SelectionDAG.cpp revision 2e141d744ee1354f49a99d9da49dbd9570e4269d
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#include "llvm/CodeGen/SelectionDAG.h" 14#include "llvm/Constants.h" 15#include "llvm/Analysis/ValueTracking.h" 16#include "llvm/Function.h" 17#include "llvm/GlobalAlias.h" 18#include "llvm/GlobalVariable.h" 19#include "llvm/Intrinsics.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/Assembly/Writer.h" 22#include "llvm/CallingConv.h" 23#include "llvm/CodeGen/MachineBasicBlock.h" 24#include "llvm/CodeGen/MachineConstantPool.h" 25#include "llvm/CodeGen/MachineFrameInfo.h" 26#include "llvm/CodeGen/MachineModuleInfo.h" 27#include "llvm/CodeGen/PseudoSourceValue.h" 28#include "llvm/Target/TargetRegisterInfo.h" 29#include "llvm/Target/TargetData.h" 30#include "llvm/Target/TargetFrameInfo.h" 31#include "llvm/Target/TargetLowering.h" 32#include "llvm/Target/TargetOptions.h" 33#include "llvm/Target/TargetInstrInfo.h" 34#include "llvm/Target/TargetIntrinsicInfo.h" 35#include "llvm/Target/TargetMachine.h" 36#include "llvm/Support/CommandLine.h" 37#include "llvm/Support/ErrorHandling.h" 38#include "llvm/Support/ManagedStatic.h" 39#include "llvm/Support/MathExtras.h" 40#include "llvm/Support/raw_ostream.h" 41#include "llvm/System/Mutex.h" 42#include "llvm/ADT/SetVector.h" 43#include "llvm/ADT/SmallPtrSet.h" 44#include "llvm/ADT/SmallSet.h" 45#include "llvm/ADT/SmallVector.h" 46#include "llvm/ADT/StringExtras.h" 47#include <algorithm> 48#include <cmath> 49using namespace llvm; 50 51/// makeVTList - Return an instance of the SDVTList struct initialized with the 52/// specified members. 53static SDVTList makeVTList(const EVT *VTs, unsigned NumVTs) { 54 SDVTList Res = {VTs, NumVTs}; 55 return Res; 56} 57 58static const fltSemantics *EVTToAPFloatSemantics(EVT VT) { 59 switch (VT.getSimpleVT().SimpleTy) { 60 default: llvm_unreachable("Unknown FP format"); 61 case MVT::f32: return &APFloat::IEEEsingle; 62 case MVT::f64: return &APFloat::IEEEdouble; 63 case MVT::f80: return &APFloat::x87DoubleExtended; 64 case MVT::f128: return &APFloat::IEEEquad; 65 case MVT::ppcf128: return &APFloat::PPCDoubleDouble; 66 } 67} 68 69SelectionDAG::DAGUpdateListener::~DAGUpdateListener() {} 70 71//===----------------------------------------------------------------------===// 72// ConstantFPSDNode Class 73//===----------------------------------------------------------------------===// 74 75/// isExactlyValue - We don't rely on operator== working on double values, as 76/// it returns true for things that are clearly not equal, like -0.0 and 0.0. 77/// As such, this method can be used to do an exact bit-for-bit comparison of 78/// two floating point values. 79bool ConstantFPSDNode::isExactlyValue(const APFloat& V) const { 80 return getValueAPF().bitwiseIsEqual(V); 81} 82 83bool ConstantFPSDNode::isValueValidForType(EVT VT, 84 const APFloat& Val) { 85 assert(VT.isFloatingPoint() && "Can only convert between FP types"); 86 87 // PPC long double cannot be converted to any other type. 88 if (VT == MVT::ppcf128 || 89 &Val.getSemantics() == &APFloat::PPCDoubleDouble) 90 return false; 91 92 // convert modifies in place, so make a copy. 93 APFloat Val2 = APFloat(Val); 94 bool losesInfo; 95 (void) Val2.convert(*EVTToAPFloatSemantics(VT), APFloat::rmNearestTiesToEven, 96 &losesInfo); 97 return !losesInfo; 98} 99 100//===----------------------------------------------------------------------===// 101// ISD Namespace 102//===----------------------------------------------------------------------===// 103 104/// isBuildVectorAllOnes - Return true if the specified node is a 105/// BUILD_VECTOR where all of the elements are ~0 or undef. 106bool ISD::isBuildVectorAllOnes(const SDNode *N) { 107 // Look through a bit convert. 108 if (N->getOpcode() == ISD::BIT_CONVERT) 109 N = N->getOperand(0).getNode(); 110 111 if (N->getOpcode() != ISD::BUILD_VECTOR) return false; 112 113 unsigned i = 0, e = N->getNumOperands(); 114 115 // Skip over all of the undef values. 116 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF) 117 ++i; 118 119 // Do not accept an all-undef vector. 120 if (i == e) return false; 121 122 // Do not accept build_vectors that aren't all constants or which have non-~0 123 // elements. 124 SDValue NotZero = N->getOperand(i); 125 if (isa<ConstantSDNode>(NotZero)) { 126 if (!cast<ConstantSDNode>(NotZero)->isAllOnesValue()) 127 return false; 128 } else if (isa<ConstantFPSDNode>(NotZero)) { 129 if (!cast<ConstantFPSDNode>(NotZero)->getValueAPF(). 130 bitcastToAPInt().isAllOnesValue()) 131 return false; 132 } else 133 return false; 134 135 // Okay, we have at least one ~0 value, check to see if the rest match or are 136 // undefs. 137 for (++i; i != e; ++i) 138 if (N->getOperand(i) != NotZero && 139 N->getOperand(i).getOpcode() != ISD::UNDEF) 140 return false; 141 return true; 142} 143 144 145/// isBuildVectorAllZeros - Return true if the specified node is a 146/// BUILD_VECTOR where all of the elements are 0 or undef. 147bool ISD::isBuildVectorAllZeros(const SDNode *N) { 148 // Look through a bit convert. 149 if (N->getOpcode() == ISD::BIT_CONVERT) 150 N = N->getOperand(0).getNode(); 151 152 if (N->getOpcode() != ISD::BUILD_VECTOR) return false; 153 154 unsigned i = 0, e = N->getNumOperands(); 155 156 // Skip over all of the undef values. 157 while (i != e && N->getOperand(i).getOpcode() == ISD::UNDEF) 158 ++i; 159 160 // Do not accept an all-undef vector. 161 if (i == e) return false; 162 163 // Do not accept build_vectors that aren't all constants or which have non-0 164 // elements. 165 SDValue Zero = N->getOperand(i); 166 if (isa<ConstantSDNode>(Zero)) { 167 if (!cast<ConstantSDNode>(Zero)->isNullValue()) 168 return false; 169 } else if (isa<ConstantFPSDNode>(Zero)) { 170 if (!cast<ConstantFPSDNode>(Zero)->getValueAPF().isPosZero()) 171 return false; 172 } else 173 return false; 174 175 // Okay, we have at least one 0 value, check to see if the rest match or are 176 // undefs. 177 for (++i; i != e; ++i) 178 if (N->getOperand(i) != Zero && 179 N->getOperand(i).getOpcode() != ISD::UNDEF) 180 return false; 181 return true; 182} 183 184/// isScalarToVector - Return true if the specified node is a 185/// ISD::SCALAR_TO_VECTOR node or a BUILD_VECTOR node where only the low 186/// element is not an undef. 187bool ISD::isScalarToVector(const SDNode *N) { 188 if (N->getOpcode() == ISD::SCALAR_TO_VECTOR) 189 return true; 190 191 if (N->getOpcode() != ISD::BUILD_VECTOR) 192 return false; 193 if (N->getOperand(0).getOpcode() == ISD::UNDEF) 194 return false; 195 unsigned NumElems = N->getNumOperands(); 196 for (unsigned i = 1; i < NumElems; ++i) { 197 SDValue V = N->getOperand(i); 198 if (V.getOpcode() != ISD::UNDEF) 199 return false; 200 } 201 return true; 202} 203 204/// getSetCCSwappedOperands - Return the operation corresponding to (Y op X) 205/// when given the operation for (X op Y). 206ISD::CondCode ISD::getSetCCSwappedOperands(ISD::CondCode Operation) { 207 // To perform this operation, we just need to swap the L and G bits of the 208 // operation. 209 unsigned OldL = (Operation >> 2) & 1; 210 unsigned OldG = (Operation >> 1) & 1; 211 return ISD::CondCode((Operation & ~6) | // Keep the N, U, E bits 212 (OldL << 1) | // New G bit 213 (OldG << 2)); // New L bit. 214} 215 216/// getSetCCInverse - Return the operation corresponding to !(X op Y), where 217/// 'op' is a valid SetCC operation. 218ISD::CondCode ISD::getSetCCInverse(ISD::CondCode Op, bool isInteger) { 219 unsigned Operation = Op; 220 if (isInteger) 221 Operation ^= 7; // Flip L, G, E bits, but not U. 222 else 223 Operation ^= 15; // Flip all of the condition bits. 224 225 if (Operation > ISD::SETTRUE2) 226 Operation &= ~8; // Don't let N and U bits get set. 227 228 return ISD::CondCode(Operation); 229} 230 231 232/// isSignedOp - For an integer comparison, return 1 if the comparison is a 233/// signed operation and 2 if the result is an unsigned comparison. Return zero 234/// if the operation does not depend on the sign of the input (setne and seteq). 235static int isSignedOp(ISD::CondCode Opcode) { 236 switch (Opcode) { 237 default: llvm_unreachable("Illegal integer setcc operation!"); 238 case ISD::SETEQ: 239 case ISD::SETNE: return 0; 240 case ISD::SETLT: 241 case ISD::SETLE: 242 case ISD::SETGT: 243 case ISD::SETGE: return 1; 244 case ISD::SETULT: 245 case ISD::SETULE: 246 case ISD::SETUGT: 247 case ISD::SETUGE: return 2; 248 } 249} 250 251/// getSetCCOrOperation - Return the result of a logical OR between different 252/// comparisons of identical values: ((X op1 Y) | (X op2 Y)). This function 253/// returns SETCC_INVALID if it is not possible to represent the resultant 254/// comparison. 255ISD::CondCode ISD::getSetCCOrOperation(ISD::CondCode Op1, ISD::CondCode Op2, 256 bool isInteger) { 257 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3) 258 // Cannot fold a signed integer setcc with an unsigned integer setcc. 259 return ISD::SETCC_INVALID; 260 261 unsigned Op = Op1 | Op2; // Combine all of the condition bits. 262 263 // If the N and U bits get set then the resultant comparison DOES suddenly 264 // care about orderedness, and is true when ordered. 265 if (Op > ISD::SETTRUE2) 266 Op &= ~16; // Clear the U bit if the N bit is set. 267 268 // Canonicalize illegal integer setcc's. 269 if (isInteger && Op == ISD::SETUNE) // e.g. SETUGT | SETULT 270 Op = ISD::SETNE; 271 272 return ISD::CondCode(Op); 273} 274 275/// getSetCCAndOperation - Return the result of a logical AND between different 276/// comparisons of identical values: ((X op1 Y) & (X op2 Y)). This 277/// function returns zero if it is not possible to represent the resultant 278/// comparison. 279ISD::CondCode ISD::getSetCCAndOperation(ISD::CondCode Op1, ISD::CondCode Op2, 280 bool isInteger) { 281 if (isInteger && (isSignedOp(Op1) | isSignedOp(Op2)) == 3) 282 // Cannot fold a signed setcc with an unsigned setcc. 283 return ISD::SETCC_INVALID; 284 285 // Combine all of the condition bits. 286 ISD::CondCode Result = ISD::CondCode(Op1 & Op2); 287 288 // Canonicalize illegal integer setcc's. 289 if (isInteger) { 290 switch (Result) { 291 default: break; 292 case ISD::SETUO : Result = ISD::SETFALSE; break; // SETUGT & SETULT 293 case ISD::SETOEQ: // SETEQ & SETU[LG]E 294 case ISD::SETUEQ: Result = ISD::SETEQ ; break; // SETUGE & SETULE 295 case ISD::SETOLT: Result = ISD::SETULT ; break; // SETULT & SETNE 296 case ISD::SETOGT: Result = ISD::SETUGT ; break; // SETUGT & SETNE 297 } 298 } 299 300 return Result; 301} 302 303const TargetMachine &SelectionDAG::getTarget() const { 304 return MF->getTarget(); 305} 306 307//===----------------------------------------------------------------------===// 308// SDNode Profile Support 309//===----------------------------------------------------------------------===// 310 311/// AddNodeIDOpcode - Add the node opcode to the NodeID data. 312/// 313static void AddNodeIDOpcode(FoldingSetNodeID &ID, unsigned OpC) { 314 ID.AddInteger(OpC); 315} 316 317/// AddNodeIDValueTypes - Value type lists are intern'd so we can represent them 318/// solely with their pointer. 319static void AddNodeIDValueTypes(FoldingSetNodeID &ID, SDVTList VTList) { 320 ID.AddPointer(VTList.VTs); 321} 322 323/// AddNodeIDOperands - Various routines for adding operands to the NodeID data. 324/// 325static void AddNodeIDOperands(FoldingSetNodeID &ID, 326 const SDValue *Ops, unsigned NumOps) { 327 for (; NumOps; --NumOps, ++Ops) { 328 ID.AddPointer(Ops->getNode()); 329 ID.AddInteger(Ops->getResNo()); 330 } 331} 332 333/// AddNodeIDOperands - Various routines for adding operands to the NodeID data. 334/// 335static void AddNodeIDOperands(FoldingSetNodeID &ID, 336 const SDUse *Ops, unsigned NumOps) { 337 for (; NumOps; --NumOps, ++Ops) { 338 ID.AddPointer(Ops->getNode()); 339 ID.AddInteger(Ops->getResNo()); 340 } 341} 342 343static void AddNodeIDNode(FoldingSetNodeID &ID, 344 unsigned short OpC, SDVTList VTList, 345 const SDValue *OpList, unsigned N) { 346 AddNodeIDOpcode(ID, OpC); 347 AddNodeIDValueTypes(ID, VTList); 348 AddNodeIDOperands(ID, OpList, N); 349} 350 351/// AddNodeIDCustom - If this is an SDNode with special info, add this info to 352/// the NodeID data. 353static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) { 354 switch (N->getOpcode()) { 355 case ISD::TargetExternalSymbol: 356 case ISD::ExternalSymbol: 357 llvm_unreachable("Should only be used on nodes with operands"); 358 default: break; // Normal nodes don't need extra info. 359 case ISD::TargetConstant: 360 case ISD::Constant: 361 ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue()); 362 break; 363 case ISD::TargetConstantFP: 364 case ISD::ConstantFP: { 365 ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue()); 366 break; 367 } 368 case ISD::TargetGlobalAddress: 369 case ISD::GlobalAddress: 370 case ISD::TargetGlobalTLSAddress: 371 case ISD::GlobalTLSAddress: { 372 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(N); 373 ID.AddPointer(GA->getGlobal()); 374 ID.AddInteger(GA->getOffset()); 375 ID.AddInteger(GA->getTargetFlags()); 376 break; 377 } 378 case ISD::BasicBlock: 379 ID.AddPointer(cast<BasicBlockSDNode>(N)->getBasicBlock()); 380 break; 381 case ISD::Register: 382 ID.AddInteger(cast<RegisterSDNode>(N)->getReg()); 383 break; 384 385 case ISD::SRCVALUE: 386 ID.AddPointer(cast<SrcValueSDNode>(N)->getValue()); 387 break; 388 case ISD::FrameIndex: 389 case ISD::TargetFrameIndex: 390 ID.AddInteger(cast<FrameIndexSDNode>(N)->getIndex()); 391 break; 392 case ISD::JumpTable: 393 case ISD::TargetJumpTable: 394 ID.AddInteger(cast<JumpTableSDNode>(N)->getIndex()); 395 ID.AddInteger(cast<JumpTableSDNode>(N)->getTargetFlags()); 396 break; 397 case ISD::ConstantPool: 398 case ISD::TargetConstantPool: { 399 const ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(N); 400 ID.AddInteger(CP->getAlignment()); 401 ID.AddInteger(CP->getOffset()); 402 if (CP->isMachineConstantPoolEntry()) 403 CP->getMachineCPVal()->AddSelectionDAGCSEId(ID); 404 else 405 ID.AddPointer(CP->getConstVal()); 406 ID.AddInteger(CP->getTargetFlags()); 407 break; 408 } 409 case ISD::LOAD: { 410 const LoadSDNode *LD = cast<LoadSDNode>(N); 411 ID.AddInteger(LD->getMemoryVT().getRawBits()); 412 ID.AddInteger(LD->getRawSubclassData()); 413 break; 414 } 415 case ISD::STORE: { 416 const StoreSDNode *ST = cast<StoreSDNode>(N); 417 ID.AddInteger(ST->getMemoryVT().getRawBits()); 418 ID.AddInteger(ST->getRawSubclassData()); 419 break; 420 } 421 case ISD::ATOMIC_CMP_SWAP: 422 case ISD::ATOMIC_SWAP: 423 case ISD::ATOMIC_LOAD_ADD: 424 case ISD::ATOMIC_LOAD_SUB: 425 case ISD::ATOMIC_LOAD_AND: 426 case ISD::ATOMIC_LOAD_OR: 427 case ISD::ATOMIC_LOAD_XOR: 428 case ISD::ATOMIC_LOAD_NAND: 429 case ISD::ATOMIC_LOAD_MIN: 430 case ISD::ATOMIC_LOAD_MAX: 431 case ISD::ATOMIC_LOAD_UMIN: 432 case ISD::ATOMIC_LOAD_UMAX: { 433 const AtomicSDNode *AT = cast<AtomicSDNode>(N); 434 ID.AddInteger(AT->getMemoryVT().getRawBits()); 435 ID.AddInteger(AT->getRawSubclassData()); 436 break; 437 } 438 case ISD::VECTOR_SHUFFLE: { 439 const ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N); 440 for (unsigned i = 0, e = N->getValueType(0).getVectorNumElements(); 441 i != e; ++i) 442 ID.AddInteger(SVN->getMaskElt(i)); 443 break; 444 } 445 case ISD::TargetBlockAddress: 446 case ISD::BlockAddress: { 447 ID.AddPointer(cast<BlockAddressSDNode>(N)->getBlockAddress()); 448 ID.AddInteger(cast<BlockAddressSDNode>(N)->getTargetFlags()); 449 break; 450 } 451 } // end switch (N->getOpcode()) 452} 453 454/// AddNodeIDNode - Generic routine for adding a nodes info to the NodeID 455/// data. 456static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) { 457 AddNodeIDOpcode(ID, N->getOpcode()); 458 // Add the return value info. 459 AddNodeIDValueTypes(ID, N->getVTList()); 460 // Add the operand info. 461 AddNodeIDOperands(ID, N->op_begin(), N->getNumOperands()); 462 463 // Handle SDNode leafs with special info. 464 AddNodeIDCustom(ID, N); 465} 466 467/// encodeMemSDNodeFlags - Generic routine for computing a value for use in 468/// the CSE map that carries volatility, indexing mode, and 469/// extension/truncation information. 470/// 471static inline unsigned 472encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile) { 473 assert((ConvType & 3) == ConvType && 474 "ConvType may not require more than 2 bits!"); 475 assert((AM & 7) == AM && 476 "AM may not require more than 3 bits!"); 477 return ConvType | 478 (AM << 2) | 479 (isVolatile << 5); 480} 481 482//===----------------------------------------------------------------------===// 483// SelectionDAG Class 484//===----------------------------------------------------------------------===// 485 486/// doNotCSE - Return true if CSE should not be performed for this node. 487static bool doNotCSE(SDNode *N) { 488 if (N->getValueType(0) == MVT::Flag) 489 return true; // Never CSE anything that produces a flag. 490 491 switch (N->getOpcode()) { 492 default: break; 493 case ISD::HANDLENODE: 494 case ISD::EH_LABEL: 495 return true; // Never CSE these nodes. 496 } 497 498 // Check that remaining values produced are not flags. 499 for (unsigned i = 1, e = N->getNumValues(); i != e; ++i) 500 if (N->getValueType(i) == MVT::Flag) 501 return true; // Never CSE anything that produces a flag. 502 503 return false; 504} 505 506/// RemoveDeadNodes - This method deletes all unreachable nodes in the 507/// SelectionDAG. 508void SelectionDAG::RemoveDeadNodes() { 509 // Create a dummy node (which is not added to allnodes), that adds a reference 510 // to the root node, preventing it from being deleted. 511 HandleSDNode Dummy(getRoot()); 512 513 SmallVector<SDNode*, 128> DeadNodes; 514 515 // Add all obviously-dead nodes to the DeadNodes worklist. 516 for (allnodes_iterator I = allnodes_begin(), E = allnodes_end(); I != E; ++I) 517 if (I->use_empty()) 518 DeadNodes.push_back(I); 519 520 RemoveDeadNodes(DeadNodes); 521 522 // If the root changed (e.g. it was a dead load, update the root). 523 setRoot(Dummy.getValue()); 524} 525 526/// RemoveDeadNodes - This method deletes the unreachable nodes in the 527/// given list, and any nodes that become unreachable as a result. 528void SelectionDAG::RemoveDeadNodes(SmallVectorImpl<SDNode *> &DeadNodes, 529 DAGUpdateListener *UpdateListener) { 530 531 // Process the worklist, deleting the nodes and adding their uses to the 532 // worklist. 533 while (!DeadNodes.empty()) { 534 SDNode *N = DeadNodes.pop_back_val(); 535 536 if (UpdateListener) 537 UpdateListener->NodeDeleted(N, 0); 538 539 // Take the node out of the appropriate CSE map. 540 RemoveNodeFromCSEMaps(N); 541 542 // Next, brutally remove the operand list. This is safe to do, as there are 543 // no cycles in the graph. 544 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) { 545 SDUse &Use = *I++; 546 SDNode *Operand = Use.getNode(); 547 Use.set(SDValue()); 548 549 // Now that we removed this operand, see if there are no uses of it left. 550 if (Operand->use_empty()) 551 DeadNodes.push_back(Operand); 552 } 553 554 DeallocateNode(N); 555 } 556} 557 558void SelectionDAG::RemoveDeadNode(SDNode *N, DAGUpdateListener *UpdateListener){ 559 SmallVector<SDNode*, 16> DeadNodes(1, N); 560 RemoveDeadNodes(DeadNodes, UpdateListener); 561} 562 563void SelectionDAG::DeleteNode(SDNode *N) { 564 // First take this out of the appropriate CSE map. 565 RemoveNodeFromCSEMaps(N); 566 567 // Finally, remove uses due to operands of this node, remove from the 568 // AllNodes list, and delete the node. 569 DeleteNodeNotInCSEMaps(N); 570} 571 572void SelectionDAG::DeleteNodeNotInCSEMaps(SDNode *N) { 573 assert(N != AllNodes.begin() && "Cannot delete the entry node!"); 574 assert(N->use_empty() && "Cannot delete a node that is not dead!"); 575 576 // Drop all of the operands and decrement used node's use counts. 577 N->DropOperands(); 578 579 DeallocateNode(N); 580} 581 582void SelectionDAG::DeallocateNode(SDNode *N) { 583 if (N->OperandsNeedDelete) 584 delete[] N->OperandList; 585 586 // Set the opcode to DELETED_NODE to help catch bugs when node 587 // memory is reallocated. 588 N->NodeType = ISD::DELETED_NODE; 589 590 NodeAllocator.Deallocate(AllNodes.remove(N)); 591} 592 593/// RemoveNodeFromCSEMaps - Take the specified node out of the CSE map that 594/// correspond to it. This is useful when we're about to delete or repurpose 595/// the node. We don't want future request for structurally identical nodes 596/// to return N anymore. 597bool SelectionDAG::RemoveNodeFromCSEMaps(SDNode *N) { 598 bool Erased = false; 599 switch (N->getOpcode()) { 600 case ISD::EntryToken: 601 llvm_unreachable("EntryToken should not be in CSEMaps!"); 602 return false; 603 case ISD::HANDLENODE: return false; // noop. 604 case ISD::CONDCODE: 605 assert(CondCodeNodes[cast<CondCodeSDNode>(N)->get()] && 606 "Cond code doesn't exist!"); 607 Erased = CondCodeNodes[cast<CondCodeSDNode>(N)->get()] != 0; 608 CondCodeNodes[cast<CondCodeSDNode>(N)->get()] = 0; 609 break; 610 case ISD::ExternalSymbol: 611 Erased = ExternalSymbols.erase(cast<ExternalSymbolSDNode>(N)->getSymbol()); 612 break; 613 case ISD::TargetExternalSymbol: { 614 ExternalSymbolSDNode *ESN = cast<ExternalSymbolSDNode>(N); 615 Erased = TargetExternalSymbols.erase( 616 std::pair<std::string,unsigned char>(ESN->getSymbol(), 617 ESN->getTargetFlags())); 618 break; 619 } 620 case ISD::VALUETYPE: { 621 EVT VT = cast<VTSDNode>(N)->getVT(); 622 if (VT.isExtended()) { 623 Erased = ExtendedValueTypeNodes.erase(VT); 624 } else { 625 Erased = ValueTypeNodes[VT.getSimpleVT().SimpleTy] != 0; 626 ValueTypeNodes[VT.getSimpleVT().SimpleTy] = 0; 627 } 628 break; 629 } 630 default: 631 // Remove it from the CSE Map. 632 Erased = CSEMap.RemoveNode(N); 633 break; 634 } 635#ifndef NDEBUG 636 // Verify that the node was actually in one of the CSE maps, unless it has a 637 // flag result (which cannot be CSE'd) or is one of the special cases that are 638 // not subject to CSE. 639 if (!Erased && N->getValueType(N->getNumValues()-1) != MVT::Flag && 640 !N->isMachineOpcode() && !doNotCSE(N)) { 641 N->dump(this); 642 errs() << "\n"; 643 llvm_unreachable("Node is not in map!"); 644 } 645#endif 646 return Erased; 647} 648 649/// AddModifiedNodeToCSEMaps - The specified node has been removed from the CSE 650/// maps and modified in place. Add it back to the CSE maps, unless an identical 651/// node already exists, in which case transfer all its users to the existing 652/// node. This transfer can potentially trigger recursive merging. 653/// 654void 655SelectionDAG::AddModifiedNodeToCSEMaps(SDNode *N, 656 DAGUpdateListener *UpdateListener) { 657 // For node types that aren't CSE'd, just act as if no identical node 658 // already exists. 659 if (!doNotCSE(N)) { 660 SDNode *Existing = CSEMap.GetOrInsertNode(N); 661 if (Existing != N) { 662 // If there was already an existing matching node, use ReplaceAllUsesWith 663 // to replace the dead one with the existing one. This can cause 664 // recursive merging of other unrelated nodes down the line. 665 ReplaceAllUsesWith(N, Existing, UpdateListener); 666 667 // N is now dead. Inform the listener if it exists and delete it. 668 if (UpdateListener) 669 UpdateListener->NodeDeleted(N, Existing); 670 DeleteNodeNotInCSEMaps(N); 671 return; 672 } 673 } 674 675 // If the node doesn't already exist, we updated it. Inform a listener if 676 // it exists. 677 if (UpdateListener) 678 UpdateListener->NodeUpdated(N); 679} 680 681/// FindModifiedNodeSlot - Find a slot for the specified node if its operands 682/// were replaced with those specified. If this node is never memoized, 683/// return null, otherwise return a pointer to the slot it would take. If a 684/// node already exists with these operands, the slot will be non-null. 685SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, SDValue Op, 686 void *&InsertPos) { 687 if (doNotCSE(N)) 688 return 0; 689 690 SDValue Ops[] = { Op }; 691 FoldingSetNodeID ID; 692 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 1); 693 AddNodeIDCustom(ID, N); 694 return CSEMap.FindNodeOrInsertPos(ID, InsertPos); 695} 696 697/// FindModifiedNodeSlot - Find a slot for the specified node if its operands 698/// were replaced with those specified. If this node is never memoized, 699/// return null, otherwise return a pointer to the slot it would take. If a 700/// node already exists with these operands, the slot will be non-null. 701SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, 702 SDValue Op1, SDValue Op2, 703 void *&InsertPos) { 704 if (doNotCSE(N)) 705 return 0; 706 707 SDValue Ops[] = { Op1, Op2 }; 708 FoldingSetNodeID ID; 709 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, 2); 710 AddNodeIDCustom(ID, N); 711 return CSEMap.FindNodeOrInsertPos(ID, InsertPos); 712} 713 714 715/// FindModifiedNodeSlot - Find a slot for the specified node if its operands 716/// were replaced with those specified. If this node is never memoized, 717/// return null, otherwise return a pointer to the slot it would take. If a 718/// node already exists with these operands, the slot will be non-null. 719SDNode *SelectionDAG::FindModifiedNodeSlot(SDNode *N, 720 const SDValue *Ops,unsigned NumOps, 721 void *&InsertPos) { 722 if (doNotCSE(N)) 723 return 0; 724 725 FoldingSetNodeID ID; 726 AddNodeIDNode(ID, N->getOpcode(), N->getVTList(), Ops, NumOps); 727 AddNodeIDCustom(ID, N); 728 return CSEMap.FindNodeOrInsertPos(ID, InsertPos); 729} 730 731/// VerifyNode - Sanity check the given node. Aborts if it is invalid. 732void SelectionDAG::VerifyNode(SDNode *N) { 733 switch (N->getOpcode()) { 734 default: 735 break; 736 case ISD::BUILD_PAIR: { 737 EVT VT = N->getValueType(0); 738 assert(N->getNumValues() == 1 && "Too many results!"); 739 assert(!VT.isVector() && (VT.isInteger() || VT.isFloatingPoint()) && 740 "Wrong return type!"); 741 assert(N->getNumOperands() == 2 && "Wrong number of operands!"); 742 assert(N->getOperand(0).getValueType() == N->getOperand(1).getValueType() && 743 "Mismatched operand types!"); 744 assert(N->getOperand(0).getValueType().isInteger() == VT.isInteger() && 745 "Wrong operand type!"); 746 assert(VT.getSizeInBits() == 2 * N->getOperand(0).getValueSizeInBits() && 747 "Wrong return type size"); 748 break; 749 } 750 case ISD::BUILD_VECTOR: { 751 assert(N->getNumValues() == 1 && "Too many results!"); 752 assert(N->getValueType(0).isVector() && "Wrong return type!"); 753 assert(N->getNumOperands() == N->getValueType(0).getVectorNumElements() && 754 "Wrong number of operands!"); 755 EVT EltVT = N->getValueType(0).getVectorElementType(); 756 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ++I) 757 assert((I->getValueType() == EltVT || 758 (EltVT.isInteger() && I->getValueType().isInteger() && 759 EltVT.bitsLE(I->getValueType()))) && 760 "Wrong operand type!"); 761 break; 762 } 763 } 764} 765 766/// getEVTAlignment - Compute the default alignment value for the 767/// given type. 768/// 769unsigned SelectionDAG::getEVTAlignment(EVT VT) const { 770 const Type *Ty = VT == MVT::iPTR ? 771 PointerType::get(Type::getInt8Ty(*getContext()), 0) : 772 VT.getTypeForEVT(*getContext()); 773 774 return TLI.getTargetData()->getABITypeAlignment(Ty); 775} 776 777// EntryNode could meaningfully have debug info if we can find it... 778SelectionDAG::SelectionDAG(TargetLowering &tli, FunctionLoweringInfo &fli) 779 : TLI(tli), FLI(fli), DW(0), 780 EntryNode(ISD::EntryToken, DebugLoc::getUnknownLoc(), 781 getVTList(MVT::Other)), Root(getEntryNode()) { 782 AllNodes.push_back(&EntryNode); 783} 784 785void SelectionDAG::init(MachineFunction &mf, MachineModuleInfo *mmi, 786 DwarfWriter *dw) { 787 MF = &mf; 788 MMI = mmi; 789 DW = dw; 790 Context = &mf.getFunction()->getContext(); 791} 792 793SelectionDAG::~SelectionDAG() { 794 allnodes_clear(); 795} 796 797void SelectionDAG::allnodes_clear() { 798 assert(&*AllNodes.begin() == &EntryNode); 799 AllNodes.remove(AllNodes.begin()); 800 while (!AllNodes.empty()) 801 DeallocateNode(AllNodes.begin()); 802} 803 804void SelectionDAG::clear() { 805 allnodes_clear(); 806 OperandAllocator.Reset(); 807 CSEMap.clear(); 808 809 ExtendedValueTypeNodes.clear(); 810 ExternalSymbols.clear(); 811 TargetExternalSymbols.clear(); 812 std::fill(CondCodeNodes.begin(), CondCodeNodes.end(), 813 static_cast<CondCodeSDNode*>(0)); 814 std::fill(ValueTypeNodes.begin(), ValueTypeNodes.end(), 815 static_cast<SDNode*>(0)); 816 817 EntryNode.UseList = 0; 818 AllNodes.push_back(&EntryNode); 819 Root = getEntryNode(); 820} 821 822SDValue SelectionDAG::getSExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) { 823 return VT.bitsGT(Op.getValueType()) ? 824 getNode(ISD::SIGN_EXTEND, DL, VT, Op) : 825 getNode(ISD::TRUNCATE, DL, VT, Op); 826} 827 828SDValue SelectionDAG::getZExtOrTrunc(SDValue Op, DebugLoc DL, EVT VT) { 829 return VT.bitsGT(Op.getValueType()) ? 830 getNode(ISD::ZERO_EXTEND, DL, VT, Op) : 831 getNode(ISD::TRUNCATE, DL, VT, Op); 832} 833 834SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, EVT VT) { 835 assert(!VT.isVector() && 836 "getZeroExtendInReg should use the vector element type instead of " 837 "the vector type!"); 838 if (Op.getValueType() == VT) return Op; 839 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits(); 840 APInt Imm = APInt::getLowBitsSet(BitWidth, 841 VT.getSizeInBits()); 842 return getNode(ISD::AND, DL, Op.getValueType(), Op, 843 getConstant(Imm, Op.getValueType())); 844} 845 846/// getNOT - Create a bitwise NOT operation as (XOR Val, -1). 847/// 848SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, EVT VT) { 849 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT; 850 SDValue NegOne = 851 getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT); 852 return getNode(ISD::XOR, DL, VT, Val, NegOne); 853} 854 855SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) { 856 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT; 857 assert((EltVT.getSizeInBits() >= 64 || 858 (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) && 859 "getConstant with a uint64_t value that doesn't fit in the type!"); 860 return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT); 861} 862 863SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) { 864 return getConstant(*ConstantInt::get(*Context, Val), VT, isT); 865} 866 867SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) { 868 assert(VT.isInteger() && "Cannot create FP integer constant!"); 869 870 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT; 871 assert(Val.getBitWidth() == EltVT.getSizeInBits() && 872 "APInt size does not match type size!"); 873 874 unsigned Opc = isT ? ISD::TargetConstant : ISD::Constant; 875 FoldingSetNodeID ID; 876 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0); 877 ID.AddPointer(&Val); 878 void *IP = 0; 879 SDNode *N = NULL; 880 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP))) 881 if (!VT.isVector()) 882 return SDValue(N, 0); 883 if (!N) { 884 N = NodeAllocator.Allocate<ConstantSDNode>(); 885 new (N) ConstantSDNode(isT, &Val, EltVT); 886 CSEMap.InsertNode(N, IP); 887 AllNodes.push_back(N); 888 } 889 890 SDValue Result(N, 0); 891 if (VT.isVector()) { 892 SmallVector<SDValue, 8> Ops; 893 Ops.assign(VT.getVectorNumElements(), Result); 894 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(), 895 VT, &Ops[0], Ops.size()); 896 } 897 return Result; 898} 899 900SDValue SelectionDAG::getIntPtrConstant(uint64_t Val, bool isTarget) { 901 return getConstant(Val, TLI.getPointerTy(), isTarget); 902} 903 904 905SDValue SelectionDAG::getConstantFP(const APFloat& V, EVT VT, bool isTarget) { 906 return getConstantFP(*ConstantFP::get(*getContext(), V), VT, isTarget); 907} 908 909SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){ 910 assert(VT.isFloatingPoint() && "Cannot create integer FP constant!"); 911 912 EVT EltVT = 913 VT.isVector() ? VT.getVectorElementType() : VT; 914 915 // Do the map lookup using the actual bit pattern for the floating point 916 // value, so that we don't have problems with 0.0 comparing equal to -0.0, and 917 // we don't have issues with SNANs. 918 unsigned Opc = isTarget ? ISD::TargetConstantFP : ISD::ConstantFP; 919 FoldingSetNodeID ID; 920 AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0); 921 ID.AddPointer(&V); 922 void *IP = 0; 923 SDNode *N = NULL; 924 if ((N = CSEMap.FindNodeOrInsertPos(ID, IP))) 925 if (!VT.isVector()) 926 return SDValue(N, 0); 927 if (!N) { 928 N = NodeAllocator.Allocate<ConstantFPSDNode>(); 929 new (N) ConstantFPSDNode(isTarget, &V, EltVT); 930 CSEMap.InsertNode(N, IP); 931 AllNodes.push_back(N); 932 } 933 934 SDValue Result(N, 0); 935 if (VT.isVector()) { 936 SmallVector<SDValue, 8> Ops; 937 Ops.assign(VT.getVectorNumElements(), Result); 938 // FIXME DebugLoc info might be appropriate here 939 Result = getNode(ISD::BUILD_VECTOR, DebugLoc::getUnknownLoc(), 940 VT, &Ops[0], Ops.size()); 941 } 942 return Result; 943} 944 945SDValue SelectionDAG::getConstantFP(double Val, EVT VT, bool isTarget) { 946 EVT EltVT = 947 VT.isVector() ? VT.getVectorElementType() : VT; 948 if (EltVT==MVT::f32) 949 return getConstantFP(APFloat((float)Val), VT, isTarget); 950 else 951 return getConstantFP(APFloat(Val), VT, isTarget); 952} 953 954SDValue SelectionDAG::getGlobalAddress(const GlobalValue *GV, 955 EVT VT, int64_t Offset, 956 bool isTargetGA, 957 unsigned char TargetFlags) { 958 assert((TargetFlags == 0 || isTargetGA) && 959 "Cannot set target flags on target-independent globals"); 960 961 // Truncate (with sign-extension) the offset value to the pointer size. 962 EVT PTy = TLI.getPointerTy(); 963 unsigned BitWidth = PTy.getSizeInBits(); 964 if (BitWidth < 64) 965 Offset = (Offset << (64 - BitWidth) >> (64 - BitWidth)); 966 967 const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV); 968 if (!GVar) { 969 // If GV is an alias then use the aliasee for determining thread-localness. 970 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV)) 971 GVar = dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false)); 972 } 973 974 unsigned Opc; 975 if (GVar && GVar->isThreadLocal()) 976 Opc = isTargetGA ? ISD::TargetGlobalTLSAddress : ISD::GlobalTLSAddress; 977 else 978 Opc = isTargetGA ? ISD::TargetGlobalAddress : ISD::GlobalAddress; 979 980 FoldingSetNodeID ID; 981 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 982 ID.AddPointer(GV); 983 ID.AddInteger(Offset); 984 ID.AddInteger(TargetFlags); 985 void *IP = 0; 986 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 987 return SDValue(E, 0); 988 SDNode *N = NodeAllocator.Allocate<GlobalAddressSDNode>(); 989 new (N) GlobalAddressSDNode(Opc, GV, VT, Offset, TargetFlags); 990 CSEMap.InsertNode(N, IP); 991 AllNodes.push_back(N); 992 return SDValue(N, 0); 993} 994 995SDValue SelectionDAG::getFrameIndex(int FI, EVT VT, bool isTarget) { 996 unsigned Opc = isTarget ? ISD::TargetFrameIndex : ISD::FrameIndex; 997 FoldingSetNodeID ID; 998 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 999 ID.AddInteger(FI); 1000 void *IP = 0; 1001 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1002 return SDValue(E, 0); 1003 SDNode *N = NodeAllocator.Allocate<FrameIndexSDNode>(); 1004 new (N) FrameIndexSDNode(FI, VT, isTarget); 1005 CSEMap.InsertNode(N, IP); 1006 AllNodes.push_back(N); 1007 return SDValue(N, 0); 1008} 1009 1010SDValue SelectionDAG::getJumpTable(int JTI, EVT VT, bool isTarget, 1011 unsigned char TargetFlags) { 1012 assert((TargetFlags == 0 || isTarget) && 1013 "Cannot set target flags on target-independent jump tables"); 1014 unsigned Opc = isTarget ? ISD::TargetJumpTable : ISD::JumpTable; 1015 FoldingSetNodeID ID; 1016 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1017 ID.AddInteger(JTI); 1018 ID.AddInteger(TargetFlags); 1019 void *IP = 0; 1020 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1021 return SDValue(E, 0); 1022 SDNode *N = NodeAllocator.Allocate<JumpTableSDNode>(); 1023 new (N) JumpTableSDNode(JTI, VT, isTarget, TargetFlags); 1024 CSEMap.InsertNode(N, IP); 1025 AllNodes.push_back(N); 1026 return SDValue(N, 0); 1027} 1028 1029SDValue SelectionDAG::getConstantPool(Constant *C, EVT VT, 1030 unsigned Alignment, int Offset, 1031 bool isTarget, 1032 unsigned char TargetFlags) { 1033 assert((TargetFlags == 0 || isTarget) && 1034 "Cannot set target flags on target-independent globals"); 1035 if (Alignment == 0) 1036 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType()); 1037 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool; 1038 FoldingSetNodeID ID; 1039 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1040 ID.AddInteger(Alignment); 1041 ID.AddInteger(Offset); 1042 ID.AddPointer(C); 1043 ID.AddInteger(TargetFlags); 1044 void *IP = 0; 1045 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1046 return SDValue(E, 0); 1047 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>(); 1048 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags); 1049 CSEMap.InsertNode(N, IP); 1050 AllNodes.push_back(N); 1051 return SDValue(N, 0); 1052} 1053 1054 1055SDValue SelectionDAG::getConstantPool(MachineConstantPoolValue *C, EVT VT, 1056 unsigned Alignment, int Offset, 1057 bool isTarget, 1058 unsigned char TargetFlags) { 1059 assert((TargetFlags == 0 || isTarget) && 1060 "Cannot set target flags on target-independent globals"); 1061 if (Alignment == 0) 1062 Alignment = TLI.getTargetData()->getPrefTypeAlignment(C->getType()); 1063 unsigned Opc = isTarget ? ISD::TargetConstantPool : ISD::ConstantPool; 1064 FoldingSetNodeID ID; 1065 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1066 ID.AddInteger(Alignment); 1067 ID.AddInteger(Offset); 1068 C->AddSelectionDAGCSEId(ID); 1069 ID.AddInteger(TargetFlags); 1070 void *IP = 0; 1071 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1072 return SDValue(E, 0); 1073 SDNode *N = NodeAllocator.Allocate<ConstantPoolSDNode>(); 1074 new (N) ConstantPoolSDNode(isTarget, C, VT, Offset, Alignment, TargetFlags); 1075 CSEMap.InsertNode(N, IP); 1076 AllNodes.push_back(N); 1077 return SDValue(N, 0); 1078} 1079 1080SDValue SelectionDAG::getBasicBlock(MachineBasicBlock *MBB) { 1081 FoldingSetNodeID ID; 1082 AddNodeIDNode(ID, ISD::BasicBlock, getVTList(MVT::Other), 0, 0); 1083 ID.AddPointer(MBB); 1084 void *IP = 0; 1085 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1086 return SDValue(E, 0); 1087 SDNode *N = NodeAllocator.Allocate<BasicBlockSDNode>(); 1088 new (N) BasicBlockSDNode(MBB); 1089 CSEMap.InsertNode(N, IP); 1090 AllNodes.push_back(N); 1091 return SDValue(N, 0); 1092} 1093 1094SDValue SelectionDAG::getValueType(EVT VT) { 1095 if (VT.isSimple() && (unsigned)VT.getSimpleVT().SimpleTy >= 1096 ValueTypeNodes.size()) 1097 ValueTypeNodes.resize(VT.getSimpleVT().SimpleTy+1); 1098 1099 SDNode *&N = VT.isExtended() ? 1100 ExtendedValueTypeNodes[VT] : ValueTypeNodes[VT.getSimpleVT().SimpleTy]; 1101 1102 if (N) return SDValue(N, 0); 1103 N = NodeAllocator.Allocate<VTSDNode>(); 1104 new (N) VTSDNode(VT); 1105 AllNodes.push_back(N); 1106 return SDValue(N, 0); 1107} 1108 1109SDValue SelectionDAG::getExternalSymbol(const char *Sym, EVT VT) { 1110 SDNode *&N = ExternalSymbols[Sym]; 1111 if (N) return SDValue(N, 0); 1112 N = NodeAllocator.Allocate<ExternalSymbolSDNode>(); 1113 new (N) ExternalSymbolSDNode(false, Sym, 0, VT); 1114 AllNodes.push_back(N); 1115 return SDValue(N, 0); 1116} 1117 1118SDValue SelectionDAG::getTargetExternalSymbol(const char *Sym, EVT VT, 1119 unsigned char TargetFlags) { 1120 SDNode *&N = 1121 TargetExternalSymbols[std::pair<std::string,unsigned char>(Sym, 1122 TargetFlags)]; 1123 if (N) return SDValue(N, 0); 1124 N = NodeAllocator.Allocate<ExternalSymbolSDNode>(); 1125 new (N) ExternalSymbolSDNode(true, Sym, TargetFlags, VT); 1126 AllNodes.push_back(N); 1127 return SDValue(N, 0); 1128} 1129 1130SDValue SelectionDAG::getCondCode(ISD::CondCode Cond) { 1131 if ((unsigned)Cond >= CondCodeNodes.size()) 1132 CondCodeNodes.resize(Cond+1); 1133 1134 if (CondCodeNodes[Cond] == 0) { 1135 CondCodeSDNode *N = NodeAllocator.Allocate<CondCodeSDNode>(); 1136 new (N) CondCodeSDNode(Cond); 1137 CondCodeNodes[Cond] = N; 1138 AllNodes.push_back(N); 1139 } 1140 return SDValue(CondCodeNodes[Cond], 0); 1141} 1142 1143// commuteShuffle - swaps the values of N1 and N2, and swaps all indices in 1144// the shuffle mask M that point at N1 to point at N2, and indices that point 1145// N2 to point at N1. 1146static void commuteShuffle(SDValue &N1, SDValue &N2, SmallVectorImpl<int> &M) { 1147 std::swap(N1, N2); 1148 int NElts = M.size(); 1149 for (int i = 0; i != NElts; ++i) { 1150 if (M[i] >= NElts) 1151 M[i] -= NElts; 1152 else if (M[i] >= 0) 1153 M[i] += NElts; 1154 } 1155} 1156 1157SDValue SelectionDAG::getVectorShuffle(EVT VT, DebugLoc dl, SDValue N1, 1158 SDValue N2, const int *Mask) { 1159 assert(N1.getValueType() == N2.getValueType() && "Invalid VECTOR_SHUFFLE"); 1160 assert(VT.isVector() && N1.getValueType().isVector() && 1161 "Vector Shuffle VTs must be a vectors"); 1162 assert(VT.getVectorElementType() == N1.getValueType().getVectorElementType() 1163 && "Vector Shuffle VTs must have same element type"); 1164 1165 // Canonicalize shuffle undef, undef -> undef 1166 if (N1.getOpcode() == ISD::UNDEF && N2.getOpcode() == ISD::UNDEF) 1167 return getUNDEF(VT); 1168 1169 // Validate that all indices in Mask are within the range of the elements 1170 // input to the shuffle. 1171 unsigned NElts = VT.getVectorNumElements(); 1172 SmallVector<int, 8> MaskVec; 1173 for (unsigned i = 0; i != NElts; ++i) { 1174 assert(Mask[i] < (int)(NElts * 2) && "Index out of range"); 1175 MaskVec.push_back(Mask[i]); 1176 } 1177 1178 // Canonicalize shuffle v, v -> v, undef 1179 if (N1 == N2) { 1180 N2 = getUNDEF(VT); 1181 for (unsigned i = 0; i != NElts; ++i) 1182 if (MaskVec[i] >= (int)NElts) MaskVec[i] -= NElts; 1183 } 1184 1185 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask. 1186 if (N1.getOpcode() == ISD::UNDEF) 1187 commuteShuffle(N1, N2, MaskVec); 1188 1189 // Canonicalize all index into lhs, -> shuffle lhs, undef 1190 // Canonicalize all index into rhs, -> shuffle rhs, undef 1191 bool AllLHS = true, AllRHS = true; 1192 bool N2Undef = N2.getOpcode() == ISD::UNDEF; 1193 for (unsigned i = 0; i != NElts; ++i) { 1194 if (MaskVec[i] >= (int)NElts) { 1195 if (N2Undef) 1196 MaskVec[i] = -1; 1197 else 1198 AllLHS = false; 1199 } else if (MaskVec[i] >= 0) { 1200 AllRHS = false; 1201 } 1202 } 1203 if (AllLHS && AllRHS) 1204 return getUNDEF(VT); 1205 if (AllLHS && !N2Undef) 1206 N2 = getUNDEF(VT); 1207 if (AllRHS) { 1208 N1 = getUNDEF(VT); 1209 commuteShuffle(N1, N2, MaskVec); 1210 } 1211 1212 // If Identity shuffle, or all shuffle in to undef, return that node. 1213 bool AllUndef = true; 1214 bool Identity = true; 1215 for (unsigned i = 0; i != NElts; ++i) { 1216 if (MaskVec[i] >= 0 && MaskVec[i] != (int)i) Identity = false; 1217 if (MaskVec[i] >= 0) AllUndef = false; 1218 } 1219 if (Identity && NElts == N1.getValueType().getVectorNumElements()) 1220 return N1; 1221 if (AllUndef) 1222 return getUNDEF(VT); 1223 1224 FoldingSetNodeID ID; 1225 SDValue Ops[2] = { N1, N2 }; 1226 AddNodeIDNode(ID, ISD::VECTOR_SHUFFLE, getVTList(VT), Ops, 2); 1227 for (unsigned i = 0; i != NElts; ++i) 1228 ID.AddInteger(MaskVec[i]); 1229 1230 void* IP = 0; 1231 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1232 return SDValue(E, 0); 1233 1234 // Allocate the mask array for the node out of the BumpPtrAllocator, since 1235 // SDNode doesn't have access to it. This memory will be "leaked" when 1236 // the node is deallocated, but recovered when the NodeAllocator is released. 1237 int *MaskAlloc = OperandAllocator.Allocate<int>(NElts); 1238 memcpy(MaskAlloc, &MaskVec[0], NElts * sizeof(int)); 1239 1240 ShuffleVectorSDNode *N = NodeAllocator.Allocate<ShuffleVectorSDNode>(); 1241 new (N) ShuffleVectorSDNode(VT, dl, N1, N2, MaskAlloc); 1242 CSEMap.InsertNode(N, IP); 1243 AllNodes.push_back(N); 1244 return SDValue(N, 0); 1245} 1246 1247SDValue SelectionDAG::getConvertRndSat(EVT VT, DebugLoc dl, 1248 SDValue Val, SDValue DTy, 1249 SDValue STy, SDValue Rnd, SDValue Sat, 1250 ISD::CvtCode Code) { 1251 // If the src and dest types are the same and the conversion is between 1252 // integer types of the same sign or two floats, no conversion is necessary. 1253 if (DTy == STy && 1254 (Code == ISD::CVT_UU || Code == ISD::CVT_SS || Code == ISD::CVT_FF)) 1255 return Val; 1256 1257 FoldingSetNodeID ID; 1258 SDValue Ops[] = { Val, DTy, STy, Rnd, Sat }; 1259 AddNodeIDNode(ID, ISD::CONVERT_RNDSAT, getVTList(VT), &Ops[0], 5); 1260 void* IP = 0; 1261 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1262 return SDValue(E, 0); 1263 CvtRndSatSDNode *N = NodeAllocator.Allocate<CvtRndSatSDNode>(); 1264 new (N) CvtRndSatSDNode(VT, dl, Ops, 5, Code); 1265 CSEMap.InsertNode(N, IP); 1266 AllNodes.push_back(N); 1267 return SDValue(N, 0); 1268} 1269 1270SDValue SelectionDAG::getRegister(unsigned RegNo, EVT VT) { 1271 FoldingSetNodeID ID; 1272 AddNodeIDNode(ID, ISD::Register, getVTList(VT), 0, 0); 1273 ID.AddInteger(RegNo); 1274 void *IP = 0; 1275 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1276 return SDValue(E, 0); 1277 SDNode *N = NodeAllocator.Allocate<RegisterSDNode>(); 1278 new (N) RegisterSDNode(RegNo, VT); 1279 CSEMap.InsertNode(N, IP); 1280 AllNodes.push_back(N); 1281 return SDValue(N, 0); 1282} 1283 1284SDValue SelectionDAG::getLabel(unsigned Opcode, DebugLoc dl, 1285 SDValue Root, 1286 unsigned LabelID) { 1287 FoldingSetNodeID ID; 1288 SDValue Ops[] = { Root }; 1289 AddNodeIDNode(ID, Opcode, getVTList(MVT::Other), &Ops[0], 1); 1290 ID.AddInteger(LabelID); 1291 void *IP = 0; 1292 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1293 return SDValue(E, 0); 1294 SDNode *N = NodeAllocator.Allocate<LabelSDNode>(); 1295 new (N) LabelSDNode(Opcode, dl, Root, LabelID); 1296 CSEMap.InsertNode(N, IP); 1297 AllNodes.push_back(N); 1298 return SDValue(N, 0); 1299} 1300 1301SDValue SelectionDAG::getBlockAddress(BlockAddress *BA, EVT VT, 1302 bool isTarget, 1303 unsigned char TargetFlags) { 1304 unsigned Opc = isTarget ? ISD::TargetBlockAddress : ISD::BlockAddress; 1305 1306 FoldingSetNodeID ID; 1307 AddNodeIDNode(ID, Opc, getVTList(VT), 0, 0); 1308 ID.AddPointer(BA); 1309 ID.AddInteger(TargetFlags); 1310 void *IP = 0; 1311 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1312 return SDValue(E, 0); 1313 SDNode *N = NodeAllocator.Allocate<BlockAddressSDNode>(); 1314 new (N) BlockAddressSDNode(Opc, VT, BA, TargetFlags); 1315 CSEMap.InsertNode(N, IP); 1316 AllNodes.push_back(N); 1317 return SDValue(N, 0); 1318} 1319 1320SDValue SelectionDAG::getSrcValue(const Value *V) { 1321 assert((!V || isa<PointerType>(V->getType())) && 1322 "SrcValue is not a pointer?"); 1323 1324 FoldingSetNodeID ID; 1325 AddNodeIDNode(ID, ISD::SRCVALUE, getVTList(MVT::Other), 0, 0); 1326 ID.AddPointer(V); 1327 1328 void *IP = 0; 1329 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 1330 return SDValue(E, 0); 1331 1332 SDNode *N = NodeAllocator.Allocate<SrcValueSDNode>(); 1333 new (N) SrcValueSDNode(V); 1334 CSEMap.InsertNode(N, IP); 1335 AllNodes.push_back(N); 1336 return SDValue(N, 0); 1337} 1338 1339/// getShiftAmountOperand - Return the specified value casted to 1340/// the target's desired shift amount type. 1341SDValue SelectionDAG::getShiftAmountOperand(SDValue Op) { 1342 EVT OpTy = Op.getValueType(); 1343 MVT ShTy = TLI.getShiftAmountTy(); 1344 if (OpTy == ShTy || OpTy.isVector()) return Op; 1345 1346 ISD::NodeType Opcode = OpTy.bitsGT(ShTy) ? ISD::TRUNCATE : ISD::ZERO_EXTEND; 1347 return getNode(Opcode, Op.getDebugLoc(), ShTy, Op); 1348} 1349 1350/// CreateStackTemporary - Create a stack temporary, suitable for holding the 1351/// specified value type. 1352SDValue SelectionDAG::CreateStackTemporary(EVT VT, unsigned minAlign) { 1353 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo(); 1354 unsigned ByteSize = VT.getStoreSize(); 1355 const Type *Ty = VT.getTypeForEVT(*getContext()); 1356 unsigned StackAlign = 1357 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), minAlign); 1358 1359 int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false); 1360 return getFrameIndex(FrameIdx, TLI.getPointerTy()); 1361} 1362 1363/// CreateStackTemporary - Create a stack temporary suitable for holding 1364/// either of the specified value types. 1365SDValue SelectionDAG::CreateStackTemporary(EVT VT1, EVT VT2) { 1366 unsigned Bytes = std::max(VT1.getStoreSizeInBits(), 1367 VT2.getStoreSizeInBits())/8; 1368 const Type *Ty1 = VT1.getTypeForEVT(*getContext()); 1369 const Type *Ty2 = VT2.getTypeForEVT(*getContext()); 1370 const TargetData *TD = TLI.getTargetData(); 1371 unsigned Align = std::max(TD->getPrefTypeAlignment(Ty1), 1372 TD->getPrefTypeAlignment(Ty2)); 1373 1374 MachineFrameInfo *FrameInfo = getMachineFunction().getFrameInfo(); 1375 int FrameIdx = FrameInfo->CreateStackObject(Bytes, Align, false); 1376 return getFrameIndex(FrameIdx, TLI.getPointerTy()); 1377} 1378 1379SDValue SelectionDAG::FoldSetCC(EVT VT, SDValue N1, 1380 SDValue N2, ISD::CondCode Cond, DebugLoc dl) { 1381 // These setcc operations always fold. 1382 switch (Cond) { 1383 default: break; 1384 case ISD::SETFALSE: 1385 case ISD::SETFALSE2: return getConstant(0, VT); 1386 case ISD::SETTRUE: 1387 case ISD::SETTRUE2: return getConstant(1, VT); 1388 1389 case ISD::SETOEQ: 1390 case ISD::SETOGT: 1391 case ISD::SETOGE: 1392 case ISD::SETOLT: 1393 case ISD::SETOLE: 1394 case ISD::SETONE: 1395 case ISD::SETO: 1396 case ISD::SETUO: 1397 case ISD::SETUEQ: 1398 case ISD::SETUNE: 1399 assert(!N1.getValueType().isInteger() && "Illegal setcc for integer!"); 1400 break; 1401 } 1402 1403 if (ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode())) { 1404 const APInt &C2 = N2C->getAPIntValue(); 1405 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1406 const APInt &C1 = N1C->getAPIntValue(); 1407 1408 switch (Cond) { 1409 default: llvm_unreachable("Unknown integer setcc!"); 1410 case ISD::SETEQ: return getConstant(C1 == C2, VT); 1411 case ISD::SETNE: return getConstant(C1 != C2, VT); 1412 case ISD::SETULT: return getConstant(C1.ult(C2), VT); 1413 case ISD::SETUGT: return getConstant(C1.ugt(C2), VT); 1414 case ISD::SETULE: return getConstant(C1.ule(C2), VT); 1415 case ISD::SETUGE: return getConstant(C1.uge(C2), VT); 1416 case ISD::SETLT: return getConstant(C1.slt(C2), VT); 1417 case ISD::SETGT: return getConstant(C1.sgt(C2), VT); 1418 case ISD::SETLE: return getConstant(C1.sle(C2), VT); 1419 case ISD::SETGE: return getConstant(C1.sge(C2), VT); 1420 } 1421 } 1422 } 1423 if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 1424 if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) { 1425 // No compile time operations on this type yet. 1426 if (N1C->getValueType(0) == MVT::ppcf128) 1427 return SDValue(); 1428 1429 APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF()); 1430 switch (Cond) { 1431 default: break; 1432 case ISD::SETEQ: if (R==APFloat::cmpUnordered) 1433 return getUNDEF(VT); 1434 // fall through 1435 case ISD::SETOEQ: return getConstant(R==APFloat::cmpEqual, VT); 1436 case ISD::SETNE: if (R==APFloat::cmpUnordered) 1437 return getUNDEF(VT); 1438 // fall through 1439 case ISD::SETONE: return getConstant(R==APFloat::cmpGreaterThan || 1440 R==APFloat::cmpLessThan, VT); 1441 case ISD::SETLT: if (R==APFloat::cmpUnordered) 1442 return getUNDEF(VT); 1443 // fall through 1444 case ISD::SETOLT: return getConstant(R==APFloat::cmpLessThan, VT); 1445 case ISD::SETGT: if (R==APFloat::cmpUnordered) 1446 return getUNDEF(VT); 1447 // fall through 1448 case ISD::SETOGT: return getConstant(R==APFloat::cmpGreaterThan, VT); 1449 case ISD::SETLE: if (R==APFloat::cmpUnordered) 1450 return getUNDEF(VT); 1451 // fall through 1452 case ISD::SETOLE: return getConstant(R==APFloat::cmpLessThan || 1453 R==APFloat::cmpEqual, VT); 1454 case ISD::SETGE: if (R==APFloat::cmpUnordered) 1455 return getUNDEF(VT); 1456 // fall through 1457 case ISD::SETOGE: return getConstant(R==APFloat::cmpGreaterThan || 1458 R==APFloat::cmpEqual, VT); 1459 case ISD::SETO: return getConstant(R!=APFloat::cmpUnordered, VT); 1460 case ISD::SETUO: return getConstant(R==APFloat::cmpUnordered, VT); 1461 case ISD::SETUEQ: return getConstant(R==APFloat::cmpUnordered || 1462 R==APFloat::cmpEqual, VT); 1463 case ISD::SETUNE: return getConstant(R!=APFloat::cmpEqual, VT); 1464 case ISD::SETULT: return getConstant(R==APFloat::cmpUnordered || 1465 R==APFloat::cmpLessThan, VT); 1466 case ISD::SETUGT: return getConstant(R==APFloat::cmpGreaterThan || 1467 R==APFloat::cmpUnordered, VT); 1468 case ISD::SETULE: return getConstant(R!=APFloat::cmpGreaterThan, VT); 1469 case ISD::SETUGE: return getConstant(R!=APFloat::cmpLessThan, VT); 1470 } 1471 } else { 1472 // Ensure that the constant occurs on the RHS. 1473 return getSetCC(dl, VT, N2, N1, ISD::getSetCCSwappedOperands(Cond)); 1474 } 1475 } 1476 1477 // Could not fold it. 1478 return SDValue(); 1479} 1480 1481/// SignBitIsZero - Return true if the sign bit of Op is known to be zero. We 1482/// use this predicate to simplify operations downstream. 1483bool SelectionDAG::SignBitIsZero(SDValue Op, unsigned Depth) const { 1484 // This predicate is not safe for vector operations. 1485 if (Op.getValueType().isVector()) 1486 return false; 1487 1488 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits(); 1489 return MaskedValueIsZero(Op, APInt::getSignBit(BitWidth), Depth); 1490} 1491 1492/// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use 1493/// this predicate to simplify operations downstream. Mask is known to be zero 1494/// for bits that V cannot have. 1495bool SelectionDAG::MaskedValueIsZero(SDValue Op, const APInt &Mask, 1496 unsigned Depth) const { 1497 APInt KnownZero, KnownOne; 1498 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth); 1499 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1500 return (KnownZero & Mask) == Mask; 1501} 1502 1503/// ComputeMaskedBits - Determine which of the bits specified in Mask are 1504/// known to be either zero or one and return them in the KnownZero/KnownOne 1505/// bitsets. This code only analyzes bits in Mask, in order to short-circuit 1506/// processing. 1507void SelectionDAG::ComputeMaskedBits(SDValue Op, const APInt &Mask, 1508 APInt &KnownZero, APInt &KnownOne, 1509 unsigned Depth) const { 1510 unsigned BitWidth = Mask.getBitWidth(); 1511 assert(BitWidth == Op.getValueType().getScalarType().getSizeInBits() && 1512 "Mask size mismatches value type size!"); 1513 1514 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything. 1515 if (Depth == 6 || Mask == 0) 1516 return; // Limit search depth. 1517 1518 APInt KnownZero2, KnownOne2; 1519 1520 switch (Op.getOpcode()) { 1521 case ISD::Constant: 1522 // We know all of the bits for a constant! 1523 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & Mask; 1524 KnownZero = ~KnownOne & Mask; 1525 return; 1526 case ISD::AND: 1527 // If either the LHS or the RHS are Zero, the result is zero. 1528 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 1529 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownZero, 1530 KnownZero2, KnownOne2, Depth+1); 1531 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1532 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1533 1534 // Output known-1 bits are only known if set in both the LHS & RHS. 1535 KnownOne &= KnownOne2; 1536 // Output known-0 are known to be clear if zero in either the LHS | RHS. 1537 KnownZero |= KnownZero2; 1538 return; 1539 case ISD::OR: 1540 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 1541 ComputeMaskedBits(Op.getOperand(0), Mask & ~KnownOne, 1542 KnownZero2, KnownOne2, Depth+1); 1543 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1544 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1545 1546 // Output known-0 bits are only known if clear in both the LHS & RHS. 1547 KnownZero &= KnownZero2; 1548 // Output known-1 are known to be set if set in either the LHS | RHS. 1549 KnownOne |= KnownOne2; 1550 return; 1551 case ISD::XOR: { 1552 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 1553 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1); 1554 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1555 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1556 1557 // Output known-0 bits are known if clear or set in both the LHS & RHS. 1558 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 1559 // Output known-1 are known to be set if set in only one of the LHS, RHS. 1560 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 1561 KnownZero = KnownZeroOut; 1562 return; 1563 } 1564 case ISD::MUL: { 1565 APInt Mask2 = APInt::getAllOnesValue(BitWidth); 1566 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero, KnownOne, Depth+1); 1567 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1); 1568 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1569 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1570 1571 // If low bits are zero in either operand, output low known-0 bits. 1572 // Also compute a conserative estimate for high known-0 bits. 1573 // More trickiness is possible, but this is sufficient for the 1574 // interesting case of alignment computation. 1575 KnownOne.clear(); 1576 unsigned TrailZ = KnownZero.countTrailingOnes() + 1577 KnownZero2.countTrailingOnes(); 1578 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() + 1579 KnownZero2.countLeadingOnes(), 1580 BitWidth) - BitWidth; 1581 1582 TrailZ = std::min(TrailZ, BitWidth); 1583 LeadZ = std::min(LeadZ, BitWidth); 1584 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) | 1585 APInt::getHighBitsSet(BitWidth, LeadZ); 1586 KnownZero &= Mask; 1587 return; 1588 } 1589 case ISD::UDIV: { 1590 // For the purposes of computing leading zeros we can conservatively 1591 // treat a udiv as a logical right shift by the power of 2 known to 1592 // be less than the denominator. 1593 APInt AllOnes = APInt::getAllOnesValue(BitWidth); 1594 ComputeMaskedBits(Op.getOperand(0), 1595 AllOnes, KnownZero2, KnownOne2, Depth+1); 1596 unsigned LeadZ = KnownZero2.countLeadingOnes(); 1597 1598 KnownOne2.clear(); 1599 KnownZero2.clear(); 1600 ComputeMaskedBits(Op.getOperand(1), 1601 AllOnes, KnownZero2, KnownOne2, Depth+1); 1602 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros(); 1603 if (RHSUnknownLeadingOnes != BitWidth) 1604 LeadZ = std::min(BitWidth, 1605 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1); 1606 1607 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask; 1608 return; 1609 } 1610 case ISD::SELECT: 1611 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero, KnownOne, Depth+1); 1612 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1); 1613 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1614 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1615 1616 // Only known if known in both the LHS and RHS. 1617 KnownOne &= KnownOne2; 1618 KnownZero &= KnownZero2; 1619 return; 1620 case ISD::SELECT_CC: 1621 ComputeMaskedBits(Op.getOperand(3), Mask, KnownZero, KnownOne, Depth+1); 1622 ComputeMaskedBits(Op.getOperand(2), Mask, KnownZero2, KnownOne2, Depth+1); 1623 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1624 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1625 1626 // Only known if known in both the LHS and RHS. 1627 KnownOne &= KnownOne2; 1628 KnownZero &= KnownZero2; 1629 return; 1630 case ISD::SADDO: 1631 case ISD::UADDO: 1632 case ISD::SSUBO: 1633 case ISD::USUBO: 1634 case ISD::SMULO: 1635 case ISD::UMULO: 1636 if (Op.getResNo() != 1) 1637 return; 1638 // The boolean result conforms to getBooleanContents. Fall through. 1639 case ISD::SETCC: 1640 // If we know the result of a setcc has the top bits zero, use this info. 1641 if (TLI.getBooleanContents() == TargetLowering::ZeroOrOneBooleanContent && 1642 BitWidth > 1) 1643 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1); 1644 return; 1645 case ISD::SHL: 1646 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0 1647 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1648 unsigned ShAmt = SA->getZExtValue(); 1649 1650 // If the shift count is an invalid immediate, don't do anything. 1651 if (ShAmt >= BitWidth) 1652 return; 1653 1654 ComputeMaskedBits(Op.getOperand(0), Mask.lshr(ShAmt), 1655 KnownZero, KnownOne, Depth+1); 1656 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1657 KnownZero <<= ShAmt; 1658 KnownOne <<= ShAmt; 1659 // low bits known zero. 1660 KnownZero |= APInt::getLowBitsSet(BitWidth, ShAmt); 1661 } 1662 return; 1663 case ISD::SRL: 1664 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0 1665 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1666 unsigned ShAmt = SA->getZExtValue(); 1667 1668 // If the shift count is an invalid immediate, don't do anything. 1669 if (ShAmt >= BitWidth) 1670 return; 1671 1672 ComputeMaskedBits(Op.getOperand(0), (Mask << ShAmt), 1673 KnownZero, KnownOne, Depth+1); 1674 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1675 KnownZero = KnownZero.lshr(ShAmt); 1676 KnownOne = KnownOne.lshr(ShAmt); 1677 1678 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask; 1679 KnownZero |= HighBits; // High bits known zero. 1680 } 1681 return; 1682 case ISD::SRA: 1683 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1684 unsigned ShAmt = SA->getZExtValue(); 1685 1686 // If the shift count is an invalid immediate, don't do anything. 1687 if (ShAmt >= BitWidth) 1688 return; 1689 1690 APInt InDemandedMask = (Mask << ShAmt); 1691 // If any of the demanded bits are produced by the sign extension, we also 1692 // demand the input sign bit. 1693 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt) & Mask; 1694 if (HighBits.getBoolValue()) 1695 InDemandedMask |= APInt::getSignBit(BitWidth); 1696 1697 ComputeMaskedBits(Op.getOperand(0), InDemandedMask, KnownZero, KnownOne, 1698 Depth+1); 1699 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1700 KnownZero = KnownZero.lshr(ShAmt); 1701 KnownOne = KnownOne.lshr(ShAmt); 1702 1703 // Handle the sign bits. 1704 APInt SignBit = APInt::getSignBit(BitWidth); 1705 SignBit = SignBit.lshr(ShAmt); // Adjust to where it is now in the mask. 1706 1707 if (KnownZero.intersects(SignBit)) { 1708 KnownZero |= HighBits; // New bits are known zero. 1709 } else if (KnownOne.intersects(SignBit)) { 1710 KnownOne |= HighBits; // New bits are known one. 1711 } 1712 } 1713 return; 1714 case ISD::SIGN_EXTEND_INREG: { 1715 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1716 unsigned EBits = EVT.getSizeInBits(); 1717 1718 // Sign extension. Compute the demanded bits in the result that are not 1719 // present in the input. 1720 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - EBits) & Mask; 1721 1722 APInt InSignBit = APInt::getSignBit(EBits); 1723 APInt InputDemandedBits = Mask & APInt::getLowBitsSet(BitWidth, EBits); 1724 1725 // If the sign extended bits are demanded, we know that the sign 1726 // bit is demanded. 1727 InSignBit.zext(BitWidth); 1728 if (NewBits.getBoolValue()) 1729 InputDemandedBits |= InSignBit; 1730 1731 ComputeMaskedBits(Op.getOperand(0), InputDemandedBits, 1732 KnownZero, KnownOne, Depth+1); 1733 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1734 1735 // If the sign bit of the input is known set or clear, then we know the 1736 // top bits of the result. 1737 if (KnownZero.intersects(InSignBit)) { // Input sign bit known clear 1738 KnownZero |= NewBits; 1739 KnownOne &= ~NewBits; 1740 } else if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1741 KnownOne |= NewBits; 1742 KnownZero &= ~NewBits; 1743 } else { // Input sign bit unknown 1744 KnownZero &= ~NewBits; 1745 KnownOne &= ~NewBits; 1746 } 1747 return; 1748 } 1749 case ISD::CTTZ: 1750 case ISD::CTLZ: 1751 case ISD::CTPOP: { 1752 unsigned LowBits = Log2_32(BitWidth)+1; 1753 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits); 1754 KnownOne.clear(); 1755 return; 1756 } 1757 case ISD::LOAD: { 1758 if (ISD::isZEXTLoad(Op.getNode())) { 1759 LoadSDNode *LD = cast<LoadSDNode>(Op); 1760 EVT VT = LD->getMemoryVT(); 1761 unsigned MemBits = VT.getSizeInBits(); 1762 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits) & Mask; 1763 } 1764 return; 1765 } 1766 case ISD::ZERO_EXTEND: { 1767 EVT InVT = Op.getOperand(0).getValueType(); 1768 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1769 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask; 1770 APInt InMask = Mask; 1771 InMask.trunc(InBits); 1772 KnownZero.trunc(InBits); 1773 KnownOne.trunc(InBits); 1774 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1775 KnownZero.zext(BitWidth); 1776 KnownOne.zext(BitWidth); 1777 KnownZero |= NewBits; 1778 return; 1779 } 1780 case ISD::SIGN_EXTEND: { 1781 EVT InVT = Op.getOperand(0).getValueType(); 1782 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1783 APInt InSignBit = APInt::getSignBit(InBits); 1784 APInt NewBits = APInt::getHighBitsSet(BitWidth, BitWidth - InBits) & Mask; 1785 APInt InMask = Mask; 1786 InMask.trunc(InBits); 1787 1788 // If any of the sign extended bits are demanded, we know that the sign 1789 // bit is demanded. Temporarily set this bit in the mask for our callee. 1790 if (NewBits.getBoolValue()) 1791 InMask |= InSignBit; 1792 1793 KnownZero.trunc(InBits); 1794 KnownOne.trunc(InBits); 1795 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1796 1797 // Note if the sign bit is known to be zero or one. 1798 bool SignBitKnownZero = KnownZero.isNegative(); 1799 bool SignBitKnownOne = KnownOne.isNegative(); 1800 assert(!(SignBitKnownZero && SignBitKnownOne) && 1801 "Sign bit can't be known to be both zero and one!"); 1802 1803 // If the sign bit wasn't actually demanded by our caller, we don't 1804 // want it set in the KnownZero and KnownOne result values. Reset the 1805 // mask and reapply it to the result values. 1806 InMask = Mask; 1807 InMask.trunc(InBits); 1808 KnownZero &= InMask; 1809 KnownOne &= InMask; 1810 1811 KnownZero.zext(BitWidth); 1812 KnownOne.zext(BitWidth); 1813 1814 // If the sign bit is known zero or one, the top bits match. 1815 if (SignBitKnownZero) 1816 KnownZero |= NewBits; 1817 else if (SignBitKnownOne) 1818 KnownOne |= NewBits; 1819 return; 1820 } 1821 case ISD::ANY_EXTEND: { 1822 EVT InVT = Op.getOperand(0).getValueType(); 1823 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1824 APInt InMask = Mask; 1825 InMask.trunc(InBits); 1826 KnownZero.trunc(InBits); 1827 KnownOne.trunc(InBits); 1828 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1829 KnownZero.zext(BitWidth); 1830 KnownOne.zext(BitWidth); 1831 return; 1832 } 1833 case ISD::TRUNCATE: { 1834 EVT InVT = Op.getOperand(0).getValueType(); 1835 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1836 APInt InMask = Mask; 1837 InMask.zext(InBits); 1838 KnownZero.zext(InBits); 1839 KnownOne.zext(InBits); 1840 ComputeMaskedBits(Op.getOperand(0), InMask, KnownZero, KnownOne, Depth+1); 1841 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1842 KnownZero.trunc(BitWidth); 1843 KnownOne.trunc(BitWidth); 1844 break; 1845 } 1846 case ISD::AssertZext: { 1847 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1848 APInt InMask = APInt::getLowBitsSet(BitWidth, VT.getSizeInBits()); 1849 ComputeMaskedBits(Op.getOperand(0), Mask & InMask, KnownZero, 1850 KnownOne, Depth+1); 1851 KnownZero |= (~InMask) & Mask; 1852 return; 1853 } 1854 case ISD::FGETSIGN: 1855 // All bits are zero except the low bit. 1856 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - 1); 1857 return; 1858 1859 case ISD::SUB: { 1860 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) { 1861 // We know that the top bits of C-X are clear if X contains less bits 1862 // than C (i.e. no wrap-around can happen). For example, 20-X is 1863 // positive if we can prove that X is >= 0 and < 16. 1864 if (CLHS->getAPIntValue().isNonNegative()) { 1865 unsigned NLZ = (CLHS->getAPIntValue()+1).countLeadingZeros(); 1866 // NLZ can't be BitWidth with no sign bit 1867 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1); 1868 ComputeMaskedBits(Op.getOperand(1), MaskV, KnownZero2, KnownOne2, 1869 Depth+1); 1870 1871 // If all of the MaskV bits are known to be zero, then we know the 1872 // output top bits are zero, because we now know that the output is 1873 // from [0-C]. 1874 if ((KnownZero2 & MaskV) == MaskV) { 1875 unsigned NLZ2 = CLHS->getAPIntValue().countLeadingZeros(); 1876 // Top bits known zero. 1877 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask; 1878 } 1879 } 1880 } 1881 } 1882 // fall through 1883 case ISD::ADD: { 1884 // Output known-0 bits are known if clear or set in both the low clear bits 1885 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the 1886 // low 3 bits clear. 1887 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes()); 1888 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1); 1889 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1890 unsigned KnownZeroOut = KnownZero2.countTrailingOnes(); 1891 1892 ComputeMaskedBits(Op.getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1); 1893 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1894 KnownZeroOut = std::min(KnownZeroOut, 1895 KnownZero2.countTrailingOnes()); 1896 1897 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut); 1898 return; 1899 } 1900 case ISD::SREM: 1901 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1902 const APInt &RA = Rem->getAPIntValue(); 1903 if (RA.isPowerOf2() || (-RA).isPowerOf2()) { 1904 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA; 1905 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth); 1906 ComputeMaskedBits(Op.getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1); 1907 1908 // If the sign bit of the first operand is zero, the sign bit of 1909 // the result is zero. If the first operand has no one bits below 1910 // the second operand's single 1 bit, its sign will be zero. 1911 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits)) 1912 KnownZero2 |= ~LowBits; 1913 1914 KnownZero |= KnownZero2 & Mask; 1915 1916 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?"); 1917 } 1918 } 1919 return; 1920 case ISD::UREM: { 1921 if (ConstantSDNode *Rem = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1922 const APInt &RA = Rem->getAPIntValue(); 1923 if (RA.isPowerOf2()) { 1924 APInt LowBits = (RA - 1); 1925 APInt Mask2 = LowBits & Mask; 1926 KnownZero |= ~LowBits & Mask; 1927 ComputeMaskedBits(Op.getOperand(0), Mask2, KnownZero, KnownOne,Depth+1); 1928 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?"); 1929 break; 1930 } 1931 } 1932 1933 // Since the result is less than or equal to either operand, any leading 1934 // zero bits in either operand must also exist in the result. 1935 APInt AllOnes = APInt::getAllOnesValue(BitWidth); 1936 ComputeMaskedBits(Op.getOperand(0), AllOnes, KnownZero, KnownOne, 1937 Depth+1); 1938 ComputeMaskedBits(Op.getOperand(1), AllOnes, KnownZero2, KnownOne2, 1939 Depth+1); 1940 1941 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(), 1942 KnownZero2.countLeadingOnes()); 1943 KnownOne.clear(); 1944 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask; 1945 return; 1946 } 1947 default: 1948 // Allow the target to implement this method for its nodes. 1949 if (Op.getOpcode() >= ISD::BUILTIN_OP_END) { 1950 case ISD::INTRINSIC_WO_CHAIN: 1951 case ISD::INTRINSIC_W_CHAIN: 1952 case ISD::INTRINSIC_VOID: 1953 TLI.computeMaskedBitsForTargetNode(Op, Mask, KnownZero, KnownOne, *this, 1954 Depth); 1955 } 1956 return; 1957 } 1958} 1959 1960/// ComputeNumSignBits - Return the number of times the sign bit of the 1961/// register is replicated into the other bits. We know that at least 1 bit 1962/// is always equal to the sign bit (itself), but other cases can give us 1963/// information. For example, immediately after an "SRA X, 2", we know that 1964/// the top 3 bits are all equal to each other, so we return 3. 1965unsigned SelectionDAG::ComputeNumSignBits(SDValue Op, unsigned Depth) const{ 1966 EVT VT = Op.getValueType(); 1967 assert(VT.isInteger() && "Invalid VT!"); 1968 unsigned VTBits = VT.getScalarType().getSizeInBits(); 1969 unsigned Tmp, Tmp2; 1970 unsigned FirstAnswer = 1; 1971 1972 if (Depth == 6) 1973 return 1; // Limit search depth. 1974 1975 switch (Op.getOpcode()) { 1976 default: break; 1977 case ISD::AssertSext: 1978 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits(); 1979 return VTBits-Tmp+1; 1980 case ISD::AssertZext: 1981 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits(); 1982 return VTBits-Tmp; 1983 1984 case ISD::Constant: { 1985 const APInt &Val = cast<ConstantSDNode>(Op)->getAPIntValue(); 1986 // If negative, return # leading ones. 1987 if (Val.isNegative()) 1988 return Val.countLeadingOnes(); 1989 1990 // Return # leading zeros. 1991 return Val.countLeadingZeros(); 1992 } 1993 1994 case ISD::SIGN_EXTEND: 1995 Tmp = VTBits-Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1996 return ComputeNumSignBits(Op.getOperand(0), Depth+1) + Tmp; 1997 1998 case ISD::SIGN_EXTEND_INREG: 1999 // Max of the input and what this extends. 2000 Tmp = cast<VTSDNode>(Op.getOperand(1))->getVT().getSizeInBits(); 2001 Tmp = VTBits-Tmp+1; 2002 2003 Tmp2 = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2004 return std::max(Tmp, Tmp2); 2005 2006 case ISD::SRA: 2007 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2008 // SRA X, C -> adds C sign bits. 2009 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2010 Tmp += C->getZExtValue(); 2011 if (Tmp > VTBits) Tmp = VTBits; 2012 } 2013 return Tmp; 2014 case ISD::SHL: 2015 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2016 // shl destroys sign bits. 2017 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2018 if (C->getZExtValue() >= VTBits || // Bad shift. 2019 C->getZExtValue() >= Tmp) break; // Shifted all sign bits out. 2020 return Tmp - C->getZExtValue(); 2021 } 2022 break; 2023 case ISD::AND: 2024 case ISD::OR: 2025 case ISD::XOR: // NOT is handled here. 2026 // Logical binary ops preserve the number of sign bits at the worst. 2027 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2028 if (Tmp != 1) { 2029 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2030 FirstAnswer = std::min(Tmp, Tmp2); 2031 // We computed what we know about the sign bits as our first 2032 // answer. Now proceed to the generic code that uses 2033 // ComputeMaskedBits, and pick whichever answer is better. 2034 } 2035 break; 2036 2037 case ISD::SELECT: 2038 Tmp = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2039 if (Tmp == 1) return 1; // Early out. 2040 Tmp2 = ComputeNumSignBits(Op.getOperand(2), Depth+1); 2041 return std::min(Tmp, Tmp2); 2042 2043 case ISD::SADDO: 2044 case ISD::UADDO: 2045 case ISD::SSUBO: 2046 case ISD::USUBO: 2047 case ISD::SMULO: 2048 case ISD::UMULO: 2049 if (Op.getResNo() != 1) 2050 break; 2051 // The boolean result conforms to getBooleanContents. Fall through. 2052 case ISD::SETCC: 2053 // If setcc returns 0/-1, all bits are sign bits. 2054 if (TLI.getBooleanContents() == 2055 TargetLowering::ZeroOrNegativeOneBooleanContent) 2056 return VTBits; 2057 break; 2058 case ISD::ROTL: 2059 case ISD::ROTR: 2060 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 2061 unsigned RotAmt = C->getZExtValue() & (VTBits-1); 2062 2063 // Handle rotate right by N like a rotate left by 32-N. 2064 if (Op.getOpcode() == ISD::ROTR) 2065 RotAmt = (VTBits-RotAmt) & (VTBits-1); 2066 2067 // If we aren't rotating out all of the known-in sign bits, return the 2068 // number that are left. This handles rotl(sext(x), 1) for example. 2069 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2070 if (Tmp > RotAmt+1) return Tmp-RotAmt; 2071 } 2072 break; 2073 case ISD::ADD: 2074 // Add can have at most one carry bit. Thus we know that the output 2075 // is, at worst, one more bit than the inputs. 2076 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2077 if (Tmp == 1) return 1; // Early out. 2078 2079 // Special case decrementing a value (ADD X, -1): 2080 if (ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(Op.getOperand(1))) 2081 if (CRHS->isAllOnesValue()) { 2082 APInt KnownZero, KnownOne; 2083 APInt Mask = APInt::getAllOnesValue(VTBits); 2084 ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1); 2085 2086 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2087 // sign bits set. 2088 if ((KnownZero | APInt(VTBits, 1)) == Mask) 2089 return VTBits; 2090 2091 // If we are subtracting one from a positive number, there is no carry 2092 // out of the result. 2093 if (KnownZero.isNegative()) 2094 return Tmp; 2095 } 2096 2097 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2098 if (Tmp2 == 1) return 1; 2099 return std::min(Tmp, Tmp2)-1; 2100 break; 2101 2102 case ISD::SUB: 2103 Tmp2 = ComputeNumSignBits(Op.getOperand(1), Depth+1); 2104 if (Tmp2 == 1) return 1; 2105 2106 // Handle NEG. 2107 if (ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(Op.getOperand(0))) 2108 if (CLHS->isNullValue()) { 2109 APInt KnownZero, KnownOne; 2110 APInt Mask = APInt::getAllOnesValue(VTBits); 2111 ComputeMaskedBits(Op.getOperand(1), Mask, KnownZero, KnownOne, Depth+1); 2112 // If the input is known to be 0 or 1, the output is 0/-1, which is all 2113 // sign bits set. 2114 if ((KnownZero | APInt(VTBits, 1)) == Mask) 2115 return VTBits; 2116 2117 // If the input is known to be positive (the sign bit is known clear), 2118 // the output of the NEG has the same number of sign bits as the input. 2119 if (KnownZero.isNegative()) 2120 return Tmp2; 2121 2122 // Otherwise, we treat this like a SUB. 2123 } 2124 2125 // Sub can have at most one carry bit. Thus we know that the output 2126 // is, at worst, one more bit than the inputs. 2127 Tmp = ComputeNumSignBits(Op.getOperand(0), Depth+1); 2128 if (Tmp == 1) return 1; // Early out. 2129 return std::min(Tmp, Tmp2)-1; 2130 break; 2131 case ISD::TRUNCATE: 2132 // FIXME: it's tricky to do anything useful for this, but it is an important 2133 // case for targets like X86. 2134 break; 2135 } 2136 2137 // Handle LOADX separately here. EXTLOAD case will fallthrough. 2138 if (Op.getOpcode() == ISD::LOAD) { 2139 LoadSDNode *LD = cast<LoadSDNode>(Op); 2140 unsigned ExtType = LD->getExtensionType(); 2141 switch (ExtType) { 2142 default: break; 2143 case ISD::SEXTLOAD: // '17' bits known 2144 Tmp = LD->getMemoryVT().getSizeInBits(); 2145 return VTBits-Tmp+1; 2146 case ISD::ZEXTLOAD: // '16' bits known 2147 Tmp = LD->getMemoryVT().getSizeInBits(); 2148 return VTBits-Tmp; 2149 } 2150 } 2151 2152 // Allow the target to implement this method for its nodes. 2153 if (Op.getOpcode() >= ISD::BUILTIN_OP_END || 2154 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 2155 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 2156 Op.getOpcode() == ISD::INTRINSIC_VOID) { 2157 unsigned NumBits = TLI.ComputeNumSignBitsForTargetNode(Op, Depth); 2158 if (NumBits > 1) FirstAnswer = std::max(FirstAnswer, NumBits); 2159 } 2160 2161 // Finally, if we can prove that the top bits of the result are 0's or 1's, 2162 // use this information. 2163 APInt KnownZero, KnownOne; 2164 APInt Mask = APInt::getAllOnesValue(VTBits); 2165 ComputeMaskedBits(Op, Mask, KnownZero, KnownOne, Depth); 2166 2167 if (KnownZero.isNegative()) { // sign bit is 0 2168 Mask = KnownZero; 2169 } else if (KnownOne.isNegative()) { // sign bit is 1; 2170 Mask = KnownOne; 2171 } else { 2172 // Nothing known. 2173 return FirstAnswer; 2174 } 2175 2176 // Okay, we know that the sign bit in Mask is set. Use CLZ to determine 2177 // the number of identical bits in the top of the input value. 2178 Mask = ~Mask; 2179 Mask <<= Mask.getBitWidth()-VTBits; 2180 // Return # leading zeros. We use 'min' here in case Val was zero before 2181 // shifting. We don't want to return '64' as for an i32 "0". 2182 return std::max(FirstAnswer, std::min(VTBits, Mask.countLeadingZeros())); 2183} 2184 2185bool SelectionDAG::isKnownNeverNaN(SDValue Op) const { 2186 // If we're told that NaNs won't happen, assume they won't. 2187 if (FiniteOnlyFPMath()) 2188 return true; 2189 2190 // If the value is a constant, we can obviously see if it is a NaN or not. 2191 if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op)) 2192 return !C->getValueAPF().isNaN(); 2193 2194 // TODO: Recognize more cases here. 2195 2196 return false; 2197} 2198 2199bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const { 2200 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2201 if (!GA) return false; 2202 if (GA->getOffset() != 0) return false; 2203 GlobalVariable *GV = dyn_cast<GlobalVariable>(GA->getGlobal()); 2204 if (!GV) return false; 2205 MachineModuleInfo *MMI = getMachineModuleInfo(); 2206 return MMI && MMI->hasDebugInfo(); 2207} 2208 2209 2210/// getShuffleScalarElt - Returns the scalar element that will make up the ith 2211/// element of the result of the vector shuffle. 2212SDValue SelectionDAG::getShuffleScalarElt(const ShuffleVectorSDNode *N, 2213 unsigned i) { 2214 EVT VT = N->getValueType(0); 2215 DebugLoc dl = N->getDebugLoc(); 2216 if (N->getMaskElt(i) < 0) 2217 return getUNDEF(VT.getVectorElementType()); 2218 unsigned Index = N->getMaskElt(i); 2219 unsigned NumElems = VT.getVectorNumElements(); 2220 SDValue V = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1); 2221 Index %= NumElems; 2222 2223 if (V.getOpcode() == ISD::BIT_CONVERT) { 2224 V = V.getOperand(0); 2225 EVT VVT = V.getValueType(); 2226 if (!VVT.isVector() || VVT.getVectorNumElements() != (unsigned)NumElems) 2227 return SDValue(); 2228 } 2229 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) 2230 return (Index == 0) ? V.getOperand(0) 2231 : getUNDEF(VT.getVectorElementType()); 2232 if (V.getOpcode() == ISD::BUILD_VECTOR) 2233 return V.getOperand(Index); 2234 if (const ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(V)) 2235 return getShuffleScalarElt(SVN, Index); 2236 return SDValue(); 2237} 2238 2239 2240/// getNode - Gets or creates the specified node. 2241/// 2242SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT) { 2243 FoldingSetNodeID ID; 2244 AddNodeIDNode(ID, Opcode, getVTList(VT), 0, 0); 2245 void *IP = 0; 2246 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 2247 return SDValue(E, 0); 2248 SDNode *N = NodeAllocator.Allocate<SDNode>(); 2249 new (N) SDNode(Opcode, DL, getVTList(VT)); 2250 CSEMap.InsertNode(N, IP); 2251 2252 AllNodes.push_back(N); 2253#ifndef NDEBUG 2254 VerifyNode(N); 2255#endif 2256 return SDValue(N, 0); 2257} 2258 2259SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, 2260 EVT VT, SDValue Operand) { 2261 // Constant fold unary operations with an integer constant operand. 2262 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Operand.getNode())) { 2263 const APInt &Val = C->getAPIntValue(); 2264 unsigned BitWidth = VT.getSizeInBits(); 2265 switch (Opcode) { 2266 default: break; 2267 case ISD::SIGN_EXTEND: 2268 return getConstant(APInt(Val).sextOrTrunc(BitWidth), VT); 2269 case ISD::ANY_EXTEND: 2270 case ISD::ZERO_EXTEND: 2271 case ISD::TRUNCATE: 2272 return getConstant(APInt(Val).zextOrTrunc(BitWidth), VT); 2273 case ISD::UINT_TO_FP: 2274 case ISD::SINT_TO_FP: { 2275 const uint64_t zero[] = {0, 0}; 2276 // No compile time operations on this type. 2277 if (VT==MVT::ppcf128) 2278 break; 2279 APFloat apf = APFloat(APInt(BitWidth, 2, zero)); 2280 (void)apf.convertFromAPInt(Val, 2281 Opcode==ISD::SINT_TO_FP, 2282 APFloat::rmNearestTiesToEven); 2283 return getConstantFP(apf, VT); 2284 } 2285 case ISD::BIT_CONVERT: 2286 if (VT == MVT::f32 && C->getValueType(0) == MVT::i32) 2287 return getConstantFP(Val.bitsToFloat(), VT); 2288 else if (VT == MVT::f64 && C->getValueType(0) == MVT::i64) 2289 return getConstantFP(Val.bitsToDouble(), VT); 2290 break; 2291 case ISD::BSWAP: 2292 return getConstant(Val.byteSwap(), VT); 2293 case ISD::CTPOP: 2294 return getConstant(Val.countPopulation(), VT); 2295 case ISD::CTLZ: 2296 return getConstant(Val.countLeadingZeros(), VT); 2297 case ISD::CTTZ: 2298 return getConstant(Val.countTrailingZeros(), VT); 2299 } 2300 } 2301 2302 // Constant fold unary operations with a floating point constant operand. 2303 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) { 2304 APFloat V = C->getValueAPF(); // make copy 2305 if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) { 2306 switch (Opcode) { 2307 case ISD::FNEG: 2308 V.changeSign(); 2309 return getConstantFP(V, VT); 2310 case ISD::FABS: 2311 V.clearSign(); 2312 return getConstantFP(V, VT); 2313 case ISD::FP_ROUND: 2314 case ISD::FP_EXTEND: { 2315 bool ignored; 2316 // This can return overflow, underflow, or inexact; we don't care. 2317 // FIXME need to be more flexible about rounding mode. 2318 (void)V.convert(*EVTToAPFloatSemantics(VT), 2319 APFloat::rmNearestTiesToEven, &ignored); 2320 return getConstantFP(V, VT); 2321 } 2322 case ISD::FP_TO_SINT: 2323 case ISD::FP_TO_UINT: { 2324 integerPart x[2]; 2325 bool ignored; 2326 assert(integerPartWidth >= 64); 2327 // FIXME need to be more flexible about rounding mode. 2328 APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(), 2329 Opcode==ISD::FP_TO_SINT, 2330 APFloat::rmTowardZero, &ignored); 2331 if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual 2332 break; 2333 APInt api(VT.getSizeInBits(), 2, x); 2334 return getConstant(api, VT); 2335 } 2336 case ISD::BIT_CONVERT: 2337 if (VT == MVT::i32 && C->getValueType(0) == MVT::f32) 2338 return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT); 2339 else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64) 2340 return getConstant(V.bitcastToAPInt().getZExtValue(), VT); 2341 break; 2342 } 2343 } 2344 } 2345 2346 unsigned OpOpcode = Operand.getNode()->getOpcode(); 2347 switch (Opcode) { 2348 case ISD::TokenFactor: 2349 case ISD::MERGE_VALUES: 2350 case ISD::CONCAT_VECTORS: 2351 return Operand; // Factor, merge or concat of one node? No need. 2352 case ISD::FP_ROUND: llvm_unreachable("Invalid method to make FP_ROUND node"); 2353 case ISD::FP_EXTEND: 2354 assert(VT.isFloatingPoint() && 2355 Operand.getValueType().isFloatingPoint() && "Invalid FP cast!"); 2356 if (Operand.getValueType() == VT) return Operand; // noop conversion. 2357 assert((!VT.isVector() || 2358 VT.getVectorNumElements() == 2359 Operand.getValueType().getVectorNumElements()) && 2360 "Vector element count mismatch!"); 2361 if (Operand.getOpcode() == ISD::UNDEF) 2362 return getUNDEF(VT); 2363 break; 2364 case ISD::SIGN_EXTEND: 2365 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2366 "Invalid SIGN_EXTEND!"); 2367 if (Operand.getValueType() == VT) return Operand; // noop extension 2368 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) && 2369 "Invalid sext node, dst < src!"); 2370 assert((!VT.isVector() || 2371 VT.getVectorNumElements() == 2372 Operand.getValueType().getVectorNumElements()) && 2373 "Vector element count mismatch!"); 2374 if (OpOpcode == ISD::SIGN_EXTEND || OpOpcode == ISD::ZERO_EXTEND) 2375 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0)); 2376 break; 2377 case ISD::ZERO_EXTEND: 2378 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2379 "Invalid ZERO_EXTEND!"); 2380 if (Operand.getValueType() == VT) return Operand; // noop extension 2381 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) && 2382 "Invalid zext node, dst < src!"); 2383 assert((!VT.isVector() || 2384 VT.getVectorNumElements() == 2385 Operand.getValueType().getVectorNumElements()) && 2386 "Vector element count mismatch!"); 2387 if (OpOpcode == ISD::ZERO_EXTEND) // (zext (zext x)) -> (zext x) 2388 return getNode(ISD::ZERO_EXTEND, DL, VT, 2389 Operand.getNode()->getOperand(0)); 2390 break; 2391 case ISD::ANY_EXTEND: 2392 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2393 "Invalid ANY_EXTEND!"); 2394 if (Operand.getValueType() == VT) return Operand; // noop extension 2395 assert(Operand.getValueType().getScalarType().bitsLT(VT.getScalarType()) && 2396 "Invalid anyext node, dst < src!"); 2397 assert((!VT.isVector() || 2398 VT.getVectorNumElements() == 2399 Operand.getValueType().getVectorNumElements()) && 2400 "Vector element count mismatch!"); 2401 if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND) 2402 // (ext (zext x)) -> (zext x) and (ext (sext x)) -> (sext x) 2403 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0)); 2404 break; 2405 case ISD::TRUNCATE: 2406 assert(VT.isInteger() && Operand.getValueType().isInteger() && 2407 "Invalid TRUNCATE!"); 2408 if (Operand.getValueType() == VT) return Operand; // noop truncate 2409 assert(Operand.getValueType().getScalarType().bitsGT(VT.getScalarType()) && 2410 "Invalid truncate node, src < dst!"); 2411 assert((!VT.isVector() || 2412 VT.getVectorNumElements() == 2413 Operand.getValueType().getVectorNumElements()) && 2414 "Vector element count mismatch!"); 2415 if (OpOpcode == ISD::TRUNCATE) 2416 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0)); 2417 else if (OpOpcode == ISD::ZERO_EXTEND || OpOpcode == ISD::SIGN_EXTEND || 2418 OpOpcode == ISD::ANY_EXTEND) { 2419 // If the source is smaller than the dest, we still need an extend. 2420 if (Operand.getNode()->getOperand(0).getValueType().getScalarType() 2421 .bitsLT(VT.getScalarType())) 2422 return getNode(OpOpcode, DL, VT, Operand.getNode()->getOperand(0)); 2423 else if (Operand.getNode()->getOperand(0).getValueType().bitsGT(VT)) 2424 return getNode(ISD::TRUNCATE, DL, VT, Operand.getNode()->getOperand(0)); 2425 else 2426 return Operand.getNode()->getOperand(0); 2427 } 2428 break; 2429 case ISD::BIT_CONVERT: 2430 // Basic sanity checking. 2431 assert(VT.getSizeInBits() == Operand.getValueType().getSizeInBits() 2432 && "Cannot BIT_CONVERT between types of different sizes!"); 2433 if (VT == Operand.getValueType()) return Operand; // noop conversion. 2434 if (OpOpcode == ISD::BIT_CONVERT) // bitconv(bitconv(x)) -> bitconv(x) 2435 return getNode(ISD::BIT_CONVERT, DL, VT, Operand.getOperand(0)); 2436 if (OpOpcode == ISD::UNDEF) 2437 return getUNDEF(VT); 2438 break; 2439 case ISD::SCALAR_TO_VECTOR: 2440 assert(VT.isVector() && !Operand.getValueType().isVector() && 2441 (VT.getVectorElementType() == Operand.getValueType() || 2442 (VT.getVectorElementType().isInteger() && 2443 Operand.getValueType().isInteger() && 2444 VT.getVectorElementType().bitsLE(Operand.getValueType()))) && 2445 "Illegal SCALAR_TO_VECTOR node!"); 2446 if (OpOpcode == ISD::UNDEF) 2447 return getUNDEF(VT); 2448 // scalar_to_vector(extract_vector_elt V, 0) -> V, top bits are undefined. 2449 if (OpOpcode == ISD::EXTRACT_VECTOR_ELT && 2450 isa<ConstantSDNode>(Operand.getOperand(1)) && 2451 Operand.getConstantOperandVal(1) == 0 && 2452 Operand.getOperand(0).getValueType() == VT) 2453 return Operand.getOperand(0); 2454 break; 2455 case ISD::FNEG: 2456 // -(X-Y) -> (Y-X) is unsafe because when X==Y, -0.0 != +0.0 2457 if (UnsafeFPMath && OpOpcode == ISD::FSUB) 2458 return getNode(ISD::FSUB, DL, VT, Operand.getNode()->getOperand(1), 2459 Operand.getNode()->getOperand(0)); 2460 if (OpOpcode == ISD::FNEG) // --X -> X 2461 return Operand.getNode()->getOperand(0); 2462 break; 2463 case ISD::FABS: 2464 if (OpOpcode == ISD::FNEG) // abs(-X) -> abs(X) 2465 return getNode(ISD::FABS, DL, VT, Operand.getNode()->getOperand(0)); 2466 break; 2467 } 2468 2469 SDNode *N; 2470 SDVTList VTs = getVTList(VT); 2471 if (VT != MVT::Flag) { // Don't CSE flag producing nodes 2472 FoldingSetNodeID ID; 2473 SDValue Ops[1] = { Operand }; 2474 AddNodeIDNode(ID, Opcode, VTs, Ops, 1); 2475 void *IP = 0; 2476 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 2477 return SDValue(E, 0); 2478 N = NodeAllocator.Allocate<UnarySDNode>(); 2479 new (N) UnarySDNode(Opcode, DL, VTs, Operand); 2480 CSEMap.InsertNode(N, IP); 2481 } else { 2482 N = NodeAllocator.Allocate<UnarySDNode>(); 2483 new (N) UnarySDNode(Opcode, DL, VTs, Operand); 2484 } 2485 2486 AllNodes.push_back(N); 2487#ifndef NDEBUG 2488 VerifyNode(N); 2489#endif 2490 return SDValue(N, 0); 2491} 2492 2493SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, 2494 EVT VT, 2495 ConstantSDNode *Cst1, 2496 ConstantSDNode *Cst2) { 2497 const APInt &C1 = Cst1->getAPIntValue(), &C2 = Cst2->getAPIntValue(); 2498 2499 switch (Opcode) { 2500 case ISD::ADD: return getConstant(C1 + C2, VT); 2501 case ISD::SUB: return getConstant(C1 - C2, VT); 2502 case ISD::MUL: return getConstant(C1 * C2, VT); 2503 case ISD::UDIV: 2504 if (C2.getBoolValue()) return getConstant(C1.udiv(C2), VT); 2505 break; 2506 case ISD::UREM: 2507 if (C2.getBoolValue()) return getConstant(C1.urem(C2), VT); 2508 break; 2509 case ISD::SDIV: 2510 if (C2.getBoolValue()) return getConstant(C1.sdiv(C2), VT); 2511 break; 2512 case ISD::SREM: 2513 if (C2.getBoolValue()) return getConstant(C1.srem(C2), VT); 2514 break; 2515 case ISD::AND: return getConstant(C1 & C2, VT); 2516 case ISD::OR: return getConstant(C1 | C2, VT); 2517 case ISD::XOR: return getConstant(C1 ^ C2, VT); 2518 case ISD::SHL: return getConstant(C1 << C2, VT); 2519 case ISD::SRL: return getConstant(C1.lshr(C2), VT); 2520 case ISD::SRA: return getConstant(C1.ashr(C2), VT); 2521 case ISD::ROTL: return getConstant(C1.rotl(C2), VT); 2522 case ISD::ROTR: return getConstant(C1.rotr(C2), VT); 2523 default: break; 2524 } 2525 2526 return SDValue(); 2527} 2528 2529SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 2530 SDValue N1, SDValue N2) { 2531 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 2532 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode()); 2533 switch (Opcode) { 2534 default: break; 2535 case ISD::TokenFactor: 2536 assert(VT == MVT::Other && N1.getValueType() == MVT::Other && 2537 N2.getValueType() == MVT::Other && "Invalid token factor!"); 2538 // Fold trivial token factors. 2539 if (N1.getOpcode() == ISD::EntryToken) return N2; 2540 if (N2.getOpcode() == ISD::EntryToken) return N1; 2541 if (N1 == N2) return N1; 2542 break; 2543 case ISD::CONCAT_VECTORS: 2544 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to 2545 // one big BUILD_VECTOR. 2546 if (N1.getOpcode() == ISD::BUILD_VECTOR && 2547 N2.getOpcode() == ISD::BUILD_VECTOR) { 2548 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end()); 2549 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end()); 2550 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size()); 2551 } 2552 break; 2553 case ISD::AND: 2554 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() && 2555 N1.getValueType() == VT && "Binary operator types must match!"); 2556 // (X & 0) -> 0. This commonly occurs when legalizing i64 values, so it's 2557 // worth handling here. 2558 if (N2C && N2C->isNullValue()) 2559 return N2; 2560 if (N2C && N2C->isAllOnesValue()) // X & -1 -> X 2561 return N1; 2562 break; 2563 case ISD::OR: 2564 case ISD::XOR: 2565 case ISD::ADD: 2566 case ISD::SUB: 2567 assert(VT.isInteger() && N1.getValueType() == N2.getValueType() && 2568 N1.getValueType() == VT && "Binary operator types must match!"); 2569 // (X ^|+- 0) -> X. This commonly occurs when legalizing i64 values, so 2570 // it's worth handling here. 2571 if (N2C && N2C->isNullValue()) 2572 return N1; 2573 break; 2574 case ISD::UDIV: 2575 case ISD::UREM: 2576 case ISD::MULHU: 2577 case ISD::MULHS: 2578 case ISD::MUL: 2579 case ISD::SDIV: 2580 case ISD::SREM: 2581 assert(VT.isInteger() && "This operator does not apply to FP types!"); 2582 // fall through 2583 case ISD::FADD: 2584 case ISD::FSUB: 2585 case ISD::FMUL: 2586 case ISD::FDIV: 2587 case ISD::FREM: 2588 if (UnsafeFPMath) { 2589 if (Opcode == ISD::FADD) { 2590 // 0+x --> x 2591 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) 2592 if (CFP->getValueAPF().isZero()) 2593 return N2; 2594 // x+0 --> x 2595 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2)) 2596 if (CFP->getValueAPF().isZero()) 2597 return N1; 2598 } else if (Opcode == ISD::FSUB) { 2599 // x-0 --> x 2600 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N2)) 2601 if (CFP->getValueAPF().isZero()) 2602 return N1; 2603 } 2604 } 2605 assert(N1.getValueType() == N2.getValueType() && 2606 N1.getValueType() == VT && "Binary operator types must match!"); 2607 break; 2608 case ISD::FCOPYSIGN: // N1 and result must match. N1/N2 need not match. 2609 assert(N1.getValueType() == VT && 2610 N1.getValueType().isFloatingPoint() && 2611 N2.getValueType().isFloatingPoint() && 2612 "Invalid FCOPYSIGN!"); 2613 break; 2614 case ISD::SHL: 2615 case ISD::SRA: 2616 case ISD::SRL: 2617 case ISD::ROTL: 2618 case ISD::ROTR: 2619 assert(VT == N1.getValueType() && 2620 "Shift operators return type must be the same as their first arg"); 2621 assert(VT.isInteger() && N2.getValueType().isInteger() && 2622 "Shifts only work on integers"); 2623 2624 // Always fold shifts of i1 values so the code generator doesn't need to 2625 // handle them. Since we know the size of the shift has to be less than the 2626 // size of the value, the shift/rotate count is guaranteed to be zero. 2627 if (VT == MVT::i1) 2628 return N1; 2629 break; 2630 case ISD::FP_ROUND_INREG: { 2631 EVT EVT = cast<VTSDNode>(N2)->getVT(); 2632 assert(VT == N1.getValueType() && "Not an inreg round!"); 2633 assert(VT.isFloatingPoint() && EVT.isFloatingPoint() && 2634 "Cannot FP_ROUND_INREG integer types"); 2635 assert(EVT.bitsLE(VT) && "Not rounding down!"); 2636 if (cast<VTSDNode>(N2)->getVT() == VT) return N1; // Not actually rounding. 2637 break; 2638 } 2639 case ISD::FP_ROUND: 2640 assert(VT.isFloatingPoint() && 2641 N1.getValueType().isFloatingPoint() && 2642 VT.bitsLE(N1.getValueType()) && 2643 isa<ConstantSDNode>(N2) && "Invalid FP_ROUND!"); 2644 if (N1.getValueType() == VT) return N1; // noop conversion. 2645 break; 2646 case ISD::AssertSext: 2647 case ISD::AssertZext: { 2648 EVT EVT = cast<VTSDNode>(N2)->getVT(); 2649 assert(VT == N1.getValueType() && "Not an inreg extend!"); 2650 assert(VT.isInteger() && EVT.isInteger() && 2651 "Cannot *_EXTEND_INREG FP types"); 2652 assert(!EVT.isVector() && 2653 "AssertSExt/AssertZExt type should be the vector element type " 2654 "rather than the vector type!"); 2655 assert(EVT.bitsLE(VT) && "Not extending!"); 2656 if (VT == EVT) return N1; // noop assertion. 2657 break; 2658 } 2659 case ISD::SIGN_EXTEND_INREG: { 2660 EVT EVT = cast<VTSDNode>(N2)->getVT(); 2661 assert(VT == N1.getValueType() && "Not an inreg extend!"); 2662 assert(VT.isInteger() && EVT.isInteger() && 2663 "Cannot *_EXTEND_INREG FP types"); 2664 assert(!EVT.isVector() && 2665 "SIGN_EXTEND_INREG type should be the vector element type rather " 2666 "than the vector type!"); 2667 assert(EVT.bitsLE(VT.getScalarType()) && "Not extending!"); 2668 if (EVT == VT) return N1; // Not actually extending 2669 2670 if (N1C) { 2671 APInt Val = N1C->getAPIntValue(); 2672 unsigned FromBits = EVT.getSizeInBits(); 2673 Val <<= Val.getBitWidth()-FromBits; 2674 Val = Val.ashr(Val.getBitWidth()-FromBits); 2675 return getConstant(Val, VT); 2676 } 2677 break; 2678 } 2679 case ISD::EXTRACT_VECTOR_ELT: 2680 // EXTRACT_VECTOR_ELT of an UNDEF is an UNDEF. 2681 if (N1.getOpcode() == ISD::UNDEF) 2682 return getUNDEF(VT); 2683 2684 // EXTRACT_VECTOR_ELT of CONCAT_VECTORS is often formed while lowering is 2685 // expanding copies of large vectors from registers. 2686 if (N2C && 2687 N1.getOpcode() == ISD::CONCAT_VECTORS && 2688 N1.getNumOperands() > 0) { 2689 unsigned Factor = 2690 N1.getOperand(0).getValueType().getVectorNumElements(); 2691 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, 2692 N1.getOperand(N2C->getZExtValue() / Factor), 2693 getConstant(N2C->getZExtValue() % Factor, 2694 N2.getValueType())); 2695 } 2696 2697 // EXTRACT_VECTOR_ELT of BUILD_VECTOR is often formed while lowering is 2698 // expanding large vector constants. 2699 if (N2C && N1.getOpcode() == ISD::BUILD_VECTOR) { 2700 SDValue Elt = N1.getOperand(N2C->getZExtValue()); 2701 EVT VEltTy = N1.getValueType().getVectorElementType(); 2702 if (Elt.getValueType() != VEltTy) { 2703 // If the vector element type is not legal, the BUILD_VECTOR operands 2704 // are promoted and implicitly truncated. Make that explicit here. 2705 Elt = getNode(ISD::TRUNCATE, DL, VEltTy, Elt); 2706 } 2707 if (VT != VEltTy) { 2708 // If the vector element type is not legal, the EXTRACT_VECTOR_ELT 2709 // result is implicitly extended. 2710 Elt = getNode(ISD::ANY_EXTEND, DL, VT, Elt); 2711 } 2712 return Elt; 2713 } 2714 2715 // EXTRACT_VECTOR_ELT of INSERT_VECTOR_ELT is often formed when vector 2716 // operations are lowered to scalars. 2717 if (N1.getOpcode() == ISD::INSERT_VECTOR_ELT) { 2718 // If the indices are the same, return the inserted element. 2719 if (N1.getOperand(2) == N2) 2720 return N1.getOperand(1); 2721 // If the indices are known different, extract the element from 2722 // the original vector. 2723 else if (isa<ConstantSDNode>(N1.getOperand(2)) && 2724 isa<ConstantSDNode>(N2)) 2725 return getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, N1.getOperand(0), N2); 2726 } 2727 break; 2728 case ISD::EXTRACT_ELEMENT: 2729 assert(N2C && (unsigned)N2C->getZExtValue() < 2 && "Bad EXTRACT_ELEMENT!"); 2730 assert(!N1.getValueType().isVector() && !VT.isVector() && 2731 (N1.getValueType().isInteger() == VT.isInteger()) && 2732 "Wrong types for EXTRACT_ELEMENT!"); 2733 2734 // EXTRACT_ELEMENT of BUILD_PAIR is often formed while legalize is expanding 2735 // 64-bit integers into 32-bit parts. Instead of building the extract of 2736 // the BUILD_PAIR, only to have legalize rip it apart, just do it now. 2737 if (N1.getOpcode() == ISD::BUILD_PAIR) 2738 return N1.getOperand(N2C->getZExtValue()); 2739 2740 // EXTRACT_ELEMENT of a constant int is also very common. 2741 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) { 2742 unsigned ElementSize = VT.getSizeInBits(); 2743 unsigned Shift = ElementSize * N2C->getZExtValue(); 2744 APInt ShiftedVal = C->getAPIntValue().lshr(Shift); 2745 return getConstant(ShiftedVal.trunc(ElementSize), VT); 2746 } 2747 break; 2748 case ISD::EXTRACT_SUBVECTOR: 2749 if (N1.getValueType() == VT) // Trivial extraction. 2750 return N1; 2751 break; 2752 } 2753 2754 if (N1C) { 2755 if (N2C) { 2756 SDValue SV = FoldConstantArithmetic(Opcode, VT, N1C, N2C); 2757 if (SV.getNode()) return SV; 2758 } else { // Cannonicalize constant to RHS if commutative 2759 if (isCommutativeBinOp(Opcode)) { 2760 std::swap(N1C, N2C); 2761 std::swap(N1, N2); 2762 } 2763 } 2764 } 2765 2766 // Constant fold FP operations. 2767 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1.getNode()); 2768 ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2.getNode()); 2769 if (N1CFP) { 2770 if (!N2CFP && isCommutativeBinOp(Opcode)) { 2771 // Cannonicalize constant to RHS if commutative 2772 std::swap(N1CFP, N2CFP); 2773 std::swap(N1, N2); 2774 } else if (N2CFP && VT != MVT::ppcf128) { 2775 APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF(); 2776 APFloat::opStatus s; 2777 switch (Opcode) { 2778 case ISD::FADD: 2779 s = V1.add(V2, APFloat::rmNearestTiesToEven); 2780 if (s != APFloat::opInvalidOp) 2781 return getConstantFP(V1, VT); 2782 break; 2783 case ISD::FSUB: 2784 s = V1.subtract(V2, APFloat::rmNearestTiesToEven); 2785 if (s!=APFloat::opInvalidOp) 2786 return getConstantFP(V1, VT); 2787 break; 2788 case ISD::FMUL: 2789 s = V1.multiply(V2, APFloat::rmNearestTiesToEven); 2790 if (s!=APFloat::opInvalidOp) 2791 return getConstantFP(V1, VT); 2792 break; 2793 case ISD::FDIV: 2794 s = V1.divide(V2, APFloat::rmNearestTiesToEven); 2795 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero) 2796 return getConstantFP(V1, VT); 2797 break; 2798 case ISD::FREM : 2799 s = V1.mod(V2, APFloat::rmNearestTiesToEven); 2800 if (s!=APFloat::opInvalidOp && s!=APFloat::opDivByZero) 2801 return getConstantFP(V1, VT); 2802 break; 2803 case ISD::FCOPYSIGN: 2804 V1.copySign(V2); 2805 return getConstantFP(V1, VT); 2806 default: break; 2807 } 2808 } 2809 } 2810 2811 // Canonicalize an UNDEF to the RHS, even over a constant. 2812 if (N1.getOpcode() == ISD::UNDEF) { 2813 if (isCommutativeBinOp(Opcode)) { 2814 std::swap(N1, N2); 2815 } else { 2816 switch (Opcode) { 2817 case ISD::FP_ROUND_INREG: 2818 case ISD::SIGN_EXTEND_INREG: 2819 case ISD::SUB: 2820 case ISD::FSUB: 2821 case ISD::FDIV: 2822 case ISD::FREM: 2823 case ISD::SRA: 2824 return N1; // fold op(undef, arg2) -> undef 2825 case ISD::UDIV: 2826 case ISD::SDIV: 2827 case ISD::UREM: 2828 case ISD::SREM: 2829 case ISD::SRL: 2830 case ISD::SHL: 2831 if (!VT.isVector()) 2832 return getConstant(0, VT); // fold op(undef, arg2) -> 0 2833 // For vectors, we can't easily build an all zero vector, just return 2834 // the LHS. 2835 return N2; 2836 } 2837 } 2838 } 2839 2840 // Fold a bunch of operators when the RHS is undef. 2841 if (N2.getOpcode() == ISD::UNDEF) { 2842 switch (Opcode) { 2843 case ISD::XOR: 2844 if (N1.getOpcode() == ISD::UNDEF) 2845 // Handle undef ^ undef -> 0 special case. This is a common 2846 // idiom (misuse). 2847 return getConstant(0, VT); 2848 // fallthrough 2849 case ISD::ADD: 2850 case ISD::ADDC: 2851 case ISD::ADDE: 2852 case ISD::SUB: 2853 case ISD::UDIV: 2854 case ISD::SDIV: 2855 case ISD::UREM: 2856 case ISD::SREM: 2857 return N2; // fold op(arg1, undef) -> undef 2858 case ISD::FADD: 2859 case ISD::FSUB: 2860 case ISD::FMUL: 2861 case ISD::FDIV: 2862 case ISD::FREM: 2863 if (UnsafeFPMath) 2864 return N2; 2865 break; 2866 case ISD::MUL: 2867 case ISD::AND: 2868 case ISD::SRL: 2869 case ISD::SHL: 2870 if (!VT.isVector()) 2871 return getConstant(0, VT); // fold op(arg1, undef) -> 0 2872 // For vectors, we can't easily build an all zero vector, just return 2873 // the LHS. 2874 return N1; 2875 case ISD::OR: 2876 if (!VT.isVector()) 2877 return getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT); 2878 // For vectors, we can't easily build an all one vector, just return 2879 // the LHS. 2880 return N1; 2881 case ISD::SRA: 2882 return N1; 2883 } 2884 } 2885 2886 // Memoize this node if possible. 2887 SDNode *N; 2888 SDVTList VTs = getVTList(VT); 2889 if (VT != MVT::Flag) { 2890 SDValue Ops[] = { N1, N2 }; 2891 FoldingSetNodeID ID; 2892 AddNodeIDNode(ID, Opcode, VTs, Ops, 2); 2893 void *IP = 0; 2894 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 2895 return SDValue(E, 0); 2896 N = NodeAllocator.Allocate<BinarySDNode>(); 2897 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2); 2898 CSEMap.InsertNode(N, IP); 2899 } else { 2900 N = NodeAllocator.Allocate<BinarySDNode>(); 2901 new (N) BinarySDNode(Opcode, DL, VTs, N1, N2); 2902 } 2903 2904 AllNodes.push_back(N); 2905#ifndef NDEBUG 2906 VerifyNode(N); 2907#endif 2908 return SDValue(N, 0); 2909} 2910 2911SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 2912 SDValue N1, SDValue N2, SDValue N3) { 2913 // Perform various simplifications. 2914 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode()); 2915 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode()); 2916 switch (Opcode) { 2917 case ISD::CONCAT_VECTORS: 2918 // A CONCAT_VECTOR with all operands BUILD_VECTOR can be simplified to 2919 // one big BUILD_VECTOR. 2920 if (N1.getOpcode() == ISD::BUILD_VECTOR && 2921 N2.getOpcode() == ISD::BUILD_VECTOR && 2922 N3.getOpcode() == ISD::BUILD_VECTOR) { 2923 SmallVector<SDValue, 16> Elts(N1.getNode()->op_begin(), N1.getNode()->op_end()); 2924 Elts.insert(Elts.end(), N2.getNode()->op_begin(), N2.getNode()->op_end()); 2925 Elts.insert(Elts.end(), N3.getNode()->op_begin(), N3.getNode()->op_end()); 2926 return getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], Elts.size()); 2927 } 2928 break; 2929 case ISD::SETCC: { 2930 // Use FoldSetCC to simplify SETCC's. 2931 SDValue Simp = FoldSetCC(VT, N1, N2, cast<CondCodeSDNode>(N3)->get(), DL); 2932 if (Simp.getNode()) return Simp; 2933 break; 2934 } 2935 case ISD::SELECT: 2936 if (N1C) { 2937 if (N1C->getZExtValue()) 2938 return N2; // select true, X, Y -> X 2939 else 2940 return N3; // select false, X, Y -> Y 2941 } 2942 2943 if (N2 == N3) return N2; // select C, X, X -> X 2944 break; 2945 case ISD::BRCOND: 2946 if (N2C) { 2947 if (N2C->getZExtValue()) // Unconditional branch 2948 return getNode(ISD::BR, DL, MVT::Other, N1, N3); 2949 else 2950 return N1; // Never-taken branch 2951 } 2952 break; 2953 case ISD::VECTOR_SHUFFLE: 2954 llvm_unreachable("should use getVectorShuffle constructor!"); 2955 break; 2956 case ISD::BIT_CONVERT: 2957 // Fold bit_convert nodes from a type to themselves. 2958 if (N1.getValueType() == VT) 2959 return N1; 2960 break; 2961 } 2962 2963 // Memoize node if it doesn't produce a flag. 2964 SDNode *N; 2965 SDVTList VTs = getVTList(VT); 2966 if (VT != MVT::Flag) { 2967 SDValue Ops[] = { N1, N2, N3 }; 2968 FoldingSetNodeID ID; 2969 AddNodeIDNode(ID, Opcode, VTs, Ops, 3); 2970 void *IP = 0; 2971 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 2972 return SDValue(E, 0); 2973 N = NodeAllocator.Allocate<TernarySDNode>(); 2974 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3); 2975 CSEMap.InsertNode(N, IP); 2976 } else { 2977 N = NodeAllocator.Allocate<TernarySDNode>(); 2978 new (N) TernarySDNode(Opcode, DL, VTs, N1, N2, N3); 2979 } 2980 AllNodes.push_back(N); 2981#ifndef NDEBUG 2982 VerifyNode(N); 2983#endif 2984 return SDValue(N, 0); 2985} 2986 2987SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 2988 SDValue N1, SDValue N2, SDValue N3, 2989 SDValue N4) { 2990 SDValue Ops[] = { N1, N2, N3, N4 }; 2991 return getNode(Opcode, DL, VT, Ops, 4); 2992} 2993 2994SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 2995 SDValue N1, SDValue N2, SDValue N3, 2996 SDValue N4, SDValue N5) { 2997 SDValue Ops[] = { N1, N2, N3, N4, N5 }; 2998 return getNode(Opcode, DL, VT, Ops, 5); 2999} 3000 3001/// getStackArgumentTokenFactor - Compute a TokenFactor to force all 3002/// the incoming stack arguments to be loaded from the stack. 3003SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) { 3004 SmallVector<SDValue, 8> ArgChains; 3005 3006 // Include the original chain at the beginning of the list. When this is 3007 // used by target LowerCall hooks, this helps legalize find the 3008 // CALLSEQ_BEGIN node. 3009 ArgChains.push_back(Chain); 3010 3011 // Add a chain value for each stack argument. 3012 for (SDNode::use_iterator U = getEntryNode().getNode()->use_begin(), 3013 UE = getEntryNode().getNode()->use_end(); U != UE; ++U) 3014 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U)) 3015 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr())) 3016 if (FI->getIndex() < 0) 3017 ArgChains.push_back(SDValue(L, 1)); 3018 3019 // Build a tokenfactor for all the chains. 3020 return getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other, 3021 &ArgChains[0], ArgChains.size()); 3022} 3023 3024/// getMemsetValue - Vectorized representation of the memset value 3025/// operand. 3026static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG, 3027 DebugLoc dl) { 3028 unsigned NumBits = VT.isVector() ? 3029 VT.getVectorElementType().getSizeInBits() : VT.getSizeInBits(); 3030 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) { 3031 APInt Val = APInt(NumBits, C->getZExtValue() & 255); 3032 unsigned Shift = 8; 3033 for (unsigned i = NumBits; i > 8; i >>= 1) { 3034 Val = (Val << Shift) | Val; 3035 Shift <<= 1; 3036 } 3037 if (VT.isInteger()) 3038 return DAG.getConstant(Val, VT); 3039 return DAG.getConstantFP(APFloat(Val), VT); 3040 } 3041 3042 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3043 Value = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Value); 3044 unsigned Shift = 8; 3045 for (unsigned i = NumBits; i > 8; i >>= 1) { 3046 Value = DAG.getNode(ISD::OR, dl, VT, 3047 DAG.getNode(ISD::SHL, dl, VT, Value, 3048 DAG.getConstant(Shift, 3049 TLI.getShiftAmountTy())), 3050 Value); 3051 Shift <<= 1; 3052 } 3053 3054 return Value; 3055} 3056 3057/// getMemsetStringVal - Similar to getMemsetValue. Except this is only 3058/// used when a memcpy is turned into a memset when the source is a constant 3059/// string ptr. 3060static SDValue getMemsetStringVal(EVT VT, DebugLoc dl, SelectionDAG &DAG, 3061 const TargetLowering &TLI, 3062 std::string &Str, unsigned Offset) { 3063 // Handle vector with all elements zero. 3064 if (Str.empty()) { 3065 if (VT.isInteger()) 3066 return DAG.getConstant(0, VT); 3067 unsigned NumElts = VT.getVectorNumElements(); 3068 MVT EltVT = (VT.getVectorElementType() == MVT::f32) ? MVT::i32 : MVT::i64; 3069 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, 3070 DAG.getConstant(0, 3071 EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts))); 3072 } 3073 3074 assert(!VT.isVector() && "Can't handle vector type here!"); 3075 unsigned NumBits = VT.getSizeInBits(); 3076 unsigned MSB = NumBits / 8; 3077 uint64_t Val = 0; 3078 if (TLI.isLittleEndian()) 3079 Offset = Offset + MSB - 1; 3080 for (unsigned i = 0; i != MSB; ++i) { 3081 Val = (Val << 8) | (unsigned char)Str[Offset]; 3082 Offset += TLI.isLittleEndian() ? -1 : 1; 3083 } 3084 return DAG.getConstant(Val, VT); 3085} 3086 3087/// getMemBasePlusOffset - Returns base and offset node for the 3088/// 3089static SDValue getMemBasePlusOffset(SDValue Base, unsigned Offset, 3090 SelectionDAG &DAG) { 3091 EVT VT = Base.getValueType(); 3092 return DAG.getNode(ISD::ADD, Base.getDebugLoc(), 3093 VT, Base, DAG.getConstant(Offset, VT)); 3094} 3095 3096/// isMemSrcFromString - Returns true if memcpy source is a string constant. 3097/// 3098static bool isMemSrcFromString(SDValue Src, std::string &Str) { 3099 unsigned SrcDelta = 0; 3100 GlobalAddressSDNode *G = NULL; 3101 if (Src.getOpcode() == ISD::GlobalAddress) 3102 G = cast<GlobalAddressSDNode>(Src); 3103 else if (Src.getOpcode() == ISD::ADD && 3104 Src.getOperand(0).getOpcode() == ISD::GlobalAddress && 3105 Src.getOperand(1).getOpcode() == ISD::Constant) { 3106 G = cast<GlobalAddressSDNode>(Src.getOperand(0)); 3107 SrcDelta = cast<ConstantSDNode>(Src.getOperand(1))->getZExtValue(); 3108 } 3109 if (!G) 3110 return false; 3111 3112 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal()); 3113 if (GV && GetConstantStringInfo(GV, Str, SrcDelta, false)) 3114 return true; 3115 3116 return false; 3117} 3118 3119/// MeetsMaxMemopRequirement - Determines if the number of memory ops required 3120/// to replace the memset / memcpy is below the threshold. It also returns the 3121/// types of the sequence of memory ops to perform memset / memcpy. 3122static 3123bool MeetsMaxMemopRequirement(std::vector<EVT> &MemOps, 3124 SDValue Dst, SDValue Src, 3125 unsigned Limit, uint64_t Size, unsigned &Align, 3126 std::string &Str, bool &isSrcStr, 3127 SelectionDAG &DAG, 3128 const TargetLowering &TLI) { 3129 isSrcStr = isMemSrcFromString(Src, Str); 3130 bool isSrcConst = isa<ConstantSDNode>(Src); 3131 EVT VT = TLI.getOptimalMemOpType(Size, Align, isSrcConst, isSrcStr, DAG); 3132 bool AllowUnalign = TLI.allowsUnalignedMemoryAccesses(VT); 3133 if (VT != MVT::iAny) { 3134 const Type *Ty = VT.getTypeForEVT(*DAG.getContext()); 3135 unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty); 3136 // If source is a string constant, this will require an unaligned load. 3137 if (NewAlign > Align && (isSrcConst || AllowUnalign)) { 3138 if (Dst.getOpcode() != ISD::FrameIndex) { 3139 // Can't change destination alignment. It requires a unaligned store. 3140 if (AllowUnalign) 3141 VT = MVT::iAny; 3142 } else { 3143 int FI = cast<FrameIndexSDNode>(Dst)->getIndex(); 3144 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 3145 if (MFI->isFixedObjectIndex(FI)) { 3146 // Can't change destination alignment. It requires a unaligned store. 3147 if (AllowUnalign) 3148 VT = MVT::iAny; 3149 } else { 3150 // Give the stack frame object a larger alignment if needed. 3151 if (MFI->getObjectAlignment(FI) < NewAlign) 3152 MFI->setObjectAlignment(FI, NewAlign); 3153 Align = NewAlign; 3154 } 3155 } 3156 } 3157 } 3158 3159 if (VT == MVT::iAny) { 3160 if (TLI.allowsUnalignedMemoryAccesses(MVT::i64)) { 3161 VT = MVT::i64; 3162 } else { 3163 switch (Align & 7) { 3164 case 0: VT = MVT::i64; break; 3165 case 4: VT = MVT::i32; break; 3166 case 2: VT = MVT::i16; break; 3167 default: VT = MVT::i8; break; 3168 } 3169 } 3170 3171 MVT LVT = MVT::i64; 3172 while (!TLI.isTypeLegal(LVT)) 3173 LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1); 3174 assert(LVT.isInteger()); 3175 3176 if (VT.bitsGT(LVT)) 3177 VT = LVT; 3178 } 3179 3180 unsigned NumMemOps = 0; 3181 while (Size != 0) { 3182 unsigned VTSize = VT.getSizeInBits() / 8; 3183 while (VTSize > Size) { 3184 // For now, only use non-vector load / store's for the left-over pieces. 3185 if (VT.isVector()) { 3186 VT = MVT::i64; 3187 while (!TLI.isTypeLegal(VT)) 3188 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1); 3189 VTSize = VT.getSizeInBits() / 8; 3190 } else { 3191 // This can result in a type that is not legal on the target, e.g. 3192 // 1 or 2 bytes on PPC. 3193 VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1); 3194 VTSize >>= 1; 3195 } 3196 } 3197 3198 if (++NumMemOps > Limit) 3199 return false; 3200 MemOps.push_back(VT); 3201 Size -= VTSize; 3202 } 3203 3204 return true; 3205} 3206 3207static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, 3208 SDValue Chain, SDValue Dst, 3209 SDValue Src, uint64_t Size, 3210 unsigned Align, bool AlwaysInline, 3211 const Value *DstSV, uint64_t DstSVOff, 3212 const Value *SrcSV, uint64_t SrcSVOff){ 3213 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3214 3215 // Expand memcpy to a series of load and store ops if the size operand falls 3216 // below a certain threshold. 3217 std::vector<EVT> MemOps; 3218 uint64_t Limit = -1ULL; 3219 if (!AlwaysInline) 3220 Limit = TLI.getMaxStoresPerMemcpy(); 3221 unsigned DstAlign = Align; // Destination alignment can change. 3222 std::string Str; 3223 bool CopyFromStr; 3224 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign, 3225 Str, CopyFromStr, DAG, TLI)) 3226 return SDValue(); 3227 3228 3229 bool isZeroStr = CopyFromStr && Str.empty(); 3230 SmallVector<SDValue, 8> OutChains; 3231 unsigned NumMemOps = MemOps.size(); 3232 uint64_t SrcOff = 0, DstOff = 0; 3233 for (unsigned i = 0; i != NumMemOps; ++i) { 3234 EVT VT = MemOps[i]; 3235 unsigned VTSize = VT.getSizeInBits() / 8; 3236 SDValue Value, Store; 3237 3238 if (CopyFromStr && (isZeroStr || !VT.isVector())) { 3239 // It's unlikely a store of a vector immediate can be done in a single 3240 // instruction. It would require a load from a constantpool first. 3241 // We also handle store a vector with all zero's. 3242 // FIXME: Handle other cases where store of vector immediate is done in 3243 // a single instruction. 3244 Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff); 3245 Store = DAG.getStore(Chain, dl, Value, 3246 getMemBasePlusOffset(Dst, DstOff, DAG), 3247 DstSV, DstSVOff + DstOff, false, DstAlign); 3248 } else { 3249 // The type might not be legal for the target. This should only happen 3250 // if the type is smaller than a legal type, as on PPC, so the right 3251 // thing to do is generate a LoadExt/StoreTrunc pair. These simplify 3252 // to Load/Store if NVT==VT. 3253 // FIXME does the case above also need this? 3254 EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT); 3255 assert(NVT.bitsGE(VT)); 3256 Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain, 3257 getMemBasePlusOffset(Src, SrcOff, DAG), 3258 SrcSV, SrcSVOff + SrcOff, VT, false, Align); 3259 Store = DAG.getTruncStore(Chain, dl, Value, 3260 getMemBasePlusOffset(Dst, DstOff, DAG), 3261 DstSV, DstSVOff + DstOff, VT, false, DstAlign); 3262 } 3263 OutChains.push_back(Store); 3264 SrcOff += VTSize; 3265 DstOff += VTSize; 3266 } 3267 3268 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3269 &OutChains[0], OutChains.size()); 3270} 3271 3272static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, 3273 SDValue Chain, SDValue Dst, 3274 SDValue Src, uint64_t Size, 3275 unsigned Align, bool AlwaysInline, 3276 const Value *DstSV, uint64_t DstSVOff, 3277 const Value *SrcSV, uint64_t SrcSVOff){ 3278 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3279 3280 // Expand memmove to a series of load and store ops if the size operand falls 3281 // below a certain threshold. 3282 std::vector<EVT> MemOps; 3283 uint64_t Limit = -1ULL; 3284 if (!AlwaysInline) 3285 Limit = TLI.getMaxStoresPerMemmove(); 3286 unsigned DstAlign = Align; // Destination alignment can change. 3287 std::string Str; 3288 bool CopyFromStr; 3289 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, Limit, Size, DstAlign, 3290 Str, CopyFromStr, DAG, TLI)) 3291 return SDValue(); 3292 3293 uint64_t SrcOff = 0, DstOff = 0; 3294 3295 SmallVector<SDValue, 8> LoadValues; 3296 SmallVector<SDValue, 8> LoadChains; 3297 SmallVector<SDValue, 8> OutChains; 3298 unsigned NumMemOps = MemOps.size(); 3299 for (unsigned i = 0; i < NumMemOps; i++) { 3300 EVT VT = MemOps[i]; 3301 unsigned VTSize = VT.getSizeInBits() / 8; 3302 SDValue Value, Store; 3303 3304 Value = DAG.getLoad(VT, dl, Chain, 3305 getMemBasePlusOffset(Src, SrcOff, DAG), 3306 SrcSV, SrcSVOff + SrcOff, false, Align); 3307 LoadValues.push_back(Value); 3308 LoadChains.push_back(Value.getValue(1)); 3309 SrcOff += VTSize; 3310 } 3311 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3312 &LoadChains[0], LoadChains.size()); 3313 OutChains.clear(); 3314 for (unsigned i = 0; i < NumMemOps; i++) { 3315 EVT VT = MemOps[i]; 3316 unsigned VTSize = VT.getSizeInBits() / 8; 3317 SDValue Value, Store; 3318 3319 Store = DAG.getStore(Chain, dl, LoadValues[i], 3320 getMemBasePlusOffset(Dst, DstOff, DAG), 3321 DstSV, DstSVOff + DstOff, false, DstAlign); 3322 OutChains.push_back(Store); 3323 DstOff += VTSize; 3324 } 3325 3326 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3327 &OutChains[0], OutChains.size()); 3328} 3329 3330static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl, 3331 SDValue Chain, SDValue Dst, 3332 SDValue Src, uint64_t Size, 3333 unsigned Align, 3334 const Value *DstSV, uint64_t DstSVOff) { 3335 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3336 3337 // Expand memset to a series of load/store ops if the size operand 3338 // falls below a certain threshold. 3339 std::vector<EVT> MemOps; 3340 std::string Str; 3341 bool CopyFromStr; 3342 if (!MeetsMaxMemopRequirement(MemOps, Dst, Src, TLI.getMaxStoresPerMemset(), 3343 Size, Align, Str, CopyFromStr, DAG, TLI)) 3344 return SDValue(); 3345 3346 SmallVector<SDValue, 8> OutChains; 3347 uint64_t DstOff = 0; 3348 3349 unsigned NumMemOps = MemOps.size(); 3350 for (unsigned i = 0; i < NumMemOps; i++) { 3351 EVT VT = MemOps[i]; 3352 unsigned VTSize = VT.getSizeInBits() / 8; 3353 SDValue Value = getMemsetValue(Src, VT, DAG, dl); 3354 SDValue Store = DAG.getStore(Chain, dl, Value, 3355 getMemBasePlusOffset(Dst, DstOff, DAG), 3356 DstSV, DstSVOff + DstOff); 3357 OutChains.push_back(Store); 3358 DstOff += VTSize; 3359 } 3360 3361 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 3362 &OutChains[0], OutChains.size()); 3363} 3364 3365SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst, 3366 SDValue Src, SDValue Size, 3367 unsigned Align, bool AlwaysInline, 3368 const Value *DstSV, uint64_t DstSVOff, 3369 const Value *SrcSV, uint64_t SrcSVOff) { 3370 3371 // Check to see if we should lower the memcpy to loads and stores first. 3372 // For cases within the target-specified limits, this is the best choice. 3373 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 3374 if (ConstantSize) { 3375 // Memcpy with size zero? Just return the original chain. 3376 if (ConstantSize->isNullValue()) 3377 return Chain; 3378 3379 SDValue Result = 3380 getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src, 3381 ConstantSize->getZExtValue(), 3382 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff); 3383 if (Result.getNode()) 3384 return Result; 3385 } 3386 3387 // Then check to see if we should lower the memcpy with target-specific 3388 // code. If the target chooses to do this, this is the next best. 3389 SDValue Result = 3390 TLI.EmitTargetCodeForMemcpy(*this, dl, Chain, Dst, Src, Size, Align, 3391 AlwaysInline, 3392 DstSV, DstSVOff, SrcSV, SrcSVOff); 3393 if (Result.getNode()) 3394 return Result; 3395 3396 // If we really need inline code and the target declined to provide it, 3397 // use a (potentially long) sequence of loads and stores. 3398 if (AlwaysInline) { 3399 assert(ConstantSize && "AlwaysInline requires a constant size!"); 3400 return getMemcpyLoadsAndStores(*this, dl, Chain, Dst, Src, 3401 ConstantSize->getZExtValue(), Align, true, 3402 DstSV, DstSVOff, SrcSV, SrcSVOff); 3403 } 3404 3405 // Emit a library call. 3406 TargetLowering::ArgListTy Args; 3407 TargetLowering::ArgListEntry Entry; 3408 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext()); 3409 Entry.Node = Dst; Args.push_back(Entry); 3410 Entry.Node = Src; Args.push_back(Entry); 3411 Entry.Node = Size; Args.push_back(Entry); 3412 // FIXME: pass in DebugLoc 3413 std::pair<SDValue,SDValue> CallResult = 3414 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()), 3415 false, false, false, false, 0, 3416 TLI.getLibcallCallingConv(RTLIB::MEMCPY), false, 3417 /*isReturnValueUsed=*/false, 3418 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMCPY), 3419 TLI.getPointerTy()), 3420 Args, *this, dl); 3421 return CallResult.second; 3422} 3423 3424SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst, 3425 SDValue Src, SDValue Size, 3426 unsigned Align, 3427 const Value *DstSV, uint64_t DstSVOff, 3428 const Value *SrcSV, uint64_t SrcSVOff) { 3429 3430 // Check to see if we should lower the memmove to loads and stores first. 3431 // For cases within the target-specified limits, this is the best choice. 3432 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 3433 if (ConstantSize) { 3434 // Memmove with size zero? Just return the original chain. 3435 if (ConstantSize->isNullValue()) 3436 return Chain; 3437 3438 SDValue Result = 3439 getMemmoveLoadsAndStores(*this, dl, Chain, Dst, Src, 3440 ConstantSize->getZExtValue(), 3441 Align, false, DstSV, DstSVOff, SrcSV, SrcSVOff); 3442 if (Result.getNode()) 3443 return Result; 3444 } 3445 3446 // Then check to see if we should lower the memmove with target-specific 3447 // code. If the target chooses to do this, this is the next best. 3448 SDValue Result = 3449 TLI.EmitTargetCodeForMemmove(*this, dl, Chain, Dst, Src, Size, Align, 3450 DstSV, DstSVOff, SrcSV, SrcSVOff); 3451 if (Result.getNode()) 3452 return Result; 3453 3454 // Emit a library call. 3455 TargetLowering::ArgListTy Args; 3456 TargetLowering::ArgListEntry Entry; 3457 Entry.Ty = TLI.getTargetData()->getIntPtrType(*getContext()); 3458 Entry.Node = Dst; Args.push_back(Entry); 3459 Entry.Node = Src; Args.push_back(Entry); 3460 Entry.Node = Size; Args.push_back(Entry); 3461 // FIXME: pass in DebugLoc 3462 std::pair<SDValue,SDValue> CallResult = 3463 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()), 3464 false, false, false, false, 0, 3465 TLI.getLibcallCallingConv(RTLIB::MEMMOVE), false, 3466 /*isReturnValueUsed=*/false, 3467 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMMOVE), 3468 TLI.getPointerTy()), 3469 Args, *this, dl); 3470 return CallResult.second; 3471} 3472 3473SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst, 3474 SDValue Src, SDValue Size, 3475 unsigned Align, 3476 const Value *DstSV, uint64_t DstSVOff) { 3477 3478 // Check to see if we should lower the memset to stores first. 3479 // For cases within the target-specified limits, this is the best choice. 3480 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 3481 if (ConstantSize) { 3482 // Memset with size zero? Just return the original chain. 3483 if (ConstantSize->isNullValue()) 3484 return Chain; 3485 3486 SDValue Result = 3487 getMemsetStores(*this, dl, Chain, Dst, Src, ConstantSize->getZExtValue(), 3488 Align, DstSV, DstSVOff); 3489 if (Result.getNode()) 3490 return Result; 3491 } 3492 3493 // Then check to see if we should lower the memset with target-specific 3494 // code. If the target chooses to do this, this is the next best. 3495 SDValue Result = 3496 TLI.EmitTargetCodeForMemset(*this, dl, Chain, Dst, Src, Size, Align, 3497 DstSV, DstSVOff); 3498 if (Result.getNode()) 3499 return Result; 3500 3501 // Emit a library call. 3502 const Type *IntPtrTy = TLI.getTargetData()->getIntPtrType(*getContext()); 3503 TargetLowering::ArgListTy Args; 3504 TargetLowering::ArgListEntry Entry; 3505 Entry.Node = Dst; Entry.Ty = IntPtrTy; 3506 Args.push_back(Entry); 3507 // Extend or truncate the argument to be an i32 value for the call. 3508 if (Src.getValueType().bitsGT(MVT::i32)) 3509 Src = getNode(ISD::TRUNCATE, dl, MVT::i32, Src); 3510 else 3511 Src = getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Src); 3512 Entry.Node = Src; 3513 Entry.Ty = Type::getInt32Ty(*getContext()); 3514 Entry.isSExt = true; 3515 Args.push_back(Entry); 3516 Entry.Node = Size; 3517 Entry.Ty = IntPtrTy; 3518 Entry.isSExt = false; 3519 Args.push_back(Entry); 3520 // FIXME: pass in DebugLoc 3521 std::pair<SDValue,SDValue> CallResult = 3522 TLI.LowerCallTo(Chain, Type::getVoidTy(*getContext()), 3523 false, false, false, false, 0, 3524 TLI.getLibcallCallingConv(RTLIB::MEMSET), false, 3525 /*isReturnValueUsed=*/false, 3526 getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET), 3527 TLI.getPointerTy()), 3528 Args, *this, dl); 3529 return CallResult.second; 3530} 3531 3532SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3533 SDValue Chain, 3534 SDValue Ptr, SDValue Cmp, 3535 SDValue Swp, const Value* PtrVal, 3536 unsigned Alignment) { 3537 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3538 Alignment = getEVTAlignment(MemVT); 3539 3540 // Check if the memory reference references a frame index 3541 if (!PtrVal) 3542 if (const FrameIndexSDNode *FI = 3543 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3544 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex()); 3545 3546 MachineFunction &MF = getMachineFunction(); 3547 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore; 3548 3549 // For now, atomics are considered to be volatile always. 3550 Flags |= MachineMemOperand::MOVolatile; 3551 3552 MachineMemOperand *MMO = 3553 MF.getMachineMemOperand(PtrVal, Flags, 0, 3554 MemVT.getStoreSize(), Alignment); 3555 3556 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Cmp, Swp, MMO); 3557} 3558 3559SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3560 SDValue Chain, 3561 SDValue Ptr, SDValue Cmp, 3562 SDValue Swp, MachineMemOperand *MMO) { 3563 assert(Opcode == ISD::ATOMIC_CMP_SWAP && "Invalid Atomic Op"); 3564 assert(Cmp.getValueType() == Swp.getValueType() && "Invalid Atomic Op Types"); 3565 3566 EVT VT = Cmp.getValueType(); 3567 3568 SDVTList VTs = getVTList(VT, MVT::Other); 3569 FoldingSetNodeID ID; 3570 ID.AddInteger(MemVT.getRawBits()); 3571 SDValue Ops[] = {Chain, Ptr, Cmp, Swp}; 3572 AddNodeIDNode(ID, Opcode, VTs, Ops, 4); 3573 void* IP = 0; 3574 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3575 cast<AtomicSDNode>(E)->refineAlignment(MMO); 3576 return SDValue(E, 0); 3577 } 3578 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>(); 3579 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Cmp, Swp, MMO); 3580 CSEMap.InsertNode(N, IP); 3581 AllNodes.push_back(N); 3582 return SDValue(N, 0); 3583} 3584 3585SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3586 SDValue Chain, 3587 SDValue Ptr, SDValue Val, 3588 const Value* PtrVal, 3589 unsigned Alignment) { 3590 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3591 Alignment = getEVTAlignment(MemVT); 3592 3593 // Check if the memory reference references a frame index 3594 if (!PtrVal) 3595 if (const FrameIndexSDNode *FI = 3596 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3597 PtrVal = PseudoSourceValue::getFixedStack(FI->getIndex()); 3598 3599 MachineFunction &MF = getMachineFunction(); 3600 unsigned Flags = MachineMemOperand::MOLoad | MachineMemOperand::MOStore; 3601 3602 // For now, atomics are considered to be volatile always. 3603 Flags |= MachineMemOperand::MOVolatile; 3604 3605 MachineMemOperand *MMO = 3606 MF.getMachineMemOperand(PtrVal, Flags, 0, 3607 MemVT.getStoreSize(), Alignment); 3608 3609 return getAtomic(Opcode, dl, MemVT, Chain, Ptr, Val, MMO); 3610} 3611 3612SDValue SelectionDAG::getAtomic(unsigned Opcode, DebugLoc dl, EVT MemVT, 3613 SDValue Chain, 3614 SDValue Ptr, SDValue Val, 3615 MachineMemOperand *MMO) { 3616 assert((Opcode == ISD::ATOMIC_LOAD_ADD || 3617 Opcode == ISD::ATOMIC_LOAD_SUB || 3618 Opcode == ISD::ATOMIC_LOAD_AND || 3619 Opcode == ISD::ATOMIC_LOAD_OR || 3620 Opcode == ISD::ATOMIC_LOAD_XOR || 3621 Opcode == ISD::ATOMIC_LOAD_NAND || 3622 Opcode == ISD::ATOMIC_LOAD_MIN || 3623 Opcode == ISD::ATOMIC_LOAD_MAX || 3624 Opcode == ISD::ATOMIC_LOAD_UMIN || 3625 Opcode == ISD::ATOMIC_LOAD_UMAX || 3626 Opcode == ISD::ATOMIC_SWAP) && 3627 "Invalid Atomic Op"); 3628 3629 EVT VT = Val.getValueType(); 3630 3631 SDVTList VTs = getVTList(VT, MVT::Other); 3632 FoldingSetNodeID ID; 3633 ID.AddInteger(MemVT.getRawBits()); 3634 SDValue Ops[] = {Chain, Ptr, Val}; 3635 AddNodeIDNode(ID, Opcode, VTs, Ops, 3); 3636 void* IP = 0; 3637 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3638 cast<AtomicSDNode>(E)->refineAlignment(MMO); 3639 return SDValue(E, 0); 3640 } 3641 SDNode* N = NodeAllocator.Allocate<AtomicSDNode>(); 3642 new (N) AtomicSDNode(Opcode, dl, VTs, MemVT, Chain, Ptr, Val, MMO); 3643 CSEMap.InsertNode(N, IP); 3644 AllNodes.push_back(N); 3645 return SDValue(N, 0); 3646} 3647 3648/// getMergeValues - Create a MERGE_VALUES node from the given operands. 3649/// Allowed to return something different (and simpler) if Simplify is true. 3650SDValue SelectionDAG::getMergeValues(const SDValue *Ops, unsigned NumOps, 3651 DebugLoc dl) { 3652 if (NumOps == 1) 3653 return Ops[0]; 3654 3655 SmallVector<EVT, 4> VTs; 3656 VTs.reserve(NumOps); 3657 for (unsigned i = 0; i < NumOps; ++i) 3658 VTs.push_back(Ops[i].getValueType()); 3659 return getNode(ISD::MERGE_VALUES, dl, getVTList(&VTs[0], NumOps), 3660 Ops, NumOps); 3661} 3662 3663SDValue 3664SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, 3665 const EVT *VTs, unsigned NumVTs, 3666 const SDValue *Ops, unsigned NumOps, 3667 EVT MemVT, const Value *srcValue, int SVOff, 3668 unsigned Align, bool Vol, 3669 bool ReadMem, bool WriteMem) { 3670 return getMemIntrinsicNode(Opcode, dl, makeVTList(VTs, NumVTs), Ops, NumOps, 3671 MemVT, srcValue, SVOff, Align, Vol, 3672 ReadMem, WriteMem); 3673} 3674 3675SDValue 3676SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList, 3677 const SDValue *Ops, unsigned NumOps, 3678 EVT MemVT, const Value *srcValue, int SVOff, 3679 unsigned Align, bool Vol, 3680 bool ReadMem, bool WriteMem) { 3681 if (Align == 0) // Ensure that codegen never sees alignment 0 3682 Align = getEVTAlignment(MemVT); 3683 3684 MachineFunction &MF = getMachineFunction(); 3685 unsigned Flags = 0; 3686 if (WriteMem) 3687 Flags |= MachineMemOperand::MOStore; 3688 if (ReadMem) 3689 Flags |= MachineMemOperand::MOLoad; 3690 if (Vol) 3691 Flags |= MachineMemOperand::MOVolatile; 3692 MachineMemOperand *MMO = 3693 MF.getMachineMemOperand(srcValue, Flags, SVOff, 3694 MemVT.getStoreSize(), Align); 3695 3696 return getMemIntrinsicNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO); 3697} 3698 3699SDValue 3700SelectionDAG::getMemIntrinsicNode(unsigned Opcode, DebugLoc dl, SDVTList VTList, 3701 const SDValue *Ops, unsigned NumOps, 3702 EVT MemVT, MachineMemOperand *MMO) { 3703 assert((Opcode == ISD::INTRINSIC_VOID || 3704 Opcode == ISD::INTRINSIC_W_CHAIN || 3705 (Opcode <= INT_MAX && 3706 (int)Opcode >= ISD::FIRST_TARGET_MEMORY_OPCODE)) && 3707 "Opcode is not a memory-accessing opcode!"); 3708 3709 // Memoize the node unless it returns a flag. 3710 MemIntrinsicSDNode *N; 3711 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) { 3712 FoldingSetNodeID ID; 3713 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps); 3714 void *IP = 0; 3715 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3716 cast<MemIntrinsicSDNode>(E)->refineAlignment(MMO); 3717 return SDValue(E, 0); 3718 } 3719 3720 N = NodeAllocator.Allocate<MemIntrinsicSDNode>(); 3721 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO); 3722 CSEMap.InsertNode(N, IP); 3723 } else { 3724 N = NodeAllocator.Allocate<MemIntrinsicSDNode>(); 3725 new (N) MemIntrinsicSDNode(Opcode, dl, VTList, Ops, NumOps, MemVT, MMO); 3726 } 3727 AllNodes.push_back(N); 3728 return SDValue(N, 0); 3729} 3730 3731SDValue 3732SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl, 3733 ISD::LoadExtType ExtType, EVT VT, SDValue Chain, 3734 SDValue Ptr, SDValue Offset, 3735 const Value *SV, int SVOffset, EVT MemVT, 3736 bool isVolatile, unsigned Alignment) { 3737 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3738 Alignment = getEVTAlignment(VT); 3739 3740 // Check if the memory reference references a frame index 3741 if (!SV) 3742 if (const FrameIndexSDNode *FI = 3743 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3744 SV = PseudoSourceValue::getFixedStack(FI->getIndex()); 3745 3746 MachineFunction &MF = getMachineFunction(); 3747 unsigned Flags = MachineMemOperand::MOLoad; 3748 if (isVolatile) 3749 Flags |= MachineMemOperand::MOVolatile; 3750 MachineMemOperand *MMO = 3751 MF.getMachineMemOperand(SV, Flags, SVOffset, 3752 MemVT.getStoreSize(), Alignment); 3753 return getLoad(AM, dl, ExtType, VT, Chain, Ptr, Offset, MemVT, MMO); 3754} 3755 3756SDValue 3757SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl, 3758 ISD::LoadExtType ExtType, EVT VT, SDValue Chain, 3759 SDValue Ptr, SDValue Offset, EVT MemVT, 3760 MachineMemOperand *MMO) { 3761 if (VT == MemVT) { 3762 ExtType = ISD::NON_EXTLOAD; 3763 } else if (ExtType == ISD::NON_EXTLOAD) { 3764 assert(VT == MemVT && "Non-extending load from different memory type!"); 3765 } else { 3766 // Extending load. 3767 assert(MemVT.getScalarType().bitsLT(VT.getScalarType()) && 3768 "Should only be an extending load, not truncating!"); 3769 assert(VT.isInteger() == MemVT.isInteger() && 3770 "Cannot convert from FP to Int or Int -> FP!"); 3771 assert(VT.isVector() == MemVT.isVector() && 3772 "Cannot use trunc store to convert to or from a vector!"); 3773 assert((!VT.isVector() || 3774 VT.getVectorNumElements() == MemVT.getVectorNumElements()) && 3775 "Cannot use trunc store to change the number of vector elements!"); 3776 } 3777 3778 bool Indexed = AM != ISD::UNINDEXED; 3779 assert((Indexed || Offset.getOpcode() == ISD::UNDEF) && 3780 "Unindexed load with an offset!"); 3781 3782 SDVTList VTs = Indexed ? 3783 getVTList(VT, Ptr.getValueType(), MVT::Other) : getVTList(VT, MVT::Other); 3784 SDValue Ops[] = { Chain, Ptr, Offset }; 3785 FoldingSetNodeID ID; 3786 AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3); 3787 ID.AddInteger(MemVT.getRawBits()); 3788 ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile())); 3789 void *IP = 0; 3790 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3791 cast<LoadSDNode>(E)->refineAlignment(MMO); 3792 return SDValue(E, 0); 3793 } 3794 SDNode *N = NodeAllocator.Allocate<LoadSDNode>(); 3795 new (N) LoadSDNode(Ops, dl, VTs, AM, ExtType, MemVT, MMO); 3796 CSEMap.InsertNode(N, IP); 3797 AllNodes.push_back(N); 3798 return SDValue(N, 0); 3799} 3800 3801SDValue SelectionDAG::getLoad(EVT VT, DebugLoc dl, 3802 SDValue Chain, SDValue Ptr, 3803 const Value *SV, int SVOffset, 3804 bool isVolatile, unsigned Alignment) { 3805 SDValue Undef = getUNDEF(Ptr.getValueType()); 3806 return getLoad(ISD::UNINDEXED, dl, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef, 3807 SV, SVOffset, VT, isVolatile, Alignment); 3808} 3809 3810SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, EVT VT, 3811 SDValue Chain, SDValue Ptr, 3812 const Value *SV, 3813 int SVOffset, EVT MemVT, 3814 bool isVolatile, unsigned Alignment) { 3815 SDValue Undef = getUNDEF(Ptr.getValueType()); 3816 return getLoad(ISD::UNINDEXED, dl, ExtType, VT, Chain, Ptr, Undef, 3817 SV, SVOffset, MemVT, isVolatile, Alignment); 3818} 3819 3820SDValue 3821SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base, 3822 SDValue Offset, ISD::MemIndexedMode AM) { 3823 LoadSDNode *LD = cast<LoadSDNode>(OrigLoad); 3824 assert(LD->getOffset().getOpcode() == ISD::UNDEF && 3825 "Load is already a indexed load!"); 3826 return getLoad(AM, dl, LD->getExtensionType(), OrigLoad.getValueType(), 3827 LD->getChain(), Base, Offset, LD->getSrcValue(), 3828 LD->getSrcValueOffset(), LD->getMemoryVT(), 3829 LD->isVolatile(), LD->getAlignment()); 3830} 3831 3832SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, 3833 SDValue Ptr, const Value *SV, int SVOffset, 3834 bool isVolatile, unsigned Alignment) { 3835 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3836 Alignment = getEVTAlignment(Val.getValueType()); 3837 3838 // Check if the memory reference references a frame index 3839 if (!SV) 3840 if (const FrameIndexSDNode *FI = 3841 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3842 SV = PseudoSourceValue::getFixedStack(FI->getIndex()); 3843 3844 MachineFunction &MF = getMachineFunction(); 3845 unsigned Flags = MachineMemOperand::MOStore; 3846 if (isVolatile) 3847 Flags |= MachineMemOperand::MOVolatile; 3848 MachineMemOperand *MMO = 3849 MF.getMachineMemOperand(SV, Flags, SVOffset, 3850 Val.getValueType().getStoreSize(), Alignment); 3851 3852 return getStore(Chain, dl, Val, Ptr, MMO); 3853} 3854 3855SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, 3856 SDValue Ptr, MachineMemOperand *MMO) { 3857 EVT VT = Val.getValueType(); 3858 SDVTList VTs = getVTList(MVT::Other); 3859 SDValue Undef = getUNDEF(Ptr.getValueType()); 3860 SDValue Ops[] = { Chain, Val, Ptr, Undef }; 3861 FoldingSetNodeID ID; 3862 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); 3863 ID.AddInteger(VT.getRawBits()); 3864 ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile())); 3865 void *IP = 0; 3866 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3867 cast<StoreSDNode>(E)->refineAlignment(MMO); 3868 return SDValue(E, 0); 3869 } 3870 SDNode *N = NodeAllocator.Allocate<StoreSDNode>(); 3871 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, false, VT, MMO); 3872 CSEMap.InsertNode(N, IP); 3873 AllNodes.push_back(N); 3874 return SDValue(N, 0); 3875} 3876 3877SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val, 3878 SDValue Ptr, const Value *SV, 3879 int SVOffset, EVT SVT, 3880 bool isVolatile, unsigned Alignment) { 3881 if (Alignment == 0) // Ensure that codegen never sees alignment 0 3882 Alignment = getEVTAlignment(SVT); 3883 3884 // Check if the memory reference references a frame index 3885 if (!SV) 3886 if (const FrameIndexSDNode *FI = 3887 dyn_cast<const FrameIndexSDNode>(Ptr.getNode())) 3888 SV = PseudoSourceValue::getFixedStack(FI->getIndex()); 3889 3890 MachineFunction &MF = getMachineFunction(); 3891 unsigned Flags = MachineMemOperand::MOStore; 3892 if (isVolatile) 3893 Flags |= MachineMemOperand::MOVolatile; 3894 MachineMemOperand *MMO = 3895 MF.getMachineMemOperand(SV, Flags, SVOffset, SVT.getStoreSize(), Alignment); 3896 3897 return getTruncStore(Chain, dl, Val, Ptr, SVT, MMO); 3898} 3899 3900SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val, 3901 SDValue Ptr, EVT SVT, 3902 MachineMemOperand *MMO) { 3903 EVT VT = Val.getValueType(); 3904 3905 if (VT == SVT) 3906 return getStore(Chain, dl, Val, Ptr, MMO); 3907 3908 assert(SVT.getScalarType().bitsLT(VT.getScalarType()) && 3909 "Should only be a truncating store, not extending!"); 3910 assert(VT.isInteger() == SVT.isInteger() && 3911 "Can't do FP-INT conversion!"); 3912 assert(VT.isVector() == SVT.isVector() && 3913 "Cannot use trunc store to convert to or from a vector!"); 3914 assert((!VT.isVector() || 3915 VT.getVectorNumElements() == SVT.getVectorNumElements()) && 3916 "Cannot use trunc store to change the number of vector elements!"); 3917 3918 SDVTList VTs = getVTList(MVT::Other); 3919 SDValue Undef = getUNDEF(Ptr.getValueType()); 3920 SDValue Ops[] = { Chain, Val, Ptr, Undef }; 3921 FoldingSetNodeID ID; 3922 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); 3923 ID.AddInteger(SVT.getRawBits()); 3924 ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile())); 3925 void *IP = 0; 3926 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { 3927 cast<StoreSDNode>(E)->refineAlignment(MMO); 3928 return SDValue(E, 0); 3929 } 3930 SDNode *N = NodeAllocator.Allocate<StoreSDNode>(); 3931 new (N) StoreSDNode(Ops, dl, VTs, ISD::UNINDEXED, true, SVT, MMO); 3932 CSEMap.InsertNode(N, IP); 3933 AllNodes.push_back(N); 3934 return SDValue(N, 0); 3935} 3936 3937SDValue 3938SelectionDAG::getIndexedStore(SDValue OrigStore, DebugLoc dl, SDValue Base, 3939 SDValue Offset, ISD::MemIndexedMode AM) { 3940 StoreSDNode *ST = cast<StoreSDNode>(OrigStore); 3941 assert(ST->getOffset().getOpcode() == ISD::UNDEF && 3942 "Store is already a indexed store!"); 3943 SDVTList VTs = getVTList(Base.getValueType(), MVT::Other); 3944 SDValue Ops[] = { ST->getChain(), ST->getValue(), Base, Offset }; 3945 FoldingSetNodeID ID; 3946 AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); 3947 ID.AddInteger(ST->getMemoryVT().getRawBits()); 3948 ID.AddInteger(ST->getRawSubclassData()); 3949 void *IP = 0; 3950 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 3951 return SDValue(E, 0); 3952 SDNode *N = NodeAllocator.Allocate<StoreSDNode>(); 3953 new (N) StoreSDNode(Ops, dl, VTs, AM, 3954 ST->isTruncatingStore(), ST->getMemoryVT(), 3955 ST->getMemOperand()); 3956 CSEMap.InsertNode(N, IP); 3957 AllNodes.push_back(N); 3958 return SDValue(N, 0); 3959} 3960 3961SDValue SelectionDAG::getVAArg(EVT VT, DebugLoc dl, 3962 SDValue Chain, SDValue Ptr, 3963 SDValue SV) { 3964 SDValue Ops[] = { Chain, Ptr, SV }; 3965 return getNode(ISD::VAARG, dl, getVTList(VT, MVT::Other), Ops, 3); 3966} 3967 3968SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 3969 const SDUse *Ops, unsigned NumOps) { 3970 switch (NumOps) { 3971 case 0: return getNode(Opcode, DL, VT); 3972 case 1: return getNode(Opcode, DL, VT, Ops[0]); 3973 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]); 3974 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]); 3975 default: break; 3976 } 3977 3978 // Copy from an SDUse array into an SDValue array for use with 3979 // the regular getNode logic. 3980 SmallVector<SDValue, 8> NewOps(Ops, Ops + NumOps); 3981 return getNode(Opcode, DL, VT, &NewOps[0], NumOps); 3982} 3983 3984SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, EVT VT, 3985 const SDValue *Ops, unsigned NumOps) { 3986 switch (NumOps) { 3987 case 0: return getNode(Opcode, DL, VT); 3988 case 1: return getNode(Opcode, DL, VT, Ops[0]); 3989 case 2: return getNode(Opcode, DL, VT, Ops[0], Ops[1]); 3990 case 3: return getNode(Opcode, DL, VT, Ops[0], Ops[1], Ops[2]); 3991 default: break; 3992 } 3993 3994 switch (Opcode) { 3995 default: break; 3996 case ISD::SELECT_CC: { 3997 assert(NumOps == 5 && "SELECT_CC takes 5 operands!"); 3998 assert(Ops[0].getValueType() == Ops[1].getValueType() && 3999 "LHS and RHS of condition must have same type!"); 4000 assert(Ops[2].getValueType() == Ops[3].getValueType() && 4001 "True and False arms of SelectCC must have same type!"); 4002 assert(Ops[2].getValueType() == VT && 4003 "select_cc node must be of same type as true and false value!"); 4004 break; 4005 } 4006 case ISD::BR_CC: { 4007 assert(NumOps == 5 && "BR_CC takes 5 operands!"); 4008 assert(Ops[2].getValueType() == Ops[3].getValueType() && 4009 "LHS/RHS of comparison should match types!"); 4010 break; 4011 } 4012 } 4013 4014 // Memoize nodes. 4015 SDNode *N; 4016 SDVTList VTs = getVTList(VT); 4017 4018 if (VT != MVT::Flag) { 4019 FoldingSetNodeID ID; 4020 AddNodeIDNode(ID, Opcode, VTs, Ops, NumOps); 4021 void *IP = 0; 4022 4023 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4024 return SDValue(E, 0); 4025 4026 N = NodeAllocator.Allocate<SDNode>(); 4027 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps); 4028 CSEMap.InsertNode(N, IP); 4029 } else { 4030 N = NodeAllocator.Allocate<SDNode>(); 4031 new (N) SDNode(Opcode, DL, VTs, Ops, NumOps); 4032 } 4033 4034 AllNodes.push_back(N); 4035#ifndef NDEBUG 4036 VerifyNode(N); 4037#endif 4038 return SDValue(N, 0); 4039} 4040 4041SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, 4042 const std::vector<EVT> &ResultTys, 4043 const SDValue *Ops, unsigned NumOps) { 4044 return getNode(Opcode, DL, getVTList(&ResultTys[0], ResultTys.size()), 4045 Ops, NumOps); 4046} 4047 4048SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, 4049 const EVT *VTs, unsigned NumVTs, 4050 const SDValue *Ops, unsigned NumOps) { 4051 if (NumVTs == 1) 4052 return getNode(Opcode, DL, VTs[0], Ops, NumOps); 4053 return getNode(Opcode, DL, makeVTList(VTs, NumVTs), Ops, NumOps); 4054} 4055 4056SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4057 const SDValue *Ops, unsigned NumOps) { 4058 if (VTList.NumVTs == 1) 4059 return getNode(Opcode, DL, VTList.VTs[0], Ops, NumOps); 4060 4061#if 0 4062 switch (Opcode) { 4063 // FIXME: figure out how to safely handle things like 4064 // int foo(int x) { return 1 << (x & 255); } 4065 // int bar() { return foo(256); } 4066 case ISD::SRA_PARTS: 4067 case ISD::SRL_PARTS: 4068 case ISD::SHL_PARTS: 4069 if (N3.getOpcode() == ISD::SIGN_EXTEND_INREG && 4070 cast<VTSDNode>(N3.getOperand(1))->getVT() != MVT::i1) 4071 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0)); 4072 else if (N3.getOpcode() == ISD::AND) 4073 if (ConstantSDNode *AndRHS = dyn_cast<ConstantSDNode>(N3.getOperand(1))) { 4074 // If the and is only masking out bits that cannot effect the shift, 4075 // eliminate the and. 4076 unsigned NumBits = VT.getSizeInBits()*2; 4077 if ((AndRHS->getValue() & (NumBits-1)) == NumBits-1) 4078 return getNode(Opcode, DL, VT, N1, N2, N3.getOperand(0)); 4079 } 4080 break; 4081 } 4082#endif 4083 4084 // Memoize the node unless it returns a flag. 4085 SDNode *N; 4086 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) { 4087 FoldingSetNodeID ID; 4088 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps); 4089 void *IP = 0; 4090 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4091 return SDValue(E, 0); 4092 if (NumOps == 1) { 4093 N = NodeAllocator.Allocate<UnarySDNode>(); 4094 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]); 4095 } else if (NumOps == 2) { 4096 N = NodeAllocator.Allocate<BinarySDNode>(); 4097 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]); 4098 } else if (NumOps == 3) { 4099 N = NodeAllocator.Allocate<TernarySDNode>(); 4100 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]); 4101 } else { 4102 N = NodeAllocator.Allocate<SDNode>(); 4103 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps); 4104 } 4105 CSEMap.InsertNode(N, IP); 4106 } else { 4107 if (NumOps == 1) { 4108 N = NodeAllocator.Allocate<UnarySDNode>(); 4109 new (N) UnarySDNode(Opcode, DL, VTList, Ops[0]); 4110 } else if (NumOps == 2) { 4111 N = NodeAllocator.Allocate<BinarySDNode>(); 4112 new (N) BinarySDNode(Opcode, DL, VTList, Ops[0], Ops[1]); 4113 } else if (NumOps == 3) { 4114 N = NodeAllocator.Allocate<TernarySDNode>(); 4115 new (N) TernarySDNode(Opcode, DL, VTList, Ops[0], Ops[1], Ops[2]); 4116 } else { 4117 N = NodeAllocator.Allocate<SDNode>(); 4118 new (N) SDNode(Opcode, DL, VTList, Ops, NumOps); 4119 } 4120 } 4121 AllNodes.push_back(N); 4122#ifndef NDEBUG 4123 VerifyNode(N); 4124#endif 4125 return SDValue(N, 0); 4126} 4127 4128SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList) { 4129 return getNode(Opcode, DL, VTList, 0, 0); 4130} 4131 4132SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4133 SDValue N1) { 4134 SDValue Ops[] = { N1 }; 4135 return getNode(Opcode, DL, VTList, Ops, 1); 4136} 4137 4138SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4139 SDValue N1, SDValue N2) { 4140 SDValue Ops[] = { N1, N2 }; 4141 return getNode(Opcode, DL, VTList, Ops, 2); 4142} 4143 4144SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4145 SDValue N1, SDValue N2, SDValue N3) { 4146 SDValue Ops[] = { N1, N2, N3 }; 4147 return getNode(Opcode, DL, VTList, Ops, 3); 4148} 4149 4150SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4151 SDValue N1, SDValue N2, SDValue N3, 4152 SDValue N4) { 4153 SDValue Ops[] = { N1, N2, N3, N4 }; 4154 return getNode(Opcode, DL, VTList, Ops, 4); 4155} 4156 4157SDValue SelectionDAG::getNode(unsigned Opcode, DebugLoc DL, SDVTList VTList, 4158 SDValue N1, SDValue N2, SDValue N3, 4159 SDValue N4, SDValue N5) { 4160 SDValue Ops[] = { N1, N2, N3, N4, N5 }; 4161 return getNode(Opcode, DL, VTList, Ops, 5); 4162} 4163 4164SDVTList SelectionDAG::getVTList(EVT VT) { 4165 return makeVTList(SDNode::getValueTypeList(VT), 1); 4166} 4167 4168SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2) { 4169 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4170 E = VTList.rend(); I != E; ++I) 4171 if (I->NumVTs == 2 && I->VTs[0] == VT1 && I->VTs[1] == VT2) 4172 return *I; 4173 4174 EVT *Array = Allocator.Allocate<EVT>(2); 4175 Array[0] = VT1; 4176 Array[1] = VT2; 4177 SDVTList Result = makeVTList(Array, 2); 4178 VTList.push_back(Result); 4179 return Result; 4180} 4181 4182SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3) { 4183 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4184 E = VTList.rend(); I != E; ++I) 4185 if (I->NumVTs == 3 && I->VTs[0] == VT1 && I->VTs[1] == VT2 && 4186 I->VTs[2] == VT3) 4187 return *I; 4188 4189 EVT *Array = Allocator.Allocate<EVT>(3); 4190 Array[0] = VT1; 4191 Array[1] = VT2; 4192 Array[2] = VT3; 4193 SDVTList Result = makeVTList(Array, 3); 4194 VTList.push_back(Result); 4195 return Result; 4196} 4197 4198SDVTList SelectionDAG::getVTList(EVT VT1, EVT VT2, EVT VT3, EVT VT4) { 4199 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4200 E = VTList.rend(); I != E; ++I) 4201 if (I->NumVTs == 4 && I->VTs[0] == VT1 && I->VTs[1] == VT2 && 4202 I->VTs[2] == VT3 && I->VTs[3] == VT4) 4203 return *I; 4204 4205 EVT *Array = Allocator.Allocate<EVT>(4); 4206 Array[0] = VT1; 4207 Array[1] = VT2; 4208 Array[2] = VT3; 4209 Array[3] = VT4; 4210 SDVTList Result = makeVTList(Array, 4); 4211 VTList.push_back(Result); 4212 return Result; 4213} 4214 4215SDVTList SelectionDAG::getVTList(const EVT *VTs, unsigned NumVTs) { 4216 switch (NumVTs) { 4217 case 0: llvm_unreachable("Cannot have nodes without results!"); 4218 case 1: return getVTList(VTs[0]); 4219 case 2: return getVTList(VTs[0], VTs[1]); 4220 case 3: return getVTList(VTs[0], VTs[1], VTs[2]); 4221 default: break; 4222 } 4223 4224 for (std::vector<SDVTList>::reverse_iterator I = VTList.rbegin(), 4225 E = VTList.rend(); I != E; ++I) { 4226 if (I->NumVTs != NumVTs || VTs[0] != I->VTs[0] || VTs[1] != I->VTs[1]) 4227 continue; 4228 4229 bool NoMatch = false; 4230 for (unsigned i = 2; i != NumVTs; ++i) 4231 if (VTs[i] != I->VTs[i]) { 4232 NoMatch = true; 4233 break; 4234 } 4235 if (!NoMatch) 4236 return *I; 4237 } 4238 4239 EVT *Array = Allocator.Allocate<EVT>(NumVTs); 4240 std::copy(VTs, VTs+NumVTs, Array); 4241 SDVTList Result = makeVTList(Array, NumVTs); 4242 VTList.push_back(Result); 4243 return Result; 4244} 4245 4246 4247/// UpdateNodeOperands - *Mutate* the specified node in-place to have the 4248/// specified operands. If the resultant node already exists in the DAG, 4249/// this does not modify the specified node, instead it returns the node that 4250/// already exists. If the resultant node does not exist in the DAG, the 4251/// input node is returned. As a degenerate case, if you specify the same 4252/// input operands as the node already has, the input node is returned. 4253SDValue SelectionDAG::UpdateNodeOperands(SDValue InN, SDValue Op) { 4254 SDNode *N = InN.getNode(); 4255 assert(N->getNumOperands() == 1 && "Update with wrong number of operands"); 4256 4257 // Check to see if there is no change. 4258 if (Op == N->getOperand(0)) return InN; 4259 4260 // See if the modified node already exists. 4261 void *InsertPos = 0; 4262 if (SDNode *Existing = FindModifiedNodeSlot(N, Op, InsertPos)) 4263 return SDValue(Existing, InN.getResNo()); 4264 4265 // Nope it doesn't. Remove the node from its current place in the maps. 4266 if (InsertPos) 4267 if (!RemoveNodeFromCSEMaps(N)) 4268 InsertPos = 0; 4269 4270 // Now we update the operands. 4271 N->OperandList[0].set(Op); 4272 4273 // If this gets put into a CSE map, add it. 4274 if (InsertPos) CSEMap.InsertNode(N, InsertPos); 4275 return InN; 4276} 4277 4278SDValue SelectionDAG:: 4279UpdateNodeOperands(SDValue InN, SDValue Op1, SDValue Op2) { 4280 SDNode *N = InN.getNode(); 4281 assert(N->getNumOperands() == 2 && "Update with wrong number of operands"); 4282 4283 // Check to see if there is no change. 4284 if (Op1 == N->getOperand(0) && Op2 == N->getOperand(1)) 4285 return InN; // No operands changed, just return the input node. 4286 4287 // See if the modified node already exists. 4288 void *InsertPos = 0; 4289 if (SDNode *Existing = FindModifiedNodeSlot(N, Op1, Op2, InsertPos)) 4290 return SDValue(Existing, InN.getResNo()); 4291 4292 // Nope it doesn't. Remove the node from its current place in the maps. 4293 if (InsertPos) 4294 if (!RemoveNodeFromCSEMaps(N)) 4295 InsertPos = 0; 4296 4297 // Now we update the operands. 4298 if (N->OperandList[0] != Op1) 4299 N->OperandList[0].set(Op1); 4300 if (N->OperandList[1] != Op2) 4301 N->OperandList[1].set(Op2); 4302 4303 // If this gets put into a CSE map, add it. 4304 if (InsertPos) CSEMap.InsertNode(N, InsertPos); 4305 return InN; 4306} 4307 4308SDValue SelectionDAG:: 4309UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, SDValue Op3) { 4310 SDValue Ops[] = { Op1, Op2, Op3 }; 4311 return UpdateNodeOperands(N, Ops, 3); 4312} 4313 4314SDValue SelectionDAG:: 4315UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, 4316 SDValue Op3, SDValue Op4) { 4317 SDValue Ops[] = { Op1, Op2, Op3, Op4 }; 4318 return UpdateNodeOperands(N, Ops, 4); 4319} 4320 4321SDValue SelectionDAG:: 4322UpdateNodeOperands(SDValue N, SDValue Op1, SDValue Op2, 4323 SDValue Op3, SDValue Op4, SDValue Op5) { 4324 SDValue Ops[] = { Op1, Op2, Op3, Op4, Op5 }; 4325 return UpdateNodeOperands(N, Ops, 5); 4326} 4327 4328SDValue SelectionDAG:: 4329UpdateNodeOperands(SDValue InN, const SDValue *Ops, unsigned NumOps) { 4330 SDNode *N = InN.getNode(); 4331 assert(N->getNumOperands() == NumOps && 4332 "Update with wrong number of operands"); 4333 4334 // Check to see if there is no change. 4335 bool AnyChange = false; 4336 for (unsigned i = 0; i != NumOps; ++i) { 4337 if (Ops[i] != N->getOperand(i)) { 4338 AnyChange = true; 4339 break; 4340 } 4341 } 4342 4343 // No operands changed, just return the input node. 4344 if (!AnyChange) return InN; 4345 4346 // See if the modified node already exists. 4347 void *InsertPos = 0; 4348 if (SDNode *Existing = FindModifiedNodeSlot(N, Ops, NumOps, InsertPos)) 4349 return SDValue(Existing, InN.getResNo()); 4350 4351 // Nope it doesn't. Remove the node from its current place in the maps. 4352 if (InsertPos) 4353 if (!RemoveNodeFromCSEMaps(N)) 4354 InsertPos = 0; 4355 4356 // Now we update the operands. 4357 for (unsigned i = 0; i != NumOps; ++i) 4358 if (N->OperandList[i] != Ops[i]) 4359 N->OperandList[i].set(Ops[i]); 4360 4361 // If this gets put into a CSE map, add it. 4362 if (InsertPos) CSEMap.InsertNode(N, InsertPos); 4363 return InN; 4364} 4365 4366/// DropOperands - Release the operands and set this node to have 4367/// zero operands. 4368void SDNode::DropOperands() { 4369 // Unlike the code in MorphNodeTo that does this, we don't need to 4370 // watch for dead nodes here. 4371 for (op_iterator I = op_begin(), E = op_end(); I != E; ) { 4372 SDUse &Use = *I++; 4373 Use.set(SDValue()); 4374 } 4375} 4376 4377/// SelectNodeTo - These are wrappers around MorphNodeTo that accept a 4378/// machine opcode. 4379/// 4380SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4381 EVT VT) { 4382 SDVTList VTs = getVTList(VT); 4383 return SelectNodeTo(N, MachineOpc, VTs, 0, 0); 4384} 4385 4386SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4387 EVT VT, SDValue Op1) { 4388 SDVTList VTs = getVTList(VT); 4389 SDValue Ops[] = { Op1 }; 4390 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1); 4391} 4392 4393SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4394 EVT VT, SDValue Op1, 4395 SDValue Op2) { 4396 SDVTList VTs = getVTList(VT); 4397 SDValue Ops[] = { Op1, Op2 }; 4398 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2); 4399} 4400 4401SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4402 EVT VT, SDValue Op1, 4403 SDValue Op2, SDValue Op3) { 4404 SDVTList VTs = getVTList(VT); 4405 SDValue Ops[] = { Op1, Op2, Op3 }; 4406 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3); 4407} 4408 4409SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4410 EVT VT, const SDValue *Ops, 4411 unsigned NumOps) { 4412 SDVTList VTs = getVTList(VT); 4413 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4414} 4415 4416SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4417 EVT VT1, EVT VT2, const SDValue *Ops, 4418 unsigned NumOps) { 4419 SDVTList VTs = getVTList(VT1, VT2); 4420 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4421} 4422 4423SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4424 EVT VT1, EVT VT2) { 4425 SDVTList VTs = getVTList(VT1, VT2); 4426 return SelectNodeTo(N, MachineOpc, VTs, (SDValue *)0, 0); 4427} 4428 4429SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4430 EVT VT1, EVT VT2, EVT VT3, 4431 const SDValue *Ops, unsigned NumOps) { 4432 SDVTList VTs = getVTList(VT1, VT2, VT3); 4433 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4434} 4435 4436SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4437 EVT VT1, EVT VT2, EVT VT3, EVT VT4, 4438 const SDValue *Ops, unsigned NumOps) { 4439 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4); 4440 return SelectNodeTo(N, MachineOpc, VTs, Ops, NumOps); 4441} 4442 4443SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4444 EVT VT1, EVT VT2, 4445 SDValue Op1) { 4446 SDVTList VTs = getVTList(VT1, VT2); 4447 SDValue Ops[] = { Op1 }; 4448 return SelectNodeTo(N, MachineOpc, VTs, Ops, 1); 4449} 4450 4451SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4452 EVT VT1, EVT VT2, 4453 SDValue Op1, SDValue Op2) { 4454 SDVTList VTs = getVTList(VT1, VT2); 4455 SDValue Ops[] = { Op1, Op2 }; 4456 return SelectNodeTo(N, MachineOpc, VTs, Ops, 2); 4457} 4458 4459SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4460 EVT VT1, EVT VT2, 4461 SDValue Op1, SDValue Op2, 4462 SDValue Op3) { 4463 SDVTList VTs = getVTList(VT1, VT2); 4464 SDValue Ops[] = { Op1, Op2, Op3 }; 4465 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3); 4466} 4467 4468SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4469 EVT VT1, EVT VT2, EVT VT3, 4470 SDValue Op1, SDValue Op2, 4471 SDValue Op3) { 4472 SDVTList VTs = getVTList(VT1, VT2, VT3); 4473 SDValue Ops[] = { Op1, Op2, Op3 }; 4474 return SelectNodeTo(N, MachineOpc, VTs, Ops, 3); 4475} 4476 4477SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, 4478 SDVTList VTs, const SDValue *Ops, 4479 unsigned NumOps) { 4480 return MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps); 4481} 4482 4483SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4484 EVT VT) { 4485 SDVTList VTs = getVTList(VT); 4486 return MorphNodeTo(N, Opc, VTs, 0, 0); 4487} 4488 4489SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4490 EVT VT, SDValue Op1) { 4491 SDVTList VTs = getVTList(VT); 4492 SDValue Ops[] = { Op1 }; 4493 return MorphNodeTo(N, Opc, VTs, Ops, 1); 4494} 4495 4496SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4497 EVT VT, SDValue Op1, 4498 SDValue Op2) { 4499 SDVTList VTs = getVTList(VT); 4500 SDValue Ops[] = { Op1, Op2 }; 4501 return MorphNodeTo(N, Opc, VTs, Ops, 2); 4502} 4503 4504SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4505 EVT VT, SDValue Op1, 4506 SDValue Op2, SDValue Op3) { 4507 SDVTList VTs = getVTList(VT); 4508 SDValue Ops[] = { Op1, Op2, Op3 }; 4509 return MorphNodeTo(N, Opc, VTs, Ops, 3); 4510} 4511 4512SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4513 EVT VT, const SDValue *Ops, 4514 unsigned NumOps) { 4515 SDVTList VTs = getVTList(VT); 4516 return MorphNodeTo(N, Opc, VTs, Ops, NumOps); 4517} 4518 4519SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4520 EVT VT1, EVT VT2, const SDValue *Ops, 4521 unsigned NumOps) { 4522 SDVTList VTs = getVTList(VT1, VT2); 4523 return MorphNodeTo(N, Opc, VTs, Ops, NumOps); 4524} 4525 4526SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4527 EVT VT1, EVT VT2) { 4528 SDVTList VTs = getVTList(VT1, VT2); 4529 return MorphNodeTo(N, Opc, VTs, (SDValue *)0, 0); 4530} 4531 4532SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4533 EVT VT1, EVT VT2, EVT VT3, 4534 const SDValue *Ops, unsigned NumOps) { 4535 SDVTList VTs = getVTList(VT1, VT2, VT3); 4536 return MorphNodeTo(N, Opc, VTs, Ops, NumOps); 4537} 4538 4539SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4540 EVT VT1, EVT VT2, 4541 SDValue Op1) { 4542 SDVTList VTs = getVTList(VT1, VT2); 4543 SDValue Ops[] = { Op1 }; 4544 return MorphNodeTo(N, Opc, VTs, Ops, 1); 4545} 4546 4547SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4548 EVT VT1, EVT VT2, 4549 SDValue Op1, SDValue Op2) { 4550 SDVTList VTs = getVTList(VT1, VT2); 4551 SDValue Ops[] = { Op1, Op2 }; 4552 return MorphNodeTo(N, Opc, VTs, Ops, 2); 4553} 4554 4555SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4556 EVT VT1, EVT VT2, 4557 SDValue Op1, SDValue Op2, 4558 SDValue Op3) { 4559 SDVTList VTs = getVTList(VT1, VT2); 4560 SDValue Ops[] = { Op1, Op2, Op3 }; 4561 return MorphNodeTo(N, Opc, VTs, Ops, 3); 4562} 4563 4564/// MorphNodeTo - These *mutate* the specified node to have the specified 4565/// return type, opcode, and operands. 4566/// 4567/// Note that MorphNodeTo returns the resultant node. If there is already a 4568/// node of the specified opcode and operands, it returns that node instead of 4569/// the current one. Note that the DebugLoc need not be the same. 4570/// 4571/// Using MorphNodeTo is faster than creating a new node and swapping it in 4572/// with ReplaceAllUsesWith both because it often avoids allocating a new 4573/// node, and because it doesn't require CSE recalculation for any of 4574/// the node's users. 4575/// 4576SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, 4577 SDVTList VTs, const SDValue *Ops, 4578 unsigned NumOps) { 4579 // If an identical node already exists, use it. 4580 void *IP = 0; 4581 if (VTs.VTs[VTs.NumVTs-1] != MVT::Flag) { 4582 FoldingSetNodeID ID; 4583 AddNodeIDNode(ID, Opc, VTs, Ops, NumOps); 4584 if (SDNode *ON = CSEMap.FindNodeOrInsertPos(ID, IP)) 4585 return ON; 4586 } 4587 4588 if (!RemoveNodeFromCSEMaps(N)) 4589 IP = 0; 4590 4591 // Start the morphing. 4592 N->NodeType = Opc; 4593 N->ValueList = VTs.VTs; 4594 N->NumValues = VTs.NumVTs; 4595 4596 // Clear the operands list, updating used nodes to remove this from their 4597 // use list. Keep track of any operands that become dead as a result. 4598 SmallPtrSet<SDNode*, 16> DeadNodeSet; 4599 for (SDNode::op_iterator I = N->op_begin(), E = N->op_end(); I != E; ) { 4600 SDUse &Use = *I++; 4601 SDNode *Used = Use.getNode(); 4602 Use.set(SDValue()); 4603 if (Used->use_empty()) 4604 DeadNodeSet.insert(Used); 4605 } 4606 4607 if (MachineSDNode *MN = dyn_cast<MachineSDNode>(N)) { 4608 // Initialize the memory references information. 4609 MN->setMemRefs(0, 0); 4610 // If NumOps is larger than the # of operands we can have in a 4611 // MachineSDNode, reallocate the operand list. 4612 if (NumOps > MN->NumOperands || !MN->OperandsNeedDelete) { 4613 if (MN->OperandsNeedDelete) 4614 delete[] MN->OperandList; 4615 if (NumOps > array_lengthof(MN->LocalOperands)) 4616 // We're creating a final node that will live unmorphed for the 4617 // remainder of the current SelectionDAG iteration, so we can allocate 4618 // the operands directly out of a pool with no recycling metadata. 4619 MN->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps), 4620 Ops, NumOps); 4621 else 4622 MN->InitOperands(MN->LocalOperands, Ops, NumOps); 4623 MN->OperandsNeedDelete = false; 4624 } else 4625 MN->InitOperands(MN->OperandList, Ops, NumOps); 4626 } else { 4627 // If NumOps is larger than the # of operands we currently have, reallocate 4628 // the operand list. 4629 if (NumOps > N->NumOperands) { 4630 if (N->OperandsNeedDelete) 4631 delete[] N->OperandList; 4632 N->InitOperands(new SDUse[NumOps], Ops, NumOps); 4633 N->OperandsNeedDelete = true; 4634 } else 4635 N->InitOperands(N->OperandList, Ops, NumOps); 4636 } 4637 4638 // Delete any nodes that are still dead after adding the uses for the 4639 // new operands. 4640 SmallVector<SDNode *, 16> DeadNodes; 4641 for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(), 4642 E = DeadNodeSet.end(); I != E; ++I) 4643 if ((*I)->use_empty()) 4644 DeadNodes.push_back(*I); 4645 RemoveDeadNodes(DeadNodes); 4646 4647 if (IP) 4648 CSEMap.InsertNode(N, IP); // Memoize the new node. 4649 return N; 4650} 4651 4652 4653/// getMachineNode - These are used for target selectors to create a new node 4654/// with specified return type(s), MachineInstr opcode, and operands. 4655/// 4656/// Note that getMachineNode returns the resultant node. If there is already a 4657/// node of the specified opcode and operands, it returns that node instead of 4658/// the current one. 4659MachineSDNode * 4660SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT) { 4661 SDVTList VTs = getVTList(VT); 4662 return getMachineNode(Opcode, dl, VTs, 0, 0); 4663} 4664 4665MachineSDNode * 4666SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, SDValue Op1) { 4667 SDVTList VTs = getVTList(VT); 4668 SDValue Ops[] = { Op1 }; 4669 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4670} 4671 4672MachineSDNode * 4673SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, 4674 SDValue Op1, SDValue Op2) { 4675 SDVTList VTs = getVTList(VT); 4676 SDValue Ops[] = { Op1, Op2 }; 4677 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4678} 4679 4680MachineSDNode * 4681SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, 4682 SDValue Op1, SDValue Op2, SDValue Op3) { 4683 SDVTList VTs = getVTList(VT); 4684 SDValue Ops[] = { Op1, Op2, Op3 }; 4685 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4686} 4687 4688MachineSDNode * 4689SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, 4690 const SDValue *Ops, unsigned NumOps) { 4691 SDVTList VTs = getVTList(VT); 4692 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4693} 4694 4695MachineSDNode * 4696SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, EVT VT2) { 4697 SDVTList VTs = getVTList(VT1, VT2); 4698 return getMachineNode(Opcode, dl, VTs, 0, 0); 4699} 4700 4701MachineSDNode * 4702SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4703 EVT VT1, EVT VT2, SDValue Op1) { 4704 SDVTList VTs = getVTList(VT1, VT2); 4705 SDValue Ops[] = { Op1 }; 4706 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4707} 4708 4709MachineSDNode * 4710SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4711 EVT VT1, EVT VT2, SDValue Op1, SDValue Op2) { 4712 SDVTList VTs = getVTList(VT1, VT2); 4713 SDValue Ops[] = { Op1, Op2 }; 4714 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4715} 4716 4717MachineSDNode * 4718SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4719 EVT VT1, EVT VT2, SDValue Op1, 4720 SDValue Op2, SDValue Op3) { 4721 SDVTList VTs = getVTList(VT1, VT2); 4722 SDValue Ops[] = { Op1, Op2, Op3 }; 4723 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4724} 4725 4726MachineSDNode * 4727SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4728 EVT VT1, EVT VT2, 4729 const SDValue *Ops, unsigned NumOps) { 4730 SDVTList VTs = getVTList(VT1, VT2); 4731 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4732} 4733 4734MachineSDNode * 4735SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4736 EVT VT1, EVT VT2, EVT VT3, 4737 SDValue Op1, SDValue Op2) { 4738 SDVTList VTs = getVTList(VT1, VT2, VT3); 4739 SDValue Ops[] = { Op1, Op2 }; 4740 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4741} 4742 4743MachineSDNode * 4744SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4745 EVT VT1, EVT VT2, EVT VT3, 4746 SDValue Op1, SDValue Op2, SDValue Op3) { 4747 SDVTList VTs = getVTList(VT1, VT2, VT3); 4748 SDValue Ops[] = { Op1, Op2, Op3 }; 4749 return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops)); 4750} 4751 4752MachineSDNode * 4753SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4754 EVT VT1, EVT VT2, EVT VT3, 4755 const SDValue *Ops, unsigned NumOps) { 4756 SDVTList VTs = getVTList(VT1, VT2, VT3); 4757 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4758} 4759 4760MachineSDNode * 4761SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, 4762 EVT VT2, EVT VT3, EVT VT4, 4763 const SDValue *Ops, unsigned NumOps) { 4764 SDVTList VTs = getVTList(VT1, VT2, VT3, VT4); 4765 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4766} 4767 4768MachineSDNode * 4769SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, 4770 const std::vector<EVT> &ResultTys, 4771 const SDValue *Ops, unsigned NumOps) { 4772 SDVTList VTs = getVTList(&ResultTys[0], ResultTys.size()); 4773 return getMachineNode(Opcode, dl, VTs, Ops, NumOps); 4774} 4775 4776MachineSDNode * 4777SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc DL, SDVTList VTs, 4778 const SDValue *Ops, unsigned NumOps) { 4779 bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Flag; 4780 MachineSDNode *N; 4781 void *IP; 4782 4783 if (DoCSE) { 4784 FoldingSetNodeID ID; 4785 AddNodeIDNode(ID, ~Opcode, VTs, Ops, NumOps); 4786 IP = 0; 4787 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4788 return cast<MachineSDNode>(E); 4789 } 4790 4791 // Allocate a new MachineSDNode. 4792 N = NodeAllocator.Allocate<MachineSDNode>(); 4793 new (N) MachineSDNode(~Opcode, DL, VTs); 4794 4795 // Initialize the operands list. 4796 if (NumOps > array_lengthof(N->LocalOperands)) 4797 // We're creating a final node that will live unmorphed for the 4798 // remainder of the current SelectionDAG iteration, so we can allocate 4799 // the operands directly out of a pool with no recycling metadata. 4800 N->InitOperands(OperandAllocator.Allocate<SDUse>(NumOps), 4801 Ops, NumOps); 4802 else 4803 N->InitOperands(N->LocalOperands, Ops, NumOps); 4804 N->OperandsNeedDelete = false; 4805 4806 if (DoCSE) 4807 CSEMap.InsertNode(N, IP); 4808 4809 AllNodes.push_back(N); 4810#ifndef NDEBUG 4811 VerifyNode(N); 4812#endif 4813 return N; 4814} 4815 4816/// getTargetExtractSubreg - A convenience function for creating 4817/// TargetInstrInfo::EXTRACT_SUBREG nodes. 4818SDValue 4819SelectionDAG::getTargetExtractSubreg(int SRIdx, DebugLoc DL, EVT VT, 4820 SDValue Operand) { 4821 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32); 4822 SDNode *Subreg = getMachineNode(TargetInstrInfo::EXTRACT_SUBREG, DL, 4823 VT, Operand, SRIdxVal); 4824 return SDValue(Subreg, 0); 4825} 4826 4827/// getTargetInsertSubreg - A convenience function for creating 4828/// TargetInstrInfo::INSERT_SUBREG nodes. 4829SDValue 4830SelectionDAG::getTargetInsertSubreg(int SRIdx, DebugLoc DL, EVT VT, 4831 SDValue Operand, SDValue Subreg) { 4832 SDValue SRIdxVal = getTargetConstant(SRIdx, MVT::i32); 4833 SDNode *Result = getMachineNode(TargetInstrInfo::INSERT_SUBREG, DL, 4834 VT, Operand, Subreg, SRIdxVal); 4835 return SDValue(Result, 0); 4836} 4837 4838/// getNodeIfExists - Get the specified node if it's already available, or 4839/// else return NULL. 4840SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList, 4841 const SDValue *Ops, unsigned NumOps) { 4842 if (VTList.VTs[VTList.NumVTs-1] != MVT::Flag) { 4843 FoldingSetNodeID ID; 4844 AddNodeIDNode(ID, Opcode, VTList, Ops, NumOps); 4845 void *IP = 0; 4846 if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) 4847 return E; 4848 } 4849 return NULL; 4850} 4851 4852/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. 4853/// This can cause recursive merging of nodes in the DAG. 4854/// 4855/// This version assumes From has a single result value. 4856/// 4857void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To, 4858 DAGUpdateListener *UpdateListener) { 4859 SDNode *From = FromN.getNode(); 4860 assert(From->getNumValues() == 1 && FromN.getResNo() == 0 && 4861 "Cannot replace with this method!"); 4862 assert(From != To.getNode() && "Cannot replace uses of with self"); 4863 4864 // Iterate over all the existing uses of From. New uses will be added 4865 // to the beginning of the use list, which we avoid visiting. 4866 // This specifically avoids visiting uses of From that arise while the 4867 // replacement is happening, because any such uses would be the result 4868 // of CSE: If an existing node looks like From after one of its operands 4869 // is replaced by To, we don't want to replace of all its users with To 4870 // too. See PR3018 for more info. 4871 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); 4872 while (UI != UE) { 4873 SDNode *User = *UI; 4874 4875 // This node is about to morph, remove its old self from the CSE maps. 4876 RemoveNodeFromCSEMaps(User); 4877 4878 // A user can appear in a use list multiple times, and when this 4879 // happens the uses are usually next to each other in the list. 4880 // To help reduce the number of CSE recomputations, process all 4881 // the uses of this user that we can find this way. 4882 do { 4883 SDUse &Use = UI.getUse(); 4884 ++UI; 4885 Use.set(To); 4886 } while (UI != UE && *UI == User); 4887 4888 // Now that we have modified User, add it back to the CSE maps. If it 4889 // already exists there, recursively merge the results together. 4890 AddModifiedNodeToCSEMaps(User, UpdateListener); 4891 } 4892} 4893 4894/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. 4895/// This can cause recursive merging of nodes in the DAG. 4896/// 4897/// This version assumes that for each value of From, there is a 4898/// corresponding value in To in the same position with the same type. 4899/// 4900void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To, 4901 DAGUpdateListener *UpdateListener) { 4902#ifndef NDEBUG 4903 for (unsigned i = 0, e = From->getNumValues(); i != e; ++i) 4904 assert((!From->hasAnyUseOfValue(i) || 4905 From->getValueType(i) == To->getValueType(i)) && 4906 "Cannot use this version of ReplaceAllUsesWith!"); 4907#endif 4908 4909 // Handle the trivial case. 4910 if (From == To) 4911 return; 4912 4913 // Iterate over just the existing users of From. See the comments in 4914 // the ReplaceAllUsesWith above. 4915 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); 4916 while (UI != UE) { 4917 SDNode *User = *UI; 4918 4919 // This node is about to morph, remove its old self from the CSE maps. 4920 RemoveNodeFromCSEMaps(User); 4921 4922 // A user can appear in a use list multiple times, and when this 4923 // happens the uses are usually next to each other in the list. 4924 // To help reduce the number of CSE recomputations, process all 4925 // the uses of this user that we can find this way. 4926 do { 4927 SDUse &Use = UI.getUse(); 4928 ++UI; 4929 Use.setNode(To); 4930 } while (UI != UE && *UI == User); 4931 4932 // Now that we have modified User, add it back to the CSE maps. If it 4933 // already exists there, recursively merge the results together. 4934 AddModifiedNodeToCSEMaps(User, UpdateListener); 4935 } 4936} 4937 4938/// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. 4939/// This can cause recursive merging of nodes in the DAG. 4940/// 4941/// This version can replace From with any result values. To must match the 4942/// number and types of values returned by From. 4943void SelectionDAG::ReplaceAllUsesWith(SDNode *From, 4944 const SDValue *To, 4945 DAGUpdateListener *UpdateListener) { 4946 if (From->getNumValues() == 1) // Handle the simple case efficiently. 4947 return ReplaceAllUsesWith(SDValue(From, 0), To[0], UpdateListener); 4948 4949 // Iterate over just the existing users of From. See the comments in 4950 // the ReplaceAllUsesWith above. 4951 SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); 4952 while (UI != UE) { 4953 SDNode *User = *UI; 4954 4955 // This node is about to morph, remove its old self from the CSE maps. 4956 RemoveNodeFromCSEMaps(User); 4957 4958 // A user can appear in a use list multiple times, and when this 4959 // happens the uses are usually next to each other in the list. 4960 // To help reduce the number of CSE recomputations, process all 4961 // the uses of this user that we can find this way. 4962 do { 4963 SDUse &Use = UI.getUse(); 4964 const SDValue &ToOp = To[Use.getResNo()]; 4965 ++UI; 4966 Use.set(ToOp); 4967 } while (UI != UE && *UI == User); 4968 4969 // Now that we have modified User, add it back to the CSE maps. If it 4970 // already exists there, recursively merge the results together. 4971 AddModifiedNodeToCSEMaps(User, UpdateListener); 4972 } 4973} 4974 4975/// ReplaceAllUsesOfValueWith - Replace any uses of From with To, leaving 4976/// uses of other values produced by From.getNode() alone. The Deleted 4977/// vector is handled the same way as for ReplaceAllUsesWith. 4978void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To, 4979 DAGUpdateListener *UpdateListener){ 4980 // Handle the really simple, really trivial case efficiently. 4981 if (From == To) return; 4982 4983 // Handle the simple, trivial, case efficiently. 4984 if (From.getNode()->getNumValues() == 1) { 4985 ReplaceAllUsesWith(From, To, UpdateListener); 4986 return; 4987 } 4988 4989 // Iterate over just the existing users of From. See the comments in 4990 // the ReplaceAllUsesWith above. 4991 SDNode::use_iterator UI = From.getNode()->use_begin(), 4992 UE = From.getNode()->use_end(); 4993 while (UI != UE) { 4994 SDNode *User = *UI; 4995 bool UserRemovedFromCSEMaps = false; 4996 4997 // A user can appear in a use list multiple times, and when this 4998 // happens the uses are usually next to each other in the list. 4999 // To help reduce the number of CSE recomputations, process all 5000 // the uses of this user that we can find this way. 5001 do { 5002 SDUse &Use = UI.getUse(); 5003 5004 // Skip uses of different values from the same node. 5005 if (Use.getResNo() != From.getResNo()) { 5006 ++UI; 5007 continue; 5008 } 5009 5010 // If this node hasn't been modified yet, it's still in the CSE maps, 5011 // so remove its old self from the CSE maps. 5012 if (!UserRemovedFromCSEMaps) { 5013 RemoveNodeFromCSEMaps(User); 5014 UserRemovedFromCSEMaps = true; 5015 } 5016 5017 ++UI; 5018 Use.set(To); 5019 } while (UI != UE && *UI == User); 5020 5021 // We are iterating over all uses of the From node, so if a use 5022 // doesn't use the specific value, no changes are made. 5023 if (!UserRemovedFromCSEMaps) 5024 continue; 5025 5026 // Now that we have modified User, add it back to the CSE maps. If it 5027 // already exists there, recursively merge the results together. 5028 AddModifiedNodeToCSEMaps(User, UpdateListener); 5029 } 5030} 5031 5032namespace { 5033 /// UseMemo - This class is used by SelectionDAG::ReplaceAllUsesOfValuesWith 5034 /// to record information about a use. 5035 struct UseMemo { 5036 SDNode *User; 5037 unsigned Index; 5038 SDUse *Use; 5039 }; 5040 5041 /// operator< - Sort Memos by User. 5042 bool operator<(const UseMemo &L, const UseMemo &R) { 5043 return (intptr_t)L.User < (intptr_t)R.User; 5044 } 5045} 5046 5047/// ReplaceAllUsesOfValuesWith - Replace any uses of From with To, leaving 5048/// uses of other values produced by From.getNode() alone. The same value 5049/// may appear in both the From and To list. The Deleted vector is 5050/// handled the same way as for ReplaceAllUsesWith. 5051void SelectionDAG::ReplaceAllUsesOfValuesWith(const SDValue *From, 5052 const SDValue *To, 5053 unsigned Num, 5054 DAGUpdateListener *UpdateListener){ 5055 // Handle the simple, trivial case efficiently. 5056 if (Num == 1) 5057 return ReplaceAllUsesOfValueWith(*From, *To, UpdateListener); 5058 5059 // Read up all the uses and make records of them. This helps 5060 // processing new uses that are introduced during the 5061 // replacement process. 5062 SmallVector<UseMemo, 4> Uses; 5063 for (unsigned i = 0; i != Num; ++i) { 5064 unsigned FromResNo = From[i].getResNo(); 5065 SDNode *FromNode = From[i].getNode(); 5066 for (SDNode::use_iterator UI = FromNode->use_begin(), 5067 E = FromNode->use_end(); UI != E; ++UI) { 5068 SDUse &Use = UI.getUse(); 5069 if (Use.getResNo() == FromResNo) { 5070 UseMemo Memo = { *UI, i, &Use }; 5071 Uses.push_back(Memo); 5072 } 5073 } 5074 } 5075 5076 // Sort the uses, so that all the uses from a given User are together. 5077 std::sort(Uses.begin(), Uses.end()); 5078 5079 for (unsigned UseIndex = 0, UseIndexEnd = Uses.size(); 5080 UseIndex != UseIndexEnd; ) { 5081 // We know that this user uses some value of From. If it is the right 5082 // value, update it. 5083 SDNode *User = Uses[UseIndex].User; 5084 5085 // This node is about to morph, remove its old self from the CSE maps. 5086 RemoveNodeFromCSEMaps(User); 5087 5088 // The Uses array is sorted, so all the uses for a given User 5089 // are next to each other in the list. 5090 // To help reduce the number of CSE recomputations, process all 5091 // the uses of this user that we can find this way. 5092 do { 5093 unsigned i = Uses[UseIndex].Index; 5094 SDUse &Use = *Uses[UseIndex].Use; 5095 ++UseIndex; 5096 5097 Use.set(To[i]); 5098 } while (UseIndex != UseIndexEnd && Uses[UseIndex].User == User); 5099 5100 // Now that we have modified User, add it back to the CSE maps. If it 5101 // already exists there, recursively merge the results together. 5102 AddModifiedNodeToCSEMaps(User, UpdateListener); 5103 } 5104} 5105 5106/// AssignTopologicalOrder - Assign a unique node id for each node in the DAG 5107/// based on their topological order. It returns the maximum id and a vector 5108/// of the SDNodes* in assigned order by reference. 5109unsigned SelectionDAG::AssignTopologicalOrder() { 5110 5111 unsigned DAGSize = 0; 5112 5113 // SortedPos tracks the progress of the algorithm. Nodes before it are 5114 // sorted, nodes after it are unsorted. When the algorithm completes 5115 // it is at the end of the list. 5116 allnodes_iterator SortedPos = allnodes_begin(); 5117 5118 // Visit all the nodes. Move nodes with no operands to the front of 5119 // the list immediately. Annotate nodes that do have operands with their 5120 // operand count. Before we do this, the Node Id fields of the nodes 5121 // may contain arbitrary values. After, the Node Id fields for nodes 5122 // before SortedPos will contain the topological sort index, and the 5123 // Node Id fields for nodes At SortedPos and after will contain the 5124 // count of outstanding operands. 5125 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ) { 5126 SDNode *N = I++; 5127 unsigned Degree = N->getNumOperands(); 5128 if (Degree == 0) { 5129 // A node with no uses, add it to the result array immediately. 5130 N->setNodeId(DAGSize++); 5131 allnodes_iterator Q = N; 5132 if (Q != SortedPos) 5133 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(Q)); 5134 ++SortedPos; 5135 } else { 5136 // Temporarily use the Node Id as scratch space for the degree count. 5137 N->setNodeId(Degree); 5138 } 5139 } 5140 5141 // Visit all the nodes. As we iterate, moves nodes into sorted order, 5142 // such that by the time the end is reached all nodes will be sorted. 5143 for (allnodes_iterator I = allnodes_begin(),E = allnodes_end(); I != E; ++I) { 5144 SDNode *N = I; 5145 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end(); 5146 UI != UE; ++UI) { 5147 SDNode *P = *UI; 5148 unsigned Degree = P->getNodeId(); 5149 --Degree; 5150 if (Degree == 0) { 5151 // All of P's operands are sorted, so P may sorted now. 5152 P->setNodeId(DAGSize++); 5153 if (P != SortedPos) 5154 SortedPos = AllNodes.insert(SortedPos, AllNodes.remove(P)); 5155 ++SortedPos; 5156 } else { 5157 // Update P's outstanding operand count. 5158 P->setNodeId(Degree); 5159 } 5160 } 5161 } 5162 5163 assert(SortedPos == AllNodes.end() && 5164 "Topological sort incomplete!"); 5165 assert(AllNodes.front().getOpcode() == ISD::EntryToken && 5166 "First node in topological sort is not the entry token!"); 5167 assert(AllNodes.front().getNodeId() == 0 && 5168 "First node in topological sort has non-zero id!"); 5169 assert(AllNodes.front().getNumOperands() == 0 && 5170 "First node in topological sort has operands!"); 5171 assert(AllNodes.back().getNodeId() == (int)DAGSize-1 && 5172 "Last node in topologic sort has unexpected id!"); 5173 assert(AllNodes.back().use_empty() && 5174 "Last node in topologic sort has users!"); 5175 assert(DAGSize == allnodes_size() && "Node count mismatch!"); 5176 return DAGSize; 5177} 5178 5179 5180 5181//===----------------------------------------------------------------------===// 5182// SDNode Class 5183//===----------------------------------------------------------------------===// 5184 5185HandleSDNode::~HandleSDNode() { 5186 DropOperands(); 5187} 5188 5189GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, const GlobalValue *GA, 5190 EVT VT, int64_t o, unsigned char TF) 5191 : SDNode(Opc, DebugLoc::getUnknownLoc(), getSDVTList(VT)), 5192 Offset(o), TargetFlags(TF) { 5193 TheGlobal = const_cast<GlobalValue*>(GA); 5194} 5195 5196MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, EVT memvt, 5197 MachineMemOperand *mmo) 5198 : SDNode(Opc, dl, VTs), MemoryVT(memvt), MMO(mmo) { 5199 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile()); 5200 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!"); 5201 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!"); 5202} 5203 5204MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, 5205 const SDValue *Ops, unsigned NumOps, EVT memvt, 5206 MachineMemOperand *mmo) 5207 : SDNode(Opc, dl, VTs, Ops, NumOps), 5208 MemoryVT(memvt), MMO(mmo) { 5209 SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile()); 5210 assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!"); 5211 assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!"); 5212} 5213 5214/// Profile - Gather unique data for the node. 5215/// 5216void SDNode::Profile(FoldingSetNodeID &ID) const { 5217 AddNodeIDNode(ID, this); 5218} 5219 5220namespace { 5221 struct EVTArray { 5222 std::vector<EVT> VTs; 5223 5224 EVTArray() { 5225 VTs.reserve(MVT::LAST_VALUETYPE); 5226 for (unsigned i = 0; i < MVT::LAST_VALUETYPE; ++i) 5227 VTs.push_back(MVT((MVT::SimpleValueType)i)); 5228 } 5229 }; 5230} 5231 5232static ManagedStatic<std::set<EVT, EVT::compareRawBits> > EVTs; 5233static ManagedStatic<EVTArray> SimpleVTArray; 5234static ManagedStatic<sys::SmartMutex<true> > VTMutex; 5235 5236/// getValueTypeList - Return a pointer to the specified value type. 5237/// 5238const EVT *SDNode::getValueTypeList(EVT VT) { 5239 if (VT.isExtended()) { 5240 sys::SmartScopedLock<true> Lock(*VTMutex); 5241 return &(*EVTs->insert(VT).first); 5242 } else { 5243 return &SimpleVTArray->VTs[VT.getSimpleVT().SimpleTy]; 5244 } 5245} 5246 5247/// hasNUsesOfValue - Return true if there are exactly NUSES uses of the 5248/// indicated value. This method ignores uses of other values defined by this 5249/// operation. 5250bool SDNode::hasNUsesOfValue(unsigned NUses, unsigned Value) const { 5251 assert(Value < getNumValues() && "Bad value!"); 5252 5253 // TODO: Only iterate over uses of a given value of the node 5254 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) { 5255 if (UI.getUse().getResNo() == Value) { 5256 if (NUses == 0) 5257 return false; 5258 --NUses; 5259 } 5260 } 5261 5262 // Found exactly the right number of uses? 5263 return NUses == 0; 5264} 5265 5266 5267/// hasAnyUseOfValue - Return true if there are any use of the indicated 5268/// value. This method ignores uses of other values defined by this operation. 5269bool SDNode::hasAnyUseOfValue(unsigned Value) const { 5270 assert(Value < getNumValues() && "Bad value!"); 5271 5272 for (SDNode::use_iterator UI = use_begin(), E = use_end(); UI != E; ++UI) 5273 if (UI.getUse().getResNo() == Value) 5274 return true; 5275 5276 return false; 5277} 5278 5279 5280/// isOnlyUserOf - Return true if this node is the only use of N. 5281/// 5282bool SDNode::isOnlyUserOf(SDNode *N) const { 5283 bool Seen = false; 5284 for (SDNode::use_iterator I = N->use_begin(), E = N->use_end(); I != E; ++I) { 5285 SDNode *User = *I; 5286 if (User == this) 5287 Seen = true; 5288 else 5289 return false; 5290 } 5291 5292 return Seen; 5293} 5294 5295/// isOperand - Return true if this node is an operand of N. 5296/// 5297bool SDValue::isOperandOf(SDNode *N) const { 5298 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 5299 if (*this == N->getOperand(i)) 5300 return true; 5301 return false; 5302} 5303 5304bool SDNode::isOperandOf(SDNode *N) const { 5305 for (unsigned i = 0, e = N->NumOperands; i != e; ++i) 5306 if (this == N->OperandList[i].getNode()) 5307 return true; 5308 return false; 5309} 5310 5311/// reachesChainWithoutSideEffects - Return true if this operand (which must 5312/// be a chain) reaches the specified operand without crossing any 5313/// side-effecting instructions. In practice, this looks through token 5314/// factors and non-volatile loads. In order to remain efficient, this only 5315/// looks a couple of nodes in, it does not do an exhaustive search. 5316bool SDValue::reachesChainWithoutSideEffects(SDValue Dest, 5317 unsigned Depth) const { 5318 if (*this == Dest) return true; 5319 5320 // Don't search too deeply, we just want to be able to see through 5321 // TokenFactor's etc. 5322 if (Depth == 0) return false; 5323 5324 // If this is a token factor, all inputs to the TF happen in parallel. If any 5325 // of the operands of the TF reach dest, then we can do the xform. 5326 if (getOpcode() == ISD::TokenFactor) { 5327 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) 5328 if (getOperand(i).reachesChainWithoutSideEffects(Dest, Depth-1)) 5329 return true; 5330 return false; 5331 } 5332 5333 // Loads don't have side effects, look through them. 5334 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(*this)) { 5335 if (!Ld->isVolatile()) 5336 return Ld->getChain().reachesChainWithoutSideEffects(Dest, Depth-1); 5337 } 5338 return false; 5339} 5340 5341/// isPredecessorOf - Return true if this node is a predecessor of N. This node 5342/// is either an operand of N or it can be reached by traversing up the operands. 5343/// NOTE: this is an expensive method. Use it carefully. 5344bool SDNode::isPredecessorOf(SDNode *N) const { 5345 SmallPtrSet<SDNode *, 32> Visited; 5346 SmallVector<SDNode *, 16> Worklist; 5347 Worklist.push_back(N); 5348 5349 do { 5350 N = Worklist.pop_back_val(); 5351 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 5352 SDNode *Op = N->getOperand(i).getNode(); 5353 if (Op == this) 5354 return true; 5355 if (Visited.insert(Op)) 5356 Worklist.push_back(Op); 5357 } 5358 } while (!Worklist.empty()); 5359 5360 return false; 5361} 5362 5363uint64_t SDNode::getConstantOperandVal(unsigned Num) const { 5364 assert(Num < NumOperands && "Invalid child # of SDNode!"); 5365 return cast<ConstantSDNode>(OperandList[Num])->getZExtValue(); 5366} 5367 5368std::string SDNode::getOperationName(const SelectionDAG *G) const { 5369 switch (getOpcode()) { 5370 default: 5371 if (getOpcode() < ISD::BUILTIN_OP_END) 5372 return "<<Unknown DAG Node>>"; 5373 if (isMachineOpcode()) { 5374 if (G) 5375 if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo()) 5376 if (getMachineOpcode() < TII->getNumOpcodes()) 5377 return TII->get(getMachineOpcode()).getName(); 5378 return "<<Unknown Machine Node>>"; 5379 } 5380 if (G) { 5381 const TargetLowering &TLI = G->getTargetLoweringInfo(); 5382 const char *Name = TLI.getTargetNodeName(getOpcode()); 5383 if (Name) return Name; 5384 return "<<Unknown Target Node>>"; 5385 } 5386 return "<<Unknown Node>>"; 5387 5388#ifndef NDEBUG 5389 case ISD::DELETED_NODE: 5390 return "<<Deleted Node!>>"; 5391#endif 5392 case ISD::PREFETCH: return "Prefetch"; 5393 case ISD::MEMBARRIER: return "MemBarrier"; 5394 case ISD::ATOMIC_CMP_SWAP: return "AtomicCmpSwap"; 5395 case ISD::ATOMIC_SWAP: return "AtomicSwap"; 5396 case ISD::ATOMIC_LOAD_ADD: return "AtomicLoadAdd"; 5397 case ISD::ATOMIC_LOAD_SUB: return "AtomicLoadSub"; 5398 case ISD::ATOMIC_LOAD_AND: return "AtomicLoadAnd"; 5399 case ISD::ATOMIC_LOAD_OR: return "AtomicLoadOr"; 5400 case ISD::ATOMIC_LOAD_XOR: return "AtomicLoadXor"; 5401 case ISD::ATOMIC_LOAD_NAND: return "AtomicLoadNand"; 5402 case ISD::ATOMIC_LOAD_MIN: return "AtomicLoadMin"; 5403 case ISD::ATOMIC_LOAD_MAX: return "AtomicLoadMax"; 5404 case ISD::ATOMIC_LOAD_UMIN: return "AtomicLoadUMin"; 5405 case ISD::ATOMIC_LOAD_UMAX: return "AtomicLoadUMax"; 5406 case ISD::PCMARKER: return "PCMarker"; 5407 case ISD::READCYCLECOUNTER: return "ReadCycleCounter"; 5408 case ISD::SRCVALUE: return "SrcValue"; 5409 case ISD::EntryToken: return "EntryToken"; 5410 case ISD::TokenFactor: return "TokenFactor"; 5411 case ISD::AssertSext: return "AssertSext"; 5412 case ISD::AssertZext: return "AssertZext"; 5413 5414 case ISD::BasicBlock: return "BasicBlock"; 5415 case ISD::VALUETYPE: return "ValueType"; 5416 case ISD::Register: return "Register"; 5417 5418 case ISD::Constant: return "Constant"; 5419 case ISD::ConstantFP: return "ConstantFP"; 5420 case ISD::GlobalAddress: return "GlobalAddress"; 5421 case ISD::GlobalTLSAddress: return "GlobalTLSAddress"; 5422 case ISD::FrameIndex: return "FrameIndex"; 5423 case ISD::JumpTable: return "JumpTable"; 5424 case ISD::GLOBAL_OFFSET_TABLE: return "GLOBAL_OFFSET_TABLE"; 5425 case ISD::RETURNADDR: return "RETURNADDR"; 5426 case ISD::FRAMEADDR: return "FRAMEADDR"; 5427 case ISD::FRAME_TO_ARGS_OFFSET: return "FRAME_TO_ARGS_OFFSET"; 5428 case ISD::EXCEPTIONADDR: return "EXCEPTIONADDR"; 5429 case ISD::LSDAADDR: return "LSDAADDR"; 5430 case ISD::EHSELECTION: return "EHSELECTION"; 5431 case ISD::EH_RETURN: return "EH_RETURN"; 5432 case ISD::ConstantPool: return "ConstantPool"; 5433 case ISD::ExternalSymbol: return "ExternalSymbol"; 5434 case ISD::BlockAddress: return "BlockAddress"; 5435 case ISD::INTRINSIC_WO_CHAIN: 5436 case ISD::INTRINSIC_VOID: 5437 case ISD::INTRINSIC_W_CHAIN: { 5438 unsigned OpNo = getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 0 : 1; 5439 unsigned IID = cast<ConstantSDNode>(getOperand(OpNo))->getZExtValue(); 5440 if (IID < Intrinsic::num_intrinsics) 5441 return Intrinsic::getName((Intrinsic::ID)IID); 5442 else if (const TargetIntrinsicInfo *TII = G->getTarget().getIntrinsicInfo()) 5443 return TII->getName(IID); 5444 llvm_unreachable("Invalid intrinsic ID"); 5445 } 5446 5447 case ISD::BUILD_VECTOR: return "BUILD_VECTOR"; 5448 case ISD::TargetConstant: return "TargetConstant"; 5449 case ISD::TargetConstantFP:return "TargetConstantFP"; 5450 case ISD::TargetGlobalAddress: return "TargetGlobalAddress"; 5451 case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress"; 5452 case ISD::TargetFrameIndex: return "TargetFrameIndex"; 5453 case ISD::TargetJumpTable: return "TargetJumpTable"; 5454 case ISD::TargetConstantPool: return "TargetConstantPool"; 5455 case ISD::TargetExternalSymbol: return "TargetExternalSymbol"; 5456 case ISD::TargetBlockAddress: return "TargetBlockAddress"; 5457 5458 case ISD::CopyToReg: return "CopyToReg"; 5459 case ISD::CopyFromReg: return "CopyFromReg"; 5460 case ISD::UNDEF: return "undef"; 5461 case ISD::MERGE_VALUES: return "merge_values"; 5462 case ISD::INLINEASM: return "inlineasm"; 5463 case ISD::EH_LABEL: return "eh_label"; 5464 case ISD::HANDLENODE: return "handlenode"; 5465 5466 // Unary operators 5467 case ISD::FABS: return "fabs"; 5468 case ISD::FNEG: return "fneg"; 5469 case ISD::FSQRT: return "fsqrt"; 5470 case ISD::FSIN: return "fsin"; 5471 case ISD::FCOS: return "fcos"; 5472 case ISD::FPOWI: return "fpowi"; 5473 case ISD::FPOW: return "fpow"; 5474 case ISD::FTRUNC: return "ftrunc"; 5475 case ISD::FFLOOR: return "ffloor"; 5476 case ISD::FCEIL: return "fceil"; 5477 case ISD::FRINT: return "frint"; 5478 case ISD::FNEARBYINT: return "fnearbyint"; 5479 5480 // Binary operators 5481 case ISD::ADD: return "add"; 5482 case ISD::SUB: return "sub"; 5483 case ISD::MUL: return "mul"; 5484 case ISD::MULHU: return "mulhu"; 5485 case ISD::MULHS: return "mulhs"; 5486 case ISD::SDIV: return "sdiv"; 5487 case ISD::UDIV: return "udiv"; 5488 case ISD::SREM: return "srem"; 5489 case ISD::UREM: return "urem"; 5490 case ISD::SMUL_LOHI: return "smul_lohi"; 5491 case ISD::UMUL_LOHI: return "umul_lohi"; 5492 case ISD::SDIVREM: return "sdivrem"; 5493 case ISD::UDIVREM: return "udivrem"; 5494 case ISD::AND: return "and"; 5495 case ISD::OR: return "or"; 5496 case ISD::XOR: return "xor"; 5497 case ISD::SHL: return "shl"; 5498 case ISD::SRA: return "sra"; 5499 case ISD::SRL: return "srl"; 5500 case ISD::ROTL: return "rotl"; 5501 case ISD::ROTR: return "rotr"; 5502 case ISD::FADD: return "fadd"; 5503 case ISD::FSUB: return "fsub"; 5504 case ISD::FMUL: return "fmul"; 5505 case ISD::FDIV: return "fdiv"; 5506 case ISD::FREM: return "frem"; 5507 case ISD::FCOPYSIGN: return "fcopysign"; 5508 case ISD::FGETSIGN: return "fgetsign"; 5509 5510 case ISD::SETCC: return "setcc"; 5511 case ISD::VSETCC: return "vsetcc"; 5512 case ISD::SELECT: return "select"; 5513 case ISD::SELECT_CC: return "select_cc"; 5514 case ISD::INSERT_VECTOR_ELT: return "insert_vector_elt"; 5515 case ISD::EXTRACT_VECTOR_ELT: return "extract_vector_elt"; 5516 case ISD::CONCAT_VECTORS: return "concat_vectors"; 5517 case ISD::EXTRACT_SUBVECTOR: return "extract_subvector"; 5518 case ISD::SCALAR_TO_VECTOR: return "scalar_to_vector"; 5519 case ISD::VECTOR_SHUFFLE: return "vector_shuffle"; 5520 case ISD::CARRY_FALSE: return "carry_false"; 5521 case ISD::ADDC: return "addc"; 5522 case ISD::ADDE: return "adde"; 5523 case ISD::SADDO: return "saddo"; 5524 case ISD::UADDO: return "uaddo"; 5525 case ISD::SSUBO: return "ssubo"; 5526 case ISD::USUBO: return "usubo"; 5527 case ISD::SMULO: return "smulo"; 5528 case ISD::UMULO: return "umulo"; 5529 case ISD::SUBC: return "subc"; 5530 case ISD::SUBE: return "sube"; 5531 case ISD::SHL_PARTS: return "shl_parts"; 5532 case ISD::SRA_PARTS: return "sra_parts"; 5533 case ISD::SRL_PARTS: return "srl_parts"; 5534 5535 // Conversion operators. 5536 case ISD::SIGN_EXTEND: return "sign_extend"; 5537 case ISD::ZERO_EXTEND: return "zero_extend"; 5538 case ISD::ANY_EXTEND: return "any_extend"; 5539 case ISD::SIGN_EXTEND_INREG: return "sign_extend_inreg"; 5540 case ISD::TRUNCATE: return "truncate"; 5541 case ISD::FP_ROUND: return "fp_round"; 5542 case ISD::FLT_ROUNDS_: return "flt_rounds"; 5543 case ISD::FP_ROUND_INREG: return "fp_round_inreg"; 5544 case ISD::FP_EXTEND: return "fp_extend"; 5545 5546 case ISD::SINT_TO_FP: return "sint_to_fp"; 5547 case ISD::UINT_TO_FP: return "uint_to_fp"; 5548 case ISD::FP_TO_SINT: return "fp_to_sint"; 5549 case ISD::FP_TO_UINT: return "fp_to_uint"; 5550 case ISD::BIT_CONVERT: return "bit_convert"; 5551 5552 case ISD::CONVERT_RNDSAT: { 5553 switch (cast<CvtRndSatSDNode>(this)->getCvtCode()) { 5554 default: llvm_unreachable("Unknown cvt code!"); 5555 case ISD::CVT_FF: return "cvt_ff"; 5556 case ISD::CVT_FS: return "cvt_fs"; 5557 case ISD::CVT_FU: return "cvt_fu"; 5558 case ISD::CVT_SF: return "cvt_sf"; 5559 case ISD::CVT_UF: return "cvt_uf"; 5560 case ISD::CVT_SS: return "cvt_ss"; 5561 case ISD::CVT_SU: return "cvt_su"; 5562 case ISD::CVT_US: return "cvt_us"; 5563 case ISD::CVT_UU: return "cvt_uu"; 5564 } 5565 } 5566 5567 // Control flow instructions 5568 case ISD::BR: return "br"; 5569 case ISD::BRIND: return "brind"; 5570 case ISD::BR_JT: return "br_jt"; 5571 case ISD::BRCOND: return "brcond"; 5572 case ISD::BR_CC: return "br_cc"; 5573 case ISD::CALLSEQ_START: return "callseq_start"; 5574 case ISD::CALLSEQ_END: return "callseq_end"; 5575 5576 // Other operators 5577 case ISD::LOAD: return "load"; 5578 case ISD::STORE: return "store"; 5579 case ISD::VAARG: return "vaarg"; 5580 case ISD::VACOPY: return "vacopy"; 5581 case ISD::VAEND: return "vaend"; 5582 case ISD::VASTART: return "vastart"; 5583 case ISD::DYNAMIC_STACKALLOC: return "dynamic_stackalloc"; 5584 case ISD::EXTRACT_ELEMENT: return "extract_element"; 5585 case ISD::BUILD_PAIR: return "build_pair"; 5586 case ISD::STACKSAVE: return "stacksave"; 5587 case ISD::STACKRESTORE: return "stackrestore"; 5588 case ISD::TRAP: return "trap"; 5589 5590 // Bit manipulation 5591 case ISD::BSWAP: return "bswap"; 5592 case ISD::CTPOP: return "ctpop"; 5593 case ISD::CTTZ: return "cttz"; 5594 case ISD::CTLZ: return "ctlz"; 5595 5596 // Trampolines 5597 case ISD::TRAMPOLINE: return "trampoline"; 5598 5599 case ISD::CONDCODE: 5600 switch (cast<CondCodeSDNode>(this)->get()) { 5601 default: llvm_unreachable("Unknown setcc condition!"); 5602 case ISD::SETOEQ: return "setoeq"; 5603 case ISD::SETOGT: return "setogt"; 5604 case ISD::SETOGE: return "setoge"; 5605 case ISD::SETOLT: return "setolt"; 5606 case ISD::SETOLE: return "setole"; 5607 case ISD::SETONE: return "setone"; 5608 5609 case ISD::SETO: return "seto"; 5610 case ISD::SETUO: return "setuo"; 5611 case ISD::SETUEQ: return "setue"; 5612 case ISD::SETUGT: return "setugt"; 5613 case ISD::SETUGE: return "setuge"; 5614 case ISD::SETULT: return "setult"; 5615 case ISD::SETULE: return "setule"; 5616 case ISD::SETUNE: return "setune"; 5617 5618 case ISD::SETEQ: return "seteq"; 5619 case ISD::SETGT: return "setgt"; 5620 case ISD::SETGE: return "setge"; 5621 case ISD::SETLT: return "setlt"; 5622 case ISD::SETLE: return "setle"; 5623 case ISD::SETNE: return "setne"; 5624 } 5625 } 5626} 5627 5628const char *SDNode::getIndexedModeName(ISD::MemIndexedMode AM) { 5629 switch (AM) { 5630 default: 5631 return ""; 5632 case ISD::PRE_INC: 5633 return "<pre-inc>"; 5634 case ISD::PRE_DEC: 5635 return "<pre-dec>"; 5636 case ISD::POST_INC: 5637 return "<post-inc>"; 5638 case ISD::POST_DEC: 5639 return "<post-dec>"; 5640 } 5641} 5642 5643std::string ISD::ArgFlagsTy::getArgFlagsString() { 5644 std::string S = "< "; 5645 5646 if (isZExt()) 5647 S += "zext "; 5648 if (isSExt()) 5649 S += "sext "; 5650 if (isInReg()) 5651 S += "inreg "; 5652 if (isSRet()) 5653 S += "sret "; 5654 if (isByVal()) 5655 S += "byval "; 5656 if (isNest()) 5657 S += "nest "; 5658 if (getByValAlign()) 5659 S += "byval-align:" + utostr(getByValAlign()) + " "; 5660 if (getOrigAlign()) 5661 S += "orig-align:" + utostr(getOrigAlign()) + " "; 5662 if (getByValSize()) 5663 S += "byval-size:" + utostr(getByValSize()) + " "; 5664 return S + ">"; 5665} 5666 5667void SDNode::dump() const { dump(0); } 5668void SDNode::dump(const SelectionDAG *G) const { 5669 print(errs(), G); 5670} 5671 5672void SDNode::print_types(raw_ostream &OS, const SelectionDAG *G) const { 5673 OS << (void*)this << ": "; 5674 5675 for (unsigned i = 0, e = getNumValues(); i != e; ++i) { 5676 if (i) OS << ","; 5677 if (getValueType(i) == MVT::Other) 5678 OS << "ch"; 5679 else 5680 OS << getValueType(i).getEVTString(); 5681 } 5682 OS << " = " << getOperationName(G); 5683} 5684 5685void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const { 5686 if (const MachineSDNode *MN = dyn_cast<MachineSDNode>(this)) { 5687 if (!MN->memoperands_empty()) { 5688 OS << "<"; 5689 OS << "Mem:"; 5690 for (MachineSDNode::mmo_iterator i = MN->memoperands_begin(), 5691 e = MN->memoperands_end(); i != e; ++i) { 5692 OS << **i; 5693 if (next(i) != e) 5694 OS << " "; 5695 } 5696 OS << ">"; 5697 } 5698 } else if (const ShuffleVectorSDNode *SVN = 5699 dyn_cast<ShuffleVectorSDNode>(this)) { 5700 OS << "<"; 5701 for (unsigned i = 0, e = ValueList[0].getVectorNumElements(); i != e; ++i) { 5702 int Idx = SVN->getMaskElt(i); 5703 if (i) OS << ","; 5704 if (Idx < 0) 5705 OS << "u"; 5706 else 5707 OS << Idx; 5708 } 5709 OS << ">"; 5710 } else if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) { 5711 OS << '<' << CSDN->getAPIntValue() << '>'; 5712 } else if (const ConstantFPSDNode *CSDN = dyn_cast<ConstantFPSDNode>(this)) { 5713 if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle) 5714 OS << '<' << CSDN->getValueAPF().convertToFloat() << '>'; 5715 else if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEdouble) 5716 OS << '<' << CSDN->getValueAPF().convertToDouble() << '>'; 5717 else { 5718 OS << "<APFloat("; 5719 CSDN->getValueAPF().bitcastToAPInt().dump(); 5720 OS << ")>"; 5721 } 5722 } else if (const GlobalAddressSDNode *GADN = 5723 dyn_cast<GlobalAddressSDNode>(this)) { 5724 int64_t offset = GADN->getOffset(); 5725 OS << '<'; 5726 WriteAsOperand(OS, GADN->getGlobal()); 5727 OS << '>'; 5728 if (offset > 0) 5729 OS << " + " << offset; 5730 else 5731 OS << " " << offset; 5732 if (unsigned int TF = GADN->getTargetFlags()) 5733 OS << " [TF=" << TF << ']'; 5734 } else if (const FrameIndexSDNode *FIDN = dyn_cast<FrameIndexSDNode>(this)) { 5735 OS << "<" << FIDN->getIndex() << ">"; 5736 } else if (const JumpTableSDNode *JTDN = dyn_cast<JumpTableSDNode>(this)) { 5737 OS << "<" << JTDN->getIndex() << ">"; 5738 if (unsigned int TF = JTDN->getTargetFlags()) 5739 OS << " [TF=" << TF << ']'; 5740 } else if (const ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(this)){ 5741 int offset = CP->getOffset(); 5742 if (CP->isMachineConstantPoolEntry()) 5743 OS << "<" << *CP->getMachineCPVal() << ">"; 5744 else 5745 OS << "<" << *CP->getConstVal() << ">"; 5746 if (offset > 0) 5747 OS << " + " << offset; 5748 else 5749 OS << " " << offset; 5750 if (unsigned int TF = CP->getTargetFlags()) 5751 OS << " [TF=" << TF << ']'; 5752 } else if (const BasicBlockSDNode *BBDN = dyn_cast<BasicBlockSDNode>(this)) { 5753 OS << "<"; 5754 const Value *LBB = (const Value*)BBDN->getBasicBlock()->getBasicBlock(); 5755 if (LBB) 5756 OS << LBB->getName() << " "; 5757 OS << (const void*)BBDN->getBasicBlock() << ">"; 5758 } else if (const RegisterSDNode *R = dyn_cast<RegisterSDNode>(this)) { 5759 if (G && R->getReg() && 5760 TargetRegisterInfo::isPhysicalRegister(R->getReg())) { 5761 OS << " %" << G->getTarget().getRegisterInfo()->getName(R->getReg()); 5762 } else { 5763 OS << " %reg" << R->getReg(); 5764 } 5765 } else if (const ExternalSymbolSDNode *ES = 5766 dyn_cast<ExternalSymbolSDNode>(this)) { 5767 OS << "'" << ES->getSymbol() << "'"; 5768 if (unsigned int TF = ES->getTargetFlags()) 5769 OS << " [TF=" << TF << ']'; 5770 } else if (const SrcValueSDNode *M = dyn_cast<SrcValueSDNode>(this)) { 5771 if (M->getValue()) 5772 OS << "<" << M->getValue() << ">"; 5773 else 5774 OS << "<null>"; 5775 } else if (const VTSDNode *N = dyn_cast<VTSDNode>(this)) { 5776 OS << ":" << N->getVT().getEVTString(); 5777 } 5778 else if (const LoadSDNode *LD = dyn_cast<LoadSDNode>(this)) { 5779 OS << "<" << *LD->getMemOperand(); 5780 5781 bool doExt = true; 5782 switch (LD->getExtensionType()) { 5783 default: doExt = false; break; 5784 case ISD::EXTLOAD: OS << ", anyext"; break; 5785 case ISD::SEXTLOAD: OS << ", sext"; break; 5786 case ISD::ZEXTLOAD: OS << ", zext"; break; 5787 } 5788 if (doExt) 5789 OS << " from " << LD->getMemoryVT().getEVTString(); 5790 5791 const char *AM = getIndexedModeName(LD->getAddressingMode()); 5792 if (*AM) 5793 OS << ", " << AM; 5794 5795 OS << ">"; 5796 } else if (const StoreSDNode *ST = dyn_cast<StoreSDNode>(this)) { 5797 OS << "<" << *ST->getMemOperand(); 5798 5799 if (ST->isTruncatingStore()) 5800 OS << ", trunc to " << ST->getMemoryVT().getEVTString(); 5801 5802 const char *AM = getIndexedModeName(ST->getAddressingMode()); 5803 if (*AM) 5804 OS << ", " << AM; 5805 5806 OS << ">"; 5807 } else if (const MemSDNode* M = dyn_cast<MemSDNode>(this)) { 5808 OS << "<" << *M->getMemOperand() << ">"; 5809 } else if (const BlockAddressSDNode *BA = 5810 dyn_cast<BlockAddressSDNode>(this)) { 5811 OS << "<"; 5812 WriteAsOperand(OS, BA->getBlockAddress()->getFunction(), false); 5813 OS << ", "; 5814 WriteAsOperand(OS, BA->getBlockAddress()->getBasicBlock(), false); 5815 OS << ">"; 5816 if (unsigned int TF = BA->getTargetFlags()) 5817 OS << " [TF=" << TF << ']'; 5818 } 5819} 5820 5821void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const { 5822 print_types(OS, G); 5823 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) { 5824 if (i) OS << ", "; else OS << " "; 5825 OS << (void*)getOperand(i).getNode(); 5826 if (unsigned RN = getOperand(i).getResNo()) 5827 OS << ":" << RN; 5828 } 5829 print_details(OS, G); 5830} 5831 5832static void DumpNodes(const SDNode *N, unsigned indent, const SelectionDAG *G) { 5833 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) 5834 if (N->getOperand(i).getNode()->hasOneUse()) 5835 DumpNodes(N->getOperand(i).getNode(), indent+2, G); 5836 else 5837 errs() << "\n" << std::string(indent+2, ' ') 5838 << (void*)N->getOperand(i).getNode() << ": <multiple use>"; 5839 5840 5841 errs() << "\n"; 5842 errs().indent(indent); 5843 N->dump(G); 5844} 5845 5846SDValue SelectionDAG::UnrollVectorOp(SDNode *N, unsigned ResNE) { 5847 assert(N->getNumValues() == 1 && 5848 "Can't unroll a vector with multiple results!"); 5849 5850 EVT VT = N->getValueType(0); 5851 unsigned NE = VT.getVectorNumElements(); 5852 EVT EltVT = VT.getVectorElementType(); 5853 DebugLoc dl = N->getDebugLoc(); 5854 5855 SmallVector<SDValue, 8> Scalars; 5856 SmallVector<SDValue, 4> Operands(N->getNumOperands()); 5857 5858 // If ResNE is 0, fully unroll the vector op. 5859 if (ResNE == 0) 5860 ResNE = NE; 5861 else if (NE > ResNE) 5862 NE = ResNE; 5863 5864 unsigned i; 5865 for (i= 0; i != NE; ++i) { 5866 for (unsigned j = 0; j != N->getNumOperands(); ++j) { 5867 SDValue Operand = N->getOperand(j); 5868 EVT OperandVT = Operand.getValueType(); 5869 if (OperandVT.isVector()) { 5870 // A vector operand; extract a single element. 5871 EVT OperandEltVT = OperandVT.getVectorElementType(); 5872 Operands[j] = getNode(ISD::EXTRACT_VECTOR_ELT, dl, 5873 OperandEltVT, 5874 Operand, 5875 getConstant(i, MVT::i32)); 5876 } else { 5877 // A scalar operand; just use it as is. 5878 Operands[j] = Operand; 5879 } 5880 } 5881 5882 switch (N->getOpcode()) { 5883 default: 5884 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, 5885 &Operands[0], Operands.size())); 5886 break; 5887 case ISD::SHL: 5888 case ISD::SRA: 5889 case ISD::SRL: 5890 case ISD::ROTL: 5891 case ISD::ROTR: 5892 Scalars.push_back(getNode(N->getOpcode(), dl, EltVT, Operands[0], 5893 getShiftAmountOperand(Operands[1]))); 5894 break; 5895 } 5896 } 5897 5898 for (; i < ResNE; ++i) 5899 Scalars.push_back(getUNDEF(EltVT)); 5900 5901 return getNode(ISD::BUILD_VECTOR, dl, 5902 EVT::getVectorVT(*getContext(), EltVT, ResNE), 5903 &Scalars[0], Scalars.size()); 5904} 5905 5906 5907/// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a 5908/// location that is 'Dist' units away from the location that the 'Base' load 5909/// is loading from. 5910bool SelectionDAG::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base, 5911 unsigned Bytes, int Dist) const { 5912 if (LD->getChain() != Base->getChain()) 5913 return false; 5914 EVT VT = LD->getValueType(0); 5915 if (VT.getSizeInBits() / 8 != Bytes) 5916 return false; 5917 5918 SDValue Loc = LD->getOperand(1); 5919 SDValue BaseLoc = Base->getOperand(1); 5920 if (Loc.getOpcode() == ISD::FrameIndex) { 5921 if (BaseLoc.getOpcode() != ISD::FrameIndex) 5922 return false; 5923 const MachineFrameInfo *MFI = getMachineFunction().getFrameInfo(); 5924 int FI = cast<FrameIndexSDNode>(Loc)->getIndex(); 5925 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex(); 5926 int FS = MFI->getObjectSize(FI); 5927 int BFS = MFI->getObjectSize(BFI); 5928 if (FS != BFS || FS != (int)Bytes) return false; 5929 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes); 5930 } 5931 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) { 5932 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1)); 5933 if (V && (V->getSExtValue() == Dist*Bytes)) 5934 return true; 5935 } 5936 5937 GlobalValue *GV1 = NULL; 5938 GlobalValue *GV2 = NULL; 5939 int64_t Offset1 = 0; 5940 int64_t Offset2 = 0; 5941 bool isGA1 = TLI.isGAPlusOffset(Loc.getNode(), GV1, Offset1); 5942 bool isGA2 = TLI.isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2); 5943 if (isGA1 && isGA2 && GV1 == GV2) 5944 return Offset1 == (Offset2 + Dist*Bytes); 5945 return false; 5946} 5947 5948 5949/// InferPtrAlignment - Infer alignment of a load / store address. Return 0 if 5950/// it cannot be inferred. 5951unsigned SelectionDAG::InferPtrAlignment(SDValue Ptr) const { 5952 // If this is a GlobalAddress + cst, return the alignment. 5953 GlobalValue *GV; 5954 int64_t GVOffset = 0; 5955 if (TLI.isGAPlusOffset(Ptr.getNode(), GV, GVOffset)) 5956 return MinAlign(GV->getAlignment(), GVOffset); 5957 5958 // If this is a direct reference to a stack slot, use information about the 5959 // stack slot's alignment. 5960 int FrameIdx = 1 << 31; 5961 int64_t FrameOffset = 0; 5962 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Ptr)) { 5963 FrameIdx = FI->getIndex(); 5964 } else if (Ptr.getOpcode() == ISD::ADD && 5965 isa<ConstantSDNode>(Ptr.getOperand(1)) && 5966 isa<FrameIndexSDNode>(Ptr.getOperand(0))) { 5967 FrameIdx = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex(); 5968 FrameOffset = Ptr.getConstantOperandVal(1); 5969 } 5970 5971 if (FrameIdx != (1 << 31)) { 5972 // FIXME: Handle FI+CST. 5973 const MachineFrameInfo &MFI = *getMachineFunction().getFrameInfo(); 5974 unsigned FIInfoAlign = MinAlign(MFI.getObjectAlignment(FrameIdx), 5975 FrameOffset); 5976 if (MFI.isFixedObjectIndex(FrameIdx)) { 5977 int64_t ObjectOffset = MFI.getObjectOffset(FrameIdx) + FrameOffset; 5978 5979 // The alignment of the frame index can be determined from its offset from 5980 // the incoming frame position. If the frame object is at offset 32 and 5981 // the stack is guaranteed to be 16-byte aligned, then we know that the 5982 // object is 16-byte aligned. 5983 unsigned StackAlign = getTarget().getFrameInfo()->getStackAlignment(); 5984 unsigned Align = MinAlign(ObjectOffset, StackAlign); 5985 5986 // Finally, the frame object itself may have a known alignment. Factor 5987 // the alignment + offset into a new alignment. For example, if we know 5988 // the FI is 8 byte aligned, but the pointer is 4 off, we really have a 5989 // 4-byte alignment of the resultant pointer. Likewise align 4 + 4-byte 5990 // offset = 4-byte alignment, align 4 + 1-byte offset = align 1, etc. 5991 return std::max(Align, FIInfoAlign); 5992 } 5993 return FIInfoAlign; 5994 } 5995 5996 return 0; 5997} 5998 5999void SelectionDAG::dump() const { 6000 errs() << "SelectionDAG has " << AllNodes.size() << " nodes:"; 6001 6002 for (allnodes_const_iterator I = allnodes_begin(), E = allnodes_end(); 6003 I != E; ++I) { 6004 const SDNode *N = I; 6005 if (!N->hasOneUse() && N != getRoot().getNode()) 6006 DumpNodes(N, 2, this); 6007 } 6008 6009 if (getRoot().getNode()) DumpNodes(getRoot().getNode(), 2, this); 6010 6011 errs() << "\n\n"; 6012} 6013 6014void SDNode::printr(raw_ostream &OS, const SelectionDAG *G) const { 6015 print_types(OS, G); 6016 print_details(OS, G); 6017} 6018 6019typedef SmallPtrSet<const SDNode *, 128> VisitedSDNodeSet; 6020static void DumpNodesr(raw_ostream &OS, const SDNode *N, unsigned indent, 6021 const SelectionDAG *G, VisitedSDNodeSet &once) { 6022 if (!once.insert(N)) // If we've been here before, return now. 6023 return; 6024 // Dump the current SDNode, but don't end the line yet. 6025 OS << std::string(indent, ' '); 6026 N->printr(OS, G); 6027 // Having printed this SDNode, walk the children: 6028 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 6029 const SDNode *child = N->getOperand(i).getNode(); 6030 if (i) OS << ","; 6031 OS << " "; 6032 if (child->getNumOperands() == 0) { 6033 // This child has no grandchildren; print it inline right here. 6034 child->printr(OS, G); 6035 once.insert(child); 6036 } else { // Just the address. FIXME: also print the child's opcode 6037 OS << (void*)child; 6038 if (unsigned RN = N->getOperand(i).getResNo()) 6039 OS << ":" << RN; 6040 } 6041 } 6042 OS << "\n"; 6043 // Dump children that have grandchildren on their own line(s). 6044 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) { 6045 const SDNode *child = N->getOperand(i).getNode(); 6046 DumpNodesr(OS, child, indent+2, G, once); 6047 } 6048} 6049 6050void SDNode::dumpr() const { 6051 VisitedSDNodeSet once; 6052 DumpNodesr(errs(), this, 0, 0, once); 6053} 6054 6055void SDNode::dumpr(const SelectionDAG *G) const { 6056 VisitedSDNodeSet once; 6057 DumpNodesr(errs(), this, 0, G, once); 6058} 6059 6060 6061// getAddressSpace - Return the address space this GlobalAddress belongs to. 6062unsigned GlobalAddressSDNode::getAddressSpace() const { 6063 return getGlobal()->getType()->getAddressSpace(); 6064} 6065 6066 6067const Type *ConstantPoolSDNode::getType() const { 6068 if (isMachineConstantPoolEntry()) 6069 return Val.MachineCPVal->getType(); 6070 return Val.ConstVal->getType(); 6071} 6072 6073bool BuildVectorSDNode::isConstantSplat(APInt &SplatValue, 6074 APInt &SplatUndef, 6075 unsigned &SplatBitSize, 6076 bool &HasAnyUndefs, 6077 unsigned MinSplatBits, 6078 bool isBigEndian) { 6079 EVT VT = getValueType(0); 6080 assert(VT.isVector() && "Expected a vector type"); 6081 unsigned sz = VT.getSizeInBits(); 6082 if (MinSplatBits > sz) 6083 return false; 6084 6085 SplatValue = APInt(sz, 0); 6086 SplatUndef = APInt(sz, 0); 6087 6088 // Get the bits. Bits with undefined values (when the corresponding element 6089 // of the vector is an ISD::UNDEF value) are set in SplatUndef and cleared 6090 // in SplatValue. If any of the values are not constant, give up and return 6091 // false. 6092 unsigned int nOps = getNumOperands(); 6093 assert(nOps > 0 && "isConstantSplat has 0-size build vector"); 6094 unsigned EltBitSize = VT.getVectorElementType().getSizeInBits(); 6095 6096 for (unsigned j = 0; j < nOps; ++j) { 6097 unsigned i = isBigEndian ? nOps-1-j : j; 6098 SDValue OpVal = getOperand(i); 6099 unsigned BitPos = j * EltBitSize; 6100 6101 if (OpVal.getOpcode() == ISD::UNDEF) 6102 SplatUndef |= APInt::getBitsSet(sz, BitPos, BitPos + EltBitSize); 6103 else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) 6104 SplatValue |= (APInt(CN->getAPIntValue()).zextOrTrunc(EltBitSize). 6105 zextOrTrunc(sz) << BitPos); 6106 else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) 6107 SplatValue |= CN->getValueAPF().bitcastToAPInt().zextOrTrunc(sz) <<BitPos; 6108 else 6109 return false; 6110 } 6111 6112 // The build_vector is all constants or undefs. Find the smallest element 6113 // size that splats the vector. 6114 6115 HasAnyUndefs = (SplatUndef != 0); 6116 while (sz > 8) { 6117 6118 unsigned HalfSize = sz / 2; 6119 APInt HighValue = APInt(SplatValue).lshr(HalfSize).trunc(HalfSize); 6120 APInt LowValue = APInt(SplatValue).trunc(HalfSize); 6121 APInt HighUndef = APInt(SplatUndef).lshr(HalfSize).trunc(HalfSize); 6122 APInt LowUndef = APInt(SplatUndef).trunc(HalfSize); 6123 6124 // If the two halves do not match (ignoring undef bits), stop here. 6125 if ((HighValue & ~LowUndef) != (LowValue & ~HighUndef) || 6126 MinSplatBits > HalfSize) 6127 break; 6128 6129 SplatValue = HighValue | LowValue; 6130 SplatUndef = HighUndef & LowUndef; 6131 6132 sz = HalfSize; 6133 } 6134 6135 SplatBitSize = sz; 6136 return true; 6137} 6138 6139bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) { 6140 // Find the first non-undef value in the shuffle mask. 6141 unsigned i, e; 6142 for (i = 0, e = VT.getVectorNumElements(); i != e && Mask[i] < 0; ++i) 6143 /* search */; 6144 6145 assert(i != e && "VECTOR_SHUFFLE node with all undef indices!"); 6146 6147 // Make sure all remaining elements are either undef or the same as the first 6148 // non-undef value. 6149 for (int Idx = Mask[i]; i != e; ++i) 6150 if (Mask[i] >= 0 && Mask[i] != Idx) 6151 return false; 6152 return true; 6153} 6154 6155