATOMIC_LOAD_AND: 430 case ISD::ATOMIC_LOAD_OR: 431 case ISD::ATOMIC_LOAD_XOR: 432 case ISD::ATOMIC_LOAD_NAND: 433 case ISD::ATOMIC_LOAD_MIN: 434 case ISD::ATOMIC_LOAD_MAX: 435 case ISD::ATOMIC_LOAD_UMIN: 436 case ISD::ATOMIC_LOAD_UMAX: { 437 const AtomicSDNode *AT = cast<AtomicSDNode>(N); 438 ID.AddInteger(AT->getMemoryVT().getRawBits()); 439 ID.AddInteger(AT->getRawSubclassData()); 440 break; 441 } 442 case ISD::VECTOR_SHUFFLE: { 443 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 444 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements(); 445 i != e; ++i) 446 ID.AddInteger(SVN->getMaskElt(i)); 447 break; 448 } 449 case ISD::TargetBlockAddress: 450 case ISD::BlockAddress: { 451 ID.AddPointer(cast<BlockAddressSDNode>(N)->getBlockAddress()); 452 ID.AddInteger(cast<BlockAddressSDNode>(N)->getTargetFlags()); 453 break; 454 } 455 } // end switch (N->getOpcode()) 456} 457 458/// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID 459/// data. 460static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) { 461 AddNodeIDOpcode(ID, N->getOpcode()); 462 // Add the return value info. 463 AddNodeIDValueTypes(ID, N->getVTList()); 464 // Add the operand info. 465 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands()); 466 467 // Handle SDNode leafs with special info. 468 AddNodeIDCustom(ID, N); 469} 470 471/// encodeMemSDNodeFlags - Generic routine for computing a value for use in 472/// the CSE map that carries volatility, temporalness, indexing mode, and 473/// extension/truncation information. 474/// 475static inline unsigned 476encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile, 477 bool isNonTemporal) { 478 assert((ConvType & 3) == ConvType && 479 "ConvType may not require more than 2 bits!"); 480 assert((AM & 7) == AM && 481 "AM may not require more than 3 bits!"); 482 return ConvType | 483 (AM << 2) | 484 (isVolatile << 5) | 485 (isNonTemporal << 6); 486} 487 488//===----------------------------------------------------------------------===// 489// SelectionDAG Class 490//===----------------------------------------------------------------------===// 491 492/// doNotCSE - Return true if CSE should not be performed for this node. 493static bool doNotCSE(SDNode *N) { 494 if (N->getValueType(0) == MVT::Flag) 495 return true; // Never CSE anything that produces a flag. 496 497 switch (N->getOpcode()) { 498 default: break; 499 case ISD::HANDLENODE: 500 case ISD::EH_LABEL: 501 return true; // Never CSE these nodes. 502 } 503 504 // Check that remaining values produced are not flags. 505 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i) 506 if (N->getValueType(i) == MVT::Flag) 507 return true; // Never CSE anything that produces a flag. 508 509 return false; 510} 511 512/// RemoveDeadNodes - This method deletes all unreachable nodes in the 513/// SelectionDAG. 514void SelectionDAG::RemoveDeadNodes() { 515 // Create a dummy node (which is not added to allnodes), that adds a reference 516 // to the root node, preventing it from being deleted. 517 HandleSDNode Dummy(getRoot()); 518 519 SmallVector<SDNode*, 128> DeadNodes; 520 521 // Add all obviously-dead nodes to the DeadNodes worklist. 522 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I) 523 if (I->use_empty()) 524 DeadNodes.push_back(I); 525 526 RemoveDeadNodes(DeadNodes); 527 528 // If the root changed (e.g. it was a dead load, update the root). 529 setRoot(Dummy.getValue()); 530} 531 532/// RemoveDeadNodes - This method deletes the unreachable nodes in the 533/// given list, and any nodes that become unreachable as a result. 534void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes, 535 DAGUpdateListener *UpdateListener) { 536 537 // Process the worklist, deleting the nodes and adding their uses to the 538 // worklist. 539 while (!DeadNodes.empty()) { 540 SDNode *N = DeadNodes.pop_back_val(); 541 542 if (UpdateListener) 543 UpdateListener->NodeDeleted(N, 0); 544 545 // Take the node out of the appropriate CSE map. 546 RemoveNodeFromCSEMaps(N); 547 548 // Next, brutally remove the operand list. This is safe to do, as there are 549 // no cycles in the graph. 550 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) { 551 SDUse &Use = *I++; 552 SDNode *Operand = Use.getNode(); 553 Use.set(SDValue()); 554 555 // Now that we removed this operand, see if there are no uses of it left. 556 if (Operand->use_empty()) 557 DeadNodes.push_back(Operand); 558 } 559 560 DeallocateNode(N); 561 } 562} 563 564void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){ 565 SmallVector<SDNode*, 16> DeadNodes(1, N); 566 RemoveDeadNodes(DeadNodes, UpdateListener); 567} 568 569void SelectionDAG::DeleteNode(SDNode *N) { 570 // First take this out of the appropriate CSE map. 571 RemoveNodeFromCSEMaps(N); 572 573 // Finally, remove uses due to operands of this node, remove from the 574 // AllNodes list, and delete the node. 575 DeleteNodeNotInCSEMaps(N); 576} 577 578void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) { 579 assert(N != AllNodes.begin() && "Cannot delete the entry node!"); 580 assert(N->use_empty() && "Cannot delete a node that is not dead!"); 581 582 // Drop all of the operands and decrement used node's use counts. 583 N->DropOperands(); 584 585 DeallocateNode(N); 586} 587 588void SelectionDAG::DeallocateNode(SDNode *N) { 589 if (N->OperandsNeedDelete) 590 delete[] N->OperandList; 591 592 // Set the opcode to DELETED_NODE to help catch bugs when node 593 // memory is reallocated. 594 N->NodeType = ISD::DELETED_NODE; 595 596 NodeAllocator.Deallocate(AllNodes.remove(N)); 597 598 // Remove the ordering of this node. 599 Ordering->remove(N); 600 601 // And its entry in the debug info table, if any. 602 DbgInfo->remove(N); 603} 604 605/// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that 606/// correspond to it. This is useful when we're about to delete or repurpose 607/// the node. We don't want future request for structurally identical nodes 608/// to return N anymore. 609bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) { 610 bool Erased = false; 611 switch (N->getOpcode()) { 612 case ISD::EntryToken: 613 llvm_unreachable("EntryToken should not be in CSEMaps!"); 614 return false; 615 case ISD::HANDLENODE: return false; // noop. 616 case ISD::CONDCODE: 617 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] && 618 "Cond code doesn't exist!"); 619 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0; 620 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0; 621 break; 622 case ISD::ExternalSymbol: 623 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol()); 624 break; 625 case ISD::TargetExternalSymbol: { 626 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N); 627 Erased = TargetExternalSymbols.erase( 628 std::pair<std::string,unsigned char>(ESN->getSymbol(), 629 ESN->getTargetFlags())); 630 break; 631 } 632 case ISD::VALUETYPE: { 633 EVT VT = cast<VTSDNode>(N)->getVT(); 634 if (VT.isExtended()) { 635 Erased = ExtendedValueTypeNodes.erase(VT); 636 } else { 637 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != 0; 638 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = 0; 639 } 640 break; 641 } 642 default: 643 // Remove it from the CSE Map. 644 Erased = CSEMap.RemoveNode(N); 645 break; 646 } 647#ifndef NDEBUG 648 // Verify that the node was actually in one of the CSE maps, unless it has a 649 // flag result (which cannot be CSE'd) or is one of the special cases that are 650 // not subject to CSE. 651 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag && 652 !N->isMachineOpcode() && !doNotCSE(N)) { 653 N->dump(this); 654 dbgs() << "\n"; 655 llvm_unreachable("Node is not in map!"); 656 } 657#endif 658 return Erased; 659} 660 661/// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE 662/// maps and modified in place. Add it back to the CSE maps, unless an identical 663/// node already exists, in which case transfer all its users to the existing 664/// node. This transfer can potentially trigger recursive merging. 665/// 666void 667SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N, 668 DAGUpdateListener *UpdateListener) { 669 // For node types that aren't CSE'd, just act as if no identical node 670 // already exists. 671 if (!doNotCSE(N)) { 672 SDNode *Existing = CSEMap.GetOrInsertNode(N); 673 if (Existing != N) { 674 // If there was already an existing matching node, use ReplaceAllUsesWith 675 // to replace the dead one with the existing one. This can cause 676 // recursive merging of other unrelated nodes down the line. 677 ReplaceAllUsesWith(N, Existing, UpdateListener); 678 679 // N is now dead. Inform the listener if it exists and delete it. 680 if (UpdateListener) 681 UpdateListener->NodeDeleted(N, Existing); 682 DeleteNodeNotInCSEMaps(N); 683 return; 684 } 685 } 686 687 // If the node doesn't already exist, we updated it. Inform a listener if 688 // it exists. 689 if (UpdateListener) 690 UpdateListener->NodeUpdated(N); 691} 692 693/// FindModifiedNodeSlot - Find a slot for the specified node if its operands 694/// were replaced with those specified. If this node is never memoized, 695/// return null, otherwise return a pointer to the slot it would take. If a 696/// node already exists with these operands, the slot will be non-null. 697SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op, 698 void *&InsertPos) { 699 if (doNotCSE(N)) 700 return 0; 701 702 SDValue Ops[] = { Op }; 703 FoldingSetNodeID ID; 704 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1); 705 AddNodeIDCustom(ID, N); 706 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos); 707 return Node; 708} 709 710/// FindModifiedNodeSlot - Find a slot for the specified node if its operands 711/// were replaced with those specified. If this node is never memoized, 712/// return null, otherwise return a pointer to the slot it would take. If a 713/// node already exists with these operands, the slot will be non-null. 714SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, 715 SDValue Op1, SDValue Op2, 716 void *&InsertPos) { 717 if (doNotCSE(N)) 718 return 0; 719 720 SDValue Ops[] = { Op1, Op2 }; 721 FoldingSetNodeID ID; 722 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2); 723 AddNodeIDCustom(ID, N); 724 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos); 725 return Node; 726} 727 728 729/// FindModifiedNodeSlot - Find a slot for the specified node if its operands 730/// were replaced with those specified. If this node is never memoized, 731/// return null, otherwise return a pointer to the slot it would take. If a 732/// node already exists with these operands, the slot will be non-null. 733SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, 734 const SDValue *Ops,unsigned NumOps, 735 void *&InsertPos) { 736 if (doNotCSE(N)) 737 return 0; 738 739 FoldingSetNodeID ID; 740 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps); 741 AddNodeIDCustom(ID, N); 742 SDNode *Node = CSEMap.FindNodeOrInsertPos(ID, InsertPos); 743 return Node; 744} 745 746/// VerifyNode - Sanity check the given node. Aborts if it is invalid. 747void SelectionDAG::VerifyNode(SDNode *N) { 748 switch (N->getOpcode()) { 749 default: 750 break; 751 case ISD::BUILD_PAIR: { 752 EVT VT = N->getValueType(0); 753 assert(N->getNumValues() == 1 && "Too many results!"); 754 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) && 755 "Wrong return type!"); 756 assert(N->getNumOperands() == 2 && "Wrong number of operands!"); 757 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() && 758 "Mismatched operand types!"); 759 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() && 760 "Wrong operand type!"); 761 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() && 762 "Wrong return type size"); 763 break; 764 } 765 case ISD::BUILD_VECTOR: { 766 assert(N->getNumValues() == 1 && "Too many results!"); 767 assert(N->getValueType(0).isVector() && "Wrong return type!"); 768 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() && 769 "Wrong number of operands!"); 770 EVT EltVT = N->getValueType(0).getVectorElementType(); 771 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I) 772 assert((I->getValueType() == EltVT || 773 (EltVT.isInteger() && I->getValueType().isInteger() && 774 EltVT.bitsLE(I->getValueType()))) && 775 "Wrong operand type!"); 776 break; 777 } 778 } 779} 780 781/// getEVTAlignment - Compute the default alignment value for the 782/// given type. 783/// 784unsigned SelectionDAG::getEVTAlignment(EVT VT) const { 785 const Type *Ty = VT == MVT::iPTR ? 786 PointerType::get(Type::getInt8Ty(*getContext()), 0) : 787 VT.getTypeForEVT(*getContext()); 788 789 return TLI.getTargetData()->getABITypeAlignment(Ty); 790} 791 792// EntryNode could meaningfully have debug info if we can find it... 793SelectionDAG::SelectionDAG(TargetLowering &tli, FunctionLoweringInfo &fli) 794 : TLI(tli), FLI(fli), DW(0), 795 EntryNode(ISD::EntryToken, DebugLoc::getUnknownLoc(), 796 getVTList(MVT::Other)), 797 Root(getEntryNode()), Ordering(0) { 798 AllNodes.push_back(&EntryNode); 799 Ordering = new SDNodeOrdering(); 800 DbgInfo = new SDDbgInfo(); 801} 802 803void SelectionDAG::init(MachineFunction &mf, MachineModuleInfo *mmi, 804 DwarfWriter *dw) { 805 MF = &mf; 806 MMI = mmi; 807 DW = dw; 808 Context = &mf.getFunction()->getContext(); 809} 810 811SelectionDAG::~SelectionDAG() { 812 allnodes_clear(); 813 delete Ordering; 814 delete DbgInfo; 815} 816 817void SelectionDAG::allnodes_clear() { 818 assert(&*AllNodes.begin() == &EntryNode); 819 AllNodes.remove(AllNodes.begin()); 820 while (!AllNodes.empty()) 821 DeallocateNode(AllNodes.begin()); 822} 823 824void SelectionDAG::clear() { 825 allnodes_clear(); 826 OperandAllocator.Reset(); 827 CSEMap.clear(); 828 829 ExtendedValueTypeNodes.clear(); 830 ExternalSymbols.clear(); 831 TargetExternalSymbols.clear(); 832 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(), 833 static_cast<CondCodeSDNode*>(0)); 834 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(), 835 static_cast<SDNode*>(0)); 836 837 EntryNode.UseList = 0; 838 AllNodes.push_back(&EntryNode); 839 Root = getEntryNode(); 840 delete Ordering; 841 Ordering = new SDNodeOrdering(); 842 delete DbgInfo; 843 DbgInfo = new SDDbgInfo(); 844} 845 846SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) { 847 return VT.bitsGT(Op.getValueType()) ? 848 getNode(ISD::SIGN_EXTEND, DL, VT, Op) : 849 getNode(ISD::TRUNCATE, DL, VT, Op); 850} 851 852SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) { 853 return VT.bitsGT(Op.getValueType()) ? 854 getNode(ISD::ZERO_EXTEND, DL, VT, Op) : 855 getNode(ISD::TRUNCATE, DL, VT, Op); 856} 857 858SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, EVT VT) { 859 assert(!VT.isVector() && 860 "getZeroExtendInReg should use the vector element type instead of " 861 "the vector type!"); 862 if (Op.getValueType() == VT) return Op; 863 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits(); 864 APInt Imm = APInt::getLowBitsSet(BitWidth, 865 VT.getSizeInBits()); 866 return getNode(ISD::AND, DL, Op.getValueType(), Op, 867 getConstant(Imm, Op.getValueType())); 868} 869 870/// getNOT - Create a bitwise NOT operation as (XOR Val, -1). 871/// 872SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, EVT VT) { 873 EVT EltVT = VT.getScalarType(); 874 SDValue NegOne = 875 getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT); 876 return getNode(ISD::XOR, DL, VT, Val, NegOne); 877} 878 879SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) { 880 EVT EltVT = VT.getScalarType(); 881 assert((EltVT.getSizeInBits() >= 64 || 882 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) && 883 "getConstant with a uint64_t value that doesn't fit in the type!"); 884 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT); 885} 886 887SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) { 888 return getConstant(*ConstantInt::get(*Context, Val), VT, isT); 889} 890 891SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) { 892 assert(VT.isInteger() && "Cannot create FP integer constant!"); 893 894 EVT EltVT = VT.getScalarType(); 895 assert(Val.getBitWidth() == EltVT.getSizeInBits() && 896 "APInt size does not match type size!"); 897 898 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant; 899 FoldingSetNodeID ID; 900 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0); 901 ID.AddPointer(&Val); 902 void *IP = 0; 903 SDNode *N = NULL; 904 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP))) 905 if (!VT.isVector()) 906 return SDValue(N, 0); 907 908 if (!N) { 909 N = NodeAllocator.Allocate<ConstantSDNode>(); 910 new (N) ConstantSDNode(isT, &Val, EltVT); 911 CSEMap.InsertNode(N, IP); 912 AllNodes.push_back(N); 913 } 914 915 SDValue Result(N, 0); 916 if (VT.isVector()) { 917 SmallVector<SDValue, 8> Ops; 918 Ops.assign(VT.getVectorNumElements(), Result); 919 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(), 920 VT, &Ops[0], Ops.size()); 921 } 922 return Result; 923} 924 925SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) { 926 return getConstant(Val, TLI.getPointerTy(), isTarget); 927} 928 929 930SDValue SelectionDAG::getConstantFP(const APFloat& V, EVT VT, bool isTarget) { 931 return getConstantFP(*ConstantFP::get(*getContext(), V), VT, isTarget); 932} 933 934SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){ 935 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!"); 936 937 EVT EltVT = VT.getScalarType(); 938 939 // Do the map lookup using the actual bit pattern for the floating point 940 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and 941 // we don't have issues with SNANs. 942 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP; 943 FoldingSetNodeID ID; 944 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0); 945 ID.AddPointer(&V); 946 void *IP = 0; 947 SDNode *N = NULL; 948 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP))) 949 if (!VT.isVector()) 950 return SDValue(N, 0); 951 952 if (!N) { 953 N = NodeAllocator.Allocate<ConstantFPSDNode>(); 954 new (N) ConstantFPSDNode(isTarget, &V, EltVT); 955 CSEMap.InsertNode(N, IP); 956 AllNodes.push_back(N); 957 } 958 959 SDValue Result(N, 0); 960 if (VT.isVector()) { 961 SmallVector<SDValue, 8> Ops; 962 Ops.assign(VT.getVectorNumElements(), Result); 963 // FIXME DebugLoc info might be appropriate here 964 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(), 965 VT, &Ops[0], Ops.size()); 966 } 967 return Result; 968} 969 970SDValue SelectionDAG::getConstantFP(double Val, EVT VT, bool isTarget) { 971 EVT EltVT = VT.getScalarType(); 972 if (EltVT==MVT::f32) 973 return getConstantFP(APFloat((float)Val), VT, isTarget); 974 else 975 return getConstantFP(APFloat(Val), VT, isTarget); 976} 977 978SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV, 979 EVT VT, int64_t Offset, 980 bool isTargetGA, 981 unsigned char TargetFlags) { 982 assert((TargetFlags == 0 || isTargetGA) && 983 "Cannot set target flags on target-independent globals"); 984 985 // Truncate (with sign-extension) the offset value to the pointer size. 986 EVT PTy = TLI.getPointerTy(); 987 unsigned BitWidth = PTy.getSizeInBits(); 988 if (BitWidth < 64) 989 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth)); 990 991 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV); 992 if (!GVar) { 993 // If GV is an alias then use the aliasee for determining thread-localness. 994 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV)) 995 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false)); 996 } 997 998 unsigned Opc; 999 if (GVar && GVar->isThreadLocal()) 1000 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress; 1001 else 1002 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress; 1003 1004 FoldingSetNodeID ID; 1005 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1006 ID.AddPointer(GV); 1007 ID.AddInteger(Offset); 1008 ID.AddInteger(TargetFlags); 1009 void *IP = 0; 1010 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1011 return SDValue(E, 0); 1012 1013 SDNode *N = NodeAllocator.Allocate<GlobalAddressSDNode>(); 1014 new (N) GlobalAddressSDNode(Opc, GV, VT, Offset, TargetFlags); 1015 CSEMap.InsertNode(N, IP); 1016 AllNodes.push_back(N); 1017 return SDValue(N, 0); 1018} 1019 1020SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) { 1021 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex; 1022 FoldingSetNodeID ID; 1023 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1024 ID.AddInteger(FI); 1025 void *IP = 0; 1026 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1027 return SDValue(E, 0); 1028 1029 SDNode *N = NodeAllocator.Allocate<FrameIndexSDNode>(); 1030 new (N) FrameIndexSDNode(FI, VT, isTarget); 1031 CSEMap.InsertNode(N, IP); 1032 AllNodes.push_back(N); 1033 return SDValue(N, 0); 1034} 1035 1036SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget, 1037 unsigned char TargetFlags) { 1038 assert((TargetFlags == 0 || isTarget) && 1039 "Cannot set target flags on target-independent jump tables"); 1040 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable; 1041 FoldingSetNodeID ID; 1042 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1043 ID.AddInteger(JTI); 1044 ID.AddInteger(TargetFlags); 1045 void *IP = 0; 1046 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1047 return SDValue(E, 0); 1048 1049 SDNode *N = NodeAllocator.Allocate<JumpTableSDNode>(); 1050 new (N) JumpTableSDNode(JTI, VT, isTarget, TargetFlags); 1051 CSEMap.InsertNode(N, IP); 1052 AllNodes.push_back(N); 1053 return SDValue(N, 0); 1054} 1055 1056SDValue SelectionDAG::getConstantPool(Constant *C, EVT VT, 1057 unsigned Alignment, int Offset, 1058 bool isTarget, 1059 unsigned char TargetFlags) { 1060 assert((TargetFlags == 0 || isTarget) && 1061 "Cannot set target flags on target-independent globals"); 1062 if (Alignment == 0) 1063 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType()); 1064 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool; 1065 FoldingSetNodeID ID; 1066 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1067 ID.AddInteger(Alignment); 1068 ID.AddInteger(Offset); 1069 ID.AddPointer(C); 1070 ID.AddInteger(TargetFlags); 1071 void *IP = 0; 1072 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1073 return SDValue(E, 0); 1074 1075 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>(); 1076 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags); 1077 CSEMap.InsertNode(N, IP); 1078 AllNodes.push_back(N); 1079 return SDValue(N, 0); 1080} 1081 1082 1083SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT, 1084 unsigned Alignment, int Offset, 1085 bool isTarget, 1086 unsigned char TargetFlags) { 1087 assert((TargetFlags == 0 || isTarget) && 1088 "Cannot set target flags on target-independent globals"); 1089 if (Alignment == 0) 1090 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType()); 1091 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool; 1092 FoldingSetNodeID ID; 1093 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1094 ID.AddInteger(Alignment); 1095 ID.AddInteger(Offset); 1096 C->AddSelectionDAGCSEId(ID); 1097 ID.AddInteger(TargetFlags); 1098 void *IP = 0; 1099 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1100 return SDValue(E, 0); 1101 1102 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>(); 1103 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags); 1104 CSEMap.InsertNode(N, IP); 1105 AllNodes.push_back(N); 1106 return SDValue(N, 0); 1107} 1108 1109SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) { 1110 FoldingSetNodeID ID; 1111 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0); 1112 ID.AddPointer(MBB); 1113 void *IP = 0; 1114 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1115 return SDValue(E, 0); 1116 1117 SDNode *N = NodeAllocator.Allocate<BasicBlockSDNode>(); 1118 new (N) BasicBlockSDNode(MBB); 1119 CSEMap.InsertNode(N, IP); 1120 AllNodes.push_back(N); 1121 return SDValue(N, 0); 1122} 1123 1124SDValue SelectionDAG::getValueType(EVT VT) { 1125 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >= 1126 ValueTypeNodes.size()) 1127 ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1); 1128 1129 SDNode *&N = VT.isExtended() ? 1130 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy]; 1131 1132 if (N) return SDValue(N, 0); 1133 N = NodeAllocator.Allocate<VTSDNode>(); 1134 new (N) VTSDNode(VT); 1135 AllNodes.push_back(N); 1136 return SDValue(N, 0); 1137} 1138 1139SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) { 1140 SDNode *&N = ExternalSymbols[Sym]; 1141 if (N) return SDValue(N, 0); 1142 N = NodeAllocator.Allocate<ExternalSymbolSDNode>(); 1143 new (N) ExternalSymbolSDNode(false, Sym, 0, VT); 1144 AllNodes.push_back(N); 1145 return SDValue(N, 0); 1146} 1147 1148SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT, 1149 unsigned char TargetFlags) { 1150 SDNode *&N = 1151 TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym, 1152 TargetFlags)]; 1153 if (N) return SDValue(N, 0); 1154 N = NodeAllocator.Allocate<ExternalSymbolSDNode>(); 1155 new (N) ExternalSymbolSDNode(true, Sym, TargetFlags, VT); 1156 AllNodes.push_back(N); 1157 return SDValue(N, 0); 1158} 1159 1160SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) { 1161 if ((unsigned)Cond >= CondCodeNodes.size()) 1162 CondCodeNodes.resize(Cond+1); 1163 1164 if (CondCodeNodes[Cond] == 0) { 1165 CondCodeSDNode *N = NodeAllocator.Allocate<CondCodeSDNode>(); 1166 new (N) CondCodeSDNode(Cond); 1167 CondCodeNodes[Cond] = N; 1168 AllNodes.push_back(N); 1169 } 1170 1171 return SDValue(CondCodeNodes[Cond], 0); 1172} 1173 1174// commuteShuffle - swaps the values of N1 and N2, and swaps all indices in 1175// the shuffle mask M that point at N1 to point at N2, and indices that point 1176// N2 to point at N1. 1177static void commuteShuffle(SDValue &N1, SDValue &N2, SmallVectorImpl<int> &M) { 1178 std::swap(N1, N2); 1179 int NElts = M.size(); 1180 for (int i = 0; i != NElts; ++i) { 1181 if (M[i] >= NElts) 1182 M[i] -= NElts; 1183 else if (M[i] >= 0) 1184 M[i] += NElts; 1185 } 1186} 1187 1188SDValue SelectionDAG::getVectorShuffle(EVT VT, DebugLoc dl, SDValue N1, 1189 SDValue N2, const int *Mask) { 1190 assert(N1.getValueType() == N2.getValueType() && "Invalid VECTOR_SHUFFLE"); 1191 assert(VT.isVector() && N1.getValueType().isVector() && 1192 "Vector Shuffle VTs must be a vectors"); 1193 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType() 1194 && "Vector Shuffle VTs must have same element type"); 1195 1196 // Canonicalize shuffle undef, undef -> undef 1197 if (N1.getOpcode() == ISD::UNDEF && N2.getOpcode() == ISD::UNDEF) 1198 return getUNDEF(VT); 1199 1200 // Validate that all indices in Mask are within the range of the elements 1201 // input to the shuffle. 1202 unsigned NElts = VT.getVectorNumElements(); 1203 SmallVector<int, 8> MaskVec; 1204 for (unsigned i = 0; i != NElts; ++i) { 1205 assert(Mask[i] < (int)(NElts * 2) && "Index out of range"); 1206 MaskVec.push_back(Mask[i]); 1207 } 1208 1209 // Canonicalize shuffle v, v -> v, undef 1210 if (N1 == N2) { 1211 N2 = getUNDEF(VT); 1212 for (unsigned i = 0; i != NElts; ++i) 1213 if (MaskVec[i] >= (int)NElts) MaskVec[i] -= NElts; 1214 } 1215 1216 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask. 1217 if (N1.getOpcode() == ISD::UNDEF) 1218 commuteShuffle(N1, N2, MaskVec); 1219 1220 // Canonicalize all index into lhs, -> shuffle lhs, undef 1221 // Canonicalize all index into rhs, -> shuffle rhs, undef 1222 bool AllLHS = true, AllRHS = true; 1223 bool N2Undef = N2.getOpcode() == ISD::UNDEF; 1224 for (unsigned i = 0; i != NElts; ++i) { 1225 if (MaskVec[i] >= (int)NElts) { 1226 if (N2Undef) 1227 MaskVec[i] = -1; 1228 else 1229 AllLHS = false; 1230 } else if (MaskVec[i] >= 0) { 1231 AllRHS = false; 1232 } 1233 } 1234 if (AllLHS && AllRHS) 1235 return getUNDEF(VT); 1236 if (AllLHS && !N2Undef) 1237 N2 = getUNDEF(VT); 1238 if (AllRHS) { 1239 N1 = getUNDEF(VT); 1240 commuteShuffle(N1, N2, MaskVec); 1241 } 1242 1243 // If Identity shuffle, or all shuffle in to undef, return that node. 1244 bool AllUndef = true; 1245 bool Identity = true; 1246 for (unsigned i = 0; i != NElts; ++i) { 1247 if (MaskVec[i] >= 0 && MaskVec[i] != (int)i) Identity = false; 1248 if (MaskVec[i] >= 0) AllUndef = false; 1249 } 1250 if (Identity && NElts == N1.getValueType().getVectorNumElements()) 1251 return N1; 1252 if (AllUndef) 1253 return getUNDEF(VT); 1254 1255 FoldingSetNodeID ID; 1256 SDValue Ops[2] = { N1, N2 }; 1257 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops, 2); 1258 for (unsigned i = 0; i != NElts; ++i) 1259 ID.AddInteger(MaskVec[i]); 1260 1261 void* IP = 0; 1262 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1263 return SDValue(E, 0); 1264 1265 // Allocate the mask array for the node out of the BumpPtrAllocator, since 1266 // SDNode doesn't have access to it. This memory will be "leaked" when 1267 // the node is deallocated, but recovered when the NodeAllocator is released. 1268 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts); 1269 memcpy(MaskAlloc, &MaskVec[0], NElts * sizeof(int)); 1270 1271 ShuffleVectorSDNode *N = NodeAllocator.Allocate<ShuffleVectorSDNode>(); 1272 new (N) ShuffleVectorSDNode(VT, dl, N1, N2, MaskAlloc); 1273 CSEMap.InsertNode(N, IP); 1274 AllNodes.push_back(N); 1275 return SDValue(N, 0); 1276} 1277 1278SDValue SelectionDAG::getConvertRndSat(EVT VT, DebugLoc dl, 1279 SDValue Val, SDValue DTy, 1280 SDValue STy, SDValue Rnd, SDValue Sat, 1281 ISD::CvtCode Code) { 1282 // If the src and dest types are the same and the conversion is between 1283 // integer types of the same sign or two floats, no conversion is necessary. 1284 if (DTy == STy && 1285 (Code == ISD::CVT_UU || Code == ISD::CVT_SS || Code == ISD::CVT_FF)) 1286 return Val; 1287 1288 FoldingSetNodeID ID; 1289 SDValue Ops[] = { Val, DTy, STy, Rnd, Sat }; 1290 AddNodeIDNode(ID, ISD::CONVERT_RNDSAT, getVTList(VT), &Ops[0], 5); 1291 void* IP = 0; 1292 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1293 return SDValue(E, 0); 1294 1295 CvtRndSatSDNode *N = NodeAllocator.Allocate<CvtRndSatSDNode>(); 1296 new (N) CvtRndSatSDNode(VT, dl, Ops, 5, Code); 1297 CSEMap.InsertNode(N, IP); 1298 AllNodes.push_back(N); 1299 return SDValue(N, 0); 1300} 1301 1302SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) { 1303 FoldingSetNodeID ID; 1304 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0); 1305 ID.AddInteger(RegNo); 1306 void *IP = 0; 1307 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1308 return SDValue(E, 0); 1309 1310 SDNode *N = NodeAllocator.Allocate<RegisterSDNode>(); 1311 new (N) RegisterSDNode(RegNo, VT); 1312 CSEMap.InsertNode(N, IP); 1313 AllNodes.push_back(N); 1314 return SDValue(N, 0); 1315} 1316 1317SDValue SelectionDAG::getEHLabel(DebugLoc dl, SDValue Root, MCSymbol *Label) { 1318 FoldingSetNodeID ID; 1319 SDValue Ops[] = { Root }; 1320 AddNodeIDNode(ID, ISD::EH_LABEL, getVTList(MVT::Other), &Ops[0], 1); 1321 ID.AddPointer(Label); 1322 void *IP = 0; 1323 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1324 return SDValue(E, 0); 1325 1326 SDNode *N = NodeAllocator.Allocate<EHLabelSDNode>(); 1327 new (N) EHLabelSDNode(dl, Root, Label); 1328 CSEMap.InsertNode(N, IP); 1329 AllNodes.push_back(N); 1330 return SDValue(N, 0); 1331} 1332 1333 1334SDValue SelectionDAG::getBlockAddress(BlockAddress *BA, EVT VT, 1335 bool isTarget, 1336 unsigned char TargetFlags) { 1337 unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress; 1338 1339 FoldingSetNodeID ID; 1340 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1341 ID.AddPointer(BA); 1342 ID.AddInteger(TargetFlags); 1343 void *IP = 0; 1344 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1345 return SDValue(E, 0); 1346 1347 SDNode *N = NodeAllocator.Allocate<BlockAddressSDNode>(); 1348 new (N) BlockAddressSDNode(Opc, VT, BA, TargetFlags); 1349 CSEMap.InsertNode(N, IP); 1350 AllNodes.push_back(N); 1351 return SDValue(N, 0); 1352} 1353 1354SDValue SelectionDAG::getSrcValue(const Value *V) { 1355 assert((!V || V->getType()->isPointerTy()) && 1356 "SrcValue is not a pointer?"); 1357 1358 FoldingSetNodeID ID; 1359 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0); 1360 ID.AddPointer(V); 1361 1362 void *IP = 0; 1363 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1364 return SDValue(E, 0); 1365 1366 SDNode *N = NodeAllocator.Allocate<SrcValueSDNode>(); 1367 new (N) SrcValueSDNode(V); 1368 CSEMap.InsertNode(N, IP); 1369 AllNodes.push_back(N); 1370 return SDValue(N, 0); 1371} 1372 1373/// getShiftAmountOperand - Return the specified value casted to 1374/// the target's desired shift amount type. 1375SDValue SelectionDAG::getShiftAmountOperand(SDValue Op) { 1376 EVT OpTy = Op.getValueType(); 1377 MVT ShTy = TLI.getShiftAmountTy(); 1378 if (OpTy == ShTy || OpTy.isVector()) return Op; 1379 1380 ISD::NodeType Opcode = OpTy.bitsGT(ShTy) ? ISD::TRUNCATE : ISD::ZERO_EXTEND; 1381 return getNode(Opcode, Op.getDebugLoc(), ShTy, Op); 1382} 1383 1384/// CreateStackTemporary - Create a stack temporary, suitable for holding the 1385/// specified value type. 1386SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) { 1387 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo(); 1388 unsigned ByteSize = VT.getStoreSize(); 1389 const Type *Ty = VT.getTypeForEVT(*getContext()); 1390 unsigned StackAlign = 1391 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign); 1392 1393 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false); 1394 return getFrameIndex(FrameIdx, TLI.getPointerTy()); 1395} 1396 1397/// CreateStackTemporary - Create a stack temporary suitable for holding 1398/// either of the specified value types. 1399SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) { 1400 unsigned Bytes = std::max(VT1.getStoreSizeInBits(), 1401 VT2.getStoreSizeInBits())/8; 1402 const Type *Ty1 = VT1.getTypeForEVT(*getContext()); 1403 const Type *Ty2 = VT2.getTypeForEVT(*getContext()); 1404 const TargetData *TD = TLI.getTargetData(); 1405 unsigned Align = std::max(TD->getPrefTypeAlignment(Ty1), 1406 TD->getPrefTypeAlignment(Ty2)); 1407 1408 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo(); 1409 int FrameIdx = FrameInfo->CreateStackObject(Bytes, Align, false); 1410 return getFrameIndex(FrameIdx, TLI.getPointerTy()); 1411} 1412 1413SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1, 1414 SDValue N2, ISD::CondCode Cond, DebugLoc dl) { 1415 // These setcc operations always fold. 1416 switch (Cond) { 1417 default: break; 1418 case ISD::SETFALSE: 1419 case ISD::SETFALSE2: return getConstant(0, VT); 1420 case ISD::SETTRUE: 1421 case ISD::SETTRUE2: return getConstant(1, VT); 1422 1423 case ISD::SETOEQ: 1424 case ISD::SETOGT: 1425 case ISD::SETOGE: 1426 case ISD::SETOLT: 1427 case ISD::SETOLE: 1428 case ISD::SETONE: 1429 case ISD::SETO: 1430 case ISD::SETUO: 1431 case ISD::SETUEQ: 1432 case ISD::SETUNE: 1433 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!"); 1434 break; 1435 } 1436 1437 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) { 1438 const APInt &C2 = N2C->getAPIntValue(); 1439 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1440 const APInt &C1 = N1C->getAPIntValue(); 1441 1442 switch (Cond) { 1443 default: llvm_unreachable("Unknown integer setcc!"); 1444 case ISD::SETEQ: return getConstant(C1 == C2, VT); 1445 case ISD::SETNE: return getConstant(C1 != C2, VT); 1446 case ISD::SETULT: return getConstant(C1.ult(C2), VT); 1447 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT); 1448 case ISD::SETULE: return getConstant(C1.ule(C2), VT); 1449 case ISD::SETUGE: return getConstant(C1.uge(C2), VT); 1450 case ISD::SETLT: return getConstant(C1.slt(C2), VT); 1451 case ISD::SETGT: return getConstant(C1.sgt(C2), VT); 1452 case ISD::SETLE: return getConstant(C1.sle(C2), VT); 1453 case ISD::SETGE: return getConstant(C1.sge(C2), VT); 1454 } 1455 } 1456 } 1457 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 1458 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) { 1459 // No compile time operations on this type yet. 1460 if (N1C->getValueType(0) == MVT::ppcf128) 1461 return SDValue(); 1462 1463 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF()); 1464 switch (Cond) { 1465 default: break; 1466 case ISD::SETEQ: if (R==APFloat::cmpUnordered) 1467 return getUNDEF(VT); 1468 // fall through 1469 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT); 1470 case ISD::SETNE: if (R==APFloat::cmpUnordered) 1471 return getUNDEF(VT); 1472 // fall through 1473 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan || 1474 R==APFloat::cmpLessThan, VT); 1475 case ISD::SETLT: if (R==APFloat::cmpUnordered) 1476 return getUNDEF(VT); 1477 // fall through 1478 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT); 1479 case ISD::SETGT: if (R==APFloat::cmpUnordered) 1480 return getUNDEF(VT); 1481 // fall through 1482 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT); 1483 case ISD::SETLE: if (R==APFloat::cmpUnordered) 1484 return getUNDEF(VT); 1485 // fall through 1486 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan || 1487 R==APFloat::cmpEqual, VT); 1488 case ISD::SETGE: if (R==APFloat::cmpUnordered) 1489 return getUNDEF(VT); 1490 // fall through 1491 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan || 1492 R==APFloat::cmpEqual, VT); 1493 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT); 1494 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT); 1495 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered || 1496 R==APFloat::cmpEqual, VT); 1497 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT); 1498 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered || 1499 R==APFloat::cmpLessThan, VT); 1500 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan || 1501 R==APFloat::cmpUnordered, VT); 1502 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT); 1503 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT); 1504 } 1505 } else { 1506 // Ensure that the constant occurs on the RHS. 1507 return getSetCC(dl, VT, N2, N1, ISD::getSetCCSwappedOperands(Cond)); 1508 } 1509 } 1510 1511 // Could not fold it. 1512 return SDValue(); 1513} 1514 1515/// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We 1516/// use this predicate to simplify operations downstream. 1517bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const { 1518 // This predicate is not safe for vector operations. 1519 if (Op.getValueType().isVector()) 1520 return false; 1521 1522 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits(); 1523 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth); 1524} 1525 1526/// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use 1527/// this predicate to simplify operations downstream. Mask is known to be zero 1528/// for bits that V cannot have. 1529bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask, 1530 unsigned Depth) const { 1531 APInt KnownZero, KnownOne; 1532 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth); 1533 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1534 return (KnownZero & Mask) == Mask; 1535} 1536 1537/// ComputeMaskedBits - Determine which of the bits specified in Mask are 1538/// known to be either zero or one and return them in the KnownZero/KnownOne 1539/// bitsets. This code only analyzes bits in Mask, in order to short-circuit 1540/// processing. 1541void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask, 1542 APInt &KnownZero, APInt &KnownOne, 1543 unsigned Depth) const { 1544 unsigned BitWidth = Mask.getBitWidth(); 1545 assert(BitWidth == Op.getValueType().getScalarType().getSizeInBits() && 1546 "Mask size mismatches value type size!"); 1547 1548 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything. 1549 if (Depth == 6 || Mask == 0) 1550 return; // Limit search depth. 1551 1552 APInt KnownZero2, KnownOne2; 1553 1554 switch (Op.getOpcode()) { 1555 case ISD::Constant: 1556 // We know all of the bits for a constant! 1557 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask; 1558 KnownZero = ~KnownOne & Mask; 1559 return; 1560 case ISD::AND: 1561 // If either the LHS or the RHS are Zero, the result is zero. 1562 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 1563 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero, 1564 KnownZero2, KnownOne2, Depth+1); 1565 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1566 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1567 1568 // Output known-1 bits are only known if set in both the LHS & RHS. 1569 KnownOne &= KnownOne2; 1570 // Output known-0 are known to be clear if zero in either the LHS | RHS. 1571 KnownZero |= KnownZero2; 1572 return; 1573 case ISD::OR: 1574 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 1575 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne, 1576 KnownZero2, KnownOne2, Depth+1); 1577 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1578 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1579 1580 // Output known-0 bits are only known if clear in both the LHS & RHS. 1581 KnownZero &= KnownZero2; 1582 // Output known-1 are known to be set if set in either the LHS | RHS. 1583 KnownOne |= KnownOne2; 1584 return; 1585 case ISD::XOR: { 1586 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 1587 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1); 1588 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1589 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1590 1591 // Output known-0 bits are known if clear or set in both the LHS & RHS. 1592 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 1593 // Output known-1 are known to be set if set in only one of the LHS, RHS. 1594 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 1595 KnownZero = KnownZeroOut; 1596 return; 1597 } 1598 case ISD::MUL: { 1599 APInt Mask2 = APInt::getAllOnesValue(BitWidth); 1600 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1); 1601 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1); 1602 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1603 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1604 1605 // If low bits are zero in either operand, output low known-0 bits. 1606 // Also compute a conserative estimate for high known-0 bits. 1607 // More trickiness is possible, but this is sufficient for the 1608 // interesting case of alignment computation. 1609 KnownOne.clear(); 1610 unsigned TrailZ = KnownZero.countTrailingOnes() + 1611 KnownZero2.countTrailingOnes(); 1612 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() + 1613 KnownZero2.countLeadingOnes(), 1614 BitWidth) - BitWidth; 1615 1616 TrailZ = std::min(TrailZ, BitWidth); 1617 LeadZ = std::min(LeadZ, BitWidth); 1618 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) | 1619 APInt::getHighBitsSet(BitWidth, LeadZ); 1620 KnownZero &= Mask; 1621 return; 1622 } 1623 case ISD::UDIV: { 1624 // For the purposes of computing leading zeros we can conservatively 1625 // treat a udiv as a logical right shift by the power of 2 known to 1626 // be less than the denominator. 1627 APInt AllOnes = APInt::getAllOnesValue(BitWidth); 1628 ComputeMaskedBits(Op.getOperand(0), 1629 AllOnes, KnownZero2, KnownOne2, Depth+1); 1630 unsigned LeadZ = KnownZero2.countLeadingOnes(); 1631 1632 KnownOne2.clear(); 1633 KnownZero2.clear(); 1634 ComputeMaskedBits(Op.getOperand(1), 1635 AllOnes, KnownZero2, KnownOne2, Depth+1); 1636 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros(); 1637 if (RHSUnknownLeadingOnes != BitWidth) 1638 LeadZ = std::min(BitWidth, 1639 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1); 1640 1641 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask; 1642 return; 1643 } 1644 case ISD::SELECT: 1645 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1); 1646 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1); 1647 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1648 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1649 1650 // Only known if known in both the LHS and RHS. 1651 KnownOne &= KnownOne2; 1652 KnownZero &= KnownZero2; 1653 return; 1654 case ISD::SELECT_CC: 1655 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1); 1656 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1); 1657 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1658 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1659 1660 // Only known if known in both the LHS and RHS. 1661 KnownOne &= KnownOne2; 1662 KnownZero &= KnownZero2; 1663 return; 1664 case ISD::SADDO: 1665 case ISD::UADDO: 1666 case ISD::SSUBO: 1667 case ISD::USUBO: 1668 case ISD::SMULO: 1669 case ISD::UMULO: 1670 if (Op.getResNo() != 1) 1671 return; 1672 // The boolean result conforms to getBooleanContents. Fall through. 1673 case ISD::SETCC: 1674 // If we know the result of a setcc has the top bits zero, use this info. 1675 if (TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent && 1676 BitWidth > 1) 1677 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1); 1678 return; 1679 case ISD::SHL: 1680 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0 1681 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1682 unsigned ShAmt = SA->getZExtValue(); 1683 1684 // If the shift count is an invalid immediate, don't do anything. 1685 if (ShAmt >= BitWidth) 1686 return; 1687 1688 ComputeMaskedBits(Op.getOperand(0), Mask.lshr(ShAmt), 1689 KnownZero, KnownOne, Depth+1); 1690 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1691 KnownZero <<= ShAmt; 1692 KnownOne <<= ShAmt; 1693 // low bits known zero. 1694 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt); 1695 } 1696 return; 1697 case ISD::SRL: 1698 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0 1699 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1700 unsigned ShAmt = SA->getZExtValue(); 1701 1702 // If the shift count is an invalid immediate, don't do anything. 1703 if (ShAmt >= BitWidth) 1704 return; 1705 1706 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt), 1707 KnownZero, KnownOne, Depth+1); 1708 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1709 KnownZero = KnownZero.lshr(ShAmt); 1710 KnownOne = KnownOne.lshr(ShAmt); 1711 1712 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask; 1713 KnownZero |= HighBits; // High bits known zero. 1714 } 1715 return; 1716 case ISD::SRA: 1717 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1718 unsigned ShAmt = SA->getZExtValue(); 1719 1720 // If the shift count is an invalid immediate, don't do anything. 1721 if (ShAmt >= BitWidth) 1722 return; 1723 1724 APInt InDemandedMask = (Mask << ShAmt); 1725 // If any of the demanded bits are produced by the sign extension, we also 1726 // demand the input sign bit. 1727 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask; 1728 if (HighBits.getBoolValue()) 1729 InDemandedMask |= APInt::getSignBit(BitWidth); 1730 1731 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne, 1732 Depth+1); 1733 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1734 KnownZero = KnownZero.lshr(ShAmt); 1735 KnownOne = KnownOne.lshr(ShAmt); 1736 1737 // Handle the sign bits. 1738 APInt SignBit = APInt::getSignBit(BitWidth); 1739 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask. 1740 1741 if (KnownZero.intersects(SignBit)) { 1742 KnownZero |= HighBits; // New bits are known zero. 1743 } else if (KnownOne.intersects(SignBit)) { 1744 KnownOne |= HighBits; // New bits are known one. 1745 } 1746 } 1747 return; 1748 case ISD::SIGN_EXTEND_INREG: { 1749 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1750 unsigned EBits = EVT.getScalarType().getSizeInBits(); 1751 1752 // Sign extension. Compute the demanded bits in the result that are not 1753 // present in the input. 1754 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask; 1755 1756 APInt InSignBit = APInt::getSignBit(EBits); 1757 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits); 1758 1759 // If the sign extended bits are demanded, we know that the sign 1760 // bit is demanded. 1761 InSignBit.zext(BitWidth); 1762 if (NewBits.getBoolValue()) 1763 InputDemandedBits |= InSignBit; 1764 1765 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits, 1766 KnownZero, KnownOne, Depth+1); 1767 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1768 1769 // If the sign bit of the input is known set or clear, then we know the 1770 // top bits of the result. 1771 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear 1772 KnownZero |= NewBits; 1773 KnownOne &= ~NewBits; 1774 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1775 KnownOne |= NewBits; 1776 KnownZero &= ~NewBits; 1777 } else { // Input sign bit unknown 1778 KnownZero &= ~NewBits; 1779 KnownOne &= ~NewBits; 1780 } 1781 return; 1782 } 1783 case ISD::CTTZ: 1784 case ISD::CTLZ: 1785 case ISD::CTPOP: { 1786 unsigned LowBits = Log2_32(BitWidth)+1; 1787 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits); 1788 KnownOne.clear(); 1789 return; 1790 } 1791 case ISD::LOAD: { 1792 if (ISD::isZEXTLoad(Op.getNode())) { 1793 LoadSDNode *LD = cast<LoadSDNode>(Op); 1794 EVT VT = LD->getMemoryVT(); 1795 unsigned MemBits = VT.getScalarType().getSizeInBits(); 1796 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask; 1797 } 1798 return; 1799 } 1800 case ISD::ZERO_EXTEND: { 1801 EVT InVT = Op.getOperand(0).getValueType(); 1802 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1803 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask; 1804 APInt InMask = Mask; 1805 InMask.trunc(InBits); 1806 KnownZero.trunc(InBits); 1807 KnownOne.trunc(InBits); 1808 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1809 KnownZero.zext(BitWidth); 1810 KnownOne.zext(BitWidth); 1811 KnownZero |= NewBits; 1812 return; 1813 } 1814 case ISD::SIGN_EXTEND: { 1815 EVT InVT = Op.getOperand(0).getValueType(); 1816 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1817 APInt InSignBit = APInt::getSignBit(InBits); 1818 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask; 1819 APInt InMask = Mask; 1820 InMask.trunc(InBits); 1821 1822 // If any of the sign extended bits are demanded, we know that the sign 1823 // bit is demanded. Temporarily set this bit in the mask for our callee. 1824 if (NewBits.getBoolValue()) 1825 InMask |= InSignBit; 1826 1827 KnownZero.trunc(InBits); 1828 KnownOne.trunc(InBits); 1829 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1830 1831 // Note if the sign bit is known to be zero or one. 1832 bool SignBitKnownZero = KnownZero.isNegative(); 1833 bool SignBitKnownOne = KnownOne.isNegative(); 1834 assert(!(SignBitKnownZero && SignBitKnownOne) && 1835 "Sign bit can't be known to be both zero and one!"); 1836 1837 // If the sign bit wasn't actually demanded by our caller, we don't 1838 // want it set in the KnownZero and KnownOne result values. Reset the 1839 // mask and reapply it to the result values. 1840 InMask = Mask; 1841 InMask.trunc(InBits); 1842 KnownZero &= InMask; 1843 KnownOne &= InMask; 1844 1845 KnownZero.zext(BitWidth); 1846 KnownOne.zext(BitWidth); 1847 1848 // If the sign bit is known zero or one, the top bits match. 1849 if (SignBitKnownZero) 1850 KnownZero |= NewBits; 1851 else if (SignBitKnownOne) 1852 KnownOne |= NewBits; 1853 return; 1854 } 1855 case ISD::ANY_EXTEND: { 1856 EVT InVT = Op.getOperand(0).getValueType(); 1857 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1858 APInt InMask = Mask; 1859 InMask.trunc(InBits); 1860 KnownZero.trunc(InBits); 1861 KnownOne.trunc(InBits); 1862 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1863 KnownZero.zext(BitWidth); 1864 KnownOne.zext(BitWidth); 1865 return; 1866 } 1867 case ISD::TRUNCATE: { 1868 EVT InVT = Op.getOperand(0).getValueType(); 1869 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1870 APInt InMask = Mask; 1871 InMask.zext(InBits); 1872 KnownZero.zext(InBits); 1873 KnownOne.zext(InBits); 1874 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1875 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1876 KnownZero.trunc(BitWidth); 1877 KnownOne.trunc(BitWidth); 1878 break; 1879 } 1880 case ISD::AssertZext: { 1881 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1882 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits()); 1883 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero, 1884 KnownOne, Depth+1); 1885 KnownZero |= (~InMask) & Mask; 1886 return; 1887 } 1888 case ISD::FGETSIGN: 1889 // All bits are zero except the low bit. 1890 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1); 1891 return; 1892 1893 case ISD::SUB: { 1894 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) { 1895 // We know that the top bits of C-X are clear if X contains less bits 1896 // than C (i.e. no wrap-around can happen). For example, 20-X is 1897 // positive if we can prove that X is >= 0 and < 16. 1898 if (CLHS->getAPIntValue().isNonNegative()) { 1899 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros(); 1900 // NLZ can't be BitWidth with no sign bit 1901 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1); 1902 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2, 1903 Depth+1); 1904 1905 // If all of the MaskV bits are known to be zero, then we know the 1906 // output top bits are zero, because we now know that the output is 1907 // from [0-C]. 1908 if ((KnownZero2 & MaskV) == MaskV) { 1909 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros(); 1910 // Top bits known zero. 1911 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask; 1912 } 1913 } 1914 } 1915 } 1916 // fall through 1917 case ISD::ADD: { 1918 // Output known-0 bits are known if clear or set in both the low clear bits 1919 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the 1920 // low 3 bits clear. 1921 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes()); 1922 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1); 1923 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1924 unsigned KnownZeroOut = KnownZero2.countTrailingOnes(); 1925 1926 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1); 1927 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1928 KnownZeroOut = std::min(KnownZeroOut, 1929 KnownZero2.countTrailingOnes()); 1930 1931 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut); 1932 return; 1933 } 1934 case ISD::SREM: 1935 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1936 const APInt &RA = Rem->getAPIntValue().abs(); 1937 if (RA.isPowerOf2()) { 1938 APInt LowBits = RA - 1; 1939 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth); 1940 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1); 1941 1942 // The low bits of the first operand are unchanged by the srem. 1943 KnownZero = KnownZero2 & LowBits; 1944 KnownOne = KnownOne2 & LowBits; 1945 1946 // If the first operand is non-negative or has all low bits zero, then 1947 // the upper bits are all zero. 1948 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits)) 1949 KnownZero |= ~LowBits; 1950 1951 // If the first operand is negative and not all low bits are zero, then 1952 // the upper bits are all one. 1953 if (KnownOne2[BitWidth-1] && ((KnownOne2 & LowBits) != 0)) 1954 KnownOne |= ~LowBits; 1955 1956 KnownZero &= Mask; 1957 KnownOne &= Mask; 1958 1959 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?"); 1960 } 1961 } 1962 return; 1963 case ISD::UREM: { 1964 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1965 const APInt &RA = Rem->getAPIntValue(); 1966 if (RA.isPowerOf2()) { 1967 APInt LowBits = (RA - 1); 1968 APInt Mask2 = LowBits & Mask; 1969 KnownZero |= ~LowBits & Mask; 1970 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero, KnownOne,Depth+1); 1971 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?"); 1972 break; 1973 } 1974 } 1975 1976 // Since the result is less than or equal to either operand, any leading 1977 // zero bits in either operand must also exist in the result. 1978 APInt AllOnes = APInt::getAllOnesValue(BitWidth); 1979 ComputeMaskedBits(Op.getOperand(0), AllOnes, KnownZero, KnownOne, 1980 Depth+1); 1981 ComputeMaskedBits(Op.getOperand(1), AllOnes, KnownZero2, KnownOne2, 1982 Depth+1); 1983 1984 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(), 1985 KnownZero2.countLeadingOnes()); 1986 KnownOne.clear(); 1987 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask; 1988 return; 1989 } 1990 default: 1991 // Allow the target to implement this method for its nodes. 1992 if (Op.getOpcode() >= ISD::BUILTIN_OP_END) { 1993 case ISD::INTRINSIC_WO_CHAIN: 1994 case ISD::INTRINSIC_W_CHAIN: 1995 case ISD::INTRINSIC_VOID: 1996 TLI.computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne, *this, 1997 Depth); 1998 } 1999 return; 2000 } 2001} 2002 2003/// ComputeNumSignBits - Return the number of times the sign bit of the 2004/// register is replicated into the other bits. We know that at least 1 bit 2005/// is always equal to the sign bit (itself), but other cases can give us 2006/// information. For example, immediately after an "SRA X, 2", we know that 2007/// the top 3 bits are all equal to each other, so we return 3. 2008unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{ 2009 EVT VT = Op.getValueType(); 2010 assert(VT.isInteger() && "Invalid VT!"); 2011 unsigned VTBits = VT.getScalarType().getSizeInBits(); 2012 unsigned Tmp, Tmp2; 2013 unsigned FirstAnswer = 1; 2014 2015 if (Depth == 6) 2016 return 1; // Limit search depth. 2017 2018 switch (Op.getOpcode()) { 2019 default: break; 2020 case ISD::AssertSext: 2021 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits(); 2022 return VTBits-Tmp+1; 2023 case ISD::AssertZext: 2024 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits(); 2025 return VTBits-Tmp; 2026 2027 case ISD::Constant: { 2028 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue(); 2029 // If negative, return # leading ones. 2030 if (Val.isNegative()) 2031 return Val.countLeadingOnes(); 2032 2033 // Return # leading zeros. 2034 return Val.countLeadingZeros(); 2035 } 2036 2037 case ISD::SIGN_EXTEND: 2038 Tmp = VTBits-Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 2039 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp; 2040 2041 case ISD::SIGN_EXTEND_INREG: 2042 // Max of the input and what this extends. 2043 Tmp = 2044 cast<VTSDNode>(Op.getOperand(1))->getVT().getScalarType().getSizeInBits(); 2045 Tmp = VTBits-Tmp+1; 2046 2047 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2048 return std::max(Tmp, Tmp2); 2049 2050 case ISD::SRA: 2051 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2052 // SRA X, C -> adds C sign bits. 2053 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2054 Tmp += C->getZExtValue(); 2055 if (Tmp > VTBits) Tmp = VTBits; 2056 } 2057 return Tmp; 2058 case ISD::SHL: 2059 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2060 // shl destroys sign bits. 2061 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2062 if (C->getZExtValue() >= VTBits || // Bad shift. 2063 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out. 2064 return Tmp - C->getZExtValue(); 2065 } 2066 break; 2067 case ISD::AND: 2068 case ISD::OR: 2069 case ISD::XOR: // NOT is handled here. 2070 // Logical binary ops preserve the number of sign bits at the worst. 2071 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2072 if (Tmp != 1) { 2073 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2074 FirstAnswer = std::min(Tmp, Tmp2); 2075 // We computed what we know about the sign bits as our first 2076 // answer. Now proceed to the generic code that uses 2077 // ComputeMaskedBits, and pick whichever answer is better. 2078 } 2079 break; 2080 2081 case ISD::SELECT: 2082 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2083 if (Tmp == 1) return 1; // Early out. 2084 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1); 2085 return std::min(Tmp, Tmp2); 2086 2087 case ISD::SADDO: 2088 case ISD::UADDO: 2089 case ISD::SSUBO: 2090 case ISD::USUBO: 2091 case ISD::SMULO: 2092 case ISD::UMULO: 2093 if (Op.getResNo() != 1) 2094 break; 2095 // The boolean result conforms to getBooleanContents. Fall through. 2096 case ISD::SETCC: 2097 // If setcc returns 0/-1, all bits are sign bits. 2098 if (TLI.getBooleanContents() == 2099 TargetLowering::ZeroOrNegativeOneBooleanContent) 2100 return VTBits; 2101 break; 2102 case ISD::ROTL: 2103 case ISD::ROTR: 2104 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2105 unsigned RotAmt = C->getZExtValue() & (VTBits-1); 2106 2107 // Handle rotate right by N like a rotate left by 32-N. 2108 if (Op.getOpcode() == ISD::ROTR) 2109 RotAmt = (VTBits-RotAmt) & (VTBits-1); 2110 2111 // If we aren't rotating out all of the known-in sign bits, return the 2112 // number that are left. This handles rotl(sext(x), 1) for example. 2113 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2114 if (Tmp > RotAmt+1) return Tmp-RotAmt; 2115 } 2116 break; 2117 case ISD::ADD: 2118 // Add can have at most one carry bit. Thus we know that the output 2119 // is, at worst, one more bit than the inputs. 2120 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2121 if (Tmp == 1) return 1; // Early out. 2122 2123 // Special case decrementing a value (ADD X, -1): 2124 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1))) 2125 if (CRHS->isAllOnesValue()) { 2126 APInt KnownZero, KnownOne; 2127 APInt Mask = APInt::getAllOnesValue(VTBits); 2128 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1); 2129 2130 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2131 // sign bits set. 2132 if ((KnownZero | APInt(VTBits, 1)) == Mask) 2133 return VTBits; 2134 2135 // If we are subtracting one from a positive number, there is no carry 2136 // out of the result. 2137 if (KnownZero.isNegative()) 2138 return Tmp; 2139 } 2140 2141 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2142 if (Tmp2 == 1) return 1; 2143 return std::min(Tmp, Tmp2)-1; 2144 break; 2145 2146 case ISD::SUB: 2147 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2148 if (Tmp2 == 1) return 1; 2149 2150 // Handle NEG. 2151 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) 2152 if (CLHS->isNullValue()) { 2153 APInt KnownZero, KnownOne; 2154 APInt Mask = APInt::getAllOnesValue(VTBits); 2155 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 2156 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2157 // sign bits set. 2158 if ((KnownZero | APInt(VTBits, 1)) == Mask) 2159 return VTBits; 2160 2161 // If the input is known to be positive (the sign bit is known clear), 2162 // the output of the NEG has the same number of sign bits as the input. 2163 if (KnownZero.isNegative()) 2164 return Tmp2; 2165 2166 // Otherwise, we treat this like a SUB. 2167 } 2168 2169 // Sub can have at most one carry bit. Thus we know that the output 2170 // is, at worst, one more bit than the inputs. 2171 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2172 if (Tmp == 1) return 1; // Early out. 2173 return std::min(Tmp, Tmp2)-1; 2174 break; 2175 case ISD::TRUNCATE: 2176 // FIXME: it's tricky to do anything useful for this, but it is an important 2177 // case for targets like X86. 2178 break; 2179 } 2180 2181 // Handle LOADX separately here. EXTLOAD case will fallthrough. 2182 if (Op.getOpcode() == ISD::LOAD) { 2183 LoadSDNode *LD = cast<LoadSDNode>(Op); 2184 unsigned ExtType = LD->getExtensionType(); 2185 switch (ExtType) { 2186 default: break; 2187 case ISD::SEXTLOAD: // '17' bits known 2188 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits(); 2189 return VTBits-Tmp+1; 2190 case ISD::ZEXTLOAD: // '16' bits known 2191 Tmp = LD->getMemoryVT().getScalarType().getSizeInBits(); 2192 return VTBits-Tmp; 2193 } 2194 } 2195 2196 // Allow the target to implement this method for its nodes. 2197 if (Op.getOpcode() >= ISD::BUILTIN_OP_END || 2198 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 2199 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 2200 Op.getOpcode() == ISD::INTRINSIC_VOID) { 2201 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth); 2202 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits); 2203 } 2204 2205 // Finally, if we can prove that the top bits of the result are 0's or 1's, 2206 // use this information. 2207 APInt KnownZero, KnownOne; 2208 APInt Mask = APInt::getAllOnesValue(VTBits); 2209 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth); 2210 2211 if (KnownZero.isNegative()) { // sign bit is 0 2212 Mask = KnownZero; 2213 } else if (KnownOne.isNegative()) { // sign bit is 1; 2214 Mask = KnownOne; 2215 } else { 2216 // Nothing known. 2217 return FirstAnswer; 2218 } 2219 2220 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine 2221 // the number of identical bits in the top of the input value. 2222 Mask = ~Mask; 2223 Mask <<= Mask.getBitWidth()-VTBits; 2224 // Return # leading zeros. We use 'min' here in case Val was zero before 2225 // shifting. We don't want to return '64' as for an i32 "0". 2226 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros())); 2227} 2228 2229bool SelectionDAG::isKnownNeverNaN(SDValue Op) const { 2230 // If we're told that NaNs won't happen, assume they won't. 2231 if (FiniteOnlyFPMath()) 2232 return true; 2233 2234 // If the value is a constant, we can obviously see if it is a NaN or not. 2235 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op)) 2236 return !C->getValueAPF().isNaN(); 2237 2238 // TODO: Recognize more cases here. 2239 2240 return false; 2241} 2242 2243bool SelectionDAG::isKnownNeverZero(SDValue Op) const { 2244 // If the value is a constant, we can obviously see if it is a zero or not. 2245 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op)) 2246 return !C->isZero(); 2247 2248 // TODO: Recognize more cases here. 2249 2250 return false; 2251} 2252 2253bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const { 2254 // Check the obvious case. 2255 if (A == B) return true; 2256 2257 // For for negative and positive zero. 2258 if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A)) 2259 if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B)) 2260 if (CA->isZero() && CB->isZero()) return true; 2261 2262 // Otherwise they may not be equal. 2263 return false; 2264} 2265 2266bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const { 2267 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2268 if (!GA) return false; 2269 if (GA->getOffset() != 0) return false; 2270 GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal()); 2271 if (!GV) return false; 2272 MachineModuleInfo *MMI = getMachineModuleInfo(); 2273 return MMI && MMI->hasDebugInfo(); 2274} 2275 2276 2277/// getShuffleScalarElt - Returns the scalar element that will make up the ith 2278/// element of the result of the vector shuffle. 2279SDValue SelectionDAG::getShuffleScalarElt(const ShuffleVectorSDNode *N, 2280 unsigned i) { 2281 EVT VT = N->getValueType(0); 2282 DebugLoc dl = N->getDebugLoc(); 2283 if (N->getMaskElt(i) < 0) 2284 return getUNDEF(VT.getVectorElementType()); 2285 unsigned Index = N->getMaskElt(i); 2286 unsigned NumElems = VT.getVectorNumElements(); 2287 SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1); 2288 Index %= NumElems; 2289 2290 if (V.getOpcode() == ISD::BIT_CONVERT) { 2291 V = V.getOperand(0); 2292 EVT VVT = V.getValueType(); 2293 if (!VVT.isVector() || VVT.getVectorNumElements() != (unsigned)NumElems) 2294 return SDValue(); 2295 } 2296 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) 2297 return (Index == 0) ? V.getOperand(0) 2298 : getUNDEF(VT.getVectorElementType()); 2299 if (V.getOpcode() == ISD::BUILD_VECTOR) 2300 return V.getOperand(Index); 2301 if (const ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(V)) 2302 return getShuffleScalarElt(SVN, Index); 2303 return SDValue(); 2304} 2305 2306 2307/// getNode - Gets or creates the specified node. 2308/// 2309SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT) { 2310 FoldingSetNodeID ID; 2311 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0); 2312 void *IP = 0; 2313 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 2314 return SDValue(E, 0); 2315 2316 SDNode *N = NodeAllocator.Allocate<SDNode>(); 2317 new (N) SDNode(Opcode, DL, getVTList(VT)); 2318 CSEMap.InsertNode(N, IP); 2319 2320 AllNodes.push_back(N); 2321#ifndef NDEBUG 2322 VerifyNode(N); 2323#endif 2324 return SDValue(N, 0); 2325} 2326 2327SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, 2328 EVT VT, SDValue Operand) { 2329 // Constant fold unary operations with an integer constant operand. 2330 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) { 2331 const APInt &Val = C->getAPIntValue(); 2332 switch (Opcode) { 2333 default: break; 2334 case ISD::SIGN_EXTEND: 2335 return getConstant(APInt(Val).sext(VT.getSizeInBits()), VT); 2336 case ISD::ANY_EXTEND: 2337 case ISD::ZERO_EXTEND: 2338 case ISD::TRUNCATE: 2339 return getConstant(APInt(Val).zextOrTrunc(VT.getSizeInBits()), VT); 2340 case ISD::UINT_TO_FP: 2341 case ISD::SINT_TO_FP: { 2342 const uint64_t zero[] = {0, 0}; 2343 // No compile time operations on ppcf128. 2344 if (VT == MVT::ppcf128) break; 2345 APFloat apf = APFloat(APInt(VT.getSizeInBits(), 2, zero)); 2346 (void)apf.convertFromAPInt(Val, 2347 Opcode==ISD::SINT_TO_FP, 2348 APFloat::rmNearestTiesToEven); 2349 return getConstantFP(apf, VT); 2350 } 2351 case ISD::BIT_CONVERT: 2352 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32) 2353 return getConstantFP(Val.bitsToFloat(), VT); 2354 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64) 2355 return getConstantFP(Val.bitsToDouble(), VT); 2356 break; 2357 case ISD::BSWAP: 2358 return getConstant(Val.byteSwap(), VT); 2359 case ISD::CTPOP: 2360 return getConstant(Val.countPopulation(), VT); 2361 case ISD::CTLZ: 2362 return getConstant(Val.countLeadingZeros(), VT); 2363 case ISD::CTTZ: 2364 return getConstant(Val.countTrailingZeros(), VT); 2365 } 2366 } 2367 2368 // Constant fold unary operations with a floating point constant operand. 2369 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) { 2370 APFloat V = C->getValueAPF(); // make copy 2371 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) { 2372 switch (Opcode) { 2373 case ISD::FNEG: 2374 V.changeSign(); 2375 return getConstantFP(V, VT); 2376 case ISD::FABS: 2377 V.clearSign(); 2378 return getConstantFP(V, VT); 2379 case ISD::FP_ROUND: 2380 case ISD::FP_EXTEND: { 2381 bool ignored; 2382 // This can return overflow, underflow, or inexact; we don't care. 2383 // FIXME need to be more flexible about rounding mode. 2384 (void)V.convert(*EVTToAPFloatSemantics(VT), 2385 APFloat::rmNearestTiesToEven, &ignored); 2386 return getConstantFP(V, VT); 2387 } 2388 case ISD::FP_TO_SINT: 2389 case ISD::FP_TO_UINT: { 2390 integerPart x[2]; 2391 bool ignored; 2392 assert(integerPartWidth >= 64); 2393 // FIXME need to be more flexible about rounding mode. 2394 APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(), 2395 Opcode==ISD::FP_TO_SINT, 2396 APFloat::rmTowardZero, &ignored); 2397 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual 2398 break; 2399 APInt api(VT.getSizeInBits(), 2, x); 2400 return getConstant(api, VT); 2401 } 2402 case ISD::BIT_CONVERT: 2403 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32) 2404 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT); 2405 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64) 2406 return getConstant(V.bitcastToAPInt().getZExtValue(), VT); 2407 break; 2408 } 2409 } 2410 } 2411 2412 unsigned OpOpcode = Operand.getNode()->getOpcode(); 2413 switch (Opcode) { 2414 case ISD::TokenFactor: 2415 case ISD::MERGE_VALUES: 2416 case ISD::CONCAT_VECTORS: 2417 return Operand; // Factor, merge or concat of one node? No need. 2418 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node"); 2419 case ISD::FP_EXTEND: 2420 assert(VT.isFloatingPoint() && 2421 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!"); 2422 if (Operand.getValueType() == VT) return Operand; // noop conversion. 2423 assert((!VT.isVector() || 2424 VT.getVectorNumElements() == 2425 Operand.getValueType().getVectorNumElements()) && 2426 "Vector element count mismatch!"); 2427 if (Operand.getOpcode() == ISD::UNDEF) 2428 return getUNDEF(VT); 2429 break; 2430 case ISD::SIGN_EXTEND: 2431 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2432 "Invalid SIGN_EXTEND!"); 2433 if (Operand.getValueType() == VT) return Operand; // noop extension 2434 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) && 2435 "Invalid sext node, dst < src!"); 2436 assert((!VT.isVector() || 2437 VT.getVectorNumElements() == 2438 Operand.getValueType().getVectorNumElements()) && 2439 "Vector element count mismatch!"); 2440 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND) 2441 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0)); 2442 break; 2443 case ISD::ZERO_EXTEND: 2444 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2445 "Invalid ZERO_EXTEND!"); 2446 if (Operand.getValueType() == VT) return Operand; // noop extension 2447 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) && 2448 "Invalid zext node, dst < src!"); 2449 assert((!VT.isVector() || 2450 VT.getVectorNumElements() == 2451 Operand.getValueType().getVectorNumElements()) && 2452 "Vector element count mismatch!"); 2453 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x) 2454 return getNode(ISD::ZERO_EXTEND, DL, VT, 2455 Operand.getNode()->getOperand(0)); 2456 break; 2457 case ISD::ANY_EXTEND: 2458 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2459 "Invalid ANY_EXTEND!"); 2460 if (Operand.getValueType() == VT) return Operand; // noop extension 2461 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) && 2462 "Invalid anyext node, dst < src!"); 2463 assert((!VT.isVector() || 2464 VT.getVectorNumElements() == 2465 Operand.getValueType().getVectorNumElements()) && 2466 "Vector element count mismatch!"); 2467 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND) 2468 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x) 2469 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0)); 2470 break; 2471 case ISD::TRUNCATE: 2472 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2473 "Invalid TRUNCATE!"); 2474 if (Operand.getValueType() == VT) return Operand; // noop truncate 2475 assert(Operand.getValueType().getScalarType().bitsGT(VT.getScalarType()) && 2476 "Invalid truncate node, src < dst!"); 2477 assert((!VT.isVector() || 2478 VT.getVectorNumElements() == 2479 Operand.getValueType().getVectorNumElements()) && 2480 "Vector element count mismatch!"); 2481 if (OpOpcode == ISD::TRUNCATE) 2482 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0)); 2483 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND || 2484 OpOpcode == ISD::ANY_EXTEND) { 2485 // If the source is smaller than the dest, we still need an extend. 2486 if (Operand.getNode()->getOperand(0).getValueType().getScalarType() 2487 .bitsLT(VT.getScalarType())) 2488 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0)); 2489 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT)) 2490 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0)); 2491 else 2492 return Operand.getNode()->getOperand(0); 2493 } 2494 break; 2495 case ISD::BIT_CONVERT: 2496 // Basic sanity checking. 2497 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits() 2498 && "Cannot BIT_CONVERT between types of different sizes!"); 2499 if (VT == Operand.getValueType()) return Operand; // noop conversion. 2500 if (OpOpcode == ISD::BIT_CONVERT) // bitconv(bitconv(x)) -> bitconv(x) 2501 return getNode(ISD::BIT_CONVERT, DL, VT, Operand.getOperand(0)); 2502 if (OpOpcode == ISD::UNDEF) 2503 return getUNDEF(VT); 2504 break; 2505 case ISD::SCALAR_TO_VECTOR: 2506 assert(VT.isVector() && !Operand.getValueType().isVector() && 2507 (VT.getVectorElementType() == Operand.getValueType() || 2508 (VT.getVectorElementType().isInteger() && 2509 Operand.getValueType().isInteger() && 2510 VT.getVectorElementType().bitsLE(Operand.getValueType()))) && 2511 "Illegal SCALAR_TO_VECTOR node!"); 2512 if (OpOpcode == ISD::UNDEF) 2513 return getUNDEF(VT); 2514 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined. 2515 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT && 2516 isa<ConstantSDNode>(Operand.getOperand(1)) && 2517 Operand.getConstantOperandVal(1) == 0 && 2518 Operand.getOperand(0).getValueType() == VT) 2519 return Operand.getOperand(0); 2520 break; 2521 case ISD::FNEG: 2522 // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0 2523 if (UnsafeFPMath && OpOpcode == ISD::FSUB) 2524 return getNode(ISD::FSUB, DL, VT, Operand.getNode()->getOperand(1), 2525 Operand.getNode()->getOperand(0)); 2526 if (OpOpcode == ISD::FNEG) // --X -> X 2527 return Operand.getNode()->getOperand(0); 2528 break; 2529 case ISD::FABS: 2530 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X) 2531 return getNode(ISD::FABS, DL, VT, Operand.getNode()->getOperand(0)); 2532 break; 2533 } 2534 2535 SDNode *N; 2536 SDVTList VTs = getVTList(VT); 2537 if (VT != MVT::Flag) { // Don't CSE flag producing nodes 2538 FoldingSetNodeID ID; 2539 SDValue Ops[1] = { Operand }; 2540 AddNodeIDNode(ID, Opcode, VTs, Ops, 1); 2541 void *IP = 0; 2542 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 2543 return SDValue(E, 0); 2544 2545 N = NodeAllocator.Allocate<UnarySDNode>(); 2546 new (N) UnarySDNode(Opcode, DL, VTs, Operand); 2547 CSEMap.InsertNode(N, IP); 2548 } else { 2549 N = NodeAllocator.Allocate<UnarySDNode>(); 2550 new (N) UnarySDNode(Opcode, DL, VTs, Operand); 2551 } 2552 2553 AllNodes.push_back(N); 2554#ifndef NDEBUG 2555 VerifyNode(N); 2556#endif 2557 return SDValue(N, 0); 2558} 2559 2560SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, 2561 EVT VT, 2562 ConstantSDNode *Cst1, 2563 ConstantSDNode *Cst2) { 2564 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue(); 2565 2566 switch (Opcode) { 2567 case ISD::ADD: return getConstant(C1 + C2, VT); 2568 case ISD::SUB: return getConstant(C1 - C2, VT); 2569 case ISD::MUL: return getConstant(C1 * C2, VT); 2570 case ISD::UDIV: 2571 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT); 2572 break; 2573 case ISD::UREM: 2574 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT); 2575 break; 2576 case ISD::SDIV: 2577 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT); 2578 break; 2579 case ISD::SREM: 2580 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT); 2581 break; 2582 case ISD::AND: return getConstant(C1 & C2, VT); 2583 case ISD::OR: return getConstant(C1 | C2, VT); 2584 case ISD::XOR: return getConstant(C1 ^ C2, VT); 2585 case ISD::SHL: return getConstant(C1 << C2, VT); 2586 case ISD::SRL: return getConstant(C1.lshr(C2), VT); 2587 case ISD::SRA: return getConstant(C1.ashr(C2), VT); 2588 case ISD::ROTL: return getConstant(C1.rotl(C2), VT); 2589 case ISD::ROTR: return getConstant(C1.rotr(C2), VT); 2590 default: break; 2591 } 2592 2593 return SDValue(); 2594} 2595 2596SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 2597 SDValue N1, SDValue N2) { 2598 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 2599 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode()); 2600 switch (Opcode) { 2601 default: break; 2602 case ISD::TokenFactor: 2603 assert(VT == MVT::Other && N1.getValueType() == MVT::Other && 2604 N2.getValueType() == MVT::Other && "Invalid token factor!"); 2605 // Fold trivial token factors. 2606 if (N1.getOpcode() == ISD::EntryToken) return N2; 2607 if (N2.getOpcode() == ISD::EntryToken) return N1; 2608 if (N1 == N2) return N1; 2609 break; 2610 case ISD::CONCAT_VECTORS: 2611 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to 2612 // one big BUILD_VECTOR. 2613 if (N1.getOpcode() == ISD::BUILD_VECTOR && 2614 N2.getOpcode() == ISD::BUILD_VECTOR) { 2615 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end()); 2616 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end()); 2617 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size()); 2618 } 2619 break; 2620 case ISD::AND: 2621 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() && 2622 N1.getValueType() == VT && "Binary operator types must match!"); 2623 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's 2624 // worth handling here. 2625 if (N2C && N2C->isNullValue()) 2626 return N2; 2627 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X 2628 return N1; 2629 break; 2630 case ISD::OR: 2631 case ISD::XOR: 2632 case ISD::ADD: 2633 case ISD::SUB: 2634 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() && 2635 N1.getValueType() == VT && "Binary operator types must match!"); 2636 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so 2637 // it's worth handling here. 2638 if (N2C && N2C->isNullValue()) 2639 return N1; 2640 break; 2641 case ISD::UDIV: 2642 case ISD::UREM: 2643 case ISD::MULHU: 2644 case ISD::MULHS: 2645 case ISD::MUL: 2646 case ISD::SDIV: 2647 case ISD::SREM: 2648 assert(VT.isInteger() && "This operator does not apply to FP types!"); 2649 // fall through 2650 case ISD::FADD: 2651 case ISD::FSUB: 2652 case ISD::FMUL: 2653 case ISD::FDIV: 2654 case ISD::FREM: 2655 if (UnsafeFPMath) { 2656 if (Opcode == ISD::FADD) { 2657 // 0+x --> x 2658 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) 2659 if (CFP->getValueAPF().isZero()) 2660 return N2; 2661 // x+0 --> x 2662 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2)) 2663 if (CFP->getValueAPF().isZero()) 2664 return N1; 2665 } else if (Opcode == ISD::FSUB) { 2666 // x-0 --> x 2667 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2)) 2668 if (CFP->getValueAPF().isZero()) 2669 return N1; 2670 } 2671 } 2672 assert(N1.getValueType() == N2.getValueType() && 2673 N1.getValueType() == VT && "Binary operator types must match!"); 2674 break; 2675 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match. 2676 assert(N1.getValueType() == VT && 2677 N1.getValueType().isFloatingPoint() && 2678 N2.getValueType().isFloatingPoint() && 2679 "Invalid FCOPYSIGN!"); 2680 break; 2681 case ISD::SHL: 2682 case ISD::SRA: 2683 case ISD::SRL: 2684 case ISD::ROTL: 2685 case ISD::ROTR: 2686 assert(VT == N1.getValueType() && 2687 "Shift operators return type must be the same as their first arg"); 2688 assert(VT.isInteger() && N2.getValueType().isInteger() && 2689 "Shifts only work on integers"); 2690 2691 // Always fold shifts of i1 values so the code generator doesn't need to 2692 // handle them. Since we know the size of the shift has to be less than the 2693 // size of the value, the shift/rotate count is guaranteed to be zero. 2694 if (VT == MVT::i1) 2695 return N1; 2696 if (N2C && N2C->isNullValue()) 2697 return N1; 2698 break; 2699 case ISD::FP_ROUND_INREG: { 2700 EVT EVT = cast<VTSDNode>(N2)->getVT(); 2701 assert(VT == N1.getValueType() && "Not an inreg round!"); 2702 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() && 2703 "Cannot FP_ROUND_INREG integer types"); 2704 assert(EVT.isVector() == VT.isVector() && 2705 "FP_ROUND_INREG type should be vector iff the operand " 2706 "type is vector!"); 2707 assert((!EVT.isVector() || 2708 EVT.getVectorNumElements() == VT.getVectorNumElements()) && 2709 "Vector element counts must match in FP_ROUND_INREG"); 2710 assert(EVT.bitsLE(VT) && "Not rounding down!"); 2711 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding. 2712 break; 2713 } 2714 case ISD::FP_ROUND: 2715 assert(VT.isFloatingPoint() && 2716 N1.getValueType().isFloatingPoint() && 2717 VT.bitsLE(N1.getValueType()) && 2718 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!"); 2719 if (N1.getValueType() == VT) return N1; // noop conversion. 2720 break; 2721 case ISD::AssertSext: 2722 case ISD::AssertZext: { 2723 EVT EVT = cast<VTSDNode>(N2)->getVT(); 2724 assert(VT == N1.getValueType() && "Not an inreg extend!"); 2725 assert(VT.isInteger() && EVT.isInteger() && 2726 "Cannot *_EXTEND_INREG FP types"); 2727 assert(!EVT.isVector() && 2728 "AssertSExt/AssertZExt type should be the vector element type " 2729 "rather than the vector type!"); 2730 assert(EVT.bitsLE(VT) && "Not extending!"); 2731 if (VT == EVT) return N1; // noop assertion. 2732 break; 2733 } 2734 case ISD::SIGN_EXTEND_INREG: { 2735 EVT EVT = cast<VTSDNode>(N2)->getVT(); 2736 assert(VT == N1.getValueType() && "Not an inreg extend!"); 2737 assert(VT.isInteger() && EVT.isInteger() && 2738 "Cannot *_EXTEND_INREG FP types"); 2739 assert(EVT.isVector() == VT.isVector() && 2740 "SIGN_EXTEND_INREG type should be vector iff the operand " 2741 "type is vector!"); 2742 assert((!EVT.isVector() || 2743 EVT.getVectorNumElements() == VT.getVectorNumElements()) && 2744 "Vector element counts must match in SIGN_EXTEND_INREG"); 2745 assert(EVT.bitsLE(VT) && "Not extending!"); 2746 if (EVT == VT) return N1; // Not actually extending 2747 2748 if (N1C) { 2749 APInt Val = N1C->getAPIntValue(); 2750 unsigned FromBits = EVT.getScalarType().getSizeInBits(); 2751 Val <<= Val.getBitWidth()-FromBits; 2752 Val = Val.ashr(Val.getBitWidth()-FromBits); 2753 return getConstant(Val, VT); 2754 } 2755 break; 2756 } 2757 case ISD::EXTRACT_VECTOR_ELT: 2758 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF. 2759 if (N1.getOpcode() == ISD::UNDEF) 2760 return getUNDEF(VT); 2761 2762 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is 2763 // expanding copies of large vectors from registers. 2764 if (N2C && 2765 N1.getOpcode() == ISD::CONCAT_VECTORS && 2766 N1.getNumOperands() > 0) { 2767 unsigned Factor = 2768 N1.getOperand(0).getValueType().getVectorNumElements(); 2769 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, 2770 N1.getOperand(N2C->getZExtValue() / Factor), 2771 getConstant(N2C->getZExtValue() % Factor, 2772 N2.getValueType())); 2773 } 2774 2775 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is 2776 // expanding large vector constants. 2777 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) { 2778 SDValue Elt = N1.getOperand(N2C->getZExtValue()); 2779 EVT VEltTy = N1.getValueType().getVectorElementType(); 2780 if (Elt.getValueType() != VEltTy) { 2781 // If the vector element type is not legal, the BUILD_VECTOR operands 2782 // are promoted and implicitly truncated. Make that explicit here. 2783 Elt = getNode(ISD::TRUNCATE, DL, VEltTy, Elt); 2784 } 2785 if (VT != VEltTy) { 2786 // If the vector element type is not legal, the EXTRACT_VECTOR_ELT 2787 // result is implicitly extended. 2788 Elt = getNode(ISD::ANY_EXTEND, DL, VT, Elt); 2789 } 2790 return Elt; 2791 } 2792 2793 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector 2794 // operations are lowered to scalars. 2795 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) { 2796 // If the indices are the same, return the inserted element else 2797 // if the indices are known different, extract the element from 2798 // the original vector. 2799 if (N1.getOperand(2) == N2) { 2800 if (VT == N1.getOperand(1).getValueType()) 2801 return N1.getOperand(1); 2802 else 2803 return getSExtOrTrunc(N1.getOperand(1), DL, VT); 2804 } else if (isa<ConstantSDNode>(N1.getOperand(2)) && 2805 isa<ConstantSDNode>(N2)) 2806 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2); 2807 } 2808 break; 2809 case ISD::EXTRACT_ELEMENT: 2810 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!"); 2811 assert(!N1.getValueType().isVector() && !VT.isVector() && 2812 (N1.getValueType().isInteger() == VT.isInteger()) && 2813 "Wrong types for EXTRACT_ELEMENT!"); 2814 2815 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding 2816 // 64-bit integers into 32-bit parts. Instead of building the extract of 2817 // the BUILD_PAIR, only to have legalize rip it apart, just do it now. 2818 if (N1.getOpcode() == ISD::BUILD_PAIR) 2819 return N1.getOperand(N2C->getZExtValue()); 2820 2821 // EXTRACT_ELEMENT of a constant int is also very common. 2822 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { 2823 unsigned ElementSize = VT.getSizeInBits(); 2824 unsigned Shift = ElementSize * N2C->getZExtValue(); 2825 APInt ShiftedVal = C->getAPIntValue().lshr(Shift); 2826 return getConstant(ShiftedVal.trunc(ElementSize), VT); 2827 } 2828 break; 2829 case ISD::EXTRACT_SUBVECTOR: 2830 if (N1.getValueType() == VT) // Trivial extraction. 2831 return N1; 2832 break; 2833 } 2834 2835 if (N1C) { 2836 if (N2C) { 2837 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C); 2838 if (SV.getNode()) return SV; 2839 } else { // Cannonicalize constant to RHS if commutative 2840 if (isCommutativeBinOp(Opcode)) { 2841 std::swap(N1C, N2C); 2842 std::swap(N1, N2); 2843 } 2844 } 2845 } 2846 2847 // Constant fold FP operations. 2848 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode()); 2849 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode()); 2850 if (N1CFP) { 2851 if (!N2CFP && isCommutativeBinOp(Opcode)) { 2852 // Cannonicalize constant to RHS if commutative 2853 std::swap(N1CFP, N2CFP); 2854 std::swap(N1, N2); 2855 } else if (N2CFP && VT != MVT::ppcf128) { 2856 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF(); 2857 APFloat::opStatus s; 2858 switch (Opcode) { 2859 case ISD::FADD: 2860 s = V1.add(V2, APFloat::rmNearestTiesToEven); 2861 if (s != APFloat::opInvalidOp) 2862 return getConstantFP(V1, VT); 2863 break; 2864 case ISD::FSUB: 2865 s = V1.subtract(V2, APFloat::rmNearestTiesToEven); 2866 if (s!=APFloat::opInvalidOp) 2867 return getConstantFP(V1, VT); 2868 break; 2869 case ISD::FMUL: 2870 s = V1.multiply(V2, APFloat::rmNearestTiesToEven); 2871 if (s!=APFloat::opInvalidOp) 2872 return getConstantFP(V1, VT); 2873 break; 2874 case ISD::FDIV: 2875 s = V1.divide(V2, APFloat::rmNearestTiesToEven); 2876 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero) 2877 return getConstantFP(V1, VT); 2878 break; 2879 case ISD::FREM : 2880 s = V1.mod(V2, APFloat::rmNearestTiesToEven); 2881 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero) 2882 return getConstantFP(V1, VT); 2883 break; 2884 case ISD::FCOPYSIGN: 2885 V1.copySign(V2); 2886 return getConstantFP(V1, VT); 2887 default: break; 2888 } 2889 } 2890 } 2891 2892 // Canonicalize an UNDEF to the RHS, even over a constant. 2893 if (N1.getOpcode() == ISD::UNDEF) { 2894 if (isCommutativeBinOp(Opcode)) { 2895 std::swap(N1, N2); 2896 } else { 2897 switch (Opcode) { 2898 case ISD::FP_ROUND_INREG: 2899 case ISD::SIGN_EXTEND_INREG: 2900 case ISD::SUB: 2901 case ISD::FSUB: 2902 case ISD::FDIV: 2903 case ISD::FREM: 2904 case ISD::SRA: 2905 return N1; // fold op(undef, arg2) -> undef 2906 case ISD::UDIV: 2907 case ISD::SDIV: 2908 case ISD::UREM: 2909 case ISD::SREM: 2910 case ISD::SRL: 2911 case ISD::SHL: 2912 if (!VT.isVector()) 2913 return getConstant(0, VT); // fold op(undef, arg2) -> 0 2914 // For vectors, we can't easily build an all zero vector, just return 2915 // the LHS. 2916 return N2; 2917 } 2918 } 2919 } 2920 2921 // Fold a bunch of operators when the RHS is undef. 2922 if (N2.getOpcode() == ISD::UNDEF) { 2923 switch (Opcode) { 2924 case ISD::XOR: 2925 if (N1.getOpcode() == ISD::UNDEF) 2926 // Handle undef ^ undef -> 0 special case. This is a common 2927 // idiom (misuse). 2928 return getConstant(0, VT); 2929 // fallthrough 2930 case ISD::ADD: 2931 case ISD::ADDC: 2932 case ISD::ADDE: 2933 case ISD::SUB: 2934 case ISD::UDIV: 2935 case ISD::SDIV: 2936 case ISD::UREM: 2937 case ISD::SREM: 2938 return N2; // fold op(arg1, undef) -> undef 2939 case ISD::FADD: 2940 case ISD::FSUB: 2941 case ISD::FMUL: 2942 case ISD::FDIV: 2943 case ISD::FREM: 2944 if (UnsafeFPMath) 2945 return N2; 2946 break; 2947 case ISD::MUL: 2948 case ISD::AND: 2949 case ISD::SRL: 2950 case ISD::SHL: 2951 if (!VT.isVector()) 2952 return getConstant(0, VT); // fold op(arg1, undef) -> 0 2953 // For vectors, we can't easily build an all zero vector, just return 2954 // the LHS. 2955 return N1; 2956 case ISD::OR: 2957 if (!VT.isVector()) 2958 return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT); 2959 // For vectors, we can't easily build an all one vector, just return 2960 // the LHS. 2961 return N1; 2962 case ISD::SRA: 2963 return N1; 2964 } 2965 } 2966 2967 // Memoize this node if possible. 2968 SDNode *N; 2969 SDVTList VTs = getVTList(VT); 2970 if (VT != MVT::Flag) { 2971 SDValue Ops[] = { N1, N2 }; 2972 FoldingSetNodeID ID; 2973 AddNodeIDNode(ID, Opcode, VTs, Ops, 2); 2974 void *IP = 0; 2975 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 2976 return SDValue(E, 0); 2977 2978 N = NodeAllocator.Allocate<BinarySDNode>(); 2979 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2); 2980 CSEMap.InsertNode(N, IP); 2981 } else { 2982 N = NodeAllocator.Allocate<BinarySDNode>(); 2983 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2); 2984 } 2985 2986 AllNodes.push_back(N); 2987#ifndef NDEBUG 2988 VerifyNode(N); 2989#endif 2990 return SDValue(N, 0); 2991} 2992 2993SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 2994 SDValue N1, SDValue N2, SDValue N3) { 2995 // Perform various simplifications. 2996 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 2997 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode()); 2998 switch (Opcode) { 2999 case ISD::CONCAT_VECTORS: 3000 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to 3001 // one big BUILD_VECTOR. 3002 if (N1.getOpcode() == ISD::BUILD_VECTOR && 3003 N2.getOpcode() == ISD::BUILD_VECTOR && 3004 N3.getOpcode() == ISD::BUILD_VECTOR) { 3005 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end()); 3006 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end()); 3007 Elts.insert(Elts.end(), N3.getNode()->op_begin(), N3.getNode()->op_end()); 3008 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size()); 3009 } 3010 break; 3011 case ISD::SETCC: { 3012 // Use FoldSetCC to simplify SETCC's. 3013 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL); 3014 if (Simp.getNode()) return Simp; 3015 break; 3016 } 3017 case ISD::SELECT: 3018 if (N1C) { 3019 if (N1C->getZExtValue()) 3020 return N2; // select true, X, Y -> X 3021 else 3022 return N3; // select false, X, Y -> Y 3023 } 3024 3025 if (N2 == N3) return N2; // select C, X, X -> X 3026 break; 3027 case ISD::BRCOND: 3028 if (N2C) { 3029 if (N2C->getZExtValue()) // Unconditional branch 3030 return getNode(ISD::BR, DL, MVT::Other, N1, N3); 3031 else 3032 return N1; // Never-taken branch 3033 } 3034 break; 3035 case ISD::VECTOR_SHUFFLE: 3036 llvm_unreachable("should use getVectorShuffle constructor!"); 3037 break; 3038 case ISD::BIT_CONVERT: 3039 // Fold bit_convert nodes from a type to themselves. 3040 if (N1.getValueType() == VT) 3041 return N1; 3042 break; 3043 } 3044 3045 // Memoize node if it doesn't produce a flag. 3046 SDNode *N; 3047 SDVTList VTs = getVTList(VT); 3048 if (VT != MVT::Flag) { 3049 SDValue Ops[] = { N1, N2, N3 }; 3050 FoldingSetNodeID ID; 3051 AddNodeIDNode(ID, Opcode, VTs, Ops, 3); 3052 void *IP = 0; 3053 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 3054 return SDValue(E, 0); 3055 3056 N = NodeAllocator.Allocate<TernarySDNode>(); 3057 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3); 3058 CSEMap.InsertNode(N, IP); 3059 } else { 3060 N = NodeAllocator.Allocate<TernarySDNode>(); 3061 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3); 3062 } 3063 3064 AllNodes.push_back(N); 3065#ifndef NDEBUG 3066 VerifyNode(N); 3067#endif 3068 return SDValue(N, 0); 3069} 3070 3071SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 3072 SDValue N1, SDValue N2, SDValue N3, 3073 SDValue N4) { 3074 SDValue Ops[] = { N1, N2, N3, N4 }; 3075 return getNode(Opcode, DL, VT, Ops, 4); 3076} 3077 3078SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 3079 SDValue N1, SDValue N2, SDValue N3, 3080 SDValue N4, SDValue N5) { 3081 SDValue Ops[] = { N1, N2, N3, N4, N5 }; 3082 return getNode(Opcode, DL, VT, Ops, 5); 3083} 3084 3085/// getStackArgumentTokenFactor - Compute a TokenFactor to force all 3086/// the incoming stack arguments to be loaded from the stack. 3087SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) { 3088 SmallVector<SDValue, 8> ArgChains; 3089 3090 // Include the original chain at the beginning of the list. When this is 3091 // used by target LowerCall hooks, this helps legalize find the 3092 // CALLSEQ_BEGIN node. 3093 ArgChains.push_back(Chain); 3094 3095 // Add a chain value for each stack argument. 3096 for (SDNode::use_iterator U = getEntryNode().getNode()->use_begin(), 3097 UE = getEntryNode().getNode()->use_end(); U != UE; ++U) 3098 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U)) 3099 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) 3100 if (FI->getIndex() < 0) 3101 ArgChains.push_back(SDValue(L, 1)); 3102 3103 // Build a tokenfactor for all the chains. 3104 return getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other, 3105 &ArgChains[0], ArgChains.size()); 3106} 3107 3108/// getMemsetValue - Vectorized representation of the memset value 3109/// operand. 3110static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG, 3111 DebugLoc dl) { 3112 unsigned NumBits = VT.getScalarType().getSizeInBits(); 3113 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) { 3114 APInt Val = APInt(NumBits, C->getZExtValue() & 255); 3115 unsigned Shift = 8; 3116 for (unsigned i = NumBits; i > 8; i >>= 1) { 3117 Val = (Val << Shift) | Val; 3118 Shift <<= 1; 3119 } 3120 if (VT.isInteger()) 3121 return DAG.getConstant(Val, VT); 3122 return DAG.getConstantFP(APFloat(Val), VT); 3123 } 3124 3125 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3126 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value); 3127 unsigned Shift = 8; 3128 for (unsigned i = NumBits; i > 8; i >>= 1) { 3129 Value = DAG.getNode(ISD::OR, dl, VT, 3130 DAG.getNode(ISD::SHL, dl, VT, Value, 3131 DAG.getConstant(Shift, 3132 TLI.getShiftAmountTy())), 3133 Value); 3134 Shift <<= 1; 3135 } 3136 3137 return Value; 3138} 3139 3140/// getMemsetStringVal - Similar to getMemsetValue. Except this is only 3141/// used when a memcpy is turned into a memset when the source is a constant 3142/// string ptr. 3143static SDValue getMemsetStringVal(EVT VT, DebugLoc dl, SelectionDAG &DAG, 3144 const TargetLowering &TLI, 3145 std::string &Str, unsigned Offset) { 3146 // Handle vector with all elements zero. 3147 if (Str.empty()) { 3148 if (VT.isInteger()) 3149 return DAG.getConstant(0, VT); 3150 unsigned NumElts = VT.getVectorNumElements(); 3151 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64; 3152 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, 3153 DAG.getConstant(0, 3154 EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts))); 3155 } 3156 3157 assert(!VT.isVector() && "Can't handle vector type here!"); 3158 unsigned NumBits = VT.getSizeInBits(); 3159 unsigned MSB = NumBits / 8; 3160 uint64_t Val = 0; 3161 if (TLI.isLittleEndian()) 3162 Offset = Offset + MSB - 1; 3163 for (unsigned i = 0; i != MSB; ++i) { 3164 Val = (Val << 8) | (unsigned char)Str[Offset]; 3165 Offset += TLI.isLittleEndian() ? -1 : 1; 3166 } 3167 return DAG.getConstant(Val, VT); 3168} 3169 3170/// getMemBasePlusOffset - Returns base and offset node for the 3171/// 3172static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset, 3173 SelectionDAG &DAG) { 3174 EVT VT = Base.getValueType(); 3175 return DAG.getNode(ISD::ADD, Base.getDebugLoc(), 3176 VT, Base, DAG.getConstant(Offset, VT)); 3177} 3178 3179/// isMemSrcFromString - Returns true if memcpy source is a string constant. 3180/// 3181static bool isMemSrcFromString(SDValue Src, std::string &Str) { 3182 unsigned SrcDelta = 0; 3183 GlobalAddressSDNode *G = NULL; 3184 if (Src.getOpcode() == ISD::GlobalAddress) 3185 G = cast<GlobalAddressSDNode>(Src); 3186 else if (Src.getOpcode() == ISD::ADD && 3187 Src.getOperand(0).getOpcode() == ISD::GlobalAddress && 3188 Src.getOperand(1).getOpcode() == ISD::Constant) { 3189 G = cast<GlobalAddressSDNode>(Src.getOperand(0)); 3190 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue(); 3191 } 3192 if (!G) 3193 return false; 3194 3195 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal()); 3196 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false)) 3197 return true; 3198 3199 return false; 3200} 3201 3202/// MeetsMaxMemopRequirement - Determines if the number of memory ops required 3203/// to replace the memset / memcpy is below the threshold. It also returns the 3204/// types of the sequence of memory ops to perform memset / memcpy. 3205static 3206bool MeetsMaxMemopRequirement(std::vector<EVT> &MemOps, 3207 SDValue Dst, SDValue Src, 3208 unsigned Limit, uint64_t Size, unsigned &Align, 3209 std::string &Str, bool &isSrcStr, 3210 SelectionDAG &DAG, 3211 const TargetLowering &TLI) { 3212 isSrcStr = isMemSrcFromString(Src, Str); 3213 bool isSrcConst = isa<ConstantSDNode>(Src); 3214 EVT VT = TLI.getOptimalMemOpType(Size, Align, isSrcConst, isSrcStr, DAG); 3215 bool AllowUnalign = TLI.allowsUnalignedMemoryAccesses(VT); 3216 if (VT != MVT::Other) { 3217 const Type *Ty = VT.getTypeForEVT(*DAG.getContext()); 3218 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty); 3219 // If source is a string constant, this will require an unaligned load. 3220 if (NewAlign > Align && (isSrcConst || AllowUnalign)) { 3221 if (Dst.getOpcode() != ISD::FrameIndex) { 3222 // Can't change destination alignment. It requires a unaligned store. 3223 if (AllowUnalign) 3224 VT = MVT::Other; 3225 } else { 3226 int FI = cast<FrameIndexSDNode>(Dst)->getIndex(); 3227 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 3228 if (MFI->isFixedObjectIndex(FI)) { 3229 // Can't change destination alignment. It requires a unaligned store. 3230 if (AllowUnalign) 3231 VT = MVT::Other; 3232 } else { 3233 // Give the stack frame object a larger alignment if needed. 3234 if (MFI->getObjectAlignment(FI) < NewAlign) 3235 MFI->setObjectAlignment(FI, NewAlign); 3236 Align = NewAlign; 3237 } 3238 } 3239 } 3240 } 3241 3242 if (VT == MVT::Other) { 3243 if (TLI.allowsUnalignedMemoryAccesses(MVT::i64)) { 3244 VT = MVT::i64; 3245 } else { 3246 switch (Align & 7) { 3247 case 0: VT = MVT::i64; break; 3248 case 4: VT = MVT::i32; break; 3249 case 2: VT = MVT::i16; break; 3250 default: VT = MVT::i8; break; 3251 } 3252 } 3253 3254 MVT LVT = MVT::i64; 3255 while (!TLI.isTypeLegal(LVT)) 3256 LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1); 3257 assert(LVT.isInteger()); 3258 3259 if (VT.bitsGT(LVT)) 3260 VT = LVT; 3261 } 3262 3263 unsigned NumMemOps = 0; 3264 while (Size != 0) { 3265 unsigned VTSize = VT.getSizeInBits() / 8; 3266 while (VTSize > Size) { 3267 // For now, only use non-vector load / store's for the left-over pieces. 3268 if (VT.isVector()) { 3269 VT = MVT::i64; 3270 while (!TLI.isTypeLegal(VT)) 3271 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1); 3272 VTSize = VT.getSizeInBits() / 8; 3273 } else { 3274 // This can result in a type that is not legal on the target, e.g. 3275 // 1 or 2 bytes on PPC. 3276 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1); 3277 VTSize >>= 1; 3278 } 3279 } 3280 3281 if (++NumMemOps > Limit) 3282 return false; 3283 MemOps.push_back(VT); 3284 Size -= VTSize; 3285 } 3286 3287 return true; 3288} 3289 3290static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, 3291 SDValue Chain, SDValue Dst, 3292 SDValue Src, uint64_t Size, 3293 unsigned Align, bool AlwaysInline, 3294 const Value *DstSV, uint64_t DstSVOff, 3295 const Value *SrcSV, uint64_t SrcSVOff){ 3296 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3297 3298 // Expand memcpy to a series of load and store ops if the size operand falls 3299 // below a certain threshold. 3300 std::vector<EVT> MemOps; 3301 uint64_t Limit = -1ULL; 3302 if (!AlwaysInline) 3303 Limit = TLI.getMaxStoresPerMemcpy(); 3304 unsigned DstAlign = Align; // Destination alignment can change. 3305 std::string Str; 3306 bool CopyFromStr; 3307 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign, 3308 Str, CopyFromStr, DAG, TLI)) 3309 return SDValue(); 3310 3311 3312 bool isZeroStr = CopyFromStr && Str.empty(); 3313 SmallVector<SDValue, 8> OutChains; 3314 unsigned NumMemOps = MemOps.size(); 3315 uint64_t SrcOff = 0, DstOff = 0; 3316 for (unsigned i = 0; i != NumMemOps; ++i) { 3317 EVT VT = MemOps[i]; 3318 unsigned VTSize = VT.getSizeInBits() / 8; 3319 SDValue Value, Store; 3320 3321 if (CopyFromStr && (isZeroStr || !VT.isVector())) { 3322 // It's unlikely a store of a vector immediate can be done in a single 3323 // instruction. It would require a load from a constantpool first. 3324 // We also handle store a vector with all zero's. 3325 // FIXME: Handle other cases where store of vector immediate is done in 3326 // a single instruction. 3327 Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff); 3328 Store = DAG.getStore(Chain, dl, Value, 3329 getMemBasePlusOffset(Dst, DstOff, DAG), 3330 DstSV, DstSVOff + DstOff, false, false, DstAlign); 3331 } else { 3332 // The type might not be legal for the target. This should only happen 3333 // if the type is smaller than a legal type, as on PPC, so the right 3334 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify 3335 // to Load/Store if NVT==VT. 3336 // FIXME does the case above also need this? 3337 EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); 3338 assert(NVT.bitsGE(VT)); 3339 Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain, 3340 getMemBasePlusOffset(Src, SrcOff, DAG), 3341 SrcSV, SrcSVOff + SrcOff, VT, false, false, Align); 3342 Store = DAG.getTruncStore(Chain, dl, Value, 3343 getMemBasePlusOffset(Dst, DstOff, DAG), 3344 DstSV, DstSVOff + DstOff, VT, false, false, 3345 DstAlign); 3346 } 3347 OutChains.push_back(Store); 3348 SrcOff += VTSize; 3349 DstOff += VTSize; 3350 } 3351 3352 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3353 &OutChains[0], OutChains.size()); 3354} 3355 3356static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, 3357 SDValue Chain, SDValue Dst, 3358 SDValue Src, uint64_t Size, 3359 unsigned Align, bool AlwaysInline, 3360 const Value *DstSV, uint64_t DstSVOff, 3361 const Value *SrcSV, uint64_t SrcSVOff){ 3362 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3363 3364 // Expand memmove to a series of load and store ops if the size operand falls 3365 // below a certain threshold. 3366 std::vector<EVT> MemOps; 3367 uint64_t Limit = -1ULL; 3368 if (!AlwaysInline) 3369 Limit = TLI.getMaxStoresPerMemmove(); 3370 unsigned DstAlign = Align; // Destination alignment can change. 3371 std::string Str; 3372 bool CopyFromStr; 3373 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign, 3374 Str, CopyFromStr, DAG, TLI)) 3375 return SDValue(); 3376 3377 uint64_t SrcOff = 0, DstOff = 0; 3378 3379 SmallVector<SDValue, 8> LoadValues; 3380 SmallVector<SDValue, 8> LoadChains; 3381 SmallVector<SDValue, 8> OutChains; 3382 unsigned NumMemOps = MemOps.size(); 3383 for (unsigned i = 0; i < NumMemOps; i++) { 3384 EVT VT = MemOps[i]; 3385 unsigned VTSize = VT.getSizeInBits() / 8; 3386 SDValue Value, Store; 3387 3388 Value = DAG.getLoad(VT, dl, Chain, 3389 getMemBasePlusOffset(Src, SrcOff, DAG), 3390 SrcSV, SrcSVOff + SrcOff, false, false, Align); 3391 LoadValues.push_back(Value); 3392 LoadChains.push_back(Value.getValue(1)); 3393 SrcOff += VTSize; 3394 } 3395 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3396 &LoadChains[0], LoadChains.size()); 3397 OutChains.clear(); 3398 for (unsigned i = 0; i < NumMemOps; i++) { 3399 EVT VT = MemOps[i]; 3400 unsigned VTSize = VT.getSizeInBits() / 8; 3401 SDValue Value, Store; 3402 3403 Store = DAG.getStore(Chain, dl, LoadValues[i], 3404 getMemBasePlusOffset(Dst, DstOff, DAG), 3405 DstSV, DstSVOff + DstOff, false, false, DstAlign); 3406 OutChains.push_back(Store); 3407 DstOff += VTSize; 3408 } 3409 3410 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3411 &OutChains[0], OutChains.size()); 3412} 3413 3414static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl, 3415 SDValue Chain, SDValue Dst, 3416 SDValue Src, uint64_t Size, 3417 unsigned Align, 3418 const Value *DstSV, uint64_t DstSVOff) { 3419 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3420 3421 // Expand memset to a series of load/store ops if the size operand 3422 // falls below a certain threshold. 3423 std::vector<EVT> MemOps; 3424 std::string Str; 3425 bool CopyFromStr; 3426 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, TLI.getMaxStoresPerMemset(), 3427 Size, Align, Str, CopyFromStr, DAG, TLI)) 3428 return SDValue(); 3429 3430 SmallVector<SDValue, 8> OutChains; 3431 uint64_t DstOff = 0; 3432 3433 unsigned NumMemOps = MemOps.size(); 3434 for (unsigned i = 0; i < NumMemOps; i++) { 3435 EVT VT = MemOps[i]; 3436 unsigned VTSize = VT.getSizeInBits() / 8; 3437 SDValue Value = getMemsetValue(Src, VT, DAG, dl); 3438 SDValue Store = DAG.getStore(Chain, dl, Value, 3439 getMemBasePlusOffset(Dst, DstOff, DAG), 3440 DstSV, DstSVOff + DstOff, false, false, 0); 3441 OutChains.push_back(Store); 3442 DstOff += VTSize; 3443 } 3444 3445 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3446 &OutChains[0], OutChains.size()); 3447} 3448 3449SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst, 3450 SDValue Src, SDValue Size, 3451 unsigned Align, bool AlwaysInline, 3452 const Value *DstSV, uint64_t DstSVOff, 3453 const Value *SrcSV, uint64_t SrcSVOff) { 3454 3455 // Check to see if we should lower the memcpy to loads and stores first. 3456 // For cases within the target-specified limits, this is the best choice. 3457 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 3458 if (ConstantSize) { 3459 // Memcpy with size zero? Just return the original chain. 3460 if (ConstantSize->isNullValue()) 3461 return Chain; 3462 3463 SDValue Result = 3464 getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src, 3465 ConstantSize->getZExtValue(), 3466 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff); 3467 if (Result.getNode()) 3468 return Result; 3469 } 3470 3471 // Then check to see if we should lower the memcpy with target-specific 3472 // code. If the target chooses to do this, this is the next best. 3473 SDValue Result = 3474 TLI.EmitTargetCodeForMemcpy(*this, dl, Chain, Dst, Src, Size, Align, 3475 AlwaysInline, 3476 DstSV, DstSVOff, SrcSV, SrcSVOff); 3477 if (Result.getNode()) 3478 return Result; 3479 3480 // If we really need inline code and the target declined to provide it, 3481 // use a (potentially long) sequence of loads and stores. 3482 if (AlwaysInline) { 3483 assert(ConstantSize && "AlwaysInline requires a constant size!"); 3484 return getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src, 3485 ConstantSize->getZExtValue(), Align, true, 3486 DstSV, DstSVOff, SrcSV, SrcSVOff); 3487 } 3488 3489 // Emit a library call. 3490 TargetLowering::ArgListTy Args; 3491 TargetLowering::ArgListEntry Entry; 3492 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext()); 3493 Entry.Node = Dst; Args.push_back(Entry); 3494 Entry.Node = Src; Args.push_back(Entry); 3495 Entry.Node = Size; Args.push_back(Entry); 3496 // FIXME: pass in DebugLoc 3497 std::pair<SDValue,SDValue> CallResult = 3498 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()), 3499 false, false, false, false, 0, 3500 TLI.getLibcallCallingConv(RTLIB::MEMCPY), false, 3501 /*isReturnValueUsed=*/false, 3502 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMCPY), 3503 TLI.getPointerTy()), 3504 Args, *this, dl); 3505 return CallResult.second; 3506} 3507 3508SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst, 3509 SDValue Src, SDValue Size, 3510 unsigned Align, 3511 const Value *DstSV, uint64_t DstSVOff, 3512 const Value *SrcSV, uint64_t SrcSVOff) { 3513 3514 // Check to see if we should lower the memmove to loads and stores first. 3515 // For cases within the target-specified limits, this is the best choice. 3516 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 3517 if (ConstantSize) { 3518 // Memmove with size zero? Just return the original chain. 3519 if (ConstantSize->isNullValue()) 3520 return Chain; 3521 3522 SDValue Result = 3523 getMemmoveLoadsAndStores(*this, dl, Chain, Dst, Src, 3524 ConstantSize->getZExtValue(), 3525 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff); 3526 if (Result.getNode()) 3527 return Result; 3528 } 3529 3530 // Then check to see if we should lower the memmove with target-specific 3531 // code. If the target chooses to do this, this is the next best. 3532 SDValue Result = 3533 TLI.EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size, Align, 3534 DstSV, DstSVOff, SrcSV, SrcSVOff); 3535 if (Result.getNode()) 3536 return Result; 3537 3538 // Emit a library call. 3539 TargetLowering::ArgListTy Args; 3540 TargetLowering::ArgListEntry Entry; 3541 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext()); 3542 Entry.Node = Dst; Args.push_back(Entry); 3543 Entry.Node = Src; Args.push_back(Entry); 3544 Entry.Node = Size; Args.push_back(Entry); 3545 // FIXME: pass in DebugLoc 3546 std::pair<SDValue,SDValue> CallResult = 3547 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()), 3548 false, false, false, false, 0, 3549 TLI.getLibcallCallingConv(RTLIB::MEMMOVE), false, 3550 /*isReturnValueUsed=*/false, 3551 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMMOVE), 3552 TLI.getPointerTy()), 3553 Args, *this, dl); 3554 return CallResult.second; 3555} 3556 3557SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst, 3558 SDValue Src, SDValue Size, 3559 unsigned Align, 3560 const Value *DstSV, uint64_t DstSVOff) { 3561 3562 // Check to see if we should lower the memset to stores first. 3563 // For cases within the target-specified limits, this is the best choice. 3564 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 3565 if (ConstantSize) { 3566 // Memset with size zero? Just return the original chain. 3567 if (ConstantSize->isNullValue()) 3568 return Chain; 3569 3570 SDValue Result = 3571 getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), 3572 Align, DstSV, DstSVOff); 3573 if (Result.getNode()) 3574 return Result; 3575 } 3576 3577 // Then check to see if we should lower the memset with target-specific 3578 // code. If the target chooses to do this, this is the next best. 3579 SDValue Result = 3580 TLI.EmitTargetCodeForMemset(*this, dl, Chain, Dst, Src, Size, Align, 3581 DstSV, DstSVOff); 3582 if (Result.getNode()) 3583 return Result; 3584 3585 // Emit a library call. 3586 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType(*getContext()); 3587 TargetLowering::ArgListTy Args; 3588 TargetLowering::ArgListEntry Entry; 3589 Entry.Node = Dst; Entry.Ty = IntPtrTy; 3590 Args.push_back(Entry); 3591 // Extend or truncate the argument to be an i32 value for the call. 3592 if (Src.getValueType().bitsGT(MVT::i32)) 3593 Src = getNode(ISD::TRUNCATE, dl, MVT::i32, Src); 3594 else 3595 Src = getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src); 3596 Entry.Node = Src; 3597 Entry.Ty = Type::getInt32Ty(*getContext()); 3598 Entry.isSExt = true; 3599 Args.push_back(Entry); 3600 Entry.Node = Size; 3601 Entry.Ty = IntPtrTy; 3602 Entry.isSExt = false; 3603 Args.push_back(Entry); 3604 // FIXME: pass in DebugLoc 3605 std::pair<SDValue,SDValue> CallResult = 3606 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()), 3607 false, false, false, false, 0, 3608 TLI.getLibcallCallingConv(RTLIB::MEMSET), false, 3609 /*isReturnValueUsed=*/false, 3610 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET), 3611 TLI.getPointerTy()), 3612 Args, *this, dl); 3613 return CallResult.second; 3614} 3615 3616SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3617 SDValue Chain, 3618 SDValue Ptr, SDValue Cmp, 3619 SDValue Swp, const Value* PtrVal, 3620 unsigned Alignment) { 3621 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3622 Alignment = getEVTAlignment(MemVT); 3623 3624 // Check if the memory reference references a frame index 3625 if (!PtrVal) 3626 if (const FrameIndexSDNode *FI = 3627 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3628 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex()); 3629 3630 MachineFunction &MF = getMachineFunction(); 3631 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore; 3632 3633 // For now, atomics are considered to be volatile always. 3634 Flags |= MachineMemOperand::MOVolatile; 3635 3636 MachineMemOperand *MMO = 3637 MF.getMachineMemOperand(PtrVal, Flags, 0, 3638 MemVT.getStoreSize(), Alignment); 3639 3640 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Cmp, Swp, MMO); 3641} 3642 3643SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3644 SDValue Chain, 3645 SDValue Ptr, SDValue Cmp, 3646 SDValue Swp, MachineMemOperand *MMO) { 3647 assert(Opcode == ISD::ATOMIC_CMP_SWAP && "Invalid Atomic Op"); 3648 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types"); 3649 3650 EVT VT = Cmp.getValueType(); 3651 3652 SDVTList VTs = getVTList(VT, MVT::Other); 3653 FoldingSetNodeID ID; 3654 ID.AddInteger(MemVT.getRawBits()); 3655 SDValue Ops[] = {Chain, Ptr, Cmp, Swp}; 3656 AddNodeIDNode(ID, Opcode, VTs, Ops, 4); 3657 void* IP = 0; 3658 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3659 cast<AtomicSDNode>(E)->refineAlignment(MMO); 3660 return SDValue(E, 0); 3661 } 3662 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>(); 3663 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Cmp, Swp, MMO); 3664 CSEMap.InsertNode(N, IP); 3665 AllNodes.push_back(N); 3666 return SDValue(N, 0); 3667} 3668 3669SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3670 SDValue Chain, 3671 SDValue Ptr, SDValue Val, 3672 const Value* PtrVal, 3673 unsigned Alignment) { 3674 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3675 Alignment = getEVTAlignment(MemVT); 3676 3677 // Check if the memory reference references a frame index 3678 if (!PtrVal) 3679 if (const FrameIndexSDNode *FI = 3680 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3681 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex()); 3682 3683 MachineFunction &MF = getMachineFunction(); 3684 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore; 3685 3686 // For now, atomics are considered to be volatile always. 3687 Flags |= MachineMemOperand::MOVolatile; 3688 3689 MachineMemOperand *MMO = 3690 MF.getMachineMemOperand(PtrVal, Flags, 0, 3691 MemVT.getStoreSize(), Alignment); 3692 3693 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Val, MMO); 3694} 3695 3696SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3697 SDValue Chain, 3698 SDValue Ptr, SDValue Val, 3699 MachineMemOperand *MMO) { 3700 assert((Opcode == ISD::ATOMIC_LOAD_ADD || 3701 Opcode == ISD::ATOMIC_LOAD_SUB || 3702 Opcode == ISD::ATOMIC_LOAD_AND || 3703 Opcode == ISD::ATOMIC_LOAD_OR || 3704 Opcode == ISD::ATOMIC_LOAD_XOR || 3705 Opcode == ISD::ATOMIC_LOAD_NAND || 3706 Opcode == ISD::ATOMIC_LOAD_MIN || 3707 Opcode == ISD::ATOMIC_LOAD_MAX || 3708 Opcode == ISD::ATOMIC_LOAD_UMIN || 3709 Opcode == ISD::ATOMIC_LOAD_UMAX || 3710 Opcode == ISD::ATOMIC_SWAP) && 3711 "Invalid Atomic Op"); 3712 3713 EVT VT = Val.getValueType(); 3714 3715 SDVTList VTs = getVTList(VT, MVT::Other); 3716 FoldingSetNodeID ID; 3717 ID.AddInteger(MemVT.getRawBits()); 3718 SDValue Ops[] = {Chain, Ptr, Val}; 3719 AddNodeIDNode(ID, Opcode, VTs, Ops, 3); 3720 void* IP = 0; 3721 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3722 cast<AtomicSDNode>(E)->refineAlignment(MMO); 3723 return SDValue(E, 0); 3724 } 3725 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>(); 3726 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Val, MMO); 3727 CSEMap.InsertNode(N, IP); 3728 AllNodes.push_back(N); 3729 return SDValue(N, 0); 3730} 3731 3732/// getMergeValues - Create a MERGE_VALUES node from the given operands. 3733/// Allowed to return something different (and simpler) if Simplify is true. 3734SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps, 3735 DebugLoc dl) { 3736 if (NumOps == 1) 3737 return Ops[0]; 3738 3739 SmallVector<EVT, 4> VTs; 3740 VTs.reserve(NumOps); 3741 for (unsigned i = 0; i < NumOps; ++i) 3742 VTs.push_back(Ops[i].getValueType()); 3743 return getNode(ISD::MERGE_VALUES, dl, getVTList(&VTs[0], NumOps), 3744 Ops, NumOps); 3745} 3746 3747SDValue 3748SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, 3749 const EVT *VTs, unsigned NumVTs, 3750 const SDValue *Ops, unsigned NumOps, 3751 EVT MemVT, const Value *srcValue, int SVOff, 3752 unsigned Align, bool Vol, 3753 bool ReadMem, bool WriteMem) { 3754 return getMemIntrinsicNode(Opcode, dl, makeVTList(VTs, NumVTs), Ops, NumOps, 3755 MemVT, srcValue, SVOff, Align, Vol, 3756 ReadMem, WriteMem); 3757} 3758 3759SDValue 3760SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList, 3761 const SDValue *Ops, unsigned NumOps, 3762 EVT MemVT, const Value *srcValue, int SVOff, 3763 unsigned Align, bool Vol, 3764 bool ReadMem, bool WriteMem) { 3765 if (Align == 0) // Ensure that codegen never sees alignment 0 3766 Align = getEVTAlignment(MemVT); 3767 3768 MachineFunction &MF = getMachineFunction(); 3769 unsigned Flags = 0; 3770 if (WriteMem) 3771 Flags |= MachineMemOperand::MOStore; 3772 if (ReadMem) 3773 Flags |= MachineMemOperand::MOLoad; 3774 if (Vol) 3775 Flags |= MachineMemOperand::MOVolatile; 3776 MachineMemOperand *MMO = 3777 MF.getMachineMemOperand(srcValue, Flags, SVOff, 3778 MemVT.getStoreSize(), Align); 3779 3780 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO); 3781} 3782 3783SDValue 3784SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList, 3785 const SDValue *Ops, unsigned NumOps, 3786 EVT MemVT, MachineMemOperand *MMO) { 3787 assert((Opcode == ISD::INTRINSIC_VOID || 3788 Opcode == ISD::INTRINSIC_W_CHAIN || 3789 (Opcode <= INT_MAX && 3790 (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) && 3791 "Opcode is not a memory-accessing opcode!"); 3792 3793 // Memoize the node unless it returns a flag. 3794 MemIntrinsicSDNode *N; 3795 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) { 3796 FoldingSetNodeID ID; 3797 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps); 3798 void *IP = 0; 3799 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3800 cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO); 3801 return SDValue(E, 0); 3802 } 3803 3804 N = NodeAllocator.Allocate<MemIntrinsicSDNode>(); 3805 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO); 3806 CSEMap.InsertNode(N, IP); 3807 } else { 3808 N = NodeAllocator.Allocate<MemIntrinsicSDNode>(); 3809 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO); 3810 } 3811 AllNodes.push_back(N); 3812 return SDValue(N, 0); 3813} 3814 3815SDValue 3816SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl, 3817 ISD::LoadExtType ExtType, EVT VT, SDValue Chain, 3818 SDValue Ptr, SDValue Offset, 3819 const Value *SV, int SVOffset, EVT MemVT, 3820 bool isVolatile, bool isNonTemporal, 3821 unsigned Alignment) { 3822 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3823 Alignment = getEVTAlignment(VT); 3824 3825 // Check if the memory reference references a frame index 3826 if (!SV) 3827 if (const FrameIndexSDNode *FI = 3828 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3829 SV = PseudoSourceValue::getFixedStack(FI->getIndex()); 3830 3831 MachineFunction &MF = getMachineFunction(); 3832 unsigned Flags = MachineMemOperand::MOLoad; 3833 if (isVolatile) 3834 Flags |= MachineMemOperand::MOVolatile; 3835 if (isNonTemporal) 3836 Flags |= MachineMemOperand::MONonTemporal; 3837 MachineMemOperand *MMO = 3838 MF.getMachineMemOperand(SV, Flags, SVOffset, 3839 MemVT.getStoreSize(), Alignment); 3840 return getLoad(AM, dl, ExtType, VT, Chain, Ptr, Offset, MemVT, MMO); 3841} 3842 3843SDValue 3844SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl, 3845 ISD::LoadExtType ExtType, EVT VT, SDValue Chain, 3846 SDValue Ptr, SDValue Offset, EVT MemVT, 3847 MachineMemOperand *MMO) { 3848 if (VT == MemVT) { 3849 ExtType = ISD::NON_EXTLOAD; 3850 } else if (ExtType == ISD::NON_EXTLOAD) { 3851 assert(VT == MemVT && "Non-extending load from different memory type!"); 3852 } else { 3853 // Extending load. 3854 assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) && 3855 "Should only be an extending load, not truncating!"); 3856 assert(VT.isInteger() == MemVT.isInteger() && 3857 "Cannot convert from FP to Int or Int -> FP!"); 3858 assert(VT.isVector() == MemVT.isVector() && 3859 "Cannot use trunc store to convert to or from a vector!"); 3860 assert((!VT.isVector() || 3861 VT.getVectorNumElements() == MemVT.getVectorNumElements()) && 3862 "Cannot use trunc store to change the number of vector elements!"); 3863 } 3864 3865 bool Indexed = AM != ISD::UNINDEXED; 3866 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) && 3867 "Unindexed load with an offset!"); 3868 3869 SDVTList VTs = Indexed ? 3870 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other); 3871 SDValue Ops[] = { Chain, Ptr, Offset }; 3872 FoldingSetNodeID ID; 3873 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3); 3874 ID.AddInteger(MemVT.getRawBits()); 3875 ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile(), 3876 MMO->isNonTemporal())); 3877 void *IP = 0; 3878 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3879 cast<LoadSDNode>(E)->refineAlignment(MMO); 3880 return SDValue(E, 0); 3881 } 3882 SDNode *N = NodeAllocator.Allocate<LoadSDNode>(); 3883 new (N) LoadSDNode(Ops, dl, VTs, AM, ExtType, MemVT, MMO); 3884 CSEMap.InsertNode(N, IP); 3885 AllNodes.push_back(N); 3886 return SDValue(N, 0); 3887} 3888 3889SDValue SelectionDAG::getLoad(EVT VT, DebugLoc dl, 3890 SDValue Chain, SDValue Ptr, 3891 const Value *SV, int SVOffset, 3892 bool isVolatile, bool isNonTemporal, 3893 unsigned Alignment) { 3894 SDValue Undef = getUNDEF(Ptr.getValueType()); 3895 return getLoad(ISD::UNINDEXED, dl, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef, 3896 SV, SVOffset, VT, isVolatile, isNonTemporal, Alignment); 3897} 3898 3899SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, EVT VT, 3900 SDValue Chain, SDValue Ptr, 3901 const Value *SV, 3902 int SVOffset, EVT MemVT, 3903 bool isVolatile, bool isNonTemporal, 3904 unsigned Alignment) { 3905 SDValue Undef = getUNDEF(Ptr.getValueType()); 3906 return getLoad(ISD::UNINDEXED, dl, ExtType, VT, Chain, Ptr, Undef, 3907 SV, SVOffset, MemVT, isVolatile, isNonTemporal, Alignment); 3908} 3909 3910SDValue 3911SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base, 3912 SDValue Offset, ISD::MemIndexedMode AM) { 3913 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad); 3914 assert(LD->getOffset().getOpcode() == ISD::UNDEF && 3915 "Load is already a indexed load!"); 3916 return getLoad(AM, dl, LD->getExtensionType(), OrigLoad.getValueType(), 3917 LD->getChain(), Base, Offset, LD->getSrcValue(), 3918 LD->getSrcValueOffset(), LD->getMemoryVT(), 3919 LD->isVolatile(), LD->isNonTemporal(), LD->getAlignment()); 3920} 3921 3922SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, 3923 SDValue Ptr, const Value *SV, int SVOffset, 3924 bool isVolatile, bool isNonTemporal, 3925 unsigned Alignment) { 3926 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3927 Alignment = getEVTAlignment(Val.getValueType()); 3928 3929 // Check if the memory reference references a frame index 3930 if (!SV) 3931 if (const FrameIndexSDNode *FI = 3932 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3933 SV = PseudoSourceValue::getFixedStack(FI->getIndex()); 3934 3935 MachineFunction &MF = getMachineFunction(); 3936 unsigned Flags = MachineMemOperand::MOStore; 3937 if (isVolatile) 3938 Flags |= MachineMemOperand::MOVolatile; 3939 if (isNonTemporal) 3940 Flags |= MachineMemOperand::MONonTemporal; 3941 MachineMemOperand *MMO = 3942 MF.getMachineMemOperand(SV, Flags, SVOffset, 3943 Val.getValueType().getStoreSize(), Alignment); 3944 3945 return getStore(Chain, dl, Val, Ptr, MMO); 3946} 3947 3948SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, 3949 SDValue Ptr, MachineMemOperand *MMO) { 3950 EVT VT = Val.getValueType(); 3951 SDVTList VTs = getVTList(MVT::Other); 3952 SDValue Undef = getUNDEF(Ptr.getValueType()); 3953 SDValue Ops[] = { Chain, Val, Ptr, Undef }; 3954 FoldingSetNodeID ID; 3955 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); 3956 ID.AddInteger(VT.getRawBits()); 3957 ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile(), 3958 MMO->isNonTemporal())); 3959 void *IP = 0; 3960 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3961 cast<StoreSDNode>(E)->refineAlignment(MMO); 3962 return SDValue(E, 0); 3963 } 3964 SDNode *N = NodeAllocator.Allocate<StoreSDNode>(); 3965 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, false, VT, MMO); 3966 CSEMap.InsertNode(N, IP); 3967 AllNodes.push_back(N); 3968 return SDValue(N, 0); 3969} 3970 3971SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val, 3972 SDValue Ptr, const Value *SV, 3973 int SVOffset, EVT SVT, 3974 bool isVolatile, bool isNonTemporal, 3975 unsigned Alignment) { 3976 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3977 Alignment = getEVTAlignment(SVT); 3978 3979 // Check if the memory reference references a frame index 3980 if (!SV) 3981 if (const FrameIndexSDNode *FI = 3982 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3983 SV = PseudoSourceValue::getFixedStack(FI->getIndex()); 3984 3985 MachineFunction &MF = getMachineFunction(); 3986 unsigned Flags = MachineMemOperand::MOStore; 3987 if (isVolatile) 3988 Flags |= MachineMemOperand::MOVolatile; 3989 if (isNonTemporal) 3990 Flags |= MachineMemOperand::MONonTemporal; 3991 MachineMemOperand *MMO = 3992 MF.getMachineMemOperand(SV, Flags, SVOffset, SVT.getStoreSize(), Alignment); 3993 3994 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO); 3995} 3996 3997SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val, 3998 SDValue Ptr, EVT SVT, 3999 MachineMemOperand *MMO) { 4000 EVT VT = Val.getValueType(); 4001 4002 if (VT == SVT) 4003 return getStore(Chain, dl, Val, Ptr, MMO); 4004 4005 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) && 4006 "Should only be a truncating store, not extending!"); 4007 assert(VT.isInteger() == SVT.isInteger() && 4008 "Can't do FP-INT conversion!"); 4009 assert(VT.isVector() == SVT.isVector() && 4010 "Cannot use trunc store to convert to or from a vector!"); 4011 assert((!VT.isVector() || 4012 VT.getVectorNumElements() == SVT.getVectorNumElements()) && 4013 "Cannot use trunc store to change the number of vector elements!"); 4014 4015 SDVTList VTs = getVTList(MVT::Other); 4016 SDValue Undef = getUNDEF(Ptr.getValueType()); 4017 SDValue Ops[] = { Chain, Val, Ptr, Undef }; 4018 FoldingSetNodeID ID; 4019 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); 4020 ID.AddInteger(SVT.getRawBits()); 4021 ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile(), 4022 MMO->isNonTemporal())); 4023 void *IP = 0; 4024 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 4025 cast<StoreSDNode>(E)->refineAlignment(MMO); 4026 return SDValue(E, 0); 4027 } 4028 SDNode *N = NodeAllocator.Allocate<StoreSDNode>(); 4029 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, true, SVT, MMO); 4030 CSEMap.InsertNode(N, IP); 4031 AllNodes.push_back(N); 4032 return SDValue(N, 0); 4033} 4034 4035SDValue 4036SelectionDAG::getIndexedStore(SDValue OrigStore, DebugLoc dl, SDValue Base, 4037 SDValue Offset, ISD::MemIndexedMode AM) { 4038 StoreSDNode *ST = cast<StoreSDNode>(OrigStore); 4039 assert(ST->getOffset().getOpcode() == ISD::UNDEF && 4040 "Store is already a indexed store!"); 4041 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other); 4042 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset }; 4043 FoldingSetNodeID ID; 4044 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); 4045 ID.AddInteger(ST->getMemoryVT().getRawBits()); 4046 ID.AddInteger(ST->getRawSubclassData()); 4047 void *IP = 0; 4048 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4049 return SDValue(E, 0); 4050 4051 SDNode *N = NodeAllocator.Allocate<StoreSDNode>(); 4052 new (N) StoreSDNode(Ops, dl, VTs, AM, 4053 ST->isTruncatingStore(), ST->getMemoryVT(), 4054 ST->getMemOperand()); 4055 CSEMap.InsertNode(N, IP); 4056 AllNodes.push_back(N); 4057 return SDValue(N, 0); 4058} 4059 4060SDValue SelectionDAG::getVAArg(EVT VT, DebugLoc dl, 4061 SDValue Chain, SDValue Ptr, 4062 SDValue SV) { 4063 SDValue Ops[] = { Chain, Ptr, SV }; 4064 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops, 3); 4065} 4066 4067SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 4068 const SDUse *Ops, unsigned NumOps) { 4069 switch (NumOps) { 4070 case 0: return getNode(Opcode, DL, VT); 4071 case 1: return getNode(Opcode, DL, VT, Ops[0]); 4072 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]); 4073 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]); 4074 default: break; 4075 } 4076 4077 // Copy from an SDUse array into an SDValue array for use with 4078 // the regular getNode logic. 4079 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps); 4080 return getNode(Opcode, DL, VT, &NewOps[0], NumOps); 4081} 4082 4083SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 4084 const SDValue *Ops, unsigned NumOps) { 4085 switch (NumOps) { 4086 case 0: return getNode(Opcode, DL, VT); 4087 case 1: return getNode(Opcode, DL, VT, Ops[0]); 4088 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]); 4089 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]); 4090 default: break; 4091 } 4092 4093 switch (Opcode) { 4094 default: break; 4095 case ISD::SELECT_CC: { 4096 assert(NumOps == 5 && "SELECT_CC takes 5 operands!"); 4097 assert(Ops[0].getValueType() == Ops[1].getValueType() && 4098 "LHS and RHS of condition must have same type!"); 4099 assert(Ops[2].getValueType() == Ops[3].getValueType() && 4100 "True and False arms of SelectCC must have same type!"); 4101 assert(Ops[2].getValueType() == VT && 4102 "select_cc node must be of same type as true and false value!"); 4103 break; 4104 } 4105 case ISD::BR_CC: { 4106 assert(NumOps == 5 && "BR_CC takes 5 operands!"); 4107 assert(Ops[2].getValueType() == Ops[3].getValueType() && 4108 "LHS/RHS of comparison should match types!"); 4109 break; 4110 } 4111 } 4112 4113 // Memoize nodes. 4114 SDNode *N; 4115 SDVTList VTs = getVTList(VT); 4116 4117 if (VT != MVT::Flag) { 4118 FoldingSetNodeID ID; 4119 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps); 4120 void *IP = 0; 4121 4122 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4123 return SDValue(E, 0); 4124 4125 N = NodeAllocator.Allocate<SDNode>(); 4126 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps); 4127 CSEMap.InsertNode(N, IP); 4128 } else { 4129 N = NodeAllocator.Allocate<SDNode>(); 4130 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps); 4131 } 4132 4133 AllNodes.push_back(N); 4134#ifndef NDEBUG 4135 VerifyNode(N); 4136#endif 4137 return SDValue(N, 0); 4138} 4139 4140SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, 4141 const std::vector<EVT> &ResultTys, 4142 const SDValue *Ops, unsigned NumOps) { 4143 return getNode(Opcode, DL, getVTList(&ResultTys[0], ResultTys.size()), 4144 Ops, NumOps); 4145} 4146 4147SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, 4148 const EVT *VTs, unsigned NumVTs, 4149 const SDValue *Ops, unsigned NumOps) { 4150 if (NumVTs == 1) 4151 return getNode(Opcode, DL, VTs[0], Ops, NumOps); 4152 return getNode(Opcode, DL, makeVTList(VTs, NumVTs), Ops, NumOps); 4153} 4154 4155SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4156 const SDValue *Ops, unsigned NumOps) { 4157 if (VTList.NumVTs == 1) 4158 return getNode(Opcode, DL, VTList.VTs[0], Ops, NumOps); 4159 4160#if 0 4161 switch (Opcode) { 4162 // FIXME: figure out how to safely handle things like 4163 // int foo(int x) { return 1 << (x & 255); } 4164 // int bar() { return foo(256); } 4165 case ISD::SRA_PARTS: 4166 case ISD::SRL_PARTS: 4167 case ISD::SHL_PARTS: 4168 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG && 4169 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1) 4170 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0)); 4171 else if (N3.getOpcode() == ISD::AND) 4172 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) { 4173 // If the and is only masking out bits that cannot effect the shift, 4174 // eliminate the and. 4175 unsigned NumBits = VT.getScalarType().getSizeInBits()*2; 4176 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1) 4177 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0)); 4178 } 4179 break; 4180 } 4181#endif 4182 4183 // Memoize the node unless it returns a flag. 4184 SDNode *N; 4185 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) { 4186 FoldingSetNodeID ID; 4187 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps); 4188 void *IP = 0; 4189 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4190 return SDValue(E, 0); 4191 4192 if (NumOps == 1) { 4193 N = NodeAllocator.Allocate<UnarySDNode>(); 4194 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]); 4195 } else if (NumOps == 2) { 4196 N = NodeAllocator.Allocate<BinarySDNode>(); 4197 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]); 4198 } else if (NumOps == 3) { 4199 N = NodeAllocator.Allocate<TernarySDNode>(); 4200 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]); 4201 } else { 4202 N = NodeAllocator.Allocate<SDNode>(); 4203 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps); 4204 } 4205 CSEMap.InsertNode(N, IP); 4206 } else { 4207 if (NumOps == 1) { 4208 N = NodeAllocator.Allocate<UnarySDNode>(); 4209 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]); 4210 } else if (NumOps == 2) { 4211 N = NodeAllocator.Allocate<BinarySDNode>(); 4212 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]); 4213 } else if (NumOps == 3) { 4214 N = NodeAllocator.Allocate<TernarySDNode>(); 4215 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]); 4216 } else { 4217 N = NodeAllocator.Allocate<SDNode>(); 4218 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps); 4219 } 4220 } 4221 AllNodes.push_back(N); 4222#ifndef NDEBUG 4223 VerifyNode(N); 4224#endif 4225 return SDValue(N, 0); 4226} 4227 4228SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList) { 4229 return getNode(Opcode, DL, VTList, 0, 0); 4230} 4231 4232SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4233 SDValue N1) { 4234 SDValue Ops[] = { N1 }; 4235 return getNode(Opcode, DL, VTList, Ops, 1); 4236} 4237 4238SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4239 SDValue N1, SDValue N2) { 4240 SDValue Ops[] = { N1, N2 }; 4241 return getNode(Opcode, DL, VTList, Ops, 2); 4242} 4243 4244SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4245 SDValue N1, SDValue N2, SDValue N3) { 4246 SDValue Ops[] = { N1, N2, N3 }; 4247 return getNode(Opcode, DL, VTList, Ops, 3); 4248} 4249 4250SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4251 SDValue N1, SDValue N2, SDValue N3, 4252 SDValue N4) { 4253 SDValue Ops[] = { N1, N2, N3, N4 }; 4254 return getNode(Opcode, DL, VTList, Ops, 4); 4255} 4256 4257SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4258 SDValue N1, SDValue N2, SDValue N3, 4259 SDValue N4, SDValue N5) { 4260 SDValue Ops[] = { N1, N2, N3, N4, N5 }; 4261 return getNode(Opcode, DL, VTList, Ops, 5); 4262} 4263 4264SDVTList SelectionDAG::getVTList(EVT VT) { 4265 return makeVTList(SDNode::getValueTypeList(VT), 1); 4266} 4267 4268SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) { 4269 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4270 E = VTList.rend(); I != E; ++I) 4271 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2) 4272 return *I; 4273 4274 EVT *Array = Allocator.Allocate<EVT>(2); 4275 Array[0] = VT1; 4276 Array[1] = VT2; 4277 SDVTList Result = makeVTList(Array, 2); 4278 VTList.push_back(Result); 4279 return Result; 4280} 4281 4282SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) { 4283 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4284 E = VTList.rend(); I != E; ++I) 4285 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 && 4286 I->VTs[2] == VT3) 4287 return *I; 4288 4289 EVT *Array = Allocator.Allocate<EVT>(3); 4290 Array[0] = VT1; 4291 Array[1] = VT2; 4292 Array[2] = VT3; 4293 SDVTList Result = makeVTList(Array, 3); 4294 VTList.push_back(Result); 4295 return Result; 4296} 4297 4298SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) { 4299 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4300 E = VTList.rend(); I != E; ++I) 4301 if (I->NumVTs == 4 && I->VTs[0] == VT1 && I->VTs[1] == VT2 && 4302 I->VTs[2] == VT3 && I->VTs[3] == VT4) 4303 return *I; 4304 4305 EVT *Array = Allocator.Allocate<EVT>(4); 4306 Array[0] = VT1; 4307 Array[1] = VT2; 4308 Array[2] = VT3; 4309 Array[3] = VT4; 4310 SDVTList Result = makeVTList(Array, 4); 4311 VTList.push_back(Result); 4312 return Result; 4313} 4314 4315SDVTList SelectionDAG::getVTList(const EVT *VTs, unsigned NumVTs) { 4316 switch (NumVTs) { 4317 case 0: llvm_unreachable("Cannot have nodes without results!"); 4318 case 1: return getVTList(VTs[0]); 4319 case 2: return getVTList(VTs[0], VTs[1]); 4320 case 3: return getVTList(VTs[0], VTs[1], VTs[2]); 4321 case 4: return getVTList(VTs[0], VTs[1], VTs[2], VTs[3]); 4322 default: break; 4323 } 4324 4325 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4326 E = VTList.rend(); I != E; ++I) { 4327 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1]) 4328 continue; 4329 4330 bool NoMatch = false; 4331 for (unsigned i = 2; i != NumVTs; ++i) 4332 if (VTs[i] != I->VTs[i]) { 4333 NoMatch = true; 4334 break; 4335 } 4336 if (!NoMatch) 4337 return *I; 4338 } 4339 4340 EVT *Array = Allocator.Allocate<EVT>(NumVTs); 4341 std::copy(VTs, VTs+NumVTs, Array); 4342 SDVTList Result = makeVTList(Array, NumVTs); 4343 VTList.push_back(Result); 4344 return Result; 4345} 4346 4347 4348/// UpdateNodeOperands - *Mutate* the specified node in-place to have the 4349/// specified operands. If the resultant node already exists in the DAG, 4350/// this does not modify the specified node, instead it returns the node that 4351/// already exists. If the resultant node does not exist in the DAG, the 4352/// input node is returned. As a degenerate case, if you specify the same 4353/// input operands as the node already has, the input node is returned. 4354SDValue SelectionDAG::UpdateNodeOperands(SDValue InN, SDValue Op) { 4355 SDNode *N = InN.getNode(); 4356 assert(N->getNumOperands() == 1 && "Update with wrong number of operands"); 4357 4358 // Check to see if there is no change. 4359 if (Op == N->getOperand(0)) return InN; 4360 4361 // See if the modified node already exists. 4362 void *InsertPos = 0; 4363 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos)) 4364 return SDValue(Existing, InN.getResNo()); 4365 4366 // Nope it doesn't. Remove the node from its current place in the maps. 4367 if (InsertPos) 4368 if (!RemoveNodeFromCSEMaps(N)) 4369 InsertPos = 0; 4370 4371 // Now we update the operands. 4372 N->OperandList[0].set(Op); 4373 4374 // If this gets put into a CSE map, add it. 4375 if (InsertPos) CSEMap.InsertNode(N, InsertPos); 4376 return InN; 4377} 4378 4379SDValue SelectionDAG:: 4380UpdateNodeOperands(SDValue InN, SDValue Op1, SDValue Op2) { 4381 SDNode *N = InN.getNode(); 4382 assert(N->getNumOperands() == 2 && "Update with wrong number of operands"); 4383 4384 // Check to see if there is no change. 4385 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1)) 4386 return InN; // No operands changed, just return the input node. 4387 4388 // See if the modified node already exists. 4389 void *InsertPos = 0; 4390 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos)) 4391 return SDValue(Existing, InN.getResNo()); 4392 4393 // Nope it doesn't. Remove the node from its current place in the maps. 4394 if (InsertPos) 4395 if (!RemoveNodeFromCSEMaps(N)) 4396 InsertPos = 0; 4397 4398 // Now we update the operands. 4399 if (N->OperandList[0] != Op1) 4400 N->OperandList[0].set(Op1); 4401 if (N->OperandList[1] != Op2) 4402 N->OperandList[1].set(Op2); 4403 4404 // If this gets put into a CSE map, add it. 4405 if (InsertPos) CSEMap.InsertNode(N, InsertPos); 4406 return InN; 4407} 4408 4409SDValue SelectionDAG:: 4410UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, SDValue Op3) { 4411 SDValue Ops[] = { Op1, Op2, Op3 }; 4412 return UpdateNodeOperands(N, Ops, 3); 4413} 4414 4415SDValue SelectionDAG:: 4416UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, 4417 SDValue Op3, SDValue Op4) { 4418 SDValue Ops[] = { Op1, Op2, Op3, Op4 }; 4419 return UpdateNodeOperands(N, Ops, 4); 4420} 4421 4422SDValue SelectionDAG:: 4423UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, 4424 SDValue Op3, SDValue Op4, SDValue Op5) { 4425 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 }; 4426 return UpdateNodeOperands(N, Ops, 5); 4427} 4428 4429SDValue SelectionDAG:: 4430UpdateNodeOperands(SDValue InN, const SDValue *Ops, unsigned NumOps) { 4431 SDNode *N = InN.getNode(); 4432 assert(N->getNumOperands() == NumOps && 4433 "Update with wrong number of operands"); 4434 4435 // Check to see if there is no change. 4436 bool AnyChange = false; 4437 for (unsigned i = 0; i != NumOps; ++i) { 4438 if (Ops[i] != N->getOperand(i)) { 4439 AnyChange = true; 4440 break; 4441 } 4442 } 4443 4444 // No operands changed, just return the input node. 4445 if (!AnyChange) return InN; 4446 4447 // See if the modified node already exists. 4448 void *InsertPos = 0; 4449 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos)) 4450 return SDValue(Existing, InN.getResNo()); 4451 4452 // Nope it doesn't. Remove the node from its current place in the maps. 4453 if (InsertPos) 4454 if (!RemoveNodeFromCSEMaps(N)) 4455 InsertPos = 0; 4456 4457 // Now we update the operands. 4458 for (unsigned i = 0; i != NumOps; ++i) 4459 if (N->OperandList[i] != Ops[i]) 4460 N->OperandList[i].set(Ops[i]); 4461 4462 // If this gets put into a CSE map, add it. 4463 if (InsertPos) CSEMap.InsertNode(N, InsertPos); 4464 return InN; 4465} 4466 4467/// DropOperands - Release the operands and set this node to have 4468/// zero operands. 4469void SDNode::DropOperands() { 4470 // Unlike the code in MorphNodeTo that does this, we don't need to 4471 // watch for dead nodes here. 4472 for (op_iterator I = op_begin(), E = op_end(); I != E; ) { 4473 SDUse &Use = *I++; 4474 Use.set(SDValue()); 4475 } 4476} 4477 4478/// SelectNodeTo - These are wrappers around MorphNodeTo that accept a 4479/// machine opcode. 4480/// 4481SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4482 EVT VT) { 4483 SDVTList VTs = getVTList(VT); 4484 return SelectNodeTo(N, MachineOpc, VTs, 0, 0); 4485} 4486 4487SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4488 EVT VT, SDValue Op1) { 4489 SDVTList VTs = getVTList(VT); 4490 SDValue Ops[] = { Op1 }; 4491 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1); 4492} 4493 4494SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4495 EVT VT, SDValue Op1, 4496 SDValue Op2) { 4497 SDVTList VTs = getVTList(VT); 4498 SDValue Ops[] = { Op1, Op2 }; 4499 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2); 4500} 4501 4502SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4503 EVT VT, SDValue Op1, 4504 SDValue Op2, SDValue Op3) { 4505 SDVTList VTs = getVTList(VT); 4506 SDValue Ops[] = { Op1, Op2, Op3 }; 4507 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3); 4508} 4509 4510SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4511 EVT VT, const SDValue *Ops, 4512 unsigned NumOps) { 4513 SDVTList VTs = getVTList(VT); 4514 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4515} 4516 4517SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4518 EVT VT1, EVT VT2, const SDValue *Ops, 4519 unsigned NumOps) { 4520 SDVTList VTs = getVTList(VT1, VT2); 4521 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4522} 4523 4524SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4525 EVT VT1, EVT VT2) { 4526 SDVTList VTs = getVTList(VT1, VT2); 4527 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0); 4528} 4529 4530SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4531 EVT VT1, EVT VT2, EVT VT3, 4532 const SDValue *Ops, unsigned NumOps) { 4533 SDVTList VTs = getVTList(VT1, VT2, VT3); 4534 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4535} 4536 4537SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4538 EVT VT1, EVT VT2, EVT VT3, EVT VT4, 4539 const SDValue *Ops, unsigned NumOps) { 4540 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4); 4541 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4542} 4543 4544SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4545 EVT VT1, EVT VT2, 4546 SDValue Op1) { 4547 SDVTList VTs = getVTList(VT1, VT2); 4548 SDValue Ops[] = { Op1 }; 4549 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1); 4550} 4551 4552SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4553 EVT VT1, EVT VT2, 4554 SDValue Op1, SDValue Op2) { 4555 SDVTList VTs = getVTList(VT1, VT2); 4556 SDValue Ops[] = { Op1, Op2 }; 4557 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2); 4558} 4559 4560SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4561 EVT VT1, EVT VT2, 4562 SDValue Op1, SDValue Op2, 4563 SDValue Op3) { 4564 SDVTList VTs = getVTList(VT1, VT2); 4565 SDValue Ops[] = { Op1, Op2, Op3 }; 4566 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3); 4567} 4568 4569SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4570 EVT VT1, EVT VT2, EVT VT3, 4571 SDValue Op1, SDValue Op2, 4572 SDValue Op3) { 4573 SDVTList VTs = getVTList(VT1, VT2, VT3); 4574 SDValue Ops[] = { Op1, Op2, Op3 }; 4575 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3); 4576} 4577 4578SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4579 SDVTList VTs, const SDValue *Ops, 4580 unsigned NumOps) { 4581 N = MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps); 4582 // Reset the NodeID to -1. 4583 N->setNodeId(-1); 4584 return N; 4585} 4586 4587/// MorphNodeTo - This *mutates* the specified node to have the specified 4588/// return type, opcode, and operands. 4589/// 4590/// Note that MorphNodeTo returns the resultant node. If there is already a 4591/// node of the specified opcode and operands, it returns that node instead of 4592/// the current one. Note that the DebugLoc need not be the same. 4593/// 4594/// Using MorphNodeTo is faster than creating a new node and swapping it in 4595/// with ReplaceAllUsesWith both because it often avoids allocating a new 4596/// node, and because it doesn't require CSE recalculation for any of 4597/// the node's users. 4598/// 4599SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4600 SDVTList VTs, const SDValue *Ops, 4601 unsigned NumOps) { 4602 // If an identical node already exists, use it. 4603 void *IP = 0; 4604 if (VTs.VTs[VTs.NumVTs-1] != MVT::Flag) { 4605 FoldingSetNodeID ID; 4606 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps); 4607 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP)) 4608 return ON; 4609 } 4610 4611 if (!RemoveNodeFromCSEMaps(N)) 4612 IP = 0; 4613 4614 // Start the morphing. 4615 N->NodeType = Opc; 4616 N->ValueList = VTs.VTs; 4617 N->NumValues = VTs.NumVTs; 4618 4619 // Clear the operands list, updating used nodes to remove this from their 4620 // use list. Keep track of any operands that become dead as a result. 4621 SmallPtrSet<SDNode*, 16> DeadNodeSet; 4622 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) { 4623 SDUse &Use = *I++; 4624 SDNode *Used = Use.getNode(); 4625 Use.set(SDValue()); 4626 if (Used->use_empty()) 4627 DeadNodeSet.insert(Used); 4628 } 4629 4630 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N)) { 4631 // Initialize the memory references information. 4632 MN->setMemRefs(0, 0); 4633 // If NumOps is larger than the # of operands we can have in a 4634 // MachineSDNode, reallocate the operand list. 4635 if (NumOps > MN->NumOperands || !MN->OperandsNeedDelete) { 4636 if (MN->OperandsNeedDelete) 4637 delete[] MN->OperandList; 4638 if (NumOps > array_lengthof(MN->LocalOperands)) 4639 // We're creating a final node that will live unmorphed for the 4640 // remainder of the current SelectionDAG iteration, so we can allocate 4641 // the operands directly out of a pool with no recycling metadata. 4642 MN->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps), 4643 Ops, NumOps); 4644 else 4645 MN->InitOperands(MN->LocalOperands, Ops, NumOps); 4646 MN->OperandsNeedDelete = false; 4647 } else 4648 MN->InitOperands(MN->OperandList, Ops, NumOps); 4649 } else { 4650 // If NumOps is larger than the # of operands we currently have, reallocate 4651 // the operand list. 4652 if (NumOps > N->NumOperands) { 4653 if (N->OperandsNeedDelete) 4654 delete[] N->OperandList; 4655 N->InitOperands(new SDUse[NumOps], Ops, NumOps); 4656 N->OperandsNeedDelete = true; 4657 } else 4658 N->InitOperands(N->OperandList, Ops, NumOps); 4659 } 4660 4661 // Delete any nodes that are still dead after adding the uses for the 4662 // new operands. 4663 if (!DeadNodeSet.empty()) { 4664 SmallVector<SDNode *, 16> DeadNodes; 4665 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(), 4666 E = DeadNodeSet.end(); I != E; ++I) 4667 if ((*I)->use_empty()) 4668 DeadNodes.push_back(*I); 4669 RemoveDeadNodes(DeadNodes); 4670 } 4671 4672 if (IP) 4673 CSEMap.InsertNode(N, IP); // Memoize the new node. 4674 return N; 4675} 4676 4677 4678/// getMachineNode - These are used for target selectors to create a new node 4679/// with specified return type(s), MachineInstr opcode, and operands. 4680/// 4681/// Note that getMachineNode returns the resultant node. If there is already a 4682/// node of the specified opcode and operands, it returns that node instead of 4683/// the current one. 4684MachineSDNode * 4685SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT) { 4686 SDVTList VTs = getVTList(VT); 4687 return getMachineNode(Opcode, dl, VTs, 0, 0); 4688} 4689 4690MachineSDNode * 4691SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, SDValue Op1) { 4692 SDVTList VTs = getVTList(VT); 4693 SDValue Ops[] = { Op1 }; 4694 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4695} 4696 4697MachineSDNode * 4698SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, 4699 SDValue Op1, SDValue Op2) { 4700 SDVTList VTs = getVTList(VT); 4701 SDValue Ops[] = { Op1, Op2 }; 4702 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4703} 4704 4705MachineSDNode * 4706SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, 4707 SDValue Op1, SDValue Op2, SDValue Op3) { 4708 SDVTList VTs = getVTList(VT); 4709 SDValue Ops[] = { Op1, Op2, Op3 }; 4710 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4711} 4712 4713MachineSDNode * 4714SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, 4715 const SDValue *Ops, unsigned NumOps) { 4716 SDVTList VTs = getVTList(VT); 4717 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4718} 4719 4720MachineSDNode * 4721SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, EVT VT2) { 4722 SDVTList VTs = getVTList(VT1, VT2); 4723 return getMachineNode(Opcode, dl, VTs, 0, 0); 4724} 4725 4726MachineSDNode * 4727SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4728 EVT VT1, EVT VT2, SDValue Op1) { 4729 SDVTList VTs = getVTList(VT1, VT2); 4730 SDValue Ops[] = { Op1 }; 4731 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4732} 4733 4734MachineSDNode * 4735SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4736 EVT VT1, EVT VT2, SDValue Op1, SDValue Op2) { 4737 SDVTList VTs = getVTList(VT1, VT2); 4738 SDValue Ops[] = { Op1, Op2 }; 4739 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4740} 4741 4742MachineSDNode * 4743SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4744 EVT VT1, EVT VT2, SDValue Op1, 4745 SDValue Op2, SDValue Op3) { 4746 SDVTList VTs = getVTList(VT1, VT2); 4747 SDValue Ops[] = { Op1, Op2, Op3 }; 4748 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4749} 4750 4751MachineSDNode * 4752SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4753 EVT VT1, EVT VT2, 4754 const SDValue *Ops, unsigned NumOps) { 4755 SDVTList VTs = getVTList(VT1, VT2); 4756 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4757} 4758 4759MachineSDNode * 4760SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4761 EVT VT1, EVT VT2, EVT VT3, 4762 SDValue Op1, SDValue Op2) { 4763 SDVTList VTs = getVTList(VT1, VT2, VT3); 4764 SDValue Ops[] = { Op1, Op2 }; 4765 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4766} 4767 4768MachineSDNode * 4769SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4770 EVT VT1, EVT VT2, EVT VT3, 4771 SDValue Op1, SDValue Op2, SDValue Op3) { 4772 SDVTList VTs = getVTList(VT1, VT2, VT3); 4773 SDValue Ops[] = { Op1, Op2, Op3 }; 4774 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4775} 4776 4777MachineSDNode * 4778SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4779 EVT VT1, EVT VT2, EVT VT3, 4780 const SDValue *Ops, unsigned NumOps) { 4781 SDVTList VTs = getVTList(VT1, VT2, VT3); 4782 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4783} 4784 4785MachineSDNode * 4786SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, 4787 EVT VT2, EVT VT3, EVT VT4, 4788 const SDValue *Ops, unsigned NumOps) { 4789 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4); 4790 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4791} 4792 4793MachineSDNode * 4794SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4795 const std::vector<EVT> &ResultTys, 4796 const SDValue *Ops, unsigned NumOps) { 4797 SDVTList VTs = getVTList(&ResultTys[0], ResultTys.size()); 4798 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4799} 4800 4801MachineSDNode * 4802SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc DL, SDVTList VTs, 4803 const SDValue *Ops, unsigned NumOps) { 4804 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Flag; 4805 MachineSDNode *N; 4806 void *IP; 4807 4808 if (DoCSE) { 4809 FoldingSetNodeID ID; 4810 AddNodeIDNode(ID, ~Opcode, VTs, Ops, NumOps); 4811 IP = 0; 4812 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4813 return cast<MachineSDNode>(E); 4814 } 4815 4816 // Allocate a new MachineSDNode. 4817 N = NodeAllocator.Allocate<MachineSDNode>(); 4818 new (N) MachineSDNode(~Opcode, DL, VTs); 4819 4820 // Initialize the operands list. 4821 if (NumOps > array_lengthof(N->LocalOperands)) 4822 // We're creating a final node that will live unmorphed for the 4823 // remainder of the current SelectionDAG iteration, so we can allocate 4824 // the operands directly out of a pool with no recycling metadata. 4825 N->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps), 4826 Ops, NumOps); 4827 else 4828 N->InitOperands(N->LocalOperands, Ops, NumOps); 4829 N->OperandsNeedDelete = false; 4830 4831 if (DoCSE) 4832 CSEMap.InsertNode(N, IP); 4833 4834 AllNodes.push_back(N); 4835#ifndef NDEBUG 4836 VerifyNode(N); 4837#endif 4838 return N; 4839} 4840 4841/// getTargetExtractSubreg - A convenience function for creating 4842/// TargetOpcode::EXTRACT_SUBREG nodes. 4843SDValue 4844SelectionDAG::getTargetExtractSubreg(int SRIdx, DebugLoc DL, EVT VT, 4845 SDValue Operand) { 4846 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32); 4847 SDNode *Subreg = getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, 4848 VT, Operand, SRIdxVal); 4849 return SDValue(Subreg, 0); 4850} 4851 4852/// getTargetInsertSubreg - A convenience function for creating 4853/// TargetOpcode::INSERT_SUBREG nodes. 4854SDValue 4855SelectionDAG::getTargetInsertSubreg(int SRIdx, DebugLoc DL, EVT VT, 4856 SDValue Operand, SDValue Subreg) { 4857 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32); 4858 SDNode *Result = getMachineNode(TargetOpcode::INSERT_SUBREG, DL, 4859 VT, Operand, Subreg, SRIdxVal); 4860 return SDValue(Result, 0); 4861} 4862 4863/// getNodeIfExists - Get the specified node if it's already available, or 4864/// else return NULL. 4865SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList, 4866 const SDValue *Ops, unsigned NumOps) { 4867 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) { 4868 FoldingSetNodeID ID; 4869 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps); 4870 void *IP = 0; 4871 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4872 return E; 4873 } 4874 return NULL; 4875} 4876 4877namespace { 4878 4879/// RAUWUpdateListener - Helper for ReplaceAllUsesWith - When the node 4880/// pointed to by a use iterator is deleted, increment the use iterator 4881/// so that it doesn't dangle. 4882/// 4883/// This class also manages a "downlink" DAGUpdateListener, to forward 4884/// messages to ReplaceAllUsesWith's callers. 4885/// 4886class RAUWUpdateListener : public SelectionDAG::DAGUpdateListener { 4887 SelectionDAG::DAGUpdateListener *DownLink; 4888 SDNode::use_iterator &UI; 4889 SDNode::use_iterator &UE; 4890 4891 virtual void NodeDeleted(SDNode *N, SDNode *E) { 4892 // Increment the iterator as needed. 4893 while (UI != UE && N == *UI) 4894 ++UI; 4895 4896 // Then forward the message. 4897 if (DownLink) DownLink->NodeDeleted(N, E); 4898 } 4899 4900 virtual void NodeUpdated(SDNode *N) { 4901 // Just forward the message. 4902 if (DownLink) DownLink->NodeUpdated(N); 4903 } 4904 4905public: 4906 RAUWUpdateListener(SelectionDAG::DAGUpdateListener *dl, 4907 SDNode::use_iterator &ui, 4908 SDNode::use_iterator &ue) 4909 : DownLink(dl), UI(ui), UE(ue) {} 4910}; 4911 4912} 4913 4914/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. 4915/// This can cause recursive merging of nodes in the DAG. 4916/// 4917/// This version assumes From has a single result value. 4918/// 4919void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To, 4920 DAGUpdateListener *UpdateListener) { 4921 SDNode *From = FromN.getNode(); 4922 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 && 4923 "Cannot replace with this method!"); 4924 assert(From != To.getNode() && "Cannot replace uses of with self"); 4925 4926 // Iterate over all the existing uses of From. New uses will be added 4927 // to the beginning of the use list, which we avoid visiting. 4928 // This specifically avoids visiting uses of From that arise while the 4929 // replacement is happening, because any such uses would be the result 4930 // of CSE: If an existing node looks like From after one of its operands 4931 // is replaced by To, we don't want to replace of all its users with To 4932 // too. See PR3018 for more info. 4933 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); 4934 RAUWUpdateListener Listener(UpdateListener, UI, UE); 4935 while (UI != UE) { 4936 SDNode *User = *UI; 4937 4938 // This node is about to morph, remove its old self from the CSE maps. 4939 RemoveNodeFromCSEMaps(User); 4940 4941 // A user can appear in a use list multiple times, and when this 4942 // happens the uses are usually next to each other in the list. 4943 // To help reduce the number of CSE recomputations, process all 4944 // the uses of this user that we can find this way. 4945 do { 4946 SDUse &Use = UI.getUse(); 4947 ++UI; 4948 Use.set(To); 4949 } while (UI != UE && *UI == User); 4950 4951 // Now that we have modified User, add it back to the CSE maps. If it 4952 // already exists there, recursively merge the results together. 4953 AddModifiedNodeToCSEMaps(User, &Listener); 4954 } 4955} 4956 4957/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. 4958/// This can cause recursive merging of nodes in the DAG. 4959/// 4960/// This version assumes that for each value of From, there is a 4961/// corresponding value in To in the same position with the same type. 4962/// 4963void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To, 4964 DAGUpdateListener *UpdateListener) { 4965#ifndef NDEBUG 4966 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i) 4967 assert((!From->hasAnyUseOfValue(i) || 4968 From->getValueType(i) == To->getValueType(i)) && 4969 "Cannot use this version of ReplaceAllUsesWith!"); 4970#endif 4971 4972 // Handle the trivial case. 4973 if (From == To) 4974 return; 4975 4976 // Iterate over just the existing users of From. See the comments in 4977 // the ReplaceAllUsesWith above. 4978 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); 4979 RAUWUpdateListener Listener(UpdateListener, UI, UE); 4980 while (UI != UE) { 4981 SDNode *User = *UI; 4982 4983 // This node is about to morph, remove its old self from the CSE maps. 4984 RemoveNodeFromCSEMaps(User); 4985 4986 // A user can appear in a use list multiple times, and when this 4987 // happens the uses are usually next to each other in the list. 4988 // To help reduce the number of CSE recomputations, process all 4989 // the uses of this user that we can find this way. 4990 do { 4991 SDUse &Use = UI.getUse(); 4992 ++UI; 4993 Use.setNode(To); 4994 } while (UI != UE && *UI == User); 4995 4996 // Now that we have modified User, add it back to the CSE maps. If it 4997 // already exists there, recursively merge the results together. 4998 AddModifiedNodeToCSEMaps(User, &Listener); 4999 } 5000} 5001 5002/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. 5003/// This can cause recursive merging of nodes in the DAG. 5004/// 5005/// This version can replace From with any result values. To must match the 5006/// number and types of values returned by From. 5007void SelectionDAG::ReplaceAllUsesWith(SDNode *From, 5008 const SDValue *To, 5009 DAGUpdateListener *UpdateListener) { 5010 if (From->getNumValues() == 1) // Handle the simple case efficiently. 5011 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener); 5012 5013 // Iterate over just the existing users of From. See the comments in 5014 // the ReplaceAllUsesWith above. 5015 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); 5016 RAUWUpdateListener Listener(UpdateListener, UI, UE); 5017 while (UI != UE) { 5018 SDNode *User = *UI; 5019 5020 // This node is about to morph, remove its old self from the CSE maps. 5021 RemoveNodeFromCSEMaps(User); 5022 5023 // A user can appear in a use list multiple times, and when this 5024 // happens the uses are usually next to each other in the list. 5025 // To help reduce the number of CSE recomputations, process all 5026 // the uses of this user that we can find this way. 5027 do { 5028 SDUse &Use = UI.getUse(); 5029 const SDValue &ToOp = To[Use.getResNo()]; 5030 ++UI; 5031 Use.set(ToOp); 5032 } while (UI != UE && *UI == User); 5033 5034 // Now that we have modified User, add it back to the CSE maps. If it 5035 // already exists there, recursively merge the results together. 5036 AddModifiedNodeToCSEMaps(User, &Listener); 5037 } 5038} 5039 5040/// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving 5041/// uses of other values produced by From.getNode() alone. The Deleted 5042/// vector is handled the same way as for ReplaceAllUsesWith. 5043void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To, 5044 DAGUpdateListener *UpdateListener){ 5045 // Handle the really simple, really trivial case efficiently. 5046 if (From == To) return; 5047 5048 // Handle the simple, trivial, case efficiently. 5049 if (From.getNode()->getNumValues() == 1) { 5050 ReplaceAllUsesWith(From, To, UpdateListener); 5051 return; 5052 } 5053 5054 // Iterate over just the existing users of From. See the comments in 5055 // the ReplaceAllUsesWith above. 5056 SDNode::use_iterator UI = From.getNode()->use_begin(), 5057 UE = From.getNode()->use_end(); 5058 RAUWUpdateListener Listener(UpdateListener, UI, UE); 5059 while (UI != UE) { 5060 SDNode *User = *UI; 5061 bool UserRemovedFromCSEMaps = false; 5062 5063 // A user can appear in a use list multiple times, and when this 5064 // happens the uses are usually next to each other in the list. 5065 // To help reduce the number of CSE recomputations, process all 5066 // the uses of this user that we can find this way. 5067 do { 5068 SDUse &Use = UI.getUse(); 5069 5070 // Skip uses of different values from the same node. 5071 if (Use.getResNo() != From.getResNo()) { 5072 ++UI; 5073 continue; 5074 } 5075 5076 // If this node hasn't been modified yet, it's still in the CSE maps, 5077 // so remove its old self from the CSE maps. 5078 if (!UserRemovedFromCSEMaps) { 5079 RemoveNodeFromCSEMaps(User); 5080 UserRemovedFromCSEMaps = true; 5081 } 5082 5083 ++UI; 5084 Use.set(To); 5085 } while (UI != UE && *UI == User); 5086 5087 // We are iterating over all uses of the From node, so if a use 5088 // doesn't use the specific value, no changes are made. 5089 if (!UserRemovedFromCSEMaps) 5090 continue; 5091 5092 // Now that we have modified User, add it back to the CSE maps. If it 5093 // already exists there, recursively merge the results together. 5094 AddModifiedNodeToCSEMaps(User, &Listener); 5095 } 5096} 5097 5098namespace { 5099 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith 5100 /// to record information about a use. 5101 struct UseMemo { 5102 SDNode *User; 5103 unsigned Index; 5104 SDUse *Use; 5105 }; 5106 5107 /// operator< - Sort Memos by User. 5108 bool operator<(const UseMemo &L, const UseMemo &R) { 5109 return (intptr_t)L.User < (intptr_t)R.User; 5110 } 5111} 5112 5113/// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving 5114/// uses of other values produced by From.getNode() alone. The same value 5115/// may appear in both the From and To list. The Deleted vector is 5116/// handled the same way as for ReplaceAllUsesWith. 5117void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From, 5118 const SDValue *To, 5119 unsigned Num, 5120 DAGUpdateListener *UpdateListener){ 5121 // Handle the simple, trivial case efficiently. 5122 if (Num == 1) 5123 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener); 5124 5125 // Read up all the uses and make records of them. This helps 5126 // processing new uses that are introduced during the 5127 // replacement process. 5128 SmallVector<UseMemo, 4> Uses; 5129 for (unsigned i = 0; i != Num; ++i) { 5130 unsigned FromResNo = From[i].getResNo(); 5131 SDNode *FromNode = From[i].getNode(); 5132 for (SDNode::use_iterator UI = FromNode->use_begin(), 5133 E = FromNode->use_end(); UI != E; ++UI) { 5134 SDUse &Use = UI.getUse(); 5135 if (Use.getResNo() == FromResNo) { 5136 UseMemo Memo = { *UI, i, &Use }; 5137 Uses.push_back(Memo); 5138 } 5139 } 5140 } 5141 5142 // Sort the uses, so that all the uses from a given User are together. 5143 std::sort(Uses.begin(), Uses.end()); 5144 5145 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size(); 5146 UseIndex != UseIndexEnd; ) { 5147 // We know that this user uses some value of From. If it is the right 5148 // value, update it. 5149 SDNode *User = Uses[UseIndex].User; 5150 5151 // This node is about to morph, remove its old self from the CSE maps. 5152 RemoveNodeFromCSEMaps(User); 5153 5154 // The Uses array is sorted, so all the uses for a given User 5155 // are next to each other in the list. 5156 // To help reduce the number of CSE recomputations, process all 5157 // the uses of this user that we can find this way. 5158 do { 5159 unsigned i = Uses[UseIndex].Index; 5160 SDUse &Use = *Uses[UseIndex].Use; 5161 ++UseIndex; 5162 5163 Use.set(To[i]); 5164 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User); 5165 5166 // Now that we have modified User, add it back to the CSE maps. If it 5167 // already exists there, recursively merge the results together. 5168 AddModifiedNodeToCSEMaps(User, UpdateListener); 5169 } 5170} 5171 5172/// AssignTopologicalOrder - Assign a unique node id for each node in the DAG 5173/// based on their topological order. It returns the maximum id and a vector 5174/// of the SDNodes* in assigned order by reference. 5175unsigned SelectionDAG::AssignTopologicalOrder() { 5176 5177 unsigned DAGSize = 0; 5178 5179 // SortedPos tracks the progress of the algorithm. Nodes before it are 5180 // sorted, nodes after it are unsorted. When the algorithm completes 5181 // it is at the end of the list. 5182 allnodes_iterator SortedPos = allnodes_begin(); 5183 5184 // Visit all the nodes. Move nodes with no operands to the front of 5185 // the list immediately. Annotate nodes that do have operands with their 5186 // operand count. Before we do this, the Node Id fields of the nodes 5187 // may contain arbitrary values. After, the Node Id fields for nodes 5188 // before SortedPos will contain the topological sort index, and the 5189 // Node Id fields for nodes At SortedPos and after will contain the 5190 // count of outstanding operands. 5191 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) { 5192 SDNode *N = I++; 5193 checkForCycles(N); 5194 unsigned Degree = N->getNumOperands(); 5195 if (Degree == 0) { 5196 // A node with no uses, add it to the result array immediately. 5197 N->setNodeId(DAGSize++); 5198 allnodes_iterator Q = N; 5199 if (Q != SortedPos) 5200 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q)); 5201 assert(SortedPos != AllNodes.end() && "Overran node list"); 5202 ++SortedPos; 5203 } else { 5204 // Temporarily use the Node Id as scratch space for the degree count. 5205 N->setNodeId(Degree); 5206 } 5207 } 5208 5209 // Visit all the nodes. As we iterate, moves nodes into sorted order, 5210 // such that by the time the end is reached all nodes will be sorted. 5211 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) { 5212 SDNode *N = I; 5213 checkForCycles(N); 5214 // N is in sorted position, so all its uses have one less operand 5215 // that needs to be sorted. 5216 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 5217 UI != UE; ++UI) { 5218 SDNode *P = *UI; 5219 unsigned Degree = P->getNodeId(); 5220 assert(Degree != 0 && "Invalid node degree"); 5221 --Degree; 5222 if (Degree == 0) { 5223 // All of P's operands are sorted, so P may sorted now. 5224 P->setNodeId(DAGSize++); 5225 if (P != SortedPos) 5226 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P)); 5227 assert(SortedPos != AllNodes.end() && "Overran node list"); 5228 ++SortedPos; 5229 } else { 5230 // Update P's outstanding operand count. 5231 P->setNodeId(Degree); 5232 } 5233 } 5234 if (I == SortedPos) { 5235#ifndef NDEBUG 5236 SDNode *S = ++I; 5237 dbgs() << "Overran sorted position:\n"; 5238 S->dumprFull(); 5239#endif 5240 llvm_unreachable(0); 5241 } 5242 } 5243 5244 assert(SortedPos == AllNodes.end() && 5245 "Topological sort incomplete!"); 5246 assert(AllNodes.front().getOpcode() == ISD::EntryToken && 5247 "First node in topological sort is not the entry token!"); 5248 assert(AllNodes.front().getNodeId() == 0 && 5249 "First node in topological sort has non-zero id!"); 5250 assert(AllNodes.front().getNumOperands() == 0 && 5251 "First node in topological sort has operands!"); 5252 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 && 5253 "Last node in topologic sort has unexpected id!"); 5254 assert(AllNodes.back().use_empty() && 5255 "Last node in topologic sort has users!"); 5256 assert(DAGSize == allnodes_size() && "Node count mismatch!"); 5257 return DAGSize; 5258} 5259 5260/// AssignOrdering - Assign an order to the SDNode. 5261void SelectionDAG::AssignOrdering(const SDNode *SD, unsigned Order) { 5262 assert(SD && "Trying to assign an order to a null node!"); 5263 Ordering->add(SD, Order); 5264} 5265 5266/// GetOrdering - Get the order for the SDNode. 5267unsigned SelectionDAG::GetOrdering(const SDNode *SD) const { 5268 assert(SD && "Trying to get the order of a null node!"); 5269 return Ordering->getOrder(SD); 5270} 5271 5272/// AssignDbgInfo - Assign debug info to the SDNode. 5273void SelectionDAG::AssignDbgInfo(SDNode* SD, SDDbgValue* db) { 5274 assert(SD && "Trying to assign dbg info to a null node!"); 5275 DbgInfo->add(SD, db); 5276 SD->setHasDebugValue(true); 5277} 5278 5279/// RememberDbgInfo - Remember debug info which is not assigned to an SDNode. 5280void SelectionDAG::RememberDbgInfo(SDDbgValue* db) { 5281 DbgInfo->add(db); 5282} 5283 5284/// GetDbgInfo - Get the debug info, if any, for the SDNode. 5285SDDbgValue* SelectionDAG::GetDbgInfo(const SDNode *SD) { 5286 assert(SD && "Trying to get the order of a null node!"); 5287 if (SD->getHasDebugValue()) 5288 return DbgInfo->getSDDbgValue(SD); 5289 return 0; 5290} 5291 5292//===----------------------------------------------------------------------===// 5293// SDNode Class 5294//===----------------------------------------------------------------------===// 5295 5296HandleSDNode::~HandleSDNode() { 5297 DropOperands(); 5298} 5299 5300GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, const GlobalValue *GA, 5301 EVT VT, int64_t o, unsigned char TF) 5302 : SDNode(Opc, DebugLoc::getUnknownLoc(), getSDVTList(VT)), 5303 Offset(o), TargetFlags(TF) { 5304 TheGlobal = const_cast<GlobalValue*>(GA); 5305} 5306 5307MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, EVT memvt, 5308 MachineMemOperand *mmo) 5309 : SDNode(Opc, dl, VTs), MemoryVT(memvt), MMO(mmo) { 5310 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(), 5311 MMO->isNonTemporal()); 5312 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!"); 5313 assert(isNonTemporal() == MMO->isNonTemporal() && 5314 "Non-temporal encoding error!"); 5315 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!"); 5316} 5317 5318MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, 5319 const SDValue *Ops, unsigned NumOps, EVT memvt, 5320 MachineMemOperand *mmo) 5321 : SDNode(Opc, dl, VTs, Ops, NumOps), 5322 MemoryVT(memvt), MMO(mmo) { 5323 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(), 5324 MMO->isNonTemporal()); 5325 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!"); 5326 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!"); 5327} 5328 5329/// Profile - Gather unique data for the node. 5330/// 5331void SDNode::Profile(FoldingSetNodeID &ID) const { 5332 AddNodeIDNode(ID, this); 5333} 5334 5335namespace { 5336 struct EVTArray { 5337 std::vector<EVT> VTs; 5338 5339 EVTArray() { 5340 VTs.reserve(MVT::LAST_VALUETYPE); 5341 for (unsigned i = 0; i < MVT::LAST_VALUETYPE; ++i) 5342 VTs.push_back(MVT((MVT::SimpleValueType)i)); 5343 } 5344 }; 5345} 5346 5347static ManagedStatic<std::set<EVT, EVT::compareRawBits> > EVTs; 5348static ManagedStatic<EVTArray> SimpleVTArray; 5349static ManagedStatic<sys::SmartMutex<true> > VTMutex; 5350 5351/// getValueTypeList - Return a pointer to the specified value type. 5352/// 5353const EVT *SDNode::getValueTypeList(EVT VT) { 5354 if (VT.isExtended()) { 5355 sys::SmartScopedLock<true> Lock(*VTMutex); 5356 return &(*EVTs->insert(VT).first); 5357 } else { 5358 return &SimpleVTArray->VTs[VT.getSimpleVT().SimpleTy]; 5359 } 5360} 5361 5362/// hasNUsesOfValue - Return true if there are exactly NUSES uses of the 5363/// indicated value. This method ignores uses of other values defined by this 5364/// operation. 5365bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const { 5366 assert(Value < getNumValues() && "Bad value!"); 5367 5368 // TODO: Only iterate over uses of a given value of the node 5369 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) { 5370 if (UI.getUse().getResNo() == Value) { 5371 if (NUses == 0) 5372 return false; 5373 --NUses; 5374 } 5375 } 5376 5377 // Found exactly the right number of uses? 5378 return NUses == 0; 5379} 5380 5381 5382/// hasAnyUseOfValue - Return true if there are any use of the indicated 5383/// value. This method ignores uses of other values defined by this operation. 5384bool SDNode::hasAnyUseOfValue(unsigned Value) const { 5385 assert(Value < getNumValues() && "Bad value!"); 5386 5387 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) 5388 if (UI.getUse().getResNo() == Value) 5389 return true; 5390 5391 return false; 5392} 5393 5394 5395/// isOnlyUserOf - Return true if this node is the only use of N. 5396/// 5397bool SDNode::isOnlyUserOf(SDNode *N) const { 5398 bool Seen = false; 5399 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) { 5400 SDNode *User = *I; 5401 if (User == this) 5402 Seen = true; 5403 else 5404 return false; 5405 } 5406 5407 return Seen; 5408} 5409 5410/// isOperand - Return true if this node is an operand of N. 5411/// 5412bool SDValue::isOperandOf(SDNode *N) const { 5413 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 5414 if (*this == N->getOperand(i)) 5415 return true; 5416 return false; 5417} 5418 5419bool SDNode::isOperandOf(SDNode *N) const { 5420 for (unsigned i = 0, e = N->NumOperands; i != e; ++i) 5421 if (this == N->OperandList[i].getNode()) 5422 return true; 5423 return false; 5424} 5425 5426/// reachesChainWithoutSideEffects - Return true if this operand (which must 5427/// be a chain) reaches the specified operand without crossing any 5428/// side-effecting instructions. In practice, this looks through token 5429/// factors and non-volatile loads. In order to remain efficient, this only 5430/// looks a couple of nodes in, it does not do an exhaustive search. 5431bool SDValue::reachesChainWithoutSideEffects(SDValue Dest, 5432 unsigned Depth) const { 5433 if (*this == Dest) return true; 5434 5435 // Don't search too deeply, we just want to be able to see through 5436 // TokenFactor's etc. 5437 if (Depth == 0) return false; 5438 5439 // If this is a token factor, all inputs to the TF happen in parallel. If any 5440 // of the operands of the TF reach dest, then we can do the xform. 5441 if (getOpcode() == ISD::TokenFactor) { 5442 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) 5443 if (getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1)) 5444 return true; 5445 return false; 5446 } 5447 5448 // Loads don't have side effects, look through them. 5449 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) { 5450 if (!Ld->isVolatile()) 5451 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1); 5452 } 5453 return false; 5454} 5455 5456/// isPredecessorOf - Return true if this node is a predecessor of N. This node 5457/// is either an operand of N or it can be reached by traversing up the operands. 5458/// NOTE: this is an expensive method. Use it carefully. 5459bool SDNode::isPredecessorOf(SDNode *N) const { 5460 SmallPtrSet<SDNode *, 32> Visited; 5461 SmallVector<SDNode *, 16> Worklist; 5462 Worklist.push_back(N); 5463 5464 do { 5465 N = Worklist.pop_back_val(); 5466 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 5467 SDNode *Op = N->getOperand(i).getNode(); 5468 if (Op == this) 5469 return true; 5470 if (Visited.insert(Op)) 5471 Worklist.push_back(Op); 5472 } 5473 } while (!Worklist.empty()); 5474 5475 return false; 5476} 5477 5478uint64_t SDNode::getConstantOperandVal(unsigned Num) const { 5479 assert(Num < NumOperands && "Invalid child # of SDNode!"); 5480 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue(); 5481} 5482 5483std::string SDNode::getOperationName(const SelectionDAG *G) const { 5484 switch (getOpcode()) { 5485 default: 5486 if (getOpcode() < ISD::BUILTIN_OP_END) 5487 return "<<Unknown DAG Node>>"; 5488 if (isMachineOpcode()) { 5489 if (G) 5490 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo()) 5491 if (getMachineOpcode() < TII->getNumOpcodes()) 5492 return TII->get(getMachineOpcode()).getName(); 5493 return "<<Unknown Machine Node #" + utostr(getOpcode()) + ">>"; 5494 } 5495 if (G) { 5496 const TargetLowering &TLI = G->getTargetLoweringInfo(); 5497 const char *Name = TLI.getTargetNodeName(getOpcode()); 5498 if (Name) return Name; 5499 return "<<Unknown Target Node #" + utostr(getOpcode()) + ">>"; 5500 } 5501 return "<<Unknown Node #" + utostr(getOpcode()) + ">>"; 5502 5503#ifndef NDEBUG 5504 case ISD::DELETED_NODE: 5505 return "<<Deleted Node!>>"; 5506#endif 5507 case ISD::PREFETCH: return "Prefetch"; 5508 case ISD::MEMBARRIER: return "MemBarrier"; 5509 case ISD::ATOMIC_CMP_SWAP: return "AtomicCmpSwap"; 5510 case ISD::ATOMIC_SWAP: return "AtomicSwap"; 5511 case ISD::ATOMIC_LOAD_ADD: return "AtomicLoadAdd"; 5512 case ISD::ATOMIC_LOAD_SUB: return "AtomicLoadSub"; 5513 case ISD::ATOMIC_LOAD_AND: return "AtomicLoadAnd"; 5514 case ISD::ATOMIC_LOAD_OR: return "AtomicLoadOr"; 5515 case ISD::ATOMIC_LOAD_XOR: return "AtomicLoadXor"; 5516 case ISD::ATOMIC_LOAD_NAND: return "AtomicLoadNand"; 5517 case ISD::ATOMIC_LOAD_MIN: return "AtomicLoadMin"; 5518 case ISD::ATOMIC_LOAD_MAX: return "AtomicLoadMax"; 5519 case ISD::ATOMIC_LOAD_UMIN: return "AtomicLoadUMin"; 5520 case ISD::ATOMIC_LOAD_UMAX: return "AtomicLoadUMax"; 5521 case ISD::PCMARKER: return "PCMarker"; 5522 case ISD::READCYCLECOUNTER: return "ReadCycleCounter"; 5523 case ISD::SRCVALUE: return "SrcValue"; 5524 case ISD::EntryToken: return "EntryToken"; 5525 case ISD::TokenFactor: return "TokenFactor"; 5526 case ISD::AssertSext: return "AssertSext"; 5527 case ISD::AssertZext: return "AssertZext"; 5528 5529 case ISD::BasicBlock: return "BasicBlock"; 5530 case ISD::VALUETYPE: return "ValueType"; 5531 case ISD::Register: return "Register"; 5532 5533 case ISD::Constant: return "Constant"; 5534 case ISD::ConstantFP: return "ConstantFP"; 5535 case ISD::GlobalAddress: return "GlobalAddress"; 5536 case ISD::GlobalTLSAddress: return "GlobalTLSAddress"; 5537 case ISD::FrameIndex: return "FrameIndex"; 5538 case ISD::JumpTable: return "JumpTable"; 5539 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE"; 5540 case ISD::RETURNADDR: return "RETURNADDR"; 5541 case ISD::FRAMEADDR: return "FRAMEADDR"; 5542 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET"; 5543 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR"; 5544 case ISD::LSDAADDR: return "LSDAADDR"; 5545 case ISD::EHSELECTION: return "EHSELECTION"; 5546 case ISD::EH_RETURN: return "EH_RETURN"; 5547 case ISD::ConstantPool: return "ConstantPool"; 5548 case ISD::ExternalSymbol: return "ExternalSymbol"; 5549 case ISD::BlockAddress: return "BlockAddress"; 5550 case ISD::INTRINSIC_WO_CHAIN: 5551 case ISD::INTRINSIC_VOID: 5552 case ISD::INTRINSIC_W_CHAIN: { 5553 unsigned OpNo = getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 0 : 1; 5554 unsigned IID = cast<ConstantSDNode>(getOperand(OpNo))->getZExtValue(); 5555 if (IID < Intrinsic::num_intrinsics) 5556 return Intrinsic::getName((Intrinsic::ID)IID); 5557 else if (const TargetIntrinsicInfo *TII = G->getTarget().getIntrinsicInfo()) 5558 return TII->getName(IID); 5559 llvm_unreachable("Invalid intrinsic ID"); 5560 } 5561 5562 case ISD::BUILD_VECTOR: return "BUILD_VECTOR"; 5563 case ISD::TargetConstant: return "TargetConstant"; 5564 case ISD::TargetConstantFP:return "TargetConstantFP"; 5565 case ISD::TargetGlobalAddress: return "TargetGlobalAddress"; 5566 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress"; 5567 case ISD::TargetFrameIndex: return "TargetFrameIndex"; 5568 case ISD::TargetJumpTable: return "TargetJumpTable"; 5569 case ISD::TargetConstantPool: return "TargetConstantPool"; 5570 case ISD::TargetExternalSymbol: return "TargetExternalSymbol"; 5571 case ISD::TargetBlockAddress: return "TargetBlockAddress"; 5572 5573 case ISD::CopyToReg: return "CopyToReg"; 5574 case ISD::CopyFromReg: return "CopyFromReg"; 5575 case ISD::UNDEF: return "undef"; 5576 case ISD::MERGE_VALUES: return "merge_values"; 5577 case ISD::INLINEASM: return "inlineasm"; 5578 case ISD::EH_LABEL: return "eh_label"; 5579 case ISD::HANDLENODE: return "handlenode"; 5580 5581 // Unary operators 5582 case ISD::FABS: return "fabs"; 5583 case ISD::FNEG: return "fneg"; 5584 case ISD::FSQRT: return "fsqrt"; 5585 case ISD::FSIN: return "fsin"; 5586 case ISD::FCOS: return "fcos"; 5587 case ISD::FPOWI: return "fpowi"; 5588 case ISD::FPOW: return "fpow"; 5589 case ISD::FTRUNC: return "ftrunc"; 5590 case ISD::FFLOOR: return "ffloor"; 5591 case ISD::FCEIL: return "fceil"; 5592 case ISD::FRINT: return "frint"; 5593 case ISD::FNEARBYINT: return "fnearbyint"; 5594 5595 // Binary operators 5596 case ISD::ADD: return "add"; 5597 case ISD::SUB: return "sub"; 5598 case ISD::MUL: return "mul"; 5599 case ISD::MULHU: return "mulhu"; 5600 case ISD::MULHS: return "mulhs"; 5601 case ISD::SDIV: return "sdiv"; 5602 case ISD::UDIV: return "udiv"; 5603 case ISD::SREM: return "srem"; 5604 case ISD::UREM: return "urem"; 5605 case ISD::SMUL_LOHI: return "smul_lohi"; 5606 case ISD::UMUL_LOHI: return "umul_lohi"; 5607 case ISD::SDIVREM: return "sdivrem"; 5608 case ISD::UDIVREM: return "udivrem"; 5609 case ISD::AND: return "and"; 5610 case ISD::OR: return "or"; 5611 case ISD::XOR: return "xor"; 5612 case ISD::SHL: return "shl"; 5613 case ISD::SRA: return "sra"; 5614 case ISD::SRL: return "srl"; 5615 case ISD::ROTL: return "rotl"; 5616 case ISD::ROTR: return "rotr"; 5617 case ISD::FADD: return "fadd"; 5618 case ISD::FSUB: return "fsub"; 5619 case ISD::FMUL: return "fmul"; 5620 case ISD::FDIV: return "fdiv"; 5621 case ISD::FREM: return "frem"; 5622 case ISD::FCOPYSIGN: return "fcopysign"; 5623 case ISD::FGETSIGN: return "fgetsign"; 5624 5625 case ISD::SETCC: return "setcc"; 5626 case ISD::VSETCC: return "vsetcc"; 5627 case ISD::SELECT: return "select"; 5628 case ISD::SELECT_CC: return "select_cc"; 5629 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt"; 5630 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt"; 5631 case ISD::CONCAT_VECTORS: return "concat_vectors"; 5632 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector"; 5633 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector"; 5634 case ISD::VECTOR_SHUFFLE: return "vector_shuffle"; 5635 case ISD::CARRY_FALSE: return "carry_false"; 5636 case ISD::ADDC: return "addc"; 5637 case ISD::ADDE: return "adde"; 5638 case ISD::SADDO: return "saddo"; 5639 case ISD::UADDO: return "uaddo"; 5640 case ISD::SSUBO: return "ssubo"; 5641 case ISD::USUBO: return "usubo"; 5642 case ISD::SMULO: return "smulo"; 5643 case ISD::UMULO: return "umulo"; 5644 case ISD::SUBC: return "subc"; 5645 case ISD::SUBE: return "sube"; 5646 case ISD::SHL_PARTS: return "shl_parts"; 5647 case ISD::SRA_PARTS: return "sra_parts"; 5648 case ISD::SRL_PARTS: return "srl_parts"; 5649 5650 // Conversion operators. 5651 case ISD::SIGN_EXTEND: return "sign_extend"; 5652 case ISD::ZERO_EXTEND: return "zero_extend"; 5653 case ISD::ANY_EXTEND: return "any_extend"; 5654 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg"; 5655 case ISD::TRUNCATE: return "truncate"; 5656 case ISD::FP_ROUND: return "fp_round"; 5657 case ISD::FLT_ROUNDS_: return "flt_rounds"; 5658 case ISD::FP_ROUND_INREG: return "fp_round_inreg"; 5659 case ISD::FP_EXTEND: return "fp_extend"; 5660 5661 case ISD::SINT_TO_FP: return "sint_to_fp"; 5662 case ISD::UINT_TO_FP: return "uint_to_fp"; 5663 case ISD::FP_TO_SINT: return "fp_to_sint"; 5664 case ISD::FP_TO_UINT: return "fp_to_uint"; 5665 case ISD::BIT_CONVERT: return "bit_convert"; 5666 case ISD::FP16_TO_FP32: return "fp16_to_fp32"; 5667 case ISD::FP32_TO_FP16: return "fp32_to_fp16"; 5668 5669 case ISD::CONVERT_RNDSAT: { 5670 switch (cast<CvtRndSatSDNode>(this)->getCvtCode()) { 5671 default: llvm_unreachable("Unknown cvt code!"); 5672 case ISD::CVT_FF: return "cvt_ff"; 5673 case ISD::CVT_FS: return "cvt_fs"; 5674 case ISD::CVT_FU: return "cvt_fu"; 5675 case ISD::CVT_SF: return "cvt_sf"; 5676 case ISD::CVT_UF: return "cvt_uf"; 5677 case ISD::CVT_SS: return "cvt_ss"; 5678 case ISD::CVT_SU: return "cvt_su"; 5679 case ISD::CVT_US: return "cvt_us"; 5680 case ISD::CVT_UU: return "cvt_uu"; 5681 } 5682 } 5683 5684 // Control flow instructions 5685 case ISD::BR: return "br"; 5686 case ISD::BRIND: return "brind"; 5687 case ISD::BR_JT: return "br_jt"; 5688 case ISD::BRCOND: return "brcond"; 5689 case ISD::BR_CC: return "br_cc"; 5690 case ISD::CALLSEQ_START: return "callseq_start"; 5691 case ISD::CALLSEQ_END: return "callseq_end"; 5692 5693 // Other operators 5694 case ISD::LOAD: return "load"; 5695 case ISD::STORE: return "store"; 5696 case ISD::VAARG: return "vaarg"; 5697 case ISD::VACOPY: return "vacopy"; 5698 case ISD::VAEND: return "vaend"; 5699 case ISD::VASTART: return "vastart"; 5700 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc"; 5701 case ISD::EXTRACT_ELEMENT: return "extract_element"; 5702 case ISD::BUILD_PAIR: return "build_pair"; 5703 case ISD::STACKSAVE: return "stacksave"; 5704 case ISD::STACKRESTORE: return "stackrestore"; 5705 case ISD::TRAP: return "trap"; 5706 5707 // Bit manipulation 5708 case ISD::BSWAP: return "bswap"; 5709 case ISD::CTPOP: return "ctpop"; 5710 case ISD::CTTZ: return "cttz"; 5711 case ISD::CTLZ: return "ctlz"; 5712 5713 // Trampolines 5714 case ISD::TRAMPOLINE: return "trampoline"; 5715 5716 case ISD::CONDCODE: 5717 switch (cast<CondCodeSDNode>(this)->get()) { 5718 default: llvm_unreachable("Unknown setcc condition!"); 5719 case ISD::SETOEQ: return "setoeq"; 5720 case ISD::SETOGT: return "setogt"; 5721 case ISD::SETOGE: return "setoge"; 5722 case ISD::SETOLT: return "setolt"; 5723 case ISD::SETOLE: return "setole"; 5724 case ISD::SETONE: return "setone"; 5725 5726 case ISD::SETO: return "seto"; 5727 case ISD::SETUO: return "setuo"; 5728 case ISD::SETUEQ: return "setue"; 5729 case ISD::SETUGT: return "setugt"; 5730 case ISD::SETUGE: return "setuge"; 5731 case ISD::SETULT: return "setult"; 5732 case ISD::SETULE: return "setule"; 5733 case ISD::SETUNE: return "setune"; 5734 5735 case ISD::SETEQ: return "seteq"; 5736 case ISD::SETGT: return "setgt"; 5737 case ISD::SETGE: return "setge"; 5738 case ISD::SETLT: return "setlt"; 5739 case ISD::SETLE: return "setle"; 5740 case ISD::SETNE: return "setne"; 5741 } 5742 } 5743} 5744 5745const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) { 5746 switch (AM) { 5747 default: 5748 return ""; 5749 case ISD::PRE_INC: 5750 return "<pre-inc>"; 5751 case ISD::PRE_DEC: 5752 return "<pre-dec>"; 5753 case ISD::POST_INC: 5754 return "<post-inc>"; 5755 case ISD::POST_DEC: 5756 return "<post-dec>"; 5757 } 5758} 5759 5760std::string ISD::ArgFlagsTy::getArgFlagsString() { 5761 std::string S = "< "; 5762 5763 if (isZExt()) 5764 S += "zext "; 5765 if (isSExt()) 5766 S += "sext "; 5767 if (isInReg()) 5768 S += "inreg "; 5769 if (isSRet()) 5770 S += "sret "; 5771 if (isByVal()) 5772 S += "byval "; 5773 if (isNest()) 5774 S += "nest "; 5775 if (getByValAlign()) 5776 S += "byval-align:" + utostr(getByValAlign()) + " "; 5777 if (getOrigAlign()) 5778 S += "orig-align:" + utostr(getOrigAlign()) + " "; 5779 if (getByValSize()) 5780 S += "byval-size:" + utostr(getByValSize()) + " "; 5781 return S + ">"; 5782} 5783 5784void SDNode::dump() const { dump(0); } 5785void SDNode::dump(const SelectionDAG *G) const { 5786 print(dbgs(), G); 5787} 5788 5789void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const { 5790 OS << (void*)this << ": "; 5791 5792 for (unsigned i = 0, e = getNumValues(); i != e; ++i) { 5793 if (i) OS << ","; 5794 if (getValueType(i) == MVT::Other) 5795 OS << "ch"; 5796 else 5797 OS << getValueType(i).getEVTString(); 5798 } 5799 OS << " = " << getOperationName(G); 5800} 5801 5802void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const { 5803 if (const MachineSDNode *MN = dyn_cast<MachineSDNode>(this)) { 5804 if (!MN->memoperands_empty()) { 5805 OS << "<"; 5806 OS << "Mem:"; 5807 for (MachineSDNode::mmo_iterator i = MN->memoperands_begin(), 5808 e = MN->memoperands_end(); i != e; ++i) { 5809 OS << **i; 5810 if (next(i) != e) 5811 OS << " "; 5812 } 5813 OS << ">"; 5814 } 5815 } else if (const ShuffleVectorSDNode *SVN = 5816 dyn_cast<ShuffleVectorSDNode>(this)) { 5817 OS << "<"; 5818 for (unsigned i = 0, e = ValueList[0].getVectorNumElements(); i != e; ++i) { 5819 int Idx = SVN->getMaskElt(i); 5820 if (i) OS << ","; 5821 if (Idx < 0) 5822 OS << "u"; 5823 else 5824 OS << Idx; 5825 } 5826 OS << ">"; 5827 } else if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) { 5828 OS << '<' << CSDN->getAPIntValue() << '>'; 5829 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) { 5830 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle) 5831 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>'; 5832 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble) 5833 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>'; 5834 else { 5835 OS << "<APFloat("; 5836 CSDN->getValueAPF().bitcastToAPInt().dump(); 5837 OS << ")>"; 5838 } 5839 } else if (const GlobalAddressSDNode *GADN = 5840 dyn_cast<GlobalAddressSDNode>(this)) { 5841 int64_t offset = GADN->getOffset(); 5842 OS << '<'; 5843 WriteAsOperand(OS, GADN->getGlobal()); 5844 OS << '>'; 5845 if (offset > 0) 5846 OS << " + " << offset; 5847 else 5848 OS << " " << offset; 5849 if (unsigned int TF = GADN->getTargetFlags()) 5850 OS << " [TF=" << TF << ']'; 5851 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) { 5852 OS << "<" << FIDN->getIndex() << ">"; 5853 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) { 5854 OS << "<" << JTDN->getIndex() << ">"; 5855 if (unsigned int TF = JTDN->getTargetFlags()) 5856 OS << " [TF=" << TF << ']'; 5857 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){ 5858 int offset = CP->getOffset(); 5859 if (CP->isMachineConstantPoolEntry()) 5860 OS << "<" << *CP->getMachineCPVal() << ">"; 5861 else 5862 OS << "<" << *CP->getConstVal() << ">"; 5863 if (offset > 0) 5864 OS << " + " << offset; 5865 else 5866 OS << " " << offset; 5867 if (unsigned int TF = CP->getTargetFlags()) 5868 OS << " [TF=" << TF << ']'; 5869 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) { 5870 OS << "<"; 5871 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock(); 5872 if (LBB) 5873 OS << LBB->getName() << " "; 5874 OS << (const void*)BBDN->getBasicBlock() << ">"; 5875 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) { 5876 if (G && R->getReg() && 5877 TargetRegisterInfo::isPhysicalRegister(R->getReg())) { 5878 OS << " %" << G->getTarget().getRegisterInfo()->getName(R->getReg()); 5879 } else { 5880 OS << " %reg" << R->getReg(); 5881 } 5882 } else if (const ExternalSymbolSDNode *ES = 5883 dyn_cast<ExternalSymbolSDNode>(this)) { 5884 OS << "'" << ES->getSymbol() << "'"; 5885 if (unsigned int TF = ES->getTargetFlags()) 5886 OS << " [TF=" << TF << ']'; 5887 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) { 5888 if (M->getValue()) 5889 OS << "<" << M->getValue() << ">"; 5890 else 5891 OS << "<null>"; 5892 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) { 5893 OS << ":" << N->getVT().getEVTString(); 5894 } 5895 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) { 5896 OS << "<" << *LD->getMemOperand(); 5897 5898 bool doExt = true; 5899 switch (LD->getExtensionType()) { 5900 default: doExt = false; break; 5901 case ISD::EXTLOAD: OS << ", anyext"; break; 5902 case ISD::SEXTLOAD: OS << ", sext"; break; 5903 case ISD::ZEXTLOAD: OS << ", zext"; break; 5904 } 5905 if (doExt) 5906 OS << " from " << LD->getMemoryVT().getEVTString(); 5907 5908 const char *AM = getIndexedModeName(LD->getAddressingMode()); 5909 if (*AM) 5910 OS << ", " << AM; 5911 5912 OS << ">"; 5913 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) { 5914 OS << "<" << *ST->getMemOperand(); 5915 5916 if (ST->isTruncatingStore()) 5917 OS << ", trunc to " << ST->getMemoryVT().getEVTString(); 5918 5919 const char *AM = getIndexedModeName(ST->getAddressingMode()); 5920 if (*AM) 5921 OS << ", " << AM; 5922 5923 OS << ">"; 5924 } else if (const MemSDNode* M = dyn_cast<MemSDNode>(this)) { 5925 OS << "<" << *M->getMemOperand() << ">"; 5926 } else if (const BlockAddressSDNode *BA = 5927 dyn_cast<BlockAddressSDNode>(this)) { 5928 OS << "<"; 5929 WriteAsOperand(OS, BA->getBlockAddress()->getFunction(), false); 5930 OS << ", "; 5931 WriteAsOperand(OS, BA->getBlockAddress()->getBasicBlock(), false); 5932 OS << ">"; 5933 if (unsigned int TF = BA->getTargetFlags()) 5934 OS << " [TF=" << TF << ']'; 5935 } 5936 5937 if (G) 5938 if (unsigned Order = G->GetOrdering(this)) 5939 OS << " [ORD=" << Order << ']'; 5940 5941 if (getNodeId() != -1) 5942 OS << " [ID=" << getNodeId() << ']'; 5943} 5944 5945void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const { 5946 print_types(OS, G); 5947 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) { 5948 if (i) OS << ", "; else OS << " "; 5949 OS << (void*)getOperand(i).getNode(); 5950 if (unsigned RN = getOperand(i).getResNo()) 5951 OS << ":" << RN; 5952 } 5953 print_details(OS, G); 5954} 5955 5956static void printrWithDepthHelper(raw_ostream &OS, const SDNode *N, 5957 const SelectionDAG *G, unsigned depth, 5958 unsigned indent) 5959{ 5960 if (depth == 0) 5961 return; 5962 5963 OS.indent(indent); 5964 5965 N->print(OS, G); 5966 5967 if (depth < 1) 5968 return; 5969 5970 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 5971 OS << '\n'; 5972 printrWithDepthHelper(OS, N->getOperand(i).getNode(), G, depth-1, indent+2); 5973 } 5974} 5975 5976void SDNode::printrWithDepth(raw_ostream &OS, const SelectionDAG *G, 5977 unsigned depth) const { 5978 printrWithDepthHelper(OS, this, G, depth, 0); 5979} 5980 5981void SDNode::printrFull(raw_ostream &OS, const SelectionDAG *G) const { 5982 // Don't print impossibly deep things. 5983 printrWithDepth(OS, G, 100); 5984} 5985 5986void SDNode::dumprWithDepth(const SelectionDAG *G, unsigned depth) const { 5987 printrWithDepth(dbgs(), G, depth); 5988} 5989 5990void SDNode::dumprFull(const SelectionDAG *G) const { 5991 // Don't print impossibly deep things. 5992 dumprWithDepth(G, 100); 5993} 5994 5995static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) { 5996 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 5997 if (N->getOperand(i).getNode()->hasOneUse()) 5998 DumpNodes(N->getOperand(i).getNode(), indent+2, G); 5999 else 6000 dbgs() << "\n" << std::string(indent+2, ' ') 6001 << (void*)N->getOperand(i).getNode() << ": <multiple use>"; 6002 6003 6004 dbgs() << "\n"; 6005 dbgs().indent(indent); 6006 N->dump(G); 6007} 6008 6009SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) { 6010 assert(N->getNumValues() == 1 && 6011 "Can't unroll a vector with multiple results!"); 6012 6013 EVT VT = N->getValueType(0); 6014 unsigned NE = VT.getVectorNumElements(); 6015 EVT EltVT = VT.getVectorElementType(); 6016 DebugLoc dl = N->getDebugLoc(); 6017 6018 SmallVector<SDValue, 8> Scalars; 6019 SmallVector<SDValue, 4> Operands(N->getNumOperands()); 6020 6021 // If ResNE is 0, fully unroll the vector op. 6022 if (ResNE == 0) 6023 ResNE = NE; 6024 else if (NE > ResNE) 6025 NE = ResNE; 6026 6027 unsigned i; 6028 for (i= 0; i != NE; ++i) { 6029 for (unsigned j = 0; j != N->getNumOperands(); ++j) { 6030 SDValue Operand = N->getOperand(j); 6031 EVT OperandVT = Operand.getValueType(); 6032 if (OperandVT.isVector()) { 6033 // A vector operand; extract a single element. 6034 EVT OperandEltVT = OperandVT.getVectorElementType(); 6035 Operands[j] = getNode(ISD::EXTRACT_VECTOR_ELT, dl, 6036 OperandEltVT, 6037 Operand, 6038 getConstant(i, MVT::i32)); 6039 } else { 6040 // A scalar operand; just use it as is. 6041 Operands[j] = Operand; 6042 } 6043 } 6044 6045 switch (N->getOpcode()) { 6046 default: 6047 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, 6048 &Operands[0], Operands.size())); 6049 break; 6050 case ISD::SHL: 6051 case ISD::SRA: 6052 case ISD::SRL: 6053 case ISD::ROTL: 6054 case ISD::ROTR: 6055 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands[0], 6056 getShiftAmountOperand(Operands[1]))); 6057 break; 6058 case ISD::SIGN_EXTEND_INREG: 6059 case ISD::FP_ROUND_INREG: { 6060 EVT ExtVT = cast<VTSDNode>(Operands[1])->getVT().getVectorElementType(); 6061 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, 6062 Operands[0], 6063 getValueType(ExtVT))); 6064 } 6065 } 6066 } 6067 6068 for (; i < ResNE; ++i) 6069 Scalars.push_back(getUNDEF(EltVT)); 6070 6071 return getNode(ISD::BUILD_VECTOR, dl, 6072 EVT::getVectorVT(*getContext(), EltVT, ResNE), 6073 &Scalars[0], Scalars.size()); 6074} 6075 6076 6077/// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a 6078/// location that is 'Dist' units away from the location that the 'Base' load 6079/// is loading from. 6080bool SelectionDAG::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base, 6081 unsigned Bytes, int Dist) const { 6082 if (LD->getChain() != Base->getChain()) 6083 return false; 6084 EVT VT = LD->getValueType(0); 6085 if (VT.getSizeInBits() / 8 != Bytes) 6086 return false; 6087 6088 SDValue Loc = LD->getOperand(1); 6089 SDValue BaseLoc = Base->getOperand(1); 6090 if (Loc.getOpcode() == ISD::FrameIndex) { 6091 if (BaseLoc.getOpcode() != ISD::FrameIndex) 6092 return false; 6093 const MachineFrameInfo *MFI = getMachineFunction().getFrameInfo(); 6094 int FI = cast<FrameIndexSDNode>(Loc)->getIndex(); 6095 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex(); 6096 int FS = MFI->getObjectSize(FI); 6097 int BFS = MFI->getObjectSize(BFI); 6098 if (FS != BFS || FS != (int)Bytes) return false; 6099 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes); 6100 } 6101 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) { 6102 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1)); 6103 if (V && (V->getSExtValue() == Dist*Bytes)) 6104 return true; 6105 } 6106 6107 GlobalValue *GV1 = NULL; 6108 GlobalValue *GV2 = NULL; 6109 int64_t Offset1 = 0; 6110 int64_t Offset2 = 0; 6111 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1); 6112 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2); 6113 if (isGA1 && isGA2 && GV1 == GV2) 6114 return Offset1 == (Offset2 + Dist*Bytes); 6115 return false; 6116} 6117 6118 6119/// InferPtrAlignment - Infer alignment of a load / store address. Return 0 if 6120/// it cannot be inferred. 6121unsigned SelectionDAG::InferPtrAlignment(SDValue Ptr) const { 6122 // If this is a GlobalAddress + cst, return the alignment. 6123 GlobalValue *GV; 6124 int64_t GVOffset = 0; 6125 if (TLI.isGAPlusOffset(Ptr.getNode(), GV, GVOffset)) 6126 return MinAlign(GV->getAlignment(), GVOffset); 6127 6128 // If this is a direct reference to a stack slot, use information about the 6129 // stack slot's alignment. 6130 int FrameIdx = 1 << 31; 6131 int64_t FrameOffset = 0; 6132 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) { 6133 FrameIdx = FI->getIndex(); 6134 } else if (Ptr.getOpcode() == ISD::ADD && 6135 isa<ConstantSDNode>(Ptr.getOperand(1)) && 6136 isa<FrameIndexSDNode>(Ptr.getOperand(0))) { 6137 FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex(); 6138 FrameOffset = Ptr.getConstantOperandVal(1); 6139 } 6140 6141 if (FrameIdx != (1 << 31)) { 6142 // FIXME: Handle FI+CST. 6143 const MachineFrameInfo &MFI = *getMachineFunction().getFrameInfo(); 6144 unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx), 6145 FrameOffset); 6146 if (MFI.isFixedObjectIndex(FrameIdx)) { 6147 int64_t ObjectOffset = MFI.getObjectOffset(FrameIdx) + FrameOffset; 6148 6149 // The alignment of the frame index can be determined from its offset from 6150 // the incoming frame position. If the frame object is at offset 32 and 6151 // the stack is guaranteed to be 16-byte aligned, then we know that the 6152 // object is 16-byte aligned. 6153 unsigned StackAlign = getTarget().getFrameInfo()->getStackAlignment(); 6154 unsigned Align = MinAlign(ObjectOffset, StackAlign); 6155 6156 // Finally, the frame object itself may have a known alignment. Factor 6157 // the alignment + offset into a new alignment. For example, if we know 6158 // the FI is 8 byte aligned, but the pointer is 4 off, we really have a 6159 // 4-byte alignment of the resultant pointer. Likewise align 4 + 4-byte 6160 // offset = 4-byte alignment, align 4 + 1-byte offset = align 1, etc. 6161 return std::max(Align, FIInfoAlign); 6162 } 6163 return FIInfoAlign; 6164 } 6165 6166 return 0; 6167} 6168 6169void SelectionDAG::dump() const { 6170 dbgs() << "SelectionDAG has " << AllNodes.size() << " nodes:"; 6171 6172 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end(); 6173 I != E; ++I) { 6174 const SDNode *N = I; 6175 if (!N->hasOneUse() && N != getRoot().getNode()) 6176 DumpNodes(N, 2, this); 6177 } 6178 6179 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this); 6180 6181 dbgs() << "\n\n"; 6182} 6183 6184void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const { 6185 print_types(OS, G); 6186 print_details(OS, G); 6187} 6188 6189typedef SmallPtrSet<const SDNode *, 128> VisitedSDNodeSet; 6190static void DumpNodesr(raw_ostream &OS, const SDNode *N, unsigned indent, 6191 const SelectionDAG *G, VisitedSDNodeSet &once) { 6192 if (!once.insert(N)) // If we've been here before, return now. 6193 return; 6194 6195 // Dump the current SDNode, but don't end the line yet. 6196 OS << std::string(indent, ' '); 6197 N->printr(OS, G); 6198 6199 // Having printed this SDNode, walk the children: 6200 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 6201 const SDNode *child = N->getOperand(i).getNode(); 6202 6203 if (i) OS << ","; 6204 OS << " "; 6205 6206 if (child->getNumOperands() == 0) { 6207 // This child has no grandchildren; print it inline right here. 6208 child->printr(OS, G); 6209 once.insert(child); 6210 } else { // Just the address. FIXME: also print the child's opcode. 6211 OS << (void*)child; 6212 if (unsigned RN = N->getOperand(i).getResNo()) 6213 OS << ":" << RN; 6214 } 6215 } 6216 6217 OS << "\n"; 6218 6219 // Dump children that have grandchildren on their own line(s). 6220 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 6221 const SDNode *child = N->getOperand(i).getNode(); 6222 DumpNodesr(OS, child, indent+2, G, once); 6223 } 6224} 6225 6226void SDNode::dumpr() const { 6227 VisitedSDNodeSet once; 6228 DumpNodesr(dbgs(), this, 0, 0, once); 6229} 6230 6231void SDNode::dumpr(const SelectionDAG *G) const { 6232 VisitedSDNodeSet once; 6233 DumpNodesr(dbgs(), this, 0, G, once); 6234} 6235 6236 6237// getAddressSpace - Return the address space this GlobalAddress belongs to. 6238unsigned GlobalAddressSDNode::getAddressSpace() const { 6239 return getGlobal()->getType()->getAddressSpace(); 6240} 6241 6242 6243const Type *ConstantPoolSDNode::getType() const { 6244 if (isMachineConstantPoolEntry()) 6245 return Val.MachineCPVal->getType(); 6246 return Val.ConstVal->getType(); 6247} 6248 6249bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue, 6250 APInt &SplatUndef, 6251 unsigned &SplatBitSize, 6252 bool &HasAnyUndefs, 6253 unsigned MinSplatBits, 6254 bool isBigEndian) { 6255 EVT VT = getValueType(0); 6256 assert(VT.isVector() && "Expected a vector type"); 6257 unsigned sz = VT.getSizeInBits(); 6258 if (MinSplatBits > sz) 6259 return false; 6260 6261 SplatValue = APInt(sz, 0); 6262 SplatUndef = APInt(sz, 0); 6263 6264 // Get the bits. Bits with undefined values (when the corresponding element 6265 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared 6266 // in SplatValue. If any of the values are not constant, give up and return 6267 // false. 6268 unsigned int nOps = getNumOperands(); 6269 assert(nOps > 0 && "isConstantSplat has 0-size build vector"); 6270 unsigned EltBitSize = VT.getVectorElementType().getSizeInBits(); 6271 6272 for (unsigned j = 0; j < nOps; ++j) { 6273 unsigned i = isBigEndian ? nOps-1-j : j; 6274 SDValue OpVal = getOperand(i); 6275 unsigned BitPos = j * EltBitSize; 6276 6277 if (OpVal.getOpcode() == ISD::UNDEF) 6278 SplatUndef |= APInt::getBitsSet(sz, BitPos, BitPos + EltBitSize); 6279 else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) 6280 SplatValue |= (APInt(CN->getAPIntValue()).zextOrTrunc(EltBitSize). 6281 zextOrTrunc(sz) << BitPos); 6282 else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) 6283 SplatValue |= CN->getValueAPF().bitcastToAPInt().zextOrTrunc(sz) <<BitPos; 6284 else 6285 return false; 6286 } 6287 6288 // The build_vector is all constants or undefs. Find the smallest element 6289 // size that splats the vector. 6290 6291 HasAnyUndefs = (SplatUndef != 0); 6292 while (sz > 8) { 6293 6294 unsigned HalfSize = sz / 2; 6295 APInt HighValue = APInt(SplatValue).lshr(HalfSize).trunc(HalfSize); 6296 APInt LowValue = APInt(SplatValue).trunc(HalfSize); 6297 APInt HighUndef = APInt(SplatUndef).lshr(HalfSize).trunc(HalfSize); 6298 APInt LowUndef = APInt(SplatUndef).trunc(HalfSize); 6299 6300 // If the two halves do not match (ignoring undef bits), stop here. 6301 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) || 6302 MinSplatBits > HalfSize) 6303 break; 6304 6305 SplatValue = HighValue | LowValue; 6306 SplatUndef = HighUndef & LowUndef; 6307 6308 sz = HalfSize; 6309 } 6310 6311 SplatBitSize = sz; 6312 return true; 6313} 6314 6315bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) { 6316 // Find the first non-undef value in the shuffle mask. 6317 unsigned i, e; 6318 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i) 6319 /* search */; 6320 6321 assert(i != e && "VECTOR_SHUFFLE node with all undef indices!"); 6322 6323 // Make sure all remaining elements are either undef or the same as the first 6324 // non-undef value. 6325 for (int Idx = Mask[i]; i != e; ++i) 6326 if (Mask[i] >= 0 && Mask[i] != Idx) 6327 return false; 6328 return true; 6329} 6330 6331#ifdef XDEBUG 6332static void checkForCyclesHelper(const SDNode *N, 6333 SmallPtrSet<const SDNode*, 32> &Visited, 6334 SmallPtrSet<const SDNode*, 32> &Checked) { 6335 // If this node has already been checked, don't check it again. 6336 if (Checked.count(N)) 6337 return; 6338 6339 // If a node has already been visited on this depth-first walk, reject it as 6340 // a cycle. 6341 if (!Visited.insert(N)) { 6342 dbgs() << "Offending node:\n"; 6343 N->dumprFull(); 6344 errs() << "Detected cycle in SelectionDAG\n"; 6345 abort(); 6346 } 6347 6348 for(unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 6349 checkForCyclesHelper(N->getOperand(i).getNode(), Visited, Checked); 6350 6351 Checked.insert(N); 6352 Visited.erase(N); 6353} 6354#endif 6355 6356void llvm::checkForCycles(const llvm::SDNode *N) { 6357#ifdef XDEBUG 6358 assert(N && "Checking nonexistant SDNode"); 6359 SmallPtrSet<const SDNode*, 32> visited; 6360 SmallPtrSet<const SDNode*, 32> checked; 6361 checkForCyclesHelper(N, visited, checked); 6362#endif 6363} 6364 6365void llvm::checkForCycles(const llvm::SelectionDAG *DAG) { 6366 checkForCycles(DAG->getRoot().getNode()); 6367} 6368