1 //===-- llvm/Target/TargetLowering.h - Target Lowering Info -----*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file describes how to lower LLVM code to machine code. This has two 11 // main components: 12 // 13 // 1. Which ValueTypes are natively supported by the target. 14 // 2. Which operations are supported for supported ValueTypes. 15 // 3. Cost thresholds for alternative implementations of certain operations. 16 // 17 // In addition it has a few other components, like information about FP 18 // immediates. 19 // 20 //===----------------------------------------------------------------------===// 21 22 #ifndef LLVM_TARGET_TARGETLOWERING_H 23 #define LLVM_TARGET_TARGETLOWERING_H 24 25 #include "llvm/CallingConv.h" 26 #include "llvm/InlineAsm.h" 27 #include "llvm/Attributes.h" 28 #include "llvm/ADT/DenseMap.h" 29 #include "llvm/Support/CallSite.h" 30 #include "llvm/CodeGen/SelectionDAGNodes.h" 31 #include "llvm/CodeGen/RuntimeLibcalls.h" 32 #include "llvm/Support/DebugLoc.h" 33 #include "llvm/Target/TargetCallingConv.h" 34 #include "llvm/Target/TargetMachine.h" 35 #include <climits> 36 #include <map> 37 #include <vector> 38 39 namespace llvm { 40 class CallInst; 41 class CCState; 42 class FastISel; 43 class FunctionLoweringInfo; 44 class ImmutableCallSite; 45 class IntrinsicInst; 46 class MachineBasicBlock; 47 class MachineFunction; 48 class MachineInstr; 49 class MachineJumpTableInfo; 50 class MCContext; 51 class MCExpr; 52 template<typename T> class SmallVectorImpl; 53 class TargetData; 54 class TargetRegisterClass; 55 class TargetLibraryInfo; 56 class TargetLoweringObjectFile; 57 class Value; 58 59 namespace Sched { 60 enum Preference { 61 None, // No preference 62 Source, // Follow source order. 63 RegPressure, // Scheduling for lowest register pressure. 64 Hybrid, // Scheduling for both latency and register pressure. 65 ILP, // Scheduling for ILP in low register pressure mode. 66 VLIW // Scheduling for VLIW targets. 67 }; 68 } 69 70 71 //===----------------------------------------------------------------------===// 72 /// TargetLowering - This class defines information used to lower LLVM code to 73 /// legal SelectionDAG operators that the target instruction selector can accept 74 /// natively. 75 /// 76 /// This class also defines callbacks that targets must implement to lower 77 /// target-specific constructs to SelectionDAG operators. 78 /// 79 class TargetLowering { 80 TargetLowering(const TargetLowering&); // DO NOT IMPLEMENT 81 void operator=(const TargetLowering&); // DO NOT IMPLEMENT 82 public: 83 /// LegalizeAction - This enum indicates whether operations are valid for a 84 /// target, and if not, what action should be used to make them valid. 85 enum LegalizeAction { 86 Legal, // The target natively supports this operation. 87 Promote, // This operation should be executed in a larger type. 88 Expand, // Try to expand this to other ops, otherwise use a libcall. 89 Custom // Use the LowerOperation hook to implement custom lowering. 90 }; 91 92 /// LegalizeTypeAction - This enum indicates whether a types are legal for a 93 /// target, and if not, what action should be used to make them valid. 94 enum LegalizeTypeAction { 95 TypeLegal, // The target natively supports this type. 96 TypePromoteInteger, // Replace this integer with a larger one. 97 TypeExpandInteger, // Split this integer into two of half the size. 98 TypeSoftenFloat, // Convert this float to a same size integer type. 99 TypeExpandFloat, // Split this float into two of half the size. 100 TypeScalarizeVector, // Replace this one-element vector with its element. 101 TypeSplitVector, // Split this vector into two of half the size. 102 TypeWidenVector // This vector should be widened into a larger vector. 103 }; 104 105 enum BooleanContent { // How the target represents true/false values. 106 UndefinedBooleanContent, // Only bit 0 counts, the rest can hold garbage. 107 ZeroOrOneBooleanContent, // All bits zero except for bit 0. 108 ZeroOrNegativeOneBooleanContent // All bits equal to bit 0. 109 }; 110 111 enum SelectSupportKind { 112 ScalarValSelect, // The target supports scalar selects (ex: cmov). 113 ScalarCondVectorVal, // The target supports selects with a scalar condition 114 // and vector values (ex: cmov). 115 VectorMaskSelect // The target supports vector selects with a vector 116 // mask (ex: x86 blends). 117 }; 118 getExtendForContent(BooleanContent Content)119 static ISD::NodeType getExtendForContent(BooleanContent Content) { 120 switch (Content) { 121 case UndefinedBooleanContent: 122 // Extend by adding rubbish bits. 123 return ISD::ANY_EXTEND; 124 case ZeroOrOneBooleanContent: 125 // Extend by adding zero bits. 126 return ISD::ZERO_EXTEND; 127 case ZeroOrNegativeOneBooleanContent: 128 // Extend by copying the sign bit. 129 return ISD::SIGN_EXTEND; 130 } 131 llvm_unreachable("Invalid content kind"); 132 } 133 134 /// NOTE: The constructor takes ownership of TLOF. 135 explicit TargetLowering(const TargetMachine &TM, 136 const TargetLoweringObjectFile *TLOF); 137 virtual ~TargetLowering(); 138 getTargetMachine()139 const TargetMachine &getTargetMachine() const { return TM; } getTargetData()140 const TargetData *getTargetData() const { return TD; } getObjFileLowering()141 const TargetLoweringObjectFile &getObjFileLowering() const { return TLOF; } 142 isBigEndian()143 bool isBigEndian() const { return !IsLittleEndian; } isLittleEndian()144 bool isLittleEndian() const { return IsLittleEndian; } getPointerTy()145 MVT getPointerTy() const { return PointerTy; } 146 virtual MVT getShiftAmountTy(EVT LHSTy) const; 147 148 /// isSelectExpensive - Return true if the select operation is expensive for 149 /// this target. isSelectExpensive()150 bool isSelectExpensive() const { return SelectIsExpensive; } 151 isSelectSupported(SelectSupportKind kind)152 virtual bool isSelectSupported(SelectSupportKind kind) const { return true; } 153 154 /// isIntDivCheap() - Return true if integer divide is usually cheaper than 155 /// a sequence of several shifts, adds, and multiplies for this target. isIntDivCheap()156 bool isIntDivCheap() const { return IntDivIsCheap; } 157 158 /// isSlowDivBypassed - Returns true if target has indicated at least one 159 /// type should be bypassed. isSlowDivBypassed()160 bool isSlowDivBypassed() const { return !BypassSlowDivTypes.empty(); } 161 162 /// getBypassSlowDivTypes - Returns map of slow types for division or 163 /// remainder with corresponding fast types getBypassSlowDivTypes()164 const DenseMap<Type *, Type *> &getBypassSlowDivTypes() const { 165 return BypassSlowDivTypes; 166 } 167 168 /// isPow2DivCheap() - Return true if pow2 div is cheaper than a chain of 169 /// srl/add/sra. isPow2DivCheap()170 bool isPow2DivCheap() const { return Pow2DivIsCheap; } 171 172 /// isJumpExpensive() - Return true if Flow Control is an expensive operation 173 /// that should be avoided. isJumpExpensive()174 bool isJumpExpensive() const { return JumpIsExpensive; } 175 176 /// isPredictableSelectExpensive - Return true if selects are only cheaper 177 /// than branches if the branch is unlikely to be predicted right. isPredictableSelectExpensive()178 bool isPredictableSelectExpensive() const { 179 return predictableSelectIsExpensive; 180 } 181 182 /// getSetCCResultType - Return the ValueType of the result of SETCC 183 /// operations. Also used to obtain the target's preferred type for 184 /// the condition operand of SELECT and BRCOND nodes. In the case of 185 /// BRCOND the argument passed is MVT::Other since there are no other 186 /// operands to get a type hint from. 187 virtual EVT getSetCCResultType(EVT VT) const; 188 189 /// getCmpLibcallReturnType - Return the ValueType for comparison 190 /// libcalls. Comparions libcalls include floating point comparion calls, 191 /// and Ordered/Unordered check calls on floating point numbers. 192 virtual 193 MVT::SimpleValueType getCmpLibcallReturnType() const; 194 195 /// getBooleanContents - For targets without i1 registers, this gives the 196 /// nature of the high-bits of boolean values held in types wider than i1. 197 /// "Boolean values" are special true/false values produced by nodes like 198 /// SETCC and consumed (as the condition) by nodes like SELECT and BRCOND. 199 /// Not to be confused with general values promoted from i1. 200 /// Some cpus distinguish between vectors of boolean and scalars; the isVec 201 /// parameter selects between the two kinds. For example on X86 a scalar 202 /// boolean should be zero extended from i1, while the elements of a vector 203 /// of booleans should be sign extended from i1. getBooleanContents(bool isVec)204 BooleanContent getBooleanContents(bool isVec) const { 205 return isVec ? BooleanVectorContents : BooleanContents; 206 } 207 208 /// getSchedulingPreference - Return target scheduling preference. getSchedulingPreference()209 Sched::Preference getSchedulingPreference() const { 210 return SchedPreferenceInfo; 211 } 212 213 /// getSchedulingPreference - Some scheduler, e.g. hybrid, can switch to 214 /// different scheduling heuristics for different nodes. This function returns 215 /// the preference (or none) for the given node. getSchedulingPreference(SDNode *)216 virtual Sched::Preference getSchedulingPreference(SDNode *) const { 217 return Sched::None; 218 } 219 220 /// getRegClassFor - Return the register class that should be used for the 221 /// specified value type. getRegClassFor(EVT VT)222 virtual const TargetRegisterClass *getRegClassFor(EVT VT) const { 223 assert(VT.isSimple() && "getRegClassFor called on illegal type!"); 224 const TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT().SimpleTy]; 225 assert(RC && "This value type is not natively supported!"); 226 return RC; 227 } 228 229 /// getRepRegClassFor - Return the 'representative' register class for the 230 /// specified value type. The 'representative' register class is the largest 231 /// legal super-reg register class for the register class of the value type. 232 /// For example, on i386 the rep register class for i8, i16, and i32 are GR32; 233 /// while the rep register class is GR64 on x86_64. getRepRegClassFor(EVT VT)234 virtual const TargetRegisterClass *getRepRegClassFor(EVT VT) const { 235 assert(VT.isSimple() && "getRepRegClassFor called on illegal type!"); 236 const TargetRegisterClass *RC = RepRegClassForVT[VT.getSimpleVT().SimpleTy]; 237 return RC; 238 } 239 240 /// getRepRegClassCostFor - Return the cost of the 'representative' register 241 /// class for the specified value type. getRepRegClassCostFor(EVT VT)242 virtual uint8_t getRepRegClassCostFor(EVT VT) const { 243 assert(VT.isSimple() && "getRepRegClassCostFor called on illegal type!"); 244 return RepRegClassCostForVT[VT.getSimpleVT().SimpleTy]; 245 } 246 247 /// isTypeLegal - Return true if the target has native support for the 248 /// specified value type. This means that it has a register that directly 249 /// holds it without promotions or expansions. isTypeLegal(EVT VT)250 bool isTypeLegal(EVT VT) const { 251 assert(!VT.isSimple() || 252 (unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT)); 253 return VT.isSimple() && RegClassForVT[VT.getSimpleVT().SimpleTy] != 0; 254 } 255 256 class ValueTypeActionImpl { 257 /// ValueTypeActions - For each value type, keep a LegalizeTypeAction enum 258 /// that indicates how instruction selection should deal with the type. 259 uint8_t ValueTypeActions[MVT::LAST_VALUETYPE]; 260 261 public: ValueTypeActionImpl()262 ValueTypeActionImpl() { 263 std::fill(ValueTypeActions, array_endof(ValueTypeActions), 0); 264 } 265 getTypeAction(MVT VT)266 LegalizeTypeAction getTypeAction(MVT VT) const { 267 return (LegalizeTypeAction)ValueTypeActions[VT.SimpleTy]; 268 } 269 setTypeAction(EVT VT,LegalizeTypeAction Action)270 void setTypeAction(EVT VT, LegalizeTypeAction Action) { 271 unsigned I = VT.getSimpleVT().SimpleTy; 272 ValueTypeActions[I] = Action; 273 } 274 }; 275 getValueTypeActions()276 const ValueTypeActionImpl &getValueTypeActions() const { 277 return ValueTypeActions; 278 } 279 280 /// getTypeAction - Return how we should legalize values of this type, either 281 /// it is already legal (return 'Legal') or we need to promote it to a larger 282 /// type (return 'Promote'), or we need to expand it into multiple registers 283 /// of smaller integer type (return 'Expand'). 'Custom' is not an option. getTypeAction(LLVMContext & Context,EVT VT)284 LegalizeTypeAction getTypeAction(LLVMContext &Context, EVT VT) const { 285 return getTypeConversion(Context, VT).first; 286 } getTypeAction(MVT VT)287 LegalizeTypeAction getTypeAction(MVT VT) const { 288 return ValueTypeActions.getTypeAction(VT); 289 } 290 291 /// getTypeToTransformTo - For types supported by the target, this is an 292 /// identity function. For types that must be promoted to larger types, this 293 /// returns the larger type to promote to. For integer types that are larger 294 /// than the largest integer register, this contains one step in the expansion 295 /// to get to the smaller register. For illegal floating point types, this 296 /// returns the integer type to transform to. getTypeToTransformTo(LLVMContext & Context,EVT VT)297 EVT getTypeToTransformTo(LLVMContext &Context, EVT VT) const { 298 return getTypeConversion(Context, VT).second; 299 } 300 301 /// getTypeToExpandTo - For types supported by the target, this is an 302 /// identity function. For types that must be expanded (i.e. integer types 303 /// that are larger than the largest integer register or illegal floating 304 /// point types), this returns the largest legal type it will be expanded to. getTypeToExpandTo(LLVMContext & Context,EVT VT)305 EVT getTypeToExpandTo(LLVMContext &Context, EVT VT) const { 306 assert(!VT.isVector()); 307 while (true) { 308 switch (getTypeAction(Context, VT)) { 309 case TypeLegal: 310 return VT; 311 case TypeExpandInteger: 312 VT = getTypeToTransformTo(Context, VT); 313 break; 314 default: 315 llvm_unreachable("Type is not legal nor is it to be expanded!"); 316 } 317 } 318 } 319 320 /// getVectorTypeBreakdown - Vector types are broken down into some number of 321 /// legal first class types. For example, EVT::v8f32 maps to 2 EVT::v4f32 322 /// with Altivec or SSE1, or 8 promoted EVT::f64 values with the X86 FP stack. 323 /// Similarly, EVT::v2i64 turns into 4 EVT::i32 values with both PPC and X86. 324 /// 325 /// This method returns the number of registers needed, and the VT for each 326 /// register. It also returns the VT and quantity of the intermediate values 327 /// before they are promoted/expanded. 328 /// 329 unsigned getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 330 EVT &IntermediateVT, 331 unsigned &NumIntermediates, 332 EVT &RegisterVT) const; 333 334 /// getTgtMemIntrinsic: Given an intrinsic, checks if on the target the 335 /// intrinsic will need to map to a MemIntrinsicNode (touches memory). If 336 /// this is the case, it returns true and store the intrinsic 337 /// information into the IntrinsicInfo that was passed to the function. 338 struct IntrinsicInfo { 339 unsigned opc; // target opcode 340 EVT memVT; // memory VT 341 const Value* ptrVal; // value representing memory location 342 int offset; // offset off of ptrVal 343 unsigned align; // alignment 344 bool vol; // is volatile? 345 bool readMem; // reads memory? 346 bool writeMem; // writes memory? 347 }; 348 getTgtMemIntrinsic(IntrinsicInfo &,const CallInst &,unsigned)349 virtual bool getTgtMemIntrinsic(IntrinsicInfo &, const CallInst &, 350 unsigned /*Intrinsic*/) const { 351 return false; 352 } 353 354 /// isFPImmLegal - Returns true if the target can instruction select the 355 /// specified FP immediate natively. If false, the legalizer will materialize 356 /// the FP immediate as a load from a constant pool. isFPImmLegal(const APFloat &,EVT)357 virtual bool isFPImmLegal(const APFloat &/*Imm*/, EVT /*VT*/) const { 358 return false; 359 } 360 361 /// isShuffleMaskLegal - Targets can use this to indicate that they only 362 /// support *some* VECTOR_SHUFFLE operations, those with specific masks. 363 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values 364 /// are assumed to be legal. isShuffleMaskLegal(const SmallVectorImpl<int> &,EVT)365 virtual bool isShuffleMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 366 EVT /*VT*/) const { 367 return true; 368 } 369 370 /// canOpTrap - Returns true if the operation can trap for the value type. 371 /// VT must be a legal type. By default, we optimistically assume most 372 /// operations don't trap except for divide and remainder. 373 virtual bool canOpTrap(unsigned Op, EVT VT) const; 374 375 /// isVectorClearMaskLegal - Similar to isShuffleMaskLegal. This is 376 /// used by Targets can use this to indicate if there is a suitable 377 /// VECTOR_SHUFFLE that can be used to replace a VAND with a constant 378 /// pool entry. isVectorClearMaskLegal(const SmallVectorImpl<int> &,EVT)379 virtual bool isVectorClearMaskLegal(const SmallVectorImpl<int> &/*Mask*/, 380 EVT /*VT*/) const { 381 return false; 382 } 383 384 /// getOperationAction - Return how this operation should be treated: either 385 /// it is legal, needs to be promoted to a larger size, needs to be 386 /// expanded to some other code sequence, or the target has a custom expander 387 /// for it. getOperationAction(unsigned Op,EVT VT)388 LegalizeAction getOperationAction(unsigned Op, EVT VT) const { 389 if (VT.isExtended()) return Expand; 390 // If a target-specific SDNode requires legalization, require the target 391 // to provide custom legalization for it. 392 if (Op > array_lengthof(OpActions[0])) return Custom; 393 unsigned I = (unsigned) VT.getSimpleVT().SimpleTy; 394 return (LegalizeAction)OpActions[I][Op]; 395 } 396 397 /// isOperationLegalOrCustom - Return true if the specified operation is 398 /// legal on this target or can be made legal with custom lowering. This 399 /// is used to help guide high-level lowering decisions. isOperationLegalOrCustom(unsigned Op,EVT VT)400 bool isOperationLegalOrCustom(unsigned Op, EVT VT) const { 401 return (VT == MVT::Other || isTypeLegal(VT)) && 402 (getOperationAction(Op, VT) == Legal || 403 getOperationAction(Op, VT) == Custom); 404 } 405 406 /// isOperationLegal - Return true if the specified operation is legal on this 407 /// target. isOperationLegal(unsigned Op,EVT VT)408 bool isOperationLegal(unsigned Op, EVT VT) const { 409 return (VT == MVT::Other || isTypeLegal(VT)) && 410 getOperationAction(Op, VT) == Legal; 411 } 412 413 /// getLoadExtAction - Return how this load with extension should be treated: 414 /// either it is legal, needs to be promoted to a larger size, needs to be 415 /// expanded to some other code sequence, or the target has a custom expander 416 /// for it. getLoadExtAction(unsigned ExtType,EVT VT)417 LegalizeAction getLoadExtAction(unsigned ExtType, EVT VT) const { 418 assert(ExtType < ISD::LAST_LOADEXT_TYPE && 419 VT.getSimpleVT() < MVT::LAST_VALUETYPE && 420 "Table isn't big enough!"); 421 return (LegalizeAction)LoadExtActions[VT.getSimpleVT().SimpleTy][ExtType]; 422 } 423 424 /// isLoadExtLegal - Return true if the specified load with extension is legal 425 /// on this target. isLoadExtLegal(unsigned ExtType,EVT VT)426 bool isLoadExtLegal(unsigned ExtType, EVT VT) const { 427 return VT.isSimple() && getLoadExtAction(ExtType, VT) == Legal; 428 } 429 430 /// getTruncStoreAction - Return how this store with truncation should be 431 /// treated: either it is legal, needs to be promoted to a larger size, needs 432 /// to be expanded to some other code sequence, or the target has a custom 433 /// expander for it. getTruncStoreAction(EVT ValVT,EVT MemVT)434 LegalizeAction getTruncStoreAction(EVT ValVT, EVT MemVT) const { 435 assert(ValVT.getSimpleVT() < MVT::LAST_VALUETYPE && 436 MemVT.getSimpleVT() < MVT::LAST_VALUETYPE && 437 "Table isn't big enough!"); 438 return (LegalizeAction)TruncStoreActions[ValVT.getSimpleVT().SimpleTy] 439 [MemVT.getSimpleVT().SimpleTy]; 440 } 441 442 /// isTruncStoreLegal - Return true if the specified store with truncation is 443 /// legal on this target. isTruncStoreLegal(EVT ValVT,EVT MemVT)444 bool isTruncStoreLegal(EVT ValVT, EVT MemVT) const { 445 return isTypeLegal(ValVT) && MemVT.isSimple() && 446 getTruncStoreAction(ValVT, MemVT) == Legal; 447 } 448 449 /// getIndexedLoadAction - Return how the indexed load should be treated: 450 /// either it is legal, needs to be promoted to a larger size, needs to be 451 /// expanded to some other code sequence, or the target has a custom expander 452 /// for it. 453 LegalizeAction getIndexedLoadAction(unsigned IdxMode,EVT VT)454 getIndexedLoadAction(unsigned IdxMode, EVT VT) const { 455 assert(IdxMode < ISD::LAST_INDEXED_MODE && 456 VT.getSimpleVT() < MVT::LAST_VALUETYPE && 457 "Table isn't big enough!"); 458 unsigned Ty = (unsigned)VT.getSimpleVT().SimpleTy; 459 return (LegalizeAction)((IndexedModeActions[Ty][IdxMode] & 0xf0) >> 4); 460 } 461 462 /// isIndexedLoadLegal - Return true if the specified indexed load is legal 463 /// on this target. isIndexedLoadLegal(unsigned IdxMode,EVT VT)464 bool isIndexedLoadLegal(unsigned IdxMode, EVT VT) const { 465 return VT.isSimple() && 466 (getIndexedLoadAction(IdxMode, VT) == Legal || 467 getIndexedLoadAction(IdxMode, VT) == Custom); 468 } 469 470 /// getIndexedStoreAction - Return how the indexed store should be treated: 471 /// either it is legal, needs to be promoted to a larger size, needs to be 472 /// expanded to some other code sequence, or the target has a custom expander 473 /// for it. 474 LegalizeAction getIndexedStoreAction(unsigned IdxMode,EVT VT)475 getIndexedStoreAction(unsigned IdxMode, EVT VT) const { 476 assert(IdxMode < ISD::LAST_INDEXED_MODE && 477 VT.getSimpleVT() < MVT::LAST_VALUETYPE && 478 "Table isn't big enough!"); 479 unsigned Ty = (unsigned)VT.getSimpleVT().SimpleTy; 480 return (LegalizeAction)(IndexedModeActions[Ty][IdxMode] & 0x0f); 481 } 482 483 /// isIndexedStoreLegal - Return true if the specified indexed load is legal 484 /// on this target. isIndexedStoreLegal(unsigned IdxMode,EVT VT)485 bool isIndexedStoreLegal(unsigned IdxMode, EVT VT) const { 486 return VT.isSimple() && 487 (getIndexedStoreAction(IdxMode, VT) == Legal || 488 getIndexedStoreAction(IdxMode, VT) == Custom); 489 } 490 491 /// getCondCodeAction - Return how the condition code should be treated: 492 /// either it is legal, needs to be expanded to some other code sequence, 493 /// or the target has a custom expander for it. 494 LegalizeAction getCondCodeAction(ISD::CondCode CC,EVT VT)495 getCondCodeAction(ISD::CondCode CC, EVT VT) const { 496 assert((unsigned)CC < array_lengthof(CondCodeActions) && 497 (unsigned)VT.getSimpleVT().SimpleTy < sizeof(CondCodeActions[0])*4 && 498 "Table isn't big enough!"); 499 LegalizeAction Action = (LegalizeAction) 500 ((CondCodeActions[CC] >> (2*VT.getSimpleVT().SimpleTy)) & 3); 501 assert(Action != Promote && "Can't promote condition code!"); 502 return Action; 503 } 504 505 /// isCondCodeLegal - Return true if the specified condition code is legal 506 /// on this target. isCondCodeLegal(ISD::CondCode CC,EVT VT)507 bool isCondCodeLegal(ISD::CondCode CC, EVT VT) const { 508 return getCondCodeAction(CC, VT) == Legal || 509 getCondCodeAction(CC, VT) == Custom; 510 } 511 512 513 /// getTypeToPromoteTo - If the action for this operation is to promote, this 514 /// method returns the ValueType to promote to. getTypeToPromoteTo(unsigned Op,EVT VT)515 EVT getTypeToPromoteTo(unsigned Op, EVT VT) const { 516 assert(getOperationAction(Op, VT) == Promote && 517 "This operation isn't promoted!"); 518 519 // See if this has an explicit type specified. 520 std::map<std::pair<unsigned, MVT::SimpleValueType>, 521 MVT::SimpleValueType>::const_iterator PTTI = 522 PromoteToType.find(std::make_pair(Op, VT.getSimpleVT().SimpleTy)); 523 if (PTTI != PromoteToType.end()) return PTTI->second; 524 525 assert((VT.isInteger() || VT.isFloatingPoint()) && 526 "Cannot autopromote this type, add it with AddPromotedToType."); 527 528 EVT NVT = VT; 529 do { 530 NVT = (MVT::SimpleValueType)(NVT.getSimpleVT().SimpleTy+1); 531 assert(NVT.isInteger() == VT.isInteger() && NVT != MVT::isVoid && 532 "Didn't find type to promote to!"); 533 } while (!isTypeLegal(NVT) || 534 getOperationAction(Op, NVT) == Promote); 535 return NVT; 536 } 537 538 /// getValueType - Return the EVT corresponding to this LLVM type. 539 /// This is fixed by the LLVM operations except for the pointer size. If 540 /// AllowUnknown is true, this will return MVT::Other for types with no EVT 541 /// counterpart (e.g. structs), otherwise it will assert. 542 EVT getValueType(Type *Ty, bool AllowUnknown = false) const { 543 // Lower scalar pointers to native pointer types. 544 if (Ty->isPointerTy()) return PointerTy; 545 546 if (Ty->isVectorTy()) { 547 VectorType *VTy = cast<VectorType>(Ty); 548 Type *Elm = VTy->getElementType(); 549 // Lower vectors of pointers to native pointer types. 550 if (Elm->isPointerTy()) 551 Elm = EVT(PointerTy).getTypeForEVT(Ty->getContext()); 552 return EVT::getVectorVT(Ty->getContext(), EVT::getEVT(Elm, false), 553 VTy->getNumElements()); 554 } 555 return EVT::getEVT(Ty, AllowUnknown); 556 } 557 558 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 559 /// function arguments in the caller parameter area. This is the actual 560 /// alignment, not its logarithm. 561 virtual unsigned getByValTypeAlignment(Type *Ty) const; 562 563 /// getRegisterType - Return the type of registers that this ValueType will 564 /// eventually require. getRegisterType(MVT VT)565 EVT getRegisterType(MVT VT) const { 566 assert((unsigned)VT.SimpleTy < array_lengthof(RegisterTypeForVT)); 567 return RegisterTypeForVT[VT.SimpleTy]; 568 } 569 570 /// getRegisterType - Return the type of registers that this ValueType will 571 /// eventually require. getRegisterType(LLVMContext & Context,EVT VT)572 EVT getRegisterType(LLVMContext &Context, EVT VT) const { 573 if (VT.isSimple()) { 574 assert((unsigned)VT.getSimpleVT().SimpleTy < 575 array_lengthof(RegisterTypeForVT)); 576 return RegisterTypeForVT[VT.getSimpleVT().SimpleTy]; 577 } 578 if (VT.isVector()) { 579 EVT VT1, RegisterVT; 580 unsigned NumIntermediates; 581 (void)getVectorTypeBreakdown(Context, VT, VT1, 582 NumIntermediates, RegisterVT); 583 return RegisterVT; 584 } 585 if (VT.isInteger()) { 586 return getRegisterType(Context, getTypeToTransformTo(Context, VT)); 587 } 588 llvm_unreachable("Unsupported extended type!"); 589 } 590 591 /// getNumRegisters - Return the number of registers that this ValueType will 592 /// eventually require. This is one for any types promoted to live in larger 593 /// registers, but may be more than one for types (like i64) that are split 594 /// into pieces. For types like i140, which are first promoted then expanded, 595 /// it is the number of registers needed to hold all the bits of the original 596 /// type. For an i140 on a 32 bit machine this means 5 registers. getNumRegisters(LLVMContext & Context,EVT VT)597 unsigned getNumRegisters(LLVMContext &Context, EVT VT) const { 598 if (VT.isSimple()) { 599 assert((unsigned)VT.getSimpleVT().SimpleTy < 600 array_lengthof(NumRegistersForVT)); 601 return NumRegistersForVT[VT.getSimpleVT().SimpleTy]; 602 } 603 if (VT.isVector()) { 604 EVT VT1, VT2; 605 unsigned NumIntermediates; 606 return getVectorTypeBreakdown(Context, VT, VT1, NumIntermediates, VT2); 607 } 608 if (VT.isInteger()) { 609 unsigned BitWidth = VT.getSizeInBits(); 610 unsigned RegWidth = getRegisterType(Context, VT).getSizeInBits(); 611 return (BitWidth + RegWidth - 1) / RegWidth; 612 } 613 llvm_unreachable("Unsupported extended type!"); 614 } 615 616 /// ShouldShrinkFPConstant - If true, then instruction selection should 617 /// seek to shrink the FP constant of the specified type to a smaller type 618 /// in order to save space and / or reduce runtime. ShouldShrinkFPConstant(EVT)619 virtual bool ShouldShrinkFPConstant(EVT) const { return true; } 620 621 /// hasTargetDAGCombine - If true, the target has custom DAG combine 622 /// transformations that it can perform for the specified node. hasTargetDAGCombine(ISD::NodeType NT)623 bool hasTargetDAGCombine(ISD::NodeType NT) const { 624 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 625 return TargetDAGCombineArray[NT >> 3] & (1 << (NT&7)); 626 } 627 628 /// This function returns the maximum number of store operations permitted 629 /// to replace a call to llvm.memset. The value is set by the target at the 630 /// performance threshold for such a replacement. If OptSize is true, 631 /// return the limit for functions that have OptSize attribute. 632 /// @brief Get maximum # of store operations permitted for llvm.memset getMaxStoresPerMemset(bool OptSize)633 unsigned getMaxStoresPerMemset(bool OptSize) const { 634 return OptSize ? maxStoresPerMemsetOptSize : maxStoresPerMemset; 635 } 636 637 /// This function returns the maximum number of store operations permitted 638 /// to replace a call to llvm.memcpy. The value is set by the target at the 639 /// performance threshold for such a replacement. If OptSize is true, 640 /// return the limit for functions that have OptSize attribute. 641 /// @brief Get maximum # of store operations permitted for llvm.memcpy getMaxStoresPerMemcpy(bool OptSize)642 unsigned getMaxStoresPerMemcpy(bool OptSize) const { 643 return OptSize ? maxStoresPerMemcpyOptSize : maxStoresPerMemcpy; 644 } 645 646 /// This function returns the maximum number of store operations permitted 647 /// to replace a call to llvm.memmove. The value is set by the target at the 648 /// performance threshold for such a replacement. If OptSize is true, 649 /// return the limit for functions that have OptSize attribute. 650 /// @brief Get maximum # of store operations permitted for llvm.memmove getMaxStoresPerMemmove(bool OptSize)651 unsigned getMaxStoresPerMemmove(bool OptSize) const { 652 return OptSize ? maxStoresPerMemmoveOptSize : maxStoresPerMemmove; 653 } 654 655 /// This function returns true if the target allows unaligned memory accesses. 656 /// of the specified type. This is used, for example, in situations where an 657 /// array copy/move/set is converted to a sequence of store operations. It's 658 /// use helps to ensure that such replacements don't generate code that causes 659 /// an alignment error (trap) on the target machine. 660 /// @brief Determine if the target supports unaligned memory accesses. allowsUnalignedMemoryAccesses(EVT)661 virtual bool allowsUnalignedMemoryAccesses(EVT) const { 662 return false; 663 } 664 665 /// This function returns true if the target would benefit from code placement 666 /// optimization. 667 /// @brief Determine if the target should perform code placement optimization. shouldOptimizeCodePlacement()668 bool shouldOptimizeCodePlacement() const { 669 return benefitFromCodePlacementOpt; 670 } 671 672 /// getOptimalMemOpType - Returns the target specific optimal type for load 673 /// and store operations as a result of memset, memcpy, and memmove 674 /// lowering. If DstAlign is zero that means it's safe to destination 675 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it 676 /// means there isn't a need to check it against alignment requirement, 677 /// probably because the source does not need to be loaded. If 678 /// 'IsZeroVal' is true, that means it's safe to return a 679 /// non-scalar-integer type, e.g. empty string source, constant, or loaded 680 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is 681 /// constant so it does not need to be loaded. 682 /// It returns EVT::Other if the type should be determined using generic 683 /// target-independent logic. getOptimalMemOpType(uint64_t,unsigned,unsigned,bool,bool,MachineFunction &)684 virtual EVT getOptimalMemOpType(uint64_t /*Size*/, 685 unsigned /*DstAlign*/, unsigned /*SrcAlign*/, 686 bool /*IsZeroVal*/, 687 bool /*MemcpyStrSrc*/, 688 MachineFunction &/*MF*/) const { 689 return MVT::Other; 690 } 691 692 /// usesUnderscoreSetJmp - Determine if we should use _setjmp or setjmp 693 /// to implement llvm.setjmp. usesUnderscoreSetJmp()694 bool usesUnderscoreSetJmp() const { 695 return UseUnderscoreSetJmp; 696 } 697 698 /// usesUnderscoreLongJmp - Determine if we should use _longjmp or longjmp 699 /// to implement llvm.longjmp. usesUnderscoreLongJmp()700 bool usesUnderscoreLongJmp() const { 701 return UseUnderscoreLongJmp; 702 } 703 704 /// supportJumpTables - return whether the target can generate code for 705 /// jump tables. supportJumpTables()706 bool supportJumpTables() const { 707 return SupportJumpTables; 708 } 709 710 /// getStackPointerRegisterToSaveRestore - If a physical register, this 711 /// specifies the register that llvm.savestack/llvm.restorestack should save 712 /// and restore. getStackPointerRegisterToSaveRestore()713 unsigned getStackPointerRegisterToSaveRestore() const { 714 return StackPointerRegisterToSaveRestore; 715 } 716 717 /// getExceptionPointerRegister - If a physical register, this returns 718 /// the register that receives the exception address on entry to a landing 719 /// pad. getExceptionPointerRegister()720 unsigned getExceptionPointerRegister() const { 721 return ExceptionPointerRegister; 722 } 723 724 /// getExceptionSelectorRegister - If a physical register, this returns 725 /// the register that receives the exception typeid on entry to a landing 726 /// pad. getExceptionSelectorRegister()727 unsigned getExceptionSelectorRegister() const { 728 return ExceptionSelectorRegister; 729 } 730 731 /// getJumpBufSize - returns the target's jmp_buf size in bytes (if never 732 /// set, the default is 200) getJumpBufSize()733 unsigned getJumpBufSize() const { 734 return JumpBufSize; 735 } 736 737 /// getJumpBufAlignment - returns the target's jmp_buf alignment in bytes 738 /// (if never set, the default is 0) getJumpBufAlignment()739 unsigned getJumpBufAlignment() const { 740 return JumpBufAlignment; 741 } 742 743 /// getMinStackArgumentAlignment - return the minimum stack alignment of an 744 /// argument. getMinStackArgumentAlignment()745 unsigned getMinStackArgumentAlignment() const { 746 return MinStackArgumentAlignment; 747 } 748 749 /// getMinFunctionAlignment - return the minimum function alignment. 750 /// getMinFunctionAlignment()751 unsigned getMinFunctionAlignment() const { 752 return MinFunctionAlignment; 753 } 754 755 /// getPrefFunctionAlignment - return the preferred function alignment. 756 /// getPrefFunctionAlignment()757 unsigned getPrefFunctionAlignment() const { 758 return PrefFunctionAlignment; 759 } 760 761 /// getPrefLoopAlignment - return the preferred loop alignment. 762 /// getPrefLoopAlignment()763 unsigned getPrefLoopAlignment() const { 764 return PrefLoopAlignment; 765 } 766 767 /// getShouldFoldAtomicFences - return whether the combiner should fold 768 /// fence MEMBARRIER instructions into the atomic intrinsic instructions. 769 /// getShouldFoldAtomicFences()770 bool getShouldFoldAtomicFences() const { 771 return ShouldFoldAtomicFences; 772 } 773 774 /// getInsertFencesFor - return whether the DAG builder should automatically 775 /// insert fences and reduce ordering for atomics. 776 /// getInsertFencesForAtomic()777 bool getInsertFencesForAtomic() const { 778 return InsertFencesForAtomic; 779 } 780 781 /// getPreIndexedAddressParts - returns true by value, base pointer and 782 /// offset pointer and addressing mode by reference if the node's address 783 /// can be legally represented as pre-indexed load / store address. getPreIndexedAddressParts(SDNode *,SDValue &,SDValue &,ISD::MemIndexedMode &,SelectionDAG &)784 virtual bool getPreIndexedAddressParts(SDNode * /*N*/, SDValue &/*Base*/, 785 SDValue &/*Offset*/, 786 ISD::MemIndexedMode &/*AM*/, 787 SelectionDAG &/*DAG*/) const { 788 return false; 789 } 790 791 /// getPostIndexedAddressParts - returns true by value, base pointer and 792 /// offset pointer and addressing mode by reference if this node can be 793 /// combined with a load / store to form a post-indexed load / store. getPostIndexedAddressParts(SDNode *,SDNode *,SDValue &,SDValue &,ISD::MemIndexedMode &,SelectionDAG &)794 virtual bool getPostIndexedAddressParts(SDNode * /*N*/, SDNode * /*Op*/, 795 SDValue &/*Base*/, SDValue &/*Offset*/, 796 ISD::MemIndexedMode &/*AM*/, 797 SelectionDAG &/*DAG*/) const { 798 return false; 799 } 800 801 /// getJumpTableEncoding - Return the entry encoding for a jump table in the 802 /// current function. The returned value is a member of the 803 /// MachineJumpTableInfo::JTEntryKind enum. 804 virtual unsigned getJumpTableEncoding() const; 805 806 virtual const MCExpr * LowerCustomJumpTableEntry(const MachineJumpTableInfo *,const MachineBasicBlock *,unsigned,MCContext &)807 LowerCustomJumpTableEntry(const MachineJumpTableInfo * /*MJTI*/, 808 const MachineBasicBlock * /*MBB*/, unsigned /*uid*/, 809 MCContext &/*Ctx*/) const { 810 llvm_unreachable("Need to implement this hook if target has custom JTIs"); 811 } 812 813 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC 814 /// jumptable. 815 virtual SDValue getPICJumpTableRelocBase(SDValue Table, 816 SelectionDAG &DAG) const; 817 818 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the 819 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an 820 /// MCExpr. 821 virtual const MCExpr * 822 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 823 unsigned JTI, MCContext &Ctx) const; 824 825 /// isOffsetFoldingLegal - Return true if folding a constant offset 826 /// with the given GlobalAddress is legal. It is frequently not legal in 827 /// PIC relocation models. 828 virtual bool isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const; 829 830 /// getStackCookieLocation - Return true if the target stores stack 831 /// protector cookies at a fixed offset in some non-standard address 832 /// space, and populates the address space and offset as 833 /// appropriate. getStackCookieLocation(unsigned &,unsigned &)834 virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/, 835 unsigned &/*Offset*/) const { 836 return false; 837 } 838 839 /// getMaximalGlobalOffset - Returns the maximal possible offset which can be 840 /// used for loads / stores from the global. getMaximalGlobalOffset()841 virtual unsigned getMaximalGlobalOffset() const { 842 return 0; 843 } 844 845 //===--------------------------------------------------------------------===// 846 // TargetLowering Optimization Methods 847 // 848 849 /// TargetLoweringOpt - A convenience struct that encapsulates a DAG, and two 850 /// SDValues for returning information from TargetLowering to its clients 851 /// that want to combine 852 struct TargetLoweringOpt { 853 SelectionDAG &DAG; 854 bool LegalTys; 855 bool LegalOps; 856 SDValue Old; 857 SDValue New; 858 TargetLoweringOptTargetLoweringOpt859 explicit TargetLoweringOpt(SelectionDAG &InDAG, 860 bool LT, bool LO) : 861 DAG(InDAG), LegalTys(LT), LegalOps(LO) {} 862 LegalTypesTargetLoweringOpt863 bool LegalTypes() const { return LegalTys; } LegalOperationsTargetLoweringOpt864 bool LegalOperations() const { return LegalOps; } 865 CombineToTargetLoweringOpt866 bool CombineTo(SDValue O, SDValue N) { 867 Old = O; 868 New = N; 869 return true; 870 } 871 872 /// ShrinkDemandedConstant - Check to see if the specified operand of the 873 /// specified instruction is a constant integer. If so, check to see if 874 /// there are any bits set in the constant that are not demanded. If so, 875 /// shrink the constant and return true. 876 bool ShrinkDemandedConstant(SDValue Op, const APInt &Demanded); 877 878 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the 879 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening 880 /// cast, but it could be generalized for targets with other types of 881 /// implicit widening casts. 882 bool ShrinkDemandedOp(SDValue Op, unsigned BitWidth, const APInt &Demanded, 883 DebugLoc dl); 884 }; 885 886 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the 887 /// DemandedMask bits of the result of Op are ever used downstream. If we can 888 /// use this information to simplify Op, create a new simplified DAG node and 889 /// return true, returning the original and new nodes in Old and New. 890 /// Otherwise, analyze the expression and return a mask of KnownOne and 891 /// KnownZero bits for the expression (used to simplify the caller). 892 /// The KnownZero/One bits may only be accurate for those bits in the 893 /// DemandedMask. 894 bool SimplifyDemandedBits(SDValue Op, const APInt &DemandedMask, 895 APInt &KnownZero, APInt &KnownOne, 896 TargetLoweringOpt &TLO, unsigned Depth = 0) const; 897 898 /// computeMaskedBitsForTargetNode - Determine which of the bits specified in 899 /// Mask are known to be either zero or one and return them in the 900 /// KnownZero/KnownOne bitsets. 901 virtual void computeMaskedBitsForTargetNode(const SDValue Op, 902 APInt &KnownZero, 903 APInt &KnownOne, 904 const SelectionDAG &DAG, 905 unsigned Depth = 0) const; 906 907 /// ComputeNumSignBitsForTargetNode - This method can be implemented by 908 /// targets that want to expose additional information about sign bits to the 909 /// DAG Combiner. 910 virtual unsigned ComputeNumSignBitsForTargetNode(SDValue Op, 911 unsigned Depth = 0) const; 912 913 struct DAGCombinerInfo { 914 void *DC; // The DAG Combiner object. 915 bool BeforeLegalize; 916 bool BeforeLegalizeOps; 917 bool CalledByLegalizer; 918 public: 919 SelectionDAG &DAG; 920 DAGCombinerInfoDAGCombinerInfo921 DAGCombinerInfo(SelectionDAG &dag, bool bl, bool blo, bool cl, void *dc) 922 : DC(dc), BeforeLegalize(bl), BeforeLegalizeOps(blo), 923 CalledByLegalizer(cl), DAG(dag) {} 924 isBeforeLegalizeDAGCombinerInfo925 bool isBeforeLegalize() const { return BeforeLegalize; } isBeforeLegalizeOpsDAGCombinerInfo926 bool isBeforeLegalizeOps() const { return BeforeLegalizeOps; } isCalledByLegalizerDAGCombinerInfo927 bool isCalledByLegalizer() const { return CalledByLegalizer; } 928 929 void AddToWorklist(SDNode *N); 930 void RemoveFromWorklist(SDNode *N); 931 SDValue CombineTo(SDNode *N, const std::vector<SDValue> &To, 932 bool AddTo = true); 933 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true); 934 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo = true); 935 936 void CommitTargetLoweringOpt(const TargetLoweringOpt &TLO); 937 }; 938 939 /// SimplifySetCC - Try to simplify a setcc built with the specified operands 940 /// and cc. If it is unable to simplify it, return a null SDValue. 941 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 942 ISD::CondCode Cond, bool foldBooleans, 943 DAGCombinerInfo &DCI, DebugLoc dl) const; 944 945 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 946 /// node is a GlobalAddress + offset. 947 virtual bool 948 isGAPlusOffset(SDNode *N, const GlobalValue* &GA, int64_t &Offset) const; 949 950 /// PerformDAGCombine - This method will be invoked for all target nodes and 951 /// for any target-independent nodes that the target has registered with 952 /// invoke it for. 953 /// 954 /// The semantics are as follows: 955 /// Return Value: 956 /// SDValue.Val == 0 - No change was made 957 /// SDValue.Val == N - N was replaced, is dead, and is already handled. 958 /// otherwise - N should be replaced by the returned Operand. 959 /// 960 /// In addition, methods provided by DAGCombinerInfo may be used to perform 961 /// more complex transformations. 962 /// 963 virtual SDValue PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const; 964 965 /// isTypeDesirableForOp - Return true if the target has native support for 966 /// the specified value type and it is 'desirable' to use the type for the 967 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16 968 /// instruction encodings are longer and some i16 instructions are slow. isTypeDesirableForOp(unsigned,EVT VT)969 virtual bool isTypeDesirableForOp(unsigned /*Opc*/, EVT VT) const { 970 // By default, assume all legal types are desirable. 971 return isTypeLegal(VT); 972 } 973 974 /// isDesirableToPromoteOp - Return true if it is profitable for dag combiner 975 /// to transform a floating point op of specified opcode to a equivalent op of 976 /// an integer type. e.g. f32 load -> i32 load can be profitable on ARM. isDesirableToTransformToIntegerOp(unsigned,EVT)977 virtual bool isDesirableToTransformToIntegerOp(unsigned /*Opc*/, 978 EVT /*VT*/) const { 979 return false; 980 } 981 982 /// IsDesirableToPromoteOp - This method query the target whether it is 983 /// beneficial for dag combiner to promote the specified node. If true, it 984 /// should return the desired promotion type by reference. IsDesirableToPromoteOp(SDValue,EVT &)985 virtual bool IsDesirableToPromoteOp(SDValue /*Op*/, EVT &/*PVT*/) const { 986 return false; 987 } 988 989 //===--------------------------------------------------------------------===// 990 // TargetLowering Configuration Methods - These methods should be invoked by 991 // the derived class constructor to configure this object for the target. 992 // 993 994 protected: 995 /// setBooleanContents - Specify how the target extends the result of a 996 /// boolean value from i1 to a wider type. See getBooleanContents. setBooleanContents(BooleanContent Ty)997 void setBooleanContents(BooleanContent Ty) { BooleanContents = Ty; } 998 /// setBooleanVectorContents - Specify how the target extends the result 999 /// of a vector boolean value from a vector of i1 to a wider type. See 1000 /// getBooleanContents. setBooleanVectorContents(BooleanContent Ty)1001 void setBooleanVectorContents(BooleanContent Ty) { 1002 BooleanVectorContents = Ty; 1003 } 1004 1005 /// setSchedulingPreference - Specify the target scheduling preference. setSchedulingPreference(Sched::Preference Pref)1006 void setSchedulingPreference(Sched::Preference Pref) { 1007 SchedPreferenceInfo = Pref; 1008 } 1009 1010 /// setUseUnderscoreSetJmp - Indicate whether this target prefers to 1011 /// use _setjmp to implement llvm.setjmp or the non _ version. 1012 /// Defaults to false. setUseUnderscoreSetJmp(bool Val)1013 void setUseUnderscoreSetJmp(bool Val) { 1014 UseUnderscoreSetJmp = Val; 1015 } 1016 1017 /// setUseUnderscoreLongJmp - Indicate whether this target prefers to 1018 /// use _longjmp to implement llvm.longjmp or the non _ version. 1019 /// Defaults to false. setUseUnderscoreLongJmp(bool Val)1020 void setUseUnderscoreLongJmp(bool Val) { 1021 UseUnderscoreLongJmp = Val; 1022 } 1023 1024 /// setSupportJumpTables - Indicate whether the target can generate code for 1025 /// jump tables. setSupportJumpTables(bool Val)1026 void setSupportJumpTables(bool Val) { 1027 SupportJumpTables = Val; 1028 } 1029 1030 /// setStackPointerRegisterToSaveRestore - If set to a physical register, this 1031 /// specifies the register that llvm.savestack/llvm.restorestack should save 1032 /// and restore. setStackPointerRegisterToSaveRestore(unsigned R)1033 void setStackPointerRegisterToSaveRestore(unsigned R) { 1034 StackPointerRegisterToSaveRestore = R; 1035 } 1036 1037 /// setExceptionPointerRegister - If set to a physical register, this sets 1038 /// the register that receives the exception address on entry to a landing 1039 /// pad. setExceptionPointerRegister(unsigned R)1040 void setExceptionPointerRegister(unsigned R) { 1041 ExceptionPointerRegister = R; 1042 } 1043 1044 /// setExceptionSelectorRegister - If set to a physical register, this sets 1045 /// the register that receives the exception typeid on entry to a landing 1046 /// pad. setExceptionSelectorRegister(unsigned R)1047 void setExceptionSelectorRegister(unsigned R) { 1048 ExceptionSelectorRegister = R; 1049 } 1050 1051 /// SelectIsExpensive - Tells the code generator not to expand operations 1052 /// into sequences that use the select operations if possible. 1053 void setSelectIsExpensive(bool isExpensive = true) { 1054 SelectIsExpensive = isExpensive; 1055 } 1056 1057 /// JumpIsExpensive - Tells the code generator not to expand sequence of 1058 /// operations into a separate sequences that increases the amount of 1059 /// flow control. 1060 void setJumpIsExpensive(bool isExpensive = true) { 1061 JumpIsExpensive = isExpensive; 1062 } 1063 1064 /// setIntDivIsCheap - Tells the code generator that integer divide is 1065 /// expensive, and if possible, should be replaced by an alternate sequence 1066 /// of instructions not containing an integer divide. 1067 void setIntDivIsCheap(bool isCheap = true) { IntDivIsCheap = isCheap; } 1068 1069 /// addBypassSlowDivType - Tells the code generator which types to bypass. addBypassSlowDivType(Type * slow_type,Type * fast_type)1070 void addBypassSlowDivType(Type *slow_type, Type *fast_type) { 1071 BypassSlowDivTypes[slow_type] = fast_type; 1072 } 1073 1074 /// setPow2DivIsCheap - Tells the code generator that it shouldn't generate 1075 /// srl/add/sra for a signed divide by power of two, and let the target handle 1076 /// it. 1077 void setPow2DivIsCheap(bool isCheap = true) { Pow2DivIsCheap = isCheap; } 1078 1079 /// addRegisterClass - Add the specified register class as an available 1080 /// regclass for the specified value type. This indicates the selector can 1081 /// handle values of that class natively. addRegisterClass(EVT VT,const TargetRegisterClass * RC)1082 void addRegisterClass(EVT VT, const TargetRegisterClass *RC) { 1083 assert((unsigned)VT.getSimpleVT().SimpleTy < array_lengthof(RegClassForVT)); 1084 AvailableRegClasses.push_back(std::make_pair(VT, RC)); 1085 RegClassForVT[VT.getSimpleVT().SimpleTy] = RC; 1086 } 1087 1088 /// findRepresentativeClass - Return the largest legal super-reg register class 1089 /// of the register class for the specified type and its associated "cost". 1090 virtual std::pair<const TargetRegisterClass*, uint8_t> 1091 findRepresentativeClass(EVT VT) const; 1092 1093 /// computeRegisterProperties - Once all of the register classes are added, 1094 /// this allows us to compute derived properties we expose. 1095 void computeRegisterProperties(); 1096 1097 /// setOperationAction - Indicate that the specified operation does not work 1098 /// with the specified type and indicate what to do about it. setOperationAction(unsigned Op,MVT VT,LegalizeAction Action)1099 void setOperationAction(unsigned Op, MVT VT, 1100 LegalizeAction Action) { 1101 assert(Op < array_lengthof(OpActions[0]) && "Table isn't big enough!"); 1102 OpActions[(unsigned)VT.SimpleTy][Op] = (uint8_t)Action; 1103 } 1104 1105 /// setLoadExtAction - Indicate that the specified load with extension does 1106 /// not work with the specified type and indicate what to do about it. setLoadExtAction(unsigned ExtType,MVT VT,LegalizeAction Action)1107 void setLoadExtAction(unsigned ExtType, MVT VT, 1108 LegalizeAction Action) { 1109 assert(ExtType < ISD::LAST_LOADEXT_TYPE && VT < MVT::LAST_VALUETYPE && 1110 "Table isn't big enough!"); 1111 LoadExtActions[VT.SimpleTy][ExtType] = (uint8_t)Action; 1112 } 1113 1114 /// setTruncStoreAction - Indicate that the specified truncating store does 1115 /// not work with the specified type and indicate what to do about it. setTruncStoreAction(MVT ValVT,MVT MemVT,LegalizeAction Action)1116 void setTruncStoreAction(MVT ValVT, MVT MemVT, 1117 LegalizeAction Action) { 1118 assert(ValVT < MVT::LAST_VALUETYPE && MemVT < MVT::LAST_VALUETYPE && 1119 "Table isn't big enough!"); 1120 TruncStoreActions[ValVT.SimpleTy][MemVT.SimpleTy] = (uint8_t)Action; 1121 } 1122 1123 /// setIndexedLoadAction - Indicate that the specified indexed load does or 1124 /// does not work with the specified type and indicate what to do abort 1125 /// it. NOTE: All indexed mode loads are initialized to Expand in 1126 /// TargetLowering.cpp setIndexedLoadAction(unsigned IdxMode,MVT VT,LegalizeAction Action)1127 void setIndexedLoadAction(unsigned IdxMode, MVT VT, 1128 LegalizeAction Action) { 1129 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE && 1130 (unsigned)Action < 0xf && "Table isn't big enough!"); 1131 // Load action are kept in the upper half. 1132 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0xf0; 1133 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action) <<4; 1134 } 1135 1136 /// setIndexedStoreAction - Indicate that the specified indexed store does or 1137 /// does not work with the specified type and indicate what to do about 1138 /// it. NOTE: All indexed mode stores are initialized to Expand in 1139 /// TargetLowering.cpp setIndexedStoreAction(unsigned IdxMode,MVT VT,LegalizeAction Action)1140 void setIndexedStoreAction(unsigned IdxMode, MVT VT, 1141 LegalizeAction Action) { 1142 assert(VT < MVT::LAST_VALUETYPE && IdxMode < ISD::LAST_INDEXED_MODE && 1143 (unsigned)Action < 0xf && "Table isn't big enough!"); 1144 // Store action are kept in the lower half. 1145 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] &= ~0x0f; 1146 IndexedModeActions[(unsigned)VT.SimpleTy][IdxMode] |= ((uint8_t)Action); 1147 } 1148 1149 /// setCondCodeAction - Indicate that the specified condition code is or isn't 1150 /// supported on the target and indicate what to do about it. setCondCodeAction(ISD::CondCode CC,MVT VT,LegalizeAction Action)1151 void setCondCodeAction(ISD::CondCode CC, MVT VT, 1152 LegalizeAction Action) { 1153 assert(VT < MVT::LAST_VALUETYPE && 1154 (unsigned)CC < array_lengthof(CondCodeActions) && 1155 "Table isn't big enough!"); 1156 CondCodeActions[(unsigned)CC] &= ~(uint64_t(3UL) << VT.SimpleTy*2); 1157 CondCodeActions[(unsigned)CC] |= (uint64_t)Action << VT.SimpleTy*2; 1158 } 1159 1160 /// AddPromotedToType - If Opc/OrigVT is specified as being promoted, the 1161 /// promotion code defaults to trying a larger integer/fp until it can find 1162 /// one that works. If that default is insufficient, this method can be used 1163 /// by the target to override the default. AddPromotedToType(unsigned Opc,MVT OrigVT,MVT DestVT)1164 void AddPromotedToType(unsigned Opc, MVT OrigVT, MVT DestVT) { 1165 PromoteToType[std::make_pair(Opc, OrigVT.SimpleTy)] = DestVT.SimpleTy; 1166 } 1167 1168 /// setTargetDAGCombine - Targets should invoke this method for each target 1169 /// independent node that they want to provide a custom DAG combiner for by 1170 /// implementing the PerformDAGCombine virtual method. setTargetDAGCombine(ISD::NodeType NT)1171 void setTargetDAGCombine(ISD::NodeType NT) { 1172 assert(unsigned(NT >> 3) < array_lengthof(TargetDAGCombineArray)); 1173 TargetDAGCombineArray[NT >> 3] |= 1 << (NT&7); 1174 } 1175 1176 /// setJumpBufSize - Set the target's required jmp_buf buffer size (in 1177 /// bytes); default is 200 setJumpBufSize(unsigned Size)1178 void setJumpBufSize(unsigned Size) { 1179 JumpBufSize = Size; 1180 } 1181 1182 /// setJumpBufAlignment - Set the target's required jmp_buf buffer 1183 /// alignment (in bytes); default is 0 setJumpBufAlignment(unsigned Align)1184 void setJumpBufAlignment(unsigned Align) { 1185 JumpBufAlignment = Align; 1186 } 1187 1188 /// setMinFunctionAlignment - Set the target's minimum function alignment (in 1189 /// log2(bytes)) setMinFunctionAlignment(unsigned Align)1190 void setMinFunctionAlignment(unsigned Align) { 1191 MinFunctionAlignment = Align; 1192 } 1193 1194 /// setPrefFunctionAlignment - Set the target's preferred function alignment. 1195 /// This should be set if there is a performance benefit to 1196 /// higher-than-minimum alignment (in log2(bytes)) setPrefFunctionAlignment(unsigned Align)1197 void setPrefFunctionAlignment(unsigned Align) { 1198 PrefFunctionAlignment = Align; 1199 } 1200 1201 /// setPrefLoopAlignment - Set the target's preferred loop alignment. Default 1202 /// alignment is zero, it means the target does not care about loop alignment. 1203 /// The alignment is specified in log2(bytes). setPrefLoopAlignment(unsigned Align)1204 void setPrefLoopAlignment(unsigned Align) { 1205 PrefLoopAlignment = Align; 1206 } 1207 1208 /// setMinStackArgumentAlignment - Set the minimum stack alignment of an 1209 /// argument (in log2(bytes)). setMinStackArgumentAlignment(unsigned Align)1210 void setMinStackArgumentAlignment(unsigned Align) { 1211 MinStackArgumentAlignment = Align; 1212 } 1213 1214 /// setShouldFoldAtomicFences - Set if the target's implementation of the 1215 /// atomic operation intrinsics includes locking. Default is false. setShouldFoldAtomicFences(bool fold)1216 void setShouldFoldAtomicFences(bool fold) { 1217 ShouldFoldAtomicFences = fold; 1218 } 1219 1220 /// setInsertFencesForAtomic - Set if the DAG builder should 1221 /// automatically insert fences and reduce the order of atomic memory 1222 /// operations to Monotonic. setInsertFencesForAtomic(bool fence)1223 void setInsertFencesForAtomic(bool fence) { 1224 InsertFencesForAtomic = fence; 1225 } 1226 1227 public: 1228 //===--------------------------------------------------------------------===// 1229 // Lowering methods - These methods must be implemented by targets so that 1230 // the SelectionDAGLowering code knows how to lower these. 1231 // 1232 1233 /// LowerFormalArguments - This hook must be implemented to lower the 1234 /// incoming (formal) arguments, described by the Ins array, into the 1235 /// specified DAG. The implementation should fill in the InVals array 1236 /// with legal-type argument values, and return the resulting token 1237 /// chain value. 1238 /// 1239 virtual SDValue LowerFormalArguments(SDValue,CallingConv::ID,bool,const SmallVectorImpl<ISD::InputArg> &,DebugLoc,SelectionDAG &,SmallVectorImpl<SDValue> &)1240 LowerFormalArguments(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 1241 bool /*isVarArg*/, 1242 const SmallVectorImpl<ISD::InputArg> &/*Ins*/, 1243 DebugLoc /*dl*/, SelectionDAG &/*DAG*/, 1244 SmallVectorImpl<SDValue> &/*InVals*/) const { 1245 llvm_unreachable("Not Implemented"); 1246 } 1247 1248 struct ArgListEntry { 1249 SDValue Node; 1250 Type* Ty; 1251 bool isSExt : 1; 1252 bool isZExt : 1; 1253 bool isInReg : 1; 1254 bool isSRet : 1; 1255 bool isNest : 1; 1256 bool isByVal : 1; 1257 uint16_t Alignment; 1258 ArgListEntryArgListEntry1259 ArgListEntry() : isSExt(false), isZExt(false), isInReg(false), 1260 isSRet(false), isNest(false), isByVal(false), Alignment(0) { } 1261 }; 1262 typedef std::vector<ArgListEntry> ArgListTy; 1263 1264 /// CallLoweringInfo - This structure contains all information that is 1265 /// necessary for lowering calls. It is passed to TLI::LowerCallTo when the 1266 /// SelectionDAG builder needs to lower a call, and targets will see this 1267 /// struct in their LowerCall implementation. 1268 struct CallLoweringInfo { 1269 SDValue Chain; 1270 Type *RetTy; 1271 bool RetSExt : 1; 1272 bool RetZExt : 1; 1273 bool IsVarArg : 1; 1274 bool IsInReg : 1; 1275 bool DoesNotReturn : 1; 1276 bool IsReturnValueUsed : 1; 1277 1278 // IsTailCall should be modified by implementations of 1279 // TargetLowering::LowerCall that perform tail call conversions. 1280 bool IsTailCall; 1281 1282 unsigned NumFixedArgs; 1283 CallingConv::ID CallConv; 1284 SDValue Callee; 1285 ArgListTy &Args; 1286 SelectionDAG &DAG; 1287 DebugLoc DL; 1288 ImmutableCallSite *CS; 1289 SmallVector<ISD::OutputArg, 32> Outs; 1290 SmallVector<SDValue, 32> OutVals; 1291 SmallVector<ISD::InputArg, 32> Ins; 1292 1293 1294 /// CallLoweringInfo - Constructs a call lowering context based on the 1295 /// ImmutableCallSite \p cs. CallLoweringInfoCallLoweringInfo1296 CallLoweringInfo(SDValue chain, Type *retTy, 1297 FunctionType *FTy, bool isTailCall, SDValue callee, 1298 ArgListTy &args, SelectionDAG &dag, DebugLoc dl, 1299 ImmutableCallSite &cs) 1300 : Chain(chain), RetTy(retTy), RetSExt(cs.paramHasAttr(0, Attribute::SExt)), 1301 RetZExt(cs.paramHasAttr(0, Attribute::ZExt)), IsVarArg(FTy->isVarArg()), 1302 IsInReg(cs.paramHasAttr(0, Attribute::InReg)), 1303 DoesNotReturn(cs.doesNotReturn()), 1304 IsReturnValueUsed(!cs.getInstruction()->use_empty()), 1305 IsTailCall(isTailCall), NumFixedArgs(FTy->getNumParams()), 1306 CallConv(cs.getCallingConv()), Callee(callee), Args(args), DAG(dag), 1307 DL(dl), CS(&cs) {} 1308 1309 /// CallLoweringInfo - Constructs a call lowering context based on the 1310 /// provided call information. CallLoweringInfoCallLoweringInfo1311 CallLoweringInfo(SDValue chain, Type *retTy, bool retSExt, bool retZExt, 1312 bool isVarArg, bool isInReg, unsigned numFixedArgs, 1313 CallingConv::ID callConv, bool isTailCall, 1314 bool doesNotReturn, bool isReturnValueUsed, SDValue callee, 1315 ArgListTy &args, SelectionDAG &dag, DebugLoc dl) 1316 : Chain(chain), RetTy(retTy), RetSExt(retSExt), RetZExt(retZExt), 1317 IsVarArg(isVarArg), IsInReg(isInReg), DoesNotReturn(doesNotReturn), 1318 IsReturnValueUsed(isReturnValueUsed), IsTailCall(isTailCall), 1319 NumFixedArgs(numFixedArgs), CallConv(callConv), Callee(callee), 1320 Args(args), DAG(dag), DL(dl), CS(NULL) {} 1321 }; 1322 1323 /// LowerCallTo - This function lowers an abstract call to a function into an 1324 /// actual call. This returns a pair of operands. The first element is the 1325 /// return value for the function (if RetTy is not VoidTy). The second 1326 /// element is the outgoing token chain. It calls LowerCall to do the actual 1327 /// lowering. 1328 std::pair<SDValue, SDValue> LowerCallTo(CallLoweringInfo &CLI) const; 1329 1330 /// LowerCall - This hook must be implemented to lower calls into the 1331 /// the specified DAG. The outgoing arguments to the call are described 1332 /// by the Outs array, and the values to be returned by the call are 1333 /// described by the Ins array. The implementation should fill in the 1334 /// InVals array with legal-type return values from the call, and return 1335 /// the resulting token chain value. 1336 virtual SDValue LowerCall(CallLoweringInfo &,SmallVectorImpl<SDValue> &)1337 LowerCall(CallLoweringInfo &/*CLI*/, 1338 SmallVectorImpl<SDValue> &/*InVals*/) const { 1339 llvm_unreachable("Not Implemented"); 1340 } 1341 1342 /// HandleByVal - Target-specific cleanup for formal ByVal parameters. HandleByVal(CCState *,unsigned &)1343 virtual void HandleByVal(CCState *, unsigned &) const {} 1344 1345 /// CanLowerReturn - This hook should be implemented to check whether the 1346 /// return values described by the Outs array can fit into the return 1347 /// registers. If false is returned, an sret-demotion is performed. 1348 /// CanLowerReturn(CallingConv::ID,MachineFunction &,bool,const SmallVectorImpl<ISD::OutputArg> &,LLVMContext &)1349 virtual bool CanLowerReturn(CallingConv::ID /*CallConv*/, 1350 MachineFunction &/*MF*/, bool /*isVarArg*/, 1351 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 1352 LLVMContext &/*Context*/) const 1353 { 1354 // Return true by default to get preexisting behavior. 1355 return true; 1356 } 1357 1358 /// LowerReturn - This hook must be implemented to lower outgoing 1359 /// return values, described by the Outs array, into the specified 1360 /// DAG. The implementation should return the resulting token chain 1361 /// value. 1362 /// 1363 virtual SDValue LowerReturn(SDValue,CallingConv::ID,bool,const SmallVectorImpl<ISD::OutputArg> &,const SmallVectorImpl<SDValue> &,DebugLoc,SelectionDAG &)1364 LowerReturn(SDValue /*Chain*/, CallingConv::ID /*CallConv*/, 1365 bool /*isVarArg*/, 1366 const SmallVectorImpl<ISD::OutputArg> &/*Outs*/, 1367 const SmallVectorImpl<SDValue> &/*OutVals*/, 1368 DebugLoc /*dl*/, SelectionDAG &/*DAG*/) const { 1369 llvm_unreachable("Not Implemented"); 1370 } 1371 1372 /// isUsedByReturnOnly - Return true if result of the specified node is used 1373 /// by a return node only. It also compute and return the input chain for the 1374 /// tail call. 1375 /// This is used to determine whether it is possible 1376 /// to codegen a libcall as tail call at legalization time. isUsedByReturnOnly(SDNode *,SDValue & Chain)1377 virtual bool isUsedByReturnOnly(SDNode *, SDValue &Chain) const { 1378 return false; 1379 } 1380 1381 /// mayBeEmittedAsTailCall - Return true if the target may be able emit the 1382 /// call instruction as a tail call. This is used by optimization passes to 1383 /// determine if it's profitable to duplicate return instructions to enable 1384 /// tailcall optimization. mayBeEmittedAsTailCall(CallInst *)1385 virtual bool mayBeEmittedAsTailCall(CallInst *) const { 1386 return false; 1387 } 1388 1389 /// getTypeForExtArgOrReturn - Return the type that should be used to zero or 1390 /// sign extend a zeroext/signext integer argument or return value. 1391 /// FIXME: Most C calling convention requires the return type to be promoted, 1392 /// but this is not true all the time, e.g. i1 on x86-64. It is also not 1393 /// necessary for non-C calling conventions. The frontend should handle this 1394 /// and include all of the necessary information. getTypeForExtArgOrReturn(LLVMContext & Context,EVT VT,ISD::NodeType)1395 virtual EVT getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT, 1396 ISD::NodeType /*ExtendKind*/) const { 1397 EVT MinVT = getRegisterType(Context, MVT::i32); 1398 return VT.bitsLT(MinVT) ? MinVT : VT; 1399 } 1400 1401 /// LowerOperationWrapper - This callback is invoked by the type legalizer 1402 /// to legalize nodes with an illegal operand type but legal result types. 1403 /// It replaces the LowerOperation callback in the type Legalizer. 1404 /// The reason we can not do away with LowerOperation entirely is that 1405 /// LegalizeDAG isn't yet ready to use this callback. 1406 /// TODO: Consider merging with ReplaceNodeResults. 1407 1408 /// The target places new result values for the node in Results (their number 1409 /// and types must exactly match those of the original return values of 1410 /// the node), or leaves Results empty, which indicates that the node is not 1411 /// to be custom lowered after all. 1412 /// The default implementation calls LowerOperation. 1413 virtual void LowerOperationWrapper(SDNode *N, 1414 SmallVectorImpl<SDValue> &Results, 1415 SelectionDAG &DAG) const; 1416 1417 /// LowerOperation - This callback is invoked for operations that are 1418 /// unsupported by the target, which are registered to use 'custom' lowering, 1419 /// and whose defined values are all legal. 1420 /// If the target has no operations that require custom lowering, it need not 1421 /// implement this. The default implementation of this aborts. 1422 virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; 1423 1424 /// ReplaceNodeResults - This callback is invoked when a node result type is 1425 /// illegal for the target, and the operation was registered to use 'custom' 1426 /// lowering for that result type. The target places new result values for 1427 /// the node in Results (their number and types must exactly match those of 1428 /// the original return values of the node), or leaves Results empty, which 1429 /// indicates that the node is not to be custom lowered after all. 1430 /// 1431 /// If the target has no operations that require custom lowering, it need not 1432 /// implement this. The default implementation aborts. ReplaceNodeResults(SDNode *,SmallVectorImpl<SDValue> &,SelectionDAG &)1433 virtual void ReplaceNodeResults(SDNode * /*N*/, 1434 SmallVectorImpl<SDValue> &/*Results*/, 1435 SelectionDAG &/*DAG*/) const { 1436 llvm_unreachable("ReplaceNodeResults not implemented for this target!"); 1437 } 1438 1439 /// getTargetNodeName() - This method returns the name of a target specific 1440 /// DAG node. 1441 virtual const char *getTargetNodeName(unsigned Opcode) const; 1442 1443 /// createFastISel - This method returns a target specific FastISel object, 1444 /// or null if the target does not support "fast" ISel. createFastISel(FunctionLoweringInfo &,const TargetLibraryInfo *)1445 virtual FastISel *createFastISel(FunctionLoweringInfo &, 1446 const TargetLibraryInfo *) const { 1447 return 0; 1448 } 1449 1450 //===--------------------------------------------------------------------===// 1451 // Inline Asm Support hooks 1452 // 1453 1454 /// ExpandInlineAsm - This hook allows the target to expand an inline asm 1455 /// call to be explicit llvm code if it wants to. This is useful for 1456 /// turning simple inline asms into LLVM intrinsics, which gives the 1457 /// compiler more information about the behavior of the code. ExpandInlineAsm(CallInst *)1458 virtual bool ExpandInlineAsm(CallInst *) const { 1459 return false; 1460 } 1461 1462 enum ConstraintType { 1463 C_Register, // Constraint represents specific register(s). 1464 C_RegisterClass, // Constraint represents any of register(s) in class. 1465 C_Memory, // Memory constraint. 1466 C_Other, // Something else. 1467 C_Unknown // Unsupported constraint. 1468 }; 1469 1470 enum ConstraintWeight { 1471 // Generic weights. 1472 CW_Invalid = -1, // No match. 1473 CW_Okay = 0, // Acceptable. 1474 CW_Good = 1, // Good weight. 1475 CW_Better = 2, // Better weight. 1476 CW_Best = 3, // Best weight. 1477 1478 // Well-known weights. 1479 CW_SpecificReg = CW_Okay, // Specific register operands. 1480 CW_Register = CW_Good, // Register operands. 1481 CW_Memory = CW_Better, // Memory operands. 1482 CW_Constant = CW_Best, // Constant operand. 1483 CW_Default = CW_Okay // Default or don't know type. 1484 }; 1485 1486 /// AsmOperandInfo - This contains information for each constraint that we are 1487 /// lowering. 1488 struct AsmOperandInfo : public InlineAsm::ConstraintInfo { 1489 /// ConstraintCode - This contains the actual string for the code, like "m". 1490 /// TargetLowering picks the 'best' code from ConstraintInfo::Codes that 1491 /// most closely matches the operand. 1492 std::string ConstraintCode; 1493 1494 /// ConstraintType - Information about the constraint code, e.g. Register, 1495 /// RegisterClass, Memory, Other, Unknown. 1496 TargetLowering::ConstraintType ConstraintType; 1497 1498 /// CallOperandval - If this is the result output operand or a 1499 /// clobber, this is null, otherwise it is the incoming operand to the 1500 /// CallInst. This gets modified as the asm is processed. 1501 Value *CallOperandVal; 1502 1503 /// ConstraintVT - The ValueType for the operand value. 1504 EVT ConstraintVT; 1505 1506 /// isMatchingInputConstraint - Return true of this is an input operand that 1507 /// is a matching constraint like "4". 1508 bool isMatchingInputConstraint() const; 1509 1510 /// getMatchedOperand - If this is an input matching constraint, this method 1511 /// returns the output operand it matches. 1512 unsigned getMatchedOperand() const; 1513 1514 /// Copy constructor for copying from an AsmOperandInfo. AsmOperandInfoAsmOperandInfo1515 AsmOperandInfo(const AsmOperandInfo &info) 1516 : InlineAsm::ConstraintInfo(info), 1517 ConstraintCode(info.ConstraintCode), 1518 ConstraintType(info.ConstraintType), 1519 CallOperandVal(info.CallOperandVal), 1520 ConstraintVT(info.ConstraintVT) { 1521 } 1522 1523 /// Copy constructor for copying from a ConstraintInfo. AsmOperandInfoAsmOperandInfo1524 AsmOperandInfo(const InlineAsm::ConstraintInfo &info) 1525 : InlineAsm::ConstraintInfo(info), 1526 ConstraintType(TargetLowering::C_Unknown), 1527 CallOperandVal(0), ConstraintVT(MVT::Other) { 1528 } 1529 }; 1530 1531 typedef std::vector<AsmOperandInfo> AsmOperandInfoVector; 1532 1533 /// ParseConstraints - Split up the constraint string from the inline 1534 /// assembly value into the specific constraints and their prefixes, 1535 /// and also tie in the associated operand values. 1536 /// If this returns an empty vector, and if the constraint string itself 1537 /// isn't empty, there was an error parsing. 1538 virtual AsmOperandInfoVector ParseConstraints(ImmutableCallSite CS) const; 1539 1540 /// Examine constraint type and operand type and determine a weight value. 1541 /// The operand object must already have been set up with the operand type. 1542 virtual ConstraintWeight getMultipleConstraintMatchWeight( 1543 AsmOperandInfo &info, int maIndex) const; 1544 1545 /// Examine constraint string and operand type and determine a weight value. 1546 /// The operand object must already have been set up with the operand type. 1547 virtual ConstraintWeight getSingleConstraintMatchWeight( 1548 AsmOperandInfo &info, const char *constraint) const; 1549 1550 /// ComputeConstraintToUse - Determines the constraint code and constraint 1551 /// type to use for the specific AsmOperandInfo, setting 1552 /// OpInfo.ConstraintCode and OpInfo.ConstraintType. If the actual operand 1553 /// being passed in is available, it can be passed in as Op, otherwise an 1554 /// empty SDValue can be passed. 1555 virtual void ComputeConstraintToUse(AsmOperandInfo &OpInfo, 1556 SDValue Op, 1557 SelectionDAG *DAG = 0) const; 1558 1559 /// getConstraintType - Given a constraint, return the type of constraint it 1560 /// is for this target. 1561 virtual ConstraintType getConstraintType(const std::string &Constraint) const; 1562 1563 /// getRegForInlineAsmConstraint - Given a physical register constraint (e.g. 1564 /// {edx}), return the register number and the register class for the 1565 /// register. 1566 /// 1567 /// Given a register class constraint, like 'r', if this corresponds directly 1568 /// to an LLVM register class, return a register of 0 and the register class 1569 /// pointer. 1570 /// 1571 /// This should only be used for C_Register constraints. On error, 1572 /// this returns a register number of 0 and a null register class pointer.. 1573 virtual std::pair<unsigned, const TargetRegisterClass*> 1574 getRegForInlineAsmConstraint(const std::string &Constraint, 1575 EVT VT) const; 1576 1577 /// LowerXConstraint - try to replace an X constraint, which matches anything, 1578 /// with another that has more specific requirements based on the type of the 1579 /// corresponding operand. This returns null if there is no replacement to 1580 /// make. 1581 virtual const char *LowerXConstraint(EVT ConstraintVT) const; 1582 1583 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 1584 /// vector. If it is invalid, don't add anything to Ops. 1585 virtual void LowerAsmOperandForConstraint(SDValue Op, std::string &Constraint, 1586 std::vector<SDValue> &Ops, 1587 SelectionDAG &DAG) const; 1588 1589 //===--------------------------------------------------------------------===// 1590 // Instruction Emitting Hooks 1591 // 1592 1593 // EmitInstrWithCustomInserter - This method should be implemented by targets 1594 // that mark instructions with the 'usesCustomInserter' flag. These 1595 // instructions are special in various ways, which require special support to 1596 // insert. The specified MachineInstr is created but not inserted into any 1597 // basic blocks, and this method is called to expand it into a sequence of 1598 // instructions, potentially also creating new basic blocks and control flow. 1599 virtual MachineBasicBlock * 1600 EmitInstrWithCustomInserter(MachineInstr *MI, MachineBasicBlock *MBB) const; 1601 1602 /// AdjustInstrPostInstrSelection - This method should be implemented by 1603 /// targets that mark instructions with the 'hasPostISelHook' flag. These 1604 /// instructions must be adjusted after instruction selection by target hooks. 1605 /// e.g. To fill in optional defs for ARM 's' setting instructions. 1606 virtual void 1607 AdjustInstrPostInstrSelection(MachineInstr *MI, SDNode *Node) const; 1608 1609 //===--------------------------------------------------------------------===// 1610 // Addressing mode description hooks (used by LSR etc). 1611 // 1612 1613 /// AddrMode - This represents an addressing mode of: 1614 /// BaseGV + BaseOffs + BaseReg + Scale*ScaleReg 1615 /// If BaseGV is null, there is no BaseGV. 1616 /// If BaseOffs is zero, there is no base offset. 1617 /// If HasBaseReg is false, there is no base register. 1618 /// If Scale is zero, there is no ScaleReg. Scale of 1 indicates a reg with 1619 /// no scale. 1620 /// 1621 struct AddrMode { 1622 GlobalValue *BaseGV; 1623 int64_t BaseOffs; 1624 bool HasBaseReg; 1625 int64_t Scale; AddrModeAddrMode1626 AddrMode() : BaseGV(0), BaseOffs(0), HasBaseReg(false), Scale(0) {} 1627 }; 1628 1629 /// GetAddrModeArguments - CodeGenPrepare sinks address calculations into the 1630 /// same BB as Load/Store instructions reading the address. This allows as 1631 /// much computation as possible to be done in the address mode for that 1632 /// operand. This hook lets targets also pass back when this should be done 1633 /// on intrinsics which load/store. GetAddrModeArguments(IntrinsicInst * I,SmallVectorImpl<Value * > & Ops,Type * & AccessTy)1634 virtual bool GetAddrModeArguments(IntrinsicInst *I, 1635 SmallVectorImpl<Value*> &Ops, 1636 Type *&AccessTy) const { 1637 return false; 1638 } 1639 1640 /// isLegalAddressingMode - Return true if the addressing mode represented by 1641 /// AM is legal for this target, for a load/store of the specified type. 1642 /// The type may be VoidTy, in which case only return true if the addressing 1643 /// mode is legal for a load/store of any legal type. 1644 /// TODO: Handle pre/postinc as well. 1645 virtual bool isLegalAddressingMode(const AddrMode &AM, Type *Ty) const; 1646 1647 /// isLegalICmpImmediate - Return true if the specified immediate is legal 1648 /// icmp immediate, that is the target has icmp instructions which can compare 1649 /// a register against the immediate without having to materialize the 1650 /// immediate into a register. isLegalICmpImmediate(int64_t)1651 virtual bool isLegalICmpImmediate(int64_t) const { 1652 return true; 1653 } 1654 1655 /// isLegalAddImmediate - Return true if the specified immediate is legal 1656 /// add immediate, that is the target has add instructions which can add 1657 /// a register with the immediate without having to materialize the 1658 /// immediate into a register. isLegalAddImmediate(int64_t)1659 virtual bool isLegalAddImmediate(int64_t) const { 1660 return true; 1661 } 1662 1663 /// isTruncateFree - Return true if it's free to truncate a value of 1664 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in 1665 /// register EAX to i16 by referencing its sub-register AX. isTruncateFree(Type *,Type *)1666 virtual bool isTruncateFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 1667 return false; 1668 } 1669 isTruncateFree(EVT,EVT)1670 virtual bool isTruncateFree(EVT /*VT1*/, EVT /*VT2*/) const { 1671 return false; 1672 } 1673 1674 /// isZExtFree - Return true if any actual instruction that defines a 1675 /// value of type Ty1 implicitly zero-extends the value to Ty2 in the result 1676 /// register. This does not necessarily include registers defined in 1677 /// unknown ways, such as incoming arguments, or copies from unknown 1678 /// virtual registers. Also, if isTruncateFree(Ty2, Ty1) is true, this 1679 /// does not necessarily apply to truncate instructions. e.g. on x86-64, 1680 /// all instructions that define 32-bit values implicit zero-extend the 1681 /// result out to 64 bits. isZExtFree(Type *,Type *)1682 virtual bool isZExtFree(Type * /*Ty1*/, Type * /*Ty2*/) const { 1683 return false; 1684 } 1685 isZExtFree(EVT,EVT)1686 virtual bool isZExtFree(EVT /*VT1*/, EVT /*VT2*/) const { 1687 return false; 1688 } 1689 1690 /// isFNegFree - Return true if an fneg operation is free to the point where 1691 /// it is never worthwhile to replace it with a bitwise operation. isFNegFree(EVT)1692 virtual bool isFNegFree(EVT) const { 1693 return false; 1694 } 1695 1696 /// isFAbsFree - Return true if an fneg operation is free to the point where 1697 /// it is never worthwhile to replace it with a bitwise operation. isFAbsFree(EVT)1698 virtual bool isFAbsFree(EVT) const { 1699 return false; 1700 } 1701 1702 /// isFMAFasterThanMulAndAdd - Return true if an FMA operation is faster than 1703 /// a pair of mul and add instructions. fmuladd intrinsics will be expanded to 1704 /// FMAs when this method returns true (and FMAs are legal), otherwise fmuladd 1705 /// is expanded to mul + add. isFMAFasterThanMulAndAdd(EVT)1706 virtual bool isFMAFasterThanMulAndAdd(EVT) const { 1707 return false; 1708 } 1709 1710 /// isNarrowingProfitable - Return true if it's profitable to narrow 1711 /// operations of type VT1 to VT2. e.g. on x86, it's profitable to narrow 1712 /// from i32 to i8 but not from i32 to i16. isNarrowingProfitable(EVT,EVT)1713 virtual bool isNarrowingProfitable(EVT /*VT1*/, EVT /*VT2*/) const { 1714 return false; 1715 } 1716 1717 //===--------------------------------------------------------------------===// 1718 // Div utility functions 1719 // 1720 SDValue BuildExactSDIV(SDValue Op1, SDValue Op2, DebugLoc dl, 1721 SelectionDAG &DAG) const; 1722 SDValue BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 1723 std::vector<SDNode*>* Created) const; 1724 SDValue BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 1725 std::vector<SDNode*>* Created) const; 1726 1727 1728 //===--------------------------------------------------------------------===// 1729 // Runtime Library hooks 1730 // 1731 1732 /// setLibcallName - Rename the default libcall routine name for the specified 1733 /// libcall. setLibcallName(RTLIB::Libcall Call,const char * Name)1734 void setLibcallName(RTLIB::Libcall Call, const char *Name) { 1735 LibcallRoutineNames[Call] = Name; 1736 } 1737 1738 /// getLibcallName - Get the libcall routine name for the specified libcall. 1739 /// getLibcallName(RTLIB::Libcall Call)1740 const char *getLibcallName(RTLIB::Libcall Call) const { 1741 return LibcallRoutineNames[Call]; 1742 } 1743 1744 /// setCmpLibcallCC - Override the default CondCode to be used to test the 1745 /// result of the comparison libcall against zero. setCmpLibcallCC(RTLIB::Libcall Call,ISD::CondCode CC)1746 void setCmpLibcallCC(RTLIB::Libcall Call, ISD::CondCode CC) { 1747 CmpLibcallCCs[Call] = CC; 1748 } 1749 1750 /// getCmpLibcallCC - Get the CondCode that's to be used to test the result of 1751 /// the comparison libcall against zero. getCmpLibcallCC(RTLIB::Libcall Call)1752 ISD::CondCode getCmpLibcallCC(RTLIB::Libcall Call) const { 1753 return CmpLibcallCCs[Call]; 1754 } 1755 1756 /// setLibcallCallingConv - Set the CallingConv that should be used for the 1757 /// specified libcall. setLibcallCallingConv(RTLIB::Libcall Call,CallingConv::ID CC)1758 void setLibcallCallingConv(RTLIB::Libcall Call, CallingConv::ID CC) { 1759 LibcallCallingConvs[Call] = CC; 1760 } 1761 1762 /// getLibcallCallingConv - Get the CallingConv that should be used for the 1763 /// specified libcall. getLibcallCallingConv(RTLIB::Libcall Call)1764 CallingConv::ID getLibcallCallingConv(RTLIB::Libcall Call) const { 1765 return LibcallCallingConvs[Call]; 1766 } 1767 1768 private: 1769 const TargetMachine &TM; 1770 const TargetData *TD; 1771 const TargetLoweringObjectFile &TLOF; 1772 1773 /// PointerTy - The type to use for pointers, usually i32 or i64. 1774 /// 1775 MVT PointerTy; 1776 1777 /// IsLittleEndian - True if this is a little endian target. 1778 /// 1779 bool IsLittleEndian; 1780 1781 /// SelectIsExpensive - Tells the code generator not to expand operations 1782 /// into sequences that use the select operations if possible. 1783 bool SelectIsExpensive; 1784 1785 /// IntDivIsCheap - Tells the code generator not to expand integer divides by 1786 /// constants into a sequence of muls, adds, and shifts. This is a hack until 1787 /// a real cost model is in place. If we ever optimize for size, this will be 1788 /// set to true unconditionally. 1789 bool IntDivIsCheap; 1790 1791 /// BypassSlowDivTypes - Tells the code generator to bypass slow divide or 1792 /// remainder instructions. For example, SlowDivBypass[i32,u8] tells the code 1793 /// generator to bypass 32-bit signed integer div/rem with an 8-bit unsigned 1794 /// integer div/rem when the operands are positive and less than 256. 1795 DenseMap <Type *, Type *> BypassSlowDivTypes; 1796 1797 /// Pow2DivIsCheap - Tells the code generator that it shouldn't generate 1798 /// srl/add/sra for a signed divide by power of two, and let the target handle 1799 /// it. 1800 bool Pow2DivIsCheap; 1801 1802 /// JumpIsExpensive - Tells the code generator that it shouldn't generate 1803 /// extra flow control instructions and should attempt to combine flow 1804 /// control instructions via predication. 1805 bool JumpIsExpensive; 1806 1807 /// UseUnderscoreSetJmp - This target prefers to use _setjmp to implement 1808 /// llvm.setjmp. Defaults to false. 1809 bool UseUnderscoreSetJmp; 1810 1811 /// UseUnderscoreLongJmp - This target prefers to use _longjmp to implement 1812 /// llvm.longjmp. Defaults to false. 1813 bool UseUnderscoreLongJmp; 1814 1815 /// SupportJumpTables - Whether the target can generate code for jumptables. 1816 /// If it's not true, then each jumptable must be lowered into if-then-else's. 1817 bool SupportJumpTables; 1818 1819 /// BooleanContents - Information about the contents of the high-bits in 1820 /// boolean values held in a type wider than i1. See getBooleanContents. 1821 BooleanContent BooleanContents; 1822 /// BooleanVectorContents - Information about the contents of the high-bits 1823 /// in boolean vector values when the element type is wider than i1. See 1824 /// getBooleanContents. 1825 BooleanContent BooleanVectorContents; 1826 1827 /// SchedPreferenceInfo - The target scheduling preference: shortest possible 1828 /// total cycles or lowest register usage. 1829 Sched::Preference SchedPreferenceInfo; 1830 1831 /// JumpBufSize - The size, in bytes, of the target's jmp_buf buffers 1832 unsigned JumpBufSize; 1833 1834 /// JumpBufAlignment - The alignment, in bytes, of the target's jmp_buf 1835 /// buffers 1836 unsigned JumpBufAlignment; 1837 1838 /// MinStackArgumentAlignment - The minimum alignment that any argument 1839 /// on the stack needs to have. 1840 /// 1841 unsigned MinStackArgumentAlignment; 1842 1843 /// MinFunctionAlignment - The minimum function alignment (used when 1844 /// optimizing for size, and to prevent explicitly provided alignment 1845 /// from leading to incorrect code). 1846 /// 1847 unsigned MinFunctionAlignment; 1848 1849 /// PrefFunctionAlignment - The preferred function alignment (used when 1850 /// alignment unspecified and optimizing for speed). 1851 /// 1852 unsigned PrefFunctionAlignment; 1853 1854 /// PrefLoopAlignment - The preferred loop alignment. 1855 /// 1856 unsigned PrefLoopAlignment; 1857 1858 /// ShouldFoldAtomicFences - Whether fencing MEMBARRIER instructions should 1859 /// be folded into the enclosed atomic intrinsic instruction by the 1860 /// combiner. 1861 bool ShouldFoldAtomicFences; 1862 1863 /// InsertFencesForAtomic - Whether the DAG builder should automatically 1864 /// insert fences and reduce ordering for atomics. (This will be set for 1865 /// for most architectures with weak memory ordering.) 1866 bool InsertFencesForAtomic; 1867 1868 /// StackPointerRegisterToSaveRestore - If set to a physical register, this 1869 /// specifies the register that llvm.savestack/llvm.restorestack should save 1870 /// and restore. 1871 unsigned StackPointerRegisterToSaveRestore; 1872 1873 /// ExceptionPointerRegister - If set to a physical register, this specifies 1874 /// the register that receives the exception address on entry to a landing 1875 /// pad. 1876 unsigned ExceptionPointerRegister; 1877 1878 /// ExceptionSelectorRegister - If set to a physical register, this specifies 1879 /// the register that receives the exception typeid on entry to a landing 1880 /// pad. 1881 unsigned ExceptionSelectorRegister; 1882 1883 /// RegClassForVT - This indicates the default register class to use for 1884 /// each ValueType the target supports natively. 1885 const TargetRegisterClass *RegClassForVT[MVT::LAST_VALUETYPE]; 1886 unsigned char NumRegistersForVT[MVT::LAST_VALUETYPE]; 1887 EVT RegisterTypeForVT[MVT::LAST_VALUETYPE]; 1888 1889 /// RepRegClassForVT - This indicates the "representative" register class to 1890 /// use for each ValueType the target supports natively. This information is 1891 /// used by the scheduler to track register pressure. By default, the 1892 /// representative register class is the largest legal super-reg register 1893 /// class of the register class of the specified type. e.g. On x86, i8, i16, 1894 /// and i32's representative class would be GR32. 1895 const TargetRegisterClass *RepRegClassForVT[MVT::LAST_VALUETYPE]; 1896 1897 /// RepRegClassCostForVT - This indicates the "cost" of the "representative" 1898 /// register class for each ValueType. The cost is used by the scheduler to 1899 /// approximate register pressure. 1900 uint8_t RepRegClassCostForVT[MVT::LAST_VALUETYPE]; 1901 1902 /// TransformToType - For any value types we are promoting or expanding, this 1903 /// contains the value type that we are changing to. For Expanded types, this 1904 /// contains one step of the expand (e.g. i64 -> i32), even if there are 1905 /// multiple steps required (e.g. i64 -> i16). For types natively supported 1906 /// by the system, this holds the same type (e.g. i32 -> i32). 1907 EVT TransformToType[MVT::LAST_VALUETYPE]; 1908 1909 /// OpActions - For each operation and each value type, keep a LegalizeAction 1910 /// that indicates how instruction selection should deal with the operation. 1911 /// Most operations are Legal (aka, supported natively by the target), but 1912 /// operations that are not should be described. Note that operations on 1913 /// non-legal value types are not described here. 1914 uint8_t OpActions[MVT::LAST_VALUETYPE][ISD::BUILTIN_OP_END]; 1915 1916 /// LoadExtActions - For each load extension type and each value type, 1917 /// keep a LegalizeAction that indicates how instruction selection should deal 1918 /// with a load of a specific value type and extension type. 1919 uint8_t LoadExtActions[MVT::LAST_VALUETYPE][ISD::LAST_LOADEXT_TYPE]; 1920 1921 /// TruncStoreActions - For each value type pair keep a LegalizeAction that 1922 /// indicates whether a truncating store of a specific value type and 1923 /// truncating type is legal. 1924 uint8_t TruncStoreActions[MVT::LAST_VALUETYPE][MVT::LAST_VALUETYPE]; 1925 1926 /// IndexedModeActions - For each indexed mode and each value type, 1927 /// keep a pair of LegalizeAction that indicates how instruction 1928 /// selection should deal with the load / store. The first dimension is the 1929 /// value_type for the reference. The second dimension represents the various 1930 /// modes for load store. 1931 uint8_t IndexedModeActions[MVT::LAST_VALUETYPE][ISD::LAST_INDEXED_MODE]; 1932 1933 /// CondCodeActions - For each condition code (ISD::CondCode) keep a 1934 /// LegalizeAction that indicates how instruction selection should 1935 /// deal with the condition code. 1936 uint64_t CondCodeActions[ISD::SETCC_INVALID]; 1937 1938 ValueTypeActionImpl ValueTypeActions; 1939 1940 typedef std::pair<LegalizeTypeAction, EVT> LegalizeKind; 1941 1942 LegalizeKind getTypeConversion(LLVMContext & Context,EVT VT)1943 getTypeConversion(LLVMContext &Context, EVT VT) const { 1944 // If this is a simple type, use the ComputeRegisterProp mechanism. 1945 if (VT.isSimple()) { 1946 assert((unsigned)VT.getSimpleVT().SimpleTy < 1947 array_lengthof(TransformToType)); 1948 EVT NVT = TransformToType[VT.getSimpleVT().SimpleTy]; 1949 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(VT.getSimpleVT()); 1950 1951 assert( 1952 (!(NVT.isSimple() && LA != TypeLegal) || 1953 ValueTypeActions.getTypeAction(NVT.getSimpleVT()) != TypePromoteInteger) 1954 && "Promote may not follow Expand or Promote"); 1955 1956 return LegalizeKind(LA, NVT); 1957 } 1958 1959 // Handle Extended Scalar Types. 1960 if (!VT.isVector()) { 1961 assert(VT.isInteger() && "Float types must be simple"); 1962 unsigned BitSize = VT.getSizeInBits(); 1963 // First promote to a power-of-two size, then expand if necessary. 1964 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 1965 EVT NVT = VT.getRoundIntegerType(Context); 1966 assert(NVT != VT && "Unable to round integer VT"); 1967 LegalizeKind NextStep = getTypeConversion(Context, NVT); 1968 // Avoid multi-step promotion. 1969 if (NextStep.first == TypePromoteInteger) return NextStep; 1970 // Return rounded integer type. 1971 return LegalizeKind(TypePromoteInteger, NVT); 1972 } 1973 1974 return LegalizeKind(TypeExpandInteger, 1975 EVT::getIntegerVT(Context, VT.getSizeInBits()/2)); 1976 } 1977 1978 // Handle vector types. 1979 unsigned NumElts = VT.getVectorNumElements(); 1980 EVT EltVT = VT.getVectorElementType(); 1981 1982 // Vectors with only one element are always scalarized. 1983 if (NumElts == 1) 1984 return LegalizeKind(TypeScalarizeVector, EltVT); 1985 1986 // Try to widen vector elements until a legal type is found. 1987 if (EltVT.isInteger()) { 1988 // Vectors with a number of elements that is not a power of two are always 1989 // widened, for example <3 x float> -> <4 x float>. 1990 if (!VT.isPow2VectorType()) { 1991 NumElts = (unsigned)NextPowerOf2(NumElts); 1992 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 1993 return LegalizeKind(TypeWidenVector, NVT); 1994 } 1995 1996 // Examine the element type. 1997 LegalizeKind LK = getTypeConversion(Context, EltVT); 1998 1999 // If type is to be expanded, split the vector. 2000 // <4 x i140> -> <2 x i140> 2001 if (LK.first == TypeExpandInteger) 2002 return LegalizeKind(TypeSplitVector, 2003 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 2004 2005 // Promote the integer element types until a legal vector type is found 2006 // or until the element integer type is too big. If a legal type was not 2007 // found, fallback to the usual mechanism of widening/splitting the 2008 // vector. 2009 while (1) { 2010 // Increase the bitwidth of the element to the next pow-of-two 2011 // (which is greater than 8 bits). 2012 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits() 2013 ).getRoundIntegerType(Context); 2014 2015 // Stop trying when getting a non-simple element type. 2016 // Note that vector elements may be greater than legal vector element 2017 // types. Example: X86 XMM registers hold 64bit element on 32bit systems. 2018 if (!EltVT.isSimple()) break; 2019 2020 // Build a new vector type and check if it is legal. 2021 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 2022 // Found a legal promoted vector type. 2023 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 2024 return LegalizeKind(TypePromoteInteger, 2025 EVT::getVectorVT(Context, EltVT, NumElts)); 2026 } 2027 } 2028 2029 // Try to widen the vector until a legal type is found. 2030 // If there is no wider legal type, split the vector. 2031 while (1) { 2032 // Round up to the next power of 2. 2033 NumElts = (unsigned)NextPowerOf2(NumElts); 2034 2035 // If there is no simple vector type with this many elements then there 2036 // cannot be a larger legal vector type. Note that this assumes that 2037 // there are no skipped intermediate vector types in the simple types. 2038 if (!EltVT.isSimple()) break; 2039 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 2040 if (LargerVector == MVT()) break; 2041 2042 // If this type is legal then widen the vector. 2043 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 2044 return LegalizeKind(TypeWidenVector, LargerVector); 2045 } 2046 2047 // Widen odd vectors to next power of two. 2048 if (!VT.isPow2VectorType()) { 2049 EVT NVT = VT.getPow2VectorType(Context); 2050 return LegalizeKind(TypeWidenVector, NVT); 2051 } 2052 2053 // Vectors with illegal element types are expanded. 2054 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 2055 return LegalizeKind(TypeSplitVector, NVT); 2056 } 2057 2058 std::vector<std::pair<EVT, const TargetRegisterClass*> > AvailableRegClasses; 2059 2060 /// TargetDAGCombineArray - Targets can specify ISD nodes that they would 2061 /// like PerformDAGCombine callbacks for by calling setTargetDAGCombine(), 2062 /// which sets a bit in this array. 2063 unsigned char 2064 TargetDAGCombineArray[(ISD::BUILTIN_OP_END+CHAR_BIT-1)/CHAR_BIT]; 2065 2066 /// PromoteToType - For operations that must be promoted to a specific type, 2067 /// this holds the destination type. This map should be sparse, so don't hold 2068 /// it as an array. 2069 /// 2070 /// Targets add entries to this map with AddPromotedToType(..), clients access 2071 /// this with getTypeToPromoteTo(..). 2072 std::map<std::pair<unsigned, MVT::SimpleValueType>, MVT::SimpleValueType> 2073 PromoteToType; 2074 2075 /// LibcallRoutineNames - Stores the name each libcall. 2076 /// 2077 const char *LibcallRoutineNames[RTLIB::UNKNOWN_LIBCALL]; 2078 2079 /// CmpLibcallCCs - The ISD::CondCode that should be used to test the result 2080 /// of each of the comparison libcall against zero. 2081 ISD::CondCode CmpLibcallCCs[RTLIB::UNKNOWN_LIBCALL]; 2082 2083 /// LibcallCallingConvs - Stores the CallingConv that should be used for each 2084 /// libcall. 2085 CallingConv::ID LibcallCallingConvs[RTLIB::UNKNOWN_LIBCALL]; 2086 2087 protected: 2088 /// When lowering \@llvm.memset this field specifies the maximum number of 2089 /// store operations that may be substituted for the call to memset. Targets 2090 /// must set this value based on the cost threshold for that target. Targets 2091 /// should assume that the memset will be done using as many of the largest 2092 /// store operations first, followed by smaller ones, if necessary, per 2093 /// alignment restrictions. For example, storing 9 bytes on a 32-bit machine 2094 /// with 16-bit alignment would result in four 2-byte stores and one 1-byte 2095 /// store. This only applies to setting a constant array of a constant size. 2096 /// @brief Specify maximum number of store instructions per memset call. 2097 unsigned maxStoresPerMemset; 2098 2099 /// Maximum number of stores operations that may be substituted for the call 2100 /// to memset, used for functions with OptSize attribute. 2101 unsigned maxStoresPerMemsetOptSize; 2102 2103 /// When lowering \@llvm.memcpy this field specifies the maximum number of 2104 /// store operations that may be substituted for a call to memcpy. Targets 2105 /// must set this value based on the cost threshold for that target. Targets 2106 /// should assume that the memcpy will be done using as many of the largest 2107 /// store operations first, followed by smaller ones, if necessary, per 2108 /// alignment restrictions. For example, storing 7 bytes on a 32-bit machine 2109 /// with 32-bit alignment would result in one 4-byte store, a one 2-byte store 2110 /// and one 1-byte store. This only applies to copying a constant array of 2111 /// constant size. 2112 /// @brief Specify maximum bytes of store instructions per memcpy call. 2113 unsigned maxStoresPerMemcpy; 2114 2115 /// Maximum number of store operations that may be substituted for a call 2116 /// to memcpy, used for functions with OptSize attribute. 2117 unsigned maxStoresPerMemcpyOptSize; 2118 2119 /// When lowering \@llvm.memmove this field specifies the maximum number of 2120 /// store instructions that may be substituted for a call to memmove. Targets 2121 /// must set this value based on the cost threshold for that target. Targets 2122 /// should assume that the memmove will be done using as many of the largest 2123 /// store operations first, followed by smaller ones, if necessary, per 2124 /// alignment restrictions. For example, moving 9 bytes on a 32-bit machine 2125 /// with 8-bit alignment would result in nine 1-byte stores. This only 2126 /// applies to copying a constant array of constant size. 2127 /// @brief Specify maximum bytes of store instructions per memmove call. 2128 unsigned maxStoresPerMemmove; 2129 2130 /// Maximum number of store instructions that may be substituted for a call 2131 /// to memmove, used for functions with OpSize attribute. 2132 unsigned maxStoresPerMemmoveOptSize; 2133 2134 /// This field specifies whether the target can benefit from code placement 2135 /// optimization. 2136 bool benefitFromCodePlacementOpt; 2137 2138 /// predictableSelectIsExpensive - Tells the code generator that select is 2139 /// more expensive than a branch if the branch is usually predicted right. 2140 bool predictableSelectIsExpensive; 2141 2142 private: 2143 /// isLegalRC - Return true if the value types that can be represented by the 2144 /// specified register class are all legal. 2145 bool isLegalRC(const TargetRegisterClass *RC) const; 2146 }; 2147 2148 /// GetReturnInfo - Given an LLVM IR type and return type attributes, 2149 /// compute the return value EVTs and flags, and optionally also 2150 /// the offsets, if the return value is being lowered to memory. 2151 void GetReturnInfo(Type* ReturnType, Attributes attr, 2152 SmallVectorImpl<ISD::OutputArg> &Outs, 2153 const TargetLowering &TLI); 2154 2155 } // end llvm namespace 2156 2157 #endif 2158