TargetLowering.h revision 732f05c41f177a0bc4d47e93a5d02120f146cb4c
1//===-- llvm/Target/TargetLowering.h - Target Lowering Info -----*- C++ -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file describes how to lower LLVM code to machine code. This has two 11// main components: 12// 13// 1. Which ValueTypes are natively supported by the target. 14// 2. Which operations are supported for supported ValueTypes. 15// 3. Cost thresholds for alternative implementations of certain operations. 16// 17// In addition it has a few other components, like information about FP 18// immediates. 19// 20//===----------------------------------------------------------------------===// 21 22#ifndef LLVM_TARGET_TARGETLOWERING_H 23#define LLVM_TARGET_TARGETLOWERING_H 24 25#include "llvm/CallingConv.h" 26#include "llvm/InlineAsm.h" 27#include "llvm/Attributes.h" 28#include "llvm/ADT/SmallPtrSet.h" 29#include "llvm/CodeGen/SelectionDAGNodes.h" 30#include "llvm/CodeGen/RuntimeLibcalls.h" 31#include "llvm/Support/DebugLoc.h" 32#include "llvm/Target/TargetCallingConv.h" 33#include "llvm/Target/TargetMachine.h" 34#include <climits> 35#include <map> 36#include <vector> 37 38namespace llvm { 39 class CallInst; 40 class CCState; 41 class FastISel; 42 class FunctionLoweringInfo; 43 class ImmutableCallSite; 44 class MachineBasicBlock; 45 class MachineFunction; 46 class MachineInstr; 47 class MachineJumpTableInfo; 48 class MCContext; 49 class MCExpr; 50 template<typename T> class SmallVectorImpl; 51 class TargetData; 52 class TargetRegisterClass; 53 class TargetLoweringObjectFile; 54 class Value; 55 56 // FIXME: should this be here? 57 namespace TLSModel { 58 enum Model { 59 GeneralDynamic, 60 LocalDynamic, 61 InitialExec, 62 LocalExec 63 }; 64 } 65 TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc); 66 67 68//===----------------------------------------------------------------------===// 69/// TargetLowering - This class defines information used to lower LLVM code to 70/// legal SelectionDAG operators that the target instruction selector can accept 71/// natively. 72/// 73/// This class also defines callbacks that targets must implement to lower 74/// target-specific constructs to SelectionDAG operators. 75/// 76class TargetLowering { 77 TargetLowering(const TargetLowering&); // DO NOT IMPLEMENT 78 void operator=(const TargetLowering&); // DO NOT IMPLEMENT 79public: 80 /// LegalizeAction - This enum indicates whether operations are valid for a 81 /// target, and if not, what action should be used to make them valid. 82 enum LegalizeAction { 83 Legal, // The target natively supports this operation. 84 Promote, // This operation should be executed in a larger type. 85 Expand, // Try to expand this to other ops, otherwise use a libcall. 86 Custom // Use the LowerOperation hook to implement custom lowering. 87 }; 88 89 /// LegalizeTypeAction - This enum indicates whether a types are legal for a 90 /// target, and if not, what action should be used to make them valid. 91 enum LegalizeTypeAction { 92 TypeLegal, // The target natively supports this type. 93 TypePromoteInteger, // Replace this integer with a larger one. 94 TypeExpandInteger, // Split this integer into two of half the size. 95 TypeSoftenFloat, // Convert this float to a same size integer type. 96 TypeExpandFloat, // Split this float into two of half the size. 97 TypeScalarizeVector, // Replace this one-element vector with its element. 98 TypeSplitVector, // Split this vector into two of half the size. 99 TypeWidenVector // This vector should be widened into a larger vector. 100 }; 101 102 enum BooleanContent { // How the target represents true/false values. 103 UndefinedBooleanContent, // Only bit 0 counts, the rest can hold garbage. 104 ZeroOrOneBooleanContent, // All bits zero except for bit 0. 105 ZeroOrNegativeOneBooleanContent // All bits equal to bit 0. 106 }; 107 108 static ISD::NodeType getExtendForContent(BooleanContent Content) { 109 switch (Content) { 110 case UndefinedBooleanContent: 111 // Extend by adding rubbish bits. 112 return ISD::ANY_EXTEND; 113 case ZeroOrOneBooleanContent: 114 // Extend by adding zero bits. 115 return ISD::ZERO_EXTEND; 116 case ZeroOrNegativeOneBooleanContent: 117 // Extend by copying the sign bit. 118 return ISD::SIGN_EXTEND; 119 } 120 llvm_unreachable("Invalid content kind"); 121 } 122 123 /// NOTE: The constructor takes ownership of TLOF. 124 explicit TargetLowering(const TargetMachine &TM, 125 const TargetLoweringObjectFile *TLOF); 126 virtual ~TargetLowering(); 127 128 const TargetMachine &getTargetMachine() const { return TM; } 129 const TargetData *getTargetData() const { return TD; } 130 const TargetLoweringObjectFile &getObjFileLowering() const { return TLOF; } 131 132 bool isBigEndian() const { return !IsLittleEndian; } 133 bool isLittleEndian() const { return IsLittleEndian; } 134 MVT getPointerTy() const { return PointerTy; } 135 virtual MVT getShiftAmountTy(EVT LHSTy) const; 136 137 /// isSelectExpensive - Return true if the select operation is expensive for 138 /// this target. 139 bool isSelectExpensive() const { return SelectIsExpensive; } 140 141 /// isIntDivCheap() - Return true if integer divide is usually cheaper than 142 /// a sequence of several shifts, adds, and multiplies for this target. 143 bool isIntDivCheap() const { return IntDivIsCheap; } 144 145 /// isPow2DivCheap() - Return true if pow2 div is cheaper than a chain of 146 /// srl/add/sra. 147 bool isPow2DivCheap() const { return Pow2DivIsCheap; } 148 149 /// isJumpExpensive() - Return true if Flow Control is an expensive operation 150 /// that should be avoided. 151 bool isJumpExpensive() const { return JumpIsExpensive; } 152 153 /// getSetCCResultType - Return the ValueType of the result of SETCC 154 /// operations. Also used to obtain the target's preferred type for 155 /// the condition operand of SELECT and BRCOND nodes. In the case of 156 /// BRCOND the argument passed is MVT::Other since there are no other 157 /// operands to get a type hint from. 158 virtual EVT getSetCCResultType(EVT VT) const; 159 160 /// getCmpLibcallReturnType - Return the ValueType for comparison 161 /// libcalls. Comparions libcalls include floating point comparion calls, 162 /// and Ordered/Unordered check calls on floating point numbers. 163 virtual 164 MVT::SimpleValueType getCmpLibcallReturnType() const; 165 166 /// getBooleanContents - For targets without i1 registers, this gives the 167 /// nature of the high-bits of boolean values held in types wider than i1. 168 /// "Boolean values" are special true/false values produced by nodes like 169 /// SETCC and consumed (as the condition) by nodes like SELECT and BRCOND. 170 /// Not to be confused with general values promoted from i1. 171 /// Some cpus distinguish between vectors of boolean and scalars; the isVec 172 /// parameter selects between the two kinds. For example on X86 a scalar 173 /// boolean should be zero extended from i1, while the elements of a vector 174 /// of booleans should be sign extended from i1. 175 BooleanContent getBooleanContents(bool isVec) const { 176 return isVec ? BooleanVectorContents : BooleanContents; 177 } 178 179 /// getSchedulingPreference - Return target scheduling preference. 180 Sched::Preference getSchedulingPreference() const { 181 return SchedPreferenceInfo; 182 } 183 184 /// getSchedulingPreference - Some scheduler, e.g. hybrid, can switch to 185 /// different scheduling heuristics for different nodes. This function returns 186 /// the preference (or none) for the given node. 187 virtual Sched::Preference getSchedulingPreference(SDNode *) const { 188 return Sched::None; 189 } 190 191 /// getRegClassFor - Return the register class that should be used for the 192 /// specified value type. 193 virtual TargetRegisterClass *getRegClassFor(EVT VT) const { 194 assert(VT.isSimple() && "getRegClassFor called on illegal type!"); 195 TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT().SimpleTy]; 196 assert(RC && "This value type is not natively supported!"); 197 return RC; 198 } 199 200 /// getRepRegClassFor - Return the 'representative' register class for the 201 /// specified value type. The 'representative' register class is the largest 202 /// legal super-reg register class for the register class of the value type. 203 /// For example, on i386 the rep register class for i8, i16, and i32 are GR32; 204 /// while the rep register class is GR64 on x86_64. 205 virtual const TargetRegisterClass *getRepRegClassFor(EVT VT) const { 206 assert(VT.isSimple() && "getRepRegClassFor called on illegal type!"); 207 const TargetRegisterClass *RC = RepRegClassForVT[VT.getSimpleVT().SimpleTy]; 208 return RC; 209 } 210 211 /// getRepRegClassCostFor - Return the cost of the 'representative' register 212 /// class for the specified value type. 213 virtual uint8_t getRepRegClassCostFor(EVT VT) const { 214 assert(VT.isSimple() && "getRepRegClassCostFor called on illegal type!"); 215 return RepRegClassCostForVT[VT.getSimpleVT().SimpleTy]; 216 } 217 218 /// isTypeLegal - Return true if the target has native support for the 219 /// specified value type. This means that it has a register that directly 220 /// holds it without promotions or expansions. 221 bool isTypeLegal(EVT VT) const { 222 assert(!VT.isSimple() || 223 (unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT)); 224 return VT.isSimple() && RegClassForVT[VT.getSimpleVT().SimpleTy] != 0; 225 } 226 227 class ValueTypeActionImpl { 228 /// ValueTypeActions - For each value type, keep a LegalizeTypeAction enum 229 /// that indicates how instruction selection should deal with the type. 230 uint8_t ValueTypeActions[MVT::LAST_VALUETYPE]; 231 232 public: 233 ValueTypeActionImpl() { 234 std::fill(ValueTypeActions, array_endof(ValueTypeActions), 0); 235 } 236 237 LegalizeTypeAction getTypeAction(MVT VT) const { 238 return (LegalizeTypeAction)ValueTypeActions[VT.SimpleTy]; 239 } 240 241 void setTypeAction(EVT VT, LegalizeTypeAction Action) { 242 unsigned I = VT.getSimpleVT().SimpleTy; 243 ValueTypeActions[I] = Action; 244 } 245 }; 246 247 const ValueTypeActionImpl &getValueTypeActions() const { 248 return ValueTypeActions; 249 } 250 251 /// getTypeAction - Return how we should legalize values of this type, either 252 /// it is already legal (return 'Legal') or we need to promote it to a larger 253 /// type (return 'Promote'), or we need to expand it into multiple registers 254 /// of smaller integer type (return 'Expand'). 'Custom' is not an option. 255 LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const { 256 return getTypeConversion(Context, VT).first; 257 } 258 LegalizeTypeAction getTypeAction(MVT VT) const { 259 return ValueTypeActions.getTypeAction(VT); 260 } 261 262 /// getTypeToTransformTo - For types supported by the target, this is an 263 /// identity function. For types that must be promoted to larger types, this 264 /// returns the larger type to promote to. For integer types that are larger 265 /// than the largest integer register, this contains one step in the expansion 266 /// to get to the smaller register. For illegal floating point types, this 267 /// returns the integer type to transform to. 268 EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const { 269 return getTypeConversion(Context, VT).second; 270 } 271 272 /// getTypeToExpandTo - For types supported by the target, this is an 273 /// identity function. For types that must be expanded (i.e. integer types 274 /// that are larger than the largest integer register or illegal floating 275 /// point types), this returns the largest legal type it will be expanded to. 276 EVT getTypeToExpandTo(LLVMContext &Context, EVT VT) const { 277 assert(!VT.isVector()); 278 while (true) { 279 switch (getTypeAction(Context, VT)) { 280 case TypeLegal: 281 return VT; 282 case TypeExpandInteger: 283 VT = getTypeToTransformTo(Context, VT); 284 break; 285 default: 286 assert(false && "Type is not legal nor is it to be expanded!"); 287 return VT; 288 } 289 } 290 return VT; 291 } 292 293 /// getVectorTypeBreakdown - Vector types are broken down into some number of 294 /// legal first class types. For example, EVT::v8f32 maps to 2 EVT::v4f32 295 /// with Altivec or SSE1, or 8 promoted EVT::f64 values with the X86 FP stack. 296 /// Similarly, EVT::v2i64 turns into 4 EVT::i32 values with both PPC and X86. 297 /// 298 /// This method returns the number of registers needed, and the VT for each 299 /// register. It also returns the VT and quantity of the intermediate values 300 /// before they are promoted/expanded. 301 /// 302 unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 303 EVT &IntermediateVT, 304 unsigned &NumIntermediates, 305 EVT &RegisterVT) const; 306 307 /// getTgtMemIntrinsic: Given an intrinsic, checks if on the target the 308 /// intrinsic will need to map to a MemIntrinsicNode (touches memory). If 309 /// this is the case, it returns true and store the intrinsic 310 /// information into the IntrinsicInfo that was passed to the function. 311 struct IntrinsicInfo { 312 unsigned opc; // target opcode 313 EVT memVT; // memory VT 314 const Value* ptrVal; // value representing memory location 315 int offset; // offset off of ptrVal 316 unsigned align; // alignment 317 bool vol; // is volatile? 318 bool readMem; // reads memory? 319 bool writeMem; // writes memory? 320 }; 321 322 virtual bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &, 323 unsigned /*Intrinsic*/) const { 324 return false; 325 } 326 327 /// isFPImmLegal - Returns true if the target can instruction select the 328 /// specified FP immediate natively. If false, the legalizer will materialize 329 /// the FP immediate as a load from a constant pool. 330 virtual bool isFPImmLegal(const APFloat &/*Imm*/, EVT /*VT*/) const { 331 return false; 332 } 333 334 /// isShuffleMaskLegal - Targets can use this to indicate that they only 335 /// support *some* VECTOR_SHUFFLE operations, those with specific masks. 336 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values 337 /// are assumed to be legal. 338 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 339 EVT /*VT*/) const { 340 return true; 341 } 342 343 /// canOpTrap - Returns true if the operation can trap for the value type. 344 /// VT must be a legal type. By default, we optimistically assume most 345 /// operations don't trap except for divide and remainder. 346 virtual bool canOpTrap(unsigned Op, EVT VT) const; 347 348 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is 349 /// used by Targets can use this to indicate if there is a suitable 350 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant 351 /// pool entry. 352 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 353 EVT /*VT*/) const { 354 return false; 355 } 356 357 /// getOperationAction - Return how this operation should be treated: either 358 /// it is legal, needs to be promoted to a larger size, needs to be 359 /// expanded to some other code sequence, or the target has a custom expander 360 /// for it. 361 LegalizeAction getOperationAction(unsigned Op, EVT VT) const { 362 if (VT.isExtended()) return Expand; 363 assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!"); 364 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy; 365 return (LegalizeAction)OpActions[I][Op]; 366 } 367 368 /// isOperationLegalOrCustom - Return true if the specified operation is 369 /// legal on this target or can be made legal with custom lowering. This 370 /// is used to help guide high-level lowering decisions. 371 bool isOperationLegalOrCustom(unsigned Op, EVT VT) const { 372 return (VT == MVT::Other || isTypeLegal(VT)) && 373 (getOperationAction(Op, VT) == Legal || 374 getOperationAction(Op, VT) == Custom); 375 } 376 377 /// isOperationLegal - Return true if the specified operation is legal on this 378 /// target. 379 bool isOperationLegal(unsigned Op, EVT VT) const { 380 return (VT == MVT::Other || isTypeLegal(VT)) && 381 getOperationAction(Op, VT) == Legal; 382 } 383 384 /// getLoadExtAction - Return how this load with extension should be treated: 385 /// either it is legal, needs to be promoted to a larger size, needs to be 386 /// expanded to some other code sequence, or the target has a custom expander 387 /// for it. 388 LegalizeAction getLoadExtAction(unsigned ExtType, EVT VT) const { 389 assert(ExtType < ISD::LAST_LOADEXT_TYPE && 390 VT.getSimpleVT() < MVT::LAST_VALUETYPE && 391 "Table isn't big enough!"); 392 return (LegalizeAction)LoadExtActions[VT.getSimpleVT().SimpleTy][ExtType]; 393 } 394 395 /// isLoadExtLegal - Return true if the specified load with extension is legal 396 /// on this target. 397 bool isLoadExtLegal(unsigned ExtType, EVT VT) const { 398 return VT.isSimple() && getLoadExtAction(ExtType, VT) == Legal; 399 } 400 401 /// getTruncStoreAction - Return how this store with truncation should be 402 /// treated: either it is legal, needs to be promoted to a larger size, needs 403 /// to be expanded to some other code sequence, or the target has a custom 404 /// expander for it. 405 LegalizeAction getTruncStoreAction(EVT ValVT, EVT MemVT) const { 406 assert(ValVT.getSimpleVT() < MVT::LAST_VALUETYPE && 407 MemVT.getSimpleVT() < MVT::LAST_VALUETYPE && 408 "Table isn't big enough!"); 409 return (LegalizeAction)TruncStoreActions[ValVT.getSimpleVT().SimpleTy] 410 [MemVT.getSimpleVT().SimpleTy]; 411 } 412 413 /// isTruncStoreLegal - Return true if the specified store with truncation is 414 /// legal on this target. 415 bool isTruncStoreLegal(EVT ValVT, EVT MemVT) const { 416 return isTypeLegal(ValVT) && MemVT.isSimple() && 417 getTruncStoreAction(ValVT, MemVT) == Legal; 418 } 419 420 /// getIndexedLoadAction - Return how the indexed load should be treated: 421 /// either it is legal, needs to be promoted to a larger size, needs to be 422 /// expanded to some other code sequence, or the target has a custom expander 423 /// for it. 424 LegalizeAction 425 getIndexedLoadAction(unsigned IdxMode, EVT VT) const { 426 assert(IdxMode < ISD::LAST_INDEXED_MODE && 427 VT.getSimpleVT() < MVT::LAST_VALUETYPE && 428 "Table isn't big enough!"); 429 unsigned Ty = (unsigned)VT.getSimpleVT().SimpleTy; 430 return (LegalizeAction)((IndexedModeActions[Ty][IdxMode] & 0xf0) >> 4); 431 } 432 433 /// isIndexedLoadLegal - Return true if the specified indexed load is legal 434 /// on this target. 435 bool isIndexedLoadLegal(unsigned IdxMode, EVT VT) const { 436 return VT.isSimple() && 437 (getIndexedLoadAction(IdxMode, VT) == Legal || 438 getIndexedLoadAction(IdxMode, VT) == Custom); 439 } 440 441 /// getIndexedStoreAction - Return how the indexed store should be treated: 442 /// either it is legal, needs to be promoted to a larger size, needs to be 443 /// expanded to some other code sequence, or the target has a custom expander 444 /// for it. 445 LegalizeAction 446 getIndexedStoreAction(unsigned IdxMode, EVT VT) const { 447 assert(IdxMode < ISD::LAST_INDEXED_MODE && 448 VT.getSimpleVT() < MVT::LAST_VALUETYPE && 449 "Table isn't big enough!"); 450 unsigned Ty = (unsigned)VT.getSimpleVT().SimpleTy; 451 return (LegalizeAction)(IndexedModeActions[Ty][IdxMode] & 0x0f); 452 } 453 454 /// isIndexedStoreLegal - Return true if the specified indexed load is legal 455 /// on this target. 456 bool isIndexedStoreLegal(unsigned IdxMode, EVT VT) const { 457 return VT.isSimple() && 458 (getIndexedStoreAction(IdxMode, VT) == Legal || 459 getIndexedStoreAction(IdxMode, VT) == Custom); 460 } 461 462 /// getCondCodeAction - Return how the condition code should be treated: 463 /// either it is legal, needs to be expanded to some other code sequence, 464 /// or the target has a custom expander for it. 465 LegalizeAction 466 getCondCodeAction(ISD::CondCode CC, EVT VT) const { 467 assert((unsigned)CC < array_lengthof(CondCodeActions) && 468 (unsigned)VT.getSimpleVT().SimpleTy < sizeof(CondCodeActions[0])*4 && 469 "Table isn't big enough!"); 470 LegalizeAction Action = (LegalizeAction) 471 ((CondCodeActions[CC] >> (2*VT.getSimpleVT().SimpleTy)) & 3); 472 assert(Action != Promote && "Can't promote condition code!"); 473 return Action; 474 } 475 476 /// isCondCodeLegal - Return true if the specified condition code is legal 477 /// on this target. 478 bool isCondCodeLegal(ISD::CondCode CC, EVT VT) const { 479 return getCondCodeAction(CC, VT) == Legal || 480 getCondCodeAction(CC, VT) == Custom; 481 } 482 483 484 /// getTypeToPromoteTo - If the action for this operation is to promote, this 485 /// method returns the ValueType to promote to. 486 EVT getTypeToPromoteTo(unsigned Op, EVT VT) const { 487 assert(getOperationAction(Op, VT) == Promote && 488 "This operation isn't promoted!"); 489 490 // See if this has an explicit type specified. 491 std::map<std::pair<unsigned, MVT::SimpleValueType>, 492 MVT::SimpleValueType>::const_iterator PTTI = 493 PromoteToType.find(std::make_pair(Op, VT.getSimpleVT().SimpleTy)); 494 if (PTTI != PromoteToType.end()) return PTTI->second; 495 496 assert((VT.isInteger() || VT.isFloatingPoint()) && 497 "Cannot autopromote this type, add it with AddPromotedToType."); 498 499 EVT NVT = VT; 500 do { 501 NVT = (MVT::SimpleValueType)(NVT.getSimpleVT().SimpleTy+1); 502 assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid && 503 "Didn't find type to promote to!"); 504 } while (!isTypeLegal(NVT) || 505 getOperationAction(Op, NVT) == Promote); 506 return NVT; 507 } 508 509 /// getValueType - Return the EVT corresponding to this LLVM type. 510 /// This is fixed by the LLVM operations except for the pointer size. If 511 /// AllowUnknown is true, this will return MVT::Other for types with no EVT 512 /// counterpart (e.g. structs), otherwise it will assert. 513 EVT getValueType(Type *Ty, bool AllowUnknown = false) const { 514 // Lower scalar pointers to native pointer types. 515 if (Ty->isPointerTy()) return PointerTy; 516 517 if (Ty->isVectorTy()) { 518 VectorType *VTy = cast<VectorType>(Ty); 519 Type *Elm = VTy->getElementType(); 520 // Lower vectors of pointers to native pointer types. 521 if (Elm->isPointerTy()) 522 Elm = EVT(PointerTy).getTypeForEVT(Ty->getContext()); 523 return EVT::getVectorVT(Ty->getContext(), EVT::getEVT(Elm, false), 524 VTy->getNumElements()); 525 } 526 return EVT::getEVT(Ty, AllowUnknown); 527 } 528 529 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 530 /// function arguments in the caller parameter area. This is the actual 531 /// alignment, not its logarithm. 532 virtual unsigned getByValTypeAlignment(Type *Ty) const; 533 534 /// getRegisterType - Return the type of registers that this ValueType will 535 /// eventually require. 536 EVT getRegisterType(MVT VT) const { 537 assert((unsigned)VT.SimpleTy < array_lengthof(RegisterTypeForVT)); 538 return RegisterTypeForVT[VT.SimpleTy]; 539 } 540 541 /// getRegisterType - Return the type of registers that this ValueType will 542 /// eventually require. 543 EVT getRegisterType(LLVMContext &Context, EVT VT) const { 544 if (VT.isSimple()) { 545 assert((unsigned)VT.getSimpleVT().SimpleTy < 546 array_lengthof(RegisterTypeForVT)); 547 return RegisterTypeForVT[VT.getSimpleVT().SimpleTy]; 548 } 549 if (VT.isVector()) { 550 EVT VT1, RegisterVT; 551 unsigned NumIntermediates; 552 (void)getVectorTypeBreakdown(Context, VT, VT1, 553 NumIntermediates, RegisterVT); 554 return RegisterVT; 555 } 556 if (VT.isInteger()) { 557 return getRegisterType(Context, getTypeToTransformTo(Context, VT)); 558 } 559 assert(0 && "Unsupported extended type!"); 560 return EVT(MVT::Other); // Not reached 561 } 562 563 /// getNumRegisters - Return the number of registers that this ValueType will 564 /// eventually require. This is one for any types promoted to live in larger 565 /// registers, but may be more than one for types (like i64) that are split 566 /// into pieces. For types like i140, which are first promoted then expanded, 567 /// it is the number of registers needed to hold all the bits of the original 568 /// type. For an i140 on a 32 bit machine this means 5 registers. 569 unsigned getNumRegisters(LLVMContext &Context, EVT VT) const { 570 if (VT.isSimple()) { 571 assert((unsigned)VT.getSimpleVT().SimpleTy < 572 array_lengthof(NumRegistersForVT)); 573 return NumRegistersForVT[VT.getSimpleVT().SimpleTy]; 574 } 575 if (VT.isVector()) { 576 EVT VT1, VT2; 577 unsigned NumIntermediates; 578 return getVectorTypeBreakdown(Context, VT, VT1, NumIntermediates, VT2); 579 } 580 if (VT.isInteger()) { 581 unsigned BitWidth = VT.getSizeInBits(); 582 unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits(); 583 return (BitWidth + RegWidth - 1) / RegWidth; 584 } 585 assert(0 && "Unsupported extended type!"); 586 return 0; // Not reached 587 } 588 589 /// ShouldShrinkFPConstant - If true, then instruction selection should 590 /// seek to shrink the FP constant of the specified type to a smaller type 591 /// in order to save space and / or reduce runtime. 592 virtual bool ShouldShrinkFPConstant(EVT) const { return true; } 593 594 /// hasTargetDAGCombine - If true, the target has custom DAG combine 595 /// transformations that it can perform for the specified node. 596 bool hasTargetDAGCombine(ISD::NodeType NT) const { 597 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 598 return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7)); 599 } 600 601 /// This function returns the maximum number of store operations permitted 602 /// to replace a call to llvm.memset. The value is set by the target at the 603 /// performance threshold for such a replacement. If OptSize is true, 604 /// return the limit for functions that have OptSize attribute. 605 /// @brief Get maximum # of store operations permitted for llvm.memset 606 unsigned getMaxStoresPerMemset(bool OptSize) const { 607 return OptSize ? maxStoresPerMemsetOptSize : maxStoresPerMemset; 608 } 609 610 /// This function returns the maximum number of store operations permitted 611 /// to replace a call to llvm.memcpy. The value is set by the target at the 612 /// performance threshold for such a replacement. If OptSize is true, 613 /// return the limit for functions that have OptSize attribute. 614 /// @brief Get maximum # of store operations permitted for llvm.memcpy 615 unsigned getMaxStoresPerMemcpy(bool OptSize) const { 616 return OptSize ? maxStoresPerMemcpyOptSize : maxStoresPerMemcpy; 617 } 618 619 /// This function returns the maximum number of store operations permitted 620 /// to replace a call to llvm.memmove. The value is set by the target at the 621 /// performance threshold for such a replacement. If OptSize is true, 622 /// return the limit for functions that have OptSize attribute. 623 /// @brief Get maximum # of store operations permitted for llvm.memmove 624 unsigned getMaxStoresPerMemmove(bool OptSize) const { 625 return OptSize ? maxStoresPerMemmoveOptSize : maxStoresPerMemmove; 626 } 627 628 /// This function returns true if the target allows unaligned memory accesses. 629 /// of the specified type. This is used, for example, in situations where an 630 /// array copy/move/set is converted to a sequence of store operations. It's 631 /// use helps to ensure that such replacements don't generate code that causes 632 /// an alignment error (trap) on the target machine. 633 /// @brief Determine if the target supports unaligned memory accesses. 634 virtual bool allowsUnalignedMemoryAccesses(EVT) const { 635 return false; 636 } 637 638 /// This function returns true if the target would benefit from code placement 639 /// optimization. 640 /// @brief Determine if the target should perform code placement optimization. 641 bool shouldOptimizeCodePlacement() const { 642 return benefitFromCodePlacementOpt; 643 } 644 645 /// getOptimalMemOpType - Returns the target specific optimal type for load 646 /// and store operations as a result of memset, memcpy, and memmove 647 /// lowering. If DstAlign is zero that means it's safe to destination 648 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it 649 /// means there isn't a need to check it against alignment requirement, 650 /// probably because the source does not need to be loaded. If 651 /// 'IsZeroVal' is true, that means it's safe to return a 652 /// non-scalar-integer type, e.g. empty string source, constant, or loaded 653 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is 654 /// constant so it does not need to be loaded. 655 /// It returns EVT::Other if the type should be determined using generic 656 /// target-independent logic. 657 virtual EVT getOptimalMemOpType(uint64_t /*Size*/, 658 unsigned /*DstAlign*/, unsigned /*SrcAlign*/, 659 bool /*IsZeroVal*/, 660 bool /*MemcpyStrSrc*/, 661 MachineFunction &/*MF*/) const { 662 return MVT::Other; 663 } 664 665 /// usesUnderscoreSetJmp - Determine if we should use _setjmp or setjmp 666 /// to implement llvm.setjmp. 667 bool usesUnderscoreSetJmp() const { 668 return UseUnderscoreSetJmp; 669 } 670 671 /// usesUnderscoreLongJmp - Determine if we should use _longjmp or longjmp 672 /// to implement llvm.longjmp. 673 bool usesUnderscoreLongJmp() const { 674 return UseUnderscoreLongJmp; 675 } 676 677 /// getStackPointerRegisterToSaveRestore - If a physical register, this 678 /// specifies the register that llvm.savestack/llvm.restorestack should save 679 /// and restore. 680 unsigned getStackPointerRegisterToSaveRestore() const { 681 return StackPointerRegisterToSaveRestore; 682 } 683 684 /// getExceptionAddressRegister - If a physical register, this returns 685 /// the register that receives the exception address on entry to a landing 686 /// pad. 687 unsigned getExceptionAddressRegister() const { 688 return ExceptionPointerRegister; 689 } 690 691 /// getExceptionSelectorRegister - If a physical register, this returns 692 /// the register that receives the exception typeid on entry to a landing 693 /// pad. 694 unsigned getExceptionSelectorRegister() const { 695 return ExceptionSelectorRegister; 696 } 697 698 /// getJumpBufSize - returns the target's jmp_buf size in bytes (if never 699 /// set, the default is 200) 700 unsigned getJumpBufSize() const { 701 return JumpBufSize; 702 } 703 704 /// getJumpBufAlignment - returns the target's jmp_buf alignment in bytes 705 /// (if never set, the default is 0) 706 unsigned getJumpBufAlignment() const { 707 return JumpBufAlignment; 708 } 709 710 /// getMinStackArgumentAlignment - return the minimum stack alignment of an 711 /// argument. 712 unsigned getMinStackArgumentAlignment() const { 713 return MinStackArgumentAlignment; 714 } 715 716 /// getMinFunctionAlignment - return the minimum function alignment. 717 /// 718 unsigned getMinFunctionAlignment() const { 719 return MinFunctionAlignment; 720 } 721 722 /// getPrefFunctionAlignment - return the preferred function alignment. 723 /// 724 unsigned getPrefFunctionAlignment() const { 725 return PrefFunctionAlignment; 726 } 727 728 /// getPrefLoopAlignment - return the preferred loop alignment. 729 /// 730 unsigned getPrefLoopAlignment() const { 731 return PrefLoopAlignment; 732 } 733 734 /// getShouldFoldAtomicFences - return whether the combiner should fold 735 /// fence MEMBARRIER instructions into the atomic intrinsic instructions. 736 /// 737 bool getShouldFoldAtomicFences() const { 738 return ShouldFoldAtomicFences; 739 } 740 741 /// getInsertFencesFor - return whether the DAG builder should automatically 742 /// insert fences and reduce ordering for atomics. 743 /// 744 bool getInsertFencesForAtomic() const { 745 return InsertFencesForAtomic; 746 } 747 748 /// getPreIndexedAddressParts - returns true by value, base pointer and 749 /// offset pointer and addressing mode by reference if the node's address 750 /// can be legally represented as pre-indexed load / store address. 751 virtual bool getPreIndexedAddressParts(SDNode * /*N*/, SDValue &/*Base*/, 752 SDValue &/*Offset*/, 753 ISD::MemIndexedMode &/*AM*/, 754 SelectionDAG &/*DAG*/) const { 755 return false; 756 } 757 758 /// getPostIndexedAddressParts - returns true by value, base pointer and 759 /// offset pointer and addressing mode by reference if this node can be 760 /// combined with a load / store to form a post-indexed load / store. 761 virtual bool getPostIndexedAddressParts(SDNode * /*N*/, SDNode * /*Op*/, 762 SDValue &/*Base*/, SDValue &/*Offset*/, 763 ISD::MemIndexedMode &/*AM*/, 764 SelectionDAG &/*DAG*/) const { 765 return false; 766 } 767 768 /// getJumpTableEncoding - Return the entry encoding for a jump table in the 769 /// current function. The returned value is a member of the 770 /// MachineJumpTableInfo::JTEntryKind enum. 771 virtual unsigned getJumpTableEncoding() const; 772 773 virtual const MCExpr * 774 LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/, 775 const MachineBasicBlock * /*MBB*/, unsigned /*uid*/, 776 MCContext &/*Ctx*/) const { 777 assert(0 && "Need to implement this hook if target has custom JTIs"); 778 return 0; 779 } 780 781 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC 782 /// jumptable. 783 virtual SDValue getPICJumpTableRelocBase(SDValue Table, 784 SelectionDAG &DAG) const; 785 786 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the 787 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an 788 /// MCExpr. 789 virtual const MCExpr * 790 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 791 unsigned JTI, MCContext &Ctx) const; 792 793 /// isOffsetFoldingLegal - Return true if folding a constant offset 794 /// with the given GlobalAddress is legal. It is frequently not legal in 795 /// PIC relocation models. 796 virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const; 797 798 /// getStackCookieLocation - Return true if the target stores stack 799 /// protector cookies at a fixed offset in some non-standard address 800 /// space, and populates the address space and offset as 801 /// appropriate. 802 virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/, 803 unsigned &/*Offset*/) const { 804 return false; 805 } 806 807 /// getMaximalGlobalOffset - Returns the maximal possible offset which can be 808 /// used for loads / stores from the global. 809 virtual unsigned getMaximalGlobalOffset() const { 810 return 0; 811 } 812 813 //===--------------------------------------------------------------------===// 814 // TargetLowering Optimization Methods 815 // 816 817 /// TargetLoweringOpt - A convenience struct that encapsulates a DAG, and two 818 /// SDValues for returning information from TargetLowering to its clients 819 /// that want to combine 820 struct TargetLoweringOpt { 821 SelectionDAG &DAG; 822 bool LegalTys; 823 bool LegalOps; 824 SDValue Old; 825 SDValue New; 826 827 explicit TargetLoweringOpt(SelectionDAG &InDAG, 828 bool LT, bool LO) : 829 DAG(InDAG), LegalTys(LT), LegalOps(LO) {} 830 831 bool LegalTypes() const { return LegalTys; } 832 bool LegalOperations() const { return LegalOps; } 833 834 bool CombineTo(SDValue O, SDValue N) { 835 Old = O; 836 New = N; 837 return true; 838 } 839 840 /// ShrinkDemandedConstant - Check to see if the specified operand of the 841 /// specified instruction is a constant integer. If so, check to see if 842 /// there are any bits set in the constant that are not demanded. If so, 843 /// shrink the constant and return true. 844 bool ShrinkDemandedConstant(SDValue Op, const APInt &Demanded); 845 846 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the 847 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening 848 /// cast, but it could be generalized for targets with other types of 849 /// implicit widening casts. 850 bool ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &Demanded, 851 DebugLoc dl); 852 }; 853 854 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the 855 /// DemandedMask bits of the result of Op are ever used downstream. If we can 856 /// use this information to simplify Op, create a new simplified DAG node and 857 /// return true, returning the original and new nodes in Old and New. 858 /// Otherwise, analyze the expression and return a mask of KnownOne and 859 /// KnownZero bits for the expression (used to simplify the caller). 860 /// The KnownZero/One bits may only be accurate for those bits in the 861 /// DemandedMask. 862 bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask, 863 APInt &KnownZero, APInt &KnownOne, 864 TargetLoweringOpt &TLO, unsigned Depth = 0) const; 865 866 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in 867 /// Mask are known to be either zero or one and return them in the 868 /// KnownZero/KnownOne bitsets. 869 virtual void computeMaskedBitsForTargetNode(const SDValue Op, 870 const APInt &Mask, 871 APInt &KnownZero, 872 APInt &KnownOne, 873 const SelectionDAG &DAG, 874 unsigned Depth = 0) const; 875 876 /// ComputeNumSignBitsForTargetNode - This method can be implemented by 877 /// targets that want to expose additional information about sign bits to the 878 /// DAG Combiner. 879 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op, 880 unsigned Depth = 0) const; 881 882 struct DAGCombinerInfo { 883 void *DC; // The DAG Combiner object. 884 bool BeforeLegalize; 885 bool BeforeLegalizeOps; 886 bool CalledByLegalizer; 887 public: 888 SelectionDAG &DAG; 889 890 DAGCombinerInfo(SelectionDAG &dag, bool bl, bool blo, bool cl, void *dc) 891 : DC(dc), BeforeLegalize(bl), BeforeLegalizeOps(blo), 892 CalledByLegalizer(cl), DAG(dag) {} 893 894 bool isBeforeLegalize() const { return BeforeLegalize; } 895 bool isBeforeLegalizeOps() const { return BeforeLegalizeOps; } 896 bool isCalledByLegalizer() const { return CalledByLegalizer; } 897 898 void AddToWorklist(SDNode *N); 899 void RemoveFromWorklist(SDNode *N); 900 SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To, 901 bool AddTo = true); 902 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true); 903 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true); 904 905 void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO); 906 }; 907 908 /// SimplifySetCC - Try to simplify a setcc built with the specified operands 909 /// and cc. If it is unable to simplify it, return a null SDValue. 910 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 911 ISD::CondCode Cond, bool foldBooleans, 912 DAGCombinerInfo &DCI, DebugLoc dl) const; 913 914 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 915 /// node is a GlobalAddress + offset. 916 virtual bool 917 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const; 918 919 /// PerformDAGCombine - This method will be invoked for all target nodes and 920 /// for any target-independent nodes that the target has registered with 921 /// invoke it for. 922 /// 923 /// The semantics are as follows: 924 /// Return Value: 925 /// SDValue.Val == 0 - No change was made 926 /// SDValue.Val == N - N was replaced, is dead, and is already handled. 927 /// otherwise - N should be replaced by the returned Operand. 928 /// 929 /// In addition, methods provided by DAGCombinerInfo may be used to perform 930 /// more complex transformations. 931 /// 932 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const; 933 934 /// isTypeDesirableForOp - Return true if the target has native support for 935 /// the specified value type and it is 'desirable' to use the type for the 936 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16 937 /// instruction encodings are longer and some i16 instructions are slow. 938 virtual bool isTypeDesirableForOp(unsigned /*Opc*/, EVT VT) const { 939 // By default, assume all legal types are desirable. 940 return isTypeLegal(VT); 941 } 942 943 /// isDesirableToPromoteOp - Return true if it is profitable for dag combiner 944 /// to transform a floating point op of specified opcode to a equivalent op of 945 /// an integer type. e.g. f32 load -> i32 load can be profitable on ARM. 946 virtual bool isDesirableToTransformToIntegerOp(unsigned /*Opc*/, 947 EVT /*VT*/) const { 948 return false; 949 } 950 951 /// IsDesirableToPromoteOp - This method query the target whether it is 952 /// beneficial for dag combiner to promote the specified node. If true, it 953 /// should return the desired promotion type by reference. 954 virtual bool IsDesirableToPromoteOp(SDValue /*Op*/, EVT &/*PVT*/) const { 955 return false; 956 } 957 958 //===--------------------------------------------------------------------===// 959 // TargetLowering Configuration Methods - These methods should be invoked by 960 // the derived class constructor to configure this object for the target. 961 // 962 963protected: 964 /// setBooleanContents - Specify how the target extends the result of a 965 /// boolean value from i1 to a wider type. See getBooleanContents. 966 void setBooleanContents(BooleanContent Ty) { BooleanContents = Ty; } 967 /// setBooleanVectorContents - Specify how the target extends the result 968 /// of a vector boolean value from a vector of i1 to a wider type. See 969 /// getBooleanContents. 970 void setBooleanVectorContents(BooleanContent Ty) { 971 BooleanVectorContents = Ty; 972 } 973 974 /// setSchedulingPreference - Specify the target scheduling preference. 975 void setSchedulingPreference(Sched::Preference Pref) { 976 SchedPreferenceInfo = Pref; 977 } 978 979 /// setUseUnderscoreSetJmp - Indicate whether this target prefers to 980 /// use _setjmp to implement llvm.setjmp or the non _ version. 981 /// Defaults to false. 982 void setUseUnderscoreSetJmp(bool Val) { 983 UseUnderscoreSetJmp = Val; 984 } 985 986 /// setUseUnderscoreLongJmp - Indicate whether this target prefers to 987 /// use _longjmp to implement llvm.longjmp or the non _ version. 988 /// Defaults to false. 989 void setUseUnderscoreLongJmp(bool Val) { 990 UseUnderscoreLongJmp = Val; 991 } 992 993 /// setStackPointerRegisterToSaveRestore - If set to a physical register, this 994 /// specifies the register that llvm.savestack/llvm.restorestack should save 995 /// and restore. 996 void setStackPointerRegisterToSaveRestore(unsigned R) { 997 StackPointerRegisterToSaveRestore = R; 998 } 999 1000 /// setExceptionPointerRegister - If set to a physical register, this sets 1001 /// the register that receives the exception address on entry to a landing 1002 /// pad. 1003 void setExceptionPointerRegister(unsigned R) { 1004 ExceptionPointerRegister = R; 1005 } 1006 1007 /// setExceptionSelectorRegister - If set to a physical register, this sets 1008 /// the register that receives the exception typeid on entry to a landing 1009 /// pad. 1010 void setExceptionSelectorRegister(unsigned R) { 1011 ExceptionSelectorRegister = R; 1012 } 1013 1014 /// SelectIsExpensive - Tells the code generator not to expand operations 1015 /// into sequences that use the select operations if possible. 1016 void setSelectIsExpensive(bool isExpensive = true) { 1017 SelectIsExpensive = isExpensive; 1018 } 1019 1020 /// JumpIsExpensive - Tells the code generator not to expand sequence of 1021 /// operations into a separate sequences that increases the amount of 1022 /// flow control. 1023 void setJumpIsExpensive(bool isExpensive = true) { 1024 JumpIsExpensive = isExpensive; 1025 } 1026 1027 /// setIntDivIsCheap - Tells the code generator that integer divide is 1028 /// expensive, and if possible, should be replaced by an alternate sequence 1029 /// of instructions not containing an integer divide. 1030 void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; } 1031 1032 /// setPow2DivIsCheap - Tells the code generator that it shouldn't generate 1033 /// srl/add/sra for a signed divide by power of two, and let the target handle 1034 /// it. 1035 void setPow2DivIsCheap(bool isCheap = true) { Pow2DivIsCheap = isCheap; } 1036 1037 /// addRegisterClass - Add the specified register class as an available 1038 /// regclass for the specified value type. This indicates the selector can 1039 /// handle values of that class natively. 1040 void addRegisterClass(EVT VT, TargetRegisterClass *RC) { 1041 assert((unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT)); 1042 AvailableRegClasses.push_back(std::make_pair(VT, RC)); 1043 RegClassForVT[VT.getSimpleVT().SimpleTy] = RC; 1044 } 1045 1046 /// findRepresentativeClass - Return the largest legal super-reg register class 1047 /// of the register class for the specified type and its associated "cost". 1048 virtual std::pair<const TargetRegisterClass*, uint8_t> 1049 findRepresentativeClass(EVT VT) const; 1050 1051 /// computeRegisterProperties - Once all of the register classes are added, 1052 /// this allows us to compute derived properties we expose. 1053 void computeRegisterProperties(); 1054 1055 /// setOperationAction - Indicate that the specified operation does not work 1056 /// with the specified type and indicate what to do about it. 1057 void setOperationAction(unsigned Op, MVT VT, 1058 LegalizeAction Action) { 1059 assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!"); 1060 OpActions[(unsigned)VT.SimpleTy][Op] = (uint8_t)Action; 1061 } 1062 1063 /// setLoadExtAction - Indicate that the specified load with extension does 1064 /// not work with the specified type and indicate what to do about it. 1065 void setLoadExtAction(unsigned ExtType, MVT VT, 1066 LegalizeAction Action) { 1067 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE && 1068 "Table isn't big enough!"); 1069 LoadExtActions[VT.SimpleTy][ExtType] = (uint8_t)Action; 1070 } 1071 1072 /// setTruncStoreAction - Indicate that the specified truncating store does 1073 /// not work with the specified type and indicate what to do about it. 1074 void setTruncStoreAction(MVT ValVT, MVT MemVT, 1075 LegalizeAction Action) { 1076 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE && 1077 "Table isn't big enough!"); 1078 TruncStoreActions[ValVT.SimpleTy][MemVT.SimpleTy] = (uint8_t)Action; 1079 } 1080 1081 /// setIndexedLoadAction - Indicate that the specified indexed load does or 1082 /// does not work with the specified type and indicate what to do abort 1083 /// it. NOTE: All indexed mode loads are initialized to Expand in 1084 /// TargetLowering.cpp 1085 void setIndexedLoadAction(unsigned IdxMode, MVT VT, 1086 LegalizeAction Action) { 1087 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE && 1088 (unsigned)Action < 0xf && "Table isn't big enough!"); 1089 // Load action are kept in the upper half. 1090 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0xf0; 1091 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action) <<4; 1092 } 1093 1094 /// setIndexedStoreAction - Indicate that the specified indexed store does or 1095 /// does not work with the specified type and indicate what to do about 1096 /// it. NOTE: All indexed mode stores are initialized to Expand in 1097 /// TargetLowering.cpp 1098 void setIndexedStoreAction(unsigned IdxMode, MVT VT, 1099 LegalizeAction Action) { 1100 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE && 1101 (unsigned)Action < 0xf && "Table isn't big enough!"); 1102 // Store action are kept in the lower half. 1103 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0x0f; 1104 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action); 1105 } 1106 1107 /// setCondCodeAction - Indicate that the specified condition code is or isn't 1108 /// supported on the target and indicate what to do about it. 1109 void setCondCodeAction(ISD::CondCode CC, MVT VT, 1110 LegalizeAction Action) { 1111 assert(VT < MVT::LAST_VALUETYPE && 1112 (unsigned)CC < array_lengthof(CondCodeActions) && 1113 "Table isn't big enough!"); 1114 CondCodeActions[(unsigned)CC] &= ~(uint64_t(3UL) << VT.SimpleTy*2); 1115 CondCodeActions[(unsigned)CC] |= (uint64_t)Action << VT.SimpleTy*2; 1116 } 1117 1118 /// AddPromotedToType - If Opc/OrigVT is specified as being promoted, the 1119 /// promotion code defaults to trying a larger integer/fp until it can find 1120 /// one that works. If that default is insufficient, this method can be used 1121 /// by the target to override the default. 1122 void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) { 1123 PromoteToType[std::make_pair(Opc, OrigVT.SimpleTy)] = DestVT.SimpleTy; 1124 } 1125 1126 /// setTargetDAGCombine - Targets should invoke this method for each target 1127 /// independent node that they want to provide a custom DAG combiner for by 1128 /// implementing the PerformDAGCombine virtual method. 1129 void setTargetDAGCombine(ISD::NodeType NT) { 1130 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 1131 TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7); 1132 } 1133 1134 /// setJumpBufSize - Set the target's required jmp_buf buffer size (in 1135 /// bytes); default is 200 1136 void setJumpBufSize(unsigned Size) { 1137 JumpBufSize = Size; 1138 } 1139 1140 /// setJumpBufAlignment - Set the target's required jmp_buf buffer 1141 /// alignment (in bytes); default is 0 1142 void setJumpBufAlignment(unsigned Align) { 1143 JumpBufAlignment = Align; 1144 } 1145 1146 /// setMinFunctionAlignment - Set the target's minimum function alignment (in 1147 /// log2(bytes)) 1148 void setMinFunctionAlignment(unsigned Align) { 1149 MinFunctionAlignment = Align; 1150 } 1151 1152 /// setPrefFunctionAlignment - Set the target's preferred function alignment. 1153 /// This should be set if there is a performance benefit to 1154 /// higher-than-minimum alignment (in log2(bytes)) 1155 void setPrefFunctionAlignment(unsigned Align) { 1156 PrefFunctionAlignment = Align; 1157 } 1158 1159 /// setPrefLoopAlignment - Set the target's preferred loop alignment. Default 1160 /// alignment is zero, it means the target does not care about loop alignment. 1161 /// The alignment is specified in log2(bytes). 1162 void setPrefLoopAlignment(unsigned Align) { 1163 PrefLoopAlignment = Align; 1164 } 1165 1166 /// setMinStackArgumentAlignment - Set the minimum stack alignment of an 1167 /// argument (in log2(bytes)). 1168 void setMinStackArgumentAlignment(unsigned Align) { 1169 MinStackArgumentAlignment = Align; 1170 } 1171 1172 /// setShouldFoldAtomicFences - Set if the target's implementation of the 1173 /// atomic operation intrinsics includes locking. Default is false. 1174 void setShouldFoldAtomicFences(bool fold) { 1175 ShouldFoldAtomicFences = fold; 1176 } 1177 1178 /// setInsertFencesForAtomic - Set if the the DAG builder should 1179 /// automatically insert fences and reduce the order of atomic memory 1180 /// operations to Monotonic. 1181 void setInsertFencesForAtomic(bool fence) { 1182 InsertFencesForAtomic = fence; 1183 } 1184 1185public: 1186 //===--------------------------------------------------------------------===// 1187 // Lowering methods - These methods must be implemented by targets so that 1188 // the SelectionDAGLowering code knows how to lower these. 1189 // 1190 1191 /// LowerFormalArguments - This hook must be implemented to lower the 1192 /// incoming (formal) arguments, described by the Ins array, into the 1193 /// specified DAG. The implementation should fill in the InVals array 1194 /// with legal-type argument values, and return the resulting token 1195 /// chain value. 1196 /// 1197 virtual SDValue 1198 LowerFormalArguments(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 1199 bool /*isVarArg*/, 1200 const SmallVectorImpl<ISD::InputArg> &/*Ins*/, 1201 DebugLoc /*dl*/, SelectionDAG &/*DAG*/, 1202 SmallVectorImpl<SDValue> &/*InVals*/) const { 1203 assert(0 && "Not Implemented"); 1204 return SDValue(); // this is here to silence compiler errors 1205 } 1206 1207 /// LowerCallTo - This function lowers an abstract call to a function into an 1208 /// actual call. This returns a pair of operands. The first element is the 1209 /// return value for the function (if RetTy is not VoidTy). The second 1210 /// element is the outgoing token chain. It calls LowerCall to do the actual 1211 /// lowering. 1212 struct ArgListEntry { 1213 SDValue Node; 1214 Type* Ty; 1215 bool isSExt : 1; 1216 bool isZExt : 1; 1217 bool isInReg : 1; 1218 bool isSRet : 1; 1219 bool isNest : 1; 1220 bool isByVal : 1; 1221 uint16_t Alignment; 1222 1223 ArgListEntry() : isSExt(false), isZExt(false), isInReg(false), 1224 isSRet(false), isNest(false), isByVal(false), Alignment(0) { } 1225 }; 1226 typedef std::vector<ArgListEntry> ArgListTy; 1227 std::pair<SDValue, SDValue> 1228 LowerCallTo(SDValue Chain, Type *RetTy, bool RetSExt, bool RetZExt, 1229 bool isVarArg, bool isInreg, unsigned NumFixedArgs, 1230 CallingConv::ID CallConv, bool isTailCall, 1231 bool isReturnValueUsed, SDValue Callee, ArgListTy &Args, 1232 SelectionDAG &DAG, DebugLoc dl) const; 1233 1234 /// LowerCall - This hook must be implemented to lower calls into the 1235 /// the specified DAG. The outgoing arguments to the call are described 1236 /// by the Outs array, and the values to be returned by the call are 1237 /// described by the Ins array. The implementation should fill in the 1238 /// InVals array with legal-type return values from the call, and return 1239 /// the resulting token chain value. 1240 virtual SDValue 1241 LowerCall(SDValue /*Chain*/, SDValue /*Callee*/, 1242 CallingConv::ID /*CallConv*/, bool /*isVarArg*/, 1243 bool &/*isTailCall*/, 1244 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 1245 const SmallVectorImpl<SDValue> &/*OutVals*/, 1246 const SmallVectorImpl<ISD::InputArg> &/*Ins*/, 1247 DebugLoc /*dl*/, SelectionDAG &/*DAG*/, 1248 SmallVectorImpl<SDValue> &/*InVals*/) const { 1249 assert(0 && "Not Implemented"); 1250 return SDValue(); // this is here to silence compiler errors 1251 } 1252 1253 /// HandleByVal - Target-specific cleanup for formal ByVal parameters. 1254 virtual void HandleByVal(CCState *, unsigned &) const {} 1255 1256 /// CanLowerReturn - This hook should be implemented to check whether the 1257 /// return values described by the Outs array can fit into the return 1258 /// registers. If false is returned, an sret-demotion is performed. 1259 /// 1260 virtual bool CanLowerReturn(CallingConv::ID /*CallConv*/, 1261 MachineFunction &/*MF*/, bool /*isVarArg*/, 1262 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 1263 LLVMContext &/*Context*/) const 1264 { 1265 // Return true by default to get preexisting behavior. 1266 return true; 1267 } 1268 1269 /// LowerReturn - This hook must be implemented to lower outgoing 1270 /// return values, described by the Outs array, into the specified 1271 /// DAG. The implementation should return the resulting token chain 1272 /// value. 1273 /// 1274 virtual SDValue 1275 LowerReturn(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 1276 bool /*isVarArg*/, 1277 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 1278 const SmallVectorImpl<SDValue> &/*OutVals*/, 1279 DebugLoc /*dl*/, SelectionDAG &/*DAG*/) const { 1280 assert(0 && "Not Implemented"); 1281 return SDValue(); // this is here to silence compiler errors 1282 } 1283 1284 /// isUsedByReturnOnly - Return true if result of the specified node is used 1285 /// by a return node only. This is used to determine whether it is possible 1286 /// to codegen a libcall as tail call at legalization time. 1287 virtual bool isUsedByReturnOnly(SDNode *) const { 1288 return false; 1289 } 1290 1291 /// mayBeEmittedAsTailCall - Return true if the target may be able emit the 1292 /// call instruction as a tail call. This is used by optimization passes to 1293 /// determine if it's profitable to duplicate return instructions to enable 1294 /// tailcall optimization. 1295 virtual bool mayBeEmittedAsTailCall(CallInst *) const { 1296 return false; 1297 } 1298 1299 /// getTypeForExtArgOrReturn - Return the type that should be used to zero or 1300 /// sign extend a zeroext/signext integer argument or return value. 1301 /// FIXME: Most C calling convention requires the return type to be promoted, 1302 /// but this is not true all the time, e.g. i1 on x86-64. It is also not 1303 /// necessary for non-C calling conventions. The frontend should handle this 1304 /// and include all of the necessary information. 1305 virtual EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT, 1306 ISD::NodeType /*ExtendKind*/) const { 1307 EVT MinVT = getRegisterType(Context, MVT::i32); 1308 return VT.bitsLT(MinVT) ? MinVT : VT; 1309 } 1310 1311 /// LowerOperationWrapper - This callback is invoked by the type legalizer 1312 /// to legalize nodes with an illegal operand type but legal result types. 1313 /// It replaces the LowerOperation callback in the type Legalizer. 1314 /// The reason we can not do away with LowerOperation entirely is that 1315 /// LegalizeDAG isn't yet ready to use this callback. 1316 /// TODO: Consider merging with ReplaceNodeResults. 1317 1318 /// The target places new result values for the node in Results (their number 1319 /// and types must exactly match those of the original return values of 1320 /// the node), or leaves Results empty, which indicates that the node is not 1321 /// to be custom lowered after all. 1322 /// The default implementation calls LowerOperation. 1323 virtual void LowerOperationWrapper(SDNode *N, 1324 SmallVectorImpl<SDValue> &Results, 1325 SelectionDAG &DAG) const; 1326 1327 /// LowerOperation - This callback is invoked for operations that are 1328 /// unsupported by the target, which are registered to use 'custom' lowering, 1329 /// and whose defined values are all legal. 1330 /// If the target has no operations that require custom lowering, it need not 1331 /// implement this. The default implementation of this aborts. 1332 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; 1333 1334 /// ReplaceNodeResults - This callback is invoked when a node result type is 1335 /// illegal for the target, and the operation was registered to use 'custom' 1336 /// lowering for that result type. The target places new result values for 1337 /// the node in Results (their number and types must exactly match those of 1338 /// the original return values of the node), or leaves Results empty, which 1339 /// indicates that the node is not to be custom lowered after all. 1340 /// 1341 /// If the target has no operations that require custom lowering, it need not 1342 /// implement this. The default implementation aborts. 1343 virtual void ReplaceNodeResults(SDNode * /*N*/, 1344 SmallVectorImpl<SDValue> &/*Results*/, 1345 SelectionDAG &/*DAG*/) const { 1346 assert(0 && "ReplaceNodeResults not implemented for this target!"); 1347 } 1348 1349 /// getTargetNodeName() - This method returns the name of a target specific 1350 /// DAG node. 1351 virtual const char *getTargetNodeName(unsigned Opcode) const; 1352 1353 /// createFastISel - This method returns a target specific FastISel object, 1354 /// or null if the target does not support "fast" ISel. 1355 virtual FastISel *createFastISel(FunctionLoweringInfo &) const { 1356 return 0; 1357 } 1358 1359 //===--------------------------------------------------------------------===// 1360 // Inline Asm Support hooks 1361 // 1362 1363 /// ExpandInlineAsm - This hook allows the target to expand an inline asm 1364 /// call to be explicit llvm code if it wants to. This is useful for 1365 /// turning simple inline asms into LLVM intrinsics, which gives the 1366 /// compiler more information about the behavior of the code. 1367 virtual bool ExpandInlineAsm(CallInst *) const { 1368 return false; 1369 } 1370 1371 enum ConstraintType { 1372 C_Register, // Constraint represents specific register(s). 1373 C_RegisterClass, // Constraint represents any of register(s) in class. 1374 C_Memory, // Memory constraint. 1375 C_Other, // Something else. 1376 C_Unknown // Unsupported constraint. 1377 }; 1378 1379 enum ConstraintWeight { 1380 // Generic weights. 1381 CW_Invalid = -1, // No match. 1382 CW_Okay = 0, // Acceptable. 1383 CW_Good = 1, // Good weight. 1384 CW_Better = 2, // Better weight. 1385 CW_Best = 3, // Best weight. 1386 1387 // Well-known weights. 1388 CW_SpecificReg = CW_Okay, // Specific register operands. 1389 CW_Register = CW_Good, // Register operands. 1390 CW_Memory = CW_Better, // Memory operands. 1391 CW_Constant = CW_Best, // Constant operand. 1392 CW_Default = CW_Okay // Default or don't know type. 1393 }; 1394 1395 /// AsmOperandInfo - This contains information for each constraint that we are 1396 /// lowering. 1397 struct AsmOperandInfo : public InlineAsm::ConstraintInfo { 1398 /// ConstraintCode - This contains the actual string for the code, like "m". 1399 /// TargetLowering picks the 'best' code from ConstraintInfo::Codes that 1400 /// most closely matches the operand. 1401 std::string ConstraintCode; 1402 1403 /// ConstraintType - Information about the constraint code, e.g. Register, 1404 /// RegisterClass, Memory, Other, Unknown. 1405 TargetLowering::ConstraintType ConstraintType; 1406 1407 /// CallOperandval - If this is the result output operand or a 1408 /// clobber, this is null, otherwise it is the incoming operand to the 1409 /// CallInst. This gets modified as the asm is processed. 1410 Value *CallOperandVal; 1411 1412 /// ConstraintVT - The ValueType for the operand value. 1413 EVT ConstraintVT; 1414 1415 /// isMatchingInputConstraint - Return true of this is an input operand that 1416 /// is a matching constraint like "4". 1417 bool isMatchingInputConstraint() const; 1418 1419 /// getMatchedOperand - If this is an input matching constraint, this method 1420 /// returns the output operand it matches. 1421 unsigned getMatchedOperand() const; 1422 1423 /// Copy constructor for copying from an AsmOperandInfo. 1424 AsmOperandInfo(const AsmOperandInfo &info) 1425 : InlineAsm::ConstraintInfo(info), 1426 ConstraintCode(info.ConstraintCode), 1427 ConstraintType(info.ConstraintType), 1428 CallOperandVal(info.CallOperandVal), 1429 ConstraintVT(info.ConstraintVT) { 1430 } 1431 1432 /// Copy constructor for copying from a ConstraintInfo. 1433 AsmOperandInfo(const InlineAsm::ConstraintInfo &info) 1434 : InlineAsm::ConstraintInfo(info), 1435 ConstraintType(TargetLowering::C_Unknown), 1436 CallOperandVal(0), ConstraintVT(MVT::Other) { 1437 } 1438 }; 1439 1440 typedef std::vector<AsmOperandInfo> AsmOperandInfoVector; 1441 1442 /// ParseConstraints - Split up the constraint string from the inline 1443 /// assembly value into the specific constraints and their prefixes, 1444 /// and also tie in the associated operand values. 1445 /// If this returns an empty vector, and if the constraint string itself 1446 /// isn't empty, there was an error parsing. 1447 virtual AsmOperandInfoVector ParseConstraints(ImmutableCallSite CS) const; 1448 1449 /// Examine constraint type and operand type and determine a weight value. 1450 /// The operand object must already have been set up with the operand type. 1451 virtual ConstraintWeight getMultipleConstraintMatchWeight( 1452 AsmOperandInfo &info, int maIndex) const; 1453 1454 /// Examine constraint string and operand type and determine a weight value. 1455 /// The operand object must already have been set up with the operand type. 1456 virtual ConstraintWeight getSingleConstraintMatchWeight( 1457 AsmOperandInfo &info, const char *constraint) const; 1458 1459 /// ComputeConstraintToUse - Determines the constraint code and constraint 1460 /// type to use for the specific AsmOperandInfo, setting 1461 /// OpInfo.ConstraintCode and OpInfo.ConstraintType. If the actual operand 1462 /// being passed in is available, it can be passed in as Op, otherwise an 1463 /// empty SDValue can be passed. 1464 virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo, 1465 SDValue Op, 1466 SelectionDAG *DAG = 0) const; 1467 1468 /// getConstraintType - Given a constraint, return the type of constraint it 1469 /// is for this target. 1470 virtual ConstraintType getConstraintType(const std::string &Constraint) const; 1471 1472 /// getRegForInlineAsmConstraint - Given a physical register constraint (e.g. 1473 /// {edx}), return the register number and the register class for the 1474 /// register. 1475 /// 1476 /// Given a register class constraint, like 'r', if this corresponds directly 1477 /// to an LLVM register class, return a register of 0 and the register class 1478 /// pointer. 1479 /// 1480 /// This should only be used for C_Register constraints. On error, 1481 /// this returns a register number of 0 and a null register class pointer.. 1482 virtual std::pair<unsigned, const TargetRegisterClass*> 1483 getRegForInlineAsmConstraint(const std::string &Constraint, 1484 EVT VT) const; 1485 1486 /// LowerXConstraint - try to replace an X constraint, which matches anything, 1487 /// with another that has more specific requirements based on the type of the 1488 /// corresponding operand. This returns null if there is no replacement to 1489 /// make. 1490 virtual const char *LowerXConstraint(EVT ConstraintVT) const; 1491 1492 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 1493 /// vector. If it is invalid, don't add anything to Ops. 1494 virtual void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, 1495 std::vector<SDValue> &Ops, 1496 SelectionDAG &DAG) const; 1497 1498 //===--------------------------------------------------------------------===// 1499 // Instruction Emitting Hooks 1500 // 1501 1502 // EmitInstrWithCustomInserter - This method should be implemented by targets 1503 // that mark instructions with the 'usesCustomInserter' flag. These 1504 // instructions are special in various ways, which require special support to 1505 // insert. The specified MachineInstr is created but not inserted into any 1506 // basic blocks, and this method is called to expand it into a sequence of 1507 // instructions, potentially also creating new basic blocks and control flow. 1508 virtual MachineBasicBlock * 1509 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const; 1510 1511 /// AdjustInstrPostInstrSelection - This method should be implemented by 1512 /// targets that mark instructions with the 'hasPostISelHook' flag. These 1513 /// instructions must be adjusted after instruction selection by target hooks. 1514 /// e.g. To fill in optional defs for ARM 's' setting instructions. 1515 virtual void 1516 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const; 1517 1518 //===--------------------------------------------------------------------===// 1519 // Addressing mode description hooks (used by LSR etc). 1520 // 1521 1522 /// AddrMode - This represents an addressing mode of: 1523 /// BaseGV + BaseOffs + BaseReg + Scale*ScaleReg 1524 /// If BaseGV is null, there is no BaseGV. 1525 /// If BaseOffs is zero, there is no base offset. 1526 /// If HasBaseReg is false, there is no base register. 1527 /// If Scale is zero, there is no ScaleReg. Scale of 1 indicates a reg with 1528 /// no scale. 1529 /// 1530 struct AddrMode { 1531 GlobalValue *BaseGV; 1532 int64_t BaseOffs; 1533 bool HasBaseReg; 1534 int64_t Scale; 1535 AddrMode() : BaseGV(0), BaseOffs(0), HasBaseReg(false), Scale(0) {} 1536 }; 1537 1538 /// isLegalAddressingMode - Return true if the addressing mode represented by 1539 /// AM is legal for this target, for a load/store of the specified type. 1540 /// The type may be VoidTy, in which case only return true if the addressing 1541 /// mode is legal for a load/store of any legal type. 1542 /// TODO: Handle pre/postinc as well. 1543 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const; 1544 1545 /// isLegalICmpImmediate - Return true if the specified immediate is legal 1546 /// icmp immediate, that is the target has icmp instructions which can compare 1547 /// a register against the immediate without having to materialize the 1548 /// immediate into a register. 1549 virtual bool isLegalICmpImmediate(int64_t) const { 1550 return true; 1551 } 1552 1553 /// isLegalAddImmediate - Return true if the specified immediate is legal 1554 /// add immediate, that is the target has add instructions which can add 1555 /// a register with the immediate without having to materialize the 1556 /// immediate into a register. 1557 virtual bool isLegalAddImmediate(int64_t) const { 1558 return true; 1559 } 1560 1561 /// isTruncateFree - Return true if it's free to truncate a value of 1562 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in 1563 /// register EAX to i16 by referencing its sub-register AX. 1564 virtual bool isTruncateFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 1565 return false; 1566 } 1567 1568 virtual bool isTruncateFree(EVT /*VT1*/, EVT /*VT2*/) const { 1569 return false; 1570 } 1571 1572 /// isZExtFree - Return true if any actual instruction that defines a 1573 /// value of type Ty1 implicitly zero-extends the value to Ty2 in the result 1574 /// register. This does not necessarily include registers defined in 1575 /// unknown ways, such as incoming arguments, or copies from unknown 1576 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this 1577 /// does not necessarily apply to truncate instructions. e.g. on x86-64, 1578 /// all instructions that define 32-bit values implicit zero-extend the 1579 /// result out to 64 bits. 1580 virtual bool isZExtFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 1581 return false; 1582 } 1583 1584 virtual bool isZExtFree(EVT /*VT1*/, EVT /*VT2*/) const { 1585 return false; 1586 } 1587 1588 /// isNarrowingProfitable - Return true if it's profitable to narrow 1589 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow 1590 /// from i32 to i8 but not from i32 to i16. 1591 virtual bool isNarrowingProfitable(EVT /*VT1*/, EVT /*VT2*/) const { 1592 return false; 1593 } 1594 1595 //===--------------------------------------------------------------------===// 1596 // Div utility functions 1597 // 1598 SDValue BuildExactSDIV(SDValue Op1, SDValue Op2, DebugLoc dl, 1599 SelectionDAG &DAG) const; 1600 SDValue BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 1601 std::vector<SDNode*>* Created) const; 1602 SDValue BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 1603 std::vector<SDNode*>* Created) const; 1604 1605 1606 //===--------------------------------------------------------------------===// 1607 // Runtime Library hooks 1608 // 1609 1610 /// setLibcallName - Rename the default libcall routine name for the specified 1611 /// libcall. 1612 void setLibcallName(RTLIB::Libcall Call, const char *Name) { 1613 LibcallRoutineNames[Call] = Name; 1614 } 1615 1616 /// getLibcallName - Get the libcall routine name for the specified libcall. 1617 /// 1618 const char *getLibcallName(RTLIB::Libcall Call) const { 1619 return LibcallRoutineNames[Call]; 1620 } 1621 1622 /// setCmpLibcallCC - Override the default CondCode to be used to test the 1623 /// result of the comparison libcall against zero. 1624 void setCmpLibcallCC(RTLIB::Libcall Call, ISD::CondCode CC) { 1625 CmpLibcallCCs[Call] = CC; 1626 } 1627 1628 /// getCmpLibcallCC - Get the CondCode that's to be used to test the result of 1629 /// the comparison libcall against zero. 1630 ISD::CondCode getCmpLibcallCC(RTLIB::Libcall Call) const { 1631 return CmpLibcallCCs[Call]; 1632 } 1633 1634 /// setLibcallCallingConv - Set the CallingConv that should be used for the 1635 /// specified libcall. 1636 void setLibcallCallingConv(RTLIB::Libcall Call, CallingConv::ID CC) { 1637 LibcallCallingConvs[Call] = CC; 1638 } 1639 1640 /// getLibcallCallingConv - Get the CallingConv that should be used for the 1641 /// specified libcall. 1642 CallingConv::ID getLibcallCallingConv(RTLIB::Libcall Call) const { 1643 return LibcallCallingConvs[Call]; 1644 } 1645 1646private: 1647 const TargetMachine &TM; 1648 const TargetData *TD; 1649 const TargetLoweringObjectFile &TLOF; 1650 1651 /// We are in the process of implementing a new TypeLegalization action 1652 /// which is the promotion of vector elements. This feature is under 1653 /// development. Until this feature is complete, it is only enabled using a 1654 /// flag. We pass this flag using a member because of circular dep issues. 1655 /// This member will be removed with the flag once we complete the transition. 1656 bool mayPromoteElements; 1657 1658 /// PointerTy - The type to use for pointers, usually i32 or i64. 1659 /// 1660 MVT PointerTy; 1661 1662 /// IsLittleEndian - True if this is a little endian target. 1663 /// 1664 bool IsLittleEndian; 1665 1666 /// SelectIsExpensive - Tells the code generator not to expand operations 1667 /// into sequences that use the select operations if possible. 1668 bool SelectIsExpensive; 1669 1670 /// IntDivIsCheap - Tells the code generator not to expand integer divides by 1671 /// constants into a sequence of muls, adds, and shifts. This is a hack until 1672 /// a real cost model is in place. If we ever optimize for size, this will be 1673 /// set to true unconditionally. 1674 bool IntDivIsCheap; 1675 1676 /// Pow2DivIsCheap - Tells the code generator that it shouldn't generate 1677 /// srl/add/sra for a signed divide by power of two, and let the target handle 1678 /// it. 1679 bool Pow2DivIsCheap; 1680 1681 /// JumpIsExpensive - Tells the code generator that it shouldn't generate 1682 /// extra flow control instructions and should attempt to combine flow 1683 /// control instructions via predication. 1684 bool JumpIsExpensive; 1685 1686 /// UseUnderscoreSetJmp - This target prefers to use _setjmp to implement 1687 /// llvm.setjmp. Defaults to false. 1688 bool UseUnderscoreSetJmp; 1689 1690 /// UseUnderscoreLongJmp - This target prefers to use _longjmp to implement 1691 /// llvm.longjmp. Defaults to false. 1692 bool UseUnderscoreLongJmp; 1693 1694 /// BooleanContents - Information about the contents of the high-bits in 1695 /// boolean values held in a type wider than i1. See getBooleanContents. 1696 BooleanContent BooleanContents; 1697 /// BooleanVectorContents - Information about the contents of the high-bits 1698 /// in boolean vector values when the element type is wider than i1. See 1699 /// getBooleanContents. 1700 BooleanContent BooleanVectorContents; 1701 1702 /// SchedPreferenceInfo - The target scheduling preference: shortest possible 1703 /// total cycles or lowest register usage. 1704 Sched::Preference SchedPreferenceInfo; 1705 1706 /// JumpBufSize - The size, in bytes, of the target's jmp_buf buffers 1707 unsigned JumpBufSize; 1708 1709 /// JumpBufAlignment - The alignment, in bytes, of the target's jmp_buf 1710 /// buffers 1711 unsigned JumpBufAlignment; 1712 1713 /// MinStackArgumentAlignment - The minimum alignment that any argument 1714 /// on the stack needs to have. 1715 /// 1716 unsigned MinStackArgumentAlignment; 1717 1718 /// MinFunctionAlignment - The minimum function alignment (used when 1719 /// optimizing for size, and to prevent explicitly provided alignment 1720 /// from leading to incorrect code). 1721 /// 1722 unsigned MinFunctionAlignment; 1723 1724 /// PrefFunctionAlignment - The preferred function alignment (used when 1725 /// alignment unspecified and optimizing for speed). 1726 /// 1727 unsigned PrefFunctionAlignment; 1728 1729 /// PrefLoopAlignment - The preferred loop alignment. 1730 /// 1731 unsigned PrefLoopAlignment; 1732 1733 /// ShouldFoldAtomicFences - Whether fencing MEMBARRIER instructions should 1734 /// be folded into the enclosed atomic intrinsic instruction by the 1735 /// combiner. 1736 bool ShouldFoldAtomicFences; 1737 1738 /// InsertFencesForAtomic - Whether the DAG builder should automatically 1739 /// insert fences and reduce ordering for atomics. (This will be set for 1740 /// for most architectures with weak memory ordering.) 1741 bool InsertFencesForAtomic; 1742 1743 /// StackPointerRegisterToSaveRestore - If set to a physical register, this 1744 /// specifies the register that llvm.savestack/llvm.restorestack should save 1745 /// and restore. 1746 unsigned StackPointerRegisterToSaveRestore; 1747 1748 /// ExceptionPointerRegister - If set to a physical register, this specifies 1749 /// the register that receives the exception address on entry to a landing 1750 /// pad. 1751 unsigned ExceptionPointerRegister; 1752 1753 /// ExceptionSelectorRegister - If set to a physical register, this specifies 1754 /// the register that receives the exception typeid on entry to a landing 1755 /// pad. 1756 unsigned ExceptionSelectorRegister; 1757 1758 /// RegClassForVT - This indicates the default register class to use for 1759 /// each ValueType the target supports natively. 1760 TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE]; 1761 unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE]; 1762 EVT RegisterTypeForVT[MVT::LAST_VALUETYPE]; 1763 1764 /// RepRegClassForVT - This indicates the "representative" register class to 1765 /// use for each ValueType the target supports natively. This information is 1766 /// used by the scheduler to track register pressure. By default, the 1767 /// representative register class is the largest legal super-reg register 1768 /// class of the register class of the specified type. e.g. On x86, i8, i16, 1769 /// and i32's representative class would be GR32. 1770 const TargetRegisterClass *RepRegClassForVT[MVT::LAST_VALUETYPE]; 1771 1772 /// RepRegClassCostForVT - This indicates the "cost" of the "representative" 1773 /// register class for each ValueType. The cost is used by the scheduler to 1774 /// approximate register pressure. 1775 uint8_t RepRegClassCostForVT[MVT::LAST_VALUETYPE]; 1776 1777 /// TransformToType - For any value types we are promoting or expanding, this 1778 /// contains the value type that we are changing to. For Expanded types, this 1779 /// contains one step of the expand (e.g. i64 -> i32), even if there are 1780 /// multiple steps required (e.g. i64 -> i16). For types natively supported 1781 /// by the system, this holds the same type (e.g. i32 -> i32). 1782 EVT TransformToType[MVT::LAST_VALUETYPE]; 1783 1784 /// OpActions - For each operation and each value type, keep a LegalizeAction 1785 /// that indicates how instruction selection should deal with the operation. 1786 /// Most operations are Legal (aka, supported natively by the target), but 1787 /// operations that are not should be described. Note that operations on 1788 /// non-legal value types are not described here. 1789 uint8_t OpActions[MVT::LAST_VALUETYPE][ISD::BUILTIN_OP_END]; 1790 1791 /// LoadExtActions - For each load extension type and each value type, 1792 /// keep a LegalizeAction that indicates how instruction selection should deal 1793 /// with a load of a specific value type and extension type. 1794 uint8_t LoadExtActions[MVT::LAST_VALUETYPE][ISD::LAST_LOADEXT_TYPE]; 1795 1796 /// TruncStoreActions - For each value type pair keep a LegalizeAction that 1797 /// indicates whether a truncating store of a specific value type and 1798 /// truncating type is legal. 1799 uint8_t TruncStoreActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE]; 1800 1801 /// IndexedModeActions - For each indexed mode and each value type, 1802 /// keep a pair of LegalizeAction that indicates how instruction 1803 /// selection should deal with the load / store. The first dimension is the 1804 /// value_type for the reference. The second dimension represents the various 1805 /// modes for load store. 1806 uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][ISD::LAST_INDEXED_MODE]; 1807 1808 /// CondCodeActions - For each condition code (ISD::CondCode) keep a 1809 /// LegalizeAction that indicates how instruction selection should 1810 /// deal with the condition code. 1811 uint64_t CondCodeActions[ISD::SETCC_INVALID]; 1812 1813 ValueTypeActionImpl ValueTypeActions; 1814 1815 typedef std::pair<LegalizeTypeAction, EVT> LegalizeKind; 1816 1817 LegalizeKind 1818 getTypeConversion(LLVMContext &Context, EVT VT) const { 1819 // If this is a simple type, use the ComputeRegisterProp mechanism. 1820 if (VT.isSimple()) { 1821 assert((unsigned)VT.getSimpleVT().SimpleTy < 1822 array_lengthof(TransformToType)); 1823 EVT NVT = TransformToType[VT.getSimpleVT().SimpleTy]; 1824 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(VT.getSimpleVT()); 1825 1826 assert( 1827 (!(NVT.isSimple() && LA != TypeLegal) || 1828 ValueTypeActions.getTypeAction(NVT.getSimpleVT()) != TypePromoteInteger) 1829 && "Promote may not follow Expand or Promote"); 1830 1831 return LegalizeKind(LA, NVT); 1832 } 1833 1834 // Handle Extended Scalar Types. 1835 if (!VT.isVector()) { 1836 assert(VT.isInteger() && "Float types must be simple"); 1837 unsigned BitSize = VT.getSizeInBits(); 1838 // First promote to a power-of-two size, then expand if necessary. 1839 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 1840 EVT NVT = VT.getRoundIntegerType(Context); 1841 assert(NVT != VT && "Unable to round integer VT"); 1842 LegalizeKind NextStep = getTypeConversion(Context, NVT); 1843 // Avoid multi-step promotion. 1844 if (NextStep.first == TypePromoteInteger) return NextStep; 1845 // Return rounded integer type. 1846 return LegalizeKind(TypePromoteInteger, NVT); 1847 } 1848 1849 return LegalizeKind(TypeExpandInteger, 1850 EVT::getIntegerVT(Context, VT.getSizeInBits()/2)); 1851 } 1852 1853 // Handle vector types. 1854 unsigned NumElts = VT.getVectorNumElements(); 1855 EVT EltVT = VT.getVectorElementType(); 1856 1857 // Vectors with only one element are always scalarized. 1858 if (NumElts == 1) 1859 return LegalizeKind(TypeScalarizeVector, EltVT); 1860 1861 // If we allow the promotion of vector elements using a flag, 1862 // then try to widen vector elements until a legal type is found. 1863 if (mayPromoteElements && EltVT.isInteger()) { 1864 // Vectors with a number of elements that is not a power of two are always 1865 // widened, for example <3 x float> -> <4 x float>. 1866 if (!VT.isPow2VectorType()) { 1867 NumElts = (unsigned)NextPowerOf2(NumElts); 1868 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 1869 return LegalizeKind(TypeWidenVector, NVT); 1870 } 1871 1872 // Examine the element type. 1873 LegalizeKind LK = getTypeConversion(Context, EltVT); 1874 1875 // If type is to be expanded, split the vector. 1876 // <4 x i140> -> <2 x i140> 1877 if (LK.first == TypeExpandInteger) 1878 return LegalizeKind(TypeSplitVector, 1879 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 1880 1881 // Promote the integer element types until a legal vector type is found 1882 // or until the element integer type is too big. If a legal type was not 1883 // found, fallback to the usual mechanism of widening/splitting the 1884 // vector. 1885 while (1) { 1886 // Increase the bitwidth of the element to the next pow-of-two 1887 // (which is greater than 8 bits). 1888 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits() 1889 ).getRoundIntegerType(Context); 1890 1891 // Stop trying when getting a non-simple element type. 1892 // Note that vector elements may be greater than legal vector element 1893 // types. Example: X86 XMM registers hold 64bit element on 32bit systems. 1894 if (!EltVT.isSimple()) break; 1895 1896 // Build a new vector type and check if it is legal. 1897 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 1898 // Found a legal promoted vector type. 1899 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 1900 return LegalizeKind(TypePromoteInteger, 1901 EVT::getVectorVT(Context, EltVT, NumElts)); 1902 } 1903 } 1904 1905 // Try to widen the vector until a legal type is found. 1906 // If there is no wider legal type, split the vector. 1907 while (1) { 1908 // Round up to the next power of 2. 1909 NumElts = (unsigned)NextPowerOf2(NumElts); 1910 1911 // If there is no simple vector type with this many elements then there 1912 // cannot be a larger legal vector type. Note that this assumes that 1913 // there are no skipped intermediate vector types in the simple types. 1914 if (!EltVT.isSimple()) break; 1915 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 1916 if (LargerVector == MVT()) break; 1917 1918 // If this type is legal then widen the vector. 1919 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 1920 return LegalizeKind(TypeWidenVector, LargerVector); 1921 } 1922 1923 // Widen odd vectors to next power of two. 1924 if (!VT.isPow2VectorType()) { 1925 EVT NVT = VT.getPow2VectorType(Context); 1926 return LegalizeKind(TypeWidenVector, NVT); 1927 } 1928 1929 // Vectors with illegal element types are expanded. 1930 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 1931 return LegalizeKind(TypeSplitVector, NVT); 1932 1933 assert(false && "Unable to handle this kind of vector type"); 1934 return LegalizeKind(TypeLegal, VT); 1935 } 1936 1937 std::vector<std::pair<EVT, TargetRegisterClass*> > AvailableRegClasses; 1938 1939 /// TargetDAGCombineArray - Targets can specify ISD nodes that they would 1940 /// like PerformDAGCombine callbacks for by calling setTargetDAGCombine(), 1941 /// which sets a bit in this array. 1942 unsigned char 1943 TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT]; 1944 1945 /// PromoteToType - For operations that must be promoted to a specific type, 1946 /// this holds the destination type. This map should be sparse, so don't hold 1947 /// it as an array. 1948 /// 1949 /// Targets add entries to this map with AddPromotedToType(..), clients access 1950 /// this with getTypeToPromoteTo(..). 1951 std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType> 1952 PromoteToType; 1953 1954 /// LibcallRoutineNames - Stores the name each libcall. 1955 /// 1956 const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL]; 1957 1958 /// CmpLibcallCCs - The ISD::CondCode that should be used to test the result 1959 /// of each of the comparison libcall against zero. 1960 ISD::CondCode CmpLibcallCCs[RTLIB::UNKNOWN_LIBCALL]; 1961 1962 /// LibcallCallingConvs - Stores the CallingConv that should be used for each 1963 /// libcall. 1964 CallingConv::ID LibcallCallingConvs[RTLIB::UNKNOWN_LIBCALL]; 1965 1966protected: 1967 /// When lowering \@llvm.memset this field specifies the maximum number of 1968 /// store operations that may be substituted for the call to memset. Targets 1969 /// must set this value based on the cost threshold for that target. Targets 1970 /// should assume that the memset will be done using as many of the largest 1971 /// store operations first, followed by smaller ones, if necessary, per 1972 /// alignment restrictions. For example, storing 9 bytes on a 32-bit machine 1973 /// with 16-bit alignment would result in four 2-byte stores and one 1-byte 1974 /// store. This only applies to setting a constant array of a constant size. 1975 /// @brief Specify maximum number of store instructions per memset call. 1976 unsigned maxStoresPerMemset; 1977 1978 /// Maximum number of stores operations that may be substituted for the call 1979 /// to memset, used for functions with OptSize attribute. 1980 unsigned maxStoresPerMemsetOptSize; 1981 1982 /// When lowering \@llvm.memcpy this field specifies the maximum number of 1983 /// store operations that may be substituted for a call to memcpy. Targets 1984 /// must set this value based on the cost threshold for that target. Targets 1985 /// should assume that the memcpy will be done using as many of the largest 1986 /// store operations first, followed by smaller ones, if necessary, per 1987 /// alignment restrictions. For example, storing 7 bytes on a 32-bit machine 1988 /// with 32-bit alignment would result in one 4-byte store, a one 2-byte store 1989 /// and one 1-byte store. This only applies to copying a constant array of 1990 /// constant size. 1991 /// @brief Specify maximum bytes of store instructions per memcpy call. 1992 unsigned maxStoresPerMemcpy; 1993 1994 /// Maximum number of store operations that may be substituted for a call 1995 /// to memcpy, used for functions with OptSize attribute. 1996 unsigned maxStoresPerMemcpyOptSize; 1997 1998 /// When lowering \@llvm.memmove this field specifies the maximum number of 1999 /// store instructions that may be substituted for a call to memmove. Targets 2000 /// must set this value based on the cost threshold for that target. Targets 2001 /// should assume that the memmove will be done using as many of the largest 2002 /// store operations first, followed by smaller ones, if necessary, per 2003 /// alignment restrictions. For example, moving 9 bytes on a 32-bit machine 2004 /// with 8-bit alignment would result in nine 1-byte stores. This only 2005 /// applies to copying a constant array of constant size. 2006 /// @brief Specify maximum bytes of store instructions per memmove call. 2007 unsigned maxStoresPerMemmove; 2008 2009 /// Maximum number of store instructions that may be substituted for a call 2010 /// to memmove, used for functions with OpSize attribute. 2011 unsigned maxStoresPerMemmoveOptSize; 2012 2013 /// This field specifies whether the target can benefit from code placement 2014 /// optimization. 2015 bool benefitFromCodePlacementOpt; 2016 2017private: 2018 /// isLegalRC - Return true if the value types that can be represented by the 2019 /// specified register class are all legal. 2020 bool isLegalRC(const TargetRegisterClass *RC) const; 2021 2022 /// hasLegalSuperRegRegClasses - Return true if the specified register class 2023 /// has one or more super-reg register classes that are legal. 2024 bool hasLegalSuperRegRegClasses(const TargetRegisterClass *RC) const; 2025}; 2026 2027/// GetReturnInfo - Given an LLVM IR type and return type attributes, 2028/// compute the return value EVTs and flags, and optionally also 2029/// the offsets, if the return value is being lowered to memory. 2030void GetReturnInfo(Type* ReturnType, Attributes attr, 2031 SmallVectorImpl<ISD::OutputArg> &Outs, 2032 const TargetLowering &TLI, 2033 SmallVectorImpl<uint64_t> *Offsets = 0); 2034 2035} // end llvm namespace 2036 2037#endif 2038