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 /// \file 11 /// This file describes how to lower LLVM code to machine code. This has two 12 /// main components: 13 /// 14 /// 1. Which ValueTypes are natively supported by the target. 15 /// 2. Which operations are supported for supported ValueTypes. 16 /// 3. Cost thresholds for alternative implementations of certain operations. 17 /// 18 /// In addition it has a few other components, like information about FP 19 /// immediates. 20 /// 21 //===----------------------------------------------------------------------===// 22 23 #ifndef LLVM_TARGET_TARGETLOWERING_H 24 #define LLVM_TARGET_TARGETLOWERING_H 25 26 #include "llvm/ADT/DenseMap.h" 27 #include "llvm/CodeGen/DAGCombine.h" 28 #include "llvm/CodeGen/RuntimeLibcalls.h" 29 #include "llvm/CodeGen/SelectionDAGNodes.h" 30 #include "llvm/IR/Attributes.h" 31 #include "llvm/IR/CallSite.h" 32 #include "llvm/IR/CallingConv.h" 33 #include "llvm/IR/InlineAsm.h" 34 #include "llvm/IR/IRBuilder.h" 35 #include "llvm/MC/MCRegisterInfo.h" 36 #include "llvm/Target/TargetCallingConv.h" 37 #include "llvm/Target/TargetMachine.h" 38 #include <climits> 39 #include <map> 40 #include <vector> 41 42 namespace llvm { 43 class CallInst; 44 class CCState; 45 class FastISel; 46 class FunctionLoweringInfo; 47 class ImmutableCallSite; 48 class IntrinsicInst; 49 class MachineBasicBlock; 50 class MachineFunction; 51 class MachineInstr; 52 class MachineJumpTableInfo; 53 class Mangler; 54 class MCContext; 55 class MCExpr; 56 class MCSymbol; 57 template<typename T> class SmallVectorImpl; 58 class DataLayout; 59 class TargetRegisterClass; 60 class TargetLibraryInfo; 61 class TargetLoweringObjectFile; 62 class Value; 63 64 namespace Sched { 65 enum Preference { 66 None, // No preference 67 Source, // Follow source order. 68 RegPressure, // Scheduling for lowest register pressure. 69 Hybrid, // Scheduling for both latency and register pressure. 70 ILP, // Scheduling for ILP in low register pressure mode. 71 VLIW // Scheduling for VLIW targets. 72 }; 73 } 74 75 /// This base class for TargetLowering contains the SelectionDAG-independent 76 /// parts that can be used from the rest of CodeGen. 77 class TargetLoweringBase { 78 TargetLoweringBase(const TargetLoweringBase&) LLVM_DELETED_FUNCTION; 79 void operator=(const TargetLoweringBase&) LLVM_DELETED_FUNCTION; 80 81 public: 82 /// This enum indicates whether operations are valid for a target, and if not, 83 /// what action should be used to make them valid. 84 enum LegalizeAction { 85 Legal, // The target natively supports this operation. 86 Promote, // This operation should be executed in a larger type. 87 Expand, // Try to expand this to other ops, otherwise use a libcall. 88 Custom // Use the LowerOperation hook to implement custom lowering. 89 }; 90 91 /// This enum indicates whether a types are legal for a target, and if not, 92 /// what action should be used to make them valid. 93 enum LegalizeTypeAction { 94 TypeLegal, // The target natively supports this type. 95 TypePromoteInteger, // Replace this integer with a larger one. 96 TypeExpandInteger, // Split this integer into two of half the size. 97 TypeSoftenFloat, // Convert this float to a same size integer type. 98 TypeExpandFloat, // Split this float into two of half the size. 99 TypeScalarizeVector, // Replace this one-element vector with its element. 100 TypeSplitVector, // Split this vector into two of half the size. 101 TypeWidenVector // This vector should be widened into a larger vector. 102 }; 103 104 /// LegalizeKind holds the legalization kind that needs to happen to EVT 105 /// in order to type-legalize it. 106 typedef std::pair<LegalizeTypeAction, EVT> LegalizeKind; 107 108 /// Enum that describes how the target represents true/false values. 109 enum BooleanContent { 110 UndefinedBooleanContent, // Only bit 0 counts, the rest can hold garbage. 111 ZeroOrOneBooleanContent, // All bits zero except for bit 0. 112 ZeroOrNegativeOneBooleanContent // All bits equal to bit 0. 113 }; 114 115 /// Enum that describes what type of support for selects the target has. 116 enum SelectSupportKind { 117 ScalarValSelect, // The target supports scalar selects (ex: cmov). 118 ScalarCondVectorVal, // The target supports selects with a scalar condition 119 // and vector values (ex: cmov). 120 VectorMaskSelect // The target supports vector selects with a vector 121 // mask (ex: x86 blends). 122 }; 123 124 static ISD::NodeType getExtendForContent(BooleanContent Content) { 125 switch (Content) { 126 case UndefinedBooleanContent: 127 // Extend by adding rubbish bits. 128 return ISD::ANY_EXTEND; 129 case ZeroOrOneBooleanContent: 130 // Extend by adding zero bits. 131 return ISD::ZERO_EXTEND; 132 case ZeroOrNegativeOneBooleanContent: 133 // Extend by copying the sign bit. 134 return ISD::SIGN_EXTEND; 135 } 136 llvm_unreachable("Invalid content kind"); 137 } 138 139 /// NOTE: The constructor takes ownership of TLOF. 140 explicit TargetLoweringBase(const TargetMachine &TM, 141 const TargetLoweringObjectFile *TLOF); 142 virtual ~TargetLoweringBase(); 143 144 protected: 145 /// \brief Initialize all of the actions to default values. 146 void initActions(); 147 148 public: 149 const TargetMachine &getTargetMachine() const { return TM; } 150 const DataLayout *getDataLayout() const { return DL; } 151 const TargetLoweringObjectFile &getObjFileLowering() const { return TLOF; } 152 153 bool isBigEndian() const { return !IsLittleEndian; } 154 bool isLittleEndian() const { return IsLittleEndian; } 155 156 /// Return the pointer type for the given address space, defaults to 157 /// the pointer type from the data layout. 158 /// FIXME: The default needs to be removed once all the code is updated. 159 virtual MVT getPointerTy(uint32_t /*AS*/ = 0) const; 160 unsigned getPointerSizeInBits(uint32_t AS = 0) const; 161 unsigned getPointerTypeSizeInBits(Type *Ty) const; 162 virtual MVT getScalarShiftAmountTy(EVT LHSTy) const; 163 164 EVT getShiftAmountTy(EVT LHSTy) const; 165 166 /// Returns the type to be used for the index operand of: 167 /// ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT, 168 /// ISD::INSERT_SUBVECTOR, and ISD::EXTRACT_SUBVECTOR 169 virtual MVT getVectorIdxTy() const { 170 return getPointerTy(); 171 } 172 173 /// Return true if the select operation is expensive for this target. 174 bool isSelectExpensive() const { return SelectIsExpensive; } 175 176 virtual bool isSelectSupported(SelectSupportKind /*kind*/) const { 177 return true; 178 } 179 180 /// Return true if multiple condition registers are available. 181 bool hasMultipleConditionRegisters() const { 182 return HasMultipleConditionRegisters; 183 } 184 185 /// Return true if the target has BitExtract instructions. 186 bool hasExtractBitsInsn() const { return HasExtractBitsInsn; } 187 188 /// Return the preferred vector type legalization action. 189 virtual TargetLoweringBase::LegalizeTypeAction 190 getPreferredVectorAction(EVT VT) const { 191 // The default action for one element vectors is to scalarize 192 if (VT.getVectorNumElements() == 1) 193 return TypeScalarizeVector; 194 // The default action for other vectors is to promote 195 return TypePromoteInteger; 196 } 197 198 // There are two general methods for expanding a BUILD_VECTOR node: 199 // 1. Use SCALAR_TO_VECTOR on the defined scalar values and then shuffle 200 // them together. 201 // 2. Build the vector on the stack and then load it. 202 // If this function returns true, then method (1) will be used, subject to 203 // the constraint that all of the necessary shuffles are legal (as determined 204 // by isShuffleMaskLegal). If this function returns false, then method (2) is 205 // always used. The vector type, and the number of defined values, are 206 // provided. 207 virtual bool 208 shouldExpandBuildVectorWithShuffles(EVT /* VT */, 209 unsigned DefinedValues) const { 210 return DefinedValues < 3; 211 } 212 213 /// Return true if integer divide is usually cheaper than a sequence of 214 /// several shifts, adds, and multiplies for this target. 215 bool isIntDivCheap() const { return IntDivIsCheap; } 216 217 /// Returns true if target has indicated at least one type should be bypassed. 218 bool isSlowDivBypassed() const { return !BypassSlowDivWidths.empty(); } 219 220 /// Returns map of slow types for division or remainder with corresponding 221 /// fast types 222 const DenseMap<unsigned int, unsigned int> &getBypassSlowDivWidths() const { 223 return BypassSlowDivWidths; 224 } 225 226 /// Return true if pow2 div is cheaper than a chain of srl/add/sra. 227 bool isPow2DivCheap() const { return Pow2DivIsCheap; } 228 229 /// Return true if Flow Control is an expensive operation that should be 230 /// avoided. 231 bool isJumpExpensive() const { return JumpIsExpensive; } 232 233 /// Return true if selects are only cheaper than branches if the branch is 234 /// unlikely to be predicted right. 235 bool isPredictableSelectExpensive() const { 236 return PredictableSelectIsExpensive; 237 } 238 239 /// isLoadBitCastBeneficial() - Return true if the following transform 240 /// is beneficial. 241 /// fold (conv (load x)) -> (load (conv*)x) 242 /// On architectures that don't natively support some vector loads efficiently, 243 /// casting the load to a smaller vector of larger types and loading 244 /// is more efficient, however, this can be undone by optimizations in 245 /// dag combiner. 246 virtual bool isLoadBitCastBeneficial(EVT /* Load */, EVT /* Bitcast */) const { 247 return true; 248 } 249 250 /// \brief Return if the target supports combining a 251 /// chain like: 252 /// \code 253 /// %andResult = and %val1, #imm-with-one-bit-set; 254 /// %icmpResult = icmp %andResult, 0 255 /// br i1 %icmpResult, label %dest1, label %dest2 256 /// \endcode 257 /// into a single machine instruction of a form like: 258 /// \code 259 /// brOnBitSet %register, #bitNumber, dest 260 /// \endcode 261 bool isMaskAndBranchFoldingLegal() const { 262 return MaskAndBranchFoldingIsLegal; 263 } 264 265 /// Return the ValueType of the result of SETCC operations. Also used to 266 /// obtain the target's preferred type for the condition operand of SELECT and 267 /// BRCOND nodes. In the case of BRCOND the argument passed is MVT::Other 268 /// since there are no other operands to get a type hint from. 269 virtual EVT getSetCCResultType(LLVMContext &Context, EVT VT) const; 270 271 /// Return the ValueType for comparison libcalls. Comparions libcalls include 272 /// floating point comparion calls, and Ordered/Unordered check calls on 273 /// floating point numbers. 274 virtual 275 MVT::SimpleValueType getCmpLibcallReturnType() const; 276 277 /// For targets without i1 registers, this gives the nature of the high-bits 278 /// of boolean values held in types wider than i1. 279 /// 280 /// "Boolean values" are special true/false values produced by nodes like 281 /// SETCC and consumed (as the condition) by nodes like SELECT and BRCOND. 282 /// Not to be confused with general values promoted from i1. Some cpus 283 /// distinguish between vectors of boolean and scalars; the isVec parameter 284 /// selects between the two kinds. For example on X86 a scalar boolean should 285 /// be zero extended from i1, while the elements of a vector of booleans 286 /// should be sign extended from i1. 287 /// 288 /// Some cpus also treat floating point types the same way as they treat 289 /// vectors instead of the way they treat scalars. 290 BooleanContent getBooleanContents(bool isVec, bool isFloat) const { 291 if (isVec) 292 return BooleanVectorContents; 293 return isFloat ? BooleanFloatContents : BooleanContents; 294 } 295 296 BooleanContent getBooleanContents(EVT Type) const { 297 return getBooleanContents(Type.isVector(), Type.isFloatingPoint()); 298 } 299 300 /// Return target scheduling preference. 301 Sched::Preference getSchedulingPreference() const { 302 return SchedPreferenceInfo; 303 } 304 305 /// Some scheduler, e.g. hybrid, can switch to different scheduling heuristics 306 /// for different nodes. This function returns the preference (or none) for 307 /// the given node. 308 virtual Sched::Preference getSchedulingPreference(SDNode *) const { 309 return Sched::None; 310 } 311 312 /// Return the register class that should be used for the specified value 313 /// type. 314 virtual const TargetRegisterClass *getRegClassFor(MVT VT) const { 315 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 316 assert(RC && "This value type is not natively supported!"); 317 return RC; 318 } 319 320 /// Return the 'representative' register class for the specified value 321 /// type. 322 /// 323 /// The 'representative' register class is the largest legal super-reg 324 /// register class for the register class of the value type. For example, on 325 /// i386 the rep register class for i8, i16, and i32 are GR32; while the rep 326 /// register class is GR64 on x86_64. 327 virtual const TargetRegisterClass *getRepRegClassFor(MVT VT) const { 328 const TargetRegisterClass *RC = RepRegClassForVT[VT.SimpleTy]; 329 return RC; 330 } 331 332 /// Return the cost of the 'representative' register class for the specified 333 /// value type. 334 virtual uint8_t getRepRegClassCostFor(MVT VT) const { 335 return RepRegClassCostForVT[VT.SimpleTy]; 336 } 337 338 /// Return true if the target has native support for the specified value type. 339 /// This means that it has a register that directly holds it without 340 /// promotions or expansions. 341 bool isTypeLegal(EVT VT) const { 342 assert(!VT.isSimple() || 343 (unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT)); 344 return VT.isSimple() && RegClassForVT[VT.getSimpleVT().SimpleTy] != nullptr; 345 } 346 347 class ValueTypeActionImpl { 348 /// ValueTypeActions - For each value type, keep a LegalizeTypeAction enum 349 /// that indicates how instruction selection should deal with the type. 350 uint8_t ValueTypeActions[MVT::LAST_VALUETYPE]; 351 352 public: 353 ValueTypeActionImpl() { 354 std::fill(std::begin(ValueTypeActions), std::end(ValueTypeActions), 0); 355 } 356 357 LegalizeTypeAction getTypeAction(MVT VT) const { 358 return (LegalizeTypeAction)ValueTypeActions[VT.SimpleTy]; 359 } 360 361 void setTypeAction(MVT VT, LegalizeTypeAction Action) { 362 unsigned I = VT.SimpleTy; 363 ValueTypeActions[I] = Action; 364 } 365 }; 366 367 const ValueTypeActionImpl &getValueTypeActions() const { 368 return ValueTypeActions; 369 } 370 371 /// Return how we should legalize values of this type, either it is already 372 /// legal (return 'Legal') or we need to promote it to a larger type (return 373 /// 'Promote'), or we need to expand it into multiple registers of smaller 374 /// integer type (return 'Expand'). 'Custom' is not an option. 375 LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const { 376 return getTypeConversion(Context, VT).first; 377 } 378 LegalizeTypeAction getTypeAction(MVT VT) const { 379 return ValueTypeActions.getTypeAction(VT); 380 } 381 382 /// For types supported by the target, this is an identity function. For 383 /// types that must be promoted to larger types, this returns the larger type 384 /// to promote to. For integer types that are larger than the largest integer 385 /// register, this contains one step in the expansion to get to the smaller 386 /// register. For illegal floating point types, this returns the integer type 387 /// to transform to. 388 EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const { 389 return getTypeConversion(Context, VT).second; 390 } 391 392 /// For types supported by the target, this is an identity function. For 393 /// types that must be expanded (i.e. integer types that are larger than the 394 /// largest integer register or illegal floating point types), this returns 395 /// the largest legal type it will be expanded to. 396 EVT getTypeToExpandTo(LLVMContext &Context, EVT VT) const { 397 assert(!VT.isVector()); 398 while (true) { 399 switch (getTypeAction(Context, VT)) { 400 case TypeLegal: 401 return VT; 402 case TypeExpandInteger: 403 VT = getTypeToTransformTo(Context, VT); 404 break; 405 default: 406 llvm_unreachable("Type is not legal nor is it to be expanded!"); 407 } 408 } 409 } 410 411 /// Vector types are broken down into some number of legal first class types. 412 /// For example, EVT::v8f32 maps to 2 EVT::v4f32 with Altivec or SSE1, or 8 413 /// promoted EVT::f64 values with the X86 FP stack. Similarly, EVT::v2i64 414 /// turns into 4 EVT::i32 values with both PPC and X86. 415 /// 416 /// This method returns the number of registers needed, and the VT for each 417 /// register. It also returns the VT and quantity of the intermediate values 418 /// before they are promoted/expanded. 419 unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 420 EVT &IntermediateVT, 421 unsigned &NumIntermediates, 422 MVT &RegisterVT) const; 423 424 struct IntrinsicInfo { 425 unsigned opc; // target opcode 426 EVT memVT; // memory VT 427 const Value* ptrVal; // value representing memory location 428 int offset; // offset off of ptrVal 429 unsigned align; // alignment 430 bool vol; // is volatile? 431 bool readMem; // reads memory? 432 bool writeMem; // writes memory? 433 }; 434 435 /// Given an intrinsic, checks if on the target the intrinsic will need to map 436 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns 437 /// true and store the intrinsic information into the IntrinsicInfo that was 438 /// passed to the function. 439 virtual bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &, 440 unsigned /*Intrinsic*/) const { 441 return false; 442 } 443 444 /// Returns true if the target can instruction select the specified FP 445 /// immediate natively. If false, the legalizer will materialize the FP 446 /// immediate as a load from a constant pool. 447 virtual bool isFPImmLegal(const APFloat &/*Imm*/, EVT /*VT*/) const { 448 return false; 449 } 450 451 /// Targets can use this to indicate that they only support *some* 452 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a 453 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to be 454 /// legal. 455 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 456 EVT /*VT*/) const { 457 return true; 458 } 459 460 /// Returns true if the operation can trap for the value type. 461 /// 462 /// VT must be a legal type. By default, we optimistically assume most 463 /// operations don't trap except for divide and remainder. 464 virtual bool canOpTrap(unsigned Op, EVT VT) const; 465 466 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to 467 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to replace 468 /// a VAND with a constant pool entry. 469 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 470 EVT /*VT*/) const { 471 return false; 472 } 473 474 /// Return how this operation should be treated: either it is legal, needs to 475 /// be promoted to a larger size, needs to be expanded to some other code 476 /// sequence, or the target has a custom expander for it. 477 LegalizeAction getOperationAction(unsigned Op, EVT VT) const { 478 if (VT.isExtended()) return Expand; 479 // If a target-specific SDNode requires legalization, require the target 480 // to provide custom legalization for it. 481 if (Op > array_lengthof(OpActions[0])) return Custom; 482 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy; 483 return (LegalizeAction)OpActions[I][Op]; 484 } 485 486 /// Return true if the specified operation is legal on this target or can be 487 /// made legal with custom lowering. This is used to help guide high-level 488 /// lowering decisions. 489 bool isOperationLegalOrCustom(unsigned Op, EVT VT) const { 490 return (VT == MVT::Other || isTypeLegal(VT)) && 491 (getOperationAction(Op, VT) == Legal || 492 getOperationAction(Op, VT) == Custom); 493 } 494 495 /// Return true if the specified operation is legal on this target or can be 496 /// made legal using promotion. This is used to help guide high-level lowering 497 /// decisions. 498 bool isOperationLegalOrPromote(unsigned Op, EVT VT) const { 499 return (VT == MVT::Other || isTypeLegal(VT)) && 500 (getOperationAction(Op, VT) == Legal || 501 getOperationAction(Op, VT) == Promote); 502 } 503 504 /// Return true if the specified operation is illegal on this target or 505 /// unlikely to be made legal with custom lowering. This is used to help guide 506 /// high-level lowering decisions. 507 bool isOperationExpand(unsigned Op, EVT VT) const { 508 return (!isTypeLegal(VT) || getOperationAction(Op, VT) == Expand); 509 } 510 511 /// Return true if the specified operation is legal on this target. 512 bool isOperationLegal(unsigned Op, EVT VT) const { 513 return (VT == MVT::Other || isTypeLegal(VT)) && 514 getOperationAction(Op, VT) == Legal; 515 } 516 517 /// Return how this load with extension should be treated: either it is legal, 518 /// needs to be promoted to a larger size, needs to be expanded to some other 519 /// code sequence, or the target has a custom expander for it. 520 LegalizeAction getLoadExtAction(unsigned ExtType, MVT VT) const { 521 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE && 522 "Table isn't big enough!"); 523 return (LegalizeAction)LoadExtActions[VT.SimpleTy][ExtType]; 524 } 525 526 /// Return true if the specified load with extension is legal on this target. 527 bool isLoadExtLegal(unsigned ExtType, EVT VT) const { 528 return VT.isSimple() && 529 getLoadExtAction(ExtType, VT.getSimpleVT()) == Legal; 530 } 531 532 /// Return how this store with truncation should be treated: either it is 533 /// legal, needs to be promoted to a larger size, needs to be expanded to some 534 /// other code sequence, or the target has a custom expander for it. 535 LegalizeAction getTruncStoreAction(MVT ValVT, MVT MemVT) const { 536 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE && 537 "Table isn't big enough!"); 538 return (LegalizeAction)TruncStoreActions[ValVT.SimpleTy] 539 [MemVT.SimpleTy]; 540 } 541 542 /// Return true if the specified store with truncation is legal on this 543 /// target. 544 bool isTruncStoreLegal(EVT ValVT, EVT MemVT) const { 545 return isTypeLegal(ValVT) && MemVT.isSimple() && 546 getTruncStoreAction(ValVT.getSimpleVT(), MemVT.getSimpleVT()) == Legal; 547 } 548 549 /// Return how the indexed load should be treated: either it is legal, needs 550 /// to be promoted to a larger size, needs to be expanded to some other code 551 /// sequence, or the target has a custom expander for it. 552 LegalizeAction 553 getIndexedLoadAction(unsigned IdxMode, MVT VT) const { 554 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE && 555 "Table isn't big enough!"); 556 unsigned Ty = (unsigned)VT.SimpleTy; 557 return (LegalizeAction)((IndexedModeActions[Ty][IdxMode] & 0xf0) >> 4); 558 } 559 560 /// Return true if the specified indexed load is legal on this target. 561 bool isIndexedLoadLegal(unsigned IdxMode, EVT VT) const { 562 return VT.isSimple() && 563 (getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Legal || 564 getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Custom); 565 } 566 567 /// Return how the indexed store should be treated: either it is legal, needs 568 /// to be promoted to a larger size, needs to be expanded to some other code 569 /// sequence, or the target has a custom expander for it. 570 LegalizeAction 571 getIndexedStoreAction(unsigned IdxMode, MVT VT) const { 572 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT < MVT::LAST_VALUETYPE && 573 "Table isn't big enough!"); 574 unsigned Ty = (unsigned)VT.SimpleTy; 575 return (LegalizeAction)(IndexedModeActions[Ty][IdxMode] & 0x0f); 576 } 577 578 /// Return true if the specified indexed load is legal on this target. 579 bool isIndexedStoreLegal(unsigned IdxMode, EVT VT) const { 580 return VT.isSimple() && 581 (getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Legal || 582 getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Custom); 583 } 584 585 /// Return how the condition code should be treated: either it is legal, needs 586 /// to be expanded to some other code sequence, or the target has a custom 587 /// expander for it. 588 LegalizeAction 589 getCondCodeAction(ISD::CondCode CC, MVT VT) const { 590 assert((unsigned)CC < array_lengthof(CondCodeActions) && 591 ((unsigned)VT.SimpleTy >> 4) < array_lengthof(CondCodeActions[0]) && 592 "Table isn't big enough!"); 593 // See setCondCodeAction for how this is encoded. 594 uint32_t Shift = 2 * (VT.SimpleTy & 0xF); 595 uint32_t Value = CondCodeActions[CC][VT.SimpleTy >> 4]; 596 LegalizeAction Action = (LegalizeAction) ((Value >> Shift) & 0x3); 597 assert(Action != Promote && "Can't promote condition code!"); 598 return Action; 599 } 600 601 /// Return true if the specified condition code is legal on this target. 602 bool isCondCodeLegal(ISD::CondCode CC, MVT VT) const { 603 return 604 getCondCodeAction(CC, VT) == Legal || 605 getCondCodeAction(CC, VT) == Custom; 606 } 607 608 609 /// If the action for this operation is to promote, this method returns the 610 /// ValueType to promote to. 611 MVT getTypeToPromoteTo(unsigned Op, MVT VT) const { 612 assert(getOperationAction(Op, VT) == Promote && 613 "This operation isn't promoted!"); 614 615 // See if this has an explicit type specified. 616 std::map<std::pair<unsigned, MVT::SimpleValueType>, 617 MVT::SimpleValueType>::const_iterator PTTI = 618 PromoteToType.find(std::make_pair(Op, VT.SimpleTy)); 619 if (PTTI != PromoteToType.end()) return PTTI->second; 620 621 assert((VT.isInteger() || VT.isFloatingPoint()) && 622 "Cannot autopromote this type, add it with AddPromotedToType."); 623 624 MVT NVT = VT; 625 do { 626 NVT = (MVT::SimpleValueType)(NVT.SimpleTy+1); 627 assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid && 628 "Didn't find type to promote to!"); 629 } while (!isTypeLegal(NVT) || 630 getOperationAction(Op, NVT) == Promote); 631 return NVT; 632 } 633 634 /// Return the EVT corresponding to this LLVM type. This is fixed by the LLVM 635 /// operations except for the pointer size. If AllowUnknown is true, this 636 /// will return MVT::Other for types with no EVT counterpart (e.g. structs), 637 /// otherwise it will assert. 638 EVT getValueType(Type *Ty, bool AllowUnknown = false) const { 639 // Lower scalar pointers to native pointer types. 640 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 641 return getPointerTy(PTy->getAddressSpace()); 642 643 if (Ty->isVectorTy()) { 644 VectorType *VTy = cast<VectorType>(Ty); 645 Type *Elm = VTy->getElementType(); 646 // Lower vectors of pointers to native pointer types. 647 if (PointerType *PT = dyn_cast<PointerType>(Elm)) { 648 EVT PointerTy(getPointerTy(PT->getAddressSpace())); 649 Elm = PointerTy.getTypeForEVT(Ty->getContext()); 650 } 651 652 return EVT::getVectorVT(Ty->getContext(), EVT::getEVT(Elm, false), 653 VTy->getNumElements()); 654 } 655 return EVT::getEVT(Ty, AllowUnknown); 656 } 657 658 /// Return the MVT corresponding to this LLVM type. See getValueType. 659 MVT getSimpleValueType(Type *Ty, bool AllowUnknown = false) const { 660 return getValueType(Ty, AllowUnknown).getSimpleVT(); 661 } 662 663 /// Return the desired alignment for ByVal or InAlloca aggregate function 664 /// arguments in the caller parameter area. This is the actual alignment, not 665 /// its logarithm. 666 virtual unsigned getByValTypeAlignment(Type *Ty) const; 667 668 /// Return the type of registers that this ValueType will eventually require. 669 MVT getRegisterType(MVT VT) const { 670 assert((unsigned)VT.SimpleTy < array_lengthof(RegisterTypeForVT)); 671 return RegisterTypeForVT[VT.SimpleTy]; 672 } 673 674 /// Return the type of registers that this ValueType will eventually require. 675 MVT getRegisterType(LLVMContext &Context, EVT VT) const { 676 if (VT.isSimple()) { 677 assert((unsigned)VT.getSimpleVT().SimpleTy < 678 array_lengthof(RegisterTypeForVT)); 679 return RegisterTypeForVT[VT.getSimpleVT().SimpleTy]; 680 } 681 if (VT.isVector()) { 682 EVT VT1; 683 MVT RegisterVT; 684 unsigned NumIntermediates; 685 (void)getVectorTypeBreakdown(Context, VT, VT1, 686 NumIntermediates, RegisterVT); 687 return RegisterVT; 688 } 689 if (VT.isInteger()) { 690 return getRegisterType(Context, getTypeToTransformTo(Context, VT)); 691 } 692 llvm_unreachable("Unsupported extended type!"); 693 } 694 695 /// Return the number of registers that this ValueType will eventually 696 /// require. 697 /// 698 /// This is one for any types promoted to live in larger registers, but may be 699 /// more than one for types (like i64) that are split into pieces. For types 700 /// like i140, which are first promoted then expanded, it is the number of 701 /// registers needed to hold all the bits of the original type. For an i140 702 /// on a 32 bit machine this means 5 registers. 703 unsigned getNumRegisters(LLVMContext &Context, EVT VT) const { 704 if (VT.isSimple()) { 705 assert((unsigned)VT.getSimpleVT().SimpleTy < 706 array_lengthof(NumRegistersForVT)); 707 return NumRegistersForVT[VT.getSimpleVT().SimpleTy]; 708 } 709 if (VT.isVector()) { 710 EVT VT1; 711 MVT VT2; 712 unsigned NumIntermediates; 713 return getVectorTypeBreakdown(Context, VT, VT1, NumIntermediates, VT2); 714 } 715 if (VT.isInteger()) { 716 unsigned BitWidth = VT.getSizeInBits(); 717 unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits(); 718 return (BitWidth + RegWidth - 1) / RegWidth; 719 } 720 llvm_unreachable("Unsupported extended type!"); 721 } 722 723 /// If true, then instruction selection should seek to shrink the FP constant 724 /// of the specified type to a smaller type in order to save space and / or 725 /// reduce runtime. 726 virtual bool ShouldShrinkFPConstant(EVT) const { return true; } 727 728 /// When splitting a value of the specified type into parts, does the Lo 729 /// or Hi part come first? This usually follows the endianness, except 730 /// for ppcf128, where the Hi part always comes first. 731 bool hasBigEndianPartOrdering(EVT VT) const { 732 return isBigEndian() || VT == MVT::ppcf128; 733 } 734 735 /// If true, the target has custom DAG combine transformations that it can 736 /// perform for the specified node. 737 bool hasTargetDAGCombine(ISD::NodeType NT) const { 738 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 739 return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7)); 740 } 741 742 /// \brief Get maximum # of store operations permitted for llvm.memset 743 /// 744 /// This function returns the maximum number of store operations permitted 745 /// to replace a call to llvm.memset. The value is set by the target at the 746 /// performance threshold for such a replacement. If OptSize is true, 747 /// return the limit for functions that have OptSize attribute. 748 unsigned getMaxStoresPerMemset(bool OptSize) const { 749 return OptSize ? MaxStoresPerMemsetOptSize : MaxStoresPerMemset; 750 } 751 752 /// \brief Get maximum # of store operations permitted for llvm.memcpy 753 /// 754 /// This function returns the maximum number of store operations permitted 755 /// to replace a call to llvm.memcpy. The value is set by the target at the 756 /// performance threshold for such a replacement. If OptSize is true, 757 /// return the limit for functions that have OptSize attribute. 758 unsigned getMaxStoresPerMemcpy(bool OptSize) const { 759 return OptSize ? MaxStoresPerMemcpyOptSize : MaxStoresPerMemcpy; 760 } 761 762 /// \brief Get maximum # of store operations permitted for llvm.memmove 763 /// 764 /// This function returns the maximum number of store operations permitted 765 /// to replace a call to llvm.memmove. The value is set by the target at the 766 /// performance threshold for such a replacement. If OptSize is true, 767 /// return the limit for functions that have OptSize attribute. 768 unsigned getMaxStoresPerMemmove(bool OptSize) const { 769 return OptSize ? MaxStoresPerMemmoveOptSize : MaxStoresPerMemmove; 770 } 771 772 /// \brief Determine if the target supports unaligned memory accesses. 773 /// 774 /// This function returns true if the target allows unaligned memory accesses 775 /// of the specified type in the given address space. If true, it also returns 776 /// whether the unaligned memory access is "fast" in the third argument by 777 /// reference. This is used, for example, in situations where an array 778 /// copy/move/set is converted to a sequence of store operations. Its use 779 /// helps to ensure that such replacements don't generate code that causes an 780 /// alignment error (trap) on the target machine. 781 virtual bool allowsUnalignedMemoryAccesses(EVT, 782 unsigned AddrSpace = 0, 783 bool * /*Fast*/ = nullptr) const { 784 return false; 785 } 786 787 /// Returns the target specific optimal type for load and store operations as 788 /// a result of memset, memcpy, and memmove lowering. 789 /// 790 /// If DstAlign is zero that means it's safe to destination alignment can 791 /// satisfy any constraint. Similarly if SrcAlign is zero it means there isn't 792 /// a need to check it against alignment requirement, probably because the 793 /// source does not need to be loaded. If 'IsMemset' is true, that means it's 794 /// expanding a memset. If 'ZeroMemset' is true, that means it's a memset of 795 /// zero. 'MemcpyStrSrc' indicates whether the memcpy source is constant so it 796 /// does not need to be loaded. It returns EVT::Other if the type should be 797 /// determined using generic target-independent logic. 798 virtual EVT getOptimalMemOpType(uint64_t /*Size*/, 799 unsigned /*DstAlign*/, unsigned /*SrcAlign*/, 800 bool /*IsMemset*/, 801 bool /*ZeroMemset*/, 802 bool /*MemcpyStrSrc*/, 803 MachineFunction &/*MF*/) const { 804 return MVT::Other; 805 } 806 807 /// Returns true if it's safe to use load / store of the specified type to 808 /// expand memcpy / memset inline. 809 /// 810 /// This is mostly true for all types except for some special cases. For 811 /// example, on X86 targets without SSE2 f64 load / store are done with fldl / 812 /// fstpl which also does type conversion. Note the specified type doesn't 813 /// have to be legal as the hook is used before type legalization. 814 virtual bool isSafeMemOpType(MVT /*VT*/) const { return true; } 815 816 /// Determine if we should use _setjmp or setjmp to implement llvm.setjmp. 817 bool usesUnderscoreSetJmp() const { 818 return UseUnderscoreSetJmp; 819 } 820 821 /// Determine if we should use _longjmp or longjmp to implement llvm.longjmp. 822 bool usesUnderscoreLongJmp() const { 823 return UseUnderscoreLongJmp; 824 } 825 826 /// Return whether the target can generate code for jump tables. 827 bool supportJumpTables() const { 828 return SupportJumpTables; 829 } 830 831 /// Return integer threshold on number of blocks to use jump tables rather 832 /// than if sequence. 833 int getMinimumJumpTableEntries() const { 834 return MinimumJumpTableEntries; 835 } 836 837 /// If a physical register, this specifies the register that 838 /// llvm.savestack/llvm.restorestack should save and restore. 839 unsigned getStackPointerRegisterToSaveRestore() const { 840 return StackPointerRegisterToSaveRestore; 841 } 842 843 /// If a physical register, this returns the register that receives the 844 /// exception address on entry to a landing pad. 845 unsigned getExceptionPointerRegister() const { 846 return ExceptionPointerRegister; 847 } 848 849 /// If a physical register, this returns the register that receives the 850 /// exception typeid on entry to a landing pad. 851 unsigned getExceptionSelectorRegister() const { 852 return ExceptionSelectorRegister; 853 } 854 855 /// Returns the target's jmp_buf size in bytes (if never set, the default is 856 /// 200) 857 unsigned getJumpBufSize() const { 858 return JumpBufSize; 859 } 860 861 /// Returns the target's jmp_buf alignment in bytes (if never set, the default 862 /// is 0) 863 unsigned getJumpBufAlignment() const { 864 return JumpBufAlignment; 865 } 866 867 /// Return the minimum stack alignment of an argument. 868 unsigned getMinStackArgumentAlignment() const { 869 return MinStackArgumentAlignment; 870 } 871 872 /// Return the minimum function alignment. 873 unsigned getMinFunctionAlignment() const { 874 return MinFunctionAlignment; 875 } 876 877 /// Return the preferred function alignment. 878 unsigned getPrefFunctionAlignment() const { 879 return PrefFunctionAlignment; 880 } 881 882 /// Return the preferred loop alignment. 883 unsigned getPrefLoopAlignment() const { 884 return PrefLoopAlignment; 885 } 886 887 /// Return whether the DAG builder should automatically insert fences and 888 /// reduce ordering for atomics. 889 bool getInsertFencesForAtomic() const { 890 return InsertFencesForAtomic; 891 } 892 893 /// Return true if the target stores stack protector cookies at a fixed offset 894 /// in some non-standard address space, and populates the address space and 895 /// offset as appropriate. 896 virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/, 897 unsigned &/*Offset*/) const { 898 return false; 899 } 900 901 /// Returns the maximal possible offset which can be used for loads / stores 902 /// from the global. 903 virtual unsigned getMaximalGlobalOffset() const { 904 return 0; 905 } 906 907 /// Returns true if a cast between SrcAS and DestAS is a noop. 908 virtual bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const { 909 return false; 910 } 911 912 //===--------------------------------------------------------------------===// 913 /// \name Helpers for TargetTransformInfo implementations 914 /// @{ 915 916 /// Get the ISD node that corresponds to the Instruction class opcode. 917 int InstructionOpcodeToISD(unsigned Opcode) const; 918 919 /// Estimate the cost of type-legalization and the legalized type. 920 std::pair<unsigned, MVT> getTypeLegalizationCost(Type *Ty) const; 921 922 /// @} 923 924 //===--------------------------------------------------------------------===// 925 /// \name Helpers for load-linked/store-conditional atomic expansion. 926 /// @{ 927 928 /// Perform a load-linked operation on Addr, returning a "Value *" with the 929 /// corresponding pointee type. This may entail some non-trivial operations to 930 /// truncate or reconstruct types that will be illegal in the backend. See 931 /// ARMISelLowering for an example implementation. 932 virtual Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr, 933 AtomicOrdering Ord) const { 934 llvm_unreachable("Load linked unimplemented on this target"); 935 } 936 937 /// Perform a store-conditional operation to Addr. Return the status of the 938 /// store. This should be 0 if the store succeeded, non-zero otherwise. 939 virtual Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val, 940 Value *Addr, AtomicOrdering Ord) const { 941 llvm_unreachable("Store conditional unimplemented on this target"); 942 } 943 944 /// Return true if the given (atomic) instruction should be expanded by the 945 /// IR-level AtomicExpandLoadLinked pass into a loop involving 946 /// load-linked/store-conditional pairs. Atomic stores will be expanded in the 947 /// same way as "atomic xchg" operations which ignore their output if needed. 948 virtual bool shouldExpandAtomicInIR(Instruction *Inst) const { 949 return false; 950 } 951 952 953 //===--------------------------------------------------------------------===// 954 // TargetLowering Configuration Methods - These methods should be invoked by 955 // the derived class constructor to configure this object for the target. 956 // 957 958 /// \brief Reset the operation actions based on target options. 959 virtual void resetOperationActions() {} 960 961 protected: 962 /// Specify how the target extends the result of integer and floating point 963 /// boolean values from i1 to a wider type. See getBooleanContents. 964 void setBooleanContents(BooleanContent Ty) { 965 BooleanContents = Ty; 966 BooleanFloatContents = Ty; 967 } 968 969 /// Specify how the target extends the result of integer and floating point 970 /// boolean values from i1 to a wider type. See getBooleanContents. 971 void setBooleanContents(BooleanContent IntTy, BooleanContent FloatTy) { 972 BooleanContents = IntTy; 973 BooleanFloatContents = FloatTy; 974 } 975 976 /// Specify how the target extends the result of a vector boolean value from a 977 /// vector of i1 to a wider type. See getBooleanContents. 978 void setBooleanVectorContents(BooleanContent Ty) { 979 BooleanVectorContents = Ty; 980 } 981 982 /// Specify the target scheduling preference. 983 void setSchedulingPreference(Sched::Preference Pref) { 984 SchedPreferenceInfo = Pref; 985 } 986 987 /// Indicate whether this target prefers to use _setjmp to implement 988 /// llvm.setjmp or the version without _. Defaults to false. 989 void setUseUnderscoreSetJmp(bool Val) { 990 UseUnderscoreSetJmp = Val; 991 } 992 993 /// Indicate whether this target prefers to use _longjmp to implement 994 /// llvm.longjmp or the version without _. Defaults to false. 995 void setUseUnderscoreLongJmp(bool Val) { 996 UseUnderscoreLongJmp = Val; 997 } 998 999 /// Indicate whether the target can generate code for jump tables. 1000 void setSupportJumpTables(bool Val) { 1001 SupportJumpTables = Val; 1002 } 1003 1004 /// Indicate the number of blocks to generate jump tables rather than if 1005 /// sequence. 1006 void setMinimumJumpTableEntries(int Val) { 1007 MinimumJumpTableEntries = Val; 1008 } 1009 1010 /// If set to a physical register, this specifies the register that 1011 /// llvm.savestack/llvm.restorestack should save and restore. 1012 void setStackPointerRegisterToSaveRestore(unsigned R) { 1013 StackPointerRegisterToSaveRestore = R; 1014 } 1015 1016 /// If set to a physical register, this sets the register that receives the 1017 /// exception address on entry to a landing pad. 1018 void setExceptionPointerRegister(unsigned R) { 1019 ExceptionPointerRegister = R; 1020 } 1021 1022 /// If set to a physical register, this sets the register that receives the 1023 /// exception typeid on entry to a landing pad. 1024 void setExceptionSelectorRegister(unsigned R) { 1025 ExceptionSelectorRegister = R; 1026 } 1027 1028 /// Tells the code generator not to expand operations into sequences that use 1029 /// the select operations if possible. 1030 void setSelectIsExpensive(bool isExpensive = true) { 1031 SelectIsExpensive = isExpensive; 1032 } 1033 1034 /// Tells the code generator that the target has multiple (allocatable) 1035 /// condition registers that can be used to store the results of comparisons 1036 /// for use by selects and conditional branches. With multiple condition 1037 /// registers, the code generator will not aggressively sink comparisons into 1038 /// the blocks of their users. 1039 void setHasMultipleConditionRegisters(bool hasManyRegs = true) { 1040 HasMultipleConditionRegisters = hasManyRegs; 1041 } 1042 1043 /// Tells the code generator that the target has BitExtract instructions. 1044 /// The code generator will aggressively sink "shift"s into the blocks of 1045 /// their users if the users will generate "and" instructions which can be 1046 /// combined with "shift" to BitExtract instructions. 1047 void setHasExtractBitsInsn(bool hasExtractInsn = true) { 1048 HasExtractBitsInsn = hasExtractInsn; 1049 } 1050 1051 /// Tells the code generator not to expand sequence of operations into a 1052 /// separate sequences that increases the amount of flow control. 1053 void setJumpIsExpensive(bool isExpensive = true) { 1054 JumpIsExpensive = isExpensive; 1055 } 1056 1057 /// Tells the code generator that integer divide is expensive, and if 1058 /// possible, should be replaced by an alternate sequence of instructions not 1059 /// containing an integer divide. 1060 void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; } 1061 1062 /// Tells the code generator which bitwidths to bypass. 1063 void addBypassSlowDiv(unsigned int SlowBitWidth, unsigned int FastBitWidth) { 1064 BypassSlowDivWidths[SlowBitWidth] = FastBitWidth; 1065 } 1066 1067 /// Tells the code generator that it shouldn't generate srl/add/sra for a 1068 /// signed divide by power of two, and let the target handle it. 1069 void setPow2DivIsCheap(bool isCheap = true) { Pow2DivIsCheap = isCheap; } 1070 1071 /// Add the specified register class as an available regclass for the 1072 /// specified value type. This indicates the selector can handle values of 1073 /// that class natively. 1074 void addRegisterClass(MVT VT, const TargetRegisterClass *RC) { 1075 assert((unsigned)VT.SimpleTy < array_lengthof(RegClassForVT)); 1076 AvailableRegClasses.push_back(std::make_pair(VT, RC)); 1077 RegClassForVT[VT.SimpleTy] = RC; 1078 } 1079 1080 /// Remove all register classes. 1081 void clearRegisterClasses() { 1082 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE * sizeof(TargetRegisterClass*)); 1083 1084 AvailableRegClasses.clear(); 1085 } 1086 1087 /// \brief Remove all operation actions. 1088 void clearOperationActions() { 1089 } 1090 1091 /// Return the largest legal super-reg register class of the register class 1092 /// for the specified type and its associated "cost". 1093 virtual std::pair<const TargetRegisterClass*, uint8_t> 1094 findRepresentativeClass(MVT VT) const; 1095 1096 /// Once all of the register classes are added, this allows us to compute 1097 /// derived properties we expose. 1098 void computeRegisterProperties(); 1099 1100 /// Indicate that the specified operation does not work with the specified 1101 /// type and indicate what to do about it. 1102 void setOperationAction(unsigned Op, MVT VT, 1103 LegalizeAction Action) { 1104 assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!"); 1105 OpActions[(unsigned)VT.SimpleTy][Op] = (uint8_t)Action; 1106 } 1107 1108 /// Indicate that the specified load with extension does not work with the 1109 /// specified type and indicate what to do about it. 1110 void setLoadExtAction(unsigned ExtType, MVT VT, 1111 LegalizeAction Action) { 1112 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE && 1113 "Table isn't big enough!"); 1114 LoadExtActions[VT.SimpleTy][ExtType] = (uint8_t)Action; 1115 } 1116 1117 /// Indicate that the specified truncating store does not work with the 1118 /// specified type and indicate what to do about it. 1119 void setTruncStoreAction(MVT ValVT, MVT MemVT, 1120 LegalizeAction Action) { 1121 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE && 1122 "Table isn't big enough!"); 1123 TruncStoreActions[ValVT.SimpleTy][MemVT.SimpleTy] = (uint8_t)Action; 1124 } 1125 1126 /// Indicate that the specified indexed load does or does not work with the 1127 /// specified type and indicate what to do abort it. 1128 /// 1129 /// NOTE: All indexed mode loads are initialized to Expand in 1130 /// TargetLowering.cpp 1131 void setIndexedLoadAction(unsigned IdxMode, MVT VT, 1132 LegalizeAction Action) { 1133 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE && 1134 (unsigned)Action < 0xf && "Table isn't big enough!"); 1135 // Load action are kept in the upper half. 1136 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0xf0; 1137 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action) <<4; 1138 } 1139 1140 /// Indicate that the specified indexed store does or does not work with the 1141 /// specified type and indicate what to do about it. 1142 /// 1143 /// NOTE: All indexed mode stores are initialized to Expand in 1144 /// TargetLowering.cpp 1145 void setIndexedStoreAction(unsigned IdxMode, MVT VT, 1146 LegalizeAction Action) { 1147 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE && 1148 (unsigned)Action < 0xf && "Table isn't big enough!"); 1149 // Store action are kept in the lower half. 1150 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0x0f; 1151 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action); 1152 } 1153 1154 /// Indicate that the specified condition code is or isn't supported on the 1155 /// target and indicate what to do about it. 1156 void setCondCodeAction(ISD::CondCode CC, MVT VT, 1157 LegalizeAction Action) { 1158 assert(VT < MVT::LAST_VALUETYPE && 1159 (unsigned)CC < array_lengthof(CondCodeActions) && 1160 "Table isn't big enough!"); 1161 /// The lower 5 bits of the SimpleTy index into Nth 2bit set from the 32-bit 1162 /// value and the upper 27 bits index into the second dimension of the array 1163 /// to select what 32-bit value to use. 1164 uint32_t Shift = 2 * (VT.SimpleTy & 0xF); 1165 CondCodeActions[CC][VT.SimpleTy >> 4] &= ~((uint32_t)0x3 << Shift); 1166 CondCodeActions[CC][VT.SimpleTy >> 4] |= (uint32_t)Action << Shift; 1167 } 1168 1169 /// If Opc/OrigVT is specified as being promoted, the promotion code defaults 1170 /// to trying a larger integer/fp until it can find one that works. If that 1171 /// default is insufficient, this method can be used by the target to override 1172 /// the default. 1173 void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) { 1174 PromoteToType[std::make_pair(Opc, OrigVT.SimpleTy)] = DestVT.SimpleTy; 1175 } 1176 1177 /// Targets should invoke this method for each target independent node that 1178 /// they want to provide a custom DAG combiner for by implementing the 1179 /// PerformDAGCombine virtual method. 1180 void setTargetDAGCombine(ISD::NodeType NT) { 1181 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 1182 TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7); 1183 } 1184 1185 /// Set the target's required jmp_buf buffer size (in bytes); default is 200 1186 void setJumpBufSize(unsigned Size) { 1187 JumpBufSize = Size; 1188 } 1189 1190 /// Set the target's required jmp_buf buffer alignment (in bytes); default is 1191 /// 0 1192 void setJumpBufAlignment(unsigned Align) { 1193 JumpBufAlignment = Align; 1194 } 1195 1196 /// Set the target's minimum function alignment (in log2(bytes)) 1197 void setMinFunctionAlignment(unsigned Align) { 1198 MinFunctionAlignment = Align; 1199 } 1200 1201 /// Set the target's preferred function alignment. This should be set if 1202 /// there is a performance benefit to higher-than-minimum alignment (in 1203 /// log2(bytes)) 1204 void setPrefFunctionAlignment(unsigned Align) { 1205 PrefFunctionAlignment = Align; 1206 } 1207 1208 /// Set the target's preferred loop alignment. Default alignment is zero, it 1209 /// means the target does not care about loop alignment. The alignment is 1210 /// specified in log2(bytes). 1211 void setPrefLoopAlignment(unsigned Align) { 1212 PrefLoopAlignment = Align; 1213 } 1214 1215 /// Set the minimum stack alignment of an argument (in log2(bytes)). 1216 void setMinStackArgumentAlignment(unsigned Align) { 1217 MinStackArgumentAlignment = Align; 1218 } 1219 1220 /// Set if the DAG builder should automatically insert fences and reduce the 1221 /// order of atomic memory operations to Monotonic. 1222 void setInsertFencesForAtomic(bool fence) { 1223 InsertFencesForAtomic = fence; 1224 } 1225 1226 public: 1227 //===--------------------------------------------------------------------===// 1228 // Addressing mode description hooks (used by LSR etc). 1229 // 1230 1231 /// CodeGenPrepare sinks address calculations into the same BB as Load/Store 1232 /// instructions reading the address. This allows as much computation as 1233 /// possible to be done in the address mode for that operand. This hook lets 1234 /// targets also pass back when this should be done on intrinsics which 1235 /// load/store. 1236 virtual bool GetAddrModeArguments(IntrinsicInst * /*I*/, 1237 SmallVectorImpl<Value*> &/*Ops*/, 1238 Type *&/*AccessTy*/) const { 1239 return false; 1240 } 1241 1242 /// This represents an addressing mode of: 1243 /// BaseGV + BaseOffs + BaseReg + Scale*ScaleReg 1244 /// If BaseGV is null, there is no BaseGV. 1245 /// If BaseOffs is zero, there is no base offset. 1246 /// If HasBaseReg is false, there is no base register. 1247 /// If Scale is zero, there is no ScaleReg. Scale of 1 indicates a reg with 1248 /// no scale. 1249 struct AddrMode { 1250 GlobalValue *BaseGV; 1251 int64_t BaseOffs; 1252 bool HasBaseReg; 1253 int64_t Scale; 1254 AddrMode() : BaseGV(nullptr), BaseOffs(0), HasBaseReg(false), Scale(0) {} 1255 }; 1256 1257 /// Return true if the addressing mode represented by AM is legal for this 1258 /// target, for a load/store of the specified type. 1259 /// 1260 /// The type may be VoidTy, in which case only return true if the addressing 1261 /// mode is legal for a load/store of any legal type. TODO: Handle 1262 /// pre/postinc as well. 1263 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const; 1264 1265 /// \brief Return the cost of the scaling factor used in the addressing mode 1266 /// represented by AM for this target, for a load/store of the specified type. 1267 /// 1268 /// If the AM is supported, the return value must be >= 0. 1269 /// If the AM is not supported, it returns a negative value. 1270 /// TODO: Handle pre/postinc as well. 1271 virtual int getScalingFactorCost(const AddrMode &AM, Type *Ty) const { 1272 // Default: assume that any scaling factor used in a legal AM is free. 1273 if (isLegalAddressingMode(AM, Ty)) return 0; 1274 return -1; 1275 } 1276 1277 /// Return true if the specified immediate is legal icmp immediate, that is 1278 /// the target has icmp instructions which can compare a register against the 1279 /// immediate without having to materialize the immediate into a register. 1280 virtual bool isLegalICmpImmediate(int64_t) const { 1281 return true; 1282 } 1283 1284 /// Return true if the specified immediate is legal add immediate, that is the 1285 /// target has add instructions which can add a register with the immediate 1286 /// without having to materialize the immediate into a register. 1287 virtual bool isLegalAddImmediate(int64_t) const { 1288 return true; 1289 } 1290 1291 /// Return true if it's significantly cheaper to shift a vector by a uniform 1292 /// scalar than by an amount which will vary across each lane. On x86, for 1293 /// example, there is a "psllw" instruction for the former case, but no simple 1294 /// instruction for a general "a << b" operation on vectors. 1295 virtual bool isVectorShiftByScalarCheap(Type *Ty) const { 1296 return false; 1297 } 1298 1299 /// Return true if it's free to truncate a value of type Ty1 to type 1300 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16 1301 /// by referencing its sub-register AX. 1302 virtual bool isTruncateFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 1303 return false; 1304 } 1305 1306 /// Return true if a truncation from Ty1 to Ty2 is permitted when deciding 1307 /// whether a call is in tail position. Typically this means that both results 1308 /// would be assigned to the same register or stack slot, but it could mean 1309 /// the target performs adequate checks of its own before proceeding with the 1310 /// tail call. 1311 virtual bool allowTruncateForTailCall(Type * /*Ty1*/, Type * /*Ty2*/) const { 1312 return false; 1313 } 1314 1315 virtual bool isTruncateFree(EVT /*VT1*/, EVT /*VT2*/) const { 1316 return false; 1317 } 1318 1319 /// Return true if any actual instruction that defines a value of type Ty1 1320 /// implicitly zero-extends the value to Ty2 in the result register. 1321 /// 1322 /// This does not necessarily include registers defined in unknown ways, such 1323 /// as incoming arguments, or copies from unknown virtual registers. Also, if 1324 /// isTruncateFree(Ty2, Ty1) is true, this does not necessarily apply to 1325 /// truncate instructions. e.g. on x86-64, all instructions that define 32-bit 1326 /// values implicit zero-extend the result out to 64 bits. 1327 virtual bool isZExtFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 1328 return false; 1329 } 1330 1331 virtual bool isZExtFree(EVT /*VT1*/, EVT /*VT2*/) const { 1332 return false; 1333 } 1334 1335 /// Return true if the target supplies and combines to a paired load 1336 /// two loaded values of type LoadedType next to each other in memory. 1337 /// RequiredAlignment gives the minimal alignment constraints that must be met 1338 /// to be able to select this paired load. 1339 /// 1340 /// This information is *not* used to generate actual paired loads, but it is 1341 /// used to generate a sequence of loads that is easier to combine into a 1342 /// paired load. 1343 /// For instance, something like this: 1344 /// a = load i64* addr 1345 /// b = trunc i64 a to i32 1346 /// c = lshr i64 a, 32 1347 /// d = trunc i64 c to i32 1348 /// will be optimized into: 1349 /// b = load i32* addr1 1350 /// d = load i32* addr2 1351 /// Where addr1 = addr2 +/- sizeof(i32). 1352 /// 1353 /// In other words, unless the target performs a post-isel load combining, 1354 /// this information should not be provided because it will generate more 1355 /// loads. 1356 virtual bool hasPairedLoad(Type * /*LoadedType*/, 1357 unsigned & /*RequiredAligment*/) const { 1358 return false; 1359 } 1360 1361 virtual bool hasPairedLoad(EVT /*LoadedType*/, 1362 unsigned & /*RequiredAligment*/) const { 1363 return false; 1364 } 1365 1366 /// Return true if zero-extending the specific node Val to type VT2 is free 1367 /// (either because it's implicitly zero-extended such as ARM ldrb / ldrh or 1368 /// because it's folded such as X86 zero-extending loads). 1369 virtual bool isZExtFree(SDValue Val, EVT VT2) const { 1370 return isZExtFree(Val.getValueType(), VT2); 1371 } 1372 1373 /// Return true if an fneg operation is free to the point where it is never 1374 /// worthwhile to replace it with a bitwise operation. 1375 virtual bool isFNegFree(EVT VT) const { 1376 assert(VT.isFloatingPoint()); 1377 return false; 1378 } 1379 1380 /// Return true if an fabs operation is free to the point where it is never 1381 /// worthwhile to replace it with a bitwise operation. 1382 virtual bool isFAbsFree(EVT VT) const { 1383 assert(VT.isFloatingPoint()); 1384 return false; 1385 } 1386 1387 /// Return true if an FMA operation is faster than a pair of fmul and fadd 1388 /// instructions. fmuladd intrinsics will be expanded to FMAs when this method 1389 /// returns true, otherwise fmuladd is expanded to fmul + fadd. 1390 /// 1391 /// NOTE: This may be called before legalization on types for which FMAs are 1392 /// not legal, but should return true if those types will eventually legalize 1393 /// to types that support FMAs. After legalization, it will only be called on 1394 /// types that support FMAs (via Legal or Custom actions) 1395 virtual bool isFMAFasterThanFMulAndFAdd(EVT) const { 1396 return false; 1397 } 1398 1399 /// Return true if it's profitable to narrow operations of type VT1 to 1400 /// VT2. e.g. on x86, it's profitable to narrow from i32 to i8 but not from 1401 /// i32 to i16. 1402 virtual bool isNarrowingProfitable(EVT /*VT1*/, EVT /*VT2*/) const { 1403 return false; 1404 } 1405 1406 /// \brief Return true if it is beneficial to convert a load of a constant to 1407 /// just the constant itself. 1408 /// On some targets it might be more efficient to use a combination of 1409 /// arithmetic instructions to materialize the constant instead of loading it 1410 /// from a constant pool. 1411 virtual bool shouldConvertConstantLoadToIntImm(const APInt &Imm, 1412 Type *Ty) const { 1413 return false; 1414 } 1415 //===--------------------------------------------------------------------===// 1416 // Runtime Library hooks 1417 // 1418 1419 /// Rename the default libcall routine name for the specified libcall. 1420 void setLibcallName(RTLIB::Libcall Call, const char *Name) { 1421 LibcallRoutineNames[Call] = Name; 1422 } 1423 1424 /// Get the libcall routine name for the specified libcall. 1425 const char *getLibcallName(RTLIB::Libcall Call) const { 1426 return LibcallRoutineNames[Call]; 1427 } 1428 1429 /// Override the default CondCode to be used to test the result of the 1430 /// comparison libcall against zero. 1431 void setCmpLibcallCC(RTLIB::Libcall Call, ISD::CondCode CC) { 1432 CmpLibcallCCs[Call] = CC; 1433 } 1434 1435 /// Get the CondCode that's to be used to test the result of the comparison 1436 /// libcall against zero. 1437 ISD::CondCode getCmpLibcallCC(RTLIB::Libcall Call) const { 1438 return CmpLibcallCCs[Call]; 1439 } 1440 1441 /// Set the CallingConv that should be used for the specified libcall. 1442 void setLibcallCallingConv(RTLIB::Libcall Call, CallingConv::ID CC) { 1443 LibcallCallingConvs[Call] = CC; 1444 } 1445 1446 /// Get the CallingConv that should be used for the specified libcall. 1447 CallingConv::ID getLibcallCallingConv(RTLIB::Libcall Call) const { 1448 return LibcallCallingConvs[Call]; 1449 } 1450 1451 private: 1452 const TargetMachine &TM; 1453 const DataLayout *DL; 1454 const TargetLoweringObjectFile &TLOF; 1455 1456 /// True if this is a little endian target. 1457 bool IsLittleEndian; 1458 1459 /// Tells the code generator not to expand operations into sequences that use 1460 /// the select operations if possible. 1461 bool SelectIsExpensive; 1462 1463 /// Tells the code generator that the target has multiple (allocatable) 1464 /// condition registers that can be used to store the results of comparisons 1465 /// for use by selects and conditional branches. With multiple condition 1466 /// registers, the code generator will not aggressively sink comparisons into 1467 /// the blocks of their users. 1468 bool HasMultipleConditionRegisters; 1469 1470 /// Tells the code generator that the target has BitExtract instructions. 1471 /// The code generator will aggressively sink "shift"s into the blocks of 1472 /// their users if the users will generate "and" instructions which can be 1473 /// combined with "shift" to BitExtract instructions. 1474 bool HasExtractBitsInsn; 1475 1476 /// Tells the code generator not to expand integer divides by constants into a 1477 /// sequence of muls, adds, and shifts. This is a hack until a real cost 1478 /// model is in place. If we ever optimize for size, this will be set to true 1479 /// unconditionally. 1480 bool IntDivIsCheap; 1481 1482 /// Tells the code generator to bypass slow divide or remainder 1483 /// instructions. For example, BypassSlowDivWidths[32,8] tells the code 1484 /// generator to bypass 32-bit integer div/rem with an 8-bit unsigned integer 1485 /// div/rem when the operands are positive and less than 256. 1486 DenseMap <unsigned int, unsigned int> BypassSlowDivWidths; 1487 1488 /// Tells the code generator that it shouldn't generate srl/add/sra for a 1489 /// signed divide by power of two, and let the target handle it. 1490 bool Pow2DivIsCheap; 1491 1492 /// Tells the code generator that it shouldn't generate extra flow control 1493 /// instructions and should attempt to combine flow control instructions via 1494 /// predication. 1495 bool JumpIsExpensive; 1496 1497 /// This target prefers to use _setjmp to implement llvm.setjmp. 1498 /// 1499 /// Defaults to false. 1500 bool UseUnderscoreSetJmp; 1501 1502 /// This target prefers to use _longjmp to implement llvm.longjmp. 1503 /// 1504 /// Defaults to false. 1505 bool UseUnderscoreLongJmp; 1506 1507 /// Whether the target can generate code for jumptables. If it's not true, 1508 /// then each jumptable must be lowered into if-then-else's. 1509 bool SupportJumpTables; 1510 1511 /// Number of blocks threshold to use jump tables. 1512 int MinimumJumpTableEntries; 1513 1514 /// Information about the contents of the high-bits in boolean values held in 1515 /// a type wider than i1. See getBooleanContents. 1516 BooleanContent BooleanContents; 1517 1518 /// Information about the contents of the high-bits in boolean values held in 1519 /// a type wider than i1. See getBooleanContents. 1520 BooleanContent BooleanFloatContents; 1521 1522 /// Information about the contents of the high-bits in boolean vector values 1523 /// when the element type is wider than i1. See getBooleanContents. 1524 BooleanContent BooleanVectorContents; 1525 1526 /// The target scheduling preference: shortest possible total cycles or lowest 1527 /// register usage. 1528 Sched::Preference SchedPreferenceInfo; 1529 1530 /// The size, in bytes, of the target's jmp_buf buffers 1531 unsigned JumpBufSize; 1532 1533 /// The alignment, in bytes, of the target's jmp_buf buffers 1534 unsigned JumpBufAlignment; 1535 1536 /// The minimum alignment that any argument on the stack needs to have. 1537 unsigned MinStackArgumentAlignment; 1538 1539 /// The minimum function alignment (used when optimizing for size, and to 1540 /// prevent explicitly provided alignment from leading to incorrect code). 1541 unsigned MinFunctionAlignment; 1542 1543 /// The preferred function alignment (used when alignment unspecified and 1544 /// optimizing for speed). 1545 unsigned PrefFunctionAlignment; 1546 1547 /// The preferred loop alignment. 1548 unsigned PrefLoopAlignment; 1549 1550 /// Whether the DAG builder should automatically insert fences and reduce 1551 /// ordering for atomics. (This will be set for for most architectures with 1552 /// weak memory ordering.) 1553 bool InsertFencesForAtomic; 1554 1555 /// If set to a physical register, this specifies the register that 1556 /// llvm.savestack/llvm.restorestack should save and restore. 1557 unsigned StackPointerRegisterToSaveRestore; 1558 1559 /// If set to a physical register, this specifies the register that receives 1560 /// the exception address on entry to a landing pad. 1561 unsigned ExceptionPointerRegister; 1562 1563 /// If set to a physical register, this specifies the register that receives 1564 /// the exception typeid on entry to a landing pad. 1565 unsigned ExceptionSelectorRegister; 1566 1567 /// This indicates the default register class to use for each ValueType the 1568 /// target supports natively. 1569 const TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE]; 1570 unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE]; 1571 MVT RegisterTypeForVT[MVT::LAST_VALUETYPE]; 1572 1573 /// This indicates the "representative" register class to use for each 1574 /// ValueType the target supports natively. This information is used by the 1575 /// scheduler to track register pressure. By default, the representative 1576 /// register class is the largest legal super-reg register class of the 1577 /// register class of the specified type. e.g. On x86, i8, i16, and i32's 1578 /// representative class would be GR32. 1579 const TargetRegisterClass *RepRegClassForVT[MVT::LAST_VALUETYPE]; 1580 1581 /// This indicates the "cost" of the "representative" register class for each 1582 /// ValueType. The cost is used by the scheduler to approximate register 1583 /// pressure. 1584 uint8_t RepRegClassCostForVT[MVT::LAST_VALUETYPE]; 1585 1586 /// For any value types we are promoting or expanding, this contains the value 1587 /// type that we are changing to. For Expanded types, this contains one step 1588 /// of the expand (e.g. i64 -> i32), even if there are multiple steps required 1589 /// (e.g. i64 -> i16). For types natively supported by the system, this holds 1590 /// the same type (e.g. i32 -> i32). 1591 MVT TransformToType[MVT::LAST_VALUETYPE]; 1592 1593 /// For each operation and each value type, keep a LegalizeAction that 1594 /// indicates how instruction selection should deal with the operation. Most 1595 /// operations are Legal (aka, supported natively by the target), but 1596 /// operations that are not should be described. Note that operations on 1597 /// non-legal value types are not described here. 1598 uint8_t OpActions[MVT::LAST_VALUETYPE][ISD::BUILTIN_OP_END]; 1599 1600 /// For each load extension type and each value type, keep a LegalizeAction 1601 /// that indicates how instruction selection should deal with a load of a 1602 /// specific value type and extension type. 1603 uint8_t LoadExtActions[MVT::LAST_VALUETYPE][ISD::LAST_LOADEXT_TYPE]; 1604 1605 /// For each value type pair keep a LegalizeAction that indicates whether a 1606 /// truncating store of a specific value type and truncating type is legal. 1607 uint8_t TruncStoreActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE]; 1608 1609 /// For each indexed mode and each value type, keep a pair of LegalizeAction 1610 /// that indicates how instruction selection should deal with the load / 1611 /// store. 1612 /// 1613 /// The first dimension is the value_type for the reference. The second 1614 /// dimension represents the various modes for load store. 1615 uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][ISD::LAST_INDEXED_MODE]; 1616 1617 /// For each condition code (ISD::CondCode) keep a LegalizeAction that 1618 /// indicates how instruction selection should deal with the condition code. 1619 /// 1620 /// Because each CC action takes up 2 bits, we need to have the array size be 1621 /// large enough to fit all of the value types. This can be done by rounding 1622 /// up the MVT::LAST_VALUETYPE value to the next multiple of 16. 1623 uint32_t CondCodeActions[ISD::SETCC_INVALID][(MVT::LAST_VALUETYPE + 15) / 16]; 1624 1625 ValueTypeActionImpl ValueTypeActions; 1626 1627 public: 1628 LegalizeKind 1629 getTypeConversion(LLVMContext &Context, EVT VT) const { 1630 // If this is a simple type, use the ComputeRegisterProp mechanism. 1631 if (VT.isSimple()) { 1632 MVT SVT = VT.getSimpleVT(); 1633 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 1634 MVT NVT = TransformToType[SVT.SimpleTy]; 1635 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 1636 1637 assert( 1638 (LA == TypeLegal || 1639 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) 1640 && "Promote may not follow Expand or Promote"); 1641 1642 if (LA == TypeSplitVector) 1643 return LegalizeKind(LA, EVT::getVectorVT(Context, 1644 SVT.getVectorElementType(), 1645 SVT.getVectorNumElements()/2)); 1646 if (LA == TypeScalarizeVector) 1647 return LegalizeKind(LA, SVT.getVectorElementType()); 1648 return LegalizeKind(LA, NVT); 1649 } 1650 1651 // Handle Extended Scalar Types. 1652 if (!VT.isVector()) { 1653 assert(VT.isInteger() && "Float types must be simple"); 1654 unsigned BitSize = VT.getSizeInBits(); 1655 // First promote to a power-of-two size, then expand if necessary. 1656 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 1657 EVT NVT = VT.getRoundIntegerType(Context); 1658 assert(NVT != VT && "Unable to round integer VT"); 1659 LegalizeKind NextStep = getTypeConversion(Context, NVT); 1660 // Avoid multi-step promotion. 1661 if (NextStep.first == TypePromoteInteger) return NextStep; 1662 // Return rounded integer type. 1663 return LegalizeKind(TypePromoteInteger, NVT); 1664 } 1665 1666 return LegalizeKind(TypeExpandInteger, 1667 EVT::getIntegerVT(Context, VT.getSizeInBits()/2)); 1668 } 1669 1670 // Handle vector types. 1671 unsigned NumElts = VT.getVectorNumElements(); 1672 EVT EltVT = VT.getVectorElementType(); 1673 1674 // Vectors with only one element are always scalarized. 1675 if (NumElts == 1) 1676 return LegalizeKind(TypeScalarizeVector, EltVT); 1677 1678 // Try to widen vector elements until the element type is a power of two and 1679 // promote it to a legal type later on, for example: 1680 // <3 x i8> -> <4 x i8> -> <4 x i32> 1681 if (EltVT.isInteger()) { 1682 // Vectors with a number of elements that is not a power of two are always 1683 // widened, for example <3 x i8> -> <4 x i8>. 1684 if (!VT.isPow2VectorType()) { 1685 NumElts = (unsigned)NextPowerOf2(NumElts); 1686 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 1687 return LegalizeKind(TypeWidenVector, NVT); 1688 } 1689 1690 // Examine the element type. 1691 LegalizeKind LK = getTypeConversion(Context, EltVT); 1692 1693 // If type is to be expanded, split the vector. 1694 // <4 x i140> -> <2 x i140> 1695 if (LK.first == TypeExpandInteger) 1696 return LegalizeKind(TypeSplitVector, 1697 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 1698 1699 // Promote the integer element types until a legal vector type is found 1700 // or until the element integer type is too big. If a legal type was not 1701 // found, fallback to the usual mechanism of widening/splitting the 1702 // vector. 1703 EVT OldEltVT = EltVT; 1704 while (1) { 1705 // Increase the bitwidth of the element to the next pow-of-two 1706 // (which is greater than 8 bits). 1707 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits() 1708 ).getRoundIntegerType(Context); 1709 1710 // Stop trying when getting a non-simple element type. 1711 // Note that vector elements may be greater than legal vector element 1712 // types. Example: X86 XMM registers hold 64bit element on 32bit 1713 // systems. 1714 if (!EltVT.isSimple()) break; 1715 1716 // Build a new vector type and check if it is legal. 1717 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 1718 // Found a legal promoted vector type. 1719 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 1720 return LegalizeKind(TypePromoteInteger, 1721 EVT::getVectorVT(Context, EltVT, NumElts)); 1722 } 1723 1724 // Reset the type to the unexpanded type if we did not find a legal vector 1725 // type with a promoted vector element type. 1726 EltVT = OldEltVT; 1727 } 1728 1729 // Try to widen the vector until a legal type is found. 1730 // If there is no wider legal type, split the vector. 1731 while (1) { 1732 // Round up to the next power of 2. 1733 NumElts = (unsigned)NextPowerOf2(NumElts); 1734 1735 // If there is no simple vector type with this many elements then there 1736 // cannot be a larger legal vector type. Note that this assumes that 1737 // there are no skipped intermediate vector types in the simple types. 1738 if (!EltVT.isSimple()) break; 1739 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 1740 if (LargerVector == MVT()) break; 1741 1742 // If this type is legal then widen the vector. 1743 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 1744 return LegalizeKind(TypeWidenVector, LargerVector); 1745 } 1746 1747 // Widen odd vectors to next power of two. 1748 if (!VT.isPow2VectorType()) { 1749 EVT NVT = VT.getPow2VectorType(Context); 1750 return LegalizeKind(TypeWidenVector, NVT); 1751 } 1752 1753 // Vectors with illegal element types are expanded. 1754 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 1755 return LegalizeKind(TypeSplitVector, NVT); 1756 } 1757 1758 private: 1759 std::vector<std::pair<MVT, const TargetRegisterClass*> > AvailableRegClasses; 1760 1761 /// Targets can specify ISD nodes that they would like PerformDAGCombine 1762 /// callbacks for by calling setTargetDAGCombine(), which sets a bit in this 1763 /// array. 1764 unsigned char 1765 TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT]; 1766 1767 /// For operations that must be promoted to a specific type, this holds the 1768 /// destination type. This map should be sparse, so don't hold it as an 1769 /// array. 1770 /// 1771 /// Targets add entries to this map with AddPromotedToType(..), clients access 1772 /// this with getTypeToPromoteTo(..). 1773 std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType> 1774 PromoteToType; 1775 1776 /// Stores the name each libcall. 1777 const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL]; 1778 1779 /// The ISD::CondCode that should be used to test the result of each of the 1780 /// comparison libcall against zero. 1781 ISD::CondCode CmpLibcallCCs[RTLIB::UNKNOWN_LIBCALL]; 1782 1783 /// Stores the CallingConv that should be used for each libcall. 1784 CallingConv::ID LibcallCallingConvs[RTLIB::UNKNOWN_LIBCALL]; 1785 1786 protected: 1787 /// \brief Specify maximum number of store instructions per memset call. 1788 /// 1789 /// When lowering \@llvm.memset this field specifies the maximum number of 1790 /// store operations that may be substituted for the call to memset. Targets 1791 /// must set this value based on the cost threshold for that target. Targets 1792 /// should assume that the memset will be done using as many of the largest 1793 /// store operations first, followed by smaller ones, if necessary, per 1794 /// alignment restrictions. For example, storing 9 bytes on a 32-bit machine 1795 /// with 16-bit alignment would result in four 2-byte stores and one 1-byte 1796 /// store. This only applies to setting a constant array of a constant size. 1797 unsigned MaxStoresPerMemset; 1798 1799 /// Maximum number of stores operations that may be substituted for the call 1800 /// to memset, used for functions with OptSize attribute. 1801 unsigned MaxStoresPerMemsetOptSize; 1802 1803 /// \brief Specify maximum bytes of store instructions per memcpy call. 1804 /// 1805 /// When lowering \@llvm.memcpy this field specifies the maximum number of 1806 /// store operations that may be substituted for a call to memcpy. Targets 1807 /// must set this value based on the cost threshold for that target. Targets 1808 /// should assume that the memcpy will be done using as many of the largest 1809 /// store operations first, followed by smaller ones, if necessary, per 1810 /// alignment restrictions. For example, storing 7 bytes on a 32-bit machine 1811 /// with 32-bit alignment would result in one 4-byte store, a one 2-byte store 1812 /// and one 1-byte store. This only applies to copying a constant array of 1813 /// constant size. 1814 unsigned MaxStoresPerMemcpy; 1815 1816 /// Maximum number of store operations that may be substituted for a call to 1817 /// memcpy, used for functions with OptSize attribute. 1818 unsigned MaxStoresPerMemcpyOptSize; 1819 1820 /// \brief Specify maximum bytes of store instructions per memmove call. 1821 /// 1822 /// When lowering \@llvm.memmove this field specifies the maximum number of 1823 /// store instructions that may be substituted for a call to memmove. Targets 1824 /// must set this value based on the cost threshold for that target. Targets 1825 /// should assume that the memmove will be done using as many of the largest 1826 /// store operations first, followed by smaller ones, if necessary, per 1827 /// alignment restrictions. For example, moving 9 bytes on a 32-bit machine 1828 /// with 8-bit alignment would result in nine 1-byte stores. This only 1829 /// applies to copying a constant array of constant size. 1830 unsigned MaxStoresPerMemmove; 1831 1832 /// Maximum number of store instructions that may be substituted for a call to 1833 /// memmove, used for functions with OpSize attribute. 1834 unsigned MaxStoresPerMemmoveOptSize; 1835 1836 /// Tells the code generator that select is more expensive than a branch if 1837 /// the branch is usually predicted right. 1838 bool PredictableSelectIsExpensive; 1839 1840 /// MaskAndBranchFoldingIsLegal - Indicates if the target supports folding 1841 /// a mask of a single bit, a compare, and a branch into a single instruction. 1842 bool MaskAndBranchFoldingIsLegal; 1843 1844 protected: 1845 /// Return true if the value types that can be represented by the specified 1846 /// register class are all legal. 1847 bool isLegalRC(const TargetRegisterClass *RC) const; 1848 1849 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 1850 /// sequence of memory operands that is recognized by PrologEpilogInserter. 1851 MachineBasicBlock *emitPatchPoint(MachineInstr *MI, MachineBasicBlock *MBB) const; 1852 }; 1853 1854 /// This class defines information used to lower LLVM code to legal SelectionDAG 1855 /// operators that the target instruction selector can accept natively. 1856 /// 1857 /// This class also defines callbacks that targets must implement to lower 1858 /// target-specific constructs to SelectionDAG operators. 1859 class TargetLowering : public TargetLoweringBase { 1860 TargetLowering(const TargetLowering&) LLVM_DELETED_FUNCTION; 1861 void operator=(const TargetLowering&) LLVM_DELETED_FUNCTION; 1862 1863 public: 1864 /// NOTE: The constructor takes ownership of TLOF. 1865 explicit TargetLowering(const TargetMachine &TM, 1866 const TargetLoweringObjectFile *TLOF); 1867 1868 /// Returns true by value, base pointer and offset pointer and addressing mode 1869 /// by reference if the node's address can be legally represented as 1870 /// pre-indexed load / store address. 1871 virtual bool getPreIndexedAddressParts(SDNode * /*N*/, SDValue &/*Base*/, 1872 SDValue &/*Offset*/, 1873 ISD::MemIndexedMode &/*AM*/, 1874 SelectionDAG &/*DAG*/) const { 1875 return false; 1876 } 1877 1878 /// Returns true by value, base pointer and offset pointer and addressing mode 1879 /// by reference if this node can be combined with a load / store to form a 1880 /// post-indexed load / store. 1881 virtual bool getPostIndexedAddressParts(SDNode * /*N*/, SDNode * /*Op*/, 1882 SDValue &/*Base*/, 1883 SDValue &/*Offset*/, 1884 ISD::MemIndexedMode &/*AM*/, 1885 SelectionDAG &/*DAG*/) const { 1886 return false; 1887 } 1888 1889 /// Return the entry encoding for a jump table in the current function. The 1890 /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum. 1891 virtual unsigned getJumpTableEncoding() const; 1892 1893 virtual const MCExpr * 1894 LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/, 1895 const MachineBasicBlock * /*MBB*/, unsigned /*uid*/, 1896 MCContext &/*Ctx*/) const { 1897 llvm_unreachable("Need to implement this hook if target has custom JTIs"); 1898 } 1899 1900 /// Returns relocation base for the given PIC jumptable. 1901 virtual SDValue getPICJumpTableRelocBase(SDValue Table, 1902 SelectionDAG &DAG) const; 1903 1904 /// This returns the relocation base for the given PIC jumptable, the same as 1905 /// getPICJumpTableRelocBase, but as an MCExpr. 1906 virtual const MCExpr * 1907 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 1908 unsigned JTI, MCContext &Ctx) const; 1909 1910 /// Return true if folding a constant offset with the given GlobalAddress is 1911 /// legal. It is frequently not legal in PIC relocation models. 1912 virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const; 1913 1914 bool isInTailCallPosition(SelectionDAG &DAG, SDNode *Node, 1915 SDValue &Chain) const; 1916 1917 void softenSetCCOperands(SelectionDAG &DAG, EVT VT, 1918 SDValue &NewLHS, SDValue &NewRHS, 1919 ISD::CondCode &CCCode, SDLoc DL) const; 1920 1921 /// Returns a pair of (return value, chain). 1922 std::pair<SDValue, SDValue> makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC, 1923 EVT RetVT, const SDValue *Ops, 1924 unsigned NumOps, bool isSigned, 1925 SDLoc dl, bool doesNotReturn = false, 1926 bool isReturnValueUsed = true) const; 1927 1928 //===--------------------------------------------------------------------===// 1929 // TargetLowering Optimization Methods 1930 // 1931 1932 /// A convenience struct that encapsulates a DAG, and two SDValues for 1933 /// returning information from TargetLowering to its clients that want to 1934 /// combine. 1935 struct TargetLoweringOpt { 1936 SelectionDAG &DAG; 1937 bool LegalTys; 1938 bool LegalOps; 1939 SDValue Old; 1940 SDValue New; 1941 1942 explicit TargetLoweringOpt(SelectionDAG &InDAG, 1943 bool LT, bool LO) : 1944 DAG(InDAG), LegalTys(LT), LegalOps(LO) {} 1945 1946 bool LegalTypes() const { return LegalTys; } 1947 bool LegalOperations() const { return LegalOps; } 1948 1949 bool CombineTo(SDValue O, SDValue N) { 1950 Old = O; 1951 New = N; 1952 return true; 1953 } 1954 1955 /// Check to see if the specified operand of the specified instruction is a 1956 /// constant integer. If so, check to see if there are any bits set in the 1957 /// constant that are not demanded. If so, shrink the constant and return 1958 /// true. 1959 bool ShrinkDemandedConstant(SDValue Op, const APInt &Demanded); 1960 1961 /// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free. This 1962 /// uses isZExtFree and ZERO_EXTEND for the widening cast, but it could be 1963 /// generalized for targets with other types of implicit widening casts. 1964 bool ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &Demanded, 1965 SDLoc dl); 1966 }; 1967 1968 /// Look at Op. At this point, we know that only the DemandedMask bits of the 1969 /// result of Op are ever used downstream. If we can use this information to 1970 /// simplify Op, create a new simplified DAG node and return true, returning 1971 /// the original and new nodes in Old and New. Otherwise, analyze the 1972 /// expression and return a mask of KnownOne and KnownZero bits for the 1973 /// expression (used to simplify the caller). The KnownZero/One bits may only 1974 /// be accurate for those bits in the DemandedMask. 1975 bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask, 1976 APInt &KnownZero, APInt &KnownOne, 1977 TargetLoweringOpt &TLO, unsigned Depth = 0) const; 1978 1979 /// Determine which of the bits specified in Mask are known to be either zero 1980 /// or one and return them in the KnownZero/KnownOne bitsets. 1981 virtual void computeKnownBitsForTargetNode(const SDValue Op, 1982 APInt &KnownZero, 1983 APInt &KnownOne, 1984 const SelectionDAG &DAG, 1985 unsigned Depth = 0) const; 1986 1987 /// This method can be implemented by targets that want to expose additional 1988 /// information about sign bits to the DAG Combiner. 1989 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op, 1990 const SelectionDAG &DAG, 1991 unsigned Depth = 0) const; 1992 1993 struct DAGCombinerInfo { 1994 void *DC; // The DAG Combiner object. 1995 CombineLevel Level; 1996 bool CalledByLegalizer; 1997 public: 1998 SelectionDAG &DAG; 1999 2000 DAGCombinerInfo(SelectionDAG &dag, CombineLevel level, bool cl, void *dc) 2001 : DC(dc), Level(level), CalledByLegalizer(cl), DAG(dag) {} 2002 2003 bool isBeforeLegalize() const { return Level == BeforeLegalizeTypes; } 2004 bool isBeforeLegalizeOps() const { return Level < AfterLegalizeVectorOps; } 2005 bool isAfterLegalizeVectorOps() const { 2006 return Level == AfterLegalizeDAG; 2007 } 2008 CombineLevel getDAGCombineLevel() { return Level; } 2009 bool isCalledByLegalizer() const { return CalledByLegalizer; } 2010 2011 void AddToWorklist(SDNode *N); 2012 void RemoveFromWorklist(SDNode *N); 2013 SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To, 2014 bool AddTo = true); 2015 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true); 2016 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true); 2017 2018 void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO); 2019 }; 2020 2021 /// Return if the N is a constant or constant vector equal to the true value 2022 /// from getBooleanContents(). 2023 bool isConstTrueVal(const SDNode *N) const; 2024 2025 /// Return if the N is a constant or constant vector equal to the false value 2026 /// from getBooleanContents(). 2027 bool isConstFalseVal(const SDNode *N) const; 2028 2029 /// Try to simplify a setcc built with the specified operands and cc. If it is 2030 /// unable to simplify it, return a null SDValue. 2031 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 2032 ISD::CondCode Cond, bool foldBooleans, 2033 DAGCombinerInfo &DCI, SDLoc dl) const; 2034 2035 /// Returns true (and the GlobalValue and the offset) if the node is a 2036 /// GlobalAddress + offset. 2037 virtual bool 2038 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const; 2039 2040 /// This method will be invoked for all target nodes and for any 2041 /// target-independent nodes that the target has registered with invoke it 2042 /// for. 2043 /// 2044 /// The semantics are as follows: 2045 /// Return Value: 2046 /// SDValue.Val == 0 - No change was made 2047 /// SDValue.Val == N - N was replaced, is dead, and is already handled. 2048 /// otherwise - N should be replaced by the returned Operand. 2049 /// 2050 /// In addition, methods provided by DAGCombinerInfo may be used to perform 2051 /// more complex transformations. 2052 /// 2053 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const; 2054 2055 /// Return true if it is profitable to move a following shift through this 2056 // node, adjusting any immediate operands as necessary to preserve semantics. 2057 // This transformation may not be desirable if it disrupts a particularly 2058 // auspicious target-specific tree (e.g. bitfield extraction in AArch64). 2059 // By default, it returns true. 2060 virtual bool isDesirableToCommuteWithShift(const SDNode *N /*Op*/) const { 2061 return true; 2062 } 2063 2064 /// Return true if the target has native support for the specified value type 2065 /// and it is 'desirable' to use the type for the given node type. e.g. On x86 2066 /// i16 is legal, but undesirable since i16 instruction encodings are longer 2067 /// and some i16 instructions are slow. 2068 virtual bool isTypeDesirableForOp(unsigned /*Opc*/, EVT VT) const { 2069 // By default, assume all legal types are desirable. 2070 return isTypeLegal(VT); 2071 } 2072 2073 /// Return true if it is profitable for dag combiner to transform a floating 2074 /// point op of specified opcode to a equivalent op of an integer 2075 /// type. e.g. f32 load -> i32 load can be profitable on ARM. 2076 virtual bool isDesirableToTransformToIntegerOp(unsigned /*Opc*/, 2077 EVT /*VT*/) const { 2078 return false; 2079 } 2080 2081 /// This method query the target whether it is beneficial for dag combiner to 2082 /// promote the specified node. If true, it should return the desired 2083 /// promotion type by reference. 2084 virtual bool IsDesirableToPromoteOp(SDValue /*Op*/, EVT &/*PVT*/) const { 2085 return false; 2086 } 2087 2088 //===--------------------------------------------------------------------===// 2089 // Lowering methods - These methods must be implemented by targets so that 2090 // the SelectionDAGBuilder code knows how to lower these. 2091 // 2092 2093 /// This hook must be implemented to lower the incoming (formal) arguments, 2094 /// described by the Ins array, into the specified DAG. The implementation 2095 /// should fill in the InVals array with legal-type argument values, and 2096 /// return the resulting token chain value. 2097 /// 2098 virtual SDValue 2099 LowerFormalArguments(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 2100 bool /*isVarArg*/, 2101 const SmallVectorImpl<ISD::InputArg> &/*Ins*/, 2102 SDLoc /*dl*/, SelectionDAG &/*DAG*/, 2103 SmallVectorImpl<SDValue> &/*InVals*/) const { 2104 llvm_unreachable("Not Implemented"); 2105 } 2106 2107 struct ArgListEntry { 2108 SDValue Node; 2109 Type* Ty; 2110 bool isSExt : 1; 2111 bool isZExt : 1; 2112 bool isInReg : 1; 2113 bool isSRet : 1; 2114 bool isNest : 1; 2115 bool isByVal : 1; 2116 bool isInAlloca : 1; 2117 bool isReturned : 1; 2118 uint16_t Alignment; 2119 2120 ArgListEntry() : isSExt(false), isZExt(false), isInReg(false), 2121 isSRet(false), isNest(false), isByVal(false), isInAlloca(false), 2122 isReturned(false), Alignment(0) { } 2123 2124 void setAttributes(ImmutableCallSite *CS, unsigned AttrIdx); 2125 }; 2126 typedef std::vector<ArgListEntry> ArgListTy; 2127 2128 /// This structure contains all information that is necessary for lowering 2129 /// calls. It is passed to TLI::LowerCallTo when the SelectionDAG builder 2130 /// needs to lower a call, and targets will see this struct in their LowerCall 2131 /// implementation. 2132 struct CallLoweringInfo { 2133 SDValue Chain; 2134 Type *RetTy; 2135 bool RetSExt : 1; 2136 bool RetZExt : 1; 2137 bool IsVarArg : 1; 2138 bool IsInReg : 1; 2139 bool DoesNotReturn : 1; 2140 bool IsReturnValueUsed : 1; 2141 2142 // IsTailCall should be modified by implementations of 2143 // TargetLowering::LowerCall that perform tail call conversions. 2144 bool IsTailCall; 2145 2146 unsigned NumFixedArgs; 2147 CallingConv::ID CallConv; 2148 SDValue Callee; 2149 ArgListTy Args; 2150 SelectionDAG &DAG; 2151 SDLoc DL; 2152 ImmutableCallSite *CS; 2153 SmallVector<ISD::OutputArg, 32> Outs; 2154 SmallVector<SDValue, 32> OutVals; 2155 SmallVector<ISD::InputArg, 32> Ins; 2156 2157 CallLoweringInfo(SelectionDAG &DAG) 2158 : RetTy(nullptr), RetSExt(false), RetZExt(false), IsVarArg(false), 2159 IsInReg(false), DoesNotReturn(false), IsReturnValueUsed(true), 2160 IsTailCall(false), NumFixedArgs(-1), CallConv(CallingConv::C), 2161 DAG(DAG), CS(nullptr) {} 2162 2163 CallLoweringInfo &setDebugLoc(SDLoc dl) { 2164 DL = dl; 2165 return *this; 2166 } 2167 2168 CallLoweringInfo &setChain(SDValue InChain) { 2169 Chain = InChain; 2170 return *this; 2171 } 2172 2173 CallLoweringInfo &setCallee(CallingConv::ID CC, Type *ResultType, 2174 SDValue Target, ArgListTy &&ArgsList, 2175 unsigned FixedArgs = -1) { 2176 RetTy = ResultType; 2177 Callee = Target; 2178 CallConv = CC; 2179 NumFixedArgs = 2180 (FixedArgs == static_cast<unsigned>(-1) ? Args.size() : FixedArgs); 2181 Args = std::move(ArgsList); 2182 return *this; 2183 } 2184 2185 CallLoweringInfo &setCallee(Type *ResultType, FunctionType *FTy, 2186 SDValue Target, ArgListTy &&ArgsList, 2187 ImmutableCallSite &Call) { 2188 RetTy = ResultType; 2189 2190 IsInReg = Call.paramHasAttr(0, Attribute::InReg); 2191 DoesNotReturn = Call.doesNotReturn(); 2192 IsVarArg = FTy->isVarArg(); 2193 IsReturnValueUsed = !Call.getInstruction()->use_empty(); 2194 RetSExt = Call.paramHasAttr(0, Attribute::SExt); 2195 RetZExt = Call.paramHasAttr(0, Attribute::ZExt); 2196 2197 Callee = Target; 2198 2199 CallConv = Call.getCallingConv(); 2200 NumFixedArgs = FTy->getNumParams(); 2201 Args = std::move(ArgsList); 2202 2203 CS = &Call; 2204 2205 return *this; 2206 } 2207 2208 CallLoweringInfo &setInRegister(bool Value = true) { 2209 IsInReg = Value; 2210 return *this; 2211 } 2212 2213 CallLoweringInfo &setNoReturn(bool Value = true) { 2214 DoesNotReturn = Value; 2215 return *this; 2216 } 2217 2218 CallLoweringInfo &setVarArg(bool Value = true) { 2219 IsVarArg = Value; 2220 return *this; 2221 } 2222 2223 CallLoweringInfo &setTailCall(bool Value = true) { 2224 IsTailCall = Value; 2225 return *this; 2226 } 2227 2228 CallLoweringInfo &setDiscardResult(bool Value = true) { 2229 IsReturnValueUsed = !Value; 2230 return *this; 2231 } 2232 2233 CallLoweringInfo &setSExtResult(bool Value = true) { 2234 RetSExt = Value; 2235 return *this; 2236 } 2237 2238 CallLoweringInfo &setZExtResult(bool Value = true) { 2239 RetZExt = Value; 2240 return *this; 2241 } 2242 2243 ArgListTy &getArgs() { 2244 return Args; 2245 } 2246 }; 2247 2248 /// This function lowers an abstract call to a function into an actual call. 2249 /// This returns a pair of operands. The first element is the return value 2250 /// for the function (if RetTy is not VoidTy). The second element is the 2251 /// outgoing token chain. It calls LowerCall to do the actual lowering. 2252 std::pair<SDValue, SDValue> LowerCallTo(CallLoweringInfo &CLI) const; 2253 2254 /// This hook must be implemented to lower calls into the the specified 2255 /// DAG. The outgoing arguments to the call are described by the Outs array, 2256 /// and the values to be returned by the call are described by the Ins 2257 /// array. The implementation should fill in the InVals array with legal-type 2258 /// return values from the call, and return the resulting token chain value. 2259 virtual SDValue 2260 LowerCall(CallLoweringInfo &/*CLI*/, 2261 SmallVectorImpl<SDValue> &/*InVals*/) const { 2262 llvm_unreachable("Not Implemented"); 2263 } 2264 2265 /// Target-specific cleanup for formal ByVal parameters. 2266 virtual void HandleByVal(CCState *, unsigned &, unsigned) const {} 2267 2268 /// This hook should be implemented to check whether the return values 2269 /// described by the Outs array can fit into the return registers. If false 2270 /// is returned, an sret-demotion is performed. 2271 virtual bool CanLowerReturn(CallingConv::ID /*CallConv*/, 2272 MachineFunction &/*MF*/, bool /*isVarArg*/, 2273 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 2274 LLVMContext &/*Context*/) const 2275 { 2276 // Return true by default to get preexisting behavior. 2277 return true; 2278 } 2279 2280 /// This hook must be implemented to lower outgoing return values, described 2281 /// by the Outs array, into the specified DAG. The implementation should 2282 /// return the resulting token chain value. 2283 virtual SDValue 2284 LowerReturn(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 2285 bool /*isVarArg*/, 2286 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 2287 const SmallVectorImpl<SDValue> &/*OutVals*/, 2288 SDLoc /*dl*/, SelectionDAG &/*DAG*/) const { 2289 llvm_unreachable("Not Implemented"); 2290 } 2291 2292 /// Return true if result of the specified node is used by a return node 2293 /// only. It also compute and return the input chain for the tail call. 2294 /// 2295 /// This is used to determine whether it is possible to codegen a libcall as 2296 /// tail call at legalization time. 2297 virtual bool isUsedByReturnOnly(SDNode *, SDValue &/*Chain*/) const { 2298 return false; 2299 } 2300 2301 /// Return true if the target may be able emit the call instruction as a tail 2302 /// call. This is used by optimization passes to determine if it's profitable 2303 /// to duplicate return instructions to enable tailcall optimization. 2304 virtual bool mayBeEmittedAsTailCall(CallInst *) const { 2305 return false; 2306 } 2307 2308 /// Return the builtin name for the __builtin___clear_cache intrinsic 2309 /// Default is to invoke the clear cache library call 2310 virtual const char * getClearCacheBuiltinName() const { 2311 return "__clear_cache"; 2312 } 2313 2314 /// Return the register ID of the name passed in. Used by named register 2315 /// global variables extension. There is no target-independent behaviour 2316 /// so the default action is to bail. 2317 virtual unsigned getRegisterByName(const char* RegName, EVT VT) const { 2318 report_fatal_error("Named registers not implemented for this target"); 2319 } 2320 2321 /// Return the type that should be used to zero or sign extend a 2322 /// zeroext/signext integer argument or return value. FIXME: Most C calling 2323 /// convention requires the return type to be promoted, but this is not true 2324 /// all the time, e.g. i1 on x86-64. It is also not necessary for non-C 2325 /// calling conventions. The frontend should handle this and include all of 2326 /// the necessary information. 2327 virtual MVT getTypeForExtArgOrReturn(MVT VT, 2328 ISD::NodeType /*ExtendKind*/) const { 2329 MVT MinVT = getRegisterType(MVT::i32); 2330 return VT.bitsLT(MinVT) ? MinVT : VT; 2331 } 2332 2333 /// For some targets, an LLVM struct type must be broken down into multiple 2334 /// simple types, but the calling convention specifies that the entire struct 2335 /// must be passed in a block of consecutive registers. 2336 virtual bool 2337 functionArgumentNeedsConsecutiveRegisters(Type *Ty, CallingConv::ID CallConv, 2338 bool isVarArg) const { 2339 return false; 2340 } 2341 2342 /// Returns a 0 terminated array of registers that can be safely used as 2343 /// scratch registers. 2344 virtual const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const { 2345 return nullptr; 2346 } 2347 2348 /// This callback is used to prepare for a volatile or atomic load. 2349 /// It takes a chain node as input and returns the chain for the load itself. 2350 /// 2351 /// Having a callback like this is necessary for targets like SystemZ, 2352 /// which allows a CPU to reuse the result of a previous load indefinitely, 2353 /// even if a cache-coherent store is performed by another CPU. The default 2354 /// implementation does nothing. 2355 virtual SDValue prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL, 2356 SelectionDAG &DAG) const { 2357 return Chain; 2358 } 2359 2360 /// This callback is invoked by the type legalizer to legalize nodes with an 2361 /// illegal operand type but legal result types. It replaces the 2362 /// LowerOperation callback in the type Legalizer. The reason we can not do 2363 /// away with LowerOperation entirely is that LegalizeDAG isn't yet ready to 2364 /// use this callback. 2365 /// 2366 /// TODO: Consider merging with ReplaceNodeResults. 2367 /// 2368 /// The target places new result values for the node in Results (their number 2369 /// and types must exactly match those of the original return values of 2370 /// the node), or leaves Results empty, which indicates that the node is not 2371 /// to be custom lowered after all. 2372 /// The default implementation calls LowerOperation. 2373 virtual void LowerOperationWrapper(SDNode *N, 2374 SmallVectorImpl<SDValue> &Results, 2375 SelectionDAG &DAG) const; 2376 2377 /// This callback is invoked for operations that are unsupported by the 2378 /// target, which are registered to use 'custom' lowering, and whose defined 2379 /// values are all legal. If the target has no operations that require custom 2380 /// lowering, it need not implement this. The default implementation of this 2381 /// aborts. 2382 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; 2383 2384 /// This callback is invoked when a node result type is illegal for the 2385 /// target, and the operation was registered to use 'custom' lowering for that 2386 /// result type. The target places new result values for the node in Results 2387 /// (their number and types must exactly match those of the original return 2388 /// values of the node), or leaves Results empty, which indicates that the 2389 /// node is not to be custom lowered after all. 2390 /// 2391 /// If the target has no operations that require custom lowering, it need not 2392 /// implement this. The default implementation aborts. 2393 virtual void ReplaceNodeResults(SDNode * /*N*/, 2394 SmallVectorImpl<SDValue> &/*Results*/, 2395 SelectionDAG &/*DAG*/) const { 2396 llvm_unreachable("ReplaceNodeResults not implemented for this target!"); 2397 } 2398 2399 /// This method returns the name of a target specific DAG node. 2400 virtual const char *getTargetNodeName(unsigned Opcode) const; 2401 2402 /// This method returns a target specific FastISel object, or null if the 2403 /// target does not support "fast" ISel. 2404 virtual FastISel *createFastISel(FunctionLoweringInfo &, 2405 const TargetLibraryInfo *) const { 2406 return nullptr; 2407 } 2408 2409 2410 bool verifyReturnAddressArgumentIsConstant(SDValue Op, 2411 SelectionDAG &DAG) const; 2412 2413 //===--------------------------------------------------------------------===// 2414 // Inline Asm Support hooks 2415 // 2416 2417 /// This hook allows the target to expand an inline asm call to be explicit 2418 /// llvm code if it wants to. This is useful for turning simple inline asms 2419 /// into LLVM intrinsics, which gives the compiler more information about the 2420 /// behavior of the code. 2421 virtual bool ExpandInlineAsm(CallInst *) const { 2422 return false; 2423 } 2424 2425 enum ConstraintType { 2426 C_Register, // Constraint represents specific register(s). 2427 C_RegisterClass, // Constraint represents any of register(s) in class. 2428 C_Memory, // Memory constraint. 2429 C_Other, // Something else. 2430 C_Unknown // Unsupported constraint. 2431 }; 2432 2433 enum ConstraintWeight { 2434 // Generic weights. 2435 CW_Invalid = -1, // No match. 2436 CW_Okay = 0, // Acceptable. 2437 CW_Good = 1, // Good weight. 2438 CW_Better = 2, // Better weight. 2439 CW_Best = 3, // Best weight. 2440 2441 // Well-known weights. 2442 CW_SpecificReg = CW_Okay, // Specific register operands. 2443 CW_Register = CW_Good, // Register operands. 2444 CW_Memory = CW_Better, // Memory operands. 2445 CW_Constant = CW_Best, // Constant operand. 2446 CW_Default = CW_Okay // Default or don't know type. 2447 }; 2448 2449 /// This contains information for each constraint that we are lowering. 2450 struct AsmOperandInfo : public InlineAsm::ConstraintInfo { 2451 /// This contains the actual string for the code, like "m". TargetLowering 2452 /// picks the 'best' code from ConstraintInfo::Codes that most closely 2453 /// matches the operand. 2454 std::string ConstraintCode; 2455 2456 /// Information about the constraint code, e.g. Register, RegisterClass, 2457 /// Memory, Other, Unknown. 2458 TargetLowering::ConstraintType ConstraintType; 2459 2460 /// If this is the result output operand or a clobber, this is null, 2461 /// otherwise it is the incoming operand to the CallInst. This gets 2462 /// modified as the asm is processed. 2463 Value *CallOperandVal; 2464 2465 /// The ValueType for the operand value. 2466 MVT ConstraintVT; 2467 2468 /// Return true of this is an input operand that is a matching constraint 2469 /// like "4". 2470 bool isMatchingInputConstraint() const; 2471 2472 /// If this is an input matching constraint, this method returns the output 2473 /// operand it matches. 2474 unsigned getMatchedOperand() const; 2475 2476 /// Copy constructor for copying from a ConstraintInfo. 2477 AsmOperandInfo(const InlineAsm::ConstraintInfo &info) 2478 : InlineAsm::ConstraintInfo(info), 2479 ConstraintType(TargetLowering::C_Unknown), 2480 CallOperandVal(nullptr), ConstraintVT(MVT::Other) { 2481 } 2482 }; 2483 2484 typedef std::vector<AsmOperandInfo> AsmOperandInfoVector; 2485 2486 /// Split up the constraint string from the inline assembly value into the 2487 /// specific constraints and their prefixes, and also tie in the associated 2488 /// operand values. If this returns an empty vector, and if the constraint 2489 /// string itself isn't empty, there was an error parsing. 2490 virtual AsmOperandInfoVector ParseConstraints(ImmutableCallSite CS) const; 2491 2492 /// Examine constraint type and operand type and determine a weight value. 2493 /// The operand object must already have been set up with the operand type. 2494 virtual ConstraintWeight getMultipleConstraintMatchWeight( 2495 AsmOperandInfo &info, int maIndex) const; 2496 2497 /// Examine constraint string and operand type and determine a weight value. 2498 /// The operand object must already have been set up with the operand type. 2499 virtual ConstraintWeight getSingleConstraintMatchWeight( 2500 AsmOperandInfo &info, const char *constraint) const; 2501 2502 /// Determines the constraint code and constraint type to use for the specific 2503 /// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType. 2504 /// If the actual operand being passed in is available, it can be passed in as 2505 /// Op, otherwise an empty SDValue can be passed. 2506 virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo, 2507 SDValue Op, 2508 SelectionDAG *DAG = nullptr) const; 2509 2510 /// Given a constraint, return the type of constraint it is for this target. 2511 virtual ConstraintType getConstraintType(const std::string &Constraint) const; 2512 2513 /// Given a physical register constraint (e.g. {edx}), return the register 2514 /// number and the register class for the register. 2515 /// 2516 /// Given a register class constraint, like 'r', if this corresponds directly 2517 /// to an LLVM register class, return a register of 0 and the register class 2518 /// pointer. 2519 /// 2520 /// This should only be used for C_Register constraints. On error, this 2521 /// returns a register number of 0 and a null register class pointer.. 2522 virtual std::pair<unsigned, const TargetRegisterClass*> 2523 getRegForInlineAsmConstraint(const std::string &Constraint, 2524 MVT VT) const; 2525 2526 /// Try to replace an X constraint, which matches anything, with another that 2527 /// has more specific requirements based on the type of the corresponding 2528 /// operand. This returns null if there is no replacement to make. 2529 virtual const char *LowerXConstraint(EVT ConstraintVT) const; 2530 2531 /// Lower the specified operand into the Ops vector. If it is invalid, don't 2532 /// add anything to Ops. 2533 virtual void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, 2534 std::vector<SDValue> &Ops, 2535 SelectionDAG &DAG) const; 2536 2537 //===--------------------------------------------------------------------===// 2538 // Div utility functions 2539 // 2540 SDValue BuildExactSDIV(SDValue Op1, SDValue Op2, SDLoc dl, 2541 SelectionDAG &DAG) const; 2542 SDValue BuildSDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, 2543 bool IsAfterLegalization, 2544 std::vector<SDNode *> *Created) const; 2545 SDValue BuildUDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, 2546 bool IsAfterLegalization, 2547 std::vector<SDNode *> *Created) const; 2548 2549 //===--------------------------------------------------------------------===// 2550 // Legalization utility functions 2551 // 2552 2553 /// Expand a MUL into two nodes. One that computes the high bits of 2554 /// the result and one that computes the low bits. 2555 /// \param HiLoVT The value type to use for the Lo and Hi nodes. 2556 /// \param LL Low bits of the LHS of the MUL. You can use this parameter 2557 /// if you want to control how low bits are extracted from the LHS. 2558 /// \param LH High bits of the LHS of the MUL. See LL for meaning. 2559 /// \param RL Low bits of the RHS of the MUL. See LL for meaning 2560 /// \param RH High bits of the RHS of the MUL. See LL for meaning. 2561 /// \returns true if the node has been expanded. false if it has not 2562 bool expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT, 2563 SelectionDAG &DAG, SDValue LL = SDValue(), 2564 SDValue LH = SDValue(), SDValue RL = SDValue(), 2565 SDValue RH = SDValue()) const; 2566 2567 //===--------------------------------------------------------------------===// 2568 // Instruction Emitting Hooks 2569 // 2570 2571 /// This method should be implemented by targets that mark instructions with 2572 /// the 'usesCustomInserter' flag. These instructions are special in various 2573 /// ways, which require special support to insert. The specified MachineInstr 2574 /// is created but not inserted into any basic blocks, and this method is 2575 /// called to expand it into a sequence of instructions, potentially also 2576 /// creating new basic blocks and control flow. 2577 virtual MachineBasicBlock * 2578 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const; 2579 2580 /// This method should be implemented by targets that mark instructions with 2581 /// the 'hasPostISelHook' flag. These instructions must be adjusted after 2582 /// instruction selection by target hooks. e.g. To fill in optional defs for 2583 /// ARM 's' setting instructions. 2584 virtual void 2585 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const; 2586 }; 2587 2588 /// Given an LLVM IR type and return type attributes, compute the return value 2589 /// EVTs and flags, and optionally also the offsets, if the return value is 2590 /// being lowered to memory. 2591 void GetReturnInfo(Type* ReturnType, AttributeSet attr, 2592 SmallVectorImpl<ISD::OutputArg> &Outs, 2593 const TargetLowering &TLI); 2594 2595 } // end llvm namespace 2596 2597 #endif 2598