SelectionDAGISel.cpp revision 7294b58ecdb430d1473d0a9220a234e62a710d54
1//===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This implements the SelectionDAGISel class. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "isel" 15#include "llvm/ADT/BitVector.h" 16#include "llvm/Analysis/AliasAnalysis.h" 17#include "llvm/CodeGen/SelectionDAGISel.h" 18#include "llvm/CodeGen/ScheduleDAG.h" 19#include "llvm/Constants.h" 20#include "llvm/CallingConv.h" 21#include "llvm/DerivedTypes.h" 22#include "llvm/Function.h" 23#include "llvm/GlobalVariable.h" 24#include "llvm/InlineAsm.h" 25#include "llvm/Instructions.h" 26#include "llvm/Intrinsics.h" 27#include "llvm/IntrinsicInst.h" 28#include "llvm/ParameterAttributes.h" 29#include "llvm/CodeGen/MachineModuleInfo.h" 30#include "llvm/CodeGen/MachineFunction.h" 31#include "llvm/CodeGen/MachineFrameInfo.h" 32#include "llvm/CodeGen/MachineJumpTableInfo.h" 33#include "llvm/CodeGen/MachineInstrBuilder.h" 34#include "llvm/CodeGen/SchedulerRegistry.h" 35#include "llvm/CodeGen/SelectionDAG.h" 36#include "llvm/CodeGen/SSARegMap.h" 37#include "llvm/Target/MRegisterInfo.h" 38#include "llvm/Target/TargetData.h" 39#include "llvm/Target/TargetFrameInfo.h" 40#include "llvm/Target/TargetInstrInfo.h" 41#include "llvm/Target/TargetLowering.h" 42#include "llvm/Target/TargetMachine.h" 43#include "llvm/Target/TargetOptions.h" 44#include "llvm/Support/MathExtras.h" 45#include "llvm/Support/Debug.h" 46#include "llvm/Support/Compiler.h" 47#include <algorithm> 48using namespace llvm; 49 50#ifndef NDEBUG 51static cl::opt<bool> 52ViewISelDAGs("view-isel-dags", cl::Hidden, 53 cl::desc("Pop up a window to show isel dags as they are selected")); 54static cl::opt<bool> 55ViewSchedDAGs("view-sched-dags", cl::Hidden, 56 cl::desc("Pop up a window to show sched dags as they are processed")); 57#else 58static const bool ViewISelDAGs = 0, ViewSchedDAGs = 0; 59#endif 60 61//===---------------------------------------------------------------------===// 62/// 63/// RegisterScheduler class - Track the registration of instruction schedulers. 64/// 65//===---------------------------------------------------------------------===// 66MachinePassRegistry RegisterScheduler::Registry; 67 68//===---------------------------------------------------------------------===// 69/// 70/// ISHeuristic command line option for instruction schedulers. 71/// 72//===---------------------------------------------------------------------===// 73namespace { 74 cl::opt<RegisterScheduler::FunctionPassCtor, false, 75 RegisterPassParser<RegisterScheduler> > 76 ISHeuristic("sched", 77 cl::init(&createDefaultScheduler), 78 cl::desc("Instruction schedulers available:")); 79 80 static RegisterScheduler 81 defaultListDAGScheduler("default", " Best scheduler for the target", 82 createDefaultScheduler); 83} // namespace 84 85namespace { struct AsmOperandInfo; } 86 87namespace { 88 /// RegsForValue - This struct represents the physical registers that a 89 /// particular value is assigned and the type information about the value. 90 /// This is needed because values can be promoted into larger registers and 91 /// expanded into multiple smaller registers than the value. 92 struct VISIBILITY_HIDDEN RegsForValue { 93 /// Regs - This list hold the register (for legal and promoted values) 94 /// or register set (for expanded values) that the value should be assigned 95 /// to. 96 std::vector<unsigned> Regs; 97 98 /// RegVT - The value type of each register. 99 /// 100 MVT::ValueType RegVT; 101 102 /// ValueVT - The value type of the LLVM value, which may be promoted from 103 /// RegVT or made from merging the two expanded parts. 104 MVT::ValueType ValueVT; 105 106 RegsForValue() : RegVT(MVT::Other), ValueVT(MVT::Other) {} 107 108 RegsForValue(unsigned Reg, MVT::ValueType regvt, MVT::ValueType valuevt) 109 : RegVT(regvt), ValueVT(valuevt) { 110 Regs.push_back(Reg); 111 } 112 RegsForValue(const std::vector<unsigned> ®s, 113 MVT::ValueType regvt, MVT::ValueType valuevt) 114 : Regs(regs), RegVT(regvt), ValueVT(valuevt) { 115 } 116 117 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from 118 /// this value and returns the result as a ValueVT value. This uses 119 /// Chain/Flag as the input and updates them for the output Chain/Flag. 120 SDOperand getCopyFromRegs(SelectionDAG &DAG, 121 SDOperand &Chain, SDOperand &Flag) const; 122 123 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 124 /// specified value into the registers specified by this object. This uses 125 /// Chain/Flag as the input and updates them for the output Chain/Flag. 126 void getCopyToRegs(SDOperand Val, SelectionDAG &DAG, 127 SDOperand &Chain, SDOperand &Flag, 128 MVT::ValueType PtrVT) const; 129 130 /// AddInlineAsmOperands - Add this value to the specified inlineasm node 131 /// operand list. This adds the code marker and includes the number of 132 /// values added into it. 133 void AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG, 134 std::vector<SDOperand> &Ops) const; 135 }; 136} 137 138namespace llvm { 139 //===--------------------------------------------------------------------===// 140 /// createDefaultScheduler - This creates an instruction scheduler appropriate 141 /// for the target. 142 ScheduleDAG* createDefaultScheduler(SelectionDAGISel *IS, 143 SelectionDAG *DAG, 144 MachineBasicBlock *BB) { 145 TargetLowering &TLI = IS->getTargetLowering(); 146 147 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) { 148 return createTDListDAGScheduler(IS, DAG, BB); 149 } else { 150 assert(TLI.getSchedulingPreference() == 151 TargetLowering::SchedulingForRegPressure && "Unknown sched type!"); 152 return createBURRListDAGScheduler(IS, DAG, BB); 153 } 154 } 155 156 157 //===--------------------------------------------------------------------===// 158 /// FunctionLoweringInfo - This contains information that is global to a 159 /// function that is used when lowering a region of the function. 160 class FunctionLoweringInfo { 161 public: 162 TargetLowering &TLI; 163 Function &Fn; 164 MachineFunction &MF; 165 SSARegMap *RegMap; 166 167 FunctionLoweringInfo(TargetLowering &TLI, Function &Fn,MachineFunction &MF); 168 169 /// MBBMap - A mapping from LLVM basic blocks to their machine code entry. 170 std::map<const BasicBlock*, MachineBasicBlock *> MBBMap; 171 172 /// ValueMap - Since we emit code for the function a basic block at a time, 173 /// we must remember which virtual registers hold the values for 174 /// cross-basic-block values. 175 DenseMap<const Value*, unsigned> ValueMap; 176 177 /// StaticAllocaMap - Keep track of frame indices for fixed sized allocas in 178 /// the entry block. This allows the allocas to be efficiently referenced 179 /// anywhere in the function. 180 std::map<const AllocaInst*, int> StaticAllocaMap; 181 182 unsigned MakeReg(MVT::ValueType VT) { 183 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT)); 184 } 185 186 /// isExportedInst - Return true if the specified value is an instruction 187 /// exported from its block. 188 bool isExportedInst(const Value *V) { 189 return ValueMap.count(V); 190 } 191 192 unsigned CreateRegForValue(const Value *V); 193 194 unsigned InitializeRegForValue(const Value *V) { 195 unsigned &R = ValueMap[V]; 196 assert(R == 0 && "Already initialized this value register!"); 197 return R = CreateRegForValue(V); 198 } 199 }; 200} 201 202/// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by 203/// PHI nodes or outside of the basic block that defines it, or used by a 204/// switch instruction, which may expand to multiple basic blocks. 205static bool isUsedOutsideOfDefiningBlock(Instruction *I) { 206 if (isa<PHINode>(I)) return true; 207 BasicBlock *BB = I->getParent(); 208 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI) 209 if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI) || 210 // FIXME: Remove switchinst special case. 211 isa<SwitchInst>(*UI)) 212 return true; 213 return false; 214} 215 216/// isOnlyUsedInEntryBlock - If the specified argument is only used in the 217/// entry block, return true. This includes arguments used by switches, since 218/// the switch may expand into multiple basic blocks. 219static bool isOnlyUsedInEntryBlock(Argument *A) { 220 BasicBlock *Entry = A->getParent()->begin(); 221 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI) 222 if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI)) 223 return false; // Use not in entry block. 224 return true; 225} 226 227FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli, 228 Function &fn, MachineFunction &mf) 229 : TLI(tli), Fn(fn), MF(mf), RegMap(MF.getSSARegMap()) { 230 231 // Create a vreg for each argument register that is not dead and is used 232 // outside of the entry block for the function. 233 for (Function::arg_iterator AI = Fn.arg_begin(), E = Fn.arg_end(); 234 AI != E; ++AI) 235 if (!isOnlyUsedInEntryBlock(AI)) 236 InitializeRegForValue(AI); 237 238 // Initialize the mapping of values to registers. This is only set up for 239 // instruction values that are used outside of the block that defines 240 // them. 241 Function::iterator BB = Fn.begin(), EB = Fn.end(); 242 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 243 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) 244 if (ConstantInt *CUI = dyn_cast<ConstantInt>(AI->getArraySize())) { 245 const Type *Ty = AI->getAllocatedType(); 246 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty); 247 unsigned Align = 248 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), 249 AI->getAlignment()); 250 251 TySize *= CUI->getZExtValue(); // Get total allocated size. 252 if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects. 253 StaticAllocaMap[AI] = 254 MF.getFrameInfo()->CreateStackObject(TySize, Align); 255 } 256 257 for (; BB != EB; ++BB) 258 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 259 if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I)) 260 if (!isa<AllocaInst>(I) || 261 !StaticAllocaMap.count(cast<AllocaInst>(I))) 262 InitializeRegForValue(I); 263 264 // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This 265 // also creates the initial PHI MachineInstrs, though none of the input 266 // operands are populated. 267 for (BB = Fn.begin(), EB = Fn.end(); BB != EB; ++BB) { 268 MachineBasicBlock *MBB = new MachineBasicBlock(BB); 269 MBBMap[BB] = MBB; 270 MF.getBasicBlockList().push_back(MBB); 271 272 // Create Machine PHI nodes for LLVM PHI nodes, lowering them as 273 // appropriate. 274 PHINode *PN; 275 for (BasicBlock::iterator I = BB->begin();(PN = dyn_cast<PHINode>(I)); ++I){ 276 if (PN->use_empty()) continue; 277 278 MVT::ValueType VT = TLI.getValueType(PN->getType()); 279 unsigned NumElements; 280 if (VT != MVT::Vector) 281 NumElements = TLI.getNumElements(VT); 282 else { 283 MVT::ValueType VT1,VT2; 284 NumElements = 285 TLI.getVectorTypeBreakdown(cast<VectorType>(PN->getType()), 286 VT1, VT2); 287 } 288 unsigned PHIReg = ValueMap[PN]; 289 assert(PHIReg && "PHI node does not have an assigned virtual register!"); 290 const TargetInstrInfo *TII = TLI.getTargetMachine().getInstrInfo(); 291 for (unsigned i = 0; i != NumElements; ++i) 292 BuildMI(MBB, TII->get(TargetInstrInfo::PHI), PHIReg+i); 293 } 294 } 295} 296 297/// CreateRegForValue - Allocate the appropriate number of virtual registers of 298/// the correctly promoted or expanded types. Assign these registers 299/// consecutive vreg numbers and return the first assigned number. 300unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) { 301 MVT::ValueType VT = TLI.getValueType(V->getType()); 302 303 // The number of multiples of registers that we need, to, e.g., split up 304 // a <2 x int64> -> 4 x i32 registers. 305 unsigned NumVectorRegs = 1; 306 307 // If this is a vector type, figure out what type it will decompose into 308 // and how many of the elements it will use. 309 if (VT == MVT::Vector) { 310 const VectorType *PTy = cast<VectorType>(V->getType()); 311 unsigned NumElts = PTy->getNumElements(); 312 MVT::ValueType EltTy = TLI.getValueType(PTy->getElementType()); 313 MVT::ValueType VecTy = getVectorType(EltTy, NumElts); 314 315 // Divide the input until we get to a supported size. This will always 316 // end with a scalar if the target doesn't support vectors. 317 while (NumElts > 1 && !TLI.isTypeLegal(VecTy)) { 318 NumElts >>= 1; 319 NumVectorRegs <<= 1; 320 VecTy = getVectorType(EltTy, NumElts); 321 } 322 323 // Check that VecTy isn't a 1-element vector. 324 if (NumElts == 1 && VecTy == MVT::Other) 325 VT = EltTy; 326 else 327 VT = VecTy; 328 } 329 330 // The common case is that we will only create one register for this 331 // value. If we have that case, create and return the virtual register. 332 unsigned NV = TLI.getNumElements(VT); 333 if (NV == 1) { 334 // If we are promoting this value, pick the next largest supported type. 335 MVT::ValueType PromotedType = TLI.getTypeToTransformTo(VT); 336 unsigned Reg = MakeReg(PromotedType); 337 // If this is a vector of supported or promoted types (e.g. 4 x i16), 338 // create all of the registers. 339 for (unsigned i = 1; i != NumVectorRegs; ++i) 340 MakeReg(PromotedType); 341 return Reg; 342 } 343 344 // If this value is represented with multiple target registers, make sure 345 // to create enough consecutive registers of the right (smaller) type. 346 VT = TLI.getTypeToExpandTo(VT); 347 unsigned R = MakeReg(VT); 348 for (unsigned i = 1; i != NV*NumVectorRegs; ++i) 349 MakeReg(VT); 350 return R; 351} 352 353//===----------------------------------------------------------------------===// 354/// SelectionDAGLowering - This is the common target-independent lowering 355/// implementation that is parameterized by a TargetLowering object. 356/// Also, targets can overload any lowering method. 357/// 358namespace llvm { 359class SelectionDAGLowering { 360 MachineBasicBlock *CurMBB; 361 362 DenseMap<const Value*, SDOperand> NodeMap; 363 364 /// PendingLoads - Loads are not emitted to the program immediately. We bunch 365 /// them up and then emit token factor nodes when possible. This allows us to 366 /// get simple disambiguation between loads without worrying about alias 367 /// analysis. 368 std::vector<SDOperand> PendingLoads; 369 370 /// Case - A struct to record the Value for a switch case, and the 371 /// case's target basic block. 372 struct Case { 373 Constant* Low; 374 Constant* High; 375 MachineBasicBlock* BB; 376 377 Case() : Low(0), High(0), BB(0) { } 378 Case(Constant* low, Constant* high, MachineBasicBlock* bb) : 379 Low(low), High(high), BB(bb) { } 380 uint64_t size() const { 381 uint64_t rHigh = cast<ConstantInt>(High)->getSExtValue(); 382 uint64_t rLow = cast<ConstantInt>(Low)->getSExtValue(); 383 return (rHigh - rLow + 1ULL); 384 } 385 }; 386 387 struct CaseBits { 388 uint64_t Mask; 389 MachineBasicBlock* BB; 390 unsigned Bits; 391 392 CaseBits(uint64_t mask, MachineBasicBlock* bb, unsigned bits): 393 Mask(mask), BB(bb), Bits(bits) { } 394 }; 395 396 typedef std::vector<Case> CaseVector; 397 typedef std::vector<CaseBits> CaseBitsVector; 398 typedef CaseVector::iterator CaseItr; 399 typedef std::pair<CaseItr, CaseItr> CaseRange; 400 401 /// CaseRec - A struct with ctor used in lowering switches to a binary tree 402 /// of conditional branches. 403 struct CaseRec { 404 CaseRec(MachineBasicBlock *bb, Constant *lt, Constant *ge, CaseRange r) : 405 CaseBB(bb), LT(lt), GE(ge), Range(r) {} 406 407 /// CaseBB - The MBB in which to emit the compare and branch 408 MachineBasicBlock *CaseBB; 409 /// LT, GE - If nonzero, we know the current case value must be less-than or 410 /// greater-than-or-equal-to these Constants. 411 Constant *LT; 412 Constant *GE; 413 /// Range - A pair of iterators representing the range of case values to be 414 /// processed at this point in the binary search tree. 415 CaseRange Range; 416 }; 417 418 typedef std::vector<CaseRec> CaseRecVector; 419 420 /// The comparison function for sorting the switch case values in the vector. 421 /// WARNING: Case ranges should be disjoint! 422 struct CaseCmp { 423 bool operator () (const Case& C1, const Case& C2) { 424 assert(isa<ConstantInt>(C1.Low) && isa<ConstantInt>(C2.High)); 425 const ConstantInt* CI1 = cast<const ConstantInt>(C1.Low); 426 const ConstantInt* CI2 = cast<const ConstantInt>(C2.High); 427 return CI1->getValue().slt(CI2->getValue()); 428 } 429 }; 430 431 struct CaseBitsCmp { 432 bool operator () (const CaseBits& C1, const CaseBits& C2) { 433 return C1.Bits > C2.Bits; 434 } 435 }; 436 437 unsigned Clusterify(CaseVector& Cases, const SwitchInst &SI); 438 439public: 440 // TLI - This is information that describes the available target features we 441 // need for lowering. This indicates when operations are unavailable, 442 // implemented with a libcall, etc. 443 TargetLowering &TLI; 444 SelectionDAG &DAG; 445 const TargetData *TD; 446 447 /// SwitchCases - Vector of CaseBlock structures used to communicate 448 /// SwitchInst code generation information. 449 std::vector<SelectionDAGISel::CaseBlock> SwitchCases; 450 /// JTCases - Vector of JumpTable structures used to communicate 451 /// SwitchInst code generation information. 452 std::vector<SelectionDAGISel::JumpTableBlock> JTCases; 453 std::vector<SelectionDAGISel::BitTestBlock> BitTestCases; 454 455 /// FuncInfo - Information about the function as a whole. 456 /// 457 FunctionLoweringInfo &FuncInfo; 458 459 SelectionDAGLowering(SelectionDAG &dag, TargetLowering &tli, 460 FunctionLoweringInfo &funcinfo) 461 : TLI(tli), DAG(dag), TD(DAG.getTarget().getTargetData()), 462 FuncInfo(funcinfo) { 463 } 464 465 /// getRoot - Return the current virtual root of the Selection DAG. 466 /// 467 SDOperand getRoot() { 468 if (PendingLoads.empty()) 469 return DAG.getRoot(); 470 471 if (PendingLoads.size() == 1) { 472 SDOperand Root = PendingLoads[0]; 473 DAG.setRoot(Root); 474 PendingLoads.clear(); 475 return Root; 476 } 477 478 // Otherwise, we have to make a token factor node. 479 SDOperand Root = DAG.getNode(ISD::TokenFactor, MVT::Other, 480 &PendingLoads[0], PendingLoads.size()); 481 PendingLoads.clear(); 482 DAG.setRoot(Root); 483 return Root; 484 } 485 486 SDOperand CopyValueToVirtualRegister(Value *V, unsigned Reg); 487 488 void visit(Instruction &I) { visit(I.getOpcode(), I); } 489 490 void visit(unsigned Opcode, User &I) { 491 // Note: this doesn't use InstVisitor, because it has to work with 492 // ConstantExpr's in addition to instructions. 493 switch (Opcode) { 494 default: assert(0 && "Unknown instruction type encountered!"); 495 abort(); 496 // Build the switch statement using the Instruction.def file. 497#define HANDLE_INST(NUM, OPCODE, CLASS) \ 498 case Instruction::OPCODE:return visit##OPCODE((CLASS&)I); 499#include "llvm/Instruction.def" 500 } 501 } 502 503 void setCurrentBasicBlock(MachineBasicBlock *MBB) { CurMBB = MBB; } 504 505 SDOperand getLoadFrom(const Type *Ty, SDOperand Ptr, 506 const Value *SV, SDOperand Root, 507 bool isVolatile, unsigned Alignment); 508 509 SDOperand getIntPtrConstant(uint64_t Val) { 510 return DAG.getConstant(Val, TLI.getPointerTy()); 511 } 512 513 SDOperand getValue(const Value *V); 514 515 void setValue(const Value *V, SDOperand NewN) { 516 SDOperand &N = NodeMap[V]; 517 assert(N.Val == 0 && "Already set a value for this node!"); 518 N = NewN; 519 } 520 521 void GetRegistersForValue(AsmOperandInfo &OpInfo, bool HasEarlyClobber, 522 std::set<unsigned> &OutputRegs, 523 std::set<unsigned> &InputRegs); 524 525 void FindMergedConditions(Value *Cond, MachineBasicBlock *TBB, 526 MachineBasicBlock *FBB, MachineBasicBlock *CurBB, 527 unsigned Opc); 528 bool isExportableFromCurrentBlock(Value *V, const BasicBlock *FromBB); 529 void ExportFromCurrentBlock(Value *V); 530 void LowerCallTo(Instruction &I, 531 const Type *CalledValueTy, unsigned CallingConv, 532 bool IsTailCall, SDOperand Callee, unsigned OpIdx); 533 534 // Terminator instructions. 535 void visitRet(ReturnInst &I); 536 void visitBr(BranchInst &I); 537 void visitSwitch(SwitchInst &I); 538 void visitUnreachable(UnreachableInst &I) { /* noop */ } 539 540 // Helpers for visitSwitch 541 bool handleSmallSwitchRange(CaseRec& CR, 542 CaseRecVector& WorkList, 543 Value* SV, 544 MachineBasicBlock* Default); 545 bool handleJTSwitchCase(CaseRec& CR, 546 CaseRecVector& WorkList, 547 Value* SV, 548 MachineBasicBlock* Default); 549 bool handleBTSplitSwitchCase(CaseRec& CR, 550 CaseRecVector& WorkList, 551 Value* SV, 552 MachineBasicBlock* Default); 553 bool handleBitTestsSwitchCase(CaseRec& CR, 554 CaseRecVector& WorkList, 555 Value* SV, 556 MachineBasicBlock* Default); 557 void visitSwitchCase(SelectionDAGISel::CaseBlock &CB); 558 void visitBitTestHeader(SelectionDAGISel::BitTestBlock &B); 559 void visitBitTestCase(MachineBasicBlock* NextMBB, 560 unsigned Reg, 561 SelectionDAGISel::BitTestCase &B); 562 void visitJumpTable(SelectionDAGISel::JumpTable &JT); 563 void visitJumpTableHeader(SelectionDAGISel::JumpTable &JT, 564 SelectionDAGISel::JumpTableHeader &JTH); 565 566 // These all get lowered before this pass. 567 void visitInvoke(InvokeInst &I); 568 void visitInvoke(InvokeInst &I, bool AsTerminator); 569 void visitUnwind(UnwindInst &I); 570 571 void visitScalarBinary(User &I, unsigned OpCode); 572 void visitVectorBinary(User &I, unsigned OpCode); 573 void visitEitherBinary(User &I, unsigned ScalarOp, unsigned VectorOp); 574 void visitShift(User &I, unsigned Opcode); 575 void visitAdd(User &I) { 576 if (isa<VectorType>(I.getType())) 577 visitVectorBinary(I, ISD::VADD); 578 else if (I.getType()->isFloatingPoint()) 579 visitScalarBinary(I, ISD::FADD); 580 else 581 visitScalarBinary(I, ISD::ADD); 582 } 583 void visitSub(User &I); 584 void visitMul(User &I) { 585 if (isa<VectorType>(I.getType())) 586 visitVectorBinary(I, ISD::VMUL); 587 else if (I.getType()->isFloatingPoint()) 588 visitScalarBinary(I, ISD::FMUL); 589 else 590 visitScalarBinary(I, ISD::MUL); 591 } 592 void visitURem(User &I) { visitScalarBinary(I, ISD::UREM); } 593 void visitSRem(User &I) { visitScalarBinary(I, ISD::SREM); } 594 void visitFRem(User &I) { visitScalarBinary(I, ISD::FREM); } 595 void visitUDiv(User &I) { visitEitherBinary(I, ISD::UDIV, ISD::VUDIV); } 596 void visitSDiv(User &I) { visitEitherBinary(I, ISD::SDIV, ISD::VSDIV); } 597 void visitFDiv(User &I) { visitEitherBinary(I, ISD::FDIV, ISD::VSDIV); } 598 void visitAnd (User &I) { visitEitherBinary(I, ISD::AND, ISD::VAND ); } 599 void visitOr (User &I) { visitEitherBinary(I, ISD::OR, ISD::VOR ); } 600 void visitXor (User &I) { visitEitherBinary(I, ISD::XOR, ISD::VXOR ); } 601 void visitShl (User &I) { visitShift(I, ISD::SHL); } 602 void visitLShr(User &I) { visitShift(I, ISD::SRL); } 603 void visitAShr(User &I) { visitShift(I, ISD::SRA); } 604 void visitICmp(User &I); 605 void visitFCmp(User &I); 606 // Visit the conversion instructions 607 void visitTrunc(User &I); 608 void visitZExt(User &I); 609 void visitSExt(User &I); 610 void visitFPTrunc(User &I); 611 void visitFPExt(User &I); 612 void visitFPToUI(User &I); 613 void visitFPToSI(User &I); 614 void visitUIToFP(User &I); 615 void visitSIToFP(User &I); 616 void visitPtrToInt(User &I); 617 void visitIntToPtr(User &I); 618 void visitBitCast(User &I); 619 620 void visitExtractElement(User &I); 621 void visitInsertElement(User &I); 622 void visitShuffleVector(User &I); 623 624 void visitGetElementPtr(User &I); 625 void visitSelect(User &I); 626 627 void visitMalloc(MallocInst &I); 628 void visitFree(FreeInst &I); 629 void visitAlloca(AllocaInst &I); 630 void visitLoad(LoadInst &I); 631 void visitStore(StoreInst &I); 632 void visitPHI(PHINode &I) { } // PHI nodes are handled specially. 633 void visitCall(CallInst &I); 634 void visitInlineAsm(CallInst &I); 635 const char *visitIntrinsicCall(CallInst &I, unsigned Intrinsic); 636 void visitTargetIntrinsic(CallInst &I, unsigned Intrinsic); 637 638 void visitVAStart(CallInst &I); 639 void visitVAArg(VAArgInst &I); 640 void visitVAEnd(CallInst &I); 641 void visitVACopy(CallInst &I); 642 643 void visitMemIntrinsic(CallInst &I, unsigned Op); 644 645 void visitUserOp1(Instruction &I) { 646 assert(0 && "UserOp1 should not exist at instruction selection time!"); 647 abort(); 648 } 649 void visitUserOp2(Instruction &I) { 650 assert(0 && "UserOp2 should not exist at instruction selection time!"); 651 abort(); 652 } 653}; 654} // end namespace llvm 655 656SDOperand SelectionDAGLowering::getValue(const Value *V) { 657 SDOperand &N = NodeMap[V]; 658 if (N.Val) return N; 659 660 const Type *VTy = V->getType(); 661 MVT::ValueType VT = TLI.getValueType(VTy); 662 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) { 663 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 664 visit(CE->getOpcode(), *CE); 665 SDOperand N1 = NodeMap[V]; 666 assert(N1.Val && "visit didn't populate the ValueMap!"); 667 return N1; 668 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(C)) { 669 return N = DAG.getGlobalAddress(GV, VT); 670 } else if (isa<ConstantPointerNull>(C)) { 671 return N = DAG.getConstant(0, TLI.getPointerTy()); 672 } else if (isa<UndefValue>(C)) { 673 if (!isa<VectorType>(VTy)) 674 return N = DAG.getNode(ISD::UNDEF, VT); 675 676 // Create a VBUILD_VECTOR of undef nodes. 677 const VectorType *PTy = cast<VectorType>(VTy); 678 unsigned NumElements = PTy->getNumElements(); 679 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType()); 680 681 SmallVector<SDOperand, 8> Ops; 682 Ops.assign(NumElements, DAG.getNode(ISD::UNDEF, PVT)); 683 684 // Create a VConstant node with generic Vector type. 685 Ops.push_back(DAG.getConstant(NumElements, MVT::i32)); 686 Ops.push_back(DAG.getValueType(PVT)); 687 return N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, 688 &Ops[0], Ops.size()); 689 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 690 return N = DAG.getConstantFP(CFP->getValue(), VT); 691 } else if (const VectorType *PTy = dyn_cast<VectorType>(VTy)) { 692 unsigned NumElements = PTy->getNumElements(); 693 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType()); 694 695 // Now that we know the number and type of the elements, push a 696 // Constant or ConstantFP node onto the ops list for each element of 697 // the packed constant. 698 SmallVector<SDOperand, 8> Ops; 699 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) { 700 for (unsigned i = 0; i != NumElements; ++i) 701 Ops.push_back(getValue(CP->getOperand(i))); 702 } else { 703 assert(isa<ConstantAggregateZero>(C) && "Unknown packed constant!"); 704 SDOperand Op; 705 if (MVT::isFloatingPoint(PVT)) 706 Op = DAG.getConstantFP(0, PVT); 707 else 708 Op = DAG.getConstant(0, PVT); 709 Ops.assign(NumElements, Op); 710 } 711 712 // Create a VBUILD_VECTOR node with generic Vector type. 713 Ops.push_back(DAG.getConstant(NumElements, MVT::i32)); 714 Ops.push_back(DAG.getValueType(PVT)); 715 return NodeMap[V] = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], 716 Ops.size()); 717 } else { 718 // Canonicalize all constant ints to be unsigned. 719 return N = DAG.getConstant(cast<ConstantInt>(C)->getZExtValue(),VT); 720 } 721 } 722 723 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 724 std::map<const AllocaInst*, int>::iterator SI = 725 FuncInfo.StaticAllocaMap.find(AI); 726 if (SI != FuncInfo.StaticAllocaMap.end()) 727 return DAG.getFrameIndex(SI->second, TLI.getPointerTy()); 728 } 729 730 unsigned InReg = FuncInfo.ValueMap[V]; 731 assert(InReg && "Value not in map!"); 732 733 // If this type is not legal, make it so now. 734 if (VT != MVT::Vector) { 735 if (TLI.getTypeAction(VT) == TargetLowering::Expand) { 736 // Source must be expanded. This input value is actually coming from the 737 // register pair InReg and InReg+1. 738 MVT::ValueType DestVT = TLI.getTypeToExpandTo(VT); 739 unsigned NumVals = TLI.getNumElements(VT); 740 N = DAG.getCopyFromReg(DAG.getEntryNode(), InReg, DestVT); 741 if (NumVals == 1) 742 N = DAG.getNode(ISD::BIT_CONVERT, VT, N); 743 else { 744 assert(NumVals == 2 && "1 to 4 (and more) expansion not implemented!"); 745 N = DAG.getNode(ISD::BUILD_PAIR, VT, N, 746 DAG.getCopyFromReg(DAG.getEntryNode(), InReg+1, DestVT)); 747 } 748 } else { 749 MVT::ValueType DestVT = TLI.getTypeToTransformTo(VT); 750 N = DAG.getCopyFromReg(DAG.getEntryNode(), InReg, DestVT); 751 if (TLI.getTypeAction(VT) == TargetLowering::Promote) // Promotion case 752 N = MVT::isFloatingPoint(VT) 753 ? DAG.getNode(ISD::FP_ROUND, VT, N) 754 : DAG.getNode(ISD::TRUNCATE, VT, N); 755 } 756 } else { 757 // Otherwise, if this is a vector, make it available as a generic vector 758 // here. 759 MVT::ValueType PTyElementVT, PTyLegalElementVT; 760 const VectorType *PTy = cast<VectorType>(VTy); 761 unsigned NE = TLI.getVectorTypeBreakdown(PTy, PTyElementVT, 762 PTyLegalElementVT); 763 764 // Build a VBUILD_VECTOR with the input registers. 765 SmallVector<SDOperand, 8> Ops; 766 if (PTyElementVT == PTyLegalElementVT) { 767 // If the value types are legal, just VBUILD the CopyFromReg nodes. 768 for (unsigned i = 0; i != NE; ++i) 769 Ops.push_back(DAG.getCopyFromReg(DAG.getEntryNode(), InReg++, 770 PTyElementVT)); 771 } else if (PTyElementVT < PTyLegalElementVT) { 772 // If the register was promoted, use TRUNCATE of FP_ROUND as appropriate. 773 for (unsigned i = 0; i != NE; ++i) { 774 SDOperand Op = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++, 775 PTyElementVT); 776 if (MVT::isFloatingPoint(PTyElementVT)) 777 Op = DAG.getNode(ISD::FP_ROUND, PTyElementVT, Op); 778 else 779 Op = DAG.getNode(ISD::TRUNCATE, PTyElementVT, Op); 780 Ops.push_back(Op); 781 } 782 } else { 783 // If the register was expanded, use BUILD_PAIR. 784 assert((NE & 1) == 0 && "Must expand into a multiple of 2 elements!"); 785 for (unsigned i = 0; i != NE/2; ++i) { 786 SDOperand Op0 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++, 787 PTyElementVT); 788 SDOperand Op1 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++, 789 PTyElementVT); 790 Ops.push_back(DAG.getNode(ISD::BUILD_PAIR, VT, Op0, Op1)); 791 } 792 } 793 794 Ops.push_back(DAG.getConstant(NE, MVT::i32)); 795 Ops.push_back(DAG.getValueType(PTyLegalElementVT)); 796 N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size()); 797 798 // Finally, use a VBIT_CONVERT to make this available as the appropriate 799 // vector type. 800 N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N, 801 DAG.getConstant(PTy->getNumElements(), 802 MVT::i32), 803 DAG.getValueType(TLI.getValueType(PTy->getElementType()))); 804 } 805 806 return N; 807} 808 809 810void SelectionDAGLowering::visitRet(ReturnInst &I) { 811 if (I.getNumOperands() == 0) { 812 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, getRoot())); 813 return; 814 } 815 SmallVector<SDOperand, 8> NewValues; 816 NewValues.push_back(getRoot()); 817 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 818 SDOperand RetOp = getValue(I.getOperand(i)); 819 820 // If this is an integer return value, we need to promote it ourselves to 821 // the full width of a register, since LegalizeOp will use ANY_EXTEND rather 822 // than sign/zero. 823 // FIXME: C calling convention requires the return type to be promoted to 824 // at least 32-bit. But this is not necessary for non-C calling conventions. 825 if (MVT::isInteger(RetOp.getValueType()) && 826 RetOp.getValueType() < MVT::i64) { 827 MVT::ValueType TmpVT; 828 if (TLI.getTypeAction(MVT::i32) == TargetLowering::Promote) 829 TmpVT = TLI.getTypeToTransformTo(MVT::i32); 830 else 831 TmpVT = MVT::i32; 832 const FunctionType *FTy = I.getParent()->getParent()->getFunctionType(); 833 const ParamAttrsList *Attrs = FTy->getParamAttrs(); 834 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 835 if (Attrs && Attrs->paramHasAttr(0, ParamAttr::SExt)) 836 ExtendKind = ISD::SIGN_EXTEND; 837 if (Attrs && Attrs->paramHasAttr(0, ParamAttr::ZExt)) 838 ExtendKind = ISD::ZERO_EXTEND; 839 RetOp = DAG.getNode(ExtendKind, TmpVT, RetOp); 840 } 841 NewValues.push_back(RetOp); 842 NewValues.push_back(DAG.getConstant(false, MVT::i32)); 843 } 844 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, 845 &NewValues[0], NewValues.size())); 846} 847 848/// ExportFromCurrentBlock - If this condition isn't known to be exported from 849/// the current basic block, add it to ValueMap now so that we'll get a 850/// CopyTo/FromReg. 851void SelectionDAGLowering::ExportFromCurrentBlock(Value *V) { 852 // No need to export constants. 853 if (!isa<Instruction>(V) && !isa<Argument>(V)) return; 854 855 // Already exported? 856 if (FuncInfo.isExportedInst(V)) return; 857 858 unsigned Reg = FuncInfo.InitializeRegForValue(V); 859 PendingLoads.push_back(CopyValueToVirtualRegister(V, Reg)); 860} 861 862bool SelectionDAGLowering::isExportableFromCurrentBlock(Value *V, 863 const BasicBlock *FromBB) { 864 // The operands of the setcc have to be in this block. We don't know 865 // how to export them from some other block. 866 if (Instruction *VI = dyn_cast<Instruction>(V)) { 867 // Can export from current BB. 868 if (VI->getParent() == FromBB) 869 return true; 870 871 // Is already exported, noop. 872 return FuncInfo.isExportedInst(V); 873 } 874 875 // If this is an argument, we can export it if the BB is the entry block or 876 // if it is already exported. 877 if (isa<Argument>(V)) { 878 if (FromBB == &FromBB->getParent()->getEntryBlock()) 879 return true; 880 881 // Otherwise, can only export this if it is already exported. 882 return FuncInfo.isExportedInst(V); 883 } 884 885 // Otherwise, constants can always be exported. 886 return true; 887} 888 889static bool InBlock(const Value *V, const BasicBlock *BB) { 890 if (const Instruction *I = dyn_cast<Instruction>(V)) 891 return I->getParent() == BB; 892 return true; 893} 894 895/// FindMergedConditions - If Cond is an expression like 896void SelectionDAGLowering::FindMergedConditions(Value *Cond, 897 MachineBasicBlock *TBB, 898 MachineBasicBlock *FBB, 899 MachineBasicBlock *CurBB, 900 unsigned Opc) { 901 // If this node is not part of the or/and tree, emit it as a branch. 902 Instruction *BOp = dyn_cast<Instruction>(Cond); 903 904 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) || 905 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() || 906 BOp->getParent() != CurBB->getBasicBlock() || 907 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) || 908 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) { 909 const BasicBlock *BB = CurBB->getBasicBlock(); 910 911 // If the leaf of the tree is a comparison, merge the condition into 912 // the caseblock. 913 if ((isa<ICmpInst>(Cond) || isa<FCmpInst>(Cond)) && 914 // The operands of the cmp have to be in this block. We don't know 915 // how to export them from some other block. If this is the first block 916 // of the sequence, no exporting is needed. 917 (CurBB == CurMBB || 918 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) && 919 isExportableFromCurrentBlock(BOp->getOperand(1), BB)))) { 920 BOp = cast<Instruction>(Cond); 921 ISD::CondCode Condition; 922 if (ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) { 923 switch (IC->getPredicate()) { 924 default: assert(0 && "Unknown icmp predicate opcode!"); 925 case ICmpInst::ICMP_EQ: Condition = ISD::SETEQ; break; 926 case ICmpInst::ICMP_NE: Condition = ISD::SETNE; break; 927 case ICmpInst::ICMP_SLE: Condition = ISD::SETLE; break; 928 case ICmpInst::ICMP_ULE: Condition = ISD::SETULE; break; 929 case ICmpInst::ICMP_SGE: Condition = ISD::SETGE; break; 930 case ICmpInst::ICMP_UGE: Condition = ISD::SETUGE; break; 931 case ICmpInst::ICMP_SLT: Condition = ISD::SETLT; break; 932 case ICmpInst::ICMP_ULT: Condition = ISD::SETULT; break; 933 case ICmpInst::ICMP_SGT: Condition = ISD::SETGT; break; 934 case ICmpInst::ICMP_UGT: Condition = ISD::SETUGT; break; 935 } 936 } else if (FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) { 937 ISD::CondCode FPC, FOC; 938 switch (FC->getPredicate()) { 939 default: assert(0 && "Unknown fcmp predicate opcode!"); 940 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break; 941 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break; 942 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break; 943 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break; 944 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break; 945 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break; 946 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break; 947 case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break; 948 case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break; 949 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break; 950 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break; 951 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break; 952 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break; 953 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break; 954 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break; 955 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break; 956 } 957 if (FiniteOnlyFPMath()) 958 Condition = FOC; 959 else 960 Condition = FPC; 961 } else { 962 Condition = ISD::SETEQ; // silence warning. 963 assert(0 && "Unknown compare instruction"); 964 } 965 966 SelectionDAGISel::CaseBlock CB(Condition, BOp->getOperand(0), 967 BOp->getOperand(1), NULL, TBB, FBB, CurBB); 968 SwitchCases.push_back(CB); 969 return; 970 } 971 972 // Create a CaseBlock record representing this branch. 973 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(), 974 NULL, TBB, FBB, CurBB); 975 SwitchCases.push_back(CB); 976 return; 977 } 978 979 980 // Create TmpBB after CurBB. 981 MachineFunction::iterator BBI = CurBB; 982 MachineBasicBlock *TmpBB = new MachineBasicBlock(CurBB->getBasicBlock()); 983 CurBB->getParent()->getBasicBlockList().insert(++BBI, TmpBB); 984 985 if (Opc == Instruction::Or) { 986 // Codegen X | Y as: 987 // jmp_if_X TBB 988 // jmp TmpBB 989 // TmpBB: 990 // jmp_if_Y TBB 991 // jmp FBB 992 // 993 994 // Emit the LHS condition. 995 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc); 996 997 // Emit the RHS condition into TmpBB. 998 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc); 999 } else { 1000 assert(Opc == Instruction::And && "Unknown merge op!"); 1001 // Codegen X & Y as: 1002 // jmp_if_X TmpBB 1003 // jmp FBB 1004 // TmpBB: 1005 // jmp_if_Y TBB 1006 // jmp FBB 1007 // 1008 // This requires creation of TmpBB after CurBB. 1009 1010 // Emit the LHS condition. 1011 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc); 1012 1013 // Emit the RHS condition into TmpBB. 1014 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc); 1015 } 1016} 1017 1018/// If the set of cases should be emitted as a series of branches, return true. 1019/// If we should emit this as a bunch of and/or'd together conditions, return 1020/// false. 1021static bool 1022ShouldEmitAsBranches(const std::vector<SelectionDAGISel::CaseBlock> &Cases) { 1023 if (Cases.size() != 2) return true; 1024 1025 // If this is two comparisons of the same values or'd or and'd together, they 1026 // will get folded into a single comparison, so don't emit two blocks. 1027 if ((Cases[0].CmpLHS == Cases[1].CmpLHS && 1028 Cases[0].CmpRHS == Cases[1].CmpRHS) || 1029 (Cases[0].CmpRHS == Cases[1].CmpLHS && 1030 Cases[0].CmpLHS == Cases[1].CmpRHS)) { 1031 return false; 1032 } 1033 1034 return true; 1035} 1036 1037void SelectionDAGLowering::visitBr(BranchInst &I) { 1038 // Update machine-CFG edges. 1039 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)]; 1040 1041 // Figure out which block is immediately after the current one. 1042 MachineBasicBlock *NextBlock = 0; 1043 MachineFunction::iterator BBI = CurMBB; 1044 if (++BBI != CurMBB->getParent()->end()) 1045 NextBlock = BBI; 1046 1047 if (I.isUnconditional()) { 1048 // If this is not a fall-through branch, emit the branch. 1049 if (Succ0MBB != NextBlock) 1050 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(), 1051 DAG.getBasicBlock(Succ0MBB))); 1052 1053 // Update machine-CFG edges. 1054 CurMBB->addSuccessor(Succ0MBB); 1055 1056 return; 1057 } 1058 1059 // If this condition is one of the special cases we handle, do special stuff 1060 // now. 1061 Value *CondVal = I.getCondition(); 1062 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)]; 1063 1064 // If this is a series of conditions that are or'd or and'd together, emit 1065 // this as a sequence of branches instead of setcc's with and/or operations. 1066 // For example, instead of something like: 1067 // cmp A, B 1068 // C = seteq 1069 // cmp D, E 1070 // F = setle 1071 // or C, F 1072 // jnz foo 1073 // Emit: 1074 // cmp A, B 1075 // je foo 1076 // cmp D, E 1077 // jle foo 1078 // 1079 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) { 1080 if (BOp->hasOneUse() && 1081 (BOp->getOpcode() == Instruction::And || 1082 BOp->getOpcode() == Instruction::Or)) { 1083 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode()); 1084 // If the compares in later blocks need to use values not currently 1085 // exported from this block, export them now. This block should always 1086 // be the first entry. 1087 assert(SwitchCases[0].ThisBB == CurMBB && "Unexpected lowering!"); 1088 1089 // Allow some cases to be rejected. 1090 if (ShouldEmitAsBranches(SwitchCases)) { 1091 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) { 1092 ExportFromCurrentBlock(SwitchCases[i].CmpLHS); 1093 ExportFromCurrentBlock(SwitchCases[i].CmpRHS); 1094 } 1095 1096 // Emit the branch for this block. 1097 visitSwitchCase(SwitchCases[0]); 1098 SwitchCases.erase(SwitchCases.begin()); 1099 return; 1100 } 1101 1102 // Okay, we decided not to do this, remove any inserted MBB's and clear 1103 // SwitchCases. 1104 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) 1105 CurMBB->getParent()->getBasicBlockList().erase(SwitchCases[i].ThisBB); 1106 1107 SwitchCases.clear(); 1108 } 1109 } 1110 1111 // Create a CaseBlock record representing this branch. 1112 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(), 1113 NULL, Succ0MBB, Succ1MBB, CurMBB); 1114 // Use visitSwitchCase to actually insert the fast branch sequence for this 1115 // cond branch. 1116 visitSwitchCase(CB); 1117} 1118 1119/// visitSwitchCase - Emits the necessary code to represent a single node in 1120/// the binary search tree resulting from lowering a switch instruction. 1121void SelectionDAGLowering::visitSwitchCase(SelectionDAGISel::CaseBlock &CB) { 1122 SDOperand Cond; 1123 SDOperand CondLHS = getValue(CB.CmpLHS); 1124 1125 // Build the setcc now. 1126 if (CB.CmpMHS == NULL) { 1127 // Fold "(X == true)" to X and "(X == false)" to !X to 1128 // handle common cases produced by branch lowering. 1129 if (CB.CmpRHS == ConstantInt::getTrue() && CB.CC == ISD::SETEQ) 1130 Cond = CondLHS; 1131 else if (CB.CmpRHS == ConstantInt::getFalse() && CB.CC == ISD::SETEQ) { 1132 SDOperand True = DAG.getConstant(1, CondLHS.getValueType()); 1133 Cond = DAG.getNode(ISD::XOR, CondLHS.getValueType(), CondLHS, True); 1134 } else 1135 Cond = DAG.getSetCC(MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC); 1136 } else { 1137 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now"); 1138 1139 uint64_t Low = cast<ConstantInt>(CB.CmpLHS)->getSExtValue(); 1140 uint64_t High = cast<ConstantInt>(CB.CmpRHS)->getSExtValue(); 1141 1142 SDOperand CmpOp = getValue(CB.CmpMHS); 1143 MVT::ValueType VT = CmpOp.getValueType(); 1144 1145 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) { 1146 Cond = DAG.getSetCC(MVT::i1, CmpOp, DAG.getConstant(High, VT), ISD::SETLE); 1147 } else { 1148 SDOperand SUB = DAG.getNode(ISD::SUB, VT, CmpOp, DAG.getConstant(Low, VT)); 1149 Cond = DAG.getSetCC(MVT::i1, SUB, 1150 DAG.getConstant(High-Low, VT), ISD::SETULE); 1151 } 1152 1153 } 1154 1155 // Set NextBlock to be the MBB immediately after the current one, if any. 1156 // This is used to avoid emitting unnecessary branches to the next block. 1157 MachineBasicBlock *NextBlock = 0; 1158 MachineFunction::iterator BBI = CurMBB; 1159 if (++BBI != CurMBB->getParent()->end()) 1160 NextBlock = BBI; 1161 1162 // If the lhs block is the next block, invert the condition so that we can 1163 // fall through to the lhs instead of the rhs block. 1164 if (CB.TrueBB == NextBlock) { 1165 std::swap(CB.TrueBB, CB.FalseBB); 1166 SDOperand True = DAG.getConstant(1, Cond.getValueType()); 1167 Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True); 1168 } 1169 SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(), Cond, 1170 DAG.getBasicBlock(CB.TrueBB)); 1171 if (CB.FalseBB == NextBlock) 1172 DAG.setRoot(BrCond); 1173 else 1174 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond, 1175 DAG.getBasicBlock(CB.FalseBB))); 1176 // Update successor info 1177 CurMBB->addSuccessor(CB.TrueBB); 1178 CurMBB->addSuccessor(CB.FalseBB); 1179} 1180 1181/// visitJumpTable - Emit JumpTable node in the current MBB 1182void SelectionDAGLowering::visitJumpTable(SelectionDAGISel::JumpTable &JT) { 1183 // Emit the code for the jump table 1184 assert(JT.Reg != -1U && "Should lower JT Header first!"); 1185 MVT::ValueType PTy = TLI.getPointerTy(); 1186 SDOperand Index = DAG.getCopyFromReg(getRoot(), JT.Reg, PTy); 1187 SDOperand Table = DAG.getJumpTable(JT.JTI, PTy); 1188 DAG.setRoot(DAG.getNode(ISD::BR_JT, MVT::Other, Index.getValue(1), 1189 Table, Index)); 1190 return; 1191} 1192 1193/// visitJumpTableHeader - This function emits necessary code to produce index 1194/// in the JumpTable from switch case. 1195void SelectionDAGLowering::visitJumpTableHeader(SelectionDAGISel::JumpTable &JT, 1196 SelectionDAGISel::JumpTableHeader &JTH) { 1197 // Subtract the lowest switch case value from the value being switched on 1198 // and conditional branch to default mbb if the result is greater than the 1199 // difference between smallest and largest cases. 1200 SDOperand SwitchOp = getValue(JTH.SValue); 1201 MVT::ValueType VT = SwitchOp.getValueType(); 1202 SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp, 1203 DAG.getConstant(JTH.First, VT)); 1204 1205 // The SDNode we just created, which holds the value being switched on 1206 // minus the the smallest case value, needs to be copied to a virtual 1207 // register so it can be used as an index into the jump table in a 1208 // subsequent basic block. This value may be smaller or larger than the 1209 // target's pointer type, and therefore require extension or truncating. 1210 if (VT > TLI.getPointerTy()) 1211 SwitchOp = DAG.getNode(ISD::TRUNCATE, TLI.getPointerTy(), SUB); 1212 else 1213 SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), SUB); 1214 1215 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy()); 1216 SDOperand CopyTo = DAG.getCopyToReg(getRoot(), JumpTableReg, SwitchOp); 1217 JT.Reg = JumpTableReg; 1218 1219 // Emit the range check for the jump table, and branch to the default 1220 // block for the switch statement if the value being switched on exceeds 1221 // the largest case in the switch. 1222 SDOperand CMP = DAG.getSetCC(TLI.getSetCCResultTy(), SUB, 1223 DAG.getConstant(JTH.Last-JTH.First,VT), 1224 ISD::SETUGT); 1225 1226 // Set NextBlock to be the MBB immediately after the current one, if any. 1227 // This is used to avoid emitting unnecessary branches to the next block. 1228 MachineBasicBlock *NextBlock = 0; 1229 MachineFunction::iterator BBI = CurMBB; 1230 if (++BBI != CurMBB->getParent()->end()) 1231 NextBlock = BBI; 1232 1233 SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, CMP, 1234 DAG.getBasicBlock(JT.Default)); 1235 1236 if (JT.MBB == NextBlock) 1237 DAG.setRoot(BrCond); 1238 else 1239 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond, 1240 DAG.getBasicBlock(JT.MBB))); 1241 1242 return; 1243} 1244 1245/// visitBitTestHeader - This function emits necessary code to produce value 1246/// suitable for "bit tests" 1247void SelectionDAGLowering::visitBitTestHeader(SelectionDAGISel::BitTestBlock &B) { 1248 // Subtract the minimum value 1249 SDOperand SwitchOp = getValue(B.SValue); 1250 MVT::ValueType VT = SwitchOp.getValueType(); 1251 SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp, 1252 DAG.getConstant(B.First, VT)); 1253 1254 // Check range 1255 SDOperand RangeCmp = DAG.getSetCC(TLI.getSetCCResultTy(), SUB, 1256 DAG.getConstant(B.Range, VT), 1257 ISD::SETUGT); 1258 1259 SDOperand ShiftOp; 1260 if (VT > TLI.getShiftAmountTy()) 1261 ShiftOp = DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), SUB); 1262 else 1263 ShiftOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getShiftAmountTy(), SUB); 1264 1265 // Make desired shift 1266 SDOperand SwitchVal = DAG.getNode(ISD::SHL, TLI.getPointerTy(), 1267 DAG.getConstant(1, TLI.getPointerTy()), 1268 ShiftOp); 1269 1270 unsigned SwitchReg = FuncInfo.MakeReg(TLI.getPointerTy()); 1271 SDOperand CopyTo = DAG.getCopyToReg(getRoot(), SwitchReg, SwitchVal); 1272 B.Reg = SwitchReg; 1273 1274 SDOperand BrRange = DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, RangeCmp, 1275 DAG.getBasicBlock(B.Default)); 1276 1277 // Set NextBlock to be the MBB immediately after the current one, if any. 1278 // This is used to avoid emitting unnecessary branches to the next block. 1279 MachineBasicBlock *NextBlock = 0; 1280 MachineFunction::iterator BBI = CurMBB; 1281 if (++BBI != CurMBB->getParent()->end()) 1282 NextBlock = BBI; 1283 1284 MachineBasicBlock* MBB = B.Cases[0].ThisBB; 1285 if (MBB == NextBlock) 1286 DAG.setRoot(BrRange); 1287 else 1288 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, CopyTo, 1289 DAG.getBasicBlock(MBB))); 1290 1291 CurMBB->addSuccessor(B.Default); 1292 CurMBB->addSuccessor(MBB); 1293 1294 return; 1295} 1296 1297/// visitBitTestCase - this function produces one "bit test" 1298void SelectionDAGLowering::visitBitTestCase(MachineBasicBlock* NextMBB, 1299 unsigned Reg, 1300 SelectionDAGISel::BitTestCase &B) { 1301 // Emit bit tests and jumps 1302 SDOperand SwitchVal = DAG.getCopyFromReg(getRoot(), Reg, TLI.getPointerTy()); 1303 1304 SDOperand AndOp = DAG.getNode(ISD::AND, TLI.getPointerTy(), 1305 SwitchVal, 1306 DAG.getConstant(B.Mask, 1307 TLI.getPointerTy())); 1308 SDOperand AndCmp = DAG.getSetCC(TLI.getSetCCResultTy(), AndOp, 1309 DAG.getConstant(0, TLI.getPointerTy()), 1310 ISD::SETNE); 1311 SDOperand BrAnd = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(), 1312 AndCmp, DAG.getBasicBlock(B.TargetBB)); 1313 1314 // Set NextBlock to be the MBB immediately after the current one, if any. 1315 // This is used to avoid emitting unnecessary branches to the next block. 1316 MachineBasicBlock *NextBlock = 0; 1317 MachineFunction::iterator BBI = CurMBB; 1318 if (++BBI != CurMBB->getParent()->end()) 1319 NextBlock = BBI; 1320 1321 if (NextMBB == NextBlock) 1322 DAG.setRoot(BrAnd); 1323 else 1324 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrAnd, 1325 DAG.getBasicBlock(NextMBB))); 1326 1327 CurMBB->addSuccessor(B.TargetBB); 1328 CurMBB->addSuccessor(NextMBB); 1329 1330 return; 1331} 1332 1333void SelectionDAGLowering::visitInvoke(InvokeInst &I) { 1334 assert(0 && "Should never be visited directly"); 1335} 1336void SelectionDAGLowering::visitInvoke(InvokeInst &I, bool AsTerminator) { 1337 // Retrieve successors. 1338 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)]; 1339 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)]; 1340 1341 if (!AsTerminator) { 1342 // Mark landing pad so that it doesn't get deleted in branch folding. 1343 LandingPad->setIsLandingPad(); 1344 1345 // Insert a label before the invoke call to mark the try range. 1346 // This can be used to detect deletion of the invoke via the 1347 // MachineModuleInfo. 1348 MachineModuleInfo *MMI = DAG.getMachineModuleInfo(); 1349 unsigned BeginLabel = MMI->NextLabelID(); 1350 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(), 1351 DAG.getConstant(BeginLabel, MVT::i32))); 1352 1353 LowerCallTo(I, I.getCalledValue()->getType(), 1354 I.getCallingConv(), 1355 false, 1356 getValue(I.getOperand(0)), 1357 3); 1358 1359 // Insert a label before the invoke call to mark the try range. 1360 // This can be used to detect deletion of the invoke via the 1361 // MachineModuleInfo. 1362 unsigned EndLabel = MMI->NextLabelID(); 1363 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(), 1364 DAG.getConstant(EndLabel, MVT::i32))); 1365 1366 // Inform MachineModuleInfo of range. 1367 MMI->addInvoke(LandingPad, BeginLabel, EndLabel); 1368 1369 // Update successor info 1370 CurMBB->addSuccessor(Return); 1371 CurMBB->addSuccessor(LandingPad); 1372 } else { 1373 // Drop into normal successor. 1374 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(), 1375 DAG.getBasicBlock(Return))); 1376 } 1377} 1378 1379void SelectionDAGLowering::visitUnwind(UnwindInst &I) { 1380} 1381 1382/// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for 1383/// small case ranges). 1384bool SelectionDAGLowering::handleSmallSwitchRange(CaseRec& CR, 1385 CaseRecVector& WorkList, 1386 Value* SV, 1387 MachineBasicBlock* Default) { 1388 Case& BackCase = *(CR.Range.second-1); 1389 1390 // Size is the number of Cases represented by this range. 1391 unsigned Size = CR.Range.second - CR.Range.first; 1392 if (Size > 3) 1393 return false; 1394 1395 // Get the MachineFunction which holds the current MBB. This is used when 1396 // inserting any additional MBBs necessary to represent the switch. 1397 MachineFunction *CurMF = CurMBB->getParent(); 1398 1399 // Figure out which block is immediately after the current one. 1400 MachineBasicBlock *NextBlock = 0; 1401 MachineFunction::iterator BBI = CR.CaseBB; 1402 1403 if (++BBI != CurMBB->getParent()->end()) 1404 NextBlock = BBI; 1405 1406 // TODO: If any two of the cases has the same destination, and if one value 1407 // is the same as the other, but has one bit unset that the other has set, 1408 // use bit manipulation to do two compares at once. For example: 1409 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)" 1410 1411 // Rearrange the case blocks so that the last one falls through if possible. 1412 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) { 1413 // The last case block won't fall through into 'NextBlock' if we emit the 1414 // branches in this order. See if rearranging a case value would help. 1415 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) { 1416 if (I->BB == NextBlock) { 1417 std::swap(*I, BackCase); 1418 break; 1419 } 1420 } 1421 } 1422 1423 // Create a CaseBlock record representing a conditional branch to 1424 // the Case's target mbb if the value being switched on SV is equal 1425 // to C. 1426 MachineBasicBlock *CurBlock = CR.CaseBB; 1427 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) { 1428 MachineBasicBlock *FallThrough; 1429 if (I != E-1) { 1430 FallThrough = new MachineBasicBlock(CurBlock->getBasicBlock()); 1431 CurMF->getBasicBlockList().insert(BBI, FallThrough); 1432 } else { 1433 // If the last case doesn't match, go to the default block. 1434 FallThrough = Default; 1435 } 1436 1437 Value *RHS, *LHS, *MHS; 1438 ISD::CondCode CC; 1439 if (I->High == I->Low) { 1440 // This is just small small case range :) containing exactly 1 case 1441 CC = ISD::SETEQ; 1442 LHS = SV; RHS = I->High; MHS = NULL; 1443 } else { 1444 CC = ISD::SETLE; 1445 LHS = I->Low; MHS = SV; RHS = I->High; 1446 } 1447 SelectionDAGISel::CaseBlock CB(CC, LHS, RHS, MHS, 1448 I->BB, FallThrough, CurBlock); 1449 1450 // If emitting the first comparison, just call visitSwitchCase to emit the 1451 // code into the current block. Otherwise, push the CaseBlock onto the 1452 // vector to be later processed by SDISel, and insert the node's MBB 1453 // before the next MBB. 1454 if (CurBlock == CurMBB) 1455 visitSwitchCase(CB); 1456 else 1457 SwitchCases.push_back(CB); 1458 1459 CurBlock = FallThrough; 1460 } 1461 1462 return true; 1463} 1464 1465static inline bool areJTsAllowed(const TargetLowering &TLI) { 1466 return (TLI.isOperationLegal(ISD::BR_JT, MVT::Other) || 1467 TLI.isOperationLegal(ISD::BRIND, MVT::Other)); 1468} 1469 1470/// handleJTSwitchCase - Emit jumptable for current switch case range 1471bool SelectionDAGLowering::handleJTSwitchCase(CaseRec& CR, 1472 CaseRecVector& WorkList, 1473 Value* SV, 1474 MachineBasicBlock* Default) { 1475 Case& FrontCase = *CR.Range.first; 1476 Case& BackCase = *(CR.Range.second-1); 1477 1478 int64_t First = cast<ConstantInt>(FrontCase.Low)->getSExtValue(); 1479 int64_t Last = cast<ConstantInt>(BackCase.High)->getSExtValue(); 1480 1481 uint64_t TSize = 0; 1482 for (CaseItr I = CR.Range.first, E = CR.Range.second; 1483 I!=E; ++I) 1484 TSize += I->size(); 1485 1486 if (!areJTsAllowed(TLI) || TSize <= 3) 1487 return false; 1488 1489 double Density = (double)TSize / (double)((Last - First) + 1ULL); 1490 if (Density < 0.4) 1491 return false; 1492 1493 DOUT << "Lowering jump table\n" 1494 << "First entry: " << First << ". Last entry: " << Last << "\n" 1495 << "Size: " << TSize << ". Density: " << Density << "\n\n"; 1496 1497 // Get the MachineFunction which holds the current MBB. This is used when 1498 // inserting any additional MBBs necessary to represent the switch. 1499 MachineFunction *CurMF = CurMBB->getParent(); 1500 1501 // Figure out which block is immediately after the current one. 1502 MachineBasicBlock *NextBlock = 0; 1503 MachineFunction::iterator BBI = CR.CaseBB; 1504 1505 if (++BBI != CurMBB->getParent()->end()) 1506 NextBlock = BBI; 1507 1508 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 1509 1510 // Create a new basic block to hold the code for loading the address 1511 // of the jump table, and jumping to it. Update successor information; 1512 // we will either branch to the default case for the switch, or the jump 1513 // table. 1514 MachineBasicBlock *JumpTableBB = new MachineBasicBlock(LLVMBB); 1515 CurMF->getBasicBlockList().insert(BBI, JumpTableBB); 1516 CR.CaseBB->addSuccessor(Default); 1517 CR.CaseBB->addSuccessor(JumpTableBB); 1518 1519 // Build a vector of destination BBs, corresponding to each target 1520 // of the jump table. If the value of the jump table slot corresponds to 1521 // a case statement, push the case's BB onto the vector, otherwise, push 1522 // the default BB. 1523 std::vector<MachineBasicBlock*> DestBBs; 1524 int64_t TEI = First; 1525 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) { 1526 int64_t Low = cast<ConstantInt>(I->Low)->getSExtValue(); 1527 int64_t High = cast<ConstantInt>(I->High)->getSExtValue(); 1528 1529 if ((Low <= TEI) && (TEI <= High)) { 1530 DestBBs.push_back(I->BB); 1531 if (TEI==High) 1532 ++I; 1533 } else { 1534 DestBBs.push_back(Default); 1535 } 1536 } 1537 1538 // Update successor info. Add one edge to each unique successor. 1539 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs()); 1540 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(), 1541 E = DestBBs.end(); I != E; ++I) { 1542 if (!SuccsHandled[(*I)->getNumber()]) { 1543 SuccsHandled[(*I)->getNumber()] = true; 1544 JumpTableBB->addSuccessor(*I); 1545 } 1546 } 1547 1548 // Create a jump table index for this jump table, or return an existing 1549 // one. 1550 unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs); 1551 1552 // Set the jump table information so that we can codegen it as a second 1553 // MachineBasicBlock 1554 SelectionDAGISel::JumpTable JT(-1U, JTI, JumpTableBB, Default); 1555 SelectionDAGISel::JumpTableHeader JTH(First, Last, SV, CR.CaseBB, 1556 (CR.CaseBB == CurMBB)); 1557 if (CR.CaseBB == CurMBB) 1558 visitJumpTableHeader(JT, JTH); 1559 1560 JTCases.push_back(SelectionDAGISel::JumpTableBlock(JTH, JT)); 1561 1562 return true; 1563} 1564 1565/// handleBTSplitSwitchCase - emit comparison and split binary search tree into 1566/// 2 subtrees. 1567bool SelectionDAGLowering::handleBTSplitSwitchCase(CaseRec& CR, 1568 CaseRecVector& WorkList, 1569 Value* SV, 1570 MachineBasicBlock* Default) { 1571 // Get the MachineFunction which holds the current MBB. This is used when 1572 // inserting any additional MBBs necessary to represent the switch. 1573 MachineFunction *CurMF = CurMBB->getParent(); 1574 1575 // Figure out which block is immediately after the current one. 1576 MachineBasicBlock *NextBlock = 0; 1577 MachineFunction::iterator BBI = CR.CaseBB; 1578 1579 if (++BBI != CurMBB->getParent()->end()) 1580 NextBlock = BBI; 1581 1582 Case& FrontCase = *CR.Range.first; 1583 Case& BackCase = *(CR.Range.second-1); 1584 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 1585 1586 // Size is the number of Cases represented by this range. 1587 unsigned Size = CR.Range.second - CR.Range.first; 1588 1589 int64_t First = cast<ConstantInt>(FrontCase.Low)->getSExtValue(); 1590 int64_t Last = cast<ConstantInt>(BackCase.High)->getSExtValue(); 1591 double FMetric = 0; 1592 CaseItr Pivot = CR.Range.first + Size/2; 1593 1594 // Select optimal pivot, maximizing sum density of LHS and RHS. This will 1595 // (heuristically) allow us to emit JumpTable's later. 1596 uint64_t TSize = 0; 1597 for (CaseItr I = CR.Range.first, E = CR.Range.second; 1598 I!=E; ++I) 1599 TSize += I->size(); 1600 1601 uint64_t LSize = FrontCase.size(); 1602 uint64_t RSize = TSize-LSize; 1603 DOUT << "Selecting best pivot: \n" 1604 << "First: " << First << ", Last: " << Last <<"\n" 1605 << "LSize: " << LSize << ", RSize: " << RSize << "\n"; 1606 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second; 1607 J!=E; ++I, ++J) { 1608 int64_t LEnd = cast<ConstantInt>(I->High)->getSExtValue(); 1609 int64_t RBegin = cast<ConstantInt>(J->Low)->getSExtValue(); 1610 assert((RBegin-LEnd>=1) && "Invalid case distance"); 1611 double LDensity = (double)LSize / (double)((LEnd - First) + 1ULL); 1612 double RDensity = (double)RSize / (double)((Last - RBegin) + 1ULL); 1613 double Metric = Log2_64(RBegin-LEnd)*(LDensity+RDensity); 1614 // Should always split in some non-trivial place 1615 DOUT <<"=>Step\n" 1616 << "LEnd: " << LEnd << ", RBegin: " << RBegin << "\n" 1617 << "LDensity: " << LDensity << ", RDensity: " << RDensity << "\n" 1618 << "Metric: " << Metric << "\n"; 1619 if (FMetric < Metric) { 1620 Pivot = J; 1621 FMetric = Metric; 1622 DOUT << "Current metric set to: " << FMetric << "\n"; 1623 } 1624 1625 LSize += J->size(); 1626 RSize -= J->size(); 1627 } 1628 if (areJTsAllowed(TLI)) { 1629 // If our case is dense we *really* should handle it earlier! 1630 assert((FMetric > 0) && "Should handle dense range earlier!"); 1631 } else { 1632 Pivot = CR.Range.first + Size/2; 1633 } 1634 1635 CaseRange LHSR(CR.Range.first, Pivot); 1636 CaseRange RHSR(Pivot, CR.Range.second); 1637 Constant *C = Pivot->Low; 1638 MachineBasicBlock *FalseBB = 0, *TrueBB = 0; 1639 1640 // We know that we branch to the LHS if the Value being switched on is 1641 // less than the Pivot value, C. We use this to optimize our binary 1642 // tree a bit, by recognizing that if SV is greater than or equal to the 1643 // LHS's Case Value, and that Case Value is exactly one less than the 1644 // Pivot's Value, then we can branch directly to the LHS's Target, 1645 // rather than creating a leaf node for it. 1646 if ((LHSR.second - LHSR.first) == 1 && 1647 LHSR.first->High == CR.GE && 1648 cast<ConstantInt>(C)->getSExtValue() == 1649 (cast<ConstantInt>(CR.GE)->getSExtValue() + 1LL)) { 1650 TrueBB = LHSR.first->BB; 1651 } else { 1652 TrueBB = new MachineBasicBlock(LLVMBB); 1653 CurMF->getBasicBlockList().insert(BBI, TrueBB); 1654 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR)); 1655 } 1656 1657 // Similar to the optimization above, if the Value being switched on is 1658 // known to be less than the Constant CR.LT, and the current Case Value 1659 // is CR.LT - 1, then we can branch directly to the target block for 1660 // the current Case Value, rather than emitting a RHS leaf node for it. 1661 if ((RHSR.second - RHSR.first) == 1 && CR.LT && 1662 cast<ConstantInt>(RHSR.first->Low)->getSExtValue() == 1663 (cast<ConstantInt>(CR.LT)->getSExtValue() - 1LL)) { 1664 FalseBB = RHSR.first->BB; 1665 } else { 1666 FalseBB = new MachineBasicBlock(LLVMBB); 1667 CurMF->getBasicBlockList().insert(BBI, FalseBB); 1668 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR)); 1669 } 1670 1671 // Create a CaseBlock record representing a conditional branch to 1672 // the LHS node if the value being switched on SV is less than C. 1673 // Otherwise, branch to LHS. 1674 SelectionDAGISel::CaseBlock CB(ISD::SETLT, SV, C, NULL, 1675 TrueBB, FalseBB, CR.CaseBB); 1676 1677 if (CR.CaseBB == CurMBB) 1678 visitSwitchCase(CB); 1679 else 1680 SwitchCases.push_back(CB); 1681 1682 return true; 1683} 1684 1685/// handleBitTestsSwitchCase - if current case range has few destination and 1686/// range span less, than machine word bitwidth, encode case range into series 1687/// of masks and emit bit tests with these masks. 1688bool SelectionDAGLowering::handleBitTestsSwitchCase(CaseRec& CR, 1689 CaseRecVector& WorkList, 1690 Value* SV, 1691 MachineBasicBlock* Default){ 1692 unsigned IntPtrBits = getSizeInBits(TLI.getPointerTy()); 1693 1694 Case& FrontCase = *CR.Range.first; 1695 Case& BackCase = *(CR.Range.second-1); 1696 1697 // Get the MachineFunction which holds the current MBB. This is used when 1698 // inserting any additional MBBs necessary to represent the switch. 1699 MachineFunction *CurMF = CurMBB->getParent(); 1700 1701 unsigned numCmps = 0; 1702 for (CaseItr I = CR.Range.first, E = CR.Range.second; 1703 I!=E; ++I) { 1704 // Single case counts one, case range - two. 1705 if (I->Low == I->High) 1706 numCmps +=1; 1707 else 1708 numCmps +=2; 1709 } 1710 1711 // Count unique destinations 1712 SmallSet<MachineBasicBlock*, 4> Dests; 1713 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { 1714 Dests.insert(I->BB); 1715 if (Dests.size() > 3) 1716 // Don't bother the code below, if there are too much unique destinations 1717 return false; 1718 } 1719 DOUT << "Total number of unique destinations: " << Dests.size() << "\n" 1720 << "Total number of comparisons: " << numCmps << "\n"; 1721 1722 // Compute span of values. 1723 Constant* minValue = FrontCase.Low; 1724 Constant* maxValue = BackCase.High; 1725 uint64_t range = cast<ConstantInt>(maxValue)->getSExtValue() - 1726 cast<ConstantInt>(minValue)->getSExtValue(); 1727 DOUT << "Compare range: " << range << "\n" 1728 << "Low bound: " << cast<ConstantInt>(minValue)->getSExtValue() << "\n" 1729 << "High bound: " << cast<ConstantInt>(maxValue)->getSExtValue() << "\n"; 1730 1731 if (range>=IntPtrBits || 1732 (!(Dests.size() == 1 && numCmps >= 3) && 1733 !(Dests.size() == 2 && numCmps >= 5) && 1734 !(Dests.size() >= 3 && numCmps >= 6))) 1735 return false; 1736 1737 DOUT << "Emitting bit tests\n"; 1738 int64_t lowBound = 0; 1739 1740 // Optimize the case where all the case values fit in a 1741 // word without having to subtract minValue. In this case, 1742 // we can optimize away the subtraction. 1743 if (cast<ConstantInt>(minValue)->getSExtValue() >= 0 && 1744 cast<ConstantInt>(maxValue)->getSExtValue() < IntPtrBits) { 1745 range = cast<ConstantInt>(maxValue)->getSExtValue(); 1746 } else { 1747 lowBound = cast<ConstantInt>(minValue)->getSExtValue(); 1748 } 1749 1750 CaseBitsVector CasesBits; 1751 unsigned i, count = 0; 1752 1753 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) { 1754 MachineBasicBlock* Dest = I->BB; 1755 for (i = 0; i < count; ++i) 1756 if (Dest == CasesBits[i].BB) 1757 break; 1758 1759 if (i == count) { 1760 assert((count < 3) && "Too much destinations to test!"); 1761 CasesBits.push_back(CaseBits(0, Dest, 0)); 1762 count++; 1763 } 1764 1765 uint64_t lo = cast<ConstantInt>(I->Low)->getSExtValue() - lowBound; 1766 uint64_t hi = cast<ConstantInt>(I->High)->getSExtValue() - lowBound; 1767 1768 for (uint64_t j = lo; j <= hi; j++) { 1769 CasesBits[i].Mask |= 1ULL << j; 1770 CasesBits[i].Bits++; 1771 } 1772 1773 } 1774 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp()); 1775 1776 SelectionDAGISel::BitTestInfo BTC; 1777 1778 // Figure out which block is immediately after the current one. 1779 MachineFunction::iterator BBI = CR.CaseBB; 1780 ++BBI; 1781 1782 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock(); 1783 1784 DOUT << "Cases:\n"; 1785 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) { 1786 DOUT << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits 1787 << ", BB: " << CasesBits[i].BB << "\n"; 1788 1789 MachineBasicBlock *CaseBB = new MachineBasicBlock(LLVMBB); 1790 CurMF->getBasicBlockList().insert(BBI, CaseBB); 1791 BTC.push_back(SelectionDAGISel::BitTestCase(CasesBits[i].Mask, 1792 CaseBB, 1793 CasesBits[i].BB)); 1794 } 1795 1796 SelectionDAGISel::BitTestBlock BTB(lowBound, range, SV, 1797 -1U, (CR.CaseBB == CurMBB), 1798 CR.CaseBB, Default, BTC); 1799 1800 if (CR.CaseBB == CurMBB) 1801 visitBitTestHeader(BTB); 1802 1803 BitTestCases.push_back(BTB); 1804 1805 return true; 1806} 1807 1808 1809// Clusterify - Transform simple list of Cases into list of CaseRange's 1810unsigned SelectionDAGLowering::Clusterify(CaseVector& Cases, 1811 const SwitchInst& SI) { 1812 unsigned numCmps = 0; 1813 1814 // Start with "simple" cases 1815 for (unsigned i = 1; i < SI.getNumSuccessors(); ++i) { 1816 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)]; 1817 Cases.push_back(Case(SI.getSuccessorValue(i), 1818 SI.getSuccessorValue(i), 1819 SMBB)); 1820 } 1821 sort(Cases.begin(), Cases.end(), CaseCmp()); 1822 1823 // Merge case into clusters 1824 if (Cases.size()>=2) 1825 for (CaseItr I=Cases.begin(), J=++(Cases.begin()), E=Cases.end(); J!=E; ) { 1826 int64_t nextValue = cast<ConstantInt>(J->Low)->getSExtValue(); 1827 int64_t currentValue = cast<ConstantInt>(I->High)->getSExtValue(); 1828 MachineBasicBlock* nextBB = J->BB; 1829 MachineBasicBlock* currentBB = I->BB; 1830 1831 // If the two neighboring cases go to the same destination, merge them 1832 // into a single case. 1833 if ((nextValue-currentValue==1) && (currentBB == nextBB)) { 1834 I->High = J->High; 1835 J = Cases.erase(J); 1836 } else { 1837 I = J++; 1838 } 1839 } 1840 1841 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) { 1842 if (I->Low != I->High) 1843 // A range counts double, since it requires two compares. 1844 ++numCmps; 1845 } 1846 1847 return numCmps; 1848} 1849 1850void SelectionDAGLowering::visitSwitch(SwitchInst &SI) { 1851 // Figure out which block is immediately after the current one. 1852 MachineBasicBlock *NextBlock = 0; 1853 MachineFunction::iterator BBI = CurMBB; 1854 1855 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()]; 1856 1857 // If there is only the default destination, branch to it if it is not the 1858 // next basic block. Otherwise, just fall through. 1859 if (SI.getNumOperands() == 2) { 1860 // Update machine-CFG edges. 1861 1862 // If this is not a fall-through branch, emit the branch. 1863 if (Default != NextBlock) 1864 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(), 1865 DAG.getBasicBlock(Default))); 1866 1867 CurMBB->addSuccessor(Default); 1868 return; 1869 } 1870 1871 // If there are any non-default case statements, create a vector of Cases 1872 // representing each one, and sort the vector so that we can efficiently 1873 // create a binary search tree from them. 1874 CaseVector Cases; 1875 unsigned numCmps = Clusterify(Cases, SI); 1876 DOUT << "Clusterify finished. Total clusters: " << Cases.size() 1877 << ". Total compares: " << numCmps << "\n"; 1878 1879 // Get the Value to be switched on and default basic blocks, which will be 1880 // inserted into CaseBlock records, representing basic blocks in the binary 1881 // search tree. 1882 Value *SV = SI.getOperand(0); 1883 1884 // Push the initial CaseRec onto the worklist 1885 CaseRecVector WorkList; 1886 WorkList.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end()))); 1887 1888 while (!WorkList.empty()) { 1889 // Grab a record representing a case range to process off the worklist 1890 CaseRec CR = WorkList.back(); 1891 WorkList.pop_back(); 1892 1893 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default)) 1894 continue; 1895 1896 // If the range has few cases (two or less) emit a series of specific 1897 // tests. 1898 if (handleSmallSwitchRange(CR, WorkList, SV, Default)) 1899 continue; 1900 1901 // If the switch has more than 5 blocks, and at least 40% dense, and the 1902 // target supports indirect branches, then emit a jump table rather than 1903 // lowering the switch to a binary tree of conditional branches. 1904 if (handleJTSwitchCase(CR, WorkList, SV, Default)) 1905 continue; 1906 1907 // Emit binary tree. We need to pick a pivot, and push left and right ranges 1908 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call. 1909 handleBTSplitSwitchCase(CR, WorkList, SV, Default); 1910 } 1911} 1912 1913 1914void SelectionDAGLowering::visitSub(User &I) { 1915 // -0.0 - X --> fneg 1916 const Type *Ty = I.getType(); 1917 if (isa<VectorType>(Ty)) { 1918 visitVectorBinary(I, ISD::VSUB); 1919 } else if (Ty->isFloatingPoint()) { 1920 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0))) 1921 if (CFP->isExactlyValue(-0.0)) { 1922 SDOperand Op2 = getValue(I.getOperand(1)); 1923 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2)); 1924 return; 1925 } 1926 visitScalarBinary(I, ISD::FSUB); 1927 } else 1928 visitScalarBinary(I, ISD::SUB); 1929} 1930 1931void SelectionDAGLowering::visitScalarBinary(User &I, unsigned OpCode) { 1932 SDOperand Op1 = getValue(I.getOperand(0)); 1933 SDOperand Op2 = getValue(I.getOperand(1)); 1934 1935 setValue(&I, DAG.getNode(OpCode, Op1.getValueType(), Op1, Op2)); 1936} 1937 1938void 1939SelectionDAGLowering::visitVectorBinary(User &I, unsigned OpCode) { 1940 assert(isa<VectorType>(I.getType())); 1941 const VectorType *Ty = cast<VectorType>(I.getType()); 1942 SDOperand Typ = DAG.getValueType(TLI.getValueType(Ty->getElementType())); 1943 1944 setValue(&I, DAG.getNode(OpCode, MVT::Vector, 1945 getValue(I.getOperand(0)), 1946 getValue(I.getOperand(1)), 1947 DAG.getConstant(Ty->getNumElements(), MVT::i32), 1948 Typ)); 1949} 1950 1951void SelectionDAGLowering::visitEitherBinary(User &I, unsigned ScalarOp, 1952 unsigned VectorOp) { 1953 if (isa<VectorType>(I.getType())) 1954 visitVectorBinary(I, VectorOp); 1955 else 1956 visitScalarBinary(I, ScalarOp); 1957} 1958 1959void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) { 1960 SDOperand Op1 = getValue(I.getOperand(0)); 1961 SDOperand Op2 = getValue(I.getOperand(1)); 1962 1963 if (TLI.getShiftAmountTy() < Op2.getValueType()) 1964 Op2 = DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), Op2); 1965 else if (TLI.getShiftAmountTy() > Op2.getValueType()) 1966 Op2 = DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), Op2); 1967 1968 setValue(&I, DAG.getNode(Opcode, Op1.getValueType(), Op1, Op2)); 1969} 1970 1971void SelectionDAGLowering::visitICmp(User &I) { 1972 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE; 1973 if (ICmpInst *IC = dyn_cast<ICmpInst>(&I)) 1974 predicate = IC->getPredicate(); 1975 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I)) 1976 predicate = ICmpInst::Predicate(IC->getPredicate()); 1977 SDOperand Op1 = getValue(I.getOperand(0)); 1978 SDOperand Op2 = getValue(I.getOperand(1)); 1979 ISD::CondCode Opcode; 1980 switch (predicate) { 1981 case ICmpInst::ICMP_EQ : Opcode = ISD::SETEQ; break; 1982 case ICmpInst::ICMP_NE : Opcode = ISD::SETNE; break; 1983 case ICmpInst::ICMP_UGT : Opcode = ISD::SETUGT; break; 1984 case ICmpInst::ICMP_UGE : Opcode = ISD::SETUGE; break; 1985 case ICmpInst::ICMP_ULT : Opcode = ISD::SETULT; break; 1986 case ICmpInst::ICMP_ULE : Opcode = ISD::SETULE; break; 1987 case ICmpInst::ICMP_SGT : Opcode = ISD::SETGT; break; 1988 case ICmpInst::ICMP_SGE : Opcode = ISD::SETGE; break; 1989 case ICmpInst::ICMP_SLT : Opcode = ISD::SETLT; break; 1990 case ICmpInst::ICMP_SLE : Opcode = ISD::SETLE; break; 1991 default: 1992 assert(!"Invalid ICmp predicate value"); 1993 Opcode = ISD::SETEQ; 1994 break; 1995 } 1996 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Opcode)); 1997} 1998 1999void SelectionDAGLowering::visitFCmp(User &I) { 2000 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE; 2001 if (FCmpInst *FC = dyn_cast<FCmpInst>(&I)) 2002 predicate = FC->getPredicate(); 2003 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I)) 2004 predicate = FCmpInst::Predicate(FC->getPredicate()); 2005 SDOperand Op1 = getValue(I.getOperand(0)); 2006 SDOperand Op2 = getValue(I.getOperand(1)); 2007 ISD::CondCode Condition, FOC, FPC; 2008 switch (predicate) { 2009 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break; 2010 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break; 2011 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break; 2012 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break; 2013 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break; 2014 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break; 2015 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break; 2016 case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break; 2017 case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break; 2018 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break; 2019 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break; 2020 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break; 2021 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break; 2022 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break; 2023 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break; 2024 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break; 2025 default: 2026 assert(!"Invalid FCmp predicate value"); 2027 FOC = FPC = ISD::SETFALSE; 2028 break; 2029 } 2030 if (FiniteOnlyFPMath()) 2031 Condition = FOC; 2032 else 2033 Condition = FPC; 2034 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Condition)); 2035} 2036 2037void SelectionDAGLowering::visitSelect(User &I) { 2038 SDOperand Cond = getValue(I.getOperand(0)); 2039 SDOperand TrueVal = getValue(I.getOperand(1)); 2040 SDOperand FalseVal = getValue(I.getOperand(2)); 2041 if (!isa<VectorType>(I.getType())) { 2042 setValue(&I, DAG.getNode(ISD::SELECT, TrueVal.getValueType(), Cond, 2043 TrueVal, FalseVal)); 2044 } else { 2045 setValue(&I, DAG.getNode(ISD::VSELECT, MVT::Vector, Cond, TrueVal, FalseVal, 2046 *(TrueVal.Val->op_end()-2), 2047 *(TrueVal.Val->op_end()-1))); 2048 } 2049} 2050 2051 2052void SelectionDAGLowering::visitTrunc(User &I) { 2053 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest). 2054 SDOperand N = getValue(I.getOperand(0)); 2055 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2056 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N)); 2057} 2058 2059void SelectionDAGLowering::visitZExt(User &I) { 2060 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 2061 // ZExt also can't be a cast to bool for same reason. So, nothing much to do 2062 SDOperand N = getValue(I.getOperand(0)); 2063 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2064 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N)); 2065} 2066 2067void SelectionDAGLowering::visitSExt(User &I) { 2068 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 2069 // SExt also can't be a cast to bool for same reason. So, nothing much to do 2070 SDOperand N = getValue(I.getOperand(0)); 2071 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2072 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, DestVT, N)); 2073} 2074 2075void SelectionDAGLowering::visitFPTrunc(User &I) { 2076 // FPTrunc is never a no-op cast, no need to check 2077 SDOperand N = getValue(I.getOperand(0)); 2078 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2079 setValue(&I, DAG.getNode(ISD::FP_ROUND, DestVT, N)); 2080} 2081 2082void SelectionDAGLowering::visitFPExt(User &I){ 2083 // FPTrunc is never a no-op cast, no need to check 2084 SDOperand N = getValue(I.getOperand(0)); 2085 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2086 setValue(&I, DAG.getNode(ISD::FP_EXTEND, DestVT, N)); 2087} 2088 2089void SelectionDAGLowering::visitFPToUI(User &I) { 2090 // FPToUI is never a no-op cast, no need to check 2091 SDOperand N = getValue(I.getOperand(0)); 2092 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2093 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, DestVT, N)); 2094} 2095 2096void SelectionDAGLowering::visitFPToSI(User &I) { 2097 // FPToSI is never a no-op cast, no need to check 2098 SDOperand N = getValue(I.getOperand(0)); 2099 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2100 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, DestVT, N)); 2101} 2102 2103void SelectionDAGLowering::visitUIToFP(User &I) { 2104 // UIToFP is never a no-op cast, no need to check 2105 SDOperand N = getValue(I.getOperand(0)); 2106 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2107 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, DestVT, N)); 2108} 2109 2110void SelectionDAGLowering::visitSIToFP(User &I){ 2111 // UIToFP is never a no-op cast, no need to check 2112 SDOperand N = getValue(I.getOperand(0)); 2113 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2114 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, DestVT, N)); 2115} 2116 2117void SelectionDAGLowering::visitPtrToInt(User &I) { 2118 // What to do depends on the size of the integer and the size of the pointer. 2119 // We can either truncate, zero extend, or no-op, accordingly. 2120 SDOperand N = getValue(I.getOperand(0)); 2121 MVT::ValueType SrcVT = N.getValueType(); 2122 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2123 SDOperand Result; 2124 if (MVT::getSizeInBits(DestVT) < MVT::getSizeInBits(SrcVT)) 2125 Result = DAG.getNode(ISD::TRUNCATE, DestVT, N); 2126 else 2127 // Note: ZERO_EXTEND can handle cases where the sizes are equal too 2128 Result = DAG.getNode(ISD::ZERO_EXTEND, DestVT, N); 2129 setValue(&I, Result); 2130} 2131 2132void SelectionDAGLowering::visitIntToPtr(User &I) { 2133 // What to do depends on the size of the integer and the size of the pointer. 2134 // We can either truncate, zero extend, or no-op, accordingly. 2135 SDOperand N = getValue(I.getOperand(0)); 2136 MVT::ValueType SrcVT = N.getValueType(); 2137 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2138 if (MVT::getSizeInBits(DestVT) < MVT::getSizeInBits(SrcVT)) 2139 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N)); 2140 else 2141 // Note: ZERO_EXTEND can handle cases where the sizes are equal too 2142 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N)); 2143} 2144 2145void SelectionDAGLowering::visitBitCast(User &I) { 2146 SDOperand N = getValue(I.getOperand(0)); 2147 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 2148 if (DestVT == MVT::Vector) { 2149 // This is a cast to a vector from something else. 2150 // Get information about the output vector. 2151 const VectorType *DestTy = cast<VectorType>(I.getType()); 2152 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType()); 2153 setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N, 2154 DAG.getConstant(DestTy->getNumElements(),MVT::i32), 2155 DAG.getValueType(EltVT))); 2156 return; 2157 } 2158 MVT::ValueType SrcVT = N.getValueType(); 2159 if (SrcVT == MVT::Vector) { 2160 // This is a cast from a vctor to something else. 2161 // Get information about the input vector. 2162 setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N)); 2163 return; 2164 } 2165 2166 // BitCast assures us that source and destination are the same size so this 2167 // is either a BIT_CONVERT or a no-op. 2168 if (DestVT != N.getValueType()) 2169 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, DestVT, N)); // convert types 2170 else 2171 setValue(&I, N); // noop cast. 2172} 2173 2174void SelectionDAGLowering::visitInsertElement(User &I) { 2175 SDOperand InVec = getValue(I.getOperand(0)); 2176 SDOperand InVal = getValue(I.getOperand(1)); 2177 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), 2178 getValue(I.getOperand(2))); 2179 2180 SDOperand Num = *(InVec.Val->op_end()-2); 2181 SDOperand Typ = *(InVec.Val->op_end()-1); 2182 setValue(&I, DAG.getNode(ISD::VINSERT_VECTOR_ELT, MVT::Vector, 2183 InVec, InVal, InIdx, Num, Typ)); 2184} 2185 2186void SelectionDAGLowering::visitExtractElement(User &I) { 2187 SDOperand InVec = getValue(I.getOperand(0)); 2188 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), 2189 getValue(I.getOperand(1))); 2190 SDOperand Typ = *(InVec.Val->op_end()-1); 2191 setValue(&I, DAG.getNode(ISD::VEXTRACT_VECTOR_ELT, 2192 TLI.getValueType(I.getType()), InVec, InIdx)); 2193} 2194 2195void SelectionDAGLowering::visitShuffleVector(User &I) { 2196 SDOperand V1 = getValue(I.getOperand(0)); 2197 SDOperand V2 = getValue(I.getOperand(1)); 2198 SDOperand Mask = getValue(I.getOperand(2)); 2199 2200 SDOperand Num = *(V1.Val->op_end()-2); 2201 SDOperand Typ = *(V2.Val->op_end()-1); 2202 setValue(&I, DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, 2203 V1, V2, Mask, Num, Typ)); 2204} 2205 2206 2207void SelectionDAGLowering::visitGetElementPtr(User &I) { 2208 SDOperand N = getValue(I.getOperand(0)); 2209 const Type *Ty = I.getOperand(0)->getType(); 2210 2211 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end(); 2212 OI != E; ++OI) { 2213 Value *Idx = *OI; 2214 if (const StructType *StTy = dyn_cast<StructType>(Ty)) { 2215 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 2216 if (Field) { 2217 // N = N + Offset 2218 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field); 2219 N = DAG.getNode(ISD::ADD, N.getValueType(), N, 2220 getIntPtrConstant(Offset)); 2221 } 2222 Ty = StTy->getElementType(Field); 2223 } else { 2224 Ty = cast<SequentialType>(Ty)->getElementType(); 2225 2226 // If this is a constant subscript, handle it quickly. 2227 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { 2228 if (CI->getZExtValue() == 0) continue; 2229 uint64_t Offs = 2230 TD->getTypeSize(Ty)*cast<ConstantInt>(CI)->getSExtValue(); 2231 N = DAG.getNode(ISD::ADD, N.getValueType(), N, getIntPtrConstant(Offs)); 2232 continue; 2233 } 2234 2235 // N = N + Idx * ElementSize; 2236 uint64_t ElementSize = TD->getTypeSize(Ty); 2237 SDOperand IdxN = getValue(Idx); 2238 2239 // If the index is smaller or larger than intptr_t, truncate or extend 2240 // it. 2241 if (IdxN.getValueType() < N.getValueType()) { 2242 IdxN = DAG.getNode(ISD::SIGN_EXTEND, N.getValueType(), IdxN); 2243 } else if (IdxN.getValueType() > N.getValueType()) 2244 IdxN = DAG.getNode(ISD::TRUNCATE, N.getValueType(), IdxN); 2245 2246 // If this is a multiply by a power of two, turn it into a shl 2247 // immediately. This is a very common case. 2248 if (isPowerOf2_64(ElementSize)) { 2249 unsigned Amt = Log2_64(ElementSize); 2250 IdxN = DAG.getNode(ISD::SHL, N.getValueType(), IdxN, 2251 DAG.getConstant(Amt, TLI.getShiftAmountTy())); 2252 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN); 2253 continue; 2254 } 2255 2256 SDOperand Scale = getIntPtrConstant(ElementSize); 2257 IdxN = DAG.getNode(ISD::MUL, N.getValueType(), IdxN, Scale); 2258 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN); 2259 } 2260 } 2261 setValue(&I, N); 2262} 2263 2264void SelectionDAGLowering::visitAlloca(AllocaInst &I) { 2265 // If this is a fixed sized alloca in the entry block of the function, 2266 // allocate it statically on the stack. 2267 if (FuncInfo.StaticAllocaMap.count(&I)) 2268 return; // getValue will auto-populate this. 2269 2270 const Type *Ty = I.getAllocatedType(); 2271 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty); 2272 unsigned Align = 2273 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty), 2274 I.getAlignment()); 2275 2276 SDOperand AllocSize = getValue(I.getArraySize()); 2277 MVT::ValueType IntPtr = TLI.getPointerTy(); 2278 if (IntPtr < AllocSize.getValueType()) 2279 AllocSize = DAG.getNode(ISD::TRUNCATE, IntPtr, AllocSize); 2280 else if (IntPtr > AllocSize.getValueType()) 2281 AllocSize = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, AllocSize); 2282 2283 AllocSize = DAG.getNode(ISD::MUL, IntPtr, AllocSize, 2284 getIntPtrConstant(TySize)); 2285 2286 // Handle alignment. If the requested alignment is less than or equal to the 2287 // stack alignment, ignore it and round the size of the allocation up to the 2288 // stack alignment size. If the size is greater than the stack alignment, we 2289 // note this in the DYNAMIC_STACKALLOC node. 2290 unsigned StackAlign = 2291 TLI.getTargetMachine().getFrameInfo()->getStackAlignment(); 2292 if (Align <= StackAlign) { 2293 Align = 0; 2294 // Add SA-1 to the size. 2295 AllocSize = DAG.getNode(ISD::ADD, AllocSize.getValueType(), AllocSize, 2296 getIntPtrConstant(StackAlign-1)); 2297 // Mask out the low bits for alignment purposes. 2298 AllocSize = DAG.getNode(ISD::AND, AllocSize.getValueType(), AllocSize, 2299 getIntPtrConstant(~(uint64_t)(StackAlign-1))); 2300 } 2301 2302 SDOperand Ops[] = { getRoot(), AllocSize, getIntPtrConstant(Align) }; 2303 const MVT::ValueType *VTs = DAG.getNodeValueTypes(AllocSize.getValueType(), 2304 MVT::Other); 2305 SDOperand DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, VTs, 2, Ops, 3); 2306 setValue(&I, DSA); 2307 DAG.setRoot(DSA.getValue(1)); 2308 2309 // Inform the Frame Information that we have just allocated a variable-sized 2310 // object. 2311 CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject(); 2312} 2313 2314void SelectionDAGLowering::visitLoad(LoadInst &I) { 2315 SDOperand Ptr = getValue(I.getOperand(0)); 2316 2317 SDOperand Root; 2318 if (I.isVolatile()) 2319 Root = getRoot(); 2320 else { 2321 // Do not serialize non-volatile loads against each other. 2322 Root = DAG.getRoot(); 2323 } 2324 2325 setValue(&I, getLoadFrom(I.getType(), Ptr, I.getOperand(0), 2326 Root, I.isVolatile(), I.getAlignment())); 2327} 2328 2329SDOperand SelectionDAGLowering::getLoadFrom(const Type *Ty, SDOperand Ptr, 2330 const Value *SV, SDOperand Root, 2331 bool isVolatile, 2332 unsigned Alignment) { 2333 SDOperand L; 2334 if (const VectorType *PTy = dyn_cast<VectorType>(Ty)) { 2335 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType()); 2336 L = DAG.getVecLoad(PTy->getNumElements(), PVT, Root, Ptr, 2337 DAG.getSrcValue(SV)); 2338 } else { 2339 L = DAG.getLoad(TLI.getValueType(Ty), Root, Ptr, SV, 0, 2340 isVolatile, Alignment); 2341 } 2342 2343 if (isVolatile) 2344 DAG.setRoot(L.getValue(1)); 2345 else 2346 PendingLoads.push_back(L.getValue(1)); 2347 2348 return L; 2349} 2350 2351 2352void SelectionDAGLowering::visitStore(StoreInst &I) { 2353 Value *SrcV = I.getOperand(0); 2354 SDOperand Src = getValue(SrcV); 2355 SDOperand Ptr = getValue(I.getOperand(1)); 2356 DAG.setRoot(DAG.getStore(getRoot(), Src, Ptr, I.getOperand(1), 0, 2357 I.isVolatile(), I.getAlignment())); 2358} 2359 2360/// IntrinsicCannotAccessMemory - Return true if the specified intrinsic cannot 2361/// access memory and has no other side effects at all. 2362static bool IntrinsicCannotAccessMemory(unsigned IntrinsicID) { 2363#define GET_NO_MEMORY_INTRINSICS 2364#include "llvm/Intrinsics.gen" 2365#undef GET_NO_MEMORY_INTRINSICS 2366 return false; 2367} 2368 2369// IntrinsicOnlyReadsMemory - Return true if the specified intrinsic doesn't 2370// have any side-effects or if it only reads memory. 2371static bool IntrinsicOnlyReadsMemory(unsigned IntrinsicID) { 2372#define GET_SIDE_EFFECT_INFO 2373#include "llvm/Intrinsics.gen" 2374#undef GET_SIDE_EFFECT_INFO 2375 return false; 2376} 2377 2378/// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC 2379/// node. 2380void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I, 2381 unsigned Intrinsic) { 2382 bool HasChain = !IntrinsicCannotAccessMemory(Intrinsic); 2383 bool OnlyLoad = HasChain && IntrinsicOnlyReadsMemory(Intrinsic); 2384 2385 // Build the operand list. 2386 SmallVector<SDOperand, 8> Ops; 2387 if (HasChain) { // If this intrinsic has side-effects, chainify it. 2388 if (OnlyLoad) { 2389 // We don't need to serialize loads against other loads. 2390 Ops.push_back(DAG.getRoot()); 2391 } else { 2392 Ops.push_back(getRoot()); 2393 } 2394 } 2395 2396 // Add the intrinsic ID as an integer operand. 2397 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy())); 2398 2399 // Add all operands of the call to the operand list. 2400 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) { 2401 SDOperand Op = getValue(I.getOperand(i)); 2402 2403 // If this is a vector type, force it to the right vector type. 2404 if (Op.getValueType() == MVT::Vector) { 2405 const VectorType *OpTy = cast<VectorType>(I.getOperand(i)->getType()); 2406 MVT::ValueType EltVT = TLI.getValueType(OpTy->getElementType()); 2407 2408 MVT::ValueType VVT = MVT::getVectorType(EltVT, OpTy->getNumElements()); 2409 assert(VVT != MVT::Other && "Intrinsic uses a non-legal type?"); 2410 Op = DAG.getNode(ISD::VBIT_CONVERT, VVT, Op); 2411 } 2412 2413 assert(TLI.isTypeLegal(Op.getValueType()) && 2414 "Intrinsic uses a non-legal type?"); 2415 Ops.push_back(Op); 2416 } 2417 2418 std::vector<MVT::ValueType> VTs; 2419 if (I.getType() != Type::VoidTy) { 2420 MVT::ValueType VT = TLI.getValueType(I.getType()); 2421 if (VT == MVT::Vector) { 2422 const VectorType *DestTy = cast<VectorType>(I.getType()); 2423 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType()); 2424 2425 VT = MVT::getVectorType(EltVT, DestTy->getNumElements()); 2426 assert(VT != MVT::Other && "Intrinsic uses a non-legal type?"); 2427 } 2428 2429 assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?"); 2430 VTs.push_back(VT); 2431 } 2432 if (HasChain) 2433 VTs.push_back(MVT::Other); 2434 2435 const MVT::ValueType *VTList = DAG.getNodeValueTypes(VTs); 2436 2437 // Create the node. 2438 SDOperand Result; 2439 if (!HasChain) 2440 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VTList, VTs.size(), 2441 &Ops[0], Ops.size()); 2442 else if (I.getType() != Type::VoidTy) 2443 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, VTList, VTs.size(), 2444 &Ops[0], Ops.size()); 2445 else 2446 Result = DAG.getNode(ISD::INTRINSIC_VOID, VTList, VTs.size(), 2447 &Ops[0], Ops.size()); 2448 2449 if (HasChain) { 2450 SDOperand Chain = Result.getValue(Result.Val->getNumValues()-1); 2451 if (OnlyLoad) 2452 PendingLoads.push_back(Chain); 2453 else 2454 DAG.setRoot(Chain); 2455 } 2456 if (I.getType() != Type::VoidTy) { 2457 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) { 2458 MVT::ValueType EVT = TLI.getValueType(PTy->getElementType()); 2459 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result, 2460 DAG.getConstant(PTy->getNumElements(), MVT::i32), 2461 DAG.getValueType(EVT)); 2462 } 2463 setValue(&I, Result); 2464 } 2465} 2466 2467/// ExtractGlobalVariable - If C is a global variable, or a bitcast of one 2468/// (possibly constant folded), return it. Otherwise return NULL. 2469static GlobalVariable *ExtractGlobalVariable (Constant *C) { 2470 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(C)) 2471 return GV; 2472 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 2473 if (CE->getOpcode() == Instruction::BitCast) 2474 return dyn_cast<GlobalVariable>(CE->getOperand(0)); 2475 else if (CE->getOpcode() == Instruction::GetElementPtr) { 2476 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i) 2477 if (!CE->getOperand(i)->isNullValue()) 2478 return NULL; 2479 return dyn_cast<GlobalVariable>(CE->getOperand(0)); 2480 } 2481 } 2482 return NULL; 2483} 2484 2485/// visitIntrinsicCall - Lower the call to the specified intrinsic function. If 2486/// we want to emit this as a call to a named external function, return the name 2487/// otherwise lower it and return null. 2488const char * 2489SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) { 2490 switch (Intrinsic) { 2491 default: 2492 // By default, turn this into a target intrinsic node. 2493 visitTargetIntrinsic(I, Intrinsic); 2494 return 0; 2495 case Intrinsic::vastart: visitVAStart(I); return 0; 2496 case Intrinsic::vaend: visitVAEnd(I); return 0; 2497 case Intrinsic::vacopy: visitVACopy(I); return 0; 2498 case Intrinsic::returnaddress: 2499 setValue(&I, DAG.getNode(ISD::RETURNADDR, TLI.getPointerTy(), 2500 getValue(I.getOperand(1)))); 2501 return 0; 2502 case Intrinsic::frameaddress: 2503 setValue(&I, DAG.getNode(ISD::FRAMEADDR, TLI.getPointerTy(), 2504 getValue(I.getOperand(1)))); 2505 return 0; 2506 case Intrinsic::setjmp: 2507 return "_setjmp"+!TLI.usesUnderscoreSetJmp(); 2508 break; 2509 case Intrinsic::longjmp: 2510 return "_longjmp"+!TLI.usesUnderscoreLongJmp(); 2511 break; 2512 case Intrinsic::memcpy_i32: 2513 case Intrinsic::memcpy_i64: 2514 visitMemIntrinsic(I, ISD::MEMCPY); 2515 return 0; 2516 case Intrinsic::memset_i32: 2517 case Intrinsic::memset_i64: 2518 visitMemIntrinsic(I, ISD::MEMSET); 2519 return 0; 2520 case Intrinsic::memmove_i32: 2521 case Intrinsic::memmove_i64: 2522 visitMemIntrinsic(I, ISD::MEMMOVE); 2523 return 0; 2524 2525 case Intrinsic::dbg_stoppoint: { 2526 MachineModuleInfo *MMI = DAG.getMachineModuleInfo(); 2527 DbgStopPointInst &SPI = cast<DbgStopPointInst>(I); 2528 if (MMI && SPI.getContext() && MMI->Verify(SPI.getContext())) { 2529 SDOperand Ops[5]; 2530 2531 Ops[0] = getRoot(); 2532 Ops[1] = getValue(SPI.getLineValue()); 2533 Ops[2] = getValue(SPI.getColumnValue()); 2534 2535 DebugInfoDesc *DD = MMI->getDescFor(SPI.getContext()); 2536 assert(DD && "Not a debug information descriptor"); 2537 CompileUnitDesc *CompileUnit = cast<CompileUnitDesc>(DD); 2538 2539 Ops[3] = DAG.getString(CompileUnit->getFileName()); 2540 Ops[4] = DAG.getString(CompileUnit->getDirectory()); 2541 2542 DAG.setRoot(DAG.getNode(ISD::LOCATION, MVT::Other, Ops, 5)); 2543 } 2544 2545 return 0; 2546 } 2547 case Intrinsic::dbg_region_start: { 2548 MachineModuleInfo *MMI = DAG.getMachineModuleInfo(); 2549 DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I); 2550 if (MMI && RSI.getContext() && MMI->Verify(RSI.getContext())) { 2551 unsigned LabelID = MMI->RecordRegionStart(RSI.getContext()); 2552 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(), 2553 DAG.getConstant(LabelID, MVT::i32))); 2554 } 2555 2556 return 0; 2557 } 2558 case Intrinsic::dbg_region_end: { 2559 MachineModuleInfo *MMI = DAG.getMachineModuleInfo(); 2560 DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I); 2561 if (MMI && REI.getContext() && MMI->Verify(REI.getContext())) { 2562 unsigned LabelID = MMI->RecordRegionEnd(REI.getContext()); 2563 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, 2564 getRoot(), DAG.getConstant(LabelID, MVT::i32))); 2565 } 2566 2567 return 0; 2568 } 2569 case Intrinsic::dbg_func_start: { 2570 MachineModuleInfo *MMI = DAG.getMachineModuleInfo(); 2571 DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I); 2572 if (MMI && FSI.getSubprogram() && 2573 MMI->Verify(FSI.getSubprogram())) { 2574 unsigned LabelID = MMI->RecordRegionStart(FSI.getSubprogram()); 2575 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, 2576 getRoot(), DAG.getConstant(LabelID, MVT::i32))); 2577 } 2578 2579 return 0; 2580 } 2581 case Intrinsic::dbg_declare: { 2582 MachineModuleInfo *MMI = DAG.getMachineModuleInfo(); 2583 DbgDeclareInst &DI = cast<DbgDeclareInst>(I); 2584 if (MMI && DI.getVariable() && MMI->Verify(DI.getVariable())) { 2585 SDOperand AddressOp = getValue(DI.getAddress()); 2586 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(AddressOp)) 2587 MMI->RecordVariable(DI.getVariable(), FI->getIndex()); 2588 } 2589 2590 return 0; 2591 } 2592 2593 case Intrinsic::eh_exception: { 2594 MachineModuleInfo *MMI = DAG.getMachineModuleInfo(); 2595 2596 if (MMI) { 2597 // Add a label to mark the beginning of the landing pad. Deletion of the 2598 // landing pad can thus be detected via the MachineModuleInfo. 2599 unsigned LabelID = MMI->addLandingPad(CurMBB); 2600 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, DAG.getEntryNode(), 2601 DAG.getConstant(LabelID, MVT::i32))); 2602 2603 // Mark exception register as live in. 2604 unsigned Reg = TLI.getExceptionAddressRegister(); 2605 if (Reg) CurMBB->addLiveIn(Reg); 2606 2607 // Insert the EXCEPTIONADDR instruction. 2608 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other); 2609 SDOperand Ops[1]; 2610 Ops[0] = DAG.getRoot(); 2611 SDOperand Op = DAG.getNode(ISD::EXCEPTIONADDR, VTs, Ops, 1); 2612 setValue(&I, Op); 2613 DAG.setRoot(Op.getValue(1)); 2614 } else { 2615 setValue(&I, DAG.getConstant(0, TLI.getPointerTy())); 2616 } 2617 return 0; 2618 } 2619 2620 case Intrinsic::eh_selector: 2621 case Intrinsic::eh_filter:{ 2622 MachineModuleInfo *MMI = DAG.getMachineModuleInfo(); 2623 2624 if (MMI) { 2625 // Inform the MachineModuleInfo of the personality for this landing pad. 2626 ConstantExpr *CE = dyn_cast<ConstantExpr>(I.getOperand(2)); 2627 assert(CE && CE->getOpcode() == Instruction::BitCast && 2628 isa<Function>(CE->getOperand(0)) && 2629 "Personality should be a function"); 2630 MMI->addPersonality(CurMBB, cast<Function>(CE->getOperand(0))); 2631 if (Intrinsic == Intrinsic::eh_filter) 2632 MMI->setIsFilterLandingPad(CurMBB); 2633 2634 // Gather all the type infos for this landing pad and pass them along to 2635 // MachineModuleInfo. 2636 std::vector<GlobalVariable *> TyInfo; 2637 for (unsigned i = 3, N = I.getNumOperands(); i < N; ++i) { 2638 Constant *C = cast<Constant>(I.getOperand(i)); 2639 GlobalVariable *GV = ExtractGlobalVariable(C); 2640 assert (GV || (isa<ConstantInt>(C) && 2641 cast<ConstantInt>(C)->isNullValue()) && 2642 "TypeInfo must be a global variable or NULL"); 2643 TyInfo.push_back(GV); 2644 } 2645 MMI->addCatchTypeInfo(CurMBB, TyInfo); 2646 2647 // Mark exception selector register as live in. 2648 unsigned Reg = TLI.getExceptionSelectorRegister(); 2649 if (Reg) CurMBB->addLiveIn(Reg); 2650 2651 // Insert the EHSELECTION instruction. 2652 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other); 2653 SDOperand Ops[2]; 2654 Ops[0] = getValue(I.getOperand(1)); 2655 Ops[1] = getRoot(); 2656 SDOperand Op = DAG.getNode(ISD::EHSELECTION, VTs, Ops, 2); 2657 setValue(&I, Op); 2658 DAG.setRoot(Op.getValue(1)); 2659 } else { 2660 setValue(&I, DAG.getConstant(0, TLI.getPointerTy())); 2661 } 2662 2663 return 0; 2664 } 2665 2666 case Intrinsic::eh_typeid_for: { 2667 MachineModuleInfo *MMI = DAG.getMachineModuleInfo(); 2668 2669 if (MMI) { 2670 // Find the type id for the given typeinfo. 2671 Constant *C = cast<Constant>(I.getOperand(1)); 2672 GlobalVariable *GV = ExtractGlobalVariable(C); 2673 assert (GV || (isa<ConstantInt>(C) && 2674 cast<ConstantInt>(C)->isNullValue()) && 2675 "TypeInfo must be a global variable or NULL"); 2676 2677 unsigned TypeID = MMI->getTypeIDFor(GV); 2678 setValue(&I, DAG.getConstant(TypeID, MVT::i32)); 2679 } else { 2680 setValue(&I, DAG.getConstant(0, MVT::i32)); 2681 } 2682 2683 return 0; 2684 } 2685 2686 case Intrinsic::sqrt_f32: 2687 case Intrinsic::sqrt_f64: 2688 setValue(&I, DAG.getNode(ISD::FSQRT, 2689 getValue(I.getOperand(1)).getValueType(), 2690 getValue(I.getOperand(1)))); 2691 return 0; 2692 case Intrinsic::powi_f32: 2693 case Intrinsic::powi_f64: 2694 setValue(&I, DAG.getNode(ISD::FPOWI, 2695 getValue(I.getOperand(1)).getValueType(), 2696 getValue(I.getOperand(1)), 2697 getValue(I.getOperand(2)))); 2698 return 0; 2699 case Intrinsic::pcmarker: { 2700 SDOperand Tmp = getValue(I.getOperand(1)); 2701 DAG.setRoot(DAG.getNode(ISD::PCMARKER, MVT::Other, getRoot(), Tmp)); 2702 return 0; 2703 } 2704 case Intrinsic::readcyclecounter: { 2705 SDOperand Op = getRoot(); 2706 SDOperand Tmp = DAG.getNode(ISD::READCYCLECOUNTER, 2707 DAG.getNodeValueTypes(MVT::i64, MVT::Other), 2, 2708 &Op, 1); 2709 setValue(&I, Tmp); 2710 DAG.setRoot(Tmp.getValue(1)); 2711 return 0; 2712 } 2713 case Intrinsic::part_select: { 2714 // Currently not implemented: just abort 2715 assert(0 && "part_select intrinsic not implemented"); 2716 abort(); 2717 } 2718 case Intrinsic::part_set: { 2719 // Currently not implemented: just abort 2720 assert(0 && "part_set intrinsic not implemented"); 2721 abort(); 2722 } 2723 case Intrinsic::bswap: 2724 setValue(&I, DAG.getNode(ISD::BSWAP, 2725 getValue(I.getOperand(1)).getValueType(), 2726 getValue(I.getOperand(1)))); 2727 return 0; 2728 case Intrinsic::cttz: { 2729 SDOperand Arg = getValue(I.getOperand(1)); 2730 MVT::ValueType Ty = Arg.getValueType(); 2731 SDOperand result = DAG.getNode(ISD::CTTZ, Ty, Arg); 2732 if (Ty < MVT::i32) 2733 result = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, result); 2734 else if (Ty > MVT::i32) 2735 result = DAG.getNode(ISD::TRUNCATE, MVT::i32, result); 2736 setValue(&I, result); 2737 return 0; 2738 } 2739 case Intrinsic::ctlz: { 2740 SDOperand Arg = getValue(I.getOperand(1)); 2741 MVT::ValueType Ty = Arg.getValueType(); 2742 SDOperand result = DAG.getNode(ISD::CTLZ, Ty, Arg); 2743 if (Ty < MVT::i32) 2744 result = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, result); 2745 else if (Ty > MVT::i32) 2746 result = DAG.getNode(ISD::TRUNCATE, MVT::i32, result); 2747 setValue(&I, result); 2748 return 0; 2749 } 2750 case Intrinsic::ctpop: { 2751 SDOperand Arg = getValue(I.getOperand(1)); 2752 MVT::ValueType Ty = Arg.getValueType(); 2753 SDOperand result = DAG.getNode(ISD::CTPOP, Ty, Arg); 2754 if (Ty < MVT::i32) 2755 result = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, result); 2756 else if (Ty > MVT::i32) 2757 result = DAG.getNode(ISD::TRUNCATE, MVT::i32, result); 2758 setValue(&I, result); 2759 return 0; 2760 } 2761 case Intrinsic::stacksave: { 2762 SDOperand Op = getRoot(); 2763 SDOperand Tmp = DAG.getNode(ISD::STACKSAVE, 2764 DAG.getNodeValueTypes(TLI.getPointerTy(), MVT::Other), 2, &Op, 1); 2765 setValue(&I, Tmp); 2766 DAG.setRoot(Tmp.getValue(1)); 2767 return 0; 2768 } 2769 case Intrinsic::stackrestore: { 2770 SDOperand Tmp = getValue(I.getOperand(1)); 2771 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, MVT::Other, getRoot(), Tmp)); 2772 return 0; 2773 } 2774 case Intrinsic::prefetch: 2775 // FIXME: Currently discarding prefetches. 2776 return 0; 2777 } 2778} 2779 2780 2781void SelectionDAGLowering::LowerCallTo(Instruction &I, 2782 const Type *CalledValueTy, 2783 unsigned CallingConv, 2784 bool IsTailCall, 2785 SDOperand Callee, unsigned OpIdx) { 2786 const PointerType *PT = cast<PointerType>(CalledValueTy); 2787 const FunctionType *FTy = cast<FunctionType>(PT->getElementType()); 2788 const ParamAttrsList *Attrs = FTy->getParamAttrs(); 2789 2790 TargetLowering::ArgListTy Args; 2791 TargetLowering::ArgListEntry Entry; 2792 Args.reserve(I.getNumOperands()); 2793 for (unsigned i = OpIdx, e = I.getNumOperands(); i != e; ++i) { 2794 Value *Arg = I.getOperand(i); 2795 SDOperand ArgNode = getValue(Arg); 2796 Entry.Node = ArgNode; Entry.Ty = Arg->getType(); 2797 2798 unsigned attrInd = i - OpIdx + 1; 2799 Entry.isSExt = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::SExt); 2800 Entry.isZExt = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::ZExt); 2801 Entry.isInReg = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::InReg); 2802 Entry.isSRet = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::StructRet); 2803 Args.push_back(Entry); 2804 } 2805 2806 std::pair<SDOperand,SDOperand> Result = 2807 TLI.LowerCallTo(getRoot(), I.getType(), 2808 Attrs && Attrs->paramHasAttr(0, ParamAttr::SExt), 2809 FTy->isVarArg(), CallingConv, IsTailCall, 2810 Callee, Args, DAG); 2811 if (I.getType() != Type::VoidTy) 2812 setValue(&I, Result.first); 2813 DAG.setRoot(Result.second); 2814} 2815 2816 2817void SelectionDAGLowering::visitCall(CallInst &I) { 2818 const char *RenameFn = 0; 2819 if (Function *F = I.getCalledFunction()) { 2820 if (F->isDeclaration()) 2821 if (unsigned IID = F->getIntrinsicID()) { 2822 RenameFn = visitIntrinsicCall(I, IID); 2823 if (!RenameFn) 2824 return; 2825 } else { // Not an LLVM intrinsic. 2826 const std::string &Name = F->getName(); 2827 if (Name[0] == 'c' && (Name == "copysign" || Name == "copysignf")) { 2828 if (I.getNumOperands() == 3 && // Basic sanity checks. 2829 I.getOperand(1)->getType()->isFloatingPoint() && 2830 I.getType() == I.getOperand(1)->getType() && 2831 I.getType() == I.getOperand(2)->getType()) { 2832 SDOperand LHS = getValue(I.getOperand(1)); 2833 SDOperand RHS = getValue(I.getOperand(2)); 2834 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, LHS.getValueType(), 2835 LHS, RHS)); 2836 return; 2837 } 2838 } else if (Name[0] == 'f' && (Name == "fabs" || Name == "fabsf")) { 2839 if (I.getNumOperands() == 2 && // Basic sanity checks. 2840 I.getOperand(1)->getType()->isFloatingPoint() && 2841 I.getType() == I.getOperand(1)->getType()) { 2842 SDOperand Tmp = getValue(I.getOperand(1)); 2843 setValue(&I, DAG.getNode(ISD::FABS, Tmp.getValueType(), Tmp)); 2844 return; 2845 } 2846 } else if (Name[0] == 's' && (Name == "sin" || Name == "sinf")) { 2847 if (I.getNumOperands() == 2 && // Basic sanity checks. 2848 I.getOperand(1)->getType()->isFloatingPoint() && 2849 I.getType() == I.getOperand(1)->getType()) { 2850 SDOperand Tmp = getValue(I.getOperand(1)); 2851 setValue(&I, DAG.getNode(ISD::FSIN, Tmp.getValueType(), Tmp)); 2852 return; 2853 } 2854 } else if (Name[0] == 'c' && (Name == "cos" || Name == "cosf")) { 2855 if (I.getNumOperands() == 2 && // Basic sanity checks. 2856 I.getOperand(1)->getType()->isFloatingPoint() && 2857 I.getType() == I.getOperand(1)->getType()) { 2858 SDOperand Tmp = getValue(I.getOperand(1)); 2859 setValue(&I, DAG.getNode(ISD::FCOS, Tmp.getValueType(), Tmp)); 2860 return; 2861 } 2862 } 2863 } 2864 } else if (isa<InlineAsm>(I.getOperand(0))) { 2865 visitInlineAsm(I); 2866 return; 2867 } 2868 2869 SDOperand Callee; 2870 if (!RenameFn) 2871 Callee = getValue(I.getOperand(0)); 2872 else 2873 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy()); 2874 2875 LowerCallTo(I, I.getCalledValue()->getType(), 2876 I.getCallingConv(), 2877 I.isTailCall(), 2878 Callee, 2879 1); 2880} 2881 2882 2883SDOperand RegsForValue::getCopyFromRegs(SelectionDAG &DAG, 2884 SDOperand &Chain, SDOperand &Flag)const{ 2885 SDOperand Val = DAG.getCopyFromReg(Chain, Regs[0], RegVT, Flag); 2886 Chain = Val.getValue(1); 2887 Flag = Val.getValue(2); 2888 2889 // If the result was expanded, copy from the top part. 2890 if (Regs.size() > 1) { 2891 assert(Regs.size() == 2 && 2892 "Cannot expand to more than 2 elts yet!"); 2893 SDOperand Hi = DAG.getCopyFromReg(Chain, Regs[1], RegVT, Flag); 2894 Chain = Hi.getValue(1); 2895 Flag = Hi.getValue(2); 2896 if (DAG.getTargetLoweringInfo().isLittleEndian()) 2897 return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Val, Hi); 2898 else 2899 return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Hi, Val); 2900 } 2901 2902 // Otherwise, if the return value was promoted or extended, truncate it to the 2903 // appropriate type. 2904 if (RegVT == ValueVT) 2905 return Val; 2906 2907 if (MVT::isVector(RegVT)) { 2908 assert(ValueVT == MVT::Vector && "Unknown vector conversion!"); 2909 return DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Val, 2910 DAG.getConstant(MVT::getVectorNumElements(RegVT), 2911 MVT::i32), 2912 DAG.getValueType(MVT::getVectorBaseType(RegVT))); 2913 } 2914 2915 if (MVT::isInteger(RegVT)) { 2916 if (ValueVT < RegVT) 2917 return DAG.getNode(ISD::TRUNCATE, ValueVT, Val); 2918 else 2919 return DAG.getNode(ISD::ANY_EXTEND, ValueVT, Val); 2920 } 2921 2922 assert(MVT::isFloatingPoint(RegVT) && MVT::isFloatingPoint(ValueVT)); 2923 return DAG.getNode(ISD::FP_ROUND, ValueVT, Val); 2924} 2925 2926/// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 2927/// specified value into the registers specified by this object. This uses 2928/// Chain/Flag as the input and updates them for the output Chain/Flag. 2929void RegsForValue::getCopyToRegs(SDOperand Val, SelectionDAG &DAG, 2930 SDOperand &Chain, SDOperand &Flag, 2931 MVT::ValueType PtrVT) const { 2932 if (Regs.size() == 1) { 2933 // If there is a single register and the types differ, this must be 2934 // a promotion. 2935 if (RegVT != ValueVT) { 2936 if (MVT::isVector(RegVT)) { 2937 assert(Val.getValueType() == MVT::Vector &&"Not a vector-vector cast?"); 2938 Val = DAG.getNode(ISD::VBIT_CONVERT, RegVT, Val); 2939 } else if (MVT::isInteger(RegVT) && MVT::isInteger(Val.getValueType())) { 2940 if (RegVT < ValueVT) 2941 Val = DAG.getNode(ISD::TRUNCATE, RegVT, Val); 2942 else 2943 Val = DAG.getNode(ISD::ANY_EXTEND, RegVT, Val); 2944 } else if (MVT::isFloatingPoint(RegVT) && 2945 MVT::isFloatingPoint(Val.getValueType())) { 2946 Val = DAG.getNode(ISD::FP_EXTEND, RegVT, Val); 2947 } else if (MVT::getSizeInBits(RegVT) == 2948 MVT::getSizeInBits(Val.getValueType())) { 2949 Val = DAG.getNode(ISD::BIT_CONVERT, RegVT, Val); 2950 } else { 2951 assert(0 && "Unknown mismatch!"); 2952 } 2953 } 2954 Chain = DAG.getCopyToReg(Chain, Regs[0], Val, Flag); 2955 Flag = Chain.getValue(1); 2956 } else { 2957 std::vector<unsigned> R(Regs); 2958 if (!DAG.getTargetLoweringInfo().isLittleEndian()) 2959 std::reverse(R.begin(), R.end()); 2960 2961 for (unsigned i = 0, e = R.size(); i != e; ++i) { 2962 SDOperand Part = DAG.getNode(ISD::EXTRACT_ELEMENT, RegVT, Val, 2963 DAG.getConstant(i, PtrVT)); 2964 Chain = DAG.getCopyToReg(Chain, R[i], Part, Flag); 2965 Flag = Chain.getValue(1); 2966 } 2967 } 2968} 2969 2970/// AddInlineAsmOperands - Add this value to the specified inlineasm node 2971/// operand list. This adds the code marker and includes the number of 2972/// values added into it. 2973void RegsForValue::AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG, 2974 std::vector<SDOperand> &Ops) const { 2975 MVT::ValueType IntPtrTy = DAG.getTargetLoweringInfo().getPointerTy(); 2976 Ops.push_back(DAG.getTargetConstant(Code | (Regs.size() << 3), IntPtrTy)); 2977 for (unsigned i = 0, e = Regs.size(); i != e; ++i) 2978 Ops.push_back(DAG.getRegister(Regs[i], RegVT)); 2979} 2980 2981/// isAllocatableRegister - If the specified register is safe to allocate, 2982/// i.e. it isn't a stack pointer or some other special register, return the 2983/// register class for the register. Otherwise, return null. 2984static const TargetRegisterClass * 2985isAllocatableRegister(unsigned Reg, MachineFunction &MF, 2986 const TargetLowering &TLI, const MRegisterInfo *MRI) { 2987 MVT::ValueType FoundVT = MVT::Other; 2988 const TargetRegisterClass *FoundRC = 0; 2989 for (MRegisterInfo::regclass_iterator RCI = MRI->regclass_begin(), 2990 E = MRI->regclass_end(); RCI != E; ++RCI) { 2991 MVT::ValueType ThisVT = MVT::Other; 2992 2993 const TargetRegisterClass *RC = *RCI; 2994 // If none of the the value types for this register class are valid, we 2995 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2996 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 2997 I != E; ++I) { 2998 if (TLI.isTypeLegal(*I)) { 2999 // If we have already found this register in a different register class, 3000 // choose the one with the largest VT specified. For example, on 3001 // PowerPC, we favor f64 register classes over f32. 3002 if (FoundVT == MVT::Other || 3003 MVT::getSizeInBits(FoundVT) < MVT::getSizeInBits(*I)) { 3004 ThisVT = *I; 3005 break; 3006 } 3007 } 3008 } 3009 3010 if (ThisVT == MVT::Other) continue; 3011 3012 // NOTE: This isn't ideal. In particular, this might allocate the 3013 // frame pointer in functions that need it (due to them not being taken 3014 // out of allocation, because a variable sized allocation hasn't been seen 3015 // yet). This is a slight code pessimization, but should still work. 3016 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF), 3017 E = RC->allocation_order_end(MF); I != E; ++I) 3018 if (*I == Reg) { 3019 // We found a matching register class. Keep looking at others in case 3020 // we find one with larger registers that this physreg is also in. 3021 FoundRC = RC; 3022 FoundVT = ThisVT; 3023 break; 3024 } 3025 } 3026 return FoundRC; 3027} 3028 3029 3030namespace { 3031/// AsmOperandInfo - This contains information for each constraint that we are 3032/// lowering. 3033struct AsmOperandInfo : public InlineAsm::ConstraintInfo { 3034 /// ConstraintCode - This contains the actual string for the code, like "m". 3035 std::string ConstraintCode; 3036 3037 /// ConstraintType - Information about the constraint code, e.g. Register, 3038 /// RegisterClass, Memory, Other, Unknown. 3039 TargetLowering::ConstraintType ConstraintType; 3040 3041 /// CallOperand/CallOperandval - If this is the result output operand or a 3042 /// clobber, this is null, otherwise it is the incoming operand to the 3043 /// CallInst. This gets modified as the asm is processed. 3044 SDOperand CallOperand; 3045 Value *CallOperandVal; 3046 3047 /// ConstraintVT - The ValueType for the operand value. 3048 MVT::ValueType ConstraintVT; 3049 3050 /// AssignedRegs - If this is a register or register class operand, this 3051 /// contains the set of register corresponding to the operand. 3052 RegsForValue AssignedRegs; 3053 3054 AsmOperandInfo(const InlineAsm::ConstraintInfo &info) 3055 : InlineAsm::ConstraintInfo(info), 3056 ConstraintType(TargetLowering::C_Unknown), 3057 CallOperand(0,0), CallOperandVal(0), ConstraintVT(MVT::Other) { 3058 } 3059 3060 void ComputeConstraintToUse(const TargetLowering &TLI); 3061 3062 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers 3063 /// busy in OutputRegs/InputRegs. 3064 void MarkAllocatedRegs(bool isOutReg, bool isInReg, 3065 std::set<unsigned> &OutputRegs, 3066 std::set<unsigned> &InputRegs) const { 3067 if (isOutReg) 3068 OutputRegs.insert(AssignedRegs.Regs.begin(), AssignedRegs.Regs.end()); 3069 if (isInReg) 3070 InputRegs.insert(AssignedRegs.Regs.begin(), AssignedRegs.Regs.end()); 3071 } 3072}; 3073} // end anon namespace. 3074 3075/// getConstraintGenerality - Return an integer indicating how general CT is. 3076static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 3077 switch (CT) { 3078 default: assert(0 && "Unknown constraint type!"); 3079 case TargetLowering::C_Other: 3080 case TargetLowering::C_Unknown: 3081 return 0; 3082 case TargetLowering::C_Register: 3083 return 1; 3084 case TargetLowering::C_RegisterClass: 3085 return 2; 3086 case TargetLowering::C_Memory: 3087 return 3; 3088 } 3089} 3090 3091void AsmOperandInfo::ComputeConstraintToUse(const TargetLowering &TLI) { 3092 assert(!Codes.empty() && "Must have at least one constraint"); 3093 3094 std::string *Current = &Codes[0]; 3095 TargetLowering::ConstraintType CurType = TLI.getConstraintType(*Current); 3096 if (Codes.size() == 1) { // Single-letter constraints ('r') are very common. 3097 ConstraintCode = *Current; 3098 ConstraintType = CurType; 3099 return; 3100 } 3101 3102 unsigned CurGenerality = getConstraintGenerality(CurType); 3103 3104 // If we have multiple constraints, try to pick the most general one ahead 3105 // of time. This isn't a wonderful solution, but handles common cases. 3106 for (unsigned j = 1, e = Codes.size(); j != e; ++j) { 3107 TargetLowering::ConstraintType ThisType = TLI.getConstraintType(Codes[j]); 3108 unsigned ThisGenerality = getConstraintGenerality(ThisType); 3109 if (ThisGenerality > CurGenerality) { 3110 // This constraint letter is more general than the previous one, 3111 // use it. 3112 CurType = ThisType; 3113 Current = &Codes[j]; 3114 CurGenerality = ThisGenerality; 3115 } 3116 } 3117 3118 ConstraintCode = *Current; 3119 ConstraintType = CurType; 3120} 3121 3122 3123void SelectionDAGLowering:: 3124GetRegistersForValue(AsmOperandInfo &OpInfo, bool HasEarlyClobber, 3125 std::set<unsigned> &OutputRegs, 3126 std::set<unsigned> &InputRegs) { 3127 // Compute whether this value requires an input register, an output register, 3128 // or both. 3129 bool isOutReg = false; 3130 bool isInReg = false; 3131 switch (OpInfo.Type) { 3132 case InlineAsm::isOutput: 3133 isOutReg = true; 3134 3135 // If this is an early-clobber output, or if there is an input 3136 // constraint that matches this, we need to reserve the input register 3137 // so no other inputs allocate to it. 3138 isInReg = OpInfo.isEarlyClobber || OpInfo.hasMatchingInput; 3139 break; 3140 case InlineAsm::isInput: 3141 isInReg = true; 3142 isOutReg = false; 3143 break; 3144 case InlineAsm::isClobber: 3145 isOutReg = true; 3146 isInReg = true; 3147 break; 3148 } 3149 3150 3151 MachineFunction &MF = DAG.getMachineFunction(); 3152 std::vector<unsigned> Regs; 3153 3154 // If this is a constraint for a single physreg, or a constraint for a 3155 // register class, find it. 3156 std::pair<unsigned, const TargetRegisterClass*> PhysReg = 3157 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode, 3158 OpInfo.ConstraintVT); 3159 3160 unsigned NumRegs = 1; 3161 if (OpInfo.ConstraintVT != MVT::Other) 3162 NumRegs = TLI.getNumElements(OpInfo.ConstraintVT); 3163 MVT::ValueType RegVT; 3164 MVT::ValueType ValueVT = OpInfo.ConstraintVT; 3165 3166 3167 // If this is a constraint for a specific physical register, like {r17}, 3168 // assign it now. 3169 if (PhysReg.first) { 3170 if (OpInfo.ConstraintVT == MVT::Other) 3171 ValueVT = *PhysReg.second->vt_begin(); 3172 3173 // Get the actual register value type. This is important, because the user 3174 // may have asked for (e.g.) the AX register in i32 type. We need to 3175 // remember that AX is actually i16 to get the right extension. 3176 RegVT = *PhysReg.second->vt_begin(); 3177 3178 // This is a explicit reference to a physical register. 3179 Regs.push_back(PhysReg.first); 3180 3181 // If this is an expanded reference, add the rest of the regs to Regs. 3182 if (NumRegs != 1) { 3183 TargetRegisterClass::iterator I = PhysReg.second->begin(); 3184 TargetRegisterClass::iterator E = PhysReg.second->end(); 3185 for (; *I != PhysReg.first; ++I) 3186 assert(I != E && "Didn't find reg!"); 3187 3188 // Already added the first reg. 3189 --NumRegs; ++I; 3190 for (; NumRegs; --NumRegs, ++I) { 3191 assert(I != E && "Ran out of registers to allocate!"); 3192 Regs.push_back(*I); 3193 } 3194 } 3195 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); 3196 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs); 3197 return; 3198 } 3199 3200 // Otherwise, if this was a reference to an LLVM register class, create vregs 3201 // for this reference. 3202 std::vector<unsigned> RegClassRegs; 3203 if (PhysReg.second) { 3204 // If this is an early clobber or tied register, our regalloc doesn't know 3205 // how to maintain the constraint. If it isn't, go ahead and create vreg 3206 // and let the regalloc do the right thing. 3207 if (!OpInfo.hasMatchingInput && !OpInfo.isEarlyClobber && 3208 // If there is some other early clobber and this is an input register, 3209 // then we are forced to pre-allocate the input reg so it doesn't 3210 // conflict with the earlyclobber. 3211 !(OpInfo.Type == InlineAsm::isInput && HasEarlyClobber)) { 3212 RegVT = *PhysReg.second->vt_begin(); 3213 3214 if (OpInfo.ConstraintVT == MVT::Other) 3215 ValueVT = RegVT; 3216 3217 // Create the appropriate number of virtual registers. 3218 SSARegMap *RegMap = MF.getSSARegMap(); 3219 for (; NumRegs; --NumRegs) 3220 Regs.push_back(RegMap->createVirtualRegister(PhysReg.second)); 3221 3222 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT); 3223 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs); 3224 return; 3225 } 3226 3227 // Otherwise, we can't allocate it. Let the code below figure out how to 3228 // maintain these constraints. 3229 RegClassRegs.assign(PhysReg.second->begin(), PhysReg.second->end()); 3230 3231 } else { 3232 // This is a reference to a register class that doesn't directly correspond 3233 // to an LLVM register class. Allocate NumRegs consecutive, available, 3234 // registers from the class. 3235 RegClassRegs = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode, 3236 OpInfo.ConstraintVT); 3237 } 3238 3239 const MRegisterInfo *MRI = DAG.getTarget().getRegisterInfo(); 3240 unsigned NumAllocated = 0; 3241 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) { 3242 unsigned Reg = RegClassRegs[i]; 3243 // See if this register is available. 3244 if ((isOutReg && OutputRegs.count(Reg)) || // Already used. 3245 (isInReg && InputRegs.count(Reg))) { // Already used. 3246 // Make sure we find consecutive registers. 3247 NumAllocated = 0; 3248 continue; 3249 } 3250 3251 // Check to see if this register is allocatable (i.e. don't give out the 3252 // stack pointer). 3253 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, MRI); 3254 if (!RC) { 3255 // Make sure we find consecutive registers. 3256 NumAllocated = 0; 3257 continue; 3258 } 3259 3260 // Okay, this register is good, we can use it. 3261 ++NumAllocated; 3262 3263 // If we allocated enough consecutive registers, succeed. 3264 if (NumAllocated == NumRegs) { 3265 unsigned RegStart = (i-NumAllocated)+1; 3266 unsigned RegEnd = i+1; 3267 // Mark all of the allocated registers used. 3268 for (unsigned i = RegStart; i != RegEnd; ++i) 3269 Regs.push_back(RegClassRegs[i]); 3270 3271 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(), 3272 OpInfo.ConstraintVT); 3273 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs); 3274 return; 3275 } 3276 } 3277 3278 // Otherwise, we couldn't allocate enough registers for this. 3279 return; 3280} 3281 3282 3283/// visitInlineAsm - Handle a call to an InlineAsm object. 3284/// 3285void SelectionDAGLowering::visitInlineAsm(CallInst &I) { 3286 InlineAsm *IA = cast<InlineAsm>(I.getOperand(0)); 3287 3288 /// ConstraintOperands - Information about all of the constraints. 3289 std::vector<AsmOperandInfo> ConstraintOperands; 3290 3291 SDOperand Chain = getRoot(); 3292 SDOperand Flag; 3293 3294 std::set<unsigned> OutputRegs, InputRegs; 3295 3296 // Do a prepass over the constraints, canonicalizing them, and building up the 3297 // ConstraintOperands list. 3298 std::vector<InlineAsm::ConstraintInfo> 3299 ConstraintInfos = IA->ParseConstraints(); 3300 3301 // SawEarlyClobber - Keep track of whether we saw an earlyclobber output 3302 // constraint. If so, we can't let the register allocator allocate any input 3303 // registers, because it will not know to avoid the earlyclobbered output reg. 3304 bool SawEarlyClobber = false; 3305 3306 unsigned OpNo = 1; // OpNo - The operand of the CallInst. 3307 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 3308 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i])); 3309 AsmOperandInfo &OpInfo = ConstraintOperands.back(); 3310 3311 MVT::ValueType OpVT = MVT::Other; 3312 3313 // Compute the value type for each operand. 3314 switch (OpInfo.Type) { 3315 case InlineAsm::isOutput: 3316 if (!OpInfo.isIndirect) { 3317 // The return value of the call is this value. As such, there is no 3318 // corresponding argument. 3319 assert(I.getType() != Type::VoidTy && "Bad inline asm!"); 3320 OpVT = TLI.getValueType(I.getType()); 3321 } else { 3322 OpInfo.CallOperandVal = I.getOperand(OpNo++); 3323 } 3324 break; 3325 case InlineAsm::isInput: 3326 OpInfo.CallOperandVal = I.getOperand(OpNo++); 3327 break; 3328 case InlineAsm::isClobber: 3329 // Nothing to do. 3330 break; 3331 } 3332 3333 // If this is an input or an indirect output, process the call argument. 3334 if (OpInfo.CallOperandVal) { 3335 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal); 3336 const Type *OpTy = OpInfo.CallOperandVal->getType(); 3337 // If this is an indirect operand, the operand is a pointer to the 3338 // accessed type. 3339 if (OpInfo.isIndirect) 3340 OpTy = cast<PointerType>(OpTy)->getElementType(); 3341 3342 // If OpTy is not a first-class value, it may be a struct/union that we 3343 // can tile with integers. 3344 if (!OpTy->isFirstClassType() && OpTy->isSized()) { 3345 unsigned BitSize = TD->getTypeSizeInBits(OpTy); 3346 switch (BitSize) { 3347 default: break; 3348 case 1: 3349 case 8: 3350 case 16: 3351 case 32: 3352 case 64: 3353 OpTy = IntegerType::get(BitSize); 3354 break; 3355 } 3356 } 3357 3358 OpVT = TLI.getValueType(OpTy, true); 3359 } 3360 3361 OpInfo.ConstraintVT = OpVT; 3362 3363 // Compute the constraint code and ConstraintType to use. 3364 OpInfo.ComputeConstraintToUse(TLI); 3365 3366 // Keep track of whether we see an earlyclobber. 3367 SawEarlyClobber |= OpInfo.isEarlyClobber; 3368 3369 // If this is a memory input, and if the operand is not indirect, do what we 3370 // need to to provide an address for the memory input. 3371 if (OpInfo.ConstraintType == TargetLowering::C_Memory && 3372 !OpInfo.isIndirect) { 3373 assert(OpInfo.Type == InlineAsm::isInput && 3374 "Can only indirectify direct input operands!"); 3375 3376 // Memory operands really want the address of the value. If we don't have 3377 // an indirect input, put it in the constpool if we can, otherwise spill 3378 // it to a stack slot. 3379 3380 // If the operand is a float, integer, or vector constant, spill to a 3381 // constant pool entry to get its address. 3382 Value *OpVal = OpInfo.CallOperandVal; 3383 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) || 3384 isa<ConstantVector>(OpVal)) { 3385 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal), 3386 TLI.getPointerTy()); 3387 } else { 3388 // Otherwise, create a stack slot and emit a store to it before the 3389 // asm. 3390 const Type *Ty = OpVal->getType(); 3391 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty); 3392 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty); 3393 MachineFunction &MF = DAG.getMachineFunction(); 3394 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align); 3395 SDOperand StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy()); 3396 Chain = DAG.getStore(Chain, OpInfo.CallOperand, StackSlot, NULL, 0); 3397 OpInfo.CallOperand = StackSlot; 3398 } 3399 3400 // There is no longer a Value* corresponding to this operand. 3401 OpInfo.CallOperandVal = 0; 3402 // It is now an indirect operand. 3403 OpInfo.isIndirect = true; 3404 } 3405 3406 // If this constraint is for a specific register, allocate it before 3407 // anything else. 3408 if (OpInfo.ConstraintType == TargetLowering::C_Register) 3409 GetRegistersForValue(OpInfo, SawEarlyClobber, OutputRegs, InputRegs); 3410 } 3411 ConstraintInfos.clear(); 3412 3413 3414 // Second pass - Loop over all of the operands, assigning virtual or physregs 3415 // to registerclass operands. 3416 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { 3417 AsmOperandInfo &OpInfo = ConstraintOperands[i]; 3418 3419 // C_Register operands have already been allocated, Other/Memory don't need 3420 // to be. 3421 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass) 3422 GetRegistersForValue(OpInfo, SawEarlyClobber, OutputRegs, InputRegs); 3423 } 3424 3425 // AsmNodeOperands - The operands for the ISD::INLINEASM node. 3426 std::vector<SDOperand> AsmNodeOperands; 3427 AsmNodeOperands.push_back(SDOperand()); // reserve space for input chain 3428 AsmNodeOperands.push_back( 3429 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), MVT::Other)); 3430 3431 3432 // Loop over all of the inputs, copying the operand values into the 3433 // appropriate registers and processing the output regs. 3434 RegsForValue RetValRegs; 3435 3436 // IndirectStoresToEmit - The set of stores to emit after the inline asm node. 3437 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit; 3438 3439 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) { 3440 AsmOperandInfo &OpInfo = ConstraintOperands[i]; 3441 3442 switch (OpInfo.Type) { 3443 case InlineAsm::isOutput: { 3444 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass && 3445 OpInfo.ConstraintType != TargetLowering::C_Register) { 3446 // Memory output, or 'other' output (e.g. 'X' constraint). 3447 assert(OpInfo.isIndirect && "Memory output must be indirect operand"); 3448 3449 // Add information to the INLINEASM node to know about this output. 3450 unsigned ResOpType = 4/*MEM*/ | (1 << 3); 3451 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32)); 3452 AsmNodeOperands.push_back(OpInfo.CallOperand); 3453 break; 3454 } 3455 3456 // Otherwise, this is a register or register class output. 3457 3458 // Copy the output from the appropriate register. Find a register that 3459 // we can use. 3460 if (OpInfo.AssignedRegs.Regs.empty()) { 3461 cerr << "Couldn't allocate output reg for contraint '" 3462 << OpInfo.ConstraintCode << "'!\n"; 3463 exit(1); 3464 } 3465 3466 if (!OpInfo.isIndirect) { 3467 // This is the result value of the call. 3468 assert(RetValRegs.Regs.empty() && 3469 "Cannot have multiple output constraints yet!"); 3470 assert(I.getType() != Type::VoidTy && "Bad inline asm!"); 3471 RetValRegs = OpInfo.AssignedRegs; 3472 } else { 3473 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs, 3474 OpInfo.CallOperandVal)); 3475 } 3476 3477 // Add information to the INLINEASM node to know that this register is 3478 // set. 3479 OpInfo.AssignedRegs.AddInlineAsmOperands(2 /*REGDEF*/, DAG, 3480 AsmNodeOperands); 3481 break; 3482 } 3483 case InlineAsm::isInput: { 3484 SDOperand InOperandVal = OpInfo.CallOperand; 3485 3486 if (isdigit(OpInfo.ConstraintCode[0])) { // Matching constraint? 3487 // If this is required to match an output register we have already set, 3488 // just use its register. 3489 unsigned OperandNo = atoi(OpInfo.ConstraintCode.c_str()); 3490 3491 // Scan until we find the definition we already emitted of this operand. 3492 // When we find it, create a RegsForValue operand. 3493 unsigned CurOp = 2; // The first operand. 3494 for (; OperandNo; --OperandNo) { 3495 // Advance to the next operand. 3496 unsigned NumOps = 3497 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue(); 3498 assert(((NumOps & 7) == 2 /*REGDEF*/ || 3499 (NumOps & 7) == 4 /*MEM*/) && 3500 "Skipped past definitions?"); 3501 CurOp += (NumOps>>3)+1; 3502 } 3503 3504 unsigned NumOps = 3505 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue(); 3506 if ((NumOps & 7) == 2 /*REGDEF*/) { 3507 // Add NumOps>>3 registers to MatchedRegs. 3508 RegsForValue MatchedRegs; 3509 MatchedRegs.ValueVT = InOperandVal.getValueType(); 3510 MatchedRegs.RegVT = AsmNodeOperands[CurOp+1].getValueType(); 3511 for (unsigned i = 0, e = NumOps>>3; i != e; ++i) { 3512 unsigned Reg = 3513 cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg(); 3514 MatchedRegs.Regs.push_back(Reg); 3515 } 3516 3517 // Use the produced MatchedRegs object to 3518 MatchedRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag, 3519 TLI.getPointerTy()); 3520 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, DAG, AsmNodeOperands); 3521 break; 3522 } else { 3523 assert((NumOps & 7) == 4/*MEM*/ && "Unknown matching constraint!"); 3524 assert(0 && "matching constraints for memory operands unimp"); 3525 } 3526 } 3527 3528 if (OpInfo.ConstraintType == TargetLowering::C_Other) { 3529 assert(!OpInfo.isIndirect && 3530 "Don't know how to handle indirect other inputs yet!"); 3531 3532 InOperandVal = TLI.isOperandValidForConstraint(InOperandVal, 3533 OpInfo.ConstraintCode[0], 3534 DAG); 3535 if (!InOperandVal.Val) { 3536 cerr << "Invalid operand for inline asm constraint '" 3537 << OpInfo.ConstraintCode << "'!\n"; 3538 exit(1); 3539 } 3540 3541 // Add information to the INLINEASM node to know about this input. 3542 unsigned ResOpType = 3 /*IMM*/ | (1 << 3); 3543 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32)); 3544 AsmNodeOperands.push_back(InOperandVal); 3545 break; 3546 } else if (OpInfo.ConstraintType == TargetLowering::C_Memory) { 3547 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!"); 3548 assert(InOperandVal.getValueType() == TLI.getPointerTy() && 3549 "Memory operands expect pointer values"); 3550 3551 // Add information to the INLINEASM node to know about this input. 3552 unsigned ResOpType = 4/*MEM*/ | (1 << 3); 3553 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32)); 3554 AsmNodeOperands.push_back(InOperandVal); 3555 break; 3556 } 3557 3558 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass || 3559 OpInfo.ConstraintType == TargetLowering::C_Register) && 3560 "Unknown constraint type!"); 3561 assert(!OpInfo.isIndirect && 3562 "Don't know how to handle indirect register inputs yet!"); 3563 3564 // Copy the input into the appropriate registers. 3565 assert(!OpInfo.AssignedRegs.Regs.empty() && 3566 "Couldn't allocate input reg!"); 3567 3568 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag, 3569 TLI.getPointerTy()); 3570 3571 OpInfo.AssignedRegs.AddInlineAsmOperands(1/*REGUSE*/, DAG, 3572 AsmNodeOperands); 3573 break; 3574 } 3575 case InlineAsm::isClobber: { 3576 // Add the clobbered value to the operand list, so that the register 3577 // allocator is aware that the physreg got clobbered. 3578 if (!OpInfo.AssignedRegs.Regs.empty()) 3579 OpInfo.AssignedRegs.AddInlineAsmOperands(2/*REGDEF*/, DAG, 3580 AsmNodeOperands); 3581 break; 3582 } 3583 } 3584 } 3585 3586 // Finish up input operands. 3587 AsmNodeOperands[0] = Chain; 3588 if (Flag.Val) AsmNodeOperands.push_back(Flag); 3589 3590 Chain = DAG.getNode(ISD::INLINEASM, 3591 DAG.getNodeValueTypes(MVT::Other, MVT::Flag), 2, 3592 &AsmNodeOperands[0], AsmNodeOperands.size()); 3593 Flag = Chain.getValue(1); 3594 3595 // If this asm returns a register value, copy the result from that register 3596 // and set it as the value of the call. 3597 if (!RetValRegs.Regs.empty()) { 3598 SDOperand Val = RetValRegs.getCopyFromRegs(DAG, Chain, Flag); 3599 3600 // If the result of the inline asm is a vector, it may have the wrong 3601 // width/num elts. Make sure to convert it to the right type with 3602 // vbit_convert. 3603 if (Val.getValueType() == MVT::Vector) { 3604 const VectorType *VTy = cast<VectorType>(I.getType()); 3605 unsigned DesiredNumElts = VTy->getNumElements(); 3606 MVT::ValueType DesiredEltVT = TLI.getValueType(VTy->getElementType()); 3607 3608 Val = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Val, 3609 DAG.getConstant(DesiredNumElts, MVT::i32), 3610 DAG.getValueType(DesiredEltVT)); 3611 } 3612 3613 setValue(&I, Val); 3614 } 3615 3616 std::vector<std::pair<SDOperand, Value*> > StoresToEmit; 3617 3618 // Process indirect outputs, first output all of the flagged copies out of 3619 // physregs. 3620 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) { 3621 RegsForValue &OutRegs = IndirectStoresToEmit[i].first; 3622 Value *Ptr = IndirectStoresToEmit[i].second; 3623 SDOperand OutVal = OutRegs.getCopyFromRegs(DAG, Chain, Flag); 3624 StoresToEmit.push_back(std::make_pair(OutVal, Ptr)); 3625 } 3626 3627 // Emit the non-flagged stores from the physregs. 3628 SmallVector<SDOperand, 8> OutChains; 3629 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) 3630 OutChains.push_back(DAG.getStore(Chain, StoresToEmit[i].first, 3631 getValue(StoresToEmit[i].second), 3632 StoresToEmit[i].second, 0)); 3633 if (!OutChains.empty()) 3634 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, 3635 &OutChains[0], OutChains.size()); 3636 DAG.setRoot(Chain); 3637} 3638 3639 3640void SelectionDAGLowering::visitMalloc(MallocInst &I) { 3641 SDOperand Src = getValue(I.getOperand(0)); 3642 3643 MVT::ValueType IntPtr = TLI.getPointerTy(); 3644 3645 if (IntPtr < Src.getValueType()) 3646 Src = DAG.getNode(ISD::TRUNCATE, IntPtr, Src); 3647 else if (IntPtr > Src.getValueType()) 3648 Src = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, Src); 3649 3650 // Scale the source by the type size. 3651 uint64_t ElementSize = TD->getTypeSize(I.getType()->getElementType()); 3652 Src = DAG.getNode(ISD::MUL, Src.getValueType(), 3653 Src, getIntPtrConstant(ElementSize)); 3654 3655 TargetLowering::ArgListTy Args; 3656 TargetLowering::ArgListEntry Entry; 3657 Entry.Node = Src; 3658 Entry.Ty = TLI.getTargetData()->getIntPtrType(); 3659 Args.push_back(Entry); 3660 3661 std::pair<SDOperand,SDOperand> Result = 3662 TLI.LowerCallTo(getRoot(), I.getType(), false, false, CallingConv::C, true, 3663 DAG.getExternalSymbol("malloc", IntPtr), 3664 Args, DAG); 3665 setValue(&I, Result.first); // Pointers always fit in registers 3666 DAG.setRoot(Result.second); 3667} 3668 3669void SelectionDAGLowering::visitFree(FreeInst &I) { 3670 TargetLowering::ArgListTy Args; 3671 TargetLowering::ArgListEntry Entry; 3672 Entry.Node = getValue(I.getOperand(0)); 3673 Entry.Ty = TLI.getTargetData()->getIntPtrType(); 3674 Args.push_back(Entry); 3675 MVT::ValueType IntPtr = TLI.getPointerTy(); 3676 std::pair<SDOperand,SDOperand> Result = 3677 TLI.LowerCallTo(getRoot(), Type::VoidTy, false, false, CallingConv::C, true, 3678 DAG.getExternalSymbol("free", IntPtr), Args, DAG); 3679 DAG.setRoot(Result.second); 3680} 3681 3682// InsertAtEndOfBasicBlock - This method should be implemented by targets that 3683// mark instructions with the 'usesCustomDAGSchedInserter' flag. These 3684// instructions are special in various ways, which require special support to 3685// insert. The specified MachineInstr is created but not inserted into any 3686// basic blocks, and the scheduler passes ownership of it to this method. 3687MachineBasicBlock *TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI, 3688 MachineBasicBlock *MBB) { 3689 cerr << "If a target marks an instruction with " 3690 << "'usesCustomDAGSchedInserter', it must implement " 3691 << "TargetLowering::InsertAtEndOfBasicBlock!\n"; 3692 abort(); 3693 return 0; 3694} 3695 3696void SelectionDAGLowering::visitVAStart(CallInst &I) { 3697 DAG.setRoot(DAG.getNode(ISD::VASTART, MVT::Other, getRoot(), 3698 getValue(I.getOperand(1)), 3699 DAG.getSrcValue(I.getOperand(1)))); 3700} 3701 3702void SelectionDAGLowering::visitVAArg(VAArgInst &I) { 3703 SDOperand V = DAG.getVAArg(TLI.getValueType(I.getType()), getRoot(), 3704 getValue(I.getOperand(0)), 3705 DAG.getSrcValue(I.getOperand(0))); 3706 setValue(&I, V); 3707 DAG.setRoot(V.getValue(1)); 3708} 3709 3710void SelectionDAGLowering::visitVAEnd(CallInst &I) { 3711 DAG.setRoot(DAG.getNode(ISD::VAEND, MVT::Other, getRoot(), 3712 getValue(I.getOperand(1)), 3713 DAG.getSrcValue(I.getOperand(1)))); 3714} 3715 3716void SelectionDAGLowering::visitVACopy(CallInst &I) { 3717 DAG.setRoot(DAG.getNode(ISD::VACOPY, MVT::Other, getRoot(), 3718 getValue(I.getOperand(1)), 3719 getValue(I.getOperand(2)), 3720 DAG.getSrcValue(I.getOperand(1)), 3721 DAG.getSrcValue(I.getOperand(2)))); 3722} 3723 3724/// ExpandScalarFormalArgs - Recursively expand the formal_argument node, either 3725/// bit_convert it or join a pair of them with a BUILD_PAIR when appropriate. 3726static SDOperand ExpandScalarFormalArgs(MVT::ValueType VT, SDNode *Arg, 3727 unsigned &i, SelectionDAG &DAG, 3728 TargetLowering &TLI) { 3729 if (TLI.getTypeAction(VT) != TargetLowering::Expand) 3730 return SDOperand(Arg, i++); 3731 3732 MVT::ValueType EVT = TLI.getTypeToTransformTo(VT); 3733 unsigned NumVals = MVT::getSizeInBits(VT) / MVT::getSizeInBits(EVT); 3734 if (NumVals == 1) { 3735 return DAG.getNode(ISD::BIT_CONVERT, VT, 3736 ExpandScalarFormalArgs(EVT, Arg, i, DAG, TLI)); 3737 } else if (NumVals == 2) { 3738 SDOperand Lo = ExpandScalarFormalArgs(EVT, Arg, i, DAG, TLI); 3739 SDOperand Hi = ExpandScalarFormalArgs(EVT, Arg, i, DAG, TLI); 3740 if (!TLI.isLittleEndian()) 3741 std::swap(Lo, Hi); 3742 return DAG.getNode(ISD::BUILD_PAIR, VT, Lo, Hi); 3743 } else { 3744 // Value scalarized into many values. Unimp for now. 3745 assert(0 && "Cannot expand i64 -> i16 yet!"); 3746 } 3747 return SDOperand(); 3748} 3749 3750/// TargetLowering::LowerArguments - This is the default LowerArguments 3751/// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all 3752/// targets are migrated to using FORMAL_ARGUMENTS, this hook should be 3753/// integrated into SDISel. 3754std::vector<SDOperand> 3755TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) { 3756 const FunctionType *FTy = F.getFunctionType(); 3757 const ParamAttrsList *Attrs = FTy->getParamAttrs(); 3758 // Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node. 3759 std::vector<SDOperand> Ops; 3760 Ops.push_back(DAG.getRoot()); 3761 Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy())); 3762 Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy())); 3763 3764 // Add one result value for each formal argument. 3765 std::vector<MVT::ValueType> RetVals; 3766 unsigned j = 1; 3767 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); 3768 I != E; ++I, ++j) { 3769 MVT::ValueType VT = getValueType(I->getType()); 3770 unsigned Flags = ISD::ParamFlags::NoFlagSet; 3771 unsigned OriginalAlignment = 3772 getTargetData()->getABITypeAlignment(I->getType()); 3773 3774 // FIXME: Distinguish between a formal with no [sz]ext attribute from one 3775 // that is zero extended! 3776 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::ZExt)) 3777 Flags &= ~(ISD::ParamFlags::SExt); 3778 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::SExt)) 3779 Flags |= ISD::ParamFlags::SExt; 3780 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::InReg)) 3781 Flags |= ISD::ParamFlags::InReg; 3782 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::StructRet)) 3783 Flags |= ISD::ParamFlags::StructReturn; 3784 Flags |= (OriginalAlignment << ISD::ParamFlags::OrigAlignmentOffs); 3785 3786 switch (getTypeAction(VT)) { 3787 default: assert(0 && "Unknown type action!"); 3788 case Legal: 3789 RetVals.push_back(VT); 3790 Ops.push_back(DAG.getConstant(Flags, MVT::i32)); 3791 break; 3792 case Promote: 3793 RetVals.push_back(getTypeToTransformTo(VT)); 3794 Ops.push_back(DAG.getConstant(Flags, MVT::i32)); 3795 break; 3796 case Expand: 3797 if (VT != MVT::Vector) { 3798 // If this is a large integer, it needs to be broken up into small 3799 // integers. Figure out what the destination type is and how many small 3800 // integers it turns into. 3801 MVT::ValueType NVT = getTypeToExpandTo(VT); 3802 unsigned NumVals = getNumElements(VT); 3803 for (unsigned i = 0; i != NumVals; ++i) { 3804 RetVals.push_back(NVT); 3805 // if it isn't first piece, alignment must be 1 3806 if (i > 0) 3807 Flags = (Flags & (~ISD::ParamFlags::OrigAlignment)) | 3808 (1 << ISD::ParamFlags::OrigAlignmentOffs); 3809 Ops.push_back(DAG.getConstant(Flags, MVT::i32)); 3810 } 3811 } else { 3812 // Otherwise, this is a vector type. We only support legal vectors 3813 // right now. 3814 unsigned NumElems = cast<VectorType>(I->getType())->getNumElements(); 3815 const Type *EltTy = cast<VectorType>(I->getType())->getElementType(); 3816 3817 // Figure out if there is a Packed type corresponding to this Vector 3818 // type. If so, convert to the vector type. 3819 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems); 3820 if (TVT != MVT::Other && isTypeLegal(TVT)) { 3821 RetVals.push_back(TVT); 3822 Ops.push_back(DAG.getConstant(Flags, MVT::i32)); 3823 } else { 3824 assert(0 && "Don't support illegal by-val vector arguments yet!"); 3825 } 3826 } 3827 break; 3828 } 3829 } 3830 3831 RetVals.push_back(MVT::Other); 3832 3833 // Create the node. 3834 SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS, 3835 DAG.getNodeValueTypes(RetVals), RetVals.size(), 3836 &Ops[0], Ops.size()).Val; 3837 3838 DAG.setRoot(SDOperand(Result, Result->getNumValues()-1)); 3839 3840 // Set up the return result vector. 3841 Ops.clear(); 3842 unsigned i = 0; 3843 unsigned Idx = 1; 3844 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; 3845 ++I, ++Idx) { 3846 MVT::ValueType VT = getValueType(I->getType()); 3847 3848 switch (getTypeAction(VT)) { 3849 default: assert(0 && "Unknown type action!"); 3850 case Legal: 3851 Ops.push_back(SDOperand(Result, i++)); 3852 break; 3853 case Promote: { 3854 SDOperand Op(Result, i++); 3855 if (MVT::isInteger(VT)) { 3856 if (Attrs && Attrs->paramHasAttr(Idx, ParamAttr::SExt)) 3857 Op = DAG.getNode(ISD::AssertSext, Op.getValueType(), Op, 3858 DAG.getValueType(VT)); 3859 else if (Attrs && Attrs->paramHasAttr(Idx, ParamAttr::ZExt)) 3860 Op = DAG.getNode(ISD::AssertZext, Op.getValueType(), Op, 3861 DAG.getValueType(VT)); 3862 Op = DAG.getNode(ISD::TRUNCATE, VT, Op); 3863 } else { 3864 assert(MVT::isFloatingPoint(VT) && "Not int or FP?"); 3865 Op = DAG.getNode(ISD::FP_ROUND, VT, Op); 3866 } 3867 Ops.push_back(Op); 3868 break; 3869 } 3870 case Expand: 3871 if (VT != MVT::Vector) { 3872 // If this is a large integer or a floating point node that needs to be 3873 // expanded, it needs to be reassembled from small integers. Figure out 3874 // what the source elt type is and how many small integers it is. 3875 Ops.push_back(ExpandScalarFormalArgs(VT, Result, i, DAG, *this)); 3876 } else { 3877 // Otherwise, this is a vector type. We only support legal vectors 3878 // right now. 3879 const VectorType *PTy = cast<VectorType>(I->getType()); 3880 unsigned NumElems = PTy->getNumElements(); 3881 const Type *EltTy = PTy->getElementType(); 3882 3883 // Figure out if there is a Packed type corresponding to this Vector 3884 // type. If so, convert to the vector type. 3885 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems); 3886 if (TVT != MVT::Other && isTypeLegal(TVT)) { 3887 SDOperand N = SDOperand(Result, i++); 3888 // Handle copies from generic vectors to registers. 3889 N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N, 3890 DAG.getConstant(NumElems, MVT::i32), 3891 DAG.getValueType(getValueType(EltTy))); 3892 Ops.push_back(N); 3893 } else { 3894 assert(0 && "Don't support illegal by-val vector arguments yet!"); 3895 abort(); 3896 } 3897 } 3898 break; 3899 } 3900 } 3901 return Ops; 3902} 3903 3904 3905/// ExpandScalarCallArgs - Recursively expand call argument node by 3906/// bit_converting it or extract a pair of elements from the larger node. 3907static void ExpandScalarCallArgs(MVT::ValueType VT, SDOperand Arg, 3908 unsigned Flags, 3909 SmallVector<SDOperand, 32> &Ops, 3910 SelectionDAG &DAG, 3911 TargetLowering &TLI, 3912 bool isFirst = true) { 3913 3914 if (TLI.getTypeAction(VT) != TargetLowering::Expand) { 3915 // if it isn't first piece, alignment must be 1 3916 if (!isFirst) 3917 Flags = (Flags & (~ISD::ParamFlags::OrigAlignment)) | 3918 (1 << ISD::ParamFlags::OrigAlignmentOffs); 3919 Ops.push_back(Arg); 3920 Ops.push_back(DAG.getConstant(Flags, MVT::i32)); 3921 return; 3922 } 3923 3924 MVT::ValueType EVT = TLI.getTypeToTransformTo(VT); 3925 unsigned NumVals = MVT::getSizeInBits(VT) / MVT::getSizeInBits(EVT); 3926 if (NumVals == 1) { 3927 Arg = DAG.getNode(ISD::BIT_CONVERT, EVT, Arg); 3928 ExpandScalarCallArgs(EVT, Arg, Flags, Ops, DAG, TLI, isFirst); 3929 } else if (NumVals == 2) { 3930 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, EVT, Arg, 3931 DAG.getConstant(0, TLI.getPointerTy())); 3932 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, EVT, Arg, 3933 DAG.getConstant(1, TLI.getPointerTy())); 3934 if (!TLI.isLittleEndian()) 3935 std::swap(Lo, Hi); 3936 ExpandScalarCallArgs(EVT, Lo, Flags, Ops, DAG, TLI, isFirst); 3937 ExpandScalarCallArgs(EVT, Hi, Flags, Ops, DAG, TLI, false); 3938 } else { 3939 // Value scalarized into many values. Unimp for now. 3940 assert(0 && "Cannot expand i64 -> i16 yet!"); 3941 } 3942} 3943 3944/// TargetLowering::LowerCallTo - This is the default LowerCallTo 3945/// implementation, which just inserts an ISD::CALL node, which is later custom 3946/// lowered by the target to something concrete. FIXME: When all targets are 3947/// migrated to using ISD::CALL, this hook should be integrated into SDISel. 3948std::pair<SDOperand, SDOperand> 3949TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy, 3950 bool RetTyIsSigned, bool isVarArg, 3951 unsigned CallingConv, bool isTailCall, 3952 SDOperand Callee, 3953 ArgListTy &Args, SelectionDAG &DAG) { 3954 SmallVector<SDOperand, 32> Ops; 3955 Ops.push_back(Chain); // Op#0 - Chain 3956 Ops.push_back(DAG.getConstant(CallingConv, getPointerTy())); // Op#1 - CC 3957 Ops.push_back(DAG.getConstant(isVarArg, getPointerTy())); // Op#2 - VarArg 3958 Ops.push_back(DAG.getConstant(isTailCall, getPointerTy())); // Op#3 - Tail 3959 Ops.push_back(Callee); 3960 3961 // Handle all of the outgoing arguments. 3962 for (unsigned i = 0, e = Args.size(); i != e; ++i) { 3963 MVT::ValueType VT = getValueType(Args[i].Ty); 3964 SDOperand Op = Args[i].Node; 3965 unsigned Flags = ISD::ParamFlags::NoFlagSet; 3966 unsigned OriginalAlignment = 3967 getTargetData()->getABITypeAlignment(Args[i].Ty); 3968 3969 if (Args[i].isSExt) 3970 Flags |= ISD::ParamFlags::SExt; 3971 if (Args[i].isZExt) 3972 Flags |= ISD::ParamFlags::ZExt; 3973 if (Args[i].isInReg) 3974 Flags |= ISD::ParamFlags::InReg; 3975 if (Args[i].isSRet) 3976 Flags |= ISD::ParamFlags::StructReturn; 3977 Flags |= OriginalAlignment << ISD::ParamFlags::OrigAlignmentOffs; 3978 3979 switch (getTypeAction(VT)) { 3980 default: assert(0 && "Unknown type action!"); 3981 case Legal: 3982 Ops.push_back(Op); 3983 Ops.push_back(DAG.getConstant(Flags, MVT::i32)); 3984 break; 3985 case Promote: 3986 if (MVT::isInteger(VT)) { 3987 unsigned ExtOp; 3988 if (Args[i].isSExt) 3989 ExtOp = ISD::SIGN_EXTEND; 3990 else if (Args[i].isZExt) 3991 ExtOp = ISD::ZERO_EXTEND; 3992 else 3993 ExtOp = ISD::ANY_EXTEND; 3994 Op = DAG.getNode(ExtOp, getTypeToTransformTo(VT), Op); 3995 } else { 3996 assert(MVT::isFloatingPoint(VT) && "Not int or FP?"); 3997 Op = DAG.getNode(ISD::FP_EXTEND, getTypeToTransformTo(VT), Op); 3998 } 3999 Ops.push_back(Op); 4000 Ops.push_back(DAG.getConstant(Flags, MVT::i32)); 4001 break; 4002 case Expand: 4003 if (VT != MVT::Vector) { 4004 // If this is a large integer, it needs to be broken down into small 4005 // integers. Figure out what the source elt type is and how many small 4006 // integers it is. 4007 ExpandScalarCallArgs(VT, Op, Flags, Ops, DAG, *this); 4008 } else { 4009 // Otherwise, this is a vector type. We only support legal vectors 4010 // right now. 4011 const VectorType *PTy = cast<VectorType>(Args[i].Ty); 4012 unsigned NumElems = PTy->getNumElements(); 4013 const Type *EltTy = PTy->getElementType(); 4014 4015 // Figure out if there is a Packed type corresponding to this Vector 4016 // type. If so, convert to the vector type. 4017 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems); 4018 if (TVT != MVT::Other && isTypeLegal(TVT)) { 4019 // Insert a VBIT_CONVERT of the MVT::Vector type to the vector type. 4020 Op = DAG.getNode(ISD::VBIT_CONVERT, TVT, Op); 4021 Ops.push_back(Op); 4022 Ops.push_back(DAG.getConstant(Flags, MVT::i32)); 4023 } else { 4024 assert(0 && "Don't support illegal by-val vector call args yet!"); 4025 abort(); 4026 } 4027 } 4028 break; 4029 } 4030 } 4031 4032 // Figure out the result value types. 4033 SmallVector<MVT::ValueType, 4> RetTys; 4034 4035 if (RetTy != Type::VoidTy) { 4036 MVT::ValueType VT = getValueType(RetTy); 4037 switch (getTypeAction(VT)) { 4038 default: assert(0 && "Unknown type action!"); 4039 case Legal: 4040 RetTys.push_back(VT); 4041 break; 4042 case Promote: 4043 RetTys.push_back(getTypeToTransformTo(VT)); 4044 break; 4045 case Expand: 4046 if (VT != MVT::Vector) { 4047 // If this is a large integer, it needs to be reassembled from small 4048 // integers. Figure out what the source elt type is and how many small 4049 // integers it is. 4050 MVT::ValueType NVT = getTypeToExpandTo(VT); 4051 unsigned NumVals = getNumElements(VT); 4052 for (unsigned i = 0; i != NumVals; ++i) 4053 RetTys.push_back(NVT); 4054 } else { 4055 // Otherwise, this is a vector type. We only support legal vectors 4056 // right now. 4057 const VectorType *PTy = cast<VectorType>(RetTy); 4058 unsigned NumElems = PTy->getNumElements(); 4059 const Type *EltTy = PTy->getElementType(); 4060 4061 // Figure out if there is a Packed type corresponding to this Vector 4062 // type. If so, convert to the vector type. 4063 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems); 4064 if (TVT != MVT::Other && isTypeLegal(TVT)) { 4065 RetTys.push_back(TVT); 4066 } else { 4067 assert(0 && "Don't support illegal by-val vector call results yet!"); 4068 abort(); 4069 } 4070 } 4071 } 4072 } 4073 4074 RetTys.push_back(MVT::Other); // Always has a chain. 4075 4076 // Finally, create the CALL node. 4077 SDOperand Res = DAG.getNode(ISD::CALL, 4078 DAG.getVTList(&RetTys[0], RetTys.size()), 4079 &Ops[0], Ops.size()); 4080 4081 // This returns a pair of operands. The first element is the 4082 // return value for the function (if RetTy is not VoidTy). The second 4083 // element is the outgoing token chain. 4084 SDOperand ResVal; 4085 if (RetTys.size() != 1) { 4086 MVT::ValueType VT = getValueType(RetTy); 4087 if (RetTys.size() == 2) { 4088 ResVal = Res; 4089 4090 // If this value was promoted, truncate it down. 4091 if (ResVal.getValueType() != VT) { 4092 if (VT == MVT::Vector) { 4093 // Insert a VBIT_CONVERT to convert from the packed result type to the 4094 // MVT::Vector type. 4095 unsigned NumElems = cast<VectorType>(RetTy)->getNumElements(); 4096 const Type *EltTy = cast<VectorType>(RetTy)->getElementType(); 4097 4098 // Figure out if there is a Packed type corresponding to this Vector 4099 // type. If so, convert to the vector type. 4100 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy),NumElems); 4101 if (TVT != MVT::Other && isTypeLegal(TVT)) { 4102 // Insert a VBIT_CONVERT of the FORMAL_ARGUMENTS to a 4103 // "N x PTyElementVT" MVT::Vector type. 4104 ResVal = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, ResVal, 4105 DAG.getConstant(NumElems, MVT::i32), 4106 DAG.getValueType(getValueType(EltTy))); 4107 } else { 4108 abort(); 4109 } 4110 } else if (MVT::isInteger(VT)) { 4111 unsigned AssertOp = ISD::AssertSext; 4112 if (!RetTyIsSigned) 4113 AssertOp = ISD::AssertZext; 4114 ResVal = DAG.getNode(AssertOp, ResVal.getValueType(), ResVal, 4115 DAG.getValueType(VT)); 4116 ResVal = DAG.getNode(ISD::TRUNCATE, VT, ResVal); 4117 } else { 4118 assert(MVT::isFloatingPoint(VT)); 4119 if (getTypeAction(VT) == Expand) 4120 ResVal = DAG.getNode(ISD::BIT_CONVERT, VT, ResVal); 4121 else 4122 ResVal = DAG.getNode(ISD::FP_ROUND, VT, ResVal); 4123 } 4124 } 4125 } else if (RetTys.size() == 3) { 4126 ResVal = DAG.getNode(ISD::BUILD_PAIR, VT, 4127 Res.getValue(0), Res.getValue(1)); 4128 4129 } else { 4130 assert(0 && "Case not handled yet!"); 4131 } 4132 } 4133 4134 return std::make_pair(ResVal, Res.getValue(Res.Val->getNumValues()-1)); 4135} 4136 4137SDOperand TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) { 4138 assert(0 && "LowerOperation not implemented for this target!"); 4139 abort(); 4140 return SDOperand(); 4141} 4142 4143SDOperand TargetLowering::CustomPromoteOperation(SDOperand Op, 4144 SelectionDAG &DAG) { 4145 assert(0 && "CustomPromoteOperation not implemented for this target!"); 4146 abort(); 4147 return SDOperand(); 4148} 4149 4150/// getMemsetValue - Vectorized representation of the memset value 4151/// operand. 4152static SDOperand getMemsetValue(SDOperand Value, MVT::ValueType VT, 4153 SelectionDAG &DAG) { 4154 MVT::ValueType CurVT = VT; 4155 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) { 4156 uint64_t Val = C->getValue() & 255; 4157 unsigned Shift = 8; 4158 while (CurVT != MVT::i8) { 4159 Val = (Val << Shift) | Val; 4160 Shift <<= 1; 4161 CurVT = (MVT::ValueType)((unsigned)CurVT - 1); 4162 } 4163 return DAG.getConstant(Val, VT); 4164 } else { 4165 Value = DAG.getNode(ISD::ZERO_EXTEND, VT, Value); 4166 unsigned Shift = 8; 4167 while (CurVT != MVT::i8) { 4168 Value = 4169 DAG.getNode(ISD::OR, VT, 4170 DAG.getNode(ISD::SHL, VT, Value, 4171 DAG.getConstant(Shift, MVT::i8)), Value); 4172 Shift <<= 1; 4173 CurVT = (MVT::ValueType)((unsigned)CurVT - 1); 4174 } 4175 4176 return Value; 4177 } 4178} 4179 4180/// getMemsetStringVal - Similar to getMemsetValue. Except this is only 4181/// used when a memcpy is turned into a memset when the source is a constant 4182/// string ptr. 4183static SDOperand getMemsetStringVal(MVT::ValueType VT, 4184 SelectionDAG &DAG, TargetLowering &TLI, 4185 std::string &Str, unsigned Offset) { 4186 uint64_t Val = 0; 4187 unsigned MSB = getSizeInBits(VT) / 8; 4188 if (TLI.isLittleEndian()) 4189 Offset = Offset + MSB - 1; 4190 for (unsigned i = 0; i != MSB; ++i) { 4191 Val = (Val << 8) | (unsigned char)Str[Offset]; 4192 Offset += TLI.isLittleEndian() ? -1 : 1; 4193 } 4194 return DAG.getConstant(Val, VT); 4195} 4196 4197/// getMemBasePlusOffset - Returns base and offset node for the 4198static SDOperand getMemBasePlusOffset(SDOperand Base, unsigned Offset, 4199 SelectionDAG &DAG, TargetLowering &TLI) { 4200 MVT::ValueType VT = Base.getValueType(); 4201 return DAG.getNode(ISD::ADD, VT, Base, DAG.getConstant(Offset, VT)); 4202} 4203 4204/// MeetsMaxMemopRequirement - Determines if the number of memory ops required 4205/// to replace the memset / memcpy is below the threshold. It also returns the 4206/// types of the sequence of memory ops to perform memset / memcpy. 4207static bool MeetsMaxMemopRequirement(std::vector<MVT::ValueType> &MemOps, 4208 unsigned Limit, uint64_t Size, 4209 unsigned Align, TargetLowering &TLI) { 4210 MVT::ValueType VT; 4211 4212 if (TLI.allowsUnalignedMemoryAccesses()) { 4213 VT = MVT::i64; 4214 } else { 4215 switch (Align & 7) { 4216 case 0: 4217 VT = MVT::i64; 4218 break; 4219 case 4: 4220 VT = MVT::i32; 4221 break; 4222 case 2: 4223 VT = MVT::i16; 4224 break; 4225 default: 4226 VT = MVT::i8; 4227 break; 4228 } 4229 } 4230 4231 MVT::ValueType LVT = MVT::i64; 4232 while (!TLI.isTypeLegal(LVT)) 4233 LVT = (MVT::ValueType)((unsigned)LVT - 1); 4234 assert(MVT::isInteger(LVT)); 4235 4236 if (VT > LVT) 4237 VT = LVT; 4238 4239 unsigned NumMemOps = 0; 4240 while (Size != 0) { 4241 unsigned VTSize = getSizeInBits(VT) / 8; 4242 while (VTSize > Size) { 4243 VT = (MVT::ValueType)((unsigned)VT - 1); 4244 VTSize >>= 1; 4245 } 4246 assert(MVT::isInteger(VT)); 4247 4248 if (++NumMemOps > Limit) 4249 return false; 4250 MemOps.push_back(VT); 4251 Size -= VTSize; 4252 } 4253 4254 return true; 4255} 4256 4257void SelectionDAGLowering::visitMemIntrinsic(CallInst &I, unsigned Op) { 4258 SDOperand Op1 = getValue(I.getOperand(1)); 4259 SDOperand Op2 = getValue(I.getOperand(2)); 4260 SDOperand Op3 = getValue(I.getOperand(3)); 4261 SDOperand Op4 = getValue(I.getOperand(4)); 4262 unsigned Align = (unsigned)cast<ConstantSDNode>(Op4)->getValue(); 4263 if (Align == 0) Align = 1; 4264 4265 if (ConstantSDNode *Size = dyn_cast<ConstantSDNode>(Op3)) { 4266 std::vector<MVT::ValueType> MemOps; 4267 4268 // Expand memset / memcpy to a series of load / store ops 4269 // if the size operand falls below a certain threshold. 4270 SmallVector<SDOperand, 8> OutChains; 4271 switch (Op) { 4272 default: break; // Do nothing for now. 4273 case ISD::MEMSET: { 4274 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemset(), 4275 Size->getValue(), Align, TLI)) { 4276 unsigned NumMemOps = MemOps.size(); 4277 unsigned Offset = 0; 4278 for (unsigned i = 0; i < NumMemOps; i++) { 4279 MVT::ValueType VT = MemOps[i]; 4280 unsigned VTSize = getSizeInBits(VT) / 8; 4281 SDOperand Value = getMemsetValue(Op2, VT, DAG); 4282 SDOperand Store = DAG.getStore(getRoot(), Value, 4283 getMemBasePlusOffset(Op1, Offset, DAG, TLI), 4284 I.getOperand(1), Offset); 4285 OutChains.push_back(Store); 4286 Offset += VTSize; 4287 } 4288 } 4289 break; 4290 } 4291 case ISD::MEMCPY: { 4292 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemcpy(), 4293 Size->getValue(), Align, TLI)) { 4294 unsigned NumMemOps = MemOps.size(); 4295 unsigned SrcOff = 0, DstOff = 0, SrcDelta = 0; 4296 GlobalAddressSDNode *G = NULL; 4297 std::string Str; 4298 bool CopyFromStr = false; 4299 4300 if (Op2.getOpcode() == ISD::GlobalAddress) 4301 G = cast<GlobalAddressSDNode>(Op2); 4302 else if (Op2.getOpcode() == ISD::ADD && 4303 Op2.getOperand(0).getOpcode() == ISD::GlobalAddress && 4304 Op2.getOperand(1).getOpcode() == ISD::Constant) { 4305 G = cast<GlobalAddressSDNode>(Op2.getOperand(0)); 4306 SrcDelta = cast<ConstantSDNode>(Op2.getOperand(1))->getValue(); 4307 } 4308 if (G) { 4309 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal()); 4310 if (GV && GV->isConstant()) { 4311 Str = GV->getStringValue(false); 4312 if (!Str.empty()) { 4313 CopyFromStr = true; 4314 SrcOff += SrcDelta; 4315 } 4316 } 4317 } 4318 4319 for (unsigned i = 0; i < NumMemOps; i++) { 4320 MVT::ValueType VT = MemOps[i]; 4321 unsigned VTSize = getSizeInBits(VT) / 8; 4322 SDOperand Value, Chain, Store; 4323 4324 if (CopyFromStr) { 4325 Value = getMemsetStringVal(VT, DAG, TLI, Str, SrcOff); 4326 Chain = getRoot(); 4327 Store = 4328 DAG.getStore(Chain, Value, 4329 getMemBasePlusOffset(Op1, DstOff, DAG, TLI), 4330 I.getOperand(1), DstOff); 4331 } else { 4332 Value = DAG.getLoad(VT, getRoot(), 4333 getMemBasePlusOffset(Op2, SrcOff, DAG, TLI), 4334 I.getOperand(2), SrcOff); 4335 Chain = Value.getValue(1); 4336 Store = 4337 DAG.getStore(Chain, Value, 4338 getMemBasePlusOffset(Op1, DstOff, DAG, TLI), 4339 I.getOperand(1), DstOff); 4340 } 4341 OutChains.push_back(Store); 4342 SrcOff += VTSize; 4343 DstOff += VTSize; 4344 } 4345 } 4346 break; 4347 } 4348 } 4349 4350 if (!OutChains.empty()) { 4351 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, 4352 &OutChains[0], OutChains.size())); 4353 return; 4354 } 4355 } 4356 4357 DAG.setRoot(DAG.getNode(Op, MVT::Other, getRoot(), Op1, Op2, Op3, Op4)); 4358} 4359 4360//===----------------------------------------------------------------------===// 4361// SelectionDAGISel code 4362//===----------------------------------------------------------------------===// 4363 4364unsigned SelectionDAGISel::MakeReg(MVT::ValueType VT) { 4365 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT)); 4366} 4367 4368void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const { 4369 AU.addRequired<AliasAnalysis>(); 4370 AU.setPreservesAll(); 4371} 4372 4373 4374 4375bool SelectionDAGISel::runOnFunction(Function &Fn) { 4376 MachineFunction &MF = MachineFunction::construct(&Fn, TLI.getTargetMachine()); 4377 RegMap = MF.getSSARegMap(); 4378 DOUT << "\n\n\n=== " << Fn.getName() << "\n"; 4379 4380 FunctionLoweringInfo FuncInfo(TLI, Fn, MF); 4381 4382 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) 4383 SelectBasicBlock(I, MF, FuncInfo); 4384 4385 // Add function live-ins to entry block live-in set. 4386 BasicBlock *EntryBB = &Fn.getEntryBlock(); 4387 BB = FuncInfo.MBBMap[EntryBB]; 4388 if (!MF.livein_empty()) 4389 for (MachineFunction::livein_iterator I = MF.livein_begin(), 4390 E = MF.livein_end(); I != E; ++I) 4391 BB->addLiveIn(I->first); 4392 4393 return true; 4394} 4395 4396SDOperand SelectionDAGLowering::CopyValueToVirtualRegister(Value *V, 4397 unsigned Reg) { 4398 SDOperand Op = getValue(V); 4399 assert((Op.getOpcode() != ISD::CopyFromReg || 4400 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && 4401 "Copy from a reg to the same reg!"); 4402 4403 // If this type is not legal, we must make sure to not create an invalid 4404 // register use. 4405 MVT::ValueType SrcVT = Op.getValueType(); 4406 MVT::ValueType DestVT = TLI.getTypeToTransformTo(SrcVT); 4407 if (SrcVT == DestVT) { 4408 return DAG.getCopyToReg(getRoot(), Reg, Op); 4409 } else if (SrcVT == MVT::Vector) { 4410 // Handle copies from generic vectors to registers. 4411 MVT::ValueType PTyElementVT, PTyLegalElementVT; 4412 unsigned NE = TLI.getVectorTypeBreakdown(cast<VectorType>(V->getType()), 4413 PTyElementVT, PTyLegalElementVT); 4414 4415 // Insert a VBIT_CONVERT of the input vector to a "N x PTyElementVT" 4416 // MVT::Vector type. 4417 Op = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Op, 4418 DAG.getConstant(NE, MVT::i32), 4419 DAG.getValueType(PTyElementVT)); 4420 4421 // Loop over all of the elements of the resultant vector, 4422 // VEXTRACT_VECTOR_ELT'ing them, converting them to PTyLegalElementVT, then 4423 // copying them into output registers. 4424 SmallVector<SDOperand, 8> OutChains; 4425 SDOperand Root = getRoot(); 4426 for (unsigned i = 0; i != NE; ++i) { 4427 SDOperand Elt = DAG.getNode(ISD::VEXTRACT_VECTOR_ELT, PTyElementVT, 4428 Op, DAG.getConstant(i, TLI.getPointerTy())); 4429 if (PTyElementVT == PTyLegalElementVT) { 4430 // Elements are legal. 4431 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt)); 4432 } else if (PTyLegalElementVT > PTyElementVT) { 4433 // Elements are promoted. 4434 if (MVT::isFloatingPoint(PTyLegalElementVT)) 4435 Elt = DAG.getNode(ISD::FP_EXTEND, PTyLegalElementVT, Elt); 4436 else 4437 Elt = DAG.getNode(ISD::ANY_EXTEND, PTyLegalElementVT, Elt); 4438 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt)); 4439 } else { 4440 // Elements are expanded. 4441 // The src value is expanded into multiple registers. 4442 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT, 4443 Elt, DAG.getConstant(0, TLI.getPointerTy())); 4444 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT, 4445 Elt, DAG.getConstant(1, TLI.getPointerTy())); 4446 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Lo)); 4447 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Hi)); 4448 } 4449 } 4450 return DAG.getNode(ISD::TokenFactor, MVT::Other, 4451 &OutChains[0], OutChains.size()); 4452 } else if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) { 4453 // The src value is promoted to the register. 4454 if (MVT::isFloatingPoint(SrcVT)) 4455 Op = DAG.getNode(ISD::FP_EXTEND, DestVT, Op); 4456 else 4457 Op = DAG.getNode(ISD::ANY_EXTEND, DestVT, Op); 4458 return DAG.getCopyToReg(getRoot(), Reg, Op); 4459 } else { 4460 DestVT = TLI.getTypeToExpandTo(SrcVT); 4461 unsigned NumVals = TLI.getNumElements(SrcVT); 4462 if (NumVals == 1) 4463 return DAG.getCopyToReg(getRoot(), Reg, 4464 DAG.getNode(ISD::BIT_CONVERT, DestVT, Op)); 4465 assert(NumVals == 2 && "1 to 4 (and more) expansion not implemented!"); 4466 // The src value is expanded into multiple registers. 4467 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT, 4468 Op, DAG.getConstant(0, TLI.getPointerTy())); 4469 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT, 4470 Op, DAG.getConstant(1, TLI.getPointerTy())); 4471 Op = DAG.getCopyToReg(getRoot(), Reg, Lo); 4472 return DAG.getCopyToReg(Op, Reg+1, Hi); 4473 } 4474} 4475 4476void SelectionDAGISel:: 4477LowerArguments(BasicBlock *LLVMBB, SelectionDAGLowering &SDL, 4478 std::vector<SDOperand> &UnorderedChains) { 4479 // If this is the entry block, emit arguments. 4480 Function &F = *LLVMBB->getParent(); 4481 FunctionLoweringInfo &FuncInfo = SDL.FuncInfo; 4482 SDOperand OldRoot = SDL.DAG.getRoot(); 4483 std::vector<SDOperand> Args = TLI.LowerArguments(F, SDL.DAG); 4484 4485 unsigned a = 0; 4486 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); 4487 AI != E; ++AI, ++a) 4488 if (!AI->use_empty()) { 4489 SDL.setValue(AI, Args[a]); 4490 4491 // If this argument is live outside of the entry block, insert a copy from 4492 // whereever we got it to the vreg that other BB's will reference it as. 4493 DenseMap<const Value*, unsigned>::iterator VMI=FuncInfo.ValueMap.find(AI); 4494 if (VMI != FuncInfo.ValueMap.end()) { 4495 SDOperand Copy = SDL.CopyValueToVirtualRegister(AI, VMI->second); 4496 UnorderedChains.push_back(Copy); 4497 } 4498 } 4499 4500 // Finally, if the target has anything special to do, allow it to do so. 4501 // FIXME: this should insert code into the DAG! 4502 EmitFunctionEntryCode(F, SDL.DAG.getMachineFunction()); 4503} 4504 4505void SelectionDAGISel::BuildSelectionDAG(SelectionDAG &DAG, BasicBlock *LLVMBB, 4506 std::vector<std::pair<MachineInstr*, unsigned> > &PHINodesToUpdate, 4507 FunctionLoweringInfo &FuncInfo) { 4508 SelectionDAGLowering SDL(DAG, TLI, FuncInfo); 4509 4510 std::vector<SDOperand> UnorderedChains; 4511 4512 // Lower any arguments needed in this block if this is the entry block. 4513 if (LLVMBB == &LLVMBB->getParent()->getEntryBlock()) 4514 LowerArguments(LLVMBB, SDL, UnorderedChains); 4515 4516 BB = FuncInfo.MBBMap[LLVMBB]; 4517 SDL.setCurrentBasicBlock(BB); 4518 4519 // Lower all of the non-terminator instructions. 4520 for (BasicBlock::iterator I = LLVMBB->begin(), E = --LLVMBB->end(); 4521 I != E; ++I) 4522 SDL.visit(*I); 4523 4524 // Lower call part of invoke. 4525 InvokeInst *Invoke = dyn_cast<InvokeInst>(LLVMBB->getTerminator()); 4526 if (Invoke) SDL.visitInvoke(*Invoke, false); 4527 4528 // Ensure that all instructions which are used outside of their defining 4529 // blocks are available as virtual registers. 4530 for (BasicBlock::iterator I = LLVMBB->begin(), E = LLVMBB->end(); I != E;++I) 4531 if (!I->use_empty() && !isa<PHINode>(I)) { 4532 DenseMap<const Value*, unsigned>::iterator VMI =FuncInfo.ValueMap.find(I); 4533 if (VMI != FuncInfo.ValueMap.end()) 4534 UnorderedChains.push_back( 4535 SDL.CopyValueToVirtualRegister(I, VMI->second)); 4536 } 4537 4538 // Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to 4539 // ensure constants are generated when needed. Remember the virtual registers 4540 // that need to be added to the Machine PHI nodes as input. We cannot just 4541 // directly add them, because expansion might result in multiple MBB's for one 4542 // BB. As such, the start of the BB might correspond to a different MBB than 4543 // the end. 4544 // 4545 TerminatorInst *TI = LLVMBB->getTerminator(); 4546 4547 // Emit constants only once even if used by multiple PHI nodes. 4548 std::map<Constant*, unsigned> ConstantsOut; 4549 4550 // Vector bool would be better, but vector<bool> is really slow. 4551 std::vector<unsigned char> SuccsHandled; 4552 if (TI->getNumSuccessors()) 4553 SuccsHandled.resize(BB->getParent()->getNumBlockIDs()); 4554 4555 // Check successor nodes PHI nodes that expect a constant to be available from 4556 // this block. 4557 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { 4558 BasicBlock *SuccBB = TI->getSuccessor(succ); 4559 if (!isa<PHINode>(SuccBB->begin())) continue; 4560 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB]; 4561 4562 // If this terminator has multiple identical successors (common for 4563 // switches), only handle each succ once. 4564 unsigned SuccMBBNo = SuccMBB->getNumber(); 4565 if (SuccsHandled[SuccMBBNo]) continue; 4566 SuccsHandled[SuccMBBNo] = true; 4567 4568 MachineBasicBlock::iterator MBBI = SuccMBB->begin(); 4569 PHINode *PN; 4570 4571 // At this point we know that there is a 1-1 correspondence between LLVM PHI 4572 // nodes and Machine PHI nodes, but the incoming operands have not been 4573 // emitted yet. 4574 for (BasicBlock::iterator I = SuccBB->begin(); 4575 (PN = dyn_cast<PHINode>(I)); ++I) { 4576 // Ignore dead phi's. 4577 if (PN->use_empty()) continue; 4578 4579 unsigned Reg; 4580 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); 4581 4582 if (Constant *C = dyn_cast<Constant>(PHIOp)) { 4583 unsigned &RegOut = ConstantsOut[C]; 4584 if (RegOut == 0) { 4585 RegOut = FuncInfo.CreateRegForValue(C); 4586 UnorderedChains.push_back( 4587 SDL.CopyValueToVirtualRegister(C, RegOut)); 4588 } 4589 Reg = RegOut; 4590 } else { 4591 Reg = FuncInfo.ValueMap[PHIOp]; 4592 if (Reg == 0) { 4593 assert(isa<AllocaInst>(PHIOp) && 4594 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && 4595 "Didn't codegen value into a register!??"); 4596 Reg = FuncInfo.CreateRegForValue(PHIOp); 4597 UnorderedChains.push_back( 4598 SDL.CopyValueToVirtualRegister(PHIOp, Reg)); 4599 } 4600 } 4601 4602 // Remember that this register needs to added to the machine PHI node as 4603 // the input for this MBB. 4604 MVT::ValueType VT = TLI.getValueType(PN->getType()); 4605 unsigned NumElements; 4606 if (VT != MVT::Vector) 4607 NumElements = TLI.getNumElements(VT); 4608 else { 4609 MVT::ValueType VT1,VT2; 4610 NumElements = 4611 TLI.getVectorTypeBreakdown(cast<VectorType>(PN->getType()), 4612 VT1, VT2); 4613 } 4614 for (unsigned i = 0, e = NumElements; i != e; ++i) 4615 PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i)); 4616 } 4617 } 4618 ConstantsOut.clear(); 4619 4620 // Turn all of the unordered chains into one factored node. 4621 if (!UnorderedChains.empty()) { 4622 SDOperand Root = SDL.getRoot(); 4623 if (Root.getOpcode() != ISD::EntryToken) { 4624 unsigned i = 0, e = UnorderedChains.size(); 4625 for (; i != e; ++i) { 4626 assert(UnorderedChains[i].Val->getNumOperands() > 1); 4627 if (UnorderedChains[i].Val->getOperand(0) == Root) 4628 break; // Don't add the root if we already indirectly depend on it. 4629 } 4630 4631 if (i == e) 4632 UnorderedChains.push_back(Root); 4633 } 4634 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, 4635 &UnorderedChains[0], UnorderedChains.size())); 4636 } 4637 4638 // Lower the terminator after the copies are emitted. 4639 if (Invoke) { 4640 // Just the branch part of invoke. 4641 SDL.visitInvoke(*Invoke, true); 4642 } else { 4643 SDL.visit(*LLVMBB->getTerminator()); 4644 } 4645 4646 // Copy over any CaseBlock records that may now exist due to SwitchInst 4647 // lowering, as well as any jump table information. 4648 SwitchCases.clear(); 4649 SwitchCases = SDL.SwitchCases; 4650 JTCases.clear(); 4651 JTCases = SDL.JTCases; 4652 BitTestCases.clear(); 4653 BitTestCases = SDL.BitTestCases; 4654 4655 // Make sure the root of the DAG is up-to-date. 4656 DAG.setRoot(SDL.getRoot()); 4657} 4658 4659void SelectionDAGISel::CodeGenAndEmitDAG(SelectionDAG &DAG) { 4660 // Get alias analysis for load/store combining. 4661 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 4662 4663 // Run the DAG combiner in pre-legalize mode. 4664 DAG.Combine(false, AA); 4665 4666 DOUT << "Lowered selection DAG:\n"; 4667 DEBUG(DAG.dump()); 4668 4669 // Second step, hack on the DAG until it only uses operations and types that 4670 // the target supports. 4671 DAG.Legalize(); 4672 4673 DOUT << "Legalized selection DAG:\n"; 4674 DEBUG(DAG.dump()); 4675 4676 // Run the DAG combiner in post-legalize mode. 4677 DAG.Combine(true, AA); 4678 4679 if (ViewISelDAGs) DAG.viewGraph(); 4680 4681 // Third, instruction select all of the operations to machine code, adding the 4682 // code to the MachineBasicBlock. 4683 InstructionSelectBasicBlock(DAG); 4684 4685 DOUT << "Selected machine code:\n"; 4686 DEBUG(BB->dump()); 4687} 4688 4689void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB, MachineFunction &MF, 4690 FunctionLoweringInfo &FuncInfo) { 4691 std::vector<std::pair<MachineInstr*, unsigned> > PHINodesToUpdate; 4692 { 4693 SelectionDAG DAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>()); 4694 CurDAG = &DAG; 4695 4696 // First step, lower LLVM code to some DAG. This DAG may use operations and 4697 // types that are not supported by the target. 4698 BuildSelectionDAG(DAG, LLVMBB, PHINodesToUpdate, FuncInfo); 4699 4700 // Second step, emit the lowered DAG as machine code. 4701 CodeGenAndEmitDAG(DAG); 4702 } 4703 4704 DOUT << "Total amount of phi nodes to update: " 4705 << PHINodesToUpdate.size() << "\n"; 4706 DEBUG(for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) 4707 DOUT << "Node " << i << " : (" << PHINodesToUpdate[i].first 4708 << ", " << PHINodesToUpdate[i].second << ")\n";); 4709 4710 // Next, now that we know what the last MBB the LLVM BB expanded is, update 4711 // PHI nodes in successors. 4712 if (SwitchCases.empty() && JTCases.empty() && BitTestCases.empty()) { 4713 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) { 4714 MachineInstr *PHI = PHINodesToUpdate[i].first; 4715 assert(PHI->getOpcode() == TargetInstrInfo::PHI && 4716 "This is not a machine PHI node that we are updating!"); 4717 PHI->addRegOperand(PHINodesToUpdate[i].second, false); 4718 PHI->addMachineBasicBlockOperand(BB); 4719 } 4720 return; 4721 } 4722 4723 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i) { 4724 // Lower header first, if it wasn't already lowered 4725 if (!BitTestCases[i].Emitted) { 4726 SelectionDAG HSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>()); 4727 CurDAG = &HSDAG; 4728 SelectionDAGLowering HSDL(HSDAG, TLI, FuncInfo); 4729 // Set the current basic block to the mbb we wish to insert the code into 4730 BB = BitTestCases[i].Parent; 4731 HSDL.setCurrentBasicBlock(BB); 4732 // Emit the code 4733 HSDL.visitBitTestHeader(BitTestCases[i]); 4734 HSDAG.setRoot(HSDL.getRoot()); 4735 CodeGenAndEmitDAG(HSDAG); 4736 } 4737 4738 for (unsigned j = 0, ej = BitTestCases[i].Cases.size(); j != ej; ++j) { 4739 SelectionDAG BSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>()); 4740 CurDAG = &BSDAG; 4741 SelectionDAGLowering BSDL(BSDAG, TLI, FuncInfo); 4742 // Set the current basic block to the mbb we wish to insert the code into 4743 BB = BitTestCases[i].Cases[j].ThisBB; 4744 BSDL.setCurrentBasicBlock(BB); 4745 // Emit the code 4746 if (j+1 != ej) 4747 BSDL.visitBitTestCase(BitTestCases[i].Cases[j+1].ThisBB, 4748 BitTestCases[i].Reg, 4749 BitTestCases[i].Cases[j]); 4750 else 4751 BSDL.visitBitTestCase(BitTestCases[i].Default, 4752 BitTestCases[i].Reg, 4753 BitTestCases[i].Cases[j]); 4754 4755 4756 BSDAG.setRoot(BSDL.getRoot()); 4757 CodeGenAndEmitDAG(BSDAG); 4758 } 4759 4760 // Update PHI Nodes 4761 for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) { 4762 MachineInstr *PHI = PHINodesToUpdate[pi].first; 4763 MachineBasicBlock *PHIBB = PHI->getParent(); 4764 assert(PHI->getOpcode() == TargetInstrInfo::PHI && 4765 "This is not a machine PHI node that we are updating!"); 4766 // This is "default" BB. We have two jumps to it. From "header" BB and 4767 // from last "case" BB. 4768 if (PHIBB == BitTestCases[i].Default) { 4769 PHI->addRegOperand(PHINodesToUpdate[pi].second, false); 4770 PHI->addMachineBasicBlockOperand(BitTestCases[i].Parent); 4771 PHI->addRegOperand(PHINodesToUpdate[pi].second, false); 4772 PHI->addMachineBasicBlockOperand(BitTestCases[i].Cases.back().ThisBB); 4773 } 4774 // One of "cases" BB. 4775 for (unsigned j = 0, ej = BitTestCases[i].Cases.size(); j != ej; ++j) { 4776 MachineBasicBlock* cBB = BitTestCases[i].Cases[j].ThisBB; 4777 if (cBB->succ_end() != 4778 std::find(cBB->succ_begin(),cBB->succ_end(), PHIBB)) { 4779 PHI->addRegOperand(PHINodesToUpdate[pi].second, false); 4780 PHI->addMachineBasicBlockOperand(cBB); 4781 } 4782 } 4783 } 4784 } 4785 4786 // If the JumpTable record is filled in, then we need to emit a jump table. 4787 // Updating the PHI nodes is tricky in this case, since we need to determine 4788 // whether the PHI is a successor of the range check MBB or the jump table MBB 4789 for (unsigned i = 0, e = JTCases.size(); i != e; ++i) { 4790 // Lower header first, if it wasn't already lowered 4791 if (!JTCases[i].first.Emitted) { 4792 SelectionDAG HSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>()); 4793 CurDAG = &HSDAG; 4794 SelectionDAGLowering HSDL(HSDAG, TLI, FuncInfo); 4795 // Set the current basic block to the mbb we wish to insert the code into 4796 BB = JTCases[i].first.HeaderBB; 4797 HSDL.setCurrentBasicBlock(BB); 4798 // Emit the code 4799 HSDL.visitJumpTableHeader(JTCases[i].second, JTCases[i].first); 4800 HSDAG.setRoot(HSDL.getRoot()); 4801 CodeGenAndEmitDAG(HSDAG); 4802 } 4803 4804 SelectionDAG JSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>()); 4805 CurDAG = &JSDAG; 4806 SelectionDAGLowering JSDL(JSDAG, TLI, FuncInfo); 4807 // Set the current basic block to the mbb we wish to insert the code into 4808 BB = JTCases[i].second.MBB; 4809 JSDL.setCurrentBasicBlock(BB); 4810 // Emit the code 4811 JSDL.visitJumpTable(JTCases[i].second); 4812 JSDAG.setRoot(JSDL.getRoot()); 4813 CodeGenAndEmitDAG(JSDAG); 4814 4815 // Update PHI Nodes 4816 for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) { 4817 MachineInstr *PHI = PHINodesToUpdate[pi].first; 4818 MachineBasicBlock *PHIBB = PHI->getParent(); 4819 assert(PHI->getOpcode() == TargetInstrInfo::PHI && 4820 "This is not a machine PHI node that we are updating!"); 4821 // "default" BB. We can go there only from header BB. 4822 if (PHIBB == JTCases[i].second.Default) { 4823 PHI->addRegOperand(PHINodesToUpdate[pi].second, false); 4824 PHI->addMachineBasicBlockOperand(JTCases[i].first.HeaderBB); 4825 } 4826 // JT BB. Just iterate over successors here 4827 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) { 4828 PHI->addRegOperand(PHINodesToUpdate[pi].second, false); 4829 PHI->addMachineBasicBlockOperand(BB); 4830 } 4831 } 4832 } 4833 4834 // If the switch block involved a branch to one of the actual successors, we 4835 // need to update PHI nodes in that block. 4836 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) { 4837 MachineInstr *PHI = PHINodesToUpdate[i].first; 4838 assert(PHI->getOpcode() == TargetInstrInfo::PHI && 4839 "This is not a machine PHI node that we are updating!"); 4840 if (BB->isSuccessor(PHI->getParent())) { 4841 PHI->addRegOperand(PHINodesToUpdate[i].second, false); 4842 PHI->addMachineBasicBlockOperand(BB); 4843 } 4844 } 4845 4846 // If we generated any switch lowering information, build and codegen any 4847 // additional DAGs necessary. 4848 for (unsigned i = 0, e = SwitchCases.size(); i != e; ++i) { 4849 SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>()); 4850 CurDAG = &SDAG; 4851 SelectionDAGLowering SDL(SDAG, TLI, FuncInfo); 4852 4853 // Set the current basic block to the mbb we wish to insert the code into 4854 BB = SwitchCases[i].ThisBB; 4855 SDL.setCurrentBasicBlock(BB); 4856 4857 // Emit the code 4858 SDL.visitSwitchCase(SwitchCases[i]); 4859 SDAG.setRoot(SDL.getRoot()); 4860 CodeGenAndEmitDAG(SDAG); 4861 4862 // Handle any PHI nodes in successors of this chunk, as if we were coming 4863 // from the original BB before switch expansion. Note that PHI nodes can 4864 // occur multiple times in PHINodesToUpdate. We have to be very careful to 4865 // handle them the right number of times. 4866 while ((BB = SwitchCases[i].TrueBB)) { // Handle LHS and RHS. 4867 for (MachineBasicBlock::iterator Phi = BB->begin(); 4868 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){ 4869 // This value for this PHI node is recorded in PHINodesToUpdate, get it. 4870 for (unsigned pn = 0; ; ++pn) { 4871 assert(pn != PHINodesToUpdate.size() && "Didn't find PHI entry!"); 4872 if (PHINodesToUpdate[pn].first == Phi) { 4873 Phi->addRegOperand(PHINodesToUpdate[pn].second, false); 4874 Phi->addMachineBasicBlockOperand(SwitchCases[i].ThisBB); 4875 break; 4876 } 4877 } 4878 } 4879 4880 // Don't process RHS if same block as LHS. 4881 if (BB == SwitchCases[i].FalseBB) 4882 SwitchCases[i].FalseBB = 0; 4883 4884 // If we haven't handled the RHS, do so now. Otherwise, we're done. 4885 SwitchCases[i].TrueBB = SwitchCases[i].FalseBB; 4886 SwitchCases[i].FalseBB = 0; 4887 } 4888 assert(SwitchCases[i].TrueBB == 0 && SwitchCases[i].FalseBB == 0); 4889 } 4890} 4891 4892 4893//===----------------------------------------------------------------------===// 4894/// ScheduleAndEmitDAG - Pick a safe ordering and emit instructions for each 4895/// target node in the graph. 4896void SelectionDAGISel::ScheduleAndEmitDAG(SelectionDAG &DAG) { 4897 if (ViewSchedDAGs) DAG.viewGraph(); 4898 4899 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault(); 4900 4901 if (!Ctor) { 4902 Ctor = ISHeuristic; 4903 RegisterScheduler::setDefault(Ctor); 4904 } 4905 4906 ScheduleDAG *SL = Ctor(this, &DAG, BB); 4907 BB = SL->Run(); 4908 delete SL; 4909} 4910 4911 4912HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() { 4913 return new HazardRecognizer(); 4914} 4915 4916//===----------------------------------------------------------------------===// 4917// Helper functions used by the generated instruction selector. 4918//===----------------------------------------------------------------------===// 4919// Calls to these methods are generated by tblgen. 4920 4921/// CheckAndMask - The isel is trying to match something like (and X, 255). If 4922/// the dag combiner simplified the 255, we still want to match. RHS is the 4923/// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value 4924/// specified in the .td file (e.g. 255). 4925bool SelectionDAGISel::CheckAndMask(SDOperand LHS, ConstantSDNode *RHS, 4926 int64_t DesiredMaskS) { 4927 uint64_t ActualMask = RHS->getValue(); 4928 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType()); 4929 4930 // If the actual mask exactly matches, success! 4931 if (ActualMask == DesiredMask) 4932 return true; 4933 4934 // If the actual AND mask is allowing unallowed bits, this doesn't match. 4935 if (ActualMask & ~DesiredMask) 4936 return false; 4937 4938 // Otherwise, the DAG Combiner may have proven that the value coming in is 4939 // either already zero or is not demanded. Check for known zero input bits. 4940 uint64_t NeededMask = DesiredMask & ~ActualMask; 4941 if (getTargetLowering().MaskedValueIsZero(LHS, NeededMask)) 4942 return true; 4943 4944 // TODO: check to see if missing bits are just not demanded. 4945 4946 // Otherwise, this pattern doesn't match. 4947 return false; 4948} 4949 4950/// CheckOrMask - The isel is trying to match something like (or X, 255). If 4951/// the dag combiner simplified the 255, we still want to match. RHS is the 4952/// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value 4953/// specified in the .td file (e.g. 255). 4954bool SelectionDAGISel::CheckOrMask(SDOperand LHS, ConstantSDNode *RHS, 4955 int64_t DesiredMaskS) { 4956 uint64_t ActualMask = RHS->getValue(); 4957 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType()); 4958 4959 // If the actual mask exactly matches, success! 4960 if (ActualMask == DesiredMask) 4961 return true; 4962 4963 // If the actual AND mask is allowing unallowed bits, this doesn't match. 4964 if (ActualMask & ~DesiredMask) 4965 return false; 4966 4967 // Otherwise, the DAG Combiner may have proven that the value coming in is 4968 // either already zero or is not demanded. Check for known zero input bits. 4969 uint64_t NeededMask = DesiredMask & ~ActualMask; 4970 4971 uint64_t KnownZero, KnownOne; 4972 getTargetLowering().ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne); 4973 4974 // If all the missing bits in the or are already known to be set, match! 4975 if ((NeededMask & KnownOne) == NeededMask) 4976 return true; 4977 4978 // TODO: check to see if missing bits are just not demanded. 4979 4980 // Otherwise, this pattern doesn't match. 4981 return false; 4982} 4983 4984 4985/// SelectInlineAsmMemoryOperands - Calls to this are automatically generated 4986/// by tblgen. Others should not call it. 4987void SelectionDAGISel:: 4988SelectInlineAsmMemoryOperands(std::vector<SDOperand> &Ops, SelectionDAG &DAG) { 4989 std::vector<SDOperand> InOps; 4990 std::swap(InOps, Ops); 4991 4992 Ops.push_back(InOps[0]); // input chain. 4993 Ops.push_back(InOps[1]); // input asm string. 4994 4995 unsigned i = 2, e = InOps.size(); 4996 if (InOps[e-1].getValueType() == MVT::Flag) 4997 --e; // Don't process a flag operand if it is here. 4998 4999 while (i != e) { 5000 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue(); 5001 if ((Flags & 7) != 4 /*MEM*/) { 5002 // Just skip over this operand, copying the operands verbatim. 5003 Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1); 5004 i += (Flags >> 3) + 1; 5005 } else { 5006 assert((Flags >> 3) == 1 && "Memory operand with multiple values?"); 5007 // Otherwise, this is a memory operand. Ask the target to select it. 5008 std::vector<SDOperand> SelOps; 5009 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps, DAG)) { 5010 cerr << "Could not match memory address. Inline asm failure!\n"; 5011 exit(1); 5012 } 5013 5014 // Add this to the output node. 5015 MVT::ValueType IntPtrTy = DAG.getTargetLoweringInfo().getPointerTy(); 5016 Ops.push_back(DAG.getTargetConstant(4/*MEM*/ | (SelOps.size() << 3), 5017 IntPtrTy)); 5018 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end()); 5019 i += 2; 5020 } 5021 } 5022 5023 // Add the flag input back if present. 5024 if (e != InOps.size()) 5025 Ops.push_back(InOps.back()); 5026} 5027 5028char SelectionDAGISel::ID = 0; 5029