SelectionDAGISel.cpp revision 47857812e29324a9d1560796a05b53d3a9217fd9
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/Analysis/AliasAnalysis.h" 16#include "llvm/CodeGen/SelectionDAGISel.h" 17#include "llvm/CodeGen/ScheduleDAG.h" 18#include "llvm/CallingConv.h" 19#include "llvm/Constants.h" 20#include "llvm/DerivedTypes.h" 21#include "llvm/Function.h" 22#include "llvm/GlobalVariable.h" 23#include "llvm/InlineAsm.h" 24#include "llvm/Instructions.h" 25#include "llvm/Intrinsics.h" 26#include "llvm/IntrinsicInst.h" 27#include "llvm/CodeGen/MachineDebugInfo.h" 28#include "llvm/CodeGen/MachineFunction.h" 29#include "llvm/CodeGen/MachineFrameInfo.h" 30#include "llvm/CodeGen/MachineJumpTableInfo.h" 31#include "llvm/CodeGen/MachineInstrBuilder.h" 32#include "llvm/CodeGen/SchedulerRegistry.h" 33#include "llvm/CodeGen/SelectionDAG.h" 34#include "llvm/CodeGen/SSARegMap.h" 35#include "llvm/Target/MRegisterInfo.h" 36#include "llvm/Target/TargetAsmInfo.h" 37#include "llvm/Target/TargetData.h" 38#include "llvm/Target/TargetFrameInfo.h" 39#include "llvm/Target/TargetInstrInfo.h" 40#include "llvm/Target/TargetLowering.h" 41#include "llvm/Target/TargetMachine.h" 42#include "llvm/Target/TargetOptions.h" 43#include "llvm/Transforms/Utils/BasicBlockUtils.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/// 64/// RegisterScheduler class - Track the registration of instruction schedulers. 65/// 66//===---------------------------------------------------------------------===// 67MachinePassRegistry RegisterScheduler::Registry; 68 69//===---------------------------------------------------------------------===// 70/// 71/// ISHeuristic command line option for instruction schedulers. 72/// 73//===---------------------------------------------------------------------===// 74namespace { 75 cl::opt<RegisterScheduler::FunctionPassCtor, false, 76 RegisterPassParser<RegisterScheduler> > 77 ISHeuristic("sched", 78 cl::init(&createDefaultScheduler), 79 cl::desc("Instruction schedulers available:")); 80 81 static RegisterScheduler 82 defaultListDAGScheduler("default", " Best scheduler for the target", 83 createDefaultScheduler); 84} // namespace 85 86namespace { 87 /// RegsForValue - This struct represents the physical registers that a 88 /// particular value is assigned and the type information about the value. 89 /// This is needed because values can be promoted into larger registers and 90 /// expanded into multiple smaller registers than the value. 91 struct VISIBILITY_HIDDEN RegsForValue { 92 /// Regs - This list hold the register (for legal and promoted values) 93 /// or register set (for expanded values) that the value should be assigned 94 /// to. 95 std::vector<unsigned> Regs; 96 97 /// RegVT - The value type of each register. 98 /// 99 MVT::ValueType RegVT; 100 101 /// ValueVT - The value type of the LLVM value, which may be promoted from 102 /// RegVT or made from merging the two expanded parts. 103 MVT::ValueType ValueVT; 104 105 RegsForValue() : RegVT(MVT::Other), ValueVT(MVT::Other) {} 106 107 RegsForValue(unsigned Reg, MVT::ValueType regvt, MVT::ValueType valuevt) 108 : RegVT(regvt), ValueVT(valuevt) { 109 Regs.push_back(Reg); 110 } 111 RegsForValue(const std::vector<unsigned> ®s, 112 MVT::ValueType regvt, MVT::ValueType valuevt) 113 : Regs(regs), RegVT(regvt), ValueVT(valuevt) { 114 } 115 116 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from 117 /// this value and returns the result as a ValueVT value. This uses 118 /// Chain/Flag as the input and updates them for the output Chain/Flag. 119 SDOperand getCopyFromRegs(SelectionDAG &DAG, 120 SDOperand &Chain, SDOperand &Flag) const; 121 122 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 123 /// specified value into the registers specified by this object. This uses 124 /// Chain/Flag as the input and updates them for the output Chain/Flag. 125 void getCopyToRegs(SDOperand Val, SelectionDAG &DAG, 126 SDOperand &Chain, SDOperand &Flag, 127 MVT::ValueType PtrVT) const; 128 129 /// AddInlineAsmOperands - Add this value to the specified inlineasm node 130 /// operand list. This adds the code marker and includes the number of 131 /// values added into it. 132 void AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG, 133 std::vector<SDOperand> &Ops) const; 134 }; 135} 136 137namespace llvm { 138 //===--------------------------------------------------------------------===// 139 /// createDefaultScheduler - This creates an instruction scheduler appropriate 140 /// for the target. 141 ScheduleDAG* createDefaultScheduler(SelectionDAGISel *IS, 142 SelectionDAG *DAG, 143 MachineBasicBlock *BB) { 144 TargetLowering &TLI = IS->getTargetLowering(); 145 146 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) { 147 return createTDListDAGScheduler(IS, DAG, BB); 148 } else { 149 assert(TLI.getSchedulingPreference() == 150 TargetLowering::SchedulingForRegPressure && "Unknown sched type!"); 151 return createBURRListDAGScheduler(IS, DAG, BB); 152 } 153 } 154 155 156 //===--------------------------------------------------------------------===// 157 /// FunctionLoweringInfo - This contains information that is global to a 158 /// function that is used when lowering a region of the function. 159 class FunctionLoweringInfo { 160 public: 161 TargetLowering &TLI; 162 Function &Fn; 163 MachineFunction &MF; 164 SSARegMap *RegMap; 165 166 FunctionLoweringInfo(TargetLowering &TLI, Function &Fn,MachineFunction &MF); 167 168 /// MBBMap - A mapping from LLVM basic blocks to their machine code entry. 169 std::map<const BasicBlock*, MachineBasicBlock *> MBBMap; 170 171 /// ValueMap - Since we emit code for the function a basic block at a time, 172 /// we must remember which virtual registers hold the values for 173 /// cross-basic-block values. 174 std::map<const Value*, unsigned> ValueMap; 175 176 /// StaticAllocaMap - Keep track of frame indices for fixed sized allocas in 177 /// the entry block. This allows the allocas to be efficiently referenced 178 /// anywhere in the function. 179 std::map<const AllocaInst*, int> StaticAllocaMap; 180 181 unsigned MakeReg(MVT::ValueType VT) { 182 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT)); 183 } 184 185 /// isExportedInst - Return true if the specified value is an instruction 186 /// exported from its block. 187 bool isExportedInst(const Value *V) { 188 return ValueMap.count(V); 189 } 190 191 unsigned CreateRegForValue(const Value *V); 192 193 unsigned InitializeRegForValue(const Value *V) { 194 unsigned &R = ValueMap[V]; 195 assert(R == 0 && "Already initialized this value register!"); 196 return R = CreateRegForValue(V); 197 } 198 }; 199} 200 201/// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by 202/// PHI nodes or outside of the basic block that defines it, or used by a 203/// switch instruction, which may expand to multiple basic blocks. 204static bool isUsedOutsideOfDefiningBlock(Instruction *I) { 205 if (isa<PHINode>(I)) return true; 206 BasicBlock *BB = I->getParent(); 207 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI) 208 if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI) || 209 // FIXME: Remove switchinst special case. 210 isa<SwitchInst>(*UI)) 211 return true; 212 return false; 213} 214 215/// isOnlyUsedInEntryBlock - If the specified argument is only used in the 216/// entry block, return true. This includes arguments used by switches, since 217/// the switch may expand into multiple basic blocks. 218static bool isOnlyUsedInEntryBlock(Argument *A) { 219 BasicBlock *Entry = A->getParent()->begin(); 220 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI) 221 if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI)) 222 return false; // Use not in entry block. 223 return true; 224} 225 226FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli, 227 Function &fn, MachineFunction &mf) 228 : TLI(tli), Fn(fn), MF(mf), RegMap(MF.getSSARegMap()) { 229 230 // Create a vreg for each argument register that is not dead and is used 231 // outside of the entry block for the function. 232 for (Function::arg_iterator AI = Fn.arg_begin(), E = Fn.arg_end(); 233 AI != E; ++AI) 234 if (!isOnlyUsedInEntryBlock(AI)) 235 InitializeRegForValue(AI); 236 237 // Initialize the mapping of values to registers. This is only set up for 238 // instruction values that are used outside of the block that defines 239 // them. 240 Function::iterator BB = Fn.begin(), EB = Fn.end(); 241 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 242 if (AllocaInst *AI = dyn_cast<AllocaInst>(I)) 243 if (ConstantInt *CUI = dyn_cast<ConstantInt>(AI->getArraySize())) { 244 const Type *Ty = AI->getAllocatedType(); 245 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty); 246 unsigned Align = 247 std::max((unsigned)TLI.getTargetData()->getTypeAlignment(Ty), 248 AI->getAlignment()); 249 250 // If the alignment of the value is smaller than the size of the 251 // value, and if the size of the value is particularly small 252 // (<= 8 bytes), round up to the size of the value for potentially 253 // better performance. 254 // 255 // FIXME: This could be made better with a preferred alignment hook in 256 // TargetData. It serves primarily to 8-byte align doubles for X86. 257 if (Align < TySize && TySize <= 8) Align = TySize; 258 TySize *= CUI->getZExtValue(); // Get total allocated size. 259 if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects. 260 StaticAllocaMap[AI] = 261 MF.getFrameInfo()->CreateStackObject((unsigned)TySize, Align); 262 } 263 264 for (; BB != EB; ++BB) 265 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 266 if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I)) 267 if (!isa<AllocaInst>(I) || 268 !StaticAllocaMap.count(cast<AllocaInst>(I))) 269 InitializeRegForValue(I); 270 271 // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This 272 // also creates the initial PHI MachineInstrs, though none of the input 273 // operands are populated. 274 for (BB = Fn.begin(), EB = Fn.end(); BB != EB; ++BB) { 275 MachineBasicBlock *MBB = new MachineBasicBlock(BB); 276 MBBMap[BB] = MBB; 277 MF.getBasicBlockList().push_back(MBB); 278 279 // Create Machine PHI nodes for LLVM PHI nodes, lowering them as 280 // appropriate. 281 PHINode *PN; 282 for (BasicBlock::iterator I = BB->begin();(PN = dyn_cast<PHINode>(I)); ++I){ 283 if (PN->use_empty()) continue; 284 285 MVT::ValueType VT = TLI.getValueType(PN->getType()); 286 unsigned NumElements; 287 if (VT != MVT::Vector) 288 NumElements = TLI.getNumElements(VT); 289 else { 290 MVT::ValueType VT1,VT2; 291 NumElements = 292 TLI.getPackedTypeBreakdown(cast<PackedType>(PN->getType()), 293 VT1, VT2); 294 } 295 unsigned PHIReg = ValueMap[PN]; 296 assert(PHIReg && "PHI node does not have an assigned virtual register!"); 297 const TargetInstrInfo *TII = TLI.getTargetMachine().getInstrInfo(); 298 for (unsigned i = 0; i != NumElements; ++i) 299 BuildMI(MBB, TII->get(TargetInstrInfo::PHI), PHIReg+i); 300 } 301 } 302} 303 304/// CreateRegForValue - Allocate the appropriate number of virtual registers of 305/// the correctly promoted or expanded types. Assign these registers 306/// consecutive vreg numbers and return the first assigned number. 307unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) { 308 MVT::ValueType VT = TLI.getValueType(V->getType()); 309 310 // The number of multiples of registers that we need, to, e.g., split up 311 // a <2 x int64> -> 4 x i32 registers. 312 unsigned NumVectorRegs = 1; 313 314 // If this is a packed type, figure out what type it will decompose into 315 // and how many of the elements it will use. 316 if (VT == MVT::Vector) { 317 const PackedType *PTy = cast<PackedType>(V->getType()); 318 unsigned NumElts = PTy->getNumElements(); 319 MVT::ValueType EltTy = TLI.getValueType(PTy->getElementType()); 320 321 // Divide the input until we get to a supported size. This will always 322 // end with a scalar if the target doesn't support vectors. 323 while (NumElts > 1 && !TLI.isTypeLegal(getVectorType(EltTy, NumElts))) { 324 NumElts >>= 1; 325 NumVectorRegs <<= 1; 326 } 327 if (NumElts == 1) 328 VT = EltTy; 329 else 330 VT = getVectorType(EltTy, NumElts); 331 } 332 333 // The common case is that we will only create one register for this 334 // value. If we have that case, create and return the virtual register. 335 unsigned NV = TLI.getNumElements(VT); 336 if (NV == 1) { 337 // If we are promoting this value, pick the next largest supported type. 338 MVT::ValueType PromotedType = TLI.getTypeToTransformTo(VT); 339 unsigned Reg = MakeReg(PromotedType); 340 // If this is a vector of supported or promoted types (e.g. 4 x i16), 341 // create all of the registers. 342 for (unsigned i = 1; i != NumVectorRegs; ++i) 343 MakeReg(PromotedType); 344 return Reg; 345 } 346 347 // If this value is represented with multiple target registers, make sure 348 // to create enough consecutive registers of the right (smaller) type. 349 VT = TLI.getTypeToExpandTo(VT); 350 unsigned R = MakeReg(VT); 351 for (unsigned i = 1; i != NV*NumVectorRegs; ++i) 352 MakeReg(VT); 353 return R; 354} 355 356//===----------------------------------------------------------------------===// 357/// SelectionDAGLowering - This is the common target-independent lowering 358/// implementation that is parameterized by a TargetLowering object. 359/// Also, targets can overload any lowering method. 360/// 361namespace llvm { 362class SelectionDAGLowering { 363 MachineBasicBlock *CurMBB; 364 365 std::map<const Value*, SDOperand> NodeMap; 366 367 /// PendingLoads - Loads are not emitted to the program immediately. We bunch 368 /// them up and then emit token factor nodes when possible. This allows us to 369 /// get simple disambiguation between loads without worrying about alias 370 /// analysis. 371 std::vector<SDOperand> PendingLoads; 372 373 /// Case - A pair of values to record the Value for a switch case, and the 374 /// case's target basic block. 375 typedef std::pair<Constant*, MachineBasicBlock*> Case; 376 typedef std::vector<Case>::iterator CaseItr; 377 typedef std::pair<CaseItr, CaseItr> CaseRange; 378 379 /// CaseRec - A struct with ctor used in lowering switches to a binary tree 380 /// of conditional branches. 381 struct CaseRec { 382 CaseRec(MachineBasicBlock *bb, Constant *lt, Constant *ge, CaseRange r) : 383 CaseBB(bb), LT(lt), GE(ge), Range(r) {} 384 385 /// CaseBB - The MBB in which to emit the compare and branch 386 MachineBasicBlock *CaseBB; 387 /// LT, GE - If nonzero, we know the current case value must be less-than or 388 /// greater-than-or-equal-to these Constants. 389 Constant *LT; 390 Constant *GE; 391 /// Range - A pair of iterators representing the range of case values to be 392 /// processed at this point in the binary search tree. 393 CaseRange Range; 394 }; 395 396 /// The comparison function for sorting Case values. 397 struct CaseCmp { 398 bool operator () (const Case& C1, const Case& C2) { 399 assert(isa<ConstantInt>(C1.first) && isa<ConstantInt>(C2.first)); 400 return cast<const ConstantInt>(C1.first)->getZExtValue() < 401 cast<const ConstantInt>(C2.first)->getZExtValue(); 402 } 403 }; 404 405public: 406 // TLI - This is information that describes the available target features we 407 // need for lowering. This indicates when operations are unavailable, 408 // implemented with a libcall, etc. 409 TargetLowering &TLI; 410 SelectionDAG &DAG; 411 const TargetData *TD; 412 413 /// SwitchCases - Vector of CaseBlock structures used to communicate 414 /// SwitchInst code generation information. 415 std::vector<SelectionDAGISel::CaseBlock> SwitchCases; 416 SelectionDAGISel::JumpTable JT; 417 418 /// FuncInfo - Information about the function as a whole. 419 /// 420 FunctionLoweringInfo &FuncInfo; 421 422 SelectionDAGLowering(SelectionDAG &dag, TargetLowering &tli, 423 FunctionLoweringInfo &funcinfo) 424 : TLI(tli), DAG(dag), TD(DAG.getTarget().getTargetData()), 425 JT(0,0,0,0), FuncInfo(funcinfo) { 426 } 427 428 /// getRoot - Return the current virtual root of the Selection DAG. 429 /// 430 SDOperand getRoot() { 431 if (PendingLoads.empty()) 432 return DAG.getRoot(); 433 434 if (PendingLoads.size() == 1) { 435 SDOperand Root = PendingLoads[0]; 436 DAG.setRoot(Root); 437 PendingLoads.clear(); 438 return Root; 439 } 440 441 // Otherwise, we have to make a token factor node. 442 SDOperand Root = DAG.getNode(ISD::TokenFactor, MVT::Other, 443 &PendingLoads[0], PendingLoads.size()); 444 PendingLoads.clear(); 445 DAG.setRoot(Root); 446 return Root; 447 } 448 449 SDOperand CopyValueToVirtualRegister(Value *V, unsigned Reg); 450 451 void visit(Instruction &I) { visit(I.getOpcode(), I); } 452 453 void visit(unsigned Opcode, User &I) { 454 // Note: this doesn't use InstVisitor, because it has to work with 455 // ConstantExpr's in addition to instructions. 456 switch (Opcode) { 457 default: assert(0 && "Unknown instruction type encountered!"); 458 abort(); 459 // Build the switch statement using the Instruction.def file. 460#define HANDLE_INST(NUM, OPCODE, CLASS) \ 461 case Instruction::OPCODE:return visit##OPCODE((CLASS&)I); 462#include "llvm/Instruction.def" 463 } 464 } 465 466 void setCurrentBasicBlock(MachineBasicBlock *MBB) { CurMBB = MBB; } 467 468 SDOperand getLoadFrom(const Type *Ty, SDOperand Ptr, 469 const Value *SV, SDOperand Root, 470 bool isVolatile); 471 472 SDOperand getIntPtrConstant(uint64_t Val) { 473 return DAG.getConstant(Val, TLI.getPointerTy()); 474 } 475 476 SDOperand getValue(const Value *V); 477 478 const SDOperand &setValue(const Value *V, SDOperand NewN) { 479 SDOperand &N = NodeMap[V]; 480 assert(N.Val == 0 && "Already set a value for this node!"); 481 return N = NewN; 482 } 483 484 RegsForValue GetRegistersForValue(const std::string &ConstrCode, 485 MVT::ValueType VT, 486 bool OutReg, bool InReg, 487 std::set<unsigned> &OutputRegs, 488 std::set<unsigned> &InputRegs); 489 490 void FindMergedConditions(Value *Cond, MachineBasicBlock *TBB, 491 MachineBasicBlock *FBB, MachineBasicBlock *CurBB, 492 unsigned Opc); 493 bool isExportableFromCurrentBlock(Value *V, const BasicBlock *FromBB); 494 void ExportFromCurrentBlock(Value *V); 495 496 // Terminator instructions. 497 void visitRet(ReturnInst &I); 498 void visitBr(BranchInst &I); 499 void visitSwitch(SwitchInst &I); 500 void visitUnreachable(UnreachableInst &I) { /* noop */ } 501 502 // Helper for visitSwitch 503 void visitSwitchCase(SelectionDAGISel::CaseBlock &CB); 504 void visitJumpTable(SelectionDAGISel::JumpTable &JT); 505 506 // These all get lowered before this pass. 507 void visitInvoke(InvokeInst &I) { assert(0 && "TODO"); } 508 void visitUnwind(UnwindInst &I) { assert(0 && "TODO"); } 509 510 void visitIntBinary(User &I, unsigned IntOp, unsigned VecOp); 511 void visitFPBinary(User &I, unsigned FPOp, unsigned VecOp); 512 void visitShift(User &I, unsigned Opcode); 513 void visitAdd(User &I) { 514 if (I.getType()->isFloatingPoint()) 515 visitFPBinary(I, ISD::FADD, ISD::VADD); 516 else 517 visitIntBinary(I, ISD::ADD, ISD::VADD); 518 } 519 void visitSub(User &I); 520 void visitMul(User &I) { 521 if (I.getType()->isFloatingPoint()) 522 visitFPBinary(I, ISD::FMUL, ISD::VMUL); 523 else 524 visitIntBinary(I, ISD::MUL, ISD::VMUL); 525 } 526 void visitURem(User &I) { visitIntBinary(I, ISD::UREM, 0); } 527 void visitSRem(User &I) { visitIntBinary(I, ISD::SREM, 0); } 528 void visitFRem(User &I) { visitFPBinary (I, ISD::FREM, 0); } 529 void visitUDiv(User &I) { visitIntBinary(I, ISD::UDIV, ISD::VUDIV); } 530 void visitSDiv(User &I) { visitIntBinary(I, ISD::SDIV, ISD::VSDIV); } 531 void visitFDiv(User &I) { visitFPBinary (I, ISD::FDIV, ISD::VSDIV); } 532 void visitAnd(User &I) { visitIntBinary(I, ISD::AND, ISD::VAND); } 533 void visitOr (User &I) { visitIntBinary(I, ISD::OR, ISD::VOR); } 534 void visitXor(User &I) { visitIntBinary(I, ISD::XOR, ISD::VXOR); } 535 void visitShl(User &I) { visitShift(I, ISD::SHL); } 536 void visitLShr(User &I) { visitShift(I, ISD::SRL); } 537 void visitAShr(User &I) { visitShift(I, ISD::SRA); } 538 void visitICmp(User &I); 539 void visitFCmp(User &I); 540 // Visit the conversion instructions 541 void visitTrunc(User &I); 542 void visitZExt(User &I); 543 void visitSExt(User &I); 544 void visitFPTrunc(User &I); 545 void visitFPExt(User &I); 546 void visitFPToUI(User &I); 547 void visitFPToSI(User &I); 548 void visitUIToFP(User &I); 549 void visitSIToFP(User &I); 550 void visitPtrToInt(User &I); 551 void visitIntToPtr(User &I); 552 void visitBitCast(User &I); 553 554 void visitExtractElement(User &I); 555 void visitInsertElement(User &I); 556 void visitShuffleVector(User &I); 557 558 void visitGetElementPtr(User &I); 559 void visitSelect(User &I); 560 561 void visitMalloc(MallocInst &I); 562 void visitFree(FreeInst &I); 563 void visitAlloca(AllocaInst &I); 564 void visitLoad(LoadInst &I); 565 void visitStore(StoreInst &I); 566 void visitPHI(PHINode &I) { } // PHI nodes are handled specially. 567 void visitCall(CallInst &I); 568 void visitInlineAsm(CallInst &I); 569 const char *visitIntrinsicCall(CallInst &I, unsigned Intrinsic); 570 void visitTargetIntrinsic(CallInst &I, unsigned Intrinsic); 571 572 void visitVAStart(CallInst &I); 573 void visitVAArg(VAArgInst &I); 574 void visitVAEnd(CallInst &I); 575 void visitVACopy(CallInst &I); 576 void visitFrameReturnAddress(CallInst &I, bool isFrameAddress); 577 578 void visitMemIntrinsic(CallInst &I, unsigned Op); 579 580 void visitUserOp1(Instruction &I) { 581 assert(0 && "UserOp1 should not exist at instruction selection time!"); 582 abort(); 583 } 584 void visitUserOp2(Instruction &I) { 585 assert(0 && "UserOp2 should not exist at instruction selection time!"); 586 abort(); 587 } 588}; 589} // end namespace llvm 590 591SDOperand SelectionDAGLowering::getValue(const Value *V) { 592 SDOperand &N = NodeMap[V]; 593 if (N.Val) return N; 594 595 const Type *VTy = V->getType(); 596 MVT::ValueType VT = TLI.getValueType(VTy); 597 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) { 598 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) { 599 visit(CE->getOpcode(), *CE); 600 assert(N.Val && "visit didn't populate the ValueMap!"); 601 return N; 602 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(C)) { 603 return N = DAG.getGlobalAddress(GV, VT); 604 } else if (isa<ConstantPointerNull>(C)) { 605 return N = DAG.getConstant(0, TLI.getPointerTy()); 606 } else if (isa<UndefValue>(C)) { 607 if (!isa<PackedType>(VTy)) 608 return N = DAG.getNode(ISD::UNDEF, VT); 609 610 // Create a VBUILD_VECTOR of undef nodes. 611 const PackedType *PTy = cast<PackedType>(VTy); 612 unsigned NumElements = PTy->getNumElements(); 613 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType()); 614 615 SmallVector<SDOperand, 8> Ops; 616 Ops.assign(NumElements, DAG.getNode(ISD::UNDEF, PVT)); 617 618 // Create a VConstant node with generic Vector type. 619 Ops.push_back(DAG.getConstant(NumElements, MVT::i32)); 620 Ops.push_back(DAG.getValueType(PVT)); 621 return N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, 622 &Ops[0], Ops.size()); 623 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) { 624 return N = DAG.getConstantFP(CFP->getValue(), VT); 625 } else if (const PackedType *PTy = dyn_cast<PackedType>(VTy)) { 626 unsigned NumElements = PTy->getNumElements(); 627 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType()); 628 629 // Now that we know the number and type of the elements, push a 630 // Constant or ConstantFP node onto the ops list for each element of 631 // the packed constant. 632 SmallVector<SDOperand, 8> Ops; 633 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(C)) { 634 for (unsigned i = 0; i != NumElements; ++i) 635 Ops.push_back(getValue(CP->getOperand(i))); 636 } else { 637 assert(isa<ConstantAggregateZero>(C) && "Unknown packed constant!"); 638 SDOperand Op; 639 if (MVT::isFloatingPoint(PVT)) 640 Op = DAG.getConstantFP(0, PVT); 641 else 642 Op = DAG.getConstant(0, PVT); 643 Ops.assign(NumElements, Op); 644 } 645 646 // Create a VBUILD_VECTOR node with generic Vector type. 647 Ops.push_back(DAG.getConstant(NumElements, MVT::i32)); 648 Ops.push_back(DAG.getValueType(PVT)); 649 return N = DAG.getNode(ISD::VBUILD_VECTOR,MVT::Vector,&Ops[0],Ops.size()); 650 } else { 651 // Canonicalize all constant ints to be unsigned. 652 return N = DAG.getConstant(cast<ConstantIntegral>(C)->getZExtValue(),VT); 653 } 654 } 655 656 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) { 657 std::map<const AllocaInst*, int>::iterator SI = 658 FuncInfo.StaticAllocaMap.find(AI); 659 if (SI != FuncInfo.StaticAllocaMap.end()) 660 return DAG.getFrameIndex(SI->second, TLI.getPointerTy()); 661 } 662 663 std::map<const Value*, unsigned>::const_iterator VMI = 664 FuncInfo.ValueMap.find(V); 665 assert(VMI != FuncInfo.ValueMap.end() && "Value not in map!"); 666 667 unsigned InReg = VMI->second; 668 669 // If this type is not legal, make it so now. 670 if (VT != MVT::Vector) { 671 if (TLI.getTypeAction(VT) == TargetLowering::Expand) { 672 // Source must be expanded. This input value is actually coming from the 673 // register pair VMI->second and VMI->second+1. 674 MVT::ValueType DestVT = TLI.getTypeToExpandTo(VT); 675 unsigned NumVals = TLI.getNumElements(VT); 676 N = DAG.getCopyFromReg(DAG.getEntryNode(), InReg, DestVT); 677 if (NumVals == 1) 678 N = DAG.getNode(ISD::BIT_CONVERT, VT, N); 679 else { 680 assert(NumVals == 2 && "1 to 4 (and more) expansion not implemented!"); 681 N = DAG.getNode(ISD::BUILD_PAIR, VT, N, 682 DAG.getCopyFromReg(DAG.getEntryNode(), InReg+1, DestVT)); 683 } 684 } else { 685 MVT::ValueType DestVT = TLI.getTypeToTransformTo(VT); 686 N = DAG.getCopyFromReg(DAG.getEntryNode(), InReg, DestVT); 687 if (TLI.getTypeAction(VT) == TargetLowering::Promote) // Promotion case 688 N = MVT::isFloatingPoint(VT) 689 ? DAG.getNode(ISD::FP_ROUND, VT, N) 690 : DAG.getNode(ISD::TRUNCATE, VT, N); 691 } 692 } else { 693 // Otherwise, if this is a vector, make it available as a generic vector 694 // here. 695 MVT::ValueType PTyElementVT, PTyLegalElementVT; 696 const PackedType *PTy = cast<PackedType>(VTy); 697 unsigned NE = TLI.getPackedTypeBreakdown(PTy, PTyElementVT, 698 PTyLegalElementVT); 699 700 // Build a VBUILD_VECTOR with the input registers. 701 SmallVector<SDOperand, 8> Ops; 702 if (PTyElementVT == PTyLegalElementVT) { 703 // If the value types are legal, just VBUILD the CopyFromReg nodes. 704 for (unsigned i = 0; i != NE; ++i) 705 Ops.push_back(DAG.getCopyFromReg(DAG.getEntryNode(), InReg++, 706 PTyElementVT)); 707 } else if (PTyElementVT < PTyLegalElementVT) { 708 // If the register was promoted, use TRUNCATE of FP_ROUND as appropriate. 709 for (unsigned i = 0; i != NE; ++i) { 710 SDOperand Op = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++, 711 PTyElementVT); 712 if (MVT::isFloatingPoint(PTyElementVT)) 713 Op = DAG.getNode(ISD::FP_ROUND, PTyElementVT, Op); 714 else 715 Op = DAG.getNode(ISD::TRUNCATE, PTyElementVT, Op); 716 Ops.push_back(Op); 717 } 718 } else { 719 // If the register was expanded, use BUILD_PAIR. 720 assert((NE & 1) == 0 && "Must expand into a multiple of 2 elements!"); 721 for (unsigned i = 0; i != NE/2; ++i) { 722 SDOperand Op0 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++, 723 PTyElementVT); 724 SDOperand Op1 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++, 725 PTyElementVT); 726 Ops.push_back(DAG.getNode(ISD::BUILD_PAIR, VT, Op0, Op1)); 727 } 728 } 729 730 Ops.push_back(DAG.getConstant(NE, MVT::i32)); 731 Ops.push_back(DAG.getValueType(PTyLegalElementVT)); 732 N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size()); 733 734 // Finally, use a VBIT_CONVERT to make this available as the appropriate 735 // vector type. 736 N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N, 737 DAG.getConstant(PTy->getNumElements(), 738 MVT::i32), 739 DAG.getValueType(TLI.getValueType(PTy->getElementType()))); 740 } 741 742 return N; 743} 744 745 746void SelectionDAGLowering::visitRet(ReturnInst &I) { 747 if (I.getNumOperands() == 0) { 748 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, getRoot())); 749 return; 750 } 751 SmallVector<SDOperand, 8> NewValues; 752 NewValues.push_back(getRoot()); 753 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { 754 SDOperand RetOp = getValue(I.getOperand(i)); 755 756 // If this is an integer return value, we need to promote it ourselves to 757 // the full width of a register, since LegalizeOp will use ANY_EXTEND rather 758 // than sign/zero. 759 // FIXME: C calling convention requires the return type to be promoted to 760 // at least 32-bit. But this is not necessary for non-C calling conventions. 761 if (MVT::isInteger(RetOp.getValueType()) && 762 RetOp.getValueType() < MVT::i64) { 763 MVT::ValueType TmpVT; 764 if (TLI.getTypeAction(MVT::i32) == TargetLowering::Promote) 765 TmpVT = TLI.getTypeToTransformTo(MVT::i32); 766 else 767 TmpVT = MVT::i32; 768 const FunctionType *FTy = I.getParent()->getParent()->getFunctionType(); 769 ISD::NodeType ExtendKind = ISD::SIGN_EXTEND; 770 if (FTy->paramHasAttr(0, FunctionType::ZExtAttribute)) 771 ExtendKind = ISD::ZERO_EXTEND; 772 RetOp = DAG.getNode(ExtendKind, TmpVT, RetOp); 773 } 774 NewValues.push_back(RetOp); 775 NewValues.push_back(DAG.getConstant(false, MVT::i32)); 776 } 777 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, 778 &NewValues[0], NewValues.size())); 779} 780 781/// ExportFromCurrentBlock - If this condition isn't known to be exported from 782/// the current basic block, add it to ValueMap now so that we'll get a 783/// CopyTo/FromReg. 784void SelectionDAGLowering::ExportFromCurrentBlock(Value *V) { 785 // No need to export constants. 786 if (!isa<Instruction>(V) && !isa<Argument>(V)) return; 787 788 // Already exported? 789 if (FuncInfo.isExportedInst(V)) return; 790 791 unsigned Reg = FuncInfo.InitializeRegForValue(V); 792 PendingLoads.push_back(CopyValueToVirtualRegister(V, Reg)); 793} 794 795bool SelectionDAGLowering::isExportableFromCurrentBlock(Value *V, 796 const BasicBlock *FromBB) { 797 // The operands of the setcc have to be in this block. We don't know 798 // how to export them from some other block. 799 if (Instruction *VI = dyn_cast<Instruction>(V)) { 800 // Can export from current BB. 801 if (VI->getParent() == FromBB) 802 return true; 803 804 // Is already exported, noop. 805 return FuncInfo.isExportedInst(V); 806 } 807 808 // If this is an argument, we can export it if the BB is the entry block or 809 // if it is already exported. 810 if (isa<Argument>(V)) { 811 if (FromBB == &FromBB->getParent()->getEntryBlock()) 812 return true; 813 814 // Otherwise, can only export this if it is already exported. 815 return FuncInfo.isExportedInst(V); 816 } 817 818 // Otherwise, constants can always be exported. 819 return true; 820} 821 822static bool InBlock(const Value *V, const BasicBlock *BB) { 823 if (const Instruction *I = dyn_cast<Instruction>(V)) 824 return I->getParent() == BB; 825 return true; 826} 827 828/// FindMergedConditions - If Cond is an expression like 829void SelectionDAGLowering::FindMergedConditions(Value *Cond, 830 MachineBasicBlock *TBB, 831 MachineBasicBlock *FBB, 832 MachineBasicBlock *CurBB, 833 unsigned Opc) { 834 // If this node is not part of the or/and tree, emit it as a branch. 835 Instruction *BOp = dyn_cast<Instruction>(Cond); 836 837 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) || 838 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() || 839 BOp->getParent() != CurBB->getBasicBlock() || 840 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) || 841 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) { 842 const BasicBlock *BB = CurBB->getBasicBlock(); 843 844 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Cond)) 845 if ((II->getIntrinsicID() == Intrinsic::isunordered_f32 || 846 II->getIntrinsicID() == Intrinsic::isunordered_f64) && 847 // The operands of the setcc have to be in this block. We don't know 848 // how to export them from some other block. If this is the first 849 // block of the sequence, no exporting is needed. 850 (CurBB == CurMBB || 851 (isExportableFromCurrentBlock(II->getOperand(1), BB) && 852 isExportableFromCurrentBlock(II->getOperand(2), BB)))) { 853 SelectionDAGISel::CaseBlock CB(ISD::SETUO, II->getOperand(1), 854 II->getOperand(2), TBB, FBB, CurBB); 855 SwitchCases.push_back(CB); 856 return; 857 } 858 859 860 // If the leaf of the tree is a comparison, merge the condition into 861 // the caseblock. 862 if ((isa<ICmpInst>(Cond) || isa<FCmpInst>(Cond)) && 863 // The operands of the cmp have to be in this block. We don't know 864 // how to export them from some other block. If this is the first block 865 // of the sequence, no exporting is needed. 866 (CurBB == CurMBB || 867 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) && 868 isExportableFromCurrentBlock(BOp->getOperand(1), BB)))) { 869 BOp = cast<Instruction>(Cond); 870 ISD::CondCode Condition; 871 if (ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) { 872 switch (IC->getPredicate()) { 873 default: assert(0 && "Unknown icmp predicate opcode!"); 874 case ICmpInst::ICMP_EQ: Condition = ISD::SETEQ; break; 875 case ICmpInst::ICMP_NE: Condition = ISD::SETNE; break; 876 case ICmpInst::ICMP_SLE: Condition = ISD::SETLE; break; 877 case ICmpInst::ICMP_ULE: Condition = ISD::SETULE; break; 878 case ICmpInst::ICMP_SGE: Condition = ISD::SETGE; break; 879 case ICmpInst::ICMP_UGE: Condition = ISD::SETUGE; break; 880 case ICmpInst::ICMP_SLT: Condition = ISD::SETLT; break; 881 case ICmpInst::ICMP_ULT: Condition = ISD::SETULT; break; 882 case ICmpInst::ICMP_SGT: Condition = ISD::SETGT; break; 883 case ICmpInst::ICMP_UGT: Condition = ISD::SETUGT; break; 884 } 885 } else if (FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) { 886 ISD::CondCode FPC, FOC; 887 switch (FC->getPredicate()) { 888 default: assert(0 && "Unknown fcmp predicate opcode!"); 889 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break; 890 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break; 891 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break; 892 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break; 893 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break; 894 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break; 895 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break; 896 case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break; 897 case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break; 898 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break; 899 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break; 900 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break; 901 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break; 902 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break; 903 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break; 904 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break; 905 } 906 if (FiniteOnlyFPMath()) 907 Condition = FOC; 908 else 909 Condition = FPC; 910 } else { 911 assert(0 && "Unknown compare instruction"); 912 } 913 914 SelectionDAGISel::CaseBlock CB(Condition, BOp->getOperand(0), 915 BOp->getOperand(1), TBB, FBB, CurBB); 916 SwitchCases.push_back(CB); 917 return; 918 } 919 920 // Create a CaseBlock record representing this branch. 921 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, Cond, ConstantBool::getTrue(), 922 TBB, FBB, CurBB); 923 SwitchCases.push_back(CB); 924 return; 925 } 926 927 928 // Create TmpBB after CurBB. 929 MachineFunction::iterator BBI = CurBB; 930 MachineBasicBlock *TmpBB = new MachineBasicBlock(CurBB->getBasicBlock()); 931 CurBB->getParent()->getBasicBlockList().insert(++BBI, TmpBB); 932 933 if (Opc == Instruction::Or) { 934 // Codegen X | Y as: 935 // jmp_if_X TBB 936 // jmp TmpBB 937 // TmpBB: 938 // jmp_if_Y TBB 939 // jmp FBB 940 // 941 942 // Emit the LHS condition. 943 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc); 944 945 // Emit the RHS condition into TmpBB. 946 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc); 947 } else { 948 assert(Opc == Instruction::And && "Unknown merge op!"); 949 // Codegen X & Y as: 950 // jmp_if_X TmpBB 951 // jmp FBB 952 // TmpBB: 953 // jmp_if_Y TBB 954 // jmp FBB 955 // 956 // This requires creation of TmpBB after CurBB. 957 958 // Emit the LHS condition. 959 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc); 960 961 // Emit the RHS condition into TmpBB. 962 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc); 963 } 964} 965 966/// If the set of cases should be emitted as a series of branches, return true. 967/// If we should emit this as a bunch of and/or'd together conditions, return 968/// false. 969static bool 970ShouldEmitAsBranches(const std::vector<SelectionDAGISel::CaseBlock> &Cases) { 971 if (Cases.size() != 2) return true; 972 973 // If this is two comparisons of the same values or'd or and'd together, they 974 // will get folded into a single comparison, so don't emit two blocks. 975 if ((Cases[0].CmpLHS == Cases[1].CmpLHS && 976 Cases[0].CmpRHS == Cases[1].CmpRHS) || 977 (Cases[0].CmpRHS == Cases[1].CmpLHS && 978 Cases[0].CmpLHS == Cases[1].CmpRHS)) { 979 return false; 980 } 981 982 return true; 983} 984 985void SelectionDAGLowering::visitBr(BranchInst &I) { 986 // Update machine-CFG edges. 987 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)]; 988 989 // Figure out which block is immediately after the current one. 990 MachineBasicBlock *NextBlock = 0; 991 MachineFunction::iterator BBI = CurMBB; 992 if (++BBI != CurMBB->getParent()->end()) 993 NextBlock = BBI; 994 995 if (I.isUnconditional()) { 996 // If this is not a fall-through branch, emit the branch. 997 if (Succ0MBB != NextBlock) 998 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(), 999 DAG.getBasicBlock(Succ0MBB))); 1000 1001 // Update machine-CFG edges. 1002 CurMBB->addSuccessor(Succ0MBB); 1003 1004 return; 1005 } 1006 1007 // If this condition is one of the special cases we handle, do special stuff 1008 // now. 1009 Value *CondVal = I.getCondition(); 1010 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)]; 1011 1012 // If this is a series of conditions that are or'd or and'd together, emit 1013 // this as a sequence of branches instead of setcc's with and/or operations. 1014 // For example, instead of something like: 1015 // cmp A, B 1016 // C = seteq 1017 // cmp D, E 1018 // F = setle 1019 // or C, F 1020 // jnz foo 1021 // Emit: 1022 // cmp A, B 1023 // je foo 1024 // cmp D, E 1025 // jle foo 1026 // 1027 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) { 1028 if (BOp->hasOneUse() && 1029 (BOp->getOpcode() == Instruction::And || 1030 BOp->getOpcode() == Instruction::Or)) { 1031 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode()); 1032 // If the compares in later blocks need to use values not currently 1033 // exported from this block, export them now. This block should always 1034 // be the first entry. 1035 assert(SwitchCases[0].ThisBB == CurMBB && "Unexpected lowering!"); 1036 1037 // Allow some cases to be rejected. 1038 if (ShouldEmitAsBranches(SwitchCases)) { 1039 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) { 1040 ExportFromCurrentBlock(SwitchCases[i].CmpLHS); 1041 ExportFromCurrentBlock(SwitchCases[i].CmpRHS); 1042 } 1043 1044 // Emit the branch for this block. 1045 visitSwitchCase(SwitchCases[0]); 1046 SwitchCases.erase(SwitchCases.begin()); 1047 return; 1048 } 1049 1050 // Okay, we decided not to do this, remove any inserted MBB's and clear 1051 // SwitchCases. 1052 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) 1053 CurMBB->getParent()->getBasicBlockList().erase(SwitchCases[i].ThisBB); 1054 1055 SwitchCases.clear(); 1056 } 1057 } 1058 1059 // Create a CaseBlock record representing this branch. 1060 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, CondVal, ConstantBool::getTrue(), 1061 Succ0MBB, Succ1MBB, CurMBB); 1062 // Use visitSwitchCase to actually insert the fast branch sequence for this 1063 // cond branch. 1064 visitSwitchCase(CB); 1065} 1066 1067/// visitSwitchCase - Emits the necessary code to represent a single node in 1068/// the binary search tree resulting from lowering a switch instruction. 1069void SelectionDAGLowering::visitSwitchCase(SelectionDAGISel::CaseBlock &CB) { 1070 SDOperand Cond; 1071 SDOperand CondLHS = getValue(CB.CmpLHS); 1072 1073 // Build the setcc now, fold "(X == true)" to X and "(X == false)" to !X to 1074 // handle common cases produced by branch lowering. 1075 if (CB.CmpRHS == ConstantBool::getTrue() && CB.CC == ISD::SETEQ) 1076 Cond = CondLHS; 1077 else if (CB.CmpRHS == ConstantBool::getFalse() && CB.CC == ISD::SETEQ) { 1078 SDOperand True = DAG.getConstant(1, CondLHS.getValueType()); 1079 Cond = DAG.getNode(ISD::XOR, CondLHS.getValueType(), CondLHS, True); 1080 } else 1081 Cond = DAG.getSetCC(MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC); 1082 1083 // Set NextBlock to be the MBB immediately after the current one, if any. 1084 // This is used to avoid emitting unnecessary branches to the next block. 1085 MachineBasicBlock *NextBlock = 0; 1086 MachineFunction::iterator BBI = CurMBB; 1087 if (++BBI != CurMBB->getParent()->end()) 1088 NextBlock = BBI; 1089 1090 // If the lhs block is the next block, invert the condition so that we can 1091 // fall through to the lhs instead of the rhs block. 1092 if (CB.TrueBB == NextBlock) { 1093 std::swap(CB.TrueBB, CB.FalseBB); 1094 SDOperand True = DAG.getConstant(1, Cond.getValueType()); 1095 Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True); 1096 } 1097 SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(), Cond, 1098 DAG.getBasicBlock(CB.TrueBB)); 1099 if (CB.FalseBB == NextBlock) 1100 DAG.setRoot(BrCond); 1101 else 1102 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond, 1103 DAG.getBasicBlock(CB.FalseBB))); 1104 // Update successor info 1105 CurMBB->addSuccessor(CB.TrueBB); 1106 CurMBB->addSuccessor(CB.FalseBB); 1107} 1108 1109void SelectionDAGLowering::visitJumpTable(SelectionDAGISel::JumpTable &JT) { 1110 // Emit the code for the jump table 1111 MVT::ValueType PTy = TLI.getPointerTy(); 1112 SDOperand Index = DAG.getCopyFromReg(getRoot(), JT.Reg, PTy); 1113 SDOperand Table = DAG.getJumpTable(JT.JTI, PTy); 1114 DAG.setRoot(DAG.getNode(ISD::BR_JT, MVT::Other, Index.getValue(1), 1115 Table, Index)); 1116 return; 1117} 1118 1119void SelectionDAGLowering::visitSwitch(SwitchInst &I) { 1120 // Figure out which block is immediately after the current one. 1121 MachineBasicBlock *NextBlock = 0; 1122 MachineFunction::iterator BBI = CurMBB; 1123 1124 if (++BBI != CurMBB->getParent()->end()) 1125 NextBlock = BBI; 1126 1127 MachineBasicBlock *Default = FuncInfo.MBBMap[I.getDefaultDest()]; 1128 1129 // If there is only the default destination, branch to it if it is not the 1130 // next basic block. Otherwise, just fall through. 1131 if (I.getNumOperands() == 2) { 1132 // Update machine-CFG edges. 1133 1134 // If this is not a fall-through branch, emit the branch. 1135 if (Default != NextBlock) 1136 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(), 1137 DAG.getBasicBlock(Default))); 1138 1139 CurMBB->addSuccessor(Default); 1140 return; 1141 } 1142 1143 // If there are any non-default case statements, create a vector of Cases 1144 // representing each one, and sort the vector so that we can efficiently 1145 // create a binary search tree from them. 1146 std::vector<Case> Cases; 1147 1148 for (unsigned i = 1; i < I.getNumSuccessors(); ++i) { 1149 MachineBasicBlock *SMBB = FuncInfo.MBBMap[I.getSuccessor(i)]; 1150 Cases.push_back(Case(I.getSuccessorValue(i), SMBB)); 1151 } 1152 1153 std::sort(Cases.begin(), Cases.end(), CaseCmp()); 1154 1155 // Get the Value to be switched on and default basic blocks, which will be 1156 // inserted into CaseBlock records, representing basic blocks in the binary 1157 // search tree. 1158 Value *SV = I.getOperand(0); 1159 1160 // Get the MachineFunction which holds the current MBB. This is used during 1161 // emission of jump tables, and when inserting any additional MBBs necessary 1162 // to represent the switch. 1163 MachineFunction *CurMF = CurMBB->getParent(); 1164 const BasicBlock *LLVMBB = CurMBB->getBasicBlock(); 1165 1166 // If the switch has few cases (two or less) emit a series of specific 1167 // tests. 1168 if (Cases.size() < 3) { 1169 // TODO: If any two of the cases has the same destination, and if one value 1170 // is the same as the other, but has one bit unset that the other has set, 1171 // use bit manipulation to do two compares at once. For example: 1172 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)" 1173 1174 // Rearrange the case blocks so that the last one falls through if possible. 1175 if (NextBlock && Default != NextBlock && Cases.back().second != NextBlock) { 1176 // The last case block won't fall through into 'NextBlock' if we emit the 1177 // branches in this order. See if rearranging a case value would help. 1178 for (unsigned i = 0, e = Cases.size()-1; i != e; ++i) { 1179 if (Cases[i].second == NextBlock) { 1180 std::swap(Cases[i], Cases.back()); 1181 break; 1182 } 1183 } 1184 } 1185 1186 // Create a CaseBlock record representing a conditional branch to 1187 // the Case's target mbb if the value being switched on SV is equal 1188 // to C. 1189 MachineBasicBlock *CurBlock = CurMBB; 1190 for (unsigned i = 0, e = Cases.size(); i != e; ++i) { 1191 MachineBasicBlock *FallThrough; 1192 if (i != e-1) { 1193 FallThrough = new MachineBasicBlock(CurMBB->getBasicBlock()); 1194 CurMF->getBasicBlockList().insert(BBI, FallThrough); 1195 } else { 1196 // If the last case doesn't match, go to the default block. 1197 FallThrough = Default; 1198 } 1199 1200 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, SV, Cases[i].first, 1201 Cases[i].second, FallThrough, CurBlock); 1202 1203 // If emitting the first comparison, just call visitSwitchCase to emit the 1204 // code into the current block. Otherwise, push the CaseBlock onto the 1205 // vector to be later processed by SDISel, and insert the node's MBB 1206 // before the next MBB. 1207 if (CurBlock == CurMBB) 1208 visitSwitchCase(CB); 1209 else 1210 SwitchCases.push_back(CB); 1211 1212 CurBlock = FallThrough; 1213 } 1214 return; 1215 } 1216 1217 // If the switch has more than 5 blocks, and at least 31.25% dense, and the 1218 // target supports indirect branches, then emit a jump table rather than 1219 // lowering the switch to a binary tree of conditional branches. 1220 if ((TLI.isOperationLegal(ISD::BR_JT, MVT::Other) || 1221 TLI.isOperationLegal(ISD::BRIND, MVT::Other)) && 1222 Cases.size() > 5) { 1223 uint64_t First =cast<ConstantIntegral>(Cases.front().first)->getZExtValue(); 1224 uint64_t Last = cast<ConstantIntegral>(Cases.back().first)->getZExtValue(); 1225 double Density = (double)Cases.size() / (double)((Last - First) + 1ULL); 1226 1227 if (Density >= 0.3125) { 1228 // Create a new basic block to hold the code for loading the address 1229 // of the jump table, and jumping to it. Update successor information; 1230 // we will either branch to the default case for the switch, or the jump 1231 // table. 1232 MachineBasicBlock *JumpTableBB = new MachineBasicBlock(LLVMBB); 1233 CurMF->getBasicBlockList().insert(BBI, JumpTableBB); 1234 CurMBB->addSuccessor(Default); 1235 CurMBB->addSuccessor(JumpTableBB); 1236 1237 // Subtract the lowest switch case value from the value being switched on 1238 // and conditional branch to default mbb if the result is greater than the 1239 // difference between smallest and largest cases. 1240 SDOperand SwitchOp = getValue(SV); 1241 MVT::ValueType VT = SwitchOp.getValueType(); 1242 SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp, 1243 DAG.getConstant(First, VT)); 1244 1245 // The SDNode we just created, which holds the value being switched on 1246 // minus the the smallest case value, needs to be copied to a virtual 1247 // register so it can be used as an index into the jump table in a 1248 // subsequent basic block. This value may be smaller or larger than the 1249 // target's pointer type, and therefore require extension or truncating. 1250 if (VT > TLI.getPointerTy()) 1251 SwitchOp = DAG.getNode(ISD::TRUNCATE, TLI.getPointerTy(), SUB); 1252 else 1253 SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), SUB); 1254 1255 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy()); 1256 SDOperand CopyTo = DAG.getCopyToReg(getRoot(), JumpTableReg, SwitchOp); 1257 1258 // Emit the range check for the jump table, and branch to the default 1259 // block for the switch statement if the value being switched on exceeds 1260 // the largest case in the switch. 1261 SDOperand CMP = DAG.getSetCC(TLI.getSetCCResultTy(), SUB, 1262 DAG.getConstant(Last-First,VT), ISD::SETUGT); 1263 DAG.setRoot(DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, CMP, 1264 DAG.getBasicBlock(Default))); 1265 1266 // Build a vector of destination BBs, corresponding to each target 1267 // of the jump table. If the value of the jump table slot corresponds to 1268 // a case statement, push the case's BB onto the vector, otherwise, push 1269 // the default BB. 1270 std::vector<MachineBasicBlock*> DestBBs; 1271 uint64_t TEI = First; 1272 for (CaseItr ii = Cases.begin(), ee = Cases.end(); ii != ee; ++TEI) 1273 if (cast<ConstantIntegral>(ii->first)->getZExtValue() == TEI) { 1274 DestBBs.push_back(ii->second); 1275 ++ii; 1276 } else { 1277 DestBBs.push_back(Default); 1278 } 1279 1280 // Update successor info. Add one edge to each unique successor. 1281 // Vector bool would be better, but vector<bool> is really slow. 1282 std::vector<unsigned char> SuccsHandled; 1283 SuccsHandled.resize(CurMBB->getParent()->getNumBlockIDs()); 1284 1285 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(), 1286 E = DestBBs.end(); I != E; ++I) { 1287 if (!SuccsHandled[(*I)->getNumber()]) { 1288 SuccsHandled[(*I)->getNumber()] = true; 1289 JumpTableBB->addSuccessor(*I); 1290 } 1291 } 1292 1293 // Create a jump table index for this jump table, or return an existing 1294 // one. 1295 unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs); 1296 1297 // Set the jump table information so that we can codegen it as a second 1298 // MachineBasicBlock 1299 JT.Reg = JumpTableReg; 1300 JT.JTI = JTI; 1301 JT.MBB = JumpTableBB; 1302 JT.Default = Default; 1303 return; 1304 } 1305 } 1306 1307 // Push the initial CaseRec onto the worklist 1308 std::vector<CaseRec> CaseVec; 1309 CaseVec.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end()))); 1310 1311 while (!CaseVec.empty()) { 1312 // Grab a record representing a case range to process off the worklist 1313 CaseRec CR = CaseVec.back(); 1314 CaseVec.pop_back(); 1315 1316 // Size is the number of Cases represented by this range. If Size is 1, 1317 // then we are processing a leaf of the binary search tree. Otherwise, 1318 // we need to pick a pivot, and push left and right ranges onto the 1319 // worklist. 1320 unsigned Size = CR.Range.second - CR.Range.first; 1321 1322 if (Size == 1) { 1323 // Create a CaseBlock record representing a conditional branch to 1324 // the Case's target mbb if the value being switched on SV is equal 1325 // to C. Otherwise, branch to default. 1326 Constant *C = CR.Range.first->first; 1327 MachineBasicBlock *Target = CR.Range.first->second; 1328 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, SV, C, Target, Default, 1329 CR.CaseBB); 1330 1331 // If the MBB representing the leaf node is the current MBB, then just 1332 // call visitSwitchCase to emit the code into the current block. 1333 // Otherwise, push the CaseBlock onto the vector to be later processed 1334 // by SDISel, and insert the node's MBB before the next MBB. 1335 if (CR.CaseBB == CurMBB) 1336 visitSwitchCase(CB); 1337 else 1338 SwitchCases.push_back(CB); 1339 } else { 1340 // split case range at pivot 1341 CaseItr Pivot = CR.Range.first + (Size / 2); 1342 CaseRange LHSR(CR.Range.first, Pivot); 1343 CaseRange RHSR(Pivot, CR.Range.second); 1344 Constant *C = Pivot->first; 1345 MachineBasicBlock *FalseBB = 0, *TrueBB = 0; 1346 1347 // We know that we branch to the LHS if the Value being switched on is 1348 // less than the Pivot value, C. We use this to optimize our binary 1349 // tree a bit, by recognizing that if SV is greater than or equal to the 1350 // LHS's Case Value, and that Case Value is exactly one less than the 1351 // Pivot's Value, then we can branch directly to the LHS's Target, 1352 // rather than creating a leaf node for it. 1353 if ((LHSR.second - LHSR.first) == 1 && 1354 LHSR.first->first == CR.GE && 1355 cast<ConstantIntegral>(C)->getZExtValue() == 1356 (cast<ConstantIntegral>(CR.GE)->getZExtValue() + 1ULL)) { 1357 TrueBB = LHSR.first->second; 1358 } else { 1359 TrueBB = new MachineBasicBlock(LLVMBB); 1360 CurMF->getBasicBlockList().insert(BBI, TrueBB); 1361 CaseVec.push_back(CaseRec(TrueBB, C, CR.GE, LHSR)); 1362 } 1363 1364 // Similar to the optimization above, if the Value being switched on is 1365 // known to be less than the Constant CR.LT, and the current Case Value 1366 // is CR.LT - 1, then we can branch directly to the target block for 1367 // the current Case Value, rather than emitting a RHS leaf node for it. 1368 if ((RHSR.second - RHSR.first) == 1 && CR.LT && 1369 cast<ConstantIntegral>(RHSR.first->first)->getZExtValue() == 1370 (cast<ConstantIntegral>(CR.LT)->getZExtValue() - 1ULL)) { 1371 FalseBB = RHSR.first->second; 1372 } else { 1373 FalseBB = new MachineBasicBlock(LLVMBB); 1374 CurMF->getBasicBlockList().insert(BBI, FalseBB); 1375 CaseVec.push_back(CaseRec(FalseBB,CR.LT,C,RHSR)); 1376 } 1377 1378 // Create a CaseBlock record representing a conditional branch to 1379 // the LHS node if the value being switched on SV is less than C. 1380 // Otherwise, branch to LHS. 1381 ISD::CondCode CC = ISD::SETULT; 1382 SelectionDAGISel::CaseBlock CB(CC, SV, C, TrueBB, FalseBB, CR.CaseBB); 1383 1384 if (CR.CaseBB == CurMBB) 1385 visitSwitchCase(CB); 1386 else 1387 SwitchCases.push_back(CB); 1388 } 1389 } 1390} 1391 1392void SelectionDAGLowering::visitSub(User &I) { 1393 // -0.0 - X --> fneg 1394 if (I.getType()->isFloatingPoint()) { 1395 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0))) 1396 if (CFP->isExactlyValue(-0.0)) { 1397 SDOperand Op2 = getValue(I.getOperand(1)); 1398 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2)); 1399 return; 1400 } 1401 visitFPBinary(I, ISD::FSUB, ISD::VSUB); 1402 } else 1403 visitIntBinary(I, ISD::SUB, ISD::VSUB); 1404} 1405 1406void 1407SelectionDAGLowering::visitIntBinary(User &I, unsigned IntOp, unsigned VecOp) { 1408 const Type *Ty = I.getType(); 1409 SDOperand Op1 = getValue(I.getOperand(0)); 1410 SDOperand Op2 = getValue(I.getOperand(1)); 1411 1412 if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) { 1413 SDOperand Num = DAG.getConstant(PTy->getNumElements(), MVT::i32); 1414 SDOperand Typ = DAG.getValueType(TLI.getValueType(PTy->getElementType())); 1415 setValue(&I, DAG.getNode(VecOp, MVT::Vector, Op1, Op2, Num, Typ)); 1416 } else { 1417 setValue(&I, DAG.getNode(IntOp, Op1.getValueType(), Op1, Op2)); 1418 } 1419} 1420 1421void 1422SelectionDAGLowering::visitFPBinary(User &I, unsigned FPOp, unsigned VecOp) { 1423 const Type *Ty = I.getType(); 1424 SDOperand Op1 = getValue(I.getOperand(0)); 1425 SDOperand Op2 = getValue(I.getOperand(1)); 1426 1427 if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) { 1428 SDOperand Num = DAG.getConstant(PTy->getNumElements(), MVT::i32); 1429 SDOperand Typ = DAG.getValueType(TLI.getValueType(PTy->getElementType())); 1430 setValue(&I, DAG.getNode(VecOp, MVT::Vector, Op1, Op2, Num, Typ)); 1431 } else { 1432 setValue(&I, DAG.getNode(FPOp, Op1.getValueType(), Op1, Op2)); 1433 } 1434} 1435 1436void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) { 1437 SDOperand Op1 = getValue(I.getOperand(0)); 1438 SDOperand Op2 = getValue(I.getOperand(1)); 1439 1440 Op2 = DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), Op2); 1441 1442 setValue(&I, DAG.getNode(Opcode, Op1.getValueType(), Op1, Op2)); 1443} 1444 1445void SelectionDAGLowering::visitICmp(User &I) { 1446 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE; 1447 if (ICmpInst *IC = dyn_cast<ICmpInst>(&I)) 1448 predicate = IC->getPredicate(); 1449 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I)) 1450 predicate = ICmpInst::Predicate(IC->getPredicate()); 1451 SDOperand Op1 = getValue(I.getOperand(0)); 1452 SDOperand Op2 = getValue(I.getOperand(1)); 1453 ISD::CondCode Opcode; 1454 switch (predicate) { 1455 case ICmpInst::ICMP_EQ : Opcode = ISD::SETEQ; break; 1456 case ICmpInst::ICMP_NE : Opcode = ISD::SETNE; break; 1457 case ICmpInst::ICMP_UGT : Opcode = ISD::SETUGT; break; 1458 case ICmpInst::ICMP_UGE : Opcode = ISD::SETUGE; break; 1459 case ICmpInst::ICMP_ULT : Opcode = ISD::SETULT; break; 1460 case ICmpInst::ICMP_ULE : Opcode = ISD::SETULE; break; 1461 case ICmpInst::ICMP_SGT : Opcode = ISD::SETGT; break; 1462 case ICmpInst::ICMP_SGE : Opcode = ISD::SETGE; break; 1463 case ICmpInst::ICMP_SLT : Opcode = ISD::SETLT; break; 1464 case ICmpInst::ICMP_SLE : Opcode = ISD::SETLE; break; 1465 default: 1466 assert(!"Invalid ICmp predicate value"); 1467 Opcode = ISD::SETEQ; 1468 break; 1469 } 1470 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Opcode)); 1471} 1472 1473void SelectionDAGLowering::visitFCmp(User &I) { 1474 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE; 1475 if (FCmpInst *FC = dyn_cast<FCmpInst>(&I)) 1476 predicate = FC->getPredicate(); 1477 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I)) 1478 predicate = FCmpInst::Predicate(FC->getPredicate()); 1479 SDOperand Op1 = getValue(I.getOperand(0)); 1480 SDOperand Op2 = getValue(I.getOperand(1)); 1481 ISD::CondCode Condition, FOC, FPC; 1482 switch (predicate) { 1483 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break; 1484 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break; 1485 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break; 1486 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break; 1487 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break; 1488 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break; 1489 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break; 1490 case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break; 1491 case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break; 1492 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break; 1493 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break; 1494 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break; 1495 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break; 1496 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break; 1497 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break; 1498 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break; 1499 default: 1500 assert(!"Invalid FCmp predicate value"); 1501 FOC = FPC = ISD::SETFALSE; 1502 break; 1503 } 1504 if (FiniteOnlyFPMath()) 1505 Condition = FOC; 1506 else 1507 Condition = FPC; 1508 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Condition)); 1509} 1510 1511void SelectionDAGLowering::visitSelect(User &I) { 1512 SDOperand Cond = getValue(I.getOperand(0)); 1513 SDOperand TrueVal = getValue(I.getOperand(1)); 1514 SDOperand FalseVal = getValue(I.getOperand(2)); 1515 if (!isa<PackedType>(I.getType())) { 1516 setValue(&I, DAG.getNode(ISD::SELECT, TrueVal.getValueType(), Cond, 1517 TrueVal, FalseVal)); 1518 } else { 1519 setValue(&I, DAG.getNode(ISD::VSELECT, MVT::Vector, Cond, TrueVal, FalseVal, 1520 *(TrueVal.Val->op_end()-2), 1521 *(TrueVal.Val->op_end()-1))); 1522 } 1523} 1524 1525 1526void SelectionDAGLowering::visitTrunc(User &I) { 1527 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest). 1528 SDOperand N = getValue(I.getOperand(0)); 1529 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1530 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N)); 1531} 1532 1533void SelectionDAGLowering::visitZExt(User &I) { 1534 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 1535 // ZExt also can't be a cast to bool for same reason. So, nothing much to do 1536 SDOperand N = getValue(I.getOperand(0)); 1537 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1538 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N)); 1539} 1540 1541void SelectionDAGLowering::visitSExt(User &I) { 1542 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest). 1543 // SExt also can't be a cast to bool for same reason. So, nothing much to do 1544 SDOperand N = getValue(I.getOperand(0)); 1545 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1546 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, DestVT, N)); 1547} 1548 1549void SelectionDAGLowering::visitFPTrunc(User &I) { 1550 // FPTrunc is never a no-op cast, no need to check 1551 SDOperand N = getValue(I.getOperand(0)); 1552 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1553 setValue(&I, DAG.getNode(ISD::FP_ROUND, DestVT, N)); 1554} 1555 1556void SelectionDAGLowering::visitFPExt(User &I){ 1557 // FPTrunc is never a no-op cast, no need to check 1558 SDOperand N = getValue(I.getOperand(0)); 1559 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1560 setValue(&I, DAG.getNode(ISD::FP_EXTEND, DestVT, N)); 1561} 1562 1563void SelectionDAGLowering::visitFPToUI(User &I) { 1564 // FPToUI is never a no-op cast, no need to check 1565 SDOperand N = getValue(I.getOperand(0)); 1566 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1567 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, DestVT, N)); 1568} 1569 1570void SelectionDAGLowering::visitFPToSI(User &I) { 1571 // FPToSI is never a no-op cast, no need to check 1572 SDOperand N = getValue(I.getOperand(0)); 1573 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1574 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, DestVT, N)); 1575} 1576 1577void SelectionDAGLowering::visitUIToFP(User &I) { 1578 // UIToFP is never a no-op cast, no need to check 1579 SDOperand N = getValue(I.getOperand(0)); 1580 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1581 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, DestVT, N)); 1582} 1583 1584void SelectionDAGLowering::visitSIToFP(User &I){ 1585 // UIToFP is never a no-op cast, no need to check 1586 SDOperand N = getValue(I.getOperand(0)); 1587 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1588 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, DestVT, N)); 1589} 1590 1591void SelectionDAGLowering::visitPtrToInt(User &I) { 1592 // What to do depends on the size of the integer and the size of the pointer. 1593 // We can either truncate, zero extend, or no-op, accordingly. 1594 SDOperand N = getValue(I.getOperand(0)); 1595 MVT::ValueType SrcVT = N.getValueType(); 1596 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1597 SDOperand Result; 1598 if (MVT::getSizeInBits(DestVT) < MVT::getSizeInBits(SrcVT)) 1599 Result = DAG.getNode(ISD::TRUNCATE, DestVT, N); 1600 else 1601 // Note: ZERO_EXTEND can handle cases where the sizes are equal too 1602 Result = DAG.getNode(ISD::ZERO_EXTEND, DestVT, N); 1603 setValue(&I, Result); 1604} 1605 1606void SelectionDAGLowering::visitIntToPtr(User &I) { 1607 // What to do depends on the size of the integer and the size of the pointer. 1608 // We can either truncate, zero extend, or no-op, accordingly. 1609 SDOperand N = getValue(I.getOperand(0)); 1610 MVT::ValueType SrcVT = N.getValueType(); 1611 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1612 if (MVT::getSizeInBits(DestVT) < MVT::getSizeInBits(SrcVT)) 1613 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N)); 1614 else 1615 // Note: ZERO_EXTEND can handle cases where the sizes are equal too 1616 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N)); 1617} 1618 1619void SelectionDAGLowering::visitBitCast(User &I) { 1620 SDOperand N = getValue(I.getOperand(0)); 1621 MVT::ValueType DestVT = TLI.getValueType(I.getType()); 1622 if (DestVT == MVT::Vector) { 1623 // This is a cast to a vector from something else. 1624 // Get information about the output vector. 1625 const PackedType *DestTy = cast<PackedType>(I.getType()); 1626 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType()); 1627 setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N, 1628 DAG.getConstant(DestTy->getNumElements(),MVT::i32), 1629 DAG.getValueType(EltVT))); 1630 return; 1631 } 1632 MVT::ValueType SrcVT = N.getValueType(); 1633 if (SrcVT == MVT::Vector) { 1634 // This is a cast from a vctor to something else. 1635 // Get information about the input vector. 1636 setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N)); 1637 return; 1638 } 1639 1640 // BitCast assures us that source and destination are the same size so this 1641 // is either a BIT_CONVERT or a no-op. 1642 if (DestVT != N.getValueType()) 1643 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, DestVT, N)); // convert types 1644 else 1645 setValue(&I, N); // noop cast. 1646} 1647 1648void SelectionDAGLowering::visitInsertElement(User &I) { 1649 SDOperand InVec = getValue(I.getOperand(0)); 1650 SDOperand InVal = getValue(I.getOperand(1)); 1651 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), 1652 getValue(I.getOperand(2))); 1653 1654 SDOperand Num = *(InVec.Val->op_end()-2); 1655 SDOperand Typ = *(InVec.Val->op_end()-1); 1656 setValue(&I, DAG.getNode(ISD::VINSERT_VECTOR_ELT, MVT::Vector, 1657 InVec, InVal, InIdx, Num, Typ)); 1658} 1659 1660void SelectionDAGLowering::visitExtractElement(User &I) { 1661 SDOperand InVec = getValue(I.getOperand(0)); 1662 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), 1663 getValue(I.getOperand(1))); 1664 SDOperand Typ = *(InVec.Val->op_end()-1); 1665 setValue(&I, DAG.getNode(ISD::VEXTRACT_VECTOR_ELT, 1666 TLI.getValueType(I.getType()), InVec, InIdx)); 1667} 1668 1669void SelectionDAGLowering::visitShuffleVector(User &I) { 1670 SDOperand V1 = getValue(I.getOperand(0)); 1671 SDOperand V2 = getValue(I.getOperand(1)); 1672 SDOperand Mask = getValue(I.getOperand(2)); 1673 1674 SDOperand Num = *(V1.Val->op_end()-2); 1675 SDOperand Typ = *(V2.Val->op_end()-1); 1676 setValue(&I, DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector, 1677 V1, V2, Mask, Num, Typ)); 1678} 1679 1680 1681void SelectionDAGLowering::visitGetElementPtr(User &I) { 1682 SDOperand N = getValue(I.getOperand(0)); 1683 const Type *Ty = I.getOperand(0)->getType(); 1684 1685 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end(); 1686 OI != E; ++OI) { 1687 Value *Idx = *OI; 1688 if (const StructType *StTy = dyn_cast<StructType>(Ty)) { 1689 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 1690 if (Field) { 1691 // N = N + Offset 1692 uint64_t Offset = TD->getStructLayout(StTy)->MemberOffsets[Field]; 1693 N = DAG.getNode(ISD::ADD, N.getValueType(), N, 1694 getIntPtrConstant(Offset)); 1695 } 1696 Ty = StTy->getElementType(Field); 1697 } else { 1698 Ty = cast<SequentialType>(Ty)->getElementType(); 1699 1700 // If this is a constant subscript, handle it quickly. 1701 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { 1702 if (CI->getZExtValue() == 0) continue; 1703 uint64_t Offs = 1704 TD->getTypeSize(Ty)*cast<ConstantInt>(CI)->getZExtValue(); 1705 N = DAG.getNode(ISD::ADD, N.getValueType(), N, getIntPtrConstant(Offs)); 1706 continue; 1707 } 1708 1709 // N = N + Idx * ElementSize; 1710 uint64_t ElementSize = TD->getTypeSize(Ty); 1711 SDOperand IdxN = getValue(Idx); 1712 1713 // If the index is smaller or larger than intptr_t, truncate or extend 1714 // it. 1715 if (IdxN.getValueType() < N.getValueType()) { 1716 IdxN = DAG.getNode(ISD::SIGN_EXTEND, N.getValueType(), IdxN); 1717 } else if (IdxN.getValueType() > N.getValueType()) 1718 IdxN = DAG.getNode(ISD::TRUNCATE, N.getValueType(), IdxN); 1719 1720 // If this is a multiply by a power of two, turn it into a shl 1721 // immediately. This is a very common case. 1722 if (isPowerOf2_64(ElementSize)) { 1723 unsigned Amt = Log2_64(ElementSize); 1724 IdxN = DAG.getNode(ISD::SHL, N.getValueType(), IdxN, 1725 DAG.getConstant(Amt, TLI.getShiftAmountTy())); 1726 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN); 1727 continue; 1728 } 1729 1730 SDOperand Scale = getIntPtrConstant(ElementSize); 1731 IdxN = DAG.getNode(ISD::MUL, N.getValueType(), IdxN, Scale); 1732 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN); 1733 } 1734 } 1735 setValue(&I, N); 1736} 1737 1738void SelectionDAGLowering::visitAlloca(AllocaInst &I) { 1739 // If this is a fixed sized alloca in the entry block of the function, 1740 // allocate it statically on the stack. 1741 if (FuncInfo.StaticAllocaMap.count(&I)) 1742 return; // getValue will auto-populate this. 1743 1744 const Type *Ty = I.getAllocatedType(); 1745 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty); 1746 unsigned Align = std::max((unsigned)TLI.getTargetData()->getTypeAlignment(Ty), 1747 I.getAlignment()); 1748 1749 SDOperand AllocSize = getValue(I.getArraySize()); 1750 MVT::ValueType IntPtr = TLI.getPointerTy(); 1751 if (IntPtr < AllocSize.getValueType()) 1752 AllocSize = DAG.getNode(ISD::TRUNCATE, IntPtr, AllocSize); 1753 else if (IntPtr > AllocSize.getValueType()) 1754 AllocSize = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, AllocSize); 1755 1756 AllocSize = DAG.getNode(ISD::MUL, IntPtr, AllocSize, 1757 getIntPtrConstant(TySize)); 1758 1759 // Handle alignment. If the requested alignment is less than or equal to the 1760 // stack alignment, ignore it and round the size of the allocation up to the 1761 // stack alignment size. If the size is greater than the stack alignment, we 1762 // note this in the DYNAMIC_STACKALLOC node. 1763 unsigned StackAlign = 1764 TLI.getTargetMachine().getFrameInfo()->getStackAlignment(); 1765 if (Align <= StackAlign) { 1766 Align = 0; 1767 // Add SA-1 to the size. 1768 AllocSize = DAG.getNode(ISD::ADD, AllocSize.getValueType(), AllocSize, 1769 getIntPtrConstant(StackAlign-1)); 1770 // Mask out the low bits for alignment purposes. 1771 AllocSize = DAG.getNode(ISD::AND, AllocSize.getValueType(), AllocSize, 1772 getIntPtrConstant(~(uint64_t)(StackAlign-1))); 1773 } 1774 1775 SDOperand Ops[] = { getRoot(), AllocSize, getIntPtrConstant(Align) }; 1776 const MVT::ValueType *VTs = DAG.getNodeValueTypes(AllocSize.getValueType(), 1777 MVT::Other); 1778 SDOperand DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, VTs, 2, Ops, 3); 1779 DAG.setRoot(setValue(&I, DSA).getValue(1)); 1780 1781 // Inform the Frame Information that we have just allocated a variable-sized 1782 // object. 1783 CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject(); 1784} 1785 1786void SelectionDAGLowering::visitLoad(LoadInst &I) { 1787 SDOperand Ptr = getValue(I.getOperand(0)); 1788 1789 SDOperand Root; 1790 if (I.isVolatile()) 1791 Root = getRoot(); 1792 else { 1793 // Do not serialize non-volatile loads against each other. 1794 Root = DAG.getRoot(); 1795 } 1796 1797 setValue(&I, getLoadFrom(I.getType(), Ptr, I.getOperand(0), 1798 Root, I.isVolatile())); 1799} 1800 1801SDOperand SelectionDAGLowering::getLoadFrom(const Type *Ty, SDOperand Ptr, 1802 const Value *SV, SDOperand Root, 1803 bool isVolatile) { 1804 SDOperand L; 1805 if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) { 1806 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType()); 1807 L = DAG.getVecLoad(PTy->getNumElements(), PVT, Root, Ptr, 1808 DAG.getSrcValue(SV)); 1809 } else { 1810 L = DAG.getLoad(TLI.getValueType(Ty), Root, Ptr, SV, 0, isVolatile); 1811 } 1812 1813 if (isVolatile) 1814 DAG.setRoot(L.getValue(1)); 1815 else 1816 PendingLoads.push_back(L.getValue(1)); 1817 1818 return L; 1819} 1820 1821 1822void SelectionDAGLowering::visitStore(StoreInst &I) { 1823 Value *SrcV = I.getOperand(0); 1824 SDOperand Src = getValue(SrcV); 1825 SDOperand Ptr = getValue(I.getOperand(1)); 1826 DAG.setRoot(DAG.getStore(getRoot(), Src, Ptr, I.getOperand(1), 0, 1827 I.isVolatile())); 1828} 1829 1830/// IntrinsicCannotAccessMemory - Return true if the specified intrinsic cannot 1831/// access memory and has no other side effects at all. 1832static bool IntrinsicCannotAccessMemory(unsigned IntrinsicID) { 1833#define GET_NO_MEMORY_INTRINSICS 1834#include "llvm/Intrinsics.gen" 1835#undef GET_NO_MEMORY_INTRINSICS 1836 return false; 1837} 1838 1839// IntrinsicOnlyReadsMemory - Return true if the specified intrinsic doesn't 1840// have any side-effects or if it only reads memory. 1841static bool IntrinsicOnlyReadsMemory(unsigned IntrinsicID) { 1842#define GET_SIDE_EFFECT_INFO 1843#include "llvm/Intrinsics.gen" 1844#undef GET_SIDE_EFFECT_INFO 1845 return false; 1846} 1847 1848/// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC 1849/// node. 1850void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I, 1851 unsigned Intrinsic) { 1852 bool HasChain = !IntrinsicCannotAccessMemory(Intrinsic); 1853 bool OnlyLoad = HasChain && IntrinsicOnlyReadsMemory(Intrinsic); 1854 1855 // Build the operand list. 1856 SmallVector<SDOperand, 8> Ops; 1857 if (HasChain) { // If this intrinsic has side-effects, chainify it. 1858 if (OnlyLoad) { 1859 // We don't need to serialize loads against other loads. 1860 Ops.push_back(DAG.getRoot()); 1861 } else { 1862 Ops.push_back(getRoot()); 1863 } 1864 } 1865 1866 // Add the intrinsic ID as an integer operand. 1867 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy())); 1868 1869 // Add all operands of the call to the operand list. 1870 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) { 1871 SDOperand Op = getValue(I.getOperand(i)); 1872 1873 // If this is a vector type, force it to the right packed type. 1874 if (Op.getValueType() == MVT::Vector) { 1875 const PackedType *OpTy = cast<PackedType>(I.getOperand(i)->getType()); 1876 MVT::ValueType EltVT = TLI.getValueType(OpTy->getElementType()); 1877 1878 MVT::ValueType VVT = MVT::getVectorType(EltVT, OpTy->getNumElements()); 1879 assert(VVT != MVT::Other && "Intrinsic uses a non-legal type?"); 1880 Op = DAG.getNode(ISD::VBIT_CONVERT, VVT, Op); 1881 } 1882 1883 assert(TLI.isTypeLegal(Op.getValueType()) && 1884 "Intrinsic uses a non-legal type?"); 1885 Ops.push_back(Op); 1886 } 1887 1888 std::vector<MVT::ValueType> VTs; 1889 if (I.getType() != Type::VoidTy) { 1890 MVT::ValueType VT = TLI.getValueType(I.getType()); 1891 if (VT == MVT::Vector) { 1892 const PackedType *DestTy = cast<PackedType>(I.getType()); 1893 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType()); 1894 1895 VT = MVT::getVectorType(EltVT, DestTy->getNumElements()); 1896 assert(VT != MVT::Other && "Intrinsic uses a non-legal type?"); 1897 } 1898 1899 assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?"); 1900 VTs.push_back(VT); 1901 } 1902 if (HasChain) 1903 VTs.push_back(MVT::Other); 1904 1905 const MVT::ValueType *VTList = DAG.getNodeValueTypes(VTs); 1906 1907 // Create the node. 1908 SDOperand Result; 1909 if (!HasChain) 1910 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VTList, VTs.size(), 1911 &Ops[0], Ops.size()); 1912 else if (I.getType() != Type::VoidTy) 1913 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, VTList, VTs.size(), 1914 &Ops[0], Ops.size()); 1915 else 1916 Result = DAG.getNode(ISD::INTRINSIC_VOID, VTList, VTs.size(), 1917 &Ops[0], Ops.size()); 1918 1919 if (HasChain) { 1920 SDOperand Chain = Result.getValue(Result.Val->getNumValues()-1); 1921 if (OnlyLoad) 1922 PendingLoads.push_back(Chain); 1923 else 1924 DAG.setRoot(Chain); 1925 } 1926 if (I.getType() != Type::VoidTy) { 1927 if (const PackedType *PTy = dyn_cast<PackedType>(I.getType())) { 1928 MVT::ValueType EVT = TLI.getValueType(PTy->getElementType()); 1929 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result, 1930 DAG.getConstant(PTy->getNumElements(), MVT::i32), 1931 DAG.getValueType(EVT)); 1932 } 1933 setValue(&I, Result); 1934 } 1935} 1936 1937/// visitIntrinsicCall - Lower the call to the specified intrinsic function. If 1938/// we want to emit this as a call to a named external function, return the name 1939/// otherwise lower it and return null. 1940const char * 1941SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) { 1942 switch (Intrinsic) { 1943 default: 1944 // By default, turn this into a target intrinsic node. 1945 visitTargetIntrinsic(I, Intrinsic); 1946 return 0; 1947 case Intrinsic::vastart: visitVAStart(I); return 0; 1948 case Intrinsic::vaend: visitVAEnd(I); return 0; 1949 case Intrinsic::vacopy: visitVACopy(I); return 0; 1950 case Intrinsic::returnaddress: visitFrameReturnAddress(I, false); return 0; 1951 case Intrinsic::frameaddress: visitFrameReturnAddress(I, true); return 0; 1952 case Intrinsic::setjmp: 1953 return "_setjmp"+!TLI.usesUnderscoreSetJmp(); 1954 break; 1955 case Intrinsic::longjmp: 1956 return "_longjmp"+!TLI.usesUnderscoreLongJmp(); 1957 break; 1958 case Intrinsic::memcpy_i32: 1959 case Intrinsic::memcpy_i64: 1960 visitMemIntrinsic(I, ISD::MEMCPY); 1961 return 0; 1962 case Intrinsic::memset_i32: 1963 case Intrinsic::memset_i64: 1964 visitMemIntrinsic(I, ISD::MEMSET); 1965 return 0; 1966 case Intrinsic::memmove_i32: 1967 case Intrinsic::memmove_i64: 1968 visitMemIntrinsic(I, ISD::MEMMOVE); 1969 return 0; 1970 1971 case Intrinsic::dbg_stoppoint: { 1972 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo(); 1973 DbgStopPointInst &SPI = cast<DbgStopPointInst>(I); 1974 if (DebugInfo && SPI.getContext() && DebugInfo->Verify(SPI.getContext())) { 1975 SDOperand Ops[5]; 1976 1977 Ops[0] = getRoot(); 1978 Ops[1] = getValue(SPI.getLineValue()); 1979 Ops[2] = getValue(SPI.getColumnValue()); 1980 1981 DebugInfoDesc *DD = DebugInfo->getDescFor(SPI.getContext()); 1982 assert(DD && "Not a debug information descriptor"); 1983 CompileUnitDesc *CompileUnit = cast<CompileUnitDesc>(DD); 1984 1985 Ops[3] = DAG.getString(CompileUnit->getFileName()); 1986 Ops[4] = DAG.getString(CompileUnit->getDirectory()); 1987 1988 DAG.setRoot(DAG.getNode(ISD::LOCATION, MVT::Other, Ops, 5)); 1989 } 1990 1991 return 0; 1992 } 1993 case Intrinsic::dbg_region_start: { 1994 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo(); 1995 DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I); 1996 if (DebugInfo && RSI.getContext() && DebugInfo->Verify(RSI.getContext())) { 1997 unsigned LabelID = DebugInfo->RecordRegionStart(RSI.getContext()); 1998 DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other, getRoot(), 1999 DAG.getConstant(LabelID, MVT::i32))); 2000 } 2001 2002 return 0; 2003 } 2004 case Intrinsic::dbg_region_end: { 2005 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo(); 2006 DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I); 2007 if (DebugInfo && REI.getContext() && DebugInfo->Verify(REI.getContext())) { 2008 unsigned LabelID = DebugInfo->RecordRegionEnd(REI.getContext()); 2009 DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other, 2010 getRoot(), DAG.getConstant(LabelID, MVT::i32))); 2011 } 2012 2013 return 0; 2014 } 2015 case Intrinsic::dbg_func_start: { 2016 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo(); 2017 DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I); 2018 if (DebugInfo && FSI.getSubprogram() && 2019 DebugInfo->Verify(FSI.getSubprogram())) { 2020 unsigned LabelID = DebugInfo->RecordRegionStart(FSI.getSubprogram()); 2021 DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other, 2022 getRoot(), DAG.getConstant(LabelID, MVT::i32))); 2023 } 2024 2025 return 0; 2026 } 2027 case Intrinsic::dbg_declare: { 2028 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo(); 2029 DbgDeclareInst &DI = cast<DbgDeclareInst>(I); 2030 if (DebugInfo && DI.getVariable() && DebugInfo->Verify(DI.getVariable())) { 2031 SDOperand AddressOp = getValue(DI.getAddress()); 2032 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(AddressOp)) 2033 DebugInfo->RecordVariable(DI.getVariable(), FI->getIndex()); 2034 } 2035 2036 return 0; 2037 } 2038 2039 case Intrinsic::isunordered_f32: 2040 case Intrinsic::isunordered_f64: 2041 setValue(&I, DAG.getSetCC(MVT::i1,getValue(I.getOperand(1)), 2042 getValue(I.getOperand(2)), ISD::SETUO)); 2043 return 0; 2044 2045 case Intrinsic::sqrt_f32: 2046 case Intrinsic::sqrt_f64: 2047 setValue(&I, DAG.getNode(ISD::FSQRT, 2048 getValue(I.getOperand(1)).getValueType(), 2049 getValue(I.getOperand(1)))); 2050 return 0; 2051 case Intrinsic::powi_f32: 2052 case Intrinsic::powi_f64: 2053 setValue(&I, DAG.getNode(ISD::FPOWI, 2054 getValue(I.getOperand(1)).getValueType(), 2055 getValue(I.getOperand(1)), 2056 getValue(I.getOperand(2)))); 2057 return 0; 2058 case Intrinsic::pcmarker: { 2059 SDOperand Tmp = getValue(I.getOperand(1)); 2060 DAG.setRoot(DAG.getNode(ISD::PCMARKER, MVT::Other, getRoot(), Tmp)); 2061 return 0; 2062 } 2063 case Intrinsic::readcyclecounter: { 2064 SDOperand Op = getRoot(); 2065 SDOperand Tmp = DAG.getNode(ISD::READCYCLECOUNTER, 2066 DAG.getNodeValueTypes(MVT::i64, MVT::Other), 2, 2067 &Op, 1); 2068 setValue(&I, Tmp); 2069 DAG.setRoot(Tmp.getValue(1)); 2070 return 0; 2071 } 2072 case Intrinsic::bswap_i16: 2073 case Intrinsic::bswap_i32: 2074 case Intrinsic::bswap_i64: 2075 setValue(&I, DAG.getNode(ISD::BSWAP, 2076 getValue(I.getOperand(1)).getValueType(), 2077 getValue(I.getOperand(1)))); 2078 return 0; 2079 case Intrinsic::cttz_i8: 2080 case Intrinsic::cttz_i16: 2081 case Intrinsic::cttz_i32: 2082 case Intrinsic::cttz_i64: 2083 setValue(&I, DAG.getNode(ISD::CTTZ, 2084 getValue(I.getOperand(1)).getValueType(), 2085 getValue(I.getOperand(1)))); 2086 return 0; 2087 case Intrinsic::ctlz_i8: 2088 case Intrinsic::ctlz_i16: 2089 case Intrinsic::ctlz_i32: 2090 case Intrinsic::ctlz_i64: 2091 setValue(&I, DAG.getNode(ISD::CTLZ, 2092 getValue(I.getOperand(1)).getValueType(), 2093 getValue(I.getOperand(1)))); 2094 return 0; 2095 case Intrinsic::ctpop_i8: 2096 case Intrinsic::ctpop_i16: 2097 case Intrinsic::ctpop_i32: 2098 case Intrinsic::ctpop_i64: 2099 setValue(&I, DAG.getNode(ISD::CTPOP, 2100 getValue(I.getOperand(1)).getValueType(), 2101 getValue(I.getOperand(1)))); 2102 return 0; 2103 case Intrinsic::stacksave: { 2104 SDOperand Op = getRoot(); 2105 SDOperand Tmp = DAG.getNode(ISD::STACKSAVE, 2106 DAG.getNodeValueTypes(TLI.getPointerTy(), MVT::Other), 2, &Op, 1); 2107 setValue(&I, Tmp); 2108 DAG.setRoot(Tmp.getValue(1)); 2109 return 0; 2110 } 2111 case Intrinsic::stackrestore: { 2112 SDOperand Tmp = getValue(I.getOperand(1)); 2113 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, MVT::Other, getRoot(), Tmp)); 2114 return 0; 2115 } 2116 case Intrinsic::prefetch: 2117 // FIXME: Currently discarding prefetches. 2118 return 0; 2119 } 2120} 2121 2122 2123void SelectionDAGLowering::visitCall(CallInst &I) { 2124 const char *RenameFn = 0; 2125 if (Function *F = I.getCalledFunction()) { 2126 if (F->isExternal()) 2127 if (unsigned IID = F->getIntrinsicID()) { 2128 RenameFn = visitIntrinsicCall(I, IID); 2129 if (!RenameFn) 2130 return; 2131 } else { // Not an LLVM intrinsic. 2132 const std::string &Name = F->getName(); 2133 if (Name[0] == 'c' && (Name == "copysign" || Name == "copysignf")) { 2134 if (I.getNumOperands() == 3 && // Basic sanity checks. 2135 I.getOperand(1)->getType()->isFloatingPoint() && 2136 I.getType() == I.getOperand(1)->getType() && 2137 I.getType() == I.getOperand(2)->getType()) { 2138 SDOperand LHS = getValue(I.getOperand(1)); 2139 SDOperand RHS = getValue(I.getOperand(2)); 2140 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, LHS.getValueType(), 2141 LHS, RHS)); 2142 return; 2143 } 2144 } else if (Name[0] == 'f' && (Name == "fabs" || Name == "fabsf")) { 2145 if (I.getNumOperands() == 2 && // Basic sanity checks. 2146 I.getOperand(1)->getType()->isFloatingPoint() && 2147 I.getType() == I.getOperand(1)->getType()) { 2148 SDOperand Tmp = getValue(I.getOperand(1)); 2149 setValue(&I, DAG.getNode(ISD::FABS, Tmp.getValueType(), Tmp)); 2150 return; 2151 } 2152 } else if (Name[0] == 's' && (Name == "sin" || Name == "sinf")) { 2153 if (I.getNumOperands() == 2 && // Basic sanity checks. 2154 I.getOperand(1)->getType()->isFloatingPoint() && 2155 I.getType() == I.getOperand(1)->getType()) { 2156 SDOperand Tmp = getValue(I.getOperand(1)); 2157 setValue(&I, DAG.getNode(ISD::FSIN, Tmp.getValueType(), Tmp)); 2158 return; 2159 } 2160 } else if (Name[0] == 'c' && (Name == "cos" || Name == "cosf")) { 2161 if (I.getNumOperands() == 2 && // Basic sanity checks. 2162 I.getOperand(1)->getType()->isFloatingPoint() && 2163 I.getType() == I.getOperand(1)->getType()) { 2164 SDOperand Tmp = getValue(I.getOperand(1)); 2165 setValue(&I, DAG.getNode(ISD::FCOS, Tmp.getValueType(), Tmp)); 2166 return; 2167 } 2168 } 2169 } 2170 } else if (isa<InlineAsm>(I.getOperand(0))) { 2171 visitInlineAsm(I); 2172 return; 2173 } 2174 2175 const PointerType *PT = cast<PointerType>(I.getCalledValue()->getType()); 2176 const FunctionType *FTy = cast<FunctionType>(PT->getElementType()); 2177 2178 SDOperand Callee; 2179 if (!RenameFn) 2180 Callee = getValue(I.getOperand(0)); 2181 else 2182 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy()); 2183 TargetLowering::ArgListTy Args; 2184 TargetLowering::ArgListEntry Entry; 2185 Args.reserve(I.getNumOperands()); 2186 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) { 2187 Value *Arg = I.getOperand(i); 2188 SDOperand ArgNode = getValue(Arg); 2189 Entry.Node = ArgNode; Entry.Ty = Arg->getType(); 2190 Entry.isSigned = FTy->paramHasAttr(i, FunctionType::SExtAttribute); 2191 Args.push_back(Entry); 2192 } 2193 2194 std::pair<SDOperand,SDOperand> Result = 2195 TLI.LowerCallTo(getRoot(), I.getType(), 2196 FTy->paramHasAttr(0,FunctionType::SExtAttribute), 2197 FTy->isVarArg(), I.getCallingConv(), I.isTailCall(), 2198 Callee, Args, DAG); 2199 if (I.getType() != Type::VoidTy) 2200 setValue(&I, Result.first); 2201 DAG.setRoot(Result.second); 2202} 2203 2204SDOperand RegsForValue::getCopyFromRegs(SelectionDAG &DAG, 2205 SDOperand &Chain, SDOperand &Flag)const{ 2206 SDOperand Val = DAG.getCopyFromReg(Chain, Regs[0], RegVT, Flag); 2207 Chain = Val.getValue(1); 2208 Flag = Val.getValue(2); 2209 2210 // If the result was expanded, copy from the top part. 2211 if (Regs.size() > 1) { 2212 assert(Regs.size() == 2 && 2213 "Cannot expand to more than 2 elts yet!"); 2214 SDOperand Hi = DAG.getCopyFromReg(Chain, Regs[1], RegVT, Flag); 2215 Chain = Hi.getValue(1); 2216 Flag = Hi.getValue(2); 2217 if (DAG.getTargetLoweringInfo().isLittleEndian()) 2218 return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Val, Hi); 2219 else 2220 return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Hi, Val); 2221 } 2222 2223 // Otherwise, if the return value was promoted or extended, truncate it to the 2224 // appropriate type. 2225 if (RegVT == ValueVT) 2226 return Val; 2227 2228 if (MVT::isInteger(RegVT)) { 2229 if (ValueVT < RegVT) 2230 return DAG.getNode(ISD::TRUNCATE, ValueVT, Val); 2231 else 2232 return DAG.getNode(ISD::ANY_EXTEND, ValueVT, Val); 2233 } else { 2234 return DAG.getNode(ISD::FP_ROUND, ValueVT, Val); 2235 } 2236} 2237 2238/// getCopyToRegs - Emit a series of CopyToReg nodes that copies the 2239/// specified value into the registers specified by this object. This uses 2240/// Chain/Flag as the input and updates them for the output Chain/Flag. 2241void RegsForValue::getCopyToRegs(SDOperand Val, SelectionDAG &DAG, 2242 SDOperand &Chain, SDOperand &Flag, 2243 MVT::ValueType PtrVT) const { 2244 if (Regs.size() == 1) { 2245 // If there is a single register and the types differ, this must be 2246 // a promotion. 2247 if (RegVT != ValueVT) { 2248 if (MVT::isInteger(RegVT)) { 2249 if (RegVT < ValueVT) 2250 Val = DAG.getNode(ISD::TRUNCATE, RegVT, Val); 2251 else 2252 Val = DAG.getNode(ISD::ANY_EXTEND, RegVT, Val); 2253 } else 2254 Val = DAG.getNode(ISD::FP_EXTEND, RegVT, Val); 2255 } 2256 Chain = DAG.getCopyToReg(Chain, Regs[0], Val, Flag); 2257 Flag = Chain.getValue(1); 2258 } else { 2259 std::vector<unsigned> R(Regs); 2260 if (!DAG.getTargetLoweringInfo().isLittleEndian()) 2261 std::reverse(R.begin(), R.end()); 2262 2263 for (unsigned i = 0, e = R.size(); i != e; ++i) { 2264 SDOperand Part = DAG.getNode(ISD::EXTRACT_ELEMENT, RegVT, Val, 2265 DAG.getConstant(i, PtrVT)); 2266 Chain = DAG.getCopyToReg(Chain, R[i], Part, Flag); 2267 Flag = Chain.getValue(1); 2268 } 2269 } 2270} 2271 2272/// AddInlineAsmOperands - Add this value to the specified inlineasm node 2273/// operand list. This adds the code marker and includes the number of 2274/// values added into it. 2275void RegsForValue::AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG, 2276 std::vector<SDOperand> &Ops) const { 2277 Ops.push_back(DAG.getConstant(Code | (Regs.size() << 3), MVT::i32)); 2278 for (unsigned i = 0, e = Regs.size(); i != e; ++i) 2279 Ops.push_back(DAG.getRegister(Regs[i], RegVT)); 2280} 2281 2282/// isAllocatableRegister - If the specified register is safe to allocate, 2283/// i.e. it isn't a stack pointer or some other special register, return the 2284/// register class for the register. Otherwise, return null. 2285static const TargetRegisterClass * 2286isAllocatableRegister(unsigned Reg, MachineFunction &MF, 2287 const TargetLowering &TLI, const MRegisterInfo *MRI) { 2288 MVT::ValueType FoundVT = MVT::Other; 2289 const TargetRegisterClass *FoundRC = 0; 2290 for (MRegisterInfo::regclass_iterator RCI = MRI->regclass_begin(), 2291 E = MRI->regclass_end(); RCI != E; ++RCI) { 2292 MVT::ValueType ThisVT = MVT::Other; 2293 2294 const TargetRegisterClass *RC = *RCI; 2295 // If none of the the value types for this register class are valid, we 2296 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2297 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 2298 I != E; ++I) { 2299 if (TLI.isTypeLegal(*I)) { 2300 // If we have already found this register in a different register class, 2301 // choose the one with the largest VT specified. For example, on 2302 // PowerPC, we favor f64 register classes over f32. 2303 if (FoundVT == MVT::Other || 2304 MVT::getSizeInBits(FoundVT) < MVT::getSizeInBits(*I)) { 2305 ThisVT = *I; 2306 break; 2307 } 2308 } 2309 } 2310 2311 if (ThisVT == MVT::Other) continue; 2312 2313 // NOTE: This isn't ideal. In particular, this might allocate the 2314 // frame pointer in functions that need it (due to them not being taken 2315 // out of allocation, because a variable sized allocation hasn't been seen 2316 // yet). This is a slight code pessimization, but should still work. 2317 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF), 2318 E = RC->allocation_order_end(MF); I != E; ++I) 2319 if (*I == Reg) { 2320 // We found a matching register class. Keep looking at others in case 2321 // we find one with larger registers that this physreg is also in. 2322 FoundRC = RC; 2323 FoundVT = ThisVT; 2324 break; 2325 } 2326 } 2327 return FoundRC; 2328} 2329 2330RegsForValue SelectionDAGLowering:: 2331GetRegistersForValue(const std::string &ConstrCode, 2332 MVT::ValueType VT, bool isOutReg, bool isInReg, 2333 std::set<unsigned> &OutputRegs, 2334 std::set<unsigned> &InputRegs) { 2335 std::pair<unsigned, const TargetRegisterClass*> PhysReg = 2336 TLI.getRegForInlineAsmConstraint(ConstrCode, VT); 2337 std::vector<unsigned> Regs; 2338 2339 unsigned NumRegs = VT != MVT::Other ? TLI.getNumElements(VT) : 1; 2340 MVT::ValueType RegVT; 2341 MVT::ValueType ValueVT = VT; 2342 2343 // If this is a constraint for a specific physical register, like {r17}, 2344 // assign it now. 2345 if (PhysReg.first) { 2346 if (VT == MVT::Other) 2347 ValueVT = *PhysReg.second->vt_begin(); 2348 2349 // Get the actual register value type. This is important, because the user 2350 // may have asked for (e.g.) the AX register in i32 type. We need to 2351 // remember that AX is actually i16 to get the right extension. 2352 RegVT = *PhysReg.second->vt_begin(); 2353 2354 // This is a explicit reference to a physical register. 2355 Regs.push_back(PhysReg.first); 2356 2357 // If this is an expanded reference, add the rest of the regs to Regs. 2358 if (NumRegs != 1) { 2359 TargetRegisterClass::iterator I = PhysReg.second->begin(); 2360 TargetRegisterClass::iterator E = PhysReg.second->end(); 2361 for (; *I != PhysReg.first; ++I) 2362 assert(I != E && "Didn't find reg!"); 2363 2364 // Already added the first reg. 2365 --NumRegs; ++I; 2366 for (; NumRegs; --NumRegs, ++I) { 2367 assert(I != E && "Ran out of registers to allocate!"); 2368 Regs.push_back(*I); 2369 } 2370 } 2371 return RegsForValue(Regs, RegVT, ValueVT); 2372 } 2373 2374 // Otherwise, if this was a reference to an LLVM register class, create vregs 2375 // for this reference. 2376 std::vector<unsigned> RegClassRegs; 2377 if (PhysReg.second) { 2378 // If this is an early clobber or tied register, our regalloc doesn't know 2379 // how to maintain the constraint. If it isn't, go ahead and create vreg 2380 // and let the regalloc do the right thing. 2381 if (!isOutReg || !isInReg) { 2382 if (VT == MVT::Other) 2383 ValueVT = *PhysReg.second->vt_begin(); 2384 RegVT = *PhysReg.second->vt_begin(); 2385 2386 // Create the appropriate number of virtual registers. 2387 SSARegMap *RegMap = DAG.getMachineFunction().getSSARegMap(); 2388 for (; NumRegs; --NumRegs) 2389 Regs.push_back(RegMap->createVirtualRegister(PhysReg.second)); 2390 2391 return RegsForValue(Regs, RegVT, ValueVT); 2392 } 2393 2394 // Otherwise, we can't allocate it. Let the code below figure out how to 2395 // maintain these constraints. 2396 RegClassRegs.assign(PhysReg.second->begin(), PhysReg.second->end()); 2397 2398 } else { 2399 // This is a reference to a register class that doesn't directly correspond 2400 // to an LLVM register class. Allocate NumRegs consecutive, available, 2401 // registers from the class. 2402 RegClassRegs = TLI.getRegClassForInlineAsmConstraint(ConstrCode, VT); 2403 } 2404 2405 const MRegisterInfo *MRI = DAG.getTarget().getRegisterInfo(); 2406 MachineFunction &MF = *CurMBB->getParent(); 2407 unsigned NumAllocated = 0; 2408 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) { 2409 unsigned Reg = RegClassRegs[i]; 2410 // See if this register is available. 2411 if ((isOutReg && OutputRegs.count(Reg)) || // Already used. 2412 (isInReg && InputRegs.count(Reg))) { // Already used. 2413 // Make sure we find consecutive registers. 2414 NumAllocated = 0; 2415 continue; 2416 } 2417 2418 // Check to see if this register is allocatable (i.e. don't give out the 2419 // stack pointer). 2420 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, MRI); 2421 if (!RC) { 2422 // Make sure we find consecutive registers. 2423 NumAllocated = 0; 2424 continue; 2425 } 2426 2427 // Okay, this register is good, we can use it. 2428 ++NumAllocated; 2429 2430 // If we allocated enough consecutive 2431 if (NumAllocated == NumRegs) { 2432 unsigned RegStart = (i-NumAllocated)+1; 2433 unsigned RegEnd = i+1; 2434 // Mark all of the allocated registers used. 2435 for (unsigned i = RegStart; i != RegEnd; ++i) { 2436 unsigned Reg = RegClassRegs[i]; 2437 Regs.push_back(Reg); 2438 if (isOutReg) OutputRegs.insert(Reg); // Mark reg used. 2439 if (isInReg) InputRegs.insert(Reg); // Mark reg used. 2440 } 2441 2442 return RegsForValue(Regs, *RC->vt_begin(), VT); 2443 } 2444 } 2445 2446 // Otherwise, we couldn't allocate enough registers for this. 2447 return RegsForValue(); 2448} 2449 2450 2451/// visitInlineAsm - Handle a call to an InlineAsm object. 2452/// 2453void SelectionDAGLowering::visitInlineAsm(CallInst &I) { 2454 InlineAsm *IA = cast<InlineAsm>(I.getOperand(0)); 2455 2456 SDOperand AsmStr = DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), 2457 MVT::Other); 2458 2459 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints(); 2460 std::vector<MVT::ValueType> ConstraintVTs; 2461 2462 /// AsmNodeOperands - A list of pairs. The first element is a register, the 2463 /// second is a bitfield where bit #0 is set if it is a use and bit #1 is set 2464 /// if it is a def of that register. 2465 std::vector<SDOperand> AsmNodeOperands; 2466 AsmNodeOperands.push_back(SDOperand()); // reserve space for input chain 2467 AsmNodeOperands.push_back(AsmStr); 2468 2469 SDOperand Chain = getRoot(); 2470 SDOperand Flag; 2471 2472 // We fully assign registers here at isel time. This is not optimal, but 2473 // should work. For register classes that correspond to LLVM classes, we 2474 // could let the LLVM RA do its thing, but we currently don't. Do a prepass 2475 // over the constraints, collecting fixed registers that we know we can't use. 2476 std::set<unsigned> OutputRegs, InputRegs; 2477 unsigned OpNum = 1; 2478 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) { 2479 assert(Constraints[i].Codes.size() == 1 && "Only handles one code so far!"); 2480 std::string &ConstraintCode = Constraints[i].Codes[0]; 2481 2482 MVT::ValueType OpVT; 2483 2484 // Compute the value type for each operand and add it to ConstraintVTs. 2485 switch (Constraints[i].Type) { 2486 case InlineAsm::isOutput: 2487 if (!Constraints[i].isIndirectOutput) { 2488 assert(I.getType() != Type::VoidTy && "Bad inline asm!"); 2489 OpVT = TLI.getValueType(I.getType()); 2490 } else { 2491 const Type *OpTy = I.getOperand(OpNum)->getType(); 2492 OpVT = TLI.getValueType(cast<PointerType>(OpTy)->getElementType()); 2493 OpNum++; // Consumes a call operand. 2494 } 2495 break; 2496 case InlineAsm::isInput: 2497 OpVT = TLI.getValueType(I.getOperand(OpNum)->getType()); 2498 OpNum++; // Consumes a call operand. 2499 break; 2500 case InlineAsm::isClobber: 2501 OpVT = MVT::Other; 2502 break; 2503 } 2504 2505 ConstraintVTs.push_back(OpVT); 2506 2507 if (TLI.getRegForInlineAsmConstraint(ConstraintCode, OpVT).first == 0) 2508 continue; // Not assigned a fixed reg. 2509 2510 // Build a list of regs that this operand uses. This always has a single 2511 // element for promoted/expanded operands. 2512 RegsForValue Regs = GetRegistersForValue(ConstraintCode, OpVT, 2513 false, false, 2514 OutputRegs, InputRegs); 2515 2516 switch (Constraints[i].Type) { 2517 case InlineAsm::isOutput: 2518 // We can't assign any other output to this register. 2519 OutputRegs.insert(Regs.Regs.begin(), Regs.Regs.end()); 2520 // If this is an early-clobber output, it cannot be assigned to the same 2521 // value as the input reg. 2522 if (Constraints[i].isEarlyClobber || Constraints[i].hasMatchingInput) 2523 InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end()); 2524 break; 2525 case InlineAsm::isInput: 2526 // We can't assign any other input to this register. 2527 InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end()); 2528 break; 2529 case InlineAsm::isClobber: 2530 // Clobbered regs cannot be used as inputs or outputs. 2531 InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end()); 2532 OutputRegs.insert(Regs.Regs.begin(), Regs.Regs.end()); 2533 break; 2534 } 2535 } 2536 2537 // Loop over all of the inputs, copying the operand values into the 2538 // appropriate registers and processing the output regs. 2539 RegsForValue RetValRegs; 2540 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit; 2541 OpNum = 1; 2542 2543 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) { 2544 assert(Constraints[i].Codes.size() == 1 && "Only handles one code so far!"); 2545 std::string &ConstraintCode = Constraints[i].Codes[0]; 2546 2547 switch (Constraints[i].Type) { 2548 case InlineAsm::isOutput: { 2549 TargetLowering::ConstraintType CTy = TargetLowering::C_RegisterClass; 2550 if (ConstraintCode.size() == 1) // not a physreg name. 2551 CTy = TLI.getConstraintType(ConstraintCode[0]); 2552 2553 if (CTy == TargetLowering::C_Memory) { 2554 // Memory output. 2555 SDOperand InOperandVal = getValue(I.getOperand(OpNum)); 2556 2557 // Check that the operand (the address to store to) isn't a float. 2558 if (!MVT::isInteger(InOperandVal.getValueType())) 2559 assert(0 && "MATCH FAIL!"); 2560 2561 if (!Constraints[i].isIndirectOutput) 2562 assert(0 && "MATCH FAIL!"); 2563 2564 OpNum++; // Consumes a call operand. 2565 2566 // Extend/truncate to the right pointer type if needed. 2567 MVT::ValueType PtrType = TLI.getPointerTy(); 2568 if (InOperandVal.getValueType() < PtrType) 2569 InOperandVal = DAG.getNode(ISD::ZERO_EXTEND, PtrType, InOperandVal); 2570 else if (InOperandVal.getValueType() > PtrType) 2571 InOperandVal = DAG.getNode(ISD::TRUNCATE, PtrType, InOperandVal); 2572 2573 // Add information to the INLINEASM node to know about this output. 2574 unsigned ResOpType = 4/*MEM*/ | (1 << 3); 2575 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32)); 2576 AsmNodeOperands.push_back(InOperandVal); 2577 break; 2578 } 2579 2580 // Otherwise, this is a register output. 2581 assert(CTy == TargetLowering::C_RegisterClass && "Unknown op type!"); 2582 2583 // If this is an early-clobber output, or if there is an input 2584 // constraint that matches this, we need to reserve the input register 2585 // so no other inputs allocate to it. 2586 bool UsesInputRegister = false; 2587 if (Constraints[i].isEarlyClobber || Constraints[i].hasMatchingInput) 2588 UsesInputRegister = true; 2589 2590 // Copy the output from the appropriate register. Find a register that 2591 // we can use. 2592 RegsForValue Regs = 2593 GetRegistersForValue(ConstraintCode, ConstraintVTs[i], 2594 true, UsesInputRegister, 2595 OutputRegs, InputRegs); 2596 if (Regs.Regs.empty()) { 2597 cerr << "Couldn't allocate output reg for contraint '" 2598 << ConstraintCode << "'!\n"; 2599 exit(1); 2600 } 2601 2602 if (!Constraints[i].isIndirectOutput) { 2603 assert(RetValRegs.Regs.empty() && 2604 "Cannot have multiple output constraints yet!"); 2605 assert(I.getType() != Type::VoidTy && "Bad inline asm!"); 2606 RetValRegs = Regs; 2607 } else { 2608 IndirectStoresToEmit.push_back(std::make_pair(Regs, 2609 I.getOperand(OpNum))); 2610 OpNum++; // Consumes a call operand. 2611 } 2612 2613 // Add information to the INLINEASM node to know that this register is 2614 // set. 2615 Regs.AddInlineAsmOperands(2 /*REGDEF*/, DAG, AsmNodeOperands); 2616 break; 2617 } 2618 case InlineAsm::isInput: { 2619 SDOperand InOperandVal = getValue(I.getOperand(OpNum)); 2620 OpNum++; // Consumes a call operand. 2621 2622 if (isdigit(ConstraintCode[0])) { // Matching constraint? 2623 // If this is required to match an output register we have already set, 2624 // just use its register. 2625 unsigned OperandNo = atoi(ConstraintCode.c_str()); 2626 2627 // Scan until we find the definition we already emitted of this operand. 2628 // When we find it, create a RegsForValue operand. 2629 unsigned CurOp = 2; // The first operand. 2630 for (; OperandNo; --OperandNo) { 2631 // Advance to the next operand. 2632 unsigned NumOps = 2633 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue(); 2634 assert(((NumOps & 7) == 2 /*REGDEF*/ || 2635 (NumOps & 7) == 4 /*MEM*/) && 2636 "Skipped past definitions?"); 2637 CurOp += (NumOps>>3)+1; 2638 } 2639 2640 unsigned NumOps = 2641 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue(); 2642 assert((NumOps & 7) == 2 /*REGDEF*/ && 2643 "Skipped past definitions?"); 2644 2645 // Add NumOps>>3 registers to MatchedRegs. 2646 RegsForValue MatchedRegs; 2647 MatchedRegs.ValueVT = InOperandVal.getValueType(); 2648 MatchedRegs.RegVT = AsmNodeOperands[CurOp+1].getValueType(); 2649 for (unsigned i = 0, e = NumOps>>3; i != e; ++i) { 2650 unsigned Reg=cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg(); 2651 MatchedRegs.Regs.push_back(Reg); 2652 } 2653 2654 // Use the produced MatchedRegs object to 2655 MatchedRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag, 2656 TLI.getPointerTy()); 2657 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, DAG, AsmNodeOperands); 2658 break; 2659 } 2660 2661 TargetLowering::ConstraintType CTy = TargetLowering::C_RegisterClass; 2662 if (ConstraintCode.size() == 1) // not a physreg name. 2663 CTy = TLI.getConstraintType(ConstraintCode[0]); 2664 2665 if (CTy == TargetLowering::C_Other) { 2666 InOperandVal = TLI.isOperandValidForConstraint(InOperandVal, 2667 ConstraintCode[0], DAG); 2668 if (!InOperandVal.Val) { 2669 cerr << "Invalid operand for inline asm constraint '" 2670 << ConstraintCode << "'!\n"; 2671 exit(1); 2672 } 2673 2674 // Add information to the INLINEASM node to know about this input. 2675 unsigned ResOpType = 3 /*IMM*/ | (1 << 3); 2676 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32)); 2677 AsmNodeOperands.push_back(InOperandVal); 2678 break; 2679 } else if (CTy == TargetLowering::C_Memory) { 2680 // Memory input. 2681 2682 // Check that the operand isn't a float. 2683 if (!MVT::isInteger(InOperandVal.getValueType())) 2684 assert(0 && "MATCH FAIL!"); 2685 2686 // Extend/truncate to the right pointer type if needed. 2687 MVT::ValueType PtrType = TLI.getPointerTy(); 2688 if (InOperandVal.getValueType() < PtrType) 2689 InOperandVal = DAG.getNode(ISD::ZERO_EXTEND, PtrType, InOperandVal); 2690 else if (InOperandVal.getValueType() > PtrType) 2691 InOperandVal = DAG.getNode(ISD::TRUNCATE, PtrType, InOperandVal); 2692 2693 // Add information to the INLINEASM node to know about this input. 2694 unsigned ResOpType = 4/*MEM*/ | (1 << 3); 2695 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32)); 2696 AsmNodeOperands.push_back(InOperandVal); 2697 break; 2698 } 2699 2700 assert(CTy == TargetLowering::C_RegisterClass && "Unknown op type!"); 2701 2702 // Copy the input into the appropriate registers. 2703 RegsForValue InRegs = 2704 GetRegistersForValue(ConstraintCode, ConstraintVTs[i], 2705 false, true, OutputRegs, InputRegs); 2706 // FIXME: should be match fail. 2707 assert(!InRegs.Regs.empty() && "Couldn't allocate input reg!"); 2708 2709 InRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag, TLI.getPointerTy()); 2710 2711 InRegs.AddInlineAsmOperands(1/*REGUSE*/, DAG, AsmNodeOperands); 2712 break; 2713 } 2714 case InlineAsm::isClobber: { 2715 RegsForValue ClobberedRegs = 2716 GetRegistersForValue(ConstraintCode, MVT::Other, false, false, 2717 OutputRegs, InputRegs); 2718 // Add the clobbered value to the operand list, so that the register 2719 // allocator is aware that the physreg got clobbered. 2720 if (!ClobberedRegs.Regs.empty()) 2721 ClobberedRegs.AddInlineAsmOperands(2/*REGDEF*/, DAG, AsmNodeOperands); 2722 break; 2723 } 2724 } 2725 } 2726 2727 // Finish up input operands. 2728 AsmNodeOperands[0] = Chain; 2729 if (Flag.Val) AsmNodeOperands.push_back(Flag); 2730 2731 Chain = DAG.getNode(ISD::INLINEASM, 2732 DAG.getNodeValueTypes(MVT::Other, MVT::Flag), 2, 2733 &AsmNodeOperands[0], AsmNodeOperands.size()); 2734 Flag = Chain.getValue(1); 2735 2736 // If this asm returns a register value, copy the result from that register 2737 // and set it as the value of the call. 2738 if (!RetValRegs.Regs.empty()) 2739 setValue(&I, RetValRegs.getCopyFromRegs(DAG, Chain, Flag)); 2740 2741 std::vector<std::pair<SDOperand, Value*> > StoresToEmit; 2742 2743 // Process indirect outputs, first output all of the flagged copies out of 2744 // physregs. 2745 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) { 2746 RegsForValue &OutRegs = IndirectStoresToEmit[i].first; 2747 Value *Ptr = IndirectStoresToEmit[i].second; 2748 SDOperand OutVal = OutRegs.getCopyFromRegs(DAG, Chain, Flag); 2749 StoresToEmit.push_back(std::make_pair(OutVal, Ptr)); 2750 } 2751 2752 // Emit the non-flagged stores from the physregs. 2753 SmallVector<SDOperand, 8> OutChains; 2754 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i) 2755 OutChains.push_back(DAG.getStore(Chain, StoresToEmit[i].first, 2756 getValue(StoresToEmit[i].second), 2757 StoresToEmit[i].second, 0)); 2758 if (!OutChains.empty()) 2759 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, 2760 &OutChains[0], OutChains.size()); 2761 DAG.setRoot(Chain); 2762} 2763 2764 2765void SelectionDAGLowering::visitMalloc(MallocInst &I) { 2766 SDOperand Src = getValue(I.getOperand(0)); 2767 2768 MVT::ValueType IntPtr = TLI.getPointerTy(); 2769 2770 if (IntPtr < Src.getValueType()) 2771 Src = DAG.getNode(ISD::TRUNCATE, IntPtr, Src); 2772 else if (IntPtr > Src.getValueType()) 2773 Src = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, Src); 2774 2775 // Scale the source by the type size. 2776 uint64_t ElementSize = TD->getTypeSize(I.getType()->getElementType()); 2777 Src = DAG.getNode(ISD::MUL, Src.getValueType(), 2778 Src, getIntPtrConstant(ElementSize)); 2779 2780 TargetLowering::ArgListTy Args; 2781 TargetLowering::ArgListEntry Entry; 2782 Entry.Node = Src; 2783 Entry.Ty = TLI.getTargetData()->getIntPtrType(); 2784 Entry.isSigned = false; 2785 Args.push_back(Entry); 2786 2787 std::pair<SDOperand,SDOperand> Result = 2788 TLI.LowerCallTo(getRoot(), I.getType(), false, false, CallingConv::C, true, 2789 DAG.getExternalSymbol("malloc", IntPtr), 2790 Args, DAG); 2791 setValue(&I, Result.first); // Pointers always fit in registers 2792 DAG.setRoot(Result.second); 2793} 2794 2795void SelectionDAGLowering::visitFree(FreeInst &I) { 2796 TargetLowering::ArgListTy Args; 2797 TargetLowering::ArgListEntry Entry; 2798 Entry.Node = getValue(I.getOperand(0)); 2799 Entry.Ty = TLI.getTargetData()->getIntPtrType(); 2800 Entry.isSigned = false; 2801 Args.push_back(Entry); 2802 MVT::ValueType IntPtr = TLI.getPointerTy(); 2803 std::pair<SDOperand,SDOperand> Result = 2804 TLI.LowerCallTo(getRoot(), Type::VoidTy, false, false, CallingConv::C, true, 2805 DAG.getExternalSymbol("free", IntPtr), Args, DAG); 2806 DAG.setRoot(Result.second); 2807} 2808 2809// InsertAtEndOfBasicBlock - This method should be implemented by targets that 2810// mark instructions with the 'usesCustomDAGSchedInserter' flag. These 2811// instructions are special in various ways, which require special support to 2812// insert. The specified MachineInstr is created but not inserted into any 2813// basic blocks, and the scheduler passes ownership of it to this method. 2814MachineBasicBlock *TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI, 2815 MachineBasicBlock *MBB) { 2816 cerr << "If a target marks an instruction with " 2817 << "'usesCustomDAGSchedInserter', it must implement " 2818 << "TargetLowering::InsertAtEndOfBasicBlock!\n"; 2819 abort(); 2820 return 0; 2821} 2822 2823void SelectionDAGLowering::visitVAStart(CallInst &I) { 2824 DAG.setRoot(DAG.getNode(ISD::VASTART, MVT::Other, getRoot(), 2825 getValue(I.getOperand(1)), 2826 DAG.getSrcValue(I.getOperand(1)))); 2827} 2828 2829void SelectionDAGLowering::visitVAArg(VAArgInst &I) { 2830 SDOperand V = DAG.getVAArg(TLI.getValueType(I.getType()), getRoot(), 2831 getValue(I.getOperand(0)), 2832 DAG.getSrcValue(I.getOperand(0))); 2833 setValue(&I, V); 2834 DAG.setRoot(V.getValue(1)); 2835} 2836 2837void SelectionDAGLowering::visitVAEnd(CallInst &I) { 2838 DAG.setRoot(DAG.getNode(ISD::VAEND, MVT::Other, getRoot(), 2839 getValue(I.getOperand(1)), 2840 DAG.getSrcValue(I.getOperand(1)))); 2841} 2842 2843void SelectionDAGLowering::visitVACopy(CallInst &I) { 2844 DAG.setRoot(DAG.getNode(ISD::VACOPY, MVT::Other, getRoot(), 2845 getValue(I.getOperand(1)), 2846 getValue(I.getOperand(2)), 2847 DAG.getSrcValue(I.getOperand(1)), 2848 DAG.getSrcValue(I.getOperand(2)))); 2849} 2850 2851/// ExpandScalarFormalArgs - Recursively expand the formal_argument node, either 2852/// bit_convert it or join a pair of them with a BUILD_PAIR when appropriate. 2853static SDOperand ExpandScalarFormalArgs(MVT::ValueType VT, SDNode *Arg, 2854 unsigned &i, SelectionDAG &DAG, 2855 TargetLowering &TLI) { 2856 if (TLI.getTypeAction(VT) != TargetLowering::Expand) 2857 return SDOperand(Arg, i++); 2858 2859 MVT::ValueType EVT = TLI.getTypeToTransformTo(VT); 2860 unsigned NumVals = MVT::getSizeInBits(VT) / MVT::getSizeInBits(EVT); 2861 if (NumVals == 1) { 2862 return DAG.getNode(ISD::BIT_CONVERT, VT, 2863 ExpandScalarFormalArgs(EVT, Arg, i, DAG, TLI)); 2864 } else if (NumVals == 2) { 2865 SDOperand Lo = ExpandScalarFormalArgs(EVT, Arg, i, DAG, TLI); 2866 SDOperand Hi = ExpandScalarFormalArgs(EVT, Arg, i, DAG, TLI); 2867 if (!TLI.isLittleEndian()) 2868 std::swap(Lo, Hi); 2869 return DAG.getNode(ISD::BUILD_PAIR, VT, Lo, Hi); 2870 } else { 2871 // Value scalarized into many values. Unimp for now. 2872 assert(0 && "Cannot expand i64 -> i16 yet!"); 2873 } 2874 return SDOperand(); 2875} 2876 2877/// TargetLowering::LowerArguments - This is the default LowerArguments 2878/// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all 2879/// targets are migrated to using FORMAL_ARGUMENTS, this hook should be 2880/// integrated into SDISel. 2881std::vector<SDOperand> 2882TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) { 2883 // Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node. 2884 std::vector<SDOperand> Ops; 2885 Ops.push_back(DAG.getRoot()); 2886 Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy())); 2887 Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy())); 2888 2889 // Add one result value for each formal argument. 2890 std::vector<MVT::ValueType> RetVals; 2891 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) { 2892 MVT::ValueType VT = getValueType(I->getType()); 2893 2894 switch (getTypeAction(VT)) { 2895 default: assert(0 && "Unknown type action!"); 2896 case Legal: 2897 RetVals.push_back(VT); 2898 break; 2899 case Promote: 2900 RetVals.push_back(getTypeToTransformTo(VT)); 2901 break; 2902 case Expand: 2903 if (VT != MVT::Vector) { 2904 // If this is a large integer, it needs to be broken up into small 2905 // integers. Figure out what the destination type is and how many small 2906 // integers it turns into. 2907 MVT::ValueType NVT = getTypeToExpandTo(VT); 2908 unsigned NumVals = getNumElements(VT); 2909 for (unsigned i = 0; i != NumVals; ++i) 2910 RetVals.push_back(NVT); 2911 } else { 2912 // Otherwise, this is a vector type. We only support legal vectors 2913 // right now. 2914 unsigned NumElems = cast<PackedType>(I->getType())->getNumElements(); 2915 const Type *EltTy = cast<PackedType>(I->getType())->getElementType(); 2916 2917 // Figure out if there is a Packed type corresponding to this Vector 2918 // type. If so, convert to the packed type. 2919 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems); 2920 if (TVT != MVT::Other && isTypeLegal(TVT)) { 2921 RetVals.push_back(TVT); 2922 } else { 2923 assert(0 && "Don't support illegal by-val vector arguments yet!"); 2924 } 2925 } 2926 break; 2927 } 2928 } 2929 2930 RetVals.push_back(MVT::Other); 2931 2932 // Create the node. 2933 SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS, 2934 DAG.getNodeValueTypes(RetVals), RetVals.size(), 2935 &Ops[0], Ops.size()).Val; 2936 2937 DAG.setRoot(SDOperand(Result, Result->getNumValues()-1)); 2938 2939 // Set up the return result vector. 2940 Ops.clear(); 2941 const FunctionType *FTy = F.getFunctionType(); 2942 unsigned i = 0; 2943 unsigned Idx = 1; 2944 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; 2945 ++I, ++Idx) { 2946 MVT::ValueType VT = getValueType(I->getType()); 2947 2948 switch (getTypeAction(VT)) { 2949 default: assert(0 && "Unknown type action!"); 2950 case Legal: 2951 Ops.push_back(SDOperand(Result, i++)); 2952 break; 2953 case Promote: { 2954 SDOperand Op(Result, i++); 2955 if (MVT::isInteger(VT)) { 2956 unsigned AssertOp = ISD::AssertSext; 2957 if (FTy->paramHasAttr(Idx, FunctionType::ZExtAttribute)) 2958 AssertOp = ISD::AssertZext; 2959 Op = DAG.getNode(AssertOp, Op.getValueType(), Op, DAG.getValueType(VT)); 2960 Op = DAG.getNode(ISD::TRUNCATE, VT, Op); 2961 } else { 2962 assert(MVT::isFloatingPoint(VT) && "Not int or FP?"); 2963 Op = DAG.getNode(ISD::FP_ROUND, VT, Op); 2964 } 2965 Ops.push_back(Op); 2966 break; 2967 } 2968 case Expand: 2969 if (VT != MVT::Vector) { 2970 // If this is a large integer or a floating point node that needs to be 2971 // expanded, it needs to be reassembled from small integers. Figure out 2972 // what the source elt type is and how many small integers it is. 2973 Ops.push_back(ExpandScalarFormalArgs(VT, Result, i, DAG, *this)); 2974 } else { 2975 // Otherwise, this is a vector type. We only support legal vectors 2976 // right now. 2977 const PackedType *PTy = cast<PackedType>(I->getType()); 2978 unsigned NumElems = PTy->getNumElements(); 2979 const Type *EltTy = PTy->getElementType(); 2980 2981 // Figure out if there is a Packed type corresponding to this Vector 2982 // type. If so, convert to the packed type. 2983 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems); 2984 if (TVT != MVT::Other && isTypeLegal(TVT)) { 2985 SDOperand N = SDOperand(Result, i++); 2986 // Handle copies from generic vectors to registers. 2987 N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N, 2988 DAG.getConstant(NumElems, MVT::i32), 2989 DAG.getValueType(getValueType(EltTy))); 2990 Ops.push_back(N); 2991 } else { 2992 assert(0 && "Don't support illegal by-val vector arguments yet!"); 2993 abort(); 2994 } 2995 } 2996 break; 2997 } 2998 } 2999 return Ops; 3000} 3001 3002 3003/// ExpandScalarCallArgs - Recursively expand call argument node by 3004/// bit_converting it or extract a pair of elements from the larger node. 3005static void ExpandScalarCallArgs(MVT::ValueType VT, SDOperand Arg, 3006 bool isSigned, 3007 SmallVector<SDOperand, 32> &Ops, 3008 SelectionDAG &DAG, 3009 TargetLowering &TLI) { 3010 if (TLI.getTypeAction(VT) != TargetLowering::Expand) { 3011 Ops.push_back(Arg); 3012 Ops.push_back(DAG.getConstant(isSigned, MVT::i32)); 3013 return; 3014 } 3015 3016 MVT::ValueType EVT = TLI.getTypeToTransformTo(VT); 3017 unsigned NumVals = MVT::getSizeInBits(VT) / MVT::getSizeInBits(EVT); 3018 if (NumVals == 1) { 3019 Arg = DAG.getNode(ISD::BIT_CONVERT, EVT, Arg); 3020 ExpandScalarCallArgs(EVT, Arg, isSigned, Ops, DAG, TLI); 3021 } else if (NumVals == 2) { 3022 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, EVT, Arg, 3023 DAG.getConstant(0, TLI.getPointerTy())); 3024 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, EVT, Arg, 3025 DAG.getConstant(1, TLI.getPointerTy())); 3026 if (!TLI.isLittleEndian()) 3027 std::swap(Lo, Hi); 3028 ExpandScalarCallArgs(EVT, Lo, isSigned, Ops, DAG, TLI); 3029 ExpandScalarCallArgs(EVT, Hi, isSigned, Ops, DAG, TLI); 3030 } else { 3031 // Value scalarized into many values. Unimp for now. 3032 assert(0 && "Cannot expand i64 -> i16 yet!"); 3033 } 3034} 3035 3036/// TargetLowering::LowerCallTo - This is the default LowerCallTo 3037/// implementation, which just inserts an ISD::CALL node, which is later custom 3038/// lowered by the target to something concrete. FIXME: When all targets are 3039/// migrated to using ISD::CALL, this hook should be integrated into SDISel. 3040std::pair<SDOperand, SDOperand> 3041TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy, 3042 bool RetTyIsSigned, bool isVarArg, 3043 unsigned CallingConv, bool isTailCall, 3044 SDOperand Callee, 3045 ArgListTy &Args, SelectionDAG &DAG) { 3046 SmallVector<SDOperand, 32> Ops; 3047 Ops.push_back(Chain); // Op#0 - Chain 3048 Ops.push_back(DAG.getConstant(CallingConv, getPointerTy())); // Op#1 - CC 3049 Ops.push_back(DAG.getConstant(isVarArg, getPointerTy())); // Op#2 - VarArg 3050 Ops.push_back(DAG.getConstant(isTailCall, getPointerTy())); // Op#3 - Tail 3051 Ops.push_back(Callee); 3052 3053 // Handle all of the outgoing arguments. 3054 for (unsigned i = 0, e = Args.size(); i != e; ++i) { 3055 MVT::ValueType VT = getValueType(Args[i].Ty); 3056 SDOperand Op = Args[i].Node; 3057 bool isSigned = Args[i].isSigned; 3058 switch (getTypeAction(VT)) { 3059 default: assert(0 && "Unknown type action!"); 3060 case Legal: 3061 Ops.push_back(Op); 3062 Ops.push_back(DAG.getConstant(isSigned, MVT::i32)); 3063 break; 3064 case Promote: 3065 if (MVT::isInteger(VT)) { 3066 unsigned ExtOp = isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND; 3067 Op = DAG.getNode(ExtOp, getTypeToTransformTo(VT), Op); 3068 } else { 3069 assert(MVT::isFloatingPoint(VT) && "Not int or FP?"); 3070 Op = DAG.getNode(ISD::FP_EXTEND, getTypeToTransformTo(VT), Op); 3071 } 3072 Ops.push_back(Op); 3073 Ops.push_back(DAG.getConstant(isSigned, MVT::i32)); 3074 break; 3075 case Expand: 3076 if (VT != MVT::Vector) { 3077 // If this is a large integer, it needs to be broken down into small 3078 // integers. Figure out what the source elt type is and how many small 3079 // integers it is. 3080 ExpandScalarCallArgs(VT, Op, isSigned, Ops, DAG, *this); 3081 } else { 3082 // Otherwise, this is a vector type. We only support legal vectors 3083 // right now. 3084 const PackedType *PTy = cast<PackedType>(Args[i].Ty); 3085 unsigned NumElems = PTy->getNumElements(); 3086 const Type *EltTy = PTy->getElementType(); 3087 3088 // Figure out if there is a Packed type corresponding to this Vector 3089 // type. If so, convert to the packed type. 3090 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems); 3091 if (TVT != MVT::Other && isTypeLegal(TVT)) { 3092 // Insert a VBIT_CONVERT of the MVT::Vector type to the packed type. 3093 Op = DAG.getNode(ISD::VBIT_CONVERT, TVT, Op); 3094 Ops.push_back(Op); 3095 Ops.push_back(DAG.getConstant(isSigned, MVT::i32)); 3096 } else { 3097 assert(0 && "Don't support illegal by-val vector call args yet!"); 3098 abort(); 3099 } 3100 } 3101 break; 3102 } 3103 } 3104 3105 // Figure out the result value types. 3106 SmallVector<MVT::ValueType, 4> RetTys; 3107 3108 if (RetTy != Type::VoidTy) { 3109 MVT::ValueType VT = getValueType(RetTy); 3110 switch (getTypeAction(VT)) { 3111 default: assert(0 && "Unknown type action!"); 3112 case Legal: 3113 RetTys.push_back(VT); 3114 break; 3115 case Promote: 3116 RetTys.push_back(getTypeToTransformTo(VT)); 3117 break; 3118 case Expand: 3119 if (VT != MVT::Vector) { 3120 // If this is a large integer, it needs to be reassembled from small 3121 // integers. Figure out what the source elt type is and how many small 3122 // integers it is. 3123 MVT::ValueType NVT = getTypeToExpandTo(VT); 3124 unsigned NumVals = getNumElements(VT); 3125 for (unsigned i = 0; i != NumVals; ++i) 3126 RetTys.push_back(NVT); 3127 } else { 3128 // Otherwise, this is a vector type. We only support legal vectors 3129 // right now. 3130 const PackedType *PTy = cast<PackedType>(RetTy); 3131 unsigned NumElems = PTy->getNumElements(); 3132 const Type *EltTy = PTy->getElementType(); 3133 3134 // Figure out if there is a Packed type corresponding to this Vector 3135 // type. If so, convert to the packed type. 3136 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems); 3137 if (TVT != MVT::Other && isTypeLegal(TVT)) { 3138 RetTys.push_back(TVT); 3139 } else { 3140 assert(0 && "Don't support illegal by-val vector call results yet!"); 3141 abort(); 3142 } 3143 } 3144 } 3145 } 3146 3147 RetTys.push_back(MVT::Other); // Always has a chain. 3148 3149 // Finally, create the CALL node. 3150 SDOperand Res = DAG.getNode(ISD::CALL, 3151 DAG.getVTList(&RetTys[0], RetTys.size()), 3152 &Ops[0], Ops.size()); 3153 3154 // This returns a pair of operands. The first element is the 3155 // return value for the function (if RetTy is not VoidTy). The second 3156 // element is the outgoing token chain. 3157 SDOperand ResVal; 3158 if (RetTys.size() != 1) { 3159 MVT::ValueType VT = getValueType(RetTy); 3160 if (RetTys.size() == 2) { 3161 ResVal = Res; 3162 3163 // If this value was promoted, truncate it down. 3164 if (ResVal.getValueType() != VT) { 3165 if (VT == MVT::Vector) { 3166 // Insert a VBITCONVERT to convert from the packed result type to the 3167 // MVT::Vector type. 3168 unsigned NumElems = cast<PackedType>(RetTy)->getNumElements(); 3169 const Type *EltTy = cast<PackedType>(RetTy)->getElementType(); 3170 3171 // Figure out if there is a Packed type corresponding to this Vector 3172 // type. If so, convert to the packed type. 3173 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems); 3174 if (TVT != MVT::Other && isTypeLegal(TVT)) { 3175 // Insert a VBIT_CONVERT of the FORMAL_ARGUMENTS to a 3176 // "N x PTyElementVT" MVT::Vector type. 3177 ResVal = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, ResVal, 3178 DAG.getConstant(NumElems, MVT::i32), 3179 DAG.getValueType(getValueType(EltTy))); 3180 } else { 3181 abort(); 3182 } 3183 } else if (MVT::isInteger(VT)) { 3184 unsigned AssertOp = ISD::AssertSext; 3185 if (!RetTyIsSigned) 3186 AssertOp = ISD::AssertZext; 3187 ResVal = DAG.getNode(AssertOp, ResVal.getValueType(), ResVal, 3188 DAG.getValueType(VT)); 3189 ResVal = DAG.getNode(ISD::TRUNCATE, VT, ResVal); 3190 } else { 3191 assert(MVT::isFloatingPoint(VT)); 3192 if (getTypeAction(VT) == Expand) 3193 ResVal = DAG.getNode(ISD::BIT_CONVERT, VT, ResVal); 3194 else 3195 ResVal = DAG.getNode(ISD::FP_ROUND, VT, ResVal); 3196 } 3197 } 3198 } else if (RetTys.size() == 3) { 3199 ResVal = DAG.getNode(ISD::BUILD_PAIR, VT, 3200 Res.getValue(0), Res.getValue(1)); 3201 3202 } else { 3203 assert(0 && "Case not handled yet!"); 3204 } 3205 } 3206 3207 return std::make_pair(ResVal, Res.getValue(Res.Val->getNumValues()-1)); 3208} 3209 3210 3211 3212// It is always conservatively correct for llvm.returnaddress and 3213// llvm.frameaddress to return 0. 3214// 3215// FIXME: Change this to insert a FRAMEADDR/RETURNADDR node, and have that be 3216// expanded to 0 if the target wants. 3217std::pair<SDOperand, SDOperand> 3218TargetLowering::LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, 3219 unsigned Depth, SelectionDAG &DAG) { 3220 return std::make_pair(DAG.getConstant(0, getPointerTy()), Chain); 3221} 3222 3223SDOperand TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) { 3224 assert(0 && "LowerOperation not implemented for this target!"); 3225 abort(); 3226 return SDOperand(); 3227} 3228 3229SDOperand TargetLowering::CustomPromoteOperation(SDOperand Op, 3230 SelectionDAG &DAG) { 3231 assert(0 && "CustomPromoteOperation not implemented for this target!"); 3232 abort(); 3233 return SDOperand(); 3234} 3235 3236void SelectionDAGLowering::visitFrameReturnAddress(CallInst &I, bool isFrame) { 3237 unsigned Depth = (unsigned)cast<ConstantInt>(I.getOperand(1))->getZExtValue(); 3238 std::pair<SDOperand,SDOperand> Result = 3239 TLI.LowerFrameReturnAddress(isFrame, getRoot(), Depth, DAG); 3240 setValue(&I, Result.first); 3241 DAG.setRoot(Result.second); 3242} 3243 3244/// getMemsetValue - Vectorized representation of the memset value 3245/// operand. 3246static SDOperand getMemsetValue(SDOperand Value, MVT::ValueType VT, 3247 SelectionDAG &DAG) { 3248 MVT::ValueType CurVT = VT; 3249 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) { 3250 uint64_t Val = C->getValue() & 255; 3251 unsigned Shift = 8; 3252 while (CurVT != MVT::i8) { 3253 Val = (Val << Shift) | Val; 3254 Shift <<= 1; 3255 CurVT = (MVT::ValueType)((unsigned)CurVT - 1); 3256 } 3257 return DAG.getConstant(Val, VT); 3258 } else { 3259 Value = DAG.getNode(ISD::ZERO_EXTEND, VT, Value); 3260 unsigned Shift = 8; 3261 while (CurVT != MVT::i8) { 3262 Value = 3263 DAG.getNode(ISD::OR, VT, 3264 DAG.getNode(ISD::SHL, VT, Value, 3265 DAG.getConstant(Shift, MVT::i8)), Value); 3266 Shift <<= 1; 3267 CurVT = (MVT::ValueType)((unsigned)CurVT - 1); 3268 } 3269 3270 return Value; 3271 } 3272} 3273 3274/// getMemsetStringVal - Similar to getMemsetValue. Except this is only 3275/// used when a memcpy is turned into a memset when the source is a constant 3276/// string ptr. 3277static SDOperand getMemsetStringVal(MVT::ValueType VT, 3278 SelectionDAG &DAG, TargetLowering &TLI, 3279 std::string &Str, unsigned Offset) { 3280 uint64_t Val = 0; 3281 unsigned MSB = getSizeInBits(VT) / 8; 3282 if (TLI.isLittleEndian()) 3283 Offset = Offset + MSB - 1; 3284 for (unsigned i = 0; i != MSB; ++i) { 3285 Val = (Val << 8) | (unsigned char)Str[Offset]; 3286 Offset += TLI.isLittleEndian() ? -1 : 1; 3287 } 3288 return DAG.getConstant(Val, VT); 3289} 3290 3291/// getMemBasePlusOffset - Returns base and offset node for the 3292static SDOperand getMemBasePlusOffset(SDOperand Base, unsigned Offset, 3293 SelectionDAG &DAG, TargetLowering &TLI) { 3294 MVT::ValueType VT = Base.getValueType(); 3295 return DAG.getNode(ISD::ADD, VT, Base, DAG.getConstant(Offset, VT)); 3296} 3297 3298/// MeetsMaxMemopRequirement - Determines if the number of memory ops required 3299/// to replace the memset / memcpy is below the threshold. It also returns the 3300/// types of the sequence of memory ops to perform memset / memcpy. 3301static bool MeetsMaxMemopRequirement(std::vector<MVT::ValueType> &MemOps, 3302 unsigned Limit, uint64_t Size, 3303 unsigned Align, TargetLowering &TLI) { 3304 MVT::ValueType VT; 3305 3306 if (TLI.allowsUnalignedMemoryAccesses()) { 3307 VT = MVT::i64; 3308 } else { 3309 switch (Align & 7) { 3310 case 0: 3311 VT = MVT::i64; 3312 break; 3313 case 4: 3314 VT = MVT::i32; 3315 break; 3316 case 2: 3317 VT = MVT::i16; 3318 break; 3319 default: 3320 VT = MVT::i8; 3321 break; 3322 } 3323 } 3324 3325 MVT::ValueType LVT = MVT::i64; 3326 while (!TLI.isTypeLegal(LVT)) 3327 LVT = (MVT::ValueType)((unsigned)LVT - 1); 3328 assert(MVT::isInteger(LVT)); 3329 3330 if (VT > LVT) 3331 VT = LVT; 3332 3333 unsigned NumMemOps = 0; 3334 while (Size != 0) { 3335 unsigned VTSize = getSizeInBits(VT) / 8; 3336 while (VTSize > Size) { 3337 VT = (MVT::ValueType)((unsigned)VT - 1); 3338 VTSize >>= 1; 3339 } 3340 assert(MVT::isInteger(VT)); 3341 3342 if (++NumMemOps > Limit) 3343 return false; 3344 MemOps.push_back(VT); 3345 Size -= VTSize; 3346 } 3347 3348 return true; 3349} 3350 3351void SelectionDAGLowering::visitMemIntrinsic(CallInst &I, unsigned Op) { 3352 SDOperand Op1 = getValue(I.getOperand(1)); 3353 SDOperand Op2 = getValue(I.getOperand(2)); 3354 SDOperand Op3 = getValue(I.getOperand(3)); 3355 SDOperand Op4 = getValue(I.getOperand(4)); 3356 unsigned Align = (unsigned)cast<ConstantSDNode>(Op4)->getValue(); 3357 if (Align == 0) Align = 1; 3358 3359 if (ConstantSDNode *Size = dyn_cast<ConstantSDNode>(Op3)) { 3360 std::vector<MVT::ValueType> MemOps; 3361 3362 // Expand memset / memcpy to a series of load / store ops 3363 // if the size operand falls below a certain threshold. 3364 SmallVector<SDOperand, 8> OutChains; 3365 switch (Op) { 3366 default: break; // Do nothing for now. 3367 case ISD::MEMSET: { 3368 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemset(), 3369 Size->getValue(), Align, TLI)) { 3370 unsigned NumMemOps = MemOps.size(); 3371 unsigned Offset = 0; 3372 for (unsigned i = 0; i < NumMemOps; i++) { 3373 MVT::ValueType VT = MemOps[i]; 3374 unsigned VTSize = getSizeInBits(VT) / 8; 3375 SDOperand Value = getMemsetValue(Op2, VT, DAG); 3376 SDOperand Store = DAG.getStore(getRoot(), Value, 3377 getMemBasePlusOffset(Op1, Offset, DAG, TLI), 3378 I.getOperand(1), Offset); 3379 OutChains.push_back(Store); 3380 Offset += VTSize; 3381 } 3382 } 3383 break; 3384 } 3385 case ISD::MEMCPY: { 3386 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemcpy(), 3387 Size->getValue(), Align, TLI)) { 3388 unsigned NumMemOps = MemOps.size(); 3389 unsigned SrcOff = 0, DstOff = 0, SrcDelta = 0; 3390 GlobalAddressSDNode *G = NULL; 3391 std::string Str; 3392 bool CopyFromStr = false; 3393 3394 if (Op2.getOpcode() == ISD::GlobalAddress) 3395 G = cast<GlobalAddressSDNode>(Op2); 3396 else if (Op2.getOpcode() == ISD::ADD && 3397 Op2.getOperand(0).getOpcode() == ISD::GlobalAddress && 3398 Op2.getOperand(1).getOpcode() == ISD::Constant) { 3399 G = cast<GlobalAddressSDNode>(Op2.getOperand(0)); 3400 SrcDelta = cast<ConstantSDNode>(Op2.getOperand(1))->getValue(); 3401 } 3402 if (G) { 3403 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal()); 3404 if (GV && GV->isConstant()) { 3405 Str = GV->getStringValue(false); 3406 if (!Str.empty()) { 3407 CopyFromStr = true; 3408 SrcOff += SrcDelta; 3409 } 3410 } 3411 } 3412 3413 for (unsigned i = 0; i < NumMemOps; i++) { 3414 MVT::ValueType VT = MemOps[i]; 3415 unsigned VTSize = getSizeInBits(VT) / 8; 3416 SDOperand Value, Chain, Store; 3417 3418 if (CopyFromStr) { 3419 Value = getMemsetStringVal(VT, DAG, TLI, Str, SrcOff); 3420 Chain = getRoot(); 3421 Store = 3422 DAG.getStore(Chain, Value, 3423 getMemBasePlusOffset(Op1, DstOff, DAG, TLI), 3424 I.getOperand(1), DstOff); 3425 } else { 3426 Value = DAG.getLoad(VT, getRoot(), 3427 getMemBasePlusOffset(Op2, SrcOff, DAG, TLI), 3428 I.getOperand(2), SrcOff); 3429 Chain = Value.getValue(1); 3430 Store = 3431 DAG.getStore(Chain, Value, 3432 getMemBasePlusOffset(Op1, DstOff, DAG, TLI), 3433 I.getOperand(1), DstOff); 3434 } 3435 OutChains.push_back(Store); 3436 SrcOff += VTSize; 3437 DstOff += VTSize; 3438 } 3439 } 3440 break; 3441 } 3442 } 3443 3444 if (!OutChains.empty()) { 3445 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, 3446 &OutChains[0], OutChains.size())); 3447 return; 3448 } 3449 } 3450 3451 DAG.setRoot(DAG.getNode(Op, MVT::Other, getRoot(), Op1, Op2, Op3, Op4)); 3452} 3453 3454//===----------------------------------------------------------------------===// 3455// SelectionDAGISel code 3456//===----------------------------------------------------------------------===// 3457 3458unsigned SelectionDAGISel::MakeReg(MVT::ValueType VT) { 3459 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT)); 3460} 3461 3462void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const { 3463 // FIXME: we only modify the CFG to split critical edges. This 3464 // updates dom and loop info. 3465 AU.addRequired<AliasAnalysis>(); 3466} 3467 3468 3469/// OptimizeNoopCopyExpression - We have determined that the specified cast 3470/// instruction is a noop copy (e.g. it's casting from one pointer type to 3471/// another, int->uint, or int->sbyte on PPC. 3472/// 3473/// Return true if any changes are made. 3474static bool OptimizeNoopCopyExpression(CastInst *CI) { 3475 BasicBlock *DefBB = CI->getParent(); 3476 3477 /// InsertedCasts - Only insert a cast in each block once. 3478 std::map<BasicBlock*, CastInst*> InsertedCasts; 3479 3480 bool MadeChange = false; 3481 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end(); 3482 UI != E; ) { 3483 Use &TheUse = UI.getUse(); 3484 Instruction *User = cast<Instruction>(*UI); 3485 3486 // Figure out which BB this cast is used in. For PHI's this is the 3487 // appropriate predecessor block. 3488 BasicBlock *UserBB = User->getParent(); 3489 if (PHINode *PN = dyn_cast<PHINode>(User)) { 3490 unsigned OpVal = UI.getOperandNo()/2; 3491 UserBB = PN->getIncomingBlock(OpVal); 3492 } 3493 3494 // Preincrement use iterator so we don't invalidate it. 3495 ++UI; 3496 3497 // If this user is in the same block as the cast, don't change the cast. 3498 if (UserBB == DefBB) continue; 3499 3500 // If we have already inserted a cast into this block, use it. 3501 CastInst *&InsertedCast = InsertedCasts[UserBB]; 3502 3503 if (!InsertedCast) { 3504 BasicBlock::iterator InsertPt = UserBB->begin(); 3505 while (isa<PHINode>(InsertPt)) ++InsertPt; 3506 3507 InsertedCast = 3508 CastInst::create(CI->getOpcode(), CI->getOperand(0), CI->getType(), "", 3509 InsertPt); 3510 MadeChange = true; 3511 } 3512 3513 // Replace a use of the cast with a use of the new casat. 3514 TheUse = InsertedCast; 3515 } 3516 3517 // If we removed all uses, nuke the cast. 3518 if (CI->use_empty()) 3519 CI->eraseFromParent(); 3520 3521 return MadeChange; 3522} 3523 3524/// InsertGEPComputeCode - Insert code into BB to compute Ptr+PtrOffset, 3525/// casting to the type of GEPI. 3526static Instruction *InsertGEPComputeCode(Instruction *&V, BasicBlock *BB, 3527 Instruction *GEPI, Value *Ptr, 3528 Value *PtrOffset) { 3529 if (V) return V; // Already computed. 3530 3531 // Figure out the insertion point 3532 BasicBlock::iterator InsertPt; 3533 if (BB == GEPI->getParent()) { 3534 // If GEP is already inserted into BB, insert right after the GEP. 3535 InsertPt = GEPI; 3536 ++InsertPt; 3537 } else { 3538 // Otherwise, insert at the top of BB, after any PHI nodes 3539 InsertPt = BB->begin(); 3540 while (isa<PHINode>(InsertPt)) ++InsertPt; 3541 } 3542 3543 // If Ptr is itself a cast, but in some other BB, emit a copy of the cast into 3544 // BB so that there is only one value live across basic blocks (the cast 3545 // operand). 3546 if (CastInst *CI = dyn_cast<CastInst>(Ptr)) 3547 if (CI->getParent() != BB && isa<PointerType>(CI->getOperand(0)->getType())) 3548 Ptr = CastInst::create(CI->getOpcode(), CI->getOperand(0), CI->getType(), 3549 "", InsertPt); 3550 3551 // Add the offset, cast it to the right type. 3552 Ptr = BinaryOperator::createAdd(Ptr, PtrOffset, "", InsertPt); 3553 // Ptr is an integer type, GEPI is pointer type ==> IntToPtr 3554 return V = CastInst::create(Instruction::IntToPtr, Ptr, GEPI->getType(), 3555 "", InsertPt); 3556} 3557 3558/// ReplaceUsesOfGEPInst - Replace all uses of RepPtr with inserted code to 3559/// compute its value. The RepPtr value can be computed with Ptr+PtrOffset. One 3560/// trivial way of doing this would be to evaluate Ptr+PtrOffset in RepPtr's 3561/// block, then ReplaceAllUsesWith'ing everything. However, we would prefer to 3562/// sink PtrOffset into user blocks where doing so will likely allow us to fold 3563/// the constant add into a load or store instruction. Additionally, if a user 3564/// is a pointer-pointer cast, we look through it to find its users. 3565static void ReplaceUsesOfGEPInst(Instruction *RepPtr, Value *Ptr, 3566 Constant *PtrOffset, BasicBlock *DefBB, 3567 GetElementPtrInst *GEPI, 3568 std::map<BasicBlock*,Instruction*> &InsertedExprs) { 3569 while (!RepPtr->use_empty()) { 3570 Instruction *User = cast<Instruction>(RepPtr->use_back()); 3571 3572 // If the user is a Pointer-Pointer cast, recurse. Only BitCast can be 3573 // used for a Pointer-Pointer cast. 3574 if (isa<BitCastInst>(User)) { 3575 ReplaceUsesOfGEPInst(User, Ptr, PtrOffset, DefBB, GEPI, InsertedExprs); 3576 3577 // Drop the use of RepPtr. The cast is dead. Don't delete it now, else we 3578 // could invalidate an iterator. 3579 User->setOperand(0, UndefValue::get(RepPtr->getType())); 3580 continue; 3581 } 3582 3583 // If this is a load of the pointer, or a store through the pointer, emit 3584 // the increment into the load/store block. 3585 Instruction *NewVal; 3586 if (isa<LoadInst>(User) || 3587 (isa<StoreInst>(User) && User->getOperand(0) != RepPtr)) { 3588 NewVal = InsertGEPComputeCode(InsertedExprs[User->getParent()], 3589 User->getParent(), GEPI, 3590 Ptr, PtrOffset); 3591 } else { 3592 // If this use is not foldable into the addressing mode, use a version 3593 // emitted in the GEP block. 3594 NewVal = InsertGEPComputeCode(InsertedExprs[DefBB], DefBB, GEPI, 3595 Ptr, PtrOffset); 3596 } 3597 3598 if (GEPI->getType() != RepPtr->getType()) { 3599 BasicBlock::iterator IP = NewVal; 3600 ++IP; 3601 // NewVal must be a GEP which must be pointer type, so BitCast 3602 NewVal = new BitCastInst(NewVal, RepPtr->getType(), "", IP); 3603 } 3604 User->replaceUsesOfWith(RepPtr, NewVal); 3605 } 3606} 3607 3608 3609/// OptimizeGEPExpression - Since we are doing basic-block-at-a-time instruction 3610/// selection, we want to be a bit careful about some things. In particular, if 3611/// we have a GEP instruction that is used in a different block than it is 3612/// defined, the addressing expression of the GEP cannot be folded into loads or 3613/// stores that use it. In this case, decompose the GEP and move constant 3614/// indices into blocks that use it. 3615static bool OptimizeGEPExpression(GetElementPtrInst *GEPI, 3616 const TargetData *TD) { 3617 // If this GEP is only used inside the block it is defined in, there is no 3618 // need to rewrite it. 3619 bool isUsedOutsideDefBB = false; 3620 BasicBlock *DefBB = GEPI->getParent(); 3621 for (Value::use_iterator UI = GEPI->use_begin(), E = GEPI->use_end(); 3622 UI != E; ++UI) { 3623 if (cast<Instruction>(*UI)->getParent() != DefBB) { 3624 isUsedOutsideDefBB = true; 3625 break; 3626 } 3627 } 3628 if (!isUsedOutsideDefBB) return false; 3629 3630 // If this GEP has no non-zero constant indices, there is nothing we can do, 3631 // ignore it. 3632 bool hasConstantIndex = false; 3633 bool hasVariableIndex = false; 3634 for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1, 3635 E = GEPI->op_end(); OI != E; ++OI) { 3636 if (ConstantInt *CI = dyn_cast<ConstantInt>(*OI)) { 3637 if (CI->getZExtValue()) { 3638 hasConstantIndex = true; 3639 break; 3640 } 3641 } else { 3642 hasVariableIndex = true; 3643 } 3644 } 3645 3646 // If this is a "GEP X, 0, 0, 0", turn this into a cast. 3647 if (!hasConstantIndex && !hasVariableIndex) { 3648 /// The GEP operand must be a pointer, so must its result -> BitCast 3649 Value *NC = new BitCastInst(GEPI->getOperand(0), GEPI->getType(), 3650 GEPI->getName(), GEPI); 3651 GEPI->replaceAllUsesWith(NC); 3652 GEPI->eraseFromParent(); 3653 return true; 3654 } 3655 3656 // If this is a GEP &Alloca, 0, 0, forward subst the frame index into uses. 3657 if (!hasConstantIndex && !isa<AllocaInst>(GEPI->getOperand(0))) 3658 return false; 3659 3660 // Otherwise, decompose the GEP instruction into multiplies and adds. Sum the 3661 // constant offset (which we now know is non-zero) and deal with it later. 3662 uint64_t ConstantOffset = 0; 3663 const Type *UIntPtrTy = TD->getIntPtrType(); 3664 Value *Ptr = new PtrToIntInst(GEPI->getOperand(0), UIntPtrTy, "", GEPI); 3665 const Type *Ty = GEPI->getOperand(0)->getType(); 3666 3667 for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1, 3668 E = GEPI->op_end(); OI != E; ++OI) { 3669 Value *Idx = *OI; 3670 if (const StructType *StTy = dyn_cast<StructType>(Ty)) { 3671 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue(); 3672 if (Field) 3673 ConstantOffset += TD->getStructLayout(StTy)->MemberOffsets[Field]; 3674 Ty = StTy->getElementType(Field); 3675 } else { 3676 Ty = cast<SequentialType>(Ty)->getElementType(); 3677 3678 // Handle constant subscripts. 3679 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) { 3680 if (CI->getZExtValue() == 0) continue; 3681 ConstantOffset += (int64_t)TD->getTypeSize(Ty)*CI->getSExtValue(); 3682 continue; 3683 } 3684 3685 // Ptr = Ptr + Idx * ElementSize; 3686 3687 // Cast Idx to UIntPtrTy if needed. 3688 Idx = CastInst::createIntegerCast(Idx, UIntPtrTy, true/*SExt*/, "", GEPI); 3689 3690 uint64_t ElementSize = TD->getTypeSize(Ty); 3691 // Mask off bits that should not be set. 3692 ElementSize &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits()); 3693 Constant *SizeCst = ConstantInt::get(UIntPtrTy, ElementSize); 3694 3695 // Multiply by the element size and add to the base. 3696 Idx = BinaryOperator::createMul(Idx, SizeCst, "", GEPI); 3697 Ptr = BinaryOperator::createAdd(Ptr, Idx, "", GEPI); 3698 } 3699 } 3700 3701 // Make sure that the offset fits in uintptr_t. 3702 ConstantOffset &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits()); 3703 Constant *PtrOffset = ConstantInt::get(UIntPtrTy, ConstantOffset); 3704 3705 // Okay, we have now emitted all of the variable index parts to the BB that 3706 // the GEP is defined in. Loop over all of the using instructions, inserting 3707 // an "add Ptr, ConstantOffset" into each block that uses it and update the 3708 // instruction to use the newly computed value, making GEPI dead. When the 3709 // user is a load or store instruction address, we emit the add into the user 3710 // block, otherwise we use a canonical version right next to the gep (these 3711 // won't be foldable as addresses, so we might as well share the computation). 3712 3713 std::map<BasicBlock*,Instruction*> InsertedExprs; 3714 ReplaceUsesOfGEPInst(GEPI, Ptr, PtrOffset, DefBB, GEPI, InsertedExprs); 3715 3716 // Finally, the GEP is dead, remove it. 3717 GEPI->eraseFromParent(); 3718 3719 return true; 3720} 3721 3722 3723/// SplitEdgeNicely - Split the critical edge from TI to it's specified 3724/// successor if it will improve codegen. We only do this if the successor has 3725/// phi nodes (otherwise critical edges are ok). If there is already another 3726/// predecessor of the succ that is empty (and thus has no phi nodes), use it 3727/// instead of introducing a new block. 3728static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) { 3729 BasicBlock *TIBB = TI->getParent(); 3730 BasicBlock *Dest = TI->getSuccessor(SuccNum); 3731 assert(isa<PHINode>(Dest->begin()) && 3732 "This should only be called if Dest has a PHI!"); 3733 3734 /// TIPHIValues - This array is lazily computed to determine the values of 3735 /// PHIs in Dest that TI would provide. 3736 std::vector<Value*> TIPHIValues; 3737 3738 // Check to see if Dest has any blocks that can be used as a split edge for 3739 // this terminator. 3740 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) { 3741 BasicBlock *Pred = *PI; 3742 // To be usable, the pred has to end with an uncond branch to the dest. 3743 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator()); 3744 if (!PredBr || !PredBr->isUnconditional() || 3745 // Must be empty other than the branch. 3746 &Pred->front() != PredBr) 3747 continue; 3748 3749 // Finally, since we know that Dest has phi nodes in it, we have to make 3750 // sure that jumping to Pred will have the same affect as going to Dest in 3751 // terms of PHI values. 3752 PHINode *PN; 3753 unsigned PHINo = 0; 3754 bool FoundMatch = true; 3755 for (BasicBlock::iterator I = Dest->begin(); 3756 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) { 3757 if (PHINo == TIPHIValues.size()) 3758 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB)); 3759 3760 // If the PHI entry doesn't work, we can't use this pred. 3761 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) { 3762 FoundMatch = false; 3763 break; 3764 } 3765 } 3766 3767 // If we found a workable predecessor, change TI to branch to Succ. 3768 if (FoundMatch) { 3769 Dest->removePredecessor(TIBB); 3770 TI->setSuccessor(SuccNum, Pred); 3771 return; 3772 } 3773 } 3774 3775 SplitCriticalEdge(TI, SuccNum, P, true); 3776} 3777 3778 3779bool SelectionDAGISel::runOnFunction(Function &Fn) { 3780 MachineFunction &MF = MachineFunction::construct(&Fn, TLI.getTargetMachine()); 3781 RegMap = MF.getSSARegMap(); 3782 DOUT << "\n\n\n=== " << Fn.getName() << "\n"; 3783 3784 // First, split all critical edges. 3785 // 3786 // In this pass we also look for GEP and cast instructions that are used 3787 // across basic blocks and rewrite them to improve basic-block-at-a-time 3788 // selection. 3789 // 3790 bool MadeChange = true; 3791 while (MadeChange) { 3792 MadeChange = false; 3793 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) { 3794 // Split all critical edges where the dest block has a PHI. 3795 TerminatorInst *BBTI = BB->getTerminator(); 3796 if (BBTI->getNumSuccessors() > 1) { 3797 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i) 3798 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) && 3799 isCriticalEdge(BBTI, i, true)) 3800 SplitEdgeNicely(BBTI, i, this); 3801 } 3802 3803 3804 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 3805 Instruction *I = BBI++; 3806 3807 if (CallInst *CI = dyn_cast<CallInst>(I)) { 3808 // If we found an inline asm expession, and if the target knows how to 3809 // lower it to normal LLVM code, do so now. 3810 if (isa<InlineAsm>(CI->getCalledValue())) 3811 if (const TargetAsmInfo *TAI = 3812 TLI.getTargetMachine().getTargetAsmInfo()) { 3813 if (TAI->ExpandInlineAsm(CI)) 3814 BBI = BB->begin(); 3815 } 3816 } else if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) { 3817 MadeChange |= OptimizeGEPExpression(GEPI, TLI.getTargetData()); 3818 } else if (CastInst *CI = dyn_cast<CastInst>(I)) { 3819 // If the source of the cast is a constant, then this should have 3820 // already been constant folded. The only reason NOT to constant fold 3821 // it is if something (e.g. LSR) was careful to place the constant 3822 // evaluation in a block other than then one that uses it (e.g. to hoist 3823 // the address of globals out of a loop). If this is the case, we don't 3824 // want to forward-subst the cast. 3825 if (isa<Constant>(CI->getOperand(0))) 3826 continue; 3827 3828 // If this is a noop copy, sink it into user blocks to reduce the number 3829 // of virtual registers that must be created and coallesced. 3830 MVT::ValueType SrcVT = TLI.getValueType(CI->getOperand(0)->getType()); 3831 MVT::ValueType DstVT = TLI.getValueType(CI->getType()); 3832 3833 // This is an fp<->int conversion? 3834 if (MVT::isInteger(SrcVT) != MVT::isInteger(DstVT)) 3835 continue; 3836 3837 // If this is an extension, it will be a zero or sign extension, which 3838 // isn't a noop. 3839 if (SrcVT < DstVT) continue; 3840 3841 // If these values will be promoted, find out what they will be promoted 3842 // to. This helps us consider truncates on PPC as noop copies when they 3843 // are. 3844 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) 3845 SrcVT = TLI.getTypeToTransformTo(SrcVT); 3846 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote) 3847 DstVT = TLI.getTypeToTransformTo(DstVT); 3848 3849 // If, after promotion, these are the same types, this is a noop copy. 3850 if (SrcVT == DstVT) 3851 MadeChange |= OptimizeNoopCopyExpression(CI); 3852 } 3853 } 3854 } 3855 } 3856 3857 FunctionLoweringInfo FuncInfo(TLI, Fn, MF); 3858 3859 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) 3860 SelectBasicBlock(I, MF, FuncInfo); 3861 3862 return true; 3863} 3864 3865SDOperand SelectionDAGLowering::CopyValueToVirtualRegister(Value *V, 3866 unsigned Reg) { 3867 SDOperand Op = getValue(V); 3868 assert((Op.getOpcode() != ISD::CopyFromReg || 3869 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) && 3870 "Copy from a reg to the same reg!"); 3871 3872 // If this type is not legal, we must make sure to not create an invalid 3873 // register use. 3874 MVT::ValueType SrcVT = Op.getValueType(); 3875 MVT::ValueType DestVT = TLI.getTypeToTransformTo(SrcVT); 3876 if (SrcVT == DestVT) { 3877 return DAG.getCopyToReg(getRoot(), Reg, Op); 3878 } else if (SrcVT == MVT::Vector) { 3879 // Handle copies from generic vectors to registers. 3880 MVT::ValueType PTyElementVT, PTyLegalElementVT; 3881 unsigned NE = TLI.getPackedTypeBreakdown(cast<PackedType>(V->getType()), 3882 PTyElementVT, PTyLegalElementVT); 3883 3884 // Insert a VBIT_CONVERT of the input vector to a "N x PTyElementVT" 3885 // MVT::Vector type. 3886 Op = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Op, 3887 DAG.getConstant(NE, MVT::i32), 3888 DAG.getValueType(PTyElementVT)); 3889 3890 // Loop over all of the elements of the resultant vector, 3891 // VEXTRACT_VECTOR_ELT'ing them, converting them to PTyLegalElementVT, then 3892 // copying them into output registers. 3893 SmallVector<SDOperand, 8> OutChains; 3894 SDOperand Root = getRoot(); 3895 for (unsigned i = 0; i != NE; ++i) { 3896 SDOperand Elt = DAG.getNode(ISD::VEXTRACT_VECTOR_ELT, PTyElementVT, 3897 Op, DAG.getConstant(i, TLI.getPointerTy())); 3898 if (PTyElementVT == PTyLegalElementVT) { 3899 // Elements are legal. 3900 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt)); 3901 } else if (PTyLegalElementVT > PTyElementVT) { 3902 // Elements are promoted. 3903 if (MVT::isFloatingPoint(PTyLegalElementVT)) 3904 Elt = DAG.getNode(ISD::FP_EXTEND, PTyLegalElementVT, Elt); 3905 else 3906 Elt = DAG.getNode(ISD::ANY_EXTEND, PTyLegalElementVT, Elt); 3907 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt)); 3908 } else { 3909 // Elements are expanded. 3910 // The src value is expanded into multiple registers. 3911 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT, 3912 Elt, DAG.getConstant(0, TLI.getPointerTy())); 3913 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT, 3914 Elt, DAG.getConstant(1, TLI.getPointerTy())); 3915 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Lo)); 3916 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Hi)); 3917 } 3918 } 3919 return DAG.getNode(ISD::TokenFactor, MVT::Other, 3920 &OutChains[0], OutChains.size()); 3921 } else if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote) { 3922 // The src value is promoted to the register. 3923 if (MVT::isFloatingPoint(SrcVT)) 3924 Op = DAG.getNode(ISD::FP_EXTEND, DestVT, Op); 3925 else 3926 Op = DAG.getNode(ISD::ANY_EXTEND, DestVT, Op); 3927 return DAG.getCopyToReg(getRoot(), Reg, Op); 3928 } else { 3929 DestVT = TLI.getTypeToExpandTo(SrcVT); 3930 unsigned NumVals = TLI.getNumElements(SrcVT); 3931 if (NumVals == 1) 3932 return DAG.getCopyToReg(getRoot(), Reg, 3933 DAG.getNode(ISD::BIT_CONVERT, DestVT, Op)); 3934 assert(NumVals == 2 && "1 to 4 (and more) expansion not implemented!"); 3935 // The src value is expanded into multiple registers. 3936 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT, 3937 Op, DAG.getConstant(0, TLI.getPointerTy())); 3938 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT, 3939 Op, DAG.getConstant(1, TLI.getPointerTy())); 3940 Op = DAG.getCopyToReg(getRoot(), Reg, Lo); 3941 return DAG.getCopyToReg(Op, Reg+1, Hi); 3942 } 3943} 3944 3945void SelectionDAGISel:: 3946LowerArguments(BasicBlock *BB, SelectionDAGLowering &SDL, 3947 std::vector<SDOperand> &UnorderedChains) { 3948 // If this is the entry block, emit arguments. 3949 Function &F = *BB->getParent(); 3950 FunctionLoweringInfo &FuncInfo = SDL.FuncInfo; 3951 SDOperand OldRoot = SDL.DAG.getRoot(); 3952 std::vector<SDOperand> Args = TLI.LowerArguments(F, SDL.DAG); 3953 3954 unsigned a = 0; 3955 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end(); 3956 AI != E; ++AI, ++a) 3957 if (!AI->use_empty()) { 3958 SDL.setValue(AI, Args[a]); 3959 3960 // If this argument is live outside of the entry block, insert a copy from 3961 // whereever we got it to the vreg that other BB's will reference it as. 3962 if (FuncInfo.ValueMap.count(AI)) { 3963 SDOperand Copy = 3964 SDL.CopyValueToVirtualRegister(AI, FuncInfo.ValueMap[AI]); 3965 UnorderedChains.push_back(Copy); 3966 } 3967 } 3968 3969 // Finally, if the target has anything special to do, allow it to do so. 3970 // FIXME: this should insert code into the DAG! 3971 EmitFunctionEntryCode(F, SDL.DAG.getMachineFunction()); 3972} 3973 3974void SelectionDAGISel::BuildSelectionDAG(SelectionDAG &DAG, BasicBlock *LLVMBB, 3975 std::vector<std::pair<MachineInstr*, unsigned> > &PHINodesToUpdate, 3976 FunctionLoweringInfo &FuncInfo) { 3977 SelectionDAGLowering SDL(DAG, TLI, FuncInfo); 3978 3979 std::vector<SDOperand> UnorderedChains; 3980 3981 // Lower any arguments needed in this block if this is the entry block. 3982 if (LLVMBB == &LLVMBB->getParent()->front()) 3983 LowerArguments(LLVMBB, SDL, UnorderedChains); 3984 3985 BB = FuncInfo.MBBMap[LLVMBB]; 3986 SDL.setCurrentBasicBlock(BB); 3987 3988 // Lower all of the non-terminator instructions. 3989 for (BasicBlock::iterator I = LLVMBB->begin(), E = --LLVMBB->end(); 3990 I != E; ++I) 3991 SDL.visit(*I); 3992 3993 // Ensure that all instructions which are used outside of their defining 3994 // blocks are available as virtual registers. 3995 for (BasicBlock::iterator I = LLVMBB->begin(), E = LLVMBB->end(); I != E;++I) 3996 if (!I->use_empty() && !isa<PHINode>(I)) { 3997 std::map<const Value*, unsigned>::iterator VMI =FuncInfo.ValueMap.find(I); 3998 if (VMI != FuncInfo.ValueMap.end()) 3999 UnorderedChains.push_back( 4000 SDL.CopyValueToVirtualRegister(I, VMI->second)); 4001 } 4002 4003 // Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to 4004 // ensure constants are generated when needed. Remember the virtual registers 4005 // that need to be added to the Machine PHI nodes as input. We cannot just 4006 // directly add them, because expansion might result in multiple MBB's for one 4007 // BB. As such, the start of the BB might correspond to a different MBB than 4008 // the end. 4009 // 4010 TerminatorInst *TI = LLVMBB->getTerminator(); 4011 4012 // Emit constants only once even if used by multiple PHI nodes. 4013 std::map<Constant*, unsigned> ConstantsOut; 4014 4015 // Vector bool would be better, but vector<bool> is really slow. 4016 std::vector<unsigned char> SuccsHandled; 4017 if (TI->getNumSuccessors()) 4018 SuccsHandled.resize(BB->getParent()->getNumBlockIDs()); 4019 4020 // Check successor nodes PHI nodes that expect a constant to be available from 4021 // this block. 4022 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) { 4023 BasicBlock *SuccBB = TI->getSuccessor(succ); 4024 if (!isa<PHINode>(SuccBB->begin())) continue; 4025 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB]; 4026 4027 // If this terminator has multiple identical successors (common for 4028 // switches), only handle each succ once. 4029 unsigned SuccMBBNo = SuccMBB->getNumber(); 4030 if (SuccsHandled[SuccMBBNo]) continue; 4031 SuccsHandled[SuccMBBNo] = true; 4032 4033 MachineBasicBlock::iterator MBBI = SuccMBB->begin(); 4034 PHINode *PN; 4035 4036 // At this point we know that there is a 1-1 correspondence between LLVM PHI 4037 // nodes and Machine PHI nodes, but the incoming operands have not been 4038 // emitted yet. 4039 for (BasicBlock::iterator I = SuccBB->begin(); 4040 (PN = dyn_cast<PHINode>(I)); ++I) { 4041 // Ignore dead phi's. 4042 if (PN->use_empty()) continue; 4043 4044 unsigned Reg; 4045 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB); 4046 4047 if (Constant *C = dyn_cast<Constant>(PHIOp)) { 4048 unsigned &RegOut = ConstantsOut[C]; 4049 if (RegOut == 0) { 4050 RegOut = FuncInfo.CreateRegForValue(C); 4051 UnorderedChains.push_back( 4052 SDL.CopyValueToVirtualRegister(C, RegOut)); 4053 } 4054 Reg = RegOut; 4055 } else { 4056 Reg = FuncInfo.ValueMap[PHIOp]; 4057 if (Reg == 0) { 4058 assert(isa<AllocaInst>(PHIOp) && 4059 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) && 4060 "Didn't codegen value into a register!??"); 4061 Reg = FuncInfo.CreateRegForValue(PHIOp); 4062 UnorderedChains.push_back( 4063 SDL.CopyValueToVirtualRegister(PHIOp, Reg)); 4064 } 4065 } 4066 4067 // Remember that this register needs to added to the machine PHI node as 4068 // the input for this MBB. 4069 MVT::ValueType VT = TLI.getValueType(PN->getType()); 4070 unsigned NumElements; 4071 if (VT != MVT::Vector) 4072 NumElements = TLI.getNumElements(VT); 4073 else { 4074 MVT::ValueType VT1,VT2; 4075 NumElements = 4076 TLI.getPackedTypeBreakdown(cast<PackedType>(PN->getType()), 4077 VT1, VT2); 4078 } 4079 for (unsigned i = 0, e = NumElements; i != e; ++i) 4080 PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i)); 4081 } 4082 } 4083 ConstantsOut.clear(); 4084 4085 // Turn all of the unordered chains into one factored node. 4086 if (!UnorderedChains.empty()) { 4087 SDOperand Root = SDL.getRoot(); 4088 if (Root.getOpcode() != ISD::EntryToken) { 4089 unsigned i = 0, e = UnorderedChains.size(); 4090 for (; i != e; ++i) { 4091 assert(UnorderedChains[i].Val->getNumOperands() > 1); 4092 if (UnorderedChains[i].Val->getOperand(0) == Root) 4093 break; // Don't add the root if we already indirectly depend on it. 4094 } 4095 4096 if (i == e) 4097 UnorderedChains.push_back(Root); 4098 } 4099 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, 4100 &UnorderedChains[0], UnorderedChains.size())); 4101 } 4102 4103 // Lower the terminator after the copies are emitted. 4104 SDL.visit(*LLVMBB->getTerminator()); 4105 4106 // Copy over any CaseBlock records that may now exist due to SwitchInst 4107 // lowering, as well as any jump table information. 4108 SwitchCases.clear(); 4109 SwitchCases = SDL.SwitchCases; 4110 JT = SDL.JT; 4111 4112 // Make sure the root of the DAG is up-to-date. 4113 DAG.setRoot(SDL.getRoot()); 4114} 4115 4116void SelectionDAGISel::CodeGenAndEmitDAG(SelectionDAG &DAG) { 4117 // Get alias analysis for load/store combining. 4118 AliasAnalysis &AA = getAnalysis<AliasAnalysis>(); 4119 4120 // Run the DAG combiner in pre-legalize mode. 4121 DAG.Combine(false, AA); 4122 4123 DOUT << "Lowered selection DAG:\n"; 4124 DEBUG(DAG.dump()); 4125 4126 // Second step, hack on the DAG until it only uses operations and types that 4127 // the target supports. 4128 DAG.Legalize(); 4129 4130 DOUT << "Legalized selection DAG:\n"; 4131 DEBUG(DAG.dump()); 4132 4133 // Run the DAG combiner in post-legalize mode. 4134 DAG.Combine(true, AA); 4135 4136 if (ViewISelDAGs) DAG.viewGraph(); 4137 4138 // Third, instruction select all of the operations to machine code, adding the 4139 // code to the MachineBasicBlock. 4140 InstructionSelectBasicBlock(DAG); 4141 4142 DOUT << "Selected machine code:\n"; 4143 DEBUG(BB->dump()); 4144} 4145 4146void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB, MachineFunction &MF, 4147 FunctionLoweringInfo &FuncInfo) { 4148 std::vector<std::pair<MachineInstr*, unsigned> > PHINodesToUpdate; 4149 { 4150 SelectionDAG DAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>()); 4151 CurDAG = &DAG; 4152 4153 // First step, lower LLVM code to some DAG. This DAG may use operations and 4154 // types that are not supported by the target. 4155 BuildSelectionDAG(DAG, LLVMBB, PHINodesToUpdate, FuncInfo); 4156 4157 // Second step, emit the lowered DAG as machine code. 4158 CodeGenAndEmitDAG(DAG); 4159 } 4160 4161 // Next, now that we know what the last MBB the LLVM BB expanded is, update 4162 // PHI nodes in successors. 4163 if (SwitchCases.empty() && JT.Reg == 0) { 4164 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) { 4165 MachineInstr *PHI = PHINodesToUpdate[i].first; 4166 assert(PHI->getOpcode() == TargetInstrInfo::PHI && 4167 "This is not a machine PHI node that we are updating!"); 4168 PHI->addRegOperand(PHINodesToUpdate[i].second, false); 4169 PHI->addMachineBasicBlockOperand(BB); 4170 } 4171 return; 4172 } 4173 4174 // If the JumpTable record is filled in, then we need to emit a jump table. 4175 // Updating the PHI nodes is tricky in this case, since we need to determine 4176 // whether the PHI is a successor of the range check MBB or the jump table MBB 4177 if (JT.Reg) { 4178 assert(SwitchCases.empty() && "Cannot have jump table and lowered switch"); 4179 SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>()); 4180 CurDAG = &SDAG; 4181 SelectionDAGLowering SDL(SDAG, TLI, FuncInfo); 4182 MachineBasicBlock *RangeBB = BB; 4183 // Set the current basic block to the mbb we wish to insert the code into 4184 BB = JT.MBB; 4185 SDL.setCurrentBasicBlock(BB); 4186 // Emit the code 4187 SDL.visitJumpTable(JT); 4188 SDAG.setRoot(SDL.getRoot()); 4189 CodeGenAndEmitDAG(SDAG); 4190 // Update PHI Nodes 4191 for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) { 4192 MachineInstr *PHI = PHINodesToUpdate[pi].first; 4193 MachineBasicBlock *PHIBB = PHI->getParent(); 4194 assert(PHI->getOpcode() == TargetInstrInfo::PHI && 4195 "This is not a machine PHI node that we are updating!"); 4196 if (PHIBB == JT.Default) { 4197 PHI->addRegOperand(PHINodesToUpdate[pi].second, false); 4198 PHI->addMachineBasicBlockOperand(RangeBB); 4199 } 4200 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) { 4201 PHI->addRegOperand(PHINodesToUpdate[pi].second, false); 4202 PHI->addMachineBasicBlockOperand(BB); 4203 } 4204 } 4205 return; 4206 } 4207 4208 // If the switch block involved a branch to one of the actual successors, we 4209 // need to update PHI nodes in that block. 4210 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) { 4211 MachineInstr *PHI = PHINodesToUpdate[i].first; 4212 assert(PHI->getOpcode() == TargetInstrInfo::PHI && 4213 "This is not a machine PHI node that we are updating!"); 4214 if (BB->isSuccessor(PHI->getParent())) { 4215 PHI->addRegOperand(PHINodesToUpdate[i].second, false); 4216 PHI->addMachineBasicBlockOperand(BB); 4217 } 4218 } 4219 4220 // If we generated any switch lowering information, build and codegen any 4221 // additional DAGs necessary. 4222 for (unsigned i = 0, e = SwitchCases.size(); i != e; ++i) { 4223 SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>()); 4224 CurDAG = &SDAG; 4225 SelectionDAGLowering SDL(SDAG, TLI, FuncInfo); 4226 4227 // Set the current basic block to the mbb we wish to insert the code into 4228 BB = SwitchCases[i].ThisBB; 4229 SDL.setCurrentBasicBlock(BB); 4230 4231 // Emit the code 4232 SDL.visitSwitchCase(SwitchCases[i]); 4233 SDAG.setRoot(SDL.getRoot()); 4234 CodeGenAndEmitDAG(SDAG); 4235 4236 // Handle any PHI nodes in successors of this chunk, as if we were coming 4237 // from the original BB before switch expansion. Note that PHI nodes can 4238 // occur multiple times in PHINodesToUpdate. We have to be very careful to 4239 // handle them the right number of times. 4240 while ((BB = SwitchCases[i].TrueBB)) { // Handle LHS and RHS. 4241 for (MachineBasicBlock::iterator Phi = BB->begin(); 4242 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){ 4243 // This value for this PHI node is recorded in PHINodesToUpdate, get it. 4244 for (unsigned pn = 0; ; ++pn) { 4245 assert(pn != PHINodesToUpdate.size() && "Didn't find PHI entry!"); 4246 if (PHINodesToUpdate[pn].first == Phi) { 4247 Phi->addRegOperand(PHINodesToUpdate[pn].second, false); 4248 Phi->addMachineBasicBlockOperand(SwitchCases[i].ThisBB); 4249 break; 4250 } 4251 } 4252 } 4253 4254 // Don't process RHS if same block as LHS. 4255 if (BB == SwitchCases[i].FalseBB) 4256 SwitchCases[i].FalseBB = 0; 4257 4258 // If we haven't handled the RHS, do so now. Otherwise, we're done. 4259 SwitchCases[i].TrueBB = SwitchCases[i].FalseBB; 4260 SwitchCases[i].FalseBB = 0; 4261 } 4262 assert(SwitchCases[i].TrueBB == 0 && SwitchCases[i].FalseBB == 0); 4263 } 4264} 4265 4266 4267//===----------------------------------------------------------------------===// 4268/// ScheduleAndEmitDAG - Pick a safe ordering and emit instructions for each 4269/// target node in the graph. 4270void SelectionDAGISel::ScheduleAndEmitDAG(SelectionDAG &DAG) { 4271 if (ViewSchedDAGs) DAG.viewGraph(); 4272 4273 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault(); 4274 4275 if (!Ctor) { 4276 Ctor = ISHeuristic; 4277 RegisterScheduler::setDefault(Ctor); 4278 } 4279 4280 ScheduleDAG *SL = Ctor(this, &DAG, BB); 4281 BB = SL->Run(); 4282 delete SL; 4283} 4284 4285 4286HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() { 4287 return new HazardRecognizer(); 4288} 4289 4290//===----------------------------------------------------------------------===// 4291// Helper functions used by the generated instruction selector. 4292//===----------------------------------------------------------------------===// 4293// Calls to these methods are generated by tblgen. 4294 4295/// CheckAndMask - The isel is trying to match something like (and X, 255). If 4296/// the dag combiner simplified the 255, we still want to match. RHS is the 4297/// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value 4298/// specified in the .td file (e.g. 255). 4299bool SelectionDAGISel::CheckAndMask(SDOperand LHS, ConstantSDNode *RHS, 4300 int64_t DesiredMaskS) { 4301 uint64_t ActualMask = RHS->getValue(); 4302 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType()); 4303 4304 // If the actual mask exactly matches, success! 4305 if (ActualMask == DesiredMask) 4306 return true; 4307 4308 // If the actual AND mask is allowing unallowed bits, this doesn't match. 4309 if (ActualMask & ~DesiredMask) 4310 return false; 4311 4312 // Otherwise, the DAG Combiner may have proven that the value coming in is 4313 // either already zero or is not demanded. Check for known zero input bits. 4314 uint64_t NeededMask = DesiredMask & ~ActualMask; 4315 if (getTargetLowering().MaskedValueIsZero(LHS, NeededMask)) 4316 return true; 4317 4318 // TODO: check to see if missing bits are just not demanded. 4319 4320 // Otherwise, this pattern doesn't match. 4321 return false; 4322} 4323 4324/// CheckOrMask - The isel is trying to match something like (or X, 255). If 4325/// the dag combiner simplified the 255, we still want to match. RHS is the 4326/// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value 4327/// specified in the .td file (e.g. 255). 4328bool SelectionDAGISel::CheckOrMask(SDOperand LHS, ConstantSDNode *RHS, 4329 int64_t DesiredMaskS) { 4330 uint64_t ActualMask = RHS->getValue(); 4331 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType()); 4332 4333 // If the actual mask exactly matches, success! 4334 if (ActualMask == DesiredMask) 4335 return true; 4336 4337 // If the actual AND mask is allowing unallowed bits, this doesn't match. 4338 if (ActualMask & ~DesiredMask) 4339 return false; 4340 4341 // Otherwise, the DAG Combiner may have proven that the value coming in is 4342 // either already zero or is not demanded. Check for known zero input bits. 4343 uint64_t NeededMask = DesiredMask & ~ActualMask; 4344 4345 uint64_t KnownZero, KnownOne; 4346 getTargetLowering().ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne); 4347 4348 // If all the missing bits in the or are already known to be set, match! 4349 if ((NeededMask & KnownOne) == NeededMask) 4350 return true; 4351 4352 // TODO: check to see if missing bits are just not demanded. 4353 4354 // Otherwise, this pattern doesn't match. 4355 return false; 4356} 4357 4358 4359/// SelectInlineAsmMemoryOperands - Calls to this are automatically generated 4360/// by tblgen. Others should not call it. 4361void SelectionDAGISel:: 4362SelectInlineAsmMemoryOperands(std::vector<SDOperand> &Ops, SelectionDAG &DAG) { 4363 std::vector<SDOperand> InOps; 4364 std::swap(InOps, Ops); 4365 4366 Ops.push_back(InOps[0]); // input chain. 4367 Ops.push_back(InOps[1]); // input asm string. 4368 4369 unsigned i = 2, e = InOps.size(); 4370 if (InOps[e-1].getValueType() == MVT::Flag) 4371 --e; // Don't process a flag operand if it is here. 4372 4373 while (i != e) { 4374 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue(); 4375 if ((Flags & 7) != 4 /*MEM*/) { 4376 // Just skip over this operand, copying the operands verbatim. 4377 Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1); 4378 i += (Flags >> 3) + 1; 4379 } else { 4380 assert((Flags >> 3) == 1 && "Memory operand with multiple values?"); 4381 // Otherwise, this is a memory operand. Ask the target to select it. 4382 std::vector<SDOperand> SelOps; 4383 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps, DAG)) { 4384 cerr << "Could not match memory address. Inline asm failure!\n"; 4385 exit(1); 4386 } 4387 4388 // Add this to the output node. 4389 Ops.push_back(DAG.getTargetConstant(4/*MEM*/ | (SelOps.size() << 3), 4390 MVT::i32)); 4391 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end()); 4392 i += 2; 4393 } 4394 } 4395 4396 // Add the flag input back if present. 4397 if (e != InOps.size()) 4398 Ops.push_back(InOps.back()); 4399} 4400