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/IRBuilder.h" 34 #include "llvm/IR/InlineAsm.h" 35 #include "llvm/IR/Instructions.h" 36 #include "llvm/MC/MCRegisterInfo.h" 37 #include "llvm/Target/TargetCallingConv.h" 38 #include "llvm/Target/TargetMachine.h" 39 #include <climits> 40 #include <map> 41 #include <vector> 42 43 namespace llvm { 44 class CallInst; 45 class CCState; 46 class FastISel; 47 class FunctionLoweringInfo; 48 class ImmutableCallSite; 49 class IntrinsicInst; 50 class MachineBasicBlock; 51 class MachineFunction; 52 class MachineInstr; 53 class MachineJumpTableInfo; 54 class MachineLoop; 55 class Mangler; 56 class MCContext; 57 class MCExpr; 58 class MCSymbol; 59 template<typename T> class SmallVectorImpl; 60 class DataLayout; 61 class TargetRegisterClass; 62 class TargetLibraryInfo; 63 class TargetLoweringObjectFile; 64 class Value; 65 66 namespace Sched { 67 enum Preference { 68 None, // No preference 69 Source, // Follow source order. 70 RegPressure, // Scheduling for lowest register pressure. 71 Hybrid, // Scheduling for both latency and register pressure. 72 ILP, // Scheduling for ILP in low register pressure mode. 73 VLIW // Scheduling for VLIW targets. 74 }; 75 } 76 77 /// This base class for TargetLowering contains the SelectionDAG-independent 78 /// parts that can be used from the rest of CodeGen. 79 class TargetLoweringBase { 80 TargetLoweringBase(const TargetLoweringBase&) = delete; 81 void operator=(const TargetLoweringBase&) = delete; 82 83 public: 84 /// This enum indicates whether operations are valid for a target, and if not, 85 /// what action should be used to make them valid. 86 enum LegalizeAction { 87 Legal, // The target natively supports this operation. 88 Promote, // This operation should be executed in a larger type. 89 Expand, // Try to expand this to other ops, otherwise use a libcall. 90 Custom // Use the LowerOperation hook to implement custom lowering. 91 }; 92 93 /// This enum indicates whether a types are legal for a target, and if not, 94 /// what action should be used to make them valid. 95 enum LegalizeTypeAction { 96 TypeLegal, // The target natively supports this type. 97 TypePromoteInteger, // Replace this integer with a larger one. 98 TypeExpandInteger, // Split this integer into two of half the size. 99 TypeSoftenFloat, // Convert this float to a same size integer type. 100 TypeExpandFloat, // Split this float into two of half the size. 101 TypeScalarizeVector, // Replace this one-element vector with its element. 102 TypeSplitVector, // Split this vector into two of half the size. 103 TypeWidenVector, // This vector should be widened into a larger vector. 104 TypePromoteFloat // Replace this float with a larger one. 105 }; 106 107 /// LegalizeKind holds the legalization kind that needs to happen to EVT 108 /// in order to type-legalize it. 109 typedef std::pair<LegalizeTypeAction, EVT> LegalizeKind; 110 111 /// Enum that describes how the target represents true/false values. 112 enum BooleanContent { 113 UndefinedBooleanContent, // Only bit 0 counts, the rest can hold garbage. 114 ZeroOrOneBooleanContent, // All bits zero except for bit 0. 115 ZeroOrNegativeOneBooleanContent // All bits equal to bit 0. 116 }; 117 118 /// Enum that describes what type of support for selects the target has. 119 enum SelectSupportKind { 120 ScalarValSelect, // The target supports scalar selects (ex: cmov). 121 ScalarCondVectorVal, // The target supports selects with a scalar condition 122 // and vector values (ex: cmov). 123 VectorMaskSelect // The target supports vector selects with a vector 124 // mask (ex: x86 blends). 125 }; 126 127 /// Enum that specifies what a AtomicRMWInst is expanded to, if at all. Exists 128 /// because different targets have different levels of support for these 129 /// atomic RMW instructions, and also have different options w.r.t. what they 130 /// should expand to. 131 enum class AtomicRMWExpansionKind { 132 None, // Don't expand the instruction. 133 LLSC, // Expand the instruction into loadlinked/storeconditional; used 134 // by ARM/AArch64. Implies `hasLoadLinkedStoreConditional` 135 // returns true. 136 CmpXChg, // Expand the instruction into cmpxchg; used by at least X86. 137 }; 138 getExtendForContent(BooleanContent Content)139 static ISD::NodeType getExtendForContent(BooleanContent Content) { 140 switch (Content) { 141 case UndefinedBooleanContent: 142 // Extend by adding rubbish bits. 143 return ISD::ANY_EXTEND; 144 case ZeroOrOneBooleanContent: 145 // Extend by adding zero bits. 146 return ISD::ZERO_EXTEND; 147 case ZeroOrNegativeOneBooleanContent: 148 // Extend by copying the sign bit. 149 return ISD::SIGN_EXTEND; 150 } 151 llvm_unreachable("Invalid content kind"); 152 } 153 154 /// NOTE: The TargetMachine owns TLOF. 155 explicit TargetLoweringBase(const TargetMachine &TM); ~TargetLoweringBase()156 virtual ~TargetLoweringBase() {} 157 158 protected: 159 /// \brief Initialize all of the actions to default values. 160 void initActions(); 161 162 public: getTargetMachine()163 const TargetMachine &getTargetMachine() const { return TM; } 164 useSoftFloat()165 virtual bool useSoftFloat() const { return false; } 166 167 /// Return the pointer type for the given address space, defaults to 168 /// the pointer type from the data layout. 169 /// FIXME: The default needs to be removed once all the code is updated. 170 MVT getPointerTy(const DataLayout &DL, uint32_t AS = 0) const { 171 return MVT::getIntegerVT(DL.getPointerSizeInBits(AS)); 172 } 173 174 /// EVT is not used in-tree, but is used by out-of-tree target. 175 /// A documentation for this function would be nice... 176 virtual MVT getScalarShiftAmountTy(const DataLayout &, EVT) const; 177 178 EVT getShiftAmountTy(EVT LHSTy, const DataLayout &DL) const; 179 180 /// Returns the type to be used for the index operand of: 181 /// ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT, 182 /// ISD::INSERT_SUBVECTOR, and ISD::EXTRACT_SUBVECTOR getVectorIdxTy(const DataLayout & DL)183 virtual MVT getVectorIdxTy(const DataLayout &DL) const { 184 return getPointerTy(DL); 185 } 186 187 /// Return true if the select operation is expensive for this target. isSelectExpensive()188 bool isSelectExpensive() const { return SelectIsExpensive; } 189 isSelectSupported(SelectSupportKind)190 virtual bool isSelectSupported(SelectSupportKind /*kind*/) const { 191 return true; 192 } 193 194 /// Return true if multiple condition registers are available. hasMultipleConditionRegisters()195 bool hasMultipleConditionRegisters() const { 196 return HasMultipleConditionRegisters; 197 } 198 199 /// Return true if the target has BitExtract instructions. hasExtractBitsInsn()200 bool hasExtractBitsInsn() const { return HasExtractBitsInsn; } 201 202 /// Return the preferred vector type legalization action. 203 virtual TargetLoweringBase::LegalizeTypeAction getPreferredVectorAction(EVT VT)204 getPreferredVectorAction(EVT VT) const { 205 // The default action for one element vectors is to scalarize 206 if (VT.getVectorNumElements() == 1) 207 return TypeScalarizeVector; 208 // The default action for other vectors is to promote 209 return TypePromoteInteger; 210 } 211 212 // There are two general methods for expanding a BUILD_VECTOR node: 213 // 1. Use SCALAR_TO_VECTOR on the defined scalar values and then shuffle 214 // them together. 215 // 2. Build the vector on the stack and then load it. 216 // If this function returns true, then method (1) will be used, subject to 217 // the constraint that all of the necessary shuffles are legal (as determined 218 // by isShuffleMaskLegal). If this function returns false, then method (2) is 219 // always used. The vector type, and the number of defined values, are 220 // provided. 221 virtual bool shouldExpandBuildVectorWithShuffles(EVT,unsigned DefinedValues)222 shouldExpandBuildVectorWithShuffles(EVT /* VT */, 223 unsigned DefinedValues) const { 224 return DefinedValues < 3; 225 } 226 227 /// Return true if integer divide is usually cheaper than a sequence of 228 /// several shifts, adds, and multiplies for this target. isIntDivCheap()229 bool isIntDivCheap() const { return IntDivIsCheap; } 230 231 /// Return true if sqrt(x) is as cheap or cheaper than 1 / rsqrt(x) isFsqrtCheap()232 bool isFsqrtCheap() const { 233 return FsqrtIsCheap; 234 } 235 236 /// Returns true if target has indicated at least one type should be bypassed. isSlowDivBypassed()237 bool isSlowDivBypassed() const { return !BypassSlowDivWidths.empty(); } 238 239 /// Returns map of slow types for division or remainder with corresponding 240 /// fast types getBypassSlowDivWidths()241 const DenseMap<unsigned int, unsigned int> &getBypassSlowDivWidths() const { 242 return BypassSlowDivWidths; 243 } 244 245 /// Return true if pow2 sdiv is cheaper than a chain of sra/srl/add/sra. isPow2SDivCheap()246 bool isPow2SDivCheap() const { return Pow2SDivIsCheap; } 247 248 /// Return true if Flow Control is an expensive operation that should be 249 /// avoided. isJumpExpensive()250 bool isJumpExpensive() const { return JumpIsExpensive; } 251 252 /// Return true if selects are only cheaper than branches if the branch is 253 /// unlikely to be predicted right. isPredictableSelectExpensive()254 bool isPredictableSelectExpensive() const { 255 return PredictableSelectIsExpensive; 256 } 257 258 /// isLoadBitCastBeneficial() - Return true if the following transform 259 /// is beneficial. 260 /// fold (conv (load x)) -> (load (conv*)x) 261 /// On architectures that don't natively support some vector loads 262 /// efficiently, casting the load to a smaller vector of larger types and 263 /// loading is more efficient, however, this can be undone by optimizations in 264 /// dag combiner. isLoadBitCastBeneficial(EVT,EVT)265 virtual bool isLoadBitCastBeneficial(EVT /* Load */, 266 EVT /* Bitcast */) const { 267 return true; 268 } 269 270 /// Return true if it is expected to be cheaper to do a store of a non-zero 271 /// vector constant with the given size and type for the address space than to 272 /// store the individual scalar element constants. storeOfVectorConstantIsCheap(EVT MemVT,unsigned NumElem,unsigned AddrSpace)273 virtual bool storeOfVectorConstantIsCheap(EVT MemVT, 274 unsigned NumElem, 275 unsigned AddrSpace) const { 276 return false; 277 } 278 279 /// \brief Return true if it is cheap to speculate a call to intrinsic cttz. isCheapToSpeculateCttz()280 virtual bool isCheapToSpeculateCttz() const { 281 return false; 282 } 283 284 /// \brief Return true if it is cheap to speculate a call to intrinsic ctlz. isCheapToSpeculateCtlz()285 virtual bool isCheapToSpeculateCtlz() const { 286 return false; 287 } 288 289 /// \brief Return if the target supports combining a 290 /// chain like: 291 /// \code 292 /// %andResult = and %val1, #imm-with-one-bit-set; 293 /// %icmpResult = icmp %andResult, 0 294 /// br i1 %icmpResult, label %dest1, label %dest2 295 /// \endcode 296 /// into a single machine instruction of a form like: 297 /// \code 298 /// brOnBitSet %register, #bitNumber, dest 299 /// \endcode isMaskAndBranchFoldingLegal()300 bool isMaskAndBranchFoldingLegal() const { 301 return MaskAndBranchFoldingIsLegal; 302 } 303 304 /// \brief Return true if the target wants to use the optimization that 305 /// turns ext(promotableInst1(...(promotableInstN(load)))) into 306 /// promotedInst1(...(promotedInstN(ext(load)))). enableExtLdPromotion()307 bool enableExtLdPromotion() const { return EnableExtLdPromotion; } 308 309 /// Return true if the target can combine store(extractelement VectorTy, 310 /// Idx). 311 /// \p Cost[out] gives the cost of that transformation when this is true. canCombineStoreAndExtract(Type * VectorTy,Value * Idx,unsigned & Cost)312 virtual bool canCombineStoreAndExtract(Type *VectorTy, Value *Idx, 313 unsigned &Cost) const { 314 return false; 315 } 316 317 /// Return true if target supports floating point exceptions. hasFloatingPointExceptions()318 bool hasFloatingPointExceptions() const { 319 return HasFloatingPointExceptions; 320 } 321 322 /// Return true if target always beneficiates from combining into FMA for a 323 /// given value type. This must typically return false on targets where FMA 324 /// takes more cycles to execute than FADD. enableAggressiveFMAFusion(EVT VT)325 virtual bool enableAggressiveFMAFusion(EVT VT) const { 326 return false; 327 } 328 329 /// Return the ValueType of the result of SETCC operations. 330 virtual EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context, 331 EVT VT) const; 332 333 /// Return the ValueType for comparison libcalls. Comparions libcalls include 334 /// floating point comparion calls, and Ordered/Unordered check calls on 335 /// floating point numbers. 336 virtual 337 MVT::SimpleValueType getCmpLibcallReturnType() const; 338 339 /// For targets without i1 registers, this gives the nature of the high-bits 340 /// of boolean values held in types wider than i1. 341 /// 342 /// "Boolean values" are special true/false values produced by nodes like 343 /// SETCC and consumed (as the condition) by nodes like SELECT and BRCOND. 344 /// Not to be confused with general values promoted from i1. Some cpus 345 /// distinguish between vectors of boolean and scalars; the isVec parameter 346 /// selects between the two kinds. For example on X86 a scalar boolean should 347 /// be zero extended from i1, while the elements of a vector of booleans 348 /// should be sign extended from i1. 349 /// 350 /// Some cpus also treat floating point types the same way as they treat 351 /// vectors instead of the way they treat scalars. getBooleanContents(bool isVec,bool isFloat)352 BooleanContent getBooleanContents(bool isVec, bool isFloat) const { 353 if (isVec) 354 return BooleanVectorContents; 355 return isFloat ? BooleanFloatContents : BooleanContents; 356 } 357 getBooleanContents(EVT Type)358 BooleanContent getBooleanContents(EVT Type) const { 359 return getBooleanContents(Type.isVector(), Type.isFloatingPoint()); 360 } 361 362 /// Return target scheduling preference. getSchedulingPreference()363 Sched::Preference getSchedulingPreference() const { 364 return SchedPreferenceInfo; 365 } 366 367 /// Some scheduler, e.g. hybrid, can switch to different scheduling heuristics 368 /// for different nodes. This function returns the preference (or none) for 369 /// the given node. getSchedulingPreference(SDNode *)370 virtual Sched::Preference getSchedulingPreference(SDNode *) const { 371 return Sched::None; 372 } 373 374 /// Return the register class that should be used for the specified value 375 /// type. getRegClassFor(MVT VT)376 virtual const TargetRegisterClass *getRegClassFor(MVT VT) const { 377 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 378 assert(RC && "This value type is not natively supported!"); 379 return RC; 380 } 381 382 /// Return the 'representative' register class for the specified value 383 /// type. 384 /// 385 /// The 'representative' register class is the largest legal super-reg 386 /// register class for the register class of the value type. For example, on 387 /// i386 the rep register class for i8, i16, and i32 are GR32; while the rep 388 /// register class is GR64 on x86_64. getRepRegClassFor(MVT VT)389 virtual const TargetRegisterClass *getRepRegClassFor(MVT VT) const { 390 const TargetRegisterClass *RC = RepRegClassForVT[VT.SimpleTy]; 391 return RC; 392 } 393 394 /// Return the cost of the 'representative' register class for the specified 395 /// value type. getRepRegClassCostFor(MVT VT)396 virtual uint8_t getRepRegClassCostFor(MVT VT) const { 397 return RepRegClassCostForVT[VT.SimpleTy]; 398 } 399 400 /// Return true if the target has native support for the specified value type. 401 /// This means that it has a register that directly holds it without 402 /// promotions or expansions. isTypeLegal(EVT VT)403 bool isTypeLegal(EVT VT) const { 404 assert(!VT.isSimple() || 405 (unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT)); 406 return VT.isSimple() && RegClassForVT[VT.getSimpleVT().SimpleTy] != nullptr; 407 } 408 409 class ValueTypeActionImpl { 410 /// ValueTypeActions - For each value type, keep a LegalizeTypeAction enum 411 /// that indicates how instruction selection should deal with the type. 412 uint8_t ValueTypeActions[MVT::LAST_VALUETYPE]; 413 414 public: ValueTypeActionImpl()415 ValueTypeActionImpl() { 416 std::fill(std::begin(ValueTypeActions), std::end(ValueTypeActions), 0); 417 } 418 getTypeAction(MVT VT)419 LegalizeTypeAction getTypeAction(MVT VT) const { 420 return (LegalizeTypeAction)ValueTypeActions[VT.SimpleTy]; 421 } 422 setTypeAction(MVT VT,LegalizeTypeAction Action)423 void setTypeAction(MVT VT, LegalizeTypeAction Action) { 424 unsigned I = VT.SimpleTy; 425 ValueTypeActions[I] = Action; 426 } 427 }; 428 getValueTypeActions()429 const ValueTypeActionImpl &getValueTypeActions() const { 430 return ValueTypeActions; 431 } 432 433 /// Return how we should legalize values of this type, either it is already 434 /// legal (return 'Legal') or we need to promote it to a larger type (return 435 /// 'Promote'), or we need to expand it into multiple registers of smaller 436 /// integer type (return 'Expand'). 'Custom' is not an option. getTypeAction(LLVMContext & Context,EVT VT)437 LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const { 438 return getTypeConversion(Context, VT).first; 439 } getTypeAction(MVT VT)440 LegalizeTypeAction getTypeAction(MVT VT) const { 441 return ValueTypeActions.getTypeAction(VT); 442 } 443 444 /// For types supported by the target, this is an identity function. For 445 /// types that must be promoted to larger types, this returns the larger type 446 /// to promote to. For integer types that are larger than the largest integer 447 /// register, this contains one step in the expansion to get to the smaller 448 /// register. For illegal floating point types, this returns the integer type 449 /// to transform to. getTypeToTransformTo(LLVMContext & Context,EVT VT)450 EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const { 451 return getTypeConversion(Context, VT).second; 452 } 453 454 /// For types supported by the target, this is an identity function. For 455 /// types that must be expanded (i.e. integer types that are larger than the 456 /// largest integer register or illegal floating point types), this returns 457 /// the largest legal type it will be expanded to. getTypeToExpandTo(LLVMContext & Context,EVT VT)458 EVT getTypeToExpandTo(LLVMContext &Context, EVT VT) const { 459 assert(!VT.isVector()); 460 while (true) { 461 switch (getTypeAction(Context, VT)) { 462 case TypeLegal: 463 return VT; 464 case TypeExpandInteger: 465 VT = getTypeToTransformTo(Context, VT); 466 break; 467 default: 468 llvm_unreachable("Type is not legal nor is it to be expanded!"); 469 } 470 } 471 } 472 473 /// Vector types are broken down into some number of legal first class types. 474 /// For example, EVT::v8f32 maps to 2 EVT::v4f32 with Altivec or SSE1, or 8 475 /// promoted EVT::f64 values with the X86 FP stack. Similarly, EVT::v2i64 476 /// turns into 4 EVT::i32 values with both PPC and X86. 477 /// 478 /// This method returns the number of registers needed, and the VT for each 479 /// register. It also returns the VT and quantity of the intermediate values 480 /// before they are promoted/expanded. 481 unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 482 EVT &IntermediateVT, 483 unsigned &NumIntermediates, 484 MVT &RegisterVT) const; 485 486 struct IntrinsicInfo { 487 unsigned opc; // target opcode 488 EVT memVT; // memory VT 489 const Value* ptrVal; // value representing memory location 490 int offset; // offset off of ptrVal 491 unsigned size; // the size of the memory location 492 // (taken from memVT if zero) 493 unsigned align; // alignment 494 bool vol; // is volatile? 495 bool readMem; // reads memory? 496 bool writeMem; // writes memory? 497 IntrinsicInfoIntrinsicInfo498 IntrinsicInfo() : opc(0), ptrVal(nullptr), offset(0), size(0), align(1), 499 vol(false), readMem(false), writeMem(false) {} 500 }; 501 502 /// Given an intrinsic, checks if on the target the intrinsic will need to map 503 /// to a MemIntrinsicNode (touches memory). If this is the case, it returns 504 /// true and store the intrinsic information into the IntrinsicInfo that was 505 /// passed to the function. getTgtMemIntrinsic(IntrinsicInfo &,const CallInst &,unsigned)506 virtual bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &, 507 unsigned /*Intrinsic*/) const { 508 return false; 509 } 510 511 /// Returns true if the target can instruction select the specified FP 512 /// immediate natively. If false, the legalizer will materialize the FP 513 /// immediate as a load from a constant pool. isFPImmLegal(const APFloat &,EVT)514 virtual bool isFPImmLegal(const APFloat &/*Imm*/, EVT /*VT*/) const { 515 return false; 516 } 517 518 /// Targets can use this to indicate that they only support *some* 519 /// VECTOR_SHUFFLE operations, those with specific masks. By default, if a 520 /// target supports the VECTOR_SHUFFLE node, all mask values are assumed to be 521 /// legal. isShuffleMaskLegal(const SmallVectorImpl<int> &,EVT)522 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 523 EVT /*VT*/) const { 524 return true; 525 } 526 527 /// Returns true if the operation can trap for the value type. 528 /// 529 /// VT must be a legal type. By default, we optimistically assume most 530 /// operations don't trap except for divide and remainder. 531 virtual bool canOpTrap(unsigned Op, EVT VT) const; 532 533 /// Similar to isShuffleMaskLegal. This is used by Targets can use this to 534 /// indicate if there is a suitable VECTOR_SHUFFLE that can be used to replace 535 /// a VAND with a constant pool entry. isVectorClearMaskLegal(const SmallVectorImpl<int> &,EVT)536 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 537 EVT /*VT*/) const { 538 return false; 539 } 540 541 /// Return how this operation should be treated: either it is legal, needs to 542 /// be promoted to a larger size, needs to be expanded to some other code 543 /// sequence, or the target has a custom expander for it. getOperationAction(unsigned Op,EVT VT)544 LegalizeAction getOperationAction(unsigned Op, EVT VT) const { 545 if (VT.isExtended()) return Expand; 546 // If a target-specific SDNode requires legalization, require the target 547 // to provide custom legalization for it. 548 if (Op > array_lengthof(OpActions[0])) return Custom; 549 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy; 550 return (LegalizeAction)OpActions[I][Op]; 551 } 552 553 /// Return true if the specified operation is legal on this target or can be 554 /// made legal with custom lowering. This is used to help guide high-level 555 /// lowering decisions. isOperationLegalOrCustom(unsigned Op,EVT VT)556 bool isOperationLegalOrCustom(unsigned Op, EVT VT) const { 557 return (VT == MVT::Other || isTypeLegal(VT)) && 558 (getOperationAction(Op, VT) == Legal || 559 getOperationAction(Op, VT) == Custom); 560 } 561 562 /// Return true if the specified operation is legal on this target or can be 563 /// made legal using promotion. This is used to help guide high-level lowering 564 /// decisions. isOperationLegalOrPromote(unsigned Op,EVT VT)565 bool isOperationLegalOrPromote(unsigned Op, EVT VT) const { 566 return (VT == MVT::Other || isTypeLegal(VT)) && 567 (getOperationAction(Op, VT) == Legal || 568 getOperationAction(Op, VT) == Promote); 569 } 570 571 /// Return true if the specified operation is illegal on this target or 572 /// unlikely to be made legal with custom lowering. This is used to help guide 573 /// high-level lowering decisions. isOperationExpand(unsigned Op,EVT VT)574 bool isOperationExpand(unsigned Op, EVT VT) const { 575 return (!isTypeLegal(VT) || getOperationAction(Op, VT) == Expand); 576 } 577 578 /// Return true if the specified operation is legal on this target. isOperationLegal(unsigned Op,EVT VT)579 bool isOperationLegal(unsigned Op, EVT VT) const { 580 return (VT == MVT::Other || isTypeLegal(VT)) && 581 getOperationAction(Op, VT) == Legal; 582 } 583 584 /// Return how this load with extension should be treated: either it is legal, 585 /// needs to be promoted to a larger size, needs to be expanded to some other 586 /// code sequence, or the target has a custom expander for it. getLoadExtAction(unsigned ExtType,EVT ValVT,EVT MemVT)587 LegalizeAction getLoadExtAction(unsigned ExtType, EVT ValVT, 588 EVT MemVT) const { 589 if (ValVT.isExtended() || MemVT.isExtended()) return Expand; 590 unsigned ValI = (unsigned) ValVT.getSimpleVT().SimpleTy; 591 unsigned MemI = (unsigned) MemVT.getSimpleVT().SimpleTy; 592 assert(ExtType < ISD::LAST_LOADEXT_TYPE && ValI < MVT::LAST_VALUETYPE && 593 MemI < MVT::LAST_VALUETYPE && "Table isn't big enough!"); 594 return (LegalizeAction)LoadExtActions[ValI][MemI][ExtType]; 595 } 596 597 /// Return true if the specified load with extension is legal on this target. isLoadExtLegal(unsigned ExtType,EVT ValVT,EVT MemVT)598 bool isLoadExtLegal(unsigned ExtType, EVT ValVT, EVT MemVT) const { 599 return ValVT.isSimple() && MemVT.isSimple() && 600 getLoadExtAction(ExtType, ValVT, MemVT) == Legal; 601 } 602 603 /// Return true if the specified load with extension is legal or custom 604 /// on this target. isLoadExtLegalOrCustom(unsigned ExtType,EVT ValVT,EVT MemVT)605 bool isLoadExtLegalOrCustom(unsigned ExtType, EVT ValVT, EVT MemVT) const { 606 return ValVT.isSimple() && MemVT.isSimple() && 607 (getLoadExtAction(ExtType, ValVT, MemVT) == Legal || 608 getLoadExtAction(ExtType, ValVT, MemVT) == Custom); 609 } 610 611 /// Return how this store with truncation should be treated: either it is 612 /// legal, needs to be promoted to a larger size, needs to be expanded to some 613 /// other code sequence, or the target has a custom expander for it. getTruncStoreAction(EVT ValVT,EVT MemVT)614 LegalizeAction getTruncStoreAction(EVT ValVT, EVT MemVT) const { 615 if (ValVT.isExtended() || MemVT.isExtended()) return Expand; 616 unsigned ValI = (unsigned) ValVT.getSimpleVT().SimpleTy; 617 unsigned MemI = (unsigned) MemVT.getSimpleVT().SimpleTy; 618 assert(ValI < MVT::LAST_VALUETYPE && MemI < MVT::LAST_VALUETYPE && 619 "Table isn't big enough!"); 620 return (LegalizeAction)TruncStoreActions[ValI][MemI]; 621 } 622 623 /// Return true if the specified store with truncation is legal on this 624 /// target. isTruncStoreLegal(EVT ValVT,EVT MemVT)625 bool isTruncStoreLegal(EVT ValVT, EVT MemVT) const { 626 return isTypeLegal(ValVT) && MemVT.isSimple() && 627 getTruncStoreAction(ValVT.getSimpleVT(), MemVT.getSimpleVT()) == Legal; 628 } 629 630 /// Return how the indexed load should be treated: either it is legal, needs 631 /// to be promoted to a larger size, needs to be expanded to some other code 632 /// sequence, or the target has a custom expander for it. 633 LegalizeAction getIndexedLoadAction(unsigned IdxMode,MVT VT)634 getIndexedLoadAction(unsigned IdxMode, MVT VT) const { 635 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT.isValid() && 636 "Table isn't big enough!"); 637 unsigned Ty = (unsigned)VT.SimpleTy; 638 return (LegalizeAction)((IndexedModeActions[Ty][IdxMode] & 0xf0) >> 4); 639 } 640 641 /// Return true if the specified indexed load is legal on this target. isIndexedLoadLegal(unsigned IdxMode,EVT VT)642 bool isIndexedLoadLegal(unsigned IdxMode, EVT VT) const { 643 return VT.isSimple() && 644 (getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Legal || 645 getIndexedLoadAction(IdxMode, VT.getSimpleVT()) == Custom); 646 } 647 648 /// Return how the indexed store should be treated: either it is legal, needs 649 /// to be promoted to a larger size, needs to be expanded to some other code 650 /// sequence, or the target has a custom expander for it. 651 LegalizeAction getIndexedStoreAction(unsigned IdxMode,MVT VT)652 getIndexedStoreAction(unsigned IdxMode, MVT VT) const { 653 assert(IdxMode < ISD::LAST_INDEXED_MODE && VT.isValid() && 654 "Table isn't big enough!"); 655 unsigned Ty = (unsigned)VT.SimpleTy; 656 return (LegalizeAction)(IndexedModeActions[Ty][IdxMode] & 0x0f); 657 } 658 659 /// Return true if the specified indexed load is legal on this target. isIndexedStoreLegal(unsigned IdxMode,EVT VT)660 bool isIndexedStoreLegal(unsigned IdxMode, EVT VT) const { 661 return VT.isSimple() && 662 (getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Legal || 663 getIndexedStoreAction(IdxMode, VT.getSimpleVT()) == Custom); 664 } 665 666 /// Return how the condition code should be treated: either it is legal, needs 667 /// to be expanded to some other code sequence, or the target has a custom 668 /// expander for it. 669 LegalizeAction getCondCodeAction(ISD::CondCode CC,MVT VT)670 getCondCodeAction(ISD::CondCode CC, MVT VT) const { 671 assert((unsigned)CC < array_lengthof(CondCodeActions) && 672 ((unsigned)VT.SimpleTy >> 4) < array_lengthof(CondCodeActions[0]) && 673 "Table isn't big enough!"); 674 // See setCondCodeAction for how this is encoded. 675 uint32_t Shift = 2 * (VT.SimpleTy & 0xF); 676 uint32_t Value = CondCodeActions[CC][VT.SimpleTy >> 4]; 677 LegalizeAction Action = (LegalizeAction) ((Value >> Shift) & 0x3); 678 assert(Action != Promote && "Can't promote condition code!"); 679 return Action; 680 } 681 682 /// Return true if the specified condition code is legal on this target. isCondCodeLegal(ISD::CondCode CC,MVT VT)683 bool isCondCodeLegal(ISD::CondCode CC, MVT VT) const { 684 return 685 getCondCodeAction(CC, VT) == Legal || 686 getCondCodeAction(CC, VT) == Custom; 687 } 688 689 690 /// If the action for this operation is to promote, this method returns the 691 /// ValueType to promote to. getTypeToPromoteTo(unsigned Op,MVT VT)692 MVT getTypeToPromoteTo(unsigned Op, MVT VT) const { 693 assert(getOperationAction(Op, VT) == Promote && 694 "This operation isn't promoted!"); 695 696 // See if this has an explicit type specified. 697 std::map<std::pair<unsigned, MVT::SimpleValueType>, 698 MVT::SimpleValueType>::const_iterator PTTI = 699 PromoteToType.find(std::make_pair(Op, VT.SimpleTy)); 700 if (PTTI != PromoteToType.end()) return PTTI->second; 701 702 assert((VT.isInteger() || VT.isFloatingPoint()) && 703 "Cannot autopromote this type, add it with AddPromotedToType."); 704 705 MVT NVT = VT; 706 do { 707 NVT = (MVT::SimpleValueType)(NVT.SimpleTy+1); 708 assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid && 709 "Didn't find type to promote to!"); 710 } while (!isTypeLegal(NVT) || 711 getOperationAction(Op, NVT) == Promote); 712 return NVT; 713 } 714 715 /// Return the EVT corresponding to this LLVM type. This is fixed by the LLVM 716 /// operations except for the pointer size. If AllowUnknown is true, this 717 /// will return MVT::Other for types with no EVT counterpart (e.g. structs), 718 /// otherwise it will assert. 719 EVT getValueType(const DataLayout &DL, Type *Ty, 720 bool AllowUnknown = false) const { 721 // Lower scalar pointers to native pointer types. 722 if (PointerType *PTy = dyn_cast<PointerType>(Ty)) 723 return getPointerTy(DL, PTy->getAddressSpace()); 724 725 if (Ty->isVectorTy()) { 726 VectorType *VTy = cast<VectorType>(Ty); 727 Type *Elm = VTy->getElementType(); 728 // Lower vectors of pointers to native pointer types. 729 if (PointerType *PT = dyn_cast<PointerType>(Elm)) { 730 EVT PointerTy(getPointerTy(DL, PT->getAddressSpace())); 731 Elm = PointerTy.getTypeForEVT(Ty->getContext()); 732 } 733 734 return EVT::getVectorVT(Ty->getContext(), EVT::getEVT(Elm, false), 735 VTy->getNumElements()); 736 } 737 return EVT::getEVT(Ty, AllowUnknown); 738 } 739 740 /// Return the MVT corresponding to this LLVM type. See getValueType. 741 MVT getSimpleValueType(const DataLayout &DL, Type *Ty, 742 bool AllowUnknown = false) const { 743 return getValueType(DL, Ty, AllowUnknown).getSimpleVT(); 744 } 745 746 /// Return the desired alignment for ByVal or InAlloca aggregate function 747 /// arguments in the caller parameter area. This is the actual alignment, not 748 /// its logarithm. 749 virtual unsigned getByValTypeAlignment(Type *Ty, const DataLayout &DL) const; 750 751 /// Return the type of registers that this ValueType will eventually require. getRegisterType(MVT VT)752 MVT getRegisterType(MVT VT) const { 753 assert((unsigned)VT.SimpleTy < array_lengthof(RegisterTypeForVT)); 754 return RegisterTypeForVT[VT.SimpleTy]; 755 } 756 757 /// Return the type of registers that this ValueType will eventually require. getRegisterType(LLVMContext & Context,EVT VT)758 MVT getRegisterType(LLVMContext &Context, EVT VT) const { 759 if (VT.isSimple()) { 760 assert((unsigned)VT.getSimpleVT().SimpleTy < 761 array_lengthof(RegisterTypeForVT)); 762 return RegisterTypeForVT[VT.getSimpleVT().SimpleTy]; 763 } 764 if (VT.isVector()) { 765 EVT VT1; 766 MVT RegisterVT; 767 unsigned NumIntermediates; 768 (void)getVectorTypeBreakdown(Context, VT, VT1, 769 NumIntermediates, RegisterVT); 770 return RegisterVT; 771 } 772 if (VT.isInteger()) { 773 return getRegisterType(Context, getTypeToTransformTo(Context, VT)); 774 } 775 llvm_unreachable("Unsupported extended type!"); 776 } 777 778 /// Return the number of registers that this ValueType will eventually 779 /// require. 780 /// 781 /// This is one for any types promoted to live in larger registers, but may be 782 /// more than one for types (like i64) that are split into pieces. For types 783 /// like i140, which are first promoted then expanded, it is the number of 784 /// registers needed to hold all the bits of the original type. For an i140 785 /// on a 32 bit machine this means 5 registers. getNumRegisters(LLVMContext & Context,EVT VT)786 unsigned getNumRegisters(LLVMContext &Context, EVT VT) const { 787 if (VT.isSimple()) { 788 assert((unsigned)VT.getSimpleVT().SimpleTy < 789 array_lengthof(NumRegistersForVT)); 790 return NumRegistersForVT[VT.getSimpleVT().SimpleTy]; 791 } 792 if (VT.isVector()) { 793 EVT VT1; 794 MVT VT2; 795 unsigned NumIntermediates; 796 return getVectorTypeBreakdown(Context, VT, VT1, NumIntermediates, VT2); 797 } 798 if (VT.isInteger()) { 799 unsigned BitWidth = VT.getSizeInBits(); 800 unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits(); 801 return (BitWidth + RegWidth - 1) / RegWidth; 802 } 803 llvm_unreachable("Unsupported extended type!"); 804 } 805 806 /// If true, then instruction selection should seek to shrink the FP constant 807 /// of the specified type to a smaller type in order to save space and / or 808 /// reduce runtime. ShouldShrinkFPConstant(EVT)809 virtual bool ShouldShrinkFPConstant(EVT) const { return true; } 810 811 // Return true if it is profitable to reduce the given load node to a smaller 812 // type. 813 // 814 // e.g. (i16 (trunc (i32 (load x))) -> i16 load x should be performed shouldReduceLoadWidth(SDNode * Load,ISD::LoadExtType ExtTy,EVT NewVT)815 virtual bool shouldReduceLoadWidth(SDNode *Load, 816 ISD::LoadExtType ExtTy, 817 EVT NewVT) const { 818 return true; 819 } 820 821 /// When splitting a value of the specified type into parts, does the Lo 822 /// or Hi part come first? This usually follows the endianness, except 823 /// for ppcf128, where the Hi part always comes first. hasBigEndianPartOrdering(EVT VT,const DataLayout & DL)824 bool hasBigEndianPartOrdering(EVT VT, const DataLayout &DL) const { 825 return DL.isBigEndian() || VT == MVT::ppcf128; 826 } 827 828 /// If true, the target has custom DAG combine transformations that it can 829 /// perform for the specified node. hasTargetDAGCombine(ISD::NodeType NT)830 bool hasTargetDAGCombine(ISD::NodeType NT) const { 831 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 832 return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7)); 833 } 834 835 /// \brief Get maximum # of store operations permitted for llvm.memset 836 /// 837 /// This function returns the maximum number of store operations permitted 838 /// to replace a call to llvm.memset. The value is set by the target at the 839 /// performance threshold for such a replacement. If OptSize is true, 840 /// return the limit for functions that have OptSize attribute. getMaxStoresPerMemset(bool OptSize)841 unsigned getMaxStoresPerMemset(bool OptSize) const { 842 return OptSize ? MaxStoresPerMemsetOptSize : MaxStoresPerMemset; 843 } 844 845 /// \brief Get maximum # of store operations permitted for llvm.memcpy 846 /// 847 /// This function returns the maximum number of store operations permitted 848 /// to replace a call to llvm.memcpy. The value is set by the target at the 849 /// performance threshold for such a replacement. If OptSize is true, 850 /// return the limit for functions that have OptSize attribute. getMaxStoresPerMemcpy(bool OptSize)851 unsigned getMaxStoresPerMemcpy(bool OptSize) const { 852 return OptSize ? MaxStoresPerMemcpyOptSize : MaxStoresPerMemcpy; 853 } 854 855 /// \brief Get maximum # of store operations permitted for llvm.memmove 856 /// 857 /// This function returns the maximum number of store operations permitted 858 /// to replace a call to llvm.memmove. The value is set by the target at the 859 /// performance threshold for such a replacement. If OptSize is true, 860 /// return the limit for functions that have OptSize attribute. getMaxStoresPerMemmove(bool OptSize)861 unsigned getMaxStoresPerMemmove(bool OptSize) const { 862 return OptSize ? MaxStoresPerMemmoveOptSize : MaxStoresPerMemmove; 863 } 864 865 /// \brief Determine if the target supports unaligned memory accesses. 866 /// 867 /// This function returns true if the target allows unaligned memory accesses 868 /// of the specified type in the given address space. If true, it also returns 869 /// whether the unaligned memory access is "fast" in the last argument by 870 /// reference. This is used, for example, in situations where an array 871 /// copy/move/set is converted to a sequence of store operations. Its use 872 /// helps to ensure that such replacements don't generate code that causes an 873 /// alignment error (trap) on the target machine. 874 virtual bool allowsMisalignedMemoryAccesses(EVT, 875 unsigned AddrSpace = 0, 876 unsigned Align = 1, 877 bool * /*Fast*/ = nullptr) const { 878 return false; 879 } 880 881 /// Returns the target specific optimal type for load and store operations as 882 /// a result of memset, memcpy, and memmove lowering. 883 /// 884 /// If DstAlign is zero that means it's safe to destination alignment can 885 /// satisfy any constraint. Similarly if SrcAlign is zero it means there isn't 886 /// a need to check it against alignment requirement, probably because the 887 /// source does not need to be loaded. If 'IsMemset' is true, that means it's 888 /// expanding a memset. If 'ZeroMemset' is true, that means it's a memset of 889 /// zero. 'MemcpyStrSrc' indicates whether the memcpy source is constant so it 890 /// does not need to be loaded. It returns EVT::Other if the type should be 891 /// determined using generic target-independent logic. getOptimalMemOpType(uint64_t,unsigned,unsigned,bool,bool,bool,MachineFunction &)892 virtual EVT getOptimalMemOpType(uint64_t /*Size*/, 893 unsigned /*DstAlign*/, unsigned /*SrcAlign*/, 894 bool /*IsMemset*/, 895 bool /*ZeroMemset*/, 896 bool /*MemcpyStrSrc*/, 897 MachineFunction &/*MF*/) const { 898 return MVT::Other; 899 } 900 901 /// Returns true if it's safe to use load / store of the specified type to 902 /// expand memcpy / memset inline. 903 /// 904 /// This is mostly true for all types except for some special cases. For 905 /// example, on X86 targets without SSE2 f64 load / store are done with fldl / 906 /// fstpl which also does type conversion. Note the specified type doesn't 907 /// have to be legal as the hook is used before type legalization. isSafeMemOpType(MVT)908 virtual bool isSafeMemOpType(MVT /*VT*/) const { return true; } 909 910 /// Determine if we should use _setjmp or setjmp to implement llvm.setjmp. usesUnderscoreSetJmp()911 bool usesUnderscoreSetJmp() const { 912 return UseUnderscoreSetJmp; 913 } 914 915 /// Determine if we should use _longjmp or longjmp to implement llvm.longjmp. usesUnderscoreLongJmp()916 bool usesUnderscoreLongJmp() const { 917 return UseUnderscoreLongJmp; 918 } 919 920 /// Return integer threshold on number of blocks to use jump tables rather 921 /// than if sequence. getMinimumJumpTableEntries()922 int getMinimumJumpTableEntries() const { 923 return MinimumJumpTableEntries; 924 } 925 926 /// If a physical register, this specifies the register that 927 /// llvm.savestack/llvm.restorestack should save and restore. getStackPointerRegisterToSaveRestore()928 unsigned getStackPointerRegisterToSaveRestore() const { 929 return StackPointerRegisterToSaveRestore; 930 } 931 932 /// If a physical register, this returns the register that receives the 933 /// exception address on entry to a landing pad. getExceptionPointerRegister()934 unsigned getExceptionPointerRegister() const { 935 return ExceptionPointerRegister; 936 } 937 938 /// If a physical register, this returns the register that receives the 939 /// exception typeid on entry to a landing pad. getExceptionSelectorRegister()940 unsigned getExceptionSelectorRegister() const { 941 return ExceptionSelectorRegister; 942 } 943 944 /// Returns the target's jmp_buf size in bytes (if never set, the default is 945 /// 200) getJumpBufSize()946 unsigned getJumpBufSize() const { 947 return JumpBufSize; 948 } 949 950 /// Returns the target's jmp_buf alignment in bytes (if never set, the default 951 /// is 0) getJumpBufAlignment()952 unsigned getJumpBufAlignment() const { 953 return JumpBufAlignment; 954 } 955 956 /// Return the minimum stack alignment of an argument. getMinStackArgumentAlignment()957 unsigned getMinStackArgumentAlignment() const { 958 return MinStackArgumentAlignment; 959 } 960 961 /// Return the minimum function alignment. getMinFunctionAlignment()962 unsigned getMinFunctionAlignment() const { 963 return MinFunctionAlignment; 964 } 965 966 /// Return the preferred function alignment. getPrefFunctionAlignment()967 unsigned getPrefFunctionAlignment() const { 968 return PrefFunctionAlignment; 969 } 970 971 /// Return the preferred loop alignment. 972 virtual unsigned getPrefLoopAlignment(MachineLoop *ML = nullptr) const { 973 return PrefLoopAlignment; 974 } 975 976 /// Return whether the DAG builder should automatically insert fences and 977 /// reduce ordering for atomics. getInsertFencesForAtomic()978 bool getInsertFencesForAtomic() const { 979 return InsertFencesForAtomic; 980 } 981 982 /// Return true if the target stores stack protector cookies at a fixed offset 983 /// in some non-standard address space, and populates the address space and 984 /// offset as appropriate. getStackCookieLocation(unsigned &,unsigned &)985 virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/, 986 unsigned &/*Offset*/) const { 987 return false; 988 } 989 990 /// Returns true if a cast between SrcAS and DestAS is a noop. isNoopAddrSpaceCast(unsigned SrcAS,unsigned DestAS)991 virtual bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const { 992 return false; 993 } 994 995 /// Return true if the pointer arguments to CI should be aligned by aligning 996 /// the object whose address is being passed. If so then MinSize is set to the 997 /// minimum size the object must be to be aligned and PrefAlign is set to the 998 /// preferred alignment. shouldAlignPointerArgs(CallInst *,unsigned &,unsigned &)999 virtual bool shouldAlignPointerArgs(CallInst * /*CI*/, unsigned & /*MinSize*/, 1000 unsigned & /*PrefAlign*/) const { 1001 return false; 1002 } 1003 1004 //===--------------------------------------------------------------------===// 1005 /// \name Helpers for TargetTransformInfo implementations 1006 /// @{ 1007 1008 /// Get the ISD node that corresponds to the Instruction class opcode. 1009 int InstructionOpcodeToISD(unsigned Opcode) const; 1010 1011 /// Estimate the cost of type-legalization and the legalized type. 1012 std::pair<unsigned, MVT> getTypeLegalizationCost(const DataLayout &DL, 1013 Type *Ty) const; 1014 1015 /// @} 1016 1017 //===--------------------------------------------------------------------===// 1018 /// \name Helpers for atomic expansion. 1019 /// @{ 1020 1021 /// True if AtomicExpandPass should use emitLoadLinked/emitStoreConditional 1022 /// and expand AtomicCmpXchgInst. hasLoadLinkedStoreConditional()1023 virtual bool hasLoadLinkedStoreConditional() const { return false; } 1024 1025 /// Perform a load-linked operation on Addr, returning a "Value *" with the 1026 /// corresponding pointee type. This may entail some non-trivial operations to 1027 /// truncate or reconstruct types that will be illegal in the backend. See 1028 /// ARMISelLowering for an example implementation. emitLoadLinked(IRBuilder<> & Builder,Value * Addr,AtomicOrdering Ord)1029 virtual Value *emitLoadLinked(IRBuilder<> &Builder, Value *Addr, 1030 AtomicOrdering Ord) const { 1031 llvm_unreachable("Load linked unimplemented on this target"); 1032 } 1033 1034 /// Perform a store-conditional operation to Addr. Return the status of the 1035 /// store. This should be 0 if the store succeeded, non-zero otherwise. emitStoreConditional(IRBuilder<> & Builder,Value * Val,Value * Addr,AtomicOrdering Ord)1036 virtual Value *emitStoreConditional(IRBuilder<> &Builder, Value *Val, 1037 Value *Addr, AtomicOrdering Ord) const { 1038 llvm_unreachable("Store conditional unimplemented on this target"); 1039 } 1040 1041 /// Inserts in the IR a target-specific intrinsic specifying a fence. 1042 /// It is called by AtomicExpandPass before expanding an 1043 /// AtomicRMW/AtomicCmpXchg/AtomicStore/AtomicLoad. 1044 /// RMW and CmpXchg set both IsStore and IsLoad to true. 1045 /// This function should either return a nullptr, or a pointer to an IR-level 1046 /// Instruction*. Even complex fence sequences can be represented by a 1047 /// single Instruction* through an intrinsic to be lowered later. 1048 /// Backends with !getInsertFencesForAtomic() should keep a no-op here. 1049 /// Backends should override this method to produce target-specific intrinsic 1050 /// for their fences. 1051 /// FIXME: Please note that the default implementation here in terms of 1052 /// IR-level fences exists for historical/compatibility reasons and is 1053 /// *unsound* ! Fences cannot, in general, be used to restore sequential 1054 /// consistency. For example, consider the following example: 1055 /// atomic<int> x = y = 0; 1056 /// int r1, r2, r3, r4; 1057 /// Thread 0: 1058 /// x.store(1); 1059 /// Thread 1: 1060 /// y.store(1); 1061 /// Thread 2: 1062 /// r1 = x.load(); 1063 /// r2 = y.load(); 1064 /// Thread 3: 1065 /// r3 = y.load(); 1066 /// r4 = x.load(); 1067 /// r1 = r3 = 1 and r2 = r4 = 0 is impossible as long as the accesses are all 1068 /// seq_cst. But if they are lowered to monotonic accesses, no amount of 1069 /// IR-level fences can prevent it. 1070 /// @{ emitLeadingFence(IRBuilder<> & Builder,AtomicOrdering Ord,bool IsStore,bool IsLoad)1071 virtual Instruction *emitLeadingFence(IRBuilder<> &Builder, 1072 AtomicOrdering Ord, bool IsStore, 1073 bool IsLoad) const { 1074 if (!getInsertFencesForAtomic()) 1075 return nullptr; 1076 1077 if (isAtLeastRelease(Ord) && IsStore) 1078 return Builder.CreateFence(Ord); 1079 else 1080 return nullptr; 1081 } 1082 emitTrailingFence(IRBuilder<> & Builder,AtomicOrdering Ord,bool IsStore,bool IsLoad)1083 virtual Instruction *emitTrailingFence(IRBuilder<> &Builder, 1084 AtomicOrdering Ord, bool IsStore, 1085 bool IsLoad) const { 1086 if (!getInsertFencesForAtomic()) 1087 return nullptr; 1088 1089 if (isAtLeastAcquire(Ord)) 1090 return Builder.CreateFence(Ord); 1091 else 1092 return nullptr; 1093 } 1094 /// @} 1095 1096 /// Returns true if the given (atomic) store should be expanded by the 1097 /// IR-level AtomicExpand pass into an "atomic xchg" which ignores its input. shouldExpandAtomicStoreInIR(StoreInst * SI)1098 virtual bool shouldExpandAtomicStoreInIR(StoreInst *SI) const { 1099 return false; 1100 } 1101 1102 /// Returns true if arguments should be sign-extended in lib calls. shouldSignExtendTypeInLibCall(EVT Type,bool IsSigned)1103 virtual bool shouldSignExtendTypeInLibCall(EVT Type, bool IsSigned) const { 1104 return IsSigned; 1105 } 1106 1107 /// Returns true if the given (atomic) load should be expanded by the 1108 /// IR-level AtomicExpand pass into a load-linked instruction 1109 /// (through emitLoadLinked()). shouldExpandAtomicLoadInIR(LoadInst * LI)1110 virtual bool shouldExpandAtomicLoadInIR(LoadInst *LI) const { return false; } 1111 1112 /// Returns how the IR-level AtomicExpand pass should expand the given 1113 /// AtomicRMW, if at all. Default is to never expand. 1114 virtual AtomicRMWExpansionKind shouldExpandAtomicRMWInIR(AtomicRMWInst *)1115 shouldExpandAtomicRMWInIR(AtomicRMWInst *) const { 1116 return AtomicRMWExpansionKind::None; 1117 } 1118 1119 /// On some platforms, an AtomicRMW that never actually modifies the value 1120 /// (such as fetch_add of 0) can be turned into a fence followed by an 1121 /// atomic load. This may sound useless, but it makes it possible for the 1122 /// processor to keep the cacheline shared, dramatically improving 1123 /// performance. And such idempotent RMWs are useful for implementing some 1124 /// kinds of locks, see for example (justification + benchmarks): 1125 /// http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf 1126 /// This method tries doing that transformation, returning the atomic load if 1127 /// it succeeds, and nullptr otherwise. 1128 /// If shouldExpandAtomicLoadInIR returns true on that load, it will undergo 1129 /// another round of expansion. 1130 virtual LoadInst * lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst * RMWI)1131 lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *RMWI) const { 1132 return nullptr; 1133 } 1134 1135 /// Returns true if we should normalize 1136 /// select(N0&N1, X, Y) => select(N0, select(N1, X, Y), Y) and 1137 /// select(N0|N1, X, Y) => select(N0, select(N1, X, Y, Y)) if it is likely 1138 /// that it saves us from materializing N0 and N1 in an integer register. 1139 /// Targets that are able to perform and/or on flags should return false here. shouldNormalizeToSelectSequence(LLVMContext & Context,EVT VT)1140 virtual bool shouldNormalizeToSelectSequence(LLVMContext &Context, 1141 EVT VT) const { 1142 // If a target has multiple condition registers, then it likely has logical 1143 // operations on those registers. 1144 if (hasMultipleConditionRegisters()) 1145 return false; 1146 // Only do the transform if the value won't be split into multiple 1147 // registers. 1148 LegalizeTypeAction Action = getTypeAction(Context, VT); 1149 return Action != TypeExpandInteger && Action != TypeExpandFloat && 1150 Action != TypeSplitVector; 1151 } 1152 1153 //===--------------------------------------------------------------------===// 1154 // TargetLowering Configuration Methods - These methods should be invoked by 1155 // the derived class constructor to configure this object for the target. 1156 // 1157 protected: 1158 /// Specify how the target extends the result of integer and floating point 1159 /// boolean values from i1 to a wider type. See getBooleanContents. setBooleanContents(BooleanContent Ty)1160 void setBooleanContents(BooleanContent Ty) { 1161 BooleanContents = Ty; 1162 BooleanFloatContents = Ty; 1163 } 1164 1165 /// Specify how the target extends the result of integer and floating point 1166 /// boolean values from i1 to a wider type. See getBooleanContents. setBooleanContents(BooleanContent IntTy,BooleanContent FloatTy)1167 void setBooleanContents(BooleanContent IntTy, BooleanContent FloatTy) { 1168 BooleanContents = IntTy; 1169 BooleanFloatContents = FloatTy; 1170 } 1171 1172 /// Specify how the target extends the result of a vector boolean value from a 1173 /// vector of i1 to a wider type. See getBooleanContents. setBooleanVectorContents(BooleanContent Ty)1174 void setBooleanVectorContents(BooleanContent Ty) { 1175 BooleanVectorContents = Ty; 1176 } 1177 1178 /// Specify the target scheduling preference. setSchedulingPreference(Sched::Preference Pref)1179 void setSchedulingPreference(Sched::Preference Pref) { 1180 SchedPreferenceInfo = Pref; 1181 } 1182 1183 /// Indicate whether this target prefers to use _setjmp to implement 1184 /// llvm.setjmp or the version without _. Defaults to false. setUseUnderscoreSetJmp(bool Val)1185 void setUseUnderscoreSetJmp(bool Val) { 1186 UseUnderscoreSetJmp = Val; 1187 } 1188 1189 /// Indicate whether this target prefers to use _longjmp to implement 1190 /// llvm.longjmp or the version without _. Defaults to false. setUseUnderscoreLongJmp(bool Val)1191 void setUseUnderscoreLongJmp(bool Val) { 1192 UseUnderscoreLongJmp = Val; 1193 } 1194 1195 /// Indicate the number of blocks to generate jump tables rather than if 1196 /// sequence. setMinimumJumpTableEntries(int Val)1197 void setMinimumJumpTableEntries(int Val) { 1198 MinimumJumpTableEntries = Val; 1199 } 1200 1201 /// If set to a physical register, this specifies the register that 1202 /// llvm.savestack/llvm.restorestack should save and restore. setStackPointerRegisterToSaveRestore(unsigned R)1203 void setStackPointerRegisterToSaveRestore(unsigned R) { 1204 StackPointerRegisterToSaveRestore = R; 1205 } 1206 1207 /// If set to a physical register, this sets the register that receives the 1208 /// exception address on entry to a landing pad. setExceptionPointerRegister(unsigned R)1209 void setExceptionPointerRegister(unsigned R) { 1210 ExceptionPointerRegister = R; 1211 } 1212 1213 /// If set to a physical register, this sets the register that receives the 1214 /// exception typeid on entry to a landing pad. setExceptionSelectorRegister(unsigned R)1215 void setExceptionSelectorRegister(unsigned R) { 1216 ExceptionSelectorRegister = R; 1217 } 1218 1219 /// Tells the code generator not to expand operations into sequences that use 1220 /// the select operations if possible. 1221 void setSelectIsExpensive(bool isExpensive = true) { 1222 SelectIsExpensive = isExpensive; 1223 } 1224 1225 /// Tells the code generator that the target has multiple (allocatable) 1226 /// condition registers that can be used to store the results of comparisons 1227 /// for use by selects and conditional branches. With multiple condition 1228 /// registers, the code generator will not aggressively sink comparisons into 1229 /// the blocks of their users. 1230 void setHasMultipleConditionRegisters(bool hasManyRegs = true) { 1231 HasMultipleConditionRegisters = hasManyRegs; 1232 } 1233 1234 /// Tells the code generator that the target has BitExtract instructions. 1235 /// The code generator will aggressively sink "shift"s into the blocks of 1236 /// their users if the users will generate "and" instructions which can be 1237 /// combined with "shift" to BitExtract instructions. 1238 void setHasExtractBitsInsn(bool hasExtractInsn = true) { 1239 HasExtractBitsInsn = hasExtractInsn; 1240 } 1241 1242 /// Tells the code generator not to expand logic operations on comparison 1243 /// predicates into separate sequences that increase the amount of flow 1244 /// control. 1245 void setJumpIsExpensive(bool isExpensive = true); 1246 1247 /// Tells the code generator that integer divide is expensive, and if 1248 /// possible, should be replaced by an alternate sequence of instructions not 1249 /// containing an integer divide. 1250 void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; } 1251 1252 /// Tells the code generator that fsqrt is cheap, and should not be replaced 1253 /// with an alternative sequence of instructions. 1254 void setFsqrtIsCheap(bool isCheap = true) { FsqrtIsCheap = isCheap; } 1255 1256 /// Tells the code generator that this target supports floating point 1257 /// exceptions and cares about preserving floating point exception behavior. 1258 void setHasFloatingPointExceptions(bool FPExceptions = true) { 1259 HasFloatingPointExceptions = FPExceptions; 1260 } 1261 1262 /// Tells the code generator which bitwidths to bypass. addBypassSlowDiv(unsigned int SlowBitWidth,unsigned int FastBitWidth)1263 void addBypassSlowDiv(unsigned int SlowBitWidth, unsigned int FastBitWidth) { 1264 BypassSlowDivWidths[SlowBitWidth] = FastBitWidth; 1265 } 1266 1267 /// Tells the code generator that it shouldn't generate sra/srl/add/sra for a 1268 /// signed divide by power of two; let the target handle it. 1269 void setPow2SDivIsCheap(bool isCheap = true) { Pow2SDivIsCheap = isCheap; } 1270 1271 /// Add the specified register class as an available regclass for the 1272 /// specified value type. This indicates the selector can handle values of 1273 /// that class natively. addRegisterClass(MVT VT,const TargetRegisterClass * RC)1274 void addRegisterClass(MVT VT, const TargetRegisterClass *RC) { 1275 assert((unsigned)VT.SimpleTy < array_lengthof(RegClassForVT)); 1276 AvailableRegClasses.push_back(std::make_pair(VT, RC)); 1277 RegClassForVT[VT.SimpleTy] = RC; 1278 } 1279 1280 /// Remove all register classes. clearRegisterClasses()1281 void clearRegisterClasses() { 1282 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE * sizeof(TargetRegisterClass*)); 1283 1284 AvailableRegClasses.clear(); 1285 } 1286 1287 /// \brief Remove all operation actions. clearOperationActions()1288 void clearOperationActions() { 1289 } 1290 1291 /// Return the largest legal super-reg register class of the register class 1292 /// for the specified type and its associated "cost". 1293 virtual std::pair<const TargetRegisterClass *, uint8_t> 1294 findRepresentativeClass(const TargetRegisterInfo *TRI, MVT VT) const; 1295 1296 /// Once all of the register classes are added, this allows us to compute 1297 /// derived properties we expose. 1298 void computeRegisterProperties(const TargetRegisterInfo *TRI); 1299 1300 /// Indicate that the specified operation does not work with the specified 1301 /// type and indicate what to do about it. setOperationAction(unsigned Op,MVT VT,LegalizeAction Action)1302 void setOperationAction(unsigned Op, MVT VT, 1303 LegalizeAction Action) { 1304 assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!"); 1305 OpActions[(unsigned)VT.SimpleTy][Op] = (uint8_t)Action; 1306 } 1307 1308 /// Indicate that the specified load with extension does not work with the 1309 /// specified type and indicate what to do about it. setLoadExtAction(unsigned ExtType,MVT ValVT,MVT MemVT,LegalizeAction Action)1310 void setLoadExtAction(unsigned ExtType, MVT ValVT, MVT MemVT, 1311 LegalizeAction Action) { 1312 assert(ExtType < ISD::LAST_LOADEXT_TYPE && ValVT.isValid() && 1313 MemVT.isValid() && "Table isn't big enough!"); 1314 LoadExtActions[ValVT.SimpleTy][MemVT.SimpleTy][ExtType] = (uint8_t)Action; 1315 } 1316 1317 /// Indicate that the specified truncating store does not work with the 1318 /// specified type and indicate what to do about it. setTruncStoreAction(MVT ValVT,MVT MemVT,LegalizeAction Action)1319 void setTruncStoreAction(MVT ValVT, MVT MemVT, 1320 LegalizeAction Action) { 1321 assert(ValVT.isValid() && MemVT.isValid() && "Table isn't big enough!"); 1322 TruncStoreActions[ValVT.SimpleTy][MemVT.SimpleTy] = (uint8_t)Action; 1323 } 1324 1325 /// Indicate that the specified indexed load does or does not work with the 1326 /// specified type and indicate what to do abort it. 1327 /// 1328 /// NOTE: All indexed mode loads are initialized to Expand in 1329 /// TargetLowering.cpp setIndexedLoadAction(unsigned IdxMode,MVT VT,LegalizeAction Action)1330 void setIndexedLoadAction(unsigned IdxMode, MVT VT, 1331 LegalizeAction Action) { 1332 assert(VT.isValid() && IdxMode < ISD::LAST_INDEXED_MODE && 1333 (unsigned)Action < 0xf && "Table isn't big enough!"); 1334 // Load action are kept in the upper half. 1335 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0xf0; 1336 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action) <<4; 1337 } 1338 1339 /// Indicate that the specified indexed store does or does not work with the 1340 /// specified type and indicate what to do about it. 1341 /// 1342 /// NOTE: All indexed mode stores are initialized to Expand in 1343 /// TargetLowering.cpp setIndexedStoreAction(unsigned IdxMode,MVT VT,LegalizeAction Action)1344 void setIndexedStoreAction(unsigned IdxMode, MVT VT, 1345 LegalizeAction Action) { 1346 assert(VT.isValid() && IdxMode < ISD::LAST_INDEXED_MODE && 1347 (unsigned)Action < 0xf && "Table isn't big enough!"); 1348 // Store action are kept in the lower half. 1349 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0x0f; 1350 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action); 1351 } 1352 1353 /// Indicate that the specified condition code is or isn't supported on the 1354 /// target and indicate what to do about it. setCondCodeAction(ISD::CondCode CC,MVT VT,LegalizeAction Action)1355 void setCondCodeAction(ISD::CondCode CC, MVT VT, 1356 LegalizeAction Action) { 1357 assert(VT.isValid() && (unsigned)CC < array_lengthof(CondCodeActions) && 1358 "Table isn't big enough!"); 1359 /// The lower 5 bits of the SimpleTy index into Nth 2bit set from the 32-bit 1360 /// value and the upper 27 bits index into the second dimension of the array 1361 /// to select what 32-bit value to use. 1362 uint32_t Shift = 2 * (VT.SimpleTy & 0xF); 1363 CondCodeActions[CC][VT.SimpleTy >> 4] &= ~((uint32_t)0x3 << Shift); 1364 CondCodeActions[CC][VT.SimpleTy >> 4] |= (uint32_t)Action << Shift; 1365 } 1366 1367 /// If Opc/OrigVT is specified as being promoted, the promotion code defaults 1368 /// to trying a larger integer/fp until it can find one that works. If that 1369 /// default is insufficient, this method can be used by the target to override 1370 /// the default. AddPromotedToType(unsigned Opc,MVT OrigVT,MVT DestVT)1371 void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) { 1372 PromoteToType[std::make_pair(Opc, OrigVT.SimpleTy)] = DestVT.SimpleTy; 1373 } 1374 1375 /// Targets should invoke this method for each target independent node that 1376 /// they want to provide a custom DAG combiner for by implementing the 1377 /// PerformDAGCombine virtual method. setTargetDAGCombine(ISD::NodeType NT)1378 void setTargetDAGCombine(ISD::NodeType NT) { 1379 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 1380 TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7); 1381 } 1382 1383 /// Set the target's required jmp_buf buffer size (in bytes); default is 200 setJumpBufSize(unsigned Size)1384 void setJumpBufSize(unsigned Size) { 1385 JumpBufSize = Size; 1386 } 1387 1388 /// Set the target's required jmp_buf buffer alignment (in bytes); default is 1389 /// 0 setJumpBufAlignment(unsigned Align)1390 void setJumpBufAlignment(unsigned Align) { 1391 JumpBufAlignment = Align; 1392 } 1393 1394 /// Set the target's minimum function alignment (in log2(bytes)) setMinFunctionAlignment(unsigned Align)1395 void setMinFunctionAlignment(unsigned Align) { 1396 MinFunctionAlignment = Align; 1397 } 1398 1399 /// Set the target's preferred function alignment. This should be set if 1400 /// there is a performance benefit to higher-than-minimum alignment (in 1401 /// log2(bytes)) setPrefFunctionAlignment(unsigned Align)1402 void setPrefFunctionAlignment(unsigned Align) { 1403 PrefFunctionAlignment = Align; 1404 } 1405 1406 /// Set the target's preferred loop alignment. Default alignment is zero, it 1407 /// means the target does not care about loop alignment. The alignment is 1408 /// specified in log2(bytes). The target may also override 1409 /// getPrefLoopAlignment to provide per-loop values. setPrefLoopAlignment(unsigned Align)1410 void setPrefLoopAlignment(unsigned Align) { 1411 PrefLoopAlignment = Align; 1412 } 1413 1414 /// Set the minimum stack alignment of an argument (in log2(bytes)). setMinStackArgumentAlignment(unsigned Align)1415 void setMinStackArgumentAlignment(unsigned Align) { 1416 MinStackArgumentAlignment = Align; 1417 } 1418 1419 /// Set if the DAG builder should automatically insert fences and reduce the 1420 /// order of atomic memory operations to Monotonic. setInsertFencesForAtomic(bool fence)1421 void setInsertFencesForAtomic(bool fence) { 1422 InsertFencesForAtomic = fence; 1423 } 1424 1425 public: 1426 //===--------------------------------------------------------------------===// 1427 // Addressing mode description hooks (used by LSR etc). 1428 // 1429 1430 /// CodeGenPrepare sinks address calculations into the same BB as Load/Store 1431 /// instructions reading the address. This allows as much computation as 1432 /// possible to be done in the address mode for that operand. This hook lets 1433 /// targets also pass back when this should be done on intrinsics which 1434 /// load/store. 1435 virtual bool GetAddrModeArguments(IntrinsicInst * /*I*/, 1436 SmallVectorImpl<Value*> &/*Ops*/, 1437 Type *&/*AccessTy*/, 1438 unsigned AddrSpace = 0) const { 1439 return false; 1440 } 1441 1442 /// This represents an addressing mode of: 1443 /// BaseGV + BaseOffs + BaseReg + Scale*ScaleReg 1444 /// If BaseGV is null, there is no BaseGV. 1445 /// If BaseOffs is zero, there is no base offset. 1446 /// If HasBaseReg is false, there is no base register. 1447 /// If Scale is zero, there is no ScaleReg. Scale of 1 indicates a reg with 1448 /// no scale. 1449 struct AddrMode { 1450 GlobalValue *BaseGV; 1451 int64_t BaseOffs; 1452 bool HasBaseReg; 1453 int64_t Scale; AddrModeAddrMode1454 AddrMode() : BaseGV(nullptr), BaseOffs(0), HasBaseReg(false), Scale(0) {} 1455 }; 1456 1457 /// Return true if the addressing mode represented by AM is legal for this 1458 /// target, for a load/store of the specified type. 1459 /// 1460 /// The type may be VoidTy, in which case only return true if the addressing 1461 /// mode is legal for a load/store of any legal type. TODO: Handle 1462 /// pre/postinc as well. 1463 /// 1464 /// If the address space cannot be determined, it will be -1. 1465 /// 1466 /// TODO: Remove default argument 1467 virtual bool isLegalAddressingMode(const DataLayout &DL, const AddrMode &AM, 1468 Type *Ty, unsigned AddrSpace) const; 1469 1470 /// \brief Return the cost of the scaling factor used in the addressing mode 1471 /// represented by AM for this target, for a load/store of the specified type. 1472 /// 1473 /// If the AM is supported, the return value must be >= 0. 1474 /// If the AM is not supported, it returns a negative value. 1475 /// TODO: Handle pre/postinc as well. 1476 /// TODO: Remove default argument 1477 virtual int getScalingFactorCost(const DataLayout &DL, const AddrMode &AM, 1478 Type *Ty, unsigned AS = 0) const { 1479 // Default: assume that any scaling factor used in a legal AM is free. 1480 if (isLegalAddressingMode(DL, AM, Ty, AS)) 1481 return 0; 1482 return -1; 1483 } 1484 1485 /// Return true if the specified immediate is legal icmp immediate, that is 1486 /// the target has icmp instructions which can compare a register against the 1487 /// immediate without having to materialize the immediate into a register. isLegalICmpImmediate(int64_t)1488 virtual bool isLegalICmpImmediate(int64_t) const { 1489 return true; 1490 } 1491 1492 /// Return true if the specified immediate is legal add immediate, that is the 1493 /// target has add instructions which can add a register with the immediate 1494 /// without having to materialize the immediate into a register. isLegalAddImmediate(int64_t)1495 virtual bool isLegalAddImmediate(int64_t) const { 1496 return true; 1497 } 1498 1499 /// Return true if it's significantly cheaper to shift a vector by a uniform 1500 /// scalar than by an amount which will vary across each lane. On x86, for 1501 /// example, there is a "psllw" instruction for the former case, but no simple 1502 /// instruction for a general "a << b" operation on vectors. isVectorShiftByScalarCheap(Type * Ty)1503 virtual bool isVectorShiftByScalarCheap(Type *Ty) const { 1504 return false; 1505 } 1506 1507 /// Return true if it's free to truncate a value of type Ty1 to type 1508 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16 1509 /// by referencing its sub-register AX. isTruncateFree(Type *,Type *)1510 virtual bool isTruncateFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 1511 return false; 1512 } 1513 1514 /// Return true if a truncation from Ty1 to Ty2 is permitted when deciding 1515 /// whether a call is in tail position. Typically this means that both results 1516 /// would be assigned to the same register or stack slot, but it could mean 1517 /// the target performs adequate checks of its own before proceeding with the 1518 /// tail call. allowTruncateForTailCall(Type *,Type *)1519 virtual bool allowTruncateForTailCall(Type * /*Ty1*/, Type * /*Ty2*/) const { 1520 return false; 1521 } 1522 isTruncateFree(EVT,EVT)1523 virtual bool isTruncateFree(EVT /*VT1*/, EVT /*VT2*/) const { 1524 return false; 1525 } 1526 isProfitableToHoist(Instruction * I)1527 virtual bool isProfitableToHoist(Instruction *I) const { return true; } 1528 1529 /// Return true if the extension represented by \p I is free. 1530 /// Unlikely the is[Z|FP]ExtFree family which is based on types, 1531 /// this method can use the context provided by \p I to decide 1532 /// whether or not \p I is free. 1533 /// This method extends the behavior of the is[Z|FP]ExtFree family. 1534 /// In other words, if is[Z|FP]Free returns true, then this method 1535 /// returns true as well. The converse is not true. 1536 /// The target can perform the adequate checks by overriding isExtFreeImpl. 1537 /// \pre \p I must be a sign, zero, or fp extension. isExtFree(const Instruction * I)1538 bool isExtFree(const Instruction *I) const { 1539 switch (I->getOpcode()) { 1540 case Instruction::FPExt: 1541 if (isFPExtFree(EVT::getEVT(I->getType()))) 1542 return true; 1543 break; 1544 case Instruction::ZExt: 1545 if (isZExtFree(I->getOperand(0)->getType(), I->getType())) 1546 return true; 1547 break; 1548 case Instruction::SExt: 1549 break; 1550 default: 1551 llvm_unreachable("Instruction is not an extension"); 1552 } 1553 return isExtFreeImpl(I); 1554 } 1555 1556 /// Return true if any actual instruction that defines a value of type Ty1 1557 /// implicitly zero-extends the value to Ty2 in the result register. 1558 /// 1559 /// This does not necessarily include registers defined in unknown ways, such 1560 /// as incoming arguments, or copies from unknown virtual registers. Also, if 1561 /// isTruncateFree(Ty2, Ty1) is true, this does not necessarily apply to 1562 /// truncate instructions. e.g. on x86-64, all instructions that define 32-bit 1563 /// values implicit zero-extend the result out to 64 bits. isZExtFree(Type *,Type *)1564 virtual bool isZExtFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 1565 return false; 1566 } 1567 isZExtFree(EVT,EVT)1568 virtual bool isZExtFree(EVT /*VT1*/, EVT /*VT2*/) const { 1569 return false; 1570 } 1571 1572 /// Return true if the target supplies and combines to a paired load 1573 /// two loaded values of type LoadedType next to each other in memory. 1574 /// RequiredAlignment gives the minimal alignment constraints that must be met 1575 /// to be able to select this paired load. 1576 /// 1577 /// This information is *not* used to generate actual paired loads, but it is 1578 /// used to generate a sequence of loads that is easier to combine into a 1579 /// paired load. 1580 /// For instance, something like this: 1581 /// a = load i64* addr 1582 /// b = trunc i64 a to i32 1583 /// c = lshr i64 a, 32 1584 /// d = trunc i64 c to i32 1585 /// will be optimized into: 1586 /// b = load i32* addr1 1587 /// d = load i32* addr2 1588 /// Where addr1 = addr2 +/- sizeof(i32). 1589 /// 1590 /// In other words, unless the target performs a post-isel load combining, 1591 /// this information should not be provided because it will generate more 1592 /// loads. hasPairedLoad(Type *,unsigned &)1593 virtual bool hasPairedLoad(Type * /*LoadedType*/, 1594 unsigned & /*RequiredAligment*/) const { 1595 return false; 1596 } 1597 hasPairedLoad(EVT,unsigned &)1598 virtual bool hasPairedLoad(EVT /*LoadedType*/, 1599 unsigned & /*RequiredAligment*/) const { 1600 return false; 1601 } 1602 1603 /// \brief Get the maximum supported factor for interleaved memory accesses. 1604 /// Default to be the minimum interleave factor: 2. getMaxSupportedInterleaveFactor()1605 virtual unsigned getMaxSupportedInterleaveFactor() const { return 2; } 1606 1607 /// \brief Lower an interleaved load to target specific intrinsics. Return 1608 /// true on success. 1609 /// 1610 /// \p LI is the vector load instruction. 1611 /// \p Shuffles is the shufflevector list to DE-interleave the loaded vector. 1612 /// \p Indices is the corresponding indices for each shufflevector. 1613 /// \p Factor is the interleave factor. lowerInterleavedLoad(LoadInst * LI,ArrayRef<ShuffleVectorInst * > Shuffles,ArrayRef<unsigned> Indices,unsigned Factor)1614 virtual bool lowerInterleavedLoad(LoadInst *LI, 1615 ArrayRef<ShuffleVectorInst *> Shuffles, 1616 ArrayRef<unsigned> Indices, 1617 unsigned Factor) const { 1618 return false; 1619 } 1620 1621 /// \brief Lower an interleaved store to target specific intrinsics. Return 1622 /// true on success. 1623 /// 1624 /// \p SI is the vector store instruction. 1625 /// \p SVI is the shufflevector to RE-interleave the stored vector. 1626 /// \p Factor is the interleave factor. lowerInterleavedStore(StoreInst * SI,ShuffleVectorInst * SVI,unsigned Factor)1627 virtual bool lowerInterleavedStore(StoreInst *SI, ShuffleVectorInst *SVI, 1628 unsigned Factor) const { 1629 return false; 1630 } 1631 1632 /// Return true if zero-extending the specific node Val to type VT2 is free 1633 /// (either because it's implicitly zero-extended such as ARM ldrb / ldrh or 1634 /// because it's folded such as X86 zero-extending loads). isZExtFree(SDValue Val,EVT VT2)1635 virtual bool isZExtFree(SDValue Val, EVT VT2) const { 1636 return isZExtFree(Val.getValueType(), VT2); 1637 } 1638 1639 /// Return true if an fpext operation is free (for instance, because 1640 /// single-precision floating-point numbers are implicitly extended to 1641 /// double-precision). isFPExtFree(EVT VT)1642 virtual bool isFPExtFree(EVT VT) const { 1643 assert(VT.isFloatingPoint()); 1644 return false; 1645 } 1646 1647 /// Return true if folding a vector load into ExtVal (a sign, zero, or any 1648 /// extend node) is profitable. isVectorLoadExtDesirable(SDValue ExtVal)1649 virtual bool isVectorLoadExtDesirable(SDValue ExtVal) const { return false; } 1650 1651 /// Return true if an fneg operation is free to the point where it is never 1652 /// worthwhile to replace it with a bitwise operation. isFNegFree(EVT VT)1653 virtual bool isFNegFree(EVT VT) const { 1654 assert(VT.isFloatingPoint()); 1655 return false; 1656 } 1657 1658 /// Return true if an fabs operation is free to the point where it is never 1659 /// worthwhile to replace it with a bitwise operation. isFAbsFree(EVT VT)1660 virtual bool isFAbsFree(EVT VT) const { 1661 assert(VT.isFloatingPoint()); 1662 return false; 1663 } 1664 1665 /// Return true if an FMA operation is faster than a pair of fmul and fadd 1666 /// instructions. fmuladd intrinsics will be expanded to FMAs when this method 1667 /// returns true, otherwise fmuladd is expanded to fmul + fadd. 1668 /// 1669 /// NOTE: This may be called before legalization on types for which FMAs are 1670 /// not legal, but should return true if those types will eventually legalize 1671 /// to types that support FMAs. After legalization, it will only be called on 1672 /// types that support FMAs (via Legal or Custom actions) isFMAFasterThanFMulAndFAdd(EVT)1673 virtual bool isFMAFasterThanFMulAndFAdd(EVT) const { 1674 return false; 1675 } 1676 1677 /// Return true if it's profitable to narrow operations of type VT1 to 1678 /// VT2. e.g. on x86, it's profitable to narrow from i32 to i8 but not from 1679 /// i32 to i16. isNarrowingProfitable(EVT,EVT)1680 virtual bool isNarrowingProfitable(EVT /*VT1*/, EVT /*VT2*/) const { 1681 return false; 1682 } 1683 1684 /// \brief Return true if it is beneficial to convert a load of a constant to 1685 /// just the constant itself. 1686 /// On some targets it might be more efficient to use a combination of 1687 /// arithmetic instructions to materialize the constant instead of loading it 1688 /// from a constant pool. shouldConvertConstantLoadToIntImm(const APInt & Imm,Type * Ty)1689 virtual bool shouldConvertConstantLoadToIntImm(const APInt &Imm, 1690 Type *Ty) const { 1691 return false; 1692 } 1693 1694 /// Return true if EXTRACT_SUBVECTOR is cheap for this result type 1695 /// with this index. This is needed because EXTRACT_SUBVECTOR usually 1696 /// has custom lowering that depends on the index of the first element, 1697 /// and only the target knows which lowering is cheap. isExtractSubvectorCheap(EVT ResVT,unsigned Index)1698 virtual bool isExtractSubvectorCheap(EVT ResVT, unsigned Index) const { 1699 return false; 1700 } 1701 1702 //===--------------------------------------------------------------------===// 1703 // Runtime Library hooks 1704 // 1705 1706 /// Rename the default libcall routine name for the specified libcall. setLibcallName(RTLIB::Libcall Call,const char * Name)1707 void setLibcallName(RTLIB::Libcall Call, const char *Name) { 1708 LibcallRoutineNames[Call] = Name; 1709 } 1710 1711 /// Get the libcall routine name for the specified libcall. getLibcallName(RTLIB::Libcall Call)1712 const char *getLibcallName(RTLIB::Libcall Call) const { 1713 return LibcallRoutineNames[Call]; 1714 } 1715 1716 /// Override the default CondCode to be used to test the result of the 1717 /// comparison libcall against zero. setCmpLibcallCC(RTLIB::Libcall Call,ISD::CondCode CC)1718 void setCmpLibcallCC(RTLIB::Libcall Call, ISD::CondCode CC) { 1719 CmpLibcallCCs[Call] = CC; 1720 } 1721 1722 /// Get the CondCode that's to be used to test the result of the comparison 1723 /// libcall against zero. getCmpLibcallCC(RTLIB::Libcall Call)1724 ISD::CondCode getCmpLibcallCC(RTLIB::Libcall Call) const { 1725 return CmpLibcallCCs[Call]; 1726 } 1727 1728 /// Set the CallingConv that should be used for the specified libcall. setLibcallCallingConv(RTLIB::Libcall Call,CallingConv::ID CC)1729 void setLibcallCallingConv(RTLIB::Libcall Call, CallingConv::ID CC) { 1730 LibcallCallingConvs[Call] = CC; 1731 } 1732 1733 /// Get the CallingConv that should be used for the specified libcall. getLibcallCallingConv(RTLIB::Libcall Call)1734 CallingConv::ID getLibcallCallingConv(RTLIB::Libcall Call) const { 1735 return LibcallCallingConvs[Call]; 1736 } 1737 1738 private: 1739 const TargetMachine &TM; 1740 1741 /// Tells the code generator not to expand operations into sequences that use 1742 /// the select operations if possible. 1743 bool SelectIsExpensive; 1744 1745 /// Tells the code generator that the target has multiple (allocatable) 1746 /// condition registers that can be used to store the results of comparisons 1747 /// for use by selects and conditional branches. With multiple condition 1748 /// registers, the code generator will not aggressively sink comparisons into 1749 /// the blocks of their users. 1750 bool HasMultipleConditionRegisters; 1751 1752 /// Tells the code generator that the target has BitExtract instructions. 1753 /// The code generator will aggressively sink "shift"s into the blocks of 1754 /// their users if the users will generate "and" instructions which can be 1755 /// combined with "shift" to BitExtract instructions. 1756 bool HasExtractBitsInsn; 1757 1758 /// Tells the code generator not to expand integer divides by constants into a 1759 /// sequence of muls, adds, and shifts. This is a hack until a real cost 1760 /// model is in place. If we ever optimize for size, this will be set to true 1761 /// unconditionally. 1762 bool IntDivIsCheap; 1763 1764 // Don't expand fsqrt with an approximation based on the inverse sqrt. 1765 bool FsqrtIsCheap; 1766 1767 /// Tells the code generator to bypass slow divide or remainder 1768 /// instructions. For example, BypassSlowDivWidths[32,8] tells the code 1769 /// generator to bypass 32-bit integer div/rem with an 8-bit unsigned integer 1770 /// div/rem when the operands are positive and less than 256. 1771 DenseMap <unsigned int, unsigned int> BypassSlowDivWidths; 1772 1773 /// Tells the code generator that it shouldn't generate sra/srl/add/sra for a 1774 /// signed divide by power of two; let the target handle it. 1775 bool Pow2SDivIsCheap; 1776 1777 /// Tells the code generator that it shouldn't generate extra flow control 1778 /// instructions and should attempt to combine flow control instructions via 1779 /// predication. 1780 bool JumpIsExpensive; 1781 1782 /// Whether the target supports or cares about preserving floating point 1783 /// exception behavior. 1784 bool HasFloatingPointExceptions; 1785 1786 /// This target prefers to use _setjmp to implement llvm.setjmp. 1787 /// 1788 /// Defaults to false. 1789 bool UseUnderscoreSetJmp; 1790 1791 /// This target prefers to use _longjmp to implement llvm.longjmp. 1792 /// 1793 /// Defaults to false. 1794 bool UseUnderscoreLongJmp; 1795 1796 /// Number of blocks threshold to use jump tables. 1797 int MinimumJumpTableEntries; 1798 1799 /// Information about the contents of the high-bits in boolean values held in 1800 /// a type wider than i1. See getBooleanContents. 1801 BooleanContent BooleanContents; 1802 1803 /// Information about the contents of the high-bits in boolean values held in 1804 /// a type wider than i1. See getBooleanContents. 1805 BooleanContent BooleanFloatContents; 1806 1807 /// Information about the contents of the high-bits in boolean vector values 1808 /// when the element type is wider than i1. See getBooleanContents. 1809 BooleanContent BooleanVectorContents; 1810 1811 /// The target scheduling preference: shortest possible total cycles or lowest 1812 /// register usage. 1813 Sched::Preference SchedPreferenceInfo; 1814 1815 /// The size, in bytes, of the target's jmp_buf buffers 1816 unsigned JumpBufSize; 1817 1818 /// The alignment, in bytes, of the target's jmp_buf buffers 1819 unsigned JumpBufAlignment; 1820 1821 /// The minimum alignment that any argument on the stack needs to have. 1822 unsigned MinStackArgumentAlignment; 1823 1824 /// The minimum function alignment (used when optimizing for size, and to 1825 /// prevent explicitly provided alignment from leading to incorrect code). 1826 unsigned MinFunctionAlignment; 1827 1828 /// The preferred function alignment (used when alignment unspecified and 1829 /// optimizing for speed). 1830 unsigned PrefFunctionAlignment; 1831 1832 /// The preferred loop alignment. 1833 unsigned PrefLoopAlignment; 1834 1835 /// Whether the DAG builder should automatically insert fences and reduce 1836 /// ordering for atomics. (This will be set for for most architectures with 1837 /// weak memory ordering.) 1838 bool InsertFencesForAtomic; 1839 1840 /// If set to a physical register, this specifies the register that 1841 /// llvm.savestack/llvm.restorestack should save and restore. 1842 unsigned StackPointerRegisterToSaveRestore; 1843 1844 /// If set to a physical register, this specifies the register that receives 1845 /// the exception address on entry to a landing pad. 1846 unsigned ExceptionPointerRegister; 1847 1848 /// If set to a physical register, this specifies the register that receives 1849 /// the exception typeid on entry to a landing pad. 1850 unsigned ExceptionSelectorRegister; 1851 1852 /// This indicates the default register class to use for each ValueType the 1853 /// target supports natively. 1854 const TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE]; 1855 unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE]; 1856 MVT RegisterTypeForVT[MVT::LAST_VALUETYPE]; 1857 1858 /// This indicates the "representative" register class to use for each 1859 /// ValueType the target supports natively. This information is used by the 1860 /// scheduler to track register pressure. By default, the representative 1861 /// register class is the largest legal super-reg register class of the 1862 /// register class of the specified type. e.g. On x86, i8, i16, and i32's 1863 /// representative class would be GR32. 1864 const TargetRegisterClass *RepRegClassForVT[MVT::LAST_VALUETYPE]; 1865 1866 /// This indicates the "cost" of the "representative" register class for each 1867 /// ValueType. The cost is used by the scheduler to approximate register 1868 /// pressure. 1869 uint8_t RepRegClassCostForVT[MVT::LAST_VALUETYPE]; 1870 1871 /// For any value types we are promoting or expanding, this contains the value 1872 /// type that we are changing to. For Expanded types, this contains one step 1873 /// of the expand (e.g. i64 -> i32), even if there are multiple steps required 1874 /// (e.g. i64 -> i16). For types natively supported by the system, this holds 1875 /// the same type (e.g. i32 -> i32). 1876 MVT TransformToType[MVT::LAST_VALUETYPE]; 1877 1878 /// For each operation and each value type, keep a LegalizeAction that 1879 /// indicates how instruction selection should deal with the operation. Most 1880 /// operations are Legal (aka, supported natively by the target), but 1881 /// operations that are not should be described. Note that operations on 1882 /// non-legal value types are not described here. 1883 uint8_t OpActions[MVT::LAST_VALUETYPE][ISD::BUILTIN_OP_END]; 1884 1885 /// For each load extension type and each value type, keep a LegalizeAction 1886 /// that indicates how instruction selection should deal with a load of a 1887 /// specific value type and extension type. 1888 uint8_t LoadExtActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE] 1889 [ISD::LAST_LOADEXT_TYPE]; 1890 1891 /// For each value type pair keep a LegalizeAction that indicates whether a 1892 /// truncating store of a specific value type and truncating type is legal. 1893 uint8_t TruncStoreActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE]; 1894 1895 /// For each indexed mode and each value type, keep a pair of LegalizeAction 1896 /// that indicates how instruction selection should deal with the load / 1897 /// store. 1898 /// 1899 /// The first dimension is the value_type for the reference. The second 1900 /// dimension represents the various modes for load store. 1901 uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][ISD::LAST_INDEXED_MODE]; 1902 1903 /// For each condition code (ISD::CondCode) keep a LegalizeAction that 1904 /// indicates how instruction selection should deal with the condition code. 1905 /// 1906 /// Because each CC action takes up 2 bits, we need to have the array size be 1907 /// large enough to fit all of the value types. This can be done by rounding 1908 /// up the MVT::LAST_VALUETYPE value to the next multiple of 16. 1909 uint32_t CondCodeActions[ISD::SETCC_INVALID][(MVT::LAST_VALUETYPE + 15) / 16]; 1910 1911 ValueTypeActionImpl ValueTypeActions; 1912 1913 private: 1914 LegalizeKind getTypeConversion(LLVMContext &Context, EVT VT) const; 1915 1916 private: 1917 std::vector<std::pair<MVT, const TargetRegisterClass*> > AvailableRegClasses; 1918 1919 /// Targets can specify ISD nodes that they would like PerformDAGCombine 1920 /// callbacks for by calling setTargetDAGCombine(), which sets a bit in this 1921 /// array. 1922 unsigned char 1923 TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT]; 1924 1925 /// For operations that must be promoted to a specific type, this holds the 1926 /// destination type. This map should be sparse, so don't hold it as an 1927 /// array. 1928 /// 1929 /// Targets add entries to this map with AddPromotedToType(..), clients access 1930 /// this with getTypeToPromoteTo(..). 1931 std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType> 1932 PromoteToType; 1933 1934 /// Stores the name each libcall. 1935 const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL]; 1936 1937 /// The ISD::CondCode that should be used to test the result of each of the 1938 /// comparison libcall against zero. 1939 ISD::CondCode CmpLibcallCCs[RTLIB::UNKNOWN_LIBCALL]; 1940 1941 /// Stores the CallingConv that should be used for each libcall. 1942 CallingConv::ID LibcallCallingConvs[RTLIB::UNKNOWN_LIBCALL]; 1943 1944 protected: 1945 /// Return true if the extension represented by \p I is free. 1946 /// \pre \p I is a sign, zero, or fp extension and 1947 /// is[Z|FP]ExtFree of the related types is not true. isExtFreeImpl(const Instruction * I)1948 virtual bool isExtFreeImpl(const Instruction *I) const { return false; } 1949 1950 /// \brief Specify maximum number of store instructions per memset call. 1951 /// 1952 /// When lowering \@llvm.memset this field specifies the maximum number of 1953 /// store operations that may be substituted for the call to memset. Targets 1954 /// must set this value based on the cost threshold for that target. Targets 1955 /// should assume that the memset will be done using as many of the largest 1956 /// store operations first, followed by smaller ones, if necessary, per 1957 /// alignment restrictions. For example, storing 9 bytes on a 32-bit machine 1958 /// with 16-bit alignment would result in four 2-byte stores and one 1-byte 1959 /// store. This only applies to setting a constant array of a constant size. 1960 unsigned MaxStoresPerMemset; 1961 1962 /// Maximum number of stores operations that may be substituted for the call 1963 /// to memset, used for functions with OptSize attribute. 1964 unsigned MaxStoresPerMemsetOptSize; 1965 1966 /// \brief Specify maximum bytes of store instructions per memcpy call. 1967 /// 1968 /// When lowering \@llvm.memcpy this field specifies the maximum number of 1969 /// store operations that may be substituted for a call to memcpy. Targets 1970 /// must set this value based on the cost threshold for that target. Targets 1971 /// should assume that the memcpy will be done using as many of the largest 1972 /// store operations first, followed by smaller ones, if necessary, per 1973 /// alignment restrictions. For example, storing 7 bytes on a 32-bit machine 1974 /// with 32-bit alignment would result in one 4-byte store, a one 2-byte store 1975 /// and one 1-byte store. This only applies to copying a constant array of 1976 /// constant size. 1977 unsigned MaxStoresPerMemcpy; 1978 1979 /// Maximum number of store operations that may be substituted for a call to 1980 /// memcpy, used for functions with OptSize attribute. 1981 unsigned MaxStoresPerMemcpyOptSize; 1982 1983 /// \brief Specify maximum bytes of store instructions per memmove call. 1984 /// 1985 /// When lowering \@llvm.memmove this field specifies the maximum number of 1986 /// store instructions that may be substituted for a call to memmove. Targets 1987 /// must set this value based on the cost threshold for that target. Targets 1988 /// should assume that the memmove will be done using as many of the largest 1989 /// store operations first, followed by smaller ones, if necessary, per 1990 /// alignment restrictions. For example, moving 9 bytes on a 32-bit machine 1991 /// with 8-bit alignment would result in nine 1-byte stores. This only 1992 /// applies to copying a constant array of constant size. 1993 unsigned MaxStoresPerMemmove; 1994 1995 /// Maximum number of store instructions that may be substituted for a call to 1996 /// memmove, used for functions with OpSize attribute. 1997 unsigned MaxStoresPerMemmoveOptSize; 1998 1999 /// Tells the code generator that select is more expensive than a branch if 2000 /// the branch is usually predicted right. 2001 bool PredictableSelectIsExpensive; 2002 2003 /// MaskAndBranchFoldingIsLegal - Indicates if the target supports folding 2004 /// a mask of a single bit, a compare, and a branch into a single instruction. 2005 bool MaskAndBranchFoldingIsLegal; 2006 2007 /// \see enableExtLdPromotion. 2008 bool EnableExtLdPromotion; 2009 2010 protected: 2011 /// Return true if the value types that can be represented by the specified 2012 /// register class are all legal. 2013 bool isLegalRC(const TargetRegisterClass *RC) const; 2014 2015 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 2016 /// sequence of memory operands that is recognized by PrologEpilogInserter. 2017 MachineBasicBlock *emitPatchPoint(MachineInstr *MI, 2018 MachineBasicBlock *MBB) const; 2019 }; 2020 2021 /// This class defines information used to lower LLVM code to legal SelectionDAG 2022 /// operators that the target instruction selector can accept natively. 2023 /// 2024 /// This class also defines callbacks that targets must implement to lower 2025 /// target-specific constructs to SelectionDAG operators. 2026 class TargetLowering : public TargetLoweringBase { 2027 TargetLowering(const TargetLowering&) = delete; 2028 void operator=(const TargetLowering&) = delete; 2029 2030 public: 2031 /// NOTE: The TargetMachine owns TLOF. 2032 explicit TargetLowering(const TargetMachine &TM); 2033 2034 /// Returns true by value, base pointer and offset pointer and addressing mode 2035 /// by reference if the node's address can be legally represented as 2036 /// pre-indexed load / store address. getPreIndexedAddressParts(SDNode *,SDValue &,SDValue &,ISD::MemIndexedMode &,SelectionDAG &)2037 virtual bool getPreIndexedAddressParts(SDNode * /*N*/, SDValue &/*Base*/, 2038 SDValue &/*Offset*/, 2039 ISD::MemIndexedMode &/*AM*/, 2040 SelectionDAG &/*DAG*/) const { 2041 return false; 2042 } 2043 2044 /// Returns true by value, base pointer and offset pointer and addressing mode 2045 /// by reference if this node can be combined with a load / store to form a 2046 /// post-indexed load / store. getPostIndexedAddressParts(SDNode *,SDNode *,SDValue &,SDValue &,ISD::MemIndexedMode &,SelectionDAG &)2047 virtual bool getPostIndexedAddressParts(SDNode * /*N*/, SDNode * /*Op*/, 2048 SDValue &/*Base*/, 2049 SDValue &/*Offset*/, 2050 ISD::MemIndexedMode &/*AM*/, 2051 SelectionDAG &/*DAG*/) const { 2052 return false; 2053 } 2054 2055 /// Return the entry encoding for a jump table in the current function. The 2056 /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum. 2057 virtual unsigned getJumpTableEncoding() const; 2058 2059 virtual const MCExpr * LowerCustomJumpTableEntry(const MachineJumpTableInfo *,const MachineBasicBlock *,unsigned,MCContext &)2060 LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/, 2061 const MachineBasicBlock * /*MBB*/, unsigned /*uid*/, 2062 MCContext &/*Ctx*/) const { 2063 llvm_unreachable("Need to implement this hook if target has custom JTIs"); 2064 } 2065 2066 /// Returns relocation base for the given PIC jumptable. 2067 virtual SDValue getPICJumpTableRelocBase(SDValue Table, 2068 SelectionDAG &DAG) const; 2069 2070 /// This returns the relocation base for the given PIC jumptable, the same as 2071 /// getPICJumpTableRelocBase, but as an MCExpr. 2072 virtual const MCExpr * 2073 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 2074 unsigned JTI, MCContext &Ctx) const; 2075 2076 /// Return true if folding a constant offset with the given GlobalAddress is 2077 /// legal. It is frequently not legal in PIC relocation models. 2078 virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const; 2079 2080 bool isInTailCallPosition(SelectionDAG &DAG, SDNode *Node, 2081 SDValue &Chain) const; 2082 2083 void softenSetCCOperands(SelectionDAG &DAG, EVT VT, 2084 SDValue &NewLHS, SDValue &NewRHS, 2085 ISD::CondCode &CCCode, SDLoc DL) const; 2086 2087 /// Returns a pair of (return value, chain). 2088 /// It is an error to pass RTLIB::UNKNOWN_LIBCALL as \p LC. 2089 std::pair<SDValue, SDValue> makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC, 2090 EVT RetVT, const SDValue *Ops, 2091 unsigned NumOps, bool isSigned, 2092 SDLoc dl, bool doesNotReturn = false, 2093 bool isReturnValueUsed = true) const; 2094 2095 //===--------------------------------------------------------------------===// 2096 // TargetLowering Optimization Methods 2097 // 2098 2099 /// A convenience struct that encapsulates a DAG, and two SDValues for 2100 /// returning information from TargetLowering to its clients that want to 2101 /// combine. 2102 struct TargetLoweringOpt { 2103 SelectionDAG &DAG; 2104 bool LegalTys; 2105 bool LegalOps; 2106 SDValue Old; 2107 SDValue New; 2108 TargetLoweringOptTargetLoweringOpt2109 explicit TargetLoweringOpt(SelectionDAG &InDAG, 2110 bool LT, bool LO) : 2111 DAG(InDAG), LegalTys(LT), LegalOps(LO) {} 2112 LegalTypesTargetLoweringOpt2113 bool LegalTypes() const { return LegalTys; } LegalOperationsTargetLoweringOpt2114 bool LegalOperations() const { return LegalOps; } 2115 CombineToTargetLoweringOpt2116 bool CombineTo(SDValue O, SDValue N) { 2117 Old = O; 2118 New = N; 2119 return true; 2120 } 2121 2122 /// Check to see if the specified operand of the specified instruction is a 2123 /// constant integer. If so, check to see if there are any bits set in the 2124 /// constant that are not demanded. If so, shrink the constant and return 2125 /// true. 2126 bool ShrinkDemandedConstant(SDValue Op, const APInt &Demanded); 2127 2128 /// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free. This 2129 /// uses isZExtFree and ZERO_EXTEND for the widening cast, but it could be 2130 /// generalized for targets with other types of implicit widening casts. 2131 bool ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &Demanded, 2132 SDLoc dl); 2133 }; 2134 2135 /// Look at Op. At this point, we know that only the DemandedMask bits of the 2136 /// result of Op are ever used downstream. If we can use this information to 2137 /// simplify Op, create a new simplified DAG node and return true, returning 2138 /// the original and new nodes in Old and New. Otherwise, analyze the 2139 /// expression and return a mask of KnownOne and KnownZero bits for the 2140 /// expression (used to simplify the caller). The KnownZero/One bits may only 2141 /// be accurate for those bits in the DemandedMask. 2142 bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask, 2143 APInt &KnownZero, APInt &KnownOne, 2144 TargetLoweringOpt &TLO, unsigned Depth = 0) const; 2145 2146 /// Determine which of the bits specified in Mask are known to be either zero 2147 /// or one and return them in the KnownZero/KnownOne bitsets. 2148 virtual void computeKnownBitsForTargetNode(const SDValue Op, 2149 APInt &KnownZero, 2150 APInt &KnownOne, 2151 const SelectionDAG &DAG, 2152 unsigned Depth = 0) const; 2153 2154 /// This method can be implemented by targets that want to expose additional 2155 /// information about sign bits to the DAG Combiner. 2156 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op, 2157 const SelectionDAG &DAG, 2158 unsigned Depth = 0) const; 2159 2160 struct DAGCombinerInfo { 2161 void *DC; // The DAG Combiner object. 2162 CombineLevel Level; 2163 bool CalledByLegalizer; 2164 public: 2165 SelectionDAG &DAG; 2166 DAGCombinerInfoDAGCombinerInfo2167 DAGCombinerInfo(SelectionDAG &dag, CombineLevel level, bool cl, void *dc) 2168 : DC(dc), Level(level), CalledByLegalizer(cl), DAG(dag) {} 2169 isBeforeLegalizeDAGCombinerInfo2170 bool isBeforeLegalize() const { return Level == BeforeLegalizeTypes; } isBeforeLegalizeOpsDAGCombinerInfo2171 bool isBeforeLegalizeOps() const { return Level < AfterLegalizeVectorOps; } isAfterLegalizeVectorOpsDAGCombinerInfo2172 bool isAfterLegalizeVectorOps() const { 2173 return Level == AfterLegalizeDAG; 2174 } getDAGCombineLevelDAGCombinerInfo2175 CombineLevel getDAGCombineLevel() { return Level; } isCalledByLegalizerDAGCombinerInfo2176 bool isCalledByLegalizer() const { return CalledByLegalizer; } 2177 2178 void AddToWorklist(SDNode *N); 2179 void RemoveFromWorklist(SDNode *N); 2180 SDValue CombineTo(SDNode *N, ArrayRef<SDValue> To, bool AddTo = true); 2181 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true); 2182 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true); 2183 2184 void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO); 2185 }; 2186 2187 /// Return if the N is a constant or constant vector equal to the true value 2188 /// from getBooleanContents(). 2189 bool isConstTrueVal(const SDNode *N) const; 2190 2191 /// Return if the N is a constant or constant vector equal to the false value 2192 /// from getBooleanContents(). 2193 bool isConstFalseVal(const SDNode *N) const; 2194 2195 /// Try to simplify a setcc built with the specified operands and cc. If it is 2196 /// unable to simplify it, return a null SDValue. 2197 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 2198 ISD::CondCode Cond, bool foldBooleans, 2199 DAGCombinerInfo &DCI, SDLoc dl) const; 2200 2201 /// Returns true (and the GlobalValue and the offset) if the node is a 2202 /// GlobalAddress + offset. 2203 virtual bool 2204 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const; 2205 2206 /// This method will be invoked for all target nodes and for any 2207 /// target-independent nodes that the target has registered with invoke it 2208 /// for. 2209 /// 2210 /// The semantics are as follows: 2211 /// Return Value: 2212 /// SDValue.Val == 0 - No change was made 2213 /// SDValue.Val == N - N was replaced, is dead, and is already handled. 2214 /// otherwise - N should be replaced by the returned Operand. 2215 /// 2216 /// In addition, methods provided by DAGCombinerInfo may be used to perform 2217 /// more complex transformations. 2218 /// 2219 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const; 2220 2221 /// Return true if it is profitable to move a following shift through this 2222 // node, adjusting any immediate operands as necessary to preserve semantics. 2223 // This transformation may not be desirable if it disrupts a particularly 2224 // auspicious target-specific tree (e.g. bitfield extraction in AArch64). 2225 // By default, it returns true. isDesirableToCommuteWithShift(const SDNode * N)2226 virtual bool isDesirableToCommuteWithShift(const SDNode *N /*Op*/) const { 2227 return true; 2228 } 2229 2230 /// Return true if the target has native support for the specified value type 2231 /// and it is 'desirable' to use the type for the given node type. e.g. On x86 2232 /// i16 is legal, but undesirable since i16 instruction encodings are longer 2233 /// and some i16 instructions are slow. isTypeDesirableForOp(unsigned,EVT VT)2234 virtual bool isTypeDesirableForOp(unsigned /*Opc*/, EVT VT) const { 2235 // By default, assume all legal types are desirable. 2236 return isTypeLegal(VT); 2237 } 2238 2239 /// Return true if it is profitable for dag combiner to transform a floating 2240 /// point op of specified opcode to a equivalent op of an integer 2241 /// type. e.g. f32 load -> i32 load can be profitable on ARM. isDesirableToTransformToIntegerOp(unsigned,EVT)2242 virtual bool isDesirableToTransformToIntegerOp(unsigned /*Opc*/, 2243 EVT /*VT*/) const { 2244 return false; 2245 } 2246 2247 /// This method query the target whether it is beneficial for dag combiner to 2248 /// promote the specified node. If true, it should return the desired 2249 /// promotion type by reference. IsDesirableToPromoteOp(SDValue,EVT &)2250 virtual bool IsDesirableToPromoteOp(SDValue /*Op*/, EVT &/*PVT*/) const { 2251 return false; 2252 } 2253 2254 //===--------------------------------------------------------------------===// 2255 // Lowering methods - These methods must be implemented by targets so that 2256 // the SelectionDAGBuilder code knows how to lower these. 2257 // 2258 2259 /// This hook must be implemented to lower the incoming (formal) arguments, 2260 /// described by the Ins array, into the specified DAG. The implementation 2261 /// should fill in the InVals array with legal-type argument values, and 2262 /// return the resulting token chain value. 2263 /// 2264 virtual SDValue LowerFormalArguments(SDValue,CallingConv::ID,bool,const SmallVectorImpl<ISD::InputArg> &,SDLoc,SelectionDAG &,SmallVectorImpl<SDValue> &)2265 LowerFormalArguments(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 2266 bool /*isVarArg*/, 2267 const SmallVectorImpl<ISD::InputArg> &/*Ins*/, 2268 SDLoc /*dl*/, SelectionDAG &/*DAG*/, 2269 SmallVectorImpl<SDValue> &/*InVals*/) const { 2270 llvm_unreachable("Not Implemented"); 2271 } 2272 2273 struct ArgListEntry { 2274 SDValue Node; 2275 Type* Ty; 2276 bool isSExt : 1; 2277 bool isZExt : 1; 2278 bool isInReg : 1; 2279 bool isSRet : 1; 2280 bool isNest : 1; 2281 bool isByVal : 1; 2282 bool isInAlloca : 1; 2283 bool isReturned : 1; 2284 uint16_t Alignment; 2285 ArgListEntryArgListEntry2286 ArgListEntry() : isSExt(false), isZExt(false), isInReg(false), 2287 isSRet(false), isNest(false), isByVal(false), isInAlloca(false), 2288 isReturned(false), Alignment(0) { } 2289 2290 void setAttributes(ImmutableCallSite *CS, unsigned AttrIdx); 2291 }; 2292 typedef std::vector<ArgListEntry> ArgListTy; 2293 2294 /// This structure contains all information that is necessary for lowering 2295 /// calls. It is passed to TLI::LowerCallTo when the SelectionDAG builder 2296 /// needs to lower a call, and targets will see this struct in their LowerCall 2297 /// implementation. 2298 struct CallLoweringInfo { 2299 SDValue Chain; 2300 Type *RetTy; 2301 bool RetSExt : 1; 2302 bool RetZExt : 1; 2303 bool IsVarArg : 1; 2304 bool IsInReg : 1; 2305 bool DoesNotReturn : 1; 2306 bool IsReturnValueUsed : 1; 2307 2308 // IsTailCall should be modified by implementations of 2309 // TargetLowering::LowerCall that perform tail call conversions. 2310 bool IsTailCall; 2311 2312 unsigned NumFixedArgs; 2313 CallingConv::ID CallConv; 2314 SDValue Callee; 2315 ArgListTy Args; 2316 SelectionDAG &DAG; 2317 SDLoc DL; 2318 ImmutableCallSite *CS; 2319 bool IsPatchPoint; 2320 SmallVector<ISD::OutputArg, 32> Outs; 2321 SmallVector<SDValue, 32> OutVals; 2322 SmallVector<ISD::InputArg, 32> Ins; 2323 CallLoweringInfoCallLoweringInfo2324 CallLoweringInfo(SelectionDAG &DAG) 2325 : RetTy(nullptr), RetSExt(false), RetZExt(false), IsVarArg(false), 2326 IsInReg(false), DoesNotReturn(false), IsReturnValueUsed(true), 2327 IsTailCall(false), NumFixedArgs(-1), CallConv(CallingConv::C), 2328 DAG(DAG), CS(nullptr), IsPatchPoint(false) {} 2329 setDebugLocCallLoweringInfo2330 CallLoweringInfo &setDebugLoc(SDLoc dl) { 2331 DL = dl; 2332 return *this; 2333 } 2334 setChainCallLoweringInfo2335 CallLoweringInfo &setChain(SDValue InChain) { 2336 Chain = InChain; 2337 return *this; 2338 } 2339 2340 CallLoweringInfo &setCallee(CallingConv::ID CC, Type *ResultType, 2341 SDValue Target, ArgListTy &&ArgsList, 2342 unsigned FixedArgs = -1) { 2343 RetTy = ResultType; 2344 Callee = Target; 2345 CallConv = CC; 2346 NumFixedArgs = 2347 (FixedArgs == static_cast<unsigned>(-1) ? Args.size() : FixedArgs); 2348 Args = std::move(ArgsList); 2349 return *this; 2350 } 2351 setCalleeCallLoweringInfo2352 CallLoweringInfo &setCallee(Type *ResultType, FunctionType *FTy, 2353 SDValue Target, ArgListTy &&ArgsList, 2354 ImmutableCallSite &Call) { 2355 RetTy = ResultType; 2356 2357 IsInReg = Call.paramHasAttr(0, Attribute::InReg); 2358 DoesNotReturn = Call.doesNotReturn(); 2359 IsVarArg = FTy->isVarArg(); 2360 IsReturnValueUsed = !Call.getInstruction()->use_empty(); 2361 RetSExt = Call.paramHasAttr(0, Attribute::SExt); 2362 RetZExt = Call.paramHasAttr(0, Attribute::ZExt); 2363 2364 Callee = Target; 2365 2366 CallConv = Call.getCallingConv(); 2367 NumFixedArgs = FTy->getNumParams(); 2368 Args = std::move(ArgsList); 2369 2370 CS = &Call; 2371 2372 return *this; 2373 } 2374 2375 CallLoweringInfo &setInRegister(bool Value = true) { 2376 IsInReg = Value; 2377 return *this; 2378 } 2379 2380 CallLoweringInfo &setNoReturn(bool Value = true) { 2381 DoesNotReturn = Value; 2382 return *this; 2383 } 2384 2385 CallLoweringInfo &setVarArg(bool Value = true) { 2386 IsVarArg = Value; 2387 return *this; 2388 } 2389 2390 CallLoweringInfo &setTailCall(bool Value = true) { 2391 IsTailCall = Value; 2392 return *this; 2393 } 2394 2395 CallLoweringInfo &setDiscardResult(bool Value = true) { 2396 IsReturnValueUsed = !Value; 2397 return *this; 2398 } 2399 2400 CallLoweringInfo &setSExtResult(bool Value = true) { 2401 RetSExt = Value; 2402 return *this; 2403 } 2404 2405 CallLoweringInfo &setZExtResult(bool Value = true) { 2406 RetZExt = Value; 2407 return *this; 2408 } 2409 2410 CallLoweringInfo &setIsPatchPoint(bool Value = true) { 2411 IsPatchPoint = Value; 2412 return *this; 2413 } 2414 getArgsCallLoweringInfo2415 ArgListTy &getArgs() { 2416 return Args; 2417 } 2418 2419 }; 2420 2421 /// This function lowers an abstract call to a function into an actual call. 2422 /// This returns a pair of operands. The first element is the return value 2423 /// for the function (if RetTy is not VoidTy). The second element is the 2424 /// outgoing token chain. It calls LowerCall to do the actual lowering. 2425 std::pair<SDValue, SDValue> LowerCallTo(CallLoweringInfo &CLI) const; 2426 2427 /// This hook must be implemented to lower calls into the specified 2428 /// DAG. The outgoing arguments to the call are described by the Outs array, 2429 /// and the values to be returned by the call are described by the Ins 2430 /// array. The implementation should fill in the InVals array with legal-type 2431 /// return values from the call, and return the resulting token chain value. 2432 virtual SDValue LowerCall(CallLoweringInfo &,SmallVectorImpl<SDValue> &)2433 LowerCall(CallLoweringInfo &/*CLI*/, 2434 SmallVectorImpl<SDValue> &/*InVals*/) const { 2435 llvm_unreachable("Not Implemented"); 2436 } 2437 2438 /// Target-specific cleanup for formal ByVal parameters. HandleByVal(CCState *,unsigned &,unsigned)2439 virtual void HandleByVal(CCState *, unsigned &, unsigned) const {} 2440 2441 /// This hook should be implemented to check whether the return values 2442 /// described by the Outs array can fit into the return registers. If false 2443 /// is returned, an sret-demotion is performed. CanLowerReturn(CallingConv::ID,MachineFunction &,bool,const SmallVectorImpl<ISD::OutputArg> &,LLVMContext &)2444 virtual bool CanLowerReturn(CallingConv::ID /*CallConv*/, 2445 MachineFunction &/*MF*/, bool /*isVarArg*/, 2446 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 2447 LLVMContext &/*Context*/) const 2448 { 2449 // Return true by default to get preexisting behavior. 2450 return true; 2451 } 2452 2453 /// This hook must be implemented to lower outgoing return values, described 2454 /// by the Outs array, into the specified DAG. The implementation should 2455 /// return the resulting token chain value. 2456 virtual SDValue LowerReturn(SDValue,CallingConv::ID,bool,const SmallVectorImpl<ISD::OutputArg> &,const SmallVectorImpl<SDValue> &,SDLoc,SelectionDAG &)2457 LowerReturn(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 2458 bool /*isVarArg*/, 2459 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 2460 const SmallVectorImpl<SDValue> &/*OutVals*/, 2461 SDLoc /*dl*/, SelectionDAG &/*DAG*/) const { 2462 llvm_unreachable("Not Implemented"); 2463 } 2464 2465 /// Return true if result of the specified node is used by a return node 2466 /// only. It also compute and return the input chain for the tail call. 2467 /// 2468 /// This is used to determine whether it is possible to codegen a libcall as 2469 /// tail call at legalization time. isUsedByReturnOnly(SDNode *,SDValue &)2470 virtual bool isUsedByReturnOnly(SDNode *, SDValue &/*Chain*/) const { 2471 return false; 2472 } 2473 2474 /// Return true if the target may be able emit the call instruction as a tail 2475 /// call. This is used by optimization passes to determine if it's profitable 2476 /// to duplicate return instructions to enable tailcall optimization. mayBeEmittedAsTailCall(CallInst *)2477 virtual bool mayBeEmittedAsTailCall(CallInst *) const { 2478 return false; 2479 } 2480 2481 /// Return the builtin name for the __builtin___clear_cache intrinsic 2482 /// Default is to invoke the clear cache library call getClearCacheBuiltinName()2483 virtual const char * getClearCacheBuiltinName() const { 2484 return "__clear_cache"; 2485 } 2486 2487 /// Return the register ID of the name passed in. Used by named register 2488 /// global variables extension. There is no target-independent behaviour 2489 /// so the default action is to bail. getRegisterByName(const char * RegName,EVT VT,SelectionDAG & DAG)2490 virtual unsigned getRegisterByName(const char* RegName, EVT VT, 2491 SelectionDAG &DAG) const { 2492 report_fatal_error("Named registers not implemented for this target"); 2493 } 2494 2495 /// Return the type that should be used to zero or sign extend a 2496 /// zeroext/signext integer argument or return value. FIXME: Most C calling 2497 /// convention requires the return type to be promoted, but this is not true 2498 /// all the time, e.g. i1 on x86-64. It is also not necessary for non-C 2499 /// calling conventions. The frontend should handle this and include all of 2500 /// the necessary information. getTypeForExtArgOrReturn(LLVMContext & Context,EVT VT,ISD::NodeType)2501 virtual EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT, 2502 ISD::NodeType /*ExtendKind*/) const { 2503 EVT MinVT = getRegisterType(Context, MVT::i32); 2504 return VT.bitsLT(MinVT) ? MinVT : VT; 2505 } 2506 2507 /// For some targets, an LLVM struct type must be broken down into multiple 2508 /// simple types, but the calling convention specifies that the entire struct 2509 /// must be passed in a block of consecutive registers. 2510 virtual bool functionArgumentNeedsConsecutiveRegisters(Type * Ty,CallingConv::ID CallConv,bool isVarArg)2511 functionArgumentNeedsConsecutiveRegisters(Type *Ty, CallingConv::ID CallConv, 2512 bool isVarArg) const { 2513 return false; 2514 } 2515 2516 /// Returns a 0 terminated array of registers that can be safely used as 2517 /// scratch registers. getScratchRegisters(CallingConv::ID CC)2518 virtual const MCPhysReg *getScratchRegisters(CallingConv::ID CC) const { 2519 return nullptr; 2520 } 2521 2522 /// This callback is used to prepare for a volatile or atomic load. 2523 /// It takes a chain node as input and returns the chain for the load itself. 2524 /// 2525 /// Having a callback like this is necessary for targets like SystemZ, 2526 /// which allows a CPU to reuse the result of a previous load indefinitely, 2527 /// even if a cache-coherent store is performed by another CPU. The default 2528 /// implementation does nothing. prepareVolatileOrAtomicLoad(SDValue Chain,SDLoc DL,SelectionDAG & DAG)2529 virtual SDValue prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL, 2530 SelectionDAG &DAG) const { 2531 return Chain; 2532 } 2533 2534 /// This callback is invoked by the type legalizer to legalize nodes with an 2535 /// illegal operand type but legal result types. It replaces the 2536 /// LowerOperation callback in the type Legalizer. The reason we can not do 2537 /// away with LowerOperation entirely is that LegalizeDAG isn't yet ready to 2538 /// use this callback. 2539 /// 2540 /// TODO: Consider merging with ReplaceNodeResults. 2541 /// 2542 /// The target places new result values for the node in Results (their number 2543 /// and types must exactly match those of the original return values of 2544 /// the node), or leaves Results empty, which indicates that the node is not 2545 /// to be custom lowered after all. 2546 /// The default implementation calls LowerOperation. 2547 virtual void LowerOperationWrapper(SDNode *N, 2548 SmallVectorImpl<SDValue> &Results, 2549 SelectionDAG &DAG) const; 2550 2551 /// This callback is invoked for operations that are unsupported by the 2552 /// target, which are registered to use 'custom' lowering, and whose defined 2553 /// values are all legal. If the target has no operations that require custom 2554 /// lowering, it need not implement this. The default implementation of this 2555 /// aborts. 2556 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; 2557 2558 /// This callback is invoked when a node result type is illegal for the 2559 /// target, and the operation was registered to use 'custom' lowering for that 2560 /// result type. The target places new result values for the node in Results 2561 /// (their number and types must exactly match those of the original return 2562 /// values of the node), or leaves Results empty, which indicates that the 2563 /// node is not to be custom lowered after all. 2564 /// 2565 /// If the target has no operations that require custom lowering, it need not 2566 /// implement this. The default implementation aborts. ReplaceNodeResults(SDNode *,SmallVectorImpl<SDValue> &,SelectionDAG &)2567 virtual void ReplaceNodeResults(SDNode * /*N*/, 2568 SmallVectorImpl<SDValue> &/*Results*/, 2569 SelectionDAG &/*DAG*/) const { 2570 llvm_unreachable("ReplaceNodeResults not implemented for this target!"); 2571 } 2572 2573 /// This method returns the name of a target specific DAG node. 2574 virtual const char *getTargetNodeName(unsigned Opcode) const; 2575 2576 /// This method returns a target specific FastISel object, or null if the 2577 /// target does not support "fast" ISel. createFastISel(FunctionLoweringInfo &,const TargetLibraryInfo *)2578 virtual FastISel *createFastISel(FunctionLoweringInfo &, 2579 const TargetLibraryInfo *) const { 2580 return nullptr; 2581 } 2582 2583 2584 bool verifyReturnAddressArgumentIsConstant(SDValue Op, 2585 SelectionDAG &DAG) const; 2586 2587 //===--------------------------------------------------------------------===// 2588 // Inline Asm Support hooks 2589 // 2590 2591 /// This hook allows the target to expand an inline asm call to be explicit 2592 /// llvm code if it wants to. This is useful for turning simple inline asms 2593 /// into LLVM intrinsics, which gives the compiler more information about the 2594 /// behavior of the code. ExpandInlineAsm(CallInst *)2595 virtual bool ExpandInlineAsm(CallInst *) const { 2596 return false; 2597 } 2598 2599 enum ConstraintType { 2600 C_Register, // Constraint represents specific register(s). 2601 C_RegisterClass, // Constraint represents any of register(s) in class. 2602 C_Memory, // Memory constraint. 2603 C_Other, // Something else. 2604 C_Unknown // Unsupported constraint. 2605 }; 2606 2607 enum ConstraintWeight { 2608 // Generic weights. 2609 CW_Invalid = -1, // No match. 2610 CW_Okay = 0, // Acceptable. 2611 CW_Good = 1, // Good weight. 2612 CW_Better = 2, // Better weight. 2613 CW_Best = 3, // Best weight. 2614 2615 // Well-known weights. 2616 CW_SpecificReg = CW_Okay, // Specific register operands. 2617 CW_Register = CW_Good, // Register operands. 2618 CW_Memory = CW_Better, // Memory operands. 2619 CW_Constant = CW_Best, // Constant operand. 2620 CW_Default = CW_Okay // Default or don't know type. 2621 }; 2622 2623 /// This contains information for each constraint that we are lowering. 2624 struct AsmOperandInfo : public InlineAsm::ConstraintInfo { 2625 /// This contains the actual string for the code, like "m". TargetLowering 2626 /// picks the 'best' code from ConstraintInfo::Codes that most closely 2627 /// matches the operand. 2628 std::string ConstraintCode; 2629 2630 /// Information about the constraint code, e.g. Register, RegisterClass, 2631 /// Memory, Other, Unknown. 2632 TargetLowering::ConstraintType ConstraintType; 2633 2634 /// If this is the result output operand or a clobber, this is null, 2635 /// otherwise it is the incoming operand to the CallInst. This gets 2636 /// modified as the asm is processed. 2637 Value *CallOperandVal; 2638 2639 /// The ValueType for the operand value. 2640 MVT ConstraintVT; 2641 2642 /// Return true of this is an input operand that is a matching constraint 2643 /// like "4". 2644 bool isMatchingInputConstraint() const; 2645 2646 /// If this is an input matching constraint, this method returns the output 2647 /// operand it matches. 2648 unsigned getMatchedOperand() const; 2649 2650 /// Copy constructor for copying from a ConstraintInfo. AsmOperandInfoAsmOperandInfo2651 AsmOperandInfo(InlineAsm::ConstraintInfo Info) 2652 : InlineAsm::ConstraintInfo(std::move(Info)), 2653 ConstraintType(TargetLowering::C_Unknown), CallOperandVal(nullptr), 2654 ConstraintVT(MVT::Other) {} 2655 }; 2656 2657 typedef std::vector<AsmOperandInfo> AsmOperandInfoVector; 2658 2659 /// Split up the constraint string from the inline assembly value into the 2660 /// specific constraints and their prefixes, and also tie in the associated 2661 /// operand values. If this returns an empty vector, and if the constraint 2662 /// string itself isn't empty, there was an error parsing. 2663 virtual AsmOperandInfoVector ParseConstraints(const DataLayout &DL, 2664 const TargetRegisterInfo *TRI, 2665 ImmutableCallSite CS) const; 2666 2667 /// Examine constraint type and operand type and determine a weight value. 2668 /// The operand object must already have been set up with the operand type. 2669 virtual ConstraintWeight getMultipleConstraintMatchWeight( 2670 AsmOperandInfo &info, int maIndex) const; 2671 2672 /// Examine constraint string and operand type and determine a weight value. 2673 /// The operand object must already have been set up with the operand type. 2674 virtual ConstraintWeight getSingleConstraintMatchWeight( 2675 AsmOperandInfo &info, const char *constraint) const; 2676 2677 /// Determines the constraint code and constraint type to use for the specific 2678 /// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType. 2679 /// If the actual operand being passed in is available, it can be passed in as 2680 /// Op, otherwise an empty SDValue can be passed. 2681 virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo, 2682 SDValue Op, 2683 SelectionDAG *DAG = nullptr) const; 2684 2685 /// Given a constraint, return the type of constraint it is for this target. 2686 virtual ConstraintType getConstraintType(StringRef Constraint) const; 2687 2688 /// Given a physical register constraint (e.g. {edx}), return the register 2689 /// number and the register class for the register. 2690 /// 2691 /// Given a register class constraint, like 'r', if this corresponds directly 2692 /// to an LLVM register class, return a register of 0 and the register class 2693 /// pointer. 2694 /// 2695 /// This should only be used for C_Register constraints. On error, this 2696 /// returns a register number of 0 and a null register class pointer. 2697 virtual std::pair<unsigned, const TargetRegisterClass *> 2698 getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI, 2699 StringRef Constraint, MVT VT) const; 2700 getInlineAsmMemConstraint(StringRef ConstraintCode)2701 virtual unsigned getInlineAsmMemConstraint(StringRef ConstraintCode) const { 2702 if (ConstraintCode == "i") 2703 return InlineAsm::Constraint_i; 2704 else if (ConstraintCode == "m") 2705 return InlineAsm::Constraint_m; 2706 return InlineAsm::Constraint_Unknown; 2707 } 2708 2709 /// Try to replace an X constraint, which matches anything, with another that 2710 /// has more specific requirements based on the type of the corresponding 2711 /// operand. This returns null if there is no replacement to make. 2712 virtual const char *LowerXConstraint(EVT ConstraintVT) const; 2713 2714 /// Lower the specified operand into the Ops vector. If it is invalid, don't 2715 /// add anything to Ops. 2716 virtual void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, 2717 std::vector<SDValue> &Ops, 2718 SelectionDAG &DAG) const; 2719 2720 //===--------------------------------------------------------------------===// 2721 // Div utility functions 2722 // 2723 SDValue BuildSDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, 2724 bool IsAfterLegalization, 2725 std::vector<SDNode *> *Created) const; 2726 SDValue BuildUDIV(SDNode *N, const APInt &Divisor, SelectionDAG &DAG, 2727 bool IsAfterLegalization, 2728 std::vector<SDNode *> *Created) const; BuildSDIVPow2(SDNode * N,const APInt & Divisor,SelectionDAG & DAG,std::vector<SDNode * > * Created)2729 virtual SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, 2730 SelectionDAG &DAG, 2731 std::vector<SDNode *> *Created) const { 2732 return SDValue(); 2733 } 2734 2735 /// Indicate whether this target prefers to combine the given number of FDIVs 2736 /// with the same divisor. combineRepeatedFPDivisors(unsigned NumUsers)2737 virtual bool combineRepeatedFPDivisors(unsigned NumUsers) const { 2738 return false; 2739 } 2740 2741 /// Hooks for building estimates in place of slower divisions and square 2742 /// roots. 2743 2744 /// Return a reciprocal square root estimate value for the input operand. 2745 /// The RefinementSteps output is the number of Newton-Raphson refinement 2746 /// iterations required to generate a sufficient (though not necessarily 2747 /// IEEE-754 compliant) estimate for the value type. 2748 /// The boolean UseOneConstNR output is used to select a Newton-Raphson 2749 /// algorithm implementation that uses one constant or two constants. 2750 /// A target may choose to implement its own refinement within this function. 2751 /// If that's true, then return '0' as the number of RefinementSteps to avoid 2752 /// any further refinement of the estimate. 2753 /// An empty SDValue return means no estimate sequence can be created. getRsqrtEstimate(SDValue Operand,DAGCombinerInfo & DCI,unsigned & RefinementSteps,bool & UseOneConstNR)2754 virtual SDValue getRsqrtEstimate(SDValue Operand, DAGCombinerInfo &DCI, 2755 unsigned &RefinementSteps, 2756 bool &UseOneConstNR) const { 2757 return SDValue(); 2758 } 2759 2760 /// Return a reciprocal estimate value for the input operand. 2761 /// The RefinementSteps output is the number of Newton-Raphson refinement 2762 /// iterations required to generate a sufficient (though not necessarily 2763 /// IEEE-754 compliant) estimate for the value type. 2764 /// A target may choose to implement its own refinement within this function. 2765 /// If that's true, then return '0' as the number of RefinementSteps to avoid 2766 /// any further refinement of the estimate. 2767 /// An empty SDValue return means no estimate sequence can be created. getRecipEstimate(SDValue Operand,DAGCombinerInfo & DCI,unsigned & RefinementSteps)2768 virtual SDValue getRecipEstimate(SDValue Operand, DAGCombinerInfo &DCI, 2769 unsigned &RefinementSteps) const { 2770 return SDValue(); 2771 } 2772 2773 //===--------------------------------------------------------------------===// 2774 // Legalization utility functions 2775 // 2776 2777 /// Expand a MUL into two nodes. One that computes the high bits of 2778 /// the result and one that computes the low bits. 2779 /// \param HiLoVT The value type to use for the Lo and Hi nodes. 2780 /// \param LL Low bits of the LHS of the MUL. You can use this parameter 2781 /// if you want to control how low bits are extracted from the LHS. 2782 /// \param LH High bits of the LHS of the MUL. See LL for meaning. 2783 /// \param RL Low bits of the RHS of the MUL. See LL for meaning 2784 /// \param RH High bits of the RHS of the MUL. See LL for meaning. 2785 /// \returns true if the node has been expanded. false if it has not 2786 bool expandMUL(SDNode *N, SDValue &Lo, SDValue &Hi, EVT HiLoVT, 2787 SelectionDAG &DAG, SDValue LL = SDValue(), 2788 SDValue LH = SDValue(), SDValue RL = SDValue(), 2789 SDValue RH = SDValue()) const; 2790 2791 /// Expand float(f32) to SINT(i64) conversion 2792 /// \param N Node to expand 2793 /// \param Result output after conversion 2794 /// \returns True, if the expansion was successful, false otherwise 2795 bool expandFP_TO_SINT(SDNode *N, SDValue &Result, SelectionDAG &DAG) const; 2796 2797 //===--------------------------------------------------------------------===// 2798 // Instruction Emitting Hooks 2799 // 2800 2801 /// This method should be implemented by targets that mark instructions with 2802 /// the 'usesCustomInserter' flag. These instructions are special in various 2803 /// ways, which require special support to insert. The specified MachineInstr 2804 /// is created but not inserted into any basic blocks, and this method is 2805 /// called to expand it into a sequence of instructions, potentially also 2806 /// creating new basic blocks and control flow. 2807 /// As long as the returned basic block is different (i.e., we created a new 2808 /// one), the custom inserter is free to modify the rest of \p MBB. 2809 virtual MachineBasicBlock * 2810 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const; 2811 2812 /// This method should be implemented by targets that mark instructions with 2813 /// the 'hasPostISelHook' flag. These instructions must be adjusted after 2814 /// instruction selection by target hooks. e.g. To fill in optional defs for 2815 /// ARM 's' setting instructions. 2816 virtual void 2817 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const; 2818 2819 /// If this function returns true, SelectionDAGBuilder emits a 2820 /// LOAD_STACK_GUARD node when it is lowering Intrinsic::stackprotector. useLoadStackGuardNode()2821 virtual bool useLoadStackGuardNode() const { 2822 return false; 2823 } 2824 }; 2825 2826 /// Given an LLVM IR type and return type attributes, compute the return value 2827 /// EVTs and flags, and optionally also the offsets, if the return value is 2828 /// being lowered to memory. 2829 void GetReturnInfo(Type *ReturnType, AttributeSet attr, 2830 SmallVectorImpl<ISD::OutputArg> &Outs, 2831 const TargetLowering &TLI, const DataLayout &DL); 2832 2833 } // end llvm namespace 2834 2835 #endif 2836